From: Vitaly Fertman <vitaly_fertman@xyratex.com>
Date: Thu, 9 Oct 2014 15:18:34 +0000 (-0400)
Subject: LU-5579 ldlm: re-sent enqueue vs lock destroy race
X-Git-Tag: 2.6.54~5
X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=commitdiff_plain;h=e77a7afe261cc2a9393adf4ffe59401c830d16ec;ds=sidebyside

LU-5579 ldlm: re-sent enqueue vs lock destroy race

Upon lock enqueue re-send, lock is pinned by ldlm_handle_enqueue0,
however it may race with client eviction or even lock cancel (if
a reply for the original RPC finally reached the client) and the
lock cannot be found by cookie anymore:

ASSERTION( lock != NULL ) failed: Invalid lock handle

Signed-off-by: Vitaly Fertman <vitaly_fertman@xyratex.com>
Change-Id: I9d8156bf78a1b83ac22ffaa1148feb43bef37b1a
Xyratex-bug-id: MRP-2094
Reviewed-on: http://review.whamcloud.com/11839
Tested-by: Jenkins
Reviewed-by: James Simmons <uja.ornl@gmail.com>
Tested-by: Maloo <hpdd-maloo@intel.com>
Reviewed-by: Oleg Drokin <oleg.drokin@intel.com>
---

diff --git a/lustre/mdt/mdt_handler.c b/lustre/mdt/mdt_handler.c
index 204ebd5..c1e5ea8 100644
--- a/lustre/mdt/mdt_handler.c
+++ b/lustre/mdt/mdt_handler.c
@@ -1351,7 +1351,7 @@ static int mdt_getattr_name_lock(struct mdt_thread_info *info,
 
 		rc = mdt_check_resent_lock(info, child, lhc);
 		if (rc < 0) {
-			RETURN(-EPROTO);
+			RETURN(rc);
 		} else if (rc > 0) {
 			mdt_lock_handle_init(lhc);
 			mdt_lock_reg_init(lhc, LCK_PR);
@@ -2329,8 +2329,14 @@ int mdt_check_resent_lock(struct mdt_thread_info *info,
 
 		lock = ldlm_handle2lock(&lhc->mlh_reg_lh);
 		LASSERT(lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT);
-		LASSERTF(lock != NULL, "Invalid lock handle "LPX64"\n",
-			 lhc->mlh_reg_lh.cookie);
+		if (lock == NULL) {
+			/* Lock is pinned by ldlm_handle_enqueue0() as it is
+			 * a resend case, however, it could be already destroyed
+			 * due to client eviction or a raced cancel RPC. */
+			LDLM_DEBUG_NOLOCK("Invalid lock handle "LPX64"\n",
+					  lhc->mlh_reg_lh.cookie);
+			RETURN(-ESTALE);
+		}
 
 		if (!fid_res_name_eq(mdt_object_fid(mo),
 				     &lock->l_resource->lr_name)) {