Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / ptlrpc / client.c
index 8fa9fff..de5f94b 100644 (file)
@@ -1,26 +1,37 @@
 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
  * vim:expandtab:shiftwidth=8:tabstop=8:
  *
- *  Copyright (c) 2002, 2003 Cluster File Systems, Inc.
+ * GPL HEADER START
  *
- *   This file is part of the Lustre file system, http://www.lustre.org
- *   Lustre is a trademark of Cluster File Systems, Inc.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
- *   You may have signed or agreed to another license before downloading
- *   this software.  If so, you are bound by the terms and conditions
- *   of that agreement, and the following does not apply to you.  See the
- *   LICENSE file included with this distribution for more information.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
  *
- *   If you did not agree to a different license, then this copy of Lustre
- *   is open source software; you can redistribute it and/or modify it
- *   under the terms of version 2 of the GNU General Public License as
- *   published by the Free Software Foundation.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
  *
- *   In either case, Lustre is distributed in the hope that it will be
- *   useful, but WITHOUT ANY WARRANTY; without even the implied warranty
- *   of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- *   license text for more details.
+ * 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.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
  */
 
 #define DEBUG_SUBSYSTEM S_RPC
@@ -60,7 +71,7 @@ struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
                 return NULL;
         }
 
-        c = ptlrpc_get_connection(peer, self, uuid);
+        c = ptlrpc_connection_get(peer, self, uuid);
         if (c) {
                 memcpy(c->c_remote_uuid.uuid,
                        uuid->uuid, sizeof(c->c_remote_uuid.uuid));
@@ -71,24 +82,6 @@ struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
         return c;
 }
 
-void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
-                                 struct obd_uuid *uuid)
-{
-        lnet_nid_t        self;
-        lnet_process_id_t peer;
-        int               err;
-
-        err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
-        if (err != 0) {
-                CERROR("cannot find peer %s!\n", uuid->uuid);
-                return;
-        }
-
-        conn->c_peer = peer;
-        conn->c_self = self;
-        return;
-}
-
 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
 {
         struct ptlrpc_bulk_desc *desc;
@@ -101,7 +94,7 @@ static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal
         cfs_waitq_init(&desc->bd_waitq);
         desc->bd_max_iov = npages;
         desc->bd_iov_count = 0;
-        desc->bd_md_h = LNET_INVALID_HANDLE;
+        LNetInvalidateHandle(&desc->bd_md_h);
         desc->bd_portal = portal;
         desc->bd_type = type;
 
@@ -214,8 +207,7 @@ void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
         idx = import_at_get_index(req->rq_import,
                                   req->rq_request_portal);
         serv_est = at_get(&at->iat_service_estimate[idx]);
-        /* add an arbitrary minimum: 125% +5 sec */
-        req->rq_timeout = serv_est + (serv_est >> 2) + 5;
+        req->rq_timeout = at_est2timeout(serv_est);
         /* We could get even fancier here, using history to predict increased
            loading... */
 
@@ -225,17 +217,19 @@ void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
 }
 
 /* Adjust max service estimate based on server value */
-static void ptlrpc_at_adj_service(struct ptlrpc_request *req)
+static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
+                                  unsigned int serv_est)
 {
         int idx;
-        unsigned int serv_est, oldse;
-        struct imp_at *at = &req->rq_import->imp_at;
+        unsigned int oldse;
+        struct imp_at *at;
 
-        LASSERT(req->rq_import);
+        /* do estimate only if is not in recovery */
+        if (!(req->rq_send_state & (LUSTRE_IMP_FULL | LUSTRE_IMP_CONNECTING)))
+                return;
 
-        /* service estimate is returned in the repmsg timeout field,
-           may be 0 on err */
-        serv_est = lustre_msg_get_timeout(req->rq_repmsg);
+        LASSERT(req->rq_import);
+        at = &req->rq_import->imp_at;
 
         idx = import_at_get_index(req->rq_import, req->rq_request_portal);
         /* max service estimates are tracked on the server side,
@@ -255,21 +249,22 @@ int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
 }
 
 /* Adjust expected network latency */
-static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req)
+static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
+                                      unsigned int service_time)
 {
-        unsigned int st, nl, oldnl;
-        struct imp_at *at = &req->rq_import->imp_at;
+        unsigned int nl, oldnl;
+        struct imp_at *at;
         time_t now = cfs_time_current_sec();
 
         LASSERT(req->rq_import);
-
-        st = lustre_msg_get_service_time(req->rq_repmsg);
+        at = &req->rq_import->imp_at;
 
         /* Network latency is total time less server processing time */
-        nl = max_t(int, now - req->rq_sent - st, 0) + 1/*st rounding*/;
-        if (st > now - req->rq_sent + 2 /* rounding */)
-                CERROR("Reported service time %u > total measured time %ld\n",
-                       st, now - req->rq_sent);
+        nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
+        if (service_time > now - req->rq_sent + 3 /* bz16408 */)
+                CWARN("Reported service time %u > total measured time "
+                      CFS_DURATION_T"\n", service_time,
+                      cfs_time_sub(now, req->rq_sent));
 
         oldnl = at_add(&at->iat_net_latency, nl);
         if (oldnl != 0)
@@ -306,44 +301,53 @@ static int unpack_reply(struct ptlrpc_request *req)
  * Handle an early reply message, called with the rq_lock held.
  * If anything goes wrong just ignore it - same as if it never happened
  */
-static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) {
-        time_t          olddl;
-        int             rc;
+static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
+{
+        struct ptlrpc_request *early_req;
+        time_t                 olddl;
+        int                    rc;
         ENTRY;
 
         req->rq_early = 0;
         spin_unlock(&req->rq_lock);
 
-        rc = sptlrpc_cli_unwrap_early_reply(req);
-        if (rc)
-                GOTO(out, rc);
+        rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
+        if (rc) {
+                spin_lock(&req->rq_lock);
+                RETURN(rc);
+        }
 
-        rc = unpack_reply(req);
-        if (rc)
-                GOTO(out_cleanup, rc);
+        rc = unpack_reply(early_req);
+        if (rc == 0) {
+                /* Expecting to increase the service time estimate here */
+                ptlrpc_at_adj_service(req,
+                        lustre_msg_get_timeout(early_req->rq_repmsg));
+                ptlrpc_at_adj_net_latency(req,
+                        lustre_msg_get_service_time(early_req->rq_repmsg));
+        }
 
-        /* Expecting to increase the service time estimate here */
-        ptlrpc_at_adj_service(req);
-        ptlrpc_at_adj_net_latency(req);
+        sptlrpc_cli_finish_early_reply(early_req);
 
-        /* Adjust the local timeout for this req */
-        ptlrpc_at_set_req_timeout(req);
+        spin_lock(&req->rq_lock);
 
-        olddl = req->rq_deadline;
-        /* server assumes it now has rq_timeout from when it sent the
-           early reply, so client should give it at least that long. */
-        req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
-                    ptlrpc_at_get_net_latency(req);
+        if (rc == 0) {
+                /* Adjust the local timeout for this req */
+                ptlrpc_at_set_req_timeout(req);
 
-        DEBUG_REQ(D_ADAPTTO, req,
-                  "Early reply #%d, new deadline in %lds (%+lds)",
-                  req->rq_early_count, req->rq_deadline -
-                  cfs_time_current_sec(), req->rq_deadline - olddl);
+                olddl = req->rq_deadline;
+                /* server assumes it now has rq_timeout from when it sent the
+                   early reply, so client should give it at least that long. */
+                req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
+                            ptlrpc_at_get_net_latency(req);
+
+                DEBUG_REQ(D_ADAPTTO, req,
+                          "Early reply #%d, new deadline in "CFS_DURATION_T"s "
+                          "("CFS_DURATION_T"s)", req->rq_early_count,
+                          cfs_time_sub(req->rq_deadline,
+                                       cfs_time_current_sec()),
+                          cfs_time_sub(req->rq_deadline, olddl));
+        }
 
-out_cleanup:
-        sptlrpc_cli_finish_early_reply(req);
-out:
-        spin_lock(&req->rq_lock);
         RETURN(rc);
 }
 
@@ -462,6 +466,7 @@ static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_po
         request->rq_reqbuf = reqbuf;
         request->rq_reqbuf_len = pool->prp_rq_size;
         request->rq_pool = pool;
+
         return request;
 }
 
@@ -477,7 +482,7 @@ static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
 
 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
                                       __u32 version, int opcode,
-                                      int count, int *lengths, char **bufs,
+                                      int count, __u32 *lengths, char **bufs,
                                       struct ptlrpc_cli_ctx *ctx)
 {
         struct obd_import  *imp = request->rq_import;
@@ -555,7 +560,7 @@ int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
 
 int ptlrpc_request_pack(struct ptlrpc_request *request,
-                        __u32 version, int opcode) 
+                        __u32 version, int opcode)
 {
         return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
 }
@@ -573,9 +578,10 @@ struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
                 OBD_ALLOC_PTR(request);
 
         if (request) {
-                LASSERT((unsigned long)imp > 0x1000);
+                LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
                 LASSERT(imp != LP_POISON);
-                LASSERT((unsigned long)imp->imp_client > 0x1000);
+                LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
+                        imp->imp_client);
                 LASSERT(imp->imp_client != LP_POISON);
 
                 request->rq_import = class_import_get(imp);
@@ -643,7 +649,7 @@ struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
 struct ptlrpc_request *
 ptlrpc_prep_req_pool(struct obd_import *imp,
                      __u32 version, int opcode,
-                     int count, int *lengths, char **bufs,
+                     int count, __u32 *lengths, char **bufs,
                      struct ptlrpc_request_pool *pool)
 {
         struct ptlrpc_request *request;
@@ -664,7 +670,7 @@ ptlrpc_prep_req_pool(struct obd_import *imp,
 
 struct ptlrpc_request *
 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
-                int *lengths, char **bufs)
+                __u32 *lengths, char **bufs)
 {
         return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
                                     NULL);
@@ -684,7 +690,7 @@ struct ptlrpc_request_set *ptlrpc_prep_set(void)
         spin_lock_init(&set->set_new_req_lock);
         CFS_INIT_LIST_HEAD(&set->set_new_requests);
         CFS_INIT_LIST_HEAD(&set->set_cblist);
-        
+
         RETURN(set);
 }
 
@@ -720,14 +726,13 @@ void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
                 if (req->rq_phase == RQ_PHASE_NEW) {
 
                         if (req->rq_interpret_reply != NULL) {
-                                int (*interpreter)(struct ptlrpc_request *,
-                                                   void *, int) =
+                                ptlrpc_interpterer_t interpreter =
                                         req->rq_interpret_reply;
 
                                 /* higher level (i.e. LOV) failed;
                                  * let the sub reqs clean up */
                                 req->rq_status = -EBADR;
-                                interpreter(req, &req->rq_async_args,
+                                interpreter(NULL, req, &req->rq_async_args,
                                             req->rq_status);
                         }
                         set->set_remaining--;
@@ -770,16 +775,36 @@ void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
         atomic_inc(&req->rq_import->imp_inflight);
 }
 
-/* lock so many callers can add things, the context that owns the set
- * is supposed to notice these and move them into the set proper. */
-void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
-                            struct ptlrpc_request *req)
+/**
+ * Lock so many callers can add things, the context that owns the set
+ * is supposed to notice these and move them into the set proper.
+ */
+int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
+                           struct ptlrpc_request *req)
 {
+        struct ptlrpc_request_set *set = pc->pc_set;
+
+        /*
+         * Let caller know that we stopped and will not handle this request.
+         * It needs to take care itself of request.
+         */
+        if (test_bit(LIOD_STOP, &pc->pc_flags))
+                return -EALREADY;
+
         spin_lock(&set->set_new_req_lock);
-        /* The set takes over the caller's request reference */
+        /*
+         * The set takes over the caller's request reference.
+         */
         list_add_tail(&req->rq_set_chain, &set->set_new_requests);
         req->rq_set = set;
         spin_unlock(&set->set_new_req_lock);
+
+        /*
+         * Let thead know that we added something and better it to wake up
+         * and process.
+         */
+        cfs_waitq_signal(&set->set_waitq);
+        return 0;
 }
 
 /*
@@ -928,7 +953,7 @@ static int after_reply(struct ptlrpc_request *req)
 
         /*
          * NB Until this point, the whole of the incoming message,
-         * including buflens, status etc is in the sender's byte order. 
+         * including buflens, status etc is in the sender's byte order.
          */
 
         rc = sptlrpc_cli_unwrap_reply(req);
@@ -938,7 +963,7 @@ static int after_reply(struct ptlrpc_request *req)
         }
 
         /*
-         * Security layer unwrap might ask resend this request. 
+         * Security layer unwrap might ask resend this request.
          */
         if (req->rq_resend)
                 RETURN(0);
@@ -961,8 +986,9 @@ static int after_reply(struct ptlrpc_request *req)
         }
 
         OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
-        ptlrpc_at_adj_service(req);
-        ptlrpc_at_adj_net_latency(req);
+        ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
+        ptlrpc_at_adj_net_latency(req,
+                                  lustre_msg_get_service_time(req->rq_repmsg));
 
         rc = ptlrpc_check_status(req);
         imp->imp_connect_error = rc;
@@ -971,7 +997,7 @@ static int after_reply(struct ptlrpc_request *req)
                 /*
                  * Either we've been evicted, or the server has failed for
                  * some reason. Try to reconnect, and if that fails, punt to
-                 * the upcall. 
+                 * the upcall.
                  */
                 if (ll_rpc_recoverable_error(rc)) {
                         if (req->rq_send_state != LUSTRE_IMP_FULL ||
@@ -983,14 +1009,14 @@ static int after_reply(struct ptlrpc_request *req)
                 }
         } else {
                 /*
-                 * Let's look if server sent slv. Do it only for RPC with 
-                 * rc == 0. 
+                 * Let's look if server sent slv. Do it only for RPC with
+                 * rc == 0.
                  */
                 ldlm_cli_update_pool(req);
         }
 
         /*
-         * Store transno in reqmsg for replay. 
+         * Store transno in reqmsg for replay.
          */
         req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
         lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
@@ -999,10 +1025,10 @@ static int after_reply(struct ptlrpc_request *req)
                 spin_lock(&imp->imp_lock);
                 /*
                  * No point in adding already-committed requests to the replay
-                 * list, we will just remove them immediately. b=9829 
+                 * list, we will just remove them immediately. b=9829
                  */
-                if (req->rq_transno != 0 && 
-                    (req->rq_transno > 
+                if (req->rq_transno != 0 &&
+                    (req->rq_transno >
                      lustre_msg_get_last_committed(req->rq_repmsg) ||
                      req->rq_replay))
                         ptlrpc_retain_replayable_request(req, imp);
@@ -1013,7 +1039,7 @@ static int after_reply(struct ptlrpc_request *req)
                 }
 
                 /*
-                 * Replay-enabled imports return commit-status information. 
+                 * Replay-enabled imports return commit-status information.
                  */
                 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
                         imp->imp_peer_committed_transno =
@@ -1035,7 +1061,7 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req)
         LASSERT(req->rq_phase == RQ_PHASE_NEW);
         if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
                 RETURN (0);
-        
+
         req->rq_phase = RQ_PHASE_RPC;
 
         imp = req->rq_import;
@@ -1080,8 +1106,6 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req)
                         req->rq_status = rc;
                         RETURN(1);
                 } else {
-                        /* here begins timeout counting */
-                        req->rq_sent = cfs_time_current_sec();
                         req->rq_wait_ctx = 1;
                         RETURN(0);
                 }
@@ -1104,7 +1128,7 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req)
 }
 
 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
-int ptlrpc_check_set(struct ptlrpc_request_set *set)
+int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
 {
         struct list_head *tmp;
         int force_timer_recalc = 0;
@@ -1182,12 +1206,6 @@ int ptlrpc_check_set(struct ptlrpc_request_set *set)
                             req->rq_waiting || req->rq_wait_ctx) {
                                 int status;
 
-                                /* rq_wait_ctx is only touched in ptlrpcd,
-                                 * no lock needed here.
-                                 */
-                                if (req->rq_wait_ctx)
-                                        goto check_ctx;
-
                                 ptlrpc_unregister_reply(req);
 
                                 spin_lock(&imp->imp_lock);
@@ -1216,7 +1234,11 @@ int ptlrpc_check_set(struct ptlrpc_request_set *set)
                                 spin_unlock(&imp->imp_lock);
 
                                 req->rq_waiting = 0;
-                                if (req->rq_resend) {
+
+                                if (req->rq_timedout||req->rq_resend) {
+                                        /* This is re-sending anyways, 
+                                         * let's mark req as resend. */
+                                        req->rq_resend = 1;
                                         lustre_msg_add_flags(req->rq_reqmsg,
                                                              MSG_RESENT);
                                         if (req->rq_bulk) {
@@ -1232,21 +1254,21 @@ int ptlrpc_check_set(struct ptlrpc_request_set *set)
                                                        old_xid, req->rq_xid);
                                         }
                                 }
-check_ctx:
+                                /*
+                                 * rq_wait_ctx is only touched by ptlrpcd,
+                                 * so no lock is needed here.
+                                 */
                                 status = sptlrpc_req_refresh_ctx(req, -1);
                                 if (status) {
                                         if (req->rq_err) {
                                                 req->rq_status = status;
                                                 force_timer_recalc = 1;
-                                        }
-                                        if (!req->rq_wait_ctx) {
-                                                /* begins timeout counting */
-                                                req->rq_sent = cfs_time_current_sec();
+                                        } else {
                                                 req->rq_wait_ctx = 1;
                                         }
+
                                         continue;
                                 } else {
-                                        req->rq_sent = 0;
                                         req->rq_wait_ctx = 0;
                                 }
 
@@ -1336,9 +1358,10 @@ check_ctx:
                         ptlrpc_unregister_bulk (req);
 
                 if (req->rq_interpret_reply != NULL) {
-                        int (*interpreter)(struct ptlrpc_request *,void *,int) =
+                        ptlrpc_interpterer_t interpreter =
                                 req->rq_interpret_reply;
-                        req->rq_status = interpreter(req, &req->rq_async_args,
+                        req->rq_status = interpreter(NULL, req,
+                                                     &req->rq_async_args,
                                                      req->rq_status);
                 }
                 req->rq_phase = RQ_PHASE_COMPLETE;
@@ -1367,17 +1390,19 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req)
         ENTRY;
 
         DEBUG_REQ(D_ERROR|D_NETERROR, req,
-                  "%s (sent at %lu, "CFS_DURATION_T"s ago)",
+                  "%s (sent at "CFS_TIME_T", "CFS_DURATION_T"s ago)",
                   req->rq_net_err ? "network error" : "timeout",
-                  (long)req->rq_sent, cfs_time_current_sec() - req->rq_sent);
+                  req->rq_sent, cfs_time_sub(cfs_time_current_sec(),
+                  req->rq_sent));
 
         if (imp) {
-                LCONSOLE_WARN("Request x"LPU64" sent from %s to NID %s %lus ago"
-                              " has timed out (limit %lus).\n", req->rq_xid,
+                LCONSOLE_WARN("Request x"LPU64" sent from %s to NID %s "
+                              CFS_DURATION_T"s ago has timed out "
+                              "(limit "CFS_DURATION_T"s).\n", req->rq_xid,
                               req->rq_import->imp_obd->obd_name,
                               libcfs_nid2str(imp->imp_connection->c_peer.nid),
-                              cfs_time_current_sec() - req->rq_sent,
-                              req->rq_deadline - req->rq_sent);
+                              cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
+                              cfs_time_sub(req->rq_deadline, req->rq_sent));
         }
 
         if (imp != NULL && obd_debug_peer_on_timeout)
@@ -1385,7 +1410,6 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req)
 
         spin_lock(&req->rq_lock);
         req->rq_timedout = 1;
-        req->rq_wait_ctx = 0;
         spin_unlock(&req->rq_lock);
 
         ptlrpc_unregister_reply (req);
@@ -1419,7 +1443,7 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req)
                 spin_unlock(&req->rq_lock);
                 RETURN(1);
         }
-        
+
         /* if a request can't be resent we can't wait for an answer after
            the timeout */
         if (req->rq_no_resend) {
@@ -1517,6 +1541,9 @@ int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
                 if (req->rq_timedout)   /* already timed out */
                         continue;
 
+                if (req->rq_wait_ctx)   /* waiting for ctx */
+                        continue;
+
                 if (req->rq_phase == RQ_PHASE_NEW)
                         deadline = req->rq_sent;
                 else
@@ -1557,7 +1584,8 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set)
                 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
                                        ptlrpc_expired_set,
                                        ptlrpc_interrupted_set, set);
-                rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
+                rc = l_wait_event(set->set_waitq,
+                                  ptlrpc_check_set(NULL, set), &lwi);
 
                 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
 
@@ -1614,6 +1642,7 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
         LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
         LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
         LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
+        LASSERTF(!request->rq_replay, "req %p\n", request);
         LASSERT(request->rq_cli_ctx);
 
         req_capsule_fini(&request->rq_pill);
@@ -1661,11 +1690,6 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
         EXIT;
 }
 
-void ptlrpc_free_req(struct ptlrpc_request *request)
-{
-        __ptlrpc_free_req(request, 0);
-}
-
 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
 {
@@ -2067,8 +2091,8 @@ repeat:
                   "-- sleeping for "CFS_DURATION_T" ticks", timeout);
         lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
                                req);
-        rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
-        if (rc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
+        brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
+        if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
                                  ptlrpc_check_and_wait_suspend(req)))
                 goto repeat;
 
@@ -2093,7 +2117,7 @@ after_send:
         if (req->rq_err) {
                 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
                           rc, req->rq_status);
-                GOTO(out, rc = -EIO);
+                GOTO(out, rc = rc ? rc : -EIO);
         }
 
         if (req->rq_intr) {
@@ -2170,7 +2194,8 @@ struct ptlrpc_replay_async_args {
         int praa_old_status;
 };
 
-static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
+static int ptlrpc_replay_interpret(const struct lu_env *env,
+                                   struct ptlrpc_request *req,
                                     void * data, int rc)
 {
         struct ptlrpc_replay_async_args *aa = data;
@@ -2242,7 +2267,7 @@ int ptlrpc_replay_req(struct ptlrpc_request *req)
         LASSERT(req->rq_bulk == NULL);
 
         LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
-        aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
+        aa = ptlrpc_req_async_args(req);
         memset(aa, 0, sizeof *aa);
 
         /* Prepare request to be resent with ptlrpcd */
@@ -2320,8 +2345,60 @@ void ptlrpc_abort_inflight(struct obd_import *imp)
         EXIT;
 }
 
-static __u64 ptlrpc_last_xid = 0;
-spinlock_t ptlrpc_last_xid_lock;
+void ptlrpc_abort_set(struct ptlrpc_request_set *set)
+{
+        struct list_head *tmp, *n;
+
+        LASSERT(set != NULL);
+
+        list_for_each_safe(tmp, n, &set->set_requests) {
+                struct ptlrpc_request *req =
+                        list_entry(tmp, struct ptlrpc_request, rq_set_chain);
+
+                spin_lock (&req->rq_lock);
+                if (req->rq_phase != RQ_PHASE_RPC) {
+                        spin_unlock (&req->rq_lock);
+                        continue;
+                }
+
+                req->rq_err = 1;
+                req->rq_status = -EINTR;
+                ptlrpc_wake_client_req(req);
+                spin_unlock (&req->rq_lock);
+        }
+}
+
+static __u64 ptlrpc_last_xid;
+static spinlock_t ptlrpc_last_xid_lock;
+
+/* Initialize the XID for the node.  This is common among all requests on
+ * this node, and only requires the property that it is monotonically
+ * increasing.  It does not need to be sequential.  Since this is also used
+ * as the RDMA match bits, it is important that a single client NOT have
+ * the same match bits for two different in-flight requests, hence we do
+ * NOT want to have an XID per target or similar.
+ *
+ * To avoid an unlikely collision between match bits after a client reboot
+ * (which would cause old to be delivered into the wrong buffer) we initialize
+ * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
+ * If the time is clearly incorrect, we instead use a 62-bit random number.
+ * In the worst case the random number will overflow 1M RPCs per second in
+ * 9133 years, or permutations thereof.
+ */
+#define YEAR_2004 (1ULL << 30)
+void ptlrpc_init_xid(void)
+{
+        time_t now = cfs_time_current_sec();
+
+        spin_lock_init(&ptlrpc_last_xid_lock);
+        if (now < YEAR_2004) {
+                ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
+                ptlrpc_last_xid >>= 2;
+                ptlrpc_last_xid |= (1ULL << 61);
+        } else {
+                ptlrpc_last_xid = (now << 20);
+        }
+}
 
 __u64 ptlrpc_next_xid(void)
 {
@@ -2334,11 +2411,16 @@ __u64 ptlrpc_next_xid(void)
 
 __u64 ptlrpc_sample_next_xid(void)
 {
+#if BITS_PER_LONG == 32
+        /* need to avoid possible word tearing on 32-bit systems */
         __u64 tmp;
         spin_lock(&ptlrpc_last_xid_lock);
         tmp = ptlrpc_last_xid + 1;
         spin_unlock(&ptlrpc_last_xid_lock);
         return tmp;
+#else
+        /* No need to lock, since returned value is racy anyways */
+        return ptlrpc_last_xid + 1;
+#endif
 }
 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
-