1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #define DEBUG_SUBSYSTEM S_RPC
41 #include <liblustre.h>
44 #include <obd_support.h>
45 #include <obd_class.h>
46 #include <lustre_lib.h>
47 #include <lustre_ha.h>
48 #include <lustre_import.h>
49 #include <lustre_req_layout.h>
51 #include "ptlrpc_internal.h"
53 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
54 struct ptlrpc_client *cl)
56 cl->cli_request_portal = req_portal;
57 cl->cli_reply_portal = rep_portal;
61 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
63 struct ptlrpc_connection *c;
65 lnet_process_id_t peer;
68 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
70 CERROR("cannot find peer %s!\n", uuid->uuid);
74 c = ptlrpc_connection_get(peer, self, uuid);
76 memcpy(c->c_remote_uuid.uuid,
77 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
80 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
85 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
87 struct ptlrpc_bulk_desc *desc;
89 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
93 cfs_spin_lock_init(&desc->bd_lock);
94 cfs_waitq_init(&desc->bd_waitq);
95 desc->bd_max_iov = npages;
96 desc->bd_iov_count = 0;
97 LNetInvalidateHandle(&desc->bd_md_h);
98 desc->bd_portal = portal;
104 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
105 int npages, int type, int portal)
107 struct obd_import *imp = req->rq_import;
108 struct ptlrpc_bulk_desc *desc;
111 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
112 desc = new_bulk(npages, type, portal);
116 desc->bd_import_generation = req->rq_import_generation;
117 desc->bd_import = class_import_get(imp);
120 desc->bd_cbid.cbid_fn = client_bulk_callback;
121 desc->bd_cbid.cbid_arg = desc;
123 /* This makes req own desc, and free it when she frees herself */
129 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
130 int npages, int type, int portal)
132 struct obd_export *exp = req->rq_export;
133 struct ptlrpc_bulk_desc *desc;
136 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
138 desc = new_bulk(npages, type, portal);
142 desc->bd_export = class_export_get(exp);
145 desc->bd_cbid.cbid_fn = server_bulk_callback;
146 desc->bd_cbid.cbid_arg = desc;
148 /* NB we don't assign rq_bulk here; server-side requests are
149 * re-used, and the handler frees the bulk desc explicitly. */
154 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
155 cfs_page_t *page, int pageoffset, int len)
157 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
158 LASSERT(page != NULL);
159 LASSERT(pageoffset >= 0);
161 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
165 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
168 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
172 LASSERT(desc != NULL);
173 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
174 LASSERT(!desc->bd_network_rw); /* network hands off or */
175 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
177 sptlrpc_enc_pool_put_pages(desc);
180 class_export_put(desc->bd_export);
182 class_import_put(desc->bd_import);
184 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
185 bd_iov[desc->bd_max_iov]));
189 /* Set server timelimit for this req */
190 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
196 LASSERT(req->rq_import);
199 /* non-AT settings */
200 req->rq_timeout = req->rq_import->imp_server_timeout ?
201 obd_timeout / 2 : obd_timeout;
203 at = &req->rq_import->imp_at;
204 idx = import_at_get_index(req->rq_import,
205 req->rq_request_portal);
206 serv_est = at_get(&at->iat_service_estimate[idx]);
207 req->rq_timeout = at_est2timeout(serv_est);
209 /* We could get even fancier here, using history to predict increased
212 /* Let the server know what this RPC timeout is by putting it in the
214 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
217 /* Adjust max service estimate based on server value */
218 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
219 unsigned int serv_est)
225 LASSERT(req->rq_import);
226 at = &req->rq_import->imp_at;
228 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
229 /* max service estimates are tracked on the server side,
230 so just keep minimal history here */
231 oldse = at_add(&at->iat_service_estimate[idx], serv_est);
233 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
234 "has changed from %d to %d\n",
235 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
236 oldse, at_get(&at->iat_service_estimate[idx]));
239 /* Expected network latency per remote node (secs) */
240 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
242 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
245 /* Adjust expected network latency */
246 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
247 unsigned int service_time)
249 unsigned int nl, oldnl;
251 time_t now = cfs_time_current_sec();
253 LASSERT(req->rq_import);
254 at = &req->rq_import->imp_at;
256 /* Network latency is total time less server processing time */
257 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
258 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
259 CWARN("Reported service time %u > total measured time "
260 CFS_DURATION_T"\n", service_time,
261 cfs_time_sub(now, req->rq_sent));
263 oldnl = at_add(&at->iat_net_latency, nl);
265 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
266 "has changed from %d to %d\n",
267 req->rq_import->imp_obd->obd_name,
269 &req->rq_import->imp_connection->c_remote_uuid),
270 oldnl, at_get(&at->iat_net_latency));
273 static int unpack_reply(struct ptlrpc_request *req)
277 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
278 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
280 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
285 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
287 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
294 * Handle an early reply message, called with the rq_lock held.
295 * If anything goes wrong just ignore it - same as if it never happened
297 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
299 struct ptlrpc_request *early_req;
305 cfs_spin_unlock(&req->rq_lock);
307 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
309 cfs_spin_lock(&req->rq_lock);
313 rc = unpack_reply(early_req);
315 /* Expecting to increase the service time estimate here */
316 ptlrpc_at_adj_service(req,
317 lustre_msg_get_timeout(early_req->rq_repmsg));
318 ptlrpc_at_adj_net_latency(req,
319 lustre_msg_get_service_time(early_req->rq_repmsg));
322 sptlrpc_cli_finish_early_reply(early_req);
324 cfs_spin_lock(&req->rq_lock);
327 /* Adjust the local timeout for this req */
328 ptlrpc_at_set_req_timeout(req);
330 olddl = req->rq_deadline;
331 /* server assumes it now has rq_timeout from when it sent the
332 early reply, so client should give it at least that long. */
333 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
334 ptlrpc_at_get_net_latency(req);
336 DEBUG_REQ(D_ADAPTTO, req,
337 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
338 "("CFS_DURATION_T"s)", req->rq_early_count,
339 cfs_time_sub(req->rq_deadline,
340 cfs_time_current_sec()),
341 cfs_time_sub(req->rq_deadline, olddl));
347 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
350 struct ptlrpc_request *req;
352 LASSERT(pool != NULL);
354 cfs_spin_lock(&pool->prp_lock);
355 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
356 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
357 cfs_list_del(&req->rq_list);
358 LASSERT(req->rq_reqbuf);
359 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
360 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
361 OBD_FREE(req, sizeof(*req));
363 cfs_spin_unlock(&pool->prp_lock);
364 OBD_FREE(pool, sizeof(*pool));
367 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
372 while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
375 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
376 size == pool->prp_rq_size,
377 "Trying to change pool size with nonempty pool "
378 "from %d to %d bytes\n", pool->prp_rq_size, size);
380 cfs_spin_lock(&pool->prp_lock);
381 pool->prp_rq_size = size;
382 for (i = 0; i < num_rq; i++) {
383 struct ptlrpc_request *req;
384 struct lustre_msg *msg;
386 cfs_spin_unlock(&pool->prp_lock);
387 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
390 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
392 OBD_FREE(req, sizeof(struct ptlrpc_request));
395 req->rq_reqbuf = msg;
396 req->rq_reqbuf_len = size;
398 cfs_spin_lock(&pool->prp_lock);
399 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
401 cfs_spin_unlock(&pool->prp_lock);
405 struct ptlrpc_request_pool *
406 ptlrpc_init_rq_pool(int num_rq, int msgsize,
407 void (*populate_pool)(struct ptlrpc_request_pool *, int))
409 struct ptlrpc_request_pool *pool;
411 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
415 /* Request next power of two for the allocation, because internally
416 kernel would do exactly this */
418 cfs_spin_lock_init(&pool->prp_lock);
419 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
420 pool->prp_rq_size = msgsize;
421 pool->prp_populate = populate_pool;
423 populate_pool(pool, num_rq);
425 if (cfs_list_empty(&pool->prp_req_list)) {
426 /* have not allocated a single request for the pool */
427 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
433 static struct ptlrpc_request *
434 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
436 struct ptlrpc_request *request;
437 struct lustre_msg *reqbuf;
442 cfs_spin_lock(&pool->prp_lock);
444 /* See if we have anything in a pool, and bail out if nothing,
445 * in writeout path, where this matters, this is safe to do, because
446 * nothing is lost in this case, and when some in-flight requests
447 * complete, this code will be called again. */
448 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
449 cfs_spin_unlock(&pool->prp_lock);
453 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
455 cfs_list_del_init(&request->rq_list);
456 cfs_spin_unlock(&pool->prp_lock);
458 LASSERT(request->rq_reqbuf);
459 LASSERT(request->rq_pool);
461 reqbuf = request->rq_reqbuf;
462 memset(request, 0, sizeof(*request));
463 request->rq_reqbuf = reqbuf;
464 request->rq_reqbuf_len = pool->prp_rq_size;
465 request->rq_pool = pool;
470 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
472 struct ptlrpc_request_pool *pool = request->rq_pool;
474 cfs_spin_lock(&pool->prp_lock);
475 LASSERT(cfs_list_empty(&request->rq_list));
476 LASSERT(!request->rq_receiving_reply);
477 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
478 cfs_spin_unlock(&pool->prp_lock);
481 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
482 __u32 version, int opcode,
483 int count, __u32 *lengths, char **bufs,
484 struct ptlrpc_cli_ctx *ctx)
486 struct obd_import *imp = request->rq_import;
491 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
493 rc = sptlrpc_req_get_ctx(request);
498 sptlrpc_req_set_flavor(request, opcode);
500 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
503 LASSERT(!request->rq_pool);
507 lustre_msg_add_version(request->rq_reqmsg, version);
508 request->rq_send_state = LUSTRE_IMP_FULL;
509 request->rq_type = PTL_RPC_MSG_REQUEST;
510 request->rq_export = NULL;
512 request->rq_req_cbid.cbid_fn = request_out_callback;
513 request->rq_req_cbid.cbid_arg = request;
515 request->rq_reply_cbid.cbid_fn = reply_in_callback;
516 request->rq_reply_cbid.cbid_arg = request;
518 request->rq_reply_deadline = 0;
519 request->rq_phase = RQ_PHASE_NEW;
520 request->rq_next_phase = RQ_PHASE_UNDEFINED;
522 request->rq_request_portal = imp->imp_client->cli_request_portal;
523 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
525 ptlrpc_at_set_req_timeout(request);
527 cfs_spin_lock_init(&request->rq_lock);
528 CFS_INIT_LIST_HEAD(&request->rq_list);
529 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
530 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
531 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
532 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
533 CFS_INIT_LIST_HEAD(&request->rq_history_list);
534 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
535 cfs_waitq_init(&request->rq_reply_waitq);
536 request->rq_xid = ptlrpc_next_xid();
537 cfs_atomic_set(&request->rq_refcount, 1);
539 lustre_msg_set_opc(request->rq_reqmsg, opcode);
543 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
545 class_import_put(imp);
549 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
550 __u32 version, int opcode, char **bufs,
551 struct ptlrpc_cli_ctx *ctx)
555 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
556 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
557 request->rq_pill.rc_area[RCL_CLIENT],
560 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
562 int ptlrpc_request_pack(struct ptlrpc_request *request,
563 __u32 version, int opcode)
565 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
569 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
570 struct ptlrpc_request_pool *pool)
572 struct ptlrpc_request *request = NULL;
575 request = ptlrpc_prep_req_from_pool(pool);
578 OBD_ALLOC_PTR(request);
581 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
582 LASSERT(imp != LP_POISON);
583 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
585 LASSERT(imp->imp_client != LP_POISON);
587 request->rq_import = class_import_get(imp);
589 CERROR("request allocation out of memory\n");
595 static struct ptlrpc_request *
596 ptlrpc_request_alloc_internal(struct obd_import *imp,
597 struct ptlrpc_request_pool * pool,
598 const struct req_format *format)
600 struct ptlrpc_request *request;
602 request = __ptlrpc_request_alloc(imp, pool);
606 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
607 req_capsule_set(&request->rq_pill, format);
611 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
612 const struct req_format *format)
614 return ptlrpc_request_alloc_internal(imp, NULL, format);
617 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
618 struct ptlrpc_request_pool * pool,
619 const struct req_format *format)
621 return ptlrpc_request_alloc_internal(imp, pool, format);
624 void ptlrpc_request_free(struct ptlrpc_request *request)
626 if (request->rq_pool)
627 __ptlrpc_free_req_to_pool(request);
629 OBD_FREE_PTR(request);
632 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
633 const struct req_format *format,
634 __u32 version, int opcode)
636 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
640 rc = ptlrpc_request_pack(req, version, opcode);
642 ptlrpc_request_free(req);
649 struct ptlrpc_request *
650 ptlrpc_prep_req_pool(struct obd_import *imp,
651 __u32 version, int opcode,
652 int count, __u32 *lengths, char **bufs,
653 struct ptlrpc_request_pool *pool)
655 struct ptlrpc_request *request;
658 request = __ptlrpc_request_alloc(imp, pool);
662 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
663 lengths, bufs, NULL);
665 ptlrpc_request_free(request);
671 struct ptlrpc_request *
672 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
673 __u32 *lengths, char **bufs)
675 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
679 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
680 unsigned int timeout,
681 ptlrpc_interpterer_t interpreter)
683 struct ptlrpc_request *request = NULL;
686 OBD_ALLOC(request, sizeof(*request));
688 CERROR("request allocation out of memory\n");
692 request->rq_send_state = LUSTRE_IMP_FULL;
693 request->rq_type = PTL_RPC_MSG_REQUEST;
694 request->rq_import = class_import_get(imp);
695 request->rq_export = NULL;
696 request->rq_import_generation = imp->imp_generation;
698 request->rq_timeout = timeout;
699 request->rq_sent = cfs_time_current_sec();
700 request->rq_deadline = request->rq_sent + timeout;
701 request->rq_reply_deadline = request->rq_deadline;
702 request->rq_interpret_reply = interpreter;
703 request->rq_phase = RQ_PHASE_RPC;
704 request->rq_next_phase = RQ_PHASE_INTERPRET;
705 /* don't want reply */
706 request->rq_receiving_reply = 0;
707 request->rq_must_unlink = 0;
708 request->rq_no_delay = request->rq_no_resend = 1;
709 request->rq_fake = 1;
711 cfs_spin_lock_init(&request->rq_lock);
712 CFS_INIT_LIST_HEAD(&request->rq_list);
713 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
714 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
715 CFS_INIT_LIST_HEAD(&request->rq_history_list);
716 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
717 cfs_waitq_init(&request->rq_reply_waitq);
719 request->rq_xid = ptlrpc_next_xid();
720 cfs_atomic_set(&request->rq_refcount, 1);
725 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
727 /* if we kill request before timeout - need adjust counter */
728 if (req->rq_phase == RQ_PHASE_RPC) {
729 struct ptlrpc_request_set *set = req->rq_set;
732 set->set_remaining --;
735 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
736 cfs_list_del_init(&req->rq_list);
740 struct ptlrpc_request_set *ptlrpc_prep_set(void)
742 struct ptlrpc_request_set *set;
745 OBD_ALLOC(set, sizeof *set);
748 CFS_INIT_LIST_HEAD(&set->set_requests);
749 cfs_waitq_init(&set->set_waitq);
750 set->set_remaining = 0;
751 cfs_spin_lock_init(&set->set_new_req_lock);
752 CFS_INIT_LIST_HEAD(&set->set_new_requests);
753 CFS_INIT_LIST_HEAD(&set->set_cblist);
758 /* Finish with this set; opposite of prep_set. */
759 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
767 /* Requests on the set should either all be completed, or all be new */
768 expected_phase = (set->set_remaining == 0) ?
769 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
770 cfs_list_for_each (tmp, &set->set_requests) {
771 struct ptlrpc_request *req =
772 cfs_list_entry(tmp, struct ptlrpc_request,
775 LASSERT(req->rq_phase == expected_phase);
779 LASSERTF(set->set_remaining == 0 || set->set_remaining == n, "%d / %d\n",
780 set->set_remaining, n);
782 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
783 struct ptlrpc_request *req =
784 cfs_list_entry(tmp, struct ptlrpc_request,
786 cfs_list_del_init(&req->rq_set_chain);
788 LASSERT(req->rq_phase == expected_phase);
790 if (req->rq_phase == RQ_PHASE_NEW) {
791 ptlrpc_req_interpret(NULL, req, -EBADR);
792 set->set_remaining--;
796 ptlrpc_req_finished (req);
799 LASSERT(set->set_remaining == 0);
801 OBD_FREE(set, sizeof(*set));
805 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
806 set_interpreter_func fn, void *data)
808 struct ptlrpc_set_cbdata *cbdata;
810 OBD_ALLOC_PTR(cbdata);
814 cbdata->psc_interpret = fn;
815 cbdata->psc_data = data;
816 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
821 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
822 struct ptlrpc_request *req)
824 /* The set takes over the caller's request reference */
825 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
827 set->set_remaining++;
831 * Lock so many callers can add things, the context that owns the set
832 * is supposed to notice these and move them into the set proper.
834 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
835 struct ptlrpc_request *req)
837 struct ptlrpc_request_set *set = pc->pc_set;
840 * Let caller know that we stopped and will not handle this request.
841 * It needs to take care itself of request.
843 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags))
846 cfs_spin_lock(&set->set_new_req_lock);
848 * The set takes over the caller's request reference.
850 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
852 cfs_spin_unlock(&set->set_new_req_lock);
854 cfs_waitq_signal(&set->set_waitq);
859 * Based on the current state of the import, determine if the request
860 * can be sent, is an error, or should be delayed.
862 * Returns true if this request should be delayed. If false, and
863 * *status is set, then the request can not be sent and *status is the
864 * error code. If false and status is 0, then request can be sent.
866 * The imp->imp_lock must be held.
868 static int ptlrpc_import_delay_req(struct obd_import *imp,
869 struct ptlrpc_request *req, int *status)
874 LASSERT (status != NULL);
877 if (req->rq_ctx_init || req->rq_ctx_fini) {
878 /* always allow ctx init/fini rpc go through */
879 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
880 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
883 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
884 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
886 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
887 imp->imp_state == LUSTRE_IMP_CONNECTING) {
888 /* allow CONNECT even if import is invalid */ ;
889 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
890 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
893 } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
894 imp->imp_obd->obd_no_recov) {
895 /* If the import has been invalidated (such as by an OST
896 * failure), and if the import(MGC) tried all of its connection
897 * list (Bug 13464), the request must fail with -ESHUTDOWN.
898 * This indicates the requests should be discarded; an -EIO
899 * may result in a resend of the request. */
900 if (!imp->imp_deactive)
901 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
902 *status = -ESHUTDOWN; /* bz 12940 */
903 } else if (req->rq_import_generation != imp->imp_generation) {
904 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
906 } else if (req->rq_send_state != imp->imp_state) {
907 /* invalidate in progress - any requests should be drop */
908 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
909 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
911 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
912 *status = -EWOULDBLOCK;
921 /* Conditionally suppress specific console messages */
922 static int ptlrpc_console_allow(struct ptlrpc_request *req)
924 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
927 /* Suppress particular reconnect errors which are to be expected. No
928 * errors are suppressed for the initial connection on an import */
929 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
930 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
932 /* Suppress timed out reconnect requests */
933 if (req->rq_timedout)
936 /* Suppress unavailable/again reconnect requests */
937 err = lustre_msg_get_status(req->rq_repmsg);
938 if (err == -ENODEV || err == -EAGAIN)
945 static int ptlrpc_check_status(struct ptlrpc_request *req)
950 err = lustre_msg_get_status(req->rq_repmsg);
951 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
952 struct obd_import *imp = req->rq_import;
953 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
954 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
955 " with %s. The %s operation failed with %d\n",
956 libcfs_nid2str(imp->imp_connection->c_peer.nid),
957 ll_opcode2str(opc), err);
958 RETURN(err < 0 ? err : -EINVAL);
962 DEBUG_REQ(D_INFO, req, "status is %d", err);
963 } else if (err > 0) {
964 /* XXX: translate this error from net to host */
965 DEBUG_REQ(D_INFO, req, "status is %d", err);
968 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
969 struct obd_import *imp = req->rq_import;
970 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
972 if (ptlrpc_console_allow(req))
973 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
974 "communicating with %s. The %s "
975 "operation failed with %d\n",
977 imp->imp_connection->c_peer.nid),
978 ll_opcode2str(opc), err);
980 RETURN(err < 0 ? err : -EINVAL);
987 * save pre-versions for replay
989 static void ptlrpc_save_versions(struct ptlrpc_request *req)
991 struct lustre_msg *repmsg = req->rq_repmsg;
992 struct lustre_msg *reqmsg = req->rq_reqmsg;
993 __u64 *versions = lustre_msg_get_versions(repmsg);
996 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1000 lustre_msg_set_versions(reqmsg, versions);
1001 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1002 versions[0], versions[1]);
1008 * Callback function called when client receives RPC reply for \a req.
1010 static int after_reply(struct ptlrpc_request *req)
1012 struct obd_import *imp = req->rq_import;
1013 struct obd_device *obd = req->rq_import->imp_obd;
1015 struct timeval work_start;
1019 LASSERT(obd != NULL);
1020 /* repbuf must be unlinked */
1021 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1023 if (req->rq_reply_truncate) {
1024 if (req->rq_no_resend) {
1025 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1026 " expected: %d, actual size: %d",
1027 req->rq_nob_received, req->rq_repbuf_len);
1031 sptlrpc_cli_free_repbuf(req);
1032 /* Pass the required reply buffer size (include
1033 * space for early reply).
1034 * NB: no need to roundup because alloc_repbuf
1035 * will roundup it */
1036 req->rq_replen = req->rq_nob_received;
1037 req->rq_nob_received = 0;
1043 * NB Until this point, the whole of the incoming message,
1044 * including buflens, status etc is in the sender's byte order.
1046 rc = sptlrpc_cli_unwrap_reply(req);
1048 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1053 * Security layer unwrap might ask resend this request.
1058 rc = unpack_reply(req);
1062 cfs_gettimeofday(&work_start);
1063 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1064 if (obd->obd_svc_stats != NULL) {
1065 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1067 ptlrpc_lprocfs_rpc_sent(req, timediff);
1070 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1071 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1072 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1073 lustre_msg_get_type(req->rq_repmsg));
1077 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1078 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1079 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1080 ptlrpc_at_adj_net_latency(req,
1081 lustre_msg_get_service_time(req->rq_repmsg));
1083 rc = ptlrpc_check_status(req);
1084 imp->imp_connect_error = rc;
1088 * Either we've been evicted, or the server has failed for
1089 * some reason. Try to reconnect, and if that fails, punt to
1092 if (ll_rpc_recoverable_error(rc)) {
1093 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1094 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1097 ptlrpc_request_handle_notconn(req);
1102 * Let's look if server sent slv. Do it only for RPC with
1105 ldlm_cli_update_pool(req);
1109 * Store transno in reqmsg for replay.
1111 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1112 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1113 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1116 if (imp->imp_replayable) {
1117 cfs_spin_lock(&imp->imp_lock);
1119 * No point in adding already-committed requests to the replay
1120 * list, we will just remove them immediately. b=9829
1122 if (req->rq_transno != 0 &&
1124 lustre_msg_get_last_committed(req->rq_repmsg) ||
1126 /** version recovery */
1127 ptlrpc_save_versions(req);
1128 ptlrpc_retain_replayable_request(req, imp);
1129 } else if (req->rq_commit_cb != NULL) {
1130 cfs_spin_unlock(&imp->imp_lock);
1131 req->rq_commit_cb(req);
1132 cfs_spin_lock(&imp->imp_lock);
1136 * Replay-enabled imports return commit-status information.
1138 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1139 imp->imp_peer_committed_transno =
1140 lustre_msg_get_last_committed(req->rq_repmsg);
1142 ptlrpc_free_committed(imp);
1143 cfs_spin_unlock(&imp->imp_lock);
1149 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1151 struct obd_import *imp;
1155 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1156 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1159 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1161 imp = req->rq_import;
1162 cfs_spin_lock(&imp->imp_lock);
1164 req->rq_import_generation = imp->imp_generation;
1166 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1167 cfs_spin_lock(&req->rq_lock);
1168 req->rq_waiting = 1;
1169 cfs_spin_unlock(&req->rq_lock);
1171 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1172 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1173 ptlrpc_import_state_name(req->rq_send_state),
1174 ptlrpc_import_state_name(imp->imp_state));
1175 LASSERT(cfs_list_empty(&req->rq_list));
1176 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1177 cfs_atomic_inc(&req->rq_import->imp_inflight);
1178 cfs_spin_unlock(&imp->imp_lock);
1183 cfs_spin_unlock(&imp->imp_lock);
1184 req->rq_status = rc;
1185 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1189 LASSERT(cfs_list_empty(&req->rq_list));
1190 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1191 cfs_atomic_inc(&req->rq_import->imp_inflight);
1192 cfs_spin_unlock(&imp->imp_lock);
1194 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1196 rc = sptlrpc_req_refresh_ctx(req, -1);
1199 req->rq_status = rc;
1202 req->rq_wait_ctx = 1;
1207 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1208 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1209 imp->imp_obd->obd_uuid.uuid,
1210 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1211 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1212 lustre_msg_get_opc(req->rq_reqmsg));
1214 rc = ptl_send_rpc(req, 0);
1216 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1217 req->rq_net_err = 1;
1223 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1224 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1227 int force_timer_recalc = 0;
1230 if (set->set_remaining == 0)
1233 cfs_list_for_each(tmp, &set->set_requests) {
1234 struct ptlrpc_request *req =
1235 cfs_list_entry(tmp, struct ptlrpc_request,
1237 struct obd_import *imp = req->rq_import;
1238 int unregistered = 0;
1241 if (req->rq_phase == RQ_PHASE_NEW &&
1242 ptlrpc_send_new_req(req)) {
1243 force_timer_recalc = 1;
1246 /* delayed send - skip */
1247 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1250 if (!(req->rq_phase == RQ_PHASE_RPC ||
1251 req->rq_phase == RQ_PHASE_BULK ||
1252 req->rq_phase == RQ_PHASE_INTERPRET ||
1253 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1254 req->rq_phase == RQ_PHASE_COMPLETE)) {
1255 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1259 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1260 LASSERT(req->rq_next_phase != req->rq_phase);
1261 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1264 * Skip processing until reply is unlinked. We
1265 * can't return to pool before that and we can't
1266 * call interpret before that. We need to make
1267 * sure that all rdma transfers finished and will
1268 * not corrupt any data.
1270 if (ptlrpc_client_recv_or_unlink(req) ||
1271 ptlrpc_client_bulk_active(req))
1275 * Turn fail_loc off to prevent it from looping
1278 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1279 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1282 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1283 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1288 * Move to next phase if reply was successfully
1291 ptlrpc_rqphase_move(req, req->rq_next_phase);
1294 if (req->rq_phase == RQ_PHASE_COMPLETE)
1297 if (req->rq_phase == RQ_PHASE_INTERPRET)
1298 GOTO(interpret, req->rq_status);
1301 * Note that this also will start async reply unlink.
1303 if (req->rq_net_err && !req->rq_timedout) {
1304 ptlrpc_expire_one_request(req, 1);
1307 * Check if we still need to wait for unlink.
1309 if (ptlrpc_client_recv_or_unlink(req) ||
1310 ptlrpc_client_bulk_active(req))
1315 req->rq_replied = 0;
1316 if (req->rq_status == 0)
1317 req->rq_status = -EIO;
1318 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1319 GOTO(interpret, req->rq_status);
1322 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1323 * so it sets rq_intr regardless of individual rpc
1324 * timeouts. The synchronous IO waiting path sets
1325 * rq_intr irrespective of whether ptlrpcd
1326 * has seen a timeout. Our policy is to only interpret
1327 * interrupted rpcs after they have timed out, so we
1328 * need to enforce that here.
1331 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1332 req->rq_wait_ctx)) {
1333 req->rq_status = -EINTR;
1334 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1335 GOTO(interpret, req->rq_status);
1338 if (req->rq_phase == RQ_PHASE_RPC) {
1339 if (req->rq_timedout || req->rq_resend ||
1340 req->rq_waiting || req->rq_wait_ctx) {
1343 if (!ptlrpc_unregister_reply(req, 1))
1346 cfs_spin_lock(&imp->imp_lock);
1347 if (ptlrpc_import_delay_req(imp, req, &status)){
1348 /* put on delay list - only if we wait
1349 * recovery finished - before send */
1350 cfs_list_del_init(&req->rq_list);
1351 cfs_list_add_tail(&req->rq_list,
1354 cfs_spin_unlock(&imp->imp_lock);
1359 req->rq_status = status;
1360 ptlrpc_rqphase_move(req,
1361 RQ_PHASE_INTERPRET);
1362 cfs_spin_unlock(&imp->imp_lock);
1363 GOTO(interpret, req->rq_status);
1365 if (req->rq_no_resend && !req->rq_wait_ctx) {
1366 req->rq_status = -ENOTCONN;
1367 ptlrpc_rqphase_move(req,
1368 RQ_PHASE_INTERPRET);
1369 cfs_spin_unlock(&imp->imp_lock);
1370 GOTO(interpret, req->rq_status);
1373 cfs_list_del_init(&req->rq_list);
1374 cfs_list_add_tail(&req->rq_list,
1375 &imp->imp_sending_list);
1377 cfs_spin_unlock(&imp->imp_lock);
1379 req->rq_waiting = 0;
1381 if (req->rq_timedout||req->rq_resend) {
1382 /* This is re-sending anyways,
1383 * let's mark req as resend. */
1388 if (!ptlrpc_unregister_bulk(req, 1))
1391 /* ensure previous bulk fails */
1392 old_xid = req->rq_xid;
1393 req->rq_xid = ptlrpc_next_xid();
1394 CDEBUG(D_HA, "resend bulk "
1397 old_xid, req->rq_xid);
1401 * rq_wait_ctx is only touched by ptlrpcd,
1402 * so no lock is needed here.
1404 status = sptlrpc_req_refresh_ctx(req, -1);
1407 req->rq_status = status;
1408 req->rq_wait_ctx = 0;
1409 force_timer_recalc = 1;
1411 req->rq_wait_ctx = 1;
1416 req->rq_wait_ctx = 0;
1419 rc = ptl_send_rpc(req, 0);
1421 DEBUG_REQ(D_HA, req, "send failed (%d)",
1423 force_timer_recalc = 1;
1424 req->rq_net_err = 1;
1426 /* need to reset the timeout */
1427 force_timer_recalc = 1;
1430 cfs_spin_lock(&req->rq_lock);
1432 if (ptlrpc_client_early(req)) {
1433 ptlrpc_at_recv_early_reply(req);
1434 cfs_spin_unlock(&req->rq_lock);
1438 /* Still waiting for a reply? */
1439 if (ptlrpc_client_recv(req)) {
1440 cfs_spin_unlock(&req->rq_lock);
1444 /* Did we actually receive a reply? */
1445 if (!ptlrpc_client_replied(req)) {
1446 cfs_spin_unlock(&req->rq_lock);
1450 cfs_spin_unlock(&req->rq_lock);
1452 /* unlink from net because we are going to
1453 * swab in-place of reply buffer */
1454 unregistered = ptlrpc_unregister_reply(req, 1);
1458 req->rq_status = after_reply(req);
1462 /* If there is no bulk associated with this request,
1463 * then we're done and should let the interpreter
1464 * process the reply. Similarly if the RPC returned
1465 * an error, and therefore the bulk will never arrive.
1467 if (req->rq_bulk == NULL || req->rq_status != 0) {
1468 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1469 GOTO(interpret, req->rq_status);
1472 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1475 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1476 if (ptlrpc_client_bulk_active(req))
1479 if (!req->rq_bulk->bd_success) {
1480 /* The RPC reply arrived OK, but the bulk screwed
1481 * up! Dead weird since the server told us the RPC
1482 * was good after getting the REPLY for her GET or
1483 * the ACK for her PUT. */
1484 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1488 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1491 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1493 /* This moves to "unregistering" phase we need to wait for
1495 if (!unregistered && !ptlrpc_unregister_reply(req, 1))
1498 if (!ptlrpc_unregister_bulk(req, 1))
1501 /* When calling interpret receiving already should be
1503 LASSERT(!req->rq_receiving_reply);
1505 ptlrpc_req_interpret(env, req, req->rq_status);
1507 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1509 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1510 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1511 imp->imp_obd->obd_uuid.uuid,
1512 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1514 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1515 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1517 cfs_spin_lock(&imp->imp_lock);
1518 /* Request already may be not on sending or delaying list. This
1519 * may happen in the case of marking it erroneous for the case
1520 * ptlrpc_import_delay_req(req, status) find it impossible to
1521 * allow sending this rpc and returns *status != 0. */
1522 if (!cfs_list_empty(&req->rq_list)) {
1523 cfs_list_del_init(&req->rq_list);
1524 cfs_atomic_dec(&imp->imp_inflight);
1526 cfs_spin_unlock(&imp->imp_lock);
1528 set->set_remaining--;
1529 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1532 /* If we hit an error, we want to recover promptly. */
1533 RETURN(set->set_remaining == 0 || force_timer_recalc);
1536 /* Return 1 if we should give up, else 0 */
1537 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1539 struct obd_import *imp = req->rq_import;
1543 cfs_spin_lock(&req->rq_lock);
1544 req->rq_timedout = 1;
1545 cfs_spin_unlock(&req->rq_lock);
1547 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req,
1548 "Request x"LPU64" sent from %s to NID %s "CFS_DURATION_T"s "
1549 "ago has %s ("CFS_DURATION_T"s prior to deadline).\n",
1550 req->rq_xid, imp ? imp->imp_obd->obd_name : "<?>",
1551 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1552 cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1553 req->rq_net_err ? "failed due to network error" : "timed out",
1554 cfs_time_sub(req->rq_deadline, req->rq_sent));
1556 if (imp != NULL && obd_debug_peer_on_timeout)
1557 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1559 ptlrpc_unregister_reply(req, async_unlink);
1560 ptlrpc_unregister_bulk(req, async_unlink);
1562 if (obd_dump_on_timeout)
1563 libcfs_debug_dumplog();
1566 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1573 cfs_atomic_inc(&imp->imp_timeouts);
1575 /* The DLM server doesn't want recovery run on its imports. */
1576 if (imp->imp_dlm_fake)
1579 /* If this request is for recovery or other primordial tasks,
1580 * then error it out here. */
1581 if (req->rq_ctx_init || req->rq_ctx_fini ||
1582 req->rq_send_state != LUSTRE_IMP_FULL ||
1583 imp->imp_obd->obd_no_recov) {
1584 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1585 ptlrpc_import_state_name(req->rq_send_state),
1586 ptlrpc_import_state_name(imp->imp_state));
1587 cfs_spin_lock(&req->rq_lock);
1588 req->rq_status = -ETIMEDOUT;
1590 cfs_spin_unlock(&req->rq_lock);
1594 /* if a request can't be resent we can't wait for an answer after
1596 if (req->rq_no_resend) {
1597 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1601 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1606 int ptlrpc_expired_set(void *data)
1608 struct ptlrpc_request_set *set = data;
1610 time_t now = cfs_time_current_sec();
1613 LASSERT(set != NULL);
1616 * A timeout expired. See which reqs it applies to...
1618 cfs_list_for_each (tmp, &set->set_requests) {
1619 struct ptlrpc_request *req =
1620 cfs_list_entry(tmp, struct ptlrpc_request,
1623 /* don't expire request waiting for context */
1624 if (req->rq_wait_ctx)
1627 /* Request in-flight? */
1628 if (!((req->rq_phase == RQ_PHASE_RPC &&
1629 !req->rq_waiting && !req->rq_resend) ||
1630 (req->rq_phase == RQ_PHASE_BULK)))
1633 if (req->rq_timedout || /* already dealt with */
1634 req->rq_deadline > now) /* not expired */
1637 /* Deal with this guy. Do it asynchronously to not block
1638 * ptlrpcd thread. */
1639 ptlrpc_expire_one_request(req, 1);
1643 * When waiting for a whole set, we always break out of the
1644 * sleep so we can recalculate the timeout, or enable interrupts
1645 * if everyone's timed out.
1650 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1652 cfs_spin_lock(&req->rq_lock);
1654 cfs_spin_unlock(&req->rq_lock);
1657 void ptlrpc_interrupted_set(void *data)
1659 struct ptlrpc_request_set *set = data;
1662 LASSERT(set != NULL);
1663 CERROR("INTERRUPTED SET %p\n", set);
1665 cfs_list_for_each(tmp, &set->set_requests) {
1666 struct ptlrpc_request *req =
1667 cfs_list_entry(tmp, struct ptlrpc_request,
1670 if (req->rq_phase != RQ_PHASE_RPC &&
1671 req->rq_phase != RQ_PHASE_UNREGISTERING)
1674 ptlrpc_mark_interrupted(req);
1679 * Get the smallest timeout in the set; this does NOT set a timeout.
1681 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1684 time_t now = cfs_time_current_sec();
1686 struct ptlrpc_request *req;
1690 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1692 cfs_list_for_each(tmp, &set->set_requests) {
1693 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1696 * Request in-flight?
1698 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1699 (req->rq_phase == RQ_PHASE_BULK) ||
1700 (req->rq_phase == RQ_PHASE_NEW)))
1704 * Already timed out.
1706 if (req->rq_timedout)
1712 if (req->rq_wait_ctx)
1715 if (req->rq_phase == RQ_PHASE_NEW)
1716 deadline = req->rq_sent;
1718 deadline = req->rq_sent + req->rq_timeout;
1720 if (deadline <= now) /* actually expired already */
1721 timeout = 1; /* ASAP */
1722 else if (timeout == 0 || timeout > deadline - now)
1723 timeout = deadline - now;
1728 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1731 struct ptlrpc_request *req;
1732 struct l_wait_info lwi;
1736 if (cfs_list_empty(&set->set_requests))
1739 cfs_list_for_each(tmp, &set->set_requests) {
1740 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1741 if (req->rq_phase == RQ_PHASE_NEW)
1742 (void)ptlrpc_send_new_req(req);
1746 timeout = ptlrpc_set_next_timeout(set);
1748 /* wait until all complete, interrupted, or an in-flight
1750 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1753 if (timeout == 0 && !cfs_signal_pending())
1755 * No requests are in-flight (ether timed out
1756 * or delayed), so we can allow interrupts.
1757 * We still want to block for a limited time,
1758 * so we allow interrupts during the timeout.
1760 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
1762 ptlrpc_interrupted_set, set);
1765 * At least one request is in flight, so no
1766 * interrupts are allowed. Wait until all
1767 * complete, or an in-flight req times out.
1769 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
1770 ptlrpc_expired_set, set);
1772 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
1774 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1776 /* -EINTR => all requests have been flagged rq_intr so next
1778 * -ETIMEDOUT => someone timed out. When all reqs have
1779 * timed out, signals are enabled allowing completion with
1781 * I don't really care if we go once more round the loop in
1782 * the error cases -eeb. */
1783 } while (rc != 0 || set->set_remaining != 0);
1785 LASSERT(set->set_remaining == 0);
1788 cfs_list_for_each(tmp, &set->set_requests) {
1789 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1791 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1792 if (req->rq_status != 0)
1793 rc = req->rq_status;
1796 if (set->set_interpret != NULL) {
1797 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1799 rc = interpreter (set, set->set_arg, rc);
1801 struct ptlrpc_set_cbdata *cbdata, *n;
1804 cfs_list_for_each_entry_safe(cbdata, n,
1805 &set->set_cblist, psc_item) {
1806 cfs_list_del_init(&cbdata->psc_item);
1807 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1810 OBD_FREE_PTR(cbdata);
1817 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1820 if (request == NULL) {
1825 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1826 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1827 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
1828 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
1829 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
1830 LASSERTF(!request->rq_replay, "req %p\n", request);
1831 LASSERT(request->rq_cli_ctx || request->rq_fake);
1833 req_capsule_fini(&request->rq_pill);
1835 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1836 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1837 if (request->rq_import != NULL) {
1839 cfs_spin_lock(&request->rq_import->imp_lock);
1840 cfs_list_del_init(&request->rq_replay_list);
1842 cfs_spin_unlock(&request->rq_import->imp_lock);
1844 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
1846 if (cfs_atomic_read(&request->rq_refcount) != 0) {
1847 DEBUG_REQ(D_ERROR, request,
1848 "freeing request with nonzero refcount");
1852 if (request->rq_repbuf != NULL)
1853 sptlrpc_cli_free_repbuf(request);
1854 if (request->rq_export != NULL) {
1855 class_export_put(request->rq_export);
1856 request->rq_export = NULL;
1858 if (request->rq_import != NULL) {
1859 class_import_put(request->rq_import);
1860 request->rq_import = NULL;
1862 if (request->rq_bulk != NULL)
1863 ptlrpc_free_bulk(request->rq_bulk);
1865 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1866 sptlrpc_cli_free_reqbuf(request);
1868 if (request->rq_cli_ctx)
1869 sptlrpc_req_put_ctx(request, !locked);
1871 if (request->rq_pool)
1872 __ptlrpc_free_req_to_pool(request);
1874 OBD_FREE(request, sizeof(*request));
1878 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1879 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1881 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1882 (void)__ptlrpc_req_finished(request, 1);
1885 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1888 if (request == NULL)
1891 if (request == LP_POISON ||
1892 request->rq_reqmsg == LP_POISON) {
1893 CERROR("dereferencing freed request (bug 575)\n");
1898 DEBUG_REQ(D_INFO, request, "refcount now %u",
1899 cfs_atomic_read(&request->rq_refcount) - 1);
1901 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
1902 __ptlrpc_free_req(request, locked);
1909 void ptlrpc_req_finished(struct ptlrpc_request *request)
1911 __ptlrpc_req_finished(request, 0);
1914 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1916 return request->rq_xid;
1918 EXPORT_SYMBOL(ptlrpc_req_xid);
1920 /* Disengage the client's reply buffer from the network
1921 * NB does _NOT_ unregister any client-side bulk.
1922 * IDEMPOTENT, but _not_ safe against concurrent callers.
1923 * The request owner (i.e. the thread doing the I/O) must call...
1925 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1929 struct l_wait_info lwi;
1934 LASSERT(!cfs_in_interrupt());
1937 * Let's setup deadline for reply unlink.
1939 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1940 async && request->rq_reply_deadline == 0)
1941 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1944 * Nothing left to do.
1946 if (!ptlrpc_client_recv_or_unlink(request))
1949 LNetMDUnlink(request->rq_reply_md_h);
1952 * Let's check it once again.
1954 if (!ptlrpc_client_recv_or_unlink(request))
1958 * Move to "Unregistering" phase as reply was not unlinked yet.
1960 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1963 * Do not wait for unlink to finish.
1969 * We have to l_wait_event() whatever the result, to give liblustre
1970 * a chance to run reply_in_callback(), and to make sure we've
1971 * unlinked before returning a req to the pool.
1973 if (request->rq_set != NULL)
1974 wq = &request->rq_set->set_waitq;
1976 wq = &request->rq_reply_waitq;
1979 /* Network access will complete in finite time but the HUGE
1980 * timeout lets us CWARN for visibility of sluggish NALs */
1981 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1982 cfs_time_seconds(1), NULL, NULL);
1983 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1986 ptlrpc_rqphase_move(request, request->rq_next_phase);
1990 LASSERT(rc == -ETIMEDOUT);
1991 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1992 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1993 request->rq_must_unlink);
1998 /* caller must hold imp->imp_lock */
1999 void ptlrpc_free_committed(struct obd_import *imp)
2001 cfs_list_t *tmp, *saved;
2002 struct ptlrpc_request *req;
2003 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2006 LASSERT(imp != NULL);
2008 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2011 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2012 imp->imp_generation == imp->imp_last_generation_checked) {
2013 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
2014 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2018 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2019 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2020 imp->imp_generation);
2021 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2022 imp->imp_last_generation_checked = imp->imp_generation;
2024 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2025 req = cfs_list_entry(tmp, struct ptlrpc_request,
2028 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2029 LASSERT(req != last_req);
2032 if (req->rq_transno == 0) {
2033 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2036 if (req->rq_import_generation < imp->imp_generation) {
2037 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2041 if (req->rq_replay) {
2042 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2046 /* not yet committed */
2047 if (req->rq_transno > imp->imp_peer_committed_transno) {
2048 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2052 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
2053 imp->imp_peer_committed_transno);
2055 cfs_spin_lock(&req->rq_lock);
2057 cfs_spin_unlock(&req->rq_lock);
2058 if (req->rq_commit_cb != NULL)
2059 req->rq_commit_cb(req);
2060 cfs_list_del_init(&req->rq_replay_list);
2061 __ptlrpc_req_finished(req, 1);
2068 void ptlrpc_cleanup_client(struct obd_import *imp)
2075 void ptlrpc_resend_req(struct ptlrpc_request *req)
2077 DEBUG_REQ(D_HA, req, "going to resend");
2078 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2079 req->rq_status = -EAGAIN;
2081 cfs_spin_lock(&req->rq_lock);
2083 req->rq_net_err = 0;
2084 req->rq_timedout = 0;
2086 __u64 old_xid = req->rq_xid;
2088 /* ensure previous bulk fails */
2089 req->rq_xid = ptlrpc_next_xid();
2090 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2091 old_xid, req->rq_xid);
2093 ptlrpc_client_wake_req(req);
2094 cfs_spin_unlock(&req->rq_lock);
2097 /* XXX: this function and rq_status are currently unused */
2098 void ptlrpc_restart_req(struct ptlrpc_request *req)
2100 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2101 req->rq_status = -ERESTARTSYS;
2103 cfs_spin_lock(&req->rq_lock);
2104 req->rq_restart = 1;
2105 req->rq_timedout = 0;
2106 ptlrpc_client_wake_req(req);
2107 cfs_spin_unlock(&req->rq_lock);
2110 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2113 cfs_atomic_inc(&req->rq_refcount);
2117 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2118 struct obd_import *imp)
2122 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2124 if (req->rq_transno == 0) {
2125 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2129 /* clear this for new requests that were resent as well
2130 as resent replayed requests. */
2131 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2133 /* don't re-add requests that have been replayed */
2134 if (!cfs_list_empty(&req->rq_replay_list))
2137 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2139 LASSERT(imp->imp_replayable);
2140 /* Balanced in ptlrpc_free_committed, usually. */
2141 ptlrpc_request_addref(req);
2142 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2143 struct ptlrpc_request *iter =
2144 cfs_list_entry(tmp, struct ptlrpc_request,
2147 /* We may have duplicate transnos if we create and then
2148 * open a file, or for closes retained if to match creating
2149 * opens, so use req->rq_xid as a secondary key.
2150 * (See bugs 684, 685, and 428.)
2151 * XXX no longer needed, but all opens need transnos!
2153 if (iter->rq_transno > req->rq_transno)
2156 if (iter->rq_transno == req->rq_transno) {
2157 LASSERT(iter->rq_xid != req->rq_xid);
2158 if (iter->rq_xid > req->rq_xid)
2162 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2166 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2169 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2171 struct ptlrpc_request_set *set;
2175 LASSERT(req->rq_set == NULL);
2176 LASSERT(!req->rq_receiving_reply);
2178 set = ptlrpc_prep_set();
2180 CERROR("Unable to allocate ptlrpc set.");
2184 /* for distributed debugging */
2185 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2187 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2188 ptlrpc_request_addref(req);
2189 ptlrpc_set_add_req(set, req);
2190 rc = ptlrpc_set_wait(set);
2191 ptlrpc_set_destroy(set);
2196 struct ptlrpc_replay_async_args {
2198 int praa_old_status;
2201 static int ptlrpc_replay_interpret(const struct lu_env *env,
2202 struct ptlrpc_request *req,
2203 void * data, int rc)
2205 struct ptlrpc_replay_async_args *aa = data;
2206 struct obd_import *imp = req->rq_import;
2209 cfs_atomic_dec(&imp->imp_replay_inflight);
2211 if (!ptlrpc_client_replied(req)) {
2212 CERROR("request replay timed out, restarting recovery\n");
2213 GOTO(out, rc = -ETIMEDOUT);
2216 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2217 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2218 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2219 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2221 /** VBR: check version failure */
2222 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2223 /** replay was failed due to version mismatch */
2224 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2225 cfs_spin_lock(&imp->imp_lock);
2226 imp->imp_vbr_failed = 1;
2227 imp->imp_no_lock_replay = 1;
2228 cfs_spin_unlock(&imp->imp_lock);
2230 /** The transno had better not change over replay. */
2231 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2232 lustre_msg_get_transno(req->rq_repmsg) ||
2233 lustre_msg_get_transno(req->rq_repmsg) == 0,
2235 lustre_msg_get_transno(req->rq_reqmsg),
2236 lustre_msg_get_transno(req->rq_repmsg));
2239 cfs_spin_lock(&imp->imp_lock);
2240 /** if replays by version then gap was occur on server, no trust to locks */
2241 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2242 imp->imp_no_lock_replay = 1;
2243 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2244 cfs_spin_unlock(&imp->imp_lock);
2245 LASSERT(imp->imp_last_replay_transno);
2247 DEBUG_REQ(D_HA, req, "got rep");
2249 /* let the callback do fixups, possibly including in the request */
2250 if (req->rq_replay_cb)
2251 req->rq_replay_cb(req);
2253 if (ptlrpc_client_replied(req) &&
2254 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2255 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2256 lustre_msg_get_status(req->rq_repmsg),
2257 aa->praa_old_status);
2259 /* Put it back for re-replay. */
2260 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2264 * Errors while replay can set transno to 0, but
2265 * imp_last_replay_transno shouldn't be set to 0 anyway
2267 if (req->rq_transno > 0) {
2268 cfs_spin_lock(&imp->imp_lock);
2269 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2270 imp->imp_last_replay_transno = req->rq_transno;
2271 cfs_spin_unlock(&imp->imp_lock);
2273 CERROR("Transno is 0 during replay!\n");
2274 /* continue with recovery */
2275 rc = ptlrpc_import_recovery_state_machine(imp);
2277 req->rq_send_state = aa->praa_old_state;
2280 /* this replay failed, so restart recovery */
2281 ptlrpc_connect_import(imp, NULL);
2286 int ptlrpc_replay_req(struct ptlrpc_request *req)
2288 struct ptlrpc_replay_async_args *aa;
2291 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2292 /* Not handling automatic bulk replay yet (or ever?) */
2293 LASSERT(req->rq_bulk == NULL);
2295 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2296 aa = ptlrpc_req_async_args(req);
2297 memset(aa, 0, sizeof *aa);
2299 /* Prepare request to be resent with ptlrpcd */
2300 aa->praa_old_state = req->rq_send_state;
2301 req->rq_send_state = LUSTRE_IMP_REPLAY;
2302 req->rq_phase = RQ_PHASE_NEW;
2303 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2305 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2307 req->rq_interpret_reply = ptlrpc_replay_interpret;
2308 /* Readjust the timeout for current conditions */
2309 ptlrpc_at_set_req_timeout(req);
2311 DEBUG_REQ(D_HA, req, "REPLAY");
2313 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2314 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2316 ptlrpcd_add_req(req, PSCOPE_OTHER);
2320 void ptlrpc_abort_inflight(struct obd_import *imp)
2322 cfs_list_t *tmp, *n;
2325 /* Make sure that no new requests get processed for this import.
2326 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2327 * this flag and then putting requests on sending_list or delayed_list.
2329 cfs_spin_lock(&imp->imp_lock);
2331 /* XXX locking? Maybe we should remove each request with the list
2332 * locked? Also, how do we know if the requests on the list are
2333 * being freed at this time?
2335 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2336 struct ptlrpc_request *req =
2337 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2339 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2341 cfs_spin_lock (&req->rq_lock);
2342 if (req->rq_import_generation < imp->imp_generation) {
2344 req->rq_status = -EINTR;
2345 ptlrpc_client_wake_req(req);
2347 cfs_spin_unlock (&req->rq_lock);
2350 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2351 struct ptlrpc_request *req =
2352 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2354 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2356 cfs_spin_lock (&req->rq_lock);
2357 if (req->rq_import_generation < imp->imp_generation) {
2359 req->rq_status = -EINTR;
2360 ptlrpc_client_wake_req(req);
2362 cfs_spin_unlock (&req->rq_lock);
2365 /* Last chance to free reqs left on the replay list, but we
2366 * will still leak reqs that haven't committed. */
2367 if (imp->imp_replayable)
2368 ptlrpc_free_committed(imp);
2370 cfs_spin_unlock(&imp->imp_lock);
2375 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2377 cfs_list_t *tmp, *pos;
2379 LASSERT(set != NULL);
2381 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2382 struct ptlrpc_request *req =
2383 cfs_list_entry(pos, struct ptlrpc_request,
2386 cfs_spin_lock(&req->rq_lock);
2387 if (req->rq_phase != RQ_PHASE_RPC) {
2388 cfs_spin_unlock(&req->rq_lock);
2393 req->rq_status = -EINTR;
2394 ptlrpc_client_wake_req(req);
2395 cfs_spin_unlock(&req->rq_lock);
2399 static __u64 ptlrpc_last_xid;
2400 static cfs_spinlock_t ptlrpc_last_xid_lock;
2402 /* Initialize the XID for the node. This is common among all requests on
2403 * this node, and only requires the property that it is monotonically
2404 * increasing. It does not need to be sequential. Since this is also used
2405 * as the RDMA match bits, it is important that a single client NOT have
2406 * the same match bits for two different in-flight requests, hence we do
2407 * NOT want to have an XID per target or similar.
2409 * To avoid an unlikely collision between match bits after a client reboot
2410 * (which would deliver old data into the wrong RDMA buffer) initialize
2411 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2412 * If the time is clearly incorrect, we instead use a 62-bit random number.
2413 * In the worst case the random number will overflow 1M RPCs per second in
2414 * 9133 years, or permutations thereof.
2416 #define YEAR_2004 (1ULL << 30)
2417 void ptlrpc_init_xid(void)
2419 time_t now = cfs_time_current_sec();
2421 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2422 if (now < YEAR_2004) {
2423 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2424 ptlrpc_last_xid >>= 2;
2425 ptlrpc_last_xid |= (1ULL << 61);
2427 ptlrpc_last_xid = (__u64)now << 20;
2431 __u64 ptlrpc_next_xid(void)
2434 cfs_spin_lock(&ptlrpc_last_xid_lock);
2435 tmp = ++ptlrpc_last_xid;
2436 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2440 __u64 ptlrpc_sample_next_xid(void)
2442 #if BITS_PER_LONG == 32
2443 /* need to avoid possible word tearing on 32-bit systems */
2445 cfs_spin_lock(&ptlrpc_last_xid_lock);
2446 tmp = ptlrpc_last_xid + 1;
2447 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2450 /* No need to lock, since returned value is racy anyways */
2451 return ptlrpc_last_xid + 1;
2454 EXPORT_SYMBOL(ptlrpc_sample_next_xid);