]> git.ipfire.org Git - people/arne_f/kernel.git/commitdiff
pnfs: lift retry logic from send_layoutget to pnfs_update_layout
authorJeff Layton <jlayton@poochiereds.net>
Tue, 17 May 2016 16:28:46 +0000 (12:28 -0400)
committerAnna Schumaker <Anna.Schumaker@Netapp.com>
Tue, 17 May 2016 19:48:12 +0000 (15:48 -0400)
If we get back something like NFS4ERR_OLD_STATEID, that will be
translated into -EAGAIN, and the do/while loop in send_layoutget
will drive the call again.

This is not quite what we want, I think. An error like that is a
sign that something has changed. That something could have been a
concurrent LAYOUTGET that would give us a usable lseg.

Lift the retry logic into pnfs_update_layout instead. That allows
us to redo the layout search, and may spare us from having to issue
an RPC.

Signed-off-by: Jeff Layton <jeff.layton@primarydata.com>
Signed-off-by: Anna Schumaker <Anna.Schumaker@Netapp.com>
fs/nfs/pnfs.c

index e3e9973b5f762b2ff0640be3ba9c2ca1135791eb..2a609af845fe75ef9e904b1dbc2cd669ae930ce5 100644 (file)
@@ -839,7 +839,6 @@ send_layoutget(struct pnfs_layout_hdr *lo,
        struct inode *ino = lo->plh_inode;
        struct nfs_server *server = NFS_SERVER(ino);
        struct nfs4_layoutget *lgp;
-       struct pnfs_layout_segment *lseg;
        loff_t i_size;
 
        dprintk("--> %s\n", __func__);
@@ -849,45 +848,30 @@ send_layoutget(struct pnfs_layout_hdr *lo,
         * store in lseg. If we race with a concurrent seqid morphing
         * op, then re-send the LAYOUTGET.
         */
-       do {
-               lgp = kzalloc(sizeof(*lgp), gfp_flags);
-               if (lgp == NULL)
-                       return NULL;
-
-               i_size = i_size_read(ino);
-
-               lgp->args.minlength = PAGE_SIZE;
-               if (lgp->args.minlength > range->length)
-                       lgp->args.minlength = range->length;
-               if (range->iomode == IOMODE_READ) {
-                       if (range->offset >= i_size)
-                               lgp->args.minlength = 0;
-                       else if (i_size - range->offset < lgp->args.minlength)
-                               lgp->args.minlength = i_size - range->offset;
-               }
-               lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
-               pnfs_copy_range(&lgp->args.range, range);
-               lgp->args.type = server->pnfs_curr_ld->id;
-               lgp->args.inode = ino;
-               lgp->args.ctx = get_nfs_open_context(ctx);
-               lgp->gfp_flags = gfp_flags;
-               lgp->cred = lo->plh_lc_cred;
+       lgp = kzalloc(sizeof(*lgp), gfp_flags);
+       if (lgp == NULL)
+               return ERR_PTR(-ENOMEM);
 
-               lseg = nfs4_proc_layoutget(lgp, gfp_flags);
-       } while (lseg == ERR_PTR(-EAGAIN));
+       i_size = i_size_read(ino);
 
-       if (IS_ERR(lseg)) {
-               if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
-                       pnfs_layout_clear_fail_bit(lo,
-                                       pnfs_iomode_to_fail_bit(range->iomode));
-                       lseg = NULL;
-               }
-       } else {
-               pnfs_layout_clear_fail_bit(lo,
-                               pnfs_iomode_to_fail_bit(range->iomode));
+       lgp->args.minlength = PAGE_SIZE;
+       if (lgp->args.minlength > range->length)
+               lgp->args.minlength = range->length;
+       if (range->iomode == IOMODE_READ) {
+               if (range->offset >= i_size)
+                       lgp->args.minlength = 0;
+               else if (i_size - range->offset < lgp->args.minlength)
+                       lgp->args.minlength = i_size - range->offset;
        }
+       lgp->args.maxcount = PNFS_LAYOUT_MAXSIZE;
+       pnfs_copy_range(&lgp->args.range, range);
+       lgp->args.type = server->pnfs_curr_ld->id;
+       lgp->args.inode = ino;
+       lgp->args.ctx = get_nfs_open_context(ctx);
+       lgp->gfp_flags = gfp_flags;
+       lgp->cred = lo->plh_lc_cred;
 
-       return lseg;
+       return nfs4_proc_layoutget(lgp, gfp_flags);
 }
 
 static void pnfs_clear_layoutcommit(struct inode *inode,
@@ -1649,6 +1633,22 @@ lookup_again:
                arg.length = PAGE_ALIGN(arg.length);
 
        lseg = send_layoutget(lo, ctx, &arg, gfp_flags);
+       if (IS_ERR(lseg)) {
+               if (lseg == ERR_PTR(-EAGAIN)) {
+                       if (first)
+                               pnfs_clear_first_layoutget(lo);
+                       pnfs_put_layout_hdr(lo);
+                       goto lookup_again;
+               }
+
+               if (!nfs_error_is_fatal(PTR_ERR(lseg))) {
+                       pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
+                       lseg = NULL;
+               }
+       } else {
+               pnfs_layout_clear_fail_bit(lo, pnfs_iomode_to_fail_bit(iomode));
+       }
+
        atomic_dec(&lo->plh_outstanding);
        trace_pnfs_update_layout(ino, pos, count, iomode, lo,
                                 PNFS_UPDATE_LAYOUT_SEND_LAYOUTGET);