Skip to contentSkip to conversation

block-copy: protect block-copy internal structures

PATCH Applied Emanuele Giuseppe Esposito 2 participants 8 messages Jun 24, 2021 – Jun 25, 2021 Block layer

lore
Share & tools
lore
#0 lore raw

v5 cover block-copy: protect block-copy internal structures

This serie of patches aims to reduce the usage of the AioContexlock in block-copy, by introducing smaller granularity locks thus on making the block layer thread safe.

This serie depends on my previous serie that brings thread safety to the smaller API used by block-copy, like ratelimit, progressmeter abd co-shared-resource.

What's missing for block-copy to be fully thread-safe is fixing the CoSleep API to allow cross-thread sleep and wakeup. Paolo is working on it.

Patch 1 provides a small refactoring, patch 2 introduces the .method field in BlockCopyState, to be used instead of .use_copy_range, .copy_size and .zeros.
Patch 3 provide a refactoring in preparation to
the lock added in patch 4 on BlockCopyTask, BlockCopyCallState and BlockCopyState. Patch 5 uses load_acquire/store_release to make sure BlockCopyCallState OUT fields are updated before finished is set to true.

Based-on: <[email protected]>

---
v5:
* Squash patch 3 (improve comments) with patch 5 (add CoMutex).
* Better comments in block-copy, drop IN/OUT/State categories
* Remove some load_acquire in patch 6, replace them with atomic reads

Emanuele Giuseppe Esposito (4):
block-copy: small refactor in block_copy_task_entry and
block_copy_common
block-copy: move progress_set_remaining in block_copy_task_end
block-copy: add CoMutex lock
block-copy: atomic .cancelled and .finished fields in
BlockCopyCallState

Paolo Bonzini (1):
block-copy: streamline choice of copy_range vs. read/write

 include/block/block-copy.h |   2 +
 block/block-copy.c         | 368 +++++++++++++++++++++++--------------
 2 files changed, 231 insertions(+), 139 deletions(-)
Signature · 1 line
2.31.1
#1 lore raw

PATCH v5 1/5 block-copy: small refactor in block_copy_task_entry and block_copy_common

Use a local variable instead of referencing BlockCopyState through a BlockCopyCallState or BlockCopyTask every time.
This is in preparation for next patches.

No functional change intended.

---
 block/block-copy.c | 14 ++++++++------
 1 file changed, 8 insertions(+), 6 deletions(-)
block/block-copy.c +8 -6
index 943e30b7e6..f0dbb4912b 100644
@@ -452,14 +452,15 @@ static void block_copy_handle_copy_range_result(BlockCopyState *s,
 static coroutine_fn int block_copy_task_entry(AioTask *task)
 {
     BlockCopyTask *t = container_of(task, BlockCopyTask, task);
+    BlockCopyState *s = t->s;
     bool error_is_read = false;
     bool copy_range = t->copy_range;
     int ret;
 
-    ret = block_copy_do_copy(t->s, t->offset, t->bytes, t->zeroes,
+    ret = block_copy_do_copy(s, t->offset, t->bytes, t->zeroes,
                              &copy_range, &error_is_read);
     if (t->copy_range) {
-        block_copy_handle_copy_range_result(t->s, copy_range);
+        block_copy_handle_copy_range_result(s, copy_range);
     }
     if (ret < 0) {
         if (!t->call_state->ret) {
@@ -467,9 +468,9 @@ static coroutine_fn int block_copy_task_entry(AioTask *task)
             t->call_state->error_is_read = error_is_read;
         }
     } else {
-        progress_work_done(t->s->progress, t->bytes);
+        progress_work_done(s->progress, t->bytes);
     }
-    co_put_to_shres(t->s->mem, t->bytes);
+    co_put_to_shres(s->mem, t->bytes);
     block_copy_task_end(t, ret);
 
     return ret;
@@ -714,14 +715,15 @@ void block_copy_kick(BlockCopyCallState *call_state)
 static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
 {
     int ret;
+    BlockCopyState *s = call_state->s;
 
-    QLIST_INSERT_HEAD(&call_state->s->calls, call_state, list);
+    QLIST_INSERT_HEAD(&s->calls, call_state, list);
 
     do {
         ret = block_copy_dirty_clusters(call_state);
 
         if (ret == 0 && !call_state->cancelled) {
-            ret = block_copy_wait_one(call_state->s, call_state->offset,
+            ret = block_copy_wait_one(s, call_state->offset,
                                       call_state->bytes);
         }
 
Signature · 1 line
2.31.1
#2 lore raw

PATCH v5 2/5 block-copy: streamline choice of copy_range vs. read/write

In reply to Emanuele Giuseppe Esposito

From: Paolo Bonzini <[email protected]>

Put the logic to determine the copy size in a separate function, so that there is a simple state machine for the possible methods of copying data from one BlockDriverState to the other.

Use .method instead of .copy_range as in-out argument, and include also .zeroes as an additional copy method.

While at it, store the common computation of block_copy_max_transfer into a new field of BlockCopyState, and make sure that we always obey max_transfer; that's more efficient even for the COPY_RANGE_READ_WRITE case.

---
 block/block-copy.c | 176 +++++++++++++++++++++++----------------------
 1 file changed, 90 insertions(+), 86 deletions(-)
block/block-copy.c +90 -86
index f0dbb4912b..bbcc53ff70 100644
@@ -28,6 +28,14 @@
 #define BLOCK_COPY_MAX_WORKERS 64
 #define BLOCK_COPY_SLICE_TIME 100000000ULL /* ns */
 
+typedef enum {
+    COPY_READ_WRITE_CLUSTER,
+    COPY_READ_WRITE,
+    COPY_WRITE_ZEROES,
+    COPY_RANGE_SMALL,
+    COPY_RANGE_FULL
+} BlockCopyMethod;
+
 static coroutine_fn int block_copy_task_entry(AioTask *task);
 
 typedef struct BlockCopyCallState {
@@ -64,8 +72,7 @@ typedef struct BlockCopyTask {
     BlockCopyCallState *call_state;
     int64_t offset;
     int64_t bytes;
-    bool zeroes;
-    bool copy_range;
+    BlockCopyMethod method;
     QLIST_ENTRY(BlockCopyTask) list;
     CoQueue wait_queue; /* coroutines blocked on this task */
 } BlockCopyTask;
@@ -86,8 +93,8 @@ typedef struct BlockCopyState {
     BdrvDirtyBitmap *copy_bitmap;
     int64_t in_flight_bytes;
     int64_t cluster_size;
-    bool use_copy_range;
-    int64_t copy_size;
+    BlockCopyMethod method;
+    int64_t max_transfer;
     uint64_t len;
     QLIST_HEAD(, BlockCopyTask) tasks; /* All tasks from all block-copy calls */
     QLIST_HEAD(, BlockCopyCallState) calls;
@@ -149,6 +156,24 @@ static bool coroutine_fn block_copy_wait_one(BlockCopyState *s, int64_t offset,
     return true;
 }
 
+static int64_t block_copy_chunk_size(BlockCopyState *s)
+{
+    switch (s->method) {
+    case COPY_READ_WRITE_CLUSTER:
+        return s->cluster_size;
+    case COPY_READ_WRITE:
+    case COPY_RANGE_SMALL:
+        return MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER),
+                   s->max_transfer);
+    case COPY_RANGE_FULL:
+        return MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_COPY_RANGE),
+                   s->max_transfer);
+    default:
+        /* Cannot have COPY_WRITE_ZEROES here.  */
+        abort();
+    }
+}
+
 /*
  * Search for the first dirty area in offset/bytes range and create task at
  * the beginning of it.
@@ -158,8 +183,9 @@ static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
                                              int64_t offset, int64_t bytes)
 {
     BlockCopyTask *task;
-    int64_t max_chunk = MIN_NON_ZERO(s->copy_size, call_state->max_chunk);
+    int64_t max_chunk;
 
+    max_chunk = MIN_NON_ZERO(block_copy_chunk_size(s), call_state->max_chunk);
     if (!bdrv_dirty_bitmap_next_dirty_area(s->copy_bitmap,
                                            offset, offset + bytes,
                                            max_chunk, &offset, &bytes))
@@ -183,7 +209,7 @@ static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
         .call_state = call_state,
         .offset = offset,
         .bytes = bytes,
-        .copy_range = s->use_copy_range,
+        .method = s->method,
     };
     qemu_co_queue_init(&task->wait_queue);
     QLIST_INSERT_HEAD(&s->tasks, task, list);
@@ -267,28 +293,28 @@ BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target,
         .len = bdrv_dirty_bitmap_size(copy_bitmap),
         .write_flags = write_flags,
         .mem = shres_create(BLOCK_COPY_MAX_MEM),
+        .max_transfer = QEMU_ALIGN_DOWN(
+                                    block_copy_max_transfer(source, target),
+                                    cluster_size),
     };
 
-    if (block_copy_max_transfer(source, target) < cluster_size) {
+    if (s->max_transfer < cluster_size) {
         /*
          * copy_range does not respect max_transfer. We don't want to bother
          * with requests smaller than block-copy cluster size, so fallback to
          * buffered copying (read and write respect max_transfer on their
          * behalf).
          */
-        s->use_copy_range = false;
-        s->copy_size = cluster_size;
+        s->method = COPY_READ_WRITE_CLUSTER;
     } else if (write_flags & BDRV_REQ_WRITE_COMPRESSED) {
         /* Compression supports only cluster-size writes and no copy-range. */
-        s->use_copy_range = false;
-        s->copy_size = cluster_size;
+        s->method = COPY_READ_WRITE_CLUSTER;
     } else {
         /*
-         * We enable copy-range, but keep small copy_size, until first
+         * If copy range enabled, start with COPY_RANGE_SMALL, until first
          * successful copy_range (look at block_copy_do_copy).
          */
-        s->use_copy_range = use_copy_range;
-        s->copy_size = MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER);
+        s->method = use_copy_range ? COPY_RANGE_SMALL : COPY_READ_WRITE;
     }
 
     ratelimit_init(&s->rate_limit);
@@ -343,17 +369,14 @@ static coroutine_fn int block_copy_task_run(AioTaskPool *pool,
  *
  * No sync here: nor bitmap neighter intersecting requests handling, only copy.
  *
- * @copy_range is an in-out argument: if *copy_range is false, copy_range is not
- * done. If *copy_range is true, copy_range is attempted. If the copy_range
- * attempt fails, the function falls back to the usual read+write and
- * *copy_range is set to false. *copy_range and zeroes must not be true
- * simultaneously.
- *
+ * @method is an in-out argument, so that copy_range can be either extended to
+ * a full-size buffer or disabled if the copy_range attempt fails.  The output
+ * value of @method should be used for subsequent tasks.
  * Returns 0 on success.
  */
 static int coroutine_fn block_copy_do_copy(BlockCopyState *s,
                                            int64_t offset, int64_t bytes,
-                                           bool zeroes, bool *copy_range,
+                                           BlockCopyMethod *method,
                                            bool *error_is_read)
 {
     int ret;
@@ -367,9 +390,9 @@ static int coroutine_fn block_copy_do_copy(BlockCopyState *s,
     assert(offset + bytes <= s->len ||
            offset + bytes == QEMU_ALIGN_UP(s->len, s->cluster_size));
     assert(nbytes < INT_MAX);
-    assert(!(*copy_range && zeroes));
 
-    if (zeroes) {
+    switch (*method) {
+    case COPY_WRITE_ZEROES:
         ret = bdrv_co_pwrite_zeroes(s->target, offset, nbytes, s->write_flags &
                                     ~BDRV_REQ_WRITE_COMPRESSED);
         if (ret < 0) {
@@ -377,76 +400,59 @@ static int coroutine_fn block_copy_do_copy(BlockCopyState *s,
             *error_is_read = false;
         }
         return ret;
-    }
 
-    if (*copy_range) {
+    case COPY_RANGE_SMALL:
+    case COPY_RANGE_FULL:
         ret = bdrv_co_copy_range(s->source, offset, s->target, offset, nbytes,
                                  0, s->write_flags);
-        if (ret < 0) {
-            trace_block_copy_copy_range_fail(s, offset, ret);
-            *copy_range = false;
-            /* Fallback to read+write with allocated buffer */
-        } else {
+        if (ret >= 0) {
+            /* Successful copy-range, increase chunk size.  */
+            *method = COPY_RANGE_FULL;
             return 0;
         }
-    }
 
-    /*
-     * In case of failed copy_range request above, we may proceed with buffered
-     * request larger than BLOCK_COPY_MAX_BUFFER. Still, further requests will
-     * be properly limited, so don't care too much. Moreover the most likely
-     * case (copy_range is unsupported for the configuration, so the very first
-     * copy_range request fails) is handled by setting large copy_size only
-     * after first successful copy_range.
-     */
+        trace_block_copy_copy_range_fail(s, offset, ret);
+        *method = COPY_READ_WRITE;
+        /* Fall through to read+write with allocated buffer */
 
-    bounce_buffer = qemu_blockalign(s->source->bs, nbytes);
+    case COPY_READ_WRITE_CLUSTER:
+    case COPY_READ_WRITE:
+        /*
+         * In case of failed copy_range request above, we may proceed with
+         * buffered request larger than BLOCK_COPY_MAX_BUFFER.
+         * Still, further requests will be properly limited, so don't care too
+         * much. Moreover the most likely case (copy_range is unsupported for
+         * the configuration, so the very first copy_range request fails)
+         * is handled by setting large copy_size only after first successful
+         * copy_range.
+         */
 
-    ret = bdrv_co_pread(s->source, offset, nbytes, bounce_buffer, 0);
-    if (ret < 0) {
-        trace_block_copy_read_fail(s, offset, ret);
-        *error_is_read = true;
-        goto out;
-    }
+        bounce_buffer = qemu_blockalign(s->source->bs, nbytes);
 
-    ret = bdrv_co_pwrite(s->target, offset, nbytes, bounce_buffer,
-                         s->write_flags);
-    if (ret < 0) {
-        trace_block_copy_write_fail(s, offset, ret);
-        *error_is_read = false;
-        goto out;
-    }
+        ret = bdrv_co_pread(s->source, offset, nbytes, bounce_buffer, 0);
+        if (ret < 0) {
+            trace_block_copy_read_fail(s, offset, ret);
+            *error_is_read = true;
+            goto out;
+        }
 
-out:
-    qemu_vfree(bounce_buffer);
+        ret = bdrv_co_pwrite(s->target, offset, nbytes, bounce_buffer,
+                             s->write_flags);
+        if (ret < 0) {
+            trace_block_copy_write_fail(s, offset, ret);
+            *error_is_read = false;
+            goto out;
+        }
 
-    return ret;
-}
+    out:
+        qemu_vfree(bounce_buffer);
+        break;
 
-static void block_copy_handle_copy_range_result(BlockCopyState *s,
-                                                bool is_success)
-{
-    if (!s->use_copy_range) {
-        /* already disabled */
-        return;
+    default:
+        abort();
     }
 
-    if (is_success) {
-        /*
-         * Successful copy-range. Now increase copy_size.  copy_range
-         * does not respect max_transfer (it's a TODO), so we factor
-         * that in here.
-         */
-        s->copy_size =
-                MIN(MAX(s->cluster_size, BLOCK_COPY_MAX_COPY_RANGE),
-                    QEMU_ALIGN_DOWN(block_copy_max_transfer(s->source,
-                                                            s->target),
-                                    s->cluster_size));
-    } else {
-        /* Copy-range failed, disable it. */
-        s->use_copy_range = false;
-        s->copy_size = MAX(s->cluster_size, BLOCK_COPY_MAX_BUFFER);
-    }
+    return ret;
 }
 
 static coroutine_fn int block_copy_task_entry(AioTask *task)
@@ -454,13 +460,12 @@ static coroutine_fn int block_copy_task_entry(AioTask *task)
     BlockCopyTask *t = container_of(task, BlockCopyTask, task);
     BlockCopyState *s = t->s;
     bool error_is_read = false;
-    bool copy_range = t->copy_range;
+    BlockCopyMethod method = t->method;
     int ret;
 
-    ret = block_copy_do_copy(s, t->offset, t->bytes, t->zeroes,
-                             &copy_range, &error_is_read);
-    if (t->copy_range) {
-        block_copy_handle_copy_range_result(s, copy_range);
+    ret = block_copy_do_copy(s, t->offset, t->bytes, &method, &error_is_read);
+    if (s->method == t->method) {
+        s->method = method;
     }
     if (ret < 0) {
         if (!t->call_state->ret) {
@@ -643,8 +648,7 @@ block_copy_dirty_clusters(BlockCopyCallState *call_state)
             continue;
         }
         if (ret & BDRV_BLOCK_ZERO) {
-            task->zeroes = true;
-            task->copy_range = false;
+            task->method = COPY_WRITE_ZEROES;
         }
 
         if (!call_state->ignore_ratelimit) {
Signature · 1 line
2.31.1
#3 lore raw

PATCH v5 3/5 block-copy: move progress_set_remaining in block_copy_task_end

In reply to Emanuele Giuseppe Esposito

Moving this function in task_end ensures to update the progress anyways, even if there is an error.

It also helps in next patch, allowing task_end to have only one critical section.

---
 block/block-copy.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)
block/block-copy.c +3 -3
index bbcc53ff70..d44c41549e 100644
@@ -248,6 +248,9 @@ static void coroutine_fn block_copy_task_end(BlockCopyTask *task, int ret)
         bdrv_set_dirty_bitmap(task->s->copy_bitmap, task->offset, task->bytes);
     }
     QLIST_REMOVE(task, list);
+    progress_set_remaining(task->s->progress,
+                           bdrv_get_dirty_count(task->s->copy_bitmap) +
+                           task->s->in_flight_bytes);
     qemu_co_queue_restart_all(&task->wait_queue);
 }
 
@@ -638,9 +641,6 @@ block_copy_dirty_clusters(BlockCopyCallState *call_state)
         }
         if (s->skip_unallocated && !(ret & BDRV_BLOCK_ALLOCATED)) {
             block_copy_task_end(task, 0);
-            progress_set_remaining(s->progress,
-                                   bdrv_get_dirty_count(s->copy_bitmap) +
-                                   s->in_flight_bytes);
             trace_block_copy_skip_range(s, task->offset, task->bytes);
             offset = task_end(task);
             bytes = end - offset;
Signature · 1 line
2.31.1
#4 lore raw

PATCH v5 4/5 block-copy: add CoMutex lock

In reply to Emanuele Giuseppe Esposito

Group various structures fields, to better understand what we need to protect with a lock and what doesn't need it.
Then, add a CoMutex to protect concurrent access of block-copy data structures. This mutex also protects .copy_bitmap, because its thread-safe API does not prevent it from assigning two tasks to the same bitmap region.

Exceptions to the lock:
- .sleep_state is handled in the series "coroutine: new sleep/wake API" and thus here left as TODO.

- .finished, .cancelled and reads to .ret and .error_is_read will be protected in the following patch, because are used also outside coroutines.

- .skip_unallocated is atomic. Including it under the mutex would increase the critical sections and make them also much more complex. We can have it as atomic since it is only written from outside and read by block-copy coroutines.

---
 block/block-copy.c | 155 +++++++++++++++++++++++++++++++++------------
 1 file changed, 116 insertions(+), 39 deletions(-)
block/block-copy.c +116 -39
index d44c41549e..52878ba57a 100644
@@ -39,7 +39,7 @@ typedef enum {
 static coroutine_fn int block_copy_task_entry(AioTask *task);
 
 typedef struct BlockCopyCallState {
-    /* IN parameters. Initialized in block_copy_async() and never changed. */
+    /* Fields initialized in block_copy_async() and never changed. */
     BlockCopyState *s;
     int64_t offset;
     int64_t bytes;
@@ -48,33 +48,60 @@ typedef struct BlockCopyCallState {
     bool ignore_ratelimit;
     BlockCopyAsyncCallbackFunc cb;
     void *cb_opaque;
-
     /* Coroutine where async block-copy is running */
     Coroutine *co;
 
-    /* To reference all call states from BlockCopyState */
-    QLIST_ENTRY(BlockCopyCallState) list;
-
-    /* State */
-    int ret;
+    /* Fields whose state changes throughout the execution */
     bool finished;
-    QemuCoSleep sleep;
+    QemuCoSleep sleep; /* TODO: protect API with a lock */
     bool cancelled;
+    /* To reference all call states from BlockCopyState */
+    QLIST_ENTRY(BlockCopyCallState) list;
 
-    /* OUT parameters */
+    /*
+     * Fields that report information about return values and erros.
+     * Protected by lock in BlockCopyState.
+     */
     bool error_is_read;
+    /*
+     * @ret is set concurrently by tasks under mutex. Only set once by first
+     * failed task (and untouched if no task failed).
+     * After finishing (call_state->finished is true), it is not modified
+     * anymore and may be safely read without mutex.
+     */
+    int ret;
 } BlockCopyCallState;
 
 typedef struct BlockCopyTask {
     AioTask task;
 
+    /*
+     * Fields initialized in block_copy_task_create()
+     * and never changed.
+     */
     BlockCopyState *s;
     BlockCopyCallState *call_state;
     int64_t offset;
-    int64_t bytes;
+    /*
+     * @method can also be set again in the while loop of
+     * block_copy_dirty_clusters(), but it is never accessed concurrently
+     * because the only other function that reads it is
+     * block_copy_task_entry() and it is invoked afterwards in the same
+     * iteration.
+     */
     BlockCopyMethod method;
-    QLIST_ENTRY(BlockCopyTask) list;
+
+    /*
+     * Fields whose state changes throughout the execution
+     * Protected by lock in BlockCopyState.
+     */
     CoQueue wait_queue; /* coroutines blocked on this task */
+    /*
+     * Only protect the case of parallel read while updating @bytes
+     * value in block_copy_task_shrink().
+     */
+    int64_t bytes;
+    QLIST_ENTRY(BlockCopyTask) list;
 } BlockCopyTask;
 
 static int64_t task_end(BlockCopyTask *task)
@@ -90,17 +117,25 @@ typedef struct BlockCopyState {
      */
     BdrvChild *source;
     BdrvChild *target;
-    BdrvDirtyBitmap *copy_bitmap;
-    int64_t in_flight_bytes;
+
+    /*
+     * Fields initialized in block_copy_state_new()
+     * and never changed.
+     */
     int64_t cluster_size;
-    BlockCopyMethod method;
     int64_t max_transfer;
     uint64_t len;
-    QLIST_HEAD(, BlockCopyTask) tasks; /* All tasks from all block-copy calls */
-    QLIST_HEAD(, BlockCopyCallState) calls;
-
     BdrvRequestFlags write_flags;
 
+    /*
+     * Fields whose state changes throughout the execution
+     * Protected by lock.
+     */
+    CoMutex lock;
+    int64_t in_flight_bytes;
+    BlockCopyMethod method;
+    QLIST_HEAD(, BlockCopyTask) tasks; /* All tasks from all block-copy calls */
+    QLIST_HEAD(, BlockCopyCallState) calls;
     /*
      * skip_unallocated:
      *
@@ -115,15 +150,15 @@ typedef struct BlockCopyState {
      * skip unallocated regions, clear them in the copy_bitmap, and invoke
      * block_copy_reset_unallocated() every time it does.
      */
-    bool skip_unallocated;
-
+    bool skip_unallocated; /* atomic */
+    /* State fields that use a thread-safe API */
+    BdrvDirtyBitmap *copy_bitmap;
     ProgressMeter *progress;
-
     SharedResource *mem;
-
     RateLimit rate_limit;
 } BlockCopyState;
 
+/* Called with lock held */
 static BlockCopyTask *find_conflicting_task(BlockCopyState *s,
                                             int64_t offset, int64_t bytes)
 {
@@ -141,6 +176,9 @@ static BlockCopyTask *find_conflicting_task(BlockCopyState *s,
 /*
  * If there are no intersecting tasks return false. Otherwise, wait for the
  * first found intersecting tasks to finish and return true.
+ *
+ * Called with lock held. May temporary release the lock.
+ * Return value of 0 proves that lock was NOT released.
  */
 static bool coroutine_fn block_copy_wait_one(BlockCopyState *s, int64_t offset,
                                              int64_t bytes)
@@ -151,11 +189,12 @@ static bool coroutine_fn block_copy_wait_one(BlockCopyState *s, int64_t offset,
         return false;
     }
 
-    qemu_co_queue_wait(&task->wait_queue, NULL);
+    qemu_co_queue_wait(&task->wait_queue, &s->lock);
 
     return true;
 }
 
+/* Called with lock held */
 static int64_t block_copy_chunk_size(BlockCopyState *s)
 {
     switch (s->method) {
@@ -178,13 +217,14 @@ static int64_t block_copy_chunk_size(BlockCopyState *s)
  * Search for the first dirty area in offset/bytes range and create task at
  * the beginning of it.
  */
-static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
-                                             BlockCopyCallState *call_state,
-                                             int64_t offset, int64_t bytes)
+static coroutine_fn BlockCopyTask *
+block_copy_task_create(BlockCopyState *s, BlockCopyCallState *call_state,
+                       int64_t offset, int64_t bytes)
 {
     BlockCopyTask *task;
     int64_t max_chunk;
 
+    QEMU_LOCK_GUARD(&s->lock);
     max_chunk = MIN_NON_ZERO(block_copy_chunk_size(s), call_state->max_chunk);
     if (!bdrv_dirty_bitmap_next_dirty_area(s->copy_bitmap,
                                            offset, offset + bytes,
@@ -227,6 +267,7 @@ static BlockCopyTask *block_copy_task_create(BlockCopyState *s,
 static void coroutine_fn block_copy_task_shrink(BlockCopyTask *task,
                                                 int64_t new_bytes)
 {
+    QEMU_LOCK_GUARD(&task->s->lock);
     if (new_bytes == task->bytes) {
         return;
     }
@@ -243,6 +284,7 @@ static void coroutine_fn block_copy_task_shrink(BlockCopyTask *task,
 
 static void coroutine_fn block_copy_task_end(BlockCopyTask *task, int ret)
 {
+    QEMU_LOCK_GUARD(&task->s->lock);
     task->s->in_flight_bytes -= task->bytes;
     if (ret < 0) {
         bdrv_set_dirty_bitmap(task->s->copy_bitmap, task->offset, task->bytes);
@@ -321,12 +363,14 @@ BlockCopyState *block_copy_state_new(BdrvChild *source, BdrvChild *target,
     }
 
     ratelimit_init(&s->rate_limit);
+    qemu_co_mutex_init(&s->lock);
     QLIST_INIT(&s->tasks);
     QLIST_INIT(&s->calls);
 
     return s;
 }
 
+/* Only set before running the job, no need for locking. */
 void block_copy_set_progress_meter(BlockCopyState *s, ProgressMeter *pm)
 {
     s->progress = pm;
@@ -467,16 +511,20 @@ static coroutine_fn int block_copy_task_entry(AioTask *task)
     int ret;
 
     ret = block_copy_do_copy(s, t->offset, t->bytes, &method, &error_is_read);
-    if (s->method == t->method) {
-        s->method = method;
-    }
-    if (ret < 0) {
-        if (!t->call_state->ret) {
-            t->call_state->ret = ret;
-            t->call_state->error_is_read = error_is_read;
+
+    WITH_QEMU_LOCK_GUARD(&s->lock) {
+        if (s->method == t->method) {
+            s->method = method;
+        }
+
+        if (ret < 0) {
+            if (!t->call_state->ret) {
+                t->call_state->ret = ret;
+                t->call_state->error_is_read = error_is_read;
+            }
+        } else {
+            progress_work_done(s->progress, t->bytes);
         }
-    } else {
-        progress_work_done(s->progress, t->bytes);
     }
     co_put_to_shres(s->mem, t->bytes);
     block_copy_task_end(t, ret);
@@ -491,7 +539,7 @@ static int block_copy_block_status(BlockCopyState *s, int64_t offset,
     BlockDriverState *base;
     int ret;
 
-    if (s->skip_unallocated) {
+    if (qatomic_read(&s->skip_unallocated)) {
         base = bdrv_backing_chain_next(s->source->bs);
     } else {
         base = NULL;
@@ -578,10 +626,12 @@ int64_t block_copy_reset_unallocated(BlockCopyState *s,
     bytes = clusters * s->cluster_size;
 
     if (!ret) {
+        qemu_co_mutex_lock(&s->lock);
         bdrv_reset_dirty_bitmap(s->copy_bitmap, offset, bytes);
         progress_set_remaining(s->progress,
                                bdrv_get_dirty_count(s->copy_bitmap) +
                                s->in_flight_bytes);
+        qemu_co_mutex_unlock(&s->lock);
     }
 
     *count = bytes;
@@ -639,7 +689,8 @@ block_copy_dirty_clusters(BlockCopyCallState *call_state)
         if (status_bytes < task->bytes) {
             block_copy_task_shrink(task, status_bytes);
         }
-        if (s->skip_unallocated && !(ret & BDRV_BLOCK_ALLOCATED)) {
+        if (qatomic_read(&s->skip_unallocated) &&
+            !(ret & BDRV_BLOCK_ALLOCATED)) {
             block_copy_task_end(task, 0);
             trace_block_copy_skip_range(s, task->offset, task->bytes);
             offset = task_end(task);
@@ -721,14 +772,38 @@ static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
     int ret;
     BlockCopyState *s = call_state->s;
 
+    qemu_co_mutex_lock(&s->lock);
     QLIST_INSERT_HEAD(&s->calls, call_state, list);
+    qemu_co_mutex_unlock(&s->lock);
 
     do {
         ret = block_copy_dirty_clusters(call_state);
 
         if (ret == 0 && !call_state->cancelled) {
-            ret = block_copy_wait_one(s, call_state->offset,
-                                      call_state->bytes);
+            WITH_QEMU_LOCK_GUARD(&s->lock) {
+                /*
+                 * Check that there is no task we still need to
+                 * wait to complete
+                 */
+                ret = block_copy_wait_one(s, call_state->offset,
+                                          call_state->bytes);
+                if (ret == 0) {
+                    /*
+                     * No pending tasks, but check again the bitmap in this
+                     * same critical section, since a task might have failed
+                     * between this and the critical section in
+                     * block_copy_dirty_clusters().
+                     *
+                     * block_copy_wait_one return value 0 also means that it
+                     * didn't relase the lock. So, we are still in the same
+                     * critical section, not interrupted by any concurrent
+                     * access to state.
+                     */
+                    ret = bdrv_dirty_bitmap_next_dirty(s->copy_bitmap,
+                                                       call_state->offset,
+                                                       call_state->bytes) >= 0;
+                }
+            }
         }
 
         /*
@@ -748,7 +823,9 @@ static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
         call_state->cb(call_state->cb_opaque);
     }
 
+    qemu_co_mutex_lock(&s->lock);
     QLIST_REMOVE(call_state, list);
+    qemu_co_mutex_unlock(&s->lock);
 
     return ret;
 }
@@ -851,7 +928,7 @@ BdrvDirtyBitmap *block_copy_dirty_bitmap(BlockCopyState *s)
 
 void block_copy_set_skip_unallocated(BlockCopyState *s, bool skip)
 {
-    s->skip_unallocated = skip;
+    qatomic_set(&s->skip_unallocated, skip);
 }
 
 void block_copy_set_speed(BlockCopyState *s, uint64_t speed)
Signature · 1 line
2.31.1
#5 lore raw

PATCH v5 5/5 block-copy: atomic .cancelled and .finished fields in BlockCopyCallState

In reply to Emanuele Giuseppe Esposito

By adding acquire/release pairs, we ensure that .ret and .error_is_read fields are written by block_copy_dirty_clusters before .finished is true, and that they are read by API user after .finished is true.

The atomic here are necessary because the fields are concurrently modified in coroutines, and read outside.

---
 include/block/block-copy.h |  2 ++
 block/block-copy.c         | 37 ++++++++++++++++++++++---------------
 2 files changed, 24 insertions(+), 15 deletions(-)
include/block/block-copy.h +2 -0
index 338f2ea7fd..5c8278895c 100644
@@ -18,6 +18,8 @@
 #include "block/block.h"
 #include "qemu/co-shared-resource.h"
 
+/* All APIs are thread-safe */
+
 typedef void (*BlockCopyAsyncCallbackFunc)(void *opaque);
 typedef struct BlockCopyState BlockCopyState;
 typedef struct BlockCopyCallState BlockCopyCallState;
block/block-copy.c +22 -15
index 52878ba57a..594d9f4aec 100644
@@ -52,9 +52,9 @@ typedef struct BlockCopyCallState {
     Coroutine *co;
 
     /* Fields whose state changes throughout the execution */
-    bool finished;
+    bool finished; /* atomic */
     QemuCoSleep sleep; /* TODO: protect API with a lock */
-    bool cancelled;
+    bool cancelled; /* atomic */
     /* To reference all call states from BlockCopyState */
     QLIST_ENTRY(BlockCopyCallState) list;
 
@@ -667,7 +667,8 @@ block_copy_dirty_clusters(BlockCopyCallState *call_state)
     assert(QEMU_IS_ALIGNED(offset, s->cluster_size));
     assert(QEMU_IS_ALIGNED(bytes, s->cluster_size));
 
-    while (bytes && aio_task_pool_status(aio) == 0 && !call_state->cancelled) {
+    while (bytes && aio_task_pool_status(aio) == 0 &&
+           !qatomic_read(&call_state->cancelled)) {
         BlockCopyTask *task;
         int64_t status_bytes;
 
@@ -779,7 +780,7 @@ static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
     do {
         ret = block_copy_dirty_clusters(call_state);
 
-        if (ret == 0 && !call_state->cancelled) {
+        if (ret == 0 && !qatomic_read(&call_state->cancelled)) {
             WITH_QEMU_LOCK_GUARD(&s->lock) {
                 /*
                  * Check that there is no task we still need to
@@ -815,9 +816,9 @@ static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
          * 2. We have waited for some intersecting block-copy request
          *    It may have failed and produced new dirty bits.
          */
-    } while (ret > 0 && !call_state->cancelled);
+    } while (ret > 0 && !qatomic_read(&call_state->cancelled));
 
-    call_state->finished = true;
+    qatomic_store_release(&call_state->finished, true);
 
     if (call_state->cb) {
         call_state->cb(call_state->cb_opaque);
@@ -880,44 +881,50 @@ void block_copy_call_free(BlockCopyCallState *call_state)
         return;
     }
 
-    assert(call_state->finished);
+    assert(qatomic_read(&call_state->finished));
     g_free(call_state);
 }
 
 bool block_copy_call_finished(BlockCopyCallState *call_state)
 {
-    return call_state->finished;
+    return qatomic_read(&call_state->finished);
 }
 
 bool block_copy_call_succeeded(BlockCopyCallState *call_state)
 {
-    return call_state->finished && !call_state->cancelled &&
-        call_state->ret == 0;
+    return qatomic_load_acquire(&call_state->finished) &&
+           !qatomic_read(&call_state->cancelled) &&
+           call_state->ret == 0;
 }
 
 bool block_copy_call_failed(BlockCopyCallState *call_state)
 {
-    return call_state->finished && !call_state->cancelled &&
-        call_state->ret < 0;
+    return qatomic_load_acquire(&call_state->finished) &&
+           !qatomic_read(&call_state->cancelled) &&
+           call_state->ret < 0;
 }
 
 bool block_copy_call_cancelled(BlockCopyCallState *call_state)
 {
-    return call_state->cancelled;
+    return qatomic_read(&call_state->cancelled);
 }
 
 int block_copy_call_status(BlockCopyCallState *call_state, bool *error_is_read)
 {
-    assert(call_state->finished);
+    assert(qatomic_load_acquire(&call_state->finished));
     if (error_is_read) {
         *error_is_read = call_state->error_is_read;
     }
     return call_state->ret;
 }
 
+/*
+ * Note that cancelling and finishing are racy.
+ * User can cancel a block-copy that is already finished.
+ */
 void block_copy_call_cancel(BlockCopyCallState *call_state)
 {
-    call_state->cancelled = true;
+    qatomic_set(&call_state->cancelled, true);
     block_copy_kick(call_state);
 }
 
Signature · 1 line
2.31.1
Reviewed-by on the v5 series
#6 lore raw
In reply to Emanuele Giuseppe Esposito
Quoted 38 lines · Emanuele Giuseppe Esposito

This serie of patches aims to reduce the usage of the AioContexlock in block-copy, by introducing smaller granularity locks thus on making the block layer thread safe.

This serie depends on my previous serie that brings thread safety to the smaller API used by block-copy, like ratelimit, progressmeter abd co-shared-resource.

What's missing for block-copy to be fully thread-safe is fixing the CoSleep API to allow cross-thread sleep and wakeup. Paolo is working on it.

Patch 1 provides a small refactoring, patch 2 introduces the .method field in BlockCopyState, to be used instead of .use_copy_range, .copy_size and .zeros.
Patch 3 provide a refactoring in preparation to
the lock added in patch 4 on BlockCopyTask, BlockCopyCallState and BlockCopyState. Patch 5 uses load_acquire/store_release to make sure BlockCopyCallState OUT fields are updated before finished is set to true.

Based-on: <[email protected]>

---
v5:
* Squash patch 3 (improve comments) with patch 5 (add CoMutex).
* Better comments in block-copy, drop IN/OUT/State categories
* Remove some load_acquire in patch 6, replace them with atomic reads

Emanuele Giuseppe Esposito (4):
block-copy: small refactor in block_copy_task_entry and
block_copy_common
block-copy: move progress_set_remaining in block_copy_task_end
block-copy: add CoMutex lock
block-copy: atomic .cancelled and .finished fields in
BlockCopyCallState

Quoted 7 lines

Paolo Bonzini (1):
block-copy: streamline choice of copy_range vs. read/write

  include/block/block-copy.h |   2 +
  block/block-copy.c         | 368 +++++++++++++++++++++++--------------
  2 files changed, 231 insertions(+), 139 deletions(-)
Signature · 2 lines
Best regards,
Vladimir
#7 lore raw
In reply to Emanuele Giuseppe Esposito
Quoted 12 lines · Emanuele Giuseppe Esposito

This serie of patches aims to reduce the usage of the AioContexlock in block-copy, by introducing smaller granularity locks thus on making the block layer thread safe.

This serie depends on my previous serie that brings thread safety to the smaller API used by block-copy, like ratelimit, progressmeter abd co-shared-resource.

What's missing for block-copy to be fully thread-safe is fixing the CoSleep API to allow cross-thread sleep and wakeup. Paolo is working on it.

Quoted 9 lines

Patch 1 provides a small refactoring, patch 2 introduces the .method field in BlockCopyState, to be used instead of .use_copy_range, .copy_size and .zeros.
Patch 3 provide a refactoring in preparation to
the lock added in patch 4 on BlockCopyTask, BlockCopyCallState and BlockCopyState. Patch 5 uses load_acquire/store_release to make sure BlockCopyCallState OUT fields are updated before finished is set to true.

Thanks, applied to the jobs branch

Signature · 2 lines
Best regards,
Vladimir