On Thu, 13 Mar, 2025 at 12:09:45 +0800, Stefan Hajnoczi wrote:
> On Sat, Mar 08, 2025 at 06:16:17PM +0800, zoudongjie wrote:
> > From: Zhu Yangyang <zhuyangyan...@huawei.com>
> >
> > The bdrv_drained_begin() function is a blocking function. In scenarios
> > where network storage
> > is used and network links fail, it may block for a long time.
> > Therefore, we add a timeout parameter to control the duration of the block.
> >
> > Since bdrv_drained_begin() has been widely adopted, both
> > bdrv_drained_begin()
> > and bdrv_drained_begin_timeout() will be retained.
> >
> > Signed-off-by: Zhu Yangyang <zhuyangyan...@huawei.com>
> > ---
> > block/io.c | 55 ++++++++++++++++++++++++++++++-------
> > include/block/aio-wait.h | 58 ++++++++++++++++++++++++++++++++++++++++
> > include/block/block-io.h | 7 +++++
> > util/aio-wait.c | 7 +++++
> > 4 files changed, 117 insertions(+), 10 deletions(-)
> >
> > diff --git a/block/io.c b/block/io.c
> > index d369b994df..03b8b2dca7 100644
> > --- a/block/io.c
> > +++ b/block/io.c
> > @@ -255,6 +255,8 @@ typedef struct {
> > bool begin;
> > bool poll;
> > BdrvChild *parent;
> > + int ret;
> > + int64_t timeout;
>
> Please put the units (milliseconds) into the variable name here and
> everywhere else in the patch to avoid confusion about units:
>
> int64_t timeout_ms;
Ok, I'm going to modify it in patch V2.
>
> > } BdrvCoDrainData;
> >
> > /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll()
> > */
> > @@ -283,6 +285,8 @@ static bool bdrv_drain_poll_top_level(BlockDriverState
> > *bs,
> > return bdrv_drain_poll(bs, ignore_parent, false);
> > }
> >
> > +static int bdrv_do_drained_begin_timeout(BlockDriverState *bs,
> > + BdrvChild *parent, bool poll, int64_t timeout);
> > static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
> > bool poll);
> > static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent);
> > @@ -296,7 +300,8 @@ static void bdrv_co_drain_bh_cb(void *opaque)
> > if (bs) {
> > bdrv_dec_in_flight(bs);
> > if (data->begin) {
> > - bdrv_do_drained_begin(bs, data->parent, data->poll);
> > + data->ret = bdrv_do_drained_begin_timeout(
> > + bs, data->parent, data->poll, data->timeout);
> > } else {
> > assert(!data->poll);
> > bdrv_do_drained_end(bs, data->parent);
> > @@ -310,10 +315,11 @@ static void bdrv_co_drain_bh_cb(void *opaque)
> > aio_co_wake(co);
> > }
> >
> > -static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs,
> > - bool begin,
> > - BdrvChild *parent,
> > - bool poll)
> > +static int coroutine_fn bdrv_co_yield_to_drain_timeout(BlockDriverState
> > *bs,
> > + bool begin,
> > + BdrvChild *parent,
> > + bool poll,
> > + int64_t timeout)
> > {
> > BdrvCoDrainData data;
> > Coroutine *self = qemu_coroutine_self();
> > @@ -329,6 +335,8 @@ static void coroutine_fn
> > bdrv_co_yield_to_drain(BlockDriverState *bs,
> > .begin = begin,
> > .parent = parent,
> > .poll = poll,
> > + .timeout = timeout,
> > + .ret = 0
> > };
> >
> > if (bs) {
> > @@ -342,16 +350,25 @@ static void coroutine_fn
> > bdrv_co_yield_to_drain(BlockDriverState *bs,
> > /* If we are resumed from some other event (such as an aio completion
> > or a
> > * timer callback), it is a bug in the caller that should be fixed. */
> > assert(data.done);
> > + return data.ret;
> > }
> >
> > -static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
> > - bool poll)
> > +static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs,
> > + bool begin,
> > + BdrvChild *parent,
> > + bool poll)
> > +{
> > + bdrv_co_yield_to_drain_timeout(bs, begin, parent, poll, -1);
>
> Is this safe on 32-bit platforms?
I'm sorry, can it be more specific here, I didn't get it.
>
> > +}
> > +
> > +static int bdrv_do_drained_begin_timeout(BlockDriverState *bs,
> > + BdrvChild *parent, bool poll, int64_t timeout_ms)
> > {
> > IO_OR_GS_CODE();
> >
> > if (qemu_in_coroutine()) {
> > - bdrv_co_yield_to_drain(bs, true, parent, poll);
> > - return;
> > + return bdrv_co_yield_to_drain_timeout(bs, true, parent, poll,
> > + timeout_ms);
> > }
> >
> > GLOBAL_STATE_CODE();
> > @@ -375,8 +392,20 @@ static void bdrv_do_drained_begin(BlockDriverState
> > *bs, BdrvChild *parent,
> > * nodes.
> > */
> > if (poll) {
> > - BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, parent));
> > + if (timeout_ms < 0) {
> > + BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, parent));
> > + } else {
> > + return BDRV_POLL_WHILE_TIMEOUT(
> > + bs, bdrv_drain_poll_top_level(bs, parent), timeout_ms);
> > + }
>
> Any reason to handle timeout_ms < 0 here instead of in
> BDRV_POLL_WHILE_TIMEOUT()? It would be more consistent to support -1 in
> BDRV_POLL_WHILE_TIMEOUT() so that you don't need to remember which
> functions/macros support timeout_ms=-1 and which dont.
Previously, BDRV_POLL_WHILE_TIMEOUT() was not done very well, aio_poll() exits
frequently because interval is used in the timer. but now I will support -1 in
BDRV_POLL_WHILE_TIMEOUT().
>
> > }
> > + return 0;
> > +}
> > +
> > +static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
> > + bool poll)
> > +{
> > + bdrv_do_drained_begin_timeout(bs, parent, poll, -1);
> > }
> >
> > void bdrv_do_drained_begin_quiesce(BlockDriverState *bs, BdrvChild *parent)
> > @@ -390,6 +419,12 @@ bdrv_drained_begin(BlockDriverState *bs)
> > IO_OR_GS_CODE();
> > bdrv_do_drained_begin(bs, NULL, true);
> > }
> > +int coroutine_mixed_fn
> > +bdrv_drained_begin_timeout(BlockDriverState *bs, int64_t timeout_ms)
> > +{
> > + IO_OR_GS_CODE();
> > + return bdrv_do_drained_begin_timeout(bs, NULL, true, timeout_ms);
> > +}
> >
> > /**
> > * This function does not poll, nor must any of its recursively called
> > diff --git a/include/block/aio-wait.h b/include/block/aio-wait.h
> > index cf5e8bde1c..efbcb9777a 100644
> > --- a/include/block/aio-wait.h
> > +++ b/include/block/aio-wait.h
> > @@ -28,6 +28,8 @@
> > #include "block/aio.h"
> > #include "qemu/main-loop.h"
> >
> > +#define AIO_WAIT_INTERVAL 10 /* ms */
> > +
> > /**
> > * AioWait:
> > *
> > @@ -56,6 +58,11 @@ typedef struct {
> > unsigned num_waiters;
> > } AioWait;
> >
> > +typedef struct {
> > + struct QEMUTimer *timer;
>
> struct is not necessary since QEMUTimer is a typedef:
>
> QEMUTimer *timer;
>
> Also, can this be a struct field instead of a pointer by using
> aio_timer_init_ms() instead of aio_timer_new()?
Ok, I'm going to modify it in patch V2.
>
> > + int64_t interval;
> > +} AioWaitTimer;
> > +
> > extern AioWait global_aio_wait;
> >
> > /**
> > @@ -99,6 +106,55 @@ extern AioWait global_aio_wait;
> > qatomic_dec(&wait_->num_waiters); \
> > waited_; })
> >
> > +/**
> > + * AIO_WAIT_WHILE_TIMEOUT:
> > + *
> > + * Refer to the implementation of AIO_WAIT_WHILE_INTERNAL,
> > + * the timeout parameter is added.
> > + */
> > +#define AIO_WAIT_WHILE_TIMEOUT(ctx, cond, timeout) ({ \
> > + int ret_ = 0; \
> > + AioWait *wait_ = &global_aio_wait; \
> > + AioContext *ctx_ = (ctx); \
> > + int64_t start_ = qemu_clock_get_ms(QEMU_CLOCK_REALTIME); \
> > + int64_t deadline_ = start_ + (timeout); \
> > + /* Ensure that the aio_poll exits periodically to check timeout. */ \
> > + AioWaitTimer *s_ = g_malloc0(sizeof(AioWaitTimer)); \
>
> Can this be allocated on the stack instead of the heap?
Yes, it's certainly better.
>
> > + s_->interval = AIO_WAIT_INTERVAL; \
> > + /* Increment wait_->num_waiters before evaluating cond. */ \
> > + qatomic_inc(&wait_->num_waiters); \
> > + /* Paired with smp_mb in aio_wait_kick(). */ \
> > + smp_mb__after_rmw(); \
> > + if (ctx_ && in_aio_context_home_thread(ctx_)) { \
> > + s_->timer = aio_timer_new(ctx_, QEMU_CLOCK_REALTIME, \
> > + SCALE_MS, aio_wait_timer_retry, s_); \
> > + aio_wait_timer_retry(s_); \
> > + while ((cond)) { \
> > + aio_poll(ctx_, true); \
> > + if (qemu_clock_get_ms(QEMU_CLOCK_REALTIME) > deadline_) { \
> > + ret_ = -ETIMEDOUT; \
> > + break; \
> > + } \
> > + } \
>
> What is the purpose of interval?
>
> I expected the timer's callback function to be an empty function that is
> called when the deadline expires. The while loop here would use
> timer_pending() to check for expiry instead of explicitly checking
> against the deadline.
This was really a good idea; it resolved some of my doubts and made everything
look better.
>
> > + } else { \
> > + s_->timer = aio_timer_new(qemu_get_aio_context(), \
> > + QEMU_CLOCK_REALTIME, SCALE_MS, aio_wait_timer_retry, s_); \
> > + aio_wait_timer_retry(s_); \
> > + while ((cond)) { \
> > + assert(qemu_get_current_aio_context() == \
> > + qemu_get_aio_context()); \
> > + aio_poll(qemu_get_aio_context(), true); \
> > + if (qemu_clock_get_ms(QEMU_CLOCK_REALTIME) > deadline_) { \
> > + ret_ = -ETIMEDOUT; \
> > + break; \
> > + } \
> > + } \
> > + } \
> > + qatomic_dec(&wait_->num_waiters); \
> > + timer_free(s_->timer); \
>
> This will need to be timer_del() when the QEMUTimer is moved onto the
> stack.
>
> > + g_free(s_); \
> > + ret_; })
> > +
> > #define AIO_WAIT_WHILE(ctx, cond) \
> > AIO_WAIT_WHILE_INTERNAL(ctx, cond)
> >
> > @@ -149,4 +205,6 @@ static inline bool
> > in_aio_context_home_thread(AioContext *ctx)
> > }
> > }
> >
> > +void aio_wait_timer_retry(void *opaque);
> > +
> > #endif /* QEMU_AIO_WAIT_H */
> > diff --git a/include/block/block-io.h b/include/block/block-io.h
> > index b49e0537dd..84f92d2b09 100644
> > --- a/include/block/block-io.h
> > +++ b/include/block/block-io.h
> > @@ -354,6 +354,11 @@ bdrv_co_copy_range(BdrvChild *src, int64_t src_offset,
> > AIO_WAIT_WHILE(bdrv_get_aio_context(bs_), \
> > cond); })
> >
> > +#define BDRV_POLL_WHILE_TIMEOUT(bs, cond, timeout) ({ \
> > + BlockDriverState *bs_ = (bs); \
> > + AIO_WAIT_WHILE_TIMEOUT(bdrv_get_aio_context(bs_), \
> > + cond, timeout); })
> > +
> > void bdrv_drain(BlockDriverState *bs);
> >
> > int co_wrapper_mixed_bdrv_rdlock
> > @@ -431,6 +436,8 @@ bdrv_drain_poll(BlockDriverState *bs, BdrvChild
> > *ignore_parent,
> > */
> > void bdrv_drained_begin(BlockDriverState *bs);
> >
> > +int bdrv_drained_begin_timeout(BlockDriverState *bs, int64_t timeout_ms);
> > +
> > /**
> > * bdrv_do_drained_begin_quiesce:
> > *
> > diff --git a/util/aio-wait.c b/util/aio-wait.c
> > index b5336cf5fd..9aed165529 100644
> > --- a/util/aio-wait.c
> > +++ b/util/aio-wait.c
> > @@ -84,3 +84,10 @@ void aio_wait_bh_oneshot(AioContext *ctx, QEMUBHFunc
> > *cb, void *opaque)
> > aio_bh_schedule_oneshot(ctx, aio_wait_bh, &data);
> > AIO_WAIT_WHILE_UNLOCKED(NULL, !data.done);
> > }
> > +
> > +void aio_wait_timer_retry(void *opaque)
> > +{
> > + AioWaitTimer *s = opaque;
> > +
> > + timer_mod(s->timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) +
> > s->interval);
> > +}
> > --
> > 2.33.0
> >
