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/
32 /** Implementation of client-side PortalRPC interfaces */
34 #define DEBUG_SUBSYSTEM S_RPC
36 #include <linux/delay.h>
37 #include <linux/random.h>
39 #include <lnet/lib-lnet.h>
40 #include <obd_support.h>
41 #include <obd_class.h>
42 #include <lustre_lib.h>
43 #include <lustre_ha.h>
44 #include <lustre_import.h>
45 #include <lustre_req_layout.h>
47 #include "ptlrpc_internal.h"
49 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
50 struct page *page, int pageoffset,
53 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
56 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
57 struct page *page, int pageoffset,
60 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
63 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
67 for (i = 0; i < desc->bd_iov_count ; i++)
68 put_page(desc->bd_vec[i].bv_page);
71 static int ptlrpc_prep_bulk_frag_pages(struct ptlrpc_bulk_desc *desc,
74 unsigned int offset = (unsigned long)frag & ~PAGE_MASK;
78 int page_len = min_t(unsigned int, PAGE_SIZE - offset,
80 unsigned long vaddr = (unsigned long)frag;
82 ptlrpc_prep_bulk_page_nopin(desc,
83 lnet_kvaddr_to_page(vaddr),
93 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
94 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
95 .release_frags = ptlrpc_release_bulk_page_pin,
97 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
99 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
100 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
101 .release_frags = ptlrpc_release_bulk_noop,
102 .add_iov_frag = ptlrpc_prep_bulk_frag_pages,
104 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
106 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
107 static int ptlrpcd_check_work(struct ptlrpc_request *req);
108 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
111 * Initialize passed in client structure \a cl.
113 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
114 struct ptlrpc_client *cl)
116 cl->cli_request_portal = req_portal;
117 cl->cli_reply_portal = rep_portal;
120 EXPORT_SYMBOL(ptlrpc_init_client);
123 * Return PortalRPC connection for remore uud \a uuid
125 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
126 lnet_nid_t nid4refnet)
128 struct ptlrpc_connection *c;
130 struct lnet_process_id peer;
134 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
135 * before accessing its values.
137 /* coverity[uninit_use_in_call] */
138 peer.nid = nid4refnet;
139 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
141 CNETERR("cannot find peer %s!\n", uuid->uuid);
145 c = ptlrpc_connection_get(peer, self, uuid);
147 memcpy(c->c_remote_uuid.uuid,
148 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
151 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
157 * Allocate and initialize new bulk descriptor on the sender.
158 * Returns pointer to the descriptor or NULL on error.
160 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
161 unsigned int max_brw,
162 enum ptlrpc_bulk_op_type type,
164 const struct ptlrpc_bulk_frag_ops *ops)
166 struct ptlrpc_bulk_desc *desc;
169 LASSERT(ops->add_kiov_frag != NULL);
171 if (max_brw > PTLRPC_BULK_OPS_COUNT)
174 if (nfrags > LNET_MAX_IOV * max_brw)
181 OBD_ALLOC_LARGE(desc->bd_vec,
182 nfrags * sizeof(*desc->bd_vec));
186 spin_lock_init(&desc->bd_lock);
187 init_waitqueue_head(&desc->bd_waitq);
188 desc->bd_max_iov = nfrags;
189 desc->bd_iov_count = 0;
190 desc->bd_portal = portal;
191 desc->bd_type = type;
192 desc->bd_md_count = 0;
193 desc->bd_nob_last = LNET_MTU;
194 desc->bd_frag_ops = ops;
195 LASSERT(max_brw > 0);
196 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
198 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
199 * node. Negotiated ocd_brw_size will always be <= this number.
201 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
202 LNetInvalidateMDHandle(&desc->bd_mds[i]);
211 * Prepare bulk descriptor for specified outgoing request \a req that
212 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
213 * the bulk to be sent. Used on client-side.
214 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
217 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
219 unsigned int max_brw,
222 const struct ptlrpc_bulk_frag_ops
225 struct obd_import *imp = req->rq_import;
226 struct ptlrpc_bulk_desc *desc;
229 LASSERT(ptlrpc_is_bulk_op_passive(type));
231 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
235 desc->bd_import = class_import_get(imp);
238 desc->bd_cbid.cbid_fn = client_bulk_callback;
239 desc->bd_cbid.cbid_arg = desc;
241 /* This makes req own desc, and free it when she frees herself */
246 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
248 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
249 struct page *page, int pageoffset, int len,
252 struct bio_vec *kiov;
254 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
255 LASSERT(page != NULL);
256 LASSERT(pageoffset >= 0);
258 LASSERT(pageoffset + len <= PAGE_SIZE);
260 kiov = &desc->bd_vec[desc->bd_iov_count];
262 if (((desc->bd_iov_count % LNET_MAX_IOV) == 0) ||
263 ((desc->bd_nob_last + len) > LNET_MTU)) {
264 desc->bd_mds_off[desc->bd_md_count] = desc->bd_iov_count;
266 desc->bd_nob_last = 0;
267 LASSERT(desc->bd_md_count <= PTLRPC_BULK_OPS_COUNT);
270 desc->bd_nob_last += len;
276 kiov->bv_page = page;
277 kiov->bv_offset = pageoffset;
280 desc->bd_iov_count++;
282 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
284 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
288 LASSERT(desc != NULL);
289 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
290 LASSERT(desc->bd_refs == 0); /* network hands off */
291 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
292 LASSERT(desc->bd_frag_ops != NULL);
294 sptlrpc_enc_pool_put_pages(desc);
297 class_export_put(desc->bd_export);
299 class_import_put(desc->bd_import);
301 if (desc->bd_frag_ops->release_frags != NULL)
302 desc->bd_frag_ops->release_frags(desc);
304 OBD_FREE_LARGE(desc->bd_vec,
305 desc->bd_max_iov * sizeof(*desc->bd_vec));
309 EXPORT_SYMBOL(ptlrpc_free_bulk);
312 * Set server timelimit for this req, i.e. how long are we willing to wait
313 * for reply before timing out this request.
315 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
317 LASSERT(req->rq_import);
320 /* non-AT settings */
322 * \a imp_server_timeout means this is reverse import and
323 * we send (currently only) ASTs to the client and cannot afford
324 * to wait too long for the reply, otherwise the other client
325 * (because of which we are sending this request) would
326 * timeout waiting for us
328 req->rq_timeout = req->rq_import->imp_server_timeout ?
329 obd_timeout / 2 : obd_timeout;
331 struct imp_at *at = &req->rq_import->imp_at;
335 idx = import_at_get_index(req->rq_import,
336 req->rq_request_portal);
337 serv_est = at_get(&at->iat_service_estimate[idx]);
339 * Currently a 32 bit value is sent over the
340 * wire for rq_timeout so please don't change this
341 * to time64_t. The work for LU-1158 will in time
342 * replace rq_timeout with a 64 bit nanosecond value
344 req->rq_timeout = at_est2timeout(serv_est);
347 * We could get even fancier here, using history to predict increased
350 * Let the server know what this RPC timeout is by putting it in the
353 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
355 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
357 /* Adjust max service estimate based on server value */
358 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
365 LASSERT(req->rq_import);
366 at = &req->rq_import->imp_at;
368 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
370 * max service estimates are tracked on the server side,
371 * so just keep minimal history here
373 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
376 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
377 req->rq_import->imp_obd->obd_name,
378 req->rq_request_portal,
379 oldse, at_get(&at->iat_service_estimate[idx]));
382 /* Expected network latency per remote node (secs) */
383 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
385 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
388 /* Adjust expected network latency */
389 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
390 timeout_t service_timeout)
392 time64_t now = ktime_get_real_seconds();
397 LASSERT(req->rq_import);
399 if (service_timeout > now - req->rq_sent + 3) {
401 * b=16408, however, this can also happen if early reply
402 * is lost and client RPC is expired and resent, early reply
403 * or reply of original RPC can still be fit in reply buffer
404 * of resent RPC, now client is measuring time from the
405 * resent time, but server sent back service time of original
408 CDEBUG_LIMIT((lustre_msg_get_flags(req->rq_reqmsg) &
409 MSG_RESENT) ? D_ADAPTTO : D_WARNING,
410 "Reported service time %u > total measured time %lld\n",
411 service_timeout, now - req->rq_sent);
415 /* Network latency is total time less server processing time,
418 nl = max_t(timeout_t, now - req->rq_sent - service_timeout, 0) + 1;
419 at = &req->rq_import->imp_at;
421 oldnl = at_measured(&at->iat_net_latency, nl);
424 "The network latency for %s (nid %s) has changed from %d to %d\n",
425 req->rq_import->imp_obd->obd_name,
426 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
427 oldnl, at_get(&at->iat_net_latency));
430 static int unpack_reply(struct ptlrpc_request *req)
434 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
435 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
437 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
443 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
445 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
453 * Handle an early reply message, called with the rq_lock held.
454 * If anything goes wrong just ignore it - same as if it never happened
456 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
457 __must_hold(&req->rq_lock)
459 struct ptlrpc_request *early_req;
460 timeout_t service_timeout;
466 spin_unlock(&req->rq_lock);
468 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
470 spin_lock(&req->rq_lock);
474 rc = unpack_reply(early_req);
476 sptlrpc_cli_finish_early_reply(early_req);
477 spin_lock(&req->rq_lock);
482 * Use new timeout value just to adjust the local value for this
483 * request, don't include it into at_history. It is unclear yet why
484 * service time increased and should it be counted or skipped, e.g.
485 * that can be recovery case or some error or server, the real reply
486 * will add all new data if it is worth to add.
488 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
489 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
491 /* Network latency can be adjusted, it is pure network delays */
492 service_timeout = lustre_msg_get_service_timeout(early_req->rq_repmsg);
493 ptlrpc_at_adj_net_latency(req, service_timeout);
495 sptlrpc_cli_finish_early_reply(early_req);
497 spin_lock(&req->rq_lock);
498 olddl = req->rq_deadline;
500 * server assumes it now has rq_timeout from when the request
501 * arrived, so the client should give it at least that long.
502 * since we don't know the arrival time we'll use the original
505 req->rq_deadline = req->rq_sent + req->rq_timeout +
506 ptlrpc_at_get_net_latency(req);
508 /* The below message is checked in replay-single.sh test_65{a,b} */
509 /* The below message is checked in sanity-{gss,krb5} test_8 */
510 DEBUG_REQ(D_ADAPTTO, req,
511 "Early reply #%d, new deadline in %llds (%llds)",
513 req->rq_deadline - ktime_get_real_seconds(),
514 req->rq_deadline - olddl);
519 static struct kmem_cache *request_cache;
521 int ptlrpc_request_cache_init(void)
523 request_cache = kmem_cache_create("ptlrpc_cache",
524 sizeof(struct ptlrpc_request),
525 0, SLAB_HWCACHE_ALIGN, NULL);
526 return request_cache ? 0 : -ENOMEM;
529 void ptlrpc_request_cache_fini(void)
531 kmem_cache_destroy(request_cache);
534 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
536 struct ptlrpc_request *req;
538 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
542 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
544 OBD_SLAB_FREE_PTR(req, request_cache);
548 * Wind down request pool \a pool.
549 * Frees all requests from the pool too
551 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
553 struct ptlrpc_request *req;
555 LASSERT(pool != NULL);
557 spin_lock(&pool->prp_lock);
558 while ((req = list_first_entry_or_null(&pool->prp_req_list,
559 struct ptlrpc_request,
561 list_del(&req->rq_list);
562 LASSERT(req->rq_reqbuf);
563 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
564 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
565 ptlrpc_request_cache_free(req);
567 spin_unlock(&pool->prp_lock);
568 OBD_FREE(pool, sizeof(*pool));
570 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
573 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
575 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
580 while (size < pool->prp_rq_size)
583 LASSERTF(list_empty(&pool->prp_req_list) ||
584 size == pool->prp_rq_size,
585 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
586 pool->prp_rq_size, size);
588 pool->prp_rq_size = size;
589 for (i = 0; i < num_rq; i++) {
590 struct ptlrpc_request *req;
591 struct lustre_msg *msg;
593 req = ptlrpc_request_cache_alloc(GFP_NOFS);
596 OBD_ALLOC_LARGE(msg, size);
598 ptlrpc_request_cache_free(req);
601 req->rq_reqbuf = msg;
602 req->rq_reqbuf_len = size;
604 spin_lock(&pool->prp_lock);
605 list_add_tail(&req->rq_list, &pool->prp_req_list);
606 spin_unlock(&pool->prp_lock);
610 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
613 * Create and initialize new request pool with given attributes:
614 * \a num_rq - initial number of requests to create for the pool
615 * \a msgsize - maximum message size possible for requests in thid pool
616 * \a populate_pool - function to be called when more requests need to be added
618 * Returns pointer to newly created pool or NULL on error.
620 struct ptlrpc_request_pool *
621 ptlrpc_init_rq_pool(int num_rq, int msgsize,
622 int (*populate_pool)(struct ptlrpc_request_pool *, int))
624 struct ptlrpc_request_pool *pool;
631 * Request next power of two for the allocation, because internally
632 * kernel would do exactly this
634 spin_lock_init(&pool->prp_lock);
635 INIT_LIST_HEAD(&pool->prp_req_list);
636 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
637 pool->prp_populate = populate_pool;
639 populate_pool(pool, num_rq);
643 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
646 * Fetches one request from pool \a pool
648 static struct ptlrpc_request *
649 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
651 struct ptlrpc_request *request;
652 struct lustre_msg *reqbuf;
657 spin_lock(&pool->prp_lock);
660 * See if we have anything in a pool, and bail out if nothing,
661 * in writeout path, where this matters, this is safe to do, because
662 * nothing is lost in this case, and when some in-flight requests
663 * complete, this code will be called again.
665 if (unlikely(list_empty(&pool->prp_req_list))) {
666 spin_unlock(&pool->prp_lock);
670 request = list_first_entry(&pool->prp_req_list, struct ptlrpc_request,
672 list_del_init(&request->rq_list);
673 spin_unlock(&pool->prp_lock);
675 LASSERT(request->rq_reqbuf);
676 LASSERT(request->rq_pool);
678 reqbuf = request->rq_reqbuf;
679 memset(request, 0, sizeof(*request));
680 request->rq_reqbuf = reqbuf;
681 request->rq_reqbuf_len = pool->prp_rq_size;
682 request->rq_pool = pool;
688 * Returns freed \a request to pool.
690 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
692 struct ptlrpc_request_pool *pool = request->rq_pool;
694 spin_lock(&pool->prp_lock);
695 LASSERT(list_empty(&request->rq_list));
696 LASSERT(!request->rq_receiving_reply);
697 list_add_tail(&request->rq_list, &pool->prp_req_list);
698 spin_unlock(&pool->prp_lock);
701 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
703 struct obd_import *imp = req->rq_import;
704 struct ptlrpc_request *iter;
706 assert_spin_locked(&imp->imp_lock);
707 LASSERT(list_empty(&req->rq_unreplied_list));
709 /* unreplied list is sorted by xid in ascending order */
710 list_for_each_entry_reverse(iter, &imp->imp_unreplied_list,
712 LASSERT(req->rq_xid != iter->rq_xid);
713 if (req->rq_xid < iter->rq_xid)
715 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
718 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
721 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
723 req->rq_xid = ptlrpc_next_xid();
724 ptlrpc_add_unreplied(req);
727 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
729 spin_lock(&req->rq_import->imp_lock);
730 ptlrpc_assign_next_xid_nolock(req);
731 spin_unlock(&req->rq_import->imp_lock);
734 static atomic64_t ptlrpc_last_xid;
736 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
738 spin_lock(&req->rq_import->imp_lock);
739 list_del_init(&req->rq_unreplied_list);
740 ptlrpc_assign_next_xid_nolock(req);
741 spin_unlock(&req->rq_import->imp_lock);
742 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
745 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
747 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
751 opc = lustre_msg_get_opc(req->rq_reqmsg);
752 tag = obd_get_mod_rpc_slot(cli, opc);
753 lustre_msg_set_tag(req->rq_reqmsg, tag);
754 ptlrpc_reassign_next_xid(req);
756 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
758 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
760 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
763 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
764 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
766 obd_put_mod_rpc_slot(cli, opc, tag);
769 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
771 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
772 __u32 version, int opcode, char **bufs,
773 struct ptlrpc_cli_ctx *ctx)
776 struct obd_import *imp;
782 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
783 imp = request->rq_import;
784 lengths = request->rq_pill.rc_area[RCL_CLIENT];
787 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
789 rc = sptlrpc_req_get_ctx(request);
793 sptlrpc_req_set_flavor(request, opcode);
795 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
800 lustre_msg_add_version(request->rq_reqmsg, version);
801 request->rq_send_state = LUSTRE_IMP_FULL;
802 request->rq_type = PTL_RPC_MSG_REQUEST;
804 request->rq_req_cbid.cbid_fn = request_out_callback;
805 request->rq_req_cbid.cbid_arg = request;
807 request->rq_reply_cbid.cbid_fn = reply_in_callback;
808 request->rq_reply_cbid.cbid_arg = request;
810 request->rq_reply_deadline = 0;
811 request->rq_bulk_deadline = 0;
812 request->rq_req_deadline = 0;
813 request->rq_phase = RQ_PHASE_NEW;
814 request->rq_next_phase = RQ_PHASE_UNDEFINED;
816 request->rq_request_portal = imp->imp_client->cli_request_portal;
817 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
819 ptlrpc_at_set_req_timeout(request);
821 lustre_msg_set_opc(request->rq_reqmsg, opcode);
823 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
824 if (cfs_fail_val == opcode) {
825 time64_t *fail_t = NULL, *fail2_t = NULL;
827 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
828 fail_t = &request->rq_bulk_deadline;
829 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
830 fail_t = &request->rq_reply_deadline;
831 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
832 fail_t = &request->rq_req_deadline;
833 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
834 fail_t = &request->rq_reply_deadline;
835 fail2_t = &request->rq_bulk_deadline;
836 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
837 time64_t now = ktime_get_real_seconds();
838 u64 xid = ((u64)now >> 4) << 24;
840 atomic64_set(&ptlrpc_last_xid, xid);
844 *fail_t = ktime_get_real_seconds() +
845 PTLRPC_REQ_LONG_UNLINK;
848 *fail2_t = ktime_get_real_seconds() +
849 PTLRPC_REQ_LONG_UNLINK;
852 * The RPC is infected, let the test to change the
855 msleep(4 * MSEC_PER_SEC);
858 ptlrpc_assign_next_xid(request);
863 LASSERT(!request->rq_pool);
864 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
866 atomic_dec(&imp->imp_reqs);
867 class_import_put(imp);
871 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
874 * Pack request buffers for network transfer, performing necessary encryption
875 * steps if necessary.
877 int ptlrpc_request_pack(struct ptlrpc_request *request,
878 __u32 version, int opcode)
880 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
882 EXPORT_SYMBOL(ptlrpc_request_pack);
885 * Helper function to allocate new request on import \a imp
886 * and possibly using existing request from pool \a pool if provided.
887 * Returns allocated request structure with import field filled or
891 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
892 struct ptlrpc_request_pool *pool)
894 struct ptlrpc_request *request = NULL;
896 request = ptlrpc_request_cache_alloc(GFP_NOFS);
898 if (!request && pool)
899 request = ptlrpc_prep_req_from_pool(pool);
902 ptlrpc_cli_req_init(request);
904 LASSERTF((unsigned long)imp > 0x1000, "%p\n", imp);
905 LASSERT(imp != LP_POISON);
906 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
908 LASSERT(imp->imp_client != LP_POISON);
910 request->rq_import = class_import_get(imp);
911 atomic_inc(&imp->imp_reqs);
913 CERROR("request allocation out of memory\n");
919 static int ptlrpc_reconnect_if_idle(struct obd_import *imp)
924 * initiate connection if needed when the import has been
925 * referenced by the new request to avoid races with disconnect.
926 * serialize this check against conditional state=IDLE
927 * in ptlrpc_disconnect_idle_interpret()
929 spin_lock(&imp->imp_lock);
930 if (imp->imp_state == LUSTRE_IMP_IDLE) {
931 imp->imp_generation++;
932 imp->imp_initiated_at = imp->imp_generation;
933 imp->imp_state = LUSTRE_IMP_NEW;
935 /* connect_import_locked releases imp_lock */
936 rc = ptlrpc_connect_import_locked(imp);
939 ptlrpc_pinger_add_import(imp);
941 spin_unlock(&imp->imp_lock);
947 * Helper function for creating a request.
948 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
949 * buffer structures according to capsule template \a format.
950 * Returns allocated request structure pointer or NULL on error.
952 static struct ptlrpc_request *
953 ptlrpc_request_alloc_internal(struct obd_import *imp,
954 struct ptlrpc_request_pool *pool,
955 const struct req_format *format)
957 struct ptlrpc_request *request;
959 request = __ptlrpc_request_alloc(imp, pool);
963 /* don't make expensive check for idling connection
964 * if it's already connected */
965 if (unlikely(imp->imp_state != LUSTRE_IMP_FULL)) {
966 if (ptlrpc_reconnect_if_idle(imp) < 0) {
967 atomic_dec(&imp->imp_reqs);
968 ptlrpc_request_free(request);
973 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
974 req_capsule_set(&request->rq_pill, format);
979 * Allocate new request structure for import \a imp and initialize its
980 * buffer structure according to capsule template \a format.
982 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
983 const struct req_format *format)
985 return ptlrpc_request_alloc_internal(imp, NULL, format);
987 EXPORT_SYMBOL(ptlrpc_request_alloc);
990 * Allocate new request structure for import \a imp from pool \a pool and
991 * initialize its buffer structure according to capsule template \a format.
993 struct ptlrpc_request *
994 ptlrpc_request_alloc_pool(struct obd_import *imp,
995 struct ptlrpc_request_pool *pool,
996 const struct req_format *format)
998 return ptlrpc_request_alloc_internal(imp, pool, format);
1000 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
1003 * For requests not from pool, free memory of the request structure.
1004 * For requests obtained from a pool earlier, return request back to pool.
1006 void ptlrpc_request_free(struct ptlrpc_request *request)
1008 if (request->rq_pool)
1009 __ptlrpc_free_req_to_pool(request);
1011 ptlrpc_request_cache_free(request);
1013 EXPORT_SYMBOL(ptlrpc_request_free);
1016 * Allocate new request for operatione \a opcode and immediatelly pack it for
1018 * Only used for simple requests like OBD_PING where the only important
1019 * part of the request is operation itself.
1020 * Returns allocated request or NULL on error.
1022 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1023 const struct req_format *format,
1024 __u32 version, int opcode)
1026 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1030 rc = ptlrpc_request_pack(req, version, opcode);
1032 ptlrpc_request_free(req);
1038 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1041 * Allocate and initialize new request set structure on the current CPT.
1042 * Returns a pointer to the newly allocated set structure or NULL on error.
1044 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1046 struct ptlrpc_request_set *set;
1050 cpt = cfs_cpt_current(cfs_cpt_tab, 0);
1051 OBD_CPT_ALLOC(set, cfs_cpt_tab, cpt, sizeof(*set));
1054 atomic_set(&set->set_refcount, 1);
1055 INIT_LIST_HEAD(&set->set_requests);
1056 init_waitqueue_head(&set->set_waitq);
1057 atomic_set(&set->set_new_count, 0);
1058 atomic_set(&set->set_remaining, 0);
1059 spin_lock_init(&set->set_new_req_lock);
1060 INIT_LIST_HEAD(&set->set_new_requests);
1061 set->set_max_inflight = UINT_MAX;
1062 set->set_producer = NULL;
1063 set->set_producer_arg = NULL;
1068 EXPORT_SYMBOL(ptlrpc_prep_set);
1071 * Allocate and initialize new request set structure with flow control
1072 * extension. This extension allows to control the number of requests in-flight
1073 * for the whole set. A callback function to generate requests must be provided
1074 * and the request set will keep the number of requests sent over the wire to
1076 * Returns a pointer to the newly allocated set structure or NULL on error.
1078 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1082 struct ptlrpc_request_set *set;
1084 set = ptlrpc_prep_set();
1088 set->set_max_inflight = max;
1089 set->set_producer = func;
1090 set->set_producer_arg = arg;
1096 * Wind down and free request set structure previously allocated with
1098 * Ensures that all requests on the set have completed and removes
1099 * all requests from the request list in a set.
1100 * If any unsent request happen to be on the list, pretends that they got
1101 * an error in flight and calls their completion handler.
1103 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1105 struct ptlrpc_request *req;
1111 /* Requests on the set should either all be completed, or all be new */
1112 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1113 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1114 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
1115 LASSERT(req->rq_phase == expected_phase);
1119 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1120 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1121 atomic_read(&set->set_remaining), n);
1123 while ((req = list_first_entry_or_null(&set->set_requests,
1124 struct ptlrpc_request,
1126 list_del_init(&req->rq_set_chain);
1128 LASSERT(req->rq_phase == expected_phase);
1130 if (req->rq_phase == RQ_PHASE_NEW) {
1131 ptlrpc_req_interpret(NULL, req, -EBADR);
1132 atomic_dec(&set->set_remaining);
1135 spin_lock(&req->rq_lock);
1137 req->rq_invalid_rqset = 0;
1138 spin_unlock(&req->rq_lock);
1140 ptlrpc_req_finished(req);
1143 LASSERT(atomic_read(&set->set_remaining) == 0);
1145 ptlrpc_reqset_put(set);
1148 EXPORT_SYMBOL(ptlrpc_set_destroy);
1151 * Add a new request to the general purpose request set.
1152 * Assumes request reference from the caller.
1154 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1155 struct ptlrpc_request *req)
1157 if (set == PTLRPCD_SET) {
1158 ptlrpcd_add_req(req);
1162 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1163 LASSERT(list_empty(&req->rq_set_chain));
1165 if (req->rq_allow_intr)
1166 set->set_allow_intr = 1;
1168 /* The set takes over the caller's request reference */
1169 list_add_tail(&req->rq_set_chain, &set->set_requests);
1171 atomic_inc(&set->set_remaining);
1172 req->rq_queued_time = ktime_get_seconds();
1175 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1177 if (set->set_producer)
1179 * If the request set has a producer callback, the RPC must be
1180 * sent straight away
1182 ptlrpc_send_new_req(req);
1184 EXPORT_SYMBOL(ptlrpc_set_add_req);
1187 * Add a request to a request with dedicated server thread
1188 * and wake the thread to make any necessary processing.
1189 * Currently only used for ptlrpcd.
1191 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1192 struct ptlrpc_request *req)
1194 struct ptlrpc_request_set *set = pc->pc_set;
1197 LASSERT(req->rq_set == NULL);
1198 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1200 spin_lock(&set->set_new_req_lock);
1202 * The set takes over the caller's request reference.
1205 req->rq_queued_time = ktime_get_seconds();
1206 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1207 count = atomic_inc_return(&set->set_new_count);
1208 spin_unlock(&set->set_new_req_lock);
1210 /* Only need to call wakeup once for the first entry. */
1212 wake_up(&set->set_waitq);
1215 * XXX: It maybe unnecessary to wakeup all the partners. But to
1216 * guarantee the async RPC can be processed ASAP, we have
1217 * no other better choice. It maybe fixed in future.
1219 for (i = 0; i < pc->pc_npartners; i++)
1220 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1225 * Based on the current state of the import, determine if the request
1226 * can be sent, is an error, or should be delayed.
1228 * Returns true if this request should be delayed. If false, and
1229 * *status is set, then the request can not be sent and *status is the
1230 * error code. If false and status is 0, then request can be sent.
1232 * The imp->imp_lock must be held.
1234 static int ptlrpc_import_delay_req(struct obd_import *imp,
1235 struct ptlrpc_request *req, int *status)
1243 if (req->rq_ctx_init || req->rq_ctx_fini) {
1244 /* always allow ctx init/fini rpc go through */
1245 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1246 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1248 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1249 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1252 * pings or MDS-equivalent STATFS may safely
1255 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1256 D_HA : D_ERROR, req, "IMP_CLOSED");
1258 } else if (ptlrpc_send_limit_expired(req)) {
1259 /* probably doesn't need to be a D_ERROR afterinitial testing */
1260 DEBUG_REQ(D_HA, req, "send limit expired");
1261 *status = -ETIMEDOUT;
1262 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1263 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1264 ;/* allow CONNECT even if import is invalid */
1265 if (atomic_read(&imp->imp_inval_count) != 0) {
1266 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1269 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1270 if (!imp->imp_deactive)
1271 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1272 *status = -ESHUTDOWN; /* b=12940 */
1273 } else if (req->rq_import_generation != imp->imp_generation) {
1274 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1276 } else if (req->rq_send_state != imp->imp_state) {
1277 /* invalidate in progress - any requests should be drop */
1278 if (atomic_read(&imp->imp_inval_count) != 0) {
1279 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1281 } else if (req->rq_no_delay &&
1282 imp->imp_generation != imp->imp_initiated_at) {
1283 /* ignore nodelay for requests initiating connections */
1285 } else if (req->rq_allow_replay &&
1286 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1287 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1288 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1289 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1290 DEBUG_REQ(D_HA, req, "allow during recovery");
1300 * Decide if the error message should be printed to the console or not.
1301 * Makes its decision based on request type, status, and failure frequency.
1303 * \param[in] req request that failed and may need a console message
1305 * \retval false if no message should be printed
1306 * \retval true if console message should be printed
1308 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1310 LASSERT(req->rq_reqmsg != NULL);
1312 /* Suppress particular reconnect errors which are to be expected. */
1313 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1314 /* Suppress timed out reconnect requests */
1315 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1320 * Suppress most unavailable/again reconnect requests, but
1321 * print occasionally so it is clear client is trying to
1322 * connect to a server where no target is running.
1324 if ((err == -ENODEV || err == -EAGAIN) &&
1325 req->rq_import->imp_conn_cnt % 30 != 20)
1329 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1330 /* -EAGAIN is normal when using POSIX flocks */
1333 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1334 (req->rq_xid & 0xf) != 10)
1335 /* Suppress most ping requests, they may fail occasionally */
1342 * Check request processing status.
1343 * Returns the status.
1345 static int ptlrpc_check_status(struct ptlrpc_request *req)
1350 rc = lustre_msg_get_status(req->rq_repmsg);
1351 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1352 struct obd_import *imp = req->rq_import;
1353 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1354 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1356 if (ptlrpc_console_allow(req, opc, rc))
1357 LCONSOLE_ERROR_MSG(0x11,
1358 "%s: operation %s to node %s failed: rc = %d\n",
1359 imp->imp_obd->obd_name,
1361 libcfs_nid2str(nid), rc);
1362 RETURN(rc < 0 ? rc : -EINVAL);
1366 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1372 * save pre-versions of objects into request for replay.
1373 * Versions are obtained from server reply.
1376 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1378 struct lustre_msg *repmsg = req->rq_repmsg;
1379 struct lustre_msg *reqmsg = req->rq_reqmsg;
1380 __u64 *versions = lustre_msg_get_versions(repmsg);
1383 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1387 lustre_msg_set_versions(reqmsg, versions);
1388 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1389 versions[0], versions[1]);
1394 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1396 struct ptlrpc_request *req;
1398 assert_spin_locked(&imp->imp_lock);
1399 if (list_empty(&imp->imp_unreplied_list))
1402 req = list_first_entry(&imp->imp_unreplied_list, struct ptlrpc_request,
1404 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1406 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1407 imp->imp_known_replied_xid = req->rq_xid - 1;
1409 return req->rq_xid - 1;
1413 * Callback function called when client receives RPC reply for \a req.
1414 * Returns 0 on success or error code.
1415 * The return alue would be assigned to req->rq_status by the caller
1416 * as request processing status.
1417 * This function also decides if the request needs to be saved for later replay.
1419 static int after_reply(struct ptlrpc_request *req)
1421 struct obd_import *imp = req->rq_import;
1422 struct obd_device *obd = req->rq_import->imp_obd;
1429 LASSERT(obd != NULL);
1430 /* repbuf must be unlinked */
1431 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1433 if (req->rq_reply_truncated) {
1434 if (ptlrpc_no_resend(req)) {
1435 DEBUG_REQ(D_ERROR, req,
1436 "reply buffer overflow, expected=%d, actual size=%d",
1437 req->rq_nob_received, req->rq_repbuf_len);
1441 sptlrpc_cli_free_repbuf(req);
1443 * Pass the required reply buffer size (include
1444 * space for early reply).
1445 * NB: no need to roundup because alloc_repbuf
1448 req->rq_replen = req->rq_nob_received;
1449 req->rq_nob_received = 0;
1450 spin_lock(&req->rq_lock);
1452 spin_unlock(&req->rq_lock);
1456 work_start = ktime_get_real();
1457 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1460 * NB Until this point, the whole of the incoming message,
1461 * including buflens, status etc is in the sender's byte order.
1463 rc = sptlrpc_cli_unwrap_reply(req);
1465 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1470 * Security layer unwrap might ask resend this request.
1475 rc = unpack_reply(req);
1479 /* retry indefinitely on EINPROGRESS */
1480 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1481 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1482 time64_t now = ktime_get_real_seconds();
1484 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1485 D_RPCTRACE, req, "resending request on EINPROGRESS");
1486 spin_lock(&req->rq_lock);
1488 spin_unlock(&req->rq_lock);
1489 req->rq_nr_resend++;
1491 /* Readjust the timeout for current conditions */
1492 ptlrpc_at_set_req_timeout(req);
1494 * delay resend to give a chance to the server to get ready.
1495 * The delay is increased by 1s on every resend and is capped to
1496 * the current request timeout (i.e. obd_timeout if AT is off,
1497 * or AT service time x 125% + 5s, see at_est2timeout)
1499 if (req->rq_nr_resend > req->rq_timeout)
1500 req->rq_sent = now + req->rq_timeout;
1502 req->rq_sent = now + req->rq_nr_resend;
1504 /* Resend for EINPROGRESS will use a new XID */
1505 spin_lock(&imp->imp_lock);
1506 list_del_init(&req->rq_unreplied_list);
1507 spin_unlock(&imp->imp_lock);
1512 if (obd->obd_svc_stats) {
1513 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1515 ptlrpc_lprocfs_rpc_sent(req, timediff);
1518 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1519 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1520 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1521 lustre_msg_get_type(req->rq_repmsg));
1525 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1526 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1527 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1528 ptlrpc_at_adj_net_latency(req,
1529 lustre_msg_get_service_timeout(req->rq_repmsg));
1531 rc = ptlrpc_check_status(req);
1535 * Either we've been evicted, or the server has failed for
1536 * some reason. Try to reconnect, and if that fails, punt to
1539 if (ptlrpc_recoverable_error(rc)) {
1540 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1541 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1544 ptlrpc_request_handle_notconn(req);
1549 * Let's look if server sent slv. Do it only for RPC with
1552 ldlm_cli_update_pool(req);
1556 * Store transno in reqmsg for replay.
1558 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1559 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1560 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1563 if (imp->imp_replayable) {
1564 spin_lock(&imp->imp_lock);
1566 * No point in adding already-committed requests to the replay
1567 * list, we will just remove them immediately. b=9829
1569 if (req->rq_transno != 0 &&
1571 lustre_msg_get_last_committed(req->rq_repmsg) ||
1573 /** version recovery */
1574 ptlrpc_save_versions(req);
1575 ptlrpc_retain_replayable_request(req, imp);
1576 } else if (req->rq_commit_cb &&
1577 list_empty(&req->rq_replay_list)) {
1579 * NB: don't call rq_commit_cb if it's already on
1580 * rq_replay_list, ptlrpc_free_committed() will call
1581 * it later, see LU-3618 for details
1583 spin_unlock(&imp->imp_lock);
1584 req->rq_commit_cb(req);
1585 spin_lock(&imp->imp_lock);
1589 * Replay-enabled imports return commit-status information.
1591 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1592 if (likely(committed > imp->imp_peer_committed_transno))
1593 imp->imp_peer_committed_transno = committed;
1595 ptlrpc_free_committed(imp);
1597 if (!list_empty(&imp->imp_replay_list)) {
1598 struct ptlrpc_request *last;
1600 last = list_entry(imp->imp_replay_list.prev,
1601 struct ptlrpc_request,
1604 * Requests with rq_replay stay on the list even if no
1605 * commit is expected.
1607 if (last->rq_transno > imp->imp_peer_committed_transno)
1608 ptlrpc_pinger_commit_expected(imp);
1611 spin_unlock(&imp->imp_lock);
1618 * Helper function to send request \a req over the network for the first time
1619 * Also adjusts request phase.
1620 * Returns 0 on success or error code.
1622 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1624 struct obd_import *imp = req->rq_import;
1629 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1631 /* do not try to go further if there is not enough memory in enc_pool */
1632 if (req->rq_sent && req->rq_bulk)
1633 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1634 pool_is_at_full_capacity())
1637 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1638 (!req->rq_generation_set ||
1639 req->rq_import_generation == imp->imp_generation))
1642 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1644 spin_lock(&imp->imp_lock);
1646 LASSERT(req->rq_xid != 0);
1647 LASSERT(!list_empty(&req->rq_unreplied_list));
1649 if (!req->rq_generation_set)
1650 req->rq_import_generation = imp->imp_generation;
1652 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1653 spin_lock(&req->rq_lock);
1654 req->rq_waiting = 1;
1655 spin_unlock(&req->rq_lock);
1657 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1658 ptlrpc_import_state_name(req->rq_send_state),
1659 ptlrpc_import_state_name(imp->imp_state));
1660 LASSERT(list_empty(&req->rq_list));
1661 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1662 atomic_inc(&req->rq_import->imp_inflight);
1663 spin_unlock(&imp->imp_lock);
1668 spin_unlock(&imp->imp_lock);
1669 req->rq_status = rc;
1670 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1674 LASSERT(list_empty(&req->rq_list));
1675 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1676 atomic_inc(&req->rq_import->imp_inflight);
1679 * find the known replied XID from the unreplied list, CONNECT
1680 * and DISCONNECT requests are skipped to make the sanity check
1681 * on server side happy. see process_req_last_xid().
1683 * For CONNECT: Because replay requests have lower XID, it'll
1684 * break the sanity check if CONNECT bump the exp_last_xid on
1687 * For DISCONNECT: Since client will abort inflight RPC before
1688 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1689 * than the inflight RPC.
1691 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1692 min_xid = ptlrpc_known_replied_xid(imp);
1693 spin_unlock(&imp->imp_lock);
1695 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1697 lustre_msg_set_status(req->rq_reqmsg, current->pid);
1699 /* If the request to be sent is an LDLM callback, do not try to
1701 * An LDLM callback is sent by a server to a client in order to make
1702 * it release a lock, on a communication channel that uses a reverse
1703 * context. It cannot be refreshed on its own, as it is the 'reverse'
1704 * (server-side) representation of a client context.
1705 * We do not care if the reverse context is expired, and want to send
1706 * the LDLM callback anyway. Once the client receives the AST, it is
1707 * its job to refresh its own context if it has expired, hence
1708 * refreshing the associated reverse context on server side, before
1709 * being able to send the LDLM_CANCEL requested by the server.
1711 if (lustre_msg_get_opc(req->rq_reqmsg) != LDLM_BL_CALLBACK &&
1712 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_CP_CALLBACK &&
1713 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_GL_CALLBACK)
1714 rc = sptlrpc_req_refresh_ctx(req, 0);
1717 req->rq_status = rc;
1720 spin_lock(&req->rq_lock);
1721 req->rq_wait_ctx = 1;
1722 spin_unlock(&req->rq_lock);
1728 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1730 imp->imp_obd->obd_uuid.uuid,
1731 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1732 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1733 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1735 rc = ptl_send_rpc(req, 0);
1736 if (rc == -ENOMEM) {
1737 spin_lock(&imp->imp_lock);
1738 if (!list_empty(&req->rq_list)) {
1739 list_del_init(&req->rq_list);
1740 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1741 wake_up(&req->rq_import->imp_recovery_waitq);
1743 spin_unlock(&imp->imp_lock);
1744 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1748 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1750 spin_lock(&req->rq_lock);
1751 req->rq_net_err = 1;
1752 spin_unlock(&req->rq_lock);
1758 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1763 LASSERT(set->set_producer != NULL);
1765 remaining = atomic_read(&set->set_remaining);
1768 * populate the ->set_requests list with requests until we
1769 * reach the maximum number of RPCs in flight for this set
1771 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1772 rc = set->set_producer(set, set->set_producer_arg);
1773 if (rc == -ENOENT) {
1774 /* no more RPC to produce */
1775 set->set_producer = NULL;
1776 set->set_producer_arg = NULL;
1781 RETURN((atomic_read(&set->set_remaining) - remaining));
1785 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1786 * and no more replies are expected.
1787 * (it is possible to get less replies than requests sent e.g. due to timed out
1788 * requests or requests that we had trouble to send out)
1790 * NOTE: This function contains a potential schedule point (cond_resched()).
1792 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1794 struct ptlrpc_request *req, *next;
1795 LIST_HEAD(comp_reqs);
1796 int force_timer_recalc = 0;
1799 if (atomic_read(&set->set_remaining) == 0)
1802 list_for_each_entry_safe(req, next, &set->set_requests,
1804 struct obd_import *imp = req->rq_import;
1805 int unregistered = 0;
1809 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1810 list_move_tail(&req->rq_set_chain, &comp_reqs);
1815 * This schedule point is mainly for the ptlrpcd caller of this
1816 * function. Most ptlrpc sets are not long-lived and unbounded
1817 * in length, but at the least the set used by the ptlrpcd is.
1818 * Since the processing time is unbounded, we need to insert an
1819 * explicit schedule point to make the thread well-behaved.
1824 * If the caller requires to allow to be interpreted by force
1825 * and it has really been interpreted, then move the request
1826 * to RQ_PHASE_INTERPRET phase in spite of what the current
1829 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1830 req->rq_status = -EINTR;
1831 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1834 * Since it is interpreted and we have to wait for
1835 * the reply to be unlinked, then use sync mode.
1839 GOTO(interpret, req->rq_status);
1842 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1843 force_timer_recalc = 1;
1845 /* delayed send - skip */
1846 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1849 /* delayed resend - skip */
1850 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1851 req->rq_sent > ktime_get_real_seconds())
1854 if (!(req->rq_phase == RQ_PHASE_RPC ||
1855 req->rq_phase == RQ_PHASE_BULK ||
1856 req->rq_phase == RQ_PHASE_INTERPRET ||
1857 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1858 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1859 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1863 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1864 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1865 LASSERT(req->rq_next_phase != req->rq_phase);
1866 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1868 if (req->rq_req_deadline &&
1869 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1870 req->rq_req_deadline = 0;
1871 if (req->rq_reply_deadline &&
1872 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1873 req->rq_reply_deadline = 0;
1874 if (req->rq_bulk_deadline &&
1875 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1876 req->rq_bulk_deadline = 0;
1879 * Skip processing until reply is unlinked. We
1880 * can't return to pool before that and we can't
1881 * call interpret before that. We need to make
1882 * sure that all rdma transfers finished and will
1883 * not corrupt any data.
1885 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1886 ptlrpc_cli_wait_unlink(req))
1888 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1889 ptlrpc_client_bulk_active(req))
1893 * Turn fail_loc off to prevent it from looping
1896 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1897 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1900 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1901 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1906 * Move to next phase if reply was successfully
1909 ptlrpc_rqphase_move(req, req->rq_next_phase);
1912 if (req->rq_phase == RQ_PHASE_INTERPRET)
1913 GOTO(interpret, req->rq_status);
1916 * Note that this also will start async reply unlink.
1918 if (req->rq_net_err && !req->rq_timedout) {
1919 ptlrpc_expire_one_request(req, 1);
1922 * Check if we still need to wait for unlink.
1924 if (ptlrpc_cli_wait_unlink(req) ||
1925 ptlrpc_client_bulk_active(req))
1927 /* If there is no need to resend, fail it now. */
1928 if (req->rq_no_resend) {
1929 if (req->rq_status == 0)
1930 req->rq_status = -EIO;
1931 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1932 GOTO(interpret, req->rq_status);
1939 spin_lock(&req->rq_lock);
1940 req->rq_replied = 0;
1941 spin_unlock(&req->rq_lock);
1942 if (req->rq_status == 0)
1943 req->rq_status = -EIO;
1944 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1945 GOTO(interpret, req->rq_status);
1949 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1950 * so it sets rq_intr regardless of individual rpc
1951 * timeouts. The synchronous IO waiting path sets
1952 * rq_intr irrespective of whether ptlrpcd
1953 * has seen a timeout. Our policy is to only interpret
1954 * interrupted rpcs after they have timed out, so we
1955 * need to enforce that here.
1958 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1959 req->rq_wait_ctx)) {
1960 req->rq_status = -EINTR;
1961 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1962 GOTO(interpret, req->rq_status);
1965 if (req->rq_phase == RQ_PHASE_RPC) {
1966 if (req->rq_timedout || req->rq_resend ||
1967 req->rq_waiting || req->rq_wait_ctx) {
1970 if (!ptlrpc_unregister_reply(req, 1)) {
1971 ptlrpc_unregister_bulk(req, 1);
1975 spin_lock(&imp->imp_lock);
1976 if (ptlrpc_import_delay_req(imp, req,
1979 * put on delay list - only if we wait
1980 * recovery finished - before send
1982 list_move_tail(&req->rq_list,
1983 &imp->imp_delayed_list);
1984 spin_unlock(&imp->imp_lock);
1989 req->rq_status = status;
1990 ptlrpc_rqphase_move(req,
1991 RQ_PHASE_INTERPRET);
1992 spin_unlock(&imp->imp_lock);
1993 GOTO(interpret, req->rq_status);
1995 /* ignore on just initiated connections */
1996 if (ptlrpc_no_resend(req) &&
1997 !req->rq_wait_ctx &&
1998 imp->imp_generation !=
1999 imp->imp_initiated_at) {
2000 req->rq_status = -ENOTCONN;
2001 ptlrpc_rqphase_move(req,
2002 RQ_PHASE_INTERPRET);
2003 spin_unlock(&imp->imp_lock);
2004 GOTO(interpret, req->rq_status);
2007 /* don't resend too fast in case of network
2010 if (ktime_get_real_seconds() < (req->rq_sent + 1)
2011 && req->rq_net_err && req->rq_timedout) {
2013 DEBUG_REQ(D_INFO, req,
2014 "throttle request");
2015 /* Don't try to resend RPC right away
2016 * as it is likely it will fail again
2017 * and ptlrpc_check_set() will be
2018 * called again, keeping this thread
2019 * busy. Instead, wait for the next
2020 * timeout. Flag it as resend to
2021 * ensure we don't wait to long.
2024 spin_unlock(&imp->imp_lock);
2028 list_move_tail(&req->rq_list,
2029 &imp->imp_sending_list);
2031 spin_unlock(&imp->imp_lock);
2033 spin_lock(&req->rq_lock);
2034 req->rq_waiting = 0;
2035 spin_unlock(&req->rq_lock);
2037 if (req->rq_timedout || req->rq_resend) {
2039 * This is re-sending anyways,
2040 * let's mark req as resend.
2042 spin_lock(&req->rq_lock);
2044 spin_unlock(&req->rq_lock);
2047 * rq_wait_ctx is only touched by ptlrpcd,
2048 * so no lock is needed here.
2050 status = sptlrpc_req_refresh_ctx(req, 0);
2053 req->rq_status = status;
2054 spin_lock(&req->rq_lock);
2055 req->rq_wait_ctx = 0;
2056 spin_unlock(&req->rq_lock);
2057 force_timer_recalc = 1;
2059 spin_lock(&req->rq_lock);
2060 req->rq_wait_ctx = 1;
2061 spin_unlock(&req->rq_lock);
2066 spin_lock(&req->rq_lock);
2067 req->rq_wait_ctx = 0;
2068 spin_unlock(&req->rq_lock);
2072 * In any case, the previous bulk should be
2073 * cleaned up to prepare for the new sending
2076 !ptlrpc_unregister_bulk(req, 1))
2079 rc = ptl_send_rpc(req, 0);
2080 if (rc == -ENOMEM) {
2081 spin_lock(&imp->imp_lock);
2082 if (!list_empty(&req->rq_list))
2083 list_del_init(&req->rq_list);
2084 spin_unlock(&imp->imp_lock);
2085 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2089 DEBUG_REQ(D_HA, req,
2090 "send failed: rc = %d", rc);
2091 force_timer_recalc = 1;
2092 spin_lock(&req->rq_lock);
2093 req->rq_net_err = 1;
2094 spin_unlock(&req->rq_lock);
2097 /* need to reset the timeout */
2098 force_timer_recalc = 1;
2101 spin_lock(&req->rq_lock);
2103 if (ptlrpc_client_early(req)) {
2104 ptlrpc_at_recv_early_reply(req);
2105 spin_unlock(&req->rq_lock);
2109 /* Still waiting for a reply? */
2110 if (ptlrpc_client_recv(req)) {
2111 spin_unlock(&req->rq_lock);
2115 /* Did we actually receive a reply? */
2116 if (!ptlrpc_client_replied(req)) {
2117 spin_unlock(&req->rq_lock);
2121 spin_unlock(&req->rq_lock);
2124 * unlink from net because we are going to
2125 * swab in-place of reply buffer
2127 unregistered = ptlrpc_unregister_reply(req, 1);
2131 req->rq_status = after_reply(req);
2136 * If there is no bulk associated with this request,
2137 * then we're done and should let the interpreter
2138 * process the reply. Similarly if the RPC returned
2139 * an error, and therefore the bulk will never arrive.
2141 if (!req->rq_bulk || req->rq_status < 0) {
2142 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2143 GOTO(interpret, req->rq_status);
2146 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2149 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2150 if (ptlrpc_client_bulk_active(req))
2153 if (req->rq_bulk->bd_failure) {
2155 * The RPC reply arrived OK, but the bulk screwed
2156 * up! Dead weird since the server told us the RPC
2157 * was good after getting the REPLY for her GET or
2158 * the ACK for her PUT.
2160 DEBUG_REQ(D_ERROR, req, "bulk transfer failed %d/%d/%d",
2162 req->rq_bulk->bd_nob,
2163 req->rq_bulk->bd_nob_transferred);
2164 req->rq_status = -EIO;
2167 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2170 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2173 * This moves to "unregistering" phase we need to wait for
2176 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2177 /* start async bulk unlink too */
2178 ptlrpc_unregister_bulk(req, 1);
2182 if (!ptlrpc_unregister_bulk(req, async))
2186 * When calling interpret receiving already should be
2189 LASSERT(!req->rq_receiving_reply);
2191 ptlrpc_req_interpret(env, req, req->rq_status);
2193 if (ptlrpcd_check_work(req)) {
2194 atomic_dec(&set->set_remaining);
2197 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2201 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2203 imp->imp_obd->obd_uuid.uuid,
2204 lustre_msg_get_status(req->rq_reqmsg),
2206 obd_import_nid2str(imp),
2207 lustre_msg_get_opc(req->rq_reqmsg),
2208 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2210 spin_lock(&imp->imp_lock);
2212 * Request already may be not on sending or delaying list. This
2213 * may happen in the case of marking it erroneous for the case
2214 * ptlrpc_import_delay_req(req, status) find it impossible to
2215 * allow sending this rpc and returns *status != 0.
2217 if (!list_empty(&req->rq_list)) {
2218 list_del_init(&req->rq_list);
2219 if (atomic_dec_and_test(&imp->imp_inflight))
2220 wake_up(&imp->imp_recovery_waitq);
2222 list_del_init(&req->rq_unreplied_list);
2223 spin_unlock(&imp->imp_lock);
2225 atomic_dec(&set->set_remaining);
2226 wake_up(&imp->imp_recovery_waitq);
2228 if (set->set_producer) {
2229 /* produce a new request if possible */
2230 if (ptlrpc_set_producer(set) > 0)
2231 force_timer_recalc = 1;
2234 * free the request that has just been completed
2235 * in order not to pollute set->set_requests
2237 list_del_init(&req->rq_set_chain);
2238 spin_lock(&req->rq_lock);
2240 req->rq_invalid_rqset = 0;
2241 spin_unlock(&req->rq_lock);
2243 /* record rq_status to compute the final status later */
2244 if (req->rq_status != 0)
2245 set->set_rc = req->rq_status;
2246 ptlrpc_req_finished(req);
2248 list_move_tail(&req->rq_set_chain, &comp_reqs);
2253 * move completed request at the head of list so it's easier for
2254 * caller to find them
2256 list_splice(&comp_reqs, &set->set_requests);
2258 /* If we hit an error, we want to recover promptly. */
2259 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2261 EXPORT_SYMBOL(ptlrpc_check_set);
2264 * Time out request \a req. is \a async_unlink is set, that means do not wait
2265 * until LNet actually confirms network buffer unlinking.
2266 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2268 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2270 struct obd_import *imp = req->rq_import;
2271 unsigned int debug_mask = D_RPCTRACE;
2275 spin_lock(&req->rq_lock);
2276 req->rq_timedout = 1;
2277 spin_unlock(&req->rq_lock);
2279 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2280 lustre_msg_get_status(req->rq_reqmsg)))
2281 debug_mask = D_WARNING;
2282 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2283 req->rq_net_err ? "failed due to network error" :
2284 ((req->rq_real_sent == 0 ||
2285 req->rq_real_sent < req->rq_sent ||
2286 req->rq_real_sent >= req->rq_deadline) ?
2287 "timed out for sent delay" : "timed out for slow reply"),
2288 req->rq_sent, req->rq_real_sent);
2290 if (imp && obd_debug_peer_on_timeout)
2291 LNetDebugPeer(imp->imp_connection->c_peer);
2293 ptlrpc_unregister_reply(req, async_unlink);
2294 ptlrpc_unregister_bulk(req, async_unlink);
2296 if (obd_dump_on_timeout)
2297 libcfs_debug_dumplog();
2300 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2304 atomic_inc(&imp->imp_timeouts);
2306 /* The DLM server doesn't want recovery run on its imports. */
2307 if (imp->imp_dlm_fake)
2311 * If this request is for recovery or other primordial tasks,
2312 * then error it out here.
2314 if (req->rq_ctx_init || req->rq_ctx_fini ||
2315 req->rq_send_state != LUSTRE_IMP_FULL ||
2316 imp->imp_obd->obd_no_recov) {
2317 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2318 ptlrpc_import_state_name(req->rq_send_state),
2319 ptlrpc_import_state_name(imp->imp_state));
2320 spin_lock(&req->rq_lock);
2321 req->rq_status = -ETIMEDOUT;
2323 spin_unlock(&req->rq_lock);
2328 * if a request can't be resent we can't wait for an answer after
2331 if (ptlrpc_no_resend(req)) {
2332 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2336 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2342 * Time out all uncompleted requests in request set pointed by \a data
2343 * This is called when a wait times out.
2345 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2347 struct ptlrpc_request *req;
2348 time64_t now = ktime_get_real_seconds();
2351 LASSERT(set != NULL);
2354 * A timeout expired. See which reqs it applies to...
2356 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2357 /* don't expire request waiting for context */
2358 if (req->rq_wait_ctx)
2361 /* Request in-flight? */
2362 if (!((req->rq_phase == RQ_PHASE_RPC &&
2363 !req->rq_waiting && !req->rq_resend) ||
2364 (req->rq_phase == RQ_PHASE_BULK)))
2367 if (req->rq_timedout || /* already dealt with */
2368 req->rq_deadline > now) /* not expired */
2372 * Deal with this guy. Do it asynchronously to not block
2375 ptlrpc_expire_one_request(req, 1);
2377 * Loops require that we resched once in a while to avoid
2378 * RCU stalls and a few other problems.
2386 * Interrupts (sets interrupted flag) all uncompleted requests in
2387 * a set \a data. This is called when a wait_event is interrupted
2390 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2392 struct ptlrpc_request *req;
2394 LASSERT(set != NULL);
2395 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2397 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2401 if (req->rq_phase != RQ_PHASE_RPC &&
2402 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2403 !req->rq_allow_intr)
2406 spin_lock(&req->rq_lock);
2408 spin_unlock(&req->rq_lock);
2413 * Get the smallest timeout in the set; this does NOT set a timeout.
2415 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2417 time64_t now = ktime_get_real_seconds();
2419 struct ptlrpc_request *req;
2423 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2424 /* Request in-flight? */
2425 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2426 (req->rq_phase == RQ_PHASE_BULK) ||
2427 (req->rq_phase == RQ_PHASE_NEW)))
2430 /* Already timed out. */
2431 if (req->rq_timedout)
2434 /* Waiting for ctx. */
2435 if (req->rq_wait_ctx)
2438 if (req->rq_phase == RQ_PHASE_NEW)
2439 deadline = req->rq_sent;
2440 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2441 deadline = req->rq_sent;
2443 deadline = req->rq_sent + req->rq_timeout;
2445 if (deadline <= now) /* actually expired already */
2446 timeout = 1; /* ASAP */
2447 else if (timeout == 0 || timeout > deadline - now)
2448 timeout = deadline - now;
2454 * Send all unset request from the set and then wait untill all
2455 * requests in the set complete (either get a reply, timeout, get an
2456 * error or otherwise be interrupted).
2457 * Returns 0 on success or error code otherwise.
2459 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2461 struct ptlrpc_request *req;
2466 if (set->set_producer)
2467 (void)ptlrpc_set_producer(set);
2469 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2470 if (req->rq_phase == RQ_PHASE_NEW)
2471 (void)ptlrpc_send_new_req(req);
2474 if (list_empty(&set->set_requests))
2478 timeout = ptlrpc_set_next_timeout(set);
2481 * wait until all complete, interrupted, or an in-flight
2484 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2487 if ((timeout == 0 && !signal_pending(current)) ||
2488 set->set_allow_intr) {
2490 * No requests are in-flight (ether timed out
2491 * or delayed), so we can allow interrupts.
2492 * We still want to block for a limited time,
2493 * so we allow interrupts during the timeout.
2495 rc = l_wait_event_abortable_timeout(
2497 ptlrpc_check_set(NULL, set),
2498 cfs_time_seconds(timeout ? timeout : 1));
2501 ptlrpc_expired_set(set);
2502 } else if (rc < 0) {
2504 ptlrpc_interrupted_set(set);
2510 * At least one request is in flight, so no
2511 * interrupts are allowed. Wait until all
2512 * complete, or an in-flight req times out.
2514 rc = wait_event_idle_timeout(
2516 ptlrpc_check_set(NULL, set),
2517 cfs_time_seconds(timeout ? timeout : 1));
2519 ptlrpc_expired_set(set);
2526 * LU-769 - if we ignored the signal because
2527 * it was already pending when we started, we
2528 * need to handle it now or we risk it being
2531 if (rc == -ETIMEDOUT &&
2532 signal_pending(current)) {
2535 siginitset(&new, LUSTRE_FATAL_SIGS);
2536 sigprocmask(SIG_BLOCK, &new, &old);
2538 * In fact we only interrupt for the
2539 * "fatal" signals like SIGINT or
2540 * SIGKILL. We still ignore less
2541 * important signals since ptlrpc set
2542 * is not easily reentrant from
2545 if (signal_pending(current))
2546 ptlrpc_interrupted_set(set);
2547 sigprocmask(SIG_SETMASK, &old, NULL);
2551 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2554 * -EINTR => all requests have been flagged rq_intr so next
2556 * -ETIMEDOUT => someone timed out. When all reqs have
2557 * timed out, signals are enabled allowing completion with
2559 * I don't really care if we go once more round the loop in
2560 * the error cases -eeb.
2562 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2563 list_for_each_entry(req, &set->set_requests,
2565 spin_lock(&req->rq_lock);
2566 req->rq_invalid_rqset = 1;
2567 spin_unlock(&req->rq_lock);
2570 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2572 LASSERT(atomic_read(&set->set_remaining) == 0);
2574 rc = set->set_rc; /* rq_status of already freed requests if any */
2575 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2576 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2577 if (req->rq_status != 0)
2578 rc = req->rq_status;
2583 EXPORT_SYMBOL(ptlrpc_set_wait);
2586 * Helper fuction for request freeing.
2587 * Called when request count reached zero and request needs to be freed.
2588 * Removes request from all sorts of sending/replay lists it might be on,
2589 * frees network buffers if any are present.
2590 * If \a locked is set, that means caller is already holding import imp_lock
2591 * and so we no longer need to reobtain it (for certain lists manipulations)
2593 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2600 LASSERT(!request->rq_srv_req);
2601 LASSERT(request->rq_export == NULL);
2602 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2603 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2604 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2605 LASSERTF(!request->rq_replay, "req %p\n", request);
2607 req_capsule_fini(&request->rq_pill);
2610 * We must take it off the imp_replay_list first. Otherwise, we'll set
2611 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2613 if (request->rq_import) {
2615 spin_lock(&request->rq_import->imp_lock);
2616 list_del_init(&request->rq_replay_list);
2617 list_del_init(&request->rq_unreplied_list);
2619 spin_unlock(&request->rq_import->imp_lock);
2621 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2623 if (atomic_read(&request->rq_refcount) != 0) {
2624 DEBUG_REQ(D_ERROR, request,
2625 "freeing request with nonzero refcount");
2629 if (request->rq_repbuf)
2630 sptlrpc_cli_free_repbuf(request);
2632 if (request->rq_import) {
2633 if (!ptlrpcd_check_work(request)) {
2634 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2635 atomic_dec(&request->rq_import->imp_reqs);
2637 class_import_put(request->rq_import);
2638 request->rq_import = NULL;
2640 if (request->rq_bulk)
2641 ptlrpc_free_bulk(request->rq_bulk);
2643 if (request->rq_reqbuf || request->rq_clrbuf)
2644 sptlrpc_cli_free_reqbuf(request);
2646 if (request->rq_cli_ctx)
2647 sptlrpc_req_put_ctx(request, !locked);
2649 if (request->rq_pool)
2650 __ptlrpc_free_req_to_pool(request);
2652 ptlrpc_request_cache_free(request);
2656 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2658 * Drop one request reference. Must be called with import imp_lock held.
2659 * When reference count drops to zero, request is freed.
2661 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2663 assert_spin_locked(&request->rq_import->imp_lock);
2664 (void)__ptlrpc_req_finished(request, 1);
2669 * Drops one reference count for request \a request.
2670 * \a locked set indicates that caller holds import imp_lock.
2671 * Frees the request whe reference count reaches zero.
2673 * \retval 1 the request is freed
2674 * \retval 0 some others still hold references on the request
2676 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2684 LASSERT(request != LP_POISON);
2685 LASSERT(request->rq_reqmsg != LP_POISON);
2687 DEBUG_REQ(D_INFO, request, "refcount now %u",
2688 atomic_read(&request->rq_refcount) - 1);
2690 spin_lock(&request->rq_lock);
2691 count = atomic_dec_return(&request->rq_refcount);
2692 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2695 * For open RPC, the client does not know the EA size (LOV, ACL, and
2696 * so on) before replied, then the client has to reserve very large
2697 * reply buffer. Such buffer will not be released until the RPC freed.
2698 * Since The open RPC is replayable, we need to keep it in the replay
2699 * list until close. If there are a lot of files opened concurrently,
2700 * then the client may be OOM.
2702 * If fact, it is unnecessary to keep reply buffer for open replay,
2703 * related EAs have already been saved via mdc_save_lovea() before
2704 * coming here. So it is safe to free the reply buffer some earlier
2705 * before releasing the RPC to avoid client OOM. LU-9514
2707 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2708 spin_lock(&request->rq_early_free_lock);
2709 sptlrpc_cli_free_repbuf(request);
2710 request->rq_repbuf = NULL;
2711 request->rq_repbuf_len = 0;
2712 request->rq_repdata = NULL;
2713 request->rq_reqdata_len = 0;
2714 spin_unlock(&request->rq_early_free_lock);
2716 spin_unlock(&request->rq_lock);
2719 __ptlrpc_free_req(request, locked);
2725 * Drops one reference count for a request.
2727 void ptlrpc_req_finished(struct ptlrpc_request *request)
2729 __ptlrpc_req_finished(request, 0);
2731 EXPORT_SYMBOL(ptlrpc_req_finished);
2734 * Returns xid of a \a request
2736 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2738 return request->rq_xid;
2740 EXPORT_SYMBOL(ptlrpc_req_xid);
2743 * Disengage the client's reply buffer from the network
2744 * NB does _NOT_ unregister any client-side bulk.
2745 * IDEMPOTENT, but _not_ safe against concurrent callers.
2746 * The request owner (i.e. the thread doing the I/O) must call...
2747 * Returns 0 on success or 1 if unregistering cannot be made.
2749 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2751 bool discard = false;
2755 LASSERT(!in_interrupt());
2757 /* Let's setup deadline for reply unlink. */
2758 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2759 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2760 request->rq_reply_deadline = ktime_get_real_seconds() +
2761 PTLRPC_REQ_LONG_UNLINK;
2764 * Nothing left to do.
2766 if (!__ptlrpc_cli_wait_unlink(request, &discard))
2769 LNetMDUnlink(request->rq_reply_md_h);
2771 if (discard) /* Discard the request-out callback */
2772 __LNetMDUnlink(request->rq_req_md_h, discard);
2775 * Let's check it once again.
2777 if (!ptlrpc_cli_wait_unlink(request))
2780 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2781 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2784 * Do not wait for unlink to finish.
2790 * We have to wait_event_idle_timeout() whatever the result, to get
2791 * a chance to run reply_in_callback(), and to make sure we've
2792 * unlinked before returning a req to the pool.
2795 wait_queue_head_t *wq = (request->rq_set) ?
2796 &request->rq_set->set_waitq :
2797 &request->rq_reply_waitq;
2798 int seconds = PTLRPC_REQ_LONG_UNLINK;
2800 * Network access will complete in finite time but the HUGE
2801 * timeout lets us CWARN for visibility of sluggish NALs
2803 while (seconds > 0 &&
2804 wait_event_idle_timeout(
2806 !ptlrpc_cli_wait_unlink(request),
2807 cfs_time_seconds(1)) == 0)
2810 ptlrpc_rqphase_move(request, request->rq_next_phase);
2814 DEBUG_REQ(D_WARNING, request,
2815 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2816 request->rq_receiving_reply,
2817 request->rq_req_unlinked,
2818 request->rq_reply_unlinked);
2823 static void ptlrpc_free_request(struct ptlrpc_request *req)
2825 spin_lock(&req->rq_lock);
2827 spin_unlock(&req->rq_lock);
2829 if (req->rq_commit_cb)
2830 req->rq_commit_cb(req);
2831 list_del_init(&req->rq_replay_list);
2833 __ptlrpc_req_finished(req, 1);
2837 * the request is committed and dropped from the replay list of its import
2839 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2841 struct obd_import *imp = req->rq_import;
2843 spin_lock(&imp->imp_lock);
2844 if (list_empty(&req->rq_replay_list)) {
2845 spin_unlock(&imp->imp_lock);
2849 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2850 if (imp->imp_replay_cursor == &req->rq_replay_list)
2851 imp->imp_replay_cursor = req->rq_replay_list.next;
2852 ptlrpc_free_request(req);
2855 spin_unlock(&imp->imp_lock);
2857 EXPORT_SYMBOL(ptlrpc_request_committed);
2860 * Iterates through replay_list on import and prunes
2861 * all requests have transno smaller than last_committed for the
2862 * import and don't have rq_replay set.
2863 * Since requests are sorted in transno order, stops when meetign first
2864 * transno bigger than last_committed.
2865 * caller must hold imp->imp_lock
2867 void ptlrpc_free_committed(struct obd_import *imp)
2869 struct ptlrpc_request *req, *saved;
2870 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2871 bool skip_committed_list = true;
2874 LASSERT(imp != NULL);
2875 assert_spin_locked(&imp->imp_lock);
2877 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2878 imp->imp_generation == imp->imp_last_generation_checked) {
2879 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2880 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2883 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2884 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2885 imp->imp_generation);
2887 if (imp->imp_generation != imp->imp_last_generation_checked ||
2888 imp->imp_last_transno_checked == 0)
2889 skip_committed_list = false;
2891 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2892 imp->imp_last_generation_checked = imp->imp_generation;
2894 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2896 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2897 LASSERT(req != last_req);
2900 if (req->rq_transno == 0) {
2901 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2904 if (req->rq_import_generation < imp->imp_generation) {
2905 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2909 /* not yet committed */
2910 if (req->rq_transno > imp->imp_peer_committed_transno) {
2911 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2915 if (req->rq_replay) {
2916 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2917 list_move_tail(&req->rq_replay_list,
2918 &imp->imp_committed_list);
2922 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2923 imp->imp_peer_committed_transno);
2925 ptlrpc_free_request(req);
2928 if (skip_committed_list)
2931 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2933 LASSERT(req->rq_transno != 0);
2934 if (req->rq_import_generation < imp->imp_generation ||
2936 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2937 req->rq_import_generation <
2938 imp->imp_generation ? "stale" : "closed");
2940 if (imp->imp_replay_cursor == &req->rq_replay_list)
2941 imp->imp_replay_cursor =
2942 req->rq_replay_list.next;
2944 ptlrpc_free_request(req);
2951 void ptlrpc_cleanup_client(struct obd_import *imp)
2958 * Schedule previously sent request for resend.
2959 * For bulk requests we assign new xid (to avoid problems with
2960 * lost replies and therefore several transfers landing into same buffer
2961 * from different sending attempts).
2963 void ptlrpc_resend_req(struct ptlrpc_request *req)
2965 DEBUG_REQ(D_HA, req, "going to resend");
2966 spin_lock(&req->rq_lock);
2969 * Request got reply but linked to the import list still.
2970 * Let ptlrpc_check_set() process it.
2972 if (ptlrpc_client_replied(req)) {
2973 spin_unlock(&req->rq_lock);
2974 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2978 req->rq_status = -EAGAIN;
2981 req->rq_net_err = 0;
2982 req->rq_timedout = 0;
2984 ptlrpc_client_wake_req(req);
2985 spin_unlock(&req->rq_lock);
2988 /* XXX: this function and rq_status are currently unused */
2989 void ptlrpc_restart_req(struct ptlrpc_request *req)
2991 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2992 req->rq_status = -ERESTARTSYS;
2994 spin_lock(&req->rq_lock);
2995 req->rq_restart = 1;
2996 req->rq_timedout = 0;
2997 ptlrpc_client_wake_req(req);
2998 spin_unlock(&req->rq_lock);
3002 * Grab additional reference on a request \a req
3004 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
3007 atomic_inc(&req->rq_refcount);
3010 EXPORT_SYMBOL(ptlrpc_request_addref);
3013 * Add a request to import replay_list.
3014 * Must be called under imp_lock
3016 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3017 struct obd_import *imp)
3019 struct ptlrpc_request *iter;
3021 assert_spin_locked(&imp->imp_lock);
3023 if (req->rq_transno == 0) {
3024 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3029 * clear this for new requests that were resent as well
3030 * as resent replayed requests.
3032 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3034 /* don't re-add requests that have been replayed */
3035 if (!list_empty(&req->rq_replay_list))
3038 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3040 spin_lock(&req->rq_lock);
3042 spin_unlock(&req->rq_lock);
3044 LASSERT(imp->imp_replayable);
3045 /* Balanced in ptlrpc_free_committed, usually. */
3046 ptlrpc_request_addref(req);
3047 list_for_each_entry_reverse(iter, &imp->imp_replay_list,
3050 * We may have duplicate transnos if we create and then
3051 * open a file, or for closes retained if to match creating
3052 * opens, so use req->rq_xid as a secondary key.
3053 * (See bugs 684, 685, and 428.)
3054 * XXX no longer needed, but all opens need transnos!
3056 if (iter->rq_transno > req->rq_transno)
3059 if (iter->rq_transno == req->rq_transno) {
3060 LASSERT(iter->rq_xid != req->rq_xid);
3061 if (iter->rq_xid > req->rq_xid)
3065 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3069 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3073 * Send request and wait until it completes.
3074 * Returns request processing status.
3076 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3078 struct ptlrpc_request_set *set;
3082 LASSERT(req->rq_set == NULL);
3083 LASSERT(!req->rq_receiving_reply);
3085 set = ptlrpc_prep_set();
3087 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3091 /* for distributed debugging */
3092 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3094 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3095 ptlrpc_request_addref(req);
3096 ptlrpc_set_add_req(set, req);
3097 rc = ptlrpc_set_wait(NULL, set);
3098 ptlrpc_set_destroy(set);
3102 EXPORT_SYMBOL(ptlrpc_queue_wait);
3105 * Callback used for replayed requests reply processing.
3106 * In case of successful reply calls registered request replay callback.
3107 * In case of error restart replay process.
3109 static int ptlrpc_replay_interpret(const struct lu_env *env,
3110 struct ptlrpc_request *req,
3113 struct ptlrpc_replay_async_args *aa = args;
3114 struct obd_import *imp = req->rq_import;
3117 atomic_dec(&imp->imp_replay_inflight);
3120 * Note: if it is bulk replay (MDS-MDS replay), then even if
3121 * server got the request, but bulk transfer timeout, let's
3122 * replay the bulk req again
3124 if (!ptlrpc_client_replied(req) ||
3126 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3127 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3128 GOTO(out, rc = -ETIMEDOUT);
3131 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3132 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3133 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3134 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3136 /** VBR: check version failure */
3137 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3138 /** replay was failed due to version mismatch */
3139 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3140 spin_lock(&imp->imp_lock);
3141 imp->imp_vbr_failed = 1;
3142 spin_unlock(&imp->imp_lock);
3143 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3145 /** The transno had better not change over replay. */
3146 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3147 lustre_msg_get_transno(req->rq_repmsg) ||
3148 lustre_msg_get_transno(req->rq_repmsg) == 0,
3150 lustre_msg_get_transno(req->rq_reqmsg),
3151 lustre_msg_get_transno(req->rq_repmsg));
3154 spin_lock(&imp->imp_lock);
3155 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3156 spin_unlock(&imp->imp_lock);
3157 LASSERT(imp->imp_last_replay_transno);
3159 /* transaction number shouldn't be bigger than the latest replayed */
3160 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3161 DEBUG_REQ(D_ERROR, req,
3162 "Reported transno=%llu is bigger than replayed=%llu",
3164 lustre_msg_get_transno(req->rq_reqmsg));
3165 GOTO(out, rc = -EINVAL);
3168 DEBUG_REQ(D_HA, req, "got reply");
3170 /* let the callback do fixups, possibly including in the request */
3171 if (req->rq_replay_cb)
3172 req->rq_replay_cb(req);
3174 if (ptlrpc_client_replied(req) &&
3175 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3176 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3177 lustre_msg_get_status(req->rq_repmsg),
3178 aa->praa_old_status);
3181 * Note: If the replay fails for MDT-MDT recovery, let's
3182 * abort all of the following requests in the replay
3183 * and sending list, because MDT-MDT update requests
3184 * are dependent on each other, see LU-7039
3186 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3187 struct ptlrpc_request *free_req;
3188 struct ptlrpc_request *tmp;
3190 spin_lock(&imp->imp_lock);
3191 list_for_each_entry_safe(free_req, tmp,
3192 &imp->imp_replay_list,
3194 ptlrpc_free_request(free_req);
3197 list_for_each_entry_safe(free_req, tmp,
3198 &imp->imp_committed_list,
3200 ptlrpc_free_request(free_req);
3203 list_for_each_entry_safe(free_req, tmp,
3204 &imp->imp_delayed_list,
3206 spin_lock(&free_req->rq_lock);
3207 free_req->rq_err = 1;
3208 free_req->rq_status = -EIO;
3209 ptlrpc_client_wake_req(free_req);
3210 spin_unlock(&free_req->rq_lock);
3213 list_for_each_entry_safe(free_req, tmp,
3214 &imp->imp_sending_list,
3216 spin_lock(&free_req->rq_lock);
3217 free_req->rq_err = 1;
3218 free_req->rq_status = -EIO;
3219 ptlrpc_client_wake_req(free_req);
3220 spin_unlock(&free_req->rq_lock);
3222 spin_unlock(&imp->imp_lock);
3225 /* Put it back for re-replay. */
3226 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3230 * Errors while replay can set transno to 0, but
3231 * imp_last_replay_transno shouldn't be set to 0 anyway
3233 if (req->rq_transno == 0)
3234 CERROR("Transno is 0 during replay!\n");
3236 /* continue with recovery */
3237 rc = ptlrpc_import_recovery_state_machine(imp);
3239 req->rq_send_state = aa->praa_old_state;
3242 /* this replay failed, so restart recovery */
3243 ptlrpc_connect_import(imp);
3249 * Prepares and queues request for replay.
3250 * Adds it to ptlrpcd queue for actual sending.
3251 * Returns 0 on success.
3253 int ptlrpc_replay_req(struct ptlrpc_request *req)
3255 struct ptlrpc_replay_async_args *aa;
3259 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3261 aa = ptlrpc_req_async_args(aa, req);
3262 memset(aa, 0, sizeof(*aa));
3264 /* Prepare request to be resent with ptlrpcd */
3265 aa->praa_old_state = req->rq_send_state;
3266 req->rq_send_state = LUSTRE_IMP_REPLAY;
3267 req->rq_phase = RQ_PHASE_NEW;
3268 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3270 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3272 req->rq_interpret_reply = ptlrpc_replay_interpret;
3273 /* Readjust the timeout for current conditions */
3274 ptlrpc_at_set_req_timeout(req);
3276 /* Tell server net_latency to calculate how long to wait for reply. */
3277 lustre_msg_set_service_timeout(req->rq_reqmsg,
3278 ptlrpc_at_get_net_latency(req));
3279 DEBUG_REQ(D_HA, req, "REPLAY");
3281 atomic_inc(&req->rq_import->imp_replay_inflight);
3282 spin_lock(&req->rq_lock);
3283 req->rq_early_free_repbuf = 0;
3284 spin_unlock(&req->rq_lock);
3285 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3287 ptlrpcd_add_req(req);
3292 * Aborts all in-flight request on import \a imp sending and delayed lists
3294 void ptlrpc_abort_inflight(struct obd_import *imp)
3296 struct ptlrpc_request *req;
3300 * Make sure that no new requests get processed for this import.
3301 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3302 * this flag and then putting requests on sending_list or delayed_list.
3304 assert_spin_locked(&imp->imp_lock);
3307 * XXX locking? Maybe we should remove each request with the list
3308 * locked? Also, how do we know if the requests on the list are
3309 * being freed at this time?
3311 list_for_each_entry(req, &imp->imp_sending_list, rq_list) {
3312 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3314 spin_lock(&req->rq_lock);
3315 if (req->rq_import_generation < imp->imp_generation) {
3317 req->rq_status = -EIO;
3318 ptlrpc_client_wake_req(req);
3320 spin_unlock(&req->rq_lock);
3323 list_for_each_entry(req, &imp->imp_delayed_list, rq_list) {
3324 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3326 spin_lock(&req->rq_lock);
3327 if (req->rq_import_generation < imp->imp_generation) {
3329 req->rq_status = -EIO;
3330 ptlrpc_client_wake_req(req);
3332 spin_unlock(&req->rq_lock);
3336 * Last chance to free reqs left on the replay list, but we
3337 * will still leak reqs that haven't committed.
3339 if (imp->imp_replayable)
3340 ptlrpc_free_committed(imp);
3346 * Abort all uncompleted requests in request set \a set
3348 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3350 struct ptlrpc_request *req;
3352 LASSERT(set != NULL);
3354 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
3355 spin_lock(&req->rq_lock);
3356 if (req->rq_phase != RQ_PHASE_RPC) {
3357 spin_unlock(&req->rq_lock);
3362 req->rq_status = -EINTR;
3363 ptlrpc_client_wake_req(req);
3364 spin_unlock(&req->rq_lock);
3369 * Initialize the XID for the node. This is common among all requests on
3370 * this node, and only requires the property that it is monotonically
3371 * increasing. It does not need to be sequential. Since this is also used
3372 * as the RDMA match bits, it is important that a single client NOT have
3373 * the same match bits for two different in-flight requests, hence we do
3374 * NOT want to have an XID per target or similar.
3376 * To avoid an unlikely collision between match bits after a client reboot
3377 * (which would deliver old data into the wrong RDMA buffer) initialize
3378 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3379 * If the time is clearly incorrect, we instead use a 62-bit random number.
3380 * In the worst case the random number will overflow 1M RPCs per second in
3381 * 9133 years, or permutations thereof.
3383 #define YEAR_2004 (1ULL << 30)
3384 void ptlrpc_init_xid(void)
3386 time64_t now = ktime_get_real_seconds();
3389 if (now < YEAR_2004) {
3390 get_random_bytes(&xid, sizeof(xid));
3392 xid |= (1ULL << 61);
3394 xid = (u64)now << 20;
3397 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3398 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3400 xid &= PTLRPC_BULK_OPS_MASK;
3401 atomic64_set(&ptlrpc_last_xid, xid);
3405 * Increase xid and returns resulting new value to the caller.
3407 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3408 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3409 * itself uses the last bulk xid needed, so the server can determine the
3410 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3411 * xid must align to a power-of-two value.
3413 * This is assumed to be true due to the initial ptlrpc_last_xid
3414 * value also being initialized to a power-of-two value. LU-1431
3416 __u64 ptlrpc_next_xid(void)
3418 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3422 * If request has a new allocated XID (new request or EINPROGRESS resend),
3423 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3424 * request to ensure previous bulk fails and avoid problems with lost replies
3425 * and therefore several transfers landing into the same buffer from different
3427 * Also, to avoid previous reply landing to a different sending attempt.
3429 void ptlrpc_set_mbits(struct ptlrpc_request *req)
3431 int md_count = req->rq_bulk ? req->rq_bulk->bd_md_count : 1;
3434 * Generate new matchbits for all resend requests, including
3437 if (req->rq_resend) {
3438 __u64 old_mbits = req->rq_mbits;
3441 * First time resend on -EINPROGRESS will generate new xid,
3442 * so we can actually use the rq_xid as rq_mbits in such case,
3443 * however, it's bit hard to distinguish such resend with a
3444 * 'resend for the -EINPROGRESS resend'. To make it simple,
3445 * we opt to generate mbits for all resend cases.
3447 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3449 req->rq_mbits = ptlrpc_next_xid();
3452 * Old version transfers rq_xid to peer as
3455 spin_lock(&req->rq_import->imp_lock);
3456 list_del_init(&req->rq_unreplied_list);
3457 ptlrpc_assign_next_xid_nolock(req);
3458 spin_unlock(&req->rq_import->imp_lock);
3459 req->rq_mbits = req->rq_xid;
3461 CDEBUG(D_HA, "resend with new mbits old x%llu new x%llu\n",
3462 old_mbits, req->rq_mbits);
3463 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3464 /* Request being sent first time, use xid as matchbits. */
3465 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3466 BULK_MBITS) || req->rq_mbits == 0)
3468 req->rq_mbits = req->rq_xid;
3470 req->rq_mbits -= md_count - 1;
3474 * Replay request, xid and matchbits have already been
3475 * correctly assigned.
3481 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3482 * that server can infer the number of bulks that were prepared,
3485 req->rq_mbits += md_count - 1;
3488 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3489 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3491 * It's ok to directly set the rq_xid here, since this xid bump
3492 * won't affect the request position in unreplied list.
3494 if (!OCD_HAS_FLAG(&req->rq_import->imp_connect_data, BULK_MBITS))
3495 req->rq_xid = req->rq_mbits;
3499 * Get a glimpse at what next xid value might have been.
3500 * Returns possible next xid.
3502 __u64 ptlrpc_sample_next_xid(void)
3504 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3506 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3509 * Functions for operating ptlrpc workers.
3511 * A ptlrpc work is a function which will be running inside ptlrpc context.
3512 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3514 * 1. after a work is created, it can be used many times, that is:
3515 * handler = ptlrpcd_alloc_work();
3516 * ptlrpcd_queue_work();
3518 * queue it again when necessary:
3519 * ptlrpcd_queue_work();
3520 * ptlrpcd_destroy_work();
3521 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3522 * it will only be queued once in any time. Also as its name implies, it may
3523 * have delay before it really runs by ptlrpcd thread.
3525 struct ptlrpc_work_async_args {
3526 int (*cb)(const struct lu_env *, void *);
3530 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3532 /* re-initialize the req */
3533 req->rq_timeout = obd_timeout;
3534 req->rq_sent = ktime_get_real_seconds();
3535 req->rq_deadline = req->rq_sent + req->rq_timeout;
3536 req->rq_phase = RQ_PHASE_INTERPRET;
3537 req->rq_next_phase = RQ_PHASE_COMPLETE;
3538 req->rq_xid = ptlrpc_next_xid();
3539 req->rq_import_generation = req->rq_import->imp_generation;
3541 ptlrpcd_add_req(req);
3544 static int work_interpreter(const struct lu_env *env,
3545 struct ptlrpc_request *req, void *args, int rc)
3547 struct ptlrpc_work_async_args *arg = args;
3549 LASSERT(ptlrpcd_check_work(req));
3550 LASSERT(arg->cb != NULL);
3552 rc = arg->cb(env, arg->cbdata);
3554 list_del_init(&req->rq_set_chain);
3557 if (atomic_dec_return(&req->rq_refcount) > 1) {
3558 atomic_set(&req->rq_refcount, 2);
3559 ptlrpcd_add_work_req(req);
3564 static int worker_format;
3566 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3568 return req->rq_pill.rc_fmt == (void *)&worker_format;
3572 * Create a work for ptlrpc.
3574 void *ptlrpcd_alloc_work(struct obd_import *imp,
3575 int (*cb)(const struct lu_env *, void *), void *cbdata)
3577 struct ptlrpc_request *req = NULL;
3578 struct ptlrpc_work_async_args *args;
3584 RETURN(ERR_PTR(-EINVAL));
3586 /* copy some code from deprecated fakereq. */
3587 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3589 CERROR("ptlrpc: run out of memory!\n");
3590 RETURN(ERR_PTR(-ENOMEM));
3593 ptlrpc_cli_req_init(req);
3595 req->rq_send_state = LUSTRE_IMP_FULL;
3596 req->rq_type = PTL_RPC_MSG_REQUEST;
3597 req->rq_import = class_import_get(imp);
3598 req->rq_interpret_reply = work_interpreter;
3599 /* don't want reply */
3600 req->rq_no_delay = req->rq_no_resend = 1;
3601 req->rq_pill.rc_fmt = (void *)&worker_format;
3603 args = ptlrpc_req_async_args(args, req);
3605 args->cbdata = cbdata;
3609 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3611 void ptlrpcd_destroy_work(void *handler)
3613 struct ptlrpc_request *req = handler;
3616 ptlrpc_req_finished(req);
3618 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3620 int ptlrpcd_queue_work(void *handler)
3622 struct ptlrpc_request *req = handler;
3625 * Check if the req is already being queued.
3627 * Here comes a trick: it lacks a way of checking if a req is being
3628 * processed reliably in ptlrpc. Here I have to use refcount of req
3629 * for this purpose. This is okay because the caller should use this
3630 * req as opaque data. - Jinshan
3632 LASSERT(atomic_read(&req->rq_refcount) > 0);
3633 if (atomic_inc_return(&req->rq_refcount) == 2)
3634 ptlrpcd_add_work_req(req);
3637 EXPORT_SYMBOL(ptlrpcd_queue_work);