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>
50 #include "ptlrpc_internal.h"
52 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
53 struct ptlrpc_client *cl)
55 cl->cli_request_portal = req_portal;
56 cl->cli_reply_portal = rep_portal;
60 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
62 struct ptlrpc_connection *c;
64 lnet_process_id_t peer;
67 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
69 CERROR("cannot find peer %s!\n", uuid->uuid);
73 c = ptlrpc_connection_get(peer, self, uuid);
75 memcpy(c->c_remote_uuid.uuid,
76 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
79 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
84 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
86 struct ptlrpc_bulk_desc *desc;
88 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
92 spin_lock_init(&desc->bd_lock);
93 cfs_waitq_init(&desc->bd_waitq);
94 desc->bd_max_iov = npages;
95 desc->bd_iov_count = 0;
96 desc->bd_md_h = LNET_INVALID_HANDLE;
97 desc->bd_portal = portal;
103 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
104 int npages, int type, int portal)
106 struct obd_import *imp = req->rq_import;
107 struct ptlrpc_bulk_desc *desc;
110 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
111 desc = new_bulk(npages, type, portal);
115 desc->bd_import_generation = req->rq_import_generation;
116 desc->bd_import = class_import_get(imp);
119 desc->bd_cbid.cbid_fn = client_bulk_callback;
120 desc->bd_cbid.cbid_arg = desc;
122 /* This makes req own desc, and free it when she frees herself */
128 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
129 int npages, int type, int portal)
131 struct obd_export *exp = req->rq_export;
132 struct ptlrpc_bulk_desc *desc;
135 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
137 desc = new_bulk(npages, type, portal);
141 desc->bd_export = class_export_get(exp);
144 desc->bd_cbid.cbid_fn = server_bulk_callback;
145 desc->bd_cbid.cbid_arg = desc;
147 /* NB we don't assign rq_bulk here; server-side requests are
148 * re-used, and the handler frees the bulk desc explicitly. */
153 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
154 cfs_page_t *page, int pageoffset, int len)
156 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
157 LASSERT(page != NULL);
158 LASSERT(pageoffset >= 0);
160 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)
173 LASSERT(desc != NULL);
174 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
175 LASSERT(!desc->bd_network_rw); /* network hands off or */
176 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
178 class_export_put(desc->bd_export);
180 class_import_put(desc->bd_import);
182 for (i = 0; i < desc->bd_iov_count ; i++)
183 cfs_page_unpin(desc->bd_iov[i].kiov_page);
185 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
186 bd_iov[desc->bd_max_iov]));
190 /* Set server timelimit for this req */
191 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
197 LASSERT(req->rq_import);
200 /* non-AT settings */
201 req->rq_timeout = req->rq_import->imp_server_timeout ?
202 obd_timeout / 2 : obd_timeout;
203 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
207 at = &req->rq_import->imp_at;
208 idx = import_at_get_index(req->rq_import,
209 req->rq_request_portal);
210 serv_est = at_get(&at->iat_service_estimate[idx]);
211 req->rq_timeout = at_est2timeout(serv_est);
212 /* We could get even fancier here, using history to predict increased
215 /* Let the server know what this RPC timeout is by putting it in the
217 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
220 /* Adjust max service estimate based on server value */
221 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
222 unsigned int serv_est)
228 /* do estimate only if is not in recovery */
229 if ((req->rq_send_state != LUSTRE_IMP_FULL) &&
230 (req->rq_send_state != LUSTRE_IMP_CONNECTING))
233 LASSERT(req->rq_import);
234 at = &req->rq_import->imp_at;
236 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
237 /* max service estimates are tracked on the server side,
238 so just keep minimal history here */
239 oldse = at_add(&at->iat_service_estimate[idx], serv_est);
241 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
242 "has changed from %d to %d\n",
243 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
244 oldse, at_get(&at->iat_service_estimate[idx]));
247 /* Expected network latency per remote node (secs) */
248 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
250 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
253 /* Adjust expected network latency */
254 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
255 unsigned int service_time)
257 unsigned int nl, oldnl;
259 time_t now = cfs_time_current_sec();
261 LASSERT(req->rq_import);
262 at = &req->rq_import->imp_at;
264 /* Network latency is total time less server processing time */
265 nl = max_t(int, now - req->rq_sent - service_time, 0) + 1/*st rounding*/;
266 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
267 CWARN("Reported service time %u > total measured time %ld\n",
268 service_time, now - req->rq_sent);
270 oldnl = at_add(&at->iat_net_latency, nl);
272 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
273 "has changed from %d to %d\n",
274 req->rq_import->imp_obd->obd_name,
276 &req->rq_import->imp_connection->c_remote_uuid),
277 oldnl, at_get(&at->iat_net_latency));
280 static int unpack_reply_common(struct ptlrpc_request *req)
284 req->rq_rep_swab_mask = 0;
285 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
287 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
292 lustre_set_rep_swabbed(req, MSG_PTLRPC_HEADER_OFF);
297 static int unpack_reply(struct ptlrpc_request *req)
301 rc = unpack_reply_common(req);
305 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
307 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
313 static inline void unpack_reply_free_msg(struct lustre_msg *msg, int len)
318 static int unpack_reply_copy_msg(struct ptlrpc_request *req,
319 struct lustre_msg **msg, int *len)
321 struct lustre_msg *msgcpy;
322 __u32 csum_calc, csum_get;
326 LASSERT_SPIN_LOCKED(&req->rq_lock);
330 /* Swabbing required when rc == 1 */
331 rc = unpack_reply_common(req);
335 lencpy = req->rq_replen;
336 spin_unlock(&req->rq_lock);
338 OBD_ALLOC(msgcpy, lencpy);
340 spin_lock(&req->rq_lock);
343 spin_lock(&req->rq_lock);
345 /* Checksum must be calculated before being unswabbed. If the magic
346 * in the copy is unswabbed discard like the checksum failure case */
347 memcpy(msgcpy, req->rq_repmsg, lencpy);
348 if (lustre_msg_need_swab(msgcpy)) {
349 DEBUG_REQ(D_NET, req, "incorrect message magic: %08x\n",
351 GOTO(err, rc = -EINVAL);
354 csum_calc = lustre_msg_calc_cksum(msgcpy);
356 /* Unpack the copy the original rq_repmsg is untouched */
357 rc = lustre_unpack_msg_ptlrpc_body(msgcpy, MSG_PTLRPC_BODY_OFF, rc);
359 DEBUG_REQ(D_ERROR, req, "unpack msg copy failed: %d", rc);
360 GOTO(err, rc = -EPROTO);
363 /* For early replies the LND may update repmsg outside req->rq_lock
364 * resulting in a checksum failure which is non-harmful */
365 csum_get = lustre_msg_get_cksum(msgcpy);
366 if (csum_calc != csum_get) {
367 DEBUG_REQ(D_NET, req, "checksum mismatch: %x != %x\n",
368 csum_calc, csum_get);
369 GOTO(err, rc = -EINVAL);
376 unpack_reply_free_msg(msgcpy, lencpy);
380 /* Handle an early reply message. To prevent a real reply from arriving
381 * and changing req->rq_repmsg this func is called under the rq_lock */
382 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req) {
383 struct lustre_msg *msg;
388 LASSERT_SPIN_LOCKED(&req->rq_lock);
391 /* All early replys for this request use a single repbuf which can
392 * be updated outside the req->rq_lock. To prevent racing we create
393 * a copy of the repmsg and verify its checksum before it is used. */
394 rc = unpack_reply_copy_msg(req, &msg, &len);
396 /* Let's just ignore it - same as if it never got here */
397 CDEBUG(D_ADAPTTO, "Discarding racing early reply: %d\n", rc);
401 /* Expecting to increase the service time estimate here */
402 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(msg));
403 ptlrpc_at_adj_net_latency(req, lustre_msg_get_service_time(msg));
405 /* Adjust the local timeout for this req */
406 ptlrpc_at_set_req_timeout(req);
408 olddl = req->rq_deadline;
409 /* Server assumes it now has rq_timeout from when it sent the
410 early reply, so client should give it at least that long. */
411 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
412 ptlrpc_at_get_net_latency(req);
414 DEBUG_REQ(D_ADAPTTO, req,
415 "Early reply #%d, new deadline in %lds (%+lds)",
416 req->rq_early_count, req->rq_deadline -
417 cfs_time_current_sec(), req->rq_deadline - olddl);
419 unpack_reply_free_msg(msg, len);
423 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
425 struct list_head *l, *tmp;
426 struct ptlrpc_request *req;
428 LASSERT(pool != NULL);
430 spin_lock(&pool->prp_lock);
431 list_for_each_safe(l, tmp, &pool->prp_req_list) {
432 req = list_entry(l, struct ptlrpc_request, rq_list);
433 list_del(&req->rq_list);
434 LASSERT(req->rq_reqmsg);
435 OBD_FREE(req->rq_reqmsg, pool->prp_rq_size);
436 OBD_FREE(req, sizeof(*req));
438 spin_unlock(&pool->prp_lock);
439 OBD_FREE(pool, sizeof(*pool));
442 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
447 while (size < pool->prp_rq_size)
450 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
451 "Trying to change pool size with nonempty pool "
452 "from %d to %d bytes\n", pool->prp_rq_size, size);
454 spin_lock(&pool->prp_lock);
455 pool->prp_rq_size = size;
456 for (i = 0; i < num_rq; i++) {
457 struct ptlrpc_request *req;
458 struct lustre_msg *msg;
460 spin_unlock(&pool->prp_lock);
461 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
464 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
466 OBD_FREE(req, sizeof(struct ptlrpc_request));
469 req->rq_reqmsg = msg;
471 spin_lock(&pool->prp_lock);
472 list_add_tail(&req->rq_list, &pool->prp_req_list);
474 spin_unlock(&pool->prp_lock);
478 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
479 void (*populate_pool)(struct ptlrpc_request_pool *, int))
481 struct ptlrpc_request_pool *pool;
483 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
487 /* Request next power of two for the allocation, because internally
488 kernel would do exactly this */
490 spin_lock_init(&pool->prp_lock);
491 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
492 pool->prp_rq_size = msgsize;
493 pool->prp_populate = populate_pool;
495 populate_pool(pool, num_rq);
497 if (list_empty(&pool->prp_req_list)) {
498 /* have not allocated a single request for the pool */
499 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
505 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
507 struct ptlrpc_request *request;
508 struct lustre_msg *reqmsg;
513 spin_lock(&pool->prp_lock);
515 /* See if we have anything in a pool, and bail out if nothing,
516 * in writeout path, where this matters, this is safe to do, because
517 * nothing is lost in this case, and when some in-flight requests
518 * complete, this code will be called again. */
519 if (unlikely(list_empty(&pool->prp_req_list))) {
520 spin_unlock(&pool->prp_lock);
524 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
526 list_del_init(&request->rq_list);
527 spin_unlock(&pool->prp_lock);
529 LASSERT(request->rq_reqmsg);
530 LASSERT(request->rq_pool);
532 reqmsg = request->rq_reqmsg;
533 memset(request, 0, sizeof(*request));
534 request->rq_reqmsg = reqmsg;
535 request->rq_pool = pool;
536 request->rq_reqlen = pool->prp_rq_size;
540 struct ptlrpc_request *
541 ptlrpc_prep_req_pool(struct obd_import *imp, __u32 version, int opcode,
542 int count, __u32 *lengths, char **bufs,
543 struct ptlrpc_request_pool *pool)
545 struct ptlrpc_request *request = NULL;
549 /* The obd disconnected */
553 LASSERT(imp != LP_POISON);
554 LASSERT((unsigned long)imp->imp_client > 0x1000);
555 LASSERT(imp->imp_client != LP_POISON);
558 request = ptlrpc_prep_req_from_pool(pool);
561 OBD_ALLOC(request, sizeof(*request));
564 CERROR("request allocation out of memory\n");
568 rc = lustre_pack_request(request, imp->imp_msg_magic, count, lengths,
571 LASSERT(!request->rq_pool);
572 OBD_FREE(request, sizeof(*request));
576 lustre_msg_add_version(request->rq_reqmsg, version);
577 request->rq_send_state = LUSTRE_IMP_FULL;
578 request->rq_type = PTL_RPC_MSG_REQUEST;
579 request->rq_import = class_import_get(imp);
580 request->rq_export = NULL;
582 request->rq_req_cbid.cbid_fn = request_out_callback;
583 request->rq_req_cbid.cbid_arg = request;
585 request->rq_reply_cbid.cbid_fn = reply_in_callback;
586 request->rq_reply_cbid.cbid_arg = request;
588 request->rq_reply_deadline = 0;
589 request->rq_phase = RQ_PHASE_NEW;
590 request->rq_next_phase = RQ_PHASE_UNDEFINED;
592 request->rq_request_portal = imp->imp_client->cli_request_portal;
593 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
595 ptlrpc_at_set_req_timeout(request);
597 spin_lock_init(&request->rq_lock);
598 CFS_INIT_LIST_HEAD(&request->rq_list);
599 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
600 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
601 CFS_INIT_LIST_HEAD(&request->rq_history_list);
602 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
603 cfs_waitq_init(&request->rq_reply_waitq);
604 request->rq_xid = ptlrpc_next_xid();
605 atomic_set(&request->rq_refcount, 1);
607 lustre_msg_set_opc(request->rq_reqmsg, opcode);
612 struct ptlrpc_request *
613 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
614 __u32 *lengths, char **bufs)
616 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
620 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
621 unsigned int timeout,
622 int (*interpreter)(struct ptlrpc_request *,
625 struct ptlrpc_request *request = NULL;
628 OBD_ALLOC(request, sizeof(*request));
630 CERROR("request allocation out of memory\n");
634 request->rq_send_state = LUSTRE_IMP_FULL;
635 request->rq_type = PTL_RPC_MSG_REQUEST;
636 request->rq_import = class_import_get(imp);
637 request->rq_export = NULL;
639 request->rq_sent = cfs_time_current_sec();
640 request->rq_reply_deadline = request->rq_sent + timeout;
641 request->rq_interpret_reply = interpreter;
642 request->rq_phase = RQ_PHASE_RPC;
643 request->rq_next_phase = RQ_PHASE_INTERPRET;
644 /* don't want reply */
645 request->rq_receiving_reply = 0;
646 request->rq_must_unlink = 0;
647 request->rq_no_delay = request->rq_no_resend = 1;
648 request->rq_fake = 1;
650 spin_lock_init(&request->rq_lock);
651 CFS_INIT_LIST_HEAD(&request->rq_list);
652 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
653 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
654 CFS_INIT_LIST_HEAD(&request->rq_history_list);
655 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
656 cfs_waitq_init(&request->rq_reply_waitq);
658 request->rq_xid = ptlrpc_next_xid();
659 atomic_set(&request->rq_refcount, 1);
664 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
666 /* if we kill request before timeout - need adjust counter */
667 if (req->rq_phase == RQ_PHASE_RPC) {
668 struct ptlrpc_request_set *set = req->rq_set;
671 set->set_remaining --;
674 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
675 list_del_init(&req->rq_list);
679 struct ptlrpc_request_set *ptlrpc_prep_set(void)
681 struct ptlrpc_request_set *set;
684 OBD_ALLOC(set, sizeof *set);
687 CFS_INIT_LIST_HEAD(&set->set_requests);
688 cfs_waitq_init(&set->set_waitq);
689 set->set_remaining = 0;
690 spin_lock_init(&set->set_new_req_lock);
691 CFS_INIT_LIST_HEAD(&set->set_new_requests);
692 CFS_INIT_LIST_HEAD(&set->set_cblist);
697 /* Finish with this set; opposite of prep_set. */
698 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
700 struct list_head *tmp;
701 struct list_head *next;
706 /* Requests on the set should either all be completed, or all be new */
707 expected_phase = (set->set_remaining == 0) ?
708 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
709 list_for_each (tmp, &set->set_requests) {
710 struct ptlrpc_request *req =
711 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
713 LASSERT(req->rq_phase == expected_phase);
717 LASSERTF(set->set_remaining == 0 || set->set_remaining == n, "%d / %d\n",
718 set->set_remaining, n);
720 list_for_each_safe(tmp, next, &set->set_requests) {
721 struct ptlrpc_request *req =
722 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
723 list_del_init(&req->rq_set_chain);
725 LASSERT(req->rq_phase == expected_phase);
727 if (req->rq_phase == RQ_PHASE_NEW) {
729 if (req->rq_interpret_reply != NULL) {
730 int (*interpreter)(struct ptlrpc_request *,
732 req->rq_interpret_reply;
734 /* higher level (i.e. LOV) failed;
735 * let the sub reqs clean up */
736 req->rq_status = -EBADR;
737 interpreter(req, &req->rq_async_args,
740 set->set_remaining--;
744 ptlrpc_req_finished (req);
747 LASSERT(set->set_remaining == 0);
749 OBD_FREE(set, sizeof(*set));
753 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
754 set_interpreter_func fn, void *data)
756 struct ptlrpc_set_cbdata *cbdata;
758 OBD_SLAB_ALLOC(cbdata, ptlrpc_cbdata_slab,
759 CFS_ALLOC_STD, sizeof(*cbdata));
763 cbdata->psc_interpret = fn;
764 cbdata->psc_data = data;
765 list_add_tail(&cbdata->psc_item, &set->set_cblist);
770 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
771 struct ptlrpc_request *req)
773 /* The set takes over the caller's request reference */
774 list_add_tail(&req->rq_set_chain, &set->set_requests);
776 set->set_remaining++;
780 * Lock so many callers can add things, the context that owns the set
781 * is supposed to notice these and move them into the set proper.
783 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
784 struct ptlrpc_request *req)
786 struct ptlrpc_request_set *set = pc->pc_set;
789 * Let caller know that we stopped and will not handle this request.
790 * It needs to take care itself of request.
792 if (test_bit(LIOD_STOP, &pc->pc_flags))
795 spin_lock(&set->set_new_req_lock);
797 * The set takes over the caller's request reference.
799 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
801 spin_unlock(&set->set_new_req_lock);
804 * Let thead know that we added something and better it to wake up
807 cfs_waitq_signal(&set->set_waitq);
812 * Based on the current state of the import, determine if the request
813 * can be sent, is an error, or should be delayed.
815 * Returns true if this request should be delayed. If false, and
816 * *status is set, then the request can not be sent and *status is the
817 * error code. If false and status is 0, then request can be sent.
819 * The imp->imp_lock must be held.
821 static int ptlrpc_import_delay_req(struct obd_import *imp,
822 struct ptlrpc_request *req, int *status)
827 LASSERT (status != NULL);
830 if (imp->imp_state == LUSTRE_IMP_NEW) {
831 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
834 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
835 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
837 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
838 imp->imp_state == LUSTRE_IMP_CONNECTING) {
839 /* allow CONNECT even if import is invalid */ ;
840 if (atomic_read(&imp->imp_inval_count) != 0) {
841 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
844 } else if ((imp->imp_invalid && (!imp->imp_recon_bk)) ||
845 imp->imp_obd->obd_no_recov) {
846 /* If the import has been invalidated (such as by an OST
847 * failure), and if the import(MGC) tried all of its connection
848 * list (Bug 13464), the request must fail with -ESHUTDOWN.
849 * This indicates the requests should be discarded; an -EIO
850 * may result in a resend of the request. */
851 if (!imp->imp_deactive)
852 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
853 *status = -ESHUTDOWN; /* bz 12940 */
854 } else if (req->rq_import_generation != imp->imp_generation) {
855 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
857 } else if (req->rq_send_state != imp->imp_state) {
858 /* invalidate in progress - any requests should be drop */
859 if (atomic_read(&imp->imp_inval_count) != 0) {
860 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
862 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
863 *status = -EWOULDBLOCK;
872 static int ptlrpc_check_reply(struct ptlrpc_request *req)
875 const char *what = "";
878 /* serialise with network callback */
879 spin_lock(&req->rq_lock);
881 if (ptlrpc_client_replied(req)) {
886 if (req->rq_net_err && !req->rq_timedout) {
888 spin_unlock(&req->rq_lock);
889 rc = ptlrpc_expire_one_request(req, 0);
890 spin_lock(&req->rq_lock);
899 if (req->rq_resend) {
904 if (req->rq_restart) {
909 if (ptlrpc_client_early(req)) {
911 ptlrpc_at_recv_early_reply(req);
912 GOTO(out, rc = 0); /* keep waiting */
917 spin_unlock(&req->rq_lock);
918 DEBUG_REQ(D_NET, req, "%src = %d for", what, rc);
922 /* Conditionally suppress specific console messages */
923 static int ptlrpc_console_allow(struct ptlrpc_request *req)
925 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
928 /* Suppress particular reconnect errors which are to be expected. No
929 * errors are suppressed for the initial connection on an import */
930 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
931 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
933 /* Suppress timed out reconnect requests */
934 if (req->rq_timedout)
937 /* Suppress unavailable/again reconnect requests */
938 err = lustre_msg_get_status(req->rq_repmsg);
939 if (err == -ENODEV || err == -EAGAIN)
946 static int ptlrpc_check_status(struct ptlrpc_request *req)
951 err = lustre_msg_get_status(req->rq_repmsg);
953 DEBUG_REQ(D_INFO, req, "status is %d", err);
954 } else if (err > 0) {
955 /* XXX: translate this error from net to host */
956 DEBUG_REQ(D_INFO, req, "status is %d", err);
959 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
960 struct obd_import *imp = req->rq_import;
961 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
963 if (ptlrpc_console_allow(req))
964 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
965 "communicating with %s. The %s "
966 "operation failed with %d\n",
968 imp->imp_connection->c_peer.nid),
969 ll_opcode2str(opc), err);
971 RETURN(err < 0 ? err : -EINVAL);
977 /* VBR: we should save pre-versions for replay*/
978 static void ptlrpc_save_versions(struct ptlrpc_request *req)
980 struct lustre_msg *repmsg = req->rq_repmsg;
981 struct lustre_msg *reqmsg = req->rq_reqmsg;
982 __u64 *versions = lustre_msg_get_versions(repmsg);
984 /* Interoperability with 1.6. This should be changed to LASSERT in HEAD */
985 if (versions == NULL)
988 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
991 lustre_msg_set_versions(reqmsg, versions);
992 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
993 versions[0], versions[1]);
998 static int after_reply(struct ptlrpc_request *req)
1000 struct obd_import *imp = req->rq_import;
1001 struct obd_device *obd = req->rq_import->imp_obd;
1003 struct timeval work_start;
1007 LASSERT(!req->rq_receiving_reply);
1010 /* NB Until this point, the whole of the incoming message,
1011 * including buflens, status etc is in the sender's byte order. */
1013 if (req->rq_reply_truncate && !req->rq_no_resend) {
1015 OBD_FREE(req->rq_repbuf, req->rq_replen);
1016 req->rq_repbuf = NULL;
1017 req->rq_replen = req->rq_nob_received;
1021 LASSERT (req->rq_nob_received <= req->rq_replen);
1022 rc = unpack_reply(req);
1026 do_gettimeofday(&work_start);
1027 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1028 if (obd->obd_svc_stats != NULL) {
1029 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1031 ptlrpc_lprocfs_rpc_sent(req, timediff);
1034 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1035 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1036 ptlrpc_at_adj_net_latency(req, lustre_msg_get_service_time(req->rq_repmsg));
1038 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1039 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1040 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1041 lustre_msg_get_type(req->rq_repmsg));
1045 rc = ptlrpc_check_status(req);
1047 /* Either we've been evicted, or the server has failed for
1048 * some reason. Try to reconnect, and if that fails, punt to
1050 if (ll_rpc_recoverable_error(rc)) {
1051 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1052 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1055 ptlrpc_request_handle_notconn(req);
1059 /* Let's look if server sent slv. Do it only for RPC with
1061 ldlm_cli_update_pool(req);
1064 /* Store transno in reqmsg for replay. */
1065 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1066 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1067 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1070 if (imp->imp_replayable) {
1071 spin_lock(&imp->imp_lock);
1072 /* no point in adding already-committed requests to the replay
1073 * list, we will just remove them immediately. b=9829 */
1074 if (req->rq_transno != 0 &&
1076 lustre_msg_get_last_committed(req->rq_repmsg) ||
1078 /* version recovery */
1079 ptlrpc_save_versions(req);
1080 ptlrpc_retain_replayable_request(req, imp);
1081 } else if (req->rq_commit_cb != NULL) {
1082 spin_unlock(&imp->imp_lock);
1083 req->rq_commit_cb(req);
1084 spin_lock(&imp->imp_lock);
1087 /* Replay-enabled imports return commit-status information. */
1088 if (lustre_msg_get_last_committed(req->rq_repmsg))
1089 imp->imp_peer_committed_transno =
1090 lustre_msg_get_last_committed(req->rq_repmsg);
1091 ptlrpc_free_committed(imp);
1093 if (req->rq_transno > imp->imp_peer_committed_transno)
1094 ptlrpc_pinger_sending_on_import(imp);
1095 spin_unlock(&imp->imp_lock);
1101 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1103 struct obd_import *imp;
1107 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1108 if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
1111 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1113 imp = req->rq_import;
1114 spin_lock(&imp->imp_lock);
1116 req->rq_import_generation = imp->imp_generation;
1118 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1119 spin_lock(&req->rq_lock);
1120 req->rq_waiting = 1;
1121 spin_unlock(&req->rq_lock);
1123 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1124 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1125 ptlrpc_import_state_name(req->rq_send_state),
1126 ptlrpc_import_state_name(imp->imp_state));
1127 LASSERT(list_empty(&req->rq_list));
1128 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1129 atomic_inc(&req->rq_import->imp_inflight);
1130 spin_unlock(&imp->imp_lock);
1135 spin_unlock(&imp->imp_lock);
1136 req->rq_status = rc;
1137 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1141 LASSERT(list_empty(&req->rq_list));
1142 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1143 atomic_inc(&req->rq_import->imp_inflight);
1144 spin_unlock(&imp->imp_lock);
1146 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1147 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1148 " %s:%s:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1149 imp->imp_obd->obd_uuid.uuid,
1150 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1151 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1152 lustre_msg_get_opc(req->rq_reqmsg));
1154 rc = ptl_send_rpc(req, 0);
1156 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1157 req->rq_net_err = 1;
1163 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
1164 int ptlrpc_check_set(struct ptlrpc_request_set *set)
1166 struct list_head *tmp;
1167 int force_timer_recalc = 0;
1170 if (set->set_remaining == 0)
1173 list_for_each(tmp, &set->set_requests) {
1174 struct ptlrpc_request *req =
1175 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1176 struct obd_import *imp = req->rq_import;
1179 if (req->rq_phase == RQ_PHASE_NEW &&
1180 ptlrpc_send_new_req(req)) {
1181 force_timer_recalc = 1;
1184 /* delayed send - skip */
1185 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1188 if (!(req->rq_phase == RQ_PHASE_RPC ||
1189 req->rq_phase == RQ_PHASE_BULK ||
1190 req->rq_phase == RQ_PHASE_INTERPRET ||
1191 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1192 req->rq_phase == RQ_PHASE_COMPLETE)) {
1193 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1197 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1198 LASSERT(req->rq_next_phase != req->rq_phase);
1199 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1201 /* Skip processing until reply is unlinked. We
1202 * can't return to pool before that and we can't
1203 * call interpret before that. We need to make
1204 * sure that all rdma transfers finished and will
1205 * not corrupt any data. */
1206 if (ptlrpc_client_recv_or_unlink(req) ||
1207 ptlrpc_client_bulk_active(req))
1210 /* Turn repl fail_loc off to prevent it from looping
1212 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1213 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK |
1217 /* Turn off bulk fail_loc. */
1218 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1219 OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK |
1223 /* Move to next phase if reply was successfully
1225 ptlrpc_rqphase_move(req, req->rq_next_phase);
1228 if (req->rq_phase == RQ_PHASE_COMPLETE)
1231 if (req->rq_phase == RQ_PHASE_INTERPRET)
1232 GOTO(interpret, req->rq_status);
1234 /* Note that this also will start async reply unlink */
1235 if (req->rq_net_err && !req->rq_timedout) {
1236 ptlrpc_expire_one_request(req, 1);
1238 /* Check if we still need to wait for unlink. */
1239 if (ptlrpc_client_recv_or_unlink(req) ||
1240 ptlrpc_client_bulk_active(req))
1245 if (req->rq_status == 0)
1246 req->rq_status = -EIO;
1247 GOTO(interpret, req->rq_status);
1250 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
1251 * will only be set after rq_timedout, but the oig waiting
1252 * path sets rq_intr irrespective of whether ptlrpcd has
1253 * seen a timeout. our policy is to only interpret
1254 * interrupted rpcs after they have timed out */
1255 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
1256 req->rq_status = -EINTR;
1257 GOTO(interpret, req->rq_status);
1260 if (req->rq_phase == RQ_PHASE_RPC) {
1261 if (req->rq_timedout||req->rq_waiting||req->rq_resend) {
1264 if (!ptlrpc_unregister_reply(req, 1))
1267 spin_lock(&imp->imp_lock);
1268 if (ptlrpc_import_delay_req(imp, req, &status)){
1269 /* put on delay list - only if we wait
1270 * recovery finished - before send */
1271 list_del_init(&req->rq_list);
1272 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1273 spin_unlock(&imp->imp_lock);
1278 req->rq_status = status;
1279 spin_unlock(&imp->imp_lock);
1280 GOTO(interpret, req->rq_status);
1282 if (req->rq_no_resend) {
1283 req->rq_status = -ENOTCONN;
1284 spin_unlock(&imp->imp_lock);
1285 GOTO(interpret, req->rq_status);
1288 list_del_init(&req->rq_list);
1289 list_add_tail(&req->rq_list,
1290 &imp->imp_sending_list);
1292 spin_unlock(&imp->imp_lock);
1294 req->rq_waiting = 0;
1296 if (req->rq_timedout||req->rq_resend) {
1297 /* This is re-sending anyways,
1298 * let's mark req as resend. */
1303 if (!ptlrpc_unregister_bulk(req, 1))
1306 /* ensure previous bulk fails */
1307 old_xid = req->rq_xid;
1308 req->rq_xid = ptlrpc_next_xid();
1309 CDEBUG(D_HA, "resend bulk "
1312 old_xid, req->rq_xid);
1316 rc = ptl_send_rpc(req, 0);
1318 DEBUG_REQ(D_HA, req, "send failed (%d)",
1320 force_timer_recalc = 1;
1321 req->rq_net_err = 1;
1323 /* need to reset the timeout */
1324 force_timer_recalc = 1;
1327 spin_lock(&req->rq_lock);
1329 if (ptlrpc_client_early(req)) {
1330 ptlrpc_at_recv_early_reply(req);
1331 spin_unlock(&req->rq_lock);
1335 /* Still waiting for a reply? */
1336 if (ptlrpc_client_recv(req)) {
1337 spin_unlock(&req->rq_lock);
1341 /* Did we actually receive a reply? */
1342 if (!ptlrpc_client_replied(req)) {
1343 spin_unlock(&req->rq_lock);
1347 spin_unlock(&req->rq_lock);
1349 req->rq_status = after_reply(req);
1353 /* If there is no bulk associated with this request,
1354 * then we're done and should let the interpreter
1355 * process the reply. Similarly if the RPC returned
1356 * an error, and therefore the bulk will never arrive.
1358 if (req->rq_bulk == NULL || req->rq_status != 0)
1359 GOTO(interpret, req->rq_status);
1361 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1364 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1365 if (ptlrpc_client_bulk_active(req))
1368 if (!req->rq_bulk->bd_success) {
1369 /* The RPC reply arrived OK, but the bulk screwed
1370 * up! Dead wierd since the server told us the RPC
1371 * was good after getting the REPLY for her GET or
1372 * the ACK for her PUT. */
1373 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1374 req->rq_status = -EIO;
1375 GOTO(interpret, req->rq_status);
1378 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1380 /* This moves to "unregistering" phase we need to wait for
1382 if (!ptlrpc_unregister_reply(req, 1))
1385 if (!ptlrpc_unregister_bulk(req, 1))
1388 /* When calling interpret receiving already should be
1390 LASSERT(!req->rq_receiving_reply);
1392 ptlrpc_req_interpret(req, req->rq_status);
1394 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1396 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1397 "opc %s:%s:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
1398 imp->imp_obd->obd_uuid.uuid,
1399 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1401 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1402 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1404 spin_lock(&imp->imp_lock);
1405 /* Request already may be not on sending or delaying list. This
1406 * may happen in the case of marking it errorneous for the case
1407 * ptlrpc_import_delay_req(req, status) find it impossible to
1408 * allow sending this rpc and returns *status != 0. */
1409 if (!list_empty(&req->rq_list)) {
1410 list_del_init(&req->rq_list);
1411 atomic_dec(&imp->imp_inflight);
1413 spin_unlock(&imp->imp_lock);
1415 set->set_remaining--;
1416 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1419 /* If we hit an error, we want to recover promptly. */
1420 RETURN(set->set_remaining == 0 || force_timer_recalc);
1423 /* Return 1 if we should give up, else 0 */
1424 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1426 struct obd_import *imp = req->rq_import;
1430 DEBUG_REQ(D_WARNING, req, "Request x"LPU64" sent from %s to NID %s"
1431 " %lus ago has %s (limit %lus).\n", req->rq_xid,
1432 imp ? imp->imp_obd->obd_name : "<?>",
1433 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1434 cfs_time_current_sec() - req->rq_sent,
1435 req->rq_net_err ? "failed due to network error" : "timed out",
1436 req->rq_deadline - req->rq_sent);
1438 if (imp != NULL && obd_debug_peer_on_timeout)
1439 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1441 spin_lock(&req->rq_lock);
1442 req->rq_timedout = 1;
1443 spin_unlock(&req->rq_lock);
1445 ptlrpc_unregister_reply(req, async_unlink);
1446 ptlrpc_unregister_bulk(req, async_unlink);
1448 if (obd_dump_on_timeout)
1449 libcfs_debug_dumplog();
1452 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1459 atomic_inc(&imp->imp_timeouts);
1461 /* The DLM server doesn't want recovery run on its imports. */
1462 if (imp->imp_dlm_fake)
1465 /* If this request is for recovery or other primordial tasks,
1466 * then error it out here. */
1467 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1468 imp->imp_obd->obd_no_recov) {
1469 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1470 ptlrpc_import_state_name(req->rq_send_state),
1471 ptlrpc_import_state_name(imp->imp_state));
1472 spin_lock(&req->rq_lock);
1473 req->rq_status = -ETIMEDOUT;
1475 spin_unlock(&req->rq_lock);
1479 /* if a request can't be resent we can't wait for an answer after
1481 if (req->rq_no_resend) {
1482 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1486 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1491 int ptlrpc_expired_set(void *data)
1493 struct ptlrpc_request_set *set = data;
1494 struct list_head *tmp;
1495 time_t now = cfs_time_current_sec();
1498 LASSERT(set != NULL);
1500 /* A timeout expired; see which reqs it applies to... */
1501 list_for_each(tmp, &set->set_requests) {
1502 struct ptlrpc_request *req =
1503 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1505 /* Request in-flight? */
1506 if (!((req->rq_phase == RQ_PHASE_RPC &&
1507 !req->rq_waiting && !req->rq_resend) ||
1508 (req->rq_phase == RQ_PHASE_BULK)))
1511 if (req->rq_timedout || /* already dealt with */
1512 req->rq_deadline > now) /* not expired */
1515 /* Deal with this guy. Do it asynchronously to not block
1516 * ptlrpcd thread. */
1517 ptlrpc_expire_one_request(req, 1);
1520 /* When waiting for a whole set, we always to break out of the
1521 * sleep so we can recalculate the timeout, or enable interrupts
1522 * if everyone's timed out. */
1526 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1528 spin_lock(&req->rq_lock);
1530 spin_unlock(&req->rq_lock);
1533 void ptlrpc_interrupted_set(void *data)
1535 struct ptlrpc_request_set *set = data;
1536 struct list_head *tmp;
1538 LASSERT(set != NULL);
1539 CERROR("INTERRUPTED SET %p\n", set);
1541 list_for_each(tmp, &set->set_requests) {
1542 struct ptlrpc_request *req =
1543 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1545 if (req->rq_phase != RQ_PHASE_RPC &&
1546 req->rq_phase != RQ_PHASE_UNREGISTERING)
1549 ptlrpc_mark_interrupted(req);
1553 /* get the smallest timeout in the set; this does NOT set a timeout. */
1554 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1556 struct list_head *tmp;
1557 time_t now = cfs_time_current_sec();
1559 struct ptlrpc_request *req;
1563 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1565 list_for_each(tmp, &set->set_requests) {
1566 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1568 /* request in-flight? */
1569 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1570 (req->rq_phase == RQ_PHASE_BULK) ||
1571 (req->rq_phase == RQ_PHASE_NEW)))
1574 /* Already timed out. */
1575 if (req->rq_timedout)
1578 if (req->rq_phase == RQ_PHASE_NEW)
1579 deadline = req->rq_sent; /* delayed send */
1581 deadline = req->rq_deadline;
1583 if (deadline <= now) { /* actually expired already */
1584 timeout = 1; /* ASAP */
1587 if ((timeout == 0) || (timeout > (deadline - now)))
1588 timeout = deadline - now;
1593 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1595 struct list_head *tmp;
1596 struct ptlrpc_request *req;
1597 struct l_wait_info lwi;
1601 if (list_empty(&set->set_requests))
1604 list_for_each(tmp, &set->set_requests) {
1605 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1606 if (req->rq_phase == RQ_PHASE_NEW)
1607 (void)ptlrpc_send_new_req(req);
1611 timeout = ptlrpc_set_next_timeout(set);
1613 /* wait until all complete, interrupted, or an in-flight
1615 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1617 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1619 ptlrpc_interrupted_set, set);
1620 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1622 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1624 /* -EINTR => all requests have been flagged rq_intr so next
1626 * -ETIMEOUTD => someone timed out. When all reqs have
1627 * timed out, signals are enabled allowing completion with
1629 * I don't really care if we go once more round the loop in
1630 * the error cases -eeb. */
1631 } while (rc != 0 || set->set_remaining != 0);
1633 LASSERT(set->set_remaining == 0);
1636 list_for_each(tmp, &set->set_requests) {
1637 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1639 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1640 if (req->rq_status != 0)
1641 rc = req->rq_status;
1644 if (set->set_interpret != NULL) {
1645 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1647 rc = interpreter (set, set->set_arg, rc);
1649 struct ptlrpc_set_cbdata *cbdata, *n;
1652 list_for_each_entry_safe(cbdata, n,
1653 &set->set_cblist, psc_item) {
1654 list_del_init(&cbdata->psc_item);
1655 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1658 OBD_SLAB_FREE(cbdata, ptlrpc_cbdata_slab,
1666 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
1668 struct ptlrpc_request_pool *pool = request->rq_pool;
1670 spin_lock(&pool->prp_lock);
1671 LASSERT(list_empty(&request->rq_list));
1672 LASSERT(!request->rq_receiving_reply);
1673 list_add_tail(&request->rq_list, &pool->prp_req_list);
1674 spin_unlock(&pool->prp_lock);
1677 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1680 if (request == NULL) {
1685 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1686 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1687 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1688 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1689 LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request);
1690 LASSERTF(!request->rq_replay, "req %p\n", request);
1692 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1693 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1694 if (request->rq_import != NULL) {
1696 spin_lock(&request->rq_import->imp_lock);
1697 list_del_init(&request->rq_replay_list);
1699 spin_unlock(&request->rq_import->imp_lock);
1701 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1703 if (atomic_read(&request->rq_refcount) != 0) {
1704 DEBUG_REQ(D_ERROR, request,
1705 "freeing request with nonzero refcount");
1709 if (request->rq_repbuf != NULL) {
1710 OBD_FREE(request->rq_repbuf, request->rq_replen);
1711 request->rq_repbuf = NULL;
1712 request->rq_repmsg = NULL;
1714 if (request->rq_export != NULL) {
1715 class_export_put(request->rq_export);
1716 request->rq_export = NULL;
1718 if (request->rq_import != NULL) {
1719 class_import_put(request->rq_import);
1720 request->rq_import = NULL;
1722 if (request->rq_bulk != NULL)
1723 ptlrpc_free_bulk(request->rq_bulk);
1725 if (request->rq_pool) {
1726 __ptlrpc_free_req_to_pool(request);
1728 if (request->rq_reqmsg != NULL) {
1729 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1730 request->rq_reqmsg = NULL;
1732 OBD_FREE(request, sizeof(*request));
1737 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1738 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1740 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1741 (void)__ptlrpc_req_finished(request, 1);
1744 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1747 if (request == NULL)
1750 if (request == LP_POISON ||
1751 request->rq_reqmsg == LP_POISON) {
1752 CERROR("dereferencing freed request (bug 575)\n");
1757 DEBUG_REQ(D_INFO, request, "refcount now %u",
1758 atomic_read(&request->rq_refcount) - 1);
1760 if (atomic_dec_and_test(&request->rq_refcount)) {
1761 __ptlrpc_free_req(request, locked);
1768 void ptlrpc_req_finished(struct ptlrpc_request *request)
1770 __ptlrpc_req_finished(request, 0);
1773 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1775 return request->rq_xid;
1777 EXPORT_SYMBOL(ptlrpc_req_xid);
1779 /* Disengage the client's reply buffer from the network
1780 * NB does _NOT_ unregister any client-side bulk.
1781 * IDEMPOTENT, but _not_ safe against concurrent callers.
1782 * The request owner (i.e. the thread doing the I/O) must call...
1784 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
1788 struct l_wait_info lwi;
1792 LASSERT(!in_interrupt());
1794 /* Let's setup deadline for reply unlink. */
1795 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
1796 async && request->rq_reply_deadline == 0)
1797 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
1799 /* Nothing left to do. */
1800 if (!ptlrpc_client_recv_or_unlink(request))
1803 LNetMDUnlink(request->rq_reply_md_h);
1805 /* Let's check it once again. */
1806 if (!ptlrpc_client_recv_or_unlink(request))
1809 /* Move to "Unregistering" phase as reply was not unlinked yet. */
1810 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
1812 /* Do not wait for unlink to finish. */
1816 /* We have to l_wait_event() whatever the result, to give liblustre
1817 * a chance to run reply_in_callback(), and to make sure we've
1818 * unlinked before returning a req to the pool */
1819 if (request->rq_set != NULL)
1820 wq = &request->rq_set->set_waitq;
1822 wq = &request->rq_reply_waitq;
1825 /* Network access will complete in finite time but the HUGE
1826 * timeout lets us CWARN for visibility of sluggish NALs */
1827 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
1828 cfs_time_seconds(1), NULL, NULL);
1829 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
1832 ptlrpc_rqphase_move(request, request->rq_next_phase);
1836 LASSERT(rc == -ETIMEDOUT);
1837 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
1838 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
1839 request->rq_must_unlink);
1844 /* caller must hold imp->imp_lock */
1845 void ptlrpc_free_committed(struct obd_import *imp)
1847 struct list_head *tmp, *saved;
1848 struct ptlrpc_request *req;
1849 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1852 LASSERT(imp != NULL);
1854 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1857 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1858 imp->imp_generation == imp->imp_last_generation_checked) {
1859 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1860 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1865 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1866 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1867 imp->imp_generation);
1868 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1869 imp->imp_last_generation_checked = imp->imp_generation;
1871 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1872 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1874 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1875 LASSERT(req != last_req);
1878 if (req->rq_import_generation < imp->imp_generation) {
1879 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1883 if (req->rq_replay) {
1884 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1888 /* not yet committed */
1889 if (req->rq_transno > imp->imp_peer_committed_transno) {
1890 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1894 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1895 imp->imp_peer_committed_transno);
1897 spin_lock(&req->rq_lock);
1899 spin_unlock(&req->rq_lock);
1900 if (req->rq_commit_cb != NULL)
1901 req->rq_commit_cb(req);
1902 list_del_init(&req->rq_replay_list);
1903 __ptlrpc_req_finished(req, 1);
1910 void ptlrpc_cleanup_client(struct obd_import *imp)
1917 void ptlrpc_resend_req(struct ptlrpc_request *req)
1919 DEBUG_REQ(D_HA, req, "going to resend");
1920 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1921 req->rq_status = -EAGAIN;
1923 spin_lock(&req->rq_lock);
1925 req->rq_net_err = 0;
1926 req->rq_timedout = 0;
1928 __u64 old_xid = req->rq_xid;
1930 /* ensure previous bulk fails */
1931 req->rq_xid = ptlrpc_next_xid();
1932 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1933 old_xid, req->rq_xid);
1935 ptlrpc_client_wake_req(req);
1936 spin_unlock(&req->rq_lock);
1939 /* XXX: this function and rq_status are currently unused */
1940 void ptlrpc_restart_req(struct ptlrpc_request *req)
1942 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1943 req->rq_status = -ERESTARTSYS;
1945 spin_lock(&req->rq_lock);
1946 req->rq_restart = 1;
1947 req->rq_timedout = 0;
1948 ptlrpc_client_wake_req(req);
1949 spin_unlock(&req->rq_lock);
1952 static void interrupted_request(void *data)
1954 struct ptlrpc_request *req = data;
1955 DEBUG_REQ(D_HA, req, "request interrupted");
1956 spin_lock(&req->rq_lock);
1958 spin_unlock(&req->rq_lock);
1961 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1964 atomic_inc(&req->rq_refcount);
1968 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1969 struct obd_import *imp)
1971 struct list_head *tmp;
1973 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1975 /* clear this for new requests that were resent as well
1976 as resent replayed requests. */
1977 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1979 /* don't re-add requests that have been replayed */
1980 if (!list_empty(&req->rq_replay_list))
1983 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1985 LASSERT(imp->imp_replayable);
1986 /* Balanced in ptlrpc_free_committed, usually. */
1987 ptlrpc_request_addref(req);
1988 list_for_each_prev(tmp, &imp->imp_replay_list) {
1989 struct ptlrpc_request *iter =
1990 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1992 /* We may have duplicate transnos if we create and then
1993 * open a file, or for closes retained if to match creating
1994 * opens, so use req->rq_xid as a secondary key.
1995 * (See bugs 684, 685, and 428.)
1996 * XXX no longer needed, but all opens need transnos!
1998 if (iter->rq_transno > req->rq_transno)
2001 if (iter->rq_transno == req->rq_transno) {
2002 LASSERT(iter->rq_xid != req->rq_xid);
2003 if (iter->rq_xid > req->rq_xid)
2007 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2011 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
2014 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2018 struct l_wait_info lwi;
2019 struct obd_import *imp = req->rq_import;
2020 cfs_duration_t timeout = CFS_TICK;
2024 LASSERT(req->rq_set == NULL);
2025 LASSERT(!req->rq_receiving_reply);
2027 /* for distributed debugging */
2028 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2029 LASSERT(imp->imp_obd != NULL);
2030 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
2031 "%s:%s:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
2032 imp->imp_obd->obd_uuid.uuid,
2033 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2034 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2035 lustre_msg_get_opc(req->rq_reqmsg));
2037 /* Mark phase here for a little debug help */
2038 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
2040 spin_lock(&imp->imp_lock);
2041 req->rq_import_generation = imp->imp_generation;
2043 if (ptlrpc_import_delay_req(imp, req, &rc)) {
2044 list_del_init(&req->rq_list);
2045 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
2046 atomic_inc(&imp->imp_inflight);
2047 spin_unlock(&imp->imp_lock);
2049 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
2051 ptlrpc_import_state_name(req->rq_send_state),
2052 ptlrpc_import_state_name(imp->imp_state));
2053 lwi = LWI_INTR(interrupted_request, req);
2054 rc = l_wait_event(req->rq_reply_waitq,
2055 (req->rq_send_state == imp->imp_state ||
2056 req->rq_err || req->rq_intr),
2058 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
2060 ptlrpc_import_state_name(imp->imp_state),
2061 ptlrpc_import_state_name(req->rq_send_state),
2062 req->rq_err, req->rq_intr);
2064 spin_lock(&imp->imp_lock);
2065 list_del_init(&req->rq_list);
2066 atomic_dec(&imp->imp_inflight);
2069 /* rq_status was set locally */
2070 rc = req->rq_status ? req->rq_status : -EIO;
2072 else if (req->rq_intr) {
2075 else if (req->rq_no_resend) {
2084 spin_unlock(&imp->imp_lock);
2085 req->rq_status = rc; // XXX this ok?
2089 if (req->rq_resend) {
2090 if (req->rq_bulk != NULL) {
2091 ptlrpc_unregister_bulk(req, 0);
2093 /* bulk requests are supposed to be
2094 * idempotent, so we are free to bump the xid
2095 * here, which we need to do before
2096 * registering the bulk again (bug 6371).
2097 * print the old xid first for sanity.
2099 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
2100 req->rq_xid = ptlrpc_next_xid();
2103 DEBUG_REQ(D_HA, req, "resending: ");
2106 /* XXX this is the same as ptlrpc_set_wait */
2107 LASSERT(list_empty(&req->rq_list));
2108 list_add_tail(&req->rq_list, &imp->imp_sending_list);
2109 atomic_inc(&imp->imp_inflight);
2110 spin_unlock(&imp->imp_lock);
2112 rc = ptl_send_rpc(req, 0);
2114 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
2116 timeoutl = req->rq_deadline - cfs_time_current_sec();
2117 timeout = (timeoutl <= 0 || rc) ? CFS_TICK :
2118 cfs_time_seconds(timeoutl);
2119 DEBUG_REQ(D_NET, req,
2120 "-- sleeping for "CFS_DURATION_T" ticks", timeout);
2121 lwi = LWI_TIMEOUT_INTR(timeout, NULL, interrupted_request, req);
2122 brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req),
2124 /* Wait again if we changed deadline */
2125 } while ((brc == -ETIMEDOUT) &&
2126 (req->rq_deadline > cfs_time_current_sec()));
2128 if ((brc == -ETIMEDOUT) && !ptlrpc_expire_one_request(req, 0)) {
2129 /* Wait forever for reconnect / replay or failure */
2130 lwi = LWI_INTR(interrupted_request, req);
2131 brc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req),
2135 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
2136 "%s:%s:%d:x"LPU64":%s:%d\n", cfs_curproc_comm(),
2137 imp->imp_obd->obd_uuid.uuid,
2138 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2139 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2140 lustre_msg_get_opc(req->rq_reqmsg));
2142 /* If the reply was received normally, this just grabs the spinlock
2143 * (ensuring the reply callback has returned), sees that
2144 * req->rq_receiving_reply is clear and returns. */
2145 ptlrpc_unregister_reply(req, 0);
2147 spin_lock(&imp->imp_lock);
2148 list_del_init(&req->rq_list);
2149 atomic_dec(&imp->imp_inflight);
2150 spin_unlock(&imp->imp_lock);
2153 DEBUG_REQ(D_RPCTRACE, req, "err rc=%d status=%d",
2154 rc, req->rq_status);
2155 rc = rc ? rc : req->rq_status;
2156 GOTO(out, rc = rc ? rc : -EIO);
2160 /* Should only be interrupted if we timed out. */
2161 if (!req->rq_timedout)
2162 DEBUG_REQ(D_ERROR, req,
2163 "rq_intr set but rq_timedout not");
2164 GOTO(out, rc = -EINTR);
2167 /* Resend if we need to */
2168 if (req->rq_resend||req->rq_timedout) {
2169 /* ...unless we were specifically told otherwise. */
2170 if (req->rq_no_resend)
2171 GOTO(out, rc = -ETIMEDOUT);
2172 spin_lock(&imp->imp_lock);
2173 /* we can have rq_timeout on dlm fake import which not support
2174 * recovery - but me need resend request on this import instead
2175 * of return error */
2180 if (!ptlrpc_client_replied(req)) {
2181 /* How can this be? -eeb */
2182 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
2184 GOTO(out, rc = req->rq_status);
2187 rc = after_reply(req);
2188 /* NB may return +ve success rc */
2189 if (req->rq_resend) {
2190 spin_lock(&imp->imp_lock);
2195 if (req->rq_bulk != NULL) {
2197 /* success so far. Note that anything going wrong
2198 * with bulk now, is EXTREMELY strange, since the
2199 * server must have believed that the bulk
2200 * tranferred OK before she replied with success to
2202 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
2203 brc = l_wait_event(req->rq_reply_waitq,
2204 !ptlrpc_client_bulk_active(req),
2206 LASSERT(brc == 0 || brc == -ETIMEDOUT);
2208 LASSERT(brc == -ETIMEDOUT);
2209 DEBUG_REQ(D_ERROR, req, "bulk timed out");
2211 } else if (!req->rq_bulk->bd_success) {
2212 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2217 ptlrpc_unregister_bulk(req, 0);
2220 LASSERT(!req->rq_receiving_reply);
2221 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2222 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
2226 struct ptlrpc_replay_async_args {
2228 int praa_old_status;
2231 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
2232 void * data, int rc)
2234 struct ptlrpc_replay_async_args *aa = data;
2235 struct obd_import *imp = req->rq_import;
2238 atomic_dec(&imp->imp_replay_inflight);
2240 if (!ptlrpc_client_replied(req)) {
2241 CERROR("request replay timed out, restarting recovery\n");
2242 GOTO(out, rc = -ETIMEDOUT);
2245 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2246 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2247 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2248 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2250 /* VBR: check version failure */
2251 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2252 /* replay was failed due to version mismatch */
2253 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2254 spin_lock(&imp->imp_lock);
2255 imp->imp_vbr_failed = 1;
2256 imp->imp_no_lock_replay = 1;
2257 spin_unlock(&imp->imp_lock);
2259 /* The transno had better not change over replay. */
2260 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
2261 lustre_msg_get_transno(req->rq_repmsg) ||
2262 lustre_msg_get_transno(req->rq_repmsg) == 0);
2265 spin_lock(&imp->imp_lock);
2266 /* if replays by version then gap was occur on server, no trust to locks */
2267 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2268 imp->imp_no_lock_replay = 1;
2269 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2270 spin_unlock(&imp->imp_lock);
2271 LASSERT(imp->imp_last_replay_transno);
2273 DEBUG_REQ(D_HA, req, "got rep");
2275 /* let the callback do fixups, possibly including in the request */
2276 if (req->rq_replay_cb)
2277 req->rq_replay_cb(req);
2279 if (ptlrpc_client_replied(req) &&
2280 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2281 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2282 lustre_msg_get_status(req->rq_repmsg),
2283 aa->praa_old_status);
2285 /* Put it back for re-replay. */
2286 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2289 /* continue with recovery */
2290 rc = ptlrpc_import_recovery_state_machine(imp);
2292 req->rq_send_state = aa->praa_old_state;
2295 /* this replay failed, so restart recovery */
2296 ptlrpc_connect_import(imp, NULL);
2302 int ptlrpc_replay_req(struct ptlrpc_request *req)
2304 struct ptlrpc_replay_async_args *aa;
2307 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2309 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2310 aa = ptlrpc_req_async_args(req);
2311 memset(aa, 0, sizeof *aa);
2313 /* Prepare request to be resent with ptlrpcd */
2314 aa->praa_old_state = req->rq_send_state;
2315 req->rq_send_state = LUSTRE_IMP_REPLAY;
2316 req->rq_phase = RQ_PHASE_NEW;
2317 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2319 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2321 req->rq_interpret_reply = ptlrpc_replay_interpret;
2322 /* Readjust the timeout for current conditions */
2323 ptlrpc_at_set_req_timeout(req);
2325 DEBUG_REQ(D_HA, req, "REPLAY");
2327 atomic_inc(&req->rq_import->imp_replay_inflight);
2328 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2330 ptlrpcd_add_req(req);
2334 void ptlrpc_abort_inflight(struct obd_import *imp)
2336 struct list_head *tmp, *n;
2339 /* Make sure that no new requests get processed for this import.
2340 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2341 * this flag and then putting requests on sending_list or delayed_list.
2343 spin_lock(&imp->imp_lock);
2345 /* XXX locking? Maybe we should remove each request with the list
2346 * locked? Also, how do we know if the requests on the list are
2347 * being freed at this time?
2349 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2350 struct ptlrpc_request *req =
2351 list_entry(tmp, struct ptlrpc_request, rq_list);
2353 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2355 spin_lock (&req->rq_lock);
2356 if (req->rq_import_generation < imp->imp_generation) {
2358 req->rq_status = -EINTR;
2359 ptlrpc_client_wake_req(req);
2361 spin_unlock (&req->rq_lock);
2364 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2365 struct ptlrpc_request *req =
2366 list_entry(tmp, struct ptlrpc_request, rq_list);
2368 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2370 spin_lock (&req->rq_lock);
2371 if (req->rq_import_generation < imp->imp_generation) {
2373 req->rq_status = -EINTR;
2374 ptlrpc_client_wake_req(req);
2376 spin_unlock (&req->rq_lock);
2379 /* Last chance to free reqs left on the replay list, but we
2380 * will still leak reqs that haven't committed. */
2381 if (imp->imp_replayable)
2382 ptlrpc_free_committed(imp);
2384 spin_unlock(&imp->imp_lock);
2389 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2391 struct list_head *tmp, *pos;
2393 LASSERT(set != NULL);
2395 list_for_each_safe(pos, tmp, &set->set_requests) {
2396 struct ptlrpc_request *req =
2397 list_entry(pos, struct ptlrpc_request, rq_set_chain);
2399 spin_lock(&req->rq_lock);
2400 if (req->rq_phase != RQ_PHASE_RPC) {
2401 spin_unlock(&req->rq_lock);
2406 req->rq_status = -EINTR;
2407 ptlrpc_client_wake_req(req);
2408 spin_unlock(&req->rq_lock);
2412 static __u64 ptlrpc_last_xid;
2413 static spinlock_t ptlrpc_last_xid_lock;
2415 /* Initialize the XID for the node. This is common among all requests on
2416 * this node, and only requires the property that it is monotonically
2417 * increasing. It does not need to be sequential. Since this is also used
2418 * as the RDMA match bits, it is important that a single client NOT have
2419 * the same match bits for two different in-flight requests, hence we do
2420 * NOT want to have an XID per target or similar.
2422 * To avoid an unlikely collision between match bits after a client reboot
2423 * (which would deliver old data into the wrong RDMA buffer) we initialize
2424 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2425 * If the time is clearly incorrect, we instead use a 62-bit random number.
2426 * In the worst case the random number will overflow 1M RPCs per second in
2427 * 9133 years, or permutations thereof.
2429 #define YEAR_2004 (1ULL << 30)
2430 void ptlrpc_init_xid(void)
2432 time_t now = cfs_time_current_sec();
2434 spin_lock_init(&ptlrpc_last_xid_lock);
2435 if (now < YEAR_2004) {
2436 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2437 ptlrpc_last_xid >>= 2;
2438 ptlrpc_last_xid |= (1ULL << 61);
2440 ptlrpc_last_xid = (__u64)now << 20;
2444 __u64 ptlrpc_next_xid(void)
2447 spin_lock(&ptlrpc_last_xid_lock);
2448 tmp = ++ptlrpc_last_xid;
2449 spin_unlock(&ptlrpc_last_xid_lock);
2453 __u64 ptlrpc_sample_next_xid(void)
2455 if (sizeof(long) < 8) {
2456 /* need to avoid possible word tearing on 32-bit systems */
2458 spin_lock(&ptlrpc_last_xid_lock);
2459 tmp = ptlrpc_last_xid + 1;
2460 spin_unlock(&ptlrpc_last_xid_lock);
2463 /* No need to lock, since returned value is racy anyways */
2464 return ptlrpc_last_xid + 1;
2466 EXPORT_SYMBOL(ptlrpc_sample_next_xid);