Whamcloud - gitweb
LU-1517 ptlrpc: throw net error to ptlrpc for bulk
[fs/lustre-release.git] / lustre / ptlrpc / events.c
index c08beb2..55cc32e 100644 (file)
@@ -1,6 +1,4 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
  * GPL HEADER START
  *
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
@@ -26,7 +24,7 @@
  * GPL HEADER END
  */
 /*
- * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
  */
 /*
@@ -65,19 +63,19 @@ void request_out_callback(lnet_event_t *ev)
                  ev->type == LNET_EVENT_UNLINK);
         LASSERT (ev->unlinked);
 
-        DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
-                  "type %d, status %d", ev->type, ev->status);
+        DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
 
         sptlrpc_request_out_callback(req);
+        req->rq_real_sent = cfs_time_current_sec();
 
         if (ev->type == LNET_EVENT_UNLINK || ev->status != 0) {
 
                 /* Failed send: make it seem like the reply timed out, just
                  * like failing sends in client.c does currently...  */
 
-                spin_lock(&req->rq_lock);
+                cfs_spin_lock(&req->rq_lock);
                 req->rq_net_err = 1;
-                spin_unlock(&req->rq_lock);
+                cfs_spin_unlock(&req->rq_lock);
 
                 ptlrpc_client_wake_req(req);
         }
@@ -96,17 +94,16 @@ void reply_in_callback(lnet_event_t *ev)
         struct ptlrpc_request *req = cbid->cbid_arg;
         ENTRY;
 
-        DEBUG_REQ((ev->status == 0) ? D_NET : D_ERROR, req,
-                  "type %d, status %d", ev->type, ev->status);
+        DEBUG_REQ(D_NET, req, "type %d, status %d", ev->type, ev->status);
 
         LASSERT (ev->type == LNET_EVENT_PUT || ev->type == LNET_EVENT_UNLINK);
         LASSERT (ev->md.start == req->rq_repbuf);
-        LASSERT (ev->mlength <= req->rq_repbuf_len);
+        LASSERT (ev->offset + ev->mlength <= req->rq_repbuf_len);
         /* We've set LNET_MD_MANAGE_REMOTE for all outgoing requests
            for adaptive timeouts' early reply. */
         LASSERT((ev->md.options & LNET_MD_MANAGE_REMOTE) != 0);
 
-        spin_lock(&req->rq_lock);
+        cfs_spin_lock(&req->rq_lock);
 
         req->rq_receiving_reply = 0;
         req->rq_early = 0;
@@ -115,9 +112,20 @@ void reply_in_callback(lnet_event_t *ev)
 
         if (ev->status)
                 goto out_wake;
+
         if (ev->type == LNET_EVENT_UNLINK) {
                 LASSERT(ev->unlinked);
-                DEBUG_REQ(D_RPCTRACE, req, "unlink");
+                DEBUG_REQ(D_NET, req, "unlink");
+                goto out_wake;
+        }
+
+        if (ev->mlength < ev->rlength ) {
+                CDEBUG(D_RPCTRACE, "truncate req %p rpc %d - %d+%d\n", req,
+                       req->rq_replen, ev->rlength, ev->offset);
+                req->rq_reply_truncate = 1;
+                req->rq_replied = 1;
+                req->rq_status = -EOVERFLOW;
+                req->rq_nob_received = ev->rlength + ev->offset;
                 goto out_wake;
         }
 
@@ -159,7 +167,7 @@ out_wake:
         /* NB don't unlock till after wakeup; req can disappear under us
          * since we don't have our own ref */
         ptlrpc_client_wake_req(req);
-        spin_unlock(&req->rq_lock);
+        cfs_spin_unlock(&req->rq_lock);
         EXIT;
 }
 
@@ -170,6 +178,7 @@ void client_bulk_callback (lnet_event_t *ev)
 {
         struct ptlrpc_cb_id     *cbid = ev->md.user_ptr;
         struct ptlrpc_bulk_desc *desc = cbid->cbid_arg;
+        struct ptlrpc_request   *req;
         ENTRY;
 
         LASSERT ((desc->bd_type == BULK_PUT_SINK &&
@@ -179,12 +188,18 @@ void client_bulk_callback (lnet_event_t *ev)
                  ev->type == LNET_EVENT_UNLINK);
         LASSERT (ev->unlinked);
 
+        if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB, CFS_FAIL_ONCE))
+                ev->status = -EIO;
+
+        if (CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2,CFS_FAIL_ONCE))
+                ev->status = -EIO;
+
         CDEBUG((ev->status == 0) ? D_NET : D_ERROR,
                "event type %d, status %d, desc %p\n",
                ev->type, ev->status, desc);
 
-        spin_lock(&desc->bd_lock);
-
+        cfs_spin_lock(&desc->bd_lock);
+        req = desc->bd_req;
         LASSERT(desc->bd_network_rw);
         desc->bd_network_rw = 0;
 
@@ -192,6 +207,11 @@ void client_bulk_callback (lnet_event_t *ev)
                 desc->bd_success = 1;
                 desc->bd_nob_transferred = ev->mlength;
                 desc->bd_sender = ev->sender;
+        } else {
+                /* start reconnect and resend if network error hit */
+                cfs_spin_lock(&req->rq_lock);
+                req->rq_net_err = 1;
+                cfs_spin_unlock(&req->rq_lock);
         }
 
         /* release the encrypted pages for write */
@@ -200,20 +220,75 @@ void client_bulk_callback (lnet_event_t *ev)
 
         /* NB don't unlock till after wakeup; desc can disappear under us
          * otherwise */
-        ptlrpc_client_wake_req(desc->bd_req);
+        ptlrpc_client_wake_req(req);
 
-        spin_unlock(&desc->bd_lock);
+        cfs_spin_unlock(&desc->bd_lock);
         EXIT;
 }
 
 /*
+ * We will have percpt request history list for ptlrpc service in upcoming
+ * patches because we don't want to be serialized by current per-service
+ * history operations. So we require history ID can (somehow) show arriving
+ * order w/o grabbing global lock, and user can sort them in userspace.
+ *
+ * This is how we generate history ID for ptlrpc_request:
+ * ----------------------------------------------------
+ * |  32 bits  |  16 bits  | (16 - X)bits  |  X bits  |
+ * ----------------------------------------------------
+ * |  seconds  | usec / 16 |   sequence    | CPT id   |
+ * ----------------------------------------------------
+ *
+ * it might not be precise but should be good enough.
+ */
+
+#define REQS_CPT_BITS(svcpt)   ((svcpt)->scp_service->srv_cpt_bits)
+
+#define REQS_SEC_SHIFT         32
+#define REQS_USEC_SHIFT                16
+#define REQS_SEQ_SHIFT(svcpt)  REQS_CPT_BITS(svcpt)
+
+static void ptlrpc_req_add_history(struct ptlrpc_service_part *svcpt,
+                                  struct ptlrpc_request *req)
+{
+       __u64   sec = req->rq_arrival_time.tv_sec;
+       __u32   usec = req->rq_arrival_time.tv_usec >> 4; /* usec / 16 */
+       __u64   new_seq;
+
+       /* set sequence ID for request and add it to history list,
+        * it must be called with hold svcpt::scp_lock */
+
+       new_seq = (sec << REQS_SEC_SHIFT) |
+                 (usec << REQS_USEC_SHIFT) | svcpt->scp_cpt;
+       if (new_seq > svcpt->scp_hist_seq) {
+               /* This handles the initial case of scp_hist_seq == 0 or
+                * we just jumped into a new time window */
+               svcpt->scp_hist_seq = new_seq;
+       } else {
+               LASSERT(REQS_SEQ_SHIFT(svcpt) < REQS_USEC_SHIFT);
+               /* NB: increase sequence number in current usec bucket,
+                * however, it's possible that we used up all bits for
+                * sequence and jumped into the next usec bucket (future time),
+                * then we hope there will be less RPCs per bucket at some
+                * point, and sequence will catch up again */
+               svcpt->scp_hist_seq += (1U << REQS_SEQ_SHIFT(svcpt));
+               new_seq = svcpt->scp_hist_seq;
+       }
+
+       req->rq_history_seq = new_seq;
+
+       cfs_list_add_tail(&req->rq_history_list, &svcpt->scp_hist_reqs);
+}
+
+/*
  * Server's incoming request callback
  */
 void request_in_callback(lnet_event_t *ev)
 {
-        struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
-        struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
-        struct ptlrpc_service             *service = rqbd->rqbd_service;
+       struct ptlrpc_cb_id               *cbid = ev->md.user_ptr;
+       struct ptlrpc_request_buffer_desc *rqbd = cbid->cbid_arg;
+       struct ptlrpc_service_part        *svcpt = rqbd->rqbd_svcpt;
+       struct ptlrpc_service             *service = svcpt->scp_service;
         struct ptlrpc_request             *req;
         ENTRY;
 
@@ -258,37 +333,35 @@ void request_in_callback(lnet_event_t *ev)
         req->rq_reqbuf = ev->md.start + ev->offset;
         if (ev->type == LNET_EVENT_PUT && ev->status == 0)
                 req->rq_reqdata_len = ev->mlength;
-        do_gettimeofday(&req->rq_arrival_time);
+        cfs_gettimeofday(&req->rq_arrival_time);
         req->rq_peer = ev->initiator;
         req->rq_self = ev->target.nid;
         req->rq_rqbd = rqbd;
         req->rq_phase = RQ_PHASE_NEW;
-#ifdef CRAY_XT3
-        req->rq_uid = ev->uid;
-#endif
-        spin_lock_init(&req->rq_lock);
+        cfs_spin_lock_init(&req->rq_lock);
         CFS_INIT_LIST_HEAD(&req->rq_timed_list);
-        atomic_set(&req->rq_refcount, 1);
+       CFS_INIT_LIST_HEAD(&req->rq_exp_list);
+        cfs_atomic_set(&req->rq_refcount, 1);
         if (ev->type == LNET_EVENT_PUT)
-                DEBUG_REQ(D_RPCTRACE, req, "incoming req");
+                CDEBUG(D_INFO, "incoming req@%p x"LPU64" msgsize %u\n",
+                       req, req->rq_xid, ev->mlength);
 
         CDEBUG(D_RPCTRACE, "peer: %s\n", libcfs_id2str(req->rq_peer));
 
-        spin_lock(&service->srv_lock);
+       cfs_spin_lock(&svcpt->scp_lock);
 
-        req->rq_history_seq = service->srv_request_seq++;
-        list_add_tail(&req->rq_history_list, &service->srv_request_history);
+       ptlrpc_req_add_history(svcpt, req);
 
-        if (ev->unlinked) {
-                service->srv_nrqbd_receiving--;
-                CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
-                       service->srv_nrqbd_receiving);
-
-                /* Normally, don't complain about 0 buffers posted; LNET won't
-                 * drop incoming reqs since we set the portal lazy */
-                if (test_req_buffer_pressure &&
-                    ev->type != LNET_EVENT_UNLINK &&
-                    service->srv_nrqbd_receiving == 0)
+       if (ev->unlinked) {
+               svcpt->scp_nrqbds_posted--;
+               CDEBUG(D_INFO, "Buffer complete: %d buffers still posted\n",
+                      svcpt->scp_nrqbds_posted);
+
+               /* Normally, don't complain about 0 buffers posted; LNET won't
+                * drop incoming reqs since we set the portal lazy */
+               if (test_req_buffer_pressure &&
+                   ev->type != LNET_EVENT_UNLINK &&
+                   svcpt->scp_nrqbds_posted == 0)
                         CWARN("All %s request buffers busy\n",
                               service->srv_name);
 
@@ -298,15 +371,15 @@ void request_in_callback(lnet_event_t *ev)
                 rqbd->rqbd_refcount++;
         }
 
-        list_add_tail(&req->rq_list, &service->srv_req_in_queue);
-        service->srv_n_queued_reqs++;
+       cfs_list_add_tail(&req->rq_list, &svcpt->scp_req_incoming);
+       svcpt->scp_nreqs_incoming++;
 
-        /* NB everything can disappear under us once the request
-         * has been queued and we unlock, so do the wake now... */
-        cfs_waitq_signal(&service->srv_waitq);
+       /* NB everything can disappear under us once the request
+        * has been queued and we unlock, so do the wake now... */
+       cfs_waitq_signal(&svcpt->scp_waitq);
 
-        spin_unlock(&service->srv_lock);
-        EXIT;
+       cfs_spin_unlock(&svcpt->scp_lock);
+       EXIT;
 }
 
 /*
@@ -314,9 +387,9 @@ void request_in_callback(lnet_event_t *ev)
  */
 void reply_out_callback(lnet_event_t *ev)
 {
-        struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
-        struct ptlrpc_reply_state *rs = cbid->cbid_arg;
-        struct ptlrpc_service     *svc = rs->rs_service;
+       struct ptlrpc_cb_id       *cbid = ev->md.user_ptr;
+       struct ptlrpc_reply_state *rs = cbid->cbid_arg;
+       struct ptlrpc_service_part *svcpt = rs->rs_svcpt;
         ENTRY;
 
         LASSERT (ev->type == LNET_EVENT_SEND ||
@@ -328,7 +401,6 @@ void reply_out_callback(lnet_event_t *ev)
                  * net's ref on 'rs' */
                 LASSERT (ev->unlinked);
                 ptlrpc_rs_decref(rs);
-                atomic_dec (&svc->srv_outstanding_replies);
                 EXIT;
                 return;
         }
@@ -338,19 +410,22 @@ void reply_out_callback(lnet_event_t *ev)
         if (ev->unlinked) {
                 /* Last network callback. The net's ref on 'rs' stays put
                  * until ptlrpc_handle_rs() is done with it */
-                spin_lock(&svc->srv_lock);
-                spin_lock(&rs->rs_lock);
-                rs->rs_on_net = 0;
-                if (!rs->rs_no_ack ||
-                    rs->rs_transno <= rs->rs_export->exp_obd->obd_last_committed)
-                        ptlrpc_schedule_difficult_reply (rs);
-                spin_unlock(&rs->rs_lock);
-                spin_unlock(&svc->srv_lock);
-        }
-
-        EXIT;
+               cfs_spin_lock(&svcpt->scp_rep_lock);
+               cfs_spin_lock(&rs->rs_lock);
+
+               rs->rs_on_net = 0;
+               if (!rs->rs_no_ack ||
+                   rs->rs_transno <=
+                   rs->rs_export->exp_obd->obd_last_committed)
+                       ptlrpc_schedule_difficult_reply(rs);
+
+               cfs_spin_unlock(&rs->rs_lock);
+               cfs_spin_unlock(&svcpt->scp_rep_lock);
+       }
+       EXIT;
 }
 
+#ifdef HAVE_SERVER_SUPPORT
 /*
  * Server's bulk completion callback
  */
@@ -371,7 +446,7 @@ void server_bulk_callback (lnet_event_t *ev)
                "event type %d, status %d, desc %p\n",
                ev->type, ev->status, desc);
 
-        spin_lock(&desc->bd_lock);
+        cfs_spin_lock(&desc->bd_lock);
 
         if ((ev->type == LNET_EVENT_ACK ||
              ev->type == LNET_EVENT_REPLY) &&
@@ -390,9 +465,10 @@ void server_bulk_callback (lnet_event_t *ev)
                 cfs_waitq_signal(&desc->bd_waitq);
         }
 
-        spin_unlock(&desc->bd_lock);
+        cfs_spin_unlock(&desc->bd_lock);
         EXIT;
 }
+#endif
 
 static void ptlrpc_master_callback(lnet_event_t *ev)
 {
@@ -405,8 +481,11 @@ static void ptlrpc_master_callback(lnet_event_t *ev)
                  callback == reply_in_callback ||
                  callback == client_bulk_callback ||
                  callback == request_in_callback ||
-                 callback == reply_out_callback ||
-                 callback == server_bulk_callback);
+                 callback == reply_out_callback
+#ifdef HAVE_SERVER_SUPPORT
+                 || callback == server_bulk_callback
+#endif
+                 );
 
         callback (ev);
 }
@@ -461,8 +540,6 @@ int ptlrpc_uuid_to_peer (struct obd_uuid *uuid,
         }
 
         CDEBUG(D_NET,"%s->%s\n", uuid->uuid, libcfs_id2str(*peer));
-        if (rc != 0)
-                CERROR("No NID found for %s\n", uuid->uuid);
         return rc;
 }
 
@@ -532,10 +609,11 @@ int ptlrpc_ni_init(void)
         /* CAVEAT EMPTOR: how we process portals events is _radically_
          * different depending on... */
 #ifdef __KERNEL__
-        /* kernel portals calls our master callback when events are added to
-         * the event queue.  In fact lustre never pulls events off this queue,
-         * so it's only sized for some debug history. */
-        rc = LNetEQAlloc(1024, ptlrpc_master_callback, &ptlrpc_eq_h);
+       /* kernel LNet calls our master callback when there are new event,
+        * because we are guaranteed to get every event via callback,
+        * so we just set EQ size to 0 to avoid overhread of serializing
+        * enqueue/dequeue operations in LNet. */
+       rc = LNetEQAlloc(0, ptlrpc_master_callback, &ptlrpc_eq_h);
 #else
         /* liblustre calls the master callback when it removes events from the
          * event queue.  The event queue has to be big enough not to drop
@@ -557,7 +635,7 @@ CFS_LIST_HEAD(liblustre_idle_callbacks);
 void *liblustre_services_callback;
 
 void *
-liblustre_register_waitidle_callback (struct list_head *callback_list,
+liblustre_register_waitidle_callback (cfs_list_t *callback_list,
                                       const char *name,
                                       int (*fn)(void *arg), void *arg)
 {
@@ -569,7 +647,7 @@ liblustre_register_waitidle_callback (struct list_head *callback_list,
         llwc->llwc_name = name;
         llwc->llwc_fn = fn;
         llwc->llwc_arg = arg;
-        list_add_tail(&llwc->llwc_list, callback_list);
+        cfs_list_add_tail(&llwc->llwc_list, callback_list);
 
         return (llwc);
 }
@@ -579,7 +657,7 @@ liblustre_deregister_waitidle_callback (void *opaque)
 {
         struct liblustre_wait_callback *llwc = opaque;
 
-        list_del(&llwc->llwc_list);
+        cfs_list_del(&llwc->llwc_list);
         OBD_FREE(llwc, sizeof(*llwc));
 }
 
@@ -641,7 +719,7 @@ int liblustre_waiting = 0;
 int
 liblustre_wait_event (int timeout)
 {
-        struct list_head               *tmp;
+        cfs_list_t                     *tmp;
         struct liblustre_wait_callback *llwc;
         int                             found_something = 0;
 
@@ -654,9 +732,10 @@ liblustre_wait_event (int timeout)
                         found_something = 1;
 
                 /* Give all registered callbacks a bite at the cherry */
-                list_for_each(tmp, &liblustre_wait_callbacks) {
-                        llwc = list_entry(tmp, struct liblustre_wait_callback,
-                                          llwc_list);
+                cfs_list_for_each(tmp, &liblustre_wait_callbacks) {
+                        llwc = cfs_list_entry(tmp,
+                                              struct liblustre_wait_callback,
+                                              llwc_list);
 
                         if (llwc->llwc_fn(llwc->llwc_arg))
                                 found_something = 1;
@@ -681,7 +760,7 @@ liblustre_wait_idle(void)
 {
         static int recursed = 0;
 
-        struct list_head               *tmp;
+        cfs_list_t                     *tmp;
         struct liblustre_wait_callback *llwc;
         int                             idle = 0;
 
@@ -693,9 +772,10 @@ liblustre_wait_idle(void)
 
                 idle = 1;
 
-                list_for_each(tmp, &liblustre_idle_callbacks) {
-                        llwc = list_entry(tmp, struct liblustre_wait_callback,
-                                          llwc_list);
+                cfs_list_for_each(tmp, &liblustre_idle_callbacks) {
+                        llwc = cfs_list_entry(tmp,
+                                              struct liblustre_wait_callback,
+                                              llwc_list);
 
                         if (!llwc->llwc_fn(llwc->llwc_arg)) {
                                 idle = 0;
@@ -723,7 +803,7 @@ int ptlrpc_init_portals(void)
                 liblustre_register_wait_callback("liblustre_check_services",
                                                  &liblustre_check_services,
                                                  NULL);
-        init_completion_module(liblustre_wait_event);
+        cfs_init_completion_module(liblustre_wait_event);
 #endif
         rc = ptlrpcd_addref();
         if (rc == 0)