4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 /** Implementation of client-side PortalRPC interfaces */
35 #define DEBUG_SUBSYSTEM S_RPC
37 #include <linux/delay.h>
38 #include <linux/random.h>
40 #include <obd_support.h>
41 #include <obd_class.h>
42 #include <lustre_lib.h>
43 #include <lustre_ha.h>
44 #include <lustre_import.h>
45 #include <lustre_req_layout.h>
47 #include "ptlrpc_internal.h"
49 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
50 struct page *page, int pageoffset,
53 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
56 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
57 struct page *page, int pageoffset,
60 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
63 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
67 for (i = 0; i < desc->bd_iov_count ; i++)
68 put_page(BD_GET_KIOV(desc, i).kiov_page);
71 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
72 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
73 .release_frags = ptlrpc_release_bulk_page_pin,
75 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
77 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
78 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
79 .release_frags = ptlrpc_release_bulk_noop,
81 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
83 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
84 .add_iov_frag = ptlrpc_prep_bulk_frag,
86 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
88 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
89 static int ptlrpcd_check_work(struct ptlrpc_request *req);
90 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
93 * Initialize passed in client structure \a cl.
95 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
96 struct ptlrpc_client *cl)
98 cl->cli_request_portal = req_portal;
99 cl->cli_reply_portal = rep_portal;
102 EXPORT_SYMBOL(ptlrpc_init_client);
105 * Return PortalRPC connection for remore uud \a uuid
107 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
108 lnet_nid_t nid4refnet)
110 struct ptlrpc_connection *c;
112 struct lnet_process_id peer;
116 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
117 * before accessing its values.
119 /* coverity[uninit_use_in_call] */
120 peer.nid = nid4refnet;
121 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
123 CNETERR("cannot find peer %s!\n", uuid->uuid);
127 c = ptlrpc_connection_get(peer, self, uuid);
129 memcpy(c->c_remote_uuid.uuid,
130 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
133 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
139 * Allocate and initialize new bulk descriptor on the sender.
140 * Returns pointer to the descriptor or NULL on error.
142 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
143 unsigned int max_brw,
144 enum ptlrpc_bulk_op_type type,
146 const struct ptlrpc_bulk_frag_ops *ops)
148 struct ptlrpc_bulk_desc *desc;
151 /* ensure that only one of KIOV or IOVEC is set but not both */
152 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
153 ops->add_kiov_frag != NULL) ||
154 (ptlrpc_is_bulk_desc_kvec(type) &&
155 ops->add_iov_frag != NULL));
160 if (type & PTLRPC_BULK_BUF_KIOV) {
161 OBD_ALLOC_LARGE(GET_KIOV(desc),
162 nfrags * sizeof(*GET_KIOV(desc)));
166 OBD_ALLOC_LARGE(GET_KVEC(desc),
167 nfrags * sizeof(*GET_KVEC(desc)));
172 spin_lock_init(&desc->bd_lock);
173 init_waitqueue_head(&desc->bd_waitq);
174 desc->bd_max_iov = nfrags;
175 desc->bd_iov_count = 0;
176 desc->bd_portal = portal;
177 desc->bd_type = type;
178 desc->bd_md_count = 0;
179 desc->bd_frag_ops = ops;
180 LASSERT(max_brw > 0);
181 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
183 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
184 * node. Negotiated ocd_brw_size will always be <= this number.
186 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
187 LNetInvalidateMDHandle(&desc->bd_mds[i]);
196 * Prepare bulk descriptor for specified outgoing request \a req that
197 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
198 * the bulk to be sent. Used on client-side.
199 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
202 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
204 unsigned int max_brw,
207 const struct ptlrpc_bulk_frag_ops
210 struct obd_import *imp = req->rq_import;
211 struct ptlrpc_bulk_desc *desc;
214 LASSERT(ptlrpc_is_bulk_op_passive(type));
216 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
220 desc->bd_import = class_import_get(imp);
223 desc->bd_cbid.cbid_fn = client_bulk_callback;
224 desc->bd_cbid.cbid_arg = desc;
226 /* This makes req own desc, and free it when she frees herself */
231 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
233 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
234 struct page *page, int pageoffset, int len,
239 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
240 LASSERT(page != NULL);
241 LASSERT(pageoffset >= 0);
243 LASSERT(pageoffset + len <= PAGE_SIZE);
244 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
246 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
253 kiov->kiov_page = page;
254 kiov->kiov_offset = pageoffset;
255 kiov->kiov_len = len;
257 desc->bd_iov_count++;
259 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
261 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
268 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
269 LASSERT(frag != NULL);
271 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
273 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
277 iovec->iov_base = frag;
278 iovec->iov_len = len;
280 desc->bd_iov_count++;
282 RETURN(desc->bd_nob);
284 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
286 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
290 LASSERT(desc != NULL);
291 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
292 LASSERT(desc->bd_md_count == 0); /* network hands off */
293 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
294 LASSERT(desc->bd_frag_ops != NULL);
296 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
297 sptlrpc_enc_pool_put_pages(desc);
300 class_export_put(desc->bd_export);
302 class_import_put(desc->bd_import);
304 if (desc->bd_frag_ops->release_frags != NULL)
305 desc->bd_frag_ops->release_frags(desc);
307 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
308 OBD_FREE_LARGE(GET_KIOV(desc),
309 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
311 OBD_FREE_LARGE(GET_KVEC(desc),
312 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
316 EXPORT_SYMBOL(ptlrpc_free_bulk);
319 * Set server timelimit for this req, i.e. how long are we willing to wait
320 * for reply before timing out this request.
322 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
328 LASSERT(req->rq_import);
331 /* non-AT settings */
333 * \a imp_server_timeout means this is reverse import and
334 * we send (currently only) ASTs to the client and cannot afford
335 * to wait too long for the reply, otherwise the other client
336 * (because of which we are sending this request) would
337 * timeout waiting for us
339 req->rq_timeout = req->rq_import->imp_server_timeout ?
340 obd_timeout / 2 : obd_timeout;
342 at = &req->rq_import->imp_at;
343 idx = import_at_get_index(req->rq_import,
344 req->rq_request_portal);
345 serv_est = at_get(&at->iat_service_estimate[idx]);
346 req->rq_timeout = at_est2timeout(serv_est);
349 * We could get even fancier here, using history to predict increased
354 * Let the server know what this RPC timeout is by putting it in the
357 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
359 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
361 /* Adjust max service estimate based on server value */
362 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
363 unsigned int serv_est)
369 LASSERT(req->rq_import);
370 at = &req->rq_import->imp_at;
372 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
374 * max service estimates are tracked on the server side,
375 * so just keep minimal history here
377 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
380 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
381 req->rq_import->imp_obd->obd_name,
382 req->rq_request_portal,
383 oldse, at_get(&at->iat_service_estimate[idx]));
386 /* Expected network latency per remote node (secs) */
387 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
389 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
392 /* Adjust expected network latency */
393 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
394 unsigned int service_time)
396 unsigned int nl, oldnl;
398 time64_t now = ktime_get_real_seconds();
400 LASSERT(req->rq_import);
402 if (service_time > now - req->rq_sent + 3) {
404 * b=16408, however, this can also happen if early reply
405 * is lost and client RPC is expired and resent, early reply
406 * or reply of original RPC can still be fit in reply buffer
407 * of resent RPC, now client is measuring time from the
408 * resent time, but server sent back service time of original
411 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
412 D_ADAPTTO : D_WARNING,
413 "Reported service time %u > total measured time %lld\n",
414 service_time, now - req->rq_sent);
418 /* Network latency is total time less server processing time */
419 nl = max_t(int, now - req->rq_sent -
420 service_time, 0) + 1; /* st rounding */
421 at = &req->rq_import->imp_at;
423 oldnl = at_measured(&at->iat_net_latency, nl);
426 "The network latency for %s (nid %s) has changed from %d to %d\n",
427 req->rq_import->imp_obd->obd_name,
428 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
429 oldnl, at_get(&at->iat_net_latency));
432 static int unpack_reply(struct ptlrpc_request *req)
436 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
437 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
439 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
445 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
447 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
455 * Handle an early reply message, called with the rq_lock held.
456 * If anything goes wrong just ignore it - same as if it never happened
458 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
459 __must_hold(&req->rq_lock)
461 struct ptlrpc_request *early_req;
467 spin_unlock(&req->rq_lock);
469 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
471 spin_lock(&req->rq_lock);
475 rc = unpack_reply(early_req);
477 sptlrpc_cli_finish_early_reply(early_req);
478 spin_lock(&req->rq_lock);
483 * Use new timeout value just to adjust the local value for this
484 * request, don't include it into at_history. It is unclear yet why
485 * service time increased and should it be counted or skipped, e.g.
486 * that can be recovery case or some error or server, the real reply
487 * will add all new data if it is worth to add.
489 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
490 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
492 /* Network latency can be adjusted, it is pure network delays */
493 ptlrpc_at_adj_net_latency(req,
494 lustre_msg_get_service_time(early_req->rq_repmsg));
496 sptlrpc_cli_finish_early_reply(early_req);
498 spin_lock(&req->rq_lock);
499 olddl = req->rq_deadline;
501 * server assumes it now has rq_timeout from when the request
502 * arrived, so the client should give it at least that long.
503 * since we don't know the arrival time we'll use the original
506 req->rq_deadline = req->rq_sent + req->rq_timeout +
507 ptlrpc_at_get_net_latency(req);
509 /* The below message is checked in replay-single.sh test_65{a,b} */
510 /* The below message is checked in sanity-{gss,krb5} test_8 */
511 DEBUG_REQ(D_ADAPTTO, req,
512 "Early reply #%d, new deadline in %llds (%llds)",
514 req->rq_deadline - ktime_get_real_seconds(),
515 req->rq_deadline - olddl);
520 static struct kmem_cache *request_cache;
522 int ptlrpc_request_cache_init(void)
524 request_cache = kmem_cache_create("ptlrpc_cache",
525 sizeof(struct ptlrpc_request),
526 0, SLAB_HWCACHE_ALIGN, NULL);
527 return request_cache ? 0 : -ENOMEM;
530 void ptlrpc_request_cache_fini(void)
532 kmem_cache_destroy(request_cache);
535 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
537 struct ptlrpc_request *req;
539 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
543 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
545 OBD_SLAB_FREE_PTR(req, request_cache);
549 * Wind down request pool \a pool.
550 * Frees all requests from the pool too
552 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
554 struct list_head *l, *tmp;
555 struct ptlrpc_request *req;
557 LASSERT(pool != NULL);
559 spin_lock(&pool->prp_lock);
560 list_for_each_safe(l, tmp, &pool->prp_req_list) {
561 req = list_entry(l, struct ptlrpc_request, rq_list);
562 list_del(&req->rq_list);
563 LASSERT(req->rq_reqbuf);
564 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
565 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
566 ptlrpc_request_cache_free(req);
568 spin_unlock(&pool->prp_lock);
569 OBD_FREE(pool, sizeof(*pool));
571 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
574 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
576 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
581 while (size < pool->prp_rq_size)
584 LASSERTF(list_empty(&pool->prp_req_list) ||
585 size == pool->prp_rq_size,
586 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
587 pool->prp_rq_size, size);
589 pool->prp_rq_size = size;
590 for (i = 0; i < num_rq; i++) {
591 struct ptlrpc_request *req;
592 struct lustre_msg *msg;
594 req = ptlrpc_request_cache_alloc(GFP_NOFS);
597 OBD_ALLOC_LARGE(msg, size);
599 ptlrpc_request_cache_free(req);
602 req->rq_reqbuf = msg;
603 req->rq_reqbuf_len = size;
605 spin_lock(&pool->prp_lock);
606 list_add_tail(&req->rq_list, &pool->prp_req_list);
607 spin_unlock(&pool->prp_lock);
611 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
614 * Create and initialize new request pool with given attributes:
615 * \a num_rq - initial number of requests to create for the pool
616 * \a msgsize - maximum message size possible for requests in thid pool
617 * \a populate_pool - function to be called when more requests need to be added
619 * Returns pointer to newly created pool or NULL on error.
621 struct ptlrpc_request_pool *
622 ptlrpc_init_rq_pool(int num_rq, int msgsize,
623 int (*populate_pool)(struct ptlrpc_request_pool *, int))
625 struct ptlrpc_request_pool *pool;
627 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
632 * Request next power of two for the allocation, because internally
633 * kernel would do exactly this
635 spin_lock_init(&pool->prp_lock);
636 INIT_LIST_HEAD(&pool->prp_req_list);
637 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
638 pool->prp_populate = populate_pool;
640 populate_pool(pool, num_rq);
644 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
647 * Fetches one request from pool \a pool
649 static struct ptlrpc_request *
650 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
652 struct ptlrpc_request *request;
653 struct lustre_msg *reqbuf;
658 spin_lock(&pool->prp_lock);
661 * See if we have anything in a pool, and bail out if nothing,
662 * in writeout path, where this matters, this is safe to do, because
663 * nothing is lost in this case, and when some in-flight requests
664 * complete, this code will be called again.
666 if (unlikely(list_empty(&pool->prp_req_list))) {
667 spin_unlock(&pool->prp_lock);
671 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
673 list_del_init(&request->rq_list);
674 spin_unlock(&pool->prp_lock);
676 LASSERT(request->rq_reqbuf);
677 LASSERT(request->rq_pool);
679 reqbuf = request->rq_reqbuf;
680 memset(request, 0, sizeof(*request));
681 request->rq_reqbuf = reqbuf;
682 request->rq_reqbuf_len = pool->prp_rq_size;
683 request->rq_pool = pool;
689 * Returns freed \a request to pool.
691 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
693 struct ptlrpc_request_pool *pool = request->rq_pool;
695 spin_lock(&pool->prp_lock);
696 LASSERT(list_empty(&request->rq_list));
697 LASSERT(!request->rq_receiving_reply);
698 list_add_tail(&request->rq_list, &pool->prp_req_list);
699 spin_unlock(&pool->prp_lock);
702 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
704 struct obd_import *imp = req->rq_import;
705 struct list_head *tmp;
706 struct ptlrpc_request *iter;
708 assert_spin_locked(&imp->imp_lock);
709 LASSERT(list_empty(&req->rq_unreplied_list));
711 /* unreplied list is sorted by xid in ascending order */
712 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
713 iter = list_entry(tmp, struct ptlrpc_request,
716 LASSERT(req->rq_xid != iter->rq_xid);
717 if (req->rq_xid < iter->rq_xid)
719 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
722 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
725 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
727 req->rq_xid = ptlrpc_next_xid();
728 ptlrpc_add_unreplied(req);
731 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
733 spin_lock(&req->rq_import->imp_lock);
734 ptlrpc_assign_next_xid_nolock(req);
735 spin_unlock(&req->rq_import->imp_lock);
738 static atomic64_t ptlrpc_last_xid;
740 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
742 spin_lock(&req->rq_import->imp_lock);
743 list_del_init(&req->rq_unreplied_list);
744 ptlrpc_assign_next_xid_nolock(req);
745 spin_unlock(&req->rq_import->imp_lock);
746 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
749 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
751 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
755 opc = lustre_msg_get_opc(req->rq_reqmsg);
756 tag = obd_get_mod_rpc_slot(cli, opc);
757 lustre_msg_set_tag(req->rq_reqmsg, tag);
758 ptlrpc_reassign_next_xid(req);
760 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
762 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
764 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
767 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
768 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
770 obd_put_mod_rpc_slot(cli, opc, tag);
773 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
775 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
776 __u32 version, int opcode, char **bufs,
777 struct ptlrpc_cli_ctx *ctx)
780 struct obd_import *imp;
786 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
787 imp = request->rq_import;
788 lengths = request->rq_pill.rc_area[RCL_CLIENT];
791 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
793 rc = sptlrpc_req_get_ctx(request);
797 sptlrpc_req_set_flavor(request, opcode);
799 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
804 lustre_msg_add_version(request->rq_reqmsg, version);
805 request->rq_send_state = LUSTRE_IMP_FULL;
806 request->rq_type = PTL_RPC_MSG_REQUEST;
808 request->rq_req_cbid.cbid_fn = request_out_callback;
809 request->rq_req_cbid.cbid_arg = request;
811 request->rq_reply_cbid.cbid_fn = reply_in_callback;
812 request->rq_reply_cbid.cbid_arg = request;
814 request->rq_reply_deadline = 0;
815 request->rq_bulk_deadline = 0;
816 request->rq_req_deadline = 0;
817 request->rq_phase = RQ_PHASE_NEW;
818 request->rq_next_phase = RQ_PHASE_UNDEFINED;
820 request->rq_request_portal = imp->imp_client->cli_request_portal;
821 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
823 ptlrpc_at_set_req_timeout(request);
825 lustre_msg_set_opc(request->rq_reqmsg, opcode);
827 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
828 if (cfs_fail_val == opcode) {
829 time64_t *fail_t = NULL, *fail2_t = NULL;
831 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
832 fail_t = &request->rq_bulk_deadline;
833 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
834 fail_t = &request->rq_reply_deadline;
835 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
836 fail_t = &request->rq_req_deadline;
837 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
838 fail_t = &request->rq_reply_deadline;
839 fail2_t = &request->rq_bulk_deadline;
840 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
841 time64_t now = ktime_get_real_seconds();
842 u64 xid = ((u64)now >> 4) << 24;
844 atomic64_set(&ptlrpc_last_xid, xid);
848 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
851 *fail2_t = ktime_get_real_seconds() +
855 * The RPC is infected, let the test to change the
858 msleep(4 * MSEC_PER_SEC);
861 ptlrpc_assign_next_xid(request);
866 LASSERT(!request->rq_pool);
867 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
869 class_import_put(imp);
873 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
876 * Pack request buffers for network transfer, performing necessary encryption
877 * steps if necessary.
879 int ptlrpc_request_pack(struct ptlrpc_request *request,
880 __u32 version, int opcode)
882 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
884 EXPORT_SYMBOL(ptlrpc_request_pack);
887 * Helper function to allocate new request on import \a imp
888 * and possibly using existing request from pool \a pool if provided.
889 * Returns allocated request structure with import field filled or
893 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
894 struct ptlrpc_request_pool *pool)
896 struct ptlrpc_request *request = NULL;
898 request = ptlrpc_request_cache_alloc(GFP_NOFS);
900 if (!request && pool)
901 request = ptlrpc_prep_req_from_pool(pool);
904 ptlrpc_cli_req_init(request);
906 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
907 LASSERT(imp != LP_POISON);
908 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
910 LASSERT(imp->imp_client != LP_POISON);
912 request->rq_import = class_import_get(imp);
914 CERROR("request allocation out of memory\n");
921 * Helper function for creating a request.
922 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
923 * buffer structures according to capsule template \a format.
924 * Returns allocated request structure pointer or NULL on error.
926 static struct ptlrpc_request *
927 ptlrpc_request_alloc_internal(struct obd_import *imp,
928 struct ptlrpc_request_pool *pool,
929 const struct req_format *format)
931 struct ptlrpc_request *request;
933 request = __ptlrpc_request_alloc(imp, pool);
938 * initiate connection if needed when the import has been
939 * referenced by the new request to avoid races with disconnect
941 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
944 CDEBUG_LIMIT(imp->imp_idle_debug,
945 "%s: reconnect after %llds idle\n",
946 imp->imp_obd->obd_name, ktime_get_real_seconds() -
947 imp->imp_last_reply_time);
948 spin_lock(&imp->imp_lock);
949 if (imp->imp_state == LUSTRE_IMP_IDLE) {
950 imp->imp_generation++;
951 imp->imp_initiated_at = imp->imp_generation;
952 imp->imp_state = LUSTRE_IMP_NEW;
954 /* connect_import_locked releases imp_lock */
955 rc = ptlrpc_connect_import_locked(imp);
957 ptlrpc_request_free(request);
960 ptlrpc_pinger_add_import(imp);
962 spin_unlock(&imp->imp_lock);
966 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
967 req_capsule_set(&request->rq_pill, format);
972 * Allocate new request structure for import \a imp and initialize its
973 * buffer structure according to capsule template \a format.
975 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
976 const struct req_format *format)
978 return ptlrpc_request_alloc_internal(imp, NULL, format);
980 EXPORT_SYMBOL(ptlrpc_request_alloc);
983 * Allocate new request structure for import \a imp from pool \a pool and
984 * initialize its buffer structure according to capsule template \a format.
986 struct ptlrpc_request *
987 ptlrpc_request_alloc_pool(struct obd_import *imp,
988 struct ptlrpc_request_pool *pool,
989 const struct req_format *format)
991 return ptlrpc_request_alloc_internal(imp, pool, format);
993 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
996 * For requests not from pool, free memory of the request structure.
997 * For requests obtained from a pool earlier, return request back to pool.
999 void ptlrpc_request_free(struct ptlrpc_request *request)
1001 if (request->rq_pool)
1002 __ptlrpc_free_req_to_pool(request);
1004 ptlrpc_request_cache_free(request);
1006 EXPORT_SYMBOL(ptlrpc_request_free);
1009 * Allocate new request for operatione \a opcode and immediatelly pack it for
1011 * Only used for simple requests like OBD_PING where the only important
1012 * part of the request is operation itself.
1013 * Returns allocated request or NULL on error.
1015 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1016 const struct req_format *format,
1017 __u32 version, int opcode)
1019 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1023 rc = ptlrpc_request_pack(req, version, opcode);
1025 ptlrpc_request_free(req);
1031 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1034 * Allocate and initialize new request set structure on the current CPT.
1035 * Returns a pointer to the newly allocated set structure or NULL on error.
1037 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1039 struct ptlrpc_request_set *set;
1043 cpt = cfs_cpt_current(cfs_cpt_table, 0);
1044 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1047 atomic_set(&set->set_refcount, 1);
1048 INIT_LIST_HEAD(&set->set_requests);
1049 init_waitqueue_head(&set->set_waitq);
1050 atomic_set(&set->set_new_count, 0);
1051 atomic_set(&set->set_remaining, 0);
1052 spin_lock_init(&set->set_new_req_lock);
1053 INIT_LIST_HEAD(&set->set_new_requests);
1054 set->set_max_inflight = UINT_MAX;
1055 set->set_producer = NULL;
1056 set->set_producer_arg = NULL;
1061 EXPORT_SYMBOL(ptlrpc_prep_set);
1064 * Allocate and initialize new request set structure with flow control
1065 * extension. This extension allows to control the number of requests in-flight
1066 * for the whole set. A callback function to generate requests must be provided
1067 * and the request set will keep the number of requests sent over the wire to
1069 * Returns a pointer to the newly allocated set structure or NULL on error.
1071 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1075 struct ptlrpc_request_set *set;
1077 set = ptlrpc_prep_set();
1081 set->set_max_inflight = max;
1082 set->set_producer = func;
1083 set->set_producer_arg = arg;
1089 * Wind down and free request set structure previously allocated with
1091 * Ensures that all requests on the set have completed and removes
1092 * all requests from the request list in a set.
1093 * If any unsent request happen to be on the list, pretends that they got
1094 * an error in flight and calls their completion handler.
1096 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1098 struct list_head *tmp;
1099 struct list_head *next;
1105 /* Requests on the set should either all be completed, or all be new */
1106 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1107 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1108 list_for_each(tmp, &set->set_requests) {
1109 struct ptlrpc_request *req =
1110 list_entry(tmp, struct ptlrpc_request,
1113 LASSERT(req->rq_phase == expected_phase);
1117 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1118 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1119 atomic_read(&set->set_remaining), n);
1121 list_for_each_safe(tmp, next, &set->set_requests) {
1122 struct ptlrpc_request *req =
1123 list_entry(tmp, struct ptlrpc_request,
1125 list_del_init(&req->rq_set_chain);
1127 LASSERT(req->rq_phase == expected_phase);
1129 if (req->rq_phase == RQ_PHASE_NEW) {
1130 ptlrpc_req_interpret(NULL, req, -EBADR);
1131 atomic_dec(&set->set_remaining);
1134 spin_lock(&req->rq_lock);
1136 req->rq_invalid_rqset = 0;
1137 spin_unlock(&req->rq_lock);
1139 ptlrpc_req_finished(req);
1142 LASSERT(atomic_read(&set->set_remaining) == 0);
1144 ptlrpc_reqset_put(set);
1147 EXPORT_SYMBOL(ptlrpc_set_destroy);
1150 * Add a new request to the general purpose request set.
1151 * Assumes request reference from the caller.
1153 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1154 struct ptlrpc_request *req)
1156 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1157 LASSERT(list_empty(&req->rq_set_chain));
1159 if (req->rq_allow_intr)
1160 set->set_allow_intr = 1;
1162 /* The set takes over the caller's request reference */
1163 list_add_tail(&req->rq_set_chain, &set->set_requests);
1165 atomic_inc(&set->set_remaining);
1166 req->rq_queued_time = ktime_get_seconds();
1169 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1171 if (set->set_producer)
1173 * If the request set has a producer callback, the RPC must be
1174 * sent straight away
1176 ptlrpc_send_new_req(req);
1178 EXPORT_SYMBOL(ptlrpc_set_add_req);
1181 * Add a request to a request with dedicated server thread
1182 * and wake the thread to make any necessary processing.
1183 * Currently only used for ptlrpcd.
1185 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1186 struct ptlrpc_request *req)
1188 struct ptlrpc_request_set *set = pc->pc_set;
1191 LASSERT(req->rq_set == NULL);
1192 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1194 spin_lock(&set->set_new_req_lock);
1196 * The set takes over the caller's request reference.
1199 req->rq_queued_time = ktime_get_seconds();
1200 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1201 count = atomic_inc_return(&set->set_new_count);
1202 spin_unlock(&set->set_new_req_lock);
1204 /* Only need to call wakeup once for the first entry. */
1206 wake_up(&set->set_waitq);
1209 * XXX: It maybe unnecessary to wakeup all the partners. But to
1210 * guarantee the async RPC can be processed ASAP, we have
1211 * no other better choice. It maybe fixed in future.
1213 for (i = 0; i < pc->pc_npartners; i++)
1214 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1219 * Based on the current state of the import, determine if the request
1220 * can be sent, is an error, or should be delayed.
1222 * Returns true if this request should be delayed. If false, and
1223 * *status is set, then the request can not be sent and *status is the
1224 * error code. If false and status is 0, then request can be sent.
1226 * The imp->imp_lock must be held.
1228 static int ptlrpc_import_delay_req(struct obd_import *imp,
1229 struct ptlrpc_request *req, int *status)
1237 if (req->rq_ctx_init || req->rq_ctx_fini) {
1238 /* always allow ctx init/fini rpc go through */
1239 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1240 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1242 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1243 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1246 * pings or MDS-equivalent STATFS may safely
1249 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1250 D_HA : D_ERROR, req, "IMP_CLOSED");
1252 } else if (ptlrpc_send_limit_expired(req)) {
1253 /* probably doesn't need to be a D_ERROR afterinitial testing */
1254 DEBUG_REQ(D_HA, req, "send limit expired");
1255 *status = -ETIMEDOUT;
1256 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1257 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1258 ;/* allow CONNECT even if import is invalid */
1259 if (atomic_read(&imp->imp_inval_count) != 0) {
1260 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1263 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1264 if (!imp->imp_deactive)
1265 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1266 *status = -ESHUTDOWN; /* b=12940 */
1267 } else if (req->rq_import_generation != imp->imp_generation) {
1268 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1270 } else if (req->rq_send_state != imp->imp_state) {
1271 /* invalidate in progress - any requests should be drop */
1272 if (atomic_read(&imp->imp_inval_count) != 0) {
1273 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1275 } else if (req->rq_no_delay &&
1276 imp->imp_generation != imp->imp_initiated_at) {
1277 /* ignore nodelay for requests initiating connections */
1278 *status = -EWOULDBLOCK;
1279 } else if (req->rq_allow_replay &&
1280 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1281 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1282 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1283 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1284 DEBUG_REQ(D_HA, req, "allow during recovery");
1294 * Decide if the error message should be printed to the console or not.
1295 * Makes its decision based on request type, status, and failure frequency.
1297 * \param[in] req request that failed and may need a console message
1299 * \retval false if no message should be printed
1300 * \retval true if console message should be printed
1302 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1304 LASSERT(req->rq_reqmsg != NULL);
1306 /* Suppress particular reconnect errors which are to be expected. */
1307 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1308 /* Suppress timed out reconnect requests */
1309 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1314 * Suppress most unavailable/again reconnect requests, but
1315 * print occasionally so it is clear client is trying to
1316 * connect to a server where no target is running.
1318 if ((err == -ENODEV || err == -EAGAIN) &&
1319 req->rq_import->imp_conn_cnt % 30 != 20)
1323 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1324 /* -EAGAIN is normal when using POSIX flocks */
1327 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1328 (req->rq_xid & 0xf) != 10)
1329 /* Suppress most ping requests, they may fail occasionally */
1336 * Check request processing status.
1337 * Returns the status.
1339 static int ptlrpc_check_status(struct ptlrpc_request *req)
1344 rc = lustre_msg_get_status(req->rq_repmsg);
1345 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1346 struct obd_import *imp = req->rq_import;
1347 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1348 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1350 if (ptlrpc_console_allow(req, opc, rc))
1351 LCONSOLE_ERROR_MSG(0x11,
1352 "%s: operation %s to node %s failed: rc = %d\n",
1353 imp->imp_obd->obd_name,
1355 libcfs_nid2str(nid), rc);
1356 RETURN(rc < 0 ? rc : -EINVAL);
1360 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1366 * save pre-versions of objects into request for replay.
1367 * Versions are obtained from server reply.
1370 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1372 struct lustre_msg *repmsg = req->rq_repmsg;
1373 struct lustre_msg *reqmsg = req->rq_reqmsg;
1374 __u64 *versions = lustre_msg_get_versions(repmsg);
1377 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1381 lustre_msg_set_versions(reqmsg, versions);
1382 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1383 versions[0], versions[1]);
1388 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1390 struct ptlrpc_request *req;
1392 assert_spin_locked(&imp->imp_lock);
1393 if (list_empty(&imp->imp_unreplied_list))
1396 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1398 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1400 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1401 imp->imp_known_replied_xid = req->rq_xid - 1;
1403 return req->rq_xid - 1;
1407 * Callback function called when client receives RPC reply for \a req.
1408 * Returns 0 on success or error code.
1409 * The return alue would be assigned to req->rq_status by the caller
1410 * as request processing status.
1411 * This function also decides if the request needs to be saved for later replay.
1413 static int after_reply(struct ptlrpc_request *req)
1415 struct obd_import *imp = req->rq_import;
1416 struct obd_device *obd = req->rq_import->imp_obd;
1423 LASSERT(obd != NULL);
1424 /* repbuf must be unlinked */
1425 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1427 if (req->rq_reply_truncated) {
1428 if (ptlrpc_no_resend(req)) {
1429 DEBUG_REQ(D_ERROR, req,
1430 "reply buffer overflow, expected=%d, actual size=%d",
1431 req->rq_nob_received, req->rq_repbuf_len);
1435 sptlrpc_cli_free_repbuf(req);
1437 * Pass the required reply buffer size (include
1438 * space for early reply).
1439 * NB: no need to roundup because alloc_repbuf
1442 req->rq_replen = req->rq_nob_received;
1443 req->rq_nob_received = 0;
1444 spin_lock(&req->rq_lock);
1446 spin_unlock(&req->rq_lock);
1450 work_start = ktime_get_real();
1451 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1454 * NB Until this point, the whole of the incoming message,
1455 * including buflens, status etc is in the sender's byte order.
1457 rc = sptlrpc_cli_unwrap_reply(req);
1459 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1464 * Security layer unwrap might ask resend this request.
1469 rc = unpack_reply(req);
1473 /* retry indefinitely on EINPROGRESS */
1474 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1475 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1476 time64_t now = ktime_get_real_seconds();
1478 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1479 D_RPCTRACE, req, "resending request on EINPROGRESS");
1480 spin_lock(&req->rq_lock);
1482 spin_unlock(&req->rq_lock);
1483 req->rq_nr_resend++;
1485 /* Readjust the timeout for current conditions */
1486 ptlrpc_at_set_req_timeout(req);
1488 * delay resend to give a chance to the server to get ready.
1489 * The delay is increased by 1s on every resend and is capped to
1490 * the current request timeout (i.e. obd_timeout if AT is off,
1491 * or AT service time x 125% + 5s, see at_est2timeout)
1493 if (req->rq_nr_resend > req->rq_timeout)
1494 req->rq_sent = now + req->rq_timeout;
1496 req->rq_sent = now + req->rq_nr_resend;
1498 /* Resend for EINPROGRESS will use a new XID */
1499 spin_lock(&imp->imp_lock);
1500 list_del_init(&req->rq_unreplied_list);
1501 spin_unlock(&imp->imp_lock);
1506 if (obd->obd_svc_stats) {
1507 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1509 ptlrpc_lprocfs_rpc_sent(req, timediff);
1512 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1513 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1514 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1515 lustre_msg_get_type(req->rq_repmsg));
1519 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1520 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1521 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1522 ptlrpc_at_adj_net_latency(req,
1523 lustre_msg_get_service_time(req->rq_repmsg));
1525 rc = ptlrpc_check_status(req);
1529 * Either we've been evicted, or the server has failed for
1530 * some reason. Try to reconnect, and if that fails, punt to
1533 if (ptlrpc_recoverable_error(rc)) {
1534 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1535 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1538 ptlrpc_request_handle_notconn(req);
1543 * Let's look if server sent slv. Do it only for RPC with
1546 ldlm_cli_update_pool(req);
1550 * Store transno in reqmsg for replay.
1552 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1553 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1554 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1557 if (imp->imp_replayable) {
1558 spin_lock(&imp->imp_lock);
1560 * No point in adding already-committed requests to the replay
1561 * list, we will just remove them immediately. b=9829
1563 if (req->rq_transno != 0 &&
1565 lustre_msg_get_last_committed(req->rq_repmsg) ||
1567 /** version recovery */
1568 ptlrpc_save_versions(req);
1569 ptlrpc_retain_replayable_request(req, imp);
1570 } else if (req->rq_commit_cb &&
1571 list_empty(&req->rq_replay_list)) {
1573 * NB: don't call rq_commit_cb if it's already on
1574 * rq_replay_list, ptlrpc_free_committed() will call
1575 * it later, see LU-3618 for details
1577 spin_unlock(&imp->imp_lock);
1578 req->rq_commit_cb(req);
1579 spin_lock(&imp->imp_lock);
1583 * Replay-enabled imports return commit-status information.
1585 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1586 if (likely(committed > imp->imp_peer_committed_transno))
1587 imp->imp_peer_committed_transno = committed;
1589 ptlrpc_free_committed(imp);
1591 if (!list_empty(&imp->imp_replay_list)) {
1592 struct ptlrpc_request *last;
1594 last = list_entry(imp->imp_replay_list.prev,
1595 struct ptlrpc_request,
1598 * Requests with rq_replay stay on the list even if no
1599 * commit is expected.
1601 if (last->rq_transno > imp->imp_peer_committed_transno)
1602 ptlrpc_pinger_commit_expected(imp);
1605 spin_unlock(&imp->imp_lock);
1612 * Helper function to send request \a req over the network for the first time
1613 * Also adjusts request phase.
1614 * Returns 0 on success or error code.
1616 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1618 struct obd_import *imp = req->rq_import;
1623 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1625 /* do not try to go further if there is not enough memory in enc_pool */
1626 if (req->rq_sent && req->rq_bulk)
1627 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1628 pool_is_at_full_capacity())
1631 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1632 (!req->rq_generation_set ||
1633 req->rq_import_generation == imp->imp_generation))
1636 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1638 spin_lock(&imp->imp_lock);
1640 LASSERT(req->rq_xid != 0);
1641 LASSERT(!list_empty(&req->rq_unreplied_list));
1643 if (!req->rq_generation_set)
1644 req->rq_import_generation = imp->imp_generation;
1646 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1647 spin_lock(&req->rq_lock);
1648 req->rq_waiting = 1;
1649 spin_unlock(&req->rq_lock);
1651 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1652 ptlrpc_import_state_name(req->rq_send_state),
1653 ptlrpc_import_state_name(imp->imp_state));
1654 LASSERT(list_empty(&req->rq_list));
1655 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1656 atomic_inc(&req->rq_import->imp_inflight);
1657 spin_unlock(&imp->imp_lock);
1662 spin_unlock(&imp->imp_lock);
1663 req->rq_status = rc;
1664 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1668 LASSERT(list_empty(&req->rq_list));
1669 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1670 atomic_inc(&req->rq_import->imp_inflight);
1673 * find the known replied XID from the unreplied list, CONNECT
1674 * and DISCONNECT requests are skipped to make the sanity check
1675 * on server side happy. see process_req_last_xid().
1677 * For CONNECT: Because replay requests have lower XID, it'll
1678 * break the sanity check if CONNECT bump the exp_last_xid on
1681 * For DISCONNECT: Since client will abort inflight RPC before
1682 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1683 * than the inflight RPC.
1685 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1686 min_xid = ptlrpc_known_replied_xid(imp);
1687 spin_unlock(&imp->imp_lock);
1689 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1691 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1693 rc = sptlrpc_req_refresh_ctx(req, -1);
1696 req->rq_status = rc;
1699 spin_lock(&req->rq_lock);
1700 req->rq_wait_ctx = 1;
1701 spin_unlock(&req->rq_lock);
1707 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1708 req, current_comm(),
1709 imp->imp_obd->obd_uuid.uuid,
1710 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1711 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1712 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1714 rc = ptl_send_rpc(req, 0);
1715 if (rc == -ENOMEM) {
1716 spin_lock(&imp->imp_lock);
1717 if (!list_empty(&req->rq_list)) {
1718 list_del_init(&req->rq_list);
1719 atomic_dec(&req->rq_import->imp_inflight);
1721 spin_unlock(&imp->imp_lock);
1722 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1726 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1728 spin_lock(&req->rq_lock);
1729 req->rq_net_err = 1;
1730 spin_unlock(&req->rq_lock);
1736 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1741 LASSERT(set->set_producer != NULL);
1743 remaining = atomic_read(&set->set_remaining);
1746 * populate the ->set_requests list with requests until we
1747 * reach the maximum number of RPCs in flight for this set
1749 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1750 rc = set->set_producer(set, set->set_producer_arg);
1751 if (rc == -ENOENT) {
1752 /* no more RPC to produce */
1753 set->set_producer = NULL;
1754 set->set_producer_arg = NULL;
1759 RETURN((atomic_read(&set->set_remaining) - remaining));
1763 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1764 * and no more replies are expected.
1765 * (it is possible to get less replies than requests sent e.g. due to timed out
1766 * requests or requests that we had trouble to send out)
1768 * NOTE: This function contains a potential schedule point (cond_resched()).
1770 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1772 struct list_head *tmp, *next;
1773 struct list_head comp_reqs;
1774 int force_timer_recalc = 0;
1777 if (atomic_read(&set->set_remaining) == 0)
1780 INIT_LIST_HEAD(&comp_reqs);
1781 list_for_each_safe(tmp, next, &set->set_requests) {
1782 struct ptlrpc_request *req =
1783 list_entry(tmp, struct ptlrpc_request,
1785 struct obd_import *imp = req->rq_import;
1786 int unregistered = 0;
1790 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1791 list_move_tail(&req->rq_set_chain, &comp_reqs);
1796 * This schedule point is mainly for the ptlrpcd caller of this
1797 * function. Most ptlrpc sets are not long-lived and unbounded
1798 * in length, but at the least the set used by the ptlrpcd is.
1799 * Since the processing time is unbounded, we need to insert an
1800 * explicit schedule point to make the thread well-behaved.
1805 * If the caller requires to allow to be interpreted by force
1806 * and it has really been interpreted, then move the request
1807 * to RQ_PHASE_INTERPRET phase in spite of what the current
1810 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1811 req->rq_status = -EINTR;
1812 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1815 * Since it is interpreted and we have to wait for
1816 * the reply to be unlinked, then use sync mode.
1820 GOTO(interpret, req->rq_status);
1823 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1824 force_timer_recalc = 1;
1826 /* delayed send - skip */
1827 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1830 /* delayed resend - skip */
1831 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1832 req->rq_sent > ktime_get_real_seconds())
1835 if (!(req->rq_phase == RQ_PHASE_RPC ||
1836 req->rq_phase == RQ_PHASE_BULK ||
1837 req->rq_phase == RQ_PHASE_INTERPRET ||
1838 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1839 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1840 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1844 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1845 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1846 LASSERT(req->rq_next_phase != req->rq_phase);
1847 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1849 if (req->rq_req_deadline &&
1850 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1851 req->rq_req_deadline = 0;
1852 if (req->rq_reply_deadline &&
1853 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1854 req->rq_reply_deadline = 0;
1855 if (req->rq_bulk_deadline &&
1856 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1857 req->rq_bulk_deadline = 0;
1860 * Skip processing until reply is unlinked. We
1861 * can't return to pool before that and we can't
1862 * call interpret before that. We need to make
1863 * sure that all rdma transfers finished and will
1864 * not corrupt any data.
1866 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1867 ptlrpc_client_recv_or_unlink(req))
1869 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1870 ptlrpc_client_bulk_active(req))
1874 * Turn fail_loc off to prevent it from looping
1877 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1878 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1881 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1882 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1887 * Move to next phase if reply was successfully
1890 ptlrpc_rqphase_move(req, req->rq_next_phase);
1893 if (req->rq_phase == RQ_PHASE_INTERPRET)
1894 GOTO(interpret, req->rq_status);
1897 * Note that this also will start async reply unlink.
1899 if (req->rq_net_err && !req->rq_timedout) {
1900 ptlrpc_expire_one_request(req, 1);
1903 * Check if we still need to wait for unlink.
1905 if (ptlrpc_client_recv_or_unlink(req) ||
1906 ptlrpc_client_bulk_active(req))
1908 /* If there is no need to resend, fail it now. */
1909 if (req->rq_no_resend) {
1910 if (req->rq_status == 0)
1911 req->rq_status = -EIO;
1912 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1913 GOTO(interpret, req->rq_status);
1920 spin_lock(&req->rq_lock);
1921 req->rq_replied = 0;
1922 spin_unlock(&req->rq_lock);
1923 if (req->rq_status == 0)
1924 req->rq_status = -EIO;
1925 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1926 GOTO(interpret, req->rq_status);
1930 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1931 * so it sets rq_intr regardless of individual rpc
1932 * timeouts. The synchronous IO waiting path sets
1933 * rq_intr irrespective of whether ptlrpcd
1934 * has seen a timeout. Our policy is to only interpret
1935 * interrupted rpcs after they have timed out, so we
1936 * need to enforce that here.
1939 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1940 req->rq_wait_ctx)) {
1941 req->rq_status = -EINTR;
1942 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1943 GOTO(interpret, req->rq_status);
1946 if (req->rq_phase == RQ_PHASE_RPC) {
1947 if (req->rq_timedout || req->rq_resend ||
1948 req->rq_waiting || req->rq_wait_ctx) {
1951 if (!ptlrpc_unregister_reply(req, 1)) {
1952 ptlrpc_unregister_bulk(req, 1);
1956 spin_lock(&imp->imp_lock);
1957 if (ptlrpc_import_delay_req(imp, req,
1960 * put on delay list - only if we wait
1961 * recovery finished - before send
1963 list_move_tail(&req->rq_list,
1964 &imp->imp_delayed_list);
1965 spin_unlock(&imp->imp_lock);
1970 req->rq_status = status;
1971 ptlrpc_rqphase_move(req,
1972 RQ_PHASE_INTERPRET);
1973 spin_unlock(&imp->imp_lock);
1974 GOTO(interpret, req->rq_status);
1976 /* ignore on just initiated connections */
1977 if (ptlrpc_no_resend(req) &&
1978 !req->rq_wait_ctx &&
1979 imp->imp_generation !=
1980 imp->imp_initiated_at) {
1981 req->rq_status = -ENOTCONN;
1982 ptlrpc_rqphase_move(req,
1983 RQ_PHASE_INTERPRET);
1984 spin_unlock(&imp->imp_lock);
1985 GOTO(interpret, req->rq_status);
1988 list_move_tail(&req->rq_list,
1989 &imp->imp_sending_list);
1991 spin_unlock(&imp->imp_lock);
1993 spin_lock(&req->rq_lock);
1994 req->rq_waiting = 0;
1995 spin_unlock(&req->rq_lock);
1997 if (req->rq_timedout || req->rq_resend) {
1999 * This is re-sending anyways,
2000 * let's mark req as resend.
2002 spin_lock(&req->rq_lock);
2004 spin_unlock(&req->rq_lock);
2007 * rq_wait_ctx is only touched by ptlrpcd,
2008 * so no lock is needed here.
2010 status = sptlrpc_req_refresh_ctx(req, -1);
2013 req->rq_status = status;
2014 spin_lock(&req->rq_lock);
2015 req->rq_wait_ctx = 0;
2016 spin_unlock(&req->rq_lock);
2017 force_timer_recalc = 1;
2019 spin_lock(&req->rq_lock);
2020 req->rq_wait_ctx = 1;
2021 spin_unlock(&req->rq_lock);
2026 spin_lock(&req->rq_lock);
2027 req->rq_wait_ctx = 0;
2028 spin_unlock(&req->rq_lock);
2032 * In any case, the previous bulk should be
2033 * cleaned up to prepare for the new sending
2036 !ptlrpc_unregister_bulk(req, 1))
2039 rc = ptl_send_rpc(req, 0);
2040 if (rc == -ENOMEM) {
2041 spin_lock(&imp->imp_lock);
2042 if (!list_empty(&req->rq_list))
2043 list_del_init(&req->rq_list);
2044 spin_unlock(&imp->imp_lock);
2045 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2049 DEBUG_REQ(D_HA, req,
2050 "send failed: rc = %d", rc);
2051 force_timer_recalc = 1;
2052 spin_lock(&req->rq_lock);
2053 req->rq_net_err = 1;
2054 spin_unlock(&req->rq_lock);
2057 /* need to reset the timeout */
2058 force_timer_recalc = 1;
2061 spin_lock(&req->rq_lock);
2063 if (ptlrpc_client_early(req)) {
2064 ptlrpc_at_recv_early_reply(req);
2065 spin_unlock(&req->rq_lock);
2069 /* Still waiting for a reply? */
2070 if (ptlrpc_client_recv(req)) {
2071 spin_unlock(&req->rq_lock);
2075 /* Did we actually receive a reply? */
2076 if (!ptlrpc_client_replied(req)) {
2077 spin_unlock(&req->rq_lock);
2081 spin_unlock(&req->rq_lock);
2084 * unlink from net because we are going to
2085 * swab in-place of reply buffer
2087 unregistered = ptlrpc_unregister_reply(req, 1);
2091 req->rq_status = after_reply(req);
2096 * If there is no bulk associated with this request,
2097 * then we're done and should let the interpreter
2098 * process the reply. Similarly if the RPC returned
2099 * an error, and therefore the bulk will never arrive.
2101 if (!req->rq_bulk || req->rq_status < 0) {
2102 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2103 GOTO(interpret, req->rq_status);
2106 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2109 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2110 if (ptlrpc_client_bulk_active(req))
2113 if (req->rq_bulk->bd_failure) {
2115 * The RPC reply arrived OK, but the bulk screwed
2116 * up! Dead weird since the server told us the RPC
2117 * was good after getting the REPLY for her GET or
2118 * the ACK for her PUT.
2120 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2121 req->rq_status = -EIO;
2124 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2127 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2130 * This moves to "unregistering" phase we need to wait for
2133 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2134 /* start async bulk unlink too */
2135 ptlrpc_unregister_bulk(req, 1);
2139 if (!ptlrpc_unregister_bulk(req, async))
2143 * When calling interpret receiving already should be
2146 LASSERT(!req->rq_receiving_reply);
2148 ptlrpc_req_interpret(env, req, req->rq_status);
2150 if (ptlrpcd_check_work(req)) {
2151 atomic_dec(&set->set_remaining);
2154 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2158 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2159 req, current_comm(),
2160 imp->imp_obd->obd_uuid.uuid,
2161 lustre_msg_get_status(req->rq_reqmsg),
2163 obd_import_nid2str(imp),
2164 lustre_msg_get_opc(req->rq_reqmsg),
2165 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2167 spin_lock(&imp->imp_lock);
2169 * Request already may be not on sending or delaying list. This
2170 * may happen in the case of marking it erroneous for the case
2171 * ptlrpc_import_delay_req(req, status) find it impossible to
2172 * allow sending this rpc and returns *status != 0.
2174 if (!list_empty(&req->rq_list)) {
2175 list_del_init(&req->rq_list);
2176 atomic_dec(&imp->imp_inflight);
2178 list_del_init(&req->rq_unreplied_list);
2179 spin_unlock(&imp->imp_lock);
2181 atomic_dec(&set->set_remaining);
2182 wake_up_all(&imp->imp_recovery_waitq);
2184 if (set->set_producer) {
2185 /* produce a new request if possible */
2186 if (ptlrpc_set_producer(set) > 0)
2187 force_timer_recalc = 1;
2190 * free the request that has just been completed
2191 * in order not to pollute set->set_requests
2193 list_del_init(&req->rq_set_chain);
2194 spin_lock(&req->rq_lock);
2196 req->rq_invalid_rqset = 0;
2197 spin_unlock(&req->rq_lock);
2199 /* record rq_status to compute the final status later */
2200 if (req->rq_status != 0)
2201 set->set_rc = req->rq_status;
2202 ptlrpc_req_finished(req);
2204 list_move_tail(&req->rq_set_chain, &comp_reqs);
2209 * move completed request at the head of list so it's easier for
2210 * caller to find them
2212 list_splice(&comp_reqs, &set->set_requests);
2214 /* If we hit an error, we want to recover promptly. */
2215 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2217 EXPORT_SYMBOL(ptlrpc_check_set);
2220 * Time out request \a req. is \a async_unlink is set, that means do not wait
2221 * until LNet actually confirms network buffer unlinking.
2222 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2224 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2226 struct obd_import *imp = req->rq_import;
2227 unsigned int debug_mask = D_RPCTRACE;
2231 spin_lock(&req->rq_lock);
2232 req->rq_timedout = 1;
2233 spin_unlock(&req->rq_lock);
2235 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2236 lustre_msg_get_status(req->rq_reqmsg)))
2237 debug_mask = D_WARNING;
2238 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2239 req->rq_net_err ? "failed due to network error" :
2240 ((req->rq_real_sent == 0 ||
2241 req->rq_real_sent < req->rq_sent ||
2242 req->rq_real_sent >= req->rq_deadline) ?
2243 "timed out for sent delay" : "timed out for slow reply"),
2244 (s64)req->rq_sent, (s64)req->rq_real_sent);
2246 if (imp && obd_debug_peer_on_timeout)
2247 LNetDebugPeer(imp->imp_connection->c_peer);
2249 ptlrpc_unregister_reply(req, async_unlink);
2250 ptlrpc_unregister_bulk(req, async_unlink);
2252 if (obd_dump_on_timeout)
2253 libcfs_debug_dumplog();
2256 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2260 atomic_inc(&imp->imp_timeouts);
2262 /* The DLM server doesn't want recovery run on its imports. */
2263 if (imp->imp_dlm_fake)
2267 * If this request is for recovery or other primordial tasks,
2268 * then error it out here.
2270 if (req->rq_ctx_init || req->rq_ctx_fini ||
2271 req->rq_send_state != LUSTRE_IMP_FULL ||
2272 imp->imp_obd->obd_no_recov) {
2273 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2274 ptlrpc_import_state_name(req->rq_send_state),
2275 ptlrpc_import_state_name(imp->imp_state));
2276 spin_lock(&req->rq_lock);
2277 req->rq_status = -ETIMEDOUT;
2279 spin_unlock(&req->rq_lock);
2284 * if a request can't be resent we can't wait for an answer after
2287 if (ptlrpc_no_resend(req)) {
2288 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2292 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2298 * Time out all uncompleted requests in request set pointed by \a data
2299 * Callback used when waiting on sets with l_wait_event.
2302 int ptlrpc_expired_set(void *data)
2304 struct ptlrpc_request_set *set = data;
2305 struct list_head *tmp;
2306 time64_t now = ktime_get_real_seconds();
2309 LASSERT(set != NULL);
2312 * A timeout expired. See which reqs it applies to...
2314 list_for_each(tmp, &set->set_requests) {
2315 struct ptlrpc_request *req =
2316 list_entry(tmp, struct ptlrpc_request,
2319 /* don't expire request waiting for context */
2320 if (req->rq_wait_ctx)
2323 /* Request in-flight? */
2324 if (!((req->rq_phase == RQ_PHASE_RPC &&
2325 !req->rq_waiting && !req->rq_resend) ||
2326 (req->rq_phase == RQ_PHASE_BULK)))
2329 if (req->rq_timedout || /* already dealt with */
2330 req->rq_deadline > now) /* not expired */
2334 * Deal with this guy. Do it asynchronously to not block
2337 ptlrpc_expire_one_request(req, 1);
2341 * When waiting for a whole set, we always break out of the
2342 * sleep so we can recalculate the timeout, or enable interrupts
2343 * if everyone's timed out.
2349 * Sets rq_intr flag in \a req under spinlock.
2351 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2353 spin_lock(&req->rq_lock);
2355 spin_unlock(&req->rq_lock);
2357 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2360 * Interrupts (sets interrupted flag) all uncompleted requests in
2361 * a set \a data. Callback for l_wait_event for interruptible waits.
2363 static void ptlrpc_interrupted_set(void *data)
2365 struct ptlrpc_request_set *set = data;
2366 struct list_head *tmp;
2368 LASSERT(set != NULL);
2369 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2371 list_for_each(tmp, &set->set_requests) {
2372 struct ptlrpc_request *req =
2373 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2378 if (req->rq_phase != RQ_PHASE_RPC &&
2379 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2380 !req->rq_allow_intr)
2383 ptlrpc_mark_interrupted(req);
2388 * Get the smallest timeout in the set; this does NOT set a timeout.
2390 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2392 struct list_head *tmp;
2393 time64_t now = ktime_get_real_seconds();
2395 struct ptlrpc_request *req;
2399 list_for_each(tmp, &set->set_requests) {
2400 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2402 /* Request in-flight? */
2403 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2404 (req->rq_phase == RQ_PHASE_BULK) ||
2405 (req->rq_phase == RQ_PHASE_NEW)))
2408 /* Already timed out. */
2409 if (req->rq_timedout)
2412 /* Waiting for ctx. */
2413 if (req->rq_wait_ctx)
2416 if (req->rq_phase == RQ_PHASE_NEW)
2417 deadline = req->rq_sent;
2418 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2419 deadline = req->rq_sent;
2421 deadline = req->rq_sent + req->rq_timeout;
2423 if (deadline <= now) /* actually expired already */
2424 timeout = 1; /* ASAP */
2425 else if (timeout == 0 || timeout > deadline - now)
2426 timeout = deadline - now;
2432 * Send all unset request from the set and then wait untill all
2433 * requests in the set complete (either get a reply, timeout, get an
2434 * error or otherwise be interrupted).
2435 * Returns 0 on success or error code otherwise.
2437 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2439 struct list_head *tmp;
2440 struct ptlrpc_request *req;
2441 struct l_wait_info lwi;
2446 if (set->set_producer)
2447 (void)ptlrpc_set_producer(set);
2449 list_for_each(tmp, &set->set_requests) {
2450 req = list_entry(tmp, struct ptlrpc_request,
2452 if (req->rq_phase == RQ_PHASE_NEW)
2453 (void)ptlrpc_send_new_req(req);
2456 if (list_empty(&set->set_requests))
2460 timeout = ptlrpc_set_next_timeout(set);
2463 * wait until all complete, interrupted, or an in-flight
2466 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2469 if ((timeout == 0 && !signal_pending(current)) ||
2470 set->set_allow_intr)
2472 * No requests are in-flight (ether timed out
2473 * or delayed), so we can allow interrupts.
2474 * We still want to block for a limited time,
2475 * so we allow interrupts during the timeout.
2477 lwi = LWI_TIMEOUT_INTR_ALL(
2478 cfs_time_seconds(timeout ? timeout : 1),
2480 ptlrpc_interrupted_set, set);
2483 * At least one request is in flight, so no
2484 * interrupts are allowed. Wait until all
2485 * complete, or an in-flight req times out.
2487 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2488 ptlrpc_expired_set, set);
2490 rc = l_wait_event(set->set_waitq,
2491 ptlrpc_check_set(NULL, set), &lwi);
2494 * LU-769 - if we ignored the signal because it was already
2495 * pending when we started, we need to handle it now or we risk
2496 * it being ignored forever
2498 if (rc == -ETIMEDOUT &&
2499 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2500 signal_pending(current)) {
2501 sigset_t blocked_sigs =
2502 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2505 * In fact we only interrupt for the "fatal" signals
2506 * like SIGINT or SIGKILL. We still ignore less
2507 * important signals since ptlrpc set is not easily
2508 * reentrant from userspace again
2510 if (signal_pending(current))
2511 ptlrpc_interrupted_set(set);
2512 cfs_restore_sigs(blocked_sigs);
2515 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2518 * -EINTR => all requests have been flagged rq_intr so next
2520 * -ETIMEDOUT => someone timed out. When all reqs have
2521 * timed out, signals are enabled allowing completion with
2523 * I don't really care if we go once more round the loop in
2524 * the error cases -eeb.
2526 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2527 list_for_each(tmp, &set->set_requests) {
2528 req = list_entry(tmp, struct ptlrpc_request,
2530 spin_lock(&req->rq_lock);
2531 req->rq_invalid_rqset = 1;
2532 spin_unlock(&req->rq_lock);
2535 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2537 LASSERT(atomic_read(&set->set_remaining) == 0);
2539 rc = set->set_rc; /* rq_status of already freed requests if any */
2540 list_for_each(tmp, &set->set_requests) {
2541 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2543 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2544 if (req->rq_status != 0)
2545 rc = req->rq_status;
2550 EXPORT_SYMBOL(ptlrpc_set_wait);
2553 * Helper fuction for request freeing.
2554 * Called when request count reached zero and request needs to be freed.
2555 * Removes request from all sorts of sending/replay lists it might be on,
2556 * frees network buffers if any are present.
2557 * If \a locked is set, that means caller is already holding import imp_lock
2558 * and so we no longer need to reobtain it (for certain lists manipulations)
2560 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2567 LASSERT(!request->rq_srv_req);
2568 LASSERT(request->rq_export == NULL);
2569 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2570 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2571 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2572 LASSERTF(!request->rq_replay, "req %p\n", request);
2574 req_capsule_fini(&request->rq_pill);
2577 * We must take it off the imp_replay_list first. Otherwise, we'll set
2578 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2580 if (request->rq_import) {
2582 spin_lock(&request->rq_import->imp_lock);
2583 list_del_init(&request->rq_replay_list);
2584 list_del_init(&request->rq_unreplied_list);
2586 spin_unlock(&request->rq_import->imp_lock);
2588 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2590 if (atomic_read(&request->rq_refcount) != 0) {
2591 DEBUG_REQ(D_ERROR, request,
2592 "freeing request with nonzero refcount");
2596 if (request->rq_repbuf)
2597 sptlrpc_cli_free_repbuf(request);
2599 if (request->rq_import) {
2600 class_import_put(request->rq_import);
2601 request->rq_import = NULL;
2603 if (request->rq_bulk)
2604 ptlrpc_free_bulk(request->rq_bulk);
2606 if (request->rq_reqbuf || request->rq_clrbuf)
2607 sptlrpc_cli_free_reqbuf(request);
2609 if (request->rq_cli_ctx)
2610 sptlrpc_req_put_ctx(request, !locked);
2612 if (request->rq_pool)
2613 __ptlrpc_free_req_to_pool(request);
2615 ptlrpc_request_cache_free(request);
2619 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2621 * Drop one request reference. Must be called with import imp_lock held.
2622 * When reference count drops to zero, request is freed.
2624 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2626 assert_spin_locked(&request->rq_import->imp_lock);
2627 (void)__ptlrpc_req_finished(request, 1);
2632 * Drops one reference count for request \a request.
2633 * \a locked set indicates that caller holds import imp_lock.
2634 * Frees the request whe reference count reaches zero.
2636 * \retval 1 the request is freed
2637 * \retval 0 some others still hold references on the request
2639 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2647 LASSERT(request != LP_POISON);
2648 LASSERT(request->rq_reqmsg != LP_POISON);
2650 DEBUG_REQ(D_INFO, request, "refcount now %u",
2651 atomic_read(&request->rq_refcount) - 1);
2653 spin_lock(&request->rq_lock);
2654 count = atomic_dec_return(&request->rq_refcount);
2655 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2658 * For open RPC, the client does not know the EA size (LOV, ACL, and
2659 * so on) before replied, then the client has to reserve very large
2660 * reply buffer. Such buffer will not be released until the RPC freed.
2661 * Since The open RPC is replayable, we need to keep it in the replay
2662 * list until close. If there are a lot of files opened concurrently,
2663 * then the client may be OOM.
2665 * If fact, it is unnecessary to keep reply buffer for open replay,
2666 * related EAs have already been saved via mdc_save_lovea() before
2667 * coming here. So it is safe to free the reply buffer some earlier
2668 * before releasing the RPC to avoid client OOM. LU-9514
2670 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2671 spin_lock(&request->rq_early_free_lock);
2672 sptlrpc_cli_free_repbuf(request);
2673 request->rq_repbuf = NULL;
2674 request->rq_repbuf_len = 0;
2675 request->rq_repdata = NULL;
2676 request->rq_reqdata_len = 0;
2677 spin_unlock(&request->rq_early_free_lock);
2679 spin_unlock(&request->rq_lock);
2682 __ptlrpc_free_req(request, locked);
2688 * Drops one reference count for a request.
2690 void ptlrpc_req_finished(struct ptlrpc_request *request)
2692 __ptlrpc_req_finished(request, 0);
2694 EXPORT_SYMBOL(ptlrpc_req_finished);
2697 * Returns xid of a \a request
2699 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2701 return request->rq_xid;
2703 EXPORT_SYMBOL(ptlrpc_req_xid);
2706 * Disengage the client's reply buffer from the network
2707 * NB does _NOT_ unregister any client-side bulk.
2708 * IDEMPOTENT, but _not_ safe against concurrent callers.
2709 * The request owner (i.e. the thread doing the I/O) must call...
2710 * Returns 0 on success or 1 if unregistering cannot be made.
2712 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2715 struct l_wait_info lwi;
2720 LASSERT(!in_interrupt());
2722 /* Let's setup deadline for reply unlink. */
2723 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2724 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2725 request->rq_reply_deadline = ktime_get_real_seconds() +
2729 * Nothing left to do.
2731 if (!ptlrpc_client_recv_or_unlink(request))
2734 LNetMDUnlink(request->rq_reply_md_h);
2737 * Let's check it once again.
2739 if (!ptlrpc_client_recv_or_unlink(request))
2742 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2743 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2746 * Do not wait for unlink to finish.
2752 * We have to l_wait_event() whatever the result, to give liblustre
2753 * a chance to run reply_in_callback(), and to make sure we've
2754 * unlinked before returning a req to the pool.
2757 /* The wq argument is ignored by user-space wait_event macros */
2758 wait_queue_head_t *wq = (request->rq_set) ?
2759 &request->rq_set->set_waitq :
2760 &request->rq_reply_waitq;
2762 * Network access will complete in finite time but the HUGE
2763 * timeout lets us CWARN for visibility of sluggish NALs
2765 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2766 cfs_time_seconds(1), NULL, NULL);
2767 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2770 ptlrpc_rqphase_move(request, request->rq_next_phase);
2774 LASSERT(rc == -ETIMEDOUT);
2775 DEBUG_REQ(D_WARNING, request,
2776 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2777 request->rq_receiving_reply,
2778 request->rq_req_unlinked,
2779 request->rq_reply_unlinked);
2784 static void ptlrpc_free_request(struct ptlrpc_request *req)
2786 spin_lock(&req->rq_lock);
2788 spin_unlock(&req->rq_lock);
2790 if (req->rq_commit_cb)
2791 req->rq_commit_cb(req);
2792 list_del_init(&req->rq_replay_list);
2794 __ptlrpc_req_finished(req, 1);
2798 * the request is committed and dropped from the replay list of its import
2800 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2802 struct obd_import *imp = req->rq_import;
2804 spin_lock(&imp->imp_lock);
2805 if (list_empty(&req->rq_replay_list)) {
2806 spin_unlock(&imp->imp_lock);
2810 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2811 if (imp->imp_replay_cursor == &req->rq_replay_list)
2812 imp->imp_replay_cursor = req->rq_replay_list.next;
2813 ptlrpc_free_request(req);
2816 spin_unlock(&imp->imp_lock);
2818 EXPORT_SYMBOL(ptlrpc_request_committed);
2821 * Iterates through replay_list on import and prunes
2822 * all requests have transno smaller than last_committed for the
2823 * import and don't have rq_replay set.
2824 * Since requests are sorted in transno order, stops when meetign first
2825 * transno bigger than last_committed.
2826 * caller must hold imp->imp_lock
2828 void ptlrpc_free_committed(struct obd_import *imp)
2830 struct ptlrpc_request *req, *saved;
2831 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2832 bool skip_committed_list = true;
2835 LASSERT(imp != NULL);
2836 assert_spin_locked(&imp->imp_lock);
2838 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2839 imp->imp_generation == imp->imp_last_generation_checked) {
2840 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2841 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2844 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2845 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2846 imp->imp_generation);
2848 if (imp->imp_generation != imp->imp_last_generation_checked ||
2849 imp->imp_last_transno_checked == 0)
2850 skip_committed_list = false;
2852 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2853 imp->imp_last_generation_checked = imp->imp_generation;
2855 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2857 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2858 LASSERT(req != last_req);
2861 if (req->rq_transno == 0) {
2862 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2865 if (req->rq_import_generation < imp->imp_generation) {
2866 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2870 /* not yet committed */
2871 if (req->rq_transno > imp->imp_peer_committed_transno) {
2872 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2876 if (req->rq_replay) {
2877 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2878 list_move_tail(&req->rq_replay_list,
2879 &imp->imp_committed_list);
2883 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2884 imp->imp_peer_committed_transno);
2886 ptlrpc_free_request(req);
2889 if (skip_committed_list)
2892 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2894 LASSERT(req->rq_transno != 0);
2895 if (req->rq_import_generation < imp->imp_generation ||
2897 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2898 req->rq_import_generation <
2899 imp->imp_generation ? "stale" : "closed");
2901 if (imp->imp_replay_cursor == &req->rq_replay_list)
2902 imp->imp_replay_cursor =
2903 req->rq_replay_list.next;
2905 ptlrpc_free_request(req);
2912 void ptlrpc_cleanup_client(struct obd_import *imp)
2919 * Schedule previously sent request for resend.
2920 * For bulk requests we assign new xid (to avoid problems with
2921 * lost replies and therefore several transfers landing into same buffer
2922 * from different sending attempts).
2924 void ptlrpc_resend_req(struct ptlrpc_request *req)
2926 DEBUG_REQ(D_HA, req, "going to resend");
2927 spin_lock(&req->rq_lock);
2930 * Request got reply but linked to the import list still.
2931 * Let ptlrpc_check_set() process it.
2933 if (ptlrpc_client_replied(req)) {
2934 spin_unlock(&req->rq_lock);
2935 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2939 req->rq_status = -EAGAIN;
2942 req->rq_net_err = 0;
2943 req->rq_timedout = 0;
2945 ptlrpc_client_wake_req(req);
2946 spin_unlock(&req->rq_lock);
2949 /* XXX: this function and rq_status are currently unused */
2950 void ptlrpc_restart_req(struct ptlrpc_request *req)
2952 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2953 req->rq_status = -ERESTARTSYS;
2955 spin_lock(&req->rq_lock);
2956 req->rq_restart = 1;
2957 req->rq_timedout = 0;
2958 ptlrpc_client_wake_req(req);
2959 spin_unlock(&req->rq_lock);
2963 * Grab additional reference on a request \a req
2965 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2968 atomic_inc(&req->rq_refcount);
2971 EXPORT_SYMBOL(ptlrpc_request_addref);
2974 * Add a request to import replay_list.
2975 * Must be called under imp_lock
2977 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2978 struct obd_import *imp)
2980 struct list_head *tmp;
2982 assert_spin_locked(&imp->imp_lock);
2984 if (req->rq_transno == 0) {
2985 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2990 * clear this for new requests that were resent as well
2991 * as resent replayed requests.
2993 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2995 /* don't re-add requests that have been replayed */
2996 if (!list_empty(&req->rq_replay_list))
2999 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3001 spin_lock(&req->rq_lock);
3003 spin_unlock(&req->rq_lock);
3005 LASSERT(imp->imp_replayable);
3006 /* Balanced in ptlrpc_free_committed, usually. */
3007 ptlrpc_request_addref(req);
3008 list_for_each_prev(tmp, &imp->imp_replay_list) {
3009 struct ptlrpc_request *iter = list_entry(tmp,
3010 struct ptlrpc_request,
3014 * We may have duplicate transnos if we create and then
3015 * open a file, or for closes retained if to match creating
3016 * opens, so use req->rq_xid as a secondary key.
3017 * (See bugs 684, 685, and 428.)
3018 * XXX no longer needed, but all opens need transnos!
3020 if (iter->rq_transno > req->rq_transno)
3023 if (iter->rq_transno == req->rq_transno) {
3024 LASSERT(iter->rq_xid != req->rq_xid);
3025 if (iter->rq_xid > req->rq_xid)
3029 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3033 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3037 * Send request and wait until it completes.
3038 * Returns request processing status.
3040 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3042 struct ptlrpc_request_set *set;
3046 LASSERT(req->rq_set == NULL);
3047 LASSERT(!req->rq_receiving_reply);
3049 set = ptlrpc_prep_set();
3051 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3055 /* for distributed debugging */
3056 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3058 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3059 ptlrpc_request_addref(req);
3060 ptlrpc_set_add_req(set, req);
3061 rc = ptlrpc_set_wait(NULL, set);
3062 ptlrpc_set_destroy(set);
3066 EXPORT_SYMBOL(ptlrpc_queue_wait);
3069 * Callback used for replayed requests reply processing.
3070 * In case of successful reply calls registered request replay callback.
3071 * In case of error restart replay process.
3073 static int ptlrpc_replay_interpret(const struct lu_env *env,
3074 struct ptlrpc_request *req,
3077 struct ptlrpc_replay_async_args *aa = args;
3078 struct obd_import *imp = req->rq_import;
3081 atomic_dec(&imp->imp_replay_inflight);
3084 * Note: if it is bulk replay (MDS-MDS replay), then even if
3085 * server got the request, but bulk transfer timeout, let's
3086 * replay the bulk req again
3088 if (!ptlrpc_client_replied(req) ||
3090 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3091 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3092 GOTO(out, rc = -ETIMEDOUT);
3095 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3096 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3097 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3098 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3100 /** VBR: check version failure */
3101 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3102 /** replay was failed due to version mismatch */
3103 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3104 spin_lock(&imp->imp_lock);
3105 imp->imp_vbr_failed = 1;
3106 spin_unlock(&imp->imp_lock);
3107 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3109 /** The transno had better not change over replay. */
3110 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3111 lustre_msg_get_transno(req->rq_repmsg) ||
3112 lustre_msg_get_transno(req->rq_repmsg) == 0,
3114 lustre_msg_get_transno(req->rq_reqmsg),
3115 lustre_msg_get_transno(req->rq_repmsg));
3118 spin_lock(&imp->imp_lock);
3119 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3120 spin_unlock(&imp->imp_lock);
3121 LASSERT(imp->imp_last_replay_transno);
3123 /* transaction number shouldn't be bigger than the latest replayed */
3124 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3125 DEBUG_REQ(D_ERROR, req,
3126 "Reported transno=%llu is bigger than replayed=%llu",
3128 lustre_msg_get_transno(req->rq_reqmsg));
3129 GOTO(out, rc = -EINVAL);
3132 DEBUG_REQ(D_HA, req, "got reply");
3134 /* let the callback do fixups, possibly including in the request */
3135 if (req->rq_replay_cb)
3136 req->rq_replay_cb(req);
3138 if (ptlrpc_client_replied(req) &&
3139 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3140 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3141 lustre_msg_get_status(req->rq_repmsg),
3142 aa->praa_old_status);
3145 * Note: If the replay fails for MDT-MDT recovery, let's
3146 * abort all of the following requests in the replay
3147 * and sending list, because MDT-MDT update requests
3148 * are dependent on each other, see LU-7039
3150 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3151 struct ptlrpc_request *free_req;
3152 struct ptlrpc_request *tmp;
3154 spin_lock(&imp->imp_lock);
3155 list_for_each_entry_safe(free_req, tmp,
3156 &imp->imp_replay_list,
3158 ptlrpc_free_request(free_req);
3161 list_for_each_entry_safe(free_req, tmp,
3162 &imp->imp_committed_list,
3164 ptlrpc_free_request(free_req);
3167 list_for_each_entry_safe(free_req, tmp,
3168 &imp->imp_delayed_list,
3170 spin_lock(&free_req->rq_lock);
3171 free_req->rq_err = 1;
3172 free_req->rq_status = -EIO;
3173 ptlrpc_client_wake_req(free_req);
3174 spin_unlock(&free_req->rq_lock);
3177 list_for_each_entry_safe(free_req, tmp,
3178 &imp->imp_sending_list,
3180 spin_lock(&free_req->rq_lock);
3181 free_req->rq_err = 1;
3182 free_req->rq_status = -EIO;
3183 ptlrpc_client_wake_req(free_req);
3184 spin_unlock(&free_req->rq_lock);
3186 spin_unlock(&imp->imp_lock);
3189 /* Put it back for re-replay. */
3190 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3194 * Errors while replay can set transno to 0, but
3195 * imp_last_replay_transno shouldn't be set to 0 anyway
3197 if (req->rq_transno == 0)
3198 CERROR("Transno is 0 during replay!\n");
3200 /* continue with recovery */
3201 rc = ptlrpc_import_recovery_state_machine(imp);
3203 req->rq_send_state = aa->praa_old_state;
3206 /* this replay failed, so restart recovery */
3207 ptlrpc_connect_import(imp);
3213 * Prepares and queues request for replay.
3214 * Adds it to ptlrpcd queue for actual sending.
3215 * Returns 0 on success.
3217 int ptlrpc_replay_req(struct ptlrpc_request *req)
3219 struct ptlrpc_replay_async_args *aa;
3223 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3225 aa = ptlrpc_req_async_args(aa, req);
3226 memset(aa, 0, sizeof(*aa));
3228 /* Prepare request to be resent with ptlrpcd */
3229 aa->praa_old_state = req->rq_send_state;
3230 req->rq_send_state = LUSTRE_IMP_REPLAY;
3231 req->rq_phase = RQ_PHASE_NEW;
3232 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3234 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3236 req->rq_interpret_reply = ptlrpc_replay_interpret;
3237 /* Readjust the timeout for current conditions */
3238 ptlrpc_at_set_req_timeout(req);
3240 /* Tell server net_latency to calculate how long to wait for reply. */
3241 lustre_msg_set_service_time(req->rq_reqmsg,
3242 ptlrpc_at_get_net_latency(req));
3243 DEBUG_REQ(D_HA, req, "REPLAY");
3245 atomic_inc(&req->rq_import->imp_replay_inflight);
3246 spin_lock(&req->rq_lock);
3247 req->rq_early_free_repbuf = 0;
3248 spin_unlock(&req->rq_lock);
3249 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3251 ptlrpcd_add_req(req);
3256 * Aborts all in-flight request on import \a imp sending and delayed lists
3258 void ptlrpc_abort_inflight(struct obd_import *imp)
3260 struct list_head *tmp, *n;
3264 * Make sure that no new requests get processed for this import.
3265 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3266 * this flag and then putting requests on sending_list or delayed_list.
3268 assert_spin_locked(&imp->imp_lock);
3271 * XXX locking? Maybe we should remove each request with the list
3272 * locked? Also, how do we know if the requests on the list are
3273 * being freed at this time?
3275 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3276 struct ptlrpc_request *req = list_entry(tmp,
3277 struct ptlrpc_request,
3280 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3282 spin_lock(&req->rq_lock);
3283 if (req->rq_import_generation < imp->imp_generation) {
3285 req->rq_status = -EIO;
3286 ptlrpc_client_wake_req(req);
3288 spin_unlock(&req->rq_lock);
3291 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3292 struct ptlrpc_request *req =
3293 list_entry(tmp, struct ptlrpc_request, rq_list);
3295 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3297 spin_lock(&req->rq_lock);
3298 if (req->rq_import_generation < imp->imp_generation) {
3300 req->rq_status = -EIO;
3301 ptlrpc_client_wake_req(req);
3303 spin_unlock(&req->rq_lock);
3307 * Last chance to free reqs left on the replay list, but we
3308 * will still leak reqs that haven't committed.
3310 if (imp->imp_replayable)
3311 ptlrpc_free_committed(imp);
3317 * Abort all uncompleted requests in request set \a set
3319 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3321 struct list_head *tmp, *pos;
3323 LASSERT(set != NULL);
3325 list_for_each_safe(pos, tmp, &set->set_requests) {
3326 struct ptlrpc_request *req =
3327 list_entry(pos, struct ptlrpc_request,
3330 spin_lock(&req->rq_lock);
3331 if (req->rq_phase != RQ_PHASE_RPC) {
3332 spin_unlock(&req->rq_lock);
3337 req->rq_status = -EINTR;
3338 ptlrpc_client_wake_req(req);
3339 spin_unlock(&req->rq_lock);
3344 * Initialize the XID for the node. This is common among all requests on
3345 * this node, and only requires the property that it is monotonically
3346 * increasing. It does not need to be sequential. Since this is also used
3347 * as the RDMA match bits, it is important that a single client NOT have
3348 * the same match bits for two different in-flight requests, hence we do
3349 * NOT want to have an XID per target or similar.
3351 * To avoid an unlikely collision between match bits after a client reboot
3352 * (which would deliver old data into the wrong RDMA buffer) initialize
3353 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3354 * If the time is clearly incorrect, we instead use a 62-bit random number.
3355 * In the worst case the random number will overflow 1M RPCs per second in
3356 * 9133 years, or permutations thereof.
3358 #define YEAR_2004 (1ULL << 30)
3359 void ptlrpc_init_xid(void)
3361 time64_t now = ktime_get_real_seconds();
3364 if (now < YEAR_2004) {
3365 get_random_bytes(&xid, sizeof(xid));
3367 xid |= (1ULL << 61);
3369 xid = (u64)now << 20;
3372 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3373 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3375 xid &= PTLRPC_BULK_OPS_MASK;
3376 atomic64_set(&ptlrpc_last_xid, xid);
3380 * Increase xid and returns resulting new value to the caller.
3382 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3383 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3384 * itself uses the last bulk xid needed, so the server can determine the
3385 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3386 * xid must align to a power-of-two value.
3388 * This is assumed to be true due to the initial ptlrpc_last_xid
3389 * value also being initialized to a power-of-two value. LU-1431
3391 __u64 ptlrpc_next_xid(void)
3393 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3397 * If request has a new allocated XID (new request or EINPROGRESS resend),
3398 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3399 * request to ensure previous bulk fails and avoid problems with lost replies
3400 * and therefore several transfers landing into the same buffer from different
3403 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3405 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3407 LASSERT(bd != NULL);
3410 * Generate new matchbits for all resend requests, including
3413 if (req->rq_resend) {
3414 __u64 old_mbits = req->rq_mbits;
3417 * First time resend on -EINPROGRESS will generate new xid,
3418 * so we can actually use the rq_xid as rq_mbits in such case,
3419 * however, it's bit hard to distinguish such resend with a
3420 * 'resend for the -EINPROGRESS resend'. To make it simple,
3421 * we opt to generate mbits for all resend cases.
3423 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3425 req->rq_mbits = ptlrpc_next_xid();
3428 * Old version transfers rq_xid to peer as
3431 spin_lock(&req->rq_import->imp_lock);
3432 list_del_init(&req->rq_unreplied_list);
3433 ptlrpc_assign_next_xid_nolock(req);
3434 spin_unlock(&req->rq_import->imp_lock);
3435 req->rq_mbits = req->rq_xid;
3437 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3438 old_mbits, req->rq_mbits);
3439 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3440 /* Request being sent first time, use xid as matchbits. */
3441 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3442 || req->rq_mbits == 0) {
3443 req->rq_mbits = req->rq_xid;
3445 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3447 req->rq_mbits -= total_md - 1;
3451 * Replay request, xid and matchbits have already been
3452 * correctly assigned.
3458 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3459 * that server can infer the number of bulks that were prepared,
3462 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3466 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3467 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3469 * It's ok to directly set the rq_xid here, since this xid bump
3470 * won't affect the request position in unreplied list.
3472 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3473 req->rq_xid = req->rq_mbits;
3477 * Get a glimpse at what next xid value might have been.
3478 * Returns possible next xid.
3480 __u64 ptlrpc_sample_next_xid(void)
3482 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3484 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3487 * Functions for operating ptlrpc workers.
3489 * A ptlrpc work is a function which will be running inside ptlrpc context.
3490 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3492 * 1. after a work is created, it can be used many times, that is:
3493 * handler = ptlrpcd_alloc_work();
3494 * ptlrpcd_queue_work();
3496 * queue it again when necessary:
3497 * ptlrpcd_queue_work();
3498 * ptlrpcd_destroy_work();
3499 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3500 * it will only be queued once in any time. Also as its name implies, it may
3501 * have delay before it really runs by ptlrpcd thread.
3503 struct ptlrpc_work_async_args {
3504 int (*cb)(const struct lu_env *, void *);
3508 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3510 /* re-initialize the req */
3511 req->rq_timeout = obd_timeout;
3512 req->rq_sent = ktime_get_real_seconds();
3513 req->rq_deadline = req->rq_sent + req->rq_timeout;
3514 req->rq_phase = RQ_PHASE_INTERPRET;
3515 req->rq_next_phase = RQ_PHASE_COMPLETE;
3516 req->rq_xid = ptlrpc_next_xid();
3517 req->rq_import_generation = req->rq_import->imp_generation;
3519 ptlrpcd_add_req(req);
3522 static int work_interpreter(const struct lu_env *env,
3523 struct ptlrpc_request *req, void *args, int rc)
3525 struct ptlrpc_work_async_args *arg = args;
3527 LASSERT(ptlrpcd_check_work(req));
3528 LASSERT(arg->cb != NULL);
3530 rc = arg->cb(env, arg->cbdata);
3532 list_del_init(&req->rq_set_chain);
3535 if (atomic_dec_return(&req->rq_refcount) > 1) {
3536 atomic_set(&req->rq_refcount, 2);
3537 ptlrpcd_add_work_req(req);
3542 static int worker_format;
3544 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3546 return req->rq_pill.rc_fmt == (void *)&worker_format;
3550 * Create a work for ptlrpc.
3552 void *ptlrpcd_alloc_work(struct obd_import *imp,
3553 int (*cb)(const struct lu_env *, void *), void *cbdata)
3555 struct ptlrpc_request *req = NULL;
3556 struct ptlrpc_work_async_args *args;
3562 RETURN(ERR_PTR(-EINVAL));
3564 /* copy some code from deprecated fakereq. */
3565 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3567 CERROR("ptlrpc: run out of memory!\n");
3568 RETURN(ERR_PTR(-ENOMEM));
3571 ptlrpc_cli_req_init(req);
3573 req->rq_send_state = LUSTRE_IMP_FULL;
3574 req->rq_type = PTL_RPC_MSG_REQUEST;
3575 req->rq_import = class_import_get(imp);
3576 req->rq_interpret_reply = work_interpreter;
3577 /* don't want reply */
3578 req->rq_no_delay = req->rq_no_resend = 1;
3579 req->rq_pill.rc_fmt = (void *)&worker_format;
3581 args = ptlrpc_req_async_args(args, req);
3583 args->cbdata = cbdata;
3587 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3589 void ptlrpcd_destroy_work(void *handler)
3591 struct ptlrpc_request *req = handler;
3594 ptlrpc_req_finished(req);
3596 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3598 int ptlrpcd_queue_work(void *handler)
3600 struct ptlrpc_request *req = handler;
3603 * Check if the req is already being queued.
3605 * Here comes a trick: it lacks a way of checking if a req is being
3606 * processed reliably in ptlrpc. Here I have to use refcount of req
3607 * for this purpose. This is okay because the caller should use this
3608 * req as opaque data. - Jinshan
3610 LASSERT(atomic_read(&req->rq_refcount) > 0);
3611 if (atomic_inc_return(&req->rq_refcount) == 2)
3612 ptlrpcd_add_work_req(req);
3615 EXPORT_SYMBOL(ptlrpcd_queue_work);