NFS: Further optimise nfs_lock_and_join_requests()
authorTrond Myklebust <trond.myklebust@primarydata.com>
Wed, 19 Jul 2017 12:23:10 +0000 (08:23 -0400)
committerTrond Myklebust <trond.myklebust@primarydata.com>
Tue, 15 Aug 2017 15:54:47 +0000 (11:54 -0400)
When locking the entire group in order to remove subrequests,
the locks are always taken in order, and with the page group
lock being taken after the page head is locked. The intention
is that:

1) The lock on the group head guarantees that requests may not
   be removed from the group (although new entries could be appended
   if we're not holding the group lock).
2) It is safe to drop and retake the page group lock while iterating
   through the list, in particular when waiting for a subrequest lock.

Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
fs/nfs/write.c

index ff7c90c7ff7953cf42c880e81324e6986c717b37..1ee5d89380d9c4fafe7664dde834e1e27eeff4d7 100644 (file)
@@ -377,31 +377,17 @@ nfs_page_group_clear_bits(struct nfs_page *req)
  *
  * returns 0 on success, < 0 on error.
  */
-static int
-nfs_unroll_locks_and_wait(struct inode *inode, struct nfs_page *head,
+static void
+nfs_unroll_locks(struct inode *inode, struct nfs_page *head,
                          struct nfs_page *req)
 {
        struct nfs_page *tmp;
-       int ret;
 
        /* relinquish all the locks successfully grabbed this run */
        for (tmp = head->wb_this_page ; tmp != req; tmp = tmp->wb_this_page)
                nfs_unlock_request(tmp);
 
        WARN_ON_ONCE(test_bit(PG_TEARDOWN, &req->wb_flags));
-
-       /* grab a ref on the request that will be waited on */
-       kref_get(&req->wb_kref);
-
-       nfs_page_group_unlock(head);
-
-       /* release ref from nfs_page_find_head_request_locked */
-       nfs_unlock_and_release_request(head);
-
-       ret = nfs_wait_on_request(req);
-       nfs_release_request(req);
-
-       return ret;
 }
 
 /*
@@ -525,18 +511,21 @@ try_again:
        total_bytes = head->wb_bytes;
        for (subreq = head->wb_this_page; subreq != head;
                        subreq = subreq->wb_this_page) {
-               if (!nfs_lock_request(subreq)) {
+
+               while (!nfs_lock_request(subreq)) {
                        /*
-                        * releases page group bit lock and
-                        * page locks and all references
+                        * Unlock page to allow nfs_page_group_sync_on_bit()
+                        * to succeed
                         */
-                       ret = nfs_unroll_locks_and_wait(inode, head,
-                               subreq);
-
-                       if (ret == 0)
-                               goto try_again;
-
-                       return ERR_PTR(ret);
+                       nfs_page_group_unlock(head);
+                       ret = nfs_wait_on_request(subreq);
+                       if (!ret)
+                               ret = nfs_page_group_lock(head, false);
+                       if (ret < 0) {
+                               nfs_unroll_locks(inode, head, subreq);
+                               nfs_unlock_and_release_request(head);
+                               return ERR_PTR(ret);
+                       }
                }
                /*
                 * Subrequests are always contiguous, non overlapping
@@ -549,7 +538,9 @@ try_again:
                            ((subreq->wb_offset + subreq->wb_bytes) >
                             (head->wb_offset + total_bytes)))) {
                        nfs_unlock_request(subreq);
-                       nfs_unroll_locks_and_wait(inode, head, subreq);
+                       nfs_unroll_locks(inode, head, subreq);
+                       nfs_page_group_unlock(head);
+                       nfs_unlock_and_release_request(head);
                        return ERR_PTR(-EIO);
                }
        }