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);
798 cfs_spin_lock(&req->rq_lock);
800 req->rq_invalid_rqset = 0;
801 cfs_spin_unlock(&req->rq_lock);
803 ptlrpc_req_finished (req);
806 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
808 OBD_FREE(set, sizeof(*set));
812 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
813 set_interpreter_func fn, void *data)
815 struct ptlrpc_set_cbdata *cbdata;
817 OBD_ALLOC_PTR(cbdata);
821 cbdata->psc_interpret = fn;
822 cbdata->psc_data = data;
823 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
828 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
829 struct ptlrpc_request *req)
831 /* The set takes over the caller's request reference */
832 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
834 cfs_atomic_inc(&set->set_remaining);
835 req->rq_queued_time = cfs_time_current(); /* Where is the best place to set this? */
839 * Lock so many callers can add things, the context that owns the set
840 * is supposed to notice these and move them into the set proper.
842 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
843 struct ptlrpc_request *req)
845 struct ptlrpc_request_set *set = pc->pc_set;
848 * Let caller know that we stopped and will not handle this request.
849 * It needs to take care itself of request.
851 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags))
854 cfs_spin_lock(&set->set_new_req_lock);
856 * The set takes over the caller's request reference.
858 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
860 cfs_spin_unlock(&set->set_new_req_lock);
862 cfs_waitq_signal(&set->set_waitq);
867 * Based on the current state of the import, determine if the request
868 * can be sent, is an error, or should be delayed.
870 * Returns true if this request should be delayed. If false, and
871 * *status is set, then the request can not be sent and *status is the
872 * error code. If false and status is 0, then request can be sent.
874 * The imp->imp_lock must be held.
876 static int ptlrpc_import_delay_req(struct obd_import *imp,
877 struct ptlrpc_request *req, int *status)
882 LASSERT (status != NULL);
885 if (req->rq_ctx_init || req->rq_ctx_fini) {
886 /* always allow ctx init/fini rpc go through */
887 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
888 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
891 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
892 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
894 } else if (imp->imp_obd->obd_no_recov) {
895 *status = -ESHUTDOWN;
896 } else if (ptlrpc_send_limit_expired(req)) {
897 /* probably doesn't need to be a D_ERROR after initial testing */
898 DEBUG_REQ(D_ERROR, req, "send limit expired ");
900 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
901 imp->imp_state == LUSTRE_IMP_CONNECTING) {
902 /* allow CONNECT even if import is invalid */ ;
903 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
904 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
907 } else if (imp->imp_invalid) {
908 if (!imp->imp_deactive)
909 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
910 *status = -ESHUTDOWN; /* bz 12940 */
911 } else if (req->rq_import_generation != imp->imp_generation) {
912 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
914 } else if (req->rq_send_state != imp->imp_state) {
915 /* invalidate in progress - any requests should be drop */
916 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
917 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
919 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
920 *status = -EWOULDBLOCK;
929 /* Conditionally suppress specific console messages */
930 static int ptlrpc_console_allow(struct ptlrpc_request *req)
932 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
935 /* Suppress particular reconnect errors which are to be expected. No
936 * errors are suppressed for the initial connection on an import */
937 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
938 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
940 /* Suppress timed out reconnect requests */
941 if (req->rq_timedout)
944 /* Suppress unavailable/again reconnect requests */
945 err = lustre_msg_get_status(req->rq_repmsg);
946 if (err == -ENODEV || err == -EAGAIN)
953 static int ptlrpc_check_status(struct ptlrpc_request *req)
958 err = lustre_msg_get_status(req->rq_repmsg);
959 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
960 struct obd_import *imp = req->rq_import;
961 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
962 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
963 " with %s. The %s operation failed with %d\n",
964 libcfs_nid2str(imp->imp_connection->c_peer.nid),
965 ll_opcode2str(opc), err);
966 RETURN(err < 0 ? err : -EINVAL);
970 DEBUG_REQ(D_INFO, req, "status is %d", err);
971 } else if (err > 0) {
972 /* XXX: translate this error from net to host */
973 DEBUG_REQ(D_INFO, req, "status is %d", err);
976 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
977 struct obd_import *imp = req->rq_import;
978 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
980 if (ptlrpc_console_allow(req))
981 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
982 "communicating with %s. The %s "
983 "operation failed with %d\n",
985 imp->imp_connection->c_peer.nid),
986 ll_opcode2str(opc), err);
988 RETURN(err < 0 ? err : -EINVAL);
995 * save pre-versions for replay
997 static void ptlrpc_save_versions(struct ptlrpc_request *req)
999 struct lustre_msg *repmsg = req->rq_repmsg;
1000 struct lustre_msg *reqmsg = req->rq_reqmsg;
1001 __u64 *versions = lustre_msg_get_versions(repmsg);
1004 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1008 lustre_msg_set_versions(reqmsg, versions);
1009 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1010 versions[0], versions[1]);
1016 * Callback function called when client receives RPC reply for \a req.
1018 static int after_reply(struct ptlrpc_request *req)
1020 struct obd_import *imp = req->rq_import;
1021 struct obd_device *obd = req->rq_import->imp_obd;
1023 struct timeval work_start;
1027 LASSERT(obd != NULL);
1028 /* repbuf must be unlinked */
1029 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1031 if (req->rq_reply_truncate) {
1032 if (ptlrpc_no_resend(req)) {
1033 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1034 " expected: %d, actual size: %d",
1035 req->rq_nob_received, req->rq_repbuf_len);
1039 sptlrpc_cli_free_repbuf(req);
1040 /* Pass the required reply buffer size (include
1041 * space for early reply).
1042 * NB: no need to roundup because alloc_repbuf
1043 * will roundup it */
1044 req->rq_replen = req->rq_nob_received;
1045 req->rq_nob_received = 0;
1051 * NB Until this point, the whole of the incoming message,
1052 * including buflens, status etc is in the sender's byte order.
1054 rc = sptlrpc_cli_unwrap_reply(req);
1056 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1061 * Security layer unwrap might ask resend this request.
1066 rc = unpack_reply(req);
1070 cfs_gettimeofday(&work_start);
1071 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1072 if (obd->obd_svc_stats != NULL) {
1073 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1075 ptlrpc_lprocfs_rpc_sent(req, timediff);
1078 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1079 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1080 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1081 lustre_msg_get_type(req->rq_repmsg));
1085 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1086 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1087 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1088 ptlrpc_at_adj_net_latency(req,
1089 lustre_msg_get_service_time(req->rq_repmsg));
1091 rc = ptlrpc_check_status(req);
1092 imp->imp_connect_error = rc;
1096 * Either we've been evicted, or the server has failed for
1097 * some reason. Try to reconnect, and if that fails, punt to
1100 if (ll_rpc_recoverable_error(rc)) {
1101 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1102 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1105 ptlrpc_request_handle_notconn(req);
1110 * Let's look if server sent slv. Do it only for RPC with
1113 ldlm_cli_update_pool(req);
1117 * Store transno in reqmsg for replay.
1119 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1120 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1121 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1124 if (imp->imp_replayable) {
1125 cfs_spin_lock(&imp->imp_lock);
1127 * No point in adding already-committed requests to the replay
1128 * list, we will just remove them immediately. b=9829
1130 if (req->rq_transno != 0 &&
1132 lustre_msg_get_last_committed(req->rq_repmsg) ||
1134 /** version recovery */
1135 ptlrpc_save_versions(req);
1136 ptlrpc_retain_replayable_request(req, imp);
1137 } else if (req->rq_commit_cb != NULL) {
1138 cfs_spin_unlock(&imp->imp_lock);
1139 req->rq_commit_cb(req);
1140 cfs_spin_lock(&imp->imp_lock);
1144 * Replay-enabled imports return commit-status information.
1146 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1147 imp->imp_peer_committed_transno =
1148 lustre_msg_get_last_committed(req->rq_repmsg);
1150 ptlrpc_free_committed(imp);
1151 cfs_spin_unlock(&imp->imp_lock);
1157 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1159 struct obd_import *imp;
1163 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1164 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1167 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1169 imp = req->rq_import;
1170 cfs_spin_lock(&imp->imp_lock);
1172 req->rq_import_generation = imp->imp_generation;
1174 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1175 cfs_spin_lock(&req->rq_lock);
1176 req->rq_waiting = 1;
1177 cfs_spin_unlock(&req->rq_lock);
1179 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1180 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1181 ptlrpc_import_state_name(req->rq_send_state),
1182 ptlrpc_import_state_name(imp->imp_state));
1183 LASSERT(cfs_list_empty(&req->rq_list));
1184 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1185 cfs_atomic_inc(&req->rq_import->imp_inflight);
1186 cfs_spin_unlock(&imp->imp_lock);
1191 cfs_spin_unlock(&imp->imp_lock);
1192 req->rq_status = rc;
1193 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1197 LASSERT(cfs_list_empty(&req->rq_list));
1198 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1199 cfs_atomic_inc(&req->rq_import->imp_inflight);
1200 cfs_spin_unlock(&imp->imp_lock);
1202 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1204 rc = sptlrpc_req_refresh_ctx(req, -1);
1207 req->rq_status = rc;
1210 req->rq_wait_ctx = 1;
1215 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1216 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1217 imp->imp_obd->obd_uuid.uuid,
1218 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1219 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1220 lustre_msg_get_opc(req->rq_reqmsg));
1222 rc = ptl_send_rpc(req, 0);
1224 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1225 req->rq_net_err = 1;
1231 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1232 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1235 int force_timer_recalc = 0;
1238 if (cfs_atomic_read(&set->set_remaining) == 0)
1241 cfs_list_for_each(tmp, &set->set_requests) {
1242 struct ptlrpc_request *req =
1243 cfs_list_entry(tmp, struct ptlrpc_request,
1245 struct obd_import *imp = req->rq_import;
1246 int unregistered = 0;
1249 if (req->rq_phase == RQ_PHASE_NEW &&
1250 ptlrpc_send_new_req(req)) {
1251 force_timer_recalc = 1;
1254 /* delayed send - skip */
1255 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1258 if (!(req->rq_phase == RQ_PHASE_RPC ||
1259 req->rq_phase == RQ_PHASE_BULK ||
1260 req->rq_phase == RQ_PHASE_INTERPRET ||
1261 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1262 req->rq_phase == RQ_PHASE_COMPLETE)) {
1263 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1267 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1268 LASSERT(req->rq_next_phase != req->rq_phase);
1269 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1272 * Skip processing until reply is unlinked. We
1273 * can't return to pool before that and we can't
1274 * call interpret before that. We need to make
1275 * sure that all rdma transfers finished and will
1276 * not corrupt any data.
1278 if (ptlrpc_client_recv_or_unlink(req) ||
1279 ptlrpc_client_bulk_active(req))
1283 * Turn fail_loc off to prevent it from looping
1286 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1287 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1290 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1291 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1296 * Move to next phase if reply was successfully
1299 ptlrpc_rqphase_move(req, req->rq_next_phase);
1302 if (req->rq_phase == RQ_PHASE_COMPLETE)
1305 if (req->rq_phase == RQ_PHASE_INTERPRET)
1306 GOTO(interpret, req->rq_status);
1309 * Note that this also will start async reply unlink.
1311 if (req->rq_net_err && !req->rq_timedout) {
1312 ptlrpc_expire_one_request(req, 1);
1315 * Check if we still need to wait for unlink.
1317 if (ptlrpc_client_recv_or_unlink(req) ||
1318 ptlrpc_client_bulk_active(req))
1323 cfs_spin_lock(&req->rq_lock);
1324 req->rq_replied = 0;
1325 cfs_spin_unlock(&req->rq_lock);
1326 if (req->rq_status == 0)
1327 req->rq_status = -EIO;
1328 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1329 GOTO(interpret, req->rq_status);
1332 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1333 * so it sets rq_intr regardless of individual rpc
1334 * timeouts. The synchronous IO waiting path sets
1335 * rq_intr irrespective of whether ptlrpcd
1336 * has seen a timeout. Our policy is to only interpret
1337 * interrupted rpcs after they have timed out, so we
1338 * need to enforce that here.
1341 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1342 req->rq_wait_ctx)) {
1343 req->rq_status = -EINTR;
1344 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1345 GOTO(interpret, req->rq_status);
1348 if (req->rq_phase == RQ_PHASE_RPC) {
1349 if (req->rq_timedout || req->rq_resend ||
1350 req->rq_waiting || req->rq_wait_ctx) {
1353 if (!ptlrpc_unregister_reply(req, 1))
1356 cfs_spin_lock(&imp->imp_lock);
1357 if (ptlrpc_import_delay_req(imp, req, &status)){
1358 /* put on delay list - only if we wait
1359 * recovery finished - before send */
1360 cfs_list_del_init(&req->rq_list);
1361 cfs_list_add_tail(&req->rq_list,
1364 cfs_spin_unlock(&imp->imp_lock);
1369 req->rq_status = status;
1370 ptlrpc_rqphase_move(req,
1371 RQ_PHASE_INTERPRET);
1372 cfs_spin_unlock(&imp->imp_lock);
1373 GOTO(interpret, req->rq_status);
1375 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1376 req->rq_status = -ENOTCONN;
1377 ptlrpc_rqphase_move(req,
1378 RQ_PHASE_INTERPRET);
1379 cfs_spin_unlock(&imp->imp_lock);
1380 GOTO(interpret, req->rq_status);
1383 cfs_list_del_init(&req->rq_list);
1384 cfs_list_add_tail(&req->rq_list,
1385 &imp->imp_sending_list);
1387 cfs_spin_unlock(&imp->imp_lock);
1389 cfs_spin_lock(&req->rq_lock);
1390 req->rq_waiting = 0;
1391 cfs_spin_unlock(&req->rq_lock);
1393 if (req->rq_timedout || req->rq_resend) {
1394 /* This is re-sending anyways,
1395 * let's mark req as resend. */
1396 cfs_spin_lock(&req->rq_lock);
1398 cfs_spin_unlock(&req->rq_lock);
1402 if (!ptlrpc_unregister_bulk(req, 1))
1405 /* ensure previous bulk fails */
1406 old_xid = req->rq_xid;
1407 req->rq_xid = ptlrpc_next_xid();
1408 CDEBUG(D_HA, "resend bulk "
1411 old_xid, req->rq_xid);
1415 * rq_wait_ctx is only touched by ptlrpcd,
1416 * so no lock is needed here.
1418 status = sptlrpc_req_refresh_ctx(req, -1);
1421 req->rq_status = status;
1422 cfs_spin_lock(&req->rq_lock);
1423 req->rq_wait_ctx = 0;
1424 cfs_spin_unlock(&req->rq_lock);
1425 force_timer_recalc = 1;
1427 cfs_spin_lock(&req->rq_lock);
1428 req->rq_wait_ctx = 1;
1429 cfs_spin_unlock(&req->rq_lock);
1434 cfs_spin_lock(&req->rq_lock);
1435 req->rq_wait_ctx = 0;
1436 cfs_spin_unlock(&req->rq_lock);
1439 rc = ptl_send_rpc(req, 0);
1441 DEBUG_REQ(D_HA, req, "send failed (%d)",
1443 force_timer_recalc = 1;
1444 cfs_spin_lock(&req->rq_lock);
1445 req->rq_net_err = 1;
1446 cfs_spin_unlock(&req->rq_lock);
1448 /* need to reset the timeout */
1449 force_timer_recalc = 1;
1452 cfs_spin_lock(&req->rq_lock);
1454 if (ptlrpc_client_early(req)) {
1455 ptlrpc_at_recv_early_reply(req);
1456 cfs_spin_unlock(&req->rq_lock);
1460 /* Still waiting for a reply? */
1461 if (ptlrpc_client_recv(req)) {
1462 cfs_spin_unlock(&req->rq_lock);
1466 /* Did we actually receive a reply? */
1467 if (!ptlrpc_client_replied(req)) {
1468 cfs_spin_unlock(&req->rq_lock);
1472 cfs_spin_unlock(&req->rq_lock);
1474 /* unlink from net because we are going to
1475 * swab in-place of reply buffer */
1476 unregistered = ptlrpc_unregister_reply(req, 1);
1480 req->rq_status = after_reply(req);
1484 /* If there is no bulk associated with this request,
1485 * then we're done and should let the interpreter
1486 * process the reply. Similarly if the RPC returned
1487 * an error, and therefore the bulk will never arrive.
1489 if (req->rq_bulk == NULL || req->rq_status != 0) {
1490 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1491 GOTO(interpret, req->rq_status);
1494 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1497 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1498 if (ptlrpc_client_bulk_active(req))
1501 if (!req->rq_bulk->bd_success) {
1502 /* The RPC reply arrived OK, but the bulk screwed
1503 * up! Dead weird since the server told us the RPC
1504 * was good after getting the REPLY for her GET or
1505 * the ACK for her PUT. */
1506 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1510 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1513 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1515 /* This moves to "unregistering" phase we need to wait for
1517 if (!unregistered && !ptlrpc_unregister_reply(req, 1))
1520 if (!ptlrpc_unregister_bulk(req, 1))
1523 /* When calling interpret receiving already should be
1525 LASSERT(!req->rq_receiving_reply);
1527 ptlrpc_req_interpret(env, req, req->rq_status);
1529 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1531 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1532 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1533 imp->imp_obd->obd_uuid.uuid,
1534 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1536 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1537 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1539 cfs_spin_lock(&imp->imp_lock);
1540 /* Request already may be not on sending or delaying list. This
1541 * may happen in the case of marking it erroneous for the case
1542 * ptlrpc_import_delay_req(req, status) find it impossible to
1543 * allow sending this rpc and returns *status != 0. */
1544 if (!cfs_list_empty(&req->rq_list)) {
1545 cfs_list_del_init(&req->rq_list);
1546 cfs_atomic_dec(&imp->imp_inflight);
1548 cfs_spin_unlock(&imp->imp_lock);
1550 cfs_atomic_dec(&set->set_remaining);
1551 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1554 /* If we hit an error, we want to recover promptly. */
1555 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1558 /* Return 1 if we should give up, else 0 */
1559 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1561 struct obd_import *imp = req->rq_import;
1565 cfs_spin_lock(&req->rq_lock);
1566 req->rq_timedout = 1;
1567 cfs_spin_unlock(&req->rq_lock);
1569 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req,
1570 "Request x"LPU64" sent from %s to NID %s "CFS_DURATION_T"s "
1571 "ago has %s ("CFS_DURATION_T"s prior to deadline).\n",
1572 req->rq_xid, imp ? imp->imp_obd->obd_name : "<?>",
1573 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1574 cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1575 req->rq_net_err ? "failed due to network error" : "timed out",
1576 cfs_time_sub(req->rq_deadline, req->rq_sent));
1578 if (imp != NULL && obd_debug_peer_on_timeout)
1579 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1581 ptlrpc_unregister_reply(req, async_unlink);
1582 ptlrpc_unregister_bulk(req, async_unlink);
1584 if (obd_dump_on_timeout)
1585 libcfs_debug_dumplog();
1588 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1595 cfs_atomic_inc(&imp->imp_timeouts);
1597 /* The DLM server doesn't want recovery run on its imports. */
1598 if (imp->imp_dlm_fake)
1601 /* If this request is for recovery or other primordial tasks,
1602 * then error it out here. */
1603 if (req->rq_ctx_init || req->rq_ctx_fini ||
1604 req->rq_send_state != LUSTRE_IMP_FULL ||
1605 imp->imp_obd->obd_no_recov) {
1606 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1607 ptlrpc_import_state_name(req->rq_send_state),
1608 ptlrpc_import_state_name(imp->imp_state));
1609 cfs_spin_lock(&req->rq_lock);
1610 req->rq_status = -ETIMEDOUT;
1612 cfs_spin_unlock(&req->rq_lock);
1616 /* if a request can't be resent we can't wait for an answer after
1618 if (ptlrpc_no_resend(req)) {
1619 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1623 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1628 int ptlrpc_expired_set(void *data)
1630 struct ptlrpc_request_set *set = data;
1632 time_t now = cfs_time_current_sec();
1635 LASSERT(set != NULL);
1638 * A timeout expired. See which reqs it applies to...
1640 cfs_list_for_each (tmp, &set->set_requests) {
1641 struct ptlrpc_request *req =
1642 cfs_list_entry(tmp, struct ptlrpc_request,
1645 /* don't expire request waiting for context */
1646 if (req->rq_wait_ctx)
1649 /* Request in-flight? */
1650 if (!((req->rq_phase == RQ_PHASE_RPC &&
1651 !req->rq_waiting && !req->rq_resend) ||
1652 (req->rq_phase == RQ_PHASE_BULK)))
1655 if (req->rq_timedout || /* already dealt with */
1656 req->rq_deadline > now) /* not expired */
1659 /* Deal with this guy. Do it asynchronously to not block
1660 * ptlrpcd thread. */
1661 ptlrpc_expire_one_request(req, 1);
1665 * When waiting for a whole set, we always break out of the
1666 * sleep so we can recalculate the timeout, or enable interrupts
1667 * if everyone's timed out.
1672 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1674 cfs_spin_lock(&req->rq_lock);
1676 cfs_spin_unlock(&req->rq_lock);
1679 void ptlrpc_interrupted_set(void *data)
1681 struct ptlrpc_request_set *set = data;
1684 LASSERT(set != NULL);
1685 CERROR("INTERRUPTED SET %p\n", set);
1687 cfs_list_for_each(tmp, &set->set_requests) {
1688 struct ptlrpc_request *req =
1689 cfs_list_entry(tmp, struct ptlrpc_request,
1692 if (req->rq_phase != RQ_PHASE_RPC &&
1693 req->rq_phase != RQ_PHASE_UNREGISTERING)
1696 ptlrpc_mark_interrupted(req);
1701 * Get the smallest timeout in the set; this does NOT set a timeout.
1703 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1706 time_t now = cfs_time_current_sec();
1708 struct ptlrpc_request *req;
1712 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1714 cfs_list_for_each(tmp, &set->set_requests) {
1715 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1718 * Request in-flight?
1720 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1721 (req->rq_phase == RQ_PHASE_BULK) ||
1722 (req->rq_phase == RQ_PHASE_NEW)))
1726 * Already timed out.
1728 if (req->rq_timedout)
1734 if (req->rq_wait_ctx)
1737 if (req->rq_phase == RQ_PHASE_NEW)
1738 deadline = req->rq_sent;
1740 deadline = req->rq_sent + req->rq_timeout;
1742 if (deadline <= now) /* actually expired already */
1743 timeout = 1; /* ASAP */
1744 else if (timeout == 0 || timeout > deadline - now)
1745 timeout = deadline - now;
1750 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1753 struct ptlrpc_request *req;
1754 struct l_wait_info lwi;
1758 if (cfs_list_empty(&set->set_requests))
1761 cfs_list_for_each(tmp, &set->set_requests) {
1762 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1763 if (req->rq_phase == RQ_PHASE_NEW)
1764 (void)ptlrpc_send_new_req(req);
1768 timeout = ptlrpc_set_next_timeout(set);
1770 /* wait until all complete, interrupted, or an in-flight
1772 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1775 if (timeout == 0 && !cfs_signal_pending())
1777 * No requests are in-flight (ether timed out
1778 * or delayed), so we can allow interrupts.
1779 * We still want to block for a limited time,
1780 * so we allow interrupts during the timeout.
1782 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
1784 ptlrpc_interrupted_set, set);
1787 * At least one request is in flight, so no
1788 * interrupts are allowed. Wait until all
1789 * complete, or an in-flight req times out.
1791 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
1792 ptlrpc_expired_set, set);
1794 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
1796 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1798 /* -EINTR => all requests have been flagged rq_intr so next
1800 * -ETIMEDOUT => someone timed out. When all reqs have
1801 * timed out, signals are enabled allowing completion with
1803 * I don't really care if we go once more round the loop in
1804 * the error cases -eeb. */
1805 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
1806 cfs_list_for_each(tmp, &set->set_requests) {
1807 req = cfs_list_entry(tmp, struct ptlrpc_request,
1809 cfs_spin_lock(&req->rq_lock);
1810 req->rq_invalid_rqset = 1;
1811 cfs_spin_unlock(&req->rq_lock);
1814 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
1816 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
1819 cfs_list_for_each(tmp, &set->set_requests) {
1820 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1822 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1823 if (req->rq_status != 0)
1824 rc = req->rq_status;
1827 if (set->set_interpret != NULL) {
1828 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1830 rc = interpreter (set, set->set_arg, rc);
1832 struct ptlrpc_set_cbdata *cbdata, *n;
1835 cfs_list_for_each_entry_safe(cbdata, n,
1836 &set->set_cblist, psc_item) {
1837 cfs_list_del_init(&cbdata->psc_item);
1838 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1841 OBD_FREE_PTR(cbdata);
1848 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1851 if (request == NULL) {
1856 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1857 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1858 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
1859 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
1860 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
1861 LASSERTF(!request->rq_replay, "req %p\n", request);
1862 LASSERT(request->rq_cli_ctx || request->rq_fake);
1864 req_capsule_fini(&request->rq_pill);
1866 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1867 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1868 if (request->rq_import != NULL) {
1870 cfs_spin_lock(&request->rq_import->imp_lock);
1871 cfs_list_del_init(&request->rq_replay_list);
1873 cfs_spin_unlock(&request->rq_import->imp_lock);
1875 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
1877 if (cfs_atomic_read(&request->rq_refcount) != 0) {
1878 DEBUG_REQ(D_ERROR, request,
1879 "freeing request with nonzero refcount");
1883 if (request->rq_repbuf != NULL)
1884 sptlrpc_cli_free_repbuf(request);
1885 if (request->rq_export != NULL) {
1886 class_export_put(request->rq_export);
1887 request->rq_export = NULL;
1889 if (request->rq_import != NULL) {
1890 class_import_put(request->rq_import);
1891 request->rq_import = NULL;
1893 if (request->rq_bulk != NULL)
1894 ptlrpc_free_bulk(request->rq_bulk);
1896 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1897 sptlrpc_cli_free_reqbuf(request);
1899 if (request->rq_cli_ctx)
1900 sptlrpc_req_put_ctx(request, !locked);
1902 if (request->rq_pool)
1903 __ptlrpc_free_req_to_pool(request);
1905 OBD_FREE(request, sizeof(*request));
1909 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1910 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1912 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1913 (void)__ptlrpc_req_finished(request, 1);
1916 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1919 if (request == NULL)
1922 if (request == LP_POISON ||
1923 request->rq_reqmsg == LP_POISON) {
1924 CERROR("dereferencing freed request (bug 575)\n");
1929 DEBUG_REQ(D_INFO, request, "refcount now %u",
1930 cfs_atomic_read(&request->rq_refcount) - 1);
1932 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
1933 __ptlrpc_free_req(request, locked);
1940 void ptlrpc_req_finished(struct ptlrpc_request *request)
1942 __ptlrpc_req_finished(request, 0);
1945 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1947 return request->rq_xid;
1949 EXPORT_SYMBOL(ptlrpc_req_xid);
1951 /* Disengage the client's reply buffer from the network
1952 * NB does _NOT_ unregister any client-side bulk.
1953 * IDEMPOTENT, but _not_ safe against concurrent callers.
1954 * The request owner (i.e. the thread doing the I/O) must call...
1956 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1960 struct l_wait_info lwi;
1965 LASSERT(!cfs_in_interrupt());
1968 * Let's setup deadline for reply unlink.
1970 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1971 async && request->rq_reply_deadline == 0)
1972 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1975 * Nothing left to do.
1977 if (!ptlrpc_client_recv_or_unlink(request))
1980 LNetMDUnlink(request->rq_reply_md_h);
1983 * Let's check it once again.
1985 if (!ptlrpc_client_recv_or_unlink(request))
1989 * Move to "Unregistering" phase as reply was not unlinked yet.
1991 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1994 * Do not wait for unlink to finish.
2000 * We have to l_wait_event() whatever the result, to give liblustre
2001 * a chance to run reply_in_callback(), and to make sure we've
2002 * unlinked before returning a req to the pool.
2004 if (request->rq_set != NULL)
2005 wq = &request->rq_set->set_waitq;
2007 wq = &request->rq_reply_waitq;
2010 /* Network access will complete in finite time but the HUGE
2011 * timeout lets us CWARN for visibility of sluggish NALs */
2012 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2013 cfs_time_seconds(1), NULL, NULL);
2014 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2017 ptlrpc_rqphase_move(request, request->rq_next_phase);
2021 LASSERT(rc == -ETIMEDOUT);
2022 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2023 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2024 request->rq_must_unlink);
2029 /* caller must hold imp->imp_lock */
2030 void ptlrpc_free_committed(struct obd_import *imp)
2032 cfs_list_t *tmp, *saved;
2033 struct ptlrpc_request *req;
2034 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2037 LASSERT(imp != NULL);
2039 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2042 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2043 imp->imp_generation == imp->imp_last_generation_checked) {
2044 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
2045 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2049 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2050 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2051 imp->imp_generation);
2052 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2053 imp->imp_last_generation_checked = imp->imp_generation;
2055 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2056 req = cfs_list_entry(tmp, struct ptlrpc_request,
2059 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2060 LASSERT(req != last_req);
2063 if (req->rq_transno == 0) {
2064 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2067 if (req->rq_import_generation < imp->imp_generation) {
2068 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2072 if (req->rq_replay) {
2073 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2077 /* not yet committed */
2078 if (req->rq_transno > imp->imp_peer_committed_transno) {
2079 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2083 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
2084 imp->imp_peer_committed_transno);
2086 cfs_spin_lock(&req->rq_lock);
2088 cfs_spin_unlock(&req->rq_lock);
2089 if (req->rq_commit_cb != NULL)
2090 req->rq_commit_cb(req);
2091 cfs_list_del_init(&req->rq_replay_list);
2092 __ptlrpc_req_finished(req, 1);
2099 void ptlrpc_cleanup_client(struct obd_import *imp)
2106 void ptlrpc_resend_req(struct ptlrpc_request *req)
2108 DEBUG_REQ(D_HA, req, "going to resend");
2109 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2110 req->rq_status = -EAGAIN;
2112 cfs_spin_lock(&req->rq_lock);
2114 req->rq_net_err = 0;
2115 req->rq_timedout = 0;
2117 __u64 old_xid = req->rq_xid;
2119 /* ensure previous bulk fails */
2120 req->rq_xid = ptlrpc_next_xid();
2121 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2122 old_xid, req->rq_xid);
2124 ptlrpc_client_wake_req(req);
2125 cfs_spin_unlock(&req->rq_lock);
2128 /* XXX: this function and rq_status are currently unused */
2129 void ptlrpc_restart_req(struct ptlrpc_request *req)
2131 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2132 req->rq_status = -ERESTARTSYS;
2134 cfs_spin_lock(&req->rq_lock);
2135 req->rq_restart = 1;
2136 req->rq_timedout = 0;
2137 ptlrpc_client_wake_req(req);
2138 cfs_spin_unlock(&req->rq_lock);
2141 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2144 cfs_atomic_inc(&req->rq_refcount);
2148 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2149 struct obd_import *imp)
2153 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2155 if (req->rq_transno == 0) {
2156 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2160 /* clear this for new requests that were resent as well
2161 as resent replayed requests. */
2162 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2164 /* don't re-add requests that have been replayed */
2165 if (!cfs_list_empty(&req->rq_replay_list))
2168 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2170 LASSERT(imp->imp_replayable);
2171 /* Balanced in ptlrpc_free_committed, usually. */
2172 ptlrpc_request_addref(req);
2173 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2174 struct ptlrpc_request *iter =
2175 cfs_list_entry(tmp, struct ptlrpc_request,
2178 /* We may have duplicate transnos if we create and then
2179 * open a file, or for closes retained if to match creating
2180 * opens, so use req->rq_xid as a secondary key.
2181 * (See bugs 684, 685, and 428.)
2182 * XXX no longer needed, but all opens need transnos!
2184 if (iter->rq_transno > req->rq_transno)
2187 if (iter->rq_transno == req->rq_transno) {
2188 LASSERT(iter->rq_xid != req->rq_xid);
2189 if (iter->rq_xid > req->rq_xid)
2193 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2197 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2200 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2202 struct ptlrpc_request_set *set;
2206 LASSERT(req->rq_set == NULL);
2207 LASSERT(!req->rq_receiving_reply);
2209 set = ptlrpc_prep_set();
2211 CERROR("Unable to allocate ptlrpc set.");
2215 /* for distributed debugging */
2216 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2218 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2219 ptlrpc_request_addref(req);
2220 ptlrpc_set_add_req(set, req);
2221 rc = ptlrpc_set_wait(set);
2222 ptlrpc_set_destroy(set);
2227 struct ptlrpc_replay_async_args {
2229 int praa_old_status;
2232 static int ptlrpc_replay_interpret(const struct lu_env *env,
2233 struct ptlrpc_request *req,
2234 void * data, int rc)
2236 struct ptlrpc_replay_async_args *aa = data;
2237 struct obd_import *imp = req->rq_import;
2240 cfs_atomic_dec(&imp->imp_replay_inflight);
2242 if (!ptlrpc_client_replied(req)) {
2243 CERROR("request replay timed out, restarting recovery\n");
2244 GOTO(out, rc = -ETIMEDOUT);
2247 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2248 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2249 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2250 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2252 /** VBR: check version failure */
2253 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2254 /** replay was failed due to version mismatch */
2255 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2256 cfs_spin_lock(&imp->imp_lock);
2257 imp->imp_vbr_failed = 1;
2258 imp->imp_no_lock_replay = 1;
2259 cfs_spin_unlock(&imp->imp_lock);
2261 /** The transno had better not change over replay. */
2262 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2263 lustre_msg_get_transno(req->rq_repmsg) ||
2264 lustre_msg_get_transno(req->rq_repmsg) == 0,
2266 lustre_msg_get_transno(req->rq_reqmsg),
2267 lustre_msg_get_transno(req->rq_repmsg));
2270 cfs_spin_lock(&imp->imp_lock);
2271 /** if replays by version then gap was occur on server, no trust to locks */
2272 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2273 imp->imp_no_lock_replay = 1;
2274 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2275 cfs_spin_unlock(&imp->imp_lock);
2276 LASSERT(imp->imp_last_replay_transno);
2278 DEBUG_REQ(D_HA, req, "got rep");
2280 /* let the callback do fixups, possibly including in the request */
2281 if (req->rq_replay_cb)
2282 req->rq_replay_cb(req);
2284 if (ptlrpc_client_replied(req) &&
2285 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2286 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2287 lustre_msg_get_status(req->rq_repmsg),
2288 aa->praa_old_status);
2290 /* Put it back for re-replay. */
2291 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2295 * Errors while replay can set transno to 0, but
2296 * imp_last_replay_transno shouldn't be set to 0 anyway
2298 if (req->rq_transno > 0) {
2299 cfs_spin_lock(&imp->imp_lock);
2300 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2301 imp->imp_last_replay_transno = req->rq_transno;
2302 cfs_spin_unlock(&imp->imp_lock);
2304 CERROR("Transno is 0 during replay!\n");
2305 /* continue with recovery */
2306 rc = ptlrpc_import_recovery_state_machine(imp);
2308 req->rq_send_state = aa->praa_old_state;
2311 /* this replay failed, so restart recovery */
2312 ptlrpc_connect_import(imp, NULL);
2317 int ptlrpc_replay_req(struct ptlrpc_request *req)
2319 struct ptlrpc_replay_async_args *aa;
2322 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2323 /* Not handling automatic bulk replay yet (or ever?) */
2324 LASSERT(req->rq_bulk == NULL);
2326 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2327 aa = ptlrpc_req_async_args(req);
2328 memset(aa, 0, sizeof *aa);
2330 /* Prepare request to be resent with ptlrpcd */
2331 aa->praa_old_state = req->rq_send_state;
2332 req->rq_send_state = LUSTRE_IMP_REPLAY;
2333 req->rq_phase = RQ_PHASE_NEW;
2334 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2336 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2338 req->rq_interpret_reply = ptlrpc_replay_interpret;
2339 /* Readjust the timeout for current conditions */
2340 ptlrpc_at_set_req_timeout(req);
2342 DEBUG_REQ(D_HA, req, "REPLAY");
2344 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2345 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2347 ptlrpcd_add_req(req, PSCOPE_OTHER);
2351 void ptlrpc_abort_inflight(struct obd_import *imp)
2353 cfs_list_t *tmp, *n;
2356 /* Make sure that no new requests get processed for this import.
2357 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2358 * this flag and then putting requests on sending_list or delayed_list.
2360 cfs_spin_lock(&imp->imp_lock);
2362 /* XXX locking? Maybe we should remove each request with the list
2363 * locked? Also, how do we know if the requests on the list are
2364 * being freed at this time?
2366 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2367 struct ptlrpc_request *req =
2368 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2370 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2372 cfs_spin_lock (&req->rq_lock);
2373 if (req->rq_import_generation < imp->imp_generation) {
2375 req->rq_status = -EINTR;
2376 ptlrpc_client_wake_req(req);
2378 cfs_spin_unlock (&req->rq_lock);
2381 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2382 struct ptlrpc_request *req =
2383 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2385 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2387 cfs_spin_lock (&req->rq_lock);
2388 if (req->rq_import_generation < imp->imp_generation) {
2390 req->rq_status = -EINTR;
2391 ptlrpc_client_wake_req(req);
2393 cfs_spin_unlock (&req->rq_lock);
2396 /* Last chance to free reqs left on the replay list, but we
2397 * will still leak reqs that haven't committed. */
2398 if (imp->imp_replayable)
2399 ptlrpc_free_committed(imp);
2401 cfs_spin_unlock(&imp->imp_lock);
2406 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2408 cfs_list_t *tmp, *pos;
2410 LASSERT(set != NULL);
2412 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2413 struct ptlrpc_request *req =
2414 cfs_list_entry(pos, struct ptlrpc_request,
2417 cfs_spin_lock(&req->rq_lock);
2418 if (req->rq_phase != RQ_PHASE_RPC) {
2419 cfs_spin_unlock(&req->rq_lock);
2424 req->rq_status = -EINTR;
2425 ptlrpc_client_wake_req(req);
2426 cfs_spin_unlock(&req->rq_lock);
2430 static __u64 ptlrpc_last_xid;
2431 static cfs_spinlock_t ptlrpc_last_xid_lock;
2433 /* Initialize the XID for the node. This is common among all requests on
2434 * this node, and only requires the property that it is monotonically
2435 * increasing. It does not need to be sequential. Since this is also used
2436 * as the RDMA match bits, it is important that a single client NOT have
2437 * the same match bits for two different in-flight requests, hence we do
2438 * NOT want to have an XID per target or similar.
2440 * To avoid an unlikely collision between match bits after a client reboot
2441 * (which would deliver old data into the wrong RDMA buffer) initialize
2442 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2443 * If the time is clearly incorrect, we instead use a 62-bit random number.
2444 * In the worst case the random number will overflow 1M RPCs per second in
2445 * 9133 years, or permutations thereof.
2447 #define YEAR_2004 (1ULL << 30)
2448 void ptlrpc_init_xid(void)
2450 time_t now = cfs_time_current_sec();
2452 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2453 if (now < YEAR_2004) {
2454 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2455 ptlrpc_last_xid >>= 2;
2456 ptlrpc_last_xid |= (1ULL << 61);
2458 ptlrpc_last_xid = (__u64)now << 20;
2462 __u64 ptlrpc_next_xid(void)
2465 cfs_spin_lock(&ptlrpc_last_xid_lock);
2466 tmp = ++ptlrpc_last_xid;
2467 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2471 __u64 ptlrpc_sample_next_xid(void)
2473 #if BITS_PER_LONG == 32
2474 /* need to avoid possible word tearing on 32-bit systems */
2476 cfs_spin_lock(&ptlrpc_last_xid_lock);
2477 tmp = ptlrpc_last_xid + 1;
2478 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2481 /* No need to lock, since returned value is racy anyways */
2482 return ptlrpc_last_xid + 1;
2485 EXPORT_SYMBOL(ptlrpc_sample_next_xid);