dm snapshot: fix header corruption race on invalidation
authorMikulas Patocka <mpatocka@redhat.com>
Fri, 4 Sep 2009 19:40:39 +0000 (20:40 +0100)
committerAlasdair G Kergon <agk@redhat.com>
Fri, 4 Sep 2009 19:40:39 +0000 (20:40 +0100)
If a persistent snapshot fills up, a race can corrupt the on-disk header
which causes a crash on any future attempt to activate the snapshot
(typically while booting).  This patch fixes the race.

When the snapshot overflows, __invalidate_snapshot is called, which calls
snapshot store method drop_snapshot. It goes to persistent_drop_snapshot that
calls write_header. write_header constructs the new header in the "area"
location.

Concurrently, an existing kcopyd job may finish, call copy_callback
and commit_exception method, that goes to persistent_commit_exception.
persistent_commit_exception doesn't do locking, relying on the fact that
callbacks are single-threaded, but it can race with snapshot invalidation and
overwrite the header that is just being written while the snapshot is being
invalidated.

The result of this race is a corrupted header being written that can
lead to a crash on further reactivation (if chunk_size is zero in the
corrupted header).

The fix is to use separate memory areas for each.

See the bug: https://bugzilla.redhat.com/show_bug.cgi?id=461506

Cc: stable@kernel.org
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
drivers/md/dm-snap-persistent.c

index 2a3d626..5d1a975 100644 (file)
@@ -105,6 +105,13 @@ struct pstore {
         */
        void *zero_area;
 
+       /*
+        * An area used for header. The header can be written
+        * concurrently with metadata (when invalidating the snapshot),
+        * so it needs a separate buffer.
+        */
+       void *header_area;
+
        /*
         * Used to keep track of which metadata area the data in
         * 'chunk' refers to.
@@ -148,16 +155,27 @@ static int alloc_area(struct pstore *ps)
         */
        ps->area = vmalloc(len);
        if (!ps->area)
-               return r;
+               goto err_area;
 
        ps->zero_area = vmalloc(len);
-       if (!ps->zero_area) {
-               vfree(ps->area);
-               return r;
-       }
+       if (!ps->zero_area)
+               goto err_zero_area;
        memset(ps->zero_area, 0, len);
 
+       ps->header_area = vmalloc(len);
+       if (!ps->header_area)
+               goto err_header_area;
+
        return 0;
+
+err_header_area:
+       vfree(ps->zero_area);
+
+err_zero_area:
+       vfree(ps->area);
+
+err_area:
+       return r;
 }
 
 static void free_area(struct pstore *ps)
@@ -169,6 +187,10 @@ static void free_area(struct pstore *ps)
        if (ps->zero_area)
                vfree(ps->zero_area);
        ps->zero_area = NULL;
+
+       if (ps->header_area)
+               vfree(ps->header_area);
+       ps->header_area = NULL;
 }
 
 struct mdata_req {
@@ -285,11 +307,11 @@ static int read_header(struct pstore *ps, int *new_snapshot)
        if (r)
                return r;
 
-       r = chunk_io(ps, ps->area, 0, READ, 1);
+       r = chunk_io(ps, ps->header_area, 0, READ, 1);
        if (r)
                goto bad;
 
-       dh = (struct disk_header *) ps->area;
+       dh = ps->header_area;
 
        if (le32_to_cpu(dh->magic) == 0) {
                *new_snapshot = 1;
@@ -339,15 +361,15 @@ static int write_header(struct pstore *ps)
 {
        struct disk_header *dh;
 
-       memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
+       memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
 
-       dh = (struct disk_header *) ps->area;
+       dh = ps->header_area;
        dh->magic = cpu_to_le32(SNAP_MAGIC);
        dh->valid = cpu_to_le32(ps->valid);
        dh->version = cpu_to_le32(ps->version);
        dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
 
-       return chunk_io(ps, ps->area, 0, WRITE, 1);
+       return chunk_io(ps, ps->header_area, 0, WRITE, 1);
 }
 
 /*
@@ -667,6 +689,8 @@ static int persistent_ctr(struct dm_exception_store *store,
        ps->valid = 1;
        ps->version = SNAPSHOT_DISK_VERSION;
        ps->area = NULL;
+       ps->zero_area = NULL;
+       ps->header_area = NULL;
        ps->next_free = 2;      /* skipping the header and first area */
        ps->current_committed = 0;