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(&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_mod_list);
538 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
539 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
540 CFS_INIT_LIST_HEAD(&request->rq_history_list);
541 cfs_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++;
781 atomic_inc(&req->rq_import->imp_inflight);
785 * Lock so many callers can add things, the context that owns the set
786 * is supposed to notice these and move them into the set proper.
788 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
789 struct ptlrpc_request *req)
791 struct ptlrpc_request_set *set = pc->pc_set;
794 * Let caller know that we stopped and will not handle this request.
795 * It needs to take care itself of request.
797 if (test_bit(LIOD_STOP, &pc->pc_flags))
800 spin_lock(&set->set_new_req_lock);
802 * The set takes over the caller's request reference.
804 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
806 spin_unlock(&set->set_new_req_lock);
809 * Let thead know that we added something and better it to wake up
812 cfs_waitq_signal(&set->set_waitq);
817 * Based on the current state of the import, determine if the request
818 * can be sent, is an error, or should be delayed.
820 * Returns true if this request should be delayed. If false, and
821 * *status is set, then the request can not be sent and *status is the
822 * error code. If false and status is 0, then request can be sent.
824 * The imp->imp_lock must be held.
826 static int ptlrpc_import_delay_req(struct obd_import *imp,
827 struct ptlrpc_request *req, int *status)
832 LASSERT (status != NULL);
835 if (req->rq_ctx_init || req->rq_ctx_fini) {
836 /* always allow ctx init/fini rpc go through */
837 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
838 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
841 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
842 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
844 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
845 imp->imp_state == LUSTRE_IMP_CONNECTING) {
846 /* allow CONNECT even if import is invalid */ ;
847 if (atomic_read(&imp->imp_inval_count) != 0) {
848 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
851 } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
852 imp->imp_obd->obd_no_recov) {
853 /* If the import has been invalidated (such as by an OST
854 * failure), and if the import(MGC) tried all of its connection
855 * list (Bug 13464), the request must fail with -ESHUTDOWN.
856 * This indicates the requests should be discarded; an -EIO
857 * may result in a resend of the request. */
858 if (!imp->imp_deactive)
859 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
860 *status = -ESHUTDOWN; /* bz 12940 */
861 } else if (req->rq_import_generation != imp->imp_generation) {
862 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
864 } else if (req->rq_send_state != imp->imp_state) {
865 /* invalidate in progress - any requests should be drop */
866 if (atomic_read(&imp->imp_inval_count) != 0) {
867 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
869 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
870 *status = -EWOULDBLOCK;
879 static int ptlrpc_check_reply(struct ptlrpc_request *req)
884 /* serialise with network callback */
885 spin_lock(&req->rq_lock);
887 if (ptlrpc_client_replied(req))
890 if (req->rq_net_err && !req->rq_timedout) {
891 spin_unlock(&req->rq_lock);
892 rc = ptlrpc_expire_one_request(req, 0);
893 spin_lock(&req->rq_lock);
906 if (ptlrpc_client_early(req)) {
907 ptlrpc_at_recv_early_reply(req);
908 GOTO(out, rc = 0); /* keep waiting */
913 spin_unlock(&req->rq_lock);
914 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
918 static int ptlrpc_check_status(struct ptlrpc_request *req)
923 err = lustre_msg_get_status(req->rq_repmsg);
924 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
925 struct obd_import *imp = req->rq_import;
926 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
927 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
928 " with %s. The %s operation failed with %d\n",
929 libcfs_nid2str(imp->imp_connection->c_peer.nid),
930 ll_opcode2str(opc), err);
931 RETURN(err < 0 ? err : -EINVAL);
935 DEBUG_REQ(D_INFO, req, "status is %d", err);
936 } else if (err > 0) {
937 /* XXX: translate this error from net to host */
938 DEBUG_REQ(D_INFO, req, "status is %d", err);
945 * Callback function called when client receives RPC reply for \a req.
947 static int after_reply(struct ptlrpc_request *req)
949 struct obd_import *imp = req->rq_import;
950 struct obd_device *obd = req->rq_import->imp_obd;
952 struct timeval work_start;
956 LASSERT(!req->rq_receiving_reply);
958 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
961 * NB Until this point, the whole of the incoming message,
962 * including buflens, status etc is in the sender's byte order.
965 rc = sptlrpc_cli_unwrap_reply(req);
967 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
972 * Security layer unwrap might ask resend this request.
977 rc = unpack_reply(req);
981 do_gettimeofday(&work_start);
982 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
983 if (obd->obd_svc_stats != NULL)
984 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
987 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
988 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
989 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
990 lustre_msg_get_type(req->rq_repmsg));
994 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
995 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
996 ptlrpc_at_adj_net_latency(req,
997 lustre_msg_get_service_time(req->rq_repmsg));
999 rc = ptlrpc_check_status(req);
1000 imp->imp_connect_error = rc;
1004 * Either we've been evicted, or the server has failed for
1005 * some reason. Try to reconnect, and if that fails, punt to
1008 if (ll_rpc_recoverable_error(rc)) {
1009 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1010 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1013 ptlrpc_request_handle_notconn(req);
1018 * Let's look if server sent slv. Do it only for RPC with
1021 ldlm_cli_update_pool(req);
1025 * Store transno in reqmsg for replay.
1027 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1028 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1030 if (req->rq_import->imp_replayable) {
1031 spin_lock(&imp->imp_lock);
1033 * No point in adding already-committed requests to the replay
1034 * list, we will just remove them immediately. b=9829
1036 if (req->rq_transno != 0 &&
1038 lustre_msg_get_last_committed(req->rq_repmsg) ||
1040 ptlrpc_retain_replayable_request(req, imp);
1041 else if (req->rq_commit_cb != NULL) {
1042 spin_unlock(&imp->imp_lock);
1043 req->rq_commit_cb(req);
1044 spin_lock(&imp->imp_lock);
1048 * Replay-enabled imports return commit-status information.
1050 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1051 imp->imp_peer_committed_transno =
1052 lustre_msg_get_last_committed(req->rq_repmsg);
1054 ptlrpc_free_committed(imp);
1055 spin_unlock(&imp->imp_lock);
1061 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1063 struct obd_import *imp;
1067 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1068 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1071 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1073 imp = req->rq_import;
1074 spin_lock(&imp->imp_lock);
1076 req->rq_import_generation = imp->imp_generation;
1078 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1079 spin_lock (&req->rq_lock);
1080 req->rq_waiting = 1;
1081 spin_unlock (&req->rq_lock);
1083 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1085 lustre_msg_get_status(req->rq_reqmsg) ,
1086 ptlrpc_import_state_name(req->rq_send_state),
1087 ptlrpc_import_state_name(imp->imp_state));
1088 LASSERT(list_empty (&req->rq_list));
1090 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1091 spin_unlock(&imp->imp_lock);
1096 spin_unlock(&imp->imp_lock);
1097 req->rq_status = rc;
1098 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1102 /* XXX this is the same as ptlrpc_queue_wait */
1103 LASSERT(list_empty(&req->rq_list));
1104 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1105 spin_unlock(&imp->imp_lock);
1107 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1109 rc = sptlrpc_req_refresh_ctx(req, -1);
1112 req->rq_status = rc;
1115 req->rq_wait_ctx = 1;
1120 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1121 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1122 imp->imp_obd->obd_uuid.uuid,
1123 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1124 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1125 lustre_msg_get_opc(req->rq_reqmsg));
1127 rc = ptl_send_rpc(req, 0);
1129 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1130 req->rq_net_err = 1;
1136 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1137 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1139 struct list_head *tmp;
1140 int force_timer_recalc = 0;
1143 if (set->set_remaining == 0)
1146 list_for_each(tmp, &set->set_requests) {
1147 struct ptlrpc_request *req =
1148 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1149 struct obd_import *imp = req->rq_import;
1152 if (req->rq_phase == RQ_PHASE_NEW &&
1153 ptlrpc_send_new_req(req)) {
1154 force_timer_recalc = 1;
1157 /* delayed send - skip */
1158 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1161 if (!(req->rq_phase == RQ_PHASE_RPC ||
1162 req->rq_phase == RQ_PHASE_BULK ||
1163 req->rq_phase == RQ_PHASE_INTERPRET ||
1164 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1165 req->rq_phase == RQ_PHASE_COMPLETE)) {
1166 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1170 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1171 LASSERT(req->rq_next_phase != req->rq_phase);
1172 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1175 * Skip processing until reply is unlinked. We
1176 * can't return to pool before that and we can't
1177 * call interpret before that. We need to make
1178 * sure that all rdma transfers finished and will
1179 * not corrupt any data.
1181 if (ptlrpc_client_recv_or_unlink(req))
1185 * Turn fail_loc off to prevent it from looping
1188 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_UNLINK |
1192 * Move to next phase if reply was successfully
1195 ptlrpc_rqphase_move(req, req->rq_next_phase);
1198 if (req->rq_phase == RQ_PHASE_COMPLETE)
1201 if (req->rq_phase == RQ_PHASE_INTERPRET)
1202 GOTO(interpret, req->rq_status);
1205 * Note that this also will start async reply unlink.
1207 if (req->rq_net_err && !req->rq_timedout) {
1208 ptlrpc_expire_one_request(req, 1);
1211 * Check if we still need to wait for unlink.
1213 if (ptlrpc_client_recv_or_unlink(req))
1218 req->rq_replied = 0;
1219 if (req->rq_status == 0)
1220 req->rq_status = -EIO;
1221 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1222 GOTO(interpret, req->rq_status);
1225 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1226 * will only be set after rq_timedout, but the synchronous IO
1227 * waiting path sets rq_intr irrespective of whether ptlrpcd
1228 * has seen a timeout. our policy is to only interpret
1229 * interrupted rpcs after they have timed out */
1230 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1231 req->rq_wait_ctx)) {
1232 req->rq_status = -EINTR;
1233 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1234 GOTO(interpret, req->rq_status);
1237 if (req->rq_phase == RQ_PHASE_RPC) {
1238 if (req->rq_timedout || req->rq_resend ||
1239 req->rq_waiting || req->rq_wait_ctx) {
1242 if (!ptlrpc_unregister_reply(req, 1))
1245 spin_lock(&imp->imp_lock);
1247 if (ptlrpc_import_delay_req(imp, req, &status)){
1248 spin_unlock(&imp->imp_lock);
1253 req->rq_status = status;
1254 ptlrpc_rqphase_move(req,
1255 RQ_PHASE_INTERPRET);
1256 spin_unlock(&imp->imp_lock);
1257 GOTO(interpret, req->rq_status);
1259 if (req->rq_no_resend && !req->rq_wait_ctx) {
1260 req->rq_status = -ENOTCONN;
1261 ptlrpc_rqphase_move(req,
1262 RQ_PHASE_INTERPRET);
1263 spin_unlock(&imp->imp_lock);
1264 GOTO(interpret, req->rq_status);
1267 list_del_init(&req->rq_list);
1268 list_add_tail(&req->rq_list,
1269 &imp->imp_sending_list);
1271 spin_unlock(&imp->imp_lock);
1273 req->rq_waiting = 0;
1275 if (req->rq_timedout||req->rq_resend) {
1276 /* This is re-sending anyways,
1277 * let's mark req as resend. */
1279 lustre_msg_add_flags(req->rq_reqmsg,
1282 __u64 old_xid = req->rq_xid;
1284 ptlrpc_unregister_bulk(req);
1286 /* ensure previous bulk fails */
1287 req->rq_xid = ptlrpc_next_xid();
1288 CDEBUG(D_HA, "resend bulk "
1291 old_xid, req->rq_xid);
1295 * rq_wait_ctx is only touched by ptlrpcd,
1296 * so no lock is needed here.
1298 status = sptlrpc_req_refresh_ctx(req, -1);
1301 req->rq_status = status;
1302 force_timer_recalc = 1;
1304 req->rq_wait_ctx = 1;
1309 req->rq_wait_ctx = 0;
1312 rc = ptl_send_rpc(req, 0);
1314 DEBUG_REQ(D_HA, req, "send failed (%d)",
1316 force_timer_recalc = 1;
1317 req->rq_net_err = 1;
1319 /* need to reset the timeout */
1320 force_timer_recalc = 1;
1323 spin_lock(&req->rq_lock);
1325 if (ptlrpc_client_early(req)) {
1326 ptlrpc_at_recv_early_reply(req);
1327 spin_unlock(&req->rq_lock);
1331 /* Still waiting for a reply? */
1332 if (ptlrpc_client_recv(req)) {
1333 spin_unlock(&req->rq_lock);
1337 /* Did we actually receive a reply? */
1338 if (!ptlrpc_client_replied(req)) {
1339 spin_unlock(&req->rq_lock);
1343 spin_unlock(&req->rq_lock);
1345 req->rq_status = after_reply(req);
1346 if (req->rq_resend) {
1347 /* Add this req to the delayed list so
1348 it can be errored if the import is
1349 evicted after recovery. */
1350 spin_lock(&imp->imp_lock);
1351 list_del_init(&req->rq_list);
1352 list_add_tail(&req->rq_list,
1353 &imp->imp_delayed_list);
1354 spin_unlock(&imp->imp_lock);
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_bulk_active(req->rq_bulk))
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 (req->rq_bulk != NULL)
1395 ptlrpc_unregister_bulk(req);
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 if (!list_empty(&req->rq_list))
1419 list_del_init(&req->rq_list);
1420 atomic_dec(&imp->imp_inflight);
1421 spin_unlock(&imp->imp_lock);
1423 set->set_remaining--;
1424 cfs_waitq_signal(&imp->imp_recovery_waitq);
1427 /* If we hit an error, we want to recover promptly. */
1428 RETURN(set->set_remaining == 0 || force_timer_recalc);
1431 /* Return 1 if we should give up, else 0 */
1432 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1434 struct obd_import *imp = req->rq_import;
1438 DEBUG_REQ(D_ERROR|D_NETERROR, req,
1439 "%s (sent at "CFS_TIME_T", "CFS_DURATION_T"s ago)",
1440 req->rq_net_err ? "network error" : "timeout",
1441 req->rq_sent, cfs_time_sub(cfs_time_current_sec(),
1445 LCONSOLE_WARN("Request x"LPU64" sent from %s to NID %s "
1446 CFS_DURATION_T"s ago has timed out "
1447 "(limit "CFS_DURATION_T"s).\n", req->rq_xid,
1448 req->rq_import->imp_obd->obd_name,
1449 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1450 cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1451 cfs_time_sub(req->rq_deadline, req->rq_sent));
1454 if (imp != NULL && obd_debug_peer_on_timeout)
1455 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1457 spin_lock(&req->rq_lock);
1458 req->rq_timedout = 1;
1459 spin_unlock(&req->rq_lock);
1461 ptlrpc_unregister_reply(req, async_unlink);
1463 if (obd_dump_on_timeout)
1464 libcfs_debug_dumplog();
1466 if (req->rq_bulk != NULL)
1467 ptlrpc_unregister_bulk (req);
1470 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1474 /* The DLM server doesn't want recovery run on its imports. */
1475 if (imp->imp_dlm_fake)
1478 /* If this request is for recovery or other primordial tasks,
1479 * then error it out here. */
1480 if (req->rq_ctx_init || req->rq_ctx_fini ||
1481 req->rq_send_state != LUSTRE_IMP_FULL ||
1482 imp->imp_obd->obd_no_recov) {
1483 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1484 ptlrpc_import_state_name(req->rq_send_state),
1485 ptlrpc_import_state_name(imp->imp_state));
1486 spin_lock(&req->rq_lock);
1487 req->rq_status = -ETIMEDOUT;
1489 spin_unlock(&req->rq_lock);
1493 /* if a request can't be resent we can't wait for an answer after
1495 if (req->rq_no_resend) {
1496 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1500 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1505 int ptlrpc_expired_set(void *data)
1507 struct ptlrpc_request_set *set = data;
1508 struct list_head *tmp;
1509 time_t now = cfs_time_current_sec();
1512 LASSERT(set != NULL);
1515 * A timeout expired. See which reqs it applies to...
1517 list_for_each (tmp, &set->set_requests) {
1518 struct ptlrpc_request *req =
1519 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1521 /* Request in-flight? */
1522 if (!((req->rq_phase &
1523 (RQ_PHASE_RPC | RQ_PHASE_UNREGISTERING) &&
1524 !req->rq_waiting && !req->rq_resend) ||
1525 (req->rq_phase == RQ_PHASE_BULK)))
1528 if (req->rq_timedout || /* already dealt with */
1529 req->rq_deadline > now) /* not expired */
1532 /* Deal with this guy. Do it asynchronously to not block
1533 * ptlrpcd thread. */
1534 ptlrpc_expire_one_request(req, 1);
1538 * When waiting for a whole set, we always to break out of the
1539 * sleep so we can recalculate the timeout, or enable interrupts
1540 * if everyone's timed out.
1545 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1547 spin_lock(&req->rq_lock);
1549 spin_unlock(&req->rq_lock);
1552 void ptlrpc_interrupted_set(void *data)
1554 struct ptlrpc_request_set *set = data;
1555 struct list_head *tmp;
1557 LASSERT(set != NULL);
1558 CERROR("INTERRUPTED SET %p\n", set);
1560 list_for_each(tmp, &set->set_requests) {
1561 struct ptlrpc_request *req =
1562 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1564 if (req->rq_phase != RQ_PHASE_RPC &&
1565 req->rq_phase != RQ_PHASE_UNREGISTERING)
1568 ptlrpc_mark_interrupted(req);
1573 * Get the smallest timeout in the set; this does NOT set a timeout.
1575 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1577 struct list_head *tmp;
1578 time_t now = cfs_time_current_sec();
1580 struct ptlrpc_request *req;
1584 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1586 list_for_each(tmp, &set->set_requests) {
1587 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1590 * Request in-flight?
1592 if (!(((req->rq_phase &
1593 (RQ_PHASE_RPC | RQ_PHASE_UNREGISTERING)) &&
1594 !req->rq_waiting) ||
1595 (req->rq_phase == RQ_PHASE_BULK) ||
1596 (req->rq_phase == RQ_PHASE_NEW)))
1600 * Check those waiting for long reply unlink every one
1603 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1609 * Already timed out.
1611 if (req->rq_timedout)
1617 if (req->rq_wait_ctx)
1620 if (req->rq_phase == RQ_PHASE_NEW)
1621 deadline = req->rq_sent;
1623 deadline = req->rq_sent + req->rq_timeout;
1625 if (deadline <= now) /* actually expired already */
1626 timeout = 1; /* ASAP */
1627 else if (timeout == 0 || timeout > deadline - now)
1628 timeout = deadline - now;
1633 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1635 struct list_head *tmp;
1636 struct ptlrpc_request *req;
1637 struct l_wait_info lwi;
1641 if (list_empty(&set->set_requests))
1644 list_for_each(tmp, &set->set_requests) {
1645 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1646 if (req->rq_phase == RQ_PHASE_NEW)
1647 (void)ptlrpc_send_new_req(req);
1651 timeout = ptlrpc_set_next_timeout(set);
1653 /* wait until all complete, interrupted, or an in-flight
1655 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1657 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1659 ptlrpc_interrupted_set, set);
1660 rc = l_wait_event(set->set_waitq,
1661 ptlrpc_check_set(NULL, set), &lwi);
1663 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1665 /* -EINTR => all requests have been flagged rq_intr so next
1667 * -ETIMEOUTD => someone timed out. When all reqs have
1668 * timed out, signals are enabled allowing completion with
1670 * I don't really care if we go once more round the loop in
1671 * the error cases -eeb. */
1672 } while (rc != 0 || set->set_remaining != 0);
1674 LASSERT(set->set_remaining == 0);
1677 list_for_each(tmp, &set->set_requests) {
1678 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1680 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1681 if (req->rq_status != 0)
1682 rc = req->rq_status;
1685 if (set->set_interpret != NULL) {
1686 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1688 rc = interpreter (set, set->set_arg, rc);
1690 struct ptlrpc_set_cbdata *cbdata, *n;
1693 list_for_each_entry_safe(cbdata, n,
1694 &set->set_cblist, psc_item) {
1695 list_del_init(&cbdata->psc_item);
1696 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1699 OBD_FREE_PTR(cbdata);
1706 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1709 if (request == NULL) {
1714 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1715 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1716 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1717 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1718 LASSERTF(!request->rq_replay, "req %p\n", request);
1719 LASSERT(request->rq_cli_ctx);
1721 req_capsule_fini(&request->rq_pill);
1723 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1724 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1725 if (request->rq_import != NULL) {
1727 spin_lock(&request->rq_import->imp_lock);
1728 list_del_init(&request->rq_mod_list);
1729 list_del_init(&request->rq_replay_list);
1731 spin_unlock(&request->rq_import->imp_lock);
1733 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1735 if (atomic_read(&request->rq_refcount) != 0) {
1736 DEBUG_REQ(D_ERROR, request,
1737 "freeing request with nonzero refcount");
1741 if (request->rq_repbuf != NULL)
1742 sptlrpc_cli_free_repbuf(request);
1743 if (request->rq_export != NULL) {
1744 class_export_put(request->rq_export);
1745 request->rq_export = NULL;
1747 if (request->rq_import != NULL) {
1748 class_import_put(request->rq_import);
1749 request->rq_import = NULL;
1751 if (request->rq_bulk != NULL)
1752 ptlrpc_free_bulk(request->rq_bulk);
1754 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1755 sptlrpc_cli_free_reqbuf(request);
1757 sptlrpc_req_put_ctx(request, !locked);
1759 if (request->rq_pool)
1760 __ptlrpc_free_req_to_pool(request);
1762 OBD_FREE(request, sizeof(*request));
1766 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1767 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1769 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1770 (void)__ptlrpc_req_finished(request, 1);
1773 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1776 if (request == NULL)
1779 if (request == LP_POISON ||
1780 request->rq_reqmsg == LP_POISON) {
1781 CERROR("dereferencing freed request (bug 575)\n");
1786 DEBUG_REQ(D_INFO, request, "refcount now %u",
1787 atomic_read(&request->rq_refcount) - 1);
1789 if (atomic_dec_and_test(&request->rq_refcount)) {
1790 __ptlrpc_free_req(request, locked);
1797 void ptlrpc_req_finished(struct ptlrpc_request *request)
1799 __ptlrpc_req_finished(request, 0);
1802 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1804 return request->rq_xid;
1806 EXPORT_SYMBOL(ptlrpc_req_xid);
1808 /* Disengage the client's reply buffer from the network
1809 * NB does _NOT_ unregister any client-side bulk.
1810 * IDEMPOTENT, but _not_ safe against concurrent callers.
1811 * The request owner (i.e. the thread doing the I/O) must call...
1813 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1817 struct l_wait_info lwi;
1822 LASSERT(!in_interrupt());
1825 * Let's setup deadline for reply unlink.
1827 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_UNLINK) &&
1828 async && request->rq_reply_deadline == 0)
1829 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1832 * Nothing left to do.
1834 if (!ptlrpc_client_recv_or_unlink(request))
1837 LNetMDUnlink(request->rq_reply_md_h);
1840 * Let's check it once again.
1842 if (!ptlrpc_client_recv_or_unlink(request))
1846 * Move to "Unregistering" phase as reply was not unlinked yet.
1848 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1851 * Do not wait for unlink to finish.
1857 * We have to l_wait_event() whatever the result, to give liblustre
1858 * a chance to run reply_in_callback(), and to make sure we've
1859 * unlinked before returning a req to the pool.
1861 if (request->rq_set != NULL)
1862 wq = &request->rq_set->set_waitq;
1864 wq = &request->rq_reply_waitq;
1867 /* Network access will complete in finite time but the HUGE
1868 * timeout lets us CWARN for visibility of sluggish NALs */
1869 lwi = LWI_TIMEOUT(cfs_time_seconds(LONG_UNLINK), NULL, NULL);
1870 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1873 ptlrpc_rqphase_move(request, request->rq_next_phase);
1877 LASSERT(rc == -ETIMEDOUT);
1878 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1879 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1880 request->rq_must_unlink);
1885 /* caller must hold imp->imp_lock */
1886 void ptlrpc_free_committed(struct obd_import *imp)
1888 struct list_head *tmp, *saved;
1889 struct ptlrpc_request *req;
1890 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1893 LASSERT(imp != NULL);
1895 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1898 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1899 imp->imp_generation == imp->imp_last_generation_checked) {
1900 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1901 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1905 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1906 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1907 imp->imp_generation);
1908 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1909 imp->imp_last_generation_checked = imp->imp_generation;
1911 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1912 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1914 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1915 LASSERT(req != last_req);
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 /* clear this for new requests that were resent as well
2036 as resent replayed requests. */
2037 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2039 /* don't re-add requests that have been replayed */
2040 if (!list_empty(&req->rq_replay_list))
2043 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2045 LASSERT(imp->imp_replayable);
2046 /* Balanced in ptlrpc_free_committed, usually. */
2047 ptlrpc_request_addref(req);
2048 list_for_each_prev(tmp, &imp->imp_replay_list) {
2049 struct ptlrpc_request *iter =
2050 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
2052 /* We may have duplicate transnos if we create and then
2053 * open a file, or for closes retained if to match creating
2054 * opens, so use req->rq_xid as a secondary key.
2055 * (See bugs 684, 685, and 428.)
2056 * XXX no longer needed, but all opens need transnos!
2058 if (iter->rq_transno > req->rq_transno)
2061 if (iter->rq_transno == req->rq_transno) {
2062 LASSERT(iter->rq_xid != req->rq_xid);
2063 if (iter->rq_xid > req->rq_xid)
2067 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2071 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
2074 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2078 struct l_wait_info lwi;
2079 struct obd_import *imp = req->rq_import;
2080 cfs_duration_t timeout = CFS_TICK;
2084 LASSERT(req->rq_set == NULL);
2085 LASSERT(!req->rq_receiving_reply);
2086 atomic_inc(&imp->imp_inflight);
2088 /* for distributed debugging */
2089 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2090 LASSERT(imp->imp_obd != NULL);
2091 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
2092 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2093 imp->imp_obd->obd_uuid.uuid,
2094 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2095 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2096 lustre_msg_get_opc(req->rq_reqmsg));
2098 /* Mark phase here for a little debug help */
2099 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
2101 spin_lock(&imp->imp_lock);
2102 req->rq_import_generation = imp->imp_generation;
2104 if (ptlrpc_import_delay_req(imp, req, &rc)) {
2105 list_del(&req->rq_list);
2107 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
2108 spin_unlock(&imp->imp_lock);
2110 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
2112 ptlrpc_import_state_name(req->rq_send_state),
2113 ptlrpc_import_state_name(imp->imp_state));
2114 lwi = LWI_INTR(interrupted_request, req);
2115 rc = l_wait_event(req->rq_reply_waitq,
2116 (req->rq_send_state == imp->imp_state ||
2117 req->rq_err || req->rq_intr),
2119 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2121 ptlrpc_import_state_name(imp->imp_state),
2122 ptlrpc_import_state_name(req->rq_send_state),
2123 req->rq_err, req->rq_intr);
2125 spin_lock(&imp->imp_lock);
2126 list_del_init(&req->rq_list);
2129 /* rq_status was set locally */
2132 else if (req->rq_intr) {
2135 else if (req->rq_no_resend) {
2136 spin_unlock(&imp->imp_lock);
2137 GOTO(out, rc = -ETIMEDOUT);
2145 list_del_init(&req->rq_list);
2146 spin_unlock(&imp->imp_lock);
2147 req->rq_status = rc; // XXX this ok?
2151 if (req->rq_resend) {
2152 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
2154 if (req->rq_bulk != NULL) {
2155 ptlrpc_unregister_bulk (req);
2157 /* bulk requests are supposed to be
2158 * idempotent, so we are free to bump the xid
2159 * here, which we need to do before
2160 * registering the bulk again (bug 6371).
2161 * print the old xid first for sanity.
2163 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2164 req->rq_xid = ptlrpc_next_xid();
2167 DEBUG_REQ(D_HA, req, "resending: ");
2170 /* XXX this is the same as ptlrpc_set_wait */
2171 LASSERT(list_empty(&req->rq_list));
2172 list_add_tail(&req->rq_list, &imp->imp_sending_list);
2173 spin_unlock(&imp->imp_lock);
2175 rc = sptlrpc_req_refresh_ctx(req, 0);
2178 /* we got fatal ctx refresh error, directly jump out
2179 * thus we can pass back the actual error code.
2181 spin_lock(&imp->imp_lock);
2182 list_del_init(&req->rq_list);
2183 spin_unlock(&imp->imp_lock);
2185 CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
2188 /* simulating we got error during send rpc */
2192 rc = ptl_send_rpc(req, 0);
2194 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2196 timeoutl = req->rq_deadline - cfs_time_current_sec();
2197 timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2198 cfs_time_seconds(timeoutl);
2199 DEBUG_REQ(D_NET, req,
2200 "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2201 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
2203 brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
2204 if (brc == -ETIMEDOUT && ((req->rq_deadline > cfs_time_current_sec()) ||
2205 ptlrpc_check_and_wait_suspend(req)))
2209 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2210 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
2211 imp->imp_obd->obd_uuid.uuid,
2212 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2213 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2214 lustre_msg_get_opc(req->rq_reqmsg));
2216 spin_lock(&imp->imp_lock);
2217 list_del_init(&req->rq_list);
2218 spin_unlock(&imp->imp_lock);
2220 /* If the reply was received normally, this just grabs the spinlock
2221 * (ensuring the reply callback has returned), sees that
2222 * req->rq_receiving_reply is clear and returns. */
2223 ptlrpc_unregister_reply(req, 0);
2226 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2227 rc, req->rq_status);
2228 GOTO(out, rc = rc ? rc : -EIO);
2232 /* Should only be interrupted if we timed out. */
2233 if (!req->rq_timedout)
2234 DEBUG_REQ(D_ERROR, req,
2235 "rq_intr set but rq_timedout not");
2236 GOTO(out, rc = -EINTR);
2239 /* Resend if we need to */
2240 if (req->rq_resend) {
2241 /* ...unless we were specifically told otherwise. */
2242 if (req->rq_no_resend)
2243 GOTO(out, rc = -ETIMEDOUT);
2244 spin_lock(&imp->imp_lock);
2248 if (req->rq_timedout) { /* non-recoverable timeout */
2249 GOTO(out, rc = -ETIMEDOUT);
2252 if (!ptlrpc_client_replied(req)) {
2253 /* How can this be? -eeb */
2254 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2256 GOTO(out, rc = req->rq_status);
2259 rc = after_reply(req);
2260 /* NB may return +ve success rc */
2261 if (req->rq_resend) {
2262 spin_lock(&imp->imp_lock);
2267 if (req->rq_bulk != NULL) {
2269 /* success so far. Note that anything going wrong
2270 * with bulk now, is EXTREMELY strange, since the
2271 * server must have believed that the bulk
2272 * tranferred OK before she replied with success to
2274 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2275 brc = l_wait_event(req->rq_reply_waitq,
2276 !ptlrpc_bulk_active(req->rq_bulk),
2278 LASSERT(brc == 0 || brc == -ETIMEDOUT);
2280 LASSERT(brc == -ETIMEDOUT);
2281 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2283 } else if (!req->rq_bulk->bd_success) {
2284 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2289 ptlrpc_unregister_bulk (req);
2292 LASSERT(!req->rq_receiving_reply);
2293 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2295 atomic_dec(&imp->imp_inflight);
2296 cfs_waitq_signal(&imp->imp_recovery_waitq);
2300 struct ptlrpc_replay_async_args {
2302 int praa_old_status;
2305 static int ptlrpc_replay_interpret(const struct lu_env *env,
2306 struct ptlrpc_request *req,
2307 void * data, int rc)
2309 struct ptlrpc_replay_async_args *aa = data;
2310 struct obd_import *imp = req->rq_import;
2313 atomic_dec(&imp->imp_replay_inflight);
2315 if (!ptlrpc_client_replied(req)) {
2316 CERROR("request replay timed out, restarting recovery\n");
2317 GOTO(out, rc = -ETIMEDOUT);
2320 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2321 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2322 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2323 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2325 /* The transno had better not change over replay. */
2326 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
2327 lustre_msg_get_transno(req->rq_repmsg));
2329 DEBUG_REQ(D_HA, req, "got rep");
2331 /* let the callback do fixups, possibly including in the request */
2332 if (req->rq_replay_cb)
2333 req->rq_replay_cb(req);
2335 if (ptlrpc_client_replied(req) &&
2336 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2337 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2338 lustre_msg_get_status(req->rq_repmsg),
2339 aa->praa_old_status);
2341 /* Put it back for re-replay. */
2342 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2346 * Errors while replay can set transno to 0, but
2347 * imp_last_replay_transno shouldn't be set to 0 anyway
2349 if (req->rq_transno > 0) {
2350 spin_lock(&imp->imp_lock);
2351 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2352 imp->imp_last_replay_transno = req->rq_transno;
2353 spin_unlock(&imp->imp_lock);
2355 CERROR("Transno is 0 during replay!\n");
2356 /* continue with recovery */
2357 rc = ptlrpc_import_recovery_state_machine(imp);
2359 req->rq_send_state = aa->praa_old_state;
2362 /* this replay failed, so restart recovery */
2363 ptlrpc_connect_import(imp, NULL);
2368 int ptlrpc_replay_req(struct ptlrpc_request *req)
2370 struct ptlrpc_replay_async_args *aa;
2373 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2374 /* Not handling automatic bulk replay yet (or ever?) */
2375 LASSERT(req->rq_bulk == NULL);
2377 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2378 aa = ptlrpc_req_async_args(req);
2379 memset(aa, 0, sizeof *aa);
2381 /* Prepare request to be resent with ptlrpcd */
2382 aa->praa_old_state = req->rq_send_state;
2383 req->rq_send_state = LUSTRE_IMP_REPLAY;
2384 req->rq_phase = RQ_PHASE_NEW;
2385 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2387 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2389 req->rq_interpret_reply = ptlrpc_replay_interpret;
2390 /* Readjust the timeout for current conditions */
2391 ptlrpc_at_set_req_timeout(req);
2393 DEBUG_REQ(D_HA, req, "REPLAY");
2395 atomic_inc(&req->rq_import->imp_replay_inflight);
2396 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2398 ptlrpcd_add_req(req, PSCOPE_OTHER);
2402 void ptlrpc_abort_inflight(struct obd_import *imp)
2404 struct list_head *tmp, *n;
2407 /* Make sure that no new requests get processed for this import.
2408 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2409 * this flag and then putting requests on sending_list or delayed_list.
2411 spin_lock(&imp->imp_lock);
2413 /* XXX locking? Maybe we should remove each request with the list
2414 * locked? Also, how do we know if the requests on the list are
2415 * being freed at this time?
2417 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2418 struct ptlrpc_request *req =
2419 list_entry(tmp, struct ptlrpc_request, rq_list);
2421 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2423 spin_lock (&req->rq_lock);
2424 if (req->rq_import_generation < imp->imp_generation) {
2426 req->rq_status = -EINTR;
2427 ptlrpc_client_wake_req(req);
2429 spin_unlock (&req->rq_lock);
2432 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2433 struct ptlrpc_request *req =
2434 list_entry(tmp, struct ptlrpc_request, rq_list);
2436 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2438 spin_lock (&req->rq_lock);
2439 if (req->rq_import_generation < imp->imp_generation) {
2441 req->rq_status = -EINTR;
2442 ptlrpc_client_wake_req(req);
2444 spin_unlock (&req->rq_lock);
2447 /* Last chance to free reqs left on the replay list, but we
2448 * will still leak reqs that haven't committed. */
2449 if (imp->imp_replayable)
2450 ptlrpc_free_committed(imp);
2452 spin_unlock(&imp->imp_lock);
2457 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2459 struct list_head *tmp, *n;
2461 LASSERT(set != NULL);
2463 list_for_each_safe(tmp, n, &set->set_requests) {
2464 struct ptlrpc_request *req =
2465 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2467 spin_lock (&req->rq_lock);
2468 if (req->rq_phase != RQ_PHASE_RPC) {
2469 spin_unlock (&req->rq_lock);
2474 req->rq_status = -EINTR;
2475 ptlrpc_client_wake_req(req);
2476 spin_unlock (&req->rq_lock);
2480 static __u64 ptlrpc_last_xid;
2481 static spinlock_t ptlrpc_last_xid_lock;
2483 /* Initialize the XID for the node. This is common among all requests on
2484 * this node, and only requires the property that it is monotonically
2485 * increasing. It does not need to be sequential. Since this is also used
2486 * as the RDMA match bits, it is important that a single client NOT have
2487 * the same match bits for two different in-flight requests, hence we do
2488 * NOT want to have an XID per target or similar.
2490 * To avoid an unlikely collision between match bits after a client reboot
2491 * (which would cause old to be delivered into the wrong buffer) we initialize
2492 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2493 * If the time is clearly incorrect, we instead use a 62-bit random number.
2494 * In the worst case the random number will overflow 1M RPCs per second in
2495 * 9133 years, or permutations thereof.
2497 #define YEAR_2004 (1ULL << 30)
2498 void ptlrpc_init_xid(void)
2500 time_t now = cfs_time_current_sec();
2502 spin_lock_init(&ptlrpc_last_xid_lock);
2503 if (now < YEAR_2004) {
2504 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2505 ptlrpc_last_xid >>= 2;
2506 ptlrpc_last_xid |= (1ULL << 61);
2508 ptlrpc_last_xid = (now << 20);
2512 __u64 ptlrpc_next_xid(void)
2515 spin_lock(&ptlrpc_last_xid_lock);
2516 tmp = ++ptlrpc_last_xid;
2517 spin_unlock(&ptlrpc_last_xid_lock);
2521 __u64 ptlrpc_sample_next_xid(void)
2523 #if BITS_PER_LONG == 32
2524 /* need to avoid possible word tearing on 32-bit systems */
2526 spin_lock(&ptlrpc_last_xid_lock);
2527 tmp = ptlrpc_last_xid + 1;
2528 spin_unlock(&ptlrpc_last_xid_lock);
2531 /* No need to lock, since returned value is racy anyways */
2532 return ptlrpc_last_xid + 1;
2535 EXPORT_SYMBOL(ptlrpc_sample_next_xid);