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, 2015, 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)
84 struct ptlrpc_connection *c;
86 lnet_process_id_t peer;
89 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
90 * before accessing its values. */
91 /* coverity[uninit_use_in_call] */
92 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
94 CNETERR("cannot find peer %s!\n", uuid->uuid);
98 c = ptlrpc_connection_get(peer, self, uuid);
100 memcpy(c->c_remote_uuid.uuid,
101 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
104 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
110 * Allocate and initialize new bulk descriptor on the sender.
111 * Returns pointer to the descriptor or NULL on error.
113 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
114 enum ptlrpc_bulk_op_type type,
116 const struct ptlrpc_bulk_frag_ops *ops)
118 struct ptlrpc_bulk_desc *desc;
121 /* ensure that only one of KIOV or IOVEC is set but not both */
122 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
123 ops->add_kiov_frag != NULL) ||
124 (ptlrpc_is_bulk_desc_kvec(type) &&
125 ops->add_iov_frag != NULL));
130 if (type & PTLRPC_BULK_BUF_KIOV) {
131 OBD_ALLOC_LARGE(GET_KIOV(desc),
132 nfrags * sizeof(*GET_KIOV(desc)));
133 if (GET_KIOV(desc) == NULL)
136 OBD_ALLOC_LARGE(GET_KVEC(desc),
137 nfrags * sizeof(*GET_KVEC(desc)));
138 if (GET_KVEC(desc) == NULL)
142 spin_lock_init(&desc->bd_lock);
143 init_waitqueue_head(&desc->bd_waitq);
144 desc->bd_max_iov = nfrags;
145 desc->bd_iov_count = 0;
146 desc->bd_portal = portal;
147 desc->bd_type = type;
148 desc->bd_md_count = 0;
149 desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
150 LASSERT(max_brw > 0);
151 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
152 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
153 * node. Negotiated ocd_brw_size will always be <= this number. */
154 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
155 LNetInvalidateHandle(&desc->bd_mds[i]);
164 * Prepare bulk descriptor for specified outgoing request \a req that
165 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
166 * the bulk to be sent. Used on client-side.
167 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
170 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
171 unsigned nfrags, unsigned max_brw,
174 const struct ptlrpc_bulk_frag_ops
177 struct obd_import *imp = req->rq_import;
178 struct ptlrpc_bulk_desc *desc;
181 LASSERT(ptlrpc_is_bulk_op_passive(type));
183 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
187 desc->bd_import_generation = req->rq_import_generation;
188 desc->bd_import = class_import_get(imp);
191 desc->bd_cbid.cbid_fn = client_bulk_callback;
192 desc->bd_cbid.cbid_arg = desc;
194 /* This makes req own desc, and free it when she frees herself */
199 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
201 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
202 struct page *page, int pageoffset, int len,
207 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
208 LASSERT(page != NULL);
209 LASSERT(pageoffset >= 0);
211 LASSERT(pageoffset + len <= PAGE_SIZE);
212 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
214 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
221 kiov->kiov_page = page;
222 kiov->kiov_offset = pageoffset;
223 kiov->kiov_len = len;
225 desc->bd_iov_count++;
227 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
229 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
235 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
236 LASSERT(frag != NULL);
238 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
240 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
244 iovec->iov_base = frag;
245 iovec->iov_len = len;
247 desc->bd_iov_count++;
249 RETURN(desc->bd_nob);
251 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
253 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
257 LASSERT(desc != NULL);
258 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
259 LASSERT(desc->bd_md_count == 0); /* network hands off */
260 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
261 LASSERT(desc->bd_frag_ops != NULL);
263 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
264 sptlrpc_enc_pool_put_pages(desc);
267 class_export_put(desc->bd_export);
269 class_import_put(desc->bd_import);
271 if (desc->bd_frag_ops->release_frags != NULL)
272 desc->bd_frag_ops->release_frags(desc);
274 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
275 OBD_FREE_LARGE(GET_KIOV(desc),
276 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
278 OBD_FREE_LARGE(GET_KVEC(desc),
279 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
283 EXPORT_SYMBOL(ptlrpc_free_bulk);
286 * Set server timelimit for this req, i.e. how long are we willing to wait
287 * for reply before timing out this request.
289 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
295 LASSERT(req->rq_import);
298 /* non-AT settings */
300 * \a imp_server_timeout means this is reverse import and
301 * we send (currently only) ASTs to the client and cannot afford
302 * to wait too long for the reply, otherwise the other client
303 * (because of which we are sending this request) would
304 * timeout waiting for us
306 req->rq_timeout = req->rq_import->imp_server_timeout ?
307 obd_timeout / 2 : obd_timeout;
309 at = &req->rq_import->imp_at;
310 idx = import_at_get_index(req->rq_import,
311 req->rq_request_portal);
312 serv_est = at_get(&at->iat_service_estimate[idx]);
313 req->rq_timeout = at_est2timeout(serv_est);
315 /* We could get even fancier here, using history to predict increased
318 /* Let the server know what this RPC timeout is by putting it in the
320 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
322 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
324 /* Adjust max service estimate based on server value */
325 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
326 unsigned int serv_est)
332 LASSERT(req->rq_import);
333 at = &req->rq_import->imp_at;
335 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
336 /* max service estimates are tracked on the server side,
337 so just keep minimal history here */
338 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
340 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
341 "has changed from %d to %d\n",
342 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
343 oldse, at_get(&at->iat_service_estimate[idx]));
346 /* Expected network latency per remote node (secs) */
347 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
349 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
352 /* Adjust expected network latency */
353 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
354 unsigned int service_time)
356 unsigned int nl, oldnl;
358 time_t now = cfs_time_current_sec();
360 LASSERT(req->rq_import);
362 if (service_time > now - req->rq_sent + 3) {
363 /* bz16408, however, this can also happen if early reply
364 * is lost and client RPC is expired and resent, early reply
365 * or reply of original RPC can still be fit in reply buffer
366 * of resent RPC, now client is measuring time from the
367 * resent time, but server sent back service time of original
370 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
371 D_ADAPTTO : D_WARNING,
372 "Reported service time %u > total measured time "
373 CFS_DURATION_T"\n", service_time,
374 cfs_time_sub(now, req->rq_sent));
378 /* Network latency is total time less server processing time */
379 nl = max_t(int, now - req->rq_sent -
380 service_time, 0) + 1; /* st rounding */
381 at = &req->rq_import->imp_at;
383 oldnl = at_measured(&at->iat_net_latency, nl);
385 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
386 "has changed from %d to %d\n",
387 req->rq_import->imp_obd->obd_name,
389 &req->rq_import->imp_connection->c_remote_uuid),
390 oldnl, at_get(&at->iat_net_latency));
393 static int unpack_reply(struct ptlrpc_request *req)
397 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
398 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
400 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
405 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
407 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
414 * Handle an early reply message, called with the rq_lock held.
415 * If anything goes wrong just ignore it - same as if it never happened
417 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
418 __must_hold(&req->rq_lock)
420 struct ptlrpc_request *early_req;
426 spin_unlock(&req->rq_lock);
428 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
430 spin_lock(&req->rq_lock);
434 rc = unpack_reply(early_req);
436 sptlrpc_cli_finish_early_reply(early_req);
437 spin_lock(&req->rq_lock);
441 /* Use new timeout value just to adjust the local value for this
442 * request, don't include it into at_history. It is unclear yet why
443 * service time increased and should it be counted or skipped, e.g.
444 * that can be recovery case or some error or server, the real reply
445 * will add all new data if it is worth to add. */
446 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
447 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
449 /* Network latency can be adjusted, it is pure network delays */
450 ptlrpc_at_adj_net_latency(req,
451 lustre_msg_get_service_time(early_req->rq_repmsg));
453 sptlrpc_cli_finish_early_reply(early_req);
455 spin_lock(&req->rq_lock);
456 olddl = req->rq_deadline;
457 /* server assumes it now has rq_timeout from when the request
458 * arrived, so the client should give it at least that long.
459 * since we don't know the arrival time we'll use the original
461 req->rq_deadline = req->rq_sent + req->rq_timeout +
462 ptlrpc_at_get_net_latency(req);
464 DEBUG_REQ(D_ADAPTTO, req,
465 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
466 "("CFS_DURATION_T"s)", req->rq_early_count,
467 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
468 cfs_time_sub(req->rq_deadline, olddl));
473 static struct kmem_cache *request_cache;
475 int ptlrpc_request_cache_init(void)
477 request_cache = kmem_cache_create("ptlrpc_cache",
478 sizeof(struct ptlrpc_request),
479 0, SLAB_HWCACHE_ALIGN, NULL);
480 return request_cache == NULL ? -ENOMEM : 0;
483 void ptlrpc_request_cache_fini(void)
485 kmem_cache_destroy(request_cache);
488 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
490 struct ptlrpc_request *req;
492 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
496 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
498 OBD_SLAB_FREE_PTR(req, request_cache);
502 * Wind down request pool \a pool.
503 * Frees all requests from the pool too
505 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
507 struct list_head *l, *tmp;
508 struct ptlrpc_request *req;
510 LASSERT(pool != NULL);
512 spin_lock(&pool->prp_lock);
513 list_for_each_safe(l, tmp, &pool->prp_req_list) {
514 req = list_entry(l, struct ptlrpc_request, rq_list);
515 list_del(&req->rq_list);
516 LASSERT(req->rq_reqbuf);
517 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
518 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
519 ptlrpc_request_cache_free(req);
521 spin_unlock(&pool->prp_lock);
522 OBD_FREE(pool, sizeof(*pool));
524 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
527 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
529 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
534 while (size < pool->prp_rq_size)
537 LASSERTF(list_empty(&pool->prp_req_list) ||
538 size == pool->prp_rq_size,
539 "Trying to change pool size with nonempty pool "
540 "from %d to %d bytes\n", pool->prp_rq_size, size);
542 spin_lock(&pool->prp_lock);
543 pool->prp_rq_size = size;
544 for (i = 0; i < num_rq; i++) {
545 struct ptlrpc_request *req;
546 struct lustre_msg *msg;
548 spin_unlock(&pool->prp_lock);
549 req = ptlrpc_request_cache_alloc(GFP_NOFS);
552 OBD_ALLOC_LARGE(msg, size);
554 ptlrpc_request_cache_free(req);
557 req->rq_reqbuf = msg;
558 req->rq_reqbuf_len = size;
560 spin_lock(&pool->prp_lock);
561 list_add_tail(&req->rq_list, &pool->prp_req_list);
563 spin_unlock(&pool->prp_lock);
566 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
569 * Create and initialize new request pool with given attributes:
570 * \a num_rq - initial number of requests to create for the pool
571 * \a msgsize - maximum message size possible for requests in thid pool
572 * \a populate_pool - function to be called when more requests need to be added
574 * Returns pointer to newly created pool or NULL on error.
576 struct ptlrpc_request_pool *
577 ptlrpc_init_rq_pool(int num_rq, int msgsize,
578 int (*populate_pool)(struct ptlrpc_request_pool *, int))
580 struct ptlrpc_request_pool *pool;
582 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
586 /* Request next power of two for the allocation, because internally
587 kernel would do exactly this */
589 spin_lock_init(&pool->prp_lock);
590 INIT_LIST_HEAD(&pool->prp_req_list);
591 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
592 pool->prp_populate = populate_pool;
594 populate_pool(pool, num_rq);
598 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
601 * Fetches one request from pool \a pool
603 static struct ptlrpc_request *
604 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
606 struct ptlrpc_request *request;
607 struct lustre_msg *reqbuf;
612 spin_lock(&pool->prp_lock);
614 /* See if we have anything in a pool, and bail out if nothing,
615 * in writeout path, where this matters, this is safe to do, because
616 * nothing is lost in this case, and when some in-flight requests
617 * complete, this code will be called again. */
618 if (unlikely(list_empty(&pool->prp_req_list))) {
619 spin_unlock(&pool->prp_lock);
623 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
625 list_del_init(&request->rq_list);
626 spin_unlock(&pool->prp_lock);
628 LASSERT(request->rq_reqbuf);
629 LASSERT(request->rq_pool);
631 reqbuf = request->rq_reqbuf;
632 memset(request, 0, sizeof(*request));
633 request->rq_reqbuf = reqbuf;
634 request->rq_reqbuf_len = pool->prp_rq_size;
635 request->rq_pool = pool;
641 * Returns freed \a request to pool.
643 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
645 struct ptlrpc_request_pool *pool = request->rq_pool;
647 spin_lock(&pool->prp_lock);
648 LASSERT(list_empty(&request->rq_list));
649 LASSERT(!request->rq_receiving_reply);
650 list_add_tail(&request->rq_list, &pool->prp_req_list);
651 spin_unlock(&pool->prp_lock);
654 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
656 struct obd_import *imp = req->rq_import;
657 struct list_head *tmp;
658 struct ptlrpc_request *iter;
660 assert_spin_locked(&imp->imp_lock);
661 LASSERT(list_empty(&req->rq_unreplied_list));
663 /* unreplied list is sorted by xid in ascending order */
664 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
665 iter = list_entry(tmp, struct ptlrpc_request,
668 LASSERT(req->rq_xid != iter->rq_xid);
669 if (req->rq_xid < iter->rq_xid)
671 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
674 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
677 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
679 req->rq_xid = ptlrpc_next_xid();
680 ptlrpc_add_unreplied(req);
683 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
685 spin_lock(&req->rq_import->imp_lock);
686 ptlrpc_assign_next_xid_nolock(req);
687 spin_unlock(&req->rq_import->imp_lock);
690 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
691 __u32 version, int opcode, char **bufs,
692 struct ptlrpc_cli_ctx *ctx)
695 struct obd_import *imp;
701 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
702 imp = request->rq_import;
703 lengths = request->rq_pill.rc_area[RCL_CLIENT];
706 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
708 rc = sptlrpc_req_get_ctx(request);
712 sptlrpc_req_set_flavor(request, opcode);
714 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
719 lustre_msg_add_version(request->rq_reqmsg, version);
720 request->rq_send_state = LUSTRE_IMP_FULL;
721 request->rq_type = PTL_RPC_MSG_REQUEST;
723 request->rq_req_cbid.cbid_fn = request_out_callback;
724 request->rq_req_cbid.cbid_arg = request;
726 request->rq_reply_cbid.cbid_fn = reply_in_callback;
727 request->rq_reply_cbid.cbid_arg = request;
729 request->rq_reply_deadline = 0;
730 request->rq_bulk_deadline = 0;
731 request->rq_req_deadline = 0;
732 request->rq_phase = RQ_PHASE_NEW;
733 request->rq_next_phase = RQ_PHASE_UNDEFINED;
735 request->rq_request_portal = imp->imp_client->cli_request_portal;
736 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
738 ptlrpc_at_set_req_timeout(request);
740 lustre_msg_set_opc(request->rq_reqmsg, opcode);
741 ptlrpc_assign_next_xid(request);
743 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
744 if (cfs_fail_val == opcode) {
745 time_t *fail_t = NULL, *fail2_t = NULL;
747 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
748 fail_t = &request->rq_bulk_deadline;
749 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
750 fail_t = &request->rq_reply_deadline;
751 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
752 fail_t = &request->rq_req_deadline;
753 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
754 fail_t = &request->rq_reply_deadline;
755 fail2_t = &request->rq_bulk_deadline;
759 *fail_t = cfs_time_current_sec() + LONG_UNLINK;
762 *fail2_t = cfs_time_current_sec() + LONG_UNLINK;
764 /* The RPC is infected, let the test to change the
766 set_current_state(TASK_UNINTERRUPTIBLE);
767 schedule_timeout(cfs_time_seconds(2));
768 set_current_state(TASK_RUNNING);
775 LASSERT(!request->rq_pool);
776 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
778 class_import_put(imp);
783 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
786 * Pack request buffers for network transfer, performing necessary encryption
787 * steps if necessary.
789 int ptlrpc_request_pack(struct ptlrpc_request *request,
790 __u32 version, int opcode)
793 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
797 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
798 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
799 * have to send old ptlrpc_body to keep interoprability with these
802 * Only three kinds of server->client RPCs so far:
807 * XXX This should be removed whenever we drop the interoprability with
808 * the these old clients.
810 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
811 opcode == LDLM_GL_CALLBACK)
812 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
813 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
817 EXPORT_SYMBOL(ptlrpc_request_pack);
820 * Helper function to allocate new request on import \a imp
821 * and possibly using existing request from pool \a pool if provided.
822 * Returns allocated request structure with import field filled or
826 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
827 struct ptlrpc_request_pool *pool)
829 struct ptlrpc_request *request = NULL;
831 request = ptlrpc_request_cache_alloc(GFP_NOFS);
833 if (!request && pool)
834 request = ptlrpc_prep_req_from_pool(pool);
837 ptlrpc_cli_req_init(request);
839 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
840 LASSERT(imp != LP_POISON);
841 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
843 LASSERT(imp->imp_client != LP_POISON);
845 request->rq_import = class_import_get(imp);
847 CERROR("request allocation out of memory\n");
854 * Helper function for creating a request.
855 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
856 * buffer structures according to capsule template \a format.
857 * Returns allocated request structure pointer or NULL on error.
859 static struct ptlrpc_request *
860 ptlrpc_request_alloc_internal(struct obd_import *imp,
861 struct ptlrpc_request_pool * pool,
862 const struct req_format *format)
864 struct ptlrpc_request *request;
866 request = __ptlrpc_request_alloc(imp, pool);
870 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
871 req_capsule_set(&request->rq_pill, format);
876 * Allocate new request structure for import \a imp and initialize its
877 * buffer structure according to capsule template \a format.
879 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
880 const struct req_format *format)
882 return ptlrpc_request_alloc_internal(imp, NULL, format);
884 EXPORT_SYMBOL(ptlrpc_request_alloc);
887 * Allocate new request structure for import \a imp from pool \a pool and
888 * initialize its buffer structure according to capsule template \a format.
890 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
891 struct ptlrpc_request_pool * pool,
892 const struct req_format *format)
894 return ptlrpc_request_alloc_internal(imp, pool, format);
896 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
899 * For requests not from pool, free memory of the request structure.
900 * For requests obtained from a pool earlier, return request back to pool.
902 void ptlrpc_request_free(struct ptlrpc_request *request)
904 if (request->rq_pool)
905 __ptlrpc_free_req_to_pool(request);
907 ptlrpc_request_cache_free(request);
909 EXPORT_SYMBOL(ptlrpc_request_free);
912 * Allocate new request for operatione \a opcode and immediatelly pack it for
914 * Only used for simple requests like OBD_PING where the only important
915 * part of the request is operation itself.
916 * Returns allocated request or NULL on error.
918 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
919 const struct req_format *format,
920 __u32 version, int opcode)
922 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
926 rc = ptlrpc_request_pack(req, version, opcode);
928 ptlrpc_request_free(req);
934 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
937 * Allocate and initialize new request set structure on the current CPT.
938 * Returns a pointer to the newly allocated set structure or NULL on error.
940 struct ptlrpc_request_set *ptlrpc_prep_set(void)
942 struct ptlrpc_request_set *set;
946 cpt = cfs_cpt_current(cfs_cpt_table, 0);
947 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
950 atomic_set(&set->set_refcount, 1);
951 INIT_LIST_HEAD(&set->set_requests);
952 init_waitqueue_head(&set->set_waitq);
953 atomic_set(&set->set_new_count, 0);
954 atomic_set(&set->set_remaining, 0);
955 spin_lock_init(&set->set_new_req_lock);
956 INIT_LIST_HEAD(&set->set_new_requests);
957 INIT_LIST_HEAD(&set->set_cblist);
958 set->set_max_inflight = UINT_MAX;
959 set->set_producer = NULL;
960 set->set_producer_arg = NULL;
965 EXPORT_SYMBOL(ptlrpc_prep_set);
968 * Allocate and initialize new request set structure with flow control
969 * extension. This extension allows to control the number of requests in-flight
970 * for the whole set. A callback function to generate requests must be provided
971 * and the request set will keep the number of requests sent over the wire to
973 * Returns a pointer to the newly allocated set structure or NULL on error.
975 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
979 struct ptlrpc_request_set *set;
981 set = ptlrpc_prep_set();
985 set->set_max_inflight = max;
986 set->set_producer = func;
987 set->set_producer_arg = arg;
993 * Wind down and free request set structure previously allocated with
995 * Ensures that all requests on the set have completed and removes
996 * all requests from the request list in a set.
997 * If any unsent request happen to be on the list, pretends that they got
998 * an error in flight and calls their completion handler.
1000 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1002 struct list_head *tmp;
1003 struct list_head *next;
1008 /* Requests on the set should either all be completed, or all be new */
1009 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1010 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1011 list_for_each(tmp, &set->set_requests) {
1012 struct ptlrpc_request *req =
1013 list_entry(tmp, struct ptlrpc_request,
1016 LASSERT(req->rq_phase == expected_phase);
1020 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1021 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1022 atomic_read(&set->set_remaining), n);
1024 list_for_each_safe(tmp, next, &set->set_requests) {
1025 struct ptlrpc_request *req =
1026 list_entry(tmp, struct ptlrpc_request,
1028 list_del_init(&req->rq_set_chain);
1030 LASSERT(req->rq_phase == expected_phase);
1032 if (req->rq_phase == RQ_PHASE_NEW) {
1033 ptlrpc_req_interpret(NULL, req, -EBADR);
1034 atomic_dec(&set->set_remaining);
1037 spin_lock(&req->rq_lock);
1039 req->rq_invalid_rqset = 0;
1040 spin_unlock(&req->rq_lock);
1042 ptlrpc_req_finished (req);
1045 LASSERT(atomic_read(&set->set_remaining) == 0);
1047 ptlrpc_reqset_put(set);
1050 EXPORT_SYMBOL(ptlrpc_set_destroy);
1053 * Add a callback function \a fn to the set.
1054 * This function would be called when all requests on this set are completed.
1055 * The function will be passed \a data argument.
1057 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1058 set_interpreter_func fn, void *data)
1060 struct ptlrpc_set_cbdata *cbdata;
1062 OBD_ALLOC_PTR(cbdata);
1066 cbdata->psc_interpret = fn;
1067 cbdata->psc_data = data;
1068 list_add_tail(&cbdata->psc_item, &set->set_cblist);
1074 * Add a new request to the general purpose request set.
1075 * Assumes request reference from the caller.
1077 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1078 struct ptlrpc_request *req)
1080 LASSERT(list_empty(&req->rq_set_chain));
1082 if (req->rq_allow_intr)
1083 set->set_allow_intr = 1;
1085 /* The set takes over the caller's request reference */
1086 list_add_tail(&req->rq_set_chain, &set->set_requests);
1088 atomic_inc(&set->set_remaining);
1089 req->rq_queued_time = cfs_time_current();
1091 if (req->rq_reqmsg != NULL)
1092 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1094 if (set->set_producer != NULL)
1095 /* If the request set has a producer callback, the RPC must be
1096 * sent straight away */
1097 ptlrpc_send_new_req(req);
1099 EXPORT_SYMBOL(ptlrpc_set_add_req);
1102 * Add a request to a request with dedicated server thread
1103 * and wake the thread to make any necessary processing.
1104 * Currently only used for ptlrpcd.
1106 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1107 struct ptlrpc_request *req)
1109 struct ptlrpc_request_set *set = pc->pc_set;
1112 LASSERT(req->rq_set == NULL);
1113 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1115 spin_lock(&set->set_new_req_lock);
1117 * The set takes over the caller's request reference.
1120 req->rq_queued_time = cfs_time_current();
1121 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1122 count = atomic_inc_return(&set->set_new_count);
1123 spin_unlock(&set->set_new_req_lock);
1125 /* Only need to call wakeup once for the first entry. */
1127 wake_up(&set->set_waitq);
1129 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1130 * guarantee the async RPC can be processed ASAP, we have
1131 * no other better choice. It maybe fixed in future. */
1132 for (i = 0; i < pc->pc_npartners; i++)
1133 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1138 * Based on the current state of the import, determine if the request
1139 * can be sent, is an error, or should be delayed.
1141 * Returns true if this request should be delayed. If false, and
1142 * *status is set, then the request can not be sent and *status is the
1143 * error code. If false and status is 0, then request can be sent.
1145 * The imp->imp_lock must be held.
1147 static int ptlrpc_import_delay_req(struct obd_import *imp,
1148 struct ptlrpc_request *req, int *status)
1153 LASSERT (status != NULL);
1156 if (req->rq_ctx_init || req->rq_ctx_fini) {
1157 /* always allow ctx init/fini rpc go through */
1158 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1159 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1161 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1162 /* pings may safely race with umount */
1163 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1164 D_HA : D_ERROR, req, "IMP_CLOSED ");
1166 } else if (ptlrpc_send_limit_expired(req)) {
1167 /* probably doesn't need to be a D_ERROR after initial testing*/
1168 DEBUG_REQ(D_HA, req, "send limit expired ");
1169 *status = -ETIMEDOUT;
1170 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1171 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1172 /* allow CONNECT even if import is invalid */ ;
1173 if (atomic_read(&imp->imp_inval_count) != 0) {
1174 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1177 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1178 if (!imp->imp_deactive)
1179 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1180 *status = -ESHUTDOWN; /* bz 12940 */
1181 } else if (req->rq_import_generation != imp->imp_generation) {
1182 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1184 } else if (req->rq_send_state != imp->imp_state) {
1185 /* invalidate in progress - any requests should be drop */
1186 if (atomic_read(&imp->imp_inval_count) != 0) {
1187 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1189 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1190 *status = -EWOULDBLOCK;
1191 } else if (req->rq_allow_replay &&
1192 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1193 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1194 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1195 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1196 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1206 * Decide if the error message should be printed to the console or not.
1207 * Makes its decision based on request type, status, and failure frequency.
1209 * \param[in] req request that failed and may need a console message
1211 * \retval false if no message should be printed
1212 * \retval true if console message should be printed
1214 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1218 LASSERT(req->rq_reqmsg != NULL);
1219 opc = lustre_msg_get_opc(req->rq_reqmsg);
1221 /* Suppress particular reconnect errors which are to be expected. */
1222 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1225 /* Suppress timed out reconnect requests */
1226 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1230 /* Suppress most unavailable/again reconnect requests, but
1231 * print occasionally so it is clear client is trying to
1232 * connect to a server where no target is running. */
1233 err = lustre_msg_get_status(req->rq_repmsg);
1234 if ((err == -ENODEV || err == -EAGAIN) &&
1235 req->rq_import->imp_conn_cnt % 30 != 20)
1243 * Check request processing status.
1244 * Returns the status.
1246 static int ptlrpc_check_status(struct ptlrpc_request *req)
1251 err = lustre_msg_get_status(req->rq_repmsg);
1252 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1253 struct obd_import *imp = req->rq_import;
1254 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1255 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1257 /* -EAGAIN is normal when using POSIX flocks */
1258 if (ptlrpc_console_allow(req) &&
1259 !(opc == LDLM_ENQUEUE && err == -EAGAIN))
1260 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1261 "failed: rc = %d\n",
1262 imp->imp_obd->obd_name,
1264 libcfs_nid2str(nid), err);
1265 RETURN(err < 0 ? err : -EINVAL);
1269 DEBUG_REQ(D_INFO, req, "status is %d", err);
1270 } else if (err > 0) {
1271 /* XXX: translate this error from net to host */
1272 DEBUG_REQ(D_INFO, req, "status is %d", err);
1279 * save pre-versions of objects into request for replay.
1280 * Versions are obtained from server reply.
1283 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1285 struct lustre_msg *repmsg = req->rq_repmsg;
1286 struct lustre_msg *reqmsg = req->rq_reqmsg;
1287 __u64 *versions = lustre_msg_get_versions(repmsg);
1290 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1294 lustre_msg_set_versions(reqmsg, versions);
1295 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1296 versions[0], versions[1]);
1301 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1303 struct ptlrpc_request *req;
1305 assert_spin_locked(&imp->imp_lock);
1306 if (list_empty(&imp->imp_unreplied_list))
1309 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1311 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1313 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1314 imp->imp_known_replied_xid = req->rq_xid - 1;
1316 return req->rq_xid - 1;
1320 * Callback function called when client receives RPC reply for \a req.
1321 * Returns 0 on success or error code.
1322 * The return alue would be assigned to req->rq_status by the caller
1323 * as request processing status.
1324 * This function also decides if the request needs to be saved for later replay.
1326 static int after_reply(struct ptlrpc_request *req)
1328 struct obd_import *imp = req->rq_import;
1329 struct obd_device *obd = req->rq_import->imp_obd;
1331 struct timeval work_start;
1336 LASSERT(obd != NULL);
1337 /* repbuf must be unlinked */
1338 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1340 if (req->rq_reply_truncated) {
1341 if (ptlrpc_no_resend(req)) {
1342 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1343 " expected: %d, actual size: %d",
1344 req->rq_nob_received, req->rq_repbuf_len);
1348 sptlrpc_cli_free_repbuf(req);
1349 /* Pass the required reply buffer size (include
1350 * space for early reply).
1351 * NB: no need to roundup because alloc_repbuf
1352 * will roundup it */
1353 req->rq_replen = req->rq_nob_received;
1354 req->rq_nob_received = 0;
1355 spin_lock(&req->rq_lock);
1357 spin_unlock(&req->rq_lock);
1361 do_gettimeofday(&work_start);
1362 timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1365 * NB Until this point, the whole of the incoming message,
1366 * including buflens, status etc is in the sender's byte order.
1368 rc = sptlrpc_cli_unwrap_reply(req);
1370 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1375 * Security layer unwrap might ask resend this request.
1380 rc = unpack_reply(req);
1384 /* retry indefinitely on EINPROGRESS */
1385 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1386 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1387 time_t now = cfs_time_current_sec();
1389 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1390 spin_lock(&req->rq_lock);
1392 spin_unlock(&req->rq_lock);
1393 req->rq_nr_resend++;
1395 /* Readjust the timeout for current conditions */
1396 ptlrpc_at_set_req_timeout(req);
1397 /* delay resend to give a chance to the server to get ready.
1398 * The delay is increased by 1s on every resend and is capped to
1399 * the current request timeout (i.e. obd_timeout if AT is off,
1400 * or AT service time x 125% + 5s, see at_est2timeout) */
1401 if (req->rq_nr_resend > req->rq_timeout)
1402 req->rq_sent = now + req->rq_timeout;
1404 req->rq_sent = now + req->rq_nr_resend;
1406 /* Resend for EINPROGRESS will use a new XID */
1407 spin_lock(&imp->imp_lock);
1408 list_del_init(&req->rq_unreplied_list);
1409 spin_unlock(&imp->imp_lock);
1414 if (obd->obd_svc_stats != NULL) {
1415 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1417 ptlrpc_lprocfs_rpc_sent(req, timediff);
1420 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1421 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1422 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1423 lustre_msg_get_type(req->rq_repmsg));
1427 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1428 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1429 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1430 ptlrpc_at_adj_net_latency(req,
1431 lustre_msg_get_service_time(req->rq_repmsg));
1433 rc = ptlrpc_check_status(req);
1434 imp->imp_connect_error = rc;
1438 * Either we've been evicted, or the server has failed for
1439 * some reason. Try to reconnect, and if that fails, punt to
1442 if (ptlrpc_recoverable_error(rc)) {
1443 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1444 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1447 ptlrpc_request_handle_notconn(req);
1452 * Let's look if server sent slv. Do it only for RPC with
1455 ldlm_cli_update_pool(req);
1459 * Store transno in reqmsg for replay.
1461 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1462 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1463 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1466 if (imp->imp_replayable) {
1467 spin_lock(&imp->imp_lock);
1469 * No point in adding already-committed requests to the replay
1470 * list, we will just remove them immediately. b=9829
1472 if (req->rq_transno != 0 &&
1474 lustre_msg_get_last_committed(req->rq_repmsg) ||
1476 /** version recovery */
1477 ptlrpc_save_versions(req);
1478 ptlrpc_retain_replayable_request(req, imp);
1479 } else if (req->rq_commit_cb != NULL &&
1480 list_empty(&req->rq_replay_list)) {
1481 /* NB: don't call rq_commit_cb if it's already on
1482 * rq_replay_list, ptlrpc_free_committed() will call
1483 * it later, see LU-3618 for details */
1484 spin_unlock(&imp->imp_lock);
1485 req->rq_commit_cb(req);
1486 spin_lock(&imp->imp_lock);
1490 * Replay-enabled imports return commit-status information.
1492 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1493 if (likely(committed > imp->imp_peer_committed_transno))
1494 imp->imp_peer_committed_transno = committed;
1496 ptlrpc_free_committed(imp);
1498 if (!list_empty(&imp->imp_replay_list)) {
1499 struct ptlrpc_request *last;
1501 last = list_entry(imp->imp_replay_list.prev,
1502 struct ptlrpc_request,
1505 * Requests with rq_replay stay on the list even if no
1506 * commit is expected.
1508 if (last->rq_transno > imp->imp_peer_committed_transno)
1509 ptlrpc_pinger_commit_expected(imp);
1512 spin_unlock(&imp->imp_lock);
1519 * Helper function to send request \a req over the network for the first time
1520 * Also adjusts request phase.
1521 * Returns 0 on success or error code.
1523 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1525 struct obd_import *imp = req->rq_import;
1530 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1532 /* do not try to go further if there is not enough memory in enc_pool */
1533 if (req->rq_sent && req->rq_bulk != NULL)
1534 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1535 pool_is_at_full_capacity())
1538 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1539 (!req->rq_generation_set ||
1540 req->rq_import_generation == imp->imp_generation))
1543 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1545 spin_lock(&imp->imp_lock);
1547 LASSERT(req->rq_xid != 0);
1548 LASSERT(!list_empty(&req->rq_unreplied_list));
1550 if (!req->rq_generation_set)
1551 req->rq_import_generation = imp->imp_generation;
1553 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1554 spin_lock(&req->rq_lock);
1555 req->rq_waiting = 1;
1556 spin_unlock(&req->rq_lock);
1558 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1559 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1560 ptlrpc_import_state_name(req->rq_send_state),
1561 ptlrpc_import_state_name(imp->imp_state));
1562 LASSERT(list_empty(&req->rq_list));
1563 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1564 atomic_inc(&req->rq_import->imp_inflight);
1565 spin_unlock(&imp->imp_lock);
1570 spin_unlock(&imp->imp_lock);
1571 req->rq_status = rc;
1572 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1576 LASSERT(list_empty(&req->rq_list));
1577 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1578 atomic_inc(&req->rq_import->imp_inflight);
1580 /* find the known replied XID from the unreplied list, CONNECT
1581 * and DISCONNECT requests are skipped to make the sanity check
1582 * on server side happy. see process_req_last_xid().
1584 * For CONNECT: Because replay requests have lower XID, it'll
1585 * break the sanity check if CONNECT bump the exp_last_xid on
1588 * For DISCONNECT: Since client will abort inflight RPC before
1589 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1590 * than the inflight RPC.
1592 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1593 min_xid = ptlrpc_known_replied_xid(imp);
1594 spin_unlock(&imp->imp_lock);
1596 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1598 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1600 rc = sptlrpc_req_refresh_ctx(req, -1);
1603 req->rq_status = rc;
1606 spin_lock(&req->rq_lock);
1607 req->rq_wait_ctx = 1;
1608 spin_unlock(&req->rq_lock);
1613 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1614 " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1615 imp->imp_obd->obd_uuid.uuid,
1616 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1617 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1618 lustre_msg_get_opc(req->rq_reqmsg));
1620 rc = ptl_send_rpc(req, 0);
1621 if (rc == -ENOMEM) {
1622 spin_lock(&imp->imp_lock);
1623 if (!list_empty(&req->rq_list)) {
1624 list_del_init(&req->rq_list);
1625 atomic_dec(&req->rq_import->imp_inflight);
1627 spin_unlock(&imp->imp_lock);
1628 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1632 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1633 spin_lock(&req->rq_lock);
1634 req->rq_net_err = 1;
1635 spin_unlock(&req->rq_lock);
1641 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1646 LASSERT(set->set_producer != NULL);
1648 remaining = atomic_read(&set->set_remaining);
1650 /* populate the ->set_requests list with requests until we
1651 * reach the maximum number of RPCs in flight for this set */
1652 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1653 rc = set->set_producer(set, set->set_producer_arg);
1654 if (rc == -ENOENT) {
1655 /* no more RPC to produce */
1656 set->set_producer = NULL;
1657 set->set_producer_arg = NULL;
1662 RETURN((atomic_read(&set->set_remaining) - remaining));
1666 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1667 * and no more replies are expected.
1668 * (it is possible to get less replies than requests sent e.g. due to timed out
1669 * requests or requests that we had trouble to send out)
1671 * NOTE: This function contains a potential schedule point (cond_resched()).
1673 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1675 struct list_head *tmp, *next;
1676 struct list_head comp_reqs;
1677 int force_timer_recalc = 0;
1680 if (atomic_read(&set->set_remaining) == 0)
1683 INIT_LIST_HEAD(&comp_reqs);
1684 list_for_each_safe(tmp, next, &set->set_requests) {
1685 struct ptlrpc_request *req =
1686 list_entry(tmp, struct ptlrpc_request,
1688 struct obd_import *imp = req->rq_import;
1689 int unregistered = 0;
1693 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1694 list_move_tail(&req->rq_set_chain, &comp_reqs);
1698 /* This schedule point is mainly for the ptlrpcd caller of this
1699 * function. Most ptlrpc sets are not long-lived and unbounded
1700 * in length, but at the least the set used by the ptlrpcd is.
1701 * Since the processing time is unbounded, we need to insert an
1702 * explicit schedule point to make the thread well-behaved.
1706 /* If the caller requires to allow to be interpreted by force
1707 * and it has really been interpreted, then move the request
1708 * to RQ_PHASE_INTERPRET phase in spite of what the current
1710 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1711 req->rq_status = -EINTR;
1712 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1714 /* Since it is interpreted and we have to wait for
1715 * the reply to be unlinked, then use sync mode. */
1718 GOTO(interpret, req->rq_status);
1721 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1722 force_timer_recalc = 1;
1724 /* delayed send - skip */
1725 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1728 /* delayed resend - skip */
1729 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1730 req->rq_sent > cfs_time_current_sec())
1733 if (!(req->rq_phase == RQ_PHASE_RPC ||
1734 req->rq_phase == RQ_PHASE_BULK ||
1735 req->rq_phase == RQ_PHASE_INTERPRET ||
1736 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1737 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1738 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1742 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1743 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1744 LASSERT(req->rq_next_phase != req->rq_phase);
1745 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1747 if (req->rq_req_deadline &&
1748 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1749 req->rq_req_deadline = 0;
1750 if (req->rq_reply_deadline &&
1751 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1752 req->rq_reply_deadline = 0;
1753 if (req->rq_bulk_deadline &&
1754 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1755 req->rq_bulk_deadline = 0;
1758 * Skip processing until reply is unlinked. We
1759 * can't return to pool before that and we can't
1760 * call interpret before that. We need to make
1761 * sure that all rdma transfers finished and will
1762 * not corrupt any data.
1764 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1765 ptlrpc_client_recv_or_unlink(req))
1767 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1768 ptlrpc_client_bulk_active(req))
1772 * Turn fail_loc off to prevent it from looping
1775 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1776 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1779 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1780 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1785 * Move to next phase if reply was successfully
1788 ptlrpc_rqphase_move(req, req->rq_next_phase);
1791 if (req->rq_phase == RQ_PHASE_INTERPRET)
1792 GOTO(interpret, req->rq_status);
1795 * Note that this also will start async reply unlink.
1797 if (req->rq_net_err && !req->rq_timedout) {
1798 ptlrpc_expire_one_request(req, 1);
1801 * Check if we still need to wait for unlink.
1803 if (ptlrpc_client_recv_or_unlink(req) ||
1804 ptlrpc_client_bulk_active(req))
1806 /* If there is no need to resend, fail it now. */
1807 if (req->rq_no_resend) {
1808 if (req->rq_status == 0)
1809 req->rq_status = -EIO;
1810 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1811 GOTO(interpret, req->rq_status);
1818 spin_lock(&req->rq_lock);
1819 req->rq_replied = 0;
1820 spin_unlock(&req->rq_lock);
1821 if (req->rq_status == 0)
1822 req->rq_status = -EIO;
1823 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1824 GOTO(interpret, req->rq_status);
1827 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1828 * so it sets rq_intr regardless of individual rpc
1829 * timeouts. The synchronous IO waiting path sets
1830 * rq_intr irrespective of whether ptlrpcd
1831 * has seen a timeout. Our policy is to only interpret
1832 * interrupted rpcs after they have timed out, so we
1833 * need to enforce that here.
1836 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1837 req->rq_wait_ctx)) {
1838 req->rq_status = -EINTR;
1839 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1840 GOTO(interpret, req->rq_status);
1843 if (req->rq_phase == RQ_PHASE_RPC) {
1844 if (req->rq_timedout || req->rq_resend ||
1845 req->rq_waiting || req->rq_wait_ctx) {
1848 if (!ptlrpc_unregister_reply(req, 1)) {
1849 ptlrpc_unregister_bulk(req, 1);
1853 spin_lock(&imp->imp_lock);
1854 if (ptlrpc_import_delay_req(imp, req, &status)){
1855 /* put on delay list - only if we wait
1856 * recovery finished - before send */
1857 list_del_init(&req->rq_list);
1858 list_add_tail(&req->rq_list,
1861 spin_unlock(&imp->imp_lock);
1866 req->rq_status = status;
1867 ptlrpc_rqphase_move(req,
1868 RQ_PHASE_INTERPRET);
1869 spin_unlock(&imp->imp_lock);
1870 GOTO(interpret, req->rq_status);
1872 if (ptlrpc_no_resend(req) &&
1873 !req->rq_wait_ctx) {
1874 req->rq_status = -ENOTCONN;
1875 ptlrpc_rqphase_move(req,
1876 RQ_PHASE_INTERPRET);
1877 spin_unlock(&imp->imp_lock);
1878 GOTO(interpret, req->rq_status);
1881 list_del_init(&req->rq_list);
1882 list_add_tail(&req->rq_list,
1883 &imp->imp_sending_list);
1885 spin_unlock(&imp->imp_lock);
1887 spin_lock(&req->rq_lock);
1888 req->rq_waiting = 0;
1889 spin_unlock(&req->rq_lock);
1891 if (req->rq_timedout || req->rq_resend) {
1892 /* This is re-sending anyways,
1893 * let's mark req as resend. */
1894 spin_lock(&req->rq_lock);
1896 spin_unlock(&req->rq_lock);
1898 if (req->rq_bulk != NULL &&
1899 !ptlrpc_unregister_bulk(req, 1))
1903 * rq_wait_ctx is only touched by ptlrpcd,
1904 * so no lock is needed here.
1906 status = sptlrpc_req_refresh_ctx(req, -1);
1909 req->rq_status = status;
1910 spin_lock(&req->rq_lock);
1911 req->rq_wait_ctx = 0;
1912 spin_unlock(&req->rq_lock);
1913 force_timer_recalc = 1;
1915 spin_lock(&req->rq_lock);
1916 req->rq_wait_ctx = 1;
1917 spin_unlock(&req->rq_lock);
1922 spin_lock(&req->rq_lock);
1923 req->rq_wait_ctx = 0;
1924 spin_unlock(&req->rq_lock);
1927 rc = ptl_send_rpc(req, 0);
1928 if (rc == -ENOMEM) {
1929 spin_lock(&imp->imp_lock);
1930 if (!list_empty(&req->rq_list))
1931 list_del_init(&req->rq_list);
1932 spin_unlock(&imp->imp_lock);
1933 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1937 DEBUG_REQ(D_HA, req,
1938 "send failed: rc = %d", rc);
1939 force_timer_recalc = 1;
1940 spin_lock(&req->rq_lock);
1941 req->rq_net_err = 1;
1942 spin_unlock(&req->rq_lock);
1945 /* need to reset the timeout */
1946 force_timer_recalc = 1;
1949 spin_lock(&req->rq_lock);
1951 if (ptlrpc_client_early(req)) {
1952 ptlrpc_at_recv_early_reply(req);
1953 spin_unlock(&req->rq_lock);
1957 /* Still waiting for a reply? */
1958 if (ptlrpc_client_recv(req)) {
1959 spin_unlock(&req->rq_lock);
1963 /* Did we actually receive a reply? */
1964 if (!ptlrpc_client_replied(req)) {
1965 spin_unlock(&req->rq_lock);
1969 spin_unlock(&req->rq_lock);
1971 /* unlink from net because we are going to
1972 * swab in-place of reply buffer */
1973 unregistered = ptlrpc_unregister_reply(req, 1);
1977 req->rq_status = after_reply(req);
1981 /* If there is no bulk associated with this request,
1982 * then we're done and should let the interpreter
1983 * process the reply. Similarly if the RPC returned
1984 * an error, and therefore the bulk will never arrive.
1986 if (req->rq_bulk == NULL || req->rq_status < 0) {
1987 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1988 GOTO(interpret, req->rq_status);
1991 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1994 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1995 if (ptlrpc_client_bulk_active(req))
1998 if (req->rq_bulk->bd_failure) {
1999 /* The RPC reply arrived OK, but the bulk screwed
2000 * up! Dead weird since the server told us the RPC
2001 * was good after getting the REPLY for her GET or
2002 * the ACK for her PUT. */
2003 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2004 req->rq_status = -EIO;
2007 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2010 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2012 /* This moves to "unregistering" phase we need to wait for
2014 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2015 /* start async bulk unlink too */
2016 ptlrpc_unregister_bulk(req, 1);
2020 if (!ptlrpc_unregister_bulk(req, async))
2023 /* When calling interpret receiving already should be
2025 LASSERT(!req->rq_receiving_reply);
2027 ptlrpc_req_interpret(env, req, req->rq_status);
2029 if (ptlrpcd_check_work(req)) {
2030 atomic_dec(&set->set_remaining);
2033 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2035 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
2036 "Completed RPC pname:cluuid:pid:xid:nid:"
2037 "opc %s:%s:%d:%llu:%s:%d\n",
2038 current_comm(), imp->imp_obd->obd_uuid.uuid,
2039 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2040 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2041 lustre_msg_get_opc(req->rq_reqmsg));
2043 spin_lock(&imp->imp_lock);
2044 /* Request already may be not on sending or delaying list. This
2045 * may happen in the case of marking it erroneous for the case
2046 * ptlrpc_import_delay_req(req, status) find it impossible to
2047 * allow sending this rpc and returns *status != 0. */
2048 if (!list_empty(&req->rq_list)) {
2049 list_del_init(&req->rq_list);
2050 atomic_dec(&imp->imp_inflight);
2052 list_del_init(&req->rq_unreplied_list);
2053 spin_unlock(&imp->imp_lock);
2055 atomic_dec(&set->set_remaining);
2056 wake_up_all(&imp->imp_recovery_waitq);
2058 if (set->set_producer) {
2059 /* produce a new request if possible */
2060 if (ptlrpc_set_producer(set) > 0)
2061 force_timer_recalc = 1;
2063 /* free the request that has just been completed
2064 * in order not to pollute set->set_requests */
2065 list_del_init(&req->rq_set_chain);
2066 spin_lock(&req->rq_lock);
2068 req->rq_invalid_rqset = 0;
2069 spin_unlock(&req->rq_lock);
2071 /* record rq_status to compute the final status later */
2072 if (req->rq_status != 0)
2073 set->set_rc = req->rq_status;
2074 ptlrpc_req_finished(req);
2076 list_move_tail(&req->rq_set_chain, &comp_reqs);
2080 /* move completed request at the head of list so it's easier for
2081 * caller to find them */
2082 list_splice(&comp_reqs, &set->set_requests);
2084 /* If we hit an error, we want to recover promptly. */
2085 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2087 EXPORT_SYMBOL(ptlrpc_check_set);
2090 * Time out request \a req. is \a async_unlink is set, that means do not wait
2091 * until LNet actually confirms network buffer unlinking.
2092 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2094 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2096 struct obd_import *imp = req->rq_import;
2100 spin_lock(&req->rq_lock);
2101 req->rq_timedout = 1;
2102 spin_unlock(&req->rq_lock);
2104 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
2105 "/real "CFS_DURATION_T"]",
2106 req->rq_net_err ? "failed due to network error" :
2107 ((req->rq_real_sent == 0 ||
2108 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
2109 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
2110 "timed out for sent delay" : "timed out for slow reply"),
2111 req->rq_sent, req->rq_real_sent);
2113 if (imp != NULL && obd_debug_peer_on_timeout)
2114 LNetDebugPeer(imp->imp_connection->c_peer);
2116 ptlrpc_unregister_reply(req, async_unlink);
2117 ptlrpc_unregister_bulk(req, async_unlink);
2119 if (obd_dump_on_timeout)
2120 libcfs_debug_dumplog();
2123 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2127 atomic_inc(&imp->imp_timeouts);
2129 /* The DLM server doesn't want recovery run on its imports. */
2130 if (imp->imp_dlm_fake)
2133 /* If this request is for recovery or other primordial tasks,
2134 * then error it out here. */
2135 if (req->rq_ctx_init || req->rq_ctx_fini ||
2136 req->rq_send_state != LUSTRE_IMP_FULL ||
2137 imp->imp_obd->obd_no_recov) {
2138 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2139 ptlrpc_import_state_name(req->rq_send_state),
2140 ptlrpc_import_state_name(imp->imp_state));
2141 spin_lock(&req->rq_lock);
2142 req->rq_status = -ETIMEDOUT;
2144 spin_unlock(&req->rq_lock);
2148 /* if a request can't be resent we can't wait for an answer after
2150 if (ptlrpc_no_resend(req)) {
2151 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2155 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2161 * Time out all uncompleted requests in request set pointed by \a data
2162 * Callback used when waiting on sets with l_wait_event.
2165 int ptlrpc_expired_set(void *data)
2167 struct ptlrpc_request_set *set = data;
2168 struct list_head *tmp;
2169 time_t now = cfs_time_current_sec();
2172 LASSERT(set != NULL);
2175 * A timeout expired. See which reqs it applies to...
2177 list_for_each(tmp, &set->set_requests) {
2178 struct ptlrpc_request *req =
2179 list_entry(tmp, struct ptlrpc_request,
2182 /* don't expire request waiting for context */
2183 if (req->rq_wait_ctx)
2186 /* Request in-flight? */
2187 if (!((req->rq_phase == RQ_PHASE_RPC &&
2188 !req->rq_waiting && !req->rq_resend) ||
2189 (req->rq_phase == RQ_PHASE_BULK)))
2192 if (req->rq_timedout || /* already dealt with */
2193 req->rq_deadline > now) /* not expired */
2196 /* Deal with this guy. Do it asynchronously to not block
2197 * ptlrpcd thread. */
2198 ptlrpc_expire_one_request(req, 1);
2202 * When waiting for a whole set, we always break out of the
2203 * sleep so we can recalculate the timeout, or enable interrupts
2204 * if everyone's timed out.
2210 * Sets rq_intr flag in \a req under spinlock.
2212 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2214 spin_lock(&req->rq_lock);
2216 spin_unlock(&req->rq_lock);
2218 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2221 * Interrupts (sets interrupted flag) all uncompleted requests in
2222 * a set \a data. Callback for l_wait_event for interruptible waits.
2224 static void ptlrpc_interrupted_set(void *data)
2226 struct ptlrpc_request_set *set = data;
2227 struct list_head *tmp;
2229 LASSERT(set != NULL);
2230 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2232 list_for_each(tmp, &set->set_requests) {
2233 struct ptlrpc_request *req =
2234 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2239 if (req->rq_phase != RQ_PHASE_RPC &&
2240 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2241 !req->rq_allow_intr)
2244 ptlrpc_mark_interrupted(req);
2249 * Get the smallest timeout in the set; this does NOT set a timeout.
2251 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2253 struct list_head *tmp;
2254 time_t now = cfs_time_current_sec();
2256 struct ptlrpc_request *req;
2260 list_for_each(tmp, &set->set_requests) {
2261 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2264 * Request in-flight?
2266 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2267 (req->rq_phase == RQ_PHASE_BULK) ||
2268 (req->rq_phase == RQ_PHASE_NEW)))
2272 * Already timed out.
2274 if (req->rq_timedout)
2280 if (req->rq_wait_ctx)
2283 if (req->rq_phase == RQ_PHASE_NEW)
2284 deadline = req->rq_sent;
2285 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2286 deadline = req->rq_sent;
2288 deadline = req->rq_sent + req->rq_timeout;
2290 if (deadline <= now) /* actually expired already */
2291 timeout = 1; /* ASAP */
2292 else if (timeout == 0 || timeout > deadline - now)
2293 timeout = deadline - now;
2299 * Send all unset request from the set and then wait untill all
2300 * requests in the set complete (either get a reply, timeout, get an
2301 * error or otherwise be interrupted).
2302 * Returns 0 on success or error code otherwise.
2304 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2306 struct list_head *tmp;
2307 struct ptlrpc_request *req;
2308 struct l_wait_info lwi;
2312 if (set->set_producer)
2313 (void)ptlrpc_set_producer(set);
2315 list_for_each(tmp, &set->set_requests) {
2316 req = list_entry(tmp, struct ptlrpc_request,
2318 if (req->rq_phase == RQ_PHASE_NEW)
2319 (void)ptlrpc_send_new_req(req);
2322 if (list_empty(&set->set_requests))
2326 timeout = ptlrpc_set_next_timeout(set);
2328 /* wait until all complete, interrupted, or an in-flight
2330 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2333 if ((timeout == 0 && !signal_pending(current)) ||
2334 set->set_allow_intr)
2335 /* No requests are in-flight (ether timed out
2336 * or delayed), so we can allow interrupts.
2337 * We still want to block for a limited time,
2338 * so we allow interrupts during the timeout. */
2339 lwi = LWI_TIMEOUT_INTR_ALL(
2340 cfs_time_seconds(timeout ? timeout : 1),
2342 ptlrpc_interrupted_set, set);
2345 * At least one request is in flight, so no
2346 * interrupts are allowed. Wait until all
2347 * complete, or an in-flight req times out.
2349 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2350 ptlrpc_expired_set, set);
2352 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2354 /* LU-769 - if we ignored the signal because it was already
2355 * pending when we started, we need to handle it now or we risk
2356 * it being ignored forever */
2357 if (rc == -ETIMEDOUT &&
2358 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2359 signal_pending(current)) {
2360 sigset_t blocked_sigs =
2361 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2363 /* In fact we only interrupt for the "fatal" signals
2364 * like SIGINT or SIGKILL. We still ignore less
2365 * important signals since ptlrpc set is not easily
2366 * reentrant from userspace again */
2367 if (signal_pending(current))
2368 ptlrpc_interrupted_set(set);
2369 cfs_restore_sigs(blocked_sigs);
2372 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2374 /* -EINTR => all requests have been flagged rq_intr so next
2376 * -ETIMEDOUT => someone timed out. When all reqs have
2377 * timed out, signals are enabled allowing completion with
2379 * I don't really care if we go once more round the loop in
2380 * the error cases -eeb. */
2381 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2382 list_for_each(tmp, &set->set_requests) {
2383 req = list_entry(tmp, struct ptlrpc_request,
2385 spin_lock(&req->rq_lock);
2386 req->rq_invalid_rqset = 1;
2387 spin_unlock(&req->rq_lock);
2390 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2392 LASSERT(atomic_read(&set->set_remaining) == 0);
2394 rc = set->set_rc; /* rq_status of already freed requests if any */
2395 list_for_each(tmp, &set->set_requests) {
2396 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2398 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2399 if (req->rq_status != 0)
2400 rc = req->rq_status;
2403 if (set->set_interpret != NULL) {
2404 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2406 rc = interpreter (set, set->set_arg, rc);
2408 struct ptlrpc_set_cbdata *cbdata, *n;
2411 list_for_each_entry_safe(cbdata, n,
2412 &set->set_cblist, psc_item) {
2413 list_del_init(&cbdata->psc_item);
2414 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2417 OBD_FREE_PTR(cbdata);
2423 EXPORT_SYMBOL(ptlrpc_set_wait);
2426 * Helper fuction for request freeing.
2427 * Called when request count reached zero and request needs to be freed.
2428 * Removes request from all sorts of sending/replay lists it might be on,
2429 * frees network buffers if any are present.
2430 * If \a locked is set, that means caller is already holding import imp_lock
2431 * and so we no longer need to reobtain it (for certain lists manipulations)
2433 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2437 if (request == NULL)
2440 LASSERT(!request->rq_srv_req);
2441 LASSERT(request->rq_export == NULL);
2442 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2443 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2444 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2445 LASSERTF(!request->rq_replay, "req %p\n", request);
2447 req_capsule_fini(&request->rq_pill);
2449 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2450 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2451 if (request->rq_import != NULL) {
2453 spin_lock(&request->rq_import->imp_lock);
2454 list_del_init(&request->rq_replay_list);
2455 list_del_init(&request->rq_unreplied_list);
2457 spin_unlock(&request->rq_import->imp_lock);
2459 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2461 if (atomic_read(&request->rq_refcount) != 0) {
2462 DEBUG_REQ(D_ERROR, request,
2463 "freeing request with nonzero refcount");
2467 if (request->rq_repbuf != NULL)
2468 sptlrpc_cli_free_repbuf(request);
2470 if (request->rq_import != NULL) {
2471 class_import_put(request->rq_import);
2472 request->rq_import = NULL;
2474 if (request->rq_bulk != NULL)
2475 ptlrpc_free_bulk(request->rq_bulk);
2477 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2478 sptlrpc_cli_free_reqbuf(request);
2480 if (request->rq_cli_ctx)
2481 sptlrpc_req_put_ctx(request, !locked);
2483 if (request->rq_pool)
2484 __ptlrpc_free_req_to_pool(request);
2486 ptlrpc_request_cache_free(request);
2490 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2492 * Drop one request reference. Must be called with import imp_lock held.
2493 * When reference count drops to zero, request is freed.
2495 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2497 assert_spin_locked(&request->rq_import->imp_lock);
2498 (void)__ptlrpc_req_finished(request, 1);
2503 * Drops one reference count for request \a request.
2504 * \a locked set indicates that caller holds import imp_lock.
2505 * Frees the request whe reference count reaches zero.
2507 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2510 if (request == NULL)
2513 if (request == LP_POISON ||
2514 request->rq_reqmsg == LP_POISON) {
2515 CERROR("dereferencing freed request (bug 575)\n");
2520 DEBUG_REQ(D_INFO, request, "refcount now %u",
2521 atomic_read(&request->rq_refcount) - 1);
2523 if (atomic_dec_and_test(&request->rq_refcount)) {
2524 __ptlrpc_free_req(request, locked);
2532 * Drops one reference count for a request.
2534 void ptlrpc_req_finished(struct ptlrpc_request *request)
2536 __ptlrpc_req_finished(request, 0);
2538 EXPORT_SYMBOL(ptlrpc_req_finished);
2541 * Returns xid of a \a request
2543 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2545 return request->rq_xid;
2547 EXPORT_SYMBOL(ptlrpc_req_xid);
2550 * Disengage the client's reply buffer from the network
2551 * NB does _NOT_ unregister any client-side bulk.
2552 * IDEMPOTENT, but _not_ safe against concurrent callers.
2553 * The request owner (i.e. the thread doing the I/O) must call...
2554 * Returns 0 on success or 1 if unregistering cannot be made.
2556 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2559 struct l_wait_info lwi;
2564 LASSERT(!in_interrupt());
2566 /* Let's setup deadline for reply unlink. */
2567 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2568 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2569 request->rq_reply_deadline =
2570 cfs_time_current_sec() + LONG_UNLINK;
2573 * Nothing left to do.
2575 if (!ptlrpc_client_recv_or_unlink(request))
2578 LNetMDUnlink(request->rq_reply_md_h);
2581 * Let's check it once again.
2583 if (!ptlrpc_client_recv_or_unlink(request))
2586 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2587 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2590 * Do not wait for unlink to finish.
2596 * We have to l_wait_event() whatever the result, to give liblustre
2597 * a chance to run reply_in_callback(), and to make sure we've
2598 * unlinked before returning a req to the pool.
2601 /* The wq argument is ignored by user-space wait_event macros */
2602 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2603 &request->rq_set->set_waitq :
2604 &request->rq_reply_waitq;
2605 /* Network access will complete in finite time but the HUGE
2606 * timeout lets us CWARN for visibility of sluggish NALs */
2607 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2608 cfs_time_seconds(1), NULL, NULL);
2609 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2612 ptlrpc_rqphase_move(request, request->rq_next_phase);
2616 LASSERT(rc == -ETIMEDOUT);
2617 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2618 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2619 request->rq_receiving_reply,
2620 request->rq_req_unlinked,
2621 request->rq_reply_unlinked);
2626 static void ptlrpc_free_request(struct ptlrpc_request *req)
2628 spin_lock(&req->rq_lock);
2630 spin_unlock(&req->rq_lock);
2632 if (req->rq_commit_cb != NULL)
2633 req->rq_commit_cb(req);
2634 list_del_init(&req->rq_replay_list);
2636 __ptlrpc_req_finished(req, 1);
2640 * the request is committed and dropped from the replay list of its import
2642 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2644 struct obd_import *imp = req->rq_import;
2646 spin_lock(&imp->imp_lock);
2647 if (list_empty(&req->rq_replay_list)) {
2648 spin_unlock(&imp->imp_lock);
2652 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2653 ptlrpc_free_request(req);
2655 spin_unlock(&imp->imp_lock);
2657 EXPORT_SYMBOL(ptlrpc_request_committed);
2660 * Iterates through replay_list on import and prunes
2661 * all requests have transno smaller than last_committed for the
2662 * import and don't have rq_replay set.
2663 * Since requests are sorted in transno order, stops when meetign first
2664 * transno bigger than last_committed.
2665 * caller must hold imp->imp_lock
2667 void ptlrpc_free_committed(struct obd_import *imp)
2669 struct ptlrpc_request *req, *saved;
2670 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2671 bool skip_committed_list = true;
2674 LASSERT(imp != NULL);
2675 assert_spin_locked(&imp->imp_lock);
2677 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2678 imp->imp_generation == imp->imp_last_generation_checked) {
2679 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2680 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2683 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2684 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2685 imp->imp_generation);
2687 if (imp->imp_generation != imp->imp_last_generation_checked ||
2688 imp->imp_last_transno_checked == 0)
2689 skip_committed_list = false;
2691 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2692 imp->imp_last_generation_checked = imp->imp_generation;
2694 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2696 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2697 LASSERT(req != last_req);
2700 if (req->rq_transno == 0) {
2701 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2704 if (req->rq_import_generation < imp->imp_generation) {
2705 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2709 /* not yet committed */
2710 if (req->rq_transno > imp->imp_peer_committed_transno) {
2711 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2715 if (req->rq_replay) {
2716 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2717 list_move_tail(&req->rq_replay_list,
2718 &imp->imp_committed_list);
2722 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2723 imp->imp_peer_committed_transno);
2725 ptlrpc_free_request(req);
2728 if (skip_committed_list)
2731 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2733 LASSERT(req->rq_transno != 0);
2734 if (req->rq_import_generation < imp->imp_generation) {
2735 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2736 ptlrpc_free_request(req);
2737 } else if (!req->rq_replay) {
2738 DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2739 ptlrpc_free_request(req);
2746 void ptlrpc_cleanup_client(struct obd_import *imp)
2753 * Schedule previously sent request for resend.
2754 * For bulk requests we assign new xid (to avoid problems with
2755 * lost replies and therefore several transfers landing into same buffer
2756 * from different sending attempts).
2758 void ptlrpc_resend_req(struct ptlrpc_request *req)
2760 DEBUG_REQ(D_HA, req, "going to resend");
2761 spin_lock(&req->rq_lock);
2763 /* Request got reply but linked to the import list still.
2764 Let ptlrpc_check_set() to process it. */
2765 if (ptlrpc_client_replied(req)) {
2766 spin_unlock(&req->rq_lock);
2767 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2771 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2772 req->rq_status = -EAGAIN;
2775 req->rq_net_err = 0;
2776 req->rq_timedout = 0;
2778 ptlrpc_client_wake_req(req);
2779 spin_unlock(&req->rq_lock);
2782 /* XXX: this function and rq_status are currently unused */
2783 void ptlrpc_restart_req(struct ptlrpc_request *req)
2785 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2786 req->rq_status = -ERESTARTSYS;
2788 spin_lock(&req->rq_lock);
2789 req->rq_restart = 1;
2790 req->rq_timedout = 0;
2791 ptlrpc_client_wake_req(req);
2792 spin_unlock(&req->rq_lock);
2796 * Grab additional reference on a request \a req
2798 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2801 atomic_inc(&req->rq_refcount);
2804 EXPORT_SYMBOL(ptlrpc_request_addref);
2807 * Add a request to import replay_list.
2808 * Must be called under imp_lock
2810 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2811 struct obd_import *imp)
2813 struct list_head *tmp;
2815 assert_spin_locked(&imp->imp_lock);
2817 if (req->rq_transno == 0) {
2818 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2822 /* clear this for new requests that were resent as well
2823 as resent replayed requests. */
2824 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2826 /* don't re-add requests that have been replayed */
2827 if (!list_empty(&req->rq_replay_list))
2830 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2832 spin_lock(&req->rq_lock);
2834 spin_unlock(&req->rq_lock);
2836 LASSERT(imp->imp_replayable);
2837 /* Balanced in ptlrpc_free_committed, usually. */
2838 ptlrpc_request_addref(req);
2839 list_for_each_prev(tmp, &imp->imp_replay_list) {
2840 struct ptlrpc_request *iter = list_entry(tmp,
2841 struct ptlrpc_request,
2844 /* We may have duplicate transnos if we create and then
2845 * open a file, or for closes retained if to match creating
2846 * opens, so use req->rq_xid as a secondary key.
2847 * (See bugs 684, 685, and 428.)
2848 * XXX no longer needed, but all opens need transnos!
2850 if (iter->rq_transno > req->rq_transno)
2853 if (iter->rq_transno == req->rq_transno) {
2854 LASSERT(iter->rq_xid != req->rq_xid);
2855 if (iter->rq_xid > req->rq_xid)
2859 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2863 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2867 * Send request and wait until it completes.
2868 * Returns request processing status.
2870 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2872 struct ptlrpc_request_set *set;
2876 LASSERT(req->rq_set == NULL);
2877 LASSERT(!req->rq_receiving_reply);
2879 set = ptlrpc_prep_set();
2881 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2885 /* for distributed debugging */
2886 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2888 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2889 ptlrpc_request_addref(req);
2890 ptlrpc_set_add_req(set, req);
2891 rc = ptlrpc_set_wait(set);
2892 ptlrpc_set_destroy(set);
2896 EXPORT_SYMBOL(ptlrpc_queue_wait);
2899 * Callback used for replayed requests reply processing.
2900 * In case of successful reply calls registered request replay callback.
2901 * In case of error restart replay process.
2903 static int ptlrpc_replay_interpret(const struct lu_env *env,
2904 struct ptlrpc_request *req,
2905 void * data, int rc)
2907 struct ptlrpc_replay_async_args *aa = data;
2908 struct obd_import *imp = req->rq_import;
2911 atomic_dec(&imp->imp_replay_inflight);
2913 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2914 * server got the request, but bulk transfer timeout, let's
2915 * replay the bulk req again */
2916 if (!ptlrpc_client_replied(req) ||
2917 (req->rq_bulk != NULL &&
2918 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2919 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2920 GOTO(out, rc = -ETIMEDOUT);
2923 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2924 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2925 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2926 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2928 /** VBR: check version failure */
2929 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2930 /** replay was failed due to version mismatch */
2931 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2932 spin_lock(&imp->imp_lock);
2933 imp->imp_vbr_failed = 1;
2934 imp->imp_no_lock_replay = 1;
2935 spin_unlock(&imp->imp_lock);
2936 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2938 /** The transno had better not change over replay. */
2939 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2940 lustre_msg_get_transno(req->rq_repmsg) ||
2941 lustre_msg_get_transno(req->rq_repmsg) == 0,
2943 lustre_msg_get_transno(req->rq_reqmsg),
2944 lustre_msg_get_transno(req->rq_repmsg));
2947 spin_lock(&imp->imp_lock);
2948 /** if replays by version then gap occur on server, no trust to locks */
2949 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2950 imp->imp_no_lock_replay = 1;
2951 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2952 spin_unlock(&imp->imp_lock);
2953 LASSERT(imp->imp_last_replay_transno);
2955 /* transaction number shouldn't be bigger than the latest replayed */
2956 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2957 DEBUG_REQ(D_ERROR, req,
2958 "Reported transno %llu is bigger than the "
2959 "replayed one: %llu", req->rq_transno,
2960 lustre_msg_get_transno(req->rq_reqmsg));
2961 GOTO(out, rc = -EINVAL);
2964 DEBUG_REQ(D_HA, req, "got rep");
2966 /* let the callback do fixups, possibly including in the request */
2967 if (req->rq_replay_cb)
2968 req->rq_replay_cb(req);
2970 if (ptlrpc_client_replied(req) &&
2971 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2972 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2973 lustre_msg_get_status(req->rq_repmsg),
2974 aa->praa_old_status);
2976 /* Note: If the replay fails for MDT-MDT recovery, let's
2977 * abort all of the following requests in the replay
2978 * and sending list, because MDT-MDT update requests
2979 * are dependent on each other, see LU-7039 */
2980 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
2981 struct ptlrpc_request *free_req;
2982 struct ptlrpc_request *tmp;
2984 spin_lock(&imp->imp_lock);
2985 list_for_each_entry_safe(free_req, tmp,
2986 &imp->imp_replay_list,
2988 ptlrpc_free_request(free_req);
2991 list_for_each_entry_safe(free_req, tmp,
2992 &imp->imp_committed_list,
2994 ptlrpc_free_request(free_req);
2997 list_for_each_entry_safe(free_req, tmp,
2998 &imp->imp_delayed_list,
3000 spin_lock(&free_req->rq_lock);
3001 free_req->rq_err = 1;
3002 free_req->rq_status = -EIO;
3003 ptlrpc_client_wake_req(free_req);
3004 spin_unlock(&free_req->rq_lock);
3007 list_for_each_entry_safe(free_req, tmp,
3008 &imp->imp_sending_list,
3010 spin_lock(&free_req->rq_lock);
3011 free_req->rq_err = 1;
3012 free_req->rq_status = -EIO;
3013 ptlrpc_client_wake_req(free_req);
3014 spin_unlock(&free_req->rq_lock);
3016 spin_unlock(&imp->imp_lock);
3019 /* Put it back for re-replay. */
3020 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3024 * Errors while replay can set transno to 0, but
3025 * imp_last_replay_transno shouldn't be set to 0 anyway
3027 if (req->rq_transno == 0)
3028 CERROR("Transno is 0 during replay!\n");
3030 /* continue with recovery */
3031 rc = ptlrpc_import_recovery_state_machine(imp);
3033 req->rq_send_state = aa->praa_old_state;
3036 /* this replay failed, so restart recovery */
3037 ptlrpc_connect_import(imp);
3043 * Prepares and queues request for replay.
3044 * Adds it to ptlrpcd queue for actual sending.
3045 * Returns 0 on success.
3047 int ptlrpc_replay_req(struct ptlrpc_request *req)
3049 struct ptlrpc_replay_async_args *aa;
3052 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3054 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
3055 aa = ptlrpc_req_async_args(req);
3056 memset(aa, 0, sizeof *aa);
3058 /* Prepare request to be resent with ptlrpcd */
3059 aa->praa_old_state = req->rq_send_state;
3060 req->rq_send_state = LUSTRE_IMP_REPLAY;
3061 req->rq_phase = RQ_PHASE_NEW;
3062 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3064 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3066 req->rq_interpret_reply = ptlrpc_replay_interpret;
3067 /* Readjust the timeout for current conditions */
3068 ptlrpc_at_set_req_timeout(req);
3070 /* Tell server the net_latency, so the server can calculate how long
3071 * it should wait for next replay */
3072 lustre_msg_set_service_time(req->rq_reqmsg,
3073 ptlrpc_at_get_net_latency(req));
3074 DEBUG_REQ(D_HA, req, "REPLAY");
3076 atomic_inc(&req->rq_import->imp_replay_inflight);
3077 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3079 ptlrpcd_add_req(req);
3084 * Aborts all in-flight request on import \a imp sending and delayed lists
3086 void ptlrpc_abort_inflight(struct obd_import *imp)
3088 struct list_head *tmp, *n;
3091 /* Make sure that no new requests get processed for this import.
3092 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3093 * this flag and then putting requests on sending_list or delayed_list.
3095 spin_lock(&imp->imp_lock);
3097 /* XXX locking? Maybe we should remove each request with the list
3098 * locked? Also, how do we know if the requests on the list are
3099 * being freed at this time?
3101 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3102 struct ptlrpc_request *req = list_entry(tmp,
3103 struct ptlrpc_request,
3106 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3108 spin_lock(&req->rq_lock);
3109 if (req->rq_import_generation < imp->imp_generation) {
3111 req->rq_status = -EIO;
3112 ptlrpc_client_wake_req(req);
3114 spin_unlock(&req->rq_lock);
3117 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3118 struct ptlrpc_request *req =
3119 list_entry(tmp, struct ptlrpc_request, rq_list);
3121 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3123 spin_lock(&req->rq_lock);
3124 if (req->rq_import_generation < imp->imp_generation) {
3126 req->rq_status = -EIO;
3127 ptlrpc_client_wake_req(req);
3129 spin_unlock(&req->rq_lock);
3132 /* Last chance to free reqs left on the replay list, but we
3133 * will still leak reqs that haven't committed. */
3134 if (imp->imp_replayable)
3135 ptlrpc_free_committed(imp);
3137 spin_unlock(&imp->imp_lock);
3143 * Abort all uncompleted requests in request set \a set
3145 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3147 struct list_head *tmp, *pos;
3149 LASSERT(set != NULL);
3151 list_for_each_safe(pos, tmp, &set->set_requests) {
3152 struct ptlrpc_request *req =
3153 list_entry(pos, struct ptlrpc_request,
3156 spin_lock(&req->rq_lock);
3157 if (req->rq_phase != RQ_PHASE_RPC) {
3158 spin_unlock(&req->rq_lock);
3163 req->rq_status = -EINTR;
3164 ptlrpc_client_wake_req(req);
3165 spin_unlock(&req->rq_lock);
3169 static __u64 ptlrpc_last_xid;
3170 static spinlock_t ptlrpc_last_xid_lock;
3173 * Initialize the XID for the node. This is common among all requests on
3174 * this node, and only requires the property that it is monotonically
3175 * increasing. It does not need to be sequential. Since this is also used
3176 * as the RDMA match bits, it is important that a single client NOT have
3177 * the same match bits for two different in-flight requests, hence we do
3178 * NOT want to have an XID per target or similar.
3180 * To avoid an unlikely collision between match bits after a client reboot
3181 * (which would deliver old data into the wrong RDMA buffer) initialize
3182 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3183 * If the time is clearly incorrect, we instead use a 62-bit random number.
3184 * In the worst case the random number will overflow 1M RPCs per second in
3185 * 9133 years, or permutations thereof.
3187 #define YEAR_2004 (1ULL << 30)
3188 void ptlrpc_init_xid(void)
3190 time_t now = cfs_time_current_sec();
3192 spin_lock_init(&ptlrpc_last_xid_lock);
3193 if (now < YEAR_2004) {
3194 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3195 ptlrpc_last_xid >>= 2;
3196 ptlrpc_last_xid |= (1ULL << 61);
3198 ptlrpc_last_xid = (__u64)now << 20;
3201 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3202 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3203 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3207 * Increase xid and returns resulting new value to the caller.
3209 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3210 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3211 * itself uses the last bulk xid needed, so the server can determine the
3212 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3213 * xid must align to a power-of-two value.
3215 * This is assumed to be true due to the initial ptlrpc_last_xid
3216 * value also being initialized to a power-of-two value. LU-1431
3218 __u64 ptlrpc_next_xid(void)
3222 spin_lock(&ptlrpc_last_xid_lock);
3223 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3224 ptlrpc_last_xid = next;
3225 spin_unlock(&ptlrpc_last_xid_lock);
3231 * If request has a new allocated XID (new request or EINPROGRESS resend),
3232 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3233 * request to ensure previous bulk fails and avoid problems with lost replies
3234 * and therefore several transfers landing into the same buffer from different
3237 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3239 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3241 LASSERT(bd != NULL);
3243 /* Generate new matchbits for all resend requests, including
3245 if (req->rq_resend) {
3246 __u64 old_mbits = req->rq_mbits;
3248 /* First time resend on -EINPROGRESS will generate new xid,
3249 * so we can actually use the rq_xid as rq_mbits in such case,
3250 * however, it's bit hard to distinguish such resend with a
3251 * 'resend for the -EINPROGRESS resend'. To make it simple,
3252 * we opt to generate mbits for all resend cases. */
3253 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
3254 req->rq_mbits = ptlrpc_next_xid();
3256 /* Old version transfers rq_xid to peer as
3258 spin_lock(&req->rq_import->imp_lock);
3259 list_del_init(&req->rq_unreplied_list);
3260 ptlrpc_assign_next_xid_nolock(req);
3261 spin_unlock(&req->rq_import->imp_lock);
3262 req->rq_mbits = req->rq_xid;
3264 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3265 old_mbits, req->rq_mbits);
3266 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3267 /* Request being sent first time, use xid as matchbits. */
3268 req->rq_mbits = req->rq_xid;
3270 /* Replay request, xid and matchbits have already been
3271 * correctly assigned. */
3275 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3276 * that server can infer the number of bulks that were prepared,
3278 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3281 /* Set rq_xid as rq_mbits to indicate the final bulk for the old
3282 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3284 * It's ok to directly set the rq_xid here, since this xid bump
3285 * won't affect the request position in unreplied list. */
3286 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3287 req->rq_xid = req->rq_mbits;
3291 * Get a glimpse at what next xid value might have been.
3292 * Returns possible next xid.
3294 __u64 ptlrpc_sample_next_xid(void)
3296 #if BITS_PER_LONG == 32
3297 /* need to avoid possible word tearing on 32-bit systems */
3300 spin_lock(&ptlrpc_last_xid_lock);
3301 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3302 spin_unlock(&ptlrpc_last_xid_lock);
3306 /* No need to lock, since returned value is racy anyways */
3307 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3310 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3313 * Functions for operating ptlrpc workers.
3315 * A ptlrpc work is a function which will be running inside ptlrpc context.
3316 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3318 * 1. after a work is created, it can be used many times, that is:
3319 * handler = ptlrpcd_alloc_work();
3320 * ptlrpcd_queue_work();
3322 * queue it again when necessary:
3323 * ptlrpcd_queue_work();
3324 * ptlrpcd_destroy_work();
3325 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3326 * it will only be queued once in any time. Also as its name implies, it may
3327 * have delay before it really runs by ptlrpcd thread.
3329 struct ptlrpc_work_async_args {
3330 int (*cb)(const struct lu_env *, void *);
3334 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3336 /* re-initialize the req */
3337 req->rq_timeout = obd_timeout;
3338 req->rq_sent = cfs_time_current_sec();
3339 req->rq_deadline = req->rq_sent + req->rq_timeout;
3340 req->rq_phase = RQ_PHASE_INTERPRET;
3341 req->rq_next_phase = RQ_PHASE_COMPLETE;
3342 req->rq_xid = ptlrpc_next_xid();
3343 req->rq_import_generation = req->rq_import->imp_generation;
3345 ptlrpcd_add_req(req);
3348 static int work_interpreter(const struct lu_env *env,
3349 struct ptlrpc_request *req, void *data, int rc)
3351 struct ptlrpc_work_async_args *arg = data;
3353 LASSERT(ptlrpcd_check_work(req));
3354 LASSERT(arg->cb != NULL);
3356 rc = arg->cb(env, arg->cbdata);
3358 list_del_init(&req->rq_set_chain);
3361 if (atomic_dec_return(&req->rq_refcount) > 1) {
3362 atomic_set(&req->rq_refcount, 2);
3363 ptlrpcd_add_work_req(req);
3368 static int worker_format;
3370 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3372 return req->rq_pill.rc_fmt == (void *)&worker_format;
3376 * Create a work for ptlrpc.
3378 void *ptlrpcd_alloc_work(struct obd_import *imp,
3379 int (*cb)(const struct lu_env *, void *), void *cbdata)
3381 struct ptlrpc_request *req = NULL;
3382 struct ptlrpc_work_async_args *args;
3388 RETURN(ERR_PTR(-EINVAL));
3390 /* copy some code from deprecated fakereq. */
3391 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3393 CERROR("ptlrpc: run out of memory!\n");
3394 RETURN(ERR_PTR(-ENOMEM));
3397 ptlrpc_cli_req_init(req);
3399 req->rq_send_state = LUSTRE_IMP_FULL;
3400 req->rq_type = PTL_RPC_MSG_REQUEST;
3401 req->rq_import = class_import_get(imp);
3402 req->rq_interpret_reply = work_interpreter;
3403 /* don't want reply */
3404 req->rq_no_delay = req->rq_no_resend = 1;
3405 req->rq_pill.rc_fmt = (void *)&worker_format;
3407 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3408 args = ptlrpc_req_async_args(req);
3410 args->cbdata = cbdata;
3414 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3416 void ptlrpcd_destroy_work(void *handler)
3418 struct ptlrpc_request *req = handler;
3421 ptlrpc_req_finished(req);
3423 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3425 int ptlrpcd_queue_work(void *handler)
3427 struct ptlrpc_request *req = handler;
3430 * Check if the req is already being queued.
3432 * Here comes a trick: it lacks a way of checking if a req is being
3433 * processed reliably in ptlrpc. Here I have to use refcount of req
3434 * for this purpose. This is okay because the caller should use this
3435 * req as opaque data. - Jinshan
3437 LASSERT(atomic_read(&req->rq_refcount) > 0);
3438 if (atomic_inc_return(&req->rq_refcount) == 2)
3439 ptlrpcd_add_work_req(req);
3442 EXPORT_SYMBOL(ptlrpcd_queue_work);