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(!req->rq_receiving_reply);
1021 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
1023 if (req->rq_reply_truncate && !req->rq_no_resend) {
1025 sptlrpc_cli_free_repbuf(req);
1026 req->rq_replen = req->rq_nob_received;
1031 * NB Until this point, the whole of the incoming message,
1032 * including buflens, status etc is in the sender's byte order.
1035 rc = sptlrpc_cli_unwrap_reply(req);
1037 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1042 * Security layer unwrap might ask resend this request.
1047 rc = unpack_reply(req);
1051 cfs_gettimeofday(&work_start);
1052 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1053 if (obd->obd_svc_stats != NULL) {
1054 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1056 ptlrpc_lprocfs_rpc_sent(req, timediff);
1059 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1060 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1061 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1062 lustre_msg_get_type(req->rq_repmsg));
1066 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1067 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1068 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1069 ptlrpc_at_adj_net_latency(req,
1070 lustre_msg_get_service_time(req->rq_repmsg));
1072 rc = ptlrpc_check_status(req);
1073 imp->imp_connect_error = rc;
1077 * Either we've been evicted, or the server has failed for
1078 * some reason. Try to reconnect, and if that fails, punt to
1081 if (ll_rpc_recoverable_error(rc)) {
1082 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1083 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1086 ptlrpc_request_handle_notconn(req);
1091 * Let's look if server sent slv. Do it only for RPC with
1094 ldlm_cli_update_pool(req);
1098 * Store transno in reqmsg for replay.
1100 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1101 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1102 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1105 if (imp->imp_replayable) {
1106 cfs_spin_lock(&imp->imp_lock);
1108 * No point in adding already-committed requests to the replay
1109 * list, we will just remove them immediately. b=9829
1111 if (req->rq_transno != 0 &&
1113 lustre_msg_get_last_committed(req->rq_repmsg) ||
1115 /** version recovery */
1116 ptlrpc_save_versions(req);
1117 ptlrpc_retain_replayable_request(req, imp);
1118 } else if (req->rq_commit_cb != NULL) {
1119 cfs_spin_unlock(&imp->imp_lock);
1120 req->rq_commit_cb(req);
1121 cfs_spin_lock(&imp->imp_lock);
1125 * Replay-enabled imports return commit-status information.
1127 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1128 imp->imp_peer_committed_transno =
1129 lustre_msg_get_last_committed(req->rq_repmsg);
1131 ptlrpc_free_committed(imp);
1132 cfs_spin_unlock(&imp->imp_lock);
1138 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1140 struct obd_import *imp;
1144 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1145 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1148 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1150 imp = req->rq_import;
1151 cfs_spin_lock(&imp->imp_lock);
1153 req->rq_import_generation = imp->imp_generation;
1155 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1156 cfs_spin_lock(&req->rq_lock);
1157 req->rq_waiting = 1;
1158 cfs_spin_unlock(&req->rq_lock);
1160 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1161 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1162 ptlrpc_import_state_name(req->rq_send_state),
1163 ptlrpc_import_state_name(imp->imp_state));
1164 LASSERT(cfs_list_empty(&req->rq_list));
1165 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1166 cfs_atomic_inc(&req->rq_import->imp_inflight);
1167 cfs_spin_unlock(&imp->imp_lock);
1172 cfs_spin_unlock(&imp->imp_lock);
1173 req->rq_status = rc;
1174 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1178 LASSERT(cfs_list_empty(&req->rq_list));
1179 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1180 cfs_atomic_inc(&req->rq_import->imp_inflight);
1181 cfs_spin_unlock(&imp->imp_lock);
1183 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1185 rc = sptlrpc_req_refresh_ctx(req, -1);
1188 req->rq_status = rc;
1191 req->rq_wait_ctx = 1;
1196 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1197 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1198 imp->imp_obd->obd_uuid.uuid,
1199 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1200 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1201 lustre_msg_get_opc(req->rq_reqmsg));
1203 rc = ptl_send_rpc(req, 0);
1205 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1206 req->rq_net_err = 1;
1212 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1213 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1216 int force_timer_recalc = 0;
1219 if (set->set_remaining == 0)
1222 cfs_list_for_each(tmp, &set->set_requests) {
1223 struct ptlrpc_request *req =
1224 cfs_list_entry(tmp, struct ptlrpc_request,
1226 struct obd_import *imp = req->rq_import;
1229 if (req->rq_phase == RQ_PHASE_NEW &&
1230 ptlrpc_send_new_req(req)) {
1231 force_timer_recalc = 1;
1234 /* delayed send - skip */
1235 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1238 if (!(req->rq_phase == RQ_PHASE_RPC ||
1239 req->rq_phase == RQ_PHASE_BULK ||
1240 req->rq_phase == RQ_PHASE_INTERPRET ||
1241 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1242 req->rq_phase == RQ_PHASE_COMPLETE)) {
1243 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1247 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1248 LASSERT(req->rq_next_phase != req->rq_phase);
1249 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1252 * Skip processing until reply is unlinked. We
1253 * can't return to pool before that and we can't
1254 * call interpret before that. We need to make
1255 * sure that all rdma transfers finished and will
1256 * not corrupt any data.
1258 if (ptlrpc_client_recv_or_unlink(req) ||
1259 ptlrpc_client_bulk_active(req))
1263 * Turn fail_loc off to prevent it from looping
1266 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1267 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1270 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1271 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1276 * Move to next phase if reply was successfully
1279 ptlrpc_rqphase_move(req, req->rq_next_phase);
1282 if (req->rq_phase == RQ_PHASE_COMPLETE)
1285 if (req->rq_phase == RQ_PHASE_INTERPRET)
1286 GOTO(interpret, req->rq_status);
1289 * Note that this also will start async reply unlink.
1291 if (req->rq_net_err && !req->rq_timedout) {
1292 ptlrpc_expire_one_request(req, 1);
1295 * Check if we still need to wait for unlink.
1297 if (ptlrpc_client_recv_or_unlink(req) ||
1298 ptlrpc_client_bulk_active(req))
1303 req->rq_replied = 0;
1304 if (req->rq_status == 0)
1305 req->rq_status = -EIO;
1306 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1307 GOTO(interpret, req->rq_status);
1310 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1311 * so it sets rq_intr regardless of individual rpc
1312 * timeouts. The synchronous IO waiting path sets
1313 * rq_intr irrespective of whether ptlrpcd
1314 * has seen a timeout. Our policy is to only interpret
1315 * interrupted rpcs after they have timed out, so we
1316 * need to enforce that here.
1319 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1320 req->rq_wait_ctx)) {
1321 req->rq_status = -EINTR;
1322 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1323 GOTO(interpret, req->rq_status);
1326 if (req->rq_phase == RQ_PHASE_RPC) {
1327 if (req->rq_timedout || req->rq_resend ||
1328 req->rq_waiting || req->rq_wait_ctx) {
1331 if (!ptlrpc_unregister_reply(req, 1))
1334 cfs_spin_lock(&imp->imp_lock);
1335 if (ptlrpc_import_delay_req(imp, req, &status)){
1336 /* put on delay list - only if we wait
1337 * recovery finished - before send */
1338 cfs_list_del_init(&req->rq_list);
1339 cfs_list_add_tail(&req->rq_list,
1342 cfs_spin_unlock(&imp->imp_lock);
1347 req->rq_status = status;
1348 ptlrpc_rqphase_move(req,
1349 RQ_PHASE_INTERPRET);
1350 cfs_spin_unlock(&imp->imp_lock);
1351 GOTO(interpret, req->rq_status);
1353 if (req->rq_no_resend && !req->rq_wait_ctx) {
1354 req->rq_status = -ENOTCONN;
1355 ptlrpc_rqphase_move(req,
1356 RQ_PHASE_INTERPRET);
1357 cfs_spin_unlock(&imp->imp_lock);
1358 GOTO(interpret, req->rq_status);
1361 cfs_list_del_init(&req->rq_list);
1362 cfs_list_add_tail(&req->rq_list,
1363 &imp->imp_sending_list);
1365 cfs_spin_unlock(&imp->imp_lock);
1367 req->rq_waiting = 0;
1369 if (req->rq_timedout||req->rq_resend) {
1370 /* This is re-sending anyways,
1371 * let's mark req as resend. */
1376 if (!ptlrpc_unregister_bulk(req, 1))
1379 /* ensure previous bulk fails */
1380 old_xid = req->rq_xid;
1381 req->rq_xid = ptlrpc_next_xid();
1382 CDEBUG(D_HA, "resend bulk "
1385 old_xid, req->rq_xid);
1389 * rq_wait_ctx is only touched by ptlrpcd,
1390 * so no lock is needed here.
1392 status = sptlrpc_req_refresh_ctx(req, -1);
1395 req->rq_status = status;
1396 req->rq_wait_ctx = 0;
1397 force_timer_recalc = 1;
1399 req->rq_wait_ctx = 1;
1404 req->rq_wait_ctx = 0;
1407 rc = ptl_send_rpc(req, 0);
1409 DEBUG_REQ(D_HA, req, "send failed (%d)",
1411 force_timer_recalc = 1;
1412 req->rq_net_err = 1;
1414 /* need to reset the timeout */
1415 force_timer_recalc = 1;
1418 cfs_spin_lock(&req->rq_lock);
1420 if (ptlrpc_client_early(req)) {
1421 ptlrpc_at_recv_early_reply(req);
1422 cfs_spin_unlock(&req->rq_lock);
1426 /* Still waiting for a reply? */
1427 if (ptlrpc_client_recv(req)) {
1428 cfs_spin_unlock(&req->rq_lock);
1432 /* Did we actually receive a reply? */
1433 if (!ptlrpc_client_replied(req)) {
1434 cfs_spin_unlock(&req->rq_lock);
1438 cfs_spin_unlock(&req->rq_lock);
1440 req->rq_status = after_reply(req);
1444 /* If there is no bulk associated with this request,
1445 * then we're done and should let the interpreter
1446 * process the reply. Similarly if the RPC returned
1447 * an error, and therefore the bulk will never arrive.
1449 if (req->rq_bulk == NULL || req->rq_status != 0) {
1450 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1451 GOTO(interpret, req->rq_status);
1454 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1457 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1458 if (ptlrpc_client_bulk_active(req))
1461 if (!req->rq_bulk->bd_success) {
1462 /* The RPC reply arrived OK, but the bulk screwed
1463 * up! Dead weird since the server told us the RPC
1464 * was good after getting the REPLY for her GET or
1465 * the ACK for her PUT. */
1466 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1470 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1473 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1475 /* This moves to "unregistering" phase we need to wait for
1477 if (!ptlrpc_unregister_reply(req, 1))
1480 if (!ptlrpc_unregister_bulk(req, 1))
1483 /* When calling interpret receiving already should be
1485 LASSERT(!req->rq_receiving_reply);
1487 ptlrpc_req_interpret(env, req, req->rq_status);
1489 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1491 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1492 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1493 imp->imp_obd->obd_uuid.uuid,
1494 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1496 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1497 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1499 cfs_spin_lock(&imp->imp_lock);
1500 /* Request already may be not on sending or delaying list. This
1501 * may happen in the case of marking it erroneous for the case
1502 * ptlrpc_import_delay_req(req, status) find it impossible to
1503 * allow sending this rpc and returns *status != 0. */
1504 if (!cfs_list_empty(&req->rq_list)) {
1505 cfs_list_del_init(&req->rq_list);
1506 cfs_atomic_dec(&imp->imp_inflight);
1508 cfs_spin_unlock(&imp->imp_lock);
1510 set->set_remaining--;
1511 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1514 /* If we hit an error, we want to recover promptly. */
1515 RETURN(set->set_remaining == 0 || force_timer_recalc);
1518 /* Return 1 if we should give up, else 0 */
1519 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1521 struct obd_import *imp = req->rq_import;
1525 cfs_spin_lock(&req->rq_lock);
1526 req->rq_timedout = 1;
1527 cfs_spin_unlock(&req->rq_lock);
1529 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req,
1530 "Request x"LPU64" sent from %s to NID %s "CFS_DURATION_T"s "
1531 "ago has %s ("CFS_DURATION_T"s prior to deadline).\n",
1532 req->rq_xid, imp ? imp->imp_obd->obd_name : "<?>",
1533 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1534 cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1535 req->rq_net_err ? "failed due to network error" : "timed out",
1536 cfs_time_sub(req->rq_deadline, req->rq_sent));
1538 if (imp != NULL && obd_debug_peer_on_timeout)
1539 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1541 ptlrpc_unregister_reply(req, async_unlink);
1542 ptlrpc_unregister_bulk(req, async_unlink);
1544 if (obd_dump_on_timeout)
1545 libcfs_debug_dumplog();
1548 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1555 cfs_atomic_inc(&imp->imp_timeouts);
1557 /* The DLM server doesn't want recovery run on its imports. */
1558 if (imp->imp_dlm_fake)
1561 /* If this request is for recovery or other primordial tasks,
1562 * then error it out here. */
1563 if (req->rq_ctx_init || req->rq_ctx_fini ||
1564 req->rq_send_state != LUSTRE_IMP_FULL ||
1565 imp->imp_obd->obd_no_recov) {
1566 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1567 ptlrpc_import_state_name(req->rq_send_state),
1568 ptlrpc_import_state_name(imp->imp_state));
1569 cfs_spin_lock(&req->rq_lock);
1570 req->rq_status = -ETIMEDOUT;
1572 cfs_spin_unlock(&req->rq_lock);
1576 /* if a request can't be resent we can't wait for an answer after
1578 if (req->rq_no_resend) {
1579 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1583 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1588 int ptlrpc_expired_set(void *data)
1590 struct ptlrpc_request_set *set = data;
1592 time_t now = cfs_time_current_sec();
1595 LASSERT(set != NULL);
1598 * A timeout expired. See which reqs it applies to...
1600 cfs_list_for_each (tmp, &set->set_requests) {
1601 struct ptlrpc_request *req =
1602 cfs_list_entry(tmp, struct ptlrpc_request,
1605 /* don't expire request waiting for context */
1606 if (req->rq_wait_ctx)
1609 /* Request in-flight? */
1610 if (!((req->rq_phase == RQ_PHASE_RPC &&
1611 !req->rq_waiting && !req->rq_resend) ||
1612 (req->rq_phase == RQ_PHASE_BULK)))
1615 if (req->rq_timedout || /* already dealt with */
1616 req->rq_deadline > now) /* not expired */
1619 /* Deal with this guy. Do it asynchronously to not block
1620 * ptlrpcd thread. */
1621 ptlrpc_expire_one_request(req, 1);
1625 * When waiting for a whole set, we always break out of the
1626 * sleep so we can recalculate the timeout, or enable interrupts
1627 * if everyone's timed out.
1632 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1634 cfs_spin_lock(&req->rq_lock);
1636 cfs_spin_unlock(&req->rq_lock);
1639 void ptlrpc_interrupted_set(void *data)
1641 struct ptlrpc_request_set *set = data;
1644 LASSERT(set != NULL);
1645 CERROR("INTERRUPTED SET %p\n", set);
1647 cfs_list_for_each(tmp, &set->set_requests) {
1648 struct ptlrpc_request *req =
1649 cfs_list_entry(tmp, struct ptlrpc_request,
1652 if (req->rq_phase != RQ_PHASE_RPC &&
1653 req->rq_phase != RQ_PHASE_UNREGISTERING)
1656 ptlrpc_mark_interrupted(req);
1661 * Get the smallest timeout in the set; this does NOT set a timeout.
1663 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1666 time_t now = cfs_time_current_sec();
1668 struct ptlrpc_request *req;
1672 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1674 cfs_list_for_each(tmp, &set->set_requests) {
1675 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1678 * Request in-flight?
1680 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1681 (req->rq_phase == RQ_PHASE_BULK) ||
1682 (req->rq_phase == RQ_PHASE_NEW)))
1686 * Already timed out.
1688 if (req->rq_timedout)
1694 if (req->rq_wait_ctx)
1697 if (req->rq_phase == RQ_PHASE_NEW)
1698 deadline = req->rq_sent;
1700 deadline = req->rq_sent + req->rq_timeout;
1702 if (deadline <= now) /* actually expired already */
1703 timeout = 1; /* ASAP */
1704 else if (timeout == 0 || timeout > deadline - now)
1705 timeout = deadline - now;
1710 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1713 struct ptlrpc_request *req;
1714 struct l_wait_info lwi;
1718 if (cfs_list_empty(&set->set_requests))
1721 cfs_list_for_each(tmp, &set->set_requests) {
1722 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1723 if (req->rq_phase == RQ_PHASE_NEW)
1724 (void)ptlrpc_send_new_req(req);
1728 timeout = ptlrpc_set_next_timeout(set);
1730 /* wait until all complete, interrupted, or an in-flight
1732 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1735 if (timeout == 0 && !cfs_signal_pending())
1737 * No requests are in-flight (ether timed out
1738 * or delayed), so we can allow interrupts.
1739 * We still want to block for a limited time,
1740 * so we allow interrupts during the timeout.
1742 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
1744 ptlrpc_interrupted_set, set);
1747 * At least one request is in flight, so no
1748 * interrupts are allowed. Wait until all
1749 * complete, or an in-flight req times out.
1751 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
1752 ptlrpc_expired_set, set);
1754 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
1756 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1758 /* -EINTR => all requests have been flagged rq_intr so next
1760 * -ETIMEDOUT => someone timed out. When all reqs have
1761 * timed out, signals are enabled allowing completion with
1763 * I don't really care if we go once more round the loop in
1764 * the error cases -eeb. */
1765 } while (rc != 0 || set->set_remaining != 0);
1767 LASSERT(set->set_remaining == 0);
1770 cfs_list_for_each(tmp, &set->set_requests) {
1771 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1773 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1774 if (req->rq_status != 0)
1775 rc = req->rq_status;
1778 if (set->set_interpret != NULL) {
1779 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1781 rc = interpreter (set, set->set_arg, rc);
1783 struct ptlrpc_set_cbdata *cbdata, *n;
1786 cfs_list_for_each_entry_safe(cbdata, n,
1787 &set->set_cblist, psc_item) {
1788 cfs_list_del_init(&cbdata->psc_item);
1789 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1792 OBD_FREE_PTR(cbdata);
1799 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1802 if (request == NULL) {
1807 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1808 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1809 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
1810 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
1811 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
1812 LASSERTF(!request->rq_replay, "req %p\n", request);
1813 LASSERT(request->rq_cli_ctx || request->rq_fake);
1815 req_capsule_fini(&request->rq_pill);
1817 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1818 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1819 if (request->rq_import != NULL) {
1821 cfs_spin_lock(&request->rq_import->imp_lock);
1822 cfs_list_del_init(&request->rq_replay_list);
1824 cfs_spin_unlock(&request->rq_import->imp_lock);
1826 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
1828 if (cfs_atomic_read(&request->rq_refcount) != 0) {
1829 DEBUG_REQ(D_ERROR, request,
1830 "freeing request with nonzero refcount");
1834 if (request->rq_repbuf != NULL)
1835 sptlrpc_cli_free_repbuf(request);
1836 if (request->rq_export != NULL) {
1837 class_export_put(request->rq_export);
1838 request->rq_export = NULL;
1840 if (request->rq_import != NULL) {
1841 class_import_put(request->rq_import);
1842 request->rq_import = NULL;
1844 if (request->rq_bulk != NULL)
1845 ptlrpc_free_bulk(request->rq_bulk);
1847 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1848 sptlrpc_cli_free_reqbuf(request);
1850 if (request->rq_cli_ctx)
1851 sptlrpc_req_put_ctx(request, !locked);
1853 if (request->rq_pool)
1854 __ptlrpc_free_req_to_pool(request);
1856 OBD_FREE(request, sizeof(*request));
1860 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1861 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1863 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1864 (void)__ptlrpc_req_finished(request, 1);
1867 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1870 if (request == NULL)
1873 if (request == LP_POISON ||
1874 request->rq_reqmsg == LP_POISON) {
1875 CERROR("dereferencing freed request (bug 575)\n");
1880 DEBUG_REQ(D_INFO, request, "refcount now %u",
1881 cfs_atomic_read(&request->rq_refcount) - 1);
1883 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
1884 __ptlrpc_free_req(request, locked);
1891 void ptlrpc_req_finished(struct ptlrpc_request *request)
1893 __ptlrpc_req_finished(request, 0);
1896 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1898 return request->rq_xid;
1900 EXPORT_SYMBOL(ptlrpc_req_xid);
1902 /* Disengage the client's reply buffer from the network
1903 * NB does _NOT_ unregister any client-side bulk.
1904 * IDEMPOTENT, but _not_ safe against concurrent callers.
1905 * The request owner (i.e. the thread doing the I/O) must call...
1907 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1911 struct l_wait_info lwi;
1916 LASSERT(!cfs_in_interrupt());
1919 * Let's setup deadline for reply unlink.
1921 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1922 async && request->rq_reply_deadline == 0)
1923 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1926 * Nothing left to do.
1928 if (!ptlrpc_client_recv_or_unlink(request))
1931 LNetMDUnlink(request->rq_reply_md_h);
1934 * Let's check it once again.
1936 if (!ptlrpc_client_recv_or_unlink(request))
1940 * Move to "Unregistering" phase as reply was not unlinked yet.
1942 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1945 * Do not wait for unlink to finish.
1951 * We have to l_wait_event() whatever the result, to give liblustre
1952 * a chance to run reply_in_callback(), and to make sure we've
1953 * unlinked before returning a req to the pool.
1955 if (request->rq_set != NULL)
1956 wq = &request->rq_set->set_waitq;
1958 wq = &request->rq_reply_waitq;
1961 /* Network access will complete in finite time but the HUGE
1962 * timeout lets us CWARN for visibility of sluggish NALs */
1963 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1964 cfs_time_seconds(1), NULL, NULL);
1965 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1968 ptlrpc_rqphase_move(request, request->rq_next_phase);
1972 LASSERT(rc == -ETIMEDOUT);
1973 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1974 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1975 request->rq_must_unlink);
1980 /* caller must hold imp->imp_lock */
1981 void ptlrpc_free_committed(struct obd_import *imp)
1983 cfs_list_t *tmp, *saved;
1984 struct ptlrpc_request *req;
1985 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1988 LASSERT(imp != NULL);
1990 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1993 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1994 imp->imp_generation == imp->imp_last_generation_checked) {
1995 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1996 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2000 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2001 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2002 imp->imp_generation);
2003 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2004 imp->imp_last_generation_checked = imp->imp_generation;
2006 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2007 req = cfs_list_entry(tmp, struct ptlrpc_request,
2010 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2011 LASSERT(req != last_req);
2014 if (req->rq_transno == 0) {
2015 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2018 if (req->rq_import_generation < imp->imp_generation) {
2019 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2023 if (req->rq_replay) {
2024 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2028 /* not yet committed */
2029 if (req->rq_transno > imp->imp_peer_committed_transno) {
2030 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2034 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
2035 imp->imp_peer_committed_transno);
2037 cfs_spin_lock(&req->rq_lock);
2039 cfs_spin_unlock(&req->rq_lock);
2040 if (req->rq_commit_cb != NULL)
2041 req->rq_commit_cb(req);
2042 cfs_list_del_init(&req->rq_replay_list);
2043 __ptlrpc_req_finished(req, 1);
2050 void ptlrpc_cleanup_client(struct obd_import *imp)
2057 void ptlrpc_resend_req(struct ptlrpc_request *req)
2059 DEBUG_REQ(D_HA, req, "going to resend");
2060 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2061 req->rq_status = -EAGAIN;
2063 cfs_spin_lock(&req->rq_lock);
2065 req->rq_net_err = 0;
2066 req->rq_timedout = 0;
2068 __u64 old_xid = req->rq_xid;
2070 /* ensure previous bulk fails */
2071 req->rq_xid = ptlrpc_next_xid();
2072 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2073 old_xid, req->rq_xid);
2075 ptlrpc_client_wake_req(req);
2076 cfs_spin_unlock(&req->rq_lock);
2079 /* XXX: this function and rq_status are currently unused */
2080 void ptlrpc_restart_req(struct ptlrpc_request *req)
2082 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2083 req->rq_status = -ERESTARTSYS;
2085 cfs_spin_lock(&req->rq_lock);
2086 req->rq_restart = 1;
2087 req->rq_timedout = 0;
2088 ptlrpc_client_wake_req(req);
2089 cfs_spin_unlock(&req->rq_lock);
2092 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2095 cfs_atomic_inc(&req->rq_refcount);
2099 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2100 struct obd_import *imp)
2104 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2106 if (req->rq_transno == 0) {
2107 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2111 /* clear this for new requests that were resent as well
2112 as resent replayed requests. */
2113 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2115 /* don't re-add requests that have been replayed */
2116 if (!cfs_list_empty(&req->rq_replay_list))
2119 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2121 LASSERT(imp->imp_replayable);
2122 /* Balanced in ptlrpc_free_committed, usually. */
2123 ptlrpc_request_addref(req);
2124 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2125 struct ptlrpc_request *iter =
2126 cfs_list_entry(tmp, struct ptlrpc_request,
2129 /* We may have duplicate transnos if we create and then
2130 * open a file, or for closes retained if to match creating
2131 * opens, so use req->rq_xid as a secondary key.
2132 * (See bugs 684, 685, and 428.)
2133 * XXX no longer needed, but all opens need transnos!
2135 if (iter->rq_transno > req->rq_transno)
2138 if (iter->rq_transno == req->rq_transno) {
2139 LASSERT(iter->rq_xid != req->rq_xid);
2140 if (iter->rq_xid > req->rq_xid)
2144 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2148 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2151 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2153 struct ptlrpc_request_set *set;
2157 LASSERT(req->rq_set == NULL);
2158 LASSERT(!req->rq_receiving_reply);
2160 set = ptlrpc_prep_set();
2162 CERROR("Unable to allocate ptlrpc set.");
2166 /* for distributed debugging */
2167 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2169 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2170 ptlrpc_request_addref(req);
2171 ptlrpc_set_add_req(set, req);
2172 rc = ptlrpc_set_wait(set);
2173 ptlrpc_set_destroy(set);
2178 struct ptlrpc_replay_async_args {
2180 int praa_old_status;
2183 static int ptlrpc_replay_interpret(const struct lu_env *env,
2184 struct ptlrpc_request *req,
2185 void * data, int rc)
2187 struct ptlrpc_replay_async_args *aa = data;
2188 struct obd_import *imp = req->rq_import;
2191 cfs_atomic_dec(&imp->imp_replay_inflight);
2193 if (!ptlrpc_client_replied(req)) {
2194 CERROR("request replay timed out, restarting recovery\n");
2195 GOTO(out, rc = -ETIMEDOUT);
2198 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2199 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2200 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2201 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2203 /** VBR: check version failure */
2204 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2205 /** replay was failed due to version mismatch */
2206 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2207 cfs_spin_lock(&imp->imp_lock);
2208 imp->imp_vbr_failed = 1;
2209 imp->imp_no_lock_replay = 1;
2210 cfs_spin_unlock(&imp->imp_lock);
2212 /** The transno had better not change over replay. */
2213 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2214 lustre_msg_get_transno(req->rq_repmsg) ||
2215 lustre_msg_get_transno(req->rq_repmsg) == 0,
2217 lustre_msg_get_transno(req->rq_reqmsg),
2218 lustre_msg_get_transno(req->rq_repmsg));
2221 cfs_spin_lock(&imp->imp_lock);
2222 /** if replays by version then gap was occur on server, no trust to locks */
2223 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2224 imp->imp_no_lock_replay = 1;
2225 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2226 cfs_spin_unlock(&imp->imp_lock);
2227 LASSERT(imp->imp_last_replay_transno);
2229 DEBUG_REQ(D_HA, req, "got rep");
2231 /* let the callback do fixups, possibly including in the request */
2232 if (req->rq_replay_cb)
2233 req->rq_replay_cb(req);
2235 if (ptlrpc_client_replied(req) &&
2236 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2237 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2238 lustre_msg_get_status(req->rq_repmsg),
2239 aa->praa_old_status);
2241 /* Put it back for re-replay. */
2242 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2246 * Errors while replay can set transno to 0, but
2247 * imp_last_replay_transno shouldn't be set to 0 anyway
2249 if (req->rq_transno > 0) {
2250 cfs_spin_lock(&imp->imp_lock);
2251 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2252 imp->imp_last_replay_transno = req->rq_transno;
2253 cfs_spin_unlock(&imp->imp_lock);
2255 CERROR("Transno is 0 during replay!\n");
2256 /* continue with recovery */
2257 rc = ptlrpc_import_recovery_state_machine(imp);
2259 req->rq_send_state = aa->praa_old_state;
2262 /* this replay failed, so restart recovery */
2263 ptlrpc_connect_import(imp, NULL);
2268 int ptlrpc_replay_req(struct ptlrpc_request *req)
2270 struct ptlrpc_replay_async_args *aa;
2273 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2274 /* Not handling automatic bulk replay yet (or ever?) */
2275 LASSERT(req->rq_bulk == NULL);
2277 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2278 aa = ptlrpc_req_async_args(req);
2279 memset(aa, 0, sizeof *aa);
2281 /* Prepare request to be resent with ptlrpcd */
2282 aa->praa_old_state = req->rq_send_state;
2283 req->rq_send_state = LUSTRE_IMP_REPLAY;
2284 req->rq_phase = RQ_PHASE_NEW;
2285 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2287 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2289 req->rq_interpret_reply = ptlrpc_replay_interpret;
2290 /* Readjust the timeout for current conditions */
2291 ptlrpc_at_set_req_timeout(req);
2293 DEBUG_REQ(D_HA, req, "REPLAY");
2295 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2296 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2298 ptlrpcd_add_req(req, PSCOPE_OTHER);
2302 void ptlrpc_abort_inflight(struct obd_import *imp)
2304 cfs_list_t *tmp, *n;
2307 /* Make sure that no new requests get processed for this import.
2308 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2309 * this flag and then putting requests on sending_list or delayed_list.
2311 cfs_spin_lock(&imp->imp_lock);
2313 /* XXX locking? Maybe we should remove each request with the list
2314 * locked? Also, how do we know if the requests on the list are
2315 * being freed at this time?
2317 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2318 struct ptlrpc_request *req =
2319 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2321 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2323 cfs_spin_lock (&req->rq_lock);
2324 if (req->rq_import_generation < imp->imp_generation) {
2326 req->rq_status = -EINTR;
2327 ptlrpc_client_wake_req(req);
2329 cfs_spin_unlock (&req->rq_lock);
2332 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2333 struct ptlrpc_request *req =
2334 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2336 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2338 cfs_spin_lock (&req->rq_lock);
2339 if (req->rq_import_generation < imp->imp_generation) {
2341 req->rq_status = -EINTR;
2342 ptlrpc_client_wake_req(req);
2344 cfs_spin_unlock (&req->rq_lock);
2347 /* Last chance to free reqs left on the replay list, but we
2348 * will still leak reqs that haven't committed. */
2349 if (imp->imp_replayable)
2350 ptlrpc_free_committed(imp);
2352 cfs_spin_unlock(&imp->imp_lock);
2357 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2359 cfs_list_t *tmp, *pos;
2361 LASSERT(set != NULL);
2363 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2364 struct ptlrpc_request *req =
2365 cfs_list_entry(pos, struct ptlrpc_request,
2368 cfs_spin_lock(&req->rq_lock);
2369 if (req->rq_phase != RQ_PHASE_RPC) {
2370 cfs_spin_unlock(&req->rq_lock);
2375 req->rq_status = -EINTR;
2376 ptlrpc_client_wake_req(req);
2377 cfs_spin_unlock(&req->rq_lock);
2381 static __u64 ptlrpc_last_xid;
2382 static cfs_spinlock_t ptlrpc_last_xid_lock;
2384 /* Initialize the XID for the node. This is common among all requests on
2385 * this node, and only requires the property that it is monotonically
2386 * increasing. It does not need to be sequential. Since this is also used
2387 * as the RDMA match bits, it is important that a single client NOT have
2388 * the same match bits for two different in-flight requests, hence we do
2389 * NOT want to have an XID per target or similar.
2391 * To avoid an unlikely collision between match bits after a client reboot
2392 * (which would deliver old data into the wrong RDMA buffer) initialize
2393 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2394 * If the time is clearly incorrect, we instead use a 62-bit random number.
2395 * In the worst case the random number will overflow 1M RPCs per second in
2396 * 9133 years, or permutations thereof.
2398 #define YEAR_2004 (1ULL << 30)
2399 void ptlrpc_init_xid(void)
2401 time_t now = cfs_time_current_sec();
2403 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2404 if (now < YEAR_2004) {
2405 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2406 ptlrpc_last_xid >>= 2;
2407 ptlrpc_last_xid |= (1ULL << 61);
2409 ptlrpc_last_xid = (__u64)now << 20;
2413 __u64 ptlrpc_next_xid(void)
2416 cfs_spin_lock(&ptlrpc_last_xid_lock);
2417 tmp = ++ptlrpc_last_xid;
2418 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2422 __u64 ptlrpc_sample_next_xid(void)
2424 #if BITS_PER_LONG == 32
2425 /* need to avoid possible word tearing on 32-bit systems */
2427 cfs_spin_lock(&ptlrpc_last_xid_lock);
2428 tmp = ptlrpc_last_xid + 1;
2429 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2432 /* No need to lock, since returned value is racy anyways */
2433 return ptlrpc_last_xid + 1;
2436 EXPORT_SYMBOL(ptlrpc_sample_next_xid);