4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 /** Implementation of client-side PortalRPC interfaces */
35 #define DEBUG_SUBSYSTEM S_RPC
37 #include <linux/delay.h>
38 #include <linux/random.h>
40 #include <lnet/lib-lnet.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
48 #include "ptlrpc_internal.h"
50 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
51 struct page *page, int pageoffset,
54 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
57 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
58 struct page *page, int pageoffset,
61 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
64 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
68 for (i = 0; i < desc->bd_iov_count ; i++)
69 put_page(BD_GET_KIOV(desc, i).kiov_page);
72 static int ptlrpc_prep_bulk_frag_pages(struct ptlrpc_bulk_desc *desc,
75 unsigned int offset = (uintptr_t)frag & ~PAGE_MASK;
79 int page_len = min_t(unsigned int, PAGE_SIZE - offset,
81 uintptr_t vaddr = (uintptr_t) frag;
83 ptlrpc_prep_bulk_page_nopin(desc,
84 lnet_kvaddr_to_page(vaddr),
94 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
95 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
96 .release_frags = ptlrpc_release_bulk_page_pin,
98 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
100 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
101 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
102 .release_frags = ptlrpc_release_bulk_noop,
103 .add_iov_frag = ptlrpc_prep_bulk_frag_pages,
105 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
107 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
108 .add_iov_frag = ptlrpc_prep_bulk_frag,
110 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
112 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
113 static int ptlrpcd_check_work(struct ptlrpc_request *req);
114 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
117 * Initialize passed in client structure \a cl.
119 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
120 struct ptlrpc_client *cl)
122 cl->cli_request_portal = req_portal;
123 cl->cli_reply_portal = rep_portal;
126 EXPORT_SYMBOL(ptlrpc_init_client);
129 * Return PortalRPC connection for remore uud \a uuid
131 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
132 lnet_nid_t nid4refnet)
134 struct ptlrpc_connection *c;
136 struct lnet_process_id peer;
140 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
141 * before accessing its values.
143 /* coverity[uninit_use_in_call] */
144 peer.nid = nid4refnet;
145 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
147 CNETERR("cannot find peer %s!\n", uuid->uuid);
151 c = ptlrpc_connection_get(peer, self, uuid);
153 memcpy(c->c_remote_uuid.uuid,
154 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
157 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
163 * Allocate and initialize new bulk descriptor on the sender.
164 * Returns pointer to the descriptor or NULL on error.
166 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
167 unsigned int max_brw,
168 enum ptlrpc_bulk_op_type type,
170 const struct ptlrpc_bulk_frag_ops *ops)
172 struct ptlrpc_bulk_desc *desc;
175 /* ensure that only one of KIOV or IOVEC is set but not both */
176 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
177 ops->add_kiov_frag != NULL) ||
178 (ptlrpc_is_bulk_desc_kvec(type) &&
179 ops->add_iov_frag != NULL));
184 if (type & PTLRPC_BULK_BUF_KIOV) {
185 OBD_ALLOC_LARGE(GET_KIOV(desc),
186 nfrags * sizeof(*GET_KIOV(desc)));
190 OBD_ALLOC_LARGE(GET_KVEC(desc),
191 nfrags * sizeof(*GET_KVEC(desc)));
196 spin_lock_init(&desc->bd_lock);
197 init_waitqueue_head(&desc->bd_waitq);
198 desc->bd_max_iov = nfrags;
199 desc->bd_iov_count = 0;
200 desc->bd_portal = portal;
201 desc->bd_type = type;
202 desc->bd_md_count = 0;
203 desc->bd_frag_ops = ops;
204 LASSERT(max_brw > 0);
205 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
207 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
208 * node. Negotiated ocd_brw_size will always be <= this number.
210 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
211 LNetInvalidateMDHandle(&desc->bd_mds[i]);
220 * Prepare bulk descriptor for specified outgoing request \a req that
221 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
222 * the bulk to be sent. Used on client-side.
223 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
226 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
228 unsigned int max_brw,
231 const struct ptlrpc_bulk_frag_ops
234 struct obd_import *imp = req->rq_import;
235 struct ptlrpc_bulk_desc *desc;
238 LASSERT(ptlrpc_is_bulk_op_passive(type));
240 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
244 desc->bd_import = class_import_get(imp);
247 desc->bd_cbid.cbid_fn = client_bulk_callback;
248 desc->bd_cbid.cbid_arg = desc;
250 /* This makes req own desc, and free it when she frees herself */
255 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
257 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
258 struct page *page, int pageoffset, int len,
263 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
264 LASSERT(page != NULL);
265 LASSERT(pageoffset >= 0);
267 LASSERT(pageoffset + len <= PAGE_SIZE);
268 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
270 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
277 kiov->kiov_page = page;
278 kiov->kiov_offset = pageoffset;
279 kiov->kiov_len = len;
281 desc->bd_iov_count++;
283 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
285 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
292 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
293 LASSERT(frag != NULL);
295 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
297 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
301 iovec->iov_base = frag;
302 iovec->iov_len = len;
304 desc->bd_iov_count++;
306 RETURN(desc->bd_nob);
308 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
310 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
314 LASSERT(desc != NULL);
315 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
316 LASSERT(desc->bd_md_count == 0); /* network hands off */
317 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
318 LASSERT(desc->bd_frag_ops != NULL);
320 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
321 sptlrpc_enc_pool_put_pages(desc);
324 class_export_put(desc->bd_export);
326 class_import_put(desc->bd_import);
328 if (desc->bd_frag_ops->release_frags != NULL)
329 desc->bd_frag_ops->release_frags(desc);
331 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
332 OBD_FREE_LARGE(GET_KIOV(desc),
333 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
335 OBD_FREE_LARGE(GET_KVEC(desc),
336 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
340 EXPORT_SYMBOL(ptlrpc_free_bulk);
343 * Set server timelimit for this req, i.e. how long are we willing to wait
344 * for reply before timing out this request.
346 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
352 LASSERT(req->rq_import);
355 /* non-AT settings */
357 * \a imp_server_timeout means this is reverse import and
358 * we send (currently only) ASTs to the client and cannot afford
359 * to wait too long for the reply, otherwise the other client
360 * (because of which we are sending this request) would
361 * timeout waiting for us
363 req->rq_timeout = req->rq_import->imp_server_timeout ?
364 obd_timeout / 2 : obd_timeout;
366 at = &req->rq_import->imp_at;
367 idx = import_at_get_index(req->rq_import,
368 req->rq_request_portal);
369 serv_est = at_get(&at->iat_service_estimate[idx]);
370 req->rq_timeout = at_est2timeout(serv_est);
373 * We could get even fancier here, using history to predict increased
378 * Let the server know what this RPC timeout is by putting it in the
381 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
383 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
385 /* Adjust max service estimate based on server value */
386 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
387 unsigned int serv_est)
393 LASSERT(req->rq_import);
394 at = &req->rq_import->imp_at;
396 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
398 * max service estimates are tracked on the server side,
399 * so just keep minimal history here
401 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
404 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
405 req->rq_import->imp_obd->obd_name,
406 req->rq_request_portal,
407 oldse, at_get(&at->iat_service_estimate[idx]));
410 /* Expected network latency per remote node (secs) */
411 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
413 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
416 /* Adjust expected network latency */
417 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
418 unsigned int service_time)
420 unsigned int nl, oldnl;
422 time64_t now = ktime_get_real_seconds();
424 LASSERT(req->rq_import);
426 if (service_time > now - req->rq_sent + 3) {
428 * b=16408, however, this can also happen if early reply
429 * is lost and client RPC is expired and resent, early reply
430 * or reply of original RPC can still be fit in reply buffer
431 * of resent RPC, now client is measuring time from the
432 * resent time, but server sent back service time of original
435 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
436 D_ADAPTTO : D_WARNING,
437 "Reported service time %u > total measured time %lld\n",
438 service_time, now - req->rq_sent);
442 /* Network latency is total time less server processing time */
443 nl = max_t(int, now - req->rq_sent -
444 service_time, 0) + 1; /* st rounding */
445 at = &req->rq_import->imp_at;
447 oldnl = at_measured(&at->iat_net_latency, nl);
450 "The network latency for %s (nid %s) has changed from %d to %d\n",
451 req->rq_import->imp_obd->obd_name,
452 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
453 oldnl, at_get(&at->iat_net_latency));
456 static int unpack_reply(struct ptlrpc_request *req)
460 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
461 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
463 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
469 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
471 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
479 * Handle an early reply message, called with the rq_lock held.
480 * If anything goes wrong just ignore it - same as if it never happened
482 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
483 __must_hold(&req->rq_lock)
485 struct ptlrpc_request *early_req;
491 spin_unlock(&req->rq_lock);
493 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
495 spin_lock(&req->rq_lock);
499 rc = unpack_reply(early_req);
501 sptlrpc_cli_finish_early_reply(early_req);
502 spin_lock(&req->rq_lock);
507 * Use new timeout value just to adjust the local value for this
508 * request, don't include it into at_history. It is unclear yet why
509 * service time increased and should it be counted or skipped, e.g.
510 * that can be recovery case or some error or server, the real reply
511 * will add all new data if it is worth to add.
513 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
514 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
516 /* Network latency can be adjusted, it is pure network delays */
517 ptlrpc_at_adj_net_latency(req,
518 lustre_msg_get_service_time(early_req->rq_repmsg));
520 sptlrpc_cli_finish_early_reply(early_req);
522 spin_lock(&req->rq_lock);
523 olddl = req->rq_deadline;
525 * server assumes it now has rq_timeout from when the request
526 * arrived, so the client should give it at least that long.
527 * since we don't know the arrival time we'll use the original
530 req->rq_deadline = req->rq_sent + req->rq_timeout +
531 ptlrpc_at_get_net_latency(req);
533 /* The below message is checked in replay-single.sh test_65{a,b} */
534 /* The below message is checked in sanity-{gss,krb5} test_8 */
535 DEBUG_REQ(D_ADAPTTO, req,
536 "Early reply #%d, new deadline in %llds (%llds)",
538 req->rq_deadline - ktime_get_real_seconds(),
539 req->rq_deadline - olddl);
544 static struct kmem_cache *request_cache;
546 int ptlrpc_request_cache_init(void)
548 request_cache = kmem_cache_create("ptlrpc_cache",
549 sizeof(struct ptlrpc_request),
550 0, SLAB_HWCACHE_ALIGN, NULL);
551 return request_cache ? 0 : -ENOMEM;
554 void ptlrpc_request_cache_fini(void)
556 kmem_cache_destroy(request_cache);
559 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
561 struct ptlrpc_request *req;
563 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
567 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
569 OBD_SLAB_FREE_PTR(req, request_cache);
573 * Wind down request pool \a pool.
574 * Frees all requests from the pool too
576 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
578 struct list_head *l, *tmp;
579 struct ptlrpc_request *req;
581 LASSERT(pool != NULL);
583 spin_lock(&pool->prp_lock);
584 list_for_each_safe(l, tmp, &pool->prp_req_list) {
585 req = list_entry(l, struct ptlrpc_request, rq_list);
586 list_del(&req->rq_list);
587 LASSERT(req->rq_reqbuf);
588 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
589 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
590 ptlrpc_request_cache_free(req);
592 spin_unlock(&pool->prp_lock);
593 OBD_FREE(pool, sizeof(*pool));
595 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
598 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
600 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
605 while (size < pool->prp_rq_size)
608 LASSERTF(list_empty(&pool->prp_req_list) ||
609 size == pool->prp_rq_size,
610 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
611 pool->prp_rq_size, size);
613 pool->prp_rq_size = size;
614 for (i = 0; i < num_rq; i++) {
615 struct ptlrpc_request *req;
616 struct lustre_msg *msg;
618 req = ptlrpc_request_cache_alloc(GFP_NOFS);
621 OBD_ALLOC_LARGE(msg, size);
623 ptlrpc_request_cache_free(req);
626 req->rq_reqbuf = msg;
627 req->rq_reqbuf_len = size;
629 spin_lock(&pool->prp_lock);
630 list_add_tail(&req->rq_list, &pool->prp_req_list);
631 spin_unlock(&pool->prp_lock);
635 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
638 * Create and initialize new request pool with given attributes:
639 * \a num_rq - initial number of requests to create for the pool
640 * \a msgsize - maximum message size possible for requests in thid pool
641 * \a populate_pool - function to be called when more requests need to be added
643 * Returns pointer to newly created pool or NULL on error.
645 struct ptlrpc_request_pool *
646 ptlrpc_init_rq_pool(int num_rq, int msgsize,
647 int (*populate_pool)(struct ptlrpc_request_pool *, int))
649 struct ptlrpc_request_pool *pool;
656 * Request next power of two for the allocation, because internally
657 * kernel would do exactly this
659 spin_lock_init(&pool->prp_lock);
660 INIT_LIST_HEAD(&pool->prp_req_list);
661 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
662 pool->prp_populate = populate_pool;
664 populate_pool(pool, num_rq);
668 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
671 * Fetches one request from pool \a pool
673 static struct ptlrpc_request *
674 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
676 struct ptlrpc_request *request;
677 struct lustre_msg *reqbuf;
682 spin_lock(&pool->prp_lock);
685 * See if we have anything in a pool, and bail out if nothing,
686 * in writeout path, where this matters, this is safe to do, because
687 * nothing is lost in this case, and when some in-flight requests
688 * complete, this code will be called again.
690 if (unlikely(list_empty(&pool->prp_req_list))) {
691 spin_unlock(&pool->prp_lock);
695 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
697 list_del_init(&request->rq_list);
698 spin_unlock(&pool->prp_lock);
700 LASSERT(request->rq_reqbuf);
701 LASSERT(request->rq_pool);
703 reqbuf = request->rq_reqbuf;
704 memset(request, 0, sizeof(*request));
705 request->rq_reqbuf = reqbuf;
706 request->rq_reqbuf_len = pool->prp_rq_size;
707 request->rq_pool = pool;
713 * Returns freed \a request to pool.
715 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
717 struct ptlrpc_request_pool *pool = request->rq_pool;
719 spin_lock(&pool->prp_lock);
720 LASSERT(list_empty(&request->rq_list));
721 LASSERT(!request->rq_receiving_reply);
722 list_add_tail(&request->rq_list, &pool->prp_req_list);
723 spin_unlock(&pool->prp_lock);
726 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
728 struct obd_import *imp = req->rq_import;
729 struct list_head *tmp;
730 struct ptlrpc_request *iter;
732 assert_spin_locked(&imp->imp_lock);
733 LASSERT(list_empty(&req->rq_unreplied_list));
735 /* unreplied list is sorted by xid in ascending order */
736 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
737 iter = list_entry(tmp, struct ptlrpc_request,
740 LASSERT(req->rq_xid != iter->rq_xid);
741 if (req->rq_xid < iter->rq_xid)
743 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
746 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
749 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
751 req->rq_xid = ptlrpc_next_xid();
752 ptlrpc_add_unreplied(req);
755 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
757 spin_lock(&req->rq_import->imp_lock);
758 ptlrpc_assign_next_xid_nolock(req);
759 spin_unlock(&req->rq_import->imp_lock);
762 static atomic64_t ptlrpc_last_xid;
764 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
766 spin_lock(&req->rq_import->imp_lock);
767 list_del_init(&req->rq_unreplied_list);
768 ptlrpc_assign_next_xid_nolock(req);
769 spin_unlock(&req->rq_import->imp_lock);
770 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
773 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
775 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
779 opc = lustre_msg_get_opc(req->rq_reqmsg);
780 tag = obd_get_mod_rpc_slot(cli, opc);
781 lustre_msg_set_tag(req->rq_reqmsg, tag);
782 ptlrpc_reassign_next_xid(req);
784 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
786 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
788 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
791 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
792 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
794 obd_put_mod_rpc_slot(cli, opc, tag);
797 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
799 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
800 __u32 version, int opcode, char **bufs,
801 struct ptlrpc_cli_ctx *ctx)
804 struct obd_import *imp;
810 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
811 imp = request->rq_import;
812 lengths = request->rq_pill.rc_area[RCL_CLIENT];
815 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
817 rc = sptlrpc_req_get_ctx(request);
821 sptlrpc_req_set_flavor(request, opcode);
823 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
828 lustre_msg_add_version(request->rq_reqmsg, version);
829 request->rq_send_state = LUSTRE_IMP_FULL;
830 request->rq_type = PTL_RPC_MSG_REQUEST;
832 request->rq_req_cbid.cbid_fn = request_out_callback;
833 request->rq_req_cbid.cbid_arg = request;
835 request->rq_reply_cbid.cbid_fn = reply_in_callback;
836 request->rq_reply_cbid.cbid_arg = request;
838 request->rq_reply_deadline = 0;
839 request->rq_bulk_deadline = 0;
840 request->rq_req_deadline = 0;
841 request->rq_phase = RQ_PHASE_NEW;
842 request->rq_next_phase = RQ_PHASE_UNDEFINED;
844 request->rq_request_portal = imp->imp_client->cli_request_portal;
845 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
847 ptlrpc_at_set_req_timeout(request);
849 lustre_msg_set_opc(request->rq_reqmsg, opcode);
851 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
852 if (cfs_fail_val == opcode) {
853 time64_t *fail_t = NULL, *fail2_t = NULL;
855 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
856 fail_t = &request->rq_bulk_deadline;
857 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
858 fail_t = &request->rq_reply_deadline;
859 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
860 fail_t = &request->rq_req_deadline;
861 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
862 fail_t = &request->rq_reply_deadline;
863 fail2_t = &request->rq_bulk_deadline;
864 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
865 time64_t now = ktime_get_real_seconds();
866 u64 xid = ((u64)now >> 4) << 24;
868 atomic64_set(&ptlrpc_last_xid, xid);
872 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
875 *fail2_t = ktime_get_real_seconds() +
879 * The RPC is infected, let the test to change the
882 msleep(4 * MSEC_PER_SEC);
885 ptlrpc_assign_next_xid(request);
890 LASSERT(!request->rq_pool);
891 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
893 class_import_put(imp);
897 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
900 * Pack request buffers for network transfer, performing necessary encryption
901 * steps if necessary.
903 int ptlrpc_request_pack(struct ptlrpc_request *request,
904 __u32 version, int opcode)
906 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
908 EXPORT_SYMBOL(ptlrpc_request_pack);
911 * Helper function to allocate new request on import \a imp
912 * and possibly using existing request from pool \a pool if provided.
913 * Returns allocated request structure with import field filled or
917 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
918 struct ptlrpc_request_pool *pool)
920 struct ptlrpc_request *request = NULL;
922 request = ptlrpc_request_cache_alloc(GFP_NOFS);
924 if (!request && pool)
925 request = ptlrpc_prep_req_from_pool(pool);
928 ptlrpc_cli_req_init(request);
930 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
931 LASSERT(imp != LP_POISON);
932 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
934 LASSERT(imp->imp_client != LP_POISON);
936 request->rq_import = class_import_get(imp);
938 CERROR("request allocation out of memory\n");
945 * Helper function for creating a request.
946 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
947 * buffer structures according to capsule template \a format.
948 * Returns allocated request structure pointer or NULL on error.
950 static struct ptlrpc_request *
951 ptlrpc_request_alloc_internal(struct obd_import *imp,
952 struct ptlrpc_request_pool *pool,
953 const struct req_format *format)
955 struct ptlrpc_request *request;
957 request = __ptlrpc_request_alloc(imp, pool);
962 * initiate connection if needed when the import has been
963 * referenced by the new request to avoid races with disconnect
965 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
968 CDEBUG_LIMIT(imp->imp_idle_debug,
969 "%s: reconnect after %llds idle\n",
970 imp->imp_obd->obd_name, ktime_get_real_seconds() -
971 imp->imp_last_reply_time);
972 spin_lock(&imp->imp_lock);
973 if (imp->imp_state == LUSTRE_IMP_IDLE) {
974 imp->imp_generation++;
975 imp->imp_initiated_at = imp->imp_generation;
976 imp->imp_state = LUSTRE_IMP_NEW;
978 /* connect_import_locked releases imp_lock */
979 rc = ptlrpc_connect_import_locked(imp);
981 ptlrpc_request_free(request);
984 ptlrpc_pinger_add_import(imp);
986 spin_unlock(&imp->imp_lock);
990 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
991 req_capsule_set(&request->rq_pill, format);
996 * Allocate new request structure for import \a imp and initialize its
997 * buffer structure according to capsule template \a format.
999 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
1000 const struct req_format *format)
1002 return ptlrpc_request_alloc_internal(imp, NULL, format);
1004 EXPORT_SYMBOL(ptlrpc_request_alloc);
1007 * Allocate new request structure for import \a imp from pool \a pool and
1008 * initialize its buffer structure according to capsule template \a format.
1010 struct ptlrpc_request *
1011 ptlrpc_request_alloc_pool(struct obd_import *imp,
1012 struct ptlrpc_request_pool *pool,
1013 const struct req_format *format)
1015 return ptlrpc_request_alloc_internal(imp, pool, format);
1017 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
1020 * For requests not from pool, free memory of the request structure.
1021 * For requests obtained from a pool earlier, return request back to pool.
1023 void ptlrpc_request_free(struct ptlrpc_request *request)
1025 if (request->rq_pool)
1026 __ptlrpc_free_req_to_pool(request);
1028 ptlrpc_request_cache_free(request);
1030 EXPORT_SYMBOL(ptlrpc_request_free);
1033 * Allocate new request for operatione \a opcode and immediatelly pack it for
1035 * Only used for simple requests like OBD_PING where the only important
1036 * part of the request is operation itself.
1037 * Returns allocated request or NULL on error.
1039 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1040 const struct req_format *format,
1041 __u32 version, int opcode)
1043 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1047 rc = ptlrpc_request_pack(req, version, opcode);
1049 ptlrpc_request_free(req);
1055 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1058 * Allocate and initialize new request set structure on the current CPT.
1059 * Returns a pointer to the newly allocated set structure or NULL on error.
1061 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1063 struct ptlrpc_request_set *set;
1067 cpt = cfs_cpt_current(cfs_cpt_table, 0);
1068 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1071 atomic_set(&set->set_refcount, 1);
1072 INIT_LIST_HEAD(&set->set_requests);
1073 init_waitqueue_head(&set->set_waitq);
1074 atomic_set(&set->set_new_count, 0);
1075 atomic_set(&set->set_remaining, 0);
1076 spin_lock_init(&set->set_new_req_lock);
1077 INIT_LIST_HEAD(&set->set_new_requests);
1078 set->set_max_inflight = UINT_MAX;
1079 set->set_producer = NULL;
1080 set->set_producer_arg = NULL;
1085 EXPORT_SYMBOL(ptlrpc_prep_set);
1088 * Allocate and initialize new request set structure with flow control
1089 * extension. This extension allows to control the number of requests in-flight
1090 * for the whole set. A callback function to generate requests must be provided
1091 * and the request set will keep the number of requests sent over the wire to
1093 * Returns a pointer to the newly allocated set structure or NULL on error.
1095 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1099 struct ptlrpc_request_set *set;
1101 set = ptlrpc_prep_set();
1105 set->set_max_inflight = max;
1106 set->set_producer = func;
1107 set->set_producer_arg = arg;
1113 * Wind down and free request set structure previously allocated with
1115 * Ensures that all requests on the set have completed and removes
1116 * all requests from the request list in a set.
1117 * If any unsent request happen to be on the list, pretends that they got
1118 * an error in flight and calls their completion handler.
1120 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1122 struct list_head *tmp;
1123 struct list_head *next;
1129 /* Requests on the set should either all be completed, or all be new */
1130 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1131 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1132 list_for_each(tmp, &set->set_requests) {
1133 struct ptlrpc_request *req =
1134 list_entry(tmp, struct ptlrpc_request,
1137 LASSERT(req->rq_phase == expected_phase);
1141 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1142 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1143 atomic_read(&set->set_remaining), n);
1145 list_for_each_safe(tmp, next, &set->set_requests) {
1146 struct ptlrpc_request *req =
1147 list_entry(tmp, struct ptlrpc_request,
1149 list_del_init(&req->rq_set_chain);
1151 LASSERT(req->rq_phase == expected_phase);
1153 if (req->rq_phase == RQ_PHASE_NEW) {
1154 ptlrpc_req_interpret(NULL, req, -EBADR);
1155 atomic_dec(&set->set_remaining);
1158 spin_lock(&req->rq_lock);
1160 req->rq_invalid_rqset = 0;
1161 spin_unlock(&req->rq_lock);
1163 ptlrpc_req_finished(req);
1166 LASSERT(atomic_read(&set->set_remaining) == 0);
1168 ptlrpc_reqset_put(set);
1171 EXPORT_SYMBOL(ptlrpc_set_destroy);
1174 * Add a new request to the general purpose request set.
1175 * Assumes request reference from the caller.
1177 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1178 struct ptlrpc_request *req)
1180 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1181 LASSERT(list_empty(&req->rq_set_chain));
1183 if (req->rq_allow_intr)
1184 set->set_allow_intr = 1;
1186 /* The set takes over the caller's request reference */
1187 list_add_tail(&req->rq_set_chain, &set->set_requests);
1189 atomic_inc(&set->set_remaining);
1190 req->rq_queued_time = ktime_get_seconds();
1193 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1195 if (set->set_producer)
1197 * If the request set has a producer callback, the RPC must be
1198 * sent straight away
1200 ptlrpc_send_new_req(req);
1202 EXPORT_SYMBOL(ptlrpc_set_add_req);
1205 * Add a request to a request with dedicated server thread
1206 * and wake the thread to make any necessary processing.
1207 * Currently only used for ptlrpcd.
1209 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1210 struct ptlrpc_request *req)
1212 struct ptlrpc_request_set *set = pc->pc_set;
1215 LASSERT(req->rq_set == NULL);
1216 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1218 spin_lock(&set->set_new_req_lock);
1220 * The set takes over the caller's request reference.
1223 req->rq_queued_time = ktime_get_seconds();
1224 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1225 count = atomic_inc_return(&set->set_new_count);
1226 spin_unlock(&set->set_new_req_lock);
1228 /* Only need to call wakeup once for the first entry. */
1230 wake_up(&set->set_waitq);
1233 * XXX: It maybe unnecessary to wakeup all the partners. But to
1234 * guarantee the async RPC can be processed ASAP, we have
1235 * no other better choice. It maybe fixed in future.
1237 for (i = 0; i < pc->pc_npartners; i++)
1238 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1243 * Based on the current state of the import, determine if the request
1244 * can be sent, is an error, or should be delayed.
1246 * Returns true if this request should be delayed. If false, and
1247 * *status is set, then the request can not be sent and *status is the
1248 * error code. If false and status is 0, then request can be sent.
1250 * The imp->imp_lock must be held.
1252 static int ptlrpc_import_delay_req(struct obd_import *imp,
1253 struct ptlrpc_request *req, int *status)
1261 if (req->rq_ctx_init || req->rq_ctx_fini) {
1262 /* always allow ctx init/fini rpc go through */
1263 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1264 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1266 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1267 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1270 * pings or MDS-equivalent STATFS may safely
1273 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1274 D_HA : D_ERROR, req, "IMP_CLOSED");
1276 } else if (ptlrpc_send_limit_expired(req)) {
1277 /* probably doesn't need to be a D_ERROR afterinitial testing */
1278 DEBUG_REQ(D_HA, req, "send limit expired");
1279 *status = -ETIMEDOUT;
1280 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1281 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1282 ;/* allow CONNECT even if import is invalid */
1283 if (atomic_read(&imp->imp_inval_count) != 0) {
1284 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1287 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1288 if (!imp->imp_deactive)
1289 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1290 *status = -ESHUTDOWN; /* b=12940 */
1291 } else if (req->rq_import_generation != imp->imp_generation) {
1292 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1294 } else if (req->rq_send_state != imp->imp_state) {
1295 /* invalidate in progress - any requests should be drop */
1296 if (atomic_read(&imp->imp_inval_count) != 0) {
1297 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1299 } else if (req->rq_no_delay &&
1300 imp->imp_generation != imp->imp_initiated_at) {
1301 /* ignore nodelay for requests initiating connections */
1302 *status = -EWOULDBLOCK;
1303 } else if (req->rq_allow_replay &&
1304 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1305 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1306 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1307 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1308 DEBUG_REQ(D_HA, req, "allow during recovery");
1318 * Decide if the error message should be printed to the console or not.
1319 * Makes its decision based on request type, status, and failure frequency.
1321 * \param[in] req request that failed and may need a console message
1323 * \retval false if no message should be printed
1324 * \retval true if console message should be printed
1326 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1328 LASSERT(req->rq_reqmsg != NULL);
1330 /* Suppress particular reconnect errors which are to be expected. */
1331 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1332 /* Suppress timed out reconnect requests */
1333 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1338 * Suppress most unavailable/again reconnect requests, but
1339 * print occasionally so it is clear client is trying to
1340 * connect to a server where no target is running.
1342 if ((err == -ENODEV || err == -EAGAIN) &&
1343 req->rq_import->imp_conn_cnt % 30 != 20)
1347 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1348 /* -EAGAIN is normal when using POSIX flocks */
1351 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1352 (req->rq_xid & 0xf) != 10)
1353 /* Suppress most ping requests, they may fail occasionally */
1360 * Check request processing status.
1361 * Returns the status.
1363 static int ptlrpc_check_status(struct ptlrpc_request *req)
1368 rc = lustre_msg_get_status(req->rq_repmsg);
1369 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1370 struct obd_import *imp = req->rq_import;
1371 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1372 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1374 if (ptlrpc_console_allow(req, opc, rc))
1375 LCONSOLE_ERROR_MSG(0x11,
1376 "%s: operation %s to node %s failed: rc = %d\n",
1377 imp->imp_obd->obd_name,
1379 libcfs_nid2str(nid), rc);
1380 RETURN(rc < 0 ? rc : -EINVAL);
1384 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1390 * save pre-versions of objects into request for replay.
1391 * Versions are obtained from server reply.
1394 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1396 struct lustre_msg *repmsg = req->rq_repmsg;
1397 struct lustre_msg *reqmsg = req->rq_reqmsg;
1398 __u64 *versions = lustre_msg_get_versions(repmsg);
1401 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1405 lustre_msg_set_versions(reqmsg, versions);
1406 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1407 versions[0], versions[1]);
1412 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1414 struct ptlrpc_request *req;
1416 assert_spin_locked(&imp->imp_lock);
1417 if (list_empty(&imp->imp_unreplied_list))
1420 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1422 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1424 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1425 imp->imp_known_replied_xid = req->rq_xid - 1;
1427 return req->rq_xid - 1;
1431 * Callback function called when client receives RPC reply for \a req.
1432 * Returns 0 on success or error code.
1433 * The return alue would be assigned to req->rq_status by the caller
1434 * as request processing status.
1435 * This function also decides if the request needs to be saved for later replay.
1437 static int after_reply(struct ptlrpc_request *req)
1439 struct obd_import *imp = req->rq_import;
1440 struct obd_device *obd = req->rq_import->imp_obd;
1447 LASSERT(obd != NULL);
1448 /* repbuf must be unlinked */
1449 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1451 if (req->rq_reply_truncated) {
1452 if (ptlrpc_no_resend(req)) {
1453 DEBUG_REQ(D_ERROR, req,
1454 "reply buffer overflow, expected=%d, actual size=%d",
1455 req->rq_nob_received, req->rq_repbuf_len);
1459 sptlrpc_cli_free_repbuf(req);
1461 * Pass the required reply buffer size (include
1462 * space for early reply).
1463 * NB: no need to roundup because alloc_repbuf
1466 req->rq_replen = req->rq_nob_received;
1467 req->rq_nob_received = 0;
1468 spin_lock(&req->rq_lock);
1470 spin_unlock(&req->rq_lock);
1474 work_start = ktime_get_real();
1475 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1478 * NB Until this point, the whole of the incoming message,
1479 * including buflens, status etc is in the sender's byte order.
1481 rc = sptlrpc_cli_unwrap_reply(req);
1483 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1488 * Security layer unwrap might ask resend this request.
1493 rc = unpack_reply(req);
1497 /* retry indefinitely on EINPROGRESS */
1498 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1499 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1500 time64_t now = ktime_get_real_seconds();
1502 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1503 D_RPCTRACE, req, "resending request on EINPROGRESS");
1504 spin_lock(&req->rq_lock);
1506 spin_unlock(&req->rq_lock);
1507 req->rq_nr_resend++;
1509 /* Readjust the timeout for current conditions */
1510 ptlrpc_at_set_req_timeout(req);
1512 * delay resend to give a chance to the server to get ready.
1513 * The delay is increased by 1s on every resend and is capped to
1514 * the current request timeout (i.e. obd_timeout if AT is off,
1515 * or AT service time x 125% + 5s, see at_est2timeout)
1517 if (req->rq_nr_resend > req->rq_timeout)
1518 req->rq_sent = now + req->rq_timeout;
1520 req->rq_sent = now + req->rq_nr_resend;
1522 /* Resend for EINPROGRESS will use a new XID */
1523 spin_lock(&imp->imp_lock);
1524 list_del_init(&req->rq_unreplied_list);
1525 spin_unlock(&imp->imp_lock);
1530 if (obd->obd_svc_stats) {
1531 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1533 ptlrpc_lprocfs_rpc_sent(req, timediff);
1536 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1537 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1538 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1539 lustre_msg_get_type(req->rq_repmsg));
1543 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1544 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1545 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1546 ptlrpc_at_adj_net_latency(req,
1547 lustre_msg_get_service_time(req->rq_repmsg));
1549 rc = ptlrpc_check_status(req);
1553 * Either we've been evicted, or the server has failed for
1554 * some reason. Try to reconnect, and if that fails, punt to
1557 if (ptlrpc_recoverable_error(rc)) {
1558 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1559 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1562 ptlrpc_request_handle_notconn(req);
1567 * Let's look if server sent slv. Do it only for RPC with
1570 ldlm_cli_update_pool(req);
1574 * Store transno in reqmsg for replay.
1576 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1577 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1578 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1581 if (imp->imp_replayable) {
1582 spin_lock(&imp->imp_lock);
1584 * No point in adding already-committed requests to the replay
1585 * list, we will just remove them immediately. b=9829
1587 if (req->rq_transno != 0 &&
1589 lustre_msg_get_last_committed(req->rq_repmsg) ||
1591 /** version recovery */
1592 ptlrpc_save_versions(req);
1593 ptlrpc_retain_replayable_request(req, imp);
1594 } else if (req->rq_commit_cb &&
1595 list_empty(&req->rq_replay_list)) {
1597 * NB: don't call rq_commit_cb if it's already on
1598 * rq_replay_list, ptlrpc_free_committed() will call
1599 * it later, see LU-3618 for details
1601 spin_unlock(&imp->imp_lock);
1602 req->rq_commit_cb(req);
1603 spin_lock(&imp->imp_lock);
1607 * Replay-enabled imports return commit-status information.
1609 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1610 if (likely(committed > imp->imp_peer_committed_transno))
1611 imp->imp_peer_committed_transno = committed;
1613 ptlrpc_free_committed(imp);
1615 if (!list_empty(&imp->imp_replay_list)) {
1616 struct ptlrpc_request *last;
1618 last = list_entry(imp->imp_replay_list.prev,
1619 struct ptlrpc_request,
1622 * Requests with rq_replay stay on the list even if no
1623 * commit is expected.
1625 if (last->rq_transno > imp->imp_peer_committed_transno)
1626 ptlrpc_pinger_commit_expected(imp);
1629 spin_unlock(&imp->imp_lock);
1636 * Helper function to send request \a req over the network for the first time
1637 * Also adjusts request phase.
1638 * Returns 0 on success or error code.
1640 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1642 struct obd_import *imp = req->rq_import;
1647 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1649 /* do not try to go further if there is not enough memory in enc_pool */
1650 if (req->rq_sent && req->rq_bulk)
1651 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1652 pool_is_at_full_capacity())
1655 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1656 (!req->rq_generation_set ||
1657 req->rq_import_generation == imp->imp_generation))
1660 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1662 spin_lock(&imp->imp_lock);
1664 LASSERT(req->rq_xid != 0);
1665 LASSERT(!list_empty(&req->rq_unreplied_list));
1667 if (!req->rq_generation_set)
1668 req->rq_import_generation = imp->imp_generation;
1670 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1671 spin_lock(&req->rq_lock);
1672 req->rq_waiting = 1;
1673 spin_unlock(&req->rq_lock);
1675 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1676 ptlrpc_import_state_name(req->rq_send_state),
1677 ptlrpc_import_state_name(imp->imp_state));
1678 LASSERT(list_empty(&req->rq_list));
1679 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1680 atomic_inc(&req->rq_import->imp_inflight);
1681 spin_unlock(&imp->imp_lock);
1686 spin_unlock(&imp->imp_lock);
1687 req->rq_status = rc;
1688 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1692 LASSERT(list_empty(&req->rq_list));
1693 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1694 atomic_inc(&req->rq_import->imp_inflight);
1697 * find the known replied XID from the unreplied list, CONNECT
1698 * and DISCONNECT requests are skipped to make the sanity check
1699 * on server side happy. see process_req_last_xid().
1701 * For CONNECT: Because replay requests have lower XID, it'll
1702 * break the sanity check if CONNECT bump the exp_last_xid on
1705 * For DISCONNECT: Since client will abort inflight RPC before
1706 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1707 * than the inflight RPC.
1709 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1710 min_xid = ptlrpc_known_replied_xid(imp);
1711 spin_unlock(&imp->imp_lock);
1713 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1715 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1717 rc = sptlrpc_req_refresh_ctx(req, -1);
1720 req->rq_status = rc;
1723 spin_lock(&req->rq_lock);
1724 req->rq_wait_ctx = 1;
1725 spin_unlock(&req->rq_lock);
1731 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1732 req, current_comm(),
1733 imp->imp_obd->obd_uuid.uuid,
1734 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1735 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1736 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1738 rc = ptl_send_rpc(req, 0);
1739 if (rc == -ENOMEM) {
1740 spin_lock(&imp->imp_lock);
1741 if (!list_empty(&req->rq_list)) {
1742 list_del_init(&req->rq_list);
1743 atomic_dec(&req->rq_import->imp_inflight);
1745 spin_unlock(&imp->imp_lock);
1746 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1750 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1752 spin_lock(&req->rq_lock);
1753 req->rq_net_err = 1;
1754 spin_unlock(&req->rq_lock);
1760 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1765 LASSERT(set->set_producer != NULL);
1767 remaining = atomic_read(&set->set_remaining);
1770 * populate the ->set_requests list with requests until we
1771 * reach the maximum number of RPCs in flight for this set
1773 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1774 rc = set->set_producer(set, set->set_producer_arg);
1775 if (rc == -ENOENT) {
1776 /* no more RPC to produce */
1777 set->set_producer = NULL;
1778 set->set_producer_arg = NULL;
1783 RETURN((atomic_read(&set->set_remaining) - remaining));
1787 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1788 * and no more replies are expected.
1789 * (it is possible to get less replies than requests sent e.g. due to timed out
1790 * requests or requests that we had trouble to send out)
1792 * NOTE: This function contains a potential schedule point (cond_resched()).
1794 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1796 struct list_head *tmp, *next;
1797 LIST_HEAD(comp_reqs);
1798 int force_timer_recalc = 0;
1801 if (atomic_read(&set->set_remaining) == 0)
1804 list_for_each_safe(tmp, next, &set->set_requests) {
1805 struct ptlrpc_request *req =
1806 list_entry(tmp, struct ptlrpc_request,
1808 struct obd_import *imp = req->rq_import;
1809 int unregistered = 0;
1813 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1814 list_move_tail(&req->rq_set_chain, &comp_reqs);
1819 * This schedule point is mainly for the ptlrpcd caller of this
1820 * function. Most ptlrpc sets are not long-lived and unbounded
1821 * in length, but at the least the set used by the ptlrpcd is.
1822 * Since the processing time is unbounded, we need to insert an
1823 * explicit schedule point to make the thread well-behaved.
1828 * If the caller requires to allow to be interpreted by force
1829 * and it has really been interpreted, then move the request
1830 * to RQ_PHASE_INTERPRET phase in spite of what the current
1833 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1834 req->rq_status = -EINTR;
1835 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1838 * Since it is interpreted and we have to wait for
1839 * the reply to be unlinked, then use sync mode.
1843 GOTO(interpret, req->rq_status);
1846 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1847 force_timer_recalc = 1;
1849 /* delayed send - skip */
1850 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1853 /* delayed resend - skip */
1854 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1855 req->rq_sent > ktime_get_real_seconds())
1858 if (!(req->rq_phase == RQ_PHASE_RPC ||
1859 req->rq_phase == RQ_PHASE_BULK ||
1860 req->rq_phase == RQ_PHASE_INTERPRET ||
1861 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1862 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1863 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1867 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1868 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1869 LASSERT(req->rq_next_phase != req->rq_phase);
1870 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1872 if (req->rq_req_deadline &&
1873 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1874 req->rq_req_deadline = 0;
1875 if (req->rq_reply_deadline &&
1876 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1877 req->rq_reply_deadline = 0;
1878 if (req->rq_bulk_deadline &&
1879 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1880 req->rq_bulk_deadline = 0;
1883 * Skip processing until reply is unlinked. We
1884 * can't return to pool before that and we can't
1885 * call interpret before that. We need to make
1886 * sure that all rdma transfers finished and will
1887 * not corrupt any data.
1889 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1890 ptlrpc_client_recv_or_unlink(req))
1892 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1893 ptlrpc_client_bulk_active(req))
1897 * Turn fail_loc off to prevent it from looping
1900 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1901 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1904 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1905 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1910 * Move to next phase if reply was successfully
1913 ptlrpc_rqphase_move(req, req->rq_next_phase);
1916 if (req->rq_phase == RQ_PHASE_INTERPRET)
1917 GOTO(interpret, req->rq_status);
1920 * Note that this also will start async reply unlink.
1922 if (req->rq_net_err && !req->rq_timedout) {
1923 ptlrpc_expire_one_request(req, 1);
1926 * Check if we still need to wait for unlink.
1928 if (ptlrpc_client_recv_or_unlink(req) ||
1929 ptlrpc_client_bulk_active(req))
1931 /* If there is no need to resend, fail it now. */
1932 if (req->rq_no_resend) {
1933 if (req->rq_status == 0)
1934 req->rq_status = -EIO;
1935 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1936 GOTO(interpret, req->rq_status);
1943 spin_lock(&req->rq_lock);
1944 req->rq_replied = 0;
1945 spin_unlock(&req->rq_lock);
1946 if (req->rq_status == 0)
1947 req->rq_status = -EIO;
1948 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1949 GOTO(interpret, req->rq_status);
1953 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1954 * so it sets rq_intr regardless of individual rpc
1955 * timeouts. The synchronous IO waiting path sets
1956 * rq_intr irrespective of whether ptlrpcd
1957 * has seen a timeout. Our policy is to only interpret
1958 * interrupted rpcs after they have timed out, so we
1959 * need to enforce that here.
1962 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1963 req->rq_wait_ctx)) {
1964 req->rq_status = -EINTR;
1965 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1966 GOTO(interpret, req->rq_status);
1969 if (req->rq_phase == RQ_PHASE_RPC) {
1970 if (req->rq_timedout || req->rq_resend ||
1971 req->rq_waiting || req->rq_wait_ctx) {
1974 if (!ptlrpc_unregister_reply(req, 1)) {
1975 ptlrpc_unregister_bulk(req, 1);
1979 spin_lock(&imp->imp_lock);
1980 if (ptlrpc_import_delay_req(imp, req,
1983 * put on delay list - only if we wait
1984 * recovery finished - before send
1986 list_move_tail(&req->rq_list,
1987 &imp->imp_delayed_list);
1988 spin_unlock(&imp->imp_lock);
1993 req->rq_status = status;
1994 ptlrpc_rqphase_move(req,
1995 RQ_PHASE_INTERPRET);
1996 spin_unlock(&imp->imp_lock);
1997 GOTO(interpret, req->rq_status);
1999 /* ignore on just initiated connections */
2000 if (ptlrpc_no_resend(req) &&
2001 !req->rq_wait_ctx &&
2002 imp->imp_generation !=
2003 imp->imp_initiated_at) {
2004 req->rq_status = -ENOTCONN;
2005 ptlrpc_rqphase_move(req,
2006 RQ_PHASE_INTERPRET);
2007 spin_unlock(&imp->imp_lock);
2008 GOTO(interpret, req->rq_status);
2011 list_move_tail(&req->rq_list,
2012 &imp->imp_sending_list);
2014 spin_unlock(&imp->imp_lock);
2016 spin_lock(&req->rq_lock);
2017 req->rq_waiting = 0;
2018 spin_unlock(&req->rq_lock);
2020 if (req->rq_timedout || req->rq_resend) {
2022 * This is re-sending anyways,
2023 * let's mark req as resend.
2025 spin_lock(&req->rq_lock);
2027 spin_unlock(&req->rq_lock);
2030 * rq_wait_ctx is only touched by ptlrpcd,
2031 * so no lock is needed here.
2033 status = sptlrpc_req_refresh_ctx(req, -1);
2036 req->rq_status = status;
2037 spin_lock(&req->rq_lock);
2038 req->rq_wait_ctx = 0;
2039 spin_unlock(&req->rq_lock);
2040 force_timer_recalc = 1;
2042 spin_lock(&req->rq_lock);
2043 req->rq_wait_ctx = 1;
2044 spin_unlock(&req->rq_lock);
2049 spin_lock(&req->rq_lock);
2050 req->rq_wait_ctx = 0;
2051 spin_unlock(&req->rq_lock);
2055 * In any case, the previous bulk should be
2056 * cleaned up to prepare for the new sending
2059 !ptlrpc_unregister_bulk(req, 1))
2062 rc = ptl_send_rpc(req, 0);
2063 if (rc == -ENOMEM) {
2064 spin_lock(&imp->imp_lock);
2065 if (!list_empty(&req->rq_list))
2066 list_del_init(&req->rq_list);
2067 spin_unlock(&imp->imp_lock);
2068 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2072 DEBUG_REQ(D_HA, req,
2073 "send failed: rc = %d", rc);
2074 force_timer_recalc = 1;
2075 spin_lock(&req->rq_lock);
2076 req->rq_net_err = 1;
2077 spin_unlock(&req->rq_lock);
2080 /* need to reset the timeout */
2081 force_timer_recalc = 1;
2084 spin_lock(&req->rq_lock);
2086 if (ptlrpc_client_early(req)) {
2087 ptlrpc_at_recv_early_reply(req);
2088 spin_unlock(&req->rq_lock);
2092 /* Still waiting for a reply? */
2093 if (ptlrpc_client_recv(req)) {
2094 spin_unlock(&req->rq_lock);
2098 /* Did we actually receive a reply? */
2099 if (!ptlrpc_client_replied(req)) {
2100 spin_unlock(&req->rq_lock);
2104 spin_unlock(&req->rq_lock);
2107 * unlink from net because we are going to
2108 * swab in-place of reply buffer
2110 unregistered = ptlrpc_unregister_reply(req, 1);
2114 req->rq_status = after_reply(req);
2119 * If there is no bulk associated with this request,
2120 * then we're done and should let the interpreter
2121 * process the reply. Similarly if the RPC returned
2122 * an error, and therefore the bulk will never arrive.
2124 if (!req->rq_bulk || req->rq_status < 0) {
2125 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2126 GOTO(interpret, req->rq_status);
2129 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2132 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2133 if (ptlrpc_client_bulk_active(req))
2136 if (req->rq_bulk->bd_failure) {
2138 * The RPC reply arrived OK, but the bulk screwed
2139 * up! Dead weird since the server told us the RPC
2140 * was good after getting the REPLY for her GET or
2141 * the ACK for her PUT.
2143 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2144 req->rq_status = -EIO;
2147 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2150 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2153 * This moves to "unregistering" phase we need to wait for
2156 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2157 /* start async bulk unlink too */
2158 ptlrpc_unregister_bulk(req, 1);
2162 if (!ptlrpc_unregister_bulk(req, async))
2166 * When calling interpret receiving already should be
2169 LASSERT(!req->rq_receiving_reply);
2171 ptlrpc_req_interpret(env, req, req->rq_status);
2173 if (ptlrpcd_check_work(req)) {
2174 atomic_dec(&set->set_remaining);
2177 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2181 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2182 req, current_comm(),
2183 imp->imp_obd->obd_uuid.uuid,
2184 lustre_msg_get_status(req->rq_reqmsg),
2186 obd_import_nid2str(imp),
2187 lustre_msg_get_opc(req->rq_reqmsg),
2188 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2190 spin_lock(&imp->imp_lock);
2192 * Request already may be not on sending or delaying list. This
2193 * may happen in the case of marking it erroneous for the case
2194 * ptlrpc_import_delay_req(req, status) find it impossible to
2195 * allow sending this rpc and returns *status != 0.
2197 if (!list_empty(&req->rq_list)) {
2198 list_del_init(&req->rq_list);
2199 atomic_dec(&imp->imp_inflight);
2201 list_del_init(&req->rq_unreplied_list);
2202 spin_unlock(&imp->imp_lock);
2204 atomic_dec(&set->set_remaining);
2205 wake_up_all(&imp->imp_recovery_waitq);
2207 if (set->set_producer) {
2208 /* produce a new request if possible */
2209 if (ptlrpc_set_producer(set) > 0)
2210 force_timer_recalc = 1;
2213 * free the request that has just been completed
2214 * in order not to pollute set->set_requests
2216 list_del_init(&req->rq_set_chain);
2217 spin_lock(&req->rq_lock);
2219 req->rq_invalid_rqset = 0;
2220 spin_unlock(&req->rq_lock);
2222 /* record rq_status to compute the final status later */
2223 if (req->rq_status != 0)
2224 set->set_rc = req->rq_status;
2225 ptlrpc_req_finished(req);
2227 list_move_tail(&req->rq_set_chain, &comp_reqs);
2232 * move completed request at the head of list so it's easier for
2233 * caller to find them
2235 list_splice(&comp_reqs, &set->set_requests);
2237 /* If we hit an error, we want to recover promptly. */
2238 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2240 EXPORT_SYMBOL(ptlrpc_check_set);
2243 * Time out request \a req. is \a async_unlink is set, that means do not wait
2244 * until LNet actually confirms network buffer unlinking.
2245 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2247 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2249 struct obd_import *imp = req->rq_import;
2250 unsigned int debug_mask = D_RPCTRACE;
2254 spin_lock(&req->rq_lock);
2255 req->rq_timedout = 1;
2256 spin_unlock(&req->rq_lock);
2258 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2259 lustre_msg_get_status(req->rq_reqmsg)))
2260 debug_mask = D_WARNING;
2261 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2262 req->rq_net_err ? "failed due to network error" :
2263 ((req->rq_real_sent == 0 ||
2264 req->rq_real_sent < req->rq_sent ||
2265 req->rq_real_sent >= req->rq_deadline) ?
2266 "timed out for sent delay" : "timed out for slow reply"),
2267 (s64)req->rq_sent, (s64)req->rq_real_sent);
2269 if (imp && obd_debug_peer_on_timeout)
2270 LNetDebugPeer(imp->imp_connection->c_peer);
2272 ptlrpc_unregister_reply(req, async_unlink);
2273 ptlrpc_unregister_bulk(req, async_unlink);
2275 if (obd_dump_on_timeout)
2276 libcfs_debug_dumplog();
2279 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2283 atomic_inc(&imp->imp_timeouts);
2285 /* The DLM server doesn't want recovery run on its imports. */
2286 if (imp->imp_dlm_fake)
2290 * If this request is for recovery or other primordial tasks,
2291 * then error it out here.
2293 if (req->rq_ctx_init || req->rq_ctx_fini ||
2294 req->rq_send_state != LUSTRE_IMP_FULL ||
2295 imp->imp_obd->obd_no_recov) {
2296 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2297 ptlrpc_import_state_name(req->rq_send_state),
2298 ptlrpc_import_state_name(imp->imp_state));
2299 spin_lock(&req->rq_lock);
2300 req->rq_status = -ETIMEDOUT;
2302 spin_unlock(&req->rq_lock);
2307 * if a request can't be resent we can't wait for an answer after
2310 if (ptlrpc_no_resend(req)) {
2311 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2315 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2321 * Time out all uncompleted requests in request set pointed by \a data
2322 * Callback used when waiting on sets with l_wait_event.
2325 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2327 struct list_head *tmp;
2328 time64_t now = ktime_get_real_seconds();
2331 LASSERT(set != NULL);
2334 * A timeout expired. See which reqs it applies to...
2336 list_for_each(tmp, &set->set_requests) {
2337 struct ptlrpc_request *req =
2338 list_entry(tmp, struct ptlrpc_request,
2341 /* don't expire request waiting for context */
2342 if (req->rq_wait_ctx)
2345 /* Request in-flight? */
2346 if (!((req->rq_phase == RQ_PHASE_RPC &&
2347 !req->rq_waiting && !req->rq_resend) ||
2348 (req->rq_phase == RQ_PHASE_BULK)))
2351 if (req->rq_timedout || /* already dealt with */
2352 req->rq_deadline > now) /* not expired */
2356 * Deal with this guy. Do it asynchronously to not block
2359 ptlrpc_expire_one_request(req, 1);
2364 * Sets rq_intr flag in \a req under spinlock.
2366 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2368 spin_lock(&req->rq_lock);
2370 spin_unlock(&req->rq_lock);
2372 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2375 * Interrupts (sets interrupted flag) all uncompleted requests in
2376 * a set \a data. Callback for l_wait_event for interruptible waits.
2378 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2380 struct list_head *tmp;
2382 LASSERT(set != NULL);
2383 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2385 list_for_each(tmp, &set->set_requests) {
2386 struct ptlrpc_request *req =
2387 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2392 if (req->rq_phase != RQ_PHASE_RPC &&
2393 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2394 !req->rq_allow_intr)
2397 ptlrpc_mark_interrupted(req);
2402 * Get the smallest timeout in the set; this does NOT set a timeout.
2404 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2406 struct list_head *tmp;
2407 time64_t now = ktime_get_real_seconds();
2409 struct ptlrpc_request *req;
2413 list_for_each(tmp, &set->set_requests) {
2414 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2416 /* Request in-flight? */
2417 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2418 (req->rq_phase == RQ_PHASE_BULK) ||
2419 (req->rq_phase == RQ_PHASE_NEW)))
2422 /* Already timed out. */
2423 if (req->rq_timedout)
2426 /* Waiting for ctx. */
2427 if (req->rq_wait_ctx)
2430 if (req->rq_phase == RQ_PHASE_NEW)
2431 deadline = req->rq_sent;
2432 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2433 deadline = req->rq_sent;
2435 deadline = req->rq_sent + req->rq_timeout;
2437 if (deadline <= now) /* actually expired already */
2438 timeout = 1; /* ASAP */
2439 else if (timeout == 0 || timeout > deadline - now)
2440 timeout = deadline - now;
2446 * Send all unset request from the set and then wait untill all
2447 * requests in the set complete (either get a reply, timeout, get an
2448 * error or otherwise be interrupted).
2449 * Returns 0 on success or error code otherwise.
2451 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2453 struct list_head *tmp;
2454 struct ptlrpc_request *req;
2459 if (set->set_producer)
2460 (void)ptlrpc_set_producer(set);
2462 list_for_each(tmp, &set->set_requests) {
2463 req = list_entry(tmp, struct ptlrpc_request,
2465 if (req->rq_phase == RQ_PHASE_NEW)
2466 (void)ptlrpc_send_new_req(req);
2469 if (list_empty(&set->set_requests))
2473 timeout = ptlrpc_set_next_timeout(set);
2476 * wait until all complete, interrupted, or an in-flight
2479 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2482 if ((timeout == 0 && !signal_pending(current)) ||
2483 set->set_allow_intr) {
2485 * No requests are in-flight (ether timed out
2486 * or delayed), so we can allow interrupts.
2487 * We still want to block for a limited time,
2488 * so we allow interrupts during the timeout.
2490 rc = l_wait_event_abortable_timeout(
2492 ptlrpc_check_set(NULL, set),
2493 cfs_time_seconds(timeout ? timeout : 1));
2496 ptlrpc_expired_set(set);
2497 } else if (rc < 0) {
2499 ptlrpc_interrupted_set(set);
2505 * At least one request is in flight, so no
2506 * interrupts are allowed. Wait until all
2507 * complete, or an in-flight req times out.
2509 rc = wait_event_idle_timeout(
2511 ptlrpc_check_set(NULL, set),
2512 cfs_time_seconds(timeout ? timeout : 1));
2514 ptlrpc_expired_set(set);
2521 * LU-769 - if we ignored the signal because
2522 * it was already pending when we started, we
2523 * need to handle it now or we risk it being
2526 if (rc == -ETIMEDOUT &&
2527 signal_pending(current)) {
2528 sigset_t blocked_sigs =
2529 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2532 * In fact we only interrupt for the
2533 * "fatal" signals like SIGINT or
2534 * SIGKILL. We still ignore less
2535 * important signals since ptlrpc set
2536 * is not easily reentrant from
2539 if (signal_pending(current))
2540 ptlrpc_interrupted_set(set);
2541 cfs_restore_sigs(blocked_sigs);
2545 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2548 * -EINTR => all requests have been flagged rq_intr so next
2550 * -ETIMEDOUT => someone timed out. When all reqs have
2551 * timed out, signals are enabled allowing completion with
2553 * I don't really care if we go once more round the loop in
2554 * the error cases -eeb.
2556 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2557 list_for_each(tmp, &set->set_requests) {
2558 req = list_entry(tmp, struct ptlrpc_request,
2560 spin_lock(&req->rq_lock);
2561 req->rq_invalid_rqset = 1;
2562 spin_unlock(&req->rq_lock);
2565 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2567 LASSERT(atomic_read(&set->set_remaining) == 0);
2569 rc = set->set_rc; /* rq_status of already freed requests if any */
2570 list_for_each(tmp, &set->set_requests) {
2571 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2573 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2574 if (req->rq_status != 0)
2575 rc = req->rq_status;
2580 EXPORT_SYMBOL(ptlrpc_set_wait);
2583 * Helper fuction for request freeing.
2584 * Called when request count reached zero and request needs to be freed.
2585 * Removes request from all sorts of sending/replay lists it might be on,
2586 * frees network buffers if any are present.
2587 * If \a locked is set, that means caller is already holding import imp_lock
2588 * and so we no longer need to reobtain it (for certain lists manipulations)
2590 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2597 LASSERT(!request->rq_srv_req);
2598 LASSERT(request->rq_export == NULL);
2599 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2600 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2601 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2602 LASSERTF(!request->rq_replay, "req %p\n", request);
2604 req_capsule_fini(&request->rq_pill);
2607 * We must take it off the imp_replay_list first. Otherwise, we'll set
2608 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2610 if (request->rq_import) {
2612 spin_lock(&request->rq_import->imp_lock);
2613 list_del_init(&request->rq_replay_list);
2614 list_del_init(&request->rq_unreplied_list);
2616 spin_unlock(&request->rq_import->imp_lock);
2618 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2620 if (atomic_read(&request->rq_refcount) != 0) {
2621 DEBUG_REQ(D_ERROR, request,
2622 "freeing request with nonzero refcount");
2626 if (request->rq_repbuf)
2627 sptlrpc_cli_free_repbuf(request);
2629 if (request->rq_import) {
2630 class_import_put(request->rq_import);
2631 request->rq_import = NULL;
2633 if (request->rq_bulk)
2634 ptlrpc_free_bulk(request->rq_bulk);
2636 if (request->rq_reqbuf || request->rq_clrbuf)
2637 sptlrpc_cli_free_reqbuf(request);
2639 if (request->rq_cli_ctx)
2640 sptlrpc_req_put_ctx(request, !locked);
2642 if (request->rq_pool)
2643 __ptlrpc_free_req_to_pool(request);
2645 ptlrpc_request_cache_free(request);
2649 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2651 * Drop one request reference. Must be called with import imp_lock held.
2652 * When reference count drops to zero, request is freed.
2654 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2656 assert_spin_locked(&request->rq_import->imp_lock);
2657 (void)__ptlrpc_req_finished(request, 1);
2662 * Drops one reference count for request \a request.
2663 * \a locked set indicates that caller holds import imp_lock.
2664 * Frees the request whe reference count reaches zero.
2666 * \retval 1 the request is freed
2667 * \retval 0 some others still hold references on the request
2669 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2677 LASSERT(request != LP_POISON);
2678 LASSERT(request->rq_reqmsg != LP_POISON);
2680 DEBUG_REQ(D_INFO, request, "refcount now %u",
2681 atomic_read(&request->rq_refcount) - 1);
2683 spin_lock(&request->rq_lock);
2684 count = atomic_dec_return(&request->rq_refcount);
2685 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2688 * For open RPC, the client does not know the EA size (LOV, ACL, and
2689 * so on) before replied, then the client has to reserve very large
2690 * reply buffer. Such buffer will not be released until the RPC freed.
2691 * Since The open RPC is replayable, we need to keep it in the replay
2692 * list until close. If there are a lot of files opened concurrently,
2693 * then the client may be OOM.
2695 * If fact, it is unnecessary to keep reply buffer for open replay,
2696 * related EAs have already been saved via mdc_save_lovea() before
2697 * coming here. So it is safe to free the reply buffer some earlier
2698 * before releasing the RPC to avoid client OOM. LU-9514
2700 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2701 spin_lock(&request->rq_early_free_lock);
2702 sptlrpc_cli_free_repbuf(request);
2703 request->rq_repbuf = NULL;
2704 request->rq_repbuf_len = 0;
2705 request->rq_repdata = NULL;
2706 request->rq_reqdata_len = 0;
2707 spin_unlock(&request->rq_early_free_lock);
2709 spin_unlock(&request->rq_lock);
2712 __ptlrpc_free_req(request, locked);
2718 * Drops one reference count for a request.
2720 void ptlrpc_req_finished(struct ptlrpc_request *request)
2722 __ptlrpc_req_finished(request, 0);
2724 EXPORT_SYMBOL(ptlrpc_req_finished);
2727 * Returns xid of a \a request
2729 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2731 return request->rq_xid;
2733 EXPORT_SYMBOL(ptlrpc_req_xid);
2736 * Disengage the client's reply buffer from the network
2737 * NB does _NOT_ unregister any client-side bulk.
2738 * IDEMPOTENT, but _not_ safe against concurrent callers.
2739 * The request owner (i.e. the thread doing the I/O) must call...
2740 * Returns 0 on success or 1 if unregistering cannot be made.
2742 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2747 LASSERT(!in_interrupt());
2749 /* Let's setup deadline for reply unlink. */
2750 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2751 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2752 request->rq_reply_deadline = ktime_get_real_seconds() +
2756 * Nothing left to do.
2758 if (!ptlrpc_client_recv_or_unlink(request))
2761 LNetMDUnlink(request->rq_reply_md_h);
2764 * Let's check it once again.
2766 if (!ptlrpc_client_recv_or_unlink(request))
2769 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2770 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2773 * Do not wait for unlink to finish.
2779 * We have to l_wait_event() whatever the result, to give liblustre
2780 * a chance to run reply_in_callback(), and to make sure we've
2781 * unlinked before returning a req to the pool.
2784 wait_queue_head_t *wq = (request->rq_set) ?
2785 &request->rq_set->set_waitq :
2786 &request->rq_reply_waitq;
2787 int seconds = LONG_UNLINK;
2789 * Network access will complete in finite time but the HUGE
2790 * timeout lets us CWARN for visibility of sluggish NALs
2792 while (seconds > 0 &&
2793 wait_event_idle_timeout(
2795 !ptlrpc_client_recv_or_unlink(request),
2796 cfs_time_seconds(1)) == 0)
2799 ptlrpc_rqphase_move(request, request->rq_next_phase);
2803 DEBUG_REQ(D_WARNING, request,
2804 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2805 request->rq_receiving_reply,
2806 request->rq_req_unlinked,
2807 request->rq_reply_unlinked);
2812 static void ptlrpc_free_request(struct ptlrpc_request *req)
2814 spin_lock(&req->rq_lock);
2816 spin_unlock(&req->rq_lock);
2818 if (req->rq_commit_cb)
2819 req->rq_commit_cb(req);
2820 list_del_init(&req->rq_replay_list);
2822 __ptlrpc_req_finished(req, 1);
2826 * the request is committed and dropped from the replay list of its import
2828 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2830 struct obd_import *imp = req->rq_import;
2832 spin_lock(&imp->imp_lock);
2833 if (list_empty(&req->rq_replay_list)) {
2834 spin_unlock(&imp->imp_lock);
2838 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2839 if (imp->imp_replay_cursor == &req->rq_replay_list)
2840 imp->imp_replay_cursor = req->rq_replay_list.next;
2841 ptlrpc_free_request(req);
2844 spin_unlock(&imp->imp_lock);
2846 EXPORT_SYMBOL(ptlrpc_request_committed);
2849 * Iterates through replay_list on import and prunes
2850 * all requests have transno smaller than last_committed for the
2851 * import and don't have rq_replay set.
2852 * Since requests are sorted in transno order, stops when meetign first
2853 * transno bigger than last_committed.
2854 * caller must hold imp->imp_lock
2856 void ptlrpc_free_committed(struct obd_import *imp)
2858 struct ptlrpc_request *req, *saved;
2859 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2860 bool skip_committed_list = true;
2863 LASSERT(imp != NULL);
2864 assert_spin_locked(&imp->imp_lock);
2866 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2867 imp->imp_generation == imp->imp_last_generation_checked) {
2868 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2869 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2872 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2873 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2874 imp->imp_generation);
2876 if (imp->imp_generation != imp->imp_last_generation_checked ||
2877 imp->imp_last_transno_checked == 0)
2878 skip_committed_list = false;
2880 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2881 imp->imp_last_generation_checked = imp->imp_generation;
2883 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2885 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2886 LASSERT(req != last_req);
2889 if (req->rq_transno == 0) {
2890 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2893 if (req->rq_import_generation < imp->imp_generation) {
2894 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2898 /* not yet committed */
2899 if (req->rq_transno > imp->imp_peer_committed_transno) {
2900 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2904 if (req->rq_replay) {
2905 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2906 list_move_tail(&req->rq_replay_list,
2907 &imp->imp_committed_list);
2911 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2912 imp->imp_peer_committed_transno);
2914 ptlrpc_free_request(req);
2917 if (skip_committed_list)
2920 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2922 LASSERT(req->rq_transno != 0);
2923 if (req->rq_import_generation < imp->imp_generation ||
2925 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2926 req->rq_import_generation <
2927 imp->imp_generation ? "stale" : "closed");
2929 if (imp->imp_replay_cursor == &req->rq_replay_list)
2930 imp->imp_replay_cursor =
2931 req->rq_replay_list.next;
2933 ptlrpc_free_request(req);
2940 void ptlrpc_cleanup_client(struct obd_import *imp)
2947 * Schedule previously sent request for resend.
2948 * For bulk requests we assign new xid (to avoid problems with
2949 * lost replies and therefore several transfers landing into same buffer
2950 * from different sending attempts).
2952 void ptlrpc_resend_req(struct ptlrpc_request *req)
2954 DEBUG_REQ(D_HA, req, "going to resend");
2955 spin_lock(&req->rq_lock);
2958 * Request got reply but linked to the import list still.
2959 * Let ptlrpc_check_set() process it.
2961 if (ptlrpc_client_replied(req)) {
2962 spin_unlock(&req->rq_lock);
2963 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2967 req->rq_status = -EAGAIN;
2970 req->rq_net_err = 0;
2971 req->rq_timedout = 0;
2973 ptlrpc_client_wake_req(req);
2974 spin_unlock(&req->rq_lock);
2977 /* XXX: this function and rq_status are currently unused */
2978 void ptlrpc_restart_req(struct ptlrpc_request *req)
2980 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2981 req->rq_status = -ERESTARTSYS;
2983 spin_lock(&req->rq_lock);
2984 req->rq_restart = 1;
2985 req->rq_timedout = 0;
2986 ptlrpc_client_wake_req(req);
2987 spin_unlock(&req->rq_lock);
2991 * Grab additional reference on a request \a req
2993 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2996 atomic_inc(&req->rq_refcount);
2999 EXPORT_SYMBOL(ptlrpc_request_addref);
3002 * Add a request to import replay_list.
3003 * Must be called under imp_lock
3005 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3006 struct obd_import *imp)
3008 struct list_head *tmp;
3010 assert_spin_locked(&imp->imp_lock);
3012 if (req->rq_transno == 0) {
3013 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3018 * clear this for new requests that were resent as well
3019 * as resent replayed requests.
3021 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3023 /* don't re-add requests that have been replayed */
3024 if (!list_empty(&req->rq_replay_list))
3027 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3029 spin_lock(&req->rq_lock);
3031 spin_unlock(&req->rq_lock);
3033 LASSERT(imp->imp_replayable);
3034 /* Balanced in ptlrpc_free_committed, usually. */
3035 ptlrpc_request_addref(req);
3036 list_for_each_prev(tmp, &imp->imp_replay_list) {
3037 struct ptlrpc_request *iter = list_entry(tmp,
3038 struct ptlrpc_request,
3042 * We may have duplicate transnos if we create and then
3043 * open a file, or for closes retained if to match creating
3044 * opens, so use req->rq_xid as a secondary key.
3045 * (See bugs 684, 685, and 428.)
3046 * XXX no longer needed, but all opens need transnos!
3048 if (iter->rq_transno > req->rq_transno)
3051 if (iter->rq_transno == req->rq_transno) {
3052 LASSERT(iter->rq_xid != req->rq_xid);
3053 if (iter->rq_xid > req->rq_xid)
3057 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3061 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3065 * Send request and wait until it completes.
3066 * Returns request processing status.
3068 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3070 struct ptlrpc_request_set *set;
3074 LASSERT(req->rq_set == NULL);
3075 LASSERT(!req->rq_receiving_reply);
3077 set = ptlrpc_prep_set();
3079 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3083 /* for distributed debugging */
3084 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3086 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3087 ptlrpc_request_addref(req);
3088 ptlrpc_set_add_req(set, req);
3089 rc = ptlrpc_set_wait(NULL, set);
3090 ptlrpc_set_destroy(set);
3094 EXPORT_SYMBOL(ptlrpc_queue_wait);
3097 * Callback used for replayed requests reply processing.
3098 * In case of successful reply calls registered request replay callback.
3099 * In case of error restart replay process.
3101 static int ptlrpc_replay_interpret(const struct lu_env *env,
3102 struct ptlrpc_request *req,
3105 struct ptlrpc_replay_async_args *aa = args;
3106 struct obd_import *imp = req->rq_import;
3109 atomic_dec(&imp->imp_replay_inflight);
3112 * Note: if it is bulk replay (MDS-MDS replay), then even if
3113 * server got the request, but bulk transfer timeout, let's
3114 * replay the bulk req again
3116 if (!ptlrpc_client_replied(req) ||
3118 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3119 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3120 GOTO(out, rc = -ETIMEDOUT);
3123 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3124 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3125 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3126 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3128 /** VBR: check version failure */
3129 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3130 /** replay was failed due to version mismatch */
3131 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3132 spin_lock(&imp->imp_lock);
3133 imp->imp_vbr_failed = 1;
3134 spin_unlock(&imp->imp_lock);
3135 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3137 /** The transno had better not change over replay. */
3138 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3139 lustre_msg_get_transno(req->rq_repmsg) ||
3140 lustre_msg_get_transno(req->rq_repmsg) == 0,
3142 lustre_msg_get_transno(req->rq_reqmsg),
3143 lustre_msg_get_transno(req->rq_repmsg));
3146 spin_lock(&imp->imp_lock);
3147 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3148 spin_unlock(&imp->imp_lock);
3149 LASSERT(imp->imp_last_replay_transno);
3151 /* transaction number shouldn't be bigger than the latest replayed */
3152 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3153 DEBUG_REQ(D_ERROR, req,
3154 "Reported transno=%llu is bigger than replayed=%llu",
3156 lustre_msg_get_transno(req->rq_reqmsg));
3157 GOTO(out, rc = -EINVAL);
3160 DEBUG_REQ(D_HA, req, "got reply");
3162 /* let the callback do fixups, possibly including in the request */
3163 if (req->rq_replay_cb)
3164 req->rq_replay_cb(req);
3166 if (ptlrpc_client_replied(req) &&
3167 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3168 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3169 lustre_msg_get_status(req->rq_repmsg),
3170 aa->praa_old_status);
3173 * Note: If the replay fails for MDT-MDT recovery, let's
3174 * abort all of the following requests in the replay
3175 * and sending list, because MDT-MDT update requests
3176 * are dependent on each other, see LU-7039
3178 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3179 struct ptlrpc_request *free_req;
3180 struct ptlrpc_request *tmp;
3182 spin_lock(&imp->imp_lock);
3183 list_for_each_entry_safe(free_req, tmp,
3184 &imp->imp_replay_list,
3186 ptlrpc_free_request(free_req);
3189 list_for_each_entry_safe(free_req, tmp,
3190 &imp->imp_committed_list,
3192 ptlrpc_free_request(free_req);
3195 list_for_each_entry_safe(free_req, tmp,
3196 &imp->imp_delayed_list,
3198 spin_lock(&free_req->rq_lock);
3199 free_req->rq_err = 1;
3200 free_req->rq_status = -EIO;
3201 ptlrpc_client_wake_req(free_req);
3202 spin_unlock(&free_req->rq_lock);
3205 list_for_each_entry_safe(free_req, tmp,
3206 &imp->imp_sending_list,
3208 spin_lock(&free_req->rq_lock);
3209 free_req->rq_err = 1;
3210 free_req->rq_status = -EIO;
3211 ptlrpc_client_wake_req(free_req);
3212 spin_unlock(&free_req->rq_lock);
3214 spin_unlock(&imp->imp_lock);
3217 /* Put it back for re-replay. */
3218 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3222 * Errors while replay can set transno to 0, but
3223 * imp_last_replay_transno shouldn't be set to 0 anyway
3225 if (req->rq_transno == 0)
3226 CERROR("Transno is 0 during replay!\n");
3228 /* continue with recovery */
3229 rc = ptlrpc_import_recovery_state_machine(imp);
3231 req->rq_send_state = aa->praa_old_state;
3234 /* this replay failed, so restart recovery */
3235 ptlrpc_connect_import(imp);
3241 * Prepares and queues request for replay.
3242 * Adds it to ptlrpcd queue for actual sending.
3243 * Returns 0 on success.
3245 int ptlrpc_replay_req(struct ptlrpc_request *req)
3247 struct ptlrpc_replay_async_args *aa;
3251 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3253 aa = ptlrpc_req_async_args(aa, req);
3254 memset(aa, 0, sizeof(*aa));
3256 /* Prepare request to be resent with ptlrpcd */
3257 aa->praa_old_state = req->rq_send_state;
3258 req->rq_send_state = LUSTRE_IMP_REPLAY;
3259 req->rq_phase = RQ_PHASE_NEW;
3260 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3262 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3264 req->rq_interpret_reply = ptlrpc_replay_interpret;
3265 /* Readjust the timeout for current conditions */
3266 ptlrpc_at_set_req_timeout(req);
3268 /* Tell server net_latency to calculate how long to wait for reply. */
3269 lustre_msg_set_service_time(req->rq_reqmsg,
3270 ptlrpc_at_get_net_latency(req));
3271 DEBUG_REQ(D_HA, req, "REPLAY");
3273 atomic_inc(&req->rq_import->imp_replay_inflight);
3274 spin_lock(&req->rq_lock);
3275 req->rq_early_free_repbuf = 0;
3276 spin_unlock(&req->rq_lock);
3277 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3279 ptlrpcd_add_req(req);
3284 * Aborts all in-flight request on import \a imp sending and delayed lists
3286 void ptlrpc_abort_inflight(struct obd_import *imp)
3288 struct list_head *tmp, *n;
3292 * Make sure that no new requests get processed for this import.
3293 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3294 * this flag and then putting requests on sending_list or delayed_list.
3296 assert_spin_locked(&imp->imp_lock);
3299 * XXX locking? Maybe we should remove each request with the list
3300 * locked? Also, how do we know if the requests on the list are
3301 * being freed at this time?
3303 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3304 struct ptlrpc_request *req = list_entry(tmp,
3305 struct ptlrpc_request,
3308 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3310 spin_lock(&req->rq_lock);
3311 if (req->rq_import_generation < imp->imp_generation) {
3313 req->rq_status = -EIO;
3314 ptlrpc_client_wake_req(req);
3316 spin_unlock(&req->rq_lock);
3319 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3320 struct ptlrpc_request *req =
3321 list_entry(tmp, struct ptlrpc_request, rq_list);
3323 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3325 spin_lock(&req->rq_lock);
3326 if (req->rq_import_generation < imp->imp_generation) {
3328 req->rq_status = -EIO;
3329 ptlrpc_client_wake_req(req);
3331 spin_unlock(&req->rq_lock);
3335 * Last chance to free reqs left on the replay list, but we
3336 * will still leak reqs that haven't committed.
3338 if (imp->imp_replayable)
3339 ptlrpc_free_committed(imp);
3345 * Abort all uncompleted requests in request set \a set
3347 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3349 struct list_head *tmp, *pos;
3351 LASSERT(set != NULL);
3353 list_for_each_safe(pos, tmp, &set->set_requests) {
3354 struct ptlrpc_request *req =
3355 list_entry(pos, struct ptlrpc_request,
3358 spin_lock(&req->rq_lock);
3359 if (req->rq_phase != RQ_PHASE_RPC) {
3360 spin_unlock(&req->rq_lock);
3365 req->rq_status = -EINTR;
3366 ptlrpc_client_wake_req(req);
3367 spin_unlock(&req->rq_lock);
3372 * Initialize the XID for the node. This is common among all requests on
3373 * this node, and only requires the property that it is monotonically
3374 * increasing. It does not need to be sequential. Since this is also used
3375 * as the RDMA match bits, it is important that a single client NOT have
3376 * the same match bits for two different in-flight requests, hence we do
3377 * NOT want to have an XID per target or similar.
3379 * To avoid an unlikely collision between match bits after a client reboot
3380 * (which would deliver old data into the wrong RDMA buffer) initialize
3381 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3382 * If the time is clearly incorrect, we instead use a 62-bit random number.
3383 * In the worst case the random number will overflow 1M RPCs per second in
3384 * 9133 years, or permutations thereof.
3386 #define YEAR_2004 (1ULL << 30)
3387 void ptlrpc_init_xid(void)
3389 time64_t now = ktime_get_real_seconds();
3392 if (now < YEAR_2004) {
3393 get_random_bytes(&xid, sizeof(xid));
3395 xid |= (1ULL << 61);
3397 xid = (u64)now << 20;
3400 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3401 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3403 xid &= PTLRPC_BULK_OPS_MASK;
3404 atomic64_set(&ptlrpc_last_xid, xid);
3408 * Increase xid and returns resulting new value to the caller.
3410 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3411 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3412 * itself uses the last bulk xid needed, so the server can determine the
3413 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3414 * xid must align to a power-of-two value.
3416 * This is assumed to be true due to the initial ptlrpc_last_xid
3417 * value also being initialized to a power-of-two value. LU-1431
3419 __u64 ptlrpc_next_xid(void)
3421 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3425 * If request has a new allocated XID (new request or EINPROGRESS resend),
3426 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3427 * request to ensure previous bulk fails and avoid problems with lost replies
3428 * and therefore several transfers landing into the same buffer from different
3431 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3433 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3435 LASSERT(bd != NULL);
3438 * Generate new matchbits for all resend requests, including
3441 if (req->rq_resend) {
3442 __u64 old_mbits = req->rq_mbits;
3445 * First time resend on -EINPROGRESS will generate new xid,
3446 * so we can actually use the rq_xid as rq_mbits in such case,
3447 * however, it's bit hard to distinguish such resend with a
3448 * 'resend for the -EINPROGRESS resend'. To make it simple,
3449 * we opt to generate mbits for all resend cases.
3451 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3453 req->rq_mbits = ptlrpc_next_xid();
3456 * Old version transfers rq_xid to peer as
3459 spin_lock(&req->rq_import->imp_lock);
3460 list_del_init(&req->rq_unreplied_list);
3461 ptlrpc_assign_next_xid_nolock(req);
3462 spin_unlock(&req->rq_import->imp_lock);
3463 req->rq_mbits = req->rq_xid;
3465 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3466 old_mbits, req->rq_mbits);
3467 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3468 /* Request being sent first time, use xid as matchbits. */
3469 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3470 || req->rq_mbits == 0) {
3471 req->rq_mbits = req->rq_xid;
3473 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3475 req->rq_mbits -= total_md - 1;
3479 * Replay request, xid and matchbits have already been
3480 * correctly assigned.
3486 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3487 * that server can infer the number of bulks that were prepared,
3490 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3494 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3495 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3497 * It's ok to directly set the rq_xid here, since this xid bump
3498 * won't affect the request position in unreplied list.
3500 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3501 req->rq_xid = req->rq_mbits;
3505 * Get a glimpse at what next xid value might have been.
3506 * Returns possible next xid.
3508 __u64 ptlrpc_sample_next_xid(void)
3510 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3512 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3515 * Functions for operating ptlrpc workers.
3517 * A ptlrpc work is a function which will be running inside ptlrpc context.
3518 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3520 * 1. after a work is created, it can be used many times, that is:
3521 * handler = ptlrpcd_alloc_work();
3522 * ptlrpcd_queue_work();
3524 * queue it again when necessary:
3525 * ptlrpcd_queue_work();
3526 * ptlrpcd_destroy_work();
3527 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3528 * it will only be queued once in any time. Also as its name implies, it may
3529 * have delay before it really runs by ptlrpcd thread.
3531 struct ptlrpc_work_async_args {
3532 int (*cb)(const struct lu_env *, void *);
3536 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3538 /* re-initialize the req */
3539 req->rq_timeout = obd_timeout;
3540 req->rq_sent = ktime_get_real_seconds();
3541 req->rq_deadline = req->rq_sent + req->rq_timeout;
3542 req->rq_phase = RQ_PHASE_INTERPRET;
3543 req->rq_next_phase = RQ_PHASE_COMPLETE;
3544 req->rq_xid = ptlrpc_next_xid();
3545 req->rq_import_generation = req->rq_import->imp_generation;
3547 ptlrpcd_add_req(req);
3550 static int work_interpreter(const struct lu_env *env,
3551 struct ptlrpc_request *req, void *args, int rc)
3553 struct ptlrpc_work_async_args *arg = args;
3555 LASSERT(ptlrpcd_check_work(req));
3556 LASSERT(arg->cb != NULL);
3558 rc = arg->cb(env, arg->cbdata);
3560 list_del_init(&req->rq_set_chain);
3563 if (atomic_dec_return(&req->rq_refcount) > 1) {
3564 atomic_set(&req->rq_refcount, 2);
3565 ptlrpcd_add_work_req(req);
3570 static int worker_format;
3572 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3574 return req->rq_pill.rc_fmt == (void *)&worker_format;
3578 * Create a work for ptlrpc.
3580 void *ptlrpcd_alloc_work(struct obd_import *imp,
3581 int (*cb)(const struct lu_env *, void *), void *cbdata)
3583 struct ptlrpc_request *req = NULL;
3584 struct ptlrpc_work_async_args *args;
3590 RETURN(ERR_PTR(-EINVAL));
3592 /* copy some code from deprecated fakereq. */
3593 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3595 CERROR("ptlrpc: run out of memory!\n");
3596 RETURN(ERR_PTR(-ENOMEM));
3599 ptlrpc_cli_req_init(req);
3601 req->rq_send_state = LUSTRE_IMP_FULL;
3602 req->rq_type = PTL_RPC_MSG_REQUEST;
3603 req->rq_import = class_import_get(imp);
3604 req->rq_interpret_reply = work_interpreter;
3605 /* don't want reply */
3606 req->rq_no_delay = req->rq_no_resend = 1;
3607 req->rq_pill.rc_fmt = (void *)&worker_format;
3609 args = ptlrpc_req_async_args(args, req);
3611 args->cbdata = cbdata;
3615 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3617 void ptlrpcd_destroy_work(void *handler)
3619 struct ptlrpc_request *req = handler;
3622 ptlrpc_req_finished(req);
3624 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3626 int ptlrpcd_queue_work(void *handler)
3628 struct ptlrpc_request *req = handler;
3631 * Check if the req is already being queued.
3633 * Here comes a trick: it lacks a way of checking if a req is being
3634 * processed reliably in ptlrpc. Here I have to use refcount of req
3635 * for this purpose. This is okay because the caller should use this
3636 * req as opaque data. - Jinshan
3638 LASSERT(atomic_read(&req->rq_refcount) > 0);
3639 if (atomic_inc_return(&req->rq_refcount) == 2)
3640 ptlrpcd_add_work_req(req);
3643 EXPORT_SYMBOL(ptlrpcd_queue_work);