Whamcloud - gitweb
LU-6496 ptlrpc: Fix wrong code indentation in plain_authorize
[fs/lustre-release.git] / lustre / ptlrpc / client.c
index 3f3e220..091cbcb 100644 (file)
@@ -27,7 +27,7 @@
  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  *
- * Copyright (c) 2011, 2013, Intel Corporation.
+ * Copyright (c) 2011, 2014, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
 /** Implementation of client-side PortalRPC interfaces */
 
 #define DEBUG_SUBSYSTEM S_RPC
-#ifndef __KERNEL__
-#include <errno.h>
-#include <signal.h>
-#include <liblustre.h>
-#endif
 
 #include <obd_support.h>
 #include <obd_class.h>
@@ -54,6 +49,7 @@
 
 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
 static int ptlrpcd_check_work(struct ptlrpc_request *req);
+static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
 
 /**
  * Initialize passed in client structure \a cl.
@@ -96,7 +92,6 @@ struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
 
         return c;
 }
-EXPORT_SYMBOL(ptlrpc_uuid_to_connection);
 
 /**
  * Allocate and initialize new bulk descriptor on the sender.
@@ -286,22 +281,35 @@ 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,
-                                      unsigned int service_time)
+void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
+                              unsigned int service_time)
 {
         unsigned int nl, oldnl;
         struct imp_at *at;
         time_t now = cfs_time_current_sec();
 
         LASSERT(req->rq_import);
-        at = &req->rq_import->imp_at;
+
+       if (service_time > now - req->rq_sent + 3) {
+               /* bz16408, however, this can also happen if early reply
+                * is lost and client RPC is expired and resent, early reply
+                * or reply of original RPC can still be fit in reply buffer
+                * of resent RPC, now client is measuring time from the
+                * resent time, but server sent back service time of original
+                * RPC.
+                */
+               CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
+                      D_ADAPTTO : D_WARNING,
+                      "Reported service time %u > total measured time "
+                      CFS_DURATION_T"\n", service_time,
+                      cfs_time_sub(now, req->rq_sent));
+               return;
+       }
 
         /* Network latency is total time less server processing time */
-        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));
+       nl = max_t(int, now - req->rq_sent -
+                       service_time, 0) + 1; /* st rounding */
+       at = &req->rq_import->imp_at;
 
         oldnl = at_measured(&at->iat_net_latency, nl);
         if (oldnl != 0)
@@ -338,6 +346,7 @@ static int unpack_reply(struct ptlrpc_request *req)
  * If anything goes wrong just ignore it - same as if it never happened
  */
 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
+__must_hold(&req->rq_lock)
 {
         struct ptlrpc_request *early_req;
         time_t                 olddl;
@@ -350,27 +359,29 @@ static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
        rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
        if (rc) {
                spin_lock(&req->rq_lock);
-                RETURN(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));
-        }
-
-        sptlrpc_cli_finish_early_reply(early_req);
+               RETURN(rc);
+       }
 
+       rc = unpack_reply(early_req);
        if (rc != 0) {
+               sptlrpc_cli_finish_early_reply(early_req);
                spin_lock(&req->rq_lock);
                RETURN(rc);
        }
 
-       /* Adjust the local timeout for this req */
-       ptlrpc_at_set_req_timeout(req);
+       /* Use new timeout value just to adjust the local value for this
+        * request, don't include it into at_history. It is unclear yet why
+        * service time increased and should it be counted or skipped, e.g.
+        * that can be recovery case or some error or server, the real reply
+        * will add all new data if it is worth to add. */
+       req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
+       lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
+
+       /* Network latency can be adjusted, it is pure network delays */
+       ptlrpc_at_adj_net_latency(req,
+                       lustre_msg_get_service_time(early_req->rq_repmsg));
+
+       sptlrpc_cli_finish_early_reply(early_req);
 
        spin_lock(&req->rq_lock);
        olddl = req->rq_deadline;
@@ -390,7 +401,7 @@ static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
        RETURN(rc);
 }
 
-struct kmem_cache *request_cache;
+static struct kmem_cache *request_cache;
 
 int ptlrpc_request_cache_init(void)
 {
@@ -405,7 +416,7 @@ void ptlrpc_request_cache_fini(void)
        kmem_cache_destroy(request_cache);
 }
 
-struct ptlrpc_request *ptlrpc_request_cache_alloc(int flags)
+struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
 {
        struct ptlrpc_request *req;
 
@@ -424,15 +435,15 @@ void ptlrpc_request_cache_free(struct ptlrpc_request *req)
  */
 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
 {
-       cfs_list_t *l, *tmp;
+       struct list_head *l, *tmp;
        struct ptlrpc_request *req;
 
        LASSERT(pool != NULL);
 
        spin_lock(&pool->prp_lock);
-       cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
-               req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
-               cfs_list_del(&req->rq_list);
+       list_for_each_safe(l, tmp, &pool->prp_req_list) {
+               req = list_entry(l, struct ptlrpc_request, rq_list);
+               list_del(&req->rq_list);
                LASSERT(req->rq_reqbuf);
                LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
                OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
@@ -454,7 +465,7 @@ void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
         while (size < pool->prp_rq_size)
                 size <<= 1;
 
-        LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
+       LASSERTF(list_empty(&pool->prp_req_list) ||
                  size == pool->prp_rq_size,
                  "Trying to change pool size with nonempty pool "
                  "from %d to %d bytes\n", pool->prp_rq_size, size);
@@ -478,7 +489,7 @@ void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
                 req->rq_reqbuf_len = size;
                 req->rq_pool = pool;
                spin_lock(&pool->prp_lock);
-               cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
+               list_add_tail(&req->rq_list, &pool->prp_req_list);
        }
        spin_unlock(&pool->prp_lock);
        return;
@@ -495,30 +506,30 @@ EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
  */
 struct ptlrpc_request_pool *
 ptlrpc_init_rq_pool(int num_rq, int msgsize,
-                    void (*populate_pool)(struct ptlrpc_request_pool *, int))
+                   void (*populate_pool)(struct ptlrpc_request_pool *, int))
 {
-        struct ptlrpc_request_pool *pool;
+       struct ptlrpc_request_pool *pool;
 
-        OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
-        if (!pool)
-                return NULL;
+       OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
+       if (!pool)
+               return NULL;
 
-        /* Request next power of two for the allocation, because internally
-           kernel would do exactly this */
+       /* Request next power of two for the allocation, because internally
+          kernel would do exactly this */
 
        spin_lock_init(&pool->prp_lock);
-        CFS_INIT_LIST_HEAD(&pool->prp_req_list);
-        pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
-        pool->prp_populate = populate_pool;
+       INIT_LIST_HEAD(&pool->prp_req_list);
+       pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
+       pool->prp_populate = populate_pool;
 
-        populate_pool(pool, num_rq);
+       populate_pool(pool, num_rq);
 
-        if (cfs_list_empty(&pool->prp_req_list)) {
-                /* have not allocated a single request for the pool */
-                OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
-                pool = NULL;
-        }
-        return pool;
+       if (list_empty(&pool->prp_req_list)) {
+               /* have not allocated a single request for the pool */
+               OBD_FREE(pool, sizeof(struct ptlrpc_request_pool));
+               pool = NULL;
+       }
+       return pool;
 }
 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
 
@@ -540,14 +551,14 @@ ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
         * in writeout path, where this matters, this is safe to do, because
         * nothing is lost in this case, and when some in-flight requests
         * complete, this code will be called again. */
-       if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
+       if (unlikely(list_empty(&pool->prp_req_list))) {
                spin_unlock(&pool->prp_lock);
                return NULL;
        }
 
-       request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
-                                rq_list);
-       cfs_list_del_init(&request->rq_list);
+       request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
+                            rq_list);
+       list_del_init(&request->rq_list);
        spin_unlock(&pool->prp_lock);
 
         LASSERT(request->rq_reqbuf);
@@ -570,9 +581,9 @@ static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
        struct ptlrpc_request_pool *pool = request->rq_pool;
 
        spin_lock(&pool->prp_lock);
-       LASSERT(cfs_list_empty(&request->rq_list));
+       LASSERT(list_empty(&request->rq_list));
        LASSERT(!request->rq_receiving_reply);
-       cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
+       list_add_tail(&request->rq_list, &pool->prp_req_list);
        spin_unlock(&pool->prp_lock);
 }
 
@@ -605,7 +616,6 @@ static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
         lustre_msg_add_version(request->rq_reqmsg, version);
         request->rq_send_state = LUSTRE_IMP_FULL;
         request->rq_type = PTL_RPC_MSG_REQUEST;
-        request->rq_export = NULL;
 
         request->rq_req_cbid.cbid_fn  = request_out_callback;
         request->rq_req_cbid.cbid_arg = request;
@@ -622,19 +632,7 @@ static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
 
         ptlrpc_at_set_req_timeout(request);
 
-       spin_lock_init(&request->rq_lock);
-       CFS_INIT_LIST_HEAD(&request->rq_list);
-       CFS_INIT_LIST_HEAD(&request->rq_timed_list);
-       CFS_INIT_LIST_HEAD(&request->rq_replay_list);
-       CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
-       CFS_INIT_LIST_HEAD(&request->rq_set_chain);
-       CFS_INIT_LIST_HEAD(&request->rq_history_list);
-       CFS_INIT_LIST_HEAD(&request->rq_exp_list);
-       init_waitqueue_head(&request->rq_reply_waitq);
-       init_waitqueue_head(&request->rq_set_waitq);
        request->rq_xid = ptlrpc_next_xid();
-       atomic_set(&request->rq_refcount, 1);
-
        lustre_msg_set_opc(request->rq_reqmsg, opcode);
 
        RETURN(0);
@@ -711,9 +709,11 @@ struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
                request = ptlrpc_request_cache_alloc(GFP_NOFS);
 
        if (request) {
+               ptlrpc_cli_req_init(request);
+
                LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
                LASSERT(imp != LP_POISON);
-               LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
+               LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
                        imp->imp_client);
                LASSERT(imp->imp_client != LP_POISON);
 
@@ -836,7 +836,6 @@ ptlrpc_prep_req_pool(struct obd_import *imp,
         }
         return request;
 }
-EXPORT_SYMBOL(ptlrpc_prep_req_pool);
 
 /**
  * Same as ptlrpc_prep_req_pool, but without pool
@@ -848,7 +847,6 @@ ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
         return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
                                     NULL);
 }
-EXPORT_SYMBOL(ptlrpc_prep_req);
 
 /**
  * Allocate and initialize new request set structure.
@@ -863,13 +861,13 @@ struct ptlrpc_request_set *ptlrpc_prep_set(void)
        if (!set)
                RETURN(NULL);
        atomic_set(&set->set_refcount, 1);
-       CFS_INIT_LIST_HEAD(&set->set_requests);
+       INIT_LIST_HEAD(&set->set_requests);
        init_waitqueue_head(&set->set_waitq);
        atomic_set(&set->set_new_count, 0);
        atomic_set(&set->set_remaining, 0);
        spin_lock_init(&set->set_new_req_lock);
-       CFS_INIT_LIST_HEAD(&set->set_new_requests);
-       CFS_INIT_LIST_HEAD(&set->set_cblist);
+       INIT_LIST_HEAD(&set->set_new_requests);
+       INIT_LIST_HEAD(&set->set_cblist);
        set->set_max_inflight = UINT_MAX;
        set->set_producer     = NULL;
        set->set_producer_arg = NULL;
@@ -903,7 +901,6 @@ struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
 
        RETURN(set);
 }
-EXPORT_SYMBOL(ptlrpc_prep_fcset);
 
 /**
  * Wind down and free request set structure previously allocated with
@@ -915,33 +912,33 @@ EXPORT_SYMBOL(ptlrpc_prep_fcset);
  */
 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
 {
-        cfs_list_t       *tmp;
-        cfs_list_t       *next;
-        int               expected_phase;
-        int               n = 0;
-        ENTRY;
+       struct list_head        *tmp;
+       struct list_head        *next;
+       int                      expected_phase;
+       int                      n = 0;
+       ENTRY;
 
-        /* Requests on the set should either all be completed, or all be new */
+       /* Requests on the set should either all be completed, or all be new */
        expected_phase = (atomic_read(&set->set_remaining) == 0) ?
-                         RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
-        cfs_list_for_each (tmp, &set->set_requests) {
-                struct ptlrpc_request *req =
-                        cfs_list_entry(tmp, struct ptlrpc_request,
-                                       rq_set_chain);
-
-                LASSERT(req->rq_phase == expected_phase);
-                n++;
-        }
+                        RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
+       list_for_each(tmp, &set->set_requests) {
+               struct ptlrpc_request *req =
+                       list_entry(tmp, struct ptlrpc_request,
+                                  rq_set_chain);
+
+               LASSERT(req->rq_phase == expected_phase);
+               n++;
+       }
 
        LASSERTF(atomic_read(&set->set_remaining) == 0 ||
                 atomic_read(&set->set_remaining) == n, "%d / %d\n",
                 atomic_read(&set->set_remaining), n);
 
-        cfs_list_for_each_safe(tmp, next, &set->set_requests) {
-                struct ptlrpc_request *req =
-                        cfs_list_entry(tmp, struct ptlrpc_request,
-                                       rq_set_chain);
-                cfs_list_del_init(&req->rq_set_chain);
+       list_for_each_safe(tmp, next, &set->set_requests) {
+               struct ptlrpc_request *req =
+                       list_entry(tmp, struct ptlrpc_request,
+                                  rq_set_chain);
+               list_del_init(&req->rq_set_chain);
 
                LASSERT(req->rq_phase == expected_phase);
 
@@ -973,19 +970,18 @@ EXPORT_SYMBOL(ptlrpc_set_destroy);
 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
                       set_interpreter_func fn, void *data)
 {
-        struct ptlrpc_set_cbdata *cbdata;
+       struct ptlrpc_set_cbdata *cbdata;
 
-        OBD_ALLOC_PTR(cbdata);
-        if (cbdata == NULL)
-                RETURN(-ENOMEM);
+       OBD_ALLOC_PTR(cbdata);
+       if (cbdata == NULL)
+               RETURN(-ENOMEM);
 
-        cbdata->psc_interpret = fn;
-        cbdata->psc_data = data;
-        cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
+       cbdata->psc_interpret = fn;
+       cbdata->psc_data = data;
+       list_add_tail(&cbdata->psc_item, &set->set_cblist);
 
-        RETURN(0);
+       RETURN(0);
 }
-EXPORT_SYMBOL(ptlrpc_set_add_cb);
 
 /**
  * Add a new request to the general purpose request set.
@@ -994,10 +990,10 @@ EXPORT_SYMBOL(ptlrpc_set_add_cb);
 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
                         struct ptlrpc_request *req)
 {
-       LASSERT(cfs_list_empty(&req->rq_set_chain));
+       LASSERT(list_empty(&req->rq_set_chain));
 
        /* The set takes over the caller's request reference */
-       cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
+       list_add_tail(&req->rq_set_chain, &set->set_requests);
        req->rq_set = set;
        atomic_inc(&set->set_remaining);
        req->rq_queued_time = cfs_time_current();
@@ -1032,7 +1028,7 @@ void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
         */
        req->rq_set = set;
        req->rq_queued_time = cfs_time_current();
-       cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
+       list_add_tail(&req->rq_set_chain, &set->set_new_requests);
        count = atomic_inc_return(&set->set_new_count);
        spin_unlock(&set->set_new_req_lock);
 
@@ -1047,7 +1043,6 @@ void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
                        wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
        }
 }
-EXPORT_SYMBOL(ptlrpc_set_add_new_req);
 
 /**
  * Based on the current state of the import, determine if the request
@@ -1079,9 +1074,9 @@ static int ptlrpc_import_delay_req(struct obd_import *imp,
                          D_HA : D_ERROR, req, "IMP_CLOSED ");
                *status = -EIO;
         } else if (ptlrpc_send_limit_expired(req)) {
-                /* probably doesn't need to be a D_ERROR after initial testing */
-                DEBUG_REQ(D_ERROR, req, "send limit expired ");
-                *status = -EIO;
+               /* probably doesn't need to be a D_ERROR after initial testing*/
+               DEBUG_REQ(D_HA, req, "send limit expired ");
+               *status = -ETIMEDOUT;
        } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
                   imp->imp_state == LUSTRE_IMP_CONNECTING) {
                /* allow CONNECT even if import is invalid */ ;
@@ -1118,35 +1113,40 @@ static int ptlrpc_import_delay_req(struct obd_import *imp,
 }
 
 /**
- * Decide if the eror message regarding provided request \a req
- * should be printed to the console or not.
- * Makes it's decision on request status and other properties.
- * Returns 1 to print error on the system console or 0 if not.
+ * Decide if the error message should be printed to the console or not.
+ * Makes its decision based on request type, status, and failure frequency.
+ *
+ * \param[in] req  request that failed and may need a console message
+ *
+ * \retval false if no message should be printed
+ * \retval true  if console message should be printed
  */
-static int ptlrpc_console_allow(struct ptlrpc_request *req)
+static bool ptlrpc_console_allow(struct ptlrpc_request *req)
 {
-        __u32 opc;
-        int err;
+       __u32 opc;
 
-        LASSERT(req->rq_reqmsg != NULL);
-        opc = lustre_msg_get_opc(req->rq_reqmsg);
+       LASSERT(req->rq_reqmsg != NULL);
+       opc = lustre_msg_get_opc(req->rq_reqmsg);
 
-        /* Suppress particular reconnect errors which are to be expected.  No
-         * errors are suppressed for the initial connection on an import */
-        if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
-            (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
+       /* Suppress particular reconnect errors which are to be expected. */
+       if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
+               int err;
 
-                /* Suppress timed out reconnect requests */
-                if (req->rq_timedout)
-                        return 0;
+               /* Suppress timed out reconnect requests */
+               if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
+                   req->rq_timedout)
+                       return false;
 
-                /* Suppress unavailable/again reconnect requests */
-                err = lustre_msg_get_status(req->rq_repmsg);
-                if (err == -ENODEV || err == -EAGAIN)
-                        return 0;
-        }
+               /* Suppress most unavailable/again reconnect requests, but
+                * print occasionally so it is clear client is trying to
+                * connect to a server where no target is running. */
+               err = lustre_msg_get_status(req->rq_repmsg);
+               if ((err == -ENODEV || err == -EAGAIN) &&
+                   req->rq_import->imp_conn_cnt % 30 != 20)
+                       return false;
+       }
 
-        return 1;
+       return true;
 }
 
 /**
@@ -1161,14 +1161,15 @@ static int ptlrpc_check_status(struct ptlrpc_request *req)
         err = lustre_msg_get_status(req->rq_repmsg);
        if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
                struct obd_import *imp = req->rq_import;
+               lnet_nid_t nid = imp->imp_connection->c_peer.nid;
                __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
+
                if (ptlrpc_console_allow(req))
-                       LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s,"
-                                          " operation %s failed with %d.\n",
-                                          imp->imp_obd->obd_name,
-                                          libcfs_nid2str(
-                                          imp->imp_connection->c_peer.nid),
-                                          ll_opcode2str(opc), err);
+                       LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
+                                          "failed: rc = %d\n",
+                                          imp->imp_obd->obd_name,
+                                          ll_opcode2str(opc),
+                                          libcfs_nid2str(nid), err);
                RETURN(err < 0 ? err : -EINVAL);
        }
 
@@ -1223,9 +1224,9 @@ static int after_reply(struct ptlrpc_request *req)
 
         LASSERT(obd != NULL);
         /* repbuf must be unlinked */
-        LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
+       LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
 
-        if (req->rq_reply_truncate) {
+       if (req->rq_reply_truncated) {
                 if (ptlrpc_no_resend(req)) {
                         DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
                                   " expected: %d, actual size: %d",
@@ -1246,6 +1247,9 @@ static int after_reply(struct ptlrpc_request *req)
                 RETURN(0);
         }
 
+       do_gettimeofday(&work_start);
+       timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
+
         /*
          * NB Until this point, the whole of the incoming message,
          * including buflens, status etc is in the sender's byte order.
@@ -1272,7 +1276,9 @@ static int after_reply(struct ptlrpc_request *req)
                time_t  now = cfs_time_current_sec();
 
                DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
+               spin_lock(&req->rq_lock);
                req->rq_resend = 1;
+               spin_unlock(&req->rq_lock);
                req->rq_nr_resend++;
 
                /* allocate new xid to avoid reply reconstruction */
@@ -1298,8 +1304,6 @@ static int after_reply(struct ptlrpc_request *req)
                RETURN(0);
        }
 
-       do_gettimeofday(&work_start);
-       timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
        if (obd->obd_svc_stats != NULL) {
                lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
                                    timediff);
@@ -1322,20 +1326,20 @@ static int after_reply(struct ptlrpc_request *req)
         rc = ptlrpc_check_status(req);
         imp->imp_connect_error = rc;
 
-        if (rc) {
-                /*
-                 * Either we've been evicted, or the server has failed for
-                 * some reason. Try to reconnect, and if that fails, punt to
-                 * the upcall.
-                 */
-                if (ll_rpc_recoverable_error(rc)) {
-                        if (req->rq_send_state != LUSTRE_IMP_FULL ||
-                            imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
-                                RETURN(rc);
-                        }
-                        ptlrpc_request_handle_notconn(req);
-                        RETURN(rc);
-                }
+       if (rc) {
+               /*
+                * Either we've been evicted, or the server has failed for
+                * some reason. Try to reconnect, and if that fails, punt to
+                * the upcall.
+                */
+               if (ptlrpc_recoverable_error(rc)) {
+                       if (req->rq_send_state != LUSTRE_IMP_FULL ||
+                           imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
+                               RETURN(rc);
+                       }
+                       ptlrpc_request_handle_notconn(req);
+                       RETURN(rc);
+               }
         } else {
                 /*
                  * Let's look if server sent slv. Do it only for RPC with
@@ -1385,12 +1389,12 @@ static int after_reply(struct ptlrpc_request *req)
 
                ptlrpc_free_committed(imp);
 
-               if (!cfs_list_empty(&imp->imp_replay_list)) {
+               if (!list_empty(&imp->imp_replay_list)) {
                        struct ptlrpc_request *last;
 
-                       last = cfs_list_entry(imp->imp_replay_list.prev,
-                                             struct ptlrpc_request,
-                                             rq_replay_list);
+                       last = list_entry(imp->imp_replay_list.prev,
+                                         struct ptlrpc_request,
+                                         rq_replay_list);
                        /*
                         * Requests with rq_replay stay on the list even if no
                         * commit is expected.
@@ -1438,8 +1442,8 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req)
                          "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
                          ptlrpc_import_state_name(req->rq_send_state),
                          ptlrpc_import_state_name(imp->imp_state));
-               LASSERT(cfs_list_empty(&req->rq_list));
-               cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
+               LASSERT(list_empty(&req->rq_list));
+               list_add_tail(&req->rq_list, &imp->imp_delayed_list);
                atomic_inc(&req->rq_import->imp_inflight);
                spin_unlock(&imp->imp_lock);
                RETURN(0);
@@ -1452,8 +1456,8 @@ static int ptlrpc_send_new_req(struct ptlrpc_request *req)
                RETURN(rc);
        }
 
-       LASSERT(cfs_list_empty(&req->rq_list));
-       cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
+       LASSERT(list_empty(&req->rq_list));
+       list_add_tail(&req->rq_list, &imp->imp_sending_list);
        atomic_inc(&req->rq_import->imp_inflight);
        spin_unlock(&imp->imp_lock);
 
@@ -1519,23 +1523,35 @@ static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
  * and no more replies are expected.
  * (it is possible to get less replies than requests sent e.g. due to timed out
  * requests or requests that we had trouble to send out)
+ *
+ * NOTE: This function contains a potential schedule point (cond_resched()).
  */
 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
 {
-        cfs_list_t *tmp, *next;
-        int force_timer_recalc = 0;
-        ENTRY;
+       struct list_head *tmp, *next;
+       struct list_head  comp_reqs;
+       int force_timer_recalc = 0;
+       ENTRY;
 
        if (atomic_read(&set->set_remaining) == 0)
-                RETURN(1);
+               RETURN(1);
+
+       INIT_LIST_HEAD(&comp_reqs);
+       list_for_each_safe(tmp, next, &set->set_requests) {
+               struct ptlrpc_request *req =
+                       list_entry(tmp, struct ptlrpc_request,
+                                  rq_set_chain);
+               struct obd_import *imp = req->rq_import;
+               int unregistered = 0;
+               int rc = 0;
 
-        cfs_list_for_each_safe(tmp, next, &set->set_requests) {
-                struct ptlrpc_request *req =
-                        cfs_list_entry(tmp, struct ptlrpc_request,
-                                       rq_set_chain);
-                struct obd_import *imp = req->rq_import;
-                int unregistered = 0;
-                int rc = 0;
+               /* This schedule point is mainly for the ptlrpcd caller of this
+                * function.  Most ptlrpc sets are not long-lived and unbounded
+                * in length, but at the least the set used by the ptlrpcd is.
+                * Since the processing time is unbounded, we need to insert an
+                * explicit schedule point to make the thread well-behaved.
+                */
+               cond_resched();
 
                 if (req->rq_phase == RQ_PHASE_NEW &&
                     ptlrpc_send_new_req(req)) {
@@ -1595,8 +1611,10 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
                         ptlrpc_rqphase_move(req, req->rq_next_phase);
                 }
 
-                if (req->rq_phase == RQ_PHASE_COMPLETE)
+                if (req->rq_phase == RQ_PHASE_COMPLETE) {
+                       list_move_tail(&req->rq_set_chain, &comp_reqs);
                         continue;
+               }
 
                 if (req->rq_phase == RQ_PHASE_INTERPRET)
                         GOTO(interpret, req->rq_status);
@@ -1636,7 +1654,7 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
 
                 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
                  * so it sets rq_intr regardless of individual rpc
-                 * timeouts. The synchronous IO waiting path sets 
+                * timeouts. The synchronous IO waiting path sets
                  * rq_intr irrespective of whether ptlrpcd
                  * has seen a timeout.  Our policy is to only interpret
                  * interrupted rpcs after they have timed out, so we
@@ -1655,15 +1673,17 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
                             req->rq_waiting || req->rq_wait_ctx) {
                                 int status;
 
-                                if (!ptlrpc_unregister_reply(req, 1))
-                                        continue;
+                               if (!ptlrpc_unregister_reply(req, 1)) {
+                                       ptlrpc_unregister_bulk(req, 1);
+                                       continue;
+                               }
 
                                spin_lock(&imp->imp_lock);
                                if (ptlrpc_import_delay_req(imp, req, &status)){
                                        /* put on delay list - only if we wait
                                         * recovery finished - before send */
-                                       cfs_list_del_init(&req->rq_list);
-                                       cfs_list_add_tail(&req->rq_list,
+                                       list_del_init(&req->rq_list);
+                                       list_add_tail(&req->rq_list,
                                                          &imp->
                                                          imp_delayed_list);
                                        spin_unlock(&imp->imp_lock);
@@ -1686,8 +1706,8 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
                                        GOTO(interpret, req->rq_status);
                                }
 
-                               cfs_list_del_init(&req->rq_list);
-                               cfs_list_add_tail(&req->rq_list,
+                               list_del_init(&req->rq_list);
+                               list_add_tail(&req->rq_list,
                                                  &imp->imp_sending_list);
 
                                spin_unlock(&imp->imp_lock);
@@ -1855,8 +1875,8 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
                 * may happen in the case of marking it erroneous for the case
                 * ptlrpc_import_delay_req(req, status) find it impossible to
                 * allow sending this rpc and returns *status != 0. */
-               if (!cfs_list_empty(&req->rq_list)) {
-                       cfs_list_del_init(&req->rq_list);
+               if (!list_empty(&req->rq_list)) {
+                       list_del_init(&req->rq_list);
                        atomic_dec(&imp->imp_inflight);
                }
                spin_unlock(&imp->imp_lock);
@@ -1871,7 +1891,7 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
 
                        /* free the request that has just been completed
                         * in order not to pollute set->set_requests */
-                       cfs_list_del_init(&req->rq_set_chain);
+                       list_del_init(&req->rq_set_chain);
                        spin_lock(&req->rq_lock);
                        req->rq_set = NULL;
                        req->rq_invalid_rqset = 0;
@@ -1881,9 +1901,15 @@ int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
                        if (req->rq_status != 0)
                                set->set_rc = req->rq_status;
                        ptlrpc_req_finished(req);
+               } else {
+                       list_move_tail(&req->rq_set_chain, &comp_reqs);
                }
        }
 
+       /* move completed request at the head of list so it's easier for
+        * caller to find them */
+       list_splice(&comp_reqs, &set->set_requests);
+
        /* If we hit an error, we want to recover promptly. */
        RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
 }
@@ -1967,20 +1993,20 @@ int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
  */
 int ptlrpc_expired_set(void *data)
 {
-        struct ptlrpc_request_set *set = data;
-        cfs_list_t                *tmp;
-        time_t                     now = cfs_time_current_sec();
-        ENTRY;
+       struct ptlrpc_request_set       *set = data;
+       struct list_head                *tmp;
+       time_t                          now = cfs_time_current_sec();
+       ENTRY;
 
-        LASSERT(set != NULL);
+       LASSERT(set != NULL);
 
-        /*
-         * A timeout expired. See which reqs it applies to...
-         */
-        cfs_list_for_each (tmp, &set->set_requests) {
-                struct ptlrpc_request *req =
-                        cfs_list_entry(tmp, struct ptlrpc_request,
-                                       rq_set_chain);
+       /*
+        * A timeout expired. See which reqs it applies to...
+        */
+       list_for_each(tmp, &set->set_requests) {
+               struct ptlrpc_request *req =
+                       list_entry(tmp, struct ptlrpc_request,
+                                  rq_set_chain);
 
                 /* don't expire request waiting for context */
                 if (req->rq_wait_ctx)
@@ -2008,7 +2034,6 @@ int ptlrpc_expired_set(void *data)
          */
         RETURN(1);
 }
-EXPORT_SYMBOL(ptlrpc_expired_set);
 
 /**
  * Sets rq_intr flag in \a req under spinlock.
@@ -2025,42 +2050,40 @@ EXPORT_SYMBOL(ptlrpc_mark_interrupted);
  * Interrupts (sets interrupted flag) all uncompleted requests in
  * a set \a data. Callback for l_wait_event for interruptible waits.
  */
-void ptlrpc_interrupted_set(void *data)
+static void ptlrpc_interrupted_set(void *data)
 {
-        struct ptlrpc_request_set *set = data;
-        cfs_list_t *tmp;
+       struct ptlrpc_request_set *set = data;
+       struct list_head *tmp;
 
-        LASSERT(set != NULL);
-        CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
+       LASSERT(set != NULL);
+       CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
 
-        cfs_list_for_each(tmp, &set->set_requests) {
-                struct ptlrpc_request *req =
-                        cfs_list_entry(tmp, struct ptlrpc_request,
-                                       rq_set_chain);
+       list_for_each(tmp, &set->set_requests) {
+               struct ptlrpc_request *req =
+                       list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 
-                if (req->rq_phase != RQ_PHASE_RPC &&
-                    req->rq_phase != RQ_PHASE_UNREGISTERING)
-                        continue;
+               if (req->rq_phase != RQ_PHASE_RPC &&
+                   req->rq_phase != RQ_PHASE_UNREGISTERING)
+                       continue;
 
-                ptlrpc_mark_interrupted(req);
-        }
+               ptlrpc_mark_interrupted(req);
+       }
 }
-EXPORT_SYMBOL(ptlrpc_interrupted_set);
 
 /**
  * Get the smallest timeout in the set; this does NOT set a timeout.
  */
 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
 {
-        cfs_list_t            *tmp;
-        time_t                 now = cfs_time_current_sec();
-        int                    timeout = 0;
-        struct ptlrpc_request *req;
-        int                    deadline;
-        ENTRY;
+       struct list_head        *tmp;
+       time_t                   now = cfs_time_current_sec();
+       int                      timeout = 0;
+       struct ptlrpc_request   *req;
+       int                      deadline;
+       ENTRY;
 
-        cfs_list_for_each(tmp, &set->set_requests) {
-                req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
+       list_for_each(tmp, &set->set_requests) {
+               req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 
                 /*
                  * Request in-flight?
@@ -2096,7 +2119,6 @@ int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
         }
         RETURN(timeout);
 }
-EXPORT_SYMBOL(ptlrpc_set_next_timeout);
 
 /**
  * Send all unset request from the set and then wait untill all
@@ -2106,7 +2128,7 @@ EXPORT_SYMBOL(ptlrpc_set_next_timeout);
  */
 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
 {
-        cfs_list_t            *tmp;
+       struct list_head            *tmp;
         struct ptlrpc_request *req;
         struct l_wait_info     lwi;
         int                    rc, timeout;
@@ -2115,14 +2137,14 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set)
        if (set->set_producer)
                (void)ptlrpc_set_producer(set);
        else
-               cfs_list_for_each(tmp, &set->set_requests) {
-                       req = cfs_list_entry(tmp, struct ptlrpc_request,
-                                            rq_set_chain);
+               list_for_each(tmp, &set->set_requests) {
+                       req = list_entry(tmp, struct ptlrpc_request,
+                                        rq_set_chain);
                        if (req->rq_phase == RQ_PHASE_NEW)
                                (void)ptlrpc_send_new_req(req);
                }
 
-        if (cfs_list_empty(&set->set_requests))
+       if (list_empty(&set->set_requests))
                 RETURN(0);
 
         do {
@@ -2133,21 +2155,21 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set)
                 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
                        set, timeout);
 
-                if (timeout == 0 && !cfs_signal_pending())
+               if (timeout == 0 && !signal_pending(current))
                         /*
                          * No requests are in-flight (ether timed out
                          * or delayed), so we can allow interrupts.
                          * We still want to block for a limited time,
                          * so we allow interrupts during the timeout.
                          */
-                        lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1), 
+                       lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
                                                    ptlrpc_expired_set,
                                                    ptlrpc_interrupted_set, set);
                 else
                         /*
                          * At least one request is in flight, so no
                          * interrupts are allowed. Wait until all
-                         * complete, or an in-flight req times out. 
+                        * complete, or an in-flight req times out.
                          */
                         lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
                                           ptlrpc_expired_set, set);
@@ -2158,7 +2180,7 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set)
                  * pending when we started, we need to handle it now or we risk
                  * it being ignored forever */
                if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
-                   cfs_signal_pending()) {
+                   signal_pending(current)) {
                        sigset_t blocked_sigs =
                                           cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
 
@@ -2166,7 +2188,7 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set)
                         * like SIGINT or SIGKILL. We still ignore less
                         * important signals since ptlrpc set is not easily
                         * reentrant from userspace again */
-                       if (cfs_signal_pending())
+                       if (signal_pending(current))
                                ptlrpc_interrupted_set(set);
                        cfs_restore_sigs(blocked_sigs);
                }
@@ -2181,9 +2203,9 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set)
                  * I don't really care if we go once more round the loop in
                  * the error cases -eeb. */
                if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
-                        cfs_list_for_each(tmp, &set->set_requests) {
-                                req = cfs_list_entry(tmp, struct ptlrpc_request,
-                                                     rq_set_chain);
+                       list_for_each(tmp, &set->set_requests) {
+                               req = list_entry(tmp, struct ptlrpc_request,
+                                                rq_set_chain);
                                spin_lock(&req->rq_lock);
                                req->rq_invalid_rqset = 1;
                                spin_unlock(&req->rq_lock);
@@ -2194,8 +2216,8 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set)
        LASSERT(atomic_read(&set->set_remaining) == 0);
 
         rc = set->set_rc; /* rq_status of already freed requests if any */
-        cfs_list_for_each(tmp, &set->set_requests) {
-                req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
+       list_for_each(tmp, &set->set_requests) {
+               req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
 
                 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
                 if (req->rq_status != 0)
@@ -2210,9 +2232,9 @@ int ptlrpc_set_wait(struct ptlrpc_request_set *set)
                 struct ptlrpc_set_cbdata *cbdata, *n;
                 int err;
 
-                cfs_list_for_each_entry_safe(cbdata, n,
+               list_for_each_entry_safe(cbdata, n,
                                          &set->set_cblist, psc_item) {
-                        cfs_list_del_init(&cbdata->psc_item);
+                       list_del_init(&cbdata->psc_item);
                         err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
                         if (err && !rc)
                                 rc = err;
@@ -2234,31 +2256,30 @@ EXPORT_SYMBOL(ptlrpc_set_wait);
  */
 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 {
-        ENTRY;
-        if (request == NULL) {
-                EXIT;
-                return;
-        }
+       ENTRY;
 
-        LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
-        LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
-        LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
-        LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
-        LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
-        LASSERTF(!request->rq_replay, "req %p\n", request);
+       if (request == NULL)
+               RETURN_EXIT;
 
-        req_capsule_fini(&request->rq_pill);
+       LASSERT(!request->rq_srv_req);
+       LASSERT(request->rq_export == NULL);
+       LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
+       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);
 
-        /* We must take it off the imp_replay_list first.  Otherwise, we'll set
-         * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
-        if (request->rq_import != NULL) {
+       req_capsule_fini(&request->rq_pill);
+
+       /* We must take it off the imp_replay_list first.  Otherwise, we'll set
+        * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
+       if (request->rq_import != NULL) {
                if (!locked)
                        spin_lock(&request->rq_import->imp_lock);
-               cfs_list_del_init(&request->rq_replay_list);
+               list_del_init(&request->rq_replay_list);
                if (!locked)
                        spin_unlock(&request->rq_import->imp_lock);
         }
-        LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
+       LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
 
        if (atomic_read(&request->rq_refcount) != 0) {
                DEBUG_REQ(D_ERROR, request,
@@ -2268,10 +2289,7 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 
         if (request->rq_repbuf != NULL)
                 sptlrpc_cli_free_repbuf(request);
-        if (request->rq_export != NULL) {
-                class_export_put(request->rq_export);
-                request->rq_export = NULL;
-        }
+
         if (request->rq_import != NULL) {
                 class_import_put(request->rq_import);
                 request->rq_import = NULL;
@@ -2295,14 +2313,13 @@ static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
 /**
  * Drop one request reference. Must be called with import imp_lock held.
- * When reference count drops to zero, reuqest is freed.
+ * When reference count drops to zero, request is freed.
  */
 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
 {
        assert_spin_locked(&request->rq_import->imp_lock);
        (void)__ptlrpc_req_finished(request, 1);
 }
-EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
 
 /**
  * Helper function
@@ -2359,7 +2376,7 @@ EXPORT_SYMBOL(ptlrpc_req_xid);
  * The request owner (i.e. the thread doing the I/O) must call...
  * Returns 0 on success or 1 if unregistering cannot be made.
  */
-int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
+static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
 {
        int                rc;
        struct l_wait_info lwi;
@@ -2407,12 +2424,10 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
          * unlinked before returning a req to the pool.
          */
         for (;;) {
-#ifdef __KERNEL__
                /* The wq argument is ignored by user-space wait_event macros */
                wait_queue_head_t *wq = (request->rq_set != NULL) ?
                                        &request->rq_set->set_waitq :
                                        &request->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),
@@ -2425,13 +2440,14 @@ int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
                 }
 
                 LASSERT(rc == -ETIMEDOUT);
-                DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
-                          "rvcng=%d unlnk=%d", request->rq_receiving_reply,
-                          request->rq_must_unlink);
+               DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
+                         "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
+                         request->rq_receiving_reply,
+                         request->rq_req_unlinked,
+                         request->rq_reply_unlinked);
         }
         RETURN(0);
 }
-EXPORT_SYMBOL(ptlrpc_unregister_reply);
 
 static void ptlrpc_free_request(struct ptlrpc_request *req)
 {
@@ -2441,7 +2457,7 @@ static void ptlrpc_free_request(struct ptlrpc_request *req)
 
        if (req->rq_commit_cb != NULL)
                req->rq_commit_cb(req);
-       cfs_list_del_init(&req->rq_replay_list);
+       list_del_init(&req->rq_replay_list);
 
        __ptlrpc_req_finished(req, 1);
 }
@@ -2454,7 +2470,7 @@ void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
        struct obd_import       *imp = req->rq_import;
 
        spin_lock(&imp->imp_lock);
-       if (cfs_list_empty(&req->rq_replay_list)) {
+       if (list_empty(&req->rq_replay_list)) {
                spin_unlock(&imp->imp_lock);
                return;
        }
@@ -2494,13 +2510,14 @@ void ptlrpc_free_committed(struct obd_import *imp)
                imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
                imp->imp_generation);
 
-       if (imp->imp_generation != imp->imp_last_generation_checked)
+       if (imp->imp_generation != imp->imp_last_generation_checked ||
+           imp->imp_last_transno_checked == 0)
                skip_committed_list = false;
 
         imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
         imp->imp_last_generation_checked = imp->imp_generation;
 
-       cfs_list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
+       list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
                                     rq_replay_list) {
                 /* XXX ok to remove when 1357 resolved - rread 05/29/03  */
                 LASSERT(req != last_req);
@@ -2523,7 +2540,7 @@ void ptlrpc_free_committed(struct obd_import *imp)
 
                if (req->rq_replay) {
                        DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
-                       cfs_list_move_tail(&req->rq_replay_list,
+                       list_move_tail(&req->rq_replay_list,
                                           &imp->imp_committed_list);
                        continue;
                }
@@ -2537,12 +2554,15 @@ free_req:
        if (skip_committed_list)
                GOTO(out, 0);
 
-       cfs_list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
+       list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
                                     rq_replay_list) {
                LASSERT(req->rq_transno != 0);
                if (req->rq_import_generation < imp->imp_generation) {
                        DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
                        ptlrpc_free_request(req);
+               } else if (!req->rq_replay) {
+                       DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
+                       ptlrpc_free_request(req);
                }
        }
 out:
@@ -2554,7 +2574,6 @@ void ptlrpc_cleanup_client(struct obd_import *imp)
         ENTRY;
         EXIT;
 }
-EXPORT_SYMBOL(ptlrpc_cleanup_client);
 
 /**
  * Schedule previously sent request for resend.
@@ -2565,10 +2584,19 @@ EXPORT_SYMBOL(ptlrpc_cleanup_client);
 void ptlrpc_resend_req(struct ptlrpc_request *req)
 {
         DEBUG_REQ(D_HA, req, "going to resend");
+       spin_lock(&req->rq_lock);
+
+       /* Request got reply but linked to the import list still.
+          Let ptlrpc_check_set() to process it. */
+       if (ptlrpc_client_replied(req)) {
+               spin_unlock(&req->rq_lock);
+               DEBUG_REQ(D_HA, req, "it has reply, so skip it");
+               return;
+       }
+
         lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
         req->rq_status = -EAGAIN;
 
-       spin_lock(&req->rq_lock);
         req->rq_resend = 1;
         req->rq_net_err = 0;
         req->rq_timedout = 0;
@@ -2583,7 +2611,6 @@ void ptlrpc_resend_req(struct ptlrpc_request *req)
         ptlrpc_client_wake_req(req);
        spin_unlock(&req->rq_lock);
 }
-EXPORT_SYMBOL(ptlrpc_resend_req);
 
 /* XXX: this function and rq_status are currently unused */
 void ptlrpc_restart_req(struct ptlrpc_request *req)
@@ -2597,7 +2624,6 @@ void ptlrpc_restart_req(struct ptlrpc_request *req)
        ptlrpc_client_wake_req(req);
        spin_unlock(&req->rq_lock);
 }
-EXPORT_SYMBOL(ptlrpc_restart_req);
 
 /**
  * Grab additional reference on a request \a req
@@ -2617,7 +2643,7 @@ EXPORT_SYMBOL(ptlrpc_request_addref);
 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
                                       struct obd_import *imp)
 {
-       cfs_list_t *tmp;
+       struct list_head *tmp;
 
        assert_spin_locked(&imp->imp_lock);
 
@@ -2630,19 +2656,19 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
            as resent replayed requests. */
         lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
 
-        /* don't re-add requests that have been replayed */
-        if (!cfs_list_empty(&req->rq_replay_list))
-                return;
+       /* don't re-add requests that have been replayed */
+       if (!list_empty(&req->rq_replay_list))
+               return;
 
-        lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
+       lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
 
-        LASSERT(imp->imp_replayable);
-        /* Balanced in ptlrpc_free_committed, usually. */
-        ptlrpc_request_addref(req);
-        cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
-                struct ptlrpc_request *iter =
-                        cfs_list_entry(tmp, struct ptlrpc_request,
-                                       rq_replay_list);
+       LASSERT(imp->imp_replayable);
+       /* Balanced in ptlrpc_free_committed, usually. */
+       ptlrpc_request_addref(req);
+       list_for_each_prev(tmp, &imp->imp_replay_list) {
+               struct ptlrpc_request *iter = list_entry(tmp,
+                                                        struct ptlrpc_request,
+                                                        rq_replay_list);
 
                 /* We may have duplicate transnos if we create and then
                  * open a file, or for closes retained if to match creating
@@ -2659,13 +2685,12 @@ void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
                                 continue;
                 }
 
-                cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
-                return;
-        }
+               list_add(&req->rq_replay_list, &iter->rq_replay_list);
+               return;
+       }
 
-        cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
+       list_add(&req->rq_replay_list, &imp->imp_replay_list);
 }
-EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
 
 /**
  * Send request and wait until it completes.
@@ -2680,11 +2705,11 @@ int ptlrpc_queue_wait(struct ptlrpc_request *req)
         LASSERT(req->rq_set == NULL);
         LASSERT(!req->rq_receiving_reply);
 
-        set = ptlrpc_prep_set();
-        if (set == NULL) {
-                CERROR("Unable to allocate ptlrpc set.");
-                RETURN(-ENOMEM);
-        }
+       set = ptlrpc_prep_set();
+       if (set == NULL) {
+               CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
+               RETURN(-ENOMEM);
+       }
 
        /* for distributed debugging */
        lustre_msg_set_status(req->rq_reqmsg, current_pid());
@@ -2699,14 +2724,9 @@ int ptlrpc_queue_wait(struct ptlrpc_request *req)
 }
 EXPORT_SYMBOL(ptlrpc_queue_wait);
 
-struct ptlrpc_replay_async_args {
-        int praa_old_state;
-        int praa_old_status;
-};
-
 /**
  * Callback used for replayed requests reply processing.
- * In case of succesful reply calls registeresd request replay callback.
+ * In case of successful reply calls registered request replay callback.
  * In case of error restart replay process.
  */
 static int ptlrpc_replay_interpret(const struct lu_env *env,
@@ -2840,29 +2860,29 @@ int ptlrpc_replay_req(struct ptlrpc_request *req)
        ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
        RETURN(0);
 }
-EXPORT_SYMBOL(ptlrpc_replay_req);
 
 /**
  * Aborts all in-flight request on import \a imp sending and delayed lists
  */
 void ptlrpc_abort_inflight(struct obd_import *imp)
 {
-        cfs_list_t *tmp, *n;
-        ENTRY;
+       struct list_head *tmp, *n;
+       ENTRY;
 
-        /* Make sure that no new requests get processed for this import.
-         * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
-         * this flag and then putting requests on sending_list or delayed_list.
-         */
+       /* Make sure that no new requests get processed for this import.
+        * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
+        * this flag and then putting requests on sending_list or delayed_list.
+        */
        spin_lock(&imp->imp_lock);
 
-        /* XXX locking?  Maybe we should remove each request with the list
-         * locked?  Also, how do we know if the requests on the list are
-         * being freed at this time?
-         */
-        cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
-                struct ptlrpc_request *req =
-                        cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
+       /* XXX locking?  Maybe we should remove each request with the list
+        * locked?  Also, how do we know if the requests on the list are
+        * being freed at this time?
+        */
+       list_for_each_safe(tmp, n, &imp->imp_sending_list) {
+               struct ptlrpc_request *req = list_entry(tmp,
+                                                       struct ptlrpc_request,
+                                                       rq_list);
 
                 DEBUG_REQ(D_RPCTRACE, req, "inflight");
 
@@ -2875,9 +2895,9 @@ void ptlrpc_abort_inflight(struct obd_import *imp)
                spin_unlock(&req->rq_lock);
        }
 
-       cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
+       list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
                struct ptlrpc_request *req =
-                       cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
+                       list_entry(tmp, struct ptlrpc_request, rq_list);
 
                DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
 
@@ -2899,21 +2919,20 @@ void ptlrpc_abort_inflight(struct obd_import *imp)
 
        EXIT;
 }
-EXPORT_SYMBOL(ptlrpc_abort_inflight);
 
 /**
  * Abort all uncompleted requests in request set \a set
  */
 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
 {
-        cfs_list_t *tmp, *pos;
+       struct list_head *tmp, *pos;
 
-        LASSERT(set != NULL);
+       LASSERT(set != NULL);
 
-        cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
-                struct ptlrpc_request *req =
-                        cfs_list_entry(pos, struct ptlrpc_request,
-                                       rq_set_chain);
+       list_for_each_safe(pos, tmp, &set->set_requests) {
+               struct ptlrpc_request *req =
+                       list_entry(pos, struct ptlrpc_request,
+                                  rq_set_chain);
 
                spin_lock(&req->rq_lock);
                if (req->rq_phase != RQ_PHASE_RPC) {
@@ -2988,7 +3007,6 @@ __u64 ptlrpc_next_xid(void)
 
        return next;
 }
-EXPORT_SYMBOL(ptlrpc_next_xid);
 
 /**
  * Get a glimpse at what next xid value might have been.
@@ -3082,7 +3100,7 @@ static int ptlrpcd_check_work(struct ptlrpc_request *req)
 void *ptlrpcd_alloc_work(struct obd_import *imp,
                         int (*cb)(const struct lu_env *, void *), void *cbdata)
 {
-       struct ptlrpc_request         *req = NULL;
+       struct ptlrpc_request         *req = NULL;
        struct ptlrpc_work_async_args *args;
        ENTRY;
 
@@ -3098,27 +3116,16 @@ void *ptlrpcd_alloc_work(struct obd_import *imp,
                RETURN(ERR_PTR(-ENOMEM));
        }
 
+       ptlrpc_cli_req_init(req);
+
        req->rq_send_state = LUSTRE_IMP_FULL;
        req->rq_type = PTL_RPC_MSG_REQUEST;
        req->rq_import = class_import_get(imp);
-       req->rq_export = NULL;
        req->rq_interpret_reply = work_interpreter;
        /* don't want reply */
-       req->rq_receiving_reply = 0;
-       req->rq_must_unlink = 0;
        req->rq_no_delay = req->rq_no_resend = 1;
        req->rq_pill.rc_fmt = (void *)&worker_format;
 
-       spin_lock_init(&req->rq_lock);
-       CFS_INIT_LIST_HEAD(&req->rq_list);
-       CFS_INIT_LIST_HEAD(&req->rq_replay_list);
-       CFS_INIT_LIST_HEAD(&req->rq_set_chain);
-       CFS_INIT_LIST_HEAD(&req->rq_history_list);
-       CFS_INIT_LIST_HEAD(&req->rq_exp_list);
-       init_waitqueue_head(&req->rq_reply_waitq);
-       init_waitqueue_head(&req->rq_set_waitq);
-       atomic_set(&req->rq_refcount, 1);
-
        CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
        args = ptlrpc_req_async_args(req);
        args->cb     = cb;