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 atomic64_t ptlrpc_last_xid;
739 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
740 __u32 version, int opcode, char **bufs,
741 struct ptlrpc_cli_ctx *ctx)
744 struct obd_import *imp;
750 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
751 imp = request->rq_import;
752 lengths = request->rq_pill.rc_area[RCL_CLIENT];
755 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
757 rc = sptlrpc_req_get_ctx(request);
761 sptlrpc_req_set_flavor(request, opcode);
763 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
768 lustre_msg_add_version(request->rq_reqmsg, version);
769 request->rq_send_state = LUSTRE_IMP_FULL;
770 request->rq_type = PTL_RPC_MSG_REQUEST;
772 request->rq_req_cbid.cbid_fn = request_out_callback;
773 request->rq_req_cbid.cbid_arg = request;
775 request->rq_reply_cbid.cbid_fn = reply_in_callback;
776 request->rq_reply_cbid.cbid_arg = request;
778 request->rq_reply_deadline = 0;
779 request->rq_bulk_deadline = 0;
780 request->rq_req_deadline = 0;
781 request->rq_phase = RQ_PHASE_NEW;
782 request->rq_next_phase = RQ_PHASE_UNDEFINED;
784 request->rq_request_portal = imp->imp_client->cli_request_portal;
785 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
787 ptlrpc_at_set_req_timeout(request);
789 lustre_msg_set_opc(request->rq_reqmsg, opcode);
791 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
792 if (cfs_fail_val == opcode) {
793 time64_t *fail_t = NULL, *fail2_t = NULL;
795 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
796 fail_t = &request->rq_bulk_deadline;
797 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
798 fail_t = &request->rq_reply_deadline;
799 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
800 fail_t = &request->rq_req_deadline;
801 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
802 fail_t = &request->rq_reply_deadline;
803 fail2_t = &request->rq_bulk_deadline;
804 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
805 time64_t now = ktime_get_real_seconds();
806 u64 xid = ((u64)now >> 4) << 24;
808 atomic64_set(&ptlrpc_last_xid, xid);
812 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
815 *fail2_t = ktime_get_real_seconds() +
819 * The RPC is infected, let the test to change the
822 msleep(4 * MSEC_PER_SEC);
825 ptlrpc_assign_next_xid(request);
830 LASSERT(!request->rq_pool);
831 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
833 class_import_put(imp);
837 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
840 * Pack request buffers for network transfer, performing necessary encryption
841 * steps if necessary.
843 int ptlrpc_request_pack(struct ptlrpc_request *request,
844 __u32 version, int opcode)
848 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
853 * For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
854 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
855 * have to send old ptlrpc_body to keep interoprability with these
858 * Only three kinds of server->client RPCs so far:
863 * XXX This should be removed whenever we drop the interoprability with
864 * the these old clients.
866 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
867 opcode == LDLM_GL_CALLBACK)
868 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
869 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
873 EXPORT_SYMBOL(ptlrpc_request_pack);
876 * Helper function to allocate new request on import \a imp
877 * and possibly using existing request from pool \a pool if provided.
878 * Returns allocated request structure with import field filled or
882 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
883 struct ptlrpc_request_pool *pool)
885 struct ptlrpc_request *request = NULL;
887 request = ptlrpc_request_cache_alloc(GFP_NOFS);
889 if (!request && pool)
890 request = ptlrpc_prep_req_from_pool(pool);
893 ptlrpc_cli_req_init(request);
895 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
896 LASSERT(imp != LP_POISON);
897 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
899 LASSERT(imp->imp_client != LP_POISON);
901 request->rq_import = class_import_get(imp);
903 CERROR("request allocation out of memory\n");
910 * Helper function for creating a request.
911 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
912 * buffer structures according to capsule template \a format.
913 * Returns allocated request structure pointer or NULL on error.
915 static struct ptlrpc_request *
916 ptlrpc_request_alloc_internal(struct obd_import *imp,
917 struct ptlrpc_request_pool *pool,
918 const struct req_format *format)
920 struct ptlrpc_request *request;
923 request = __ptlrpc_request_alloc(imp, pool);
928 * initiate connection if needed when the import has been
929 * referenced by the new request to avoid races with disconnect
931 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
934 CDEBUG_LIMIT(imp->imp_idle_debug,
935 "%s: reconnect after %llds idle\n",
936 imp->imp_obd->obd_name, ktime_get_real_seconds() -
937 imp->imp_last_reply_time);
938 spin_lock(&imp->imp_lock);
939 if (imp->imp_state == LUSTRE_IMP_IDLE) {
940 imp->imp_generation++;
941 imp->imp_initiated_at = imp->imp_generation;
942 imp->imp_state = LUSTRE_IMP_NEW;
945 spin_unlock(&imp->imp_lock);
947 rc = ptlrpc_connect_import(imp);
949 ptlrpc_request_free(request);
952 ptlrpc_pinger_add_import(imp);
956 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
957 req_capsule_set(&request->rq_pill, format);
962 * Allocate new request structure for import \a imp and initialize its
963 * buffer structure according to capsule template \a format.
965 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
966 const struct req_format *format)
968 return ptlrpc_request_alloc_internal(imp, NULL, format);
970 EXPORT_SYMBOL(ptlrpc_request_alloc);
973 * Allocate new request structure for import \a imp from pool \a pool and
974 * initialize its buffer structure according to capsule template \a format.
976 struct ptlrpc_request *
977 ptlrpc_request_alloc_pool(struct obd_import *imp,
978 struct ptlrpc_request_pool *pool,
979 const struct req_format *format)
981 return ptlrpc_request_alloc_internal(imp, pool, format);
983 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
986 * For requests not from pool, free memory of the request structure.
987 * For requests obtained from a pool earlier, return request back to pool.
989 void ptlrpc_request_free(struct ptlrpc_request *request)
991 if (request->rq_pool)
992 __ptlrpc_free_req_to_pool(request);
994 ptlrpc_request_cache_free(request);
996 EXPORT_SYMBOL(ptlrpc_request_free);
999 * Allocate new request for operatione \a opcode and immediatelly pack it for
1001 * Only used for simple requests like OBD_PING where the only important
1002 * part of the request is operation itself.
1003 * Returns allocated request or NULL on error.
1005 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1006 const struct req_format *format,
1007 __u32 version, int opcode)
1009 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1013 rc = ptlrpc_request_pack(req, version, opcode);
1015 ptlrpc_request_free(req);
1021 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1024 * Allocate and initialize new request set structure on the current CPT.
1025 * Returns a pointer to the newly allocated set structure or NULL on error.
1027 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1029 struct ptlrpc_request_set *set;
1033 cpt = cfs_cpt_current(cfs_cpt_table, 0);
1034 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1037 atomic_set(&set->set_refcount, 1);
1038 INIT_LIST_HEAD(&set->set_requests);
1039 init_waitqueue_head(&set->set_waitq);
1040 atomic_set(&set->set_new_count, 0);
1041 atomic_set(&set->set_remaining, 0);
1042 spin_lock_init(&set->set_new_req_lock);
1043 INIT_LIST_HEAD(&set->set_new_requests);
1044 set->set_max_inflight = UINT_MAX;
1045 set->set_producer = NULL;
1046 set->set_producer_arg = NULL;
1051 EXPORT_SYMBOL(ptlrpc_prep_set);
1054 * Allocate and initialize new request set structure with flow control
1055 * extension. This extension allows to control the number of requests in-flight
1056 * for the whole set. A callback function to generate requests must be provided
1057 * and the request set will keep the number of requests sent over the wire to
1059 * Returns a pointer to the newly allocated set structure or NULL on error.
1061 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1065 struct ptlrpc_request_set *set;
1067 set = ptlrpc_prep_set();
1071 set->set_max_inflight = max;
1072 set->set_producer = func;
1073 set->set_producer_arg = arg;
1079 * Wind down and free request set structure previously allocated with
1081 * Ensures that all requests on the set have completed and removes
1082 * all requests from the request list in a set.
1083 * If any unsent request happen to be on the list, pretends that they got
1084 * an error in flight and calls their completion handler.
1086 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1088 struct list_head *tmp;
1089 struct list_head *next;
1095 /* Requests on the set should either all be completed, or all be new */
1096 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1097 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1098 list_for_each(tmp, &set->set_requests) {
1099 struct ptlrpc_request *req =
1100 list_entry(tmp, struct ptlrpc_request,
1103 LASSERT(req->rq_phase == expected_phase);
1107 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1108 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1109 atomic_read(&set->set_remaining), n);
1111 list_for_each_safe(tmp, next, &set->set_requests) {
1112 struct ptlrpc_request *req =
1113 list_entry(tmp, struct ptlrpc_request,
1115 list_del_init(&req->rq_set_chain);
1117 LASSERT(req->rq_phase == expected_phase);
1119 if (req->rq_phase == RQ_PHASE_NEW) {
1120 ptlrpc_req_interpret(NULL, req, -EBADR);
1121 atomic_dec(&set->set_remaining);
1124 spin_lock(&req->rq_lock);
1126 req->rq_invalid_rqset = 0;
1127 spin_unlock(&req->rq_lock);
1129 ptlrpc_req_finished(req);
1132 LASSERT(atomic_read(&set->set_remaining) == 0);
1134 ptlrpc_reqset_put(set);
1137 EXPORT_SYMBOL(ptlrpc_set_destroy);
1140 * Add a new request to the general purpose request set.
1141 * Assumes request reference from the caller.
1143 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1144 struct ptlrpc_request *req)
1146 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1147 LASSERT(list_empty(&req->rq_set_chain));
1149 if (req->rq_allow_intr)
1150 set->set_allow_intr = 1;
1152 /* The set takes over the caller's request reference */
1153 list_add_tail(&req->rq_set_chain, &set->set_requests);
1155 atomic_inc(&set->set_remaining);
1156 req->rq_queued_time = ktime_get_seconds();
1159 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1161 if (set->set_producer)
1163 * If the request set has a producer callback, the RPC must be
1164 * sent straight away
1166 ptlrpc_send_new_req(req);
1168 EXPORT_SYMBOL(ptlrpc_set_add_req);
1171 * Add a request to a request with dedicated server thread
1172 * and wake the thread to make any necessary processing.
1173 * Currently only used for ptlrpcd.
1175 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1176 struct ptlrpc_request *req)
1178 struct ptlrpc_request_set *set = pc->pc_set;
1181 LASSERT(req->rq_set == NULL);
1182 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1184 spin_lock(&set->set_new_req_lock);
1186 * The set takes over the caller's request reference.
1189 req->rq_queued_time = ktime_get_seconds();
1190 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1191 count = atomic_inc_return(&set->set_new_count);
1192 spin_unlock(&set->set_new_req_lock);
1194 /* Only need to call wakeup once for the first entry. */
1196 wake_up(&set->set_waitq);
1199 * XXX: It maybe unnecessary to wakeup all the partners. But to
1200 * guarantee the async RPC can be processed ASAP, we have
1201 * no other better choice. It maybe fixed in future.
1203 for (i = 0; i < pc->pc_npartners; i++)
1204 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1209 * Based on the current state of the import, determine if the request
1210 * can be sent, is an error, or should be delayed.
1212 * Returns true if this request should be delayed. If false, and
1213 * *status is set, then the request can not be sent and *status is the
1214 * error code. If false and status is 0, then request can be sent.
1216 * The imp->imp_lock must be held.
1218 static int ptlrpc_import_delay_req(struct obd_import *imp,
1219 struct ptlrpc_request *req, int *status)
1227 if (req->rq_ctx_init || req->rq_ctx_fini) {
1228 /* always allow ctx init/fini rpc go through */
1229 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1230 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1232 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1233 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1236 * pings or MDS-equivalent STATFS may safely
1239 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1240 D_HA : D_ERROR, req, "IMP_CLOSED ");
1242 } else if (ptlrpc_send_limit_expired(req)) {
1243 /* probably doesn't need to be a D_ERROR afterinitial testing */
1244 DEBUG_REQ(D_HA, req, "send limit expired ");
1245 *status = -ETIMEDOUT;
1246 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1247 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1248 ;/* allow CONNECT even if import is invalid */
1249 if (atomic_read(&imp->imp_inval_count) != 0) {
1250 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1253 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1254 if (!imp->imp_deactive)
1255 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1256 *status = -ESHUTDOWN; /* b=12940 */
1257 } else if (req->rq_import_generation != imp->imp_generation) {
1258 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1260 } else if (req->rq_send_state != imp->imp_state) {
1261 /* invalidate in progress - any requests should be drop */
1262 if (atomic_read(&imp->imp_inval_count) != 0) {
1263 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1265 } else if (req->rq_no_delay &&
1266 imp->imp_generation != imp->imp_initiated_at) {
1267 /* ignore nodelay for requests initiating connections */
1268 *status = -EWOULDBLOCK;
1269 } else if (req->rq_allow_replay &&
1270 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1271 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1272 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1273 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1274 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1284 * Decide if the error message should be printed to the console or not.
1285 * Makes its decision based on request type, status, and failure frequency.
1287 * \param[in] req request that failed and may need a console message
1289 * \retval false if no message should be printed
1290 * \retval true if console message should be printed
1292 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1294 LASSERT(req->rq_reqmsg != NULL);
1296 /* Suppress particular reconnect errors which are to be expected. */
1297 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1298 /* Suppress timed out reconnect requests */
1299 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1304 * Suppress most unavailable/again reconnect requests, but
1305 * print occasionally so it is clear client is trying to
1306 * connect to a server where no target is running.
1308 if ((err == -ENODEV || err == -EAGAIN) &&
1309 req->rq_import->imp_conn_cnt % 30 != 20)
1313 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1314 /* -EAGAIN is normal when using POSIX flocks */
1317 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1318 (req->rq_xid & 0xf) != 10)
1319 /* Suppress most ping requests, they may fail occasionally */
1326 * Check request processing status.
1327 * Returns the status.
1329 static int ptlrpc_check_status(struct ptlrpc_request *req)
1334 err = lustre_msg_get_status(req->rq_repmsg);
1335 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1336 struct obd_import *imp = req->rq_import;
1337 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1338 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1340 if (ptlrpc_console_allow(req, opc, err))
1341 LCONSOLE_ERROR_MSG(0x11,
1342 "%s: operation %s to node %s failed: rc = %d\n",
1343 imp->imp_obd->obd_name,
1345 libcfs_nid2str(nid), err);
1346 RETURN(err < 0 ? err : -EINVAL);
1350 DEBUG_REQ(D_INFO, req, "status is %d", err);
1351 } else if (err > 0) {
1352 /* XXX: translate this error from net to host */
1353 DEBUG_REQ(D_INFO, req, "status is %d", err);
1360 * save pre-versions of objects into request for replay.
1361 * Versions are obtained from server reply.
1364 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1366 struct lustre_msg *repmsg = req->rq_repmsg;
1367 struct lustre_msg *reqmsg = req->rq_reqmsg;
1368 __u64 *versions = lustre_msg_get_versions(repmsg);
1371 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1375 lustre_msg_set_versions(reqmsg, versions);
1376 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1377 versions[0], versions[1]);
1382 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1384 struct ptlrpc_request *req;
1386 assert_spin_locked(&imp->imp_lock);
1387 if (list_empty(&imp->imp_unreplied_list))
1390 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1392 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1394 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1395 imp->imp_known_replied_xid = req->rq_xid - 1;
1397 return req->rq_xid - 1;
1401 * Callback function called when client receives RPC reply for \a req.
1402 * Returns 0 on success or error code.
1403 * The return alue would be assigned to req->rq_status by the caller
1404 * as request processing status.
1405 * This function also decides if the request needs to be saved for later replay.
1407 static int after_reply(struct ptlrpc_request *req)
1409 struct obd_import *imp = req->rq_import;
1410 struct obd_device *obd = req->rq_import->imp_obd;
1417 LASSERT(obd != NULL);
1418 /* repbuf must be unlinked */
1419 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1421 if (req->rq_reply_truncated) {
1422 if (ptlrpc_no_resend(req)) {
1423 DEBUG_REQ(D_ERROR, req,
1424 "reply buffer overflow, expected: %d, actual size: %d",
1425 req->rq_nob_received, req->rq_repbuf_len);
1429 sptlrpc_cli_free_repbuf(req);
1431 * Pass the required reply buffer size (include
1432 * space for early reply).
1433 * NB: no need to roundup because alloc_repbuf
1436 req->rq_replen = req->rq_nob_received;
1437 req->rq_nob_received = 0;
1438 spin_lock(&req->rq_lock);
1440 spin_unlock(&req->rq_lock);
1444 work_start = ktime_get_real();
1445 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1448 * NB Until this point, the whole of the incoming message,
1449 * including buflens, status etc is in the sender's byte order.
1451 rc = sptlrpc_cli_unwrap_reply(req);
1453 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1458 * Security layer unwrap might ask resend this request.
1463 rc = unpack_reply(req);
1467 /* retry indefinitely on EINPROGRESS */
1468 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1469 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1470 time64_t now = ktime_get_real_seconds();
1472 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1473 spin_lock(&req->rq_lock);
1475 spin_unlock(&req->rq_lock);
1476 req->rq_nr_resend++;
1478 /* Readjust the timeout for current conditions */
1479 ptlrpc_at_set_req_timeout(req);
1481 * delay resend to give a chance to the server to get ready.
1482 * The delay is increased by 1s on every resend and is capped to
1483 * the current request timeout (i.e. obd_timeout if AT is off,
1484 * or AT service time x 125% + 5s, see at_est2timeout)
1486 if (req->rq_nr_resend > req->rq_timeout)
1487 req->rq_sent = now + req->rq_timeout;
1489 req->rq_sent = now + req->rq_nr_resend;
1491 /* Resend for EINPROGRESS will use a new XID */
1492 spin_lock(&imp->imp_lock);
1493 list_del_init(&req->rq_unreplied_list);
1494 spin_unlock(&imp->imp_lock);
1499 if (obd->obd_svc_stats) {
1500 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1502 ptlrpc_lprocfs_rpc_sent(req, timediff);
1505 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1506 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1507 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1508 lustre_msg_get_type(req->rq_repmsg));
1512 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1513 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1514 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1515 ptlrpc_at_adj_net_latency(req,
1516 lustre_msg_get_service_time(req->rq_repmsg));
1518 rc = ptlrpc_check_status(req);
1522 * Either we've been evicted, or the server has failed for
1523 * some reason. Try to reconnect, and if that fails, punt to
1526 if (ptlrpc_recoverable_error(rc)) {
1527 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1528 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1531 ptlrpc_request_handle_notconn(req);
1536 * Let's look if server sent slv. Do it only for RPC with
1539 ldlm_cli_update_pool(req);
1543 * Store transno in reqmsg for replay.
1545 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1546 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1547 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1550 if (imp->imp_replayable) {
1551 spin_lock(&imp->imp_lock);
1553 * No point in adding already-committed requests to the replay
1554 * list, we will just remove them immediately. b=9829
1556 if (req->rq_transno != 0 &&
1558 lustre_msg_get_last_committed(req->rq_repmsg) ||
1560 /** version recovery */
1561 ptlrpc_save_versions(req);
1562 ptlrpc_retain_replayable_request(req, imp);
1563 } else if (req->rq_commit_cb &&
1564 list_empty(&req->rq_replay_list)) {
1566 * NB: don't call rq_commit_cb if it's already on
1567 * rq_replay_list, ptlrpc_free_committed() will call
1568 * it later, see LU-3618 for details
1570 spin_unlock(&imp->imp_lock);
1571 req->rq_commit_cb(req);
1572 spin_lock(&imp->imp_lock);
1576 * Replay-enabled imports return commit-status information.
1578 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1579 if (likely(committed > imp->imp_peer_committed_transno))
1580 imp->imp_peer_committed_transno = committed;
1582 ptlrpc_free_committed(imp);
1584 if (!list_empty(&imp->imp_replay_list)) {
1585 struct ptlrpc_request *last;
1587 last = list_entry(imp->imp_replay_list.prev,
1588 struct ptlrpc_request,
1591 * Requests with rq_replay stay on the list even if no
1592 * commit is expected.
1594 if (last->rq_transno > imp->imp_peer_committed_transno)
1595 ptlrpc_pinger_commit_expected(imp);
1598 spin_unlock(&imp->imp_lock);
1605 * Helper function to send request \a req over the network for the first time
1606 * Also adjusts request phase.
1607 * Returns 0 on success or error code.
1609 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1611 struct obd_import *imp = req->rq_import;
1616 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1618 /* do not try to go further if there is not enough memory in enc_pool */
1619 if (req->rq_sent && req->rq_bulk)
1620 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1621 pool_is_at_full_capacity())
1624 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1625 (!req->rq_generation_set ||
1626 req->rq_import_generation == imp->imp_generation))
1629 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1631 spin_lock(&imp->imp_lock);
1633 LASSERT(req->rq_xid != 0);
1634 LASSERT(!list_empty(&req->rq_unreplied_list));
1636 if (!req->rq_generation_set)
1637 req->rq_import_generation = imp->imp_generation;
1639 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1640 spin_lock(&req->rq_lock);
1641 req->rq_waiting = 1;
1642 spin_unlock(&req->rq_lock);
1644 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1645 ptlrpc_import_state_name(req->rq_send_state),
1646 ptlrpc_import_state_name(imp->imp_state));
1647 LASSERT(list_empty(&req->rq_list));
1648 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1649 atomic_inc(&req->rq_import->imp_inflight);
1650 spin_unlock(&imp->imp_lock);
1655 spin_unlock(&imp->imp_lock);
1656 req->rq_status = rc;
1657 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1661 LASSERT(list_empty(&req->rq_list));
1662 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1663 atomic_inc(&req->rq_import->imp_inflight);
1666 * find the known replied XID from the unreplied list, CONNECT
1667 * and DISCONNECT requests are skipped to make the sanity check
1668 * on server side happy. see process_req_last_xid().
1670 * For CONNECT: Because replay requests have lower XID, it'll
1671 * break the sanity check if CONNECT bump the exp_last_xid on
1674 * For DISCONNECT: Since client will abort inflight RPC before
1675 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1676 * than the inflight RPC.
1678 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1679 min_xid = ptlrpc_known_replied_xid(imp);
1680 spin_unlock(&imp->imp_lock);
1682 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1684 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1686 rc = sptlrpc_req_refresh_ctx(req, -1);
1689 req->rq_status = rc;
1692 spin_lock(&req->rq_lock);
1693 req->rq_wait_ctx = 1;
1694 spin_unlock(&req->rq_lock);
1700 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1701 req, current_comm(),
1702 imp->imp_obd->obd_uuid.uuid,
1703 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1704 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1705 lustre_msg_get_jobid(req->rq_reqmsg));
1707 rc = ptl_send_rpc(req, 0);
1708 if (rc == -ENOMEM) {
1709 spin_lock(&imp->imp_lock);
1710 if (!list_empty(&req->rq_list)) {
1711 list_del_init(&req->rq_list);
1712 atomic_dec(&req->rq_import->imp_inflight);
1714 spin_unlock(&imp->imp_lock);
1715 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1719 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1720 spin_lock(&req->rq_lock);
1721 req->rq_net_err = 1;
1722 spin_unlock(&req->rq_lock);
1728 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1733 LASSERT(set->set_producer != NULL);
1735 remaining = atomic_read(&set->set_remaining);
1738 * populate the ->set_requests list with requests until we
1739 * reach the maximum number of RPCs in flight for this set
1741 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1742 rc = set->set_producer(set, set->set_producer_arg);
1743 if (rc == -ENOENT) {
1744 /* no more RPC to produce */
1745 set->set_producer = NULL;
1746 set->set_producer_arg = NULL;
1751 RETURN((atomic_read(&set->set_remaining) - remaining));
1755 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1756 * and no more replies are expected.
1757 * (it is possible to get less replies than requests sent e.g. due to timed out
1758 * requests or requests that we had trouble to send out)
1760 * NOTE: This function contains a potential schedule point (cond_resched()).
1762 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1764 struct list_head *tmp, *next;
1765 struct list_head comp_reqs;
1766 int force_timer_recalc = 0;
1769 if (atomic_read(&set->set_remaining) == 0)
1772 INIT_LIST_HEAD(&comp_reqs);
1773 list_for_each_safe(tmp, next, &set->set_requests) {
1774 struct ptlrpc_request *req =
1775 list_entry(tmp, struct ptlrpc_request,
1777 struct obd_import *imp = req->rq_import;
1778 int unregistered = 0;
1782 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1783 list_move_tail(&req->rq_set_chain, &comp_reqs);
1788 * This schedule point is mainly for the ptlrpcd caller of this
1789 * function. Most ptlrpc sets are not long-lived and unbounded
1790 * in length, but at the least the set used by the ptlrpcd is.
1791 * Since the processing time is unbounded, we need to insert an
1792 * explicit schedule point to make the thread well-behaved.
1797 * If the caller requires to allow to be interpreted by force
1798 * and it has really been interpreted, then move the request
1799 * to RQ_PHASE_INTERPRET phase in spite of what the current
1802 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1803 req->rq_status = -EINTR;
1804 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1807 * Since it is interpreted and we have to wait for
1808 * the reply to be unlinked, then use sync mode.
1812 GOTO(interpret, req->rq_status);
1815 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1816 force_timer_recalc = 1;
1818 /* delayed send - skip */
1819 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1822 /* delayed resend - skip */
1823 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1824 req->rq_sent > ktime_get_real_seconds())
1827 if (!(req->rq_phase == RQ_PHASE_RPC ||
1828 req->rq_phase == RQ_PHASE_BULK ||
1829 req->rq_phase == RQ_PHASE_INTERPRET ||
1830 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1831 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1832 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1836 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1837 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1838 LASSERT(req->rq_next_phase != req->rq_phase);
1839 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1841 if (req->rq_req_deadline &&
1842 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1843 req->rq_req_deadline = 0;
1844 if (req->rq_reply_deadline &&
1845 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1846 req->rq_reply_deadline = 0;
1847 if (req->rq_bulk_deadline &&
1848 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1849 req->rq_bulk_deadline = 0;
1852 * Skip processing until reply is unlinked. We
1853 * can't return to pool before that and we can't
1854 * call interpret before that. We need to make
1855 * sure that all rdma transfers finished and will
1856 * not corrupt any data.
1858 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1859 ptlrpc_client_recv_or_unlink(req))
1861 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1862 ptlrpc_client_bulk_active(req))
1866 * Turn fail_loc off to prevent it from looping
1869 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1870 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1873 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1874 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1879 * Move to next phase if reply was successfully
1882 ptlrpc_rqphase_move(req, req->rq_next_phase);
1885 if (req->rq_phase == RQ_PHASE_INTERPRET)
1886 GOTO(interpret, req->rq_status);
1889 * Note that this also will start async reply unlink.
1891 if (req->rq_net_err && !req->rq_timedout) {
1892 ptlrpc_expire_one_request(req, 1);
1895 * Check if we still need to wait for unlink.
1897 if (ptlrpc_client_recv_or_unlink(req) ||
1898 ptlrpc_client_bulk_active(req))
1900 /* If there is no need to resend, fail it now. */
1901 if (req->rq_no_resend) {
1902 if (req->rq_status == 0)
1903 req->rq_status = -EIO;
1904 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1905 GOTO(interpret, req->rq_status);
1912 spin_lock(&req->rq_lock);
1913 req->rq_replied = 0;
1914 spin_unlock(&req->rq_lock);
1915 if (req->rq_status == 0)
1916 req->rq_status = -EIO;
1917 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1918 GOTO(interpret, req->rq_status);
1922 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1923 * so it sets rq_intr regardless of individual rpc
1924 * timeouts. The synchronous IO waiting path sets
1925 * rq_intr irrespective of whether ptlrpcd
1926 * has seen a timeout. Our policy is to only interpret
1927 * interrupted rpcs after they have timed out, so we
1928 * need to enforce that here.
1931 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1932 req->rq_wait_ctx)) {
1933 req->rq_status = -EINTR;
1934 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1935 GOTO(interpret, req->rq_status);
1938 if (req->rq_phase == RQ_PHASE_RPC) {
1939 if (req->rq_timedout || req->rq_resend ||
1940 req->rq_waiting || req->rq_wait_ctx) {
1943 if (!ptlrpc_unregister_reply(req, 1)) {
1944 ptlrpc_unregister_bulk(req, 1);
1948 spin_lock(&imp->imp_lock);
1949 if (ptlrpc_import_delay_req(imp, req,
1952 * put on delay list - only if we wait
1953 * recovery finished - before send
1955 list_move_tail(&req->rq_list,
1956 &imp->imp_delayed_list);
1957 spin_unlock(&imp->imp_lock);
1962 req->rq_status = status;
1963 ptlrpc_rqphase_move(req,
1964 RQ_PHASE_INTERPRET);
1965 spin_unlock(&imp->imp_lock);
1966 GOTO(interpret, req->rq_status);
1968 /* ignore on just initiated connections */
1969 if (ptlrpc_no_resend(req) &&
1970 !req->rq_wait_ctx &&
1971 imp->imp_generation !=
1972 imp->imp_initiated_at) {
1973 req->rq_status = -ENOTCONN;
1974 ptlrpc_rqphase_move(req,
1975 RQ_PHASE_INTERPRET);
1976 spin_unlock(&imp->imp_lock);
1977 GOTO(interpret, req->rq_status);
1980 list_move_tail(&req->rq_list,
1981 &imp->imp_sending_list);
1983 spin_unlock(&imp->imp_lock);
1985 spin_lock(&req->rq_lock);
1986 req->rq_waiting = 0;
1987 spin_unlock(&req->rq_lock);
1989 if (req->rq_timedout || req->rq_resend) {
1991 * This is re-sending anyways,
1992 * let's mark req as resend.
1994 spin_lock(&req->rq_lock);
1996 spin_unlock(&req->rq_lock);
1999 * rq_wait_ctx is only touched by ptlrpcd,
2000 * so no lock is needed here.
2002 status = sptlrpc_req_refresh_ctx(req, -1);
2005 req->rq_status = status;
2006 spin_lock(&req->rq_lock);
2007 req->rq_wait_ctx = 0;
2008 spin_unlock(&req->rq_lock);
2009 force_timer_recalc = 1;
2011 spin_lock(&req->rq_lock);
2012 req->rq_wait_ctx = 1;
2013 spin_unlock(&req->rq_lock);
2018 spin_lock(&req->rq_lock);
2019 req->rq_wait_ctx = 0;
2020 spin_unlock(&req->rq_lock);
2024 * In any case, the previous bulk should be
2025 * cleaned up to prepare for the new sending
2028 !ptlrpc_unregister_bulk(req, 1))
2031 rc = ptl_send_rpc(req, 0);
2032 if (rc == -ENOMEM) {
2033 spin_lock(&imp->imp_lock);
2034 if (!list_empty(&req->rq_list))
2035 list_del_init(&req->rq_list);
2036 spin_unlock(&imp->imp_lock);
2037 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2041 DEBUG_REQ(D_HA, req,
2042 "send failed: rc = %d", rc);
2043 force_timer_recalc = 1;
2044 spin_lock(&req->rq_lock);
2045 req->rq_net_err = 1;
2046 spin_unlock(&req->rq_lock);
2049 /* need to reset the timeout */
2050 force_timer_recalc = 1;
2053 spin_lock(&req->rq_lock);
2055 if (ptlrpc_client_early(req)) {
2056 ptlrpc_at_recv_early_reply(req);
2057 spin_unlock(&req->rq_lock);
2061 /* Still waiting for a reply? */
2062 if (ptlrpc_client_recv(req)) {
2063 spin_unlock(&req->rq_lock);
2067 /* Did we actually receive a reply? */
2068 if (!ptlrpc_client_replied(req)) {
2069 spin_unlock(&req->rq_lock);
2073 spin_unlock(&req->rq_lock);
2076 * unlink from net because we are going to
2077 * swab in-place of reply buffer
2079 unregistered = ptlrpc_unregister_reply(req, 1);
2083 req->rq_status = after_reply(req);
2088 * If there is no bulk associated with this request,
2089 * then we're done and should let the interpreter
2090 * process the reply. Similarly if the RPC returned
2091 * an error, and therefore the bulk will never arrive.
2093 if (!req->rq_bulk || req->rq_status < 0) {
2094 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2095 GOTO(interpret, req->rq_status);
2098 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2101 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2102 if (ptlrpc_client_bulk_active(req))
2105 if (req->rq_bulk->bd_failure) {
2107 * The RPC reply arrived OK, but the bulk screwed
2108 * up! Dead weird since the server told us the RPC
2109 * was good after getting the REPLY for her GET or
2110 * the ACK for her PUT.
2112 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2113 req->rq_status = -EIO;
2116 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2119 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2122 * This moves to "unregistering" phase we need to wait for
2125 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2126 /* start async bulk unlink too */
2127 ptlrpc_unregister_bulk(req, 1);
2131 if (!ptlrpc_unregister_bulk(req, async))
2135 * When calling interpret receiving already should be
2138 LASSERT(!req->rq_receiving_reply);
2140 ptlrpc_req_interpret(env, req, req->rq_status);
2142 if (ptlrpcd_check_work(req)) {
2143 atomic_dec(&set->set_remaining);
2146 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2150 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2151 req, current_comm(),
2152 imp->imp_obd->obd_uuid.uuid,
2153 lustre_msg_get_status(req->rq_reqmsg),
2155 obd_import_nid2str(imp),
2156 lustre_msg_get_opc(req->rq_reqmsg),
2157 lustre_msg_get_jobid(req->rq_reqmsg));
2159 spin_lock(&imp->imp_lock);
2161 * Request already may be not on sending or delaying list. This
2162 * may happen in the case of marking it erroneous for the case
2163 * ptlrpc_import_delay_req(req, status) find it impossible to
2164 * allow sending this rpc and returns *status != 0.
2166 if (!list_empty(&req->rq_list)) {
2167 list_del_init(&req->rq_list);
2168 atomic_dec(&imp->imp_inflight);
2170 list_del_init(&req->rq_unreplied_list);
2171 spin_unlock(&imp->imp_lock);
2173 atomic_dec(&set->set_remaining);
2174 wake_up_all(&imp->imp_recovery_waitq);
2176 if (set->set_producer) {
2177 /* produce a new request if possible */
2178 if (ptlrpc_set_producer(set) > 0)
2179 force_timer_recalc = 1;
2182 * free the request that has just been completed
2183 * in order not to pollute set->set_requests
2185 list_del_init(&req->rq_set_chain);
2186 spin_lock(&req->rq_lock);
2188 req->rq_invalid_rqset = 0;
2189 spin_unlock(&req->rq_lock);
2191 /* record rq_status to compute the final status later */
2192 if (req->rq_status != 0)
2193 set->set_rc = req->rq_status;
2194 ptlrpc_req_finished(req);
2196 list_move_tail(&req->rq_set_chain, &comp_reqs);
2201 * move completed request at the head of list so it's easier for
2202 * caller to find them
2204 list_splice(&comp_reqs, &set->set_requests);
2206 /* If we hit an error, we want to recover promptly. */
2207 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2209 EXPORT_SYMBOL(ptlrpc_check_set);
2212 * Time out request \a req. is \a async_unlink is set, that means do not wait
2213 * until LNet actually confirms network buffer unlinking.
2214 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2216 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2218 struct obd_import *imp = req->rq_import;
2219 unsigned int debug_mask = D_RPCTRACE;
2223 spin_lock(&req->rq_lock);
2224 req->rq_timedout = 1;
2225 spin_unlock(&req->rq_lock);
2227 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2228 lustre_msg_get_status(req->rq_reqmsg)))
2229 debug_mask = D_WARNING;
2230 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2231 req->rq_net_err ? "failed due to network error" :
2232 ((req->rq_real_sent == 0 ||
2233 req->rq_real_sent < req->rq_sent ||
2234 req->rq_real_sent >= req->rq_deadline) ?
2235 "timed out for sent delay" : "timed out for slow reply"),
2236 (s64)req->rq_sent, (s64)req->rq_real_sent);
2238 if (imp && obd_debug_peer_on_timeout)
2239 LNetDebugPeer(imp->imp_connection->c_peer);
2241 ptlrpc_unregister_reply(req, async_unlink);
2242 ptlrpc_unregister_bulk(req, async_unlink);
2244 if (obd_dump_on_timeout)
2245 libcfs_debug_dumplog();
2248 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2252 atomic_inc(&imp->imp_timeouts);
2254 /* The DLM server doesn't want recovery run on its imports. */
2255 if (imp->imp_dlm_fake)
2259 * If this request is for recovery or other primordial tasks,
2260 * then error it out here.
2262 if (req->rq_ctx_init || req->rq_ctx_fini ||
2263 req->rq_send_state != LUSTRE_IMP_FULL ||
2264 imp->imp_obd->obd_no_recov) {
2265 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2266 ptlrpc_import_state_name(req->rq_send_state),
2267 ptlrpc_import_state_name(imp->imp_state));
2268 spin_lock(&req->rq_lock);
2269 req->rq_status = -ETIMEDOUT;
2271 spin_unlock(&req->rq_lock);
2276 * if a request can't be resent we can't wait for an answer after
2279 if (ptlrpc_no_resend(req)) {
2280 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2284 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2290 * Time out all uncompleted requests in request set pointed by \a data
2291 * Callback used when waiting on sets with l_wait_event.
2294 int ptlrpc_expired_set(void *data)
2296 struct ptlrpc_request_set *set = data;
2297 struct list_head *tmp;
2298 time64_t now = ktime_get_real_seconds();
2301 LASSERT(set != NULL);
2304 * A timeout expired. See which reqs it applies to...
2306 list_for_each(tmp, &set->set_requests) {
2307 struct ptlrpc_request *req =
2308 list_entry(tmp, struct ptlrpc_request,
2311 /* don't expire request waiting for context */
2312 if (req->rq_wait_ctx)
2315 /* Request in-flight? */
2316 if (!((req->rq_phase == RQ_PHASE_RPC &&
2317 !req->rq_waiting && !req->rq_resend) ||
2318 (req->rq_phase == RQ_PHASE_BULK)))
2321 if (req->rq_timedout || /* already dealt with */
2322 req->rq_deadline > now) /* not expired */
2326 * Deal with this guy. Do it asynchronously to not block
2329 ptlrpc_expire_one_request(req, 1);
2333 * When waiting for a whole set, we always break out of the
2334 * sleep so we can recalculate the timeout, or enable interrupts
2335 * if everyone's timed out.
2341 * Sets rq_intr flag in \a req under spinlock.
2343 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2345 spin_lock(&req->rq_lock);
2347 spin_unlock(&req->rq_lock);
2349 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2352 * Interrupts (sets interrupted flag) all uncompleted requests in
2353 * a set \a data. Callback for l_wait_event for interruptible waits.
2355 static void ptlrpc_interrupted_set(void *data)
2357 struct ptlrpc_request_set *set = data;
2358 struct list_head *tmp;
2360 LASSERT(set != NULL);
2361 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2363 list_for_each(tmp, &set->set_requests) {
2364 struct ptlrpc_request *req =
2365 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2370 if (req->rq_phase != RQ_PHASE_RPC &&
2371 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2372 !req->rq_allow_intr)
2375 ptlrpc_mark_interrupted(req);
2380 * Get the smallest timeout in the set; this does NOT set a timeout.
2382 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2384 struct list_head *tmp;
2385 time64_t now = ktime_get_real_seconds();
2387 struct ptlrpc_request *req;
2391 list_for_each(tmp, &set->set_requests) {
2392 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2394 /* Request in-flight? */
2395 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2396 (req->rq_phase == RQ_PHASE_BULK) ||
2397 (req->rq_phase == RQ_PHASE_NEW)))
2400 /* Already timed out. */
2401 if (req->rq_timedout)
2404 /* Waiting for ctx. */
2405 if (req->rq_wait_ctx)
2408 if (req->rq_phase == RQ_PHASE_NEW)
2409 deadline = req->rq_sent;
2410 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2411 deadline = req->rq_sent;
2413 deadline = req->rq_sent + req->rq_timeout;
2415 if (deadline <= now) /* actually expired already */
2416 timeout = 1; /* ASAP */
2417 else if (timeout == 0 || timeout > deadline - now)
2418 timeout = deadline - now;
2424 * Send all unset request from the set and then wait untill all
2425 * requests in the set complete (either get a reply, timeout, get an
2426 * error or otherwise be interrupted).
2427 * Returns 0 on success or error code otherwise.
2429 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2431 struct list_head *tmp;
2432 struct ptlrpc_request *req;
2433 struct l_wait_info lwi;
2438 if (set->set_producer)
2439 (void)ptlrpc_set_producer(set);
2441 list_for_each(tmp, &set->set_requests) {
2442 req = list_entry(tmp, struct ptlrpc_request,
2444 if (req->rq_phase == RQ_PHASE_NEW)
2445 (void)ptlrpc_send_new_req(req);
2448 if (list_empty(&set->set_requests))
2452 timeout = ptlrpc_set_next_timeout(set);
2455 * wait until all complete, interrupted, or an in-flight
2458 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2461 if ((timeout == 0 && !signal_pending(current)) ||
2462 set->set_allow_intr)
2464 * No requests are in-flight (ether timed out
2465 * or delayed), so we can allow interrupts.
2466 * We still want to block for a limited time,
2467 * so we allow interrupts during the timeout.
2469 lwi = LWI_TIMEOUT_INTR_ALL(
2470 cfs_time_seconds(timeout ? timeout : 1),
2472 ptlrpc_interrupted_set, set);
2475 * At least one request is in flight, so no
2476 * interrupts are allowed. Wait until all
2477 * complete, or an in-flight req times out.
2479 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2480 ptlrpc_expired_set, set);
2482 rc = l_wait_event(set->set_waitq,
2483 ptlrpc_check_set(NULL, set), &lwi);
2486 * LU-769 - if we ignored the signal because it was already
2487 * pending when we started, we need to handle it now or we risk
2488 * it being ignored forever
2490 if (rc == -ETIMEDOUT &&
2491 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2492 signal_pending(current)) {
2493 sigset_t blocked_sigs =
2494 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2497 * In fact we only interrupt for the "fatal" signals
2498 * like SIGINT or SIGKILL. We still ignore less
2499 * important signals since ptlrpc set is not easily
2500 * reentrant from userspace again
2502 if (signal_pending(current))
2503 ptlrpc_interrupted_set(set);
2504 cfs_restore_sigs(blocked_sigs);
2507 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2510 * -EINTR => all requests have been flagged rq_intr so next
2512 * -ETIMEDOUT => someone timed out. When all reqs have
2513 * timed out, signals are enabled allowing completion with
2515 * I don't really care if we go once more round the loop in
2516 * the error cases -eeb.
2518 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2519 list_for_each(tmp, &set->set_requests) {
2520 req = list_entry(tmp, struct ptlrpc_request,
2522 spin_lock(&req->rq_lock);
2523 req->rq_invalid_rqset = 1;
2524 spin_unlock(&req->rq_lock);
2527 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2529 LASSERT(atomic_read(&set->set_remaining) == 0);
2531 rc = set->set_rc; /* rq_status of already freed requests if any */
2532 list_for_each(tmp, &set->set_requests) {
2533 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2535 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2536 if (req->rq_status != 0)
2537 rc = req->rq_status;
2542 EXPORT_SYMBOL(ptlrpc_set_wait);
2545 * Helper fuction for request freeing.
2546 * Called when request count reached zero and request needs to be freed.
2547 * Removes request from all sorts of sending/replay lists it might be on,
2548 * frees network buffers if any are present.
2549 * If \a locked is set, that means caller is already holding import imp_lock
2550 * and so we no longer need to reobtain it (for certain lists manipulations)
2552 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2559 LASSERT(!request->rq_srv_req);
2560 LASSERT(request->rq_export == NULL);
2561 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2562 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2563 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2564 LASSERTF(!request->rq_replay, "req %p\n", request);
2566 req_capsule_fini(&request->rq_pill);
2569 * We must take it off the imp_replay_list first. Otherwise, we'll set
2570 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2572 if (request->rq_import) {
2574 spin_lock(&request->rq_import->imp_lock);
2575 list_del_init(&request->rq_replay_list);
2576 list_del_init(&request->rq_unreplied_list);
2578 spin_unlock(&request->rq_import->imp_lock);
2580 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2582 if (atomic_read(&request->rq_refcount) != 0) {
2583 DEBUG_REQ(D_ERROR, request,
2584 "freeing request with nonzero refcount");
2588 if (request->rq_repbuf)
2589 sptlrpc_cli_free_repbuf(request);
2591 if (request->rq_import) {
2592 class_import_put(request->rq_import);
2593 request->rq_import = NULL;
2595 if (request->rq_bulk)
2596 ptlrpc_free_bulk(request->rq_bulk);
2598 if (request->rq_reqbuf || request->rq_clrbuf)
2599 sptlrpc_cli_free_reqbuf(request);
2601 if (request->rq_cli_ctx)
2602 sptlrpc_req_put_ctx(request, !locked);
2604 if (request->rq_pool)
2605 __ptlrpc_free_req_to_pool(request);
2607 ptlrpc_request_cache_free(request);
2611 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2613 * Drop one request reference. Must be called with import imp_lock held.
2614 * When reference count drops to zero, request is freed.
2616 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2618 assert_spin_locked(&request->rq_import->imp_lock);
2619 (void)__ptlrpc_req_finished(request, 1);
2624 * Drops one reference count for request \a request.
2625 * \a locked set indicates that caller holds import imp_lock.
2626 * Frees the request whe reference count reaches zero.
2628 * \retval 1 the request is freed
2629 * \retval 0 some others still hold references on the request
2631 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2639 LASSERT(request != LP_POISON);
2640 LASSERT(request->rq_reqmsg != LP_POISON);
2642 DEBUG_REQ(D_INFO, request, "refcount now %u",
2643 atomic_read(&request->rq_refcount) - 1);
2645 spin_lock(&request->rq_lock);
2646 count = atomic_dec_return(&request->rq_refcount);
2647 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2650 * For open RPC, the client does not know the EA size (LOV, ACL, and
2651 * so on) before replied, then the client has to reserve very large
2652 * reply buffer. Such buffer will not be released until the RPC freed.
2653 * Since The open RPC is replayable, we need to keep it in the replay
2654 * list until close. If there are a lot of files opened concurrently,
2655 * then the client may be OOM.
2657 * If fact, it is unnecessary to keep reply buffer for open replay,
2658 * related EAs have already been saved via mdc_save_lovea() before
2659 * coming here. So it is safe to free the reply buffer some earlier
2660 * before releasing the RPC to avoid client OOM. LU-9514
2662 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2663 spin_lock(&request->rq_early_free_lock);
2664 sptlrpc_cli_free_repbuf(request);
2665 request->rq_repbuf = NULL;
2666 request->rq_repbuf_len = 0;
2667 request->rq_repdata = NULL;
2668 request->rq_reqdata_len = 0;
2669 spin_unlock(&request->rq_early_free_lock);
2671 spin_unlock(&request->rq_lock);
2674 __ptlrpc_free_req(request, locked);
2680 * Drops one reference count for a request.
2682 void ptlrpc_req_finished(struct ptlrpc_request *request)
2684 __ptlrpc_req_finished(request, 0);
2686 EXPORT_SYMBOL(ptlrpc_req_finished);
2689 * Returns xid of a \a request
2691 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2693 return request->rq_xid;
2695 EXPORT_SYMBOL(ptlrpc_req_xid);
2698 * Disengage the client's reply buffer from the network
2699 * NB does _NOT_ unregister any client-side bulk.
2700 * IDEMPOTENT, but _not_ safe against concurrent callers.
2701 * The request owner (i.e. the thread doing the I/O) must call...
2702 * Returns 0 on success or 1 if unregistering cannot be made.
2704 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2707 struct l_wait_info lwi;
2712 LASSERT(!in_interrupt());
2714 /* Let's setup deadline for reply unlink. */
2715 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2716 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2717 request->rq_reply_deadline = ktime_get_real_seconds() +
2721 * Nothing left to do.
2723 if (!ptlrpc_client_recv_or_unlink(request))
2726 LNetMDUnlink(request->rq_reply_md_h);
2729 * Let's check it once again.
2731 if (!ptlrpc_client_recv_or_unlink(request))
2734 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2735 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2738 * Do not wait for unlink to finish.
2744 * We have to l_wait_event() whatever the result, to give liblustre
2745 * a chance to run reply_in_callback(), and to make sure we've
2746 * unlinked before returning a req to the pool.
2749 /* The wq argument is ignored by user-space wait_event macros */
2750 wait_queue_head_t *wq = (request->rq_set) ?
2751 &request->rq_set->set_waitq :
2752 &request->rq_reply_waitq;
2754 * Network access will complete in finite time but the HUGE
2755 * timeout lets us CWARN for visibility of sluggish NALs
2757 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2758 cfs_time_seconds(1), NULL, NULL);
2759 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2762 ptlrpc_rqphase_move(request, request->rq_next_phase);
2766 LASSERT(rc == -ETIMEDOUT);
2767 DEBUG_REQ(D_WARNING, request,
2768 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2769 request->rq_receiving_reply,
2770 request->rq_req_unlinked,
2771 request->rq_reply_unlinked);
2776 static void ptlrpc_free_request(struct ptlrpc_request *req)
2778 spin_lock(&req->rq_lock);
2780 spin_unlock(&req->rq_lock);
2782 if (req->rq_commit_cb)
2783 req->rq_commit_cb(req);
2784 list_del_init(&req->rq_replay_list);
2786 __ptlrpc_req_finished(req, 1);
2790 * the request is committed and dropped from the replay list of its import
2792 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2794 struct obd_import *imp = req->rq_import;
2796 spin_lock(&imp->imp_lock);
2797 if (list_empty(&req->rq_replay_list)) {
2798 spin_unlock(&imp->imp_lock);
2802 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2803 if (imp->imp_replay_cursor == &req->rq_replay_list)
2804 imp->imp_replay_cursor = req->rq_replay_list.next;
2805 ptlrpc_free_request(req);
2808 spin_unlock(&imp->imp_lock);
2810 EXPORT_SYMBOL(ptlrpc_request_committed);
2813 * Iterates through replay_list on import and prunes
2814 * all requests have transno smaller than last_committed for the
2815 * import and don't have rq_replay set.
2816 * Since requests are sorted in transno order, stops when meetign first
2817 * transno bigger than last_committed.
2818 * caller must hold imp->imp_lock
2820 void ptlrpc_free_committed(struct obd_import *imp)
2822 struct ptlrpc_request *req, *saved;
2823 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2824 bool skip_committed_list = true;
2827 LASSERT(imp != NULL);
2828 assert_spin_locked(&imp->imp_lock);
2830 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2831 imp->imp_generation == imp->imp_last_generation_checked) {
2832 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2833 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2836 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2837 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2838 imp->imp_generation);
2840 if (imp->imp_generation != imp->imp_last_generation_checked ||
2841 imp->imp_last_transno_checked == 0)
2842 skip_committed_list = false;
2844 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2845 imp->imp_last_generation_checked = imp->imp_generation;
2847 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2849 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2850 LASSERT(req != last_req);
2853 if (req->rq_transno == 0) {
2854 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2857 if (req->rq_import_generation < imp->imp_generation) {
2858 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2862 /* not yet committed */
2863 if (req->rq_transno > imp->imp_peer_committed_transno) {
2864 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2868 if (req->rq_replay) {
2869 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2870 list_move_tail(&req->rq_replay_list,
2871 &imp->imp_committed_list);
2875 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2876 imp->imp_peer_committed_transno);
2878 ptlrpc_free_request(req);
2881 if (skip_committed_list)
2884 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2886 LASSERT(req->rq_transno != 0);
2887 if (req->rq_import_generation < imp->imp_generation ||
2889 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2890 req->rq_import_generation <
2891 imp->imp_generation ? "stale" : "closed");
2893 if (imp->imp_replay_cursor == &req->rq_replay_list)
2894 imp->imp_replay_cursor =
2895 req->rq_replay_list.next;
2897 ptlrpc_free_request(req);
2904 void ptlrpc_cleanup_client(struct obd_import *imp)
2911 * Schedule previously sent request for resend.
2912 * For bulk requests we assign new xid (to avoid problems with
2913 * lost replies and therefore several transfers landing into same buffer
2914 * from different sending attempts).
2916 void ptlrpc_resend_req(struct ptlrpc_request *req)
2918 DEBUG_REQ(D_HA, req, "going to resend");
2919 spin_lock(&req->rq_lock);
2922 * Request got reply but linked to the import list still.
2923 * Let ptlrpc_check_set() process it.
2925 if (ptlrpc_client_replied(req)) {
2926 spin_unlock(&req->rq_lock);
2927 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2931 req->rq_status = -EAGAIN;
2934 req->rq_net_err = 0;
2935 req->rq_timedout = 0;
2937 ptlrpc_client_wake_req(req);
2938 spin_unlock(&req->rq_lock);
2941 /* XXX: this function and rq_status are currently unused */
2942 void ptlrpc_restart_req(struct ptlrpc_request *req)
2944 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2945 req->rq_status = -ERESTARTSYS;
2947 spin_lock(&req->rq_lock);
2948 req->rq_restart = 1;
2949 req->rq_timedout = 0;
2950 ptlrpc_client_wake_req(req);
2951 spin_unlock(&req->rq_lock);
2955 * Grab additional reference on a request \a req
2957 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2960 atomic_inc(&req->rq_refcount);
2963 EXPORT_SYMBOL(ptlrpc_request_addref);
2966 * Add a request to import replay_list.
2967 * Must be called under imp_lock
2969 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2970 struct obd_import *imp)
2972 struct list_head *tmp;
2974 assert_spin_locked(&imp->imp_lock);
2976 if (req->rq_transno == 0) {
2977 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2982 * clear this for new requests that were resent as well
2983 * as resent replayed requests.
2985 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2987 /* don't re-add requests that have been replayed */
2988 if (!list_empty(&req->rq_replay_list))
2991 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2993 spin_lock(&req->rq_lock);
2995 spin_unlock(&req->rq_lock);
2997 LASSERT(imp->imp_replayable);
2998 /* Balanced in ptlrpc_free_committed, usually. */
2999 ptlrpc_request_addref(req);
3000 list_for_each_prev(tmp, &imp->imp_replay_list) {
3001 struct ptlrpc_request *iter = list_entry(tmp,
3002 struct ptlrpc_request,
3006 * We may have duplicate transnos if we create and then
3007 * open a file, or for closes retained if to match creating
3008 * opens, so use req->rq_xid as a secondary key.
3009 * (See bugs 684, 685, and 428.)
3010 * XXX no longer needed, but all opens need transnos!
3012 if (iter->rq_transno > req->rq_transno)
3015 if (iter->rq_transno == req->rq_transno) {
3016 LASSERT(iter->rq_xid != req->rq_xid);
3017 if (iter->rq_xid > req->rq_xid)
3021 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3025 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3029 * Send request and wait until it completes.
3030 * Returns request processing status.
3032 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3034 struct ptlrpc_request_set *set;
3038 LASSERT(req->rq_set == NULL);
3039 LASSERT(!req->rq_receiving_reply);
3041 set = ptlrpc_prep_set();
3043 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3047 /* for distributed debugging */
3048 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3050 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3051 ptlrpc_request_addref(req);
3052 ptlrpc_set_add_req(set, req);
3053 rc = ptlrpc_set_wait(NULL, set);
3054 ptlrpc_set_destroy(set);
3058 EXPORT_SYMBOL(ptlrpc_queue_wait);
3061 * Callback used for replayed requests reply processing.
3062 * In case of successful reply calls registered request replay callback.
3063 * In case of error restart replay process.
3065 static int ptlrpc_replay_interpret(const struct lu_env *env,
3066 struct ptlrpc_request *req,
3069 struct ptlrpc_replay_async_args *aa = args;
3070 struct obd_import *imp = req->rq_import;
3073 atomic_dec(&imp->imp_replay_inflight);
3076 * Note: if it is bulk replay (MDS-MDS replay), then even if
3077 * server got the request, but bulk transfer timeout, let's
3078 * replay the bulk req again
3080 if (!ptlrpc_client_replied(req) ||
3082 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3083 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
3084 GOTO(out, rc = -ETIMEDOUT);
3087 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3088 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3089 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3090 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3092 /** VBR: check version failure */
3093 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3094 /** replay was failed due to version mismatch */
3095 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
3096 spin_lock(&imp->imp_lock);
3097 imp->imp_vbr_failed = 1;
3098 spin_unlock(&imp->imp_lock);
3099 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3101 /** The transno had better not change over replay. */
3102 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3103 lustre_msg_get_transno(req->rq_repmsg) ||
3104 lustre_msg_get_transno(req->rq_repmsg) == 0,
3106 lustre_msg_get_transno(req->rq_reqmsg),
3107 lustre_msg_get_transno(req->rq_repmsg));
3110 spin_lock(&imp->imp_lock);
3111 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3112 spin_unlock(&imp->imp_lock);
3113 LASSERT(imp->imp_last_replay_transno);
3115 /* transaction number shouldn't be bigger than the latest replayed */
3116 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3117 DEBUG_REQ(D_ERROR, req,
3118 "Reported transno %llu is bigger than the replayed one: %llu",
3120 lustre_msg_get_transno(req->rq_reqmsg));
3121 GOTO(out, rc = -EINVAL);
3124 DEBUG_REQ(D_HA, req, "got rep");
3126 /* let the callback do fixups, possibly including in the request */
3127 if (req->rq_replay_cb)
3128 req->rq_replay_cb(req);
3130 if (ptlrpc_client_replied(req) &&
3131 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3132 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3133 lustre_msg_get_status(req->rq_repmsg),
3134 aa->praa_old_status);
3137 * Note: If the replay fails for MDT-MDT recovery, let's
3138 * abort all of the following requests in the replay
3139 * and sending list, because MDT-MDT update requests
3140 * are dependent on each other, see LU-7039
3142 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3143 struct ptlrpc_request *free_req;
3144 struct ptlrpc_request *tmp;
3146 spin_lock(&imp->imp_lock);
3147 list_for_each_entry_safe(free_req, tmp,
3148 &imp->imp_replay_list,
3150 ptlrpc_free_request(free_req);
3153 list_for_each_entry_safe(free_req, tmp,
3154 &imp->imp_committed_list,
3156 ptlrpc_free_request(free_req);
3159 list_for_each_entry_safe(free_req, tmp,
3160 &imp->imp_delayed_list,
3162 spin_lock(&free_req->rq_lock);
3163 free_req->rq_err = 1;
3164 free_req->rq_status = -EIO;
3165 ptlrpc_client_wake_req(free_req);
3166 spin_unlock(&free_req->rq_lock);
3169 list_for_each_entry_safe(free_req, tmp,
3170 &imp->imp_sending_list,
3172 spin_lock(&free_req->rq_lock);
3173 free_req->rq_err = 1;
3174 free_req->rq_status = -EIO;
3175 ptlrpc_client_wake_req(free_req);
3176 spin_unlock(&free_req->rq_lock);
3178 spin_unlock(&imp->imp_lock);
3181 /* Put it back for re-replay. */
3182 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3186 * Errors while replay can set transno to 0, but
3187 * imp_last_replay_transno shouldn't be set to 0 anyway
3189 if (req->rq_transno == 0)
3190 CERROR("Transno is 0 during replay!\n");
3192 /* continue with recovery */
3193 rc = ptlrpc_import_recovery_state_machine(imp);
3195 req->rq_send_state = aa->praa_old_state;
3198 /* this replay failed, so restart recovery */
3199 ptlrpc_connect_import(imp);
3205 * Prepares and queues request for replay.
3206 * Adds it to ptlrpcd queue for actual sending.
3207 * Returns 0 on success.
3209 int ptlrpc_replay_req(struct ptlrpc_request *req)
3211 struct ptlrpc_replay_async_args *aa;
3215 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3217 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3218 aa = ptlrpc_req_async_args(req);
3219 memset(aa, 0, sizeof(*aa));
3221 /* Prepare request to be resent with ptlrpcd */
3222 aa->praa_old_state = req->rq_send_state;
3223 req->rq_send_state = LUSTRE_IMP_REPLAY;
3224 req->rq_phase = RQ_PHASE_NEW;
3225 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3227 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3229 req->rq_interpret_reply = ptlrpc_replay_interpret;
3230 /* Readjust the timeout for current conditions */
3231 ptlrpc_at_set_req_timeout(req);
3233 /* Tell server net_latency to calculate how long to wait for reply. */
3234 lustre_msg_set_service_time(req->rq_reqmsg,
3235 ptlrpc_at_get_net_latency(req));
3236 DEBUG_REQ(D_HA, req, "REPLAY");
3238 atomic_inc(&req->rq_import->imp_replay_inflight);
3239 spin_lock(&req->rq_lock);
3240 req->rq_early_free_repbuf = 0;
3241 spin_unlock(&req->rq_lock);
3242 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3244 ptlrpcd_add_req(req);
3249 * Aborts all in-flight request on import \a imp sending and delayed lists
3251 void ptlrpc_abort_inflight(struct obd_import *imp)
3253 struct list_head *tmp, *n;
3257 * Make sure that no new requests get processed for this import.
3258 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3259 * this flag and then putting requests on sending_list or delayed_list.
3261 spin_lock(&imp->imp_lock);
3264 * XXX locking? Maybe we should remove each request with the list
3265 * locked? Also, how do we know if the requests on the list are
3266 * being freed at this time?
3268 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3269 struct ptlrpc_request *req = list_entry(tmp,
3270 struct ptlrpc_request,
3273 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3275 spin_lock(&req->rq_lock);
3276 if (req->rq_import_generation < imp->imp_generation) {
3278 req->rq_status = -EIO;
3279 ptlrpc_client_wake_req(req);
3281 spin_unlock(&req->rq_lock);
3284 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3285 struct ptlrpc_request *req =
3286 list_entry(tmp, struct ptlrpc_request, rq_list);
3288 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3290 spin_lock(&req->rq_lock);
3291 if (req->rq_import_generation < imp->imp_generation) {
3293 req->rq_status = -EIO;
3294 ptlrpc_client_wake_req(req);
3296 spin_unlock(&req->rq_lock);
3300 * Last chance to free reqs left on the replay list, but we
3301 * will still leak reqs that haven't committed.
3303 if (imp->imp_replayable)
3304 ptlrpc_free_committed(imp);
3306 spin_unlock(&imp->imp_lock);
3312 * Abort all uncompleted requests in request set \a set
3314 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3316 struct list_head *tmp, *pos;
3318 LASSERT(set != NULL);
3320 list_for_each_safe(pos, tmp, &set->set_requests) {
3321 struct ptlrpc_request *req =
3322 list_entry(pos, struct ptlrpc_request,
3325 spin_lock(&req->rq_lock);
3326 if (req->rq_phase != RQ_PHASE_RPC) {
3327 spin_unlock(&req->rq_lock);
3332 req->rq_status = -EINTR;
3333 ptlrpc_client_wake_req(req);
3334 spin_unlock(&req->rq_lock);
3339 * Initialize the XID for the node. This is common among all requests on
3340 * this node, and only requires the property that it is monotonically
3341 * increasing. It does not need to be sequential. Since this is also used
3342 * as the RDMA match bits, it is important that a single client NOT have
3343 * the same match bits for two different in-flight requests, hence we do
3344 * NOT want to have an XID per target or similar.
3346 * To avoid an unlikely collision between match bits after a client reboot
3347 * (which would deliver old data into the wrong RDMA buffer) initialize
3348 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3349 * If the time is clearly incorrect, we instead use a 62-bit random number.
3350 * In the worst case the random number will overflow 1M RPCs per second in
3351 * 9133 years, or permutations thereof.
3353 #define YEAR_2004 (1ULL << 30)
3354 void ptlrpc_init_xid(void)
3356 time64_t now = ktime_get_real_seconds();
3359 if (now < YEAR_2004) {
3360 get_random_bytes(&xid, sizeof(xid));
3362 xid |= (1ULL << 61);
3364 xid = (u64)now << 20;
3367 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3368 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3369 xid &= PTLRPC_BULK_OPS_MASK;
3370 atomic64_set(&ptlrpc_last_xid, xid);
3374 * Increase xid and returns resulting new value to the caller.
3376 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3377 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3378 * itself uses the last bulk xid needed, so the server can determine the
3379 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3380 * xid must align to a power-of-two value.
3382 * This is assumed to be true due to the initial ptlrpc_last_xid
3383 * value also being initialized to a power-of-two value. LU-1431
3385 __u64 ptlrpc_next_xid(void)
3387 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3391 * If request has a new allocated XID (new request or EINPROGRESS resend),
3392 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3393 * request to ensure previous bulk fails and avoid problems with lost replies
3394 * and therefore several transfers landing into the same buffer from different
3397 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3399 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3401 LASSERT(bd != NULL);
3404 * Generate new matchbits for all resend requests, including
3407 if (req->rq_resend) {
3408 __u64 old_mbits = req->rq_mbits;
3411 * First time resend on -EINPROGRESS will generate new xid,
3412 * so we can actually use the rq_xid as rq_mbits in such case,
3413 * however, it's bit hard to distinguish such resend with a
3414 * 'resend for the -EINPROGRESS resend'. To make it simple,
3415 * we opt to generate mbits for all resend cases.
3417 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3419 req->rq_mbits = ptlrpc_next_xid();
3422 * Old version transfers rq_xid to peer as
3425 spin_lock(&req->rq_import->imp_lock);
3426 list_del_init(&req->rq_unreplied_list);
3427 ptlrpc_assign_next_xid_nolock(req);
3428 spin_unlock(&req->rq_import->imp_lock);
3429 req->rq_mbits = req->rq_xid;
3431 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3432 old_mbits, req->rq_mbits);
3433 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3434 /* Request being sent first time, use xid as matchbits. */
3435 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3436 || req->rq_mbits == 0) {
3437 req->rq_mbits = req->rq_xid;
3439 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3441 req->rq_mbits -= total_md - 1;
3445 * Replay request, xid and matchbits have already been
3446 * correctly assigned.
3452 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3453 * that server can infer the number of bulks that were prepared,
3456 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3460 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3461 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3463 * It's ok to directly set the rq_xid here, since this xid bump
3464 * won't affect the request position in unreplied list.
3466 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3467 req->rq_xid = req->rq_mbits;
3471 * Get a glimpse at what next xid value might have been.
3472 * Returns possible next xid.
3474 __u64 ptlrpc_sample_next_xid(void)
3476 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3478 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3481 * Functions for operating ptlrpc workers.
3483 * A ptlrpc work is a function which will be running inside ptlrpc context.
3484 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3486 * 1. after a work is created, it can be used many times, that is:
3487 * handler = ptlrpcd_alloc_work();
3488 * ptlrpcd_queue_work();
3490 * queue it again when necessary:
3491 * ptlrpcd_queue_work();
3492 * ptlrpcd_destroy_work();
3493 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3494 * it will only be queued once in any time. Also as its name implies, it may
3495 * have delay before it really runs by ptlrpcd thread.
3497 struct ptlrpc_work_async_args {
3498 int (*cb)(const struct lu_env *, void *);
3502 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3504 /* re-initialize the req */
3505 req->rq_timeout = obd_timeout;
3506 req->rq_sent = ktime_get_real_seconds();
3507 req->rq_deadline = req->rq_sent + req->rq_timeout;
3508 req->rq_phase = RQ_PHASE_INTERPRET;
3509 req->rq_next_phase = RQ_PHASE_COMPLETE;
3510 req->rq_xid = ptlrpc_next_xid();
3511 req->rq_import_generation = req->rq_import->imp_generation;
3513 ptlrpcd_add_req(req);
3516 static int work_interpreter(const struct lu_env *env,
3517 struct ptlrpc_request *req, void *args, int rc)
3519 struct ptlrpc_work_async_args *arg = args;
3521 LASSERT(ptlrpcd_check_work(req));
3522 LASSERT(arg->cb != NULL);
3524 rc = arg->cb(env, arg->cbdata);
3526 list_del_init(&req->rq_set_chain);
3529 if (atomic_dec_return(&req->rq_refcount) > 1) {
3530 atomic_set(&req->rq_refcount, 2);
3531 ptlrpcd_add_work_req(req);
3536 static int worker_format;
3538 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3540 return req->rq_pill.rc_fmt == (void *)&worker_format;
3544 * Create a work for ptlrpc.
3546 void *ptlrpcd_alloc_work(struct obd_import *imp,
3547 int (*cb)(const struct lu_env *, void *), void *cbdata)
3549 struct ptlrpc_request *req = NULL;
3550 struct ptlrpc_work_async_args *args;
3556 RETURN(ERR_PTR(-EINVAL));
3558 /* copy some code from deprecated fakereq. */
3559 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3561 CERROR("ptlrpc: run out of memory!\n");
3562 RETURN(ERR_PTR(-ENOMEM));
3565 ptlrpc_cli_req_init(req);
3567 req->rq_send_state = LUSTRE_IMP_FULL;
3568 req->rq_type = PTL_RPC_MSG_REQUEST;
3569 req->rq_import = class_import_get(imp);
3570 req->rq_interpret_reply = work_interpreter;
3571 /* don't want reply */
3572 req->rq_no_delay = req->rq_no_resend = 1;
3573 req->rq_pill.rc_fmt = (void *)&worker_format;
3575 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3576 args = ptlrpc_req_async_args(req);
3578 args->cbdata = cbdata;
3582 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3584 void ptlrpcd_destroy_work(void *handler)
3586 struct ptlrpc_request *req = handler;
3589 ptlrpc_req_finished(req);
3591 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3593 int ptlrpcd_queue_work(void *handler)
3595 struct ptlrpc_request *req = handler;
3598 * Check if the req is already being queued.
3600 * Here comes a trick: it lacks a way of checking if a req is being
3601 * processed reliably in ptlrpc. Here I have to use refcount of req
3602 * for this purpose. This is okay because the caller should use this
3603 * req as opaque data. - Jinshan
3605 LASSERT(atomic_read(&req->rq_refcount) > 0);
3606 if (atomic_inc_return(&req->rq_refcount) == 2)
3607 ptlrpcd_add_work_req(req);
3610 EXPORT_SYMBOL(ptlrpcd_queue_work);