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 cfs_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;
203 at = &req->rq_import->imp_at;
204 idx = import_at_get_index(req->rq_import,
205 req->rq_request_portal);
206 serv_est = at_get(&at->iat_service_estimate[idx]);
207 req->rq_timeout = at_est2timeout(serv_est);
209 /* We could get even fancier here, using history to predict increased
212 /* Let the server know what this RPC timeout is by putting it in the
214 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
217 /* Adjust max service estimate based on server value */
218 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
219 unsigned int serv_est)
225 LASSERT(req->rq_import);
226 at = &req->rq_import->imp_at;
228 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
229 /* max service estimates are tracked on the server side,
230 so just keep minimal history here */
231 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
233 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
234 "has changed from %d to %d\n",
235 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
236 oldse, at_get(&at->iat_service_estimate[idx]));
239 /* Expected network latency per remote node (secs) */
240 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
242 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
245 /* Adjust expected network latency */
246 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
247 unsigned int service_time)
249 unsigned int nl, oldnl;
251 time_t now = cfs_time_current_sec();
253 LASSERT(req->rq_import);
254 at = &req->rq_import->imp_at;
256 /* Network latency is total time less server processing time */
257 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
258 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
259 CWARN("Reported service time %u > total measured time "
260 CFS_DURATION_T"\n", service_time,
261 cfs_time_sub(now, req->rq_sent));
263 oldnl = at_measured(&at->iat_net_latency, nl);
265 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
266 "has changed from %d to %d\n",
267 req->rq_import->imp_obd->obd_name,
269 &req->rq_import->imp_connection->c_remote_uuid),
270 oldnl, at_get(&at->iat_net_latency));
273 static int unpack_reply(struct ptlrpc_request *req)
277 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
278 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
280 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
285 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
287 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
294 * Handle an early reply message, called with the rq_lock held.
295 * If anything goes wrong just ignore it - same as if it never happened
297 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
299 struct ptlrpc_request *early_req;
305 cfs_spin_unlock(&req->rq_lock);
307 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
309 cfs_spin_lock(&req->rq_lock);
313 rc = unpack_reply(early_req);
315 /* Expecting to increase the service time estimate here */
316 ptlrpc_at_adj_service(req,
317 lustre_msg_get_timeout(early_req->rq_repmsg));
318 ptlrpc_at_adj_net_latency(req,
319 lustre_msg_get_service_time(early_req->rq_repmsg));
322 sptlrpc_cli_finish_early_reply(early_req);
324 cfs_spin_lock(&req->rq_lock);
327 /* Adjust the local timeout for this req */
328 ptlrpc_at_set_req_timeout(req);
330 olddl = req->rq_deadline;
331 /* server assumes it now has rq_timeout from when it sent the
332 early reply, so client should give it at least that long. */
333 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
334 ptlrpc_at_get_net_latency(req);
336 DEBUG_REQ(D_ADAPTTO, req,
337 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
338 "("CFS_DURATION_T"s)", req->rq_early_count,
339 cfs_time_sub(req->rq_deadline,
340 cfs_time_current_sec()),
341 cfs_time_sub(req->rq_deadline, olddl));
347 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
350 struct ptlrpc_request *req;
352 LASSERT(pool != NULL);
354 cfs_spin_lock(&pool->prp_lock);
355 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
356 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
357 cfs_list_del(&req->rq_list);
358 LASSERT(req->rq_reqbuf);
359 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
360 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
361 OBD_FREE(req, sizeof(*req));
363 cfs_spin_unlock(&pool->prp_lock);
364 OBD_FREE(pool, sizeof(*pool));
367 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
372 while (size < pool->prp_rq_size)
375 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
376 size == pool->prp_rq_size,
377 "Trying to change pool size with nonempty pool "
378 "from %d to %d bytes\n", pool->prp_rq_size, size);
380 cfs_spin_lock(&pool->prp_lock);
381 pool->prp_rq_size = size;
382 for (i = 0; i < num_rq; i++) {
383 struct ptlrpc_request *req;
384 struct lustre_msg *msg;
386 cfs_spin_unlock(&pool->prp_lock);
387 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
390 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
392 OBD_FREE(req, sizeof(struct ptlrpc_request));
395 req->rq_reqbuf = msg;
396 req->rq_reqbuf_len = size;
398 cfs_spin_lock(&pool->prp_lock);
399 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
401 cfs_spin_unlock(&pool->prp_lock);
405 struct ptlrpc_request_pool *
406 ptlrpc_init_rq_pool(int num_rq, int msgsize,
407 void (*populate_pool)(struct ptlrpc_request_pool *, int))
409 struct ptlrpc_request_pool *pool;
411 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
415 /* Request next power of two for the allocation, because internally
416 kernel would do exactly this */
418 cfs_spin_lock_init(&pool->prp_lock);
419 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
420 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
421 pool->prp_populate = populate_pool;
423 populate_pool(pool, num_rq);
425 if (cfs_list_empty(&pool->prp_req_list)) {
426 /* have not allocated a single request for the pool */
427 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
433 static struct ptlrpc_request *
434 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
436 struct ptlrpc_request *request;
437 struct lustre_msg *reqbuf;
442 cfs_spin_lock(&pool->prp_lock);
444 /* See if we have anything in a pool, and bail out if nothing,
445 * in writeout path, where this matters, this is safe to do, because
446 * nothing is lost in this case, and when some in-flight requests
447 * complete, this code will be called again. */
448 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
449 cfs_spin_unlock(&pool->prp_lock);
453 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
455 cfs_list_del_init(&request->rq_list);
456 cfs_spin_unlock(&pool->prp_lock);
458 LASSERT(request->rq_reqbuf);
459 LASSERT(request->rq_pool);
461 reqbuf = request->rq_reqbuf;
462 memset(request, 0, sizeof(*request));
463 request->rq_reqbuf = reqbuf;
464 request->rq_reqbuf_len = pool->prp_rq_size;
465 request->rq_pool = pool;
470 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
472 struct ptlrpc_request_pool *pool = request->rq_pool;
474 cfs_spin_lock(&pool->prp_lock);
475 LASSERT(cfs_list_empty(&request->rq_list));
476 LASSERT(!request->rq_receiving_reply);
477 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
478 cfs_spin_unlock(&pool->prp_lock);
481 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
482 __u32 version, int opcode,
483 int count, __u32 *lengths, char **bufs,
484 struct ptlrpc_cli_ctx *ctx)
486 struct obd_import *imp = request->rq_import;
491 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
493 rc = sptlrpc_req_get_ctx(request);
498 sptlrpc_req_set_flavor(request, opcode);
500 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
503 LASSERT(!request->rq_pool);
507 lustre_msg_add_version(request->rq_reqmsg, version);
508 request->rq_send_state = LUSTRE_IMP_FULL;
509 request->rq_type = PTL_RPC_MSG_REQUEST;
510 request->rq_export = NULL;
512 request->rq_req_cbid.cbid_fn = request_out_callback;
513 request->rq_req_cbid.cbid_arg = request;
515 request->rq_reply_cbid.cbid_fn = reply_in_callback;
516 request->rq_reply_cbid.cbid_arg = request;
518 request->rq_reply_deadline = 0;
519 request->rq_phase = RQ_PHASE_NEW;
520 request->rq_next_phase = RQ_PHASE_UNDEFINED;
522 request->rq_request_portal = imp->imp_client->cli_request_portal;
523 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
525 ptlrpc_at_set_req_timeout(request);
527 cfs_spin_lock_init(&request->rq_lock);
528 CFS_INIT_LIST_HEAD(&request->rq_list);
529 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
530 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
531 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
532 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
533 CFS_INIT_LIST_HEAD(&request->rq_history_list);
534 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
535 cfs_waitq_init(&request->rq_reply_waitq);
536 cfs_waitq_init(&request->rq_set_waitq);
537 request->rq_xid = ptlrpc_next_xid();
538 cfs_atomic_set(&request->rq_refcount, 1);
540 lustre_msg_set_opc(request->rq_reqmsg, opcode);
544 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
546 class_import_put(imp);
550 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
551 __u32 version, int opcode, char **bufs,
552 struct ptlrpc_cli_ctx *ctx)
556 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
557 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
558 request->rq_pill.rc_area[RCL_CLIENT],
561 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
563 int ptlrpc_request_pack(struct ptlrpc_request *request,
564 __u32 version, int opcode)
566 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
570 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
571 struct ptlrpc_request_pool *pool)
573 struct ptlrpc_request *request = NULL;
576 request = ptlrpc_prep_req_from_pool(pool);
579 OBD_ALLOC_PTR(request);
582 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
583 LASSERT(imp != LP_POISON);
584 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
586 LASSERT(imp->imp_client != LP_POISON);
588 request->rq_import = class_import_get(imp);
590 CERROR("request allocation out of memory\n");
596 static struct ptlrpc_request *
597 ptlrpc_request_alloc_internal(struct obd_import *imp,
598 struct ptlrpc_request_pool * pool,
599 const struct req_format *format)
601 struct ptlrpc_request *request;
603 request = __ptlrpc_request_alloc(imp, pool);
607 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
608 req_capsule_set(&request->rq_pill, format);
612 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
613 const struct req_format *format)
615 return ptlrpc_request_alloc_internal(imp, NULL, format);
618 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
619 struct ptlrpc_request_pool * pool,
620 const struct req_format *format)
622 return ptlrpc_request_alloc_internal(imp, pool, format);
625 void ptlrpc_request_free(struct ptlrpc_request *request)
627 if (request->rq_pool)
628 __ptlrpc_free_req_to_pool(request);
630 OBD_FREE_PTR(request);
633 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
634 const struct req_format *format,
635 __u32 version, int opcode)
637 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
641 rc = ptlrpc_request_pack(req, version, opcode);
643 ptlrpc_request_free(req);
650 struct ptlrpc_request *
651 ptlrpc_prep_req_pool(struct obd_import *imp,
652 __u32 version, int opcode,
653 int count, __u32 *lengths, char **bufs,
654 struct ptlrpc_request_pool *pool)
656 struct ptlrpc_request *request;
659 request = __ptlrpc_request_alloc(imp, pool);
663 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
664 lengths, bufs, NULL);
666 ptlrpc_request_free(request);
672 struct ptlrpc_request *
673 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
674 __u32 *lengths, char **bufs)
676 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
680 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
681 unsigned int timeout,
682 ptlrpc_interpterer_t interpreter)
684 struct ptlrpc_request *request = NULL;
687 OBD_ALLOC(request, sizeof(*request));
689 CERROR("request allocation out of memory\n");
693 request->rq_send_state = LUSTRE_IMP_FULL;
694 request->rq_type = PTL_RPC_MSG_REQUEST;
695 request->rq_import = class_import_get(imp);
696 request->rq_export = NULL;
697 request->rq_import_generation = imp->imp_generation;
699 request->rq_timeout = timeout;
700 request->rq_sent = cfs_time_current_sec();
701 request->rq_deadline = request->rq_sent + timeout;
702 request->rq_reply_deadline = request->rq_deadline;
703 request->rq_interpret_reply = interpreter;
704 request->rq_phase = RQ_PHASE_RPC;
705 request->rq_next_phase = RQ_PHASE_INTERPRET;
706 /* don't want reply */
707 request->rq_receiving_reply = 0;
708 request->rq_must_unlink = 0;
709 request->rq_no_delay = request->rq_no_resend = 1;
710 request->rq_fake = 1;
712 cfs_spin_lock_init(&request->rq_lock);
713 CFS_INIT_LIST_HEAD(&request->rq_list);
714 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
715 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
716 CFS_INIT_LIST_HEAD(&request->rq_history_list);
717 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
718 cfs_waitq_init(&request->rq_reply_waitq);
719 cfs_waitq_init(&request->rq_set_waitq);
721 request->rq_xid = ptlrpc_next_xid();
722 cfs_atomic_set(&request->rq_refcount, 1);
727 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
729 /* if we kill request before timeout - need adjust counter */
730 if (req->rq_phase == RQ_PHASE_RPC) {
731 struct ptlrpc_request_set *set = req->rq_set;
734 cfs_atomic_dec(&set->set_remaining);
737 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
738 cfs_list_del_init(&req->rq_list);
742 struct ptlrpc_request_set *ptlrpc_prep_set(void)
744 struct ptlrpc_request_set *set;
747 OBD_ALLOC(set, sizeof *set);
750 CFS_INIT_LIST_HEAD(&set->set_requests);
751 cfs_waitq_init(&set->set_waitq);
752 cfs_atomic_set(&set->set_remaining, 0);
753 cfs_spin_lock_init(&set->set_new_req_lock);
754 CFS_INIT_LIST_HEAD(&set->set_new_requests);
755 CFS_INIT_LIST_HEAD(&set->set_cblist);
760 /* Finish with this set; opposite of prep_set. */
761 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
769 /* Requests on the set should either all be completed, or all be new */
770 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
771 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
772 cfs_list_for_each (tmp, &set->set_requests) {
773 struct ptlrpc_request *req =
774 cfs_list_entry(tmp, struct ptlrpc_request,
777 LASSERT(req->rq_phase == expected_phase);
781 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
782 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
783 cfs_atomic_read(&set->set_remaining), n);
785 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
786 struct ptlrpc_request *req =
787 cfs_list_entry(tmp, struct ptlrpc_request,
789 cfs_list_del_init(&req->rq_set_chain);
791 LASSERT(req->rq_phase == expected_phase);
793 if (req->rq_phase == RQ_PHASE_NEW) {
794 ptlrpc_req_interpret(NULL, req, -EBADR);
795 cfs_atomic_dec(&set->set_remaining);
799 ptlrpc_req_finished (req);
802 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
804 OBD_FREE(set, sizeof(*set));
808 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
809 set_interpreter_func fn, void *data)
811 struct ptlrpc_set_cbdata *cbdata;
813 OBD_ALLOC_PTR(cbdata);
817 cbdata->psc_interpret = fn;
818 cbdata->psc_data = data;
819 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
824 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
825 struct ptlrpc_request *req)
827 /* The set takes over the caller's request reference */
828 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
830 cfs_atomic_inc(&set->set_remaining);
834 * Lock so many callers can add things, the context that owns the set
835 * is supposed to notice these and move them into the set proper.
837 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
838 struct ptlrpc_request *req)
840 struct ptlrpc_request_set *set = pc->pc_set;
843 * Let caller know that we stopped and will not handle this request.
844 * It needs to take care itself of request.
846 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags))
849 cfs_spin_lock(&set->set_new_req_lock);
851 * The set takes over the caller's request reference.
853 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
855 cfs_spin_unlock(&set->set_new_req_lock);
857 cfs_waitq_signal(&set->set_waitq);
862 * Based on the current state of the import, determine if the request
863 * can be sent, is an error, or should be delayed.
865 * Returns true if this request should be delayed. If false, and
866 * *status is set, then the request can not be sent and *status is the
867 * error code. If false and status is 0, then request can be sent.
869 * The imp->imp_lock must be held.
871 static int ptlrpc_import_delay_req(struct obd_import *imp,
872 struct ptlrpc_request *req, int *status)
877 LASSERT (status != NULL);
880 if (req->rq_ctx_init || req->rq_ctx_fini) {
881 /* always allow ctx init/fini rpc go through */
882 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
883 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
886 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
887 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
889 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
890 imp->imp_state == LUSTRE_IMP_CONNECTING) {
891 /* allow CONNECT even if import is invalid */ ;
892 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
893 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
896 } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
897 imp->imp_obd->obd_no_recov) {
898 /* If the import has been invalidated (such as by an OST
899 * failure), and if the import(MGC) tried all of its connection
900 * list (Bug 13464), the request must fail with -ESHUTDOWN.
901 * This indicates the requests should be discarded; an -EIO
902 * may result in a resend of the request. */
903 if (!imp->imp_deactive)
904 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
905 *status = -ESHUTDOWN; /* bz 12940 */
906 } else if (req->rq_import_generation != imp->imp_generation) {
907 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
909 } else if (req->rq_send_state != imp->imp_state) {
910 /* invalidate in progress - any requests should be drop */
911 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
912 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
914 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
915 *status = -EWOULDBLOCK;
924 /* Conditionally suppress specific console messages */
925 static int ptlrpc_console_allow(struct ptlrpc_request *req)
927 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
930 /* Suppress particular reconnect errors which are to be expected. No
931 * errors are suppressed for the initial connection on an import */
932 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
933 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
935 /* Suppress timed out reconnect requests */
936 if (req->rq_timedout)
939 /* Suppress unavailable/again reconnect requests */
940 err = lustre_msg_get_status(req->rq_repmsg);
941 if (err == -ENODEV || err == -EAGAIN)
948 static int ptlrpc_check_status(struct ptlrpc_request *req)
953 err = lustre_msg_get_status(req->rq_repmsg);
954 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
955 struct obd_import *imp = req->rq_import;
956 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
957 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
958 " with %s. The %s operation failed with %d\n",
959 libcfs_nid2str(imp->imp_connection->c_peer.nid),
960 ll_opcode2str(opc), err);
961 RETURN(err < 0 ? err : -EINVAL);
965 DEBUG_REQ(D_INFO, req, "status is %d", err);
966 } else if (err > 0) {
967 /* XXX: translate this error from net to host */
968 DEBUG_REQ(D_INFO, req, "status is %d", err);
971 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
972 struct obd_import *imp = req->rq_import;
973 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
975 if (ptlrpc_console_allow(req))
976 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
977 "communicating with %s. The %s "
978 "operation failed with %d\n",
980 imp->imp_connection->c_peer.nid),
981 ll_opcode2str(opc), err);
983 RETURN(err < 0 ? err : -EINVAL);
990 * save pre-versions for replay
992 static void ptlrpc_save_versions(struct ptlrpc_request *req)
994 struct lustre_msg *repmsg = req->rq_repmsg;
995 struct lustre_msg *reqmsg = req->rq_reqmsg;
996 __u64 *versions = lustre_msg_get_versions(repmsg);
999 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1003 lustre_msg_set_versions(reqmsg, versions);
1004 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1005 versions[0], versions[1]);
1011 * Callback function called when client receives RPC reply for \a req.
1013 static int after_reply(struct ptlrpc_request *req)
1015 struct obd_import *imp = req->rq_import;
1016 struct obd_device *obd = req->rq_import->imp_obd;
1018 struct timeval work_start;
1022 LASSERT(obd != NULL);
1023 /* repbuf must be unlinked */
1024 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1026 if (req->rq_reply_truncate) {
1027 if (req->rq_no_resend) {
1028 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1029 " expected: %d, actual size: %d",
1030 req->rq_nob_received, req->rq_repbuf_len);
1034 sptlrpc_cli_free_repbuf(req);
1035 /* Pass the required reply buffer size (include
1036 * space for early reply).
1037 * NB: no need to roundup because alloc_repbuf
1038 * will roundup it */
1039 req->rq_replen = req->rq_nob_received;
1040 req->rq_nob_received = 0;
1046 * NB Until this point, the whole of the incoming message,
1047 * including buflens, status etc is in the sender's byte order.
1049 rc = sptlrpc_cli_unwrap_reply(req);
1051 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1056 * Security layer unwrap might ask resend this request.
1061 rc = unpack_reply(req);
1065 cfs_gettimeofday(&work_start);
1066 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1067 if (obd->obd_svc_stats != NULL) {
1068 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1070 ptlrpc_lprocfs_rpc_sent(req, timediff);
1073 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1074 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1075 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1076 lustre_msg_get_type(req->rq_repmsg));
1080 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1081 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1082 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1083 ptlrpc_at_adj_net_latency(req,
1084 lustre_msg_get_service_time(req->rq_repmsg));
1086 rc = ptlrpc_check_status(req);
1087 imp->imp_connect_error = rc;
1091 * Either we've been evicted, or the server has failed for
1092 * some reason. Try to reconnect, and if that fails, punt to
1095 if (ll_rpc_recoverable_error(rc)) {
1096 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1097 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1100 ptlrpc_request_handle_notconn(req);
1105 * Let's look if server sent slv. Do it only for RPC with
1108 ldlm_cli_update_pool(req);
1112 * Store transno in reqmsg for replay.
1114 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1115 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1116 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1119 if (imp->imp_replayable) {
1120 cfs_spin_lock(&imp->imp_lock);
1122 * No point in adding already-committed requests to the replay
1123 * list, we will just remove them immediately. b=9829
1125 if (req->rq_transno != 0 &&
1127 lustre_msg_get_last_committed(req->rq_repmsg) ||
1129 /** version recovery */
1130 ptlrpc_save_versions(req);
1131 ptlrpc_retain_replayable_request(req, imp);
1132 } else if (req->rq_commit_cb != NULL) {
1133 cfs_spin_unlock(&imp->imp_lock);
1134 req->rq_commit_cb(req);
1135 cfs_spin_lock(&imp->imp_lock);
1139 * Replay-enabled imports return commit-status information.
1141 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1142 imp->imp_peer_committed_transno =
1143 lustre_msg_get_last_committed(req->rq_repmsg);
1145 ptlrpc_free_committed(imp);
1146 cfs_spin_unlock(&imp->imp_lock);
1152 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1154 struct obd_import *imp;
1158 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1159 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1162 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1164 imp = req->rq_import;
1165 cfs_spin_lock(&imp->imp_lock);
1167 req->rq_import_generation = imp->imp_generation;
1169 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1170 cfs_spin_lock(&req->rq_lock);
1171 req->rq_waiting = 1;
1172 cfs_spin_unlock(&req->rq_lock);
1174 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1175 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1176 ptlrpc_import_state_name(req->rq_send_state),
1177 ptlrpc_import_state_name(imp->imp_state));
1178 LASSERT(cfs_list_empty(&req->rq_list));
1179 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1180 cfs_atomic_inc(&req->rq_import->imp_inflight);
1181 cfs_spin_unlock(&imp->imp_lock);
1186 cfs_spin_unlock(&imp->imp_lock);
1187 req->rq_status = rc;
1188 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1192 LASSERT(cfs_list_empty(&req->rq_list));
1193 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1194 cfs_atomic_inc(&req->rq_import->imp_inflight);
1195 cfs_spin_unlock(&imp->imp_lock);
1197 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1199 rc = sptlrpc_req_refresh_ctx(req, -1);
1202 req->rq_status = rc;
1205 req->rq_wait_ctx = 1;
1210 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1211 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1212 imp->imp_obd->obd_uuid.uuid,
1213 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1214 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1215 lustre_msg_get_opc(req->rq_reqmsg));
1217 rc = ptl_send_rpc(req, 0);
1219 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1220 req->rq_net_err = 1;
1226 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1227 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1230 int force_timer_recalc = 0;
1233 if (cfs_atomic_read(&set->set_remaining) == 0)
1236 cfs_list_for_each(tmp, &set->set_requests) {
1237 struct ptlrpc_request *req =
1238 cfs_list_entry(tmp, struct ptlrpc_request,
1240 struct obd_import *imp = req->rq_import;
1241 int unregistered = 0;
1244 if (req->rq_phase == RQ_PHASE_NEW &&
1245 ptlrpc_send_new_req(req)) {
1246 force_timer_recalc = 1;
1249 /* delayed send - skip */
1250 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1253 if (!(req->rq_phase == RQ_PHASE_RPC ||
1254 req->rq_phase == RQ_PHASE_BULK ||
1255 req->rq_phase == RQ_PHASE_INTERPRET ||
1256 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1257 req->rq_phase == RQ_PHASE_COMPLETE)) {
1258 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1262 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1263 LASSERT(req->rq_next_phase != req->rq_phase);
1264 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1267 * Skip processing until reply is unlinked. We
1268 * can't return to pool before that and we can't
1269 * call interpret before that. We need to make
1270 * sure that all rdma transfers finished and will
1271 * not corrupt any data.
1273 if (ptlrpc_client_recv_or_unlink(req) ||
1274 ptlrpc_client_bulk_active(req))
1278 * Turn fail_loc off to prevent it from looping
1281 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1282 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1285 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1286 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1291 * Move to next phase if reply was successfully
1294 ptlrpc_rqphase_move(req, req->rq_next_phase);
1297 if (req->rq_phase == RQ_PHASE_COMPLETE)
1300 if (req->rq_phase == RQ_PHASE_INTERPRET)
1301 GOTO(interpret, req->rq_status);
1304 * Note that this also will start async reply unlink.
1306 if (req->rq_net_err && !req->rq_timedout) {
1307 ptlrpc_expire_one_request(req, 1);
1310 * Check if we still need to wait for unlink.
1312 if (ptlrpc_client_recv_or_unlink(req) ||
1313 ptlrpc_client_bulk_active(req))
1318 cfs_spin_lock(&req->rq_lock);
1319 req->rq_replied = 0;
1320 cfs_spin_unlock(&req->rq_lock);
1321 if (req->rq_status == 0)
1322 req->rq_status = -EIO;
1323 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1324 GOTO(interpret, req->rq_status);
1327 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1328 * so it sets rq_intr regardless of individual rpc
1329 * timeouts. The synchronous IO waiting path sets
1330 * rq_intr irrespective of whether ptlrpcd
1331 * has seen a timeout. Our policy is to only interpret
1332 * interrupted rpcs after they have timed out, so we
1333 * need to enforce that here.
1336 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1337 req->rq_wait_ctx)) {
1338 req->rq_status = -EINTR;
1339 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1340 GOTO(interpret, req->rq_status);
1343 if (req->rq_phase == RQ_PHASE_RPC) {
1344 if (req->rq_timedout || req->rq_resend ||
1345 req->rq_waiting || req->rq_wait_ctx) {
1348 if (!ptlrpc_unregister_reply(req, 1))
1351 cfs_spin_lock(&imp->imp_lock);
1352 if (ptlrpc_import_delay_req(imp, req, &status)){
1353 /* put on delay list - only if we wait
1354 * recovery finished - before send */
1355 cfs_list_del_init(&req->rq_list);
1356 cfs_list_add_tail(&req->rq_list,
1359 cfs_spin_unlock(&imp->imp_lock);
1364 req->rq_status = status;
1365 ptlrpc_rqphase_move(req,
1366 RQ_PHASE_INTERPRET);
1367 cfs_spin_unlock(&imp->imp_lock);
1368 GOTO(interpret, req->rq_status);
1370 if (req->rq_no_resend && !req->rq_wait_ctx) {
1371 req->rq_status = -ENOTCONN;
1372 ptlrpc_rqphase_move(req,
1373 RQ_PHASE_INTERPRET);
1374 cfs_spin_unlock(&imp->imp_lock);
1375 GOTO(interpret, req->rq_status);
1378 cfs_list_del_init(&req->rq_list);
1379 cfs_list_add_tail(&req->rq_list,
1380 &imp->imp_sending_list);
1382 cfs_spin_unlock(&imp->imp_lock);
1384 cfs_spin_lock(&req->rq_lock);
1385 req->rq_waiting = 0;
1386 cfs_spin_unlock(&req->rq_lock);
1388 if (req->rq_timedout||req->rq_resend) {
1389 /* This is re-sending anyways,
1390 * let's mark req as resend. */
1391 cfs_spin_lock(&req->rq_lock);
1393 cfs_spin_unlock(&req->rq_lock);
1397 if (!ptlrpc_unregister_bulk(req, 1))
1400 /* ensure previous bulk fails */
1401 old_xid = req->rq_xid;
1402 req->rq_xid = ptlrpc_next_xid();
1403 CDEBUG(D_HA, "resend bulk "
1406 old_xid, req->rq_xid);
1410 * rq_wait_ctx is only touched by ptlrpcd,
1411 * so no lock is needed here.
1413 status = sptlrpc_req_refresh_ctx(req, -1);
1416 req->rq_status = status;
1417 cfs_spin_lock(&req->rq_lock);
1418 req->rq_wait_ctx = 0;
1419 cfs_spin_unlock(&req->rq_lock);
1420 force_timer_recalc = 1;
1422 cfs_spin_lock(&req->rq_lock);
1423 req->rq_wait_ctx = 1;
1424 cfs_spin_unlock(&req->rq_lock);
1429 cfs_spin_lock(&req->rq_lock);
1430 req->rq_wait_ctx = 0;
1431 cfs_spin_unlock(&req->rq_lock);
1434 rc = ptl_send_rpc(req, 0);
1436 DEBUG_REQ(D_HA, req, "send failed (%d)",
1438 force_timer_recalc = 1;
1439 cfs_spin_lock(&req->rq_lock);
1440 req->rq_net_err = 1;
1441 cfs_spin_unlock(&req->rq_lock);
1443 /* need to reset the timeout */
1444 force_timer_recalc = 1;
1447 cfs_spin_lock(&req->rq_lock);
1449 if (ptlrpc_client_early(req)) {
1450 ptlrpc_at_recv_early_reply(req);
1451 cfs_spin_unlock(&req->rq_lock);
1455 /* Still waiting for a reply? */
1456 if (ptlrpc_client_recv(req)) {
1457 cfs_spin_unlock(&req->rq_lock);
1461 /* Did we actually receive a reply? */
1462 if (!ptlrpc_client_replied(req)) {
1463 cfs_spin_unlock(&req->rq_lock);
1467 cfs_spin_unlock(&req->rq_lock);
1469 /* unlink from net because we are going to
1470 * swab in-place of reply buffer */
1471 unregistered = ptlrpc_unregister_reply(req, 1);
1475 req->rq_status = after_reply(req);
1479 /* If there is no bulk associated with this request,
1480 * then we're done and should let the interpreter
1481 * process the reply. Similarly if the RPC returned
1482 * an error, and therefore the bulk will never arrive.
1484 if (req->rq_bulk == NULL || req->rq_status != 0) {
1485 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1486 GOTO(interpret, req->rq_status);
1489 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1492 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1493 if (ptlrpc_client_bulk_active(req))
1496 if (!req->rq_bulk->bd_success) {
1497 /* The RPC reply arrived OK, but the bulk screwed
1498 * up! Dead weird since the server told us the RPC
1499 * was good after getting the REPLY for her GET or
1500 * the ACK for her PUT. */
1501 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1505 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1508 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1510 /* This moves to "unregistering" phase we need to wait for
1512 if (!unregistered && !ptlrpc_unregister_reply(req, 1))
1515 if (!ptlrpc_unregister_bulk(req, 1))
1518 /* When calling interpret receiving already should be
1520 LASSERT(!req->rq_receiving_reply);
1522 ptlrpc_req_interpret(env, req, req->rq_status);
1524 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1526 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1527 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1528 imp->imp_obd->obd_uuid.uuid,
1529 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1531 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1532 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1534 cfs_spin_lock(&imp->imp_lock);
1535 /* Request already may be not on sending or delaying list. This
1536 * may happen in the case of marking it erroneous for the case
1537 * ptlrpc_import_delay_req(req, status) find it impossible to
1538 * allow sending this rpc and returns *status != 0. */
1539 if (!cfs_list_empty(&req->rq_list)) {
1540 cfs_list_del_init(&req->rq_list);
1541 cfs_atomic_dec(&imp->imp_inflight);
1543 cfs_spin_unlock(&imp->imp_lock);
1545 cfs_atomic_dec(&set->set_remaining);
1546 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1549 /* If we hit an error, we want to recover promptly. */
1550 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1553 /* Return 1 if we should give up, else 0 */
1554 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1556 struct obd_import *imp = req->rq_import;
1560 cfs_spin_lock(&req->rq_lock);
1561 req->rq_timedout = 1;
1562 cfs_spin_unlock(&req->rq_lock);
1564 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req,
1565 "Request x"LPU64" sent from %s to NID %s "CFS_DURATION_T"s "
1566 "ago has %s ("CFS_DURATION_T"s prior to deadline).\n",
1567 req->rq_xid, imp ? imp->imp_obd->obd_name : "<?>",
1568 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1569 cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1570 req->rq_net_err ? "failed due to network error" : "timed out",
1571 cfs_time_sub(req->rq_deadline, req->rq_sent));
1573 if (imp != NULL && obd_debug_peer_on_timeout)
1574 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1576 ptlrpc_unregister_reply(req, async_unlink);
1577 ptlrpc_unregister_bulk(req, async_unlink);
1579 if (obd_dump_on_timeout)
1580 libcfs_debug_dumplog();
1583 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1590 cfs_atomic_inc(&imp->imp_timeouts);
1592 /* The DLM server doesn't want recovery run on its imports. */
1593 if (imp->imp_dlm_fake)
1596 /* If this request is for recovery or other primordial tasks,
1597 * then error it out here. */
1598 if (req->rq_ctx_init || req->rq_ctx_fini ||
1599 req->rq_send_state != LUSTRE_IMP_FULL ||
1600 imp->imp_obd->obd_no_recov) {
1601 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1602 ptlrpc_import_state_name(req->rq_send_state),
1603 ptlrpc_import_state_name(imp->imp_state));
1604 cfs_spin_lock(&req->rq_lock);
1605 req->rq_status = -ETIMEDOUT;
1607 cfs_spin_unlock(&req->rq_lock);
1611 /* if a request can't be resent we can't wait for an answer after
1613 if (req->rq_no_resend) {
1614 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1618 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1623 int ptlrpc_expired_set(void *data)
1625 struct ptlrpc_request_set *set = data;
1627 time_t now = cfs_time_current_sec();
1630 LASSERT(set != NULL);
1633 * A timeout expired. See which reqs it applies to...
1635 cfs_list_for_each (tmp, &set->set_requests) {
1636 struct ptlrpc_request *req =
1637 cfs_list_entry(tmp, struct ptlrpc_request,
1640 /* don't expire request waiting for context */
1641 if (req->rq_wait_ctx)
1644 /* Request in-flight? */
1645 if (!((req->rq_phase == RQ_PHASE_RPC &&
1646 !req->rq_waiting && !req->rq_resend) ||
1647 (req->rq_phase == RQ_PHASE_BULK)))
1650 if (req->rq_timedout || /* already dealt with */
1651 req->rq_deadline > now) /* not expired */
1654 /* Deal with this guy. Do it asynchronously to not block
1655 * ptlrpcd thread. */
1656 ptlrpc_expire_one_request(req, 1);
1660 * When waiting for a whole set, we always break out of the
1661 * sleep so we can recalculate the timeout, or enable interrupts
1662 * if everyone's timed out.
1667 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1669 cfs_spin_lock(&req->rq_lock);
1671 cfs_spin_unlock(&req->rq_lock);
1674 void ptlrpc_interrupted_set(void *data)
1676 struct ptlrpc_request_set *set = data;
1679 LASSERT(set != NULL);
1680 CERROR("INTERRUPTED SET %p\n", set);
1682 cfs_list_for_each(tmp, &set->set_requests) {
1683 struct ptlrpc_request *req =
1684 cfs_list_entry(tmp, struct ptlrpc_request,
1687 if (req->rq_phase != RQ_PHASE_RPC &&
1688 req->rq_phase != RQ_PHASE_UNREGISTERING)
1691 ptlrpc_mark_interrupted(req);
1696 * Get the smallest timeout in the set; this does NOT set a timeout.
1698 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1701 time_t now = cfs_time_current_sec();
1703 struct ptlrpc_request *req;
1707 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1709 cfs_list_for_each(tmp, &set->set_requests) {
1710 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1713 * Request in-flight?
1715 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1716 (req->rq_phase == RQ_PHASE_BULK) ||
1717 (req->rq_phase == RQ_PHASE_NEW)))
1721 * Already timed out.
1723 if (req->rq_timedout)
1729 if (req->rq_wait_ctx)
1732 if (req->rq_phase == RQ_PHASE_NEW)
1733 deadline = req->rq_sent;
1735 deadline = req->rq_sent + req->rq_timeout;
1737 if (deadline <= now) /* actually expired already */
1738 timeout = 1; /* ASAP */
1739 else if (timeout == 0 || timeout > deadline - now)
1740 timeout = deadline - now;
1745 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1748 struct ptlrpc_request *req;
1749 struct l_wait_info lwi;
1753 if (cfs_list_empty(&set->set_requests))
1756 cfs_list_for_each(tmp, &set->set_requests) {
1757 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1758 if (req->rq_phase == RQ_PHASE_NEW)
1759 (void)ptlrpc_send_new_req(req);
1763 timeout = ptlrpc_set_next_timeout(set);
1765 /* wait until all complete, interrupted, or an in-flight
1767 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1770 if (timeout == 0 && !cfs_signal_pending())
1772 * No requests are in-flight (ether timed out
1773 * or delayed), so we can allow interrupts.
1774 * We still want to block for a limited time,
1775 * so we allow interrupts during the timeout.
1777 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
1779 ptlrpc_interrupted_set, set);
1782 * At least one request is in flight, so no
1783 * interrupts are allowed. Wait until all
1784 * complete, or an in-flight req times out.
1786 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
1787 ptlrpc_expired_set, set);
1789 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
1791 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1793 /* -EINTR => all requests have been flagged rq_intr so next
1795 * -ETIMEDOUT => someone timed out. When all reqs have
1796 * timed out, signals are enabled allowing completion with
1798 * I don't really care if we go once more round the loop in
1799 * the error cases -eeb. */
1800 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
1801 cfs_list_for_each(tmp, &set->set_requests) {
1802 req = cfs_list_entry(tmp, struct ptlrpc_request,
1804 cfs_spin_lock(&req->rq_lock);
1805 req->rq_invalid_rqset = 1;
1806 cfs_spin_unlock(&req->rq_lock);
1809 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
1811 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
1814 cfs_list_for_each(tmp, &set->set_requests) {
1815 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1817 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1818 if (req->rq_status != 0)
1819 rc = req->rq_status;
1822 if (set->set_interpret != NULL) {
1823 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1825 rc = interpreter (set, set->set_arg, rc);
1827 struct ptlrpc_set_cbdata *cbdata, *n;
1830 cfs_list_for_each_entry_safe(cbdata, n,
1831 &set->set_cblist, psc_item) {
1832 cfs_list_del_init(&cbdata->psc_item);
1833 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1836 OBD_FREE_PTR(cbdata);
1843 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1846 if (request == NULL) {
1851 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1852 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1853 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
1854 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
1855 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
1856 LASSERTF(!request->rq_replay, "req %p\n", request);
1857 LASSERT(request->rq_cli_ctx || request->rq_fake);
1859 req_capsule_fini(&request->rq_pill);
1861 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1862 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1863 if (request->rq_import != NULL) {
1865 cfs_spin_lock(&request->rq_import->imp_lock);
1866 cfs_list_del_init(&request->rq_replay_list);
1868 cfs_spin_unlock(&request->rq_import->imp_lock);
1870 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
1872 if (cfs_atomic_read(&request->rq_refcount) != 0) {
1873 DEBUG_REQ(D_ERROR, request,
1874 "freeing request with nonzero refcount");
1878 if (request->rq_repbuf != NULL)
1879 sptlrpc_cli_free_repbuf(request);
1880 if (request->rq_export != NULL) {
1881 class_export_put(request->rq_export);
1882 request->rq_export = NULL;
1884 if (request->rq_import != NULL) {
1885 class_import_put(request->rq_import);
1886 request->rq_import = NULL;
1888 if (request->rq_bulk != NULL)
1889 ptlrpc_free_bulk(request->rq_bulk);
1891 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1892 sptlrpc_cli_free_reqbuf(request);
1894 if (request->rq_cli_ctx)
1895 sptlrpc_req_put_ctx(request, !locked);
1897 if (request->rq_pool)
1898 __ptlrpc_free_req_to_pool(request);
1900 OBD_FREE(request, sizeof(*request));
1904 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1905 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1907 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1908 (void)__ptlrpc_req_finished(request, 1);
1911 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1914 if (request == NULL)
1917 if (request == LP_POISON ||
1918 request->rq_reqmsg == LP_POISON) {
1919 CERROR("dereferencing freed request (bug 575)\n");
1924 DEBUG_REQ(D_INFO, request, "refcount now %u",
1925 cfs_atomic_read(&request->rq_refcount) - 1);
1927 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
1928 __ptlrpc_free_req(request, locked);
1935 void ptlrpc_req_finished(struct ptlrpc_request *request)
1937 __ptlrpc_req_finished(request, 0);
1940 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1942 return request->rq_xid;
1944 EXPORT_SYMBOL(ptlrpc_req_xid);
1946 /* Disengage the client's reply buffer from the network
1947 * NB does _NOT_ unregister any client-side bulk.
1948 * IDEMPOTENT, but _not_ safe against concurrent callers.
1949 * The request owner (i.e. the thread doing the I/O) must call...
1951 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1955 struct l_wait_info lwi;
1960 LASSERT(!cfs_in_interrupt());
1963 * Let's setup deadline for reply unlink.
1965 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1966 async && request->rq_reply_deadline == 0)
1967 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1970 * Nothing left to do.
1972 if (!ptlrpc_client_recv_or_unlink(request))
1975 LNetMDUnlink(request->rq_reply_md_h);
1978 * Let's check it once again.
1980 if (!ptlrpc_client_recv_or_unlink(request))
1984 * Move to "Unregistering" phase as reply was not unlinked yet.
1986 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1989 * Do not wait for unlink to finish.
1995 * We have to l_wait_event() whatever the result, to give liblustre
1996 * a chance to run reply_in_callback(), and to make sure we've
1997 * unlinked before returning a req to the pool.
1999 if (request->rq_set != NULL)
2000 wq = &request->rq_set->set_waitq;
2002 wq = &request->rq_reply_waitq;
2005 /* Network access will complete in finite time but the HUGE
2006 * timeout lets us CWARN for visibility of sluggish NALs */
2007 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2008 cfs_time_seconds(1), NULL, NULL);
2009 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2012 ptlrpc_rqphase_move(request, request->rq_next_phase);
2016 LASSERT(rc == -ETIMEDOUT);
2017 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2018 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2019 request->rq_must_unlink);
2024 /* caller must hold imp->imp_lock */
2025 void ptlrpc_free_committed(struct obd_import *imp)
2027 cfs_list_t *tmp, *saved;
2028 struct ptlrpc_request *req;
2029 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2032 LASSERT(imp != NULL);
2034 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2037 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2038 imp->imp_generation == imp->imp_last_generation_checked) {
2039 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
2040 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2044 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2045 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2046 imp->imp_generation);
2047 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2048 imp->imp_last_generation_checked = imp->imp_generation;
2050 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2051 req = cfs_list_entry(tmp, struct ptlrpc_request,
2054 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2055 LASSERT(req != last_req);
2058 if (req->rq_transno == 0) {
2059 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2062 if (req->rq_import_generation < imp->imp_generation) {
2063 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2067 if (req->rq_replay) {
2068 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2072 /* not yet committed */
2073 if (req->rq_transno > imp->imp_peer_committed_transno) {
2074 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2078 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
2079 imp->imp_peer_committed_transno);
2081 cfs_spin_lock(&req->rq_lock);
2083 cfs_spin_unlock(&req->rq_lock);
2084 if (req->rq_commit_cb != NULL)
2085 req->rq_commit_cb(req);
2086 cfs_list_del_init(&req->rq_replay_list);
2087 __ptlrpc_req_finished(req, 1);
2094 void ptlrpc_cleanup_client(struct obd_import *imp)
2101 void ptlrpc_resend_req(struct ptlrpc_request *req)
2103 DEBUG_REQ(D_HA, req, "going to resend");
2104 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2105 req->rq_status = -EAGAIN;
2107 cfs_spin_lock(&req->rq_lock);
2109 req->rq_net_err = 0;
2110 req->rq_timedout = 0;
2112 __u64 old_xid = req->rq_xid;
2114 /* ensure previous bulk fails */
2115 req->rq_xid = ptlrpc_next_xid();
2116 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2117 old_xid, req->rq_xid);
2119 ptlrpc_client_wake_req(req);
2120 cfs_spin_unlock(&req->rq_lock);
2123 /* XXX: this function and rq_status are currently unused */
2124 void ptlrpc_restart_req(struct ptlrpc_request *req)
2126 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2127 req->rq_status = -ERESTARTSYS;
2129 cfs_spin_lock(&req->rq_lock);
2130 req->rq_restart = 1;
2131 req->rq_timedout = 0;
2132 ptlrpc_client_wake_req(req);
2133 cfs_spin_unlock(&req->rq_lock);
2136 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2139 cfs_atomic_inc(&req->rq_refcount);
2143 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2144 struct obd_import *imp)
2148 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2150 if (req->rq_transno == 0) {
2151 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2155 /* clear this for new requests that were resent as well
2156 as resent replayed requests. */
2157 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2159 /* don't re-add requests that have been replayed */
2160 if (!cfs_list_empty(&req->rq_replay_list))
2163 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2165 LASSERT(imp->imp_replayable);
2166 /* Balanced in ptlrpc_free_committed, usually. */
2167 ptlrpc_request_addref(req);
2168 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2169 struct ptlrpc_request *iter =
2170 cfs_list_entry(tmp, struct ptlrpc_request,
2173 /* We may have duplicate transnos if we create and then
2174 * open a file, or for closes retained if to match creating
2175 * opens, so use req->rq_xid as a secondary key.
2176 * (See bugs 684, 685, and 428.)
2177 * XXX no longer needed, but all opens need transnos!
2179 if (iter->rq_transno > req->rq_transno)
2182 if (iter->rq_transno == req->rq_transno) {
2183 LASSERT(iter->rq_xid != req->rq_xid);
2184 if (iter->rq_xid > req->rq_xid)
2188 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2192 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2195 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2197 struct ptlrpc_request_set *set;
2201 LASSERT(req->rq_set == NULL);
2202 LASSERT(!req->rq_receiving_reply);
2204 set = ptlrpc_prep_set();
2206 CERROR("Unable to allocate ptlrpc set.");
2210 /* for distributed debugging */
2211 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2213 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2214 ptlrpc_request_addref(req);
2215 ptlrpc_set_add_req(set, req);
2216 rc = ptlrpc_set_wait(set);
2217 ptlrpc_set_destroy(set);
2222 struct ptlrpc_replay_async_args {
2224 int praa_old_status;
2227 static int ptlrpc_replay_interpret(const struct lu_env *env,
2228 struct ptlrpc_request *req,
2229 void * data, int rc)
2231 struct ptlrpc_replay_async_args *aa = data;
2232 struct obd_import *imp = req->rq_import;
2235 cfs_atomic_dec(&imp->imp_replay_inflight);
2237 if (!ptlrpc_client_replied(req)) {
2238 CERROR("request replay timed out, restarting recovery\n");
2239 GOTO(out, rc = -ETIMEDOUT);
2242 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2243 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2244 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2245 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2247 /** VBR: check version failure */
2248 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2249 /** replay was failed due to version mismatch */
2250 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2251 cfs_spin_lock(&imp->imp_lock);
2252 imp->imp_vbr_failed = 1;
2253 imp->imp_no_lock_replay = 1;
2254 cfs_spin_unlock(&imp->imp_lock);
2256 /** The transno had better not change over replay. */
2257 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2258 lustre_msg_get_transno(req->rq_repmsg) ||
2259 lustre_msg_get_transno(req->rq_repmsg) == 0,
2261 lustre_msg_get_transno(req->rq_reqmsg),
2262 lustre_msg_get_transno(req->rq_repmsg));
2265 cfs_spin_lock(&imp->imp_lock);
2266 /** if replays by version then gap was occur on server, no trust to locks */
2267 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2268 imp->imp_no_lock_replay = 1;
2269 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2270 cfs_spin_unlock(&imp->imp_lock);
2271 LASSERT(imp->imp_last_replay_transno);
2273 DEBUG_REQ(D_HA, req, "got rep");
2275 /* let the callback do fixups, possibly including in the request */
2276 if (req->rq_replay_cb)
2277 req->rq_replay_cb(req);
2279 if (ptlrpc_client_replied(req) &&
2280 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2281 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2282 lustre_msg_get_status(req->rq_repmsg),
2283 aa->praa_old_status);
2285 /* Put it back for re-replay. */
2286 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2290 * Errors while replay can set transno to 0, but
2291 * imp_last_replay_transno shouldn't be set to 0 anyway
2293 if (req->rq_transno > 0) {
2294 cfs_spin_lock(&imp->imp_lock);
2295 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2296 imp->imp_last_replay_transno = req->rq_transno;
2297 cfs_spin_unlock(&imp->imp_lock);
2299 CERROR("Transno is 0 during replay!\n");
2300 /* continue with recovery */
2301 rc = ptlrpc_import_recovery_state_machine(imp);
2303 req->rq_send_state = aa->praa_old_state;
2306 /* this replay failed, so restart recovery */
2307 ptlrpc_connect_import(imp, NULL);
2312 int ptlrpc_replay_req(struct ptlrpc_request *req)
2314 struct ptlrpc_replay_async_args *aa;
2317 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2318 /* Not handling automatic bulk replay yet (or ever?) */
2319 LASSERT(req->rq_bulk == NULL);
2321 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2322 aa = ptlrpc_req_async_args(req);
2323 memset(aa, 0, sizeof *aa);
2325 /* Prepare request to be resent with ptlrpcd */
2326 aa->praa_old_state = req->rq_send_state;
2327 req->rq_send_state = LUSTRE_IMP_REPLAY;
2328 req->rq_phase = RQ_PHASE_NEW;
2329 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2331 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2333 req->rq_interpret_reply = ptlrpc_replay_interpret;
2334 /* Readjust the timeout for current conditions */
2335 ptlrpc_at_set_req_timeout(req);
2337 DEBUG_REQ(D_HA, req, "REPLAY");
2339 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2340 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2342 ptlrpcd_add_req(req, PSCOPE_OTHER);
2346 void ptlrpc_abort_inflight(struct obd_import *imp)
2348 cfs_list_t *tmp, *n;
2351 /* Make sure that no new requests get processed for this import.
2352 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2353 * this flag and then putting requests on sending_list or delayed_list.
2355 cfs_spin_lock(&imp->imp_lock);
2357 /* XXX locking? Maybe we should remove each request with the list
2358 * locked? Also, how do we know if the requests on the list are
2359 * being freed at this time?
2361 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2362 struct ptlrpc_request *req =
2363 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2365 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2367 cfs_spin_lock (&req->rq_lock);
2368 if (req->rq_import_generation < imp->imp_generation) {
2370 req->rq_status = -EINTR;
2371 ptlrpc_client_wake_req(req);
2373 cfs_spin_unlock (&req->rq_lock);
2376 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2377 struct ptlrpc_request *req =
2378 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2380 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2382 cfs_spin_lock (&req->rq_lock);
2383 if (req->rq_import_generation < imp->imp_generation) {
2385 req->rq_status = -EINTR;
2386 ptlrpc_client_wake_req(req);
2388 cfs_spin_unlock (&req->rq_lock);
2391 /* Last chance to free reqs left on the replay list, but we
2392 * will still leak reqs that haven't committed. */
2393 if (imp->imp_replayable)
2394 ptlrpc_free_committed(imp);
2396 cfs_spin_unlock(&imp->imp_lock);
2401 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2403 cfs_list_t *tmp, *pos;
2405 LASSERT(set != NULL);
2407 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2408 struct ptlrpc_request *req =
2409 cfs_list_entry(pos, struct ptlrpc_request,
2412 cfs_spin_lock(&req->rq_lock);
2413 if (req->rq_phase != RQ_PHASE_RPC) {
2414 cfs_spin_unlock(&req->rq_lock);
2419 req->rq_status = -EINTR;
2420 ptlrpc_client_wake_req(req);
2421 cfs_spin_unlock(&req->rq_lock);
2425 static __u64 ptlrpc_last_xid;
2426 static cfs_spinlock_t ptlrpc_last_xid_lock;
2428 /* Initialize the XID for the node. This is common among all requests on
2429 * this node, and only requires the property that it is monotonically
2430 * increasing. It does not need to be sequential. Since this is also used
2431 * as the RDMA match bits, it is important that a single client NOT have
2432 * the same match bits for two different in-flight requests, hence we do
2433 * NOT want to have an XID per target or similar.
2435 * To avoid an unlikely collision between match bits after a client reboot
2436 * (which would deliver old data into the wrong RDMA buffer) initialize
2437 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2438 * If the time is clearly incorrect, we instead use a 62-bit random number.
2439 * In the worst case the random number will overflow 1M RPCs per second in
2440 * 9133 years, or permutations thereof.
2442 #define YEAR_2004 (1ULL << 30)
2443 void ptlrpc_init_xid(void)
2445 time_t now = cfs_time_current_sec();
2447 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2448 if (now < YEAR_2004) {
2449 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2450 ptlrpc_last_xid >>= 2;
2451 ptlrpc_last_xid |= (1ULL << 61);
2453 ptlrpc_last_xid = (__u64)now << 20;
2457 __u64 ptlrpc_next_xid(void)
2460 cfs_spin_lock(&ptlrpc_last_xid_lock);
2461 tmp = ++ptlrpc_last_xid;
2462 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2466 __u64 ptlrpc_sample_next_xid(void)
2468 #if BITS_PER_LONG == 32
2469 /* need to avoid possible word tearing on 32-bit systems */
2471 cfs_spin_lock(&ptlrpc_last_xid_lock);
2472 tmp = ptlrpc_last_xid + 1;
2473 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2476 /* No need to lock, since returned value is racy anyways */
2477 return ptlrpc_last_xid + 1;
2480 EXPORT_SYMBOL(ptlrpc_sample_next_xid);