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 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
50 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
51 .release_frags = ptlrpc_release_bulk_page_pin,
53 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
55 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
56 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
57 .release_frags = ptlrpc_release_bulk_noop,
59 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
61 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
62 .add_iov_frag = ptlrpc_prep_bulk_frag,
64 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
66 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
67 static int ptlrpcd_check_work(struct ptlrpc_request *req);
68 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
71 * Initialize passed in client structure \a cl.
73 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
74 struct ptlrpc_client *cl)
76 cl->cli_request_portal = req_portal;
77 cl->cli_reply_portal = rep_portal;
80 EXPORT_SYMBOL(ptlrpc_init_client);
83 * Return PortalRPC connection for remore uud \a uuid
85 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
86 lnet_nid_t nid4refnet)
88 struct ptlrpc_connection *c;
90 struct lnet_process_id peer;
94 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
95 * before accessing its values.
97 /* coverity[uninit_use_in_call] */
98 peer.nid = nid4refnet;
99 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
101 CNETERR("cannot find peer %s!\n", uuid->uuid);
105 c = ptlrpc_connection_get(peer, self, uuid);
107 memcpy(c->c_remote_uuid.uuid,
108 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
111 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
117 * Allocate and initialize new bulk descriptor on the sender.
118 * Returns pointer to the descriptor or NULL on error.
120 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
121 unsigned int max_brw,
122 enum ptlrpc_bulk_op_type type,
124 const struct ptlrpc_bulk_frag_ops *ops)
126 struct ptlrpc_bulk_desc *desc;
129 /* ensure that only one of KIOV or IOVEC is set but not both */
130 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
131 ops->add_kiov_frag != NULL) ||
132 (ptlrpc_is_bulk_desc_kvec(type) &&
133 ops->add_iov_frag != NULL));
138 if (type & PTLRPC_BULK_BUF_KIOV) {
139 OBD_ALLOC_LARGE(GET_KIOV(desc),
140 nfrags * sizeof(*GET_KIOV(desc)));
144 OBD_ALLOC_LARGE(GET_KVEC(desc),
145 nfrags * sizeof(*GET_KVEC(desc)));
150 spin_lock_init(&desc->bd_lock);
151 init_waitqueue_head(&desc->bd_waitq);
152 desc->bd_max_iov = nfrags;
153 desc->bd_iov_count = 0;
154 desc->bd_portal = portal;
155 desc->bd_type = type;
156 desc->bd_md_count = 0;
157 desc->bd_frag_ops = ops;
158 LASSERT(max_brw > 0);
159 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
161 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
162 * node. Negotiated ocd_brw_size will always be <= this number.
164 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
165 LNetInvalidateMDHandle(&desc->bd_mds[i]);
174 * Prepare bulk descriptor for specified outgoing request \a req that
175 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
176 * the bulk to be sent. Used on client-side.
177 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
180 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
182 unsigned int max_brw,
185 const struct ptlrpc_bulk_frag_ops
188 struct obd_import *imp = req->rq_import;
189 struct ptlrpc_bulk_desc *desc;
192 LASSERT(ptlrpc_is_bulk_op_passive(type));
194 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
198 desc->bd_import_generation = req->rq_import_generation;
199 desc->bd_import = class_import_get(imp);
202 desc->bd_cbid.cbid_fn = client_bulk_callback;
203 desc->bd_cbid.cbid_arg = desc;
205 /* This makes req own desc, and free it when she frees herself */
210 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
212 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
213 struct page *page, int pageoffset, int len,
218 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
219 LASSERT(page != NULL);
220 LASSERT(pageoffset >= 0);
222 LASSERT(pageoffset + len <= PAGE_SIZE);
223 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
225 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
232 kiov->kiov_page = page;
233 kiov->kiov_offset = pageoffset;
234 kiov->kiov_len = len;
236 desc->bd_iov_count++;
238 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
240 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
247 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
248 LASSERT(frag != NULL);
250 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
252 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
256 iovec->iov_base = frag;
257 iovec->iov_len = len;
259 desc->bd_iov_count++;
261 RETURN(desc->bd_nob);
263 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
265 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
269 LASSERT(desc != NULL);
270 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
271 LASSERT(desc->bd_md_count == 0); /* network hands off */
272 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
273 LASSERT(desc->bd_frag_ops != NULL);
275 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
276 sptlrpc_enc_pool_put_pages(desc);
279 class_export_put(desc->bd_export);
281 class_import_put(desc->bd_import);
283 if (desc->bd_frag_ops->release_frags != NULL)
284 desc->bd_frag_ops->release_frags(desc);
286 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
287 OBD_FREE_LARGE(GET_KIOV(desc),
288 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
290 OBD_FREE_LARGE(GET_KVEC(desc),
291 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
295 EXPORT_SYMBOL(ptlrpc_free_bulk);
298 * Set server timelimit for this req, i.e. how long are we willing to wait
299 * for reply before timing out this request.
301 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
307 LASSERT(req->rq_import);
310 /* non-AT settings */
312 * \a imp_server_timeout means this is reverse import and
313 * we send (currently only) ASTs to the client and cannot afford
314 * to wait too long for the reply, otherwise the other client
315 * (because of which we are sending this request) would
316 * timeout waiting for us
318 req->rq_timeout = req->rq_import->imp_server_timeout ?
319 obd_timeout / 2 : obd_timeout;
321 at = &req->rq_import->imp_at;
322 idx = import_at_get_index(req->rq_import,
323 req->rq_request_portal);
324 serv_est = at_get(&at->iat_service_estimate[idx]);
325 req->rq_timeout = at_est2timeout(serv_est);
328 * We could get even fancier here, using history to predict increased
333 * Let the server know what this RPC timeout is by putting it in the
336 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
338 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
340 /* Adjust max service estimate based on server value */
341 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
342 unsigned int serv_est)
348 LASSERT(req->rq_import);
349 at = &req->rq_import->imp_at;
351 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
353 * max service estimates are tracked on the server side,
354 * so just keep minimal history here
356 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
359 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
360 req->rq_import->imp_obd->obd_name,
361 req->rq_request_portal,
362 oldse, at_get(&at->iat_service_estimate[idx]));
365 /* Expected network latency per remote node (secs) */
366 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
368 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
371 /* Adjust expected network latency */
372 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
373 unsigned int service_time)
375 unsigned int nl, oldnl;
377 time64_t now = ktime_get_real_seconds();
379 LASSERT(req->rq_import);
381 if (service_time > now - req->rq_sent + 3) {
383 * b=16408, however, this can also happen if early reply
384 * is lost and client RPC is expired and resent, early reply
385 * or reply of original RPC can still be fit in reply buffer
386 * of resent RPC, now client is measuring time from the
387 * resent time, but server sent back service time of original
390 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
391 D_ADAPTTO : D_WARNING,
392 "Reported service time %u > total measured time %lld\n",
393 service_time, now - req->rq_sent);
397 /* Network latency is total time less server processing time */
398 nl = max_t(int, now - req->rq_sent -
399 service_time, 0) + 1; /* st rounding */
400 at = &req->rq_import->imp_at;
402 oldnl = at_measured(&at->iat_net_latency, nl);
405 "The network latency for %s (nid %s) has changed from %d to %d\n",
406 req->rq_import->imp_obd->obd_name,
407 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
408 oldnl, at_get(&at->iat_net_latency));
411 static int unpack_reply(struct ptlrpc_request *req)
415 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
416 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
418 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
423 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
425 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
432 * Handle an early reply message, called with the rq_lock held.
433 * If anything goes wrong just ignore it - same as if it never happened
435 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
436 __must_hold(&req->rq_lock)
438 struct ptlrpc_request *early_req;
444 spin_unlock(&req->rq_lock);
446 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
448 spin_lock(&req->rq_lock);
452 rc = unpack_reply(early_req);
454 sptlrpc_cli_finish_early_reply(early_req);
455 spin_lock(&req->rq_lock);
460 * Use new timeout value just to adjust the local value for this
461 * request, don't include it into at_history. It is unclear yet why
462 * service time increased and should it be counted or skipped, e.g.
463 * that can be recovery case or some error or server, the real reply
464 * will add all new data if it is worth to add.
466 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
467 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
469 /* Network latency can be adjusted, it is pure network delays */
470 ptlrpc_at_adj_net_latency(req,
471 lustre_msg_get_service_time(early_req->rq_repmsg));
473 sptlrpc_cli_finish_early_reply(early_req);
475 spin_lock(&req->rq_lock);
476 olddl = req->rq_deadline;
478 * server assumes it now has rq_timeout from when the request
479 * arrived, so the client should give it at least that long.
480 * since we don't know the arrival time we'll use the original
483 req->rq_deadline = req->rq_sent + req->rq_timeout +
484 ptlrpc_at_get_net_latency(req);
486 DEBUG_REQ(D_ADAPTTO, req,
487 "Early reply #%d, new deadline in %llds (%llds)",
489 req->rq_deadline - ktime_get_real_seconds(),
490 req->rq_deadline - olddl);
495 static struct kmem_cache *request_cache;
497 int ptlrpc_request_cache_init(void)
499 request_cache = kmem_cache_create("ptlrpc_cache",
500 sizeof(struct ptlrpc_request),
501 0, SLAB_HWCACHE_ALIGN, NULL);
502 return request_cache ? 0 : -ENOMEM;
505 void ptlrpc_request_cache_fini(void)
507 kmem_cache_destroy(request_cache);
510 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
512 struct ptlrpc_request *req;
514 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
518 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
520 OBD_SLAB_FREE_PTR(req, request_cache);
524 * Wind down request pool \a pool.
525 * Frees all requests from the pool too
527 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
529 struct list_head *l, *tmp;
530 struct ptlrpc_request *req;
532 LASSERT(pool != NULL);
534 spin_lock(&pool->prp_lock);
535 list_for_each_safe(l, tmp, &pool->prp_req_list) {
536 req = list_entry(l, struct ptlrpc_request, rq_list);
537 list_del(&req->rq_list);
538 LASSERT(req->rq_reqbuf);
539 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
540 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
541 ptlrpc_request_cache_free(req);
543 spin_unlock(&pool->prp_lock);
544 OBD_FREE(pool, sizeof(*pool));
546 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
549 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
551 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
556 while (size < pool->prp_rq_size)
559 LASSERTF(list_empty(&pool->prp_req_list) ||
560 size == pool->prp_rq_size,
561 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
562 pool->prp_rq_size, size);
564 spin_lock(&pool->prp_lock);
565 pool->prp_rq_size = size;
566 for (i = 0; i < num_rq; i++) {
567 struct ptlrpc_request *req;
568 struct lustre_msg *msg;
570 spin_unlock(&pool->prp_lock);
571 req = ptlrpc_request_cache_alloc(GFP_NOFS);
574 OBD_ALLOC_LARGE(msg, size);
576 ptlrpc_request_cache_free(req);
579 req->rq_reqbuf = msg;
580 req->rq_reqbuf_len = size;
582 spin_lock(&pool->prp_lock);
583 list_add_tail(&req->rq_list, &pool->prp_req_list);
585 spin_unlock(&pool->prp_lock);
588 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
591 * Create and initialize new request pool with given attributes:
592 * \a num_rq - initial number of requests to create for the pool
593 * \a msgsize - maximum message size possible for requests in thid pool
594 * \a populate_pool - function to be called when more requests need to be added
596 * Returns pointer to newly created pool or NULL on error.
598 struct ptlrpc_request_pool *
599 ptlrpc_init_rq_pool(int num_rq, int msgsize,
600 int (*populate_pool)(struct ptlrpc_request_pool *, int))
602 struct ptlrpc_request_pool *pool;
604 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
609 * Request next power of two for the allocation, because internally
610 * kernel would do exactly this
612 spin_lock_init(&pool->prp_lock);
613 INIT_LIST_HEAD(&pool->prp_req_list);
614 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
615 pool->prp_populate = populate_pool;
617 populate_pool(pool, num_rq);
621 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
624 * Fetches one request from pool \a pool
626 static struct ptlrpc_request *
627 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
629 struct ptlrpc_request *request;
630 struct lustre_msg *reqbuf;
635 spin_lock(&pool->prp_lock);
638 * See if we have anything in a pool, and bail out if nothing,
639 * in writeout path, where this matters, this is safe to do, because
640 * nothing is lost in this case, and when some in-flight requests
641 * complete, this code will be called again.
643 if (unlikely(list_empty(&pool->prp_req_list))) {
644 spin_unlock(&pool->prp_lock);
648 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
650 list_del_init(&request->rq_list);
651 spin_unlock(&pool->prp_lock);
653 LASSERT(request->rq_reqbuf);
654 LASSERT(request->rq_pool);
656 reqbuf = request->rq_reqbuf;
657 memset(request, 0, sizeof(*request));
658 request->rq_reqbuf = reqbuf;
659 request->rq_reqbuf_len = pool->prp_rq_size;
660 request->rq_pool = pool;
666 * Returns freed \a request to pool.
668 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
670 struct ptlrpc_request_pool *pool = request->rq_pool;
672 spin_lock(&pool->prp_lock);
673 LASSERT(list_empty(&request->rq_list));
674 LASSERT(!request->rq_receiving_reply);
675 list_add_tail(&request->rq_list, &pool->prp_req_list);
676 spin_unlock(&pool->prp_lock);
679 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
681 struct obd_import *imp = req->rq_import;
682 struct list_head *tmp;
683 struct ptlrpc_request *iter;
685 assert_spin_locked(&imp->imp_lock);
686 LASSERT(list_empty(&req->rq_unreplied_list));
688 /* unreplied list is sorted by xid in ascending order */
689 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
690 iter = list_entry(tmp, struct ptlrpc_request,
693 LASSERT(req->rq_xid != iter->rq_xid);
694 if (req->rq_xid < iter->rq_xid)
696 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
699 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
702 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
704 req->rq_xid = ptlrpc_next_xid();
705 ptlrpc_add_unreplied(req);
708 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
710 spin_lock(&req->rq_import->imp_lock);
711 ptlrpc_assign_next_xid_nolock(req);
712 spin_unlock(&req->rq_import->imp_lock);
715 static __u64 ptlrpc_last_xid;
716 static spinlock_t ptlrpc_last_xid_lock;
718 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
719 __u32 version, int opcode, char **bufs,
720 struct ptlrpc_cli_ctx *ctx)
723 struct obd_import *imp;
729 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
730 imp = request->rq_import;
731 lengths = request->rq_pill.rc_area[RCL_CLIENT];
734 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
736 rc = sptlrpc_req_get_ctx(request);
740 sptlrpc_req_set_flavor(request, opcode);
742 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
747 lustre_msg_add_version(request->rq_reqmsg, version);
748 request->rq_send_state = LUSTRE_IMP_FULL;
749 request->rq_type = PTL_RPC_MSG_REQUEST;
751 request->rq_req_cbid.cbid_fn = request_out_callback;
752 request->rq_req_cbid.cbid_arg = request;
754 request->rq_reply_cbid.cbid_fn = reply_in_callback;
755 request->rq_reply_cbid.cbid_arg = request;
757 request->rq_reply_deadline = 0;
758 request->rq_bulk_deadline = 0;
759 request->rq_req_deadline = 0;
760 request->rq_phase = RQ_PHASE_NEW;
761 request->rq_next_phase = RQ_PHASE_UNDEFINED;
763 request->rq_request_portal = imp->imp_client->cli_request_portal;
764 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
766 ptlrpc_at_set_req_timeout(request);
768 lustre_msg_set_opc(request->rq_reqmsg, opcode);
770 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
771 if (cfs_fail_val == opcode) {
772 time64_t *fail_t = NULL, *fail2_t = NULL;
774 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
775 fail_t = &request->rq_bulk_deadline;
776 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
777 fail_t = &request->rq_reply_deadline;
778 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
779 fail_t = &request->rq_req_deadline;
780 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
781 fail_t = &request->rq_reply_deadline;
782 fail2_t = &request->rq_bulk_deadline;
783 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
784 time64_t now = ktime_get_real_seconds();
785 spin_lock(&ptlrpc_last_xid_lock);
786 ptlrpc_last_xid = ((__u64)now >> 4) << 24;
787 spin_unlock(&ptlrpc_last_xid_lock);
791 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
794 *fail2_t = ktime_get_real_seconds() +
798 * The RPC is infected, let the test to change the
801 msleep(4 * MSEC_PER_SEC);
804 ptlrpc_assign_next_xid(request);
809 LASSERT(!request->rq_pool);
810 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
812 class_import_put(imp);
816 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
819 * Pack request buffers for network transfer, performing necessary encryption
820 * steps if necessary.
822 int ptlrpc_request_pack(struct ptlrpc_request *request,
823 __u32 version, int opcode)
827 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
832 * For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
833 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
834 * have to send old ptlrpc_body to keep interoprability with these
837 * Only three kinds of server->client RPCs so far:
842 * XXX This should be removed whenever we drop the interoprability with
843 * the these old clients.
845 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
846 opcode == LDLM_GL_CALLBACK)
847 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
848 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
852 EXPORT_SYMBOL(ptlrpc_request_pack);
855 * Helper function to allocate new request on import \a imp
856 * and possibly using existing request from pool \a pool if provided.
857 * Returns allocated request structure with import field filled or
861 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
862 struct ptlrpc_request_pool *pool)
864 struct ptlrpc_request *request = NULL;
866 request = ptlrpc_request_cache_alloc(GFP_NOFS);
868 if (!request && pool)
869 request = ptlrpc_prep_req_from_pool(pool);
872 ptlrpc_cli_req_init(request);
874 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
875 LASSERT(imp != LP_POISON);
876 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
878 LASSERT(imp->imp_client != LP_POISON);
880 request->rq_import = class_import_get(imp);
882 CERROR("request allocation out of memory\n");
889 * Helper function for creating a request.
890 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
891 * buffer structures according to capsule template \a format.
892 * Returns allocated request structure pointer or NULL on error.
894 static struct ptlrpc_request *
895 ptlrpc_request_alloc_internal(struct obd_import *imp,
896 struct ptlrpc_request_pool *pool,
897 const struct req_format *format)
899 struct ptlrpc_request *request;
902 request = __ptlrpc_request_alloc(imp, pool);
907 * initiate connection if needed when the import has been
908 * referenced by the new request to avoid races with disconnect
910 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
913 CDEBUG_LIMIT(imp->imp_idle_debug,
914 "%s: reconnect after %llds idle\n",
915 imp->imp_obd->obd_name, ktime_get_real_seconds() -
916 imp->imp_last_reply_time);
917 spin_lock(&imp->imp_lock);
918 if (imp->imp_state == LUSTRE_IMP_IDLE) {
919 imp->imp_generation++;
920 imp->imp_initiated_at = imp->imp_generation;
921 imp->imp_state = LUSTRE_IMP_NEW;
924 spin_unlock(&imp->imp_lock);
926 rc = ptlrpc_connect_import(imp);
928 ptlrpc_request_free(request);
931 ptlrpc_pinger_add_import(imp);
935 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
936 req_capsule_set(&request->rq_pill, format);
941 * Allocate new request structure for import \a imp and initialize its
942 * buffer structure according to capsule template \a format.
944 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
945 const struct req_format *format)
947 return ptlrpc_request_alloc_internal(imp, NULL, format);
949 EXPORT_SYMBOL(ptlrpc_request_alloc);
952 * Allocate new request structure for import \a imp from pool \a pool and
953 * initialize its buffer structure according to capsule template \a format.
955 struct ptlrpc_request *
956 ptlrpc_request_alloc_pool(struct obd_import *imp,
957 struct ptlrpc_request_pool *pool,
958 const struct req_format *format)
960 return ptlrpc_request_alloc_internal(imp, pool, format);
962 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
965 * For requests not from pool, free memory of the request structure.
966 * For requests obtained from a pool earlier, return request back to pool.
968 void ptlrpc_request_free(struct ptlrpc_request *request)
970 if (request->rq_pool)
971 __ptlrpc_free_req_to_pool(request);
973 ptlrpc_request_cache_free(request);
975 EXPORT_SYMBOL(ptlrpc_request_free);
978 * Allocate new request for operatione \a opcode and immediatelly pack it for
980 * Only used for simple requests like OBD_PING where the only important
981 * part of the request is operation itself.
982 * Returns allocated request or NULL on error.
984 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
985 const struct req_format *format,
986 __u32 version, int opcode)
988 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
992 rc = ptlrpc_request_pack(req, version, opcode);
994 ptlrpc_request_free(req);
1000 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1003 * Allocate and initialize new request set structure on the current CPT.
1004 * Returns a pointer to the newly allocated set structure or NULL on error.
1006 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1008 struct ptlrpc_request_set *set;
1012 cpt = cfs_cpt_current(cfs_cpt_table, 0);
1013 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1016 atomic_set(&set->set_refcount, 1);
1017 INIT_LIST_HEAD(&set->set_requests);
1018 init_waitqueue_head(&set->set_waitq);
1019 atomic_set(&set->set_new_count, 0);
1020 atomic_set(&set->set_remaining, 0);
1021 spin_lock_init(&set->set_new_req_lock);
1022 INIT_LIST_HEAD(&set->set_new_requests);
1023 set->set_max_inflight = UINT_MAX;
1024 set->set_producer = NULL;
1025 set->set_producer_arg = NULL;
1030 EXPORT_SYMBOL(ptlrpc_prep_set);
1033 * Allocate and initialize new request set structure with flow control
1034 * extension. This extension allows to control the number of requests in-flight
1035 * for the whole set. A callback function to generate requests must be provided
1036 * and the request set will keep the number of requests sent over the wire to
1038 * Returns a pointer to the newly allocated set structure or NULL on error.
1040 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1044 struct ptlrpc_request_set *set;
1046 set = ptlrpc_prep_set();
1050 set->set_max_inflight = max;
1051 set->set_producer = func;
1052 set->set_producer_arg = arg;
1058 * Wind down and free request set structure previously allocated with
1060 * Ensures that all requests on the set have completed and removes
1061 * all requests from the request list in a set.
1062 * If any unsent request happen to be on the list, pretends that they got
1063 * an error in flight and calls their completion handler.
1065 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1067 struct list_head *tmp;
1068 struct list_head *next;
1074 /* Requests on the set should either all be completed, or all be new */
1075 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1076 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1077 list_for_each(tmp, &set->set_requests) {
1078 struct ptlrpc_request *req =
1079 list_entry(tmp, struct ptlrpc_request,
1082 LASSERT(req->rq_phase == expected_phase);
1086 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1087 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1088 atomic_read(&set->set_remaining), n);
1090 list_for_each_safe(tmp, next, &set->set_requests) {
1091 struct ptlrpc_request *req =
1092 list_entry(tmp, struct ptlrpc_request,
1094 list_del_init(&req->rq_set_chain);
1096 LASSERT(req->rq_phase == expected_phase);
1098 if (req->rq_phase == RQ_PHASE_NEW) {
1099 ptlrpc_req_interpret(NULL, req, -EBADR);
1100 atomic_dec(&set->set_remaining);
1103 spin_lock(&req->rq_lock);
1105 req->rq_invalid_rqset = 0;
1106 spin_unlock(&req->rq_lock);
1108 ptlrpc_req_finished(req);
1111 LASSERT(atomic_read(&set->set_remaining) == 0);
1113 ptlrpc_reqset_put(set);
1116 EXPORT_SYMBOL(ptlrpc_set_destroy);
1119 * Add a new request to the general purpose request set.
1120 * Assumes request reference from the caller.
1122 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1123 struct ptlrpc_request *req)
1125 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1126 LASSERT(list_empty(&req->rq_set_chain));
1128 if (req->rq_allow_intr)
1129 set->set_allow_intr = 1;
1131 /* The set takes over the caller's request reference */
1132 list_add_tail(&req->rq_set_chain, &set->set_requests);
1134 atomic_inc(&set->set_remaining);
1135 req->rq_queued_time = ktime_get_seconds();
1138 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1140 if (set->set_producer)
1142 * If the request set has a producer callback, the RPC must be
1143 * sent straight away
1145 ptlrpc_send_new_req(req);
1147 EXPORT_SYMBOL(ptlrpc_set_add_req);
1150 * Add a request to a request with dedicated server thread
1151 * and wake the thread to make any necessary processing.
1152 * Currently only used for ptlrpcd.
1154 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1155 struct ptlrpc_request *req)
1157 struct ptlrpc_request_set *set = pc->pc_set;
1160 LASSERT(req->rq_set == NULL);
1161 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1163 spin_lock(&set->set_new_req_lock);
1165 * The set takes over the caller's request reference.
1168 req->rq_queued_time = ktime_get_seconds();
1169 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1170 count = atomic_inc_return(&set->set_new_count);
1171 spin_unlock(&set->set_new_req_lock);
1173 /* Only need to call wakeup once for the first entry. */
1175 wake_up(&set->set_waitq);
1178 * XXX: It maybe unnecessary to wakeup all the partners. But to
1179 * guarantee the async RPC can be processed ASAP, we have
1180 * no other better choice. It maybe fixed in future.
1182 for (i = 0; i < pc->pc_npartners; i++)
1183 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1188 * Based on the current state of the import, determine if the request
1189 * can be sent, is an error, or should be delayed.
1191 * Returns true if this request should be delayed. If false, and
1192 * *status is set, then the request can not be sent and *status is the
1193 * error code. If false and status is 0, then request can be sent.
1195 * The imp->imp_lock must be held.
1197 static int ptlrpc_import_delay_req(struct obd_import *imp,
1198 struct ptlrpc_request *req, int *status)
1206 if (req->rq_ctx_init || req->rq_ctx_fini) {
1207 /* always allow ctx init/fini rpc go through */
1208 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1209 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1211 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1212 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1215 * pings or MDS-equivalent STATFS may safely
1218 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1219 D_HA : D_ERROR, req, "IMP_CLOSED ");
1221 } else if (ptlrpc_send_limit_expired(req)) {
1222 /* probably doesn't need to be a D_ERROR afterinitial testing */
1223 DEBUG_REQ(D_HA, req, "send limit expired ");
1224 *status = -ETIMEDOUT;
1225 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1226 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1227 ;/* allow CONNECT even if import is invalid */
1228 if (atomic_read(&imp->imp_inval_count) != 0) {
1229 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1232 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1233 if (!imp->imp_deactive)
1234 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1235 *status = -ESHUTDOWN; /* b=12940 */
1236 } else if (req->rq_import_generation != imp->imp_generation) {
1237 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1239 } else if (req->rq_send_state != imp->imp_state) {
1240 /* invalidate in progress - any requests should be drop */
1241 if (atomic_read(&imp->imp_inval_count) != 0) {
1242 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1244 } else if (req->rq_no_delay &&
1245 imp->imp_generation != imp->imp_initiated_at) {
1246 /* ignore nodelay for requests initiating connections */
1247 *status = -EWOULDBLOCK;
1248 } else if (req->rq_allow_replay &&
1249 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1250 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1251 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1252 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1253 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1263 * Decide if the error message should be printed to the console or not.
1264 * Makes its decision based on request type, status, and failure frequency.
1266 * \param[in] req request that failed and may need a console message
1268 * \retval false if no message should be printed
1269 * \retval true if console message should be printed
1271 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1273 LASSERT(req->rq_reqmsg != NULL);
1275 /* Suppress particular reconnect errors which are to be expected. */
1276 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1277 /* Suppress timed out reconnect requests */
1278 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1283 * Suppress most unavailable/again reconnect requests, but
1284 * print occasionally so it is clear client is trying to
1285 * connect to a server where no target is running.
1287 if ((err == -ENODEV || err == -EAGAIN) &&
1288 req->rq_import->imp_conn_cnt % 30 != 20)
1292 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1293 /* -EAGAIN is normal when using POSIX flocks */
1296 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1297 (req->rq_xid & 0xf) != 10)
1298 /* Suppress most ping requests, they may fail occasionally */
1305 * Check request processing status.
1306 * Returns the status.
1308 static int ptlrpc_check_status(struct ptlrpc_request *req)
1313 err = lustre_msg_get_status(req->rq_repmsg);
1314 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1315 struct obd_import *imp = req->rq_import;
1316 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1317 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1319 if (ptlrpc_console_allow(req, opc, err))
1320 LCONSOLE_ERROR_MSG(0x11,
1321 "%s: operation %s to node %s failed: rc = %d\n",
1322 imp->imp_obd->obd_name,
1324 libcfs_nid2str(nid), err);
1325 RETURN(err < 0 ? err : -EINVAL);
1329 DEBUG_REQ(D_INFO, req, "status is %d", err);
1330 } else if (err > 0) {
1331 /* XXX: translate this error from net to host */
1332 DEBUG_REQ(D_INFO, req, "status is %d", err);
1339 * save pre-versions of objects into request for replay.
1340 * Versions are obtained from server reply.
1343 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1345 struct lustre_msg *repmsg = req->rq_repmsg;
1346 struct lustre_msg *reqmsg = req->rq_reqmsg;
1347 __u64 *versions = lustre_msg_get_versions(repmsg);
1350 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1354 lustre_msg_set_versions(reqmsg, versions);
1355 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1356 versions[0], versions[1]);
1361 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1363 struct ptlrpc_request *req;
1365 assert_spin_locked(&imp->imp_lock);
1366 if (list_empty(&imp->imp_unreplied_list))
1369 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1371 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1373 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1374 imp->imp_known_replied_xid = req->rq_xid - 1;
1376 return req->rq_xid - 1;
1380 * Callback function called when client receives RPC reply for \a req.
1381 * Returns 0 on success or error code.
1382 * The return alue would be assigned to req->rq_status by the caller
1383 * as request processing status.
1384 * This function also decides if the request needs to be saved for later replay.
1386 static int after_reply(struct ptlrpc_request *req)
1388 struct obd_import *imp = req->rq_import;
1389 struct obd_device *obd = req->rq_import->imp_obd;
1396 LASSERT(obd != NULL);
1397 /* repbuf must be unlinked */
1398 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1400 if (req->rq_reply_truncated) {
1401 if (ptlrpc_no_resend(req)) {
1402 DEBUG_REQ(D_ERROR, req,
1403 "reply buffer overflow, expected: %d, actual size: %d",
1404 req->rq_nob_received, req->rq_repbuf_len);
1408 sptlrpc_cli_free_repbuf(req);
1410 * Pass the required reply buffer size (include
1411 * space for early reply).
1412 * NB: no need to roundup because alloc_repbuf
1415 req->rq_replen = req->rq_nob_received;
1416 req->rq_nob_received = 0;
1417 spin_lock(&req->rq_lock);
1419 spin_unlock(&req->rq_lock);
1423 work_start = ktime_get_real();
1424 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1427 * NB Until this point, the whole of the incoming message,
1428 * including buflens, status etc is in the sender's byte order.
1430 rc = sptlrpc_cli_unwrap_reply(req);
1432 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1437 * Security layer unwrap might ask resend this request.
1442 rc = unpack_reply(req);
1446 /* retry indefinitely on EINPROGRESS */
1447 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1448 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1449 time64_t now = ktime_get_real_seconds();
1451 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1452 spin_lock(&req->rq_lock);
1454 spin_unlock(&req->rq_lock);
1455 req->rq_nr_resend++;
1457 /* Readjust the timeout for current conditions */
1458 ptlrpc_at_set_req_timeout(req);
1460 * delay resend to give a chance to the server to get ready.
1461 * The delay is increased by 1s on every resend and is capped to
1462 * the current request timeout (i.e. obd_timeout if AT is off,
1463 * or AT service time x 125% + 5s, see at_est2timeout)
1465 if (req->rq_nr_resend > req->rq_timeout)
1466 req->rq_sent = now + req->rq_timeout;
1468 req->rq_sent = now + req->rq_nr_resend;
1470 /* Resend for EINPROGRESS will use a new XID */
1471 spin_lock(&imp->imp_lock);
1472 list_del_init(&req->rq_unreplied_list);
1473 spin_unlock(&imp->imp_lock);
1478 if (obd->obd_svc_stats) {
1479 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1481 ptlrpc_lprocfs_rpc_sent(req, timediff);
1484 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1485 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1486 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1487 lustre_msg_get_type(req->rq_repmsg));
1491 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1492 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1493 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1494 ptlrpc_at_adj_net_latency(req,
1495 lustre_msg_get_service_time(req->rq_repmsg));
1497 rc = ptlrpc_check_status(req);
1501 * Either we've been evicted, or the server has failed for
1502 * some reason. Try to reconnect, and if that fails, punt to
1505 if (ptlrpc_recoverable_error(rc)) {
1506 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1507 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1510 ptlrpc_request_handle_notconn(req);
1515 * Let's look if server sent slv. Do it only for RPC with
1518 ldlm_cli_update_pool(req);
1522 * Store transno in reqmsg for replay.
1524 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1525 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1526 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1529 if (imp->imp_replayable) {
1530 spin_lock(&imp->imp_lock);
1532 * No point in adding already-committed requests to the replay
1533 * list, we will just remove them immediately. b=9829
1535 if (req->rq_transno != 0 &&
1537 lustre_msg_get_last_committed(req->rq_repmsg) ||
1539 /** version recovery */
1540 ptlrpc_save_versions(req);
1541 ptlrpc_retain_replayable_request(req, imp);
1542 } else if (req->rq_commit_cb &&
1543 list_empty(&req->rq_replay_list)) {
1545 * NB: don't call rq_commit_cb if it's already on
1546 * rq_replay_list, ptlrpc_free_committed() will call
1547 * it later, see LU-3618 for details
1549 spin_unlock(&imp->imp_lock);
1550 req->rq_commit_cb(req);
1551 spin_lock(&imp->imp_lock);
1555 * Replay-enabled imports return commit-status information.
1557 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1558 if (likely(committed > imp->imp_peer_committed_transno))
1559 imp->imp_peer_committed_transno = committed;
1561 ptlrpc_free_committed(imp);
1563 if (!list_empty(&imp->imp_replay_list)) {
1564 struct ptlrpc_request *last;
1566 last = list_entry(imp->imp_replay_list.prev,
1567 struct ptlrpc_request,
1570 * Requests with rq_replay stay on the list even if no
1571 * commit is expected.
1573 if (last->rq_transno > imp->imp_peer_committed_transno)
1574 ptlrpc_pinger_commit_expected(imp);
1577 spin_unlock(&imp->imp_lock);
1584 * Helper function to send request \a req over the network for the first time
1585 * Also adjusts request phase.
1586 * Returns 0 on success or error code.
1588 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1590 struct obd_import *imp = req->rq_import;
1595 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1597 /* do not try to go further if there is not enough memory in enc_pool */
1598 if (req->rq_sent && req->rq_bulk)
1599 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1600 pool_is_at_full_capacity())
1603 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1604 (!req->rq_generation_set ||
1605 req->rq_import_generation == imp->imp_generation))
1608 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1610 spin_lock(&imp->imp_lock);
1612 LASSERT(req->rq_xid != 0);
1613 LASSERT(!list_empty(&req->rq_unreplied_list));
1615 if (!req->rq_generation_set)
1616 req->rq_import_generation = imp->imp_generation;
1618 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1619 spin_lock(&req->rq_lock);
1620 req->rq_waiting = 1;
1621 spin_unlock(&req->rq_lock);
1623 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1624 ptlrpc_import_state_name(req->rq_send_state),
1625 ptlrpc_import_state_name(imp->imp_state));
1626 LASSERT(list_empty(&req->rq_list));
1627 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1628 atomic_inc(&req->rq_import->imp_inflight);
1629 spin_unlock(&imp->imp_lock);
1634 spin_unlock(&imp->imp_lock);
1635 req->rq_status = rc;
1636 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1640 LASSERT(list_empty(&req->rq_list));
1641 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1642 atomic_inc(&req->rq_import->imp_inflight);
1645 * find the known replied XID from the unreplied list, CONNECT
1646 * and DISCONNECT requests are skipped to make the sanity check
1647 * on server side happy. see process_req_last_xid().
1649 * For CONNECT: Because replay requests have lower XID, it'll
1650 * break the sanity check if CONNECT bump the exp_last_xid on
1653 * For DISCONNECT: Since client will abort inflight RPC before
1654 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1655 * than the inflight RPC.
1657 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1658 min_xid = ptlrpc_known_replied_xid(imp);
1659 spin_unlock(&imp->imp_lock);
1661 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1663 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1665 rc = sptlrpc_req_refresh_ctx(req, -1);
1668 req->rq_status = rc;
1671 spin_lock(&req->rq_lock);
1672 req->rq_wait_ctx = 1;
1673 spin_unlock(&req->rq_lock);
1679 "Sending RPC pname:cluuid:pid:xid:nid:opc %s:%s:%d:%llu:%s:%d\n",
1681 imp->imp_obd->obd_uuid.uuid,
1682 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1683 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg));
1685 rc = ptl_send_rpc(req, 0);
1686 if (rc == -ENOMEM) {
1687 spin_lock(&imp->imp_lock);
1688 if (!list_empty(&req->rq_list)) {
1689 list_del_init(&req->rq_list);
1690 atomic_dec(&req->rq_import->imp_inflight);
1692 spin_unlock(&imp->imp_lock);
1693 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1697 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1698 spin_lock(&req->rq_lock);
1699 req->rq_net_err = 1;
1700 spin_unlock(&req->rq_lock);
1706 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1711 LASSERT(set->set_producer != NULL);
1713 remaining = atomic_read(&set->set_remaining);
1716 * populate the ->set_requests list with requests until we
1717 * reach the maximum number of RPCs in flight for this set
1719 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1720 rc = set->set_producer(set, set->set_producer_arg);
1721 if (rc == -ENOENT) {
1722 /* no more RPC to produce */
1723 set->set_producer = NULL;
1724 set->set_producer_arg = NULL;
1729 RETURN((atomic_read(&set->set_remaining) - remaining));
1733 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1734 * and no more replies are expected.
1735 * (it is possible to get less replies than requests sent e.g. due to timed out
1736 * requests or requests that we had trouble to send out)
1738 * NOTE: This function contains a potential schedule point (cond_resched()).
1740 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1742 struct list_head *tmp, *next;
1743 struct list_head comp_reqs;
1744 int force_timer_recalc = 0;
1747 if (atomic_read(&set->set_remaining) == 0)
1750 INIT_LIST_HEAD(&comp_reqs);
1751 list_for_each_safe(tmp, next, &set->set_requests) {
1752 struct ptlrpc_request *req =
1753 list_entry(tmp, struct ptlrpc_request,
1755 struct obd_import *imp = req->rq_import;
1756 int unregistered = 0;
1760 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1761 list_move_tail(&req->rq_set_chain, &comp_reqs);
1766 * This schedule point is mainly for the ptlrpcd caller of this
1767 * function. Most ptlrpc sets are not long-lived and unbounded
1768 * in length, but at the least the set used by the ptlrpcd is.
1769 * Since the processing time is unbounded, we need to insert an
1770 * explicit schedule point to make the thread well-behaved.
1775 * If the caller requires to allow to be interpreted by force
1776 * and it has really been interpreted, then move the request
1777 * to RQ_PHASE_INTERPRET phase in spite of what the current
1780 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1781 req->rq_status = -EINTR;
1782 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1785 * Since it is interpreted and we have to wait for
1786 * the reply to be unlinked, then use sync mode.
1790 GOTO(interpret, req->rq_status);
1793 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1794 force_timer_recalc = 1;
1796 /* delayed send - skip */
1797 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1800 /* delayed resend - skip */
1801 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1802 req->rq_sent > ktime_get_real_seconds())
1805 if (!(req->rq_phase == RQ_PHASE_RPC ||
1806 req->rq_phase == RQ_PHASE_BULK ||
1807 req->rq_phase == RQ_PHASE_INTERPRET ||
1808 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1809 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1810 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1814 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1815 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1816 LASSERT(req->rq_next_phase != req->rq_phase);
1817 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1819 if (req->rq_req_deadline &&
1820 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1821 req->rq_req_deadline = 0;
1822 if (req->rq_reply_deadline &&
1823 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1824 req->rq_reply_deadline = 0;
1825 if (req->rq_bulk_deadline &&
1826 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1827 req->rq_bulk_deadline = 0;
1830 * Skip processing until reply is unlinked. We
1831 * can't return to pool before that and we can't
1832 * call interpret before that. We need to make
1833 * sure that all rdma transfers finished and will
1834 * not corrupt any data.
1836 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1837 ptlrpc_client_recv_or_unlink(req))
1839 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1840 ptlrpc_client_bulk_active(req))
1844 * Turn fail_loc off to prevent it from looping
1847 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1848 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1851 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1852 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1857 * Move to next phase if reply was successfully
1860 ptlrpc_rqphase_move(req, req->rq_next_phase);
1863 if (req->rq_phase == RQ_PHASE_INTERPRET)
1864 GOTO(interpret, req->rq_status);
1867 * Note that this also will start async reply unlink.
1869 if (req->rq_net_err && !req->rq_timedout) {
1870 ptlrpc_expire_one_request(req, 1);
1873 * Check if we still need to wait for unlink.
1875 if (ptlrpc_client_recv_or_unlink(req) ||
1876 ptlrpc_client_bulk_active(req))
1878 /* If there is no need to resend, fail it now. */
1879 if (req->rq_no_resend) {
1880 if (req->rq_status == 0)
1881 req->rq_status = -EIO;
1882 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1883 GOTO(interpret, req->rq_status);
1890 spin_lock(&req->rq_lock);
1891 req->rq_replied = 0;
1892 spin_unlock(&req->rq_lock);
1893 if (req->rq_status == 0)
1894 req->rq_status = -EIO;
1895 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1896 GOTO(interpret, req->rq_status);
1900 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1901 * so it sets rq_intr regardless of individual rpc
1902 * timeouts. The synchronous IO waiting path sets
1903 * rq_intr irrespective of whether ptlrpcd
1904 * has seen a timeout. Our policy is to only interpret
1905 * interrupted rpcs after they have timed out, so we
1906 * need to enforce that here.
1909 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1910 req->rq_wait_ctx)) {
1911 req->rq_status = -EINTR;
1912 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1913 GOTO(interpret, req->rq_status);
1916 if (req->rq_phase == RQ_PHASE_RPC) {
1917 if (req->rq_timedout || req->rq_resend ||
1918 req->rq_waiting || req->rq_wait_ctx) {
1921 if (!ptlrpc_unregister_reply(req, 1)) {
1922 ptlrpc_unregister_bulk(req, 1);
1926 spin_lock(&imp->imp_lock);
1927 if (ptlrpc_import_delay_req(imp, req,
1930 * put on delay list - only if we wait
1931 * recovery finished - before send
1933 list_del_init(&req->rq_list);
1934 list_add_tail(&req->rq_list,
1935 &imp->imp_delayed_list);
1936 spin_unlock(&imp->imp_lock);
1941 req->rq_status = status;
1942 ptlrpc_rqphase_move(req,
1943 RQ_PHASE_INTERPRET);
1944 spin_unlock(&imp->imp_lock);
1945 GOTO(interpret, req->rq_status);
1947 /* ignore on just initiated connections */
1948 if (ptlrpc_no_resend(req) &&
1949 !req->rq_wait_ctx &&
1950 imp->imp_generation !=
1951 imp->imp_initiated_at) {
1952 req->rq_status = -ENOTCONN;
1953 ptlrpc_rqphase_move(req,
1954 RQ_PHASE_INTERPRET);
1955 spin_unlock(&imp->imp_lock);
1956 GOTO(interpret, req->rq_status);
1959 list_del_init(&req->rq_list);
1960 list_add_tail(&req->rq_list,
1961 &imp->imp_sending_list);
1963 spin_unlock(&imp->imp_lock);
1965 spin_lock(&req->rq_lock);
1966 req->rq_waiting = 0;
1967 spin_unlock(&req->rq_lock);
1969 if (req->rq_timedout || req->rq_resend) {
1971 * This is re-sending anyways,
1972 * let's mark req as resend.
1974 spin_lock(&req->rq_lock);
1976 spin_unlock(&req->rq_lock);
1979 * rq_wait_ctx is only touched by ptlrpcd,
1980 * so no lock is needed here.
1982 status = sptlrpc_req_refresh_ctx(req, -1);
1985 req->rq_status = status;
1986 spin_lock(&req->rq_lock);
1987 req->rq_wait_ctx = 0;
1988 spin_unlock(&req->rq_lock);
1989 force_timer_recalc = 1;
1991 spin_lock(&req->rq_lock);
1992 req->rq_wait_ctx = 1;
1993 spin_unlock(&req->rq_lock);
1998 spin_lock(&req->rq_lock);
1999 req->rq_wait_ctx = 0;
2000 spin_unlock(&req->rq_lock);
2004 * In any case, the previous bulk should be
2005 * cleaned up to prepare for the new sending
2008 !ptlrpc_unregister_bulk(req, 1))
2011 rc = ptl_send_rpc(req, 0);
2012 if (rc == -ENOMEM) {
2013 spin_lock(&imp->imp_lock);
2014 if (!list_empty(&req->rq_list))
2015 list_del_init(&req->rq_list);
2016 spin_unlock(&imp->imp_lock);
2017 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2021 DEBUG_REQ(D_HA, req,
2022 "send failed: rc = %d", rc);
2023 force_timer_recalc = 1;
2024 spin_lock(&req->rq_lock);
2025 req->rq_net_err = 1;
2026 spin_unlock(&req->rq_lock);
2029 /* need to reset the timeout */
2030 force_timer_recalc = 1;
2033 spin_lock(&req->rq_lock);
2035 if (ptlrpc_client_early(req)) {
2036 ptlrpc_at_recv_early_reply(req);
2037 spin_unlock(&req->rq_lock);
2041 /* Still waiting for a reply? */
2042 if (ptlrpc_client_recv(req)) {
2043 spin_unlock(&req->rq_lock);
2047 /* Did we actually receive a reply? */
2048 if (!ptlrpc_client_replied(req)) {
2049 spin_unlock(&req->rq_lock);
2053 spin_unlock(&req->rq_lock);
2056 * unlink from net because we are going to
2057 * swab in-place of reply buffer
2059 unregistered = ptlrpc_unregister_reply(req, 1);
2063 req->rq_status = after_reply(req);
2068 * If there is no bulk associated with this request,
2069 * then we're done and should let the interpreter
2070 * process the reply. Similarly if the RPC returned
2071 * an error, and therefore the bulk will never arrive.
2073 if (!req->rq_bulk || req->rq_status < 0) {
2074 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2075 GOTO(interpret, req->rq_status);
2078 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2081 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2082 if (ptlrpc_client_bulk_active(req))
2085 if (req->rq_bulk->bd_failure) {
2087 * The RPC reply arrived OK, but the bulk screwed
2088 * up! Dead weird since the server told us the RPC
2089 * was good after getting the REPLY for her GET or
2090 * the ACK for her PUT.
2092 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2093 req->rq_status = -EIO;
2096 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2099 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2102 * This moves to "unregistering" phase we need to wait for
2105 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2106 /* start async bulk unlink too */
2107 ptlrpc_unregister_bulk(req, 1);
2111 if (!ptlrpc_unregister_bulk(req, async))
2115 * When calling interpret receiving already should be
2118 LASSERT(!req->rq_receiving_reply);
2120 ptlrpc_req_interpret(env, req, req->rq_status);
2122 if (ptlrpcd_check_work(req)) {
2123 atomic_dec(&set->set_remaining);
2126 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2130 "Completed RPC pname:cluuid:pid:xid:nid:opc %s:%s:%d:%llu:%s:%d\n",
2132 imp->imp_obd->obd_uuid.uuid,
2133 lustre_msg_get_status(req->rq_reqmsg),
2135 obd_import_nid2str(imp),
2136 lustre_msg_get_opc(req->rq_reqmsg));
2138 spin_lock(&imp->imp_lock);
2140 * Request already may be not on sending or delaying list. This
2141 * may happen in the case of marking it erroneous for the case
2142 * ptlrpc_import_delay_req(req, status) find it impossible to
2143 * allow sending this rpc and returns *status != 0.
2145 if (!list_empty(&req->rq_list)) {
2146 list_del_init(&req->rq_list);
2147 atomic_dec(&imp->imp_inflight);
2149 list_del_init(&req->rq_unreplied_list);
2150 spin_unlock(&imp->imp_lock);
2152 atomic_dec(&set->set_remaining);
2153 wake_up_all(&imp->imp_recovery_waitq);
2155 if (set->set_producer) {
2156 /* produce a new request if possible */
2157 if (ptlrpc_set_producer(set) > 0)
2158 force_timer_recalc = 1;
2161 * free the request that has just been completed
2162 * in order not to pollute set->set_requests
2164 list_del_init(&req->rq_set_chain);
2165 spin_lock(&req->rq_lock);
2167 req->rq_invalid_rqset = 0;
2168 spin_unlock(&req->rq_lock);
2170 /* record rq_status to compute the final status later */
2171 if (req->rq_status != 0)
2172 set->set_rc = req->rq_status;
2173 ptlrpc_req_finished(req);
2175 list_move_tail(&req->rq_set_chain, &comp_reqs);
2180 * move completed request at the head of list so it's easier for
2181 * caller to find them
2183 list_splice(&comp_reqs, &set->set_requests);
2185 /* If we hit an error, we want to recover promptly. */
2186 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2188 EXPORT_SYMBOL(ptlrpc_check_set);
2191 * Time out request \a req. is \a async_unlink is set, that means do not wait
2192 * until LNet actually confirms network buffer unlinking.
2193 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2195 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2197 struct obd_import *imp = req->rq_import;
2198 unsigned int debug_mask = D_RPCTRACE;
2202 spin_lock(&req->rq_lock);
2203 req->rq_timedout = 1;
2204 spin_unlock(&req->rq_lock);
2206 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2207 lustre_msg_get_status(req->rq_reqmsg)))
2208 debug_mask = D_WARNING;
2209 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2210 req->rq_net_err ? "failed due to network error" :
2211 ((req->rq_real_sent == 0 ||
2212 req->rq_real_sent < req->rq_sent ||
2213 req->rq_real_sent >= req->rq_deadline) ?
2214 "timed out for sent delay" : "timed out for slow reply"),
2215 (s64)req->rq_sent, (s64)req->rq_real_sent);
2217 if (imp && obd_debug_peer_on_timeout)
2218 LNetDebugPeer(imp->imp_connection->c_peer);
2220 ptlrpc_unregister_reply(req, async_unlink);
2221 ptlrpc_unregister_bulk(req, async_unlink);
2223 if (obd_dump_on_timeout)
2224 libcfs_debug_dumplog();
2227 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2231 atomic_inc(&imp->imp_timeouts);
2233 /* The DLM server doesn't want recovery run on its imports. */
2234 if (imp->imp_dlm_fake)
2238 * If this request is for recovery or other primordial tasks,
2239 * then error it out here.
2241 if (req->rq_ctx_init || req->rq_ctx_fini ||
2242 req->rq_send_state != LUSTRE_IMP_FULL ||
2243 imp->imp_obd->obd_no_recov) {
2244 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2245 ptlrpc_import_state_name(req->rq_send_state),
2246 ptlrpc_import_state_name(imp->imp_state));
2247 spin_lock(&req->rq_lock);
2248 req->rq_status = -ETIMEDOUT;
2250 spin_unlock(&req->rq_lock);
2255 * if a request can't be resent we can't wait for an answer after
2258 if (ptlrpc_no_resend(req)) {
2259 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2263 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2269 * Time out all uncompleted requests in request set pointed by \a data
2270 * Callback used when waiting on sets with l_wait_event.
2273 int ptlrpc_expired_set(void *data)
2275 struct ptlrpc_request_set *set = data;
2276 struct list_head *tmp;
2277 time64_t now = ktime_get_real_seconds();
2280 LASSERT(set != NULL);
2283 * A timeout expired. See which reqs it applies to...
2285 list_for_each(tmp, &set->set_requests) {
2286 struct ptlrpc_request *req =
2287 list_entry(tmp, struct ptlrpc_request,
2290 /* don't expire request waiting for context */
2291 if (req->rq_wait_ctx)
2294 /* Request in-flight? */
2295 if (!((req->rq_phase == RQ_PHASE_RPC &&
2296 !req->rq_waiting && !req->rq_resend) ||
2297 (req->rq_phase == RQ_PHASE_BULK)))
2300 if (req->rq_timedout || /* already dealt with */
2301 req->rq_deadline > now) /* not expired */
2305 * Deal with this guy. Do it asynchronously to not block
2308 ptlrpc_expire_one_request(req, 1);
2312 * When waiting for a whole set, we always break out of the
2313 * sleep so we can recalculate the timeout, or enable interrupts
2314 * if everyone's timed out.
2320 * Sets rq_intr flag in \a req under spinlock.
2322 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2324 spin_lock(&req->rq_lock);
2326 spin_unlock(&req->rq_lock);
2328 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2331 * Interrupts (sets interrupted flag) all uncompleted requests in
2332 * a set \a data. Callback for l_wait_event for interruptible waits.
2334 static void ptlrpc_interrupted_set(void *data)
2336 struct ptlrpc_request_set *set = data;
2337 struct list_head *tmp;
2339 LASSERT(set != NULL);
2340 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2342 list_for_each(tmp, &set->set_requests) {
2343 struct ptlrpc_request *req =
2344 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2349 if (req->rq_phase != RQ_PHASE_RPC &&
2350 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2351 !req->rq_allow_intr)
2354 ptlrpc_mark_interrupted(req);
2359 * Get the smallest timeout in the set; this does NOT set a timeout.
2361 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2363 struct list_head *tmp;
2364 time64_t now = ktime_get_real_seconds();
2366 struct ptlrpc_request *req;
2370 list_for_each(tmp, &set->set_requests) {
2371 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2373 /* Request in-flight? */
2374 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2375 (req->rq_phase == RQ_PHASE_BULK) ||
2376 (req->rq_phase == RQ_PHASE_NEW)))
2379 /* Already timed out. */
2380 if (req->rq_timedout)
2383 /* Waiting for ctx. */
2384 if (req->rq_wait_ctx)
2387 if (req->rq_phase == RQ_PHASE_NEW)
2388 deadline = req->rq_sent;
2389 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2390 deadline = req->rq_sent;
2392 deadline = req->rq_sent + req->rq_timeout;
2394 if (deadline <= now) /* actually expired already */
2395 timeout = 1; /* ASAP */
2396 else if (timeout == 0 || timeout > deadline - now)
2397 timeout = deadline - now;
2403 * Send all unset request from the set and then wait untill all
2404 * requests in the set complete (either get a reply, timeout, get an
2405 * error or otherwise be interrupted).
2406 * Returns 0 on success or error code otherwise.
2408 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2410 struct list_head *tmp;
2411 struct ptlrpc_request *req;
2412 struct l_wait_info lwi;
2417 if (set->set_producer)
2418 (void)ptlrpc_set_producer(set);
2420 list_for_each(tmp, &set->set_requests) {
2421 req = list_entry(tmp, struct ptlrpc_request,
2423 if (req->rq_phase == RQ_PHASE_NEW)
2424 (void)ptlrpc_send_new_req(req);
2427 if (list_empty(&set->set_requests))
2431 timeout = ptlrpc_set_next_timeout(set);
2434 * wait until all complete, interrupted, or an in-flight
2437 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2440 if ((timeout == 0 && !signal_pending(current)) ||
2441 set->set_allow_intr)
2443 * No requests are in-flight (ether timed out
2444 * or delayed), so we can allow interrupts.
2445 * We still want to block for a limited time,
2446 * so we allow interrupts during the timeout.
2448 lwi = LWI_TIMEOUT_INTR_ALL(
2449 cfs_time_seconds(timeout ? timeout : 1),
2451 ptlrpc_interrupted_set, set);
2454 * At least one request is in flight, so no
2455 * interrupts are allowed. Wait until all
2456 * complete, or an in-flight req times out.
2458 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2459 ptlrpc_expired_set, set);
2461 rc = l_wait_event(set->set_waitq,
2462 ptlrpc_check_set(NULL, set), &lwi);
2465 * LU-769 - if we ignored the signal because it was already
2466 * pending when we started, we need to handle it now or we risk
2467 * it being ignored forever
2469 if (rc == -ETIMEDOUT &&
2470 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2471 signal_pending(current)) {
2472 sigset_t blocked_sigs =
2473 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2476 * In fact we only interrupt for the "fatal" signals
2477 * like SIGINT or SIGKILL. We still ignore less
2478 * important signals since ptlrpc set is not easily
2479 * reentrant from userspace again
2481 if (signal_pending(current))
2482 ptlrpc_interrupted_set(set);
2483 cfs_restore_sigs(blocked_sigs);
2486 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2489 * -EINTR => all requests have been flagged rq_intr so next
2491 * -ETIMEDOUT => someone timed out. When all reqs have
2492 * timed out, signals are enabled allowing completion with
2494 * I don't really care if we go once more round the loop in
2495 * the error cases -eeb.
2497 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2498 list_for_each(tmp, &set->set_requests) {
2499 req = list_entry(tmp, struct ptlrpc_request,
2501 spin_lock(&req->rq_lock);
2502 req->rq_invalid_rqset = 1;
2503 spin_unlock(&req->rq_lock);
2506 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2508 LASSERT(atomic_read(&set->set_remaining) == 0);
2510 rc = set->set_rc; /* rq_status of already freed requests if any */
2511 list_for_each(tmp, &set->set_requests) {
2512 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2514 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2515 if (req->rq_status != 0)
2516 rc = req->rq_status;
2521 EXPORT_SYMBOL(ptlrpc_set_wait);
2524 * Helper fuction for request freeing.
2525 * Called when request count reached zero and request needs to be freed.
2526 * Removes request from all sorts of sending/replay lists it might be on,
2527 * frees network buffers if any are present.
2528 * If \a locked is set, that means caller is already holding import imp_lock
2529 * and so we no longer need to reobtain it (for certain lists manipulations)
2531 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2538 LASSERT(!request->rq_srv_req);
2539 LASSERT(request->rq_export == NULL);
2540 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2541 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2542 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2543 LASSERTF(!request->rq_replay, "req %p\n", request);
2545 req_capsule_fini(&request->rq_pill);
2548 * We must take it off the imp_replay_list first. Otherwise, we'll set
2549 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2551 if (request->rq_import) {
2553 spin_lock(&request->rq_import->imp_lock);
2554 list_del_init(&request->rq_replay_list);
2555 list_del_init(&request->rq_unreplied_list);
2557 spin_unlock(&request->rq_import->imp_lock);
2559 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2561 if (atomic_read(&request->rq_refcount) != 0) {
2562 DEBUG_REQ(D_ERROR, request,
2563 "freeing request with nonzero refcount");
2567 if (request->rq_repbuf)
2568 sptlrpc_cli_free_repbuf(request);
2570 if (request->rq_import) {
2571 class_import_put(request->rq_import);
2572 request->rq_import = NULL;
2574 if (request->rq_bulk)
2575 ptlrpc_free_bulk(request->rq_bulk);
2577 if (request->rq_reqbuf || request->rq_clrbuf)
2578 sptlrpc_cli_free_reqbuf(request);
2580 if (request->rq_cli_ctx)
2581 sptlrpc_req_put_ctx(request, !locked);
2583 if (request->rq_pool)
2584 __ptlrpc_free_req_to_pool(request);
2586 ptlrpc_request_cache_free(request);
2590 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2592 * Drop one request reference. Must be called with import imp_lock held.
2593 * When reference count drops to zero, request is freed.
2595 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2597 assert_spin_locked(&request->rq_import->imp_lock);
2598 (void)__ptlrpc_req_finished(request, 1);
2603 * Drops one reference count for request \a request.
2604 * \a locked set indicates that caller holds import imp_lock.
2605 * Frees the request whe reference count reaches zero.
2607 * \retval 1 the request is freed
2608 * \retval 0 some others still hold references on the request
2610 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2618 LASSERT(request != LP_POISON);
2619 LASSERT(request->rq_reqmsg != LP_POISON);
2621 DEBUG_REQ(D_INFO, request, "refcount now %u",
2622 atomic_read(&request->rq_refcount) - 1);
2624 spin_lock(&request->rq_lock);
2625 count = atomic_dec_return(&request->rq_refcount);
2626 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2629 * For open RPC, the client does not know the EA size (LOV, ACL, and
2630 * so on) before replied, then the client has to reserve very large
2631 * reply buffer. Such buffer will not be released until the RPC freed.
2632 * Since The open RPC is replayable, we need to keep it in the replay
2633 * list until close. If there are a lot of files opened concurrently,
2634 * then the client may be OOM.
2636 * If fact, it is unnecessary to keep reply buffer for open replay,
2637 * related EAs have already been saved via mdc_save_lovea() before
2638 * coming here. So it is safe to free the reply buffer some earlier
2639 * before releasing the RPC to avoid client OOM. LU-9514
2641 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2642 spin_lock(&request->rq_early_free_lock);
2643 sptlrpc_cli_free_repbuf(request);
2644 request->rq_repbuf = NULL;
2645 request->rq_repbuf_len = 0;
2646 request->rq_repdata = NULL;
2647 request->rq_reqdata_len = 0;
2648 spin_unlock(&request->rq_early_free_lock);
2650 spin_unlock(&request->rq_lock);
2653 __ptlrpc_free_req(request, locked);
2659 * Drops one reference count for a request.
2661 void ptlrpc_req_finished(struct ptlrpc_request *request)
2663 __ptlrpc_req_finished(request, 0);
2665 EXPORT_SYMBOL(ptlrpc_req_finished);
2668 * Returns xid of a \a request
2670 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2672 return request->rq_xid;
2674 EXPORT_SYMBOL(ptlrpc_req_xid);
2677 * Disengage the client's reply buffer from the network
2678 * NB does _NOT_ unregister any client-side bulk.
2679 * IDEMPOTENT, but _not_ safe against concurrent callers.
2680 * The request owner (i.e. the thread doing the I/O) must call...
2681 * Returns 0 on success or 1 if unregistering cannot be made.
2683 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2686 struct l_wait_info lwi;
2691 LASSERT(!in_interrupt());
2693 /* Let's setup deadline for reply unlink. */
2694 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2695 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2696 request->rq_reply_deadline = ktime_get_real_seconds() +
2700 * Nothing left to do.
2702 if (!ptlrpc_client_recv_or_unlink(request))
2705 LNetMDUnlink(request->rq_reply_md_h);
2708 * Let's check it once again.
2710 if (!ptlrpc_client_recv_or_unlink(request))
2713 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2714 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2717 * Do not wait for unlink to finish.
2723 * We have to l_wait_event() whatever the result, to give liblustre
2724 * a chance to run reply_in_callback(), and to make sure we've
2725 * unlinked before returning a req to the pool.
2728 /* The wq argument is ignored by user-space wait_event macros */
2729 wait_queue_head_t *wq = (request->rq_set) ?
2730 &request->rq_set->set_waitq :
2731 &request->rq_reply_waitq;
2733 * Network access will complete in finite time but the HUGE
2734 * timeout lets us CWARN for visibility of sluggish NALs
2736 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2737 cfs_time_seconds(1), NULL, NULL);
2738 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2741 ptlrpc_rqphase_move(request, request->rq_next_phase);
2745 LASSERT(rc == -ETIMEDOUT);
2746 DEBUG_REQ(D_WARNING, request,
2747 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2748 request->rq_receiving_reply,
2749 request->rq_req_unlinked,
2750 request->rq_reply_unlinked);
2755 static void ptlrpc_free_request(struct ptlrpc_request *req)
2757 spin_lock(&req->rq_lock);
2759 spin_unlock(&req->rq_lock);
2761 if (req->rq_commit_cb)
2762 req->rq_commit_cb(req);
2763 list_del_init(&req->rq_replay_list);
2765 __ptlrpc_req_finished(req, 1);
2769 * the request is committed and dropped from the replay list of its import
2771 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2773 struct obd_import *imp = req->rq_import;
2775 spin_lock(&imp->imp_lock);
2776 if (list_empty(&req->rq_replay_list)) {
2777 spin_unlock(&imp->imp_lock);
2781 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2782 if (imp->imp_replay_cursor == &req->rq_replay_list)
2783 imp->imp_replay_cursor = req->rq_replay_list.next;
2784 ptlrpc_free_request(req);
2787 spin_unlock(&imp->imp_lock);
2789 EXPORT_SYMBOL(ptlrpc_request_committed);
2792 * Iterates through replay_list on import and prunes
2793 * all requests have transno smaller than last_committed for the
2794 * import and don't have rq_replay set.
2795 * Since requests are sorted in transno order, stops when meetign first
2796 * transno bigger than last_committed.
2797 * caller must hold imp->imp_lock
2799 void ptlrpc_free_committed(struct obd_import *imp)
2801 struct ptlrpc_request *req, *saved;
2802 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2803 bool skip_committed_list = true;
2806 LASSERT(imp != NULL);
2807 assert_spin_locked(&imp->imp_lock);
2809 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2810 imp->imp_generation == imp->imp_last_generation_checked) {
2811 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2812 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2815 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2816 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2817 imp->imp_generation);
2819 if (imp->imp_generation != imp->imp_last_generation_checked ||
2820 imp->imp_last_transno_checked == 0)
2821 skip_committed_list = false;
2823 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2824 imp->imp_last_generation_checked = imp->imp_generation;
2826 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2828 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2829 LASSERT(req != last_req);
2832 if (req->rq_transno == 0) {
2833 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2836 if (req->rq_import_generation < imp->imp_generation) {
2837 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2841 /* not yet committed */
2842 if (req->rq_transno > imp->imp_peer_committed_transno) {
2843 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2847 if (req->rq_replay) {
2848 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2849 list_move_tail(&req->rq_replay_list,
2850 &imp->imp_committed_list);
2854 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2855 imp->imp_peer_committed_transno);
2857 ptlrpc_free_request(req);
2860 if (skip_committed_list)
2863 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2865 LASSERT(req->rq_transno != 0);
2866 if (req->rq_import_generation < imp->imp_generation ||
2868 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2869 req->rq_import_generation <
2870 imp->imp_generation ? "stale" : "closed");
2872 if (imp->imp_replay_cursor == &req->rq_replay_list)
2873 imp->imp_replay_cursor =
2874 req->rq_replay_list.next;
2876 ptlrpc_free_request(req);
2883 void ptlrpc_cleanup_client(struct obd_import *imp)
2890 * Schedule previously sent request for resend.
2891 * For bulk requests we assign new xid (to avoid problems with
2892 * lost replies and therefore several transfers landing into same buffer
2893 * from different sending attempts).
2895 void ptlrpc_resend_req(struct ptlrpc_request *req)
2897 DEBUG_REQ(D_HA, req, "going to resend");
2898 spin_lock(&req->rq_lock);
2901 * Request got reply but linked to the import list still.
2902 * Let ptlrpc_check_set() process it.
2904 if (ptlrpc_client_replied(req)) {
2905 spin_unlock(&req->rq_lock);
2906 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2910 req->rq_status = -EAGAIN;
2913 req->rq_net_err = 0;
2914 req->rq_timedout = 0;
2916 ptlrpc_client_wake_req(req);
2917 spin_unlock(&req->rq_lock);
2920 /* XXX: this function and rq_status are currently unused */
2921 void ptlrpc_restart_req(struct ptlrpc_request *req)
2923 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2924 req->rq_status = -ERESTARTSYS;
2926 spin_lock(&req->rq_lock);
2927 req->rq_restart = 1;
2928 req->rq_timedout = 0;
2929 ptlrpc_client_wake_req(req);
2930 spin_unlock(&req->rq_lock);
2934 * Grab additional reference on a request \a req
2936 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2939 atomic_inc(&req->rq_refcount);
2942 EXPORT_SYMBOL(ptlrpc_request_addref);
2945 * Add a request to import replay_list.
2946 * Must be called under imp_lock
2948 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2949 struct obd_import *imp)
2951 struct list_head *tmp;
2953 assert_spin_locked(&imp->imp_lock);
2955 if (req->rq_transno == 0) {
2956 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2961 * clear this for new requests that were resent as well
2962 * as resent replayed requests.
2964 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2966 /* don't re-add requests that have been replayed */
2967 if (!list_empty(&req->rq_replay_list))
2970 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2972 spin_lock(&req->rq_lock);
2974 spin_unlock(&req->rq_lock);
2976 LASSERT(imp->imp_replayable);
2977 /* Balanced in ptlrpc_free_committed, usually. */
2978 ptlrpc_request_addref(req);
2979 list_for_each_prev(tmp, &imp->imp_replay_list) {
2980 struct ptlrpc_request *iter = list_entry(tmp,
2981 struct ptlrpc_request,
2985 * We may have duplicate transnos if we create and then
2986 * open a file, or for closes retained if to match creating
2987 * opens, so use req->rq_xid as a secondary key.
2988 * (See bugs 684, 685, and 428.)
2989 * XXX no longer needed, but all opens need transnos!
2991 if (iter->rq_transno > req->rq_transno)
2994 if (iter->rq_transno == req->rq_transno) {
2995 LASSERT(iter->rq_xid != req->rq_xid);
2996 if (iter->rq_xid > req->rq_xid)
3000 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3004 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3008 * Send request and wait until it completes.
3009 * Returns request processing status.
3011 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3013 struct ptlrpc_request_set *set;
3017 LASSERT(req->rq_set == NULL);
3018 LASSERT(!req->rq_receiving_reply);
3020 set = ptlrpc_prep_set();
3022 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3026 /* for distributed debugging */
3027 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3029 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3030 ptlrpc_request_addref(req);
3031 ptlrpc_set_add_req(set, req);
3032 rc = ptlrpc_set_wait(NULL, set);
3033 ptlrpc_set_destroy(set);
3037 EXPORT_SYMBOL(ptlrpc_queue_wait);
3040 * Callback used for replayed requests reply processing.
3041 * In case of successful reply calls registered request replay callback.
3042 * In case of error restart replay process.
3044 static int ptlrpc_replay_interpret(const struct lu_env *env,
3045 struct ptlrpc_request *req,
3048 struct ptlrpc_replay_async_args *aa = args;
3049 struct obd_import *imp = req->rq_import;
3052 atomic_dec(&imp->imp_replay_inflight);
3055 * Note: if it is bulk replay (MDS-MDS replay), then even if
3056 * server got the request, but bulk transfer timeout, let's
3057 * replay the bulk req again
3059 if (!ptlrpc_client_replied(req) ||
3061 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3062 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
3063 GOTO(out, rc = -ETIMEDOUT);
3066 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3067 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3068 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3069 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3071 /** VBR: check version failure */
3072 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3073 /** replay was failed due to version mismatch */
3074 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
3075 spin_lock(&imp->imp_lock);
3076 imp->imp_vbr_failed = 1;
3077 spin_unlock(&imp->imp_lock);
3078 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3080 /** The transno had better not change over replay. */
3081 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3082 lustre_msg_get_transno(req->rq_repmsg) ||
3083 lustre_msg_get_transno(req->rq_repmsg) == 0,
3085 lustre_msg_get_transno(req->rq_reqmsg),
3086 lustre_msg_get_transno(req->rq_repmsg));
3089 spin_lock(&imp->imp_lock);
3090 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3091 spin_unlock(&imp->imp_lock);
3092 LASSERT(imp->imp_last_replay_transno);
3094 /* transaction number shouldn't be bigger than the latest replayed */
3095 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3096 DEBUG_REQ(D_ERROR, req,
3097 "Reported transno %llu is bigger than the replayed one: %llu",
3099 lustre_msg_get_transno(req->rq_reqmsg));
3100 GOTO(out, rc = -EINVAL);
3103 DEBUG_REQ(D_HA, req, "got rep");
3105 /* let the callback do fixups, possibly including in the request */
3106 if (req->rq_replay_cb)
3107 req->rq_replay_cb(req);
3109 if (ptlrpc_client_replied(req) &&
3110 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3111 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3112 lustre_msg_get_status(req->rq_repmsg),
3113 aa->praa_old_status);
3116 * Note: If the replay fails for MDT-MDT recovery, let's
3117 * abort all of the following requests in the replay
3118 * and sending list, because MDT-MDT update requests
3119 * are dependent on each other, see LU-7039
3121 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3122 struct ptlrpc_request *free_req;
3123 struct ptlrpc_request *tmp;
3125 spin_lock(&imp->imp_lock);
3126 list_for_each_entry_safe(free_req, tmp,
3127 &imp->imp_replay_list,
3129 ptlrpc_free_request(free_req);
3132 list_for_each_entry_safe(free_req, tmp,
3133 &imp->imp_committed_list,
3135 ptlrpc_free_request(free_req);
3138 list_for_each_entry_safe(free_req, tmp,
3139 &imp->imp_delayed_list,
3141 spin_lock(&free_req->rq_lock);
3142 free_req->rq_err = 1;
3143 free_req->rq_status = -EIO;
3144 ptlrpc_client_wake_req(free_req);
3145 spin_unlock(&free_req->rq_lock);
3148 list_for_each_entry_safe(free_req, tmp,
3149 &imp->imp_sending_list,
3151 spin_lock(&free_req->rq_lock);
3152 free_req->rq_err = 1;
3153 free_req->rq_status = -EIO;
3154 ptlrpc_client_wake_req(free_req);
3155 spin_unlock(&free_req->rq_lock);
3157 spin_unlock(&imp->imp_lock);
3160 /* Put it back for re-replay. */
3161 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3165 * Errors while replay can set transno to 0, but
3166 * imp_last_replay_transno shouldn't be set to 0 anyway
3168 if (req->rq_transno == 0)
3169 CERROR("Transno is 0 during replay!\n");
3171 /* continue with recovery */
3172 rc = ptlrpc_import_recovery_state_machine(imp);
3174 req->rq_send_state = aa->praa_old_state;
3177 /* this replay failed, so restart recovery */
3178 ptlrpc_connect_import(imp);
3184 * Prepares and queues request for replay.
3185 * Adds it to ptlrpcd queue for actual sending.
3186 * Returns 0 on success.
3188 int ptlrpc_replay_req(struct ptlrpc_request *req)
3190 struct ptlrpc_replay_async_args *aa;
3194 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3196 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3197 aa = ptlrpc_req_async_args(req);
3198 memset(aa, 0, sizeof(*aa));
3200 /* Prepare request to be resent with ptlrpcd */
3201 aa->praa_old_state = req->rq_send_state;
3202 req->rq_send_state = LUSTRE_IMP_REPLAY;
3203 req->rq_phase = RQ_PHASE_NEW;
3204 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3206 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3208 req->rq_interpret_reply = ptlrpc_replay_interpret;
3209 /* Readjust the timeout for current conditions */
3210 ptlrpc_at_set_req_timeout(req);
3212 /* Tell server net_latency to calculate how long to wait for reply. */
3213 lustre_msg_set_service_time(req->rq_reqmsg,
3214 ptlrpc_at_get_net_latency(req));
3215 DEBUG_REQ(D_HA, req, "REPLAY");
3217 atomic_inc(&req->rq_import->imp_replay_inflight);
3218 spin_lock(&req->rq_lock);
3219 req->rq_early_free_repbuf = 0;
3220 spin_unlock(&req->rq_lock);
3221 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3223 ptlrpcd_add_req(req);
3228 * Aborts all in-flight request on import \a imp sending and delayed lists
3230 void ptlrpc_abort_inflight(struct obd_import *imp)
3232 struct list_head *tmp, *n;
3236 * Make sure that no new requests get processed for this import.
3237 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3238 * this flag and then putting requests on sending_list or delayed_list.
3240 spin_lock(&imp->imp_lock);
3243 * XXX locking? Maybe we should remove each request with the list
3244 * locked? Also, how do we know if the requests on the list are
3245 * being freed at this time?
3247 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3248 struct ptlrpc_request *req = list_entry(tmp,
3249 struct ptlrpc_request,
3252 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3254 spin_lock(&req->rq_lock);
3255 if (req->rq_import_generation < imp->imp_generation) {
3257 req->rq_status = -EIO;
3258 ptlrpc_client_wake_req(req);
3260 spin_unlock(&req->rq_lock);
3263 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3264 struct ptlrpc_request *req =
3265 list_entry(tmp, struct ptlrpc_request, rq_list);
3267 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3269 spin_lock(&req->rq_lock);
3270 if (req->rq_import_generation < imp->imp_generation) {
3272 req->rq_status = -EIO;
3273 ptlrpc_client_wake_req(req);
3275 spin_unlock(&req->rq_lock);
3279 * Last chance to free reqs left on the replay list, but we
3280 * will still leak reqs that haven't committed.
3282 if (imp->imp_replayable)
3283 ptlrpc_free_committed(imp);
3285 spin_unlock(&imp->imp_lock);
3291 * Abort all uncompleted requests in request set \a set
3293 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3295 struct list_head *tmp, *pos;
3297 LASSERT(set != NULL);
3299 list_for_each_safe(pos, tmp, &set->set_requests) {
3300 struct ptlrpc_request *req =
3301 list_entry(pos, struct ptlrpc_request,
3304 spin_lock(&req->rq_lock);
3305 if (req->rq_phase != RQ_PHASE_RPC) {
3306 spin_unlock(&req->rq_lock);
3311 req->rq_status = -EINTR;
3312 ptlrpc_client_wake_req(req);
3313 spin_unlock(&req->rq_lock);
3318 * Initialize the XID for the node. This is common among all requests on
3319 * this node, and only requires the property that it is monotonically
3320 * increasing. It does not need to be sequential. Since this is also used
3321 * as the RDMA match bits, it is important that a single client NOT have
3322 * the same match bits for two different in-flight requests, hence we do
3323 * NOT want to have an XID per target or similar.
3325 * To avoid an unlikely collision between match bits after a client reboot
3326 * (which would deliver old data into the wrong RDMA buffer) initialize
3327 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3328 * If the time is clearly incorrect, we instead use a 62-bit random number.
3329 * In the worst case the random number will overflow 1M RPCs per second in
3330 * 9133 years, or permutations thereof.
3332 #define YEAR_2004 (1ULL << 30)
3333 void ptlrpc_init_xid(void)
3335 time64_t now = ktime_get_real_seconds();
3337 spin_lock_init(&ptlrpc_last_xid_lock);
3338 if (now < YEAR_2004) {
3339 get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3340 ptlrpc_last_xid >>= 2;
3341 ptlrpc_last_xid |= (1ULL << 61);
3343 ptlrpc_last_xid = (__u64)now << 20;
3346 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3347 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3348 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3352 * Increase xid and returns resulting new value to the caller.
3354 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3355 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3356 * itself uses the last bulk xid needed, so the server can determine the
3357 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3358 * xid must align to a power-of-two value.
3360 * This is assumed to be true due to the initial ptlrpc_last_xid
3361 * value also being initialized to a power-of-two value. LU-1431
3363 __u64 ptlrpc_next_xid(void)
3367 spin_lock(&ptlrpc_last_xid_lock);
3368 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3369 ptlrpc_last_xid = next;
3370 spin_unlock(&ptlrpc_last_xid_lock);
3376 * If request has a new allocated XID (new request or EINPROGRESS resend),
3377 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3378 * request to ensure previous bulk fails and avoid problems with lost replies
3379 * and therefore several transfers landing into the same buffer from different
3382 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3384 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3386 LASSERT(bd != NULL);
3389 * Generate new matchbits for all resend requests, including
3392 if (req->rq_resend) {
3393 __u64 old_mbits = req->rq_mbits;
3396 * First time resend on -EINPROGRESS will generate new xid,
3397 * so we can actually use the rq_xid as rq_mbits in such case,
3398 * however, it's bit hard to distinguish such resend with a
3399 * 'resend for the -EINPROGRESS resend'. To make it simple,
3400 * we opt to generate mbits for all resend cases.
3402 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3404 req->rq_mbits = ptlrpc_next_xid();
3407 * Old version transfers rq_xid to peer as
3410 spin_lock(&req->rq_import->imp_lock);
3411 list_del_init(&req->rq_unreplied_list);
3412 ptlrpc_assign_next_xid_nolock(req);
3413 spin_unlock(&req->rq_import->imp_lock);
3414 req->rq_mbits = req->rq_xid;
3416 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3417 old_mbits, req->rq_mbits);
3418 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3419 /* Request being sent first time, use xid as matchbits. */
3420 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3421 || req->rq_mbits == 0) {
3422 req->rq_mbits = req->rq_xid;
3424 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3426 req->rq_mbits -= total_md - 1;
3430 * Replay request, xid and matchbits have already been
3431 * correctly assigned.
3437 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3438 * that server can infer the number of bulks that were prepared,
3441 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3445 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3446 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3448 * It's ok to directly set the rq_xid here, since this xid bump
3449 * won't affect the request position in unreplied list.
3451 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3452 req->rq_xid = req->rq_mbits;
3456 * Get a glimpse at what next xid value might have been.
3457 * Returns possible next xid.
3459 __u64 ptlrpc_sample_next_xid(void)
3461 #if BITS_PER_LONG == 32
3462 /* need to avoid possible word tearing on 32-bit systems */
3465 spin_lock(&ptlrpc_last_xid_lock);
3466 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3467 spin_unlock(&ptlrpc_last_xid_lock);
3471 /* No need to lock, since returned value is racy anyways */
3472 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3475 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3478 * Functions for operating ptlrpc workers.
3480 * A ptlrpc work is a function which will be running inside ptlrpc context.
3481 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3483 * 1. after a work is created, it can be used many times, that is:
3484 * handler = ptlrpcd_alloc_work();
3485 * ptlrpcd_queue_work();
3487 * queue it again when necessary:
3488 * ptlrpcd_queue_work();
3489 * ptlrpcd_destroy_work();
3490 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3491 * it will only be queued once in any time. Also as its name implies, it may
3492 * have delay before it really runs by ptlrpcd thread.
3494 struct ptlrpc_work_async_args {
3495 int (*cb)(const struct lu_env *, void *);
3499 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3501 /* re-initialize the req */
3502 req->rq_timeout = obd_timeout;
3503 req->rq_sent = ktime_get_real_seconds();
3504 req->rq_deadline = req->rq_sent + req->rq_timeout;
3505 req->rq_phase = RQ_PHASE_INTERPRET;
3506 req->rq_next_phase = RQ_PHASE_COMPLETE;
3507 req->rq_xid = ptlrpc_next_xid();
3508 req->rq_import_generation = req->rq_import->imp_generation;
3510 ptlrpcd_add_req(req);
3513 static int work_interpreter(const struct lu_env *env,
3514 struct ptlrpc_request *req, void *args, int rc)
3516 struct ptlrpc_work_async_args *arg = args;
3518 LASSERT(ptlrpcd_check_work(req));
3519 LASSERT(arg->cb != NULL);
3521 rc = arg->cb(env, arg->cbdata);
3523 list_del_init(&req->rq_set_chain);
3526 if (atomic_dec_return(&req->rq_refcount) > 1) {
3527 atomic_set(&req->rq_refcount, 2);
3528 ptlrpcd_add_work_req(req);
3533 static int worker_format;
3535 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3537 return req->rq_pill.rc_fmt == (void *)&worker_format;
3541 * Create a work for ptlrpc.
3543 void *ptlrpcd_alloc_work(struct obd_import *imp,
3544 int (*cb)(const struct lu_env *, void *), void *cbdata)
3546 struct ptlrpc_request *req = NULL;
3547 struct ptlrpc_work_async_args *args;
3553 RETURN(ERR_PTR(-EINVAL));
3555 /* copy some code from deprecated fakereq. */
3556 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3558 CERROR("ptlrpc: run out of memory!\n");
3559 RETURN(ERR_PTR(-ENOMEM));
3562 ptlrpc_cli_req_init(req);
3564 req->rq_send_state = LUSTRE_IMP_FULL;
3565 req->rq_type = PTL_RPC_MSG_REQUEST;
3566 req->rq_import = class_import_get(imp);
3567 req->rq_interpret_reply = work_interpreter;
3568 /* don't want reply */
3569 req->rq_no_delay = req->rq_no_resend = 1;
3570 req->rq_pill.rc_fmt = (void *)&worker_format;
3572 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3573 args = ptlrpc_req_async_args(req);
3575 args->cbdata = cbdata;
3579 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3581 void ptlrpcd_destroy_work(void *handler)
3583 struct ptlrpc_request *req = handler;
3586 ptlrpc_req_finished(req);
3588 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3590 int ptlrpcd_queue_work(void *handler)
3592 struct ptlrpc_request *req = handler;
3595 * Check if the req is already being queued.
3597 * Here comes a trick: it lacks a way of checking if a req is being
3598 * processed reliably in ptlrpc. Here I have to use refcount of req
3599 * for this purpose. This is okay because the caller should use this
3600 * req as opaque data. - Jinshan
3602 LASSERT(atomic_read(&req->rq_refcount) > 0);
3603 if (atomic_inc_return(&req->rq_refcount) == 2)
3604 ptlrpcd_add_work_req(req);
3607 EXPORT_SYMBOL(ptlrpcd_queue_work);