1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
41 #include <liblustre.h>
44 #include <obd_support.h>
45 #include <obd_class.h>
46 #include <lustre_lib.h>
47 #include <lustre_ha.h>
48 #include <lustre_import.h>
50 #include "ptlrpc_internal.h"
52 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
53 struct ptlrpc_client *cl)
55 cl->cli_request_portal = req_portal;
56 cl->cli_reply_portal = rep_portal;
60 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
62 struct ptlrpc_connection *c;
64 lnet_process_id_t peer;
67 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
69 CNETERR("cannot find peer %s!\n", uuid->uuid);
73 c = ptlrpc_connection_get(peer, self, uuid);
75 memcpy(c->c_remote_uuid.uuid,
76 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
79 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
84 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
86 struct ptlrpc_bulk_desc *desc;
88 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
92 spin_lock_init(&desc->bd_lock);
93 cfs_waitq_init(&desc->bd_waitq);
94 desc->bd_max_iov = npages;
95 desc->bd_iov_count = 0;
96 desc->bd_md_h = LNET_INVALID_HANDLE;
97 desc->bd_portal = portal;
103 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
104 int npages, int type, int portal)
106 struct obd_import *imp = req->rq_import;
107 struct ptlrpc_bulk_desc *desc;
110 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
111 desc = new_bulk(npages, type, portal);
115 desc->bd_import_generation = req->rq_import_generation;
116 desc->bd_import = class_import_get(imp);
119 desc->bd_cbid.cbid_fn = client_bulk_callback;
120 desc->bd_cbid.cbid_arg = desc;
122 /* This makes req own desc, and free it when she frees herself */
128 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
129 int npages, int type, int portal)
131 struct obd_export *exp = req->rq_export;
132 struct ptlrpc_bulk_desc *desc;
135 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
137 desc = new_bulk(npages, type, portal);
141 desc->bd_export = class_export_get(exp);
144 desc->bd_cbid.cbid_fn = server_bulk_callback;
145 desc->bd_cbid.cbid_arg = desc;
147 /* NB we don't assign rq_bulk here; server-side requests are
148 * re-used, and the handler frees the bulk desc explicitly. */
153 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
154 cfs_page_t *page, int pageoffset, int len)
156 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
157 LASSERT(page != NULL);
158 LASSERT(pageoffset >= 0);
160 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
165 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
168 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
173 LASSERT(desc != NULL);
174 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
175 LASSERT(!desc->bd_network_rw); /* network hands off or */
176 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
178 class_export_put(desc->bd_export);
180 class_import_put(desc->bd_import);
182 for (i = 0; i < desc->bd_iov_count ; i++)
183 cfs_page_unpin(desc->bd_iov[i].kiov_page);
185 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
186 bd_iov[desc->bd_max_iov]));
190 /* Set server timelimit for this req */
191 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
197 LASSERT(req->rq_import);
200 /* non-AT settings */
201 req->rq_timeout = req->rq_import->imp_server_timeout ?
202 obd_timeout / 2 : obd_timeout;
204 at = &req->rq_import->imp_at;
205 idx = import_at_get_index(req->rq_import,
206 req->rq_request_portal);
207 serv_est = at_get(&at->iat_service_estimate[idx]);
208 req->rq_timeout = at_est2timeout(serv_est);
211 /* We could get even fancier here, using history to predict increased
214 /* Let the server know what this RPC timeout is by putting it in the
216 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
219 /* Adjust max service estimate based on server value */
220 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
221 unsigned int serv_est)
227 LASSERT(req->rq_import);
228 at = &req->rq_import->imp_at;
230 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
231 /* max service estimates are tracked on the server side,
232 so just keep minimal history here */
233 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
235 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
236 "has changed from %d to %d\n",
237 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
238 oldse, at_get(&at->iat_service_estimate[idx]));
241 /* Expected network latency per remote node (secs) */
242 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
244 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
247 /* Adjust expected network latency */
248 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
249 unsigned int service_time)
251 unsigned int nl, oldnl;
253 time_t now = cfs_time_current_sec();
255 LASSERT(req->rq_import);
256 at = &req->rq_import->imp_at;
258 /* Network latency is total time less server processing time */
259 nl = max_t(int, now - req->rq_sent - service_time, 0) + 1/*st rounding*/;
260 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
261 CWARN("Reported service time %u > total measured time %ld\n",
262 service_time, now - req->rq_sent);
264 oldnl = at_measured(&at->iat_net_latency, nl);
266 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
267 "has changed from %d to %d\n",
268 req->rq_import->imp_obd->obd_name,
270 &req->rq_import->imp_connection->c_remote_uuid),
271 oldnl, at_get(&at->iat_net_latency));
274 static int unpack_reply_common(struct ptlrpc_request *req)
278 req->rq_rep_swab_mask = 0;
279 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
281 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
286 lustre_set_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
291 static int unpack_reply(struct ptlrpc_request *req)
295 rc = unpack_reply_common(req);
299 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
301 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
307 static inline void unpack_reply_free_msg(struct lustre_msg *msg, int len)
312 static int unpack_reply_copy_msg(struct ptlrpc_request *req,
313 struct lustre_msg **msg, int *len)
315 struct lustre_msg *msgcpy;
316 __u32 csum_calc, csum_get;
320 LASSERT_SPIN_LOCKED(&req->rq_lock);
324 /* Swabbing required when rc == 1 */
325 rc = unpack_reply_common(req);
329 lencpy = req->rq_replen;
330 spin_unlock(&req->rq_lock);
332 OBD_ALLOC(msgcpy, lencpy);
334 spin_lock(&req->rq_lock);
337 spin_lock(&req->rq_lock);
339 /* Checksum must be calculated before being unswabbed. If the magic
340 * in the copy is unswabbed discard like the checksum failure case */
341 memcpy(msgcpy, req->rq_repmsg, lencpy);
342 if (lustre_msg_need_swab(msgcpy)) {
343 DEBUG_REQ(D_NET, req, "incorrect message magic: %08x\n",
345 GOTO(err, rc = -EINVAL);
348 csum_calc = lustre_msg_calc_cksum(msgcpy);
350 /* Unpack the copy the original rq_repmsg is untouched */
351 rc = lustre_unpack_msg_ptlrpc_body(msgcpy, MSG_PTLRPC_BODY_OFF, rc);
353 DEBUG_REQ(D_ERROR, req, "unpack msg copy failed: %d", rc);
354 GOTO(err, rc = -EPROTO);
357 /* For early replies the LND may update repmsg outside req->rq_lock
358 * resulting in a checksum failure which is non-harmful */
359 csum_get = lustre_msg_get_cksum(msgcpy);
360 if (csum_calc != csum_get) {
361 DEBUG_REQ(D_NET, req, "checksum mismatch: %x != %x\n",
362 csum_calc, csum_get);
363 GOTO(err, rc = -EINVAL);
370 unpack_reply_free_msg(msgcpy, lencpy);
374 /* Handle an early reply message. To prevent a real reply from arriving
375 * and changing req->rq_repmsg this func is called under the rq_lock */
376 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) {
377 struct lustre_msg *msg;
382 LASSERT_SPIN_LOCKED(&req->rq_lock);
385 /* All early replys for this request use a single repbuf which can
386 * be updated outside the req->rq_lock. To prevent racing we create
387 * a copy of the repmsg and verify its checksum before it is used. */
388 rc = unpack_reply_copy_msg(req, &msg, &len);
390 /* Let's just ignore it - same as if it never got here */
391 CDEBUG(D_ADAPTTO, "Discarding racing early reply: %d\n", rc);
395 /* Expecting to increase the service time estimate here */
396 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(msg));
397 ptlrpc_at_adj_net_latency(req, lustre_msg_get_service_time(msg));
398 /* Adjust the local timeout for this req */
399 ptlrpc_at_set_req_timeout(req);
401 olddl = req->rq_deadline;
402 /* Server assumes it now has rq_timeout from when it sent the
403 early reply, so client should give it at least that long. */
404 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
405 ptlrpc_at_get_net_latency(req);
407 DEBUG_REQ(D_ADAPTTO, req,
408 "Early reply #%d, new deadline in %lds (%+lds)",
409 req->rq_early_count, req->rq_deadline -
410 cfs_time_current_sec(), req->rq_deadline - olddl);
412 unpack_reply_free_msg(msg, len);
416 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
418 struct list_head *l, *tmp;
419 struct ptlrpc_request *req;
421 LASSERT(pool != NULL);
423 spin_lock(&pool->prp_lock);
424 list_for_each_safe(l, tmp, &pool->prp_req_list) {
425 req = list_entry(l, struct ptlrpc_request, rq_list);
426 list_del(&req->rq_list);
427 LASSERT(req->rq_reqmsg);
428 OBD_FREE(req->rq_reqmsg, pool->prp_rq_size);
429 OBD_FREE(req, sizeof(*req));
431 spin_unlock(&pool->prp_lock);
432 OBD_FREE(pool, sizeof(*pool));
435 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
440 while (size < pool->prp_rq_size)
443 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
444 "Trying to change pool size with nonempty pool "
445 "from %d to %d bytes\n", pool->prp_rq_size, size);
447 spin_lock(&pool->prp_lock);
448 pool->prp_rq_size = size;
449 for (i = 0; i < num_rq; i++) {
450 struct ptlrpc_request *req;
451 struct lustre_msg *msg;
453 spin_unlock(&pool->prp_lock);
454 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
457 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
459 OBD_FREE(req, sizeof(struct ptlrpc_request));
462 req->rq_reqmsg = msg;
464 spin_lock(&pool->prp_lock);
465 list_add_tail(&req->rq_list, &pool->prp_req_list);
467 spin_unlock(&pool->prp_lock);
471 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
472 void (*populate_pool)(struct ptlrpc_request_pool *, int))
474 struct ptlrpc_request_pool *pool;
476 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
480 /* Request next power of two for the allocation, because internally
481 kernel would do exactly this */
483 spin_lock_init(&pool->prp_lock);
484 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
485 pool->prp_rq_size = msgsize;
486 pool->prp_populate = populate_pool;
488 populate_pool(pool, num_rq);
490 if (list_empty(&pool->prp_req_list)) {
491 /* have not allocated a single request for the pool */
492 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
498 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
500 struct ptlrpc_request *request;
501 struct lustre_msg *reqmsg;
506 spin_lock(&pool->prp_lock);
508 /* See if we have anything in a pool, and bail out if nothing,
509 * in writeout path, where this matters, this is safe to do, because
510 * nothing is lost in this case, and when some in-flight requests
511 * complete, this code will be called again. */
512 if (unlikely(list_empty(&pool->prp_req_list))) {
513 spin_unlock(&pool->prp_lock);
517 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
519 list_del_init(&request->rq_list);
520 spin_unlock(&pool->prp_lock);
522 LASSERT(request->rq_reqmsg);
523 LASSERT(request->rq_pool);
525 reqmsg = request->rq_reqmsg;
526 memset(request, 0, sizeof(*request));
527 request->rq_reqmsg = reqmsg;
528 request->rq_pool = pool;
529 request->rq_reqlen = pool->prp_rq_size;
533 struct ptlrpc_request *
534 ptlrpc_prep_req_pool(struct obd_import *imp, __u32 version, int opcode,
535 int count, __u32 *lengths, char **bufs,
536 struct ptlrpc_request_pool *pool)
538 struct ptlrpc_request *request = NULL;
542 /* The obd disconnected */
546 LASSERT(imp != LP_POISON);
547 LASSERT((unsigned long)imp->imp_client > 0x1000);
548 LASSERT(imp->imp_client != LP_POISON);
551 request = ptlrpc_prep_req_from_pool(pool);
554 OBD_ALLOC(request, sizeof(*request));
557 CERROR("request allocation out of memory\n");
561 rc = lustre_pack_request(request, imp->imp_msg_magic, count, lengths,
564 LASSERT(!request->rq_pool);
565 OBD_FREE(request, sizeof(*request));
569 lustre_msg_add_version(request->rq_reqmsg, version);
570 request->rq_send_state = LUSTRE_IMP_FULL;
571 request->rq_type = PTL_RPC_MSG_REQUEST;
572 request->rq_import = class_import_get(imp);
573 request->rq_export = NULL;
575 request->rq_req_cbid.cbid_fn = request_out_callback;
576 request->rq_req_cbid.cbid_arg = request;
578 request->rq_reply_cbid.cbid_fn = reply_in_callback;
579 request->rq_reply_cbid.cbid_arg = request;
581 request->rq_reply_deadline = 0;
582 request->rq_phase = RQ_PHASE_NEW;
583 request->rq_next_phase = RQ_PHASE_UNDEFINED;
585 request->rq_request_portal = imp->imp_client->cli_request_portal;
586 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
588 ptlrpc_at_set_req_timeout(request);
590 spin_lock_init(&request->rq_lock);
591 CFS_INIT_LIST_HEAD(&request->rq_list);
592 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
593 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
594 CFS_INIT_LIST_HEAD(&request->rq_history_list);
595 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
596 cfs_waitq_init(&request->rq_reply_waitq);
597 cfs_waitq_init(&request->rq_set_waitq);
598 request->rq_xid = ptlrpc_next_xid();
599 atomic_set(&request->rq_refcount, 1);
601 lustre_msg_set_opc(request->rq_reqmsg, opcode);
606 struct ptlrpc_request *
607 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
608 __u32 *lengths, char **bufs)
610 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
614 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
615 unsigned int timeout,
616 int (*interpreter)(struct ptlrpc_request *,
619 struct ptlrpc_request *request = NULL;
622 OBD_ALLOC(request, sizeof(*request));
624 CERROR("request allocation out of memory\n");
628 request->rq_send_state = LUSTRE_IMP_FULL;
629 request->rq_type = PTL_RPC_MSG_REQUEST;
630 request->rq_import = class_import_get(imp);
631 request->rq_export = NULL;
632 request->rq_import_generation = imp->imp_generation;
634 request->rq_timeout = timeout;
635 request->rq_sent = cfs_time_current_sec();
636 request->rq_deadline = request->rq_sent + timeout;
637 request->rq_reply_deadline = request->rq_deadline;
638 request->rq_interpret_reply = interpreter;
639 request->rq_phase = RQ_PHASE_RPC;
640 request->rq_next_phase = RQ_PHASE_INTERPRET;
641 /* don't want reply */
642 request->rq_receiving_reply = 0;
643 request->rq_must_unlink = 0;
644 request->rq_no_delay = request->rq_no_resend = 1;
645 request->rq_fake = 1;
647 spin_lock_init(&request->rq_lock);
648 CFS_INIT_LIST_HEAD(&request->rq_list);
649 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
650 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
651 CFS_INIT_LIST_HEAD(&request->rq_history_list);
652 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
653 cfs_waitq_init(&request->rq_reply_waitq);
654 cfs_waitq_init(&request->rq_set_waitq);
656 request->rq_xid = ptlrpc_next_xid();
657 atomic_set(&request->rq_refcount, 1);
662 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
664 /* if we kill request before timeout - need adjust counter */
665 if (req->rq_phase == RQ_PHASE_RPC) {
666 struct ptlrpc_request_set *set = req->rq_set;
669 atomic_dec(&set->set_remaining);
672 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
673 list_del_init(&req->rq_list);
677 struct ptlrpc_request_set *ptlrpc_prep_set(void)
679 struct ptlrpc_request_set *set;
682 OBD_ALLOC(set, sizeof *set);
685 CFS_INIT_LIST_HEAD(&set->set_requests);
686 cfs_waitq_init(&set->set_waitq);
687 atomic_set(&set->set_remaining, 0);
688 spin_lock_init(&set->set_new_req_lock);
689 CFS_INIT_LIST_HEAD(&set->set_new_requests);
690 CFS_INIT_LIST_HEAD(&set->set_cblist);
695 /* Finish with this set; opposite of prep_set. */
696 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
698 struct list_head *tmp;
699 struct list_head *next;
704 /* Requests on the set should either all be completed, or all be new */
705 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
706 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
707 list_for_each (tmp, &set->set_requests) {
708 struct ptlrpc_request *req =
709 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
711 LASSERT(req->rq_phase == expected_phase);
715 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
716 atomic_read(&set->set_remaining) == n, "%d / %d\n",
717 atomic_read(&set->set_remaining), n);
719 list_for_each_safe(tmp, next, &set->set_requests) {
720 struct ptlrpc_request *req =
721 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
722 list_del_init(&req->rq_set_chain);
724 LASSERT(req->rq_phase == expected_phase);
726 if (req->rq_phase == RQ_PHASE_NEW) {
728 if (req->rq_interpret_reply != NULL) {
729 int (*interpreter)(struct ptlrpc_request *,
731 req->rq_interpret_reply;
733 /* higher level (i.e. LOV) failed;
734 * let the sub reqs clean up */
735 req->rq_status = -EBADR;
736 interpreter(req, &req->rq_async_args,
739 atomic_dec(&set->set_remaining);
742 spin_lock(&req->rq_lock);
744 req->rq_invalid_rqset = 0;
745 spin_unlock(&req->rq_lock);
747 cfs_waitq_signal(&req->rq_set_waitq);
748 ptlrpc_req_finished (req);
751 LASSERT(atomic_read(&set->set_remaining) == 0);
753 OBD_FREE(set, sizeof(*set));
757 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
758 set_interpreter_func fn, void *data)
760 struct ptlrpc_set_cbdata *cbdata;
762 OBD_SLAB_ALLOC(cbdata, ptlrpc_cbdata_slab,
763 CFS_ALLOC_STD, sizeof(*cbdata));
767 cbdata->psc_interpret = fn;
768 cbdata->psc_data = data;
769 list_add_tail(&cbdata->psc_item, &set->set_cblist);
774 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
775 struct ptlrpc_request *req)
777 /* The set takes over the caller's request reference */
778 LASSERT(list_empty(&req->rq_set_chain));
779 list_add_tail(&req->rq_set_chain, &set->set_requests);
781 atomic_inc(&set->set_remaining);
785 * Lock so many callers can add things, the context that owns the set
786 * is supposed to notice these and move them into the set proper.
788 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
789 struct ptlrpc_request *req)
791 struct ptlrpc_request_set *set = pc->pc_set;
794 * Let caller know that we stopped and will not handle this request.
795 * It needs to take care itself of request.
797 if (test_bit(LIOD_STOP, &pc->pc_flags))
800 spin_lock(&set->set_new_req_lock);
802 * The set takes over the caller's request reference.
804 LASSERT(list_empty(&req->rq_set_chain));
805 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
807 spin_unlock(&set->set_new_req_lock);
810 * Let thead know that we added something and better it to wake up
813 cfs_waitq_signal(&set->set_waitq);
818 * Based on the current state of the import, determine if the request
819 * can be sent, is an error, or should be delayed.
821 * Returns true if this request should be delayed. If false, and
822 * *status is set, then the request can not be sent and *status is the
823 * error code. If false and status is 0, then request can be sent.
825 * The imp->imp_lock must be held.
827 static int ptlrpc_import_delay_req(struct obd_import *imp,
828 struct ptlrpc_request *req, int *status)
833 LASSERT (status != NULL);
836 if (imp->imp_state == LUSTRE_IMP_NEW) {
837 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
840 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
841 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
843 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
844 imp->imp_state == LUSTRE_IMP_CONNECTING) {
845 /* allow CONNECT even if import is invalid */ ;
846 if (atomic_read(&imp->imp_inval_count) != 0) {
847 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
850 } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
851 imp->imp_obd->obd_no_recov) {
852 /* If the import has been invalidated (such as by an OST
853 * failure), and if the import(MGC) tried all of its connection
854 * list (Bug 13464), the request must fail with -ESHUTDOWN.
855 * This indicates the requests should be discarded; an -EIO
856 * may result in a resend of the request. */
857 if (!imp->imp_deactive)
858 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
859 *status = -ESHUTDOWN; /* bz 12940 */
860 } else if (req->rq_import_generation != imp->imp_generation) {
861 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
863 } else if (req->rq_send_state != imp->imp_state) {
864 /* invalidate in progress - any requests should be drop */
865 if (atomic_read(&imp->imp_inval_count) != 0) {
866 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
868 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
869 *status = -EWOULDBLOCK;
878 static int ptlrpc_check_reply(struct ptlrpc_request *req)
881 const char *what = "";
884 /* serialise with network callback */
885 spin_lock(&req->rq_lock);
887 if (ptlrpc_client_replied(req)) {
892 if (req->rq_net_err && !req->rq_timedout) {
894 spin_unlock(&req->rq_lock);
895 rc = ptlrpc_expire_one_request(req, 0);
896 spin_lock(&req->rq_lock);
905 if (req->rq_resend) {
910 if (req->rq_restart) {
915 if (ptlrpc_client_early(req)) {
917 ptlrpc_at_recv_early_reply(req);
918 GOTO(out, rc = 0); /* keep waiting */
923 spin_unlock(&req->rq_lock);
924 DEBUG_REQ(D_NET, req, "%src = %d for", what, rc);
928 /* Conditionally suppress specific console messages */
929 static int ptlrpc_console_allow(struct ptlrpc_request *req)
931 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
934 /* Suppress particular reconnect errors which are to be expected. No
935 * errors are suppressed for the initial connection on an import */
936 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
937 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
939 /* Suppress timed out reconnect requests */
940 if (req->rq_timedout)
943 /* Suppress unavailable/again reconnect requests */
944 err = lustre_msg_get_status(req->rq_repmsg);
945 if (err == -ENODEV || err == -EAGAIN)
952 static int ptlrpc_check_status(struct ptlrpc_request *req)
957 err = lustre_msg_get_status(req->rq_repmsg);
959 DEBUG_REQ(D_INFO, req, "status is %d", err);
960 } else if (err > 0) {
961 /* XXX: translate this error from net to host */
962 DEBUG_REQ(D_INFO, req, "status is %d", err);
965 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
966 struct obd_import *imp = req->rq_import;
967 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
969 if (ptlrpc_console_allow(req))
970 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
971 "communicating with %s. The %s "
972 "operation failed with %d\n",
974 imp->imp_connection->c_peer.nid),
975 ll_opcode2str(opc), err);
977 RETURN(err < 0 ? err : -EINVAL);
983 /* VBR: we should save pre-versions for replay*/
984 static void ptlrpc_save_versions(struct ptlrpc_request *req)
986 struct lustre_msg *repmsg = req->rq_repmsg;
987 struct lustre_msg *reqmsg = req->rq_reqmsg;
988 __u64 *versions = lustre_msg_get_versions(repmsg);
990 /* Interoperability with 1.6. This should be changed to LASSERT in HEAD */
991 if (versions == NULL)
994 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
997 lustre_msg_set_versions(reqmsg, versions);
998 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
999 versions[0], versions[1]);
1004 static int after_reply(struct ptlrpc_request *req)
1006 struct obd_import *imp = req->rq_import;
1007 struct obd_device *obd = req->rq_import->imp_obd;
1009 struct timeval work_start;
1013 /* repbuf must be unlinked */
1014 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1017 /* NB Until this point, the whole of the incoming message,
1018 * including buflens, status etc is in the sender's byte order. */
1020 if (req->rq_reply_truncate) {
1021 if (req->rq_no_resend) {
1022 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1023 " expected: %d, actual size: %d",
1024 req->rq_nob_received, req->rq_replen);
1028 OBD_FREE(req->rq_repbuf, req->rq_replen);
1029 req->rq_repbuf = NULL;
1030 /* Pass the required reply buffer size (include
1031 * space for early reply) */
1032 req->rq_replen = size_round(req->rq_nob_received);
1033 req->rq_nob_received = 0;
1035 spin_lock(&req->rq_lock);
1037 spin_unlock(&req->rq_lock);
1041 LASSERT ((char *)req->rq_repmsg + req->rq_nob_received <=
1042 (char *)req->rq_repbuf + req->rq_replen);
1043 rc = unpack_reply(req);
1047 do_gettimeofday(&work_start);
1048 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1049 if (obd->obd_svc_stats != NULL) {
1050 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1052 ptlrpc_lprocfs_rpc_sent(req, timediff);
1055 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1056 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1057 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1058 ptlrpc_at_adj_net_latency(req, lustre_msg_get_service_time(req->rq_repmsg));
1060 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1061 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1062 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1063 lustre_msg_get_type(req->rq_repmsg));
1067 rc = ptlrpc_check_status(req);
1069 /* Either we've been evicted, or the server has failed for
1070 * some reason. Try to reconnect, and if that fails, punt to
1072 if (ll_rpc_recoverable_error(rc)) {
1073 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1074 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1077 ptlrpc_request_handle_notconn(req);
1081 /* Let's look if server sent slv. Do it only for RPC with
1083 ldlm_cli_update_pool(req);
1086 /* Store transno in reqmsg for replay. */
1087 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1088 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1089 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1092 if (imp->imp_replayable) {
1093 spin_lock(&imp->imp_lock);
1094 /* no point in adding already-committed requests to the replay
1095 * list, we will just remove them immediately. b=9829 */
1096 if (req->rq_transno != 0 &&
1098 lustre_msg_get_last_committed(req->rq_repmsg) ||
1100 /* version recovery */
1101 ptlrpc_save_versions(req);
1102 ptlrpc_retain_replayable_request(req, imp);
1103 } else if (req->rq_commit_cb != NULL) {
1104 spin_unlock(&imp->imp_lock);
1105 req->rq_commit_cb(req);
1106 spin_lock(&imp->imp_lock);
1109 /* Replay-enabled imports return commit-status information. */
1110 if (lustre_msg_get_last_committed(req->rq_repmsg))
1111 imp->imp_peer_committed_transno =
1112 lustre_msg_get_last_committed(req->rq_repmsg);
1113 ptlrpc_free_committed(imp);
1115 if (req->rq_transno > imp->imp_peer_committed_transno)
1116 ptlrpc_pinger_commit_expected(imp);
1117 spin_unlock(&imp->imp_lock);
1123 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1125 struct obd_import *imp;
1129 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1130 if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
1133 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1135 imp = req->rq_import;
1136 spin_lock(&imp->imp_lock);
1138 req->rq_import_generation = imp->imp_generation;
1140 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1141 spin_lock(&req->rq_lock);
1142 req->rq_waiting = 1;
1143 spin_unlock(&req->rq_lock);
1145 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1146 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1147 ptlrpc_import_state_name(req->rq_send_state),
1148 ptlrpc_import_state_name(imp->imp_state));
1149 LASSERT(list_empty(&req->rq_list));
1150 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1151 atomic_inc(&req->rq_import->imp_inflight);
1152 spin_unlock(&imp->imp_lock);
1157 spin_unlock(&imp->imp_lock);
1158 req->rq_status = rc;
1159 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1163 LASSERT(list_empty(&req->rq_list));
1164 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1165 atomic_inc(&req->rq_import->imp_inflight);
1166 spin_unlock(&imp->imp_lock);
1168 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1169 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1170 " %s:%s:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1171 imp->imp_obd->obd_uuid.uuid,
1172 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1173 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1174 lustre_msg_get_opc(req->rq_reqmsg));
1176 rc = ptl_send_rpc(req, 0);
1178 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1180 spin_lock(&req->rq_lock);
1181 req->rq_net_err = 1;
1182 spin_unlock(&req->rq_lock);
1188 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1189 int ptlrpc_check_set(struct ptlrpc_request_set *set)
1191 struct list_head *tmp;
1192 int force_timer_recalc = 0;
1195 if (atomic_read(&set->set_remaining) == 0)
1198 list_for_each(tmp, &set->set_requests) {
1199 struct ptlrpc_request *req =
1200 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1201 struct obd_import *imp = req->rq_import;
1202 int unregistered = 0;
1205 if (req->rq_phase == RQ_PHASE_NEW &&
1206 ptlrpc_send_new_req(req)) {
1207 force_timer_recalc = 1;
1210 /* delayed send - skip */
1211 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1214 if (!(req->rq_phase == RQ_PHASE_RPC ||
1215 req->rq_phase == RQ_PHASE_BULK ||
1216 req->rq_phase == RQ_PHASE_INTERPRET ||
1217 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1218 req->rq_phase == RQ_PHASE_COMPLETE)) {
1219 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1223 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1224 LASSERT(req->rq_next_phase != req->rq_phase);
1225 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1227 /* Skip processing until reply is unlinked. We
1228 * can't return to pool before that and we can't
1229 * call interpret before that. We need to make
1230 * sure that all rdma transfers finished and will
1231 * not corrupt any data. */
1232 if (ptlrpc_client_recv_or_unlink(req) ||
1233 ptlrpc_client_bulk_active(req))
1236 /* Turn repl fail_loc off to prevent it from looping
1238 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1239 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK |
1243 /* Turn off bulk fail_loc. */
1244 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1245 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK |
1249 /* Move to next phase if reply was successfully
1251 ptlrpc_rqphase_move(req, req->rq_next_phase);
1254 if (req->rq_phase == RQ_PHASE_COMPLETE)
1257 if (req->rq_phase == RQ_PHASE_INTERPRET)
1258 GOTO(interpret, req->rq_status);
1260 /* Note that this also will start async reply unlink */
1261 if (req->rq_net_err && !req->rq_timedout) {
1262 ptlrpc_expire_one_request(req, 1);
1264 /* Check if we still need to wait for unlink. */
1265 if (ptlrpc_client_recv_or_unlink(req) ||
1266 ptlrpc_client_bulk_active(req))
1271 if (req->rq_status == 0)
1272 req->rq_status = -EIO;
1273 GOTO(interpret, req->rq_status);
1276 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1277 * will only be set after rq_timedout, but the oig waiting
1278 * path sets rq_intr irrespective of whether ptlrpcd has
1279 * seen a timeout. our policy is to only interpret
1280 * interrupted rpcs after they have timed out */
1281 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
1282 req->rq_status = -EINTR;
1283 GOTO(interpret, req->rq_status);
1286 if (req->rq_phase == RQ_PHASE_RPC) {
1287 if (req->rq_timedout||req->rq_waiting||req->rq_resend) {
1290 if (!ptlrpc_unregister_reply(req, 1))
1293 spin_lock(&imp->imp_lock);
1294 if (ptlrpc_import_delay_req(imp, req, &status)){
1295 /* put on delay list - only if we wait
1296 * recovery finished - before send */
1297 list_del_init(&req->rq_list);
1298 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1299 spin_unlock(&imp->imp_lock);
1304 req->rq_status = status;
1305 spin_unlock(&imp->imp_lock);
1306 GOTO(interpret, req->rq_status);
1308 if (req->rq_no_resend) {
1309 req->rq_status = -ENOTCONN;
1310 spin_unlock(&imp->imp_lock);
1311 GOTO(interpret, req->rq_status);
1314 list_del_init(&req->rq_list);
1315 list_add_tail(&req->rq_list,
1316 &imp->imp_sending_list);
1318 spin_unlock(&imp->imp_lock);
1320 spin_lock(&req->rq_lock);
1321 req->rq_waiting = 0;
1322 spin_unlock(&req->rq_lock);
1324 if (req->rq_timedout||req->rq_resend) {
1325 /* This is re-sending anyways,
1326 * let's mark req as resend. */
1327 spin_lock(&req->rq_lock);
1329 spin_unlock(&req->rq_lock);
1333 if (!ptlrpc_unregister_bulk(req, 1))
1336 /* ensure previous bulk fails */
1337 old_xid = req->rq_xid;
1338 req->rq_xid = ptlrpc_next_xid();
1339 CDEBUG(D_HA, "resend bulk "
1342 old_xid, req->rq_xid);
1346 rc = ptl_send_rpc(req, 0);
1348 DEBUG_REQ(D_HA, req, "send failed (%d)",
1350 force_timer_recalc = 1;
1352 spin_lock(&req->rq_lock);
1353 req->rq_net_err = 1;
1354 spin_unlock(&req->rq_lock);
1356 /* need to reset the timeout */
1357 force_timer_recalc = 1;
1360 spin_lock(&req->rq_lock);
1362 if (ptlrpc_client_early(req)) {
1363 ptlrpc_at_recv_early_reply(req);
1364 spin_unlock(&req->rq_lock);
1368 /* Still waiting for a reply? */
1369 if (ptlrpc_client_recv(req)) {
1370 spin_unlock(&req->rq_lock);
1374 /* Did we actually receive a reply? */
1375 if (!ptlrpc_client_replied(req)) {
1376 spin_unlock(&req->rq_lock);
1380 spin_unlock(&req->rq_lock);
1382 /* unlink from net because we are going to
1383 * swab in-place of reply buffer */
1384 unregistered = ptlrpc_unregister_reply(req, 1);
1388 req->rq_status = after_reply(req);
1392 /* If there is no bulk associated with this request,
1393 * then we're done and should let the interpreter
1394 * process the reply. Similarly if the RPC returned
1395 * an error, and therefore the bulk will never arrive.
1397 if (req->rq_bulk == NULL || req->rq_status != 0)
1398 GOTO(interpret, req->rq_status);
1400 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1403 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1404 if (ptlrpc_client_bulk_active(req))
1407 if (!req->rq_bulk->bd_success) {
1408 /* The RPC reply arrived OK, but the bulk screwed
1409 * up! Dead wierd since the server told us the RPC
1410 * was good after getting the REPLY for her GET or
1411 * the ACK for her PUT. */
1412 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1413 req->rq_status = -EIO;
1414 GOTO(interpret, req->rq_status);
1417 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1419 /* This moves to "unregistering" phase we need to wait for
1421 if (!unregistered && !ptlrpc_unregister_reply(req, 1))
1424 if (!ptlrpc_unregister_bulk(req, 1))
1427 /* When calling interpret receiving already should be
1429 LASSERT(!req->rq_receiving_reply);
1431 ptlrpc_req_interpret(req, req->rq_status);
1433 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1435 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1436 "opc %s:%s:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1437 imp->imp_obd->obd_uuid.uuid,
1438 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1440 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1441 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1443 spin_lock(&imp->imp_lock);
1444 /* Request already may be not on sending or delaying list. This
1445 * may happen in the case of marking it errorneous for the case
1446 * ptlrpc_import_delay_req(req, status) find it impossible to
1447 * allow sending this rpc and returns *status != 0. */
1448 if (!list_empty(&req->rq_list)) {
1449 list_del_init(&req->rq_list);
1450 atomic_dec(&imp->imp_inflight);
1452 spin_unlock(&imp->imp_lock);
1454 atomic_dec(&set->set_remaining);
1455 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1458 /* If we hit an error, we want to recover promptly. */
1459 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1462 /* Return 1 if we should give up, else 0 */
1463 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1465 struct obd_import *imp = req->rq_import;
1469 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req,
1470 "Request x"LPU64" sent from %s to NID %s"
1471 " %lus ago has %s (%lds prior to deadline).\n", req->rq_xid,
1472 imp ? imp->imp_obd->obd_name : "<?>",
1473 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1474 cfs_time_current_sec() - req->rq_sent,
1475 req->rq_net_err ? "failed due to network error" : "timed out",
1476 req->rq_deadline - req->rq_sent);
1478 if (imp != NULL && obd_debug_peer_on_timeout)
1479 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1481 spin_lock(&req->rq_lock);
1482 req->rq_timedout = 1;
1483 spin_unlock(&req->rq_lock);
1485 ptlrpc_unregister_reply(req, async_unlink);
1486 ptlrpc_unregister_bulk(req, async_unlink);
1488 if (obd_dump_on_timeout)
1489 libcfs_debug_dumplog();
1492 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1499 atomic_inc(&imp->imp_timeouts);
1501 /* The DLM server doesn't want recovery run on its imports. */
1502 if (imp->imp_dlm_fake)
1505 /* If this request is for recovery or other primordial tasks,
1506 * then error it out here. */
1507 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1508 imp->imp_obd->obd_no_recov) {
1509 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1510 ptlrpc_import_state_name(req->rq_send_state),
1511 ptlrpc_import_state_name(imp->imp_state));
1512 spin_lock(&req->rq_lock);
1513 req->rq_status = -ETIMEDOUT;
1515 spin_unlock(&req->rq_lock);
1519 /* if a request can't be resent we can't wait for an answer after
1521 if (req->rq_no_resend) {
1522 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1526 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1531 int ptlrpc_expired_set(void *data)
1533 struct ptlrpc_request_set *set = data;
1534 struct list_head *tmp;
1535 time_t now = cfs_time_current_sec();
1538 LASSERT(set != NULL);
1540 /* A timeout expired; see which reqs it applies to... */
1541 list_for_each(tmp, &set->set_requests) {
1542 struct ptlrpc_request *req =
1543 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1545 /* Request in-flight? */
1546 if (!((req->rq_phase == RQ_PHASE_RPC &&
1547 !req->rq_waiting && !req->rq_resend) ||
1548 (req->rq_phase == RQ_PHASE_BULK)))
1551 if (req->rq_timedout || /* already dealt with */
1552 req->rq_deadline > now) /* not expired */
1555 /* Deal with this guy. Do it asynchronously to not block
1556 * ptlrpcd thread. */
1557 ptlrpc_expire_one_request(req, 1);
1560 /* When waiting for a whole set, we always to break out of the
1561 * sleep so we can recalculate the timeout, or enable interrupts
1562 * if everyone's timed out. */
1566 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1568 spin_lock(&req->rq_lock);
1570 spin_unlock(&req->rq_lock);
1573 void ptlrpc_interrupted_set(void *data)
1575 struct ptlrpc_request_set *set = data;
1576 struct list_head *tmp;
1578 LASSERT(set != NULL);
1579 CERROR("INTERRUPTED SET %p\n", set);
1581 list_for_each(tmp, &set->set_requests) {
1582 struct ptlrpc_request *req =
1583 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1585 if (req->rq_phase != RQ_PHASE_RPC &&
1586 req->rq_phase != RQ_PHASE_UNREGISTERING)
1589 ptlrpc_mark_interrupted(req);
1593 /* get the smallest timeout in the set; this does NOT set a timeout. */
1594 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1596 struct list_head *tmp;
1597 time_t now = cfs_time_current_sec();
1599 struct ptlrpc_request *req;
1603 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1605 list_for_each(tmp, &set->set_requests) {
1606 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1608 /* request in-flight? */
1609 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1610 (req->rq_phase == RQ_PHASE_BULK) ||
1611 (req->rq_phase == RQ_PHASE_NEW)))
1614 /* Already timed out. */
1615 if (req->rq_timedout)
1618 if (req->rq_phase == RQ_PHASE_NEW)
1619 deadline = req->rq_sent; /* delayed send */
1621 deadline = req->rq_deadline;
1623 if (deadline <= now) { /* actually expired already */
1624 timeout = 1; /* ASAP */
1627 if ((timeout == 0) || (timeout > (deadline - now)))
1628 timeout = deadline - now;
1633 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1635 struct list_head *tmp;
1636 struct ptlrpc_request *req;
1637 struct l_wait_info lwi;
1641 if (list_empty(&set->set_requests))
1644 list_for_each(tmp, &set->set_requests) {
1645 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1646 if (req->rq_phase == RQ_PHASE_NEW)
1647 (void)ptlrpc_send_new_req(req);
1651 timeout = ptlrpc_set_next_timeout(set);
1653 /* wait until all complete, interrupted, or an in-flight
1655 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1657 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1659 ptlrpc_interrupted_set, set);
1660 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1662 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1664 /* -EINTR => all requests have been flagged rq_intr so next
1666 * -ETIMEOUTD => someone timed out. When all reqs have
1667 * timed out, signals are enabled allowing completion with
1669 * I don't really care if we go once more round the loop in
1670 * the error cases -eeb. */
1671 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
1672 list_for_each(tmp, &set->set_requests) {
1673 req = list_entry(tmp, struct ptlrpc_request,
1675 spin_lock(&req->rq_lock);
1676 req->rq_invalid_rqset = 1;
1677 spin_unlock(&req->rq_lock);
1680 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
1682 LASSERT(atomic_read(&set->set_remaining) == 0);
1685 list_for_each(tmp, &set->set_requests) {
1686 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1688 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1689 if (req->rq_status != 0)
1690 rc = req->rq_status;
1693 if (set->set_interpret != NULL) {
1694 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1696 rc = interpreter (set, set->set_arg, rc);
1698 struct ptlrpc_set_cbdata *cbdata, *n;
1701 list_for_each_entry_safe(cbdata, n,
1702 &set->set_cblist, psc_item) {
1703 list_del_init(&cbdata->psc_item);
1704 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1707 OBD_SLAB_FREE(cbdata, ptlrpc_cbdata_slab,
1715 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
1717 struct ptlrpc_request_pool *pool = request->rq_pool;
1719 spin_lock(&pool->prp_lock);
1720 LASSERT(list_empty(&request->rq_list));
1721 LASSERT(!request->rq_receiving_reply);
1722 list_add_tail(&request->rq_list, &pool->prp_req_list);
1723 spin_unlock(&pool->prp_lock);
1726 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1729 if (request == NULL) {
1734 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1735 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1736 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1737 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1738 LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request);
1739 LASSERTF(!request->rq_replay, "req %p\n", request);
1741 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1742 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1743 if (request->rq_import != NULL) {
1745 spin_lock(&request->rq_import->imp_lock);
1746 list_del_init(&request->rq_replay_list);
1748 spin_unlock(&request->rq_import->imp_lock);
1750 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1752 if (atomic_read(&request->rq_refcount) != 0) {
1753 DEBUG_REQ(D_ERROR, request,
1754 "freeing request with nonzero refcount");
1758 if (request->rq_repbuf != NULL) {
1759 OBD_FREE(request->rq_repbuf, request->rq_replen);
1760 request->rq_repbuf = NULL;
1761 request->rq_repmsg = NULL;
1763 if (request->rq_export != NULL) {
1764 class_export_put(request->rq_export);
1765 request->rq_export = NULL;
1767 if (request->rq_import != NULL) {
1768 class_import_put(request->rq_import);
1769 request->rq_import = NULL;
1771 if (request->rq_bulk != NULL)
1772 ptlrpc_free_bulk(request->rq_bulk);
1774 if (request->rq_pool) {
1775 __ptlrpc_free_req_to_pool(request);
1777 if (request->rq_reqmsg != NULL) {
1778 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1779 request->rq_reqmsg = NULL;
1781 OBD_FREE(request, sizeof(*request));
1786 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1787 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1789 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1790 (void)__ptlrpc_req_finished(request, 1);
1793 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1796 if (request == NULL)
1799 if (request == LP_POISON ||
1800 request->rq_reqmsg == LP_POISON) {
1801 CERROR("dereferencing freed request (bug 575)\n");
1806 DEBUG_REQ(D_INFO, request, "refcount now %u",
1807 atomic_read(&request->rq_refcount) - 1);
1809 if (atomic_dec_and_test(&request->rq_refcount)) {
1810 __ptlrpc_free_req(request, locked);
1817 void ptlrpc_req_finished(struct ptlrpc_request *request)
1819 __ptlrpc_req_finished(request, 0);
1822 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1824 return request->rq_xid;
1826 EXPORT_SYMBOL(ptlrpc_req_xid);
1828 /* Disengage the client's reply buffer from the network
1829 * NB does _NOT_ unregister any client-side bulk.
1830 * IDEMPOTENT, but _not_ safe against concurrent callers.
1831 * The request owner (i.e. the thread doing the I/O) must call...
1833 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1837 struct l_wait_info lwi;
1841 LASSERT(!in_interrupt());
1843 /* Let's setup deadline for reply unlink. */
1844 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1845 async && request->rq_reply_deadline == 0)
1846 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1848 /* Nothing left to do. */
1849 if (!ptlrpc_client_recv_or_unlink(request))
1852 LNetMDUnlink(request->rq_reply_md_h);
1854 /* Let's check it once again. */
1855 if (!ptlrpc_client_recv_or_unlink(request))
1858 /* Move to "Unregistering" phase as reply was not unlinked yet. */
1859 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1861 /* Do not wait for unlink to finish. */
1865 /* We have to l_wait_event() whatever the result, to give liblustre
1866 * a chance to run reply_in_callback(), and to make sure we've
1867 * unlinked before returning a req to the pool */
1868 if (request->rq_set != NULL)
1869 wq = &request->rq_set->set_waitq;
1871 wq = &request->rq_reply_waitq;
1874 /* Network access will complete in finite time but the HUGE
1875 * timeout lets us CWARN for visibility of sluggish NALs */
1876 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1877 cfs_time_seconds(1), NULL, NULL);
1878 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1881 ptlrpc_rqphase_move(request, request->rq_next_phase);
1885 LASSERT(rc == -ETIMEDOUT);
1886 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1887 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1888 request->rq_must_unlink);
1893 /* caller must hold imp->imp_lock */
1894 void ptlrpc_free_committed(struct obd_import *imp)
1896 struct list_head *tmp, *saved;
1897 struct ptlrpc_request *req;
1898 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1901 LASSERT(imp != NULL);
1903 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1906 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1907 imp->imp_generation == imp->imp_last_generation_checked) {
1908 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
1909 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1914 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1915 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1916 imp->imp_generation);
1917 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1918 imp->imp_last_generation_checked = imp->imp_generation;
1920 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1921 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1923 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1924 LASSERT(req != last_req);
1927 if (req->rq_import_generation < imp->imp_generation) {
1928 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1932 if (req->rq_replay) {
1933 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1937 /* not yet committed */
1938 if (req->rq_transno > imp->imp_peer_committed_transno) {
1939 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1943 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
1944 imp->imp_peer_committed_transno);
1946 spin_lock(&req->rq_lock);
1948 spin_unlock(&req->rq_lock);
1949 if (req->rq_commit_cb != NULL)
1950 req->rq_commit_cb(req);
1951 list_del_init(&req->rq_replay_list);
1952 __ptlrpc_req_finished(req, 1);
1959 void ptlrpc_cleanup_client(struct obd_import *imp)
1966 void ptlrpc_resend_req(struct ptlrpc_request *req)
1968 DEBUG_REQ(D_HA, req, "going to resend");
1969 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1970 req->rq_status = -EAGAIN;
1972 spin_lock(&req->rq_lock);
1974 req->rq_net_err = 0;
1975 req->rq_timedout = 0;
1977 __u64 old_xid = req->rq_xid;
1979 /* ensure previous bulk fails */
1980 req->rq_xid = ptlrpc_next_xid();
1981 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1982 old_xid, req->rq_xid);
1984 ptlrpc_client_wake_req(req);
1985 spin_unlock(&req->rq_lock);
1988 /* XXX: this function and rq_status are currently unused */
1989 void ptlrpc_restart_req(struct ptlrpc_request *req)
1991 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1992 req->rq_status = -ERESTARTSYS;
1994 spin_lock(&req->rq_lock);
1995 req->rq_restart = 1;
1996 req->rq_timedout = 0;
1997 ptlrpc_client_wake_req(req);
1998 spin_unlock(&req->rq_lock);
2001 static void interrupted_request(void *data)
2003 struct ptlrpc_request *req = data;
2004 DEBUG_REQ(D_HA, req, "request interrupted");
2005 spin_lock(&req->rq_lock);
2007 spin_unlock(&req->rq_lock);
2010 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2013 atomic_inc(&req->rq_refcount);
2017 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2018 struct obd_import *imp)
2020 struct list_head *tmp;
2022 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2024 /* clear this for new requests that were resent as well
2025 as resent replayed requests. */
2026 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2028 /* don't re-add requests that have been replayed */
2029 if (!list_empty(&req->rq_replay_list))
2032 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2034 LASSERT(imp->imp_replayable);
2035 /* Balanced in ptlrpc_free_committed, usually. */
2036 ptlrpc_request_addref(req);
2037 list_for_each_prev(tmp, &imp->imp_replay_list) {
2038 struct ptlrpc_request *iter =
2039 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
2041 /* We may have duplicate transnos if we create and then
2042 * open a file, or for closes retained if to match creating
2043 * opens, so use req->rq_xid as a secondary key.
2044 * (See bugs 684, 685, and 428.)
2045 * XXX no longer needed, but all opens need transnos!
2047 if (iter->rq_transno > req->rq_transno)
2050 if (iter->rq_transno == req->rq_transno) {
2051 LASSERT(iter->rq_xid != req->rq_xid);
2052 if (iter->rq_xid > req->rq_xid)
2056 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2060 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2063 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2067 struct l_wait_info lwi;
2068 struct obd_import *imp = req->rq_import;
2069 cfs_duration_t timeout = CFS_TICK;
2073 LASSERT(req->rq_set == NULL);
2074 LASSERT(!req->rq_receiving_reply);
2076 /* for distributed debugging */
2077 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2078 LASSERT(imp->imp_obd != NULL);
2079 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
2080 "%s:%s:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
2081 imp->imp_obd->obd_uuid.uuid,
2082 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2083 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2084 lustre_msg_get_opc(req->rq_reqmsg));
2086 /* Mark phase here for a little debug help */
2087 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
2089 spin_lock(&imp->imp_lock);
2090 req->rq_import_generation = imp->imp_generation;
2092 if (ptlrpc_import_delay_req(imp, req, &rc)) {
2093 list_del_init(&req->rq_list);
2094 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
2095 atomic_inc(&imp->imp_inflight);
2096 spin_unlock(&imp->imp_lock);
2098 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
2100 ptlrpc_import_state_name(req->rq_send_state),
2101 ptlrpc_import_state_name(imp->imp_state));
2102 lwi = LWI_INTR(interrupted_request, req);
2103 rc = l_wait_event(req->rq_reply_waitq,
2104 (req->rq_send_state == imp->imp_state ||
2105 req->rq_err || req->rq_intr),
2107 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2109 ptlrpc_import_state_name(imp->imp_state),
2110 ptlrpc_import_state_name(req->rq_send_state),
2111 req->rq_err, req->rq_intr);
2113 spin_lock(&imp->imp_lock);
2114 list_del_init(&req->rq_list);
2115 atomic_dec(&imp->imp_inflight);
2118 /* rq_status was set locally */
2119 rc = req->rq_status ? req->rq_status : -EIO;
2121 else if (req->rq_intr) {
2124 else if (req->rq_no_resend) {
2133 spin_unlock(&imp->imp_lock);
2134 req->rq_status = rc; // XXX this ok?
2138 if (req->rq_resend) {
2139 if (req->rq_bulk != NULL) {
2140 ptlrpc_unregister_bulk(req, 0);
2142 /* bulk requests are supposed to be
2143 * idempotent, so we are free to bump the xid
2144 * here, which we need to do before
2145 * registering the bulk again (bug 6371).
2146 * print the old xid first for sanity.
2148 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2149 req->rq_xid = ptlrpc_next_xid();
2152 DEBUG_REQ(D_HA, req, "resending: ");
2155 /* XXX this is the same as ptlrpc_set_wait */
2156 LASSERT(list_empty(&req->rq_list));
2157 list_add_tail(&req->rq_list, &imp->imp_sending_list);
2158 atomic_inc(&imp->imp_inflight);
2159 spin_unlock(&imp->imp_lock);
2161 rc = ptl_send_rpc(req, 0);
2163 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2165 timeoutl = req->rq_deadline - cfs_time_current_sec();
2166 timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2167 cfs_time_seconds(timeoutl);
2168 DEBUG_REQ(D_NET, req,
2169 "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2170 lwi = LWI_TIMEOUT_INTR(timeout, NULL, interrupted_request, req);
2171 brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req),
2173 /* Wait again if we changed deadline */
2174 } while ((brc == -ETIMEDOUT) &&
2175 (req->rq_deadline > cfs_time_current_sec()));
2177 if ((brc == -ETIMEDOUT) && !ptlrpc_expire_one_request(req, 0)) {
2178 /* Wait forever for reconnect / replay or failure */
2179 lwi = LWI_INTR(interrupted_request, req);
2180 brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req),
2184 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2185 "%s:%s:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
2186 imp->imp_obd->obd_uuid.uuid,
2187 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2188 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2189 lustre_msg_get_opc(req->rq_reqmsg));
2191 /* If the reply was received normally, this just grabs the spinlock
2192 * (ensuring the reply callback has returned), sees that
2193 * req->rq_receiving_reply is clear and returns. */
2194 ptlrpc_unregister_reply(req, 0);
2196 spin_lock(&imp->imp_lock);
2197 list_del_init(&req->rq_list);
2198 atomic_dec(&imp->imp_inflight);
2199 spin_unlock(&imp->imp_lock);
2202 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2203 rc, req->rq_status);
2204 rc = rc ? rc : req->rq_status;
2205 GOTO(out, rc = rc ? rc : -EIO);
2209 /* Should only be interrupted if we timed out. */
2210 if (!req->rq_timedout)
2211 DEBUG_REQ(D_ERROR, req,
2212 "rq_intr set but rq_timedout not");
2213 GOTO(out, rc = -EINTR);
2216 /* Resend if we need to */
2217 if (req->rq_resend||req->rq_timedout) {
2218 /* ...unless we were specifically told otherwise. */
2219 if (req->rq_no_resend)
2220 GOTO(out, rc = -ETIMEDOUT);
2221 spin_lock(&imp->imp_lock);
2222 /* we can have rq_timeout on dlm fake import which not support
2223 * recovery - but me need resend request on this import instead
2224 * of return error */
2225 spin_lock(&req->rq_lock);
2227 spin_unlock(&req->rq_lock);
2231 if (!ptlrpc_client_replied(req)) {
2232 /* How can this be? -eeb */
2233 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2235 GOTO(out, rc = req->rq_status);
2238 rc = after_reply(req);
2239 /* NB may return +ve success rc */
2240 if (req->rq_resend) {
2241 spin_lock(&imp->imp_lock);
2246 if (req->rq_bulk != NULL) {
2248 /* success so far. Note that anything going wrong
2249 * with bulk now, is EXTREMELY strange, since the
2250 * server must have believed that the bulk
2251 * tranferred OK before she replied with success to
2253 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2254 brc = l_wait_event(req->rq_reply_waitq,
2255 !ptlrpc_client_bulk_active(req),
2257 LASSERT(brc == 0 || brc == -ETIMEDOUT);
2259 LASSERT(brc == -ETIMEDOUT);
2260 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2262 } else if (!req->rq_bulk->bd_success) {
2263 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2268 ptlrpc_unregister_bulk(req, 0);
2271 LASSERT(!req->rq_receiving_reply);
2272 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2273 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
2277 struct ptlrpc_replay_async_args {
2279 int praa_old_status;
2282 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
2283 void * data, int rc)
2285 struct ptlrpc_replay_async_args *aa = data;
2286 struct obd_import *imp = req->rq_import;
2289 atomic_dec(&imp->imp_replay_inflight);
2291 if (!ptlrpc_client_replied(req)) {
2292 CERROR("request replay timed out, restarting recovery\n");
2293 GOTO(out, rc = -ETIMEDOUT);
2296 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2297 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2298 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2299 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2301 /* VBR: check version failure */
2302 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2303 /* replay was failed due to version mismatch */
2304 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2305 spin_lock(&imp->imp_lock);
2306 imp->imp_vbr_failed = 1;
2307 spin_unlock(&imp->imp_lock);
2309 /* The transno had better not change over replay. */
2310 if (unlikely(lustre_msg_get_transno(req->rq_reqmsg) !=
2311 lustre_msg_get_transno(req->rq_repmsg) &&
2312 lustre_msg_get_transno(req->rq_repmsg) != 0)) {
2313 DEBUG_REQ(D_ERROR, req, "Transno has changed over "
2314 "replay ("LPU64"/"LPU64")\n",
2315 lustre_msg_get_transno(req->rq_reqmsg),
2316 lustre_msg_get_transno(req->rq_repmsg));
2321 spin_lock(&imp->imp_lock);
2322 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2323 spin_unlock(&imp->imp_lock);
2324 LASSERT(imp->imp_last_replay_transno);
2326 DEBUG_REQ(D_HA, req, "got rep");
2328 /* let the callback do fixups, possibly including in the request */
2329 if (req->rq_replay_cb)
2330 req->rq_replay_cb(req);
2332 if (ptlrpc_client_replied(req) &&
2333 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2334 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2335 lustre_msg_get_status(req->rq_repmsg),
2336 aa->praa_old_status);
2338 /* Put it back for re-replay. */
2339 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2342 /* continue with recovery */
2343 rc = ptlrpc_import_recovery_state_machine(imp);
2345 req->rq_send_state = aa->praa_old_state;
2348 /* this replay failed, so restart recovery */
2349 ptlrpc_connect_import(imp, NULL);
2355 int ptlrpc_replay_req(struct ptlrpc_request *req)
2357 struct ptlrpc_replay_async_args *aa;
2360 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2362 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2363 aa = ptlrpc_req_async_args(req);
2364 memset(aa, 0, sizeof *aa);
2366 /* Prepare request to be resent with ptlrpcd */
2367 aa->praa_old_state = req->rq_send_state;
2368 req->rq_send_state = LUSTRE_IMP_REPLAY;
2369 req->rq_phase = RQ_PHASE_NEW;
2370 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2372 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2374 req->rq_interpret_reply = ptlrpc_replay_interpret;
2375 /* Readjust the timeout for current conditions */
2376 ptlrpc_at_set_req_timeout(req);
2378 DEBUG_REQ(D_HA, req, "REPLAY");
2380 atomic_inc(&req->rq_import->imp_replay_inflight);
2381 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2383 ptlrpcd_add_req(req);
2387 void ptlrpc_abort_inflight(struct obd_import *imp)
2389 struct list_head *tmp, *n;
2392 /* Make sure that no new requests get processed for this import.
2393 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2394 * this flag and then putting requests on sending_list or delayed_list.
2396 spin_lock(&imp->imp_lock);
2398 /* XXX locking? Maybe we should remove each request with the list
2399 * locked? Also, how do we know if the requests on the list are
2400 * being freed at this time?
2402 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2403 struct ptlrpc_request *req =
2404 list_entry(tmp, struct ptlrpc_request, rq_list);
2406 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2408 spin_lock (&req->rq_lock);
2409 if (req->rq_import_generation < imp->imp_generation) {
2411 req->rq_status = -EINTR;
2412 ptlrpc_client_wake_req(req);
2414 spin_unlock (&req->rq_lock);
2417 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2418 struct ptlrpc_request *req =
2419 list_entry(tmp, struct ptlrpc_request, rq_list);
2421 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2423 spin_lock (&req->rq_lock);
2424 if (req->rq_import_generation < imp->imp_generation) {
2426 req->rq_status = -EINTR;
2427 ptlrpc_client_wake_req(req);
2429 spin_unlock (&req->rq_lock);
2432 /* Last chance to free reqs left on the replay list, but we
2433 * will still leak reqs that haven't committed. */
2434 if (imp->imp_replayable)
2435 ptlrpc_free_committed(imp);
2437 spin_unlock(&imp->imp_lock);
2442 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2444 struct list_head *tmp, *pos;
2446 LASSERT(set != NULL);
2448 list_for_each_safe(pos, tmp, &set->set_requests) {
2449 struct ptlrpc_request *req =
2450 list_entry(pos, struct ptlrpc_request, rq_set_chain);
2452 spin_lock(&req->rq_lock);
2453 if (req->rq_phase != RQ_PHASE_RPC) {
2454 spin_unlock(&req->rq_lock);
2459 req->rq_status = -EINTR;
2460 ptlrpc_client_wake_req(req);
2461 spin_unlock(&req->rq_lock);
2465 static __u64 ptlrpc_last_xid;
2466 static spinlock_t ptlrpc_last_xid_lock;
2468 /* Initialize the XID for the node. This is common among all requests on
2469 * this node, and only requires the property that it is monotonically
2470 * increasing. It does not need to be sequential. Since this is also used
2471 * as the RDMA match bits, it is important that a single client NOT have
2472 * the same match bits for two different in-flight requests, hence we do
2473 * NOT want to have an XID per target or similar.
2475 * To avoid an unlikely collision between match bits after a client reboot
2476 * (which would deliver old data into the wrong RDMA buffer) we initialize
2477 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2478 * If the time is clearly incorrect, we instead use a 62-bit random number.
2479 * In the worst case the random number will overflow 1M RPCs per second in
2480 * 9133 years, or permutations thereof.
2482 #define YEAR_2004 (1ULL << 30)
2483 void ptlrpc_init_xid(void)
2485 time_t now = cfs_time_current_sec();
2487 spin_lock_init(&ptlrpc_last_xid_lock);
2488 if (now < YEAR_2004) {
2489 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2490 ptlrpc_last_xid >>= 2;
2491 ptlrpc_last_xid |= (1ULL << 61);
2493 ptlrpc_last_xid = (__u64)now << 20;
2497 __u64 ptlrpc_next_xid(void)
2500 spin_lock(&ptlrpc_last_xid_lock);
2501 tmp = ++ptlrpc_last_xid;
2502 spin_unlock(&ptlrpc_last_xid_lock);
2506 __u64 ptlrpc_sample_next_xid(void)
2508 if (sizeof(long) < 8) {
2509 /* need to avoid possible word tearing on 32-bit systems */
2511 spin_lock(&ptlrpc_last_xid_lock);
2512 tmp = ptlrpc_last_xid + 1;
2513 spin_unlock(&ptlrpc_last_xid_lock);
2516 /* No need to lock, since returned value is racy anyways */
2517 return ptlrpc_last_xid + 1;
2519 EXPORT_SYMBOL(ptlrpc_sample_next_xid);