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 <obd_support.h>
41 #include <obd_class.h>
42 #include <lustre_lib.h>
43 #include <lustre_ha.h>
44 #include <lustre_import.h>
45 #include <lustre_req_layout.h>
47 #include "ptlrpc_internal.h"
49 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
50 struct page *page, int pageoffset,
53 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
56 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
57 struct page *page, int pageoffset,
60 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
63 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
67 for (i = 0; i < desc->bd_iov_count ; i++)
68 put_page(BD_GET_KIOV(desc, i).kiov_page);
71 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
72 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
73 .release_frags = ptlrpc_release_bulk_page_pin,
75 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
77 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
78 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
79 .release_frags = ptlrpc_release_bulk_noop,
81 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
83 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
84 .add_iov_frag = ptlrpc_prep_bulk_frag,
86 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
88 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
89 static int ptlrpcd_check_work(struct ptlrpc_request *req);
90 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
93 * Initialize passed in client structure \a cl.
95 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
96 struct ptlrpc_client *cl)
98 cl->cli_request_portal = req_portal;
99 cl->cli_reply_portal = rep_portal;
102 EXPORT_SYMBOL(ptlrpc_init_client);
105 * Return PortalRPC connection for remore uud \a uuid
107 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
108 lnet_nid_t nid4refnet)
110 struct ptlrpc_connection *c;
112 struct lnet_process_id peer;
116 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
117 * before accessing its values.
119 /* coverity[uninit_use_in_call] */
120 peer.nid = nid4refnet;
121 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
123 CNETERR("cannot find peer %s!\n", uuid->uuid);
127 c = ptlrpc_connection_get(peer, self, uuid);
129 memcpy(c->c_remote_uuid.uuid,
130 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
133 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
139 * Allocate and initialize new bulk descriptor on the sender.
140 * Returns pointer to the descriptor or NULL on error.
142 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
143 unsigned int max_brw,
144 enum ptlrpc_bulk_op_type type,
146 const struct ptlrpc_bulk_frag_ops *ops)
148 struct ptlrpc_bulk_desc *desc;
151 /* ensure that only one of KIOV or IOVEC is set but not both */
152 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
153 ops->add_kiov_frag != NULL) ||
154 (ptlrpc_is_bulk_desc_kvec(type) &&
155 ops->add_iov_frag != NULL));
160 if (type & PTLRPC_BULK_BUF_KIOV) {
161 OBD_ALLOC_LARGE(GET_KIOV(desc),
162 nfrags * sizeof(*GET_KIOV(desc)));
166 OBD_ALLOC_LARGE(GET_KVEC(desc),
167 nfrags * sizeof(*GET_KVEC(desc)));
172 spin_lock_init(&desc->bd_lock);
173 init_waitqueue_head(&desc->bd_waitq);
174 desc->bd_max_iov = nfrags;
175 desc->bd_iov_count = 0;
176 desc->bd_portal = portal;
177 desc->bd_type = type;
178 desc->bd_md_count = 0;
179 desc->bd_frag_ops = ops;
180 LASSERT(max_brw > 0);
181 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
183 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
184 * node. Negotiated ocd_brw_size will always be <= this number.
186 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
187 LNetInvalidateMDHandle(&desc->bd_mds[i]);
196 * Prepare bulk descriptor for specified outgoing request \a req that
197 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
198 * the bulk to be sent. Used on client-side.
199 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
202 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
204 unsigned int max_brw,
207 const struct ptlrpc_bulk_frag_ops
210 struct obd_import *imp = req->rq_import;
211 struct ptlrpc_bulk_desc *desc;
214 LASSERT(ptlrpc_is_bulk_op_passive(type));
216 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
220 desc->bd_import_generation = req->rq_import_generation;
221 desc->bd_import = class_import_get(imp);
224 desc->bd_cbid.cbid_fn = client_bulk_callback;
225 desc->bd_cbid.cbid_arg = desc;
227 /* This makes req own desc, and free it when she frees herself */
232 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
234 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
235 struct page *page, int pageoffset, int len,
240 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
241 LASSERT(page != NULL);
242 LASSERT(pageoffset >= 0);
244 LASSERT(pageoffset + len <= PAGE_SIZE);
245 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
247 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
254 kiov->kiov_page = page;
255 kiov->kiov_offset = pageoffset;
256 kiov->kiov_len = len;
258 desc->bd_iov_count++;
260 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
262 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
269 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
270 LASSERT(frag != NULL);
272 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
274 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
278 iovec->iov_base = frag;
279 iovec->iov_len = len;
281 desc->bd_iov_count++;
283 RETURN(desc->bd_nob);
285 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
287 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
291 LASSERT(desc != NULL);
292 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
293 LASSERT(desc->bd_md_count == 0); /* network hands off */
294 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
295 LASSERT(desc->bd_frag_ops != NULL);
297 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
298 sptlrpc_enc_pool_put_pages(desc);
301 class_export_put(desc->bd_export);
303 class_import_put(desc->bd_import);
305 if (desc->bd_frag_ops->release_frags != NULL)
306 desc->bd_frag_ops->release_frags(desc);
308 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
309 OBD_FREE_LARGE(GET_KIOV(desc),
310 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
312 OBD_FREE_LARGE(GET_KVEC(desc),
313 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
317 EXPORT_SYMBOL(ptlrpc_free_bulk);
320 * Set server timelimit for this req, i.e. how long are we willing to wait
321 * for reply before timing out this request.
323 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
329 LASSERT(req->rq_import);
332 /* non-AT settings */
334 * \a imp_server_timeout means this is reverse import and
335 * we send (currently only) ASTs to the client and cannot afford
336 * to wait too long for the reply, otherwise the other client
337 * (because of which we are sending this request) would
338 * timeout waiting for us
340 req->rq_timeout = req->rq_import->imp_server_timeout ?
341 obd_timeout / 2 : obd_timeout;
343 at = &req->rq_import->imp_at;
344 idx = import_at_get_index(req->rq_import,
345 req->rq_request_portal);
346 serv_est = at_get(&at->iat_service_estimate[idx]);
347 req->rq_timeout = at_est2timeout(serv_est);
350 * We could get even fancier here, using history to predict increased
355 * Let the server know what this RPC timeout is by putting it in the
358 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
360 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
362 /* Adjust max service estimate based on server value */
363 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
364 unsigned int serv_est)
370 LASSERT(req->rq_import);
371 at = &req->rq_import->imp_at;
373 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
375 * max service estimates are tracked on the server side,
376 * so just keep minimal history here
378 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
381 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
382 req->rq_import->imp_obd->obd_name,
383 req->rq_request_portal,
384 oldse, at_get(&at->iat_service_estimate[idx]));
387 /* Expected network latency per remote node (secs) */
388 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
390 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
393 /* Adjust expected network latency */
394 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
395 unsigned int service_time)
397 unsigned int nl, oldnl;
399 time64_t now = ktime_get_real_seconds();
401 LASSERT(req->rq_import);
403 if (service_time > now - req->rq_sent + 3) {
405 * b=16408, however, this can also happen if early reply
406 * is lost and client RPC is expired and resent, early reply
407 * or reply of original RPC can still be fit in reply buffer
408 * of resent RPC, now client is measuring time from the
409 * resent time, but server sent back service time of original
412 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
413 D_ADAPTTO : D_WARNING,
414 "Reported service time %u > total measured time %lld\n",
415 service_time, now - req->rq_sent);
419 /* Network latency is total time less server processing time */
420 nl = max_t(int, now - req->rq_sent -
421 service_time, 0) + 1; /* st rounding */
422 at = &req->rq_import->imp_at;
424 oldnl = at_measured(&at->iat_net_latency, nl);
427 "The network latency for %s (nid %s) has changed from %d to %d\n",
428 req->rq_import->imp_obd->obd_name,
429 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
430 oldnl, at_get(&at->iat_net_latency));
433 static int unpack_reply(struct ptlrpc_request *req)
437 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
438 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
440 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
445 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
447 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
454 * Handle an early reply message, called with the rq_lock held.
455 * If anything goes wrong just ignore it - same as if it never happened
457 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
458 __must_hold(&req->rq_lock)
460 struct ptlrpc_request *early_req;
466 spin_unlock(&req->rq_lock);
468 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
470 spin_lock(&req->rq_lock);
474 rc = unpack_reply(early_req);
476 sptlrpc_cli_finish_early_reply(early_req);
477 spin_lock(&req->rq_lock);
482 * Use new timeout value just to adjust the local value for this
483 * request, don't include it into at_history. It is unclear yet why
484 * service time increased and should it be counted or skipped, e.g.
485 * that can be recovery case or some error or server, the real reply
486 * will add all new data if it is worth to add.
488 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
489 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
491 /* Network latency can be adjusted, it is pure network delays */
492 ptlrpc_at_adj_net_latency(req,
493 lustre_msg_get_service_time(early_req->rq_repmsg));
495 sptlrpc_cli_finish_early_reply(early_req);
497 spin_lock(&req->rq_lock);
498 olddl = req->rq_deadline;
500 * server assumes it now has rq_timeout from when the request
501 * arrived, so the client should give it at least that long.
502 * since we don't know the arrival time we'll use the original
505 req->rq_deadline = req->rq_sent + req->rq_timeout +
506 ptlrpc_at_get_net_latency(req);
508 DEBUG_REQ(D_ADAPTTO, req,
509 "Early reply #%d, new deadline in %llds (%llds)",
511 req->rq_deadline - ktime_get_real_seconds(),
512 req->rq_deadline - olddl);
517 static struct kmem_cache *request_cache;
519 int ptlrpc_request_cache_init(void)
521 request_cache = kmem_cache_create("ptlrpc_cache",
522 sizeof(struct ptlrpc_request),
523 0, SLAB_HWCACHE_ALIGN, NULL);
524 return request_cache ? 0 : -ENOMEM;
527 void ptlrpc_request_cache_fini(void)
529 kmem_cache_destroy(request_cache);
532 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
534 struct ptlrpc_request *req;
536 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
540 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
542 OBD_SLAB_FREE_PTR(req, request_cache);
546 * Wind down request pool \a pool.
547 * Frees all requests from the pool too
549 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
551 struct list_head *l, *tmp;
552 struct ptlrpc_request *req;
554 LASSERT(pool != NULL);
556 spin_lock(&pool->prp_lock);
557 list_for_each_safe(l, tmp, &pool->prp_req_list) {
558 req = list_entry(l, struct ptlrpc_request, rq_list);
559 list_del(&req->rq_list);
560 LASSERT(req->rq_reqbuf);
561 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
562 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
563 ptlrpc_request_cache_free(req);
565 spin_unlock(&pool->prp_lock);
566 OBD_FREE(pool, sizeof(*pool));
568 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
571 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
573 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
578 while (size < pool->prp_rq_size)
581 LASSERTF(list_empty(&pool->prp_req_list) ||
582 size == pool->prp_rq_size,
583 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
584 pool->prp_rq_size, size);
586 spin_lock(&pool->prp_lock);
587 pool->prp_rq_size = size;
588 for (i = 0; i < num_rq; i++) {
589 struct ptlrpc_request *req;
590 struct lustre_msg *msg;
592 spin_unlock(&pool->prp_lock);
593 req = ptlrpc_request_cache_alloc(GFP_NOFS);
596 OBD_ALLOC_LARGE(msg, size);
598 ptlrpc_request_cache_free(req);
601 req->rq_reqbuf = msg;
602 req->rq_reqbuf_len = size;
604 spin_lock(&pool->prp_lock);
605 list_add_tail(&req->rq_list, &pool->prp_req_list);
607 spin_unlock(&pool->prp_lock);
610 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
613 * Create and initialize new request pool with given attributes:
614 * \a num_rq - initial number of requests to create for the pool
615 * \a msgsize - maximum message size possible for requests in thid pool
616 * \a populate_pool - function to be called when more requests need to be added
618 * Returns pointer to newly created pool or NULL on error.
620 struct ptlrpc_request_pool *
621 ptlrpc_init_rq_pool(int num_rq, int msgsize,
622 int (*populate_pool)(struct ptlrpc_request_pool *, int))
624 struct ptlrpc_request_pool *pool;
626 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
631 * Request next power of two for the allocation, because internally
632 * kernel would do exactly this
634 spin_lock_init(&pool->prp_lock);
635 INIT_LIST_HEAD(&pool->prp_req_list);
636 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
637 pool->prp_populate = populate_pool;
639 populate_pool(pool, num_rq);
643 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
646 * Fetches one request from pool \a pool
648 static struct ptlrpc_request *
649 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
651 struct ptlrpc_request *request;
652 struct lustre_msg *reqbuf;
657 spin_lock(&pool->prp_lock);
660 * See if we have anything in a pool, and bail out if nothing,
661 * in writeout path, where this matters, this is safe to do, because
662 * nothing is lost in this case, and when some in-flight requests
663 * complete, this code will be called again.
665 if (unlikely(list_empty(&pool->prp_req_list))) {
666 spin_unlock(&pool->prp_lock);
670 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
672 list_del_init(&request->rq_list);
673 spin_unlock(&pool->prp_lock);
675 LASSERT(request->rq_reqbuf);
676 LASSERT(request->rq_pool);
678 reqbuf = request->rq_reqbuf;
679 memset(request, 0, sizeof(*request));
680 request->rq_reqbuf = reqbuf;
681 request->rq_reqbuf_len = pool->prp_rq_size;
682 request->rq_pool = pool;
688 * Returns freed \a request to pool.
690 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
692 struct ptlrpc_request_pool *pool = request->rq_pool;
694 spin_lock(&pool->prp_lock);
695 LASSERT(list_empty(&request->rq_list));
696 LASSERT(!request->rq_receiving_reply);
697 list_add_tail(&request->rq_list, &pool->prp_req_list);
698 spin_unlock(&pool->prp_lock);
701 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
703 struct obd_import *imp = req->rq_import;
704 struct list_head *tmp;
705 struct ptlrpc_request *iter;
707 assert_spin_locked(&imp->imp_lock);
708 LASSERT(list_empty(&req->rq_unreplied_list));
710 /* unreplied list is sorted by xid in ascending order */
711 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
712 iter = list_entry(tmp, struct ptlrpc_request,
715 LASSERT(req->rq_xid != iter->rq_xid);
716 if (req->rq_xid < iter->rq_xid)
718 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
721 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
724 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
726 req->rq_xid = ptlrpc_next_xid();
727 ptlrpc_add_unreplied(req);
730 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
732 spin_lock(&req->rq_import->imp_lock);
733 ptlrpc_assign_next_xid_nolock(req);
734 spin_unlock(&req->rq_import->imp_lock);
737 static __u64 ptlrpc_last_xid;
738 static spinlock_t ptlrpc_last_xid_lock;
740 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
741 __u32 version, int opcode, char **bufs,
742 struct ptlrpc_cli_ctx *ctx)
745 struct obd_import *imp;
751 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
752 imp = request->rq_import;
753 lengths = request->rq_pill.rc_area[RCL_CLIENT];
756 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
758 rc = sptlrpc_req_get_ctx(request);
762 sptlrpc_req_set_flavor(request, opcode);
764 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
769 lustre_msg_add_version(request->rq_reqmsg, version);
770 request->rq_send_state = LUSTRE_IMP_FULL;
771 request->rq_type = PTL_RPC_MSG_REQUEST;
773 request->rq_req_cbid.cbid_fn = request_out_callback;
774 request->rq_req_cbid.cbid_arg = request;
776 request->rq_reply_cbid.cbid_fn = reply_in_callback;
777 request->rq_reply_cbid.cbid_arg = request;
779 request->rq_reply_deadline = 0;
780 request->rq_bulk_deadline = 0;
781 request->rq_req_deadline = 0;
782 request->rq_phase = RQ_PHASE_NEW;
783 request->rq_next_phase = RQ_PHASE_UNDEFINED;
785 request->rq_request_portal = imp->imp_client->cli_request_portal;
786 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
788 ptlrpc_at_set_req_timeout(request);
790 lustre_msg_set_opc(request->rq_reqmsg, opcode);
792 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
793 if (cfs_fail_val == opcode) {
794 time64_t *fail_t = NULL, *fail2_t = NULL;
796 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
797 fail_t = &request->rq_bulk_deadline;
798 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
799 fail_t = &request->rq_reply_deadline;
800 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
801 fail_t = &request->rq_req_deadline;
802 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
803 fail_t = &request->rq_reply_deadline;
804 fail2_t = &request->rq_bulk_deadline;
805 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
806 time64_t now = ktime_get_real_seconds();
807 spin_lock(&ptlrpc_last_xid_lock);
808 ptlrpc_last_xid = ((__u64)now >> 4) << 24;
809 spin_unlock(&ptlrpc_last_xid_lock);
813 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
816 *fail2_t = ktime_get_real_seconds() +
820 * The RPC is infected, let the test to change the
823 msleep(4 * MSEC_PER_SEC);
826 ptlrpc_assign_next_xid(request);
831 LASSERT(!request->rq_pool);
832 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
834 class_import_put(imp);
838 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
841 * Pack request buffers for network transfer, performing necessary encryption
842 * steps if necessary.
844 int ptlrpc_request_pack(struct ptlrpc_request *request,
845 __u32 version, int opcode)
849 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
854 * For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
855 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
856 * have to send old ptlrpc_body to keep interoprability with these
859 * Only three kinds of server->client RPCs so far:
864 * XXX This should be removed whenever we drop the interoprability with
865 * the these old clients.
867 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
868 opcode == LDLM_GL_CALLBACK)
869 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
870 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
874 EXPORT_SYMBOL(ptlrpc_request_pack);
877 * Helper function to allocate new request on import \a imp
878 * and possibly using existing request from pool \a pool if provided.
879 * Returns allocated request structure with import field filled or
883 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
884 struct ptlrpc_request_pool *pool)
886 struct ptlrpc_request *request = NULL;
888 request = ptlrpc_request_cache_alloc(GFP_NOFS);
890 if (!request && pool)
891 request = ptlrpc_prep_req_from_pool(pool);
894 ptlrpc_cli_req_init(request);
896 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
897 LASSERT(imp != LP_POISON);
898 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
900 LASSERT(imp->imp_client != LP_POISON);
902 request->rq_import = class_import_get(imp);
904 CERROR("request allocation out of memory\n");
911 * Helper function for creating a request.
912 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
913 * buffer structures according to capsule template \a format.
914 * Returns allocated request structure pointer or NULL on error.
916 static struct ptlrpc_request *
917 ptlrpc_request_alloc_internal(struct obd_import *imp,
918 struct ptlrpc_request_pool *pool,
919 const struct req_format *format)
921 struct ptlrpc_request *request;
924 request = __ptlrpc_request_alloc(imp, pool);
929 * initiate connection if needed when the import has been
930 * referenced by the new request to avoid races with disconnect
932 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
935 CDEBUG_LIMIT(imp->imp_idle_debug,
936 "%s: reconnect after %llds idle\n",
937 imp->imp_obd->obd_name, ktime_get_real_seconds() -
938 imp->imp_last_reply_time);
939 spin_lock(&imp->imp_lock);
940 if (imp->imp_state == LUSTRE_IMP_IDLE) {
941 imp->imp_generation++;
942 imp->imp_initiated_at = imp->imp_generation;
943 imp->imp_state = LUSTRE_IMP_NEW;
946 spin_unlock(&imp->imp_lock);
948 rc = ptlrpc_connect_import(imp);
950 ptlrpc_request_free(request);
953 ptlrpc_pinger_add_import(imp);
957 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
958 req_capsule_set(&request->rq_pill, format);
963 * Allocate new request structure for import \a imp and initialize its
964 * buffer structure according to capsule template \a format.
966 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
967 const struct req_format *format)
969 return ptlrpc_request_alloc_internal(imp, NULL, format);
971 EXPORT_SYMBOL(ptlrpc_request_alloc);
974 * Allocate new request structure for import \a imp from pool \a pool and
975 * initialize its buffer structure according to capsule template \a format.
977 struct ptlrpc_request *
978 ptlrpc_request_alloc_pool(struct obd_import *imp,
979 struct ptlrpc_request_pool *pool,
980 const struct req_format *format)
982 return ptlrpc_request_alloc_internal(imp, pool, format);
984 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
987 * For requests not from pool, free memory of the request structure.
988 * For requests obtained from a pool earlier, return request back to pool.
990 void ptlrpc_request_free(struct ptlrpc_request *request)
992 if (request->rq_pool)
993 __ptlrpc_free_req_to_pool(request);
995 ptlrpc_request_cache_free(request);
997 EXPORT_SYMBOL(ptlrpc_request_free);
1000 * Allocate new request for operatione \a opcode and immediatelly pack it for
1002 * Only used for simple requests like OBD_PING where the only important
1003 * part of the request is operation itself.
1004 * Returns allocated request or NULL on error.
1006 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1007 const struct req_format *format,
1008 __u32 version, int opcode)
1010 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1014 rc = ptlrpc_request_pack(req, version, opcode);
1016 ptlrpc_request_free(req);
1022 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1025 * Allocate and initialize new request set structure on the current CPT.
1026 * Returns a pointer to the newly allocated set structure or NULL on error.
1028 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1030 struct ptlrpc_request_set *set;
1034 cpt = cfs_cpt_current(cfs_cpt_table, 0);
1035 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1038 atomic_set(&set->set_refcount, 1);
1039 INIT_LIST_HEAD(&set->set_requests);
1040 init_waitqueue_head(&set->set_waitq);
1041 atomic_set(&set->set_new_count, 0);
1042 atomic_set(&set->set_remaining, 0);
1043 spin_lock_init(&set->set_new_req_lock);
1044 INIT_LIST_HEAD(&set->set_new_requests);
1045 set->set_max_inflight = UINT_MAX;
1046 set->set_producer = NULL;
1047 set->set_producer_arg = NULL;
1052 EXPORT_SYMBOL(ptlrpc_prep_set);
1055 * Allocate and initialize new request set structure with flow control
1056 * extension. This extension allows to control the number of requests in-flight
1057 * for the whole set. A callback function to generate requests must be provided
1058 * and the request set will keep the number of requests sent over the wire to
1060 * Returns a pointer to the newly allocated set structure or NULL on error.
1062 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1066 struct ptlrpc_request_set *set;
1068 set = ptlrpc_prep_set();
1072 set->set_max_inflight = max;
1073 set->set_producer = func;
1074 set->set_producer_arg = arg;
1080 * Wind down and free request set structure previously allocated with
1082 * Ensures that all requests on the set have completed and removes
1083 * all requests from the request list in a set.
1084 * If any unsent request happen to be on the list, pretends that they got
1085 * an error in flight and calls their completion handler.
1087 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1089 struct list_head *tmp;
1090 struct list_head *next;
1096 /* Requests on the set should either all be completed, or all be new */
1097 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1098 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1099 list_for_each(tmp, &set->set_requests) {
1100 struct ptlrpc_request *req =
1101 list_entry(tmp, struct ptlrpc_request,
1104 LASSERT(req->rq_phase == expected_phase);
1108 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1109 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1110 atomic_read(&set->set_remaining), n);
1112 list_for_each_safe(tmp, next, &set->set_requests) {
1113 struct ptlrpc_request *req =
1114 list_entry(tmp, struct ptlrpc_request,
1116 list_del_init(&req->rq_set_chain);
1118 LASSERT(req->rq_phase == expected_phase);
1120 if (req->rq_phase == RQ_PHASE_NEW) {
1121 ptlrpc_req_interpret(NULL, req, -EBADR);
1122 atomic_dec(&set->set_remaining);
1125 spin_lock(&req->rq_lock);
1127 req->rq_invalid_rqset = 0;
1128 spin_unlock(&req->rq_lock);
1130 ptlrpc_req_finished(req);
1133 LASSERT(atomic_read(&set->set_remaining) == 0);
1135 ptlrpc_reqset_put(set);
1138 EXPORT_SYMBOL(ptlrpc_set_destroy);
1141 * Add a new request to the general purpose request set.
1142 * Assumes request reference from the caller.
1144 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1145 struct ptlrpc_request *req)
1147 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1148 LASSERT(list_empty(&req->rq_set_chain));
1150 if (req->rq_allow_intr)
1151 set->set_allow_intr = 1;
1153 /* The set takes over the caller's request reference */
1154 list_add_tail(&req->rq_set_chain, &set->set_requests);
1156 atomic_inc(&set->set_remaining);
1157 req->rq_queued_time = ktime_get_seconds();
1160 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1162 if (set->set_producer)
1164 * If the request set has a producer callback, the RPC must be
1165 * sent straight away
1167 ptlrpc_send_new_req(req);
1169 EXPORT_SYMBOL(ptlrpc_set_add_req);
1172 * Add a request to a request with dedicated server thread
1173 * and wake the thread to make any necessary processing.
1174 * Currently only used for ptlrpcd.
1176 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1177 struct ptlrpc_request *req)
1179 struct ptlrpc_request_set *set = pc->pc_set;
1182 LASSERT(req->rq_set == NULL);
1183 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1185 spin_lock(&set->set_new_req_lock);
1187 * The set takes over the caller's request reference.
1190 req->rq_queued_time = ktime_get_seconds();
1191 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1192 count = atomic_inc_return(&set->set_new_count);
1193 spin_unlock(&set->set_new_req_lock);
1195 /* Only need to call wakeup once for the first entry. */
1197 wake_up(&set->set_waitq);
1200 * XXX: It maybe unnecessary to wakeup all the partners. But to
1201 * guarantee the async RPC can be processed ASAP, we have
1202 * no other better choice. It maybe fixed in future.
1204 for (i = 0; i < pc->pc_npartners; i++)
1205 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1210 * Based on the current state of the import, determine if the request
1211 * can be sent, is an error, or should be delayed.
1213 * Returns true if this request should be delayed. If false, and
1214 * *status is set, then the request can not be sent and *status is the
1215 * error code. If false and status is 0, then request can be sent.
1217 * The imp->imp_lock must be held.
1219 static int ptlrpc_import_delay_req(struct obd_import *imp,
1220 struct ptlrpc_request *req, int *status)
1228 if (req->rq_ctx_init || req->rq_ctx_fini) {
1229 /* always allow ctx init/fini rpc go through */
1230 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1231 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1233 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1234 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1237 * pings or MDS-equivalent STATFS may safely
1240 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1241 D_HA : D_ERROR, req, "IMP_CLOSED ");
1243 } else if (ptlrpc_send_limit_expired(req)) {
1244 /* probably doesn't need to be a D_ERROR afterinitial testing */
1245 DEBUG_REQ(D_HA, req, "send limit expired ");
1246 *status = -ETIMEDOUT;
1247 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1248 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1249 ;/* allow CONNECT even if import is invalid */
1250 if (atomic_read(&imp->imp_inval_count) != 0) {
1251 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1254 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1255 if (!imp->imp_deactive)
1256 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1257 *status = -ESHUTDOWN; /* b=12940 */
1258 } else if (req->rq_import_generation != imp->imp_generation) {
1259 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1261 } else if (req->rq_send_state != imp->imp_state) {
1262 /* invalidate in progress - any requests should be drop */
1263 if (atomic_read(&imp->imp_inval_count) != 0) {
1264 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1266 } else if (req->rq_no_delay &&
1267 imp->imp_generation != imp->imp_initiated_at) {
1268 /* ignore nodelay for requests initiating connections */
1269 *status = -EWOULDBLOCK;
1270 } else if (req->rq_allow_replay &&
1271 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1272 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1273 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1274 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1275 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1285 * Decide if the error message should be printed to the console or not.
1286 * Makes its decision based on request type, status, and failure frequency.
1288 * \param[in] req request that failed and may need a console message
1290 * \retval false if no message should be printed
1291 * \retval true if console message should be printed
1293 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1295 LASSERT(req->rq_reqmsg != NULL);
1297 /* Suppress particular reconnect errors which are to be expected. */
1298 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1299 /* Suppress timed out reconnect requests */
1300 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1305 * Suppress most unavailable/again reconnect requests, but
1306 * print occasionally so it is clear client is trying to
1307 * connect to a server where no target is running.
1309 if ((err == -ENODEV || err == -EAGAIN) &&
1310 req->rq_import->imp_conn_cnt % 30 != 20)
1314 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1315 /* -EAGAIN is normal when using POSIX flocks */
1318 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1319 (req->rq_xid & 0xf) != 10)
1320 /* Suppress most ping requests, they may fail occasionally */
1327 * Check request processing status.
1328 * Returns the status.
1330 static int ptlrpc_check_status(struct ptlrpc_request *req)
1335 err = lustre_msg_get_status(req->rq_repmsg);
1336 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1337 struct obd_import *imp = req->rq_import;
1338 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1339 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1341 if (ptlrpc_console_allow(req, opc, err))
1342 LCONSOLE_ERROR_MSG(0x11,
1343 "%s: operation %s to node %s failed: rc = %d\n",
1344 imp->imp_obd->obd_name,
1346 libcfs_nid2str(nid), err);
1347 RETURN(err < 0 ? err : -EINVAL);
1351 DEBUG_REQ(D_INFO, req, "status is %d", err);
1352 } else if (err > 0) {
1353 /* XXX: translate this error from net to host */
1354 DEBUG_REQ(D_INFO, req, "status is %d", err);
1361 * save pre-versions of objects into request for replay.
1362 * Versions are obtained from server reply.
1365 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1367 struct lustre_msg *repmsg = req->rq_repmsg;
1368 struct lustre_msg *reqmsg = req->rq_reqmsg;
1369 __u64 *versions = lustre_msg_get_versions(repmsg);
1372 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1376 lustre_msg_set_versions(reqmsg, versions);
1377 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1378 versions[0], versions[1]);
1383 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1385 struct ptlrpc_request *req;
1387 assert_spin_locked(&imp->imp_lock);
1388 if (list_empty(&imp->imp_unreplied_list))
1391 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1393 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1395 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1396 imp->imp_known_replied_xid = req->rq_xid - 1;
1398 return req->rq_xid - 1;
1402 * Callback function called when client receives RPC reply for \a req.
1403 * Returns 0 on success or error code.
1404 * The return alue would be assigned to req->rq_status by the caller
1405 * as request processing status.
1406 * This function also decides if the request needs to be saved for later replay.
1408 static int after_reply(struct ptlrpc_request *req)
1410 struct obd_import *imp = req->rq_import;
1411 struct obd_device *obd = req->rq_import->imp_obd;
1418 LASSERT(obd != NULL);
1419 /* repbuf must be unlinked */
1420 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1422 if (req->rq_reply_truncated) {
1423 if (ptlrpc_no_resend(req)) {
1424 DEBUG_REQ(D_ERROR, req,
1425 "reply buffer overflow, expected: %d, actual size: %d",
1426 req->rq_nob_received, req->rq_repbuf_len);
1430 sptlrpc_cli_free_repbuf(req);
1432 * Pass the required reply buffer size (include
1433 * space for early reply).
1434 * NB: no need to roundup because alloc_repbuf
1437 req->rq_replen = req->rq_nob_received;
1438 req->rq_nob_received = 0;
1439 spin_lock(&req->rq_lock);
1441 spin_unlock(&req->rq_lock);
1445 work_start = ktime_get_real();
1446 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1449 * NB Until this point, the whole of the incoming message,
1450 * including buflens, status etc is in the sender's byte order.
1452 rc = sptlrpc_cli_unwrap_reply(req);
1454 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1459 * Security layer unwrap might ask resend this request.
1464 rc = unpack_reply(req);
1468 /* retry indefinitely on EINPROGRESS */
1469 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1470 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1471 time64_t now = ktime_get_real_seconds();
1473 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1474 spin_lock(&req->rq_lock);
1476 spin_unlock(&req->rq_lock);
1477 req->rq_nr_resend++;
1479 /* Readjust the timeout for current conditions */
1480 ptlrpc_at_set_req_timeout(req);
1482 * delay resend to give a chance to the server to get ready.
1483 * The delay is increased by 1s on every resend and is capped to
1484 * the current request timeout (i.e. obd_timeout if AT is off,
1485 * or AT service time x 125% + 5s, see at_est2timeout)
1487 if (req->rq_nr_resend > req->rq_timeout)
1488 req->rq_sent = now + req->rq_timeout;
1490 req->rq_sent = now + req->rq_nr_resend;
1492 /* Resend for EINPROGRESS will use a new XID */
1493 spin_lock(&imp->imp_lock);
1494 list_del_init(&req->rq_unreplied_list);
1495 spin_unlock(&imp->imp_lock);
1500 if (obd->obd_svc_stats) {
1501 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1503 ptlrpc_lprocfs_rpc_sent(req, timediff);
1506 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1507 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1508 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1509 lustre_msg_get_type(req->rq_repmsg));
1513 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1514 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1515 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1516 ptlrpc_at_adj_net_latency(req,
1517 lustre_msg_get_service_time(req->rq_repmsg));
1519 rc = ptlrpc_check_status(req);
1523 * Either we've been evicted, or the server has failed for
1524 * some reason. Try to reconnect, and if that fails, punt to
1527 if (ptlrpc_recoverable_error(rc)) {
1528 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1529 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1532 ptlrpc_request_handle_notconn(req);
1537 * Let's look if server sent slv. Do it only for RPC with
1540 ldlm_cli_update_pool(req);
1544 * Store transno in reqmsg for replay.
1546 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1547 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1548 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1551 if (imp->imp_replayable) {
1552 spin_lock(&imp->imp_lock);
1554 * No point in adding already-committed requests to the replay
1555 * list, we will just remove them immediately. b=9829
1557 if (req->rq_transno != 0 &&
1559 lustre_msg_get_last_committed(req->rq_repmsg) ||
1561 /** version recovery */
1562 ptlrpc_save_versions(req);
1563 ptlrpc_retain_replayable_request(req, imp);
1564 } else if (req->rq_commit_cb &&
1565 list_empty(&req->rq_replay_list)) {
1567 * NB: don't call rq_commit_cb if it's already on
1568 * rq_replay_list, ptlrpc_free_committed() will call
1569 * it later, see LU-3618 for details
1571 spin_unlock(&imp->imp_lock);
1572 req->rq_commit_cb(req);
1573 spin_lock(&imp->imp_lock);
1577 * Replay-enabled imports return commit-status information.
1579 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1580 if (likely(committed > imp->imp_peer_committed_transno))
1581 imp->imp_peer_committed_transno = committed;
1583 ptlrpc_free_committed(imp);
1585 if (!list_empty(&imp->imp_replay_list)) {
1586 struct ptlrpc_request *last;
1588 last = list_entry(imp->imp_replay_list.prev,
1589 struct ptlrpc_request,
1592 * Requests with rq_replay stay on the list even if no
1593 * commit is expected.
1595 if (last->rq_transno > imp->imp_peer_committed_transno)
1596 ptlrpc_pinger_commit_expected(imp);
1599 spin_unlock(&imp->imp_lock);
1606 * Helper function to send request \a req over the network for the first time
1607 * Also adjusts request phase.
1608 * Returns 0 on success or error code.
1610 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1612 struct obd_import *imp = req->rq_import;
1617 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1619 /* do not try to go further if there is not enough memory in enc_pool */
1620 if (req->rq_sent && req->rq_bulk)
1621 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1622 pool_is_at_full_capacity())
1625 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1626 (!req->rq_generation_set ||
1627 req->rq_import_generation == imp->imp_generation))
1630 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1632 spin_lock(&imp->imp_lock);
1634 LASSERT(req->rq_xid != 0);
1635 LASSERT(!list_empty(&req->rq_unreplied_list));
1637 if (!req->rq_generation_set)
1638 req->rq_import_generation = imp->imp_generation;
1640 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1641 spin_lock(&req->rq_lock);
1642 req->rq_waiting = 1;
1643 spin_unlock(&req->rq_lock);
1645 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1646 ptlrpc_import_state_name(req->rq_send_state),
1647 ptlrpc_import_state_name(imp->imp_state));
1648 LASSERT(list_empty(&req->rq_list));
1649 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1650 atomic_inc(&req->rq_import->imp_inflight);
1651 spin_unlock(&imp->imp_lock);
1656 spin_unlock(&imp->imp_lock);
1657 req->rq_status = rc;
1658 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1662 LASSERT(list_empty(&req->rq_list));
1663 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1664 atomic_inc(&req->rq_import->imp_inflight);
1667 * find the known replied XID from the unreplied list, CONNECT
1668 * and DISCONNECT requests are skipped to make the sanity check
1669 * on server side happy. see process_req_last_xid().
1671 * For CONNECT: Because replay requests have lower XID, it'll
1672 * break the sanity check if CONNECT bump the exp_last_xid on
1675 * For DISCONNECT: Since client will abort inflight RPC before
1676 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1677 * than the inflight RPC.
1679 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1680 min_xid = ptlrpc_known_replied_xid(imp);
1681 spin_unlock(&imp->imp_lock);
1683 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1685 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1687 rc = sptlrpc_req_refresh_ctx(req, -1);
1690 req->rq_status = rc;
1693 spin_lock(&req->rq_lock);
1694 req->rq_wait_ctx = 1;
1695 spin_unlock(&req->rq_lock);
1701 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1702 req, current_comm(),
1703 imp->imp_obd->obd_uuid.uuid,
1704 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1705 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1706 lustre_msg_get_jobid(req->rq_reqmsg));
1708 rc = ptl_send_rpc(req, 0);
1709 if (rc == -ENOMEM) {
1710 spin_lock(&imp->imp_lock);
1711 if (!list_empty(&req->rq_list)) {
1712 list_del_init(&req->rq_list);
1713 atomic_dec(&req->rq_import->imp_inflight);
1715 spin_unlock(&imp->imp_lock);
1716 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1720 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1721 spin_lock(&req->rq_lock);
1722 req->rq_net_err = 1;
1723 spin_unlock(&req->rq_lock);
1729 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1734 LASSERT(set->set_producer != NULL);
1736 remaining = atomic_read(&set->set_remaining);
1739 * populate the ->set_requests list with requests until we
1740 * reach the maximum number of RPCs in flight for this set
1742 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1743 rc = set->set_producer(set, set->set_producer_arg);
1744 if (rc == -ENOENT) {
1745 /* no more RPC to produce */
1746 set->set_producer = NULL;
1747 set->set_producer_arg = NULL;
1752 RETURN((atomic_read(&set->set_remaining) - remaining));
1756 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1757 * and no more replies are expected.
1758 * (it is possible to get less replies than requests sent e.g. due to timed out
1759 * requests or requests that we had trouble to send out)
1761 * NOTE: This function contains a potential schedule point (cond_resched()).
1763 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1765 struct list_head *tmp, *next;
1766 struct list_head comp_reqs;
1767 int force_timer_recalc = 0;
1770 if (atomic_read(&set->set_remaining) == 0)
1773 INIT_LIST_HEAD(&comp_reqs);
1774 list_for_each_safe(tmp, next, &set->set_requests) {
1775 struct ptlrpc_request *req =
1776 list_entry(tmp, struct ptlrpc_request,
1778 struct obd_import *imp = req->rq_import;
1779 int unregistered = 0;
1783 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1784 list_move_tail(&req->rq_set_chain, &comp_reqs);
1789 * This schedule point is mainly for the ptlrpcd caller of this
1790 * function. Most ptlrpc sets are not long-lived and unbounded
1791 * in length, but at the least the set used by the ptlrpcd is.
1792 * Since the processing time is unbounded, we need to insert an
1793 * explicit schedule point to make the thread well-behaved.
1798 * If the caller requires to allow to be interpreted by force
1799 * and it has really been interpreted, then move the request
1800 * to RQ_PHASE_INTERPRET phase in spite of what the current
1803 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1804 req->rq_status = -EINTR;
1805 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1808 * Since it is interpreted and we have to wait for
1809 * the reply to be unlinked, then use sync mode.
1813 GOTO(interpret, req->rq_status);
1816 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1817 force_timer_recalc = 1;
1819 /* delayed send - skip */
1820 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1823 /* delayed resend - skip */
1824 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1825 req->rq_sent > ktime_get_real_seconds())
1828 if (!(req->rq_phase == RQ_PHASE_RPC ||
1829 req->rq_phase == RQ_PHASE_BULK ||
1830 req->rq_phase == RQ_PHASE_INTERPRET ||
1831 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1832 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1833 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1837 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1838 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1839 LASSERT(req->rq_next_phase != req->rq_phase);
1840 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1842 if (req->rq_req_deadline &&
1843 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1844 req->rq_req_deadline = 0;
1845 if (req->rq_reply_deadline &&
1846 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1847 req->rq_reply_deadline = 0;
1848 if (req->rq_bulk_deadline &&
1849 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1850 req->rq_bulk_deadline = 0;
1853 * Skip processing until reply is unlinked. We
1854 * can't return to pool before that and we can't
1855 * call interpret before that. We need to make
1856 * sure that all rdma transfers finished and will
1857 * not corrupt any data.
1859 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1860 ptlrpc_client_recv_or_unlink(req))
1862 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1863 ptlrpc_client_bulk_active(req))
1867 * Turn fail_loc off to prevent it from looping
1870 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1871 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1874 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1875 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1880 * Move to next phase if reply was successfully
1883 ptlrpc_rqphase_move(req, req->rq_next_phase);
1886 if (req->rq_phase == RQ_PHASE_INTERPRET)
1887 GOTO(interpret, req->rq_status);
1890 * Note that this also will start async reply unlink.
1892 if (req->rq_net_err && !req->rq_timedout) {
1893 ptlrpc_expire_one_request(req, 1);
1896 * Check if we still need to wait for unlink.
1898 if (ptlrpc_client_recv_or_unlink(req) ||
1899 ptlrpc_client_bulk_active(req))
1901 /* If there is no need to resend, fail it now. */
1902 if (req->rq_no_resend) {
1903 if (req->rq_status == 0)
1904 req->rq_status = -EIO;
1905 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1906 GOTO(interpret, req->rq_status);
1913 spin_lock(&req->rq_lock);
1914 req->rq_replied = 0;
1915 spin_unlock(&req->rq_lock);
1916 if (req->rq_status == 0)
1917 req->rq_status = -EIO;
1918 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1919 GOTO(interpret, req->rq_status);
1923 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1924 * so it sets rq_intr regardless of individual rpc
1925 * timeouts. The synchronous IO waiting path sets
1926 * rq_intr irrespective of whether ptlrpcd
1927 * has seen a timeout. Our policy is to only interpret
1928 * interrupted rpcs after they have timed out, so we
1929 * need to enforce that here.
1932 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1933 req->rq_wait_ctx)) {
1934 req->rq_status = -EINTR;
1935 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1936 GOTO(interpret, req->rq_status);
1939 if (req->rq_phase == RQ_PHASE_RPC) {
1940 if (req->rq_timedout || req->rq_resend ||
1941 req->rq_waiting || req->rq_wait_ctx) {
1944 if (!ptlrpc_unregister_reply(req, 1)) {
1945 ptlrpc_unregister_bulk(req, 1);
1949 spin_lock(&imp->imp_lock);
1950 if (ptlrpc_import_delay_req(imp, req,
1953 * put on delay list - only if we wait
1954 * recovery finished - before send
1956 list_del_init(&req->rq_list);
1957 list_add_tail(&req->rq_list,
1958 &imp->imp_delayed_list);
1959 spin_unlock(&imp->imp_lock);
1964 req->rq_status = status;
1965 ptlrpc_rqphase_move(req,
1966 RQ_PHASE_INTERPRET);
1967 spin_unlock(&imp->imp_lock);
1968 GOTO(interpret, req->rq_status);
1970 /* ignore on just initiated connections */
1971 if (ptlrpc_no_resend(req) &&
1972 !req->rq_wait_ctx &&
1973 imp->imp_generation !=
1974 imp->imp_initiated_at) {
1975 req->rq_status = -ENOTCONN;
1976 ptlrpc_rqphase_move(req,
1977 RQ_PHASE_INTERPRET);
1978 spin_unlock(&imp->imp_lock);
1979 GOTO(interpret, req->rq_status);
1982 list_del_init(&req->rq_list);
1983 list_add_tail(&req->rq_list,
1984 &imp->imp_sending_list);
1986 spin_unlock(&imp->imp_lock);
1988 spin_lock(&req->rq_lock);
1989 req->rq_waiting = 0;
1990 spin_unlock(&req->rq_lock);
1992 if (req->rq_timedout || req->rq_resend) {
1994 * This is re-sending anyways,
1995 * let's mark req as resend.
1997 spin_lock(&req->rq_lock);
1999 spin_unlock(&req->rq_lock);
2002 * rq_wait_ctx is only touched by ptlrpcd,
2003 * so no lock is needed here.
2005 status = sptlrpc_req_refresh_ctx(req, -1);
2008 req->rq_status = status;
2009 spin_lock(&req->rq_lock);
2010 req->rq_wait_ctx = 0;
2011 spin_unlock(&req->rq_lock);
2012 force_timer_recalc = 1;
2014 spin_lock(&req->rq_lock);
2015 req->rq_wait_ctx = 1;
2016 spin_unlock(&req->rq_lock);
2021 spin_lock(&req->rq_lock);
2022 req->rq_wait_ctx = 0;
2023 spin_unlock(&req->rq_lock);
2027 * In any case, the previous bulk should be
2028 * cleaned up to prepare for the new sending
2031 !ptlrpc_unregister_bulk(req, 1))
2034 rc = ptl_send_rpc(req, 0);
2035 if (rc == -ENOMEM) {
2036 spin_lock(&imp->imp_lock);
2037 if (!list_empty(&req->rq_list))
2038 list_del_init(&req->rq_list);
2039 spin_unlock(&imp->imp_lock);
2040 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2044 DEBUG_REQ(D_HA, req,
2045 "send failed: rc = %d", rc);
2046 force_timer_recalc = 1;
2047 spin_lock(&req->rq_lock);
2048 req->rq_net_err = 1;
2049 spin_unlock(&req->rq_lock);
2052 /* need to reset the timeout */
2053 force_timer_recalc = 1;
2056 spin_lock(&req->rq_lock);
2058 if (ptlrpc_client_early(req)) {
2059 ptlrpc_at_recv_early_reply(req);
2060 spin_unlock(&req->rq_lock);
2064 /* Still waiting for a reply? */
2065 if (ptlrpc_client_recv(req)) {
2066 spin_unlock(&req->rq_lock);
2070 /* Did we actually receive a reply? */
2071 if (!ptlrpc_client_replied(req)) {
2072 spin_unlock(&req->rq_lock);
2076 spin_unlock(&req->rq_lock);
2079 * unlink from net because we are going to
2080 * swab in-place of reply buffer
2082 unregistered = ptlrpc_unregister_reply(req, 1);
2086 req->rq_status = after_reply(req);
2091 * If there is no bulk associated with this request,
2092 * then we're done and should let the interpreter
2093 * process the reply. Similarly if the RPC returned
2094 * an error, and therefore the bulk will never arrive.
2096 if (!req->rq_bulk || req->rq_status < 0) {
2097 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2098 GOTO(interpret, req->rq_status);
2101 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2104 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2105 if (ptlrpc_client_bulk_active(req))
2108 if (req->rq_bulk->bd_failure) {
2110 * The RPC reply arrived OK, but the bulk screwed
2111 * up! Dead weird since the server told us the RPC
2112 * was good after getting the REPLY for her GET or
2113 * the ACK for her PUT.
2115 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2116 req->rq_status = -EIO;
2119 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2122 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2125 * This moves to "unregistering" phase we need to wait for
2128 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2129 /* start async bulk unlink too */
2130 ptlrpc_unregister_bulk(req, 1);
2134 if (!ptlrpc_unregister_bulk(req, async))
2138 * When calling interpret receiving already should be
2141 LASSERT(!req->rq_receiving_reply);
2143 ptlrpc_req_interpret(env, req, req->rq_status);
2145 if (ptlrpcd_check_work(req)) {
2146 atomic_dec(&set->set_remaining);
2149 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2153 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2154 req, current_comm(),
2155 imp->imp_obd->obd_uuid.uuid,
2156 lustre_msg_get_status(req->rq_reqmsg),
2158 obd_import_nid2str(imp),
2159 lustre_msg_get_opc(req->rq_reqmsg),
2160 lustre_msg_get_jobid(req->rq_reqmsg));
2162 spin_lock(&imp->imp_lock);
2164 * Request already may be not on sending or delaying list. This
2165 * may happen in the case of marking it erroneous for the case
2166 * ptlrpc_import_delay_req(req, status) find it impossible to
2167 * allow sending this rpc and returns *status != 0.
2169 if (!list_empty(&req->rq_list)) {
2170 list_del_init(&req->rq_list);
2171 atomic_dec(&imp->imp_inflight);
2173 list_del_init(&req->rq_unreplied_list);
2174 spin_unlock(&imp->imp_lock);
2176 atomic_dec(&set->set_remaining);
2177 wake_up_all(&imp->imp_recovery_waitq);
2179 if (set->set_producer) {
2180 /* produce a new request if possible */
2181 if (ptlrpc_set_producer(set) > 0)
2182 force_timer_recalc = 1;
2185 * free the request that has just been completed
2186 * in order not to pollute set->set_requests
2188 list_del_init(&req->rq_set_chain);
2189 spin_lock(&req->rq_lock);
2191 req->rq_invalid_rqset = 0;
2192 spin_unlock(&req->rq_lock);
2194 /* record rq_status to compute the final status later */
2195 if (req->rq_status != 0)
2196 set->set_rc = req->rq_status;
2197 ptlrpc_req_finished(req);
2199 list_move_tail(&req->rq_set_chain, &comp_reqs);
2204 * move completed request at the head of list so it's easier for
2205 * caller to find them
2207 list_splice(&comp_reqs, &set->set_requests);
2209 /* If we hit an error, we want to recover promptly. */
2210 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2212 EXPORT_SYMBOL(ptlrpc_check_set);
2215 * Time out request \a req. is \a async_unlink is set, that means do not wait
2216 * until LNet actually confirms network buffer unlinking.
2217 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2219 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2221 struct obd_import *imp = req->rq_import;
2222 unsigned int debug_mask = D_RPCTRACE;
2226 spin_lock(&req->rq_lock);
2227 req->rq_timedout = 1;
2228 spin_unlock(&req->rq_lock);
2230 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2231 lustre_msg_get_status(req->rq_reqmsg)))
2232 debug_mask = D_WARNING;
2233 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2234 req->rq_net_err ? "failed due to network error" :
2235 ((req->rq_real_sent == 0 ||
2236 req->rq_real_sent < req->rq_sent ||
2237 req->rq_real_sent >= req->rq_deadline) ?
2238 "timed out for sent delay" : "timed out for slow reply"),
2239 (s64)req->rq_sent, (s64)req->rq_real_sent);
2241 if (imp && obd_debug_peer_on_timeout)
2242 LNetDebugPeer(imp->imp_connection->c_peer);
2244 ptlrpc_unregister_reply(req, async_unlink);
2245 ptlrpc_unregister_bulk(req, async_unlink);
2247 if (obd_dump_on_timeout)
2248 libcfs_debug_dumplog();
2251 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2255 atomic_inc(&imp->imp_timeouts);
2257 /* The DLM server doesn't want recovery run on its imports. */
2258 if (imp->imp_dlm_fake)
2262 * If this request is for recovery or other primordial tasks,
2263 * then error it out here.
2265 if (req->rq_ctx_init || req->rq_ctx_fini ||
2266 req->rq_send_state != LUSTRE_IMP_FULL ||
2267 imp->imp_obd->obd_no_recov) {
2268 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2269 ptlrpc_import_state_name(req->rq_send_state),
2270 ptlrpc_import_state_name(imp->imp_state));
2271 spin_lock(&req->rq_lock);
2272 req->rq_status = -ETIMEDOUT;
2274 spin_unlock(&req->rq_lock);
2279 * if a request can't be resent we can't wait for an answer after
2282 if (ptlrpc_no_resend(req)) {
2283 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2287 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2293 * Time out all uncompleted requests in request set pointed by \a data
2294 * Callback used when waiting on sets with l_wait_event.
2297 int ptlrpc_expired_set(void *data)
2299 struct ptlrpc_request_set *set = data;
2300 struct list_head *tmp;
2301 time64_t now = ktime_get_real_seconds();
2304 LASSERT(set != NULL);
2307 * A timeout expired. See which reqs it applies to...
2309 list_for_each(tmp, &set->set_requests) {
2310 struct ptlrpc_request *req =
2311 list_entry(tmp, struct ptlrpc_request,
2314 /* don't expire request waiting for context */
2315 if (req->rq_wait_ctx)
2318 /* Request in-flight? */
2319 if (!((req->rq_phase == RQ_PHASE_RPC &&
2320 !req->rq_waiting && !req->rq_resend) ||
2321 (req->rq_phase == RQ_PHASE_BULK)))
2324 if (req->rq_timedout || /* already dealt with */
2325 req->rq_deadline > now) /* not expired */
2329 * Deal with this guy. Do it asynchronously to not block
2332 ptlrpc_expire_one_request(req, 1);
2336 * When waiting for a whole set, we always break out of the
2337 * sleep so we can recalculate the timeout, or enable interrupts
2338 * if everyone's timed out.
2344 * Sets rq_intr flag in \a req under spinlock.
2346 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2348 spin_lock(&req->rq_lock);
2350 spin_unlock(&req->rq_lock);
2352 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2355 * Interrupts (sets interrupted flag) all uncompleted requests in
2356 * a set \a data. Callback for l_wait_event for interruptible waits.
2358 static void ptlrpc_interrupted_set(void *data)
2360 struct ptlrpc_request_set *set = data;
2361 struct list_head *tmp;
2363 LASSERT(set != NULL);
2364 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2366 list_for_each(tmp, &set->set_requests) {
2367 struct ptlrpc_request *req =
2368 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2373 if (req->rq_phase != RQ_PHASE_RPC &&
2374 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2375 !req->rq_allow_intr)
2378 ptlrpc_mark_interrupted(req);
2383 * Get the smallest timeout in the set; this does NOT set a timeout.
2385 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2387 struct list_head *tmp;
2388 time64_t now = ktime_get_real_seconds();
2390 struct ptlrpc_request *req;
2394 list_for_each(tmp, &set->set_requests) {
2395 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2397 /* Request in-flight? */
2398 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2399 (req->rq_phase == RQ_PHASE_BULK) ||
2400 (req->rq_phase == RQ_PHASE_NEW)))
2403 /* Already timed out. */
2404 if (req->rq_timedout)
2407 /* Waiting for ctx. */
2408 if (req->rq_wait_ctx)
2411 if (req->rq_phase == RQ_PHASE_NEW)
2412 deadline = req->rq_sent;
2413 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2414 deadline = req->rq_sent;
2416 deadline = req->rq_sent + req->rq_timeout;
2418 if (deadline <= now) /* actually expired already */
2419 timeout = 1; /* ASAP */
2420 else if (timeout == 0 || timeout > deadline - now)
2421 timeout = deadline - now;
2427 * Send all unset request from the set and then wait untill all
2428 * requests in the set complete (either get a reply, timeout, get an
2429 * error or otherwise be interrupted).
2430 * Returns 0 on success or error code otherwise.
2432 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2434 struct list_head *tmp;
2435 struct ptlrpc_request *req;
2436 struct l_wait_info lwi;
2441 if (set->set_producer)
2442 (void)ptlrpc_set_producer(set);
2444 list_for_each(tmp, &set->set_requests) {
2445 req = list_entry(tmp, struct ptlrpc_request,
2447 if (req->rq_phase == RQ_PHASE_NEW)
2448 (void)ptlrpc_send_new_req(req);
2451 if (list_empty(&set->set_requests))
2455 timeout = ptlrpc_set_next_timeout(set);
2458 * wait until all complete, interrupted, or an in-flight
2461 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2464 if ((timeout == 0 && !signal_pending(current)) ||
2465 set->set_allow_intr)
2467 * No requests are in-flight (ether timed out
2468 * or delayed), so we can allow interrupts.
2469 * We still want to block for a limited time,
2470 * so we allow interrupts during the timeout.
2472 lwi = LWI_TIMEOUT_INTR_ALL(
2473 cfs_time_seconds(timeout ? timeout : 1),
2475 ptlrpc_interrupted_set, set);
2478 * At least one request is in flight, so no
2479 * interrupts are allowed. Wait until all
2480 * complete, or an in-flight req times out.
2482 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2483 ptlrpc_expired_set, set);
2485 rc = l_wait_event(set->set_waitq,
2486 ptlrpc_check_set(NULL, set), &lwi);
2489 * LU-769 - if we ignored the signal because it was already
2490 * pending when we started, we need to handle it now or we risk
2491 * it being ignored forever
2493 if (rc == -ETIMEDOUT &&
2494 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2495 signal_pending(current)) {
2496 sigset_t blocked_sigs =
2497 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2500 * In fact we only interrupt for the "fatal" signals
2501 * like SIGINT or SIGKILL. We still ignore less
2502 * important signals since ptlrpc set is not easily
2503 * reentrant from userspace again
2505 if (signal_pending(current))
2506 ptlrpc_interrupted_set(set);
2507 cfs_restore_sigs(blocked_sigs);
2510 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2513 * -EINTR => all requests have been flagged rq_intr so next
2515 * -ETIMEDOUT => someone timed out. When all reqs have
2516 * timed out, signals are enabled allowing completion with
2518 * I don't really care if we go once more round the loop in
2519 * the error cases -eeb.
2521 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2522 list_for_each(tmp, &set->set_requests) {
2523 req = list_entry(tmp, struct ptlrpc_request,
2525 spin_lock(&req->rq_lock);
2526 req->rq_invalid_rqset = 1;
2527 spin_unlock(&req->rq_lock);
2530 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2532 LASSERT(atomic_read(&set->set_remaining) == 0);
2534 rc = set->set_rc; /* rq_status of already freed requests if any */
2535 list_for_each(tmp, &set->set_requests) {
2536 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2538 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2539 if (req->rq_status != 0)
2540 rc = req->rq_status;
2545 EXPORT_SYMBOL(ptlrpc_set_wait);
2548 * Helper fuction for request freeing.
2549 * Called when request count reached zero and request needs to be freed.
2550 * Removes request from all sorts of sending/replay lists it might be on,
2551 * frees network buffers if any are present.
2552 * If \a locked is set, that means caller is already holding import imp_lock
2553 * and so we no longer need to reobtain it (for certain lists manipulations)
2555 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2562 LASSERT(!request->rq_srv_req);
2563 LASSERT(request->rq_export == NULL);
2564 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2565 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2566 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2567 LASSERTF(!request->rq_replay, "req %p\n", request);
2569 req_capsule_fini(&request->rq_pill);
2572 * We must take it off the imp_replay_list first. Otherwise, we'll set
2573 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2575 if (request->rq_import) {
2577 spin_lock(&request->rq_import->imp_lock);
2578 list_del_init(&request->rq_replay_list);
2579 list_del_init(&request->rq_unreplied_list);
2581 spin_unlock(&request->rq_import->imp_lock);
2583 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2585 if (atomic_read(&request->rq_refcount) != 0) {
2586 DEBUG_REQ(D_ERROR, request,
2587 "freeing request with nonzero refcount");
2591 if (request->rq_repbuf)
2592 sptlrpc_cli_free_repbuf(request);
2594 if (request->rq_import) {
2595 class_import_put(request->rq_import);
2596 request->rq_import = NULL;
2598 if (request->rq_bulk)
2599 ptlrpc_free_bulk(request->rq_bulk);
2601 if (request->rq_reqbuf || request->rq_clrbuf)
2602 sptlrpc_cli_free_reqbuf(request);
2604 if (request->rq_cli_ctx)
2605 sptlrpc_req_put_ctx(request, !locked);
2607 if (request->rq_pool)
2608 __ptlrpc_free_req_to_pool(request);
2610 ptlrpc_request_cache_free(request);
2614 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2616 * Drop one request reference. Must be called with import imp_lock held.
2617 * When reference count drops to zero, request is freed.
2619 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2621 assert_spin_locked(&request->rq_import->imp_lock);
2622 (void)__ptlrpc_req_finished(request, 1);
2627 * Drops one reference count for request \a request.
2628 * \a locked set indicates that caller holds import imp_lock.
2629 * Frees the request whe reference count reaches zero.
2631 * \retval 1 the request is freed
2632 * \retval 0 some others still hold references on the request
2634 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2642 LASSERT(request != LP_POISON);
2643 LASSERT(request->rq_reqmsg != LP_POISON);
2645 DEBUG_REQ(D_INFO, request, "refcount now %u",
2646 atomic_read(&request->rq_refcount) - 1);
2648 spin_lock(&request->rq_lock);
2649 count = atomic_dec_return(&request->rq_refcount);
2650 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2653 * For open RPC, the client does not know the EA size (LOV, ACL, and
2654 * so on) before replied, then the client has to reserve very large
2655 * reply buffer. Such buffer will not be released until the RPC freed.
2656 * Since The open RPC is replayable, we need to keep it in the replay
2657 * list until close. If there are a lot of files opened concurrently,
2658 * then the client may be OOM.
2660 * If fact, it is unnecessary to keep reply buffer for open replay,
2661 * related EAs have already been saved via mdc_save_lovea() before
2662 * coming here. So it is safe to free the reply buffer some earlier
2663 * before releasing the RPC to avoid client OOM. LU-9514
2665 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2666 spin_lock(&request->rq_early_free_lock);
2667 sptlrpc_cli_free_repbuf(request);
2668 request->rq_repbuf = NULL;
2669 request->rq_repbuf_len = 0;
2670 request->rq_repdata = NULL;
2671 request->rq_reqdata_len = 0;
2672 spin_unlock(&request->rq_early_free_lock);
2674 spin_unlock(&request->rq_lock);
2677 __ptlrpc_free_req(request, locked);
2683 * Drops one reference count for a request.
2685 void ptlrpc_req_finished(struct ptlrpc_request *request)
2687 __ptlrpc_req_finished(request, 0);
2689 EXPORT_SYMBOL(ptlrpc_req_finished);
2692 * Returns xid of a \a request
2694 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2696 return request->rq_xid;
2698 EXPORT_SYMBOL(ptlrpc_req_xid);
2701 * Disengage the client's reply buffer from the network
2702 * NB does _NOT_ unregister any client-side bulk.
2703 * IDEMPOTENT, but _not_ safe against concurrent callers.
2704 * The request owner (i.e. the thread doing the I/O) must call...
2705 * Returns 0 on success or 1 if unregistering cannot be made.
2707 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2710 struct l_wait_info lwi;
2715 LASSERT(!in_interrupt());
2717 /* Let's setup deadline for reply unlink. */
2718 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2719 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2720 request->rq_reply_deadline = ktime_get_real_seconds() +
2724 * Nothing left to do.
2726 if (!ptlrpc_client_recv_or_unlink(request))
2729 LNetMDUnlink(request->rq_reply_md_h);
2732 * Let's check it once again.
2734 if (!ptlrpc_client_recv_or_unlink(request))
2737 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2738 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2741 * Do not wait for unlink to finish.
2747 * We have to l_wait_event() whatever the result, to give liblustre
2748 * a chance to run reply_in_callback(), and to make sure we've
2749 * unlinked before returning a req to the pool.
2752 /* The wq argument is ignored by user-space wait_event macros */
2753 wait_queue_head_t *wq = (request->rq_set) ?
2754 &request->rq_set->set_waitq :
2755 &request->rq_reply_waitq;
2757 * Network access will complete in finite time but the HUGE
2758 * timeout lets us CWARN for visibility of sluggish NALs
2760 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2761 cfs_time_seconds(1), NULL, NULL);
2762 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2765 ptlrpc_rqphase_move(request, request->rq_next_phase);
2769 LASSERT(rc == -ETIMEDOUT);
2770 DEBUG_REQ(D_WARNING, request,
2771 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2772 request->rq_receiving_reply,
2773 request->rq_req_unlinked,
2774 request->rq_reply_unlinked);
2779 static void ptlrpc_free_request(struct ptlrpc_request *req)
2781 spin_lock(&req->rq_lock);
2783 spin_unlock(&req->rq_lock);
2785 if (req->rq_commit_cb)
2786 req->rq_commit_cb(req);
2787 list_del_init(&req->rq_replay_list);
2789 __ptlrpc_req_finished(req, 1);
2793 * the request is committed and dropped from the replay list of its import
2795 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2797 struct obd_import *imp = req->rq_import;
2799 spin_lock(&imp->imp_lock);
2800 if (list_empty(&req->rq_replay_list)) {
2801 spin_unlock(&imp->imp_lock);
2805 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2806 if (imp->imp_replay_cursor == &req->rq_replay_list)
2807 imp->imp_replay_cursor = req->rq_replay_list.next;
2808 ptlrpc_free_request(req);
2811 spin_unlock(&imp->imp_lock);
2813 EXPORT_SYMBOL(ptlrpc_request_committed);
2816 * Iterates through replay_list on import and prunes
2817 * all requests have transno smaller than last_committed for the
2818 * import and don't have rq_replay set.
2819 * Since requests are sorted in transno order, stops when meetign first
2820 * transno bigger than last_committed.
2821 * caller must hold imp->imp_lock
2823 void ptlrpc_free_committed(struct obd_import *imp)
2825 struct ptlrpc_request *req, *saved;
2826 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2827 bool skip_committed_list = true;
2830 LASSERT(imp != NULL);
2831 assert_spin_locked(&imp->imp_lock);
2833 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2834 imp->imp_generation == imp->imp_last_generation_checked) {
2835 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2836 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2839 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2840 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2841 imp->imp_generation);
2843 if (imp->imp_generation != imp->imp_last_generation_checked ||
2844 imp->imp_last_transno_checked == 0)
2845 skip_committed_list = false;
2847 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2848 imp->imp_last_generation_checked = imp->imp_generation;
2850 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2852 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2853 LASSERT(req != last_req);
2856 if (req->rq_transno == 0) {
2857 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2860 if (req->rq_import_generation < imp->imp_generation) {
2861 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2865 /* not yet committed */
2866 if (req->rq_transno > imp->imp_peer_committed_transno) {
2867 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2871 if (req->rq_replay) {
2872 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2873 list_move_tail(&req->rq_replay_list,
2874 &imp->imp_committed_list);
2878 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2879 imp->imp_peer_committed_transno);
2881 ptlrpc_free_request(req);
2884 if (skip_committed_list)
2887 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2889 LASSERT(req->rq_transno != 0);
2890 if (req->rq_import_generation < imp->imp_generation ||
2892 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2893 req->rq_import_generation <
2894 imp->imp_generation ? "stale" : "closed");
2896 if (imp->imp_replay_cursor == &req->rq_replay_list)
2897 imp->imp_replay_cursor =
2898 req->rq_replay_list.next;
2900 ptlrpc_free_request(req);
2907 void ptlrpc_cleanup_client(struct obd_import *imp)
2914 * Schedule previously sent request for resend.
2915 * For bulk requests we assign new xid (to avoid problems with
2916 * lost replies and therefore several transfers landing into same buffer
2917 * from different sending attempts).
2919 void ptlrpc_resend_req(struct ptlrpc_request *req)
2921 DEBUG_REQ(D_HA, req, "going to resend");
2922 spin_lock(&req->rq_lock);
2925 * Request got reply but linked to the import list still.
2926 * Let ptlrpc_check_set() process it.
2928 if (ptlrpc_client_replied(req)) {
2929 spin_unlock(&req->rq_lock);
2930 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2934 req->rq_status = -EAGAIN;
2937 req->rq_net_err = 0;
2938 req->rq_timedout = 0;
2940 ptlrpc_client_wake_req(req);
2941 spin_unlock(&req->rq_lock);
2944 /* XXX: this function and rq_status are currently unused */
2945 void ptlrpc_restart_req(struct ptlrpc_request *req)
2947 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2948 req->rq_status = -ERESTARTSYS;
2950 spin_lock(&req->rq_lock);
2951 req->rq_restart = 1;
2952 req->rq_timedout = 0;
2953 ptlrpc_client_wake_req(req);
2954 spin_unlock(&req->rq_lock);
2958 * Grab additional reference on a request \a req
2960 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2963 atomic_inc(&req->rq_refcount);
2966 EXPORT_SYMBOL(ptlrpc_request_addref);
2969 * Add a request to import replay_list.
2970 * Must be called under imp_lock
2972 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2973 struct obd_import *imp)
2975 struct list_head *tmp;
2977 assert_spin_locked(&imp->imp_lock);
2979 if (req->rq_transno == 0) {
2980 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2985 * clear this for new requests that were resent as well
2986 * as resent replayed requests.
2988 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2990 /* don't re-add requests that have been replayed */
2991 if (!list_empty(&req->rq_replay_list))
2994 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2996 spin_lock(&req->rq_lock);
2998 spin_unlock(&req->rq_lock);
3000 LASSERT(imp->imp_replayable);
3001 /* Balanced in ptlrpc_free_committed, usually. */
3002 ptlrpc_request_addref(req);
3003 list_for_each_prev(tmp, &imp->imp_replay_list) {
3004 struct ptlrpc_request *iter = list_entry(tmp,
3005 struct ptlrpc_request,
3009 * We may have duplicate transnos if we create and then
3010 * open a file, or for closes retained if to match creating
3011 * opens, so use req->rq_xid as a secondary key.
3012 * (See bugs 684, 685, and 428.)
3013 * XXX no longer needed, but all opens need transnos!
3015 if (iter->rq_transno > req->rq_transno)
3018 if (iter->rq_transno == req->rq_transno) {
3019 LASSERT(iter->rq_xid != req->rq_xid);
3020 if (iter->rq_xid > req->rq_xid)
3024 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3028 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3032 * Send request and wait until it completes.
3033 * Returns request processing status.
3035 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3037 struct ptlrpc_request_set *set;
3041 LASSERT(req->rq_set == NULL);
3042 LASSERT(!req->rq_receiving_reply);
3044 set = ptlrpc_prep_set();
3046 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3050 /* for distributed debugging */
3051 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3053 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3054 ptlrpc_request_addref(req);
3055 ptlrpc_set_add_req(set, req);
3056 rc = ptlrpc_set_wait(NULL, set);
3057 ptlrpc_set_destroy(set);
3061 EXPORT_SYMBOL(ptlrpc_queue_wait);
3064 * Callback used for replayed requests reply processing.
3065 * In case of successful reply calls registered request replay callback.
3066 * In case of error restart replay process.
3068 static int ptlrpc_replay_interpret(const struct lu_env *env,
3069 struct ptlrpc_request *req,
3072 struct ptlrpc_replay_async_args *aa = args;
3073 struct obd_import *imp = req->rq_import;
3076 atomic_dec(&imp->imp_replay_inflight);
3079 * Note: if it is bulk replay (MDS-MDS replay), then even if
3080 * server got the request, but bulk transfer timeout, let's
3081 * replay the bulk req again
3083 if (!ptlrpc_client_replied(req) ||
3085 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3086 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
3087 GOTO(out, rc = -ETIMEDOUT);
3090 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3091 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3092 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3093 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3095 /** VBR: check version failure */
3096 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3097 /** replay was failed due to version mismatch */
3098 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
3099 spin_lock(&imp->imp_lock);
3100 imp->imp_vbr_failed = 1;
3101 spin_unlock(&imp->imp_lock);
3102 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3104 /** The transno had better not change over replay. */
3105 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3106 lustre_msg_get_transno(req->rq_repmsg) ||
3107 lustre_msg_get_transno(req->rq_repmsg) == 0,
3109 lustre_msg_get_transno(req->rq_reqmsg),
3110 lustre_msg_get_transno(req->rq_repmsg));
3113 spin_lock(&imp->imp_lock);
3114 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3115 spin_unlock(&imp->imp_lock);
3116 LASSERT(imp->imp_last_replay_transno);
3118 /* transaction number shouldn't be bigger than the latest replayed */
3119 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3120 DEBUG_REQ(D_ERROR, req,
3121 "Reported transno %llu is bigger than the replayed one: %llu",
3123 lustre_msg_get_transno(req->rq_reqmsg));
3124 GOTO(out, rc = -EINVAL);
3127 DEBUG_REQ(D_HA, req, "got rep");
3129 /* let the callback do fixups, possibly including in the request */
3130 if (req->rq_replay_cb)
3131 req->rq_replay_cb(req);
3133 if (ptlrpc_client_replied(req) &&
3134 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3135 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3136 lustre_msg_get_status(req->rq_repmsg),
3137 aa->praa_old_status);
3140 * Note: If the replay fails for MDT-MDT recovery, let's
3141 * abort all of the following requests in the replay
3142 * and sending list, because MDT-MDT update requests
3143 * are dependent on each other, see LU-7039
3145 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3146 struct ptlrpc_request *free_req;
3147 struct ptlrpc_request *tmp;
3149 spin_lock(&imp->imp_lock);
3150 list_for_each_entry_safe(free_req, tmp,
3151 &imp->imp_replay_list,
3153 ptlrpc_free_request(free_req);
3156 list_for_each_entry_safe(free_req, tmp,
3157 &imp->imp_committed_list,
3159 ptlrpc_free_request(free_req);
3162 list_for_each_entry_safe(free_req, tmp,
3163 &imp->imp_delayed_list,
3165 spin_lock(&free_req->rq_lock);
3166 free_req->rq_err = 1;
3167 free_req->rq_status = -EIO;
3168 ptlrpc_client_wake_req(free_req);
3169 spin_unlock(&free_req->rq_lock);
3172 list_for_each_entry_safe(free_req, tmp,
3173 &imp->imp_sending_list,
3175 spin_lock(&free_req->rq_lock);
3176 free_req->rq_err = 1;
3177 free_req->rq_status = -EIO;
3178 ptlrpc_client_wake_req(free_req);
3179 spin_unlock(&free_req->rq_lock);
3181 spin_unlock(&imp->imp_lock);
3184 /* Put it back for re-replay. */
3185 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3189 * Errors while replay can set transno to 0, but
3190 * imp_last_replay_transno shouldn't be set to 0 anyway
3192 if (req->rq_transno == 0)
3193 CERROR("Transno is 0 during replay!\n");
3195 /* continue with recovery */
3196 rc = ptlrpc_import_recovery_state_machine(imp);
3198 req->rq_send_state = aa->praa_old_state;
3201 /* this replay failed, so restart recovery */
3202 ptlrpc_connect_import(imp);
3208 * Prepares and queues request for replay.
3209 * Adds it to ptlrpcd queue for actual sending.
3210 * Returns 0 on success.
3212 int ptlrpc_replay_req(struct ptlrpc_request *req)
3214 struct ptlrpc_replay_async_args *aa;
3218 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3220 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3221 aa = ptlrpc_req_async_args(req);
3222 memset(aa, 0, sizeof(*aa));
3224 /* Prepare request to be resent with ptlrpcd */
3225 aa->praa_old_state = req->rq_send_state;
3226 req->rq_send_state = LUSTRE_IMP_REPLAY;
3227 req->rq_phase = RQ_PHASE_NEW;
3228 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3230 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3232 req->rq_interpret_reply = ptlrpc_replay_interpret;
3233 /* Readjust the timeout for current conditions */
3234 ptlrpc_at_set_req_timeout(req);
3236 /* Tell server net_latency to calculate how long to wait for reply. */
3237 lustre_msg_set_service_time(req->rq_reqmsg,
3238 ptlrpc_at_get_net_latency(req));
3239 DEBUG_REQ(D_HA, req, "REPLAY");
3241 atomic_inc(&req->rq_import->imp_replay_inflight);
3242 spin_lock(&req->rq_lock);
3243 req->rq_early_free_repbuf = 0;
3244 spin_unlock(&req->rq_lock);
3245 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3247 ptlrpcd_add_req(req);
3252 * Aborts all in-flight request on import \a imp sending and delayed lists
3254 void ptlrpc_abort_inflight(struct obd_import *imp)
3256 struct list_head *tmp, *n;
3260 * Make sure that no new requests get processed for this import.
3261 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3262 * this flag and then putting requests on sending_list or delayed_list.
3264 spin_lock(&imp->imp_lock);
3267 * XXX locking? Maybe we should remove each request with the list
3268 * locked? Also, how do we know if the requests on the list are
3269 * being freed at this time?
3271 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3272 struct ptlrpc_request *req = list_entry(tmp,
3273 struct ptlrpc_request,
3276 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3278 spin_lock(&req->rq_lock);
3279 if (req->rq_import_generation < imp->imp_generation) {
3281 req->rq_status = -EIO;
3282 ptlrpc_client_wake_req(req);
3284 spin_unlock(&req->rq_lock);
3287 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3288 struct ptlrpc_request *req =
3289 list_entry(tmp, struct ptlrpc_request, rq_list);
3291 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3293 spin_lock(&req->rq_lock);
3294 if (req->rq_import_generation < imp->imp_generation) {
3296 req->rq_status = -EIO;
3297 ptlrpc_client_wake_req(req);
3299 spin_unlock(&req->rq_lock);
3303 * Last chance to free reqs left on the replay list, but we
3304 * will still leak reqs that haven't committed.
3306 if (imp->imp_replayable)
3307 ptlrpc_free_committed(imp);
3309 spin_unlock(&imp->imp_lock);
3315 * Abort all uncompleted requests in request set \a set
3317 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3319 struct list_head *tmp, *pos;
3321 LASSERT(set != NULL);
3323 list_for_each_safe(pos, tmp, &set->set_requests) {
3324 struct ptlrpc_request *req =
3325 list_entry(pos, struct ptlrpc_request,
3328 spin_lock(&req->rq_lock);
3329 if (req->rq_phase != RQ_PHASE_RPC) {
3330 spin_unlock(&req->rq_lock);
3335 req->rq_status = -EINTR;
3336 ptlrpc_client_wake_req(req);
3337 spin_unlock(&req->rq_lock);
3342 * Initialize the XID for the node. This is common among all requests on
3343 * this node, and only requires the property that it is monotonically
3344 * increasing. It does not need to be sequential. Since this is also used
3345 * as the RDMA match bits, it is important that a single client NOT have
3346 * the same match bits for two different in-flight requests, hence we do
3347 * NOT want to have an XID per target or similar.
3349 * To avoid an unlikely collision between match bits after a client reboot
3350 * (which would deliver old data into the wrong RDMA buffer) initialize
3351 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3352 * If the time is clearly incorrect, we instead use a 62-bit random number.
3353 * In the worst case the random number will overflow 1M RPCs per second in
3354 * 9133 years, or permutations thereof.
3356 #define YEAR_2004 (1ULL << 30)
3357 void ptlrpc_init_xid(void)
3359 time64_t now = ktime_get_real_seconds();
3361 spin_lock_init(&ptlrpc_last_xid_lock);
3362 if (now < YEAR_2004) {
3363 get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3364 ptlrpc_last_xid >>= 2;
3365 ptlrpc_last_xid |= (1ULL << 61);
3367 ptlrpc_last_xid = (__u64)now << 20;
3370 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3371 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3372 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3376 * Increase xid and returns resulting new value to the caller.
3378 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3379 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3380 * itself uses the last bulk xid needed, so the server can determine the
3381 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3382 * xid must align to a power-of-two value.
3384 * This is assumed to be true due to the initial ptlrpc_last_xid
3385 * value also being initialized to a power-of-two value. LU-1431
3387 __u64 ptlrpc_next_xid(void)
3391 spin_lock(&ptlrpc_last_xid_lock);
3392 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3393 ptlrpc_last_xid = next;
3394 spin_unlock(&ptlrpc_last_xid_lock);
3400 * If request has a new allocated XID (new request or EINPROGRESS resend),
3401 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3402 * request to ensure previous bulk fails and avoid problems with lost replies
3403 * and therefore several transfers landing into the same buffer from different
3406 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3408 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3410 LASSERT(bd != NULL);
3413 * Generate new matchbits for all resend requests, including
3416 if (req->rq_resend) {
3417 __u64 old_mbits = req->rq_mbits;
3420 * First time resend on -EINPROGRESS will generate new xid,
3421 * so we can actually use the rq_xid as rq_mbits in such case,
3422 * however, it's bit hard to distinguish such resend with a
3423 * 'resend for the -EINPROGRESS resend'. To make it simple,
3424 * we opt to generate mbits for all resend cases.
3426 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3428 req->rq_mbits = ptlrpc_next_xid();
3431 * Old version transfers rq_xid to peer as
3434 spin_lock(&req->rq_import->imp_lock);
3435 list_del_init(&req->rq_unreplied_list);
3436 ptlrpc_assign_next_xid_nolock(req);
3437 spin_unlock(&req->rq_import->imp_lock);
3438 req->rq_mbits = req->rq_xid;
3440 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3441 old_mbits, req->rq_mbits);
3442 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3443 /* Request being sent first time, use xid as matchbits. */
3444 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3445 || req->rq_mbits == 0) {
3446 req->rq_mbits = req->rq_xid;
3448 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3450 req->rq_mbits -= total_md - 1;
3454 * Replay request, xid and matchbits have already been
3455 * correctly assigned.
3461 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3462 * that server can infer the number of bulks that were prepared,
3465 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3469 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3470 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3472 * It's ok to directly set the rq_xid here, since this xid bump
3473 * won't affect the request position in unreplied list.
3475 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3476 req->rq_xid = req->rq_mbits;
3480 * Get a glimpse at what next xid value might have been.
3481 * Returns possible next xid.
3483 __u64 ptlrpc_sample_next_xid(void)
3485 #if BITS_PER_LONG == 32
3486 /* need to avoid possible word tearing on 32-bit systems */
3489 spin_lock(&ptlrpc_last_xid_lock);
3490 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3491 spin_unlock(&ptlrpc_last_xid_lock);
3495 /* No need to lock, since returned value is racy anyways */
3496 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3499 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3502 * Functions for operating ptlrpc workers.
3504 * A ptlrpc work is a function which will be running inside ptlrpc context.
3505 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3507 * 1. after a work is created, it can be used many times, that is:
3508 * handler = ptlrpcd_alloc_work();
3509 * ptlrpcd_queue_work();
3511 * queue it again when necessary:
3512 * ptlrpcd_queue_work();
3513 * ptlrpcd_destroy_work();
3514 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3515 * it will only be queued once in any time. Also as its name implies, it may
3516 * have delay before it really runs by ptlrpcd thread.
3518 struct ptlrpc_work_async_args {
3519 int (*cb)(const struct lu_env *, void *);
3523 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3525 /* re-initialize the req */
3526 req->rq_timeout = obd_timeout;
3527 req->rq_sent = ktime_get_real_seconds();
3528 req->rq_deadline = req->rq_sent + req->rq_timeout;
3529 req->rq_phase = RQ_PHASE_INTERPRET;
3530 req->rq_next_phase = RQ_PHASE_COMPLETE;
3531 req->rq_xid = ptlrpc_next_xid();
3532 req->rq_import_generation = req->rq_import->imp_generation;
3534 ptlrpcd_add_req(req);
3537 static int work_interpreter(const struct lu_env *env,
3538 struct ptlrpc_request *req, void *args, int rc)
3540 struct ptlrpc_work_async_args *arg = args;
3542 LASSERT(ptlrpcd_check_work(req));
3543 LASSERT(arg->cb != NULL);
3545 rc = arg->cb(env, arg->cbdata);
3547 list_del_init(&req->rq_set_chain);
3550 if (atomic_dec_return(&req->rq_refcount) > 1) {
3551 atomic_set(&req->rq_refcount, 2);
3552 ptlrpcd_add_work_req(req);
3557 static int worker_format;
3559 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3561 return req->rq_pill.rc_fmt == (void *)&worker_format;
3565 * Create a work for ptlrpc.
3567 void *ptlrpcd_alloc_work(struct obd_import *imp,
3568 int (*cb)(const struct lu_env *, void *), void *cbdata)
3570 struct ptlrpc_request *req = NULL;
3571 struct ptlrpc_work_async_args *args;
3577 RETURN(ERR_PTR(-EINVAL));
3579 /* copy some code from deprecated fakereq. */
3580 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3582 CERROR("ptlrpc: run out of memory!\n");
3583 RETURN(ERR_PTR(-ENOMEM));
3586 ptlrpc_cli_req_init(req);
3588 req->rq_send_state = LUSTRE_IMP_FULL;
3589 req->rq_type = PTL_RPC_MSG_REQUEST;
3590 req->rq_import = class_import_get(imp);
3591 req->rq_interpret_reply = work_interpreter;
3592 /* don't want reply */
3593 req->rq_no_delay = req->rq_no_resend = 1;
3594 req->rq_pill.rc_fmt = (void *)&worker_format;
3596 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3597 args = ptlrpc_req_async_args(req);
3599 args->cbdata = cbdata;
3603 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3605 void ptlrpcd_destroy_work(void *handler)
3607 struct ptlrpc_request *req = handler;
3610 ptlrpc_req_finished(req);
3612 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3614 int ptlrpcd_queue_work(void *handler)
3616 struct ptlrpc_request *req = handler;
3619 * Check if the req is already being queued.
3621 * Here comes a trick: it lacks a way of checking if a req is being
3622 * processed reliably in ptlrpc. Here I have to use refcount of req
3623 * for this purpose. This is okay because the caller should use this
3624 * req as opaque data. - Jinshan
3626 LASSERT(atomic_read(&req->rq_refcount) > 0);
3627 if (atomic_inc_return(&req->rq_refcount) == 2)
3628 ptlrpcd_add_work_req(req);
3631 EXPORT_SYMBOL(ptlrpcd_queue_work);