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 <obd_support.h>
38 #include <obd_class.h>
39 #include <lustre_lib.h>
40 #include <lustre_ha.h>
41 #include <lustre_import.h>
42 #include <lustre_req_layout.h>
44 #include "ptlrpc_internal.h"
46 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
47 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
48 .release_frags = ptlrpc_release_bulk_page_pin,
50 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
52 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
53 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
54 .release_frags = ptlrpc_release_bulk_noop,
56 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
58 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
59 .add_iov_frag = ptlrpc_prep_bulk_frag,
61 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
63 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
64 static int ptlrpcd_check_work(struct ptlrpc_request *req);
65 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
68 * Initialize passed in client structure \a cl.
70 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
71 struct ptlrpc_client *cl)
73 cl->cli_request_portal = req_portal;
74 cl->cli_reply_portal = rep_portal;
77 EXPORT_SYMBOL(ptlrpc_init_client);
80 * Return PortalRPC connection for remore uud \a uuid
82 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
83 lnet_nid_t nid4refnet)
85 struct ptlrpc_connection *c;
87 struct lnet_process_id peer;
90 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
91 * before accessing its values. */
92 /* coverity[uninit_use_in_call] */
93 peer.nid = nid4refnet;
94 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
96 CNETERR("cannot find peer %s!\n", uuid->uuid);
100 c = ptlrpc_connection_get(peer, self, uuid);
102 memcpy(c->c_remote_uuid.uuid,
103 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
106 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
112 * Allocate and initialize new bulk descriptor on the sender.
113 * Returns pointer to the descriptor or NULL on error.
115 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
116 enum ptlrpc_bulk_op_type type,
118 const struct ptlrpc_bulk_frag_ops *ops)
120 struct ptlrpc_bulk_desc *desc;
123 /* ensure that only one of KIOV or IOVEC is set but not both */
124 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
125 ops->add_kiov_frag != NULL) ||
126 (ptlrpc_is_bulk_desc_kvec(type) &&
127 ops->add_iov_frag != NULL));
132 if (type & PTLRPC_BULK_BUF_KIOV) {
133 OBD_ALLOC_LARGE(GET_KIOV(desc),
134 nfrags * sizeof(*GET_KIOV(desc)));
135 if (GET_KIOV(desc) == NULL)
138 OBD_ALLOC_LARGE(GET_KVEC(desc),
139 nfrags * sizeof(*GET_KVEC(desc)));
140 if (GET_KVEC(desc) == NULL)
144 spin_lock_init(&desc->bd_lock);
145 init_waitqueue_head(&desc->bd_waitq);
146 desc->bd_max_iov = nfrags;
147 desc->bd_iov_count = 0;
148 desc->bd_portal = portal;
149 desc->bd_type = type;
150 desc->bd_md_count = 0;
151 desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
152 LASSERT(max_brw > 0);
153 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
154 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
155 * node. Negotiated ocd_brw_size will always be <= this number. */
156 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
157 LNetInvalidateMDHandle(&desc->bd_mds[i]);
166 * Prepare bulk descriptor for specified outgoing request \a req that
167 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
168 * the bulk to be sent. Used on client-side.
169 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
172 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
173 unsigned nfrags, unsigned max_brw,
176 const struct ptlrpc_bulk_frag_ops
179 struct obd_import *imp = req->rq_import;
180 struct ptlrpc_bulk_desc *desc;
183 LASSERT(ptlrpc_is_bulk_op_passive(type));
185 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
189 desc->bd_import_generation = req->rq_import_generation;
190 desc->bd_import = class_import_get(imp);
193 desc->bd_cbid.cbid_fn = client_bulk_callback;
194 desc->bd_cbid.cbid_arg = desc;
196 /* This makes req own desc, and free it when she frees herself */
201 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
203 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
204 struct page *page, int pageoffset, int len,
209 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
210 LASSERT(page != NULL);
211 LASSERT(pageoffset >= 0);
213 LASSERT(pageoffset + len <= PAGE_SIZE);
214 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
216 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
223 kiov->kiov_page = page;
224 kiov->kiov_offset = pageoffset;
225 kiov->kiov_len = len;
227 desc->bd_iov_count++;
229 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
231 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
237 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
238 LASSERT(frag != NULL);
240 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
242 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
246 iovec->iov_base = frag;
247 iovec->iov_len = len;
249 desc->bd_iov_count++;
251 RETURN(desc->bd_nob);
253 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
255 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
259 LASSERT(desc != NULL);
260 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
261 LASSERT(desc->bd_md_count == 0); /* network hands off */
262 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
263 LASSERT(desc->bd_frag_ops != NULL);
265 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
266 sptlrpc_enc_pool_put_pages(desc);
269 class_export_put(desc->bd_export);
271 class_import_put(desc->bd_import);
273 if (desc->bd_frag_ops->release_frags != NULL)
274 desc->bd_frag_ops->release_frags(desc);
276 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
277 OBD_FREE_LARGE(GET_KIOV(desc),
278 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
280 OBD_FREE_LARGE(GET_KVEC(desc),
281 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
285 EXPORT_SYMBOL(ptlrpc_free_bulk);
288 * Set server timelimit for this req, i.e. how long are we willing to wait
289 * for reply before timing out this request.
291 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
297 LASSERT(req->rq_import);
300 /* non-AT settings */
302 * \a imp_server_timeout means this is reverse import and
303 * we send (currently only) ASTs to the client and cannot afford
304 * to wait too long for the reply, otherwise the other client
305 * (because of which we are sending this request) would
306 * timeout waiting for us
308 req->rq_timeout = req->rq_import->imp_server_timeout ?
309 obd_timeout / 2 : obd_timeout;
311 at = &req->rq_import->imp_at;
312 idx = import_at_get_index(req->rq_import,
313 req->rq_request_portal);
314 serv_est = at_get(&at->iat_service_estimate[idx]);
315 req->rq_timeout = at_est2timeout(serv_est);
317 /* We could get even fancier here, using history to predict increased
320 /* Let the server know what this RPC timeout is by putting it in the
322 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
324 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
326 /* Adjust max service estimate based on server value */
327 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
328 unsigned int serv_est)
334 LASSERT(req->rq_import);
335 at = &req->rq_import->imp_at;
337 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
338 /* max service estimates are tracked on the server side,
339 so just keep minimal history here */
340 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
342 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
343 "has changed from %d to %d\n",
344 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
345 oldse, at_get(&at->iat_service_estimate[idx]));
348 /* Expected network latency per remote node (secs) */
349 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
351 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
354 /* Adjust expected network latency */
355 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
356 unsigned int service_time)
358 unsigned int nl, oldnl;
360 time64_t now = ktime_get_real_seconds();
362 LASSERT(req->rq_import);
364 if (service_time > now - req->rq_sent + 3) {
365 /* bz16408, however, this can also happen if early reply
366 * is lost and client RPC is expired and resent, early reply
367 * or reply of original RPC can still be fit in reply buffer
368 * of resent RPC, now client is measuring time from the
369 * resent time, but server sent back service time of original
372 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
373 D_ADAPTTO : D_WARNING,
374 "Reported service time %u > total measured time %lld\n",
375 service_time, now - req->rq_sent);
379 /* Network latency is total time less server processing time */
380 nl = max_t(int, now - req->rq_sent -
381 service_time, 0) + 1; /* st rounding */
382 at = &req->rq_import->imp_at;
384 oldnl = at_measured(&at->iat_net_latency, nl);
386 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
387 "has changed from %d to %d\n",
388 req->rq_import->imp_obd->obd_name,
390 &req->rq_import->imp_connection->c_remote_uuid),
391 oldnl, at_get(&at->iat_net_latency));
394 static int unpack_reply(struct ptlrpc_request *req)
398 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
399 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
401 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
406 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
408 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
415 * Handle an early reply message, called with the rq_lock held.
416 * If anything goes wrong just ignore it - same as if it never happened
418 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
419 __must_hold(&req->rq_lock)
421 struct ptlrpc_request *early_req;
427 spin_unlock(&req->rq_lock);
429 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
431 spin_lock(&req->rq_lock);
435 rc = unpack_reply(early_req);
437 sptlrpc_cli_finish_early_reply(early_req);
438 spin_lock(&req->rq_lock);
442 /* Use new timeout value just to adjust the local value for this
443 * request, don't include it into at_history. It is unclear yet why
444 * service time increased and should it be counted or skipped, e.g.
445 * that can be recovery case or some error or server, the real reply
446 * will add all new data if it is worth to add. */
447 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
448 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
450 /* Network latency can be adjusted, it is pure network delays */
451 ptlrpc_at_adj_net_latency(req,
452 lustre_msg_get_service_time(early_req->rq_repmsg));
454 sptlrpc_cli_finish_early_reply(early_req);
456 spin_lock(&req->rq_lock);
457 olddl = req->rq_deadline;
458 /* server assumes it now has rq_timeout from when the request
459 * arrived, so the client should give it at least that long.
460 * since we don't know the arrival time we'll use the original
462 req->rq_deadline = req->rq_sent + req->rq_timeout +
463 ptlrpc_at_get_net_latency(req);
465 DEBUG_REQ(D_ADAPTTO, req,
466 "Early reply #%d, new deadline in %llds (%llds)",
468 req->rq_deadline - ktime_get_real_seconds(),
469 req->rq_deadline - olddl);
474 static struct kmem_cache *request_cache;
476 int ptlrpc_request_cache_init(void)
478 request_cache = kmem_cache_create("ptlrpc_cache",
479 sizeof(struct ptlrpc_request),
480 0, SLAB_HWCACHE_ALIGN, NULL);
481 return request_cache == NULL ? -ENOMEM : 0;
484 void ptlrpc_request_cache_fini(void)
486 kmem_cache_destroy(request_cache);
489 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
491 struct ptlrpc_request *req;
493 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
497 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
499 OBD_SLAB_FREE_PTR(req, request_cache);
503 * Wind down request pool \a pool.
504 * Frees all requests from the pool too
506 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
508 struct list_head *l, *tmp;
509 struct ptlrpc_request *req;
511 LASSERT(pool != NULL);
513 spin_lock(&pool->prp_lock);
514 list_for_each_safe(l, tmp, &pool->prp_req_list) {
515 req = list_entry(l, struct ptlrpc_request, rq_list);
516 list_del(&req->rq_list);
517 LASSERT(req->rq_reqbuf);
518 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
519 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
520 ptlrpc_request_cache_free(req);
522 spin_unlock(&pool->prp_lock);
523 OBD_FREE(pool, sizeof(*pool));
525 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
528 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
530 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
535 while (size < pool->prp_rq_size)
538 LASSERTF(list_empty(&pool->prp_req_list) ||
539 size == pool->prp_rq_size,
540 "Trying to change pool size with nonempty pool "
541 "from %d to %d bytes\n", pool->prp_rq_size, size);
543 spin_lock(&pool->prp_lock);
544 pool->prp_rq_size = size;
545 for (i = 0; i < num_rq; i++) {
546 struct ptlrpc_request *req;
547 struct lustre_msg *msg;
549 spin_unlock(&pool->prp_lock);
550 req = ptlrpc_request_cache_alloc(GFP_NOFS);
553 OBD_ALLOC_LARGE(msg, size);
555 ptlrpc_request_cache_free(req);
558 req->rq_reqbuf = msg;
559 req->rq_reqbuf_len = size;
561 spin_lock(&pool->prp_lock);
562 list_add_tail(&req->rq_list, &pool->prp_req_list);
564 spin_unlock(&pool->prp_lock);
567 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
570 * Create and initialize new request pool with given attributes:
571 * \a num_rq - initial number of requests to create for the pool
572 * \a msgsize - maximum message size possible for requests in thid pool
573 * \a populate_pool - function to be called when more requests need to be added
575 * Returns pointer to newly created pool or NULL on error.
577 struct ptlrpc_request_pool *
578 ptlrpc_init_rq_pool(int num_rq, int msgsize,
579 int (*populate_pool)(struct ptlrpc_request_pool *, int))
581 struct ptlrpc_request_pool *pool;
583 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
587 /* Request next power of two for the allocation, because internally
588 kernel would do exactly this */
590 spin_lock_init(&pool->prp_lock);
591 INIT_LIST_HEAD(&pool->prp_req_list);
592 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
593 pool->prp_populate = populate_pool;
595 populate_pool(pool, num_rq);
599 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
602 * Fetches one request from pool \a pool
604 static struct ptlrpc_request *
605 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
607 struct ptlrpc_request *request;
608 struct lustre_msg *reqbuf;
613 spin_lock(&pool->prp_lock);
615 /* See if we have anything in a pool, and bail out if nothing,
616 * in writeout path, where this matters, this is safe to do, because
617 * nothing is lost in this case, and when some in-flight requests
618 * complete, this code will be called again. */
619 if (unlikely(list_empty(&pool->prp_req_list))) {
620 spin_unlock(&pool->prp_lock);
624 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
626 list_del_init(&request->rq_list);
627 spin_unlock(&pool->prp_lock);
629 LASSERT(request->rq_reqbuf);
630 LASSERT(request->rq_pool);
632 reqbuf = request->rq_reqbuf;
633 memset(request, 0, sizeof(*request));
634 request->rq_reqbuf = reqbuf;
635 request->rq_reqbuf_len = pool->prp_rq_size;
636 request->rq_pool = pool;
642 * Returns freed \a request to pool.
644 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
646 struct ptlrpc_request_pool *pool = request->rq_pool;
648 spin_lock(&pool->prp_lock);
649 LASSERT(list_empty(&request->rq_list));
650 LASSERT(!request->rq_receiving_reply);
651 list_add_tail(&request->rq_list, &pool->prp_req_list);
652 spin_unlock(&pool->prp_lock);
655 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
657 struct obd_import *imp = req->rq_import;
658 struct list_head *tmp;
659 struct ptlrpc_request *iter;
661 assert_spin_locked(&imp->imp_lock);
662 LASSERT(list_empty(&req->rq_unreplied_list));
664 /* unreplied list is sorted by xid in ascending order */
665 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
666 iter = list_entry(tmp, struct ptlrpc_request,
669 LASSERT(req->rq_xid != iter->rq_xid);
670 if (req->rq_xid < iter->rq_xid)
672 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
675 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
678 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
680 req->rq_xid = ptlrpc_next_xid();
681 ptlrpc_add_unreplied(req);
684 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
686 spin_lock(&req->rq_import->imp_lock);
687 ptlrpc_assign_next_xid_nolock(req);
688 spin_unlock(&req->rq_import->imp_lock);
691 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
692 __u32 version, int opcode, char **bufs,
693 struct ptlrpc_cli_ctx *ctx)
696 struct obd_import *imp;
702 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
703 imp = request->rq_import;
704 lengths = request->rq_pill.rc_area[RCL_CLIENT];
707 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
709 rc = sptlrpc_req_get_ctx(request);
713 sptlrpc_req_set_flavor(request, opcode);
715 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
720 lustre_msg_add_version(request->rq_reqmsg, version);
721 request->rq_send_state = LUSTRE_IMP_FULL;
722 request->rq_type = PTL_RPC_MSG_REQUEST;
724 request->rq_req_cbid.cbid_fn = request_out_callback;
725 request->rq_req_cbid.cbid_arg = request;
727 request->rq_reply_cbid.cbid_fn = reply_in_callback;
728 request->rq_reply_cbid.cbid_arg = request;
730 request->rq_reply_deadline = 0;
731 request->rq_bulk_deadline = 0;
732 request->rq_req_deadline = 0;
733 request->rq_phase = RQ_PHASE_NEW;
734 request->rq_next_phase = RQ_PHASE_UNDEFINED;
736 request->rq_request_portal = imp->imp_client->cli_request_portal;
737 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
739 ptlrpc_at_set_req_timeout(request);
741 lustre_msg_set_opc(request->rq_reqmsg, opcode);
742 ptlrpc_assign_next_xid(request);
744 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
745 if (cfs_fail_val == opcode) {
746 time64_t *fail_t = NULL, *fail2_t = NULL;
748 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
749 fail_t = &request->rq_bulk_deadline;
750 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
751 fail_t = &request->rq_reply_deadline;
752 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
753 fail_t = &request->rq_req_deadline;
754 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
755 fail_t = &request->rq_reply_deadline;
756 fail2_t = &request->rq_bulk_deadline;
760 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
763 *fail2_t = ktime_get_real_seconds() +
767 * The RPC is infected, let the test to change the
770 msleep(4 * MSEC_PER_SEC);
777 LASSERT(!request->rq_pool);
778 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
780 class_import_put(imp);
785 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
788 * Pack request buffers for network transfer, performing necessary encryption
789 * steps if necessary.
791 int ptlrpc_request_pack(struct ptlrpc_request *request,
792 __u32 version, int opcode)
795 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
799 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
800 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
801 * have to send old ptlrpc_body to keep interoprability with these
804 * Only three kinds of server->client RPCs so far:
809 * XXX This should be removed whenever we drop the interoprability with
810 * the these old clients.
812 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
813 opcode == LDLM_GL_CALLBACK)
814 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
815 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
819 EXPORT_SYMBOL(ptlrpc_request_pack);
822 * Helper function to allocate new request on import \a imp
823 * and possibly using existing request from pool \a pool if provided.
824 * Returns allocated request structure with import field filled or
828 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
829 struct ptlrpc_request_pool *pool)
831 struct ptlrpc_request *request = NULL;
833 request = ptlrpc_request_cache_alloc(GFP_NOFS);
835 if (!request && pool)
836 request = ptlrpc_prep_req_from_pool(pool);
839 ptlrpc_cli_req_init(request);
841 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
842 LASSERT(imp != LP_POISON);
843 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
845 LASSERT(imp->imp_client != LP_POISON);
847 request->rq_import = class_import_get(imp);
849 CERROR("request allocation out of memory\n");
856 * Helper function for creating a request.
857 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
858 * buffer structures according to capsule template \a format.
859 * Returns allocated request structure pointer or NULL on error.
861 static struct ptlrpc_request *
862 ptlrpc_request_alloc_internal(struct obd_import *imp,
863 struct ptlrpc_request_pool * pool,
864 const struct req_format *format)
866 struct ptlrpc_request *request;
868 request = __ptlrpc_request_alloc(imp, pool);
872 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
873 req_capsule_set(&request->rq_pill, format);
878 * Allocate new request structure for import \a imp and initialize its
879 * buffer structure according to capsule template \a format.
881 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
882 const struct req_format *format)
884 return ptlrpc_request_alloc_internal(imp, NULL, format);
886 EXPORT_SYMBOL(ptlrpc_request_alloc);
889 * Allocate new request structure for import \a imp from pool \a pool and
890 * initialize its buffer structure according to capsule template \a format.
892 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
893 struct ptlrpc_request_pool * pool,
894 const struct req_format *format)
896 return ptlrpc_request_alloc_internal(imp, pool, format);
898 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
901 * For requests not from pool, free memory of the request structure.
902 * For requests obtained from a pool earlier, return request back to pool.
904 void ptlrpc_request_free(struct ptlrpc_request *request)
906 if (request->rq_pool)
907 __ptlrpc_free_req_to_pool(request);
909 ptlrpc_request_cache_free(request);
911 EXPORT_SYMBOL(ptlrpc_request_free);
914 * Allocate new request for operatione \a opcode and immediatelly pack it for
916 * Only used for simple requests like OBD_PING where the only important
917 * part of the request is operation itself.
918 * Returns allocated request or NULL on error.
920 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
921 const struct req_format *format,
922 __u32 version, int opcode)
924 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
928 rc = ptlrpc_request_pack(req, version, opcode);
930 ptlrpc_request_free(req);
936 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
939 * Allocate and initialize new request set structure on the current CPT.
940 * Returns a pointer to the newly allocated set structure or NULL on error.
942 struct ptlrpc_request_set *ptlrpc_prep_set(void)
944 struct ptlrpc_request_set *set;
948 cpt = cfs_cpt_current(cfs_cpt_table, 0);
949 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
952 atomic_set(&set->set_refcount, 1);
953 INIT_LIST_HEAD(&set->set_requests);
954 init_waitqueue_head(&set->set_waitq);
955 atomic_set(&set->set_new_count, 0);
956 atomic_set(&set->set_remaining, 0);
957 spin_lock_init(&set->set_new_req_lock);
958 INIT_LIST_HEAD(&set->set_new_requests);
959 set->set_max_inflight = UINT_MAX;
960 set->set_producer = NULL;
961 set->set_producer_arg = NULL;
966 EXPORT_SYMBOL(ptlrpc_prep_set);
969 * Allocate and initialize new request set structure with flow control
970 * extension. This extension allows to control the number of requests in-flight
971 * for the whole set. A callback function to generate requests must be provided
972 * and the request set will keep the number of requests sent over the wire to
974 * Returns a pointer to the newly allocated set structure or NULL on error.
976 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
980 struct ptlrpc_request_set *set;
982 set = ptlrpc_prep_set();
986 set->set_max_inflight = max;
987 set->set_producer = func;
988 set->set_producer_arg = arg;
994 * Wind down and free request set structure previously allocated with
996 * Ensures that all requests on the set have completed and removes
997 * all requests from the request list in a set.
998 * If any unsent request happen to be on the list, pretends that they got
999 * an error in flight and calls their completion handler.
1001 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1003 struct list_head *tmp;
1004 struct list_head *next;
1009 /* Requests on the set should either all be completed, or all be new */
1010 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1011 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1012 list_for_each(tmp, &set->set_requests) {
1013 struct ptlrpc_request *req =
1014 list_entry(tmp, struct ptlrpc_request,
1017 LASSERT(req->rq_phase == expected_phase);
1021 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1022 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1023 atomic_read(&set->set_remaining), n);
1025 list_for_each_safe(tmp, next, &set->set_requests) {
1026 struct ptlrpc_request *req =
1027 list_entry(tmp, struct ptlrpc_request,
1029 list_del_init(&req->rq_set_chain);
1031 LASSERT(req->rq_phase == expected_phase);
1033 if (req->rq_phase == RQ_PHASE_NEW) {
1034 ptlrpc_req_interpret(NULL, req, -EBADR);
1035 atomic_dec(&set->set_remaining);
1038 spin_lock(&req->rq_lock);
1040 req->rq_invalid_rqset = 0;
1041 spin_unlock(&req->rq_lock);
1043 ptlrpc_req_finished (req);
1046 LASSERT(atomic_read(&set->set_remaining) == 0);
1048 ptlrpc_reqset_put(set);
1051 EXPORT_SYMBOL(ptlrpc_set_destroy);
1054 * Add a new request to the general purpose request set.
1055 * Assumes request reference from the caller.
1057 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1058 struct ptlrpc_request *req)
1060 LASSERT(list_empty(&req->rq_set_chain));
1062 if (req->rq_allow_intr)
1063 set->set_allow_intr = 1;
1065 /* The set takes over the caller's request reference */
1066 list_add_tail(&req->rq_set_chain, &set->set_requests);
1068 atomic_inc(&set->set_remaining);
1069 req->rq_queued_time = ktime_get_seconds();
1071 if (req->rq_reqmsg != NULL)
1072 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1074 if (set->set_producer != NULL)
1075 /* If the request set has a producer callback, the RPC must be
1076 * sent straight away */
1077 ptlrpc_send_new_req(req);
1079 EXPORT_SYMBOL(ptlrpc_set_add_req);
1082 * Add a request to a request with dedicated server thread
1083 * and wake the thread to make any necessary processing.
1084 * Currently only used for ptlrpcd.
1086 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1087 struct ptlrpc_request *req)
1089 struct ptlrpc_request_set *set = pc->pc_set;
1092 LASSERT(req->rq_set == NULL);
1093 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1095 spin_lock(&set->set_new_req_lock);
1097 * The set takes over the caller's request reference.
1100 req->rq_queued_time = ktime_get_seconds();
1101 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1102 count = atomic_inc_return(&set->set_new_count);
1103 spin_unlock(&set->set_new_req_lock);
1105 /* Only need to call wakeup once for the first entry. */
1107 wake_up(&set->set_waitq);
1109 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1110 * guarantee the async RPC can be processed ASAP, we have
1111 * no other better choice. It maybe fixed in future. */
1112 for (i = 0; i < pc->pc_npartners; i++)
1113 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1118 * Based on the current state of the import, determine if the request
1119 * can be sent, is an error, or should be delayed.
1121 * Returns true if this request should be delayed. If false, and
1122 * *status is set, then the request can not be sent and *status is the
1123 * error code. If false and status is 0, then request can be sent.
1125 * The imp->imp_lock must be held.
1127 static int ptlrpc_import_delay_req(struct obd_import *imp,
1128 struct ptlrpc_request *req, int *status)
1133 LASSERT (status != NULL);
1136 if (req->rq_ctx_init || req->rq_ctx_fini) {
1137 /* always allow ctx init/fini rpc go through */
1138 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1139 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1141 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1142 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1144 /* pings or MDS-equivalent STATFS may safely race with umount */
1145 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1146 D_HA : D_ERROR, req, "IMP_CLOSED ");
1148 } else if (ptlrpc_send_limit_expired(req)) {
1149 /* probably doesn't need to be a D_ERROR after initial testing*/
1150 DEBUG_REQ(D_HA, req, "send limit expired ");
1151 *status = -ETIMEDOUT;
1152 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1153 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1154 /* allow CONNECT even if import is invalid */ ;
1155 if (atomic_read(&imp->imp_inval_count) != 0) {
1156 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1159 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1160 if (!imp->imp_deactive)
1161 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1162 *status = -ESHUTDOWN; /* bz 12940 */
1163 } else if (req->rq_import_generation != imp->imp_generation) {
1164 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1166 } else if (req->rq_send_state != imp->imp_state) {
1167 /* invalidate in progress - any requests should be drop */
1168 if (atomic_read(&imp->imp_inval_count) != 0) {
1169 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1171 } else if (req->rq_no_delay) {
1172 *status = -EWOULDBLOCK;
1173 } else if (req->rq_allow_replay &&
1174 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1175 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1176 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1177 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1178 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1188 * Decide if the error message should be printed to the console or not.
1189 * Makes its decision based on request type, status, and failure frequency.
1191 * \param[in] req request that failed and may need a console message
1193 * \retval false if no message should be printed
1194 * \retval true if console message should be printed
1196 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1198 LASSERT(req->rq_reqmsg != NULL);
1200 /* Suppress particular reconnect errors which are to be expected. */
1201 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1203 /* Suppress timed out reconnect requests */
1204 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1208 /* Suppress most unavailable/again reconnect requests, but
1209 * print occasionally so it is clear client is trying to
1210 * connect to a server where no target is running. */
1211 if ((err == -ENODEV || err == -EAGAIN) &&
1212 req->rq_import->imp_conn_cnt % 30 != 20)
1216 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1217 /* -EAGAIN is normal when using POSIX flocks */
1220 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1221 (req->rq_xid & 0xf) != 10)
1222 /* Suppress most ping requests, they may fail occasionally */
1229 * Check request processing status.
1230 * Returns the status.
1232 static int ptlrpc_check_status(struct ptlrpc_request *req)
1237 err = lustre_msg_get_status(req->rq_repmsg);
1238 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1239 struct obd_import *imp = req->rq_import;
1240 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1241 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1243 if (ptlrpc_console_allow(req, opc, err))
1244 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1245 "failed: rc = %d\n",
1246 imp->imp_obd->obd_name,
1248 libcfs_nid2str(nid), err);
1249 RETURN(err < 0 ? err : -EINVAL);
1253 DEBUG_REQ(D_INFO, req, "status is %d", err);
1254 } else if (err > 0) {
1255 /* XXX: translate this error from net to host */
1256 DEBUG_REQ(D_INFO, req, "status is %d", err);
1263 * save pre-versions of objects into request for replay.
1264 * Versions are obtained from server reply.
1267 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1269 struct lustre_msg *repmsg = req->rq_repmsg;
1270 struct lustre_msg *reqmsg = req->rq_reqmsg;
1271 __u64 *versions = lustre_msg_get_versions(repmsg);
1274 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1278 lustre_msg_set_versions(reqmsg, versions);
1279 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1280 versions[0], versions[1]);
1285 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1287 struct ptlrpc_request *req;
1289 assert_spin_locked(&imp->imp_lock);
1290 if (list_empty(&imp->imp_unreplied_list))
1293 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1295 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1297 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1298 imp->imp_known_replied_xid = req->rq_xid - 1;
1300 return req->rq_xid - 1;
1304 * Callback function called when client receives RPC reply for \a req.
1305 * Returns 0 on success or error code.
1306 * The return alue would be assigned to req->rq_status by the caller
1307 * as request processing status.
1308 * This function also decides if the request needs to be saved for later replay.
1310 static int after_reply(struct ptlrpc_request *req)
1312 struct obd_import *imp = req->rq_import;
1313 struct obd_device *obd = req->rq_import->imp_obd;
1320 LASSERT(obd != NULL);
1321 /* repbuf must be unlinked */
1322 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1324 if (req->rq_reply_truncated) {
1325 if (ptlrpc_no_resend(req)) {
1326 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1327 " expected: %d, actual size: %d",
1328 req->rq_nob_received, req->rq_repbuf_len);
1332 sptlrpc_cli_free_repbuf(req);
1333 /* Pass the required reply buffer size (include
1334 * space for early reply).
1335 * NB: no need to roundup because alloc_repbuf
1336 * will roundup it */
1337 req->rq_replen = req->rq_nob_received;
1338 req->rq_nob_received = 0;
1339 spin_lock(&req->rq_lock);
1341 spin_unlock(&req->rq_lock);
1345 work_start = ktime_get_real();
1346 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1349 * NB Until this point, the whole of the incoming message,
1350 * including buflens, status etc is in the sender's byte order.
1352 rc = sptlrpc_cli_unwrap_reply(req);
1354 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1359 * Security layer unwrap might ask resend this request.
1364 rc = unpack_reply(req);
1368 /* retry indefinitely on EINPROGRESS */
1369 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1370 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1371 time64_t now = ktime_get_real_seconds();
1373 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1374 spin_lock(&req->rq_lock);
1376 spin_unlock(&req->rq_lock);
1377 req->rq_nr_resend++;
1379 /* Readjust the timeout for current conditions */
1380 ptlrpc_at_set_req_timeout(req);
1381 /* delay resend to give a chance to the server to get ready.
1382 * The delay is increased by 1s on every resend and is capped to
1383 * the current request timeout (i.e. obd_timeout if AT is off,
1384 * or AT service time x 125% + 5s, see at_est2timeout) */
1385 if (req->rq_nr_resend > req->rq_timeout)
1386 req->rq_sent = now + req->rq_timeout;
1388 req->rq_sent = now + req->rq_nr_resend;
1390 /* Resend for EINPROGRESS will use a new XID */
1391 spin_lock(&imp->imp_lock);
1392 list_del_init(&req->rq_unreplied_list);
1393 spin_unlock(&imp->imp_lock);
1398 if (obd->obd_svc_stats != NULL) {
1399 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1401 ptlrpc_lprocfs_rpc_sent(req, timediff);
1404 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1405 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1406 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1407 lustre_msg_get_type(req->rq_repmsg));
1411 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1412 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1413 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1414 ptlrpc_at_adj_net_latency(req,
1415 lustre_msg_get_service_time(req->rq_repmsg));
1417 rc = ptlrpc_check_status(req);
1418 imp->imp_connect_error = rc;
1422 * Either we've been evicted, or the server has failed for
1423 * some reason. Try to reconnect, and if that fails, punt to
1426 if (ptlrpc_recoverable_error(rc)) {
1427 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1428 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1431 ptlrpc_request_handle_notconn(req);
1436 * Let's look if server sent slv. Do it only for RPC with
1439 ldlm_cli_update_pool(req);
1443 * Store transno in reqmsg for replay.
1445 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1446 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1447 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1450 if (imp->imp_replayable) {
1451 spin_lock(&imp->imp_lock);
1453 * No point in adding already-committed requests to the replay
1454 * list, we will just remove them immediately. b=9829
1456 if (req->rq_transno != 0 &&
1458 lustre_msg_get_last_committed(req->rq_repmsg) ||
1460 /** version recovery */
1461 ptlrpc_save_versions(req);
1462 ptlrpc_retain_replayable_request(req, imp);
1463 } else if (req->rq_commit_cb != NULL &&
1464 list_empty(&req->rq_replay_list)) {
1465 /* NB: don't call rq_commit_cb if it's already on
1466 * rq_replay_list, ptlrpc_free_committed() will call
1467 * it later, see LU-3618 for details */
1468 spin_unlock(&imp->imp_lock);
1469 req->rq_commit_cb(req);
1470 spin_lock(&imp->imp_lock);
1474 * Replay-enabled imports return commit-status information.
1476 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1477 if (likely(committed > imp->imp_peer_committed_transno))
1478 imp->imp_peer_committed_transno = committed;
1480 ptlrpc_free_committed(imp);
1482 if (!list_empty(&imp->imp_replay_list)) {
1483 struct ptlrpc_request *last;
1485 last = list_entry(imp->imp_replay_list.prev,
1486 struct ptlrpc_request,
1489 * Requests with rq_replay stay on the list even if no
1490 * commit is expected.
1492 if (last->rq_transno > imp->imp_peer_committed_transno)
1493 ptlrpc_pinger_commit_expected(imp);
1496 spin_unlock(&imp->imp_lock);
1503 * Helper function to send request \a req over the network for the first time
1504 * Also adjusts request phase.
1505 * Returns 0 on success or error code.
1507 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1509 struct obd_import *imp = req->rq_import;
1514 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1516 /* do not try to go further if there is not enough memory in enc_pool */
1517 if (req->rq_sent && req->rq_bulk != NULL)
1518 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1519 pool_is_at_full_capacity())
1522 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1523 (!req->rq_generation_set ||
1524 req->rq_import_generation == imp->imp_generation))
1527 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1529 spin_lock(&imp->imp_lock);
1531 LASSERT(req->rq_xid != 0);
1532 LASSERT(!list_empty(&req->rq_unreplied_list));
1534 if (!req->rq_generation_set)
1535 req->rq_import_generation = imp->imp_generation;
1537 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1538 spin_lock(&req->rq_lock);
1539 req->rq_waiting = 1;
1540 spin_unlock(&req->rq_lock);
1542 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1543 ptlrpc_import_state_name(req->rq_send_state),
1544 ptlrpc_import_state_name(imp->imp_state));
1545 LASSERT(list_empty(&req->rq_list));
1546 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1547 atomic_inc(&req->rq_import->imp_inflight);
1548 spin_unlock(&imp->imp_lock);
1553 spin_unlock(&imp->imp_lock);
1554 req->rq_status = rc;
1555 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1559 LASSERT(list_empty(&req->rq_list));
1560 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1561 atomic_inc(&req->rq_import->imp_inflight);
1563 /* find the known replied XID from the unreplied list, CONNECT
1564 * and DISCONNECT requests are skipped to make the sanity check
1565 * on server side happy. see process_req_last_xid().
1567 * For CONNECT: Because replay requests have lower XID, it'll
1568 * break the sanity check if CONNECT bump the exp_last_xid on
1571 * For DISCONNECT: Since client will abort inflight RPC before
1572 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1573 * than the inflight RPC.
1575 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1576 min_xid = ptlrpc_known_replied_xid(imp);
1577 spin_unlock(&imp->imp_lock);
1579 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1581 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1583 rc = sptlrpc_req_refresh_ctx(req, -1);
1586 req->rq_status = rc;
1589 spin_lock(&req->rq_lock);
1590 req->rq_wait_ctx = 1;
1591 spin_unlock(&req->rq_lock);
1596 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1597 " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1598 imp->imp_obd->obd_uuid.uuid,
1599 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1600 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1601 lustre_msg_get_opc(req->rq_reqmsg));
1603 rc = ptl_send_rpc(req, 0);
1604 if (rc == -ENOMEM) {
1605 spin_lock(&imp->imp_lock);
1606 if (!list_empty(&req->rq_list)) {
1607 list_del_init(&req->rq_list);
1608 atomic_dec(&req->rq_import->imp_inflight);
1610 spin_unlock(&imp->imp_lock);
1611 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1615 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1616 spin_lock(&req->rq_lock);
1617 req->rq_net_err = 1;
1618 spin_unlock(&req->rq_lock);
1624 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1629 LASSERT(set->set_producer != NULL);
1631 remaining = atomic_read(&set->set_remaining);
1633 /* populate the ->set_requests list with requests until we
1634 * reach the maximum number of RPCs in flight for this set */
1635 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1636 rc = set->set_producer(set, set->set_producer_arg);
1637 if (rc == -ENOENT) {
1638 /* no more RPC to produce */
1639 set->set_producer = NULL;
1640 set->set_producer_arg = NULL;
1645 RETURN((atomic_read(&set->set_remaining) - remaining));
1649 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1650 * and no more replies are expected.
1651 * (it is possible to get less replies than requests sent e.g. due to timed out
1652 * requests or requests that we had trouble to send out)
1654 * NOTE: This function contains a potential schedule point (cond_resched()).
1656 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1658 struct list_head *tmp, *next;
1659 struct list_head comp_reqs;
1660 int force_timer_recalc = 0;
1663 if (atomic_read(&set->set_remaining) == 0)
1666 INIT_LIST_HEAD(&comp_reqs);
1667 list_for_each_safe(tmp, next, &set->set_requests) {
1668 struct ptlrpc_request *req =
1669 list_entry(tmp, struct ptlrpc_request,
1671 struct obd_import *imp = req->rq_import;
1672 int unregistered = 0;
1676 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1677 list_move_tail(&req->rq_set_chain, &comp_reqs);
1681 /* This schedule point is mainly for the ptlrpcd caller of this
1682 * function. Most ptlrpc sets are not long-lived and unbounded
1683 * in length, but at the least the set used by the ptlrpcd is.
1684 * Since the processing time is unbounded, we need to insert an
1685 * explicit schedule point to make the thread well-behaved.
1689 /* If the caller requires to allow to be interpreted by force
1690 * and it has really been interpreted, then move the request
1691 * to RQ_PHASE_INTERPRET phase in spite of what the current
1693 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1694 req->rq_status = -EINTR;
1695 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1697 /* Since it is interpreted and we have to wait for
1698 * the reply to be unlinked, then use sync mode. */
1701 GOTO(interpret, req->rq_status);
1704 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1705 force_timer_recalc = 1;
1707 /* delayed send - skip */
1708 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1711 /* delayed resend - skip */
1712 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1713 req->rq_sent > ktime_get_real_seconds())
1716 if (!(req->rq_phase == RQ_PHASE_RPC ||
1717 req->rq_phase == RQ_PHASE_BULK ||
1718 req->rq_phase == RQ_PHASE_INTERPRET ||
1719 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1720 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1721 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1725 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1726 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1727 LASSERT(req->rq_next_phase != req->rq_phase);
1728 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1730 if (req->rq_req_deadline &&
1731 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1732 req->rq_req_deadline = 0;
1733 if (req->rq_reply_deadline &&
1734 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1735 req->rq_reply_deadline = 0;
1736 if (req->rq_bulk_deadline &&
1737 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1738 req->rq_bulk_deadline = 0;
1741 * Skip processing until reply is unlinked. We
1742 * can't return to pool before that and we can't
1743 * call interpret before that. We need to make
1744 * sure that all rdma transfers finished and will
1745 * not corrupt any data.
1747 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1748 ptlrpc_client_recv_or_unlink(req))
1750 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1751 ptlrpc_client_bulk_active(req))
1755 * Turn fail_loc off to prevent it from looping
1758 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1759 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1762 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1763 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1768 * Move to next phase if reply was successfully
1771 ptlrpc_rqphase_move(req, req->rq_next_phase);
1774 if (req->rq_phase == RQ_PHASE_INTERPRET)
1775 GOTO(interpret, req->rq_status);
1778 * Note that this also will start async reply unlink.
1780 if (req->rq_net_err && !req->rq_timedout) {
1781 ptlrpc_expire_one_request(req, 1);
1784 * Check if we still need to wait for unlink.
1786 if (ptlrpc_client_recv_or_unlink(req) ||
1787 ptlrpc_client_bulk_active(req))
1789 /* If there is no need to resend, fail it now. */
1790 if (req->rq_no_resend) {
1791 if (req->rq_status == 0)
1792 req->rq_status = -EIO;
1793 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1794 GOTO(interpret, req->rq_status);
1801 spin_lock(&req->rq_lock);
1802 req->rq_replied = 0;
1803 spin_unlock(&req->rq_lock);
1804 if (req->rq_status == 0)
1805 req->rq_status = -EIO;
1806 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1807 GOTO(interpret, req->rq_status);
1810 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1811 * so it sets rq_intr regardless of individual rpc
1812 * timeouts. The synchronous IO waiting path sets
1813 * rq_intr irrespective of whether ptlrpcd
1814 * has seen a timeout. Our policy is to only interpret
1815 * interrupted rpcs after they have timed out, so we
1816 * need to enforce that here.
1819 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1820 req->rq_wait_ctx)) {
1821 req->rq_status = -EINTR;
1822 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1823 GOTO(interpret, req->rq_status);
1826 if (req->rq_phase == RQ_PHASE_RPC) {
1827 if (req->rq_timedout || req->rq_resend ||
1828 req->rq_waiting || req->rq_wait_ctx) {
1831 if (!ptlrpc_unregister_reply(req, 1)) {
1832 ptlrpc_unregister_bulk(req, 1);
1836 spin_lock(&imp->imp_lock);
1837 if (ptlrpc_import_delay_req(imp, req, &status)){
1838 /* put on delay list - only if we wait
1839 * recovery finished - before send */
1840 list_del_init(&req->rq_list);
1841 list_add_tail(&req->rq_list,
1844 spin_unlock(&imp->imp_lock);
1849 req->rq_status = status;
1850 ptlrpc_rqphase_move(req,
1851 RQ_PHASE_INTERPRET);
1852 spin_unlock(&imp->imp_lock);
1853 GOTO(interpret, req->rq_status);
1855 if (ptlrpc_no_resend(req) &&
1856 !req->rq_wait_ctx) {
1857 req->rq_status = -ENOTCONN;
1858 ptlrpc_rqphase_move(req,
1859 RQ_PHASE_INTERPRET);
1860 spin_unlock(&imp->imp_lock);
1861 GOTO(interpret, req->rq_status);
1864 list_del_init(&req->rq_list);
1865 list_add_tail(&req->rq_list,
1866 &imp->imp_sending_list);
1868 spin_unlock(&imp->imp_lock);
1870 spin_lock(&req->rq_lock);
1871 req->rq_waiting = 0;
1872 spin_unlock(&req->rq_lock);
1874 if (req->rq_timedout || req->rq_resend) {
1875 /* This is re-sending anyways,
1876 * let's mark req as resend. */
1877 spin_lock(&req->rq_lock);
1879 spin_unlock(&req->rq_lock);
1881 if (req->rq_bulk != NULL &&
1882 !ptlrpc_unregister_bulk(req, 1))
1886 * rq_wait_ctx is only touched by ptlrpcd,
1887 * so no lock is needed here.
1889 status = sptlrpc_req_refresh_ctx(req, -1);
1892 req->rq_status = status;
1893 spin_lock(&req->rq_lock);
1894 req->rq_wait_ctx = 0;
1895 spin_unlock(&req->rq_lock);
1896 force_timer_recalc = 1;
1898 spin_lock(&req->rq_lock);
1899 req->rq_wait_ctx = 1;
1900 spin_unlock(&req->rq_lock);
1905 spin_lock(&req->rq_lock);
1906 req->rq_wait_ctx = 0;
1907 spin_unlock(&req->rq_lock);
1910 rc = ptl_send_rpc(req, 0);
1911 if (rc == -ENOMEM) {
1912 spin_lock(&imp->imp_lock);
1913 if (!list_empty(&req->rq_list))
1914 list_del_init(&req->rq_list);
1915 spin_unlock(&imp->imp_lock);
1916 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1920 DEBUG_REQ(D_HA, req,
1921 "send failed: rc = %d", rc);
1922 force_timer_recalc = 1;
1923 spin_lock(&req->rq_lock);
1924 req->rq_net_err = 1;
1925 spin_unlock(&req->rq_lock);
1928 /* need to reset the timeout */
1929 force_timer_recalc = 1;
1932 spin_lock(&req->rq_lock);
1934 if (ptlrpc_client_early(req)) {
1935 ptlrpc_at_recv_early_reply(req);
1936 spin_unlock(&req->rq_lock);
1940 /* Still waiting for a reply? */
1941 if (ptlrpc_client_recv(req)) {
1942 spin_unlock(&req->rq_lock);
1946 /* Did we actually receive a reply? */
1947 if (!ptlrpc_client_replied(req)) {
1948 spin_unlock(&req->rq_lock);
1952 spin_unlock(&req->rq_lock);
1954 /* unlink from net because we are going to
1955 * swab in-place of reply buffer */
1956 unregistered = ptlrpc_unregister_reply(req, 1);
1960 req->rq_status = after_reply(req);
1964 /* If there is no bulk associated with this request,
1965 * then we're done and should let the interpreter
1966 * process the reply. Similarly if the RPC returned
1967 * an error, and therefore the bulk will never arrive.
1969 if (req->rq_bulk == NULL || req->rq_status < 0) {
1970 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1971 GOTO(interpret, req->rq_status);
1974 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1977 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1978 if (ptlrpc_client_bulk_active(req))
1981 if (req->rq_bulk->bd_failure) {
1982 /* The RPC reply arrived OK, but the bulk screwed
1983 * up! Dead weird since the server told us the RPC
1984 * was good after getting the REPLY for her GET or
1985 * the ACK for her PUT. */
1986 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1987 req->rq_status = -EIO;
1990 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1993 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1995 /* This moves to "unregistering" phase we need to wait for
1997 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
1998 /* start async bulk unlink too */
1999 ptlrpc_unregister_bulk(req, 1);
2003 if (!ptlrpc_unregister_bulk(req, async))
2006 /* When calling interpret receiving already should be
2008 LASSERT(!req->rq_receiving_reply);
2010 ptlrpc_req_interpret(env, req, req->rq_status);
2012 if (ptlrpcd_check_work(req)) {
2013 atomic_dec(&set->set_remaining);
2016 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2018 if (req->rq_reqmsg != NULL)
2020 "Completed RPC pname:cluuid:pid:xid:nid:"
2021 "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
2022 imp->imp_obd->obd_uuid.uuid,
2023 lustre_msg_get_status(req->rq_reqmsg),
2025 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2026 lustre_msg_get_opc(req->rq_reqmsg));
2028 spin_lock(&imp->imp_lock);
2029 /* Request already may be not on sending or delaying list. This
2030 * may happen in the case of marking it erroneous for the case
2031 * ptlrpc_import_delay_req(req, status) find it impossible to
2032 * allow sending this rpc and returns *status != 0. */
2033 if (!list_empty(&req->rq_list)) {
2034 list_del_init(&req->rq_list);
2035 atomic_dec(&imp->imp_inflight);
2037 list_del_init(&req->rq_unreplied_list);
2038 spin_unlock(&imp->imp_lock);
2040 atomic_dec(&set->set_remaining);
2041 wake_up_all(&imp->imp_recovery_waitq);
2043 if (set->set_producer) {
2044 /* produce a new request if possible */
2045 if (ptlrpc_set_producer(set) > 0)
2046 force_timer_recalc = 1;
2048 /* free the request that has just been completed
2049 * in order not to pollute set->set_requests */
2050 list_del_init(&req->rq_set_chain);
2051 spin_lock(&req->rq_lock);
2053 req->rq_invalid_rqset = 0;
2054 spin_unlock(&req->rq_lock);
2056 /* record rq_status to compute the final status later */
2057 if (req->rq_status != 0)
2058 set->set_rc = req->rq_status;
2059 ptlrpc_req_finished(req);
2061 list_move_tail(&req->rq_set_chain, &comp_reqs);
2065 /* move completed request at the head of list so it's easier for
2066 * caller to find them */
2067 list_splice(&comp_reqs, &set->set_requests);
2069 /* If we hit an error, we want to recover promptly. */
2070 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2072 EXPORT_SYMBOL(ptlrpc_check_set);
2075 * Time out request \a req. is \a async_unlink is set, that means do not wait
2076 * until LNet actually confirms network buffer unlinking.
2077 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2079 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2081 struct obd_import *imp = req->rq_import;
2082 unsigned int debug_mask = D_RPCTRACE;
2086 spin_lock(&req->rq_lock);
2087 req->rq_timedout = 1;
2088 spin_unlock(&req->rq_lock);
2090 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2091 lustre_msg_get_status(req->rq_reqmsg)))
2092 debug_mask = D_WARNING;
2093 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2094 req->rq_net_err ? "failed due to network error" :
2095 ((req->rq_real_sent == 0 ||
2096 req->rq_real_sent < req->rq_sent ||
2097 req->rq_real_sent >= req->rq_deadline) ?
2098 "timed out for sent delay" : "timed out for slow reply"),
2099 (s64)req->rq_sent, (s64)req->rq_real_sent);
2101 if (imp != NULL && obd_debug_peer_on_timeout)
2102 LNetDebugPeer(imp->imp_connection->c_peer);
2104 ptlrpc_unregister_reply(req, async_unlink);
2105 ptlrpc_unregister_bulk(req, async_unlink);
2107 if (obd_dump_on_timeout)
2108 libcfs_debug_dumplog();
2111 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2115 atomic_inc(&imp->imp_timeouts);
2117 /* The DLM server doesn't want recovery run on its imports. */
2118 if (imp->imp_dlm_fake)
2121 /* If this request is for recovery or other primordial tasks,
2122 * then error it out here. */
2123 if (req->rq_ctx_init || req->rq_ctx_fini ||
2124 req->rq_send_state != LUSTRE_IMP_FULL ||
2125 imp->imp_obd->obd_no_recov) {
2126 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2127 ptlrpc_import_state_name(req->rq_send_state),
2128 ptlrpc_import_state_name(imp->imp_state));
2129 spin_lock(&req->rq_lock);
2130 req->rq_status = -ETIMEDOUT;
2132 spin_unlock(&req->rq_lock);
2136 /* if a request can't be resent we can't wait for an answer after
2138 if (ptlrpc_no_resend(req)) {
2139 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2143 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2149 * Time out all uncompleted requests in request set pointed by \a data
2150 * Callback used when waiting on sets with l_wait_event.
2153 int ptlrpc_expired_set(void *data)
2155 struct ptlrpc_request_set *set = data;
2156 struct list_head *tmp;
2157 time64_t now = ktime_get_real_seconds();
2160 LASSERT(set != NULL);
2163 * A timeout expired. See which reqs it applies to...
2165 list_for_each(tmp, &set->set_requests) {
2166 struct ptlrpc_request *req =
2167 list_entry(tmp, struct ptlrpc_request,
2170 /* don't expire request waiting for context */
2171 if (req->rq_wait_ctx)
2174 /* Request in-flight? */
2175 if (!((req->rq_phase == RQ_PHASE_RPC &&
2176 !req->rq_waiting && !req->rq_resend) ||
2177 (req->rq_phase == RQ_PHASE_BULK)))
2180 if (req->rq_timedout || /* already dealt with */
2181 req->rq_deadline > now) /* not expired */
2184 /* Deal with this guy. Do it asynchronously to not block
2185 * ptlrpcd thread. */
2186 ptlrpc_expire_one_request(req, 1);
2190 * When waiting for a whole set, we always break out of the
2191 * sleep so we can recalculate the timeout, or enable interrupts
2192 * if everyone's timed out.
2198 * Sets rq_intr flag in \a req under spinlock.
2200 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2202 spin_lock(&req->rq_lock);
2204 spin_unlock(&req->rq_lock);
2206 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2209 * Interrupts (sets interrupted flag) all uncompleted requests in
2210 * a set \a data. Callback for l_wait_event for interruptible waits.
2212 static void ptlrpc_interrupted_set(void *data)
2214 struct ptlrpc_request_set *set = data;
2215 struct list_head *tmp;
2217 LASSERT(set != NULL);
2218 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2220 list_for_each(tmp, &set->set_requests) {
2221 struct ptlrpc_request *req =
2222 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2227 if (req->rq_phase != RQ_PHASE_RPC &&
2228 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2229 !req->rq_allow_intr)
2232 ptlrpc_mark_interrupted(req);
2237 * Get the smallest timeout in the set; this does NOT set a timeout.
2239 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2241 struct list_head *tmp;
2242 time64_t now = ktime_get_real_seconds();
2244 struct ptlrpc_request *req;
2248 list_for_each(tmp, &set->set_requests) {
2249 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2252 * Request in-flight?
2254 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2255 (req->rq_phase == RQ_PHASE_BULK) ||
2256 (req->rq_phase == RQ_PHASE_NEW)))
2260 * Already timed out.
2262 if (req->rq_timedout)
2268 if (req->rq_wait_ctx)
2271 if (req->rq_phase == RQ_PHASE_NEW)
2272 deadline = req->rq_sent;
2273 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2274 deadline = req->rq_sent;
2276 deadline = req->rq_sent + req->rq_timeout;
2278 if (deadline <= now) /* actually expired already */
2279 timeout = 1; /* ASAP */
2280 else if (timeout == 0 || timeout > deadline - now)
2281 timeout = deadline - now;
2287 * Send all unset request from the set and then wait untill all
2288 * requests in the set complete (either get a reply, timeout, get an
2289 * error or otherwise be interrupted).
2290 * Returns 0 on success or error code otherwise.
2292 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2294 struct list_head *tmp;
2295 struct ptlrpc_request *req;
2296 struct l_wait_info lwi;
2301 if (set->set_producer)
2302 (void)ptlrpc_set_producer(set);
2304 list_for_each(tmp, &set->set_requests) {
2305 req = list_entry(tmp, struct ptlrpc_request,
2307 if (req->rq_phase == RQ_PHASE_NEW)
2308 (void)ptlrpc_send_new_req(req);
2311 if (list_empty(&set->set_requests))
2315 timeout = ptlrpc_set_next_timeout(set);
2317 /* wait until all complete, interrupted, or an in-flight
2319 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2322 if ((timeout == 0 && !signal_pending(current)) ||
2323 set->set_allow_intr)
2324 /* No requests are in-flight (ether timed out
2325 * or delayed), so we can allow interrupts.
2326 * We still want to block for a limited time,
2327 * so we allow interrupts during the timeout. */
2328 lwi = LWI_TIMEOUT_INTR_ALL(
2329 cfs_time_seconds(timeout ? timeout : 1),
2331 ptlrpc_interrupted_set, set);
2334 * At least one request is in flight, so no
2335 * interrupts are allowed. Wait until all
2336 * complete, or an in-flight req times out.
2338 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2339 ptlrpc_expired_set, set);
2341 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2343 /* LU-769 - if we ignored the signal because it was already
2344 * pending when we started, we need to handle it now or we risk
2345 * it being ignored forever */
2346 if (rc == -ETIMEDOUT &&
2347 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2348 signal_pending(current)) {
2349 sigset_t blocked_sigs =
2350 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2352 /* In fact we only interrupt for the "fatal" signals
2353 * like SIGINT or SIGKILL. We still ignore less
2354 * important signals since ptlrpc set is not easily
2355 * reentrant from userspace again */
2356 if (signal_pending(current))
2357 ptlrpc_interrupted_set(set);
2358 cfs_restore_sigs(blocked_sigs);
2361 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2363 /* -EINTR => all requests have been flagged rq_intr so next
2365 * -ETIMEDOUT => someone timed out. When all reqs have
2366 * timed out, signals are enabled allowing completion with
2368 * I don't really care if we go once more round the loop in
2369 * the error cases -eeb. */
2370 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2371 list_for_each(tmp, &set->set_requests) {
2372 req = list_entry(tmp, struct ptlrpc_request,
2374 spin_lock(&req->rq_lock);
2375 req->rq_invalid_rqset = 1;
2376 spin_unlock(&req->rq_lock);
2379 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2381 LASSERT(atomic_read(&set->set_remaining) == 0);
2383 rc = set->set_rc; /* rq_status of already freed requests if any */
2384 list_for_each(tmp, &set->set_requests) {
2385 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2387 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2388 if (req->rq_status != 0)
2389 rc = req->rq_status;
2394 EXPORT_SYMBOL(ptlrpc_set_wait);
2397 * Helper fuction for request freeing.
2398 * Called when request count reached zero and request needs to be freed.
2399 * Removes request from all sorts of sending/replay lists it might be on,
2400 * frees network buffers if any are present.
2401 * If \a locked is set, that means caller is already holding import imp_lock
2402 * and so we no longer need to reobtain it (for certain lists manipulations)
2404 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2408 if (request == NULL)
2411 LASSERT(!request->rq_srv_req);
2412 LASSERT(request->rq_export == NULL);
2413 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2414 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2415 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2416 LASSERTF(!request->rq_replay, "req %p\n", request);
2418 req_capsule_fini(&request->rq_pill);
2420 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2421 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2422 if (request->rq_import != NULL) {
2424 spin_lock(&request->rq_import->imp_lock);
2425 list_del_init(&request->rq_replay_list);
2426 list_del_init(&request->rq_unreplied_list);
2428 spin_unlock(&request->rq_import->imp_lock);
2430 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2432 if (atomic_read(&request->rq_refcount) != 0) {
2433 DEBUG_REQ(D_ERROR, request,
2434 "freeing request with nonzero refcount");
2438 if (request->rq_repbuf != NULL)
2439 sptlrpc_cli_free_repbuf(request);
2441 if (request->rq_import != NULL) {
2442 class_import_put(request->rq_import);
2443 request->rq_import = NULL;
2445 if (request->rq_bulk != NULL)
2446 ptlrpc_free_bulk(request->rq_bulk);
2448 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2449 sptlrpc_cli_free_reqbuf(request);
2451 if (request->rq_cli_ctx)
2452 sptlrpc_req_put_ctx(request, !locked);
2454 if (request->rq_pool)
2455 __ptlrpc_free_req_to_pool(request);
2457 ptlrpc_request_cache_free(request);
2461 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2463 * Drop one request reference. Must be called with import imp_lock held.
2464 * When reference count drops to zero, request is freed.
2466 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2468 assert_spin_locked(&request->rq_import->imp_lock);
2469 (void)__ptlrpc_req_finished(request, 1);
2474 * Drops one reference count for request \a request.
2475 * \a locked set indicates that caller holds import imp_lock.
2476 * Frees the request whe reference count reaches zero.
2478 * \retval 1 the request is freed
2479 * \retval 0 some others still hold references on the request
2481 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2489 LASSERT(request != LP_POISON);
2490 LASSERT(request->rq_reqmsg != LP_POISON);
2492 DEBUG_REQ(D_INFO, request, "refcount now %u",
2493 atomic_read(&request->rq_refcount) - 1);
2495 spin_lock(&request->rq_lock);
2496 count = atomic_dec_return(&request->rq_refcount);
2497 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2499 /* For open RPC, the client does not know the EA size (LOV, ACL, and
2500 * so on) before replied, then the client has to reserve very large
2501 * reply buffer. Such buffer will not be released until the RPC freed.
2502 * Since The open RPC is replayable, we need to keep it in the replay
2503 * list until close. If there are a lot of files opened concurrently,
2504 * then the client may be OOM.
2506 * If fact, it is unnecessary to keep reply buffer for open replay,
2507 * related EAs have already been saved via mdc_save_lovea() before
2508 * coming here. So it is safe to free the reply buffer some earlier
2509 * before releasing the RPC to avoid client OOM. LU-9514 */
2510 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2511 spin_lock(&request->rq_early_free_lock);
2512 sptlrpc_cli_free_repbuf(request);
2513 request->rq_repbuf = NULL;
2514 request->rq_repbuf_len = 0;
2515 request->rq_repdata = NULL;
2516 request->rq_reqdata_len = 0;
2517 spin_unlock(&request->rq_early_free_lock);
2519 spin_unlock(&request->rq_lock);
2522 __ptlrpc_free_req(request, locked);
2528 * Drops one reference count for a request.
2530 void ptlrpc_req_finished(struct ptlrpc_request *request)
2532 __ptlrpc_req_finished(request, 0);
2534 EXPORT_SYMBOL(ptlrpc_req_finished);
2537 * Returns xid of a \a request
2539 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2541 return request->rq_xid;
2543 EXPORT_SYMBOL(ptlrpc_req_xid);
2546 * Disengage the client's reply buffer from the network
2547 * NB does _NOT_ unregister any client-side bulk.
2548 * IDEMPOTENT, but _not_ safe against concurrent callers.
2549 * The request owner (i.e. the thread doing the I/O) must call...
2550 * Returns 0 on success or 1 if unregistering cannot be made.
2552 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2555 struct l_wait_info lwi;
2560 LASSERT(!in_interrupt());
2562 /* Let's setup deadline for reply unlink. */
2563 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2564 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2565 request->rq_reply_deadline = ktime_get_real_seconds() +
2569 * Nothing left to do.
2571 if (!ptlrpc_client_recv_or_unlink(request))
2574 LNetMDUnlink(request->rq_reply_md_h);
2577 * Let's check it once again.
2579 if (!ptlrpc_client_recv_or_unlink(request))
2582 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2583 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2586 * Do not wait for unlink to finish.
2592 * We have to l_wait_event() whatever the result, to give liblustre
2593 * a chance to run reply_in_callback(), and to make sure we've
2594 * unlinked before returning a req to the pool.
2597 /* The wq argument is ignored by user-space wait_event macros */
2598 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2599 &request->rq_set->set_waitq :
2600 &request->rq_reply_waitq;
2601 /* Network access will complete in finite time but the HUGE
2602 * timeout lets us CWARN for visibility of sluggish NALs */
2603 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2604 cfs_time_seconds(1), NULL, NULL);
2605 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2608 ptlrpc_rqphase_move(request, request->rq_next_phase);
2612 LASSERT(rc == -ETIMEDOUT);
2613 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2614 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2615 request->rq_receiving_reply,
2616 request->rq_req_unlinked,
2617 request->rq_reply_unlinked);
2622 static void ptlrpc_free_request(struct ptlrpc_request *req)
2624 spin_lock(&req->rq_lock);
2626 spin_unlock(&req->rq_lock);
2628 if (req->rq_commit_cb != NULL)
2629 req->rq_commit_cb(req);
2630 list_del_init(&req->rq_replay_list);
2632 __ptlrpc_req_finished(req, 1);
2636 * the request is committed and dropped from the replay list of its import
2638 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2640 struct obd_import *imp = req->rq_import;
2642 spin_lock(&imp->imp_lock);
2643 if (list_empty(&req->rq_replay_list)) {
2644 spin_unlock(&imp->imp_lock);
2648 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2649 ptlrpc_free_request(req);
2651 spin_unlock(&imp->imp_lock);
2653 EXPORT_SYMBOL(ptlrpc_request_committed);
2656 * Iterates through replay_list on import and prunes
2657 * all requests have transno smaller than last_committed for the
2658 * import and don't have rq_replay set.
2659 * Since requests are sorted in transno order, stops when meetign first
2660 * transno bigger than last_committed.
2661 * caller must hold imp->imp_lock
2663 void ptlrpc_free_committed(struct obd_import *imp)
2665 struct ptlrpc_request *req, *saved;
2666 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2667 bool skip_committed_list = true;
2670 LASSERT(imp != NULL);
2671 assert_spin_locked(&imp->imp_lock);
2673 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2674 imp->imp_generation == imp->imp_last_generation_checked) {
2675 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2676 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2679 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2680 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2681 imp->imp_generation);
2683 if (imp->imp_generation != imp->imp_last_generation_checked ||
2684 imp->imp_last_transno_checked == 0)
2685 skip_committed_list = false;
2687 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2688 imp->imp_last_generation_checked = imp->imp_generation;
2690 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2692 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2693 LASSERT(req != last_req);
2696 if (req->rq_transno == 0) {
2697 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2700 if (req->rq_import_generation < imp->imp_generation) {
2701 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2705 /* not yet committed */
2706 if (req->rq_transno > imp->imp_peer_committed_transno) {
2707 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2711 if (req->rq_replay) {
2712 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2713 list_move_tail(&req->rq_replay_list,
2714 &imp->imp_committed_list);
2718 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2719 imp->imp_peer_committed_transno);
2721 ptlrpc_free_request(req);
2724 if (skip_committed_list)
2727 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2729 LASSERT(req->rq_transno != 0);
2730 if (req->rq_import_generation < imp->imp_generation ||
2732 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2733 req->rq_import_generation <
2734 imp->imp_generation ? "stale" : "closed");
2736 if (imp->imp_replay_cursor == &req->rq_replay_list)
2737 imp->imp_replay_cursor =
2738 req->rq_replay_list.next;
2740 ptlrpc_free_request(req);
2747 void ptlrpc_cleanup_client(struct obd_import *imp)
2754 * Schedule previously sent request for resend.
2755 * For bulk requests we assign new xid (to avoid problems with
2756 * lost replies and therefore several transfers landing into same buffer
2757 * from different sending attempts).
2759 void ptlrpc_resend_req(struct ptlrpc_request *req)
2761 DEBUG_REQ(D_HA, req, "going to resend");
2762 spin_lock(&req->rq_lock);
2764 /* Request got reply but linked to the import list still.
2765 Let ptlrpc_check_set() to process it. */
2766 if (ptlrpc_client_replied(req)) {
2767 spin_unlock(&req->rq_lock);
2768 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2772 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2773 req->rq_status = -EAGAIN;
2776 req->rq_net_err = 0;
2777 req->rq_timedout = 0;
2779 ptlrpc_client_wake_req(req);
2780 spin_unlock(&req->rq_lock);
2783 /* XXX: this function and rq_status are currently unused */
2784 void ptlrpc_restart_req(struct ptlrpc_request *req)
2786 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2787 req->rq_status = -ERESTARTSYS;
2789 spin_lock(&req->rq_lock);
2790 req->rq_restart = 1;
2791 req->rq_timedout = 0;
2792 ptlrpc_client_wake_req(req);
2793 spin_unlock(&req->rq_lock);
2797 * Grab additional reference on a request \a req
2799 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2802 atomic_inc(&req->rq_refcount);
2805 EXPORT_SYMBOL(ptlrpc_request_addref);
2808 * Add a request to import replay_list.
2809 * Must be called under imp_lock
2811 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2812 struct obd_import *imp)
2814 struct list_head *tmp;
2816 assert_spin_locked(&imp->imp_lock);
2818 if (req->rq_transno == 0) {
2819 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2823 /* clear this for new requests that were resent as well
2824 as resent replayed requests. */
2825 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2827 /* don't re-add requests that have been replayed */
2828 if (!list_empty(&req->rq_replay_list))
2831 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2833 spin_lock(&req->rq_lock);
2835 spin_unlock(&req->rq_lock);
2837 LASSERT(imp->imp_replayable);
2838 /* Balanced in ptlrpc_free_committed, usually. */
2839 ptlrpc_request_addref(req);
2840 list_for_each_prev(tmp, &imp->imp_replay_list) {
2841 struct ptlrpc_request *iter = list_entry(tmp,
2842 struct ptlrpc_request,
2845 /* We may have duplicate transnos if we create and then
2846 * open a file, or for closes retained if to match creating
2847 * opens, so use req->rq_xid as a secondary key.
2848 * (See bugs 684, 685, and 428.)
2849 * XXX no longer needed, but all opens need transnos!
2851 if (iter->rq_transno > req->rq_transno)
2854 if (iter->rq_transno == req->rq_transno) {
2855 LASSERT(iter->rq_xid != req->rq_xid);
2856 if (iter->rq_xid > req->rq_xid)
2860 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2864 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2868 * Send request and wait until it completes.
2869 * Returns request processing status.
2871 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2873 struct ptlrpc_request_set *set;
2877 LASSERT(req->rq_set == NULL);
2878 LASSERT(!req->rq_receiving_reply);
2880 set = ptlrpc_prep_set();
2882 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2886 /* for distributed debugging */
2887 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2889 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2890 ptlrpc_request_addref(req);
2891 ptlrpc_set_add_req(set, req);
2892 rc = ptlrpc_set_wait(set);
2893 ptlrpc_set_destroy(set);
2897 EXPORT_SYMBOL(ptlrpc_queue_wait);
2900 * Callback used for replayed requests reply processing.
2901 * In case of successful reply calls registered request replay callback.
2902 * In case of error restart replay process.
2904 static int ptlrpc_replay_interpret(const struct lu_env *env,
2905 struct ptlrpc_request *req,
2906 void * data, int rc)
2908 struct ptlrpc_replay_async_args *aa = data;
2909 struct obd_import *imp = req->rq_import;
2912 atomic_dec(&imp->imp_replay_inflight);
2914 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2915 * server got the request, but bulk transfer timeout, let's
2916 * replay the bulk req again */
2917 if (!ptlrpc_client_replied(req) ||
2918 (req->rq_bulk != NULL &&
2919 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2920 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2921 GOTO(out, rc = -ETIMEDOUT);
2924 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2925 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2926 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2927 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2929 /** VBR: check version failure */
2930 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2931 /** replay was failed due to version mismatch */
2932 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2933 spin_lock(&imp->imp_lock);
2934 imp->imp_vbr_failed = 1;
2935 spin_unlock(&imp->imp_lock);
2936 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2938 /** The transno had better not change over replay. */
2939 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2940 lustre_msg_get_transno(req->rq_repmsg) ||
2941 lustre_msg_get_transno(req->rq_repmsg) == 0,
2943 lustre_msg_get_transno(req->rq_reqmsg),
2944 lustre_msg_get_transno(req->rq_repmsg));
2947 spin_lock(&imp->imp_lock);
2948 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2949 spin_unlock(&imp->imp_lock);
2950 LASSERT(imp->imp_last_replay_transno);
2952 /* transaction number shouldn't be bigger than the latest replayed */
2953 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2954 DEBUG_REQ(D_ERROR, req,
2955 "Reported transno %llu is bigger than the "
2956 "replayed one: %llu", req->rq_transno,
2957 lustre_msg_get_transno(req->rq_reqmsg));
2958 GOTO(out, rc = -EINVAL);
2961 DEBUG_REQ(D_HA, req, "got rep");
2963 /* let the callback do fixups, possibly including in the request */
2964 if (req->rq_replay_cb)
2965 req->rq_replay_cb(req);
2967 if (ptlrpc_client_replied(req) &&
2968 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2969 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2970 lustre_msg_get_status(req->rq_repmsg),
2971 aa->praa_old_status);
2973 /* Note: If the replay fails for MDT-MDT recovery, let's
2974 * abort all of the following requests in the replay
2975 * and sending list, because MDT-MDT update requests
2976 * are dependent on each other, see LU-7039 */
2977 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
2978 struct ptlrpc_request *free_req;
2979 struct ptlrpc_request *tmp;
2981 spin_lock(&imp->imp_lock);
2982 list_for_each_entry_safe(free_req, tmp,
2983 &imp->imp_replay_list,
2985 ptlrpc_free_request(free_req);
2988 list_for_each_entry_safe(free_req, tmp,
2989 &imp->imp_committed_list,
2991 ptlrpc_free_request(free_req);
2994 list_for_each_entry_safe(free_req, tmp,
2995 &imp->imp_delayed_list,
2997 spin_lock(&free_req->rq_lock);
2998 free_req->rq_err = 1;
2999 free_req->rq_status = -EIO;
3000 ptlrpc_client_wake_req(free_req);
3001 spin_unlock(&free_req->rq_lock);
3004 list_for_each_entry_safe(free_req, tmp,
3005 &imp->imp_sending_list,
3007 spin_lock(&free_req->rq_lock);
3008 free_req->rq_err = 1;
3009 free_req->rq_status = -EIO;
3010 ptlrpc_client_wake_req(free_req);
3011 spin_unlock(&free_req->rq_lock);
3013 spin_unlock(&imp->imp_lock);
3016 /* Put it back for re-replay. */
3017 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3021 * Errors while replay can set transno to 0, but
3022 * imp_last_replay_transno shouldn't be set to 0 anyway
3024 if (req->rq_transno == 0)
3025 CERROR("Transno is 0 during replay!\n");
3027 /* continue with recovery */
3028 rc = ptlrpc_import_recovery_state_machine(imp);
3030 req->rq_send_state = aa->praa_old_state;
3033 /* this replay failed, so restart recovery */
3034 ptlrpc_connect_import(imp);
3040 * Prepares and queues request for replay.
3041 * Adds it to ptlrpcd queue for actual sending.
3042 * Returns 0 on success.
3044 int ptlrpc_replay_req(struct ptlrpc_request *req)
3046 struct ptlrpc_replay_async_args *aa;
3050 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3052 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3053 aa = ptlrpc_req_async_args(req);
3054 memset(aa, 0, sizeof(*aa));
3056 /* Prepare request to be resent with ptlrpcd */
3057 aa->praa_old_state = req->rq_send_state;
3058 req->rq_send_state = LUSTRE_IMP_REPLAY;
3059 req->rq_phase = RQ_PHASE_NEW;
3060 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3062 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3064 req->rq_interpret_reply = ptlrpc_replay_interpret;
3065 /* Readjust the timeout for current conditions */
3066 ptlrpc_at_set_req_timeout(req);
3068 /* Tell server the net_latency, so the server can calculate how long
3069 * it should wait for next replay */
3070 lustre_msg_set_service_time(req->rq_reqmsg,
3071 ptlrpc_at_get_net_latency(req));
3072 DEBUG_REQ(D_HA, req, "REPLAY");
3074 atomic_inc(&req->rq_import->imp_replay_inflight);
3075 spin_lock(&req->rq_lock);
3076 req->rq_early_free_repbuf = 0;
3077 spin_unlock(&req->rq_lock);
3078 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3080 ptlrpcd_add_req(req);
3085 * Aborts all in-flight request on import \a imp sending and delayed lists
3087 void ptlrpc_abort_inflight(struct obd_import *imp)
3089 struct list_head *tmp, *n;
3092 /* Make sure that no new requests get processed for this import.
3093 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3094 * this flag and then putting requests on sending_list or delayed_list.
3096 spin_lock(&imp->imp_lock);
3098 /* XXX locking? Maybe we should remove each request with the list
3099 * locked? Also, how do we know if the requests on the list are
3100 * being freed at this time?
3102 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3103 struct ptlrpc_request *req = list_entry(tmp,
3104 struct ptlrpc_request,
3107 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3109 spin_lock(&req->rq_lock);
3110 if (req->rq_import_generation < imp->imp_generation) {
3112 req->rq_status = -EIO;
3113 ptlrpc_client_wake_req(req);
3115 spin_unlock(&req->rq_lock);
3118 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3119 struct ptlrpc_request *req =
3120 list_entry(tmp, struct ptlrpc_request, rq_list);
3122 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3124 spin_lock(&req->rq_lock);
3125 if (req->rq_import_generation < imp->imp_generation) {
3127 req->rq_status = -EIO;
3128 ptlrpc_client_wake_req(req);
3130 spin_unlock(&req->rq_lock);
3133 /* Last chance to free reqs left on the replay list, but we
3134 * will still leak reqs that haven't committed. */
3135 if (imp->imp_replayable)
3136 ptlrpc_free_committed(imp);
3138 spin_unlock(&imp->imp_lock);
3144 * Abort all uncompleted requests in request set \a set
3146 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3148 struct list_head *tmp, *pos;
3150 LASSERT(set != NULL);
3152 list_for_each_safe(pos, tmp, &set->set_requests) {
3153 struct ptlrpc_request *req =
3154 list_entry(pos, struct ptlrpc_request,
3157 spin_lock(&req->rq_lock);
3158 if (req->rq_phase != RQ_PHASE_RPC) {
3159 spin_unlock(&req->rq_lock);
3164 req->rq_status = -EINTR;
3165 ptlrpc_client_wake_req(req);
3166 spin_unlock(&req->rq_lock);
3170 static __u64 ptlrpc_last_xid;
3171 static spinlock_t ptlrpc_last_xid_lock;
3174 * Initialize the XID for the node. This is common among all requests on
3175 * this node, and only requires the property that it is monotonically
3176 * increasing. It does not need to be sequential. Since this is also used
3177 * as the RDMA match bits, it is important that a single client NOT have
3178 * the same match bits for two different in-flight requests, hence we do
3179 * NOT want to have an XID per target or similar.
3181 * To avoid an unlikely collision between match bits after a client reboot
3182 * (which would deliver old data into the wrong RDMA buffer) initialize
3183 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3184 * If the time is clearly incorrect, we instead use a 62-bit random number.
3185 * In the worst case the random number will overflow 1M RPCs per second in
3186 * 9133 years, or permutations thereof.
3188 #define YEAR_2004 (1ULL << 30)
3189 void ptlrpc_init_xid(void)
3191 time64_t now = ktime_get_real_seconds();
3193 spin_lock_init(&ptlrpc_last_xid_lock);
3194 if (now < YEAR_2004) {
3195 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3196 ptlrpc_last_xid >>= 2;
3197 ptlrpc_last_xid |= (1ULL << 61);
3199 ptlrpc_last_xid = (__u64)now << 20;
3202 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3203 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3204 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3208 * Increase xid and returns resulting new value to the caller.
3210 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3211 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3212 * itself uses the last bulk xid needed, so the server can determine the
3213 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3214 * xid must align to a power-of-two value.
3216 * This is assumed to be true due to the initial ptlrpc_last_xid
3217 * value also being initialized to a power-of-two value. LU-1431
3219 __u64 ptlrpc_next_xid(void)
3223 spin_lock(&ptlrpc_last_xid_lock);
3224 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3225 ptlrpc_last_xid = next;
3226 spin_unlock(&ptlrpc_last_xid_lock);
3232 * If request has a new allocated XID (new request or EINPROGRESS resend),
3233 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3234 * request to ensure previous bulk fails and avoid problems with lost replies
3235 * and therefore several transfers landing into the same buffer from different
3238 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3240 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3242 LASSERT(bd != NULL);
3244 /* Generate new matchbits for all resend requests, including
3246 if (req->rq_resend) {
3247 __u64 old_mbits = req->rq_mbits;
3249 /* First time resend on -EINPROGRESS will generate new xid,
3250 * so we can actually use the rq_xid as rq_mbits in such case,
3251 * however, it's bit hard to distinguish such resend with a
3252 * 'resend for the -EINPROGRESS resend'. To make it simple,
3253 * we opt to generate mbits for all resend cases. */
3254 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
3255 req->rq_mbits = ptlrpc_next_xid();
3257 /* Old version transfers rq_xid to peer as
3259 spin_lock(&req->rq_import->imp_lock);
3260 list_del_init(&req->rq_unreplied_list);
3261 ptlrpc_assign_next_xid_nolock(req);
3262 spin_unlock(&req->rq_import->imp_lock);
3263 req->rq_mbits = req->rq_xid;
3265 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3266 old_mbits, req->rq_mbits);
3267 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3268 /* Request being sent first time, use xid as matchbits. */
3269 req->rq_mbits = req->rq_xid;
3271 /* Replay request, xid and matchbits have already been
3272 * correctly assigned. */
3276 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3277 * that server can infer the number of bulks that were prepared,
3279 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3282 /* Set rq_xid as rq_mbits to indicate the final bulk for the old
3283 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3285 * It's ok to directly set the rq_xid here, since this xid bump
3286 * won't affect the request position in unreplied list. */
3287 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3288 req->rq_xid = req->rq_mbits;
3292 * Get a glimpse at what next xid value might have been.
3293 * Returns possible next xid.
3295 __u64 ptlrpc_sample_next_xid(void)
3297 #if BITS_PER_LONG == 32
3298 /* need to avoid possible word tearing on 32-bit systems */
3301 spin_lock(&ptlrpc_last_xid_lock);
3302 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3303 spin_unlock(&ptlrpc_last_xid_lock);
3307 /* No need to lock, since returned value is racy anyways */
3308 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3311 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3314 * Functions for operating ptlrpc workers.
3316 * A ptlrpc work is a function which will be running inside ptlrpc context.
3317 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3319 * 1. after a work is created, it can be used many times, that is:
3320 * handler = ptlrpcd_alloc_work();
3321 * ptlrpcd_queue_work();
3323 * queue it again when necessary:
3324 * ptlrpcd_queue_work();
3325 * ptlrpcd_destroy_work();
3326 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3327 * it will only be queued once in any time. Also as its name implies, it may
3328 * have delay before it really runs by ptlrpcd thread.
3330 struct ptlrpc_work_async_args {
3331 int (*cb)(const struct lu_env *, void *);
3335 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3337 /* re-initialize the req */
3338 req->rq_timeout = obd_timeout;
3339 req->rq_sent = ktime_get_real_seconds();
3340 req->rq_deadline = req->rq_sent + req->rq_timeout;
3341 req->rq_phase = RQ_PHASE_INTERPRET;
3342 req->rq_next_phase = RQ_PHASE_COMPLETE;
3343 req->rq_xid = ptlrpc_next_xid();
3344 req->rq_import_generation = req->rq_import->imp_generation;
3346 ptlrpcd_add_req(req);
3349 static int work_interpreter(const struct lu_env *env,
3350 struct ptlrpc_request *req, void *data, int rc)
3352 struct ptlrpc_work_async_args *arg = data;
3354 LASSERT(ptlrpcd_check_work(req));
3355 LASSERT(arg->cb != NULL);
3357 rc = arg->cb(env, arg->cbdata);
3359 list_del_init(&req->rq_set_chain);
3362 if (atomic_dec_return(&req->rq_refcount) > 1) {
3363 atomic_set(&req->rq_refcount, 2);
3364 ptlrpcd_add_work_req(req);
3369 static int worker_format;
3371 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3373 return req->rq_pill.rc_fmt == (void *)&worker_format;
3377 * Create a work for ptlrpc.
3379 void *ptlrpcd_alloc_work(struct obd_import *imp,
3380 int (*cb)(const struct lu_env *, void *), void *cbdata)
3382 struct ptlrpc_request *req = NULL;
3383 struct ptlrpc_work_async_args *args;
3389 RETURN(ERR_PTR(-EINVAL));
3391 /* copy some code from deprecated fakereq. */
3392 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3394 CERROR("ptlrpc: run out of memory!\n");
3395 RETURN(ERR_PTR(-ENOMEM));
3398 ptlrpc_cli_req_init(req);
3400 req->rq_send_state = LUSTRE_IMP_FULL;
3401 req->rq_type = PTL_RPC_MSG_REQUEST;
3402 req->rq_import = class_import_get(imp);
3403 req->rq_interpret_reply = work_interpreter;
3404 /* don't want reply */
3405 req->rq_no_delay = req->rq_no_resend = 1;
3406 req->rq_pill.rc_fmt = (void *)&worker_format;
3408 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3409 args = ptlrpc_req_async_args(req);
3411 args->cbdata = cbdata;
3415 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3417 void ptlrpcd_destroy_work(void *handler)
3419 struct ptlrpc_request *req = handler;
3422 ptlrpc_req_finished(req);
3424 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3426 int ptlrpcd_queue_work(void *handler)
3428 struct ptlrpc_request *req = handler;
3431 * Check if the req is already being queued.
3433 * Here comes a trick: it lacks a way of checking if a req is being
3434 * processed reliably in ptlrpc. Here I have to use refcount of req
3435 * for this purpose. This is okay because the caller should use this
3436 * req as opaque data. - Jinshan
3438 LASSERT(atomic_read(&req->rq_refcount) > 0);
3439 if (atomic_inc_return(&req->rq_refcount) == 2)
3440 ptlrpcd_add_work_req(req);
3443 EXPORT_SYMBOL(ptlrpcd_queue_work);