Whamcloud - gitweb
LU-8645 ptlrpc: update imp_known_replied_xid on resend-replay
[fs/lustre-release.git] / lustre / ptlrpc / recover.c
index ec6e6e0..02f79d0 100644 (file)
  *
  * You should have received a copy of the GNU General Public License
  * version 2 along with this program; If not, see
- * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
- *
- * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
- * CA 95054 USA or visit www.sun.com if you need additional information or
- * have any questions.
+ * http://www.gnu.org/licenses/gpl-2.0.html
  *
  * GPL HEADER END
  */
@@ -27,7 +23,7 @@
  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, Whamcloud, Inc.
+ * Copyright (c) 2011, 2015, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
  */
 
 #define DEBUG_SUBSYSTEM S_RPC
-#ifdef __KERNEL__
-# include <libcfs/libcfs.h>
-#else
-# include <liblustre.h>
-#endif
-
+#include <linux/list.h>
+#include <libcfs/libcfs.h>
 #include <obd_support.h>
 #include <lustre_ha.h>
 #include <lustre_net.h>
 #include <lustre_import.h>
 #include <lustre_export.h>
 #include <obd.h>
-#include <obd_ost.h>
 #include <obd_class.h>
-#include <obd_lov.h> /* for IOC_LOV_SET_OSC_ACTIVE */
-#include <libcfs/list.h>
 
 #include "ptlrpc_internal.h"
 
@@ -79,7 +68,7 @@ void ptlrpc_initiate_recovery(struct obd_import *imp)
 int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
 {
         int rc = 0;
-        cfs_list_t *tmp, *pos;
+       struct list_head *tmp, *pos;
         struct ptlrpc_request *req = NULL;
         __u64 last_transno;
         ENTRY;
@@ -95,7 +84,7 @@ int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
        last_transno = imp->imp_last_replay_transno;
        spin_unlock(&imp->imp_lock);
 
-        CDEBUG(D_HA, "import %p from %s committed "LPU64" last "LPU64"\n",
+       CDEBUG(D_HA, "import %p from %s committed %llu last %llu\n",
                imp, obd2cli_tgt(imp->imp_obd),
                imp->imp_peer_committed_transno, last_transno);
 
@@ -114,38 +103,88 @@ int ptlrpc_replay_next(struct obd_import *imp, int *inflight)
          * imp_lock is being held by ptlrpc_replay, but it's not. it's
          * just a little race...
          */
-        cfs_list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
-                req = cfs_list_entry(tmp, struct ptlrpc_request,
-                                     rq_replay_list);
-
-                /* If need to resend the last sent transno (because a
-                   reconnect has occurred), then stop on the matching
-                   req and send it again. If, however, the last sent
-                   transno has been committed then we continue replay
-                   from the next request. */
-                if (req->rq_transno > last_transno) {
-                        if (imp->imp_resend_replay)
-                                lustre_msg_add_flags(req->rq_reqmsg,
-                                                     MSG_RESENT);
-                        break;
-                }
-                req = NULL;
-        }
+
+       /* Replay all the committed open requests on committed_list first */
+       if (!list_empty(&imp->imp_committed_list)) {
+               tmp = imp->imp_committed_list.prev;
+               req = list_entry(tmp, struct ptlrpc_request,
+                                    rq_replay_list);
+
+               /* The last request on committed_list hasn't been replayed */
+               if (req->rq_transno > last_transno) {
+                       /* Since the imp_committed_list is immutable before
+                        * all of it's requests being replayed, it's safe to
+                        * use a cursor to accelerate the search */
+                       if (!imp->imp_resend_replay ||
+                           imp->imp_replay_cursor == &imp->imp_committed_list)
+                               imp->imp_replay_cursor =
+                                       imp->imp_replay_cursor->next;
+
+                       while (imp->imp_replay_cursor !=
+                              &imp->imp_committed_list) {
+                               req = list_entry(imp->imp_replay_cursor,
+                                                    struct ptlrpc_request,
+                                                    rq_replay_list);
+                               if (req->rq_transno > last_transno)
+                                       break;
+
+                               req = NULL;
+                               imp->imp_replay_cursor =
+                                       imp->imp_replay_cursor->next;
+                       }
+               } else {
+                       /* All requests on committed_list have been replayed */
+                       imp->imp_replay_cursor = &imp->imp_committed_list;
+                       req = NULL;
+               }
+       }
+
+       /* All the requests in committed list have been replayed, let's replay
+        * the imp_replay_list */
+       if (req == NULL) {
+               list_for_each_safe(tmp, pos, &imp->imp_replay_list) {
+                       req = list_entry(tmp, struct ptlrpc_request,
+                                            rq_replay_list);
+
+                       if (req->rq_transno > last_transno)
+                               break;
+                       req = NULL;
+               }
+       }
+
+       /* If need to resend the last sent transno (because a reconnect
+        * has occurred), then stop on the matching req and send it again.
+        * If, however, the last sent transno has been committed then we
+        * continue replay from the next request. */
+       if (req != NULL && imp->imp_resend_replay)
+               lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
 
        spin_lock(&imp->imp_lock);
+       /* The resend replay request may have been removed from the
+        * unreplied list. */
+       if (req != NULL && imp->imp_resend_replay &&
+           list_empty(&req->rq_unreplied_list)) {
+               ptlrpc_add_unreplied(req);
+               imp->imp_known_replied_xid = ptlrpc_known_replied_xid(imp);
+       }
+
        imp->imp_resend_replay = 0;
        spin_unlock(&imp->imp_lock);
 
-        if (req != NULL) {
-                rc = ptlrpc_replay_req(req);
-                if (rc) {
-                        CERROR("recovery replay error %d for req "
-                               LPU64"\n", rc, req->rq_xid);
-                        RETURN(rc);
-                }
-                *inflight = 1;
-        }
-        RETURN(rc);
+       if (req != NULL) {
+               /* The request should have been added back in unreplied list
+                * by ptlrpc_prepare_replay(). */
+               LASSERT(!list_empty(&req->rq_unreplied_list));
+
+               rc = ptlrpc_replay_req(req);
+               if (rc) {
+                       CERROR("recovery replay error %d for req "
+                              "%llu\n", rc, req->rq_xid);
+                       RETURN(rc);
+               }
+               *inflight = 1;
+       }
+       RETURN(rc);
 }
 
 /**
@@ -170,19 +209,21 @@ int ptlrpc_resend(struct obd_import *imp)
                 RETURN(-1);
         }
 
-        cfs_list_for_each_entry_safe(req, next, &imp->imp_sending_list,
-                                     rq_list) {
-                LASSERTF((long)req > CFS_PAGE_SIZE && req != LP_POISON,
-                         "req %p bad\n", req);
-                LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
-                if (!ptlrpc_no_resend(req))
-                        ptlrpc_resend_req(req);
-        }
+       list_for_each_entry_safe(req, next, &imp->imp_sending_list, rq_list) {
+               LASSERTF((long)req > PAGE_SIZE && req != LP_POISON,
+                        "req %p bad\n", req);
+               LASSERTF(req->rq_type != LI_POISON, "req %p freed\n", req);
+
+               /* If the request is allowed to be sent during replay and it
+                * is not timeout yet, then it does not need to be resent. */
+               if (!ptlrpc_no_resend(req) &&
+                   (req->rq_timedout || !req->rq_allow_replay))
+                       ptlrpc_resend_req(req);
+       }
        spin_unlock(&imp->imp_lock);
 
        RETURN(0);
 }
-EXPORT_SYMBOL(ptlrpc_resend);
 
 /**
  * Go through all requests in delayed list and wake their threads
@@ -190,19 +231,18 @@ EXPORT_SYMBOL(ptlrpc_resend);
  */
 void ptlrpc_wake_delayed(struct obd_import *imp)
 {
-       cfs_list_t *tmp, *pos;
+       struct list_head *tmp, *pos;
        struct ptlrpc_request *req;
 
        spin_lock(&imp->imp_lock);
-       cfs_list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
-               req = cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
+       list_for_each_safe(tmp, pos, &imp->imp_delayed_list) {
+               req = list_entry(tmp, struct ptlrpc_request, rq_list);
 
                DEBUG_REQ(D_HA, req, "waking (set %p):", req->rq_set);
                ptlrpc_client_wake_req(req);
        }
        spin_unlock(&imp->imp_lock);
 }
-EXPORT_SYMBOL(ptlrpc_wake_delayed);
 
 void ptlrpc_request_handle_notconn(struct ptlrpc_request *failed_req)
 {
@@ -295,7 +335,7 @@ int ptlrpc_recover_import(struct obd_import *imp, char *new_uuid, int async)
 
        spin_lock(&imp->imp_lock);
        if (imp->imp_state == LUSTRE_IMP_NEW || imp->imp_deactive ||
-           cfs_atomic_read(&imp->imp_inval_count))
+           atomic_read(&imp->imp_inval_count))
                rc = -EINVAL;
        spin_unlock(&imp->imp_lock);
         if (rc)
@@ -351,11 +391,14 @@ EXPORT_SYMBOL(ptlrpc_recover_import);
 int ptlrpc_import_in_recovery(struct obd_import *imp)
 {
        int in_recovery = 1;
+
        spin_lock(&imp->imp_lock);
        if (imp->imp_state == LUSTRE_IMP_FULL ||
            imp->imp_state == LUSTRE_IMP_CLOSED ||
-           imp->imp_state == LUSTRE_IMP_DISCON)
+           imp->imp_state == LUSTRE_IMP_DISCON ||
+           imp->imp_obd->obd_no_recov)
                in_recovery = 0;
        spin_unlock(&imp->imp_lock);
+
        return in_recovery;
 }