1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
41 #include <liblustre.h>
44 #include <obd_support.h>
45 #include <obd_class.h>
46 #include <lustre_lib.h>
47 #include <lustre_ha.h>
48 #include <lustre_import.h>
49 #include <lustre_req_layout.h>
51 #include "ptlrpc_internal.h"
53 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
54 struct ptlrpc_client *cl)
56 cl->cli_request_portal = req_portal;
57 cl->cli_reply_portal = rep_portal;
61 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
63 struct ptlrpc_connection *c;
65 lnet_process_id_t peer;
68 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
70 CERROR("cannot find peer %s!\n", uuid->uuid);
74 c = ptlrpc_connection_get(peer, self, uuid);
76 memcpy(c->c_remote_uuid.uuid,
77 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
80 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
85 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
87 struct ptlrpc_bulk_desc *desc;
89 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
93 spin_lock_init(&desc->bd_lock);
94 cfs_waitq_init(&desc->bd_waitq);
95 desc->bd_max_iov = npages;
96 desc->bd_iov_count = 0;
97 LNetInvalidateHandle(&desc->bd_md_h);
98 desc->bd_portal = portal;
104 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
105 int npages, int type, int portal)
107 struct obd_import *imp = req->rq_import;
108 struct ptlrpc_bulk_desc *desc;
111 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
112 desc = new_bulk(npages, type, portal);
116 desc->bd_import_generation = req->rq_import_generation;
117 desc->bd_import = class_import_get(imp);
120 desc->bd_cbid.cbid_fn = client_bulk_callback;
121 desc->bd_cbid.cbid_arg = desc;
123 /* This makes req own desc, and free it when she frees herself */
129 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
130 int npages, int type, int portal)
132 struct obd_export *exp = req->rq_export;
133 struct ptlrpc_bulk_desc *desc;
136 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
138 desc = new_bulk(npages, type, portal);
142 desc->bd_export = class_export_get(exp);
145 desc->bd_cbid.cbid_fn = server_bulk_callback;
146 desc->bd_cbid.cbid_arg = desc;
148 /* NB we don't assign rq_bulk here; server-side requests are
149 * re-used, and the handler frees the bulk desc explicitly. */
154 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
155 cfs_page_t *page, int pageoffset, int len)
157 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
158 LASSERT(page != NULL);
159 LASSERT(pageoffset >= 0);
161 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
165 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
168 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
172 LASSERT(desc != NULL);
173 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
174 LASSERT(!desc->bd_network_rw); /* network hands off or */
175 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
177 sptlrpc_enc_pool_put_pages(desc);
180 class_export_put(desc->bd_export);
182 class_import_put(desc->bd_import);
184 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
185 bd_iov[desc->bd_max_iov]));
189 /* Set server timelimit for this req */
190 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
196 LASSERT(req->rq_import);
199 /* non-AT settings */
200 req->rq_timeout = req->rq_import->imp_server_timeout ?
201 obd_timeout / 2 : obd_timeout;
202 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
206 at = &req->rq_import->imp_at;
207 idx = import_at_get_index(req->rq_import,
208 req->rq_request_portal);
209 serv_est = at_get(&at->iat_service_estimate[idx]);
210 req->rq_timeout = at_est2timeout(serv_est);
211 /* We could get even fancier here, using history to predict increased
214 /* Let the server know what this RPC timeout is by putting it in the
216 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
219 /* Adjust max service estimate based on server value */
220 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
221 unsigned int serv_est)
227 /* do estimate only if is not in recovery */
228 if ((req->rq_send_state != LUSTRE_IMP_FULL) &&
229 (req->rq_send_state != LUSTRE_IMP_CONNECTING))
232 LASSERT(req->rq_import);
233 at = &req->rq_import->imp_at;
235 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
236 /* max service estimates are tracked on the server side,
237 so just keep minimal history here */
238 oldse = at_add(&at->iat_service_estimate[idx], serv_est);
240 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
241 "has changed from %d to %d\n",
242 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
243 oldse, at_get(&at->iat_service_estimate[idx]));
246 /* Expected network latency per remote node (secs) */
247 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
249 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
252 /* Adjust expected network latency */
253 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
254 unsigned int service_time)
256 unsigned int nl, oldnl;
258 time_t now = cfs_time_current_sec();
260 LASSERT(req->rq_import);
261 at = &req->rq_import->imp_at;
263 /* Network latency is total time less server processing time */
264 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
265 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
266 CWARN("Reported service time %u > total measured time "
267 CFS_DURATION_T"\n", service_time,
268 cfs_time_sub(now, req->rq_sent));
270 oldnl = at_add(&at->iat_net_latency, nl);
272 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
273 "has changed from %d to %d\n",
274 req->rq_import->imp_obd->obd_name,
276 &req->rq_import->imp_connection->c_remote_uuid),
277 oldnl, at_get(&at->iat_net_latency));
280 static int unpack_reply(struct ptlrpc_request *req)
284 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
285 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
287 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
292 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
294 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
301 * Handle an early reply message, called with the rq_lock held.
302 * If anything goes wrong just ignore it - same as if it never happened
304 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
306 struct ptlrpc_request *early_req;
312 spin_unlock(&req->rq_lock);
314 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
316 spin_lock(&req->rq_lock);
320 rc = unpack_reply(early_req);
322 /* Expecting to increase the service time estimate here */
323 ptlrpc_at_adj_service(req,
324 lustre_msg_get_timeout(early_req->rq_repmsg));
325 ptlrpc_at_adj_net_latency(req,
326 lustre_msg_get_service_time(early_req->rq_repmsg));
329 sptlrpc_cli_finish_early_reply(early_req);
331 spin_lock(&req->rq_lock);
334 /* Adjust the local timeout for this req */
335 ptlrpc_at_set_req_timeout(req);
337 olddl = req->rq_deadline;
338 /* server assumes it now has rq_timeout from when it sent the
339 early reply, so client should give it at least that long. */
340 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
341 ptlrpc_at_get_net_latency(req);
343 DEBUG_REQ(D_ADAPTTO, req,
344 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
345 "("CFS_DURATION_T"s)", req->rq_early_count,
346 cfs_time_sub(req->rq_deadline,
347 cfs_time_current_sec()),
348 cfs_time_sub(req->rq_deadline, olddl));
354 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
356 struct list_head *l, *tmp;
357 struct ptlrpc_request *req;
359 LASSERT(pool != NULL);
361 spin_lock(&pool->prp_lock);
362 list_for_each_safe(l, tmp, &pool->prp_req_list) {
363 req = list_entry(l, struct ptlrpc_request, rq_list);
364 list_del(&req->rq_list);
365 LASSERT(req->rq_reqbuf);
366 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
367 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
368 OBD_FREE(req, sizeof(*req));
370 spin_unlock(&pool->prp_lock);
371 OBD_FREE(pool, sizeof(*pool));
374 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
379 while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
382 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
383 "Trying to change pool size with nonempty pool "
384 "from %d to %d bytes\n", pool->prp_rq_size, size);
386 spin_lock(&pool->prp_lock);
387 pool->prp_rq_size = size;
388 for (i = 0; i < num_rq; i++) {
389 struct ptlrpc_request *req;
390 struct lustre_msg *msg;
392 spin_unlock(&pool->prp_lock);
393 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
396 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
398 OBD_FREE(req, sizeof(struct ptlrpc_request));
401 req->rq_reqbuf = msg;
402 req->rq_reqbuf_len = size;
404 spin_lock(&pool->prp_lock);
405 list_add_tail(&req->rq_list, &pool->prp_req_list);
407 spin_unlock(&pool->prp_lock);
411 struct ptlrpc_request_pool *
412 ptlrpc_init_rq_pool(int num_rq, int msgsize,
413 void (*populate_pool)(struct ptlrpc_request_pool *, int))
415 struct ptlrpc_request_pool *pool;
417 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
421 /* Request next power of two for the allocation, because internally
422 kernel would do exactly this */
424 spin_lock_init(&pool->prp_lock);
425 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
426 pool->prp_rq_size = msgsize;
427 pool->prp_populate = populate_pool;
429 populate_pool(pool, num_rq);
431 if (list_empty(&pool->prp_req_list)) {
432 /* have not allocated a single request for the pool */
433 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
439 static struct ptlrpc_request *
440 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
442 struct ptlrpc_request *request;
443 struct lustre_msg *reqbuf;
448 spin_lock(&pool->prp_lock);
450 /* See if we have anything in a pool, and bail out if nothing,
451 * in writeout path, where this matters, this is safe to do, because
452 * nothing is lost in this case, and when some in-flight requests
453 * complete, this code will be called again. */
454 if (unlikely(list_empty(&pool->prp_req_list))) {
455 spin_unlock(&pool->prp_lock);
459 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
461 list_del_init(&request->rq_list);
462 spin_unlock(&pool->prp_lock);
464 LASSERT(request->rq_reqbuf);
465 LASSERT(request->rq_pool);
467 reqbuf = request->rq_reqbuf;
468 memset(request, 0, sizeof(*request));
469 request->rq_reqbuf = reqbuf;
470 request->rq_reqbuf_len = pool->prp_rq_size;
471 request->rq_pool = pool;
476 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
478 struct ptlrpc_request_pool *pool = request->rq_pool;
480 spin_lock(&pool->prp_lock);
481 LASSERT(list_empty(&request->rq_list));
482 LASSERT(!request->rq_receiving_reply);
483 list_add_tail(&request->rq_list, &pool->prp_req_list);
484 spin_unlock(&pool->prp_lock);
487 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
488 __u32 version, int opcode,
489 int count, __u32 *lengths, char **bufs,
490 struct ptlrpc_cli_ctx *ctx)
492 struct obd_import *imp = request->rq_import;
497 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
499 rc = sptlrpc_req_get_ctx(request);
504 sptlrpc_req_set_flavor(request, opcode);
506 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
509 LASSERT(!request->rq_pool);
513 lustre_msg_add_version(request->rq_reqmsg, version);
514 request->rq_send_state = LUSTRE_IMP_FULL;
515 request->rq_type = PTL_RPC_MSG_REQUEST;
516 request->rq_export = NULL;
518 request->rq_req_cbid.cbid_fn = request_out_callback;
519 request->rq_req_cbid.cbid_arg = request;
521 request->rq_reply_cbid.cbid_fn = reply_in_callback;
522 request->rq_reply_cbid.cbid_arg = request;
524 request->rq_reply_deadline = 0;
525 request->rq_phase = RQ_PHASE_NEW;
526 request->rq_next_phase = RQ_PHASE_UNDEFINED;
528 request->rq_request_portal = imp->imp_client->cli_request_portal;
529 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
531 ptlrpc_at_set_req_timeout(request);
533 spin_lock_init(&request->rq_lock);
534 CFS_INIT_LIST_HEAD(&request->rq_list);
535 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
536 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
537 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
538 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
539 CFS_INIT_LIST_HEAD(&request->rq_history_list);
540 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
541 cfs_waitq_init(&request->rq_reply_waitq);
542 request->rq_xid = ptlrpc_next_xid();
543 atomic_set(&request->rq_refcount, 1);
545 lustre_msg_set_opc(request->rq_reqmsg, opcode);
549 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
551 class_import_put(imp);
555 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
556 __u32 version, int opcode, char **bufs,
557 struct ptlrpc_cli_ctx *ctx)
561 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
562 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
563 request->rq_pill.rc_area[RCL_CLIENT],
566 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
568 int ptlrpc_request_pack(struct ptlrpc_request *request,
569 __u32 version, int opcode)
571 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
575 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
576 struct ptlrpc_request_pool *pool)
578 struct ptlrpc_request *request = NULL;
581 request = ptlrpc_prep_req_from_pool(pool);
584 OBD_ALLOC_PTR(request);
587 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
588 LASSERT(imp != LP_POISON);
589 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
591 LASSERT(imp->imp_client != LP_POISON);
593 request->rq_import = class_import_get(imp);
595 CERROR("request allocation out of memory\n");
601 static struct ptlrpc_request *
602 ptlrpc_request_alloc_internal(struct obd_import *imp,
603 struct ptlrpc_request_pool * pool,
604 const struct req_format *format)
606 struct ptlrpc_request *request;
608 request = __ptlrpc_request_alloc(imp, pool);
612 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
613 req_capsule_set(&request->rq_pill, format);
617 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
618 const struct req_format *format)
620 return ptlrpc_request_alloc_internal(imp, NULL, format);
623 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
624 struct ptlrpc_request_pool * pool,
625 const struct req_format *format)
627 return ptlrpc_request_alloc_internal(imp, pool, format);
630 void ptlrpc_request_free(struct ptlrpc_request *request)
632 if (request->rq_pool)
633 __ptlrpc_free_req_to_pool(request);
635 OBD_FREE_PTR(request);
638 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
639 const struct req_format *format,
640 __u32 version, int opcode)
642 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
646 rc = ptlrpc_request_pack(req, version, opcode);
648 ptlrpc_request_free(req);
655 struct ptlrpc_request *
656 ptlrpc_prep_req_pool(struct obd_import *imp,
657 __u32 version, int opcode,
658 int count, __u32 *lengths, char **bufs,
659 struct ptlrpc_request_pool *pool)
661 struct ptlrpc_request *request;
664 request = __ptlrpc_request_alloc(imp, pool);
668 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
669 lengths, bufs, NULL);
671 ptlrpc_request_free(request);
677 struct ptlrpc_request *
678 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
679 __u32 *lengths, char **bufs)
681 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
685 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
686 unsigned int timeout,
687 ptlrpc_interpterer_t interpreter)
689 struct ptlrpc_request *request = NULL;
692 OBD_ALLOC(request, sizeof(*request));
694 CERROR("request allocation out of memory\n");
698 request->rq_send_state = LUSTRE_IMP_FULL;
699 request->rq_type = PTL_RPC_MSG_REQUEST;
700 request->rq_import = class_import_get(imp);
701 request->rq_export = NULL;
702 request->rq_import_generation = imp->imp_generation;
704 request->rq_timeout = timeout;
705 request->rq_sent = cfs_time_current_sec();
706 request->rq_deadline = request->rq_sent + timeout;
707 request->rq_reply_deadline = request->rq_deadline;
708 request->rq_interpret_reply = interpreter;
709 request->rq_phase = RQ_PHASE_RPC;
710 request->rq_next_phase = RQ_PHASE_INTERPRET;
711 /* don't want reply */
712 request->rq_receiving_reply = 0;
713 request->rq_must_unlink = 0;
714 request->rq_no_delay = request->rq_no_resend = 1;
715 request->rq_fake = 1;
717 spin_lock_init(&request->rq_lock);
718 CFS_INIT_LIST_HEAD(&request->rq_list);
719 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
720 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
721 CFS_INIT_LIST_HEAD(&request->rq_history_list);
722 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
723 cfs_waitq_init(&request->rq_reply_waitq);
725 request->rq_xid = ptlrpc_next_xid();
726 atomic_set(&request->rq_refcount, 1);
731 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
733 /* if we kill request before timeout - need adjust counter */
734 if (req->rq_phase == RQ_PHASE_RPC) {
735 struct ptlrpc_request_set *set = req->rq_set;
738 set->set_remaining --;
741 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
742 list_del_init(&req->rq_list);
746 struct ptlrpc_request_set *ptlrpc_prep_set(void)
748 struct ptlrpc_request_set *set;
751 OBD_ALLOC(set, sizeof *set);
754 CFS_INIT_LIST_HEAD(&set->set_requests);
755 cfs_waitq_init(&set->set_waitq);
756 set->set_remaining = 0;
757 spin_lock_init(&set->set_new_req_lock);
758 CFS_INIT_LIST_HEAD(&set->set_new_requests);
759 CFS_INIT_LIST_HEAD(&set->set_cblist);
764 /* Finish with this set; opposite of prep_set. */
765 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
767 struct list_head *tmp;
768 struct list_head *next;
773 /* Requests on the set should either all be completed, or all be new */
774 expected_phase = (set->set_remaining == 0) ?
775 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
776 list_for_each (tmp, &set->set_requests) {
777 struct ptlrpc_request *req =
778 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
780 LASSERT(req->rq_phase == expected_phase);
784 LASSERTF(set->set_remaining == 0 || set->set_remaining == n, "%d / %d\n",
785 set->set_remaining, n);
787 list_for_each_safe(tmp, next, &set->set_requests) {
788 struct ptlrpc_request *req =
789 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
790 list_del_init(&req->rq_set_chain);
792 LASSERT(req->rq_phase == expected_phase);
794 if (req->rq_phase == RQ_PHASE_NEW) {
795 ptlrpc_req_interpret(NULL, req, -EBADR);
796 set->set_remaining--;
800 ptlrpc_req_finished (req);
803 LASSERT(set->set_remaining == 0);
805 OBD_FREE(set, sizeof(*set));
809 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
810 set_interpreter_func fn, void *data)
812 struct ptlrpc_set_cbdata *cbdata;
814 OBD_ALLOC_PTR(cbdata);
818 cbdata->psc_interpret = fn;
819 cbdata->psc_data = data;
820 list_add_tail(&cbdata->psc_item, &set->set_cblist);
825 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
826 struct ptlrpc_request *req)
828 /* The set takes over the caller's request reference */
829 list_add_tail(&req->rq_set_chain, &set->set_requests);
831 set->set_remaining++;
835 * Lock so many callers can add things, the context that owns the set
836 * is supposed to notice these and move them into the set proper.
838 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
839 struct ptlrpc_request *req)
841 struct ptlrpc_request_set *set = pc->pc_set;
844 * Let caller know that we stopped and will not handle this request.
845 * It needs to take care itself of request.
847 if (test_bit(LIOD_STOP, &pc->pc_flags))
850 spin_lock(&set->set_new_req_lock);
852 * The set takes over the caller's request reference.
854 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
856 spin_unlock(&set->set_new_req_lock);
858 cfs_waitq_signal(&set->set_waitq);
863 * Based on the current state of the import, determine if the request
864 * can be sent, is an error, or should be delayed.
866 * Returns true if this request should be delayed. If false, and
867 * *status is set, then the request can not be sent and *status is the
868 * error code. If false and status is 0, then request can be sent.
870 * The imp->imp_lock must be held.
872 static int ptlrpc_import_delay_req(struct obd_import *imp,
873 struct ptlrpc_request *req, int *status)
878 LASSERT (status != NULL);
881 if (req->rq_ctx_init || req->rq_ctx_fini) {
882 /* always allow ctx init/fini rpc go through */
883 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
884 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
887 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
888 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
890 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
891 imp->imp_state == LUSTRE_IMP_CONNECTING) {
892 /* allow CONNECT even if import is invalid */ ;
893 if (atomic_read(&imp->imp_inval_count) != 0) {
894 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
897 } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
898 imp->imp_obd->obd_no_recov) {
899 /* If the import has been invalidated (such as by an OST
900 * failure), and if the import(MGC) tried all of its connection
901 * list (Bug 13464), the request must fail with -ESHUTDOWN.
902 * This indicates the requests should be discarded; an -EIO
903 * may result in a resend of the request. */
904 if (!imp->imp_deactive)
905 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
906 *status = -ESHUTDOWN; /* bz 12940 */
907 } else if (req->rq_import_generation != imp->imp_generation) {
908 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
910 } else if (req->rq_send_state != imp->imp_state) {
911 /* invalidate in progress - any requests should be drop */
912 if (atomic_read(&imp->imp_inval_count) != 0) {
913 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
915 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
916 *status = -EWOULDBLOCK;
925 static int ptlrpc_check_reply(struct ptlrpc_request *req)
930 /* serialise with network callback */
931 spin_lock(&req->rq_lock);
933 if (ptlrpc_client_replied(req))
936 if (req->rq_net_err && !req->rq_timedout) {
937 spin_unlock(&req->rq_lock);
938 rc = ptlrpc_expire_one_request(req, 0);
939 spin_lock(&req->rq_lock);
952 if (ptlrpc_client_early(req)) {
953 ptlrpc_at_recv_early_reply(req);
954 GOTO(out, rc = 0); /* keep waiting */
959 spin_unlock(&req->rq_lock);
960 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
964 /* Conditionally suppress specific console messages */
965 static int ptlrpc_console_allow(struct ptlrpc_request *req)
967 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
970 /* Suppress particular reconnect errors which are to be expected. No
971 * errors are suppressed for the initial connection on an import */
972 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
973 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
975 /* Suppress timed out reconnect requests */
976 if (req->rq_timedout)
979 /* Suppress unavailable/again reconnect requests */
980 err = lustre_msg_get_status(req->rq_repmsg);
981 if (err == -ENODEV || err == -EAGAIN)
989 static int ptlrpc_check_status(struct ptlrpc_request *req)
994 err = lustre_msg_get_status(req->rq_repmsg);
995 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
996 struct obd_import *imp = req->rq_import;
997 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
998 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
999 " with %s. The %s operation failed with %d\n",
1000 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1001 ll_opcode2str(opc), err);
1002 RETURN(err < 0 ? err : -EINVAL);
1006 DEBUG_REQ(D_INFO, req, "status is %d", err);
1007 } else if (err > 0) {
1008 /* XXX: translate this error from net to host */
1009 DEBUG_REQ(D_INFO, req, "status is %d", err);
1012 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1013 struct obd_import *imp = req->rq_import;
1014 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1016 if (ptlrpc_console_allow(req))
1017 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1018 "communicating with %s. The %s "
1019 "operation failed with %d\n",
1021 imp->imp_connection->c_peer.nid),
1022 ll_opcode2str(opc), err);
1024 RETURN(err < 0 ? err : -EINVAL);
1031 * save pre-versions for replay
1033 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1035 struct lustre_msg *repmsg = req->rq_repmsg;
1036 struct lustre_msg *reqmsg = req->rq_reqmsg;
1037 __u64 *versions = lustre_msg_get_versions(repmsg);
1040 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1044 lustre_msg_set_versions(reqmsg, versions);
1045 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1046 versions[0], versions[1]);
1052 * Callback function called when client receives RPC reply for \a req.
1054 static int after_reply(struct ptlrpc_request *req)
1056 struct obd_import *imp = req->rq_import;
1057 struct obd_device *obd = req->rq_import->imp_obd;
1059 struct timeval work_start;
1063 LASSERT(!req->rq_receiving_reply);
1065 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
1067 if (req->rq_reply_truncate && !req->rq_no_resend) {
1069 sptlrpc_cli_free_repbuf(req);
1070 req->rq_replen = req->rq_nob_received;
1075 * NB Until this point, the whole of the incoming message,
1076 * including buflens, status etc is in the sender's byte order.
1079 rc = sptlrpc_cli_unwrap_reply(req);
1081 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1086 * Security layer unwrap might ask resend this request.
1091 rc = unpack_reply(req);
1095 do_gettimeofday(&work_start);
1096 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1097 if (obd->obd_svc_stats != NULL) {
1098 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1100 ptlrpc_lprocfs_rpc_sent(req, timediff);
1103 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1104 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1105 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1106 lustre_msg_get_type(req->rq_repmsg));
1110 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1111 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1112 ptlrpc_at_adj_net_latency(req,
1113 lustre_msg_get_service_time(req->rq_repmsg));
1115 rc = ptlrpc_check_status(req);
1116 imp->imp_connect_error = rc;
1120 * Either we've been evicted, or the server has failed for
1121 * some reason. Try to reconnect, and if that fails, punt to
1124 if (ll_rpc_recoverable_error(rc)) {
1125 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1126 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1129 ptlrpc_request_handle_notconn(req);
1134 * Let's look if server sent slv. Do it only for RPC with
1137 ldlm_cli_update_pool(req);
1141 * Store transno in reqmsg for replay.
1143 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1144 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1145 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1148 if (imp->imp_replayable) {
1149 spin_lock(&imp->imp_lock);
1151 * No point in adding already-committed requests to the replay
1152 * list, we will just remove them immediately. b=9829
1154 if (req->rq_transno != 0 &&
1156 lustre_msg_get_last_committed(req->rq_repmsg) ||
1158 /** version recovery */
1159 ptlrpc_save_versions(req);
1160 ptlrpc_retain_replayable_request(req, imp);
1161 } else if (req->rq_commit_cb != NULL) {
1162 spin_unlock(&imp->imp_lock);
1163 req->rq_commit_cb(req);
1164 spin_lock(&imp->imp_lock);
1168 * Replay-enabled imports return commit-status information.
1170 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1171 imp->imp_peer_committed_transno =
1172 lustre_msg_get_last_committed(req->rq_repmsg);
1174 ptlrpc_free_committed(imp);
1175 spin_unlock(&imp->imp_lock);
1181 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1183 struct obd_import *imp;
1187 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1188 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1191 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1193 imp = req->rq_import;
1194 spin_lock(&imp->imp_lock);
1196 req->rq_import_generation = imp->imp_generation;
1198 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1199 spin_lock(&req->rq_lock);
1200 req->rq_waiting = 1;
1201 spin_unlock(&req->rq_lock);
1203 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1204 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1205 ptlrpc_import_state_name(req->rq_send_state),
1206 ptlrpc_import_state_name(imp->imp_state));
1207 LASSERT(list_empty(&req->rq_list));
1208 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1209 atomic_inc(&req->rq_import->imp_inflight);
1210 spin_unlock(&imp->imp_lock);
1215 spin_unlock(&imp->imp_lock);
1216 req->rq_status = rc;
1217 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1221 LASSERT(list_empty(&req->rq_list));
1222 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1223 atomic_inc(&req->rq_import->imp_inflight);
1224 spin_unlock(&imp->imp_lock);
1226 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1228 rc = sptlrpc_req_refresh_ctx(req, -1);
1231 req->rq_status = rc;
1234 req->rq_wait_ctx = 1;
1239 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1240 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1241 imp->imp_obd->obd_uuid.uuid,
1242 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1243 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1244 lustre_msg_get_opc(req->rq_reqmsg));
1246 rc = ptl_send_rpc(req, 0);
1248 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1249 req->rq_net_err = 1;
1255 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1256 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1258 struct list_head *tmp;
1259 int force_timer_recalc = 0;
1262 if (set->set_remaining == 0)
1265 list_for_each(tmp, &set->set_requests) {
1266 struct ptlrpc_request *req =
1267 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1268 struct obd_import *imp = req->rq_import;
1271 if (req->rq_phase == RQ_PHASE_NEW &&
1272 ptlrpc_send_new_req(req)) {
1273 force_timer_recalc = 1;
1276 /* delayed send - skip */
1277 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1280 if (!(req->rq_phase == RQ_PHASE_RPC ||
1281 req->rq_phase == RQ_PHASE_BULK ||
1282 req->rq_phase == RQ_PHASE_INTERPRET ||
1283 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1284 req->rq_phase == RQ_PHASE_COMPLETE)) {
1285 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1289 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1290 LASSERT(req->rq_next_phase != req->rq_phase);
1291 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1294 * Skip processing until reply is unlinked. We
1295 * can't return to pool before that and we can't
1296 * call interpret before that. We need to make
1297 * sure that all rdma transfers finished and will
1298 * not corrupt any data.
1300 if (ptlrpc_client_recv_or_unlink(req) ||
1301 ptlrpc_client_bulk_active(req))
1305 * Turn fail_loc off to prevent it from looping
1308 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1309 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1312 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1313 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1318 * Move to next phase if reply was successfully
1321 ptlrpc_rqphase_move(req, req->rq_next_phase);
1324 if (req->rq_phase == RQ_PHASE_COMPLETE)
1327 if (req->rq_phase == RQ_PHASE_INTERPRET)
1328 GOTO(interpret, req->rq_status);
1331 * Note that this also will start async reply unlink.
1333 if (req->rq_net_err && !req->rq_timedout) {
1334 ptlrpc_expire_one_request(req, 1);
1337 * Check if we still need to wait for unlink.
1339 if (ptlrpc_client_recv_or_unlink(req) ||
1340 ptlrpc_client_bulk_active(req))
1345 req->rq_replied = 0;
1346 if (req->rq_status == 0)
1347 req->rq_status = -EIO;
1348 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1349 GOTO(interpret, req->rq_status);
1352 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1353 * will only be set after rq_timedout, but the synchronous IO
1354 * waiting path sets rq_intr irrespective of whether ptlrpcd
1355 * has seen a timeout. our policy is to only interpret
1356 * interrupted rpcs after they have timed out */
1357 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1358 req->rq_wait_ctx)) {
1359 req->rq_status = -EINTR;
1360 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1361 GOTO(interpret, req->rq_status);
1364 if (req->rq_phase == RQ_PHASE_RPC) {
1365 if (req->rq_timedout || req->rq_resend ||
1366 req->rq_waiting || req->rq_wait_ctx) {
1369 if (!ptlrpc_unregister_reply(req, 1))
1372 spin_lock(&imp->imp_lock);
1373 if (ptlrpc_import_delay_req(imp, req, &status)){
1374 /* put on delay list - only if we wait
1375 * recovery finished - before send */
1376 list_del_init(&req->rq_list);
1377 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1378 spin_unlock(&imp->imp_lock);
1383 req->rq_status = status;
1384 ptlrpc_rqphase_move(req,
1385 RQ_PHASE_INTERPRET);
1386 spin_unlock(&imp->imp_lock);
1387 GOTO(interpret, req->rq_status);
1389 if (req->rq_no_resend && !req->rq_wait_ctx) {
1390 req->rq_status = -ENOTCONN;
1391 ptlrpc_rqphase_move(req,
1392 RQ_PHASE_INTERPRET);
1393 spin_unlock(&imp->imp_lock);
1394 GOTO(interpret, req->rq_status);
1397 list_del_init(&req->rq_list);
1398 list_add_tail(&req->rq_list,
1399 &imp->imp_sending_list);
1401 spin_unlock(&imp->imp_lock);
1403 req->rq_waiting = 0;
1405 if (req->rq_timedout||req->rq_resend) {
1406 /* This is re-sending anyways,
1407 * let's mark req as resend. */
1412 if (!ptlrpc_unregister_bulk(req, 1))
1415 /* ensure previous bulk fails */
1416 old_xid = req->rq_xid;
1417 req->rq_xid = ptlrpc_next_xid();
1418 CDEBUG(D_HA, "resend bulk "
1421 old_xid, req->rq_xid);
1425 * rq_wait_ctx is only touched by ptlrpcd,
1426 * so no lock is needed here.
1428 status = sptlrpc_req_refresh_ctx(req, -1);
1431 req->rq_status = status;
1432 req->rq_wait_ctx = 0;
1433 force_timer_recalc = 1;
1435 req->rq_wait_ctx = 1;
1440 req->rq_wait_ctx = 0;
1443 rc = ptl_send_rpc(req, 0);
1445 DEBUG_REQ(D_HA, req, "send failed (%d)",
1447 force_timer_recalc = 1;
1448 req->rq_net_err = 1;
1450 /* need to reset the timeout */
1451 force_timer_recalc = 1;
1454 spin_lock(&req->rq_lock);
1456 if (ptlrpc_client_early(req)) {
1457 ptlrpc_at_recv_early_reply(req);
1458 spin_unlock(&req->rq_lock);
1462 /* Still waiting for a reply? */
1463 if (ptlrpc_client_recv(req)) {
1464 spin_unlock(&req->rq_lock);
1468 /* Did we actually receive a reply? */
1469 if (!ptlrpc_client_replied(req)) {
1470 spin_unlock(&req->rq_lock);
1474 spin_unlock(&req->rq_lock);
1476 req->rq_status = after_reply(req);
1480 /* If there is no bulk associated with this request,
1481 * then we're done and should let the interpreter
1482 * process the reply. Similarly if the RPC returned
1483 * an error, and therefore the bulk will never arrive.
1485 if (req->rq_bulk == NULL || req->rq_status != 0) {
1486 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1487 GOTO(interpret, req->rq_status);
1490 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1493 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1494 if (ptlrpc_client_bulk_active(req))
1497 if (!req->rq_bulk->bd_success) {
1498 /* The RPC reply arrived OK, but the bulk screwed
1499 * up! Dead weird since the server told us the RPC
1500 * was good after getting the REPLY for her GET or
1501 * the ACK for her PUT. */
1502 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1506 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1509 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1511 /* This moves to "unregistering" phase we need to wait for
1513 if (!ptlrpc_unregister_reply(req, 1))
1516 if (!ptlrpc_unregister_bulk(req, 1))
1519 /* When calling interpret receiving already should be
1521 LASSERT(!req->rq_receiving_reply);
1523 ptlrpc_req_interpret(env, req, req->rq_status);
1525 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1527 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1528 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1529 imp->imp_obd->obd_uuid.uuid,
1530 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1532 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1533 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1535 spin_lock(&imp->imp_lock);
1536 /* Request already may be not on sending or delaying list. This
1537 * may happen in the case of marking it erroneous for the case
1538 * ptlrpc_import_delay_req(req, status) find it impossible to
1539 * allow sending this rpc and returns *status != 0. */
1540 if (!list_empty(&req->rq_list)) {
1541 list_del_init(&req->rq_list);
1542 atomic_dec(&imp->imp_inflight);
1544 spin_unlock(&imp->imp_lock);
1546 set->set_remaining--;
1547 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1550 /* If we hit an error, we want to recover promptly. */
1551 RETURN(set->set_remaining == 0 || force_timer_recalc);
1554 /* Return 1 if we should give up, else 0 */
1555 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1557 struct obd_import *imp = req->rq_import;
1561 spin_lock(&req->rq_lock);
1562 req->rq_timedout = 1;
1563 spin_unlock(&req->rq_lock);
1565 DEBUG_REQ(D_WARNING, req, "Request x"LPU64" sent from %s to NID %s "
1566 CFS_DURATION_T"s ago has %s (limit "CFS_DURATION_T"s).\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 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 spin_lock(&req->rq_lock);
1605 req->rq_status = -ETIMEDOUT;
1607 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;
1626 struct list_head *tmp;
1627 time_t now = cfs_time_current_sec();
1630 LASSERT(set != NULL);
1633 * A timeout expired. See which reqs it applies to...
1635 list_for_each (tmp, &set->set_requests) {
1636 struct ptlrpc_request *req =
1637 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1639 /* don't expire request waiting for context */
1640 if (req->rq_wait_ctx)
1643 /* Request in-flight? */
1644 if (!((req->rq_phase == RQ_PHASE_RPC &&
1645 !req->rq_waiting && !req->rq_resend) ||
1646 (req->rq_phase == RQ_PHASE_BULK)))
1649 if (req->rq_timedout || /* already dealt with */
1650 req->rq_deadline > now) /* not expired */
1653 /* Deal with this guy. Do it asynchronously to not block
1654 * ptlrpcd thread. */
1655 ptlrpc_expire_one_request(req, 1);
1659 * When waiting for a whole set, we always break out of the
1660 * sleep so we can recalculate the timeout, or enable interrupts
1661 * if everyone's timed out.
1666 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1668 spin_lock(&req->rq_lock);
1670 spin_unlock(&req->rq_lock);
1673 void ptlrpc_interrupted_set(void *data)
1675 struct ptlrpc_request_set *set = data;
1676 struct list_head *tmp;
1678 LASSERT(set != NULL);
1679 CERROR("INTERRUPTED SET %p\n", set);
1681 list_for_each(tmp, &set->set_requests) {
1682 struct ptlrpc_request *req =
1683 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1685 if (req->rq_phase != RQ_PHASE_RPC &&
1686 req->rq_phase != RQ_PHASE_UNREGISTERING)
1689 ptlrpc_mark_interrupted(req);
1694 * Get the smallest timeout in the set; this does NOT set a timeout.
1696 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1698 struct list_head *tmp;
1699 time_t now = cfs_time_current_sec();
1701 struct ptlrpc_request *req;
1705 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1707 list_for_each(tmp, &set->set_requests) {
1708 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1711 * Request in-flight?
1713 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1714 (req->rq_phase == RQ_PHASE_BULK) ||
1715 (req->rq_phase == RQ_PHASE_NEW)))
1719 * Already timed out.
1721 if (req->rq_timedout)
1727 if (req->rq_wait_ctx)
1730 if (req->rq_phase == RQ_PHASE_NEW)
1731 deadline = req->rq_sent;
1733 deadline = req->rq_sent + req->rq_timeout;
1735 if (deadline <= now) /* actually expired already */
1736 timeout = 1; /* ASAP */
1737 else if (timeout == 0 || timeout > deadline - now)
1738 timeout = deadline - now;
1743 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1745 struct list_head *tmp;
1746 struct ptlrpc_request *req;
1747 struct l_wait_info lwi;
1751 if (list_empty(&set->set_requests))
1754 list_for_each(tmp, &set->set_requests) {
1755 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1756 if (req->rq_phase == RQ_PHASE_NEW)
1757 (void)ptlrpc_send_new_req(req);
1761 timeout = ptlrpc_set_next_timeout(set);
1763 /* wait until all complete, interrupted, or an in-flight
1765 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1767 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1769 ptlrpc_interrupted_set, set);
1770 rc = l_wait_event(set->set_waitq,
1771 ptlrpc_check_set(NULL, set), &lwi);
1773 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1775 /* -EINTR => all requests have been flagged rq_intr so next
1777 * -ETIMEDOUT => someone timed out. When all reqs have
1778 * timed out, signals are enabled allowing completion with
1780 * I don't really care if we go once more round the loop in
1781 * the error cases -eeb. */
1782 } while (rc != 0 || set->set_remaining != 0);
1784 LASSERT(set->set_remaining == 0);
1787 list_for_each(tmp, &set->set_requests) {
1788 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1790 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1791 if (req->rq_status != 0)
1792 rc = req->rq_status;
1795 if (set->set_interpret != NULL) {
1796 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1798 rc = interpreter (set, set->set_arg, rc);
1800 struct ptlrpc_set_cbdata *cbdata, *n;
1803 list_for_each_entry_safe(cbdata, n,
1804 &set->set_cblist, psc_item) {
1805 list_del_init(&cbdata->psc_item);
1806 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1809 OBD_FREE_PTR(cbdata);
1816 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1819 if (request == NULL) {
1824 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1825 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1826 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1827 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1828 LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request);
1829 LASSERTF(!request->rq_replay, "req %p\n", request);
1830 LASSERT(request->rq_cli_ctx || request->rq_fake);
1832 req_capsule_fini(&request->rq_pill);
1834 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1835 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1836 if (request->rq_import != NULL) {
1838 spin_lock(&request->rq_import->imp_lock);
1839 list_del_init(&request->rq_replay_list);
1841 spin_unlock(&request->rq_import->imp_lock);
1843 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1845 if (atomic_read(&request->rq_refcount) != 0) {
1846 DEBUG_REQ(D_ERROR, request,
1847 "freeing request with nonzero refcount");
1851 if (request->rq_repbuf != NULL)
1852 sptlrpc_cli_free_repbuf(request);
1853 if (request->rq_export != NULL) {
1854 class_export_put(request->rq_export);
1855 request->rq_export = NULL;
1857 if (request->rq_import != NULL) {
1858 class_import_put(request->rq_import);
1859 request->rq_import = NULL;
1861 if (request->rq_bulk != NULL)
1862 ptlrpc_free_bulk(request->rq_bulk);
1864 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1865 sptlrpc_cli_free_reqbuf(request);
1867 if (request->rq_cli_ctx)
1868 sptlrpc_req_put_ctx(request, !locked);
1870 if (request->rq_pool)
1871 __ptlrpc_free_req_to_pool(request);
1873 OBD_FREE(request, sizeof(*request));
1877 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1878 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1880 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1881 (void)__ptlrpc_req_finished(request, 1);
1884 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1887 if (request == NULL)
1890 if (request == LP_POISON ||
1891 request->rq_reqmsg == LP_POISON) {
1892 CERROR("dereferencing freed request (bug 575)\n");
1897 DEBUG_REQ(D_INFO, request, "refcount now %u",
1898 atomic_read(&request->rq_refcount) - 1);
1900 if (atomic_dec_and_test(&request->rq_refcount)) {
1901 __ptlrpc_free_req(request, locked);
1908 void ptlrpc_req_finished(struct ptlrpc_request *request)
1910 __ptlrpc_req_finished(request, 0);
1913 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1915 return request->rq_xid;
1917 EXPORT_SYMBOL(ptlrpc_req_xid);
1919 /* Disengage the client's reply buffer from the network
1920 * NB does _NOT_ unregister any client-side bulk.
1921 * IDEMPOTENT, but _not_ safe against concurrent callers.
1922 * The request owner (i.e. the thread doing the I/O) must call...
1924 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1928 struct l_wait_info lwi;
1933 LASSERT(!in_interrupt());
1936 * Let's setup deadline for reply unlink.
1938 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1939 async && request->rq_reply_deadline == 0)
1940 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1943 * Nothing left to do.
1945 if (!ptlrpc_client_recv_or_unlink(request))
1948 LNetMDUnlink(request->rq_reply_md_h);
1951 * Let's check it once again.
1953 if (!ptlrpc_client_recv_or_unlink(request))
1957 * Move to "Unregistering" phase as reply was not unlinked yet.
1959 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1962 * Do not wait for unlink to finish.
1968 * We have to l_wait_event() whatever the result, to give liblustre
1969 * a chance to run reply_in_callback(), and to make sure we've
1970 * unlinked before returning a req to the pool.
1972 if (request->rq_set != NULL)
1973 wq = &request->rq_set->set_waitq;
1975 wq = &request->rq_reply_waitq;
1978 /* Network access will complete in finite time but the HUGE
1979 * timeout lets us CWARN for visibility of sluggish NALs */
1980 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1981 cfs_time_seconds(1), NULL, NULL);
1982 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1985 ptlrpc_rqphase_move(request, request->rq_next_phase);
1989 LASSERT(rc == -ETIMEDOUT);
1990 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1991 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1992 request->rq_must_unlink);
1997 /* caller must hold imp->imp_lock */
1998 void ptlrpc_free_committed(struct obd_import *imp)
2000 struct list_head *tmp, *saved;
2001 struct ptlrpc_request *req;
2002 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2005 LASSERT(imp != NULL);
2007 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2010 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2011 imp->imp_generation == imp->imp_last_generation_checked) {
2012 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
2013 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2017 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2018 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2019 imp->imp_generation);
2020 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2021 imp->imp_last_generation_checked = imp->imp_generation;
2023 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2024 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
2026 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2027 LASSERT(req != last_req);
2030 if (req->rq_transno == 0) {
2031 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2034 if (req->rq_import_generation < imp->imp_generation) {
2035 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2039 if (req->rq_replay) {
2040 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2044 /* not yet committed */
2045 if (req->rq_transno > imp->imp_peer_committed_transno) {
2046 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2050 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
2051 imp->imp_peer_committed_transno);
2053 spin_lock(&req->rq_lock);
2055 spin_unlock(&req->rq_lock);
2056 if (req->rq_commit_cb != NULL)
2057 req->rq_commit_cb(req);
2058 list_del_init(&req->rq_replay_list);
2059 __ptlrpc_req_finished(req, 1);
2066 void ptlrpc_cleanup_client(struct obd_import *imp)
2073 void ptlrpc_resend_req(struct ptlrpc_request *req)
2075 DEBUG_REQ(D_HA, req, "going to resend");
2076 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2077 req->rq_status = -EAGAIN;
2079 spin_lock(&req->rq_lock);
2081 req->rq_net_err = 0;
2082 req->rq_timedout = 0;
2084 __u64 old_xid = req->rq_xid;
2086 /* ensure previous bulk fails */
2087 req->rq_xid = ptlrpc_next_xid();
2088 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2089 old_xid, req->rq_xid);
2091 ptlrpc_client_wake_req(req);
2092 spin_unlock(&req->rq_lock);
2095 /* XXX: this function and rq_status are currently unused */
2096 void ptlrpc_restart_req(struct ptlrpc_request *req)
2098 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2099 req->rq_status = -ERESTARTSYS;
2101 spin_lock(&req->rq_lock);
2102 req->rq_restart = 1;
2103 req->rq_timedout = 0;
2104 ptlrpc_client_wake_req(req);
2105 spin_unlock(&req->rq_lock);
2108 static int expired_request(void *data)
2110 struct ptlrpc_request *req = data;
2114 * Some failure can suspend regular timeouts.
2116 if (ptlrpc_check_suspend())
2120 * Deadline may have changed with an early reply.
2122 if (req->rq_deadline > cfs_time_current_sec())
2125 RETURN(ptlrpc_expire_one_request(req, 0));
2128 static void interrupted_request(void *data)
2130 struct ptlrpc_request *req = data;
2131 DEBUG_REQ(D_HA, req, "request interrupted");
2132 spin_lock(&req->rq_lock);
2134 spin_unlock(&req->rq_lock);
2137 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2140 atomic_inc(&req->rq_refcount);
2144 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2145 struct obd_import *imp)
2147 struct list_head *tmp;
2149 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2151 if (req->rq_transno == 0) {
2152 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2156 /* clear this for new requests that were resent as well
2157 as resent replayed requests. */
2158 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2160 /* don't re-add requests that have been replayed */
2161 if (!list_empty(&req->rq_replay_list))
2164 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2166 LASSERT(imp->imp_replayable);
2167 /* Balanced in ptlrpc_free_committed, usually. */
2168 ptlrpc_request_addref(req);
2169 list_for_each_prev(tmp, &imp->imp_replay_list) {
2170 struct ptlrpc_request *iter =
2171 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
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 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2192 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
2195 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2199 struct l_wait_info lwi;
2200 struct obd_import *imp = req->rq_import;
2201 cfs_duration_t timeout = CFS_TICK;
2205 LASSERT(req->rq_set == NULL);
2206 LASSERT(!req->rq_receiving_reply);
2208 /* for distributed debugging */
2209 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2210 LASSERT(imp->imp_obd != NULL);
2211 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
2212 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2213 imp->imp_obd->obd_uuid.uuid,
2214 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2215 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2216 lustre_msg_get_opc(req->rq_reqmsg));
2218 /* Mark phase here for a little debug help */
2219 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
2221 spin_lock(&imp->imp_lock);
2222 req->rq_import_generation = imp->imp_generation;
2224 if (ptlrpc_import_delay_req(imp, req, &rc)) {
2225 list_del_init(&req->rq_list);
2226 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
2227 atomic_inc(&imp->imp_inflight);
2228 spin_unlock(&imp->imp_lock);
2230 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
2232 ptlrpc_import_state_name(req->rq_send_state),
2233 ptlrpc_import_state_name(imp->imp_state));
2234 lwi = LWI_INTR(interrupted_request, req);
2235 rc = l_wait_event(req->rq_reply_waitq,
2236 (req->rq_send_state == imp->imp_state ||
2237 req->rq_err || req->rq_intr),
2239 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2241 ptlrpc_import_state_name(imp->imp_state),
2242 ptlrpc_import_state_name(req->rq_send_state),
2243 req->rq_err, req->rq_intr);
2245 spin_lock(&imp->imp_lock);
2246 list_del_init(&req->rq_list);
2247 atomic_dec(&imp->imp_inflight);
2250 /* rq_status was set locally */
2251 rc = req->rq_status ? req->rq_status : -EIO;
2253 else if (req->rq_intr) {
2256 else if (req->rq_no_resend) {
2264 spin_unlock(&imp->imp_lock);
2265 req->rq_status = rc; // XXX this ok?
2269 if (req->rq_resend) {
2270 if (req->rq_bulk != NULL) {
2271 ptlrpc_unregister_bulk(req, 0);
2273 /* bulk requests are supposed to be
2274 * idempotent, so we are free to bump the xid
2275 * here, which we need to do before
2276 * registering the bulk again (bug 6371).
2277 * print the old xid first for sanity.
2279 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2280 req->rq_xid = ptlrpc_next_xid();
2283 DEBUG_REQ(D_HA, req, "resending: ");
2286 /* XXX this is the same as ptlrpc_set_wait */
2287 LASSERT(list_empty(&req->rq_list));
2288 list_add_tail(&req->rq_list, &imp->imp_sending_list);
2289 atomic_inc(&imp->imp_inflight);
2290 spin_unlock(&imp->imp_lock);
2292 rc = sptlrpc_req_refresh_ctx(req, 0);
2295 /* we got fatal ctx refresh error, directly jump out
2296 * thus we can pass back the actual error code.
2298 spin_lock(&imp->imp_lock);
2299 list_del_init(&req->rq_list);
2300 atomic_dec(&imp->imp_inflight);
2301 spin_unlock(&imp->imp_lock);
2303 CERROR("Failed to refresh ctx of req %p: %d\n",
2307 /* simulating we got error during send rpc */
2311 rc = ptl_send_rpc(req, 0);
2313 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2315 timeoutl = req->rq_deadline - cfs_time_current_sec();
2316 timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2317 cfs_time_seconds(timeoutl);
2318 DEBUG_REQ(D_NET, req,
2319 "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2320 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
2322 brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
2323 if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
2324 ptlrpc_check_and_wait_suspend(req)))
2328 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2329 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2330 imp->imp_obd->obd_uuid.uuid,
2331 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2332 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2333 lustre_msg_get_opc(req->rq_reqmsg));
2335 /* If the reply was received normally, this just grabs the spinlock
2336 * (ensuring the reply callback has returned), sees that
2337 * req->rq_receiving_reply is clear and returns. */
2338 ptlrpc_unregister_reply(req, 0);
2340 spin_lock(&imp->imp_lock);
2341 list_del_init(&req->rq_list);
2342 atomic_dec(&imp->imp_inflight);
2343 spin_unlock(&imp->imp_lock);
2346 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2347 rc, req->rq_status);
2348 rc = rc ? rc : req->rq_status;
2349 GOTO(out, rc = rc ? rc : -EIO);
2353 /* Should only be interrupted if we timed out. */
2354 if (!req->rq_timedout)
2355 DEBUG_REQ(D_ERROR, req,
2356 "rq_intr set but rq_timedout not");
2357 GOTO(out, rc = -EINTR);
2360 /* Resend if we need to */
2361 if (req->rq_resend||req->rq_timedout) {
2362 /* ...unless we were specifically told otherwise. */
2363 if (req->rq_no_resend)
2364 GOTO(out, rc = -ETIMEDOUT);
2365 spin_lock(&imp->imp_lock);
2366 /* we can have rq_timeout on dlm fake import which not support
2367 * recovery - but me need resend request on this import instead
2368 * of return error */
2373 if (!ptlrpc_client_replied(req)) {
2374 /* How can this be? -eeb */
2375 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2377 GOTO(out, rc = req->rq_status);
2380 rc = after_reply(req);
2381 /* NB may return +ve success rc */
2382 if (req->rq_resend) {
2383 spin_lock(&imp->imp_lock);
2388 if (req->rq_bulk != NULL) {
2390 /* success so far. Note that anything going wrong
2391 * with bulk now, is EXTREMELY strange, since the
2392 * server must have believed that the bulk
2393 * transferred OK before she replied with success to
2395 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2396 brc = l_wait_event(req->rq_reply_waitq,
2397 !ptlrpc_client_bulk_active(req),
2399 LASSERT(brc == 0 || brc == -ETIMEDOUT);
2401 LASSERT(brc == -ETIMEDOUT);
2402 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2404 } else if (!req->rq_bulk->bd_success) {
2405 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2410 ptlrpc_unregister_bulk(req, 0);
2413 LASSERT(!req->rq_receiving_reply);
2414 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2415 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
2419 struct ptlrpc_replay_async_args {
2421 int praa_old_status;
2424 static int ptlrpc_replay_interpret(const struct lu_env *env,
2425 struct ptlrpc_request *req,
2426 void * data, int rc)
2428 struct ptlrpc_replay_async_args *aa = data;
2429 struct obd_import *imp = req->rq_import;
2432 atomic_dec(&imp->imp_replay_inflight);
2434 if (!ptlrpc_client_replied(req)) {
2435 CERROR("request replay timed out, restarting recovery\n");
2436 GOTO(out, rc = -ETIMEDOUT);
2439 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2440 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2441 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2442 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2444 /** VBR: check version failure */
2445 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2446 /** replay was failed due to version mismatch */
2447 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2448 spin_lock(&imp->imp_lock);
2449 imp->imp_vbr_failed = 1;
2450 imp->imp_no_lock_replay = 1;
2451 spin_unlock(&imp->imp_lock);
2453 /** The transno had better not change over replay. */
2454 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2455 lustre_msg_get_transno(req->rq_repmsg) ||
2456 lustre_msg_get_transno(req->rq_repmsg) == 0,
2458 lustre_msg_get_transno(req->rq_reqmsg),
2459 lustre_msg_get_transno(req->rq_repmsg));
2462 spin_lock(&imp->imp_lock);
2463 /** if replays by version then gap was occur on server, no trust to locks */
2464 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2465 imp->imp_no_lock_replay = 1;
2466 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2467 spin_unlock(&imp->imp_lock);
2468 LASSERT(imp->imp_last_replay_transno);
2470 DEBUG_REQ(D_HA, req, "got rep");
2472 /* let the callback do fixups, possibly including in the request */
2473 if (req->rq_replay_cb)
2474 req->rq_replay_cb(req);
2476 if (ptlrpc_client_replied(req) &&
2477 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2478 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2479 lustre_msg_get_status(req->rq_repmsg),
2480 aa->praa_old_status);
2482 /* Put it back for re-replay. */
2483 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2487 * Errors while replay can set transno to 0, but
2488 * imp_last_replay_transno shouldn't be set to 0 anyway
2490 if (req->rq_transno > 0) {
2491 spin_lock(&imp->imp_lock);
2492 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2493 imp->imp_last_replay_transno = req->rq_transno;
2494 spin_unlock(&imp->imp_lock);
2496 CERROR("Transno is 0 during replay!\n");
2497 /* continue with recovery */
2498 rc = ptlrpc_import_recovery_state_machine(imp);
2500 req->rq_send_state = aa->praa_old_state;
2503 /* this replay failed, so restart recovery */
2504 ptlrpc_connect_import(imp, NULL);
2509 int ptlrpc_replay_req(struct ptlrpc_request *req)
2511 struct ptlrpc_replay_async_args *aa;
2514 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2515 /* Not handling automatic bulk replay yet (or ever?) */
2516 LASSERT(req->rq_bulk == NULL);
2518 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2519 aa = ptlrpc_req_async_args(req);
2520 memset(aa, 0, sizeof *aa);
2522 /* Prepare request to be resent with ptlrpcd */
2523 aa->praa_old_state = req->rq_send_state;
2524 req->rq_send_state = LUSTRE_IMP_REPLAY;
2525 req->rq_phase = RQ_PHASE_NEW;
2526 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2528 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2530 req->rq_interpret_reply = ptlrpc_replay_interpret;
2531 /* Readjust the timeout for current conditions */
2532 ptlrpc_at_set_req_timeout(req);
2534 DEBUG_REQ(D_HA, req, "REPLAY");
2536 atomic_inc(&req->rq_import->imp_replay_inflight);
2537 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2539 ptlrpcd_add_req(req, PSCOPE_OTHER);
2543 void ptlrpc_abort_inflight(struct obd_import *imp)
2545 struct list_head *tmp, *n;
2548 /* Make sure that no new requests get processed for this import.
2549 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2550 * this flag and then putting requests on sending_list or delayed_list.
2552 spin_lock(&imp->imp_lock);
2554 /* XXX locking? Maybe we should remove each request with the list
2555 * locked? Also, how do we know if the requests on the list are
2556 * being freed at this time?
2558 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2559 struct ptlrpc_request *req =
2560 list_entry(tmp, struct ptlrpc_request, rq_list);
2562 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2564 spin_lock (&req->rq_lock);
2565 if (req->rq_import_generation < imp->imp_generation) {
2567 req->rq_status = -EINTR;
2568 ptlrpc_client_wake_req(req);
2570 spin_unlock (&req->rq_lock);
2573 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2574 struct ptlrpc_request *req =
2575 list_entry(tmp, struct ptlrpc_request, rq_list);
2577 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2579 spin_lock (&req->rq_lock);
2580 if (req->rq_import_generation < imp->imp_generation) {
2582 req->rq_status = -EINTR;
2583 ptlrpc_client_wake_req(req);
2585 spin_unlock (&req->rq_lock);
2588 /* Last chance to free reqs left on the replay list, but we
2589 * will still leak reqs that haven't committed. */
2590 if (imp->imp_replayable)
2591 ptlrpc_free_committed(imp);
2593 spin_unlock(&imp->imp_lock);
2598 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2600 struct list_head *tmp, *pos;
2602 LASSERT(set != NULL);
2604 list_for_each_safe(pos, tmp, &set->set_requests) {
2605 struct ptlrpc_request *req =
2606 list_entry(pos, struct ptlrpc_request, rq_set_chain);
2608 spin_lock(&req->rq_lock);
2609 if (req->rq_phase != RQ_PHASE_RPC) {
2610 spin_unlock(&req->rq_lock);
2615 req->rq_status = -EINTR;
2616 ptlrpc_client_wake_req(req);
2617 spin_unlock(&req->rq_lock);
2621 static __u64 ptlrpc_last_xid;
2622 static spinlock_t ptlrpc_last_xid_lock;
2624 /* Initialize the XID for the node. This is common among all requests on
2625 * this node, and only requires the property that it is monotonically
2626 * increasing. It does not need to be sequential. Since this is also used
2627 * as the RDMA match bits, it is important that a single client NOT have
2628 * the same match bits for two different in-flight requests, hence we do
2629 * NOT want to have an XID per target or similar.
2631 * To avoid an unlikely collision between match bits after a client reboot
2632 * (which would deliver old data into the wrong RDMA buffer) initialize
2633 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2634 * If the time is clearly incorrect, we instead use a 62-bit random number.
2635 * In the worst case the random number will overflow 1M RPCs per second in
2636 * 9133 years, or permutations thereof.
2638 #define YEAR_2004 (1ULL << 30)
2639 void ptlrpc_init_xid(void)
2641 time_t now = cfs_time_current_sec();
2643 spin_lock_init(&ptlrpc_last_xid_lock);
2644 if (now < YEAR_2004) {
2645 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2646 ptlrpc_last_xid >>= 2;
2647 ptlrpc_last_xid |= (1ULL << 61);
2649 ptlrpc_last_xid = (__u64)now << 20;
2653 __u64 ptlrpc_next_xid(void)
2656 spin_lock(&ptlrpc_last_xid_lock);
2657 tmp = ++ptlrpc_last_xid;
2658 spin_unlock(&ptlrpc_last_xid_lock);
2662 __u64 ptlrpc_sample_next_xid(void)
2664 #if BITS_PER_LONG == 32
2665 /* need to avoid possible word tearing on 32-bit systems */
2667 spin_lock(&ptlrpc_last_xid_lock);
2668 tmp = ptlrpc_last_xid + 1;
2669 spin_unlock(&ptlrpc_last_xid_lock);
2672 /* No need to lock, since returned value is racy anyways */
2673 return ptlrpc_last_xid + 1;
2676 EXPORT_SYMBOL(ptlrpc_sample_next_xid);