From: Chris Horn <hornc@cray.com>
Date: Sat, 2 May 2020 15:37:15 +0000 (-0500)
Subject: LU-13509 ptlrpc: Clear bd_registered in ptlrpc_unregister_bulk
X-Git-Tag: 2.13.55~175
X-Git-Url: https://git.whamcloud.com/?p=fs%2Flustre-release.git;a=commitdiff_plain;h=15057a17ca1e2a0e796cfada6abeb28aa77679cf

LU-13509 ptlrpc: Clear bd_registered in ptlrpc_unregister_bulk

The patch for LU-12816 https://review.whamcloud.com/36309 has us
clearing the bd_registered flag in ptl_send_rpc(). This flag is set
in ptlrpc_register_bulk(), so it makes sense for us to clear it in
ptlrpc_unregister_bulk(). When we're cleaning up in ptl_send_rpc()
we can be sure the flag is cleared with the call to
ptlrpc_unregister_bulk().

This commit also adds a test case for the LU-12816 bug.

Fixes: e6225c07ce4c ("LU-12816 ptlrpc: ptlrpc_register_bulk LBUG on ENOMEM")
Signed-off-by: Chris Horn <hornc@cray.com>
Change-Id: Iabaf109aaf72894cd5acbcacbb0299929ea1a146
Reviewed-on: https://review.whamcloud.com/38457
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Shaun Tancheff <shaun.tancheff@hpe.com>
Reviewed-by: Wang Shilong <wshilong@ddn.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>
---

diff --git a/lustre/include/obd_support.h b/lustre/include/obd_support.h
index 57ec62f..d4b3585 100644
--- a/lustre/include/obd_support.h
+++ b/lustre/include/obd_support.h
@@ -448,6 +448,7 @@ extern char obd_jobid_var[];
 #define OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK 0x51c
 #define OBD_FAIL_PTLRPC_CLIENT_BULK_CB3  0x520
 #define OBD_FAIL_PTLRPC_BULK_ATTACH      0x521
+#define OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH      0x522
 #define OBD_FAIL_PTLRPC_RESEND_RACE	 0x525
 #define OBD_FAIL_PTLRPC_ROUND_XID	 0x530
 #define OBD_FAIL_PTLRPC_CONNECT_RACE	 0x531
diff --git a/lustre/ptlrpc/niobuf.c b/lustre/ptlrpc/niobuf.c
index 43b9db0..bca143e 100644
--- a/lustre/ptlrpc/niobuf.c
+++ b/lustre/ptlrpc/niobuf.c
@@ -446,6 +446,9 @@ int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
 
 	LASSERT(!in_interrupt());     /* might sleep */
 
+	if (desc)
+		desc->bd_registered = 0;
+
 	/* Let's setup deadline for reply unlink. */
 	if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK) &&
 	    async && req->rq_bulk_deadline == 0 && cfs_fail_val == 0)
@@ -827,9 +830,16 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
 			request->rq_repmsg = NULL;
 		}
 
-		reply_me = LNetMEAttach(request->rq_reply_portal,
-					connection->c_peer, request->rq_xid, 0,
-					LNET_UNLINK, LNET_INS_AFTER);
+		if (request->rq_bulk &&
+		    OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH)) {
+			reply_me = ERR_PTR(-ENOMEM);
+		} else {
+			reply_me = LNetMEAttach(request->rq_reply_portal,
+						connection->c_peer,
+						request->rq_xid, 0,
+						LNET_UNLINK, LNET_INS_AFTER);
+		}
+
 		if (IS_ERR(reply_me)) {
 			rc = PTR_ERR(reply_me);
 			CERROR("LNetMEAttach failed: %d\n", rc);
@@ -931,9 +941,7 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
  cleanup_bulk:
 	/* We do sync unlink here as there was no real transfer here so
 	 * the chance to have long unlink to sluggish net is smaller here. */
-        ptlrpc_unregister_bulk(request, 0);
-	if (request->rq_bulk != NULL)
-		request->rq_bulk->bd_registered = 0;
+	ptlrpc_unregister_bulk(request, 0);
  out:
 	if (rc == -ENOMEM) {
 		/* set rq_sent so that this request is treated
diff --git a/lustre/tests/sanity.sh b/lustre/tests/sanity.sh
index e07a051..1fde0e9 100755
--- a/lustre/tests/sanity.sh
+++ b/lustre/tests/sanity.sh
@@ -22788,6 +22788,14 @@ test_423() {
 }
 run_test 423 "statfs should return a right data"
 
+test_424() {
+#define OBD_FAIL_PTLRPC_BULK_REPLY_ATTACH      0x522 | OBD_FAIL_ONCE
+	$LCTL set_param fail_loc=0x80000522
+	dd if=/dev/zero of=$DIR/$tfile bs=2M count=1 oflag=sync
+	rm -f $DIR/$tfile
+}
+run_test 424 "simulate ENOMEM in ptl_send_rpc bulk reply ME attach"
+
 prep_801() {
 	[[ $MDS1_VERSION -lt $(version_code 2.9.55) ]] ||
 	[[ $OST1_VERSION -lt $(version_code 2.9.55) ]] &&