[PATCH v2 0/2] qapi/throttle: Fix qmp_block_set_io_throttle blocked for too long

[PATCH v2 0/2] qapi/throttle: Fix qmp_block_set_io_throttle blocked for too long

[zoudongjie via]

From: Zhu Yangyang <zhuyangyang14@huawei.com>

First of all, thanks to Stefan Hajnoczi for his suggestions,
I made the following changes in this patch:

  1. Support 0 in BDRV_POLL_WHILE_TIMEOUT(), 0 means infinite.
  2. Use uint64_t timeout_ns instead of int64 timeout to name variables.
  3. Use timer_pending() to check for expiry instead of explicitly checking
     against the deadline for BDRV_POLL_WHILE_TIMEOUT().
  4. Add documentation for bdrv_drained_begin_timeout(), note that bdrv_drained_end()
     must be called when -ETIMEDOUT is returned.
  5. Add a timeout parameter to the qmp_block_set_io_throttle() instead of hardcoding
     the timeout, and the default value is 0, mean an infinite timeout.

v1 patch link:
https://lore.kernel.org/qemu-devel/20250308101618.721954-1-zoudongjie@huawei.com/

Zhu Yangyang (2):
  io/block: Refactoring the bdrv_drained_begin() function and implement
    a timeout mechanism.
  qapi/throttle: add timeout parameter for qmp_block_set_io_throttle()

 block/block-backend.c                       | 14 ++++-
 block/io.c                                  | 58 +++++++++++++++++----
 block/qapi-system.c                         | 10 +++-
 include/block/aio-wait.h                    | 49 +++++++++++++++++
 include/block/block-io.h                    | 22 +++++++-
 include/system/block-backend-global-state.h |  1 +
 qapi/block-core.json                        |  5 +-
 util/aio-wait.c                             |  5 ++
 8 files changed, 150 insertions(+), 14 deletions(-)

-- 
2.33.0

[PATCH v2 1/2] io/block: Refactoring the bdrv_drained_begin() function and implement a timeout mechanism.

[zoudongjie via]

From: Zhu Yangyang <zhuyangyang14@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 <zhuyangyang14@huawei.com>
---
 block/io.c               | 58 +++++++++++++++++++++++++++++++++-------
 include/block/aio-wait.h | 49 +++++++++++++++++++++++++++++++++
 include/block/block-io.h | 22 ++++++++++++++-
 util/aio-wait.c          |  5 ++++
 4 files changed, 123 insertions(+), 11 deletions(-)

diff --git a/block/io.c b/block/io.c
index 1ba8d1aeea..912b76c4a4 100644
--- a/block/io.c
+++ b/block/io.c
@@ -255,6 +255,8 @@ typedef struct {
     bool begin;
     bool poll;
     BdrvChild *parent;
+    uint64_t timeout_ns;
+    int ret;
 } BdrvCoDrainData;
 
 /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll() */
@@ -283,6 +285,10 @@ 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,
+                                         uint64_t timeout_ns);
 static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
                                   bool poll);
 static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent);
@@ -296,7 +302,9 @@ 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_ns);
         } else {
             assert(!data->poll);
             bdrv_do_drained_end(bs, data->parent);
@@ -310,10 +318,11 @@ 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,
+                                                       uint64_t timeout_ns)
 {
     BdrvCoDrainData data;
     Coroutine *self = qemu_coroutine_self();
@@ -329,6 +338,8 @@ static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs,
         .begin = begin,
         .parent = parent,
         .poll = poll,
+        .timeout_ns = timeout_ns,
+        .ret = 0
     };
 
     if (bs) {
@@ -342,16 +353,27 @@ 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, 0);
+}
+
+static int bdrv_do_drained_begin_timeout(BlockDriverState *bs,
+                                         BdrvChild *parent,
+                                         bool poll,
+                                         uint64_t timeout_ns)
 {
     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_ns);
     }
 
     GLOBAL_STATE_CODE();
@@ -375,8 +397,17 @@ static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
      * nodes.
      */
     if (poll) {
-        BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, parent));
+        return BDRV_POLL_WHILE_TIMEOUT(bs,
+                                       bdrv_drain_poll_top_level(bs, parent),
+                                       timeout_ns);
     }
+    return 0;
+}
+
+static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
+                                  bool poll)
+{
+    bdrv_do_drained_begin_timeout(bs, parent, poll, 0);
 }
 
 void bdrv_do_drained_begin_quiesce(BlockDriverState *bs, BdrvChild *parent)
@@ -391,6 +422,13 @@ bdrv_drained_begin(BlockDriverState *bs)
     bdrv_do_drained_begin(bs, NULL, true);
 }
 
+int coroutine_mixed_fn
+bdrv_drained_begin_timeout(BlockDriverState *bs, uint64_t timeout_ns)
+{
+    IO_OR_GS_CODE();
+    return bdrv_do_drained_begin_timeout(bs, NULL, true, timeout_ns);
+}
+
 /**
  * This function does not poll, nor must any of its recursively called
  * functions.
diff --git a/include/block/aio-wait.h b/include/block/aio-wait.h
index cf5e8bde1c..24c31df620 100644
--- a/include/block/aio-wait.h
+++ b/include/block/aio-wait.h
@@ -99,6 +99,53 @@ 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.
+ * @timeout_ns: maximum duration to wait, in nanoseconds, except the value
+ *              is unsigned, 0 means infinite.
+ */
+#define AIO_WAIT_WHILE_TIMEOUT(ctx, cond, timeout_ns) ({                 \
+    int ret_ = 0;                                                        \
+    uint64_t timeout_ = (timeout_ns);                                    \
+    AioWait *wait_ = &global_aio_wait;                                   \
+    AioContext *ctx_ = (ctx);                                            \
+    AioContext *current_ctx_ = NULL;                                     \
+    QEMUTimer timer_;                                                    \
+    /* 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_)) {                      \
+        current_ctx_ = ctx_;                                             \
+    } else {                                                             \
+        assert(qemu_get_current_aio_context() ==                         \
+               qemu_get_aio_context());                                  \
+        current_ctx_ = qemu_get_aio_context();                           \
+    }                                                                    \
+    if (timeout_ > 0) {                                                  \
+        timer_init_full(&timer_, &current_ctx_->tlg,                     \
+                        QEMU_CLOCK_REALTIME,                             \
+                        SCALE_NS, 0, aio_wait_timer_cb, NULL);           \
+        timer_mod_ns(&timer_,                                            \
+                     qemu_clock_get_ns(QEMU_CLOCK_REALTIME) +            \
+                     timeout_);                                          \
+    }                                                                    \
+    while ((cond)) {                                                     \
+        aio_poll(current_ctx_, true);                                    \
+        if (timeout_ > 0 && !timer_pending(&timer_)) {                   \
+            ret_ = -ETIMEDOUT;                                           \
+            break;                                                       \
+        }                                                                \
+    }                                                                    \
+    if (timeout_ > 0) {                                                  \
+        timer_del(&timer_);                                              \
+    }                                                                    \
+    qatomic_dec(&wait_->num_waiters);                                    \
+    ret_; })
+
 #define AIO_WAIT_WHILE(ctx, cond)                                  \
     AIO_WAIT_WHILE_INTERNAL(ctx, cond)
 
@@ -149,4 +196,6 @@ static inline bool in_aio_context_home_thread(AioContext *ctx)
     }
 }
 
+void aio_wait_timer_cb(void *opaque);
+
 #endif /* QEMU_AIO_WAIT_H */
diff --git a/include/block/block-io.h b/include/block/block-io.h
index b49e0537dd..844e9cf350 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_ns) ({   \
+    BlockDriverState *bs_ = (bs);                          \
+    AIO_WAIT_WHILE_TIMEOUT(bdrv_get_aio_context(bs_),      \
+                           cond, timeout_ns); })
+
 void bdrv_drain(BlockDriverState *bs);
 
 int co_wrapper_mixed_bdrv_rdlock
@@ -432,7 +437,22 @@ bdrv_drain_poll(BlockDriverState *bs, BdrvChild *ignore_parent,
 void bdrv_drained_begin(BlockDriverState *bs);
 
 /**
- * bdrv_do_drained_begin_quiesce:
+ * bdrv_drained_begin_timeout:
+ *
+ * Added timeout parameter for bdrv_drained_begin() to make a time limited.
+ *
+ * @timeout_ns: maximum duration to wait; 0 means infinite, equal to call
+ *              bdrv_drained_begin().
+ *
+ * Returns: 0 if succeeded; -ETIMEDOUT when a timeout occurs.
+ *
+ * Note: when the timeout fails, we've already begin aquiesced section, so we
+ * still need to call bdrv_drained_end() to end the quiescent section.
+ */
+int bdrv_drained_begin_timeout(BlockDriverState *bs, uint64_t timeout_ns);
+
+/**
+ * bdrv_do_drained_badegin_quiesce:
  *
  * Quiesces a BDS like bdrv_drained_begin(), but does not wait for already
  * running requests to complete.
diff --git a/util/aio-wait.c b/util/aio-wait.c
index b5336cf5fd..64c3714fb8 100644
--- a/util/aio-wait.c
+++ b/util/aio-wait.c
@@ -84,3 +84,8 @@ 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_cb(void *opaque)
+{
+    /* The point is to make AIO_WAIT_WHILE_TIMEOUT()'s aio_poll() return */
+}
-- 
2.33.0

[PATCH v2 2/2] qapi/throttle: add timeout parameter for qmp_block_set_io_throttle()

[zoudongjie via]

From: Zhu Yangyang <zhuyangyang14@huawei.com>

Calling qmp_block_set_io_throttle() will be blocked for a long time
when a network disk is configured and the network failure is just about
to occur.

Therefore, we add a timeout parameter for qmp_block_set_io_throttle to control
its execution duration.

The default value of timeout is 0, that is infinite wait, consistent with
previous behavior.

Signed-off-by: Zhu Yangyang <zhuyangyang14@huawei.com>
---
 block/block-backend.c                       | 14 +++++++++++++-
 block/qapi-system.c                         | 10 +++++++++-
 include/system/block-backend-global-state.h |  1 +
 qapi/block-core.json                        |  5 ++++-
 4 files changed, 27 insertions(+), 3 deletions(-)

diff --git a/block/block-backend.c b/block/block-backend.c
index a402db13f2..fc1554ea55 100644
--- a/block/block-backend.c
+++ b/block/block-backend.c
@@ -2679,20 +2679,32 @@ void blk_set_io_limits(BlockBackend *blk, ThrottleConfig *cfg)
 }
 
 void blk_io_limits_disable(BlockBackend *blk)
+{
+    blk_io_limits_disable_timeout(blk, 0);
+}
+
+int blk_io_limits_disable_timeout(BlockBackend *blk, uint64_t timeout_ns)
 {
     BlockDriverState *bs = blk_bs(blk);
     ThrottleGroupMember *tgm = &blk->public.throttle_group_member;
+    int ret = 0;
+
     assert(tgm->throttle_state);
     GLOBAL_STATE_CODE();
     if (bs) {
         bdrv_ref(bs);
-        bdrv_drained_begin(bs);
+        ret = bdrv_drained_begin_timeout(bs, timeout_ns);
+        if (ret < 0) {
+            goto out;
+        }
     }
     throttle_group_unregister_tgm(tgm);
+out:
     if (bs) {
         bdrv_drained_end(bs);
         bdrv_unref(bs);
     }
+    return ret;
 }
 
 /* should be called before blk_set_io_limits if a limit is set */
diff --git a/block/qapi-system.c b/block/qapi-system.c
index 54b7409b2b..3b7ba0959c 100644
--- a/block/qapi-system.c
+++ b/block/qapi-system.c
@@ -423,6 +423,7 @@ void qmp_block_set_io_throttle(BlockIOThrottle *arg, Error **errp)
     ThrottleConfig cfg;
     BlockDriverState *bs;
     BlockBackend *blk;
+    uint64_t timeout_ns;
 
     blk = qmp_get_blk(arg->device, arg->id, errp);
     if (!blk) {
@@ -444,6 +445,10 @@ void qmp_block_set_io_throttle(BlockIOThrottle *arg, Error **errp)
     cfg.buckets[THROTTLE_OPS_READ].avg  = arg->iops_rd;
     cfg.buckets[THROTTLE_OPS_WRITE].avg = arg->iops_wr;
 
+    timeout_ns = 0; /* 0 means infinite */
+    if (arg->has_timeout) {
+        timeout_ns = arg->timeout * NANOSECONDS_PER_SECOND;
+    }
     if (arg->has_bps_max) {
         cfg.buckets[THROTTLE_BPS_TOTAL].max = arg->bps_max;
     }
@@ -502,7 +507,10 @@ void qmp_block_set_io_throttle(BlockIOThrottle *arg, Error **errp)
         blk_set_io_limits(blk, &cfg);
     } else if (blk_get_public(blk)->throttle_group_member.throttle_state) {
         /* If all throttling settings are set to 0, disable I/O limits */
-        blk_io_limits_disable(blk);
+        if (blk_io_limits_disable_timeout(blk, timeout_ns) < 0) {
+            error_setg(errp, "Blk io limits disable timeout");
+            return;
+        }
     }
 }
 
diff --git a/include/system/block-backend-global-state.h b/include/system/block-backend-global-state.h
index 35b5e8380b..8a2a6cbda4 100644
--- a/include/system/block-backend-global-state.h
+++ b/include/system/block-backend-global-state.h
@@ -110,6 +110,7 @@ int blk_probe_geometry(BlockBackend *blk, HDGeometry *geo);
 
 void blk_set_io_limits(BlockBackend *blk, ThrottleConfig *cfg);
 void blk_io_limits_disable(BlockBackend *blk);
+int blk_io_limits_disable_timeout(BlockBackend *blk, uint64_t timeout_ns);
 void blk_io_limits_enable(BlockBackend *blk, const char *group);
 void blk_io_limits_update_group(BlockBackend *blk, const char *group);
 void blk_set_force_allow_inactivate(BlockBackend *blk);
diff --git a/qapi/block-core.json b/qapi/block-core.json
index b1937780e1..88ef593efd 100644
--- a/qapi/block-core.json
+++ b/qapi/block-core.json
@@ -2626,6 +2626,9 @@
 #
 # @group: throttle group name (Since 2.4)
 #
+# @timeout: In seconds. Timeout for block_set_io_throttle,
+#     0 means no limit. Defaults to 0. (Since 10.0)
+#
 # Features:
 #
 # @deprecated: Member @device is deprecated.  Use @id instead.
@@ -2642,7 +2645,7 @@
             '*bps_max_length': 'int', '*bps_rd_max_length': 'int',
             '*bps_wr_max_length': 'int', '*iops_max_length': 'int',
             '*iops_rd_max_length': 'int', '*iops_wr_max_length': 'int',
-            '*iops_size': 'int', '*group': 'str' } }
+            '*iops_size': 'int', '*group': 'str', '*timeout': 'uint32'} }
 
 ##
 # @ThrottleLimits:
-- 
2.33.0

Re: [PATCH v2 1/2] io/block: Refactoring the bdrv_drained_begin() function and implement a timeout mechanism.

[Stefan Hajnoczi]

[-- Attachment #1: Type: text/plain, Size: 12371 bytes --]

On Fri, Mar 21, 2025 at 03:09:16PM +0800, zoudongjie wrote:
> From: Zhu Yangyang <zhuyangyang14@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 <zhuyangyang14@huawei.com>
> ---
>  block/io.c               | 58 +++++++++++++++++++++++++++++++++-------
>  include/block/aio-wait.h | 49 +++++++++++++++++++++++++++++++++
>  include/block/block-io.h | 22 ++++++++++++++-
>  util/aio-wait.c          |  5 ++++
>  4 files changed, 123 insertions(+), 11 deletions(-)
> 
> diff --git a/block/io.c b/block/io.c
> index 1ba8d1aeea..912b76c4a4 100644
> --- a/block/io.c
> +++ b/block/io.c
> @@ -255,6 +255,8 @@ typedef struct {
>      bool begin;
>      bool poll;
>      BdrvChild *parent;
> +    uint64_t timeout_ns;
> +    int ret;
>  } BdrvCoDrainData;
>  
>  /* Returns true if BDRV_POLL_WHILE() should go into a blocking aio_poll() */
> @@ -283,6 +285,10 @@ 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,
> +                                         uint64_t timeout_ns);
>  static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
>                                    bool poll);
>  static void bdrv_do_drained_end(BlockDriverState *bs, BdrvChild *parent);
> @@ -296,7 +302,9 @@ 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_ns);
>          } else {
>              assert(!data->poll);
>              bdrv_do_drained_end(bs, data->parent);
> @@ -310,10 +318,11 @@ 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,
> +                                                       uint64_t timeout_ns)
>  {
>      BdrvCoDrainData data;
>      Coroutine *self = qemu_coroutine_self();
> @@ -329,6 +338,8 @@ static void coroutine_fn bdrv_co_yield_to_drain(BlockDriverState *bs,
>          .begin = begin,
>          .parent = parent,
>          .poll = poll,
> +        .timeout_ns = timeout_ns,
> +        .ret = 0
>      };
>  
>      if (bs) {
> @@ -342,16 +353,27 @@ 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, 0);
> +}
> +
> +static int bdrv_do_drained_begin_timeout(BlockDriverState *bs,
> +                                         BdrvChild *parent,
> +                                         bool poll,
> +                                         uint64_t timeout_ns)
>  {
>      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_ns);
>      }
>  
>      GLOBAL_STATE_CODE();
> @@ -375,8 +397,17 @@ static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
>       * nodes.
>       */
>      if (poll) {
> -        BDRV_POLL_WHILE(bs, bdrv_drain_poll_top_level(bs, parent));
> +        return BDRV_POLL_WHILE_TIMEOUT(bs,
> +                                       bdrv_drain_poll_top_level(bs, parent),
> +                                       timeout_ns);
>      }
> +    return 0;
> +}
> +
> +static void bdrv_do_drained_begin(BlockDriverState *bs, BdrvChild *parent,
> +                                  bool poll)
> +{
> +    bdrv_do_drained_begin_timeout(bs, parent, poll, 0);
>  }
>  
>  void bdrv_do_drained_begin_quiesce(BlockDriverState *bs, BdrvChild *parent)
> @@ -391,6 +422,13 @@ bdrv_drained_begin(BlockDriverState *bs)
>      bdrv_do_drained_begin(bs, NULL, true);
>  }
>  
> +int coroutine_mixed_fn
> +bdrv_drained_begin_timeout(BlockDriverState *bs, uint64_t timeout_ns)
> +{
> +    IO_OR_GS_CODE();
> +    return bdrv_do_drained_begin_timeout(bs, NULL, true, timeout_ns);
> +}
> +
>  /**
>   * This function does not poll, nor must any of its recursively called
>   * functions.
> diff --git a/include/block/aio-wait.h b/include/block/aio-wait.h
> index cf5e8bde1c..24c31df620 100644
> --- a/include/block/aio-wait.h
> +++ b/include/block/aio-wait.h
> @@ -99,6 +99,53 @@ 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.
> + * @timeout_ns: maximum duration to wait, in nanoseconds, except the value
> + *              is unsigned, 0 means infinite.
> + */
> +#define AIO_WAIT_WHILE_TIMEOUT(ctx, cond, timeout_ns) ({                 \
> +    int ret_ = 0;                                                        \
> +    uint64_t timeout_ = (timeout_ns);                                    \
> +    AioWait *wait_ = &global_aio_wait;                                   \
> +    AioContext *ctx_ = (ctx);                                            \
> +    AioContext *current_ctx_ = NULL;                                     \
> +    QEMUTimer timer_;                                                    \
> +    /* 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_)) {                      \
> +        current_ctx_ = ctx_;                                             \
> +    } else {                                                             \
> +        assert(qemu_get_current_aio_context() ==                         \
> +               qemu_get_aio_context());                                  \
> +        current_ctx_ = qemu_get_aio_context();                           \
> +    }                                                                    \
> +    if (timeout_ > 0) {                                                  \
> +        timer_init_full(&timer_, &current_ctx_->tlg,                     \
> +                        QEMU_CLOCK_REALTIME,                             \
> +                        SCALE_NS, 0, aio_wait_timer_cb, NULL);           \
> +        timer_mod_ns(&timer_,                                            \
> +                     qemu_clock_get_ns(QEMU_CLOCK_REALTIME) +            \
> +                     timeout_);                                          \
> +    }                                                                    \
> +    while ((cond)) {                                                     \
> +        aio_poll(current_ctx_, true);                                    \
> +        if (timeout_ > 0 && !timer_pending(&timer_)) {                   \
> +            ret_ = -ETIMEDOUT;                                           \
> +            break;                                                       \
> +        }                                                                \
> +    }                                                                    \
> +    if (timeout_ > 0) {                                                  \
> +        timer_del(&timer_);                                              \
> +    }                                                                    \
> +    qatomic_dec(&wait_->num_waiters);                                    \
> +    ret_; })
> +
>  #define AIO_WAIT_WHILE(ctx, cond)                                  \
>      AIO_WAIT_WHILE_INTERNAL(ctx, cond)

It would be nice to unify AIO_WAIT_WHILE_INTERNAL() and
AIO_WAIT_WHILE_TIMEOUT(). I don't see any caller that uses the waited_
return value from AIO_WAIT_WHILE_INTERNAL(), so I think it's possible to
replace AIO_WAIT_WHILE_INTERNAL() with AIO_WAIT_WHILE_TIMEOUT(.., ..., 0).

>  
> @@ -149,4 +196,6 @@ static inline bool in_aio_context_home_thread(AioContext *ctx)
>      }
>  }
>  
> +void aio_wait_timer_cb(void *opaque);
> +
>  #endif /* QEMU_AIO_WAIT_H */
> diff --git a/include/block/block-io.h b/include/block/block-io.h
> index b49e0537dd..844e9cf350 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_ns) ({   \
> +    BlockDriverState *bs_ = (bs);                          \
> +    AIO_WAIT_WHILE_TIMEOUT(bdrv_get_aio_context(bs_),      \
> +                           cond, timeout_ns); })
> +
>  void bdrv_drain(BlockDriverState *bs);
>  
>  int co_wrapper_mixed_bdrv_rdlock
> @@ -432,7 +437,22 @@ bdrv_drain_poll(BlockDriverState *bs, BdrvChild *ignore_parent,
>  void bdrv_drained_begin(BlockDriverState *bs);
>  
>  /**
> - * bdrv_do_drained_begin_quiesce:
> + * bdrv_drained_begin_timeout:
> + *
> + * Added timeout parameter for bdrv_drained_begin() to make a time limited.
> + *
> + * @timeout_ns: maximum duration to wait; 0 means infinite, equal to call
> + *              bdrv_drained_begin().
> + *
> + * Returns: 0 if succeeded; -ETIMEDOUT when a timeout occurs.
> + *
> + * Note: when the timeout fails, we've already begin aquiesced section, so we
> + * still need to call bdrv_drained_end() to end the quiescent section.
> + */
> +int bdrv_drained_begin_timeout(BlockDriverState *bs, uint64_t timeout_ns);
> +
> +/**
> + * bdrv_do_drained_badegin_quiesce:
>   *
>   * Quiesces a BDS like bdrv_drained_begin(), but does not wait for already
>   * running requests to complete.
> diff --git a/util/aio-wait.c b/util/aio-wait.c
> index b5336cf5fd..64c3714fb8 100644
> --- a/util/aio-wait.c
> +++ b/util/aio-wait.c
> @@ -84,3 +84,8 @@ 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_cb(void *opaque)
> +{
> +    /* The point is to make AIO_WAIT_WHILE_TIMEOUT()'s aio_poll() return */
> +}
> -- 
> 2.33.0
> 

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH v2 2/2] qapi/throttle: add timeout parameter for qmp_block_set_io_throttle()

[Stefan Hajnoczi]

[-- Attachment #1: Type: text/plain, Size: 912 bytes --]

On Fri, Mar 21, 2025 at 03:09:17PM +0800, zoudongjie wrote:
> From: Zhu Yangyang <zhuyangyang14@huawei.com>
> 
> Calling qmp_block_set_io_throttle() will be blocked for a long time
> when a network disk is configured and the network failure is just about
> to occur.
> 
> Therefore, we add a timeout parameter for qmp_block_set_io_throttle to control
> its execution duration.
> 
> The default value of timeout is 0, that is infinite wait, consistent with
> previous behavior.
> 
> Signed-off-by: Zhu Yangyang <zhuyangyang14@huawei.com>
> ---
>  block/block-backend.c                       | 14 +++++++++++++-
>  block/qapi-system.c                         | 10 +++++++++-
>  include/system/block-backend-global-state.h |  1 +
>  qapi/block-core.json                        |  5 ++++-
>  4 files changed, 27 insertions(+), 3 deletions(-)

Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 488 bytes --]

Re: [PATCH v2 1/2] io/block: Refactoring the bdrv_drained_begin() function and implement a timeout mechanism.

[zoudongjie via]

From: Zhu Yangyang <zhuyangyang14@huawei.com>

On Mon, 24 Mar 2025 14:09:13 -0400, Stefan Hajnoczi wrote:
> On Fri, Mar 21, 2025 at 03:09:16PM +0800, zoudongjie wrote:
> > From: Zhu Yangyang <zhuyangyang14@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.
> ...
> 
> > +/**
> > + * AIO_WAIT_WHILE_TIMEOUT:
> > + *
> > + * Refer to the implementation of AIO_WAIT_WHILE_INTERNAL,
> > + * the timeout parameter is added.
> > + * @timeout_ns: maximum duration to wait, in nanoseconds, except the value
> > + *              is unsigned, 0 means infinite.
> > + */
> > +#define AIO_WAIT_WHILE_TIMEOUT(ctx, cond, timeout_ns) ({                 \
> > +    int ret_ = 0;                                                        \
> > +    uint64_t timeout_ = (timeout_ns);                                    \
> > +    AioWait *wait_ = &global_aio_wait;                                   \
> > +    AioContext *ctx_ = (ctx);                                            \
> > +    AioContext *current_ctx_ = NULL;                                     \
> > +    QEMUTimer timer_;                                                    \
> > +    /* 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_)) {                      \
> > +        current_ctx_ = ctx_;                                             \
> > +    } else {                                                             \
> > +        assert(qemu_get_current_aio_context() ==                         \
> > +               qemu_get_aio_context());                                  \
> > +        current_ctx_ = qemu_get_aio_context();                           \
> > +    }                                                                    \
> > +    if (timeout_ > 0) {                                                  \
> > +        timer_init_full(&timer_, &current_ctx_->tlg,                     \
> > +                        QEMU_CLOCK_REALTIME,                             \
> > +                        SCALE_NS, 0, aio_wait_timer_cb, NULL);           \
> > +        timer_mod_ns(&timer_,                                            \
> > +                     qemu_clock_get_ns(QEMU_CLOCK_REALTIME) +            \
> > +                     timeout_);                                          \
> > +    }                                                                    \
> > +    while ((cond)) {                                                     \
> > +        aio_poll(current_ctx_, true);                                    \
> > +        if (timeout_ > 0 && !timer_pending(&timer_)) {                   \
> > +            ret_ = -ETIMEDOUT;                                           \
> > +            break;                                                       \
> > +        }                                                                \
> > +    }                                                                    \
> > +    if (timeout_ > 0) {                                                  \
> > +        timer_del(&timer_);                                              \
> > +    }                                                                    \
> > +    qatomic_dec(&wait_->num_waiters);                                    \
> > +    ret_; })
> > +
> >  #define AIO_WAIT_WHILE(ctx, cond)                                  \
> >      AIO_WAIT_WHILE_INTERNAL(ctx, cond)
>
> It would be nice to unify AIO_WAIT_WHILE_INTERNAL() and
> AIO_WAIT_WHILE_TIMEOUT(). I don't see any caller that uses the waited_
> return value from AIO_WAIT_WHILE_INTERNAL(), so I think it's possible to
> replace AIO_WAIT_WHILE_INTERNAL() with AIO_WAIT_WHILE_TIMEOUT(.., ..., 0).
>

Yes, there is no difference between AIO_WAIT_WHILE_INTERNAL() and
AIO_WAIT_WHILE_TIMEOUT() except the return value. It's right to keep the
code uniform. I will modify it in the next version.