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, 2016, 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 INIT_LIST_HEAD(&set->set_cblist);
960 set->set_max_inflight = UINT_MAX;
961 set->set_producer = NULL;
962 set->set_producer_arg = NULL;
967 EXPORT_SYMBOL(ptlrpc_prep_set);
970 * Allocate and initialize new request set structure with flow control
971 * extension. This extension allows to control the number of requests in-flight
972 * for the whole set. A callback function to generate requests must be provided
973 * and the request set will keep the number of requests sent over the wire to
975 * Returns a pointer to the newly allocated set structure or NULL on error.
977 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
981 struct ptlrpc_request_set *set;
983 set = ptlrpc_prep_set();
987 set->set_max_inflight = max;
988 set->set_producer = func;
989 set->set_producer_arg = arg;
995 * Wind down and free request set structure previously allocated with
997 * Ensures that all requests on the set have completed and removes
998 * all requests from the request list in a set.
999 * If any unsent request happen to be on the list, pretends that they got
1000 * an error in flight and calls their completion handler.
1002 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1004 struct list_head *tmp;
1005 struct list_head *next;
1010 /* Requests on the set should either all be completed, or all be new */
1011 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1012 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1013 list_for_each(tmp, &set->set_requests) {
1014 struct ptlrpc_request *req =
1015 list_entry(tmp, struct ptlrpc_request,
1018 LASSERT(req->rq_phase == expected_phase);
1022 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1023 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1024 atomic_read(&set->set_remaining), n);
1026 list_for_each_safe(tmp, next, &set->set_requests) {
1027 struct ptlrpc_request *req =
1028 list_entry(tmp, struct ptlrpc_request,
1030 list_del_init(&req->rq_set_chain);
1032 LASSERT(req->rq_phase == expected_phase);
1034 if (req->rq_phase == RQ_PHASE_NEW) {
1035 ptlrpc_req_interpret(NULL, req, -EBADR);
1036 atomic_dec(&set->set_remaining);
1039 spin_lock(&req->rq_lock);
1041 req->rq_invalid_rqset = 0;
1042 spin_unlock(&req->rq_lock);
1044 ptlrpc_req_finished (req);
1047 LASSERT(atomic_read(&set->set_remaining) == 0);
1049 ptlrpc_reqset_put(set);
1052 EXPORT_SYMBOL(ptlrpc_set_destroy);
1055 * Add a callback function \a fn to the set.
1056 * This function would be called when all requests on this set are completed.
1057 * The function will be passed \a data argument.
1059 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1060 set_interpreter_func fn, void *data)
1062 struct ptlrpc_set_cbdata *cbdata;
1064 OBD_ALLOC_PTR(cbdata);
1068 cbdata->psc_interpret = fn;
1069 cbdata->psc_data = data;
1070 list_add_tail(&cbdata->psc_item, &set->set_cblist);
1076 * Add a new request to the general purpose request set.
1077 * Assumes request reference from the caller.
1079 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1080 struct ptlrpc_request *req)
1082 LASSERT(list_empty(&req->rq_set_chain));
1084 if (req->rq_allow_intr)
1085 set->set_allow_intr = 1;
1087 /* The set takes over the caller's request reference */
1088 list_add_tail(&req->rq_set_chain, &set->set_requests);
1090 atomic_inc(&set->set_remaining);
1091 req->rq_queued_time = cfs_time_current();
1093 if (req->rq_reqmsg != NULL)
1094 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1096 if (set->set_producer != NULL)
1097 /* If the request set has a producer callback, the RPC must be
1098 * sent straight away */
1099 ptlrpc_send_new_req(req);
1101 EXPORT_SYMBOL(ptlrpc_set_add_req);
1104 * Add a request to a request with dedicated server thread
1105 * and wake the thread to make any necessary processing.
1106 * Currently only used for ptlrpcd.
1108 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1109 struct ptlrpc_request *req)
1111 struct ptlrpc_request_set *set = pc->pc_set;
1114 LASSERT(req->rq_set == NULL);
1115 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1117 spin_lock(&set->set_new_req_lock);
1119 * The set takes over the caller's request reference.
1122 req->rq_queued_time = cfs_time_current();
1123 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1124 count = atomic_inc_return(&set->set_new_count);
1125 spin_unlock(&set->set_new_req_lock);
1127 /* Only need to call wakeup once for the first entry. */
1129 wake_up(&set->set_waitq);
1131 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1132 * guarantee the async RPC can be processed ASAP, we have
1133 * no other better choice. It maybe fixed in future. */
1134 for (i = 0; i < pc->pc_npartners; i++)
1135 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1140 * Based on the current state of the import, determine if the request
1141 * can be sent, is an error, or should be delayed.
1143 * Returns true if this request should be delayed. If false, and
1144 * *status is set, then the request can not be sent and *status is the
1145 * error code. If false and status is 0, then request can be sent.
1147 * The imp->imp_lock must be held.
1149 static int ptlrpc_import_delay_req(struct obd_import *imp,
1150 struct ptlrpc_request *req, int *status)
1155 LASSERT (status != NULL);
1158 if (req->rq_ctx_init || req->rq_ctx_fini) {
1159 /* always allow ctx init/fini rpc go through */
1160 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1161 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1163 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1164 /* pings may safely race with umount */
1165 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1166 D_HA : D_ERROR, req, "IMP_CLOSED ");
1168 } else if (ptlrpc_send_limit_expired(req)) {
1169 /* probably doesn't need to be a D_ERROR after initial testing*/
1170 DEBUG_REQ(D_HA, req, "send limit expired ");
1171 *status = -ETIMEDOUT;
1172 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1173 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1174 /* allow CONNECT even if import is invalid */ ;
1175 if (atomic_read(&imp->imp_inval_count) != 0) {
1176 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1179 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1180 if (!imp->imp_deactive)
1181 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1182 *status = -ESHUTDOWN; /* bz 12940 */
1183 } else if (req->rq_import_generation != imp->imp_generation) {
1184 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1186 } else if (req->rq_send_state != imp->imp_state) {
1187 /* invalidate in progress - any requests should be drop */
1188 if (atomic_read(&imp->imp_inval_count) != 0) {
1189 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1191 } else if (req->rq_no_delay) {
1192 *status = -EWOULDBLOCK;
1193 } else if (req->rq_allow_replay &&
1194 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1195 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1196 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1197 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1198 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1208 * Decide if the error message should be printed to the console or not.
1209 * Makes its decision based on request type, status, and failure frequency.
1211 * \param[in] req request that failed and may need a console message
1213 * \retval false if no message should be printed
1214 * \retval true if console message should be printed
1216 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1220 LASSERT(req->rq_reqmsg != NULL);
1221 opc = lustre_msg_get_opc(req->rq_reqmsg);
1223 /* Suppress particular reconnect errors which are to be expected. */
1224 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1227 /* Suppress timed out reconnect requests */
1228 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1232 /* Suppress most unavailable/again reconnect requests, but
1233 * print occasionally so it is clear client is trying to
1234 * connect to a server where no target is running. */
1235 err = lustre_msg_get_status(req->rq_repmsg);
1236 if ((err == -ENODEV || err == -EAGAIN) &&
1237 req->rq_import->imp_conn_cnt % 30 != 20)
1245 * Check request processing status.
1246 * Returns the status.
1248 static int ptlrpc_check_status(struct ptlrpc_request *req)
1253 err = lustre_msg_get_status(req->rq_repmsg);
1254 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1255 struct obd_import *imp = req->rq_import;
1256 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1257 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1259 /* -EAGAIN is normal when using POSIX flocks */
1260 if (ptlrpc_console_allow(req) &&
1261 !(opc == LDLM_ENQUEUE && err == -EAGAIN))
1262 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1263 "failed: rc = %d\n",
1264 imp->imp_obd->obd_name,
1266 libcfs_nid2str(nid), err);
1267 RETURN(err < 0 ? err : -EINVAL);
1271 DEBUG_REQ(D_INFO, req, "status is %d", err);
1272 } else if (err > 0) {
1273 /* XXX: translate this error from net to host */
1274 DEBUG_REQ(D_INFO, req, "status is %d", err);
1281 * save pre-versions of objects into request for replay.
1282 * Versions are obtained from server reply.
1285 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1287 struct lustre_msg *repmsg = req->rq_repmsg;
1288 struct lustre_msg *reqmsg = req->rq_reqmsg;
1289 __u64 *versions = lustre_msg_get_versions(repmsg);
1292 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1296 lustre_msg_set_versions(reqmsg, versions);
1297 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1298 versions[0], versions[1]);
1303 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1305 struct ptlrpc_request *req;
1307 assert_spin_locked(&imp->imp_lock);
1308 if (list_empty(&imp->imp_unreplied_list))
1311 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1313 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1315 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1316 imp->imp_known_replied_xid = req->rq_xid - 1;
1318 return req->rq_xid - 1;
1322 * Callback function called when client receives RPC reply for \a req.
1323 * Returns 0 on success or error code.
1324 * The return alue would be assigned to req->rq_status by the caller
1325 * as request processing status.
1326 * This function also decides if the request needs to be saved for later replay.
1328 static int after_reply(struct ptlrpc_request *req)
1330 struct obd_import *imp = req->rq_import;
1331 struct obd_device *obd = req->rq_import->imp_obd;
1338 LASSERT(obd != NULL);
1339 /* repbuf must be unlinked */
1340 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1342 if (req->rq_reply_truncated) {
1343 if (ptlrpc_no_resend(req)) {
1344 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1345 " expected: %d, actual size: %d",
1346 req->rq_nob_received, req->rq_repbuf_len);
1350 sptlrpc_cli_free_repbuf(req);
1351 /* Pass the required reply buffer size (include
1352 * space for early reply).
1353 * NB: no need to roundup because alloc_repbuf
1354 * will roundup it */
1355 req->rq_replen = req->rq_nob_received;
1356 req->rq_nob_received = 0;
1357 spin_lock(&req->rq_lock);
1359 spin_unlock(&req->rq_lock);
1363 work_start = ktime_get_real();
1364 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1367 * NB Until this point, the whole of the incoming message,
1368 * including buflens, status etc is in the sender's byte order.
1370 rc = sptlrpc_cli_unwrap_reply(req);
1372 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1377 * Security layer unwrap might ask resend this request.
1382 rc = unpack_reply(req);
1386 /* retry indefinitely on EINPROGRESS */
1387 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1388 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1389 time64_t now = ktime_get_real_seconds();
1391 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1392 spin_lock(&req->rq_lock);
1394 spin_unlock(&req->rq_lock);
1395 req->rq_nr_resend++;
1397 /* Readjust the timeout for current conditions */
1398 ptlrpc_at_set_req_timeout(req);
1399 /* delay resend to give a chance to the server to get ready.
1400 * The delay is increased by 1s on every resend and is capped to
1401 * the current request timeout (i.e. obd_timeout if AT is off,
1402 * or AT service time x 125% + 5s, see at_est2timeout) */
1403 if (req->rq_nr_resend > req->rq_timeout)
1404 req->rq_sent = now + req->rq_timeout;
1406 req->rq_sent = now + req->rq_nr_resend;
1408 /* Resend for EINPROGRESS will use a new XID */
1409 spin_lock(&imp->imp_lock);
1410 list_del_init(&req->rq_unreplied_list);
1411 spin_unlock(&imp->imp_lock);
1416 if (obd->obd_svc_stats != NULL) {
1417 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1419 ptlrpc_lprocfs_rpc_sent(req, timediff);
1422 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1423 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1424 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1425 lustre_msg_get_type(req->rq_repmsg));
1429 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1430 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1431 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1432 ptlrpc_at_adj_net_latency(req,
1433 lustre_msg_get_service_time(req->rq_repmsg));
1435 rc = ptlrpc_check_status(req);
1436 imp->imp_connect_error = rc;
1440 * Either we've been evicted, or the server has failed for
1441 * some reason. Try to reconnect, and if that fails, punt to
1444 if (ptlrpc_recoverable_error(rc)) {
1445 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1446 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1449 ptlrpc_request_handle_notconn(req);
1454 * Let's look if server sent slv. Do it only for RPC with
1457 ldlm_cli_update_pool(req);
1461 * Store transno in reqmsg for replay.
1463 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1464 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1465 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1468 if (imp->imp_replayable) {
1469 spin_lock(&imp->imp_lock);
1471 * No point in adding already-committed requests to the replay
1472 * list, we will just remove them immediately. b=9829
1474 if (req->rq_transno != 0 &&
1476 lustre_msg_get_last_committed(req->rq_repmsg) ||
1478 /** version recovery */
1479 ptlrpc_save_versions(req);
1480 ptlrpc_retain_replayable_request(req, imp);
1481 } else if (req->rq_commit_cb != NULL &&
1482 list_empty(&req->rq_replay_list)) {
1483 /* NB: don't call rq_commit_cb if it's already on
1484 * rq_replay_list, ptlrpc_free_committed() will call
1485 * it later, see LU-3618 for details */
1486 spin_unlock(&imp->imp_lock);
1487 req->rq_commit_cb(req);
1488 spin_lock(&imp->imp_lock);
1492 * Replay-enabled imports return commit-status information.
1494 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1495 if (likely(committed > imp->imp_peer_committed_transno))
1496 imp->imp_peer_committed_transno = committed;
1498 ptlrpc_free_committed(imp);
1500 if (!list_empty(&imp->imp_replay_list)) {
1501 struct ptlrpc_request *last;
1503 last = list_entry(imp->imp_replay_list.prev,
1504 struct ptlrpc_request,
1507 * Requests with rq_replay stay on the list even if no
1508 * commit is expected.
1510 if (last->rq_transno > imp->imp_peer_committed_transno)
1511 ptlrpc_pinger_commit_expected(imp);
1514 spin_unlock(&imp->imp_lock);
1521 * Helper function to send request \a req over the network for the first time
1522 * Also adjusts request phase.
1523 * Returns 0 on success or error code.
1525 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1527 struct obd_import *imp = req->rq_import;
1532 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1534 /* do not try to go further if there is not enough memory in enc_pool */
1535 if (req->rq_sent && req->rq_bulk != NULL)
1536 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1537 pool_is_at_full_capacity())
1540 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1541 (!req->rq_generation_set ||
1542 req->rq_import_generation == imp->imp_generation))
1545 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1547 spin_lock(&imp->imp_lock);
1549 LASSERT(req->rq_xid != 0);
1550 LASSERT(!list_empty(&req->rq_unreplied_list));
1552 if (!req->rq_generation_set)
1553 req->rq_import_generation = imp->imp_generation;
1555 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1556 spin_lock(&req->rq_lock);
1557 req->rq_waiting = 1;
1558 spin_unlock(&req->rq_lock);
1560 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1561 ptlrpc_import_state_name(req->rq_send_state),
1562 ptlrpc_import_state_name(imp->imp_state));
1563 LASSERT(list_empty(&req->rq_list));
1564 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1565 atomic_inc(&req->rq_import->imp_inflight);
1566 spin_unlock(&imp->imp_lock);
1571 spin_unlock(&imp->imp_lock);
1572 req->rq_status = rc;
1573 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1577 LASSERT(list_empty(&req->rq_list));
1578 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1579 atomic_inc(&req->rq_import->imp_inflight);
1581 /* find the known replied XID from the unreplied list, CONNECT
1582 * and DISCONNECT requests are skipped to make the sanity check
1583 * on server side happy. see process_req_last_xid().
1585 * For CONNECT: Because replay requests have lower XID, it'll
1586 * break the sanity check if CONNECT bump the exp_last_xid on
1589 * For DISCONNECT: Since client will abort inflight RPC before
1590 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1591 * than the inflight RPC.
1593 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1594 min_xid = ptlrpc_known_replied_xid(imp);
1595 spin_unlock(&imp->imp_lock);
1597 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1599 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1601 rc = sptlrpc_req_refresh_ctx(req, -1);
1604 req->rq_status = rc;
1607 spin_lock(&req->rq_lock);
1608 req->rq_wait_ctx = 1;
1609 spin_unlock(&req->rq_lock);
1614 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1615 " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1616 imp->imp_obd->obd_uuid.uuid,
1617 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1618 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1619 lustre_msg_get_opc(req->rq_reqmsg));
1621 rc = ptl_send_rpc(req, 0);
1622 if (rc == -ENOMEM) {
1623 spin_lock(&imp->imp_lock);
1624 if (!list_empty(&req->rq_list)) {
1625 list_del_init(&req->rq_list);
1626 atomic_dec(&req->rq_import->imp_inflight);
1628 spin_unlock(&imp->imp_lock);
1629 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1633 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1634 spin_lock(&req->rq_lock);
1635 req->rq_net_err = 1;
1636 spin_unlock(&req->rq_lock);
1642 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1647 LASSERT(set->set_producer != NULL);
1649 remaining = atomic_read(&set->set_remaining);
1651 /* populate the ->set_requests list with requests until we
1652 * reach the maximum number of RPCs in flight for this set */
1653 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1654 rc = set->set_producer(set, set->set_producer_arg);
1655 if (rc == -ENOENT) {
1656 /* no more RPC to produce */
1657 set->set_producer = NULL;
1658 set->set_producer_arg = NULL;
1663 RETURN((atomic_read(&set->set_remaining) - remaining));
1667 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1668 * and no more replies are expected.
1669 * (it is possible to get less replies than requests sent e.g. due to timed out
1670 * requests or requests that we had trouble to send out)
1672 * NOTE: This function contains a potential schedule point (cond_resched()).
1674 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1676 struct list_head *tmp, *next;
1677 struct list_head comp_reqs;
1678 int force_timer_recalc = 0;
1681 if (atomic_read(&set->set_remaining) == 0)
1684 INIT_LIST_HEAD(&comp_reqs);
1685 list_for_each_safe(tmp, next, &set->set_requests) {
1686 struct ptlrpc_request *req =
1687 list_entry(tmp, struct ptlrpc_request,
1689 struct obd_import *imp = req->rq_import;
1690 int unregistered = 0;
1694 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1695 list_move_tail(&req->rq_set_chain, &comp_reqs);
1699 /* This schedule point is mainly for the ptlrpcd caller of this
1700 * function. Most ptlrpc sets are not long-lived and unbounded
1701 * in length, but at the least the set used by the ptlrpcd is.
1702 * Since the processing time is unbounded, we need to insert an
1703 * explicit schedule point to make the thread well-behaved.
1707 /* If the caller requires to allow to be interpreted by force
1708 * and it has really been interpreted, then move the request
1709 * to RQ_PHASE_INTERPRET phase in spite of what the current
1711 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1712 req->rq_status = -EINTR;
1713 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1715 /* Since it is interpreted and we have to wait for
1716 * the reply to be unlinked, then use sync mode. */
1719 GOTO(interpret, req->rq_status);
1722 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1723 force_timer_recalc = 1;
1725 /* delayed send - skip */
1726 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1729 /* delayed resend - skip */
1730 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1731 req->rq_sent > ktime_get_real_seconds())
1734 if (!(req->rq_phase == RQ_PHASE_RPC ||
1735 req->rq_phase == RQ_PHASE_BULK ||
1736 req->rq_phase == RQ_PHASE_INTERPRET ||
1737 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1738 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1739 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1743 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1744 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1745 LASSERT(req->rq_next_phase != req->rq_phase);
1746 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1748 if (req->rq_req_deadline &&
1749 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1750 req->rq_req_deadline = 0;
1751 if (req->rq_reply_deadline &&
1752 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1753 req->rq_reply_deadline = 0;
1754 if (req->rq_bulk_deadline &&
1755 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1756 req->rq_bulk_deadline = 0;
1759 * Skip processing until reply is unlinked. We
1760 * can't return to pool before that and we can't
1761 * call interpret before that. We need to make
1762 * sure that all rdma transfers finished and will
1763 * not corrupt any data.
1765 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1766 ptlrpc_client_recv_or_unlink(req))
1768 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1769 ptlrpc_client_bulk_active(req))
1773 * Turn fail_loc off to prevent it from looping
1776 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1777 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1780 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1781 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1786 * Move to next phase if reply was successfully
1789 ptlrpc_rqphase_move(req, req->rq_next_phase);
1792 if (req->rq_phase == RQ_PHASE_INTERPRET)
1793 GOTO(interpret, req->rq_status);
1796 * Note that this also will start async reply unlink.
1798 if (req->rq_net_err && !req->rq_timedout) {
1799 ptlrpc_expire_one_request(req, 1);
1802 * Check if we still need to wait for unlink.
1804 if (ptlrpc_client_recv_or_unlink(req) ||
1805 ptlrpc_client_bulk_active(req))
1807 /* If there is no need to resend, fail it now. */
1808 if (req->rq_no_resend) {
1809 if (req->rq_status == 0)
1810 req->rq_status = -EIO;
1811 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1812 GOTO(interpret, req->rq_status);
1819 spin_lock(&req->rq_lock);
1820 req->rq_replied = 0;
1821 spin_unlock(&req->rq_lock);
1822 if (req->rq_status == 0)
1823 req->rq_status = -EIO;
1824 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1825 GOTO(interpret, req->rq_status);
1828 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1829 * so it sets rq_intr regardless of individual rpc
1830 * timeouts. The synchronous IO waiting path sets
1831 * rq_intr irrespective of whether ptlrpcd
1832 * has seen a timeout. Our policy is to only interpret
1833 * interrupted rpcs after they have timed out, so we
1834 * need to enforce that here.
1837 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1838 req->rq_wait_ctx)) {
1839 req->rq_status = -EINTR;
1840 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1841 GOTO(interpret, req->rq_status);
1844 if (req->rq_phase == RQ_PHASE_RPC) {
1845 if (req->rq_timedout || req->rq_resend ||
1846 req->rq_waiting || req->rq_wait_ctx) {
1849 if (!ptlrpc_unregister_reply(req, 1)) {
1850 ptlrpc_unregister_bulk(req, 1);
1854 spin_lock(&imp->imp_lock);
1855 if (ptlrpc_import_delay_req(imp, req, &status)){
1856 /* put on delay list - only if we wait
1857 * recovery finished - before send */
1858 list_del_init(&req->rq_list);
1859 list_add_tail(&req->rq_list,
1862 spin_unlock(&imp->imp_lock);
1867 req->rq_status = status;
1868 ptlrpc_rqphase_move(req,
1869 RQ_PHASE_INTERPRET);
1870 spin_unlock(&imp->imp_lock);
1871 GOTO(interpret, req->rq_status);
1873 if (ptlrpc_no_resend(req) &&
1874 !req->rq_wait_ctx) {
1875 req->rq_status = -ENOTCONN;
1876 ptlrpc_rqphase_move(req,
1877 RQ_PHASE_INTERPRET);
1878 spin_unlock(&imp->imp_lock);
1879 GOTO(interpret, req->rq_status);
1882 list_del_init(&req->rq_list);
1883 list_add_tail(&req->rq_list,
1884 &imp->imp_sending_list);
1886 spin_unlock(&imp->imp_lock);
1888 spin_lock(&req->rq_lock);
1889 req->rq_waiting = 0;
1890 spin_unlock(&req->rq_lock);
1892 if (req->rq_timedout || req->rq_resend) {
1893 /* This is re-sending anyways,
1894 * let's mark req as resend. */
1895 spin_lock(&req->rq_lock);
1897 spin_unlock(&req->rq_lock);
1899 if (req->rq_bulk != NULL &&
1900 !ptlrpc_unregister_bulk(req, 1))
1904 * rq_wait_ctx is only touched by ptlrpcd,
1905 * so no lock is needed here.
1907 status = sptlrpc_req_refresh_ctx(req, -1);
1910 req->rq_status = status;
1911 spin_lock(&req->rq_lock);
1912 req->rq_wait_ctx = 0;
1913 spin_unlock(&req->rq_lock);
1914 force_timer_recalc = 1;
1916 spin_lock(&req->rq_lock);
1917 req->rq_wait_ctx = 1;
1918 spin_unlock(&req->rq_lock);
1923 spin_lock(&req->rq_lock);
1924 req->rq_wait_ctx = 0;
1925 spin_unlock(&req->rq_lock);
1928 rc = ptl_send_rpc(req, 0);
1929 if (rc == -ENOMEM) {
1930 spin_lock(&imp->imp_lock);
1931 if (!list_empty(&req->rq_list))
1932 list_del_init(&req->rq_list);
1933 spin_unlock(&imp->imp_lock);
1934 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1938 DEBUG_REQ(D_HA, req,
1939 "send failed: rc = %d", rc);
1940 force_timer_recalc = 1;
1941 spin_lock(&req->rq_lock);
1942 req->rq_net_err = 1;
1943 spin_unlock(&req->rq_lock);
1946 /* need to reset the timeout */
1947 force_timer_recalc = 1;
1950 spin_lock(&req->rq_lock);
1952 if (ptlrpc_client_early(req)) {
1953 ptlrpc_at_recv_early_reply(req);
1954 spin_unlock(&req->rq_lock);
1958 /* Still waiting for a reply? */
1959 if (ptlrpc_client_recv(req)) {
1960 spin_unlock(&req->rq_lock);
1964 /* Did we actually receive a reply? */
1965 if (!ptlrpc_client_replied(req)) {
1966 spin_unlock(&req->rq_lock);
1970 spin_unlock(&req->rq_lock);
1972 /* unlink from net because we are going to
1973 * swab in-place of reply buffer */
1974 unregistered = ptlrpc_unregister_reply(req, 1);
1978 req->rq_status = after_reply(req);
1982 /* If there is no bulk associated with this request,
1983 * then we're done and should let the interpreter
1984 * process the reply. Similarly if the RPC returned
1985 * an error, and therefore the bulk will never arrive.
1987 if (req->rq_bulk == NULL || req->rq_status < 0) {
1988 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1989 GOTO(interpret, req->rq_status);
1992 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1995 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1996 if (ptlrpc_client_bulk_active(req))
1999 if (req->rq_bulk->bd_failure) {
2000 /* The RPC reply arrived OK, but the bulk screwed
2001 * up! Dead weird since the server told us the RPC
2002 * was good after getting the REPLY for her GET or
2003 * the ACK for her PUT. */
2004 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2005 req->rq_status = -EIO;
2008 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2011 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2013 /* This moves to "unregistering" phase we need to wait for
2015 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2016 /* start async bulk unlink too */
2017 ptlrpc_unregister_bulk(req, 1);
2021 if (!ptlrpc_unregister_bulk(req, async))
2024 /* When calling interpret receiving already should be
2026 LASSERT(!req->rq_receiving_reply);
2028 ptlrpc_req_interpret(env, req, req->rq_status);
2030 if (ptlrpcd_check_work(req)) {
2031 atomic_dec(&set->set_remaining);
2034 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2036 if (req->rq_reqmsg != NULL)
2038 "Completed RPC pname:cluuid:pid:xid:nid:"
2039 "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
2040 imp->imp_obd->obd_uuid.uuid,
2041 lustre_msg_get_status(req->rq_reqmsg),
2043 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2044 lustre_msg_get_opc(req->rq_reqmsg));
2046 spin_lock(&imp->imp_lock);
2047 /* Request already may be not on sending or delaying list. This
2048 * may happen in the case of marking it erroneous for the case
2049 * ptlrpc_import_delay_req(req, status) find it impossible to
2050 * allow sending this rpc and returns *status != 0. */
2051 if (!list_empty(&req->rq_list)) {
2052 list_del_init(&req->rq_list);
2053 atomic_dec(&imp->imp_inflight);
2055 list_del_init(&req->rq_unreplied_list);
2056 spin_unlock(&imp->imp_lock);
2058 atomic_dec(&set->set_remaining);
2059 wake_up_all(&imp->imp_recovery_waitq);
2061 if (set->set_producer) {
2062 /* produce a new request if possible */
2063 if (ptlrpc_set_producer(set) > 0)
2064 force_timer_recalc = 1;
2066 /* free the request that has just been completed
2067 * in order not to pollute set->set_requests */
2068 list_del_init(&req->rq_set_chain);
2069 spin_lock(&req->rq_lock);
2071 req->rq_invalid_rqset = 0;
2072 spin_unlock(&req->rq_lock);
2074 /* record rq_status to compute the final status later */
2075 if (req->rq_status != 0)
2076 set->set_rc = req->rq_status;
2077 ptlrpc_req_finished(req);
2079 list_move_tail(&req->rq_set_chain, &comp_reqs);
2083 /* move completed request at the head of list so it's easier for
2084 * caller to find them */
2085 list_splice(&comp_reqs, &set->set_requests);
2087 /* If we hit an error, we want to recover promptly. */
2088 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2090 EXPORT_SYMBOL(ptlrpc_check_set);
2093 * Time out request \a req. is \a async_unlink is set, that means do not wait
2094 * until LNet actually confirms network buffer unlinking.
2095 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2097 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2099 struct obd_import *imp = req->rq_import;
2103 spin_lock(&req->rq_lock);
2104 req->rq_timedout = 1;
2105 spin_unlock(&req->rq_lock);
2107 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent %lld/real %lld]",
2108 req->rq_net_err ? "failed due to network error" :
2109 ((req->rq_real_sent == 0 ||
2110 req->rq_real_sent < req->rq_sent ||
2111 req->rq_real_sent >= req->rq_deadline) ?
2112 "timed out for sent delay" : "timed out for slow reply"),
2113 (s64)req->rq_sent, (s64)req->rq_real_sent);
2115 if (imp != NULL && obd_debug_peer_on_timeout)
2116 LNetDebugPeer(imp->imp_connection->c_peer);
2118 ptlrpc_unregister_reply(req, async_unlink);
2119 ptlrpc_unregister_bulk(req, async_unlink);
2121 if (obd_dump_on_timeout)
2122 libcfs_debug_dumplog();
2125 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2129 atomic_inc(&imp->imp_timeouts);
2131 /* The DLM server doesn't want recovery run on its imports. */
2132 if (imp->imp_dlm_fake)
2135 /* If this request is for recovery or other primordial tasks,
2136 * then error it out here. */
2137 if (req->rq_ctx_init || req->rq_ctx_fini ||
2138 req->rq_send_state != LUSTRE_IMP_FULL ||
2139 imp->imp_obd->obd_no_recov) {
2140 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2141 ptlrpc_import_state_name(req->rq_send_state),
2142 ptlrpc_import_state_name(imp->imp_state));
2143 spin_lock(&req->rq_lock);
2144 req->rq_status = -ETIMEDOUT;
2146 spin_unlock(&req->rq_lock);
2150 /* if a request can't be resent we can't wait for an answer after
2152 if (ptlrpc_no_resend(req)) {
2153 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2157 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2163 * Time out all uncompleted requests in request set pointed by \a data
2164 * Callback used when waiting on sets with l_wait_event.
2167 int ptlrpc_expired_set(void *data)
2169 struct ptlrpc_request_set *set = data;
2170 struct list_head *tmp;
2171 time64_t now = ktime_get_real_seconds();
2174 LASSERT(set != NULL);
2177 * A timeout expired. See which reqs it applies to...
2179 list_for_each(tmp, &set->set_requests) {
2180 struct ptlrpc_request *req =
2181 list_entry(tmp, struct ptlrpc_request,
2184 /* don't expire request waiting for context */
2185 if (req->rq_wait_ctx)
2188 /* Request in-flight? */
2189 if (!((req->rq_phase == RQ_PHASE_RPC &&
2190 !req->rq_waiting && !req->rq_resend) ||
2191 (req->rq_phase == RQ_PHASE_BULK)))
2194 if (req->rq_timedout || /* already dealt with */
2195 req->rq_deadline > now) /* not expired */
2198 /* Deal with this guy. Do it asynchronously to not block
2199 * ptlrpcd thread. */
2200 ptlrpc_expire_one_request(req, 1);
2204 * When waiting for a whole set, we always break out of the
2205 * sleep so we can recalculate the timeout, or enable interrupts
2206 * if everyone's timed out.
2212 * Sets rq_intr flag in \a req under spinlock.
2214 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2216 spin_lock(&req->rq_lock);
2218 spin_unlock(&req->rq_lock);
2220 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2223 * Interrupts (sets interrupted flag) all uncompleted requests in
2224 * a set \a data. Callback for l_wait_event for interruptible waits.
2226 static void ptlrpc_interrupted_set(void *data)
2228 struct ptlrpc_request_set *set = data;
2229 struct list_head *tmp;
2231 LASSERT(set != NULL);
2232 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2234 list_for_each(tmp, &set->set_requests) {
2235 struct ptlrpc_request *req =
2236 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2241 if (req->rq_phase != RQ_PHASE_RPC &&
2242 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2243 !req->rq_allow_intr)
2246 ptlrpc_mark_interrupted(req);
2251 * Get the smallest timeout in the set; this does NOT set a timeout.
2253 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2255 struct list_head *tmp;
2256 time64_t now = ktime_get_real_seconds();
2258 struct ptlrpc_request *req;
2262 list_for_each(tmp, &set->set_requests) {
2263 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2266 * Request in-flight?
2268 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2269 (req->rq_phase == RQ_PHASE_BULK) ||
2270 (req->rq_phase == RQ_PHASE_NEW)))
2274 * Already timed out.
2276 if (req->rq_timedout)
2282 if (req->rq_wait_ctx)
2285 if (req->rq_phase == RQ_PHASE_NEW)
2286 deadline = req->rq_sent;
2287 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2288 deadline = req->rq_sent;
2290 deadline = req->rq_sent + req->rq_timeout;
2292 if (deadline <= now) /* actually expired already */
2293 timeout = 1; /* ASAP */
2294 else if (timeout == 0 || timeout > deadline - now)
2295 timeout = deadline - now;
2301 * Send all unset request from the set and then wait untill all
2302 * requests in the set complete (either get a reply, timeout, get an
2303 * error or otherwise be interrupted).
2304 * Returns 0 on success or error code otherwise.
2306 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2308 struct list_head *tmp;
2309 struct ptlrpc_request *req;
2310 struct l_wait_info lwi;
2314 if (set->set_producer)
2315 (void)ptlrpc_set_producer(set);
2317 list_for_each(tmp, &set->set_requests) {
2318 req = list_entry(tmp, struct ptlrpc_request,
2320 if (req->rq_phase == RQ_PHASE_NEW)
2321 (void)ptlrpc_send_new_req(req);
2324 if (list_empty(&set->set_requests))
2328 timeout = ptlrpc_set_next_timeout(set);
2330 /* wait until all complete, interrupted, or an in-flight
2332 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2335 if ((timeout == 0 && !signal_pending(current)) ||
2336 set->set_allow_intr)
2337 /* No requests are in-flight (ether timed out
2338 * or delayed), so we can allow interrupts.
2339 * We still want to block for a limited time,
2340 * so we allow interrupts during the timeout. */
2341 lwi = LWI_TIMEOUT_INTR_ALL(
2342 cfs_time_seconds(timeout ? timeout : 1),
2344 ptlrpc_interrupted_set, set);
2347 * At least one request is in flight, so no
2348 * interrupts are allowed. Wait until all
2349 * complete, or an in-flight req times out.
2351 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2352 ptlrpc_expired_set, set);
2354 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2356 /* LU-769 - if we ignored the signal because it was already
2357 * pending when we started, we need to handle it now or we risk
2358 * it being ignored forever */
2359 if (rc == -ETIMEDOUT &&
2360 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2361 signal_pending(current)) {
2362 sigset_t blocked_sigs =
2363 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2365 /* In fact we only interrupt for the "fatal" signals
2366 * like SIGINT or SIGKILL. We still ignore less
2367 * important signals since ptlrpc set is not easily
2368 * reentrant from userspace again */
2369 if (signal_pending(current))
2370 ptlrpc_interrupted_set(set);
2371 cfs_restore_sigs(blocked_sigs);
2374 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2376 /* -EINTR => all requests have been flagged rq_intr so next
2378 * -ETIMEDOUT => someone timed out. When all reqs have
2379 * timed out, signals are enabled allowing completion with
2381 * I don't really care if we go once more round the loop in
2382 * the error cases -eeb. */
2383 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2384 list_for_each(tmp, &set->set_requests) {
2385 req = list_entry(tmp, struct ptlrpc_request,
2387 spin_lock(&req->rq_lock);
2388 req->rq_invalid_rqset = 1;
2389 spin_unlock(&req->rq_lock);
2392 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2394 LASSERT(atomic_read(&set->set_remaining) == 0);
2396 rc = set->set_rc; /* rq_status of already freed requests if any */
2397 list_for_each(tmp, &set->set_requests) {
2398 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2400 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2401 if (req->rq_status != 0)
2402 rc = req->rq_status;
2405 if (set->set_interpret != NULL) {
2406 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2408 rc = interpreter (set, set->set_arg, rc);
2410 struct ptlrpc_set_cbdata *cbdata, *n;
2413 list_for_each_entry_safe(cbdata, n,
2414 &set->set_cblist, psc_item) {
2415 list_del_init(&cbdata->psc_item);
2416 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2419 OBD_FREE_PTR(cbdata);
2425 EXPORT_SYMBOL(ptlrpc_set_wait);
2428 * Helper fuction for request freeing.
2429 * Called when request count reached zero and request needs to be freed.
2430 * Removes request from all sorts of sending/replay lists it might be on,
2431 * frees network buffers if any are present.
2432 * If \a locked is set, that means caller is already holding import imp_lock
2433 * and so we no longer need to reobtain it (for certain lists manipulations)
2435 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2439 if (request == NULL)
2442 LASSERT(!request->rq_srv_req);
2443 LASSERT(request->rq_export == NULL);
2444 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2445 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2446 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2447 LASSERTF(!request->rq_replay, "req %p\n", request);
2449 req_capsule_fini(&request->rq_pill);
2451 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2452 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2453 if (request->rq_import != NULL) {
2455 spin_lock(&request->rq_import->imp_lock);
2456 list_del_init(&request->rq_replay_list);
2457 list_del_init(&request->rq_unreplied_list);
2459 spin_unlock(&request->rq_import->imp_lock);
2461 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2463 if (atomic_read(&request->rq_refcount) != 0) {
2464 DEBUG_REQ(D_ERROR, request,
2465 "freeing request with nonzero refcount");
2469 if (request->rq_repbuf != NULL)
2470 sptlrpc_cli_free_repbuf(request);
2472 if (request->rq_import != NULL) {
2473 class_import_put(request->rq_import);
2474 request->rq_import = NULL;
2476 if (request->rq_bulk != NULL)
2477 ptlrpc_free_bulk(request->rq_bulk);
2479 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2480 sptlrpc_cli_free_reqbuf(request);
2482 if (request->rq_cli_ctx)
2483 sptlrpc_req_put_ctx(request, !locked);
2485 if (request->rq_pool)
2486 __ptlrpc_free_req_to_pool(request);
2488 ptlrpc_request_cache_free(request);
2492 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2494 * Drop one request reference. Must be called with import imp_lock held.
2495 * When reference count drops to zero, request is freed.
2497 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2499 assert_spin_locked(&request->rq_import->imp_lock);
2500 (void)__ptlrpc_req_finished(request, 1);
2505 * Drops one reference count for request \a request.
2506 * \a locked set indicates that caller holds import imp_lock.
2507 * Frees the request whe reference count reaches zero.
2509 * \retval 1 the request is freed
2510 * \retval 0 some others still hold references on the request
2512 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2520 LASSERT(request != LP_POISON);
2521 LASSERT(request->rq_reqmsg != LP_POISON);
2523 DEBUG_REQ(D_INFO, request, "refcount now %u",
2524 atomic_read(&request->rq_refcount) - 1);
2526 spin_lock(&request->rq_lock);
2527 count = atomic_dec_return(&request->rq_refcount);
2528 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2530 /* For open RPC, the client does not know the EA size (LOV, ACL, and
2531 * so on) before replied, then the client has to reserve very large
2532 * reply buffer. Such buffer will not be released until the RPC freed.
2533 * Since The open RPC is replayable, we need to keep it in the replay
2534 * list until close. If there are a lot of files opened concurrently,
2535 * then the client may be OOM.
2537 * If fact, it is unnecessary to keep reply buffer for open replay,
2538 * related EAs have already been saved via mdc_save_lovea() before
2539 * coming here. So it is safe to free the reply buffer some earlier
2540 * before releasing the RPC to avoid client OOM. LU-9514 */
2541 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2542 spin_lock(&request->rq_early_free_lock);
2543 sptlrpc_cli_free_repbuf(request);
2544 request->rq_repbuf = NULL;
2545 request->rq_repbuf_len = 0;
2546 request->rq_repdata = NULL;
2547 request->rq_reqdata_len = 0;
2548 spin_unlock(&request->rq_early_free_lock);
2550 spin_unlock(&request->rq_lock);
2553 __ptlrpc_free_req(request, locked);
2559 * Drops one reference count for a request.
2561 void ptlrpc_req_finished(struct ptlrpc_request *request)
2563 __ptlrpc_req_finished(request, 0);
2565 EXPORT_SYMBOL(ptlrpc_req_finished);
2568 * Returns xid of a \a request
2570 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2572 return request->rq_xid;
2574 EXPORT_SYMBOL(ptlrpc_req_xid);
2577 * Disengage the client's reply buffer from the network
2578 * NB does _NOT_ unregister any client-side bulk.
2579 * IDEMPOTENT, but _not_ safe against concurrent callers.
2580 * The request owner (i.e. the thread doing the I/O) must call...
2581 * Returns 0 on success or 1 if unregistering cannot be made.
2583 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2586 struct l_wait_info lwi;
2591 LASSERT(!in_interrupt());
2593 /* Let's setup deadline for reply unlink. */
2594 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2595 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2596 request->rq_reply_deadline = ktime_get_real_seconds() +
2600 * Nothing left to do.
2602 if (!ptlrpc_client_recv_or_unlink(request))
2605 LNetMDUnlink(request->rq_reply_md_h);
2608 * Let's check it once again.
2610 if (!ptlrpc_client_recv_or_unlink(request))
2613 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2614 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2617 * Do not wait for unlink to finish.
2623 * We have to l_wait_event() whatever the result, to give liblustre
2624 * a chance to run reply_in_callback(), and to make sure we've
2625 * unlinked before returning a req to the pool.
2628 /* The wq argument is ignored by user-space wait_event macros */
2629 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2630 &request->rq_set->set_waitq :
2631 &request->rq_reply_waitq;
2632 /* Network access will complete in finite time but the HUGE
2633 * timeout lets us CWARN for visibility of sluggish NALs */
2634 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2635 cfs_time_seconds(1), NULL, NULL);
2636 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2639 ptlrpc_rqphase_move(request, request->rq_next_phase);
2643 LASSERT(rc == -ETIMEDOUT);
2644 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2645 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2646 request->rq_receiving_reply,
2647 request->rq_req_unlinked,
2648 request->rq_reply_unlinked);
2653 static void ptlrpc_free_request(struct ptlrpc_request *req)
2655 spin_lock(&req->rq_lock);
2657 spin_unlock(&req->rq_lock);
2659 if (req->rq_commit_cb != NULL)
2660 req->rq_commit_cb(req);
2661 list_del_init(&req->rq_replay_list);
2663 __ptlrpc_req_finished(req, 1);
2667 * the request is committed and dropped from the replay list of its import
2669 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2671 struct obd_import *imp = req->rq_import;
2673 spin_lock(&imp->imp_lock);
2674 if (list_empty(&req->rq_replay_list)) {
2675 spin_unlock(&imp->imp_lock);
2679 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2680 ptlrpc_free_request(req);
2682 spin_unlock(&imp->imp_lock);
2684 EXPORT_SYMBOL(ptlrpc_request_committed);
2687 * Iterates through replay_list on import and prunes
2688 * all requests have transno smaller than last_committed for the
2689 * import and don't have rq_replay set.
2690 * Since requests are sorted in transno order, stops when meetign first
2691 * transno bigger than last_committed.
2692 * caller must hold imp->imp_lock
2694 void ptlrpc_free_committed(struct obd_import *imp)
2696 struct ptlrpc_request *req, *saved;
2697 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2698 bool skip_committed_list = true;
2701 LASSERT(imp != NULL);
2702 assert_spin_locked(&imp->imp_lock);
2704 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2705 imp->imp_generation == imp->imp_last_generation_checked) {
2706 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2707 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2710 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2711 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2712 imp->imp_generation);
2714 if (imp->imp_generation != imp->imp_last_generation_checked ||
2715 imp->imp_last_transno_checked == 0)
2716 skip_committed_list = false;
2718 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2719 imp->imp_last_generation_checked = imp->imp_generation;
2721 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2723 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2724 LASSERT(req != last_req);
2727 if (req->rq_transno == 0) {
2728 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2731 if (req->rq_import_generation < imp->imp_generation) {
2732 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2736 /* not yet committed */
2737 if (req->rq_transno > imp->imp_peer_committed_transno) {
2738 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2742 if (req->rq_replay) {
2743 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2744 list_move_tail(&req->rq_replay_list,
2745 &imp->imp_committed_list);
2749 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2750 imp->imp_peer_committed_transno);
2752 ptlrpc_free_request(req);
2755 if (skip_committed_list)
2758 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2760 LASSERT(req->rq_transno != 0);
2761 if (req->rq_import_generation < imp->imp_generation ||
2763 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2764 req->rq_import_generation <
2765 imp->imp_generation ? "stale" : "closed");
2767 if (imp->imp_replay_cursor == &req->rq_replay_list)
2768 imp->imp_replay_cursor =
2769 req->rq_replay_list.next;
2771 ptlrpc_free_request(req);
2778 void ptlrpc_cleanup_client(struct obd_import *imp)
2785 * Schedule previously sent request for resend.
2786 * For bulk requests we assign new xid (to avoid problems with
2787 * lost replies and therefore several transfers landing into same buffer
2788 * from different sending attempts).
2790 void ptlrpc_resend_req(struct ptlrpc_request *req)
2792 DEBUG_REQ(D_HA, req, "going to resend");
2793 spin_lock(&req->rq_lock);
2795 /* Request got reply but linked to the import list still.
2796 Let ptlrpc_check_set() to process it. */
2797 if (ptlrpc_client_replied(req)) {
2798 spin_unlock(&req->rq_lock);
2799 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2803 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2804 req->rq_status = -EAGAIN;
2807 req->rq_net_err = 0;
2808 req->rq_timedout = 0;
2810 ptlrpc_client_wake_req(req);
2811 spin_unlock(&req->rq_lock);
2814 /* XXX: this function and rq_status are currently unused */
2815 void ptlrpc_restart_req(struct ptlrpc_request *req)
2817 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2818 req->rq_status = -ERESTARTSYS;
2820 spin_lock(&req->rq_lock);
2821 req->rq_restart = 1;
2822 req->rq_timedout = 0;
2823 ptlrpc_client_wake_req(req);
2824 spin_unlock(&req->rq_lock);
2828 * Grab additional reference on a request \a req
2830 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2833 atomic_inc(&req->rq_refcount);
2836 EXPORT_SYMBOL(ptlrpc_request_addref);
2839 * Add a request to import replay_list.
2840 * Must be called under imp_lock
2842 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2843 struct obd_import *imp)
2845 struct list_head *tmp;
2847 assert_spin_locked(&imp->imp_lock);
2849 if (req->rq_transno == 0) {
2850 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2854 /* clear this for new requests that were resent as well
2855 as resent replayed requests. */
2856 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2858 /* don't re-add requests that have been replayed */
2859 if (!list_empty(&req->rq_replay_list))
2862 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2864 spin_lock(&req->rq_lock);
2866 spin_unlock(&req->rq_lock);
2868 LASSERT(imp->imp_replayable);
2869 /* Balanced in ptlrpc_free_committed, usually. */
2870 ptlrpc_request_addref(req);
2871 list_for_each_prev(tmp, &imp->imp_replay_list) {
2872 struct ptlrpc_request *iter = list_entry(tmp,
2873 struct ptlrpc_request,
2876 /* We may have duplicate transnos if we create and then
2877 * open a file, or for closes retained if to match creating
2878 * opens, so use req->rq_xid as a secondary key.
2879 * (See bugs 684, 685, and 428.)
2880 * XXX no longer needed, but all opens need transnos!
2882 if (iter->rq_transno > req->rq_transno)
2885 if (iter->rq_transno == req->rq_transno) {
2886 LASSERT(iter->rq_xid != req->rq_xid);
2887 if (iter->rq_xid > req->rq_xid)
2891 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2895 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2899 * Send request and wait until it completes.
2900 * Returns request processing status.
2902 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2904 struct ptlrpc_request_set *set;
2908 LASSERT(req->rq_set == NULL);
2909 LASSERT(!req->rq_receiving_reply);
2911 set = ptlrpc_prep_set();
2913 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2917 /* for distributed debugging */
2918 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2920 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2921 ptlrpc_request_addref(req);
2922 ptlrpc_set_add_req(set, req);
2923 rc = ptlrpc_set_wait(set);
2924 ptlrpc_set_destroy(set);
2928 EXPORT_SYMBOL(ptlrpc_queue_wait);
2931 * Callback used for replayed requests reply processing.
2932 * In case of successful reply calls registered request replay callback.
2933 * In case of error restart replay process.
2935 static int ptlrpc_replay_interpret(const struct lu_env *env,
2936 struct ptlrpc_request *req,
2937 void * data, int rc)
2939 struct ptlrpc_replay_async_args *aa = data;
2940 struct obd_import *imp = req->rq_import;
2943 atomic_dec(&imp->imp_replay_inflight);
2945 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2946 * server got the request, but bulk transfer timeout, let's
2947 * replay the bulk req again */
2948 if (!ptlrpc_client_replied(req) ||
2949 (req->rq_bulk != NULL &&
2950 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2951 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2952 GOTO(out, rc = -ETIMEDOUT);
2955 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2956 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2957 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2958 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2960 /** VBR: check version failure */
2961 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2962 /** replay was failed due to version mismatch */
2963 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2964 spin_lock(&imp->imp_lock);
2965 imp->imp_vbr_failed = 1;
2966 imp->imp_no_lock_replay = 1;
2967 spin_unlock(&imp->imp_lock);
2968 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2970 /** The transno had better not change over replay. */
2971 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2972 lustre_msg_get_transno(req->rq_repmsg) ||
2973 lustre_msg_get_transno(req->rq_repmsg) == 0,
2975 lustre_msg_get_transno(req->rq_reqmsg),
2976 lustre_msg_get_transno(req->rq_repmsg));
2979 spin_lock(&imp->imp_lock);
2980 /** if replays by version then gap occur on server, no trust to locks */
2981 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2982 imp->imp_no_lock_replay = 1;
2983 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2984 spin_unlock(&imp->imp_lock);
2985 LASSERT(imp->imp_last_replay_transno);
2987 /* transaction number shouldn't be bigger than the latest replayed */
2988 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2989 DEBUG_REQ(D_ERROR, req,
2990 "Reported transno %llu is bigger than the "
2991 "replayed one: %llu", req->rq_transno,
2992 lustre_msg_get_transno(req->rq_reqmsg));
2993 GOTO(out, rc = -EINVAL);
2996 DEBUG_REQ(D_HA, req, "got rep");
2998 /* let the callback do fixups, possibly including in the request */
2999 if (req->rq_replay_cb)
3000 req->rq_replay_cb(req);
3002 if (ptlrpc_client_replied(req) &&
3003 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3004 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3005 lustre_msg_get_status(req->rq_repmsg),
3006 aa->praa_old_status);
3008 /* Note: If the replay fails for MDT-MDT recovery, let's
3009 * abort all of the following requests in the replay
3010 * and sending list, because MDT-MDT update requests
3011 * are dependent on each other, see LU-7039 */
3012 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3013 struct ptlrpc_request *free_req;
3014 struct ptlrpc_request *tmp;
3016 spin_lock(&imp->imp_lock);
3017 list_for_each_entry_safe(free_req, tmp,
3018 &imp->imp_replay_list,
3020 ptlrpc_free_request(free_req);
3023 list_for_each_entry_safe(free_req, tmp,
3024 &imp->imp_committed_list,
3026 ptlrpc_free_request(free_req);
3029 list_for_each_entry_safe(free_req, tmp,
3030 &imp->imp_delayed_list,
3032 spin_lock(&free_req->rq_lock);
3033 free_req->rq_err = 1;
3034 free_req->rq_status = -EIO;
3035 ptlrpc_client_wake_req(free_req);
3036 spin_unlock(&free_req->rq_lock);
3039 list_for_each_entry_safe(free_req, tmp,
3040 &imp->imp_sending_list,
3042 spin_lock(&free_req->rq_lock);
3043 free_req->rq_err = 1;
3044 free_req->rq_status = -EIO;
3045 ptlrpc_client_wake_req(free_req);
3046 spin_unlock(&free_req->rq_lock);
3048 spin_unlock(&imp->imp_lock);
3051 /* Put it back for re-replay. */
3052 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3056 * Errors while replay can set transno to 0, but
3057 * imp_last_replay_transno shouldn't be set to 0 anyway
3059 if (req->rq_transno == 0)
3060 CERROR("Transno is 0 during replay!\n");
3062 /* continue with recovery */
3063 rc = ptlrpc_import_recovery_state_machine(imp);
3065 req->rq_send_state = aa->praa_old_state;
3068 /* this replay failed, so restart recovery */
3069 ptlrpc_connect_import(imp);
3075 * Prepares and queues request for replay.
3076 * Adds it to ptlrpcd queue for actual sending.
3077 * Returns 0 on success.
3079 int ptlrpc_replay_req(struct ptlrpc_request *req)
3081 struct ptlrpc_replay_async_args *aa;
3084 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3086 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
3087 aa = ptlrpc_req_async_args(req);
3088 memset(aa, 0, sizeof *aa);
3090 /* Prepare request to be resent with ptlrpcd */
3091 aa->praa_old_state = req->rq_send_state;
3092 req->rq_send_state = LUSTRE_IMP_REPLAY;
3093 req->rq_phase = RQ_PHASE_NEW;
3094 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3096 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3098 req->rq_interpret_reply = ptlrpc_replay_interpret;
3099 /* Readjust the timeout for current conditions */
3100 ptlrpc_at_set_req_timeout(req);
3102 /* Tell server the net_latency, so the server can calculate how long
3103 * it should wait for next replay */
3104 lustre_msg_set_service_time(req->rq_reqmsg,
3105 ptlrpc_at_get_net_latency(req));
3106 DEBUG_REQ(D_HA, req, "REPLAY");
3108 atomic_inc(&req->rq_import->imp_replay_inflight);
3109 spin_lock(&req->rq_lock);
3110 req->rq_early_free_repbuf = 0;
3111 spin_unlock(&req->rq_lock);
3112 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3114 ptlrpcd_add_req(req);
3119 * Aborts all in-flight request on import \a imp sending and delayed lists
3121 void ptlrpc_abort_inflight(struct obd_import *imp)
3123 struct list_head *tmp, *n;
3126 /* Make sure that no new requests get processed for this import.
3127 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3128 * this flag and then putting requests on sending_list or delayed_list.
3130 spin_lock(&imp->imp_lock);
3132 /* XXX locking? Maybe we should remove each request with the list
3133 * locked? Also, how do we know if the requests on the list are
3134 * being freed at this time?
3136 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3137 struct ptlrpc_request *req = list_entry(tmp,
3138 struct ptlrpc_request,
3141 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3143 spin_lock(&req->rq_lock);
3144 if (req->rq_import_generation < imp->imp_generation) {
3146 req->rq_status = -EIO;
3147 ptlrpc_client_wake_req(req);
3149 spin_unlock(&req->rq_lock);
3152 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3153 struct ptlrpc_request *req =
3154 list_entry(tmp, struct ptlrpc_request, rq_list);
3156 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3158 spin_lock(&req->rq_lock);
3159 if (req->rq_import_generation < imp->imp_generation) {
3161 req->rq_status = -EIO;
3162 ptlrpc_client_wake_req(req);
3164 spin_unlock(&req->rq_lock);
3167 /* Last chance to free reqs left on the replay list, but we
3168 * will still leak reqs that haven't committed. */
3169 if (imp->imp_replayable)
3170 ptlrpc_free_committed(imp);
3172 spin_unlock(&imp->imp_lock);
3178 * Abort all uncompleted requests in request set \a set
3180 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3182 struct list_head *tmp, *pos;
3184 LASSERT(set != NULL);
3186 list_for_each_safe(pos, tmp, &set->set_requests) {
3187 struct ptlrpc_request *req =
3188 list_entry(pos, struct ptlrpc_request,
3191 spin_lock(&req->rq_lock);
3192 if (req->rq_phase != RQ_PHASE_RPC) {
3193 spin_unlock(&req->rq_lock);
3198 req->rq_status = -EINTR;
3199 ptlrpc_client_wake_req(req);
3200 spin_unlock(&req->rq_lock);
3204 static __u64 ptlrpc_last_xid;
3205 static spinlock_t ptlrpc_last_xid_lock;
3208 * Initialize the XID for the node. This is common among all requests on
3209 * this node, and only requires the property that it is monotonically
3210 * increasing. It does not need to be sequential. Since this is also used
3211 * as the RDMA match bits, it is important that a single client NOT have
3212 * the same match bits for two different in-flight requests, hence we do
3213 * NOT want to have an XID per target or similar.
3215 * To avoid an unlikely collision between match bits after a client reboot
3216 * (which would deliver old data into the wrong RDMA buffer) initialize
3217 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3218 * If the time is clearly incorrect, we instead use a 62-bit random number.
3219 * In the worst case the random number will overflow 1M RPCs per second in
3220 * 9133 years, or permutations thereof.
3222 #define YEAR_2004 (1ULL << 30)
3223 void ptlrpc_init_xid(void)
3225 time64_t now = ktime_get_real_seconds();
3227 spin_lock_init(&ptlrpc_last_xid_lock);
3228 if (now < YEAR_2004) {
3229 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3230 ptlrpc_last_xid >>= 2;
3231 ptlrpc_last_xid |= (1ULL << 61);
3233 ptlrpc_last_xid = (__u64)now << 20;
3236 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3237 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3238 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3242 * Increase xid and returns resulting new value to the caller.
3244 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3245 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3246 * itself uses the last bulk xid needed, so the server can determine the
3247 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3248 * xid must align to a power-of-two value.
3250 * This is assumed to be true due to the initial ptlrpc_last_xid
3251 * value also being initialized to a power-of-two value. LU-1431
3253 __u64 ptlrpc_next_xid(void)
3257 spin_lock(&ptlrpc_last_xid_lock);
3258 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3259 ptlrpc_last_xid = next;
3260 spin_unlock(&ptlrpc_last_xid_lock);
3266 * If request has a new allocated XID (new request or EINPROGRESS resend),
3267 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3268 * request to ensure previous bulk fails and avoid problems with lost replies
3269 * and therefore several transfers landing into the same buffer from different
3272 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3274 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3276 LASSERT(bd != NULL);
3278 /* Generate new matchbits for all resend requests, including
3280 if (req->rq_resend) {
3281 __u64 old_mbits = req->rq_mbits;
3283 /* First time resend on -EINPROGRESS will generate new xid,
3284 * so we can actually use the rq_xid as rq_mbits in such case,
3285 * however, it's bit hard to distinguish such resend with a
3286 * 'resend for the -EINPROGRESS resend'. To make it simple,
3287 * we opt to generate mbits for all resend cases. */
3288 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
3289 req->rq_mbits = ptlrpc_next_xid();
3291 /* Old version transfers rq_xid to peer as
3293 spin_lock(&req->rq_import->imp_lock);
3294 list_del_init(&req->rq_unreplied_list);
3295 ptlrpc_assign_next_xid_nolock(req);
3296 spin_unlock(&req->rq_import->imp_lock);
3297 req->rq_mbits = req->rq_xid;
3299 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3300 old_mbits, req->rq_mbits);
3301 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3302 /* Request being sent first time, use xid as matchbits. */
3303 req->rq_mbits = req->rq_xid;
3305 /* Replay request, xid and matchbits have already been
3306 * correctly assigned. */
3310 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3311 * that server can infer the number of bulks that were prepared,
3313 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3316 /* Set rq_xid as rq_mbits to indicate the final bulk for the old
3317 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3319 * It's ok to directly set the rq_xid here, since this xid bump
3320 * won't affect the request position in unreplied list. */
3321 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3322 req->rq_xid = req->rq_mbits;
3326 * Get a glimpse at what next xid value might have been.
3327 * Returns possible next xid.
3329 __u64 ptlrpc_sample_next_xid(void)
3331 #if BITS_PER_LONG == 32
3332 /* need to avoid possible word tearing on 32-bit systems */
3335 spin_lock(&ptlrpc_last_xid_lock);
3336 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3337 spin_unlock(&ptlrpc_last_xid_lock);
3341 /* No need to lock, since returned value is racy anyways */
3342 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3345 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3348 * Functions for operating ptlrpc workers.
3350 * A ptlrpc work is a function which will be running inside ptlrpc context.
3351 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3353 * 1. after a work is created, it can be used many times, that is:
3354 * handler = ptlrpcd_alloc_work();
3355 * ptlrpcd_queue_work();
3357 * queue it again when necessary:
3358 * ptlrpcd_queue_work();
3359 * ptlrpcd_destroy_work();
3360 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3361 * it will only be queued once in any time. Also as its name implies, it may
3362 * have delay before it really runs by ptlrpcd thread.
3364 struct ptlrpc_work_async_args {
3365 int (*cb)(const struct lu_env *, void *);
3369 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3371 /* re-initialize the req */
3372 req->rq_timeout = obd_timeout;
3373 req->rq_sent = ktime_get_real_seconds();
3374 req->rq_deadline = req->rq_sent + req->rq_timeout;
3375 req->rq_phase = RQ_PHASE_INTERPRET;
3376 req->rq_next_phase = RQ_PHASE_COMPLETE;
3377 req->rq_xid = ptlrpc_next_xid();
3378 req->rq_import_generation = req->rq_import->imp_generation;
3380 ptlrpcd_add_req(req);
3383 static int work_interpreter(const struct lu_env *env,
3384 struct ptlrpc_request *req, void *data, int rc)
3386 struct ptlrpc_work_async_args *arg = data;
3388 LASSERT(ptlrpcd_check_work(req));
3389 LASSERT(arg->cb != NULL);
3391 rc = arg->cb(env, arg->cbdata);
3393 list_del_init(&req->rq_set_chain);
3396 if (atomic_dec_return(&req->rq_refcount) > 1) {
3397 atomic_set(&req->rq_refcount, 2);
3398 ptlrpcd_add_work_req(req);
3403 static int worker_format;
3405 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3407 return req->rq_pill.rc_fmt == (void *)&worker_format;
3411 * Create a work for ptlrpc.
3413 void *ptlrpcd_alloc_work(struct obd_import *imp,
3414 int (*cb)(const struct lu_env *, void *), void *cbdata)
3416 struct ptlrpc_request *req = NULL;
3417 struct ptlrpc_work_async_args *args;
3423 RETURN(ERR_PTR(-EINVAL));
3425 /* copy some code from deprecated fakereq. */
3426 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3428 CERROR("ptlrpc: run out of memory!\n");
3429 RETURN(ERR_PTR(-ENOMEM));
3432 ptlrpc_cli_req_init(req);
3434 req->rq_send_state = LUSTRE_IMP_FULL;
3435 req->rq_type = PTL_RPC_MSG_REQUEST;
3436 req->rq_import = class_import_get(imp);
3437 req->rq_interpret_reply = work_interpreter;
3438 /* don't want reply */
3439 req->rq_no_delay = req->rq_no_resend = 1;
3440 req->rq_pill.rc_fmt = (void *)&worker_format;
3442 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3443 args = ptlrpc_req_async_args(req);
3445 args->cbdata = cbdata;
3449 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3451 void ptlrpcd_destroy_work(void *handler)
3453 struct ptlrpc_request *req = handler;
3456 ptlrpc_req_finished(req);
3458 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3460 int ptlrpcd_queue_work(void *handler)
3462 struct ptlrpc_request *req = handler;
3465 * Check if the req is already being queued.
3467 * Here comes a trick: it lacks a way of checking if a req is being
3468 * processed reliably in ptlrpc. Here I have to use refcount of req
3469 * for this purpose. This is okay because the caller should use this
3470 * req as opaque data. - Jinshan
3472 LASSERT(atomic_read(&req->rq_refcount) > 0);
3473 if (atomic_inc_return(&req->rq_refcount) == 2)
3474 ptlrpcd_add_work_req(req);
3477 EXPORT_SYMBOL(ptlrpcd_queue_work);