1 | The following changes since commit 6c769690ac845fa62642a5f93b4e4bd906adab95: | 1 | The following changes since commit 801f3db7564dcce8a37a70833c0abe40ec19f8ce: |
---|---|---|---|
2 | 2 | ||
3 | Merge remote-tracking branch 'remotes/vsementsov/tags/pull-simplebench-2021-05-04' into staging (2021-05-21 12:02:34 +0100) | 3 | Merge remote-tracking branch 'remotes/philmd/tags/kconfig-20210720' into staging (2021-07-20 19:30:28 +0100) |
4 | 4 | ||
5 | are available in the Git repository at: | 5 | are available in the Git repository at: |
6 | 6 | ||
7 | https://gitlab.com/stefanha/qemu.git tags/block-pull-request | 7 | https://gitlab.com/stefanha/qemu.git tags/block-pull-request |
8 | 8 | ||
9 | for you to fetch changes up to 0a6f0c76a030710780ce10d6347a70f098024d21: | 9 | for you to fetch changes up to d7ddd0a1618a75b31dc308bb37365ce1da972154: |
10 | 10 | ||
11 | coroutine-sleep: introduce qemu_co_sleep (2021-05-21 18:22:33 +0100) | 11 | linux-aio: limit the batch size using `aio-max-batch` parameter (2021-07-21 13:47:50 +0100) |
12 | 12 | ||
13 | ---------------------------------------------------------------- | 13 | ---------------------------------------------------------------- |
14 | Pull request | 14 | Pull request |
15 | 15 | ||
16 | (Resent due to an email preparation mistake.) | 16 | Stefano's performance regression fix for commit 2558cb8dd4 ("linux-aio: |
17 | increasing MAX_EVENTS to a larger hardcoded value"). | ||
17 | 18 | ||
18 | ---------------------------------------------------------------- | 19 | ---------------------------------------------------------------- |
19 | 20 | ||
20 | Paolo Bonzini (6): | 21 | Stefano Garzarella (3): |
21 | coroutine-sleep: use a stack-allocated timer | 22 | iothread: generalize iothread_set_param/iothread_get_param |
22 | coroutine-sleep: disallow NULL QemuCoSleepState** argument | 23 | iothread: add aio-max-batch parameter |
23 | coroutine-sleep: allow qemu_co_sleep_wake that wakes nothing | 24 | linux-aio: limit the batch size using `aio-max-batch` parameter |
24 | coroutine-sleep: move timer out of QemuCoSleepState | ||
25 | coroutine-sleep: replace QemuCoSleepState pointer with struct in the | ||
26 | API | ||
27 | coroutine-sleep: introduce qemu_co_sleep | ||
28 | 25 | ||
29 | Philippe Mathieu-Daudé (1): | 26 | qapi/misc.json | 6 ++- |
30 | bitops.h: Improve find_xxx_bit() documentation | 27 | qapi/qom.json | 7 +++- |
31 | 28 | include/block/aio.h | 12 ++++++ | |
32 | Zenghui Yu (1): | 29 | include/sysemu/iothread.h | 3 ++ |
33 | multi-process: Initialize variables declared with g_auto* | 30 | block/linux-aio.c | 9 ++++- |
34 | 31 | iothread.c | 82 ++++++++++++++++++++++++++++++++++----- | |
35 | include/qemu/bitops.h | 15 ++++++-- | 32 | monitor/hmp-cmds.c | 2 + |
36 | include/qemu/coroutine.h | 27 ++++++++----- | 33 | util/aio-posix.c | 12 ++++++ |
37 | block/block-copy.c | 10 ++--- | 34 | util/aio-win32.c | 5 +++ |
38 | block/nbd.c | 14 +++---- | 35 | util/async.c | 2 + |
39 | hw/remote/memory.c | 5 +-- | 36 | qemu-options.hx | 8 +++- |
40 | hw/remote/proxy.c | 3 +- | 37 | 11 files changed, 134 insertions(+), 14 deletions(-) |
41 | util/qemu-coroutine-sleep.c | 75 +++++++++++++++++++------------------ | ||
42 | 7 files changed, 79 insertions(+), 70 deletions(-) | ||
43 | 38 | ||
44 | -- | 39 | -- |
45 | 2.31.1 | 40 | 2.31.1 |
46 | 41 | diff view generated by jsdifflib |
Deleted patch | |||
---|---|---|---|
1 | From: Zenghui Yu <yuzenghui@huawei.com> | ||
2 | 1 | ||
3 | Quote docs/devel/style.rst (section "Automatic memory deallocation"): | ||
4 | |||
5 | * Variables declared with g_auto* MUST always be initialized, | ||
6 | otherwise the cleanup function will use uninitialized stack memory | ||
7 | |||
8 | Initialize @name properly to get rid of the compilation error (using | ||
9 | gcc-7.3.0 on CentOS): | ||
10 | |||
11 | ../hw/remote/proxy.c: In function 'pci_proxy_dev_realize': | ||
12 | /usr/include/glib-2.0/glib/glib-autocleanups.h:28:3: error: 'name' may be used uninitialized in this function [-Werror=maybe-uninitialized] | ||
13 | g_free (*pp); | ||
14 | ^~~~~~~~~~~~ | ||
15 | ../hw/remote/proxy.c:350:30: note: 'name' was declared here | ||
16 | g_autofree char *name; | ||
17 | ^~~~ | ||
18 | |||
19 | Signed-off-by: Zenghui Yu <yuzenghui@huawei.com> | ||
20 | Reviewed-by: Jagannathan Raman <jag.raman@oracle.com> | ||
21 | Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com> | ||
22 | Reviewed-by: Miroslav Rezanina <mrezanin@redhat.com> | ||
23 | Message-id: 20210312112143.1369-1-yuzenghui@huawei.com | ||
24 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
25 | --- | ||
26 | hw/remote/memory.c | 5 ++--- | ||
27 | hw/remote/proxy.c | 3 +-- | ||
28 | 2 files changed, 3 insertions(+), 5 deletions(-) | ||
29 | |||
30 | diff --git a/hw/remote/memory.c b/hw/remote/memory.c | ||
31 | index XXXXXXX..XXXXXXX 100644 | ||
32 | --- a/hw/remote/memory.c | ||
33 | +++ b/hw/remote/memory.c | ||
34 | @@ -XXX,XX +XXX,XX @@ void remote_sysmem_reconfig(MPQemuMsg *msg, Error **errp) | ||
35 | |||
36 | remote_sysmem_reset(); | ||
37 | |||
38 | - for (region = 0; region < msg->num_fds; region++) { | ||
39 | - g_autofree char *name; | ||
40 | + for (region = 0; region < msg->num_fds; region++, suffix++) { | ||
41 | + g_autofree char *name = g_strdup_printf("remote-mem-%u", suffix); | ||
42 | subregion = g_new(MemoryRegion, 1); | ||
43 | - name = g_strdup_printf("remote-mem-%u", suffix++); | ||
44 | memory_region_init_ram_from_fd(subregion, NULL, | ||
45 | name, sysmem_info->sizes[region], | ||
46 | true, msg->fds[region], | ||
47 | diff --git a/hw/remote/proxy.c b/hw/remote/proxy.c | ||
48 | index XXXXXXX..XXXXXXX 100644 | ||
49 | --- a/hw/remote/proxy.c | ||
50 | +++ b/hw/remote/proxy.c | ||
51 | @@ -XXX,XX +XXX,XX @@ static void probe_pci_info(PCIDevice *dev, Error **errp) | ||
52 | PCI_BASE_ADDRESS_SPACE_IO : PCI_BASE_ADDRESS_SPACE_MEMORY; | ||
53 | |||
54 | if (size) { | ||
55 | - g_autofree char *name; | ||
56 | + g_autofree char *name = g_strdup_printf("bar-region-%d", i); | ||
57 | pdev->region[i].dev = pdev; | ||
58 | pdev->region[i].present = true; | ||
59 | if (type == PCI_BASE_ADDRESS_SPACE_MEMORY) { | ||
60 | pdev->region[i].memory = true; | ||
61 | } | ||
62 | - name = g_strdup_printf("bar-region-%d", i); | ||
63 | memory_region_init_io(&pdev->region[i].mr, OBJECT(pdev), | ||
64 | &proxy_mr_ops, &pdev->region[i], | ||
65 | name, size); | ||
66 | -- | ||
67 | 2.31.1 | ||
68 | diff view generated by jsdifflib |
Deleted patch | |||
---|---|---|---|
1 | From: Philippe Mathieu-Daudé <philmd@redhat.com> | ||
2 | 1 | ||
3 | Document the following functions return the bitmap size | ||
4 | if no matching bit is found: | ||
5 | |||
6 | - find_first_bit | ||
7 | - find_next_bit | ||
8 | - find_last_bit | ||
9 | - find_first_zero_bit | ||
10 | - find_next_zero_bit | ||
11 | |||
12 | Reviewed-by: Richard Henderson <richard.henderson@linaro.org> | ||
13 | Signed-off-by: Philippe Mathieu-Daudé <philmd@redhat.com> | ||
14 | Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
15 | Message-id: 20210510200758.2623154-2-philmd@redhat.com | ||
16 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
17 | --- | ||
18 | include/qemu/bitops.h | 15 ++++++++++++--- | ||
19 | 1 file changed, 12 insertions(+), 3 deletions(-) | ||
20 | |||
21 | diff --git a/include/qemu/bitops.h b/include/qemu/bitops.h | ||
22 | index XXXXXXX..XXXXXXX 100644 | ||
23 | --- a/include/qemu/bitops.h | ||
24 | +++ b/include/qemu/bitops.h | ||
25 | @@ -XXX,XX +XXX,XX @@ static inline int test_bit(long nr, const unsigned long *addr) | ||
26 | * @addr: The address to start the search at | ||
27 | * @size: The maximum size to search | ||
28 | * | ||
29 | - * Returns the bit number of the first set bit, or size. | ||
30 | + * Returns the bit number of the last set bit, | ||
31 | + * or @size if there is no set bit in the bitmap. | ||
32 | */ | ||
33 | unsigned long find_last_bit(const unsigned long *addr, | ||
34 | unsigned long size); | ||
35 | @@ -XXX,XX +XXX,XX @@ unsigned long find_last_bit(const unsigned long *addr, | ||
36 | * @addr: The address to base the search on | ||
37 | * @offset: The bitnumber to start searching at | ||
38 | * @size: The bitmap size in bits | ||
39 | + * | ||
40 | + * Returns the bit number of the next set bit, | ||
41 | + * or @size if there are no further set bits in the bitmap. | ||
42 | */ | ||
43 | unsigned long find_next_bit(const unsigned long *addr, | ||
44 | unsigned long size, | ||
45 | @@ -XXX,XX +XXX,XX @@ unsigned long find_next_bit(const unsigned long *addr, | ||
46 | * @addr: The address to base the search on | ||
47 | * @offset: The bitnumber to start searching at | ||
48 | * @size: The bitmap size in bits | ||
49 | + * | ||
50 | + * Returns the bit number of the next cleared bit, | ||
51 | + * or @size if there are no further clear bits in the bitmap. | ||
52 | */ | ||
53 | |||
54 | unsigned long find_next_zero_bit(const unsigned long *addr, | ||
55 | @@ -XXX,XX +XXX,XX @@ unsigned long find_next_zero_bit(const unsigned long *addr, | ||
56 | * @addr: The address to start the search at | ||
57 | * @size: The maximum size to search | ||
58 | * | ||
59 | - * Returns the bit number of the first set bit. | ||
60 | + * Returns the bit number of the first set bit, | ||
61 | + * or @size if there is no set bit in the bitmap. | ||
62 | */ | ||
63 | static inline unsigned long find_first_bit(const unsigned long *addr, | ||
64 | unsigned long size) | ||
65 | @@ -XXX,XX +XXX,XX @@ static inline unsigned long find_first_bit(const unsigned long *addr, | ||
66 | * @addr: The address to start the search at | ||
67 | * @size: The maximum size to search | ||
68 | * | ||
69 | - * Returns the bit number of the first cleared bit. | ||
70 | + * Returns the bit number of the first cleared bit, | ||
71 | + * or @size if there is no clear bit in the bitmap. | ||
72 | */ | ||
73 | static inline unsigned long find_first_zero_bit(const unsigned long *addr, | ||
74 | unsigned long size) | ||
75 | -- | ||
76 | 2.31.1 | ||
77 | diff view generated by jsdifflib |
Deleted patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
2 | 1 | ||
3 | The lifetime of the timer is well-known (it cannot outlive | ||
4 | qemu_co_sleep_ns_wakeable, because it's deleted by the time the | ||
5 | coroutine resumes), so it is not necessary to place it on the heap. | ||
6 | |||
7 | Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> | ||
8 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
9 | Message-id: 20210517100548.28806-2-pbonzini@redhat.com | ||
10 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
11 | --- | ||
12 | util/qemu-coroutine-sleep.c | 9 ++++----- | ||
13 | 1 file changed, 4 insertions(+), 5 deletions(-) | ||
14 | |||
15 | diff --git a/util/qemu-coroutine-sleep.c b/util/qemu-coroutine-sleep.c | ||
16 | index XXXXXXX..XXXXXXX 100644 | ||
17 | --- a/util/qemu-coroutine-sleep.c | ||
18 | +++ b/util/qemu-coroutine-sleep.c | ||
19 | @@ -XXX,XX +XXX,XX @@ static const char *qemu_co_sleep_ns__scheduled = "qemu_co_sleep_ns"; | ||
20 | |||
21 | struct QemuCoSleepState { | ||
22 | Coroutine *co; | ||
23 | - QEMUTimer *ts; | ||
24 | + QEMUTimer ts; | ||
25 | QemuCoSleepState **user_state_pointer; | ||
26 | }; | ||
27 | |||
28 | @@ -XXX,XX +XXX,XX @@ void qemu_co_sleep_wake(QemuCoSleepState *sleep_state) | ||
29 | if (sleep_state->user_state_pointer) { | ||
30 | *sleep_state->user_state_pointer = NULL; | ||
31 | } | ||
32 | - timer_del(sleep_state->ts); | ||
33 | + timer_del(&sleep_state->ts); | ||
34 | aio_co_wake(sleep_state->co); | ||
35 | } | ||
36 | |||
37 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
38 | AioContext *ctx = qemu_get_current_aio_context(); | ||
39 | QemuCoSleepState state = { | ||
40 | .co = qemu_coroutine_self(), | ||
41 | - .ts = aio_timer_new(ctx, type, SCALE_NS, co_sleep_cb, &state), | ||
42 | .user_state_pointer = sleep_state, | ||
43 | }; | ||
44 | |||
45 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
46 | abort(); | ||
47 | } | ||
48 | |||
49 | + aio_timer_init(ctx, &state.ts, type, SCALE_NS, co_sleep_cb, &state); | ||
50 | if (sleep_state) { | ||
51 | *sleep_state = &state; | ||
52 | } | ||
53 | - timer_mod(state.ts, qemu_clock_get_ns(type) + ns); | ||
54 | + timer_mod(&state.ts, qemu_clock_get_ns(type) + ns); | ||
55 | qemu_coroutine_yield(); | ||
56 | if (sleep_state) { | ||
57 | /* | ||
58 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
59 | */ | ||
60 | assert(*sleep_state == NULL); | ||
61 | } | ||
62 | - timer_free(state.ts); | ||
63 | } | ||
64 | -- | ||
65 | 2.31.1 | ||
66 | diff view generated by jsdifflib |
1 | From: Paolo Bonzini <pbonzini@redhat.com> | 1 | From: Stefano Garzarella <sgarzare@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | This simplification is enabled by the previous patch. Now aio_co_wake | 3 | Changes in preparation for next patches where we add a new |
4 | will only be called once, therefore we do not care about a spurious | 4 | parameter not related to the poll mechanism. |
5 | firing of the timer after a qemu_co_sleep_wake. | ||
6 | 5 | ||
7 | Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> | 6 | Let's add two new generic functions (iothread_set_param and |
8 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | 7 | iothread_get_param) that we use to set and get IOThread |
9 | Message-id: 20210517100548.28806-5-pbonzini@redhat.com | 8 | parameters. |
9 | |||
10 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | ||
11 | Message-id: 20210721094211.69853-2-sgarzare@redhat.com | ||
10 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 12 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
11 | --- | 13 | --- |
12 | util/qemu-coroutine-sleep.c | 8 ++++---- | 14 | iothread.c | 27 +++++++++++++++++++++++---- |
13 | 1 file changed, 4 insertions(+), 4 deletions(-) | 15 | 1 file changed, 23 insertions(+), 4 deletions(-) |
14 | 16 | ||
15 | diff --git a/util/qemu-coroutine-sleep.c b/util/qemu-coroutine-sleep.c | 17 | diff --git a/iothread.c b/iothread.c |
16 | index XXXXXXX..XXXXXXX 100644 | 18 | index XXXXXXX..XXXXXXX 100644 |
17 | --- a/util/qemu-coroutine-sleep.c | 19 | --- a/iothread.c |
18 | +++ b/util/qemu-coroutine-sleep.c | 20 | +++ b/iothread.c |
19 | @@ -XXX,XX +XXX,XX @@ static const char *qemu_co_sleep_ns__scheduled = "qemu_co_sleep_ns"; | 21 | @@ -XXX,XX +XXX,XX @@ static PollParamInfo poll_shrink_info = { |
20 | 22 | "poll-shrink", offsetof(IOThread, poll_shrink), | |
21 | struct QemuCoSleepState { | ||
22 | Coroutine *co; | ||
23 | - QEMUTimer ts; | ||
24 | QemuCoSleepState **user_state_pointer; | ||
25 | }; | 23 | }; |
26 | 24 | ||
27 | @@ -XXX,XX +XXX,XX @@ void qemu_co_sleep_wake(QemuCoSleepState *sleep_state) | 25 | -static void iothread_get_poll_param(Object *obj, Visitor *v, |
28 | 26 | +static void iothread_get_param(Object *obj, Visitor *v, | |
29 | assert(scheduled == qemu_co_sleep_ns__scheduled); | 27 | const char *name, void *opaque, Error **errp) |
30 | *sleep_state->user_state_pointer = NULL; | 28 | { |
31 | - timer_del(&sleep_state->ts); | 29 | IOThread *iothread = IOTHREAD(obj); |
32 | aio_co_wake(sleep_state->co); | 30 | @@ -XXX,XX +XXX,XX @@ static void iothread_get_poll_param(Object *obj, Visitor *v, |
31 | visit_type_int64(v, name, field, errp); | ||
32 | } | ||
33 | |||
34 | -static void iothread_set_poll_param(Object *obj, Visitor *v, | ||
35 | +static bool iothread_set_param(Object *obj, Visitor *v, | ||
36 | const char *name, void *opaque, Error **errp) | ||
37 | { | ||
38 | IOThread *iothread = IOTHREAD(obj); | ||
39 | @@ -XXX,XX +XXX,XX @@ static void iothread_set_poll_param(Object *obj, Visitor *v, | ||
40 | int64_t value; | ||
41 | |||
42 | if (!visit_type_int64(v, name, &value, errp)) { | ||
43 | - return; | ||
44 | + return false; | ||
33 | } | 45 | } |
34 | } | 46 | |
35 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | 47 | if (value < 0) { |
36 | QemuCoSleepState **sleep_state) | 48 | error_setg(errp, "%s value must be in range [0, %" PRId64 "]", |
37 | { | 49 | info->name, INT64_MAX); |
38 | AioContext *ctx = qemu_get_current_aio_context(); | 50 | - return; |
39 | + QEMUTimer ts; | 51 | + return false; |
40 | QemuCoSleepState state = { | ||
41 | .co = qemu_coroutine_self(), | ||
42 | .user_state_pointer = sleep_state, | ||
43 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
44 | abort(); | ||
45 | } | 52 | } |
46 | 53 | ||
47 | - aio_timer_init(ctx, &state.ts, type, SCALE_NS, co_sleep_cb, sleep_state); | 54 | *field = value; |
48 | + aio_timer_init(ctx, &ts, type, SCALE_NS, co_sleep_cb, sleep_state); | 55 | |
49 | *sleep_state = &state; | 56 | + return true; |
50 | - timer_mod(&state.ts, qemu_clock_get_ns(type) + ns); | 57 | +} |
51 | + timer_mod(&ts, qemu_clock_get_ns(type) + ns); | 58 | + |
52 | qemu_coroutine_yield(); | 59 | +static void iothread_get_poll_param(Object *obj, Visitor *v, |
53 | + timer_del(&ts); | 60 | + const char *name, void *opaque, Error **errp) |
54 | 61 | +{ | |
55 | /* qemu_co_sleep_wake clears *sleep_state before resuming this coroutine. */ | 62 | + |
56 | assert(*sleep_state == NULL); | 63 | + iothread_get_param(obj, v, name, opaque, errp); |
64 | +} | ||
65 | + | ||
66 | +static void iothread_set_poll_param(Object *obj, Visitor *v, | ||
67 | + const char *name, void *opaque, Error **errp) | ||
68 | +{ | ||
69 | + IOThread *iothread = IOTHREAD(obj); | ||
70 | + | ||
71 | + if (!iothread_set_param(obj, v, name, opaque, errp)) { | ||
72 | + return; | ||
73 | + } | ||
74 | + | ||
75 | if (iothread->ctx) { | ||
76 | aio_context_set_poll_params(iothread->ctx, | ||
77 | iothread->poll_max_ns, | ||
57 | -- | 78 | -- |
58 | 2.31.1 | 79 | 2.31.1 |
59 | 80 | diff view generated by jsdifflib |
1 | From: Paolo Bonzini <pbonzini@redhat.com> | 1 | From: Stefano Garzarella <sgarzare@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | Allow using QemuCoSleep to sleep forever until woken by qemu_co_sleep_wake. | 3 | The `aio-max-batch` parameter will be propagated to AIO engines |
4 | This makes the logic of qemu_co_sleep_ns_wakeable easy to understand. | 4 | and it will be used to control the maximum number of queued requests. |
5 | 5 | ||
6 | In the future we will introduce an API that can work even if the | 6 | When there are in queue a number of requests equal to `aio-max-batch`, |
7 | sleep and wake happen from different threads. For now, initializing | 7 | the engine invokes the system call to forward the requests to the kernel. |
8 | w->to_wake after timer_mod is fine because the timer can only fire in | 8 | |
9 | the same AioContext. | 9 | This parameter allows us to control the maximum batch size to reduce |
10 | 10 | the latency that requests might accumulate while queued in the AIO | |
11 | Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> | 11 | engine queue. |
12 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | 12 | |
13 | Message-id: 20210517100548.28806-7-pbonzini@redhat.com | 13 | If `aio-max-batch` is equal to 0 (default value), the AIO engine will |
14 | use its default maximum batch size value. | ||
15 | |||
16 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | ||
17 | Message-id: 20210721094211.69853-3-sgarzare@redhat.com | ||
14 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 18 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
15 | --- | 19 | --- |
16 | include/qemu/coroutine.h | 5 +++++ | 20 | qapi/misc.json | 6 ++++- |
17 | util/qemu-coroutine-sleep.c | 26 +++++++++++++++++++------- | 21 | qapi/qom.json | 7 ++++- |
18 | 2 files changed, 24 insertions(+), 7 deletions(-) | 22 | include/block/aio.h | 12 +++++++++ |
19 | 23 | include/sysemu/iothread.h | 3 +++ | |
20 | diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h | 24 | iothread.c | 55 +++++++++++++++++++++++++++++++++++---- |
21 | index XXXXXXX..XXXXXXX 100644 | 25 | monitor/hmp-cmds.c | 2 ++ |
22 | --- a/include/qemu/coroutine.h | 26 | util/aio-posix.c | 12 +++++++++ |
23 | +++ b/include/qemu/coroutine.h | 27 | util/aio-win32.c | 5 ++++ |
24 | @@ -XXX,XX +XXX,XX @@ typedef struct QemuCoSleep { | 28 | util/async.c | 2 ++ |
25 | void coroutine_fn qemu_co_sleep_ns_wakeable(QemuCoSleep *w, | 29 | qemu-options.hx | 8 ++++-- |
26 | QEMUClockType type, int64_t ns); | 30 | 10 files changed, 103 insertions(+), 9 deletions(-) |
31 | |||
32 | diff --git a/qapi/misc.json b/qapi/misc.json | ||
33 | index XXXXXXX..XXXXXXX 100644 | ||
34 | --- a/qapi/misc.json | ||
35 | +++ b/qapi/misc.json | ||
36 | @@ -XXX,XX +XXX,XX @@ | ||
37 | # @poll-shrink: how many ns will be removed from polling time, 0 means that | ||
38 | # it's not configured (since 2.9) | ||
39 | # | ||
40 | +# @aio-max-batch: maximum number of requests in a batch for the AIO engine, | ||
41 | +# 0 means that the engine will use its default (since 6.1) | ||
42 | +# | ||
43 | # Since: 2.0 | ||
44 | ## | ||
45 | { 'struct': 'IOThreadInfo', | ||
46 | @@ -XXX,XX +XXX,XX @@ | ||
47 | 'thread-id': 'int', | ||
48 | 'poll-max-ns': 'int', | ||
49 | 'poll-grow': 'int', | ||
50 | - 'poll-shrink': 'int' } } | ||
51 | + 'poll-shrink': 'int', | ||
52 | + 'aio-max-batch': 'int' } } | ||
53 | |||
54 | ## | ||
55 | # @query-iothreads: | ||
56 | diff --git a/qapi/qom.json b/qapi/qom.json | ||
57 | index XXXXXXX..XXXXXXX 100644 | ||
58 | --- a/qapi/qom.json | ||
59 | +++ b/qapi/qom.json | ||
60 | @@ -XXX,XX +XXX,XX @@ | ||
61 | # algorithm detects it is spending too long polling without | ||
62 | # encountering events. 0 selects a default behaviour (default: 0) | ||
63 | # | ||
64 | +# @aio-max-batch: maximum number of requests in a batch for the AIO engine, | ||
65 | +# 0 means that the engine will use its default | ||
66 | +# (default:0, since 6.1) | ||
67 | +# | ||
68 | # Since: 2.0 | ||
69 | ## | ||
70 | { 'struct': 'IothreadProperties', | ||
71 | 'data': { '*poll-max-ns': 'int', | ||
72 | '*poll-grow': 'int', | ||
73 | - '*poll-shrink': 'int' } } | ||
74 | + '*poll-shrink': 'int', | ||
75 | + '*aio-max-batch': 'int' } } | ||
76 | |||
77 | ## | ||
78 | # @MemoryBackendProperties: | ||
79 | diff --git a/include/block/aio.h b/include/block/aio.h | ||
80 | index XXXXXXX..XXXXXXX 100644 | ||
81 | --- a/include/block/aio.h | ||
82 | +++ b/include/block/aio.h | ||
83 | @@ -XXX,XX +XXX,XX @@ struct AioContext { | ||
84 | int64_t poll_grow; /* polling time growth factor */ | ||
85 | int64_t poll_shrink; /* polling time shrink factor */ | ||
86 | |||
87 | + /* AIO engine parameters */ | ||
88 | + int64_t aio_max_batch; /* maximum number of requests in a batch */ | ||
89 | + | ||
90 | /* | ||
91 | * List of handlers participating in userspace polling. Protected by | ||
92 | * ctx->list_lock. Iterated and modified mostly by the event loop thread | ||
93 | @@ -XXX,XX +XXX,XX @@ void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns, | ||
94 | int64_t grow, int64_t shrink, | ||
95 | Error **errp); | ||
27 | 96 | ||
28 | +/** | 97 | +/** |
29 | + * Yield the coroutine until the next call to qemu_co_sleep_wake. | 98 | + * aio_context_set_aio_params: |
99 | + * @ctx: the aio context | ||
100 | + * @max_batch: maximum number of requests in a batch, 0 means that the | ||
101 | + * engine will use its default | ||
30 | + */ | 102 | + */ |
31 | +void coroutine_fn qemu_co_sleep(QemuCoSleep *w); | 103 | +void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch, |
32 | + | 104 | + Error **errp); |
33 | static inline void coroutine_fn qemu_co_sleep_ns(QEMUClockType type, int64_t ns) | 105 | + |
106 | #endif | ||
107 | diff --git a/include/sysemu/iothread.h b/include/sysemu/iothread.h | ||
108 | index XXXXXXX..XXXXXXX 100644 | ||
109 | --- a/include/sysemu/iothread.h | ||
110 | +++ b/include/sysemu/iothread.h | ||
111 | @@ -XXX,XX +XXX,XX @@ struct IOThread { | ||
112 | int64_t poll_max_ns; | ||
113 | int64_t poll_grow; | ||
114 | int64_t poll_shrink; | ||
115 | + | ||
116 | + /* AioContext AIO engine parameters */ | ||
117 | + int64_t aio_max_batch; | ||
118 | }; | ||
119 | typedef struct IOThread IOThread; | ||
120 | |||
121 | diff --git a/iothread.c b/iothread.c | ||
122 | index XXXXXXX..XXXXXXX 100644 | ||
123 | --- a/iothread.c | ||
124 | +++ b/iothread.c | ||
125 | @@ -XXX,XX +XXX,XX @@ static void iothread_init_gcontext(IOThread *iothread) | ||
126 | iothread->main_loop = g_main_loop_new(iothread->worker_context, TRUE); | ||
127 | } | ||
128 | |||
129 | +static void iothread_set_aio_context_params(IOThread *iothread, Error **errp) | ||
130 | +{ | ||
131 | + ERRP_GUARD(); | ||
132 | + | ||
133 | + aio_context_set_poll_params(iothread->ctx, | ||
134 | + iothread->poll_max_ns, | ||
135 | + iothread->poll_grow, | ||
136 | + iothread->poll_shrink, | ||
137 | + errp); | ||
138 | + if (*errp) { | ||
139 | + return; | ||
140 | + } | ||
141 | + | ||
142 | + aio_context_set_aio_params(iothread->ctx, | ||
143 | + iothread->aio_max_batch, | ||
144 | + errp); | ||
145 | +} | ||
146 | + | ||
147 | static void iothread_complete(UserCreatable *obj, Error **errp) | ||
34 | { | 148 | { |
35 | QemuCoSleep w = { 0 }; | 149 | Error *local_error = NULL; |
36 | diff --git a/util/qemu-coroutine-sleep.c b/util/qemu-coroutine-sleep.c | 150 | @@ -XXX,XX +XXX,XX @@ static void iothread_complete(UserCreatable *obj, Error **errp) |
37 | index XXXXXXX..XXXXXXX 100644 | 151 | */ |
38 | --- a/util/qemu-coroutine-sleep.c | 152 | iothread_init_gcontext(iothread); |
39 | +++ b/util/qemu-coroutine-sleep.c | 153 | |
40 | @@ -XXX,XX +XXX,XX @@ static void co_sleep_cb(void *opaque) | 154 | - aio_context_set_poll_params(iothread->ctx, |
41 | qemu_co_sleep_wake(w); | 155 | - iothread->poll_max_ns, |
42 | } | 156 | - iothread->poll_grow, |
43 | 157 | - iothread->poll_shrink, | |
44 | -void coroutine_fn qemu_co_sleep_ns_wakeable(QemuCoSleep *w, | 158 | - &local_error); |
45 | - QEMUClockType type, int64_t ns) | 159 | + iothread_set_aio_context_params(iothread, &local_error); |
46 | +void coroutine_fn qemu_co_sleep(QemuCoSleep *w) | 160 | if (local_error) { |
161 | error_propagate(errp, local_error); | ||
162 | aio_context_unref(iothread->ctx); | ||
163 | @@ -XXX,XX +XXX,XX @@ static PollParamInfo poll_grow_info = { | ||
164 | static PollParamInfo poll_shrink_info = { | ||
165 | "poll-shrink", offsetof(IOThread, poll_shrink), | ||
166 | }; | ||
167 | +static PollParamInfo aio_max_batch_info = { | ||
168 | + "aio-max-batch", offsetof(IOThread, aio_max_batch), | ||
169 | +}; | ||
170 | |||
171 | static void iothread_get_param(Object *obj, Visitor *v, | ||
172 | const char *name, void *opaque, Error **errp) | ||
173 | @@ -XXX,XX +XXX,XX @@ static void iothread_set_poll_param(Object *obj, Visitor *v, | ||
174 | } | ||
175 | } | ||
176 | |||
177 | +static void iothread_get_aio_param(Object *obj, Visitor *v, | ||
178 | + const char *name, void *opaque, Error **errp) | ||
179 | +{ | ||
180 | + | ||
181 | + iothread_get_param(obj, v, name, opaque, errp); | ||
182 | +} | ||
183 | + | ||
184 | +static void iothread_set_aio_param(Object *obj, Visitor *v, | ||
185 | + const char *name, void *opaque, Error **errp) | ||
186 | +{ | ||
187 | + IOThread *iothread = IOTHREAD(obj); | ||
188 | + | ||
189 | + if (!iothread_set_param(obj, v, name, opaque, errp)) { | ||
190 | + return; | ||
191 | + } | ||
192 | + | ||
193 | + if (iothread->ctx) { | ||
194 | + aio_context_set_aio_params(iothread->ctx, | ||
195 | + iothread->aio_max_batch, | ||
196 | + errp); | ||
197 | + } | ||
198 | +} | ||
199 | + | ||
200 | static void iothread_class_init(ObjectClass *klass, void *class_data) | ||
47 | { | 201 | { |
48 | Coroutine *co = qemu_coroutine_self(); | 202 | UserCreatableClass *ucc = USER_CREATABLE_CLASS(klass); |
49 | - AioContext *ctx = qemu_get_current_aio_context(); | 203 | @@ -XXX,XX +XXX,XX @@ static void iothread_class_init(ObjectClass *klass, void *class_data) |
50 | - QEMUTimer ts; | 204 | iothread_get_poll_param, |
51 | 205 | iothread_set_poll_param, | |
52 | const char *scheduled = qatomic_cmpxchg(&co->scheduled, NULL, | 206 | NULL, &poll_shrink_info); |
53 | qemu_co_sleep_ns__scheduled); | 207 | + object_class_property_add(klass, "aio-max-batch", "int", |
54 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QemuCoSleep *w, | 208 | + iothread_get_aio_param, |
209 | + iothread_set_aio_param, | ||
210 | + NULL, &aio_max_batch_info); | ||
211 | } | ||
212 | |||
213 | static const TypeInfo iothread_info = { | ||
214 | @@ -XXX,XX +XXX,XX @@ static int query_one_iothread(Object *object, void *opaque) | ||
215 | info->poll_max_ns = iothread->poll_max_ns; | ||
216 | info->poll_grow = iothread->poll_grow; | ||
217 | info->poll_shrink = iothread->poll_shrink; | ||
218 | + info->aio_max_batch = iothread->aio_max_batch; | ||
219 | |||
220 | QAPI_LIST_APPEND(*tail, info); | ||
221 | return 0; | ||
222 | diff --git a/monitor/hmp-cmds.c b/monitor/hmp-cmds.c | ||
223 | index XXXXXXX..XXXXXXX 100644 | ||
224 | --- a/monitor/hmp-cmds.c | ||
225 | +++ b/monitor/hmp-cmds.c | ||
226 | @@ -XXX,XX +XXX,XX @@ void hmp_info_iothreads(Monitor *mon, const QDict *qdict) | ||
227 | monitor_printf(mon, " poll-max-ns=%" PRId64 "\n", value->poll_max_ns); | ||
228 | monitor_printf(mon, " poll-grow=%" PRId64 "\n", value->poll_grow); | ||
229 | monitor_printf(mon, " poll-shrink=%" PRId64 "\n", value->poll_shrink); | ||
230 | + monitor_printf(mon, " aio-max-batch=%" PRId64 "\n", | ||
231 | + value->aio_max_batch); | ||
55 | } | 232 | } |
56 | 233 | ||
57 | w->to_wake = co; | 234 | qapi_free_IOThreadInfoList(info_list); |
58 | - aio_timer_init(ctx, &ts, type, SCALE_NS, co_sleep_cb, w), | 235 | diff --git a/util/aio-posix.c b/util/aio-posix.c |
59 | - timer_mod(&ts, qemu_clock_get_ns(type) + ns); | 236 | index XXXXXXX..XXXXXXX 100644 |
60 | qemu_coroutine_yield(); | 237 | --- a/util/aio-posix.c |
61 | - timer_del(&ts); | 238 | +++ b/util/aio-posix.c |
62 | 239 | @@ -XXX,XX +XXX,XX @@ void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns, | |
63 | /* w->to_wake is cleared before resuming this coroutine. */ | 240 | |
64 | assert(w->to_wake == NULL); | 241 | aio_notify(ctx); |
65 | } | 242 | } |
66 | + | 243 | + |
67 | +void coroutine_fn qemu_co_sleep_ns_wakeable(QemuCoSleep *w, | 244 | +void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch, |
68 | + QEMUClockType type, int64_t ns) | 245 | + Error **errp) |
69 | +{ | 246 | +{ |
70 | + AioContext *ctx = qemu_get_current_aio_context(); | ||
71 | + QEMUTimer ts; | ||
72 | + | ||
73 | + aio_timer_init(ctx, &ts, type, SCALE_NS, co_sleep_cb, w); | ||
74 | + timer_mod(&ts, qemu_clock_get_ns(type) + ns); | ||
75 | + | ||
76 | + /* | 247 | + /* |
77 | + * The timer will fire in the current AiOContext, so the callback | 248 | + * No thread synchronization here, it doesn't matter if an incorrect value |
78 | + * must happen after qemu_co_sleep yields and there is no race | 249 | + * is used once. |
79 | + * between timer_mod and qemu_co_sleep. | ||
80 | + */ | 250 | + */ |
81 | + qemu_co_sleep(w); | 251 | + ctx->aio_max_batch = max_batch; |
82 | + timer_del(&ts); | 252 | + |
83 | +} | 253 | + aio_notify(ctx); |
254 | +} | ||
255 | diff --git a/util/aio-win32.c b/util/aio-win32.c | ||
256 | index XXXXXXX..XXXXXXX 100644 | ||
257 | --- a/util/aio-win32.c | ||
258 | +++ b/util/aio-win32.c | ||
259 | @@ -XXX,XX +XXX,XX @@ void aio_context_set_poll_params(AioContext *ctx, int64_t max_ns, | ||
260 | error_setg(errp, "AioContext polling is not implemented on Windows"); | ||
261 | } | ||
262 | } | ||
263 | + | ||
264 | +void aio_context_set_aio_params(AioContext *ctx, int64_t max_batch, | ||
265 | + Error **errp) | ||
266 | +{ | ||
267 | +} | ||
268 | diff --git a/util/async.c b/util/async.c | ||
269 | index XXXXXXX..XXXXXXX 100644 | ||
270 | --- a/util/async.c | ||
271 | +++ b/util/async.c | ||
272 | @@ -XXX,XX +XXX,XX @@ AioContext *aio_context_new(Error **errp) | ||
273 | ctx->poll_grow = 0; | ||
274 | ctx->poll_shrink = 0; | ||
275 | |||
276 | + ctx->aio_max_batch = 0; | ||
277 | + | ||
278 | return ctx; | ||
279 | fail: | ||
280 | g_source_destroy(&ctx->source); | ||
281 | diff --git a/qemu-options.hx b/qemu-options.hx | ||
282 | index XXXXXXX..XXXXXXX 100644 | ||
283 | --- a/qemu-options.hx | ||
284 | +++ b/qemu-options.hx | ||
285 | @@ -XXX,XX +XXX,XX @@ SRST | ||
286 | |||
287 | CN=laptop.example.com,O=Example Home,L=London,ST=London,C=GB | ||
288 | |||
289 | - ``-object iothread,id=id,poll-max-ns=poll-max-ns,poll-grow=poll-grow,poll-shrink=poll-shrink`` | ||
290 | + ``-object iothread,id=id,poll-max-ns=poll-max-ns,poll-grow=poll-grow,poll-shrink=poll-shrink,aio-max-batch=aio-max-batch`` | ||
291 | Creates a dedicated event loop thread that devices can be | ||
292 | assigned to. This is known as an IOThread. By default device | ||
293 | emulation happens in vCPU threads or the main event loop thread. | ||
294 | @@ -XXX,XX +XXX,XX @@ SRST | ||
295 | the polling time when the algorithm detects it is spending too | ||
296 | long polling without encountering events. | ||
297 | |||
298 | - The polling parameters can be modified at run-time using the | ||
299 | + The ``aio-max-batch`` parameter is the maximum number of requests | ||
300 | + in a batch for the AIO engine, 0 means that the engine will use | ||
301 | + its default. | ||
302 | + | ||
303 | + The IOThread parameters can be modified at run-time using the | ||
304 | ``qom-set`` command (where ``iothread1`` is the IOThread's | ||
305 | ``id``): | ||
306 | |||
84 | -- | 307 | -- |
85 | 2.31.1 | 308 | 2.31.1 |
86 | 309 | diff view generated by jsdifflib |
1 | From: Paolo Bonzini <pbonzini@redhat.com> | 1 | From: Stefano Garzarella <sgarzare@redhat.com> |
---|---|---|---|
2 | 2 | ||
3 | Simplify the code by removing conditionals. qemu_co_sleep_ns | 3 | When there are multiple queues attached to the same AIO context, |
4 | can simply point the argument to an on-stack temporary. | 4 | some requests may experience high latency, since in the worst case |
5 | the AIO engine queue is only flushed when it is full (MAX_EVENTS) or | ||
6 | there are no more queues plugged. | ||
5 | 7 | ||
6 | Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> | 8 | Commit 2558cb8dd4 ("linux-aio: increasing MAX_EVENTS to a larger |
7 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | 9 | hardcoded value") changed MAX_EVENTS from 128 to 1024, to increase |
8 | Message-id: 20210517100548.28806-3-pbonzini@redhat.com | 10 | the number of in-flight requests. But this change also increased |
11 | the potential maximum batch to 1024 elements. | ||
12 | |||
13 | When there is a single queue attached to the AIO context, the issue | ||
14 | is mitigated from laio_io_unplug() that will flush the queue every | ||
15 | time is invoked since there can't be others queue plugged. | ||
16 | |||
17 | Let's use the new `aio-max-batch` IOThread parameter to mitigate | ||
18 | this issue, limiting the number of requests in a batch. | ||
19 | |||
20 | We also define a default value (32): this value is obtained running | ||
21 | some benchmarks and it represents a good tradeoff between the latency | ||
22 | increase while a request is queued and the cost of the io_submit(2) | ||
23 | system call. | ||
24 | |||
25 | Signed-off-by: Stefano Garzarella <sgarzare@redhat.com> | ||
26 | Message-id: 20210721094211.69853-4-sgarzare@redhat.com | ||
9 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | 27 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> |
10 | --- | 28 | --- |
11 | include/qemu/coroutine.h | 5 +++-- | 29 | block/linux-aio.c | 9 ++++++++- |
12 | util/qemu-coroutine-sleep.c | 18 +++++------------- | 30 | 1 file changed, 8 insertions(+), 1 deletion(-) |
13 | 2 files changed, 8 insertions(+), 15 deletions(-) | ||
14 | 31 | ||
15 | diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h | 32 | diff --git a/block/linux-aio.c b/block/linux-aio.c |
16 | index XXXXXXX..XXXXXXX 100644 | 33 | index XXXXXXX..XXXXXXX 100644 |
17 | --- a/include/qemu/coroutine.h | 34 | --- a/block/linux-aio.c |
18 | +++ b/include/qemu/coroutine.h | 35 | +++ b/block/linux-aio.c |
19 | @@ -XXX,XX +XXX,XX @@ typedef struct QemuCoSleepState QemuCoSleepState; | 36 | @@ -XXX,XX +XXX,XX @@ |
20 | 37 | */ | |
21 | /** | 38 | #define MAX_EVENTS 1024 |
22 | * Yield the coroutine for a given duration. During this yield, @sleep_state | 39 | |
23 | - * (if not NULL) is set to an opaque pointer, which may be used for | 40 | +/* Maximum number of requests in a batch. (default value) */ |
24 | + * is set to an opaque pointer, which may be used for | 41 | +#define DEFAULT_MAX_BATCH 32 |
25 | * qemu_co_sleep_wake(). Be careful, the pointer is set back to zero when the | 42 | + |
26 | * timer fires. Don't save the obtained value to other variables and don't call | 43 | struct qemu_laiocb { |
27 | * qemu_co_sleep_wake from another aio context. | 44 | Coroutine *co; |
28 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | 45 | LinuxAioState *ctx; |
29 | QemuCoSleepState **sleep_state); | 46 | @@ -XXX,XX +XXX,XX @@ static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset, |
30 | static inline void coroutine_fn qemu_co_sleep_ns(QEMUClockType type, int64_t ns) | 47 | LinuxAioState *s = laiocb->ctx; |
31 | { | 48 | struct iocb *iocbs = &laiocb->iocb; |
32 | - qemu_co_sleep_ns_wakeable(type, ns, NULL); | 49 | QEMUIOVector *qiov = laiocb->qiov; |
33 | + QemuCoSleepState *unused = NULL; | 50 | + int64_t max_batch = s->aio_context->aio_max_batch ?: DEFAULT_MAX_BATCH; |
34 | + qemu_co_sleep_ns_wakeable(type, ns, &unused); | 51 | + |
35 | } | 52 | + /* limit the batch with the number of available events */ |
36 | 53 | + max_batch = MIN_NON_ZERO(MAX_EVENTS - s->io_q.in_flight, max_batch); | |
37 | /** | 54 | |
38 | diff --git a/util/qemu-coroutine-sleep.c b/util/qemu-coroutine-sleep.c | 55 | switch (type) { |
39 | index XXXXXXX..XXXXXXX 100644 | 56 | case QEMU_AIO_WRITE: |
40 | --- a/util/qemu-coroutine-sleep.c | 57 | @@ -XXX,XX +XXX,XX @@ static int laio_do_submit(int fd, struct qemu_laiocb *laiocb, off_t offset, |
41 | +++ b/util/qemu-coroutine-sleep.c | 58 | s->io_q.in_queue++; |
42 | @@ -XXX,XX +XXX,XX @@ void qemu_co_sleep_wake(QemuCoSleepState *sleep_state) | 59 | if (!s->io_q.blocked && |
43 | qemu_co_sleep_ns__scheduled, NULL); | 60 | (!s->io_q.plugged || |
44 | 61 | - s->io_q.in_flight + s->io_q.in_queue >= MAX_EVENTS)) { | |
45 | assert(scheduled == qemu_co_sleep_ns__scheduled); | 62 | + s->io_q.in_queue >= max_batch)) { |
46 | - if (sleep_state->user_state_pointer) { | 63 | ioq_submit(s); |
47 | - *sleep_state->user_state_pointer = NULL; | ||
48 | - } | ||
49 | + *sleep_state->user_state_pointer = NULL; | ||
50 | timer_del(&sleep_state->ts); | ||
51 | aio_co_wake(sleep_state->co); | ||
52 | } | ||
53 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
54 | } | 64 | } |
55 | 65 | ||
56 | aio_timer_init(ctx, &state.ts, type, SCALE_NS, co_sleep_cb, &state); | ||
57 | - if (sleep_state) { | ||
58 | - *sleep_state = &state; | ||
59 | - } | ||
60 | + *sleep_state = &state; | ||
61 | timer_mod(&state.ts, qemu_clock_get_ns(type) + ns); | ||
62 | qemu_coroutine_yield(); | ||
63 | - if (sleep_state) { | ||
64 | - /* | ||
65 | - * Note that *sleep_state is cleared during qemu_co_sleep_wake | ||
66 | - * before resuming this coroutine. | ||
67 | - */ | ||
68 | - assert(*sleep_state == NULL); | ||
69 | - } | ||
70 | + | ||
71 | + /* qemu_co_sleep_wake clears *sleep_state before resuming this coroutine. */ | ||
72 | + assert(*sleep_state == NULL); | ||
73 | } | ||
74 | -- | 66 | -- |
75 | 2.31.1 | 67 | 2.31.1 |
76 | 68 | diff view generated by jsdifflib |
Deleted patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
2 | 1 | ||
3 | All callers of qemu_co_sleep_wake are checking whether they are passing | ||
4 | a NULL argument inside the pointer-to-pointer: do the check in | ||
5 | qemu_co_sleep_wake itself. | ||
6 | |||
7 | As a side effect, qemu_co_sleep_wake can be called more than once and | ||
8 | it will only wake the coroutine once; after the first time, the argument | ||
9 | will be set to NULL via *sleep_state->user_state_pointer. However, this | ||
10 | would not be safe unless co_sleep_cb keeps using the QemuCoSleepState* | ||
11 | directly, so make it go through the pointer-to-pointer instead. | ||
12 | |||
13 | Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> | ||
14 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
15 | Message-id: 20210517100548.28806-4-pbonzini@redhat.com | ||
16 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
17 | --- | ||
18 | block/block-copy.c | 4 +--- | ||
19 | block/nbd.c | 8 ++------ | ||
20 | util/qemu-coroutine-sleep.c | 21 ++++++++++++--------- | ||
21 | 3 files changed, 15 insertions(+), 18 deletions(-) | ||
22 | |||
23 | diff --git a/block/block-copy.c b/block/block-copy.c | ||
24 | index XXXXXXX..XXXXXXX 100644 | ||
25 | --- a/block/block-copy.c | ||
26 | +++ b/block/block-copy.c | ||
27 | @@ -XXX,XX +XXX,XX @@ out: | ||
28 | |||
29 | void block_copy_kick(BlockCopyCallState *call_state) | ||
30 | { | ||
31 | - if (call_state->sleep_state) { | ||
32 | - qemu_co_sleep_wake(call_state->sleep_state); | ||
33 | - } | ||
34 | + qemu_co_sleep_wake(call_state->sleep_state); | ||
35 | } | ||
36 | |||
37 | /* | ||
38 | diff --git a/block/nbd.c b/block/nbd.c | ||
39 | index XXXXXXX..XXXXXXX 100644 | ||
40 | --- a/block/nbd.c | ||
41 | +++ b/block/nbd.c | ||
42 | @@ -XXX,XX +XXX,XX @@ static void coroutine_fn nbd_client_co_drain_begin(BlockDriverState *bs) | ||
43 | BDRVNBDState *s = (BDRVNBDState *)bs->opaque; | ||
44 | |||
45 | s->drained = true; | ||
46 | - if (s->connection_co_sleep_ns_state) { | ||
47 | - qemu_co_sleep_wake(s->connection_co_sleep_ns_state); | ||
48 | - } | ||
49 | + qemu_co_sleep_wake(s->connection_co_sleep_ns_state); | ||
50 | |||
51 | nbd_co_establish_connection_cancel(bs, false); | ||
52 | |||
53 | @@ -XXX,XX +XXX,XX @@ static void nbd_teardown_connection(BlockDriverState *bs) | ||
54 | |||
55 | s->state = NBD_CLIENT_QUIT; | ||
56 | if (s->connection_co) { | ||
57 | - if (s->connection_co_sleep_ns_state) { | ||
58 | - qemu_co_sleep_wake(s->connection_co_sleep_ns_state); | ||
59 | - } | ||
60 | + qemu_co_sleep_wake(s->connection_co_sleep_ns_state); | ||
61 | nbd_co_establish_connection_cancel(bs, true); | ||
62 | } | ||
63 | if (qemu_in_coroutine()) { | ||
64 | diff --git a/util/qemu-coroutine-sleep.c b/util/qemu-coroutine-sleep.c | ||
65 | index XXXXXXX..XXXXXXX 100644 | ||
66 | --- a/util/qemu-coroutine-sleep.c | ||
67 | +++ b/util/qemu-coroutine-sleep.c | ||
68 | @@ -XXX,XX +XXX,XX @@ struct QemuCoSleepState { | ||
69 | |||
70 | void qemu_co_sleep_wake(QemuCoSleepState *sleep_state) | ||
71 | { | ||
72 | - /* Write of schedule protected by barrier write in aio_co_schedule */ | ||
73 | - const char *scheduled = qatomic_cmpxchg(&sleep_state->co->scheduled, | ||
74 | - qemu_co_sleep_ns__scheduled, NULL); | ||
75 | + if (sleep_state) { | ||
76 | + /* Write of schedule protected by barrier write in aio_co_schedule */ | ||
77 | + const char *scheduled = qatomic_cmpxchg(&sleep_state->co->scheduled, | ||
78 | + qemu_co_sleep_ns__scheduled, NULL); | ||
79 | |||
80 | - assert(scheduled == qemu_co_sleep_ns__scheduled); | ||
81 | - *sleep_state->user_state_pointer = NULL; | ||
82 | - timer_del(&sleep_state->ts); | ||
83 | - aio_co_wake(sleep_state->co); | ||
84 | + assert(scheduled == qemu_co_sleep_ns__scheduled); | ||
85 | + *sleep_state->user_state_pointer = NULL; | ||
86 | + timer_del(&sleep_state->ts); | ||
87 | + aio_co_wake(sleep_state->co); | ||
88 | + } | ||
89 | } | ||
90 | |||
91 | static void co_sleep_cb(void *opaque) | ||
92 | { | ||
93 | - qemu_co_sleep_wake(opaque); | ||
94 | + QemuCoSleepState **sleep_state = opaque; | ||
95 | + qemu_co_sleep_wake(*sleep_state); | ||
96 | } | ||
97 | |||
98 | void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
99 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
100 | abort(); | ||
101 | } | ||
102 | |||
103 | - aio_timer_init(ctx, &state.ts, type, SCALE_NS, co_sleep_cb, &state); | ||
104 | + aio_timer_init(ctx, &state.ts, type, SCALE_NS, co_sleep_cb, sleep_state); | ||
105 | *sleep_state = &state; | ||
106 | timer_mod(&state.ts, qemu_clock_get_ns(type) + ns); | ||
107 | qemu_coroutine_yield(); | ||
108 | -- | ||
109 | 2.31.1 | ||
110 | diff view generated by jsdifflib |
Deleted patch | |||
---|---|---|---|
1 | From: Paolo Bonzini <pbonzini@redhat.com> | ||
2 | 1 | ||
3 | Right now, users of qemu_co_sleep_ns_wakeable are simply passing | ||
4 | a pointer to QemuCoSleepState by reference to the function. But | ||
5 | QemuCoSleepState really is just a Coroutine*; making the | ||
6 | content of the struct public is just as efficient and lets us | ||
7 | skip the user_state_pointer indirection. | ||
8 | |||
9 | Since the usage is changed, take the occasion to rename the | ||
10 | struct to QemuCoSleep. | ||
11 | |||
12 | Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com> | ||
13 | Signed-off-by: Paolo Bonzini <pbonzini@redhat.com> | ||
14 | Message-id: 20210517100548.28806-6-pbonzini@redhat.com | ||
15 | Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com> | ||
16 | --- | ||
17 | include/qemu/coroutine.h | 23 +++++++++++---------- | ||
18 | block/block-copy.c | 8 ++++---- | ||
19 | block/nbd.c | 10 ++++----- | ||
20 | util/qemu-coroutine-sleep.c | 41 ++++++++++++++++--------------------- | ||
21 | 4 files changed, 39 insertions(+), 43 deletions(-) | ||
22 | |||
23 | diff --git a/include/qemu/coroutine.h b/include/qemu/coroutine.h | ||
24 | index XXXXXXX..XXXXXXX 100644 | ||
25 | --- a/include/qemu/coroutine.h | ||
26 | +++ b/include/qemu/coroutine.h | ||
27 | @@ -XXX,XX +XXX,XX @@ void qemu_co_rwlock_wrlock(CoRwlock *lock); | ||
28 | */ | ||
29 | void qemu_co_rwlock_unlock(CoRwlock *lock); | ||
30 | |||
31 | -typedef struct QemuCoSleepState QemuCoSleepState; | ||
32 | +typedef struct QemuCoSleep { | ||
33 | + Coroutine *to_wake; | ||
34 | +} QemuCoSleep; | ||
35 | |||
36 | /** | ||
37 | - * Yield the coroutine for a given duration. During this yield, @sleep_state | ||
38 | - * is set to an opaque pointer, which may be used for | ||
39 | - * qemu_co_sleep_wake(). Be careful, the pointer is set back to zero when the | ||
40 | - * timer fires. Don't save the obtained value to other variables and don't call | ||
41 | - * qemu_co_sleep_wake from another aio context. | ||
42 | + * Yield the coroutine for a given duration. Initializes @w so that, | ||
43 | + * during this yield, it can be passed to qemu_co_sleep_wake() to | ||
44 | + * terminate the sleep. | ||
45 | */ | ||
46 | -void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
47 | - QemuCoSleepState **sleep_state); | ||
48 | +void coroutine_fn qemu_co_sleep_ns_wakeable(QemuCoSleep *w, | ||
49 | + QEMUClockType type, int64_t ns); | ||
50 | + | ||
51 | static inline void coroutine_fn qemu_co_sleep_ns(QEMUClockType type, int64_t ns) | ||
52 | { | ||
53 | - QemuCoSleepState *unused = NULL; | ||
54 | - qemu_co_sleep_ns_wakeable(type, ns, &unused); | ||
55 | + QemuCoSleep w = { 0 }; | ||
56 | + qemu_co_sleep_ns_wakeable(&w, type, ns); | ||
57 | } | ||
58 | |||
59 | /** | ||
60 | @@ -XXX,XX +XXX,XX @@ static inline void coroutine_fn qemu_co_sleep_ns(QEMUClockType type, int64_t ns) | ||
61 | * qemu_co_sleep_ns() and should be checked to be non-NULL before calling | ||
62 | * qemu_co_sleep_wake(). | ||
63 | */ | ||
64 | -void qemu_co_sleep_wake(QemuCoSleepState *sleep_state); | ||
65 | +void qemu_co_sleep_wake(QemuCoSleep *w); | ||
66 | |||
67 | /** | ||
68 | * Yield until a file descriptor becomes readable | ||
69 | diff --git a/block/block-copy.c b/block/block-copy.c | ||
70 | index XXXXXXX..XXXXXXX 100644 | ||
71 | --- a/block/block-copy.c | ||
72 | +++ b/block/block-copy.c | ||
73 | @@ -XXX,XX +XXX,XX @@ typedef struct BlockCopyCallState { | ||
74 | /* State */ | ||
75 | int ret; | ||
76 | bool finished; | ||
77 | - QemuCoSleepState *sleep_state; | ||
78 | + QemuCoSleep sleep; | ||
79 | bool cancelled; | ||
80 | |||
81 | /* OUT parameters */ | ||
82 | @@ -XXX,XX +XXX,XX @@ block_copy_dirty_clusters(BlockCopyCallState *call_state) | ||
83 | if (ns > 0) { | ||
84 | block_copy_task_end(task, -EAGAIN); | ||
85 | g_free(task); | ||
86 | - qemu_co_sleep_ns_wakeable(QEMU_CLOCK_REALTIME, ns, | ||
87 | - &call_state->sleep_state); | ||
88 | + qemu_co_sleep_ns_wakeable(&call_state->sleep, | ||
89 | + QEMU_CLOCK_REALTIME, ns); | ||
90 | continue; | ||
91 | } | ||
92 | } | ||
93 | @@ -XXX,XX +XXX,XX @@ out: | ||
94 | |||
95 | void block_copy_kick(BlockCopyCallState *call_state) | ||
96 | { | ||
97 | - qemu_co_sleep_wake(call_state->sleep_state); | ||
98 | + qemu_co_sleep_wake(&call_state->sleep); | ||
99 | } | ||
100 | |||
101 | /* | ||
102 | diff --git a/block/nbd.c b/block/nbd.c | ||
103 | index XXXXXXX..XXXXXXX 100644 | ||
104 | --- a/block/nbd.c | ||
105 | +++ b/block/nbd.c | ||
106 | @@ -XXX,XX +XXX,XX @@ typedef struct BDRVNBDState { | ||
107 | CoQueue free_sema; | ||
108 | Coroutine *connection_co; | ||
109 | Coroutine *teardown_co; | ||
110 | - QemuCoSleepState *connection_co_sleep_ns_state; | ||
111 | + QemuCoSleep reconnect_sleep; | ||
112 | bool drained; | ||
113 | bool wait_drained_end; | ||
114 | int in_flight; | ||
115 | @@ -XXX,XX +XXX,XX @@ static void coroutine_fn nbd_client_co_drain_begin(BlockDriverState *bs) | ||
116 | BDRVNBDState *s = (BDRVNBDState *)bs->opaque; | ||
117 | |||
118 | s->drained = true; | ||
119 | - qemu_co_sleep_wake(s->connection_co_sleep_ns_state); | ||
120 | + qemu_co_sleep_wake(&s->reconnect_sleep); | ||
121 | |||
122 | nbd_co_establish_connection_cancel(bs, false); | ||
123 | |||
124 | @@ -XXX,XX +XXX,XX @@ static void nbd_teardown_connection(BlockDriverState *bs) | ||
125 | |||
126 | s->state = NBD_CLIENT_QUIT; | ||
127 | if (s->connection_co) { | ||
128 | - qemu_co_sleep_wake(s->connection_co_sleep_ns_state); | ||
129 | + qemu_co_sleep_wake(&s->reconnect_sleep); | ||
130 | nbd_co_establish_connection_cancel(bs, true); | ||
131 | } | ||
132 | if (qemu_in_coroutine()) { | ||
133 | @@ -XXX,XX +XXX,XX @@ static coroutine_fn void nbd_co_reconnect_loop(BDRVNBDState *s) | ||
134 | } | ||
135 | bdrv_inc_in_flight(s->bs); | ||
136 | } else { | ||
137 | - qemu_co_sleep_ns_wakeable(QEMU_CLOCK_REALTIME, timeout, | ||
138 | - &s->connection_co_sleep_ns_state); | ||
139 | + qemu_co_sleep_ns_wakeable(&s->reconnect_sleep, | ||
140 | + QEMU_CLOCK_REALTIME, timeout); | ||
141 | if (s->drained) { | ||
142 | continue; | ||
143 | } | ||
144 | diff --git a/util/qemu-coroutine-sleep.c b/util/qemu-coroutine-sleep.c | ||
145 | index XXXXXXX..XXXXXXX 100644 | ||
146 | --- a/util/qemu-coroutine-sleep.c | ||
147 | +++ b/util/qemu-coroutine-sleep.c | ||
148 | @@ -XXX,XX +XXX,XX @@ | ||
149 | |||
150 | static const char *qemu_co_sleep_ns__scheduled = "qemu_co_sleep_ns"; | ||
151 | |||
152 | -struct QemuCoSleepState { | ||
153 | +void qemu_co_sleep_wake(QemuCoSleep *w) | ||
154 | +{ | ||
155 | Coroutine *co; | ||
156 | - QemuCoSleepState **user_state_pointer; | ||
157 | -}; | ||
158 | |||
159 | -void qemu_co_sleep_wake(QemuCoSleepState *sleep_state) | ||
160 | -{ | ||
161 | - if (sleep_state) { | ||
162 | + co = w->to_wake; | ||
163 | + w->to_wake = NULL; | ||
164 | + if (co) { | ||
165 | /* Write of schedule protected by barrier write in aio_co_schedule */ | ||
166 | - const char *scheduled = qatomic_cmpxchg(&sleep_state->co->scheduled, | ||
167 | + const char *scheduled = qatomic_cmpxchg(&co->scheduled, | ||
168 | qemu_co_sleep_ns__scheduled, NULL); | ||
169 | |||
170 | assert(scheduled == qemu_co_sleep_ns__scheduled); | ||
171 | - *sleep_state->user_state_pointer = NULL; | ||
172 | - aio_co_wake(sleep_state->co); | ||
173 | + aio_co_wake(co); | ||
174 | } | ||
175 | } | ||
176 | |||
177 | static void co_sleep_cb(void *opaque) | ||
178 | { | ||
179 | - QemuCoSleepState **sleep_state = opaque; | ||
180 | - qemu_co_sleep_wake(*sleep_state); | ||
181 | + QemuCoSleep *w = opaque; | ||
182 | + qemu_co_sleep_wake(w); | ||
183 | } | ||
184 | |||
185 | -void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
186 | - QemuCoSleepState **sleep_state) | ||
187 | +void coroutine_fn qemu_co_sleep_ns_wakeable(QemuCoSleep *w, | ||
188 | + QEMUClockType type, int64_t ns) | ||
189 | { | ||
190 | + Coroutine *co = qemu_coroutine_self(); | ||
191 | AioContext *ctx = qemu_get_current_aio_context(); | ||
192 | QEMUTimer ts; | ||
193 | - QemuCoSleepState state = { | ||
194 | - .co = qemu_coroutine_self(), | ||
195 | - .user_state_pointer = sleep_state, | ||
196 | - }; | ||
197 | |||
198 | - const char *scheduled = qatomic_cmpxchg(&state.co->scheduled, NULL, | ||
199 | - qemu_co_sleep_ns__scheduled); | ||
200 | + const char *scheduled = qatomic_cmpxchg(&co->scheduled, NULL, | ||
201 | + qemu_co_sleep_ns__scheduled); | ||
202 | if (scheduled) { | ||
203 | fprintf(stderr, | ||
204 | "%s: Co-routine was already scheduled in '%s'\n", | ||
205 | @@ -XXX,XX +XXX,XX @@ void coroutine_fn qemu_co_sleep_ns_wakeable(QEMUClockType type, int64_t ns, | ||
206 | abort(); | ||
207 | } | ||
208 | |||
209 | - aio_timer_init(ctx, &ts, type, SCALE_NS, co_sleep_cb, sleep_state); | ||
210 | - *sleep_state = &state; | ||
211 | + w->to_wake = co; | ||
212 | + aio_timer_init(ctx, &ts, type, SCALE_NS, co_sleep_cb, w), | ||
213 | timer_mod(&ts, qemu_clock_get_ns(type) + ns); | ||
214 | qemu_coroutine_yield(); | ||
215 | timer_del(&ts); | ||
216 | |||
217 | - /* qemu_co_sleep_wake clears *sleep_state before resuming this coroutine. */ | ||
218 | - assert(*sleep_state == NULL); | ||
219 | + /* w->to_wake is cleared before resuming this coroutine. */ | ||
220 | + assert(w->to_wake == NULL); | ||
221 | } | ||
222 | -- | ||
223 | 2.31.1 | ||
224 | diff view generated by jsdifflib |