qcow2: Fix order of refcount updates in qcow2_snapshot_goto

The refcount updates must be moved so that in the worst case we can get
cluster leaks, but refcounts may never be too low.

Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Reviewed-by: Stefan Hajnoczi <stefanha@linux.vnet.ibm.com>

2 files changed, 50 insertions, 18 deletions

diff --git a/block/qcow2-refcount.c b/block/qcow2-refcount.c
index 9605367..2db2ede 100644
--- a/block/qcow2-refcount.c
+++ b/block/qcow2-refcount.c
@@ -700,6 +700,10 @@ int qcow2_update_snapshot_refcount(BlockDriverState *bs,
     l2_table = NULL;
     l1_table = NULL;
     l1_size2 = l1_size * sizeof(uint64_t);
+
+    /* WARNING: qcow2_snapshot_goto relies on this function not using the
+     * l1_table_offset when it is the current s->l1_table_offset! Be careful
+     * when changing this! */
     if (l1_table_offset != s->l1_table_offset) {
         if (l1_size2 != 0) {
             l1_table = g_malloc0(align_offset(l1_size2, 512));
@@ -819,7 +823,8 @@ fail:
     qcow2_cache_set_writethrough(bs, s->refcount_block_cache,
         old_refcount_writethrough);
 
-    if (l1_modified) {
+    /* Update L1 only if it isn't deleted anyway (addend = -1) */
+    if (addend >= 0 && l1_modified) {
         for(i = 0; i < l1_size; i++)
             cpu_to_be64s(&l1_table[i]);
         if (bdrv_pwrite_sync(bs->file, l1_table_offset, l1_table,
diff --git a/block/qcow2-snapshot.c b/block/qcow2-snapshot.c
index 1ee22eb..9fb3ff0 100644
--- a/block/qcow2-snapshot.c
+++ b/block/qcow2-snapshot.c
@@ -387,6 +387,7 @@ int qcow2_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
     int i, snapshot_index;
     int cur_l1_bytes, sn_l1_bytes;
     int ret;
+    uint64_t *sn_l1_table = NULL;
 
     /* Search the snapshot */
     snapshot_index = find_snapshot_by_id_or_name(bs, snapshot_id);
@@ -395,14 +396,6 @@ int qcow2_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
     }
     sn = &s->snapshots[snapshot_index];
 
-    /* Decrease refcount of clusters of current L1 table.
-     * FIXME This is too early! */
-    ret = qcow2_update_snapshot_refcount(bs, s->l1_table_offset,
-                                         s->l1_size, -1);
-    if (ret < 0) {
-        goto fail;
-    }
-
     /*
      * Make sure that the current L1 table is big enough to contain the whole
      * L1 table of the snapshot. If the snapshot L1 table is smaller, the
@@ -416,33 +409,66 @@ int qcow2_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
     cur_l1_bytes = s->l1_size * sizeof(uint64_t);
     sn_l1_bytes = sn->l1_size * sizeof(uint64_t);
 
-    if (cur_l1_bytes > sn_l1_bytes) {
-        memset(s->l1_table + sn->l1_size, 0, cur_l1_bytes - sn_l1_bytes);
-    }
-
     /*
      * Copy the snapshot L1 table to the current L1 table.
      *
      * Before overwriting the old current L1 table on disk, make sure to
      * increase all refcounts for the clusters referenced by the new one.
+     * Decrease the refcount referenced by the old one only when the L1
+     * table is overwritten.
      */
-    ret = bdrv_pread(bs->file, sn->l1_table_offset, s->l1_table, sn_l1_bytes);
+    sn_l1_table = g_malloc0(cur_l1_bytes);
+
+    ret = bdrv_pread(bs->file, sn->l1_table_offset, sn_l1_table, sn_l1_bytes);
+    if (ret < 0) {
+        goto fail;
+    }
+
+    ret = qcow2_update_snapshot_refcount(bs, sn->l1_table_offset,
+                                         sn->l1_size, 1);
     if (ret < 0) {
         goto fail;
     }
 
-    ret = bdrv_pwrite_sync(bs->file, s->l1_table_offset, s->l1_table,
+    ret = bdrv_pwrite_sync(bs->file, s->l1_table_offset, sn_l1_table,
                            cur_l1_bytes);
     if (ret < 0) {
         goto fail;
     }
 
+    /*
+     * Decrease refcount of clusters of current L1 table.
+     *
+     * At this point, the in-memory s->l1_table points to the old L1 table,
+     * whereas on disk we already have the new one.
+     *
+     * qcow2_update_snapshot_refcount special cases the current L1 table to use
+     * the in-memory data instead of really using the offset to load a new one,
+     * which is why this works.
+     */
+    ret = qcow2_update_snapshot_refcount(bs, s->l1_table_offset,
+                                         s->l1_size, -1);
+
+    /*
+     * Now update the in-memory L1 table to be in sync with the on-disk one. We
+     * need to do this even if updating refcounts failed.
+     */
     for(i = 0;i < s->l1_size; i++) {
-        be64_to_cpus(&s->l1_table[i]);
+        s->l1_table[i] = be64_to_cpu(sn_l1_table[i]);
     }
 
-    /* FIXME This is too late! */
-    ret = qcow2_update_snapshot_refcount(bs, s->l1_table_offset, s->l1_size, 1);
+    if (ret < 0) {
+        goto fail;
+    }
+
+    g_free(sn_l1_table);
+    sn_l1_table = NULL;
+
+    /*
+     * Update QCOW_OFLAG_COPIED in the active L1 table (it may have changed
+     * when we decreased the refcount of the old snapshot.
+     */
+    ret = qcow2_update_snapshot_refcount(bs, s->l1_table_offset, s->l1_size, 0);
     if (ret < 0) {
         goto fail;
     }
@@ -456,6 +482,7 @@ int qcow2_snapshot_goto(BlockDriverState *bs, const char *snapshot_id)
     return 0;
 
 fail:
+    g_free(sn_l1_table);
     return ret;
 }