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 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");
748 EXPORT_SYMBOL(ptlrpc_reassign_next_xid);
750 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
751 __u32 version, int opcode, char **bufs,
752 struct ptlrpc_cli_ctx *ctx)
755 struct obd_import *imp;
761 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
762 imp = request->rq_import;
763 lengths = request->rq_pill.rc_area[RCL_CLIENT];
766 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
768 rc = sptlrpc_req_get_ctx(request);
772 sptlrpc_req_set_flavor(request, opcode);
774 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
779 lustre_msg_add_version(request->rq_reqmsg, version);
780 request->rq_send_state = LUSTRE_IMP_FULL;
781 request->rq_type = PTL_RPC_MSG_REQUEST;
783 request->rq_req_cbid.cbid_fn = request_out_callback;
784 request->rq_req_cbid.cbid_arg = request;
786 request->rq_reply_cbid.cbid_fn = reply_in_callback;
787 request->rq_reply_cbid.cbid_arg = request;
789 request->rq_reply_deadline = 0;
790 request->rq_bulk_deadline = 0;
791 request->rq_req_deadline = 0;
792 request->rq_phase = RQ_PHASE_NEW;
793 request->rq_next_phase = RQ_PHASE_UNDEFINED;
795 request->rq_request_portal = imp->imp_client->cli_request_portal;
796 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
798 ptlrpc_at_set_req_timeout(request);
800 lustre_msg_set_opc(request->rq_reqmsg, opcode);
802 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
803 if (cfs_fail_val == opcode) {
804 time64_t *fail_t = NULL, *fail2_t = NULL;
806 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
807 fail_t = &request->rq_bulk_deadline;
808 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
809 fail_t = &request->rq_reply_deadline;
810 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
811 fail_t = &request->rq_req_deadline;
812 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
813 fail_t = &request->rq_reply_deadline;
814 fail2_t = &request->rq_bulk_deadline;
815 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
816 time64_t now = ktime_get_real_seconds();
817 u64 xid = ((u64)now >> 4) << 24;
819 atomic64_set(&ptlrpc_last_xid, xid);
823 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
826 *fail2_t = ktime_get_real_seconds() +
830 * The RPC is infected, let the test to change the
833 msleep(4 * MSEC_PER_SEC);
836 ptlrpc_assign_next_xid(request);
841 LASSERT(!request->rq_pool);
842 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
844 class_import_put(imp);
848 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
851 * Pack request buffers for network transfer, performing necessary encryption
852 * steps if necessary.
854 int ptlrpc_request_pack(struct ptlrpc_request *request,
855 __u32 version, int opcode)
857 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
859 EXPORT_SYMBOL(ptlrpc_request_pack);
862 * Helper function to allocate new request on import \a imp
863 * and possibly using existing request from pool \a pool if provided.
864 * Returns allocated request structure with import field filled or
868 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
869 struct ptlrpc_request_pool *pool)
871 struct ptlrpc_request *request = NULL;
873 request = ptlrpc_request_cache_alloc(GFP_NOFS);
875 if (!request && pool)
876 request = ptlrpc_prep_req_from_pool(pool);
879 ptlrpc_cli_req_init(request);
881 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
882 LASSERT(imp != LP_POISON);
883 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
885 LASSERT(imp->imp_client != LP_POISON);
887 request->rq_import = class_import_get(imp);
889 CERROR("request allocation out of memory\n");
896 * Helper function for creating a request.
897 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
898 * buffer structures according to capsule template \a format.
899 * Returns allocated request structure pointer or NULL on error.
901 static struct ptlrpc_request *
902 ptlrpc_request_alloc_internal(struct obd_import *imp,
903 struct ptlrpc_request_pool *pool,
904 const struct req_format *format)
906 struct ptlrpc_request *request;
908 request = __ptlrpc_request_alloc(imp, pool);
913 * initiate connection if needed when the import has been
914 * referenced by the new request to avoid races with disconnect
916 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
919 CDEBUG_LIMIT(imp->imp_idle_debug,
920 "%s: reconnect after %llds idle\n",
921 imp->imp_obd->obd_name, ktime_get_real_seconds() -
922 imp->imp_last_reply_time);
923 spin_lock(&imp->imp_lock);
924 if (imp->imp_state == LUSTRE_IMP_IDLE) {
925 imp->imp_generation++;
926 imp->imp_initiated_at = imp->imp_generation;
927 imp->imp_state = LUSTRE_IMP_NEW;
929 /* connect_import_locked releases imp_lock */
930 rc = ptlrpc_connect_import_locked(imp);
932 ptlrpc_request_free(request);
935 ptlrpc_pinger_add_import(imp);
937 spin_unlock(&imp->imp_lock);
941 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
942 req_capsule_set(&request->rq_pill, format);
947 * Allocate new request structure for import \a imp and initialize its
948 * buffer structure according to capsule template \a format.
950 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
951 const struct req_format *format)
953 return ptlrpc_request_alloc_internal(imp, NULL, format);
955 EXPORT_SYMBOL(ptlrpc_request_alloc);
958 * Allocate new request structure for import \a imp from pool \a pool and
959 * initialize its buffer structure according to capsule template \a format.
961 struct ptlrpc_request *
962 ptlrpc_request_alloc_pool(struct obd_import *imp,
963 struct ptlrpc_request_pool *pool,
964 const struct req_format *format)
966 return ptlrpc_request_alloc_internal(imp, pool, format);
968 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
971 * For requests not from pool, free memory of the request structure.
972 * For requests obtained from a pool earlier, return request back to pool.
974 void ptlrpc_request_free(struct ptlrpc_request *request)
976 if (request->rq_pool)
977 __ptlrpc_free_req_to_pool(request);
979 ptlrpc_request_cache_free(request);
981 EXPORT_SYMBOL(ptlrpc_request_free);
984 * Allocate new request for operatione \a opcode and immediatelly pack it for
986 * Only used for simple requests like OBD_PING where the only important
987 * part of the request is operation itself.
988 * Returns allocated request or NULL on error.
990 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
991 const struct req_format *format,
992 __u32 version, int opcode)
994 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
998 rc = ptlrpc_request_pack(req, version, opcode);
1000 ptlrpc_request_free(req);
1006 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1009 * Allocate and initialize new request set structure on the current CPT.
1010 * Returns a pointer to the newly allocated set structure or NULL on error.
1012 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1014 struct ptlrpc_request_set *set;
1018 cpt = cfs_cpt_current(cfs_cpt_table, 0);
1019 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1022 atomic_set(&set->set_refcount, 1);
1023 INIT_LIST_HEAD(&set->set_requests);
1024 init_waitqueue_head(&set->set_waitq);
1025 atomic_set(&set->set_new_count, 0);
1026 atomic_set(&set->set_remaining, 0);
1027 spin_lock_init(&set->set_new_req_lock);
1028 INIT_LIST_HEAD(&set->set_new_requests);
1029 set->set_max_inflight = UINT_MAX;
1030 set->set_producer = NULL;
1031 set->set_producer_arg = NULL;
1036 EXPORT_SYMBOL(ptlrpc_prep_set);
1039 * Allocate and initialize new request set structure with flow control
1040 * extension. This extension allows to control the number of requests in-flight
1041 * for the whole set. A callback function to generate requests must be provided
1042 * and the request set will keep the number of requests sent over the wire to
1044 * Returns a pointer to the newly allocated set structure or NULL on error.
1046 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1050 struct ptlrpc_request_set *set;
1052 set = ptlrpc_prep_set();
1056 set->set_max_inflight = max;
1057 set->set_producer = func;
1058 set->set_producer_arg = arg;
1064 * Wind down and free request set structure previously allocated with
1066 * Ensures that all requests on the set have completed and removes
1067 * all requests from the request list in a set.
1068 * If any unsent request happen to be on the list, pretends that they got
1069 * an error in flight and calls their completion handler.
1071 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1073 struct list_head *tmp;
1074 struct list_head *next;
1080 /* Requests on the set should either all be completed, or all be new */
1081 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1082 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1083 list_for_each(tmp, &set->set_requests) {
1084 struct ptlrpc_request *req =
1085 list_entry(tmp, struct ptlrpc_request,
1088 LASSERT(req->rq_phase == expected_phase);
1092 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1093 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1094 atomic_read(&set->set_remaining), n);
1096 list_for_each_safe(tmp, next, &set->set_requests) {
1097 struct ptlrpc_request *req =
1098 list_entry(tmp, struct ptlrpc_request,
1100 list_del_init(&req->rq_set_chain);
1102 LASSERT(req->rq_phase == expected_phase);
1104 if (req->rq_phase == RQ_PHASE_NEW) {
1105 ptlrpc_req_interpret(NULL, req, -EBADR);
1106 atomic_dec(&set->set_remaining);
1109 spin_lock(&req->rq_lock);
1111 req->rq_invalid_rqset = 0;
1112 spin_unlock(&req->rq_lock);
1114 ptlrpc_req_finished(req);
1117 LASSERT(atomic_read(&set->set_remaining) == 0);
1119 ptlrpc_reqset_put(set);
1122 EXPORT_SYMBOL(ptlrpc_set_destroy);
1125 * Add a new request to the general purpose request set.
1126 * Assumes request reference from the caller.
1128 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1129 struct ptlrpc_request *req)
1131 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1132 LASSERT(list_empty(&req->rq_set_chain));
1134 if (req->rq_allow_intr)
1135 set->set_allow_intr = 1;
1137 /* The set takes over the caller's request reference */
1138 list_add_tail(&req->rq_set_chain, &set->set_requests);
1140 atomic_inc(&set->set_remaining);
1141 req->rq_queued_time = ktime_get_seconds();
1144 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1146 if (set->set_producer)
1148 * If the request set has a producer callback, the RPC must be
1149 * sent straight away
1151 ptlrpc_send_new_req(req);
1153 EXPORT_SYMBOL(ptlrpc_set_add_req);
1156 * Add a request to a request with dedicated server thread
1157 * and wake the thread to make any necessary processing.
1158 * Currently only used for ptlrpcd.
1160 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1161 struct ptlrpc_request *req)
1163 struct ptlrpc_request_set *set = pc->pc_set;
1166 LASSERT(req->rq_set == NULL);
1167 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1169 spin_lock(&set->set_new_req_lock);
1171 * The set takes over the caller's request reference.
1174 req->rq_queued_time = ktime_get_seconds();
1175 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1176 count = atomic_inc_return(&set->set_new_count);
1177 spin_unlock(&set->set_new_req_lock);
1179 /* Only need to call wakeup once for the first entry. */
1181 wake_up(&set->set_waitq);
1184 * XXX: It maybe unnecessary to wakeup all the partners. But to
1185 * guarantee the async RPC can be processed ASAP, we have
1186 * no other better choice. It maybe fixed in future.
1188 for (i = 0; i < pc->pc_npartners; i++)
1189 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1194 * Based on the current state of the import, determine if the request
1195 * can be sent, is an error, or should be delayed.
1197 * Returns true if this request should be delayed. If false, and
1198 * *status is set, then the request can not be sent and *status is the
1199 * error code. If false and status is 0, then request can be sent.
1201 * The imp->imp_lock must be held.
1203 static int ptlrpc_import_delay_req(struct obd_import *imp,
1204 struct ptlrpc_request *req, int *status)
1212 if (req->rq_ctx_init || req->rq_ctx_fini) {
1213 /* always allow ctx init/fini rpc go through */
1214 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1215 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1217 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1218 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1221 * pings or MDS-equivalent STATFS may safely
1224 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1225 D_HA : D_ERROR, req, "IMP_CLOSED");
1227 } else if (ptlrpc_send_limit_expired(req)) {
1228 /* probably doesn't need to be a D_ERROR afterinitial testing */
1229 DEBUG_REQ(D_HA, req, "send limit expired");
1230 *status = -ETIMEDOUT;
1231 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1232 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1233 ;/* allow CONNECT even if import is invalid */
1234 if (atomic_read(&imp->imp_inval_count) != 0) {
1235 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1238 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1239 if (!imp->imp_deactive)
1240 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1241 *status = -ESHUTDOWN; /* b=12940 */
1242 } else if (req->rq_import_generation != imp->imp_generation) {
1243 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1245 } else if (req->rq_send_state != imp->imp_state) {
1246 /* invalidate in progress - any requests should be drop */
1247 if (atomic_read(&imp->imp_inval_count) != 0) {
1248 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1250 } else if (req->rq_no_delay &&
1251 imp->imp_generation != imp->imp_initiated_at) {
1252 /* ignore nodelay for requests initiating connections */
1253 *status = -EWOULDBLOCK;
1254 } else if (req->rq_allow_replay &&
1255 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1256 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1257 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1258 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1259 DEBUG_REQ(D_HA, req, "allow during recovery");
1269 * Decide if the error message should be printed to the console or not.
1270 * Makes its decision based on request type, status, and failure frequency.
1272 * \param[in] req request that failed and may need a console message
1274 * \retval false if no message should be printed
1275 * \retval true if console message should be printed
1277 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1279 LASSERT(req->rq_reqmsg != NULL);
1281 /* Suppress particular reconnect errors which are to be expected. */
1282 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1283 /* Suppress timed out reconnect requests */
1284 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1289 * Suppress most unavailable/again reconnect requests, but
1290 * print occasionally so it is clear client is trying to
1291 * connect to a server where no target is running.
1293 if ((err == -ENODEV || err == -EAGAIN) &&
1294 req->rq_import->imp_conn_cnt % 30 != 20)
1298 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1299 /* -EAGAIN is normal when using POSIX flocks */
1302 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1303 (req->rq_xid & 0xf) != 10)
1304 /* Suppress most ping requests, they may fail occasionally */
1311 * Check request processing status.
1312 * Returns the status.
1314 static int ptlrpc_check_status(struct ptlrpc_request *req)
1319 rc = lustre_msg_get_status(req->rq_repmsg);
1320 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1321 struct obd_import *imp = req->rq_import;
1322 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1323 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1325 if (ptlrpc_console_allow(req, opc, rc))
1326 LCONSOLE_ERROR_MSG(0x11,
1327 "%s: operation %s to node %s failed: rc = %d\n",
1328 imp->imp_obd->obd_name,
1330 libcfs_nid2str(nid), rc);
1331 RETURN(rc < 0 ? rc : -EINVAL);
1335 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1341 * save pre-versions of objects into request for replay.
1342 * Versions are obtained from server reply.
1345 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1347 struct lustre_msg *repmsg = req->rq_repmsg;
1348 struct lustre_msg *reqmsg = req->rq_reqmsg;
1349 __u64 *versions = lustre_msg_get_versions(repmsg);
1352 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1356 lustre_msg_set_versions(reqmsg, versions);
1357 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1358 versions[0], versions[1]);
1363 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1365 struct ptlrpc_request *req;
1367 assert_spin_locked(&imp->imp_lock);
1368 if (list_empty(&imp->imp_unreplied_list))
1371 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1373 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1375 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1376 imp->imp_known_replied_xid = req->rq_xid - 1;
1378 return req->rq_xid - 1;
1382 * Callback function called when client receives RPC reply for \a req.
1383 * Returns 0 on success or error code.
1384 * The return alue would be assigned to req->rq_status by the caller
1385 * as request processing status.
1386 * This function also decides if the request needs to be saved for later replay.
1388 static int after_reply(struct ptlrpc_request *req)
1390 struct obd_import *imp = req->rq_import;
1391 struct obd_device *obd = req->rq_import->imp_obd;
1398 LASSERT(obd != NULL);
1399 /* repbuf must be unlinked */
1400 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1402 if (req->rq_reply_truncated) {
1403 if (ptlrpc_no_resend(req)) {
1404 DEBUG_REQ(D_ERROR, req,
1405 "reply buffer overflow, expected=%d, actual size=%d",
1406 req->rq_nob_received, req->rq_repbuf_len);
1410 sptlrpc_cli_free_repbuf(req);
1412 * Pass the required reply buffer size (include
1413 * space for early reply).
1414 * NB: no need to roundup because alloc_repbuf
1417 req->rq_replen = req->rq_nob_received;
1418 req->rq_nob_received = 0;
1419 spin_lock(&req->rq_lock);
1421 spin_unlock(&req->rq_lock);
1425 work_start = ktime_get_real();
1426 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1429 * NB Until this point, the whole of the incoming message,
1430 * including buflens, status etc is in the sender's byte order.
1432 rc = sptlrpc_cli_unwrap_reply(req);
1434 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1439 * Security layer unwrap might ask resend this request.
1444 rc = unpack_reply(req);
1448 /* retry indefinitely on EINPROGRESS */
1449 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1450 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1451 time64_t now = ktime_get_real_seconds();
1453 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1454 D_RPCTRACE, req, "resending request on EINPROGRESS");
1455 spin_lock(&req->rq_lock);
1457 spin_unlock(&req->rq_lock);
1458 req->rq_nr_resend++;
1460 /* Readjust the timeout for current conditions */
1461 ptlrpc_at_set_req_timeout(req);
1463 * delay resend to give a chance to the server to get ready.
1464 * The delay is increased by 1s on every resend and is capped to
1465 * the current request timeout (i.e. obd_timeout if AT is off,
1466 * or AT service time x 125% + 5s, see at_est2timeout)
1468 if (req->rq_nr_resend > req->rq_timeout)
1469 req->rq_sent = now + req->rq_timeout;
1471 req->rq_sent = now + req->rq_nr_resend;
1473 /* Resend for EINPROGRESS will use a new XID */
1474 spin_lock(&imp->imp_lock);
1475 list_del_init(&req->rq_unreplied_list);
1476 spin_unlock(&imp->imp_lock);
1481 if (obd->obd_svc_stats) {
1482 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1484 ptlrpc_lprocfs_rpc_sent(req, timediff);
1487 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1488 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1489 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1490 lustre_msg_get_type(req->rq_repmsg));
1494 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1495 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1496 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1497 ptlrpc_at_adj_net_latency(req,
1498 lustre_msg_get_service_time(req->rq_repmsg));
1500 rc = ptlrpc_check_status(req);
1504 * Either we've been evicted, or the server has failed for
1505 * some reason. Try to reconnect, and if that fails, punt to
1508 if (ptlrpc_recoverable_error(rc)) {
1509 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1510 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1513 ptlrpc_request_handle_notconn(req);
1518 * Let's look if server sent slv. Do it only for RPC with
1521 ldlm_cli_update_pool(req);
1525 * Store transno in reqmsg for replay.
1527 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1528 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1529 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1532 if (imp->imp_replayable) {
1533 spin_lock(&imp->imp_lock);
1535 * No point in adding already-committed requests to the replay
1536 * list, we will just remove them immediately. b=9829
1538 if (req->rq_transno != 0 &&
1540 lustre_msg_get_last_committed(req->rq_repmsg) ||
1542 /** version recovery */
1543 ptlrpc_save_versions(req);
1544 ptlrpc_retain_replayable_request(req, imp);
1545 } else if (req->rq_commit_cb &&
1546 list_empty(&req->rq_replay_list)) {
1548 * NB: don't call rq_commit_cb if it's already on
1549 * rq_replay_list, ptlrpc_free_committed() will call
1550 * it later, see LU-3618 for details
1552 spin_unlock(&imp->imp_lock);
1553 req->rq_commit_cb(req);
1554 spin_lock(&imp->imp_lock);
1558 * Replay-enabled imports return commit-status information.
1560 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1561 if (likely(committed > imp->imp_peer_committed_transno))
1562 imp->imp_peer_committed_transno = committed;
1564 ptlrpc_free_committed(imp);
1566 if (!list_empty(&imp->imp_replay_list)) {
1567 struct ptlrpc_request *last;
1569 last = list_entry(imp->imp_replay_list.prev,
1570 struct ptlrpc_request,
1573 * Requests with rq_replay stay on the list even if no
1574 * commit is expected.
1576 if (last->rq_transno > imp->imp_peer_committed_transno)
1577 ptlrpc_pinger_commit_expected(imp);
1580 spin_unlock(&imp->imp_lock);
1587 * Helper function to send request \a req over the network for the first time
1588 * Also adjusts request phase.
1589 * Returns 0 on success or error code.
1591 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1593 struct obd_import *imp = req->rq_import;
1598 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1600 /* do not try to go further if there is not enough memory in enc_pool */
1601 if (req->rq_sent && req->rq_bulk)
1602 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1603 pool_is_at_full_capacity())
1606 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1607 (!req->rq_generation_set ||
1608 req->rq_import_generation == imp->imp_generation))
1611 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1613 spin_lock(&imp->imp_lock);
1615 LASSERT(req->rq_xid != 0);
1616 LASSERT(!list_empty(&req->rq_unreplied_list));
1618 if (!req->rq_generation_set)
1619 req->rq_import_generation = imp->imp_generation;
1621 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1622 spin_lock(&req->rq_lock);
1623 req->rq_waiting = 1;
1624 spin_unlock(&req->rq_lock);
1626 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1627 ptlrpc_import_state_name(req->rq_send_state),
1628 ptlrpc_import_state_name(imp->imp_state));
1629 LASSERT(list_empty(&req->rq_list));
1630 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1631 atomic_inc(&req->rq_import->imp_inflight);
1632 spin_unlock(&imp->imp_lock);
1637 spin_unlock(&imp->imp_lock);
1638 req->rq_status = rc;
1639 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1643 LASSERT(list_empty(&req->rq_list));
1644 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1645 atomic_inc(&req->rq_import->imp_inflight);
1648 * find the known replied XID from the unreplied list, CONNECT
1649 * and DISCONNECT requests are skipped to make the sanity check
1650 * on server side happy. see process_req_last_xid().
1652 * For CONNECT: Because replay requests have lower XID, it'll
1653 * break the sanity check if CONNECT bump the exp_last_xid on
1656 * For DISCONNECT: Since client will abort inflight RPC before
1657 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1658 * than the inflight RPC.
1660 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1661 min_xid = ptlrpc_known_replied_xid(imp);
1662 spin_unlock(&imp->imp_lock);
1664 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1666 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1668 rc = sptlrpc_req_refresh_ctx(req, -1);
1671 req->rq_status = rc;
1674 spin_lock(&req->rq_lock);
1675 req->rq_wait_ctx = 1;
1676 spin_unlock(&req->rq_lock);
1682 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1683 req, current_comm(),
1684 imp->imp_obd->obd_uuid.uuid,
1685 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1686 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1687 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1689 rc = ptl_send_rpc(req, 0);
1690 if (rc == -ENOMEM) {
1691 spin_lock(&imp->imp_lock);
1692 if (!list_empty(&req->rq_list)) {
1693 list_del_init(&req->rq_list);
1694 atomic_dec(&req->rq_import->imp_inflight);
1696 spin_unlock(&imp->imp_lock);
1697 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1701 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1703 spin_lock(&req->rq_lock);
1704 req->rq_net_err = 1;
1705 spin_unlock(&req->rq_lock);
1711 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1716 LASSERT(set->set_producer != NULL);
1718 remaining = atomic_read(&set->set_remaining);
1721 * populate the ->set_requests list with requests until we
1722 * reach the maximum number of RPCs in flight for this set
1724 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1725 rc = set->set_producer(set, set->set_producer_arg);
1726 if (rc == -ENOENT) {
1727 /* no more RPC to produce */
1728 set->set_producer = NULL;
1729 set->set_producer_arg = NULL;
1734 RETURN((atomic_read(&set->set_remaining) - remaining));
1738 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1739 * and no more replies are expected.
1740 * (it is possible to get less replies than requests sent e.g. due to timed out
1741 * requests or requests that we had trouble to send out)
1743 * NOTE: This function contains a potential schedule point (cond_resched()).
1745 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1747 struct list_head *tmp, *next;
1748 struct list_head comp_reqs;
1749 int force_timer_recalc = 0;
1752 if (atomic_read(&set->set_remaining) == 0)
1755 INIT_LIST_HEAD(&comp_reqs);
1756 list_for_each_safe(tmp, next, &set->set_requests) {
1757 struct ptlrpc_request *req =
1758 list_entry(tmp, struct ptlrpc_request,
1760 struct obd_import *imp = req->rq_import;
1761 int unregistered = 0;
1765 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1766 list_move_tail(&req->rq_set_chain, &comp_reqs);
1771 * This schedule point is mainly for the ptlrpcd caller of this
1772 * function. Most ptlrpc sets are not long-lived and unbounded
1773 * in length, but at the least the set used by the ptlrpcd is.
1774 * Since the processing time is unbounded, we need to insert an
1775 * explicit schedule point to make the thread well-behaved.
1780 * If the caller requires to allow to be interpreted by force
1781 * and it has really been interpreted, then move the request
1782 * to RQ_PHASE_INTERPRET phase in spite of what the current
1785 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1786 req->rq_status = -EINTR;
1787 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1790 * Since it is interpreted and we have to wait for
1791 * the reply to be unlinked, then use sync mode.
1795 GOTO(interpret, req->rq_status);
1798 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1799 force_timer_recalc = 1;
1801 /* delayed send - skip */
1802 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1805 /* delayed resend - skip */
1806 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1807 req->rq_sent > ktime_get_real_seconds())
1810 if (!(req->rq_phase == RQ_PHASE_RPC ||
1811 req->rq_phase == RQ_PHASE_BULK ||
1812 req->rq_phase == RQ_PHASE_INTERPRET ||
1813 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1814 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1815 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1819 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1820 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1821 LASSERT(req->rq_next_phase != req->rq_phase);
1822 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1824 if (req->rq_req_deadline &&
1825 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1826 req->rq_req_deadline = 0;
1827 if (req->rq_reply_deadline &&
1828 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1829 req->rq_reply_deadline = 0;
1830 if (req->rq_bulk_deadline &&
1831 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1832 req->rq_bulk_deadline = 0;
1835 * Skip processing until reply is unlinked. We
1836 * can't return to pool before that and we can't
1837 * call interpret before that. We need to make
1838 * sure that all rdma transfers finished and will
1839 * not corrupt any data.
1841 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1842 ptlrpc_client_recv_or_unlink(req))
1844 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1845 ptlrpc_client_bulk_active(req))
1849 * Turn fail_loc off to prevent it from looping
1852 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1853 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1856 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1857 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1862 * Move to next phase if reply was successfully
1865 ptlrpc_rqphase_move(req, req->rq_next_phase);
1868 if (req->rq_phase == RQ_PHASE_INTERPRET)
1869 GOTO(interpret, req->rq_status);
1872 * Note that this also will start async reply unlink.
1874 if (req->rq_net_err && !req->rq_timedout) {
1875 ptlrpc_expire_one_request(req, 1);
1878 * Check if we still need to wait for unlink.
1880 if (ptlrpc_client_recv_or_unlink(req) ||
1881 ptlrpc_client_bulk_active(req))
1883 /* If there is no need to resend, fail it now. */
1884 if (req->rq_no_resend) {
1885 if (req->rq_status == 0)
1886 req->rq_status = -EIO;
1887 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1888 GOTO(interpret, req->rq_status);
1895 spin_lock(&req->rq_lock);
1896 req->rq_replied = 0;
1897 spin_unlock(&req->rq_lock);
1898 if (req->rq_status == 0)
1899 req->rq_status = -EIO;
1900 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1901 GOTO(interpret, req->rq_status);
1905 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1906 * so it sets rq_intr regardless of individual rpc
1907 * timeouts. The synchronous IO waiting path sets
1908 * rq_intr irrespective of whether ptlrpcd
1909 * has seen a timeout. Our policy is to only interpret
1910 * interrupted rpcs after they have timed out, so we
1911 * need to enforce that here.
1914 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1915 req->rq_wait_ctx)) {
1916 req->rq_status = -EINTR;
1917 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1918 GOTO(interpret, req->rq_status);
1921 if (req->rq_phase == RQ_PHASE_RPC) {
1922 if (req->rq_timedout || req->rq_resend ||
1923 req->rq_waiting || req->rq_wait_ctx) {
1926 if (!ptlrpc_unregister_reply(req, 1)) {
1927 ptlrpc_unregister_bulk(req, 1);
1931 spin_lock(&imp->imp_lock);
1932 if (ptlrpc_import_delay_req(imp, req,
1935 * put on delay list - only if we wait
1936 * recovery finished - before send
1938 list_move_tail(&req->rq_list,
1939 &imp->imp_delayed_list);
1940 spin_unlock(&imp->imp_lock);
1945 req->rq_status = status;
1946 ptlrpc_rqphase_move(req,
1947 RQ_PHASE_INTERPRET);
1948 spin_unlock(&imp->imp_lock);
1949 GOTO(interpret, req->rq_status);
1951 /* ignore on just initiated connections */
1952 if (ptlrpc_no_resend(req) &&
1953 !req->rq_wait_ctx &&
1954 imp->imp_generation !=
1955 imp->imp_initiated_at) {
1956 req->rq_status = -ENOTCONN;
1957 ptlrpc_rqphase_move(req,
1958 RQ_PHASE_INTERPRET);
1959 spin_unlock(&imp->imp_lock);
1960 GOTO(interpret, req->rq_status);
1963 list_move_tail(&req->rq_list,
1964 &imp->imp_sending_list);
1966 spin_unlock(&imp->imp_lock);
1968 spin_lock(&req->rq_lock);
1969 req->rq_waiting = 0;
1970 spin_unlock(&req->rq_lock);
1972 if (req->rq_timedout || req->rq_resend) {
1974 * This is re-sending anyways,
1975 * let's mark req as resend.
1977 spin_lock(&req->rq_lock);
1979 spin_unlock(&req->rq_lock);
1982 * rq_wait_ctx is only touched by ptlrpcd,
1983 * so no lock is needed here.
1985 status = sptlrpc_req_refresh_ctx(req, -1);
1988 req->rq_status = status;
1989 spin_lock(&req->rq_lock);
1990 req->rq_wait_ctx = 0;
1991 spin_unlock(&req->rq_lock);
1992 force_timer_recalc = 1;
1994 spin_lock(&req->rq_lock);
1995 req->rq_wait_ctx = 1;
1996 spin_unlock(&req->rq_lock);
2001 spin_lock(&req->rq_lock);
2002 req->rq_wait_ctx = 0;
2003 spin_unlock(&req->rq_lock);
2007 * In any case, the previous bulk should be
2008 * cleaned up to prepare for the new sending
2011 !ptlrpc_unregister_bulk(req, 1))
2014 rc = ptl_send_rpc(req, 0);
2015 if (rc == -ENOMEM) {
2016 spin_lock(&imp->imp_lock);
2017 if (!list_empty(&req->rq_list))
2018 list_del_init(&req->rq_list);
2019 spin_unlock(&imp->imp_lock);
2020 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2024 DEBUG_REQ(D_HA, req,
2025 "send failed: rc = %d", rc);
2026 force_timer_recalc = 1;
2027 spin_lock(&req->rq_lock);
2028 req->rq_net_err = 1;
2029 spin_unlock(&req->rq_lock);
2032 /* need to reset the timeout */
2033 force_timer_recalc = 1;
2036 spin_lock(&req->rq_lock);
2038 if (ptlrpc_client_early(req)) {
2039 ptlrpc_at_recv_early_reply(req);
2040 spin_unlock(&req->rq_lock);
2044 /* Still waiting for a reply? */
2045 if (ptlrpc_client_recv(req)) {
2046 spin_unlock(&req->rq_lock);
2050 /* Did we actually receive a reply? */
2051 if (!ptlrpc_client_replied(req)) {
2052 spin_unlock(&req->rq_lock);
2056 spin_unlock(&req->rq_lock);
2059 * unlink from net because we are going to
2060 * swab in-place of reply buffer
2062 unregistered = ptlrpc_unregister_reply(req, 1);
2066 req->rq_status = after_reply(req);
2071 * If there is no bulk associated with this request,
2072 * then we're done and should let the interpreter
2073 * process the reply. Similarly if the RPC returned
2074 * an error, and therefore the bulk will never arrive.
2076 if (!req->rq_bulk || req->rq_status < 0) {
2077 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2078 GOTO(interpret, req->rq_status);
2081 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2084 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2085 if (ptlrpc_client_bulk_active(req))
2088 if (req->rq_bulk->bd_failure) {
2090 * The RPC reply arrived OK, but the bulk screwed
2091 * up! Dead weird since the server told us the RPC
2092 * was good after getting the REPLY for her GET or
2093 * the ACK for her PUT.
2095 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2096 req->rq_status = -EIO;
2099 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2102 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2105 * This moves to "unregistering" phase we need to wait for
2108 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2109 /* start async bulk unlink too */
2110 ptlrpc_unregister_bulk(req, 1);
2114 if (!ptlrpc_unregister_bulk(req, async))
2118 * When calling interpret receiving already should be
2121 LASSERT(!req->rq_receiving_reply);
2123 ptlrpc_req_interpret(env, req, req->rq_status);
2125 if (ptlrpcd_check_work(req)) {
2126 atomic_dec(&set->set_remaining);
2129 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2133 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2134 req, current_comm(),
2135 imp->imp_obd->obd_uuid.uuid,
2136 lustre_msg_get_status(req->rq_reqmsg),
2138 obd_import_nid2str(imp),
2139 lustre_msg_get_opc(req->rq_reqmsg),
2140 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2142 spin_lock(&imp->imp_lock);
2144 * Request already may be not on sending or delaying list. This
2145 * may happen in the case of marking it erroneous for the case
2146 * ptlrpc_import_delay_req(req, status) find it impossible to
2147 * allow sending this rpc and returns *status != 0.
2149 if (!list_empty(&req->rq_list)) {
2150 list_del_init(&req->rq_list);
2151 atomic_dec(&imp->imp_inflight);
2153 list_del_init(&req->rq_unreplied_list);
2154 spin_unlock(&imp->imp_lock);
2156 atomic_dec(&set->set_remaining);
2157 wake_up_all(&imp->imp_recovery_waitq);
2159 if (set->set_producer) {
2160 /* produce a new request if possible */
2161 if (ptlrpc_set_producer(set) > 0)
2162 force_timer_recalc = 1;
2165 * free the request that has just been completed
2166 * in order not to pollute set->set_requests
2168 list_del_init(&req->rq_set_chain);
2169 spin_lock(&req->rq_lock);
2171 req->rq_invalid_rqset = 0;
2172 spin_unlock(&req->rq_lock);
2174 /* record rq_status to compute the final status later */
2175 if (req->rq_status != 0)
2176 set->set_rc = req->rq_status;
2177 ptlrpc_req_finished(req);
2179 list_move_tail(&req->rq_set_chain, &comp_reqs);
2184 * move completed request at the head of list so it's easier for
2185 * caller to find them
2187 list_splice(&comp_reqs, &set->set_requests);
2189 /* If we hit an error, we want to recover promptly. */
2190 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2192 EXPORT_SYMBOL(ptlrpc_check_set);
2195 * Time out request \a req. is \a async_unlink is set, that means do not wait
2196 * until LNet actually confirms network buffer unlinking.
2197 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2199 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2201 struct obd_import *imp = req->rq_import;
2202 unsigned int debug_mask = D_RPCTRACE;
2206 spin_lock(&req->rq_lock);
2207 req->rq_timedout = 1;
2208 spin_unlock(&req->rq_lock);
2210 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2211 lustre_msg_get_status(req->rq_reqmsg)))
2212 debug_mask = D_WARNING;
2213 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2214 req->rq_net_err ? "failed due to network error" :
2215 ((req->rq_real_sent == 0 ||
2216 req->rq_real_sent < req->rq_sent ||
2217 req->rq_real_sent >= req->rq_deadline) ?
2218 "timed out for sent delay" : "timed out for slow reply"),
2219 (s64)req->rq_sent, (s64)req->rq_real_sent);
2221 if (imp && obd_debug_peer_on_timeout)
2222 LNetDebugPeer(imp->imp_connection->c_peer);
2224 ptlrpc_unregister_reply(req, async_unlink);
2225 ptlrpc_unregister_bulk(req, async_unlink);
2227 if (obd_dump_on_timeout)
2228 libcfs_debug_dumplog();
2231 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2235 atomic_inc(&imp->imp_timeouts);
2237 /* The DLM server doesn't want recovery run on its imports. */
2238 if (imp->imp_dlm_fake)
2242 * If this request is for recovery or other primordial tasks,
2243 * then error it out here.
2245 if (req->rq_ctx_init || req->rq_ctx_fini ||
2246 req->rq_send_state != LUSTRE_IMP_FULL ||
2247 imp->imp_obd->obd_no_recov) {
2248 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2249 ptlrpc_import_state_name(req->rq_send_state),
2250 ptlrpc_import_state_name(imp->imp_state));
2251 spin_lock(&req->rq_lock);
2252 req->rq_status = -ETIMEDOUT;
2254 spin_unlock(&req->rq_lock);
2259 * if a request can't be resent we can't wait for an answer after
2262 if (ptlrpc_no_resend(req)) {
2263 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2267 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2273 * Time out all uncompleted requests in request set pointed by \a data
2274 * Callback used when waiting on sets with l_wait_event.
2277 int ptlrpc_expired_set(void *data)
2279 struct ptlrpc_request_set *set = data;
2280 struct list_head *tmp;
2281 time64_t now = ktime_get_real_seconds();
2284 LASSERT(set != NULL);
2287 * A timeout expired. See which reqs it applies to...
2289 list_for_each(tmp, &set->set_requests) {
2290 struct ptlrpc_request *req =
2291 list_entry(tmp, struct ptlrpc_request,
2294 /* don't expire request waiting for context */
2295 if (req->rq_wait_ctx)
2298 /* Request in-flight? */
2299 if (!((req->rq_phase == RQ_PHASE_RPC &&
2300 !req->rq_waiting && !req->rq_resend) ||
2301 (req->rq_phase == RQ_PHASE_BULK)))
2304 if (req->rq_timedout || /* already dealt with */
2305 req->rq_deadline > now) /* not expired */
2309 * Deal with this guy. Do it asynchronously to not block
2312 ptlrpc_expire_one_request(req, 1);
2316 * When waiting for a whole set, we always break out of the
2317 * sleep so we can recalculate the timeout, or enable interrupts
2318 * if everyone's timed out.
2324 * Sets rq_intr flag in \a req under spinlock.
2326 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2328 spin_lock(&req->rq_lock);
2330 spin_unlock(&req->rq_lock);
2332 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2335 * Interrupts (sets interrupted flag) all uncompleted requests in
2336 * a set \a data. Callback for l_wait_event for interruptible waits.
2338 static void ptlrpc_interrupted_set(void *data)
2340 struct ptlrpc_request_set *set = data;
2341 struct list_head *tmp;
2343 LASSERT(set != NULL);
2344 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2346 list_for_each(tmp, &set->set_requests) {
2347 struct ptlrpc_request *req =
2348 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2353 if (req->rq_phase != RQ_PHASE_RPC &&
2354 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2355 !req->rq_allow_intr)
2358 ptlrpc_mark_interrupted(req);
2363 * Get the smallest timeout in the set; this does NOT set a timeout.
2365 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2367 struct list_head *tmp;
2368 time64_t now = ktime_get_real_seconds();
2370 struct ptlrpc_request *req;
2374 list_for_each(tmp, &set->set_requests) {
2375 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2377 /* Request in-flight? */
2378 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2379 (req->rq_phase == RQ_PHASE_BULK) ||
2380 (req->rq_phase == RQ_PHASE_NEW)))
2383 /* Already timed out. */
2384 if (req->rq_timedout)
2387 /* Waiting for ctx. */
2388 if (req->rq_wait_ctx)
2391 if (req->rq_phase == RQ_PHASE_NEW)
2392 deadline = req->rq_sent;
2393 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2394 deadline = req->rq_sent;
2396 deadline = req->rq_sent + req->rq_timeout;
2398 if (deadline <= now) /* actually expired already */
2399 timeout = 1; /* ASAP */
2400 else if (timeout == 0 || timeout > deadline - now)
2401 timeout = deadline - now;
2407 * Send all unset request from the set and then wait untill all
2408 * requests in the set complete (either get a reply, timeout, get an
2409 * error or otherwise be interrupted).
2410 * Returns 0 on success or error code otherwise.
2412 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2414 struct list_head *tmp;
2415 struct ptlrpc_request *req;
2416 struct l_wait_info lwi;
2421 if (set->set_producer)
2422 (void)ptlrpc_set_producer(set);
2424 list_for_each(tmp, &set->set_requests) {
2425 req = list_entry(tmp, struct ptlrpc_request,
2427 if (req->rq_phase == RQ_PHASE_NEW)
2428 (void)ptlrpc_send_new_req(req);
2431 if (list_empty(&set->set_requests))
2435 timeout = ptlrpc_set_next_timeout(set);
2438 * wait until all complete, interrupted, or an in-flight
2441 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2444 if ((timeout == 0 && !signal_pending(current)) ||
2445 set->set_allow_intr)
2447 * No requests are in-flight (ether timed out
2448 * or delayed), so we can allow interrupts.
2449 * We still want to block for a limited time,
2450 * so we allow interrupts during the timeout.
2452 lwi = LWI_TIMEOUT_INTR_ALL(
2453 cfs_time_seconds(timeout ? timeout : 1),
2455 ptlrpc_interrupted_set, set);
2458 * At least one request is in flight, so no
2459 * interrupts are allowed. Wait until all
2460 * complete, or an in-flight req times out.
2462 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2463 ptlrpc_expired_set, set);
2465 rc = l_wait_event(set->set_waitq,
2466 ptlrpc_check_set(NULL, set), &lwi);
2469 * LU-769 - if we ignored the signal because it was already
2470 * pending when we started, we need to handle it now or we risk
2471 * it being ignored forever
2473 if (rc == -ETIMEDOUT &&
2474 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2475 signal_pending(current)) {
2476 sigset_t blocked_sigs =
2477 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2480 * In fact we only interrupt for the "fatal" signals
2481 * like SIGINT or SIGKILL. We still ignore less
2482 * important signals since ptlrpc set is not easily
2483 * reentrant from userspace again
2485 if (signal_pending(current))
2486 ptlrpc_interrupted_set(set);
2487 cfs_restore_sigs(blocked_sigs);
2490 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2493 * -EINTR => all requests have been flagged rq_intr so next
2495 * -ETIMEDOUT => someone timed out. When all reqs have
2496 * timed out, signals are enabled allowing completion with
2498 * I don't really care if we go once more round the loop in
2499 * the error cases -eeb.
2501 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2502 list_for_each(tmp, &set->set_requests) {
2503 req = list_entry(tmp, struct ptlrpc_request,
2505 spin_lock(&req->rq_lock);
2506 req->rq_invalid_rqset = 1;
2507 spin_unlock(&req->rq_lock);
2510 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2512 LASSERT(atomic_read(&set->set_remaining) == 0);
2514 rc = set->set_rc; /* rq_status of already freed requests if any */
2515 list_for_each(tmp, &set->set_requests) {
2516 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2518 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2519 if (req->rq_status != 0)
2520 rc = req->rq_status;
2525 EXPORT_SYMBOL(ptlrpc_set_wait);
2528 * Helper fuction for request freeing.
2529 * Called when request count reached zero and request needs to be freed.
2530 * Removes request from all sorts of sending/replay lists it might be on,
2531 * frees network buffers if any are present.
2532 * If \a locked is set, that means caller is already holding import imp_lock
2533 * and so we no longer need to reobtain it (for certain lists manipulations)
2535 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2542 LASSERT(!request->rq_srv_req);
2543 LASSERT(request->rq_export == NULL);
2544 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2545 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2546 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2547 LASSERTF(!request->rq_replay, "req %p\n", request);
2549 req_capsule_fini(&request->rq_pill);
2552 * We must take it off the imp_replay_list first. Otherwise, we'll set
2553 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2555 if (request->rq_import) {
2557 spin_lock(&request->rq_import->imp_lock);
2558 list_del_init(&request->rq_replay_list);
2559 list_del_init(&request->rq_unreplied_list);
2561 spin_unlock(&request->rq_import->imp_lock);
2563 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2565 if (atomic_read(&request->rq_refcount) != 0) {
2566 DEBUG_REQ(D_ERROR, request,
2567 "freeing request with nonzero refcount");
2571 if (request->rq_repbuf)
2572 sptlrpc_cli_free_repbuf(request);
2574 if (request->rq_import) {
2575 class_import_put(request->rq_import);
2576 request->rq_import = NULL;
2578 if (request->rq_bulk)
2579 ptlrpc_free_bulk(request->rq_bulk);
2581 if (request->rq_reqbuf || request->rq_clrbuf)
2582 sptlrpc_cli_free_reqbuf(request);
2584 if (request->rq_cli_ctx)
2585 sptlrpc_req_put_ctx(request, !locked);
2587 if (request->rq_pool)
2588 __ptlrpc_free_req_to_pool(request);
2590 ptlrpc_request_cache_free(request);
2594 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2596 * Drop one request reference. Must be called with import imp_lock held.
2597 * When reference count drops to zero, request is freed.
2599 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2601 assert_spin_locked(&request->rq_import->imp_lock);
2602 (void)__ptlrpc_req_finished(request, 1);
2607 * Drops one reference count for request \a request.
2608 * \a locked set indicates that caller holds import imp_lock.
2609 * Frees the request whe reference count reaches zero.
2611 * \retval 1 the request is freed
2612 * \retval 0 some others still hold references on the request
2614 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2622 LASSERT(request != LP_POISON);
2623 LASSERT(request->rq_reqmsg != LP_POISON);
2625 DEBUG_REQ(D_INFO, request, "refcount now %u",
2626 atomic_read(&request->rq_refcount) - 1);
2628 spin_lock(&request->rq_lock);
2629 count = atomic_dec_return(&request->rq_refcount);
2630 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2633 * For open RPC, the client does not know the EA size (LOV, ACL, and
2634 * so on) before replied, then the client has to reserve very large
2635 * reply buffer. Such buffer will not be released until the RPC freed.
2636 * Since The open RPC is replayable, we need to keep it in the replay
2637 * list until close. If there are a lot of files opened concurrently,
2638 * then the client may be OOM.
2640 * If fact, it is unnecessary to keep reply buffer for open replay,
2641 * related EAs have already been saved via mdc_save_lovea() before
2642 * coming here. So it is safe to free the reply buffer some earlier
2643 * before releasing the RPC to avoid client OOM. LU-9514
2645 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2646 spin_lock(&request->rq_early_free_lock);
2647 sptlrpc_cli_free_repbuf(request);
2648 request->rq_repbuf = NULL;
2649 request->rq_repbuf_len = 0;
2650 request->rq_repdata = NULL;
2651 request->rq_reqdata_len = 0;
2652 spin_unlock(&request->rq_early_free_lock);
2654 spin_unlock(&request->rq_lock);
2657 __ptlrpc_free_req(request, locked);
2663 * Drops one reference count for a request.
2665 void ptlrpc_req_finished(struct ptlrpc_request *request)
2667 __ptlrpc_req_finished(request, 0);
2669 EXPORT_SYMBOL(ptlrpc_req_finished);
2672 * Returns xid of a \a request
2674 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2676 return request->rq_xid;
2678 EXPORT_SYMBOL(ptlrpc_req_xid);
2681 * Disengage the client's reply buffer from the network
2682 * NB does _NOT_ unregister any client-side bulk.
2683 * IDEMPOTENT, but _not_ safe against concurrent callers.
2684 * The request owner (i.e. the thread doing the I/O) must call...
2685 * Returns 0 on success or 1 if unregistering cannot be made.
2687 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2690 struct l_wait_info lwi;
2695 LASSERT(!in_interrupt());
2697 /* Let's setup deadline for reply unlink. */
2698 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2699 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2700 request->rq_reply_deadline = ktime_get_real_seconds() +
2704 * Nothing left to do.
2706 if (!ptlrpc_client_recv_or_unlink(request))
2709 LNetMDUnlink(request->rq_reply_md_h);
2712 * Let's check it once again.
2714 if (!ptlrpc_client_recv_or_unlink(request))
2717 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2718 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2721 * Do not wait for unlink to finish.
2727 * We have to l_wait_event() whatever the result, to give liblustre
2728 * a chance to run reply_in_callback(), and to make sure we've
2729 * unlinked before returning a req to the pool.
2732 /* The wq argument is ignored by user-space wait_event macros */
2733 wait_queue_head_t *wq = (request->rq_set) ?
2734 &request->rq_set->set_waitq :
2735 &request->rq_reply_waitq;
2737 * Network access will complete in finite time but the HUGE
2738 * timeout lets us CWARN for visibility of sluggish NALs
2740 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2741 cfs_time_seconds(1), NULL, NULL);
2742 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2745 ptlrpc_rqphase_move(request, request->rq_next_phase);
2749 LASSERT(rc == -ETIMEDOUT);
2750 DEBUG_REQ(D_WARNING, request,
2751 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2752 request->rq_receiving_reply,
2753 request->rq_req_unlinked,
2754 request->rq_reply_unlinked);
2759 static void ptlrpc_free_request(struct ptlrpc_request *req)
2761 spin_lock(&req->rq_lock);
2763 spin_unlock(&req->rq_lock);
2765 if (req->rq_commit_cb)
2766 req->rq_commit_cb(req);
2767 list_del_init(&req->rq_replay_list);
2769 __ptlrpc_req_finished(req, 1);
2773 * the request is committed and dropped from the replay list of its import
2775 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2777 struct obd_import *imp = req->rq_import;
2779 spin_lock(&imp->imp_lock);
2780 if (list_empty(&req->rq_replay_list)) {
2781 spin_unlock(&imp->imp_lock);
2785 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2786 if (imp->imp_replay_cursor == &req->rq_replay_list)
2787 imp->imp_replay_cursor = req->rq_replay_list.next;
2788 ptlrpc_free_request(req);
2791 spin_unlock(&imp->imp_lock);
2793 EXPORT_SYMBOL(ptlrpc_request_committed);
2796 * Iterates through replay_list on import and prunes
2797 * all requests have transno smaller than last_committed for the
2798 * import and don't have rq_replay set.
2799 * Since requests are sorted in transno order, stops when meetign first
2800 * transno bigger than last_committed.
2801 * caller must hold imp->imp_lock
2803 void ptlrpc_free_committed(struct obd_import *imp)
2805 struct ptlrpc_request *req, *saved;
2806 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2807 bool skip_committed_list = true;
2810 LASSERT(imp != NULL);
2811 assert_spin_locked(&imp->imp_lock);
2813 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2814 imp->imp_generation == imp->imp_last_generation_checked) {
2815 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2816 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2819 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2820 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2821 imp->imp_generation);
2823 if (imp->imp_generation != imp->imp_last_generation_checked ||
2824 imp->imp_last_transno_checked == 0)
2825 skip_committed_list = false;
2827 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2828 imp->imp_last_generation_checked = imp->imp_generation;
2830 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2832 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2833 LASSERT(req != last_req);
2836 if (req->rq_transno == 0) {
2837 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2840 if (req->rq_import_generation < imp->imp_generation) {
2841 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2845 /* not yet committed */
2846 if (req->rq_transno > imp->imp_peer_committed_transno) {
2847 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2851 if (req->rq_replay) {
2852 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2853 list_move_tail(&req->rq_replay_list,
2854 &imp->imp_committed_list);
2858 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2859 imp->imp_peer_committed_transno);
2861 ptlrpc_free_request(req);
2864 if (skip_committed_list)
2867 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2869 LASSERT(req->rq_transno != 0);
2870 if (req->rq_import_generation < imp->imp_generation ||
2872 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2873 req->rq_import_generation <
2874 imp->imp_generation ? "stale" : "closed");
2876 if (imp->imp_replay_cursor == &req->rq_replay_list)
2877 imp->imp_replay_cursor =
2878 req->rq_replay_list.next;
2880 ptlrpc_free_request(req);
2887 void ptlrpc_cleanup_client(struct obd_import *imp)
2894 * Schedule previously sent request for resend.
2895 * For bulk requests we assign new xid (to avoid problems with
2896 * lost replies and therefore several transfers landing into same buffer
2897 * from different sending attempts).
2899 void ptlrpc_resend_req(struct ptlrpc_request *req)
2901 DEBUG_REQ(D_HA, req, "going to resend");
2902 spin_lock(&req->rq_lock);
2905 * Request got reply but linked to the import list still.
2906 * Let ptlrpc_check_set() process it.
2908 if (ptlrpc_client_replied(req)) {
2909 spin_unlock(&req->rq_lock);
2910 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2914 req->rq_status = -EAGAIN;
2917 req->rq_net_err = 0;
2918 req->rq_timedout = 0;
2920 ptlrpc_client_wake_req(req);
2921 spin_unlock(&req->rq_lock);
2924 /* XXX: this function and rq_status are currently unused */
2925 void ptlrpc_restart_req(struct ptlrpc_request *req)
2927 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2928 req->rq_status = -ERESTARTSYS;
2930 spin_lock(&req->rq_lock);
2931 req->rq_restart = 1;
2932 req->rq_timedout = 0;
2933 ptlrpc_client_wake_req(req);
2934 spin_unlock(&req->rq_lock);
2938 * Grab additional reference on a request \a req
2940 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2943 atomic_inc(&req->rq_refcount);
2946 EXPORT_SYMBOL(ptlrpc_request_addref);
2949 * Add a request to import replay_list.
2950 * Must be called under imp_lock
2952 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2953 struct obd_import *imp)
2955 struct list_head *tmp;
2957 assert_spin_locked(&imp->imp_lock);
2959 if (req->rq_transno == 0) {
2960 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2965 * clear this for new requests that were resent as well
2966 * as resent replayed requests.
2968 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2970 /* don't re-add requests that have been replayed */
2971 if (!list_empty(&req->rq_replay_list))
2974 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2976 spin_lock(&req->rq_lock);
2978 spin_unlock(&req->rq_lock);
2980 LASSERT(imp->imp_replayable);
2981 /* Balanced in ptlrpc_free_committed, usually. */
2982 ptlrpc_request_addref(req);
2983 list_for_each_prev(tmp, &imp->imp_replay_list) {
2984 struct ptlrpc_request *iter = list_entry(tmp,
2985 struct ptlrpc_request,
2989 * We may have duplicate transnos if we create and then
2990 * open a file, or for closes retained if to match creating
2991 * opens, so use req->rq_xid as a secondary key.
2992 * (See bugs 684, 685, and 428.)
2993 * XXX no longer needed, but all opens need transnos!
2995 if (iter->rq_transno > req->rq_transno)
2998 if (iter->rq_transno == req->rq_transno) {
2999 LASSERT(iter->rq_xid != req->rq_xid);
3000 if (iter->rq_xid > req->rq_xid)
3004 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3008 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3012 * Send request and wait until it completes.
3013 * Returns request processing status.
3015 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3017 struct ptlrpc_request_set *set;
3021 LASSERT(req->rq_set == NULL);
3022 LASSERT(!req->rq_receiving_reply);
3024 set = ptlrpc_prep_set();
3026 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3030 /* for distributed debugging */
3031 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3033 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3034 ptlrpc_request_addref(req);
3035 ptlrpc_set_add_req(set, req);
3036 rc = ptlrpc_set_wait(NULL, set);
3037 ptlrpc_set_destroy(set);
3041 EXPORT_SYMBOL(ptlrpc_queue_wait);
3044 * Callback used for replayed requests reply processing.
3045 * In case of successful reply calls registered request replay callback.
3046 * In case of error restart replay process.
3048 static int ptlrpc_replay_interpret(const struct lu_env *env,
3049 struct ptlrpc_request *req,
3052 struct ptlrpc_replay_async_args *aa = args;
3053 struct obd_import *imp = req->rq_import;
3056 atomic_dec(&imp->imp_replay_inflight);
3059 * Note: if it is bulk replay (MDS-MDS replay), then even if
3060 * server got the request, but bulk transfer timeout, let's
3061 * replay the bulk req again
3063 if (!ptlrpc_client_replied(req) ||
3065 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3066 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3067 GOTO(out, rc = -ETIMEDOUT);
3070 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3071 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3072 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3073 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3075 /** VBR: check version failure */
3076 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3077 /** replay was failed due to version mismatch */
3078 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3079 spin_lock(&imp->imp_lock);
3080 imp->imp_vbr_failed = 1;
3081 spin_unlock(&imp->imp_lock);
3082 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3084 /** The transno had better not change over replay. */
3085 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3086 lustre_msg_get_transno(req->rq_repmsg) ||
3087 lustre_msg_get_transno(req->rq_repmsg) == 0,
3089 lustre_msg_get_transno(req->rq_reqmsg),
3090 lustre_msg_get_transno(req->rq_repmsg));
3093 spin_lock(&imp->imp_lock);
3094 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3095 spin_unlock(&imp->imp_lock);
3096 LASSERT(imp->imp_last_replay_transno);
3098 /* transaction number shouldn't be bigger than the latest replayed */
3099 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3100 DEBUG_REQ(D_ERROR, req,
3101 "Reported transno=%llu is bigger than replayed=%llu",
3103 lustre_msg_get_transno(req->rq_reqmsg));
3104 GOTO(out, rc = -EINVAL);
3107 DEBUG_REQ(D_HA, req, "got reply");
3109 /* let the callback do fixups, possibly including in the request */
3110 if (req->rq_replay_cb)
3111 req->rq_replay_cb(req);
3113 if (ptlrpc_client_replied(req) &&
3114 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3115 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3116 lustre_msg_get_status(req->rq_repmsg),
3117 aa->praa_old_status);
3120 * Note: If the replay fails for MDT-MDT recovery, let's
3121 * abort all of the following requests in the replay
3122 * and sending list, because MDT-MDT update requests
3123 * are dependent on each other, see LU-7039
3125 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3126 struct ptlrpc_request *free_req;
3127 struct ptlrpc_request *tmp;
3129 spin_lock(&imp->imp_lock);
3130 list_for_each_entry_safe(free_req, tmp,
3131 &imp->imp_replay_list,
3133 ptlrpc_free_request(free_req);
3136 list_for_each_entry_safe(free_req, tmp,
3137 &imp->imp_committed_list,
3139 ptlrpc_free_request(free_req);
3142 list_for_each_entry_safe(free_req, tmp,
3143 &imp->imp_delayed_list,
3145 spin_lock(&free_req->rq_lock);
3146 free_req->rq_err = 1;
3147 free_req->rq_status = -EIO;
3148 ptlrpc_client_wake_req(free_req);
3149 spin_unlock(&free_req->rq_lock);
3152 list_for_each_entry_safe(free_req, tmp,
3153 &imp->imp_sending_list,
3155 spin_lock(&free_req->rq_lock);
3156 free_req->rq_err = 1;
3157 free_req->rq_status = -EIO;
3158 ptlrpc_client_wake_req(free_req);
3159 spin_unlock(&free_req->rq_lock);
3161 spin_unlock(&imp->imp_lock);
3164 /* Put it back for re-replay. */
3165 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3169 * Errors while replay can set transno to 0, but
3170 * imp_last_replay_transno shouldn't be set to 0 anyway
3172 if (req->rq_transno == 0)
3173 CERROR("Transno is 0 during replay!\n");
3175 /* continue with recovery */
3176 rc = ptlrpc_import_recovery_state_machine(imp);
3178 req->rq_send_state = aa->praa_old_state;
3181 /* this replay failed, so restart recovery */
3182 ptlrpc_connect_import(imp);
3188 * Prepares and queues request for replay.
3189 * Adds it to ptlrpcd queue for actual sending.
3190 * Returns 0 on success.
3192 int ptlrpc_replay_req(struct ptlrpc_request *req)
3194 struct ptlrpc_replay_async_args *aa;
3198 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3200 aa = ptlrpc_req_async_args(aa, req);
3201 memset(aa, 0, sizeof(*aa));
3203 /* Prepare request to be resent with ptlrpcd */
3204 aa->praa_old_state = req->rq_send_state;
3205 req->rq_send_state = LUSTRE_IMP_REPLAY;
3206 req->rq_phase = RQ_PHASE_NEW;
3207 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3209 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3211 req->rq_interpret_reply = ptlrpc_replay_interpret;
3212 /* Readjust the timeout for current conditions */
3213 ptlrpc_at_set_req_timeout(req);
3215 /* Tell server net_latency to calculate how long to wait for reply. */
3216 lustre_msg_set_service_time(req->rq_reqmsg,
3217 ptlrpc_at_get_net_latency(req));
3218 DEBUG_REQ(D_HA, req, "REPLAY");
3220 atomic_inc(&req->rq_import->imp_replay_inflight);
3221 spin_lock(&req->rq_lock);
3222 req->rq_early_free_repbuf = 0;
3223 spin_unlock(&req->rq_lock);
3224 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3226 ptlrpcd_add_req(req);
3231 * Aborts all in-flight request on import \a imp sending and delayed lists
3233 void ptlrpc_abort_inflight(struct obd_import *imp)
3235 struct list_head *tmp, *n;
3239 * Make sure that no new requests get processed for this import.
3240 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3241 * this flag and then putting requests on sending_list or delayed_list.
3243 assert_spin_locked(&imp->imp_lock);
3246 * XXX locking? Maybe we should remove each request with the list
3247 * locked? Also, how do we know if the requests on the list are
3248 * being freed at this time?
3250 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3251 struct ptlrpc_request *req = list_entry(tmp,
3252 struct ptlrpc_request,
3255 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3257 spin_lock(&req->rq_lock);
3258 if (req->rq_import_generation < imp->imp_generation) {
3260 req->rq_status = -EIO;
3261 ptlrpc_client_wake_req(req);
3263 spin_unlock(&req->rq_lock);
3266 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3267 struct ptlrpc_request *req =
3268 list_entry(tmp, struct ptlrpc_request, rq_list);
3270 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3272 spin_lock(&req->rq_lock);
3273 if (req->rq_import_generation < imp->imp_generation) {
3275 req->rq_status = -EIO;
3276 ptlrpc_client_wake_req(req);
3278 spin_unlock(&req->rq_lock);
3282 * Last chance to free reqs left on the replay list, but we
3283 * will still leak reqs that haven't committed.
3285 if (imp->imp_replayable)
3286 ptlrpc_free_committed(imp);
3292 * Abort all uncompleted requests in request set \a set
3294 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3296 struct list_head *tmp, *pos;
3298 LASSERT(set != NULL);
3300 list_for_each_safe(pos, tmp, &set->set_requests) {
3301 struct ptlrpc_request *req =
3302 list_entry(pos, struct ptlrpc_request,
3305 spin_lock(&req->rq_lock);
3306 if (req->rq_phase != RQ_PHASE_RPC) {
3307 spin_unlock(&req->rq_lock);
3312 req->rq_status = -EINTR;
3313 ptlrpc_client_wake_req(req);
3314 spin_unlock(&req->rq_lock);
3319 * Initialize the XID for the node. This is common among all requests on
3320 * this node, and only requires the property that it is monotonically
3321 * increasing. It does not need to be sequential. Since this is also used
3322 * as the RDMA match bits, it is important that a single client NOT have
3323 * the same match bits for two different in-flight requests, hence we do
3324 * NOT want to have an XID per target or similar.
3326 * To avoid an unlikely collision between match bits after a client reboot
3327 * (which would deliver old data into the wrong RDMA buffer) initialize
3328 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3329 * If the time is clearly incorrect, we instead use a 62-bit random number.
3330 * In the worst case the random number will overflow 1M RPCs per second in
3331 * 9133 years, or permutations thereof.
3333 #define YEAR_2004 (1ULL << 30)
3334 void ptlrpc_init_xid(void)
3336 time64_t now = ktime_get_real_seconds();
3339 if (now < YEAR_2004) {
3340 get_random_bytes(&xid, sizeof(xid));
3342 xid |= (1ULL << 61);
3344 xid = (u64)now << 20;
3347 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3348 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3349 xid &= PTLRPC_BULK_OPS_MASK;
3350 atomic64_set(&ptlrpc_last_xid, xid);
3354 * Increase xid and returns resulting new value to the caller.
3356 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3357 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3358 * itself uses the last bulk xid needed, so the server can determine the
3359 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3360 * xid must align to a power-of-two value.
3362 * This is assumed to be true due to the initial ptlrpc_last_xid
3363 * value also being initialized to a power-of-two value. LU-1431
3365 __u64 ptlrpc_next_xid(void)
3367 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3371 * If request has a new allocated XID (new request or EINPROGRESS resend),
3372 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3373 * request to ensure previous bulk fails and avoid problems with lost replies
3374 * and therefore several transfers landing into the same buffer from different
3377 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3379 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3381 LASSERT(bd != NULL);
3384 * Generate new matchbits for all resend requests, including
3387 if (req->rq_resend) {
3388 __u64 old_mbits = req->rq_mbits;
3391 * First time resend on -EINPROGRESS will generate new xid,
3392 * so we can actually use the rq_xid as rq_mbits in such case,
3393 * however, it's bit hard to distinguish such resend with a
3394 * 'resend for the -EINPROGRESS resend'. To make it simple,
3395 * we opt to generate mbits for all resend cases.
3397 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3399 req->rq_mbits = ptlrpc_next_xid();
3402 * Old version transfers rq_xid to peer as
3405 spin_lock(&req->rq_import->imp_lock);
3406 list_del_init(&req->rq_unreplied_list);
3407 ptlrpc_assign_next_xid_nolock(req);
3408 spin_unlock(&req->rq_import->imp_lock);
3409 req->rq_mbits = req->rq_xid;
3411 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3412 old_mbits, req->rq_mbits);
3413 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3414 /* Request being sent first time, use xid as matchbits. */
3415 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3416 || req->rq_mbits == 0) {
3417 req->rq_mbits = req->rq_xid;
3419 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3421 req->rq_mbits -= total_md - 1;
3425 * Replay request, xid and matchbits have already been
3426 * correctly assigned.
3432 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3433 * that server can infer the number of bulks that were prepared,
3436 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3440 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3441 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3443 * It's ok to directly set the rq_xid here, since this xid bump
3444 * won't affect the request position in unreplied list.
3446 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3447 req->rq_xid = req->rq_mbits;
3451 * Get a glimpse at what next xid value might have been.
3452 * Returns possible next xid.
3454 __u64 ptlrpc_sample_next_xid(void)
3456 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3458 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3461 * Functions for operating ptlrpc workers.
3463 * A ptlrpc work is a function which will be running inside ptlrpc context.
3464 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3466 * 1. after a work is created, it can be used many times, that is:
3467 * handler = ptlrpcd_alloc_work();
3468 * ptlrpcd_queue_work();
3470 * queue it again when necessary:
3471 * ptlrpcd_queue_work();
3472 * ptlrpcd_destroy_work();
3473 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3474 * it will only be queued once in any time. Also as its name implies, it may
3475 * have delay before it really runs by ptlrpcd thread.
3477 struct ptlrpc_work_async_args {
3478 int (*cb)(const struct lu_env *, void *);
3482 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3484 /* re-initialize the req */
3485 req->rq_timeout = obd_timeout;
3486 req->rq_sent = ktime_get_real_seconds();
3487 req->rq_deadline = req->rq_sent + req->rq_timeout;
3488 req->rq_phase = RQ_PHASE_INTERPRET;
3489 req->rq_next_phase = RQ_PHASE_COMPLETE;
3490 req->rq_xid = ptlrpc_next_xid();
3491 req->rq_import_generation = req->rq_import->imp_generation;
3493 ptlrpcd_add_req(req);
3496 static int work_interpreter(const struct lu_env *env,
3497 struct ptlrpc_request *req, void *args, int rc)
3499 struct ptlrpc_work_async_args *arg = args;
3501 LASSERT(ptlrpcd_check_work(req));
3502 LASSERT(arg->cb != NULL);
3504 rc = arg->cb(env, arg->cbdata);
3506 list_del_init(&req->rq_set_chain);
3509 if (atomic_dec_return(&req->rq_refcount) > 1) {
3510 atomic_set(&req->rq_refcount, 2);
3511 ptlrpcd_add_work_req(req);
3516 static int worker_format;
3518 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3520 return req->rq_pill.rc_fmt == (void *)&worker_format;
3524 * Create a work for ptlrpc.
3526 void *ptlrpcd_alloc_work(struct obd_import *imp,
3527 int (*cb)(const struct lu_env *, void *), void *cbdata)
3529 struct ptlrpc_request *req = NULL;
3530 struct ptlrpc_work_async_args *args;
3536 RETURN(ERR_PTR(-EINVAL));
3538 /* copy some code from deprecated fakereq. */
3539 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3541 CERROR("ptlrpc: run out of memory!\n");
3542 RETURN(ERR_PTR(-ENOMEM));
3545 ptlrpc_cli_req_init(req);
3547 req->rq_send_state = LUSTRE_IMP_FULL;
3548 req->rq_type = PTL_RPC_MSG_REQUEST;
3549 req->rq_import = class_import_get(imp);
3550 req->rq_interpret_reply = work_interpreter;
3551 /* don't want reply */
3552 req->rq_no_delay = req->rq_no_resend = 1;
3553 req->rq_pill.rc_fmt = (void *)&worker_format;
3555 args = ptlrpc_req_async_args(args, req);
3557 args->cbdata = cbdata;
3561 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3563 void ptlrpcd_destroy_work(void *handler)
3565 struct ptlrpc_request *req = handler;
3568 ptlrpc_req_finished(req);
3570 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3572 int ptlrpcd_queue_work(void *handler)
3574 struct ptlrpc_request *req = handler;
3577 * Check if the req is already being queued.
3579 * Here comes a trick: it lacks a way of checking if a req is being
3580 * processed reliably in ptlrpc. Here I have to use refcount of req
3581 * for this purpose. This is okay because the caller should use this
3582 * req as opaque data. - Jinshan
3584 LASSERT(atomic_read(&req->rq_refcount) > 0);
3585 if (atomic_inc_return(&req->rq_refcount) == 2)
3586 ptlrpcd_add_work_req(req);
3589 EXPORT_SYMBOL(ptlrpcd_queue_work);