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);
176 OBD_ALLOC_LARGE(desc->bd_vec,
177 nfrags * sizeof(*desc->bd_vec));
181 spin_lock_init(&desc->bd_lock);
182 init_waitqueue_head(&desc->bd_waitq);
183 desc->bd_max_iov = nfrags;
184 desc->bd_iov_count = 0;
185 desc->bd_portal = portal;
186 desc->bd_type = type;
187 desc->bd_md_count = 0;
188 desc->bd_frag_ops = ops;
189 LASSERT(max_brw > 0);
190 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
192 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
193 * node. Negotiated ocd_brw_size will always be <= this number.
195 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
196 LNetInvalidateMDHandle(&desc->bd_mds[i]);
205 * Prepare bulk descriptor for specified outgoing request \a req that
206 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
207 * the bulk to be sent. Used on client-side.
208 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
211 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
213 unsigned int max_brw,
216 const struct ptlrpc_bulk_frag_ops
219 struct obd_import *imp = req->rq_import;
220 struct ptlrpc_bulk_desc *desc;
223 LASSERT(ptlrpc_is_bulk_op_passive(type));
225 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
229 desc->bd_import = class_import_get(imp);
232 desc->bd_cbid.cbid_fn = client_bulk_callback;
233 desc->bd_cbid.cbid_arg = desc;
235 /* This makes req own desc, and free it when she frees herself */
240 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
242 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
243 struct page *page, int pageoffset, int len,
246 struct bio_vec *kiov;
248 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
249 LASSERT(page != NULL);
250 LASSERT(pageoffset >= 0);
252 LASSERT(pageoffset + len <= PAGE_SIZE);
254 kiov = &desc->bd_vec[desc->bd_iov_count];
261 kiov->bv_page = page;
262 kiov->bv_offset = pageoffset;
265 desc->bd_iov_count++;
267 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
269 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
273 LASSERT(desc != NULL);
274 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
275 LASSERT(desc->bd_md_count == 0); /* network hands off */
276 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
277 LASSERT(desc->bd_frag_ops != NULL);
279 sptlrpc_enc_pool_put_pages(desc);
282 class_export_put(desc->bd_export);
284 class_import_put(desc->bd_import);
286 if (desc->bd_frag_ops->release_frags != NULL)
287 desc->bd_frag_ops->release_frags(desc);
289 OBD_FREE_LARGE(desc->bd_vec,
290 desc->bd_max_iov * sizeof(*desc->bd_vec));
294 EXPORT_SYMBOL(ptlrpc_free_bulk);
297 * Set server timelimit for this req, i.e. how long are we willing to wait
298 * for reply before timing out this request.
300 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
302 LASSERT(req->rq_import);
305 /* non-AT settings */
307 * \a imp_server_timeout means this is reverse import and
308 * we send (currently only) ASTs to the client and cannot afford
309 * to wait too long for the reply, otherwise the other client
310 * (because of which we are sending this request) would
311 * timeout waiting for us
313 req->rq_timeout = req->rq_import->imp_server_timeout ?
314 obd_timeout / 2 : obd_timeout;
316 struct imp_at *at = &req->rq_import->imp_at;
320 idx = import_at_get_index(req->rq_import,
321 req->rq_request_portal);
322 serv_est = at_get(&at->iat_service_estimate[idx]);
324 * Currently a 32 bit value is sent over the
325 * wire for rq_timeout so please don't change this
326 * to time64_t. The work for LU-1158 will in time
327 * replace rq_timeout with a 64 bit nanosecond value
329 req->rq_timeout = at_est2timeout(serv_est);
332 * We could get even fancier here, using history to predict increased
335 * Let the server know what this RPC timeout is by putting it in the
338 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
340 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
342 /* Adjust max service estimate based on server value */
343 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
350 LASSERT(req->rq_import);
351 at = &req->rq_import->imp_at;
353 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
355 * max service estimates are tracked on the server side,
356 * so just keep minimal history here
358 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
361 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
362 req->rq_import->imp_obd->obd_name,
363 req->rq_request_portal,
364 oldse, at_get(&at->iat_service_estimate[idx]));
367 /* Expected network latency per remote node (secs) */
368 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
370 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
373 /* Adjust expected network latency */
374 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
375 timeout_t service_timeout)
377 time64_t now = ktime_get_real_seconds();
382 LASSERT(req->rq_import);
384 if (service_timeout > now - req->rq_sent + 3) {
386 * b=16408, however, this can also happen if early reply
387 * is lost and client RPC is expired and resent, early reply
388 * or reply of original RPC can still be fit in reply buffer
389 * of resent RPC, now client is measuring time from the
390 * resent time, but server sent back service time of original
393 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
394 D_ADAPTTO : D_WARNING,
395 "Reported service time %u > total measured time %lld\n",
396 service_timeout, now - req->rq_sent);
400 /* Network latency is total time less server processing time,
403 nl = max_t(timeout_t, now - req->rq_sent - service_timeout, 0) + 1;
404 at = &req->rq_import->imp_at;
406 oldnl = at_measured(&at->iat_net_latency, nl);
409 "The network latency for %s (nid %s) has changed from %d to %d\n",
410 req->rq_import->imp_obd->obd_name,
411 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
412 oldnl, at_get(&at->iat_net_latency));
415 static int unpack_reply(struct ptlrpc_request *req)
419 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
420 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
422 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
428 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
430 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
438 * Handle an early reply message, called with the rq_lock held.
439 * If anything goes wrong just ignore it - same as if it never happened
441 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
442 __must_hold(&req->rq_lock)
444 struct ptlrpc_request *early_req;
445 timeout_t service_timeout;
451 spin_unlock(&req->rq_lock);
453 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
455 spin_lock(&req->rq_lock);
459 rc = unpack_reply(early_req);
461 sptlrpc_cli_finish_early_reply(early_req);
462 spin_lock(&req->rq_lock);
467 * Use new timeout value just to adjust the local value for this
468 * request, don't include it into at_history. It is unclear yet why
469 * service time increased and should it be counted or skipped, e.g.
470 * that can be recovery case or some error or server, the real reply
471 * will add all new data if it is worth to add.
473 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
474 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
476 /* Network latency can be adjusted, it is pure network delays */
477 service_timeout = lustre_msg_get_service_timeout(early_req->rq_repmsg);
478 ptlrpc_at_adj_net_latency(req, service_timeout);
480 sptlrpc_cli_finish_early_reply(early_req);
482 spin_lock(&req->rq_lock);
483 olddl = req->rq_deadline;
485 * server assumes it now has rq_timeout from when the request
486 * arrived, so the client should give it at least that long.
487 * since we don't know the arrival time we'll use the original
490 req->rq_deadline = req->rq_sent + req->rq_timeout +
491 ptlrpc_at_get_net_latency(req);
493 /* The below message is checked in replay-single.sh test_65{a,b} */
494 /* The below message is checked in sanity-{gss,krb5} test_8 */
495 DEBUG_REQ(D_ADAPTTO, req,
496 "Early reply #%d, new deadline in %llds (%llds)",
498 req->rq_deadline - ktime_get_real_seconds(),
499 req->rq_deadline - olddl);
504 static struct kmem_cache *request_cache;
506 int ptlrpc_request_cache_init(void)
508 request_cache = kmem_cache_create("ptlrpc_cache",
509 sizeof(struct ptlrpc_request),
510 0, SLAB_HWCACHE_ALIGN, NULL);
511 return request_cache ? 0 : -ENOMEM;
514 void ptlrpc_request_cache_fini(void)
516 kmem_cache_destroy(request_cache);
519 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
521 struct ptlrpc_request *req;
523 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
527 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
529 OBD_SLAB_FREE_PTR(req, request_cache);
533 * Wind down request pool \a pool.
534 * Frees all requests from the pool too
536 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
538 struct list_head *l, *tmp;
539 struct ptlrpc_request *req;
541 LASSERT(pool != NULL);
543 spin_lock(&pool->prp_lock);
544 list_for_each_safe(l, tmp, &pool->prp_req_list) {
545 req = list_entry(l, struct ptlrpc_request, rq_list);
546 list_del(&req->rq_list);
547 LASSERT(req->rq_reqbuf);
548 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
549 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
550 ptlrpc_request_cache_free(req);
552 spin_unlock(&pool->prp_lock);
553 OBD_FREE(pool, sizeof(*pool));
555 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
558 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
560 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
565 while (size < pool->prp_rq_size)
568 LASSERTF(list_empty(&pool->prp_req_list) ||
569 size == pool->prp_rq_size,
570 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
571 pool->prp_rq_size, size);
573 pool->prp_rq_size = size;
574 for (i = 0; i < num_rq; i++) {
575 struct ptlrpc_request *req;
576 struct lustre_msg *msg;
578 req = ptlrpc_request_cache_alloc(GFP_NOFS);
581 OBD_ALLOC_LARGE(msg, size);
583 ptlrpc_request_cache_free(req);
586 req->rq_reqbuf = msg;
587 req->rq_reqbuf_len = size;
589 spin_lock(&pool->prp_lock);
590 list_add_tail(&req->rq_list, &pool->prp_req_list);
591 spin_unlock(&pool->prp_lock);
595 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
598 * Create and initialize new request pool with given attributes:
599 * \a num_rq - initial number of requests to create for the pool
600 * \a msgsize - maximum message size possible for requests in thid pool
601 * \a populate_pool - function to be called when more requests need to be added
603 * Returns pointer to newly created pool or NULL on error.
605 struct ptlrpc_request_pool *
606 ptlrpc_init_rq_pool(int num_rq, int msgsize,
607 int (*populate_pool)(struct ptlrpc_request_pool *, int))
609 struct ptlrpc_request_pool *pool;
616 * Request next power of two for the allocation, because internally
617 * kernel would do exactly this
619 spin_lock_init(&pool->prp_lock);
620 INIT_LIST_HEAD(&pool->prp_req_list);
621 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
622 pool->prp_populate = populate_pool;
624 populate_pool(pool, num_rq);
628 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
631 * Fetches one request from pool \a pool
633 static struct ptlrpc_request *
634 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
636 struct ptlrpc_request *request;
637 struct lustre_msg *reqbuf;
642 spin_lock(&pool->prp_lock);
645 * See if we have anything in a pool, and bail out if nothing,
646 * in writeout path, where this matters, this is safe to do, because
647 * nothing is lost in this case, and when some in-flight requests
648 * complete, this code will be called again.
650 if (unlikely(list_empty(&pool->prp_req_list))) {
651 spin_unlock(&pool->prp_lock);
655 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
657 list_del_init(&request->rq_list);
658 spin_unlock(&pool->prp_lock);
660 LASSERT(request->rq_reqbuf);
661 LASSERT(request->rq_pool);
663 reqbuf = request->rq_reqbuf;
664 memset(request, 0, sizeof(*request));
665 request->rq_reqbuf = reqbuf;
666 request->rq_reqbuf_len = pool->prp_rq_size;
667 request->rq_pool = pool;
673 * Returns freed \a request to pool.
675 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
677 struct ptlrpc_request_pool *pool = request->rq_pool;
679 spin_lock(&pool->prp_lock);
680 LASSERT(list_empty(&request->rq_list));
681 LASSERT(!request->rq_receiving_reply);
682 list_add_tail(&request->rq_list, &pool->prp_req_list);
683 spin_unlock(&pool->prp_lock);
686 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
688 struct obd_import *imp = req->rq_import;
689 struct list_head *tmp;
690 struct ptlrpc_request *iter;
692 assert_spin_locked(&imp->imp_lock);
693 LASSERT(list_empty(&req->rq_unreplied_list));
695 /* unreplied list is sorted by xid in ascending order */
696 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
697 iter = list_entry(tmp, struct ptlrpc_request,
700 LASSERT(req->rq_xid != iter->rq_xid);
701 if (req->rq_xid < iter->rq_xid)
703 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
706 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
709 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
711 req->rq_xid = ptlrpc_next_xid();
712 ptlrpc_add_unreplied(req);
715 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
717 spin_lock(&req->rq_import->imp_lock);
718 ptlrpc_assign_next_xid_nolock(req);
719 spin_unlock(&req->rq_import->imp_lock);
722 static atomic64_t ptlrpc_last_xid;
724 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
726 spin_lock(&req->rq_import->imp_lock);
727 list_del_init(&req->rq_unreplied_list);
728 ptlrpc_assign_next_xid_nolock(req);
729 spin_unlock(&req->rq_import->imp_lock);
730 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
733 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
735 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
739 opc = lustre_msg_get_opc(req->rq_reqmsg);
740 tag = obd_get_mod_rpc_slot(cli, opc);
741 lustre_msg_set_tag(req->rq_reqmsg, tag);
742 ptlrpc_reassign_next_xid(req);
744 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
746 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
748 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
751 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
752 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
754 obd_put_mod_rpc_slot(cli, opc, tag);
757 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
759 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
760 __u32 version, int opcode, char **bufs,
761 struct ptlrpc_cli_ctx *ctx)
764 struct obd_import *imp;
770 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
771 imp = request->rq_import;
772 lengths = request->rq_pill.rc_area[RCL_CLIENT];
775 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
777 rc = sptlrpc_req_get_ctx(request);
781 sptlrpc_req_set_flavor(request, opcode);
783 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
788 lustre_msg_add_version(request->rq_reqmsg, version);
789 request->rq_send_state = LUSTRE_IMP_FULL;
790 request->rq_type = PTL_RPC_MSG_REQUEST;
792 request->rq_req_cbid.cbid_fn = request_out_callback;
793 request->rq_req_cbid.cbid_arg = request;
795 request->rq_reply_cbid.cbid_fn = reply_in_callback;
796 request->rq_reply_cbid.cbid_arg = request;
798 request->rq_reply_deadline = 0;
799 request->rq_bulk_deadline = 0;
800 request->rq_req_deadline = 0;
801 request->rq_phase = RQ_PHASE_NEW;
802 request->rq_next_phase = RQ_PHASE_UNDEFINED;
804 request->rq_request_portal = imp->imp_client->cli_request_portal;
805 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
807 ptlrpc_at_set_req_timeout(request);
809 lustre_msg_set_opc(request->rq_reqmsg, opcode);
811 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
812 if (cfs_fail_val == opcode) {
813 time64_t *fail_t = NULL, *fail2_t = NULL;
815 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
816 fail_t = &request->rq_bulk_deadline;
817 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
818 fail_t = &request->rq_reply_deadline;
819 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
820 fail_t = &request->rq_req_deadline;
821 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
822 fail_t = &request->rq_reply_deadline;
823 fail2_t = &request->rq_bulk_deadline;
824 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
825 time64_t now = ktime_get_real_seconds();
826 u64 xid = ((u64)now >> 4) << 24;
828 atomic64_set(&ptlrpc_last_xid, xid);
832 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
835 *fail2_t = ktime_get_real_seconds() +
839 * The RPC is infected, let the test to change the
842 msleep(4 * MSEC_PER_SEC);
845 ptlrpc_assign_next_xid(request);
850 LASSERT(!request->rq_pool);
851 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
853 class_import_put(imp);
857 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
860 * Pack request buffers for network transfer, performing necessary encryption
861 * steps if necessary.
863 int ptlrpc_request_pack(struct ptlrpc_request *request,
864 __u32 version, int opcode)
866 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
868 EXPORT_SYMBOL(ptlrpc_request_pack);
871 * Helper function to allocate new request on import \a imp
872 * and possibly using existing request from pool \a pool if provided.
873 * Returns allocated request structure with import field filled or
877 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
878 struct ptlrpc_request_pool *pool)
880 struct ptlrpc_request *request = NULL;
882 request = ptlrpc_request_cache_alloc(GFP_NOFS);
884 if (!request && pool)
885 request = ptlrpc_prep_req_from_pool(pool);
888 ptlrpc_cli_req_init(request);
890 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
891 LASSERT(imp != LP_POISON);
892 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
894 LASSERT(imp->imp_client != LP_POISON);
896 request->rq_import = class_import_get(imp);
898 CERROR("request allocation out of memory\n");
905 * Helper function for creating a request.
906 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
907 * buffer structures according to capsule template \a format.
908 * Returns allocated request structure pointer or NULL on error.
910 static struct ptlrpc_request *
911 ptlrpc_request_alloc_internal(struct obd_import *imp,
912 struct ptlrpc_request_pool *pool,
913 const struct req_format *format)
915 struct ptlrpc_request *request;
917 request = __ptlrpc_request_alloc(imp, pool);
922 * initiate connection if needed when the import has been
923 * referenced by the new request to avoid races with disconnect
925 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
928 CDEBUG_LIMIT(imp->imp_idle_debug,
929 "%s: reconnect after %llds idle\n",
930 imp->imp_obd->obd_name, ktime_get_real_seconds() -
931 imp->imp_last_reply_time);
932 spin_lock(&imp->imp_lock);
933 if (imp->imp_state == LUSTRE_IMP_IDLE) {
934 imp->imp_generation++;
935 imp->imp_initiated_at = imp->imp_generation;
936 imp->imp_state = LUSTRE_IMP_NEW;
938 /* connect_import_locked releases imp_lock */
939 rc = ptlrpc_connect_import_locked(imp);
941 ptlrpc_request_free(request);
944 ptlrpc_pinger_add_import(imp);
946 spin_unlock(&imp->imp_lock);
950 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
951 req_capsule_set(&request->rq_pill, format);
956 * Allocate new request structure for import \a imp and initialize its
957 * buffer structure according to capsule template \a format.
959 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
960 const struct req_format *format)
962 return ptlrpc_request_alloc_internal(imp, NULL, format);
964 EXPORT_SYMBOL(ptlrpc_request_alloc);
967 * Allocate new request structure for import \a imp from pool \a pool and
968 * initialize its buffer structure according to capsule template \a format.
970 struct ptlrpc_request *
971 ptlrpc_request_alloc_pool(struct obd_import *imp,
972 struct ptlrpc_request_pool *pool,
973 const struct req_format *format)
975 return ptlrpc_request_alloc_internal(imp, pool, format);
977 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
980 * For requests not from pool, free memory of the request structure.
981 * For requests obtained from a pool earlier, return request back to pool.
983 void ptlrpc_request_free(struct ptlrpc_request *request)
985 if (request->rq_pool)
986 __ptlrpc_free_req_to_pool(request);
988 ptlrpc_request_cache_free(request);
990 EXPORT_SYMBOL(ptlrpc_request_free);
993 * Allocate new request for operatione \a opcode and immediatelly pack it for
995 * Only used for simple requests like OBD_PING where the only important
996 * part of the request is operation itself.
997 * Returns allocated request or NULL on error.
999 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1000 const struct req_format *format,
1001 __u32 version, int opcode)
1003 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1007 rc = ptlrpc_request_pack(req, version, opcode);
1009 ptlrpc_request_free(req);
1015 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1018 * Allocate and initialize new request set structure on the current CPT.
1019 * Returns a pointer to the newly allocated set structure or NULL on error.
1021 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1023 struct ptlrpc_request_set *set;
1027 cpt = cfs_cpt_current(cfs_cpt_tab, 0);
1028 OBD_CPT_ALLOC(set, cfs_cpt_tab, cpt, sizeof(*set));
1031 atomic_set(&set->set_refcount, 1);
1032 INIT_LIST_HEAD(&set->set_requests);
1033 init_waitqueue_head(&set->set_waitq);
1034 atomic_set(&set->set_new_count, 0);
1035 atomic_set(&set->set_remaining, 0);
1036 spin_lock_init(&set->set_new_req_lock);
1037 INIT_LIST_HEAD(&set->set_new_requests);
1038 set->set_max_inflight = UINT_MAX;
1039 set->set_producer = NULL;
1040 set->set_producer_arg = NULL;
1045 EXPORT_SYMBOL(ptlrpc_prep_set);
1048 * Allocate and initialize new request set structure with flow control
1049 * extension. This extension allows to control the number of requests in-flight
1050 * for the whole set. A callback function to generate requests must be provided
1051 * and the request set will keep the number of requests sent over the wire to
1053 * Returns a pointer to the newly allocated set structure or NULL on error.
1055 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1059 struct ptlrpc_request_set *set;
1061 set = ptlrpc_prep_set();
1065 set->set_max_inflight = max;
1066 set->set_producer = func;
1067 set->set_producer_arg = arg;
1073 * Wind down and free request set structure previously allocated with
1075 * Ensures that all requests on the set have completed and removes
1076 * all requests from the request list in a set.
1077 * If any unsent request happen to be on the list, pretends that they got
1078 * an error in flight and calls their completion handler.
1080 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1082 struct list_head *tmp;
1083 struct list_head *next;
1089 /* Requests on the set should either all be completed, or all be new */
1090 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1091 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1092 list_for_each(tmp, &set->set_requests) {
1093 struct ptlrpc_request *req =
1094 list_entry(tmp, struct ptlrpc_request,
1097 LASSERT(req->rq_phase == expected_phase);
1101 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1102 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1103 atomic_read(&set->set_remaining), n);
1105 list_for_each_safe(tmp, next, &set->set_requests) {
1106 struct ptlrpc_request *req =
1107 list_entry(tmp, struct ptlrpc_request,
1109 list_del_init(&req->rq_set_chain);
1111 LASSERT(req->rq_phase == expected_phase);
1113 if (req->rq_phase == RQ_PHASE_NEW) {
1114 ptlrpc_req_interpret(NULL, req, -EBADR);
1115 atomic_dec(&set->set_remaining);
1118 spin_lock(&req->rq_lock);
1120 req->rq_invalid_rqset = 0;
1121 spin_unlock(&req->rq_lock);
1123 ptlrpc_req_finished(req);
1126 LASSERT(atomic_read(&set->set_remaining) == 0);
1128 ptlrpc_reqset_put(set);
1131 EXPORT_SYMBOL(ptlrpc_set_destroy);
1134 * Add a new request to the general purpose request set.
1135 * Assumes request reference from the caller.
1137 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1138 struct ptlrpc_request *req)
1140 if (set == PTLRPCD_SET) {
1141 ptlrpcd_add_req(req);
1145 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1146 LASSERT(list_empty(&req->rq_set_chain));
1148 if (req->rq_allow_intr)
1149 set->set_allow_intr = 1;
1151 /* The set takes over the caller's request reference */
1152 list_add_tail(&req->rq_set_chain, &set->set_requests);
1154 atomic_inc(&set->set_remaining);
1155 req->rq_queued_time = ktime_get_seconds();
1158 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1160 if (set->set_producer)
1162 * If the request set has a producer callback, the RPC must be
1163 * sent straight away
1165 ptlrpc_send_new_req(req);
1167 EXPORT_SYMBOL(ptlrpc_set_add_req);
1170 * Add a request to a request with dedicated server thread
1171 * and wake the thread to make any necessary processing.
1172 * Currently only used for ptlrpcd.
1174 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1175 struct ptlrpc_request *req)
1177 struct ptlrpc_request_set *set = pc->pc_set;
1180 LASSERT(req->rq_set == NULL);
1181 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1183 spin_lock(&set->set_new_req_lock);
1185 * The set takes over the caller's request reference.
1188 req->rq_queued_time = ktime_get_seconds();
1189 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1190 count = atomic_inc_return(&set->set_new_count);
1191 spin_unlock(&set->set_new_req_lock);
1193 /* Only need to call wakeup once for the first entry. */
1195 wake_up(&set->set_waitq);
1198 * XXX: It maybe unnecessary to wakeup all the partners. But to
1199 * guarantee the async RPC can be processed ASAP, we have
1200 * no other better choice. It maybe fixed in future.
1202 for (i = 0; i < pc->pc_npartners; i++)
1203 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1208 * Based on the current state of the import, determine if the request
1209 * can be sent, is an error, or should be delayed.
1211 * Returns true if this request should be delayed. If false, and
1212 * *status is set, then the request can not be sent and *status is the
1213 * error code. If false and status is 0, then request can be sent.
1215 * The imp->imp_lock must be held.
1217 static int ptlrpc_import_delay_req(struct obd_import *imp,
1218 struct ptlrpc_request *req, int *status)
1226 if (req->rq_ctx_init || req->rq_ctx_fini) {
1227 /* always allow ctx init/fini rpc go through */
1228 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1229 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1231 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1232 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1235 * pings or MDS-equivalent STATFS may safely
1238 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1239 D_HA : D_ERROR, req, "IMP_CLOSED");
1241 } else if (ptlrpc_send_limit_expired(req)) {
1242 /* probably doesn't need to be a D_ERROR afterinitial testing */
1243 DEBUG_REQ(D_HA, req, "send limit expired");
1244 *status = -ETIMEDOUT;
1245 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1246 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1247 ;/* allow CONNECT even if import is invalid */
1248 if (atomic_read(&imp->imp_inval_count) != 0) {
1249 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1252 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1253 if (!imp->imp_deactive)
1254 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1255 *status = -ESHUTDOWN; /* b=12940 */
1256 } else if (req->rq_import_generation != imp->imp_generation) {
1257 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1259 } else if (req->rq_send_state != imp->imp_state) {
1260 /* invalidate in progress - any requests should be drop */
1261 if (atomic_read(&imp->imp_inval_count) != 0) {
1262 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1264 } else if (req->rq_no_delay &&
1265 imp->imp_generation != imp->imp_initiated_at) {
1266 /* ignore nodelay for requests initiating connections */
1267 *status = -EWOULDBLOCK;
1268 } else if (req->rq_allow_replay &&
1269 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1270 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1271 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1272 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1273 DEBUG_REQ(D_HA, req, "allow during recovery");
1283 * Decide if the error message should be printed to the console or not.
1284 * Makes its decision based on request type, status, and failure frequency.
1286 * \param[in] req request that failed and may need a console message
1288 * \retval false if no message should be printed
1289 * \retval true if console message should be printed
1291 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1293 LASSERT(req->rq_reqmsg != NULL);
1295 /* Suppress particular reconnect errors which are to be expected. */
1296 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1297 /* Suppress timed out reconnect requests */
1298 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1303 * Suppress most unavailable/again reconnect requests, but
1304 * print occasionally so it is clear client is trying to
1305 * connect to a server where no target is running.
1307 if ((err == -ENODEV || err == -EAGAIN) &&
1308 req->rq_import->imp_conn_cnt % 30 != 20)
1312 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1313 /* -EAGAIN is normal when using POSIX flocks */
1316 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1317 (req->rq_xid & 0xf) != 10)
1318 /* Suppress most ping requests, they may fail occasionally */
1325 * Check request processing status.
1326 * Returns the status.
1328 static int ptlrpc_check_status(struct ptlrpc_request *req)
1333 rc = lustre_msg_get_status(req->rq_repmsg);
1334 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1335 struct obd_import *imp = req->rq_import;
1336 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1337 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1339 if (ptlrpc_console_allow(req, opc, rc))
1340 LCONSOLE_ERROR_MSG(0x11,
1341 "%s: operation %s to node %s failed: rc = %d\n",
1342 imp->imp_obd->obd_name,
1344 libcfs_nid2str(nid), rc);
1345 RETURN(rc < 0 ? rc : -EINVAL);
1349 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1355 * save pre-versions of objects into request for replay.
1356 * Versions are obtained from server reply.
1359 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1361 struct lustre_msg *repmsg = req->rq_repmsg;
1362 struct lustre_msg *reqmsg = req->rq_reqmsg;
1363 __u64 *versions = lustre_msg_get_versions(repmsg);
1366 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1370 lustre_msg_set_versions(reqmsg, versions);
1371 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1372 versions[0], versions[1]);
1377 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1379 struct ptlrpc_request *req;
1381 assert_spin_locked(&imp->imp_lock);
1382 if (list_empty(&imp->imp_unreplied_list))
1385 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1387 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1389 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1390 imp->imp_known_replied_xid = req->rq_xid - 1;
1392 return req->rq_xid - 1;
1396 * Callback function called when client receives RPC reply for \a req.
1397 * Returns 0 on success or error code.
1398 * The return alue would be assigned to req->rq_status by the caller
1399 * as request processing status.
1400 * This function also decides if the request needs to be saved for later replay.
1402 static int after_reply(struct ptlrpc_request *req)
1404 struct obd_import *imp = req->rq_import;
1405 struct obd_device *obd = req->rq_import->imp_obd;
1412 LASSERT(obd != NULL);
1413 /* repbuf must be unlinked */
1414 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1416 if (req->rq_reply_truncated) {
1417 if (ptlrpc_no_resend(req)) {
1418 DEBUG_REQ(D_ERROR, req,
1419 "reply buffer overflow, expected=%d, actual size=%d",
1420 req->rq_nob_received, req->rq_repbuf_len);
1424 sptlrpc_cli_free_repbuf(req);
1426 * Pass the required reply buffer size (include
1427 * space for early reply).
1428 * NB: no need to roundup because alloc_repbuf
1431 req->rq_replen = req->rq_nob_received;
1432 req->rq_nob_received = 0;
1433 spin_lock(&req->rq_lock);
1435 spin_unlock(&req->rq_lock);
1439 work_start = ktime_get_real();
1440 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1443 * NB Until this point, the whole of the incoming message,
1444 * including buflens, status etc is in the sender's byte order.
1446 rc = sptlrpc_cli_unwrap_reply(req);
1448 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1453 * Security layer unwrap might ask resend this request.
1458 rc = unpack_reply(req);
1462 /* retry indefinitely on EINPROGRESS */
1463 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1464 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1465 time64_t now = ktime_get_real_seconds();
1467 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1468 D_RPCTRACE, req, "resending request on EINPROGRESS");
1469 spin_lock(&req->rq_lock);
1471 spin_unlock(&req->rq_lock);
1472 req->rq_nr_resend++;
1474 /* Readjust the timeout for current conditions */
1475 ptlrpc_at_set_req_timeout(req);
1477 * delay resend to give a chance to the server to get ready.
1478 * The delay is increased by 1s on every resend and is capped to
1479 * the current request timeout (i.e. obd_timeout if AT is off,
1480 * or AT service time x 125% + 5s, see at_est2timeout)
1482 if (req->rq_nr_resend > req->rq_timeout)
1483 req->rq_sent = now + req->rq_timeout;
1485 req->rq_sent = now + req->rq_nr_resend;
1487 /* Resend for EINPROGRESS will use a new XID */
1488 spin_lock(&imp->imp_lock);
1489 list_del_init(&req->rq_unreplied_list);
1490 spin_unlock(&imp->imp_lock);
1495 if (obd->obd_svc_stats) {
1496 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1498 ptlrpc_lprocfs_rpc_sent(req, timediff);
1501 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1502 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1503 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1504 lustre_msg_get_type(req->rq_repmsg));
1508 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1509 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1510 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1511 ptlrpc_at_adj_net_latency(req,
1512 lustre_msg_get_service_timeout(req->rq_repmsg));
1514 rc = ptlrpc_check_status(req);
1518 * Either we've been evicted, or the server has failed for
1519 * some reason. Try to reconnect, and if that fails, punt to
1522 if (ptlrpc_recoverable_error(rc)) {
1523 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1524 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1527 ptlrpc_request_handle_notconn(req);
1532 * Let's look if server sent slv. Do it only for RPC with
1535 ldlm_cli_update_pool(req);
1539 * Store transno in reqmsg for replay.
1541 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1542 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1543 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1546 if (imp->imp_replayable) {
1547 spin_lock(&imp->imp_lock);
1549 * No point in adding already-committed requests to the replay
1550 * list, we will just remove them immediately. b=9829
1552 if (req->rq_transno != 0 &&
1554 lustre_msg_get_last_committed(req->rq_repmsg) ||
1556 /** version recovery */
1557 ptlrpc_save_versions(req);
1558 ptlrpc_retain_replayable_request(req, imp);
1559 } else if (req->rq_commit_cb &&
1560 list_empty(&req->rq_replay_list)) {
1562 * NB: don't call rq_commit_cb if it's already on
1563 * rq_replay_list, ptlrpc_free_committed() will call
1564 * it later, see LU-3618 for details
1566 spin_unlock(&imp->imp_lock);
1567 req->rq_commit_cb(req);
1568 spin_lock(&imp->imp_lock);
1572 * Replay-enabled imports return commit-status information.
1574 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1575 if (likely(committed > imp->imp_peer_committed_transno))
1576 imp->imp_peer_committed_transno = committed;
1578 ptlrpc_free_committed(imp);
1580 if (!list_empty(&imp->imp_replay_list)) {
1581 struct ptlrpc_request *last;
1583 last = list_entry(imp->imp_replay_list.prev,
1584 struct ptlrpc_request,
1587 * Requests with rq_replay stay on the list even if no
1588 * commit is expected.
1590 if (last->rq_transno > imp->imp_peer_committed_transno)
1591 ptlrpc_pinger_commit_expected(imp);
1594 spin_unlock(&imp->imp_lock);
1601 * Helper function to send request \a req over the network for the first time
1602 * Also adjusts request phase.
1603 * Returns 0 on success or error code.
1605 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1607 struct obd_import *imp = req->rq_import;
1612 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1614 /* do not try to go further if there is not enough memory in enc_pool */
1615 if (req->rq_sent && req->rq_bulk)
1616 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1617 pool_is_at_full_capacity())
1620 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1621 (!req->rq_generation_set ||
1622 req->rq_import_generation == imp->imp_generation))
1625 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1627 spin_lock(&imp->imp_lock);
1629 LASSERT(req->rq_xid != 0);
1630 LASSERT(!list_empty(&req->rq_unreplied_list));
1632 if (!req->rq_generation_set)
1633 req->rq_import_generation = imp->imp_generation;
1635 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1636 spin_lock(&req->rq_lock);
1637 req->rq_waiting = 1;
1638 spin_unlock(&req->rq_lock);
1640 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1641 ptlrpc_import_state_name(req->rq_send_state),
1642 ptlrpc_import_state_name(imp->imp_state));
1643 LASSERT(list_empty(&req->rq_list));
1644 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1645 atomic_inc(&req->rq_import->imp_inflight);
1646 spin_unlock(&imp->imp_lock);
1651 spin_unlock(&imp->imp_lock);
1652 req->rq_status = rc;
1653 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1657 LASSERT(list_empty(&req->rq_list));
1658 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1659 atomic_inc(&req->rq_import->imp_inflight);
1662 * find the known replied XID from the unreplied list, CONNECT
1663 * and DISCONNECT requests are skipped to make the sanity check
1664 * on server side happy. see process_req_last_xid().
1666 * For CONNECT: Because replay requests have lower XID, it'll
1667 * break the sanity check if CONNECT bump the exp_last_xid on
1670 * For DISCONNECT: Since client will abort inflight RPC before
1671 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1672 * than the inflight RPC.
1674 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1675 min_xid = ptlrpc_known_replied_xid(imp);
1676 spin_unlock(&imp->imp_lock);
1678 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1680 lustre_msg_set_status(req->rq_reqmsg, current->pid);
1682 rc = sptlrpc_req_refresh_ctx(req, 0);
1685 req->rq_status = rc;
1688 spin_lock(&req->rq_lock);
1689 req->rq_wait_ctx = 1;
1690 spin_unlock(&req->rq_lock);
1696 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1698 imp->imp_obd->obd_uuid.uuid,
1699 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1700 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1701 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1703 rc = ptl_send_rpc(req, 0);
1704 if (rc == -ENOMEM) {
1705 spin_lock(&imp->imp_lock);
1706 if (!list_empty(&req->rq_list)) {
1707 list_del_init(&req->rq_list);
1708 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1709 wake_up(&req->rq_import->imp_recovery_waitq);
1711 spin_unlock(&imp->imp_lock);
1712 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1716 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1718 spin_lock(&req->rq_lock);
1719 req->rq_net_err = 1;
1720 spin_unlock(&req->rq_lock);
1726 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1731 LASSERT(set->set_producer != NULL);
1733 remaining = atomic_read(&set->set_remaining);
1736 * populate the ->set_requests list with requests until we
1737 * reach the maximum number of RPCs in flight for this set
1739 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1740 rc = set->set_producer(set, set->set_producer_arg);
1741 if (rc == -ENOENT) {
1742 /* no more RPC to produce */
1743 set->set_producer = NULL;
1744 set->set_producer_arg = NULL;
1749 RETURN((atomic_read(&set->set_remaining) - remaining));
1753 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1754 * and no more replies are expected.
1755 * (it is possible to get less replies than requests sent e.g. due to timed out
1756 * requests or requests that we had trouble to send out)
1758 * NOTE: This function contains a potential schedule point (cond_resched()).
1760 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1762 struct list_head *tmp, *next;
1763 LIST_HEAD(comp_reqs);
1764 int force_timer_recalc = 0;
1767 if (atomic_read(&set->set_remaining) == 0)
1770 list_for_each_safe(tmp, next, &set->set_requests) {
1771 struct ptlrpc_request *req =
1772 list_entry(tmp, struct ptlrpc_request,
1774 struct obd_import *imp = req->rq_import;
1775 int unregistered = 0;
1779 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1780 list_move_tail(&req->rq_set_chain, &comp_reqs);
1785 * This schedule point is mainly for the ptlrpcd caller of this
1786 * function. Most ptlrpc sets are not long-lived and unbounded
1787 * in length, but at the least the set used by the ptlrpcd is.
1788 * Since the processing time is unbounded, we need to insert an
1789 * explicit schedule point to make the thread well-behaved.
1794 * If the caller requires to allow to be interpreted by force
1795 * and it has really been interpreted, then move the request
1796 * to RQ_PHASE_INTERPRET phase in spite of what the current
1799 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1800 req->rq_status = -EINTR;
1801 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1804 * Since it is interpreted and we have to wait for
1805 * the reply to be unlinked, then use sync mode.
1809 GOTO(interpret, req->rq_status);
1812 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1813 force_timer_recalc = 1;
1815 /* delayed send - skip */
1816 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1819 /* delayed resend - skip */
1820 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1821 req->rq_sent > ktime_get_real_seconds())
1824 if (!(req->rq_phase == RQ_PHASE_RPC ||
1825 req->rq_phase == RQ_PHASE_BULK ||
1826 req->rq_phase == RQ_PHASE_INTERPRET ||
1827 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1828 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1829 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1833 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1834 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1835 LASSERT(req->rq_next_phase != req->rq_phase);
1836 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1838 if (req->rq_req_deadline &&
1839 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1840 req->rq_req_deadline = 0;
1841 if (req->rq_reply_deadline &&
1842 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1843 req->rq_reply_deadline = 0;
1844 if (req->rq_bulk_deadline &&
1845 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1846 req->rq_bulk_deadline = 0;
1849 * Skip processing until reply is unlinked. We
1850 * can't return to pool before that and we can't
1851 * call interpret before that. We need to make
1852 * sure that all rdma transfers finished and will
1853 * not corrupt any data.
1855 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1856 ptlrpc_client_recv_or_unlink(req))
1858 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1859 ptlrpc_client_bulk_active(req))
1863 * Turn fail_loc off to prevent it from looping
1866 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1867 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1870 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1871 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1876 * Move to next phase if reply was successfully
1879 ptlrpc_rqphase_move(req, req->rq_next_phase);
1882 if (req->rq_phase == RQ_PHASE_INTERPRET)
1883 GOTO(interpret, req->rq_status);
1886 * Note that this also will start async reply unlink.
1888 if (req->rq_net_err && !req->rq_timedout) {
1889 ptlrpc_expire_one_request(req, 1);
1892 * Check if we still need to wait for unlink.
1894 if (ptlrpc_client_recv_or_unlink(req) ||
1895 ptlrpc_client_bulk_active(req))
1897 /* If there is no need to resend, fail it now. */
1898 if (req->rq_no_resend) {
1899 if (req->rq_status == 0)
1900 req->rq_status = -EIO;
1901 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1902 GOTO(interpret, req->rq_status);
1909 spin_lock(&req->rq_lock);
1910 req->rq_replied = 0;
1911 spin_unlock(&req->rq_lock);
1912 if (req->rq_status == 0)
1913 req->rq_status = -EIO;
1914 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1915 GOTO(interpret, req->rq_status);
1919 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1920 * so it sets rq_intr regardless of individual rpc
1921 * timeouts. The synchronous IO waiting path sets
1922 * rq_intr irrespective of whether ptlrpcd
1923 * has seen a timeout. Our policy is to only interpret
1924 * interrupted rpcs after they have timed out, so we
1925 * need to enforce that here.
1928 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1929 req->rq_wait_ctx)) {
1930 req->rq_status = -EINTR;
1931 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1932 GOTO(interpret, req->rq_status);
1935 if (req->rq_phase == RQ_PHASE_RPC) {
1936 if (req->rq_timedout || req->rq_resend ||
1937 req->rq_waiting || req->rq_wait_ctx) {
1940 if (!ptlrpc_unregister_reply(req, 1)) {
1941 ptlrpc_unregister_bulk(req, 1);
1945 spin_lock(&imp->imp_lock);
1946 if (ptlrpc_import_delay_req(imp, req,
1949 * put on delay list - only if we wait
1950 * recovery finished - before send
1952 list_move_tail(&req->rq_list,
1953 &imp->imp_delayed_list);
1954 spin_unlock(&imp->imp_lock);
1959 req->rq_status = status;
1960 ptlrpc_rqphase_move(req,
1961 RQ_PHASE_INTERPRET);
1962 spin_unlock(&imp->imp_lock);
1963 GOTO(interpret, req->rq_status);
1965 /* ignore on just initiated connections */
1966 if (ptlrpc_no_resend(req) &&
1967 !req->rq_wait_ctx &&
1968 imp->imp_generation !=
1969 imp->imp_initiated_at) {
1970 req->rq_status = -ENOTCONN;
1971 ptlrpc_rqphase_move(req,
1972 RQ_PHASE_INTERPRET);
1973 spin_unlock(&imp->imp_lock);
1974 GOTO(interpret, req->rq_status);
1977 list_move_tail(&req->rq_list,
1978 &imp->imp_sending_list);
1980 spin_unlock(&imp->imp_lock);
1982 spin_lock(&req->rq_lock);
1983 req->rq_waiting = 0;
1984 spin_unlock(&req->rq_lock);
1986 if (req->rq_timedout || req->rq_resend) {
1988 * This is re-sending anyways,
1989 * let's mark req as resend.
1991 spin_lock(&req->rq_lock);
1993 spin_unlock(&req->rq_lock);
1996 * rq_wait_ctx is only touched by ptlrpcd,
1997 * so no lock is needed here.
1999 status = sptlrpc_req_refresh_ctx(req, 0);
2002 req->rq_status = status;
2003 spin_lock(&req->rq_lock);
2004 req->rq_wait_ctx = 0;
2005 spin_unlock(&req->rq_lock);
2006 force_timer_recalc = 1;
2008 spin_lock(&req->rq_lock);
2009 req->rq_wait_ctx = 1;
2010 spin_unlock(&req->rq_lock);
2015 spin_lock(&req->rq_lock);
2016 req->rq_wait_ctx = 0;
2017 spin_unlock(&req->rq_lock);
2021 * In any case, the previous bulk should be
2022 * cleaned up to prepare for the new sending
2025 !ptlrpc_unregister_bulk(req, 1))
2028 rc = ptl_send_rpc(req, 0);
2029 if (rc == -ENOMEM) {
2030 spin_lock(&imp->imp_lock);
2031 if (!list_empty(&req->rq_list))
2032 list_del_init(&req->rq_list);
2033 spin_unlock(&imp->imp_lock);
2034 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2038 DEBUG_REQ(D_HA, req,
2039 "send failed: rc = %d", rc);
2040 force_timer_recalc = 1;
2041 spin_lock(&req->rq_lock);
2042 req->rq_net_err = 1;
2043 spin_unlock(&req->rq_lock);
2046 /* need to reset the timeout */
2047 force_timer_recalc = 1;
2050 spin_lock(&req->rq_lock);
2052 if (ptlrpc_client_early(req)) {
2053 ptlrpc_at_recv_early_reply(req);
2054 spin_unlock(&req->rq_lock);
2058 /* Still waiting for a reply? */
2059 if (ptlrpc_client_recv(req)) {
2060 spin_unlock(&req->rq_lock);
2064 /* Did we actually receive a reply? */
2065 if (!ptlrpc_client_replied(req)) {
2066 spin_unlock(&req->rq_lock);
2070 spin_unlock(&req->rq_lock);
2073 * unlink from net because we are going to
2074 * swab in-place of reply buffer
2076 unregistered = ptlrpc_unregister_reply(req, 1);
2080 req->rq_status = after_reply(req);
2085 * If there is no bulk associated with this request,
2086 * then we're done and should let the interpreter
2087 * process the reply. Similarly if the RPC returned
2088 * an error, and therefore the bulk will never arrive.
2090 if (!req->rq_bulk || req->rq_status < 0) {
2091 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2092 GOTO(interpret, req->rq_status);
2095 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2098 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2099 if (ptlrpc_client_bulk_active(req))
2102 if (req->rq_bulk->bd_failure) {
2104 * The RPC reply arrived OK, but the bulk screwed
2105 * up! Dead weird since the server told us the RPC
2106 * was good after getting the REPLY for her GET or
2107 * the ACK for her PUT.
2109 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2110 req->rq_status = -EIO;
2113 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2116 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2119 * This moves to "unregistering" phase we need to wait for
2122 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2123 /* start async bulk unlink too */
2124 ptlrpc_unregister_bulk(req, 1);
2128 if (!ptlrpc_unregister_bulk(req, async))
2132 * When calling interpret receiving already should be
2135 LASSERT(!req->rq_receiving_reply);
2137 ptlrpc_req_interpret(env, req, req->rq_status);
2139 if (ptlrpcd_check_work(req)) {
2140 atomic_dec(&set->set_remaining);
2143 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2147 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2149 imp->imp_obd->obd_uuid.uuid,
2150 lustre_msg_get_status(req->rq_reqmsg),
2152 obd_import_nid2str(imp),
2153 lustre_msg_get_opc(req->rq_reqmsg),
2154 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2156 spin_lock(&imp->imp_lock);
2158 * Request already may be not on sending or delaying list. This
2159 * may happen in the case of marking it erroneous for the case
2160 * ptlrpc_import_delay_req(req, status) find it impossible to
2161 * allow sending this rpc and returns *status != 0.
2163 if (!list_empty(&req->rq_list)) {
2164 list_del_init(&req->rq_list);
2165 if (atomic_dec_and_test(&imp->imp_inflight))
2166 wake_up(&imp->imp_recovery_waitq);
2168 list_del_init(&req->rq_unreplied_list);
2169 spin_unlock(&imp->imp_lock);
2171 atomic_dec(&set->set_remaining);
2172 wake_up(&imp->imp_recovery_waitq);
2174 if (set->set_producer) {
2175 /* produce a new request if possible */
2176 if (ptlrpc_set_producer(set) > 0)
2177 force_timer_recalc = 1;
2180 * free the request that has just been completed
2181 * in order not to pollute set->set_requests
2183 list_del_init(&req->rq_set_chain);
2184 spin_lock(&req->rq_lock);
2186 req->rq_invalid_rqset = 0;
2187 spin_unlock(&req->rq_lock);
2189 /* record rq_status to compute the final status later */
2190 if (req->rq_status != 0)
2191 set->set_rc = req->rq_status;
2192 ptlrpc_req_finished(req);
2194 list_move_tail(&req->rq_set_chain, &comp_reqs);
2199 * move completed request at the head of list so it's easier for
2200 * caller to find them
2202 list_splice(&comp_reqs, &set->set_requests);
2204 /* If we hit an error, we want to recover promptly. */
2205 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2207 EXPORT_SYMBOL(ptlrpc_check_set);
2210 * Time out request \a req. is \a async_unlink is set, that means do not wait
2211 * until LNet actually confirms network buffer unlinking.
2212 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2214 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2216 struct obd_import *imp = req->rq_import;
2217 unsigned int debug_mask = D_RPCTRACE;
2221 spin_lock(&req->rq_lock);
2222 req->rq_timedout = 1;
2223 spin_unlock(&req->rq_lock);
2225 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2226 lustre_msg_get_status(req->rq_reqmsg)))
2227 debug_mask = D_WARNING;
2228 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2229 req->rq_net_err ? "failed due to network error" :
2230 ((req->rq_real_sent == 0 ||
2231 req->rq_real_sent < req->rq_sent ||
2232 req->rq_real_sent >= req->rq_deadline) ?
2233 "timed out for sent delay" : "timed out for slow reply"),
2234 req->rq_sent, req->rq_real_sent);
2236 if (imp && obd_debug_peer_on_timeout)
2237 LNetDebugPeer(imp->imp_connection->c_peer);
2239 ptlrpc_unregister_reply(req, async_unlink);
2240 ptlrpc_unregister_bulk(req, async_unlink);
2242 if (obd_dump_on_timeout)
2243 libcfs_debug_dumplog();
2246 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2250 atomic_inc(&imp->imp_timeouts);
2252 /* The DLM server doesn't want recovery run on its imports. */
2253 if (imp->imp_dlm_fake)
2257 * If this request is for recovery or other primordial tasks,
2258 * then error it out here.
2260 if (req->rq_ctx_init || req->rq_ctx_fini ||
2261 req->rq_send_state != LUSTRE_IMP_FULL ||
2262 imp->imp_obd->obd_no_recov) {
2263 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2264 ptlrpc_import_state_name(req->rq_send_state),
2265 ptlrpc_import_state_name(imp->imp_state));
2266 spin_lock(&req->rq_lock);
2267 req->rq_status = -ETIMEDOUT;
2269 spin_unlock(&req->rq_lock);
2274 * if a request can't be resent we can't wait for an answer after
2277 if (ptlrpc_no_resend(req)) {
2278 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2282 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2288 * Time out all uncompleted requests in request set pointed by \a data
2289 * This is called when a wait times out.
2291 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2293 struct list_head *tmp;
2294 time64_t now = ktime_get_real_seconds();
2297 LASSERT(set != NULL);
2300 * A timeout expired. See which reqs it applies to...
2302 list_for_each(tmp, &set->set_requests) {
2303 struct ptlrpc_request *req =
2304 list_entry(tmp, struct ptlrpc_request,
2307 /* don't expire request waiting for context */
2308 if (req->rq_wait_ctx)
2311 /* Request in-flight? */
2312 if (!((req->rq_phase == RQ_PHASE_RPC &&
2313 !req->rq_waiting && !req->rq_resend) ||
2314 (req->rq_phase == RQ_PHASE_BULK)))
2317 if (req->rq_timedout || /* already dealt with */
2318 req->rq_deadline > now) /* not expired */
2322 * Deal with this guy. Do it asynchronously to not block
2325 ptlrpc_expire_one_request(req, 1);
2330 * Interrupts (sets interrupted flag) all uncompleted requests in
2331 * a set \a data. This is called when a wait_event is interrupted
2334 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2336 struct list_head *tmp;
2338 LASSERT(set != NULL);
2339 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2341 list_for_each(tmp, &set->set_requests) {
2342 struct ptlrpc_request *req =
2343 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2348 if (req->rq_phase != RQ_PHASE_RPC &&
2349 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2350 !req->rq_allow_intr)
2353 spin_lock(&req->rq_lock);
2355 spin_unlock(&req->rq_lock);
2360 * Get the smallest timeout in the set; this does NOT set a timeout.
2362 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2364 struct list_head *tmp;
2365 time64_t now = ktime_get_real_seconds();
2367 struct ptlrpc_request *req;
2371 list_for_each(tmp, &set->set_requests) {
2372 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2374 /* Request in-flight? */
2375 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2376 (req->rq_phase == RQ_PHASE_BULK) ||
2377 (req->rq_phase == RQ_PHASE_NEW)))
2380 /* Already timed out. */
2381 if (req->rq_timedout)
2384 /* Waiting for ctx. */
2385 if (req->rq_wait_ctx)
2388 if (req->rq_phase == RQ_PHASE_NEW)
2389 deadline = req->rq_sent;
2390 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2391 deadline = req->rq_sent;
2393 deadline = req->rq_sent + req->rq_timeout;
2395 if (deadline <= now) /* actually expired already */
2396 timeout = 1; /* ASAP */
2397 else if (timeout == 0 || timeout > deadline - now)
2398 timeout = deadline - now;
2404 * Send all unset request from the set and then wait untill all
2405 * requests in the set complete (either get a reply, timeout, get an
2406 * error or otherwise be interrupted).
2407 * Returns 0 on success or error code otherwise.
2409 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2411 struct list_head *tmp;
2412 struct ptlrpc_request *req;
2417 if (set->set_producer)
2418 (void)ptlrpc_set_producer(set);
2420 list_for_each(tmp, &set->set_requests) {
2421 req = list_entry(tmp, struct ptlrpc_request,
2423 if (req->rq_phase == RQ_PHASE_NEW)
2424 (void)ptlrpc_send_new_req(req);
2427 if (list_empty(&set->set_requests))
2431 timeout = ptlrpc_set_next_timeout(set);
2434 * wait until all complete, interrupted, or an in-flight
2437 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2440 if ((timeout == 0 && !signal_pending(current)) ||
2441 set->set_allow_intr) {
2443 * No requests are in-flight (ether timed out
2444 * or delayed), so we can allow interrupts.
2445 * We still want to block for a limited time,
2446 * so we allow interrupts during the timeout.
2448 rc = l_wait_event_abortable_timeout(
2450 ptlrpc_check_set(NULL, set),
2451 cfs_time_seconds(timeout ? timeout : 1));
2454 ptlrpc_expired_set(set);
2455 } else if (rc < 0) {
2457 ptlrpc_interrupted_set(set);
2463 * At least one request is in flight, so no
2464 * interrupts are allowed. Wait until all
2465 * complete, or an in-flight req times out.
2467 rc = wait_event_idle_timeout(
2469 ptlrpc_check_set(NULL, set),
2470 cfs_time_seconds(timeout ? timeout : 1));
2472 ptlrpc_expired_set(set);
2479 * LU-769 - if we ignored the signal because
2480 * it was already pending when we started, we
2481 * need to handle it now or we risk it being
2484 if (rc == -ETIMEDOUT &&
2485 signal_pending(current)) {
2486 sigset_t blocked_sigs;
2488 cfs_block_sigsinv(LUSTRE_FATAL_SIGS,
2491 * In fact we only interrupt for the
2492 * "fatal" signals like SIGINT or
2493 * SIGKILL. We still ignore less
2494 * important signals since ptlrpc set
2495 * is not easily reentrant from
2498 if (signal_pending(current))
2499 ptlrpc_interrupted_set(set);
2500 cfs_restore_sigs(&blocked_sigs);
2504 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2507 * -EINTR => all requests have been flagged rq_intr so next
2509 * -ETIMEDOUT => someone timed out. When all reqs have
2510 * timed out, signals are enabled allowing completion with
2512 * I don't really care if we go once more round the loop in
2513 * the error cases -eeb.
2515 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2516 list_for_each(tmp, &set->set_requests) {
2517 req = list_entry(tmp, struct ptlrpc_request,
2519 spin_lock(&req->rq_lock);
2520 req->rq_invalid_rqset = 1;
2521 spin_unlock(&req->rq_lock);
2524 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2526 LASSERT(atomic_read(&set->set_remaining) == 0);
2528 rc = set->set_rc; /* rq_status of already freed requests if any */
2529 list_for_each(tmp, &set->set_requests) {
2530 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2532 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2533 if (req->rq_status != 0)
2534 rc = req->rq_status;
2539 EXPORT_SYMBOL(ptlrpc_set_wait);
2542 * Helper fuction for request freeing.
2543 * Called when request count reached zero and request needs to be freed.
2544 * Removes request from all sorts of sending/replay lists it might be on,
2545 * frees network buffers if any are present.
2546 * If \a locked is set, that means caller is already holding import imp_lock
2547 * and so we no longer need to reobtain it (for certain lists manipulations)
2549 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2556 LASSERT(!request->rq_srv_req);
2557 LASSERT(request->rq_export == NULL);
2558 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2559 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2560 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2561 LASSERTF(!request->rq_replay, "req %p\n", request);
2563 req_capsule_fini(&request->rq_pill);
2566 * We must take it off the imp_replay_list first. Otherwise, we'll set
2567 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2569 if (request->rq_import) {
2571 spin_lock(&request->rq_import->imp_lock);
2572 list_del_init(&request->rq_replay_list);
2573 list_del_init(&request->rq_unreplied_list);
2575 spin_unlock(&request->rq_import->imp_lock);
2577 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2579 if (atomic_read(&request->rq_refcount) != 0) {
2580 DEBUG_REQ(D_ERROR, request,
2581 "freeing request with nonzero refcount");
2585 if (request->rq_repbuf)
2586 sptlrpc_cli_free_repbuf(request);
2588 if (request->rq_import) {
2589 class_import_put(request->rq_import);
2590 request->rq_import = NULL;
2592 if (request->rq_bulk)
2593 ptlrpc_free_bulk(request->rq_bulk);
2595 if (request->rq_reqbuf || request->rq_clrbuf)
2596 sptlrpc_cli_free_reqbuf(request);
2598 if (request->rq_cli_ctx)
2599 sptlrpc_req_put_ctx(request, !locked);
2601 if (request->rq_pool)
2602 __ptlrpc_free_req_to_pool(request);
2604 ptlrpc_request_cache_free(request);
2608 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2610 * Drop one request reference. Must be called with import imp_lock held.
2611 * When reference count drops to zero, request is freed.
2613 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2615 assert_spin_locked(&request->rq_import->imp_lock);
2616 (void)__ptlrpc_req_finished(request, 1);
2621 * Drops one reference count for request \a request.
2622 * \a locked set indicates that caller holds import imp_lock.
2623 * Frees the request whe reference count reaches zero.
2625 * \retval 1 the request is freed
2626 * \retval 0 some others still hold references on the request
2628 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2636 LASSERT(request != LP_POISON);
2637 LASSERT(request->rq_reqmsg != LP_POISON);
2639 DEBUG_REQ(D_INFO, request, "refcount now %u",
2640 atomic_read(&request->rq_refcount) - 1);
2642 spin_lock(&request->rq_lock);
2643 count = atomic_dec_return(&request->rq_refcount);
2644 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2647 * For open RPC, the client does not know the EA size (LOV, ACL, and
2648 * so on) before replied, then the client has to reserve very large
2649 * reply buffer. Such buffer will not be released until the RPC freed.
2650 * Since The open RPC is replayable, we need to keep it in the replay
2651 * list until close. If there are a lot of files opened concurrently,
2652 * then the client may be OOM.
2654 * If fact, it is unnecessary to keep reply buffer for open replay,
2655 * related EAs have already been saved via mdc_save_lovea() before
2656 * coming here. So it is safe to free the reply buffer some earlier
2657 * before releasing the RPC to avoid client OOM. LU-9514
2659 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2660 spin_lock(&request->rq_early_free_lock);
2661 sptlrpc_cli_free_repbuf(request);
2662 request->rq_repbuf = NULL;
2663 request->rq_repbuf_len = 0;
2664 request->rq_repdata = NULL;
2665 request->rq_reqdata_len = 0;
2666 spin_unlock(&request->rq_early_free_lock);
2668 spin_unlock(&request->rq_lock);
2671 __ptlrpc_free_req(request, locked);
2677 * Drops one reference count for a request.
2679 void ptlrpc_req_finished(struct ptlrpc_request *request)
2681 __ptlrpc_req_finished(request, 0);
2683 EXPORT_SYMBOL(ptlrpc_req_finished);
2686 * Returns xid of a \a request
2688 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2690 return request->rq_xid;
2692 EXPORT_SYMBOL(ptlrpc_req_xid);
2695 * Disengage the client's reply buffer from the network
2696 * NB does _NOT_ unregister any client-side bulk.
2697 * IDEMPOTENT, but _not_ safe against concurrent callers.
2698 * The request owner (i.e. the thread doing the I/O) must call...
2699 * Returns 0 on success or 1 if unregistering cannot be made.
2701 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2706 LASSERT(!in_interrupt());
2708 /* Let's setup deadline for reply unlink. */
2709 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2710 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2711 request->rq_reply_deadline = ktime_get_real_seconds() +
2715 * Nothing left to do.
2717 if (!ptlrpc_client_recv_or_unlink(request))
2720 LNetMDUnlink(request->rq_reply_md_h);
2723 * Let's check it once again.
2725 if (!ptlrpc_client_recv_or_unlink(request))
2728 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2729 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2732 * Do not wait for unlink to finish.
2738 * We have to wait_event_idle_timeout() whatever the result, to get
2739 * a chance to run reply_in_callback(), and to make sure we've
2740 * unlinked before returning a req to the pool.
2743 wait_queue_head_t *wq = (request->rq_set) ?
2744 &request->rq_set->set_waitq :
2745 &request->rq_reply_waitq;
2746 int seconds = LONG_UNLINK;
2748 * Network access will complete in finite time but the HUGE
2749 * timeout lets us CWARN for visibility of sluggish NALs
2751 while (seconds > 0 &&
2752 wait_event_idle_timeout(
2754 !ptlrpc_client_recv_or_unlink(request),
2755 cfs_time_seconds(1)) == 0)
2758 ptlrpc_rqphase_move(request, request->rq_next_phase);
2762 DEBUG_REQ(D_WARNING, request,
2763 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2764 request->rq_receiving_reply,
2765 request->rq_req_unlinked,
2766 request->rq_reply_unlinked);
2771 static void ptlrpc_free_request(struct ptlrpc_request *req)
2773 spin_lock(&req->rq_lock);
2775 spin_unlock(&req->rq_lock);
2777 if (req->rq_commit_cb)
2778 req->rq_commit_cb(req);
2779 list_del_init(&req->rq_replay_list);
2781 __ptlrpc_req_finished(req, 1);
2785 * the request is committed and dropped from the replay list of its import
2787 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2789 struct obd_import *imp = req->rq_import;
2791 spin_lock(&imp->imp_lock);
2792 if (list_empty(&req->rq_replay_list)) {
2793 spin_unlock(&imp->imp_lock);
2797 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2798 if (imp->imp_replay_cursor == &req->rq_replay_list)
2799 imp->imp_replay_cursor = req->rq_replay_list.next;
2800 ptlrpc_free_request(req);
2803 spin_unlock(&imp->imp_lock);
2805 EXPORT_SYMBOL(ptlrpc_request_committed);
2808 * Iterates through replay_list on import and prunes
2809 * all requests have transno smaller than last_committed for the
2810 * import and don't have rq_replay set.
2811 * Since requests are sorted in transno order, stops when meetign first
2812 * transno bigger than last_committed.
2813 * caller must hold imp->imp_lock
2815 void ptlrpc_free_committed(struct obd_import *imp)
2817 struct ptlrpc_request *req, *saved;
2818 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2819 bool skip_committed_list = true;
2822 LASSERT(imp != NULL);
2823 assert_spin_locked(&imp->imp_lock);
2825 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2826 imp->imp_generation == imp->imp_last_generation_checked) {
2827 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2828 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2831 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2832 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2833 imp->imp_generation);
2835 if (imp->imp_generation != imp->imp_last_generation_checked ||
2836 imp->imp_last_transno_checked == 0)
2837 skip_committed_list = false;
2839 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2840 imp->imp_last_generation_checked = imp->imp_generation;
2842 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2844 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2845 LASSERT(req != last_req);
2848 if (req->rq_transno == 0) {
2849 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2852 if (req->rq_import_generation < imp->imp_generation) {
2853 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2857 /* not yet committed */
2858 if (req->rq_transno > imp->imp_peer_committed_transno) {
2859 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2863 if (req->rq_replay) {
2864 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2865 list_move_tail(&req->rq_replay_list,
2866 &imp->imp_committed_list);
2870 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2871 imp->imp_peer_committed_transno);
2873 ptlrpc_free_request(req);
2876 if (skip_committed_list)
2879 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2881 LASSERT(req->rq_transno != 0);
2882 if (req->rq_import_generation < imp->imp_generation ||
2884 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2885 req->rq_import_generation <
2886 imp->imp_generation ? "stale" : "closed");
2888 if (imp->imp_replay_cursor == &req->rq_replay_list)
2889 imp->imp_replay_cursor =
2890 req->rq_replay_list.next;
2892 ptlrpc_free_request(req);
2899 void ptlrpc_cleanup_client(struct obd_import *imp)
2906 * Schedule previously sent request for resend.
2907 * For bulk requests we assign new xid (to avoid problems with
2908 * lost replies and therefore several transfers landing into same buffer
2909 * from different sending attempts).
2911 void ptlrpc_resend_req(struct ptlrpc_request *req)
2913 DEBUG_REQ(D_HA, req, "going to resend");
2914 spin_lock(&req->rq_lock);
2917 * Request got reply but linked to the import list still.
2918 * Let ptlrpc_check_set() process it.
2920 if (ptlrpc_client_replied(req)) {
2921 spin_unlock(&req->rq_lock);
2922 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2926 req->rq_status = -EAGAIN;
2929 req->rq_net_err = 0;
2930 req->rq_timedout = 0;
2932 ptlrpc_client_wake_req(req);
2933 spin_unlock(&req->rq_lock);
2936 /* XXX: this function and rq_status are currently unused */
2937 void ptlrpc_restart_req(struct ptlrpc_request *req)
2939 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2940 req->rq_status = -ERESTARTSYS;
2942 spin_lock(&req->rq_lock);
2943 req->rq_restart = 1;
2944 req->rq_timedout = 0;
2945 ptlrpc_client_wake_req(req);
2946 spin_unlock(&req->rq_lock);
2950 * Grab additional reference on a request \a req
2952 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2955 atomic_inc(&req->rq_refcount);
2958 EXPORT_SYMBOL(ptlrpc_request_addref);
2961 * Add a request to import replay_list.
2962 * Must be called under imp_lock
2964 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2965 struct obd_import *imp)
2967 struct list_head *tmp;
2969 assert_spin_locked(&imp->imp_lock);
2971 if (req->rq_transno == 0) {
2972 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2977 * clear this for new requests that were resent as well
2978 * as resent replayed requests.
2980 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2982 /* don't re-add requests that have been replayed */
2983 if (!list_empty(&req->rq_replay_list))
2986 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2988 spin_lock(&req->rq_lock);
2990 spin_unlock(&req->rq_lock);
2992 LASSERT(imp->imp_replayable);
2993 /* Balanced in ptlrpc_free_committed, usually. */
2994 ptlrpc_request_addref(req);
2995 list_for_each_prev(tmp, &imp->imp_replay_list) {
2996 struct ptlrpc_request *iter = list_entry(tmp,
2997 struct ptlrpc_request,
3001 * We may have duplicate transnos if we create and then
3002 * open a file, or for closes retained if to match creating
3003 * opens, so use req->rq_xid as a secondary key.
3004 * (See bugs 684, 685, and 428.)
3005 * XXX no longer needed, but all opens need transnos!
3007 if (iter->rq_transno > req->rq_transno)
3010 if (iter->rq_transno == req->rq_transno) {
3011 LASSERT(iter->rq_xid != req->rq_xid);
3012 if (iter->rq_xid > req->rq_xid)
3016 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3020 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3024 * Send request and wait until it completes.
3025 * Returns request processing status.
3027 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3029 struct ptlrpc_request_set *set;
3033 LASSERT(req->rq_set == NULL);
3034 LASSERT(!req->rq_receiving_reply);
3036 set = ptlrpc_prep_set();
3038 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3042 /* for distributed debugging */
3043 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3045 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3046 ptlrpc_request_addref(req);
3047 ptlrpc_set_add_req(set, req);
3048 rc = ptlrpc_set_wait(NULL, set);
3049 ptlrpc_set_destroy(set);
3053 EXPORT_SYMBOL(ptlrpc_queue_wait);
3056 * Callback used for replayed requests reply processing.
3057 * In case of successful reply calls registered request replay callback.
3058 * In case of error restart replay process.
3060 static int ptlrpc_replay_interpret(const struct lu_env *env,
3061 struct ptlrpc_request *req,
3064 struct ptlrpc_replay_async_args *aa = args;
3065 struct obd_import *imp = req->rq_import;
3068 atomic_dec(&imp->imp_replay_inflight);
3071 * Note: if it is bulk replay (MDS-MDS replay), then even if
3072 * server got the request, but bulk transfer timeout, let's
3073 * replay the bulk req again
3075 if (!ptlrpc_client_replied(req) ||
3077 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3078 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3079 GOTO(out, rc = -ETIMEDOUT);
3082 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3083 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3084 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3085 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3087 /** VBR: check version failure */
3088 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3089 /** replay was failed due to version mismatch */
3090 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3091 spin_lock(&imp->imp_lock);
3092 imp->imp_vbr_failed = 1;
3093 spin_unlock(&imp->imp_lock);
3094 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3096 /** The transno had better not change over replay. */
3097 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3098 lustre_msg_get_transno(req->rq_repmsg) ||
3099 lustre_msg_get_transno(req->rq_repmsg) == 0,
3101 lustre_msg_get_transno(req->rq_reqmsg),
3102 lustre_msg_get_transno(req->rq_repmsg));
3105 spin_lock(&imp->imp_lock);
3106 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3107 spin_unlock(&imp->imp_lock);
3108 LASSERT(imp->imp_last_replay_transno);
3110 /* transaction number shouldn't be bigger than the latest replayed */
3111 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3112 DEBUG_REQ(D_ERROR, req,
3113 "Reported transno=%llu is bigger than replayed=%llu",
3115 lustre_msg_get_transno(req->rq_reqmsg));
3116 GOTO(out, rc = -EINVAL);
3119 DEBUG_REQ(D_HA, req, "got reply");
3121 /* let the callback do fixups, possibly including in the request */
3122 if (req->rq_replay_cb)
3123 req->rq_replay_cb(req);
3125 if (ptlrpc_client_replied(req) &&
3126 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3127 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3128 lustre_msg_get_status(req->rq_repmsg),
3129 aa->praa_old_status);
3132 * Note: If the replay fails for MDT-MDT recovery, let's
3133 * abort all of the following requests in the replay
3134 * and sending list, because MDT-MDT update requests
3135 * are dependent on each other, see LU-7039
3137 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3138 struct ptlrpc_request *free_req;
3139 struct ptlrpc_request *tmp;
3141 spin_lock(&imp->imp_lock);
3142 list_for_each_entry_safe(free_req, tmp,
3143 &imp->imp_replay_list,
3145 ptlrpc_free_request(free_req);
3148 list_for_each_entry_safe(free_req, tmp,
3149 &imp->imp_committed_list,
3151 ptlrpc_free_request(free_req);
3154 list_for_each_entry_safe(free_req, tmp,
3155 &imp->imp_delayed_list,
3157 spin_lock(&free_req->rq_lock);
3158 free_req->rq_err = 1;
3159 free_req->rq_status = -EIO;
3160 ptlrpc_client_wake_req(free_req);
3161 spin_unlock(&free_req->rq_lock);
3164 list_for_each_entry_safe(free_req, tmp,
3165 &imp->imp_sending_list,
3167 spin_lock(&free_req->rq_lock);
3168 free_req->rq_err = 1;
3169 free_req->rq_status = -EIO;
3170 ptlrpc_client_wake_req(free_req);
3171 spin_unlock(&free_req->rq_lock);
3173 spin_unlock(&imp->imp_lock);
3176 /* Put it back for re-replay. */
3177 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3181 * Errors while replay can set transno to 0, but
3182 * imp_last_replay_transno shouldn't be set to 0 anyway
3184 if (req->rq_transno == 0)
3185 CERROR("Transno is 0 during replay!\n");
3187 /* continue with recovery */
3188 rc = ptlrpc_import_recovery_state_machine(imp);
3190 req->rq_send_state = aa->praa_old_state;
3193 /* this replay failed, so restart recovery */
3194 ptlrpc_connect_import(imp);
3200 * Prepares and queues request for replay.
3201 * Adds it to ptlrpcd queue for actual sending.
3202 * Returns 0 on success.
3204 int ptlrpc_replay_req(struct ptlrpc_request *req)
3206 struct ptlrpc_replay_async_args *aa;
3210 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3212 aa = ptlrpc_req_async_args(aa, req);
3213 memset(aa, 0, sizeof(*aa));
3215 /* Prepare request to be resent with ptlrpcd */
3216 aa->praa_old_state = req->rq_send_state;
3217 req->rq_send_state = LUSTRE_IMP_REPLAY;
3218 req->rq_phase = RQ_PHASE_NEW;
3219 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3221 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3223 req->rq_interpret_reply = ptlrpc_replay_interpret;
3224 /* Readjust the timeout for current conditions */
3225 ptlrpc_at_set_req_timeout(req);
3227 /* Tell server net_latency to calculate how long to wait for reply. */
3228 lustre_msg_set_service_timeout(req->rq_reqmsg,
3229 ptlrpc_at_get_net_latency(req));
3230 DEBUG_REQ(D_HA, req, "REPLAY");
3232 atomic_inc(&req->rq_import->imp_replay_inflight);
3233 spin_lock(&req->rq_lock);
3234 req->rq_early_free_repbuf = 0;
3235 spin_unlock(&req->rq_lock);
3236 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3238 ptlrpcd_add_req(req);
3243 * Aborts all in-flight request on import \a imp sending and delayed lists
3245 void ptlrpc_abort_inflight(struct obd_import *imp)
3247 struct list_head *tmp, *n;
3251 * Make sure that no new requests get processed for this import.
3252 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3253 * this flag and then putting requests on sending_list or delayed_list.
3255 assert_spin_locked(&imp->imp_lock);
3258 * XXX locking? Maybe we should remove each request with the list
3259 * locked? Also, how do we know if the requests on the list are
3260 * being freed at this time?
3262 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3263 struct ptlrpc_request *req = list_entry(tmp,
3264 struct ptlrpc_request,
3267 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3269 spin_lock(&req->rq_lock);
3270 if (req->rq_import_generation < imp->imp_generation) {
3272 req->rq_status = -EIO;
3273 ptlrpc_client_wake_req(req);
3275 spin_unlock(&req->rq_lock);
3278 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3279 struct ptlrpc_request *req =
3280 list_entry(tmp, struct ptlrpc_request, rq_list);
3282 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3284 spin_lock(&req->rq_lock);
3285 if (req->rq_import_generation < imp->imp_generation) {
3287 req->rq_status = -EIO;
3288 ptlrpc_client_wake_req(req);
3290 spin_unlock(&req->rq_lock);
3294 * Last chance to free reqs left on the replay list, but we
3295 * will still leak reqs that haven't committed.
3297 if (imp->imp_replayable)
3298 ptlrpc_free_committed(imp);
3304 * Abort all uncompleted requests in request set \a set
3306 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3308 struct list_head *tmp, *pos;
3310 LASSERT(set != NULL);
3312 list_for_each_safe(pos, tmp, &set->set_requests) {
3313 struct ptlrpc_request *req =
3314 list_entry(pos, struct ptlrpc_request,
3317 spin_lock(&req->rq_lock);
3318 if (req->rq_phase != RQ_PHASE_RPC) {
3319 spin_unlock(&req->rq_lock);
3324 req->rq_status = -EINTR;
3325 ptlrpc_client_wake_req(req);
3326 spin_unlock(&req->rq_lock);
3331 * Initialize the XID for the node. This is common among all requests on
3332 * this node, and only requires the property that it is monotonically
3333 * increasing. It does not need to be sequential. Since this is also used
3334 * as the RDMA match bits, it is important that a single client NOT have
3335 * the same match bits for two different in-flight requests, hence we do
3336 * NOT want to have an XID per target or similar.
3338 * To avoid an unlikely collision between match bits after a client reboot
3339 * (which would deliver old data into the wrong RDMA buffer) initialize
3340 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3341 * If the time is clearly incorrect, we instead use a 62-bit random number.
3342 * In the worst case the random number will overflow 1M RPCs per second in
3343 * 9133 years, or permutations thereof.
3345 #define YEAR_2004 (1ULL << 30)
3346 void ptlrpc_init_xid(void)
3348 time64_t now = ktime_get_real_seconds();
3351 if (now < YEAR_2004) {
3352 get_random_bytes(&xid, sizeof(xid));
3354 xid |= (1ULL << 61);
3356 xid = (u64)now << 20;
3359 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3360 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3362 xid &= PTLRPC_BULK_OPS_MASK;
3363 atomic64_set(&ptlrpc_last_xid, xid);
3367 * Increase xid and returns resulting new value to the caller.
3369 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3370 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3371 * itself uses the last bulk xid needed, so the server can determine the
3372 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3373 * xid must align to a power-of-two value.
3375 * This is assumed to be true due to the initial ptlrpc_last_xid
3376 * value also being initialized to a power-of-two value. LU-1431
3378 __u64 ptlrpc_next_xid(void)
3380 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3384 * If request has a new allocated XID (new request or EINPROGRESS resend),
3385 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3386 * request to ensure previous bulk fails and avoid problems with lost replies
3387 * and therefore several transfers landing into the same buffer from different
3390 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3392 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3394 LASSERT(bd != NULL);
3397 * Generate new matchbits for all resend requests, including
3400 if (req->rq_resend) {
3401 __u64 old_mbits = req->rq_mbits;
3404 * First time resend on -EINPROGRESS will generate new xid,
3405 * so we can actually use the rq_xid as rq_mbits in such case,
3406 * however, it's bit hard to distinguish such resend with a
3407 * 'resend for the -EINPROGRESS resend'. To make it simple,
3408 * we opt to generate mbits for all resend cases.
3410 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data,
3412 req->rq_mbits = ptlrpc_next_xid();
3415 * Old version transfers rq_xid to peer as
3418 spin_lock(&req->rq_import->imp_lock);
3419 list_del_init(&req->rq_unreplied_list);
3420 ptlrpc_assign_next_xid_nolock(req);
3421 spin_unlock(&req->rq_import->imp_lock);
3422 req->rq_mbits = req->rq_xid;
3424 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3425 old_mbits, req->rq_mbits);
3426 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3427 /* Request being sent first time, use xid as matchbits. */
3428 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3429 || req->rq_mbits == 0) {
3430 req->rq_mbits = req->rq_xid;
3432 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3434 req->rq_mbits -= total_md - 1;
3438 * Replay request, xid and matchbits have already been
3439 * correctly assigned.
3445 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3446 * that server can infer the number of bulks that were prepared,
3449 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3453 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3454 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3456 * It's ok to directly set the rq_xid here, since this xid bump
3457 * won't affect the request position in unreplied list.
3459 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3460 req->rq_xid = req->rq_mbits;
3464 * Get a glimpse at what next xid value might have been.
3465 * Returns possible next xid.
3467 __u64 ptlrpc_sample_next_xid(void)
3469 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3471 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3474 * Functions for operating ptlrpc workers.
3476 * A ptlrpc work is a function which will be running inside ptlrpc context.
3477 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3479 * 1. after a work is created, it can be used many times, that is:
3480 * handler = ptlrpcd_alloc_work();
3481 * ptlrpcd_queue_work();
3483 * queue it again when necessary:
3484 * ptlrpcd_queue_work();
3485 * ptlrpcd_destroy_work();
3486 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3487 * it will only be queued once in any time. Also as its name implies, it may
3488 * have delay before it really runs by ptlrpcd thread.
3490 struct ptlrpc_work_async_args {
3491 int (*cb)(const struct lu_env *, void *);
3495 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3497 /* re-initialize the req */
3498 req->rq_timeout = obd_timeout;
3499 req->rq_sent = ktime_get_real_seconds();
3500 req->rq_deadline = req->rq_sent + req->rq_timeout;
3501 req->rq_phase = RQ_PHASE_INTERPRET;
3502 req->rq_next_phase = RQ_PHASE_COMPLETE;
3503 req->rq_xid = ptlrpc_next_xid();
3504 req->rq_import_generation = req->rq_import->imp_generation;
3506 ptlrpcd_add_req(req);
3509 static int work_interpreter(const struct lu_env *env,
3510 struct ptlrpc_request *req, void *args, int rc)
3512 struct ptlrpc_work_async_args *arg = args;
3514 LASSERT(ptlrpcd_check_work(req));
3515 LASSERT(arg->cb != NULL);
3517 rc = arg->cb(env, arg->cbdata);
3519 list_del_init(&req->rq_set_chain);
3522 if (atomic_dec_return(&req->rq_refcount) > 1) {
3523 atomic_set(&req->rq_refcount, 2);
3524 ptlrpcd_add_work_req(req);
3529 static int worker_format;
3531 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3533 return req->rq_pill.rc_fmt == (void *)&worker_format;
3537 * Create a work for ptlrpc.
3539 void *ptlrpcd_alloc_work(struct obd_import *imp,
3540 int (*cb)(const struct lu_env *, void *), void *cbdata)
3542 struct ptlrpc_request *req = NULL;
3543 struct ptlrpc_work_async_args *args;
3549 RETURN(ERR_PTR(-EINVAL));
3551 /* copy some code from deprecated fakereq. */
3552 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3554 CERROR("ptlrpc: run out of memory!\n");
3555 RETURN(ERR_PTR(-ENOMEM));
3558 ptlrpc_cli_req_init(req);
3560 req->rq_send_state = LUSTRE_IMP_FULL;
3561 req->rq_type = PTL_RPC_MSG_REQUEST;
3562 req->rq_import = class_import_get(imp);
3563 req->rq_interpret_reply = work_interpreter;
3564 /* don't want reply */
3565 req->rq_no_delay = req->rq_no_resend = 1;
3566 req->rq_pill.rc_fmt = (void *)&worker_format;
3568 args = ptlrpc_req_async_args(args, req);
3570 args->cbdata = cbdata;
3574 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3576 void ptlrpcd_destroy_work(void *handler)
3578 struct ptlrpc_request *req = handler;
3581 ptlrpc_req_finished(req);
3583 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3585 int ptlrpcd_queue_work(void *handler)
3587 struct ptlrpc_request *req = handler;
3590 * Check if the req is already being queued.
3592 * Here comes a trick: it lacks a way of checking if a req is being
3593 * processed reliably in ptlrpc. Here I have to use refcount of req
3594 * for this purpose. This is okay because the caller should use this
3595 * req as opaque data. - Jinshan
3597 LASSERT(atomic_read(&req->rq_refcount) > 0);
3598 if (atomic_inc_return(&req->rq_refcount) == 2)
3599 ptlrpcd_add_work_req(req);
3602 EXPORT_SYMBOL(ptlrpcd_queue_work);