staging: lustre: llite: Trust creates in revalidate too.

By forcing creates to always go via lookup we lose some
important caching benefits too.
Instead let's trust creates with positive cached entries.

Then we have 3 possible outcomes:
1. Negative dentry - we go via atomic_open and do the create
   by name there.
2. Positive dentry, no contention - we just go straight to
   ll_intent_file_open and open by fid.
3. positive dentry, contention - by the time we reach the server,
   the inode is gone. We get ENOENT which is unacceptable to return
   from create. But since we know it's a create, we substitute it
   with ESTALE and VFS retries again with LOOKUP_REVAL set, we catch
   that in revalidate and force a lookup (same path as before this
   patch).

Signed-off-by: Oleg Drokin <oleg.drokin@intel.com>
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-8371
Reviewed-on: http://review.whamcloud.com/21168
Reviewed-by: James Simmons <uja.ornl@yahoo.com>
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Signed-off-by: James Simmons <jsimmons@infradead.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

2 files changed, 16 insertions, 8 deletions

diff --git a/drivers/staging/lustre/lustre/llite/dcache.c b/drivers/staging/lustre/lustre/llite/dcache.c
index 65bf0c401b44..966f580e26fb 100644
--- a/drivers/staging/lustre/lustre/llite/dcache.c
+++ b/drivers/staging/lustre/lustre/llite/dcache.c
@@ -247,17 +247,14 @@ static int ll_revalidate_dentry(struct dentry *dentry,
                 return 1;
 
         /*
-         * if open&create is set, talk to MDS to make sure file is created if
-         * necessary, because we can't do this in ->open() later since that's
-         * called on an inode. return 0 here to let lookup to handle this.
+         * VFS warns us that this is the second go around and previous
+         * operation failed (most likely open|creat), so this time
+         * we better talk to the server via the lookup path by name,
+         * not by fid.
          */
-        if ((lookup_flags & (LOOKUP_OPEN | LOOKUP_CREATE)) ==
-            (LOOKUP_OPEN | LOOKUP_CREATE))
+        if (lookup_flags & LOOKUP_REVAL)
                 return 0;
 
-        if (lookup_flags & (LOOKUP_PARENT | LOOKUP_OPEN | LOOKUP_CREATE))
-                return 1;
-
         if (!dentry_may_statahead(dir, dentry))
                 return 1;
 
diff --git a/drivers/staging/lustre/lustre/llite/file.c b/drivers/staging/lustre/lustre/llite/file.c
index a1e51a54779e..987090119559 100644
--- a/drivers/staging/lustre/lustre/llite/file.c
+++ b/drivers/staging/lustre/lustre/llite/file.c
@@ -417,6 +417,17 @@ out:
         ptlrpc_req_finished(req);
         ll_intent_drop_lock(itp);
 
+        /*
+         * We did open by fid, but by the time we got to the server,
+         * the object disappeared. If this is a create, we cannot really
+         * tell the userspace that the file it was trying to create
+         * does not exist. Instead let's return -ESTALE, and the VFS will
+         * retry the create with LOOKUP_REVAL that we are going to catch
+         * in ll_revalidate_dentry() and use lookup then.
+         */
+        if (rc == -ENOENT && itp->it_op & IT_CREAT)
+                rc = -ESTALE;
+
         return rc;
 }