4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 /** Implementation of client-side PortalRPC interfaces */
35 #define DEBUG_SUBSYSTEM S_RPC
37 #include <linux/delay.h>
38 #include <linux/random.h>
40 #include <lnet/lib-lnet.h>
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
48 #include "ptlrpc_internal.h"
50 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
51 struct page *page, int pageoffset,
54 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
57 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
58 struct page *page, int pageoffset,
61 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
64 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
68 for (i = 0; i < desc->bd_iov_count ; i++)
69 put_page(desc->bd_vec[i].bv_page);
72 static int ptlrpc_prep_bulk_frag_pages(struct ptlrpc_bulk_desc *desc,
75 unsigned int offset = (unsigned long)frag & ~PAGE_MASK;
79 int page_len = min_t(unsigned int, PAGE_SIZE - offset,
81 unsigned long vaddr = (unsigned long)frag;
83 ptlrpc_prep_bulk_page_nopin(desc,
84 lnet_kvaddr_to_page(vaddr),
94 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
95 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
96 .release_frags = ptlrpc_release_bulk_page_pin,
98 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
100 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
101 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
102 .release_frags = ptlrpc_release_bulk_noop,
103 .add_iov_frag = ptlrpc_prep_bulk_frag_pages,
105 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
107 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
108 static int ptlrpcd_check_work(struct ptlrpc_request *req);
109 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
112 * Initialize passed in client structure \a cl.
114 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
115 struct ptlrpc_client *cl)
117 cl->cli_request_portal = req_portal;
118 cl->cli_reply_portal = rep_portal;
121 EXPORT_SYMBOL(ptlrpc_init_client);
124 * Return PortalRPC connection for remore uud \a uuid
126 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
127 lnet_nid_t nid4refnet)
129 struct ptlrpc_connection *c;
131 struct lnet_process_id peer;
135 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
136 * before accessing its values.
138 /* coverity[uninit_use_in_call] */
139 peer.nid = nid4refnet;
140 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
142 CNETERR("cannot find peer %s!\n", uuid->uuid);
146 c = ptlrpc_connection_get(peer, self, uuid);
148 memcpy(c->c_remote_uuid.uuid,
149 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
152 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
158 * Allocate and initialize new bulk descriptor on the sender.
159 * Returns pointer to the descriptor or NULL on error.
161 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
162 unsigned int max_brw,
163 enum ptlrpc_bulk_op_type type,
165 const struct ptlrpc_bulk_frag_ops *ops)
167 struct ptlrpc_bulk_desc *desc;
170 LASSERT(ops->add_kiov_frag != NULL);
172 if (max_brw > PTLRPC_BULK_OPS_COUNT)
175 if (nfrags > LNET_MAX_IOV * max_brw)
182 OBD_ALLOC_LARGE(desc->bd_vec,
183 nfrags * sizeof(*desc->bd_vec));
187 spin_lock_init(&desc->bd_lock);
188 init_waitqueue_head(&desc->bd_waitq);
189 desc->bd_max_iov = nfrags;
190 desc->bd_iov_count = 0;
191 desc->bd_portal = portal;
192 desc->bd_type = type;
193 desc->bd_md_count = 0;
194 desc->bd_nob_last = LNET_MTU;
195 desc->bd_frag_ops = ops;
196 LASSERT(max_brw > 0);
197 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
199 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
200 * node. Negotiated ocd_brw_size will always be <= this number.
202 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
203 LNetInvalidateMDHandle(&desc->bd_mds[i]);
212 * Prepare bulk descriptor for specified outgoing request \a req that
213 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
214 * the bulk to be sent. Used on client-side.
215 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
218 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
220 unsigned int max_brw,
223 const struct ptlrpc_bulk_frag_ops
226 struct obd_import *imp = req->rq_import;
227 struct ptlrpc_bulk_desc *desc;
230 LASSERT(ptlrpc_is_bulk_op_passive(type));
232 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
236 desc->bd_import = class_import_get(imp);
239 desc->bd_cbid.cbid_fn = client_bulk_callback;
240 desc->bd_cbid.cbid_arg = desc;
242 /* This makes req own desc, and free it when she frees herself */
247 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
249 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
250 struct page *page, int pageoffset, int len,
253 struct bio_vec *kiov;
255 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
256 LASSERT(page != NULL);
257 LASSERT(pageoffset >= 0);
259 LASSERT(pageoffset + len <= PAGE_SIZE);
261 kiov = &desc->bd_vec[desc->bd_iov_count];
263 if (((desc->bd_iov_count % LNET_MAX_IOV) == 0) ||
264 ((desc->bd_nob_last + len) > LNET_MTU)) {
265 desc->bd_mds_off[desc->bd_md_count] = desc->bd_iov_count;
267 desc->bd_nob_last = 0;
268 LASSERT(desc->bd_md_count <= PTLRPC_BULK_OPS_COUNT);
271 desc->bd_nob_last += len;
277 kiov->bv_page = page;
278 kiov->bv_offset = pageoffset;
281 desc->bd_iov_count++;
283 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
285 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
289 LASSERT(desc != NULL);
290 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
291 LASSERT(desc->bd_refs == 0); /* network hands off */
292 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
293 LASSERT(desc->bd_frag_ops != NULL);
295 sptlrpc_enc_pool_put_pages(desc);
298 class_export_put(desc->bd_export);
300 class_import_put(desc->bd_import);
302 if (desc->bd_frag_ops->release_frags != NULL)
303 desc->bd_frag_ops->release_frags(desc);
305 OBD_FREE_LARGE(desc->bd_vec,
306 desc->bd_max_iov * sizeof(*desc->bd_vec));
310 EXPORT_SYMBOL(ptlrpc_free_bulk);
313 * Set server timelimit for this req, i.e. how long are we willing to wait
314 * for reply before timing out this request.
316 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
318 LASSERT(req->rq_import);
321 /* non-AT settings */
323 * \a imp_server_timeout means this is reverse import and
324 * we send (currently only) ASTs to the client and cannot afford
325 * to wait too long for the reply, otherwise the other client
326 * (because of which we are sending this request) would
327 * timeout waiting for us
329 req->rq_timeout = req->rq_import->imp_server_timeout ?
330 obd_timeout / 2 : obd_timeout;
332 struct imp_at *at = &req->rq_import->imp_at;
336 idx = import_at_get_index(req->rq_import,
337 req->rq_request_portal);
338 serv_est = at_get(&at->iat_service_estimate[idx]);
340 * Currently a 32 bit value is sent over the
341 * wire for rq_timeout so please don't change this
342 * to time64_t. The work for LU-1158 will in time
343 * replace rq_timeout with a 64 bit nanosecond value
345 req->rq_timeout = at_est2timeout(serv_est);
348 * We could get even fancier here, using history to predict increased
351 * Let the server know what this RPC timeout is by putting it in the
354 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
356 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
358 /* Adjust max service estimate based on server value */
359 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
366 LASSERT(req->rq_import);
367 at = &req->rq_import->imp_at;
369 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
371 * max service estimates are tracked on the server side,
372 * so just keep minimal history here
374 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
377 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
378 req->rq_import->imp_obd->obd_name,
379 req->rq_request_portal,
380 oldse, at_get(&at->iat_service_estimate[idx]));
383 /* Expected network latency per remote node (secs) */
384 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
386 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
389 /* Adjust expected network latency */
390 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
391 timeout_t service_timeout)
393 time64_t now = ktime_get_real_seconds();
398 LASSERT(req->rq_import);
400 if (service_timeout > now - req->rq_sent + 3) {
402 * b=16408, however, this can also happen if early reply
403 * is lost and client RPC is expired and resent, early reply
404 * or reply of original RPC can still be fit in reply buffer
405 * of resent RPC, now client is measuring time from the
406 * resent time, but server sent back service time of original
409 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
410 D_ADAPTTO : D_WARNING,
411 "Reported service time %u > total measured time %lld\n",
412 service_timeout, now - req->rq_sent);
416 /* Network latency is total time less server processing time,
419 nl = max_t(timeout_t, now - req->rq_sent - service_timeout, 0) + 1;
420 at = &req->rq_import->imp_at;
422 oldnl = at_measured(&at->iat_net_latency, nl);
425 "The network latency for %s (nid %s) has changed from %d to %d\n",
426 req->rq_import->imp_obd->obd_name,
427 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
428 oldnl, at_get(&at->iat_net_latency));
431 static int unpack_reply(struct ptlrpc_request *req)
435 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
436 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
438 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
444 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
446 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
454 * Handle an early reply message, called with the rq_lock held.
455 * If anything goes wrong just ignore it - same as if it never happened
457 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
458 __must_hold(&req->rq_lock)
460 struct ptlrpc_request *early_req;
461 timeout_t service_timeout;
467 spin_unlock(&req->rq_lock);
469 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
471 spin_lock(&req->rq_lock);
475 rc = unpack_reply(early_req);
477 sptlrpc_cli_finish_early_reply(early_req);
478 spin_lock(&req->rq_lock);
483 * Use new timeout value just to adjust the local value for this
484 * request, don't include it into at_history. It is unclear yet why
485 * service time increased and should it be counted or skipped, e.g.
486 * that can be recovery case or some error or server, the real reply
487 * will add all new data if it is worth to add.
489 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
490 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
492 /* Network latency can be adjusted, it is pure network delays */
493 service_timeout = lustre_msg_get_service_timeout(early_req->rq_repmsg);
494 ptlrpc_at_adj_net_latency(req, service_timeout);
496 sptlrpc_cli_finish_early_reply(early_req);
498 spin_lock(&req->rq_lock);
499 olddl = req->rq_deadline;
501 * server assumes it now has rq_timeout from when the request
502 * arrived, so the client should give it at least that long.
503 * since we don't know the arrival time we'll use the original
506 req->rq_deadline = req->rq_sent + req->rq_timeout +
507 ptlrpc_at_get_net_latency(req);
509 /* The below message is checked in replay-single.sh test_65{a,b} */
510 /* The below message is checked in sanity-{gss,krb5} test_8 */
511 DEBUG_REQ(D_ADAPTTO, req,
512 "Early reply #%d, new deadline in %llds (%llds)",
514 req->rq_deadline - ktime_get_real_seconds(),
515 req->rq_deadline - olddl);
520 static struct kmem_cache *request_cache;
522 int ptlrpc_request_cache_init(void)
524 request_cache = kmem_cache_create("ptlrpc_cache",
525 sizeof(struct ptlrpc_request),
526 0, SLAB_HWCACHE_ALIGN, NULL);
527 return request_cache ? 0 : -ENOMEM;
530 void ptlrpc_request_cache_fini(void)
532 kmem_cache_destroy(request_cache);
535 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
537 struct ptlrpc_request *req;
539 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
543 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
545 OBD_SLAB_FREE_PTR(req, request_cache);
549 * Wind down request pool \a pool.
550 * Frees all requests from the pool too
552 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
554 struct list_head *l, *tmp;
555 struct ptlrpc_request *req;
557 LASSERT(pool != NULL);
559 spin_lock(&pool->prp_lock);
560 list_for_each_safe(l, tmp, &pool->prp_req_list) {
561 req = list_entry(l, struct ptlrpc_request, rq_list);
562 list_del(&req->rq_list);
563 LASSERT(req->rq_reqbuf);
564 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
565 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
566 ptlrpc_request_cache_free(req);
568 spin_unlock(&pool->prp_lock);
569 OBD_FREE(pool, sizeof(*pool));
571 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
574 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
576 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
581 while (size < pool->prp_rq_size)
584 LASSERTF(list_empty(&pool->prp_req_list) ||
585 size == pool->prp_rq_size,
586 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
587 pool->prp_rq_size, size);
589 pool->prp_rq_size = size;
590 for (i = 0; i < num_rq; i++) {
591 struct ptlrpc_request *req;
592 struct lustre_msg *msg;
594 req = ptlrpc_request_cache_alloc(GFP_NOFS);
597 OBD_ALLOC_LARGE(msg, size);
599 ptlrpc_request_cache_free(req);
602 req->rq_reqbuf = msg;
603 req->rq_reqbuf_len = size;
605 spin_lock(&pool->prp_lock);
606 list_add_tail(&req->rq_list, &pool->prp_req_list);
607 spin_unlock(&pool->prp_lock);
611 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
614 * Create and initialize new request pool with given attributes:
615 * \a num_rq - initial number of requests to create for the pool
616 * \a msgsize - maximum message size possible for requests in thid pool
617 * \a populate_pool - function to be called when more requests need to be added
619 * Returns pointer to newly created pool or NULL on error.
621 struct ptlrpc_request_pool *
622 ptlrpc_init_rq_pool(int num_rq, int msgsize,
623 int (*populate_pool)(struct ptlrpc_request_pool *, int))
625 struct ptlrpc_request_pool *pool;
632 * Request next power of two for the allocation, because internally
633 * kernel would do exactly this
635 spin_lock_init(&pool->prp_lock);
636 INIT_LIST_HEAD(&pool->prp_req_list);
637 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
638 pool->prp_populate = populate_pool;
640 populate_pool(pool, num_rq);
644 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
647 * Fetches one request from pool \a pool
649 static struct ptlrpc_request *
650 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
652 struct ptlrpc_request *request;
653 struct lustre_msg *reqbuf;
658 spin_lock(&pool->prp_lock);
661 * See if we have anything in a pool, and bail out if nothing,
662 * in writeout path, where this matters, this is safe to do, because
663 * nothing is lost in this case, and when some in-flight requests
664 * complete, this code will be called again.
666 if (unlikely(list_empty(&pool->prp_req_list))) {
667 spin_unlock(&pool->prp_lock);
671 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
673 list_del_init(&request->rq_list);
674 spin_unlock(&pool->prp_lock);
676 LASSERT(request->rq_reqbuf);
677 LASSERT(request->rq_pool);
679 reqbuf = request->rq_reqbuf;
680 memset(request, 0, sizeof(*request));
681 request->rq_reqbuf = reqbuf;
682 request->rq_reqbuf_len = pool->prp_rq_size;
683 request->rq_pool = pool;
689 * Returns freed \a request to pool.
691 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
693 struct ptlrpc_request_pool *pool = request->rq_pool;
695 spin_lock(&pool->prp_lock);
696 LASSERT(list_empty(&request->rq_list));
697 LASSERT(!request->rq_receiving_reply);
698 list_add_tail(&request->rq_list, &pool->prp_req_list);
699 spin_unlock(&pool->prp_lock);
702 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
704 struct obd_import *imp = req->rq_import;
705 struct list_head *tmp;
706 struct ptlrpc_request *iter;
708 assert_spin_locked(&imp->imp_lock);
709 LASSERT(list_empty(&req->rq_unreplied_list));
711 /* unreplied list is sorted by xid in ascending order */
712 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
713 iter = list_entry(tmp, struct ptlrpc_request,
716 LASSERT(req->rq_xid != iter->rq_xid);
717 if (req->rq_xid < iter->rq_xid)
719 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
722 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
725 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
727 req->rq_xid = ptlrpc_next_xid();
728 ptlrpc_add_unreplied(req);
731 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
733 spin_lock(&req->rq_import->imp_lock);
734 ptlrpc_assign_next_xid_nolock(req);
735 spin_unlock(&req->rq_import->imp_lock);
738 static atomic64_t ptlrpc_last_xid;
740 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
742 spin_lock(&req->rq_import->imp_lock);
743 list_del_init(&req->rq_unreplied_list);
744 ptlrpc_assign_next_xid_nolock(req);
745 spin_unlock(&req->rq_import->imp_lock);
746 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
749 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
751 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
755 opc = lustre_msg_get_opc(req->rq_reqmsg);
756 tag = obd_get_mod_rpc_slot(cli, opc);
757 lustre_msg_set_tag(req->rq_reqmsg, tag);
758 ptlrpc_reassign_next_xid(req);
760 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
762 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
764 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
767 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
768 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
770 obd_put_mod_rpc_slot(cli, opc, tag);
773 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
775 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
776 __u32 version, int opcode, char **bufs,
777 struct ptlrpc_cli_ctx *ctx)
780 struct obd_import *imp;
786 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
787 imp = request->rq_import;
788 lengths = request->rq_pill.rc_area[RCL_CLIENT];
791 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
793 rc = sptlrpc_req_get_ctx(request);
797 sptlrpc_req_set_flavor(request, opcode);
799 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
804 lustre_msg_add_version(request->rq_reqmsg, version);
805 request->rq_send_state = LUSTRE_IMP_FULL;
806 request->rq_type = PTL_RPC_MSG_REQUEST;
808 request->rq_req_cbid.cbid_fn = request_out_callback;
809 request->rq_req_cbid.cbid_arg = request;
811 request->rq_reply_cbid.cbid_fn = reply_in_callback;
812 request->rq_reply_cbid.cbid_arg = request;
814 request->rq_reply_deadline = 0;
815 request->rq_bulk_deadline = 0;
816 request->rq_req_deadline = 0;
817 request->rq_phase = RQ_PHASE_NEW;
818 request->rq_next_phase = RQ_PHASE_UNDEFINED;
820 request->rq_request_portal = imp->imp_client->cli_request_portal;
821 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
823 ptlrpc_at_set_req_timeout(request);
825 lustre_msg_set_opc(request->rq_reqmsg, opcode);
827 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
828 if (cfs_fail_val == opcode) {
829 time64_t *fail_t = NULL, *fail2_t = NULL;
831 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
832 fail_t = &request->rq_bulk_deadline;
833 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
834 fail_t = &request->rq_reply_deadline;
835 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
836 fail_t = &request->rq_req_deadline;
837 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
838 fail_t = &request->rq_reply_deadline;
839 fail2_t = &request->rq_bulk_deadline;
840 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
841 time64_t now = ktime_get_real_seconds();
842 u64 xid = ((u64)now >> 4) << 24;
844 atomic64_set(&ptlrpc_last_xid, xid);
848 *fail_t = ktime_get_real_seconds() +
849 PTLRPC_REQ_LONG_UNLINK;
852 *fail2_t = ktime_get_real_seconds() +
853 PTLRPC_REQ_LONG_UNLINK;
856 * The RPC is infected, let the test to change the
859 msleep(4 * MSEC_PER_SEC);
862 ptlrpc_assign_next_xid(request);
867 LASSERT(!request->rq_pool);
868 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
870 atomic_dec(&imp->imp_reqs);
871 class_import_put(imp);
875 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
878 * Pack request buffers for network transfer, performing necessary encryption
879 * steps if necessary.
881 int ptlrpc_request_pack(struct ptlrpc_request *request,
882 __u32 version, int opcode)
884 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
886 EXPORT_SYMBOL(ptlrpc_request_pack);
889 * Helper function to allocate new request on import \a imp
890 * and possibly using existing request from pool \a pool if provided.
891 * Returns allocated request structure with import field filled or
895 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
896 struct ptlrpc_request_pool *pool)
898 struct ptlrpc_request *request = NULL;
900 request = ptlrpc_request_cache_alloc(GFP_NOFS);
902 if (!request && pool)
903 request = ptlrpc_prep_req_from_pool(pool);
906 ptlrpc_cli_req_init(request);
908 LASSERTF((unsigned long)imp > 0x1000, "%p\n", imp);
909 LASSERT(imp != LP_POISON);
910 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
912 LASSERT(imp->imp_client != LP_POISON);
914 request->rq_import = class_import_get(imp);
915 atomic_inc(&imp->imp_reqs);
917 CERROR("request allocation out of memory\n");
923 static int ptlrpc_reconnect_if_idle(struct obd_import *imp)
928 * initiate connection if needed when the import has been
929 * referenced by the new request to avoid races with disconnect.
930 * serialize this check against conditional state=IDLE
931 * in ptlrpc_disconnect_idle_interpret()
933 spin_lock(&imp->imp_lock);
934 if (imp->imp_state == LUSTRE_IMP_IDLE) {
935 imp->imp_generation++;
936 imp->imp_initiated_at = imp->imp_generation;
937 imp->imp_state = LUSTRE_IMP_NEW;
939 /* connect_import_locked releases imp_lock */
940 rc = ptlrpc_connect_import_locked(imp);
943 ptlrpc_pinger_add_import(imp);
945 spin_unlock(&imp->imp_lock);
951 * Helper function for creating a request.
952 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
953 * buffer structures according to capsule template \a format.
954 * Returns allocated request structure pointer or NULL on error.
956 static struct ptlrpc_request *
957 ptlrpc_request_alloc_internal(struct obd_import *imp,
958 struct ptlrpc_request_pool *pool,
959 const struct req_format *format)
961 struct ptlrpc_request *request;
963 request = __ptlrpc_request_alloc(imp, pool);
967 /* don't make expensive check for idling connection
968 * if it's already connected */
969 if (unlikely(imp->imp_state != LUSTRE_IMP_FULL)) {
970 if (ptlrpc_reconnect_if_idle(imp) < 0) {
971 atomic_dec(&imp->imp_reqs);
972 ptlrpc_request_free(request);
977 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
978 req_capsule_set(&request->rq_pill, format);
983 * Allocate new request structure for import \a imp and initialize its
984 * buffer structure according to capsule template \a format.
986 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
987 const struct req_format *format)
989 return ptlrpc_request_alloc_internal(imp, NULL, format);
991 EXPORT_SYMBOL(ptlrpc_request_alloc);
994 * Allocate new request structure for import \a imp from pool \a pool and
995 * initialize its buffer structure according to capsule template \a format.
997 struct ptlrpc_request *
998 ptlrpc_request_alloc_pool(struct obd_import *imp,
999 struct ptlrpc_request_pool *pool,
1000 const struct req_format *format)
1002 return ptlrpc_request_alloc_internal(imp, pool, format);
1004 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
1007 * For requests not from pool, free memory of the request structure.
1008 * For requests obtained from a pool earlier, return request back to pool.
1010 void ptlrpc_request_free(struct ptlrpc_request *request)
1012 if (request->rq_pool)
1013 __ptlrpc_free_req_to_pool(request);
1015 ptlrpc_request_cache_free(request);
1017 EXPORT_SYMBOL(ptlrpc_request_free);
1020 * Allocate new request for operatione \a opcode and immediatelly pack it for
1022 * Only used for simple requests like OBD_PING where the only important
1023 * part of the request is operation itself.
1024 * Returns allocated request or NULL on error.
1026 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1027 const struct req_format *format,
1028 __u32 version, int opcode)
1030 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1034 rc = ptlrpc_request_pack(req, version, opcode);
1036 ptlrpc_request_free(req);
1042 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1045 * Allocate and initialize new request set structure on the current CPT.
1046 * Returns a pointer to the newly allocated set structure or NULL on error.
1048 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1050 struct ptlrpc_request_set *set;
1054 cpt = cfs_cpt_current(cfs_cpt_tab, 0);
1055 OBD_CPT_ALLOC(set, cfs_cpt_tab, cpt, sizeof(*set));
1058 atomic_set(&set->set_refcount, 1);
1059 INIT_LIST_HEAD(&set->set_requests);
1060 init_waitqueue_head(&set->set_waitq);
1061 atomic_set(&set->set_new_count, 0);
1062 atomic_set(&set->set_remaining, 0);
1063 spin_lock_init(&set->set_new_req_lock);
1064 INIT_LIST_HEAD(&set->set_new_requests);
1065 set->set_max_inflight = UINT_MAX;
1066 set->set_producer = NULL;
1067 set->set_producer_arg = NULL;
1072 EXPORT_SYMBOL(ptlrpc_prep_set);
1075 * Allocate and initialize new request set structure with flow control
1076 * extension. This extension allows to control the number of requests in-flight
1077 * for the whole set. A callback function to generate requests must be provided
1078 * and the request set will keep the number of requests sent over the wire to
1080 * Returns a pointer to the newly allocated set structure or NULL on error.
1082 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1086 struct ptlrpc_request_set *set;
1088 set = ptlrpc_prep_set();
1092 set->set_max_inflight = max;
1093 set->set_producer = func;
1094 set->set_producer_arg = arg;
1100 * Wind down and free request set structure previously allocated with
1102 * Ensures that all requests on the set have completed and removes
1103 * all requests from the request list in a set.
1104 * If any unsent request happen to be on the list, pretends that they got
1105 * an error in flight and calls their completion handler.
1107 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1109 struct list_head *tmp;
1110 struct list_head *next;
1116 /* Requests on the set should either all be completed, or all be new */
1117 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1118 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1119 list_for_each(tmp, &set->set_requests) {
1120 struct ptlrpc_request *req =
1121 list_entry(tmp, struct ptlrpc_request,
1124 LASSERT(req->rq_phase == expected_phase);
1128 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1129 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1130 atomic_read(&set->set_remaining), n);
1132 list_for_each_safe(tmp, next, &set->set_requests) {
1133 struct ptlrpc_request *req =
1134 list_entry(tmp, struct ptlrpc_request,
1136 list_del_init(&req->rq_set_chain);
1138 LASSERT(req->rq_phase == expected_phase);
1140 if (req->rq_phase == RQ_PHASE_NEW) {
1141 ptlrpc_req_interpret(NULL, req, -EBADR);
1142 atomic_dec(&set->set_remaining);
1145 spin_lock(&req->rq_lock);
1147 req->rq_invalid_rqset = 0;
1148 spin_unlock(&req->rq_lock);
1150 ptlrpc_req_finished(req);
1153 LASSERT(atomic_read(&set->set_remaining) == 0);
1155 ptlrpc_reqset_put(set);
1158 EXPORT_SYMBOL(ptlrpc_set_destroy);
1161 * Add a new request to the general purpose request set.
1162 * Assumes request reference from the caller.
1164 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1165 struct ptlrpc_request *req)
1167 if (set == PTLRPCD_SET) {
1168 ptlrpcd_add_req(req);
1172 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1173 LASSERT(list_empty(&req->rq_set_chain));
1175 if (req->rq_allow_intr)
1176 set->set_allow_intr = 1;
1178 /* The set takes over the caller's request reference */
1179 list_add_tail(&req->rq_set_chain, &set->set_requests);
1181 atomic_inc(&set->set_remaining);
1182 req->rq_queued_time = ktime_get_seconds();
1185 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1187 if (set->set_producer)
1189 * If the request set has a producer callback, the RPC must be
1190 * sent straight away
1192 ptlrpc_send_new_req(req);
1194 EXPORT_SYMBOL(ptlrpc_set_add_req);
1197 * Add a request to a request with dedicated server thread
1198 * and wake the thread to make any necessary processing.
1199 * Currently only used for ptlrpcd.
1201 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1202 struct ptlrpc_request *req)
1204 struct ptlrpc_request_set *set = pc->pc_set;
1207 LASSERT(req->rq_set == NULL);
1208 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1210 spin_lock(&set->set_new_req_lock);
1212 * The set takes over the caller's request reference.
1215 req->rq_queued_time = ktime_get_seconds();
1216 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1217 count = atomic_inc_return(&set->set_new_count);
1218 spin_unlock(&set->set_new_req_lock);
1220 /* Only need to call wakeup once for the first entry. */
1222 wake_up(&set->set_waitq);
1225 * XXX: It maybe unnecessary to wakeup all the partners. But to
1226 * guarantee the async RPC can be processed ASAP, we have
1227 * no other better choice. It maybe fixed in future.
1229 for (i = 0; i < pc->pc_npartners; i++)
1230 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1235 * Based on the current state of the import, determine if the request
1236 * can be sent, is an error, or should be delayed.
1238 * Returns true if this request should be delayed. If false, and
1239 * *status is set, then the request can not be sent and *status is the
1240 * error code. If false and status is 0, then request can be sent.
1242 * The imp->imp_lock must be held.
1244 static int ptlrpc_import_delay_req(struct obd_import *imp,
1245 struct ptlrpc_request *req, int *status)
1253 if (req->rq_ctx_init || req->rq_ctx_fini) {
1254 /* always allow ctx init/fini rpc go through */
1255 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1256 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1258 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1259 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1262 * pings or MDS-equivalent STATFS may safely
1265 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1266 D_HA : D_ERROR, req, "IMP_CLOSED");
1268 } else if (ptlrpc_send_limit_expired(req)) {
1269 /* probably doesn't need to be a D_ERROR afterinitial testing */
1270 DEBUG_REQ(D_HA, req, "send limit expired");
1271 *status = -ETIMEDOUT;
1272 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1273 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1274 ;/* allow CONNECT even if import is invalid */
1275 if (atomic_read(&imp->imp_inval_count) != 0) {
1276 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1279 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1280 if (!imp->imp_deactive)
1281 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1282 *status = -ESHUTDOWN; /* b=12940 */
1283 } else if (req->rq_import_generation != imp->imp_generation) {
1284 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1286 } else if (req->rq_send_state != imp->imp_state) {
1287 /* invalidate in progress - any requests should be drop */
1288 if (atomic_read(&imp->imp_inval_count) != 0) {
1289 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1291 } else if (req->rq_no_delay &&
1292 imp->imp_generation != imp->imp_initiated_at) {
1293 /* ignore nodelay for requests initiating connections */
1295 } else if (req->rq_allow_replay &&
1296 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1297 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1298 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1299 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1300 DEBUG_REQ(D_HA, req, "allow during recovery");
1310 * Decide if the error message should be printed to the console or not.
1311 * Makes its decision based on request type, status, and failure frequency.
1313 * \param[in] req request that failed and may need a console message
1315 * \retval false if no message should be printed
1316 * \retval true if console message should be printed
1318 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1320 LASSERT(req->rq_reqmsg != NULL);
1322 /* Suppress particular reconnect errors which are to be expected. */
1323 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1324 /* Suppress timed out reconnect requests */
1325 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1330 * Suppress most unavailable/again reconnect requests, but
1331 * print occasionally so it is clear client is trying to
1332 * connect to a server where no target is running.
1334 if ((err == -ENODEV || err == -EAGAIN) &&
1335 req->rq_import->imp_conn_cnt % 30 != 20)
1339 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1340 /* -EAGAIN is normal when using POSIX flocks */
1343 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1344 (req->rq_xid & 0xf) != 10)
1345 /* Suppress most ping requests, they may fail occasionally */
1352 * Check request processing status.
1353 * Returns the status.
1355 static int ptlrpc_check_status(struct ptlrpc_request *req)
1360 rc = lustre_msg_get_status(req->rq_repmsg);
1361 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1362 struct obd_import *imp = req->rq_import;
1363 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1364 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1366 if (ptlrpc_console_allow(req, opc, rc))
1367 LCONSOLE_ERROR_MSG(0x11,
1368 "%s: operation %s to node %s failed: rc = %d\n",
1369 imp->imp_obd->obd_name,
1371 libcfs_nid2str(nid), rc);
1372 RETURN(rc < 0 ? rc : -EINVAL);
1376 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1382 * save pre-versions of objects into request for replay.
1383 * Versions are obtained from server reply.
1386 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1388 struct lustre_msg *repmsg = req->rq_repmsg;
1389 struct lustre_msg *reqmsg = req->rq_reqmsg;
1390 __u64 *versions = lustre_msg_get_versions(repmsg);
1393 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1397 lustre_msg_set_versions(reqmsg, versions);
1398 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1399 versions[0], versions[1]);
1404 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1406 struct ptlrpc_request *req;
1408 assert_spin_locked(&imp->imp_lock);
1409 if (list_empty(&imp->imp_unreplied_list))
1412 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1414 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1416 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1417 imp->imp_known_replied_xid = req->rq_xid - 1;
1419 return req->rq_xid - 1;
1423 * Callback function called when client receives RPC reply for \a req.
1424 * Returns 0 on success or error code.
1425 * The return alue would be assigned to req->rq_status by the caller
1426 * as request processing status.
1427 * This function also decides if the request needs to be saved for later replay.
1429 static int after_reply(struct ptlrpc_request *req)
1431 struct obd_import *imp = req->rq_import;
1432 struct obd_device *obd = req->rq_import->imp_obd;
1439 LASSERT(obd != NULL);
1440 /* repbuf must be unlinked */
1441 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1443 if (req->rq_reply_truncated) {
1444 if (ptlrpc_no_resend(req)) {
1445 DEBUG_REQ(D_ERROR, req,
1446 "reply buffer overflow, expected=%d, actual size=%d",
1447 req->rq_nob_received, req->rq_repbuf_len);
1451 sptlrpc_cli_free_repbuf(req);
1453 * Pass the required reply buffer size (include
1454 * space for early reply).
1455 * NB: no need to roundup because alloc_repbuf
1458 req->rq_replen = req->rq_nob_received;
1459 req->rq_nob_received = 0;
1460 spin_lock(&req->rq_lock);
1462 spin_unlock(&req->rq_lock);
1466 work_start = ktime_get_real();
1467 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1470 * NB Until this point, the whole of the incoming message,
1471 * including buflens, status etc is in the sender's byte order.
1473 rc = sptlrpc_cli_unwrap_reply(req);
1475 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1480 * Security layer unwrap might ask resend this request.
1485 rc = unpack_reply(req);
1489 /* retry indefinitely on EINPROGRESS */
1490 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1491 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1492 time64_t now = ktime_get_real_seconds();
1494 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1495 D_RPCTRACE, req, "resending request on EINPROGRESS");
1496 spin_lock(&req->rq_lock);
1498 spin_unlock(&req->rq_lock);
1499 req->rq_nr_resend++;
1501 /* Readjust the timeout for current conditions */
1502 ptlrpc_at_set_req_timeout(req);
1504 * delay resend to give a chance to the server to get ready.
1505 * The delay is increased by 1s on every resend and is capped to
1506 * the current request timeout (i.e. obd_timeout if AT is off,
1507 * or AT service time x 125% + 5s, see at_est2timeout)
1509 if (req->rq_nr_resend > req->rq_timeout)
1510 req->rq_sent = now + req->rq_timeout;
1512 req->rq_sent = now + req->rq_nr_resend;
1514 /* Resend for EINPROGRESS will use a new XID */
1515 spin_lock(&imp->imp_lock);
1516 list_del_init(&req->rq_unreplied_list);
1517 spin_unlock(&imp->imp_lock);
1522 if (obd->obd_svc_stats) {
1523 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1525 ptlrpc_lprocfs_rpc_sent(req, timediff);
1528 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1529 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1530 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1531 lustre_msg_get_type(req->rq_repmsg));
1535 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1536 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1537 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1538 ptlrpc_at_adj_net_latency(req,
1539 lustre_msg_get_service_timeout(req->rq_repmsg));
1541 rc = ptlrpc_check_status(req);
1545 * Either we've been evicted, or the server has failed for
1546 * some reason. Try to reconnect, and if that fails, punt to
1549 if (ptlrpc_recoverable_error(rc)) {
1550 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1551 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1554 ptlrpc_request_handle_notconn(req);
1559 * Let's look if server sent slv. Do it only for RPC with
1562 ldlm_cli_update_pool(req);
1566 * Store transno in reqmsg for replay.
1568 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1569 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1570 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1573 if (imp->imp_replayable) {
1574 spin_lock(&imp->imp_lock);
1576 * No point in adding already-committed requests to the replay
1577 * list, we will just remove them immediately. b=9829
1579 if (req->rq_transno != 0 &&
1581 lustre_msg_get_last_committed(req->rq_repmsg) ||
1583 /** version recovery */
1584 ptlrpc_save_versions(req);
1585 ptlrpc_retain_replayable_request(req, imp);
1586 } else if (req->rq_commit_cb &&
1587 list_empty(&req->rq_replay_list)) {
1589 * NB: don't call rq_commit_cb if it's already on
1590 * rq_replay_list, ptlrpc_free_committed() will call
1591 * it later, see LU-3618 for details
1593 spin_unlock(&imp->imp_lock);
1594 req->rq_commit_cb(req);
1595 spin_lock(&imp->imp_lock);
1599 * Replay-enabled imports return commit-status information.
1601 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1602 if (likely(committed > imp->imp_peer_committed_transno))
1603 imp->imp_peer_committed_transno = committed;
1605 ptlrpc_free_committed(imp);
1607 if (!list_empty(&imp->imp_replay_list)) {
1608 struct ptlrpc_request *last;
1610 last = list_entry(imp->imp_replay_list.prev,
1611 struct ptlrpc_request,
1614 * Requests with rq_replay stay on the list even if no
1615 * commit is expected.
1617 if (last->rq_transno > imp->imp_peer_committed_transno)
1618 ptlrpc_pinger_commit_expected(imp);
1621 spin_unlock(&imp->imp_lock);
1628 * Helper function to send request \a req over the network for the first time
1629 * Also adjusts request phase.
1630 * Returns 0 on success or error code.
1632 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1634 struct obd_import *imp = req->rq_import;
1639 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1641 /* do not try to go further if there is not enough memory in enc_pool */
1642 if (req->rq_sent && req->rq_bulk)
1643 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1644 pool_is_at_full_capacity())
1647 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1648 (!req->rq_generation_set ||
1649 req->rq_import_generation == imp->imp_generation))
1652 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1654 spin_lock(&imp->imp_lock);
1656 LASSERT(req->rq_xid != 0);
1657 LASSERT(!list_empty(&req->rq_unreplied_list));
1659 if (!req->rq_generation_set)
1660 req->rq_import_generation = imp->imp_generation;
1662 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1663 spin_lock(&req->rq_lock);
1664 req->rq_waiting = 1;
1665 spin_unlock(&req->rq_lock);
1667 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1668 ptlrpc_import_state_name(req->rq_send_state),
1669 ptlrpc_import_state_name(imp->imp_state));
1670 LASSERT(list_empty(&req->rq_list));
1671 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1672 atomic_inc(&req->rq_import->imp_inflight);
1673 spin_unlock(&imp->imp_lock);
1678 spin_unlock(&imp->imp_lock);
1679 req->rq_status = rc;
1680 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1684 LASSERT(list_empty(&req->rq_list));
1685 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1686 atomic_inc(&req->rq_import->imp_inflight);
1689 * find the known replied XID from the unreplied list, CONNECT
1690 * and DISCONNECT requests are skipped to make the sanity check
1691 * on server side happy. see process_req_last_xid().
1693 * For CONNECT: Because replay requests have lower XID, it'll
1694 * break the sanity check if CONNECT bump the exp_last_xid on
1697 * For DISCONNECT: Since client will abort inflight RPC before
1698 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1699 * than the inflight RPC.
1701 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1702 min_xid = ptlrpc_known_replied_xid(imp);
1703 spin_unlock(&imp->imp_lock);
1705 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1707 lustre_msg_set_status(req->rq_reqmsg, current->pid);
1709 rc = sptlrpc_req_refresh_ctx(req, 0);
1712 req->rq_status = rc;
1715 spin_lock(&req->rq_lock);
1716 req->rq_wait_ctx = 1;
1717 spin_unlock(&req->rq_lock);
1723 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1725 imp->imp_obd->obd_uuid.uuid,
1726 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1727 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1728 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1730 rc = ptl_send_rpc(req, 0);
1731 if (rc == -ENOMEM) {
1732 spin_lock(&imp->imp_lock);
1733 if (!list_empty(&req->rq_list)) {
1734 list_del_init(&req->rq_list);
1735 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1736 wake_up(&req->rq_import->imp_recovery_waitq);
1738 spin_unlock(&imp->imp_lock);
1739 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1743 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1745 spin_lock(&req->rq_lock);
1746 req->rq_net_err = 1;
1747 spin_unlock(&req->rq_lock);
1753 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1758 LASSERT(set->set_producer != NULL);
1760 remaining = atomic_read(&set->set_remaining);
1763 * populate the ->set_requests list with requests until we
1764 * reach the maximum number of RPCs in flight for this set
1766 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1767 rc = set->set_producer(set, set->set_producer_arg);
1768 if (rc == -ENOENT) {
1769 /* no more RPC to produce */
1770 set->set_producer = NULL;
1771 set->set_producer_arg = NULL;
1776 RETURN((atomic_read(&set->set_remaining) - remaining));
1780 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1781 * and no more replies are expected.
1782 * (it is possible to get less replies than requests sent e.g. due to timed out
1783 * requests or requests that we had trouble to send out)
1785 * NOTE: This function contains a potential schedule point (cond_resched()).
1787 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1789 struct list_head *tmp, *next;
1790 LIST_HEAD(comp_reqs);
1791 int force_timer_recalc = 0;
1794 if (atomic_read(&set->set_remaining) == 0)
1797 list_for_each_safe(tmp, next, &set->set_requests) {
1798 struct ptlrpc_request *req =
1799 list_entry(tmp, struct ptlrpc_request,
1801 struct obd_import *imp = req->rq_import;
1802 int unregistered = 0;
1806 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1807 list_move_tail(&req->rq_set_chain, &comp_reqs);
1812 * This schedule point is mainly for the ptlrpcd caller of this
1813 * function. Most ptlrpc sets are not long-lived and unbounded
1814 * in length, but at the least the set used by the ptlrpcd is.
1815 * Since the processing time is unbounded, we need to insert an
1816 * explicit schedule point to make the thread well-behaved.
1821 * If the caller requires to allow to be interpreted by force
1822 * and it has really been interpreted, then move the request
1823 * to RQ_PHASE_INTERPRET phase in spite of what the current
1826 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1827 req->rq_status = -EINTR;
1828 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1831 * Since it is interpreted and we have to wait for
1832 * the reply to be unlinked, then use sync mode.
1836 GOTO(interpret, req->rq_status);
1839 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1840 force_timer_recalc = 1;
1842 /* delayed send - skip */
1843 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1846 /* delayed resend - skip */
1847 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1848 req->rq_sent > ktime_get_real_seconds())
1851 if (!(req->rq_phase == RQ_PHASE_RPC ||
1852 req->rq_phase == RQ_PHASE_BULK ||
1853 req->rq_phase == RQ_PHASE_INTERPRET ||
1854 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1855 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1856 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1860 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1861 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1862 LASSERT(req->rq_next_phase != req->rq_phase);
1863 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1865 if (req->rq_req_deadline &&
1866 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1867 req->rq_req_deadline = 0;
1868 if (req->rq_reply_deadline &&
1869 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1870 req->rq_reply_deadline = 0;
1871 if (req->rq_bulk_deadline &&
1872 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1873 req->rq_bulk_deadline = 0;
1876 * Skip processing until reply is unlinked. We
1877 * can't return to pool before that and we can't
1878 * call interpret before that. We need to make
1879 * sure that all rdma transfers finished and will
1880 * not corrupt any data.
1882 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1883 ptlrpc_cli_wait_unlink(req))
1885 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1886 ptlrpc_client_bulk_active(req))
1890 * Turn fail_loc off to prevent it from looping
1893 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1894 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1897 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1898 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1903 * Move to next phase if reply was successfully
1906 ptlrpc_rqphase_move(req, req->rq_next_phase);
1909 if (req->rq_phase == RQ_PHASE_INTERPRET)
1910 GOTO(interpret, req->rq_status);
1913 * Note that this also will start async reply unlink.
1915 if (req->rq_net_err && !req->rq_timedout) {
1916 ptlrpc_expire_one_request(req, 1);
1919 * Check if we still need to wait for unlink.
1921 if (ptlrpc_cli_wait_unlink(req) ||
1922 ptlrpc_client_bulk_active(req))
1924 /* If there is no need to resend, fail it now. */
1925 if (req->rq_no_resend) {
1926 if (req->rq_status == 0)
1927 req->rq_status = -EIO;
1928 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1929 GOTO(interpret, req->rq_status);
1936 spin_lock(&req->rq_lock);
1937 req->rq_replied = 0;
1938 spin_unlock(&req->rq_lock);
1939 if (req->rq_status == 0)
1940 req->rq_status = -EIO;
1941 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1942 GOTO(interpret, req->rq_status);
1946 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1947 * so it sets rq_intr regardless of individual rpc
1948 * timeouts. The synchronous IO waiting path sets
1949 * rq_intr irrespective of whether ptlrpcd
1950 * has seen a timeout. Our policy is to only interpret
1951 * interrupted rpcs after they have timed out, so we
1952 * need to enforce that here.
1955 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1956 req->rq_wait_ctx)) {
1957 req->rq_status = -EINTR;
1958 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1959 GOTO(interpret, req->rq_status);
1962 if (req->rq_phase == RQ_PHASE_RPC) {
1963 if (req->rq_timedout || req->rq_resend ||
1964 req->rq_waiting || req->rq_wait_ctx) {
1967 if (!ptlrpc_unregister_reply(req, 1)) {
1968 ptlrpc_unregister_bulk(req, 1);
1972 spin_lock(&imp->imp_lock);
1973 if (ptlrpc_import_delay_req(imp, req,
1976 * put on delay list - only if we wait
1977 * recovery finished - before send
1979 list_move_tail(&req->rq_list,
1980 &imp->imp_delayed_list);
1981 spin_unlock(&imp->imp_lock);
1986 req->rq_status = status;
1987 ptlrpc_rqphase_move(req,
1988 RQ_PHASE_INTERPRET);
1989 spin_unlock(&imp->imp_lock);
1990 GOTO(interpret, req->rq_status);
1992 /* ignore on just initiated connections */
1993 if (ptlrpc_no_resend(req) &&
1994 !req->rq_wait_ctx &&
1995 imp->imp_generation !=
1996 imp->imp_initiated_at) {
1997 req->rq_status = -ENOTCONN;
1998 ptlrpc_rqphase_move(req,
1999 RQ_PHASE_INTERPRET);
2000 spin_unlock(&imp->imp_lock);
2001 GOTO(interpret, req->rq_status);
2004 /* don't resend too fast in case of network
2007 if (ktime_get_real_seconds() < (req->rq_sent + 1)
2008 && req->rq_net_err && req->rq_timedout) {
2010 DEBUG_REQ(D_INFO, req,
2011 "throttle request");
2012 /* Don't try to resend RPC right away
2013 * as it is likely it will fail again
2014 * and ptlrpc_check_set() will be
2015 * called again, keeping this thread
2016 * busy. Instead, wait for the next
2017 * timeout. Flag it as resend to
2018 * ensure we don't wait to long.
2021 spin_unlock(&imp->imp_lock);
2025 list_move_tail(&req->rq_list,
2026 &imp->imp_sending_list);
2028 spin_unlock(&imp->imp_lock);
2030 spin_lock(&req->rq_lock);
2031 req->rq_waiting = 0;
2032 spin_unlock(&req->rq_lock);
2034 if (req->rq_timedout || req->rq_resend) {
2036 * This is re-sending anyways,
2037 * let's mark req as resend.
2039 spin_lock(&req->rq_lock);
2041 spin_unlock(&req->rq_lock);
2044 * rq_wait_ctx is only touched by ptlrpcd,
2045 * so no lock is needed here.
2047 status = sptlrpc_req_refresh_ctx(req, 0);
2050 req->rq_status = status;
2051 spin_lock(&req->rq_lock);
2052 req->rq_wait_ctx = 0;
2053 spin_unlock(&req->rq_lock);
2054 force_timer_recalc = 1;
2056 spin_lock(&req->rq_lock);
2057 req->rq_wait_ctx = 1;
2058 spin_unlock(&req->rq_lock);
2063 spin_lock(&req->rq_lock);
2064 req->rq_wait_ctx = 0;
2065 spin_unlock(&req->rq_lock);
2069 * In any case, the previous bulk should be
2070 * cleaned up to prepare for the new sending
2073 !ptlrpc_unregister_bulk(req, 1))
2076 rc = ptl_send_rpc(req, 0);
2077 if (rc == -ENOMEM) {
2078 spin_lock(&imp->imp_lock);
2079 if (!list_empty(&req->rq_list))
2080 list_del_init(&req->rq_list);
2081 spin_unlock(&imp->imp_lock);
2082 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2086 DEBUG_REQ(D_HA, req,
2087 "send failed: rc = %d", rc);
2088 force_timer_recalc = 1;
2089 spin_lock(&req->rq_lock);
2090 req->rq_net_err = 1;
2091 spin_unlock(&req->rq_lock);
2094 /* need to reset the timeout */
2095 force_timer_recalc = 1;
2098 spin_lock(&req->rq_lock);
2100 if (ptlrpc_client_early(req)) {
2101 ptlrpc_at_recv_early_reply(req);
2102 spin_unlock(&req->rq_lock);
2106 /* Still waiting for a reply? */
2107 if (ptlrpc_client_recv(req)) {
2108 spin_unlock(&req->rq_lock);
2112 /* Did we actually receive a reply? */
2113 if (!ptlrpc_client_replied(req)) {
2114 spin_unlock(&req->rq_lock);
2118 spin_unlock(&req->rq_lock);
2121 * unlink from net because we are going to
2122 * swab in-place of reply buffer
2124 unregistered = ptlrpc_unregister_reply(req, 1);
2128 req->rq_status = after_reply(req);
2133 * If there is no bulk associated with this request,
2134 * then we're done and should let the interpreter
2135 * process the reply. Similarly if the RPC returned
2136 * an error, and therefore the bulk will never arrive.
2138 if (!req->rq_bulk || req->rq_status < 0) {
2139 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2140 GOTO(interpret, req->rq_status);
2143 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2146 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2147 if (ptlrpc_client_bulk_active(req))
2150 if (req->rq_bulk->bd_failure) {
2152 * The RPC reply arrived OK, but the bulk screwed
2153 * up! Dead weird since the server told us the RPC
2154 * was good after getting the REPLY for her GET or
2155 * the ACK for her PUT.
2157 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2158 req->rq_status = -EIO;
2161 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2164 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2167 * This moves to "unregistering" phase we need to wait for
2170 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2171 /* start async bulk unlink too */
2172 ptlrpc_unregister_bulk(req, 1);
2176 if (!ptlrpc_unregister_bulk(req, async))
2180 * When calling interpret receiving already should be
2183 LASSERT(!req->rq_receiving_reply);
2185 ptlrpc_req_interpret(env, req, req->rq_status);
2187 if (ptlrpcd_check_work(req)) {
2188 atomic_dec(&set->set_remaining);
2191 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2195 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2197 imp->imp_obd->obd_uuid.uuid,
2198 lustre_msg_get_status(req->rq_reqmsg),
2200 obd_import_nid2str(imp),
2201 lustre_msg_get_opc(req->rq_reqmsg),
2202 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2204 spin_lock(&imp->imp_lock);
2206 * Request already may be not on sending or delaying list. This
2207 * may happen in the case of marking it erroneous for the case
2208 * ptlrpc_import_delay_req(req, status) find it impossible to
2209 * allow sending this rpc and returns *status != 0.
2211 if (!list_empty(&req->rq_list)) {
2212 list_del_init(&req->rq_list);
2213 if (atomic_dec_and_test(&imp->imp_inflight))
2214 wake_up(&imp->imp_recovery_waitq);
2216 list_del_init(&req->rq_unreplied_list);
2217 spin_unlock(&imp->imp_lock);
2219 atomic_dec(&set->set_remaining);
2220 wake_up(&imp->imp_recovery_waitq);
2222 if (set->set_producer) {
2223 /* produce a new request if possible */
2224 if (ptlrpc_set_producer(set) > 0)
2225 force_timer_recalc = 1;
2228 * free the request that has just been completed
2229 * in order not to pollute set->set_requests
2231 list_del_init(&req->rq_set_chain);
2232 spin_lock(&req->rq_lock);
2234 req->rq_invalid_rqset = 0;
2235 spin_unlock(&req->rq_lock);
2237 /* record rq_status to compute the final status later */
2238 if (req->rq_status != 0)
2239 set->set_rc = req->rq_status;
2240 ptlrpc_req_finished(req);
2242 list_move_tail(&req->rq_set_chain, &comp_reqs);
2247 * move completed request at the head of list so it's easier for
2248 * caller to find them
2250 list_splice(&comp_reqs, &set->set_requests);
2252 /* If we hit an error, we want to recover promptly. */
2253 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2255 EXPORT_SYMBOL(ptlrpc_check_set);
2258 * Time out request \a req. is \a async_unlink is set, that means do not wait
2259 * until LNet actually confirms network buffer unlinking.
2260 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2262 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2264 struct obd_import *imp = req->rq_import;
2265 unsigned int debug_mask = D_RPCTRACE;
2269 spin_lock(&req->rq_lock);
2270 req->rq_timedout = 1;
2271 spin_unlock(&req->rq_lock);
2273 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2274 lustre_msg_get_status(req->rq_reqmsg)))
2275 debug_mask = D_WARNING;
2276 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2277 req->rq_net_err ? "failed due to network error" :
2278 ((req->rq_real_sent == 0 ||
2279 req->rq_real_sent < req->rq_sent ||
2280 req->rq_real_sent >= req->rq_deadline) ?
2281 "timed out for sent delay" : "timed out for slow reply"),
2282 req->rq_sent, req->rq_real_sent);
2284 if (imp && obd_debug_peer_on_timeout)
2285 LNetDebugPeer(imp->imp_connection->c_peer);
2287 ptlrpc_unregister_reply(req, async_unlink);
2288 ptlrpc_unregister_bulk(req, async_unlink);
2290 if (obd_dump_on_timeout)
2291 libcfs_debug_dumplog();
2294 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2298 atomic_inc(&imp->imp_timeouts);
2300 /* The DLM server doesn't want recovery run on its imports. */
2301 if (imp->imp_dlm_fake)
2305 * If this request is for recovery or other primordial tasks,
2306 * then error it out here.
2308 if (req->rq_ctx_init || req->rq_ctx_fini ||
2309 req->rq_send_state != LUSTRE_IMP_FULL ||
2310 imp->imp_obd->obd_no_recov) {
2311 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2312 ptlrpc_import_state_name(req->rq_send_state),
2313 ptlrpc_import_state_name(imp->imp_state));
2314 spin_lock(&req->rq_lock);
2315 req->rq_status = -ETIMEDOUT;
2317 spin_unlock(&req->rq_lock);
2322 * if a request can't be resent we can't wait for an answer after
2325 if (ptlrpc_no_resend(req)) {
2326 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2330 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2336 * Time out all uncompleted requests in request set pointed by \a data
2337 * This is called when a wait times out.
2339 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2341 struct list_head *tmp;
2342 time64_t now = ktime_get_real_seconds();
2345 LASSERT(set != NULL);
2348 * A timeout expired. See which reqs it applies to...
2350 list_for_each(tmp, &set->set_requests) {
2351 struct ptlrpc_request *req =
2352 list_entry(tmp, struct ptlrpc_request,
2355 /* don't expire request waiting for context */
2356 if (req->rq_wait_ctx)
2359 /* Request in-flight? */
2360 if (!((req->rq_phase == RQ_PHASE_RPC &&
2361 !req->rq_waiting && !req->rq_resend) ||
2362 (req->rq_phase == RQ_PHASE_BULK)))
2365 if (req->rq_timedout || /* already dealt with */
2366 req->rq_deadline > now) /* not expired */
2370 * Deal with this guy. Do it asynchronously to not block
2373 ptlrpc_expire_one_request(req, 1);
2375 * Loops require that we resched once in a while to avoid
2376 * RCU stalls and a few other problems.
2384 * Interrupts (sets interrupted flag) all uncompleted requests in
2385 * a set \a data. This is called when a wait_event is interrupted
2388 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2390 struct list_head *tmp;
2392 LASSERT(set != NULL);
2393 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2395 list_for_each(tmp, &set->set_requests) {
2396 struct ptlrpc_request *req =
2397 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2402 if (req->rq_phase != RQ_PHASE_RPC &&
2403 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2404 !req->rq_allow_intr)
2407 spin_lock(&req->rq_lock);
2409 spin_unlock(&req->rq_lock);
2414 * Get the smallest timeout in the set; this does NOT set a timeout.
2416 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2418 struct list_head *tmp;
2419 time64_t now = ktime_get_real_seconds();
2421 struct ptlrpc_request *req;
2425 list_for_each(tmp, &set->set_requests) {
2426 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2428 /* Request in-flight? */
2429 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2430 (req->rq_phase == RQ_PHASE_BULK) ||
2431 (req->rq_phase == RQ_PHASE_NEW)))
2434 /* Already timed out. */
2435 if (req->rq_timedout)
2438 /* Waiting for ctx. */
2439 if (req->rq_wait_ctx)
2442 if (req->rq_phase == RQ_PHASE_NEW)
2443 deadline = req->rq_sent;
2444 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2445 deadline = req->rq_sent;
2447 deadline = req->rq_sent + req->rq_timeout;
2449 if (deadline <= now) /* actually expired already */
2450 timeout = 1; /* ASAP */
2451 else if (timeout == 0 || timeout > deadline - now)
2452 timeout = deadline - now;
2458 * Send all unset request from the set and then wait untill all
2459 * requests in the set complete (either get a reply, timeout, get an
2460 * error or otherwise be interrupted).
2461 * Returns 0 on success or error code otherwise.
2463 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2465 struct list_head *tmp;
2466 struct ptlrpc_request *req;
2471 if (set->set_producer)
2472 (void)ptlrpc_set_producer(set);
2474 list_for_each(tmp, &set->set_requests) {
2475 req = list_entry(tmp, struct ptlrpc_request,
2477 if (req->rq_phase == RQ_PHASE_NEW)
2478 (void)ptlrpc_send_new_req(req);
2481 if (list_empty(&set->set_requests))
2485 timeout = ptlrpc_set_next_timeout(set);
2488 * wait until all complete, interrupted, or an in-flight
2491 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2494 if ((timeout == 0 && !signal_pending(current)) ||
2495 set->set_allow_intr) {
2497 * No requests are in-flight (ether timed out
2498 * or delayed), so we can allow interrupts.
2499 * We still want to block for a limited time,
2500 * so we allow interrupts during the timeout.
2502 rc = l_wait_event_abortable_timeout(
2504 ptlrpc_check_set(NULL, set),
2505 cfs_time_seconds(timeout ? timeout : 1));
2508 ptlrpc_expired_set(set);
2509 } else if (rc < 0) {
2511 ptlrpc_interrupted_set(set);
2517 * At least one request is in flight, so no
2518 * interrupts are allowed. Wait until all
2519 * complete, or an in-flight req times out.
2521 rc = wait_event_idle_timeout(
2523 ptlrpc_check_set(NULL, set),
2524 cfs_time_seconds(timeout ? timeout : 1));
2526 ptlrpc_expired_set(set);
2533 * LU-769 - if we ignored the signal because
2534 * it was already pending when we started, we
2535 * need to handle it now or we risk it being
2538 if (rc == -ETIMEDOUT &&
2539 signal_pending(current)) {
2542 siginitset(&new, LUSTRE_FATAL_SIGS);
2543 sigprocmask(SIG_BLOCK, &new, &old);
2545 * In fact we only interrupt for the
2546 * "fatal" signals like SIGINT or
2547 * SIGKILL. We still ignore less
2548 * important signals since ptlrpc set
2549 * is not easily reentrant from
2552 if (signal_pending(current))
2553 ptlrpc_interrupted_set(set);
2554 sigprocmask(SIG_SETMASK, &old, NULL);
2558 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2561 * -EINTR => all requests have been flagged rq_intr so next
2563 * -ETIMEDOUT => someone timed out. When all reqs have
2564 * timed out, signals are enabled allowing completion with
2566 * I don't really care if we go once more round the loop in
2567 * the error cases -eeb.
2569 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2570 list_for_each(tmp, &set->set_requests) {
2571 req = list_entry(tmp, struct ptlrpc_request,
2573 spin_lock(&req->rq_lock);
2574 req->rq_invalid_rqset = 1;
2575 spin_unlock(&req->rq_lock);
2578 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2580 LASSERT(atomic_read(&set->set_remaining) == 0);
2582 rc = set->set_rc; /* rq_status of already freed requests if any */
2583 list_for_each(tmp, &set->set_requests) {
2584 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2586 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2587 if (req->rq_status != 0)
2588 rc = req->rq_status;
2593 EXPORT_SYMBOL(ptlrpc_set_wait);
2596 * Helper fuction for request freeing.
2597 * Called when request count reached zero and request needs to be freed.
2598 * Removes request from all sorts of sending/replay lists it might be on,
2599 * frees network buffers if any are present.
2600 * If \a locked is set, that means caller is already holding import imp_lock
2601 * and so we no longer need to reobtain it (for certain lists manipulations)
2603 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2610 LASSERT(!request->rq_srv_req);
2611 LASSERT(request->rq_export == NULL);
2612 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2613 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2614 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2615 LASSERTF(!request->rq_replay, "req %p\n", request);
2617 req_capsule_fini(&request->rq_pill);
2620 * We must take it off the imp_replay_list first. Otherwise, we'll set
2621 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2623 if (request->rq_import) {
2625 spin_lock(&request->rq_import->imp_lock);
2626 list_del_init(&request->rq_replay_list);
2627 list_del_init(&request->rq_unreplied_list);
2629 spin_unlock(&request->rq_import->imp_lock);
2631 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2633 if (atomic_read(&request->rq_refcount) != 0) {
2634 DEBUG_REQ(D_ERROR, request,
2635 "freeing request with nonzero refcount");
2639 if (request->rq_repbuf)
2640 sptlrpc_cli_free_repbuf(request);
2642 if (request->rq_import) {
2643 if (!ptlrpcd_check_work(request)) {
2644 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2645 atomic_dec(&request->rq_import->imp_reqs);
2647 class_import_put(request->rq_import);
2648 request->rq_import = NULL;
2650 if (request->rq_bulk)
2651 ptlrpc_free_bulk(request->rq_bulk);
2653 if (request->rq_reqbuf || request->rq_clrbuf)
2654 sptlrpc_cli_free_reqbuf(request);
2656 if (request->rq_cli_ctx)
2657 sptlrpc_req_put_ctx(request, !locked);
2659 if (request->rq_pool)
2660 __ptlrpc_free_req_to_pool(request);
2662 ptlrpc_request_cache_free(request);
2666 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2668 * Drop one request reference. Must be called with import imp_lock held.
2669 * When reference count drops to zero, request is freed.
2671 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2673 assert_spin_locked(&request->rq_import->imp_lock);
2674 (void)__ptlrpc_req_finished(request, 1);
2679 * Drops one reference count for request \a request.
2680 * \a locked set indicates that caller holds import imp_lock.
2681 * Frees the request whe reference count reaches zero.
2683 * \retval 1 the request is freed
2684 * \retval 0 some others still hold references on the request
2686 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2694 LASSERT(request != LP_POISON);
2695 LASSERT(request->rq_reqmsg != LP_POISON);
2697 DEBUG_REQ(D_INFO, request, "refcount now %u",
2698 atomic_read(&request->rq_refcount) - 1);
2700 spin_lock(&request->rq_lock);
2701 count = atomic_dec_return(&request->rq_refcount);
2702 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2705 * For open RPC, the client does not know the EA size (LOV, ACL, and
2706 * so on) before replied, then the client has to reserve very large
2707 * reply buffer. Such buffer will not be released until the RPC freed.
2708 * Since The open RPC is replayable, we need to keep it in the replay
2709 * list until close. If there are a lot of files opened concurrently,
2710 * then the client may be OOM.
2712 * If fact, it is unnecessary to keep reply buffer for open replay,
2713 * related EAs have already been saved via mdc_save_lovea() before
2714 * coming here. So it is safe to free the reply buffer some earlier
2715 * before releasing the RPC to avoid client OOM. LU-9514
2717 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2718 spin_lock(&request->rq_early_free_lock);
2719 sptlrpc_cli_free_repbuf(request);
2720 request->rq_repbuf = NULL;
2721 request->rq_repbuf_len = 0;
2722 request->rq_repdata = NULL;
2723 request->rq_reqdata_len = 0;
2724 spin_unlock(&request->rq_early_free_lock);
2726 spin_unlock(&request->rq_lock);
2729 __ptlrpc_free_req(request, locked);
2735 * Drops one reference count for a request.
2737 void ptlrpc_req_finished(struct ptlrpc_request *request)
2739 __ptlrpc_req_finished(request, 0);
2741 EXPORT_SYMBOL(ptlrpc_req_finished);
2744 * Returns xid of a \a request
2746 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2748 return request->rq_xid;
2750 EXPORT_SYMBOL(ptlrpc_req_xid);
2753 * Disengage the client's reply buffer from the network
2754 * NB does _NOT_ unregister any client-side bulk.
2755 * IDEMPOTENT, but _not_ safe against concurrent callers.
2756 * The request owner (i.e. the thread doing the I/O) must call...
2757 * Returns 0 on success or 1 if unregistering cannot be made.
2759 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2761 bool discard = false;
2765 LASSERT(!in_interrupt());
2767 /* Let's setup deadline for reply unlink. */
2768 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2769 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2770 request->rq_reply_deadline = ktime_get_real_seconds() +
2771 PTLRPC_REQ_LONG_UNLINK;
2774 * Nothing left to do.
2776 if (!__ptlrpc_cli_wait_unlink(request, &discard))
2779 LNetMDUnlink(request->rq_reply_md_h);
2781 if (discard) /* Discard the request-out callback */
2782 __LNetMDUnlink(request->rq_req_md_h, discard);
2785 * Let's check it once again.
2787 if (!ptlrpc_cli_wait_unlink(request))
2790 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2791 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2794 * Do not wait for unlink to finish.
2800 * We have to wait_event_idle_timeout() whatever the result, to get
2801 * a chance to run reply_in_callback(), and to make sure we've
2802 * unlinked before returning a req to the pool.
2805 wait_queue_head_t *wq = (request->rq_set) ?
2806 &request->rq_set->set_waitq :
2807 &request->rq_reply_waitq;
2808 int seconds = PTLRPC_REQ_LONG_UNLINK;
2810 * Network access will complete in finite time but the HUGE
2811 * timeout lets us CWARN for visibility of sluggish NALs
2813 while (seconds > 0 &&
2814 wait_event_idle_timeout(
2816 !ptlrpc_cli_wait_unlink(request),
2817 cfs_time_seconds(1)) == 0)
2820 ptlrpc_rqphase_move(request, request->rq_next_phase);
2824 DEBUG_REQ(D_WARNING, request,
2825 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2826 request->rq_receiving_reply,
2827 request->rq_req_unlinked,
2828 request->rq_reply_unlinked);
2833 static void ptlrpc_free_request(struct ptlrpc_request *req)
2835 spin_lock(&req->rq_lock);
2837 spin_unlock(&req->rq_lock);
2839 if (req->rq_commit_cb)
2840 req->rq_commit_cb(req);
2841 list_del_init(&req->rq_replay_list);
2843 __ptlrpc_req_finished(req, 1);
2847 * the request is committed and dropped from the replay list of its import
2849 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2851 struct obd_import *imp = req->rq_import;
2853 spin_lock(&imp->imp_lock);
2854 if (list_empty(&req->rq_replay_list)) {
2855 spin_unlock(&imp->imp_lock);
2859 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2860 if (imp->imp_replay_cursor == &req->rq_replay_list)
2861 imp->imp_replay_cursor = req->rq_replay_list.next;
2862 ptlrpc_free_request(req);
2865 spin_unlock(&imp->imp_lock);
2867 EXPORT_SYMBOL(ptlrpc_request_committed);
2870 * Iterates through replay_list on import and prunes
2871 * all requests have transno smaller than last_committed for the
2872 * import and don't have rq_replay set.
2873 * Since requests are sorted in transno order, stops when meetign first
2874 * transno bigger than last_committed.
2875 * caller must hold imp->imp_lock
2877 void ptlrpc_free_committed(struct obd_import *imp)
2879 struct ptlrpc_request *req, *saved;
2880 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2881 bool skip_committed_list = true;
2884 LASSERT(imp != NULL);
2885 assert_spin_locked(&imp->imp_lock);
2887 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2888 imp->imp_generation == imp->imp_last_generation_checked) {
2889 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2890 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2893 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2894 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2895 imp->imp_generation);
2897 if (imp->imp_generation != imp->imp_last_generation_checked ||
2898 imp->imp_last_transno_checked == 0)
2899 skip_committed_list = false;
2901 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2902 imp->imp_last_generation_checked = imp->imp_generation;
2904 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2906 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2907 LASSERT(req != last_req);
2910 if (req->rq_transno == 0) {
2911 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2914 if (req->rq_import_generation < imp->imp_generation) {
2915 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2919 /* not yet committed */
2920 if (req->rq_transno > imp->imp_peer_committed_transno) {
2921 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2925 if (req->rq_replay) {
2926 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2927 list_move_tail(&req->rq_replay_list,
2928 &imp->imp_committed_list);
2932 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2933 imp->imp_peer_committed_transno);
2935 ptlrpc_free_request(req);
2938 if (skip_committed_list)
2941 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2943 LASSERT(req->rq_transno != 0);
2944 if (req->rq_import_generation < imp->imp_generation ||
2946 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2947 req->rq_import_generation <
2948 imp->imp_generation ? "stale" : "closed");
2950 if (imp->imp_replay_cursor == &req->rq_replay_list)
2951 imp->imp_replay_cursor =
2952 req->rq_replay_list.next;
2954 ptlrpc_free_request(req);
2961 void ptlrpc_cleanup_client(struct obd_import *imp)
2968 * Schedule previously sent request for resend.
2969 * For bulk requests we assign new xid (to avoid problems with
2970 * lost replies and therefore several transfers landing into same buffer
2971 * from different sending attempts).
2973 void ptlrpc_resend_req(struct ptlrpc_request *req)
2975 DEBUG_REQ(D_HA, req, "going to resend");
2976 spin_lock(&req->rq_lock);
2979 * Request got reply but linked to the import list still.
2980 * Let ptlrpc_check_set() process it.
2982 if (ptlrpc_client_replied(req)) {
2983 spin_unlock(&req->rq_lock);
2984 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2988 req->rq_status = -EAGAIN;
2991 req->rq_net_err = 0;
2992 req->rq_timedout = 0;
2994 ptlrpc_client_wake_req(req);
2995 spin_unlock(&req->rq_lock);
2998 /* XXX: this function and rq_status are currently unused */
2999 void ptlrpc_restart_req(struct ptlrpc_request *req)
3001 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
3002 req->rq_status = -ERESTARTSYS;
3004 spin_lock(&req->rq_lock);
3005 req->rq_restart = 1;
3006 req->rq_timedout = 0;
3007 ptlrpc_client_wake_req(req);
3008 spin_unlock(&req->rq_lock);
3012 * Grab additional reference on a request \a req
3014 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
3017 atomic_inc(&req->rq_refcount);
3020 EXPORT_SYMBOL(ptlrpc_request_addref);
3023 * Add a request to import replay_list.
3024 * Must be called under imp_lock
3026 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3027 struct obd_import *imp)
3029 struct list_head *tmp;
3031 assert_spin_locked(&imp->imp_lock);
3033 if (req->rq_transno == 0) {
3034 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3039 * clear this for new requests that were resent as well
3040 * as resent replayed requests.
3042 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3044 /* don't re-add requests that have been replayed */
3045 if (!list_empty(&req->rq_replay_list))
3048 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3050 spin_lock(&req->rq_lock);
3052 spin_unlock(&req->rq_lock);
3054 LASSERT(imp->imp_replayable);
3055 /* Balanced in ptlrpc_free_committed, usually. */
3056 ptlrpc_request_addref(req);
3057 list_for_each_prev(tmp, &imp->imp_replay_list) {
3058 struct ptlrpc_request *iter = list_entry(tmp,
3059 struct ptlrpc_request,
3063 * We may have duplicate transnos if we create and then
3064 * open a file, or for closes retained if to match creating
3065 * opens, so use req->rq_xid as a secondary key.
3066 * (See bugs 684, 685, and 428.)
3067 * XXX no longer needed, but all opens need transnos!
3069 if (iter->rq_transno > req->rq_transno)
3072 if (iter->rq_transno == req->rq_transno) {
3073 LASSERT(iter->rq_xid != req->rq_xid);
3074 if (iter->rq_xid > req->rq_xid)
3078 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3082 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3086 * Send request and wait until it completes.
3087 * Returns request processing status.
3089 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3091 struct ptlrpc_request_set *set;
3095 LASSERT(req->rq_set == NULL);
3096 LASSERT(!req->rq_receiving_reply);
3098 set = ptlrpc_prep_set();
3100 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3104 /* for distributed debugging */
3105 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3107 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3108 ptlrpc_request_addref(req);
3109 ptlrpc_set_add_req(set, req);
3110 rc = ptlrpc_set_wait(NULL, set);
3111 ptlrpc_set_destroy(set);
3115 EXPORT_SYMBOL(ptlrpc_queue_wait);
3118 * Callback used for replayed requests reply processing.
3119 * In case of successful reply calls registered request replay callback.
3120 * In case of error restart replay process.
3122 static int ptlrpc_replay_interpret(const struct lu_env *env,
3123 struct ptlrpc_request *req,
3126 struct ptlrpc_replay_async_args *aa = args;
3127 struct obd_import *imp = req->rq_import;
3130 atomic_dec(&imp->imp_replay_inflight);
3133 * Note: if it is bulk replay (MDS-MDS replay), then even if
3134 * server got the request, but bulk transfer timeout, let's
3135 * replay the bulk req again
3137 if (!ptlrpc_client_replied(req) ||
3139 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3140 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3141 GOTO(out, rc = -ETIMEDOUT);
3144 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3145 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3146 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3147 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3149 /** VBR: check version failure */
3150 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3151 /** replay was failed due to version mismatch */
3152 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3153 spin_lock(&imp->imp_lock);
3154 imp->imp_vbr_failed = 1;
3155 spin_unlock(&imp->imp_lock);
3156 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3158 /** The transno had better not change over replay. */
3159 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3160 lustre_msg_get_transno(req->rq_repmsg) ||
3161 lustre_msg_get_transno(req->rq_repmsg) == 0,
3163 lustre_msg_get_transno(req->rq_reqmsg),
3164 lustre_msg_get_transno(req->rq_repmsg));
3167 spin_lock(&imp->imp_lock);
3168 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3169 spin_unlock(&imp->imp_lock);
3170 LASSERT(imp->imp_last_replay_transno);
3172 /* transaction number shouldn't be bigger than the latest replayed */
3173 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3174 DEBUG_REQ(D_ERROR, req,
3175 "Reported transno=%llu is bigger than replayed=%llu",
3177 lustre_msg_get_transno(req->rq_reqmsg));
3178 GOTO(out, rc = -EINVAL);
3181 DEBUG_REQ(D_HA, req, "got reply");
3183 /* let the callback do fixups, possibly including in the request */
3184 if (req->rq_replay_cb)
3185 req->rq_replay_cb(req);
3187 if (ptlrpc_client_replied(req) &&
3188 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3189 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3190 lustre_msg_get_status(req->rq_repmsg),
3191 aa->praa_old_status);
3194 * Note: If the replay fails for MDT-MDT recovery, let's
3195 * abort all of the following requests in the replay
3196 * and sending list, because MDT-MDT update requests
3197 * are dependent on each other, see LU-7039
3199 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3200 struct ptlrpc_request *free_req;
3201 struct ptlrpc_request *tmp;
3203 spin_lock(&imp->imp_lock);
3204 list_for_each_entry_safe(free_req, tmp,
3205 &imp->imp_replay_list,
3207 ptlrpc_free_request(free_req);
3210 list_for_each_entry_safe(free_req, tmp,
3211 &imp->imp_committed_list,
3213 ptlrpc_free_request(free_req);
3216 list_for_each_entry_safe(free_req, tmp,
3217 &imp->imp_delayed_list,
3219 spin_lock(&free_req->rq_lock);
3220 free_req->rq_err = 1;
3221 free_req->rq_status = -EIO;
3222 ptlrpc_client_wake_req(free_req);
3223 spin_unlock(&free_req->rq_lock);
3226 list_for_each_entry_safe(free_req, tmp,
3227 &imp->imp_sending_list,
3229 spin_lock(&free_req->rq_lock);
3230 free_req->rq_err = 1;
3231 free_req->rq_status = -EIO;
3232 ptlrpc_client_wake_req(free_req);
3233 spin_unlock(&free_req->rq_lock);
3235 spin_unlock(&imp->imp_lock);
3238 /* Put it back for re-replay. */
3239 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3243 * Errors while replay can set transno to 0, but
3244 * imp_last_replay_transno shouldn't be set to 0 anyway
3246 if (req->rq_transno == 0)
3247 CERROR("Transno is 0 during replay!\n");
3249 /* continue with recovery */
3250 rc = ptlrpc_import_recovery_state_machine(imp);
3252 req->rq_send_state = aa->praa_old_state;
3255 /* this replay failed, so restart recovery */
3256 ptlrpc_connect_import(imp);
3262 * Prepares and queues request for replay.
3263 * Adds it to ptlrpcd queue for actual sending.
3264 * Returns 0 on success.
3266 int ptlrpc_replay_req(struct ptlrpc_request *req)
3268 struct ptlrpc_replay_async_args *aa;
3272 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3274 aa = ptlrpc_req_async_args(aa, req);
3275 memset(aa, 0, sizeof(*aa));
3277 /* Prepare request to be resent with ptlrpcd */
3278 aa->praa_old_state = req->rq_send_state;
3279 req->rq_send_state = LUSTRE_IMP_REPLAY;
3280 req->rq_phase = RQ_PHASE_NEW;
3281 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3283 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3285 req->rq_interpret_reply = ptlrpc_replay_interpret;
3286 /* Readjust the timeout for current conditions */
3287 ptlrpc_at_set_req_timeout(req);
3289 /* Tell server net_latency to calculate how long to wait for reply. */
3290 lustre_msg_set_service_timeout(req->rq_reqmsg,
3291 ptlrpc_at_get_net_latency(req));
3292 DEBUG_REQ(D_HA, req, "REPLAY");
3294 atomic_inc(&req->rq_import->imp_replay_inflight);
3295 spin_lock(&req->rq_lock);
3296 req->rq_early_free_repbuf = 0;
3297 spin_unlock(&req->rq_lock);
3298 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3300 ptlrpcd_add_req(req);
3305 * Aborts all in-flight request on import \a imp sending and delayed lists
3307 void ptlrpc_abort_inflight(struct obd_import *imp)
3309 struct list_head *tmp, *n;
3313 * Make sure that no new requests get processed for this import.
3314 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3315 * this flag and then putting requests on sending_list or delayed_list.
3317 assert_spin_locked(&imp->imp_lock);
3320 * XXX locking? Maybe we should remove each request with the list
3321 * locked? Also, how do we know if the requests on the list are
3322 * being freed at this time?
3324 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3325 struct ptlrpc_request *req = list_entry(tmp,
3326 struct ptlrpc_request,
3329 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3331 spin_lock(&req->rq_lock);
3332 if (req->rq_import_generation < imp->imp_generation) {
3334 req->rq_status = -EIO;
3335 ptlrpc_client_wake_req(req);
3337 spin_unlock(&req->rq_lock);
3340 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3341 struct ptlrpc_request *req =
3342 list_entry(tmp, struct ptlrpc_request, rq_list);
3344 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3346 spin_lock(&req->rq_lock);
3347 if (req->rq_import_generation < imp->imp_generation) {
3349 req->rq_status = -EIO;
3350 ptlrpc_client_wake_req(req);
3352 spin_unlock(&req->rq_lock);
3356 * Last chance to free reqs left on the replay list, but we
3357 * will still leak reqs that haven't committed.
3359 if (imp->imp_replayable)
3360 ptlrpc_free_committed(imp);
3366 * Abort all uncompleted requests in request set \a set
3368 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3370 struct list_head *tmp, *pos;
3372 LASSERT(set != NULL);
3374 list_for_each_safe(pos, tmp, &set->set_requests) {
3375 struct ptlrpc_request *req =
3376 list_entry(pos, struct ptlrpc_request,
3379 spin_lock(&req->rq_lock);
3380 if (req->rq_phase != RQ_PHASE_RPC) {
3381 spin_unlock(&req->rq_lock);
3386 req->rq_status = -EINTR;
3387 ptlrpc_client_wake_req(req);
3388 spin_unlock(&req->rq_lock);
3393 * Initialize the XID for the node. This is common among all requests on
3394 * this node, and only requires the property that it is monotonically
3395 * increasing. It does not need to be sequential. Since this is also used
3396 * as the RDMA match bits, it is important that a single client NOT have
3397 * the same match bits for two different in-flight requests, hence we do
3398 * NOT want to have an XID per target or similar.
3400 * To avoid an unlikely collision between match bits after a client reboot
3401 * (which would deliver old data into the wrong RDMA buffer) initialize
3402 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3403 * If the time is clearly incorrect, we instead use a 62-bit random number.
3404 * In the worst case the random number will overflow 1M RPCs per second in
3405 * 9133 years, or permutations thereof.
3407 #define YEAR_2004 (1ULL << 30)
3408 void ptlrpc_init_xid(void)
3410 time64_t now = ktime_get_real_seconds();
3413 if (now < YEAR_2004) {
3414 get_random_bytes(&xid, sizeof(xid));
3416 xid |= (1ULL << 61);
3418 xid = (u64)now << 20;
3421 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3422 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3424 xid &= PTLRPC_BULK_OPS_MASK;
3425 atomic64_set(&ptlrpc_last_xid, xid);
3429 * Increase xid and returns resulting new value to the caller.
3431 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3432 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3433 * itself uses the last bulk xid needed, so the server can determine the
3434 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3435 * xid must align to a power-of-two value.
3437 * This is assumed to be true due to the initial ptlrpc_last_xid
3438 * value also being initialized to a power-of-two value. LU-1431
3440 __u64 ptlrpc_next_xid(void)
3442 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3446 * If request has a new allocated XID (new request or EINPROGRESS resend),
3447 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3448 * request to ensure previous bulk fails and avoid problems with lost replies
3449 * and therefore several transfers landing into the same buffer from different
3452 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3454 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3456 LASSERT(bd != NULL);
3459 * Generate new matchbits for all resend requests, including
3462 if (req->rq_resend) {
3463 __u64 old_mbits = req->rq_mbits;
3466 * First time resend on -EINPROGRESS will generate new xid,
3467 * so we can actually use the rq_xid as rq_mbits in such case,
3468 * however, it's bit hard to distinguish such resend with a
3469 * 'resend for the -EINPROGRESS resend'. To make it simple,
3470 * we opt to generate mbits for all resend cases.
3472 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3474 req->rq_mbits = ptlrpc_next_xid();
3477 * Old version transfers rq_xid to peer as
3480 spin_lock(&req->rq_import->imp_lock);
3481 list_del_init(&req->rq_unreplied_list);
3482 ptlrpc_assign_next_xid_nolock(req);
3483 spin_unlock(&req->rq_import->imp_lock);
3484 req->rq_mbits = req->rq_xid;
3486 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3487 old_mbits, req->rq_mbits);
3488 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3489 /* Request being sent first time, use xid as matchbits. */
3490 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3491 || req->rq_mbits == 0) {
3492 req->rq_mbits = req->rq_xid;
3494 req->rq_mbits -= bd->bd_md_count - 1;
3498 * Replay request, xid and matchbits have already been
3499 * correctly assigned.
3505 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3506 * that server can infer the number of bulks that were prepared,
3509 req->rq_mbits += bd->bd_md_count - 1;
3512 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3513 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3515 * It's ok to directly set the rq_xid here, since this xid bump
3516 * won't affect the request position in unreplied list.
3518 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3519 req->rq_xid = req->rq_mbits;
3523 * Get a glimpse at what next xid value might have been.
3524 * Returns possible next xid.
3526 __u64 ptlrpc_sample_next_xid(void)
3528 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3530 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3533 * Functions for operating ptlrpc workers.
3535 * A ptlrpc work is a function which will be running inside ptlrpc context.
3536 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3538 * 1. after a work is created, it can be used many times, that is:
3539 * handler = ptlrpcd_alloc_work();
3540 * ptlrpcd_queue_work();
3542 * queue it again when necessary:
3543 * ptlrpcd_queue_work();
3544 * ptlrpcd_destroy_work();
3545 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3546 * it will only be queued once in any time. Also as its name implies, it may
3547 * have delay before it really runs by ptlrpcd thread.
3549 struct ptlrpc_work_async_args {
3550 int (*cb)(const struct lu_env *, void *);
3554 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3556 /* re-initialize the req */
3557 req->rq_timeout = obd_timeout;
3558 req->rq_sent = ktime_get_real_seconds();
3559 req->rq_deadline = req->rq_sent + req->rq_timeout;
3560 req->rq_phase = RQ_PHASE_INTERPRET;
3561 req->rq_next_phase = RQ_PHASE_COMPLETE;
3562 req->rq_xid = ptlrpc_next_xid();
3563 req->rq_import_generation = req->rq_import->imp_generation;
3565 ptlrpcd_add_req(req);
3568 static int work_interpreter(const struct lu_env *env,
3569 struct ptlrpc_request *req, void *args, int rc)
3571 struct ptlrpc_work_async_args *arg = args;
3573 LASSERT(ptlrpcd_check_work(req));
3574 LASSERT(arg->cb != NULL);
3576 rc = arg->cb(env, arg->cbdata);
3578 list_del_init(&req->rq_set_chain);
3581 if (atomic_dec_return(&req->rq_refcount) > 1) {
3582 atomic_set(&req->rq_refcount, 2);
3583 ptlrpcd_add_work_req(req);
3588 static int worker_format;
3590 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3592 return req->rq_pill.rc_fmt == (void *)&worker_format;
3596 * Create a work for ptlrpc.
3598 void *ptlrpcd_alloc_work(struct obd_import *imp,
3599 int (*cb)(const struct lu_env *, void *), void *cbdata)
3601 struct ptlrpc_request *req = NULL;
3602 struct ptlrpc_work_async_args *args;
3608 RETURN(ERR_PTR(-EINVAL));
3610 /* copy some code from deprecated fakereq. */
3611 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3613 CERROR("ptlrpc: run out of memory!\n");
3614 RETURN(ERR_PTR(-ENOMEM));
3617 ptlrpc_cli_req_init(req);
3619 req->rq_send_state = LUSTRE_IMP_FULL;
3620 req->rq_type = PTL_RPC_MSG_REQUEST;
3621 req->rq_import = class_import_get(imp);
3622 req->rq_interpret_reply = work_interpreter;
3623 /* don't want reply */
3624 req->rq_no_delay = req->rq_no_resend = 1;
3625 req->rq_pill.rc_fmt = (void *)&worker_format;
3627 args = ptlrpc_req_async_args(args, req);
3629 args->cbdata = cbdata;
3633 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3635 void ptlrpcd_destroy_work(void *handler)
3637 struct ptlrpc_request *req = handler;
3640 ptlrpc_req_finished(req);
3642 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3644 int ptlrpcd_queue_work(void *handler)
3646 struct ptlrpc_request *req = handler;
3649 * Check if the req is already being queued.
3651 * Here comes a trick: it lacks a way of checking if a req is being
3652 * processed reliably in ptlrpc. Here I have to use refcount of req
3653 * for this purpose. This is okay because the caller should use this
3654 * req as opaque data. - Jinshan
3656 LASSERT(atomic_read(&req->rq_refcount) > 0);
3657 if (atomic_inc_return(&req->rq_refcount) == 2)
3658 ptlrpcd_add_work_req(req);
3661 EXPORT_SYMBOL(ptlrpcd_queue_work);