block/block-copy: add ratelimit to block-copy

We are going to directly use one async block-copy operation for backup
job, so we need rate limiter.

We want to maintain current backup behavior: only background copying is
limited and copy-before-write operations only participate in limit
calculation. Therefore we need one rate limiter for block-copy state
and boolean flag for block-copy call state for actual limitation.

Note, that we can't just calculate each chunk in limiter after
successful copying: it will not save us from starting a lot of async
sub-requests which will exceed limit too much. Instead let's use the
following scheme on sub-request creation:
1. If at the moment limit is not exceeded, create the request and
account it immediately.
2. If at the moment limit is already exceeded, drop create sub-request
and handle limit instead (by sleep).
With this approach we'll never exceed the limit more than by one
sub-request (which pretty much matches current backup behavior).

Note also, that if there is in-flight block-copy async call,
block_copy_kick() should be used after set-speed to apply new setup
faster. For that block_copy_kick() published in this patch.

Signed-off-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
Reviewed-by: Max Reitz <mreitz@redhat.com>
Message-Id: <20210116214705.822267-7-vsementsov@virtuozzo.com>
Signed-off-by: Max Reitz <mreitz@redhat.com>

3 files changed, 47 insertions, 3 deletions

diff --git a/block/backup-top.c b/block/backup-top.c
index 789acf6..779956d 100644
--- a/block/backup-top.c
+++ b/block/backup-top.c
@@ -61,7 +61,7 @@ static coroutine_fn int backup_top_cbw(BlockDriverState *bs, uint64_t offset,
     off = QEMU_ALIGN_DOWN(offset, s->cluster_size);
     end = QEMU_ALIGN_UP(offset + bytes, s->cluster_size);
 
-    return block_copy(s->bcs, off, end - off, NULL);
+    return block_copy(s->bcs, off, end - off, true, NULL);
 }
 
 static int coroutine_fn backup_top_co_pdiscard(BlockDriverState *bs,
diff --git a/block/backup.c b/block/backup.c
index 4b07e91..09ff5a9 100644
--- a/block/backup.c
+++ b/block/backup.c
@@ -72,7 +72,7 @@ static int coroutine_fn backup_do_cow(BackupBlockJob *job,
 
     trace_backup_do_cow_enter(job, start, offset, bytes);
 
-    ret = block_copy(job->bcs, start, end - start, error_is_read);
+    ret = block_copy(job->bcs, start, end - start, true, error_is_read);
 
     trace_backup_do_cow_return(job, offset, bytes, ret);
 
diff --git a/block/block-copy.c b/block/block-copy.c
index 6bf1735..fa27450 100644
--- a/block/block-copy.c
+++ b/block/block-copy.c
@@ -26,6 +26,7 @@
 #define BLOCK_COPY_MAX_BUFFER (1 * MiB)
 #define BLOCK_COPY_MAX_MEM (128 * MiB)
 #define BLOCK_COPY_MAX_WORKERS 64
+#define BLOCK_COPY_SLICE_TIME 100000000ULL /* ns */
 
 static coroutine_fn int block_copy_task_entry(AioTask *task);
 
@@ -36,6 +37,7 @@ typedef struct BlockCopyCallState {
     int64_t bytes;
     int max_workers;
     int64_t max_chunk;
+    bool ignore_ratelimit;
     BlockCopyAsyncCallbackFunc cb;
     void *cb_opaque;
 
@@ -48,6 +50,7 @@ typedef struct BlockCopyCallState {
     /* State */
     int ret;
     bool finished;
+    QemuCoSleepState *sleep_state;
 
     /* OUT parameters */
     bool error_is_read;
@@ -111,6 +114,9 @@ typedef struct BlockCopyState {
     void *progress_opaque;
 
     SharedResource *mem;
+
+    uint64_t speed;
+    RateLimit rate_limit;
 } BlockCopyState;
 
 static BlockCopyTask *find_conflicting_task(BlockCopyState *s,
@@ -623,6 +629,21 @@ block_copy_dirty_clusters(BlockCopyCallState *call_state)
         }
         task->zeroes = ret & BDRV_BLOCK_ZERO;
 
+        if (s->speed) {
+            if (!call_state->ignore_ratelimit) {
+                uint64_t ns = ratelimit_calculate_delay(&s->rate_limit, 0);
+                if (ns > 0) {
+                    block_copy_task_end(task, -EAGAIN);
+                    g_free(task);
+                    qemu_co_sleep_ns_wakeable(QEMU_CLOCK_REALTIME, ns,
+                                              &call_state->sleep_state);
+                    continue;
+                }
+            }
+
+            ratelimit_calculate_delay(&s->rate_limit, task->bytes);
+        }
+
         trace_block_copy_process(s, task->offset);
 
         co_get_from_shres(s->mem, task->bytes);
@@ -661,6 +682,13 @@ out:
     return ret < 0 ? ret : found_dirty;
 }
 
+void block_copy_kick(BlockCopyCallState *call_state)
+{
+    if (call_state->sleep_state) {
+        qemu_co_sleep_wake(call_state->sleep_state);
+    }
+}
+
 /*
  * block_copy_common
  *
@@ -707,12 +735,13 @@ static int coroutine_fn block_copy_common(BlockCopyCallState *call_state)
 }
 
 int coroutine_fn block_copy(BlockCopyState *s, int64_t start, int64_t bytes,
-                            bool *error_is_read)
+                            bool ignore_ratelimit, bool *error_is_read)
 {
     BlockCopyCallState call_state = {
         .s = s,
         .offset = start,
         .bytes = bytes,
+        .ignore_ratelimit = ignore_ratelimit,
         .max_workers = BLOCK_COPY_MAX_WORKERS,
     };
 
@@ -798,3 +827,18 @@ void block_copy_set_skip_unallocated(BlockCopyState *s, bool skip)
 {
     s->skip_unallocated = skip;
 }
+
+void block_copy_set_speed(BlockCopyState *s, uint64_t speed)
+{
+    s->speed = speed;
+    if (speed > 0) {
+        ratelimit_set_speed(&s->rate_limit, speed, BLOCK_COPY_SLICE_TIME);
+    }
+
+    /*
+     * Note: it's good to kick all call states from here, but it should be done
+     * only from a coroutine, to not crash if s->calls list changed while
+     * entering one call. So for now, the only user of this function kicks its
+     * only one call_state by hand.
+     */
+}