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 2008 Sun Microsystems, Inc. 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>
49 #include <lustre_req_layout.h>
51 #include "ptlrpc_internal.h"
53 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
54 struct ptlrpc_client *cl)
56 cl->cli_request_portal = req_portal;
57 cl->cli_reply_portal = rep_portal;
61 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
63 struct ptlrpc_connection *c;
65 lnet_process_id_t peer;
68 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
70 CERROR("cannot find peer %s!\n", uuid->uuid);
74 c = ptlrpc_connection_get(peer, self, uuid);
76 memcpy(c->c_remote_uuid.uuid,
77 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
80 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
85 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
87 struct ptlrpc_bulk_desc *desc;
89 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
93 spin_lock_init(&desc->bd_lock);
94 cfs_waitq_init(&desc->bd_waitq);
95 desc->bd_max_iov = npages;
96 desc->bd_iov_count = 0;
97 LNetInvalidateHandle(&desc->bd_md_h);
98 desc->bd_portal = portal;
104 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
105 int npages, int type, int portal)
107 struct obd_import *imp = req->rq_import;
108 struct ptlrpc_bulk_desc *desc;
111 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
112 desc = new_bulk(npages, type, portal);
116 desc->bd_import_generation = req->rq_import_generation;
117 desc->bd_import = class_import_get(imp);
120 desc->bd_cbid.cbid_fn = client_bulk_callback;
121 desc->bd_cbid.cbid_arg = desc;
123 /* This makes req own desc, and free it when she frees herself */
129 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
130 int npages, int type, int portal)
132 struct obd_export *exp = req->rq_export;
133 struct ptlrpc_bulk_desc *desc;
136 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
138 desc = new_bulk(npages, type, portal);
142 desc->bd_export = class_export_get(exp);
145 desc->bd_cbid.cbid_fn = server_bulk_callback;
146 desc->bd_cbid.cbid_arg = desc;
148 /* NB we don't assign rq_bulk here; server-side requests are
149 * re-used, and the handler frees the bulk desc explicitly. */
154 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
155 cfs_page_t *page, int pageoffset, int len)
157 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
158 LASSERT(page != NULL);
159 LASSERT(pageoffset >= 0);
161 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)
172 LASSERT(desc != NULL);
173 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
174 LASSERT(!desc->bd_network_rw); /* network hands off or */
175 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
177 sptlrpc_enc_pool_put_pages(desc);
180 class_export_put(desc->bd_export);
182 class_import_put(desc->bd_import);
184 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
185 bd_iov[desc->bd_max_iov]));
189 /* Set server timelimit for this req */
190 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
196 LASSERT(req->rq_import);
199 /* non-AT settings */
200 req->rq_timeout = req->rq_import->imp_server_timeout ?
201 obd_timeout / 2 : obd_timeout;
202 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
206 at = &req->rq_import->imp_at;
207 idx = import_at_get_index(req->rq_import,
208 req->rq_request_portal);
209 serv_est = at_get(&at->iat_service_estimate[idx]);
210 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 /* do estimate only if is not in recovery */
228 if ((req->rq_send_state != LUSTRE_IMP_FULL) &&
229 (req->rq_send_state != LUSTRE_IMP_CONNECTING))
232 LASSERT(req->rq_import);
233 at = &req->rq_import->imp_at;
235 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
236 /* max service estimates are tracked on the server side,
237 so just keep minimal history here */
238 oldse = at_add(&at->iat_service_estimate[idx], serv_est);
240 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
241 "has changed from %d to %d\n",
242 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
243 oldse, at_get(&at->iat_service_estimate[idx]));
246 /* Expected network latency per remote node (secs) */
247 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
249 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
252 /* Adjust expected network latency */
253 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
254 unsigned int service_time)
256 unsigned int nl, oldnl;
258 time_t now = cfs_time_current_sec();
260 LASSERT(req->rq_import);
261 at = &req->rq_import->imp_at;
263 /* Network latency is total time less server processing time */
264 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
265 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
266 CWARN("Reported service time %u > total measured time "
267 CFS_DURATION_T"\n", service_time,
268 cfs_time_sub(now, req->rq_sent));
270 oldnl = at_add(&at->iat_net_latency, nl);
272 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
273 "has changed from %d to %d\n",
274 req->rq_import->imp_obd->obd_name,
276 &req->rq_import->imp_connection->c_remote_uuid),
277 oldnl, at_get(&at->iat_net_latency));
280 static int unpack_reply(struct ptlrpc_request *req)
284 /* Clear reply swab mask; we may have already swabbed an early reply */
285 req->rq_rep_swab_mask = 0;
287 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
289 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
293 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
295 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
302 * Handle an early reply message, called with the rq_lock held.
303 * If anything goes wrong just ignore it - same as if it never happened
305 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
307 struct ptlrpc_request *early_req;
313 spin_unlock(&req->rq_lock);
315 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
317 spin_lock(&req->rq_lock);
321 rc = unpack_reply(early_req);
323 /* Expecting to increase the service time estimate here */
324 ptlrpc_at_adj_service(req,
325 lustre_msg_get_timeout(early_req->rq_repmsg));
326 ptlrpc_at_adj_net_latency(req,
327 lustre_msg_get_service_time(early_req->rq_repmsg));
330 sptlrpc_cli_finish_early_reply(early_req);
332 spin_lock(&req->rq_lock);
335 /* Adjust the local timeout for this req */
336 ptlrpc_at_set_req_timeout(req);
338 olddl = req->rq_deadline;
339 /* server assumes it now has rq_timeout from when it sent the
340 early reply, so client should give it at least that long. */
341 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
342 ptlrpc_at_get_net_latency(req);
344 DEBUG_REQ(D_ADAPTTO, req,
345 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
346 "("CFS_DURATION_T"s)", req->rq_early_count,
347 cfs_time_sub(req->rq_deadline,
348 cfs_time_current_sec()),
349 cfs_time_sub(req->rq_deadline, olddl));
355 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
357 struct list_head *l, *tmp;
358 struct ptlrpc_request *req;
360 LASSERT(pool != NULL);
362 spin_lock(&pool->prp_lock);
363 list_for_each_safe(l, tmp, &pool->prp_req_list) {
364 req = list_entry(l, struct ptlrpc_request, rq_list);
365 list_del(&req->rq_list);
366 LASSERT(req->rq_reqbuf);
367 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
368 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
369 OBD_FREE(req, sizeof(*req));
371 spin_unlock(&pool->prp_lock);
372 OBD_FREE(pool, sizeof(*pool));
375 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
380 while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
383 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
384 "Trying to change pool size with nonempty pool "
385 "from %d to %d bytes\n", pool->prp_rq_size, size);
387 spin_lock(&pool->prp_lock);
388 pool->prp_rq_size = size;
389 for (i = 0; i < num_rq; i++) {
390 struct ptlrpc_request *req;
391 struct lustre_msg *msg;
393 spin_unlock(&pool->prp_lock);
394 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
397 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
399 OBD_FREE(req, sizeof(struct ptlrpc_request));
402 req->rq_reqbuf = msg;
403 req->rq_reqbuf_len = size;
405 spin_lock(&pool->prp_lock);
406 list_add_tail(&req->rq_list, &pool->prp_req_list);
408 spin_unlock(&pool->prp_lock);
412 struct ptlrpc_request_pool *
413 ptlrpc_init_rq_pool(int num_rq, int msgsize,
414 void (*populate_pool)(struct ptlrpc_request_pool *, int))
416 struct ptlrpc_request_pool *pool;
418 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
422 /* Request next power of two for the allocation, because internally
423 kernel would do exactly this */
425 spin_lock_init(&pool->prp_lock);
426 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
427 pool->prp_rq_size = msgsize;
428 pool->prp_populate = populate_pool;
430 populate_pool(pool, num_rq);
432 if (list_empty(&pool->prp_req_list)) {
433 /* have not allocated a single request for the pool */
434 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
440 static struct ptlrpc_request *
441 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
443 struct ptlrpc_request *request;
444 struct lustre_msg *reqbuf;
449 spin_lock(&pool->prp_lock);
451 /* See if we have anything in a pool, and bail out if nothing,
452 * in writeout path, where this matters, this is safe to do, because
453 * nothing is lost in this case, and when some in-flight requests
454 * complete, this code will be called again. */
455 if (unlikely(list_empty(&pool->prp_req_list))) {
456 spin_unlock(&pool->prp_lock);
460 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
462 list_del_init(&request->rq_list);
463 spin_unlock(&pool->prp_lock);
465 LASSERT(request->rq_reqbuf);
466 LASSERT(request->rq_pool);
468 reqbuf = request->rq_reqbuf;
469 memset(request, 0, sizeof(*request));
470 request->rq_reqbuf = reqbuf;
471 request->rq_reqbuf_len = pool->prp_rq_size;
472 request->rq_pool = pool;
477 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
479 struct ptlrpc_request_pool *pool = request->rq_pool;
481 spin_lock(&pool->prp_lock);
482 LASSERT(list_empty(&request->rq_list));
483 LASSERT(!request->rq_receiving_reply);
484 list_add_tail(&request->rq_list, &pool->prp_req_list);
485 spin_unlock(&pool->prp_lock);
488 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
489 __u32 version, int opcode,
490 int count, __u32 *lengths, char **bufs,
491 struct ptlrpc_cli_ctx *ctx)
493 struct obd_import *imp = request->rq_import;
498 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
500 rc = sptlrpc_req_get_ctx(request);
505 sptlrpc_req_set_flavor(request, opcode);
507 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
510 LASSERT(!request->rq_pool);
514 lustre_msg_add_version(request->rq_reqmsg, version);
515 request->rq_send_state = LUSTRE_IMP_FULL;
516 request->rq_type = PTL_RPC_MSG_REQUEST;
517 request->rq_export = NULL;
519 request->rq_req_cbid.cbid_fn = request_out_callback;
520 request->rq_req_cbid.cbid_arg = request;
522 request->rq_reply_cbid.cbid_fn = reply_in_callback;
523 request->rq_reply_cbid.cbid_arg = request;
525 request->rq_reply_deadline = 0;
526 request->rq_phase = RQ_PHASE_NEW;
527 request->rq_next_phase = RQ_PHASE_UNDEFINED;
529 request->rq_request_portal = imp->imp_client->cli_request_portal;
530 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
532 ptlrpc_at_set_req_timeout(request);
534 spin_lock_init(&request->rq_lock);
535 CFS_INIT_LIST_HEAD(&request->rq_list);
536 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
537 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
538 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
539 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
540 CFS_INIT_LIST_HEAD(&request->rq_history_list);
541 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
542 cfs_waitq_init(&request->rq_reply_waitq);
543 request->rq_xid = ptlrpc_next_xid();
544 atomic_set(&request->rq_refcount, 1);
546 lustre_msg_set_opc(request->rq_reqmsg, opcode);
550 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
552 class_import_put(imp);
556 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
557 __u32 version, int opcode, char **bufs,
558 struct ptlrpc_cli_ctx *ctx)
562 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
563 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
564 request->rq_pill.rc_area[RCL_CLIENT],
567 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
569 int ptlrpc_request_pack(struct ptlrpc_request *request,
570 __u32 version, int opcode)
572 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
576 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
577 struct ptlrpc_request_pool *pool)
579 struct ptlrpc_request *request = NULL;
582 request = ptlrpc_prep_req_from_pool(pool);
585 OBD_ALLOC_PTR(request);
588 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
589 LASSERT(imp != LP_POISON);
590 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
592 LASSERT(imp->imp_client != LP_POISON);
594 request->rq_import = class_import_get(imp);
596 CERROR("request allocation out of memory\n");
602 static struct ptlrpc_request *
603 ptlrpc_request_alloc_internal(struct obd_import *imp,
604 struct ptlrpc_request_pool * pool,
605 const struct req_format *format)
607 struct ptlrpc_request *request;
609 request = __ptlrpc_request_alloc(imp, pool);
613 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
614 req_capsule_set(&request->rq_pill, format);
618 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
619 const struct req_format *format)
621 return ptlrpc_request_alloc_internal(imp, NULL, format);
624 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
625 struct ptlrpc_request_pool * pool,
626 const struct req_format *format)
628 return ptlrpc_request_alloc_internal(imp, pool, format);
631 void ptlrpc_request_free(struct ptlrpc_request *request)
633 if (request->rq_pool)
634 __ptlrpc_free_req_to_pool(request);
636 OBD_FREE_PTR(request);
639 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
640 const struct req_format *format,
641 __u32 version, int opcode)
643 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
647 rc = ptlrpc_request_pack(req, version, opcode);
649 ptlrpc_request_free(req);
656 struct ptlrpc_request *
657 ptlrpc_prep_req_pool(struct obd_import *imp,
658 __u32 version, int opcode,
659 int count, __u32 *lengths, char **bufs,
660 struct ptlrpc_request_pool *pool)
662 struct ptlrpc_request *request;
665 request = __ptlrpc_request_alloc(imp, pool);
669 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
670 lengths, bufs, NULL);
672 ptlrpc_request_free(request);
678 struct ptlrpc_request *
679 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
680 __u32 *lengths, char **bufs)
682 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
686 struct ptlrpc_request_set *ptlrpc_prep_set(void)
688 struct ptlrpc_request_set *set;
691 OBD_ALLOC(set, sizeof *set);
694 CFS_INIT_LIST_HEAD(&set->set_requests);
695 cfs_waitq_init(&set->set_waitq);
696 set->set_remaining = 0;
697 spin_lock_init(&set->set_new_req_lock);
698 CFS_INIT_LIST_HEAD(&set->set_new_requests);
699 CFS_INIT_LIST_HEAD(&set->set_cblist);
704 /* Finish with this set; opposite of prep_set. */
705 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
707 struct list_head *tmp;
708 struct list_head *next;
713 /* Requests on the set should either all be completed, or all be new */
714 expected_phase = (set->set_remaining == 0) ?
715 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
716 list_for_each (tmp, &set->set_requests) {
717 struct ptlrpc_request *req =
718 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
720 LASSERT(req->rq_phase == expected_phase);
724 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
726 list_for_each_safe(tmp, next, &set->set_requests) {
727 struct ptlrpc_request *req =
728 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
729 list_del_init(&req->rq_set_chain);
731 LASSERT(req->rq_phase == expected_phase);
733 if (req->rq_phase == RQ_PHASE_NEW) {
735 if (req->rq_interpret_reply != NULL) {
736 ptlrpc_interpterer_t interpreter =
737 req->rq_interpret_reply;
739 /* higher level (i.e. LOV) failed;
740 * let the sub reqs clean up */
741 req->rq_status = -EBADR;
742 interpreter(NULL, req, &req->rq_async_args,
745 set->set_remaining--;
749 ptlrpc_req_finished (req);
752 LASSERT(set->set_remaining == 0);
754 OBD_FREE(set, sizeof(*set));
758 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
759 set_interpreter_func fn, void *data)
761 struct ptlrpc_set_cbdata *cbdata;
763 OBD_ALLOC_PTR(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 list_add_tail(&req->rq_set_chain, &set->set_requests);
780 set->set_remaining++;
784 * Lock so many callers can add things, the context that owns the set
785 * is supposed to notice these and move them into the set proper.
787 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
788 struct ptlrpc_request *req)
790 struct ptlrpc_request_set *set = pc->pc_set;
793 * Let caller know that we stopped and will not handle this request.
794 * It needs to take care itself of request.
796 if (test_bit(LIOD_STOP, &pc->pc_flags))
799 spin_lock(&set->set_new_req_lock);
801 * The set takes over the caller's request reference.
803 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
805 spin_unlock(&set->set_new_req_lock);
808 * Let thead know that we added something and better it to wake up
811 cfs_waitq_signal(&set->set_waitq);
816 * Based on the current state of the import, determine if the request
817 * can be sent, is an error, or should be delayed.
819 * Returns true if this request should be delayed. If false, and
820 * *status is set, then the request can not be sent and *status is the
821 * error code. If false and status is 0, then request can be sent.
823 * The imp->imp_lock must be held.
825 static int ptlrpc_import_delay_req(struct obd_import *imp,
826 struct ptlrpc_request *req, int *status)
831 LASSERT (status != NULL);
834 if (req->rq_ctx_init || req->rq_ctx_fini) {
835 /* always allow ctx init/fini rpc go through */
836 } else 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)
883 /* serialise with network callback */
884 spin_lock(&req->rq_lock);
886 if (ptlrpc_client_replied(req))
889 if (req->rq_net_err && !req->rq_timedout) {
890 spin_unlock(&req->rq_lock);
891 rc = ptlrpc_expire_one_request(req, 0);
892 spin_lock(&req->rq_lock);
905 if (ptlrpc_client_early(req)) {
906 ptlrpc_at_recv_early_reply(req);
907 GOTO(out, rc = 0); /* keep waiting */
912 spin_unlock(&req->rq_lock);
913 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
917 /* Conditionally suppress specific console messages */
918 static int ptlrpc_console_allow(struct ptlrpc_request *req)
920 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
923 /* Suppress particular reconnect errors which are to be expected. No
924 * errors are suppressed for the initial connection on an import */
925 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
926 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
928 /* Suppress timed out reconnect requests */
929 if (req->rq_timedout)
932 /* Suppress unavailable/again reconnect requests */
933 err = lustre_msg_get_status(req->rq_repmsg);
934 if (err == -ENODEV || err == -EAGAIN)
942 static int ptlrpc_check_status(struct ptlrpc_request *req)
947 err = lustre_msg_get_status(req->rq_repmsg);
948 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
949 struct obd_import *imp = req->rq_import;
950 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
951 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
952 " with %s. The %s operation failed with %d\n",
953 libcfs_nid2str(imp->imp_connection->c_peer.nid),
954 ll_opcode2str(opc), err);
955 RETURN(err < 0 ? err : -EINVAL);
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);
984 * save pre-versions for replay
986 static void ptlrpc_save_versions(struct ptlrpc_request *req)
988 struct lustre_msg *repmsg = req->rq_repmsg;
989 struct lustre_msg *reqmsg = req->rq_reqmsg;
990 __u64 *versions = lustre_msg_get_versions(repmsg);
993 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]);
1005 * Callback function called when client receives RPC reply for \a req.
1007 static int after_reply(struct ptlrpc_request *req)
1009 struct obd_import *imp = req->rq_import;
1010 struct obd_device *obd = req->rq_import->imp_obd;
1012 struct timeval work_start;
1016 LASSERT(!req->rq_receiving_reply);
1018 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
1021 * NB Until this point, the whole of the incoming message,
1022 * including buflens, status etc is in the sender's byte order.
1025 rc = sptlrpc_cli_unwrap_reply(req);
1027 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1032 * Security layer unwrap might ask resend this request.
1037 rc = unpack_reply(req);
1041 do_gettimeofday(&work_start);
1042 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1043 if (obd->obd_svc_stats != NULL) {
1044 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1046 ptlrpc_lprocfs_rpc_sent(req, timediff);
1049 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1050 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1051 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1052 lustre_msg_get_type(req->rq_repmsg));
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,
1059 lustre_msg_get_service_time(req->rq_repmsg));
1061 rc = ptlrpc_check_status(req);
1062 imp->imp_connect_error = rc;
1066 * Either we've been evicted, or the server has failed for
1067 * some reason. Try to reconnect, and if that fails, punt to
1070 if (ll_rpc_recoverable_error(rc)) {
1071 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1072 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1075 ptlrpc_request_handle_notconn(req);
1080 * Let's look if server sent slv. Do it only for RPC with
1083 ldlm_cli_update_pool(req);
1087 * Store transno in reqmsg for replay.
1089 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1090 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1091 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1094 if (imp->imp_replayable) {
1095 spin_lock(&imp->imp_lock);
1097 * No point in adding already-committed requests to the replay
1098 * list, we will just remove them immediately. b=9829
1100 if (req->rq_transno != 0 &&
1102 lustre_msg_get_last_committed(req->rq_repmsg) ||
1104 /** version recovery */
1105 ptlrpc_save_versions(req);
1106 ptlrpc_retain_replayable_request(req, imp);
1107 } else if (req->rq_commit_cb != NULL) {
1108 spin_unlock(&imp->imp_lock);
1109 req->rq_commit_cb(req);
1110 spin_lock(&imp->imp_lock);
1114 * Replay-enabled imports return commit-status information.
1116 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1117 imp->imp_peer_committed_transno =
1118 lustre_msg_get_last_committed(req->rq_repmsg);
1120 ptlrpc_free_committed(imp);
1121 spin_unlock(&imp->imp_lock);
1127 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1129 struct obd_import *imp;
1133 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1134 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1137 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1139 imp = req->rq_import;
1140 spin_lock(&imp->imp_lock);
1142 req->rq_import_generation = imp->imp_generation;
1144 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1145 spin_lock(&req->rq_lock);
1146 req->rq_waiting = 1;
1147 spin_unlock(&req->rq_lock);
1149 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1150 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1151 ptlrpc_import_state_name(req->rq_send_state),
1152 ptlrpc_import_state_name(imp->imp_state));
1153 LASSERT(list_empty(&req->rq_list));
1154 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1155 atomic_inc(&req->rq_import->imp_inflight);
1156 spin_unlock(&imp->imp_lock);
1161 spin_unlock(&imp->imp_lock);
1162 req->rq_status = rc;
1163 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1167 LASSERT(list_empty(&req->rq_list));
1168 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1169 atomic_inc(&req->rq_import->imp_inflight);
1170 spin_unlock(&imp->imp_lock);
1172 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1174 rc = sptlrpc_req_refresh_ctx(req, -1);
1177 req->rq_status = rc;
1180 req->rq_wait_ctx = 1;
1185 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1186 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1187 imp->imp_obd->obd_uuid.uuid,
1188 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1189 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1190 lustre_msg_get_opc(req->rq_reqmsg));
1192 rc = ptl_send_rpc(req, 0);
1194 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1195 req->rq_net_err = 1;
1201 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1202 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1204 struct list_head *tmp;
1205 int force_timer_recalc = 0;
1208 if (set->set_remaining == 0)
1211 list_for_each(tmp, &set->set_requests) {
1212 struct ptlrpc_request *req =
1213 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1214 struct obd_import *imp = req->rq_import;
1217 if (req->rq_phase == RQ_PHASE_NEW &&
1218 ptlrpc_send_new_req(req)) {
1219 force_timer_recalc = 1;
1222 /* delayed send - skip */
1223 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1226 if (!(req->rq_phase == RQ_PHASE_RPC ||
1227 req->rq_phase == RQ_PHASE_BULK ||
1228 req->rq_phase == RQ_PHASE_INTERPRET ||
1229 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1230 req->rq_phase == RQ_PHASE_COMPLETE)) {
1231 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1235 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1236 LASSERT(req->rq_next_phase != req->rq_phase);
1237 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1240 * Skip processing until reply is unlinked. We
1241 * can't return to pool before that and we can't
1242 * call interpret before that. We need to make
1243 * sure that all rdma transfers finished and will
1244 * not corrupt any data.
1246 if (ptlrpc_client_recv_or_unlink(req) ||
1247 ptlrpc_client_bulk_active(req))
1251 * Turn fail_loc off to prevent it from looping
1254 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1255 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1258 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1259 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1264 * Move to next phase if reply was successfully
1267 ptlrpc_rqphase_move(req, req->rq_next_phase);
1270 if (req->rq_phase == RQ_PHASE_COMPLETE)
1273 if (req->rq_phase == RQ_PHASE_INTERPRET)
1274 GOTO(interpret, req->rq_status);
1277 * Note that this also will start async reply unlink.
1279 if (req->rq_net_err && !req->rq_timedout) {
1280 ptlrpc_expire_one_request(req, 1);
1283 * Check if we still need to wait for unlink.
1285 if (ptlrpc_client_recv_or_unlink(req) ||
1286 ptlrpc_client_bulk_active(req))
1291 req->rq_replied = 0;
1292 if (req->rq_status == 0)
1293 req->rq_status = -EIO;
1294 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1295 GOTO(interpret, req->rq_status);
1298 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1299 * will only be set after rq_timedout, but the synchronous IO
1300 * waiting path sets rq_intr irrespective of whether ptlrpcd
1301 * has seen a timeout. our policy is to only interpret
1302 * interrupted rpcs after they have timed out */
1303 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1304 req->rq_wait_ctx)) {
1305 req->rq_status = -EINTR;
1306 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1307 GOTO(interpret, req->rq_status);
1310 if (req->rq_phase == RQ_PHASE_RPC) {
1311 if (req->rq_timedout || req->rq_resend ||
1312 req->rq_waiting || req->rq_wait_ctx) {
1315 if (!ptlrpc_unregister_reply(req, 1))
1318 spin_lock(&imp->imp_lock);
1319 if (ptlrpc_import_delay_req(imp, req, &status)){
1320 /* put on delay list - only if we wait
1321 * recovery finished - before send */
1322 list_del_init(&req->rq_list);
1323 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1324 spin_unlock(&imp->imp_lock);
1329 req->rq_status = status;
1330 ptlrpc_rqphase_move(req,
1331 RQ_PHASE_INTERPRET);
1332 spin_unlock(&imp->imp_lock);
1333 GOTO(interpret, req->rq_status);
1335 if (req->rq_no_resend && !req->rq_wait_ctx) {
1336 req->rq_status = -ENOTCONN;
1337 ptlrpc_rqphase_move(req,
1338 RQ_PHASE_INTERPRET);
1339 spin_unlock(&imp->imp_lock);
1340 GOTO(interpret, req->rq_status);
1343 list_del_init(&req->rq_list);
1344 list_add_tail(&req->rq_list,
1345 &imp->imp_sending_list);
1347 spin_unlock(&imp->imp_lock);
1349 req->rq_waiting = 0;
1351 if (req->rq_timedout||req->rq_resend) {
1352 /* This is re-sending anyways,
1353 * let's mark req as resend. */
1358 if (!ptlrpc_unregister_bulk(req, 1))
1361 /* ensure previous bulk fails */
1362 old_xid = req->rq_xid;
1363 req->rq_xid = ptlrpc_next_xid();
1364 CDEBUG(D_HA, "resend bulk "
1367 old_xid, req->rq_xid);
1371 * rq_wait_ctx is only touched by ptlrpcd,
1372 * so no lock is needed here.
1374 status = sptlrpc_req_refresh_ctx(req, -1);
1377 req->rq_status = status;
1378 req->rq_wait_ctx = 0;
1379 force_timer_recalc = 1;
1381 req->rq_wait_ctx = 1;
1386 req->rq_wait_ctx = 0;
1389 rc = ptl_send_rpc(req, 0);
1391 DEBUG_REQ(D_HA, req, "send failed (%d)",
1393 force_timer_recalc = 1;
1394 req->rq_net_err = 1;
1396 /* need to reset the timeout */
1397 force_timer_recalc = 1;
1400 spin_lock(&req->rq_lock);
1402 if (ptlrpc_client_early(req)) {
1403 ptlrpc_at_recv_early_reply(req);
1404 spin_unlock(&req->rq_lock);
1408 /* Still waiting for a reply? */
1409 if (ptlrpc_client_recv(req)) {
1410 spin_unlock(&req->rq_lock);
1414 /* Did we actually receive a reply? */
1415 if (!ptlrpc_client_replied(req)) {
1416 spin_unlock(&req->rq_lock);
1420 spin_unlock(&req->rq_lock);
1422 req->rq_status = after_reply(req);
1426 /* If there is no bulk associated with this request,
1427 * then we're done and should let the interpreter
1428 * process the reply. Similarly if the RPC returned
1429 * an error, and therefore the bulk will never arrive.
1431 if (req->rq_bulk == NULL || req->rq_status != 0) {
1432 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1433 GOTO(interpret, req->rq_status);
1436 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1439 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1440 if (ptlrpc_client_bulk_active(req))
1443 if (!req->rq_bulk->bd_success) {
1444 /* The RPC reply arrived OK, but the bulk screwed
1445 * up! Dead wierd since the server told us the RPC
1446 * was good after getting the REPLY for her GET or
1447 * the ACK for her PUT. */
1448 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1452 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1455 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1457 /* This moves to "unregistering" phase we need to wait for
1459 if (!ptlrpc_unregister_reply(req, 1))
1462 if (!ptlrpc_unregister_bulk(req, 1))
1465 /* When calling interpret receiving already should be
1467 LASSERT(!req->rq_receiving_reply);
1469 if (req->rq_interpret_reply != NULL) {
1470 ptlrpc_interpterer_t interpreter =
1471 req->rq_interpret_reply;
1472 req->rq_status = interpreter(env, req,
1473 &req->rq_async_args,
1476 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1478 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1479 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1480 imp->imp_obd->obd_uuid.uuid,
1481 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1482 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1483 lustre_msg_get_opc(req->rq_reqmsg));
1485 spin_lock(&imp->imp_lock);
1486 /* Request already may be not on sending or delaying list. This
1487 * may happen in the case of marking it errorneous for the case
1488 * ptlrpc_import_delay_req(req, status) find it impossible to
1489 * allow sending this rpc and returns *status != 0. */
1490 if (!list_empty(&req->rq_list)) {
1491 list_del_init(&req->rq_list);
1492 atomic_dec(&imp->imp_inflight);
1494 spin_unlock(&imp->imp_lock);
1496 set->set_remaining--;
1497 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1500 /* If we hit an error, we want to recover promptly. */
1501 RETURN(set->set_remaining == 0 || force_timer_recalc);
1504 /* Return 1 if we should give up, else 0 */
1505 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1507 struct obd_import *imp = req->rq_import;
1511 spin_lock(&req->rq_lock);
1512 req->rq_timedout = 1;
1513 spin_unlock(&req->rq_lock);
1515 DEBUG_REQ(D_WARNING, req, "Request x"LPU64" sent from %s to NID %s "
1516 CFS_DURATION_T"s ago has %s (limit "CFS_DURATION_T"s).\n",
1517 req->rq_xid, imp ? imp->imp_obd->obd_name : "<?>",
1518 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1519 cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1520 req->rq_net_err ? "failed due to network error" : "timed out",
1521 cfs_time_sub(req->rq_deadline, req->rq_sent));
1523 if (imp != NULL && obd_debug_peer_on_timeout)
1524 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1526 ptlrpc_unregister_reply(req, async_unlink);
1527 ptlrpc_unregister_bulk(req, async_unlink);
1529 if (obd_dump_on_timeout)
1530 libcfs_debug_dumplog();
1533 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1537 atomic_inc(&imp->imp_timeouts);
1539 /* The DLM server doesn't want recovery run on its imports. */
1540 if (imp->imp_dlm_fake)
1543 /* If this request is for recovery or other primordial tasks,
1544 * then error it out here. */
1545 if (req->rq_ctx_init || req->rq_ctx_fini ||
1546 req->rq_send_state != LUSTRE_IMP_FULL ||
1547 imp->imp_obd->obd_no_recov) {
1548 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1549 ptlrpc_import_state_name(req->rq_send_state),
1550 ptlrpc_import_state_name(imp->imp_state));
1551 spin_lock(&req->rq_lock);
1552 req->rq_status = -ETIMEDOUT;
1554 spin_unlock(&req->rq_lock);
1558 /* if a request can't be resent we can't wait for an answer after
1560 if (req->rq_no_resend) {
1561 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1565 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1570 int ptlrpc_expired_set(void *data)
1572 struct ptlrpc_request_set *set = data;
1573 struct list_head *tmp;
1574 time_t now = cfs_time_current_sec();
1577 LASSERT(set != NULL);
1580 * A timeout expired. See which reqs it applies to...
1582 list_for_each (tmp, &set->set_requests) {
1583 struct ptlrpc_request *req =
1584 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1586 /* don't expire request waiting for context */
1587 if (req->rq_wait_ctx)
1590 /* Request in-flight? */
1591 if (!((req->rq_phase == RQ_PHASE_RPC &&
1592 !req->rq_waiting && !req->rq_resend) ||
1593 (req->rq_phase == RQ_PHASE_BULK)))
1596 if (req->rq_timedout || /* already dealt with */
1597 req->rq_deadline > now) /* not expired */
1600 /* Deal with this guy. Do it asynchronously to not block
1601 * ptlrpcd thread. */
1602 ptlrpc_expire_one_request(req, 1);
1606 * When waiting for a whole set, we always to break out of the
1607 * sleep so we can recalculate the timeout, or enable interrupts
1608 * if everyone's timed out.
1613 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1615 spin_lock(&req->rq_lock);
1617 spin_unlock(&req->rq_lock);
1620 void ptlrpc_interrupted_set(void *data)
1622 struct ptlrpc_request_set *set = data;
1623 struct list_head *tmp;
1625 LASSERT(set != NULL);
1626 CERROR("INTERRUPTED SET %p\n", set);
1628 list_for_each(tmp, &set->set_requests) {
1629 struct ptlrpc_request *req =
1630 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1632 if (req->rq_phase != RQ_PHASE_RPC &&
1633 req->rq_phase != RQ_PHASE_UNREGISTERING)
1636 ptlrpc_mark_interrupted(req);
1641 * Get the smallest timeout in the set; this does NOT set a timeout.
1643 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1645 struct list_head *tmp;
1646 time_t now = cfs_time_current_sec();
1648 struct ptlrpc_request *req;
1652 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1654 list_for_each(tmp, &set->set_requests) {
1655 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1658 * Request in-flight?
1660 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1661 (req->rq_phase == RQ_PHASE_BULK) ||
1662 (req->rq_phase == RQ_PHASE_NEW)))
1666 * Already timed out.
1668 if (req->rq_timedout)
1674 if (req->rq_wait_ctx)
1677 if (req->rq_phase == RQ_PHASE_NEW)
1678 deadline = req->rq_sent;
1680 deadline = req->rq_sent + req->rq_timeout;
1682 if (deadline <= now) /* actually expired already */
1683 timeout = 1; /* ASAP */
1684 else if (timeout == 0 || timeout > deadline - now)
1685 timeout = deadline - now;
1690 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1692 struct list_head *tmp;
1693 struct ptlrpc_request *req;
1694 struct l_wait_info lwi;
1698 if (list_empty(&set->set_requests))
1701 list_for_each(tmp, &set->set_requests) {
1702 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1703 if (req->rq_phase == RQ_PHASE_NEW)
1704 (void)ptlrpc_send_new_req(req);
1708 timeout = ptlrpc_set_next_timeout(set);
1710 /* wait until all complete, interrupted, or an in-flight
1712 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1714 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1716 ptlrpc_interrupted_set, set);
1717 rc = l_wait_event(set->set_waitq,
1718 ptlrpc_check_set(NULL, set), &lwi);
1720 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1722 /* -EINTR => all requests have been flagged rq_intr so next
1724 * -ETIMEOUTD => someone timed out. When all reqs have
1725 * timed out, signals are enabled allowing completion with
1727 * I don't really care if we go once more round the loop in
1728 * the error cases -eeb. */
1729 } while (rc != 0 || set->set_remaining != 0);
1731 LASSERT(set->set_remaining == 0);
1734 list_for_each(tmp, &set->set_requests) {
1735 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1737 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1738 if (req->rq_status != 0)
1739 rc = req->rq_status;
1742 if (set->set_interpret != NULL) {
1743 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1745 rc = interpreter (set, set->set_arg, rc);
1747 struct ptlrpc_set_cbdata *cbdata, *n;
1750 list_for_each_entry_safe(cbdata, n,
1751 &set->set_cblist, psc_item) {
1752 list_del_init(&cbdata->psc_item);
1753 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1756 OBD_FREE_PTR(cbdata);
1763 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1766 if (request == NULL) {
1771 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1772 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1773 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1774 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1775 LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request);
1776 LASSERTF(!request->rq_replay, "req %p\n", request);
1777 LASSERT(request->rq_cli_ctx);
1779 req_capsule_fini(&request->rq_pill);
1781 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1782 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1783 if (request->rq_import != NULL) {
1785 spin_lock(&request->rq_import->imp_lock);
1786 list_del_init(&request->rq_replay_list);
1788 spin_unlock(&request->rq_import->imp_lock);
1790 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1792 if (atomic_read(&request->rq_refcount) != 0) {
1793 DEBUG_REQ(D_ERROR, request,
1794 "freeing request with nonzero refcount");
1798 if (request->rq_repbuf != NULL)
1799 sptlrpc_cli_free_repbuf(request);
1800 if (request->rq_export != NULL) {
1801 class_export_put(request->rq_export);
1802 request->rq_export = NULL;
1804 if (request->rq_import != NULL) {
1805 class_import_put(request->rq_import);
1806 request->rq_import = NULL;
1808 if (request->rq_bulk != NULL)
1809 ptlrpc_free_bulk(request->rq_bulk);
1811 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1812 sptlrpc_cli_free_reqbuf(request);
1814 sptlrpc_req_put_ctx(request, !locked);
1816 if (request->rq_pool)
1817 __ptlrpc_free_req_to_pool(request);
1819 OBD_FREE(request, sizeof(*request));
1823 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1824 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1826 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1827 (void)__ptlrpc_req_finished(request, 1);
1830 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1833 if (request == NULL)
1836 if (request == LP_POISON ||
1837 request->rq_reqmsg == LP_POISON) {
1838 CERROR("dereferencing freed request (bug 575)\n");
1843 DEBUG_REQ(D_INFO, request, "refcount now %u",
1844 atomic_read(&request->rq_refcount) - 1);
1846 if (atomic_dec_and_test(&request->rq_refcount)) {
1847 __ptlrpc_free_req(request, locked);
1854 void ptlrpc_req_finished(struct ptlrpc_request *request)
1856 __ptlrpc_req_finished(request, 0);
1859 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1861 return request->rq_xid;
1863 EXPORT_SYMBOL(ptlrpc_req_xid);
1865 /* Disengage the client's reply buffer from the network
1866 * NB does _NOT_ unregister any client-side bulk.
1867 * IDEMPOTENT, but _not_ safe against concurrent callers.
1868 * The request owner (i.e. the thread doing the I/O) must call...
1870 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1874 struct l_wait_info lwi;
1879 LASSERT(!in_interrupt());
1882 * Let's setup deadline for reply unlink.
1884 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1885 async && request->rq_reply_deadline == 0)
1886 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1889 * Nothing left to do.
1891 if (!ptlrpc_client_recv_or_unlink(request))
1894 LNetMDUnlink(request->rq_reply_md_h);
1897 * Let's check it once again.
1899 if (!ptlrpc_client_recv_or_unlink(request))
1903 * Move to "Unregistering" phase as reply was not unlinked yet.
1905 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1908 * Do not wait for unlink to finish.
1914 * We have to l_wait_event() whatever the result, to give liblustre
1915 * a chance to run reply_in_callback(), and to make sure we've
1916 * unlinked before returning a req to the pool.
1918 if (request->rq_set != NULL)
1919 wq = &request->rq_set->set_waitq;
1921 wq = &request->rq_reply_waitq;
1924 /* Network access will complete in finite time but the HUGE
1925 * timeout lets us CWARN for visibility of sluggish NALs */
1926 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1927 cfs_time_seconds(1), NULL, NULL);
1928 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1931 ptlrpc_rqphase_move(request, request->rq_next_phase);
1935 LASSERT(rc == -ETIMEDOUT);
1936 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1937 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1938 request->rq_must_unlink);
1943 /* caller must hold imp->imp_lock */
1944 void ptlrpc_free_committed(struct obd_import *imp)
1946 struct list_head *tmp, *saved;
1947 struct ptlrpc_request *req;
1948 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1951 LASSERT(imp != NULL);
1953 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1956 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1957 imp->imp_generation == imp->imp_last_generation_checked) {
1958 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1959 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1963 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1964 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1965 imp->imp_generation);
1966 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1967 imp->imp_last_generation_checked = imp->imp_generation;
1969 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1970 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1972 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1973 LASSERT(req != last_req);
1976 if (req->rq_transno == 0) {
1977 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
1980 if (req->rq_import_generation < imp->imp_generation) {
1981 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1985 if (req->rq_replay) {
1986 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1990 /* not yet committed */
1991 if (req->rq_transno > imp->imp_peer_committed_transno) {
1992 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1996 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1997 imp->imp_peer_committed_transno);
1999 spin_lock(&req->rq_lock);
2001 spin_unlock(&req->rq_lock);
2002 if (req->rq_commit_cb != NULL)
2003 req->rq_commit_cb(req);
2004 list_del_init(&req->rq_replay_list);
2005 __ptlrpc_req_finished(req, 1);
2012 void ptlrpc_cleanup_client(struct obd_import *imp)
2019 void ptlrpc_resend_req(struct ptlrpc_request *req)
2021 DEBUG_REQ(D_HA, req, "going to resend");
2022 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2023 req->rq_status = -EAGAIN;
2025 spin_lock(&req->rq_lock);
2027 req->rq_net_err = 0;
2028 req->rq_timedout = 0;
2030 __u64 old_xid = req->rq_xid;
2032 /* ensure previous bulk fails */
2033 req->rq_xid = ptlrpc_next_xid();
2034 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2035 old_xid, req->rq_xid);
2037 ptlrpc_client_wake_req(req);
2038 spin_unlock(&req->rq_lock);
2041 /* XXX: this function and rq_status are currently unused */
2042 void ptlrpc_restart_req(struct ptlrpc_request *req)
2044 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2045 req->rq_status = -ERESTARTSYS;
2047 spin_lock(&req->rq_lock);
2048 req->rq_restart = 1;
2049 req->rq_timedout = 0;
2050 ptlrpc_client_wake_req(req);
2051 spin_unlock(&req->rq_lock);
2054 static int expired_request(void *data)
2056 struct ptlrpc_request *req = data;
2060 * Some failure can suspend regular timeouts.
2062 if (ptlrpc_check_suspend())
2066 * Deadline may have changed with an early reply.
2068 if (req->rq_deadline > cfs_time_current_sec())
2071 RETURN(ptlrpc_expire_one_request(req, 0));
2074 static void interrupted_request(void *data)
2076 struct ptlrpc_request *req = data;
2077 DEBUG_REQ(D_HA, req, "request interrupted");
2078 spin_lock(&req->rq_lock);
2080 spin_unlock(&req->rq_lock);
2083 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2086 atomic_inc(&req->rq_refcount);
2090 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2091 struct obd_import *imp)
2093 struct list_head *tmp;
2095 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2097 if (req->rq_transno == 0) {
2098 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2102 /* clear this for new requests that were resent as well
2103 as resent replayed requests. */
2104 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2106 /* don't re-add requests that have been replayed */
2107 if (!list_empty(&req->rq_replay_list))
2110 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2112 LASSERT(imp->imp_replayable);
2113 /* Balanced in ptlrpc_free_committed, usually. */
2114 ptlrpc_request_addref(req);
2115 list_for_each_prev(tmp, &imp->imp_replay_list) {
2116 struct ptlrpc_request *iter =
2117 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
2119 /* We may have duplicate transnos if we create and then
2120 * open a file, or for closes retained if to match creating
2121 * opens, so use req->rq_xid as a secondary key.
2122 * (See bugs 684, 685, and 428.)
2123 * XXX no longer needed, but all opens need transnos!
2125 if (iter->rq_transno > req->rq_transno)
2128 if (iter->rq_transno == req->rq_transno) {
2129 LASSERT(iter->rq_xid != req->rq_xid);
2130 if (iter->rq_xid > req->rq_xid)
2134 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2138 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
2141 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2145 struct l_wait_info lwi;
2146 struct obd_import *imp = req->rq_import;
2147 cfs_duration_t timeout = CFS_TICK;
2151 LASSERT(req->rq_set == NULL);
2152 LASSERT(!req->rq_receiving_reply);
2154 /* for distributed debugging */
2155 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2156 LASSERT(imp->imp_obd != NULL);
2157 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
2158 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2159 imp->imp_obd->obd_uuid.uuid,
2160 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2161 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2162 lustre_msg_get_opc(req->rq_reqmsg));
2164 /* Mark phase here for a little debug help */
2165 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
2167 spin_lock(&imp->imp_lock);
2168 req->rq_import_generation = imp->imp_generation;
2170 if (ptlrpc_import_delay_req(imp, req, &rc)) {
2171 list_del_init(&req->rq_list);
2172 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
2173 atomic_inc(&imp->imp_inflight);
2174 spin_unlock(&imp->imp_lock);
2176 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
2178 ptlrpc_import_state_name(req->rq_send_state),
2179 ptlrpc_import_state_name(imp->imp_state));
2180 lwi = LWI_INTR(interrupted_request, req);
2181 rc = l_wait_event(req->rq_reply_waitq,
2182 (req->rq_send_state == imp->imp_state ||
2183 req->rq_err || req->rq_intr),
2185 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2187 ptlrpc_import_state_name(imp->imp_state),
2188 ptlrpc_import_state_name(req->rq_send_state),
2189 req->rq_err, req->rq_intr);
2191 spin_lock(&imp->imp_lock);
2192 list_del_init(&req->rq_list);
2193 atomic_dec(&imp->imp_inflight);
2196 /* rq_status was set locally */
2199 else if (req->rq_intr) {
2202 else if (req->rq_no_resend) {
2203 spin_unlock(&imp->imp_lock);
2204 GOTO(out, rc = -ETIMEDOUT);
2212 spin_unlock(&imp->imp_lock);
2213 req->rq_status = rc; // XXX this ok?
2217 if (req->rq_resend) {
2218 if (req->rq_bulk != NULL) {
2219 ptlrpc_unregister_bulk(req, 0);
2221 /* bulk requests are supposed to be
2222 * idempotent, so we are free to bump the xid
2223 * here, which we need to do before
2224 * registering the bulk again (bug 6371).
2225 * print the old xid first for sanity.
2227 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2228 req->rq_xid = ptlrpc_next_xid();
2231 DEBUG_REQ(D_HA, req, "resending: ");
2234 /* XXX this is the same as ptlrpc_set_wait */
2235 LASSERT(list_empty(&req->rq_list));
2236 list_add_tail(&req->rq_list, &imp->imp_sending_list);
2237 atomic_inc(&imp->imp_inflight);
2238 spin_unlock(&imp->imp_lock);
2240 rc = sptlrpc_req_refresh_ctx(req, 0);
2243 /* we got fatal ctx refresh error, directly jump out
2244 * thus we can pass back the actual error code.
2246 spin_lock(&imp->imp_lock);
2247 list_del_init(&req->rq_list);
2248 atomic_dec(&imp->imp_inflight);
2249 spin_unlock(&imp->imp_lock);
2251 CERROR("Failed to refresh ctx of req %p: %d\n",
2255 /* simulating we got error during send rpc */
2259 rc = ptl_send_rpc(req, 0);
2261 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2263 timeoutl = req->rq_deadline - cfs_time_current_sec();
2264 timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2265 cfs_time_seconds(timeoutl);
2266 DEBUG_REQ(D_NET, req,
2267 "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2268 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
2270 brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
2271 if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
2272 ptlrpc_check_and_wait_suspend(req)))
2276 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2277 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2278 imp->imp_obd->obd_uuid.uuid,
2279 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2280 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2281 lustre_msg_get_opc(req->rq_reqmsg));
2283 /* If the reply was received normally, this just grabs the spinlock
2284 * (ensuring the reply callback has returned), sees that
2285 * req->rq_receiving_reply is clear and returns. */
2286 ptlrpc_unregister_reply(req, 0);
2288 spin_lock(&imp->imp_lock);
2289 list_del_init(&req->rq_list);
2290 atomic_dec(&imp->imp_inflight);
2291 spin_unlock(&imp->imp_lock);
2294 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2295 rc, req->rq_status);
2296 GOTO(out, rc = rc ? rc : -EIO);
2300 /* Should only be interrupted if we timed out. */
2301 if (!req->rq_timedout)
2302 DEBUG_REQ(D_ERROR, req,
2303 "rq_intr set but rq_timedout not");
2304 GOTO(out, rc = -EINTR);
2307 /* Resend if we need to */
2308 if (req->rq_resend||req->rq_timedout) {
2309 /* ...unless we were specifically told otherwise. */
2310 if (req->rq_no_resend)
2311 GOTO(out, rc = -ETIMEDOUT);
2312 spin_lock(&imp->imp_lock);
2313 /* we can have rq_timeout on dlm fake import which not support
2314 * recovery - but me need resend request on this import instead
2315 * of return error */
2320 if (!ptlrpc_client_replied(req)) {
2321 /* How can this be? -eeb */
2322 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2324 GOTO(out, rc = req->rq_status);
2327 rc = after_reply(req);
2328 /* NB may return +ve success rc */
2329 if (req->rq_resend) {
2330 spin_lock(&imp->imp_lock);
2335 if (req->rq_bulk != NULL) {
2337 /* success so far. Note that anything going wrong
2338 * with bulk now, is EXTREMELY strange, since the
2339 * server must have believed that the bulk
2340 * tranferred OK before she replied with success to
2342 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2343 brc = l_wait_event(req->rq_reply_waitq,
2344 !ptlrpc_client_bulk_active(req),
2346 LASSERT(brc == 0 || brc == -ETIMEDOUT);
2348 LASSERT(brc == -ETIMEDOUT);
2349 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2351 } else if (!req->rq_bulk->bd_success) {
2352 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2357 ptlrpc_unregister_bulk(req, 0);
2360 LASSERT(!req->rq_receiving_reply);
2361 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2362 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
2366 struct ptlrpc_replay_async_args {
2368 int praa_old_status;
2371 static int ptlrpc_replay_interpret(const struct lu_env *env,
2372 struct ptlrpc_request *req,
2373 void * data, int rc)
2375 struct ptlrpc_replay_async_args *aa = data;
2376 struct obd_import *imp = req->rq_import;
2379 atomic_dec(&imp->imp_replay_inflight);
2381 if (!ptlrpc_client_replied(req)) {
2382 CERROR("request replay timed out, restarting recovery\n");
2383 GOTO(out, rc = -ETIMEDOUT);
2386 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2387 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2388 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2389 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2391 /** VBR: check version failure */
2392 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2393 /** replay was failed due to version mismatch */
2394 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2395 spin_lock(&imp->imp_lock);
2396 imp->imp_vbr_failed = 1;
2397 imp->imp_no_lock_replay = 1;
2398 spin_unlock(&imp->imp_lock);
2400 /** The transno had better not change over replay. */
2401 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2402 lustre_msg_get_transno(req->rq_repmsg) ||
2403 lustre_msg_get_transno(req->rq_repmsg) == 0,
2405 lustre_msg_get_transno(req->rq_reqmsg),
2406 lustre_msg_get_transno(req->rq_repmsg));
2409 spin_lock(&imp->imp_lock);
2410 /** if replays by version then gap was occur on server, no trust to locks */
2411 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2412 imp->imp_no_lock_replay = 1;
2413 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2414 spin_unlock(&imp->imp_lock);
2415 LASSERT(imp->imp_last_replay_transno);
2417 DEBUG_REQ(D_HA, req, "got rep");
2419 /* let the callback do fixups, possibly including in the request */
2420 if (req->rq_replay_cb)
2421 req->rq_replay_cb(req);
2423 if (ptlrpc_client_replied(req) &&
2424 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2425 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2426 lustre_msg_get_status(req->rq_repmsg),
2427 aa->praa_old_status);
2429 /* Put it back for re-replay. */
2430 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2434 * Errors while replay can set transno to 0, but
2435 * imp_last_replay_transno shouldn't be set to 0 anyway
2437 if (req->rq_transno > 0) {
2438 spin_lock(&imp->imp_lock);
2439 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2440 imp->imp_last_replay_transno = req->rq_transno;
2441 spin_unlock(&imp->imp_lock);
2443 CERROR("Transno is 0 during replay!\n");
2444 /* continue with recovery */
2445 rc = ptlrpc_import_recovery_state_machine(imp);
2447 req->rq_send_state = aa->praa_old_state;
2450 /* this replay failed, so restart recovery */
2451 ptlrpc_connect_import(imp, NULL);
2456 int ptlrpc_replay_req(struct ptlrpc_request *req)
2458 struct ptlrpc_replay_async_args *aa;
2461 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2462 /* Not handling automatic bulk replay yet (or ever?) */
2463 LASSERT(req->rq_bulk == NULL);
2465 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2466 aa = ptlrpc_req_async_args(req);
2467 memset(aa, 0, sizeof *aa);
2469 /* Prepare request to be resent with ptlrpcd */
2470 aa->praa_old_state = req->rq_send_state;
2471 req->rq_send_state = LUSTRE_IMP_REPLAY;
2472 req->rq_phase = RQ_PHASE_NEW;
2473 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2475 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2477 req->rq_interpret_reply = ptlrpc_replay_interpret;
2478 /* Readjust the timeout for current conditions */
2479 ptlrpc_at_set_req_timeout(req);
2481 DEBUG_REQ(D_HA, req, "REPLAY");
2483 atomic_inc(&req->rq_import->imp_replay_inflight);
2484 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2486 ptlrpcd_add_req(req, PSCOPE_OTHER);
2490 void ptlrpc_abort_inflight(struct obd_import *imp)
2492 struct list_head *tmp, *n;
2495 /* Make sure that no new requests get processed for this import.
2496 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2497 * this flag and then putting requests on sending_list or delayed_list.
2499 spin_lock(&imp->imp_lock);
2501 /* XXX locking? Maybe we should remove each request with the list
2502 * locked? Also, how do we know if the requests on the list are
2503 * being freed at this time?
2505 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2506 struct ptlrpc_request *req =
2507 list_entry(tmp, struct ptlrpc_request, rq_list);
2509 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2511 spin_lock (&req->rq_lock);
2512 if (req->rq_import_generation < imp->imp_generation) {
2514 req->rq_status = -EINTR;
2515 ptlrpc_client_wake_req(req);
2517 spin_unlock (&req->rq_lock);
2520 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2521 struct ptlrpc_request *req =
2522 list_entry(tmp, struct ptlrpc_request, rq_list);
2524 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2526 spin_lock (&req->rq_lock);
2527 if (req->rq_import_generation < imp->imp_generation) {
2529 req->rq_status = -EINTR;
2530 ptlrpc_client_wake_req(req);
2532 spin_unlock (&req->rq_lock);
2535 /* Last chance to free reqs left on the replay list, but we
2536 * will still leak reqs that haven't committed. */
2537 if (imp->imp_replayable)
2538 ptlrpc_free_committed(imp);
2540 spin_unlock(&imp->imp_lock);
2545 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2547 struct list_head *tmp, *pos;
2549 LASSERT(set != NULL);
2551 list_for_each_safe(pos, tmp, &set->set_requests) {
2552 struct ptlrpc_request *req =
2553 list_entry(pos, struct ptlrpc_request, rq_set_chain);
2555 spin_lock(&req->rq_lock);
2556 if (req->rq_phase != RQ_PHASE_RPC) {
2557 spin_unlock(&req->rq_lock);
2562 req->rq_status = -EINTR;
2563 ptlrpc_client_wake_req(req);
2564 spin_unlock(&req->rq_lock);
2568 static __u64 ptlrpc_last_xid;
2569 static spinlock_t ptlrpc_last_xid_lock;
2571 /* Initialize the XID for the node. This is common among all requests on
2572 * this node, and only requires the property that it is monotonically
2573 * increasing. It does not need to be sequential. Since this is also used
2574 * as the RDMA match bits, it is important that a single client NOT have
2575 * the same match bits for two different in-flight requests, hence we do
2576 * NOT want to have an XID per target or similar.
2578 * To avoid an unlikely collision between match bits after a client reboot
2579 * (which would deliver old data into the wrong RDMA buffer) initialize
2580 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2581 * If the time is clearly incorrect, we instead use a 62-bit random number.
2582 * In the worst case the random number will overflow 1M RPCs per second in
2583 * 9133 years, or permutations thereof.
2585 #define YEAR_2004 (1ULL << 30)
2586 void ptlrpc_init_xid(void)
2588 time_t now = cfs_time_current_sec();
2590 spin_lock_init(&ptlrpc_last_xid_lock);
2591 if (now < YEAR_2004) {
2592 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2593 ptlrpc_last_xid >>= 2;
2594 ptlrpc_last_xid |= (1ULL << 61);
2596 ptlrpc_last_xid = (__u64)now << 20;
2600 __u64 ptlrpc_next_xid(void)
2603 spin_lock(&ptlrpc_last_xid_lock);
2604 tmp = ++ptlrpc_last_xid;
2605 spin_unlock(&ptlrpc_last_xid_lock);
2609 __u64 ptlrpc_sample_next_xid(void)
2611 #if BITS_PER_LONG == 32
2612 /* need to avoid possible word tearing on 32-bit systems */
2614 spin_lock(&ptlrpc_last_xid_lock);
2615 tmp = ptlrpc_last_xid + 1;
2616 spin_unlock(&ptlrpc_last_xid_lock);
2619 /* No need to lock, since returned value is racy anyways */
2620 return ptlrpc_last_xid + 1;
2623 EXPORT_SYMBOL(ptlrpc_sample_next_xid);