[PATCH v2 2/8] staging/lustre/mdc: fix panic at mdc_free_open()

From: Oleg Drokin
Date: Wed Aug 24 2016 - 11:14:56 EST


From: Alexander Boyko <alexander.boyko@xxxxxxxxxxx>

Assertion was happened for open request when rq_replay is set
to 1.
ASSERTION(mod->mod_open_req->rq_replay == 0)
But this situation is not fatal for client, and could happened
when mdc_close() failed.
The fix allow to free such requests. If mdc_close fail, MDS doesn`t
receive close request from client. And in a worst case client would
be evicted.

The test recreates issue when mdc_close failed and
client asserts:
ASSERTION( mod->mod_open_req->rq_replay == 0 ) failed

Signed-off-by: Alexander Boyko <alexander.boyko@xxxxxxxxxxx>
Seagate-bug-id: MRP-3156
Reviewed-on: http://review.whamcloud.com/17495
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-5282
Reviewed-by: Alex Zhuravlev <alexey.zhuravlev@xxxxxxxxx>
Reviewed-by: Andreas Dilger <andreas.dilger@xxxxxxxxx>
Signed-off-by: Oleg Drokin <green@xxxxxxxxxxxxxx>
---
.../staging/lustre/lustre/include/obd_support.h | 1 +
drivers/staging/lustre/lustre/mdc/mdc_request.c | 56 ++++++++++++++--------
2 files changed, 38 insertions(+), 19 deletions(-)

diff --git a/drivers/staging/lustre/lustre/include/obd_support.h b/drivers/staging/lustre/lustre/include/obd_support.h
index 0c29a33..4a9fe88 100644
--- a/drivers/staging/lustre/lustre/include/obd_support.h
+++ b/drivers/staging/lustre/lustre/include/obd_support.h
@@ -402,6 +402,7 @@ extern char obd_jobid_var[];
#define OBD_FAIL_MDC_GETATTR_ENQUEUE 0x803
#define OBD_FAIL_MDC_RPCS_SEM 0x804
#define OBD_FAIL_MDC_LIGHTWEIGHT 0x805
+#define OBD_FAIL_MDC_CLOSE 0x806

#define OBD_FAIL_MGS 0x900
#define OBD_FAIL_MGS_ALL_REQUEST_NET 0x901
diff --git a/drivers/staging/lustre/lustre/mdc/mdc_request.c b/drivers/staging/lustre/lustre/mdc/mdc_request.c
index 91c0b45..313889a 100644
--- a/drivers/staging/lustre/lustre/mdc/mdc_request.c
+++ b/drivers/staging/lustre/lustre/mdc/mdc_request.c
@@ -677,9 +677,15 @@ static void mdc_free_open(struct md_open_data *mod)
imp_connect_disp_stripe(mod->mod_open_req->rq_import))
committed = 1;

- LASSERT(mod->mod_open_req->rq_replay == 0);
-
- DEBUG_REQ(D_RPCTRACE, mod->mod_open_req, "free open request\n");
+ /*
+ * No reason to asssert here if the open request has
+ * rq_replay == 1. It means that mdc_close failed, and
+ * close request wasn`t sent. It is not fatal to client.
+ * The worst thing is eviction if the client gets open lock
+ */
+ DEBUG_REQ(D_RPCTRACE, mod->mod_open_req,
+ "free open request rq_replay = %d\n",
+ mod->mod_open_req->rq_replay);

ptlrpc_request_committed(mod->mod_open_req, committed);
if (mod->mod_close_req)
@@ -749,22 +755,10 @@ static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
}

*request = NULL;
- req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);
- if (!req)
- return -ENOMEM;
-
- rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
- if (rc) {
- ptlrpc_request_free(req);
- return rc;
- }
-
- /* To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
- * portal whose threads are not taking any DLM locks and are therefore
- * always progressing
- */
- req->rq_request_portal = MDS_READPAGE_PORTAL;
- ptlrpc_at_set_req_timeout(req);
+ if (OBD_FAIL_CHECK(OBD_FAIL_MDC_CLOSE))
+ req = NULL;
+ else
+ req = ptlrpc_request_alloc(class_exp2cliimp(exp), req_fmt);

/* Ensure that this close's handle is fixed up during replay. */
if (likely(mod)) {
@@ -785,6 +779,29 @@ static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
CDEBUG(D_HA,
"couldn't find open req; expecting close error\n");
}
+ if (!req) {
+ /*
+ * TODO: repeat close after errors
+ */
+ CWARN("%s: close of FID "DFID" failed, file reference will be dropped when this client unmounts or is evicted\n",
+ obd->obd_name, PFID(&op_data->op_fid1));
+ rc = -ENOMEM;
+ goto out;
+ }
+
+ rc = ptlrpc_request_pack(req, LUSTRE_MDS_VERSION, MDS_CLOSE);
+ if (rc) {
+ ptlrpc_request_free(req);
+ goto out;
+ }
+
+ /*
+ * To avoid a livelock (bug 7034), we need to send CLOSE RPCs to a
+ * portal whose threads are not taking any DLM locks and are therefore
+ * always progressing
+ */
+ req->rq_request_portal = MDS_READPAGE_PORTAL;
+ ptlrpc_at_set_req_timeout(req);

mdc_close_pack(req, op_data);

@@ -830,6 +847,7 @@ static int mdc_close(struct obd_export *exp, struct md_op_data *op_data,
}
}

+out:
if (mod) {
if (rc != 0)
mod->mod_close_req = NULL;
--
2.7.4