1
The following changes since commit dac03af5d5482ec7ee9c23db467bb7230b33c0d9:
1
The following changes since commit 56f9e46b841c7be478ca038d8d4085d776ab4b0d:
2
2
3
Merge remote-tracking branch 'remotes/rth/tags/pull-axp-20190825' into staging (2019-08-27 10:00:51 +0100)
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
https://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 5396234b96a2ac743f48644529771498e036e698:
9
for you to fetch changes up to a7b91d35bab97a2d3e779d0c64c9b837b52a6cf7:
10
10
11
block/qcow2: implement .bdrv_co_pwritev(_compressed)_part (2019-08-27 14:58:42 +0100)
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
v2:
17
* Rebased to resolve scsi conflicts
18
16
----------------------------------------------------------------
19
----------------------------------------------------------------
17
20
18
Vladimir Sementsov-Ogievskiy (12):
21
Paolo Bonzini (24):
19
util/iov: introduce qemu_iovec_init_extended
22
block: move AioContext, QEMUTimer, main-loop to libqemuutil
20
util/iov: improve qemu_iovec_is_zero
23
aio: introduce aio_co_schedule and aio_co_wake
21
block/io: refactor padding
24
block-backend: allow blk_prw from coroutine context
22
block: define .*_part io handlers in BlockDriver
25
test-thread-pool: use generic AioContext infrastructure
23
block/io: bdrv_co_do_copy_on_readv: use and support qiov_offset
26
io: add methods to set I/O handlers on AioContext
24
block/io: bdrv_co_do_copy_on_readv: lazy allocation
27
io: make qio_channel_yield aware of AioContexts
25
block/io: bdrv_aligned_preadv: use and support qiov_offset
28
nbd: convert to use qio_channel_yield
26
block/io: bdrv_aligned_pwritev: use and support qiov_offset
29
coroutine-lock: reschedule coroutine on the AioContext it was running
27
block/io: introduce bdrv_co_p{read, write}v_part
30
on
28
block/qcow2: refactor qcow2_co_preadv to use buffer-based io
31
blkdebug: reschedule coroutine on the AioContext it is running on
29
block/qcow2: implement .bdrv_co_preadv_part
32
qed: introduce qed_aio_start_io and qed_aio_next_io_cb
30
block/qcow2: implement .bdrv_co_pwritev(_compressed)_part
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
31
48
32
block/qcow2.h | 1 +
49
Makefile.objs | 4 -
33
include/block/block_int.h | 21 ++
50
stubs/Makefile.objs | 1 +
34
include/qemu/iov.h | 10 +-
51
tests/Makefile.include | 19 +-
35
block/backup.c | 2 +-
52
util/Makefile.objs | 6 +-
36
block/io.c | 541 +++++++++++++++++++++++---------------
53
block/nbd-client.h | 2 +-
37
block/qcow2-cluster.c | 14 +-
54
block/qed.h | 3 +
38
block/qcow2.c | 131 +++++----
55
include/block/aio.h | 38 ++-
39
qemu-img.c | 4 +-
56
include/block/block_int.h | 64 +++--
40
util/iov.c | 153 +++++++++--
57
include/io/channel.h | 72 +++++-
41
9 files changed, 568 insertions(+), 309 deletions(-)
58
include/qemu/coroutine.h | 84 ++++---
59
include/qemu/coroutine_int.h | 11 +-
60
include/sysemu/block-backend.h | 14 +-
61
tests/iothread.h | 25 ++
62
block/backup.c | 2 +-
63
block/blkdebug.c | 9 +-
64
block/blkreplay.c | 2 +-
65
block/block-backend.c | 13 +-
66
block/curl.c | 44 +++-
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%)
42
129
43
--
130
--
44
2.21.0
131
2.9.3
45
132
46
133
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Use and support new API in bdrv_aligned_pwritev.
3
AioContext is fairly self contained, the only dependency is QEMUTimer but
4
4
that in turn doesn't need anything else. So move them out of block-obj-y
5
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
5
to avoid introducing a dependency from io/ to block-obj-y.
6
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
6
7
Message-id: 20190604161514.262241-9-vsementsov@virtuozzo.com
7
main-loop and its dependency iohandler also need to be moved, because
8
Message-Id: <20190604161514.262241-9-vsementsov@virtuozzo.com>
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
9
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
17
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
10
---
18
---
11
block/io.c | 27 +++++++++++++--------------
19
Makefile.objs | 4 ---
12
1 file changed, 13 insertions(+), 14 deletions(-)
20
stubs/Makefile.objs | 1 +
13
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
14
diff --git a/block/io.c b/block/io.c
143
diff --git a/block/io.c b/block/io.c
15
index XXXXXXX..XXXXXXX 100644
144
index XXXXXXX..XXXXXXX 100644
16
--- a/block/io.c
145
--- a/block/io.c
17
+++ b/block/io.c
146
+++ b/block/io.c
18
@@ -XXX,XX +XXX,XX @@ bdrv_co_write_req_finish(BdrvChild *child, int64_t offset, uint64_t bytes,
147
@@ -XXX,XX +XXX,XX @@ BlockAIOCB *bdrv_aio_flush(BlockDriverState *bs,
19
*/
148
return &acb->common;
20
static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
149
}
21
BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
150
22
- int64_t align, QEMUIOVector *qiov, int flags)
151
-void *qemu_aio_get(const AIOCBInfo *aiocb_info, BlockDriverState *bs,
23
+ int64_t align, QEMUIOVector *qiov, size_t qiov_offset, int flags)
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,
24
{
226
{
25
BlockDriverState *bs = child->bs;
227
abort();
26
BlockDriver *drv = bs->drv;
228
}
27
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
229
-
28
assert(is_power_of_2(align));
230
-void aio_set_fd_handler(AioContext *ctx,
29
assert((offset & (align - 1)) == 0);
231
- int fd,
30
assert((bytes & (align - 1)) == 0);
232
- bool is_external,
31
- assert(!qiov || bytes == qiov->size);
233
- IOHandler *io_read,
32
+ assert(!qiov || qiov_offset + bytes <= qiov->size);
234
- IOHandler *io_write,
33
max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
235
- AioPollFn *io_poll,
34
align);
236
- void *opaque)
35
237
-{
36
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
238
- abort();
37
239
-}
38
if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
240
diff --git a/aio-posix.c b/util/aio-posix.c
39
!(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes &&
241
similarity index 99%
40
- qemu_iovec_is_zero(qiov, 0, qiov->size)) {
242
rename from aio-posix.c
41
+ qemu_iovec_is_zero(qiov, qiov_offset, bytes)) {
243
rename to util/aio-posix.c
42
flags |= BDRV_REQ_ZERO_WRITE;
244
index XXXXXXX..XXXXXXX 100644
43
if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
245
--- a/aio-posix.c
44
flags |= BDRV_REQ_MAY_UNMAP;
246
+++ b/util/aio-posix.c
45
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
247
@@ -XXX,XX +XXX,XX @@
46
bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO);
248
#include "qemu/rcu_queue.h"
47
ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
249
#include "qemu/sockets.h"
48
} else if (flags & BDRV_REQ_WRITE_COMPRESSED) {
250
#include "qemu/cutils.h"
49
- ret = bdrv_driver_pwritev_compressed(bs, offset, bytes, qiov, 0);
251
-#include "trace-root.h"
50
+ ret = bdrv_driver_pwritev_compressed(bs, offset, bytes,
252
+#include "trace.h"
51
+ qiov, qiov_offset);
253
#ifdef CONFIG_EPOLL_CREATE1
52
} else if (bytes <= max_transfer) {
254
#include <sys/epoll.h>
53
bdrv_debug_event(bs, BLKDBG_PWRITEV);
255
#endif
54
- ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, 0, flags);
256
diff --git a/aio-win32.c b/util/aio-win32.c
55
+ ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, qiov_offset, flags);
257
similarity index 100%
56
} else {
258
rename from aio-win32.c
57
bdrv_debug_event(bs, BLKDBG_PWRITEV);
259
rename to util/aio-win32.c
58
while (bytes_remaining) {
260
diff --git a/util/aiocb.c b/util/aiocb.c
59
int num = MIN(bytes_remaining, max_transfer);
261
new file mode 100644
60
- QEMUIOVector local_qiov;
262
index XXXXXXX..XXXXXXX
61
int local_flags = flags;
263
--- /dev/null
62
264
+++ b/util/aiocb.c
63
assert(num);
265
@@ -XXX,XX +XXX,XX @@
64
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
266
+/*
65
* need to flush on the last iteration */
267
+ * BlockAIOCB allocation
66
local_flags &= ~BDRV_REQ_FUA;
268
+ *
67
}
269
+ * Copyright (c) 2003-2017 Fabrice Bellard and other QEMU contributors
68
- qemu_iovec_init(&local_qiov, qiov->niov);
270
+ *
69
- qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
271
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
70
272
+ * of this software and associated documentation files (the "Software"), to deal
71
ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
273
+ * in the Software without restriction, including without limitation the rights
72
- num, &local_qiov, 0, local_flags);
274
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
73
- qemu_iovec_destroy(&local_qiov);
275
+ * copies of the Software, and to permit persons to whom the Software is
74
+ num, qiov, bytes - bytes_remaining,
276
+ * furnished to do so, subject to the following conditions:
75
+ local_flags);
277
+ *
76
if (ret < 0) {
278
+ * The above copyright notice and this permission notice shall be included in
77
break;
279
+ * all copies or substantial portions of the Software.
78
}
280
+ *
79
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
281
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
80
282
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
81
qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes);
283
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
82
ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes,
284
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
83
- align, &local_qiov,
285
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
84
+ align, &local_qiov, 0,
286
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
85
flags & ~BDRV_REQ_ZERO_WRITE);
287
+ * THE SOFTWARE.
86
if (ret < 0 || pad.merge_reads) {
288
+ */
87
/* Error or all work is done */
289
+
88
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
290
+#include "qemu/osdep.h"
89
/* Write the aligned part in the middle. */
291
+#include "block/aio.h"
90
uint64_t aligned_bytes = bytes & ~(align - 1);
292
+
91
ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
293
+void *qemu_aio_get(const AIOCBInfo *aiocb_info, BlockDriverState *bs,
92
- NULL, flags);
294
+ BlockCompletionFunc *cb, void *opaque)
93
+ NULL, 0, flags);
295
+{
94
if (ret < 0) {
296
+ BlockAIOCB *acb;
95
goto out;
297
+
96
}
298
+ acb = g_malloc(aiocb_info->aiocb_size);
97
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
299
+ acb->aiocb_info = aiocb_info;
98
300
+ acb->bs = bs;
99
qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align);
301
+ acb->cb = cb;
100
ret = bdrv_aligned_pwritev(child, req, offset, align, align,
302
+ acb->opaque = opaque;
101
- &local_qiov, flags & ~BDRV_REQ_ZERO_WRITE);
303
+ acb->refcnt = 1;
102
+ &local_qiov, 0,
304
+ return acb;
103
+ flags & ~BDRV_REQ_ZERO_WRITE);
305
+}
104
}
306
+
105
307
+void qemu_aio_ref(void *p)
106
out:
308
+{
107
@@ -XXX,XX +XXX,XX @@ int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
309
+ BlockAIOCB *acb = p;
108
}
310
+ acb->refcnt++;
109
311
+}
110
ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
312
+
111
- qiov, flags);
313
+void qemu_aio_unref(void *p)
112
+ qiov, 0, flags);
314
+{
113
315
+ BlockAIOCB *acb = p;
114
bdrv_padding_destroy(&pad);
316
+ assert(acb->refcnt > 0);
115
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"
116
--
409
--
117
2.21.0
410
2.9.3
118
411
119
412
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
We have similar padding code in bdrv_co_pwritev,
3
aio_co_wake provides the infrastructure to start a coroutine on a "home"
4
bdrv_co_do_pwrite_zeroes and bdrv_co_preadv. Let's combine and unify
4
AioContext. It will be used by CoMutex and CoQueue, so that coroutines
5
it.
5
don't jump from one context to another when they go to sleep on a
6
6
mutex or waitqueue. However, it can also be used as a more efficient
7
[Squashed in Vladimir's qemu-iotests 077 fix
7
alternative to one-shot bottom halves, and saves the effort of tracking
8
--Stefan]
8
which AioContext a coroutine is running on.
9
9
10
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
10
aio_co_schedule is the part of aio_co_wake that starts a coroutine
11
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
11
on a remove AioContext, but it is also useful to implement e.g.
12
Message-id: 20190604161514.262241-4-vsementsov@virtuozzo.com
12
bdrv_set_aio_context callbacks.
13
Message-Id: <20190604161514.262241-4-vsementsov@virtuozzo.com>
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
14
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
27
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
15
---
28
---
16
block/io.c | 365 +++++++++++++++++++++++++++++------------------------
29
tests/Makefile.include | 8 +-
17
1 file changed, 200 insertions(+), 165 deletions(-)
30
include/block/aio.h | 32 +++++++
18
31
include/qemu/coroutine_int.h | 11 ++-
19
diff --git a/block/io.c b/block/io.c
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
20
index XXXXXXX..XXXXXXX 100644
44
index XXXXXXX..XXXXXXX 100644
21
--- a/block/io.c
45
--- a/tests/Makefile.include
22
+++ b/block/io.c
46
+++ b/tests/Makefile.include
23
@@ -XXX,XX +XXX,XX @@ out:
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) $(test-util-obj-y)
77
diff --git a/include/block/aio.h b/include/block/aio.h
78
index XXXXXXX..XXXXXXX 100644
79
--- a/include/block/aio.h
80
+++ b/include/block/aio.h
81
@@ -XXX,XX +XXX,XX @@ typedef void QEMUBHFunc(void *opaque);
82
typedef bool AioPollFn(void *opaque);
83
typedef void IOHandler(void *opaque);
84
85
+struct Coroutine;
86
struct ThreadPool;
87
struct LinuxAioState;
88
89
@@ -XXX,XX +XXX,XX @@ struct AioContext {
90
bool notified;
91
EventNotifier notifier;
92
93
+ QSLIST_HEAD(, Coroutine) scheduled_coroutines;
94
+ QEMUBH *co_schedule_bh;
95
+
96
/* Thread pool for performing work and receiving completion callbacks.
97
* Has its own locking.
98
*/
99
@@ -XXX,XX +XXX,XX @@ static inline bool aio_node_check(AioContext *ctx, bool is_external)
24
}
100
}
25
101
26
/*
102
/**
27
- * Handle a read request in coroutine context
103
+ * aio_co_schedule:
28
+ * Request padding
104
+ * @ctx: the aio context
29
+ *
105
+ * @co: the coroutine
30
+ * |<---- align ----->| |<----- align ---->|
106
+ *
31
+ * |<- head ->|<------------- bytes ------------->|<-- tail -->|
107
+ * Start a coroutine on a remote AioContext.
32
+ * | | | | | |
108
+ *
33
+ * -*----------$-------*-------- ... --------*-----$------------*---
109
+ * The coroutine must not be entered by anyone else while aio_co_schedule()
34
+ * | | | | | |
110
+ * is active. In addition the coroutine must have yielded unless ctx
35
+ * | offset | | end |
111
+ * is the context in which the coroutine is running (i.e. the value of
36
+ * ALIGN_DOWN(offset) ALIGN_UP(offset) ALIGN_DOWN(end) ALIGN_UP(end)
112
+ * qemu_get_current_aio_context() from the coroutine itself).
37
+ * [buf ... ) [tail_buf )
113
+ */
38
+ *
114
+void aio_co_schedule(AioContext *ctx, struct Coroutine *co);
39
+ * @buf is an aligned allocation needed to store @head and @tail paddings. @head
115
+
40
+ * is placed at the beginning of @buf and @tail at the @end.
116
+/**
41
+ *
117
+ * aio_co_wake:
42
+ * @tail_buf is a pointer to sub-buffer, corresponding to align-sized chunk
118
+ * @co: the coroutine
43
+ * around tail, if tail exists.
119
+ *
44
+ *
120
+ * Restart a coroutine on the AioContext where it was running last, thus
45
+ * @merge_reads is true for small requests,
121
+ * preventing coroutines from jumping from one context to another when they
46
+ * if @buf_len == @head + bytes + @tail. In this case it is possible that both
122
+ * go to sleep.
47
+ * head and tail exist but @buf_len == align and @tail_buf == @buf.
123
+ *
48
*/
124
+ * aio_co_wake may be executed either in coroutine or non-coroutine
49
+typedef struct BdrvRequestPadding {
125
+ * context. The coroutine must not be entered by anyone else while
50
+ uint8_t *buf;
126
+ * aio_co_wake() is active.
51
+ size_t buf_len;
127
+ */
52
+ uint8_t *tail_buf;
128
+void aio_co_wake(struct Coroutine *co);
53
+ size_t head;
129
+
54
+ size_t tail;
130
+/**
55
+ bool merge_reads;
131
* Return the AioContext whose event loop runs in the current thread.
56
+ QEMUIOVector local_qiov;
132
*
57
+} BdrvRequestPadding;
133
* If called from an IOThread this will be the IOThread's AioContext. If
58
+
134
diff --git a/include/qemu/coroutine_int.h b/include/qemu/coroutine_int.h
59
+static bool bdrv_init_padding(BlockDriverState *bs,
135
index XXXXXXX..XXXXXXX 100644
60
+ int64_t offset, int64_t bytes,
136
--- a/include/qemu/coroutine_int.h
61
+ BdrvRequestPadding *pad)
137
+++ b/include/qemu/coroutine_int.h
62
+{
138
@@ -XXX,XX +XXX,XX @@ struct Coroutine {
63
+ uint64_t align = bs->bl.request_alignment;
139
CoroutineEntry *entry;
64
+ size_t sum;
140
void *entry_arg;
65
+
141
Coroutine *caller;
66
+ memset(pad, 0, sizeof(*pad));
142
+
67
+
143
+ /* Only used when the coroutine has terminated. */
68
+ pad->head = offset & (align - 1);
144
QSLIST_ENTRY(Coroutine) pool_next;
69
+ pad->tail = ((offset + bytes) & (align - 1));
145
+
70
+ if (pad->tail) {
146
size_t locks_held;
71
+ pad->tail = align - pad->tail;
147
72
+ }
148
- /* Coroutines that should be woken up when we yield or terminate */
73
+
149
+ /* Coroutines that should be woken up when we yield or terminate.
74
+ if ((!pad->head && !pad->tail) || !bytes) {
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);
75
+ return false;
420
+ return false;
76
+ }
421
+ }
77
+
422
+
78
+ sum = pad->head + bytes + pad->tail;
423
+ if (n == id) {
79
+ pad->buf_len = (sum > align && pad->head && pad->tail) ? 2 * align : align;
424
+ atomic_inc(&count_here);
80
+ pad->buf = qemu_blockalign(bs, pad->buf_len);
425
+ } else {
81
+ pad->merge_reads = sum == pad->buf_len;
426
+ atomic_inc(&count_other);
82
+ if (pad->tail) {
427
+ }
83
+ pad->tail_buf = pad->buf + pad->buf_len - align;
428
+
84
+ }
429
+ aio_co_schedule(ctx[n], co);
85
+
86
+ return true;
430
+ return true;
87
+}
431
+}
88
+
432
+
89
+static int bdrv_padding_rmw_read(BdrvChild *child,
433
+static void finish_cb(void *opaque)
90
+ BdrvTrackedRequest *req,
434
+{
91
+ BdrvRequestPadding *pad,
435
+ schedule_next(id);
92
+ bool zero_middle)
436
+}
93
+{
437
+
94
+ QEMUIOVector local_qiov;
438
+static coroutine_fn void test_multi_co_schedule_entry(void *opaque)
95
+ BlockDriverState *bs = child->bs;
439
+{
96
+ uint64_t align = bs->bl.request_alignment;
440
+ g_assert(to_schedule[id] == NULL);
97
+ int ret;
441
+ atomic_mb_set(&to_schedule[id], qemu_coroutine_self());
98
+
442
+
99
+ assert(req->serialising && pad->buf);
443
+ while (!atomic_mb_read(&now_stopping)) {
100
+
444
+ int n;
101
+ if (pad->head || pad->merge_reads) {
445
+
102
+ uint64_t bytes = pad->merge_reads ? pad->buf_len : align;
446
+ n = g_test_rand_int_range(0, NUM_CONTEXTS);
103
+
447
+ schedule_next(n);
104
+ qemu_iovec_init_buf(&local_qiov, pad->buf, bytes);
448
+ qemu_coroutine_yield();
105
+
449
+
106
+ if (pad->head) {
450
+ g_assert(to_schedule[id] == NULL);
107
+ bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
451
+ atomic_mb_set(&to_schedule[id], qemu_coroutine_self());
108
+ }
452
+ }
109
+ if (pad->merge_reads && pad->tail) {
453
+}
110
+ bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
454
+
111
+ }
455
+
112
+ ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes,
456
+static void test_multi_co_schedule(int seconds)
113
+ align, &local_qiov, 0);
457
+{
114
+ if (ret < 0) {
458
+ int i;
115
+ return ret;
459
+
116
+ }
460
+ count_here = count_other = count_retry = 0;
117
+ if (pad->head) {
461
+ now_stopping = false;
118
+ bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
462
+
119
+ }
463
+ create_aio_contexts();
120
+ if (pad->merge_reads && pad->tail) {
464
+ for (i = 0; i < NUM_CONTEXTS; i++) {
121
+ bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
465
+ Coroutine *co1 = qemu_coroutine_create(test_multi_co_schedule_entry, NULL);
122
+ }
466
+ aio_co_schedule(ctx[i], co1);
123
+
467
+ }
124
+ if (pad->merge_reads) {
468
+
125
+ goto zero_mem;
469
+ g_usleep(seconds * 1000000);
126
+ }
470
+
127
+ }
471
+ atomic_mb_set(&now_stopping, true);
128
+
472
+ for (i = 0; i < NUM_CONTEXTS; i++) {
129
+ if (pad->tail) {
473
+ ctx_run(i, finish_cb, NULL);
130
+ qemu_iovec_init_buf(&local_qiov, pad->tail_buf, align);
474
+ to_schedule[i] = NULL;
131
+
475
+ }
132
+ bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
476
+
133
+ ret = bdrv_aligned_preadv(
477
+ join_aio_contexts();
134
+ child, req,
478
+ g_test_message("scheduled %d, queued %d, retry %d, total %d\n",
135
+ req->overlap_offset + req->overlap_bytes - align,
479
+ count_other, count_here, count_retry,
136
+ align, align, &local_qiov, 0);
480
+ count_here + count_other + count_retry);
137
+ if (ret < 0) {
481
+}
138
+ return ret;
482
+
139
+ }
483
+static void test_multi_co_schedule_1(void)
140
+ bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
484
+{
141
+ }
485
+ test_multi_co_schedule(1);
142
+
486
+}
143
+zero_mem:
487
+
144
+ if (zero_middle) {
488
+static void test_multi_co_schedule_10(void)
145
+ memset(pad->buf + pad->head, 0, pad->buf_len - pad->head - pad->tail);
489
+{
146
+ }
490
+ test_multi_co_schedule(10);
147
+
491
+}
148
+ return 0;
492
+
149
+}
493
+/* End of tests. */
150
+
494
+
151
+static void bdrv_padding_destroy(BdrvRequestPadding *pad)
495
+int main(int argc, char **argv)
152
+{
496
+{
153
+ if (pad->buf) {
497
+ init_clocks();
154
+ qemu_vfree(pad->buf);
498
+
155
+ qemu_iovec_destroy(&pad->local_qiov);
499
+ g_test_init(&argc, &argv, NULL);
156
+ }
500
+ g_test_add_func("/aio/multi/lifecycle", test_lifecycle);
157
+}
501
+ if (g_test_quick()) {
158
+
502
+ g_test_add_func("/aio/multi/schedule", test_multi_co_schedule_1);
159
+/*
503
+ } else {
160
+ * bdrv_pad_request
504
+ g_test_add_func("/aio/multi/schedule", test_multi_co_schedule_10);
161
+ *
505
+ }
162
+ * Exchange request parameters with padded request if needed. Don't include RMW
506
+ return g_test_run();
163
+ * read of padding, bdrv_padding_rmw_read() should be called separately if
507
+}
164
+ * needed.
508
diff --git a/util/async.c b/util/async.c
165
+ *
509
index XXXXXXX..XXXXXXX 100644
166
+ * All parameters except @bs are in-out: they represent original request at
510
--- a/util/async.c
167
+ * function call and padded (if padding needed) at function finish.
511
+++ b/util/async.c
168
+ *
512
@@ -XXX,XX +XXX,XX @@
169
+ * Function always succeeds.
513
#include "qemu/main-loop.h"
170
+ */
514
#include "qemu/atomic.h"
171
+static bool bdrv_pad_request(BlockDriverState *bs, QEMUIOVector **qiov,
515
#include "block/raw-aio.h"
172
+ int64_t *offset, unsigned int *bytes,
516
+#include "qemu/coroutine_int.h"
173
+ BdrvRequestPadding *pad)
517
+#include "trace.h"
174
+{
518
175
+ if (!bdrv_init_padding(bs, *offset, *bytes, pad)) {
519
/***********************************************************/
176
+ return false;
520
/* bottom halves (can be seen as timers which expire ASAP) */
177
+ }
521
@@ -XXX,XX +XXX,XX @@ aio_ctx_finalize(GSource *source)
178
+
522
}
179
+ qemu_iovec_init_extended(&pad->local_qiov, pad->buf, pad->head,
523
#endif
180
+ *qiov, 0, *bytes,
524
181
+ pad->buf + pad->buf_len - pad->tail, pad->tail);
525
+ assert(QSLIST_EMPTY(&ctx->scheduled_coroutines));
182
+ *bytes += pad->head + pad->tail;
526
+ qemu_bh_delete(ctx->co_schedule_bh);
183
+ *offset -= pad->head;
527
+
184
+ *qiov = &pad->local_qiov;
528
qemu_lockcnt_lock(&ctx->list_lock);
185
+
529
assert(!qemu_lockcnt_count(&ctx->list_lock));
186
+ return true;
530
while (ctx->first_bh) {
187
+}
531
@@ -XXX,XX +XXX,XX @@ static bool event_notifier_poll(void *opaque)
188
+
532
return atomic_read(&ctx->notified);
189
int coroutine_fn bdrv_co_preadv(BdrvChild *child,
533
}
190
int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
534
191
BdrvRequestFlags flags)
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)
192
{
558
{
193
BlockDriverState *bs = child->bs;
194
- BlockDriver *drv = bs->drv;
195
BdrvTrackedRequest req;
196
-
197
- uint64_t align = bs->bl.request_alignment;
198
- uint8_t *head_buf = NULL;
199
- uint8_t *tail_buf = NULL;
200
- QEMUIOVector local_qiov;
201
- bool use_local_qiov = false;
202
+ BdrvRequestPadding pad;
203
int ret;
559
int ret;
204
560
@@ -XXX,XX +XXX,XX @@ AioContext *aio_context_new(Error **errp)
205
- trace_bdrv_co_preadv(child->bs, offset, bytes, flags);
206
-
207
- if (!drv) {
208
- return -ENOMEDIUM;
209
- }
210
+ trace_bdrv_co_preadv(bs, offset, bytes, flags);
211
212
ret = bdrv_check_byte_request(bs, offset, bytes);
213
if (ret < 0) {
214
@@ -XXX,XX +XXX,XX @@ int coroutine_fn bdrv_co_preadv(BdrvChild *child,
215
flags |= BDRV_REQ_COPY_ON_READ;
216
}
561
}
217
562
g_source_set_can_recurse(&ctx->source, true);
218
- /* Align read if necessary by padding qiov */
563
qemu_lockcnt_init(&ctx->list_lock);
219
- if (offset & (align - 1)) {
564
+
220
- head_buf = qemu_blockalign(bs, align);
565
+ ctx->co_schedule_bh = aio_bh_new(ctx, co_schedule_bh_cb, ctx);
221
- qemu_iovec_init(&local_qiov, qiov->niov + 2);
566
+ QSLIST_INIT(&ctx->scheduled_coroutines);
222
- qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1));
567
+
223
- qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
568
aio_set_event_notifier(ctx, &ctx->notifier,
224
- use_local_qiov = true;
569
false,
225
-
570
(EventNotifierHandler *)
226
- bytes += offset & (align - 1);
571
@@ -XXX,XX +XXX,XX @@ fail:
227
- offset = offset & ~(align - 1);
572
return NULL;
228
- }
229
-
230
- if ((offset + bytes) & (align - 1)) {
231
- if (!use_local_qiov) {
232
- qemu_iovec_init(&local_qiov, qiov->niov + 1);
233
- qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
234
- use_local_qiov = true;
235
- }
236
- tail_buf = qemu_blockalign(bs, align);
237
- qemu_iovec_add(&local_qiov, tail_buf,
238
- align - ((offset + bytes) & (align - 1)));
239
-
240
- bytes = ROUND_UP(bytes, align);
241
- }
242
+ bdrv_pad_request(bs, &qiov, &offset, &bytes, &pad);
243
244
tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
245
- ret = bdrv_aligned_preadv(child, &req, offset, bytes, align,
246
- use_local_qiov ? &local_qiov : qiov,
247
- flags);
248
+ ret = bdrv_aligned_preadv(child, &req, offset, bytes,
249
+ bs->bl.request_alignment,
250
+ qiov, flags);
251
tracked_request_end(&req);
252
bdrv_dec_in_flight(bs);
253
254
- if (use_local_qiov) {
255
- qemu_iovec_destroy(&local_qiov);
256
- qemu_vfree(head_buf);
257
- qemu_vfree(tail_buf);
258
- }
259
+ bdrv_padding_destroy(&pad);
260
261
return ret;
262
}
573
}
263
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
574
264
BdrvTrackedRequest *req)
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)
265
{
610
{
266
BlockDriverState *bs = child->bs;
611
g_source_ref(&ctx->source);
267
- uint8_t *buf = NULL;
612
diff --git a/util/qemu-coroutine.c b/util/qemu-coroutine.c
268
QEMUIOVector local_qiov;
613
index XXXXXXX..XXXXXXX 100644
269
uint64_t align = bs->bl.request_alignment;
614
--- a/util/qemu-coroutine.c
270
- unsigned int head_padding_bytes, tail_padding_bytes;
615
+++ b/util/qemu-coroutine.c
271
int ret = 0;
616
@@ -XXX,XX +XXX,XX @@
272
+ bool padding;
617
#include "qemu/atomic.h"
273
+ BdrvRequestPadding pad;
618
#include "qemu/coroutine.h"
274
619
#include "qemu/coroutine_int.h"
275
- head_padding_bytes = offset & (align - 1);
620
+#include "block/aio.h"
276
- tail_padding_bytes = (align - (offset + bytes)) & (align - 1);
621
277
-
622
enum {
278
-
623
POOL_BATCH_SIZE = 64,
279
- assert(flags & BDRV_REQ_ZERO_WRITE);
624
@@ -XXX,XX +XXX,XX @@ void qemu_coroutine_enter(Coroutine *co)
280
- if (head_padding_bytes || tail_padding_bytes) {
281
- buf = qemu_blockalign(bs, align);
282
- qemu_iovec_init_buf(&local_qiov, buf, align);
283
- }
284
- if (head_padding_bytes) {
285
- uint64_t zero_bytes = MIN(bytes, align - head_padding_bytes);
286
-
287
- /* RMW the unaligned part before head. */
288
+ padding = bdrv_init_padding(bs, offset, bytes, &pad);
289
+ if (padding) {
290
mark_request_serialising(req, align);
291
wait_serialising_requests(req);
292
- bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
293
- ret = bdrv_aligned_preadv(child, req, offset & ~(align - 1), align,
294
- align, &local_qiov, 0);
295
- if (ret < 0) {
296
- goto fail;
297
- }
298
- bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
299
300
- memset(buf + head_padding_bytes, 0, zero_bytes);
301
- ret = bdrv_aligned_pwritev(child, req, offset & ~(align - 1), align,
302
- align, &local_qiov,
303
- flags & ~BDRV_REQ_ZERO_WRITE);
304
- if (ret < 0) {
305
- goto fail;
306
+ bdrv_padding_rmw_read(child, req, &pad, true);
307
+
308
+ if (pad.head || pad.merge_reads) {
309
+ int64_t aligned_offset = offset & ~(align - 1);
310
+ int64_t write_bytes = pad.merge_reads ? pad.buf_len : align;
311
+
312
+ qemu_iovec_init_buf(&local_qiov, pad.buf, write_bytes);
313
+ ret = bdrv_aligned_pwritev(child, req, aligned_offset, write_bytes,
314
+ align, &local_qiov,
315
+ flags & ~BDRV_REQ_ZERO_WRITE);
316
+ if (ret < 0 || pad.merge_reads) {
317
+ /* Error or all work is done */
318
+ goto out;
319
+ }
320
+ offset += write_bytes - pad.head;
321
+ bytes -= write_bytes - pad.head;
322
}
323
- offset += zero_bytes;
324
- bytes -= zero_bytes;
325
}
625
}
326
626
327
assert(!bytes || (offset & (align - 1)) == 0);
627
co->caller = self;
328
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
628
+ co->ctx = qemu_get_current_aio_context();
329
ret = bdrv_aligned_pwritev(child, req, offset, aligned_bytes, align,
629
+
330
NULL, flags);
630
+ /* Store co->ctx before anything that stores co. Matches
331
if (ret < 0) {
631
+ * barrier in aio_co_wake.
332
- goto fail;
632
+ */
333
+ goto out;
633
+ smp_wmb();
334
}
634
+
335
bytes -= aligned_bytes;
635
ret = qemu_coroutine_switch(self, co, COROUTINE_ENTER);
336
offset += aligned_bytes;
636
337
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_zero_pwritev(BdrvChild *child,
637
qemu_co_queue_run_restart(co);
338
638
diff --git a/util/trace-events b/util/trace-events
339
assert(!bytes || (offset & (align - 1)) == 0);
639
index XXXXXXX..XXXXXXX 100644
340
if (bytes) {
640
--- a/util/trace-events
341
- assert(align == tail_padding_bytes + bytes);
641
+++ b/util/trace-events
342
- /* RMW the unaligned part after tail. */
642
@@ -XXX,XX +XXX,XX @@ run_poll_handlers_end(void *ctx, bool progress) "ctx %p progress %d"
343
- mark_request_serialising(req, align);
643
poll_shrink(void *ctx, int64_t old, int64_t new) "ctx %p old %"PRId64" new %"PRId64
344
- wait_serialising_requests(req);
644
poll_grow(void *ctx, int64_t old, int64_t new) "ctx %p old %"PRId64" new %"PRId64
345
- bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
645
346
- ret = bdrv_aligned_preadv(child, req, offset, align,
646
+# util/async.c
347
- align, &local_qiov, 0);
647
+aio_co_schedule(void *ctx, void *co) "ctx %p co %p"
348
- if (ret < 0) {
648
+aio_co_schedule_bh_cb(void *ctx, void *co) "ctx %p co %p"
349
- goto fail;
649
+
350
- }
650
# util/thread-pool.c
351
- bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
651
thread_pool_submit(void *pool, void *req, void *opaque) "pool %p req %p opaque %p"
352
+ assert(align == pad.tail + bytes);
652
thread_pool_complete(void *pool, void *req, void *opaque, int ret) "pool %p req %p opaque %p ret %d"
353
354
- memset(buf, 0, bytes);
355
+ qemu_iovec_init_buf(&local_qiov, pad.tail_buf, align);
356
ret = bdrv_aligned_pwritev(child, req, offset, align, align,
357
&local_qiov, flags & ~BDRV_REQ_ZERO_WRITE);
358
}
359
-fail:
360
- qemu_vfree(buf);
361
+
362
+out:
363
+ bdrv_padding_destroy(&pad);
364
+
365
return ret;
366
-
367
}
368
369
/*
370
@@ -XXX,XX +XXX,XX @@ int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
371
BlockDriverState *bs = child->bs;
372
BdrvTrackedRequest req;
373
uint64_t align = bs->bl.request_alignment;
374
- uint8_t *head_buf = NULL;
375
- uint8_t *tail_buf = NULL;
376
- QEMUIOVector local_qiov;
377
- bool use_local_qiov = false;
378
+ BdrvRequestPadding pad;
379
int ret;
380
381
trace_bdrv_co_pwritev(child->bs, offset, bytes, flags);
382
@@ -XXX,XX +XXX,XX @@ int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
383
goto out;
384
}
385
386
- if (offset & (align - 1)) {
387
- QEMUIOVector head_qiov;
388
-
389
+ if (bdrv_pad_request(bs, &qiov, &offset, &bytes, &pad)) {
390
mark_request_serialising(&req, align);
391
wait_serialising_requests(&req);
392
-
393
- head_buf = qemu_blockalign(bs, align);
394
- qemu_iovec_init_buf(&head_qiov, head_buf, align);
395
-
396
- bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_HEAD);
397
- ret = bdrv_aligned_preadv(child, &req, offset & ~(align - 1), align,
398
- align, &head_qiov, 0);
399
- if (ret < 0) {
400
- goto fail;
401
- }
402
- bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_HEAD);
403
-
404
- qemu_iovec_init(&local_qiov, qiov->niov + 2);
405
- qemu_iovec_add(&local_qiov, head_buf, offset & (align - 1));
406
- qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
407
- use_local_qiov = true;
408
-
409
- bytes += offset & (align - 1);
410
- offset = offset & ~(align - 1);
411
-
412
- /* We have read the tail already if the request is smaller
413
- * than one aligned block.
414
- */
415
- if (bytes < align) {
416
- qemu_iovec_add(&local_qiov, head_buf + bytes, align - bytes);
417
- bytes = align;
418
- }
419
- }
420
-
421
- if ((offset + bytes) & (align - 1)) {
422
- QEMUIOVector tail_qiov;
423
- size_t tail_bytes;
424
- bool waited;
425
-
426
- mark_request_serialising(&req, align);
427
- waited = wait_serialising_requests(&req);
428
- assert(!waited || !use_local_qiov);
429
-
430
- tail_buf = qemu_blockalign(bs, align);
431
- qemu_iovec_init_buf(&tail_qiov, tail_buf, align);
432
-
433
- bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
434
- ret = bdrv_aligned_preadv(child, &req, (offset + bytes) & ~(align - 1),
435
- align, align, &tail_qiov, 0);
436
- if (ret < 0) {
437
- goto fail;
438
- }
439
- bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_AFTER_TAIL);
440
-
441
- if (!use_local_qiov) {
442
- qemu_iovec_init(&local_qiov, qiov->niov + 1);
443
- qemu_iovec_concat(&local_qiov, qiov, 0, qiov->size);
444
- use_local_qiov = true;
445
- }
446
-
447
- tail_bytes = (offset + bytes) & (align - 1);
448
- qemu_iovec_add(&local_qiov, tail_buf + tail_bytes, align - tail_bytes);
449
-
450
- bytes = ROUND_UP(bytes, align);
451
+ bdrv_padding_rmw_read(child, &req, &pad, false);
452
}
453
454
ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
455
- use_local_qiov ? &local_qiov : qiov,
456
- flags);
457
+ qiov, flags);
458
459
-fail:
460
+ bdrv_padding_destroy(&pad);
461
462
- if (use_local_qiov) {
463
- qemu_iovec_destroy(&local_qiov);
464
- }
465
- qemu_vfree(head_buf);
466
- qemu_vfree(tail_buf);
467
out:
468
tracked_request_end(&req);
469
bdrv_dec_in_flight(bs);
470
+
471
return ret;
472
}
473
474
--
653
--
475
2.21.0
654
2.9.3
476
655
477
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
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Implement and use new interface to get rid of hd_qiov.
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.
4
7
5
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
8
While doing the change, replace custom bottom half with aio_co_schedule.
6
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
9
7
Message-id: 20190604161514.262241-12-vsementsov@virtuozzo.com
10
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
8
Message-Id: <20190604161514.262241-12-vsementsov@virtuozzo.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
9
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
15
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
10
---
16
---
11
block/qcow2-cluster.c | 5 +++--
17
block/blkdebug.c | 9 +--------
12
block/qcow2.c | 49 +++++++++++++++++++------------------------
18
1 file changed, 1 insertion(+), 8 deletions(-)
13
2 files changed, 25 insertions(+), 29 deletions(-)
14
19
15
diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
20
diff --git a/block/blkdebug.c b/block/blkdebug.c
16
index XXXXXXX..XXXXXXX 100644
21
index XXXXXXX..XXXXXXX 100644
17
--- a/block/qcow2-cluster.c
22
--- a/block/blkdebug.c
18
+++ b/block/qcow2-cluster.c
23
+++ b/block/blkdebug.c
19
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn do_perform_cow_read(BlockDriverState *bs,
20
* interface. This avoids double I/O throttling and request tracking,
21
* which can lead to deadlock when block layer copy-on-read is enabled.
22
*/
23
- ret = bs->drv->bdrv_co_preadv(bs, src_cluster_offset + offset_in_cluster,
24
- qiov->size, qiov, 0);
25
+ ret = bs->drv->bdrv_co_preadv_part(bs,
26
+ src_cluster_offset + offset_in_cluster,
27
+ qiov->size, qiov, 0, 0);
28
if (ret < 0) {
29
return ret;
30
}
31
diff --git a/block/qcow2.c b/block/qcow2.c
32
index XXXXXXX..XXXXXXX 100644
33
--- a/block/qcow2.c
34
+++ b/block/qcow2.c
35
@@ -XXX,XX +XXX,XX @@ qcow2_co_preadv_compressed(BlockDriverState *bs,
36
uint64_t file_cluster_offset,
37
uint64_t offset,
38
uint64_t bytes,
39
- QEMUIOVector *qiov);
40
+ QEMUIOVector *qiov,
41
+ size_t qiov_offset);
42
43
static int qcow2_probe(const uint8_t *buf, int buf_size, const char *filename)
44
{
45
@@ -XXX,XX +XXX,XX @@ out:
24
@@ -XXX,XX +XXX,XX @@ out:
46
return ret;
25
return ret;
47
}
26
}
48
27
49
-static coroutine_fn int qcow2_co_preadv(BlockDriverState *bs, uint64_t offset,
28
-static void error_callback_bh(void *opaque)
50
- uint64_t bytes, QEMUIOVector *qiov,
29
-{
51
- int flags)
30
- Coroutine *co = opaque;
52
+static coroutine_fn int qcow2_co_preadv_part(BlockDriverState *bs,
31
- qemu_coroutine_enter(co);
53
+ uint64_t offset, uint64_t bytes,
32
-}
54
+ QEMUIOVector *qiov,
33
-
55
+ size_t qiov_offset, int flags)
34
static int inject_error(BlockDriverState *bs, BlkdebugRule *rule)
56
{
35
{
57
BDRVQcow2State *s = bs->opaque;
36
BDRVBlkdebugState *s = bs->opaque;
58
int offset_in_cluster;
37
@@ -XXX,XX +XXX,XX @@ static int inject_error(BlockDriverState *bs, BlkdebugRule *rule)
59
int ret;
60
unsigned int cur_bytes; /* number of bytes in current iteration */
61
uint64_t cluster_offset = 0;
62
- uint64_t bytes_done = 0;
63
- QEMUIOVector hd_qiov;
64
uint8_t *cluster_data = NULL;
65
66
- qemu_iovec_init(&hd_qiov, qiov->niov);
67
-
68
while (bytes != 0) {
69
70
/* prepare next request */
71
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_preadv(BlockDriverState *bs, uint64_t offset,
72
73
offset_in_cluster = offset_into_cluster(s, offset);
74
75
- qemu_iovec_reset(&hd_qiov);
76
- qemu_iovec_concat(&hd_qiov, qiov, bytes_done, cur_bytes);
77
-
78
switch (ret) {
79
case QCOW2_CLUSTER_UNALLOCATED:
80
81
if (bs->backing) {
82
BLKDBG_EVENT(bs->file, BLKDBG_READ_BACKING_AIO);
83
- ret = bdrv_co_preadv(bs->backing, offset, cur_bytes,
84
- &hd_qiov, 0);
85
+ ret = bdrv_co_preadv_part(bs->backing, offset, cur_bytes,
86
+ qiov, qiov_offset, 0);
87
if (ret < 0) {
88
goto fail;
89
}
90
} else {
91
/* Note: in this case, no need to wait */
92
- qemu_iovec_memset(&hd_qiov, 0, 0, cur_bytes);
93
+ qemu_iovec_memset(qiov, qiov_offset, 0, cur_bytes);
94
}
95
break;
96
97
case QCOW2_CLUSTER_ZERO_PLAIN:
98
case QCOW2_CLUSTER_ZERO_ALLOC:
99
- qemu_iovec_memset(&hd_qiov, 0, 0, cur_bytes);
100
+ qemu_iovec_memset(qiov, qiov_offset, 0, cur_bytes);
101
break;
102
103
case QCOW2_CLUSTER_COMPRESSED:
104
ret = qcow2_co_preadv_compressed(bs, cluster_offset,
105
offset, cur_bytes,
106
- &hd_qiov);
107
+ qiov, qiov_offset);
108
if (ret < 0) {
109
goto fail;
110
}
111
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_preadv(BlockDriverState *bs, uint64_t offset,
112
ret = -EIO;
113
goto fail;
114
}
115
- qemu_iovec_from_buf(qiov, bytes_done, cluster_data, cur_bytes);
116
+ qemu_iovec_from_buf(qiov, qiov_offset, cluster_data, cur_bytes);
117
} else {
118
BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
119
- ret = bdrv_co_preadv(s->data_file,
120
- cluster_offset + offset_in_cluster,
121
- cur_bytes, &hd_qiov, 0);
122
+ ret = bdrv_co_preadv_part(s->data_file,
123
+ cluster_offset + offset_in_cluster,
124
+ cur_bytes, qiov, qiov_offset, 0);
125
if (ret < 0) {
126
goto fail;
127
}
128
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_preadv(BlockDriverState *bs, uint64_t offset,
129
130
bytes -= cur_bytes;
131
offset += cur_bytes;
132
- bytes_done += cur_bytes;
133
+ qiov_offset += cur_bytes;
134
}
38
}
135
ret = 0;
39
136
40
if (!immediately) {
137
fail:
41
- aio_bh_schedule_oneshot(bdrv_get_aio_context(bs), error_callback_bh,
138
- qemu_iovec_destroy(&hd_qiov);
42
- qemu_coroutine_self());
139
qemu_vfree(cluster_data);
43
+ aio_co_schedule(qemu_get_current_aio_context(), qemu_coroutine_self());
140
44
qemu_coroutine_yield();
141
return ret;
142
@@ -XXX,XX +XXX,XX @@ qcow2_co_preadv_compressed(BlockDriverState *bs,
143
uint64_t file_cluster_offset,
144
uint64_t offset,
145
uint64_t bytes,
146
- QEMUIOVector *qiov)
147
+ QEMUIOVector *qiov,
148
+ size_t qiov_offset)
149
{
150
BDRVQcow2State *s = bs->opaque;
151
int ret = 0, csize, nb_csectors;
152
@@ -XXX,XX +XXX,XX @@ qcow2_co_preadv_compressed(BlockDriverState *bs,
153
goto fail;
154
}
45
}
155
46
156
- qemu_iovec_from_buf(qiov, 0, out_buf + offset_in_cluster, bytes);
157
+ qemu_iovec_from_buf(qiov, qiov_offset, out_buf + offset_in_cluster, bytes);
158
159
fail:
160
qemu_vfree(out_buf);
161
@@ -XXX,XX +XXX,XX @@ static int qcow2_load_vmstate(BlockDriverState *bs, QEMUIOVector *qiov,
162
BDRVQcow2State *s = bs->opaque;
163
164
BLKDBG_EVENT(bs->file, BLKDBG_VMSTATE_LOAD);
165
- return bs->drv->bdrv_co_preadv(bs, qcow2_vm_state_offset(s) + pos,
166
- qiov->size, qiov, 0);
167
+ return bs->drv->bdrv_co_preadv_part(bs, qcow2_vm_state_offset(s) + pos,
168
+ qiov->size, qiov, 0, 0);
169
}
170
171
/*
172
@@ -XXX,XX +XXX,XX @@ BlockDriver bdrv_qcow2 = {
173
.bdrv_has_zero_init_truncate = bdrv_has_zero_init_1,
174
.bdrv_co_block_status = qcow2_co_block_status,
175
176
- .bdrv_co_preadv = qcow2_co_preadv,
177
+ .bdrv_co_preadv_part = qcow2_co_preadv_part,
178
.bdrv_co_pwritev = qcow2_co_pwritev,
179
.bdrv_co_flush_to_os = qcow2_co_flush_to_os,
180
181
--
47
--
182
2.21.0
48
2.9.3
183
49
184
50
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Introduce new initialization API, to create requests with padding. Will
3
qed_aio_start_io and qed_aio_next_io will not have to acquire/release
4
be used in the following patch. New API uses qemu_iovec_init_buf if
4
the AioContext, while qed_aio_next_io_cb will. Split the functionality
5
resulting io vector has only one element, to avoid extra allocations.
5
and gain a little type-safety in the process.
6
So, we need to update qemu_iovec_destroy to support destroying such
7
QIOVs.
8
6
9
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
7
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
10
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
8
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
11
Message-id: 20190604161514.262241-2-vsementsov@virtuozzo.com
9
Reviewed-by: Fam Zheng <famz@redhat.com>
12
Message-Id: <20190604161514.262241-2-vsementsov@virtuozzo.com>
10
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
11
Message-id: 20170213135235.12274-11-pbonzini@redhat.com
13
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
12
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
14
---
13
---
15
include/qemu/iov.h | 7 +++
14
block/qed.c | 39 +++++++++++++++++++++++++--------------
16
util/iov.c | 112 +++++++++++++++++++++++++++++++++++++++++++--
15
1 file changed, 25 insertions(+), 14 deletions(-)
17
2 files changed, 114 insertions(+), 5 deletions(-)
18
16
19
diff --git a/include/qemu/iov.h b/include/qemu/iov.h
17
diff --git a/block/qed.c b/block/qed.c
20
index XXXXXXX..XXXXXXX 100644
18
index XXXXXXX..XXXXXXX 100644
21
--- a/include/qemu/iov.h
19
--- a/block/qed.c
22
+++ b/include/qemu/iov.h
20
+++ b/block/qed.c
23
@@ -XXX,XX +XXX,XX @@ static inline void *qemu_iovec_buf(QEMUIOVector *qiov)
21
@@ -XXX,XX +XXX,XX @@ static CachedL2Table *qed_new_l2_table(BDRVQEDState *s)
24
22
return l2_table;
25
void qemu_iovec_init(QEMUIOVector *qiov, int alloc_hint);
26
void qemu_iovec_init_external(QEMUIOVector *qiov, struct iovec *iov, int niov);
27
+void qemu_iovec_init_extended(
28
+ QEMUIOVector *qiov,
29
+ void *head_buf, size_t head_len,
30
+ QEMUIOVector *mid_qiov, size_t mid_offset, size_t mid_len,
31
+ void *tail_buf, size_t tail_len);
32
+void qemu_iovec_init_slice(QEMUIOVector *qiov, QEMUIOVector *source,
33
+ size_t offset, size_t len);
34
void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len);
35
void qemu_iovec_concat(QEMUIOVector *dst,
36
QEMUIOVector *src, size_t soffset, size_t sbytes);
37
diff --git a/util/iov.c b/util/iov.c
38
index XXXXXXX..XXXXXXX 100644
39
--- a/util/iov.c
40
+++ b/util/iov.c
41
@@ -XXX,XX +XXX,XX @@ void qemu_iovec_concat(QEMUIOVector *dst,
42
qemu_iovec_concat_iov(dst, src->iov, src->niov, soffset, sbytes);
43
}
23
}
44
24
45
+/*
25
-static void qed_aio_next_io(void *opaque, int ret);
46
+ * qiov_find_iov
26
+static void qed_aio_next_io(QEDAIOCB *acb, int ret);
47
+ *
27
+
48
+ * Return pointer to iovec structure, where byte at @offset in original vector
28
+static void qed_aio_start_io(QEDAIOCB *acb)
49
+ * @iov exactly is.
50
+ * Set @remaining_offset to be offset inside that iovec to the same byte.
51
+ */
52
+static struct iovec *iov_skip_offset(struct iovec *iov, size_t offset,
53
+ size_t *remaining_offset)
54
+{
29
+{
55
+ while (offset > 0 && offset >= iov->iov_len) {
30
+ qed_aio_next_io(acb, 0);
56
+ offset -= iov->iov_len;
57
+ iov++;
58
+ }
59
+ *remaining_offset = offset;
60
+
61
+ return iov;
62
+}
31
+}
63
+
32
+
64
+/*
33
+static void qed_aio_next_io_cb(void *opaque, int ret)
65
+ * qiov_slice
66
+ *
67
+ * Find subarray of iovec's, containing requested range. @head would
68
+ * be offset in first iov (returned by the function), @tail would be
69
+ * count of extra bytes in last iovec (returned iov + @niov - 1).
70
+ */
71
+static struct iovec *qiov_slice(QEMUIOVector *qiov,
72
+ size_t offset, size_t len,
73
+ size_t *head, size_t *tail, int *niov)
74
+{
34
+{
75
+ struct iovec *iov, *end_iov;
35
+ QEDAIOCB *acb = opaque;
76
+
36
+
77
+ assert(offset + len <= qiov->size);
37
+ qed_aio_next_io(acb, ret);
78
+
79
+ iov = iov_skip_offset(qiov->iov, offset, head);
80
+ end_iov = iov_skip_offset(iov, *head + len, tail);
81
+
82
+ if (*tail > 0) {
83
+ assert(*tail < end_iov->iov_len);
84
+ *tail = end_iov->iov_len - *tail;
85
+ end_iov++;
86
+ }
87
+
88
+ *niov = end_iov - iov;
89
+
90
+ return iov;
91
+}
38
+}
92
+
39
93
+/*
40
static void qed_plug_allocating_write_reqs(BDRVQEDState *s)
94
+ * Compile new iovec, combining @head_buf buffer, sub-qiov of @mid_qiov,
41
{
95
+ * and @tail_buf buffer into new qiov.
42
@@ -XXX,XX +XXX,XX @@ static void qed_unplug_allocating_write_reqs(BDRVQEDState *s)
96
+ */
43
97
+void qemu_iovec_init_extended(
44
acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
98
+ QEMUIOVector *qiov,
45
if (acb) {
99
+ void *head_buf, size_t head_len,
46
- qed_aio_next_io(acb, 0);
100
+ QEMUIOVector *mid_qiov, size_t mid_offset, size_t mid_len,
47
+ qed_aio_start_io(acb);
101
+ void *tail_buf, size_t tail_len)
48
}
102
+{
49
}
103
+ size_t mid_head, mid_tail;
50
104
+ int total_niov, mid_niov = 0;
51
@@ -XXX,XX +XXX,XX @@ static void qed_aio_complete(QEDAIOCB *acb, int ret)
105
+ struct iovec *p, *mid_iov;
52
QSIMPLEQ_REMOVE_HEAD(&s->allocating_write_reqs, next);
106
+
53
acb = QSIMPLEQ_FIRST(&s->allocating_write_reqs);
107
+ if (mid_len) {
54
if (acb) {
108
+ mid_iov = qiov_slice(mid_qiov, mid_offset, mid_len,
55
- qed_aio_next_io(acb, 0);
109
+ &mid_head, &mid_tail, &mid_niov);
56
+ qed_aio_start_io(acb);
110
+ }
57
} else if (s->header.features & QED_F_NEED_CHECK) {
111
+
58
qed_start_need_check_timer(s);
112
+ total_niov = !!head_len + mid_niov + !!tail_len;
59
}
113
+ if (total_niov == 1) {
60
@@ -XXX,XX +XXX,XX @@ static void qed_commit_l2_update(void *opaque, int ret)
114
+ qemu_iovec_init_buf(qiov, NULL, 0);
61
acb->request.l2_table = qed_find_l2_cache_entry(&s->l2_cache, l2_offset);
115
+ p = &qiov->local_iov;
62
assert(acb->request.l2_table != NULL);
116
+ } else {
63
117
+ qiov->niov = qiov->nalloc = total_niov;
64
- qed_aio_next_io(opaque, ret);
118
+ qiov->size = head_len + mid_len + tail_len;
65
+ qed_aio_next_io(acb, ret);
119
+ p = qiov->iov = g_new(struct iovec, qiov->niov);
66
}
120
+ }
67
121
+
68
/**
122
+ if (head_len) {
69
@@ -XXX,XX +XXX,XX @@ static void qed_aio_write_l2_update(QEDAIOCB *acb, int ret, uint64_t offset)
123
+ p->iov_base = head_buf;
70
if (need_alloc) {
124
+ p->iov_len = head_len;
71
/* Write out the whole new L2 table */
125
+ p++;
72
qed_write_l2_table(s, &acb->request, 0, s->table_nelems, true,
126
+ }
73
- qed_aio_write_l1_update, acb);
127
+
74
+ qed_aio_write_l1_update, acb);
128
+ if (mid_len) {
75
} else {
129
+ memcpy(p, mid_iov, mid_niov * sizeof(*p));
76
/* Write out only the updated part of the L2 table */
130
+ p[0].iov_base = (uint8_t *)p[0].iov_base + mid_head;
77
qed_write_l2_table(s, &acb->request, index, acb->cur_nclusters, false,
131
+ p[0].iov_len -= mid_head;
78
- qed_aio_next_io, acb);
132
+ p[mid_niov - 1].iov_len -= mid_tail;
79
+ qed_aio_next_io_cb, acb);
133
+ p += mid_niov;
80
}
134
+ }
81
return;
135
+
82
136
+ if (tail_len) {
83
@@ -XXX,XX +XXX,XX @@ static void qed_aio_write_main(void *opaque, int ret)
137
+ p->iov_base = tail_buf;
84
}
138
+ p->iov_len = tail_len;
85
139
+ }
86
if (acb->find_cluster_ret == QED_CLUSTER_FOUND) {
140
+}
87
- next_fn = qed_aio_next_io;
141
+
88
+ next_fn = qed_aio_next_io_cb;
142
/*
89
} else {
143
* Check if the contents of the iovecs are all zero
90
if (s->bs->backing) {
91
next_fn = qed_aio_write_flush_before_l2_update;
92
@@ -XXX,XX +XXX,XX @@ static void qed_aio_write_alloc(QEDAIOCB *acb, size_t len)
93
if (acb->flags & QED_AIOCB_ZERO) {
94
/* Skip ahead if the clusters are already zero */
95
if (acb->find_cluster_ret == QED_CLUSTER_ZERO) {
96
- qed_aio_next_io(acb, 0);
97
+ qed_aio_start_io(acb);
98
return;
99
}
100
101
@@ -XXX,XX +XXX,XX @@ static void qed_aio_read_data(void *opaque, int ret,
102
/* Handle zero cluster and backing file reads */
103
if (ret == QED_CLUSTER_ZERO) {
104
qemu_iovec_memset(&acb->cur_qiov, 0, 0, acb->cur_qiov.size);
105
- qed_aio_next_io(acb, 0);
106
+ qed_aio_start_io(acb);
107
return;
108
} else if (ret != QED_CLUSTER_FOUND) {
109
qed_read_backing_file(s, acb->cur_pos, &acb->cur_qiov,
110
- &acb->backing_qiov, qed_aio_next_io, acb);
111
+ &acb->backing_qiov, qed_aio_next_io_cb, acb);
112
return;
113
}
114
115
BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
116
bdrv_aio_readv(bs->file, offset / BDRV_SECTOR_SIZE,
117
&acb->cur_qiov, acb->cur_qiov.size / BDRV_SECTOR_SIZE,
118
- qed_aio_next_io, acb);
119
+ qed_aio_next_io_cb, acb);
120
return;
121
122
err:
123
@@ -XXX,XX +XXX,XX @@ err:
124
/**
125
* Begin next I/O or complete the request
144
*/
126
*/
145
@@ -XXX,XX +XXX,XX @@ bool qemu_iovec_is_zero(QEMUIOVector *qiov)
127
-static void qed_aio_next_io(void *opaque, int ret)
146
return true;
128
+static void qed_aio_next_io(QEDAIOCB *acb, int ret)
129
{
130
- QEDAIOCB *acb = opaque;
131
BDRVQEDState *s = acb_to_s(acb);
132
QEDFindClusterFunc *io_fn = (acb->flags & QED_AIOCB_WRITE) ?
133
qed_aio_write_data : qed_aio_read_data;
134
@@ -XXX,XX +XXX,XX @@ static BlockAIOCB *qed_aio_setup(BlockDriverState *bs,
135
qemu_iovec_init(&acb->cur_qiov, qiov->niov);
136
137
/* Start request */
138
- qed_aio_next_io(acb, 0);
139
+ qed_aio_start_io(acb);
140
return &acb->common;
147
}
141
}
148
142
149
+void qemu_iovec_init_slice(QEMUIOVector *qiov, QEMUIOVector *source,
150
+ size_t offset, size_t len)
151
+{
152
+ qemu_iovec_init_extended(qiov, NULL, 0, source, offset, len, NULL, 0);
153
+}
154
+
155
void qemu_iovec_destroy(QEMUIOVector *qiov)
156
{
157
- assert(qiov->nalloc != -1);
158
+ if (qiov->nalloc != -1) {
159
+ g_free(qiov->iov);
160
+ }
161
162
- qemu_iovec_reset(qiov);
163
- g_free(qiov->iov);
164
- qiov->nalloc = 0;
165
- qiov->iov = NULL;
166
+ memset(qiov, 0, sizeof(*qiov));
167
}
168
169
void qemu_iovec_reset(QEMUIOVector *qiov)
170
--
143
--
171
2.21.0
144
2.9.3
172
145
173
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
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Allocate bounce_buffer only if it is really needed. Also, sub-optimize
3
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
4
allocation size (why not?).
4
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
5
5
Reviewed-by: Fam Zheng <famz@redhat.com>
6
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
6
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
7
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
7
Message-id: 20170213135235.12274-13-pbonzini@redhat.com
8
Message-id: 20190604161514.262241-7-vsementsov@virtuozzo.com
9
Message-Id: <20190604161514.262241-7-vsementsov@virtuozzo.com>
10
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
8
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
11
---
9
---
12
block/io.c | 21 ++++++++++++---------
10
block/qed.h | 3 +++
13
1 file changed, 12 insertions(+), 9 deletions(-)
11
block/curl.c | 2 ++
14
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
15
diff --git a/block/io.c b/block/io.c
52
diff --git a/block/io.c b/block/io.c
16
index XXXXXXX..XXXXXXX 100644
53
index XXXXXXX..XXXXXXX 100644
17
--- a/block/io.c
54
--- a/block/io.c
18
+++ b/block/io.c
55
+++ b/block/io.c
19
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
56
@@ -XXX,XX +XXX,XX @@ void bdrv_aio_cancel(BlockAIOCB *acb)
20
* modifying the image file. This is critical for zero-copy guest I/O
57
if (acb->aiocb_info->get_aio_context) {
21
* where anything might happen inside guest memory.
58
aio_poll(acb->aiocb_info->get_aio_context(acb), true);
22
*/
59
} else if (acb->bs) {
23
- void *bounce_buffer;
60
+ /* qemu_aio_ref and qemu_aio_unref are not thread-safe, so
24
+ void *bounce_buffer = NULL;
61
+ * assert that we're not using an I/O thread. Thread-safe
25
62
+ * code should use bdrv_aio_cancel_async exclusively.
26
BlockDriver *drv = bs->drv;
63
+ */
27
int64_t cluster_offset;
64
+ assert(bdrv_get_aio_context(acb->bs) == qemu_get_aio_context());
28
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
65
aio_poll(bdrv_get_aio_context(acb->bs), true);
29
trace_bdrv_co_do_copy_on_readv(bs, offset, bytes,
66
} else {
30
cluster_offset, cluster_bytes);
67
abort();
31
68
diff --git a/block/iscsi.c b/block/iscsi.c
32
- bounce_buffer = qemu_try_blockalign(bs,
69
index XXXXXXX..XXXXXXX 100644
33
- MIN(MIN(max_transfer, cluster_bytes),
70
--- a/block/iscsi.c
34
- MAX_BOUNCE_BUFFER));
71
+++ b/block/iscsi.c
35
- if (bounce_buffer == NULL) {
72
@@ -XXX,XX +XXX,XX @@ static void iscsi_retry_timer_expired(void *opaque)
36
- ret = -ENOMEM;
73
struct IscsiTask *iTask = opaque;
37
- goto err;
74
iTask->complete = 1;
38
- }
75
if (iTask->co) {
39
-
76
- qemu_coroutine_enter(iTask->co);
40
while (cluster_bytes) {
77
+ aio_co_wake(iTask->co);
41
int64_t pnum;
78
}
42
79
}
43
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
80
44
81
@@ -XXX,XX +XXX,XX @@ static void iscsi_nop_timed_event(void *opaque)
45
/* Must copy-on-read; use the bounce buffer */
82
{
46
pnum = MIN(pnum, MAX_BOUNCE_BUFFER);
83
IscsiLun *iscsilun = opaque;
47
+ if (!bounce_buffer) {
84
48
+ int64_t max_we_need = MAX(pnum, cluster_bytes - pnum);
85
+ aio_context_acquire(iscsilun->aio_context);
49
+ int64_t max_allowed = MIN(max_transfer, MAX_BOUNCE_BUFFER);
86
if (iscsi_get_nops_in_flight(iscsilun->iscsi) >= MAX_NOP_FAILURES) {
50
+ int64_t bounce_buffer_len = MIN(max_we_need, max_allowed);
87
error_report("iSCSI: NOP timeout. Reconnecting...");
51
+
88
iscsilun->request_timed_out = true;
52
+ bounce_buffer = qemu_try_blockalign(bs, bounce_buffer_len);
89
} else if (iscsi_nop_out_async(iscsilun->iscsi, NULL, NULL, 0, NULL) != 0) {
53
+ if (!bounce_buffer) {
90
error_report("iSCSI: failed to sent NOP-Out. Disabling NOP messages.");
54
+ ret = -ENOMEM;
91
- return;
55
+ goto err;
92
+ goto out;
56
+ }
93
}
57
+ }
94
58
qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum);
95
timer_mod(iscsilun->nop_timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) + NOP_INTERVAL);
59
96
iscsi_set_events(iscsilun);
60
ret = bdrv_driver_preadv(bs, cluster_offset, pnum,
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,
61
--
201
--
62
2.21.0
202
2.9.3
63
203
64
204
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Use buffer based io in encrypted case.
3
This covers both file descriptor callbacks and polling callbacks,
4
since they execute related code.
4
5
5
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
6
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
6
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
7
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
7
Message-id: 20190604161514.262241-11-vsementsov@virtuozzo.com
8
Reviewed-by: Fam Zheng <famz@redhat.com>
8
Message-Id: <20190604161514.262241-11-vsementsov@virtuozzo.com>
9
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
10
Message-id: 20170213135235.12274-14-pbonzini@redhat.com
9
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
11
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
10
---
12
---
11
block/qcow2.c | 28 ++++++++++++++++------------
13
block/curl.c | 16 +++++++++++++---
12
1 file changed, 16 insertions(+), 12 deletions(-)
14
block/iscsi.c | 4 ++++
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(-)
13
25
14
diff --git a/block/qcow2.c b/block/qcow2.c
26
diff --git a/block/curl.c b/block/curl.c
15
index XXXXXXX..XXXXXXX 100644
27
index XXXXXXX..XXXXXXX 100644
16
--- a/block/qcow2.c
28
--- a/block/curl.c
17
+++ b/block/qcow2.c
29
+++ b/block/curl.c
18
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_preadv(BlockDriverState *bs, uint64_t offset,
30
@@ -XXX,XX +XXX,XX @@ static void curl_multi_check_completion(BDRVCURLState *s)
19
}
31
}
20
32
}
21
assert(cur_bytes <= QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);
33
22
- qemu_iovec_reset(&hd_qiov);
34
-static void curl_multi_do(void *arg)
23
- qemu_iovec_add(&hd_qiov, cluster_data, cur_bytes);
35
+static void curl_multi_do_locked(CURLState *s)
24
- }
36
{
25
37
- CURLState *s = (CURLState *)arg;
26
- BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
38
CURLSocket *socket, *next_socket;
27
- ret = bdrv_co_preadv(s->data_file,
39
int running;
28
- cluster_offset + offset_in_cluster,
40
int r;
29
- cur_bytes, &hd_qiov, 0);
41
@@ -XXX,XX +XXX,XX @@ static void curl_multi_do(void *arg)
30
- if (ret < 0) {
42
}
31
- goto fail;
43
}
32
- }
44
33
- if (bs->encrypted) {
45
+static void curl_multi_do(void *arg)
34
- assert(s->crypto);
46
+{
35
+ BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
47
+ CURLState *s = (CURLState *)arg;
36
+ ret = bdrv_co_pread(s->data_file,
48
+
37
+ cluster_offset + offset_in_cluster,
49
+ aio_context_acquire(s->s->aio_context);
38
+ cur_bytes, cluster_data, 0);
50
+ curl_multi_do_locked(s);
39
+ if (ret < 0) {
51
+ aio_context_release(s->s->aio_context);
40
+ goto fail;
52
+}
41
+ }
53
+
42
+
54
static void curl_multi_read(void *arg)
43
assert((offset & (BDRV_SECTOR_SIZE - 1)) == 0);
55
{
44
assert((cur_bytes & (BDRV_SECTOR_SIZE - 1)) == 0);
56
CURLState *s = (CURLState *)arg;
45
if (qcow2_co_decrypt(bs, cluster_offset, offset,
57
46
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_preadv(BlockDriverState *bs, uint64_t offset,
58
- curl_multi_do(arg);
47
goto fail;
59
+ aio_context_acquire(s->s->aio_context);
48
}
60
+ curl_multi_do_locked(s);
49
qemu_iovec_from_buf(qiov, bytes_done, cluster_data, cur_bytes);
61
curl_multi_check_completion(s->s);
50
+ } else {
62
+ aio_context_release(s->s->aio_context);
51
+ BLKDBG_EVENT(bs->file, BLKDBG_READ_AIO);
63
}
52
+ ret = bdrv_co_preadv(s->data_file,
64
53
+ cluster_offset + offset_in_cluster,
65
static void curl_multi_timeout_do(void *arg)
54
+ cur_bytes, &hd_qiov, 0);
66
diff --git a/block/iscsi.c b/block/iscsi.c
55
+ if (ret < 0) {
67
index XXXXXXX..XXXXXXX 100644
56
+ goto fail;
68
--- a/block/iscsi.c
57
+ }
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);
334
diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c
335
index XXXXXXX..XXXXXXX 100644
336
--- a/hw/block/virtio-blk.c
337
+++ b/hw/block/virtio-blk.c
338
@@ -XXX,XX +XXX,XX @@ static void virtio_blk_ioctl_complete(void *opaque, int status)
339
{
340
VirtIOBlockIoctlReq *ioctl_req = opaque;
341
VirtIOBlockReq *req = ioctl_req->req;
342
- VirtIODevice *vdev = VIRTIO_DEVICE(req->dev);
343
+ VirtIOBlock *s = req->dev;
344
+ VirtIODevice *vdev = VIRTIO_DEVICE(s);
345
struct virtio_scsi_inhdr *scsi;
346
struct sg_io_hdr *hdr;
347
348
@@ -XXX,XX +XXX,XX @@ bool virtio_blk_handle_vq(VirtIOBlock *s, VirtQueue *vq)
349
MultiReqBuffer mrb = {};
350
bool progress = false;
351
352
+ aio_context_acquire(blk_get_aio_context(s->blk));
353
blk_io_plug(s->blk);
354
355
do {
356
@@ -XXX,XX +XXX,XX @@ bool virtio_blk_handle_vq(VirtIOBlock *s, VirtQueue *vq)
357
}
358
359
blk_io_unplug(s->blk);
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;
58
}
468
}
59
break;
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
}
60
475
61
--
476
--
62
2.21.0
477
2.9.3
63
478
64
479
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Use and support new API in bdrv_co_do_copy_on_readv. Note that in case
3
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
4
of allocated-in-top we need to shrink read size to MIN(..) by hand, as
4
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
5
pre-patch this was actually done implicitly by qemu_iovec_concat (and
5
Reviewed-by: Fam Zheng <famz@redhat.com>
6
we used local_qiov.size).
6
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
7
7
Message-id: 20170213135235.12274-15-pbonzini@redhat.com
8
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
9
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
10
Message-id: 20190604161514.262241-6-vsementsov@virtuozzo.com
11
Message-Id: <20190604161514.262241-6-vsementsov@virtuozzo.com>
12
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
8
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
13
---
9
---
14
block/io.c | 18 +++++++++---------
10
block/archipelago.c | 3 +++
15
1 file changed, 9 insertions(+), 9 deletions(-)
11
block/blkreplay.c | 2 +-
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(-)
16
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/blkreplay.c b/block/blkreplay.c
46
index XXXXXXX..XXXXXXX 100755
47
--- a/block/blkreplay.c
48
+++ b/block/blkreplay.c
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)
17
diff --git a/block/io.c b/block/io.c
186
diff --git a/block/io.c b/block/io.c
18
index XXXXXXX..XXXXXXX 100644
187
index XXXXXXX..XXXXXXX 100644
19
--- a/block/io.c
188
--- a/block/io.c
20
+++ b/block/io.c
189
+++ b/block/io.c
21
@@ -XXX,XX +XXX,XX @@ bdrv_driver_pwritev_compressed(BlockDriverState *bs, uint64_t offset,
190
@@ -XXX,XX +XXX,XX @@ static void bdrv_co_drain_bh_cb(void *opaque)
22
191
bdrv_dec_in_flight(bs);
23
static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
192
bdrv_drained_begin(bs);
24
int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
193
data->done = true;
25
- int flags)
194
- qemu_coroutine_enter(co);
26
+ size_t qiov_offset, int flags)
195
+ aio_co_wake(co);
27
{
196
}
28
BlockDriverState *bs = child->bs;
197
29
198
static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs)
30
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
199
@@ -XXX,XX +XXX,XX @@ static void bdrv_co_complete(BlockAIOCBCoroutine *acb)
31
void *bounce_buffer;
200
static void bdrv_co_em_bh(void *opaque)
32
201
{
33
BlockDriver *drv = bs->drv;
202
BlockAIOCBCoroutine *acb = opaque;
34
- QEMUIOVector local_qiov;
203
+ BlockDriverState *bs = acb->common.bs;
35
int64_t cluster_offset;
204
+ AioContext *ctx = bdrv_get_aio_context(bs);
36
int64_t cluster_bytes;
205
37
size_t skip_bytes;
206
assert(!acb->need_bh);
38
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
207
+ aio_context_acquire(ctx);
39
assert(skip_bytes < pnum);
208
bdrv_co_complete(acb);
40
209
+ aio_context_release(ctx);
41
if (ret <= 0) {
210
}
42
+ QEMUIOVector local_qiov;
211
43
+
212
static void bdrv_co_maybe_schedule_bh(BlockAIOCBCoroutine *acb)
44
/* Must copy-on-read; use the bounce buffer */
213
diff --git a/block/iscsi.c b/block/iscsi.c
45
pnum = MIN(pnum, MAX_BOUNCE_BUFFER);
214
index XXXXXXX..XXXXXXX 100644
46
qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum);
215
--- a/block/iscsi.c
47
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
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
418
diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c
419
index XXXXXXX..XXXXXXX 100644
420
--- a/hw/block/virtio-blk.c
421
+++ b/hw/block/virtio-blk.c
422
@@ -XXX,XX +XXX,XX @@ static void virtio_blk_dma_restart_bh(void *opaque)
423
424
s->rq = NULL;
425
426
+ aio_context_acquire(blk_get_aio_context(s->conf.conf.blk));
427
while (req) {
428
VirtIOBlockReq *next = req->next;
429
if (virtio_blk_handle_request(req, &mrb)) {
430
@@ -XXX,XX +XXX,XX @@ static void virtio_blk_dma_restart_bh(void *opaque)
431
if (mrb.num_reqs) {
432
virtio_blk_submit_multireq(s->blk, &mrb);
433
}
434
+ aio_context_release(blk_get_aio_context(s->conf.conf.blk));
435
}
436
437
static void virtio_blk_dma_restart_cb(void *opaque, int running,
438
diff --git a/hw/scsi/scsi-bus.c b/hw/scsi/scsi-bus.c
439
index XXXXXXX..XXXXXXX 100644
440
--- a/hw/scsi/scsi-bus.c
441
+++ b/hw/scsi/scsi-bus.c
442
@@ -XXX,XX +XXX,XX @@ static void scsi_dma_restart_bh(void *opaque)
443
qemu_bh_delete(s->bh);
444
s->bh = NULL;
445
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)
451
}
452
scsi_req_unref(req);
453
}
454
+ aio_context_release(blk_get_aio_context(s->conf.blk));
455
}
456
457
void scsi_req_retry(SCSIRequest *req)
458
diff --git a/util/async.c b/util/async.c
459
index XXXXXXX..XXXXXXX 100644
460
--- a/util/async.c
461
+++ b/util/async.c
462
@@ -XXX,XX +XXX,XX @@ int aio_bh_poll(AioContext *ctx)
463
ret = 1;
48
}
464
}
49
465
bh->idle = 0;
50
if (!(flags & BDRV_REQ_PREFETCH)) {
466
- aio_context_acquire(ctx);
51
- qemu_iovec_from_buf(qiov, progress, bounce_buffer + skip_bytes,
467
aio_bh_call(bh);
52
+ qemu_iovec_from_buf(qiov, qiov_offset + progress,
468
- aio_context_release(ctx);
53
+ bounce_buffer + skip_bytes,
54
pnum - skip_bytes);
55
}
56
} else if (!(flags & BDRV_REQ_PREFETCH)) {
57
/* Read directly into the destination */
58
- qemu_iovec_init(&local_qiov, qiov->niov);
59
- qemu_iovec_concat(&local_qiov, qiov, progress, pnum - skip_bytes);
60
- ret = bdrv_driver_preadv(bs, offset + progress, local_qiov.size,
61
- &local_qiov, 0, 0);
62
- qemu_iovec_destroy(&local_qiov);
63
+ ret = bdrv_driver_preadv(bs, offset + progress,
64
+ MIN(pnum - skip_bytes, bytes - progress),
65
+ qiov, qiov_offset + progress, 0);
66
if (ret < 0) {
67
goto err;
68
}
69
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
70
}
469
}
71
470
if (bh->deleted) {
72
if (!ret || pnum != bytes) {
471
deleted = true;
73
- ret = bdrv_co_do_copy_on_readv(child, offset, bytes, qiov, flags);
472
@@ -XXX,XX +XXX,XX @@ static void co_schedule_bh_cb(void *opaque)
74
+ ret = bdrv_co_do_copy_on_readv(child, offset, bytes, qiov, 0, flags);
473
Coroutine *co = QSLIST_FIRST(&straight);
75
goto out;
474
QSLIST_REMOVE_HEAD(&straight, co_scheduled_next);
76
} else if (flags & BDRV_REQ_PREFETCH) {
475
trace_aio_co_schedule_bh_cb(ctx, co);
77
goto out;
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)
78
--
502
--
79
2.21.0
503
2.9.3
80
504
81
505
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Add handlers supporting qiov_offset parameter:
3
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
4
bdrv_co_preadv_part
4
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
5
bdrv_co_pwritev_part
5
Reviewed-by: Fam Zheng <famz@redhat.com>
6
bdrv_co_pwritev_compressed_part
6
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
7
This is used to reduce need of defining local_qiovs and hd_qiovs in all
7
Message-id: 20170213135235.12274-16-pbonzini@redhat.com
8
corners of block layer code. The following patches will increase usage
9
of this new API part by part.
10
11
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
12
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
13
Message-id: 20190604161514.262241-5-vsementsov@virtuozzo.com
14
Message-Id: <20190604161514.262241-5-vsementsov@virtuozzo.com>
15
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
8
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
16
---
9
---
17
include/block/block_int.h | 15 ++++++
10
block/archipelago.c | 3 ---
18
block/backup.c | 2 +-
11
block/block-backend.c | 7 -------
19
block/io.c | 96 +++++++++++++++++++++++++++++++--------
12
block/curl.c | 2 +-
20
qemu-img.c | 4 +-
13
block/io.c | 6 +-----
21
4 files changed, 95 insertions(+), 22 deletions(-)
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(-)
22
28
23
diff --git a/include/block/block_int.h b/include/block/block_int.h
29
diff --git a/block/archipelago.c b/block/archipelago.c
24
index XXXXXXX..XXXXXXX 100644
30
index XXXXXXX..XXXXXXX 100644
25
--- a/include/block/block_int.h
31
--- a/block/archipelago.c
26
+++ b/include/block/block_int.h
32
+++ b/block/archipelago.c
27
@@ -XXX,XX +XXX,XX @@ struct BlockDriver {
33
@@ -XXX,XX +XXX,XX @@ static void qemu_archipelago_complete_aio(void *opaque)
28
*/
34
{
29
int coroutine_fn (*bdrv_co_preadv)(BlockDriverState *bs,
35
AIORequestData *reqdata = (AIORequestData *) opaque;
30
uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags);
36
ArchipelagoAIOCB *aio_cb = (ArchipelagoAIOCB *) reqdata->aio_cb;
31
+ int coroutine_fn (*bdrv_co_preadv_part)(BlockDriverState *bs,
37
- AioContext *ctx = bdrv_get_aio_context(aio_cb->common.bs);
32
+ uint64_t offset, uint64_t bytes,
38
33
+ QEMUIOVector *qiov, size_t qiov_offset, int flags);
39
- aio_context_acquire(ctx);
34
int coroutine_fn (*bdrv_co_writev)(BlockDriverState *bs,
40
aio_cb->common.cb(aio_cb->common.opaque, aio_cb->ret);
35
int64_t sector_num, int nb_sectors, QEMUIOVector *qiov, int flags);
41
- aio_context_release(ctx);
36
/**
42
aio_cb->status = 0;
37
@@ -XXX,XX +XXX,XX @@ struct BlockDriver {
43
38
*/
44
qemu_aio_unref(aio_cb);
39
int coroutine_fn (*bdrv_co_pwritev)(BlockDriverState *bs,
45
diff --git a/block/block-backend.c b/block/block-backend.c
40
uint64_t offset, uint64_t bytes, QEMUIOVector *qiov, int flags);
46
index XXXXXXX..XXXXXXX 100644
41
+ int coroutine_fn (*bdrv_co_pwritev_part)(BlockDriverState *bs,
47
--- a/block/block-backend.c
42
+ uint64_t offset, uint64_t bytes,
48
+++ b/block/block-backend.c
43
+ QEMUIOVector *qiov, size_t qiov_offset, int flags);
49
@@ -XXX,XX +XXX,XX @@ int blk_make_zero(BlockBackend *blk, BdrvRequestFlags flags)
44
50
static void error_callback_bh(void *opaque)
45
/*
51
{
46
* Efficiently zero a region of the disk image. Typically an image format
52
struct BlockBackendAIOCB *acb = opaque;
47
@@ -XXX,XX +XXX,XX @@ struct BlockDriver {
53
- AioContext *ctx = bdrv_get_aio_context(acb->common.bs);
48
54
49
int coroutine_fn (*bdrv_co_pwritev_compressed)(BlockDriverState *bs,
55
bdrv_dec_in_flight(acb->common.bs);
50
uint64_t offset, uint64_t bytes, QEMUIOVector *qiov);
56
- aio_context_acquire(ctx);
51
+ int coroutine_fn (*bdrv_co_pwritev_compressed_part)(BlockDriverState *bs,
57
acb->common.cb(acb->common.opaque, acb->ret);
52
+ uint64_t offset, uint64_t bytes, QEMUIOVector *qiov,
58
- aio_context_release(ctx);
53
+ size_t qiov_offset);
59
qemu_aio_unref(acb);
54
60
}
55
int (*bdrv_snapshot_create)(BlockDriverState *bs,
61
56
QEMUSnapshotInfo *sn_info);
62
@@ -XXX,XX +XXX,XX @@ static void blk_aio_complete(BlkAioEmAIOCB *acb)
57
@@ -XXX,XX +XXX,XX @@ struct BlockDriver {
63
static void blk_aio_complete_bh(void *opaque)
58
const char *const *strong_runtime_opts;
64
{
59
};
65
BlkAioEmAIOCB *acb = opaque;
60
66
- AioContext *ctx = bdrv_get_aio_context(acb->common.bs);
61
+static inline bool block_driver_can_compress(BlockDriver *drv)
67
-
62
+{
68
assert(acb->has_returned);
63
+ return drv->bdrv_co_pwritev_compressed ||
69
- aio_context_acquire(ctx);
64
+ drv->bdrv_co_pwritev_compressed_part;
70
blk_aio_complete(acb);
65
+}
71
- aio_context_release(ctx);
66
+
72
}
67
typedef struct BlockLimits {
73
68
/* Alignment requirement, in bytes, for offset/length of I/O
74
static BlockAIOCB *blk_aio_prwv(BlockBackend *blk, int64_t offset, int bytes,
69
* requests. Must be a power of 2 less than INT_MAX; defaults to
75
diff --git a/block/curl.c b/block/curl.c
70
diff --git a/block/backup.c b/block/backup.c
76
index XXXXXXX..XXXXXXX 100644
71
index XXXXXXX..XXXXXXX 100644
77
--- a/block/curl.c
72
--- a/block/backup.c
78
+++ b/block/curl.c
73
+++ b/block/backup.c
79
@@ -XXX,XX +XXX,XX @@ static void curl_readv_bh_cb(void *p)
74
@@ -XXX,XX +XXX,XX @@ BlockJob *backup_job_create(const char *job_id, BlockDriverState *bs,
80
curl_multi_socket_action(s->multi, CURL_SOCKET_TIMEOUT, 0, &running);
75
return NULL;
81
76
}
82
out:
77
83
+ aio_context_release(ctx);
78
- if (compress && target->drv->bdrv_co_pwritev_compressed == NULL) {
84
if (ret != -EINPROGRESS) {
79
+ if (compress && !block_driver_can_compress(target->drv)) {
85
acb->common.cb(acb->common.opaque, ret);
80
error_setg(errp, "Compression is not supported for this drive %s",
86
qemu_aio_unref(acb);
81
bdrv_get_device_name(target));
87
}
82
return NULL;
88
- aio_context_release(ctx);
89
}
90
91
static BlockAIOCB *curl_aio_readv(BlockDriverState *bs,
83
diff --git a/block/io.c b/block/io.c
92
diff --git a/block/io.c b/block/io.c
84
index XXXXXXX..XXXXXXX 100644
93
index XXXXXXX..XXXXXXX 100644
85
--- a/block/io.c
94
--- a/block/io.c
86
+++ b/block/io.c
95
+++ b/block/io.c
87
@@ -XXX,XX +XXX,XX @@ void bdrv_refresh_limits(BlockDriverState *bs, Error **errp)
88
89
/* Default alignment based on whether driver has byte interface */
90
bs->bl.request_alignment = (drv->bdrv_co_preadv ||
91
- drv->bdrv_aio_preadv) ? 1 : 512;
92
+ drv->bdrv_aio_preadv ||
93
+ drv->bdrv_co_preadv_part) ? 1 : 512;
94
95
/* Take some limits from the children as a default */
96
if (bs->file) {
97
@@ -XXX,XX +XXX,XX @@ static void bdrv_co_io_em_complete(void *opaque, int ret)
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
}
98
103
99
static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
104
static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
100
uint64_t offset, uint64_t bytes,
105
@@ -XXX,XX +XXX,XX @@ static void bdrv_co_complete(BlockAIOCBCoroutine *acb)
101
- QEMUIOVector *qiov, int flags)
106
static void bdrv_co_em_bh(void *opaque)
102
+ QEMUIOVector *qiov,
107
{
103
+ size_t qiov_offset, int flags)
108
BlockAIOCBCoroutine *acb = opaque;
104
{
109
- BlockDriverState *bs = acb->common.bs;
105
BlockDriver *drv = bs->drv;
110
- AioContext *ctx = bdrv_get_aio_context(bs);
106
int64_t sector_num;
111
107
unsigned int nb_sectors;
112
assert(!acb->need_bh);
108
+ QEMUIOVector local_qiov;
113
- aio_context_acquire(ctx);
109
+ int ret;
114
bdrv_co_complete(acb);
110
115
- aio_context_release(ctx);
111
assert(!(flags & ~BDRV_REQ_MASK));
116
}
112
assert(!(flags & BDRV_REQ_NO_FALLBACK));
117
113
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
118
static void bdrv_co_maybe_schedule_bh(BlockAIOCBCoroutine *acb)
114
return -ENOMEDIUM;
119
diff --git a/block/iscsi.c b/block/iscsi.c
115
}
120
index XXXXXXX..XXXXXXX 100644
116
121
--- a/block/iscsi.c
117
+ if (drv->bdrv_co_preadv_part) {
122
+++ b/block/iscsi.c
118
+ return drv->bdrv_co_preadv_part(bs, offset, bytes, qiov, qiov_offset,
123
@@ -XXX,XX +XXX,XX @@ static void
119
+ flags);
124
iscsi_bh_cb(void *p)
120
+ }
125
{
121
+
126
IscsiAIOCB *acb = p;
122
+ if (qiov_offset > 0 || bytes != qiov->size) {
127
- AioContext *ctx = bdrv_get_aio_context(acb->common.bs);
123
+ qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
128
124
+ qiov = &local_qiov;
129
qemu_bh_delete(acb->bh);
125
+ }
130
126
+
131
g_free(acb->buf);
127
if (drv->bdrv_co_preadv) {
132
acb->buf = NULL;
128
- return drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags);
133
129
+ ret = drv->bdrv_co_preadv(bs, offset, bytes, qiov, flags);
134
- aio_context_acquire(ctx);
130
+ goto out;
135
acb->common.cb(acb->common.opaque, acb->status);
131
}
136
- aio_context_release(ctx);
132
137
133
if (drv->bdrv_aio_preadv) {
138
if (acb->task != NULL) {
134
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
139
scsi_free_scsi_task(acb->task);
135
acb = drv->bdrv_aio_preadv(bs, offset, bytes, qiov, flags,
140
diff --git a/block/linux-aio.c b/block/linux-aio.c
136
bdrv_co_io_em_complete, &co);
141
index XXXXXXX..XXXXXXX 100644
137
if (acb == NULL) {
142
--- a/block/linux-aio.c
138
- return -EIO;
143
+++ b/block/linux-aio.c
139
+ ret = -EIO;
144
@@ -XXX,XX +XXX,XX @@ static inline ssize_t io_event_ret(struct io_event *ev)
140
+ goto out;
145
*/
141
} else {
146
static void qemu_laio_process_completion(struct qemu_laiocb *laiocb)
142
qemu_coroutine_yield();
147
{
143
- return co.ret;
148
- LinuxAioState *s = laiocb->ctx;
144
+ ret = co.ret;
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
393
diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c
394
index XXXXXXX..XXXXXXX 100644
395
--- a/hw/block/virtio-blk.c
396
+++ b/hw/block/virtio-blk.c
397
@@ -XXX,XX +XXX,XX @@ static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error,
398
static void virtio_blk_rw_complete(void *opaque, int ret)
399
{
400
VirtIOBlockReq *next = opaque;
401
+ VirtIOBlock *s = next->dev;
402
403
+ aio_context_acquire(blk_get_aio_context(s->conf.conf.blk));
404
while (next) {
405
VirtIOBlockReq *req = next;
406
next = req->mr_next;
407
@@ -XXX,XX +XXX,XX @@ static void virtio_blk_rw_complete(void *opaque, int ret)
408
block_acct_done(blk_get_stats(req->dev->blk), &req->acct);
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;
145
+ goto out;
423
+ goto out;
146
}
424
}
147
}
425
}
148
426
149
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_driver_preadv(BlockDriverState *bs,
427
virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
150
assert(bytes <= BDRV_REQUEST_MAX_BYTES);
428
block_acct_done(blk_get_stats(req->dev->blk), &req->acct);
151
assert(drv->bdrv_co_readv);
429
virtio_blk_free_request(req);
152
153
- return drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
154
+ ret = drv->bdrv_co_readv(bs, sector_num, nb_sectors, qiov);
155
+
430
+
156
+out:
431
+out:
157
+ if (qiov == &local_qiov) {
432
+ aio_context_release(blk_get_aio_context(s->conf.conf.blk));
158
+ qemu_iovec_destroy(&local_qiov);
433
}
159
+ }
434
160
+
435
#ifdef __linux__
161
+ return ret;
436
@@ -XXX,XX +XXX,XX @@ static void virtio_blk_ioctl_complete(void *opaque, int status)
162
}
437
virtio_stl_p(vdev, &scsi->data_len, hdr->dxfer_len);
163
438
164
static int coroutine_fn bdrv_driver_pwritev(BlockDriverState *bs,
439
out:
165
uint64_t offset, uint64_t bytes,
440
+ aio_context_acquire(blk_get_aio_context(s->conf.conf.blk));
166
- QEMUIOVector *qiov, int flags)
441
virtio_blk_req_complete(req, status);
167
+ QEMUIOVector *qiov,
442
virtio_blk_free_request(req);
168
+ size_t qiov_offset, int flags)
443
+ aio_context_release(blk_get_aio_context(s->conf.conf.blk));
169
{
444
g_free(ioctl_req);
170
BlockDriver *drv = bs->drv;
445
}
171
int64_t sector_num;
446
172
unsigned int nb_sectors;
447
diff --git a/hw/scsi/scsi-disk.c b/hw/scsi/scsi-disk.c
173
+ QEMUIOVector local_qiov;
448
index XXXXXXX..XXXXXXX 100644
174
int ret;
449
--- a/hw/scsi/scsi-disk.c
175
450
+++ b/hw/scsi/scsi-disk.c
176
assert(!(flags & ~BDRV_REQ_MASK));
451
@@ -XXX,XX +XXX,XX @@ static void scsi_aio_complete(void *opaque, int ret)
177
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_driver_pwritev(BlockDriverState *bs,
452
178
return -ENOMEDIUM;
453
assert(r->req.aiocb != NULL);
179
}
454
r->req.aiocb = NULL;
180
455
+ aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk));
181
+ if (drv->bdrv_co_pwritev_part) {
456
if (scsi_disk_req_check_error(r, ret, true)) {
182
+ ret = drv->bdrv_co_pwritev_part(bs, offset, bytes, qiov, qiov_offset,
457
goto done;
183
+ flags & bs->supported_write_flags);
458
}
184
+ flags &= ~bs->supported_write_flags;
459
@@ -XXX,XX +XXX,XX @@ static void scsi_aio_complete(void *opaque, int ret)
185
+ goto emulate_flags;
460
scsi_req_complete(&r->req, GOOD);
186
+ }
461
187
+
462
done:
188
+ if (qiov_offset > 0 || bytes != qiov->size) {
463
+ aio_context_release(blk_get_aio_context(s->qdev.conf.blk));
189
+ qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
464
scsi_req_unref(&r->req);
190
+ qiov = &local_qiov;
465
}
191
+ }
466
192
+
467
@@ -XXX,XX +XXX,XX @@ static void scsi_dma_complete(void *opaque, int ret)
193
if (drv->bdrv_co_pwritev) {
468
assert(r->req.aiocb != NULL);
194
ret = drv->bdrv_co_pwritev(bs, offset, bytes, qiov,
469
r->req.aiocb = NULL;
195
flags & bs->supported_write_flags);
470
196
@@ -XXX,XX +XXX,XX @@ emulate_flags:
471
+ aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk));
197
ret = bdrv_co_flush(bs);
472
if (ret < 0) {
198
}
473
block_acct_failed(blk_get_stats(s->qdev.conf.blk), &r->acct);
199
474
} else {
200
+ if (qiov == &local_qiov) {
475
block_acct_done(blk_get_stats(s->qdev.conf.blk), &r->acct);
201
+ qemu_iovec_destroy(&local_qiov);
476
}
202
+ }
477
scsi_dma_complete_noio(r, ret);
203
+
478
+ aio_context_release(blk_get_aio_context(s->qdev.conf.blk));
204
return ret;
479
}
205
}
480
206
481
static void scsi_read_complete(void * opaque, int ret)
207
static int coroutine_fn
482
@@ -XXX,XX +XXX,XX @@ static void scsi_read_complete(void * opaque, int ret)
208
bdrv_driver_pwritev_compressed(BlockDriverState *bs, uint64_t offset,
483
209
- uint64_t bytes, QEMUIOVector *qiov)
484
assert(r->req.aiocb != NULL);
210
+ uint64_t bytes, QEMUIOVector *qiov,
485
r->req.aiocb = NULL;
211
+ size_t qiov_offset)
486
+ aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk));
212
{
487
if (scsi_disk_req_check_error(r, ret, true)) {
213
BlockDriver *drv = bs->drv;
488
goto done;
214
+ QEMUIOVector local_qiov;
489
}
215
+ int ret;
490
@@ -XXX,XX +XXX,XX @@ static void scsi_read_complete(void * opaque, int ret)
216
491
217
if (!drv) {
492
done:
218
return -ENOMEDIUM;
493
scsi_req_unref(&r->req);
219
}
494
+ aio_context_release(blk_get_aio_context(s->qdev.conf.blk));
220
495
}
221
- if (!drv->bdrv_co_pwritev_compressed) {
496
222
+ if (!block_driver_can_compress(drv)) {
497
/* Actually issue a read to the block device. */
223
return -ENOTSUP;
498
@@ -XXX,XX +XXX,XX @@ static void scsi_do_read_cb(void *opaque, int ret)
224
}
499
assert (r->req.aiocb != NULL);
225
500
r->req.aiocb = NULL;
226
- return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov);
501
227
+ if (drv->bdrv_co_pwritev_compressed_part) {
502
+ aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk));
228
+ return drv->bdrv_co_pwritev_compressed_part(bs, offset, bytes,
503
if (ret < 0) {
229
+ qiov, qiov_offset);
504
block_acct_failed(blk_get_stats(s->qdev.conf.blk), &r->acct);
230
+ }
505
} else {
231
+
506
block_acct_done(blk_get_stats(s->qdev.conf.blk), &r->acct);
232
+ if (qiov_offset == 0) {
507
}
233
+ return drv->bdrv_co_pwritev_compressed(bs, offset, bytes, qiov);
508
scsi_do_read(opaque, ret);
234
+ }
509
+ aio_context_release(blk_get_aio_context(s->qdev.conf.blk));
235
+
510
}
236
+ qemu_iovec_init_slice(&local_qiov, qiov, qiov_offset, bytes);
511
237
+ ret = drv->bdrv_co_pwritev_compressed(bs, offset, bytes, &local_qiov);
512
/* Read more data from scsi device into buffer. */
238
+ qemu_iovec_destroy(&local_qiov);
513
@@ -XXX,XX +XXX,XX @@ static void scsi_write_complete(void * opaque, int ret)
239
+
514
assert (r->req.aiocb != NULL);
240
+ return ret;
515
r->req.aiocb = NULL;
241
}
516
242
517
+ aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk));
243
static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
518
if (ret < 0) {
244
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
519
block_acct_failed(blk_get_stats(s->qdev.conf.blk), &r->acct);
245
qemu_iovec_init_buf(&local_qiov, bounce_buffer, pnum);
520
} else {
246
521
block_acct_done(blk_get_stats(s->qdev.conf.blk), &r->acct);
247
ret = bdrv_driver_preadv(bs, cluster_offset, pnum,
522
}
248
- &local_qiov, 0);
523
scsi_write_complete_noio(r, ret);
249
+ &local_qiov, 0, 0);
524
+ aio_context_release(blk_get_aio_context(s->qdev.conf.blk));
250
if (ret < 0) {
525
}
251
goto err;
526
252
}
527
static void scsi_write_data(SCSIRequest *req)
253
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
528
@@ -XXX,XX +XXX,XX @@ static void scsi_unmap_complete(void *opaque, int ret)
254
* necessary to flush even in cache=writethrough mode.
529
{
255
*/
530
UnmapCBData *data = opaque;
256
ret = bdrv_driver_pwritev(bs, cluster_offset, pnum,
531
SCSIDiskReq *r = data->r;
257
- &local_qiov,
532
+ SCSIDiskState *s = DO_UPCAST(SCSIDiskState, qdev, r->req.dev);
258
+ &local_qiov, 0,
533
259
BDRV_REQ_WRITE_UNCHANGED);
534
assert(r->req.aiocb != NULL);
260
}
535
r->req.aiocb = NULL;
261
536
262
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_copy_on_readv(BdrvChild *child,
537
+ aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk));
263
qemu_iovec_init(&local_qiov, qiov->niov);
538
scsi_unmap_complete_noio(data, ret);
264
qemu_iovec_concat(&local_qiov, qiov, progress, pnum - skip_bytes);
539
+ aio_context_release(blk_get_aio_context(s->qdev.conf.blk));
265
ret = bdrv_driver_preadv(bs, offset + progress, local_qiov.size,
540
}
266
- &local_qiov, 0);
541
267
+ &local_qiov, 0, 0);
542
static void scsi_disk_emulate_unmap(SCSIDiskReq *r, uint8_t *inbuf)
268
qemu_iovec_destroy(&local_qiov);
543
@@ -XXX,XX +XXX,XX @@ static void scsi_write_same_complete(void *opaque, int ret)
269
if (ret < 0) {
544
270
goto err;
545
assert(r->req.aiocb != NULL);
271
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
546
r->req.aiocb = NULL;
272
547
+ aio_context_acquire(blk_get_aio_context(s->qdev.conf.blk));
273
max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
548
if (scsi_disk_req_check_error(r, ret, true)) {
274
if (bytes <= max_bytes && bytes <= max_transfer) {
549
goto done;
275
- ret = bdrv_driver_preadv(bs, offset, bytes, qiov, 0);
550
}
276
+ ret = bdrv_driver_preadv(bs, offset, bytes, qiov, 0, 0);
551
@@ -XXX,XX +XXX,XX @@ done:
277
goto out;
552
scsi_req_unref(&r->req);
278
}
553
qemu_vfree(data->iov.iov_base);
279
554
g_free(data);
280
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
555
+ aio_context_release(blk_get_aio_context(s->qdev.conf.blk));
281
qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
556
}
282
557
283
ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining,
558
static void scsi_disk_emulate_write_same(SCSIDiskReq *r, uint8_t *inbuf)
284
- num, &local_qiov, 0);
559
diff --git a/hw/scsi/scsi-generic.c b/hw/scsi/scsi-generic.c
285
+ num, &local_qiov, 0, 0);
560
index XXXXXXX..XXXXXXX 100644
286
max_bytes -= num;
561
--- a/hw/scsi/scsi-generic.c
287
qemu_iovec_destroy(&local_qiov);
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;
288
} else {
646
} else {
289
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_co_do_pwrite_zeroes(BlockDriverState *bs,
647
@@ -XXX,XX +XXX,XX @@ static void thread_pool_co_cb(void *opaque, int ret)
290
}
648
ThreadPoolCo *co = opaque;
291
qemu_iovec_init_buf(&qiov, buf, num);
649
292
650
co->ret = ret;
293
- ret = bdrv_driver_pwritev(bs, offset, num, &qiov, write_flags);
651
- qemu_coroutine_enter(co->co);
294
+ ret = bdrv_driver_pwritev(bs, offset, num, &qiov, 0, write_flags);
652
+ aio_co_wake(co->co);
295
653
}
296
/* Keep bounce buffer around if it is big enough for all
654
297
* all future requests.
655
int coroutine_fn thread_pool_submit_co(ThreadPool *pool, ThreadPoolFunc *func,
298
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
299
bdrv_debug_event(bs, BLKDBG_PWRITEV_ZERO);
300
ret = bdrv_co_do_pwrite_zeroes(bs, offset, bytes, flags);
301
} else if (flags & BDRV_REQ_WRITE_COMPRESSED) {
302
- ret = bdrv_driver_pwritev_compressed(bs, offset, bytes, qiov);
303
+ ret = bdrv_driver_pwritev_compressed(bs, offset, bytes, qiov, 0);
304
} else if (bytes <= max_transfer) {
305
bdrv_debug_event(bs, BLKDBG_PWRITEV);
306
- ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, flags);
307
+ ret = bdrv_driver_pwritev(bs, offset, bytes, qiov, 0, flags);
308
} else {
309
bdrv_debug_event(bs, BLKDBG_PWRITEV);
310
while (bytes_remaining) {
311
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
312
qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
313
314
ret = bdrv_driver_pwritev(bs, offset + bytes - bytes_remaining,
315
- num, &local_qiov, local_flags);
316
+ num, &local_qiov, 0, local_flags);
317
qemu_iovec_destroy(&local_qiov);
318
if (ret < 0) {
319
break;
320
diff --git a/qemu-img.c b/qemu-img.c
321
index XXXXXXX..XXXXXXX 100644
322
--- a/qemu-img.c
323
+++ b/qemu-img.c
324
@@ -XXX,XX +XXX,XX @@ static int img_convert(int argc, char **argv)
325
const char *preallocation =
326
qemu_opt_get(opts, BLOCK_OPT_PREALLOC);
327
328
- if (drv && !drv->bdrv_co_pwritev_compressed) {
329
+ if (drv && !block_driver_can_compress(drv)) {
330
error_report("Compression not supported for this file format");
331
ret = -1;
332
goto out;
333
@@ -XXX,XX +XXX,XX @@ static int img_convert(int argc, char **argv)
334
}
335
out_bs = blk_bs(s.target);
336
337
- if (s.compressed && !out_bs->drv->bdrv_co_pwritev_compressed) {
338
+ if (s.compressed && !block_driver_can_compress(out_bs->drv)) {
339
error_report("Compression not supported for this file format");
340
ret = -1;
341
goto out;
342
--
656
--
343
2.21.0
657
2.9.3
344
658
345
659
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
We'll need to check a part of qiov soon, so implement it now.
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.
4
7
5
Optimization with align down to 4 * sizeof(long) is dropped due to:
8
aio_dispatch can now become void.
6
1. It is strange: it aligns length of the buffer, but where is a
7
guarantee that buffer pointer is aligned itself?
8
2. buffer_is_zero() is a better place for optimizations and it has
9
them.
10
9
11
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
10
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
12
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
11
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
13
Message-id: 20190604161514.262241-3-vsementsov@virtuozzo.com
12
Reviewed-by: Fam Zheng <famz@redhat.com>
14
Message-Id: <20190604161514.262241-3-vsementsov@virtuozzo.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>
15
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
16
---
16
---
17
include/qemu/iov.h | 2 +-
17
include/block/aio.h | 6 +-----
18
block/io.c | 2 +-
18
util/aio-posix.c | 44 ++++++++++++++------------------------------
19
util/iov.c | 31 +++++++++++++++++++------------
19
util/aio-win32.c | 13 ++++---------
20
3 files changed, 21 insertions(+), 14 deletions(-)
20
util/async.c | 2 +-
21
4 files changed, 20 insertions(+), 45 deletions(-)
21
22
22
diff --git a/include/qemu/iov.h b/include/qemu/iov.h
23
diff --git a/include/block/aio.h b/include/block/aio.h
23
index XXXXXXX..XXXXXXX 100644
24
index XXXXXXX..XXXXXXX 100644
24
--- a/include/qemu/iov.h
25
--- a/include/block/aio.h
25
+++ b/include/qemu/iov.h
26
+++ b/include/block/aio.h
26
@@ -XXX,XX +XXX,XX @@ void qemu_iovec_concat(QEMUIOVector *dst,
27
@@ -XXX,XX +XXX,XX @@ bool aio_pending(AioContext *ctx);
27
size_t qemu_iovec_concat_iov(QEMUIOVector *dst,
28
/* Dispatch any pending callbacks from the GSource attached to the AioContext.
28
struct iovec *src_iov, unsigned int src_cnt,
29
*
29
size_t soffset, size_t sbytes);
30
* This is used internally in the implementation of the GSource.
30
-bool qemu_iovec_is_zero(QEMUIOVector *qiov);
31
- *
31
+bool qemu_iovec_is_zero(QEMUIOVector *qiov, size_t qiov_offeset, size_t bytes);
32
- * @dispatch_fds: true to process fds, false to skip them
32
void qemu_iovec_destroy(QEMUIOVector *qiov);
33
- * (can be used as an optimization by callers that know there
33
void qemu_iovec_reset(QEMUIOVector *qiov);
34
- * are no fds ready)
34
size_t qemu_iovec_to_buf(QEMUIOVector *qiov, size_t offset,
35
*/
35
diff --git a/block/io.c b/block/io.c
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
36
index XXXXXXX..XXXXXXX 100644
42
index XXXXXXX..XXXXXXX 100644
37
--- a/block/io.c
43
--- a/util/aio-posix.c
38
+++ b/block/io.c
44
+++ b/util/aio-posix.c
39
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_pwritev(BdrvChild *child,
45
@@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx)
40
46
AioHandler *node, *tmp;
41
if (!ret && bs->detect_zeroes != BLOCKDEV_DETECT_ZEROES_OPTIONS_OFF &&
47
bool progress = false;
42
!(flags & BDRV_REQ_ZERO_WRITE) && drv->bdrv_co_pwrite_zeroes &&
48
43
- qemu_iovec_is_zero(qiov)) {
49
- /*
44
+ qemu_iovec_is_zero(qiov, 0, qiov->size)) {
50
- * We have to walk very carefully in case aio_set_fd_handler is
45
flags |= BDRV_REQ_ZERO_WRITE;
51
- * called while we're walking.
46
if (bs->detect_zeroes == BLOCKDEV_DETECT_ZEROES_OPTIONS_UNMAP) {
52
- */
47
flags |= BDRV_REQ_MAY_UNMAP;
53
- qemu_lockcnt_inc(&ctx->list_lock);
48
diff --git a/util/iov.c b/util/iov.c
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
49
index XXXXXXX..XXXXXXX 100644
119
index XXXXXXX..XXXXXXX 100644
50
--- a/util/iov.c
120
--- a/util/aio-win32.c
51
+++ b/util/iov.c
121
+++ b/util/aio-win32.c
52
@@ -XXX,XX +XXX,XX @@ void qemu_iovec_init_extended(
122
@@ -XXX,XX +XXX,XX @@ static bool aio_dispatch_handlers(AioContext *ctx, HANDLE event)
123
return progress;
53
}
124
}
54
125
55
/*
126
-bool aio_dispatch(AioContext *ctx, bool dispatch_fds)
56
- * Check if the contents of the iovecs are all zero
127
+void aio_dispatch(AioContext *ctx)
57
+ * Check if the contents of subrange of qiov data is all zeroes.
58
*/
59
-bool qemu_iovec_is_zero(QEMUIOVector *qiov)
60
+bool qemu_iovec_is_zero(QEMUIOVector *qiov, size_t offset, size_t bytes)
61
{
128
{
62
- int i;
129
- bool progress;
63
- for (i = 0; i < qiov->niov; i++) {
130
-
64
- size_t offs = QEMU_ALIGN_DOWN(qiov->iov[i].iov_len, 4 * sizeof(long));
131
- progress = aio_bh_poll(ctx);
65
- uint8_t *ptr = qiov->iov[i].iov_base;
132
- if (dispatch_fds) {
66
- if (offs && !buffer_is_zero(qiov->iov[i].iov_base, offs)) {
133
- progress |= aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE);
67
+ struct iovec *iov;
134
- }
68
+ size_t current_offset;
135
- progress |= timerlistgroup_run_timers(&ctx->tlg);
69
+
136
- return progress;
70
+ assert(offset + bytes <= qiov->size);
137
+ aio_bh_poll(ctx);
71
+
138
+ aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE);
72
+ iov = iov_skip_offset(qiov->iov, offset, &current_offset);
139
+ timerlistgroup_run_timers(&ctx->tlg);
73
+
140
}
74
+ while (bytes) {
141
75
+ uint8_t *base = (uint8_t *)iov->iov_base + current_offset;
142
bool aio_poll(AioContext *ctx, bool blocking)
76
+ size_t len = MIN(iov->iov_len - current_offset, bytes);
143
diff --git a/util/async.c b/util/async.c
77
+
144
index XXXXXXX..XXXXXXX 100644
78
+ if (!buffer_is_zero(base, len)) {
145
--- a/util/async.c
79
return false;
146
+++ b/util/async.c
80
}
147
@@ -XXX,XX +XXX,XX @@ aio_ctx_dispatch(GSource *source,
81
- for (; offs < qiov->iov[i].iov_len; offs++) {
148
AioContext *ctx = (AioContext *) source;
82
- if (ptr[offs]) {
149
83
- return false;
150
assert(callback == NULL);
84
- }
151
- aio_dispatch(ctx, true);
85
- }
152
+ aio_dispatch(ctx);
86
+
87
+ current_offset = 0;
88
+ bytes -= len;
89
+ iov++;
90
}
91
+
92
return true;
153
return true;
93
}
154
}
94
155
95
--
156
--
96
2.21.0
157
2.9.3
97
158
98
159
diff view generated by jsdifflib
New patch
1
From: Paolo Bonzini <pbonzini@redhat.com>
1
2
3
Pull the increment/decrement pair out of aio_bh_poll and into the
4
callers.
5
6
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
7
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
8
Reviewed-by: Fam Zheng <famz@redhat.com>
9
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
10
Message-id: 20170213135235.12274-18-pbonzini@redhat.com
11
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
12
---
13
util/aio-posix.c | 8 +++-----
14
util/aio-win32.c | 8 ++++----
15
util/async.c | 12 ++++++------
16
3 files changed, 13 insertions(+), 15 deletions(-)
17
18
diff --git a/util/aio-posix.c b/util/aio-posix.c
19
index XXXXXXX..XXXXXXX 100644
20
--- a/util/aio-posix.c
21
+++ b/util/aio-posix.c
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;
71
}
72
73
void aio_dispatch(AioContext *ctx)
74
{
75
+ qemu_lockcnt_inc(&ctx->list_lock);
76
aio_bh_poll(ctx);
77
aio_dispatch_handlers(ctx, INVALID_HANDLE_VALUE);
78
+ qemu_lockcnt_dec(&ctx->list_lock);
79
timerlistgroup_run_timers(&ctx->tlg);
80
}
81
82
@@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking)
83
}
84
}
85
86
- qemu_lockcnt_dec(&ctx->list_lock);
87
first = true;
88
89
/* ctx->notifier is always registered. */
90
@@ -XXX,XX +XXX,XX @@ bool aio_poll(AioContext *ctx, bool blocking)
91
progress |= aio_dispatch_handlers(ctx, event);
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;
138
}
139
}
140
- qemu_lockcnt_unlock(&ctx->list_lock);
141
+ qemu_lockcnt_inc_and_unlock(&ctx->list_lock);
142
}
143
return ret;
144
}
145
--
146
2.9.3
147
148
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Introduce extended variants of bdrv_co_preadv and bdrv_co_pwritev
3
Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
4
with qiov_offset parameter.
4
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
5
5
Reviewed-by: Fam Zheng <famz@redhat.com>
6
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
6
Reviewed-by: Daniel P. Berrange <berrange@redhat.com>
7
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
7
Message-id: 20170213135235.12274-19-pbonzini@redhat.com
8
Message-id: 20190604161514.262241-10-vsementsov@virtuozzo.com
9
Message-Id: <20190604161514.262241-10-vsementsov@virtuozzo.com>
10
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
8
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
11
---
9
---
12
include/block/block_int.h | 6 ++++++
10
include/block/block_int.h | 64 +++++++++++++++++++++++++-----------------
13
block/io.c | 29 +++++++++++++++++++++++------
11
include/sysemu/block-backend.h | 14 ++++++---
14
2 files changed, 29 insertions(+), 6 deletions(-)
12
2 files changed, 49 insertions(+), 29 deletions(-)
15
13
16
diff --git a/include/block/block_int.h b/include/block/block_int.h
14
diff --git a/include/block/block_int.h b/include/block/block_int.h
17
index XXXXXXX..XXXXXXX 100644
15
index XXXXXXX..XXXXXXX 100644
18
--- a/include/block/block_int.h
16
--- a/include/block/block_int.h
19
+++ b/include/block/block_int.h
17
+++ b/include/block/block_int.h
20
@@ -XXX,XX +XXX,XX @@ extern BlockDriver bdrv_qcow2;
18
@@ -XXX,XX +XXX,XX @@ struct BdrvChild {
21
int coroutine_fn bdrv_co_preadv(BdrvChild *child,
19
* copied as well.
22
int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
20
*/
23
BdrvRequestFlags flags);
21
struct BlockDriverState {
24
+int coroutine_fn bdrv_co_preadv_part(BdrvChild *child,
22
- int64_t total_sectors; /* if we are reading a disk image, give its
25
+ int64_t offset, unsigned int bytes,
23
- size in sectors */
26
+ QEMUIOVector *qiov, size_t qiov_offset, BdrvRequestFlags flags);
24
+ /* Protected by big QEMU lock or read-only after opening. No special
27
int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
25
+ * locking needed during I/O...
28
int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
26
+ */
29
BdrvRequestFlags flags);
27
int open_flags; /* flags used to open the file, re-used for re-open */
30
+int coroutine_fn bdrv_co_pwritev_part(BdrvChild *child,
28
bool read_only; /* if true, the media is read only */
31
+ int64_t offset, unsigned int bytes,
29
bool encrypted; /* if true, the media is encrypted */
32
+ QEMUIOVector *qiov, size_t qiov_offset, BdrvRequestFlags flags);
30
@@ -XXX,XX +XXX,XX @@ struct BlockDriverState {
33
31
bool sg; /* if true, the device is a /dev/sg* */
34
static inline int coroutine_fn bdrv_co_pread(BdrvChild *child,
32
bool probed; /* if true, format was probed rather than specified */
35
int64_t offset, unsigned int bytes, void *buf, BdrvRequestFlags flags)
33
36
diff --git a/block/io.c b/block/io.c
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 */
81
+
82
+ /* If true, copy read backing sectors into image. Can be >1 if more
83
+ * than one client has requested copy-on-read.
84
+ */
85
+ int copy_on_read;
86
+
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;
92
+
93
+ /* Callback before write request is processed */
94
+ NotifierWithReturnList before_write_notifiers;
95
+
96
+ /* number of in-flight requests; overall and serialising */
97
+ unsigned int in_flight;
98
+ unsigned int serialising_in_flight;
99
+
100
+ bool wakeup;
101
+
102
+ /* Offset after the highest byte written to */
103
+ uint64_t wr_highest_offset;
104
+
105
/* threshold limit for writes, in bytes. "High water mark". */
106
uint64_t write_threshold_offset;
107
NotifierWithReturn write_threshold_notifier;
108
@@ -XXX,XX +XXX,XX @@ struct BlockDriverState {
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 */
117
+
118
+ QLIST_HEAD(, BdrvDirtyBitmap) dirty_bitmaps;
119
+
120
+ /* do we need to tell the quest if we have a volatile write cache? */
121
+ int enable_write_cache;
122
+
123
int quiesce_counter;
124
};
125
126
diff --git a/include/sysemu/block-backend.h b/include/sysemu/block-backend.h
37
index XXXXXXX..XXXXXXX 100644
127
index XXXXXXX..XXXXXXX 100644
38
--- a/block/io.c
128
--- a/include/sysemu/block-backend.h
39
+++ b/block/io.c
129
+++ b/include/sysemu/block-backend.h
40
@@ -XXX,XX +XXX,XX @@ static void bdrv_padding_destroy(BdrvRequestPadding *pad)
130
@@ -XXX,XX +XXX,XX @@ typedef struct BlockDevOps {
41
*
131
* fields that must be public. This is in particular for QLIST_ENTRY() and
42
* Function always succeeds.
132
* friends so that BlockBackends can be kept in lists outside block-backend.c */
43
*/
133
typedef struct BlockBackendPublic {
44
-static bool bdrv_pad_request(BlockDriverState *bs, QEMUIOVector **qiov,
134
- /* I/O throttling.
45
+static bool bdrv_pad_request(BlockDriverState *bs,
135
- * throttle_state tells us if this BlockBackend has I/O limits configured.
46
+ QEMUIOVector **qiov, size_t *qiov_offset,
136
- * io_limits_disabled tells us if they are currently being enforced */
47
int64_t *offset, unsigned int *bytes,
137
+ /* I/O throttling has its own locking, but also some fields are
48
BdrvRequestPadding *pad)
138
+ * protected by the AioContext lock.
49
{
139
+ */
50
@@ -XXX,XX +XXX,XX @@ static bool bdrv_pad_request(BlockDriverState *bs, QEMUIOVector **qiov,
51
}
52
53
qemu_iovec_init_extended(&pad->local_qiov, pad->buf, pad->head,
54
- *qiov, 0, *bytes,
55
+ *qiov, *qiov_offset, *bytes,
56
pad->buf + pad->buf_len - pad->tail, pad->tail);
57
*bytes += pad->head + pad->tail;
58
*offset -= pad->head;
59
*qiov = &pad->local_qiov;
60
+ *qiov_offset = 0;
61
62
return true;
63
}
64
@@ -XXX,XX +XXX,XX @@ static bool bdrv_pad_request(BlockDriverState *bs, QEMUIOVector **qiov,
65
int coroutine_fn bdrv_co_preadv(BdrvChild *child,
66
int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
67
BdrvRequestFlags flags)
68
+{
69
+ return bdrv_co_preadv_part(child, offset, bytes, qiov, 0, flags);
70
+}
71
+
140
+
72
+int coroutine_fn bdrv_co_preadv_part(BdrvChild *child,
141
+ /* Protected by AioContext lock. */
73
+ int64_t offset, unsigned int bytes,
142
CoQueue throttled_reqs[2];
74
+ QEMUIOVector *qiov, size_t qiov_offset,
75
+ BdrvRequestFlags flags)
76
{
77
BlockDriverState *bs = child->bs;
78
BdrvTrackedRequest req;
79
@@ -XXX,XX +XXX,XX @@ int coroutine_fn bdrv_co_preadv(BdrvChild *child,
80
flags |= BDRV_REQ_COPY_ON_READ;
81
}
82
83
- bdrv_pad_request(bs, &qiov, &offset, &bytes, &pad);
84
+ bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad);
85
86
tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
87
ret = bdrv_aligned_preadv(child, &req, offset, bytes,
88
bs->bl.request_alignment,
89
- qiov, 0, flags);
90
+ qiov, qiov_offset, flags);
91
tracked_request_end(&req);
92
bdrv_dec_in_flight(bs);
93
94
@@ -XXX,XX +XXX,XX @@ out:
95
int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
96
int64_t offset, unsigned int bytes, QEMUIOVector *qiov,
97
BdrvRequestFlags flags)
98
+{
99
+ return bdrv_co_pwritev_part(child, offset, bytes, qiov, 0, flags);
100
+}
101
+
143
+
102
+int coroutine_fn bdrv_co_pwritev_part(BdrvChild *child,
144
+ /* Nonzero if the I/O limits are currently being ignored; generally
103
+ int64_t offset, unsigned int bytes, QEMUIOVector *qiov, size_t qiov_offset,
145
+ * it is zero. */
104
+ BdrvRequestFlags flags)
146
unsigned int io_limits_disabled;
105
{
147
106
BlockDriverState *bs = child->bs;
148
/* The following fields are protected by the ThrottleGroup lock.
107
BdrvTrackedRequest req;
149
- * See the ThrottleGroup documentation for details. */
108
@@ -XXX,XX +XXX,XX @@ int coroutine_fn bdrv_co_pwritev(BdrvChild *child,
150
+ * See the ThrottleGroup documentation for details.
109
goto out;
151
+ * throttle_state tells us if I/O limits are configured. */
110
}
152
ThrottleState *throttle_state;
111
153
ThrottleTimers throttle_timers;
112
- if (bdrv_pad_request(bs, &qiov, &offset, &bytes, &pad)) {
154
unsigned pending_reqs[2];
113
+ if (bdrv_pad_request(bs, &qiov, &qiov_offset, &offset, &bytes, &pad)) {
114
mark_request_serialising(&req, align);
115
wait_serialising_requests(&req);
116
bdrv_padding_rmw_read(child, &req, &pad, false);
117
}
118
119
ret = bdrv_aligned_pwritev(child, &req, offset, bytes, align,
120
- qiov, 0, flags);
121
+ qiov, qiov_offset, flags);
122
123
bdrv_padding_destroy(&pad);
124
125
--
155
--
126
2.21.0
156
2.9.3
127
157
128
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: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Use and support new API in bdrv_co_do_copy_on_readv.
3
Running a very small critical section on pthread_mutex_t and CoMutex
4
shows that pthread_mutex_t is much faster because it doesn't actually
5
go to sleep. What happens is that the critical section is shorter
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.
4
9
5
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
10
This only works with one waiters; because CoMutex is fair, it will
6
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
11
always have more waits and wakeups than a pthread_mutex_t.
7
Message-id: 20190604161514.262241-8-vsementsov@virtuozzo.com
12
8
Message-Id: <20190604161514.262241-8-vsementsov@virtuozzo.com>
13
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
14
Reviewed-by: Fam Zheng <famz@redhat.com>
15
Message-id: 20170213181244.16297-3-pbonzini@redhat.com
9
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
16
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
10
---
17
---
11
block/io.c | 21 ++++++++-------------
18
include/qemu/coroutine.h | 5 +++++
12
1 file changed, 8 insertions(+), 13 deletions(-)
19
util/qemu-coroutine-lock.c | 51 ++++++++++++++++++++++++++++++++++++++++------
20
util/qemu-coroutine.c | 2 +-
21
3 files changed, 51 insertions(+), 7 deletions(-)
13
22
14
diff --git a/block/io.c b/block/io.c
23
diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h
15
index XXXXXXX..XXXXXXX 100644
24
index XXXXXXX..XXXXXXX 100644
16
--- a/block/io.c
25
--- a/include/qemu/coroutine.h
17
+++ b/block/io.c
26
+++ b/include/qemu/coroutine.h
18
@@ -XXX,XX +XXX,XX @@ err:
27
@@ -XXX,XX +XXX,XX @@ typedef struct CoMutex {
19
*/
28
*/
20
static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
29
unsigned locked;
21
BdrvTrackedRequest *req, int64_t offset, unsigned int bytes,
30
22
- int64_t align, QEMUIOVector *qiov, int flags)
31
+ /* Context that is holding the lock. Useful to avoid spinning
23
+ int64_t align, QEMUIOVector *qiov, size_t qiov_offset, int flags)
32
+ * when two coroutines on the same AioContext try to get the lock. :)
33
+ */
34
+ AioContext *ctx;
35
+
36
/* A queue of waiters. Elements are added atomically in front of
37
* from_push. to_pop is only populated, and popped from, by whoever
38
* is in charge of the next wakeup. This can be an unlocker or,
39
diff --git a/util/qemu-coroutine-lock.c b/util/qemu-coroutine-lock.c
40
index XXXXXXX..XXXXXXX 100644
41
--- a/util/qemu-coroutine-lock.c
42
+++ b/util/qemu-coroutine-lock.c
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)
24
{
68
{
25
BlockDriverState *bs = child->bs;
69
Coroutine *self = qemu_coroutine_self();
26
int64_t total_bytes, max_bytes;
70
CoWaitRecord w;
27
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
71
@@ -XXX,XX +XXX,XX @@ static void coroutine_fn qemu_co_mutex_lock_slowpath(CoMutex *mutex)
28
assert(is_power_of_2(align));
72
if (co == self) {
29
assert((offset & (align - 1)) == 0);
73
/* We got the lock ourselves! */
30
assert((bytes & (align - 1)) == 0);
74
assert(to_wake == &w);
31
- assert(!qiov || bytes == qiov->size);
75
+ mutex->ctx = ctx;
32
assert((bs->open_flags & BDRV_O_NO_IO) == 0);
76
return;
33
max_transfer = QEMU_ALIGN_DOWN(MIN_NON_ZERO(bs->bl.max_transfer, INT_MAX),
34
align);
35
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
36
}
77
}
37
78
38
if (!ret || pnum != bytes) {
79
- aio_co_wake(co);
39
- ret = bdrv_co_do_copy_on_readv(child, offset, bytes, qiov, 0, flags);
80
+ qemu_co_mutex_wake(mutex, co);
40
+ ret = bdrv_co_do_copy_on_readv(child, offset, bytes,
41
+ qiov, qiov_offset, flags);
42
goto out;
43
} else if (flags & BDRV_REQ_PREFETCH) {
44
goto out;
45
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
46
47
max_bytes = ROUND_UP(MAX(0, total_bytes - offset), align);
48
if (bytes <= max_bytes && bytes <= max_transfer) {
49
- ret = bdrv_driver_preadv(bs, offset, bytes, qiov, 0, 0);
50
+ ret = bdrv_driver_preadv(bs, offset, bytes, qiov, qiov_offset, 0);
51
goto out;
52
}
81
}
53
82
54
@@ -XXX,XX +XXX,XX @@ static int coroutine_fn bdrv_aligned_preadv(BdrvChild *child,
83
qemu_coroutine_yield();
55
int num;
84
@@ -XXX,XX +XXX,XX @@ static void coroutine_fn qemu_co_mutex_lock_slowpath(CoMutex *mutex)
56
85
57
if (max_bytes) {
86
void coroutine_fn qemu_co_mutex_lock(CoMutex *mutex)
58
- QEMUIOVector local_qiov;
87
{
59
-
88
+ AioContext *ctx = qemu_get_current_aio_context();
60
num = MIN(bytes_remaining, MIN(max_bytes, max_transfer));
89
Coroutine *self = qemu_coroutine_self();
61
assert(num);
90
+ int waiters, i;
62
- qemu_iovec_init(&local_qiov, qiov->niov);
91
63
- qemu_iovec_concat(&local_qiov, qiov, bytes - bytes_remaining, num);
92
- if (atomic_fetch_inc(&mutex->locked) == 0) {
64
93
+ /* Running a very small critical section on pthread_mutex_t and CoMutex
65
ret = bdrv_driver_preadv(bs, offset + bytes - bytes_remaining,
94
+ * shows that pthread_mutex_t is much faster because it doesn't actually
66
- num, &local_qiov, 0, 0);
95
+ * go to sleep. What happens is that the critical section is shorter
67
+ num, qiov, bytes - bytes_remaining, 0);
96
+ * than the latency of entering the kernel and thus FUTEX_WAIT always
68
max_bytes -= num;
97
+ * fails. With CoMutex there is no such latency but you still want to
69
- qemu_iovec_destroy(&local_qiov);
98
+ * avoid wait and wakeup. So introduce it artificially.
70
} else {
99
+ */
71
num = bytes_remaining;
100
+ i = 0;
72
ret = qemu_iovec_memset(qiov, bytes - bytes_remaining, 0,
101
+retry_fast_path:
73
@@ -XXX,XX +XXX,XX @@ static int bdrv_padding_rmw_read(BdrvChild *child,
102
+ waiters = atomic_cmpxchg(&mutex->locked, 0, 1);
74
bdrv_debug_event(bs, BLKDBG_PWRITEV_RMW_TAIL);
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);
114
+ }
115
+
116
+ if (waiters == 0) {
117
/* Uncontended. */
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;
75
}
142
}
76
ret = bdrv_aligned_preadv(child, req, req->overlap_offset, bytes,
143
77
- align, &local_qiov, 0);
144
diff --git a/util/qemu-coroutine.c b/util/qemu-coroutine.c
78
+ align, &local_qiov, 0, 0);
145
index XXXXXXX..XXXXXXX 100644
79
if (ret < 0) {
146
--- a/util/qemu-coroutine.c
80
return ret;
147
+++ b/util/qemu-coroutine.c
81
}
148
@@ -XXX,XX +XXX,XX @@ void qemu_coroutine_enter(Coroutine *co)
82
@@ -XXX,XX +XXX,XX @@ static int bdrv_padding_rmw_read(BdrvChild *child,
149
co->ctx = qemu_get_current_aio_context();
83
ret = bdrv_aligned_preadv(
150
84
child, req,
151
/* Store co->ctx before anything that stores co. Matches
85
req->overlap_offset + req->overlap_bytes - align,
152
- * barrier in aio_co_wake.
86
- align, align, &local_qiov, 0);
153
+ * barrier in aio_co_wake and qemu_co_mutex_wake.
87
+ align, align, &local_qiov, 0, 0);
154
*/
88
if (ret < 0) {
155
smp_wmb();
89
return ret;
90
}
91
@@ -XXX,XX +XXX,XX @@ int coroutine_fn bdrv_co_preadv(BdrvChild *child,
92
tracked_request_begin(&req, bs, offset, bytes, BDRV_TRACKED_READ);
93
ret = bdrv_aligned_preadv(child, &req, offset, bytes,
94
bs->bl.request_alignment,
95
- qiov, flags);
96
+ qiov, 0, flags);
97
tracked_request_end(&req);
98
bdrv_dec_in_flight(bs);
99
156
100
--
157
--
101
2.21.0
158
2.9.3
102
159
103
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
New patch
1
From: Paolo Bonzini <pbonzini@redhat.com>
1
2
3
This will avoid forward references in the next patch. It is also
4
more logical because CoQueue is not anymore the basic primitive.
5
6
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
7
Reviewed-by: Fam Zheng <famz@redhat.com>
8
Message-id: 20170213181244.16297-5-pbonzini@redhat.com
9
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
10
---
11
include/qemu/coroutine.h | 89 ++++++++++++++++++++++++------------------------
12
1 file changed, 44 insertions(+), 45 deletions(-)
13
14
diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h
15
index XXXXXXX..XXXXXXX 100644
16
--- a/include/qemu/coroutine.h
17
+++ b/include/qemu/coroutine.h
18
@@ -XXX,XX +XXX,XX @@ bool qemu_in_coroutine(void);
19
*/
20
bool qemu_coroutine_entered(Coroutine *co);
21
22
-
23
-/**
24
- * CoQueues are a mechanism to queue coroutines in order to continue executing
25
- * them later. They provide the fundamental primitives on which coroutine locks
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;
121
--
122
2.9.3
123
124
diff view generated by jsdifflib
1
From: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
1
From: Paolo Bonzini <pbonzini@redhat.com>
2
2
3
Implement and use new interface to get rid of hd_qiov.
3
All that CoQueue needs in order to become thread-safe is help
4
4
from an external mutex. Add this to the API.
5
Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
5
6
Acked-by: Stefan Hajnoczi <stefanha@redhat.com>
6
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
7
Message-id: 20190604161514.262241-13-vsementsov@virtuozzo.com
7
Reviewed-by: Fam Zheng <famz@redhat.com>
8
Message-Id: <20190604161514.262241-13-vsementsov@virtuozzo.com>
8
Message-id: 20170213181244.16297-6-pbonzini@redhat.com
9
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
9
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
10
---
10
---
11
block/qcow2.h | 1 +
11
include/qemu/coroutine.h | 8 +++++---
12
include/qemu/iov.h | 1 +
12
block/backup.c | 2 +-
13
block/qcow2-cluster.c | 9 ++++---
13
block/io.c | 4 ++--
14
block/qcow2.c | 60 +++++++++++++++++++++----------------------
14
block/nbd-client.c | 2 +-
15
util/iov.c | 10 ++++++++
15
block/qcow2-cluster.c | 4 +---
16
5 files changed, 48 insertions(+), 33 deletions(-)
16
block/sheepdog.c | 2 +-
17
17
block/throttle-groups.c | 2 +-
18
diff --git a/block/qcow2.h b/block/qcow2.h
18
hw/9pfs/9p.c | 2 +-
19
index XXXXXXX..XXXXXXX 100644
19
util/qemu-coroutine-lock.c | 24 +++++++++++++++++++++---
20
--- a/block/qcow2.h
20
9 files changed, 34 insertions(+), 16 deletions(-)
21
+++ b/block/qcow2.h
21
22
@@ -XXX,XX +XXX,XX @@ typedef struct QCowL2Meta
22
diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h
23
* from @cow_start and @cow_end into one single write operation.
23
index XXXXXXX..XXXXXXX 100644
24
*/
24
--- a/include/qemu/coroutine.h
25
QEMUIOVector *data_qiov;
25
+++ b/include/qemu/coroutine.h
26
+ size_t data_qiov_offset;
26
@@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_mutex_unlock(CoMutex *mutex);
27
27
28
/** Pointer to next L2Meta of the same write request */
28
/**
29
struct QCowL2Meta *next;
29
* CoQueues are a mechanism to queue coroutines in order to continue executing
30
diff --git a/include/qemu/iov.h b/include/qemu/iov.h
30
- * them later.
31
index XXXXXXX..XXXXXXX 100644
31
+ * them later. They are similar to condition variables, but they need help
32
--- a/include/qemu/iov.h
32
+ * from an external mutex in order to maintain thread-safety.
33
+++ b/include/qemu/iov.h
33
*/
34
@@ -XXX,XX +XXX,XX @@ void qemu_iovec_init_extended(
34
typedef struct CoQueue {
35
void *tail_buf, size_t tail_len);
35
QSIMPLEQ_HEAD(, Coroutine) entries;
36
void qemu_iovec_init_slice(QEMUIOVector *qiov, QEMUIOVector *source,
36
@@ -XXX,XX +XXX,XX @@ void qemu_co_queue_init(CoQueue *queue);
37
size_t offset, size_t len);
37
38
+int qemu_iovec_subvec_niov(QEMUIOVector *qiov, size_t offset, size_t len);
38
/**
39
void qemu_iovec_add(QEMUIOVector *qiov, void *base, size_t len);
39
* Adds the current coroutine to the CoQueue and transfers control to the
40
void qemu_iovec_concat(QEMUIOVector *dst,
40
- * caller of the coroutine.
41
QEMUIOVector *src, size_t soffset, size_t sbytes);
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++;
42
diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
97
diff --git a/block/qcow2-cluster.c b/block/qcow2-cluster.c
43
index XXXXXXX..XXXXXXX 100644
98
index XXXXXXX..XXXXXXX 100644
44
--- a/block/qcow2-cluster.c
99
--- a/block/qcow2-cluster.c
45
+++ b/block/qcow2-cluster.c
100
+++ b/block/qcow2-cluster.c
46
@@ -XXX,XX +XXX,XX @@ static int perform_cow(BlockDriverState *bs, QCowL2Meta *m)
101
@@ -XXX,XX +XXX,XX @@ static int handle_dependencies(BlockDriverState *bs, uint64_t guest_offset,
47
assert(start->nb_bytes <= UINT_MAX - end->nb_bytes);
102
if (bytes == 0) {
48
assert(start->nb_bytes + end->nb_bytes <= UINT_MAX - data_bytes);
103
/* Wait for the dependency to complete. We need to recheck
49
assert(start->offset + start->nb_bytes <= end->offset);
104
* the free/allocated clusters when we continue. */
50
- assert(!m->data_qiov || m->data_qiov->size == data_bytes);
105
- qemu_co_mutex_unlock(&s->lock);
51
106
- qemu_co_queue_wait(&old_alloc->dependent_requests);
52
if ((start->nb_bytes == 0 && end->nb_bytes == 0) || m->skip_cow) {
107
- qemu_co_mutex_lock(&s->lock);
53
return 0;
108
+ qemu_co_queue_wait(&old_alloc->dependent_requests, &s->lock);
54
@@ -XXX,XX +XXX,XX @@ static int perform_cow(BlockDriverState *bs, QCowL2Meta *m)
109
return -EAGAIN;
55
/* The part of the buffer where the end region is located */
110
}
56
end_buffer = start_buffer + buffer_size - end->nb_bytes;
57
58
- qemu_iovec_init(&qiov, 2 + (m->data_qiov ? m->data_qiov->niov : 0));
59
+ qemu_iovec_init(&qiov, 2 + (m->data_qiov ?
60
+ qemu_iovec_subvec_niov(m->data_qiov,
61
+ m->data_qiov_offset,
62
+ data_bytes)
63
+ : 0));
64
65
qemu_co_mutex_unlock(&s->lock);
66
/* First we read the existing data from both COW regions. We
67
@@ -XXX,XX +XXX,XX @@ static int perform_cow(BlockDriverState *bs, QCowL2Meta *m)
68
if (start->nb_bytes) {
69
qemu_iovec_add(&qiov, start_buffer, start->nb_bytes);
70
}
111
}
71
- qemu_iovec_concat(&qiov, m->data_qiov, 0, data_bytes);
112
diff --git a/block/sheepdog.c b/block/sheepdog.c
72
+ qemu_iovec_concat(&qiov, m->data_qiov, m->data_qiov_offset, data_bytes);
113
index XXXXXXX..XXXXXXX 100644
73
if (end->nb_bytes) {
114
--- a/block/sheepdog.c
74
qemu_iovec_add(&qiov, end_buffer, end->nb_bytes);
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;
75
}
123
}
76
diff --git a/block/qcow2.c b/block/qcow2.c
124
}
77
index XXXXXXX..XXXXXXX 100644
125
diff --git a/block/throttle-groups.c b/block/throttle-groups.c
78
--- a/block/qcow2.c
126
index XXXXXXX..XXXXXXX 100644
79
+++ b/block/qcow2.c
127
--- a/block/throttle-groups.c
80
@@ -XXX,XX +XXX,XX @@ fail:
128
+++ b/block/throttle-groups.c
81
/* Check if it's possible to merge a write request with the writing of
129
@@ -XXX,XX +XXX,XX @@ void coroutine_fn throttle_group_co_io_limits_intercept(BlockBackend *blk,
82
* the data from the COW regions */
130
if (must_wait || blkp->pending_reqs[is_write]) {
83
static bool merge_cow(uint64_t offset, unsigned bytes,
131
blkp->pending_reqs[is_write]++;
84
- QEMUIOVector *hd_qiov, QCowL2Meta *l2meta)
132
qemu_mutex_unlock(&tg->lock);
85
+ QEMUIOVector *qiov, size_t qiov_offset,
133
- qemu_co_queue_wait(&blkp->throttled_reqs[is_write]);
86
+ QCowL2Meta *l2meta)
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)
87
{
161
{
88
QCowL2Meta *m;
162
Coroutine *self = qemu_coroutine_self();
89
163
QSIMPLEQ_INSERT_TAIL(&queue->entries, self, co_queue_next);
90
@@ -XXX,XX +XXX,XX @@ static bool merge_cow(uint64_t offset, unsigned bytes,
164
+
91
165
+ if (mutex) {
92
/* Make sure that adding both COW regions to the QEMUIOVector
166
+ qemu_co_mutex_unlock(mutex);
93
* does not exceed IOV_MAX */
167
+ }
94
- if (hd_qiov->niov > IOV_MAX - 2) {
168
+
95
+ if (qemu_iovec_subvec_niov(qiov, qiov_offset, bytes) > IOV_MAX - 2) {
169
+ /* There is no race condition here. Other threads will call
96
continue;
170
+ * aio_co_schedule on our AioContext, which can reenter this
97
}
171
+ * coroutine but only after this yield and after the main loop
98
172
+ * has gone through the next iteration.
99
- m->data_qiov = hd_qiov;
173
+ */
100
+ m->data_qiov = qiov;
174
qemu_coroutine_yield();
101
+ m->data_qiov_offset = qiov_offset;
175
assert(qemu_in_coroutine());
102
return true;
176
+
103
}
177
+ /* TODO: OSv implements wait morphing here, where the wakeup
104
178
+ * primitive automatically places the woken coroutine on the
105
@@ -XXX,XX +XXX,XX @@ static int handle_alloc_space(BlockDriverState *bs, QCowL2Meta *l2meta)
179
+ * mutex's queue. This avoids the thundering herd effect.
106
return 0;
180
+ */
181
+ if (mutex) {
182
+ qemu_co_mutex_lock(mutex);
183
+ }
107
}
184
}
108
185
109
-static coroutine_fn int qcow2_co_pwritev(BlockDriverState *bs, uint64_t offset,
186
/**
110
- uint64_t bytes, QEMUIOVector *qiov,
187
@@ -XXX,XX +XXX,XX @@ void qemu_co_rwlock_rdlock(CoRwlock *lock)
111
- int flags)
188
Coroutine *self = qemu_coroutine_self();
112
+static coroutine_fn int qcow2_co_pwritev_part(
189
113
+ BlockDriverState *bs, uint64_t offset, uint64_t bytes,
190
while (lock->writer) {
114
+ QEMUIOVector *qiov, size_t qiov_offset, int flags)
191
- qemu_co_queue_wait(&lock->queue);
115
{
192
+ qemu_co_queue_wait(&lock->queue, NULL);
116
BDRVQcow2State *s = bs->opaque;
193
}
117
int offset_in_cluster;
194
lock->reader++;
118
int ret;
195
self->locks_held++;
119
unsigned int cur_bytes; /* number of sectors in current iteration */
196
@@ -XXX,XX +XXX,XX @@ void qemu_co_rwlock_wrlock(CoRwlock *lock)
120
uint64_t cluster_offset;
197
Coroutine *self = qemu_coroutine_self();
121
- QEMUIOVector hd_qiov;
198
122
+ QEMUIOVector encrypted_qiov;
199
while (lock->writer || lock->reader) {
123
uint64_t bytes_done = 0;
200
- qemu_co_queue_wait(&lock->queue);
124
uint8_t *cluster_data = NULL;
201
+ qemu_co_queue_wait(&lock->queue, NULL);
125
QCowL2Meta *l2meta = NULL;
202
}
126
203
lock->writer = true;
127
trace_qcow2_writev_start_req(qemu_coroutine_self(), offset, bytes);
204
self->locks_held++;
128
129
- qemu_iovec_init(&hd_qiov, qiov->niov);
130
-
131
qemu_co_mutex_lock(&s->lock);
132
133
while (bytes != 0) {
134
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_pwritev(BlockDriverState *bs, uint64_t offset,
135
136
qemu_co_mutex_unlock(&s->lock);
137
138
- qemu_iovec_reset(&hd_qiov);
139
- qemu_iovec_concat(&hd_qiov, qiov, bytes_done, cur_bytes);
140
-
141
if (bs->encrypted) {
142
assert(s->crypto);
143
if (!cluster_data) {
144
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_pwritev(BlockDriverState *bs, uint64_t offset,
145
}
146
}
147
148
- assert(hd_qiov.size <=
149
- QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);
150
- qemu_iovec_to_buf(&hd_qiov, 0, cluster_data, hd_qiov.size);
151
+ assert(cur_bytes <= QCOW_MAX_CRYPT_CLUSTERS * s->cluster_size);
152
+ qemu_iovec_to_buf(qiov, qiov_offset + bytes_done,
153
+ cluster_data, cur_bytes);
154
155
if (qcow2_co_encrypt(bs, cluster_offset, offset,
156
cluster_data, cur_bytes) < 0) {
157
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_pwritev(BlockDriverState *bs, uint64_t offset,
158
goto out_unlocked;
159
}
160
161
- qemu_iovec_reset(&hd_qiov);
162
- qemu_iovec_add(&hd_qiov, cluster_data, cur_bytes);
163
+ qemu_iovec_init_buf(&encrypted_qiov, cluster_data, cur_bytes);
164
}
165
166
/* Try to efficiently initialize the physical space with zeroes */
167
@@ -XXX,XX +XXX,XX @@ static coroutine_fn int qcow2_co_pwritev(BlockDriverState *bs, uint64_t offset,
168
* writing of the guest data together with that of the COW regions.
169
* If it's not possible (or not necessary) then write the
170
* guest data now. */
171
- if (!merge_cow(offset, cur_bytes, &hd_qiov, l2meta)) {
172
+ if (!merge_cow(offset, cur_bytes,
173
+ bs->encrypted ? &encrypted_qiov : qiov,
174
+ bs->encrypted ? 0 : qiov_offset + bytes_done, l2meta))
175
+ {
176
BLKDBG_EVENT(bs->file, BLKDBG_WRITE_AIO);
177
trace_qcow2_writev_data(qemu_coroutine_self(),
178
cluster_offset + offset_in_cluster);
179
- ret = bdrv_co_pwritev(s->data_file,
180
- cluster_offset + offset_in_cluster,
181
- cur_bytes, &hd_qiov, 0);
182
+ ret = bdrv_co_pwritev_part(
183
+ s->data_file, cluster_offset + offset_in_cluster, cur_bytes,
184
+ bs->encrypted ? &encrypted_qiov : qiov,
185
+ bs->encrypted ? 0 : qiov_offset + bytes_done, 0);
186
if (ret < 0) {
187
goto out_unlocked;
188
}
189
@@ -XXX,XX +XXX,XX @@ out_locked:
190
191
qemu_co_mutex_unlock(&s->lock);
192
193
- qemu_iovec_destroy(&hd_qiov);
194
qemu_vfree(cluster_data);
195
trace_qcow2_writev_done_req(qemu_coroutine_self(), ret);
196
197
@@ -XXX,XX +XXX,XX @@ fail:
198
/* XXX: put compressed sectors first, then all the cluster aligned
199
tables to avoid losing bytes in alignment */
200
static coroutine_fn int
201
-qcow2_co_pwritev_compressed(BlockDriverState *bs, uint64_t offset,
202
- uint64_t bytes, QEMUIOVector *qiov)
203
+qcow2_co_pwritev_compressed_part(BlockDriverState *bs,
204
+ uint64_t offset, uint64_t bytes,
205
+ QEMUIOVector *qiov, size_t qiov_offset)
206
{
207
BDRVQcow2State *s = bs->opaque;
208
int ret;
209
@@ -XXX,XX +XXX,XX @@ qcow2_co_pwritev_compressed(BlockDriverState *bs, uint64_t offset,
210
/* Zero-pad last write if image size is not cluster aligned */
211
memset(buf + bytes, 0, s->cluster_size - bytes);
212
}
213
- qemu_iovec_to_buf(qiov, 0, buf, bytes);
214
+ qemu_iovec_to_buf(qiov, qiov_offset, buf, bytes);
215
216
out_buf = g_malloc(s->cluster_size);
217
218
@@ -XXX,XX +XXX,XX @@ qcow2_co_pwritev_compressed(BlockDriverState *bs, uint64_t offset,
219
buf, s->cluster_size);
220
if (out_len == -ENOMEM) {
221
/* could not compress: write normal cluster */
222
- ret = qcow2_co_pwritev(bs, offset, bytes, qiov, 0);
223
+ ret = qcow2_co_pwritev_part(bs, offset, bytes, qiov, qiov_offset, 0);
224
if (ret < 0) {
225
goto fail;
226
}
227
@@ -XXX,XX +XXX,XX @@ static int qcow2_save_vmstate(BlockDriverState *bs, QEMUIOVector *qiov,
228
BDRVQcow2State *s = bs->opaque;
229
230
BLKDBG_EVENT(bs->file, BLKDBG_VMSTATE_SAVE);
231
- return bs->drv->bdrv_co_pwritev(bs, qcow2_vm_state_offset(s) + pos,
232
- qiov->size, qiov, 0);
233
+ return bs->drv->bdrv_co_pwritev_part(bs, qcow2_vm_state_offset(s) + pos,
234
+ qiov->size, qiov, 0, 0);
235
}
236
237
static int qcow2_load_vmstate(BlockDriverState *bs, QEMUIOVector *qiov,
238
@@ -XXX,XX +XXX,XX @@ BlockDriver bdrv_qcow2 = {
239
.bdrv_co_block_status = qcow2_co_block_status,
240
241
.bdrv_co_preadv_part = qcow2_co_preadv_part,
242
- .bdrv_co_pwritev = qcow2_co_pwritev,
243
+ .bdrv_co_pwritev_part = qcow2_co_pwritev_part,
244
.bdrv_co_flush_to_os = qcow2_co_flush_to_os,
245
246
.bdrv_co_pwrite_zeroes = qcow2_co_pwrite_zeroes,
247
@@ -XXX,XX +XXX,XX @@ BlockDriver bdrv_qcow2 = {
248
.bdrv_co_copy_range_from = qcow2_co_copy_range_from,
249
.bdrv_co_copy_range_to = qcow2_co_copy_range_to,
250
.bdrv_co_truncate = qcow2_co_truncate,
251
- .bdrv_co_pwritev_compressed = qcow2_co_pwritev_compressed,
252
+ .bdrv_co_pwritev_compressed_part = qcow2_co_pwritev_compressed_part,
253
.bdrv_make_empty = qcow2_make_empty,
254
255
.bdrv_snapshot_create = qcow2_snapshot_create,
256
diff --git a/util/iov.c b/util/iov.c
257
index XXXXXXX..XXXXXXX 100644
258
--- a/util/iov.c
259
+++ b/util/iov.c
260
@@ -XXX,XX +XXX,XX @@ static struct iovec *qiov_slice(QEMUIOVector *qiov,
261
return iov;
262
}
263
264
+int qemu_iovec_subvec_niov(QEMUIOVector *qiov, size_t offset, size_t len)
265
+{
266
+ size_t head, tail;
267
+ int niov;
268
+
269
+ qiov_slice(qiov, offset, len, &head, &tail, &niov);
270
+
271
+ return niov;
272
+}
273
+
274
/*
275
* Compile new iovec, combining @head_buf buffer, sub-qiov of @mid_qiov,
276
* and @tail_buf buffer into new qiov.
277
--
205
--
278
2.21.0
206
2.9.3
279
207
280
208
diff view generated by jsdifflib
New patch
1
From: Paolo Bonzini <pbonzini@redhat.com>
1
2
3
This adds a CoMutex around the existing CoQueue. Because the write-side
4
can just take CoMutex, the old "writer" field is not necessary anymore.
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.
8
9
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
10
Reviewed-by: Fam Zheng <famz@redhat.com>
11
Message-id: 20170213181244.16297-7-pbonzini@redhat.com
12
Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
13
---
14
include/qemu/coroutine.h | 3 ++-
15
util/qemu-coroutine-lock.c | 35 ++++++++++++++++++++++++-----------
16
2 files changed, 26 insertions(+), 12 deletions(-)
17
18
diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h
19
index XXXXXXX..XXXXXXX 100644
20
--- a/include/qemu/coroutine.h
21
+++ b/include/qemu/coroutine.h
22
@@ -XXX,XX +XXX,XX @@ bool qemu_co_queue_empty(CoQueue *queue);
23
24
25
typedef struct CoRwlock {
26
- bool writer;
27
+ int pending_writer;
28
int reader;
29
+ CoMutex mutex;
30
CoQueue queue;
31
} CoRwlock;
32
33
diff --git a/util/qemu-coroutine-lock.c b/util/qemu-coroutine-lock.c
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
}
106
--
107
2.9.3
108
109
diff view generated by jsdifflib