pnfs: don't check sequence on new stateids in layoutget
authorChristoph Hellwig <hch@lst.de>
Thu, 21 Aug 2014 16:09:20 +0000 (11:09 -0500)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 10 Sep 2014 19:47:01 +0000 (12:47 -0700)
When layoutget returns an entirely new layout stateid it should not
check the generation counter as the new stateid will start with a new
counter entirely unrelated to old one.

The current behavior causes constant layoutget failures against a block
server which allocates a new stateid after an recall that removed all
outstanding layouts.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/pnfs.c

index 0e0364ca877ddc5d5b63d2d59c1e4fd618a2646f..ff7fabe3352914f8e0f8d9d66a2028500abdb2c0 100644 (file)
@@ -682,17 +682,6 @@ static bool pnfs_seqid_is_newer(u32 s1, u32 s2)
        return (s32)(s1 - s2) > 0;
 }
 
-static void
-pnfs_verify_layout_stateid(struct pnfs_layout_hdr *lo,
-               const nfs4_stateid *new,
-               struct list_head *free_me_list)
-{
-       if (nfs4_stateid_match_other(&lo->plh_stateid, new))
-               return;
-       /* Layout is new! Kill existing layout segments */
-       pnfs_mark_matching_lsegs_invalid(lo, free_me_list, NULL);
-}
-
 /* update lo->plh_stateid with new if is more recent */
 void
 pnfs_set_layout_stateid(struct pnfs_layout_hdr *lo, const nfs4_stateid *new,
@@ -1367,16 +1356,29 @@ pnfs_layout_process(struct nfs4_layoutget *lgp)
                goto out_forget_reply;
        }
 
-       if (pnfs_layoutgets_blocked(lo, 1) ||
-           pnfs_layout_stateid_blocked(lo, &res->stateid)) {
+       if (pnfs_layoutgets_blocked(lo, 1)) {
                dprintk("%s forget reply due to state\n", __func__);
                goto out_forget_reply;
        }
 
-       /* Check that the new stateid matches the old stateid */
-       pnfs_verify_layout_stateid(lo, &res->stateid, &free_me);
-       /* Done processing layoutget. Set the layout stateid */
-       pnfs_set_layout_stateid(lo, &res->stateid, false);
+       if (nfs4_stateid_match_other(&lo->plh_stateid, &res->stateid)) {
+               /* existing state ID, make sure the sequence number matches. */
+               if (pnfs_layout_stateid_blocked(lo, &res->stateid)) {
+                       dprintk("%s forget reply due to sequence\n", __func__);
+                       goto out_forget_reply;
+               }
+               pnfs_set_layout_stateid(lo, &res->stateid, false);
+       } else {
+               /*
+                * We got an entirely new state ID.  Mark all segments for the
+                * inode invalid, and don't bother validating the stateid
+                * sequence number.
+                */
+               pnfs_mark_matching_lsegs_invalid(lo, &free_me, NULL);
+
+               nfs4_stateid_copy(&lo->plh_stateid, &res->stateid);
+               lo->plh_barrier = be32_to_cpu(res->stateid.seqid);
+       }
 
        pnfs_get_lseg(lseg);
        pnfs_layout_insert_lseg(lo, lseg);