4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 /** Implementation of client-side PortalRPC interfaces */
35 #define DEBUG_SUBSYSTEM S_RPC
37 #include <linux/delay.h>
38 #include <linux/random.h>
40 #include <lnet/lib-lnet.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
48 #include "ptlrpc_internal.h"
50 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
51 struct page *page, int pageoffset,
54 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
57 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
58 struct page *page, int pageoffset,
61 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
64 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
68 for (i = 0; i < desc->bd_iov_count ; i++)
69 put_page(BD_GET_KIOV(desc, i).kiov_page);
72 static int ptlrpc_prep_bulk_frag_pages(struct ptlrpc_bulk_desc *desc,
75 unsigned int offset = (unsigned long)frag & ~PAGE_MASK;
79 int page_len = min_t(unsigned int, PAGE_SIZE - offset,
81 unsigned long vaddr = (unsigned long)frag;
83 ptlrpc_prep_bulk_page_nopin(desc,
84 lnet_kvaddr_to_page(vaddr),
94 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
95 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
96 .release_frags = ptlrpc_release_bulk_page_pin,
98 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
100 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
101 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
102 .release_frags = ptlrpc_release_bulk_noop,
103 .add_iov_frag = ptlrpc_prep_bulk_frag_pages,
105 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
107 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
108 static int ptlrpcd_check_work(struct ptlrpc_request *req);
109 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
112 * Initialize passed in client structure \a cl.
114 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
115 struct ptlrpc_client *cl)
117 cl->cli_request_portal = req_portal;
118 cl->cli_reply_portal = rep_portal;
121 EXPORT_SYMBOL(ptlrpc_init_client);
124 * Return PortalRPC connection for remore uud \a uuid
126 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
127 lnet_nid_t nid4refnet)
129 struct ptlrpc_connection *c;
131 struct lnet_process_id peer;
135 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
136 * before accessing its values.
138 /* coverity[uninit_use_in_call] */
139 peer.nid = nid4refnet;
140 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
142 CNETERR("cannot find peer %s!\n", uuid->uuid);
146 c = ptlrpc_connection_get(peer, self, uuid);
148 memcpy(c->c_remote_uuid.uuid,
149 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
152 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
158 * Allocate and initialize new bulk descriptor on the sender.
159 * Returns pointer to the descriptor or NULL on error.
161 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
162 unsigned int max_brw,
163 enum ptlrpc_bulk_op_type type,
165 const struct ptlrpc_bulk_frag_ops *ops)
167 struct ptlrpc_bulk_desc *desc;
170 LASSERT(ops->add_kiov_frag != NULL);
176 OBD_ALLOC_LARGE(GET_KIOV(desc),
177 nfrags * sizeof(*GET_KIOV(desc)));
181 spin_lock_init(&desc->bd_lock);
182 init_waitqueue_head(&desc->bd_waitq);
183 desc->bd_max_iov = nfrags;
184 desc->bd_iov_count = 0;
185 desc->bd_portal = portal;
186 desc->bd_type = type;
187 desc->bd_md_count = 0;
188 desc->bd_frag_ops = ops;
189 LASSERT(max_brw > 0);
190 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
192 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
193 * node. Negotiated ocd_brw_size will always be <= this number.
195 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
196 LNetInvalidateMDHandle(&desc->bd_mds[i]);
205 * Prepare bulk descriptor for specified outgoing request \a req that
206 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
207 * the bulk to be sent. Used on client-side.
208 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
211 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
213 unsigned int max_brw,
216 const struct ptlrpc_bulk_frag_ops
219 struct obd_import *imp = req->rq_import;
220 struct ptlrpc_bulk_desc *desc;
223 LASSERT(ptlrpc_is_bulk_op_passive(type));
225 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
229 desc->bd_import = class_import_get(imp);
232 desc->bd_cbid.cbid_fn = client_bulk_callback;
233 desc->bd_cbid.cbid_arg = desc;
235 /* This makes req own desc, and free it when she frees herself */
240 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
242 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
243 struct page *page, int pageoffset, int len,
248 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
249 LASSERT(page != NULL);
250 LASSERT(pageoffset >= 0);
252 LASSERT(pageoffset + len <= PAGE_SIZE);
254 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
261 kiov->kiov_page = page;
262 kiov->kiov_offset = pageoffset;
263 kiov->kiov_len = len;
265 desc->bd_iov_count++;
267 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
269 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
273 LASSERT(desc != NULL);
274 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
275 LASSERT(desc->bd_md_count == 0); /* network hands off */
276 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
277 LASSERT(desc->bd_frag_ops != NULL);
279 sptlrpc_enc_pool_put_pages(desc);
282 class_export_put(desc->bd_export);
284 class_import_put(desc->bd_import);
286 if (desc->bd_frag_ops->release_frags != NULL)
287 desc->bd_frag_ops->release_frags(desc);
289 OBD_FREE_LARGE(GET_KIOV(desc),
290 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
294 EXPORT_SYMBOL(ptlrpc_free_bulk);
297 * Set server timelimit for this req, i.e. how long are we willing to wait
298 * for reply before timing out this request.
300 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
306 LASSERT(req->rq_import);
309 /* non-AT settings */
311 * \a imp_server_timeout means this is reverse import and
312 * we send (currently only) ASTs to the client and cannot afford
313 * to wait too long for the reply, otherwise the other client
314 * (because of which we are sending this request) would
315 * timeout waiting for us
317 req->rq_timeout = req->rq_import->imp_server_timeout ?
318 obd_timeout / 2 : obd_timeout;
320 at = &req->rq_import->imp_at;
321 idx = import_at_get_index(req->rq_import,
322 req->rq_request_portal);
323 serv_est = at_get(&at->iat_service_estimate[idx]);
324 req->rq_timeout = at_est2timeout(serv_est);
327 * We could get even fancier here, using history to predict increased
332 * Let the server know what this RPC timeout is by putting it in the
335 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
337 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
339 /* Adjust max service estimate based on server value */
340 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
341 unsigned int serv_est)
347 LASSERT(req->rq_import);
348 at = &req->rq_import->imp_at;
350 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
352 * max service estimates are tracked on the server side,
353 * so just keep minimal history here
355 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
358 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
359 req->rq_import->imp_obd->obd_name,
360 req->rq_request_portal,
361 oldse, at_get(&at->iat_service_estimate[idx]));
364 /* Expected network latency per remote node (secs) */
365 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
367 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
370 /* Adjust expected network latency */
371 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
372 unsigned int service_time)
374 unsigned int nl, oldnl;
376 time64_t now = ktime_get_real_seconds();
378 LASSERT(req->rq_import);
380 if (service_time > now - req->rq_sent + 3) {
382 * b=16408, however, this can also happen if early reply
383 * is lost and client RPC is expired and resent, early reply
384 * or reply of original RPC can still be fit in reply buffer
385 * of resent RPC, now client is measuring time from the
386 * resent time, but server sent back service time of original
389 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
390 D_ADAPTTO : D_WARNING,
391 "Reported service time %u > total measured time %lld\n",
392 service_time, now - req->rq_sent);
396 /* Network latency is total time less server processing time */
397 nl = max_t(int, now - req->rq_sent -
398 service_time, 0) + 1; /* st rounding */
399 at = &req->rq_import->imp_at;
401 oldnl = at_measured(&at->iat_net_latency, nl);
404 "The network latency for %s (nid %s) has changed from %d to %d\n",
405 req->rq_import->imp_obd->obd_name,
406 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
407 oldnl, at_get(&at->iat_net_latency));
410 static int unpack_reply(struct ptlrpc_request *req)
414 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
415 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
417 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
423 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
425 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
433 * Handle an early reply message, called with the rq_lock held.
434 * If anything goes wrong just ignore it - same as if it never happened
436 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
437 __must_hold(&req->rq_lock)
439 struct ptlrpc_request *early_req;
445 spin_unlock(&req->rq_lock);
447 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
449 spin_lock(&req->rq_lock);
453 rc = unpack_reply(early_req);
455 sptlrpc_cli_finish_early_reply(early_req);
456 spin_lock(&req->rq_lock);
461 * Use new timeout value just to adjust the local value for this
462 * request, don't include it into at_history. It is unclear yet why
463 * service time increased and should it be counted or skipped, e.g.
464 * that can be recovery case or some error or server, the real reply
465 * will add all new data if it is worth to add.
467 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
468 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
470 /* Network latency can be adjusted, it is pure network delays */
471 ptlrpc_at_adj_net_latency(req,
472 lustre_msg_get_service_time(early_req->rq_repmsg));
474 sptlrpc_cli_finish_early_reply(early_req);
476 spin_lock(&req->rq_lock);
477 olddl = req->rq_deadline;
479 * server assumes it now has rq_timeout from when the request
480 * arrived, so the client should give it at least that long.
481 * since we don't know the arrival time we'll use the original
484 req->rq_deadline = req->rq_sent + req->rq_timeout +
485 ptlrpc_at_get_net_latency(req);
487 /* The below message is checked in replay-single.sh test_65{a,b} */
488 /* The below message is checked in sanity-{gss,krb5} test_8 */
489 DEBUG_REQ(D_ADAPTTO, req,
490 "Early reply #%d, new deadline in %llds (%llds)",
492 req->rq_deadline - ktime_get_real_seconds(),
493 req->rq_deadline - olddl);
498 static struct kmem_cache *request_cache;
500 int ptlrpc_request_cache_init(void)
502 request_cache = kmem_cache_create("ptlrpc_cache",
503 sizeof(struct ptlrpc_request),
504 0, SLAB_HWCACHE_ALIGN, NULL);
505 return request_cache ? 0 : -ENOMEM;
508 void ptlrpc_request_cache_fini(void)
510 kmem_cache_destroy(request_cache);
513 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
515 struct ptlrpc_request *req;
517 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
521 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
523 OBD_SLAB_FREE_PTR(req, request_cache);
527 * Wind down request pool \a pool.
528 * Frees all requests from the pool too
530 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
532 struct list_head *l, *tmp;
533 struct ptlrpc_request *req;
535 LASSERT(pool != NULL);
537 spin_lock(&pool->prp_lock);
538 list_for_each_safe(l, tmp, &pool->prp_req_list) {
539 req = list_entry(l, struct ptlrpc_request, rq_list);
540 list_del(&req->rq_list);
541 LASSERT(req->rq_reqbuf);
542 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
543 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
544 ptlrpc_request_cache_free(req);
546 spin_unlock(&pool->prp_lock);
547 OBD_FREE(pool, sizeof(*pool));
549 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
552 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
554 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
559 while (size < pool->prp_rq_size)
562 LASSERTF(list_empty(&pool->prp_req_list) ||
563 size == pool->prp_rq_size,
564 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
565 pool->prp_rq_size, size);
567 pool->prp_rq_size = size;
568 for (i = 0; i < num_rq; i++) {
569 struct ptlrpc_request *req;
570 struct lustre_msg *msg;
572 req = ptlrpc_request_cache_alloc(GFP_NOFS);
575 OBD_ALLOC_LARGE(msg, size);
577 ptlrpc_request_cache_free(req);
580 req->rq_reqbuf = msg;
581 req->rq_reqbuf_len = size;
583 spin_lock(&pool->prp_lock);
584 list_add_tail(&req->rq_list, &pool->prp_req_list);
585 spin_unlock(&pool->prp_lock);
589 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
592 * Create and initialize new request pool with given attributes:
593 * \a num_rq - initial number of requests to create for the pool
594 * \a msgsize - maximum message size possible for requests in thid pool
595 * \a populate_pool - function to be called when more requests need to be added
597 * Returns pointer to newly created pool or NULL on error.
599 struct ptlrpc_request_pool *
600 ptlrpc_init_rq_pool(int num_rq, int msgsize,
601 int (*populate_pool)(struct ptlrpc_request_pool *, int))
603 struct ptlrpc_request_pool *pool;
610 * Request next power of two for the allocation, because internally
611 * kernel would do exactly this
613 spin_lock_init(&pool->prp_lock);
614 INIT_LIST_HEAD(&pool->prp_req_list);
615 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
616 pool->prp_populate = populate_pool;
618 populate_pool(pool, num_rq);
622 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
625 * Fetches one request from pool \a pool
627 static struct ptlrpc_request *
628 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
630 struct ptlrpc_request *request;
631 struct lustre_msg *reqbuf;
636 spin_lock(&pool->prp_lock);
639 * See if we have anything in a pool, and bail out if nothing,
640 * in writeout path, where this matters, this is safe to do, because
641 * nothing is lost in this case, and when some in-flight requests
642 * complete, this code will be called again.
644 if (unlikely(list_empty(&pool->prp_req_list))) {
645 spin_unlock(&pool->prp_lock);
649 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
651 list_del_init(&request->rq_list);
652 spin_unlock(&pool->prp_lock);
654 LASSERT(request->rq_reqbuf);
655 LASSERT(request->rq_pool);
657 reqbuf = request->rq_reqbuf;
658 memset(request, 0, sizeof(*request));
659 request->rq_reqbuf = reqbuf;
660 request->rq_reqbuf_len = pool->prp_rq_size;
661 request->rq_pool = pool;
667 * Returns freed \a request to pool.
669 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
671 struct ptlrpc_request_pool *pool = request->rq_pool;
673 spin_lock(&pool->prp_lock);
674 LASSERT(list_empty(&request->rq_list));
675 LASSERT(!request->rq_receiving_reply);
676 list_add_tail(&request->rq_list, &pool->prp_req_list);
677 spin_unlock(&pool->prp_lock);
680 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
682 struct obd_import *imp = req->rq_import;
683 struct list_head *tmp;
684 struct ptlrpc_request *iter;
686 assert_spin_locked(&imp->imp_lock);
687 LASSERT(list_empty(&req->rq_unreplied_list));
689 /* unreplied list is sorted by xid in ascending order */
690 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
691 iter = list_entry(tmp, struct ptlrpc_request,
694 LASSERT(req->rq_xid != iter->rq_xid);
695 if (req->rq_xid < iter->rq_xid)
697 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
700 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
703 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
705 req->rq_xid = ptlrpc_next_xid();
706 ptlrpc_add_unreplied(req);
709 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
711 spin_lock(&req->rq_import->imp_lock);
712 ptlrpc_assign_next_xid_nolock(req);
713 spin_unlock(&req->rq_import->imp_lock);
716 static atomic64_t ptlrpc_last_xid;
718 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
720 spin_lock(&req->rq_import->imp_lock);
721 list_del_init(&req->rq_unreplied_list);
722 ptlrpc_assign_next_xid_nolock(req);
723 spin_unlock(&req->rq_import->imp_lock);
724 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
727 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
729 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
733 opc = lustre_msg_get_opc(req->rq_reqmsg);
734 tag = obd_get_mod_rpc_slot(cli, opc);
735 lustre_msg_set_tag(req->rq_reqmsg, tag);
736 ptlrpc_reassign_next_xid(req);
738 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
740 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
742 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
745 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
746 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
748 obd_put_mod_rpc_slot(cli, opc, tag);
751 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
753 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
754 __u32 version, int opcode, char **bufs,
755 struct ptlrpc_cli_ctx *ctx)
758 struct obd_import *imp;
764 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
765 imp = request->rq_import;
766 lengths = request->rq_pill.rc_area[RCL_CLIENT];
769 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
771 rc = sptlrpc_req_get_ctx(request);
775 sptlrpc_req_set_flavor(request, opcode);
777 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
782 lustre_msg_add_version(request->rq_reqmsg, version);
783 request->rq_send_state = LUSTRE_IMP_FULL;
784 request->rq_type = PTL_RPC_MSG_REQUEST;
786 request->rq_req_cbid.cbid_fn = request_out_callback;
787 request->rq_req_cbid.cbid_arg = request;
789 request->rq_reply_cbid.cbid_fn = reply_in_callback;
790 request->rq_reply_cbid.cbid_arg = request;
792 request->rq_reply_deadline = 0;
793 request->rq_bulk_deadline = 0;
794 request->rq_req_deadline = 0;
795 request->rq_phase = RQ_PHASE_NEW;
796 request->rq_next_phase = RQ_PHASE_UNDEFINED;
798 request->rq_request_portal = imp->imp_client->cli_request_portal;
799 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
801 ptlrpc_at_set_req_timeout(request);
803 lustre_msg_set_opc(request->rq_reqmsg, opcode);
805 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
806 if (cfs_fail_val == opcode) {
807 time64_t *fail_t = NULL, *fail2_t = NULL;
809 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
810 fail_t = &request->rq_bulk_deadline;
811 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
812 fail_t = &request->rq_reply_deadline;
813 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
814 fail_t = &request->rq_req_deadline;
815 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
816 fail_t = &request->rq_reply_deadline;
817 fail2_t = &request->rq_bulk_deadline;
818 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
819 time64_t now = ktime_get_real_seconds();
820 u64 xid = ((u64)now >> 4) << 24;
822 atomic64_set(&ptlrpc_last_xid, xid);
826 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
829 *fail2_t = ktime_get_real_seconds() +
833 * The RPC is infected, let the test to change the
836 msleep(4 * MSEC_PER_SEC);
839 ptlrpc_assign_next_xid(request);
844 LASSERT(!request->rq_pool);
845 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
847 class_import_put(imp);
851 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
854 * Pack request buffers for network transfer, performing necessary encryption
855 * steps if necessary.
857 int ptlrpc_request_pack(struct ptlrpc_request *request,
858 __u32 version, int opcode)
860 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
862 EXPORT_SYMBOL(ptlrpc_request_pack);
865 * Helper function to allocate new request on import \a imp
866 * and possibly using existing request from pool \a pool if provided.
867 * Returns allocated request structure with import field filled or
871 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
872 struct ptlrpc_request_pool *pool)
874 struct ptlrpc_request *request = NULL;
876 request = ptlrpc_request_cache_alloc(GFP_NOFS);
878 if (!request && pool)
879 request = ptlrpc_prep_req_from_pool(pool);
882 ptlrpc_cli_req_init(request);
884 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
885 LASSERT(imp != LP_POISON);
886 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
888 LASSERT(imp->imp_client != LP_POISON);
890 request->rq_import = class_import_get(imp);
892 CERROR("request allocation out of memory\n");
899 * Helper function for creating a request.
900 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
901 * buffer structures according to capsule template \a format.
902 * Returns allocated request structure pointer or NULL on error.
904 static struct ptlrpc_request *
905 ptlrpc_request_alloc_internal(struct obd_import *imp,
906 struct ptlrpc_request_pool *pool,
907 const struct req_format *format)
909 struct ptlrpc_request *request;
911 request = __ptlrpc_request_alloc(imp, pool);
916 * initiate connection if needed when the import has been
917 * referenced by the new request to avoid races with disconnect
919 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
922 CDEBUG_LIMIT(imp->imp_idle_debug,
923 "%s: reconnect after %llds idle\n",
924 imp->imp_obd->obd_name, ktime_get_real_seconds() -
925 imp->imp_last_reply_time);
926 spin_lock(&imp->imp_lock);
927 if (imp->imp_state == LUSTRE_IMP_IDLE) {
928 imp->imp_generation++;
929 imp->imp_initiated_at = imp->imp_generation;
930 imp->imp_state = LUSTRE_IMP_NEW;
932 /* connect_import_locked releases imp_lock */
933 rc = ptlrpc_connect_import_locked(imp);
935 ptlrpc_request_free(request);
938 ptlrpc_pinger_add_import(imp);
940 spin_unlock(&imp->imp_lock);
944 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
945 req_capsule_set(&request->rq_pill, format);
950 * Allocate new request structure for import \a imp and initialize its
951 * buffer structure according to capsule template \a format.
953 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
954 const struct req_format *format)
956 return ptlrpc_request_alloc_internal(imp, NULL, format);
958 EXPORT_SYMBOL(ptlrpc_request_alloc);
961 * Allocate new request structure for import \a imp from pool \a pool and
962 * initialize its buffer structure according to capsule template \a format.
964 struct ptlrpc_request *
965 ptlrpc_request_alloc_pool(struct obd_import *imp,
966 struct ptlrpc_request_pool *pool,
967 const struct req_format *format)
969 return ptlrpc_request_alloc_internal(imp, pool, format);
971 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
974 * For requests not from pool, free memory of the request structure.
975 * For requests obtained from a pool earlier, return request back to pool.
977 void ptlrpc_request_free(struct ptlrpc_request *request)
979 if (request->rq_pool)
980 __ptlrpc_free_req_to_pool(request);
982 ptlrpc_request_cache_free(request);
984 EXPORT_SYMBOL(ptlrpc_request_free);
987 * Allocate new request for operatione \a opcode and immediatelly pack it for
989 * Only used for simple requests like OBD_PING where the only important
990 * part of the request is operation itself.
991 * Returns allocated request or NULL on error.
993 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
994 const struct req_format *format,
995 __u32 version, int opcode)
997 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1001 rc = ptlrpc_request_pack(req, version, opcode);
1003 ptlrpc_request_free(req);
1009 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1012 * Allocate and initialize new request set structure on the current CPT.
1013 * Returns a pointer to the newly allocated set structure or NULL on error.
1015 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1017 struct ptlrpc_request_set *set;
1021 cpt = cfs_cpt_current(cfs_cpt_tab, 0);
1022 OBD_CPT_ALLOC(set, cfs_cpt_tab, cpt, sizeof(*set));
1025 atomic_set(&set->set_refcount, 1);
1026 INIT_LIST_HEAD(&set->set_requests);
1027 init_waitqueue_head(&set->set_waitq);
1028 atomic_set(&set->set_new_count, 0);
1029 atomic_set(&set->set_remaining, 0);
1030 spin_lock_init(&set->set_new_req_lock);
1031 INIT_LIST_HEAD(&set->set_new_requests);
1032 set->set_max_inflight = UINT_MAX;
1033 set->set_producer = NULL;
1034 set->set_producer_arg = NULL;
1039 EXPORT_SYMBOL(ptlrpc_prep_set);
1042 * Allocate and initialize new request set structure with flow control
1043 * extension. This extension allows to control the number of requests in-flight
1044 * for the whole set. A callback function to generate requests must be provided
1045 * and the request set will keep the number of requests sent over the wire to
1047 * Returns a pointer to the newly allocated set structure or NULL on error.
1049 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1053 struct ptlrpc_request_set *set;
1055 set = ptlrpc_prep_set();
1059 set->set_max_inflight = max;
1060 set->set_producer = func;
1061 set->set_producer_arg = arg;
1067 * Wind down and free request set structure previously allocated with
1069 * Ensures that all requests on the set have completed and removes
1070 * all requests from the request list in a set.
1071 * If any unsent request happen to be on the list, pretends that they got
1072 * an error in flight and calls their completion handler.
1074 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1076 struct list_head *tmp;
1077 struct list_head *next;
1083 /* Requests on the set should either all be completed, or all be new */
1084 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1085 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1086 list_for_each(tmp, &set->set_requests) {
1087 struct ptlrpc_request *req =
1088 list_entry(tmp, struct ptlrpc_request,
1091 LASSERT(req->rq_phase == expected_phase);
1095 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1096 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1097 atomic_read(&set->set_remaining), n);
1099 list_for_each_safe(tmp, next, &set->set_requests) {
1100 struct ptlrpc_request *req =
1101 list_entry(tmp, struct ptlrpc_request,
1103 list_del_init(&req->rq_set_chain);
1105 LASSERT(req->rq_phase == expected_phase);
1107 if (req->rq_phase == RQ_PHASE_NEW) {
1108 ptlrpc_req_interpret(NULL, req, -EBADR);
1109 atomic_dec(&set->set_remaining);
1112 spin_lock(&req->rq_lock);
1114 req->rq_invalid_rqset = 0;
1115 spin_unlock(&req->rq_lock);
1117 ptlrpc_req_finished(req);
1120 LASSERT(atomic_read(&set->set_remaining) == 0);
1122 ptlrpc_reqset_put(set);
1125 EXPORT_SYMBOL(ptlrpc_set_destroy);
1128 * Add a new request to the general purpose request set.
1129 * Assumes request reference from the caller.
1131 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1132 struct ptlrpc_request *req)
1134 if (set == PTLRPCD_SET) {
1135 ptlrpcd_add_req(req);
1139 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1140 LASSERT(list_empty(&req->rq_set_chain));
1142 if (req->rq_allow_intr)
1143 set->set_allow_intr = 1;
1145 /* The set takes over the caller's request reference */
1146 list_add_tail(&req->rq_set_chain, &set->set_requests);
1148 atomic_inc(&set->set_remaining);
1149 req->rq_queued_time = ktime_get_seconds();
1152 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1154 if (set->set_producer)
1156 * If the request set has a producer callback, the RPC must be
1157 * sent straight away
1159 ptlrpc_send_new_req(req);
1161 EXPORT_SYMBOL(ptlrpc_set_add_req);
1164 * Add a request to a request with dedicated server thread
1165 * and wake the thread to make any necessary processing.
1166 * Currently only used for ptlrpcd.
1168 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1169 struct ptlrpc_request *req)
1171 struct ptlrpc_request_set *set = pc->pc_set;
1174 LASSERT(req->rq_set == NULL);
1175 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1177 spin_lock(&set->set_new_req_lock);
1179 * The set takes over the caller's request reference.
1182 req->rq_queued_time = ktime_get_seconds();
1183 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1184 count = atomic_inc_return(&set->set_new_count);
1185 spin_unlock(&set->set_new_req_lock);
1187 /* Only need to call wakeup once for the first entry. */
1189 wake_up(&set->set_waitq);
1192 * XXX: It maybe unnecessary to wakeup all the partners. But to
1193 * guarantee the async RPC can be processed ASAP, we have
1194 * no other better choice. It maybe fixed in future.
1196 for (i = 0; i < pc->pc_npartners; i++)
1197 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1202 * Based on the current state of the import, determine if the request
1203 * can be sent, is an error, or should be delayed.
1205 * Returns true if this request should be delayed. If false, and
1206 * *status is set, then the request can not be sent and *status is the
1207 * error code. If false and status is 0, then request can be sent.
1209 * The imp->imp_lock must be held.
1211 static int ptlrpc_import_delay_req(struct obd_import *imp,
1212 struct ptlrpc_request *req, int *status)
1220 if (req->rq_ctx_init || req->rq_ctx_fini) {
1221 /* always allow ctx init/fini rpc go through */
1222 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1223 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1225 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1226 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1229 * pings or MDS-equivalent STATFS may safely
1232 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1233 D_HA : D_ERROR, req, "IMP_CLOSED");
1235 } else if (ptlrpc_send_limit_expired(req)) {
1236 /* probably doesn't need to be a D_ERROR afterinitial testing */
1237 DEBUG_REQ(D_HA, req, "send limit expired");
1238 *status = -ETIMEDOUT;
1239 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1240 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1241 ;/* allow CONNECT even if import is invalid */
1242 if (atomic_read(&imp->imp_inval_count) != 0) {
1243 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1246 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1247 if (!imp->imp_deactive)
1248 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1249 *status = -ESHUTDOWN; /* b=12940 */
1250 } else if (req->rq_import_generation != imp->imp_generation) {
1251 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1253 } else if (req->rq_send_state != imp->imp_state) {
1254 /* invalidate in progress - any requests should be drop */
1255 if (atomic_read(&imp->imp_inval_count) != 0) {
1256 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1258 } else if (req->rq_no_delay &&
1259 imp->imp_generation != imp->imp_initiated_at) {
1260 /* ignore nodelay for requests initiating connections */
1261 *status = -EWOULDBLOCK;
1262 } else if (req->rq_allow_replay &&
1263 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1264 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1265 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1266 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1267 DEBUG_REQ(D_HA, req, "allow during recovery");
1277 * Decide if the error message should be printed to the console or not.
1278 * Makes its decision based on request type, status, and failure frequency.
1280 * \param[in] req request that failed and may need a console message
1282 * \retval false if no message should be printed
1283 * \retval true if console message should be printed
1285 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1287 LASSERT(req->rq_reqmsg != NULL);
1289 /* Suppress particular reconnect errors which are to be expected. */
1290 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1291 /* Suppress timed out reconnect requests */
1292 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1297 * Suppress most unavailable/again reconnect requests, but
1298 * print occasionally so it is clear client is trying to
1299 * connect to a server where no target is running.
1301 if ((err == -ENODEV || err == -EAGAIN) &&
1302 req->rq_import->imp_conn_cnt % 30 != 20)
1306 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1307 /* -EAGAIN is normal when using POSIX flocks */
1310 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1311 (req->rq_xid & 0xf) != 10)
1312 /* Suppress most ping requests, they may fail occasionally */
1319 * Check request processing status.
1320 * Returns the status.
1322 static int ptlrpc_check_status(struct ptlrpc_request *req)
1327 rc = lustre_msg_get_status(req->rq_repmsg);
1328 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1329 struct obd_import *imp = req->rq_import;
1330 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1331 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1333 if (ptlrpc_console_allow(req, opc, rc))
1334 LCONSOLE_ERROR_MSG(0x11,
1335 "%s: operation %s to node %s failed: rc = %d\n",
1336 imp->imp_obd->obd_name,
1338 libcfs_nid2str(nid), rc);
1339 RETURN(rc < 0 ? rc : -EINVAL);
1343 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1349 * save pre-versions of objects into request for replay.
1350 * Versions are obtained from server reply.
1353 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1355 struct lustre_msg *repmsg = req->rq_repmsg;
1356 struct lustre_msg *reqmsg = req->rq_reqmsg;
1357 __u64 *versions = lustre_msg_get_versions(repmsg);
1360 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1364 lustre_msg_set_versions(reqmsg, versions);
1365 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1366 versions[0], versions[1]);
1371 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1373 struct ptlrpc_request *req;
1375 assert_spin_locked(&imp->imp_lock);
1376 if (list_empty(&imp->imp_unreplied_list))
1379 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1381 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1383 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1384 imp->imp_known_replied_xid = req->rq_xid - 1;
1386 return req->rq_xid - 1;
1390 * Callback function called when client receives RPC reply for \a req.
1391 * Returns 0 on success or error code.
1392 * The return alue would be assigned to req->rq_status by the caller
1393 * as request processing status.
1394 * This function also decides if the request needs to be saved for later replay.
1396 static int after_reply(struct ptlrpc_request *req)
1398 struct obd_import *imp = req->rq_import;
1399 struct obd_device *obd = req->rq_import->imp_obd;
1406 LASSERT(obd != NULL);
1407 /* repbuf must be unlinked */
1408 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1410 if (req->rq_reply_truncated) {
1411 if (ptlrpc_no_resend(req)) {
1412 DEBUG_REQ(D_ERROR, req,
1413 "reply buffer overflow, expected=%d, actual size=%d",
1414 req->rq_nob_received, req->rq_repbuf_len);
1418 sptlrpc_cli_free_repbuf(req);
1420 * Pass the required reply buffer size (include
1421 * space for early reply).
1422 * NB: no need to roundup because alloc_repbuf
1425 req->rq_replen = req->rq_nob_received;
1426 req->rq_nob_received = 0;
1427 spin_lock(&req->rq_lock);
1429 spin_unlock(&req->rq_lock);
1433 work_start = ktime_get_real();
1434 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1437 * NB Until this point, the whole of the incoming message,
1438 * including buflens, status etc is in the sender's byte order.
1440 rc = sptlrpc_cli_unwrap_reply(req);
1442 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1447 * Security layer unwrap might ask resend this request.
1452 rc = unpack_reply(req);
1456 /* retry indefinitely on EINPROGRESS */
1457 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1458 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1459 time64_t now = ktime_get_real_seconds();
1461 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1462 D_RPCTRACE, req, "resending request on EINPROGRESS");
1463 spin_lock(&req->rq_lock);
1465 spin_unlock(&req->rq_lock);
1466 req->rq_nr_resend++;
1468 /* Readjust the timeout for current conditions */
1469 ptlrpc_at_set_req_timeout(req);
1471 * delay resend to give a chance to the server to get ready.
1472 * The delay is increased by 1s on every resend and is capped to
1473 * the current request timeout (i.e. obd_timeout if AT is off,
1474 * or AT service time x 125% + 5s, see at_est2timeout)
1476 if (req->rq_nr_resend > req->rq_timeout)
1477 req->rq_sent = now + req->rq_timeout;
1479 req->rq_sent = now + req->rq_nr_resend;
1481 /* Resend for EINPROGRESS will use a new XID */
1482 spin_lock(&imp->imp_lock);
1483 list_del_init(&req->rq_unreplied_list);
1484 spin_unlock(&imp->imp_lock);
1489 if (obd->obd_svc_stats) {
1490 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1492 ptlrpc_lprocfs_rpc_sent(req, timediff);
1495 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1496 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1497 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1498 lustre_msg_get_type(req->rq_repmsg));
1502 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1503 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1504 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1505 ptlrpc_at_adj_net_latency(req,
1506 lustre_msg_get_service_time(req->rq_repmsg));
1508 rc = ptlrpc_check_status(req);
1512 * Either we've been evicted, or the server has failed for
1513 * some reason. Try to reconnect, and if that fails, punt to
1516 if (ptlrpc_recoverable_error(rc)) {
1517 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1518 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1521 ptlrpc_request_handle_notconn(req);
1526 * Let's look if server sent slv. Do it only for RPC with
1529 ldlm_cli_update_pool(req);
1533 * Store transno in reqmsg for replay.
1535 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1536 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1537 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1540 if (imp->imp_replayable) {
1541 spin_lock(&imp->imp_lock);
1543 * No point in adding already-committed requests to the replay
1544 * list, we will just remove them immediately. b=9829
1546 if (req->rq_transno != 0 &&
1548 lustre_msg_get_last_committed(req->rq_repmsg) ||
1550 /** version recovery */
1551 ptlrpc_save_versions(req);
1552 ptlrpc_retain_replayable_request(req, imp);
1553 } else if (req->rq_commit_cb &&
1554 list_empty(&req->rq_replay_list)) {
1556 * NB: don't call rq_commit_cb if it's already on
1557 * rq_replay_list, ptlrpc_free_committed() will call
1558 * it later, see LU-3618 for details
1560 spin_unlock(&imp->imp_lock);
1561 req->rq_commit_cb(req);
1562 spin_lock(&imp->imp_lock);
1566 * Replay-enabled imports return commit-status information.
1568 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1569 if (likely(committed > imp->imp_peer_committed_transno))
1570 imp->imp_peer_committed_transno = committed;
1572 ptlrpc_free_committed(imp);
1574 if (!list_empty(&imp->imp_replay_list)) {
1575 struct ptlrpc_request *last;
1577 last = list_entry(imp->imp_replay_list.prev,
1578 struct ptlrpc_request,
1581 * Requests with rq_replay stay on the list even if no
1582 * commit is expected.
1584 if (last->rq_transno > imp->imp_peer_committed_transno)
1585 ptlrpc_pinger_commit_expected(imp);
1588 spin_unlock(&imp->imp_lock);
1595 * Helper function to send request \a req over the network for the first time
1596 * Also adjusts request phase.
1597 * Returns 0 on success or error code.
1599 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1601 struct obd_import *imp = req->rq_import;
1606 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1608 /* do not try to go further if there is not enough memory in enc_pool */
1609 if (req->rq_sent && req->rq_bulk)
1610 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1611 pool_is_at_full_capacity())
1614 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1615 (!req->rq_generation_set ||
1616 req->rq_import_generation == imp->imp_generation))
1619 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1621 spin_lock(&imp->imp_lock);
1623 LASSERT(req->rq_xid != 0);
1624 LASSERT(!list_empty(&req->rq_unreplied_list));
1626 if (!req->rq_generation_set)
1627 req->rq_import_generation = imp->imp_generation;
1629 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1630 spin_lock(&req->rq_lock);
1631 req->rq_waiting = 1;
1632 spin_unlock(&req->rq_lock);
1634 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1635 ptlrpc_import_state_name(req->rq_send_state),
1636 ptlrpc_import_state_name(imp->imp_state));
1637 LASSERT(list_empty(&req->rq_list));
1638 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1639 atomic_inc(&req->rq_import->imp_inflight);
1640 spin_unlock(&imp->imp_lock);
1645 spin_unlock(&imp->imp_lock);
1646 req->rq_status = rc;
1647 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1651 LASSERT(list_empty(&req->rq_list));
1652 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1653 atomic_inc(&req->rq_import->imp_inflight);
1656 * find the known replied XID from the unreplied list, CONNECT
1657 * and DISCONNECT requests are skipped to make the sanity check
1658 * on server side happy. see process_req_last_xid().
1660 * For CONNECT: Because replay requests have lower XID, it'll
1661 * break the sanity check if CONNECT bump the exp_last_xid on
1664 * For DISCONNECT: Since client will abort inflight RPC before
1665 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1666 * than the inflight RPC.
1668 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1669 min_xid = ptlrpc_known_replied_xid(imp);
1670 spin_unlock(&imp->imp_lock);
1672 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1674 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1676 rc = sptlrpc_req_refresh_ctx(req, -1);
1679 req->rq_status = rc;
1682 spin_lock(&req->rq_lock);
1683 req->rq_wait_ctx = 1;
1684 spin_unlock(&req->rq_lock);
1690 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1691 req, current_comm(),
1692 imp->imp_obd->obd_uuid.uuid,
1693 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1694 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1695 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1697 rc = ptl_send_rpc(req, 0);
1698 if (rc == -ENOMEM) {
1699 spin_lock(&imp->imp_lock);
1700 if (!list_empty(&req->rq_list)) {
1701 list_del_init(&req->rq_list);
1702 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1703 wake_up(&req->rq_import->imp_recovery_waitq);
1705 spin_unlock(&imp->imp_lock);
1706 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1710 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1712 spin_lock(&req->rq_lock);
1713 req->rq_net_err = 1;
1714 spin_unlock(&req->rq_lock);
1720 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1725 LASSERT(set->set_producer != NULL);
1727 remaining = atomic_read(&set->set_remaining);
1730 * populate the ->set_requests list with requests until we
1731 * reach the maximum number of RPCs in flight for this set
1733 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1734 rc = set->set_producer(set, set->set_producer_arg);
1735 if (rc == -ENOENT) {
1736 /* no more RPC to produce */
1737 set->set_producer = NULL;
1738 set->set_producer_arg = NULL;
1743 RETURN((atomic_read(&set->set_remaining) - remaining));
1747 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1748 * and no more replies are expected.
1749 * (it is possible to get less replies than requests sent e.g. due to timed out
1750 * requests or requests that we had trouble to send out)
1752 * NOTE: This function contains a potential schedule point (cond_resched()).
1754 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1756 struct list_head *tmp, *next;
1757 LIST_HEAD(comp_reqs);
1758 int force_timer_recalc = 0;
1761 if (atomic_read(&set->set_remaining) == 0)
1764 list_for_each_safe(tmp, next, &set->set_requests) {
1765 struct ptlrpc_request *req =
1766 list_entry(tmp, struct ptlrpc_request,
1768 struct obd_import *imp = req->rq_import;
1769 int unregistered = 0;
1773 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1774 list_move_tail(&req->rq_set_chain, &comp_reqs);
1779 * This schedule point is mainly for the ptlrpcd caller of this
1780 * function. Most ptlrpc sets are not long-lived and unbounded
1781 * in length, but at the least the set used by the ptlrpcd is.
1782 * Since the processing time is unbounded, we need to insert an
1783 * explicit schedule point to make the thread well-behaved.
1788 * If the caller requires to allow to be interpreted by force
1789 * and it has really been interpreted, then move the request
1790 * to RQ_PHASE_INTERPRET phase in spite of what the current
1793 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1794 req->rq_status = -EINTR;
1795 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1798 * Since it is interpreted and we have to wait for
1799 * the reply to be unlinked, then use sync mode.
1803 GOTO(interpret, req->rq_status);
1806 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1807 force_timer_recalc = 1;
1809 /* delayed send - skip */
1810 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1813 /* delayed resend - skip */
1814 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1815 req->rq_sent > ktime_get_real_seconds())
1818 if (!(req->rq_phase == RQ_PHASE_RPC ||
1819 req->rq_phase == RQ_PHASE_BULK ||
1820 req->rq_phase == RQ_PHASE_INTERPRET ||
1821 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1822 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1823 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1827 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1828 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1829 LASSERT(req->rq_next_phase != req->rq_phase);
1830 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1832 if (req->rq_req_deadline &&
1833 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1834 req->rq_req_deadline = 0;
1835 if (req->rq_reply_deadline &&
1836 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1837 req->rq_reply_deadline = 0;
1838 if (req->rq_bulk_deadline &&
1839 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1840 req->rq_bulk_deadline = 0;
1843 * Skip processing until reply is unlinked. We
1844 * can't return to pool before that and we can't
1845 * call interpret before that. We need to make
1846 * sure that all rdma transfers finished and will
1847 * not corrupt any data.
1849 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1850 ptlrpc_client_recv_or_unlink(req))
1852 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1853 ptlrpc_client_bulk_active(req))
1857 * Turn fail_loc off to prevent it from looping
1860 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1861 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1864 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1865 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1870 * Move to next phase if reply was successfully
1873 ptlrpc_rqphase_move(req, req->rq_next_phase);
1876 if (req->rq_phase == RQ_PHASE_INTERPRET)
1877 GOTO(interpret, req->rq_status);
1880 * Note that this also will start async reply unlink.
1882 if (req->rq_net_err && !req->rq_timedout) {
1883 ptlrpc_expire_one_request(req, 1);
1886 * Check if we still need to wait for unlink.
1888 if (ptlrpc_client_recv_or_unlink(req) ||
1889 ptlrpc_client_bulk_active(req))
1891 /* If there is no need to resend, fail it now. */
1892 if (req->rq_no_resend) {
1893 if (req->rq_status == 0)
1894 req->rq_status = -EIO;
1895 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1896 GOTO(interpret, req->rq_status);
1903 spin_lock(&req->rq_lock);
1904 req->rq_replied = 0;
1905 spin_unlock(&req->rq_lock);
1906 if (req->rq_status == 0)
1907 req->rq_status = -EIO;
1908 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1909 GOTO(interpret, req->rq_status);
1913 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1914 * so it sets rq_intr regardless of individual rpc
1915 * timeouts. The synchronous IO waiting path sets
1916 * rq_intr irrespective of whether ptlrpcd
1917 * has seen a timeout. Our policy is to only interpret
1918 * interrupted rpcs after they have timed out, so we
1919 * need to enforce that here.
1922 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1923 req->rq_wait_ctx)) {
1924 req->rq_status = -EINTR;
1925 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1926 GOTO(interpret, req->rq_status);
1929 if (req->rq_phase == RQ_PHASE_RPC) {
1930 if (req->rq_timedout || req->rq_resend ||
1931 req->rq_waiting || req->rq_wait_ctx) {
1934 if (!ptlrpc_unregister_reply(req, 1)) {
1935 ptlrpc_unregister_bulk(req, 1);
1939 spin_lock(&imp->imp_lock);
1940 if (ptlrpc_import_delay_req(imp, req,
1943 * put on delay list - only if we wait
1944 * recovery finished - before send
1946 list_move_tail(&req->rq_list,
1947 &imp->imp_delayed_list);
1948 spin_unlock(&imp->imp_lock);
1953 req->rq_status = status;
1954 ptlrpc_rqphase_move(req,
1955 RQ_PHASE_INTERPRET);
1956 spin_unlock(&imp->imp_lock);
1957 GOTO(interpret, req->rq_status);
1959 /* ignore on just initiated connections */
1960 if (ptlrpc_no_resend(req) &&
1961 !req->rq_wait_ctx &&
1962 imp->imp_generation !=
1963 imp->imp_initiated_at) {
1964 req->rq_status = -ENOTCONN;
1965 ptlrpc_rqphase_move(req,
1966 RQ_PHASE_INTERPRET);
1967 spin_unlock(&imp->imp_lock);
1968 GOTO(interpret, req->rq_status);
1971 list_move_tail(&req->rq_list,
1972 &imp->imp_sending_list);
1974 spin_unlock(&imp->imp_lock);
1976 spin_lock(&req->rq_lock);
1977 req->rq_waiting = 0;
1978 spin_unlock(&req->rq_lock);
1980 if (req->rq_timedout || req->rq_resend) {
1982 * This is re-sending anyways,
1983 * let's mark req as resend.
1985 spin_lock(&req->rq_lock);
1987 spin_unlock(&req->rq_lock);
1990 * rq_wait_ctx is only touched by ptlrpcd,
1991 * so no lock is needed here.
1993 status = sptlrpc_req_refresh_ctx(req, -1);
1996 req->rq_status = status;
1997 spin_lock(&req->rq_lock);
1998 req->rq_wait_ctx = 0;
1999 spin_unlock(&req->rq_lock);
2000 force_timer_recalc = 1;
2002 spin_lock(&req->rq_lock);
2003 req->rq_wait_ctx = 1;
2004 spin_unlock(&req->rq_lock);
2009 spin_lock(&req->rq_lock);
2010 req->rq_wait_ctx = 0;
2011 spin_unlock(&req->rq_lock);
2015 * In any case, the previous bulk should be
2016 * cleaned up to prepare for the new sending
2019 !ptlrpc_unregister_bulk(req, 1))
2022 rc = ptl_send_rpc(req, 0);
2023 if (rc == -ENOMEM) {
2024 spin_lock(&imp->imp_lock);
2025 if (!list_empty(&req->rq_list))
2026 list_del_init(&req->rq_list);
2027 spin_unlock(&imp->imp_lock);
2028 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2032 DEBUG_REQ(D_HA, req,
2033 "send failed: rc = %d", rc);
2034 force_timer_recalc = 1;
2035 spin_lock(&req->rq_lock);
2036 req->rq_net_err = 1;
2037 spin_unlock(&req->rq_lock);
2040 /* need to reset the timeout */
2041 force_timer_recalc = 1;
2044 spin_lock(&req->rq_lock);
2046 if (ptlrpc_client_early(req)) {
2047 ptlrpc_at_recv_early_reply(req);
2048 spin_unlock(&req->rq_lock);
2052 /* Still waiting for a reply? */
2053 if (ptlrpc_client_recv(req)) {
2054 spin_unlock(&req->rq_lock);
2058 /* Did we actually receive a reply? */
2059 if (!ptlrpc_client_replied(req)) {
2060 spin_unlock(&req->rq_lock);
2064 spin_unlock(&req->rq_lock);
2067 * unlink from net because we are going to
2068 * swab in-place of reply buffer
2070 unregistered = ptlrpc_unregister_reply(req, 1);
2074 req->rq_status = after_reply(req);
2079 * If there is no bulk associated with this request,
2080 * then we're done and should let the interpreter
2081 * process the reply. Similarly if the RPC returned
2082 * an error, and therefore the bulk will never arrive.
2084 if (!req->rq_bulk || req->rq_status < 0) {
2085 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2086 GOTO(interpret, req->rq_status);
2089 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2092 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2093 if (ptlrpc_client_bulk_active(req))
2096 if (req->rq_bulk->bd_failure) {
2098 * The RPC reply arrived OK, but the bulk screwed
2099 * up! Dead weird since the server told us the RPC
2100 * was good after getting the REPLY for her GET or
2101 * the ACK for her PUT.
2103 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2104 req->rq_status = -EIO;
2107 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2110 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2113 * This moves to "unregistering" phase we need to wait for
2116 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2117 /* start async bulk unlink too */
2118 ptlrpc_unregister_bulk(req, 1);
2122 if (!ptlrpc_unregister_bulk(req, async))
2126 * When calling interpret receiving already should be
2129 LASSERT(!req->rq_receiving_reply);
2131 ptlrpc_req_interpret(env, req, req->rq_status);
2133 if (ptlrpcd_check_work(req)) {
2134 atomic_dec(&set->set_remaining);
2137 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2141 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2142 req, current_comm(),
2143 imp->imp_obd->obd_uuid.uuid,
2144 lustre_msg_get_status(req->rq_reqmsg),
2146 obd_import_nid2str(imp),
2147 lustre_msg_get_opc(req->rq_reqmsg),
2148 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2150 spin_lock(&imp->imp_lock);
2152 * Request already may be not on sending or delaying list. This
2153 * may happen in the case of marking it erroneous for the case
2154 * ptlrpc_import_delay_req(req, status) find it impossible to
2155 * allow sending this rpc and returns *status != 0.
2157 if (!list_empty(&req->rq_list)) {
2158 list_del_init(&req->rq_list);
2159 if (atomic_dec_and_test(&imp->imp_inflight))
2160 wake_up(&imp->imp_recovery_waitq);
2162 list_del_init(&req->rq_unreplied_list);
2163 spin_unlock(&imp->imp_lock);
2165 atomic_dec(&set->set_remaining);
2166 wake_up(&imp->imp_recovery_waitq);
2168 if (set->set_producer) {
2169 /* produce a new request if possible */
2170 if (ptlrpc_set_producer(set) > 0)
2171 force_timer_recalc = 1;
2174 * free the request that has just been completed
2175 * in order not to pollute set->set_requests
2177 list_del_init(&req->rq_set_chain);
2178 spin_lock(&req->rq_lock);
2180 req->rq_invalid_rqset = 0;
2181 spin_unlock(&req->rq_lock);
2183 /* record rq_status to compute the final status later */
2184 if (req->rq_status != 0)
2185 set->set_rc = req->rq_status;
2186 ptlrpc_req_finished(req);
2188 list_move_tail(&req->rq_set_chain, &comp_reqs);
2193 * move completed request at the head of list so it's easier for
2194 * caller to find them
2196 list_splice(&comp_reqs, &set->set_requests);
2198 /* If we hit an error, we want to recover promptly. */
2199 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2201 EXPORT_SYMBOL(ptlrpc_check_set);
2204 * Time out request \a req. is \a async_unlink is set, that means do not wait
2205 * until LNet actually confirms network buffer unlinking.
2206 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2208 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2210 struct obd_import *imp = req->rq_import;
2211 unsigned int debug_mask = D_RPCTRACE;
2215 spin_lock(&req->rq_lock);
2216 req->rq_timedout = 1;
2217 spin_unlock(&req->rq_lock);
2219 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2220 lustre_msg_get_status(req->rq_reqmsg)))
2221 debug_mask = D_WARNING;
2222 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2223 req->rq_net_err ? "failed due to network error" :
2224 ((req->rq_real_sent == 0 ||
2225 req->rq_real_sent < req->rq_sent ||
2226 req->rq_real_sent >= req->rq_deadline) ?
2227 "timed out for sent delay" : "timed out for slow reply"),
2228 (s64)req->rq_sent, (s64)req->rq_real_sent);
2230 if (imp && obd_debug_peer_on_timeout)
2231 LNetDebugPeer(imp->imp_connection->c_peer);
2233 ptlrpc_unregister_reply(req, async_unlink);
2234 ptlrpc_unregister_bulk(req, async_unlink);
2236 if (obd_dump_on_timeout)
2237 libcfs_debug_dumplog();
2240 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2244 atomic_inc(&imp->imp_timeouts);
2246 /* The DLM server doesn't want recovery run on its imports. */
2247 if (imp->imp_dlm_fake)
2251 * If this request is for recovery or other primordial tasks,
2252 * then error it out here.
2254 if (req->rq_ctx_init || req->rq_ctx_fini ||
2255 req->rq_send_state != LUSTRE_IMP_FULL ||
2256 imp->imp_obd->obd_no_recov) {
2257 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2258 ptlrpc_import_state_name(req->rq_send_state),
2259 ptlrpc_import_state_name(imp->imp_state));
2260 spin_lock(&req->rq_lock);
2261 req->rq_status = -ETIMEDOUT;
2263 spin_unlock(&req->rq_lock);
2268 * if a request can't be resent we can't wait for an answer after
2271 if (ptlrpc_no_resend(req)) {
2272 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2276 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2282 * Time out all uncompleted requests in request set pointed by \a data
2283 * Callback used when waiting on sets with l_wait_event.
2286 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2288 struct list_head *tmp;
2289 time64_t now = ktime_get_real_seconds();
2292 LASSERT(set != NULL);
2295 * A timeout expired. See which reqs it applies to...
2297 list_for_each(tmp, &set->set_requests) {
2298 struct ptlrpc_request *req =
2299 list_entry(tmp, struct ptlrpc_request,
2302 /* don't expire request waiting for context */
2303 if (req->rq_wait_ctx)
2306 /* Request in-flight? */
2307 if (!((req->rq_phase == RQ_PHASE_RPC &&
2308 !req->rq_waiting && !req->rq_resend) ||
2309 (req->rq_phase == RQ_PHASE_BULK)))
2312 if (req->rq_timedout || /* already dealt with */
2313 req->rq_deadline > now) /* not expired */
2317 * Deal with this guy. Do it asynchronously to not block
2320 ptlrpc_expire_one_request(req, 1);
2325 * Interrupts (sets interrupted flag) all uncompleted requests in
2326 * a set \a data. Callback for l_wait_event for interruptible waits.
2328 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2330 struct list_head *tmp;
2332 LASSERT(set != NULL);
2333 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2335 list_for_each(tmp, &set->set_requests) {
2336 struct ptlrpc_request *req =
2337 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2342 if (req->rq_phase != RQ_PHASE_RPC &&
2343 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2344 !req->rq_allow_intr)
2347 spin_lock(&req->rq_lock);
2349 spin_unlock(&req->rq_lock);
2354 * Get the smallest timeout in the set; this does NOT set a timeout.
2356 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2358 struct list_head *tmp;
2359 time64_t now = ktime_get_real_seconds();
2361 struct ptlrpc_request *req;
2365 list_for_each(tmp, &set->set_requests) {
2366 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2368 /* Request in-flight? */
2369 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2370 (req->rq_phase == RQ_PHASE_BULK) ||
2371 (req->rq_phase == RQ_PHASE_NEW)))
2374 /* Already timed out. */
2375 if (req->rq_timedout)
2378 /* Waiting for ctx. */
2379 if (req->rq_wait_ctx)
2382 if (req->rq_phase == RQ_PHASE_NEW)
2383 deadline = req->rq_sent;
2384 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2385 deadline = req->rq_sent;
2387 deadline = req->rq_sent + req->rq_timeout;
2389 if (deadline <= now) /* actually expired already */
2390 timeout = 1; /* ASAP */
2391 else if (timeout == 0 || timeout > deadline - now)
2392 timeout = deadline - now;
2398 * Send all unset request from the set and then wait untill all
2399 * requests in the set complete (either get a reply, timeout, get an
2400 * error or otherwise be interrupted).
2401 * Returns 0 on success or error code otherwise.
2403 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2405 struct list_head *tmp;
2406 struct ptlrpc_request *req;
2411 if (set->set_producer)
2412 (void)ptlrpc_set_producer(set);
2414 list_for_each(tmp, &set->set_requests) {
2415 req = list_entry(tmp, struct ptlrpc_request,
2417 if (req->rq_phase == RQ_PHASE_NEW)
2418 (void)ptlrpc_send_new_req(req);
2421 if (list_empty(&set->set_requests))
2425 timeout = ptlrpc_set_next_timeout(set);
2428 * wait until all complete, interrupted, or an in-flight
2431 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2434 if ((timeout == 0 && !signal_pending(current)) ||
2435 set->set_allow_intr) {
2437 * No requests are in-flight (ether timed out
2438 * or delayed), so we can allow interrupts.
2439 * We still want to block for a limited time,
2440 * so we allow interrupts during the timeout.
2442 rc = l_wait_event_abortable_timeout(
2444 ptlrpc_check_set(NULL, set),
2445 cfs_time_seconds(timeout ? timeout : 1));
2448 ptlrpc_expired_set(set);
2449 } else if (rc < 0) {
2451 ptlrpc_interrupted_set(set);
2457 * At least one request is in flight, so no
2458 * interrupts are allowed. Wait until all
2459 * complete, or an in-flight req times out.
2461 rc = wait_event_idle_timeout(
2463 ptlrpc_check_set(NULL, set),
2464 cfs_time_seconds(timeout ? timeout : 1));
2466 ptlrpc_expired_set(set);
2473 * LU-769 - if we ignored the signal because
2474 * it was already pending when we started, we
2475 * need to handle it now or we risk it being
2478 if (rc == -ETIMEDOUT &&
2479 signal_pending(current)) {
2480 sigset_t blocked_sigs =
2481 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2484 * In fact we only interrupt for the
2485 * "fatal" signals like SIGINT or
2486 * SIGKILL. We still ignore less
2487 * important signals since ptlrpc set
2488 * is not easily reentrant from
2491 if (signal_pending(current))
2492 ptlrpc_interrupted_set(set);
2493 cfs_restore_sigs(blocked_sigs);
2497 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2500 * -EINTR => all requests have been flagged rq_intr so next
2502 * -ETIMEDOUT => someone timed out. When all reqs have
2503 * timed out, signals are enabled allowing completion with
2505 * I don't really care if we go once more round the loop in
2506 * the error cases -eeb.
2508 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2509 list_for_each(tmp, &set->set_requests) {
2510 req = list_entry(tmp, struct ptlrpc_request,
2512 spin_lock(&req->rq_lock);
2513 req->rq_invalid_rqset = 1;
2514 spin_unlock(&req->rq_lock);
2517 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2519 LASSERT(atomic_read(&set->set_remaining) == 0);
2521 rc = set->set_rc; /* rq_status of already freed requests if any */
2522 list_for_each(tmp, &set->set_requests) {
2523 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2525 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2526 if (req->rq_status != 0)
2527 rc = req->rq_status;
2532 EXPORT_SYMBOL(ptlrpc_set_wait);
2535 * Helper fuction for request freeing.
2536 * Called when request count reached zero and request needs to be freed.
2537 * Removes request from all sorts of sending/replay lists it might be on,
2538 * frees network buffers if any are present.
2539 * If \a locked is set, that means caller is already holding import imp_lock
2540 * and so we no longer need to reobtain it (for certain lists manipulations)
2542 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2549 LASSERT(!request->rq_srv_req);
2550 LASSERT(request->rq_export == NULL);
2551 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2552 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2553 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2554 LASSERTF(!request->rq_replay, "req %p\n", request);
2556 req_capsule_fini(&request->rq_pill);
2559 * We must take it off the imp_replay_list first. Otherwise, we'll set
2560 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2562 if (request->rq_import) {
2564 spin_lock(&request->rq_import->imp_lock);
2565 list_del_init(&request->rq_replay_list);
2566 list_del_init(&request->rq_unreplied_list);
2568 spin_unlock(&request->rq_import->imp_lock);
2570 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2572 if (atomic_read(&request->rq_refcount) != 0) {
2573 DEBUG_REQ(D_ERROR, request,
2574 "freeing request with nonzero refcount");
2578 if (request->rq_repbuf)
2579 sptlrpc_cli_free_repbuf(request);
2581 if (request->rq_import) {
2582 class_import_put(request->rq_import);
2583 request->rq_import = NULL;
2585 if (request->rq_bulk)
2586 ptlrpc_free_bulk(request->rq_bulk);
2588 if (request->rq_reqbuf || request->rq_clrbuf)
2589 sptlrpc_cli_free_reqbuf(request);
2591 if (request->rq_cli_ctx)
2592 sptlrpc_req_put_ctx(request, !locked);
2594 if (request->rq_pool)
2595 __ptlrpc_free_req_to_pool(request);
2597 ptlrpc_request_cache_free(request);
2601 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2603 * Drop one request reference. Must be called with import imp_lock held.
2604 * When reference count drops to zero, request is freed.
2606 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2608 assert_spin_locked(&request->rq_import->imp_lock);
2609 (void)__ptlrpc_req_finished(request, 1);
2614 * Drops one reference count for request \a request.
2615 * \a locked set indicates that caller holds import imp_lock.
2616 * Frees the request whe reference count reaches zero.
2618 * \retval 1 the request is freed
2619 * \retval 0 some others still hold references on the request
2621 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2629 LASSERT(request != LP_POISON);
2630 LASSERT(request->rq_reqmsg != LP_POISON);
2632 DEBUG_REQ(D_INFO, request, "refcount now %u",
2633 atomic_read(&request->rq_refcount) - 1);
2635 spin_lock(&request->rq_lock);
2636 count = atomic_dec_return(&request->rq_refcount);
2637 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2640 * For open RPC, the client does not know the EA size (LOV, ACL, and
2641 * so on) before replied, then the client has to reserve very large
2642 * reply buffer. Such buffer will not be released until the RPC freed.
2643 * Since The open RPC is replayable, we need to keep it in the replay
2644 * list until close. If there are a lot of files opened concurrently,
2645 * then the client may be OOM.
2647 * If fact, it is unnecessary to keep reply buffer for open replay,
2648 * related EAs have already been saved via mdc_save_lovea() before
2649 * coming here. So it is safe to free the reply buffer some earlier
2650 * before releasing the RPC to avoid client OOM. LU-9514
2652 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2653 spin_lock(&request->rq_early_free_lock);
2654 sptlrpc_cli_free_repbuf(request);
2655 request->rq_repbuf = NULL;
2656 request->rq_repbuf_len = 0;
2657 request->rq_repdata = NULL;
2658 request->rq_reqdata_len = 0;
2659 spin_unlock(&request->rq_early_free_lock);
2661 spin_unlock(&request->rq_lock);
2664 __ptlrpc_free_req(request, locked);
2670 * Drops one reference count for a request.
2672 void ptlrpc_req_finished(struct ptlrpc_request *request)
2674 __ptlrpc_req_finished(request, 0);
2676 EXPORT_SYMBOL(ptlrpc_req_finished);
2679 * Returns xid of a \a request
2681 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2683 return request->rq_xid;
2685 EXPORT_SYMBOL(ptlrpc_req_xid);
2688 * Disengage the client's reply buffer from the network
2689 * NB does _NOT_ unregister any client-side bulk.
2690 * IDEMPOTENT, but _not_ safe against concurrent callers.
2691 * The request owner (i.e. the thread doing the I/O) must call...
2692 * Returns 0 on success or 1 if unregistering cannot be made.
2694 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2699 LASSERT(!in_interrupt());
2701 /* Let's setup deadline for reply unlink. */
2702 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2703 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2704 request->rq_reply_deadline = ktime_get_real_seconds() +
2708 * Nothing left to do.
2710 if (!ptlrpc_client_recv_or_unlink(request))
2713 LNetMDUnlink(request->rq_reply_md_h);
2716 * Let's check it once again.
2718 if (!ptlrpc_client_recv_or_unlink(request))
2721 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2722 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2725 * Do not wait for unlink to finish.
2731 * We have to wait_event_idle_timeout() whatever the result, to get
2732 * a chance to run reply_in_callback(), and to make sure we've
2733 * unlinked before returning a req to the pool.
2736 wait_queue_head_t *wq = (request->rq_set) ?
2737 &request->rq_set->set_waitq :
2738 &request->rq_reply_waitq;
2739 int seconds = LONG_UNLINK;
2741 * Network access will complete in finite time but the HUGE
2742 * timeout lets us CWARN for visibility of sluggish NALs
2744 while (seconds > 0 &&
2745 wait_event_idle_timeout(
2747 !ptlrpc_client_recv_or_unlink(request),
2748 cfs_time_seconds(1)) == 0)
2751 ptlrpc_rqphase_move(request, request->rq_next_phase);
2755 DEBUG_REQ(D_WARNING, request,
2756 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2757 request->rq_receiving_reply,
2758 request->rq_req_unlinked,
2759 request->rq_reply_unlinked);
2764 static void ptlrpc_free_request(struct ptlrpc_request *req)
2766 spin_lock(&req->rq_lock);
2768 spin_unlock(&req->rq_lock);
2770 if (req->rq_commit_cb)
2771 req->rq_commit_cb(req);
2772 list_del_init(&req->rq_replay_list);
2774 __ptlrpc_req_finished(req, 1);
2778 * the request is committed and dropped from the replay list of its import
2780 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2782 struct obd_import *imp = req->rq_import;
2784 spin_lock(&imp->imp_lock);
2785 if (list_empty(&req->rq_replay_list)) {
2786 spin_unlock(&imp->imp_lock);
2790 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2791 if (imp->imp_replay_cursor == &req->rq_replay_list)
2792 imp->imp_replay_cursor = req->rq_replay_list.next;
2793 ptlrpc_free_request(req);
2796 spin_unlock(&imp->imp_lock);
2798 EXPORT_SYMBOL(ptlrpc_request_committed);
2801 * Iterates through replay_list on import and prunes
2802 * all requests have transno smaller than last_committed for the
2803 * import and don't have rq_replay set.
2804 * Since requests are sorted in transno order, stops when meetign first
2805 * transno bigger than last_committed.
2806 * caller must hold imp->imp_lock
2808 void ptlrpc_free_committed(struct obd_import *imp)
2810 struct ptlrpc_request *req, *saved;
2811 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2812 bool skip_committed_list = true;
2815 LASSERT(imp != NULL);
2816 assert_spin_locked(&imp->imp_lock);
2818 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2819 imp->imp_generation == imp->imp_last_generation_checked) {
2820 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2821 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2824 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2825 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2826 imp->imp_generation);
2828 if (imp->imp_generation != imp->imp_last_generation_checked ||
2829 imp->imp_last_transno_checked == 0)
2830 skip_committed_list = false;
2832 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2833 imp->imp_last_generation_checked = imp->imp_generation;
2835 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2837 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2838 LASSERT(req != last_req);
2841 if (req->rq_transno == 0) {
2842 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2845 if (req->rq_import_generation < imp->imp_generation) {
2846 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2850 /* not yet committed */
2851 if (req->rq_transno > imp->imp_peer_committed_transno) {
2852 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2856 if (req->rq_replay) {
2857 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2858 list_move_tail(&req->rq_replay_list,
2859 &imp->imp_committed_list);
2863 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2864 imp->imp_peer_committed_transno);
2866 ptlrpc_free_request(req);
2869 if (skip_committed_list)
2872 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2874 LASSERT(req->rq_transno != 0);
2875 if (req->rq_import_generation < imp->imp_generation ||
2877 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2878 req->rq_import_generation <
2879 imp->imp_generation ? "stale" : "closed");
2881 if (imp->imp_replay_cursor == &req->rq_replay_list)
2882 imp->imp_replay_cursor =
2883 req->rq_replay_list.next;
2885 ptlrpc_free_request(req);
2892 void ptlrpc_cleanup_client(struct obd_import *imp)
2899 * Schedule previously sent request for resend.
2900 * For bulk requests we assign new xid (to avoid problems with
2901 * lost replies and therefore several transfers landing into same buffer
2902 * from different sending attempts).
2904 void ptlrpc_resend_req(struct ptlrpc_request *req)
2906 DEBUG_REQ(D_HA, req, "going to resend");
2907 spin_lock(&req->rq_lock);
2910 * Request got reply but linked to the import list still.
2911 * Let ptlrpc_check_set() process it.
2913 if (ptlrpc_client_replied(req)) {
2914 spin_unlock(&req->rq_lock);
2915 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2919 req->rq_status = -EAGAIN;
2922 req->rq_net_err = 0;
2923 req->rq_timedout = 0;
2925 ptlrpc_client_wake_req(req);
2926 spin_unlock(&req->rq_lock);
2929 /* XXX: this function and rq_status are currently unused */
2930 void ptlrpc_restart_req(struct ptlrpc_request *req)
2932 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2933 req->rq_status = -ERESTARTSYS;
2935 spin_lock(&req->rq_lock);
2936 req->rq_restart = 1;
2937 req->rq_timedout = 0;
2938 ptlrpc_client_wake_req(req);
2939 spin_unlock(&req->rq_lock);
2943 * Grab additional reference on a request \a req
2945 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2948 atomic_inc(&req->rq_refcount);
2951 EXPORT_SYMBOL(ptlrpc_request_addref);
2954 * Add a request to import replay_list.
2955 * Must be called under imp_lock
2957 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2958 struct obd_import *imp)
2960 struct list_head *tmp;
2962 assert_spin_locked(&imp->imp_lock);
2964 if (req->rq_transno == 0) {
2965 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2970 * clear this for new requests that were resent as well
2971 * as resent replayed requests.
2973 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2975 /* don't re-add requests that have been replayed */
2976 if (!list_empty(&req->rq_replay_list))
2979 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2981 spin_lock(&req->rq_lock);
2983 spin_unlock(&req->rq_lock);
2985 LASSERT(imp->imp_replayable);
2986 /* Balanced in ptlrpc_free_committed, usually. */
2987 ptlrpc_request_addref(req);
2988 list_for_each_prev(tmp, &imp->imp_replay_list) {
2989 struct ptlrpc_request *iter = list_entry(tmp,
2990 struct ptlrpc_request,
2994 * We may have duplicate transnos if we create and then
2995 * open a file, or for closes retained if to match creating
2996 * opens, so use req->rq_xid as a secondary key.
2997 * (See bugs 684, 685, and 428.)
2998 * XXX no longer needed, but all opens need transnos!
3000 if (iter->rq_transno > req->rq_transno)
3003 if (iter->rq_transno == req->rq_transno) {
3004 LASSERT(iter->rq_xid != req->rq_xid);
3005 if (iter->rq_xid > req->rq_xid)
3009 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3013 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3017 * Send request and wait until it completes.
3018 * Returns request processing status.
3020 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3022 struct ptlrpc_request_set *set;
3026 LASSERT(req->rq_set == NULL);
3027 LASSERT(!req->rq_receiving_reply);
3029 set = ptlrpc_prep_set();
3031 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3035 /* for distributed debugging */
3036 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3038 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3039 ptlrpc_request_addref(req);
3040 ptlrpc_set_add_req(set, req);
3041 rc = ptlrpc_set_wait(NULL, set);
3042 ptlrpc_set_destroy(set);
3046 EXPORT_SYMBOL(ptlrpc_queue_wait);
3049 * Callback used for replayed requests reply processing.
3050 * In case of successful reply calls registered request replay callback.
3051 * In case of error restart replay process.
3053 static int ptlrpc_replay_interpret(const struct lu_env *env,
3054 struct ptlrpc_request *req,
3057 struct ptlrpc_replay_async_args *aa = args;
3058 struct obd_import *imp = req->rq_import;
3061 atomic_dec(&imp->imp_replay_inflight);
3064 * Note: if it is bulk replay (MDS-MDS replay), then even if
3065 * server got the request, but bulk transfer timeout, let's
3066 * replay the bulk req again
3068 if (!ptlrpc_client_replied(req) ||
3070 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3071 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3072 GOTO(out, rc = -ETIMEDOUT);
3075 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3076 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3077 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3078 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3080 /** VBR: check version failure */
3081 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3082 /** replay was failed due to version mismatch */
3083 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3084 spin_lock(&imp->imp_lock);
3085 imp->imp_vbr_failed = 1;
3086 spin_unlock(&imp->imp_lock);
3087 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3089 /** The transno had better not change over replay. */
3090 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3091 lustre_msg_get_transno(req->rq_repmsg) ||
3092 lustre_msg_get_transno(req->rq_repmsg) == 0,
3094 lustre_msg_get_transno(req->rq_reqmsg),
3095 lustre_msg_get_transno(req->rq_repmsg));
3098 spin_lock(&imp->imp_lock);
3099 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3100 spin_unlock(&imp->imp_lock);
3101 LASSERT(imp->imp_last_replay_transno);
3103 /* transaction number shouldn't be bigger than the latest replayed */
3104 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3105 DEBUG_REQ(D_ERROR, req,
3106 "Reported transno=%llu is bigger than replayed=%llu",
3108 lustre_msg_get_transno(req->rq_reqmsg));
3109 GOTO(out, rc = -EINVAL);
3112 DEBUG_REQ(D_HA, req, "got reply");
3114 /* let the callback do fixups, possibly including in the request */
3115 if (req->rq_replay_cb)
3116 req->rq_replay_cb(req);
3118 if (ptlrpc_client_replied(req) &&
3119 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3120 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3121 lustre_msg_get_status(req->rq_repmsg),
3122 aa->praa_old_status);
3125 * Note: If the replay fails for MDT-MDT recovery, let's
3126 * abort all of the following requests in the replay
3127 * and sending list, because MDT-MDT update requests
3128 * are dependent on each other, see LU-7039
3130 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3131 struct ptlrpc_request *free_req;
3132 struct ptlrpc_request *tmp;
3134 spin_lock(&imp->imp_lock);
3135 list_for_each_entry_safe(free_req, tmp,
3136 &imp->imp_replay_list,
3138 ptlrpc_free_request(free_req);
3141 list_for_each_entry_safe(free_req, tmp,
3142 &imp->imp_committed_list,
3144 ptlrpc_free_request(free_req);
3147 list_for_each_entry_safe(free_req, tmp,
3148 &imp->imp_delayed_list,
3150 spin_lock(&free_req->rq_lock);
3151 free_req->rq_err = 1;
3152 free_req->rq_status = -EIO;
3153 ptlrpc_client_wake_req(free_req);
3154 spin_unlock(&free_req->rq_lock);
3157 list_for_each_entry_safe(free_req, tmp,
3158 &imp->imp_sending_list,
3160 spin_lock(&free_req->rq_lock);
3161 free_req->rq_err = 1;
3162 free_req->rq_status = -EIO;
3163 ptlrpc_client_wake_req(free_req);
3164 spin_unlock(&free_req->rq_lock);
3166 spin_unlock(&imp->imp_lock);
3169 /* Put it back for re-replay. */
3170 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3174 * Errors while replay can set transno to 0, but
3175 * imp_last_replay_transno shouldn't be set to 0 anyway
3177 if (req->rq_transno == 0)
3178 CERROR("Transno is 0 during replay!\n");
3180 /* continue with recovery */
3181 rc = ptlrpc_import_recovery_state_machine(imp);
3183 req->rq_send_state = aa->praa_old_state;
3186 /* this replay failed, so restart recovery */
3187 ptlrpc_connect_import(imp);
3193 * Prepares and queues request for replay.
3194 * Adds it to ptlrpcd queue for actual sending.
3195 * Returns 0 on success.
3197 int ptlrpc_replay_req(struct ptlrpc_request *req)
3199 struct ptlrpc_replay_async_args *aa;
3203 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3205 aa = ptlrpc_req_async_args(aa, req);
3206 memset(aa, 0, sizeof(*aa));
3208 /* Prepare request to be resent with ptlrpcd */
3209 aa->praa_old_state = req->rq_send_state;
3210 req->rq_send_state = LUSTRE_IMP_REPLAY;
3211 req->rq_phase = RQ_PHASE_NEW;
3212 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3214 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3216 req->rq_interpret_reply = ptlrpc_replay_interpret;
3217 /* Readjust the timeout for current conditions */
3218 ptlrpc_at_set_req_timeout(req);
3220 /* Tell server net_latency to calculate how long to wait for reply. */
3221 lustre_msg_set_service_time(req->rq_reqmsg,
3222 ptlrpc_at_get_net_latency(req));
3223 DEBUG_REQ(D_HA, req, "REPLAY");
3225 atomic_inc(&req->rq_import->imp_replay_inflight);
3226 spin_lock(&req->rq_lock);
3227 req->rq_early_free_repbuf = 0;
3228 spin_unlock(&req->rq_lock);
3229 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3231 ptlrpcd_add_req(req);
3236 * Aborts all in-flight request on import \a imp sending and delayed lists
3238 void ptlrpc_abort_inflight(struct obd_import *imp)
3240 struct list_head *tmp, *n;
3244 * Make sure that no new requests get processed for this import.
3245 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3246 * this flag and then putting requests on sending_list or delayed_list.
3248 assert_spin_locked(&imp->imp_lock);
3251 * XXX locking? Maybe we should remove each request with the list
3252 * locked? Also, how do we know if the requests on the list are
3253 * being freed at this time?
3255 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3256 struct ptlrpc_request *req = list_entry(tmp,
3257 struct ptlrpc_request,
3260 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3262 spin_lock(&req->rq_lock);
3263 if (req->rq_import_generation < imp->imp_generation) {
3265 req->rq_status = -EIO;
3266 ptlrpc_client_wake_req(req);
3268 spin_unlock(&req->rq_lock);
3271 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3272 struct ptlrpc_request *req =
3273 list_entry(tmp, struct ptlrpc_request, rq_list);
3275 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3277 spin_lock(&req->rq_lock);
3278 if (req->rq_import_generation < imp->imp_generation) {
3280 req->rq_status = -EIO;
3281 ptlrpc_client_wake_req(req);
3283 spin_unlock(&req->rq_lock);
3287 * Last chance to free reqs left on the replay list, but we
3288 * will still leak reqs that haven't committed.
3290 if (imp->imp_replayable)
3291 ptlrpc_free_committed(imp);
3297 * Abort all uncompleted requests in request set \a set
3299 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3301 struct list_head *tmp, *pos;
3303 LASSERT(set != NULL);
3305 list_for_each_safe(pos, tmp, &set->set_requests) {
3306 struct ptlrpc_request *req =
3307 list_entry(pos, struct ptlrpc_request,
3310 spin_lock(&req->rq_lock);
3311 if (req->rq_phase != RQ_PHASE_RPC) {
3312 spin_unlock(&req->rq_lock);
3317 req->rq_status = -EINTR;
3318 ptlrpc_client_wake_req(req);
3319 spin_unlock(&req->rq_lock);
3324 * Initialize the XID for the node. This is common among all requests on
3325 * this node, and only requires the property that it is monotonically
3326 * increasing. It does not need to be sequential. Since this is also used
3327 * as the RDMA match bits, it is important that a single client NOT have
3328 * the same match bits for two different in-flight requests, hence we do
3329 * NOT want to have an XID per target or similar.
3331 * To avoid an unlikely collision between match bits after a client reboot
3332 * (which would deliver old data into the wrong RDMA buffer) initialize
3333 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3334 * If the time is clearly incorrect, we instead use a 62-bit random number.
3335 * In the worst case the random number will overflow 1M RPCs per second in
3336 * 9133 years, or permutations thereof.
3338 #define YEAR_2004 (1ULL << 30)
3339 void ptlrpc_init_xid(void)
3341 time64_t now = ktime_get_real_seconds();
3344 if (now < YEAR_2004) {
3345 get_random_bytes(&xid, sizeof(xid));
3347 xid |= (1ULL << 61);
3349 xid = (u64)now << 20;
3352 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3353 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3355 xid &= PTLRPC_BULK_OPS_MASK;
3356 atomic64_set(&ptlrpc_last_xid, xid);
3360 * Increase xid and returns resulting new value to the caller.
3362 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3363 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3364 * itself uses the last bulk xid needed, so the server can determine the
3365 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3366 * xid must align to a power-of-two value.
3368 * This is assumed to be true due to the initial ptlrpc_last_xid
3369 * value also being initialized to a power-of-two value. LU-1431
3371 __u64 ptlrpc_next_xid(void)
3373 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3377 * If request has a new allocated XID (new request or EINPROGRESS resend),
3378 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3379 * request to ensure previous bulk fails and avoid problems with lost replies
3380 * and therefore several transfers landing into the same buffer from different
3383 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3385 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3387 LASSERT(bd != NULL);
3390 * Generate new matchbits for all resend requests, including
3393 if (req->rq_resend) {
3394 __u64 old_mbits = req->rq_mbits;
3397 * First time resend on -EINPROGRESS will generate new xid,
3398 * so we can actually use the rq_xid as rq_mbits in such case,
3399 * however, it's bit hard to distinguish such resend with a
3400 * 'resend for the -EINPROGRESS resend'. To make it simple,
3401 * we opt to generate mbits for all resend cases.
3403 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3405 req->rq_mbits = ptlrpc_next_xid();
3408 * Old version transfers rq_xid to peer as
3411 spin_lock(&req->rq_import->imp_lock);
3412 list_del_init(&req->rq_unreplied_list);
3413 ptlrpc_assign_next_xid_nolock(req);
3414 spin_unlock(&req->rq_import->imp_lock);
3415 req->rq_mbits = req->rq_xid;
3417 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3418 old_mbits, req->rq_mbits);
3419 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3420 /* Request being sent first time, use xid as matchbits. */
3421 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3422 || req->rq_mbits == 0) {
3423 req->rq_mbits = req->rq_xid;
3425 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3427 req->rq_mbits -= total_md - 1;
3431 * Replay request, xid and matchbits have already been
3432 * correctly assigned.
3438 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3439 * that server can infer the number of bulks that were prepared,
3442 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3446 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3447 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3449 * It's ok to directly set the rq_xid here, since this xid bump
3450 * won't affect the request position in unreplied list.
3452 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3453 req->rq_xid = req->rq_mbits;
3457 * Get a glimpse at what next xid value might have been.
3458 * Returns possible next xid.
3460 __u64 ptlrpc_sample_next_xid(void)
3462 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3464 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3467 * Functions for operating ptlrpc workers.
3469 * A ptlrpc work is a function which will be running inside ptlrpc context.
3470 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3472 * 1. after a work is created, it can be used many times, that is:
3473 * handler = ptlrpcd_alloc_work();
3474 * ptlrpcd_queue_work();
3476 * queue it again when necessary:
3477 * ptlrpcd_queue_work();
3478 * ptlrpcd_destroy_work();
3479 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3480 * it will only be queued once in any time. Also as its name implies, it may
3481 * have delay before it really runs by ptlrpcd thread.
3483 struct ptlrpc_work_async_args {
3484 int (*cb)(const struct lu_env *, void *);
3488 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3490 /* re-initialize the req */
3491 req->rq_timeout = obd_timeout;
3492 req->rq_sent = ktime_get_real_seconds();
3493 req->rq_deadline = req->rq_sent + req->rq_timeout;
3494 req->rq_phase = RQ_PHASE_INTERPRET;
3495 req->rq_next_phase = RQ_PHASE_COMPLETE;
3496 req->rq_xid = ptlrpc_next_xid();
3497 req->rq_import_generation = req->rq_import->imp_generation;
3499 ptlrpcd_add_req(req);
3502 static int work_interpreter(const struct lu_env *env,
3503 struct ptlrpc_request *req, void *args, int rc)
3505 struct ptlrpc_work_async_args *arg = args;
3507 LASSERT(ptlrpcd_check_work(req));
3508 LASSERT(arg->cb != NULL);
3510 rc = arg->cb(env, arg->cbdata);
3512 list_del_init(&req->rq_set_chain);
3515 if (atomic_dec_return(&req->rq_refcount) > 1) {
3516 atomic_set(&req->rq_refcount, 2);
3517 ptlrpcd_add_work_req(req);
3522 static int worker_format;
3524 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3526 return req->rq_pill.rc_fmt == (void *)&worker_format;
3530 * Create a work for ptlrpc.
3532 void *ptlrpcd_alloc_work(struct obd_import *imp,
3533 int (*cb)(const struct lu_env *, void *), void *cbdata)
3535 struct ptlrpc_request *req = NULL;
3536 struct ptlrpc_work_async_args *args;
3542 RETURN(ERR_PTR(-EINVAL));
3544 /* copy some code from deprecated fakereq. */
3545 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3547 CERROR("ptlrpc: run out of memory!\n");
3548 RETURN(ERR_PTR(-ENOMEM));
3551 ptlrpc_cli_req_init(req);
3553 req->rq_send_state = LUSTRE_IMP_FULL;
3554 req->rq_type = PTL_RPC_MSG_REQUEST;
3555 req->rq_import = class_import_get(imp);
3556 req->rq_interpret_reply = work_interpreter;
3557 /* don't want reply */
3558 req->rq_no_delay = req->rq_no_resend = 1;
3559 req->rq_pill.rc_fmt = (void *)&worker_format;
3561 args = ptlrpc_req_async_args(args, req);
3563 args->cbdata = cbdata;
3567 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3569 void ptlrpcd_destroy_work(void *handler)
3571 struct ptlrpc_request *req = handler;
3574 ptlrpc_req_finished(req);
3576 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3578 int ptlrpcd_queue_work(void *handler)
3580 struct ptlrpc_request *req = handler;
3583 * Check if the req is already being queued.
3585 * Here comes a trick: it lacks a way of checking if a req is being
3586 * processed reliably in ptlrpc. Here I have to use refcount of req
3587 * for this purpose. This is okay because the caller should use this
3588 * req as opaque data. - Jinshan
3590 LASSERT(atomic_read(&req->rq_refcount) > 0);
3591 if (atomic_inc_return(&req->rq_refcount) == 2)
3592 ptlrpcd_add_work_req(req);
3595 EXPORT_SYMBOL(ptlrpcd_queue_work);