Commit 23078637 authored Jul 20, 2015 by Yan, Zheng Committed by Ilya Dryomov Sep 08, 2015

ceph: fix queuing inode to mdsdir's snaprealm



During MDS failovers, MClientSnap message may cause kclient to move
some inodes from root directory's snaprealm to mdsdir's snaprealm
and queue snapshots for these inodes. For a FS has never created any
snapshot, both root directory's snaprealm and mdsdir's snaprealm
share the same snapshot contexts (both are ceph_empty_snapc). This
confuses ceph_put_wrbuffer_cap_refs(), make it unable to distinguish
snapshot buffers from head buffers.

The fix is do not use ceph_empty_snapc as snaprealm's cached context.

Signed-off-by: Yan, Zheng <zyan@redhat.com>

parent 68931622

fs/ceph/snap.c

+0 −7

Original line number	Diff line number	Diff line
		@@ -338,12 +338,6 @@ static int build_snap_context(struct ceph_snap_realm *realm)
		return 0;
		}

		if (num == 0 && realm->seq == ceph_empty_snapc->seq) {
		ceph_get_snap_context(ceph_empty_snapc);
		snapc = ceph_empty_snapc;
		goto done;
		}

		/* alloc new snap context */
		err = -ENOMEM;
		if (num > (SIZE_MAX - sizeof(*snapc)) / sizeof(u64))
		@@ -381,7 +375,6 @@ static int build_snap_context(struct ceph_snap_realm *realm)
		realm->ino, realm, snapc, snapc->seq,
		(unsigned int) snapc->num_snaps);

		done:
		ceph_put_snap_context(realm->cached_context);
		realm->cached_context = snapc;
		return 0;

Admin message