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 <lnet/lib-lnet.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
48 #include "ptlrpc_internal.h"
50 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
51 struct page *page, int pageoffset,
54 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
57 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
58 struct page *page, int pageoffset,
61 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
64 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
68 for (i = 0; i < desc->bd_iov_count ; i++)
69 put_page(desc->bd_vec[i].bv_page);
72 static int ptlrpc_prep_bulk_frag_pages(struct ptlrpc_bulk_desc *desc,
75 unsigned int offset = (unsigned long)frag & ~PAGE_MASK;
79 int page_len = min_t(unsigned int, PAGE_SIZE - offset,
81 unsigned long vaddr = (unsigned long)frag;
83 ptlrpc_prep_bulk_page_nopin(desc,
84 lnet_kvaddr_to_page(vaddr),
94 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
95 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
96 .release_frags = ptlrpc_release_bulk_page_pin,
98 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
100 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
101 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
102 .release_frags = ptlrpc_release_bulk_noop,
103 .add_iov_frag = ptlrpc_prep_bulk_frag_pages,
105 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
107 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
108 static int ptlrpcd_check_work(struct ptlrpc_request *req);
109 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
112 * Initialize passed in client structure \a cl.
114 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
115 struct ptlrpc_client *cl)
117 cl->cli_request_portal = req_portal;
118 cl->cli_reply_portal = rep_portal;
121 EXPORT_SYMBOL(ptlrpc_init_client);
124 * Return PortalRPC connection for remore uud \a uuid
126 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
127 lnet_nid_t nid4refnet)
129 struct ptlrpc_connection *c;
131 struct lnet_process_id peer;
135 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
136 * before accessing its values.
138 /* coverity[uninit_use_in_call] */
139 peer.nid = nid4refnet;
140 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
142 CNETERR("cannot find peer %s!\n", uuid->uuid);
146 c = ptlrpc_connection_get(peer, self, uuid);
148 memcpy(c->c_remote_uuid.uuid,
149 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
152 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
158 * Allocate and initialize new bulk descriptor on the sender.
159 * Returns pointer to the descriptor or NULL on error.
161 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
162 unsigned int max_brw,
163 enum ptlrpc_bulk_op_type type,
165 const struct ptlrpc_bulk_frag_ops *ops)
167 struct ptlrpc_bulk_desc *desc;
170 LASSERT(ops->add_kiov_frag != NULL);
172 if (max_brw > PTLRPC_BULK_OPS_COUNT)
175 if (nfrags > LNET_MAX_IOV * max_brw)
182 OBD_ALLOC_LARGE(desc->bd_vec,
183 nfrags * sizeof(*desc->bd_vec));
187 spin_lock_init(&desc->bd_lock);
188 init_waitqueue_head(&desc->bd_waitq);
189 desc->bd_max_iov = nfrags;
190 desc->bd_iov_count = 0;
191 desc->bd_portal = portal;
192 desc->bd_type = type;
193 desc->bd_md_count = 0;
194 desc->bd_nob_last = LNET_MTU;
195 desc->bd_frag_ops = ops;
196 LASSERT(max_brw > 0);
197 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
199 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
200 * node. Negotiated ocd_brw_size will always be <= this number.
202 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
203 LNetInvalidateMDHandle(&desc->bd_mds[i]);
212 * Prepare bulk descriptor for specified outgoing request \a req that
213 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
214 * the bulk to be sent. Used on client-side.
215 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
218 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
220 unsigned int max_brw,
223 const struct ptlrpc_bulk_frag_ops
226 struct obd_import *imp = req->rq_import;
227 struct ptlrpc_bulk_desc *desc;
230 LASSERT(ptlrpc_is_bulk_op_passive(type));
232 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
236 desc->bd_import = class_import_get(imp);
239 desc->bd_cbid.cbid_fn = client_bulk_callback;
240 desc->bd_cbid.cbid_arg = desc;
242 /* This makes req own desc, and free it when she frees herself */
247 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
249 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
250 struct page *page, int pageoffset, int len,
253 struct bio_vec *kiov;
255 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
256 LASSERT(page != NULL);
257 LASSERT(pageoffset >= 0);
259 LASSERT(pageoffset + len <= PAGE_SIZE);
261 kiov = &desc->bd_vec[desc->bd_iov_count];
263 if (((desc->bd_iov_count % LNET_MAX_IOV) == 0) ||
264 ((desc->bd_nob_last + len) > LNET_MTU)) {
265 desc->bd_mds_off[desc->bd_md_count] = desc->bd_iov_count;
267 desc->bd_nob_last = 0;
268 LASSERT(desc->bd_md_count <= PTLRPC_BULK_OPS_COUNT);
271 desc->bd_nob_last += len;
277 kiov->bv_page = page;
278 kiov->bv_offset = pageoffset;
281 desc->bd_iov_count++;
283 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
285 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
289 LASSERT(desc != NULL);
290 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
291 LASSERT(desc->bd_refs == 0); /* network hands off */
292 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
293 LASSERT(desc->bd_frag_ops != NULL);
295 sptlrpc_enc_pool_put_pages(desc);
298 class_export_put(desc->bd_export);
300 class_import_put(desc->bd_import);
302 if (desc->bd_frag_ops->release_frags != NULL)
303 desc->bd_frag_ops->release_frags(desc);
305 OBD_FREE_LARGE(desc->bd_vec,
306 desc->bd_max_iov * sizeof(*desc->bd_vec));
310 EXPORT_SYMBOL(ptlrpc_free_bulk);
313 * Set server timelimit for this req, i.e. how long are we willing to wait
314 * for reply before timing out this request.
316 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
318 LASSERT(req->rq_import);
321 /* non-AT settings */
323 * \a imp_server_timeout means this is reverse import and
324 * we send (currently only) ASTs to the client and cannot afford
325 * to wait too long for the reply, otherwise the other client
326 * (because of which we are sending this request) would
327 * timeout waiting for us
329 req->rq_timeout = req->rq_import->imp_server_timeout ?
330 obd_timeout / 2 : obd_timeout;
332 struct imp_at *at = &req->rq_import->imp_at;
336 idx = import_at_get_index(req->rq_import,
337 req->rq_request_portal);
338 serv_est = at_get(&at->iat_service_estimate[idx]);
340 * Currently a 32 bit value is sent over the
341 * wire for rq_timeout so please don't change this
342 * to time64_t. The work for LU-1158 will in time
343 * replace rq_timeout with a 64 bit nanosecond value
345 req->rq_timeout = at_est2timeout(serv_est);
348 * We could get even fancier here, using history to predict increased
351 * Let the server know what this RPC timeout is by putting it in the
354 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
356 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
358 /* Adjust max service estimate based on server value */
359 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
366 LASSERT(req->rq_import);
367 at = &req->rq_import->imp_at;
369 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
371 * max service estimates are tracked on the server side,
372 * so just keep minimal history here
374 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
377 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
378 req->rq_import->imp_obd->obd_name,
379 req->rq_request_portal,
380 oldse, at_get(&at->iat_service_estimate[idx]));
383 /* Expected network latency per remote node (secs) */
384 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
386 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
389 /* Adjust expected network latency */
390 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
391 timeout_t service_timeout)
393 time64_t now = ktime_get_real_seconds();
398 LASSERT(req->rq_import);
400 if (service_timeout > now - req->rq_sent + 3) {
402 * b=16408, however, this can also happen if early reply
403 * is lost and client RPC is expired and resent, early reply
404 * or reply of original RPC can still be fit in reply buffer
405 * of resent RPC, now client is measuring time from the
406 * resent time, but server sent back service time of original
409 CDEBUG_LIMIT((lustre_msg_get_flags(req->rq_reqmsg) &
410 MSG_RESENT) ? D_ADAPTTO : D_WARNING,
411 "Reported service time %u > total measured time %lld\n",
412 service_timeout, now - req->rq_sent);
416 /* Network latency is total time less server processing time,
419 nl = max_t(timeout_t, now - req->rq_sent - service_timeout, 0) + 1;
420 at = &req->rq_import->imp_at;
422 oldnl = at_measured(&at->iat_net_latency, nl);
425 "The network latency for %s (nid %s) has changed from %d to %d\n",
426 req->rq_import->imp_obd->obd_name,
427 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
428 oldnl, at_get(&at->iat_net_latency));
431 static int unpack_reply(struct ptlrpc_request *req)
435 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
436 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
438 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
444 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
446 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
454 * Handle an early reply message, called with the rq_lock held.
455 * If anything goes wrong just ignore it - same as if it never happened
457 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
458 __must_hold(&req->rq_lock)
460 struct ptlrpc_request *early_req;
461 timeout_t service_timeout;
467 spin_unlock(&req->rq_lock);
469 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
471 spin_lock(&req->rq_lock);
475 rc = unpack_reply(early_req);
477 sptlrpc_cli_finish_early_reply(early_req);
478 spin_lock(&req->rq_lock);
483 * Use new timeout value just to adjust the local value for this
484 * request, don't include it into at_history. It is unclear yet why
485 * service time increased and should it be counted or skipped, e.g.
486 * that can be recovery case or some error or server, the real reply
487 * will add all new data if it is worth to add.
489 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
490 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
492 /* Network latency can be adjusted, it is pure network delays */
493 service_timeout = lustre_msg_get_service_timeout(early_req->rq_repmsg);
494 ptlrpc_at_adj_net_latency(req, service_timeout);
496 sptlrpc_cli_finish_early_reply(early_req);
498 spin_lock(&req->rq_lock);
499 olddl = req->rq_deadline;
501 * server assumes it now has rq_timeout from when the request
502 * arrived, so the client should give it at least that long.
503 * since we don't know the arrival time we'll use the original
506 req->rq_deadline = req->rq_sent + req->rq_timeout +
507 ptlrpc_at_get_net_latency(req);
509 /* The below message is checked in replay-single.sh test_65{a,b} */
510 /* The below message is checked in sanity-{gss,krb5} test_8 */
511 DEBUG_REQ(D_ADAPTTO, req,
512 "Early reply #%d, new deadline in %llds (%llds)",
514 req->rq_deadline - ktime_get_real_seconds(),
515 req->rq_deadline - olddl);
520 static struct kmem_cache *request_cache;
522 int ptlrpc_request_cache_init(void)
524 request_cache = kmem_cache_create("ptlrpc_cache",
525 sizeof(struct ptlrpc_request),
526 0, SLAB_HWCACHE_ALIGN, NULL);
527 return request_cache ? 0 : -ENOMEM;
530 void ptlrpc_request_cache_fini(void)
532 kmem_cache_destroy(request_cache);
535 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
537 struct ptlrpc_request *req;
539 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
543 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
545 OBD_SLAB_FREE_PTR(req, request_cache);
549 * Wind down request pool \a pool.
550 * Frees all requests from the pool too
552 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
554 struct ptlrpc_request *req;
556 LASSERT(pool != NULL);
558 spin_lock(&pool->prp_lock);
559 while ((req = list_first_entry_or_null(&pool->prp_req_list,
560 struct ptlrpc_request,
562 list_del(&req->rq_list);
563 LASSERT(req->rq_reqbuf);
564 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
565 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
566 ptlrpc_request_cache_free(req);
568 spin_unlock(&pool->prp_lock);
569 OBD_FREE(pool, sizeof(*pool));
571 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
574 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
576 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
581 while (size < pool->prp_rq_size)
584 LASSERTF(list_empty(&pool->prp_req_list) ||
585 size == pool->prp_rq_size,
586 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
587 pool->prp_rq_size, size);
589 pool->prp_rq_size = size;
590 for (i = 0; i < num_rq; i++) {
591 struct ptlrpc_request *req;
592 struct lustre_msg *msg;
594 req = ptlrpc_request_cache_alloc(GFP_NOFS);
597 OBD_ALLOC_LARGE(msg, size);
599 ptlrpc_request_cache_free(req);
602 req->rq_reqbuf = msg;
603 req->rq_reqbuf_len = size;
605 spin_lock(&pool->prp_lock);
606 list_add_tail(&req->rq_list, &pool->prp_req_list);
607 spin_unlock(&pool->prp_lock);
611 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
614 * Create and initialize new request pool with given attributes:
615 * \a num_rq - initial number of requests to create for the pool
616 * \a msgsize - maximum message size possible for requests in thid pool
617 * \a populate_pool - function to be called when more requests need to be added
619 * Returns pointer to newly created pool or NULL on error.
621 struct ptlrpc_request_pool *
622 ptlrpc_init_rq_pool(int num_rq, int msgsize,
623 int (*populate_pool)(struct ptlrpc_request_pool *, int))
625 struct ptlrpc_request_pool *pool;
632 * Request next power of two for the allocation, because internally
633 * kernel would do exactly this
635 spin_lock_init(&pool->prp_lock);
636 INIT_LIST_HEAD(&pool->prp_req_list);
637 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
638 pool->prp_populate = populate_pool;
640 populate_pool(pool, num_rq);
644 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
647 * Fetches one request from pool \a pool
649 static struct ptlrpc_request *
650 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
652 struct ptlrpc_request *request;
653 struct lustre_msg *reqbuf;
658 spin_lock(&pool->prp_lock);
661 * See if we have anything in a pool, and bail out if nothing,
662 * in writeout path, where this matters, this is safe to do, because
663 * nothing is lost in this case, and when some in-flight requests
664 * complete, this code will be called again.
666 if (unlikely(list_empty(&pool->prp_req_list))) {
667 spin_unlock(&pool->prp_lock);
671 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
673 list_del_init(&request->rq_list);
674 spin_unlock(&pool->prp_lock);
676 LASSERT(request->rq_reqbuf);
677 LASSERT(request->rq_pool);
679 reqbuf = request->rq_reqbuf;
680 memset(request, 0, sizeof(*request));
681 request->rq_reqbuf = reqbuf;
682 request->rq_reqbuf_len = pool->prp_rq_size;
683 request->rq_pool = pool;
689 * Returns freed \a request to pool.
691 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
693 struct ptlrpc_request_pool *pool = request->rq_pool;
695 spin_lock(&pool->prp_lock);
696 LASSERT(list_empty(&request->rq_list));
697 LASSERT(!request->rq_receiving_reply);
698 list_add_tail(&request->rq_list, &pool->prp_req_list);
699 spin_unlock(&pool->prp_lock);
702 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
704 struct obd_import *imp = req->rq_import;
705 struct ptlrpc_request *iter;
707 assert_spin_locked(&imp->imp_lock);
708 LASSERT(list_empty(&req->rq_unreplied_list));
710 /* unreplied list is sorted by xid in ascending order */
711 list_for_each_entry_reverse(iter, &imp->imp_unreplied_list,
713 LASSERT(req->rq_xid != iter->rq_xid);
714 if (req->rq_xid < iter->rq_xid)
716 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
719 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
722 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
724 req->rq_xid = ptlrpc_next_xid();
725 ptlrpc_add_unreplied(req);
728 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
730 spin_lock(&req->rq_import->imp_lock);
731 ptlrpc_assign_next_xid_nolock(req);
732 spin_unlock(&req->rq_import->imp_lock);
735 static atomic64_t ptlrpc_last_xid;
737 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
739 spin_lock(&req->rq_import->imp_lock);
740 list_del_init(&req->rq_unreplied_list);
741 ptlrpc_assign_next_xid_nolock(req);
742 spin_unlock(&req->rq_import->imp_lock);
743 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
746 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
748 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
752 opc = lustre_msg_get_opc(req->rq_reqmsg);
753 tag = obd_get_mod_rpc_slot(cli, opc);
754 lustre_msg_set_tag(req->rq_reqmsg, tag);
755 ptlrpc_reassign_next_xid(req);
757 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
759 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
761 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
764 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
765 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
767 obd_put_mod_rpc_slot(cli, opc, tag);
770 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
772 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
773 __u32 version, int opcode, char **bufs,
774 struct ptlrpc_cli_ctx *ctx)
777 struct obd_import *imp;
783 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
784 imp = request->rq_import;
785 lengths = request->rq_pill.rc_area[RCL_CLIENT];
788 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
790 rc = sptlrpc_req_get_ctx(request);
794 sptlrpc_req_set_flavor(request, opcode);
796 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
801 lustre_msg_add_version(request->rq_reqmsg, version);
802 request->rq_send_state = LUSTRE_IMP_FULL;
803 request->rq_type = PTL_RPC_MSG_REQUEST;
805 request->rq_req_cbid.cbid_fn = request_out_callback;
806 request->rq_req_cbid.cbid_arg = request;
808 request->rq_reply_cbid.cbid_fn = reply_in_callback;
809 request->rq_reply_cbid.cbid_arg = request;
811 request->rq_reply_deadline = 0;
812 request->rq_bulk_deadline = 0;
813 request->rq_req_deadline = 0;
814 request->rq_phase = RQ_PHASE_NEW;
815 request->rq_next_phase = RQ_PHASE_UNDEFINED;
817 request->rq_request_portal = imp->imp_client->cli_request_portal;
818 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
820 ptlrpc_at_set_req_timeout(request);
822 lustre_msg_set_opc(request->rq_reqmsg, opcode);
824 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
825 if (cfs_fail_val == opcode) {
826 time64_t *fail_t = NULL, *fail2_t = NULL;
828 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
829 fail_t = &request->rq_bulk_deadline;
830 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
831 fail_t = &request->rq_reply_deadline;
832 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
833 fail_t = &request->rq_req_deadline;
834 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
835 fail_t = &request->rq_reply_deadline;
836 fail2_t = &request->rq_bulk_deadline;
837 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
838 time64_t now = ktime_get_real_seconds();
839 u64 xid = ((u64)now >> 4) << 24;
841 atomic64_set(&ptlrpc_last_xid, xid);
845 *fail_t = ktime_get_real_seconds() +
846 PTLRPC_REQ_LONG_UNLINK;
849 *fail2_t = ktime_get_real_seconds() +
850 PTLRPC_REQ_LONG_UNLINK;
853 * The RPC is infected, let the test to change the
856 msleep(4 * MSEC_PER_SEC);
859 ptlrpc_assign_next_xid(request);
864 LASSERT(!request->rq_pool);
865 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
867 atomic_dec(&imp->imp_reqs);
868 class_import_put(imp);
872 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
875 * Pack request buffers for network transfer, performing necessary encryption
876 * steps if necessary.
878 int ptlrpc_request_pack(struct ptlrpc_request *request,
879 __u32 version, int opcode)
881 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
883 EXPORT_SYMBOL(ptlrpc_request_pack);
886 * Helper function to allocate new request on import \a imp
887 * and possibly using existing request from pool \a pool if provided.
888 * Returns allocated request structure with import field filled or
892 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
893 struct ptlrpc_request_pool *pool)
895 struct ptlrpc_request *request = NULL;
897 request = ptlrpc_request_cache_alloc(GFP_NOFS);
899 if (!request && pool)
900 request = ptlrpc_prep_req_from_pool(pool);
903 ptlrpc_cli_req_init(request);
905 LASSERTF((unsigned long)imp > 0x1000, "%p\n", imp);
906 LASSERT(imp != LP_POISON);
907 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
909 LASSERT(imp->imp_client != LP_POISON);
911 request->rq_import = class_import_get(imp);
912 atomic_inc(&imp->imp_reqs);
914 CERROR("request allocation out of memory\n");
920 static int ptlrpc_reconnect_if_idle(struct obd_import *imp)
925 * initiate connection if needed when the import has been
926 * referenced by the new request to avoid races with disconnect.
927 * serialize this check against conditional state=IDLE
928 * in ptlrpc_disconnect_idle_interpret()
930 spin_lock(&imp->imp_lock);
931 if (imp->imp_state == LUSTRE_IMP_IDLE) {
932 imp->imp_generation++;
933 imp->imp_initiated_at = imp->imp_generation;
934 imp->imp_state = LUSTRE_IMP_NEW;
936 /* connect_import_locked releases imp_lock */
937 rc = ptlrpc_connect_import_locked(imp);
940 ptlrpc_pinger_add_import(imp);
942 spin_unlock(&imp->imp_lock);
948 * Helper function for creating a request.
949 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
950 * buffer structures according to capsule template \a format.
951 * Returns allocated request structure pointer or NULL on error.
953 static struct ptlrpc_request *
954 ptlrpc_request_alloc_internal(struct obd_import *imp,
955 struct ptlrpc_request_pool *pool,
956 const struct req_format *format)
958 struct ptlrpc_request *request;
960 request = __ptlrpc_request_alloc(imp, pool);
964 /* don't make expensive check for idling connection
965 * if it's already connected */
966 if (unlikely(imp->imp_state != LUSTRE_IMP_FULL)) {
967 if (ptlrpc_reconnect_if_idle(imp) < 0) {
968 atomic_dec(&imp->imp_reqs);
969 ptlrpc_request_free(request);
974 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
975 req_capsule_set(&request->rq_pill, format);
980 * Allocate new request structure for import \a imp and initialize its
981 * buffer structure according to capsule template \a format.
983 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
984 const struct req_format *format)
986 return ptlrpc_request_alloc_internal(imp, NULL, format);
988 EXPORT_SYMBOL(ptlrpc_request_alloc);
991 * Allocate new request structure for import \a imp from pool \a pool and
992 * initialize its buffer structure according to capsule template \a format.
994 struct ptlrpc_request *
995 ptlrpc_request_alloc_pool(struct obd_import *imp,
996 struct ptlrpc_request_pool *pool,
997 const struct req_format *format)
999 return ptlrpc_request_alloc_internal(imp, pool, format);
1001 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
1004 * For requests not from pool, free memory of the request structure.
1005 * For requests obtained from a pool earlier, return request back to pool.
1007 void ptlrpc_request_free(struct ptlrpc_request *request)
1009 if (request->rq_pool)
1010 __ptlrpc_free_req_to_pool(request);
1012 ptlrpc_request_cache_free(request);
1014 EXPORT_SYMBOL(ptlrpc_request_free);
1017 * Allocate new request for operatione \a opcode and immediatelly pack it for
1019 * Only used for simple requests like OBD_PING where the only important
1020 * part of the request is operation itself.
1021 * Returns allocated request or NULL on error.
1023 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1024 const struct req_format *format,
1025 __u32 version, int opcode)
1027 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1031 rc = ptlrpc_request_pack(req, version, opcode);
1033 ptlrpc_request_free(req);
1039 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1042 * Allocate and initialize new request set structure on the current CPT.
1043 * Returns a pointer to the newly allocated set structure or NULL on error.
1045 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1047 struct ptlrpc_request_set *set;
1051 cpt = cfs_cpt_current(cfs_cpt_tab, 0);
1052 OBD_CPT_ALLOC(set, cfs_cpt_tab, cpt, sizeof(*set));
1055 atomic_set(&set->set_refcount, 1);
1056 INIT_LIST_HEAD(&set->set_requests);
1057 init_waitqueue_head(&set->set_waitq);
1058 atomic_set(&set->set_new_count, 0);
1059 atomic_set(&set->set_remaining, 0);
1060 spin_lock_init(&set->set_new_req_lock);
1061 INIT_LIST_HEAD(&set->set_new_requests);
1062 set->set_max_inflight = UINT_MAX;
1063 set->set_producer = NULL;
1064 set->set_producer_arg = NULL;
1069 EXPORT_SYMBOL(ptlrpc_prep_set);
1072 * Allocate and initialize new request set structure with flow control
1073 * extension. This extension allows to control the number of requests in-flight
1074 * for the whole set. A callback function to generate requests must be provided
1075 * and the request set will keep the number of requests sent over the wire to
1077 * Returns a pointer to the newly allocated set structure or NULL on error.
1079 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1083 struct ptlrpc_request_set *set;
1085 set = ptlrpc_prep_set();
1089 set->set_max_inflight = max;
1090 set->set_producer = func;
1091 set->set_producer_arg = arg;
1097 * Wind down and free request set structure previously allocated with
1099 * Ensures that all requests on the set have completed and removes
1100 * all requests from the request list in a set.
1101 * If any unsent request happen to be on the list, pretends that they got
1102 * an error in flight and calls their completion handler.
1104 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1106 struct ptlrpc_request *req;
1112 /* Requests on the set should either all be completed, or all be new */
1113 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1114 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1115 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
1116 LASSERT(req->rq_phase == expected_phase);
1120 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1121 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1122 atomic_read(&set->set_remaining), n);
1124 while ((req = list_first_entry_or_null(&set->set_requests,
1125 struct ptlrpc_request,
1127 list_del_init(&req->rq_set_chain);
1129 LASSERT(req->rq_phase == expected_phase);
1131 if (req->rq_phase == RQ_PHASE_NEW) {
1132 ptlrpc_req_interpret(NULL, req, -EBADR);
1133 atomic_dec(&set->set_remaining);
1136 spin_lock(&req->rq_lock);
1138 req->rq_invalid_rqset = 0;
1139 spin_unlock(&req->rq_lock);
1141 ptlrpc_req_finished(req);
1144 LASSERT(atomic_read(&set->set_remaining) == 0);
1146 ptlrpc_reqset_put(set);
1149 EXPORT_SYMBOL(ptlrpc_set_destroy);
1152 * Add a new request to the general purpose request set.
1153 * Assumes request reference from the caller.
1155 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1156 struct ptlrpc_request *req)
1158 if (set == PTLRPCD_SET) {
1159 ptlrpcd_add_req(req);
1163 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1164 LASSERT(list_empty(&req->rq_set_chain));
1166 if (req->rq_allow_intr)
1167 set->set_allow_intr = 1;
1169 /* The set takes over the caller's request reference */
1170 list_add_tail(&req->rq_set_chain, &set->set_requests);
1172 atomic_inc(&set->set_remaining);
1173 req->rq_queued_time = ktime_get_seconds();
1176 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1178 if (set->set_producer)
1180 * If the request set has a producer callback, the RPC must be
1181 * sent straight away
1183 ptlrpc_send_new_req(req);
1185 EXPORT_SYMBOL(ptlrpc_set_add_req);
1188 * Add a request to a request with dedicated server thread
1189 * and wake the thread to make any necessary processing.
1190 * Currently only used for ptlrpcd.
1192 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1193 struct ptlrpc_request *req)
1195 struct ptlrpc_request_set *set = pc->pc_set;
1198 LASSERT(req->rq_set == NULL);
1199 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1201 spin_lock(&set->set_new_req_lock);
1203 * The set takes over the caller's request reference.
1206 req->rq_queued_time = ktime_get_seconds();
1207 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1208 count = atomic_inc_return(&set->set_new_count);
1209 spin_unlock(&set->set_new_req_lock);
1211 /* Only need to call wakeup once for the first entry. */
1213 wake_up(&set->set_waitq);
1216 * XXX: It maybe unnecessary to wakeup all the partners. But to
1217 * guarantee the async RPC can be processed ASAP, we have
1218 * no other better choice. It maybe fixed in future.
1220 for (i = 0; i < pc->pc_npartners; i++)
1221 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1226 * Based on the current state of the import, determine if the request
1227 * can be sent, is an error, or should be delayed.
1229 * Returns true if this request should be delayed. If false, and
1230 * *status is set, then the request can not be sent and *status is the
1231 * error code. If false and status is 0, then request can be sent.
1233 * The imp->imp_lock must be held.
1235 static int ptlrpc_import_delay_req(struct obd_import *imp,
1236 struct ptlrpc_request *req, int *status)
1244 if (req->rq_ctx_init || req->rq_ctx_fini) {
1245 /* always allow ctx init/fini rpc go through */
1246 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1247 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1249 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1250 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1253 * pings or MDS-equivalent STATFS may safely
1256 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1257 D_HA : D_ERROR, req, "IMP_CLOSED");
1259 } else if (ptlrpc_send_limit_expired(req)) {
1260 /* probably doesn't need to be a D_ERROR afterinitial testing */
1261 DEBUG_REQ(D_HA, req, "send limit expired");
1262 *status = -ETIMEDOUT;
1263 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1264 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1265 ;/* allow CONNECT even if import is invalid */
1266 if (atomic_read(&imp->imp_inval_count) != 0) {
1267 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1270 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1271 if (!imp->imp_deactive)
1272 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1273 *status = -ESHUTDOWN; /* b=12940 */
1274 } else if (req->rq_import_generation != imp->imp_generation) {
1275 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1277 } else if (req->rq_send_state != imp->imp_state) {
1278 /* invalidate in progress - any requests should be drop */
1279 if (atomic_read(&imp->imp_inval_count) != 0) {
1280 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1282 } else if (req->rq_no_delay &&
1283 imp->imp_generation != imp->imp_initiated_at) {
1284 /* ignore nodelay for requests initiating connections */
1286 } else if (req->rq_allow_replay &&
1287 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1288 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1289 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1290 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1291 DEBUG_REQ(D_HA, req, "allow during recovery");
1301 * Decide if the error message should be printed to the console or not.
1302 * Makes its decision based on request type, status, and failure frequency.
1304 * \param[in] req request that failed and may need a console message
1306 * \retval false if no message should be printed
1307 * \retval true if console message should be printed
1309 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1311 LASSERT(req->rq_reqmsg != NULL);
1313 /* Suppress particular reconnect errors which are to be expected. */
1314 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1315 /* Suppress timed out reconnect requests */
1316 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1321 * Suppress most unavailable/again reconnect requests, but
1322 * print occasionally so it is clear client is trying to
1323 * connect to a server where no target is running.
1325 if ((err == -ENODEV || err == -EAGAIN) &&
1326 req->rq_import->imp_conn_cnt % 30 != 20)
1330 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1331 /* -EAGAIN is normal when using POSIX flocks */
1334 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1335 (req->rq_xid & 0xf) != 10)
1336 /* Suppress most ping requests, they may fail occasionally */
1343 * Check request processing status.
1344 * Returns the status.
1346 static int ptlrpc_check_status(struct ptlrpc_request *req)
1351 rc = lustre_msg_get_status(req->rq_repmsg);
1352 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1353 struct obd_import *imp = req->rq_import;
1354 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1355 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1357 if (ptlrpc_console_allow(req, opc, rc))
1358 LCONSOLE_ERROR_MSG(0x11,
1359 "%s: operation %s to node %s failed: rc = %d\n",
1360 imp->imp_obd->obd_name,
1362 libcfs_nid2str(nid), rc);
1363 RETURN(rc < 0 ? rc : -EINVAL);
1367 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1373 * save pre-versions of objects into request for replay.
1374 * Versions are obtained from server reply.
1377 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1379 struct lustre_msg *repmsg = req->rq_repmsg;
1380 struct lustre_msg *reqmsg = req->rq_reqmsg;
1381 __u64 *versions = lustre_msg_get_versions(repmsg);
1384 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1388 lustre_msg_set_versions(reqmsg, versions);
1389 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1390 versions[0], versions[1]);
1395 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1397 struct ptlrpc_request *req;
1399 assert_spin_locked(&imp->imp_lock);
1400 if (list_empty(&imp->imp_unreplied_list))
1403 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1405 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1407 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1408 imp->imp_known_replied_xid = req->rq_xid - 1;
1410 return req->rq_xid - 1;
1414 * Callback function called when client receives RPC reply for \a req.
1415 * Returns 0 on success or error code.
1416 * The return alue would be assigned to req->rq_status by the caller
1417 * as request processing status.
1418 * This function also decides if the request needs to be saved for later replay.
1420 static int after_reply(struct ptlrpc_request *req)
1422 struct obd_import *imp = req->rq_import;
1423 struct obd_device *obd = req->rq_import->imp_obd;
1430 LASSERT(obd != NULL);
1431 /* repbuf must be unlinked */
1432 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1434 if (req->rq_reply_truncated) {
1435 if (ptlrpc_no_resend(req)) {
1436 DEBUG_REQ(D_ERROR, req,
1437 "reply buffer overflow, expected=%d, actual size=%d",
1438 req->rq_nob_received, req->rq_repbuf_len);
1442 sptlrpc_cli_free_repbuf(req);
1444 * Pass the required reply buffer size (include
1445 * space for early reply).
1446 * NB: no need to roundup because alloc_repbuf
1449 req->rq_replen = req->rq_nob_received;
1450 req->rq_nob_received = 0;
1451 spin_lock(&req->rq_lock);
1453 spin_unlock(&req->rq_lock);
1457 work_start = ktime_get_real();
1458 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1461 * NB Until this point, the whole of the incoming message,
1462 * including buflens, status etc is in the sender's byte order.
1464 rc = sptlrpc_cli_unwrap_reply(req);
1466 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1471 * Security layer unwrap might ask resend this request.
1476 rc = unpack_reply(req);
1480 /* retry indefinitely on EINPROGRESS */
1481 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1482 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1483 time64_t now = ktime_get_real_seconds();
1485 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1486 D_RPCTRACE, req, "resending request on EINPROGRESS");
1487 spin_lock(&req->rq_lock);
1489 spin_unlock(&req->rq_lock);
1490 req->rq_nr_resend++;
1492 /* Readjust the timeout for current conditions */
1493 ptlrpc_at_set_req_timeout(req);
1495 * delay resend to give a chance to the server to get ready.
1496 * The delay is increased by 1s on every resend and is capped to
1497 * the current request timeout (i.e. obd_timeout if AT is off,
1498 * or AT service time x 125% + 5s, see at_est2timeout)
1500 if (req->rq_nr_resend > req->rq_timeout)
1501 req->rq_sent = now + req->rq_timeout;
1503 req->rq_sent = now + req->rq_nr_resend;
1505 /* Resend for EINPROGRESS will use a new XID */
1506 spin_lock(&imp->imp_lock);
1507 list_del_init(&req->rq_unreplied_list);
1508 spin_unlock(&imp->imp_lock);
1513 if (obd->obd_svc_stats) {
1514 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1516 ptlrpc_lprocfs_rpc_sent(req, timediff);
1519 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1520 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1521 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1522 lustre_msg_get_type(req->rq_repmsg));
1526 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1527 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1528 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1529 ptlrpc_at_adj_net_latency(req,
1530 lustre_msg_get_service_timeout(req->rq_repmsg));
1532 rc = ptlrpc_check_status(req);
1536 * Either we've been evicted, or the server has failed for
1537 * some reason. Try to reconnect, and if that fails, punt to
1540 if (ptlrpc_recoverable_error(rc)) {
1541 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1542 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1545 ptlrpc_request_handle_notconn(req);
1550 * Let's look if server sent slv. Do it only for RPC with
1553 ldlm_cli_update_pool(req);
1557 * Store transno in reqmsg for replay.
1559 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1560 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1561 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1564 if (imp->imp_replayable) {
1565 spin_lock(&imp->imp_lock);
1567 * No point in adding already-committed requests to the replay
1568 * list, we will just remove them immediately. b=9829
1570 if (req->rq_transno != 0 &&
1572 lustre_msg_get_last_committed(req->rq_repmsg) ||
1574 /** version recovery */
1575 ptlrpc_save_versions(req);
1576 ptlrpc_retain_replayable_request(req, imp);
1577 } else if (req->rq_commit_cb &&
1578 list_empty(&req->rq_replay_list)) {
1580 * NB: don't call rq_commit_cb if it's already on
1581 * rq_replay_list, ptlrpc_free_committed() will call
1582 * it later, see LU-3618 for details
1584 spin_unlock(&imp->imp_lock);
1585 req->rq_commit_cb(req);
1586 spin_lock(&imp->imp_lock);
1590 * Replay-enabled imports return commit-status information.
1592 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1593 if (likely(committed > imp->imp_peer_committed_transno))
1594 imp->imp_peer_committed_transno = committed;
1596 ptlrpc_free_committed(imp);
1598 if (!list_empty(&imp->imp_replay_list)) {
1599 struct ptlrpc_request *last;
1601 last = list_entry(imp->imp_replay_list.prev,
1602 struct ptlrpc_request,
1605 * Requests with rq_replay stay on the list even if no
1606 * commit is expected.
1608 if (last->rq_transno > imp->imp_peer_committed_transno)
1609 ptlrpc_pinger_commit_expected(imp);
1612 spin_unlock(&imp->imp_lock);
1619 * Helper function to send request \a req over the network for the first time
1620 * Also adjusts request phase.
1621 * Returns 0 on success or error code.
1623 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1625 struct obd_import *imp = req->rq_import;
1630 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1632 /* do not try to go further if there is not enough memory in enc_pool */
1633 if (req->rq_sent && req->rq_bulk)
1634 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1635 pool_is_at_full_capacity())
1638 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1639 (!req->rq_generation_set ||
1640 req->rq_import_generation == imp->imp_generation))
1643 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1645 spin_lock(&imp->imp_lock);
1647 LASSERT(req->rq_xid != 0);
1648 LASSERT(!list_empty(&req->rq_unreplied_list));
1650 if (!req->rq_generation_set)
1651 req->rq_import_generation = imp->imp_generation;
1653 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1654 spin_lock(&req->rq_lock);
1655 req->rq_waiting = 1;
1656 spin_unlock(&req->rq_lock);
1658 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1659 ptlrpc_import_state_name(req->rq_send_state),
1660 ptlrpc_import_state_name(imp->imp_state));
1661 LASSERT(list_empty(&req->rq_list));
1662 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1663 atomic_inc(&req->rq_import->imp_inflight);
1664 spin_unlock(&imp->imp_lock);
1669 spin_unlock(&imp->imp_lock);
1670 req->rq_status = rc;
1671 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1675 LASSERT(list_empty(&req->rq_list));
1676 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1677 atomic_inc(&req->rq_import->imp_inflight);
1680 * find the known replied XID from the unreplied list, CONNECT
1681 * and DISCONNECT requests are skipped to make the sanity check
1682 * on server side happy. see process_req_last_xid().
1684 * For CONNECT: Because replay requests have lower XID, it'll
1685 * break the sanity check if CONNECT bump the exp_last_xid on
1688 * For DISCONNECT: Since client will abort inflight RPC before
1689 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1690 * than the inflight RPC.
1692 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1693 min_xid = ptlrpc_known_replied_xid(imp);
1694 spin_unlock(&imp->imp_lock);
1696 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1698 lustre_msg_set_status(req->rq_reqmsg, current->pid);
1700 /* If the request to be sent is an LDLM callback, do not try to
1702 * An LDLM callback is sent by a server to a client in order to make
1703 * it release a lock, on a communication channel that uses a reverse
1704 * context. It cannot be refreshed on its own, as it is the 'reverse'
1705 * (server-side) representation of a client context.
1706 * We do not care if the reverse context is expired, and want to send
1707 * the LDLM callback anyway. Once the client receives the AST, it is
1708 * its job to refresh its own context if it has expired, hence
1709 * refreshing the associated reverse context on server side, before
1710 * being able to send the LDLM_CANCEL requested by the server.
1712 if (lustre_msg_get_opc(req->rq_reqmsg) != LDLM_BL_CALLBACK &&
1713 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_CP_CALLBACK &&
1714 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_GL_CALLBACK)
1715 rc = sptlrpc_req_refresh_ctx(req, 0);
1718 req->rq_status = rc;
1721 spin_lock(&req->rq_lock);
1722 req->rq_wait_ctx = 1;
1723 spin_unlock(&req->rq_lock);
1729 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1731 imp->imp_obd->obd_uuid.uuid,
1732 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1733 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1734 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1736 rc = ptl_send_rpc(req, 0);
1737 if (rc == -ENOMEM) {
1738 spin_lock(&imp->imp_lock);
1739 if (!list_empty(&req->rq_list)) {
1740 list_del_init(&req->rq_list);
1741 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1742 wake_up(&req->rq_import->imp_recovery_waitq);
1744 spin_unlock(&imp->imp_lock);
1745 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1749 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1751 spin_lock(&req->rq_lock);
1752 req->rq_net_err = 1;
1753 spin_unlock(&req->rq_lock);
1759 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1764 LASSERT(set->set_producer != NULL);
1766 remaining = atomic_read(&set->set_remaining);
1769 * populate the ->set_requests list with requests until we
1770 * reach the maximum number of RPCs in flight for this set
1772 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1773 rc = set->set_producer(set, set->set_producer_arg);
1774 if (rc == -ENOENT) {
1775 /* no more RPC to produce */
1776 set->set_producer = NULL;
1777 set->set_producer_arg = NULL;
1782 RETURN((atomic_read(&set->set_remaining) - remaining));
1786 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1787 * and no more replies are expected.
1788 * (it is possible to get less replies than requests sent e.g. due to timed out
1789 * requests or requests that we had trouble to send out)
1791 * NOTE: This function contains a potential schedule point (cond_resched()).
1793 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1795 struct ptlrpc_request *req, *next;
1796 LIST_HEAD(comp_reqs);
1797 int force_timer_recalc = 0;
1800 if (atomic_read(&set->set_remaining) == 0)
1803 list_for_each_entry_safe(req, next, &set->set_requests,
1805 struct obd_import *imp = req->rq_import;
1806 int unregistered = 0;
1810 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1811 list_move_tail(&req->rq_set_chain, &comp_reqs);
1816 * This schedule point is mainly for the ptlrpcd caller of this
1817 * function. Most ptlrpc sets are not long-lived and unbounded
1818 * in length, but at the least the set used by the ptlrpcd is.
1819 * Since the processing time is unbounded, we need to insert an
1820 * explicit schedule point to make the thread well-behaved.
1825 * If the caller requires to allow to be interpreted by force
1826 * and it has really been interpreted, then move the request
1827 * to RQ_PHASE_INTERPRET phase in spite of what the current
1830 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1831 req->rq_status = -EINTR;
1832 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1835 * Since it is interpreted and we have to wait for
1836 * the reply to be unlinked, then use sync mode.
1840 GOTO(interpret, req->rq_status);
1843 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1844 force_timer_recalc = 1;
1846 /* delayed send - skip */
1847 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1850 /* delayed resend - skip */
1851 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1852 req->rq_sent > ktime_get_real_seconds())
1855 if (!(req->rq_phase == RQ_PHASE_RPC ||
1856 req->rq_phase == RQ_PHASE_BULK ||
1857 req->rq_phase == RQ_PHASE_INTERPRET ||
1858 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1859 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1860 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1864 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1865 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1866 LASSERT(req->rq_next_phase != req->rq_phase);
1867 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1869 if (req->rq_req_deadline &&
1870 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1871 req->rq_req_deadline = 0;
1872 if (req->rq_reply_deadline &&
1873 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1874 req->rq_reply_deadline = 0;
1875 if (req->rq_bulk_deadline &&
1876 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1877 req->rq_bulk_deadline = 0;
1880 * Skip processing until reply is unlinked. We
1881 * can't return to pool before that and we can't
1882 * call interpret before that. We need to make
1883 * sure that all rdma transfers finished and will
1884 * not corrupt any data.
1886 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1887 ptlrpc_cli_wait_unlink(req))
1889 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1890 ptlrpc_client_bulk_active(req))
1894 * Turn fail_loc off to prevent it from looping
1897 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1898 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1901 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1902 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1907 * Move to next phase if reply was successfully
1910 ptlrpc_rqphase_move(req, req->rq_next_phase);
1913 if (req->rq_phase == RQ_PHASE_INTERPRET)
1914 GOTO(interpret, req->rq_status);
1917 * Note that this also will start async reply unlink.
1919 if (req->rq_net_err && !req->rq_timedout) {
1920 ptlrpc_expire_one_request(req, 1);
1923 * Check if we still need to wait for unlink.
1925 if (ptlrpc_cli_wait_unlink(req) ||
1926 ptlrpc_client_bulk_active(req))
1928 /* If there is no need to resend, fail it now. */
1929 if (req->rq_no_resend) {
1930 if (req->rq_status == 0)
1931 req->rq_status = -EIO;
1932 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1933 GOTO(interpret, req->rq_status);
1940 spin_lock(&req->rq_lock);
1941 req->rq_replied = 0;
1942 spin_unlock(&req->rq_lock);
1943 if (req->rq_status == 0)
1944 req->rq_status = -EIO;
1945 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1946 GOTO(interpret, req->rq_status);
1950 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1951 * so it sets rq_intr regardless of individual rpc
1952 * timeouts. The synchronous IO waiting path sets
1953 * rq_intr irrespective of whether ptlrpcd
1954 * has seen a timeout. Our policy is to only interpret
1955 * interrupted rpcs after they have timed out, so we
1956 * need to enforce that here.
1959 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1960 req->rq_wait_ctx)) {
1961 req->rq_status = -EINTR;
1962 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1963 GOTO(interpret, req->rq_status);
1966 if (req->rq_phase == RQ_PHASE_RPC) {
1967 if (req->rq_timedout || req->rq_resend ||
1968 req->rq_waiting || req->rq_wait_ctx) {
1971 if (!ptlrpc_unregister_reply(req, 1)) {
1972 ptlrpc_unregister_bulk(req, 1);
1976 spin_lock(&imp->imp_lock);
1977 if (ptlrpc_import_delay_req(imp, req,
1980 * put on delay list - only if we wait
1981 * recovery finished - before send
1983 list_move_tail(&req->rq_list,
1984 &imp->imp_delayed_list);
1985 spin_unlock(&imp->imp_lock);
1990 req->rq_status = status;
1991 ptlrpc_rqphase_move(req,
1992 RQ_PHASE_INTERPRET);
1993 spin_unlock(&imp->imp_lock);
1994 GOTO(interpret, req->rq_status);
1996 /* ignore on just initiated connections */
1997 if (ptlrpc_no_resend(req) &&
1998 !req->rq_wait_ctx &&
1999 imp->imp_generation !=
2000 imp->imp_initiated_at) {
2001 req->rq_status = -ENOTCONN;
2002 ptlrpc_rqphase_move(req,
2003 RQ_PHASE_INTERPRET);
2004 spin_unlock(&imp->imp_lock);
2005 GOTO(interpret, req->rq_status);
2008 /* don't resend too fast in case of network
2011 if (ktime_get_real_seconds() < (req->rq_sent + 1)
2012 && req->rq_net_err && req->rq_timedout) {
2014 DEBUG_REQ(D_INFO, req,
2015 "throttle request");
2016 /* Don't try to resend RPC right away
2017 * as it is likely it will fail again
2018 * and ptlrpc_check_set() will be
2019 * called again, keeping this thread
2020 * busy. Instead, wait for the next
2021 * timeout. Flag it as resend to
2022 * ensure we don't wait to long.
2025 spin_unlock(&imp->imp_lock);
2029 list_move_tail(&req->rq_list,
2030 &imp->imp_sending_list);
2032 spin_unlock(&imp->imp_lock);
2034 spin_lock(&req->rq_lock);
2035 req->rq_waiting = 0;
2036 spin_unlock(&req->rq_lock);
2038 if (req->rq_timedout || req->rq_resend) {
2040 * This is re-sending anyways,
2041 * let's mark req as resend.
2043 spin_lock(&req->rq_lock);
2045 spin_unlock(&req->rq_lock);
2048 * rq_wait_ctx is only touched by ptlrpcd,
2049 * so no lock is needed here.
2051 status = sptlrpc_req_refresh_ctx(req, 0);
2054 req->rq_status = status;
2055 spin_lock(&req->rq_lock);
2056 req->rq_wait_ctx = 0;
2057 spin_unlock(&req->rq_lock);
2058 force_timer_recalc = 1;
2060 spin_lock(&req->rq_lock);
2061 req->rq_wait_ctx = 1;
2062 spin_unlock(&req->rq_lock);
2067 spin_lock(&req->rq_lock);
2068 req->rq_wait_ctx = 0;
2069 spin_unlock(&req->rq_lock);
2073 * In any case, the previous bulk should be
2074 * cleaned up to prepare for the new sending
2077 !ptlrpc_unregister_bulk(req, 1))
2080 rc = ptl_send_rpc(req, 0);
2081 if (rc == -ENOMEM) {
2082 spin_lock(&imp->imp_lock);
2083 if (!list_empty(&req->rq_list))
2084 list_del_init(&req->rq_list);
2085 spin_unlock(&imp->imp_lock);
2086 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2090 DEBUG_REQ(D_HA, req,
2091 "send failed: rc = %d", rc);
2092 force_timer_recalc = 1;
2093 spin_lock(&req->rq_lock);
2094 req->rq_net_err = 1;
2095 spin_unlock(&req->rq_lock);
2098 /* need to reset the timeout */
2099 force_timer_recalc = 1;
2102 spin_lock(&req->rq_lock);
2104 if (ptlrpc_client_early(req)) {
2105 ptlrpc_at_recv_early_reply(req);
2106 spin_unlock(&req->rq_lock);
2110 /* Still waiting for a reply? */
2111 if (ptlrpc_client_recv(req)) {
2112 spin_unlock(&req->rq_lock);
2116 /* Did we actually receive a reply? */
2117 if (!ptlrpc_client_replied(req)) {
2118 spin_unlock(&req->rq_lock);
2122 spin_unlock(&req->rq_lock);
2125 * unlink from net because we are going to
2126 * swab in-place of reply buffer
2128 unregistered = ptlrpc_unregister_reply(req, 1);
2132 req->rq_status = after_reply(req);
2137 * If there is no bulk associated with this request,
2138 * then we're done and should let the interpreter
2139 * process the reply. Similarly if the RPC returned
2140 * an error, and therefore the bulk will never arrive.
2142 if (!req->rq_bulk || req->rq_status < 0) {
2143 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2144 GOTO(interpret, req->rq_status);
2147 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2150 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2151 if (ptlrpc_client_bulk_active(req))
2154 if (req->rq_bulk->bd_failure) {
2156 * The RPC reply arrived OK, but the bulk screwed
2157 * up! Dead weird since the server told us the RPC
2158 * was good after getting the REPLY for her GET or
2159 * the ACK for her PUT.
2161 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2162 req->rq_status = -EIO;
2165 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2168 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2171 * This moves to "unregistering" phase we need to wait for
2174 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2175 /* start async bulk unlink too */
2176 ptlrpc_unregister_bulk(req, 1);
2180 if (!ptlrpc_unregister_bulk(req, async))
2184 * When calling interpret receiving already should be
2187 LASSERT(!req->rq_receiving_reply);
2189 ptlrpc_req_interpret(env, req, req->rq_status);
2191 if (ptlrpcd_check_work(req)) {
2192 atomic_dec(&set->set_remaining);
2195 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2199 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2201 imp->imp_obd->obd_uuid.uuid,
2202 lustre_msg_get_status(req->rq_reqmsg),
2204 obd_import_nid2str(imp),
2205 lustre_msg_get_opc(req->rq_reqmsg),
2206 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2208 spin_lock(&imp->imp_lock);
2210 * Request already may be not on sending or delaying list. This
2211 * may happen in the case of marking it erroneous for the case
2212 * ptlrpc_import_delay_req(req, status) find it impossible to
2213 * allow sending this rpc and returns *status != 0.
2215 if (!list_empty(&req->rq_list)) {
2216 list_del_init(&req->rq_list);
2217 if (atomic_dec_and_test(&imp->imp_inflight))
2218 wake_up(&imp->imp_recovery_waitq);
2220 list_del_init(&req->rq_unreplied_list);
2221 spin_unlock(&imp->imp_lock);
2223 atomic_dec(&set->set_remaining);
2224 wake_up(&imp->imp_recovery_waitq);
2226 if (set->set_producer) {
2227 /* produce a new request if possible */
2228 if (ptlrpc_set_producer(set) > 0)
2229 force_timer_recalc = 1;
2232 * free the request that has just been completed
2233 * in order not to pollute set->set_requests
2235 list_del_init(&req->rq_set_chain);
2236 spin_lock(&req->rq_lock);
2238 req->rq_invalid_rqset = 0;
2239 spin_unlock(&req->rq_lock);
2241 /* record rq_status to compute the final status later */
2242 if (req->rq_status != 0)
2243 set->set_rc = req->rq_status;
2244 ptlrpc_req_finished(req);
2246 list_move_tail(&req->rq_set_chain, &comp_reqs);
2251 * move completed request at the head of list so it's easier for
2252 * caller to find them
2254 list_splice(&comp_reqs, &set->set_requests);
2256 /* If we hit an error, we want to recover promptly. */
2257 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2259 EXPORT_SYMBOL(ptlrpc_check_set);
2262 * Time out request \a req. is \a async_unlink is set, that means do not wait
2263 * until LNet actually confirms network buffer unlinking.
2264 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2266 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2268 struct obd_import *imp = req->rq_import;
2269 unsigned int debug_mask = D_RPCTRACE;
2273 spin_lock(&req->rq_lock);
2274 req->rq_timedout = 1;
2275 spin_unlock(&req->rq_lock);
2277 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2278 lustre_msg_get_status(req->rq_reqmsg)))
2279 debug_mask = D_WARNING;
2280 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2281 req->rq_net_err ? "failed due to network error" :
2282 ((req->rq_real_sent == 0 ||
2283 req->rq_real_sent < req->rq_sent ||
2284 req->rq_real_sent >= req->rq_deadline) ?
2285 "timed out for sent delay" : "timed out for slow reply"),
2286 req->rq_sent, req->rq_real_sent);
2288 if (imp && obd_debug_peer_on_timeout)
2289 LNetDebugPeer(imp->imp_connection->c_peer);
2291 ptlrpc_unregister_reply(req, async_unlink);
2292 ptlrpc_unregister_bulk(req, async_unlink);
2294 if (obd_dump_on_timeout)
2295 libcfs_debug_dumplog();
2298 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2302 atomic_inc(&imp->imp_timeouts);
2304 /* The DLM server doesn't want recovery run on its imports. */
2305 if (imp->imp_dlm_fake)
2309 * If this request is for recovery or other primordial tasks,
2310 * then error it out here.
2312 if (req->rq_ctx_init || req->rq_ctx_fini ||
2313 req->rq_send_state != LUSTRE_IMP_FULL ||
2314 imp->imp_obd->obd_no_recov) {
2315 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2316 ptlrpc_import_state_name(req->rq_send_state),
2317 ptlrpc_import_state_name(imp->imp_state));
2318 spin_lock(&req->rq_lock);
2319 req->rq_status = -ETIMEDOUT;
2321 spin_unlock(&req->rq_lock);
2326 * if a request can't be resent we can't wait for an answer after
2329 if (ptlrpc_no_resend(req)) {
2330 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2334 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2340 * Time out all uncompleted requests in request set pointed by \a data
2341 * This is called when a wait times out.
2343 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2345 struct ptlrpc_request *req;
2346 time64_t now = ktime_get_real_seconds();
2349 LASSERT(set != NULL);
2352 * A timeout expired. See which reqs it applies to...
2354 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2355 /* don't expire request waiting for context */
2356 if (req->rq_wait_ctx)
2359 /* Request in-flight? */
2360 if (!((req->rq_phase == RQ_PHASE_RPC &&
2361 !req->rq_waiting && !req->rq_resend) ||
2362 (req->rq_phase == RQ_PHASE_BULK)))
2365 if (req->rq_timedout || /* already dealt with */
2366 req->rq_deadline > now) /* not expired */
2370 * Deal with this guy. Do it asynchronously to not block
2373 ptlrpc_expire_one_request(req, 1);
2375 * Loops require that we resched once in a while to avoid
2376 * RCU stalls and a few other problems.
2384 * Interrupts (sets interrupted flag) all uncompleted requests in
2385 * a set \a data. This is called when a wait_event is interrupted
2388 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2390 struct ptlrpc_request *req;
2392 LASSERT(set != NULL);
2393 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2395 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2399 if (req->rq_phase != RQ_PHASE_RPC &&
2400 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2401 !req->rq_allow_intr)
2404 spin_lock(&req->rq_lock);
2406 spin_unlock(&req->rq_lock);
2411 * Get the smallest timeout in the set; this does NOT set a timeout.
2413 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2415 time64_t now = ktime_get_real_seconds();
2417 struct ptlrpc_request *req;
2421 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2422 /* Request in-flight? */
2423 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2424 (req->rq_phase == RQ_PHASE_BULK) ||
2425 (req->rq_phase == RQ_PHASE_NEW)))
2428 /* Already timed out. */
2429 if (req->rq_timedout)
2432 /* Waiting for ctx. */
2433 if (req->rq_wait_ctx)
2436 if (req->rq_phase == RQ_PHASE_NEW)
2437 deadline = req->rq_sent;
2438 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2439 deadline = req->rq_sent;
2441 deadline = req->rq_sent + req->rq_timeout;
2443 if (deadline <= now) /* actually expired already */
2444 timeout = 1; /* ASAP */
2445 else if (timeout == 0 || timeout > deadline - now)
2446 timeout = deadline - now;
2452 * Send all unset request from the set and then wait untill all
2453 * requests in the set complete (either get a reply, timeout, get an
2454 * error or otherwise be interrupted).
2455 * Returns 0 on success or error code otherwise.
2457 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2459 struct ptlrpc_request *req;
2464 if (set->set_producer)
2465 (void)ptlrpc_set_producer(set);
2467 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2468 if (req->rq_phase == RQ_PHASE_NEW)
2469 (void)ptlrpc_send_new_req(req);
2472 if (list_empty(&set->set_requests))
2476 timeout = ptlrpc_set_next_timeout(set);
2479 * wait until all complete, interrupted, or an in-flight
2482 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2485 if ((timeout == 0 && !signal_pending(current)) ||
2486 set->set_allow_intr) {
2488 * No requests are in-flight (ether timed out
2489 * or delayed), so we can allow interrupts.
2490 * We still want to block for a limited time,
2491 * so we allow interrupts during the timeout.
2493 rc = l_wait_event_abortable_timeout(
2495 ptlrpc_check_set(NULL, set),
2496 cfs_time_seconds(timeout ? timeout : 1));
2499 ptlrpc_expired_set(set);
2500 } else if (rc < 0) {
2502 ptlrpc_interrupted_set(set);
2508 * At least one request is in flight, so no
2509 * interrupts are allowed. Wait until all
2510 * complete, or an in-flight req times out.
2512 rc = wait_event_idle_timeout(
2514 ptlrpc_check_set(NULL, set),
2515 cfs_time_seconds(timeout ? timeout : 1));
2517 ptlrpc_expired_set(set);
2524 * LU-769 - if we ignored the signal because
2525 * it was already pending when we started, we
2526 * need to handle it now or we risk it being
2529 if (rc == -ETIMEDOUT &&
2530 signal_pending(current)) {
2533 siginitset(&new, LUSTRE_FATAL_SIGS);
2534 sigprocmask(SIG_BLOCK, &new, &old);
2536 * In fact we only interrupt for the
2537 * "fatal" signals like SIGINT or
2538 * SIGKILL. We still ignore less
2539 * important signals since ptlrpc set
2540 * is not easily reentrant from
2543 if (signal_pending(current))
2544 ptlrpc_interrupted_set(set);
2545 sigprocmask(SIG_SETMASK, &old, NULL);
2549 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2552 * -EINTR => all requests have been flagged rq_intr so next
2554 * -ETIMEDOUT => someone timed out. When all reqs have
2555 * timed out, signals are enabled allowing completion with
2557 * I don't really care if we go once more round the loop in
2558 * the error cases -eeb.
2560 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2561 list_for_each_entry(req, &set->set_requests,
2563 spin_lock(&req->rq_lock);
2564 req->rq_invalid_rqset = 1;
2565 spin_unlock(&req->rq_lock);
2568 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2570 LASSERT(atomic_read(&set->set_remaining) == 0);
2572 rc = set->set_rc; /* rq_status of already freed requests if any */
2573 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2574 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2575 if (req->rq_status != 0)
2576 rc = req->rq_status;
2581 EXPORT_SYMBOL(ptlrpc_set_wait);
2584 * Helper fuction for request freeing.
2585 * Called when request count reached zero and request needs to be freed.
2586 * Removes request from all sorts of sending/replay lists it might be on,
2587 * frees network buffers if any are present.
2588 * If \a locked is set, that means caller is already holding import imp_lock
2589 * and so we no longer need to reobtain it (for certain lists manipulations)
2591 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2598 LASSERT(!request->rq_srv_req);
2599 LASSERT(request->rq_export == NULL);
2600 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2601 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2602 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2603 LASSERTF(!request->rq_replay, "req %p\n", request);
2605 req_capsule_fini(&request->rq_pill);
2608 * We must take it off the imp_replay_list first. Otherwise, we'll set
2609 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2611 if (request->rq_import) {
2613 spin_lock(&request->rq_import->imp_lock);
2614 list_del_init(&request->rq_replay_list);
2615 list_del_init(&request->rq_unreplied_list);
2617 spin_unlock(&request->rq_import->imp_lock);
2619 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2621 if (atomic_read(&request->rq_refcount) != 0) {
2622 DEBUG_REQ(D_ERROR, request,
2623 "freeing request with nonzero refcount");
2627 if (request->rq_repbuf)
2628 sptlrpc_cli_free_repbuf(request);
2630 if (request->rq_import) {
2631 if (!ptlrpcd_check_work(request)) {
2632 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2633 atomic_dec(&request->rq_import->imp_reqs);
2635 class_import_put(request->rq_import);
2636 request->rq_import = NULL;
2638 if (request->rq_bulk)
2639 ptlrpc_free_bulk(request->rq_bulk);
2641 if (request->rq_reqbuf || request->rq_clrbuf)
2642 sptlrpc_cli_free_reqbuf(request);
2644 if (request->rq_cli_ctx)
2645 sptlrpc_req_put_ctx(request, !locked);
2647 if (request->rq_pool)
2648 __ptlrpc_free_req_to_pool(request);
2650 ptlrpc_request_cache_free(request);
2654 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2656 * Drop one request reference. Must be called with import imp_lock held.
2657 * When reference count drops to zero, request is freed.
2659 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2661 assert_spin_locked(&request->rq_import->imp_lock);
2662 (void)__ptlrpc_req_finished(request, 1);
2667 * Drops one reference count for request \a request.
2668 * \a locked set indicates that caller holds import imp_lock.
2669 * Frees the request whe reference count reaches zero.
2671 * \retval 1 the request is freed
2672 * \retval 0 some others still hold references on the request
2674 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2682 LASSERT(request != LP_POISON);
2683 LASSERT(request->rq_reqmsg != LP_POISON);
2685 DEBUG_REQ(D_INFO, request, "refcount now %u",
2686 atomic_read(&request->rq_refcount) - 1);
2688 spin_lock(&request->rq_lock);
2689 count = atomic_dec_return(&request->rq_refcount);
2690 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2693 * For open RPC, the client does not know the EA size (LOV, ACL, and
2694 * so on) before replied, then the client has to reserve very large
2695 * reply buffer. Such buffer will not be released until the RPC freed.
2696 * Since The open RPC is replayable, we need to keep it in the replay
2697 * list until close. If there are a lot of files opened concurrently,
2698 * then the client may be OOM.
2700 * If fact, it is unnecessary to keep reply buffer for open replay,
2701 * related EAs have already been saved via mdc_save_lovea() before
2702 * coming here. So it is safe to free the reply buffer some earlier
2703 * before releasing the RPC to avoid client OOM. LU-9514
2705 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2706 spin_lock(&request->rq_early_free_lock);
2707 sptlrpc_cli_free_repbuf(request);
2708 request->rq_repbuf = NULL;
2709 request->rq_repbuf_len = 0;
2710 request->rq_repdata = NULL;
2711 request->rq_reqdata_len = 0;
2712 spin_unlock(&request->rq_early_free_lock);
2714 spin_unlock(&request->rq_lock);
2717 __ptlrpc_free_req(request, locked);
2723 * Drops one reference count for a request.
2725 void ptlrpc_req_finished(struct ptlrpc_request *request)
2727 __ptlrpc_req_finished(request, 0);
2729 EXPORT_SYMBOL(ptlrpc_req_finished);
2732 * Returns xid of a \a request
2734 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2736 return request->rq_xid;
2738 EXPORT_SYMBOL(ptlrpc_req_xid);
2741 * Disengage the client's reply buffer from the network
2742 * NB does _NOT_ unregister any client-side bulk.
2743 * IDEMPOTENT, but _not_ safe against concurrent callers.
2744 * The request owner (i.e. the thread doing the I/O) must call...
2745 * Returns 0 on success or 1 if unregistering cannot be made.
2747 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2749 bool discard = false;
2753 LASSERT(!in_interrupt());
2755 /* Let's setup deadline for reply unlink. */
2756 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2757 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2758 request->rq_reply_deadline = ktime_get_real_seconds() +
2759 PTLRPC_REQ_LONG_UNLINK;
2762 * Nothing left to do.
2764 if (!__ptlrpc_cli_wait_unlink(request, &discard))
2767 LNetMDUnlink(request->rq_reply_md_h);
2769 if (discard) /* Discard the request-out callback */
2770 __LNetMDUnlink(request->rq_req_md_h, discard);
2773 * Let's check it once again.
2775 if (!ptlrpc_cli_wait_unlink(request))
2778 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2779 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2782 * Do not wait for unlink to finish.
2788 * We have to wait_event_idle_timeout() whatever the result, to get
2789 * a chance to run reply_in_callback(), and to make sure we've
2790 * unlinked before returning a req to the pool.
2793 wait_queue_head_t *wq = (request->rq_set) ?
2794 &request->rq_set->set_waitq :
2795 &request->rq_reply_waitq;
2796 int seconds = PTLRPC_REQ_LONG_UNLINK;
2798 * Network access will complete in finite time but the HUGE
2799 * timeout lets us CWARN for visibility of sluggish NALs
2801 while (seconds > 0 &&
2802 wait_event_idle_timeout(
2804 !ptlrpc_cli_wait_unlink(request),
2805 cfs_time_seconds(1)) == 0)
2808 ptlrpc_rqphase_move(request, request->rq_next_phase);
2812 DEBUG_REQ(D_WARNING, request,
2813 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2814 request->rq_receiving_reply,
2815 request->rq_req_unlinked,
2816 request->rq_reply_unlinked);
2821 static void ptlrpc_free_request(struct ptlrpc_request *req)
2823 spin_lock(&req->rq_lock);
2825 spin_unlock(&req->rq_lock);
2827 if (req->rq_commit_cb)
2828 req->rq_commit_cb(req);
2829 list_del_init(&req->rq_replay_list);
2831 __ptlrpc_req_finished(req, 1);
2835 * the request is committed and dropped from the replay list of its import
2837 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2839 struct obd_import *imp = req->rq_import;
2841 spin_lock(&imp->imp_lock);
2842 if (list_empty(&req->rq_replay_list)) {
2843 spin_unlock(&imp->imp_lock);
2847 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2848 if (imp->imp_replay_cursor == &req->rq_replay_list)
2849 imp->imp_replay_cursor = req->rq_replay_list.next;
2850 ptlrpc_free_request(req);
2853 spin_unlock(&imp->imp_lock);
2855 EXPORT_SYMBOL(ptlrpc_request_committed);
2858 * Iterates through replay_list on import and prunes
2859 * all requests have transno smaller than last_committed for the
2860 * import and don't have rq_replay set.
2861 * Since requests are sorted in transno order, stops when meetign first
2862 * transno bigger than last_committed.
2863 * caller must hold imp->imp_lock
2865 void ptlrpc_free_committed(struct obd_import *imp)
2867 struct ptlrpc_request *req, *saved;
2868 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2869 bool skip_committed_list = true;
2872 LASSERT(imp != NULL);
2873 assert_spin_locked(&imp->imp_lock);
2875 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2876 imp->imp_generation == imp->imp_last_generation_checked) {
2877 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2878 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2881 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2882 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2883 imp->imp_generation);
2885 if (imp->imp_generation != imp->imp_last_generation_checked ||
2886 imp->imp_last_transno_checked == 0)
2887 skip_committed_list = false;
2889 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2890 imp->imp_last_generation_checked = imp->imp_generation;
2892 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2894 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2895 LASSERT(req != last_req);
2898 if (req->rq_transno == 0) {
2899 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2902 if (req->rq_import_generation < imp->imp_generation) {
2903 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2907 /* not yet committed */
2908 if (req->rq_transno > imp->imp_peer_committed_transno) {
2909 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2913 if (req->rq_replay) {
2914 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2915 list_move_tail(&req->rq_replay_list,
2916 &imp->imp_committed_list);
2920 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2921 imp->imp_peer_committed_transno);
2923 ptlrpc_free_request(req);
2926 if (skip_committed_list)
2929 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2931 LASSERT(req->rq_transno != 0);
2932 if (req->rq_import_generation < imp->imp_generation ||
2934 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2935 req->rq_import_generation <
2936 imp->imp_generation ? "stale" : "closed");
2938 if (imp->imp_replay_cursor == &req->rq_replay_list)
2939 imp->imp_replay_cursor =
2940 req->rq_replay_list.next;
2942 ptlrpc_free_request(req);
2949 void ptlrpc_cleanup_client(struct obd_import *imp)
2956 * Schedule previously sent request for resend.
2957 * For bulk requests we assign new xid (to avoid problems with
2958 * lost replies and therefore several transfers landing into same buffer
2959 * from different sending attempts).
2961 void ptlrpc_resend_req(struct ptlrpc_request *req)
2963 DEBUG_REQ(D_HA, req, "going to resend");
2964 spin_lock(&req->rq_lock);
2967 * Request got reply but linked to the import list still.
2968 * Let ptlrpc_check_set() process it.
2970 if (ptlrpc_client_replied(req)) {
2971 spin_unlock(&req->rq_lock);
2972 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2976 req->rq_status = -EAGAIN;
2979 req->rq_net_err = 0;
2980 req->rq_timedout = 0;
2982 ptlrpc_client_wake_req(req);
2983 spin_unlock(&req->rq_lock);
2986 /* XXX: this function and rq_status are currently unused */
2987 void ptlrpc_restart_req(struct ptlrpc_request *req)
2989 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2990 req->rq_status = -ERESTARTSYS;
2992 spin_lock(&req->rq_lock);
2993 req->rq_restart = 1;
2994 req->rq_timedout = 0;
2995 ptlrpc_client_wake_req(req);
2996 spin_unlock(&req->rq_lock);
3000 * Grab additional reference on a request \a req
3002 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
3005 atomic_inc(&req->rq_refcount);
3008 EXPORT_SYMBOL(ptlrpc_request_addref);
3011 * Add a request to import replay_list.
3012 * Must be called under imp_lock
3014 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3015 struct obd_import *imp)
3017 struct ptlrpc_request *iter;
3019 assert_spin_locked(&imp->imp_lock);
3021 if (req->rq_transno == 0) {
3022 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3027 * clear this for new requests that were resent as well
3028 * as resent replayed requests.
3030 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3032 /* don't re-add requests that have been replayed */
3033 if (!list_empty(&req->rq_replay_list))
3036 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3038 spin_lock(&req->rq_lock);
3040 spin_unlock(&req->rq_lock);
3042 LASSERT(imp->imp_replayable);
3043 /* Balanced in ptlrpc_free_committed, usually. */
3044 ptlrpc_request_addref(req);
3045 list_for_each_entry_reverse(iter, &imp->imp_replay_list,
3048 * We may have duplicate transnos if we create and then
3049 * open a file, or for closes retained if to match creating
3050 * opens, so use req->rq_xid as a secondary key.
3051 * (See bugs 684, 685, and 428.)
3052 * XXX no longer needed, but all opens need transnos!
3054 if (iter->rq_transno > req->rq_transno)
3057 if (iter->rq_transno == req->rq_transno) {
3058 LASSERT(iter->rq_xid != req->rq_xid);
3059 if (iter->rq_xid > req->rq_xid)
3063 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3067 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3071 * Send request and wait until it completes.
3072 * Returns request processing status.
3074 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3076 struct ptlrpc_request_set *set;
3080 LASSERT(req->rq_set == NULL);
3081 LASSERT(!req->rq_receiving_reply);
3083 set = ptlrpc_prep_set();
3085 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3089 /* for distributed debugging */
3090 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3092 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3093 ptlrpc_request_addref(req);
3094 ptlrpc_set_add_req(set, req);
3095 rc = ptlrpc_set_wait(NULL, set);
3096 ptlrpc_set_destroy(set);
3100 EXPORT_SYMBOL(ptlrpc_queue_wait);
3103 * Callback used for replayed requests reply processing.
3104 * In case of successful reply calls registered request replay callback.
3105 * In case of error restart replay process.
3107 static int ptlrpc_replay_interpret(const struct lu_env *env,
3108 struct ptlrpc_request *req,
3111 struct ptlrpc_replay_async_args *aa = args;
3112 struct obd_import *imp = req->rq_import;
3115 atomic_dec(&imp->imp_replay_inflight);
3118 * Note: if it is bulk replay (MDS-MDS replay), then even if
3119 * server got the request, but bulk transfer timeout, let's
3120 * replay the bulk req again
3122 if (!ptlrpc_client_replied(req) ||
3124 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3125 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3126 GOTO(out, rc = -ETIMEDOUT);
3129 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3130 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3131 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3132 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3134 /** VBR: check version failure */
3135 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3136 /** replay was failed due to version mismatch */
3137 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3138 spin_lock(&imp->imp_lock);
3139 imp->imp_vbr_failed = 1;
3140 spin_unlock(&imp->imp_lock);
3141 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3143 /** The transno had better not change over replay. */
3144 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3145 lustre_msg_get_transno(req->rq_repmsg) ||
3146 lustre_msg_get_transno(req->rq_repmsg) == 0,
3148 lustre_msg_get_transno(req->rq_reqmsg),
3149 lustre_msg_get_transno(req->rq_repmsg));
3152 spin_lock(&imp->imp_lock);
3153 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3154 spin_unlock(&imp->imp_lock);
3155 LASSERT(imp->imp_last_replay_transno);
3157 /* transaction number shouldn't be bigger than the latest replayed */
3158 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3159 DEBUG_REQ(D_ERROR, req,
3160 "Reported transno=%llu is bigger than replayed=%llu",
3162 lustre_msg_get_transno(req->rq_reqmsg));
3163 GOTO(out, rc = -EINVAL);
3166 DEBUG_REQ(D_HA, req, "got reply");
3168 /* let the callback do fixups, possibly including in the request */
3169 if (req->rq_replay_cb)
3170 req->rq_replay_cb(req);
3172 if (ptlrpc_client_replied(req) &&
3173 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3174 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3175 lustre_msg_get_status(req->rq_repmsg),
3176 aa->praa_old_status);
3179 * Note: If the replay fails for MDT-MDT recovery, let's
3180 * abort all of the following requests in the replay
3181 * and sending list, because MDT-MDT update requests
3182 * are dependent on each other, see LU-7039
3184 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3185 struct ptlrpc_request *free_req;
3186 struct ptlrpc_request *tmp;
3188 spin_lock(&imp->imp_lock);
3189 list_for_each_entry_safe(free_req, tmp,
3190 &imp->imp_replay_list,
3192 ptlrpc_free_request(free_req);
3195 list_for_each_entry_safe(free_req, tmp,
3196 &imp->imp_committed_list,
3198 ptlrpc_free_request(free_req);
3201 list_for_each_entry_safe(free_req, tmp,
3202 &imp->imp_delayed_list,
3204 spin_lock(&free_req->rq_lock);
3205 free_req->rq_err = 1;
3206 free_req->rq_status = -EIO;
3207 ptlrpc_client_wake_req(free_req);
3208 spin_unlock(&free_req->rq_lock);
3211 list_for_each_entry_safe(free_req, tmp,
3212 &imp->imp_sending_list,
3214 spin_lock(&free_req->rq_lock);
3215 free_req->rq_err = 1;
3216 free_req->rq_status = -EIO;
3217 ptlrpc_client_wake_req(free_req);
3218 spin_unlock(&free_req->rq_lock);
3220 spin_unlock(&imp->imp_lock);
3223 /* Put it back for re-replay. */
3224 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3228 * Errors while replay can set transno to 0, but
3229 * imp_last_replay_transno shouldn't be set to 0 anyway
3231 if (req->rq_transno == 0)
3232 CERROR("Transno is 0 during replay!\n");
3234 /* continue with recovery */
3235 rc = ptlrpc_import_recovery_state_machine(imp);
3237 req->rq_send_state = aa->praa_old_state;
3240 /* this replay failed, so restart recovery */
3241 ptlrpc_connect_import(imp);
3247 * Prepares and queues request for replay.
3248 * Adds it to ptlrpcd queue for actual sending.
3249 * Returns 0 on success.
3251 int ptlrpc_replay_req(struct ptlrpc_request *req)
3253 struct ptlrpc_replay_async_args *aa;
3257 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3259 aa = ptlrpc_req_async_args(aa, req);
3260 memset(aa, 0, sizeof(*aa));
3262 /* Prepare request to be resent with ptlrpcd */
3263 aa->praa_old_state = req->rq_send_state;
3264 req->rq_send_state = LUSTRE_IMP_REPLAY;
3265 req->rq_phase = RQ_PHASE_NEW;
3266 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3268 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3270 req->rq_interpret_reply = ptlrpc_replay_interpret;
3271 /* Readjust the timeout for current conditions */
3272 ptlrpc_at_set_req_timeout(req);
3274 /* Tell server net_latency to calculate how long to wait for reply. */
3275 lustre_msg_set_service_timeout(req->rq_reqmsg,
3276 ptlrpc_at_get_net_latency(req));
3277 DEBUG_REQ(D_HA, req, "REPLAY");
3279 atomic_inc(&req->rq_import->imp_replay_inflight);
3280 spin_lock(&req->rq_lock);
3281 req->rq_early_free_repbuf = 0;
3282 spin_unlock(&req->rq_lock);
3283 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3285 ptlrpcd_add_req(req);
3290 * Aborts all in-flight request on import \a imp sending and delayed lists
3292 void ptlrpc_abort_inflight(struct obd_import *imp)
3294 struct ptlrpc_request *req;
3298 * Make sure that no new requests get processed for this import.
3299 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3300 * this flag and then putting requests on sending_list or delayed_list.
3302 assert_spin_locked(&imp->imp_lock);
3305 * XXX locking? Maybe we should remove each request with the list
3306 * locked? Also, how do we know if the requests on the list are
3307 * being freed at this time?
3309 list_for_each_entry(req, &imp->imp_sending_list, rq_list) {
3310 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3312 spin_lock(&req->rq_lock);
3313 if (req->rq_import_generation < imp->imp_generation) {
3315 req->rq_status = -EIO;
3316 ptlrpc_client_wake_req(req);
3318 spin_unlock(&req->rq_lock);
3321 list_for_each_entry(req, &imp->imp_delayed_list, rq_list) {
3322 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3324 spin_lock(&req->rq_lock);
3325 if (req->rq_import_generation < imp->imp_generation) {
3327 req->rq_status = -EIO;
3328 ptlrpc_client_wake_req(req);
3330 spin_unlock(&req->rq_lock);
3334 * Last chance to free reqs left on the replay list, but we
3335 * will still leak reqs that haven't committed.
3337 if (imp->imp_replayable)
3338 ptlrpc_free_committed(imp);
3344 * Abort all uncompleted requests in request set \a set
3346 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3348 struct ptlrpc_request *req;
3350 LASSERT(set != NULL);
3352 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
3353 spin_lock(&req->rq_lock);
3354 if (req->rq_phase != RQ_PHASE_RPC) {
3355 spin_unlock(&req->rq_lock);
3360 req->rq_status = -EINTR;
3361 ptlrpc_client_wake_req(req);
3362 spin_unlock(&req->rq_lock);
3367 * Initialize the XID for the node. This is common among all requests on
3368 * this node, and only requires the property that it is monotonically
3369 * increasing. It does not need to be sequential. Since this is also used
3370 * as the RDMA match bits, it is important that a single client NOT have
3371 * the same match bits for two different in-flight requests, hence we do
3372 * NOT want to have an XID per target or similar.
3374 * To avoid an unlikely collision between match bits after a client reboot
3375 * (which would deliver old data into the wrong RDMA buffer) initialize
3376 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3377 * If the time is clearly incorrect, we instead use a 62-bit random number.
3378 * In the worst case the random number will overflow 1M RPCs per second in
3379 * 9133 years, or permutations thereof.
3381 #define YEAR_2004 (1ULL << 30)
3382 void ptlrpc_init_xid(void)
3384 time64_t now = ktime_get_real_seconds();
3387 if (now < YEAR_2004) {
3388 get_random_bytes(&xid, sizeof(xid));
3390 xid |= (1ULL << 61);
3392 xid = (u64)now << 20;
3395 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3396 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3398 xid &= PTLRPC_BULK_OPS_MASK;
3399 atomic64_set(&ptlrpc_last_xid, xid);
3403 * Increase xid and returns resulting new value to the caller.
3405 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3406 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3407 * itself uses the last bulk xid needed, so the server can determine the
3408 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3409 * xid must align to a power-of-two value.
3411 * This is assumed to be true due to the initial ptlrpc_last_xid
3412 * value also being initialized to a power-of-two value. LU-1431
3414 __u64 ptlrpc_next_xid(void)
3416 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3420 * If request has a new allocated XID (new request or EINPROGRESS resend),
3421 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3422 * request to ensure previous bulk fails and avoid problems with lost replies
3423 * and therefore several transfers landing into the same buffer from different
3426 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3428 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3430 LASSERT(bd != NULL);
3433 * Generate new matchbits for all resend requests, including
3436 if (req->rq_resend) {
3437 __u64 old_mbits = req->rq_mbits;
3440 * First time resend on -EINPROGRESS will generate new xid,
3441 * so we can actually use the rq_xid as rq_mbits in such case,
3442 * however, it's bit hard to distinguish such resend with a
3443 * 'resend for the -EINPROGRESS resend'. To make it simple,
3444 * we opt to generate mbits for all resend cases.
3446 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3448 req->rq_mbits = ptlrpc_next_xid();
3451 * Old version transfers rq_xid to peer as
3454 spin_lock(&req->rq_import->imp_lock);
3455 list_del_init(&req->rq_unreplied_list);
3456 ptlrpc_assign_next_xid_nolock(req);
3457 spin_unlock(&req->rq_import->imp_lock);
3458 req->rq_mbits = req->rq_xid;
3460 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3461 old_mbits, req->rq_mbits);
3462 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3463 /* Request being sent first time, use xid as matchbits. */
3464 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3465 || req->rq_mbits == 0) {
3466 req->rq_mbits = req->rq_xid;
3468 req->rq_mbits -= bd->bd_md_count - 1;
3472 * Replay request, xid and matchbits have already been
3473 * correctly assigned.
3479 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3480 * that server can infer the number of bulks that were prepared,
3483 req->rq_mbits += bd->bd_md_count - 1;
3486 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3487 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3489 * It's ok to directly set the rq_xid here, since this xid bump
3490 * won't affect the request position in unreplied list.
3492 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3493 req->rq_xid = req->rq_mbits;
3497 * Get a glimpse at what next xid value might have been.
3498 * Returns possible next xid.
3500 __u64 ptlrpc_sample_next_xid(void)
3502 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3504 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3507 * Functions for operating ptlrpc workers.
3509 * A ptlrpc work is a function which will be running inside ptlrpc context.
3510 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3512 * 1. after a work is created, it can be used many times, that is:
3513 * handler = ptlrpcd_alloc_work();
3514 * ptlrpcd_queue_work();
3516 * queue it again when necessary:
3517 * ptlrpcd_queue_work();
3518 * ptlrpcd_destroy_work();
3519 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3520 * it will only be queued once in any time. Also as its name implies, it may
3521 * have delay before it really runs by ptlrpcd thread.
3523 struct ptlrpc_work_async_args {
3524 int (*cb)(const struct lu_env *, void *);
3528 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3530 /* re-initialize the req */
3531 req->rq_timeout = obd_timeout;
3532 req->rq_sent = ktime_get_real_seconds();
3533 req->rq_deadline = req->rq_sent + req->rq_timeout;
3534 req->rq_phase = RQ_PHASE_INTERPRET;
3535 req->rq_next_phase = RQ_PHASE_COMPLETE;
3536 req->rq_xid = ptlrpc_next_xid();
3537 req->rq_import_generation = req->rq_import->imp_generation;
3539 ptlrpcd_add_req(req);
3542 static int work_interpreter(const struct lu_env *env,
3543 struct ptlrpc_request *req, void *args, int rc)
3545 struct ptlrpc_work_async_args *arg = args;
3547 LASSERT(ptlrpcd_check_work(req));
3548 LASSERT(arg->cb != NULL);
3550 rc = arg->cb(env, arg->cbdata);
3552 list_del_init(&req->rq_set_chain);
3555 if (atomic_dec_return(&req->rq_refcount) > 1) {
3556 atomic_set(&req->rq_refcount, 2);
3557 ptlrpcd_add_work_req(req);
3562 static int worker_format;
3564 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3566 return req->rq_pill.rc_fmt == (void *)&worker_format;
3570 * Create a work for ptlrpc.
3572 void *ptlrpcd_alloc_work(struct obd_import *imp,
3573 int (*cb)(const struct lu_env *, void *), void *cbdata)
3575 struct ptlrpc_request *req = NULL;
3576 struct ptlrpc_work_async_args *args;
3582 RETURN(ERR_PTR(-EINVAL));
3584 /* copy some code from deprecated fakereq. */
3585 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3587 CERROR("ptlrpc: run out of memory!\n");
3588 RETURN(ERR_PTR(-ENOMEM));
3591 ptlrpc_cli_req_init(req);
3593 req->rq_send_state = LUSTRE_IMP_FULL;
3594 req->rq_type = PTL_RPC_MSG_REQUEST;
3595 req->rq_import = class_import_get(imp);
3596 req->rq_interpret_reply = work_interpreter;
3597 /* don't want reply */
3598 req->rq_no_delay = req->rq_no_resend = 1;
3599 req->rq_pill.rc_fmt = (void *)&worker_format;
3601 args = ptlrpc_req_async_args(args, req);
3603 args->cbdata = cbdata;
3607 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3609 void ptlrpcd_destroy_work(void *handler)
3611 struct ptlrpc_request *req = handler;
3614 ptlrpc_req_finished(req);
3616 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3618 int ptlrpcd_queue_work(void *handler)
3620 struct ptlrpc_request *req = handler;
3623 * Check if the req is already being queued.
3625 * Here comes a trick: it lacks a way of checking if a req is being
3626 * processed reliably in ptlrpc. Here I have to use refcount of req
3627 * for this purpose. This is okay because the caller should use this
3628 * req as opaque data. - Jinshan
3630 LASSERT(atomic_read(&req->rq_refcount) > 0);
3631 if (atomic_inc_return(&req->rq_refcount) == 2)
3632 ptlrpcd_add_work_req(req);
3635 EXPORT_SYMBOL(ptlrpcd_queue_work);