4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 /** Implementation of client-side PortalRPC interfaces */
35 #define DEBUG_SUBSYSTEM S_RPC
37 #include <linux/delay.h>
38 #include <linux/random.h>
40 #include <obd_support.h>
41 #include <obd_class.h>
42 #include <lustre_lib.h>
43 #include <lustre_ha.h>
44 #include <lustre_import.h>
45 #include <lustre_req_layout.h>
47 #include "ptlrpc_internal.h"
49 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
50 struct page *page, int pageoffset,
53 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
56 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
57 struct page *page, int pageoffset,
60 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
63 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
67 for (i = 0; i < desc->bd_iov_count ; i++)
68 put_page(BD_GET_KIOV(desc, i).kiov_page);
71 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
72 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
73 .release_frags = ptlrpc_release_bulk_page_pin,
75 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
77 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
78 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
79 .release_frags = ptlrpc_release_bulk_noop,
81 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
83 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
84 .add_iov_frag = ptlrpc_prep_bulk_frag,
86 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
88 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
89 static int ptlrpcd_check_work(struct ptlrpc_request *req);
90 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
93 * Initialize passed in client structure \a cl.
95 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
96 struct ptlrpc_client *cl)
98 cl->cli_request_portal = req_portal;
99 cl->cli_reply_portal = rep_portal;
102 EXPORT_SYMBOL(ptlrpc_init_client);
105 * Return PortalRPC connection for remore uud \a uuid
107 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
108 lnet_nid_t nid4refnet)
110 struct ptlrpc_connection *c;
112 struct lnet_process_id peer;
116 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
117 * before accessing its values.
119 /* coverity[uninit_use_in_call] */
120 peer.nid = nid4refnet;
121 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
123 CNETERR("cannot find peer %s!\n", uuid->uuid);
127 c = ptlrpc_connection_get(peer, self, uuid);
129 memcpy(c->c_remote_uuid.uuid,
130 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
133 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
139 * Allocate and initialize new bulk descriptor on the sender.
140 * Returns pointer to the descriptor or NULL on error.
142 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
143 unsigned int max_brw,
144 enum ptlrpc_bulk_op_type type,
146 const struct ptlrpc_bulk_frag_ops *ops)
148 struct ptlrpc_bulk_desc *desc;
151 /* ensure that only one of KIOV or IOVEC is set but not both */
152 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
153 ops->add_kiov_frag != NULL) ||
154 (ptlrpc_is_bulk_desc_kvec(type) &&
155 ops->add_iov_frag != NULL));
160 if (type & PTLRPC_BULK_BUF_KIOV) {
161 OBD_ALLOC_LARGE(GET_KIOV(desc),
162 nfrags * sizeof(*GET_KIOV(desc)));
166 OBD_ALLOC_LARGE(GET_KVEC(desc),
167 nfrags * sizeof(*GET_KVEC(desc)));
172 spin_lock_init(&desc->bd_lock);
173 init_waitqueue_head(&desc->bd_waitq);
174 desc->bd_max_iov = nfrags;
175 desc->bd_iov_count = 0;
176 desc->bd_portal = portal;
177 desc->bd_type = type;
178 desc->bd_md_count = 0;
179 desc->bd_frag_ops = ops;
180 LASSERT(max_brw > 0);
181 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
183 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
184 * node. Negotiated ocd_brw_size will always be <= this number.
186 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
187 LNetInvalidateMDHandle(&desc->bd_mds[i]);
196 * Prepare bulk descriptor for specified outgoing request \a req that
197 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
198 * the bulk to be sent. Used on client-side.
199 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
202 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
204 unsigned int max_brw,
207 const struct ptlrpc_bulk_frag_ops
210 struct obd_import *imp = req->rq_import;
211 struct ptlrpc_bulk_desc *desc;
214 LASSERT(ptlrpc_is_bulk_op_passive(type));
216 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
220 desc->bd_import_generation = req->rq_import_generation;
221 desc->bd_import = class_import_get(imp);
224 desc->bd_cbid.cbid_fn = client_bulk_callback;
225 desc->bd_cbid.cbid_arg = desc;
227 /* This makes req own desc, and free it when she frees herself */
232 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
234 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
235 struct page *page, int pageoffset, int len,
240 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
241 LASSERT(page != NULL);
242 LASSERT(pageoffset >= 0);
244 LASSERT(pageoffset + len <= PAGE_SIZE);
245 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
247 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
254 kiov->kiov_page = page;
255 kiov->kiov_offset = pageoffset;
256 kiov->kiov_len = len;
258 desc->bd_iov_count++;
260 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
262 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
269 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
270 LASSERT(frag != NULL);
272 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
274 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
278 iovec->iov_base = frag;
279 iovec->iov_len = len;
281 desc->bd_iov_count++;
283 RETURN(desc->bd_nob);
285 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
287 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
291 LASSERT(desc != NULL);
292 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
293 LASSERT(desc->bd_md_count == 0); /* network hands off */
294 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
295 LASSERT(desc->bd_frag_ops != NULL);
297 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
298 sptlrpc_enc_pool_put_pages(desc);
301 class_export_put(desc->bd_export);
303 class_import_put(desc->bd_import);
305 if (desc->bd_frag_ops->release_frags != NULL)
306 desc->bd_frag_ops->release_frags(desc);
308 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
309 OBD_FREE_LARGE(GET_KIOV(desc),
310 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
312 OBD_FREE_LARGE(GET_KVEC(desc),
313 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
317 EXPORT_SYMBOL(ptlrpc_free_bulk);
320 * Set server timelimit for this req, i.e. how long are we willing to wait
321 * for reply before timing out this request.
323 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
329 LASSERT(req->rq_import);
332 /* non-AT settings */
334 * \a imp_server_timeout means this is reverse import and
335 * we send (currently only) ASTs to the client and cannot afford
336 * to wait too long for the reply, otherwise the other client
337 * (because of which we are sending this request) would
338 * timeout waiting for us
340 req->rq_timeout = req->rq_import->imp_server_timeout ?
341 obd_timeout / 2 : obd_timeout;
343 at = &req->rq_import->imp_at;
344 idx = import_at_get_index(req->rq_import,
345 req->rq_request_portal);
346 serv_est = at_get(&at->iat_service_estimate[idx]);
347 req->rq_timeout = at_est2timeout(serv_est);
350 * We could get even fancier here, using history to predict increased
355 * Let the server know what this RPC timeout is by putting it in the
358 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
360 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
362 /* Adjust max service estimate based on server value */
363 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
364 unsigned int serv_est)
370 LASSERT(req->rq_import);
371 at = &req->rq_import->imp_at;
373 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
375 * max service estimates are tracked on the server side,
376 * so just keep minimal history here
378 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
381 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
382 req->rq_import->imp_obd->obd_name,
383 req->rq_request_portal,
384 oldse, at_get(&at->iat_service_estimate[idx]));
387 /* Expected network latency per remote node (secs) */
388 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
390 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
393 /* Adjust expected network latency */
394 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
395 unsigned int service_time)
397 unsigned int nl, oldnl;
399 time64_t now = ktime_get_real_seconds();
401 LASSERT(req->rq_import);
403 if (service_time > now - req->rq_sent + 3) {
405 * b=16408, however, this can also happen if early reply
406 * is lost and client RPC is expired and resent, early reply
407 * or reply of original RPC can still be fit in reply buffer
408 * of resent RPC, now client is measuring time from the
409 * resent time, but server sent back service time of original
412 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
413 D_ADAPTTO : D_WARNING,
414 "Reported service time %u > total measured time %lld\n",
415 service_time, now - req->rq_sent);
419 /* Network latency is total time less server processing time */
420 nl = max_t(int, now - req->rq_sent -
421 service_time, 0) + 1; /* st rounding */
422 at = &req->rq_import->imp_at;
424 oldnl = at_measured(&at->iat_net_latency, nl);
427 "The network latency for %s (nid %s) has changed from %d to %d\n",
428 req->rq_import->imp_obd->obd_name,
429 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
430 oldnl, at_get(&at->iat_net_latency));
433 static int unpack_reply(struct ptlrpc_request *req)
437 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
438 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
440 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
445 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
447 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
454 * Handle an early reply message, called with the rq_lock held.
455 * If anything goes wrong just ignore it - same as if it never happened
457 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
458 __must_hold(&req->rq_lock)
460 struct ptlrpc_request *early_req;
466 spin_unlock(&req->rq_lock);
468 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
470 spin_lock(&req->rq_lock);
474 rc = unpack_reply(early_req);
476 sptlrpc_cli_finish_early_reply(early_req);
477 spin_lock(&req->rq_lock);
482 * Use new timeout value just to adjust the local value for this
483 * request, don't include it into at_history. It is unclear yet why
484 * service time increased and should it be counted or skipped, e.g.
485 * that can be recovery case or some error or server, the real reply
486 * will add all new data if it is worth to add.
488 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
489 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
491 /* Network latency can be adjusted, it is pure network delays */
492 ptlrpc_at_adj_net_latency(req,
493 lustre_msg_get_service_time(early_req->rq_repmsg));
495 sptlrpc_cli_finish_early_reply(early_req);
497 spin_lock(&req->rq_lock);
498 olddl = req->rq_deadline;
500 * server assumes it now has rq_timeout from when the request
501 * arrived, so the client should give it at least that long.
502 * since we don't know the arrival time we'll use the original
505 req->rq_deadline = req->rq_sent + req->rq_timeout +
506 ptlrpc_at_get_net_latency(req);
508 DEBUG_REQ(D_ADAPTTO, req,
509 "Early reply #%d, new deadline in %llds (%llds)",
511 req->rq_deadline - ktime_get_real_seconds(),
512 req->rq_deadline - olddl);
517 static struct kmem_cache *request_cache;
519 int ptlrpc_request_cache_init(void)
521 request_cache = kmem_cache_create("ptlrpc_cache",
522 sizeof(struct ptlrpc_request),
523 0, SLAB_HWCACHE_ALIGN, NULL);
524 return request_cache ? 0 : -ENOMEM;
527 void ptlrpc_request_cache_fini(void)
529 kmem_cache_destroy(request_cache);
532 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
534 struct ptlrpc_request *req;
536 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
540 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
542 OBD_SLAB_FREE_PTR(req, request_cache);
546 * Wind down request pool \a pool.
547 * Frees all requests from the pool too
549 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
551 struct list_head *l, *tmp;
552 struct ptlrpc_request *req;
554 LASSERT(pool != NULL);
556 spin_lock(&pool->prp_lock);
557 list_for_each_safe(l, tmp, &pool->prp_req_list) {
558 req = list_entry(l, struct ptlrpc_request, rq_list);
559 list_del(&req->rq_list);
560 LASSERT(req->rq_reqbuf);
561 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
562 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
563 ptlrpc_request_cache_free(req);
565 spin_unlock(&pool->prp_lock);
566 OBD_FREE(pool, sizeof(*pool));
568 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
571 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
573 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
578 while (size < pool->prp_rq_size)
581 LASSERTF(list_empty(&pool->prp_req_list) ||
582 size == pool->prp_rq_size,
583 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
584 pool->prp_rq_size, size);
586 pool->prp_rq_size = size;
587 for (i = 0; i < num_rq; i++) {
588 struct ptlrpc_request *req;
589 struct lustre_msg *msg;
591 req = ptlrpc_request_cache_alloc(GFP_NOFS);
594 OBD_ALLOC_LARGE(msg, size);
596 ptlrpc_request_cache_free(req);
599 req->rq_reqbuf = msg;
600 req->rq_reqbuf_len = size;
602 spin_lock(&pool->prp_lock);
603 list_add_tail(&req->rq_list, &pool->prp_req_list);
604 spin_unlock(&pool->prp_lock);
608 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
611 * Create and initialize new request pool with given attributes:
612 * \a num_rq - initial number of requests to create for the pool
613 * \a msgsize - maximum message size possible for requests in thid pool
614 * \a populate_pool - function to be called when more requests need to be added
616 * Returns pointer to newly created pool or NULL on error.
618 struct ptlrpc_request_pool *
619 ptlrpc_init_rq_pool(int num_rq, int msgsize,
620 int (*populate_pool)(struct ptlrpc_request_pool *, int))
622 struct ptlrpc_request_pool *pool;
624 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
629 * Request next power of two for the allocation, because internally
630 * kernel would do exactly this
632 spin_lock_init(&pool->prp_lock);
633 INIT_LIST_HEAD(&pool->prp_req_list);
634 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
635 pool->prp_populate = populate_pool;
637 populate_pool(pool, num_rq);
641 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
644 * Fetches one request from pool \a pool
646 static struct ptlrpc_request *
647 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
649 struct ptlrpc_request *request;
650 struct lustre_msg *reqbuf;
655 spin_lock(&pool->prp_lock);
658 * See if we have anything in a pool, and bail out if nothing,
659 * in writeout path, where this matters, this is safe to do, because
660 * nothing is lost in this case, and when some in-flight requests
661 * complete, this code will be called again.
663 if (unlikely(list_empty(&pool->prp_req_list))) {
664 spin_unlock(&pool->prp_lock);
668 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
670 list_del_init(&request->rq_list);
671 spin_unlock(&pool->prp_lock);
673 LASSERT(request->rq_reqbuf);
674 LASSERT(request->rq_pool);
676 reqbuf = request->rq_reqbuf;
677 memset(request, 0, sizeof(*request));
678 request->rq_reqbuf = reqbuf;
679 request->rq_reqbuf_len = pool->prp_rq_size;
680 request->rq_pool = pool;
686 * Returns freed \a request to pool.
688 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
690 struct ptlrpc_request_pool *pool = request->rq_pool;
692 spin_lock(&pool->prp_lock);
693 LASSERT(list_empty(&request->rq_list));
694 LASSERT(!request->rq_receiving_reply);
695 list_add_tail(&request->rq_list, &pool->prp_req_list);
696 spin_unlock(&pool->prp_lock);
699 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
701 struct obd_import *imp = req->rq_import;
702 struct list_head *tmp;
703 struct ptlrpc_request *iter;
705 assert_spin_locked(&imp->imp_lock);
706 LASSERT(list_empty(&req->rq_unreplied_list));
708 /* unreplied list is sorted by xid in ascending order */
709 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
710 iter = list_entry(tmp, struct ptlrpc_request,
713 LASSERT(req->rq_xid != iter->rq_xid);
714 if (req->rq_xid < iter->rq_xid)
716 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
719 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
722 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
724 req->rq_xid = ptlrpc_next_xid();
725 ptlrpc_add_unreplied(req);
728 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
730 spin_lock(&req->rq_import->imp_lock);
731 ptlrpc_assign_next_xid_nolock(req);
732 spin_unlock(&req->rq_import->imp_lock);
735 static atomic64_t ptlrpc_last_xid;
737 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
738 __u32 version, int opcode, char **bufs,
739 struct ptlrpc_cli_ctx *ctx)
742 struct obd_import *imp;
748 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
749 imp = request->rq_import;
750 lengths = request->rq_pill.rc_area[RCL_CLIENT];
753 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
755 rc = sptlrpc_req_get_ctx(request);
759 sptlrpc_req_set_flavor(request, opcode);
761 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
766 lustre_msg_add_version(request->rq_reqmsg, version);
767 request->rq_send_state = LUSTRE_IMP_FULL;
768 request->rq_type = PTL_RPC_MSG_REQUEST;
770 request->rq_req_cbid.cbid_fn = request_out_callback;
771 request->rq_req_cbid.cbid_arg = request;
773 request->rq_reply_cbid.cbid_fn = reply_in_callback;
774 request->rq_reply_cbid.cbid_arg = request;
776 request->rq_reply_deadline = 0;
777 request->rq_bulk_deadline = 0;
778 request->rq_req_deadline = 0;
779 request->rq_phase = RQ_PHASE_NEW;
780 request->rq_next_phase = RQ_PHASE_UNDEFINED;
782 request->rq_request_portal = imp->imp_client->cli_request_portal;
783 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
785 ptlrpc_at_set_req_timeout(request);
787 lustre_msg_set_opc(request->rq_reqmsg, opcode);
789 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
790 if (cfs_fail_val == opcode) {
791 time64_t *fail_t = NULL, *fail2_t = NULL;
793 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
794 fail_t = &request->rq_bulk_deadline;
795 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
796 fail_t = &request->rq_reply_deadline;
797 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
798 fail_t = &request->rq_req_deadline;
799 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
800 fail_t = &request->rq_reply_deadline;
801 fail2_t = &request->rq_bulk_deadline;
802 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
803 time64_t now = ktime_get_real_seconds();
804 u64 xid = ((u64)now >> 4) << 24;
806 atomic64_set(&ptlrpc_last_xid, xid);
810 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
813 *fail2_t = ktime_get_real_seconds() +
817 * The RPC is infected, let the test to change the
820 msleep(4 * MSEC_PER_SEC);
823 ptlrpc_assign_next_xid(request);
828 LASSERT(!request->rq_pool);
829 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
831 class_import_put(imp);
835 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
838 * Pack request buffers for network transfer, performing necessary encryption
839 * steps if necessary.
841 int ptlrpc_request_pack(struct ptlrpc_request *request,
842 __u32 version, int opcode)
846 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
851 * For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
852 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
853 * have to send old ptlrpc_body to keep interoprability with these
856 * Only three kinds of server->client RPCs so far:
861 * XXX This should be removed whenever we drop the interoprability with
862 * the these old clients.
864 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
865 opcode == LDLM_GL_CALLBACK)
866 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
867 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
871 EXPORT_SYMBOL(ptlrpc_request_pack);
874 * Helper function to allocate new request on import \a imp
875 * and possibly using existing request from pool \a pool if provided.
876 * Returns allocated request structure with import field filled or
880 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
881 struct ptlrpc_request_pool *pool)
883 struct ptlrpc_request *request = NULL;
885 request = ptlrpc_request_cache_alloc(GFP_NOFS);
887 if (!request && pool)
888 request = ptlrpc_prep_req_from_pool(pool);
891 ptlrpc_cli_req_init(request);
893 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
894 LASSERT(imp != LP_POISON);
895 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
897 LASSERT(imp->imp_client != LP_POISON);
899 request->rq_import = class_import_get(imp);
901 CERROR("request allocation out of memory\n");
908 * Helper function for creating a request.
909 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
910 * buffer structures according to capsule template \a format.
911 * Returns allocated request structure pointer or NULL on error.
913 static struct ptlrpc_request *
914 ptlrpc_request_alloc_internal(struct obd_import *imp,
915 struct ptlrpc_request_pool *pool,
916 const struct req_format *format)
918 struct ptlrpc_request *request;
921 request = __ptlrpc_request_alloc(imp, pool);
926 * initiate connection if needed when the import has been
927 * referenced by the new request to avoid races with disconnect
929 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
932 CDEBUG_LIMIT(imp->imp_idle_debug,
933 "%s: reconnect after %llds idle\n",
934 imp->imp_obd->obd_name, ktime_get_real_seconds() -
935 imp->imp_last_reply_time);
936 spin_lock(&imp->imp_lock);
937 if (imp->imp_state == LUSTRE_IMP_IDLE) {
938 imp->imp_generation++;
939 imp->imp_initiated_at = imp->imp_generation;
940 imp->imp_state = LUSTRE_IMP_NEW;
943 spin_unlock(&imp->imp_lock);
945 rc = ptlrpc_connect_import(imp);
947 ptlrpc_request_free(request);
950 ptlrpc_pinger_add_import(imp);
954 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
955 req_capsule_set(&request->rq_pill, format);
960 * Allocate new request structure for import \a imp and initialize its
961 * buffer structure according to capsule template \a format.
963 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
964 const struct req_format *format)
966 return ptlrpc_request_alloc_internal(imp, NULL, format);
968 EXPORT_SYMBOL(ptlrpc_request_alloc);
971 * Allocate new request structure for import \a imp from pool \a pool and
972 * initialize its buffer structure according to capsule template \a format.
974 struct ptlrpc_request *
975 ptlrpc_request_alloc_pool(struct obd_import *imp,
976 struct ptlrpc_request_pool *pool,
977 const struct req_format *format)
979 return ptlrpc_request_alloc_internal(imp, pool, format);
981 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
984 * For requests not from pool, free memory of the request structure.
985 * For requests obtained from a pool earlier, return request back to pool.
987 void ptlrpc_request_free(struct ptlrpc_request *request)
989 if (request->rq_pool)
990 __ptlrpc_free_req_to_pool(request);
992 ptlrpc_request_cache_free(request);
994 EXPORT_SYMBOL(ptlrpc_request_free);
997 * Allocate new request for operatione \a opcode and immediatelly pack it for
999 * Only used for simple requests like OBD_PING where the only important
1000 * part of the request is operation itself.
1001 * Returns allocated request or NULL on error.
1003 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1004 const struct req_format *format,
1005 __u32 version, int opcode)
1007 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1011 rc = ptlrpc_request_pack(req, version, opcode);
1013 ptlrpc_request_free(req);
1019 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1022 * Allocate and initialize new request set structure on the current CPT.
1023 * Returns a pointer to the newly allocated set structure or NULL on error.
1025 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1027 struct ptlrpc_request_set *set;
1031 cpt = cfs_cpt_current(cfs_cpt_table, 0);
1032 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1035 atomic_set(&set->set_refcount, 1);
1036 INIT_LIST_HEAD(&set->set_requests);
1037 init_waitqueue_head(&set->set_waitq);
1038 atomic_set(&set->set_new_count, 0);
1039 atomic_set(&set->set_remaining, 0);
1040 spin_lock_init(&set->set_new_req_lock);
1041 INIT_LIST_HEAD(&set->set_new_requests);
1042 set->set_max_inflight = UINT_MAX;
1043 set->set_producer = NULL;
1044 set->set_producer_arg = NULL;
1049 EXPORT_SYMBOL(ptlrpc_prep_set);
1052 * Allocate and initialize new request set structure with flow control
1053 * extension. This extension allows to control the number of requests in-flight
1054 * for the whole set. A callback function to generate requests must be provided
1055 * and the request set will keep the number of requests sent over the wire to
1057 * Returns a pointer to the newly allocated set structure or NULL on error.
1059 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1063 struct ptlrpc_request_set *set;
1065 set = ptlrpc_prep_set();
1069 set->set_max_inflight = max;
1070 set->set_producer = func;
1071 set->set_producer_arg = arg;
1077 * Wind down and free request set structure previously allocated with
1079 * Ensures that all requests on the set have completed and removes
1080 * all requests from the request list in a set.
1081 * If any unsent request happen to be on the list, pretends that they got
1082 * an error in flight and calls their completion handler.
1084 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1086 struct list_head *tmp;
1087 struct list_head *next;
1093 /* Requests on the set should either all be completed, or all be new */
1094 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1095 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1096 list_for_each(tmp, &set->set_requests) {
1097 struct ptlrpc_request *req =
1098 list_entry(tmp, struct ptlrpc_request,
1101 LASSERT(req->rq_phase == expected_phase);
1105 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1106 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1107 atomic_read(&set->set_remaining), n);
1109 list_for_each_safe(tmp, next, &set->set_requests) {
1110 struct ptlrpc_request *req =
1111 list_entry(tmp, struct ptlrpc_request,
1113 list_del_init(&req->rq_set_chain);
1115 LASSERT(req->rq_phase == expected_phase);
1117 if (req->rq_phase == RQ_PHASE_NEW) {
1118 ptlrpc_req_interpret(NULL, req, -EBADR);
1119 atomic_dec(&set->set_remaining);
1122 spin_lock(&req->rq_lock);
1124 req->rq_invalid_rqset = 0;
1125 spin_unlock(&req->rq_lock);
1127 ptlrpc_req_finished(req);
1130 LASSERT(atomic_read(&set->set_remaining) == 0);
1132 ptlrpc_reqset_put(set);
1135 EXPORT_SYMBOL(ptlrpc_set_destroy);
1138 * Add a new request to the general purpose request set.
1139 * Assumes request reference from the caller.
1141 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1142 struct ptlrpc_request *req)
1144 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1145 LASSERT(list_empty(&req->rq_set_chain));
1147 if (req->rq_allow_intr)
1148 set->set_allow_intr = 1;
1150 /* The set takes over the caller's request reference */
1151 list_add_tail(&req->rq_set_chain, &set->set_requests);
1153 atomic_inc(&set->set_remaining);
1154 req->rq_queued_time = ktime_get_seconds();
1157 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1159 if (set->set_producer)
1161 * If the request set has a producer callback, the RPC must be
1162 * sent straight away
1164 ptlrpc_send_new_req(req);
1166 EXPORT_SYMBOL(ptlrpc_set_add_req);
1169 * Add a request to a request with dedicated server thread
1170 * and wake the thread to make any necessary processing.
1171 * Currently only used for ptlrpcd.
1173 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1174 struct ptlrpc_request *req)
1176 struct ptlrpc_request_set *set = pc->pc_set;
1179 LASSERT(req->rq_set == NULL);
1180 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1182 spin_lock(&set->set_new_req_lock);
1184 * The set takes over the caller's request reference.
1187 req->rq_queued_time = ktime_get_seconds();
1188 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1189 count = atomic_inc_return(&set->set_new_count);
1190 spin_unlock(&set->set_new_req_lock);
1192 /* Only need to call wakeup once for the first entry. */
1194 wake_up(&set->set_waitq);
1197 * XXX: It maybe unnecessary to wakeup all the partners. But to
1198 * guarantee the async RPC can be processed ASAP, we have
1199 * no other better choice. It maybe fixed in future.
1201 for (i = 0; i < pc->pc_npartners; i++)
1202 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1207 * Based on the current state of the import, determine if the request
1208 * can be sent, is an error, or should be delayed.
1210 * Returns true if this request should be delayed. If false, and
1211 * *status is set, then the request can not be sent and *status is the
1212 * error code. If false and status is 0, then request can be sent.
1214 * The imp->imp_lock must be held.
1216 static int ptlrpc_import_delay_req(struct obd_import *imp,
1217 struct ptlrpc_request *req, int *status)
1225 if (req->rq_ctx_init || req->rq_ctx_fini) {
1226 /* always allow ctx init/fini rpc go through */
1227 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1228 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1230 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1231 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1234 * pings or MDS-equivalent STATFS may safely
1237 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1238 D_HA : D_ERROR, req, "IMP_CLOSED ");
1240 } else if (ptlrpc_send_limit_expired(req)) {
1241 /* probably doesn't need to be a D_ERROR afterinitial testing */
1242 DEBUG_REQ(D_HA, req, "send limit expired ");
1243 *status = -ETIMEDOUT;
1244 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1245 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1246 ;/* allow CONNECT even if import is invalid */
1247 if (atomic_read(&imp->imp_inval_count) != 0) {
1248 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1251 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1252 if (!imp->imp_deactive)
1253 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1254 *status = -ESHUTDOWN; /* b=12940 */
1255 } else if (req->rq_import_generation != imp->imp_generation) {
1256 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1258 } else if (req->rq_send_state != imp->imp_state) {
1259 /* invalidate in progress - any requests should be drop */
1260 if (atomic_read(&imp->imp_inval_count) != 0) {
1261 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1263 } else if (req->rq_no_delay &&
1264 imp->imp_generation != imp->imp_initiated_at) {
1265 /* ignore nodelay for requests initiating connections */
1266 *status = -EWOULDBLOCK;
1267 } else if (req->rq_allow_replay &&
1268 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1269 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1270 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1271 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1272 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1282 * Decide if the error message should be printed to the console or not.
1283 * Makes its decision based on request type, status, and failure frequency.
1285 * \param[in] req request that failed and may need a console message
1287 * \retval false if no message should be printed
1288 * \retval true if console message should be printed
1290 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1292 LASSERT(req->rq_reqmsg != NULL);
1294 /* Suppress particular reconnect errors which are to be expected. */
1295 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1296 /* Suppress timed out reconnect requests */
1297 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1302 * Suppress most unavailable/again reconnect requests, but
1303 * print occasionally so it is clear client is trying to
1304 * connect to a server where no target is running.
1306 if ((err == -ENODEV || err == -EAGAIN) &&
1307 req->rq_import->imp_conn_cnt % 30 != 20)
1311 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1312 /* -EAGAIN is normal when using POSIX flocks */
1315 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1316 (req->rq_xid & 0xf) != 10)
1317 /* Suppress most ping requests, they may fail occasionally */
1324 * Check request processing status.
1325 * Returns the status.
1327 static int ptlrpc_check_status(struct ptlrpc_request *req)
1332 err = lustre_msg_get_status(req->rq_repmsg);
1333 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1334 struct obd_import *imp = req->rq_import;
1335 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1336 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1338 if (ptlrpc_console_allow(req, opc, err))
1339 LCONSOLE_ERROR_MSG(0x11,
1340 "%s: operation %s to node %s failed: rc = %d\n",
1341 imp->imp_obd->obd_name,
1343 libcfs_nid2str(nid), err);
1344 RETURN(err < 0 ? err : -EINVAL);
1348 DEBUG_REQ(D_INFO, req, "status is %d", err);
1349 } else if (err > 0) {
1350 /* XXX: translate this error from net to host */
1351 DEBUG_REQ(D_INFO, req, "status is %d", err);
1358 * save pre-versions of objects into request for replay.
1359 * Versions are obtained from server reply.
1362 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1364 struct lustre_msg *repmsg = req->rq_repmsg;
1365 struct lustre_msg *reqmsg = req->rq_reqmsg;
1366 __u64 *versions = lustre_msg_get_versions(repmsg);
1369 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1373 lustre_msg_set_versions(reqmsg, versions);
1374 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1375 versions[0], versions[1]);
1380 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1382 struct ptlrpc_request *req;
1384 assert_spin_locked(&imp->imp_lock);
1385 if (list_empty(&imp->imp_unreplied_list))
1388 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1390 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1392 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1393 imp->imp_known_replied_xid = req->rq_xid - 1;
1395 return req->rq_xid - 1;
1399 * Callback function called when client receives RPC reply for \a req.
1400 * Returns 0 on success or error code.
1401 * The return alue would be assigned to req->rq_status by the caller
1402 * as request processing status.
1403 * This function also decides if the request needs to be saved for later replay.
1405 static int after_reply(struct ptlrpc_request *req)
1407 struct obd_import *imp = req->rq_import;
1408 struct obd_device *obd = req->rq_import->imp_obd;
1415 LASSERT(obd != NULL);
1416 /* repbuf must be unlinked */
1417 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1419 if (req->rq_reply_truncated) {
1420 if (ptlrpc_no_resend(req)) {
1421 DEBUG_REQ(D_ERROR, req,
1422 "reply buffer overflow, expected: %d, actual size: %d",
1423 req->rq_nob_received, req->rq_repbuf_len);
1427 sptlrpc_cli_free_repbuf(req);
1429 * Pass the required reply buffer size (include
1430 * space for early reply).
1431 * NB: no need to roundup because alloc_repbuf
1434 req->rq_replen = req->rq_nob_received;
1435 req->rq_nob_received = 0;
1436 spin_lock(&req->rq_lock);
1438 spin_unlock(&req->rq_lock);
1442 work_start = ktime_get_real();
1443 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1446 * NB Until this point, the whole of the incoming message,
1447 * including buflens, status etc is in the sender's byte order.
1449 rc = sptlrpc_cli_unwrap_reply(req);
1451 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1456 * Security layer unwrap might ask resend this request.
1461 rc = unpack_reply(req);
1465 /* retry indefinitely on EINPROGRESS */
1466 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1467 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1468 time64_t now = ktime_get_real_seconds();
1470 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1471 spin_lock(&req->rq_lock);
1473 spin_unlock(&req->rq_lock);
1474 req->rq_nr_resend++;
1476 /* Readjust the timeout for current conditions */
1477 ptlrpc_at_set_req_timeout(req);
1479 * delay resend to give a chance to the server to get ready.
1480 * The delay is increased by 1s on every resend and is capped to
1481 * the current request timeout (i.e. obd_timeout if AT is off,
1482 * or AT service time x 125% + 5s, see at_est2timeout)
1484 if (req->rq_nr_resend > req->rq_timeout)
1485 req->rq_sent = now + req->rq_timeout;
1487 req->rq_sent = now + req->rq_nr_resend;
1489 /* Resend for EINPROGRESS will use a new XID */
1490 spin_lock(&imp->imp_lock);
1491 list_del_init(&req->rq_unreplied_list);
1492 spin_unlock(&imp->imp_lock);
1497 if (obd->obd_svc_stats) {
1498 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1500 ptlrpc_lprocfs_rpc_sent(req, timediff);
1503 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1504 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1505 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1506 lustre_msg_get_type(req->rq_repmsg));
1510 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1511 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1512 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1513 ptlrpc_at_adj_net_latency(req,
1514 lustre_msg_get_service_time(req->rq_repmsg));
1516 rc = ptlrpc_check_status(req);
1520 * Either we've been evicted, or the server has failed for
1521 * some reason. Try to reconnect, and if that fails, punt to
1524 if (ptlrpc_recoverable_error(rc)) {
1525 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1526 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1529 ptlrpc_request_handle_notconn(req);
1534 * Let's look if server sent slv. Do it only for RPC with
1537 ldlm_cli_update_pool(req);
1541 * Store transno in reqmsg for replay.
1543 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1544 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1545 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1548 if (imp->imp_replayable) {
1549 spin_lock(&imp->imp_lock);
1551 * No point in adding already-committed requests to the replay
1552 * list, we will just remove them immediately. b=9829
1554 if (req->rq_transno != 0 &&
1556 lustre_msg_get_last_committed(req->rq_repmsg) ||
1558 /** version recovery */
1559 ptlrpc_save_versions(req);
1560 ptlrpc_retain_replayable_request(req, imp);
1561 } else if (req->rq_commit_cb &&
1562 list_empty(&req->rq_replay_list)) {
1564 * NB: don't call rq_commit_cb if it's already on
1565 * rq_replay_list, ptlrpc_free_committed() will call
1566 * it later, see LU-3618 for details
1568 spin_unlock(&imp->imp_lock);
1569 req->rq_commit_cb(req);
1570 spin_lock(&imp->imp_lock);
1574 * Replay-enabled imports return commit-status information.
1576 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1577 if (likely(committed > imp->imp_peer_committed_transno))
1578 imp->imp_peer_committed_transno = committed;
1580 ptlrpc_free_committed(imp);
1582 if (!list_empty(&imp->imp_replay_list)) {
1583 struct ptlrpc_request *last;
1585 last = list_entry(imp->imp_replay_list.prev,
1586 struct ptlrpc_request,
1589 * Requests with rq_replay stay on the list even if no
1590 * commit is expected.
1592 if (last->rq_transno > imp->imp_peer_committed_transno)
1593 ptlrpc_pinger_commit_expected(imp);
1596 spin_unlock(&imp->imp_lock);
1603 * Helper function to send request \a req over the network for the first time
1604 * Also adjusts request phase.
1605 * Returns 0 on success or error code.
1607 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1609 struct obd_import *imp = req->rq_import;
1614 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1616 /* do not try to go further if there is not enough memory in enc_pool */
1617 if (req->rq_sent && req->rq_bulk)
1618 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1619 pool_is_at_full_capacity())
1622 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1623 (!req->rq_generation_set ||
1624 req->rq_import_generation == imp->imp_generation))
1627 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1629 spin_lock(&imp->imp_lock);
1631 LASSERT(req->rq_xid != 0);
1632 LASSERT(!list_empty(&req->rq_unreplied_list));
1634 if (!req->rq_generation_set)
1635 req->rq_import_generation = imp->imp_generation;
1637 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1638 spin_lock(&req->rq_lock);
1639 req->rq_waiting = 1;
1640 spin_unlock(&req->rq_lock);
1642 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1643 ptlrpc_import_state_name(req->rq_send_state),
1644 ptlrpc_import_state_name(imp->imp_state));
1645 LASSERT(list_empty(&req->rq_list));
1646 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1647 atomic_inc(&req->rq_import->imp_inflight);
1648 spin_unlock(&imp->imp_lock);
1653 spin_unlock(&imp->imp_lock);
1654 req->rq_status = rc;
1655 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1659 LASSERT(list_empty(&req->rq_list));
1660 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1661 atomic_inc(&req->rq_import->imp_inflight);
1664 * find the known replied XID from the unreplied list, CONNECT
1665 * and DISCONNECT requests are skipped to make the sanity check
1666 * on server side happy. see process_req_last_xid().
1668 * For CONNECT: Because replay requests have lower XID, it'll
1669 * break the sanity check if CONNECT bump the exp_last_xid on
1672 * For DISCONNECT: Since client will abort inflight RPC before
1673 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1674 * than the inflight RPC.
1676 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1677 min_xid = ptlrpc_known_replied_xid(imp);
1678 spin_unlock(&imp->imp_lock);
1680 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1682 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1684 rc = sptlrpc_req_refresh_ctx(req, -1);
1687 req->rq_status = rc;
1690 spin_lock(&req->rq_lock);
1691 req->rq_wait_ctx = 1;
1692 spin_unlock(&req->rq_lock);
1698 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1699 req, current_comm(),
1700 imp->imp_obd->obd_uuid.uuid,
1701 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1702 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1703 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1705 rc = ptl_send_rpc(req, 0);
1706 if (rc == -ENOMEM) {
1707 spin_lock(&imp->imp_lock);
1708 if (!list_empty(&req->rq_list)) {
1709 list_del_init(&req->rq_list);
1710 atomic_dec(&req->rq_import->imp_inflight);
1712 spin_unlock(&imp->imp_lock);
1713 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1717 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1718 spin_lock(&req->rq_lock);
1719 req->rq_net_err = 1;
1720 spin_unlock(&req->rq_lock);
1726 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1731 LASSERT(set->set_producer != NULL);
1733 remaining = atomic_read(&set->set_remaining);
1736 * populate the ->set_requests list with requests until we
1737 * reach the maximum number of RPCs in flight for this set
1739 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1740 rc = set->set_producer(set, set->set_producer_arg);
1741 if (rc == -ENOENT) {
1742 /* no more RPC to produce */
1743 set->set_producer = NULL;
1744 set->set_producer_arg = NULL;
1749 RETURN((atomic_read(&set->set_remaining) - remaining));
1753 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1754 * and no more replies are expected.
1755 * (it is possible to get less replies than requests sent e.g. due to timed out
1756 * requests or requests that we had trouble to send out)
1758 * NOTE: This function contains a potential schedule point (cond_resched()).
1760 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1762 struct list_head *tmp, *next;
1763 struct list_head comp_reqs;
1764 int force_timer_recalc = 0;
1767 if (atomic_read(&set->set_remaining) == 0)
1770 INIT_LIST_HEAD(&comp_reqs);
1771 list_for_each_safe(tmp, next, &set->set_requests) {
1772 struct ptlrpc_request *req =
1773 list_entry(tmp, struct ptlrpc_request,
1775 struct obd_import *imp = req->rq_import;
1776 int unregistered = 0;
1780 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1781 list_move_tail(&req->rq_set_chain, &comp_reqs);
1786 * This schedule point is mainly for the ptlrpcd caller of this
1787 * function. Most ptlrpc sets are not long-lived and unbounded
1788 * in length, but at the least the set used by the ptlrpcd is.
1789 * Since the processing time is unbounded, we need to insert an
1790 * explicit schedule point to make the thread well-behaved.
1795 * If the caller requires to allow to be interpreted by force
1796 * and it has really been interpreted, then move the request
1797 * to RQ_PHASE_INTERPRET phase in spite of what the current
1800 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1801 req->rq_status = -EINTR;
1802 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1805 * Since it is interpreted and we have to wait for
1806 * the reply to be unlinked, then use sync mode.
1810 GOTO(interpret, req->rq_status);
1813 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1814 force_timer_recalc = 1;
1816 /* delayed send - skip */
1817 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1820 /* delayed resend - skip */
1821 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1822 req->rq_sent > ktime_get_real_seconds())
1825 if (!(req->rq_phase == RQ_PHASE_RPC ||
1826 req->rq_phase == RQ_PHASE_BULK ||
1827 req->rq_phase == RQ_PHASE_INTERPRET ||
1828 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1829 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1830 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1834 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1835 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1836 LASSERT(req->rq_next_phase != req->rq_phase);
1837 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1839 if (req->rq_req_deadline &&
1840 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1841 req->rq_req_deadline = 0;
1842 if (req->rq_reply_deadline &&
1843 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1844 req->rq_reply_deadline = 0;
1845 if (req->rq_bulk_deadline &&
1846 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1847 req->rq_bulk_deadline = 0;
1850 * Skip processing until reply is unlinked. We
1851 * can't return to pool before that and we can't
1852 * call interpret before that. We need to make
1853 * sure that all rdma transfers finished and will
1854 * not corrupt any data.
1856 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1857 ptlrpc_client_recv_or_unlink(req))
1859 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1860 ptlrpc_client_bulk_active(req))
1864 * Turn fail_loc off to prevent it from looping
1867 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1868 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1871 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1872 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1877 * Move to next phase if reply was successfully
1880 ptlrpc_rqphase_move(req, req->rq_next_phase);
1883 if (req->rq_phase == RQ_PHASE_INTERPRET)
1884 GOTO(interpret, req->rq_status);
1887 * Note that this also will start async reply unlink.
1889 if (req->rq_net_err && !req->rq_timedout) {
1890 ptlrpc_expire_one_request(req, 1);
1893 * Check if we still need to wait for unlink.
1895 if (ptlrpc_client_recv_or_unlink(req) ||
1896 ptlrpc_client_bulk_active(req))
1898 /* If there is no need to resend, fail it now. */
1899 if (req->rq_no_resend) {
1900 if (req->rq_status == 0)
1901 req->rq_status = -EIO;
1902 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1903 GOTO(interpret, req->rq_status);
1910 spin_lock(&req->rq_lock);
1911 req->rq_replied = 0;
1912 spin_unlock(&req->rq_lock);
1913 if (req->rq_status == 0)
1914 req->rq_status = -EIO;
1915 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1916 GOTO(interpret, req->rq_status);
1920 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1921 * so it sets rq_intr regardless of individual rpc
1922 * timeouts. The synchronous IO waiting path sets
1923 * rq_intr irrespective of whether ptlrpcd
1924 * has seen a timeout. Our policy is to only interpret
1925 * interrupted rpcs after they have timed out, so we
1926 * need to enforce that here.
1929 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1930 req->rq_wait_ctx)) {
1931 req->rq_status = -EINTR;
1932 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1933 GOTO(interpret, req->rq_status);
1936 if (req->rq_phase == RQ_PHASE_RPC) {
1937 if (req->rq_timedout || req->rq_resend ||
1938 req->rq_waiting || req->rq_wait_ctx) {
1941 if (!ptlrpc_unregister_reply(req, 1)) {
1942 ptlrpc_unregister_bulk(req, 1);
1946 spin_lock(&imp->imp_lock);
1947 if (ptlrpc_import_delay_req(imp, req,
1950 * put on delay list - only if we wait
1951 * recovery finished - before send
1953 list_move_tail(&req->rq_list,
1954 &imp->imp_delayed_list);
1955 spin_unlock(&imp->imp_lock);
1960 req->rq_status = status;
1961 ptlrpc_rqphase_move(req,
1962 RQ_PHASE_INTERPRET);
1963 spin_unlock(&imp->imp_lock);
1964 GOTO(interpret, req->rq_status);
1966 /* ignore on just initiated connections */
1967 if (ptlrpc_no_resend(req) &&
1968 !req->rq_wait_ctx &&
1969 imp->imp_generation !=
1970 imp->imp_initiated_at) {
1971 req->rq_status = -ENOTCONN;
1972 ptlrpc_rqphase_move(req,
1973 RQ_PHASE_INTERPRET);
1974 spin_unlock(&imp->imp_lock);
1975 GOTO(interpret, req->rq_status);
1978 list_move_tail(&req->rq_list,
1979 &imp->imp_sending_list);
1981 spin_unlock(&imp->imp_lock);
1983 spin_lock(&req->rq_lock);
1984 req->rq_waiting = 0;
1985 spin_unlock(&req->rq_lock);
1987 if (req->rq_timedout || req->rq_resend) {
1989 * This is re-sending anyways,
1990 * let's mark req as resend.
1992 spin_lock(&req->rq_lock);
1994 spin_unlock(&req->rq_lock);
1997 * rq_wait_ctx is only touched by ptlrpcd,
1998 * so no lock is needed here.
2000 status = sptlrpc_req_refresh_ctx(req, -1);
2003 req->rq_status = status;
2004 spin_lock(&req->rq_lock);
2005 req->rq_wait_ctx = 0;
2006 spin_unlock(&req->rq_lock);
2007 force_timer_recalc = 1;
2009 spin_lock(&req->rq_lock);
2010 req->rq_wait_ctx = 1;
2011 spin_unlock(&req->rq_lock);
2016 spin_lock(&req->rq_lock);
2017 req->rq_wait_ctx = 0;
2018 spin_unlock(&req->rq_lock);
2022 * In any case, the previous bulk should be
2023 * cleaned up to prepare for the new sending
2026 !ptlrpc_unregister_bulk(req, 1))
2029 rc = ptl_send_rpc(req, 0);
2030 if (rc == -ENOMEM) {
2031 spin_lock(&imp->imp_lock);
2032 if (!list_empty(&req->rq_list))
2033 list_del_init(&req->rq_list);
2034 spin_unlock(&imp->imp_lock);
2035 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2039 DEBUG_REQ(D_HA, req,
2040 "send failed: rc = %d", rc);
2041 force_timer_recalc = 1;
2042 spin_lock(&req->rq_lock);
2043 req->rq_net_err = 1;
2044 spin_unlock(&req->rq_lock);
2047 /* need to reset the timeout */
2048 force_timer_recalc = 1;
2051 spin_lock(&req->rq_lock);
2053 if (ptlrpc_client_early(req)) {
2054 ptlrpc_at_recv_early_reply(req);
2055 spin_unlock(&req->rq_lock);
2059 /* Still waiting for a reply? */
2060 if (ptlrpc_client_recv(req)) {
2061 spin_unlock(&req->rq_lock);
2065 /* Did we actually receive a reply? */
2066 if (!ptlrpc_client_replied(req)) {
2067 spin_unlock(&req->rq_lock);
2071 spin_unlock(&req->rq_lock);
2074 * unlink from net because we are going to
2075 * swab in-place of reply buffer
2077 unregistered = ptlrpc_unregister_reply(req, 1);
2081 req->rq_status = after_reply(req);
2086 * If there is no bulk associated with this request,
2087 * then we're done and should let the interpreter
2088 * process the reply. Similarly if the RPC returned
2089 * an error, and therefore the bulk will never arrive.
2091 if (!req->rq_bulk || req->rq_status < 0) {
2092 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2093 GOTO(interpret, req->rq_status);
2096 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2099 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2100 if (ptlrpc_client_bulk_active(req))
2103 if (req->rq_bulk->bd_failure) {
2105 * The RPC reply arrived OK, but the bulk screwed
2106 * up! Dead weird since the server told us the RPC
2107 * was good after getting the REPLY for her GET or
2108 * the ACK for her PUT.
2110 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2111 req->rq_status = -EIO;
2114 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2117 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2120 * This moves to "unregistering" phase we need to wait for
2123 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2124 /* start async bulk unlink too */
2125 ptlrpc_unregister_bulk(req, 1);
2129 if (!ptlrpc_unregister_bulk(req, async))
2133 * When calling interpret receiving already should be
2136 LASSERT(!req->rq_receiving_reply);
2138 ptlrpc_req_interpret(env, req, req->rq_status);
2140 if (ptlrpcd_check_work(req)) {
2141 atomic_dec(&set->set_remaining);
2144 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2148 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2149 req, current_comm(),
2150 imp->imp_obd->obd_uuid.uuid,
2151 lustre_msg_get_status(req->rq_reqmsg),
2153 obd_import_nid2str(imp),
2154 lustre_msg_get_opc(req->rq_reqmsg),
2155 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2157 spin_lock(&imp->imp_lock);
2159 * Request already may be not on sending or delaying list. This
2160 * may happen in the case of marking it erroneous for the case
2161 * ptlrpc_import_delay_req(req, status) find it impossible to
2162 * allow sending this rpc and returns *status != 0.
2164 if (!list_empty(&req->rq_list)) {
2165 list_del_init(&req->rq_list);
2166 atomic_dec(&imp->imp_inflight);
2168 list_del_init(&req->rq_unreplied_list);
2169 spin_unlock(&imp->imp_lock);
2171 atomic_dec(&set->set_remaining);
2172 wake_up_all(&imp->imp_recovery_waitq);
2174 if (set->set_producer) {
2175 /* produce a new request if possible */
2176 if (ptlrpc_set_producer(set) > 0)
2177 force_timer_recalc = 1;
2180 * free the request that has just been completed
2181 * in order not to pollute set->set_requests
2183 list_del_init(&req->rq_set_chain);
2184 spin_lock(&req->rq_lock);
2186 req->rq_invalid_rqset = 0;
2187 spin_unlock(&req->rq_lock);
2189 /* record rq_status to compute the final status later */
2190 if (req->rq_status != 0)
2191 set->set_rc = req->rq_status;
2192 ptlrpc_req_finished(req);
2194 list_move_tail(&req->rq_set_chain, &comp_reqs);
2199 * move completed request at the head of list so it's easier for
2200 * caller to find them
2202 list_splice(&comp_reqs, &set->set_requests);
2204 /* If we hit an error, we want to recover promptly. */
2205 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2207 EXPORT_SYMBOL(ptlrpc_check_set);
2210 * Time out request \a req. is \a async_unlink is set, that means do not wait
2211 * until LNet actually confirms network buffer unlinking.
2212 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2214 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2216 struct obd_import *imp = req->rq_import;
2217 unsigned int debug_mask = D_RPCTRACE;
2221 spin_lock(&req->rq_lock);
2222 req->rq_timedout = 1;
2223 spin_unlock(&req->rq_lock);
2225 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2226 lustre_msg_get_status(req->rq_reqmsg)))
2227 debug_mask = D_WARNING;
2228 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2229 req->rq_net_err ? "failed due to network error" :
2230 ((req->rq_real_sent == 0 ||
2231 req->rq_real_sent < req->rq_sent ||
2232 req->rq_real_sent >= req->rq_deadline) ?
2233 "timed out for sent delay" : "timed out for slow reply"),
2234 (s64)req->rq_sent, (s64)req->rq_real_sent);
2236 if (imp && obd_debug_peer_on_timeout)
2237 LNetDebugPeer(imp->imp_connection->c_peer);
2239 ptlrpc_unregister_reply(req, async_unlink);
2240 ptlrpc_unregister_bulk(req, async_unlink);
2242 if (obd_dump_on_timeout)
2243 libcfs_debug_dumplog();
2246 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2250 atomic_inc(&imp->imp_timeouts);
2252 /* The DLM server doesn't want recovery run on its imports. */
2253 if (imp->imp_dlm_fake)
2257 * If this request is for recovery or other primordial tasks,
2258 * then error it out here.
2260 if (req->rq_ctx_init || req->rq_ctx_fini ||
2261 req->rq_send_state != LUSTRE_IMP_FULL ||
2262 imp->imp_obd->obd_no_recov) {
2263 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2264 ptlrpc_import_state_name(req->rq_send_state),
2265 ptlrpc_import_state_name(imp->imp_state));
2266 spin_lock(&req->rq_lock);
2267 req->rq_status = -ETIMEDOUT;
2269 spin_unlock(&req->rq_lock);
2274 * if a request can't be resent we can't wait for an answer after
2277 if (ptlrpc_no_resend(req)) {
2278 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2282 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2288 * Time out all uncompleted requests in request set pointed by \a data
2289 * Callback used when waiting on sets with l_wait_event.
2292 int ptlrpc_expired_set(void *data)
2294 struct ptlrpc_request_set *set = data;
2295 struct list_head *tmp;
2296 time64_t now = ktime_get_real_seconds();
2299 LASSERT(set != NULL);
2302 * A timeout expired. See which reqs it applies to...
2304 list_for_each(tmp, &set->set_requests) {
2305 struct ptlrpc_request *req =
2306 list_entry(tmp, struct ptlrpc_request,
2309 /* don't expire request waiting for context */
2310 if (req->rq_wait_ctx)
2313 /* Request in-flight? */
2314 if (!((req->rq_phase == RQ_PHASE_RPC &&
2315 !req->rq_waiting && !req->rq_resend) ||
2316 (req->rq_phase == RQ_PHASE_BULK)))
2319 if (req->rq_timedout || /* already dealt with */
2320 req->rq_deadline > now) /* not expired */
2324 * Deal with this guy. Do it asynchronously to not block
2327 ptlrpc_expire_one_request(req, 1);
2331 * When waiting for a whole set, we always break out of the
2332 * sleep so we can recalculate the timeout, or enable interrupts
2333 * if everyone's timed out.
2339 * Sets rq_intr flag in \a req under spinlock.
2341 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2343 spin_lock(&req->rq_lock);
2345 spin_unlock(&req->rq_lock);
2347 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2350 * Interrupts (sets interrupted flag) all uncompleted requests in
2351 * a set \a data. Callback for l_wait_event for interruptible waits.
2353 static void ptlrpc_interrupted_set(void *data)
2355 struct ptlrpc_request_set *set = data;
2356 struct list_head *tmp;
2358 LASSERT(set != NULL);
2359 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2361 list_for_each(tmp, &set->set_requests) {
2362 struct ptlrpc_request *req =
2363 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2368 if (req->rq_phase != RQ_PHASE_RPC &&
2369 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2370 !req->rq_allow_intr)
2373 ptlrpc_mark_interrupted(req);
2378 * Get the smallest timeout in the set; this does NOT set a timeout.
2380 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2382 struct list_head *tmp;
2383 time64_t now = ktime_get_real_seconds();
2385 struct ptlrpc_request *req;
2389 list_for_each(tmp, &set->set_requests) {
2390 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2392 /* Request in-flight? */
2393 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2394 (req->rq_phase == RQ_PHASE_BULK) ||
2395 (req->rq_phase == RQ_PHASE_NEW)))
2398 /* Already timed out. */
2399 if (req->rq_timedout)
2402 /* Waiting for ctx. */
2403 if (req->rq_wait_ctx)
2406 if (req->rq_phase == RQ_PHASE_NEW)
2407 deadline = req->rq_sent;
2408 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2409 deadline = req->rq_sent;
2411 deadline = req->rq_sent + req->rq_timeout;
2413 if (deadline <= now) /* actually expired already */
2414 timeout = 1; /* ASAP */
2415 else if (timeout == 0 || timeout > deadline - now)
2416 timeout = deadline - now;
2422 * Send all unset request from the set and then wait untill all
2423 * requests in the set complete (either get a reply, timeout, get an
2424 * error or otherwise be interrupted).
2425 * Returns 0 on success or error code otherwise.
2427 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2429 struct list_head *tmp;
2430 struct ptlrpc_request *req;
2431 struct l_wait_info lwi;
2436 if (set->set_producer)
2437 (void)ptlrpc_set_producer(set);
2439 list_for_each(tmp, &set->set_requests) {
2440 req = list_entry(tmp, struct ptlrpc_request,
2442 if (req->rq_phase == RQ_PHASE_NEW)
2443 (void)ptlrpc_send_new_req(req);
2446 if (list_empty(&set->set_requests))
2450 timeout = ptlrpc_set_next_timeout(set);
2453 * wait until all complete, interrupted, or an in-flight
2456 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2459 if ((timeout == 0 && !signal_pending(current)) ||
2460 set->set_allow_intr)
2462 * No requests are in-flight (ether timed out
2463 * or delayed), so we can allow interrupts.
2464 * We still want to block for a limited time,
2465 * so we allow interrupts during the timeout.
2467 lwi = LWI_TIMEOUT_INTR_ALL(
2468 cfs_time_seconds(timeout ? timeout : 1),
2470 ptlrpc_interrupted_set, set);
2473 * At least one request is in flight, so no
2474 * interrupts are allowed. Wait until all
2475 * complete, or an in-flight req times out.
2477 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2478 ptlrpc_expired_set, set);
2480 rc = l_wait_event(set->set_waitq,
2481 ptlrpc_check_set(NULL, set), &lwi);
2484 * LU-769 - if we ignored the signal because it was already
2485 * pending when we started, we need to handle it now or we risk
2486 * it being ignored forever
2488 if (rc == -ETIMEDOUT &&
2489 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2490 signal_pending(current)) {
2491 sigset_t blocked_sigs =
2492 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2495 * In fact we only interrupt for the "fatal" signals
2496 * like SIGINT or SIGKILL. We still ignore less
2497 * important signals since ptlrpc set is not easily
2498 * reentrant from userspace again
2500 if (signal_pending(current))
2501 ptlrpc_interrupted_set(set);
2502 cfs_restore_sigs(blocked_sigs);
2505 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2508 * -EINTR => all requests have been flagged rq_intr so next
2510 * -ETIMEDOUT => someone timed out. When all reqs have
2511 * timed out, signals are enabled allowing completion with
2513 * I don't really care if we go once more round the loop in
2514 * the error cases -eeb.
2516 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2517 list_for_each(tmp, &set->set_requests) {
2518 req = list_entry(tmp, struct ptlrpc_request,
2520 spin_lock(&req->rq_lock);
2521 req->rq_invalid_rqset = 1;
2522 spin_unlock(&req->rq_lock);
2525 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2527 LASSERT(atomic_read(&set->set_remaining) == 0);
2529 rc = set->set_rc; /* rq_status of already freed requests if any */
2530 list_for_each(tmp, &set->set_requests) {
2531 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2533 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2534 if (req->rq_status != 0)
2535 rc = req->rq_status;
2540 EXPORT_SYMBOL(ptlrpc_set_wait);
2543 * Helper fuction for request freeing.
2544 * Called when request count reached zero and request needs to be freed.
2545 * Removes request from all sorts of sending/replay lists it might be on,
2546 * frees network buffers if any are present.
2547 * If \a locked is set, that means caller is already holding import imp_lock
2548 * and so we no longer need to reobtain it (for certain lists manipulations)
2550 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2557 LASSERT(!request->rq_srv_req);
2558 LASSERT(request->rq_export == NULL);
2559 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2560 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2561 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2562 LASSERTF(!request->rq_replay, "req %p\n", request);
2564 req_capsule_fini(&request->rq_pill);
2567 * We must take it off the imp_replay_list first. Otherwise, we'll set
2568 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2570 if (request->rq_import) {
2572 spin_lock(&request->rq_import->imp_lock);
2573 list_del_init(&request->rq_replay_list);
2574 list_del_init(&request->rq_unreplied_list);
2576 spin_unlock(&request->rq_import->imp_lock);
2578 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2580 if (atomic_read(&request->rq_refcount) != 0) {
2581 DEBUG_REQ(D_ERROR, request,
2582 "freeing request with nonzero refcount");
2586 if (request->rq_repbuf)
2587 sptlrpc_cli_free_repbuf(request);
2589 if (request->rq_import) {
2590 class_import_put(request->rq_import);
2591 request->rq_import = NULL;
2593 if (request->rq_bulk)
2594 ptlrpc_free_bulk(request->rq_bulk);
2596 if (request->rq_reqbuf || request->rq_clrbuf)
2597 sptlrpc_cli_free_reqbuf(request);
2599 if (request->rq_cli_ctx)
2600 sptlrpc_req_put_ctx(request, !locked);
2602 if (request->rq_pool)
2603 __ptlrpc_free_req_to_pool(request);
2605 ptlrpc_request_cache_free(request);
2609 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2611 * Drop one request reference. Must be called with import imp_lock held.
2612 * When reference count drops to zero, request is freed.
2614 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2616 assert_spin_locked(&request->rq_import->imp_lock);
2617 (void)__ptlrpc_req_finished(request, 1);
2622 * Drops one reference count for request \a request.
2623 * \a locked set indicates that caller holds import imp_lock.
2624 * Frees the request whe reference count reaches zero.
2626 * \retval 1 the request is freed
2627 * \retval 0 some others still hold references on the request
2629 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2637 LASSERT(request != LP_POISON);
2638 LASSERT(request->rq_reqmsg != LP_POISON);
2640 DEBUG_REQ(D_INFO, request, "refcount now %u",
2641 atomic_read(&request->rq_refcount) - 1);
2643 spin_lock(&request->rq_lock);
2644 count = atomic_dec_return(&request->rq_refcount);
2645 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2648 * For open RPC, the client does not know the EA size (LOV, ACL, and
2649 * so on) before replied, then the client has to reserve very large
2650 * reply buffer. Such buffer will not be released until the RPC freed.
2651 * Since The open RPC is replayable, we need to keep it in the replay
2652 * list until close. If there are a lot of files opened concurrently,
2653 * then the client may be OOM.
2655 * If fact, it is unnecessary to keep reply buffer for open replay,
2656 * related EAs have already been saved via mdc_save_lovea() before
2657 * coming here. So it is safe to free the reply buffer some earlier
2658 * before releasing the RPC to avoid client OOM. LU-9514
2660 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2661 spin_lock(&request->rq_early_free_lock);
2662 sptlrpc_cli_free_repbuf(request);
2663 request->rq_repbuf = NULL;
2664 request->rq_repbuf_len = 0;
2665 request->rq_repdata = NULL;
2666 request->rq_reqdata_len = 0;
2667 spin_unlock(&request->rq_early_free_lock);
2669 spin_unlock(&request->rq_lock);
2672 __ptlrpc_free_req(request, locked);
2678 * Drops one reference count for a request.
2680 void ptlrpc_req_finished(struct ptlrpc_request *request)
2682 __ptlrpc_req_finished(request, 0);
2684 EXPORT_SYMBOL(ptlrpc_req_finished);
2687 * Returns xid of a \a request
2689 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2691 return request->rq_xid;
2693 EXPORT_SYMBOL(ptlrpc_req_xid);
2696 * Disengage the client's reply buffer from the network
2697 * NB does _NOT_ unregister any client-side bulk.
2698 * IDEMPOTENT, but _not_ safe against concurrent callers.
2699 * The request owner (i.e. the thread doing the I/O) must call...
2700 * Returns 0 on success or 1 if unregistering cannot be made.
2702 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2705 struct l_wait_info lwi;
2710 LASSERT(!in_interrupt());
2712 /* Let's setup deadline for reply unlink. */
2713 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2714 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2715 request->rq_reply_deadline = ktime_get_real_seconds() +
2719 * Nothing left to do.
2721 if (!ptlrpc_client_recv_or_unlink(request))
2724 LNetMDUnlink(request->rq_reply_md_h);
2727 * Let's check it once again.
2729 if (!ptlrpc_client_recv_or_unlink(request))
2732 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2733 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2736 * Do not wait for unlink to finish.
2742 * We have to l_wait_event() whatever the result, to give liblustre
2743 * a chance to run reply_in_callback(), and to make sure we've
2744 * unlinked before returning a req to the pool.
2747 /* The wq argument is ignored by user-space wait_event macros */
2748 wait_queue_head_t *wq = (request->rq_set) ?
2749 &request->rq_set->set_waitq :
2750 &request->rq_reply_waitq;
2752 * Network access will complete in finite time but the HUGE
2753 * timeout lets us CWARN for visibility of sluggish NALs
2755 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2756 cfs_time_seconds(1), NULL, NULL);
2757 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2760 ptlrpc_rqphase_move(request, request->rq_next_phase);
2764 LASSERT(rc == -ETIMEDOUT);
2765 DEBUG_REQ(D_WARNING, request,
2766 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2767 request->rq_receiving_reply,
2768 request->rq_req_unlinked,
2769 request->rq_reply_unlinked);
2774 static void ptlrpc_free_request(struct ptlrpc_request *req)
2776 spin_lock(&req->rq_lock);
2778 spin_unlock(&req->rq_lock);
2780 if (req->rq_commit_cb)
2781 req->rq_commit_cb(req);
2782 list_del_init(&req->rq_replay_list);
2784 __ptlrpc_req_finished(req, 1);
2788 * the request is committed and dropped from the replay list of its import
2790 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2792 struct obd_import *imp = req->rq_import;
2794 spin_lock(&imp->imp_lock);
2795 if (list_empty(&req->rq_replay_list)) {
2796 spin_unlock(&imp->imp_lock);
2800 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2801 if (imp->imp_replay_cursor == &req->rq_replay_list)
2802 imp->imp_replay_cursor = req->rq_replay_list.next;
2803 ptlrpc_free_request(req);
2806 spin_unlock(&imp->imp_lock);
2808 EXPORT_SYMBOL(ptlrpc_request_committed);
2811 * Iterates through replay_list on import and prunes
2812 * all requests have transno smaller than last_committed for the
2813 * import and don't have rq_replay set.
2814 * Since requests are sorted in transno order, stops when meetign first
2815 * transno bigger than last_committed.
2816 * caller must hold imp->imp_lock
2818 void ptlrpc_free_committed(struct obd_import *imp)
2820 struct ptlrpc_request *req, *saved;
2821 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2822 bool skip_committed_list = true;
2825 LASSERT(imp != NULL);
2826 assert_spin_locked(&imp->imp_lock);
2828 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2829 imp->imp_generation == imp->imp_last_generation_checked) {
2830 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2831 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2834 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2835 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2836 imp->imp_generation);
2838 if (imp->imp_generation != imp->imp_last_generation_checked ||
2839 imp->imp_last_transno_checked == 0)
2840 skip_committed_list = false;
2842 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2843 imp->imp_last_generation_checked = imp->imp_generation;
2845 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2847 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2848 LASSERT(req != last_req);
2851 if (req->rq_transno == 0) {
2852 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2855 if (req->rq_import_generation < imp->imp_generation) {
2856 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2860 /* not yet committed */
2861 if (req->rq_transno > imp->imp_peer_committed_transno) {
2862 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2866 if (req->rq_replay) {
2867 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2868 list_move_tail(&req->rq_replay_list,
2869 &imp->imp_committed_list);
2873 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2874 imp->imp_peer_committed_transno);
2876 ptlrpc_free_request(req);
2879 if (skip_committed_list)
2882 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2884 LASSERT(req->rq_transno != 0);
2885 if (req->rq_import_generation < imp->imp_generation ||
2887 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2888 req->rq_import_generation <
2889 imp->imp_generation ? "stale" : "closed");
2891 if (imp->imp_replay_cursor == &req->rq_replay_list)
2892 imp->imp_replay_cursor =
2893 req->rq_replay_list.next;
2895 ptlrpc_free_request(req);
2902 void ptlrpc_cleanup_client(struct obd_import *imp)
2909 * Schedule previously sent request for resend.
2910 * For bulk requests we assign new xid (to avoid problems with
2911 * lost replies and therefore several transfers landing into same buffer
2912 * from different sending attempts).
2914 void ptlrpc_resend_req(struct ptlrpc_request *req)
2916 DEBUG_REQ(D_HA, req, "going to resend");
2917 spin_lock(&req->rq_lock);
2920 * Request got reply but linked to the import list still.
2921 * Let ptlrpc_check_set() process it.
2923 if (ptlrpc_client_replied(req)) {
2924 spin_unlock(&req->rq_lock);
2925 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2929 req->rq_status = -EAGAIN;
2932 req->rq_net_err = 0;
2933 req->rq_timedout = 0;
2935 ptlrpc_client_wake_req(req);
2936 spin_unlock(&req->rq_lock);
2939 /* XXX: this function and rq_status are currently unused */
2940 void ptlrpc_restart_req(struct ptlrpc_request *req)
2942 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2943 req->rq_status = -ERESTARTSYS;
2945 spin_lock(&req->rq_lock);
2946 req->rq_restart = 1;
2947 req->rq_timedout = 0;
2948 ptlrpc_client_wake_req(req);
2949 spin_unlock(&req->rq_lock);
2953 * Grab additional reference on a request \a req
2955 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2958 atomic_inc(&req->rq_refcount);
2961 EXPORT_SYMBOL(ptlrpc_request_addref);
2964 * Add a request to import replay_list.
2965 * Must be called under imp_lock
2967 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2968 struct obd_import *imp)
2970 struct list_head *tmp;
2972 assert_spin_locked(&imp->imp_lock);
2974 if (req->rq_transno == 0) {
2975 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2980 * clear this for new requests that were resent as well
2981 * as resent replayed requests.
2983 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2985 /* don't re-add requests that have been replayed */
2986 if (!list_empty(&req->rq_replay_list))
2989 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2991 spin_lock(&req->rq_lock);
2993 spin_unlock(&req->rq_lock);
2995 LASSERT(imp->imp_replayable);
2996 /* Balanced in ptlrpc_free_committed, usually. */
2997 ptlrpc_request_addref(req);
2998 list_for_each_prev(tmp, &imp->imp_replay_list) {
2999 struct ptlrpc_request *iter = list_entry(tmp,
3000 struct ptlrpc_request,
3004 * We may have duplicate transnos if we create and then
3005 * open a file, or for closes retained if to match creating
3006 * opens, so use req->rq_xid as a secondary key.
3007 * (See bugs 684, 685, and 428.)
3008 * XXX no longer needed, but all opens need transnos!
3010 if (iter->rq_transno > req->rq_transno)
3013 if (iter->rq_transno == req->rq_transno) {
3014 LASSERT(iter->rq_xid != req->rq_xid);
3015 if (iter->rq_xid > req->rq_xid)
3019 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3023 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3027 * Send request and wait until it completes.
3028 * Returns request processing status.
3030 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3032 struct ptlrpc_request_set *set;
3036 LASSERT(req->rq_set == NULL);
3037 LASSERT(!req->rq_receiving_reply);
3039 set = ptlrpc_prep_set();
3041 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3045 /* for distributed debugging */
3046 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3048 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3049 ptlrpc_request_addref(req);
3050 ptlrpc_set_add_req(set, req);
3051 rc = ptlrpc_set_wait(NULL, set);
3052 ptlrpc_set_destroy(set);
3056 EXPORT_SYMBOL(ptlrpc_queue_wait);
3059 * Callback used for replayed requests reply processing.
3060 * In case of successful reply calls registered request replay callback.
3061 * In case of error restart replay process.
3063 static int ptlrpc_replay_interpret(const struct lu_env *env,
3064 struct ptlrpc_request *req,
3067 struct ptlrpc_replay_async_args *aa = args;
3068 struct obd_import *imp = req->rq_import;
3071 atomic_dec(&imp->imp_replay_inflight);
3074 * Note: if it is bulk replay (MDS-MDS replay), then even if
3075 * server got the request, but bulk transfer timeout, let's
3076 * replay the bulk req again
3078 if (!ptlrpc_client_replied(req) ||
3080 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3081 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
3082 GOTO(out, rc = -ETIMEDOUT);
3085 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3086 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3087 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3088 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3090 /** VBR: check version failure */
3091 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3092 /** replay was failed due to version mismatch */
3093 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
3094 spin_lock(&imp->imp_lock);
3095 imp->imp_vbr_failed = 1;
3096 spin_unlock(&imp->imp_lock);
3097 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3099 /** The transno had better not change over replay. */
3100 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3101 lustre_msg_get_transno(req->rq_repmsg) ||
3102 lustre_msg_get_transno(req->rq_repmsg) == 0,
3104 lustre_msg_get_transno(req->rq_reqmsg),
3105 lustre_msg_get_transno(req->rq_repmsg));
3108 spin_lock(&imp->imp_lock);
3109 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3110 spin_unlock(&imp->imp_lock);
3111 LASSERT(imp->imp_last_replay_transno);
3113 /* transaction number shouldn't be bigger than the latest replayed */
3114 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3115 DEBUG_REQ(D_ERROR, req,
3116 "Reported transno %llu is bigger than the replayed one: %llu",
3118 lustre_msg_get_transno(req->rq_reqmsg));
3119 GOTO(out, rc = -EINVAL);
3122 DEBUG_REQ(D_HA, req, "got rep");
3124 /* let the callback do fixups, possibly including in the request */
3125 if (req->rq_replay_cb)
3126 req->rq_replay_cb(req);
3128 if (ptlrpc_client_replied(req) &&
3129 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3130 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3131 lustre_msg_get_status(req->rq_repmsg),
3132 aa->praa_old_status);
3135 * Note: If the replay fails for MDT-MDT recovery, let's
3136 * abort all of the following requests in the replay
3137 * and sending list, because MDT-MDT update requests
3138 * are dependent on each other, see LU-7039
3140 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3141 struct ptlrpc_request *free_req;
3142 struct ptlrpc_request *tmp;
3144 spin_lock(&imp->imp_lock);
3145 list_for_each_entry_safe(free_req, tmp,
3146 &imp->imp_replay_list,
3148 ptlrpc_free_request(free_req);
3151 list_for_each_entry_safe(free_req, tmp,
3152 &imp->imp_committed_list,
3154 ptlrpc_free_request(free_req);
3157 list_for_each_entry_safe(free_req, tmp,
3158 &imp->imp_delayed_list,
3160 spin_lock(&free_req->rq_lock);
3161 free_req->rq_err = 1;
3162 free_req->rq_status = -EIO;
3163 ptlrpc_client_wake_req(free_req);
3164 spin_unlock(&free_req->rq_lock);
3167 list_for_each_entry_safe(free_req, tmp,
3168 &imp->imp_sending_list,
3170 spin_lock(&free_req->rq_lock);
3171 free_req->rq_err = 1;
3172 free_req->rq_status = -EIO;
3173 ptlrpc_client_wake_req(free_req);
3174 spin_unlock(&free_req->rq_lock);
3176 spin_unlock(&imp->imp_lock);
3179 /* Put it back for re-replay. */
3180 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3184 * Errors while replay can set transno to 0, but
3185 * imp_last_replay_transno shouldn't be set to 0 anyway
3187 if (req->rq_transno == 0)
3188 CERROR("Transno is 0 during replay!\n");
3190 /* continue with recovery */
3191 rc = ptlrpc_import_recovery_state_machine(imp);
3193 req->rq_send_state = aa->praa_old_state;
3196 /* this replay failed, so restart recovery */
3197 ptlrpc_connect_import(imp);
3203 * Prepares and queues request for replay.
3204 * Adds it to ptlrpcd queue for actual sending.
3205 * Returns 0 on success.
3207 int ptlrpc_replay_req(struct ptlrpc_request *req)
3209 struct ptlrpc_replay_async_args *aa;
3213 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3215 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3216 aa = ptlrpc_req_async_args(req);
3217 memset(aa, 0, sizeof(*aa));
3219 /* Prepare request to be resent with ptlrpcd */
3220 aa->praa_old_state = req->rq_send_state;
3221 req->rq_send_state = LUSTRE_IMP_REPLAY;
3222 req->rq_phase = RQ_PHASE_NEW;
3223 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3225 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3227 req->rq_interpret_reply = ptlrpc_replay_interpret;
3228 /* Readjust the timeout for current conditions */
3229 ptlrpc_at_set_req_timeout(req);
3231 /* Tell server net_latency to calculate how long to wait for reply. */
3232 lustre_msg_set_service_time(req->rq_reqmsg,
3233 ptlrpc_at_get_net_latency(req));
3234 DEBUG_REQ(D_HA, req, "REPLAY");
3236 atomic_inc(&req->rq_import->imp_replay_inflight);
3237 spin_lock(&req->rq_lock);
3238 req->rq_early_free_repbuf = 0;
3239 spin_unlock(&req->rq_lock);
3240 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3242 ptlrpcd_add_req(req);
3247 * Aborts all in-flight request on import \a imp sending and delayed lists
3249 void ptlrpc_abort_inflight(struct obd_import *imp)
3251 struct list_head *tmp, *n;
3255 * Make sure that no new requests get processed for this import.
3256 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3257 * this flag and then putting requests on sending_list or delayed_list.
3259 assert_spin_locked(&imp->imp_lock);
3262 * XXX locking? Maybe we should remove each request with the list
3263 * locked? Also, how do we know if the requests on the list are
3264 * being freed at this time?
3266 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3267 struct ptlrpc_request *req = list_entry(tmp,
3268 struct ptlrpc_request,
3271 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3273 spin_lock(&req->rq_lock);
3274 if (req->rq_import_generation < imp->imp_generation) {
3276 req->rq_status = -EIO;
3277 ptlrpc_client_wake_req(req);
3279 spin_unlock(&req->rq_lock);
3282 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3283 struct ptlrpc_request *req =
3284 list_entry(tmp, struct ptlrpc_request, rq_list);
3286 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3288 spin_lock(&req->rq_lock);
3289 if (req->rq_import_generation < imp->imp_generation) {
3291 req->rq_status = -EIO;
3292 ptlrpc_client_wake_req(req);
3294 spin_unlock(&req->rq_lock);
3298 * Last chance to free reqs left on the replay list, but we
3299 * will still leak reqs that haven't committed.
3301 if (imp->imp_replayable)
3302 ptlrpc_free_committed(imp);
3308 * Abort all uncompleted requests in request set \a set
3310 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3312 struct list_head *tmp, *pos;
3314 LASSERT(set != NULL);
3316 list_for_each_safe(pos, tmp, &set->set_requests) {
3317 struct ptlrpc_request *req =
3318 list_entry(pos, struct ptlrpc_request,
3321 spin_lock(&req->rq_lock);
3322 if (req->rq_phase != RQ_PHASE_RPC) {
3323 spin_unlock(&req->rq_lock);
3328 req->rq_status = -EINTR;
3329 ptlrpc_client_wake_req(req);
3330 spin_unlock(&req->rq_lock);
3335 * Initialize the XID for the node. This is common among all requests on
3336 * this node, and only requires the property that it is monotonically
3337 * increasing. It does not need to be sequential. Since this is also used
3338 * as the RDMA match bits, it is important that a single client NOT have
3339 * the same match bits for two different in-flight requests, hence we do
3340 * NOT want to have an XID per target or similar.
3342 * To avoid an unlikely collision between match bits after a client reboot
3343 * (which would deliver old data into the wrong RDMA buffer) initialize
3344 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3345 * If the time is clearly incorrect, we instead use a 62-bit random number.
3346 * In the worst case the random number will overflow 1M RPCs per second in
3347 * 9133 years, or permutations thereof.
3349 #define YEAR_2004 (1ULL << 30)
3350 void ptlrpc_init_xid(void)
3352 time64_t now = ktime_get_real_seconds();
3355 if (now < YEAR_2004) {
3356 get_random_bytes(&xid, sizeof(xid));
3358 xid |= (1ULL << 61);
3360 xid = (u64)now << 20;
3363 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3364 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3365 xid &= PTLRPC_BULK_OPS_MASK;
3366 atomic64_set(&ptlrpc_last_xid, xid);
3370 * Increase xid and returns resulting new value to the caller.
3372 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3373 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3374 * itself uses the last bulk xid needed, so the server can determine the
3375 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3376 * xid must align to a power-of-two value.
3378 * This is assumed to be true due to the initial ptlrpc_last_xid
3379 * value also being initialized to a power-of-two value. LU-1431
3381 __u64 ptlrpc_next_xid(void)
3383 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3387 * If request has a new allocated XID (new request or EINPROGRESS resend),
3388 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3389 * request to ensure previous bulk fails and avoid problems with lost replies
3390 * and therefore several transfers landing into the same buffer from different
3393 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3395 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3397 LASSERT(bd != NULL);
3400 * Generate new matchbits for all resend requests, including
3403 if (req->rq_resend) {
3404 __u64 old_mbits = req->rq_mbits;
3407 * First time resend on -EINPROGRESS will generate new xid,
3408 * so we can actually use the rq_xid as rq_mbits in such case,
3409 * however, it's bit hard to distinguish such resend with a
3410 * 'resend for the -EINPROGRESS resend'. To make it simple,
3411 * we opt to generate mbits for all resend cases.
3413 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3415 req->rq_mbits = ptlrpc_next_xid();
3418 * Old version transfers rq_xid to peer as
3421 spin_lock(&req->rq_import->imp_lock);
3422 list_del_init(&req->rq_unreplied_list);
3423 ptlrpc_assign_next_xid_nolock(req);
3424 spin_unlock(&req->rq_import->imp_lock);
3425 req->rq_mbits = req->rq_xid;
3427 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3428 old_mbits, req->rq_mbits);
3429 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3430 /* Request being sent first time, use xid as matchbits. */
3431 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3432 || req->rq_mbits == 0) {
3433 req->rq_mbits = req->rq_xid;
3435 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3437 req->rq_mbits -= total_md - 1;
3441 * Replay request, xid and matchbits have already been
3442 * correctly assigned.
3448 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3449 * that server can infer the number of bulks that were prepared,
3452 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3456 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3457 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3459 * It's ok to directly set the rq_xid here, since this xid bump
3460 * won't affect the request position in unreplied list.
3462 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3463 req->rq_xid = req->rq_mbits;
3467 * Get a glimpse at what next xid value might have been.
3468 * Returns possible next xid.
3470 __u64 ptlrpc_sample_next_xid(void)
3472 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3474 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3477 * Functions for operating ptlrpc workers.
3479 * A ptlrpc work is a function which will be running inside ptlrpc context.
3480 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3482 * 1. after a work is created, it can be used many times, that is:
3483 * handler = ptlrpcd_alloc_work();
3484 * ptlrpcd_queue_work();
3486 * queue it again when necessary:
3487 * ptlrpcd_queue_work();
3488 * ptlrpcd_destroy_work();
3489 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3490 * it will only be queued once in any time. Also as its name implies, it may
3491 * have delay before it really runs by ptlrpcd thread.
3493 struct ptlrpc_work_async_args {
3494 int (*cb)(const struct lu_env *, void *);
3498 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3500 /* re-initialize the req */
3501 req->rq_timeout = obd_timeout;
3502 req->rq_sent = ktime_get_real_seconds();
3503 req->rq_deadline = req->rq_sent + req->rq_timeout;
3504 req->rq_phase = RQ_PHASE_INTERPRET;
3505 req->rq_next_phase = RQ_PHASE_COMPLETE;
3506 req->rq_xid = ptlrpc_next_xid();
3507 req->rq_import_generation = req->rq_import->imp_generation;
3509 ptlrpcd_add_req(req);
3512 static int work_interpreter(const struct lu_env *env,
3513 struct ptlrpc_request *req, void *args, int rc)
3515 struct ptlrpc_work_async_args *arg = args;
3517 LASSERT(ptlrpcd_check_work(req));
3518 LASSERT(arg->cb != NULL);
3520 rc = arg->cb(env, arg->cbdata);
3522 list_del_init(&req->rq_set_chain);
3525 if (atomic_dec_return(&req->rq_refcount) > 1) {
3526 atomic_set(&req->rq_refcount, 2);
3527 ptlrpcd_add_work_req(req);
3532 static int worker_format;
3534 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3536 return req->rq_pill.rc_fmt == (void *)&worker_format;
3540 * Create a work for ptlrpc.
3542 void *ptlrpcd_alloc_work(struct obd_import *imp,
3543 int (*cb)(const struct lu_env *, void *), void *cbdata)
3545 struct ptlrpc_request *req = NULL;
3546 struct ptlrpc_work_async_args *args;
3552 RETURN(ERR_PTR(-EINVAL));
3554 /* copy some code from deprecated fakereq. */
3555 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3557 CERROR("ptlrpc: run out of memory!\n");
3558 RETURN(ERR_PTR(-ENOMEM));
3561 ptlrpc_cli_req_init(req);
3563 req->rq_send_state = LUSTRE_IMP_FULL;
3564 req->rq_type = PTL_RPC_MSG_REQUEST;
3565 req->rq_import = class_import_get(imp);
3566 req->rq_interpret_reply = work_interpreter;
3567 /* don't want reply */
3568 req->rq_no_delay = req->rq_no_resend = 1;
3569 req->rq_pill.rc_fmt = (void *)&worker_format;
3571 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3572 args = ptlrpc_req_async_args(req);
3574 args->cbdata = cbdata;
3578 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3580 void ptlrpcd_destroy_work(void *handler)
3582 struct ptlrpc_request *req = handler;
3585 ptlrpc_req_finished(req);
3587 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3589 int ptlrpcd_queue_work(void *handler)
3591 struct ptlrpc_request *req = handler;
3594 * Check if the req is already being queued.
3596 * Here comes a trick: it lacks a way of checking if a req is being
3597 * processed reliably in ptlrpc. Here I have to use refcount of req
3598 * for this purpose. This is okay because the caller should use this
3599 * req as opaque data. - Jinshan
3601 LASSERT(atomic_read(&req->rq_refcount) > 0);
3602 if (atomic_inc_return(&req->rq_refcount) == 2)
3603 ptlrpcd_add_work_req(req);
3606 EXPORT_SYMBOL(ptlrpcd_queue_work);