]> Git Repo - linux.git/commitdiff
rbd: get snapshot context after exclusive lock is ensured to be held
authorIlya Dryomov <[email protected]>
Mon, 5 Jun 2023 14:33:35 +0000 (16:33 +0200)
committerIlya Dryomov <[email protected]>
Tue, 6 Jun 2023 07:54:27 +0000 (09:54 +0200)
Move capturing the snapshot context into the image request state
machine, after exclusive lock is ensured to be held for the duration of
dealing with the image request.  This is needed to ensure correctness
of fast-diff states (OBJECT_EXISTS vs OBJECT_EXISTS_CLEAN) and object
deltas computed based off of them.  Otherwise the object map that is
forked for the snapshot isn't guaranteed to accurately reflect the
contents of the snapshot when the snapshot is taken under I/O.  This
breaks differential backup and snapshot-based mirroring use cases with
fast-diff enabled: since some object deltas may be incomplete, the
destination image may get corrupted.

Cc: [email protected]
Link: https://tracker.ceph.com/issues/61472
Signed-off-by: Ilya Dryomov <[email protected]>
Reviewed-by: Dongsheng Yang <[email protected]>
drivers/block/rbd.c

index 6c847db6ee2cd68dd1d7b885a0b34ecb5f2cab61..632751ddb2870846926391f77986ca1e44a0a5bc 100644 (file)
@@ -1336,6 +1336,8 @@ static bool rbd_obj_is_tail(struct rbd_obj_request *obj_req)
  */
 static void rbd_obj_set_copyup_enabled(struct rbd_obj_request *obj_req)
 {
+       rbd_assert(obj_req->img_request->snapc);
+
        if (obj_req->img_request->op_type == OBJ_OP_DISCARD) {
                dout("%s %p objno %llu discard\n", __func__, obj_req,
                     obj_req->ex.oe_objno);
@@ -1456,6 +1458,7 @@ __rbd_obj_add_osd_request(struct rbd_obj_request *obj_req,
 static struct ceph_osd_request *
 rbd_obj_add_osd_request(struct rbd_obj_request *obj_req, int num_ops)
 {
+       rbd_assert(obj_req->img_request->snapc);
        return __rbd_obj_add_osd_request(obj_req, obj_req->img_request->snapc,
                                         num_ops);
 }
@@ -1592,15 +1595,18 @@ static void rbd_img_request_init(struct rbd_img_request *img_request,
        mutex_init(&img_request->state_mutex);
 }
 
+/*
+ * Only snap_id is captured here, for reads.  For writes, snapshot
+ * context is captured in rbd_img_object_requests() after exclusive
+ * lock is ensured to be held.
+ */
 static void rbd_img_capture_header(struct rbd_img_request *img_req)
 {
        struct rbd_device *rbd_dev = img_req->rbd_dev;
 
        lockdep_assert_held(&rbd_dev->header_rwsem);
 
-       if (rbd_img_is_write(img_req))
-               img_req->snapc = ceph_get_snap_context(rbd_dev->header.snapc);
-       else
+       if (!rbd_img_is_write(img_req))
                img_req->snap_id = rbd_dev->spec->snap_id;
 
        if (rbd_dev_parent_get(rbd_dev))
@@ -3482,9 +3488,19 @@ static int rbd_img_exclusive_lock(struct rbd_img_request *img_req)
 
 static void rbd_img_object_requests(struct rbd_img_request *img_req)
 {
+       struct rbd_device *rbd_dev = img_req->rbd_dev;
        struct rbd_obj_request *obj_req;
 
        rbd_assert(!img_req->pending.result && !img_req->pending.num_pending);
+       rbd_assert(!need_exclusive_lock(img_req) ||
+                  __rbd_is_lock_owner(rbd_dev));
+
+       if (rbd_img_is_write(img_req)) {
+               rbd_assert(!img_req->snapc);
+               down_read(&rbd_dev->header_rwsem);
+               img_req->snapc = ceph_get_snap_context(rbd_dev->header.snapc);
+               up_read(&rbd_dev->header_rwsem);
+       }
 
        for_each_obj_request(img_req, obj_req) {
                int result = 0;
@@ -3502,7 +3518,6 @@ static void rbd_img_object_requests(struct rbd_img_request *img_req)
 
 static bool rbd_img_advance(struct rbd_img_request *img_req, int *result)
 {
-       struct rbd_device *rbd_dev = img_req->rbd_dev;
        int ret;
 
 again:
@@ -3523,9 +3538,6 @@ again:
                if (*result)
                        return true;
 
-               rbd_assert(!need_exclusive_lock(img_req) ||
-                          __rbd_is_lock_owner(rbd_dev));
-
                rbd_img_object_requests(img_req);
                if (!img_req->pending.num_pending) {
                        *result = img_req->pending.result;
@@ -3987,6 +3999,10 @@ static int rbd_post_acquire_action(struct rbd_device *rbd_dev)
 {
        int ret;
 
+       ret = rbd_dev_refresh(rbd_dev);
+       if (ret)
+               return ret;
+
        if (rbd_dev->header.features & RBD_FEATURE_OBJECT_MAP) {
                ret = rbd_object_map_open(rbd_dev);
                if (ret)
This page took 0.082384 seconds and 4 git commands to generate.