1 | The following changes since commit e47f81b617684c4546af286d307b69014a83538a: | 1 | The following changes since commit 56f9e46b841c7be478ca038d8d4085d776ab4b0d: |
---|---|---|---|
2 | 2 | ||
3 | Merge remote-tracking branch 'remotes/thibault/tags/samuel-thibault' into staging (2019-02-07 18:53:25 +0000) | 3 | Merge remote-tracking branch 'remotes/armbru/tags/pull-qapi-2017-02-20' into staging (2017-02-20 17:42:47 +0000) |
4 | 4 | ||
5 | are available in the Git repository at: | 5 | are available in the git repository at: |
6 | 6 | ||
7 | git://github.com/stefanha/qemu.git tags/block-pull-request | 7 | git://github.com/stefanha/qemu.git tags/block-pull-request |
8 | 8 | ||
9 | for you to fetch changes up to 55140166667bb555c5d05165b93b25557d2e6397: | 9 | for you to fetch changes up to a7b91d35bab97a2d3e779d0c64c9b837b52a6cf7: |
10 | 10 | ||
11 | tests/virtio-blk: add test for WRITE_ZEROES command (2019-02-11 11:58:17 +0800) | 11 | coroutine-lock: make CoRwlock thread-safe and fair (2017-02-21 11:39:40 +0000) |
12 | 12 | ||
13 | ---------------------------------------------------------------- | 13 | ---------------------------------------------------------------- |
14 | Pull request | 14 | Pull request |
15 | 15 | ||
16 | User-visible changes: | 16 | v2: |
17 | 17 | * Rebased to resolve scsi conflicts | |
18 | * virtio-blk: DISCARD and WRITE_ZEROES support | ||
19 | 18 | ||
20 | ---------------------------------------------------------------- | 19 | ---------------------------------------------------------------- |
21 | 20 | ||
22 | Peter Xu (1): | 21 | Paolo Bonzini (24): |
23 | iothread: fix iothread hang when stop too soon | 22 | block: move AioContext, QEMUTimer, main-loop to libqemuutil |
23 | aio: introduce aio_co_schedule and aio_co_wake | ||
24 | block-backend: allow blk_prw from coroutine context | ||
25 | test-thread-pool: use generic AioContext infrastructure | ||
26 | io: add methods to set I/O handlers on AioContext | ||
27 | io: make qio_channel_yield aware of AioContexts | ||
28 | nbd: convert to use qio_channel_yield | ||
29 | coroutine-lock: reschedule coroutine on the AioContext it was running | ||
30 | on | ||
31 | blkdebug: reschedule coroutine on the AioContext it is running on | ||
32 | qed: introduce qed_aio_start_io and qed_aio_next_io_cb | ||
33 | aio: push aio_context_acquire/release down to dispatching | ||
34 | block: explicitly acquire aiocontext in timers that need it | ||
35 | block: explicitly acquire aiocontext in callbacks that need it | ||
36 | block: explicitly acquire aiocontext in bottom halves that need it | ||
37 | block: explicitly acquire aiocontext in aio callbacks that need it | ||
38 | aio-posix: partially inline aio_dispatch into aio_poll | ||
39 | async: remove unnecessary inc/dec pairs | ||
40 | block: document fields protected by AioContext lock | ||
41 | coroutine-lock: make CoMutex thread-safe | ||
42 | coroutine-lock: add limited spinning to CoMutex | ||
43 | test-aio-multithread: add performance comparison with thread-based | ||
44 | mutexes | ||
45 | coroutine-lock: place CoMutex before CoQueue in header | ||
46 | coroutine-lock: add mutex argument to CoQueue APIs | ||
47 | coroutine-lock: make CoRwlock thread-safe and fair | ||
24 | 48 | ||
25 | Stefano Garzarella (7): | 49 | Makefile.objs | 4 - |
26 | virtio-blk: cleanup using VirtIOBlock *s and VirtIODevice *vdev | 50 | stubs/Makefile.objs | 1 + |
27 | virtio-blk: add acct_failed param to virtio_blk_handle_rw_error() | 51 | tests/Makefile.include | 19 +- |
28 | virtio-blk: add host_features field in VirtIOBlock | 52 | util/Makefile.objs | 6 +- |
29 | virtio-blk: add "discard" and "write-zeroes" properties | 53 | block/nbd-client.h | 2 +- |
30 | virtio-blk: add DISCARD and WRITE_ZEROES features | 54 | block/qed.h | 3 + |
31 | tests/virtio-blk: change assert on data_size in virtio_blk_request() | 55 | include/block/aio.h | 38 ++- |
32 | tests/virtio-blk: add test for WRITE_ZEROES command | 56 | include/block/block_int.h | 64 +++-- |
33 | 57 | include/io/channel.h | 72 +++++- | |
34 | Vladimir Sementsov-Ogievskiy (1): | 58 | include/qemu/coroutine.h | 84 ++++--- |
35 | qemugdb/coroutine: fix arch_prctl has unknown return type | 59 | include/qemu/coroutine_int.h | 11 +- |
36 | 60 | include/sysemu/block-backend.h | 14 +- | |
37 | include/hw/virtio/virtio-blk.h | 5 +- | 61 | tests/iothread.h | 25 ++ |
38 | hw/block/virtio-blk.c | 236 +++++++++++++++++++++++++++++---- | 62 | block/backup.c | 2 +- |
39 | hw/core/machine.c | 2 + | 63 | block/blkdebug.c | 9 +- |
40 | iothread.c | 6 +- | 64 | block/blkreplay.c | 2 +- |
41 | tests/virtio-blk-test.c | 75 ++++++++++- | 65 | block/block-backend.c | 13 +- |
42 | scripts/qemugdb/coroutine.py | 2 +- | 66 | block/curl.c | 44 +++- |
43 | 6 files changed, 297 insertions(+), 29 deletions(-) | 67 | block/gluster.c | 9 +- |
68 | block/io.c | 42 +--- | ||
69 | block/iscsi.c | 15 +- | ||
70 | block/linux-aio.c | 10 +- | ||
71 | block/mirror.c | 12 +- | ||
72 | block/nbd-client.c | 119 +++++---- | ||
73 | block/nfs.c | 9 +- | ||
74 | block/qcow2-cluster.c | 4 +- | ||
75 | block/qed-cluster.c | 2 + | ||
76 | block/qed-table.c | 12 +- | ||
77 | block/qed.c | 58 +++-- | ||
78 | block/sheepdog.c | 31 +-- | ||
79 | block/ssh.c | 29 +-- | ||
80 | block/throttle-groups.c | 4 +- | ||
81 | block/win32-aio.c | 9 +- | ||
82 | dma-helpers.c | 2 + | ||
83 | hw/9pfs/9p.c | 2 +- | ||
84 | hw/block/virtio-blk.c | 19 +- | ||
85 | hw/scsi/scsi-bus.c | 2 + | ||
86 | hw/scsi/scsi-disk.c | 15 ++ | ||
87 | hw/scsi/scsi-generic.c | 20 +- | ||
88 | hw/scsi/virtio-scsi.c | 7 + | ||
89 | io/channel-command.c | 13 + | ||
90 | io/channel-file.c | 11 + | ||
91 | io/channel-socket.c | 16 +- | ||
92 | io/channel-tls.c | 12 + | ||
93 | io/channel-watch.c | 6 + | ||
94 | io/channel.c | 97 ++++++-- | ||
95 | nbd/client.c | 2 +- | ||
96 | nbd/common.c | 9 +- | ||
97 | nbd/server.c | 94 +++----- | ||
98 | stubs/linux-aio.c | 32 +++ | ||
99 | stubs/set-fd-handler.c | 11 - | ||
100 | tests/iothread.c | 91 +++++++ | ||
101 | tests/test-aio-multithread.c | 463 ++++++++++++++++++++++++++++++++++++ | ||
102 | tests/test-thread-pool.c | 12 +- | ||
103 | aio-posix.c => util/aio-posix.c | 62 ++--- | ||
104 | aio-win32.c => util/aio-win32.c | 30 +-- | ||
105 | util/aiocb.c | 55 +++++ | ||
106 | async.c => util/async.c | 84 ++++++- | ||
107 | iohandler.c => util/iohandler.c | 0 | ||
108 | main-loop.c => util/main-loop.c | 0 | ||
109 | util/qemu-coroutine-lock.c | 254 ++++++++++++++++++-- | ||
110 | util/qemu-coroutine-sleep.c | 2 +- | ||
111 | util/qemu-coroutine.c | 8 + | ||
112 | qemu-timer.c => util/qemu-timer.c | 0 | ||
113 | thread-pool.c => util/thread-pool.c | 8 +- | ||
114 | trace-events | 11 - | ||
115 | util/trace-events | 17 +- | ||
116 | 67 files changed, 1712 insertions(+), 533 deletions(-) | ||
117 | create mode 100644 tests/iothread.h | ||
118 | create mode 100644 stubs/linux-aio.c | ||
119 | create mode 100644 tests/iothread.c | ||
120 | create mode 100644 tests/test-aio-multithread.c | ||
121 | rename aio-posix.c => util/aio-posix.c (94%) | ||
122 | rename aio-win32.c => util/aio-win32.c (95%) | ||
123 | create mode 100644 util/aiocb.c | ||
124 | rename async.c => util/async.c (82%) | ||
125 | rename iohandler.c => util/iohandler.c (100%) | ||
126 | rename main-loop.c => util/main-loop.c (100%) | ||
127 | rename qemu-timer.c => util/qemu-timer.c (100%) | ||
128 | rename thread-pool.c => util/thread-pool.c (97%) | ||
44 | 129 | ||
45 | -- | 130 | -- |
46 | 2.20.1 | 131 | 2.9.3 |
47 | 132 | ||
48 | 133 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | 1 | From: Paolo Bonzini <pbonzini@redhat.com> | |
2 | |||
3 | AioContext is fairly self contained, the only dependency is QEMUTimer but | ||
4 | that in turn doesn't need anything else. So move them out of block-obj-y | ||
5 | to avoid introducing a dependency from io/ to block-obj-y. | ||
6 | |||
7 | main-loop and its dependency iohandler also need to be moved, because | ||
8 | later in this series io/ will call iohandler_get_aio_context. | ||
9 | |||
10 | [Changed copyright "the QEMU team" to "other QEMU contributors" as | ||
11 | suggested by Daniel Berrange and agreed by Paolo. | ||
12 | --Stefan] | ||
13 | |||
14 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
15 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
16 | Message-id: 20170213135235.12274-2-pbonzini@redhat.com | ||
17 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
18 | --- | ||
19 | Makefile.objs | 4 --- | ||
20 | stubs/Makefile.objs | 1 + | ||
21 | tests/Makefile.include | 11 ++++---- | ||
22 | util/Makefile.objs | 6 +++- | ||
23 | block/io.c | 29 ------------------- | ||
24 | stubs/linux-aio.c | 32 +++++++++++++++++++++ | ||
25 | stubs/set-fd-handler.c | 11 -------- | ||
26 | aio-posix.c => util/aio-posix.c | 2 +- | ||
27 | aio-win32.c => util/aio-win32.c | 0 | ||
28 | util/aiocb.c | 55 +++++++++++++++++++++++++++++++++++++ | ||
29 | async.c => util/async.c | 3 +- | ||
30 | iohandler.c => util/iohandler.c | 0 | ||
31 | main-loop.c => util/main-loop.c | 0 | ||
32 | qemu-timer.c => util/qemu-timer.c | 0 | ||
33 | thread-pool.c => util/thread-pool.c | 2 +- | ||
34 | trace-events | 11 -------- | ||
35 | util/trace-events | 11 ++++++++ | ||
36 | 17 files changed, 114 insertions(+), 64 deletions(-) | ||
37 | create mode 100644 stubs/linux-aio.c | ||
38 | rename aio-posix.c => util/aio-posix.c (99%) | ||
39 | rename aio-win32.c => util/aio-win32.c (100%) | ||
40 | create mode 100644 util/aiocb.c | ||
41 | rename async.c => util/async.c (99%) | ||
42 | rename iohandler.c => util/iohandler.c (100%) | ||
43 | rename main-loop.c => util/main-loop.c (100%) | ||
44 | rename qemu-timer.c => util/qemu-timer.c (100%) | ||
45 | rename thread-pool.c => util/thread-pool.c (99%) | ||
46 | |||
47 | diff --git a/Makefile.objs b/Makefile.objs | ||
48 | index XXXXXXX..XXXXXXX 100644 | ||
49 | --- a/Makefile.objs | ||
50 | +++ b/Makefile.objs | ||
51 | @@ -XXX,XX +XXX,XX @@ chardev-obj-y = chardev/ | ||
52 | ####################################################################### | ||
53 | # block-obj-y is code used by both qemu system emulation and qemu-img | ||
54 | |||
55 | -block-obj-y = async.o thread-pool.o | ||
56 | block-obj-y += nbd/ | ||
57 | block-obj-y += block.o blockjob.o | ||
58 | -block-obj-y += main-loop.o iohandler.o qemu-timer.o | ||
59 | -block-obj-$(CONFIG_POSIX) += aio-posix.o | ||
60 | -block-obj-$(CONFIG_WIN32) += aio-win32.o | ||
61 | block-obj-y += block/ | ||
62 | block-obj-y += qemu-io-cmds.o | ||
63 | block-obj-$(CONFIG_REPLICATION) += replication.o | ||
64 | diff --git a/stubs/Makefile.objs b/stubs/Makefile.objs | ||
65 | index XXXXXXX..XXXXXXX 100644 | ||
66 | --- a/stubs/Makefile.objs | ||
67 | +++ b/stubs/Makefile.objs | ||
68 | @@ -XXX,XX +XXX,XX @@ stub-obj-y += get-vm-name.o | ||
69 | stub-obj-y += iothread.o | ||
70 | stub-obj-y += iothread-lock.o | ||
71 | stub-obj-y += is-daemonized.o | ||
72 | +stub-obj-$(CONFIG_LINUX_AIO) += linux-aio.o | ||
73 | stub-obj-y += machine-init-done.o | ||
74 | stub-obj-y += migr-blocker.o | ||
75 | stub-obj-y += monitor.o | ||
76 | diff --git a/tests/Makefile.include b/tests/Makefile.include | ||
77 | index XXXXXXX..XXXXXXX 100644 | ||
78 | --- a/tests/Makefile.include | ||
79 | +++ b/tests/Makefile.include | ||
80 | @@ -XXX,XX +XXX,XX @@ check-unit-y += tests/test-visitor-serialization$(EXESUF) | ||
81 | check-unit-y += tests/test-iov$(EXESUF) | ||
82 | gcov-files-test-iov-y = util/iov.c | ||
83 | check-unit-y += tests/test-aio$(EXESUF) | ||
84 | +gcov-files-test-aio-y = util/async.c util/qemu-timer.o | ||
85 | +gcov-files-test-aio-$(CONFIG_WIN32) += util/aio-win32.c | ||
86 | +gcov-files-test-aio-$(CONFIG_POSIX) += util/aio-posix.c | ||
87 | check-unit-y += tests/test-throttle$(EXESUF) | ||
88 | gcov-files-test-aio-$(CONFIG_WIN32) = aio-win32.c | ||
89 | gcov-files-test-aio-$(CONFIG_POSIX) = aio-posix.c | ||
90 | @@ -XXX,XX +XXX,XX @@ tests/check-qjson$(EXESUF): tests/check-qjson.o $(test-util-obj-y) | ||
91 | tests/check-qom-interface$(EXESUF): tests/check-qom-interface.o $(test-qom-obj-y) | ||
92 | tests/check-qom-proplist$(EXESUF): tests/check-qom-proplist.o $(test-qom-obj-y) | ||
93 | |||
94 | -tests/test-char$(EXESUF): tests/test-char.o qemu-timer.o \ | ||
95 | - $(test-util-obj-y) $(qtest-obj-y) $(test-block-obj-y) $(chardev-obj-y) | ||
96 | +tests/test-char$(EXESUF): tests/test-char.o $(test-util-obj-y) $(qtest-obj-y) $(test-io-obj-y) $(chardev-obj-y) | ||
97 | tests/test-coroutine$(EXESUF): tests/test-coroutine.o $(test-block-obj-y) | ||
98 | tests/test-aio$(EXESUF): tests/test-aio.o $(test-block-obj-y) | ||
99 | tests/test-throttle$(EXESUF): tests/test-throttle.o $(test-block-obj-y) | ||
100 | @@ -XXX,XX +XXX,XX @@ tests/test-vmstate$(EXESUF): tests/test-vmstate.o \ | ||
101 | migration/vmstate.o migration/qemu-file.o \ | ||
102 | migration/qemu-file-channel.o migration/qjson.o \ | ||
103 | $(test-io-obj-y) | ||
104 | -tests/test-timed-average$(EXESUF): tests/test-timed-average.o qemu-timer.o \ | ||
105 | - $(test-util-obj-y) | ||
106 | +tests/test-timed-average$(EXESUF): tests/test-timed-average.o $(test-util-obj-y) | ||
107 | tests/test-base64$(EXESUF): tests/test-base64.o \ | ||
108 | libqemuutil.a libqemustub.a | ||
109 | tests/ptimer-test$(EXESUF): tests/ptimer-test.o tests/ptimer-test-stubs.o hw/core/ptimer.o libqemustub.a | ||
110 | @@ -XXX,XX +XXX,XX @@ tests/usb-hcd-ehci-test$(EXESUF): tests/usb-hcd-ehci-test.o $(libqos-usb-obj-y) | ||
111 | tests/usb-hcd-xhci-test$(EXESUF): tests/usb-hcd-xhci-test.o $(libqos-usb-obj-y) | ||
112 | tests/pc-cpu-test$(EXESUF): tests/pc-cpu-test.o | ||
113 | tests/postcopy-test$(EXESUF): tests/postcopy-test.o | ||
114 | -tests/vhost-user-test$(EXESUF): tests/vhost-user-test.o qemu-timer.o \ | ||
115 | +tests/vhost-user-test$(EXESUF): tests/vhost-user-test.o $(test-util-obj-y) \ | ||
116 | $(qtest-obj-y) $(test-io-obj-y) $(libqos-virtio-obj-y) $(libqos-pc-obj-y) \ | ||
117 | $(chardev-obj-y) | ||
118 | tests/qemu-iotests/socket_scm_helper$(EXESUF): tests/qemu-iotests/socket_scm_helper.o | ||
119 | diff --git a/util/Makefile.objs b/util/Makefile.objs | ||
120 | index XXXXXXX..XXXXXXX 100644 | ||
121 | --- a/util/Makefile.objs | ||
122 | +++ b/util/Makefile.objs | ||
123 | @@ -XXX,XX +XXX,XX @@ | ||
124 | util-obj-y = osdep.o cutils.o unicode.o qemu-timer-common.o | ||
125 | util-obj-y += bufferiszero.o | ||
126 | util-obj-y += lockcnt.o | ||
127 | +util-obj-y += aiocb.o async.o thread-pool.o qemu-timer.o | ||
128 | +util-obj-y += main-loop.o iohandler.o | ||
129 | +util-obj-$(CONFIG_POSIX) += aio-posix.o | ||
130 | util-obj-$(CONFIG_POSIX) += compatfd.o | ||
131 | util-obj-$(CONFIG_POSIX) += event_notifier-posix.o | ||
132 | util-obj-$(CONFIG_POSIX) += mmap-alloc.o | ||
133 | util-obj-$(CONFIG_POSIX) += oslib-posix.o | ||
134 | util-obj-$(CONFIG_POSIX) += qemu-openpty.o | ||
135 | util-obj-$(CONFIG_POSIX) += qemu-thread-posix.o | ||
136 | -util-obj-$(CONFIG_WIN32) += event_notifier-win32.o | ||
137 | util-obj-$(CONFIG_POSIX) += memfd.o | ||
138 | +util-obj-$(CONFIG_WIN32) += aio-win32.o | ||
139 | +util-obj-$(CONFIG_WIN32) += event_notifier-win32.o | ||
140 | util-obj-$(CONFIG_WIN32) += oslib-win32.o | ||
141 | util-obj-$(CONFIG_WIN32) += qemu-thread-win32.o | ||
142 | util-obj-y += envlist.o path.o module.o | ||
143 | diff --git a/block/io.c b/block/io.c | ||
144 | index XXXXXXX..XXXXXXX 100644 | ||
145 | --- a/block/io.c | ||
146 | +++ b/block/io.c | ||
147 | @@ -XXX,XX +XXX,XX @@ BlockAIOCB *bdrv_aio_flush(BlockDriverState *bs, | ||
148 | return &acb->common; | ||
149 | } | ||
150 | |||
151 | -void *qemu_aio_get(const AIOCBInfo *aiocb_info, BlockDriverState *bs, | ||
152 | - BlockCompletionFunc *cb, void *opaque) | ||
153 | -{ | ||
154 | - BlockAIOCB *acb; | ||
155 | - | ||
156 | - acb = g_malloc(aiocb_info->aiocb_size); | ||
157 | - acb->aiocb_info = aiocb_info; | ||
158 | - acb->bs = bs; | ||
159 | - acb->cb = cb; | ||
160 | - acb->opaque = opaque; | ||
161 | - acb->refcnt = 1; | ||
162 | - return acb; | ||
163 | -} | ||
164 | - | ||
165 | -void qemu_aio_ref(void *p) | ||
166 | -{ | ||
167 | - BlockAIOCB *acb = p; | ||
168 | - acb->refcnt++; | ||
169 | -} | ||
170 | - | ||
171 | -void qemu_aio_unref(void *p) | ||
172 | -{ | ||
173 | - BlockAIOCB *acb = p; | ||
174 | - assert(acb->refcnt > 0); | ||
175 | - if (--acb->refcnt == 0) { | ||
176 | - g_free(acb); | ||
177 | - } | ||
178 | -} | ||
179 | - | ||
180 | /**************************************************************/ | ||
181 | /* Coroutine block device emulation */ | ||
182 | |||
183 | diff --git a/stubs/linux-aio.c b/stubs/linux-aio.c | ||
184 | new file mode 100644 | ||
185 | index XXXXXXX..XXXXXXX | ||
186 | --- /dev/null | ||
187 | +++ b/stubs/linux-aio.c | ||
188 | @@ -XXX,XX +XXX,XX @@ | ||
189 | +/* | ||
190 | + * Linux native AIO support. | ||
191 | + * | ||
192 | + * Copyright (C) 2009 IBM, Corp. | ||
193 | + * Copyright (C) 2009 Red Hat, Inc. | ||
194 | + * | ||
195 | + * This work is licensed under the terms of the GNU GPL, version 2 or later. | ||
196 | + * See the COPYING file in the top-level directory. | ||
197 | + */ | ||
198 | +#include "qemu/osdep.h" | ||
199 | +#include "block/aio.h" | ||
200 | +#include "block/raw-aio.h" | ||
201 | + | ||
202 | +void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context) | ||
203 | +{ | ||
204 | + abort(); | ||
205 | +} | ||
206 | + | ||
207 | +void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context) | ||
208 | +{ | ||
209 | + abort(); | ||
210 | +} | ||
211 | + | ||
212 | +LinuxAioState *laio_init(void) | ||
213 | +{ | ||
214 | + abort(); | ||
215 | +} | ||
216 | + | ||
217 | +void laio_cleanup(LinuxAioState *s) | ||
218 | +{ | ||
219 | + abort(); | ||
220 | +} | ||
221 | diff --git a/stubs/set-fd-handler.c b/stubs/set-fd-handler.c | ||
222 | index XXXXXXX..XXXXXXX 100644 | ||
223 | --- a/stubs/set-fd-handler.c | ||
224 | +++ b/stubs/set-fd-handler.c | ||
225 | @@ -XXX,XX +XXX,XX @@ void qemu_set_fd_handler(int fd, | ||
226 | { | ||
227 | abort(); | ||
228 | } | ||
229 | - | ||
230 | -void aio_set_fd_handler(AioContext *ctx, | ||
231 | - int fd, | ||
232 | - bool is_external, | ||
233 | - IOHandler *io_read, | ||
234 | - IOHandler *io_write, | ||
235 | - AioPollFn *io_poll, | ||
236 | - void *opaque) | ||
237 | -{ | ||
238 | - abort(); | ||
239 | -} | ||
240 | diff --git a/aio-posix.c b/util/aio-posix.c | ||
241 | similarity index 99% | ||
242 | rename from aio-posix.c | ||
243 | rename to util/aio-posix.c | ||
244 | index XXXXXXX..XXXXXXX 100644 | ||
245 | --- a/aio-posix.c | ||
246 | +++ b/util/aio-posix.c | ||
247 | @@ -XXX,XX +XXX,XX @@ | ||
248 | #include "qemu/rcu_queue.h" | ||
249 | #include "qemu/sockets.h" | ||
250 | #include "qemu/cutils.h" | ||
251 | -#include "trace-root.h" | ||
252 | +#include "trace.h" | ||
253 | #ifdef CONFIG_EPOLL_CREATE1 | ||
254 | #include <sys/epoll.h> | ||
255 | #endif | ||
256 | diff --git a/aio-win32.c b/util/aio-win32.c | ||
257 | similarity index 100% | ||
258 | rename from aio-win32.c | ||
259 | rename to util/aio-win32.c | ||
260 | diff --git a/util/aiocb.c b/util/aiocb.c | ||
261 | new file mode 100644 | ||
262 | index XXXXXXX..XXXXXXX | ||
263 | --- /dev/null | ||
264 | +++ b/util/aiocb.c | ||
265 | @@ -XXX,XX +XXX,XX @@ | ||
266 | +/* | ||
267 | + * BlockAIOCB allocation | ||
268 | + * | ||
269 | + * Copyright (c) 2003-2017 Fabrice Bellard and other QEMU contributors | ||
270 | + * | ||
271 | + * Permission is hereby granted, free of charge, to any person obtaining a copy | ||
272 | + * of this software and associated documentation files (the "Software"), to deal | ||
273 | + * in the Software without restriction, including without limitation the rights | ||
274 | + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell | ||
275 | + * copies of the Software, and to permit persons to whom the Software is | ||
276 | + * furnished to do so, subject to the following conditions: | ||
277 | + * | ||
278 | + * The above copyright notice and this permission notice shall be included in | ||
279 | + * all copies or substantial portions of the Software. | ||
280 | + * | ||
281 | + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | ||
282 | + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | ||
283 | + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | ||
284 | + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | ||
285 | + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, | ||
286 | + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN | ||
287 | + * THE SOFTWARE. | ||
288 | + */ | ||
289 | + | ||
290 | +#include "qemu/osdep.h" | ||
291 | +#include "block/aio.h" | ||
292 | + | ||
293 | +void *qemu_aio_get(const AIOCBInfo *aiocb_info, BlockDriverState *bs, | ||
294 | + BlockCompletionFunc *cb, void *opaque) | ||
295 | +{ | ||
296 | + BlockAIOCB *acb; | ||
297 | + | ||
298 | + acb = g_malloc(aiocb_info->aiocb_size); | ||
299 | + acb->aiocb_info = aiocb_info; | ||
300 | + acb->bs = bs; | ||
301 | + acb->cb = cb; | ||
302 | + acb->opaque = opaque; | ||
303 | + acb->refcnt = 1; | ||
304 | + return acb; | ||
305 | +} | ||
306 | + | ||
307 | +void qemu_aio_ref(void *p) | ||
308 | +{ | ||
309 | + BlockAIOCB *acb = p; | ||
310 | + acb->refcnt++; | ||
311 | +} | ||
312 | + | ||
313 | +void qemu_aio_unref(void *p) | ||
314 | +{ | ||
315 | + BlockAIOCB *acb = p; | ||
316 | + assert(acb->refcnt > 0); | ||
317 | + if (--acb->refcnt == 0) { | ||
318 | + g_free(acb); | ||
319 | + } | ||
320 | +} | ||
321 | diff --git a/async.c b/util/async.c | ||
322 | similarity index 99% | ||
323 | rename from async.c | ||
324 | rename to util/async.c | ||
325 | index XXXXXXX..XXXXXXX 100644 | ||
326 | --- a/async.c | ||
327 | +++ b/util/async.c | ||
328 | @@ -XXX,XX +XXX,XX @@ | ||
329 | /* | ||
330 | - * QEMU System Emulator | ||
331 | + * Data plane event loop | ||
332 | * | ||
333 | * Copyright (c) 2003-2008 Fabrice Bellard | ||
334 | + * Copyright (c) 2009-2017 QEMU contributors | ||
335 | * | ||
336 | * Permission is hereby granted, free of charge, to any person obtaining a copy | ||
337 | * of this software and associated documentation files (the "Software"), to deal | ||
338 | diff --git a/iohandler.c b/util/iohandler.c | ||
339 | similarity index 100% | ||
340 | rename from iohandler.c | ||
341 | rename to util/iohandler.c | ||
342 | diff --git a/main-loop.c b/util/main-loop.c | ||
343 | similarity index 100% | ||
344 | rename from main-loop.c | ||
345 | rename to util/main-loop.c | ||
346 | diff --git a/qemu-timer.c b/util/qemu-timer.c | ||
347 | similarity index 100% | ||
348 | rename from qemu-timer.c | ||
349 | rename to util/qemu-timer.c | ||
350 | diff --git a/thread-pool.c b/util/thread-pool.c | ||
351 | similarity index 99% | ||
352 | rename from thread-pool.c | ||
353 | rename to util/thread-pool.c | ||
354 | index XXXXXXX..XXXXXXX 100644 | ||
355 | --- a/thread-pool.c | ||
356 | +++ b/util/thread-pool.c | ||
357 | @@ -XXX,XX +XXX,XX @@ | ||
358 | #include "qemu/queue.h" | ||
359 | #include "qemu/thread.h" | ||
360 | #include "qemu/coroutine.h" | ||
361 | -#include "trace-root.h" | ||
362 | +#include "trace.h" | ||
363 | #include "block/thread-pool.h" | ||
364 | #include "qemu/main-loop.h" | ||
365 | |||
366 | diff --git a/trace-events b/trace-events | ||
367 | index XXXXXXX..XXXXXXX 100644 | ||
368 | --- a/trace-events | ||
369 | +++ b/trace-events | ||
370 | @@ -XXX,XX +XXX,XX @@ | ||
371 | # | ||
372 | # The <format-string> should be a sprintf()-compatible format string. | ||
373 | |||
374 | -# aio-posix.c | ||
375 | -run_poll_handlers_begin(void *ctx, int64_t max_ns) "ctx %p max_ns %"PRId64 | ||
376 | -run_poll_handlers_end(void *ctx, bool progress) "ctx %p progress %d" | ||
377 | -poll_shrink(void *ctx, int64_t old, int64_t new) "ctx %p old %"PRId64" new %"PRId64 | ||
378 | -poll_grow(void *ctx, int64_t old, int64_t new) "ctx %p old %"PRId64" new %"PRId64 | ||
379 | - | ||
380 | -# thread-pool.c | ||
381 | -thread_pool_submit(void *pool, void *req, void *opaque) "pool %p req %p opaque %p" | ||
382 | -thread_pool_complete(void *pool, void *req, void *opaque, int ret) "pool %p req %p opaque %p ret %d" | ||
383 | -thread_pool_cancel(void *req, void *opaque) "req %p opaque %p" | ||
384 | - | ||
385 | # ioport.c | ||
386 | cpu_in(unsigned int addr, char size, unsigned int val) "addr %#x(%c) value %u" | ||
387 | cpu_out(unsigned int addr, char size, unsigned int val) "addr %#x(%c) value %u" | ||
388 | diff --git a/util/trace-events b/util/trace-events | ||
389 | index XXXXXXX..XXXXXXX 100644 | ||
390 | --- a/util/trace-events | ||
391 | +++ b/util/trace-events | ||
392 | @@ -XXX,XX +XXX,XX @@ | ||
393 | # See docs/tracing.txt for syntax documentation. | ||
394 | |||
395 | +# util/aio-posix.c | ||
396 | +run_poll_handlers_begin(void *ctx, int64_t max_ns) "ctx %p max_ns %"PRId64 | ||
397 | +run_poll_handlers_end(void *ctx, bool progress) "ctx %p progress %d" | ||
398 | +poll_shrink(void *ctx, int64_t old, int64_t new) "ctx %p old %"PRId64" new %"PRId64 | ||
399 | +poll_grow(void *ctx, int64_t old, int64_t new) "ctx %p old %"PRId64" new %"PRId64 | ||
400 | + | ||
401 | +# util/thread-pool.c | ||
402 | +thread_pool_submit(void *pool, void *req, void *opaque) "pool %p req %p opaque %p" | ||
403 | +thread_pool_complete(void *pool, void *req, void *opaque, int ret) "pool %p req %p opaque %p ret %d" | ||
404 | +thread_pool_cancel(void *req, void *opaque) "req %p opaque %p" | ||
405 | + | ||
406 | # util/buffer.c | ||
407 | buffer_resize(const char *buf, size_t olen, size_t len) "%s: old %zd, new %zd" | ||
408 | buffer_move_empty(const char *buf, size_t len, const char *from) "%s: %zd bytes from %s" | ||
409 | -- | ||
410 | 2.9.3 | ||
411 | |||
412 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | 1 | From: Paolo Bonzini <pbonzini@redhat.com> | |
2 | |||
3 | aio_co_wake provides the infrastructure to start a coroutine on a "home" | ||
4 | AioContext. It will be used by CoMutex and CoQueue, so that coroutines | ||
5 | don't jump from one context to another when they go to sleep on a | ||
6 | mutex or waitqueue. However, it can also be used as a more efficient | ||
7 | alternative to one-shot bottom halves, and saves the effort of tracking | ||
8 | which AioContext a coroutine is running on. | ||
9 | |||
10 | aio_co_schedule is the part of aio_co_wake that starts a coroutine | ||
11 | on a remove AioContext, but it is also useful to implement e.g. | ||
12 | bdrv_set_aio_context callbacks. | ||
13 | |||
14 | The implementation of aio_co_schedule is based on a lock-free | ||
15 | multiple-producer, single-consumer queue. The multiple producers use | ||
16 | cmpxchg to add to a LIFO stack. The consumer (a per-AioContext bottom | ||
17 | half) grabs all items added so far, inverts the list to make it FIFO, | ||
18 | and goes through it one item at a time until it's empty. The data | ||
19 | structure was inspired by OSv, which uses it in the very code we'll | ||
20 | "port" to QEMU for the thread-safe CoMutex. | ||
21 | |||
22 | Most of the new code is really tests. | ||
23 | |||
24 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
25 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
26 | Message-id: 20170213135235.12274-3-pbonzini@redhat.com | ||
27 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
28 | --- | ||
29 | tests/Makefile.include | 8 +- | ||
30 | include/block/aio.h | 32 +++++++ | ||
31 | include/qemu/coroutine_int.h | 11 ++- | ||
32 | tests/iothread.h | 25 +++++ | ||
33 | tests/iothread.c | 91 ++++++++++++++++++ | ||
34 | tests/test-aio-multithread.c | 213 +++++++++++++++++++++++++++++++++++++++++++ | ||
35 | util/async.c | 65 +++++++++++++ | ||
36 | util/qemu-coroutine.c | 8 ++ | ||
37 | util/trace-events | 4 + | ||
38 | 9 files changed, 453 insertions(+), 4 deletions(-) | ||
39 | create mode 100644 tests/iothread.h | ||
40 | create mode 100644 tests/iothread.c | ||
41 | create mode 100644 tests/test-aio-multithread.c | ||
42 | |||
43 | diff --git a/tests/Makefile.include b/tests/Makefile.include | ||
44 | index XXXXXXX..XXXXXXX 100644 | ||
45 | --- a/tests/Makefile.include | ||
46 | +++ b/tests/Makefile.include | ||
47 | @@ -XXX,XX +XXX,XX @@ check-unit-y += tests/test-aio$(EXESUF) | ||
48 | gcov-files-test-aio-y = util/async.c util/qemu-timer.o | ||
49 | gcov-files-test-aio-$(CONFIG_WIN32) += util/aio-win32.c | ||
50 | gcov-files-test-aio-$(CONFIG_POSIX) += util/aio-posix.c | ||
51 | +check-unit-y += tests/test-aio-multithread$(EXESUF) | ||
52 | +gcov-files-test-aio-multithread-y = $(gcov-files-test-aio-y) | ||
53 | +gcov-files-test-aio-multithread-y += util/qemu-coroutine.c tests/iothread.c | ||
54 | check-unit-y += tests/test-throttle$(EXESUF) | ||
55 | -gcov-files-test-aio-$(CONFIG_WIN32) = aio-win32.c | ||
56 | -gcov-files-test-aio-$(CONFIG_POSIX) = aio-posix.c | ||
57 | check-unit-y += tests/test-thread-pool$(EXESUF) | ||
58 | gcov-files-test-thread-pool-y = thread-pool.c | ||
59 | gcov-files-test-hbitmap-y = util/hbitmap.c | ||
60 | @@ -XXX,XX +XXX,XX @@ test-qapi-obj-y = tests/test-qapi-visit.o tests/test-qapi-types.o \ | ||
61 | $(test-qom-obj-y) | ||
62 | test-crypto-obj-y = $(crypto-obj-y) $(test-qom-obj-y) | ||
63 | test-io-obj-y = $(io-obj-y) $(test-crypto-obj-y) | ||
64 | -test-block-obj-y = $(block-obj-y) $(test-io-obj-y) | ||
65 | +test-block-obj-y = $(block-obj-y) $(test-io-obj-y) tests/iothread.o | ||
66 | |||
67 | tests/check-qint$(EXESUF): tests/check-qint.o $(test-util-obj-y) | ||
68 | tests/check-qstring$(EXESUF): tests/check-qstring.o $(test-util-obj-y) | ||
69 | @@ -XXX,XX +XXX,XX @@ tests/check-qom-proplist$(EXESUF): tests/check-qom-proplist.o $(test-qom-obj-y) | ||
70 | tests/test-char$(EXESUF): tests/test-char.o $(test-util-obj-y) $(qtest-obj-y) $(test-io-obj-y) $(chardev-obj-y) | ||
71 | tests/test-coroutine$(EXESUF): tests/test-coroutine.o $(test-block-obj-y) | ||
72 | tests/test-aio$(EXESUF): tests/test-aio.o $(test-block-obj-y) | ||
73 | +tests/test-aio-multithread$(EXESUF): tests/test-aio-multithread.o $(test-block-obj-y) | ||
74 | tests/test-throttle$(EXESUF): tests/test-throttle.o $(test-block-obj-y) | ||
75 | tests/test-blockjob$(EXESUF): tests/test-blockjob.o $(test-block-obj-y) $(test-util-obj-y) | ||
76 | tests/test-blockjob-txn$(EXESUF): tests/test-blockjob-txn.o $(test-block-obj-y) $(test-util-obj-y) | ||
77 | diff --git a/include/block/aio.h b/include/block/aio.h | ||
78 | index XXXXXXX..XXXXXXX 100644 | ||
79 | --- a/include/block/aio.h | ||
80 | +++ b/include/block/aio.h | ||
81 | @@ -XXX,XX +XXX,XX @@ typedef void QEMUBHFunc(void *opaque); | ||
82 | typedef bool AioPollFn(void *opaque); | ||
83 | typedef void IOHandler(void *opaque); | ||
84 | |||
85 | +struct Coroutine; | ||
86 | struct ThreadPool; | ||
87 | struct LinuxAioState; | ||
88 | |||
89 | @@ -XXX,XX +XXX,XX @@ struct AioContext { | ||
90 | bool notified; | ||
91 | EventNotifier notifier; | ||
92 | |||
93 | + QSLIST_HEAD(, Coroutine) scheduled_coroutines; | ||
94 | + QEMUBH *co_schedule_bh; | ||
95 | + | ||
96 | /* Thread pool for performing work and receiving completion callbacks. | ||
97 | * Has its own locking. | ||
98 | */ | ||
99 | @@ -XXX,XX +XXX,XX @@ static inline bool aio_node_check(AioContext *ctx, bool is_external) | ||
100 | } | ||
101 | |||
102 | /** | ||
103 | + * aio_co_schedule: | ||
104 | + * @ctx: the aio context | ||
105 | + * @co: the coroutine | ||
106 | + * | ||
107 | + * Start a coroutine on a remote AioContext. | ||
108 | + * | ||
109 | + * The coroutine must not be entered by anyone else while aio_co_schedule() | ||
110 | + * is active. In addition the coroutine must have yielded unless ctx | ||
111 | + * is the context in which the coroutine is running (i.e. the value of | ||
112 | + * qemu_get_current_aio_context() from the coroutine itself). | ||
113 | + */ | ||
114 | +void aio_co_schedule(AioContext *ctx, struct Coroutine *co); | ||
115 | + | ||
116 | +/** | ||
117 | + * aio_co_wake: | ||
118 | + * @co: the coroutine | ||
119 | + * | ||
120 | + * Restart a coroutine on the AioContext where it was running last, thus | ||
121 | + * preventing coroutines from jumping from one context to another when they | ||
122 | + * go to sleep. | ||
123 | + * | ||
124 | + * aio_co_wake may be executed either in coroutine or non-coroutine | ||
125 | + * context. The coroutine must not be entered by anyone else while | ||
126 | + * aio_co_wake() is active. | ||
127 | + */ | ||
128 | +void aio_co_wake(struct Coroutine *co); | ||
129 | + | ||
130 | +/** | ||
131 | * Return the AioContext whose event loop runs in the current thread. | ||
132 | * | ||
133 | * If called from an IOThread this will be the IOThread's AioContext. If | ||
134 | diff --git a/include/qemu/coroutine_int.h b/include/qemu/coroutine_int.h | ||
135 | index XXXXXXX..XXXXXXX 100644 | ||
136 | --- a/include/qemu/coroutine_int.h | ||
137 | +++ b/include/qemu/coroutine_int.h | ||
138 | @@ -XXX,XX +XXX,XX @@ struct Coroutine { | ||
139 | CoroutineEntry *entry; | ||
140 | void *entry_arg; | ||
141 | Coroutine *caller; | ||
142 | + | ||
143 | + /* Only used when the coroutine has terminated. */ | ||
144 | QSLIST_ENTRY(Coroutine) pool_next; | ||
145 | + | ||
146 | size_t locks_held; | ||
147 | |||
148 | - /* Coroutines that should be woken up when we yield or terminate */ | ||
149 | + /* Coroutines that should be woken up when we yield or terminate. | ||
150 | + * Only used when the coroutine is running. | ||
151 | + */ | ||
152 | QSIMPLEQ_HEAD(, Coroutine) co_queue_wakeup; | ||
153 | + | ||
154 | + /* Only used when the coroutine has yielded. */ | ||
155 | + AioContext *ctx; | ||
156 | QSIMPLEQ_ENTRY(Coroutine) co_queue_next; | ||
157 | + QSLIST_ENTRY(Coroutine) co_scheduled_next; | ||
158 | }; | ||
159 | |||
160 | Coroutine *qemu_coroutine_new(void); | ||
161 | diff --git a/tests/iothread.h b/tests/iothread.h | ||
162 | new file mode 100644 | ||
163 | index XXXXXXX..XXXXXXX | ||
164 | --- /dev/null | ||
165 | +++ b/tests/iothread.h | ||
166 | @@ -XXX,XX +XXX,XX @@ | ||
167 | +/* | ||
168 | + * Event loop thread implementation for unit tests | ||
169 | + * | ||
170 | + * Copyright Red Hat Inc., 2013, 2016 | ||
171 | + * | ||
172 | + * Authors: | ||
173 | + * Stefan Hajnoczi <stefanha@redhat.com> | ||
174 | + * Paolo Bonzini <pbonzini@redhat.com> | ||
175 | + * | ||
176 | + * This work is licensed under the terms of the GNU GPL, version 2 or later. | ||
177 | + * See the COPYING file in the top-level directory. | ||
178 | + */ | ||
179 | +#ifndef TEST_IOTHREAD_H | ||
180 | +#define TEST_IOTHREAD_H | ||
181 | + | ||
182 | +#include "block/aio.h" | ||
183 | +#include "qemu/thread.h" | ||
184 | + | ||
185 | +typedef struct IOThread IOThread; | ||
186 | + | ||
187 | +IOThread *iothread_new(void); | ||
188 | +void iothread_join(IOThread *iothread); | ||
189 | +AioContext *iothread_get_aio_context(IOThread *iothread); | ||
190 | + | ||
191 | +#endif | ||
192 | diff --git a/tests/iothread.c b/tests/iothread.c | ||
193 | new file mode 100644 | ||
194 | index XXXXXXX..XXXXXXX | ||
195 | --- /dev/null | ||
196 | +++ b/tests/iothread.c | ||
197 | @@ -XXX,XX +XXX,XX @@ | ||
198 | +/* | ||
199 | + * Event loop thread implementation for unit tests | ||
200 | + * | ||
201 | + * Copyright Red Hat Inc., 2013, 2016 | ||
202 | + * | ||
203 | + * Authors: | ||
204 | + * Stefan Hajnoczi <stefanha@redhat.com> | ||
205 | + * Paolo Bonzini <pbonzini@redhat.com> | ||
206 | + * | ||
207 | + * This work is licensed under the terms of the GNU GPL, version 2 or later. | ||
208 | + * See the COPYING file in the top-level directory. | ||
209 | + * | ||
210 | + */ | ||
211 | + | ||
212 | +#include "qemu/osdep.h" | ||
213 | +#include "qapi/error.h" | ||
214 | +#include "block/aio.h" | ||
215 | +#include "qemu/main-loop.h" | ||
216 | +#include "qemu/rcu.h" | ||
217 | +#include "iothread.h" | ||
218 | + | ||
219 | +struct IOThread { | ||
220 | + AioContext *ctx; | ||
221 | + | ||
222 | + QemuThread thread; | ||
223 | + QemuMutex init_done_lock; | ||
224 | + QemuCond init_done_cond; /* is thread initialization done? */ | ||
225 | + bool stopping; | ||
226 | +}; | ||
227 | + | ||
228 | +static __thread IOThread *my_iothread; | ||
229 | + | ||
230 | +AioContext *qemu_get_current_aio_context(void) | ||
231 | +{ | ||
232 | + return my_iothread ? my_iothread->ctx : qemu_get_aio_context(); | ||
233 | +} | ||
234 | + | ||
235 | +static void *iothread_run(void *opaque) | ||
236 | +{ | ||
237 | + IOThread *iothread = opaque; | ||
238 | + | ||
239 | + rcu_register_thread(); | ||
240 | + | ||
241 | + my_iothread = iothread; | ||
242 | + qemu_mutex_lock(&iothread->init_done_lock); | ||
243 | + iothread->ctx = aio_context_new(&error_abort); | ||
244 | + qemu_cond_signal(&iothread->init_done_cond); | ||
245 | + qemu_mutex_unlock(&iothread->init_done_lock); | ||
246 | + | ||
247 | + while (!atomic_read(&iothread->stopping)) { | ||
248 | + aio_poll(iothread->ctx, true); | ||
249 | + } | ||
250 | + | ||
251 | + rcu_unregister_thread(); | ||
252 | + return NULL; | ||
253 | +} | ||
254 | + | ||
255 | +void iothread_join(IOThread *iothread) | ||
256 | +{ | ||
257 | + iothread->stopping = true; | ||
258 | + aio_notify(iothread->ctx); | ||
259 | + qemu_thread_join(&iothread->thread); | ||
260 | + qemu_cond_destroy(&iothread->init_done_cond); | ||
261 | + qemu_mutex_destroy(&iothread->init_done_lock); | ||
262 | + aio_context_unref(iothread->ctx); | ||
263 | + g_free(iothread); | ||
264 | +} | ||
265 | + | ||
266 | +IOThread *iothread_new(void) | ||
267 | +{ | ||
268 | + IOThread *iothread = g_new0(IOThread, 1); | ||
269 | + | ||
270 | + qemu_mutex_init(&iothread->init_done_lock); | ||
271 | + qemu_cond_init(&iothread->init_done_cond); | ||
272 | + qemu_thread_create(&iothread->thread, NULL, iothread_run, | ||
273 | + iothread, QEMU_THREAD_JOINABLE); | ||
274 | + | ||
275 | + /* Wait for initialization to complete */ | ||
276 | + qemu_mutex_lock(&iothread->init_done_lock); | ||
277 | + while (iothread->ctx == NULL) { | ||
278 | + qemu_cond_wait(&iothread->init_done_cond, | ||
279 | + &iothread->init_done_lock); | ||
280 | + } | ||
281 | + qemu_mutex_unlock(&iothread->init_done_lock); | ||
282 | + return iothread; | ||
283 | +} | ||
284 | + | ||
285 | +AioContext *iothread_get_aio_context(IOThread *iothread) | ||
286 | +{ | ||
287 | + return iothread->ctx; | ||
288 | +} | ||
289 | diff --git a/tests/test-aio-multithread.c b/tests/test-aio-multithread.c | ||
290 | new file mode 100644 | ||
291 | index XXXXXXX..XXXXXXX | ||
292 | --- /dev/null | ||
293 | +++ b/tests/test-aio-multithread.c | ||
294 | @@ -XXX,XX +XXX,XX @@ | ||
295 | +/* | ||
296 | + * AioContext multithreading tests | ||
297 | + * | ||
298 | + * Copyright Red Hat, Inc. 2016 | ||
299 | + * | ||
300 | + * Authors: | ||
301 | + * Paolo Bonzini <pbonzini@redhat.com> | ||
302 | + * | ||
303 | + * This work is licensed under the terms of the GNU LGPL, version 2 or later. | ||
304 | + * See the COPYING.LIB file in the top-level directory. | ||
305 | + */ | ||
306 | + | ||
307 | +#include "qemu/osdep.h" | ||
308 | +#include <glib.h> | ||
309 | +#include "block/aio.h" | ||
310 | +#include "qapi/error.h" | ||
311 | +#include "qemu/coroutine.h" | ||
312 | +#include "qemu/thread.h" | ||
313 | +#include "qemu/error-report.h" | ||
314 | +#include "iothread.h" | ||
315 | + | ||
316 | +/* AioContext management */ | ||
317 | + | ||
318 | +#define NUM_CONTEXTS 5 | ||
319 | + | ||
320 | +static IOThread *threads[NUM_CONTEXTS]; | ||
321 | +static AioContext *ctx[NUM_CONTEXTS]; | ||
322 | +static __thread int id = -1; | ||
323 | + | ||
324 | +static QemuEvent done_event; | ||
325 | + | ||
326 | +/* Run a function synchronously on a remote iothread. */ | ||
327 | + | ||
328 | +typedef struct CtxRunData { | ||
329 | + QEMUBHFunc *cb; | ||
330 | + void *arg; | ||
331 | +} CtxRunData; | ||
332 | + | ||
333 | +static void ctx_run_bh_cb(void *opaque) | ||
334 | +{ | ||
335 | + CtxRunData *data = opaque; | ||
336 | + | ||
337 | + data->cb(data->arg); | ||
338 | + qemu_event_set(&done_event); | ||
339 | +} | ||
340 | + | ||
341 | +static void ctx_run(int i, QEMUBHFunc *cb, void *opaque) | ||
342 | +{ | ||
343 | + CtxRunData data = { | ||
344 | + .cb = cb, | ||
345 | + .arg = opaque | ||
346 | + }; | ||
347 | + | ||
348 | + qemu_event_reset(&done_event); | ||
349 | + aio_bh_schedule_oneshot(ctx[i], ctx_run_bh_cb, &data); | ||
350 | + qemu_event_wait(&done_event); | ||
351 | +} | ||
352 | + | ||
353 | +/* Starting the iothreads. */ | ||
354 | + | ||
355 | +static void set_id_cb(void *opaque) | ||
356 | +{ | ||
357 | + int *i = opaque; | ||
358 | + | ||
359 | + id = *i; | ||
360 | +} | ||
361 | + | ||
362 | +static void create_aio_contexts(void) | ||
363 | +{ | ||
364 | + int i; | ||
365 | + | ||
366 | + for (i = 0; i < NUM_CONTEXTS; i++) { | ||
367 | + threads[i] = iothread_new(); | ||
368 | + ctx[i] = iothread_get_aio_context(threads[i]); | ||
369 | + } | ||
370 | + | ||
371 | + qemu_event_init(&done_event, false); | ||
372 | + for (i = 0; i < NUM_CONTEXTS; i++) { | ||
373 | + ctx_run(i, set_id_cb, &i); | ||
374 | + } | ||
375 | +} | ||
376 | + | ||
377 | +/* Stopping the iothreads. */ | ||
378 | + | ||
379 | +static void join_aio_contexts(void) | ||
380 | +{ | ||
381 | + int i; | ||
382 | + | ||
383 | + for (i = 0; i < NUM_CONTEXTS; i++) { | ||
384 | + aio_context_ref(ctx[i]); | ||
385 | + } | ||
386 | + for (i = 0; i < NUM_CONTEXTS; i++) { | ||
387 | + iothread_join(threads[i]); | ||
388 | + } | ||
389 | + for (i = 0; i < NUM_CONTEXTS; i++) { | ||
390 | + aio_context_unref(ctx[i]); | ||
391 | + } | ||
392 | + qemu_event_destroy(&done_event); | ||
393 | +} | ||
394 | + | ||
395 | +/* Basic test for the stuff above. */ | ||
396 | + | ||
397 | +static void test_lifecycle(void) | ||
398 | +{ | ||
399 | + create_aio_contexts(); | ||
400 | + join_aio_contexts(); | ||
401 | +} | ||
402 | + | ||
403 | +/* aio_co_schedule test. */ | ||
404 | + | ||
405 | +static Coroutine *to_schedule[NUM_CONTEXTS]; | ||
406 | + | ||
407 | +static bool now_stopping; | ||
408 | + | ||
409 | +static int count_retry; | ||
410 | +static int count_here; | ||
411 | +static int count_other; | ||
412 | + | ||
413 | +static bool schedule_next(int n) | ||
414 | +{ | ||
415 | + Coroutine *co; | ||
416 | + | ||
417 | + co = atomic_xchg(&to_schedule[n], NULL); | ||
418 | + if (!co) { | ||
419 | + atomic_inc(&count_retry); | ||
420 | + return false; | ||
421 | + } | ||
422 | + | ||
423 | + if (n == id) { | ||
424 | + atomic_inc(&count_here); | ||
425 | + } else { | ||
426 | + atomic_inc(&count_other); | ||
427 | + } | ||
428 | + | ||
429 | + aio_co_schedule(ctx[n], co); | ||
430 | + return true; | ||
431 | +} | ||
432 | + | ||
433 | +static void finish_cb(void *opaque) | ||
434 | +{ | ||
435 | + schedule_next(id); | ||
436 | +} | ||
437 | + | ||
438 | +static coroutine_fn void test_multi_co_schedule_entry(void *opaque) | ||
439 | +{ | ||
440 | + g_assert(to_schedule[id] == NULL); | ||
441 | + atomic_mb_set(&to_schedule[id], qemu_coroutine_self()); | ||
442 | + | ||
443 | + while (!atomic_mb_read(&now_stopping)) { | ||
444 | + int n; | ||
445 | + | ||
446 | + n = g_test_rand_int_range(0, NUM_CONTEXTS); | ||
447 | + schedule_next(n); | ||
448 | + qemu_coroutine_yield(); | ||
449 | + | ||
450 | + g_assert(to_schedule[id] == NULL); | ||
451 | + atomic_mb_set(&to_schedule[id], qemu_coroutine_self()); | ||
452 | + } | ||
453 | +} | ||
454 | + | ||
455 | + | ||
456 | +static void test_multi_co_schedule(int seconds) | ||
457 | +{ | ||
458 | + int i; | ||
459 | + | ||
460 | + count_here = count_other = count_retry = 0; | ||
461 | + now_stopping = false; | ||
462 | + | ||
463 | + create_aio_contexts(); | ||
464 | + for (i = 0; i < NUM_CONTEXTS; i++) { | ||
465 | + Coroutine *co1 = qemu_coroutine_create(test_multi_co_schedule_entry, NULL); | ||
466 | + aio_co_schedule(ctx[i], co1); | ||
467 | + } | ||
468 | + | ||
469 | + g_usleep(seconds * 1000000); | ||
470 | + | ||
471 | + atomic_mb_set(&now_stopping, true); | ||
472 | + for (i = 0; i < NUM_CONTEXTS; i++) { | ||
473 | + ctx_run(i, finish_cb, NULL); | ||
474 | + to_schedule[i] = NULL; | ||
475 | + } | ||
476 | + | ||
477 | + join_aio_contexts(); | ||
478 | + g_test_message("scheduled %d, queued %d, retry %d, total %d\n", | ||
479 | + count_other, count_here, count_retry, | ||
480 | + count_here + count_other + count_retry); | ||
481 | +} | ||
482 | + | ||
483 | +static void test_multi_co_schedule_1(void) | ||
484 | +{ | ||
485 | + test_multi_co_schedule(1); | ||
486 | +} | ||
487 | + | ||
488 | +static void test_multi_co_schedule_10(void) | ||
489 | +{ | ||
490 | + test_multi_co_schedule(10); | ||
491 | +} | ||
492 | + | ||
493 | +/* End of tests. */ | ||
494 | + | ||
495 | +int main(int argc, char **argv) | ||
496 | +{ | ||
497 | + init_clocks(); | ||
498 | + | ||
499 | + g_test_init(&argc, &argv, NULL); | ||
500 | + g_test_add_func("/aio/multi/lifecycle", test_lifecycle); | ||
501 | + if (g_test_quick()) { | ||
502 | + g_test_add_func("/aio/multi/schedule", test_multi_co_schedule_1); | ||
503 | + } else { | ||
504 | + g_test_add_func("/aio/multi/schedule", test_multi_co_schedule_10); | ||
505 | + } | ||
506 | + return g_test_run(); | ||
507 | +} | ||
508 | diff --git a/util/async.c b/util/async.c | ||
509 | index XXXXXXX..XXXXXXX 100644 | ||
510 | --- a/util/async.c | ||
511 | +++ b/util/async.c | ||
512 | @@ -XXX,XX +XXX,XX @@ | ||
513 | #include "qemu/main-loop.h" | ||
514 | #include "qemu/atomic.h" | ||
515 | #include "block/raw-aio.h" | ||
516 | +#include "qemu/coroutine_int.h" | ||
517 | +#include "trace.h" | ||
518 | |||
519 | /***********************************************************/ | ||
520 | /* bottom halves (can be seen as timers which expire ASAP) */ | ||
521 | @@ -XXX,XX +XXX,XX @@ aio_ctx_finalize(GSource *source) | ||
522 | } | ||
523 | #endif | ||
524 | |||
525 | + assert(QSLIST_EMPTY(&ctx->scheduled_coroutines)); | ||
526 | + qemu_bh_delete(ctx->co_schedule_bh); | ||
527 | + | ||
528 | qemu_lockcnt_lock(&ctx->list_lock); | ||
529 | assert(!qemu_lockcnt_count(&ctx->list_lock)); | ||
530 | while (ctx->first_bh) { | ||
531 | @@ -XXX,XX +XXX,XX @@ static bool event_notifier_poll(void *opaque) | ||
532 | return atomic_read(&ctx->notified); | ||
533 | } | ||
534 | |||
535 | +static void co_schedule_bh_cb(void *opaque) | ||
536 | +{ | ||
537 | + AioContext *ctx = opaque; | ||
538 | + QSLIST_HEAD(, Coroutine) straight, reversed; | ||
539 | + | ||
540 | + QSLIST_MOVE_ATOMIC(&reversed, &ctx->scheduled_coroutines); | ||
541 | + QSLIST_INIT(&straight); | ||
542 | + | ||
543 | + while (!QSLIST_EMPTY(&reversed)) { | ||
544 | + Coroutine *co = QSLIST_FIRST(&reversed); | ||
545 | + QSLIST_REMOVE_HEAD(&reversed, co_scheduled_next); | ||
546 | + QSLIST_INSERT_HEAD(&straight, co, co_scheduled_next); | ||
547 | + } | ||
548 | + | ||
549 | + while (!QSLIST_EMPTY(&straight)) { | ||
550 | + Coroutine *co = QSLIST_FIRST(&straight); | ||
551 | + QSLIST_REMOVE_HEAD(&straight, co_scheduled_next); | ||
552 | + trace_aio_co_schedule_bh_cb(ctx, co); | ||
553 | + qemu_coroutine_enter(co); | ||
554 | + } | ||
555 | +} | ||
556 | + | ||
557 | AioContext *aio_context_new(Error **errp) | ||
558 | { | ||
559 | int ret; | ||
560 | @@ -XXX,XX +XXX,XX @@ AioContext *aio_context_new(Error **errp) | ||
561 | } | ||
562 | g_source_set_can_recurse(&ctx->source, true); | ||
563 | qemu_lockcnt_init(&ctx->list_lock); | ||
564 | + | ||
565 | + ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx); | ||
566 | + QSLIST_INIT(&ctx->scheduled_coroutines); | ||
567 | + | ||
568 | aio_set_event_notifier(ctx, &ctx->notifier, | ||
569 | false, | ||
570 | (EventNotifierHandler *) | ||
571 | @@ -XXX,XX +XXX,XX @@ fail: | ||
572 | return NULL; | ||
573 | } | ||
574 | |||
575 | +void aio_co_schedule(AioContext *ctx, Coroutine *co) | ||
576 | +{ | ||
577 | + trace_aio_co_schedule(ctx, co); | ||
578 | + QSLIST_INSERT_HEAD_ATOMIC(&ctx->scheduled_coroutines, | ||
579 | + co, co_scheduled_next); | ||
580 | + qemu_bh_schedule(ctx->co_schedule_bh); | ||
581 | +} | ||
582 | + | ||
583 | +void aio_co_wake(struct Coroutine *co) | ||
584 | +{ | ||
585 | + AioContext *ctx; | ||
586 | + | ||
587 | + /* Read coroutine before co->ctx. Matches smp_wmb in | ||
588 | + * qemu_coroutine_enter. | ||
589 | + */ | ||
590 | + smp_read_barrier_depends(); | ||
591 | + ctx = atomic_read(&co->ctx); | ||
592 | + | ||
593 | + if (ctx != qemu_get_current_aio_context()) { | ||
594 | + aio_co_schedule(ctx, co); | ||
595 | + return; | ||
596 | + } | ||
597 | + | ||
598 | + if (qemu_in_coroutine()) { | ||
599 | + Coroutine *self = qemu_coroutine_self(); | ||
600 | + assert(self != co); | ||
601 | + QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, co, co_queue_next); | ||
602 | + } else { | ||
603 | + aio_context_acquire(ctx); | ||
604 | + qemu_coroutine_enter(co); | ||
605 | + aio_context_release(ctx); | ||
606 | + } | ||
607 | +} | ||
608 | + | ||
609 | void aio_context_ref(AioContext *ctx) | ||
610 | { | ||
611 | g_source_ref(&ctx->source); | ||
612 | diff --git a/util/qemu-coroutine.c b/util/qemu-coroutine.c | ||
613 | index XXXXXXX..XXXXXXX 100644 | ||
614 | --- a/util/qemu-coroutine.c | ||
615 | +++ b/util/qemu-coroutine.c | ||
616 | @@ -XXX,XX +XXX,XX @@ | ||
617 | #include "qemu/atomic.h" | ||
618 | #include "qemu/coroutine.h" | ||
619 | #include "qemu/coroutine_int.h" | ||
620 | +#include "block/aio.h" | ||
621 | |||
622 | enum { | ||
623 | POOL_BATCH_SIZE = 64, | ||
624 | @@ -XXX,XX +XXX,XX @@ void qemu_coroutine_enter(Coroutine *co) | ||
625 | } | ||
626 | |||
627 | co->caller = self; | ||
628 | + co->ctx = qemu_get_current_aio_context(); | ||
629 | + | ||
630 | + /* Store co->ctx before anything that stores co. Matches | ||
631 | + * barrier in aio_co_wake. | ||
632 | + */ | ||
633 | + smp_wmb(); | ||
634 | + | ||
635 | ret = qemu_coroutine_switch(self, co, COROUTINE_ENTER); | ||
636 | |||
637 | qemu_co_queue_run_restart(co); | ||
638 | diff --git a/util/trace-events b/util/trace-events | ||
639 | index XXXXXXX..XXXXXXX 100644 | ||
640 | --- a/util/trace-events | ||
641 | +++ b/util/trace-events | ||
642 | @@ -XXX,XX +XXX,XX @@ run_poll_handlers_end(void *ctx, bool progress) "ctx %p progress %d" | ||
643 | poll_shrink(void *ctx, int64_t old, int64_t new) "ctx %p old %"PRId64" new %"PRId64 | ||
644 | poll_grow(void *ctx, int64_t old, int64_t new) "ctx %p old %"PRId64" new %"PRId64 | ||
645 | |||
646 | +# util/async.c | ||
647 | +aio_co_schedule(void *ctx, void *co) "ctx %p co %p" | ||
648 | +aio_co_schedule_bh_cb(void *ctx, void *co) "ctx %p co %p" | ||
649 | + | ||
650 | # util/thread-pool.c | ||
651 | thread_pool_submit(void *pool, void *req, void *opaque) "pool %p req %p opaque %p" | ||
652 | thread_pool_complete(void *pool, void *req, void *opaque, int ret) "pool %p req %p opaque %p ret %d" | ||
653 | -- | ||
654 | 2.9.3 | ||
655 | |||
656 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
1 | 2 | ||
3 | qcow2_create2 calls this. Do not run a nested event loop, as that | ||
4 | breaks when aio_co_wake tries to queue the coroutine on the co_queue_wakeup | ||
5 | list of the currently running one. | ||
6 | |||
7 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
8 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
9 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
10 | Message-id: 20170213135235.12274-4-pbonzini@redhat.com | ||
11 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
12 | --- | ||
13 | block/block-backend.c | 12 ++++++++---- | ||
14 | 1 file changed, 8 insertions(+), 4 deletions(-) | ||
15 | |||
16 | diff --git a/block/block-backend.c b/block/block-backend.c | ||
17 | index XXXXXXX..XXXXXXX 100644 | ||
18 | --- a/block/block-backend.c | ||
19 | +++ b/block/block-backend.c | ||
20 | @@ -XXX,XX +XXX,XX @@ static int blk_prw(BlockBackend *blk, int64_t offset, uint8_t *buf, | ||
21 | { | ||
22 | QEMUIOVector qiov; | ||
23 | struct iovec iov; | ||
24 | - Coroutine *co; | ||
25 | BlkRwCo rwco; | ||
26 | |||
27 | iov = (struct iovec) { | ||
28 | @@ -XXX,XX +XXX,XX @@ static int blk_prw(BlockBackend *blk, int64_t offset, uint8_t *buf, | ||
29 | .ret = NOT_DONE, | ||
30 | }; | ||
31 | |||
32 | - co = qemu_coroutine_create(co_entry, &rwco); | ||
33 | - qemu_coroutine_enter(co); | ||
34 | - BDRV_POLL_WHILE(blk_bs(blk), rwco.ret == NOT_DONE); | ||
35 | + if (qemu_in_coroutine()) { | ||
36 | + /* Fast-path if already in coroutine context */ | ||
37 | + co_entry(&rwco); | ||
38 | + } else { | ||
39 | + Coroutine *co = qemu_coroutine_create(co_entry, &rwco); | ||
40 | + qemu_coroutine_enter(co); | ||
41 | + BDRV_POLL_WHILE(blk_bs(blk), rwco.ret == NOT_DONE); | ||
42 | + } | ||
43 | |||
44 | return rwco.ret; | ||
45 | } | ||
46 | -- | ||
47 | 2.9.3 | ||
48 | |||
49 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
1 | 2 | ||
3 | Once the thread pool starts using aio_co_wake, it will also need | ||
4 | qemu_get_current_aio_context(). Make test-thread-pool create | ||
5 | an AioContext with qemu_init_main_loop, so that stubs/iothread.c | ||
6 | and tests/iothread.c can provide the rest. | ||
7 | |||
8 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
9 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
10 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
11 | Message-id: 20170213135235.12274-5-pbonzini@redhat.com | ||
12 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
13 | --- | ||
14 | tests/test-thread-pool.c | 12 +++--------- | ||
15 | 1 file changed, 3 insertions(+), 9 deletions(-) | ||
16 | |||
17 | diff --git a/tests/test-thread-pool.c b/tests/test-thread-pool.c | ||
18 | index XXXXXXX..XXXXXXX 100644 | ||
19 | --- a/tests/test-thread-pool.c | ||
20 | +++ b/tests/test-thread-pool.c | ||
21 | @@ -XXX,XX +XXX,XX @@ | ||
22 | #include "qapi/error.h" | ||
23 | #include "qemu/timer.h" | ||
24 | #include "qemu/error-report.h" | ||
25 | +#include "qemu/main-loop.h" | ||
26 | |||
27 | static AioContext *ctx; | ||
28 | static ThreadPool *pool; | ||
29 | @@ -XXX,XX +XXX,XX @@ static void test_cancel_async(void) | ||
30 | int main(int argc, char **argv) | ||
31 | { | ||
32 | int ret; | ||
33 | - Error *local_error = NULL; | ||
34 | |||
35 | - init_clocks(); | ||
36 | - | ||
37 | - ctx = aio_context_new(&local_error); | ||
38 | - if (!ctx) { | ||
39 | - error_reportf_err(local_error, "Failed to create AIO Context: "); | ||
40 | - exit(1); | ||
41 | - } | ||
42 | + qemu_init_main_loop(&error_abort); | ||
43 | + ctx = qemu_get_current_aio_context(); | ||
44 | pool = aio_get_thread_pool(ctx); | ||
45 | |||
46 | g_test_init(&argc, &argv, NULL); | ||
47 | @@ -XXX,XX +XXX,XX @@ int main(int argc, char **argv) | ||
48 | |||
49 | ret = g_test_run(); | ||
50 | |||
51 | - aio_context_unref(ctx); | ||
52 | return ret; | ||
53 | } | ||
54 | -- | ||
55 | 2.9.3 | ||
56 | |||
57 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | 1 | From: Paolo Bonzini <pbonzini@redhat.com> | |
2 | |||
3 | This is in preparation for making qio_channel_yield work on | ||
4 | AioContexts other than the main one. | ||
5 | |||
6 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> | ||
7 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
8 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
9 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
10 | Message-id: 20170213135235.12274-6-pbonzini@redhat.com | ||
11 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
12 | --- | ||
13 | include/io/channel.h | 25 +++++++++++++++++++++++++ | ||
14 | io/channel-command.c | 13 +++++++++++++ | ||
15 | io/channel-file.c | 11 +++++++++++ | ||
16 | io/channel-socket.c | 16 +++++++++++----- | ||
17 | io/channel-tls.c | 12 ++++++++++++ | ||
18 | io/channel-watch.c | 6 ++++++ | ||
19 | io/channel.c | 11 +++++++++++ | ||
20 | 7 files changed, 89 insertions(+), 5 deletions(-) | ||
21 | |||
22 | diff --git a/include/io/channel.h b/include/io/channel.h | ||
23 | index XXXXXXX..XXXXXXX 100644 | ||
24 | --- a/include/io/channel.h | ||
25 | +++ b/include/io/channel.h | ||
26 | @@ -XXX,XX +XXX,XX @@ | ||
27 | |||
28 | #include "qemu-common.h" | ||
29 | #include "qom/object.h" | ||
30 | +#include "block/aio.h" | ||
31 | |||
32 | #define TYPE_QIO_CHANNEL "qio-channel" | ||
33 | #define QIO_CHANNEL(obj) \ | ||
34 | @@ -XXX,XX +XXX,XX @@ struct QIOChannelClass { | ||
35 | off_t offset, | ||
36 | int whence, | ||
37 | Error **errp); | ||
38 | + void (*io_set_aio_fd_handler)(QIOChannel *ioc, | ||
39 | + AioContext *ctx, | ||
40 | + IOHandler *io_read, | ||
41 | + IOHandler *io_write, | ||
42 | + void *opaque); | ||
43 | }; | ||
44 | |||
45 | /* General I/O handling functions */ | ||
46 | @@ -XXX,XX +XXX,XX @@ void qio_channel_yield(QIOChannel *ioc, | ||
47 | void qio_channel_wait(QIOChannel *ioc, | ||
48 | GIOCondition condition); | ||
49 | |||
50 | +/** | ||
51 | + * qio_channel_set_aio_fd_handler: | ||
52 | + * @ioc: the channel object | ||
53 | + * @ctx: the AioContext to set the handlers on | ||
54 | + * @io_read: the read handler | ||
55 | + * @io_write: the write handler | ||
56 | + * @opaque: the opaque value passed to the handler | ||
57 | + * | ||
58 | + * This is used internally by qio_channel_yield(). It can | ||
59 | + * be used by channel implementations to forward the handlers | ||
60 | + * to another channel (e.g. from #QIOChannelTLS to the | ||
61 | + * underlying socket). | ||
62 | + */ | ||
63 | +void qio_channel_set_aio_fd_handler(QIOChannel *ioc, | ||
64 | + AioContext *ctx, | ||
65 | + IOHandler *io_read, | ||
66 | + IOHandler *io_write, | ||
67 | + void *opaque); | ||
68 | + | ||
69 | #endif /* QIO_CHANNEL_H */ | ||
70 | diff --git a/io/channel-command.c b/io/channel-command.c | ||
71 | index XXXXXXX..XXXXXXX 100644 | ||
72 | --- a/io/channel-command.c | ||
73 | +++ b/io/channel-command.c | ||
74 | @@ -XXX,XX +XXX,XX @@ static int qio_channel_command_close(QIOChannel *ioc, | ||
75 | } | ||
76 | |||
77 | |||
78 | +static void qio_channel_command_set_aio_fd_handler(QIOChannel *ioc, | ||
79 | + AioContext *ctx, | ||
80 | + IOHandler *io_read, | ||
81 | + IOHandler *io_write, | ||
82 | + void *opaque) | ||
83 | +{ | ||
84 | + QIOChannelCommand *cioc = QIO_CHANNEL_COMMAND(ioc); | ||
85 | + aio_set_fd_handler(ctx, cioc->readfd, false, io_read, NULL, NULL, opaque); | ||
86 | + aio_set_fd_handler(ctx, cioc->writefd, false, NULL, io_write, NULL, opaque); | ||
87 | +} | ||
88 | + | ||
89 | + | ||
90 | static GSource *qio_channel_command_create_watch(QIOChannel *ioc, | ||
91 | GIOCondition condition) | ||
92 | { | ||
93 | @@ -XXX,XX +XXX,XX @@ static void qio_channel_command_class_init(ObjectClass *klass, | ||
94 | ioc_klass->io_set_blocking = qio_channel_command_set_blocking; | ||
95 | ioc_klass->io_close = qio_channel_command_close; | ||
96 | ioc_klass->io_create_watch = qio_channel_command_create_watch; | ||
97 | + ioc_klass->io_set_aio_fd_handler = qio_channel_command_set_aio_fd_handler; | ||
98 | } | ||
99 | |||
100 | static const TypeInfo qio_channel_command_info = { | ||
101 | diff --git a/io/channel-file.c b/io/channel-file.c | ||
102 | index XXXXXXX..XXXXXXX 100644 | ||
103 | --- a/io/channel-file.c | ||
104 | +++ b/io/channel-file.c | ||
105 | @@ -XXX,XX +XXX,XX @@ static int qio_channel_file_close(QIOChannel *ioc, | ||
106 | } | ||
107 | |||
108 | |||
109 | +static void qio_channel_file_set_aio_fd_handler(QIOChannel *ioc, | ||
110 | + AioContext *ctx, | ||
111 | + IOHandler *io_read, | ||
112 | + IOHandler *io_write, | ||
113 | + void *opaque) | ||
114 | +{ | ||
115 | + QIOChannelFile *fioc = QIO_CHANNEL_FILE(ioc); | ||
116 | + aio_set_fd_handler(ctx, fioc->fd, false, io_read, io_write, NULL, opaque); | ||
117 | +} | ||
118 | + | ||
119 | static GSource *qio_channel_file_create_watch(QIOChannel *ioc, | ||
120 | GIOCondition condition) | ||
121 | { | ||
122 | @@ -XXX,XX +XXX,XX @@ static void qio_channel_file_class_init(ObjectClass *klass, | ||
123 | ioc_klass->io_seek = qio_channel_file_seek; | ||
124 | ioc_klass->io_close = qio_channel_file_close; | ||
125 | ioc_klass->io_create_watch = qio_channel_file_create_watch; | ||
126 | + ioc_klass->io_set_aio_fd_handler = qio_channel_file_set_aio_fd_handler; | ||
127 | } | ||
128 | |||
129 | static const TypeInfo qio_channel_file_info = { | ||
130 | diff --git a/io/channel-socket.c b/io/channel-socket.c | ||
131 | index XXXXXXX..XXXXXXX 100644 | ||
132 | --- a/io/channel-socket.c | ||
133 | +++ b/io/channel-socket.c | ||
134 | @@ -XXX,XX +XXX,XX @@ qio_channel_socket_set_blocking(QIOChannel *ioc, | ||
135 | qemu_set_block(sioc->fd); | ||
136 | } else { | ||
137 | qemu_set_nonblock(sioc->fd); | ||
138 | -#ifdef WIN32 | ||
139 | - WSAEventSelect(sioc->fd, ioc->event, | ||
140 | - FD_READ | FD_ACCEPT | FD_CLOSE | | ||
141 | - FD_CONNECT | FD_WRITE | FD_OOB); | ||
142 | -#endif | ||
143 | } | ||
144 | return 0; | ||
145 | } | ||
146 | @@ -XXX,XX +XXX,XX @@ qio_channel_socket_shutdown(QIOChannel *ioc, | ||
147 | return 0; | ||
148 | } | ||
149 | |||
150 | +static void qio_channel_socket_set_aio_fd_handler(QIOChannel *ioc, | ||
151 | + AioContext *ctx, | ||
152 | + IOHandler *io_read, | ||
153 | + IOHandler *io_write, | ||
154 | + void *opaque) | ||
155 | +{ | ||
156 | + QIOChannelSocket *sioc = QIO_CHANNEL_SOCKET(ioc); | ||
157 | + aio_set_fd_handler(ctx, sioc->fd, false, io_read, io_write, NULL, opaque); | ||
158 | +} | ||
159 | + | ||
160 | static GSource *qio_channel_socket_create_watch(QIOChannel *ioc, | ||
161 | GIOCondition condition) | ||
162 | { | ||
163 | @@ -XXX,XX +XXX,XX @@ static void qio_channel_socket_class_init(ObjectClass *klass, | ||
164 | ioc_klass->io_set_cork = qio_channel_socket_set_cork; | ||
165 | ioc_klass->io_set_delay = qio_channel_socket_set_delay; | ||
166 | ioc_klass->io_create_watch = qio_channel_socket_create_watch; | ||
167 | + ioc_klass->io_set_aio_fd_handler = qio_channel_socket_set_aio_fd_handler; | ||
168 | } | ||
169 | |||
170 | static const TypeInfo qio_channel_socket_info = { | ||
171 | diff --git a/io/channel-tls.c b/io/channel-tls.c | ||
172 | index XXXXXXX..XXXXXXX 100644 | ||
173 | --- a/io/channel-tls.c | ||
174 | +++ b/io/channel-tls.c | ||
175 | @@ -XXX,XX +XXX,XX @@ static int qio_channel_tls_close(QIOChannel *ioc, | ||
176 | return qio_channel_close(tioc->master, errp); | ||
177 | } | ||
178 | |||
179 | +static void qio_channel_tls_set_aio_fd_handler(QIOChannel *ioc, | ||
180 | + AioContext *ctx, | ||
181 | + IOHandler *io_read, | ||
182 | + IOHandler *io_write, | ||
183 | + void *opaque) | ||
184 | +{ | ||
185 | + QIOChannelTLS *tioc = QIO_CHANNEL_TLS(ioc); | ||
186 | + | ||
187 | + qio_channel_set_aio_fd_handler(tioc->master, ctx, io_read, io_write, opaque); | ||
188 | +} | ||
189 | + | ||
190 | static GSource *qio_channel_tls_create_watch(QIOChannel *ioc, | ||
191 | GIOCondition condition) | ||
192 | { | ||
193 | @@ -XXX,XX +XXX,XX @@ static void qio_channel_tls_class_init(ObjectClass *klass, | ||
194 | ioc_klass->io_close = qio_channel_tls_close; | ||
195 | ioc_klass->io_shutdown = qio_channel_tls_shutdown; | ||
196 | ioc_klass->io_create_watch = qio_channel_tls_create_watch; | ||
197 | + ioc_klass->io_set_aio_fd_handler = qio_channel_tls_set_aio_fd_handler; | ||
198 | } | ||
199 | |||
200 | static const TypeInfo qio_channel_tls_info = { | ||
201 | diff --git a/io/channel-watch.c b/io/channel-watch.c | ||
202 | index XXXXXXX..XXXXXXX 100644 | ||
203 | --- a/io/channel-watch.c | ||
204 | +++ b/io/channel-watch.c | ||
205 | @@ -XXX,XX +XXX,XX @@ GSource *qio_channel_create_socket_watch(QIOChannel *ioc, | ||
206 | GSource *source; | ||
207 | QIOChannelSocketSource *ssource; | ||
208 | |||
209 | +#ifdef WIN32 | ||
210 | + WSAEventSelect(socket, ioc->event, | ||
211 | + FD_READ | FD_ACCEPT | FD_CLOSE | | ||
212 | + FD_CONNECT | FD_WRITE | FD_OOB); | ||
213 | +#endif | ||
214 | + | ||
215 | source = g_source_new(&qio_channel_socket_source_funcs, | ||
216 | sizeof(QIOChannelSocketSource)); | ||
217 | ssource = (QIOChannelSocketSource *)source; | ||
218 | diff --git a/io/channel.c b/io/channel.c | ||
219 | index XXXXXXX..XXXXXXX 100644 | ||
220 | --- a/io/channel.c | ||
221 | +++ b/io/channel.c | ||
222 | @@ -XXX,XX +XXX,XX @@ GSource *qio_channel_create_watch(QIOChannel *ioc, | ||
223 | } | ||
224 | |||
225 | |||
226 | +void qio_channel_set_aio_fd_handler(QIOChannel *ioc, | ||
227 | + AioContext *ctx, | ||
228 | + IOHandler *io_read, | ||
229 | + IOHandler *io_write, | ||
230 | + void *opaque) | ||
231 | +{ | ||
232 | + QIOChannelClass *klass = QIO_CHANNEL_GET_CLASS(ioc); | ||
233 | + | ||
234 | + klass->io_set_aio_fd_handler(ioc, ctx, io_read, io_write, opaque); | ||
235 | +} | ||
236 | + | ||
237 | guint qio_channel_add_watch(QIOChannel *ioc, | ||
238 | GIOCondition condition, | ||
239 | QIOChannelFunc func, | ||
240 | -- | ||
241 | 2.9.3 | ||
242 | |||
243 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | 1 | From: Paolo Bonzini <pbonzini@redhat.com> | |
2 | |||
3 | Support separate coroutines for reading and writing, and place the | ||
4 | read/write handlers on the AioContext that the QIOChannel is registered | ||
5 | with. | ||
6 | |||
7 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> | ||
8 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
9 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
10 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
11 | Message-id: 20170213135235.12274-7-pbonzini@redhat.com | ||
12 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
13 | --- | ||
14 | include/io/channel.h | 47 ++++++++++++++++++++++++++-- | ||
15 | io/channel.c | 86 +++++++++++++++++++++++++++++++++++++++------------- | ||
16 | 2 files changed, 109 insertions(+), 24 deletions(-) | ||
17 | |||
18 | diff --git a/include/io/channel.h b/include/io/channel.h | ||
19 | index XXXXXXX..XXXXXXX 100644 | ||
20 | --- a/include/io/channel.h | ||
21 | +++ b/include/io/channel.h | ||
22 | @@ -XXX,XX +XXX,XX @@ | ||
23 | |||
24 | #include "qemu-common.h" | ||
25 | #include "qom/object.h" | ||
26 | +#include "qemu/coroutine.h" | ||
27 | #include "block/aio.h" | ||
28 | |||
29 | #define TYPE_QIO_CHANNEL "qio-channel" | ||
30 | @@ -XXX,XX +XXX,XX @@ struct QIOChannel { | ||
31 | Object parent; | ||
32 | unsigned int features; /* bitmask of QIOChannelFeatures */ | ||
33 | char *name; | ||
34 | + AioContext *ctx; | ||
35 | + Coroutine *read_coroutine; | ||
36 | + Coroutine *write_coroutine; | ||
37 | #ifdef _WIN32 | ||
38 | HANDLE event; /* For use with GSource on Win32 */ | ||
39 | #endif | ||
40 | @@ -XXX,XX +XXX,XX @@ guint qio_channel_add_watch(QIOChannel *ioc, | ||
41 | |||
42 | |||
43 | /** | ||
44 | + * qio_channel_attach_aio_context: | ||
45 | + * @ioc: the channel object | ||
46 | + * @ctx: the #AioContext to set the handlers on | ||
47 | + * | ||
48 | + * Request that qio_channel_yield() sets I/O handlers on | ||
49 | + * the given #AioContext. If @ctx is %NULL, qio_channel_yield() | ||
50 | + * uses QEMU's main thread event loop. | ||
51 | + * | ||
52 | + * You can move a #QIOChannel from one #AioContext to another even if | ||
53 | + * I/O handlers are set for a coroutine. However, #QIOChannel provides | ||
54 | + * no synchronization between the calls to qio_channel_yield() and | ||
55 | + * qio_channel_attach_aio_context(). | ||
56 | + * | ||
57 | + * Therefore you should first call qio_channel_detach_aio_context() | ||
58 | + * to ensure that the coroutine is not entered concurrently. Then, | ||
59 | + * while the coroutine has yielded, call qio_channel_attach_aio_context(), | ||
60 | + * and then aio_co_schedule() to place the coroutine on the new | ||
61 | + * #AioContext. The calls to qio_channel_detach_aio_context() | ||
62 | + * and qio_channel_attach_aio_context() should be protected with | ||
63 | + * aio_context_acquire() and aio_context_release(). | ||
64 | + */ | ||
65 | +void qio_channel_attach_aio_context(QIOChannel *ioc, | ||
66 | + AioContext *ctx); | ||
67 | + | ||
68 | +/** | ||
69 | + * qio_channel_detach_aio_context: | ||
70 | + * @ioc: the channel object | ||
71 | + * | ||
72 | + * Disable any I/O handlers set by qio_channel_yield(). With the | ||
73 | + * help of aio_co_schedule(), this allows moving a coroutine that was | ||
74 | + * paused by qio_channel_yield() to another context. | ||
75 | + */ | ||
76 | +void qio_channel_detach_aio_context(QIOChannel *ioc); | ||
77 | + | ||
78 | +/** | ||
79 | * qio_channel_yield: | ||
80 | * @ioc: the channel object | ||
81 | * @condition: the I/O condition to wait for | ||
82 | * | ||
83 | - * Yields execution from the current coroutine until | ||
84 | - * the condition indicated by @condition becomes | ||
85 | - * available. | ||
86 | + * Yields execution from the current coroutine until the condition | ||
87 | + * indicated by @condition becomes available. @condition must | ||
88 | + * be either %G_IO_IN or %G_IO_OUT; it cannot contain both. In | ||
89 | + * addition, no two coroutine can be waiting on the same condition | ||
90 | + * and channel at the same time. | ||
91 | * | ||
92 | * This must only be called from coroutine context | ||
93 | */ | ||
94 | diff --git a/io/channel.c b/io/channel.c | ||
95 | index XXXXXXX..XXXXXXX 100644 | ||
96 | --- a/io/channel.c | ||
97 | +++ b/io/channel.c | ||
98 | @@ -XXX,XX +XXX,XX @@ | ||
99 | #include "qemu/osdep.h" | ||
100 | #include "io/channel.h" | ||
101 | #include "qapi/error.h" | ||
102 | -#include "qemu/coroutine.h" | ||
103 | +#include "qemu/main-loop.h" | ||
104 | |||
105 | bool qio_channel_has_feature(QIOChannel *ioc, | ||
106 | QIOChannelFeature feature) | ||
107 | @@ -XXX,XX +XXX,XX @@ off_t qio_channel_io_seek(QIOChannel *ioc, | ||
108 | } | ||
109 | |||
110 | |||
111 | -typedef struct QIOChannelYieldData QIOChannelYieldData; | ||
112 | -struct QIOChannelYieldData { | ||
113 | - QIOChannel *ioc; | ||
114 | - Coroutine *co; | ||
115 | -}; | ||
116 | +static void qio_channel_set_aio_fd_handlers(QIOChannel *ioc); | ||
117 | |||
118 | +static void qio_channel_restart_read(void *opaque) | ||
119 | +{ | ||
120 | + QIOChannel *ioc = opaque; | ||
121 | + Coroutine *co = ioc->read_coroutine; | ||
122 | + | ||
123 | + ioc->read_coroutine = NULL; | ||
124 | + qio_channel_set_aio_fd_handlers(ioc); | ||
125 | + aio_co_wake(co); | ||
126 | +} | ||
127 | |||
128 | -static gboolean qio_channel_yield_enter(QIOChannel *ioc, | ||
129 | - GIOCondition condition, | ||
130 | - gpointer opaque) | ||
131 | +static void qio_channel_restart_write(void *opaque) | ||
132 | { | ||
133 | - QIOChannelYieldData *data = opaque; | ||
134 | - qemu_coroutine_enter(data->co); | ||
135 | - return FALSE; | ||
136 | + QIOChannel *ioc = opaque; | ||
137 | + Coroutine *co = ioc->write_coroutine; | ||
138 | + | ||
139 | + ioc->write_coroutine = NULL; | ||
140 | + qio_channel_set_aio_fd_handlers(ioc); | ||
141 | + aio_co_wake(co); | ||
142 | } | ||
143 | |||
144 | +static void qio_channel_set_aio_fd_handlers(QIOChannel *ioc) | ||
145 | +{ | ||
146 | + IOHandler *rd_handler = NULL, *wr_handler = NULL; | ||
147 | + AioContext *ctx; | ||
148 | + | ||
149 | + if (ioc->read_coroutine) { | ||
150 | + rd_handler = qio_channel_restart_read; | ||
151 | + } | ||
152 | + if (ioc->write_coroutine) { | ||
153 | + wr_handler = qio_channel_restart_write; | ||
154 | + } | ||
155 | + | ||
156 | + ctx = ioc->ctx ? ioc->ctx : iohandler_get_aio_context(); | ||
157 | + qio_channel_set_aio_fd_handler(ioc, ctx, rd_handler, wr_handler, ioc); | ||
158 | +} | ||
159 | + | ||
160 | +void qio_channel_attach_aio_context(QIOChannel *ioc, | ||
161 | + AioContext *ctx) | ||
162 | +{ | ||
163 | + AioContext *old_ctx; | ||
164 | + if (ioc->ctx == ctx) { | ||
165 | + return; | ||
166 | + } | ||
167 | + | ||
168 | + old_ctx = ioc->ctx ? ioc->ctx : iohandler_get_aio_context(); | ||
169 | + qio_channel_set_aio_fd_handler(ioc, old_ctx, NULL, NULL, NULL); | ||
170 | + ioc->ctx = ctx; | ||
171 | + qio_channel_set_aio_fd_handlers(ioc); | ||
172 | +} | ||
173 | + | ||
174 | +void qio_channel_detach_aio_context(QIOChannel *ioc) | ||
175 | +{ | ||
176 | + ioc->read_coroutine = NULL; | ||
177 | + ioc->write_coroutine = NULL; | ||
178 | + qio_channel_set_aio_fd_handlers(ioc); | ||
179 | + ioc->ctx = NULL; | ||
180 | +} | ||
181 | |||
182 | void coroutine_fn qio_channel_yield(QIOChannel *ioc, | ||
183 | GIOCondition condition) | ||
184 | { | ||
185 | - QIOChannelYieldData data; | ||
186 | - | ||
187 | assert(qemu_in_coroutine()); | ||
188 | - data.ioc = ioc; | ||
189 | - data.co = qemu_coroutine_self(); | ||
190 | - qio_channel_add_watch(ioc, | ||
191 | - condition, | ||
192 | - qio_channel_yield_enter, | ||
193 | - &data, | ||
194 | - NULL); | ||
195 | + if (condition == G_IO_IN) { | ||
196 | + assert(!ioc->read_coroutine); | ||
197 | + ioc->read_coroutine = qemu_coroutine_self(); | ||
198 | + } else if (condition == G_IO_OUT) { | ||
199 | + assert(!ioc->write_coroutine); | ||
200 | + ioc->write_coroutine = qemu_coroutine_self(); | ||
201 | + } else { | ||
202 | + abort(); | ||
203 | + } | ||
204 | + qio_channel_set_aio_fd_handlers(ioc); | ||
205 | qemu_coroutine_yield(); | ||
206 | } | ||
207 | |||
208 | -- | ||
209 | 2.9.3 | ||
210 | |||
211 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
1 | 2 | ||
3 | In the client, read the reply headers from a coroutine, switching the | ||
4 | read side between the "read header" coroutine and the I/O coroutine that | ||
5 | reads the body of the reply. | ||
6 | |||
7 | In the server, if the server can read more requests it will create a new | ||
8 | "read request" coroutine as soon as a request has been read. Otherwise, | ||
9 | the new coroutine is created in nbd_request_put. | ||
10 | |||
11 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
12 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
13 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
14 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> | ||
15 | Message-id: 20170213135235.12274-8-pbonzini@redhat.com | ||
16 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
17 | --- | ||
18 | block/nbd-client.h | 2 +- | ||
19 | block/nbd-client.c | 117 ++++++++++++++++++++++++----------------------------- | ||
20 | nbd/client.c | 2 +- | ||
21 | nbd/common.c | 9 +---- | ||
22 | nbd/server.c | 94 +++++++++++++----------------------------- | ||
23 | 5 files changed, 83 insertions(+), 141 deletions(-) | ||
24 | |||
25 | diff --git a/block/nbd-client.h b/block/nbd-client.h | ||
26 | index XXXXXXX..XXXXXXX 100644 | ||
27 | --- a/block/nbd-client.h | ||
28 | +++ b/block/nbd-client.h | ||
29 | @@ -XXX,XX +XXX,XX @@ typedef struct NBDClientSession { | ||
30 | |||
31 | CoMutex send_mutex; | ||
32 | CoQueue free_sema; | ||
33 | - Coroutine *send_coroutine; | ||
34 | + Coroutine *read_reply_co; | ||
35 | int in_flight; | ||
36 | |||
37 | Coroutine *recv_coroutine[MAX_NBD_REQUESTS]; | ||
38 | diff --git a/block/nbd-client.c b/block/nbd-client.c | ||
39 | index XXXXXXX..XXXXXXX 100644 | ||
40 | --- a/block/nbd-client.c | ||
41 | +++ b/block/nbd-client.c | ||
42 | @@ -XXX,XX +XXX,XX @@ | ||
43 | #define HANDLE_TO_INDEX(bs, handle) ((handle) ^ ((uint64_t)(intptr_t)bs)) | ||
44 | #define INDEX_TO_HANDLE(bs, index) ((index) ^ ((uint64_t)(intptr_t)bs)) | ||
45 | |||
46 | -static void nbd_recv_coroutines_enter_all(NBDClientSession *s) | ||
47 | +static void nbd_recv_coroutines_enter_all(BlockDriverState *bs) | ||
48 | { | ||
49 | + NBDClientSession *s = nbd_get_client_session(bs); | ||
50 | int i; | ||
51 | |||
52 | for (i = 0; i < MAX_NBD_REQUESTS; i++) { | ||
53 | @@ -XXX,XX +XXX,XX @@ static void nbd_recv_coroutines_enter_all(NBDClientSession *s) | ||
54 | qemu_coroutine_enter(s->recv_coroutine[i]); | ||
55 | } | ||
56 | } | ||
57 | + BDRV_POLL_WHILE(bs, s->read_reply_co); | ||
58 | } | ||
59 | |||
60 | static void nbd_teardown_connection(BlockDriverState *bs) | ||
61 | @@ -XXX,XX +XXX,XX @@ static void nbd_teardown_connection(BlockDriverState *bs) | ||
62 | qio_channel_shutdown(client->ioc, | ||
63 | QIO_CHANNEL_SHUTDOWN_BOTH, | ||
64 | NULL); | ||
65 | - nbd_recv_coroutines_enter_all(client); | ||
66 | + nbd_recv_coroutines_enter_all(bs); | ||
67 | |||
68 | nbd_client_detach_aio_context(bs); | ||
69 | object_unref(OBJECT(client->sioc)); | ||
70 | @@ -XXX,XX +XXX,XX @@ static void nbd_teardown_connection(BlockDriverState *bs) | ||
71 | client->ioc = NULL; | ||
72 | } | ||
73 | |||
74 | -static void nbd_reply_ready(void *opaque) | ||
75 | +static coroutine_fn void nbd_read_reply_entry(void *opaque) | ||
76 | { | ||
77 | - BlockDriverState *bs = opaque; | ||
78 | - NBDClientSession *s = nbd_get_client_session(bs); | ||
79 | + NBDClientSession *s = opaque; | ||
80 | uint64_t i; | ||
81 | int ret; | ||
82 | |||
83 | - if (!s->ioc) { /* Already closed */ | ||
84 | - return; | ||
85 | - } | ||
86 | - | ||
87 | - if (s->reply.handle == 0) { | ||
88 | - /* No reply already in flight. Fetch a header. It is possible | ||
89 | - * that another thread has done the same thing in parallel, so | ||
90 | - * the socket is not readable anymore. | ||
91 | - */ | ||
92 | + for (;;) { | ||
93 | + assert(s->reply.handle == 0); | ||
94 | ret = nbd_receive_reply(s->ioc, &s->reply); | ||
95 | - if (ret == -EAGAIN) { | ||
96 | - return; | ||
97 | - } | ||
98 | if (ret < 0) { | ||
99 | - s->reply.handle = 0; | ||
100 | - goto fail; | ||
101 | + break; | ||
102 | } | ||
103 | - } | ||
104 | |||
105 | - /* There's no need for a mutex on the receive side, because the | ||
106 | - * handler acts as a synchronization point and ensures that only | ||
107 | - * one coroutine is called until the reply finishes. */ | ||
108 | - i = HANDLE_TO_INDEX(s, s->reply.handle); | ||
109 | - if (i >= MAX_NBD_REQUESTS) { | ||
110 | - goto fail; | ||
111 | - } | ||
112 | + /* There's no need for a mutex on the receive side, because the | ||
113 | + * handler acts as a synchronization point and ensures that only | ||
114 | + * one coroutine is called until the reply finishes. | ||
115 | + */ | ||
116 | + i = HANDLE_TO_INDEX(s, s->reply.handle); | ||
117 | + if (i >= MAX_NBD_REQUESTS || !s->recv_coroutine[i]) { | ||
118 | + break; | ||
119 | + } | ||
120 | |||
121 | - if (s->recv_coroutine[i]) { | ||
122 | - qemu_coroutine_enter(s->recv_coroutine[i]); | ||
123 | - return; | ||
124 | + /* We're woken up by the recv_coroutine itself. Note that there | ||
125 | + * is no race between yielding and reentering read_reply_co. This | ||
126 | + * is because: | ||
127 | + * | ||
128 | + * - if recv_coroutine[i] runs on the same AioContext, it is only | ||
129 | + * entered after we yield | ||
130 | + * | ||
131 | + * - if recv_coroutine[i] runs on a different AioContext, reentering | ||
132 | + * read_reply_co happens through a bottom half, which can only | ||
133 | + * run after we yield. | ||
134 | + */ | ||
135 | + aio_co_wake(s->recv_coroutine[i]); | ||
136 | + qemu_coroutine_yield(); | ||
137 | } | ||
138 | - | ||
139 | -fail: | ||
140 | - nbd_teardown_connection(bs); | ||
141 | -} | ||
142 | - | ||
143 | -static void nbd_restart_write(void *opaque) | ||
144 | -{ | ||
145 | - BlockDriverState *bs = opaque; | ||
146 | - | ||
147 | - qemu_coroutine_enter(nbd_get_client_session(bs)->send_coroutine); | ||
148 | + s->read_reply_co = NULL; | ||
149 | } | ||
150 | |||
151 | static int nbd_co_send_request(BlockDriverState *bs, | ||
152 | @@ -XXX,XX +XXX,XX @@ static int nbd_co_send_request(BlockDriverState *bs, | ||
153 | QEMUIOVector *qiov) | ||
154 | { | ||
155 | NBDClientSession *s = nbd_get_client_session(bs); | ||
156 | - AioContext *aio_context; | ||
157 | int rc, ret, i; | ||
158 | |||
159 | qemu_co_mutex_lock(&s->send_mutex); | ||
160 | @@ -XXX,XX +XXX,XX @@ static int nbd_co_send_request(BlockDriverState *bs, | ||
161 | return -EPIPE; | ||
162 | } | ||
163 | |||
164 | - s->send_coroutine = qemu_coroutine_self(); | ||
165 | - aio_context = bdrv_get_aio_context(bs); | ||
166 | - | ||
167 | - aio_set_fd_handler(aio_context, s->sioc->fd, false, | ||
168 | - nbd_reply_ready, nbd_restart_write, NULL, bs); | ||
169 | if (qiov) { | ||
170 | qio_channel_set_cork(s->ioc, true); | ||
171 | rc = nbd_send_request(s->ioc, request); | ||
172 | @@ -XXX,XX +XXX,XX @@ static int nbd_co_send_request(BlockDriverState *bs, | ||
173 | } else { | ||
174 | rc = nbd_send_request(s->ioc, request); | ||
175 | } | ||
176 | - aio_set_fd_handler(aio_context, s->sioc->fd, false, | ||
177 | - nbd_reply_ready, NULL, NULL, bs); | ||
178 | - s->send_coroutine = NULL; | ||
179 | qemu_co_mutex_unlock(&s->send_mutex); | ||
180 | return rc; | ||
181 | } | ||
182 | @@ -XXX,XX +XXX,XX @@ static void nbd_co_receive_reply(NBDClientSession *s, | ||
183 | { | ||
184 | int ret; | ||
185 | |||
186 | - /* Wait until we're woken up by the read handler. TODO: perhaps | ||
187 | - * peek at the next reply and avoid yielding if it's ours? */ | ||
188 | + /* Wait until we're woken up by nbd_read_reply_entry. */ | ||
189 | qemu_coroutine_yield(); | ||
190 | *reply = s->reply; | ||
191 | if (reply->handle != request->handle || | ||
192 | @@ -XXX,XX +XXX,XX @@ static void nbd_coroutine_start(NBDClientSession *s, | ||
193 | /* s->recv_coroutine[i] is set as soon as we get the send_lock. */ | ||
194 | } | ||
195 | |||
196 | -static void nbd_coroutine_end(NBDClientSession *s, | ||
197 | +static void nbd_coroutine_end(BlockDriverState *bs, | ||
198 | NBDRequest *request) | ||
199 | { | ||
200 | + NBDClientSession *s = nbd_get_client_session(bs); | ||
201 | int i = HANDLE_TO_INDEX(s, request->handle); | ||
202 | + | ||
203 | s->recv_coroutine[i] = NULL; | ||
204 | - if (s->in_flight-- == MAX_NBD_REQUESTS) { | ||
205 | - qemu_co_queue_next(&s->free_sema); | ||
206 | + s->in_flight--; | ||
207 | + qemu_co_queue_next(&s->free_sema); | ||
208 | + | ||
209 | + /* Kick the read_reply_co to get the next reply. */ | ||
210 | + if (s->read_reply_co) { | ||
211 | + aio_co_wake(s->read_reply_co); | ||
212 | } | ||
213 | } | ||
214 | |||
215 | @@ -XXX,XX +XXX,XX @@ int nbd_client_co_preadv(BlockDriverState *bs, uint64_t offset, | ||
216 | } else { | ||
217 | nbd_co_receive_reply(client, &request, &reply, qiov); | ||
218 | } | ||
219 | - nbd_coroutine_end(client, &request); | ||
220 | + nbd_coroutine_end(bs, &request); | ||
221 | return -reply.error; | ||
222 | } | ||
223 | |||
224 | @@ -XXX,XX +XXX,XX @@ int nbd_client_co_pwritev(BlockDriverState *bs, uint64_t offset, | ||
225 | } else { | ||
226 | nbd_co_receive_reply(client, &request, &reply, NULL); | ||
227 | } | ||
228 | - nbd_coroutine_end(client, &request); | ||
229 | + nbd_coroutine_end(bs, &request); | ||
230 | return -reply.error; | ||
231 | } | ||
232 | |||
233 | @@ -XXX,XX +XXX,XX @@ int nbd_client_co_pwrite_zeroes(BlockDriverState *bs, int64_t offset, | ||
234 | } else { | ||
235 | nbd_co_receive_reply(client, &request, &reply, NULL); | ||
236 | } | ||
237 | - nbd_coroutine_end(client, &request); | ||
238 | + nbd_coroutine_end(bs, &request); | ||
239 | return -reply.error; | ||
240 | } | ||
241 | |||
242 | @@ -XXX,XX +XXX,XX @@ int nbd_client_co_flush(BlockDriverState *bs) | ||
243 | } else { | ||
244 | nbd_co_receive_reply(client, &request, &reply, NULL); | ||
245 | } | ||
246 | - nbd_coroutine_end(client, &request); | ||
247 | + nbd_coroutine_end(bs, &request); | ||
248 | return -reply.error; | ||
249 | } | ||
250 | |||
251 | @@ -XXX,XX +XXX,XX @@ int nbd_client_co_pdiscard(BlockDriverState *bs, int64_t offset, int count) | ||
252 | } else { | ||
253 | nbd_co_receive_reply(client, &request, &reply, NULL); | ||
254 | } | ||
255 | - nbd_coroutine_end(client, &request); | ||
256 | + nbd_coroutine_end(bs, &request); | ||
257 | return -reply.error; | ||
258 | |||
259 | } | ||
260 | |||
261 | void nbd_client_detach_aio_context(BlockDriverState *bs) | ||
262 | { | ||
263 | - aio_set_fd_handler(bdrv_get_aio_context(bs), | ||
264 | - nbd_get_client_session(bs)->sioc->fd, | ||
265 | - false, NULL, NULL, NULL, NULL); | ||
266 | + NBDClientSession *client = nbd_get_client_session(bs); | ||
267 | + qio_channel_detach_aio_context(QIO_CHANNEL(client->sioc)); | ||
268 | } | ||
269 | |||
270 | void nbd_client_attach_aio_context(BlockDriverState *bs, | ||
271 | AioContext *new_context) | ||
272 | { | ||
273 | - aio_set_fd_handler(new_context, nbd_get_client_session(bs)->sioc->fd, | ||
274 | - false, nbd_reply_ready, NULL, NULL, bs); | ||
275 | + NBDClientSession *client = nbd_get_client_session(bs); | ||
276 | + qio_channel_attach_aio_context(QIO_CHANNEL(client->sioc), new_context); | ||
277 | + aio_co_schedule(new_context, client->read_reply_co); | ||
278 | } | ||
279 | |||
280 | void nbd_client_close(BlockDriverState *bs) | ||
281 | @@ -XXX,XX +XXX,XX @@ int nbd_client_init(BlockDriverState *bs, | ||
282 | /* Now that we're connected, set the socket to be non-blocking and | ||
283 | * kick the reply mechanism. */ | ||
284 | qio_channel_set_blocking(QIO_CHANNEL(sioc), false, NULL); | ||
285 | - | ||
286 | + client->read_reply_co = qemu_coroutine_create(nbd_read_reply_entry, client); | ||
287 | nbd_client_attach_aio_context(bs, bdrv_get_aio_context(bs)); | ||
288 | |||
289 | logout("Established connection with NBD server\n"); | ||
290 | diff --git a/nbd/client.c b/nbd/client.c | ||
291 | index XXXXXXX..XXXXXXX 100644 | ||
292 | --- a/nbd/client.c | ||
293 | +++ b/nbd/client.c | ||
294 | @@ -XXX,XX +XXX,XX @@ ssize_t nbd_receive_reply(QIOChannel *ioc, NBDReply *reply) | ||
295 | ssize_t ret; | ||
296 | |||
297 | ret = read_sync(ioc, buf, sizeof(buf)); | ||
298 | - if (ret < 0) { | ||
299 | + if (ret <= 0) { | ||
300 | return ret; | ||
301 | } | ||
302 | |||
303 | diff --git a/nbd/common.c b/nbd/common.c | ||
304 | index XXXXXXX..XXXXXXX 100644 | ||
305 | --- a/nbd/common.c | ||
306 | +++ b/nbd/common.c | ||
307 | @@ -XXX,XX +XXX,XX @@ ssize_t nbd_wr_syncv(QIOChannel *ioc, | ||
308 | } | ||
309 | if (len == QIO_CHANNEL_ERR_BLOCK) { | ||
310 | if (qemu_in_coroutine()) { | ||
311 | - /* XXX figure out if we can create a variant on | ||
312 | - * qio_channel_yield() that works with AIO contexts | ||
313 | - * and consider using that in this branch */ | ||
314 | - qemu_coroutine_yield(); | ||
315 | - } else if (done) { | ||
316 | - /* XXX this is needed by nbd_reply_ready. */ | ||
317 | - qio_channel_wait(ioc, | ||
318 | - do_read ? G_IO_IN : G_IO_OUT); | ||
319 | + qio_channel_yield(ioc, do_read ? G_IO_IN : G_IO_OUT); | ||
320 | } else { | ||
321 | return -EAGAIN; | ||
322 | } | ||
323 | diff --git a/nbd/server.c b/nbd/server.c | ||
324 | index XXXXXXX..XXXXXXX 100644 | ||
325 | --- a/nbd/server.c | ||
326 | +++ b/nbd/server.c | ||
327 | @@ -XXX,XX +XXX,XX @@ struct NBDClient { | ||
328 | CoMutex send_lock; | ||
329 | Coroutine *send_coroutine; | ||
330 | |||
331 | - bool can_read; | ||
332 | - | ||
333 | QTAILQ_ENTRY(NBDClient) next; | ||
334 | int nb_requests; | ||
335 | bool closing; | ||
336 | @@ -XXX,XX +XXX,XX @@ struct NBDClient { | ||
337 | |||
338 | /* That's all folks */ | ||
339 | |||
340 | -static void nbd_set_handlers(NBDClient *client); | ||
341 | -static void nbd_unset_handlers(NBDClient *client); | ||
342 | -static void nbd_update_can_read(NBDClient *client); | ||
343 | +static void nbd_client_receive_next_request(NBDClient *client); | ||
344 | |||
345 | static gboolean nbd_negotiate_continue(QIOChannel *ioc, | ||
346 | GIOCondition condition, | ||
347 | @@ -XXX,XX +XXX,XX @@ void nbd_client_put(NBDClient *client) | ||
348 | */ | ||
349 | assert(client->closing); | ||
350 | |||
351 | - nbd_unset_handlers(client); | ||
352 | + qio_channel_detach_aio_context(client->ioc); | ||
353 | object_unref(OBJECT(client->sioc)); | ||
354 | object_unref(OBJECT(client->ioc)); | ||
355 | if (client->tlscreds) { | ||
356 | @@ -XXX,XX +XXX,XX @@ static NBDRequestData *nbd_request_get(NBDClient *client) | ||
357 | |||
358 | assert(client->nb_requests <= MAX_NBD_REQUESTS - 1); | ||
359 | client->nb_requests++; | ||
360 | - nbd_update_can_read(client); | ||
361 | |||
362 | req = g_new0(NBDRequestData, 1); | ||
363 | nbd_client_get(client); | ||
364 | @@ -XXX,XX +XXX,XX @@ static void nbd_request_put(NBDRequestData *req) | ||
365 | g_free(req); | ||
366 | |||
367 | client->nb_requests--; | ||
368 | - nbd_update_can_read(client); | ||
369 | + nbd_client_receive_next_request(client); | ||
370 | + | ||
371 | nbd_client_put(client); | ||
372 | } | ||
373 | |||
374 | @@ -XXX,XX +XXX,XX @@ static void blk_aio_attached(AioContext *ctx, void *opaque) | ||
375 | exp->ctx = ctx; | ||
376 | |||
377 | QTAILQ_FOREACH(client, &exp->clients, next) { | ||
378 | - nbd_set_handlers(client); | ||
379 | + qio_channel_attach_aio_context(client->ioc, ctx); | ||
380 | + if (client->recv_coroutine) { | ||
381 | + aio_co_schedule(ctx, client->recv_coroutine); | ||
382 | + } | ||
383 | + if (client->send_coroutine) { | ||
384 | + aio_co_schedule(ctx, client->send_coroutine); | ||
385 | + } | ||
386 | } | ||
387 | } | ||
388 | |||
389 | @@ -XXX,XX +XXX,XX @@ static void blk_aio_detach(void *opaque) | ||
390 | TRACE("Export %s: Detaching clients from AIO context %p\n", exp->name, exp->ctx); | ||
391 | |||
392 | QTAILQ_FOREACH(client, &exp->clients, next) { | ||
393 | - nbd_unset_handlers(client); | ||
394 | + qio_channel_detach_aio_context(client->ioc); | ||
395 | } | ||
396 | |||
397 | exp->ctx = NULL; | ||
398 | @@ -XXX,XX +XXX,XX @@ static ssize_t nbd_co_send_reply(NBDRequestData *req, NBDReply *reply, | ||
399 | g_assert(qemu_in_coroutine()); | ||
400 | qemu_co_mutex_lock(&client->send_lock); | ||
401 | client->send_coroutine = qemu_coroutine_self(); | ||
402 | - nbd_set_handlers(client); | ||
403 | |||
404 | if (!len) { | ||
405 | rc = nbd_send_reply(client->ioc, reply); | ||
406 | @@ -XXX,XX +XXX,XX @@ static ssize_t nbd_co_send_reply(NBDRequestData *req, NBDReply *reply, | ||
407 | } | ||
408 | |||
409 | client->send_coroutine = NULL; | ||
410 | - nbd_set_handlers(client); | ||
411 | qemu_co_mutex_unlock(&client->send_lock); | ||
412 | return rc; | ||
413 | } | ||
414 | @@ -XXX,XX +XXX,XX @@ static ssize_t nbd_co_receive_request(NBDRequestData *req, | ||
415 | ssize_t rc; | ||
416 | |||
417 | g_assert(qemu_in_coroutine()); | ||
418 | - client->recv_coroutine = qemu_coroutine_self(); | ||
419 | - nbd_update_can_read(client); | ||
420 | - | ||
421 | + assert(client->recv_coroutine == qemu_coroutine_self()); | ||
422 | rc = nbd_receive_request(client->ioc, request); | ||
423 | if (rc < 0) { | ||
424 | if (rc != -EAGAIN) { | ||
425 | @@ -XXX,XX +XXX,XX @@ static ssize_t nbd_co_receive_request(NBDRequestData *req, | ||
426 | |||
427 | out: | ||
428 | client->recv_coroutine = NULL; | ||
429 | - nbd_update_can_read(client); | ||
430 | + nbd_client_receive_next_request(client); | ||
431 | |||
432 | return rc; | ||
433 | } | ||
434 | |||
435 | -static void nbd_trip(void *opaque) | ||
436 | +/* Owns a reference to the NBDClient passed as opaque. */ | ||
437 | +static coroutine_fn void nbd_trip(void *opaque) | ||
438 | { | ||
439 | NBDClient *client = opaque; | ||
440 | NBDExport *exp = client->exp; | ||
441 | NBDRequestData *req; | ||
442 | - NBDRequest request; | ||
443 | + NBDRequest request = { 0 }; /* GCC thinks it can be used uninitialized */ | ||
444 | NBDReply reply; | ||
445 | ssize_t ret; | ||
446 | int flags; | ||
447 | |||
448 | TRACE("Reading request."); | ||
449 | if (client->closing) { | ||
450 | + nbd_client_put(client); | ||
451 | return; | ||
452 | } | ||
453 | |||
454 | @@ -XXX,XX +XXX,XX @@ static void nbd_trip(void *opaque) | ||
455 | |||
456 | done: | ||
457 | nbd_request_put(req); | ||
458 | + nbd_client_put(client); | ||
459 | return; | ||
460 | |||
461 | out: | ||
462 | nbd_request_put(req); | ||
463 | client_close(client); | ||
464 | + nbd_client_put(client); | ||
465 | } | ||
466 | |||
467 | -static void nbd_read(void *opaque) | ||
468 | +static void nbd_client_receive_next_request(NBDClient *client) | ||
469 | { | ||
470 | - NBDClient *client = opaque; | ||
471 | - | ||
472 | - if (client->recv_coroutine) { | ||
473 | - qemu_coroutine_enter(client->recv_coroutine); | ||
474 | - } else { | ||
475 | - qemu_coroutine_enter(qemu_coroutine_create(nbd_trip, client)); | ||
476 | - } | ||
477 | -} | ||
478 | - | ||
479 | -static void nbd_restart_write(void *opaque) | ||
480 | -{ | ||
481 | - NBDClient *client = opaque; | ||
482 | - | ||
483 | - qemu_coroutine_enter(client->send_coroutine); | ||
484 | -} | ||
485 | - | ||
486 | -static void nbd_set_handlers(NBDClient *client) | ||
487 | -{ | ||
488 | - if (client->exp && client->exp->ctx) { | ||
489 | - aio_set_fd_handler(client->exp->ctx, client->sioc->fd, true, | ||
490 | - client->can_read ? nbd_read : NULL, | ||
491 | - client->send_coroutine ? nbd_restart_write : NULL, | ||
492 | - NULL, client); | ||
493 | - } | ||
494 | -} | ||
495 | - | ||
496 | -static void nbd_unset_handlers(NBDClient *client) | ||
497 | -{ | ||
498 | - if (client->exp && client->exp->ctx) { | ||
499 | - aio_set_fd_handler(client->exp->ctx, client->sioc->fd, true, NULL, | ||
500 | - NULL, NULL, NULL); | ||
501 | - } | ||
502 | -} | ||
503 | - | ||
504 | -static void nbd_update_can_read(NBDClient *client) | ||
505 | -{ | ||
506 | - bool can_read = client->recv_coroutine || | ||
507 | - client->nb_requests < MAX_NBD_REQUESTS; | ||
508 | - | ||
509 | - if (can_read != client->can_read) { | ||
510 | - client->can_read = can_read; | ||
511 | - nbd_set_handlers(client); | ||
512 | - | ||
513 | - /* There is no need to invoke aio_notify(), since aio_set_fd_handler() | ||
514 | - * in nbd_set_handlers() will have taken care of that */ | ||
515 | + if (!client->recv_coroutine && client->nb_requests < MAX_NBD_REQUESTS) { | ||
516 | + nbd_client_get(client); | ||
517 | + client->recv_coroutine = qemu_coroutine_create(nbd_trip, client); | ||
518 | + aio_co_schedule(client->exp->ctx, client->recv_coroutine); | ||
519 | } | ||
520 | } | ||
521 | |||
522 | @@ -XXX,XX +XXX,XX @@ static coroutine_fn void nbd_co_client_start(void *opaque) | ||
523 | goto out; | ||
524 | } | ||
525 | qemu_co_mutex_init(&client->send_lock); | ||
526 | - nbd_set_handlers(client); | ||
527 | |||
528 | if (exp) { | ||
529 | QTAILQ_INSERT_TAIL(&exp->clients, client, next); | ||
530 | } | ||
531 | + | ||
532 | + nbd_client_receive_next_request(client); | ||
533 | + | ||
534 | out: | ||
535 | g_free(data); | ||
536 | } | ||
537 | @@ -XXX,XX +XXX,XX @@ void nbd_client_new(NBDExport *exp, | ||
538 | object_ref(OBJECT(client->sioc)); | ||
539 | client->ioc = QIO_CHANNEL(sioc); | ||
540 | object_ref(OBJECT(client->ioc)); | ||
541 | - client->can_read = true; | ||
542 | client->close = close_fn; | ||
543 | |||
544 | data->client = client; | ||
545 | -- | ||
546 | 2.9.3 | ||
547 | |||
548 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
1 | 2 | ||
3 | As a small step towards the introduction of multiqueue, we want | ||
4 | coroutines to remain on the same AioContext that started them, | ||
5 | unless they are moved explicitly with e.g. aio_co_schedule. This patch | ||
6 | avoids that coroutines switch AioContext when they use a CoMutex. | ||
7 | For now it does not make much of a difference, because the CoMutex | ||
8 | is not thread-safe and the AioContext itself is used to protect the | ||
9 | CoMutex from concurrent access. However, this is going to change. | ||
10 | |||
11 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
12 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
13 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
14 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> | ||
15 | Message-id: 20170213135235.12274-9-pbonzini@redhat.com | ||
16 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
17 | --- | ||
18 | util/qemu-coroutine-lock.c | 5 ++--- | ||
19 | util/trace-events | 1 - | ||
20 | 2 files changed, 2 insertions(+), 4 deletions(-) | ||
21 | |||
22 | diff --git a/util/qemu-coroutine-lock.c b/util/qemu-coroutine-lock.c | ||
23 | index XXXXXXX..XXXXXXX 100644 | ||
24 | --- a/util/qemu-coroutine-lock.c | ||
25 | +++ b/util/qemu-coroutine-lock.c | ||
26 | @@ -XXX,XX +XXX,XX @@ | ||
27 | #include "qemu/coroutine.h" | ||
28 | #include "qemu/coroutine_int.h" | ||
29 | #include "qemu/queue.h" | ||
30 | +#include "block/aio.h" | ||
31 | #include "trace.h" | ||
32 | |||
33 | void qemu_co_queue_init(CoQueue *queue) | ||
34 | @@ -XXX,XX +XXX,XX @@ void qemu_co_queue_run_restart(Coroutine *co) | ||
35 | |||
36 | static bool qemu_co_queue_do_restart(CoQueue *queue, bool single) | ||
37 | { | ||
38 | - Coroutine *self = qemu_coroutine_self(); | ||
39 | Coroutine *next; | ||
40 | |||
41 | if (QSIMPLEQ_EMPTY(&queue->entries)) { | ||
42 | @@ -XXX,XX +XXX,XX @@ static bool qemu_co_queue_do_restart(CoQueue *queue, bool single) | ||
43 | |||
44 | while ((next = QSIMPLEQ_FIRST(&queue->entries)) != NULL) { | ||
45 | QSIMPLEQ_REMOVE_HEAD(&queue->entries, co_queue_next); | ||
46 | - QSIMPLEQ_INSERT_TAIL(&self->co_queue_wakeup, next, co_queue_next); | ||
47 | - trace_qemu_co_queue_next(next); | ||
48 | + aio_co_wake(next); | ||
49 | if (single) { | ||
50 | break; | ||
51 | } | ||
52 | diff --git a/util/trace-events b/util/trace-events | ||
53 | index XXXXXXX..XXXXXXX 100644 | ||
54 | --- a/util/trace-events | ||
55 | +++ b/util/trace-events | ||
56 | @@ -XXX,XX +XXX,XX @@ qemu_coroutine_terminate(void *co) "self %p" | ||
57 | |||
58 | # util/qemu-coroutine-lock.c | ||
59 | qemu_co_queue_run_restart(void *co) "co %p" | ||
60 | -qemu_co_queue_next(void *nxt) "next %p" | ||
61 | qemu_co_mutex_lock_entry(void *mutex, void *self) "mutex %p self %p" | ||
62 | qemu_co_mutex_lock_return(void *mutex, void *self) "mutex %p self %p" | ||
63 | qemu_co_mutex_unlock_entry(void *mutex, void *self) "mutex %p self %p" | ||
64 | -- | ||
65 | 2.9.3 | ||
66 | |||
67 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
1 | 2 | ||
3 | Keep the coroutine on the same AioContext. Without this change, | ||
4 | there would be a race between yielding the coroutine and reentering it. | ||
5 | While the race cannot happen now, because the code only runs from a single | ||
6 | AioContext, this will change with multiqueue support in the block layer. | ||
7 | |||
8 | While doing the change, replace custom bottom half with aio_co_schedule. | ||
9 | |||
10 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
11 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
12 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
13 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> | ||
14 | Message-id: 20170213135235.12274-10-pbonzini@redhat.com | ||
15 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
16 | --- | ||
17 | block/blkdebug.c | 9 +-------- | ||
18 | 1 file changed, 1 insertion(+), 8 deletions(-) | ||
19 | |||
20 | diff --git a/block/blkdebug.c b/block/blkdebug.c | ||
21 | index XXXXXXX..XXXXXXX 100644 | ||
22 | --- a/block/blkdebug.c | ||
23 | +++ b/block/blkdebug.c | ||
24 | @@ -XXX,XX +XXX,XX @@ out: | ||
25 | return ret; | ||
26 | } | ||
27 | |||
28 | -static void error_callback_bh(void *opaque) | ||
29 | -{ | ||
30 | - Coroutine *co = opaque; | ||
31 | - qemu_coroutine_enter(co); | ||
32 | -} | ||
33 | - | ||
34 | static int inject_error(BlockDriverState *bs, BlkdebugRule *rule) | ||
35 | { | ||
36 | BDRVBlkdebugState *s = bs->opaque; | ||
37 | @@ -XXX,XX +XXX,XX @@ static int inject_error(BlockDriverState *bs, BlkdebugRule *rule) | ||
38 | } | ||
39 | |||
40 | if (!immediately) { | ||
41 | - aio_bh_schedule_oneshot(bdrv_get_aio_context(bs), error_callback_bh, | ||
42 | - qemu_coroutine_self()); | ||
43 | + aio_co_schedule(qemu_get_current_aio_context(), qemu_coroutine_self()); | ||
44 | qemu_coroutine_yield(); | ||
45 | } | ||
46 | |||
47 | -- | ||
48 | 2.9.3 | ||
49 | |||
50 | diff view generated by jsdifflib |
1 | From: Stefano Garzarella <sgarzare@redhat.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | This patch adds the support of DISCARD and WRITE_ZEROES commands, | 3 | qed_aio_start_io and qed_aio_next_io will not have to acquire/release |
4 | that have been introduced in the virtio-blk protocol to have | 4 | the AioContext, while qed_aio_next_io_cb will. Split the functionality |
5 | better performance when using SSD backend. | 5 | and gain a little type-safety in the process. |
6 | 6 | ||
7 | We support only one segment per request since multiple segments | ||
8 | are not widely used and there are no userspace APIs that allow | ||
9 | applications to submit multiple segments in a single call. | ||
10 | |||
11 | Reviewed-by: Michael S. Tsirkin <mst@redhat.com> | ||
12 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | 7 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> |
13 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | 8 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
14 | Acked-by: Pankaj Gupta <pagupta@redhat.com> | 9 | Reviewed-by: Fam Zheng <famz@redhat.com> |
15 | Message-id: 20190208134950.187665-5-sgarzare@redhat.com | 10 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> |
11 | Message-id: 20170213135235.12274-11-pbonzini@redhat.com | ||
16 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 12 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
17 | --- | 13 | --- |
18 | include/hw/virtio/virtio-blk.h | 2 + | 14 | block/qed.c | 39 +++++++++++++++++++++++++-------------- |
19 | hw/block/virtio-blk.c | 184 +++++++++++++++++++++++++++++++++ | 15 | 1 file changed, 25 insertions(+), 14 deletions(-) |
20 | 2 files changed, 186 insertions(+) | ||
21 | 16 | ||
22 | diff --git a/include/hw/virtio/virtio-blk.h b/include/hw/virtio/virtio-blk.h | 17 | diff --git a/block/qed.c b/block/qed.c |
23 | index XXXXXXX..XXXXXXX 100644 | 18 | index XXXXXXX..XXXXXXX 100644 |
24 | --- a/include/hw/virtio/virtio-blk.h | 19 | --- a/block/qed.c |
25 | +++ b/include/hw/virtio/virtio-blk.h | 20 | +++ b/block/qed.c |
26 | @@ -XXX,XX +XXX,XX @@ struct VirtIOBlkConf | 21 | @@ -XXX,XX +XXX,XX @@ static CachedL2Table *qed_new_l2_table(BDRVQEDState *s) |
27 | uint32_t request_merging; | 22 | return l2_table; |
28 | uint16_t num_queues; | ||
29 | uint16_t queue_size; | ||
30 | + uint32_t max_discard_sectors; | ||
31 | + uint32_t max_write_zeroes_sectors; | ||
32 | }; | ||
33 | |||
34 | struct VirtIOBlockDataPlane; | ||
35 | diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c | ||
36 | index XXXXXXX..XXXXXXX 100644 | ||
37 | --- a/hw/block/virtio-blk.c | ||
38 | +++ b/hw/block/virtio-blk.c | ||
39 | @@ -XXX,XX +XXX,XX @@ out: | ||
40 | aio_context_release(blk_get_aio_context(s->conf.conf.blk)); | ||
41 | } | 23 | } |
42 | 24 | ||
43 | +static void virtio_blk_discard_write_zeroes_complete(void *opaque, int ret) | 25 | -static void qed_aio_next_io(void *opaque, int ret); |
26 | +static void qed_aio_next_io(QEDAIOCB *acb, int ret); | ||
27 | + | ||
28 | +static void qed_aio_start_io(QEDAIOCB *acb) | ||
44 | +{ | 29 | +{ |
45 | + VirtIOBlockReq *req = opaque; | 30 | + qed_aio_next_io(acb, 0); |
46 | + VirtIOBlock *s = req->dev; | ||
47 | + bool is_write_zeroes = (virtio_ldl_p(VIRTIO_DEVICE(s), &req->out.type) & | ||
48 | + ~VIRTIO_BLK_T_BARRIER) == VIRTIO_BLK_T_WRITE_ZEROES; | ||
49 | + | ||
50 | + aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); | ||
51 | + if (ret) { | ||
52 | + if (virtio_blk_handle_rw_error(req, -ret, false, is_write_zeroes)) { | ||
53 | + goto out; | ||
54 | + } | ||
55 | + } | ||
56 | + | ||
57 | + virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); | ||
58 | + if (is_write_zeroes) { | ||
59 | + block_acct_done(blk_get_stats(s->blk), &req->acct); | ||
60 | + } | ||
61 | + virtio_blk_free_request(req); | ||
62 | + | ||
63 | +out: | ||
64 | + aio_context_release(blk_get_aio_context(s->conf.conf.blk)); | ||
65 | +} | 31 | +} |
66 | + | 32 | + |
67 | #ifdef __linux__ | 33 | +static void qed_aio_next_io_cb(void *opaque, int ret) |
68 | 34 | +{ | |
69 | typedef struct { | 35 | + QEDAIOCB *acb = opaque; |
70 | @@ -XXX,XX +XXX,XX @@ static bool virtio_blk_sect_range_ok(VirtIOBlock *dev, | 36 | + |
71 | return true; | 37 | + qed_aio_next_io(acb, ret); |
38 | +} | ||
39 | |||
40 | static void qed_plug_allocating_write_reqs(BDRVQEDState *s) | ||
41 | { | ||
42 | @@ -XXX,XX +XXX,XX @@ static void qed_unplug_allocating_write_reqs(BDRVQEDState *s) | ||
43 | |||
44 | acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); | ||
45 | if (acb) { | ||
46 | - qed_aio_next_io(acb, 0); | ||
47 | + qed_aio_start_io(acb); | ||
48 | } | ||
72 | } | 49 | } |
73 | 50 | ||
74 | +static uint8_t virtio_blk_handle_discard_write_zeroes(VirtIOBlockReq *req, | 51 | @@ -XXX,XX +XXX,XX @@ static void qed_aio_complete(QEDAIOCB *acb, int ret) |
75 | + struct virtio_blk_discard_write_zeroes *dwz_hdr, bool is_write_zeroes) | 52 | QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next); |
76 | +{ | 53 | acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs); |
77 | + VirtIOBlock *s = req->dev; | 54 | if (acb) { |
78 | + VirtIODevice *vdev = VIRTIO_DEVICE(s); | 55 | - qed_aio_next_io(acb, 0); |
79 | + uint64_t sector; | 56 | + qed_aio_start_io(acb); |
80 | + uint32_t num_sectors, flags, max_sectors; | 57 | } else if (s->header.features & QED_F_NEED_CHECK) { |
81 | + uint8_t err_status; | 58 | qed_start_need_check_timer(s); |
82 | + int bytes; | 59 | } |
83 | + | 60 | @@ -XXX,XX +XXX,XX @@ static void qed_commit_l2_update(void *opaque, int ret) |
84 | + sector = virtio_ldq_p(vdev, &dwz_hdr->sector); | 61 | acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset); |
85 | + num_sectors = virtio_ldl_p(vdev, &dwz_hdr->num_sectors); | 62 | assert(acb->request.l2_table != NULL); |
86 | + flags = virtio_ldl_p(vdev, &dwz_hdr->flags); | 63 | |
87 | + max_sectors = is_write_zeroes ? s->conf.max_write_zeroes_sectors : | 64 | - qed_aio_next_io(opaque, ret); |
88 | + s->conf.max_discard_sectors; | 65 | + qed_aio_next_io(acb, ret); |
89 | + | 66 | } |
90 | + /* | 67 | |
91 | + * max_sectors is at most BDRV_REQUEST_MAX_SECTORS, this check | 68 | /** |
92 | + * make us sure that "num_sectors << BDRV_SECTOR_BITS" can fit in | 69 | @@ -XXX,XX +XXX,XX @@ static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset) |
93 | + * the integer variable. | 70 | if (need_alloc) { |
94 | + */ | 71 | /* Write out the whole new L2 table */ |
95 | + if (unlikely(num_sectors > max_sectors)) { | 72 | qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true, |
96 | + err_status = VIRTIO_BLK_S_IOERR; | 73 | - qed_aio_write_l1_update, acb); |
97 | + goto err; | 74 | + qed_aio_write_l1_update, acb); |
98 | + } | 75 | } else { |
99 | + | 76 | /* Write out only the updated part of the L2 table */ |
100 | + bytes = num_sectors << BDRV_SECTOR_BITS; | 77 | qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false, |
101 | + | 78 | - qed_aio_next_io, acb); |
102 | + if (unlikely(!virtio_blk_sect_range_ok(s, sector, bytes))) { | 79 | + qed_aio_next_io_cb, acb); |
103 | + err_status = VIRTIO_BLK_S_IOERR; | ||
104 | + goto err; | ||
105 | + } | ||
106 | + | ||
107 | + /* | ||
108 | + * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for discard | ||
109 | + * and write zeroes commands if any unknown flag is set. | ||
110 | + */ | ||
111 | + if (unlikely(flags & ~VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) { | ||
112 | + err_status = VIRTIO_BLK_S_UNSUPP; | ||
113 | + goto err; | ||
114 | + } | ||
115 | + | ||
116 | + if (is_write_zeroes) { /* VIRTIO_BLK_T_WRITE_ZEROES */ | ||
117 | + int blk_aio_flags = 0; | ||
118 | + | ||
119 | + if (flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP) { | ||
120 | + blk_aio_flags |= BDRV_REQ_MAY_UNMAP; | ||
121 | + } | ||
122 | + | ||
123 | + block_acct_start(blk_get_stats(s->blk), &req->acct, bytes, | ||
124 | + BLOCK_ACCT_WRITE); | ||
125 | + | ||
126 | + blk_aio_pwrite_zeroes(s->blk, sector << BDRV_SECTOR_BITS, | ||
127 | + bytes, blk_aio_flags, | ||
128 | + virtio_blk_discard_write_zeroes_complete, req); | ||
129 | + } else { /* VIRTIO_BLK_T_DISCARD */ | ||
130 | + /* | ||
131 | + * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for | ||
132 | + * discard commands if the unmap flag is set. | ||
133 | + */ | ||
134 | + if (unlikely(flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) { | ||
135 | + err_status = VIRTIO_BLK_S_UNSUPP; | ||
136 | + goto err; | ||
137 | + } | ||
138 | + | ||
139 | + blk_aio_pdiscard(s->blk, sector << BDRV_SECTOR_BITS, bytes, | ||
140 | + virtio_blk_discard_write_zeroes_complete, req); | ||
141 | + } | ||
142 | + | ||
143 | + return VIRTIO_BLK_S_OK; | ||
144 | + | ||
145 | +err: | ||
146 | + if (is_write_zeroes) { | ||
147 | + block_acct_invalid(blk_get_stats(s->blk), BLOCK_ACCT_WRITE); | ||
148 | + } | ||
149 | + return err_status; | ||
150 | +} | ||
151 | + | ||
152 | static int virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb) | ||
153 | { | ||
154 | uint32_t type; | ||
155 | @@ -XXX,XX +XXX,XX @@ static int virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb) | ||
156 | virtio_blk_free_request(req); | ||
157 | break; | ||
158 | } | 80 | } |
159 | + /* | 81 | return; |
160 | + * VIRTIO_BLK_T_DISCARD and VIRTIO_BLK_T_WRITE_ZEROES are defined with | 82 | |
161 | + * VIRTIO_BLK_T_OUT flag set. We masked this flag in the switch statement, | 83 | @@ -XXX,XX +XXX,XX @@ static void qed_aio_write_main(void *opaque, int ret) |
162 | + * so we must mask it for these requests, then we will check if it is set. | 84 | } |
163 | + */ | 85 | |
164 | + case VIRTIO_BLK_T_DISCARD & ~VIRTIO_BLK_T_OUT: | 86 | if (acb->find_cluster_ret == QED_CLUSTER_FOUND) { |
165 | + case VIRTIO_BLK_T_WRITE_ZEROES & ~VIRTIO_BLK_T_OUT: | 87 | - next_fn = qed_aio_next_io; |
166 | + { | 88 | + next_fn = qed_aio_next_io_cb; |
167 | + struct virtio_blk_discard_write_zeroes dwz_hdr; | 89 | } else { |
168 | + size_t out_len = iov_size(out_iov, out_num); | 90 | if (s->bs->backing) { |
169 | + bool is_write_zeroes = (type & ~VIRTIO_BLK_T_BARRIER) == | 91 | next_fn = qed_aio_write_flush_before_l2_update; |
170 | + VIRTIO_BLK_T_WRITE_ZEROES; | 92 | @@ -XXX,XX +XXX,XX @@ static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len) |
171 | + uint8_t err_status; | 93 | if (acb->flags & QED_AIOCB_ZERO) { |
172 | + | 94 | /* Skip ahead if the clusters are already zero */ |
173 | + /* | 95 | if (acb->find_cluster_ret == QED_CLUSTER_ZERO) { |
174 | + * Unsupported if VIRTIO_BLK_T_OUT is not set or the request contains | 96 | - qed_aio_next_io(acb, 0); |
175 | + * more than one segment. | 97 | + qed_aio_start_io(acb); |
176 | + */ | 98 | return; |
177 | + if (unlikely(!(type & VIRTIO_BLK_T_OUT) || | 99 | } |
178 | + out_len > sizeof(dwz_hdr))) { | 100 | |
179 | + virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP); | 101 | @@ -XXX,XX +XXX,XX @@ static void qed_aio_read_data(void *opaque, int ret, |
180 | + virtio_blk_free_request(req); | 102 | /* Handle zero cluster and backing file reads */ |
181 | + return 0; | 103 | if (ret == QED_CLUSTER_ZERO) { |
182 | + } | 104 | qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size); |
183 | + | 105 | - qed_aio_next_io(acb, 0); |
184 | + if (unlikely(iov_to_buf(out_iov, out_num, 0, &dwz_hdr, | 106 | + qed_aio_start_io(acb); |
185 | + sizeof(dwz_hdr)) != sizeof(dwz_hdr))) { | 107 | return; |
186 | + virtio_error(vdev, "virtio-blk discard/write_zeroes header" | 108 | } else if (ret != QED_CLUSTER_FOUND) { |
187 | + " too short"); | 109 | qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov, |
188 | + return -1; | 110 | - &acb->backing_qiov, qed_aio_next_io, acb); |
189 | + } | 111 | + &acb->backing_qiov, qed_aio_next_io_cb, acb); |
190 | + | ||
191 | + err_status = virtio_blk_handle_discard_write_zeroes(req, &dwz_hdr, | ||
192 | + is_write_zeroes); | ||
193 | + if (err_status != VIRTIO_BLK_S_OK) { | ||
194 | + virtio_blk_req_complete(req, err_status); | ||
195 | + virtio_blk_free_request(req); | ||
196 | + } | ||
197 | + | ||
198 | + break; | ||
199 | + } | ||
200 | default: | ||
201 | virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP); | ||
202 | virtio_blk_free_request(req); | ||
203 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config) | ||
204 | blkcfg.alignment_offset = 0; | ||
205 | blkcfg.wce = blk_enable_write_cache(s->blk); | ||
206 | virtio_stw_p(vdev, &blkcfg.num_queues, s->conf.num_queues); | ||
207 | + if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_DISCARD)) { | ||
208 | + virtio_stl_p(vdev, &blkcfg.max_discard_sectors, | ||
209 | + s->conf.max_discard_sectors); | ||
210 | + virtio_stl_p(vdev, &blkcfg.discard_sector_alignment, | ||
211 | + blk_size >> BDRV_SECTOR_BITS); | ||
212 | + /* | ||
213 | + * We support only one segment per request since multiple segments | ||
214 | + * are not widely used and there are no userspace APIs that allow | ||
215 | + * applications to submit multiple segments in a single call. | ||
216 | + */ | ||
217 | + virtio_stl_p(vdev, &blkcfg.max_discard_seg, 1); | ||
218 | + } | ||
219 | + if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_WRITE_ZEROES)) { | ||
220 | + virtio_stl_p(vdev, &blkcfg.max_write_zeroes_sectors, | ||
221 | + s->conf.max_write_zeroes_sectors); | ||
222 | + blkcfg.write_zeroes_may_unmap = 1; | ||
223 | + virtio_stl_p(vdev, &blkcfg.max_write_zeroes_seg, 1); | ||
224 | + } | ||
225 | memcpy(config, &blkcfg, sizeof(struct virtio_blk_config)); | ||
226 | } | ||
227 | |||
228 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_device_realize(DeviceState *dev, Error **errp) | ||
229 | return; | 112 | return; |
230 | } | 113 | } |
231 | 114 | ||
232 | + if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_DISCARD) && | 115 | BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO); |
233 | + (!conf->max_discard_sectors || | 116 | bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE, |
234 | + conf->max_discard_sectors > BDRV_REQUEST_MAX_SECTORS)) { | 117 | &acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE, |
235 | + error_setg(errp, "invalid max-discard-sectors property (%" PRIu32 ")" | 118 | - qed_aio_next_io, acb); |
236 | + ", must be between 1 and %d", | 119 | + qed_aio_next_io_cb, acb); |
237 | + conf->max_discard_sectors, (int)BDRV_REQUEST_MAX_SECTORS); | 120 | return; |
238 | + return; | 121 | |
239 | + } | 122 | err: |
240 | + | 123 | @@ -XXX,XX +XXX,XX @@ err: |
241 | + if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_WRITE_ZEROES) && | 124 | /** |
242 | + (!conf->max_write_zeroes_sectors || | 125 | * Begin next I/O or complete the request |
243 | + conf->max_write_zeroes_sectors > BDRV_REQUEST_MAX_SECTORS)) { | 126 | */ |
244 | + error_setg(errp, "invalid max-write-zeroes-sectors property (%" PRIu32 | 127 | -static void qed_aio_next_io(void *opaque, int ret) |
245 | + "), must be between 1 and %d", | 128 | +static void qed_aio_next_io(QEDAIOCB *acb, int ret) |
246 | + conf->max_write_zeroes_sectors, | 129 | { |
247 | + (int)BDRV_REQUEST_MAX_SECTORS); | 130 | - QEDAIOCB *acb = opaque; |
248 | + return; | 131 | BDRVQEDState *s = acb_to_s(acb); |
249 | + } | 132 | QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ? |
250 | + | 133 | qed_aio_write_data : qed_aio_read_data; |
251 | virtio_init(vdev, "virtio-blk", VIRTIO_ID_BLOCK, | 134 | @@ -XXX,XX +XXX,XX @@ static BlockAIOCB *qed_aio_setup(BlockDriverState *bs, |
252 | sizeof(struct virtio_blk_config)); | 135 | qemu_iovec_init(&acb->cur_qiov, qiov->niov); |
253 | 136 | ||
254 | @@ -XXX,XX +XXX,XX @@ static Property virtio_blk_properties[] = { | 137 | /* Start request */ |
255 | VIRTIO_BLK_F_DISCARD, true), | 138 | - qed_aio_next_io(acb, 0); |
256 | DEFINE_PROP_BIT64("write-zeroes", VirtIOBlock, host_features, | 139 | + qed_aio_start_io(acb); |
257 | VIRTIO_BLK_F_WRITE_ZEROES, true), | 140 | return &acb->common; |
258 | + DEFINE_PROP_UINT32("max-discard-sectors", VirtIOBlock, | 141 | } |
259 | + conf.max_discard_sectors, BDRV_REQUEST_MAX_SECTORS), | ||
260 | + DEFINE_PROP_UINT32("max-write-zeroes-sectors", VirtIOBlock, | ||
261 | + conf.max_write_zeroes_sectors, BDRV_REQUEST_MAX_SECTORS), | ||
262 | DEFINE_PROP_END_OF_LIST(), | ||
263 | }; | ||
264 | 142 | ||
265 | -- | 143 | -- |
266 | 2.20.1 | 144 | 2.9.3 |
267 | 145 | ||
268 | 146 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
1 | 2 | ||
3 | The AioContext data structures are now protected by list_lock and/or | ||
4 | they are walked with FOREACH_RCU primitives. There is no need anymore | ||
5 | to acquire the AioContext for the entire duration of aio_dispatch. | ||
6 | Instead, just acquire it before and after invoking the callbacks. | ||
7 | The next step is then to push it further down. | ||
8 | |||
9 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
10 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
11 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
12 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> | ||
13 | Message-id: 20170213135235.12274-12-pbonzini@redhat.com | ||
14 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
15 | --- | ||
16 | util/aio-posix.c | 25 +++++++++++-------------- | ||
17 | util/aio-win32.c | 15 +++++++-------- | ||
18 | util/async.c | 2 ++ | ||
19 | 3 files changed, 20 insertions(+), 22 deletions(-) | ||
20 | |||
21 | diff --git a/util/aio-posix.c b/util/aio-posix.c | ||
22 | index XXXXXXX..XXXXXXX 100644 | ||
23 | --- a/util/aio-posix.c | ||
24 | +++ b/util/aio-posix.c | ||
25 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx) | ||
26 | (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) && | ||
27 | aio_node_check(ctx, node->is_external) && | ||
28 | node->io_read) { | ||
29 | + aio_context_acquire(ctx); | ||
30 | node->io_read(node->opaque); | ||
31 | + aio_context_release(ctx); | ||
32 | |||
33 | /* aio_notify() does not count as progress */ | ||
34 | if (node->opaque != &ctx->notifier) { | ||
35 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx) | ||
36 | (revents & (G_IO_OUT | G_IO_ERR)) && | ||
37 | aio_node_check(ctx, node->is_external) && | ||
38 | node->io_write) { | ||
39 | + aio_context_acquire(ctx); | ||
40 | node->io_write(node->opaque); | ||
41 | + aio_context_release(ctx); | ||
42 | progress = true; | ||
43 | } | ||
44 | |||
45 | @@ -XXX,XX +XXX,XX @@ bool aio_dispatch(AioContext *ctx, bool dispatch_fds) | ||
46 | } | ||
47 | |||
48 | /* Run our timers */ | ||
49 | + aio_context_acquire(ctx); | ||
50 | progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
51 | + aio_context_release(ctx); | ||
52 | |||
53 | return progress; | ||
54 | } | ||
55 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
56 | int64_t timeout; | ||
57 | int64_t start = 0; | ||
58 | |||
59 | - aio_context_acquire(ctx); | ||
60 | - progress = false; | ||
61 | - | ||
62 | /* aio_notify can avoid the expensive event_notifier_set if | ||
63 | * everything (file descriptors, bottom halves, timers) will | ||
64 | * be re-evaluated before the next blocking poll(). This is | ||
65 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
66 | start = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | ||
67 | } | ||
68 | |||
69 | - if (try_poll_mode(ctx, blocking)) { | ||
70 | - progress = true; | ||
71 | - } else { | ||
72 | + aio_context_acquire(ctx); | ||
73 | + progress = try_poll_mode(ctx, blocking); | ||
74 | + aio_context_release(ctx); | ||
75 | + | ||
76 | + if (!progress) { | ||
77 | assert(npfd == 0); | ||
78 | |||
79 | /* fill pollfds */ | ||
80 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
81 | timeout = blocking ? aio_compute_timeout(ctx) : 0; | ||
82 | |||
83 | /* wait until next event */ | ||
84 | - if (timeout) { | ||
85 | - aio_context_release(ctx); | ||
86 | - } | ||
87 | if (aio_epoll_check_poll(ctx, pollfds, npfd, timeout)) { | ||
88 | AioHandler epoll_handler; | ||
89 | |||
90 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
91 | } else { | ||
92 | ret = qemu_poll_ns(pollfds, npfd, timeout); | ||
93 | } | ||
94 | - if (timeout) { | ||
95 | - aio_context_acquire(ctx); | ||
96 | - } | ||
97 | } | ||
98 | |||
99 | if (blocking) { | ||
100 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
101 | progress = true; | ||
102 | } | ||
103 | |||
104 | - aio_context_release(ctx); | ||
105 | - | ||
106 | return progress; | ||
107 | } | ||
108 | |||
109 | diff --git a/util/aio-win32.c b/util/aio-win32.c | ||
110 | index XXXXXXX..XXXXXXX 100644 | ||
111 | --- a/util/aio-win32.c | ||
112 | +++ b/util/aio-win32.c | ||
113 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) | ||
114 | (revents || event_notifier_get_handle(node->e) == event) && | ||
115 | node->io_notify) { | ||
116 | node->pfd.revents = 0; | ||
117 | + aio_context_acquire(ctx); | ||
118 | node->io_notify(node->e); | ||
119 | + aio_context_release(ctx); | ||
120 | |||
121 | /* aio_notify() does not count as progress */ | ||
122 | if (node->e != &ctx->notifier) { | ||
123 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) | ||
124 | (node->io_read || node->io_write)) { | ||
125 | node->pfd.revents = 0; | ||
126 | if ((revents & G_IO_IN) && node->io_read) { | ||
127 | + aio_context_acquire(ctx); | ||
128 | node->io_read(node->opaque); | ||
129 | + aio_context_release(ctx); | ||
130 | progress = true; | ||
131 | } | ||
132 | if ((revents & G_IO_OUT) && node->io_write) { | ||
133 | + aio_context_acquire(ctx); | ||
134 | node->io_write(node->opaque); | ||
135 | + aio_context_release(ctx); | ||
136 | progress = true; | ||
137 | } | ||
138 | |||
139 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
140 | int count; | ||
141 | int timeout; | ||
142 | |||
143 | - aio_context_acquire(ctx); | ||
144 | progress = false; | ||
145 | |||
146 | /* aio_notify can avoid the expensive event_notifier_set if | ||
147 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
148 | |||
149 | timeout = blocking && !have_select_revents | ||
150 | ? qemu_timeout_ns_to_ms(aio_compute_timeout(ctx)) : 0; | ||
151 | - if (timeout) { | ||
152 | - aio_context_release(ctx); | ||
153 | - } | ||
154 | ret = WaitForMultipleObjects(count, events, FALSE, timeout); | ||
155 | if (blocking) { | ||
156 | assert(first); | ||
157 | atomic_sub(&ctx->notify_me, 2); | ||
158 | } | ||
159 | - if (timeout) { | ||
160 | - aio_context_acquire(ctx); | ||
161 | - } | ||
162 | |||
163 | if (first) { | ||
164 | aio_notify_accept(ctx); | ||
165 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
166 | progress |= aio_dispatch_handlers(ctx, event); | ||
167 | } while (count > 0); | ||
168 | |||
169 | + aio_context_acquire(ctx); | ||
170 | progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
171 | - | ||
172 | aio_context_release(ctx); | ||
173 | return progress; | ||
174 | } | ||
175 | diff --git a/util/async.c b/util/async.c | ||
176 | index XXXXXXX..XXXXXXX 100644 | ||
177 | --- a/util/async.c | ||
178 | +++ b/util/async.c | ||
179 | @@ -XXX,XX +XXX,XX @@ int aio_bh_poll(AioContext *ctx) | ||
180 | ret = 1; | ||
181 | } | ||
182 | bh->idle = 0; | ||
183 | + aio_context_acquire(ctx); | ||
184 | aio_bh_call(bh); | ||
185 | + aio_context_release(ctx); | ||
186 | } | ||
187 | if (bh->deleted) { | ||
188 | deleted = true; | ||
189 | -- | ||
190 | 2.9.3 | ||
191 | |||
192 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | 1 | From: Paolo Bonzini <pbonzini@redhat.com> | |
2 | |||
3 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
4 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
5 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
6 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> | ||
7 | Message-id: 20170213135235.12274-13-pbonzini@redhat.com | ||
8 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
9 | --- | ||
10 | block/qed.h | 3 +++ | ||
11 | block/curl.c | 2 ++ | ||
12 | block/io.c | 5 +++++ | ||
13 | block/iscsi.c | 8 ++++++-- | ||
14 | block/null.c | 4 ++++ | ||
15 | block/qed.c | 12 ++++++++++++ | ||
16 | block/throttle-groups.c | 2 ++ | ||
17 | util/aio-posix.c | 2 -- | ||
18 | util/aio-win32.c | 2 -- | ||
19 | util/qemu-coroutine-sleep.c | 2 +- | ||
20 | 10 files changed, 35 insertions(+), 7 deletions(-) | ||
21 | |||
22 | diff --git a/block/qed.h b/block/qed.h | ||
23 | index XXXXXXX..XXXXXXX 100644 | ||
24 | --- a/block/qed.h | ||
25 | +++ b/block/qed.h | ||
26 | @@ -XXX,XX +XXX,XX @@ enum { | ||
27 | */ | ||
28 | typedef void QEDFindClusterFunc(void *opaque, int ret, uint64_t offset, size_t len); | ||
29 | |||
30 | +void qed_acquire(BDRVQEDState *s); | ||
31 | +void qed_release(BDRVQEDState *s); | ||
32 | + | ||
33 | /** | ||
34 | * Generic callback for chaining async callbacks | ||
35 | */ | ||
36 | diff --git a/block/curl.c b/block/curl.c | ||
37 | index XXXXXXX..XXXXXXX 100644 | ||
38 | --- a/block/curl.c | ||
39 | +++ b/block/curl.c | ||
40 | @@ -XXX,XX +XXX,XX @@ static void curl_multi_timeout_do(void *arg) | ||
41 | return; | ||
42 | } | ||
43 | |||
44 | + aio_context_acquire(s->aio_context); | ||
45 | curl_multi_socket_action(s->multi, CURL_SOCKET_TIMEOUT, 0, &running); | ||
46 | |||
47 | curl_multi_check_completion(s); | ||
48 | + aio_context_release(s->aio_context); | ||
49 | #else | ||
50 | abort(); | ||
51 | #endif | ||
52 | diff --git a/block/io.c b/block/io.c | ||
53 | index XXXXXXX..XXXXXXX 100644 | ||
54 | --- a/block/io.c | ||
55 | +++ b/block/io.c | ||
56 | @@ -XXX,XX +XXX,XX @@ void bdrv_aio_cancel(BlockAIOCB *acb) | ||
57 | if (acb->aiocb_info->get_aio_context) { | ||
58 | aio_poll(acb->aiocb_info->get_aio_context(acb), true); | ||
59 | } else if (acb->bs) { | ||
60 | + /* qemu_aio_ref and qemu_aio_unref are not thread-safe, so | ||
61 | + * assert that we're not using an I/O thread. Thread-safe | ||
62 | + * code should use bdrv_aio_cancel_async exclusively. | ||
63 | + */ | ||
64 | + assert(bdrv_get_aio_context(acb->bs) == qemu_get_aio_context()); | ||
65 | aio_poll(bdrv_get_aio_context(acb->bs), true); | ||
66 | } else { | ||
67 | abort(); | ||
68 | diff --git a/block/iscsi.c b/block/iscsi.c | ||
69 | index XXXXXXX..XXXXXXX 100644 | ||
70 | --- a/block/iscsi.c | ||
71 | +++ b/block/iscsi.c | ||
72 | @@ -XXX,XX +XXX,XX @@ static void iscsi_retry_timer_expired(void *opaque) | ||
73 | struct IscsiTask *iTask = opaque; | ||
74 | iTask->complete = 1; | ||
75 | if (iTask->co) { | ||
76 | - qemu_coroutine_enter(iTask->co); | ||
77 | + aio_co_wake(iTask->co); | ||
78 | } | ||
79 | } | ||
80 | |||
81 | @@ -XXX,XX +XXX,XX @@ static void iscsi_nop_timed_event(void *opaque) | ||
82 | { | ||
83 | IscsiLun *iscsilun = opaque; | ||
84 | |||
85 | + aio_context_acquire(iscsilun->aio_context); | ||
86 | if (iscsi_get_nops_in_flight(iscsilun->iscsi) >= MAX_NOP_FAILURES) { | ||
87 | error_report("iSCSI: NOP timeout. Reconnecting..."); | ||
88 | iscsilun->request_timed_out = true; | ||
89 | } else if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) { | ||
90 | error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages."); | ||
91 | - return; | ||
92 | + goto out; | ||
93 | } | ||
94 | |||
95 | timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL); | ||
96 | iscsi_set_events(iscsilun); | ||
97 | + | ||
98 | +out: | ||
99 | + aio_context_release(iscsilun->aio_context); | ||
100 | } | ||
101 | |||
102 | static void iscsi_readcapacity_sync(IscsiLun *iscsilun, Error **errp) | ||
103 | diff --git a/block/null.c b/block/null.c | ||
104 | index XXXXXXX..XXXXXXX 100644 | ||
105 | --- a/block/null.c | ||
106 | +++ b/block/null.c | ||
107 | @@ -XXX,XX +XXX,XX @@ static void null_bh_cb(void *opaque) | ||
108 | static void null_timer_cb(void *opaque) | ||
109 | { | ||
110 | NullAIOCB *acb = opaque; | ||
111 | + AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
112 | + | ||
113 | + aio_context_acquire(ctx); | ||
114 | acb->common.cb(acb->common.opaque, 0); | ||
115 | + aio_context_release(ctx); | ||
116 | timer_deinit(&acb->timer); | ||
117 | qemu_aio_unref(acb); | ||
118 | } | ||
119 | diff --git a/block/qed.c b/block/qed.c | ||
120 | index XXXXXXX..XXXXXXX 100644 | ||
121 | --- a/block/qed.c | ||
122 | +++ b/block/qed.c | ||
123 | @@ -XXX,XX +XXX,XX @@ static void qed_need_check_timer_cb(void *opaque) | ||
124 | |||
125 | trace_qed_need_check_timer_cb(s); | ||
126 | |||
127 | + qed_acquire(s); | ||
128 | qed_plug_allocating_write_reqs(s); | ||
129 | |||
130 | /* Ensure writes are on disk before clearing flag */ | ||
131 | bdrv_aio_flush(s->bs->file->bs, qed_clear_need_check, s); | ||
132 | + qed_release(s); | ||
133 | +} | ||
134 | + | ||
135 | +void qed_acquire(BDRVQEDState *s) | ||
136 | +{ | ||
137 | + aio_context_acquire(bdrv_get_aio_context(s->bs)); | ||
138 | +} | ||
139 | + | ||
140 | +void qed_release(BDRVQEDState *s) | ||
141 | +{ | ||
142 | + aio_context_release(bdrv_get_aio_context(s->bs)); | ||
143 | } | ||
144 | |||
145 | static void qed_start_need_check_timer(BDRVQEDState *s) | ||
146 | diff --git a/block/throttle-groups.c b/block/throttle-groups.c | ||
147 | index XXXXXXX..XXXXXXX 100644 | ||
148 | --- a/block/throttle-groups.c | ||
149 | +++ b/block/throttle-groups.c | ||
150 | @@ -XXX,XX +XXX,XX @@ static void timer_cb(BlockBackend *blk, bool is_write) | ||
151 | qemu_mutex_unlock(&tg->lock); | ||
152 | |||
153 | /* Run the request that was waiting for this timer */ | ||
154 | + aio_context_acquire(blk_get_aio_context(blk)); | ||
155 | empty_queue = !qemu_co_enter_next(&blkp->throttled_reqs[is_write]); | ||
156 | + aio_context_release(blk_get_aio_context(blk)); | ||
157 | |||
158 | /* If the request queue was empty then we have to take care of | ||
159 | * scheduling the next one */ | ||
160 | diff --git a/util/aio-posix.c b/util/aio-posix.c | ||
161 | index XXXXXXX..XXXXXXX 100644 | ||
162 | --- a/util/aio-posix.c | ||
163 | +++ b/util/aio-posix.c | ||
164 | @@ -XXX,XX +XXX,XX @@ bool aio_dispatch(AioContext *ctx, bool dispatch_fds) | ||
165 | } | ||
166 | |||
167 | /* Run our timers */ | ||
168 | - aio_context_acquire(ctx); | ||
169 | progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
170 | - aio_context_release(ctx); | ||
171 | |||
172 | return progress; | ||
173 | } | ||
174 | diff --git a/util/aio-win32.c b/util/aio-win32.c | ||
175 | index XXXXXXX..XXXXXXX 100644 | ||
176 | --- a/util/aio-win32.c | ||
177 | +++ b/util/aio-win32.c | ||
178 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
179 | progress |= aio_dispatch_handlers(ctx, event); | ||
180 | } while (count > 0); | ||
181 | |||
182 | - aio_context_acquire(ctx); | ||
183 | progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
184 | - aio_context_release(ctx); | ||
185 | return progress; | ||
186 | } | ||
187 | |||
188 | diff --git a/util/qemu-coroutine-sleep.c b/util/qemu-coroutine-sleep.c | ||
189 | index XXXXXXX..XXXXXXX 100644 | ||
190 | --- a/util/qemu-coroutine-sleep.c | ||
191 | +++ b/util/qemu-coroutine-sleep.c | ||
192 | @@ -XXX,XX +XXX,XX @@ static void co_sleep_cb(void *opaque) | ||
193 | { | ||
194 | CoSleepCB *sleep_cb = opaque; | ||
195 | |||
196 | - qemu_coroutine_enter(sleep_cb->co); | ||
197 | + aio_co_wake(sleep_cb->co); | ||
198 | } | ||
199 | |||
200 | void coroutine_fn co_aio_sleep_ns(AioContext *ctx, QEMUClockType type, | ||
201 | -- | ||
202 | 2.9.3 | ||
203 | |||
204 | diff view generated by jsdifflib |
1 | From: Stefano Garzarella <sgarzare@redhat.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | In order to avoid migration issues, we enable DISCARD and | 3 | This covers both file descriptor callbacks and polling callbacks, |
4 | WRITE_ZEROES features only for machine type >= 4.0 | 4 | since they execute related code. |
5 | 5 | ||
6 | As discussed with Michael S. Tsirkin and Stefan Hajnoczi on the | ||
7 | list [1], DISCARD operation should not have security implications | ||
8 | (eg. page cache attacks), so we can enable it by default. | ||
9 | |||
10 | [1] https://lists.gnu.org/archive/html/qemu-devel/2019-02/msg00504.html | ||
11 | |||
12 | Suggested-by: Dr. David Alan Gilbert <dgilbert@redhat.com> | ||
13 | Reviewed-by: Michael S. Tsirkin <mst@redhat.com> | ||
14 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | 6 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> |
15 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | 7 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
16 | Acked-by: Pankaj Gupta <pagupta@redhat.com> | 8 | Reviewed-by: Fam Zheng <famz@redhat.com> |
17 | Message-id: 20190208134950.187665-4-sgarzare@redhat.com | 9 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> |
10 | Message-id: 20170213135235.12274-14-pbonzini@redhat.com | ||
18 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 11 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
19 | --- | 12 | --- |
20 | hw/block/virtio-blk.c | 4 ++++ | 13 | block/curl.c | 16 +++++++++++++--- |
21 | hw/core/machine.c | 2 ++ | 14 | block/iscsi.c | 4 ++++ |
22 | 2 files changed, 6 insertions(+) | 15 | block/linux-aio.c | 4 ++++ |
16 | block/nfs.c | 6 ++++++ | ||
17 | block/sheepdog.c | 29 +++++++++++++++-------------- | ||
18 | block/ssh.c | 29 +++++++++-------------------- | ||
19 | block/win32-aio.c | 10 ++++++---- | ||
20 | hw/block/virtio-blk.c | 5 ++++- | ||
21 | hw/scsi/virtio-scsi.c | 7 +++++++ | ||
22 | util/aio-posix.c | 7 ------- | ||
23 | util/aio-win32.c | 6 ------ | ||
24 | 11 files changed, 68 insertions(+), 55 deletions(-) | ||
23 | 25 | ||
26 | diff --git a/block/curl.c b/block/curl.c | ||
27 | index XXXXXXX..XXXXXXX 100644 | ||
28 | --- a/block/curl.c | ||
29 | +++ b/block/curl.c | ||
30 | @@ -XXX,XX +XXX,XX @@ static void curl_multi_check_completion(BDRVCURLState *s) | ||
31 | } | ||
32 | } | ||
33 | |||
34 | -static void curl_multi_do(void *arg) | ||
35 | +static void curl_multi_do_locked(CURLState *s) | ||
36 | { | ||
37 | - CURLState *s = (CURLState *)arg; | ||
38 | CURLSocket *socket, *next_socket; | ||
39 | int running; | ||
40 | int r; | ||
41 | @@ -XXX,XX +XXX,XX @@ static void curl_multi_do(void *arg) | ||
42 | } | ||
43 | } | ||
44 | |||
45 | +static void curl_multi_do(void *arg) | ||
46 | +{ | ||
47 | + CURLState *s = (CURLState *)arg; | ||
48 | + | ||
49 | + aio_context_acquire(s->s->aio_context); | ||
50 | + curl_multi_do_locked(s); | ||
51 | + aio_context_release(s->s->aio_context); | ||
52 | +} | ||
53 | + | ||
54 | static void curl_multi_read(void *arg) | ||
55 | { | ||
56 | CURLState *s = (CURLState *)arg; | ||
57 | |||
58 | - curl_multi_do(arg); | ||
59 | + aio_context_acquire(s->s->aio_context); | ||
60 | + curl_multi_do_locked(s); | ||
61 | curl_multi_check_completion(s->s); | ||
62 | + aio_context_release(s->s->aio_context); | ||
63 | } | ||
64 | |||
65 | static void curl_multi_timeout_do(void *arg) | ||
66 | diff --git a/block/iscsi.c b/block/iscsi.c | ||
67 | index XXXXXXX..XXXXXXX 100644 | ||
68 | --- a/block/iscsi.c | ||
69 | +++ b/block/iscsi.c | ||
70 | @@ -XXX,XX +XXX,XX @@ iscsi_process_read(void *arg) | ||
71 | IscsiLun *iscsilun = arg; | ||
72 | struct iscsi_context *iscsi = iscsilun->iscsi; | ||
73 | |||
74 | + aio_context_acquire(iscsilun->aio_context); | ||
75 | iscsi_service(iscsi, POLLIN); | ||
76 | iscsi_set_events(iscsilun); | ||
77 | + aio_context_release(iscsilun->aio_context); | ||
78 | } | ||
79 | |||
80 | static void | ||
81 | @@ -XXX,XX +XXX,XX @@ iscsi_process_write(void *arg) | ||
82 | IscsiLun *iscsilun = arg; | ||
83 | struct iscsi_context *iscsi = iscsilun->iscsi; | ||
84 | |||
85 | + aio_context_acquire(iscsilun->aio_context); | ||
86 | iscsi_service(iscsi, POLLOUT); | ||
87 | iscsi_set_events(iscsilun); | ||
88 | + aio_context_release(iscsilun->aio_context); | ||
89 | } | ||
90 | |||
91 | static int64_t sector_lun2qemu(int64_t sector, IscsiLun *iscsilun) | ||
92 | diff --git a/block/linux-aio.c b/block/linux-aio.c | ||
93 | index XXXXXXX..XXXXXXX 100644 | ||
94 | --- a/block/linux-aio.c | ||
95 | +++ b/block/linux-aio.c | ||
96 | @@ -XXX,XX +XXX,XX @@ static void qemu_laio_completion_cb(EventNotifier *e) | ||
97 | LinuxAioState *s = container_of(e, LinuxAioState, e); | ||
98 | |||
99 | if (event_notifier_test_and_clear(&s->e)) { | ||
100 | + aio_context_acquire(s->aio_context); | ||
101 | qemu_laio_process_completions_and_submit(s); | ||
102 | + aio_context_release(s->aio_context); | ||
103 | } | ||
104 | } | ||
105 | |||
106 | @@ -XXX,XX +XXX,XX @@ static bool qemu_laio_poll_cb(void *opaque) | ||
107 | return false; | ||
108 | } | ||
109 | |||
110 | + aio_context_acquire(s->aio_context); | ||
111 | qemu_laio_process_completions_and_submit(s); | ||
112 | + aio_context_release(s->aio_context); | ||
113 | return true; | ||
114 | } | ||
115 | |||
116 | diff --git a/block/nfs.c b/block/nfs.c | ||
117 | index XXXXXXX..XXXXXXX 100644 | ||
118 | --- a/block/nfs.c | ||
119 | +++ b/block/nfs.c | ||
120 | @@ -XXX,XX +XXX,XX @@ static void nfs_set_events(NFSClient *client) | ||
121 | static void nfs_process_read(void *arg) | ||
122 | { | ||
123 | NFSClient *client = arg; | ||
124 | + | ||
125 | + aio_context_acquire(client->aio_context); | ||
126 | nfs_service(client->context, POLLIN); | ||
127 | nfs_set_events(client); | ||
128 | + aio_context_release(client->aio_context); | ||
129 | } | ||
130 | |||
131 | static void nfs_process_write(void *arg) | ||
132 | { | ||
133 | NFSClient *client = arg; | ||
134 | + | ||
135 | + aio_context_acquire(client->aio_context); | ||
136 | nfs_service(client->context, POLLOUT); | ||
137 | nfs_set_events(client); | ||
138 | + aio_context_release(client->aio_context); | ||
139 | } | ||
140 | |||
141 | static void nfs_co_init_task(BlockDriverState *bs, NFSRPC *task) | ||
142 | diff --git a/block/sheepdog.c b/block/sheepdog.c | ||
143 | index XXXXXXX..XXXXXXX 100644 | ||
144 | --- a/block/sheepdog.c | ||
145 | +++ b/block/sheepdog.c | ||
146 | @@ -XXX,XX +XXX,XX @@ static coroutine_fn int send_co_req(int sockfd, SheepdogReq *hdr, void *data, | ||
147 | return ret; | ||
148 | } | ||
149 | |||
150 | -static void restart_co_req(void *opaque) | ||
151 | -{ | ||
152 | - Coroutine *co = opaque; | ||
153 | - | ||
154 | - qemu_coroutine_enter(co); | ||
155 | -} | ||
156 | - | ||
157 | typedef struct SheepdogReqCo { | ||
158 | int sockfd; | ||
159 | BlockDriverState *bs; | ||
160 | @@ -XXX,XX +XXX,XX @@ typedef struct SheepdogReqCo { | ||
161 | unsigned int *rlen; | ||
162 | int ret; | ||
163 | bool finished; | ||
164 | + Coroutine *co; | ||
165 | } SheepdogReqCo; | ||
166 | |||
167 | +static void restart_co_req(void *opaque) | ||
168 | +{ | ||
169 | + SheepdogReqCo *srco = opaque; | ||
170 | + | ||
171 | + aio_co_wake(srco->co); | ||
172 | +} | ||
173 | + | ||
174 | static coroutine_fn void do_co_req(void *opaque) | ||
175 | { | ||
176 | int ret; | ||
177 | - Coroutine *co; | ||
178 | SheepdogReqCo *srco = opaque; | ||
179 | int sockfd = srco->sockfd; | ||
180 | SheepdogReq *hdr = srco->hdr; | ||
181 | @@ -XXX,XX +XXX,XX @@ static coroutine_fn void do_co_req(void *opaque) | ||
182 | unsigned int *wlen = srco->wlen; | ||
183 | unsigned int *rlen = srco->rlen; | ||
184 | |||
185 | - co = qemu_coroutine_self(); | ||
186 | + srco->co = qemu_coroutine_self(); | ||
187 | aio_set_fd_handler(srco->aio_context, sockfd, false, | ||
188 | - NULL, restart_co_req, NULL, co); | ||
189 | + NULL, restart_co_req, NULL, srco); | ||
190 | |||
191 | ret = send_co_req(sockfd, hdr, data, wlen); | ||
192 | if (ret < 0) { | ||
193 | @@ -XXX,XX +XXX,XX @@ static coroutine_fn void do_co_req(void *opaque) | ||
194 | } | ||
195 | |||
196 | aio_set_fd_handler(srco->aio_context, sockfd, false, | ||
197 | - restart_co_req, NULL, NULL, co); | ||
198 | + restart_co_req, NULL, NULL, srco); | ||
199 | |||
200 | ret = qemu_co_recv(sockfd, hdr, sizeof(*hdr)); | ||
201 | if (ret != sizeof(*hdr)) { | ||
202 | @@ -XXX,XX +XXX,XX @@ out: | ||
203 | aio_set_fd_handler(srco->aio_context, sockfd, false, | ||
204 | NULL, NULL, NULL, NULL); | ||
205 | |||
206 | + srco->co = NULL; | ||
207 | srco->ret = ret; | ||
208 | srco->finished = true; | ||
209 | if (srco->bs) { | ||
210 | @@ -XXX,XX +XXX,XX @@ static void coroutine_fn aio_read_response(void *opaque) | ||
211 | * We've finished all requests which belong to the AIOCB, so | ||
212 | * we can switch back to sd_co_readv/writev now. | ||
213 | */ | ||
214 | - qemu_coroutine_enter(acb->coroutine); | ||
215 | + aio_co_wake(acb->coroutine); | ||
216 | } | ||
217 | |||
218 | return; | ||
219 | @@ -XXX,XX +XXX,XX @@ static void co_read_response(void *opaque) | ||
220 | s->co_recv = qemu_coroutine_create(aio_read_response, opaque); | ||
221 | } | ||
222 | |||
223 | - qemu_coroutine_enter(s->co_recv); | ||
224 | + aio_co_wake(s->co_recv); | ||
225 | } | ||
226 | |||
227 | static void co_write_request(void *opaque) | ||
228 | { | ||
229 | BDRVSheepdogState *s = opaque; | ||
230 | |||
231 | - qemu_coroutine_enter(s->co_send); | ||
232 | + aio_co_wake(s->co_send); | ||
233 | } | ||
234 | |||
235 | /* | ||
236 | diff --git a/block/ssh.c b/block/ssh.c | ||
237 | index XXXXXXX..XXXXXXX 100644 | ||
238 | --- a/block/ssh.c | ||
239 | +++ b/block/ssh.c | ||
240 | @@ -XXX,XX +XXX,XX @@ static void restart_coroutine(void *opaque) | ||
241 | |||
242 | DPRINTF("co=%p", co); | ||
243 | |||
244 | - qemu_coroutine_enter(co); | ||
245 | + aio_co_wake(co); | ||
246 | } | ||
247 | |||
248 | -static coroutine_fn void set_fd_handler(BDRVSSHState *s, BlockDriverState *bs) | ||
249 | +/* A non-blocking call returned EAGAIN, so yield, ensuring the | ||
250 | + * handlers are set up so that we'll be rescheduled when there is an | ||
251 | + * interesting event on the socket. | ||
252 | + */ | ||
253 | +static coroutine_fn void co_yield(BDRVSSHState *s, BlockDriverState *bs) | ||
254 | { | ||
255 | int r; | ||
256 | IOHandler *rd_handler = NULL, *wr_handler = NULL; | ||
257 | @@ -XXX,XX +XXX,XX @@ static coroutine_fn void set_fd_handler(BDRVSSHState *s, BlockDriverState *bs) | ||
258 | |||
259 | aio_set_fd_handler(bdrv_get_aio_context(bs), s->sock, | ||
260 | false, rd_handler, wr_handler, NULL, co); | ||
261 | -} | ||
262 | - | ||
263 | -static coroutine_fn void clear_fd_handler(BDRVSSHState *s, | ||
264 | - BlockDriverState *bs) | ||
265 | -{ | ||
266 | - DPRINTF("s->sock=%d", s->sock); | ||
267 | - aio_set_fd_handler(bdrv_get_aio_context(bs), s->sock, | ||
268 | - false, NULL, NULL, NULL, NULL); | ||
269 | -} | ||
270 | - | ||
271 | -/* A non-blocking call returned EAGAIN, so yield, ensuring the | ||
272 | - * handlers are set up so that we'll be rescheduled when there is an | ||
273 | - * interesting event on the socket. | ||
274 | - */ | ||
275 | -static coroutine_fn void co_yield(BDRVSSHState *s, BlockDriverState *bs) | ||
276 | -{ | ||
277 | - set_fd_handler(s, bs); | ||
278 | qemu_coroutine_yield(); | ||
279 | - clear_fd_handler(s, bs); | ||
280 | + DPRINTF("s->sock=%d - back", s->sock); | ||
281 | + aio_set_fd_handler(bdrv_get_aio_context(bs), s->sock, false, | ||
282 | + NULL, NULL, NULL, NULL); | ||
283 | } | ||
284 | |||
285 | /* SFTP has a function `libssh2_sftp_seek64' which seeks to a position | ||
286 | diff --git a/block/win32-aio.c b/block/win32-aio.c | ||
287 | index XXXXXXX..XXXXXXX 100644 | ||
288 | --- a/block/win32-aio.c | ||
289 | +++ b/block/win32-aio.c | ||
290 | @@ -XXX,XX +XXX,XX @@ struct QEMUWin32AIOState { | ||
291 | HANDLE hIOCP; | ||
292 | EventNotifier e; | ||
293 | int count; | ||
294 | - bool is_aio_context_attached; | ||
295 | + AioContext *aio_ctx; | ||
296 | }; | ||
297 | |||
298 | typedef struct QEMUWin32AIOCB { | ||
299 | @@ -XXX,XX +XXX,XX @@ static void win32_aio_process_completion(QEMUWin32AIOState *s, | ||
300 | } | ||
301 | |||
302 | |||
303 | + aio_context_acquire(s->aio_ctx); | ||
304 | waiocb->common.cb(waiocb->common.opaque, ret); | ||
305 | + aio_context_release(s->aio_ctx); | ||
306 | qemu_aio_unref(waiocb); | ||
307 | } | ||
308 | |||
309 | @@ -XXX,XX +XXX,XX @@ void win32_aio_detach_aio_context(QEMUWin32AIOState *aio, | ||
310 | AioContext *old_context) | ||
311 | { | ||
312 | aio_set_event_notifier(old_context, &aio->e, false, NULL, NULL); | ||
313 | - aio->is_aio_context_attached = false; | ||
314 | + aio->aio_ctx = NULL; | ||
315 | } | ||
316 | |||
317 | void win32_aio_attach_aio_context(QEMUWin32AIOState *aio, | ||
318 | AioContext *new_context) | ||
319 | { | ||
320 | - aio->is_aio_context_attached = true; | ||
321 | + aio->aio_ctx = new_context; | ||
322 | aio_set_event_notifier(new_context, &aio->e, false, | ||
323 | win32_aio_completion_cb, NULL); | ||
324 | } | ||
325 | @@ -XXX,XX +XXX,XX @@ out_free_state: | ||
326 | |||
327 | void win32_aio_cleanup(QEMUWin32AIOState *aio) | ||
328 | { | ||
329 | - assert(!aio->is_aio_context_attached); | ||
330 | + assert(!aio->aio_ctx); | ||
331 | CloseHandle(aio->hIOCP); | ||
332 | event_notifier_cleanup(&aio->e); | ||
333 | g_free(aio); | ||
24 | diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c | 334 | diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c |
25 | index XXXXXXX..XXXXXXX 100644 | 335 | index XXXXXXX..XXXXXXX 100644 |
26 | --- a/hw/block/virtio-blk.c | 336 | --- a/hw/block/virtio-blk.c |
27 | +++ b/hw/block/virtio-blk.c | 337 | +++ b/hw/block/virtio-blk.c |
28 | @@ -XXX,XX +XXX,XX @@ static Property virtio_blk_properties[] = { | 338 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_ioctl_complete(void *opaque, int status) |
29 | DEFINE_PROP_UINT16("queue-size", VirtIOBlock, conf.queue_size, 128), | 339 | { |
30 | DEFINE_PROP_LINK("iothread", VirtIOBlock, conf.iothread, TYPE_IOTHREAD, | 340 | VirtIOBlockIoctlReq *ioctl_req = opaque; |
31 | IOThread *), | 341 | VirtIOBlockReq *req = ioctl_req->req; |
32 | + DEFINE_PROP_BIT64("discard", VirtIOBlock, host_features, | 342 | - VirtIODevice *vdev = VIRTIO_DEVICE(req->dev); |
33 | + VIRTIO_BLK_F_DISCARD, true), | 343 | + VirtIOBlock *s = req->dev; |
34 | + DEFINE_PROP_BIT64("write-zeroes", VirtIOBlock, host_features, | 344 | + VirtIODevice *vdev = VIRTIO_DEVICE(s); |
35 | + VIRTIO_BLK_F_WRITE_ZEROES, true), | 345 | struct virtio_scsi_inhdr *scsi; |
36 | DEFINE_PROP_END_OF_LIST(), | 346 | struct sg_io_hdr *hdr; |
37 | }; | 347 | |
38 | 348 | @@ -XXX,XX +XXX,XX @@ bool virtio_blk_handle_vq(VirtIOBlock *s, VirtQueue *vq) | |
39 | diff --git a/hw/core/machine.c b/hw/core/machine.c | 349 | MultiReqBuffer mrb = {}; |
40 | index XXXXXXX..XXXXXXX 100644 | 350 | bool progress = false; |
41 | --- a/hw/core/machine.c | 351 | |
42 | +++ b/hw/core/machine.c | 352 | + aio_context_acquire(blk_get_aio_context(s->blk)); |
43 | @@ -XXX,XX +XXX,XX @@ GlobalProperty hw_compat_3_1[] = { | 353 | blk_io_plug(s->blk); |
44 | { "usb-kbd", "serial", "42" }, | 354 | |
45 | { "usb-mouse", "serial", "42" }, | 355 | do { |
46 | { "usb-kbd", "serial", "42" }, | 356 | @@ -XXX,XX +XXX,XX @@ bool virtio_blk_handle_vq(VirtIOBlock *s, VirtQueue *vq) |
47 | + { "virtio-blk-device", "discard", "false" }, | 357 | } |
48 | + { "virtio-blk-device", "write-zeroes", "false" }, | 358 | |
49 | }; | 359 | blk_io_unplug(s->blk); |
50 | const size_t hw_compat_3_1_len = G_N_ELEMENTS(hw_compat_3_1); | 360 | + aio_context_release(blk_get_aio_context(s->blk)); |
361 | return progress; | ||
362 | } | ||
363 | |||
364 | diff --git a/hw/scsi/virtio-scsi.c b/hw/scsi/virtio-scsi.c | ||
365 | index XXXXXXX..XXXXXXX 100644 | ||
366 | --- a/hw/scsi/virtio-scsi.c | ||
367 | +++ b/hw/scsi/virtio-scsi.c | ||
368 | @@ -XXX,XX +XXX,XX @@ bool virtio_scsi_handle_ctrl_vq(VirtIOSCSI *s, VirtQueue *vq) | ||
369 | VirtIOSCSIReq *req; | ||
370 | bool progress = false; | ||
371 | |||
372 | + virtio_scsi_acquire(s); | ||
373 | while ((req = virtio_scsi_pop_req(s, vq))) { | ||
374 | progress = true; | ||
375 | virtio_scsi_handle_ctrl_req(s, req); | ||
376 | } | ||
377 | + virtio_scsi_release(s); | ||
378 | return progress; | ||
379 | } | ||
380 | |||
381 | @@ -XXX,XX +XXX,XX @@ bool virtio_scsi_handle_cmd_vq(VirtIOSCSI *s, VirtQueue *vq) | ||
382 | |||
383 | QTAILQ_HEAD(, VirtIOSCSIReq) reqs = QTAILQ_HEAD_INITIALIZER(reqs); | ||
384 | |||
385 | + virtio_scsi_acquire(s); | ||
386 | do { | ||
387 | virtio_queue_set_notification(vq, 0); | ||
388 | |||
389 | @@ -XXX,XX +XXX,XX @@ bool virtio_scsi_handle_cmd_vq(VirtIOSCSI *s, VirtQueue *vq) | ||
390 | QTAILQ_FOREACH_SAFE(req, &reqs, next, next) { | ||
391 | virtio_scsi_handle_cmd_req_submit(s, req); | ||
392 | } | ||
393 | + virtio_scsi_release(s); | ||
394 | return progress; | ||
395 | } | ||
396 | |||
397 | @@ -XXX,XX +XXX,XX @@ out: | ||
398 | |||
399 | bool virtio_scsi_handle_event_vq(VirtIOSCSI *s, VirtQueue *vq) | ||
400 | { | ||
401 | + virtio_scsi_acquire(s); | ||
402 | if (s->events_dropped) { | ||
403 | virtio_scsi_push_event(s, NULL, VIRTIO_SCSI_T_NO_EVENT, 0); | ||
404 | + virtio_scsi_release(s); | ||
405 | return true; | ||
406 | } | ||
407 | + virtio_scsi_release(s); | ||
408 | return false; | ||
409 | } | ||
410 | |||
411 | diff --git a/util/aio-posix.c b/util/aio-posix.c | ||
412 | index XXXXXXX..XXXXXXX 100644 | ||
413 | --- a/util/aio-posix.c | ||
414 | +++ b/util/aio-posix.c | ||
415 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx) | ||
416 | (revents & (G_IO_IN | G_IO_HUP | G_IO_ERR)) && | ||
417 | aio_node_check(ctx, node->is_external) && | ||
418 | node->io_read) { | ||
419 | - aio_context_acquire(ctx); | ||
420 | node->io_read(node->opaque); | ||
421 | - aio_context_release(ctx); | ||
422 | |||
423 | /* aio_notify() does not count as progress */ | ||
424 | if (node->opaque != &ctx->notifier) { | ||
425 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx) | ||
426 | (revents & (G_IO_OUT | G_IO_ERR)) && | ||
427 | aio_node_check(ctx, node->is_external) && | ||
428 | node->io_write) { | ||
429 | - aio_context_acquire(ctx); | ||
430 | node->io_write(node->opaque); | ||
431 | - aio_context_release(ctx); | ||
432 | progress = true; | ||
433 | } | ||
434 | |||
435 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
436 | start = qemu_clock_get_ns(QEMU_CLOCK_REALTIME); | ||
437 | } | ||
438 | |||
439 | - aio_context_acquire(ctx); | ||
440 | progress = try_poll_mode(ctx, blocking); | ||
441 | - aio_context_release(ctx); | ||
442 | - | ||
443 | if (!progress) { | ||
444 | assert(npfd == 0); | ||
445 | |||
446 | diff --git a/util/aio-win32.c b/util/aio-win32.c | ||
447 | index XXXXXXX..XXXXXXX 100644 | ||
448 | --- a/util/aio-win32.c | ||
449 | +++ b/util/aio-win32.c | ||
450 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) | ||
451 | (revents || event_notifier_get_handle(node->e) == event) && | ||
452 | node->io_notify) { | ||
453 | node->pfd.revents = 0; | ||
454 | - aio_context_acquire(ctx); | ||
455 | node->io_notify(node->e); | ||
456 | - aio_context_release(ctx); | ||
457 | |||
458 | /* aio_notify() does not count as progress */ | ||
459 | if (node->e != &ctx->notifier) { | ||
460 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) | ||
461 | (node->io_read || node->io_write)) { | ||
462 | node->pfd.revents = 0; | ||
463 | if ((revents & G_IO_IN) && node->io_read) { | ||
464 | - aio_context_acquire(ctx); | ||
465 | node->io_read(node->opaque); | ||
466 | - aio_context_release(ctx); | ||
467 | progress = true; | ||
468 | } | ||
469 | if ((revents & G_IO_OUT) && node->io_write) { | ||
470 | - aio_context_acquire(ctx); | ||
471 | node->io_write(node->opaque); | ||
472 | - aio_context_release(ctx); | ||
473 | progress = true; | ||
474 | } | ||
51 | 475 | ||
52 | -- | 476 | -- |
53 | 2.20.1 | 477 | 2.9.3 |
54 | 478 | ||
55 | 479 | diff view generated by jsdifflib |
1 | From: Stefano Garzarella <sgarzare@redhat.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | Since configurable features for virtio-blk are growing, this patch | ||
4 | adds host_features field in the struct VirtIOBlock. (as in virtio-net) | ||
5 | In this way, we can avoid to add new fields for new properties and | ||
6 | we can directly set VIRTIO_BLK_F* flags in the host_features. | ||
7 | |||
8 | We update "config-wce" and "scsi" property definition to use the new | ||
9 | host_features field without change the behaviour. | ||
10 | |||
11 | Suggested-by: Michael S. Tsirkin <mst@redhat.com> | ||
12 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | 3 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> |
13 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | 4 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
14 | Acked-by: Pankaj Gupta <pagupta@redhat.com> | 5 | Reviewed-by: Fam Zheng <famz@redhat.com> |
15 | Message-id: 20190208134950.187665-3-sgarzare@redhat.com | 6 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> |
7 | Message-id: 20170213135235.12274-15-pbonzini@redhat.com | ||
16 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 8 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
17 | --- | 9 | --- |
18 | include/hw/virtio/virtio-blk.h | 3 +-- | 10 | block/archipelago.c | 3 +++ |
19 | hw/block/virtio-blk.c | 16 +++++++++------- | 11 | block/blkreplay.c | 2 +- |
20 | 2 files changed, 10 insertions(+), 9 deletions(-) | 12 | block/block-backend.c | 6 ++++++ |
13 | block/curl.c | 26 ++++++++++++++++++-------- | ||
14 | block/gluster.c | 9 +-------- | ||
15 | block/io.c | 6 +++++- | ||
16 | block/iscsi.c | 6 +++++- | ||
17 | block/linux-aio.c | 15 +++++++++------ | ||
18 | block/nfs.c | 3 ++- | ||
19 | block/null.c | 4 ++++ | ||
20 | block/qed.c | 3 +++ | ||
21 | block/rbd.c | 4 ++++ | ||
22 | dma-helpers.c | 2 ++ | ||
23 | hw/block/virtio-blk.c | 2 ++ | ||
24 | hw/scsi/scsi-bus.c | 2 ++ | ||
25 | util/async.c | 4 ++-- | ||
26 | util/thread-pool.c | 2 ++ | ||
27 | 17 files changed, 71 insertions(+), 28 deletions(-) | ||
21 | 28 | ||
22 | diff --git a/include/hw/virtio/virtio-blk.h b/include/hw/virtio/virtio-blk.h | 29 | diff --git a/block/archipelago.c b/block/archipelago.c |
23 | index XXXXXXX..XXXXXXX 100644 | 30 | index XXXXXXX..XXXXXXX 100644 |
24 | --- a/include/hw/virtio/virtio-blk.h | 31 | --- a/block/archipelago.c |
25 | +++ b/include/hw/virtio/virtio-blk.h | 32 | +++ b/block/archipelago.c |
26 | @@ -XXX,XX +XXX,XX @@ struct VirtIOBlkConf | 33 | @@ -XXX,XX +XXX,XX @@ static void qemu_archipelago_complete_aio(void *opaque) |
27 | BlockConf conf; | 34 | { |
28 | IOThread *iothread; | 35 | AIORequestData *reqdata = (AIORequestData *) opaque; |
29 | char *serial; | 36 | ArchipelagoAIOCB *aio_cb = (ArchipelagoAIOCB *) reqdata->aio_cb; |
30 | - uint32_t scsi; | 37 | + AioContext *ctx = bdrv_get_aio_context(aio_cb->common.bs); |
31 | - uint32_t config_wce; | 38 | |
32 | uint32_t request_merging; | 39 | + aio_context_acquire(ctx); |
33 | uint16_t num_queues; | 40 | aio_cb->common.cb(aio_cb->common.opaque, aio_cb->ret); |
34 | uint16_t queue_size; | 41 | + aio_context_release(ctx); |
35 | @@ -XXX,XX +XXX,XX @@ typedef struct VirtIOBlock { | 42 | aio_cb->status = 0; |
36 | bool dataplane_disabled; | 43 | |
37 | bool dataplane_started; | 44 | qemu_aio_unref(aio_cb); |
38 | struct VirtIOBlockDataPlane *dataplane; | 45 | diff --git a/block/blkreplay.c b/block/blkreplay.c |
39 | + uint64_t host_features; | 46 | index XXXXXXX..XXXXXXX 100755 |
40 | } VirtIOBlock; | 47 | --- a/block/blkreplay.c |
41 | 48 | +++ b/block/blkreplay.c | |
42 | typedef struct VirtIOBlockReq { | 49 | @@ -XXX,XX +XXX,XX @@ static int64_t blkreplay_getlength(BlockDriverState *bs) |
50 | static void blkreplay_bh_cb(void *opaque) | ||
51 | { | ||
52 | Request *req = opaque; | ||
53 | - qemu_coroutine_enter(req->co); | ||
54 | + aio_co_wake(req->co); | ||
55 | qemu_bh_delete(req->bh); | ||
56 | g_free(req); | ||
57 | } | ||
58 | diff --git a/block/block-backend.c b/block/block-backend.c | ||
59 | index XXXXXXX..XXXXXXX 100644 | ||
60 | --- a/block/block-backend.c | ||
61 | +++ b/block/block-backend.c | ||
62 | @@ -XXX,XX +XXX,XX @@ int blk_make_zero(BlockBackend *blk, BdrvRequestFlags flags) | ||
63 | static void error_callback_bh(void *opaque) | ||
64 | { | ||
65 | struct BlockBackendAIOCB *acb = opaque; | ||
66 | + AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
67 | |||
68 | bdrv_dec_in_flight(acb->common.bs); | ||
69 | + aio_context_acquire(ctx); | ||
70 | acb->common.cb(acb->common.opaque, acb->ret); | ||
71 | + aio_context_release(ctx); | ||
72 | qemu_aio_unref(acb); | ||
73 | } | ||
74 | |||
75 | @@ -XXX,XX +XXX,XX @@ static void blk_aio_complete(BlkAioEmAIOCB *acb) | ||
76 | static void blk_aio_complete_bh(void *opaque) | ||
77 | { | ||
78 | BlkAioEmAIOCB *acb = opaque; | ||
79 | + AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
80 | |||
81 | assert(acb->has_returned); | ||
82 | + aio_context_acquire(ctx); | ||
83 | blk_aio_complete(acb); | ||
84 | + aio_context_release(ctx); | ||
85 | } | ||
86 | |||
87 | static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int bytes, | ||
88 | diff --git a/block/curl.c b/block/curl.c | ||
89 | index XXXXXXX..XXXXXXX 100644 | ||
90 | --- a/block/curl.c | ||
91 | +++ b/block/curl.c | ||
92 | @@ -XXX,XX +XXX,XX @@ static void curl_readv_bh_cb(void *p) | ||
93 | { | ||
94 | CURLState *state; | ||
95 | int running; | ||
96 | + int ret = -EINPROGRESS; | ||
97 | |||
98 | CURLAIOCB *acb = p; | ||
99 | - BDRVCURLState *s = acb->common.bs->opaque; | ||
100 | + BlockDriverState *bs = acb->common.bs; | ||
101 | + BDRVCURLState *s = bs->opaque; | ||
102 | + AioContext *ctx = bdrv_get_aio_context(bs); | ||
103 | |||
104 | size_t start = acb->sector_num * BDRV_SECTOR_SIZE; | ||
105 | size_t end; | ||
106 | |||
107 | + aio_context_acquire(ctx); | ||
108 | + | ||
109 | // In case we have the requested data already (e.g. read-ahead), | ||
110 | // we can just call the callback and be done. | ||
111 | switch (curl_find_buf(s, start, acb->nb_sectors * BDRV_SECTOR_SIZE, acb)) { | ||
112 | @@ -XXX,XX +XXX,XX @@ static void curl_readv_bh_cb(void *p) | ||
113 | qemu_aio_unref(acb); | ||
114 | // fall through | ||
115 | case FIND_RET_WAIT: | ||
116 | - return; | ||
117 | + goto out; | ||
118 | default: | ||
119 | break; | ||
120 | } | ||
121 | @@ -XXX,XX +XXX,XX @@ static void curl_readv_bh_cb(void *p) | ||
122 | // No cache found, so let's start a new request | ||
123 | state = curl_init_state(acb->common.bs, s); | ||
124 | if (!state) { | ||
125 | - acb->common.cb(acb->common.opaque, -EIO); | ||
126 | - qemu_aio_unref(acb); | ||
127 | - return; | ||
128 | + ret = -EIO; | ||
129 | + goto out; | ||
130 | } | ||
131 | |||
132 | acb->start = 0; | ||
133 | @@ -XXX,XX +XXX,XX @@ static void curl_readv_bh_cb(void *p) | ||
134 | state->orig_buf = g_try_malloc(state->buf_len); | ||
135 | if (state->buf_len && state->orig_buf == NULL) { | ||
136 | curl_clean_state(state); | ||
137 | - acb->common.cb(acb->common.opaque, -ENOMEM); | ||
138 | - qemu_aio_unref(acb); | ||
139 | - return; | ||
140 | + ret = -ENOMEM; | ||
141 | + goto out; | ||
142 | } | ||
143 | state->acb[0] = acb; | ||
144 | |||
145 | @@ -XXX,XX +XXX,XX @@ static void curl_readv_bh_cb(void *p) | ||
146 | |||
147 | /* Tell curl it needs to kick things off */ | ||
148 | curl_multi_socket_action(s->multi, CURL_SOCKET_TIMEOUT, 0, &running); | ||
149 | + | ||
150 | +out: | ||
151 | + if (ret != -EINPROGRESS) { | ||
152 | + acb->common.cb(acb->common.opaque, ret); | ||
153 | + qemu_aio_unref(acb); | ||
154 | + } | ||
155 | + aio_context_release(ctx); | ||
156 | } | ||
157 | |||
158 | static BlockAIOCB *curl_aio_readv(BlockDriverState *bs, | ||
159 | diff --git a/block/gluster.c b/block/gluster.c | ||
160 | index XXXXXXX..XXXXXXX 100644 | ||
161 | --- a/block/gluster.c | ||
162 | +++ b/block/gluster.c | ||
163 | @@ -XXX,XX +XXX,XX @@ static struct glfs *qemu_gluster_init(BlockdevOptionsGluster *gconf, | ||
164 | return qemu_gluster_glfs_init(gconf, errp); | ||
165 | } | ||
166 | |||
167 | -static void qemu_gluster_complete_aio(void *opaque) | ||
168 | -{ | ||
169 | - GlusterAIOCB *acb = (GlusterAIOCB *)opaque; | ||
170 | - | ||
171 | - qemu_coroutine_enter(acb->coroutine); | ||
172 | -} | ||
173 | - | ||
174 | /* | ||
175 | * AIO callback routine called from GlusterFS thread. | ||
176 | */ | ||
177 | @@ -XXX,XX +XXX,XX @@ static void gluster_finish_aiocb(struct glfs_fd *fd, ssize_t ret, void *arg) | ||
178 | acb->ret = -EIO; /* Partial read/write - fail it */ | ||
179 | } | ||
180 | |||
181 | - aio_bh_schedule_oneshot(acb->aio_context, qemu_gluster_complete_aio, acb); | ||
182 | + aio_co_schedule(acb->aio_context, acb->coroutine); | ||
183 | } | ||
184 | |||
185 | static void qemu_gluster_parse_flags(int bdrv_flags, int *open_flags) | ||
186 | diff --git a/block/io.c b/block/io.c | ||
187 | index XXXXXXX..XXXXXXX 100644 | ||
188 | --- a/block/io.c | ||
189 | +++ b/block/io.c | ||
190 | @@ -XXX,XX +XXX,XX @@ static void bdrv_co_drain_bh_cb(void *opaque) | ||
191 | bdrv_dec_in_flight(bs); | ||
192 | bdrv_drained_begin(bs); | ||
193 | data->done = true; | ||
194 | - qemu_coroutine_enter(co); | ||
195 | + aio_co_wake(co); | ||
196 | } | ||
197 | |||
198 | static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs) | ||
199 | @@ -XXX,XX +XXX,XX @@ static void bdrv_co_complete(BlockAIOCBCoroutine *acb) | ||
200 | static void bdrv_co_em_bh(void *opaque) | ||
201 | { | ||
202 | BlockAIOCBCoroutine *acb = opaque; | ||
203 | + BlockDriverState *bs = acb->common.bs; | ||
204 | + AioContext *ctx = bdrv_get_aio_context(bs); | ||
205 | |||
206 | assert(!acb->need_bh); | ||
207 | + aio_context_acquire(ctx); | ||
208 | bdrv_co_complete(acb); | ||
209 | + aio_context_release(ctx); | ||
210 | } | ||
211 | |||
212 | static void bdrv_co_maybe_schedule_bh(BlockAIOCBCoroutine *acb) | ||
213 | diff --git a/block/iscsi.c b/block/iscsi.c | ||
214 | index XXXXXXX..XXXXXXX 100644 | ||
215 | --- a/block/iscsi.c | ||
216 | +++ b/block/iscsi.c | ||
217 | @@ -XXX,XX +XXX,XX @@ static void | ||
218 | iscsi_bh_cb(void *p) | ||
219 | { | ||
220 | IscsiAIOCB *acb = p; | ||
221 | + AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
222 | |||
223 | qemu_bh_delete(acb->bh); | ||
224 | |||
225 | g_free(acb->buf); | ||
226 | acb->buf = NULL; | ||
227 | |||
228 | + aio_context_acquire(ctx); | ||
229 | acb->common.cb(acb->common.opaque, acb->status); | ||
230 | + aio_context_release(ctx); | ||
231 | |||
232 | if (acb->task != NULL) { | ||
233 | scsi_free_scsi_task(acb->task); | ||
234 | @@ -XXX,XX +XXX,XX @@ iscsi_schedule_bh(IscsiAIOCB *acb) | ||
235 | static void iscsi_co_generic_bh_cb(void *opaque) | ||
236 | { | ||
237 | struct IscsiTask *iTask = opaque; | ||
238 | + | ||
239 | iTask->complete = 1; | ||
240 | - qemu_coroutine_enter(iTask->co); | ||
241 | + aio_co_wake(iTask->co); | ||
242 | } | ||
243 | |||
244 | static void iscsi_retry_timer_expired(void *opaque) | ||
245 | diff --git a/block/linux-aio.c b/block/linux-aio.c | ||
246 | index XXXXXXX..XXXXXXX 100644 | ||
247 | --- a/block/linux-aio.c | ||
248 | +++ b/block/linux-aio.c | ||
249 | @@ -XXX,XX +XXX,XX @@ struct LinuxAioState { | ||
250 | io_context_t ctx; | ||
251 | EventNotifier e; | ||
252 | |||
253 | - /* io queue for submit at batch */ | ||
254 | + /* io queue for submit at batch. Protected by AioContext lock. */ | ||
255 | LaioQueue io_q; | ||
256 | |||
257 | - /* I/O completion processing */ | ||
258 | + /* I/O completion processing. Only runs in I/O thread. */ | ||
259 | QEMUBH *completion_bh; | ||
260 | int event_idx; | ||
261 | int event_max; | ||
262 | @@ -XXX,XX +XXX,XX @@ static inline ssize_t io_event_ret(struct io_event *ev) | ||
263 | */ | ||
264 | static void qemu_laio_process_completion(struct qemu_laiocb *laiocb) | ||
265 | { | ||
266 | + LinuxAioState *s = laiocb->ctx; | ||
267 | int ret; | ||
268 | |||
269 | ret = laiocb->ret; | ||
270 | @@ -XXX,XX +XXX,XX @@ static void qemu_laio_process_completion(struct qemu_laiocb *laiocb) | ||
271 | } | ||
272 | |||
273 | laiocb->ret = ret; | ||
274 | + aio_context_acquire(s->aio_context); | ||
275 | if (laiocb->co) { | ||
276 | /* If the coroutine is already entered it must be in ioq_submit() and | ||
277 | * will notice laio->ret has been filled in when it eventually runs | ||
278 | @@ -XXX,XX +XXX,XX @@ static void qemu_laio_process_completion(struct qemu_laiocb *laiocb) | ||
279 | laiocb->common.cb(laiocb->common.opaque, ret); | ||
280 | qemu_aio_unref(laiocb); | ||
281 | } | ||
282 | + aio_context_release(s->aio_context); | ||
283 | } | ||
284 | |||
285 | /** | ||
286 | @@ -XXX,XX +XXX,XX @@ static void qemu_laio_process_completions(LinuxAioState *s) | ||
287 | static void qemu_laio_process_completions_and_submit(LinuxAioState *s) | ||
288 | { | ||
289 | qemu_laio_process_completions(s); | ||
290 | + | ||
291 | + aio_context_acquire(s->aio_context); | ||
292 | if (!s->io_q.plugged && !QSIMPLEQ_EMPTY(&s->io_q.pending)) { | ||
293 | ioq_submit(s); | ||
294 | } | ||
295 | + aio_context_release(s->aio_context); | ||
296 | } | ||
297 | |||
298 | static void qemu_laio_completion_bh(void *opaque) | ||
299 | @@ -XXX,XX +XXX,XX @@ static void qemu_laio_completion_cb(EventNotifier *e) | ||
300 | LinuxAioState *s = container_of(e, LinuxAioState, e); | ||
301 | |||
302 | if (event_notifier_test_and_clear(&s->e)) { | ||
303 | - aio_context_acquire(s->aio_context); | ||
304 | qemu_laio_process_completions_and_submit(s); | ||
305 | - aio_context_release(s->aio_context); | ||
306 | } | ||
307 | } | ||
308 | |||
309 | @@ -XXX,XX +XXX,XX @@ static bool qemu_laio_poll_cb(void *opaque) | ||
310 | return false; | ||
311 | } | ||
312 | |||
313 | - aio_context_acquire(s->aio_context); | ||
314 | qemu_laio_process_completions_and_submit(s); | ||
315 | - aio_context_release(s->aio_context); | ||
316 | return true; | ||
317 | } | ||
318 | |||
319 | @@ -XXX,XX +XXX,XX @@ void laio_detach_aio_context(LinuxAioState *s, AioContext *old_context) | ||
320 | { | ||
321 | aio_set_event_notifier(old_context, &s->e, false, NULL, NULL); | ||
322 | qemu_bh_delete(s->completion_bh); | ||
323 | + s->aio_context = NULL; | ||
324 | } | ||
325 | |||
326 | void laio_attach_aio_context(LinuxAioState *s, AioContext *new_context) | ||
327 | diff --git a/block/nfs.c b/block/nfs.c | ||
328 | index XXXXXXX..XXXXXXX 100644 | ||
329 | --- a/block/nfs.c | ||
330 | +++ b/block/nfs.c | ||
331 | @@ -XXX,XX +XXX,XX @@ static void nfs_co_init_task(BlockDriverState *bs, NFSRPC *task) | ||
332 | static void nfs_co_generic_bh_cb(void *opaque) | ||
333 | { | ||
334 | NFSRPC *task = opaque; | ||
335 | + | ||
336 | task->complete = 1; | ||
337 | - qemu_coroutine_enter(task->co); | ||
338 | + aio_co_wake(task->co); | ||
339 | } | ||
340 | |||
341 | static void | ||
342 | diff --git a/block/null.c b/block/null.c | ||
343 | index XXXXXXX..XXXXXXX 100644 | ||
344 | --- a/block/null.c | ||
345 | +++ b/block/null.c | ||
346 | @@ -XXX,XX +XXX,XX @@ static const AIOCBInfo null_aiocb_info = { | ||
347 | static void null_bh_cb(void *opaque) | ||
348 | { | ||
349 | NullAIOCB *acb = opaque; | ||
350 | + AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
351 | + | ||
352 | + aio_context_acquire(ctx); | ||
353 | acb->common.cb(acb->common.opaque, 0); | ||
354 | + aio_context_release(ctx); | ||
355 | qemu_aio_unref(acb); | ||
356 | } | ||
357 | |||
358 | diff --git a/block/qed.c b/block/qed.c | ||
359 | index XXXXXXX..XXXXXXX 100644 | ||
360 | --- a/block/qed.c | ||
361 | +++ b/block/qed.c | ||
362 | @@ -XXX,XX +XXX,XX @@ static void qed_update_l2_table(BDRVQEDState *s, QEDTable *table, int index, | ||
363 | static void qed_aio_complete_bh(void *opaque) | ||
364 | { | ||
365 | QEDAIOCB *acb = opaque; | ||
366 | + BDRVQEDState *s = acb_to_s(acb); | ||
367 | BlockCompletionFunc *cb = acb->common.cb; | ||
368 | void *user_opaque = acb->common.opaque; | ||
369 | int ret = acb->bh_ret; | ||
370 | @@ -XXX,XX +XXX,XX @@ static void qed_aio_complete_bh(void *opaque) | ||
371 | qemu_aio_unref(acb); | ||
372 | |||
373 | /* Invoke callback */ | ||
374 | + qed_acquire(s); | ||
375 | cb(user_opaque, ret); | ||
376 | + qed_release(s); | ||
377 | } | ||
378 | |||
379 | static void qed_aio_complete(QEDAIOCB *acb, int ret) | ||
380 | diff --git a/block/rbd.c b/block/rbd.c | ||
381 | index XXXXXXX..XXXXXXX 100644 | ||
382 | --- a/block/rbd.c | ||
383 | +++ b/block/rbd.c | ||
384 | @@ -XXX,XX +XXX,XX @@ shutdown: | ||
385 | static void qemu_rbd_complete_aio(RADOSCB *rcb) | ||
386 | { | ||
387 | RBDAIOCB *acb = rcb->acb; | ||
388 | + AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
389 | int64_t r; | ||
390 | |||
391 | r = rcb->ret; | ||
392 | @@ -XXX,XX +XXX,XX @@ static void qemu_rbd_complete_aio(RADOSCB *rcb) | ||
393 | qemu_iovec_from_buf(acb->qiov, 0, acb->bounce, acb->qiov->size); | ||
394 | } | ||
395 | qemu_vfree(acb->bounce); | ||
396 | + | ||
397 | + aio_context_acquire(ctx); | ||
398 | acb->common.cb(acb->common.opaque, (acb->ret > 0 ? 0 : acb->ret)); | ||
399 | + aio_context_release(ctx); | ||
400 | |||
401 | qemu_aio_unref(acb); | ||
402 | } | ||
403 | diff --git a/dma-helpers.c b/dma-helpers.c | ||
404 | index XXXXXXX..XXXXXXX 100644 | ||
405 | --- a/dma-helpers.c | ||
406 | +++ b/dma-helpers.c | ||
407 | @@ -XXX,XX +XXX,XX @@ static void dma_blk_cb(void *opaque, int ret) | ||
408 | QEMU_ALIGN_DOWN(dbs->iov.size, dbs->align)); | ||
409 | } | ||
410 | |||
411 | + aio_context_acquire(dbs->ctx); | ||
412 | dbs->acb = dbs->io_func(dbs->offset, &dbs->iov, | ||
413 | dma_blk_cb, dbs, dbs->io_func_opaque); | ||
414 | + aio_context_release(dbs->ctx); | ||
415 | assert(dbs->acb); | ||
416 | } | ||
417 | |||
43 | diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c | 418 | diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c |
44 | index XXXXXXX..XXXXXXX 100644 | 419 | index XXXXXXX..XXXXXXX 100644 |
45 | --- a/hw/block/virtio-blk.c | 420 | --- a/hw/block/virtio-blk.c |
46 | +++ b/hw/block/virtio-blk.c | 421 | +++ b/hw/block/virtio-blk.c |
47 | @@ -XXX,XX +XXX,XX @@ static int virtio_blk_handle_scsi_req(VirtIOBlockReq *req) | 422 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_dma_restart_bh(void *opaque) |
48 | */ | 423 | |
49 | scsi = (void *)elem->in_sg[elem->in_num - 2].iov_base; | 424 | s->rq = NULL; |
50 | 425 | ||
51 | - if (!blk->conf.scsi) { | 426 | + aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); |
52 | + if (!virtio_has_feature(blk->host_features, VIRTIO_BLK_F_SCSI)) { | 427 | while (req) { |
53 | status = VIRTIO_BLK_S_UNSUPP; | 428 | VirtIOBlockReq *next = req->next; |
54 | goto fail; | 429 | if (virtio_blk_handle_request(req, &mrb)) { |
55 | } | 430 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_dma_restart_bh(void *opaque) |
56 | @@ -XXX,XX +XXX,XX @@ static uint64_t virtio_blk_get_features(VirtIODevice *vdev, uint64_t features, | 431 | if (mrb.num_reqs) { |
57 | { | 432 | virtio_blk_submit_multireq(s->blk, &mrb); |
58 | VirtIOBlock *s = VIRTIO_BLK(vdev); | 433 | } |
59 | 434 | + aio_context_release(blk_get_aio_context(s->conf.conf.blk)); | |
60 | + /* Firstly sync all virtio-blk possible supported features */ | 435 | } |
61 | + features |= s->host_features; | 436 | |
62 | + | 437 | static void virtio_blk_dma_restart_cb(void *opaque, int running, |
63 | virtio_add_feature(&features, VIRTIO_BLK_F_SEG_MAX); | 438 | diff --git a/hw/scsi/scsi-bus.c b/hw/scsi/scsi-bus.c |
64 | virtio_add_feature(&features, VIRTIO_BLK_F_GEOMETRY); | 439 | index XXXXXXX..XXXXXXX 100644 |
65 | virtio_add_feature(&features, VIRTIO_BLK_F_TOPOLOGY); | 440 | --- a/hw/scsi/scsi-bus.c |
66 | virtio_add_feature(&features, VIRTIO_BLK_F_BLK_SIZE); | 441 | +++ b/hw/scsi/scsi-bus.c |
67 | if (virtio_has_feature(features, VIRTIO_F_VERSION_1)) { | 442 | @@ -XXX,XX +XXX,XX @@ static void scsi_dma_restart_bh(void *opaque) |
68 | - if (s->conf.scsi) { | 443 | qemu_bh_delete(s->bh); |
69 | + if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_SCSI)) { | 444 | s->bh = NULL; |
70 | error_setg(errp, "Please set scsi=off for virtio-blk devices in order to use virtio 1.0"); | 445 | |
71 | return 0; | 446 | + aio_context_acquire(blk_get_aio_context(s->conf.blk)); |
447 | QTAILQ_FOREACH_SAFE(req, &s->requests, next, next) { | ||
448 | scsi_req_ref(req); | ||
449 | if (req->retry) { | ||
450 | @@ -XXX,XX +XXX,XX @@ static void scsi_dma_restart_bh(void *opaque) | ||
72 | } | 451 | } |
73 | @@ -XXX,XX +XXX,XX @@ static uint64_t virtio_blk_get_features(VirtIODevice *vdev, uint64_t features, | 452 | scsi_req_unref(req); |
74 | virtio_add_feature(&features, VIRTIO_BLK_F_SCSI); | 453 | } |
75 | } | 454 | + aio_context_release(blk_get_aio_context(s->conf.blk)); |
76 | 455 | } | |
77 | - if (s->conf.config_wce) { | 456 | |
78 | - virtio_add_feature(&features, VIRTIO_BLK_F_CONFIG_WCE); | 457 | void scsi_req_retry(SCSIRequest *req) |
79 | - } | 458 | diff --git a/util/async.c b/util/async.c |
80 | if (blk_enable_write_cache(s->blk)) { | 459 | index XXXXXXX..XXXXXXX 100644 |
81 | virtio_add_feature(&features, VIRTIO_BLK_F_WCE); | 460 | --- a/util/async.c |
82 | } | 461 | +++ b/util/async.c |
83 | @@ -XXX,XX +XXX,XX @@ static Property virtio_blk_properties[] = { | 462 | @@ -XXX,XX +XXX,XX @@ int aio_bh_poll(AioContext *ctx) |
84 | DEFINE_BLOCK_ERROR_PROPERTIES(VirtIOBlock, conf.conf), | 463 | ret = 1; |
85 | DEFINE_BLOCK_CHS_PROPERTIES(VirtIOBlock, conf.conf), | 464 | } |
86 | DEFINE_PROP_STRING("serial", VirtIOBlock, conf.serial), | 465 | bh->idle = 0; |
87 | - DEFINE_PROP_BIT("config-wce", VirtIOBlock, conf.config_wce, 0, true), | 466 | - aio_context_acquire(ctx); |
88 | + DEFINE_PROP_BIT64("config-wce", VirtIOBlock, host_features, | 467 | aio_bh_call(bh); |
89 | + VIRTIO_BLK_F_CONFIG_WCE, true), | 468 | - aio_context_release(ctx); |
90 | #ifdef __linux__ | 469 | } |
91 | - DEFINE_PROP_BIT("scsi", VirtIOBlock, conf.scsi, 0, false), | 470 | if (bh->deleted) { |
92 | + DEFINE_PROP_BIT64("scsi", VirtIOBlock, host_features, | 471 | deleted = true; |
93 | + VIRTIO_BLK_F_SCSI, false), | 472 | @@ -XXX,XX +XXX,XX @@ static void co_schedule_bh_cb(void *opaque) |
94 | #endif | 473 | Coroutine *co = QSLIST_FIRST(&straight); |
95 | DEFINE_PROP_BIT("request-merging", VirtIOBlock, conf.request_merging, 0, | 474 | QSLIST_REMOVE_HEAD(&straight, co_scheduled_next); |
96 | true), | 475 | trace_aio_co_schedule_bh_cb(ctx, co); |
476 | + aio_context_acquire(ctx); | ||
477 | qemu_coroutine_enter(co); | ||
478 | + aio_context_release(ctx); | ||
479 | } | ||
480 | } | ||
481 | |||
482 | diff --git a/util/thread-pool.c b/util/thread-pool.c | ||
483 | index XXXXXXX..XXXXXXX 100644 | ||
484 | --- a/util/thread-pool.c | ||
485 | +++ b/util/thread-pool.c | ||
486 | @@ -XXX,XX +XXX,XX @@ static void thread_pool_completion_bh(void *opaque) | ||
487 | ThreadPool *pool = opaque; | ||
488 | ThreadPoolElement *elem, *next; | ||
489 | |||
490 | + aio_context_acquire(pool->ctx); | ||
491 | restart: | ||
492 | QLIST_FOREACH_SAFE(elem, &pool->head, all, next) { | ||
493 | if (elem->state != THREAD_DONE) { | ||
494 | @@ -XXX,XX +XXX,XX @@ restart: | ||
495 | qemu_aio_unref(elem); | ||
496 | } | ||
497 | } | ||
498 | + aio_context_release(pool->ctx); | ||
499 | } | ||
500 | |||
501 | static void thread_pool_cancel(BlockAIOCB *acb) | ||
97 | -- | 502 | -- |
98 | 2.20.1 | 503 | 2.9.3 |
99 | 504 | ||
100 | 505 | diff view generated by jsdifflib |
1 | From: Stefano Garzarella <sgarzare@redhat.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | In several part we still using req->dev or VIRTIO_DEVICE(req->dev) | 3 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> |
4 | when we have already defined s and vdev pointers: | 4 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
5 | VirtIOBlock *s = req->dev; | 5 | Reviewed-by: Fam Zheng <famz@redhat.com> |
6 | VirtIODevice *vdev = VIRTIO_DEVICE(s); | 6 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> |
7 | 7 | Message-id: 20170213135235.12274-16-pbonzini@redhat.com | |
8 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | ||
9 | Reviewed-by: Liam Merwick <liam.merwick@oracle.com> | ||
10 | Message-id: 20190208142347.214815-1-sgarzare@redhat.com | ||
11 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 8 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
12 | --- | 9 | --- |
13 | hw/block/virtio-blk.c | 22 +++++++++------------- | 10 | block/archipelago.c | 3 --- |
14 | 1 file changed, 9 insertions(+), 13 deletions(-) | 11 | block/block-backend.c | 7 ------- |
12 | block/curl.c | 2 +- | ||
13 | block/io.c | 6 +----- | ||
14 | block/iscsi.c | 3 --- | ||
15 | block/linux-aio.c | 5 +---- | ||
16 | block/mirror.c | 12 +++++++++--- | ||
17 | block/null.c | 8 -------- | ||
18 | block/qed-cluster.c | 2 ++ | ||
19 | block/qed-table.c | 12 ++++++++++-- | ||
20 | block/qed.c | 4 ++-- | ||
21 | block/rbd.c | 4 ---- | ||
22 | block/win32-aio.c | 3 --- | ||
23 | hw/block/virtio-blk.c | 12 +++++++++++- | ||
24 | hw/scsi/scsi-disk.c | 15 +++++++++++++++ | ||
25 | hw/scsi/scsi-generic.c | 20 +++++++++++++++++--- | ||
26 | util/thread-pool.c | 4 +++- | ||
27 | 17 files changed, 72 insertions(+), 50 deletions(-) | ||
15 | 28 | ||
29 | diff --git a/block/archipelago.c b/block/archipelago.c | ||
30 | index XXXXXXX..XXXXXXX 100644 | ||
31 | --- a/block/archipelago.c | ||
32 | +++ b/block/archipelago.c | ||
33 | @@ -XXX,XX +XXX,XX @@ static void qemu_archipelago_complete_aio(void *opaque) | ||
34 | { | ||
35 | AIORequestData *reqdata = (AIORequestData *) opaque; | ||
36 | ArchipelagoAIOCB *aio_cb = (ArchipelagoAIOCB *) reqdata->aio_cb; | ||
37 | - AioContext *ctx = bdrv_get_aio_context(aio_cb->common.bs); | ||
38 | |||
39 | - aio_context_acquire(ctx); | ||
40 | aio_cb->common.cb(aio_cb->common.opaque, aio_cb->ret); | ||
41 | - aio_context_release(ctx); | ||
42 | aio_cb->status = 0; | ||
43 | |||
44 | qemu_aio_unref(aio_cb); | ||
45 | diff --git a/block/block-backend.c b/block/block-backend.c | ||
46 | index XXXXXXX..XXXXXXX 100644 | ||
47 | --- a/block/block-backend.c | ||
48 | +++ b/block/block-backend.c | ||
49 | @@ -XXX,XX +XXX,XX @@ int blk_make_zero(BlockBackend *blk, BdrvRequestFlags flags) | ||
50 | static void error_callback_bh(void *opaque) | ||
51 | { | ||
52 | struct BlockBackendAIOCB *acb = opaque; | ||
53 | - AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
54 | |||
55 | bdrv_dec_in_flight(acb->common.bs); | ||
56 | - aio_context_acquire(ctx); | ||
57 | acb->common.cb(acb->common.opaque, acb->ret); | ||
58 | - aio_context_release(ctx); | ||
59 | qemu_aio_unref(acb); | ||
60 | } | ||
61 | |||
62 | @@ -XXX,XX +XXX,XX @@ static void blk_aio_complete(BlkAioEmAIOCB *acb) | ||
63 | static void blk_aio_complete_bh(void *opaque) | ||
64 | { | ||
65 | BlkAioEmAIOCB *acb = opaque; | ||
66 | - AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
67 | - | ||
68 | assert(acb->has_returned); | ||
69 | - aio_context_acquire(ctx); | ||
70 | blk_aio_complete(acb); | ||
71 | - aio_context_release(ctx); | ||
72 | } | ||
73 | |||
74 | static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int bytes, | ||
75 | diff --git a/block/curl.c b/block/curl.c | ||
76 | index XXXXXXX..XXXXXXX 100644 | ||
77 | --- a/block/curl.c | ||
78 | +++ b/block/curl.c | ||
79 | @@ -XXX,XX +XXX,XX @@ static void curl_readv_bh_cb(void *p) | ||
80 | curl_multi_socket_action(s->multi, CURL_SOCKET_TIMEOUT, 0, &running); | ||
81 | |||
82 | out: | ||
83 | + aio_context_release(ctx); | ||
84 | if (ret != -EINPROGRESS) { | ||
85 | acb->common.cb(acb->common.opaque, ret); | ||
86 | qemu_aio_unref(acb); | ||
87 | } | ||
88 | - aio_context_release(ctx); | ||
89 | } | ||
90 | |||
91 | static BlockAIOCB *curl_aio_readv(BlockDriverState *bs, | ||
92 | diff --git a/block/io.c b/block/io.c | ||
93 | index XXXXXXX..XXXXXXX 100644 | ||
94 | --- a/block/io.c | ||
95 | +++ b/block/io.c | ||
96 | @@ -XXX,XX +XXX,XX @@ static void bdrv_co_io_em_complete(void *opaque, int ret) | ||
97 | CoroutineIOCompletion *co = opaque; | ||
98 | |||
99 | co->ret = ret; | ||
100 | - qemu_coroutine_enter(co->coroutine); | ||
101 | + aio_co_wake(co->coroutine); | ||
102 | } | ||
103 | |||
104 | static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs, | ||
105 | @@ -XXX,XX +XXX,XX @@ static void bdrv_co_complete(BlockAIOCBCoroutine *acb) | ||
106 | static void bdrv_co_em_bh(void *opaque) | ||
107 | { | ||
108 | BlockAIOCBCoroutine *acb = opaque; | ||
109 | - BlockDriverState *bs = acb->common.bs; | ||
110 | - AioContext *ctx = bdrv_get_aio_context(bs); | ||
111 | |||
112 | assert(!acb->need_bh); | ||
113 | - aio_context_acquire(ctx); | ||
114 | bdrv_co_complete(acb); | ||
115 | - aio_context_release(ctx); | ||
116 | } | ||
117 | |||
118 | static void bdrv_co_maybe_schedule_bh(BlockAIOCBCoroutine *acb) | ||
119 | diff --git a/block/iscsi.c b/block/iscsi.c | ||
120 | index XXXXXXX..XXXXXXX 100644 | ||
121 | --- a/block/iscsi.c | ||
122 | +++ b/block/iscsi.c | ||
123 | @@ -XXX,XX +XXX,XX @@ static void | ||
124 | iscsi_bh_cb(void *p) | ||
125 | { | ||
126 | IscsiAIOCB *acb = p; | ||
127 | - AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
128 | |||
129 | qemu_bh_delete(acb->bh); | ||
130 | |||
131 | g_free(acb->buf); | ||
132 | acb->buf = NULL; | ||
133 | |||
134 | - aio_context_acquire(ctx); | ||
135 | acb->common.cb(acb->common.opaque, acb->status); | ||
136 | - aio_context_release(ctx); | ||
137 | |||
138 | if (acb->task != NULL) { | ||
139 | scsi_free_scsi_task(acb->task); | ||
140 | diff --git a/block/linux-aio.c b/block/linux-aio.c | ||
141 | index XXXXXXX..XXXXXXX 100644 | ||
142 | --- a/block/linux-aio.c | ||
143 | +++ b/block/linux-aio.c | ||
144 | @@ -XXX,XX +XXX,XX @@ static inline ssize_t io_event_ret(struct io_event *ev) | ||
145 | */ | ||
146 | static void qemu_laio_process_completion(struct qemu_laiocb *laiocb) | ||
147 | { | ||
148 | - LinuxAioState *s = laiocb->ctx; | ||
149 | int ret; | ||
150 | |||
151 | ret = laiocb->ret; | ||
152 | @@ -XXX,XX +XXX,XX @@ static void qemu_laio_process_completion(struct qemu_laiocb *laiocb) | ||
153 | } | ||
154 | |||
155 | laiocb->ret = ret; | ||
156 | - aio_context_acquire(s->aio_context); | ||
157 | if (laiocb->co) { | ||
158 | /* If the coroutine is already entered it must be in ioq_submit() and | ||
159 | * will notice laio->ret has been filled in when it eventually runs | ||
160 | @@ -XXX,XX +XXX,XX @@ static void qemu_laio_process_completion(struct qemu_laiocb *laiocb) | ||
161 | * that! | ||
162 | */ | ||
163 | if (!qemu_coroutine_entered(laiocb->co)) { | ||
164 | - qemu_coroutine_enter(laiocb->co); | ||
165 | + aio_co_wake(laiocb->co); | ||
166 | } | ||
167 | } else { | ||
168 | laiocb->common.cb(laiocb->common.opaque, ret); | ||
169 | qemu_aio_unref(laiocb); | ||
170 | } | ||
171 | - aio_context_release(s->aio_context); | ||
172 | } | ||
173 | |||
174 | /** | ||
175 | diff --git a/block/mirror.c b/block/mirror.c | ||
176 | index XXXXXXX..XXXXXXX 100644 | ||
177 | --- a/block/mirror.c | ||
178 | +++ b/block/mirror.c | ||
179 | @@ -XXX,XX +XXX,XX @@ static void mirror_write_complete(void *opaque, int ret) | ||
180 | { | ||
181 | MirrorOp *op = opaque; | ||
182 | MirrorBlockJob *s = op->s; | ||
183 | + | ||
184 | + aio_context_acquire(blk_get_aio_context(s->common.blk)); | ||
185 | if (ret < 0) { | ||
186 | BlockErrorAction action; | ||
187 | |||
188 | @@ -XXX,XX +XXX,XX @@ static void mirror_write_complete(void *opaque, int ret) | ||
189 | } | ||
190 | } | ||
191 | mirror_iteration_done(op, ret); | ||
192 | + aio_context_release(blk_get_aio_context(s->common.blk)); | ||
193 | } | ||
194 | |||
195 | static void mirror_read_complete(void *opaque, int ret) | ||
196 | { | ||
197 | MirrorOp *op = opaque; | ||
198 | MirrorBlockJob *s = op->s; | ||
199 | + | ||
200 | + aio_context_acquire(blk_get_aio_context(s->common.blk)); | ||
201 | if (ret < 0) { | ||
202 | BlockErrorAction action; | ||
203 | |||
204 | @@ -XXX,XX +XXX,XX @@ static void mirror_read_complete(void *opaque, int ret) | ||
205 | } | ||
206 | |||
207 | mirror_iteration_done(op, ret); | ||
208 | - return; | ||
209 | + } else { | ||
210 | + blk_aio_pwritev(s->target, op->sector_num * BDRV_SECTOR_SIZE, &op->qiov, | ||
211 | + 0, mirror_write_complete, op); | ||
212 | } | ||
213 | - blk_aio_pwritev(s->target, op->sector_num * BDRV_SECTOR_SIZE, &op->qiov, | ||
214 | - 0, mirror_write_complete, op); | ||
215 | + aio_context_release(blk_get_aio_context(s->common.blk)); | ||
216 | } | ||
217 | |||
218 | static inline void mirror_clip_sectors(MirrorBlockJob *s, | ||
219 | diff --git a/block/null.c b/block/null.c | ||
220 | index XXXXXXX..XXXXXXX 100644 | ||
221 | --- a/block/null.c | ||
222 | +++ b/block/null.c | ||
223 | @@ -XXX,XX +XXX,XX @@ static const AIOCBInfo null_aiocb_info = { | ||
224 | static void null_bh_cb(void *opaque) | ||
225 | { | ||
226 | NullAIOCB *acb = opaque; | ||
227 | - AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
228 | - | ||
229 | - aio_context_acquire(ctx); | ||
230 | acb->common.cb(acb->common.opaque, 0); | ||
231 | - aio_context_release(ctx); | ||
232 | qemu_aio_unref(acb); | ||
233 | } | ||
234 | |||
235 | static void null_timer_cb(void *opaque) | ||
236 | { | ||
237 | NullAIOCB *acb = opaque; | ||
238 | - AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
239 | - | ||
240 | - aio_context_acquire(ctx); | ||
241 | acb->common.cb(acb->common.opaque, 0); | ||
242 | - aio_context_release(ctx); | ||
243 | timer_deinit(&acb->timer); | ||
244 | qemu_aio_unref(acb); | ||
245 | } | ||
246 | diff --git a/block/qed-cluster.c b/block/qed-cluster.c | ||
247 | index XXXXXXX..XXXXXXX 100644 | ||
248 | --- a/block/qed-cluster.c | ||
249 | +++ b/block/qed-cluster.c | ||
250 | @@ -XXX,XX +XXX,XX @@ static void qed_find_cluster_cb(void *opaque, int ret) | ||
251 | unsigned int index; | ||
252 | unsigned int n; | ||
253 | |||
254 | + qed_acquire(s); | ||
255 | if (ret) { | ||
256 | goto out; | ||
257 | } | ||
258 | @@ -XXX,XX +XXX,XX @@ static void qed_find_cluster_cb(void *opaque, int ret) | ||
259 | |||
260 | out: | ||
261 | find_cluster_cb->cb(find_cluster_cb->opaque, ret, offset, len); | ||
262 | + qed_release(s); | ||
263 | g_free(find_cluster_cb); | ||
264 | } | ||
265 | |||
266 | diff --git a/block/qed-table.c b/block/qed-table.c | ||
267 | index XXXXXXX..XXXXXXX 100644 | ||
268 | --- a/block/qed-table.c | ||
269 | +++ b/block/qed-table.c | ||
270 | @@ -XXX,XX +XXX,XX @@ static void qed_read_table_cb(void *opaque, int ret) | ||
271 | { | ||
272 | QEDReadTableCB *read_table_cb = opaque; | ||
273 | QEDTable *table = read_table_cb->table; | ||
274 | + BDRVQEDState *s = read_table_cb->s; | ||
275 | int noffsets = read_table_cb->qiov.size / sizeof(uint64_t); | ||
276 | int i; | ||
277 | |||
278 | @@ -XXX,XX +XXX,XX @@ static void qed_read_table_cb(void *opaque, int ret) | ||
279 | } | ||
280 | |||
281 | /* Byteswap offsets */ | ||
282 | + qed_acquire(s); | ||
283 | for (i = 0; i < noffsets; i++) { | ||
284 | table->offsets[i] = le64_to_cpu(table->offsets[i]); | ||
285 | } | ||
286 | + qed_release(s); | ||
287 | |||
288 | out: | ||
289 | /* Completion */ | ||
290 | - trace_qed_read_table_cb(read_table_cb->s, read_table_cb->table, ret); | ||
291 | + trace_qed_read_table_cb(s, read_table_cb->table, ret); | ||
292 | gencb_complete(&read_table_cb->gencb, ret); | ||
293 | } | ||
294 | |||
295 | @@ -XXX,XX +XXX,XX @@ typedef struct { | ||
296 | static void qed_write_table_cb(void *opaque, int ret) | ||
297 | { | ||
298 | QEDWriteTableCB *write_table_cb = opaque; | ||
299 | + BDRVQEDState *s = write_table_cb->s; | ||
300 | |||
301 | - trace_qed_write_table_cb(write_table_cb->s, | ||
302 | + trace_qed_write_table_cb(s, | ||
303 | write_table_cb->orig_table, | ||
304 | write_table_cb->flush, | ||
305 | ret); | ||
306 | @@ -XXX,XX +XXX,XX @@ static void qed_write_table_cb(void *opaque, int ret) | ||
307 | if (write_table_cb->flush) { | ||
308 | /* We still need to flush first */ | ||
309 | write_table_cb->flush = false; | ||
310 | + qed_acquire(s); | ||
311 | bdrv_aio_flush(write_table_cb->s->bs, qed_write_table_cb, | ||
312 | write_table_cb); | ||
313 | + qed_release(s); | ||
314 | return; | ||
315 | } | ||
316 | |||
317 | @@ -XXX,XX +XXX,XX @@ static void qed_read_l2_table_cb(void *opaque, int ret) | ||
318 | CachedL2Table *l2_table = request->l2_table; | ||
319 | uint64_t l2_offset = read_l2_table_cb->l2_offset; | ||
320 | |||
321 | + qed_acquire(s); | ||
322 | if (ret) { | ||
323 | /* can't trust loaded L2 table anymore */ | ||
324 | qed_unref_l2_cache_entry(l2_table); | ||
325 | @@ -XXX,XX +XXX,XX @@ static void qed_read_l2_table_cb(void *opaque, int ret) | ||
326 | request->l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset); | ||
327 | assert(request->l2_table != NULL); | ||
328 | } | ||
329 | + qed_release(s); | ||
330 | |||
331 | gencb_complete(&read_l2_table_cb->gencb, ret); | ||
332 | } | ||
333 | diff --git a/block/qed.c b/block/qed.c | ||
334 | index XXXXXXX..XXXXXXX 100644 | ||
335 | --- a/block/qed.c | ||
336 | +++ b/block/qed.c | ||
337 | @@ -XXX,XX +XXX,XX @@ static void qed_is_allocated_cb(void *opaque, int ret, uint64_t offset, size_t l | ||
338 | } | ||
339 | |||
340 | if (cb->co) { | ||
341 | - qemu_coroutine_enter(cb->co); | ||
342 | + aio_co_wake(cb->co); | ||
343 | } | ||
344 | } | ||
345 | |||
346 | @@ -XXX,XX +XXX,XX @@ static void coroutine_fn qed_co_pwrite_zeroes_cb(void *opaque, int ret) | ||
347 | cb->done = true; | ||
348 | cb->ret = ret; | ||
349 | if (cb->co) { | ||
350 | - qemu_coroutine_enter(cb->co); | ||
351 | + aio_co_wake(cb->co); | ||
352 | } | ||
353 | } | ||
354 | |||
355 | diff --git a/block/rbd.c b/block/rbd.c | ||
356 | index XXXXXXX..XXXXXXX 100644 | ||
357 | --- a/block/rbd.c | ||
358 | +++ b/block/rbd.c | ||
359 | @@ -XXX,XX +XXX,XX @@ shutdown: | ||
360 | static void qemu_rbd_complete_aio(RADOSCB *rcb) | ||
361 | { | ||
362 | RBDAIOCB *acb = rcb->acb; | ||
363 | - AioContext *ctx = bdrv_get_aio_context(acb->common.bs); | ||
364 | int64_t r; | ||
365 | |||
366 | r = rcb->ret; | ||
367 | @@ -XXX,XX +XXX,XX @@ static void qemu_rbd_complete_aio(RADOSCB *rcb) | ||
368 | qemu_iovec_from_buf(acb->qiov, 0, acb->bounce, acb->qiov->size); | ||
369 | } | ||
370 | qemu_vfree(acb->bounce); | ||
371 | - | ||
372 | - aio_context_acquire(ctx); | ||
373 | acb->common.cb(acb->common.opaque, (acb->ret > 0 ? 0 : acb->ret)); | ||
374 | - aio_context_release(ctx); | ||
375 | |||
376 | qemu_aio_unref(acb); | ||
377 | } | ||
378 | diff --git a/block/win32-aio.c b/block/win32-aio.c | ||
379 | index XXXXXXX..XXXXXXX 100644 | ||
380 | --- a/block/win32-aio.c | ||
381 | +++ b/block/win32-aio.c | ||
382 | @@ -XXX,XX +XXX,XX @@ static void win32_aio_process_completion(QEMUWin32AIOState *s, | ||
383 | qemu_vfree(waiocb->buf); | ||
384 | } | ||
385 | |||
386 | - | ||
387 | - aio_context_acquire(s->aio_ctx); | ||
388 | waiocb->common.cb(waiocb->common.opaque, ret); | ||
389 | - aio_context_release(s->aio_ctx); | ||
390 | qemu_aio_unref(waiocb); | ||
391 | } | ||
392 | |||
16 | diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c | 393 | diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c |
17 | index XXXXXXX..XXXXXXX 100644 | 394 | index XXXXXXX..XXXXXXX 100644 |
18 | --- a/hw/block/virtio-blk.c | 395 | --- a/hw/block/virtio-blk.c |
19 | +++ b/hw/block/virtio-blk.c | 396 | +++ b/hw/block/virtio-blk.c |
20 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_req_complete(VirtIOBlockReq *req, unsigned char status) | 397 | @@ -XXX,XX +XXX,XX @@ static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error, |
21 | static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error, | 398 | static void virtio_blk_rw_complete(void *opaque, int ret) |
22 | bool is_read) | 399 | { |
23 | { | 400 | VirtIOBlockReq *next = opaque; |
24 | - BlockErrorAction action = blk_get_error_action(req->dev->blk, | 401 | + VirtIOBlock *s = next->dev; |
25 | - is_read, error); | 402 | |
26 | VirtIOBlock *s = req->dev; | 403 | + aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); |
27 | + BlockErrorAction action = blk_get_error_action(s->blk, is_read, error); | 404 | while (next) { |
28 | 405 | VirtIOBlockReq *req = next; | |
29 | if (action == BLOCK_ERROR_ACTION_STOP) { | 406 | next = req->mr_next; |
30 | /* Break the link as the next request is going to be parsed from the | 407 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_rw_complete(void *opaque, int ret) |
31 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_flush_complete(void *opaque, int ret) | 408 | block_acct_done(blk_get_stats(req->dev->blk), &req->acct); |
32 | } | 409 | virtio_blk_free_request(req); |
410 | } | ||
411 | + aio_context_release(blk_get_aio_context(s->conf.conf.blk)); | ||
412 | } | ||
413 | |||
414 | static void virtio_blk_flush_complete(void *opaque, int ret) | ||
415 | { | ||
416 | VirtIOBlockReq *req = opaque; | ||
417 | + VirtIOBlock *s = req->dev; | ||
418 | |||
419 | + aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); | ||
420 | if (ret) { | ||
421 | if (virtio_blk_handle_rw_error(req, -ret, 0)) { | ||
422 | - return; | ||
423 | + goto out; | ||
424 | } | ||
425 | } | ||
33 | 426 | ||
34 | virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); | 427 | virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); |
35 | - block_acct_done(blk_get_stats(req->dev->blk), &req->acct); | 428 | block_acct_done(blk_get_stats(req->dev->blk), &req->acct); |
36 | + block_acct_done(blk_get_stats(s->blk), &req->acct); | ||
37 | virtio_blk_free_request(req); | 429 | virtio_blk_free_request(req); |
430 | + | ||
431 | +out: | ||
432 | + aio_context_release(blk_get_aio_context(s->conf.conf.blk)); | ||
433 | } | ||
434 | |||
435 | #ifdef __linux__ | ||
436 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_ioctl_complete(void *opaque, int status) | ||
437 | virtio_stl_p(vdev, &scsi->data_len, hdr->dxfer_len); | ||
38 | 438 | ||
39 | out: | 439 | out: |
40 | @@ -XXX,XX +XXX,XX @@ static int virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb) | 440 | + aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); |
41 | - sizeof(struct virtio_blk_inhdr); | 441 | virtio_blk_req_complete(req, status); |
42 | iov_discard_back(in_iov, &in_num, sizeof(struct virtio_blk_inhdr)); | 442 | virtio_blk_free_request(req); |
43 | 443 | + aio_context_release(blk_get_aio_context(s->conf.conf.blk)); | |
44 | - type = virtio_ldl_p(VIRTIO_DEVICE(req->dev), &req->out.type); | 444 | g_free(ioctl_req); |
45 | + type = virtio_ldl_p(vdev, &req->out.type); | 445 | } |
46 | 446 | ||
47 | /* VIRTIO_BLK_T_OUT defines the command direction. VIRTIO_BLK_T_BARRIER | 447 | diff --git a/hw/scsi/scsi-disk.c b/hw/scsi/scsi-disk.c |
48 | * is an optional flag. Although a guest should not send this flag if | 448 | index XXXXXXX..XXXXXXX 100644 |
49 | @@ -XXX,XX +XXX,XX @@ static int virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb) | 449 | --- a/hw/scsi/scsi-disk.c |
50 | case VIRTIO_BLK_T_IN: | 450 | +++ b/hw/scsi/scsi-disk.c |
51 | { | 451 | @@ -XXX,XX +XXX,XX @@ static void scsi_aio_complete(void *opaque, int ret) |
52 | bool is_write = type & VIRTIO_BLK_T_OUT; | 452 | |
53 | - req->sector_num = virtio_ldq_p(VIRTIO_DEVICE(req->dev), | 453 | assert(r->req.aiocb != NULL); |
54 | - &req->out.sector); | 454 | r->req.aiocb = NULL; |
55 | + req->sector_num = virtio_ldq_p(vdev, &req->out.sector); | 455 | + aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk)); |
56 | 456 | if (scsi_disk_req_check_error(r, ret, true)) { | |
57 | if (is_write) { | 457 | goto done; |
58 | qemu_iovec_init_external(&req->qiov, out_iov, out_num); | 458 | } |
59 | @@ -XXX,XX +XXX,XX @@ static int virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb) | 459 | @@ -XXX,XX +XXX,XX @@ static void scsi_aio_complete(void *opaque, int ret) |
60 | req->qiov.size / BDRV_SECTOR_SIZE); | 460 | scsi_req_complete(&r->req, GOOD); |
61 | } | 461 | |
62 | 462 | done: | |
63 | - if (!virtio_blk_sect_range_ok(req->dev, req->sector_num, | 463 | + aio_context_release(blk_get_aio_context(s->qdev.conf.blk)); |
64 | - req->qiov.size)) { | 464 | scsi_req_unref(&r->req); |
65 | + if (!virtio_blk_sect_range_ok(s, req->sector_num, req->qiov.size)) { | 465 | } |
66 | virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR); | 466 | |
67 | - block_acct_invalid(blk_get_stats(req->dev->blk), | 467 | @@ -XXX,XX +XXX,XX @@ static void scsi_dma_complete(void *opaque, int ret) |
68 | + block_acct_invalid(blk_get_stats(s->blk), | 468 | assert(r->req.aiocb != NULL); |
69 | is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ); | 469 | r->req.aiocb = NULL; |
70 | virtio_blk_free_request(req); | 470 | |
71 | return 0; | 471 | + aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk)); |
72 | } | 472 | if (ret < 0) { |
73 | 473 | block_acct_failed(blk_get_stats(s->qdev.conf.blk), &r->acct); | |
74 | - block_acct_start(blk_get_stats(req->dev->blk), | 474 | } else { |
75 | - &req->acct, req->qiov.size, | 475 | block_acct_done(blk_get_stats(s->qdev.conf.blk), &r->acct); |
76 | + block_acct_start(blk_get_stats(s->blk), &req->acct, req->qiov.size, | 476 | } |
77 | is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ); | 477 | scsi_dma_complete_noio(r, ret); |
78 | 478 | + aio_context_release(blk_get_aio_context(s->qdev.conf.blk)); | |
79 | /* merge would exceed maximum number of requests or IO direction | 479 | } |
80 | * changes */ | 480 | |
81 | if (mrb->num_reqs > 0 && (mrb->num_reqs == VIRTIO_BLK_MAX_MERGE_REQS || | 481 | static void scsi_read_complete(void * opaque, int ret) |
82 | is_write != mrb->is_write || | 482 | @@ -XXX,XX +XXX,XX @@ static void scsi_read_complete(void * opaque, int ret) |
83 | - !req->dev->conf.request_merging)) { | 483 | |
84 | - virtio_blk_submit_multireq(req->dev->blk, mrb); | 484 | assert(r->req.aiocb != NULL); |
85 | + !s->conf.request_merging)) { | 485 | r->req.aiocb = NULL; |
86 | + virtio_blk_submit_multireq(s->blk, mrb); | 486 | + aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk)); |
87 | } | 487 | if (scsi_disk_req_check_error(r, ret, true)) { |
88 | 488 | goto done; | |
89 | assert(mrb->num_reqs < VIRTIO_BLK_MAX_MERGE_REQS); | 489 | } |
490 | @@ -XXX,XX +XXX,XX @@ static void scsi_read_complete(void * opaque, int ret) | ||
491 | |||
492 | done: | ||
493 | scsi_req_unref(&r->req); | ||
494 | + aio_context_release(blk_get_aio_context(s->qdev.conf.blk)); | ||
495 | } | ||
496 | |||
497 | /* Actually issue a read to the block device. */ | ||
498 | @@ -XXX,XX +XXX,XX @@ static void scsi_do_read_cb(void *opaque, int ret) | ||
499 | assert (r->req.aiocb != NULL); | ||
500 | r->req.aiocb = NULL; | ||
501 | |||
502 | + aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk)); | ||
503 | if (ret < 0) { | ||
504 | block_acct_failed(blk_get_stats(s->qdev.conf.blk), &r->acct); | ||
505 | } else { | ||
506 | block_acct_done(blk_get_stats(s->qdev.conf.blk), &r->acct); | ||
507 | } | ||
508 | scsi_do_read(opaque, ret); | ||
509 | + aio_context_release(blk_get_aio_context(s->qdev.conf.blk)); | ||
510 | } | ||
511 | |||
512 | /* Read more data from scsi device into buffer. */ | ||
513 | @@ -XXX,XX +XXX,XX @@ static void scsi_write_complete(void * opaque, int ret) | ||
514 | assert (r->req.aiocb != NULL); | ||
515 | r->req.aiocb = NULL; | ||
516 | |||
517 | + aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk)); | ||
518 | if (ret < 0) { | ||
519 | block_acct_failed(blk_get_stats(s->qdev.conf.blk), &r->acct); | ||
520 | } else { | ||
521 | block_acct_done(blk_get_stats(s->qdev.conf.blk), &r->acct); | ||
522 | } | ||
523 | scsi_write_complete_noio(r, ret); | ||
524 | + aio_context_release(blk_get_aio_context(s->qdev.conf.blk)); | ||
525 | } | ||
526 | |||
527 | static void scsi_write_data(SCSIRequest *req) | ||
528 | @@ -XXX,XX +XXX,XX @@ static void scsi_unmap_complete(void *opaque, int ret) | ||
529 | { | ||
530 | UnmapCBData *data = opaque; | ||
531 | SCSIDiskReq *r = data->r; | ||
532 | + SCSIDiskState *s = DO_UPCAST(SCSIDiskState, qdev, r->req.dev); | ||
533 | |||
534 | assert(r->req.aiocb != NULL); | ||
535 | r->req.aiocb = NULL; | ||
536 | |||
537 | + aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk)); | ||
538 | scsi_unmap_complete_noio(data, ret); | ||
539 | + aio_context_release(blk_get_aio_context(s->qdev.conf.blk)); | ||
540 | } | ||
541 | |||
542 | static void scsi_disk_emulate_unmap(SCSIDiskReq *r, uint8_t *inbuf) | ||
543 | @@ -XXX,XX +XXX,XX @@ static void scsi_write_same_complete(void *opaque, int ret) | ||
544 | |||
545 | assert(r->req.aiocb != NULL); | ||
546 | r->req.aiocb = NULL; | ||
547 | + aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk)); | ||
548 | if (scsi_disk_req_check_error(r, ret, true)) { | ||
549 | goto done; | ||
550 | } | ||
551 | @@ -XXX,XX +XXX,XX @@ done: | ||
552 | scsi_req_unref(&r->req); | ||
553 | qemu_vfree(data->iov.iov_base); | ||
554 | g_free(data); | ||
555 | + aio_context_release(blk_get_aio_context(s->qdev.conf.blk)); | ||
556 | } | ||
557 | |||
558 | static void scsi_disk_emulate_write_same(SCSIDiskReq *r, uint8_t *inbuf) | ||
559 | diff --git a/hw/scsi/scsi-generic.c b/hw/scsi/scsi-generic.c | ||
560 | index XXXXXXX..XXXXXXX 100644 | ||
561 | --- a/hw/scsi/scsi-generic.c | ||
562 | +++ b/hw/scsi/scsi-generic.c | ||
563 | @@ -XXX,XX +XXX,XX @@ done: | ||
564 | static void scsi_command_complete(void *opaque, int ret) | ||
565 | { | ||
566 | SCSIGenericReq *r = (SCSIGenericReq *)opaque; | ||
567 | + SCSIDevice *s = r->req.dev; | ||
568 | |||
569 | assert(r->req.aiocb != NULL); | ||
570 | r->req.aiocb = NULL; | ||
571 | + | ||
572 | + aio_context_acquire(blk_get_aio_context(s->conf.blk)); | ||
573 | scsi_command_complete_noio(r, ret); | ||
574 | + aio_context_release(blk_get_aio_context(s->conf.blk)); | ||
575 | } | ||
576 | |||
577 | static int execute_command(BlockBackend *blk, | ||
578 | @@ -XXX,XX +XXX,XX @@ static void scsi_read_complete(void * opaque, int ret) | ||
579 | assert(r->req.aiocb != NULL); | ||
580 | r->req.aiocb = NULL; | ||
581 | |||
582 | + aio_context_acquire(blk_get_aio_context(s->conf.blk)); | ||
583 | + | ||
584 | if (ret || r->req.io_canceled) { | ||
585 | scsi_command_complete_noio(r, ret); | ||
586 | - return; | ||
587 | + goto done; | ||
588 | } | ||
589 | |||
590 | len = r->io_header.dxfer_len - r->io_header.resid; | ||
591 | @@ -XXX,XX +XXX,XX @@ static void scsi_read_complete(void * opaque, int ret) | ||
592 | r->len = -1; | ||
593 | if (len == 0) { | ||
594 | scsi_command_complete_noio(r, 0); | ||
595 | - return; | ||
596 | + goto done; | ||
597 | } | ||
598 | |||
599 | /* Snoop READ CAPACITY output to set the blocksize. */ | ||
600 | @@ -XXX,XX +XXX,XX @@ static void scsi_read_complete(void * opaque, int ret) | ||
601 | } | ||
602 | scsi_req_data(&r->req, len); | ||
603 | scsi_req_unref(&r->req); | ||
604 | + | ||
605 | +done: | ||
606 | + aio_context_release(blk_get_aio_context(s->conf.blk)); | ||
607 | } | ||
608 | |||
609 | /* Read more data from scsi device into buffer. */ | ||
610 | @@ -XXX,XX +XXX,XX @@ static void scsi_write_complete(void * opaque, int ret) | ||
611 | assert(r->req.aiocb != NULL); | ||
612 | r->req.aiocb = NULL; | ||
613 | |||
614 | + aio_context_acquire(blk_get_aio_context(s->conf.blk)); | ||
615 | + | ||
616 | if (ret || r->req.io_canceled) { | ||
617 | scsi_command_complete_noio(r, ret); | ||
618 | - return; | ||
619 | + goto done; | ||
620 | } | ||
621 | |||
622 | if (r->req.cmd.buf[0] == MODE_SELECT && r->req.cmd.buf[4] == 12 && | ||
623 | @@ -XXX,XX +XXX,XX @@ static void scsi_write_complete(void * opaque, int ret) | ||
624 | } | ||
625 | |||
626 | scsi_command_complete_noio(r, ret); | ||
627 | + | ||
628 | +done: | ||
629 | + aio_context_release(blk_get_aio_context(s->conf.blk)); | ||
630 | } | ||
631 | |||
632 | /* Write data to a scsi device. Returns nonzero on failure. | ||
633 | diff --git a/util/thread-pool.c b/util/thread-pool.c | ||
634 | index XXXXXXX..XXXXXXX 100644 | ||
635 | --- a/util/thread-pool.c | ||
636 | +++ b/util/thread-pool.c | ||
637 | @@ -XXX,XX +XXX,XX @@ restart: | ||
638 | */ | ||
639 | qemu_bh_schedule(pool->completion_bh); | ||
640 | |||
641 | + aio_context_release(pool->ctx); | ||
642 | elem->common.cb(elem->common.opaque, elem->ret); | ||
643 | + aio_context_acquire(pool->ctx); | ||
644 | qemu_aio_unref(elem); | ||
645 | goto restart; | ||
646 | } else { | ||
647 | @@ -XXX,XX +XXX,XX @@ static void thread_pool_co_cb(void *opaque, int ret) | ||
648 | ThreadPoolCo *co = opaque; | ||
649 | |||
650 | co->ret = ret; | ||
651 | - qemu_coroutine_enter(co->co); | ||
652 | + aio_co_wake(co->co); | ||
653 | } | ||
654 | |||
655 | int coroutine_fn thread_pool_submit_co(ThreadPool *pool, ThreadPoolFunc *func, | ||
90 | -- | 656 | -- |
91 | 2.20.1 | 657 | 2.9.3 |
92 | 658 | ||
93 | 659 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
1 | 2 | ||
3 | This patch prepares for the removal of unnecessary lockcnt inc/dec pairs. | ||
4 | Extract the dispatching loop for file descriptor handlers into a new | ||
5 | function aio_dispatch_handlers, and then inline aio_dispatch into | ||
6 | aio_poll. | ||
7 | |||
8 | aio_dispatch can now become void. | ||
9 | |||
10 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
11 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
12 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
13 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> | ||
14 | Message-id: 20170213135235.12274-17-pbonzini@redhat.com | ||
15 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
16 | --- | ||
17 | include/block/aio.h | 6 +----- | ||
18 | util/aio-posix.c | 44 ++++++++++++++------------------------------ | ||
19 | util/aio-win32.c | 13 ++++--------- | ||
20 | util/async.c | 2 +- | ||
21 | 4 files changed, 20 insertions(+), 45 deletions(-) | ||
22 | |||
23 | diff --git a/include/block/aio.h b/include/block/aio.h | ||
24 | index XXXXXXX..XXXXXXX 100644 | ||
25 | --- a/include/block/aio.h | ||
26 | +++ b/include/block/aio.h | ||
27 | @@ -XXX,XX +XXX,XX @@ bool aio_pending(AioContext *ctx); | ||
28 | /* Dispatch any pending callbacks from the GSource attached to the AioContext. | ||
29 | * | ||
30 | * This is used internally in the implementation of the GSource. | ||
31 | - * | ||
32 | - * @dispatch_fds: true to process fds, false to skip them | ||
33 | - * (can be used as an optimization by callers that know there | ||
34 | - * are no fds ready) | ||
35 | */ | ||
36 | -bool aio_dispatch(AioContext *ctx, bool dispatch_fds); | ||
37 | +void aio_dispatch(AioContext *ctx); | ||
38 | |||
39 | /* Progress in completing AIO work to occur. This can issue new pending | ||
40 | * aio as a result of executing I/O completion or bh callbacks. | ||
41 | diff --git a/util/aio-posix.c b/util/aio-posix.c | ||
42 | index XXXXXXX..XXXXXXX 100644 | ||
43 | --- a/util/aio-posix.c | ||
44 | +++ b/util/aio-posix.c | ||
45 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx) | ||
46 | AioHandler *node, *tmp; | ||
47 | bool progress = false; | ||
48 | |||
49 | - /* | ||
50 | - * We have to walk very carefully in case aio_set_fd_handler is | ||
51 | - * called while we're walking. | ||
52 | - */ | ||
53 | - qemu_lockcnt_inc(&ctx->list_lock); | ||
54 | - | ||
55 | QLIST_FOREACH_SAFE_RCU(node, &ctx->aio_handlers, node, tmp) { | ||
56 | int revents; | ||
57 | |||
58 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx) | ||
59 | } | ||
60 | } | ||
61 | |||
62 | - qemu_lockcnt_dec(&ctx->list_lock); | ||
63 | return progress; | ||
64 | } | ||
65 | |||
66 | -/* | ||
67 | - * Note that dispatch_fds == false has the side-effect of post-poning the | ||
68 | - * freeing of deleted handlers. | ||
69 | - */ | ||
70 | -bool aio_dispatch(AioContext *ctx, bool dispatch_fds) | ||
71 | +void aio_dispatch(AioContext *ctx) | ||
72 | { | ||
73 | - bool progress; | ||
74 | + aio_bh_poll(ctx); | ||
75 | |||
76 | - /* | ||
77 | - * If there are callbacks left that have been queued, we need to call them. | ||
78 | - * Do not call select in this case, because it is possible that the caller | ||
79 | - * does not need a complete flush (as is the case for aio_poll loops). | ||
80 | - */ | ||
81 | - progress = aio_bh_poll(ctx); | ||
82 | + qemu_lockcnt_inc(&ctx->list_lock); | ||
83 | + aio_dispatch_handlers(ctx); | ||
84 | + qemu_lockcnt_dec(&ctx->list_lock); | ||
85 | |||
86 | - if (dispatch_fds) { | ||
87 | - progress |= aio_dispatch_handlers(ctx); | ||
88 | - } | ||
89 | - | ||
90 | - /* Run our timers */ | ||
91 | - progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
92 | - | ||
93 | - return progress; | ||
94 | + timerlistgroup_run_timers(&ctx->tlg); | ||
95 | } | ||
96 | |||
97 | /* These thread-local variables are used only in a small part of aio_poll | ||
98 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
99 | npfd = 0; | ||
100 | qemu_lockcnt_dec(&ctx->list_lock); | ||
101 | |||
102 | - /* Run dispatch even if there were no readable fds to run timers */ | ||
103 | - if (aio_dispatch(ctx, ret > 0)) { | ||
104 | - progress = true; | ||
105 | + progress |= aio_bh_poll(ctx); | ||
106 | + | ||
107 | + if (ret > 0) { | ||
108 | + qemu_lockcnt_inc(&ctx->list_lock); | ||
109 | + progress |= aio_dispatch_handlers(ctx); | ||
110 | + qemu_lockcnt_dec(&ctx->list_lock); | ||
111 | } | ||
112 | |||
113 | + progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
114 | + | ||
115 | return progress; | ||
116 | } | ||
117 | |||
118 | diff --git a/util/aio-win32.c b/util/aio-win32.c | ||
119 | index XXXXXXX..XXXXXXX 100644 | ||
120 | --- a/util/aio-win32.c | ||
121 | +++ b/util/aio-win32.c | ||
122 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) | ||
123 | return progress; | ||
124 | } | ||
125 | |||
126 | -bool aio_dispatch(AioContext *ctx, bool dispatch_fds) | ||
127 | +void aio_dispatch(AioContext *ctx) | ||
128 | { | ||
129 | - bool progress; | ||
130 | - | ||
131 | - progress = aio_bh_poll(ctx); | ||
132 | - if (dispatch_fds) { | ||
133 | - progress |= aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE); | ||
134 | - } | ||
135 | - progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
136 | - return progress; | ||
137 | + aio_bh_poll(ctx); | ||
138 | + aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE); | ||
139 | + timerlistgroup_run_timers(&ctx->tlg); | ||
140 | } | ||
141 | |||
142 | bool aio_poll(AioContext *ctx, bool blocking) | ||
143 | diff --git a/util/async.c b/util/async.c | ||
144 | index XXXXXXX..XXXXXXX 100644 | ||
145 | --- a/util/async.c | ||
146 | +++ b/util/async.c | ||
147 | @@ -XXX,XX +XXX,XX @@ aio_ctx_dispatch(GSource *source, | ||
148 | AioContext *ctx = (AioContext *) source; | ||
149 | |||
150 | assert(callback == NULL); | ||
151 | - aio_dispatch(ctx, true); | ||
152 | + aio_dispatch(ctx); | ||
153 | return true; | ||
154 | } | ||
155 | |||
156 | -- | ||
157 | 2.9.3 | ||
158 | |||
159 | diff view generated by jsdifflib |
1 | From: Stefano Garzarella <sgarzare@redhat.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | We add acct_failed param in order to use virtio_blk_handle_rw_error() | 3 | Pull the increment/decrement pair out of aio_bh_poll and into the |
4 | also when is not required to call block_acct_failed(). (eg. a discard | 4 | callers. |
5 | operation is failed) | ||
6 | 5 | ||
7 | Reviewed-by: Michael S. Tsirkin <mst@redhat.com> | ||
8 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | 6 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> |
9 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | 7 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
10 | Acked-by: Pankaj Gupta <pagupta@redhat.com> | 8 | Reviewed-by: Fam Zheng <famz@redhat.com> |
11 | Message-id: 20190208134950.187665-2-sgarzare@redhat.com | 9 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> |
10 | Message-id: 20170213135235.12274-18-pbonzini@redhat.com | ||
12 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 11 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
13 | --- | 12 | --- |
14 | hw/block/virtio-blk.c | 10 ++++++---- | 13 | util/aio-posix.c | 8 +++----- |
15 | 1 file changed, 6 insertions(+), 4 deletions(-) | 14 | util/aio-win32.c | 8 ++++---- |
15 | util/async.c | 12 ++++++------ | ||
16 | 3 files changed, 13 insertions(+), 15 deletions(-) | ||
16 | 17 | ||
17 | diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c | 18 | diff --git a/util/aio-posix.c b/util/aio-posix.c |
18 | index XXXXXXX..XXXXXXX 100644 | 19 | index XXXXXXX..XXXXXXX 100644 |
19 | --- a/hw/block/virtio-blk.c | 20 | --- a/util/aio-posix.c |
20 | +++ b/hw/block/virtio-blk.c | 21 | +++ b/util/aio-posix.c |
21 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_req_complete(VirtIOBlockReq *req, unsigned char status) | 22 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx) |
23 | |||
24 | void aio_dispatch(AioContext *ctx) | ||
25 | { | ||
26 | + qemu_lockcnt_inc(&ctx->list_lock); | ||
27 | aio_bh_poll(ctx); | ||
28 | - | ||
29 | - qemu_lockcnt_inc(&ctx->list_lock); | ||
30 | aio_dispatch_handlers(ctx); | ||
31 | qemu_lockcnt_dec(&ctx->list_lock); | ||
32 | |||
33 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) | ||
34 | } | ||
35 | |||
36 | npfd = 0; | ||
37 | - qemu_lockcnt_dec(&ctx->list_lock); | ||
38 | |||
39 | progress |= aio_bh_poll(ctx); | ||
40 | |||
41 | if (ret > 0) { | ||
42 | - qemu_lockcnt_inc(&ctx->list_lock); | ||
43 | progress |= aio_dispatch_handlers(ctx); | ||
44 | - qemu_lockcnt_dec(&ctx->list_lock); | ||
45 | } | ||
46 | |||
47 | + qemu_lockcnt_dec(&ctx->list_lock); | ||
48 | + | ||
49 | progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
50 | |||
51 | return progress; | ||
52 | diff --git a/util/aio-win32.c b/util/aio-win32.c | ||
53 | index XXXXXXX..XXXXXXX 100644 | ||
54 | --- a/util/aio-win32.c | ||
55 | +++ b/util/aio-win32.c | ||
56 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) | ||
57 | bool progress = false; | ||
58 | AioHandler *tmp; | ||
59 | |||
60 | - qemu_lockcnt_inc(&ctx->list_lock); | ||
61 | - | ||
62 | /* | ||
63 | * We have to walk very carefully in case aio_set_fd_handler is | ||
64 | * called while we're walking. | ||
65 | @@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event) | ||
66 | } | ||
67 | } | ||
68 | |||
69 | - qemu_lockcnt_dec(&ctx->list_lock); | ||
70 | return progress; | ||
22 | } | 71 | } |
23 | 72 | ||
24 | static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error, | 73 | void aio_dispatch(AioContext *ctx) |
25 | - bool is_read) | ||
26 | + bool is_read, bool acct_failed) | ||
27 | { | 74 | { |
28 | VirtIOBlock *s = req->dev; | 75 | + qemu_lockcnt_inc(&ctx->list_lock); |
29 | BlockErrorAction action = blk_get_error_action(s->blk, is_read, error); | 76 | aio_bh_poll(ctx); |
30 | @@ -XXX,XX +XXX,XX @@ static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error, | 77 | aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE); |
31 | s->rq = req; | 78 | + qemu_lockcnt_dec(&ctx->list_lock); |
32 | } else if (action == BLOCK_ERROR_ACTION_REPORT) { | 79 | timerlistgroup_run_timers(&ctx->tlg); |
33 | virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR); | 80 | } |
34 | - block_acct_failed(blk_get_stats(s->blk), &req->acct); | 81 | |
35 | + if (acct_failed) { | 82 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) |
36 | + block_acct_failed(blk_get_stats(s->blk), &req->acct); | 83 | } |
37 | + } | ||
38 | virtio_blk_free_request(req); | ||
39 | } | 84 | } |
40 | 85 | ||
41 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_rw_complete(void *opaque, int ret) | 86 | - qemu_lockcnt_dec(&ctx->list_lock); |
42 | * the memory until the request is completed (which will | 87 | first = true; |
43 | * happen on the other side of the migration). | 88 | |
44 | */ | 89 | /* ctx->notifier is always registered. */ |
45 | - if (virtio_blk_handle_rw_error(req, -ret, is_read)) { | 90 | @@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking) |
46 | + if (virtio_blk_handle_rw_error(req, -ret, is_read, true)) { | 91 | progress |= aio_dispatch_handlers(ctx, event); |
47 | continue; | 92 | } while (count > 0); |
93 | |||
94 | + qemu_lockcnt_dec(&ctx->list_lock); | ||
95 | + | ||
96 | progress |= timerlistgroup_run_timers(&ctx->tlg); | ||
97 | return progress; | ||
98 | } | ||
99 | diff --git a/util/async.c b/util/async.c | ||
100 | index XXXXXXX..XXXXXXX 100644 | ||
101 | --- a/util/async.c | ||
102 | +++ b/util/async.c | ||
103 | @@ -XXX,XX +XXX,XX @@ void aio_bh_call(QEMUBH *bh) | ||
104 | bh->cb(bh->opaque); | ||
105 | } | ||
106 | |||
107 | -/* Multiple occurrences of aio_bh_poll cannot be called concurrently */ | ||
108 | +/* Multiple occurrences of aio_bh_poll cannot be called concurrently. | ||
109 | + * The count in ctx->list_lock is incremented before the call, and is | ||
110 | + * not affected by the call. | ||
111 | + */ | ||
112 | int aio_bh_poll(AioContext *ctx) | ||
113 | { | ||
114 | QEMUBH *bh, **bhp, *next; | ||
115 | int ret; | ||
116 | bool deleted = false; | ||
117 | |||
118 | - qemu_lockcnt_inc(&ctx->list_lock); | ||
119 | - | ||
120 | ret = 0; | ||
121 | for (bh = atomic_rcu_read(&ctx->first_bh); bh; bh = next) { | ||
122 | next = atomic_rcu_read(&bh->next); | ||
123 | @@ -XXX,XX +XXX,XX @@ int aio_bh_poll(AioContext *ctx) | ||
124 | |||
125 | /* remove deleted bhs */ | ||
126 | if (!deleted) { | ||
127 | - qemu_lockcnt_dec(&ctx->list_lock); | ||
128 | return ret; | ||
129 | } | ||
130 | |||
131 | - if (qemu_lockcnt_dec_and_lock(&ctx->list_lock)) { | ||
132 | + if (qemu_lockcnt_dec_if_lock(&ctx->list_lock)) { | ||
133 | bhp = &ctx->first_bh; | ||
134 | while (*bhp) { | ||
135 | bh = *bhp; | ||
136 | @@ -XXX,XX +XXX,XX @@ int aio_bh_poll(AioContext *ctx) | ||
137 | bhp = &bh->next; | ||
48 | } | 138 | } |
49 | } | 139 | } |
50 | @@ -XXX,XX +XXX,XX @@ static void virtio_blk_flush_complete(void *opaque, int ret) | 140 | - qemu_lockcnt_unlock(&ctx->list_lock); |
51 | 141 | + qemu_lockcnt_inc_and_unlock(&ctx->list_lock); | |
52 | aio_context_acquire(blk_get_aio_context(s->conf.conf.blk)); | ||
53 | if (ret) { | ||
54 | - if (virtio_blk_handle_rw_error(req, -ret, 0)) { | ||
55 | + if (virtio_blk_handle_rw_error(req, -ret, 0, true)) { | ||
56 | goto out; | ||
57 | } | ||
58 | } | 142 | } |
143 | return ret; | ||
144 | } | ||
59 | -- | 145 | -- |
60 | 2.20.1 | 146 | 2.9.3 |
61 | 147 | ||
62 | 148 | diff view generated by jsdifflib |
1 | From: Stefano Garzarella <sgarzare@redhat.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | If the WRITE_ZEROES feature is enabled, we check this command | ||
4 | in the test_basic(). | ||
5 | |||
6 | Reviewed-by: Michael S. Tsirkin <mst@redhat.com> | ||
7 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | 3 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> |
8 | Acked-by: Thomas Huth <thuth@redhat.com> | 4 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
9 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | 5 | Reviewed-by: Fam Zheng <famz@redhat.com> |
10 | Acked-by: Pankaj Gupta <pagupta@redhat.com> | 6 | Reviewed-by: Daniel P. Berrange <berrange@redhat.com> |
11 | Message-id: 20190208134950.187665-7-sgarzare@redhat.com | 7 | Message-id: 20170213135235.12274-19-pbonzini@redhat.com |
12 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 8 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
13 | --- | 9 | --- |
14 | tests/virtio-blk-test.c | 60 +++++++++++++++++++++++++++++++++++++++++ | 10 | include/block/block_int.h | 64 +++++++++++++++++++++++++----------------- |
15 | 1 file changed, 60 insertions(+) | 11 | include/sysemu/block-backend.h | 14 ++++++--- |
12 | 2 files changed, 49 insertions(+), 29 deletions(-) | ||
16 | 13 | ||
17 | diff --git a/tests/virtio-blk-test.c b/tests/virtio-blk-test.c | 14 | diff --git a/include/block/block_int.h b/include/block/block_int.h |
18 | index XXXXXXX..XXXXXXX 100644 | 15 | index XXXXXXX..XXXXXXX 100644 |
19 | --- a/tests/virtio-blk-test.c | 16 | --- a/include/block/block_int.h |
20 | +++ b/tests/virtio-blk-test.c | 17 | +++ b/include/block/block_int.h |
21 | @@ -XXX,XX +XXX,XX @@ static void test_basic(QVirtioDevice *dev, QGuestAllocator *alloc, | 18 | @@ -XXX,XX +XXX,XX @@ struct BdrvChild { |
22 | 19 | * copied as well. | |
23 | guest_free(alloc, req_addr); | 20 | */ |
24 | 21 | struct BlockDriverState { | |
25 | + if (features & (1u << VIRTIO_BLK_F_WRITE_ZEROES)) { | 22 | - int64_t total_sectors; /* if we are reading a disk image, give its |
26 | + struct virtio_blk_discard_write_zeroes dwz_hdr; | 23 | - size in sectors */ |
27 | + void *expected; | 24 | + /* Protected by big QEMU lock or read-only after opening. No special |
25 | + * locking needed during I/O... | ||
26 | + */ | ||
27 | int open_flags; /* flags used to open the file, re-used for re-open */ | ||
28 | bool read_only; /* if true, the media is read only */ | ||
29 | bool encrypted; /* if true, the media is encrypted */ | ||
30 | @@ -XXX,XX +XXX,XX @@ struct BlockDriverState { | ||
31 | bool sg; /* if true, the device is a /dev/sg* */ | ||
32 | bool probed; /* if true, format was probed rather than specified */ | ||
33 | |||
34 | - int copy_on_read; /* if nonzero, copy read backing sectors into image. | ||
35 | - note this is a reference count */ | ||
36 | - | ||
37 | - CoQueue flush_queue; /* Serializing flush queue */ | ||
38 | - bool active_flush_req; /* Flush request in flight? */ | ||
39 | - unsigned int write_gen; /* Current data generation */ | ||
40 | - unsigned int flushed_gen; /* Flushed write generation */ | ||
41 | - | ||
42 | BlockDriver *drv; /* NULL means no media */ | ||
43 | void *opaque; | ||
44 | |||
45 | @@ -XXX,XX +XXX,XX @@ struct BlockDriverState { | ||
46 | BdrvChild *backing; | ||
47 | BdrvChild *file; | ||
48 | |||
49 | - /* Callback before write request is processed */ | ||
50 | - NotifierWithReturnList before_write_notifiers; | ||
51 | - | ||
52 | - /* number of in-flight requests; overall and serialising */ | ||
53 | - unsigned int in_flight; | ||
54 | - unsigned int serialising_in_flight; | ||
55 | - | ||
56 | - bool wakeup; | ||
57 | - | ||
58 | - /* Offset after the highest byte written to */ | ||
59 | - uint64_t wr_highest_offset; | ||
60 | - | ||
61 | /* I/O Limits */ | ||
62 | BlockLimits bl; | ||
63 | |||
64 | @@ -XXX,XX +XXX,XX @@ struct BlockDriverState { | ||
65 | QTAILQ_ENTRY(BlockDriverState) bs_list; | ||
66 | /* element of the list of monitor-owned BDS */ | ||
67 | QTAILQ_ENTRY(BlockDriverState) monitor_list; | ||
68 | - QLIST_HEAD(, BdrvDirtyBitmap) dirty_bitmaps; | ||
69 | int refcnt; | ||
70 | |||
71 | - QLIST_HEAD(, BdrvTrackedRequest) tracked_requests; | ||
72 | - | ||
73 | /* operation blockers */ | ||
74 | QLIST_HEAD(, BdrvOpBlocker) op_blockers[BLOCK_OP_TYPE_MAX]; | ||
75 | |||
76 | @@ -XXX,XX +XXX,XX @@ struct BlockDriverState { | ||
77 | /* The error object in use for blocking operations on backing_hd */ | ||
78 | Error *backing_blocker; | ||
79 | |||
80 | + /* Protected by AioContext lock */ | ||
28 | + | 81 | + |
29 | + /* | 82 | + /* If true, copy read backing sectors into image. Can be >1 if more |
30 | + * WRITE_ZEROES request on the same sector of previous test where | 83 | + * than one client has requested copy-on-read. |
31 | + * we wrote "TEST". | 84 | + */ |
32 | + */ | 85 | + int copy_on_read; |
33 | + req.type = VIRTIO_BLK_T_WRITE_ZEROES; | ||
34 | + req.data = (char *) &dwz_hdr; | ||
35 | + dwz_hdr.sector = 0; | ||
36 | + dwz_hdr.num_sectors = 1; | ||
37 | + dwz_hdr.flags = 0; | ||
38 | + | 86 | + |
39 | + req_addr = virtio_blk_request(alloc, dev, &req, sizeof(dwz_hdr)); | 87 | + /* If we are reading a disk image, give its size in sectors. |
88 | + * Generally read-only; it is written to by load_vmstate and save_vmstate, | ||
89 | + * but the block layer is quiescent during those. | ||
90 | + */ | ||
91 | + int64_t total_sectors; | ||
40 | + | 92 | + |
41 | + free_head = qvirtqueue_add(vq, req_addr, 16, false, true); | 93 | + /* Callback before write request is processed */ |
42 | + qvirtqueue_add(vq, req_addr + 16, sizeof(dwz_hdr), false, true); | 94 | + NotifierWithReturnList before_write_notifiers; |
43 | + qvirtqueue_add(vq, req_addr + 16 + sizeof(dwz_hdr), 1, true, false); | ||
44 | + | 95 | + |
45 | + qvirtqueue_kick(dev, vq, free_head); | 96 | + /* number of in-flight requests; overall and serialising */ |
97 | + unsigned int in_flight; | ||
98 | + unsigned int serialising_in_flight; | ||
46 | + | 99 | + |
47 | + qvirtio_wait_used_elem(dev, vq, free_head, NULL, | 100 | + bool wakeup; |
48 | + QVIRTIO_BLK_TIMEOUT_US); | ||
49 | + status = readb(req_addr + 16 + sizeof(dwz_hdr)); | ||
50 | + g_assert_cmpint(status, ==, 0); | ||
51 | + | 101 | + |
52 | + guest_free(alloc, req_addr); | 102 | + /* Offset after the highest byte written to */ |
103 | + uint64_t wr_highest_offset; | ||
53 | + | 104 | + |
54 | + /* Read request to check if the sector contains all zeroes */ | 105 | /* threshold limit for writes, in bytes. "High water mark". */ |
55 | + req.type = VIRTIO_BLK_T_IN; | 106 | uint64_t write_threshold_offset; |
56 | + req.ioprio = 1; | 107 | NotifierWithReturn write_threshold_notifier; |
57 | + req.sector = 0; | 108 | @@ -XXX,XX +XXX,XX @@ struct BlockDriverState { |
58 | + req.data = g_malloc0(512); | 109 | /* counter for nested bdrv_io_plug */ |
110 | unsigned io_plugged; | ||
111 | |||
112 | + QLIST_HEAD(, BdrvTrackedRequest) tracked_requests; | ||
113 | + CoQueue flush_queue; /* Serializing flush queue */ | ||
114 | + bool active_flush_req; /* Flush request in flight? */ | ||
115 | + unsigned int write_gen; /* Current data generation */ | ||
116 | + unsigned int flushed_gen; /* Flushed write generation */ | ||
59 | + | 117 | + |
60 | + req_addr = virtio_blk_request(alloc, dev, &req, 512); | 118 | + QLIST_HEAD(, BdrvDirtyBitmap) dirty_bitmaps; |
61 | + | 119 | + |
62 | + g_free(req.data); | 120 | + /* do we need to tell the quest if we have a volatile write cache? */ |
121 | + int enable_write_cache; | ||
63 | + | 122 | + |
64 | + free_head = qvirtqueue_add(vq, req_addr, 16, false, true); | 123 | int quiesce_counter; |
65 | + qvirtqueue_add(vq, req_addr + 16, 512, true, true); | 124 | }; |
66 | + qvirtqueue_add(vq, req_addr + 528, 1, true, false); | 125 | |
126 | diff --git a/include/sysemu/block-backend.h b/include/sysemu/block-backend.h | ||
127 | index XXXXXXX..XXXXXXX 100644 | ||
128 | --- a/include/sysemu/block-backend.h | ||
129 | +++ b/include/sysemu/block-backend.h | ||
130 | @@ -XXX,XX +XXX,XX @@ typedef struct BlockDevOps { | ||
131 | * fields that must be public. This is in particular for QLIST_ENTRY() and | ||
132 | * friends so that BlockBackends can be kept in lists outside block-backend.c */ | ||
133 | typedef struct BlockBackendPublic { | ||
134 | - /* I/O throttling. | ||
135 | - * throttle_state tells us if this BlockBackend has I/O limits configured. | ||
136 | - * io_limits_disabled tells us if they are currently being enforced */ | ||
137 | + /* I/O throttling has its own locking, but also some fields are | ||
138 | + * protected by the AioContext lock. | ||
139 | + */ | ||
67 | + | 140 | + |
68 | + qvirtqueue_kick(dev, vq, free_head); | 141 | + /* Protected by AioContext lock. */ |
142 | CoQueue throttled_reqs[2]; | ||
69 | + | 143 | + |
70 | + qvirtio_wait_used_elem(dev, vq, free_head, NULL, | 144 | + /* Nonzero if the I/O limits are currently being ignored; generally |
71 | + QVIRTIO_BLK_TIMEOUT_US); | 145 | + * it is zero. */ |
72 | + status = readb(req_addr + 528); | 146 | unsigned int io_limits_disabled; |
73 | + g_assert_cmpint(status, ==, 0); | 147 | |
74 | + | 148 | /* The following fields are protected by the ThrottleGroup lock. |
75 | + data = g_malloc(512); | 149 | - * See the ThrottleGroup documentation for details. */ |
76 | + expected = g_malloc0(512); | 150 | + * See the ThrottleGroup documentation for details. |
77 | + memread(req_addr + 16, data, 512); | 151 | + * throttle_state tells us if I/O limits are configured. */ |
78 | + g_assert_cmpmem(data, 512, expected, 512); | 152 | ThrottleState *throttle_state; |
79 | + g_free(expected); | 153 | ThrottleTimers throttle_timers; |
80 | + g_free(data); | 154 | unsigned pending_reqs[2]; |
81 | + | ||
82 | + guest_free(alloc, req_addr); | ||
83 | + } | ||
84 | + | ||
85 | if (features & (1u << VIRTIO_F_ANY_LAYOUT)) { | ||
86 | /* Write and read with 2 descriptor layout */ | ||
87 | /* Write request */ | ||
88 | -- | 155 | -- |
89 | 2.20.1 | 156 | 2.9.3 |
90 | 157 | ||
91 | 158 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | 1 | From: Paolo Bonzini <pbonzini@redhat.com> | |
2 | |||
3 | This uses the lock-free mutex described in the paper '"Blocking without | ||
4 | Locking", or LFTHREADS: A lock-free thread library' by Gidenstam and | ||
5 | Papatriantafilou. The same technique is used in OSv, and in fact | ||
6 | the code is essentially a conversion to C of OSv's code. | ||
7 | |||
8 | [Added missing coroutine_fn in tests/test-aio-multithread.c. | ||
9 | --Stefan] | ||
10 | |||
11 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
12 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
13 | Message-id: 20170213181244.16297-2-pbonzini@redhat.com | ||
14 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
15 | --- | ||
16 | include/qemu/coroutine.h | 17 ++++- | ||
17 | tests/test-aio-multithread.c | 86 ++++++++++++++++++++++++ | ||
18 | util/qemu-coroutine-lock.c | 155 ++++++++++++++++++++++++++++++++++++++++--- | ||
19 | util/trace-events | 1 + | ||
20 | 4 files changed, 246 insertions(+), 13 deletions(-) | ||
21 | |||
22 | diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h | ||
23 | index XXXXXXX..XXXXXXX 100644 | ||
24 | --- a/include/qemu/coroutine.h | ||
25 | +++ b/include/qemu/coroutine.h | ||
26 | @@ -XXX,XX +XXX,XX @@ bool qemu_co_queue_empty(CoQueue *queue); | ||
27 | /** | ||
28 | * Provides a mutex that can be used to synchronise coroutines | ||
29 | */ | ||
30 | +struct CoWaitRecord; | ||
31 | typedef struct CoMutex { | ||
32 | - bool locked; | ||
33 | + /* Count of pending lockers; 0 for a free mutex, 1 for an | ||
34 | + * uncontended mutex. | ||
35 | + */ | ||
36 | + unsigned locked; | ||
37 | + | ||
38 | + /* A queue of waiters. Elements are added atomically in front of | ||
39 | + * from_push. to_pop is only populated, and popped from, by whoever | ||
40 | + * is in charge of the next wakeup. This can be an unlocker or, | ||
41 | + * through the handoff protocol, a locker that is about to go to sleep. | ||
42 | + */ | ||
43 | + QSLIST_HEAD(, CoWaitRecord) from_push, to_pop; | ||
44 | + | ||
45 | + unsigned handoff, sequence; | ||
46 | + | ||
47 | Coroutine *holder; | ||
48 | - CoQueue queue; | ||
49 | } CoMutex; | ||
50 | |||
51 | /** | ||
52 | diff --git a/tests/test-aio-multithread.c b/tests/test-aio-multithread.c | ||
53 | index XXXXXXX..XXXXXXX 100644 | ||
54 | --- a/tests/test-aio-multithread.c | ||
55 | +++ b/tests/test-aio-multithread.c | ||
56 | @@ -XXX,XX +XXX,XX @@ static void test_multi_co_schedule_10(void) | ||
57 | test_multi_co_schedule(10); | ||
58 | } | ||
59 | |||
60 | +/* CoMutex thread-safety. */ | ||
61 | + | ||
62 | +static uint32_t atomic_counter; | ||
63 | +static uint32_t running; | ||
64 | +static uint32_t counter; | ||
65 | +static CoMutex comutex; | ||
66 | + | ||
67 | +static void coroutine_fn test_multi_co_mutex_entry(void *opaque) | ||
68 | +{ | ||
69 | + while (!atomic_mb_read(&now_stopping)) { | ||
70 | + qemu_co_mutex_lock(&comutex); | ||
71 | + counter++; | ||
72 | + qemu_co_mutex_unlock(&comutex); | ||
73 | + | ||
74 | + /* Increase atomic_counter *after* releasing the mutex. Otherwise | ||
75 | + * there is a chance (it happens about 1 in 3 runs) that the iothread | ||
76 | + * exits before the coroutine is woken up, causing a spurious | ||
77 | + * assertion failure. | ||
78 | + */ | ||
79 | + atomic_inc(&atomic_counter); | ||
80 | + } | ||
81 | + atomic_dec(&running); | ||
82 | +} | ||
83 | + | ||
84 | +static void test_multi_co_mutex(int threads, int seconds) | ||
85 | +{ | ||
86 | + int i; | ||
87 | + | ||
88 | + qemu_co_mutex_init(&comutex); | ||
89 | + counter = 0; | ||
90 | + atomic_counter = 0; | ||
91 | + now_stopping = false; | ||
92 | + | ||
93 | + create_aio_contexts(); | ||
94 | + assert(threads <= NUM_CONTEXTS); | ||
95 | + running = threads; | ||
96 | + for (i = 0; i < threads; i++) { | ||
97 | + Coroutine *co1 = qemu_coroutine_create(test_multi_co_mutex_entry, NULL); | ||
98 | + aio_co_schedule(ctx[i], co1); | ||
99 | + } | ||
100 | + | ||
101 | + g_usleep(seconds * 1000000); | ||
102 | + | ||
103 | + atomic_mb_set(&now_stopping, true); | ||
104 | + while (running > 0) { | ||
105 | + g_usleep(100000); | ||
106 | + } | ||
107 | + | ||
108 | + join_aio_contexts(); | ||
109 | + g_test_message("%d iterations/second\n", counter / seconds); | ||
110 | + g_assert_cmpint(counter, ==, atomic_counter); | ||
111 | +} | ||
112 | + | ||
113 | +/* Testing with NUM_CONTEXTS threads focuses on the queue. The mutex however | ||
114 | + * is too contended (and the threads spend too much time in aio_poll) | ||
115 | + * to actually stress the handoff protocol. | ||
116 | + */ | ||
117 | +static void test_multi_co_mutex_1(void) | ||
118 | +{ | ||
119 | + test_multi_co_mutex(NUM_CONTEXTS, 1); | ||
120 | +} | ||
121 | + | ||
122 | +static void test_multi_co_mutex_10(void) | ||
123 | +{ | ||
124 | + test_multi_co_mutex(NUM_CONTEXTS, 10); | ||
125 | +} | ||
126 | + | ||
127 | +/* Testing with fewer threads stresses the handoff protocol too. Still, the | ||
128 | + * case where the locker _can_ pick up a handoff is very rare, happening | ||
129 | + * about 10 times in 1 million, so increase the runtime a bit compared to | ||
130 | + * other "quick" testcases that only run for 1 second. | ||
131 | + */ | ||
132 | +static void test_multi_co_mutex_2_3(void) | ||
133 | +{ | ||
134 | + test_multi_co_mutex(2, 3); | ||
135 | +} | ||
136 | + | ||
137 | +static void test_multi_co_mutex_2_30(void) | ||
138 | +{ | ||
139 | + test_multi_co_mutex(2, 30); | ||
140 | +} | ||
141 | + | ||
142 | /* End of tests. */ | ||
143 | |||
144 | int main(int argc, char **argv) | ||
145 | @@ -XXX,XX +XXX,XX @@ int main(int argc, char **argv) | ||
146 | g_test_add_func("/aio/multi/lifecycle", test_lifecycle); | ||
147 | if (g_test_quick()) { | ||
148 | g_test_add_func("/aio/multi/schedule", test_multi_co_schedule_1); | ||
149 | + g_test_add_func("/aio/multi/mutex/contended", test_multi_co_mutex_1); | ||
150 | + g_test_add_func("/aio/multi/mutex/handoff", test_multi_co_mutex_2_3); | ||
151 | } else { | ||
152 | g_test_add_func("/aio/multi/schedule", test_multi_co_schedule_10); | ||
153 | + g_test_add_func("/aio/multi/mutex/contended", test_multi_co_mutex_10); | ||
154 | + g_test_add_func("/aio/multi/mutex/handoff", test_multi_co_mutex_2_30); | ||
155 | } | ||
156 | return g_test_run(); | ||
157 | } | ||
158 | diff --git a/util/qemu-coroutine-lock.c b/util/qemu-coroutine-lock.c | ||
159 | index XXXXXXX..XXXXXXX 100644 | ||
160 | --- a/util/qemu-coroutine-lock.c | ||
161 | +++ b/util/qemu-coroutine-lock.c | ||
162 | @@ -XXX,XX +XXX,XX @@ | ||
163 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, | ||
164 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN | ||
165 | * THE SOFTWARE. | ||
166 | + * | ||
167 | + * The lock-free mutex implementation is based on OSv | ||
168 | + * (core/lfmutex.cc, include/lockfree/mutex.hh). | ||
169 | + * Copyright (C) 2013 Cloudius Systems, Ltd. | ||
170 | */ | ||
171 | |||
172 | #include "qemu/osdep.h" | ||
173 | @@ -XXX,XX +XXX,XX @@ bool qemu_co_queue_empty(CoQueue *queue) | ||
174 | return QSIMPLEQ_FIRST(&queue->entries) == NULL; | ||
175 | } | ||
176 | |||
177 | +/* The wait records are handled with a multiple-producer, single-consumer | ||
178 | + * lock-free queue. There cannot be two concurrent pop_waiter() calls | ||
179 | + * because pop_waiter() can only be called while mutex->handoff is zero. | ||
180 | + * This can happen in three cases: | ||
181 | + * - in qemu_co_mutex_unlock, before the hand-off protocol has started. | ||
182 | + * In this case, qemu_co_mutex_lock will see mutex->handoff == 0 and | ||
183 | + * not take part in the handoff. | ||
184 | + * - in qemu_co_mutex_lock, if it steals the hand-off responsibility from | ||
185 | + * qemu_co_mutex_unlock. In this case, qemu_co_mutex_unlock will fail | ||
186 | + * the cmpxchg (it will see either 0 or the next sequence value) and | ||
187 | + * exit. The next hand-off cannot begin until qemu_co_mutex_lock has | ||
188 | + * woken up someone. | ||
189 | + * - in qemu_co_mutex_unlock, if it takes the hand-off token itself. | ||
190 | + * In this case another iteration starts with mutex->handoff == 0; | ||
191 | + * a concurrent qemu_co_mutex_lock will fail the cmpxchg, and | ||
192 | + * qemu_co_mutex_unlock will go back to case (1). | ||
193 | + * | ||
194 | + * The following functions manage this queue. | ||
195 | + */ | ||
196 | +typedef struct CoWaitRecord { | ||
197 | + Coroutine *co; | ||
198 | + QSLIST_ENTRY(CoWaitRecord) next; | ||
199 | +} CoWaitRecord; | ||
200 | + | ||
201 | +static void push_waiter(CoMutex *mutex, CoWaitRecord *w) | ||
202 | +{ | ||
203 | + w->co = qemu_coroutine_self(); | ||
204 | + QSLIST_INSERT_HEAD_ATOMIC(&mutex->from_push, w, next); | ||
205 | +} | ||
206 | + | ||
207 | +static void move_waiters(CoMutex *mutex) | ||
208 | +{ | ||
209 | + QSLIST_HEAD(, CoWaitRecord) reversed; | ||
210 | + QSLIST_MOVE_ATOMIC(&reversed, &mutex->from_push); | ||
211 | + while (!QSLIST_EMPTY(&reversed)) { | ||
212 | + CoWaitRecord *w = QSLIST_FIRST(&reversed); | ||
213 | + QSLIST_REMOVE_HEAD(&reversed, next); | ||
214 | + QSLIST_INSERT_HEAD(&mutex->to_pop, w, next); | ||
215 | + } | ||
216 | +} | ||
217 | + | ||
218 | +static CoWaitRecord *pop_waiter(CoMutex *mutex) | ||
219 | +{ | ||
220 | + CoWaitRecord *w; | ||
221 | + | ||
222 | + if (QSLIST_EMPTY(&mutex->to_pop)) { | ||
223 | + move_waiters(mutex); | ||
224 | + if (QSLIST_EMPTY(&mutex->to_pop)) { | ||
225 | + return NULL; | ||
226 | + } | ||
227 | + } | ||
228 | + w = QSLIST_FIRST(&mutex->to_pop); | ||
229 | + QSLIST_REMOVE_HEAD(&mutex->to_pop, next); | ||
230 | + return w; | ||
231 | +} | ||
232 | + | ||
233 | +static bool has_waiters(CoMutex *mutex) | ||
234 | +{ | ||
235 | + return QSLIST_EMPTY(&mutex->to_pop) || QSLIST_EMPTY(&mutex->from_push); | ||
236 | +} | ||
237 | + | ||
238 | void qemu_co_mutex_init(CoMutex *mutex) | ||
239 | { | ||
240 | memset(mutex, 0, sizeof(*mutex)); | ||
241 | - qemu_co_queue_init(&mutex->queue); | ||
242 | } | ||
243 | |||
244 | -void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex) | ||
245 | +static void coroutine_fn qemu_co_mutex_lock_slowpath(CoMutex *mutex) | ||
246 | { | ||
247 | Coroutine *self = qemu_coroutine_self(); | ||
248 | + CoWaitRecord w; | ||
249 | + unsigned old_handoff; | ||
250 | |||
251 | trace_qemu_co_mutex_lock_entry(mutex, self); | ||
252 | + w.co = self; | ||
253 | + push_waiter(mutex, &w); | ||
254 | |||
255 | - while (mutex->locked) { | ||
256 | - qemu_co_queue_wait(&mutex->queue); | ||
257 | + /* This is the "Responsibility Hand-Off" protocol; a lock() picks from | ||
258 | + * a concurrent unlock() the responsibility of waking somebody up. | ||
259 | + */ | ||
260 | + old_handoff = atomic_mb_read(&mutex->handoff); | ||
261 | + if (old_handoff && | ||
262 | + has_waiters(mutex) && | ||
263 | + atomic_cmpxchg(&mutex->handoff, old_handoff, 0) == old_handoff) { | ||
264 | + /* There can be no concurrent pops, because there can be only | ||
265 | + * one active handoff at a time. | ||
266 | + */ | ||
267 | + CoWaitRecord *to_wake = pop_waiter(mutex); | ||
268 | + Coroutine *co = to_wake->co; | ||
269 | + if (co == self) { | ||
270 | + /* We got the lock ourselves! */ | ||
271 | + assert(to_wake == &w); | ||
272 | + return; | ||
273 | + } | ||
274 | + | ||
275 | + aio_co_wake(co); | ||
276 | } | ||
277 | |||
278 | - mutex->locked = true; | ||
279 | - mutex->holder = self; | ||
280 | - self->locks_held++; | ||
281 | - | ||
282 | + qemu_coroutine_yield(); | ||
283 | trace_qemu_co_mutex_lock_return(mutex, self); | ||
284 | } | ||
285 | |||
286 | +void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex) | ||
287 | +{ | ||
288 | + Coroutine *self = qemu_coroutine_self(); | ||
289 | + | ||
290 | + if (atomic_fetch_inc(&mutex->locked) == 0) { | ||
291 | + /* Uncontended. */ | ||
292 | + trace_qemu_co_mutex_lock_uncontended(mutex, self); | ||
293 | + } else { | ||
294 | + qemu_co_mutex_lock_slowpath(mutex); | ||
295 | + } | ||
296 | + mutex->holder = self; | ||
297 | + self->locks_held++; | ||
298 | +} | ||
299 | + | ||
300 | void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex) | ||
301 | { | ||
302 | Coroutine *self = qemu_coroutine_self(); | ||
303 | |||
304 | trace_qemu_co_mutex_unlock_entry(mutex, self); | ||
305 | |||
306 | - assert(mutex->locked == true); | ||
307 | + assert(mutex->locked); | ||
308 | assert(mutex->holder == self); | ||
309 | assert(qemu_in_coroutine()); | ||
310 | |||
311 | - mutex->locked = false; | ||
312 | mutex->holder = NULL; | ||
313 | self->locks_held--; | ||
314 | - qemu_co_queue_next(&mutex->queue); | ||
315 | + if (atomic_fetch_dec(&mutex->locked) == 1) { | ||
316 | + /* No waiting qemu_co_mutex_lock(). Pfew, that was easy! */ | ||
317 | + return; | ||
318 | + } | ||
319 | + | ||
320 | + for (;;) { | ||
321 | + CoWaitRecord *to_wake = pop_waiter(mutex); | ||
322 | + unsigned our_handoff; | ||
323 | + | ||
324 | + if (to_wake) { | ||
325 | + Coroutine *co = to_wake->co; | ||
326 | + aio_co_wake(co); | ||
327 | + break; | ||
328 | + } | ||
329 | + | ||
330 | + /* Some concurrent lock() is in progress (we know this because | ||
331 | + * mutex->locked was >1) but it hasn't yet put itself on the wait | ||
332 | + * queue. Pick a sequence number for the handoff protocol (not 0). | ||
333 | + */ | ||
334 | + if (++mutex->sequence == 0) { | ||
335 | + mutex->sequence = 1; | ||
336 | + } | ||
337 | + | ||
338 | + our_handoff = mutex->sequence; | ||
339 | + atomic_mb_set(&mutex->handoff, our_handoff); | ||
340 | + if (!has_waiters(mutex)) { | ||
341 | + /* The concurrent lock has not added itself yet, so it | ||
342 | + * will be able to pick our handoff. | ||
343 | + */ | ||
344 | + break; | ||
345 | + } | ||
346 | + | ||
347 | + /* Try to do the handoff protocol ourselves; if somebody else has | ||
348 | + * already taken it, however, we're done and they're responsible. | ||
349 | + */ | ||
350 | + if (atomic_cmpxchg(&mutex->handoff, our_handoff, 0) != our_handoff) { | ||
351 | + break; | ||
352 | + } | ||
353 | + } | ||
354 | |||
355 | trace_qemu_co_mutex_unlock_return(mutex, self); | ||
356 | } | ||
357 | diff --git a/util/trace-events b/util/trace-events | ||
358 | index XXXXXXX..XXXXXXX 100644 | ||
359 | --- a/util/trace-events | ||
360 | +++ b/util/trace-events | ||
361 | @@ -XXX,XX +XXX,XX @@ qemu_coroutine_terminate(void *co) "self %p" | ||
362 | |||
363 | # util/qemu-coroutine-lock.c | ||
364 | qemu_co_queue_run_restart(void *co) "co %p" | ||
365 | +qemu_co_mutex_lock_uncontended(void *mutex, void *self) "mutex %p self %p" | ||
366 | qemu_co_mutex_lock_entry(void *mutex, void *self) "mutex %p self %p" | ||
367 | qemu_co_mutex_lock_return(void *mutex, void *self) "mutex %p self %p" | ||
368 | qemu_co_mutex_unlock_entry(void *mutex, void *self) "mutex %p self %p" | ||
369 | -- | ||
370 | 2.9.3 | ||
371 | |||
372 | diff view generated by jsdifflib |
1 | From: Stefano Garzarella <sgarzare@redhat.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | The size of data in the virtio_blk_request must be a multiple | 3 | Running a very small critical section on pthread_mutex_t and CoMutex |
4 | of 512 bytes for IN and OUT requests, or a multiple of the size | 4 | shows that pthread_mutex_t is much faster because it doesn't actually |
5 | of struct virtio_blk_discard_write_zeroes for DISCARD and | 5 | go to sleep. What happens is that the critical section is shorter |
6 | WRITE_ZEROES requests. | 6 | than the latency of entering the kernel and thus FUTEX_WAIT always |
7 | fails. With CoMutex there is no such latency but you still want to | ||
8 | avoid wait and wakeup. So introduce it artificially. | ||
7 | 9 | ||
8 | Reviewed-by: Michael S. Tsirkin <mst@redhat.com> | 10 | This only works with one waiters; because CoMutex is fair, it will |
9 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | 11 | always have more waits and wakeups than a pthread_mutex_t. |
10 | Reviewed-by: Thomas Huth <thuth@redhat.com> | 12 | |
11 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | 13 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
12 | Acked-by: Pankaj Gupta <pagupta@redhat.com> | 14 | Reviewed-by: Fam Zheng <famz@redhat.com> |
13 | Message-id: 20190208134950.187665-6-sgarzare@redhat.com | 15 | Message-id: 20170213181244.16297-3-pbonzini@redhat.com |
14 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 16 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
15 | --- | 17 | --- |
16 | tests/virtio-blk-test.c | 15 ++++++++++++++- | 18 | include/qemu/coroutine.h | 5 +++++ |
17 | 1 file changed, 14 insertions(+), 1 deletion(-) | 19 | util/qemu-coroutine-lock.c | 51 ++++++++++++++++++++++++++++++++++++++++------ |
20 | util/qemu-coroutine.c | 2 +- | ||
21 | 3 files changed, 51 insertions(+), 7 deletions(-) | ||
18 | 22 | ||
19 | diff --git a/tests/virtio-blk-test.c b/tests/virtio-blk-test.c | 23 | diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h |
20 | index XXXXXXX..XXXXXXX 100644 | 24 | index XXXXXXX..XXXXXXX 100644 |
21 | --- a/tests/virtio-blk-test.c | 25 | --- a/include/qemu/coroutine.h |
22 | +++ b/tests/virtio-blk-test.c | 26 | +++ b/include/qemu/coroutine.h |
23 | @@ -XXX,XX +XXX,XX @@ static uint64_t virtio_blk_request(QGuestAllocator *alloc, QVirtioDevice *d, | 27 | @@ -XXX,XX +XXX,XX @@ typedef struct CoMutex { |
24 | uint64_t addr; | 28 | */ |
25 | uint8_t status = 0xFF; | 29 | unsigned locked; |
26 | 30 | ||
27 | - g_assert_cmpuint(data_size % 512, ==, 0); | 31 | + /* Context that is holding the lock. Useful to avoid spinning |
28 | + switch (req->type) { | 32 | + * when two coroutines on the same AioContext try to get the lock. :) |
29 | + case VIRTIO_BLK_T_IN: | 33 | + */ |
30 | + case VIRTIO_BLK_T_OUT: | 34 | + AioContext *ctx; |
31 | + g_assert_cmpuint(data_size % 512, ==, 0); | 35 | + |
32 | + break; | 36 | /* A queue of waiters. Elements are added atomically in front of |
33 | + case VIRTIO_BLK_T_DISCARD: | 37 | * from_push. to_pop is only populated, and popped from, by whoever |
34 | + case VIRTIO_BLK_T_WRITE_ZEROES: | 38 | * is in charge of the next wakeup. This can be an unlocker or, |
35 | + g_assert_cmpuint(data_size % | 39 | diff --git a/util/qemu-coroutine-lock.c b/util/qemu-coroutine-lock.c |
36 | + sizeof(struct virtio_blk_discard_write_zeroes), ==, 0); | 40 | index XXXXXXX..XXXXXXX 100644 |
37 | + break; | 41 | --- a/util/qemu-coroutine-lock.c |
38 | + default: | 42 | +++ b/util/qemu-coroutine-lock.c |
39 | + g_assert_cmpuint(data_size, ==, 0); | 43 | @@ -XXX,XX +XXX,XX @@ |
44 | #include "qemu-common.h" | ||
45 | #include "qemu/coroutine.h" | ||
46 | #include "qemu/coroutine_int.h" | ||
47 | +#include "qemu/processor.h" | ||
48 | #include "qemu/queue.h" | ||
49 | #include "block/aio.h" | ||
50 | #include "trace.h" | ||
51 | @@ -XXX,XX +XXX,XX @@ void qemu_co_mutex_init(CoMutex *mutex) | ||
52 | memset(mutex, 0, sizeof(*mutex)); | ||
53 | } | ||
54 | |||
55 | -static void coroutine_fn qemu_co_mutex_lock_slowpath(CoMutex *mutex) | ||
56 | +static void coroutine_fn qemu_co_mutex_wake(CoMutex *mutex, Coroutine *co) | ||
57 | +{ | ||
58 | + /* Read co before co->ctx; pairs with smp_wmb() in | ||
59 | + * qemu_coroutine_enter(). | ||
60 | + */ | ||
61 | + smp_read_barrier_depends(); | ||
62 | + mutex->ctx = co->ctx; | ||
63 | + aio_co_wake(co); | ||
64 | +} | ||
65 | + | ||
66 | +static void coroutine_fn qemu_co_mutex_lock_slowpath(AioContext *ctx, | ||
67 | + CoMutex *mutex) | ||
68 | { | ||
69 | Coroutine *self = qemu_coroutine_self(); | ||
70 | CoWaitRecord w; | ||
71 | @@ -XXX,XX +XXX,XX @@ static void coroutine_fn qemu_co_mutex_lock_slowpath(CoMutex *mutex) | ||
72 | if (co == self) { | ||
73 | /* We got the lock ourselves! */ | ||
74 | assert(to_wake == &w); | ||
75 | + mutex->ctx = ctx; | ||
76 | return; | ||
77 | } | ||
78 | |||
79 | - aio_co_wake(co); | ||
80 | + qemu_co_mutex_wake(mutex, co); | ||
81 | } | ||
82 | |||
83 | qemu_coroutine_yield(); | ||
84 | @@ -XXX,XX +XXX,XX @@ static void coroutine_fn qemu_co_mutex_lock_slowpath(CoMutex *mutex) | ||
85 | |||
86 | void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex) | ||
87 | { | ||
88 | + AioContext *ctx = qemu_get_current_aio_context(); | ||
89 | Coroutine *self = qemu_coroutine_self(); | ||
90 | + int waiters, i; | ||
91 | |||
92 | - if (atomic_fetch_inc(&mutex->locked) == 0) { | ||
93 | + /* Running a very small critical section on pthread_mutex_t and CoMutex | ||
94 | + * shows that pthread_mutex_t is much faster because it doesn't actually | ||
95 | + * go to sleep. What happens is that the critical section is shorter | ||
96 | + * than the latency of entering the kernel and thus FUTEX_WAIT always | ||
97 | + * fails. With CoMutex there is no such latency but you still want to | ||
98 | + * avoid wait and wakeup. So introduce it artificially. | ||
99 | + */ | ||
100 | + i = 0; | ||
101 | +retry_fast_path: | ||
102 | + waiters = atomic_cmpxchg(&mutex->locked, 0, 1); | ||
103 | + if (waiters != 0) { | ||
104 | + while (waiters == 1 && ++i < 1000) { | ||
105 | + if (atomic_read(&mutex->ctx) == ctx) { | ||
106 | + break; | ||
107 | + } | ||
108 | + if (atomic_read(&mutex->locked) == 0) { | ||
109 | + goto retry_fast_path; | ||
110 | + } | ||
111 | + cpu_relax(); | ||
112 | + } | ||
113 | + waiters = atomic_fetch_inc(&mutex->locked); | ||
40 | + } | 114 | + } |
41 | + | 115 | + |
42 | addr = guest_alloc(alloc, sizeof(*req) + data_size); | 116 | + if (waiters == 0) { |
43 | 117 | /* Uncontended. */ | |
44 | virtio_blk_fix_request(d, req); | 118 | trace_qemu_co_mutex_lock_uncontended(mutex, self); |
119 | + mutex->ctx = ctx; | ||
120 | } else { | ||
121 | - qemu_co_mutex_lock_slowpath(mutex); | ||
122 | + qemu_co_mutex_lock_slowpath(ctx, mutex); | ||
123 | } | ||
124 | mutex->holder = self; | ||
125 | self->locks_held++; | ||
126 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex) | ||
127 | assert(mutex->holder == self); | ||
128 | assert(qemu_in_coroutine()); | ||
129 | |||
130 | + mutex->ctx = NULL; | ||
131 | mutex->holder = NULL; | ||
132 | self->locks_held--; | ||
133 | if (atomic_fetch_dec(&mutex->locked) == 1) { | ||
134 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex) | ||
135 | unsigned our_handoff; | ||
136 | |||
137 | if (to_wake) { | ||
138 | - Coroutine *co = to_wake->co; | ||
139 | - aio_co_wake(co); | ||
140 | + qemu_co_mutex_wake(mutex, to_wake->co); | ||
141 | break; | ||
142 | } | ||
143 | |||
144 | diff --git a/util/qemu-coroutine.c b/util/qemu-coroutine.c | ||
145 | index XXXXXXX..XXXXXXX 100644 | ||
146 | --- a/util/qemu-coroutine.c | ||
147 | +++ b/util/qemu-coroutine.c | ||
148 | @@ -XXX,XX +XXX,XX @@ void qemu_coroutine_enter(Coroutine *co) | ||
149 | co->ctx = qemu_get_current_aio_context(); | ||
150 | |||
151 | /* Store co->ctx before anything that stores co. Matches | ||
152 | - * barrier in aio_co_wake. | ||
153 | + * barrier in aio_co_wake and qemu_co_mutex_wake. | ||
154 | */ | ||
155 | smp_wmb(); | ||
156 | |||
45 | -- | 157 | -- |
46 | 2.20.1 | 158 | 2.9.3 |
47 | 159 | ||
48 | 160 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | 1 | From: Paolo Bonzini <pbonzini@redhat.com> | |
2 | |||
3 | Add two implementations of the same benchmark as the previous patch, | ||
4 | but using pthreads. One uses a normal QemuMutex, the other is Linux | ||
5 | only and implements a fair mutex based on MCS locks and futexes. | ||
6 | This shows that the slower performance of the 5-thread case is due to | ||
7 | the fairness of CoMutex, rather than to coroutines. If fairness does | ||
8 | not matter, as is the case with two threads, CoMutex can actually be | ||
9 | faster than pthreads. | ||
10 | |||
11 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
12 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
13 | Message-id: 20170213181244.16297-4-pbonzini@redhat.com | ||
14 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
15 | --- | ||
16 | tests/test-aio-multithread.c | 164 +++++++++++++++++++++++++++++++++++++++++++ | ||
17 | 1 file changed, 164 insertions(+) | ||
18 | |||
19 | diff --git a/tests/test-aio-multithread.c b/tests/test-aio-multithread.c | ||
20 | index XXXXXXX..XXXXXXX 100644 | ||
21 | --- a/tests/test-aio-multithread.c | ||
22 | +++ b/tests/test-aio-multithread.c | ||
23 | @@ -XXX,XX +XXX,XX @@ static void test_multi_co_mutex_2_30(void) | ||
24 | test_multi_co_mutex(2, 30); | ||
25 | } | ||
26 | |||
27 | +/* Same test with fair mutexes, for performance comparison. */ | ||
28 | + | ||
29 | +#ifdef CONFIG_LINUX | ||
30 | +#include "qemu/futex.h" | ||
31 | + | ||
32 | +/* The nodes for the mutex reside in this structure (on which we try to avoid | ||
33 | + * false sharing). The head of the mutex is in the "mutex_head" variable. | ||
34 | + */ | ||
35 | +static struct { | ||
36 | + int next, locked; | ||
37 | + int padding[14]; | ||
38 | +} nodes[NUM_CONTEXTS] __attribute__((__aligned__(64))); | ||
39 | + | ||
40 | +static int mutex_head = -1; | ||
41 | + | ||
42 | +static void mcs_mutex_lock(void) | ||
43 | +{ | ||
44 | + int prev; | ||
45 | + | ||
46 | + nodes[id].next = -1; | ||
47 | + nodes[id].locked = 1; | ||
48 | + prev = atomic_xchg(&mutex_head, id); | ||
49 | + if (prev != -1) { | ||
50 | + atomic_set(&nodes[prev].next, id); | ||
51 | + qemu_futex_wait(&nodes[id].locked, 1); | ||
52 | + } | ||
53 | +} | ||
54 | + | ||
55 | +static void mcs_mutex_unlock(void) | ||
56 | +{ | ||
57 | + int next; | ||
58 | + if (nodes[id].next == -1) { | ||
59 | + if (atomic_read(&mutex_head) == id && | ||
60 | + atomic_cmpxchg(&mutex_head, id, -1) == id) { | ||
61 | + /* Last item in the list, exit. */ | ||
62 | + return; | ||
63 | + } | ||
64 | + while (atomic_read(&nodes[id].next) == -1) { | ||
65 | + /* mcs_mutex_lock did the xchg, but has not updated | ||
66 | + * nodes[prev].next yet. | ||
67 | + */ | ||
68 | + } | ||
69 | + } | ||
70 | + | ||
71 | + /* Wake up the next in line. */ | ||
72 | + next = nodes[id].next; | ||
73 | + nodes[next].locked = 0; | ||
74 | + qemu_futex_wake(&nodes[next].locked, 1); | ||
75 | +} | ||
76 | + | ||
77 | +static void test_multi_fair_mutex_entry(void *opaque) | ||
78 | +{ | ||
79 | + while (!atomic_mb_read(&now_stopping)) { | ||
80 | + mcs_mutex_lock(); | ||
81 | + counter++; | ||
82 | + mcs_mutex_unlock(); | ||
83 | + atomic_inc(&atomic_counter); | ||
84 | + } | ||
85 | + atomic_dec(&running); | ||
86 | +} | ||
87 | + | ||
88 | +static void test_multi_fair_mutex(int threads, int seconds) | ||
89 | +{ | ||
90 | + int i; | ||
91 | + | ||
92 | + assert(mutex_head == -1); | ||
93 | + counter = 0; | ||
94 | + atomic_counter = 0; | ||
95 | + now_stopping = false; | ||
96 | + | ||
97 | + create_aio_contexts(); | ||
98 | + assert(threads <= NUM_CONTEXTS); | ||
99 | + running = threads; | ||
100 | + for (i = 0; i < threads; i++) { | ||
101 | + Coroutine *co1 = qemu_coroutine_create(test_multi_fair_mutex_entry, NULL); | ||
102 | + aio_co_schedule(ctx[i], co1); | ||
103 | + } | ||
104 | + | ||
105 | + g_usleep(seconds * 1000000); | ||
106 | + | ||
107 | + atomic_mb_set(&now_stopping, true); | ||
108 | + while (running > 0) { | ||
109 | + g_usleep(100000); | ||
110 | + } | ||
111 | + | ||
112 | + join_aio_contexts(); | ||
113 | + g_test_message("%d iterations/second\n", counter / seconds); | ||
114 | + g_assert_cmpint(counter, ==, atomic_counter); | ||
115 | +} | ||
116 | + | ||
117 | +static void test_multi_fair_mutex_1(void) | ||
118 | +{ | ||
119 | + test_multi_fair_mutex(NUM_CONTEXTS, 1); | ||
120 | +} | ||
121 | + | ||
122 | +static void test_multi_fair_mutex_10(void) | ||
123 | +{ | ||
124 | + test_multi_fair_mutex(NUM_CONTEXTS, 10); | ||
125 | +} | ||
126 | +#endif | ||
127 | + | ||
128 | +/* Same test with pthread mutexes, for performance comparison and | ||
129 | + * portability. */ | ||
130 | + | ||
131 | +static QemuMutex mutex; | ||
132 | + | ||
133 | +static void test_multi_mutex_entry(void *opaque) | ||
134 | +{ | ||
135 | + while (!atomic_mb_read(&now_stopping)) { | ||
136 | + qemu_mutex_lock(&mutex); | ||
137 | + counter++; | ||
138 | + qemu_mutex_unlock(&mutex); | ||
139 | + atomic_inc(&atomic_counter); | ||
140 | + } | ||
141 | + atomic_dec(&running); | ||
142 | +} | ||
143 | + | ||
144 | +static void test_multi_mutex(int threads, int seconds) | ||
145 | +{ | ||
146 | + int i; | ||
147 | + | ||
148 | + qemu_mutex_init(&mutex); | ||
149 | + counter = 0; | ||
150 | + atomic_counter = 0; | ||
151 | + now_stopping = false; | ||
152 | + | ||
153 | + create_aio_contexts(); | ||
154 | + assert(threads <= NUM_CONTEXTS); | ||
155 | + running = threads; | ||
156 | + for (i = 0; i < threads; i++) { | ||
157 | + Coroutine *co1 = qemu_coroutine_create(test_multi_mutex_entry, NULL); | ||
158 | + aio_co_schedule(ctx[i], co1); | ||
159 | + } | ||
160 | + | ||
161 | + g_usleep(seconds * 1000000); | ||
162 | + | ||
163 | + atomic_mb_set(&now_stopping, true); | ||
164 | + while (running > 0) { | ||
165 | + g_usleep(100000); | ||
166 | + } | ||
167 | + | ||
168 | + join_aio_contexts(); | ||
169 | + g_test_message("%d iterations/second\n", counter / seconds); | ||
170 | + g_assert_cmpint(counter, ==, atomic_counter); | ||
171 | +} | ||
172 | + | ||
173 | +static void test_multi_mutex_1(void) | ||
174 | +{ | ||
175 | + test_multi_mutex(NUM_CONTEXTS, 1); | ||
176 | +} | ||
177 | + | ||
178 | +static void test_multi_mutex_10(void) | ||
179 | +{ | ||
180 | + test_multi_mutex(NUM_CONTEXTS, 10); | ||
181 | +} | ||
182 | + | ||
183 | /* End of tests. */ | ||
184 | |||
185 | int main(int argc, char **argv) | ||
186 | @@ -XXX,XX +XXX,XX @@ int main(int argc, char **argv) | ||
187 | g_test_add_func("/aio/multi/schedule", test_multi_co_schedule_1); | ||
188 | g_test_add_func("/aio/multi/mutex/contended", test_multi_co_mutex_1); | ||
189 | g_test_add_func("/aio/multi/mutex/handoff", test_multi_co_mutex_2_3); | ||
190 | +#ifdef CONFIG_LINUX | ||
191 | + g_test_add_func("/aio/multi/mutex/mcs", test_multi_fair_mutex_1); | ||
192 | +#endif | ||
193 | + g_test_add_func("/aio/multi/mutex/pthread", test_multi_mutex_1); | ||
194 | } else { | ||
195 | g_test_add_func("/aio/multi/schedule", test_multi_co_schedule_10); | ||
196 | g_test_add_func("/aio/multi/mutex/contended", test_multi_co_mutex_10); | ||
197 | g_test_add_func("/aio/multi/mutex/handoff", test_multi_co_mutex_2_30); | ||
198 | +#ifdef CONFIG_LINUX | ||
199 | + g_test_add_func("/aio/multi/mutex/mcs", test_multi_fair_mutex_10); | ||
200 | +#endif | ||
201 | + g_test_add_func("/aio/multi/mutex/pthread", test_multi_mutex_10); | ||
202 | } | ||
203 | return g_test_run(); | ||
204 | } | ||
205 | -- | ||
206 | 2.9.3 | ||
207 | |||
208 | diff view generated by jsdifflib |
1 | From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | qemu coroutine command results in following error output: | 3 | This will avoid forward references in the next patch. It is also |
4 | more logical because CoQueue is not anymore the basic primitive. | ||
4 | 5 | ||
5 | Python Exception <class 'gdb.error'> 'arch_prctl' has unknown return | 6 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
6 | type; cast the call to its declared return type: Error occurred in | 7 | Reviewed-by: Fam Zheng <famz@redhat.com> |
7 | Python command: 'arch_prctl' has unknown return type; cast the call to | 8 | Message-id: 20170213181244.16297-5-pbonzini@redhat.com |
8 | its declared return type | ||
9 | |||
10 | Fix it by giving it what it wants: arch_prctl return type. | ||
11 | |||
12 | Information on the topic: | ||
13 | https://sourceware.org/gdb/onlinedocs/gdb/Calling.html | ||
14 | |||
15 | Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> | ||
16 | Message-id: 20190206151425.105871-1-vsementsov@virtuozzo.com | ||
17 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 9 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
18 | --- | 10 | --- |
19 | scripts/qemugdb/coroutine.py | 2 +- | 11 | include/qemu/coroutine.h | 89 ++++++++++++++++++++++++------------------------ |
20 | 1 file changed, 1 insertion(+), 1 deletion(-) | 12 | 1 file changed, 44 insertions(+), 45 deletions(-) |
21 | 13 | ||
22 | diff --git a/scripts/qemugdb/coroutine.py b/scripts/qemugdb/coroutine.py | 14 | diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h |
23 | index XXXXXXX..XXXXXXX 100644 | 15 | index XXXXXXX..XXXXXXX 100644 |
24 | --- a/scripts/qemugdb/coroutine.py | 16 | --- a/include/qemu/coroutine.h |
25 | +++ b/scripts/qemugdb/coroutine.py | 17 | +++ b/include/qemu/coroutine.h |
26 | @@ -XXX,XX +XXX,XX @@ def get_fs_base(): | 18 | @@ -XXX,XX +XXX,XX @@ bool qemu_in_coroutine(void); |
27 | pthread_self().''' | 19 | */ |
28 | # %rsp - 120 is scratch space according to the SystemV ABI | 20 | bool qemu_coroutine_entered(Coroutine *co); |
29 | old = gdb.parse_and_eval('*(uint64_t*)($rsp - 120)') | 21 | |
30 | - gdb.execute('call arch_prctl(0x1003, $rsp - 120)', False, True) | 22 | - |
31 | + gdb.execute('call (int)arch_prctl(0x1003, $rsp - 120)', False, True) | 23 | -/** |
32 | fs_base = gdb.parse_and_eval('*(uint64_t*)($rsp - 120)') | 24 | - * CoQueues are a mechanism to queue coroutines in order to continue executing |
33 | gdb.execute('set *(uint64_t*)($rsp - 120) = %s' % old, False, True) | 25 | - * them later. They provide the fundamental primitives on which coroutine locks |
34 | return fs_base | 26 | - * are built. |
27 | - */ | ||
28 | -typedef struct CoQueue { | ||
29 | - QSIMPLEQ_HEAD(, Coroutine) entries; | ||
30 | -} CoQueue; | ||
31 | - | ||
32 | -/** | ||
33 | - * Initialise a CoQueue. This must be called before any other operation is used | ||
34 | - * on the CoQueue. | ||
35 | - */ | ||
36 | -void qemu_co_queue_init(CoQueue *queue); | ||
37 | - | ||
38 | -/** | ||
39 | - * Adds the current coroutine to the CoQueue and transfers control to the | ||
40 | - * caller of the coroutine. | ||
41 | - */ | ||
42 | -void coroutine_fn qemu_co_queue_wait(CoQueue *queue); | ||
43 | - | ||
44 | -/** | ||
45 | - * Restarts the next coroutine in the CoQueue and removes it from the queue. | ||
46 | - * | ||
47 | - * Returns true if a coroutine was restarted, false if the queue is empty. | ||
48 | - */ | ||
49 | -bool coroutine_fn qemu_co_queue_next(CoQueue *queue); | ||
50 | - | ||
51 | -/** | ||
52 | - * Restarts all coroutines in the CoQueue and leaves the queue empty. | ||
53 | - */ | ||
54 | -void coroutine_fn qemu_co_queue_restart_all(CoQueue *queue); | ||
55 | - | ||
56 | -/** | ||
57 | - * Enter the next coroutine in the queue | ||
58 | - */ | ||
59 | -bool qemu_co_enter_next(CoQueue *queue); | ||
60 | - | ||
61 | -/** | ||
62 | - * Checks if the CoQueue is empty. | ||
63 | - */ | ||
64 | -bool qemu_co_queue_empty(CoQueue *queue); | ||
65 | - | ||
66 | - | ||
67 | /** | ||
68 | * Provides a mutex that can be used to synchronise coroutines | ||
69 | */ | ||
70 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex); | ||
71 | */ | ||
72 | void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex); | ||
73 | |||
74 | + | ||
75 | +/** | ||
76 | + * CoQueues are a mechanism to queue coroutines in order to continue executing | ||
77 | + * them later. | ||
78 | + */ | ||
79 | +typedef struct CoQueue { | ||
80 | + QSIMPLEQ_HEAD(, Coroutine) entries; | ||
81 | +} CoQueue; | ||
82 | + | ||
83 | +/** | ||
84 | + * Initialise a CoQueue. This must be called before any other operation is used | ||
85 | + * on the CoQueue. | ||
86 | + */ | ||
87 | +void qemu_co_queue_init(CoQueue *queue); | ||
88 | + | ||
89 | +/** | ||
90 | + * Adds the current coroutine to the CoQueue and transfers control to the | ||
91 | + * caller of the coroutine. | ||
92 | + */ | ||
93 | +void coroutine_fn qemu_co_queue_wait(CoQueue *queue); | ||
94 | + | ||
95 | +/** | ||
96 | + * Restarts the next coroutine in the CoQueue and removes it from the queue. | ||
97 | + * | ||
98 | + * Returns true if a coroutine was restarted, false if the queue is empty. | ||
99 | + */ | ||
100 | +bool coroutine_fn qemu_co_queue_next(CoQueue *queue); | ||
101 | + | ||
102 | +/** | ||
103 | + * Restarts all coroutines in the CoQueue and leaves the queue empty. | ||
104 | + */ | ||
105 | +void coroutine_fn qemu_co_queue_restart_all(CoQueue *queue); | ||
106 | + | ||
107 | +/** | ||
108 | + * Enter the next coroutine in the queue | ||
109 | + */ | ||
110 | +bool qemu_co_enter_next(CoQueue *queue); | ||
111 | + | ||
112 | +/** | ||
113 | + * Checks if the CoQueue is empty. | ||
114 | + */ | ||
115 | +bool qemu_co_queue_empty(CoQueue *queue); | ||
116 | + | ||
117 | + | ||
118 | typedef struct CoRwlock { | ||
119 | bool writer; | ||
120 | int reader; | ||
35 | -- | 121 | -- |
36 | 2.20.1 | 122 | 2.9.3 |
37 | 123 | ||
38 | 124 | diff view generated by jsdifflib |
New patch | |||
---|---|---|---|
1 | 1 | From: Paolo Bonzini <pbonzini@redhat.com> | |
2 | |||
3 | All that CoQueue needs in order to become thread-safe is help | ||
4 | from an external mutex. Add this to the API. | ||
5 | |||
6 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
7 | Reviewed-by: Fam Zheng <famz@redhat.com> | ||
8 | Message-id: 20170213181244.16297-6-pbonzini@redhat.com | ||
9 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
10 | --- | ||
11 | include/qemu/coroutine.h | 8 +++++--- | ||
12 | block/backup.c | 2 +- | ||
13 | block/io.c | 4 ++-- | ||
14 | block/nbd-client.c | 2 +- | ||
15 | block/qcow2-cluster.c | 4 +--- | ||
16 | block/sheepdog.c | 2 +- | ||
17 | block/throttle-groups.c | 2 +- | ||
18 | hw/9pfs/9p.c | 2 +- | ||
19 | util/qemu-coroutine-lock.c | 24 +++++++++++++++++++++--- | ||
20 | 9 files changed, 34 insertions(+), 16 deletions(-) | ||
21 | |||
22 | diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h | ||
23 | index XXXXXXX..XXXXXXX 100644 | ||
24 | --- a/include/qemu/coroutine.h | ||
25 | +++ b/include/qemu/coroutine.h | ||
26 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex); | ||
27 | |||
28 | /** | ||
29 | * CoQueues are a mechanism to queue coroutines in order to continue executing | ||
30 | - * them later. | ||
31 | + * them later. They are similar to condition variables, but they need help | ||
32 | + * from an external mutex in order to maintain thread-safety. | ||
33 | */ | ||
34 | typedef struct CoQueue { | ||
35 | QSIMPLEQ_HEAD(, Coroutine) entries; | ||
36 | @@ -XXX,XX +XXX,XX @@ void qemu_co_queue_init(CoQueue *queue); | ||
37 | |||
38 | /** | ||
39 | * Adds the current coroutine to the CoQueue and transfers control to the | ||
40 | - * caller of the coroutine. | ||
41 | + * caller of the coroutine. The mutex is unlocked during the wait and | ||
42 | + * locked again afterwards. | ||
43 | */ | ||
44 | -void coroutine_fn qemu_co_queue_wait(CoQueue *queue); | ||
45 | +void coroutine_fn qemu_co_queue_wait(CoQueue *queue, CoMutex *mutex); | ||
46 | |||
47 | /** | ||
48 | * Restarts the next coroutine in the CoQueue and removes it from the queue. | ||
49 | diff --git a/block/backup.c b/block/backup.c | ||
50 | index XXXXXXX..XXXXXXX 100644 | ||
51 | --- a/block/backup.c | ||
52 | +++ b/block/backup.c | ||
53 | @@ -XXX,XX +XXX,XX @@ static void coroutine_fn wait_for_overlapping_requests(BackupBlockJob *job, | ||
54 | retry = false; | ||
55 | QLIST_FOREACH(req, &job->inflight_reqs, list) { | ||
56 | if (end > req->start && start < req->end) { | ||
57 | - qemu_co_queue_wait(&req->wait_queue); | ||
58 | + qemu_co_queue_wait(&req->wait_queue, NULL); | ||
59 | retry = true; | ||
60 | break; | ||
61 | } | ||
62 | diff --git a/block/io.c b/block/io.c | ||
63 | index XXXXXXX..XXXXXXX 100644 | ||
64 | --- a/block/io.c | ||
65 | +++ b/block/io.c | ||
66 | @@ -XXX,XX +XXX,XX @@ static bool coroutine_fn wait_serialising_requests(BdrvTrackedRequest *self) | ||
67 | * (instead of producing a deadlock in the former case). */ | ||
68 | if (!req->waiting_for) { | ||
69 | self->waiting_for = req; | ||
70 | - qemu_co_queue_wait(&req->wait_queue); | ||
71 | + qemu_co_queue_wait(&req->wait_queue, NULL); | ||
72 | self->waiting_for = NULL; | ||
73 | retry = true; | ||
74 | waited = true; | ||
75 | @@ -XXX,XX +XXX,XX @@ int coroutine_fn bdrv_co_flush(BlockDriverState *bs) | ||
76 | |||
77 | /* Wait until any previous flushes are completed */ | ||
78 | while (bs->active_flush_req) { | ||
79 | - qemu_co_queue_wait(&bs->flush_queue); | ||
80 | + qemu_co_queue_wait(&bs->flush_queue, NULL); | ||
81 | } | ||
82 | |||
83 | bs->active_flush_req = true; | ||
84 | diff --git a/block/nbd-client.c b/block/nbd-client.c | ||
85 | index XXXXXXX..XXXXXXX 100644 | ||
86 | --- a/block/nbd-client.c | ||
87 | +++ b/block/nbd-client.c | ||
88 | @@ -XXX,XX +XXX,XX @@ static void nbd_coroutine_start(NBDClientSession *s, | ||
89 | /* Poor man semaphore. The free_sema is locked when no other request | ||
90 | * can be accepted, and unlocked after receiving one reply. */ | ||
91 | if (s->in_flight == MAX_NBD_REQUESTS) { | ||
92 | - qemu_co_queue_wait(&s->free_sema); | ||
93 | + qemu_co_queue_wait(&s->free_sema, NULL); | ||
94 | assert(s->in_flight < MAX_NBD_REQUESTS); | ||
95 | } | ||
96 | s->in_flight++; | ||
97 | diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c | ||
98 | index XXXXXXX..XXXXXXX 100644 | ||
99 | --- a/block/qcow2-cluster.c | ||
100 | +++ b/block/qcow2-cluster.c | ||
101 | @@ -XXX,XX +XXX,XX @@ static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset, | ||
102 | if (bytes == 0) { | ||
103 | /* Wait for the dependency to complete. We need to recheck | ||
104 | * the free/allocated clusters when we continue. */ | ||
105 | - qemu_co_mutex_unlock(&s->lock); | ||
106 | - qemu_co_queue_wait(&old_alloc->dependent_requests); | ||
107 | - qemu_co_mutex_lock(&s->lock); | ||
108 | + qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock); | ||
109 | return -EAGAIN; | ||
110 | } | ||
111 | } | ||
112 | diff --git a/block/sheepdog.c b/block/sheepdog.c | ||
113 | index XXXXXXX..XXXXXXX 100644 | ||
114 | --- a/block/sheepdog.c | ||
115 | +++ b/block/sheepdog.c | ||
116 | @@ -XXX,XX +XXX,XX @@ static void wait_for_overlapping_aiocb(BDRVSheepdogState *s, SheepdogAIOCB *acb) | ||
117 | retry: | ||
118 | QLIST_FOREACH(cb, &s->inflight_aiocb_head, aiocb_siblings) { | ||
119 | if (AIOCBOverlapping(acb, cb)) { | ||
120 | - qemu_co_queue_wait(&s->overlapping_queue); | ||
121 | + qemu_co_queue_wait(&s->overlapping_queue, NULL); | ||
122 | goto retry; | ||
123 | } | ||
124 | } | ||
125 | diff --git a/block/throttle-groups.c b/block/throttle-groups.c | ||
126 | index XXXXXXX..XXXXXXX 100644 | ||
127 | --- a/block/throttle-groups.c | ||
128 | +++ b/block/throttle-groups.c | ||
129 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn throttle_group_co_io_limits_intercept(BlockBackend *blk, | ||
130 | if (must_wait || blkp->pending_reqs[is_write]) { | ||
131 | blkp->pending_reqs[is_write]++; | ||
132 | qemu_mutex_unlock(&tg->lock); | ||
133 | - qemu_co_queue_wait(&blkp->throttled_reqs[is_write]); | ||
134 | + qemu_co_queue_wait(&blkp->throttled_reqs[is_write], NULL); | ||
135 | qemu_mutex_lock(&tg->lock); | ||
136 | blkp->pending_reqs[is_write]--; | ||
137 | } | ||
138 | diff --git a/hw/9pfs/9p.c b/hw/9pfs/9p.c | ||
139 | index XXXXXXX..XXXXXXX 100644 | ||
140 | --- a/hw/9pfs/9p.c | ||
141 | +++ b/hw/9pfs/9p.c | ||
142 | @@ -XXX,XX +XXX,XX @@ static void coroutine_fn v9fs_flush(void *opaque) | ||
143 | /* | ||
144 | * Wait for pdu to complete. | ||
145 | */ | ||
146 | - qemu_co_queue_wait(&cancel_pdu->complete); | ||
147 | + qemu_co_queue_wait(&cancel_pdu->complete, NULL); | ||
148 | cancel_pdu->cancelled = 0; | ||
149 | pdu_free(cancel_pdu); | ||
150 | } | ||
151 | diff --git a/util/qemu-coroutine-lock.c b/util/qemu-coroutine-lock.c | ||
152 | index XXXXXXX..XXXXXXX 100644 | ||
153 | --- a/util/qemu-coroutine-lock.c | ||
154 | +++ b/util/qemu-coroutine-lock.c | ||
155 | @@ -XXX,XX +XXX,XX @@ void qemu_co_queue_init(CoQueue *queue) | ||
156 | QSIMPLEQ_INIT(&queue->entries); | ||
157 | } | ||
158 | |||
159 | -void coroutine_fn qemu_co_queue_wait(CoQueue *queue) | ||
160 | +void coroutine_fn qemu_co_queue_wait(CoQueue *queue, CoMutex *mutex) | ||
161 | { | ||
162 | Coroutine *self = qemu_coroutine_self(); | ||
163 | QSIMPLEQ_INSERT_TAIL(&queue->entries, self, co_queue_next); | ||
164 | + | ||
165 | + if (mutex) { | ||
166 | + qemu_co_mutex_unlock(mutex); | ||
167 | + } | ||
168 | + | ||
169 | + /* There is no race condition here. Other threads will call | ||
170 | + * aio_co_schedule on our AioContext, which can reenter this | ||
171 | + * coroutine but only after this yield and after the main loop | ||
172 | + * has gone through the next iteration. | ||
173 | + */ | ||
174 | qemu_coroutine_yield(); | ||
175 | assert(qemu_in_coroutine()); | ||
176 | + | ||
177 | + /* TODO: OSv implements wait morphing here, where the wakeup | ||
178 | + * primitive automatically places the woken coroutine on the | ||
179 | + * mutex's queue. This avoids the thundering herd effect. | ||
180 | + */ | ||
181 | + if (mutex) { | ||
182 | + qemu_co_mutex_lock(mutex); | ||
183 | + } | ||
184 | } | ||
185 | |||
186 | /** | ||
187 | @@ -XXX,XX +XXX,XX @@ void qemu_co_rwlock_rdlock(CoRwlock *lock) | ||
188 | Coroutine *self = qemu_coroutine_self(); | ||
189 | |||
190 | while (lock->writer) { | ||
191 | - qemu_co_queue_wait(&lock->queue); | ||
192 | + qemu_co_queue_wait(&lock->queue, NULL); | ||
193 | } | ||
194 | lock->reader++; | ||
195 | self->locks_held++; | ||
196 | @@ -XXX,XX +XXX,XX @@ void qemu_co_rwlock_wrlock(CoRwlock *lock) | ||
197 | Coroutine *self = qemu_coroutine_self(); | ||
198 | |||
199 | while (lock->writer || lock->reader) { | ||
200 | - qemu_co_queue_wait(&lock->queue); | ||
201 | + qemu_co_queue_wait(&lock->queue, NULL); | ||
202 | } | ||
203 | lock->writer = true; | ||
204 | self->locks_held++; | ||
205 | -- | ||
206 | 2.9.3 | ||
207 | |||
208 | diff view generated by jsdifflib |
1 | From: Peter Xu <peterx@redhat.com> | 1 | From: Paolo Bonzini <pbonzini@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | Lukas reported an hard to reproduce QMP iothread hang on s390 that | 3 | This adds a CoMutex around the existing CoQueue. Because the write-side |
4 | QEMU might hang at pthread_join() of the QMP monitor iothread before | 4 | can just take CoMutex, the old "writer" field is not necessary anymore. |
5 | quitting: | 5 | Instead of removing it altogether, count the number of pending writers |
6 | during a read-side critical section and forbid further readers from | ||
7 | entering. | ||
6 | 8 | ||
7 | Thread 1 | 9 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> |
8 | #0 0x000003ffad10932c in pthread_join | 10 | Reviewed-by: Fam Zheng <famz@redhat.com> |
9 | #1 0x0000000109e95750 in qemu_thread_join | 11 | Message-id: 20170213181244.16297-7-pbonzini@redhat.com |
10 | at /home/thuth/devel/qemu/util/qemu-thread-posix.c:570 | ||
11 | #2 0x0000000109c95a1c in iothread_stop | ||
12 | #3 0x0000000109bb0874 in monitor_cleanup | ||
13 | #4 0x0000000109b55042 in main | ||
14 | |||
15 | While the iothread is still in the main loop: | ||
16 | |||
17 | Thread 4 | ||
18 | #0 0x000003ffad0010e4 in ?? | ||
19 | #1 0x000003ffad553958 in g_main_context_iterate.isra.19 | ||
20 | #2 0x000003ffad553d90 in g_main_loop_run | ||
21 | #3 0x0000000109c9585a in iothread_run | ||
22 | at /home/thuth/devel/qemu/iothread.c:74 | ||
23 | #4 0x0000000109e94752 in qemu_thread_start | ||
24 | at /home/thuth/devel/qemu/util/qemu-thread-posix.c:502 | ||
25 | #5 0x000003ffad10825a in start_thread | ||
26 | #6 0x000003ffad00dcf2 in thread_start | ||
27 | |||
28 | IMHO it's because there's a race between the main thread and iothread | ||
29 | when stopping the thread in following sequence: | ||
30 | |||
31 | main thread iothread | ||
32 | =========== ============== | ||
33 | aio_poll() | ||
34 | iothread_get_g_main_context | ||
35 | set iothread->worker_context | ||
36 | iothread_stop | ||
37 | schedule iothread_stop_bh | ||
38 | execute iothread_stop_bh [1] | ||
39 | set iothread->running=false | ||
40 | (since main_loop==NULL so | ||
41 | skip to quit main loop. | ||
42 | Note: although main_loop is | ||
43 | NULL but worker_context is | ||
44 | not!) | ||
45 | atomic_read(&iothread->worker_context) [2] | ||
46 | create main_loop object | ||
47 | g_main_loop_run() [3] | ||
48 | pthread_join() [4] | ||
49 | |||
50 | We can see that when execute iothread_stop_bh() at [1] it's possible | ||
51 | that main_loop is still NULL because it's only created until the first | ||
52 | check of the worker_context later at [2]. Then the iothread will hang | ||
53 | in the main loop [3] and it'll starve the main thread too [4]. | ||
54 | |||
55 | Here the simple solution should be that we check again the "running" | ||
56 | variable before check against worker_context. | ||
57 | |||
58 | CC: Thomas Huth <thuth@redhat.com> | ||
59 | CC: Dr. David Alan Gilbert <dgilbert@redhat.com> | ||
60 | CC: Stefan Hajnoczi <stefanha@redhat.com> | ||
61 | CC: Lukáš Doktor <ldoktor@redhat.com> | ||
62 | CC: Markus Armbruster <armbru@redhat.com> | ||
63 | CC: Eric Blake <eblake@redhat.com> | ||
64 | CC: Paolo Bonzini <pbonzini@redhat.com> | ||
65 | Reported-by: Lukáš Doktor <ldoktor@redhat.com> | ||
66 | Signed-off-by: Peter Xu <peterx@redhat.com> | ||
67 | Tested-by: Thomas Huth <thuth@redhat.com> | ||
68 | Message-id: 20190129051432.22023-1-peterx@redhat.com | ||
69 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 12 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
70 | --- | 13 | --- |
71 | iothread.c | 6 +++++- | 14 | include/qemu/coroutine.h | 3 ++- |
72 | 1 file changed, 5 insertions(+), 1 deletion(-) | 15 | util/qemu-coroutine-lock.c | 35 ++++++++++++++++++++++++----------- |
16 | 2 files changed, 26 insertions(+), 12 deletions(-) | ||
73 | 17 | ||
74 | diff --git a/iothread.c b/iothread.c | 18 | diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h |
75 | index XXXXXXX..XXXXXXX 100644 | 19 | index XXXXXXX..XXXXXXX 100644 |
76 | --- a/iothread.c | 20 | --- a/include/qemu/coroutine.h |
77 | +++ b/iothread.c | 21 | +++ b/include/qemu/coroutine.h |
78 | @@ -XXX,XX +XXX,XX @@ static void *iothread_run(void *opaque) | 22 | @@ -XXX,XX +XXX,XX @@ bool qemu_co_queue_empty(CoQueue *queue); |
79 | while (iothread->running) { | 23 | |
80 | aio_poll(iothread->ctx, true); | 24 | |
81 | 25 | typedef struct CoRwlock { | |
82 | - if (atomic_read(&iothread->worker_context)) { | 26 | - bool writer; |
83 | + /* | 27 | + int pending_writer; |
84 | + * We must check the running state again in case it was | 28 | int reader; |
85 | + * changed in previous aio_poll() | 29 | + CoMutex mutex; |
86 | + */ | 30 | CoQueue queue; |
87 | + if (iothread->running && atomic_read(&iothread->worker_context)) { | 31 | } CoRwlock; |
88 | GMainLoop *loop; | 32 | |
89 | 33 | diff --git a/util/qemu-coroutine-lock.c b/util/qemu-coroutine-lock.c | |
90 | g_main_context_push_thread_default(iothread->worker_context); | 34 | index XXXXXXX..XXXXXXX 100644 |
35 | --- a/util/qemu-coroutine-lock.c | ||
36 | +++ b/util/qemu-coroutine-lock.c | ||
37 | @@ -XXX,XX +XXX,XX @@ void qemu_co_rwlock_init(CoRwlock *lock) | ||
38 | { | ||
39 | memset(lock, 0, sizeof(*lock)); | ||
40 | qemu_co_queue_init(&lock->queue); | ||
41 | + qemu_co_mutex_init(&lock->mutex); | ||
42 | } | ||
43 | |||
44 | void qemu_co_rwlock_rdlock(CoRwlock *lock) | ||
45 | { | ||
46 | Coroutine *self = qemu_coroutine_self(); | ||
47 | |||
48 | - while (lock->writer) { | ||
49 | - qemu_co_queue_wait(&lock->queue, NULL); | ||
50 | + qemu_co_mutex_lock(&lock->mutex); | ||
51 | + /* For fairness, wait if a writer is in line. */ | ||
52 | + while (lock->pending_writer) { | ||
53 | + qemu_co_queue_wait(&lock->queue, &lock->mutex); | ||
54 | } | ||
55 | lock->reader++; | ||
56 | + qemu_co_mutex_unlock(&lock->mutex); | ||
57 | + | ||
58 | + /* The rest of the read-side critical section is run without the mutex. */ | ||
59 | self->locks_held++; | ||
60 | } | ||
61 | |||
62 | @@ -XXX,XX +XXX,XX @@ void qemu_co_rwlock_unlock(CoRwlock *lock) | ||
63 | Coroutine *self = qemu_coroutine_self(); | ||
64 | |||
65 | assert(qemu_in_coroutine()); | ||
66 | - if (lock->writer) { | ||
67 | - lock->writer = false; | ||
68 | + if (!lock->reader) { | ||
69 | + /* The critical section started in qemu_co_rwlock_wrlock. */ | ||
70 | qemu_co_queue_restart_all(&lock->queue); | ||
71 | } else { | ||
72 | + self->locks_held--; | ||
73 | + | ||
74 | + qemu_co_mutex_lock(&lock->mutex); | ||
75 | lock->reader--; | ||
76 | assert(lock->reader >= 0); | ||
77 | /* Wakeup only one waiting writer */ | ||
78 | @@ -XXX,XX +XXX,XX @@ void qemu_co_rwlock_unlock(CoRwlock *lock) | ||
79 | qemu_co_queue_next(&lock->queue); | ||
80 | } | ||
81 | } | ||
82 | - self->locks_held--; | ||
83 | + qemu_co_mutex_unlock(&lock->mutex); | ||
84 | } | ||
85 | |||
86 | void qemu_co_rwlock_wrlock(CoRwlock *lock) | ||
87 | { | ||
88 | - Coroutine *self = qemu_coroutine_self(); | ||
89 | - | ||
90 | - while (lock->writer || lock->reader) { | ||
91 | - qemu_co_queue_wait(&lock->queue, NULL); | ||
92 | + qemu_co_mutex_lock(&lock->mutex); | ||
93 | + lock->pending_writer++; | ||
94 | + while (lock->reader) { | ||
95 | + qemu_co_queue_wait(&lock->queue, &lock->mutex); | ||
96 | } | ||
97 | - lock->writer = true; | ||
98 | - self->locks_held++; | ||
99 | + lock->pending_writer--; | ||
100 | + | ||
101 | + /* The rest of the write-side critical section is run with | ||
102 | + * the mutex taken, so that lock->reader remains zero. | ||
103 | + * There is no need to update self->locks_held. | ||
104 | + */ | ||
105 | } | ||
91 | -- | 106 | -- |
92 | 2.20.1 | 107 | 2.9.3 |
93 | 108 | ||
94 | 109 | diff view generated by jsdifflib |