posix: fix GFID assignment race between mkdir and lookup

In lookup, treat inodes which have recent ctime and no GFID as
though they are still getting created by another thread and
return ENOENT

Signed-off-by: Anand Avati <avati@gluster.com>

BUG: 2994 ([glusterfs-3.2.1qa2]: untar and rm in parallel hangs untar)
URL: http://bugs.gluster.com/cgi-bin/bugzilla3/show_bug.cgi?id=2994

1 files changed, 69 insertions, 1 deletions

diff --git a/xlators/storage/posix/src/posix.c b/xlators/storage/posix/src/posix.c
index f95c9f8ffbc..2472ff90c34 100644
--- a/xlators/storage/posix/src/posix.c
+++ b/xlators/storage/posix/src/posix.c
@@ -405,6 +405,74 @@ out:
 }
 
 
+static int
+is_fresh_file (struct stat *stat)
+{
+        struct timeval tv;
+
+        gettimeofday (&tv, NULL);
+
+        if ((stat->st_ctime >= (tv.tv_sec - 1))
+            && (stat->st_ctime <= tv.tv_sec))
+                return 1;
+
+        return 0;
+}
+
+
+int
+posix_gfid_heal (xlator_t *this, const char *path, dict_t *xattr_req)
+{
+        /* The purpose of this function is to prevent a race
+           where an inode creation FOP (like mkdir/mknod/create etc)
+           races with lookup in the following way:
+
+                   {create thread}       |    {lookup thread}
+                                         |
+                                         t0
+                      mkdir ("name")     |
+                                         t1
+                                         |     posix_gfid_set ("name", 2);
+                                         t2
+             posix_gfid_set ("name", 1); |
+                                         t3
+                      lstat ("name");    |     lstat ("name");
+
+          In the above case mkdir FOP would have resulted with GFID 2 while
+          it should have been GFID 1. It matters in the case where GFID would
+          have gotten set to 1 on other subvolumes of replciate/distribute
+
+          The "solution" here is that, if we detect lookup is attempting to
+          set a GFID on a file which is created very recently, but does not
+          yet have a GFID (i.e, between t1 and t2), then "fake" it as though
+          posix_gfid_heal was called at t0 instead.
+        */
+
+        uuid_t       uuid_curr;
+        int          ret = 0;
+        struct stat  stat = {0, };
+
+        if (!xattr_req)
+                goto out;
+
+        if (sys_lstat (path, &stat) != 0)
+                goto out;
+
+        ret = sys_lgetxattr (path, GFID_XATTR_KEY, uuid_curr, 16);
+        if (ret != 16) {
+                if (is_fresh_file (&stat)) {
+                        ret = -1;
+                        errno = ENOENT;
+                        goto out;
+                }
+        }
+
+        ret = posix_gfid_set (this, path, xattr_req);
+out:
+        return ret;
+}
+
+
 int32_t
 posix_lookup (call_frame_t *frame, xlator_t *this,
               loc_t *loc, dict_t *xattr_req)
@@ -426,7 +494,7 @@ posix_lookup (call_frame_t *frame, xlator_t *this,
 
         MAKE_REAL_PATH (real_path, this, loc->path);
 
-        posix_gfid_set (this, real_path, xattr_req);
+        posix_gfid_heal (this, real_path, xattr_req);
 
         op_ret   = posix_lstat_with_gfid (this, real_path, &buf);
         op_errno = errno;