Commit fde382d1 authored Apr 30, 2019 by Tomasz Zawadzki Committed by Jim Harris Apr 30, 2019

blobstore: release same cluster as claimed during initial insert



When new writes come from different threads, cluster allocations can
happen many times at once. The corresponding cluster number for the map
is determined via _spdk_bs_allocate_cluster() and kept in ctx->new_cluster.

The cluster itself is inserted into the map only on md_thread.
When there is conflict of two threads allocating same cluster,
message is returned to the losting thread to release the cluster.

Before this patch, on such failure the cluster to release
was calcualted from the page. This resulted in releasing the
cluster claim for thread that actually won it.

This patch makes it so that cluster allocated and save in ctx
is used instead.

Change-Id: Id10811b887f673f9b89e41e0637d4422f1d7270d
Signed-off-by: Tomasz Zawadzki <tomasz.zawadzki@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/452625


Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>

parent e600f096

lib/blob/blobstore.c

+1 −5

Original line number	Diff line number	Diff line
		@@ -1589,17 +1589,13 @@ _spdk_blob_insert_cluster_cpl(void *cb_arg, int bserrno)
		struct spdk_blob_copy_cluster_ctx *ctx = cb_arg;

		if (bserrno) {
		uint32_t cluster_number;

		if (bserrno == -EEXIST) {
		/* The metadata insert failed because another thread
		* allocated the cluster first. Free our cluster
		* but continue without error. */
		bserrno = 0;
		}

		cluster_number = _spdk_bs_page_to_cluster(ctx->blob->bs, ctx->page);
		_spdk_bs_release_cluster(ctx->blob->bs, cluster_number);
		_spdk_bs_release_cluster(ctx->blob->bs, ctx->new_cluster);
		}

		spdk_bs_sequence_finish(ctx->seq, bserrno);