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 | LUSTRE_IMP_CONNECTING)))
231 LASSERT(req->rq_import);
232 at = &req->rq_import->imp_at;
234 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
235 /* max service estimates are tracked on the server side,
236 so just keep minimal history here */
237 oldse = at_add(&at->iat_service_estimate[idx], serv_est);
239 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
240 "has changed from %d to %d\n",
241 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
242 oldse, at_get(&at->iat_service_estimate[idx]));
245 /* Expected network latency per remote node (secs) */
246 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
248 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
251 /* Adjust expected network latency */
252 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
253 unsigned int service_time)
255 unsigned int nl, oldnl;
257 time_t now = cfs_time_current_sec();
259 LASSERT(req->rq_import);
260 at = &req->rq_import->imp_at;
262 /* Network latency is total time less server processing time */
263 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
264 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
265 CWARN("Reported service time %u > total measured time "
266 CFS_DURATION_T"\n", service_time,
267 cfs_time_sub(now, req->rq_sent));
269 oldnl = at_add(&at->iat_net_latency, nl);
271 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
272 "has changed from %d to %d\n",
273 req->rq_import->imp_obd->obd_name,
275 &req->rq_import->imp_connection->c_remote_uuid),
276 oldnl, at_get(&at->iat_net_latency));
279 static int unpack_reply(struct ptlrpc_request *req)
283 /* Clear reply swab mask; we may have already swabbed an early reply */
284 req->rq_rep_swab_mask = 0;
286 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
288 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_set *ptlrpc_prep_set(void)
687 struct ptlrpc_request_set *set;
690 OBD_ALLOC(set, sizeof *set);
693 CFS_INIT_LIST_HEAD(&set->set_requests);
694 cfs_waitq_init(&set->set_waitq);
695 set->set_remaining = 0;
696 spin_lock_init(&set->set_new_req_lock);
697 CFS_INIT_LIST_HEAD(&set->set_new_requests);
698 CFS_INIT_LIST_HEAD(&set->set_cblist);
703 /* Finish with this set; opposite of prep_set. */
704 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
706 struct list_head *tmp;
707 struct list_head *next;
712 /* Requests on the set should either all be completed, or all be new */
713 expected_phase = (set->set_remaining == 0) ?
714 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
715 list_for_each (tmp, &set->set_requests) {
716 struct ptlrpc_request *req =
717 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
719 LASSERT(req->rq_phase == expected_phase);
723 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
725 list_for_each_safe(tmp, next, &set->set_requests) {
726 struct ptlrpc_request *req =
727 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
728 list_del_init(&req->rq_set_chain);
730 LASSERT(req->rq_phase == expected_phase);
732 if (req->rq_phase == RQ_PHASE_NEW) {
734 if (req->rq_interpret_reply != NULL) {
735 ptlrpc_interpterer_t interpreter =
736 req->rq_interpret_reply;
738 /* higher level (i.e. LOV) failed;
739 * let the sub reqs clean up */
740 req->rq_status = -EBADR;
741 interpreter(NULL, req, &req->rq_async_args,
744 set->set_remaining--;
748 ptlrpc_req_finished (req);
751 LASSERT(set->set_remaining == 0);
753 OBD_FREE(set, sizeof(*set));
757 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
758 set_interpreter_func fn, void *data)
760 struct ptlrpc_set_cbdata *cbdata;
762 OBD_ALLOC_PTR(cbdata);
766 cbdata->psc_interpret = fn;
767 cbdata->psc_data = data;
768 list_add_tail(&cbdata->psc_item, &set->set_cblist);
773 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
774 struct ptlrpc_request *req)
776 /* The set takes over the caller's request reference */
777 list_add_tail(&req->rq_set_chain, &set->set_requests);
779 set->set_remaining++;
783 * Lock so many callers can add things, the context that owns the set
784 * is supposed to notice these and move them into the set proper.
786 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
787 struct ptlrpc_request *req)
789 struct ptlrpc_request_set *set = pc->pc_set;
792 * Let caller know that we stopped and will not handle this request.
793 * It needs to take care itself of request.
795 if (test_bit(LIOD_STOP, &pc->pc_flags))
798 spin_lock(&set->set_new_req_lock);
800 * The set takes over the caller's request reference.
802 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
804 spin_unlock(&set->set_new_req_lock);
807 * Let thead know that we added something and better it to wake up
810 cfs_waitq_signal(&set->set_waitq);
815 * Based on the current state of the import, determine if the request
816 * can be sent, is an error, or should be delayed.
818 * Returns true if this request should be delayed. If false, and
819 * *status is set, then the request can not be sent and *status is the
820 * error code. If false and status is 0, then request can be sent.
822 * The imp->imp_lock must be held.
824 static int ptlrpc_import_delay_req(struct obd_import *imp,
825 struct ptlrpc_request *req, int *status)
830 LASSERT (status != NULL);
833 if (req->rq_ctx_init || req->rq_ctx_fini) {
834 /* always allow ctx init/fini rpc go through */
835 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
836 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
839 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
840 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
842 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
843 imp->imp_state == LUSTRE_IMP_CONNECTING) {
844 /* allow CONNECT even if import is invalid */ ;
845 if (atomic_read(&imp->imp_inval_count) != 0) {
846 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
849 } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
850 imp->imp_obd->obd_no_recov) {
851 /* If the import has been invalidated (such as by an OST
852 * failure), and if the import(MGC) tried all of its connection
853 * list (Bug 13464), the request must fail with -ESHUTDOWN.
854 * This indicates the requests should be discarded; an -EIO
855 * may result in a resend of the request. */
856 if (!imp->imp_deactive)
857 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
858 *status = -ESHUTDOWN; /* bz 12940 */
859 } else if (req->rq_import_generation != imp->imp_generation) {
860 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
862 } else if (req->rq_send_state != imp->imp_state) {
863 /* invalidate in progress - any requests should be drop */
864 if (atomic_read(&imp->imp_inval_count) != 0) {
865 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
867 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
868 *status = -EWOULDBLOCK;
877 static int ptlrpc_check_reply(struct ptlrpc_request *req)
882 /* serialise with network callback */
883 spin_lock(&req->rq_lock);
885 if (ptlrpc_client_replied(req))
888 if (req->rq_net_err && !req->rq_timedout) {
889 spin_unlock(&req->rq_lock);
890 rc = ptlrpc_expire_one_request(req, 0);
891 spin_lock(&req->rq_lock);
904 if (ptlrpc_client_early(req)) {
905 ptlrpc_at_recv_early_reply(req);
906 GOTO(out, rc = 0); /* keep waiting */
911 spin_unlock(&req->rq_lock);
912 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
916 static int ptlrpc_check_status(struct ptlrpc_request *req)
921 err = lustre_msg_get_status(req->rq_repmsg);
922 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
923 struct obd_import *imp = req->rq_import;
924 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
925 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
926 " with %s. The %s operation failed with %d\n",
927 libcfs_nid2str(imp->imp_connection->c_peer.nid),
928 ll_opcode2str(opc), err);
929 RETURN(err < 0 ? err : -EINVAL);
933 DEBUG_REQ(D_INFO, req, "status is %d", err);
934 } else if (err > 0) {
935 /* XXX: translate this error from net to host */
936 DEBUG_REQ(D_INFO, req, "status is %d", err);
943 * Callback function called when client receives RPC reply for \a req.
945 static int after_reply(struct ptlrpc_request *req)
947 struct obd_import *imp = req->rq_import;
948 struct obd_device *obd = req->rq_import->imp_obd;
950 struct timeval work_start;
954 LASSERT(!req->rq_receiving_reply);
956 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
959 * NB Until this point, the whole of the incoming message,
960 * including buflens, status etc is in the sender's byte order.
963 rc = sptlrpc_cli_unwrap_reply(req);
965 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
970 * Security layer unwrap might ask resend this request.
975 rc = unpack_reply(req);
979 do_gettimeofday(&work_start);
980 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
981 if (obd->obd_svc_stats != NULL)
982 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
985 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
986 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
987 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
988 lustre_msg_get_type(req->rq_repmsg));
992 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
993 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
994 ptlrpc_at_adj_net_latency(req,
995 lustre_msg_get_service_time(req->rq_repmsg));
997 rc = ptlrpc_check_status(req);
998 imp->imp_connect_error = rc;
1002 * Either we've been evicted, or the server has failed for
1003 * some reason. Try to reconnect, and if that fails, punt to
1006 if (ll_rpc_recoverable_error(rc)) {
1007 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1008 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1011 ptlrpc_request_handle_notconn(req);
1016 * Let's look if server sent slv. Do it only for RPC with
1019 ldlm_cli_update_pool(req);
1023 * Store transno in reqmsg for replay.
1025 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1026 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1028 if (req->rq_import->imp_replayable) {
1029 spin_lock(&imp->imp_lock);
1031 * No point in adding already-committed requests to the replay
1032 * list, we will just remove them immediately. b=9829
1034 if (req->rq_transno != 0 &&
1036 lustre_msg_get_last_committed(req->rq_repmsg) ||
1038 ptlrpc_retain_replayable_request(req, imp);
1039 else if (req->rq_commit_cb != NULL) {
1040 spin_unlock(&imp->imp_lock);
1041 req->rq_commit_cb(req);
1042 spin_lock(&imp->imp_lock);
1046 * Replay-enabled imports return commit-status information.
1048 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1049 imp->imp_peer_committed_transno =
1050 lustre_msg_get_last_committed(req->rq_repmsg);
1052 ptlrpc_free_committed(imp);
1053 spin_unlock(&imp->imp_lock);
1059 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1061 struct obd_import *imp;
1065 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1066 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1069 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1071 imp = req->rq_import;
1072 spin_lock(&imp->imp_lock);
1074 req->rq_import_generation = imp->imp_generation;
1076 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1077 spin_lock(&req->rq_lock);
1078 req->rq_waiting = 1;
1079 spin_unlock(&req->rq_lock);
1081 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1082 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1083 ptlrpc_import_state_name(req->rq_send_state),
1084 ptlrpc_import_state_name(imp->imp_state));
1085 LASSERT(list_empty(&req->rq_list));
1086 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1087 atomic_inc(&req->rq_import->imp_inflight);
1088 spin_unlock(&imp->imp_lock);
1093 spin_unlock(&imp->imp_lock);
1094 req->rq_status = rc;
1095 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1099 LASSERT(list_empty(&req->rq_list));
1100 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1101 atomic_inc(&req->rq_import->imp_inflight);
1102 spin_unlock(&imp->imp_lock);
1104 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1106 rc = sptlrpc_req_refresh_ctx(req, -1);
1109 req->rq_status = rc;
1112 req->rq_wait_ctx = 1;
1117 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1118 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1119 imp->imp_obd->obd_uuid.uuid,
1120 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1121 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1122 lustre_msg_get_opc(req->rq_reqmsg));
1124 rc = ptl_send_rpc(req, 0);
1126 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1127 req->rq_net_err = 1;
1133 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1134 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1136 struct list_head *tmp;
1137 int force_timer_recalc = 0;
1140 if (set->set_remaining == 0)
1143 list_for_each(tmp, &set->set_requests) {
1144 struct ptlrpc_request *req =
1145 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1146 struct obd_import *imp = req->rq_import;
1149 if (req->rq_phase == RQ_PHASE_NEW &&
1150 ptlrpc_send_new_req(req)) {
1151 force_timer_recalc = 1;
1154 /* delayed send - skip */
1155 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1158 if (!(req->rq_phase == RQ_PHASE_RPC ||
1159 req->rq_phase == RQ_PHASE_BULK ||
1160 req->rq_phase == RQ_PHASE_INTERPRET ||
1161 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1162 req->rq_phase == RQ_PHASE_COMPLETE)) {
1163 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1167 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1168 LASSERT(req->rq_next_phase != req->rq_phase);
1169 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1172 * Skip processing until reply is unlinked. We
1173 * can't return to pool before that and we can't
1174 * call interpret before that. We need to make
1175 * sure that all rdma transfers finished and will
1176 * not corrupt any data.
1178 if (ptlrpc_client_recv_or_unlink(req) ||
1179 ptlrpc_client_bulk_active(req))
1183 * Turn fail_loc off to prevent it from looping
1186 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1187 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1190 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1191 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1196 * Move to next phase if reply was successfully
1199 ptlrpc_rqphase_move(req, req->rq_next_phase);
1202 if (req->rq_phase == RQ_PHASE_COMPLETE)
1205 if (req->rq_phase == RQ_PHASE_INTERPRET)
1206 GOTO(interpret, req->rq_status);
1209 * Note that this also will start async reply unlink.
1211 if (req->rq_net_err && !req->rq_timedout) {
1212 ptlrpc_expire_one_request(req, 1);
1215 * Check if we still need to wait for unlink.
1217 if (ptlrpc_client_recv_or_unlink(req) ||
1218 ptlrpc_client_bulk_active(req))
1223 req->rq_replied = 0;
1224 if (req->rq_status == 0)
1225 req->rq_status = -EIO;
1226 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1227 GOTO(interpret, req->rq_status);
1230 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1231 * will only be set after rq_timedout, but the synchronous IO
1232 * waiting path sets rq_intr irrespective of whether ptlrpcd
1233 * has seen a timeout. our policy is to only interpret
1234 * interrupted rpcs after they have timed out */
1235 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1236 req->rq_wait_ctx)) {
1237 req->rq_status = -EINTR;
1238 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1239 GOTO(interpret, req->rq_status);
1242 if (req->rq_phase == RQ_PHASE_RPC) {
1243 if (req->rq_timedout || req->rq_resend ||
1244 req->rq_waiting || req->rq_wait_ctx) {
1247 if (!ptlrpc_unregister_reply(req, 1))
1250 spin_lock(&imp->imp_lock);
1251 if (ptlrpc_import_delay_req(imp, req, &status)){
1252 /* put on delay list - only if we wait
1253 * recovery finished - before send */
1254 list_del_init(&req->rq_list);
1255 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1256 spin_unlock(&imp->imp_lock);
1261 req->rq_status = status;
1262 ptlrpc_rqphase_move(req,
1263 RQ_PHASE_INTERPRET);
1264 spin_unlock(&imp->imp_lock);
1265 GOTO(interpret, req->rq_status);
1267 if (req->rq_no_resend && !req->rq_wait_ctx) {
1268 req->rq_status = -ENOTCONN;
1269 ptlrpc_rqphase_move(req,
1270 RQ_PHASE_INTERPRET);
1271 spin_unlock(&imp->imp_lock);
1272 GOTO(interpret, req->rq_status);
1275 list_del_init(&req->rq_list);
1276 list_add_tail(&req->rq_list,
1277 &imp->imp_sending_list);
1279 spin_unlock(&imp->imp_lock);
1281 req->rq_waiting = 0;
1283 if (req->rq_timedout||req->rq_resend) {
1284 /* This is re-sending anyways,
1285 * let's mark req as resend. */
1290 if (!ptlrpc_unregister_bulk(req, 1))
1293 /* ensure previous bulk fails */
1294 old_xid = req->rq_xid;
1295 req->rq_xid = ptlrpc_next_xid();
1296 CDEBUG(D_HA, "resend bulk "
1299 old_xid, req->rq_xid);
1303 * rq_wait_ctx is only touched by ptlrpcd,
1304 * so no lock is needed here.
1306 status = sptlrpc_req_refresh_ctx(req, -1);
1309 req->rq_status = status;
1310 req->rq_wait_ctx = 0;
1311 force_timer_recalc = 1;
1313 req->rq_wait_ctx = 1;
1318 req->rq_wait_ctx = 0;
1321 rc = ptl_send_rpc(req, 0);
1323 DEBUG_REQ(D_HA, req, "send failed (%d)",
1325 force_timer_recalc = 1;
1326 req->rq_net_err = 1;
1328 /* need to reset the timeout */
1329 force_timer_recalc = 1;
1332 spin_lock(&req->rq_lock);
1334 if (ptlrpc_client_early(req)) {
1335 ptlrpc_at_recv_early_reply(req);
1336 spin_unlock(&req->rq_lock);
1340 /* Still waiting for a reply? */
1341 if (ptlrpc_client_recv(req)) {
1342 spin_unlock(&req->rq_lock);
1346 /* Did we actually receive a reply? */
1347 if (!ptlrpc_client_replied(req)) {
1348 spin_unlock(&req->rq_lock);
1352 spin_unlock(&req->rq_lock);
1354 req->rq_status = after_reply(req);
1358 /* If there is no bulk associated with this request,
1359 * then we're done and should let the interpreter
1360 * process the reply. Similarly if the RPC returned
1361 * an error, and therefore the bulk will never arrive.
1363 if (req->rq_bulk == NULL || req->rq_status != 0) {
1364 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1365 GOTO(interpret, req->rq_status);
1368 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1371 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1372 if (ptlrpc_client_bulk_active(req))
1375 if (!req->rq_bulk->bd_success) {
1376 /* The RPC reply arrived OK, but the bulk screwed
1377 * up! Dead wierd since the server told us the RPC
1378 * was good after getting the REPLY for her GET or
1379 * the ACK for her PUT. */
1380 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1384 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1387 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1389 /* This moves to "unregistering" phase we need to wait for
1391 if (!ptlrpc_unregister_reply(req, 1))
1394 if (!ptlrpc_unregister_bulk(req, 1))
1397 /* When calling interpret receiving already should be
1399 LASSERT(!req->rq_receiving_reply);
1401 if (req->rq_interpret_reply != NULL) {
1402 ptlrpc_interpterer_t interpreter =
1403 req->rq_interpret_reply;
1404 req->rq_status = interpreter(env, req,
1405 &req->rq_async_args,
1408 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1410 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1411 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1412 imp->imp_obd->obd_uuid.uuid,
1413 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1414 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1415 lustre_msg_get_opc(req->rq_reqmsg));
1417 spin_lock(&imp->imp_lock);
1418 /* Request already may be not on sending or delaying list. This
1419 * may happen in the case of marking it errorneous for the case
1420 * ptlrpc_import_delay_req(req, status) find it impossible to
1421 * allow sending this rpc and returns *status != 0. */
1422 if (!list_empty(&req->rq_list)) {
1423 list_del_init(&req->rq_list);
1424 atomic_dec(&imp->imp_inflight);
1426 spin_unlock(&imp->imp_lock);
1428 set->set_remaining--;
1429 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1432 /* If we hit an error, we want to recover promptly. */
1433 RETURN(set->set_remaining == 0 || force_timer_recalc);
1436 /* Return 1 if we should give up, else 0 */
1437 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1439 struct obd_import *imp = req->rq_import;
1443 DEBUG_REQ(D_ERROR|D_NETERROR, req,
1444 "%s (sent at "CFS_TIME_T", "CFS_DURATION_T"s ago)",
1445 req->rq_net_err ? "network error" : "timeout",
1446 req->rq_sent, cfs_time_sub(cfs_time_current_sec(),
1450 LCONSOLE_WARN("Request x"LPU64" sent from %s to NID %s "
1451 CFS_DURATION_T"s ago has timed out "
1452 "(limit "CFS_DURATION_T"s).\n", req->rq_xid,
1453 req->rq_import->imp_obd->obd_name,
1454 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1455 cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1456 cfs_time_sub(req->rq_deadline, req->rq_sent));
1459 if (imp != NULL && obd_debug_peer_on_timeout)
1460 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1462 spin_lock(&req->rq_lock);
1463 req->rq_timedout = 1;
1464 spin_unlock(&req->rq_lock);
1466 ptlrpc_unregister_reply(req, async_unlink);
1467 ptlrpc_unregister_bulk(req, async_unlink);
1469 if (obd_dump_on_timeout)
1470 libcfs_debug_dumplog();
1473 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1477 /* The DLM server doesn't want recovery run on its imports. */
1478 if (imp->imp_dlm_fake)
1481 /* If this request is for recovery or other primordial tasks,
1482 * then error it out here. */
1483 if (req->rq_ctx_init || req->rq_ctx_fini ||
1484 req->rq_send_state != LUSTRE_IMP_FULL ||
1485 imp->imp_obd->obd_no_recov) {
1486 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1487 ptlrpc_import_state_name(req->rq_send_state),
1488 ptlrpc_import_state_name(imp->imp_state));
1489 spin_lock(&req->rq_lock);
1490 req->rq_status = -ETIMEDOUT;
1492 spin_unlock(&req->rq_lock);
1496 /* if a request can't be resent we can't wait for an answer after
1498 if (req->rq_no_resend) {
1499 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1503 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1508 int ptlrpc_expired_set(void *data)
1510 struct ptlrpc_request_set *set = data;
1511 struct list_head *tmp;
1512 time_t now = cfs_time_current_sec();
1515 LASSERT(set != NULL);
1518 * A timeout expired. See which reqs it applies to...
1520 list_for_each (tmp, &set->set_requests) {
1521 struct ptlrpc_request *req =
1522 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1524 /* don't expire request waiting for context */
1525 if (req->rq_wait_ctx)
1528 /* Request in-flight? */
1529 if (!((req->rq_phase == RQ_PHASE_RPC &&
1530 !req->rq_waiting && !req->rq_resend) ||
1531 (req->rq_phase == RQ_PHASE_BULK)))
1534 if (req->rq_timedout || /* already dealt with */
1535 req->rq_deadline > now) /* not expired */
1538 /* Deal with this guy. Do it asynchronously to not block
1539 * ptlrpcd thread. */
1540 ptlrpc_expire_one_request(req, 1);
1544 * When waiting for a whole set, we always to break out of the
1545 * sleep so we can recalculate the timeout, or enable interrupts
1546 * if everyone's timed out.
1551 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1553 spin_lock(&req->rq_lock);
1555 spin_unlock(&req->rq_lock);
1558 void ptlrpc_interrupted_set(void *data)
1560 struct ptlrpc_request_set *set = data;
1561 struct list_head *tmp;
1563 LASSERT(set != NULL);
1564 CERROR("INTERRUPTED SET %p\n", set);
1566 list_for_each(tmp, &set->set_requests) {
1567 struct ptlrpc_request *req =
1568 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1570 if (req->rq_phase != RQ_PHASE_RPC &&
1571 req->rq_phase != RQ_PHASE_UNREGISTERING)
1574 ptlrpc_mark_interrupted(req);
1579 * Get the smallest timeout in the set; this does NOT set a timeout.
1581 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1583 struct list_head *tmp;
1584 time_t now = cfs_time_current_sec();
1586 struct ptlrpc_request *req;
1590 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1592 list_for_each(tmp, &set->set_requests) {
1593 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1596 * Request in-flight?
1598 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1599 (req->rq_phase == RQ_PHASE_BULK) ||
1600 (req->rq_phase == RQ_PHASE_NEW)))
1604 * Already timed out.
1606 if (req->rq_timedout)
1612 if (req->rq_wait_ctx)
1615 if (req->rq_phase == RQ_PHASE_NEW)
1616 deadline = req->rq_sent;
1618 deadline = req->rq_sent + req->rq_timeout;
1620 if (deadline <= now) /* actually expired already */
1621 timeout = 1; /* ASAP */
1622 else if (timeout == 0 || timeout > deadline - now)
1623 timeout = deadline - now;
1628 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1630 struct list_head *tmp;
1631 struct ptlrpc_request *req;
1632 struct l_wait_info lwi;
1636 if (list_empty(&set->set_requests))
1639 list_for_each(tmp, &set->set_requests) {
1640 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1641 if (req->rq_phase == RQ_PHASE_NEW)
1642 (void)ptlrpc_send_new_req(req);
1646 timeout = ptlrpc_set_next_timeout(set);
1648 /* wait until all complete, interrupted, or an in-flight
1650 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1652 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1654 ptlrpc_interrupted_set, set);
1655 rc = l_wait_event(set->set_waitq,
1656 ptlrpc_check_set(NULL, set), &lwi);
1658 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1660 /* -EINTR => all requests have been flagged rq_intr so next
1662 * -ETIMEOUTD => someone timed out. When all reqs have
1663 * timed out, signals are enabled allowing completion with
1665 * I don't really care if we go once more round the loop in
1666 * the error cases -eeb. */
1667 } while (rc != 0 || set->set_remaining != 0);
1669 LASSERT(set->set_remaining == 0);
1672 list_for_each(tmp, &set->set_requests) {
1673 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1675 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1676 if (req->rq_status != 0)
1677 rc = req->rq_status;
1680 if (set->set_interpret != NULL) {
1681 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1683 rc = interpreter (set, set->set_arg, rc);
1685 struct ptlrpc_set_cbdata *cbdata, *n;
1688 list_for_each_entry_safe(cbdata, n,
1689 &set->set_cblist, psc_item) {
1690 list_del_init(&cbdata->psc_item);
1691 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1694 OBD_FREE_PTR(cbdata);
1701 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1704 if (request == NULL) {
1709 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1710 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1711 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1712 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1713 LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request);
1714 LASSERTF(!request->rq_replay, "req %p\n", request);
1715 LASSERT(request->rq_cli_ctx);
1717 req_capsule_fini(&request->rq_pill);
1719 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1720 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1721 if (request->rq_import != NULL) {
1723 spin_lock(&request->rq_import->imp_lock);
1724 list_del_init(&request->rq_replay_list);
1726 spin_unlock(&request->rq_import->imp_lock);
1728 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1730 if (atomic_read(&request->rq_refcount) != 0) {
1731 DEBUG_REQ(D_ERROR, request,
1732 "freeing request with nonzero refcount");
1736 if (request->rq_repbuf != NULL)
1737 sptlrpc_cli_free_repbuf(request);
1738 if (request->rq_export != NULL) {
1739 class_export_put(request->rq_export);
1740 request->rq_export = NULL;
1742 if (request->rq_import != NULL) {
1743 class_import_put(request->rq_import);
1744 request->rq_import = NULL;
1746 if (request->rq_bulk != NULL)
1747 ptlrpc_free_bulk(request->rq_bulk);
1749 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1750 sptlrpc_cli_free_reqbuf(request);
1752 sptlrpc_req_put_ctx(request, !locked);
1754 if (request->rq_pool)
1755 __ptlrpc_free_req_to_pool(request);
1757 OBD_FREE(request, sizeof(*request));
1761 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1762 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1764 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1765 (void)__ptlrpc_req_finished(request, 1);
1768 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1771 if (request == NULL)
1774 if (request == LP_POISON ||
1775 request->rq_reqmsg == LP_POISON) {
1776 CERROR("dereferencing freed request (bug 575)\n");
1781 DEBUG_REQ(D_INFO, request, "refcount now %u",
1782 atomic_read(&request->rq_refcount) - 1);
1784 if (atomic_dec_and_test(&request->rq_refcount)) {
1785 __ptlrpc_free_req(request, locked);
1792 void ptlrpc_req_finished(struct ptlrpc_request *request)
1794 __ptlrpc_req_finished(request, 0);
1797 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1799 return request->rq_xid;
1801 EXPORT_SYMBOL(ptlrpc_req_xid);
1803 /* Disengage the client's reply buffer from the network
1804 * NB does _NOT_ unregister any client-side bulk.
1805 * IDEMPOTENT, but _not_ safe against concurrent callers.
1806 * The request owner (i.e. the thread doing the I/O) must call...
1808 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1812 struct l_wait_info lwi;
1817 LASSERT(!in_interrupt());
1820 * Let's setup deadline for reply unlink.
1822 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1823 async && request->rq_reply_deadline == 0)
1824 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1827 * Nothing left to do.
1829 if (!ptlrpc_client_recv_or_unlink(request))
1832 LNetMDUnlink(request->rq_reply_md_h);
1835 * Let's check it once again.
1837 if (!ptlrpc_client_recv_or_unlink(request))
1841 * Move to "Unregistering" phase as reply was not unlinked yet.
1843 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1846 * Do not wait for unlink to finish.
1852 * We have to l_wait_event() whatever the result, to give liblustre
1853 * a chance to run reply_in_callback(), and to make sure we've
1854 * unlinked before returning a req to the pool.
1856 if (request->rq_set != NULL)
1857 wq = &request->rq_set->set_waitq;
1859 wq = &request->rq_reply_waitq;
1862 /* Network access will complete in finite time but the HUGE
1863 * timeout lets us CWARN for visibility of sluggish NALs */
1864 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1865 cfs_time_seconds(1), NULL, NULL);
1866 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1869 ptlrpc_rqphase_move(request, request->rq_next_phase);
1873 LASSERT(rc == -ETIMEDOUT);
1874 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1875 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1876 request->rq_must_unlink);
1881 /* caller must hold imp->imp_lock */
1882 void ptlrpc_free_committed(struct obd_import *imp)
1884 struct list_head *tmp, *saved;
1885 struct ptlrpc_request *req;
1886 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1889 LASSERT(imp != NULL);
1891 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1894 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1895 imp->imp_generation == imp->imp_last_generation_checked) {
1896 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1897 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1901 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1902 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1903 imp->imp_generation);
1904 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1905 imp->imp_last_generation_checked = imp->imp_generation;
1907 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1908 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1910 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1911 LASSERT(req != last_req);
1914 if (req->rq_transno == 0) {
1915 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
1918 if (req->rq_import_generation < imp->imp_generation) {
1919 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1923 if (req->rq_replay) {
1924 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1928 /* not yet committed */
1929 if (req->rq_transno > imp->imp_peer_committed_transno) {
1930 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1934 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1935 imp->imp_peer_committed_transno);
1937 spin_lock(&req->rq_lock);
1939 spin_unlock(&req->rq_lock);
1940 if (req->rq_commit_cb != NULL)
1941 req->rq_commit_cb(req);
1942 list_del_init(&req->rq_replay_list);
1943 __ptlrpc_req_finished(req, 1);
1950 void ptlrpc_cleanup_client(struct obd_import *imp)
1957 void ptlrpc_resend_req(struct ptlrpc_request *req)
1959 DEBUG_REQ(D_HA, req, "going to resend");
1960 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1961 req->rq_status = -EAGAIN;
1963 spin_lock(&req->rq_lock);
1965 req->rq_net_err = 0;
1966 req->rq_timedout = 0;
1968 __u64 old_xid = req->rq_xid;
1970 /* ensure previous bulk fails */
1971 req->rq_xid = ptlrpc_next_xid();
1972 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1973 old_xid, req->rq_xid);
1975 ptlrpc_client_wake_req(req);
1976 spin_unlock(&req->rq_lock);
1979 /* XXX: this function and rq_status are currently unused */
1980 void ptlrpc_restart_req(struct ptlrpc_request *req)
1982 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1983 req->rq_status = -ERESTARTSYS;
1985 spin_lock(&req->rq_lock);
1986 req->rq_restart = 1;
1987 req->rq_timedout = 0;
1988 ptlrpc_client_wake_req(req);
1989 spin_unlock(&req->rq_lock);
1992 static int expired_request(void *data)
1994 struct ptlrpc_request *req = data;
1998 * Some failure can suspend regular timeouts.
2000 if (ptlrpc_check_suspend())
2004 * Deadline may have changed with an early reply.
2006 if (req->rq_deadline > cfs_time_current_sec())
2009 RETURN(ptlrpc_expire_one_request(req, 0));
2012 static void interrupted_request(void *data)
2014 struct ptlrpc_request *req = data;
2015 DEBUG_REQ(D_HA, req, "request interrupted");
2016 spin_lock(&req->rq_lock);
2018 spin_unlock(&req->rq_lock);
2021 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2024 atomic_inc(&req->rq_refcount);
2028 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2029 struct obd_import *imp)
2031 struct list_head *tmp;
2033 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2035 if (req->rq_transno == 0) {
2036 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2040 /* clear this for new requests that were resent as well
2041 as resent replayed requests. */
2042 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2044 /* don't re-add requests that have been replayed */
2045 if (!list_empty(&req->rq_replay_list))
2048 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2050 LASSERT(imp->imp_replayable);
2051 /* Balanced in ptlrpc_free_committed, usually. */
2052 ptlrpc_request_addref(req);
2053 list_for_each_prev(tmp, &imp->imp_replay_list) {
2054 struct ptlrpc_request *iter =
2055 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
2057 /* We may have duplicate transnos if we create and then
2058 * open a file, or for closes retained if to match creating
2059 * opens, so use req->rq_xid as a secondary key.
2060 * (See bugs 684, 685, and 428.)
2061 * XXX no longer needed, but all opens need transnos!
2063 if (iter->rq_transno > req->rq_transno)
2066 if (iter->rq_transno == req->rq_transno) {
2067 LASSERT(iter->rq_xid != req->rq_xid);
2068 if (iter->rq_xid > req->rq_xid)
2072 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2076 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
2079 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2083 struct l_wait_info lwi;
2084 struct obd_import *imp = req->rq_import;
2085 cfs_duration_t timeout = CFS_TICK;
2089 LASSERT(req->rq_set == NULL);
2090 LASSERT(!req->rq_receiving_reply);
2092 /* for distributed debugging */
2093 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2094 LASSERT(imp->imp_obd != NULL);
2095 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
2096 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2097 imp->imp_obd->obd_uuid.uuid,
2098 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2099 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2100 lustre_msg_get_opc(req->rq_reqmsg));
2102 /* Mark phase here for a little debug help */
2103 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
2105 spin_lock(&imp->imp_lock);
2106 req->rq_import_generation = imp->imp_generation;
2108 if (ptlrpc_import_delay_req(imp, req, &rc)) {
2109 list_del_init(&req->rq_list);
2110 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
2111 atomic_inc(&imp->imp_inflight);
2112 spin_unlock(&imp->imp_lock);
2114 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
2116 ptlrpc_import_state_name(req->rq_send_state),
2117 ptlrpc_import_state_name(imp->imp_state));
2118 lwi = LWI_INTR(interrupted_request, req);
2119 rc = l_wait_event(req->rq_reply_waitq,
2120 (req->rq_send_state == imp->imp_state ||
2121 req->rq_err || req->rq_intr),
2123 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2125 ptlrpc_import_state_name(imp->imp_state),
2126 ptlrpc_import_state_name(req->rq_send_state),
2127 req->rq_err, req->rq_intr);
2129 spin_lock(&imp->imp_lock);
2130 list_del_init(&req->rq_list);
2131 atomic_dec(&imp->imp_inflight);
2134 /* rq_status was set locally */
2137 else if (req->rq_intr) {
2140 else if (req->rq_no_resend) {
2141 spin_unlock(&imp->imp_lock);
2142 GOTO(out, rc = -ETIMEDOUT);
2150 spin_unlock(&imp->imp_lock);
2151 req->rq_status = rc; // XXX this ok?
2155 if (req->rq_resend) {
2156 if (req->rq_bulk != NULL) {
2157 ptlrpc_unregister_bulk(req, 0);
2159 /* bulk requests are supposed to be
2160 * idempotent, so we are free to bump the xid
2161 * here, which we need to do before
2162 * registering the bulk again (bug 6371).
2163 * print the old xid first for sanity.
2165 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2166 req->rq_xid = ptlrpc_next_xid();
2169 DEBUG_REQ(D_HA, req, "resending: ");
2172 /* XXX this is the same as ptlrpc_set_wait */
2173 LASSERT(list_empty(&req->rq_list));
2174 list_add_tail(&req->rq_list, &imp->imp_sending_list);
2175 atomic_inc(&imp->imp_inflight);
2176 spin_unlock(&imp->imp_lock);
2178 rc = sptlrpc_req_refresh_ctx(req, 0);
2181 /* we got fatal ctx refresh error, directly jump out
2182 * thus we can pass back the actual error code.
2184 spin_lock(&imp->imp_lock);
2185 list_del_init(&req->rq_list);
2186 atomic_dec(&imp->imp_inflight);
2187 spin_unlock(&imp->imp_lock);
2189 CERROR("Failed to refresh ctx of req %p: %d\n",
2193 /* simulating we got error during send rpc */
2197 rc = ptl_send_rpc(req, 0);
2199 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2201 timeoutl = req->rq_deadline - cfs_time_current_sec();
2202 timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2203 cfs_time_seconds(timeoutl);
2204 DEBUG_REQ(D_NET, req,
2205 "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2206 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
2208 brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
2209 if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
2210 ptlrpc_check_and_wait_suspend(req)))
2214 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2215 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2216 imp->imp_obd->obd_uuid.uuid,
2217 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2218 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2219 lustre_msg_get_opc(req->rq_reqmsg));
2221 /* If the reply was received normally, this just grabs the spinlock
2222 * (ensuring the reply callback has returned), sees that
2223 * req->rq_receiving_reply is clear and returns. */
2224 ptlrpc_unregister_reply(req, 0);
2226 spin_lock(&imp->imp_lock);
2227 list_del_init(&req->rq_list);
2228 atomic_dec(&imp->imp_inflight);
2229 spin_unlock(&imp->imp_lock);
2232 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2233 rc, req->rq_status);
2234 GOTO(out, rc = rc ? rc : -EIO);
2238 /* Should only be interrupted if we timed out. */
2239 if (!req->rq_timedout)
2240 DEBUG_REQ(D_ERROR, req,
2241 "rq_intr set but rq_timedout not");
2242 GOTO(out, rc = -EINTR);
2245 /* Resend if we need to */
2246 if (req->rq_resend||req->rq_timedout) {
2247 /* ...unless we were specifically told otherwise. */
2248 if (req->rq_no_resend)
2249 GOTO(out, rc = -ETIMEDOUT);
2250 spin_lock(&imp->imp_lock);
2251 /* we can have rq_timeout on dlm fake import which not support
2252 * recovery - but me need resend request on this import instead
2253 * of return error */
2258 if (!ptlrpc_client_replied(req)) {
2259 /* How can this be? -eeb */
2260 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2262 GOTO(out, rc = req->rq_status);
2265 rc = after_reply(req);
2266 /* NB may return +ve success rc */
2267 if (req->rq_resend) {
2268 spin_lock(&imp->imp_lock);
2273 if (req->rq_bulk != NULL) {
2275 /* success so far. Note that anything going wrong
2276 * with bulk now, is EXTREMELY strange, since the
2277 * server must have believed that the bulk
2278 * tranferred OK before she replied with success to
2280 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2281 brc = l_wait_event(req->rq_reply_waitq,
2282 !ptlrpc_client_bulk_active(req),
2284 LASSERT(brc == 0 || brc == -ETIMEDOUT);
2286 LASSERT(brc == -ETIMEDOUT);
2287 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2289 } else if (!req->rq_bulk->bd_success) {
2290 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2295 ptlrpc_unregister_bulk(req, 0);
2298 LASSERT(!req->rq_receiving_reply);
2299 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2300 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
2304 struct ptlrpc_replay_async_args {
2306 int praa_old_status;
2309 static int ptlrpc_replay_interpret(const struct lu_env *env,
2310 struct ptlrpc_request *req,
2311 void * data, int rc)
2313 struct ptlrpc_replay_async_args *aa = data;
2314 struct obd_import *imp = req->rq_import;
2317 atomic_dec(&imp->imp_replay_inflight);
2319 if (!ptlrpc_client_replied(req)) {
2320 CERROR("request replay timed out, restarting recovery\n");
2321 GOTO(out, rc = -ETIMEDOUT);
2324 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2325 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2326 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2327 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2329 /* The transno had better not change over replay. */
2330 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
2331 lustre_msg_get_transno(req->rq_repmsg));
2333 DEBUG_REQ(D_HA, req, "got rep");
2335 /* let the callback do fixups, possibly including in the request */
2336 if (req->rq_replay_cb)
2337 req->rq_replay_cb(req);
2339 if (ptlrpc_client_replied(req) &&
2340 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2341 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2342 lustre_msg_get_status(req->rq_repmsg),
2343 aa->praa_old_status);
2345 /* Put it back for re-replay. */
2346 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2350 * Errors while replay can set transno to 0, but
2351 * imp_last_replay_transno shouldn't be set to 0 anyway
2353 if (req->rq_transno > 0) {
2354 spin_lock(&imp->imp_lock);
2355 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2356 imp->imp_last_replay_transno = req->rq_transno;
2357 spin_unlock(&imp->imp_lock);
2359 CERROR("Transno is 0 during replay!\n");
2360 /* continue with recovery */
2361 rc = ptlrpc_import_recovery_state_machine(imp);
2363 req->rq_send_state = aa->praa_old_state;
2366 /* this replay failed, so restart recovery */
2367 ptlrpc_connect_import(imp, NULL);
2372 int ptlrpc_replay_req(struct ptlrpc_request *req)
2374 struct ptlrpc_replay_async_args *aa;
2377 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2378 /* Not handling automatic bulk replay yet (or ever?) */
2379 LASSERT(req->rq_bulk == NULL);
2381 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2382 aa = ptlrpc_req_async_args(req);
2383 memset(aa, 0, sizeof *aa);
2385 /* Prepare request to be resent with ptlrpcd */
2386 aa->praa_old_state = req->rq_send_state;
2387 req->rq_send_state = LUSTRE_IMP_REPLAY;
2388 req->rq_phase = RQ_PHASE_NEW;
2389 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2391 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2393 req->rq_interpret_reply = ptlrpc_replay_interpret;
2394 /* Readjust the timeout for current conditions */
2395 ptlrpc_at_set_req_timeout(req);
2397 DEBUG_REQ(D_HA, req, "REPLAY");
2399 atomic_inc(&req->rq_import->imp_replay_inflight);
2400 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2402 ptlrpcd_add_req(req, PSCOPE_OTHER);
2406 void ptlrpc_abort_inflight(struct obd_import *imp)
2408 struct list_head *tmp, *n;
2411 /* Make sure that no new requests get processed for this import.
2412 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2413 * this flag and then putting requests on sending_list or delayed_list.
2415 spin_lock(&imp->imp_lock);
2417 /* XXX locking? Maybe we should remove each request with the list
2418 * locked? Also, how do we know if the requests on the list are
2419 * being freed at this time?
2421 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2422 struct ptlrpc_request *req =
2423 list_entry(tmp, struct ptlrpc_request, rq_list);
2425 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2427 spin_lock (&req->rq_lock);
2428 if (req->rq_import_generation < imp->imp_generation) {
2430 req->rq_status = -EINTR;
2431 ptlrpc_client_wake_req(req);
2433 spin_unlock (&req->rq_lock);
2436 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2437 struct ptlrpc_request *req =
2438 list_entry(tmp, struct ptlrpc_request, rq_list);
2440 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2442 spin_lock (&req->rq_lock);
2443 if (req->rq_import_generation < imp->imp_generation) {
2445 req->rq_status = -EINTR;
2446 ptlrpc_client_wake_req(req);
2448 spin_unlock (&req->rq_lock);
2451 /* Last chance to free reqs left on the replay list, but we
2452 * will still leak reqs that haven't committed. */
2453 if (imp->imp_replayable)
2454 ptlrpc_free_committed(imp);
2456 spin_unlock(&imp->imp_lock);
2461 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2463 struct list_head *tmp, *pos;
2465 LASSERT(set != NULL);
2467 list_for_each_safe(pos, tmp, &set->set_requests) {
2468 struct ptlrpc_request *req =
2469 list_entry(pos, struct ptlrpc_request, rq_set_chain);
2471 spin_lock(&req->rq_lock);
2472 if (req->rq_phase != RQ_PHASE_RPC) {
2473 spin_unlock(&req->rq_lock);
2478 req->rq_status = -EINTR;
2479 ptlrpc_client_wake_req(req);
2480 spin_unlock(&req->rq_lock);
2484 static __u64 ptlrpc_last_xid;
2485 static spinlock_t ptlrpc_last_xid_lock;
2487 /* Initialize the XID for the node. This is common among all requests on
2488 * this node, and only requires the property that it is monotonically
2489 * increasing. It does not need to be sequential. Since this is also used
2490 * as the RDMA match bits, it is important that a single client NOT have
2491 * the same match bits for two different in-flight requests, hence we do
2492 * NOT want to have an XID per target or similar.
2494 * To avoid an unlikely collision between match bits after a client reboot
2495 * (which would deliver old data into the wrong RDMA buffer) initialize
2496 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2497 * If the time is clearly incorrect, we instead use a 62-bit random number.
2498 * In the worst case the random number will overflow 1M RPCs per second in
2499 * 9133 years, or permutations thereof.
2501 #define YEAR_2004 (1ULL << 30)
2502 void ptlrpc_init_xid(void)
2504 time_t now = cfs_time_current_sec();
2506 spin_lock_init(&ptlrpc_last_xid_lock);
2507 if (now < YEAR_2004) {
2508 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2509 ptlrpc_last_xid >>= 2;
2510 ptlrpc_last_xid |= (1ULL << 61);
2512 ptlrpc_last_xid = (__u64)now << 20;
2516 __u64 ptlrpc_next_xid(void)
2519 spin_lock(&ptlrpc_last_xid_lock);
2520 tmp = ++ptlrpc_last_xid;
2521 spin_unlock(&ptlrpc_last_xid_lock);
2525 __u64 ptlrpc_sample_next_xid(void)
2527 #if BITS_PER_LONG == 32
2528 /* need to avoid possible word tearing on 32-bit systems */
2530 spin_lock(&ptlrpc_last_xid_lock);
2531 tmp = ptlrpc_last_xid + 1;
2532 spin_unlock(&ptlrpc_last_xid_lock);
2535 /* No need to lock, since returned value is racy anyways */
2536 return ptlrpc_last_xid + 1;
2539 EXPORT_SYMBOL(ptlrpc_sample_next_xid);