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: rc = %d",
446 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
448 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
456 * Handle an early reply message, called with the rq_lock held.
457 * If anything goes wrong just ignore it - same as if it never happened
459 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
460 __must_hold(&req->rq_lock)
462 struct ptlrpc_request *early_req;
468 spin_unlock(&req->rq_lock);
470 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
472 spin_lock(&req->rq_lock);
476 rc = unpack_reply(early_req);
478 sptlrpc_cli_finish_early_reply(early_req);
479 spin_lock(&req->rq_lock);
484 * Use new timeout value just to adjust the local value for this
485 * request, don't include it into at_history. It is unclear yet why
486 * service time increased and should it be counted or skipped, e.g.
487 * that can be recovery case or some error or server, the real reply
488 * will add all new data if it is worth to add.
490 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
491 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
493 /* Network latency can be adjusted, it is pure network delays */
494 ptlrpc_at_adj_net_latency(req,
495 lustre_msg_get_service_time(early_req->rq_repmsg));
497 sptlrpc_cli_finish_early_reply(early_req);
499 spin_lock(&req->rq_lock);
500 olddl = req->rq_deadline;
502 * server assumes it now has rq_timeout from when the request
503 * arrived, so the client should give it at least that long.
504 * since we don't know the arrival time we'll use the original
507 req->rq_deadline = req->rq_sent + req->rq_timeout +
508 ptlrpc_at_get_net_latency(req);
510 /* The below message is checked in replay-single.sh test_65{a,b} */
511 /* The below message is checked in sanity-{gss,krb5} test_8 */
512 DEBUG_REQ(D_ADAPTTO, req,
513 "Early reply #%d, new deadline in %llds (%llds)",
515 req->rq_deadline - ktime_get_real_seconds(),
516 req->rq_deadline - olddl);
521 static struct kmem_cache *request_cache;
523 int ptlrpc_request_cache_init(void)
525 request_cache = kmem_cache_create("ptlrpc_cache",
526 sizeof(struct ptlrpc_request),
527 0, SLAB_HWCACHE_ALIGN, NULL);
528 return request_cache ? 0 : -ENOMEM;
531 void ptlrpc_request_cache_fini(void)
533 kmem_cache_destroy(request_cache);
536 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
538 struct ptlrpc_request *req;
540 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
544 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
546 OBD_SLAB_FREE_PTR(req, request_cache);
550 * Wind down request pool \a pool.
551 * Frees all requests from the pool too
553 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
555 struct list_head *l, *tmp;
556 struct ptlrpc_request *req;
558 LASSERT(pool != NULL);
560 spin_lock(&pool->prp_lock);
561 list_for_each_safe(l, tmp, &pool->prp_req_list) {
562 req = list_entry(l, struct ptlrpc_request, rq_list);
563 list_del(&req->rq_list);
564 LASSERT(req->rq_reqbuf);
565 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
566 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
567 ptlrpc_request_cache_free(req);
569 spin_unlock(&pool->prp_lock);
570 OBD_FREE(pool, sizeof(*pool));
572 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
575 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
577 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
582 while (size < pool->prp_rq_size)
585 LASSERTF(list_empty(&pool->prp_req_list) ||
586 size == pool->prp_rq_size,
587 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
588 pool->prp_rq_size, size);
590 pool->prp_rq_size = size;
591 for (i = 0; i < num_rq; i++) {
592 struct ptlrpc_request *req;
593 struct lustre_msg *msg;
595 req = ptlrpc_request_cache_alloc(GFP_NOFS);
598 OBD_ALLOC_LARGE(msg, size);
600 ptlrpc_request_cache_free(req);
603 req->rq_reqbuf = msg;
604 req->rq_reqbuf_len = size;
606 spin_lock(&pool->prp_lock);
607 list_add_tail(&req->rq_list, &pool->prp_req_list);
608 spin_unlock(&pool->prp_lock);
612 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
615 * Create and initialize new request pool with given attributes:
616 * \a num_rq - initial number of requests to create for the pool
617 * \a msgsize - maximum message size possible for requests in thid pool
618 * \a populate_pool - function to be called when more requests need to be added
620 * Returns pointer to newly created pool or NULL on error.
622 struct ptlrpc_request_pool *
623 ptlrpc_init_rq_pool(int num_rq, int msgsize,
624 int (*populate_pool)(struct ptlrpc_request_pool *, int))
626 struct ptlrpc_request_pool *pool;
628 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
633 * Request next power of two for the allocation, because internally
634 * kernel would do exactly this
636 spin_lock_init(&pool->prp_lock);
637 INIT_LIST_HEAD(&pool->prp_req_list);
638 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
639 pool->prp_populate = populate_pool;
641 populate_pool(pool, num_rq);
645 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
648 * Fetches one request from pool \a pool
650 static struct ptlrpc_request *
651 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
653 struct ptlrpc_request *request;
654 struct lustre_msg *reqbuf;
659 spin_lock(&pool->prp_lock);
662 * See if we have anything in a pool, and bail out if nothing,
663 * in writeout path, where this matters, this is safe to do, because
664 * nothing is lost in this case, and when some in-flight requests
665 * complete, this code will be called again.
667 if (unlikely(list_empty(&pool->prp_req_list))) {
668 spin_unlock(&pool->prp_lock);
672 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
674 list_del_init(&request->rq_list);
675 spin_unlock(&pool->prp_lock);
677 LASSERT(request->rq_reqbuf);
678 LASSERT(request->rq_pool);
680 reqbuf = request->rq_reqbuf;
681 memset(request, 0, sizeof(*request));
682 request->rq_reqbuf = reqbuf;
683 request->rq_reqbuf_len = pool->prp_rq_size;
684 request->rq_pool = pool;
690 * Returns freed \a request to pool.
692 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
694 struct ptlrpc_request_pool *pool = request->rq_pool;
696 spin_lock(&pool->prp_lock);
697 LASSERT(list_empty(&request->rq_list));
698 LASSERT(!request->rq_receiving_reply);
699 list_add_tail(&request->rq_list, &pool->prp_req_list);
700 spin_unlock(&pool->prp_lock);
703 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
705 struct obd_import *imp = req->rq_import;
706 struct list_head *tmp;
707 struct ptlrpc_request *iter;
709 assert_spin_locked(&imp->imp_lock);
710 LASSERT(list_empty(&req->rq_unreplied_list));
712 /* unreplied list is sorted by xid in ascending order */
713 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
714 iter = list_entry(tmp, struct ptlrpc_request,
717 LASSERT(req->rq_xid != iter->rq_xid);
718 if (req->rq_xid < iter->rq_xid)
720 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
723 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
726 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
728 req->rq_xid = ptlrpc_next_xid();
729 ptlrpc_add_unreplied(req);
732 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
734 spin_lock(&req->rq_import->imp_lock);
735 ptlrpc_assign_next_xid_nolock(req);
736 spin_unlock(&req->rq_import->imp_lock);
739 static atomic64_t ptlrpc_last_xid;
741 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
742 __u32 version, int opcode, char **bufs,
743 struct ptlrpc_cli_ctx *ctx)
746 struct obd_import *imp;
752 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
753 imp = request->rq_import;
754 lengths = request->rq_pill.rc_area[RCL_CLIENT];
757 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
759 rc = sptlrpc_req_get_ctx(request);
763 sptlrpc_req_set_flavor(request, opcode);
765 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
770 lustre_msg_add_version(request->rq_reqmsg, version);
771 request->rq_send_state = LUSTRE_IMP_FULL;
772 request->rq_type = PTL_RPC_MSG_REQUEST;
774 request->rq_req_cbid.cbid_fn = request_out_callback;
775 request->rq_req_cbid.cbid_arg = request;
777 request->rq_reply_cbid.cbid_fn = reply_in_callback;
778 request->rq_reply_cbid.cbid_arg = request;
780 request->rq_reply_deadline = 0;
781 request->rq_bulk_deadline = 0;
782 request->rq_req_deadline = 0;
783 request->rq_phase = RQ_PHASE_NEW;
784 request->rq_next_phase = RQ_PHASE_UNDEFINED;
786 request->rq_request_portal = imp->imp_client->cli_request_portal;
787 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
789 ptlrpc_at_set_req_timeout(request);
791 lustre_msg_set_opc(request->rq_reqmsg, opcode);
793 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
794 if (cfs_fail_val == opcode) {
795 time64_t *fail_t = NULL, *fail2_t = NULL;
797 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
798 fail_t = &request->rq_bulk_deadline;
799 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
800 fail_t = &request->rq_reply_deadline;
801 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
802 fail_t = &request->rq_req_deadline;
803 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
804 fail_t = &request->rq_reply_deadline;
805 fail2_t = &request->rq_bulk_deadline;
806 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
807 time64_t now = ktime_get_real_seconds();
808 u64 xid = ((u64)now >> 4) << 24;
810 atomic64_set(&ptlrpc_last_xid, xid);
814 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
817 *fail2_t = ktime_get_real_seconds() +
821 * The RPC is infected, let the test to change the
824 msleep(4 * MSEC_PER_SEC);
827 ptlrpc_assign_next_xid(request);
832 LASSERT(!request->rq_pool);
833 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
835 class_import_put(imp);
839 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
842 * Pack request buffers for network transfer, performing necessary encryption
843 * steps if necessary.
845 int ptlrpc_request_pack(struct ptlrpc_request *request,
846 __u32 version, int opcode)
848 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
850 EXPORT_SYMBOL(ptlrpc_request_pack);
853 * Helper function to allocate new request on import \a imp
854 * and possibly using existing request from pool \a pool if provided.
855 * Returns allocated request structure with import field filled or
859 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
860 struct ptlrpc_request_pool *pool)
862 struct ptlrpc_request *request = NULL;
864 request = ptlrpc_request_cache_alloc(GFP_NOFS);
866 if (!request && pool)
867 request = ptlrpc_prep_req_from_pool(pool);
870 ptlrpc_cli_req_init(request);
872 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
873 LASSERT(imp != LP_POISON);
874 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
876 LASSERT(imp->imp_client != LP_POISON);
878 request->rq_import = class_import_get(imp);
880 CERROR("request allocation out of memory\n");
887 * Helper function for creating a request.
888 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
889 * buffer structures according to capsule template \a format.
890 * Returns allocated request structure pointer or NULL on error.
892 static struct ptlrpc_request *
893 ptlrpc_request_alloc_internal(struct obd_import *imp,
894 struct ptlrpc_request_pool *pool,
895 const struct req_format *format)
897 struct ptlrpc_request *request;
900 request = __ptlrpc_request_alloc(imp, pool);
905 * initiate connection if needed when the import has been
906 * referenced by the new request to avoid races with disconnect
908 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
911 CDEBUG_LIMIT(imp->imp_idle_debug,
912 "%s: reconnect after %llds idle\n",
913 imp->imp_obd->obd_name, ktime_get_real_seconds() -
914 imp->imp_last_reply_time);
915 spin_lock(&imp->imp_lock);
916 if (imp->imp_state == LUSTRE_IMP_IDLE) {
917 imp->imp_generation++;
918 imp->imp_initiated_at = imp->imp_generation;
919 imp->imp_state = LUSTRE_IMP_NEW;
922 spin_unlock(&imp->imp_lock);
924 rc = ptlrpc_connect_import(imp);
926 ptlrpc_request_free(request);
929 ptlrpc_pinger_add_import(imp);
933 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
934 req_capsule_set(&request->rq_pill, format);
939 * Allocate new request structure for import \a imp and initialize its
940 * buffer structure according to capsule template \a format.
942 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
943 const struct req_format *format)
945 return ptlrpc_request_alloc_internal(imp, NULL, format);
947 EXPORT_SYMBOL(ptlrpc_request_alloc);
950 * Allocate new request structure for import \a imp from pool \a pool and
951 * initialize its buffer structure according to capsule template \a format.
953 struct ptlrpc_request *
954 ptlrpc_request_alloc_pool(struct obd_import *imp,
955 struct ptlrpc_request_pool *pool,
956 const struct req_format *format)
958 return ptlrpc_request_alloc_internal(imp, pool, format);
960 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
963 * For requests not from pool, free memory of the request structure.
964 * For requests obtained from a pool earlier, return request back to pool.
966 void ptlrpc_request_free(struct ptlrpc_request *request)
968 if (request->rq_pool)
969 __ptlrpc_free_req_to_pool(request);
971 ptlrpc_request_cache_free(request);
973 EXPORT_SYMBOL(ptlrpc_request_free);
976 * Allocate new request for operatione \a opcode and immediatelly pack it for
978 * Only used for simple requests like OBD_PING where the only important
979 * part of the request is operation itself.
980 * Returns allocated request or NULL on error.
982 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
983 const struct req_format *format,
984 __u32 version, int opcode)
986 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
990 rc = ptlrpc_request_pack(req, version, opcode);
992 ptlrpc_request_free(req);
998 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1001 * Allocate and initialize new request set structure on the current CPT.
1002 * Returns a pointer to the newly allocated set structure or NULL on error.
1004 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1006 struct ptlrpc_request_set *set;
1010 cpt = cfs_cpt_current(cfs_cpt_table, 0);
1011 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof(*set));
1014 atomic_set(&set->set_refcount, 1);
1015 INIT_LIST_HEAD(&set->set_requests);
1016 init_waitqueue_head(&set->set_waitq);
1017 atomic_set(&set->set_new_count, 0);
1018 atomic_set(&set->set_remaining, 0);
1019 spin_lock_init(&set->set_new_req_lock);
1020 INIT_LIST_HEAD(&set->set_new_requests);
1021 set->set_max_inflight = UINT_MAX;
1022 set->set_producer = NULL;
1023 set->set_producer_arg = NULL;
1028 EXPORT_SYMBOL(ptlrpc_prep_set);
1031 * Allocate and initialize new request set structure with flow control
1032 * extension. This extension allows to control the number of requests in-flight
1033 * for the whole set. A callback function to generate requests must be provided
1034 * and the request set will keep the number of requests sent over the wire to
1036 * Returns a pointer to the newly allocated set structure or NULL on error.
1038 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1042 struct ptlrpc_request_set *set;
1044 set = ptlrpc_prep_set();
1048 set->set_max_inflight = max;
1049 set->set_producer = func;
1050 set->set_producer_arg = arg;
1056 * Wind down and free request set structure previously allocated with
1058 * Ensures that all requests on the set have completed and removes
1059 * all requests from the request list in a set.
1060 * If any unsent request happen to be on the list, pretends that they got
1061 * an error in flight and calls their completion handler.
1063 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1065 struct list_head *tmp;
1066 struct list_head *next;
1072 /* Requests on the set should either all be completed, or all be new */
1073 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1074 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1075 list_for_each(tmp, &set->set_requests) {
1076 struct ptlrpc_request *req =
1077 list_entry(tmp, struct ptlrpc_request,
1080 LASSERT(req->rq_phase == expected_phase);
1084 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1085 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1086 atomic_read(&set->set_remaining), n);
1088 list_for_each_safe(tmp, next, &set->set_requests) {
1089 struct ptlrpc_request *req =
1090 list_entry(tmp, struct ptlrpc_request,
1092 list_del_init(&req->rq_set_chain);
1094 LASSERT(req->rq_phase == expected_phase);
1096 if (req->rq_phase == RQ_PHASE_NEW) {
1097 ptlrpc_req_interpret(NULL, req, -EBADR);
1098 atomic_dec(&set->set_remaining);
1101 spin_lock(&req->rq_lock);
1103 req->rq_invalid_rqset = 0;
1104 spin_unlock(&req->rq_lock);
1106 ptlrpc_req_finished(req);
1109 LASSERT(atomic_read(&set->set_remaining) == 0);
1111 ptlrpc_reqset_put(set);
1114 EXPORT_SYMBOL(ptlrpc_set_destroy);
1117 * Add a new request to the general purpose request set.
1118 * Assumes request reference from the caller.
1120 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1121 struct ptlrpc_request *req)
1123 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1124 LASSERT(list_empty(&req->rq_set_chain));
1126 if (req->rq_allow_intr)
1127 set->set_allow_intr = 1;
1129 /* The set takes over the caller's request reference */
1130 list_add_tail(&req->rq_set_chain, &set->set_requests);
1132 atomic_inc(&set->set_remaining);
1133 req->rq_queued_time = ktime_get_seconds();
1136 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1138 if (set->set_producer)
1140 * If the request set has a producer callback, the RPC must be
1141 * sent straight away
1143 ptlrpc_send_new_req(req);
1145 EXPORT_SYMBOL(ptlrpc_set_add_req);
1148 * Add a request to a request with dedicated server thread
1149 * and wake the thread to make any necessary processing.
1150 * Currently only used for ptlrpcd.
1152 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1153 struct ptlrpc_request *req)
1155 struct ptlrpc_request_set *set = pc->pc_set;
1158 LASSERT(req->rq_set == NULL);
1159 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1161 spin_lock(&set->set_new_req_lock);
1163 * The set takes over the caller's request reference.
1166 req->rq_queued_time = ktime_get_seconds();
1167 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1168 count = atomic_inc_return(&set->set_new_count);
1169 spin_unlock(&set->set_new_req_lock);
1171 /* Only need to call wakeup once for the first entry. */
1173 wake_up(&set->set_waitq);
1176 * XXX: It maybe unnecessary to wakeup all the partners. But to
1177 * guarantee the async RPC can be processed ASAP, we have
1178 * no other better choice. It maybe fixed in future.
1180 for (i = 0; i < pc->pc_npartners; i++)
1181 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1186 * Based on the current state of the import, determine if the request
1187 * can be sent, is an error, or should be delayed.
1189 * Returns true if this request should be delayed. If false, and
1190 * *status is set, then the request can not be sent and *status is the
1191 * error code. If false and status is 0, then request can be sent.
1193 * The imp->imp_lock must be held.
1195 static int ptlrpc_import_delay_req(struct obd_import *imp,
1196 struct ptlrpc_request *req, int *status)
1204 if (req->rq_ctx_init || req->rq_ctx_fini) {
1205 /* always allow ctx init/fini rpc go through */
1206 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1207 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1209 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1210 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1213 * pings or MDS-equivalent STATFS may safely
1216 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1217 D_HA : D_ERROR, req, "IMP_CLOSED");
1219 } else if (ptlrpc_send_limit_expired(req)) {
1220 /* probably doesn't need to be a D_ERROR afterinitial testing */
1221 DEBUG_REQ(D_HA, req, "send limit expired");
1222 *status = -ETIMEDOUT;
1223 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1224 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1225 ;/* allow CONNECT even if import is invalid */
1226 if (atomic_read(&imp->imp_inval_count) != 0) {
1227 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1230 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1231 if (!imp->imp_deactive)
1232 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1233 *status = -ESHUTDOWN; /* b=12940 */
1234 } else if (req->rq_import_generation != imp->imp_generation) {
1235 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1237 } else if (req->rq_send_state != imp->imp_state) {
1238 /* invalidate in progress - any requests should be drop */
1239 if (atomic_read(&imp->imp_inval_count) != 0) {
1240 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1242 } else if (req->rq_no_delay &&
1243 imp->imp_generation != imp->imp_initiated_at) {
1244 /* ignore nodelay for requests initiating connections */
1245 *status = -EWOULDBLOCK;
1246 } else if (req->rq_allow_replay &&
1247 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1248 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1249 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1250 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1251 DEBUG_REQ(D_HA, req, "allow during recovery");
1261 * Decide if the error message should be printed to the console or not.
1262 * Makes its decision based on request type, status, and failure frequency.
1264 * \param[in] req request that failed and may need a console message
1266 * \retval false if no message should be printed
1267 * \retval true if console message should be printed
1269 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1271 LASSERT(req->rq_reqmsg != NULL);
1273 /* Suppress particular reconnect errors which are to be expected. */
1274 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1275 /* Suppress timed out reconnect requests */
1276 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1281 * Suppress most unavailable/again reconnect requests, but
1282 * print occasionally so it is clear client is trying to
1283 * connect to a server where no target is running.
1285 if ((err == -ENODEV || err == -EAGAIN) &&
1286 req->rq_import->imp_conn_cnt % 30 != 20)
1290 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1291 /* -EAGAIN is normal when using POSIX flocks */
1294 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1295 (req->rq_xid & 0xf) != 10)
1296 /* Suppress most ping requests, they may fail occasionally */
1303 * Check request processing status.
1304 * Returns the status.
1306 static int ptlrpc_check_status(struct ptlrpc_request *req)
1311 rc = lustre_msg_get_status(req->rq_repmsg);
1312 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1313 struct obd_import *imp = req->rq_import;
1314 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1315 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1317 if (ptlrpc_console_allow(req, opc, rc))
1318 LCONSOLE_ERROR_MSG(0x11,
1319 "%s: operation %s to node %s failed: rc = %d\n",
1320 imp->imp_obd->obd_name,
1322 libcfs_nid2str(nid), rc);
1323 RETURN(rc < 0 ? rc : -EINVAL);
1327 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1333 * save pre-versions of objects into request for replay.
1334 * Versions are obtained from server reply.
1337 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1339 struct lustre_msg *repmsg = req->rq_repmsg;
1340 struct lustre_msg *reqmsg = req->rq_reqmsg;
1341 __u64 *versions = lustre_msg_get_versions(repmsg);
1344 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1348 lustre_msg_set_versions(reqmsg, versions);
1349 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1350 versions[0], versions[1]);
1355 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1357 struct ptlrpc_request *req;
1359 assert_spin_locked(&imp->imp_lock);
1360 if (list_empty(&imp->imp_unreplied_list))
1363 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1365 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1367 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1368 imp->imp_known_replied_xid = req->rq_xid - 1;
1370 return req->rq_xid - 1;
1374 * Callback function called when client receives RPC reply for \a req.
1375 * Returns 0 on success or error code.
1376 * The return alue would be assigned to req->rq_status by the caller
1377 * as request processing status.
1378 * This function also decides if the request needs to be saved for later replay.
1380 static int after_reply(struct ptlrpc_request *req)
1382 struct obd_import *imp = req->rq_import;
1383 struct obd_device *obd = req->rq_import->imp_obd;
1390 LASSERT(obd != NULL);
1391 /* repbuf must be unlinked */
1392 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1394 if (req->rq_reply_truncated) {
1395 if (ptlrpc_no_resend(req)) {
1396 DEBUG_REQ(D_ERROR, req,
1397 "reply buffer overflow, expected=%d, actual size=%d",
1398 req->rq_nob_received, req->rq_repbuf_len);
1402 sptlrpc_cli_free_repbuf(req);
1404 * Pass the required reply buffer size (include
1405 * space for early reply).
1406 * NB: no need to roundup because alloc_repbuf
1409 req->rq_replen = req->rq_nob_received;
1410 req->rq_nob_received = 0;
1411 spin_lock(&req->rq_lock);
1413 spin_unlock(&req->rq_lock);
1417 work_start = ktime_get_real();
1418 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1421 * NB Until this point, the whole of the incoming message,
1422 * including buflens, status etc is in the sender's byte order.
1424 rc = sptlrpc_cli_unwrap_reply(req);
1426 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1431 * Security layer unwrap might ask resend this request.
1436 rc = unpack_reply(req);
1440 /* retry indefinitely on EINPROGRESS */
1441 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1442 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1443 time64_t now = ktime_get_real_seconds();
1445 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1446 D_RPCTRACE, req, "resending request on EINPROGRESS");
1447 spin_lock(&req->rq_lock);
1449 spin_unlock(&req->rq_lock);
1450 req->rq_nr_resend++;
1452 /* Readjust the timeout for current conditions */
1453 ptlrpc_at_set_req_timeout(req);
1455 * delay resend to give a chance to the server to get ready.
1456 * The delay is increased by 1s on every resend and is capped to
1457 * the current request timeout (i.e. obd_timeout if AT is off,
1458 * or AT service time x 125% + 5s, see at_est2timeout)
1460 if (req->rq_nr_resend > req->rq_timeout)
1461 req->rq_sent = now + req->rq_timeout;
1463 req->rq_sent = now + req->rq_nr_resend;
1465 /* Resend for EINPROGRESS will use a new XID */
1466 spin_lock(&imp->imp_lock);
1467 list_del_init(&req->rq_unreplied_list);
1468 spin_unlock(&imp->imp_lock);
1473 if (obd->obd_svc_stats) {
1474 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1476 ptlrpc_lprocfs_rpc_sent(req, timediff);
1479 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1480 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1481 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1482 lustre_msg_get_type(req->rq_repmsg));
1486 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1487 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1488 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1489 ptlrpc_at_adj_net_latency(req,
1490 lustre_msg_get_service_time(req->rq_repmsg));
1492 rc = ptlrpc_check_status(req);
1496 * Either we've been evicted, or the server has failed for
1497 * some reason. Try to reconnect, and if that fails, punt to
1500 if (ptlrpc_recoverable_error(rc)) {
1501 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1502 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1505 ptlrpc_request_handle_notconn(req);
1510 * Let's look if server sent slv. Do it only for RPC with
1513 ldlm_cli_update_pool(req);
1517 * Store transno in reqmsg for replay.
1519 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1520 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1521 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1524 if (imp->imp_replayable) {
1525 spin_lock(&imp->imp_lock);
1527 * No point in adding already-committed requests to the replay
1528 * list, we will just remove them immediately. b=9829
1530 if (req->rq_transno != 0 &&
1532 lustre_msg_get_last_committed(req->rq_repmsg) ||
1534 /** version recovery */
1535 ptlrpc_save_versions(req);
1536 ptlrpc_retain_replayable_request(req, imp);
1537 } else if (req->rq_commit_cb &&
1538 list_empty(&req->rq_replay_list)) {
1540 * NB: don't call rq_commit_cb if it's already on
1541 * rq_replay_list, ptlrpc_free_committed() will call
1542 * it later, see LU-3618 for details
1544 spin_unlock(&imp->imp_lock);
1545 req->rq_commit_cb(req);
1546 spin_lock(&imp->imp_lock);
1550 * Replay-enabled imports return commit-status information.
1552 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1553 if (likely(committed > imp->imp_peer_committed_transno))
1554 imp->imp_peer_committed_transno = committed;
1556 ptlrpc_free_committed(imp);
1558 if (!list_empty(&imp->imp_replay_list)) {
1559 struct ptlrpc_request *last;
1561 last = list_entry(imp->imp_replay_list.prev,
1562 struct ptlrpc_request,
1565 * Requests with rq_replay stay on the list even if no
1566 * commit is expected.
1568 if (last->rq_transno > imp->imp_peer_committed_transno)
1569 ptlrpc_pinger_commit_expected(imp);
1572 spin_unlock(&imp->imp_lock);
1579 * Helper function to send request \a req over the network for the first time
1580 * Also adjusts request phase.
1581 * Returns 0 on success or error code.
1583 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1585 struct obd_import *imp = req->rq_import;
1590 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1592 /* do not try to go further if there is not enough memory in enc_pool */
1593 if (req->rq_sent && req->rq_bulk)
1594 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1595 pool_is_at_full_capacity())
1598 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1599 (!req->rq_generation_set ||
1600 req->rq_import_generation == imp->imp_generation))
1603 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1605 spin_lock(&imp->imp_lock);
1607 LASSERT(req->rq_xid != 0);
1608 LASSERT(!list_empty(&req->rq_unreplied_list));
1610 if (!req->rq_generation_set)
1611 req->rq_import_generation = imp->imp_generation;
1613 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1614 spin_lock(&req->rq_lock);
1615 req->rq_waiting = 1;
1616 spin_unlock(&req->rq_lock);
1618 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1619 ptlrpc_import_state_name(req->rq_send_state),
1620 ptlrpc_import_state_name(imp->imp_state));
1621 LASSERT(list_empty(&req->rq_list));
1622 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1623 atomic_inc(&req->rq_import->imp_inflight);
1624 spin_unlock(&imp->imp_lock);
1629 spin_unlock(&imp->imp_lock);
1630 req->rq_status = rc;
1631 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1635 LASSERT(list_empty(&req->rq_list));
1636 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1637 atomic_inc(&req->rq_import->imp_inflight);
1640 * find the known replied XID from the unreplied list, CONNECT
1641 * and DISCONNECT requests are skipped to make the sanity check
1642 * on server side happy. see process_req_last_xid().
1644 * For CONNECT: Because replay requests have lower XID, it'll
1645 * break the sanity check if CONNECT bump the exp_last_xid on
1648 * For DISCONNECT: Since client will abort inflight RPC before
1649 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1650 * than the inflight RPC.
1652 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1653 min_xid = ptlrpc_known_replied_xid(imp);
1654 spin_unlock(&imp->imp_lock);
1656 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1658 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1660 rc = sptlrpc_req_refresh_ctx(req, -1);
1663 req->rq_status = rc;
1666 spin_lock(&req->rq_lock);
1667 req->rq_wait_ctx = 1;
1668 spin_unlock(&req->rq_lock);
1674 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1675 req, current_comm(),
1676 imp->imp_obd->obd_uuid.uuid,
1677 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1678 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1679 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1681 rc = ptl_send_rpc(req, 0);
1682 if (rc == -ENOMEM) {
1683 spin_lock(&imp->imp_lock);
1684 if (!list_empty(&req->rq_list)) {
1685 list_del_init(&req->rq_list);
1686 atomic_dec(&req->rq_import->imp_inflight);
1688 spin_unlock(&imp->imp_lock);
1689 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1693 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1695 spin_lock(&req->rq_lock);
1696 req->rq_net_err = 1;
1697 spin_unlock(&req->rq_lock);
1703 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1708 LASSERT(set->set_producer != NULL);
1710 remaining = atomic_read(&set->set_remaining);
1713 * populate the ->set_requests list with requests until we
1714 * reach the maximum number of RPCs in flight for this set
1716 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1717 rc = set->set_producer(set, set->set_producer_arg);
1718 if (rc == -ENOENT) {
1719 /* no more RPC to produce */
1720 set->set_producer = NULL;
1721 set->set_producer_arg = NULL;
1726 RETURN((atomic_read(&set->set_remaining) - remaining));
1730 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1731 * and no more replies are expected.
1732 * (it is possible to get less replies than requests sent e.g. due to timed out
1733 * requests or requests that we had trouble to send out)
1735 * NOTE: This function contains a potential schedule point (cond_resched()).
1737 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1739 struct list_head *tmp, *next;
1740 struct list_head comp_reqs;
1741 int force_timer_recalc = 0;
1744 if (atomic_read(&set->set_remaining) == 0)
1747 INIT_LIST_HEAD(&comp_reqs);
1748 list_for_each_safe(tmp, next, &set->set_requests) {
1749 struct ptlrpc_request *req =
1750 list_entry(tmp, struct ptlrpc_request,
1752 struct obd_import *imp = req->rq_import;
1753 int unregistered = 0;
1757 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1758 list_move_tail(&req->rq_set_chain, &comp_reqs);
1763 * This schedule point is mainly for the ptlrpcd caller of this
1764 * function. Most ptlrpc sets are not long-lived and unbounded
1765 * in length, but at the least the set used by the ptlrpcd is.
1766 * Since the processing time is unbounded, we need to insert an
1767 * explicit schedule point to make the thread well-behaved.
1772 * If the caller requires to allow to be interpreted by force
1773 * and it has really been interpreted, then move the request
1774 * to RQ_PHASE_INTERPRET phase in spite of what the current
1777 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1778 req->rq_status = -EINTR;
1779 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1782 * Since it is interpreted and we have to wait for
1783 * the reply to be unlinked, then use sync mode.
1787 GOTO(interpret, req->rq_status);
1790 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1791 force_timer_recalc = 1;
1793 /* delayed send - skip */
1794 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1797 /* delayed resend - skip */
1798 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1799 req->rq_sent > ktime_get_real_seconds())
1802 if (!(req->rq_phase == RQ_PHASE_RPC ||
1803 req->rq_phase == RQ_PHASE_BULK ||
1804 req->rq_phase == RQ_PHASE_INTERPRET ||
1805 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1806 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1807 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1811 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1812 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1813 LASSERT(req->rq_next_phase != req->rq_phase);
1814 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1816 if (req->rq_req_deadline &&
1817 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1818 req->rq_req_deadline = 0;
1819 if (req->rq_reply_deadline &&
1820 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1821 req->rq_reply_deadline = 0;
1822 if (req->rq_bulk_deadline &&
1823 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1824 req->rq_bulk_deadline = 0;
1827 * Skip processing until reply is unlinked. We
1828 * can't return to pool before that and we can't
1829 * call interpret before that. We need to make
1830 * sure that all rdma transfers finished and will
1831 * not corrupt any data.
1833 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1834 ptlrpc_client_recv_or_unlink(req))
1836 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1837 ptlrpc_client_bulk_active(req))
1841 * Turn fail_loc off to prevent it from looping
1844 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1845 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1848 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1849 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1854 * Move to next phase if reply was successfully
1857 ptlrpc_rqphase_move(req, req->rq_next_phase);
1860 if (req->rq_phase == RQ_PHASE_INTERPRET)
1861 GOTO(interpret, req->rq_status);
1864 * Note that this also will start async reply unlink.
1866 if (req->rq_net_err && !req->rq_timedout) {
1867 ptlrpc_expire_one_request(req, 1);
1870 * Check if we still need to wait for unlink.
1872 if (ptlrpc_client_recv_or_unlink(req) ||
1873 ptlrpc_client_bulk_active(req))
1875 /* If there is no need to resend, fail it now. */
1876 if (req->rq_no_resend) {
1877 if (req->rq_status == 0)
1878 req->rq_status = -EIO;
1879 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1880 GOTO(interpret, req->rq_status);
1887 spin_lock(&req->rq_lock);
1888 req->rq_replied = 0;
1889 spin_unlock(&req->rq_lock);
1890 if (req->rq_status == 0)
1891 req->rq_status = -EIO;
1892 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1893 GOTO(interpret, req->rq_status);
1897 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1898 * so it sets rq_intr regardless of individual rpc
1899 * timeouts. The synchronous IO waiting path sets
1900 * rq_intr irrespective of whether ptlrpcd
1901 * has seen a timeout. Our policy is to only interpret
1902 * interrupted rpcs after they have timed out, so we
1903 * need to enforce that here.
1906 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1907 req->rq_wait_ctx)) {
1908 req->rq_status = -EINTR;
1909 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1910 GOTO(interpret, req->rq_status);
1913 if (req->rq_phase == RQ_PHASE_RPC) {
1914 if (req->rq_timedout || req->rq_resend ||
1915 req->rq_waiting || req->rq_wait_ctx) {
1918 if (!ptlrpc_unregister_reply(req, 1)) {
1919 ptlrpc_unregister_bulk(req, 1);
1923 spin_lock(&imp->imp_lock);
1924 if (ptlrpc_import_delay_req(imp, req,
1927 * put on delay list - only if we wait
1928 * recovery finished - before send
1930 list_move_tail(&req->rq_list,
1931 &imp->imp_delayed_list);
1932 spin_unlock(&imp->imp_lock);
1937 req->rq_status = status;
1938 ptlrpc_rqphase_move(req,
1939 RQ_PHASE_INTERPRET);
1940 spin_unlock(&imp->imp_lock);
1941 GOTO(interpret, req->rq_status);
1943 /* ignore on just initiated connections */
1944 if (ptlrpc_no_resend(req) &&
1945 !req->rq_wait_ctx &&
1946 imp->imp_generation !=
1947 imp->imp_initiated_at) {
1948 req->rq_status = -ENOTCONN;
1949 ptlrpc_rqphase_move(req,
1950 RQ_PHASE_INTERPRET);
1951 spin_unlock(&imp->imp_lock);
1952 GOTO(interpret, req->rq_status);
1955 list_move_tail(&req->rq_list,
1956 &imp->imp_sending_list);
1958 spin_unlock(&imp->imp_lock);
1960 spin_lock(&req->rq_lock);
1961 req->rq_waiting = 0;
1962 spin_unlock(&req->rq_lock);
1964 if (req->rq_timedout || req->rq_resend) {
1966 * This is re-sending anyways,
1967 * let's mark req as resend.
1969 spin_lock(&req->rq_lock);
1971 spin_unlock(&req->rq_lock);
1974 * rq_wait_ctx is only touched by ptlrpcd,
1975 * so no lock is needed here.
1977 status = sptlrpc_req_refresh_ctx(req, -1);
1980 req->rq_status = status;
1981 spin_lock(&req->rq_lock);
1982 req->rq_wait_ctx = 0;
1983 spin_unlock(&req->rq_lock);
1984 force_timer_recalc = 1;
1986 spin_lock(&req->rq_lock);
1987 req->rq_wait_ctx = 1;
1988 spin_unlock(&req->rq_lock);
1993 spin_lock(&req->rq_lock);
1994 req->rq_wait_ctx = 0;
1995 spin_unlock(&req->rq_lock);
1999 * In any case, the previous bulk should be
2000 * cleaned up to prepare for the new sending
2003 !ptlrpc_unregister_bulk(req, 1))
2006 rc = ptl_send_rpc(req, 0);
2007 if (rc == -ENOMEM) {
2008 spin_lock(&imp->imp_lock);
2009 if (!list_empty(&req->rq_list))
2010 list_del_init(&req->rq_list);
2011 spin_unlock(&imp->imp_lock);
2012 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2016 DEBUG_REQ(D_HA, req,
2017 "send failed: rc = %d", rc);
2018 force_timer_recalc = 1;
2019 spin_lock(&req->rq_lock);
2020 req->rq_net_err = 1;
2021 spin_unlock(&req->rq_lock);
2024 /* need to reset the timeout */
2025 force_timer_recalc = 1;
2028 spin_lock(&req->rq_lock);
2030 if (ptlrpc_client_early(req)) {
2031 ptlrpc_at_recv_early_reply(req);
2032 spin_unlock(&req->rq_lock);
2036 /* Still waiting for a reply? */
2037 if (ptlrpc_client_recv(req)) {
2038 spin_unlock(&req->rq_lock);
2042 /* Did we actually receive a reply? */
2043 if (!ptlrpc_client_replied(req)) {
2044 spin_unlock(&req->rq_lock);
2048 spin_unlock(&req->rq_lock);
2051 * unlink from net because we are going to
2052 * swab in-place of reply buffer
2054 unregistered = ptlrpc_unregister_reply(req, 1);
2058 req->rq_status = after_reply(req);
2063 * If there is no bulk associated with this request,
2064 * then we're done and should let the interpreter
2065 * process the reply. Similarly if the RPC returned
2066 * an error, and therefore the bulk will never arrive.
2068 if (!req->rq_bulk || req->rq_status < 0) {
2069 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2070 GOTO(interpret, req->rq_status);
2073 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2076 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2077 if (ptlrpc_client_bulk_active(req))
2080 if (req->rq_bulk->bd_failure) {
2082 * The RPC reply arrived OK, but the bulk screwed
2083 * up! Dead weird since the server told us the RPC
2084 * was good after getting the REPLY for her GET or
2085 * the ACK for her PUT.
2087 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2088 req->rq_status = -EIO;
2091 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2094 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2097 * This moves to "unregistering" phase we need to wait for
2100 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2101 /* start async bulk unlink too */
2102 ptlrpc_unregister_bulk(req, 1);
2106 if (!ptlrpc_unregister_bulk(req, async))
2110 * When calling interpret receiving already should be
2113 LASSERT(!req->rq_receiving_reply);
2115 ptlrpc_req_interpret(env, req, req->rq_status);
2117 if (ptlrpcd_check_work(req)) {
2118 atomic_dec(&set->set_remaining);
2121 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2125 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2126 req, current_comm(),
2127 imp->imp_obd->obd_uuid.uuid,
2128 lustre_msg_get_status(req->rq_reqmsg),
2130 obd_import_nid2str(imp),
2131 lustre_msg_get_opc(req->rq_reqmsg),
2132 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2134 spin_lock(&imp->imp_lock);
2136 * Request already may be not on sending or delaying list. This
2137 * may happen in the case of marking it erroneous for the case
2138 * ptlrpc_import_delay_req(req, status) find it impossible to
2139 * allow sending this rpc and returns *status != 0.
2141 if (!list_empty(&req->rq_list)) {
2142 list_del_init(&req->rq_list);
2143 atomic_dec(&imp->imp_inflight);
2145 list_del_init(&req->rq_unreplied_list);
2146 spin_unlock(&imp->imp_lock);
2148 atomic_dec(&set->set_remaining);
2149 wake_up_all(&imp->imp_recovery_waitq);
2151 if (set->set_producer) {
2152 /* produce a new request if possible */
2153 if (ptlrpc_set_producer(set) > 0)
2154 force_timer_recalc = 1;
2157 * free the request that has just been completed
2158 * in order not to pollute set->set_requests
2160 list_del_init(&req->rq_set_chain);
2161 spin_lock(&req->rq_lock);
2163 req->rq_invalid_rqset = 0;
2164 spin_unlock(&req->rq_lock);
2166 /* record rq_status to compute the final status later */
2167 if (req->rq_status != 0)
2168 set->set_rc = req->rq_status;
2169 ptlrpc_req_finished(req);
2171 list_move_tail(&req->rq_set_chain, &comp_reqs);
2176 * move completed request at the head of list so it's easier for
2177 * caller to find them
2179 list_splice(&comp_reqs, &set->set_requests);
2181 /* If we hit an error, we want to recover promptly. */
2182 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2184 EXPORT_SYMBOL(ptlrpc_check_set);
2187 * Time out request \a req. is \a async_unlink is set, that means do not wait
2188 * until LNet actually confirms network buffer unlinking.
2189 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2191 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2193 struct obd_import *imp = req->rq_import;
2194 unsigned int debug_mask = D_RPCTRACE;
2198 spin_lock(&req->rq_lock);
2199 req->rq_timedout = 1;
2200 spin_unlock(&req->rq_lock);
2202 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2203 lustre_msg_get_status(req->rq_reqmsg)))
2204 debug_mask = D_WARNING;
2205 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2206 req->rq_net_err ? "failed due to network error" :
2207 ((req->rq_real_sent == 0 ||
2208 req->rq_real_sent < req->rq_sent ||
2209 req->rq_real_sent >= req->rq_deadline) ?
2210 "timed out for sent delay" : "timed out for slow reply"),
2211 (s64)req->rq_sent, (s64)req->rq_real_sent);
2213 if (imp && obd_debug_peer_on_timeout)
2214 LNetDebugPeer(imp->imp_connection->c_peer);
2216 ptlrpc_unregister_reply(req, async_unlink);
2217 ptlrpc_unregister_bulk(req, async_unlink);
2219 if (obd_dump_on_timeout)
2220 libcfs_debug_dumplog();
2223 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2227 atomic_inc(&imp->imp_timeouts);
2229 /* The DLM server doesn't want recovery run on its imports. */
2230 if (imp->imp_dlm_fake)
2234 * If this request is for recovery or other primordial tasks,
2235 * then error it out here.
2237 if (req->rq_ctx_init || req->rq_ctx_fini ||
2238 req->rq_send_state != LUSTRE_IMP_FULL ||
2239 imp->imp_obd->obd_no_recov) {
2240 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2241 ptlrpc_import_state_name(req->rq_send_state),
2242 ptlrpc_import_state_name(imp->imp_state));
2243 spin_lock(&req->rq_lock);
2244 req->rq_status = -ETIMEDOUT;
2246 spin_unlock(&req->rq_lock);
2251 * if a request can't be resent we can't wait for an answer after
2254 if (ptlrpc_no_resend(req)) {
2255 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2259 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2265 * Time out all uncompleted requests in request set pointed by \a data
2266 * Callback used when waiting on sets with l_wait_event.
2269 int ptlrpc_expired_set(void *data)
2271 struct ptlrpc_request_set *set = data;
2272 struct list_head *tmp;
2273 time64_t now = ktime_get_real_seconds();
2276 LASSERT(set != NULL);
2279 * A timeout expired. See which reqs it applies to...
2281 list_for_each(tmp, &set->set_requests) {
2282 struct ptlrpc_request *req =
2283 list_entry(tmp, struct ptlrpc_request,
2286 /* don't expire request waiting for context */
2287 if (req->rq_wait_ctx)
2290 /* Request in-flight? */
2291 if (!((req->rq_phase == RQ_PHASE_RPC &&
2292 !req->rq_waiting && !req->rq_resend) ||
2293 (req->rq_phase == RQ_PHASE_BULK)))
2296 if (req->rq_timedout || /* already dealt with */
2297 req->rq_deadline > now) /* not expired */
2301 * Deal with this guy. Do it asynchronously to not block
2304 ptlrpc_expire_one_request(req, 1);
2308 * When waiting for a whole set, we always break out of the
2309 * sleep so we can recalculate the timeout, or enable interrupts
2310 * if everyone's timed out.
2316 * Sets rq_intr flag in \a req under spinlock.
2318 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2320 spin_lock(&req->rq_lock);
2322 spin_unlock(&req->rq_lock);
2324 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2327 * Interrupts (sets interrupted flag) all uncompleted requests in
2328 * a set \a data. Callback for l_wait_event for interruptible waits.
2330 static void ptlrpc_interrupted_set(void *data)
2332 struct ptlrpc_request_set *set = data;
2333 struct list_head *tmp;
2335 LASSERT(set != NULL);
2336 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2338 list_for_each(tmp, &set->set_requests) {
2339 struct ptlrpc_request *req =
2340 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2345 if (req->rq_phase != RQ_PHASE_RPC &&
2346 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2347 !req->rq_allow_intr)
2350 ptlrpc_mark_interrupted(req);
2355 * Get the smallest timeout in the set; this does NOT set a timeout.
2357 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2359 struct list_head *tmp;
2360 time64_t now = ktime_get_real_seconds();
2362 struct ptlrpc_request *req;
2366 list_for_each(tmp, &set->set_requests) {
2367 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2369 /* Request in-flight? */
2370 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2371 (req->rq_phase == RQ_PHASE_BULK) ||
2372 (req->rq_phase == RQ_PHASE_NEW)))
2375 /* Already timed out. */
2376 if (req->rq_timedout)
2379 /* Waiting for ctx. */
2380 if (req->rq_wait_ctx)
2383 if (req->rq_phase == RQ_PHASE_NEW)
2384 deadline = req->rq_sent;
2385 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2386 deadline = req->rq_sent;
2388 deadline = req->rq_sent + req->rq_timeout;
2390 if (deadline <= now) /* actually expired already */
2391 timeout = 1; /* ASAP */
2392 else if (timeout == 0 || timeout > deadline - now)
2393 timeout = deadline - now;
2399 * Send all unset request from the set and then wait untill all
2400 * requests in the set complete (either get a reply, timeout, get an
2401 * error or otherwise be interrupted).
2402 * Returns 0 on success or error code otherwise.
2404 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2406 struct list_head *tmp;
2407 struct ptlrpc_request *req;
2408 struct l_wait_info lwi;
2413 if (set->set_producer)
2414 (void)ptlrpc_set_producer(set);
2416 list_for_each(tmp, &set->set_requests) {
2417 req = list_entry(tmp, struct ptlrpc_request,
2419 if (req->rq_phase == RQ_PHASE_NEW)
2420 (void)ptlrpc_send_new_req(req);
2423 if (list_empty(&set->set_requests))
2427 timeout = ptlrpc_set_next_timeout(set);
2430 * wait until all complete, interrupted, or an in-flight
2433 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2436 if ((timeout == 0 && !signal_pending(current)) ||
2437 set->set_allow_intr)
2439 * No requests are in-flight (ether timed out
2440 * or delayed), so we can allow interrupts.
2441 * We still want to block for a limited time,
2442 * so we allow interrupts during the timeout.
2444 lwi = LWI_TIMEOUT_INTR_ALL(
2445 cfs_time_seconds(timeout ? timeout : 1),
2447 ptlrpc_interrupted_set, set);
2450 * At least one request is in flight, so no
2451 * interrupts are allowed. Wait until all
2452 * complete, or an in-flight req times out.
2454 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2455 ptlrpc_expired_set, set);
2457 rc = l_wait_event(set->set_waitq,
2458 ptlrpc_check_set(NULL, set), &lwi);
2461 * LU-769 - if we ignored the signal because it was already
2462 * pending when we started, we need to handle it now or we risk
2463 * it being ignored forever
2465 if (rc == -ETIMEDOUT &&
2466 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2467 signal_pending(current)) {
2468 sigset_t blocked_sigs =
2469 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2472 * In fact we only interrupt for the "fatal" signals
2473 * like SIGINT or SIGKILL. We still ignore less
2474 * important signals since ptlrpc set is not easily
2475 * reentrant from userspace again
2477 if (signal_pending(current))
2478 ptlrpc_interrupted_set(set);
2479 cfs_restore_sigs(blocked_sigs);
2482 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2485 * -EINTR => all requests have been flagged rq_intr so next
2487 * -ETIMEDOUT => someone timed out. When all reqs have
2488 * timed out, signals are enabled allowing completion with
2490 * I don't really care if we go once more round the loop in
2491 * the error cases -eeb.
2493 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2494 list_for_each(tmp, &set->set_requests) {
2495 req = list_entry(tmp, struct ptlrpc_request,
2497 spin_lock(&req->rq_lock);
2498 req->rq_invalid_rqset = 1;
2499 spin_unlock(&req->rq_lock);
2502 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2504 LASSERT(atomic_read(&set->set_remaining) == 0);
2506 rc = set->set_rc; /* rq_status of already freed requests if any */
2507 list_for_each(tmp, &set->set_requests) {
2508 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2510 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2511 if (req->rq_status != 0)
2512 rc = req->rq_status;
2517 EXPORT_SYMBOL(ptlrpc_set_wait);
2520 * Helper fuction for request freeing.
2521 * Called when request count reached zero and request needs to be freed.
2522 * Removes request from all sorts of sending/replay lists it might be on,
2523 * frees network buffers if any are present.
2524 * If \a locked is set, that means caller is already holding import imp_lock
2525 * and so we no longer need to reobtain it (for certain lists manipulations)
2527 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2534 LASSERT(!request->rq_srv_req);
2535 LASSERT(request->rq_export == NULL);
2536 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2537 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2538 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2539 LASSERTF(!request->rq_replay, "req %p\n", request);
2541 req_capsule_fini(&request->rq_pill);
2544 * We must take it off the imp_replay_list first. Otherwise, we'll set
2545 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2547 if (request->rq_import) {
2549 spin_lock(&request->rq_import->imp_lock);
2550 list_del_init(&request->rq_replay_list);
2551 list_del_init(&request->rq_unreplied_list);
2553 spin_unlock(&request->rq_import->imp_lock);
2555 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2557 if (atomic_read(&request->rq_refcount) != 0) {
2558 DEBUG_REQ(D_ERROR, request,
2559 "freeing request with nonzero refcount");
2563 if (request->rq_repbuf)
2564 sptlrpc_cli_free_repbuf(request);
2566 if (request->rq_import) {
2567 class_import_put(request->rq_import);
2568 request->rq_import = NULL;
2570 if (request->rq_bulk)
2571 ptlrpc_free_bulk(request->rq_bulk);
2573 if (request->rq_reqbuf || request->rq_clrbuf)
2574 sptlrpc_cli_free_reqbuf(request);
2576 if (request->rq_cli_ctx)
2577 sptlrpc_req_put_ctx(request, !locked);
2579 if (request->rq_pool)
2580 __ptlrpc_free_req_to_pool(request);
2582 ptlrpc_request_cache_free(request);
2586 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2588 * Drop one request reference. Must be called with import imp_lock held.
2589 * When reference count drops to zero, request is freed.
2591 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2593 assert_spin_locked(&request->rq_import->imp_lock);
2594 (void)__ptlrpc_req_finished(request, 1);
2599 * Drops one reference count for request \a request.
2600 * \a locked set indicates that caller holds import imp_lock.
2601 * Frees the request whe reference count reaches zero.
2603 * \retval 1 the request is freed
2604 * \retval 0 some others still hold references on the request
2606 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2614 LASSERT(request != LP_POISON);
2615 LASSERT(request->rq_reqmsg != LP_POISON);
2617 DEBUG_REQ(D_INFO, request, "refcount now %u",
2618 atomic_read(&request->rq_refcount) - 1);
2620 spin_lock(&request->rq_lock);
2621 count = atomic_dec_return(&request->rq_refcount);
2622 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2625 * For open RPC, the client does not know the EA size (LOV, ACL, and
2626 * so on) before replied, then the client has to reserve very large
2627 * reply buffer. Such buffer will not be released until the RPC freed.
2628 * Since The open RPC is replayable, we need to keep it in the replay
2629 * list until close. If there are a lot of files opened concurrently,
2630 * then the client may be OOM.
2632 * If fact, it is unnecessary to keep reply buffer for open replay,
2633 * related EAs have already been saved via mdc_save_lovea() before
2634 * coming here. So it is safe to free the reply buffer some earlier
2635 * before releasing the RPC to avoid client OOM. LU-9514
2637 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2638 spin_lock(&request->rq_early_free_lock);
2639 sptlrpc_cli_free_repbuf(request);
2640 request->rq_repbuf = NULL;
2641 request->rq_repbuf_len = 0;
2642 request->rq_repdata = NULL;
2643 request->rq_reqdata_len = 0;
2644 spin_unlock(&request->rq_early_free_lock);
2646 spin_unlock(&request->rq_lock);
2649 __ptlrpc_free_req(request, locked);
2655 * Drops one reference count for a request.
2657 void ptlrpc_req_finished(struct ptlrpc_request *request)
2659 __ptlrpc_req_finished(request, 0);
2661 EXPORT_SYMBOL(ptlrpc_req_finished);
2664 * Returns xid of a \a request
2666 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2668 return request->rq_xid;
2670 EXPORT_SYMBOL(ptlrpc_req_xid);
2673 * Disengage the client's reply buffer from the network
2674 * NB does _NOT_ unregister any client-side bulk.
2675 * IDEMPOTENT, but _not_ safe against concurrent callers.
2676 * The request owner (i.e. the thread doing the I/O) must call...
2677 * Returns 0 on success or 1 if unregistering cannot be made.
2679 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2682 struct l_wait_info lwi;
2687 LASSERT(!in_interrupt());
2689 /* Let's setup deadline for reply unlink. */
2690 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2691 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2692 request->rq_reply_deadline = ktime_get_real_seconds() +
2696 * Nothing left to do.
2698 if (!ptlrpc_client_recv_or_unlink(request))
2701 LNetMDUnlink(request->rq_reply_md_h);
2704 * Let's check it once again.
2706 if (!ptlrpc_client_recv_or_unlink(request))
2709 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2710 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2713 * Do not wait for unlink to finish.
2719 * We have to l_wait_event() whatever the result, to give liblustre
2720 * a chance to run reply_in_callback(), and to make sure we've
2721 * unlinked before returning a req to the pool.
2724 /* The wq argument is ignored by user-space wait_event macros */
2725 wait_queue_head_t *wq = (request->rq_set) ?
2726 &request->rq_set->set_waitq :
2727 &request->rq_reply_waitq;
2729 * Network access will complete in finite time but the HUGE
2730 * timeout lets us CWARN for visibility of sluggish NALs
2732 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2733 cfs_time_seconds(1), NULL, NULL);
2734 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2737 ptlrpc_rqphase_move(request, request->rq_next_phase);
2741 LASSERT(rc == -ETIMEDOUT);
2742 DEBUG_REQ(D_WARNING, request,
2743 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2744 request->rq_receiving_reply,
2745 request->rq_req_unlinked,
2746 request->rq_reply_unlinked);
2751 static void ptlrpc_free_request(struct ptlrpc_request *req)
2753 spin_lock(&req->rq_lock);
2755 spin_unlock(&req->rq_lock);
2757 if (req->rq_commit_cb)
2758 req->rq_commit_cb(req);
2759 list_del_init(&req->rq_replay_list);
2761 __ptlrpc_req_finished(req, 1);
2765 * the request is committed and dropped from the replay list of its import
2767 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2769 struct obd_import *imp = req->rq_import;
2771 spin_lock(&imp->imp_lock);
2772 if (list_empty(&req->rq_replay_list)) {
2773 spin_unlock(&imp->imp_lock);
2777 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2778 if (imp->imp_replay_cursor == &req->rq_replay_list)
2779 imp->imp_replay_cursor = req->rq_replay_list.next;
2780 ptlrpc_free_request(req);
2783 spin_unlock(&imp->imp_lock);
2785 EXPORT_SYMBOL(ptlrpc_request_committed);
2788 * Iterates through replay_list on import and prunes
2789 * all requests have transno smaller than last_committed for the
2790 * import and don't have rq_replay set.
2791 * Since requests are sorted in transno order, stops when meetign first
2792 * transno bigger than last_committed.
2793 * caller must hold imp->imp_lock
2795 void ptlrpc_free_committed(struct obd_import *imp)
2797 struct ptlrpc_request *req, *saved;
2798 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2799 bool skip_committed_list = true;
2802 LASSERT(imp != NULL);
2803 assert_spin_locked(&imp->imp_lock);
2805 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2806 imp->imp_generation == imp->imp_last_generation_checked) {
2807 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2808 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2811 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2812 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2813 imp->imp_generation);
2815 if (imp->imp_generation != imp->imp_last_generation_checked ||
2816 imp->imp_last_transno_checked == 0)
2817 skip_committed_list = false;
2819 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2820 imp->imp_last_generation_checked = imp->imp_generation;
2822 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2824 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2825 LASSERT(req != last_req);
2828 if (req->rq_transno == 0) {
2829 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2832 if (req->rq_import_generation < imp->imp_generation) {
2833 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2837 /* not yet committed */
2838 if (req->rq_transno > imp->imp_peer_committed_transno) {
2839 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2843 if (req->rq_replay) {
2844 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2845 list_move_tail(&req->rq_replay_list,
2846 &imp->imp_committed_list);
2850 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2851 imp->imp_peer_committed_transno);
2853 ptlrpc_free_request(req);
2856 if (skip_committed_list)
2859 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2861 LASSERT(req->rq_transno != 0);
2862 if (req->rq_import_generation < imp->imp_generation ||
2864 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2865 req->rq_import_generation <
2866 imp->imp_generation ? "stale" : "closed");
2868 if (imp->imp_replay_cursor == &req->rq_replay_list)
2869 imp->imp_replay_cursor =
2870 req->rq_replay_list.next;
2872 ptlrpc_free_request(req);
2879 void ptlrpc_cleanup_client(struct obd_import *imp)
2886 * Schedule previously sent request for resend.
2887 * For bulk requests we assign new xid (to avoid problems with
2888 * lost replies and therefore several transfers landing into same buffer
2889 * from different sending attempts).
2891 void ptlrpc_resend_req(struct ptlrpc_request *req)
2893 DEBUG_REQ(D_HA, req, "going to resend");
2894 spin_lock(&req->rq_lock);
2897 * Request got reply but linked to the import list still.
2898 * Let ptlrpc_check_set() process it.
2900 if (ptlrpc_client_replied(req)) {
2901 spin_unlock(&req->rq_lock);
2902 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2906 req->rq_status = -EAGAIN;
2909 req->rq_net_err = 0;
2910 req->rq_timedout = 0;
2912 ptlrpc_client_wake_req(req);
2913 spin_unlock(&req->rq_lock);
2916 /* XXX: this function and rq_status are currently unused */
2917 void ptlrpc_restart_req(struct ptlrpc_request *req)
2919 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2920 req->rq_status = -ERESTARTSYS;
2922 spin_lock(&req->rq_lock);
2923 req->rq_restart = 1;
2924 req->rq_timedout = 0;
2925 ptlrpc_client_wake_req(req);
2926 spin_unlock(&req->rq_lock);
2930 * Grab additional reference on a request \a req
2932 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2935 atomic_inc(&req->rq_refcount);
2938 EXPORT_SYMBOL(ptlrpc_request_addref);
2941 * Add a request to import replay_list.
2942 * Must be called under imp_lock
2944 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2945 struct obd_import *imp)
2947 struct list_head *tmp;
2949 assert_spin_locked(&imp->imp_lock);
2951 if (req->rq_transno == 0) {
2952 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2957 * clear this for new requests that were resent as well
2958 * as resent replayed requests.
2960 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2962 /* don't re-add requests that have been replayed */
2963 if (!list_empty(&req->rq_replay_list))
2966 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2968 spin_lock(&req->rq_lock);
2970 spin_unlock(&req->rq_lock);
2972 LASSERT(imp->imp_replayable);
2973 /* Balanced in ptlrpc_free_committed, usually. */
2974 ptlrpc_request_addref(req);
2975 list_for_each_prev(tmp, &imp->imp_replay_list) {
2976 struct ptlrpc_request *iter = list_entry(tmp,
2977 struct ptlrpc_request,
2981 * We may have duplicate transnos if we create and then
2982 * open a file, or for closes retained if to match creating
2983 * opens, so use req->rq_xid as a secondary key.
2984 * (See bugs 684, 685, and 428.)
2985 * XXX no longer needed, but all opens need transnos!
2987 if (iter->rq_transno > req->rq_transno)
2990 if (iter->rq_transno == req->rq_transno) {
2991 LASSERT(iter->rq_xid != req->rq_xid);
2992 if (iter->rq_xid > req->rq_xid)
2996 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3000 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3004 * Send request and wait until it completes.
3005 * Returns request processing status.
3007 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3009 struct ptlrpc_request_set *set;
3013 LASSERT(req->rq_set == NULL);
3014 LASSERT(!req->rq_receiving_reply);
3016 set = ptlrpc_prep_set();
3018 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3022 /* for distributed debugging */
3023 lustre_msg_set_status(req->rq_reqmsg, current_pid());
3025 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3026 ptlrpc_request_addref(req);
3027 ptlrpc_set_add_req(set, req);
3028 rc = ptlrpc_set_wait(NULL, set);
3029 ptlrpc_set_destroy(set);
3033 EXPORT_SYMBOL(ptlrpc_queue_wait);
3036 * Callback used for replayed requests reply processing.
3037 * In case of successful reply calls registered request replay callback.
3038 * In case of error restart replay process.
3040 static int ptlrpc_replay_interpret(const struct lu_env *env,
3041 struct ptlrpc_request *req,
3044 struct ptlrpc_replay_async_args *aa = args;
3045 struct obd_import *imp = req->rq_import;
3048 atomic_dec(&imp->imp_replay_inflight);
3051 * Note: if it is bulk replay (MDS-MDS replay), then even if
3052 * server got the request, but bulk transfer timeout, let's
3053 * replay the bulk req again
3055 if (!ptlrpc_client_replied(req) ||
3057 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3058 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3059 GOTO(out, rc = -ETIMEDOUT);
3062 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3063 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3064 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3065 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3067 /** VBR: check version failure */
3068 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3069 /** replay was failed due to version mismatch */
3070 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3071 spin_lock(&imp->imp_lock);
3072 imp->imp_vbr_failed = 1;
3073 spin_unlock(&imp->imp_lock);
3074 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3076 /** The transno had better not change over replay. */
3077 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3078 lustre_msg_get_transno(req->rq_repmsg) ||
3079 lustre_msg_get_transno(req->rq_repmsg) == 0,
3081 lustre_msg_get_transno(req->rq_reqmsg),
3082 lustre_msg_get_transno(req->rq_repmsg));
3085 spin_lock(&imp->imp_lock);
3086 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3087 spin_unlock(&imp->imp_lock);
3088 LASSERT(imp->imp_last_replay_transno);
3090 /* transaction number shouldn't be bigger than the latest replayed */
3091 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3092 DEBUG_REQ(D_ERROR, req,
3093 "Reported transno=%llu is bigger than replayed=%llu",
3095 lustre_msg_get_transno(req->rq_reqmsg));
3096 GOTO(out, rc = -EINVAL);
3099 DEBUG_REQ(D_HA, req, "got reply");
3101 /* let the callback do fixups, possibly including in the request */
3102 if (req->rq_replay_cb)
3103 req->rq_replay_cb(req);
3105 if (ptlrpc_client_replied(req) &&
3106 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3107 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3108 lustre_msg_get_status(req->rq_repmsg),
3109 aa->praa_old_status);
3112 * Note: If the replay fails for MDT-MDT recovery, let's
3113 * abort all of the following requests in the replay
3114 * and sending list, because MDT-MDT update requests
3115 * are dependent on each other, see LU-7039
3117 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3118 struct ptlrpc_request *free_req;
3119 struct ptlrpc_request *tmp;
3121 spin_lock(&imp->imp_lock);
3122 list_for_each_entry_safe(free_req, tmp,
3123 &imp->imp_replay_list,
3125 ptlrpc_free_request(free_req);
3128 list_for_each_entry_safe(free_req, tmp,
3129 &imp->imp_committed_list,
3131 ptlrpc_free_request(free_req);
3134 list_for_each_entry_safe(free_req, tmp,
3135 &imp->imp_delayed_list,
3137 spin_lock(&free_req->rq_lock);
3138 free_req->rq_err = 1;
3139 free_req->rq_status = -EIO;
3140 ptlrpc_client_wake_req(free_req);
3141 spin_unlock(&free_req->rq_lock);
3144 list_for_each_entry_safe(free_req, tmp,
3145 &imp->imp_sending_list,
3147 spin_lock(&free_req->rq_lock);
3148 free_req->rq_err = 1;
3149 free_req->rq_status = -EIO;
3150 ptlrpc_client_wake_req(free_req);
3151 spin_unlock(&free_req->rq_lock);
3153 spin_unlock(&imp->imp_lock);
3156 /* Put it back for re-replay. */
3157 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3161 * Errors while replay can set transno to 0, but
3162 * imp_last_replay_transno shouldn't be set to 0 anyway
3164 if (req->rq_transno == 0)
3165 CERROR("Transno is 0 during replay!\n");
3167 /* continue with recovery */
3168 rc = ptlrpc_import_recovery_state_machine(imp);
3170 req->rq_send_state = aa->praa_old_state;
3173 /* this replay failed, so restart recovery */
3174 ptlrpc_connect_import(imp);
3180 * Prepares and queues request for replay.
3181 * Adds it to ptlrpcd queue for actual sending.
3182 * Returns 0 on success.
3184 int ptlrpc_replay_req(struct ptlrpc_request *req)
3186 struct ptlrpc_replay_async_args *aa;
3190 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3192 aa = ptlrpc_req_async_args(aa, req);
3193 memset(aa, 0, sizeof(*aa));
3195 /* Prepare request to be resent with ptlrpcd */
3196 aa->praa_old_state = req->rq_send_state;
3197 req->rq_send_state = LUSTRE_IMP_REPLAY;
3198 req->rq_phase = RQ_PHASE_NEW;
3199 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3201 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3203 req->rq_interpret_reply = ptlrpc_replay_interpret;
3204 /* Readjust the timeout for current conditions */
3205 ptlrpc_at_set_req_timeout(req);
3207 /* Tell server net_latency to calculate how long to wait for reply. */
3208 lustre_msg_set_service_time(req->rq_reqmsg,
3209 ptlrpc_at_get_net_latency(req));
3210 DEBUG_REQ(D_HA, req, "REPLAY");
3212 atomic_inc(&req->rq_import->imp_replay_inflight);
3213 spin_lock(&req->rq_lock);
3214 req->rq_early_free_repbuf = 0;
3215 spin_unlock(&req->rq_lock);
3216 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3218 ptlrpcd_add_req(req);
3223 * Aborts all in-flight request on import \a imp sending and delayed lists
3225 void ptlrpc_abort_inflight(struct obd_import *imp)
3227 struct list_head *tmp, *n;
3231 * Make sure that no new requests get processed for this import.
3232 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3233 * this flag and then putting requests on sending_list or delayed_list.
3235 assert_spin_locked(&imp->imp_lock);
3238 * XXX locking? Maybe we should remove each request with the list
3239 * locked? Also, how do we know if the requests on the list are
3240 * being freed at this time?
3242 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3243 struct ptlrpc_request *req = list_entry(tmp,
3244 struct ptlrpc_request,
3247 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3249 spin_lock(&req->rq_lock);
3250 if (req->rq_import_generation < imp->imp_generation) {
3252 req->rq_status = -EIO;
3253 ptlrpc_client_wake_req(req);
3255 spin_unlock(&req->rq_lock);
3258 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3259 struct ptlrpc_request *req =
3260 list_entry(tmp, struct ptlrpc_request, rq_list);
3262 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3264 spin_lock(&req->rq_lock);
3265 if (req->rq_import_generation < imp->imp_generation) {
3267 req->rq_status = -EIO;
3268 ptlrpc_client_wake_req(req);
3270 spin_unlock(&req->rq_lock);
3274 * Last chance to free reqs left on the replay list, but we
3275 * will still leak reqs that haven't committed.
3277 if (imp->imp_replayable)
3278 ptlrpc_free_committed(imp);
3284 * Abort all uncompleted requests in request set \a set
3286 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3288 struct list_head *tmp, *pos;
3290 LASSERT(set != NULL);
3292 list_for_each_safe(pos, tmp, &set->set_requests) {
3293 struct ptlrpc_request *req =
3294 list_entry(pos, struct ptlrpc_request,
3297 spin_lock(&req->rq_lock);
3298 if (req->rq_phase != RQ_PHASE_RPC) {
3299 spin_unlock(&req->rq_lock);
3304 req->rq_status = -EINTR;
3305 ptlrpc_client_wake_req(req);
3306 spin_unlock(&req->rq_lock);
3311 * Initialize the XID for the node. This is common among all requests on
3312 * this node, and only requires the property that it is monotonically
3313 * increasing. It does not need to be sequential. Since this is also used
3314 * as the RDMA match bits, it is important that a single client NOT have
3315 * the same match bits for two different in-flight requests, hence we do
3316 * NOT want to have an XID per target or similar.
3318 * To avoid an unlikely collision between match bits after a client reboot
3319 * (which would deliver old data into the wrong RDMA buffer) initialize
3320 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3321 * If the time is clearly incorrect, we instead use a 62-bit random number.
3322 * In the worst case the random number will overflow 1M RPCs per second in
3323 * 9133 years, or permutations thereof.
3325 #define YEAR_2004 (1ULL << 30)
3326 void ptlrpc_init_xid(void)
3328 time64_t now = ktime_get_real_seconds();
3331 if (now < YEAR_2004) {
3332 get_random_bytes(&xid, sizeof(xid));
3334 xid |= (1ULL << 61);
3336 xid = (u64)now << 20;
3339 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3340 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3341 xid &= PTLRPC_BULK_OPS_MASK;
3342 atomic64_set(&ptlrpc_last_xid, xid);
3346 * Increase xid and returns resulting new value to the caller.
3348 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3349 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3350 * itself uses the last bulk xid needed, so the server can determine the
3351 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3352 * xid must align to a power-of-two value.
3354 * This is assumed to be true due to the initial ptlrpc_last_xid
3355 * value also being initialized to a power-of-two value. LU-1431
3357 __u64 ptlrpc_next_xid(void)
3359 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3363 * If request has a new allocated XID (new request or EINPROGRESS resend),
3364 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3365 * request to ensure previous bulk fails and avoid problems with lost replies
3366 * and therefore several transfers landing into the same buffer from different
3369 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3371 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3373 LASSERT(bd != NULL);
3376 * Generate new matchbits for all resend requests, including
3379 if (req->rq_resend) {
3380 __u64 old_mbits = req->rq_mbits;
3383 * First time resend on -EINPROGRESS will generate new xid,
3384 * so we can actually use the rq_xid as rq_mbits in such case,
3385 * however, it's bit hard to distinguish such resend with a
3386 * 'resend for the -EINPROGRESS resend'. To make it simple,
3387 * we opt to generate mbits for all resend cases.
3389 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3391 req->rq_mbits = ptlrpc_next_xid();
3394 * Old version transfers rq_xid to peer as
3397 spin_lock(&req->rq_import->imp_lock);
3398 list_del_init(&req->rq_unreplied_list);
3399 ptlrpc_assign_next_xid_nolock(req);
3400 spin_unlock(&req->rq_import->imp_lock);
3401 req->rq_mbits = req->rq_xid;
3403 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3404 old_mbits, req->rq_mbits);
3405 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3406 /* Request being sent first time, use xid as matchbits. */
3407 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3408 || req->rq_mbits == 0) {
3409 req->rq_mbits = req->rq_xid;
3411 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3413 req->rq_mbits -= total_md - 1;
3417 * Replay request, xid and matchbits have already been
3418 * correctly assigned.
3424 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3425 * that server can infer the number of bulks that were prepared,
3428 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3432 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3433 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3435 * It's ok to directly set the rq_xid here, since this xid bump
3436 * won't affect the request position in unreplied list.
3438 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3439 req->rq_xid = req->rq_mbits;
3443 * Get a glimpse at what next xid value might have been.
3444 * Returns possible next xid.
3446 __u64 ptlrpc_sample_next_xid(void)
3448 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3450 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3453 * Functions for operating ptlrpc workers.
3455 * A ptlrpc work is a function which will be running inside ptlrpc context.
3456 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3458 * 1. after a work is created, it can be used many times, that is:
3459 * handler = ptlrpcd_alloc_work();
3460 * ptlrpcd_queue_work();
3462 * queue it again when necessary:
3463 * ptlrpcd_queue_work();
3464 * ptlrpcd_destroy_work();
3465 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3466 * it will only be queued once in any time. Also as its name implies, it may
3467 * have delay before it really runs by ptlrpcd thread.
3469 struct ptlrpc_work_async_args {
3470 int (*cb)(const struct lu_env *, void *);
3474 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3476 /* re-initialize the req */
3477 req->rq_timeout = obd_timeout;
3478 req->rq_sent = ktime_get_real_seconds();
3479 req->rq_deadline = req->rq_sent + req->rq_timeout;
3480 req->rq_phase = RQ_PHASE_INTERPRET;
3481 req->rq_next_phase = RQ_PHASE_COMPLETE;
3482 req->rq_xid = ptlrpc_next_xid();
3483 req->rq_import_generation = req->rq_import->imp_generation;
3485 ptlrpcd_add_req(req);
3488 static int work_interpreter(const struct lu_env *env,
3489 struct ptlrpc_request *req, void *args, int rc)
3491 struct ptlrpc_work_async_args *arg = args;
3493 LASSERT(ptlrpcd_check_work(req));
3494 LASSERT(arg->cb != NULL);
3496 rc = arg->cb(env, arg->cbdata);
3498 list_del_init(&req->rq_set_chain);
3501 if (atomic_dec_return(&req->rq_refcount) > 1) {
3502 atomic_set(&req->rq_refcount, 2);
3503 ptlrpcd_add_work_req(req);
3508 static int worker_format;
3510 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3512 return req->rq_pill.rc_fmt == (void *)&worker_format;
3516 * Create a work for ptlrpc.
3518 void *ptlrpcd_alloc_work(struct obd_import *imp,
3519 int (*cb)(const struct lu_env *, void *), void *cbdata)
3521 struct ptlrpc_request *req = NULL;
3522 struct ptlrpc_work_async_args *args;
3528 RETURN(ERR_PTR(-EINVAL));
3530 /* copy some code from deprecated fakereq. */
3531 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3533 CERROR("ptlrpc: run out of memory!\n");
3534 RETURN(ERR_PTR(-ENOMEM));
3537 ptlrpc_cli_req_init(req);
3539 req->rq_send_state = LUSTRE_IMP_FULL;
3540 req->rq_type = PTL_RPC_MSG_REQUEST;
3541 req->rq_import = class_import_get(imp);
3542 req->rq_interpret_reply = work_interpreter;
3543 /* don't want reply */
3544 req->rq_no_delay = req->rq_no_resend = 1;
3545 req->rq_pill.rc_fmt = (void *)&worker_format;
3547 args = ptlrpc_req_async_args(args, req);
3549 args->cbdata = cbdata;
3553 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3555 void ptlrpcd_destroy_work(void *handler)
3557 struct ptlrpc_request *req = handler;
3560 ptlrpc_req_finished(req);
3562 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3564 int ptlrpcd_queue_work(void *handler)
3566 struct ptlrpc_request *req = handler;
3569 * Check if the req is already being queued.
3571 * Here comes a trick: it lacks a way of checking if a req is being
3572 * processed reliably in ptlrpc. Here I have to use refcount of req
3573 * for this purpose. This is okay because the caller should use this
3574 * req as opaque data. - Jinshan
3576 LASSERT(atomic_read(&req->rq_refcount) > 0);
3577 if (atomic_inc_return(&req->rq_refcount) == 2)
3578 ptlrpcd_add_work_req(req);
3581 EXPORT_SYMBOL(ptlrpcd_queue_work);