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((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
410 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 rc = sptlrpc_req_refresh_ctx(req, 0);
1703 req->rq_status = rc;
1706 spin_lock(&req->rq_lock);
1707 req->rq_wait_ctx = 1;
1708 spin_unlock(&req->rq_lock);
1714 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1716 imp->imp_obd->obd_uuid.uuid,
1717 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1718 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1719 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1721 rc = ptl_send_rpc(req, 0);
1722 if (rc == -ENOMEM) {
1723 spin_lock(&imp->imp_lock);
1724 if (!list_empty(&req->rq_list)) {
1725 list_del_init(&req->rq_list);
1726 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1727 wake_up(&req->rq_import->imp_recovery_waitq);
1729 spin_unlock(&imp->imp_lock);
1730 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1734 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1736 spin_lock(&req->rq_lock);
1737 req->rq_net_err = 1;
1738 spin_unlock(&req->rq_lock);
1744 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1749 LASSERT(set->set_producer != NULL);
1751 remaining = atomic_read(&set->set_remaining);
1754 * populate the ->set_requests list with requests until we
1755 * reach the maximum number of RPCs in flight for this set
1757 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1758 rc = set->set_producer(set, set->set_producer_arg);
1759 if (rc == -ENOENT) {
1760 /* no more RPC to produce */
1761 set->set_producer = NULL;
1762 set->set_producer_arg = NULL;
1767 RETURN((atomic_read(&set->set_remaining) - remaining));
1771 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1772 * and no more replies are expected.
1773 * (it is possible to get less replies than requests sent e.g. due to timed out
1774 * requests or requests that we had trouble to send out)
1776 * NOTE: This function contains a potential schedule point (cond_resched()).
1778 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1780 struct ptlrpc_request *req, *next;
1781 LIST_HEAD(comp_reqs);
1782 int force_timer_recalc = 0;
1785 if (atomic_read(&set->set_remaining) == 0)
1788 list_for_each_entry_safe(req, next, &set->set_requests,
1790 struct obd_import *imp = req->rq_import;
1791 int unregistered = 0;
1795 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1796 list_move_tail(&req->rq_set_chain, &comp_reqs);
1801 * This schedule point is mainly for the ptlrpcd caller of this
1802 * function. Most ptlrpc sets are not long-lived and unbounded
1803 * in length, but at the least the set used by the ptlrpcd is.
1804 * Since the processing time is unbounded, we need to insert an
1805 * explicit schedule point to make the thread well-behaved.
1810 * If the caller requires to allow to be interpreted by force
1811 * and it has really been interpreted, then move the request
1812 * to RQ_PHASE_INTERPRET phase in spite of what the current
1815 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1816 req->rq_status = -EINTR;
1817 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1820 * Since it is interpreted and we have to wait for
1821 * the reply to be unlinked, then use sync mode.
1825 GOTO(interpret, req->rq_status);
1828 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1829 force_timer_recalc = 1;
1831 /* delayed send - skip */
1832 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1835 /* delayed resend - skip */
1836 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1837 req->rq_sent > ktime_get_real_seconds())
1840 if (!(req->rq_phase == RQ_PHASE_RPC ||
1841 req->rq_phase == RQ_PHASE_BULK ||
1842 req->rq_phase == RQ_PHASE_INTERPRET ||
1843 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1844 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1845 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1849 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1850 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1851 LASSERT(req->rq_next_phase != req->rq_phase);
1852 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1854 if (req->rq_req_deadline &&
1855 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1856 req->rq_req_deadline = 0;
1857 if (req->rq_reply_deadline &&
1858 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1859 req->rq_reply_deadline = 0;
1860 if (req->rq_bulk_deadline &&
1861 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1862 req->rq_bulk_deadline = 0;
1865 * Skip processing until reply is unlinked. We
1866 * can't return to pool before that and we can't
1867 * call interpret before that. We need to make
1868 * sure that all rdma transfers finished and will
1869 * not corrupt any data.
1871 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1872 ptlrpc_cli_wait_unlink(req))
1874 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1875 ptlrpc_client_bulk_active(req))
1879 * Turn fail_loc off to prevent it from looping
1882 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1883 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1886 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1887 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1892 * Move to next phase if reply was successfully
1895 ptlrpc_rqphase_move(req, req->rq_next_phase);
1898 if (req->rq_phase == RQ_PHASE_INTERPRET)
1899 GOTO(interpret, req->rq_status);
1902 * Note that this also will start async reply unlink.
1904 if (req->rq_net_err && !req->rq_timedout) {
1905 ptlrpc_expire_one_request(req, 1);
1908 * Check if we still need to wait for unlink.
1910 if (ptlrpc_cli_wait_unlink(req) ||
1911 ptlrpc_client_bulk_active(req))
1913 /* If there is no need to resend, fail it now. */
1914 if (req->rq_no_resend) {
1915 if (req->rq_status == 0)
1916 req->rq_status = -EIO;
1917 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1918 GOTO(interpret, req->rq_status);
1925 spin_lock(&req->rq_lock);
1926 req->rq_replied = 0;
1927 spin_unlock(&req->rq_lock);
1928 if (req->rq_status == 0)
1929 req->rq_status = -EIO;
1930 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1931 GOTO(interpret, req->rq_status);
1935 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1936 * so it sets rq_intr regardless of individual rpc
1937 * timeouts. The synchronous IO waiting path sets
1938 * rq_intr irrespective of whether ptlrpcd
1939 * has seen a timeout. Our policy is to only interpret
1940 * interrupted rpcs after they have timed out, so we
1941 * need to enforce that here.
1944 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1945 req->rq_wait_ctx)) {
1946 req->rq_status = -EINTR;
1947 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1948 GOTO(interpret, req->rq_status);
1951 if (req->rq_phase == RQ_PHASE_RPC) {
1952 if (req->rq_timedout || req->rq_resend ||
1953 req->rq_waiting || req->rq_wait_ctx) {
1956 if (!ptlrpc_unregister_reply(req, 1)) {
1957 ptlrpc_unregister_bulk(req, 1);
1961 spin_lock(&imp->imp_lock);
1962 if (ptlrpc_import_delay_req(imp, req,
1965 * put on delay list - only if we wait
1966 * recovery finished - before send
1968 list_move_tail(&req->rq_list,
1969 &imp->imp_delayed_list);
1970 spin_unlock(&imp->imp_lock);
1975 req->rq_status = status;
1976 ptlrpc_rqphase_move(req,
1977 RQ_PHASE_INTERPRET);
1978 spin_unlock(&imp->imp_lock);
1979 GOTO(interpret, req->rq_status);
1981 /* ignore on just initiated connections */
1982 if (ptlrpc_no_resend(req) &&
1983 !req->rq_wait_ctx &&
1984 imp->imp_generation !=
1985 imp->imp_initiated_at) {
1986 req->rq_status = -ENOTCONN;
1987 ptlrpc_rqphase_move(req,
1988 RQ_PHASE_INTERPRET);
1989 spin_unlock(&imp->imp_lock);
1990 GOTO(interpret, req->rq_status);
1993 /* don't resend too fast in case of network
1996 if (ktime_get_real_seconds() < (req->rq_sent + 1)
1997 && req->rq_net_err && req->rq_timedout) {
1999 DEBUG_REQ(D_INFO, req,
2000 "throttle request");
2001 /* Don't try to resend RPC right away
2002 * as it is likely it will fail again
2003 * and ptlrpc_check_set() will be
2004 * called again, keeping this thread
2005 * busy. Instead, wait for the next
2006 * timeout. Flag it as resend to
2007 * ensure we don't wait to long.
2010 spin_unlock(&imp->imp_lock);
2014 list_move_tail(&req->rq_list,
2015 &imp->imp_sending_list);
2017 spin_unlock(&imp->imp_lock);
2019 spin_lock(&req->rq_lock);
2020 req->rq_waiting = 0;
2021 spin_unlock(&req->rq_lock);
2023 if (req->rq_timedout || req->rq_resend) {
2025 * This is re-sending anyways,
2026 * let's mark req as resend.
2028 spin_lock(&req->rq_lock);
2030 spin_unlock(&req->rq_lock);
2033 * rq_wait_ctx is only touched by ptlrpcd,
2034 * so no lock is needed here.
2036 status = sptlrpc_req_refresh_ctx(req, 0);
2039 req->rq_status = status;
2040 spin_lock(&req->rq_lock);
2041 req->rq_wait_ctx = 0;
2042 spin_unlock(&req->rq_lock);
2043 force_timer_recalc = 1;
2045 spin_lock(&req->rq_lock);
2046 req->rq_wait_ctx = 1;
2047 spin_unlock(&req->rq_lock);
2052 spin_lock(&req->rq_lock);
2053 req->rq_wait_ctx = 0;
2054 spin_unlock(&req->rq_lock);
2058 * In any case, the previous bulk should be
2059 * cleaned up to prepare for the new sending
2062 !ptlrpc_unregister_bulk(req, 1))
2065 rc = ptl_send_rpc(req, 0);
2066 if (rc == -ENOMEM) {
2067 spin_lock(&imp->imp_lock);
2068 if (!list_empty(&req->rq_list))
2069 list_del_init(&req->rq_list);
2070 spin_unlock(&imp->imp_lock);
2071 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2075 DEBUG_REQ(D_HA, req,
2076 "send failed: rc = %d", rc);
2077 force_timer_recalc = 1;
2078 spin_lock(&req->rq_lock);
2079 req->rq_net_err = 1;
2080 spin_unlock(&req->rq_lock);
2083 /* need to reset the timeout */
2084 force_timer_recalc = 1;
2087 spin_lock(&req->rq_lock);
2089 if (ptlrpc_client_early(req)) {
2090 ptlrpc_at_recv_early_reply(req);
2091 spin_unlock(&req->rq_lock);
2095 /* Still waiting for a reply? */
2096 if (ptlrpc_client_recv(req)) {
2097 spin_unlock(&req->rq_lock);
2101 /* Did we actually receive a reply? */
2102 if (!ptlrpc_client_replied(req)) {
2103 spin_unlock(&req->rq_lock);
2107 spin_unlock(&req->rq_lock);
2110 * unlink from net because we are going to
2111 * swab in-place of reply buffer
2113 unregistered = ptlrpc_unregister_reply(req, 1);
2117 req->rq_status = after_reply(req);
2122 * If there is no bulk associated with this request,
2123 * then we're done and should let the interpreter
2124 * process the reply. Similarly if the RPC returned
2125 * an error, and therefore the bulk will never arrive.
2127 if (!req->rq_bulk || req->rq_status < 0) {
2128 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2129 GOTO(interpret, req->rq_status);
2132 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2135 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2136 if (ptlrpc_client_bulk_active(req))
2139 if (req->rq_bulk->bd_failure) {
2141 * The RPC reply arrived OK, but the bulk screwed
2142 * up! Dead weird since the server told us the RPC
2143 * was good after getting the REPLY for her GET or
2144 * the ACK for her PUT.
2146 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2147 req->rq_status = -EIO;
2150 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2153 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2156 * This moves to "unregistering" phase we need to wait for
2159 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2160 /* start async bulk unlink too */
2161 ptlrpc_unregister_bulk(req, 1);
2165 if (!ptlrpc_unregister_bulk(req, async))
2169 * When calling interpret receiving already should be
2172 LASSERT(!req->rq_receiving_reply);
2174 ptlrpc_req_interpret(env, req, req->rq_status);
2176 if (ptlrpcd_check_work(req)) {
2177 atomic_dec(&set->set_remaining);
2180 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2184 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2186 imp->imp_obd->obd_uuid.uuid,
2187 lustre_msg_get_status(req->rq_reqmsg),
2189 obd_import_nid2str(imp),
2190 lustre_msg_get_opc(req->rq_reqmsg),
2191 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2193 spin_lock(&imp->imp_lock);
2195 * Request already may be not on sending or delaying list. This
2196 * may happen in the case of marking it erroneous for the case
2197 * ptlrpc_import_delay_req(req, status) find it impossible to
2198 * allow sending this rpc and returns *status != 0.
2200 if (!list_empty(&req->rq_list)) {
2201 list_del_init(&req->rq_list);
2202 if (atomic_dec_and_test(&imp->imp_inflight))
2203 wake_up(&imp->imp_recovery_waitq);
2205 list_del_init(&req->rq_unreplied_list);
2206 spin_unlock(&imp->imp_lock);
2208 atomic_dec(&set->set_remaining);
2209 wake_up(&imp->imp_recovery_waitq);
2211 if (set->set_producer) {
2212 /* produce a new request if possible */
2213 if (ptlrpc_set_producer(set) > 0)
2214 force_timer_recalc = 1;
2217 * free the request that has just been completed
2218 * in order not to pollute set->set_requests
2220 list_del_init(&req->rq_set_chain);
2221 spin_lock(&req->rq_lock);
2223 req->rq_invalid_rqset = 0;
2224 spin_unlock(&req->rq_lock);
2226 /* record rq_status to compute the final status later */
2227 if (req->rq_status != 0)
2228 set->set_rc = req->rq_status;
2229 ptlrpc_req_finished(req);
2231 list_move_tail(&req->rq_set_chain, &comp_reqs);
2236 * move completed request at the head of list so it's easier for
2237 * caller to find them
2239 list_splice(&comp_reqs, &set->set_requests);
2241 /* If we hit an error, we want to recover promptly. */
2242 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2244 EXPORT_SYMBOL(ptlrpc_check_set);
2247 * Time out request \a req. is \a async_unlink is set, that means do not wait
2248 * until LNet actually confirms network buffer unlinking.
2249 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2251 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2253 struct obd_import *imp = req->rq_import;
2254 unsigned int debug_mask = D_RPCTRACE;
2258 spin_lock(&req->rq_lock);
2259 req->rq_timedout = 1;
2260 spin_unlock(&req->rq_lock);
2262 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2263 lustre_msg_get_status(req->rq_reqmsg)))
2264 debug_mask = D_WARNING;
2265 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2266 req->rq_net_err ? "failed due to network error" :
2267 ((req->rq_real_sent == 0 ||
2268 req->rq_real_sent < req->rq_sent ||
2269 req->rq_real_sent >= req->rq_deadline) ?
2270 "timed out for sent delay" : "timed out for slow reply"),
2271 req->rq_sent, req->rq_real_sent);
2273 if (imp && obd_debug_peer_on_timeout)
2274 LNetDebugPeer(imp->imp_connection->c_peer);
2276 ptlrpc_unregister_reply(req, async_unlink);
2277 ptlrpc_unregister_bulk(req, async_unlink);
2279 if (obd_dump_on_timeout)
2280 libcfs_debug_dumplog();
2283 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2287 atomic_inc(&imp->imp_timeouts);
2289 /* The DLM server doesn't want recovery run on its imports. */
2290 if (imp->imp_dlm_fake)
2294 * If this request is for recovery or other primordial tasks,
2295 * then error it out here.
2297 if (req->rq_ctx_init || req->rq_ctx_fini ||
2298 req->rq_send_state != LUSTRE_IMP_FULL ||
2299 imp->imp_obd->obd_no_recov) {
2300 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2301 ptlrpc_import_state_name(req->rq_send_state),
2302 ptlrpc_import_state_name(imp->imp_state));
2303 spin_lock(&req->rq_lock);
2304 req->rq_status = -ETIMEDOUT;
2306 spin_unlock(&req->rq_lock);
2311 * if a request can't be resent we can't wait for an answer after
2314 if (ptlrpc_no_resend(req)) {
2315 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2319 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2325 * Time out all uncompleted requests in request set pointed by \a data
2326 * This is called when a wait times out.
2328 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2330 struct ptlrpc_request *req;
2331 time64_t now = ktime_get_real_seconds();
2334 LASSERT(set != NULL);
2337 * A timeout expired. See which reqs it applies to...
2339 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2340 /* don't expire request waiting for context */
2341 if (req->rq_wait_ctx)
2344 /* Request in-flight? */
2345 if (!((req->rq_phase == RQ_PHASE_RPC &&
2346 !req->rq_waiting && !req->rq_resend) ||
2347 (req->rq_phase == RQ_PHASE_BULK)))
2350 if (req->rq_timedout || /* already dealt with */
2351 req->rq_deadline > now) /* not expired */
2355 * Deal with this guy. Do it asynchronously to not block
2358 ptlrpc_expire_one_request(req, 1);
2360 * Loops require that we resched once in a while to avoid
2361 * RCU stalls and a few other problems.
2369 * Interrupts (sets interrupted flag) all uncompleted requests in
2370 * a set \a data. This is called when a wait_event is interrupted
2373 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2375 struct ptlrpc_request *req;
2377 LASSERT(set != NULL);
2378 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2380 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2384 if (req->rq_phase != RQ_PHASE_RPC &&
2385 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2386 !req->rq_allow_intr)
2389 spin_lock(&req->rq_lock);
2391 spin_unlock(&req->rq_lock);
2396 * Get the smallest timeout in the set; this does NOT set a timeout.
2398 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2400 time64_t now = ktime_get_real_seconds();
2402 struct ptlrpc_request *req;
2406 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2407 /* Request in-flight? */
2408 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2409 (req->rq_phase == RQ_PHASE_BULK) ||
2410 (req->rq_phase == RQ_PHASE_NEW)))
2413 /* Already timed out. */
2414 if (req->rq_timedout)
2417 /* Waiting for ctx. */
2418 if (req->rq_wait_ctx)
2421 if (req->rq_phase == RQ_PHASE_NEW)
2422 deadline = req->rq_sent;
2423 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2424 deadline = req->rq_sent;
2426 deadline = req->rq_sent + req->rq_timeout;
2428 if (deadline <= now) /* actually expired already */
2429 timeout = 1; /* ASAP */
2430 else if (timeout == 0 || timeout > deadline - now)
2431 timeout = deadline - now;
2437 * Send all unset request from the set and then wait untill all
2438 * requests in the set complete (either get a reply, timeout, get an
2439 * error or otherwise be interrupted).
2440 * Returns 0 on success or error code otherwise.
2442 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2444 struct ptlrpc_request *req;
2449 if (set->set_producer)
2450 (void)ptlrpc_set_producer(set);
2452 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2453 if (req->rq_phase == RQ_PHASE_NEW)
2454 (void)ptlrpc_send_new_req(req);
2457 if (list_empty(&set->set_requests))
2461 timeout = ptlrpc_set_next_timeout(set);
2464 * wait until all complete, interrupted, or an in-flight
2467 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2470 if ((timeout == 0 && !signal_pending(current)) ||
2471 set->set_allow_intr) {
2473 * No requests are in-flight (ether timed out
2474 * or delayed), so we can allow interrupts.
2475 * We still want to block for a limited time,
2476 * so we allow interrupts during the timeout.
2478 rc = l_wait_event_abortable_timeout(
2480 ptlrpc_check_set(NULL, set),
2481 cfs_time_seconds(timeout ? timeout : 1));
2484 ptlrpc_expired_set(set);
2485 } else if (rc < 0) {
2487 ptlrpc_interrupted_set(set);
2493 * At least one request is in flight, so no
2494 * interrupts are allowed. Wait until all
2495 * complete, or an in-flight req times out.
2497 rc = wait_event_idle_timeout(
2499 ptlrpc_check_set(NULL, set),
2500 cfs_time_seconds(timeout ? timeout : 1));
2502 ptlrpc_expired_set(set);
2509 * LU-769 - if we ignored the signal because
2510 * it was already pending when we started, we
2511 * need to handle it now or we risk it being
2514 if (rc == -ETIMEDOUT &&
2515 signal_pending(current)) {
2518 siginitset(&new, LUSTRE_FATAL_SIGS);
2519 sigprocmask(SIG_BLOCK, &new, &old);
2521 * In fact we only interrupt for the
2522 * "fatal" signals like SIGINT or
2523 * SIGKILL. We still ignore less
2524 * important signals since ptlrpc set
2525 * is not easily reentrant from
2528 if (signal_pending(current))
2529 ptlrpc_interrupted_set(set);
2530 sigprocmask(SIG_SETMASK, &old, NULL);
2534 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2537 * -EINTR => all requests have been flagged rq_intr so next
2539 * -ETIMEDOUT => someone timed out. When all reqs have
2540 * timed out, signals are enabled allowing completion with
2542 * I don't really care if we go once more round the loop in
2543 * the error cases -eeb.
2545 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2546 list_for_each_entry(req, &set->set_requests,
2548 spin_lock(&req->rq_lock);
2549 req->rq_invalid_rqset = 1;
2550 spin_unlock(&req->rq_lock);
2553 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2555 LASSERT(atomic_read(&set->set_remaining) == 0);
2557 rc = set->set_rc; /* rq_status of already freed requests if any */
2558 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2559 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2560 if (req->rq_status != 0)
2561 rc = req->rq_status;
2566 EXPORT_SYMBOL(ptlrpc_set_wait);
2569 * Helper fuction for request freeing.
2570 * Called when request count reached zero and request needs to be freed.
2571 * Removes request from all sorts of sending/replay lists it might be on,
2572 * frees network buffers if any are present.
2573 * If \a locked is set, that means caller is already holding import imp_lock
2574 * and so we no longer need to reobtain it (for certain lists manipulations)
2576 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2583 LASSERT(!request->rq_srv_req);
2584 LASSERT(request->rq_export == NULL);
2585 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2586 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2587 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2588 LASSERTF(!request->rq_replay, "req %p\n", request);
2590 req_capsule_fini(&request->rq_pill);
2593 * We must take it off the imp_replay_list first. Otherwise, we'll set
2594 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2596 if (request->rq_import) {
2598 spin_lock(&request->rq_import->imp_lock);
2599 list_del_init(&request->rq_replay_list);
2600 list_del_init(&request->rq_unreplied_list);
2602 spin_unlock(&request->rq_import->imp_lock);
2604 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2606 if (atomic_read(&request->rq_refcount) != 0) {
2607 DEBUG_REQ(D_ERROR, request,
2608 "freeing request with nonzero refcount");
2612 if (request->rq_repbuf)
2613 sptlrpc_cli_free_repbuf(request);
2615 if (request->rq_import) {
2616 if (!ptlrpcd_check_work(request)) {
2617 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2618 atomic_dec(&request->rq_import->imp_reqs);
2620 class_import_put(request->rq_import);
2621 request->rq_import = NULL;
2623 if (request->rq_bulk)
2624 ptlrpc_free_bulk(request->rq_bulk);
2626 if (request->rq_reqbuf || request->rq_clrbuf)
2627 sptlrpc_cli_free_reqbuf(request);
2629 if (request->rq_cli_ctx)
2630 sptlrpc_req_put_ctx(request, !locked);
2632 if (request->rq_pool)
2633 __ptlrpc_free_req_to_pool(request);
2635 ptlrpc_request_cache_free(request);
2639 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2641 * Drop one request reference. Must be called with import imp_lock held.
2642 * When reference count drops to zero, request is freed.
2644 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2646 assert_spin_locked(&request->rq_import->imp_lock);
2647 (void)__ptlrpc_req_finished(request, 1);
2652 * Drops one reference count for request \a request.
2653 * \a locked set indicates that caller holds import imp_lock.
2654 * Frees the request whe reference count reaches zero.
2656 * \retval 1 the request is freed
2657 * \retval 0 some others still hold references on the request
2659 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2667 LASSERT(request != LP_POISON);
2668 LASSERT(request->rq_reqmsg != LP_POISON);
2670 DEBUG_REQ(D_INFO, request, "refcount now %u",
2671 atomic_read(&request->rq_refcount) - 1);
2673 spin_lock(&request->rq_lock);
2674 count = atomic_dec_return(&request->rq_refcount);
2675 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2678 * For open RPC, the client does not know the EA size (LOV, ACL, and
2679 * so on) before replied, then the client has to reserve very large
2680 * reply buffer. Such buffer will not be released until the RPC freed.
2681 * Since The open RPC is replayable, we need to keep it in the replay
2682 * list until close. If there are a lot of files opened concurrently,
2683 * then the client may be OOM.
2685 * If fact, it is unnecessary to keep reply buffer for open replay,
2686 * related EAs have already been saved via mdc_save_lovea() before
2687 * coming here. So it is safe to free the reply buffer some earlier
2688 * before releasing the RPC to avoid client OOM. LU-9514
2690 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2691 spin_lock(&request->rq_early_free_lock);
2692 sptlrpc_cli_free_repbuf(request);
2693 request->rq_repbuf = NULL;
2694 request->rq_repbuf_len = 0;
2695 request->rq_repdata = NULL;
2696 request->rq_reqdata_len = 0;
2697 spin_unlock(&request->rq_early_free_lock);
2699 spin_unlock(&request->rq_lock);
2702 __ptlrpc_free_req(request, locked);
2708 * Drops one reference count for a request.
2710 void ptlrpc_req_finished(struct ptlrpc_request *request)
2712 __ptlrpc_req_finished(request, 0);
2714 EXPORT_SYMBOL(ptlrpc_req_finished);
2717 * Returns xid of a \a request
2719 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2721 return request->rq_xid;
2723 EXPORT_SYMBOL(ptlrpc_req_xid);
2726 * Disengage the client's reply buffer from the network
2727 * NB does _NOT_ unregister any client-side bulk.
2728 * IDEMPOTENT, but _not_ safe against concurrent callers.
2729 * The request owner (i.e. the thread doing the I/O) must call...
2730 * Returns 0 on success or 1 if unregistering cannot be made.
2732 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2734 bool discard = false;
2738 LASSERT(!in_interrupt());
2740 /* Let's setup deadline for reply unlink. */
2741 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2742 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2743 request->rq_reply_deadline = ktime_get_real_seconds() +
2744 PTLRPC_REQ_LONG_UNLINK;
2747 * Nothing left to do.
2749 if (!__ptlrpc_cli_wait_unlink(request, &discard))
2752 LNetMDUnlink(request->rq_reply_md_h);
2754 if (discard) /* Discard the request-out callback */
2755 __LNetMDUnlink(request->rq_req_md_h, discard);
2758 * Let's check it once again.
2760 if (!ptlrpc_cli_wait_unlink(request))
2763 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2764 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2767 * Do not wait for unlink to finish.
2773 * We have to wait_event_idle_timeout() whatever the result, to get
2774 * a chance to run reply_in_callback(), and to make sure we've
2775 * unlinked before returning a req to the pool.
2778 wait_queue_head_t *wq = (request->rq_set) ?
2779 &request->rq_set->set_waitq :
2780 &request->rq_reply_waitq;
2781 int seconds = PTLRPC_REQ_LONG_UNLINK;
2783 * Network access will complete in finite time but the HUGE
2784 * timeout lets us CWARN for visibility of sluggish NALs
2786 while (seconds > 0 &&
2787 wait_event_idle_timeout(
2789 !ptlrpc_cli_wait_unlink(request),
2790 cfs_time_seconds(1)) == 0)
2793 ptlrpc_rqphase_move(request, request->rq_next_phase);
2797 DEBUG_REQ(D_WARNING, request,
2798 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2799 request->rq_receiving_reply,
2800 request->rq_req_unlinked,
2801 request->rq_reply_unlinked);
2806 static void ptlrpc_free_request(struct ptlrpc_request *req)
2808 spin_lock(&req->rq_lock);
2810 spin_unlock(&req->rq_lock);
2812 if (req->rq_commit_cb)
2813 req->rq_commit_cb(req);
2814 list_del_init(&req->rq_replay_list);
2816 __ptlrpc_req_finished(req, 1);
2820 * the request is committed and dropped from the replay list of its import
2822 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2824 struct obd_import *imp = req->rq_import;
2826 spin_lock(&imp->imp_lock);
2827 if (list_empty(&req->rq_replay_list)) {
2828 spin_unlock(&imp->imp_lock);
2832 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2833 if (imp->imp_replay_cursor == &req->rq_replay_list)
2834 imp->imp_replay_cursor = req->rq_replay_list.next;
2835 ptlrpc_free_request(req);
2838 spin_unlock(&imp->imp_lock);
2840 EXPORT_SYMBOL(ptlrpc_request_committed);
2843 * Iterates through replay_list on import and prunes
2844 * all requests have transno smaller than last_committed for the
2845 * import and don't have rq_replay set.
2846 * Since requests are sorted in transno order, stops when meetign first
2847 * transno bigger than last_committed.
2848 * caller must hold imp->imp_lock
2850 void ptlrpc_free_committed(struct obd_import *imp)
2852 struct ptlrpc_request *req, *saved;
2853 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2854 bool skip_committed_list = true;
2857 LASSERT(imp != NULL);
2858 assert_spin_locked(&imp->imp_lock);
2860 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2861 imp->imp_generation == imp->imp_last_generation_checked) {
2862 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2863 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2866 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2867 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2868 imp->imp_generation);
2870 if (imp->imp_generation != imp->imp_last_generation_checked ||
2871 imp->imp_last_transno_checked == 0)
2872 skip_committed_list = false;
2874 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2875 imp->imp_last_generation_checked = imp->imp_generation;
2877 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2879 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2880 LASSERT(req != last_req);
2883 if (req->rq_transno == 0) {
2884 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2887 if (req->rq_import_generation < imp->imp_generation) {
2888 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2892 /* not yet committed */
2893 if (req->rq_transno > imp->imp_peer_committed_transno) {
2894 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2898 if (req->rq_replay) {
2899 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2900 list_move_tail(&req->rq_replay_list,
2901 &imp->imp_committed_list);
2905 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2906 imp->imp_peer_committed_transno);
2908 ptlrpc_free_request(req);
2911 if (skip_committed_list)
2914 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2916 LASSERT(req->rq_transno != 0);
2917 if (req->rq_import_generation < imp->imp_generation ||
2919 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2920 req->rq_import_generation <
2921 imp->imp_generation ? "stale" : "closed");
2923 if (imp->imp_replay_cursor == &req->rq_replay_list)
2924 imp->imp_replay_cursor =
2925 req->rq_replay_list.next;
2927 ptlrpc_free_request(req);
2934 void ptlrpc_cleanup_client(struct obd_import *imp)
2941 * Schedule previously sent request for resend.
2942 * For bulk requests we assign new xid (to avoid problems with
2943 * lost replies and therefore several transfers landing into same buffer
2944 * from different sending attempts).
2946 void ptlrpc_resend_req(struct ptlrpc_request *req)
2948 DEBUG_REQ(D_HA, req, "going to resend");
2949 spin_lock(&req->rq_lock);
2952 * Request got reply but linked to the import list still.
2953 * Let ptlrpc_check_set() process it.
2955 if (ptlrpc_client_replied(req)) {
2956 spin_unlock(&req->rq_lock);
2957 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2961 req->rq_status = -EAGAIN;
2964 req->rq_net_err = 0;
2965 req->rq_timedout = 0;
2967 ptlrpc_client_wake_req(req);
2968 spin_unlock(&req->rq_lock);
2971 /* XXX: this function and rq_status are currently unused */
2972 void ptlrpc_restart_req(struct ptlrpc_request *req)
2974 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2975 req->rq_status = -ERESTARTSYS;
2977 spin_lock(&req->rq_lock);
2978 req->rq_restart = 1;
2979 req->rq_timedout = 0;
2980 ptlrpc_client_wake_req(req);
2981 spin_unlock(&req->rq_lock);
2985 * Grab additional reference on a request \a req
2987 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2990 atomic_inc(&req->rq_refcount);
2993 EXPORT_SYMBOL(ptlrpc_request_addref);
2996 * Add a request to import replay_list.
2997 * Must be called under imp_lock
2999 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3000 struct obd_import *imp)
3002 struct ptlrpc_request *iter;
3004 assert_spin_locked(&imp->imp_lock);
3006 if (req->rq_transno == 0) {
3007 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3012 * clear this for new requests that were resent as well
3013 * as resent replayed requests.
3015 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3017 /* don't re-add requests that have been replayed */
3018 if (!list_empty(&req->rq_replay_list))
3021 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3023 spin_lock(&req->rq_lock);
3025 spin_unlock(&req->rq_lock);
3027 LASSERT(imp->imp_replayable);
3028 /* Balanced in ptlrpc_free_committed, usually. */
3029 ptlrpc_request_addref(req);
3030 list_for_each_entry_reverse(iter, &imp->imp_replay_list,
3033 * We may have duplicate transnos if we create and then
3034 * open a file, or for closes retained if to match creating
3035 * opens, so use req->rq_xid as a secondary key.
3036 * (See bugs 684, 685, and 428.)
3037 * XXX no longer needed, but all opens need transnos!
3039 if (iter->rq_transno > req->rq_transno)
3042 if (iter->rq_transno == req->rq_transno) {
3043 LASSERT(iter->rq_xid != req->rq_xid);
3044 if (iter->rq_xid > req->rq_xid)
3048 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3052 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3056 * Send request and wait until it completes.
3057 * Returns request processing status.
3059 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3061 struct ptlrpc_request_set *set;
3065 LASSERT(req->rq_set == NULL);
3066 LASSERT(!req->rq_receiving_reply);
3068 set = ptlrpc_prep_set();
3070 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3074 /* for distributed debugging */
3075 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3077 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3078 ptlrpc_request_addref(req);
3079 ptlrpc_set_add_req(set, req);
3080 rc = ptlrpc_set_wait(NULL, set);
3081 ptlrpc_set_destroy(set);
3085 EXPORT_SYMBOL(ptlrpc_queue_wait);
3088 * Callback used for replayed requests reply processing.
3089 * In case of successful reply calls registered request replay callback.
3090 * In case of error restart replay process.
3092 static int ptlrpc_replay_interpret(const struct lu_env *env,
3093 struct ptlrpc_request *req,
3096 struct ptlrpc_replay_async_args *aa = args;
3097 struct obd_import *imp = req->rq_import;
3100 atomic_dec(&imp->imp_replay_inflight);
3103 * Note: if it is bulk replay (MDS-MDS replay), then even if
3104 * server got the request, but bulk transfer timeout, let's
3105 * replay the bulk req again
3107 if (!ptlrpc_client_replied(req) ||
3109 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3110 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3111 GOTO(out, rc = -ETIMEDOUT);
3114 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3115 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3116 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3117 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3119 /** VBR: check version failure */
3120 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3121 /** replay was failed due to version mismatch */
3122 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3123 spin_lock(&imp->imp_lock);
3124 imp->imp_vbr_failed = 1;
3125 spin_unlock(&imp->imp_lock);
3126 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3128 /** The transno had better not change over replay. */
3129 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3130 lustre_msg_get_transno(req->rq_repmsg) ||
3131 lustre_msg_get_transno(req->rq_repmsg) == 0,
3133 lustre_msg_get_transno(req->rq_reqmsg),
3134 lustre_msg_get_transno(req->rq_repmsg));
3137 spin_lock(&imp->imp_lock);
3138 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3139 spin_unlock(&imp->imp_lock);
3140 LASSERT(imp->imp_last_replay_transno);
3142 /* transaction number shouldn't be bigger than the latest replayed */
3143 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3144 DEBUG_REQ(D_ERROR, req,
3145 "Reported transno=%llu is bigger than replayed=%llu",
3147 lustre_msg_get_transno(req->rq_reqmsg));
3148 GOTO(out, rc = -EINVAL);
3151 DEBUG_REQ(D_HA, req, "got reply");
3153 /* let the callback do fixups, possibly including in the request */
3154 if (req->rq_replay_cb)
3155 req->rq_replay_cb(req);
3157 if (ptlrpc_client_replied(req) &&
3158 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3159 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3160 lustre_msg_get_status(req->rq_repmsg),
3161 aa->praa_old_status);
3164 * Note: If the replay fails for MDT-MDT recovery, let's
3165 * abort all of the following requests in the replay
3166 * and sending list, because MDT-MDT update requests
3167 * are dependent on each other, see LU-7039
3169 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3170 struct ptlrpc_request *free_req;
3171 struct ptlrpc_request *tmp;
3173 spin_lock(&imp->imp_lock);
3174 list_for_each_entry_safe(free_req, tmp,
3175 &imp->imp_replay_list,
3177 ptlrpc_free_request(free_req);
3180 list_for_each_entry_safe(free_req, tmp,
3181 &imp->imp_committed_list,
3183 ptlrpc_free_request(free_req);
3186 list_for_each_entry_safe(free_req, tmp,
3187 &imp->imp_delayed_list,
3189 spin_lock(&free_req->rq_lock);
3190 free_req->rq_err = 1;
3191 free_req->rq_status = -EIO;
3192 ptlrpc_client_wake_req(free_req);
3193 spin_unlock(&free_req->rq_lock);
3196 list_for_each_entry_safe(free_req, tmp,
3197 &imp->imp_sending_list,
3199 spin_lock(&free_req->rq_lock);
3200 free_req->rq_err = 1;
3201 free_req->rq_status = -EIO;
3202 ptlrpc_client_wake_req(free_req);
3203 spin_unlock(&free_req->rq_lock);
3205 spin_unlock(&imp->imp_lock);
3208 /* Put it back for re-replay. */
3209 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3213 * Errors while replay can set transno to 0, but
3214 * imp_last_replay_transno shouldn't be set to 0 anyway
3216 if (req->rq_transno == 0)
3217 CERROR("Transno is 0 during replay!\n");
3219 /* continue with recovery */
3220 rc = ptlrpc_import_recovery_state_machine(imp);
3222 req->rq_send_state = aa->praa_old_state;
3225 /* this replay failed, so restart recovery */
3226 ptlrpc_connect_import(imp);
3232 * Prepares and queues request for replay.
3233 * Adds it to ptlrpcd queue for actual sending.
3234 * Returns 0 on success.
3236 int ptlrpc_replay_req(struct ptlrpc_request *req)
3238 struct ptlrpc_replay_async_args *aa;
3242 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3244 aa = ptlrpc_req_async_args(aa, req);
3245 memset(aa, 0, sizeof(*aa));
3247 /* Prepare request to be resent with ptlrpcd */
3248 aa->praa_old_state = req->rq_send_state;
3249 req->rq_send_state = LUSTRE_IMP_REPLAY;
3250 req->rq_phase = RQ_PHASE_NEW;
3251 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3253 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3255 req->rq_interpret_reply = ptlrpc_replay_interpret;
3256 /* Readjust the timeout for current conditions */
3257 ptlrpc_at_set_req_timeout(req);
3259 /* Tell server net_latency to calculate how long to wait for reply. */
3260 lustre_msg_set_service_timeout(req->rq_reqmsg,
3261 ptlrpc_at_get_net_latency(req));
3262 DEBUG_REQ(D_HA, req, "REPLAY");
3264 atomic_inc(&req->rq_import->imp_replay_inflight);
3265 spin_lock(&req->rq_lock);
3266 req->rq_early_free_repbuf = 0;
3267 spin_unlock(&req->rq_lock);
3268 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3270 ptlrpcd_add_req(req);
3275 * Aborts all in-flight request on import \a imp sending and delayed lists
3277 void ptlrpc_abort_inflight(struct obd_import *imp)
3279 struct ptlrpc_request *req, *n;
3283 * Make sure that no new requests get processed for this import.
3284 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3285 * this flag and then putting requests on sending_list or delayed_list.
3287 assert_spin_locked(&imp->imp_lock);
3290 * XXX locking? Maybe we should remove each request with the list
3291 * locked? Also, how do we know if the requests on the list are
3292 * being freed at this time?
3294 list_for_each_entry_safe(req, n, &imp->imp_sending_list, rq_list) {
3295 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3297 spin_lock(&req->rq_lock);
3298 if (req->rq_import_generation < imp->imp_generation) {
3300 req->rq_status = -EIO;
3301 ptlrpc_client_wake_req(req);
3303 spin_unlock(&req->rq_lock);
3306 list_for_each_entry_safe(req, n, &imp->imp_delayed_list, rq_list) {
3307 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3309 spin_lock(&req->rq_lock);
3310 if (req->rq_import_generation < imp->imp_generation) {
3312 req->rq_status = -EIO;
3313 ptlrpc_client_wake_req(req);
3315 spin_unlock(&req->rq_lock);
3319 * Last chance to free reqs left on the replay list, but we
3320 * will still leak reqs that haven't committed.
3322 if (imp->imp_replayable)
3323 ptlrpc_free_committed(imp);
3329 * Abort all uncompleted requests in request set \a set
3331 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3333 struct ptlrpc_request *req, *tmp;
3335 LASSERT(set != NULL);
3337 list_for_each_entry_safe(req, tmp, &set->set_requests, rq_set_chain) {
3338 spin_lock(&req->rq_lock);
3339 if (req->rq_phase != RQ_PHASE_RPC) {
3340 spin_unlock(&req->rq_lock);
3345 req->rq_status = -EINTR;
3346 ptlrpc_client_wake_req(req);
3347 spin_unlock(&req->rq_lock);
3352 * Initialize the XID for the node. This is common among all requests on
3353 * this node, and only requires the property that it is monotonically
3354 * increasing. It does not need to be sequential. Since this is also used
3355 * as the RDMA match bits, it is important that a single client NOT have
3356 * the same match bits for two different in-flight requests, hence we do
3357 * NOT want to have an XID per target or similar.
3359 * To avoid an unlikely collision between match bits after a client reboot
3360 * (which would deliver old data into the wrong RDMA buffer) initialize
3361 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3362 * If the time is clearly incorrect, we instead use a 62-bit random number.
3363 * In the worst case the random number will overflow 1M RPCs per second in
3364 * 9133 years, or permutations thereof.
3366 #define YEAR_2004 (1ULL << 30)
3367 void ptlrpc_init_xid(void)
3369 time64_t now = ktime_get_real_seconds();
3372 if (now < YEAR_2004) {
3373 get_random_bytes(&xid, sizeof(xid));
3375 xid |= (1ULL << 61);
3377 xid = (u64)now << 20;
3380 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3381 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3383 xid &= PTLRPC_BULK_OPS_MASK;
3384 atomic64_set(&ptlrpc_last_xid, xid);
3388 * Increase xid and returns resulting new value to the caller.
3390 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3391 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3392 * itself uses the last bulk xid needed, so the server can determine the
3393 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3394 * xid must align to a power-of-two value.
3396 * This is assumed to be true due to the initial ptlrpc_last_xid
3397 * value also being initialized to a power-of-two value. LU-1431
3399 __u64 ptlrpc_next_xid(void)
3401 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3405 * If request has a new allocated XID (new request or EINPROGRESS resend),
3406 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3407 * request to ensure previous bulk fails and avoid problems with lost replies
3408 * and therefore several transfers landing into the same buffer from different
3411 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3413 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3415 LASSERT(bd != NULL);
3418 * Generate new matchbits for all resend requests, including
3421 if (req->rq_resend) {
3422 __u64 old_mbits = req->rq_mbits;
3425 * First time resend on -EINPROGRESS will generate new xid,
3426 * so we can actually use the rq_xid as rq_mbits in such case,
3427 * however, it's bit hard to distinguish such resend with a
3428 * 'resend for the -EINPROGRESS resend'. To make it simple,
3429 * we opt to generate mbits for all resend cases.
3431 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3433 req->rq_mbits = ptlrpc_next_xid();
3436 * Old version transfers rq_xid to peer as
3439 spin_lock(&req->rq_import->imp_lock);
3440 list_del_init(&req->rq_unreplied_list);
3441 ptlrpc_assign_next_xid_nolock(req);
3442 spin_unlock(&req->rq_import->imp_lock);
3443 req->rq_mbits = req->rq_xid;
3445 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3446 old_mbits, req->rq_mbits);
3447 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3448 /* Request being sent first time, use xid as matchbits. */
3449 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3450 || req->rq_mbits == 0) {
3451 req->rq_mbits = req->rq_xid;
3453 req->rq_mbits -= bd->bd_md_count - 1;
3457 * Replay request, xid and matchbits have already been
3458 * correctly assigned.
3464 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3465 * that server can infer the number of bulks that were prepared,
3468 req->rq_mbits += bd->bd_md_count - 1;
3471 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3472 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3474 * It's ok to directly set the rq_xid here, since this xid bump
3475 * won't affect the request position in unreplied list.
3477 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3478 req->rq_xid = req->rq_mbits;
3482 * Get a glimpse at what next xid value might have been.
3483 * Returns possible next xid.
3485 __u64 ptlrpc_sample_next_xid(void)
3487 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3489 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3492 * Functions for operating ptlrpc workers.
3494 * A ptlrpc work is a function which will be running inside ptlrpc context.
3495 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3497 * 1. after a work is created, it can be used many times, that is:
3498 * handler = ptlrpcd_alloc_work();
3499 * ptlrpcd_queue_work();
3501 * queue it again when necessary:
3502 * ptlrpcd_queue_work();
3503 * ptlrpcd_destroy_work();
3504 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3505 * it will only be queued once in any time. Also as its name implies, it may
3506 * have delay before it really runs by ptlrpcd thread.
3508 struct ptlrpc_work_async_args {
3509 int (*cb)(const struct lu_env *, void *);
3513 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3515 /* re-initialize the req */
3516 req->rq_timeout = obd_timeout;
3517 req->rq_sent = ktime_get_real_seconds();
3518 req->rq_deadline = req->rq_sent + req->rq_timeout;
3519 req->rq_phase = RQ_PHASE_INTERPRET;
3520 req->rq_next_phase = RQ_PHASE_COMPLETE;
3521 req->rq_xid = ptlrpc_next_xid();
3522 req->rq_import_generation = req->rq_import->imp_generation;
3524 ptlrpcd_add_req(req);
3527 static int work_interpreter(const struct lu_env *env,
3528 struct ptlrpc_request *req, void *args, int rc)
3530 struct ptlrpc_work_async_args *arg = args;
3532 LASSERT(ptlrpcd_check_work(req));
3533 LASSERT(arg->cb != NULL);
3535 rc = arg->cb(env, arg->cbdata);
3537 list_del_init(&req->rq_set_chain);
3540 if (atomic_dec_return(&req->rq_refcount) > 1) {
3541 atomic_set(&req->rq_refcount, 2);
3542 ptlrpcd_add_work_req(req);
3547 static int worker_format;
3549 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3551 return req->rq_pill.rc_fmt == (void *)&worker_format;
3555 * Create a work for ptlrpc.
3557 void *ptlrpcd_alloc_work(struct obd_import *imp,
3558 int (*cb)(const struct lu_env *, void *), void *cbdata)
3560 struct ptlrpc_request *req = NULL;
3561 struct ptlrpc_work_async_args *args;
3567 RETURN(ERR_PTR(-EINVAL));
3569 /* copy some code from deprecated fakereq. */
3570 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3572 CERROR("ptlrpc: run out of memory!\n");
3573 RETURN(ERR_PTR(-ENOMEM));
3576 ptlrpc_cli_req_init(req);
3578 req->rq_send_state = LUSTRE_IMP_FULL;
3579 req->rq_type = PTL_RPC_MSG_REQUEST;
3580 req->rq_import = class_import_get(imp);
3581 req->rq_interpret_reply = work_interpreter;
3582 /* don't want reply */
3583 req->rq_no_delay = req->rq_no_resend = 1;
3584 req->rq_pill.rc_fmt = (void *)&worker_format;
3586 args = ptlrpc_req_async_args(args, req);
3588 args->cbdata = cbdata;
3592 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3594 void ptlrpcd_destroy_work(void *handler)
3596 struct ptlrpc_request *req = handler;
3599 ptlrpc_req_finished(req);
3601 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3603 int ptlrpcd_queue_work(void *handler)
3605 struct ptlrpc_request *req = handler;
3608 * Check if the req is already being queued.
3610 * Here comes a trick: it lacks a way of checking if a req is being
3611 * processed reliably in ptlrpc. Here I have to use refcount of req
3612 * for this purpose. This is okay because the caller should use this
3613 * req as opaque data. - Jinshan
3615 LASSERT(atomic_read(&req->rq_refcount) > 0);
3616 if (atomic_inc_return(&req->rq_refcount) == 2)
3617 ptlrpcd_add_work_req(req);
3620 EXPORT_SYMBOL(ptlrpcd_queue_work);