]> git.proxmox.com Git - mirror_qemu.git/blobdiff - block/commit.c
block/nvme: Add driver statistics for access alignment and hw errors
[mirror_qemu.git] / block / commit.c
index 3b46ca7f97071b76c2f04983d5c8ca5de5a2e68c..1e85c306cc415342c44103a5f74e6beb9158a770 100644 (file)
@@ -37,37 +37,20 @@ typedef struct CommitBlockJob {
     BlockBackend *top;
     BlockBackend *base;
     BlockDriverState *base_bs;
+    BlockDriverState *base_overlay;
     BlockdevOnError on_error;
     bool base_read_only;
+    bool chain_frozen;
     char *backing_file_str;
 } CommitBlockJob;
 
-static int coroutine_fn commit_populate(BlockBackend *bs, BlockBackend *base,
-                                        int64_t offset, uint64_t bytes,
-                                        void *buf)
-{
-    int ret = 0;
-    QEMUIOVector qiov = QEMU_IOVEC_INIT_BUF(qiov, buf, bytes);
-
-    assert(bytes < SIZE_MAX);
-
-    ret = blk_co_preadv(bs, offset, qiov.size, &qiov, 0);
-    if (ret < 0) {
-        return ret;
-    }
-
-    ret = blk_co_pwritev(base, offset, qiov.size, &qiov, 0);
-    if (ret < 0) {
-        return ret;
-    }
-
-    return 0;
-}
-
 static int commit_prepare(Job *job)
 {
     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
 
+    bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
+    s->chain_frozen = false;
+
     /* Remove base node parent that still uses BLK_PERM_WRITE/RESIZE before
      * the normal backing chain can be restored. */
     blk_unref(s->base);
@@ -84,6 +67,10 @@ static void commit_abort(Job *job)
     CommitBlockJob *s = container_of(job, CommitBlockJob, common.job);
     BlockDriverState *top_bs = blk_bs(s->top);
 
+    if (s->chain_frozen) {
+        bdrv_unfreeze_backing_chain(s->commit_top_bs, s->base_bs);
+    }
+
     /* Make sure commit_top_bs and top stay around until bdrv_replace_node() */
     bdrv_ref(top_bs);
     bdrv_ref(s->commit_top_bs);
@@ -103,9 +90,7 @@ static void commit_abort(Job *job)
      * XXX Can (or should) we somehow keep 'consistent read' blocked even
      * after the failed/cancelled commit job is gone? If we already wrote
      * something to base, the intermediate images aren't valid any more. */
-    bdrv_child_try_set_perm(s->commit_top_bs->backing, 0, BLK_PERM_ALL,
-                            &error_abort);
-    bdrv_replace_node(s->commit_top_bs, backing_bs(s->commit_top_bs),
+    bdrv_replace_node(s->commit_top_bs, s->commit_top_bs->backing->bs,
                       &error_abort);
 
     bdrv_unref(s->commit_top_bs);
@@ -135,7 +120,6 @@ static int coroutine_fn commit_run(Job *job, Error **errp)
     int ret = 0;
     int64_t n = 0; /* bytes */
     void *buf = NULL;
-    int bytes_written = 0;
     int64_t len, base_len;
 
     ret = len = blk_getlength(s->top);
@@ -150,7 +134,7 @@ static int coroutine_fn commit_run(Job *job, Error **errp)
     }
 
     if (base_len < len) {
-        ret = blk_truncate(s->base, len, PREALLOC_MODE_OFF, NULL);
+        ret = blk_truncate(s->base, len, false, PREALLOC_MODE_OFF, 0, NULL);
         if (ret) {
             goto out;
         }
@@ -160,6 +144,7 @@ static int coroutine_fn commit_run(Job *job, Error **errp)
 
     for (offset = 0; offset < len; offset += n) {
         bool copy;
+        bool error_in_source = true;
 
         /* Note that even when no rate limit is applied we need to yield
          * with no pending I/O here so that bdrv_drain_all() returns.
@@ -169,17 +154,25 @@ static int coroutine_fn commit_run(Job *job, Error **errp)
             break;
         }
         /* Copy if allocated above the base */
-        ret = bdrv_is_allocated_above(blk_bs(s->top), blk_bs(s->base),
+        ret = bdrv_is_allocated_above(blk_bs(s->top), s->base_overlay, true,
                                       offset, COMMIT_BUFFER_SIZE, &n);
         copy = (ret == 1);
         trace_commit_one_iteration(s, offset, n, ret);
         if (copy) {
-            ret = commit_populate(s->top, s->base, offset, n, buf);
-            bytes_written += n;
+            assert(n < SIZE_MAX);
+
+            ret = blk_co_pread(s->top, offset, n, buf, 0);
+            if (ret >= 0) {
+                ret = blk_co_pwrite(s->base, offset, n, buf, 0);
+                if (ret < 0) {
+                    error_in_source = false;
+                }
+            }
         }
         if (ret < 0) {
             BlockErrorAction action =
-                block_job_error_action(&s->common, false, s->on_error, -ret);
+                block_job_error_action(&s->common, s->on_error,
+                                       error_in_source, -ret);
             if (action == BLOCK_ERROR_ACTION_REPORT) {
                 goto out;
             } else {
@@ -211,7 +204,6 @@ static const BlockJobDriver commit_job_driver = {
         .job_type      = JOB_TYPE_COMMIT,
         .free          = block_job_free,
         .user_resume   = block_job_user_resume,
-        .drain         = block_job_drain,
         .run           = commit_run,
         .prepare       = commit_prepare,
         .abort         = commit_abort,
@@ -232,7 +224,7 @@ static void bdrv_commit_top_refresh_filename(BlockDriverState *bs)
 }
 
 static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
-                                       const BdrvChildRole *role,
+                                       BdrvChildRole role,
                                        BlockReopenQueue *reopen_queue,
                                        uint64_t perm, uint64_t shared,
                                        uint64_t *nperm, uint64_t *nshared)
@@ -246,9 +238,10 @@ static void bdrv_commit_top_child_perm(BlockDriverState *bs, BdrvChild *c,
 static BlockDriver bdrv_commit_top = {
     .format_name                = "commit_top",
     .bdrv_co_preadv             = bdrv_commit_top_preadv,
-    .bdrv_co_block_status       = bdrv_co_block_status_from_backing,
     .bdrv_refresh_filename      = bdrv_commit_top_refresh_filename,
     .bdrv_child_perm            = bdrv_commit_top_child_perm,
+
+    .is_filter                  = true,
 };
 
 void commit_start(const char *job_id, BlockDriverState *bs,
@@ -260,15 +253,35 @@ void commit_start(const char *job_id, BlockDriverState *bs,
     CommitBlockJob *s;
     BlockDriverState *iter;
     BlockDriverState *commit_top_bs = NULL;
+    BlockDriverState *filtered_base;
     Error *local_err = NULL;
+    int64_t base_size, top_size;
+    uint64_t base_perms, iter_shared_perms;
     int ret;
 
     assert(top != bs);
-    if (top == base) {
+    if (bdrv_skip_filters(top) == bdrv_skip_filters(base)) {
         error_setg(errp, "Invalid files for merge: top and base are the same");
         return;
     }
 
+    base_size = bdrv_getlength(base);
+    if (base_size < 0) {
+        error_setg_errno(errp, -base_size, "Could not inquire base image size");
+        return;
+    }
+
+    top_size = bdrv_getlength(top);
+    if (top_size < 0) {
+        error_setg_errno(errp, -top_size, "Could not inquire top image size");
+        return;
+    }
+
+    base_perms = BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE;
+    if (base_size < top_size) {
+        base_perms |= BLK_PERM_RESIZE;
+    }
+
     s = block_job_create(job_id, &commit_job_driver, NULL, bs, 0, BLK_PERM_ALL,
                          speed, creation_flags, NULL, NULL, errp);
     if (!s) {
@@ -293,51 +306,75 @@ void commit_start(const char *job_id, BlockDriverState *bs,
     if (!filter_node_name) {
         commit_top_bs->implicit = true;
     }
+
+    /* So that we can always drop this node */
+    commit_top_bs->never_freeze = true;
+
     commit_top_bs->total_sectors = top->total_sectors;
-    bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(top));
 
-    bdrv_set_backing_hd(commit_top_bs, top, &local_err);
-    if (local_err) {
-        bdrv_unref(commit_top_bs);
-        commit_top_bs = NULL;
-        error_propagate(errp, local_err);
-        goto fail;
-    }
-    bdrv_replace_node(top, commit_top_bs, &local_err);
+    bdrv_append(commit_top_bs, top, &local_err);
     if (local_err) {
-        bdrv_unref(commit_top_bs);
         commit_top_bs = NULL;
         error_propagate(errp, local_err);
         goto fail;
     }
 
     s->commit_top_bs = commit_top_bs;
-    bdrv_unref(commit_top_bs);
 
-    /* Block all nodes between top and base, because they will
-     * disappear from the chain after this operation. */
-    assert(bdrv_chain_contains(top, base));
-    for (iter = top; iter != base; iter = backing_bs(iter)) {
-        /* XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
-         * at s->base (if writes are blocked for a node, they are also blocked
-         * for its backing file). The other options would be a second filter
-         * driver above s->base. */
+    /*
+     * Block all nodes between top and base, because they will
+     * disappear from the chain after this operation.
+     * Note that this assumes that the user is fine with removing all
+     * nodes (including R/W filters) between top and base.  Assuring
+     * this is the responsibility of the interface (i.e. whoever calls
+     * commit_start()).
+     */
+    s->base_overlay = bdrv_find_overlay(top, base);
+    assert(s->base_overlay);
+
+    /*
+     * The topmost node with
+     * bdrv_skip_filters(filtered_base) == bdrv_skip_filters(base)
+     */
+    filtered_base = bdrv_cow_bs(s->base_overlay);
+    assert(bdrv_skip_filters(filtered_base) == bdrv_skip_filters(base));
+
+    /*
+     * XXX BLK_PERM_WRITE needs to be allowed so we don't block ourselves
+     * at s->base (if writes are blocked for a node, they are also blocked
+     * for its backing file). The other options would be a second filter
+     * driver above s->base.
+     */
+    iter_shared_perms = BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE;
+
+    for (iter = top; iter != base; iter = bdrv_filter_or_cow_bs(iter)) {
+        if (iter == filtered_base) {
+            /*
+             * From here on, all nodes are filters on the base.  This
+             * allows us to share BLK_PERM_CONSISTENT_READ.
+             */
+            iter_shared_perms |= BLK_PERM_CONSISTENT_READ;
+        }
+
         ret = block_job_add_bdrv(&s->common, "intermediate node", iter, 0,
-                                 BLK_PERM_WRITE_UNCHANGED | BLK_PERM_WRITE,
-                                 errp);
+                                 iter_shared_perms, errp);
         if (ret < 0) {
             goto fail;
         }
     }
 
+    if (bdrv_freeze_backing_chain(commit_top_bs, base, errp) < 0) {
+        goto fail;
+    }
+    s->chain_frozen = true;
+
     ret = block_job_add_bdrv(&s->common, "base", base, 0, BLK_PERM_ALL, errp);
     if (ret < 0) {
         goto fail;
     }
 
-    s->base = blk_new(BLK_PERM_CONSISTENT_READ
-                      | BLK_PERM_WRITE
-                      | BLK_PERM_RESIZE,
+    s->base = blk_new(s->common.job.aio_context,
+                      base_perms,
                       BLK_PERM_CONSISTENT_READ
                       | BLK_PERM_GRAPH_MOD
                       | BLK_PERM_WRITE_UNCHANGED);
@@ -345,14 +382,16 @@ void commit_start(const char *job_id, BlockDriverState *bs,
     if (ret < 0) {
         goto fail;
     }
+    blk_set_disable_request_queuing(s->base, true);
     s->base_bs = base;
 
     /* Required permissions are already taken with block_job_add_bdrv() */
-    s->top = blk_new(0, BLK_PERM_ALL);
+    s->top = blk_new(s->common.job.aio_context, 0, BLK_PERM_ALL);
     ret = blk_insert_bs(s->top, top, errp);
     if (ret < 0) {
         goto fail;
     }
+    blk_set_disable_request_queuing(s->top, true);
 
     s->backing_file_str = g_strdup(backing_file_str);
     s->on_error = on_error;
@@ -362,12 +401,18 @@ void commit_start(const char *job_id, BlockDriverState *bs,
     return;
 
 fail:
+    if (s->chain_frozen) {
+        bdrv_unfreeze_backing_chain(commit_top_bs, base);
+    }
     if (s->base) {
         blk_unref(s->base);
     }
     if (s->top) {
         blk_unref(s->top);
     }
+    if (s->base_read_only) {
+        bdrv_reopen_set_read_only(base, true, NULL);
+    }
     job_early_fail(&s->common.job);
     /* commit_top_bs has to be replaced after deleting the block job,
      * otherwise this would fail because of lack of permissions. */
@@ -386,6 +431,7 @@ int bdrv_commit(BlockDriverState *bs)
     BlockDriverState *backing_file_bs = NULL;
     BlockDriverState *commit_top_bs = NULL;
     BlockDriver *drv = bs->drv;
+    AioContext *ctx;
     int64_t offset, length, backing_length;
     int ro;
     int64_t n;
@@ -396,25 +442,31 @@ int bdrv_commit(BlockDriverState *bs)
     if (!drv)
         return -ENOMEDIUM;
 
-    if (!bs->backing) {
+    backing_file_bs = bdrv_cow_bs(bs);
+
+    if (!backing_file_bs) {
         return -ENOTSUP;
     }
 
     if (bdrv_op_is_blocked(bs, BLOCK_OP_TYPE_COMMIT_SOURCE, NULL) ||
-        bdrv_op_is_blocked(bs->backing->bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL)) {
+        bdrv_op_is_blocked(backing_file_bs, BLOCK_OP_TYPE_COMMIT_TARGET, NULL))
+    {
         return -EBUSY;
     }
 
-    ro = bs->backing->bs->read_only;
+    ro = backing_file_bs->read_only;
 
     if (ro) {
-        if (bdrv_reopen_set_read_only(bs->backing->bs, false, NULL)) {
+        if (bdrv_reopen_set_read_only(backing_file_bs, false, NULL)) {
             return -EACCES;
         }
     }
 
-    src = blk_new(BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
-    backing = blk_new(BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
+    ctx = bdrv_get_aio_context(bs);
+    /* WRITE_UNCHANGED is required for bdrv_make_empty() */
+    src = blk_new(ctx, BLK_PERM_CONSISTENT_READ | BLK_PERM_WRITE_UNCHANGED,
+                  BLK_PERM_ALL);
+    backing = blk_new(ctx, BLK_PERM_WRITE | BLK_PERM_RESIZE, BLK_PERM_ALL);
 
     ret = blk_insert_bs(src, bs, &local_err);
     if (ret < 0) {
@@ -423,15 +475,12 @@ int bdrv_commit(BlockDriverState *bs)
     }
 
     /* Insert commit_top block node above backing, so we can write to it */
-    backing_file_bs = backing_bs(bs);
-
     commit_top_bs = bdrv_new_open_driver(&bdrv_commit_top, NULL, BDRV_O_RDWR,
                                          &local_err);
     if (commit_top_bs == NULL) {
         error_report_err(local_err);
         goto ro_cleanup;
     }
-    bdrv_set_aio_context(commit_top_bs, bdrv_get_aio_context(backing_file_bs));
 
     bdrv_set_backing_hd(commit_top_bs, backing_file_bs, &error_abort);
     bdrv_set_backing_hd(bs, commit_top_bs, &error_abort);
@@ -458,7 +507,8 @@ int bdrv_commit(BlockDriverState *bs)
      * grow the backing file image if possible.  If not possible,
      * we must return an error */
     if (length > backing_length) {
-        ret = blk_truncate(backing, length, PREALLOC_MODE_OFF, &local_err);
+        ret = blk_truncate(backing, length, false, PREALLOC_MODE_OFF, 0,
+                           &local_err);
         if (ret < 0) {
             error_report_err(local_err);
             goto ro_cleanup;
@@ -491,14 +541,14 @@ int bdrv_commit(BlockDriverState *bs)
         }
     }
 
-    if (drv->bdrv_make_empty) {
-        ret = drv->bdrv_make_empty(bs);
-        if (ret < 0) {
-            goto ro_cleanup;
-        }
-        blk_flush(src);
+    ret = blk_make_empty(src, NULL);
+    /* Ignore -ENOTSUP */
+    if (ret < 0 && ret != -ENOTSUP) {
+        goto ro_cleanup;
     }
 
+    blk_flush(src);
+
     /*
      * Make sure all data we wrote to the backing device is actually
      * stable on disk.
@@ -510,7 +560,7 @@ ro_cleanup:
     qemu_vfree(buf);
 
     blk_unref(backing);
-    if (backing_file_bs) {
+    if (bdrv_cow_bs(bs) != backing_file_bs) {
         bdrv_set_backing_hd(bs, backing_file_bs, &error_abort);
     }
     bdrv_unref(commit_top_bs);
@@ -518,7 +568,7 @@ ro_cleanup:
 
     if (ro) {
         /* ignoring error return here */
-        bdrv_reopen_set_read_only(bs->backing->bs, true, NULL);
+        bdrv_reopen_set_read_only(backing_file_bs, true, NULL);
     }
 
     return ret;