Whamcloud - gitweb
LU-5528 ptlrpc: fix race between connect vs resend
[fs/lustre-release.git] / lustre / ptlrpc / niobuf.c
index f779054..c44fe83 100644 (file)
@@ -35,9 +35,6 @@
  */
 
 #define DEBUG_SUBSYSTEM S_RPC
-#ifndef __KERNEL__
-#include <liblustre.h>
-#endif
 #include <obd_support.h>
 #include <lustre_net.h>
 #include <lustre_lib.h>
@@ -458,12 +455,10 @@ int ptlrpc_unregister_bulk(struct ptlrpc_request *req, int async)
                 RETURN(0);
 
         for (;;) {
-#ifdef __KERNEL__
                /* The wq argument is ignored by user-space wait_event macros */
                wait_queue_head_t *wq = (req->rq_set != NULL) ?
                                        &req->rq_set->set_waitq :
                                        &req->rq_reply_waitq;
-#endif
                 /* Network access will complete in finite time but the HUGE
                  * timeout lets us CWARN for visibility of sluggish NALs */
                 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
@@ -678,7 +673,8 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
         struct ptlrpc_connection *connection;
         lnet_handle_me_t  reply_me_h;
         lnet_md_t         reply_md;
-        struct obd_device *obd = request->rq_import->imp_obd;
+       struct obd_import *imp = request->rq_import;
+       struct obd_device *obd = imp->imp_obd;
         ENTRY;
 
         if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_DROP_RPC))
@@ -691,7 +687,7 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
          * cleanly from the previous attempt */
         LASSERT(!request->rq_receiving_reply);
        LASSERT(!((lustre_msg_get_flags(request->rq_reqmsg) & MSG_REPLAY) &&
-               (request->rq_import->imp_state == LUSTRE_IMP_FULL)));
+               (imp->imp_state == LUSTRE_IMP_FULL)));
 
        if (unlikely(obd != NULL && obd->obd_fail)) {
                CDEBUG(D_HA, "muting rpc for failed imp obd %s\n",
@@ -704,19 +700,28 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
                 RETURN(-ENODEV);
         }
 
-        connection = request->rq_import->imp_connection;
-
-        lustre_msg_set_handle(request->rq_reqmsg,
-                              &request->rq_import->imp_remote_handle);
-        lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
-        lustre_msg_set_conn_cnt(request->rq_reqmsg,
-                                request->rq_import->imp_conn_cnt);
-        lustre_msghdr_set_flags(request->rq_reqmsg,
-                                request->rq_import->imp_msghdr_flags);
-
-        if (request->rq_resend)
-                lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
-
+       connection = imp->imp_connection;
+
+       lustre_msg_set_handle(request->rq_reqmsg,
+                             &imp->imp_remote_handle);
+       lustre_msg_set_type(request->rq_reqmsg, PTL_RPC_MSG_REQUEST);
+       lustre_msg_set_conn_cnt(request->rq_reqmsg,
+                               imp->imp_conn_cnt);
+       lustre_msghdr_set_flags(request->rq_reqmsg,
+                               imp->imp_msghdr_flags);
+
+       /** For enabled AT all request should have AT_SUPPORT in the
+        * FULL import state when OBD_CONNECT_AT is set */
+       LASSERT(AT_OFF || imp->imp_state != LUSTRE_IMP_FULL ||
+               (imp->imp_msghdr_flags & MSGHDR_AT_SUPPORT) ||
+               !(imp->imp_connect_data.ocd_connect_flags &
+               OBD_CONNECT_AT));
+
+       if (request->rq_resend) {
+               lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
+               if (request->rq_resend_cb != NULL)
+                       request->rq_resend_cb(request, &request->rq_async_args);
+       }
         if (request->rq_memalloc)
                 mpflag = cfs_memory_pressure_get_and_set();
 
@@ -763,18 +768,18 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
         }
 
        spin_lock(&request->rq_lock);
-        /* If the MD attach succeeds, there _will_ be a reply_in callback */
-        request->rq_receiving_reply = !noreply;
-        /* We are responsible for unlinking the reply buffer */
-        request->rq_must_unlink = !noreply;
-        /* Clear any flags that may be present from previous sends. */
+       /* We are responsible for unlinking the reply buffer */
+       request->rq_reply_unlinked = noreply;
+       request->rq_receiving_reply = !noreply;
+       /* Clear any flags that may be present from previous sends. */
+       request->rq_req_unlinked = 0;
         request->rq_replied = 0;
         request->rq_err = 0;
         request->rq_timedout = 0;
         request->rq_net_err = 0;
         request->rq_resend = 0;
         request->rq_restart = 0;
-        request->rq_reply_truncate = 0;
+       request->rq_reply_truncated = 0;
        spin_unlock(&request->rq_lock);
 
         if (!noreply) {
@@ -789,8 +794,8 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
                 reply_md.user_ptr  = &request->rq_reply_cbid;
                 reply_md.eq_handle = ptlrpc_eq_h;
 
-                /* We must see the unlink callback to unset rq_must_unlink,
-                   so we can't auto-unlink */
+               /* We must see the unlink callback to set rq_reply_unlinked,
+                * so we can't auto-unlink */
                 rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN,
                                   &request->rq_reply_md_h);
                 if (rc != 0) {
@@ -813,18 +818,18 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
         ptlrpc_request_addref(request);
        if (obd != NULL && obd->obd_svc_stats != NULL)
                lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQACTIVE_CNTR,
-                       atomic_read(&request->rq_import->imp_inflight));
+                       atomic_read(&imp->imp_inflight));
 
        OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_DELAY_SEND, request->rq_timeout + 5);
 
-       do_gettimeofday(&request->rq_arrival_time);
+       do_gettimeofday(&request->rq_sent_tv);
        request->rq_sent = cfs_time_current_sec();
        /* We give the server rq_timeout secs to process the req, and
           add the network latency for our local timeout. */
         request->rq_deadline = request->rq_sent + request->rq_timeout +
                 ptlrpc_at_get_net_latency(request);
 
-        ptlrpc_pinger_sending_on_import(request->rq_import);
+       ptlrpc_pinger_sending_on_import(imp);
 
         DEBUG_REQ(D_INFO, request, "send flg=%x",
                   lustre_msg_get_flags(request->rq_reqmsg));
@@ -834,9 +839,10 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
                           connection,
                           request->rq_request_portal,
                           request->rq_xid, 0);
-        if (rc == 0)
-                GOTO(out, rc);
+       if (likely(rc == 0))
+               GOTO(out, rc);
 
+       request->rq_req_unlinked = 1;
         ptlrpc_req_finished(request);
         if (noreply)
                 GOTO(out, rc);