4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 /** Implementation of client-side PortalRPC interfaces */
35 #define DEBUG_SUBSYSTEM S_RPC
37 #include <linux/delay.h>
38 #include <obd_support.h>
39 #include <obd_class.h>
40 #include <lustre_lib.h>
41 #include <lustre_ha.h>
42 #include <lustre_import.h>
43 #include <lustre_req_layout.h>
45 #include "ptlrpc_internal.h"
47 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
48 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
49 .release_frags = ptlrpc_release_bulk_page_pin,
51 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
53 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
54 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
55 .release_frags = ptlrpc_release_bulk_noop,
57 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
59 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
60 .add_iov_frag = ptlrpc_prep_bulk_frag,
62 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
64 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
65 static int ptlrpcd_check_work(struct ptlrpc_request *req);
66 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
69 * Initialize passed in client structure \a cl.
71 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
72 struct ptlrpc_client *cl)
74 cl->cli_request_portal = req_portal;
75 cl->cli_reply_portal = rep_portal;
78 EXPORT_SYMBOL(ptlrpc_init_client);
81 * Return PortalRPC connection for remore uud \a uuid
83 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
84 lnet_nid_t nid4refnet)
86 struct ptlrpc_connection *c;
88 struct lnet_process_id peer;
91 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
92 * before accessing its values. */
93 /* coverity[uninit_use_in_call] */
94 peer.nid = nid4refnet;
95 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
97 CNETERR("cannot find peer %s!\n", uuid->uuid);
101 c = ptlrpc_connection_get(peer, self, uuid);
103 memcpy(c->c_remote_uuid.uuid,
104 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
107 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
113 * Allocate and initialize new bulk descriptor on the sender.
114 * Returns pointer to the descriptor or NULL on error.
116 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
117 enum ptlrpc_bulk_op_type type,
119 const struct ptlrpc_bulk_frag_ops *ops)
121 struct ptlrpc_bulk_desc *desc;
124 /* ensure that only one of KIOV or IOVEC is set but not both */
125 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
126 ops->add_kiov_frag != NULL) ||
127 (ptlrpc_is_bulk_desc_kvec(type) &&
128 ops->add_iov_frag != NULL));
133 if (type & PTLRPC_BULK_BUF_KIOV) {
134 OBD_ALLOC_LARGE(GET_KIOV(desc),
135 nfrags * sizeof(*GET_KIOV(desc)));
136 if (GET_KIOV(desc) == NULL)
139 OBD_ALLOC_LARGE(GET_KVEC(desc),
140 nfrags * sizeof(*GET_KVEC(desc)));
141 if (GET_KVEC(desc) == NULL)
145 spin_lock_init(&desc->bd_lock);
146 init_waitqueue_head(&desc->bd_waitq);
147 desc->bd_max_iov = nfrags;
148 desc->bd_iov_count = 0;
149 desc->bd_portal = portal;
150 desc->bd_type = type;
151 desc->bd_md_count = 0;
152 desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
153 LASSERT(max_brw > 0);
154 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
155 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
156 * node. Negotiated ocd_brw_size will always be <= this number. */
157 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
158 LNetInvalidateMDHandle(&desc->bd_mds[i]);
167 * Prepare bulk descriptor for specified outgoing request \a req that
168 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
169 * the bulk to be sent. Used on client-side.
170 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
173 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
174 unsigned nfrags, unsigned max_brw,
177 const struct ptlrpc_bulk_frag_ops
180 struct obd_import *imp = req->rq_import;
181 struct ptlrpc_bulk_desc *desc;
184 LASSERT(ptlrpc_is_bulk_op_passive(type));
186 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
190 desc->bd_import_generation = req->rq_import_generation;
191 desc->bd_import = class_import_get(imp);
194 desc->bd_cbid.cbid_fn = client_bulk_callback;
195 desc->bd_cbid.cbid_arg = desc;
197 /* This makes req own desc, and free it when she frees herself */
202 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
204 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
205 struct page *page, int pageoffset, int len,
210 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
211 LASSERT(page != NULL);
212 LASSERT(pageoffset >= 0);
214 LASSERT(pageoffset + len <= PAGE_SIZE);
215 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
217 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
224 kiov->kiov_page = page;
225 kiov->kiov_offset = pageoffset;
226 kiov->kiov_len = len;
228 desc->bd_iov_count++;
230 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
232 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
238 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
239 LASSERT(frag != NULL);
241 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
243 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
247 iovec->iov_base = frag;
248 iovec->iov_len = len;
250 desc->bd_iov_count++;
252 RETURN(desc->bd_nob);
254 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
256 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
260 LASSERT(desc != NULL);
261 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
262 LASSERT(desc->bd_md_count == 0); /* network hands off */
263 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
264 LASSERT(desc->bd_frag_ops != NULL);
266 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
267 sptlrpc_enc_pool_put_pages(desc);
270 class_export_put(desc->bd_export);
272 class_import_put(desc->bd_import);
274 if (desc->bd_frag_ops->release_frags != NULL)
275 desc->bd_frag_ops->release_frags(desc);
277 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
278 OBD_FREE_LARGE(GET_KIOV(desc),
279 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
281 OBD_FREE_LARGE(GET_KVEC(desc),
282 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
286 EXPORT_SYMBOL(ptlrpc_free_bulk);
289 * Set server timelimit for this req, i.e. how long are we willing to wait
290 * for reply before timing out this request.
292 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
298 LASSERT(req->rq_import);
301 /* non-AT settings */
303 * \a imp_server_timeout means this is reverse import and
304 * we send (currently only) ASTs to the client and cannot afford
305 * to wait too long for the reply, otherwise the other client
306 * (because of which we are sending this request) would
307 * timeout waiting for us
309 req->rq_timeout = req->rq_import->imp_server_timeout ?
310 obd_timeout / 2 : obd_timeout;
312 at = &req->rq_import->imp_at;
313 idx = import_at_get_index(req->rq_import,
314 req->rq_request_portal);
315 serv_est = at_get(&at->iat_service_estimate[idx]);
316 req->rq_timeout = at_est2timeout(serv_est);
318 /* We could get even fancier here, using history to predict increased
321 /* Let the server know what this RPC timeout is by putting it in the
323 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
325 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
327 /* Adjust max service estimate based on server value */
328 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
329 unsigned int serv_est)
335 LASSERT(req->rq_import);
336 at = &req->rq_import->imp_at;
338 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
339 /* max service estimates are tracked on the server side,
340 so just keep minimal history here */
341 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
343 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
344 "has changed from %d to %d\n",
345 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
346 oldse, at_get(&at->iat_service_estimate[idx]));
349 /* Expected network latency per remote node (secs) */
350 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
352 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
355 /* Adjust expected network latency */
356 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
357 unsigned int service_time)
359 unsigned int nl, oldnl;
361 time64_t now = ktime_get_real_seconds();
363 LASSERT(req->rq_import);
365 if (service_time > now - req->rq_sent + 3) {
366 /* bz16408, however, this can also happen if early reply
367 * is lost and client RPC is expired and resent, early reply
368 * or reply of original RPC can still be fit in reply buffer
369 * of resent RPC, now client is measuring time from the
370 * resent time, but server sent back service time of original
373 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
374 D_ADAPTTO : D_WARNING,
375 "Reported service time %u > total measured time %lld\n",
376 service_time, now - req->rq_sent);
380 /* Network latency is total time less server processing time */
381 nl = max_t(int, now - req->rq_sent -
382 service_time, 0) + 1; /* st rounding */
383 at = &req->rq_import->imp_at;
385 oldnl = at_measured(&at->iat_net_latency, nl);
387 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
388 "has changed from %d to %d\n",
389 req->rq_import->imp_obd->obd_name,
391 &req->rq_import->imp_connection->c_remote_uuid),
392 oldnl, at_get(&at->iat_net_latency));
395 static int unpack_reply(struct ptlrpc_request *req)
399 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
400 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
402 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
407 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
409 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
416 * Handle an early reply message, called with the rq_lock held.
417 * If anything goes wrong just ignore it - same as if it never happened
419 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
420 __must_hold(&req->rq_lock)
422 struct ptlrpc_request *early_req;
428 spin_unlock(&req->rq_lock);
430 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
432 spin_lock(&req->rq_lock);
436 rc = unpack_reply(early_req);
438 sptlrpc_cli_finish_early_reply(early_req);
439 spin_lock(&req->rq_lock);
443 /* Use new timeout value just to adjust the local value for this
444 * request, don't include it into at_history. It is unclear yet why
445 * service time increased and should it be counted or skipped, e.g.
446 * that can be recovery case or some error or server, the real reply
447 * will add all new data if it is worth to add. */
448 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
449 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
451 /* Network latency can be adjusted, it is pure network delays */
452 ptlrpc_at_adj_net_latency(req,
453 lustre_msg_get_service_time(early_req->rq_repmsg));
455 sptlrpc_cli_finish_early_reply(early_req);
457 spin_lock(&req->rq_lock);
458 olddl = req->rq_deadline;
459 /* server assumes it now has rq_timeout from when the request
460 * arrived, so the client should give it at least that long.
461 * since we don't know the arrival time we'll use the original
463 req->rq_deadline = req->rq_sent + req->rq_timeout +
464 ptlrpc_at_get_net_latency(req);
466 DEBUG_REQ(D_ADAPTTO, req,
467 "Early reply #%d, new deadline in %llds (%llds)",
469 req->rq_deadline - ktime_get_real_seconds(),
470 req->rq_deadline - olddl);
475 static struct kmem_cache *request_cache;
477 int ptlrpc_request_cache_init(void)
479 request_cache = kmem_cache_create("ptlrpc_cache",
480 sizeof(struct ptlrpc_request),
481 0, SLAB_HWCACHE_ALIGN, NULL);
482 return request_cache == NULL ? -ENOMEM : 0;
485 void ptlrpc_request_cache_fini(void)
487 kmem_cache_destroy(request_cache);
490 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
492 struct ptlrpc_request *req;
494 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
498 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
500 OBD_SLAB_FREE_PTR(req, request_cache);
504 * Wind down request pool \a pool.
505 * Frees all requests from the pool too
507 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
509 struct list_head *l, *tmp;
510 struct ptlrpc_request *req;
512 LASSERT(pool != NULL);
514 spin_lock(&pool->prp_lock);
515 list_for_each_safe(l, tmp, &pool->prp_req_list) {
516 req = list_entry(l, struct ptlrpc_request, rq_list);
517 list_del(&req->rq_list);
518 LASSERT(req->rq_reqbuf);
519 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
520 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
521 ptlrpc_request_cache_free(req);
523 spin_unlock(&pool->prp_lock);
524 OBD_FREE(pool, sizeof(*pool));
526 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
529 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
531 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
536 while (size < pool->prp_rq_size)
539 LASSERTF(list_empty(&pool->prp_req_list) ||
540 size == pool->prp_rq_size,
541 "Trying to change pool size with nonempty pool "
542 "from %d to %d bytes\n", pool->prp_rq_size, size);
544 spin_lock(&pool->prp_lock);
545 pool->prp_rq_size = size;
546 for (i = 0; i < num_rq; i++) {
547 struct ptlrpc_request *req;
548 struct lustre_msg *msg;
550 spin_unlock(&pool->prp_lock);
551 req = ptlrpc_request_cache_alloc(GFP_NOFS);
554 OBD_ALLOC_LARGE(msg, size);
556 ptlrpc_request_cache_free(req);
559 req->rq_reqbuf = msg;
560 req->rq_reqbuf_len = size;
562 spin_lock(&pool->prp_lock);
563 list_add_tail(&req->rq_list, &pool->prp_req_list);
565 spin_unlock(&pool->prp_lock);
568 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
571 * Create and initialize new request pool with given attributes:
572 * \a num_rq - initial number of requests to create for the pool
573 * \a msgsize - maximum message size possible for requests in thid pool
574 * \a populate_pool - function to be called when more requests need to be added
576 * Returns pointer to newly created pool or NULL on error.
578 struct ptlrpc_request_pool *
579 ptlrpc_init_rq_pool(int num_rq, int msgsize,
580 int (*populate_pool)(struct ptlrpc_request_pool *, int))
582 struct ptlrpc_request_pool *pool;
584 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
588 /* Request next power of two for the allocation, because internally
589 kernel would do exactly this */
591 spin_lock_init(&pool->prp_lock);
592 INIT_LIST_HEAD(&pool->prp_req_list);
593 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
594 pool->prp_populate = populate_pool;
596 populate_pool(pool, num_rq);
600 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
603 * Fetches one request from pool \a pool
605 static struct ptlrpc_request *
606 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
608 struct ptlrpc_request *request;
609 struct lustre_msg *reqbuf;
614 spin_lock(&pool->prp_lock);
616 /* See if we have anything in a pool, and bail out if nothing,
617 * in writeout path, where this matters, this is safe to do, because
618 * nothing is lost in this case, and when some in-flight requests
619 * complete, this code will be called again. */
620 if (unlikely(list_empty(&pool->prp_req_list))) {
621 spin_unlock(&pool->prp_lock);
625 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
627 list_del_init(&request->rq_list);
628 spin_unlock(&pool->prp_lock);
630 LASSERT(request->rq_reqbuf);
631 LASSERT(request->rq_pool);
633 reqbuf = request->rq_reqbuf;
634 memset(request, 0, sizeof(*request));
635 request->rq_reqbuf = reqbuf;
636 request->rq_reqbuf_len = pool->prp_rq_size;
637 request->rq_pool = pool;
643 * Returns freed \a request to pool.
645 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
647 struct ptlrpc_request_pool *pool = request->rq_pool;
649 spin_lock(&pool->prp_lock);
650 LASSERT(list_empty(&request->rq_list));
651 LASSERT(!request->rq_receiving_reply);
652 list_add_tail(&request->rq_list, &pool->prp_req_list);
653 spin_unlock(&pool->prp_lock);
656 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
658 struct obd_import *imp = req->rq_import;
659 struct list_head *tmp;
660 struct ptlrpc_request *iter;
662 assert_spin_locked(&imp->imp_lock);
663 LASSERT(list_empty(&req->rq_unreplied_list));
665 /* unreplied list is sorted by xid in ascending order */
666 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
667 iter = list_entry(tmp, struct ptlrpc_request,
670 LASSERT(req->rq_xid != iter->rq_xid);
671 if (req->rq_xid < iter->rq_xid)
673 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
676 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
679 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
681 req->rq_xid = ptlrpc_next_xid();
682 ptlrpc_add_unreplied(req);
685 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
687 spin_lock(&req->rq_import->imp_lock);
688 ptlrpc_assign_next_xid_nolock(req);
689 spin_unlock(&req->rq_import->imp_lock);
692 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
693 __u32 version, int opcode, char **bufs,
694 struct ptlrpc_cli_ctx *ctx)
697 struct obd_import *imp;
703 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
704 imp = request->rq_import;
705 lengths = request->rq_pill.rc_area[RCL_CLIENT];
708 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
710 rc = sptlrpc_req_get_ctx(request);
714 sptlrpc_req_set_flavor(request, opcode);
716 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
721 lustre_msg_add_version(request->rq_reqmsg, version);
722 request->rq_send_state = LUSTRE_IMP_FULL;
723 request->rq_type = PTL_RPC_MSG_REQUEST;
725 request->rq_req_cbid.cbid_fn = request_out_callback;
726 request->rq_req_cbid.cbid_arg = request;
728 request->rq_reply_cbid.cbid_fn = reply_in_callback;
729 request->rq_reply_cbid.cbid_arg = request;
731 request->rq_reply_deadline = 0;
732 request->rq_bulk_deadline = 0;
733 request->rq_req_deadline = 0;
734 request->rq_phase = RQ_PHASE_NEW;
735 request->rq_next_phase = RQ_PHASE_UNDEFINED;
737 request->rq_request_portal = imp->imp_client->cli_request_portal;
738 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
740 ptlrpc_at_set_req_timeout(request);
742 lustre_msg_set_opc(request->rq_reqmsg, opcode);
743 ptlrpc_assign_next_xid(request);
745 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
746 if (cfs_fail_val == opcode) {
747 time64_t *fail_t = NULL, *fail2_t = NULL;
749 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
750 fail_t = &request->rq_bulk_deadline;
751 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
752 fail_t = &request->rq_reply_deadline;
753 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
754 fail_t = &request->rq_req_deadline;
755 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
756 fail_t = &request->rq_reply_deadline;
757 fail2_t = &request->rq_bulk_deadline;
761 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
764 *fail2_t = ktime_get_real_seconds() +
768 * The RPC is infected, let the test to change the
771 msleep(4 * MSEC_PER_SEC);
778 LASSERT(!request->rq_pool);
779 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
781 class_import_put(imp);
786 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
789 * Pack request buffers for network transfer, performing necessary encryption
790 * steps if necessary.
792 int ptlrpc_request_pack(struct ptlrpc_request *request,
793 __u32 version, int opcode)
796 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
800 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
801 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
802 * have to send old ptlrpc_body to keep interoprability with these
805 * Only three kinds of server->client RPCs so far:
810 * XXX This should be removed whenever we drop the interoprability with
811 * the these old clients.
813 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
814 opcode == LDLM_GL_CALLBACK)
815 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
816 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
820 EXPORT_SYMBOL(ptlrpc_request_pack);
823 * Helper function to allocate new request on import \a imp
824 * and possibly using existing request from pool \a pool if provided.
825 * Returns allocated request structure with import field filled or
829 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
830 struct ptlrpc_request_pool *pool)
832 struct ptlrpc_request *request = NULL;
834 request = ptlrpc_request_cache_alloc(GFP_NOFS);
836 if (!request && pool)
837 request = ptlrpc_prep_req_from_pool(pool);
840 ptlrpc_cli_req_init(request);
842 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
843 LASSERT(imp != LP_POISON);
844 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
846 LASSERT(imp->imp_client != LP_POISON);
848 request->rq_import = class_import_get(imp);
850 CERROR("request allocation out of memory\n");
857 * Helper function for creating a request.
858 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
859 * buffer structures according to capsule template \a format.
860 * Returns allocated request structure pointer or NULL on error.
862 static struct ptlrpc_request *
863 ptlrpc_request_alloc_internal(struct obd_import *imp,
864 struct ptlrpc_request_pool * pool,
865 const struct req_format *format)
867 struct ptlrpc_request *request;
869 request = __ptlrpc_request_alloc(imp, pool);
873 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
874 req_capsule_set(&request->rq_pill, format);
879 * Allocate new request structure for import \a imp and initialize its
880 * buffer structure according to capsule template \a format.
882 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
883 const struct req_format *format)
885 return ptlrpc_request_alloc_internal(imp, NULL, format);
887 EXPORT_SYMBOL(ptlrpc_request_alloc);
890 * Allocate new request structure for import \a imp from pool \a pool and
891 * initialize its buffer structure according to capsule template \a format.
893 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
894 struct ptlrpc_request_pool * pool,
895 const struct req_format *format)
897 return ptlrpc_request_alloc_internal(imp, pool, format);
899 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
902 * For requests not from pool, free memory of the request structure.
903 * For requests obtained from a pool earlier, return request back to pool.
905 void ptlrpc_request_free(struct ptlrpc_request *request)
907 if (request->rq_pool)
908 __ptlrpc_free_req_to_pool(request);
910 ptlrpc_request_cache_free(request);
912 EXPORT_SYMBOL(ptlrpc_request_free);
915 * Allocate new request for operatione \a opcode and immediatelly pack it for
917 * Only used for simple requests like OBD_PING where the only important
918 * part of the request is operation itself.
919 * Returns allocated request or NULL on error.
921 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
922 const struct req_format *format,
923 __u32 version, int opcode)
925 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
929 rc = ptlrpc_request_pack(req, version, opcode);
931 ptlrpc_request_free(req);
937 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
940 * Allocate and initialize new request set structure on the current CPT.
941 * Returns a pointer to the newly allocated set structure or NULL on error.
943 struct ptlrpc_request_set *ptlrpc_prep_set(void)
945 struct ptlrpc_request_set *set;
949 cpt = cfs_cpt_current(cfs_cpt_table, 0);
950 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
953 atomic_set(&set->set_refcount, 1);
954 INIT_LIST_HEAD(&set->set_requests);
955 init_waitqueue_head(&set->set_waitq);
956 atomic_set(&set->set_new_count, 0);
957 atomic_set(&set->set_remaining, 0);
958 spin_lock_init(&set->set_new_req_lock);
959 INIT_LIST_HEAD(&set->set_new_requests);
960 set->set_max_inflight = UINT_MAX;
961 set->set_producer = NULL;
962 set->set_producer_arg = NULL;
967 EXPORT_SYMBOL(ptlrpc_prep_set);
970 * Allocate and initialize new request set structure with flow control
971 * extension. This extension allows to control the number of requests in-flight
972 * for the whole set. A callback function to generate requests must be provided
973 * and the request set will keep the number of requests sent over the wire to
975 * Returns a pointer to the newly allocated set structure or NULL on error.
977 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
981 struct ptlrpc_request_set *set;
983 set = ptlrpc_prep_set();
987 set->set_max_inflight = max;
988 set->set_producer = func;
989 set->set_producer_arg = arg;
995 * Wind down and free request set structure previously allocated with
997 * Ensures that all requests on the set have completed and removes
998 * all requests from the request list in a set.
999 * If any unsent request happen to be on the list, pretends that they got
1000 * an error in flight and calls their completion handler.
1002 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1004 struct list_head *tmp;
1005 struct list_head *next;
1010 /* Requests on the set should either all be completed, or all be new */
1011 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1012 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1013 list_for_each(tmp, &set->set_requests) {
1014 struct ptlrpc_request *req =
1015 list_entry(tmp, struct ptlrpc_request,
1018 LASSERT(req->rq_phase == expected_phase);
1022 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1023 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1024 atomic_read(&set->set_remaining), n);
1026 list_for_each_safe(tmp, next, &set->set_requests) {
1027 struct ptlrpc_request *req =
1028 list_entry(tmp, struct ptlrpc_request,
1030 list_del_init(&req->rq_set_chain);
1032 LASSERT(req->rq_phase == expected_phase);
1034 if (req->rq_phase == RQ_PHASE_NEW) {
1035 ptlrpc_req_interpret(NULL, req, -EBADR);
1036 atomic_dec(&set->set_remaining);
1039 spin_lock(&req->rq_lock);
1041 req->rq_invalid_rqset = 0;
1042 spin_unlock(&req->rq_lock);
1044 ptlrpc_req_finished (req);
1047 LASSERT(atomic_read(&set->set_remaining) == 0);
1049 ptlrpc_reqset_put(set);
1052 EXPORT_SYMBOL(ptlrpc_set_destroy);
1055 * Add a new request to the general purpose request set.
1056 * Assumes request reference from the caller.
1058 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1059 struct ptlrpc_request *req)
1061 LASSERT(list_empty(&req->rq_set_chain));
1063 if (req->rq_allow_intr)
1064 set->set_allow_intr = 1;
1066 /* The set takes over the caller's request reference */
1067 list_add_tail(&req->rq_set_chain, &set->set_requests);
1069 atomic_inc(&set->set_remaining);
1070 req->rq_queued_time = ktime_get_seconds();
1072 if (req->rq_reqmsg != NULL)
1073 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1075 if (set->set_producer != NULL)
1076 /* If the request set has a producer callback, the RPC must be
1077 * sent straight away */
1078 ptlrpc_send_new_req(req);
1080 EXPORT_SYMBOL(ptlrpc_set_add_req);
1083 * Add a request to a request with dedicated server thread
1084 * and wake the thread to make any necessary processing.
1085 * Currently only used for ptlrpcd.
1087 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1088 struct ptlrpc_request *req)
1090 struct ptlrpc_request_set *set = pc->pc_set;
1093 LASSERT(req->rq_set == NULL);
1094 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1096 spin_lock(&set->set_new_req_lock);
1098 * The set takes over the caller's request reference.
1101 req->rq_queued_time = ktime_get_seconds();
1102 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1103 count = atomic_inc_return(&set->set_new_count);
1104 spin_unlock(&set->set_new_req_lock);
1106 /* Only need to call wakeup once for the first entry. */
1108 wake_up(&set->set_waitq);
1110 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1111 * guarantee the async RPC can be processed ASAP, we have
1112 * no other better choice. It maybe fixed in future. */
1113 for (i = 0; i < pc->pc_npartners; i++)
1114 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1119 * Based on the current state of the import, determine if the request
1120 * can be sent, is an error, or should be delayed.
1122 * Returns true if this request should be delayed. If false, and
1123 * *status is set, then the request can not be sent and *status is the
1124 * error code. If false and status is 0, then request can be sent.
1126 * The imp->imp_lock must be held.
1128 static int ptlrpc_import_delay_req(struct obd_import *imp,
1129 struct ptlrpc_request *req, int *status)
1134 LASSERT (status != NULL);
1137 if (req->rq_ctx_init || req->rq_ctx_fini) {
1138 /* always allow ctx init/fini rpc go through */
1139 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1140 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1142 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1143 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1145 /* pings or MDS-equivalent STATFS may safely race with umount */
1146 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1147 D_HA : D_ERROR, req, "IMP_CLOSED ");
1149 } else if (ptlrpc_send_limit_expired(req)) {
1150 /* probably doesn't need to be a D_ERROR after initial testing*/
1151 DEBUG_REQ(D_HA, req, "send limit expired ");
1152 *status = -ETIMEDOUT;
1153 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1154 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1155 /* allow CONNECT even if import is invalid */ ;
1156 if (atomic_read(&imp->imp_inval_count) != 0) {
1157 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1160 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1161 if (!imp->imp_deactive)
1162 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1163 *status = -ESHUTDOWN; /* bz 12940 */
1164 } else if (req->rq_import_generation != imp->imp_generation) {
1165 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1167 } else if (req->rq_send_state != imp->imp_state) {
1168 /* invalidate in progress - any requests should be drop */
1169 if (atomic_read(&imp->imp_inval_count) != 0) {
1170 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1172 } else if (req->rq_no_delay) {
1173 *status = -EWOULDBLOCK;
1174 } else if (req->rq_allow_replay &&
1175 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1176 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1177 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1178 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1179 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1189 * Decide if the error message should be printed to the console or not.
1190 * Makes its decision based on request type, status, and failure frequency.
1192 * \param[in] req request that failed and may need a console message
1194 * \retval false if no message should be printed
1195 * \retval true if console message should be printed
1197 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1199 LASSERT(req->rq_reqmsg != NULL);
1201 /* Suppress particular reconnect errors which are to be expected. */
1202 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1204 /* Suppress timed out reconnect requests */
1205 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1209 /* Suppress most unavailable/again reconnect requests, but
1210 * print occasionally so it is clear client is trying to
1211 * connect to a server where no target is running. */
1212 if ((err == -ENODEV || err == -EAGAIN) &&
1213 req->rq_import->imp_conn_cnt % 30 != 20)
1217 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1218 /* -EAGAIN is normal when using POSIX flocks */
1221 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1222 (req->rq_xid & 0xf) != 10)
1223 /* Suppress most ping requests, they may fail occasionally */
1230 * Check request processing status.
1231 * Returns the status.
1233 static int ptlrpc_check_status(struct ptlrpc_request *req)
1238 err = lustre_msg_get_status(req->rq_repmsg);
1239 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1240 struct obd_import *imp = req->rq_import;
1241 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1242 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1244 if (ptlrpc_console_allow(req, opc, err))
1245 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1246 "failed: rc = %d\n",
1247 imp->imp_obd->obd_name,
1249 libcfs_nid2str(nid), err);
1250 RETURN(err < 0 ? err : -EINVAL);
1254 DEBUG_REQ(D_INFO, req, "status is %d", err);
1255 } else if (err > 0) {
1256 /* XXX: translate this error from net to host */
1257 DEBUG_REQ(D_INFO, req, "status is %d", err);
1264 * save pre-versions of objects into request for replay.
1265 * Versions are obtained from server reply.
1268 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1270 struct lustre_msg *repmsg = req->rq_repmsg;
1271 struct lustre_msg *reqmsg = req->rq_reqmsg;
1272 __u64 *versions = lustre_msg_get_versions(repmsg);
1275 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1279 lustre_msg_set_versions(reqmsg, versions);
1280 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1281 versions[0], versions[1]);
1286 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1288 struct ptlrpc_request *req;
1290 assert_spin_locked(&imp->imp_lock);
1291 if (list_empty(&imp->imp_unreplied_list))
1294 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1296 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1298 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1299 imp->imp_known_replied_xid = req->rq_xid - 1;
1301 return req->rq_xid - 1;
1305 * Callback function called when client receives RPC reply for \a req.
1306 * Returns 0 on success or error code.
1307 * The return alue would be assigned to req->rq_status by the caller
1308 * as request processing status.
1309 * This function also decides if the request needs to be saved for later replay.
1311 static int after_reply(struct ptlrpc_request *req)
1313 struct obd_import *imp = req->rq_import;
1314 struct obd_device *obd = req->rq_import->imp_obd;
1321 LASSERT(obd != NULL);
1322 /* repbuf must be unlinked */
1323 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1325 if (req->rq_reply_truncated) {
1326 if (ptlrpc_no_resend(req)) {
1327 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1328 " expected: %d, actual size: %d",
1329 req->rq_nob_received, req->rq_repbuf_len);
1333 sptlrpc_cli_free_repbuf(req);
1334 /* Pass the required reply buffer size (include
1335 * space for early reply).
1336 * NB: no need to roundup because alloc_repbuf
1337 * will roundup it */
1338 req->rq_replen = req->rq_nob_received;
1339 req->rq_nob_received = 0;
1340 spin_lock(&req->rq_lock);
1342 spin_unlock(&req->rq_lock);
1346 work_start = ktime_get_real();
1347 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1350 * NB Until this point, the whole of the incoming message,
1351 * including buflens, status etc is in the sender's byte order.
1353 rc = sptlrpc_cli_unwrap_reply(req);
1355 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1360 * Security layer unwrap might ask resend this request.
1365 rc = unpack_reply(req);
1369 /* retry indefinitely on EINPROGRESS */
1370 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1371 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1372 time64_t now = ktime_get_real_seconds();
1374 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1375 spin_lock(&req->rq_lock);
1377 spin_unlock(&req->rq_lock);
1378 req->rq_nr_resend++;
1380 /* Readjust the timeout for current conditions */
1381 ptlrpc_at_set_req_timeout(req);
1382 /* delay resend to give a chance to the server to get ready.
1383 * The delay is increased by 1s on every resend and is capped to
1384 * the current request timeout (i.e. obd_timeout if AT is off,
1385 * or AT service time x 125% + 5s, see at_est2timeout) */
1386 if (req->rq_nr_resend > req->rq_timeout)
1387 req->rq_sent = now + req->rq_timeout;
1389 req->rq_sent = now + req->rq_nr_resend;
1391 /* Resend for EINPROGRESS will use a new XID */
1392 spin_lock(&imp->imp_lock);
1393 list_del_init(&req->rq_unreplied_list);
1394 spin_unlock(&imp->imp_lock);
1399 if (obd->obd_svc_stats != NULL) {
1400 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1402 ptlrpc_lprocfs_rpc_sent(req, timediff);
1405 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1406 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1407 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1408 lustre_msg_get_type(req->rq_repmsg));
1412 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1413 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1414 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1415 ptlrpc_at_adj_net_latency(req,
1416 lustre_msg_get_service_time(req->rq_repmsg));
1418 rc = ptlrpc_check_status(req);
1419 imp->imp_connect_error = rc;
1423 * Either we've been evicted, or the server has failed for
1424 * some reason. Try to reconnect, and if that fails, punt to
1427 if (ptlrpc_recoverable_error(rc)) {
1428 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1429 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1432 ptlrpc_request_handle_notconn(req);
1437 * Let's look if server sent slv. Do it only for RPC with
1440 ldlm_cli_update_pool(req);
1444 * Store transno in reqmsg for replay.
1446 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1447 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1448 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1451 if (imp->imp_replayable) {
1452 spin_lock(&imp->imp_lock);
1454 * No point in adding already-committed requests to the replay
1455 * list, we will just remove them immediately. b=9829
1457 if (req->rq_transno != 0 &&
1459 lustre_msg_get_last_committed(req->rq_repmsg) ||
1461 /** version recovery */
1462 ptlrpc_save_versions(req);
1463 ptlrpc_retain_replayable_request(req, imp);
1464 } else if (req->rq_commit_cb != NULL &&
1465 list_empty(&req->rq_replay_list)) {
1466 /* NB: don't call rq_commit_cb if it's already on
1467 * rq_replay_list, ptlrpc_free_committed() will call
1468 * it later, see LU-3618 for details */
1469 spin_unlock(&imp->imp_lock);
1470 req->rq_commit_cb(req);
1471 spin_lock(&imp->imp_lock);
1475 * Replay-enabled imports return commit-status information.
1477 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1478 if (likely(committed > imp->imp_peer_committed_transno))
1479 imp->imp_peer_committed_transno = committed;
1481 ptlrpc_free_committed(imp);
1483 if (!list_empty(&imp->imp_replay_list)) {
1484 struct ptlrpc_request *last;
1486 last = list_entry(imp->imp_replay_list.prev,
1487 struct ptlrpc_request,
1490 * Requests with rq_replay stay on the list even if no
1491 * commit is expected.
1493 if (last->rq_transno > imp->imp_peer_committed_transno)
1494 ptlrpc_pinger_commit_expected(imp);
1497 spin_unlock(&imp->imp_lock);
1504 * Helper function to send request \a req over the network for the first time
1505 * Also adjusts request phase.
1506 * Returns 0 on success or error code.
1508 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1510 struct obd_import *imp = req->rq_import;
1515 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1517 /* do not try to go further if there is not enough memory in enc_pool */
1518 if (req->rq_sent && req->rq_bulk != NULL)
1519 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1520 pool_is_at_full_capacity())
1523 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1524 (!req->rq_generation_set ||
1525 req->rq_import_generation == imp->imp_generation))
1528 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1530 spin_lock(&imp->imp_lock);
1532 LASSERT(req->rq_xid != 0);
1533 LASSERT(!list_empty(&req->rq_unreplied_list));
1535 if (!req->rq_generation_set)
1536 req->rq_import_generation = imp->imp_generation;
1538 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1539 spin_lock(&req->rq_lock);
1540 req->rq_waiting = 1;
1541 spin_unlock(&req->rq_lock);
1543 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1544 ptlrpc_import_state_name(req->rq_send_state),
1545 ptlrpc_import_state_name(imp->imp_state));
1546 LASSERT(list_empty(&req->rq_list));
1547 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1548 atomic_inc(&req->rq_import->imp_inflight);
1549 spin_unlock(&imp->imp_lock);
1554 spin_unlock(&imp->imp_lock);
1555 req->rq_status = rc;
1556 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1560 LASSERT(list_empty(&req->rq_list));
1561 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1562 atomic_inc(&req->rq_import->imp_inflight);
1564 /* find the known replied XID from the unreplied list, CONNECT
1565 * and DISCONNECT requests are skipped to make the sanity check
1566 * on server side happy. see process_req_last_xid().
1568 * For CONNECT: Because replay requests have lower XID, it'll
1569 * break the sanity check if CONNECT bump the exp_last_xid on
1572 * For DISCONNECT: Since client will abort inflight RPC before
1573 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1574 * than the inflight RPC.
1576 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1577 min_xid = ptlrpc_known_replied_xid(imp);
1578 spin_unlock(&imp->imp_lock);
1580 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1582 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1584 rc = sptlrpc_req_refresh_ctx(req, -1);
1587 req->rq_status = rc;
1590 spin_lock(&req->rq_lock);
1591 req->rq_wait_ctx = 1;
1592 spin_unlock(&req->rq_lock);
1597 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1598 " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1599 imp->imp_obd->obd_uuid.uuid,
1600 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1601 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1602 lustre_msg_get_opc(req->rq_reqmsg));
1604 rc = ptl_send_rpc(req, 0);
1605 if (rc == -ENOMEM) {
1606 spin_lock(&imp->imp_lock);
1607 if (!list_empty(&req->rq_list)) {
1608 list_del_init(&req->rq_list);
1609 atomic_dec(&req->rq_import->imp_inflight);
1611 spin_unlock(&imp->imp_lock);
1612 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1616 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1617 spin_lock(&req->rq_lock);
1618 req->rq_net_err = 1;
1619 spin_unlock(&req->rq_lock);
1625 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1630 LASSERT(set->set_producer != NULL);
1632 remaining = atomic_read(&set->set_remaining);
1634 /* populate the ->set_requests list with requests until we
1635 * reach the maximum number of RPCs in flight for this set */
1636 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1637 rc = set->set_producer(set, set->set_producer_arg);
1638 if (rc == -ENOENT) {
1639 /* no more RPC to produce */
1640 set->set_producer = NULL;
1641 set->set_producer_arg = NULL;
1646 RETURN((atomic_read(&set->set_remaining) - remaining));
1650 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1651 * and no more replies are expected.
1652 * (it is possible to get less replies than requests sent e.g. due to timed out
1653 * requests or requests that we had trouble to send out)
1655 * NOTE: This function contains a potential schedule point (cond_resched()).
1657 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1659 struct list_head *tmp, *next;
1660 struct list_head comp_reqs;
1661 int force_timer_recalc = 0;
1664 if (atomic_read(&set->set_remaining) == 0)
1667 INIT_LIST_HEAD(&comp_reqs);
1668 list_for_each_safe(tmp, next, &set->set_requests) {
1669 struct ptlrpc_request *req =
1670 list_entry(tmp, struct ptlrpc_request,
1672 struct obd_import *imp = req->rq_import;
1673 int unregistered = 0;
1677 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1678 list_move_tail(&req->rq_set_chain, &comp_reqs);
1682 /* This schedule point is mainly for the ptlrpcd caller of this
1683 * function. Most ptlrpc sets are not long-lived and unbounded
1684 * in length, but at the least the set used by the ptlrpcd is.
1685 * Since the processing time is unbounded, we need to insert an
1686 * explicit schedule point to make the thread well-behaved.
1690 /* If the caller requires to allow to be interpreted by force
1691 * and it has really been interpreted, then move the request
1692 * to RQ_PHASE_INTERPRET phase in spite of what the current
1694 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1695 req->rq_status = -EINTR;
1696 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1698 /* Since it is interpreted and we have to wait for
1699 * the reply to be unlinked, then use sync mode. */
1702 GOTO(interpret, req->rq_status);
1705 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1706 force_timer_recalc = 1;
1708 /* delayed send - skip */
1709 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1712 /* delayed resend - skip */
1713 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1714 req->rq_sent > ktime_get_real_seconds())
1717 if (!(req->rq_phase == RQ_PHASE_RPC ||
1718 req->rq_phase == RQ_PHASE_BULK ||
1719 req->rq_phase == RQ_PHASE_INTERPRET ||
1720 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1721 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1722 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1726 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1727 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1728 LASSERT(req->rq_next_phase != req->rq_phase);
1729 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1731 if (req->rq_req_deadline &&
1732 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1733 req->rq_req_deadline = 0;
1734 if (req->rq_reply_deadline &&
1735 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1736 req->rq_reply_deadline = 0;
1737 if (req->rq_bulk_deadline &&
1738 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1739 req->rq_bulk_deadline = 0;
1742 * Skip processing until reply is unlinked. We
1743 * can't return to pool before that and we can't
1744 * call interpret before that. We need to make
1745 * sure that all rdma transfers finished and will
1746 * not corrupt any data.
1748 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1749 ptlrpc_client_recv_or_unlink(req))
1751 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1752 ptlrpc_client_bulk_active(req))
1756 * Turn fail_loc off to prevent it from looping
1759 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1760 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1763 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1764 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1769 * Move to next phase if reply was successfully
1772 ptlrpc_rqphase_move(req, req->rq_next_phase);
1775 if (req->rq_phase == RQ_PHASE_INTERPRET)
1776 GOTO(interpret, req->rq_status);
1779 * Note that this also will start async reply unlink.
1781 if (req->rq_net_err && !req->rq_timedout) {
1782 ptlrpc_expire_one_request(req, 1);
1785 * Check if we still need to wait for unlink.
1787 if (ptlrpc_client_recv_or_unlink(req) ||
1788 ptlrpc_client_bulk_active(req))
1790 /* If there is no need to resend, fail it now. */
1791 if (req->rq_no_resend) {
1792 if (req->rq_status == 0)
1793 req->rq_status = -EIO;
1794 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1795 GOTO(interpret, req->rq_status);
1802 spin_lock(&req->rq_lock);
1803 req->rq_replied = 0;
1804 spin_unlock(&req->rq_lock);
1805 if (req->rq_status == 0)
1806 req->rq_status = -EIO;
1807 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1808 GOTO(interpret, req->rq_status);
1811 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1812 * so it sets rq_intr regardless of individual rpc
1813 * timeouts. The synchronous IO waiting path sets
1814 * rq_intr irrespective of whether ptlrpcd
1815 * has seen a timeout. Our policy is to only interpret
1816 * interrupted rpcs after they have timed out, so we
1817 * need to enforce that here.
1820 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1821 req->rq_wait_ctx)) {
1822 req->rq_status = -EINTR;
1823 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1824 GOTO(interpret, req->rq_status);
1827 if (req->rq_phase == RQ_PHASE_RPC) {
1828 if (req->rq_timedout || req->rq_resend ||
1829 req->rq_waiting || req->rq_wait_ctx) {
1832 if (!ptlrpc_unregister_reply(req, 1)) {
1833 ptlrpc_unregister_bulk(req, 1);
1837 spin_lock(&imp->imp_lock);
1838 if (ptlrpc_import_delay_req(imp, req, &status)){
1839 /* put on delay list - only if we wait
1840 * recovery finished - before send */
1841 list_del_init(&req->rq_list);
1842 list_add_tail(&req->rq_list,
1845 spin_unlock(&imp->imp_lock);
1850 req->rq_status = status;
1851 ptlrpc_rqphase_move(req,
1852 RQ_PHASE_INTERPRET);
1853 spin_unlock(&imp->imp_lock);
1854 GOTO(interpret, req->rq_status);
1856 if (ptlrpc_no_resend(req) &&
1857 !req->rq_wait_ctx) {
1858 req->rq_status = -ENOTCONN;
1859 ptlrpc_rqphase_move(req,
1860 RQ_PHASE_INTERPRET);
1861 spin_unlock(&imp->imp_lock);
1862 GOTO(interpret, req->rq_status);
1865 list_del_init(&req->rq_list);
1866 list_add_tail(&req->rq_list,
1867 &imp->imp_sending_list);
1869 spin_unlock(&imp->imp_lock);
1871 spin_lock(&req->rq_lock);
1872 req->rq_waiting = 0;
1873 spin_unlock(&req->rq_lock);
1875 if (req->rq_timedout || req->rq_resend) {
1876 /* This is re-sending anyways,
1877 * let's mark req as resend. */
1878 spin_lock(&req->rq_lock);
1880 spin_unlock(&req->rq_lock);
1882 if (req->rq_bulk != NULL &&
1883 !ptlrpc_unregister_bulk(req, 1))
1887 * rq_wait_ctx is only touched by ptlrpcd,
1888 * so no lock is needed here.
1890 status = sptlrpc_req_refresh_ctx(req, -1);
1893 req->rq_status = status;
1894 spin_lock(&req->rq_lock);
1895 req->rq_wait_ctx = 0;
1896 spin_unlock(&req->rq_lock);
1897 force_timer_recalc = 1;
1899 spin_lock(&req->rq_lock);
1900 req->rq_wait_ctx = 1;
1901 spin_unlock(&req->rq_lock);
1906 spin_lock(&req->rq_lock);
1907 req->rq_wait_ctx = 0;
1908 spin_unlock(&req->rq_lock);
1911 rc = ptl_send_rpc(req, 0);
1912 if (rc == -ENOMEM) {
1913 spin_lock(&imp->imp_lock);
1914 if (!list_empty(&req->rq_list))
1915 list_del_init(&req->rq_list);
1916 spin_unlock(&imp->imp_lock);
1917 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1921 DEBUG_REQ(D_HA, req,
1922 "send failed: rc = %d", rc);
1923 force_timer_recalc = 1;
1924 spin_lock(&req->rq_lock);
1925 req->rq_net_err = 1;
1926 spin_unlock(&req->rq_lock);
1929 /* need to reset the timeout */
1930 force_timer_recalc = 1;
1933 spin_lock(&req->rq_lock);
1935 if (ptlrpc_client_early(req)) {
1936 ptlrpc_at_recv_early_reply(req);
1937 spin_unlock(&req->rq_lock);
1941 /* Still waiting for a reply? */
1942 if (ptlrpc_client_recv(req)) {
1943 spin_unlock(&req->rq_lock);
1947 /* Did we actually receive a reply? */
1948 if (!ptlrpc_client_replied(req)) {
1949 spin_unlock(&req->rq_lock);
1953 spin_unlock(&req->rq_lock);
1955 /* unlink from net because we are going to
1956 * swab in-place of reply buffer */
1957 unregistered = ptlrpc_unregister_reply(req, 1);
1961 req->rq_status = after_reply(req);
1965 /* If there is no bulk associated with this request,
1966 * then we're done and should let the interpreter
1967 * process the reply. Similarly if the RPC returned
1968 * an error, and therefore the bulk will never arrive.
1970 if (req->rq_bulk == NULL || req->rq_status < 0) {
1971 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1972 GOTO(interpret, req->rq_status);
1975 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1978 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1979 if (ptlrpc_client_bulk_active(req))
1982 if (req->rq_bulk->bd_failure) {
1983 /* The RPC reply arrived OK, but the bulk screwed
1984 * up! Dead weird since the server told us the RPC
1985 * was good after getting the REPLY for her GET or
1986 * the ACK for her PUT. */
1987 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1988 req->rq_status = -EIO;
1991 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1994 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1996 /* This moves to "unregistering" phase we need to wait for
1998 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
1999 /* start async bulk unlink too */
2000 ptlrpc_unregister_bulk(req, 1);
2004 if (!ptlrpc_unregister_bulk(req, async))
2007 /* When calling interpret receiving already should be
2009 LASSERT(!req->rq_receiving_reply);
2011 ptlrpc_req_interpret(env, req, req->rq_status);
2013 if (ptlrpcd_check_work(req)) {
2014 atomic_dec(&set->set_remaining);
2017 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2019 if (req->rq_reqmsg != NULL)
2021 "Completed RPC pname:cluuid:pid:xid:nid:"
2022 "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
2023 imp->imp_obd->obd_uuid.uuid,
2024 lustre_msg_get_status(req->rq_reqmsg),
2026 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2027 lustre_msg_get_opc(req->rq_reqmsg));
2029 spin_lock(&imp->imp_lock);
2030 /* Request already may be not on sending or delaying list. This
2031 * may happen in the case of marking it erroneous for the case
2032 * ptlrpc_import_delay_req(req, status) find it impossible to
2033 * allow sending this rpc and returns *status != 0. */
2034 if (!list_empty(&req->rq_list)) {
2035 list_del_init(&req->rq_list);
2036 atomic_dec(&imp->imp_inflight);
2038 list_del_init(&req->rq_unreplied_list);
2039 spin_unlock(&imp->imp_lock);
2041 atomic_dec(&set->set_remaining);
2042 wake_up_all(&imp->imp_recovery_waitq);
2044 if (set->set_producer) {
2045 /* produce a new request if possible */
2046 if (ptlrpc_set_producer(set) > 0)
2047 force_timer_recalc = 1;
2049 /* free the request that has just been completed
2050 * in order not to pollute set->set_requests */
2051 list_del_init(&req->rq_set_chain);
2052 spin_lock(&req->rq_lock);
2054 req->rq_invalid_rqset = 0;
2055 spin_unlock(&req->rq_lock);
2057 /* record rq_status to compute the final status later */
2058 if (req->rq_status != 0)
2059 set->set_rc = req->rq_status;
2060 ptlrpc_req_finished(req);
2062 list_move_tail(&req->rq_set_chain, &comp_reqs);
2066 /* move completed request at the head of list so it's easier for
2067 * caller to find them */
2068 list_splice(&comp_reqs, &set->set_requests);
2070 /* If we hit an error, we want to recover promptly. */
2071 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2073 EXPORT_SYMBOL(ptlrpc_check_set);
2076 * Time out request \a req. is \a async_unlink is set, that means do not wait
2077 * until LNet actually confirms network buffer unlinking.
2078 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2080 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2082 struct obd_import *imp = req->rq_import;
2083 unsigned int debug_mask = D_RPCTRACE;
2087 spin_lock(&req->rq_lock);
2088 req->rq_timedout = 1;
2089 spin_unlock(&req->rq_lock);
2091 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2092 lustre_msg_get_status(req->rq_reqmsg)))
2093 debug_mask = D_WARNING;
2094 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2095 req->rq_net_err ? "failed due to network error" :
2096 ((req->rq_real_sent == 0 ||
2097 req->rq_real_sent < req->rq_sent ||
2098 req->rq_real_sent >= req->rq_deadline) ?
2099 "timed out for sent delay" : "timed out for slow reply"),
2100 (s64)req->rq_sent, (s64)req->rq_real_sent);
2102 if (imp != NULL && obd_debug_peer_on_timeout)
2103 LNetDebugPeer(imp->imp_connection->c_peer);
2105 ptlrpc_unregister_reply(req, async_unlink);
2106 ptlrpc_unregister_bulk(req, async_unlink);
2108 if (obd_dump_on_timeout)
2109 libcfs_debug_dumplog();
2112 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2116 atomic_inc(&imp->imp_timeouts);
2118 /* The DLM server doesn't want recovery run on its imports. */
2119 if (imp->imp_dlm_fake)
2122 /* If this request is for recovery or other primordial tasks,
2123 * then error it out here. */
2124 if (req->rq_ctx_init || req->rq_ctx_fini ||
2125 req->rq_send_state != LUSTRE_IMP_FULL ||
2126 imp->imp_obd->obd_no_recov) {
2127 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2128 ptlrpc_import_state_name(req->rq_send_state),
2129 ptlrpc_import_state_name(imp->imp_state));
2130 spin_lock(&req->rq_lock);
2131 req->rq_status = -ETIMEDOUT;
2133 spin_unlock(&req->rq_lock);
2137 /* if a request can't be resent we can't wait for an answer after
2139 if (ptlrpc_no_resend(req)) {
2140 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2144 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2150 * Time out all uncompleted requests in request set pointed by \a data
2151 * Callback used when waiting on sets with l_wait_event.
2154 int ptlrpc_expired_set(void *data)
2156 struct ptlrpc_request_set *set = data;
2157 struct list_head *tmp;
2158 time64_t now = ktime_get_real_seconds();
2161 LASSERT(set != NULL);
2164 * A timeout expired. See which reqs it applies to...
2166 list_for_each(tmp, &set->set_requests) {
2167 struct ptlrpc_request *req =
2168 list_entry(tmp, struct ptlrpc_request,
2171 /* don't expire request waiting for context */
2172 if (req->rq_wait_ctx)
2175 /* Request in-flight? */
2176 if (!((req->rq_phase == RQ_PHASE_RPC &&
2177 !req->rq_waiting && !req->rq_resend) ||
2178 (req->rq_phase == RQ_PHASE_BULK)))
2181 if (req->rq_timedout || /* already dealt with */
2182 req->rq_deadline > now) /* not expired */
2185 /* Deal with this guy. Do it asynchronously to not block
2186 * ptlrpcd thread. */
2187 ptlrpc_expire_one_request(req, 1);
2191 * When waiting for a whole set, we always break out of the
2192 * sleep so we can recalculate the timeout, or enable interrupts
2193 * if everyone's timed out.
2199 * Sets rq_intr flag in \a req under spinlock.
2201 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2203 spin_lock(&req->rq_lock);
2205 spin_unlock(&req->rq_lock);
2207 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2210 * Interrupts (sets interrupted flag) all uncompleted requests in
2211 * a set \a data. Callback for l_wait_event for interruptible waits.
2213 static void ptlrpc_interrupted_set(void *data)
2215 struct ptlrpc_request_set *set = data;
2216 struct list_head *tmp;
2218 LASSERT(set != NULL);
2219 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2221 list_for_each(tmp, &set->set_requests) {
2222 struct ptlrpc_request *req =
2223 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2228 if (req->rq_phase != RQ_PHASE_RPC &&
2229 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2230 !req->rq_allow_intr)
2233 ptlrpc_mark_interrupted(req);
2238 * Get the smallest timeout in the set; this does NOT set a timeout.
2240 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2242 struct list_head *tmp;
2243 time64_t now = ktime_get_real_seconds();
2245 struct ptlrpc_request *req;
2249 list_for_each(tmp, &set->set_requests) {
2250 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2253 * Request in-flight?
2255 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2256 (req->rq_phase == RQ_PHASE_BULK) ||
2257 (req->rq_phase == RQ_PHASE_NEW)))
2261 * Already timed out.
2263 if (req->rq_timedout)
2269 if (req->rq_wait_ctx)
2272 if (req->rq_phase == RQ_PHASE_NEW)
2273 deadline = req->rq_sent;
2274 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2275 deadline = req->rq_sent;
2277 deadline = req->rq_sent + req->rq_timeout;
2279 if (deadline <= now) /* actually expired already */
2280 timeout = 1; /* ASAP */
2281 else if (timeout == 0 || timeout > deadline - now)
2282 timeout = deadline - now;
2288 * Send all unset request from the set and then wait untill all
2289 * requests in the set complete (either get a reply, timeout, get an
2290 * error or otherwise be interrupted).
2291 * Returns 0 on success or error code otherwise.
2293 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2295 struct list_head *tmp;
2296 struct ptlrpc_request *req;
2297 struct l_wait_info lwi;
2302 if (set->set_producer)
2303 (void)ptlrpc_set_producer(set);
2305 list_for_each(tmp, &set->set_requests) {
2306 req = list_entry(tmp, struct ptlrpc_request,
2308 if (req->rq_phase == RQ_PHASE_NEW)
2309 (void)ptlrpc_send_new_req(req);
2312 if (list_empty(&set->set_requests))
2316 timeout = ptlrpc_set_next_timeout(set);
2318 /* wait until all complete, interrupted, or an in-flight
2320 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2323 if ((timeout == 0 && !signal_pending(current)) ||
2324 set->set_allow_intr)
2325 /* No requests are in-flight (ether timed out
2326 * or delayed), so we can allow interrupts.
2327 * We still want to block for a limited time,
2328 * so we allow interrupts during the timeout. */
2329 lwi = LWI_TIMEOUT_INTR_ALL(
2330 cfs_time_seconds(timeout ? timeout : 1),
2332 ptlrpc_interrupted_set, set);
2335 * At least one request is in flight, so no
2336 * interrupts are allowed. Wait until all
2337 * complete, or an in-flight req times out.
2339 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2340 ptlrpc_expired_set, set);
2342 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2344 /* LU-769 - if we ignored the signal because it was already
2345 * pending when we started, we need to handle it now or we risk
2346 * it being ignored forever */
2347 if (rc == -ETIMEDOUT &&
2348 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2349 signal_pending(current)) {
2350 sigset_t blocked_sigs =
2351 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2353 /* In fact we only interrupt for the "fatal" signals
2354 * like SIGINT or SIGKILL. We still ignore less
2355 * important signals since ptlrpc set is not easily
2356 * reentrant from userspace again */
2357 if (signal_pending(current))
2358 ptlrpc_interrupted_set(set);
2359 cfs_restore_sigs(blocked_sigs);
2362 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2364 /* -EINTR => all requests have been flagged rq_intr so next
2366 * -ETIMEDOUT => someone timed out. When all reqs have
2367 * timed out, signals are enabled allowing completion with
2369 * I don't really care if we go once more round the loop in
2370 * the error cases -eeb. */
2371 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2372 list_for_each(tmp, &set->set_requests) {
2373 req = list_entry(tmp, struct ptlrpc_request,
2375 spin_lock(&req->rq_lock);
2376 req->rq_invalid_rqset = 1;
2377 spin_unlock(&req->rq_lock);
2380 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2382 LASSERT(atomic_read(&set->set_remaining) == 0);
2384 rc = set->set_rc; /* rq_status of already freed requests if any */
2385 list_for_each(tmp, &set->set_requests) {
2386 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2388 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2389 if (req->rq_status != 0)
2390 rc = req->rq_status;
2395 EXPORT_SYMBOL(ptlrpc_set_wait);
2398 * Helper fuction for request freeing.
2399 * Called when request count reached zero and request needs to be freed.
2400 * Removes request from all sorts of sending/replay lists it might be on,
2401 * frees network buffers if any are present.
2402 * If \a locked is set, that means caller is already holding import imp_lock
2403 * and so we no longer need to reobtain it (for certain lists manipulations)
2405 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2409 if (request == NULL)
2412 LASSERT(!request->rq_srv_req);
2413 LASSERT(request->rq_export == NULL);
2414 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2415 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2416 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2417 LASSERTF(!request->rq_replay, "req %p\n", request);
2419 req_capsule_fini(&request->rq_pill);
2421 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2422 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2423 if (request->rq_import != NULL) {
2425 spin_lock(&request->rq_import->imp_lock);
2426 list_del_init(&request->rq_replay_list);
2427 list_del_init(&request->rq_unreplied_list);
2429 spin_unlock(&request->rq_import->imp_lock);
2431 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2433 if (atomic_read(&request->rq_refcount) != 0) {
2434 DEBUG_REQ(D_ERROR, request,
2435 "freeing request with nonzero refcount");
2439 if (request->rq_repbuf != NULL)
2440 sptlrpc_cli_free_repbuf(request);
2442 if (request->rq_import != NULL) {
2443 class_import_put(request->rq_import);
2444 request->rq_import = NULL;
2446 if (request->rq_bulk != NULL)
2447 ptlrpc_free_bulk(request->rq_bulk);
2449 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2450 sptlrpc_cli_free_reqbuf(request);
2452 if (request->rq_cli_ctx)
2453 sptlrpc_req_put_ctx(request, !locked);
2455 if (request->rq_pool)
2456 __ptlrpc_free_req_to_pool(request);
2458 ptlrpc_request_cache_free(request);
2462 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2464 * Drop one request reference. Must be called with import imp_lock held.
2465 * When reference count drops to zero, request is freed.
2467 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2469 assert_spin_locked(&request->rq_import->imp_lock);
2470 (void)__ptlrpc_req_finished(request, 1);
2475 * Drops one reference count for request \a request.
2476 * \a locked set indicates that caller holds import imp_lock.
2477 * Frees the request whe reference count reaches zero.
2479 * \retval 1 the request is freed
2480 * \retval 0 some others still hold references on the request
2482 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2490 LASSERT(request != LP_POISON);
2491 LASSERT(request->rq_reqmsg != LP_POISON);
2493 DEBUG_REQ(D_INFO, request, "refcount now %u",
2494 atomic_read(&request->rq_refcount) - 1);
2496 spin_lock(&request->rq_lock);
2497 count = atomic_dec_return(&request->rq_refcount);
2498 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2500 /* For open RPC, the client does not know the EA size (LOV, ACL, and
2501 * so on) before replied, then the client has to reserve very large
2502 * reply buffer. Such buffer will not be released until the RPC freed.
2503 * Since The open RPC is replayable, we need to keep it in the replay
2504 * list until close. If there are a lot of files opened concurrently,
2505 * then the client may be OOM.
2507 * If fact, it is unnecessary to keep reply buffer for open replay,
2508 * related EAs have already been saved via mdc_save_lovea() before
2509 * coming here. So it is safe to free the reply buffer some earlier
2510 * before releasing the RPC to avoid client OOM. LU-9514 */
2511 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2512 spin_lock(&request->rq_early_free_lock);
2513 sptlrpc_cli_free_repbuf(request);
2514 request->rq_repbuf = NULL;
2515 request->rq_repbuf_len = 0;
2516 request->rq_repdata = NULL;
2517 request->rq_reqdata_len = 0;
2518 spin_unlock(&request->rq_early_free_lock);
2520 spin_unlock(&request->rq_lock);
2523 __ptlrpc_free_req(request, locked);
2529 * Drops one reference count for a request.
2531 void ptlrpc_req_finished(struct ptlrpc_request *request)
2533 __ptlrpc_req_finished(request, 0);
2535 EXPORT_SYMBOL(ptlrpc_req_finished);
2538 * Returns xid of a \a request
2540 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2542 return request->rq_xid;
2544 EXPORT_SYMBOL(ptlrpc_req_xid);
2547 * Disengage the client's reply buffer from the network
2548 * NB does _NOT_ unregister any client-side bulk.
2549 * IDEMPOTENT, but _not_ safe against concurrent callers.
2550 * The request owner (i.e. the thread doing the I/O) must call...
2551 * Returns 0 on success or 1 if unregistering cannot be made.
2553 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2556 struct l_wait_info lwi;
2561 LASSERT(!in_interrupt());
2563 /* Let's setup deadline for reply unlink. */
2564 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2565 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2566 request->rq_reply_deadline = ktime_get_real_seconds() +
2570 * Nothing left to do.
2572 if (!ptlrpc_client_recv_or_unlink(request))
2575 LNetMDUnlink(request->rq_reply_md_h);
2578 * Let's check it once again.
2580 if (!ptlrpc_client_recv_or_unlink(request))
2583 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2584 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2587 * Do not wait for unlink to finish.
2593 * We have to l_wait_event() whatever the result, to give liblustre
2594 * a chance to run reply_in_callback(), and to make sure we've
2595 * unlinked before returning a req to the pool.
2598 /* The wq argument is ignored by user-space wait_event macros */
2599 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2600 &request->rq_set->set_waitq :
2601 &request->rq_reply_waitq;
2602 /* Network access will complete in finite time but the HUGE
2603 * timeout lets us CWARN for visibility of sluggish NALs */
2604 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2605 cfs_time_seconds(1), NULL, NULL);
2606 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2609 ptlrpc_rqphase_move(request, request->rq_next_phase);
2613 LASSERT(rc == -ETIMEDOUT);
2614 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2615 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2616 request->rq_receiving_reply,
2617 request->rq_req_unlinked,
2618 request->rq_reply_unlinked);
2623 static void ptlrpc_free_request(struct ptlrpc_request *req)
2625 spin_lock(&req->rq_lock);
2627 spin_unlock(&req->rq_lock);
2629 if (req->rq_commit_cb != NULL)
2630 req->rq_commit_cb(req);
2631 list_del_init(&req->rq_replay_list);
2633 __ptlrpc_req_finished(req, 1);
2637 * the request is committed and dropped from the replay list of its import
2639 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2641 struct obd_import *imp = req->rq_import;
2643 spin_lock(&imp->imp_lock);
2644 if (list_empty(&req->rq_replay_list)) {
2645 spin_unlock(&imp->imp_lock);
2649 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2650 ptlrpc_free_request(req);
2652 spin_unlock(&imp->imp_lock);
2654 EXPORT_SYMBOL(ptlrpc_request_committed);
2657 * Iterates through replay_list on import and prunes
2658 * all requests have transno smaller than last_committed for the
2659 * import and don't have rq_replay set.
2660 * Since requests are sorted in transno order, stops when meetign first
2661 * transno bigger than last_committed.
2662 * caller must hold imp->imp_lock
2664 void ptlrpc_free_committed(struct obd_import *imp)
2666 struct ptlrpc_request *req, *saved;
2667 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2668 bool skip_committed_list = true;
2671 LASSERT(imp != NULL);
2672 assert_spin_locked(&imp->imp_lock);
2674 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2675 imp->imp_generation == imp->imp_last_generation_checked) {
2676 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2677 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2680 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2681 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2682 imp->imp_generation);
2684 if (imp->imp_generation != imp->imp_last_generation_checked ||
2685 imp->imp_last_transno_checked == 0)
2686 skip_committed_list = false;
2688 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2689 imp->imp_last_generation_checked = imp->imp_generation;
2691 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2693 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2694 LASSERT(req != last_req);
2697 if (req->rq_transno == 0) {
2698 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2701 if (req->rq_import_generation < imp->imp_generation) {
2702 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2706 /* not yet committed */
2707 if (req->rq_transno > imp->imp_peer_committed_transno) {
2708 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2712 if (req->rq_replay) {
2713 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2714 list_move_tail(&req->rq_replay_list,
2715 &imp->imp_committed_list);
2719 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2720 imp->imp_peer_committed_transno);
2722 ptlrpc_free_request(req);
2725 if (skip_committed_list)
2728 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2730 LASSERT(req->rq_transno != 0);
2731 if (req->rq_import_generation < imp->imp_generation ||
2733 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2734 req->rq_import_generation <
2735 imp->imp_generation ? "stale" : "closed");
2737 if (imp->imp_replay_cursor == &req->rq_replay_list)
2738 imp->imp_replay_cursor =
2739 req->rq_replay_list.next;
2741 ptlrpc_free_request(req);
2748 void ptlrpc_cleanup_client(struct obd_import *imp)
2755 * Schedule previously sent request for resend.
2756 * For bulk requests we assign new xid (to avoid problems with
2757 * lost replies and therefore several transfers landing into same buffer
2758 * from different sending attempts).
2760 void ptlrpc_resend_req(struct ptlrpc_request *req)
2762 DEBUG_REQ(D_HA, req, "going to resend");
2763 spin_lock(&req->rq_lock);
2765 /* Request got reply but linked to the import list still.
2766 Let ptlrpc_check_set() to process it. */
2767 if (ptlrpc_client_replied(req)) {
2768 spin_unlock(&req->rq_lock);
2769 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2773 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2774 req->rq_status = -EAGAIN;
2777 req->rq_net_err = 0;
2778 req->rq_timedout = 0;
2780 ptlrpc_client_wake_req(req);
2781 spin_unlock(&req->rq_lock);
2784 /* XXX: this function and rq_status are currently unused */
2785 void ptlrpc_restart_req(struct ptlrpc_request *req)
2787 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2788 req->rq_status = -ERESTARTSYS;
2790 spin_lock(&req->rq_lock);
2791 req->rq_restart = 1;
2792 req->rq_timedout = 0;
2793 ptlrpc_client_wake_req(req);
2794 spin_unlock(&req->rq_lock);
2798 * Grab additional reference on a request \a req
2800 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2803 atomic_inc(&req->rq_refcount);
2806 EXPORT_SYMBOL(ptlrpc_request_addref);
2809 * Add a request to import replay_list.
2810 * Must be called under imp_lock
2812 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2813 struct obd_import *imp)
2815 struct list_head *tmp;
2817 assert_spin_locked(&imp->imp_lock);
2819 if (req->rq_transno == 0) {
2820 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2824 /* clear this for new requests that were resent as well
2825 as resent replayed requests. */
2826 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2828 /* don't re-add requests that have been replayed */
2829 if (!list_empty(&req->rq_replay_list))
2832 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2834 spin_lock(&req->rq_lock);
2836 spin_unlock(&req->rq_lock);
2838 LASSERT(imp->imp_replayable);
2839 /* Balanced in ptlrpc_free_committed, usually. */
2840 ptlrpc_request_addref(req);
2841 list_for_each_prev(tmp, &imp->imp_replay_list) {
2842 struct ptlrpc_request *iter = list_entry(tmp,
2843 struct ptlrpc_request,
2846 /* We may have duplicate transnos if we create and then
2847 * open a file, or for closes retained if to match creating
2848 * opens, so use req->rq_xid as a secondary key.
2849 * (See bugs 684, 685, and 428.)
2850 * XXX no longer needed, but all opens need transnos!
2852 if (iter->rq_transno > req->rq_transno)
2855 if (iter->rq_transno == req->rq_transno) {
2856 LASSERT(iter->rq_xid != req->rq_xid);
2857 if (iter->rq_xid > req->rq_xid)
2861 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2865 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2869 * Send request and wait until it completes.
2870 * Returns request processing status.
2872 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2874 struct ptlrpc_request_set *set;
2878 LASSERT(req->rq_set == NULL);
2879 LASSERT(!req->rq_receiving_reply);
2881 set = ptlrpc_prep_set();
2883 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2887 /* for distributed debugging */
2888 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2890 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2891 ptlrpc_request_addref(req);
2892 ptlrpc_set_add_req(set, req);
2893 rc = ptlrpc_set_wait(set);
2894 ptlrpc_set_destroy(set);
2898 EXPORT_SYMBOL(ptlrpc_queue_wait);
2901 * Callback used for replayed requests reply processing.
2902 * In case of successful reply calls registered request replay callback.
2903 * In case of error restart replay process.
2905 static int ptlrpc_replay_interpret(const struct lu_env *env,
2906 struct ptlrpc_request *req,
2907 void * data, int rc)
2909 struct ptlrpc_replay_async_args *aa = data;
2910 struct obd_import *imp = req->rq_import;
2913 atomic_dec(&imp->imp_replay_inflight);
2915 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2916 * server got the request, but bulk transfer timeout, let's
2917 * replay the bulk req again */
2918 if (!ptlrpc_client_replied(req) ||
2919 (req->rq_bulk != NULL &&
2920 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2921 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2922 GOTO(out, rc = -ETIMEDOUT);
2925 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2926 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2927 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2928 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2930 /** VBR: check version failure */
2931 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2932 /** replay was failed due to version mismatch */
2933 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2934 spin_lock(&imp->imp_lock);
2935 imp->imp_vbr_failed = 1;
2936 spin_unlock(&imp->imp_lock);
2937 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2939 /** The transno had better not change over replay. */
2940 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2941 lustre_msg_get_transno(req->rq_repmsg) ||
2942 lustre_msg_get_transno(req->rq_repmsg) == 0,
2944 lustre_msg_get_transno(req->rq_reqmsg),
2945 lustre_msg_get_transno(req->rq_repmsg));
2948 spin_lock(&imp->imp_lock);
2949 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2950 spin_unlock(&imp->imp_lock);
2951 LASSERT(imp->imp_last_replay_transno);
2953 /* transaction number shouldn't be bigger than the latest replayed */
2954 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2955 DEBUG_REQ(D_ERROR, req,
2956 "Reported transno %llu is bigger than the "
2957 "replayed one: %llu", req->rq_transno,
2958 lustre_msg_get_transno(req->rq_reqmsg));
2959 GOTO(out, rc = -EINVAL);
2962 DEBUG_REQ(D_HA, req, "got rep");
2964 /* let the callback do fixups, possibly including in the request */
2965 if (req->rq_replay_cb)
2966 req->rq_replay_cb(req);
2968 if (ptlrpc_client_replied(req) &&
2969 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2970 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2971 lustre_msg_get_status(req->rq_repmsg),
2972 aa->praa_old_status);
2974 /* Note: If the replay fails for MDT-MDT recovery, let's
2975 * abort all of the following requests in the replay
2976 * and sending list, because MDT-MDT update requests
2977 * are dependent on each other, see LU-7039 */
2978 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
2979 struct ptlrpc_request *free_req;
2980 struct ptlrpc_request *tmp;
2982 spin_lock(&imp->imp_lock);
2983 list_for_each_entry_safe(free_req, tmp,
2984 &imp->imp_replay_list,
2986 ptlrpc_free_request(free_req);
2989 list_for_each_entry_safe(free_req, tmp,
2990 &imp->imp_committed_list,
2992 ptlrpc_free_request(free_req);
2995 list_for_each_entry_safe(free_req, tmp,
2996 &imp->imp_delayed_list,
2998 spin_lock(&free_req->rq_lock);
2999 free_req->rq_err = 1;
3000 free_req->rq_status = -EIO;
3001 ptlrpc_client_wake_req(free_req);
3002 spin_unlock(&free_req->rq_lock);
3005 list_for_each_entry_safe(free_req, tmp,
3006 &imp->imp_sending_list,
3008 spin_lock(&free_req->rq_lock);
3009 free_req->rq_err = 1;
3010 free_req->rq_status = -EIO;
3011 ptlrpc_client_wake_req(free_req);
3012 spin_unlock(&free_req->rq_lock);
3014 spin_unlock(&imp->imp_lock);
3017 /* Put it back for re-replay. */
3018 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3022 * Errors while replay can set transno to 0, but
3023 * imp_last_replay_transno shouldn't be set to 0 anyway
3025 if (req->rq_transno == 0)
3026 CERROR("Transno is 0 during replay!\n");
3028 /* continue with recovery */
3029 rc = ptlrpc_import_recovery_state_machine(imp);
3031 req->rq_send_state = aa->praa_old_state;
3034 /* this replay failed, so restart recovery */
3035 ptlrpc_connect_import(imp);
3041 * Prepares and queues request for replay.
3042 * Adds it to ptlrpcd queue for actual sending.
3043 * Returns 0 on success.
3045 int ptlrpc_replay_req(struct ptlrpc_request *req)
3047 struct ptlrpc_replay_async_args *aa;
3051 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3053 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3054 aa = ptlrpc_req_async_args(req);
3055 memset(aa, 0, sizeof(*aa));
3057 /* Prepare request to be resent with ptlrpcd */
3058 aa->praa_old_state = req->rq_send_state;
3059 req->rq_send_state = LUSTRE_IMP_REPLAY;
3060 req->rq_phase = RQ_PHASE_NEW;
3061 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3063 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3065 req->rq_interpret_reply = ptlrpc_replay_interpret;
3066 /* Readjust the timeout for current conditions */
3067 ptlrpc_at_set_req_timeout(req);
3069 /* Tell server the net_latency, so the server can calculate how long
3070 * it should wait for next replay */
3071 lustre_msg_set_service_time(req->rq_reqmsg,
3072 ptlrpc_at_get_net_latency(req));
3073 DEBUG_REQ(D_HA, req, "REPLAY");
3075 atomic_inc(&req->rq_import->imp_replay_inflight);
3076 spin_lock(&req->rq_lock);
3077 req->rq_early_free_repbuf = 0;
3078 spin_unlock(&req->rq_lock);
3079 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3081 ptlrpcd_add_req(req);
3086 * Aborts all in-flight request on import \a imp sending and delayed lists
3088 void ptlrpc_abort_inflight(struct obd_import *imp)
3090 struct list_head *tmp, *n;
3093 /* Make sure that no new requests get processed for this import.
3094 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3095 * this flag and then putting requests on sending_list or delayed_list.
3097 spin_lock(&imp->imp_lock);
3099 /* XXX locking? Maybe we should remove each request with the list
3100 * locked? Also, how do we know if the requests on the list are
3101 * being freed at this time?
3103 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3104 struct ptlrpc_request *req = list_entry(tmp,
3105 struct ptlrpc_request,
3108 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3110 spin_lock(&req->rq_lock);
3111 if (req->rq_import_generation < imp->imp_generation) {
3113 req->rq_status = -EIO;
3114 ptlrpc_client_wake_req(req);
3116 spin_unlock(&req->rq_lock);
3119 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3120 struct ptlrpc_request *req =
3121 list_entry(tmp, struct ptlrpc_request, rq_list);
3123 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3125 spin_lock(&req->rq_lock);
3126 if (req->rq_import_generation < imp->imp_generation) {
3128 req->rq_status = -EIO;
3129 ptlrpc_client_wake_req(req);
3131 spin_unlock(&req->rq_lock);
3134 /* Last chance to free reqs left on the replay list, but we
3135 * will still leak reqs that haven't committed. */
3136 if (imp->imp_replayable)
3137 ptlrpc_free_committed(imp);
3139 spin_unlock(&imp->imp_lock);
3145 * Abort all uncompleted requests in request set \a set
3147 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3149 struct list_head *tmp, *pos;
3151 LASSERT(set != NULL);
3153 list_for_each_safe(pos, tmp, &set->set_requests) {
3154 struct ptlrpc_request *req =
3155 list_entry(pos, struct ptlrpc_request,
3158 spin_lock(&req->rq_lock);
3159 if (req->rq_phase != RQ_PHASE_RPC) {
3160 spin_unlock(&req->rq_lock);
3165 req->rq_status = -EINTR;
3166 ptlrpc_client_wake_req(req);
3167 spin_unlock(&req->rq_lock);
3171 static __u64 ptlrpc_last_xid;
3172 static spinlock_t ptlrpc_last_xid_lock;
3175 * Initialize the XID for the node. This is common among all requests on
3176 * this node, and only requires the property that it is monotonically
3177 * increasing. It does not need to be sequential. Since this is also used
3178 * as the RDMA match bits, it is important that a single client NOT have
3179 * the same match bits for two different in-flight requests, hence we do
3180 * NOT want to have an XID per target or similar.
3182 * To avoid an unlikely collision between match bits after a client reboot
3183 * (which would deliver old data into the wrong RDMA buffer) initialize
3184 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3185 * If the time is clearly incorrect, we instead use a 62-bit random number.
3186 * In the worst case the random number will overflow 1M RPCs per second in
3187 * 9133 years, or permutations thereof.
3189 #define YEAR_2004 (1ULL << 30)
3190 void ptlrpc_init_xid(void)
3192 time64_t now = ktime_get_real_seconds();
3194 spin_lock_init(&ptlrpc_last_xid_lock);
3195 if (now < YEAR_2004) {
3196 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3197 ptlrpc_last_xid >>= 2;
3198 ptlrpc_last_xid |= (1ULL << 61);
3200 ptlrpc_last_xid = (__u64)now << 20;
3203 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3204 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3205 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3209 * Increase xid and returns resulting new value to the caller.
3211 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3212 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3213 * itself uses the last bulk xid needed, so the server can determine the
3214 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3215 * xid must align to a power-of-two value.
3217 * This is assumed to be true due to the initial ptlrpc_last_xid
3218 * value also being initialized to a power-of-two value. LU-1431
3220 __u64 ptlrpc_next_xid(void)
3224 spin_lock(&ptlrpc_last_xid_lock);
3225 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3226 ptlrpc_last_xid = next;
3227 spin_unlock(&ptlrpc_last_xid_lock);
3233 * If request has a new allocated XID (new request or EINPROGRESS resend),
3234 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3235 * request to ensure previous bulk fails and avoid problems with lost replies
3236 * and therefore several transfers landing into the same buffer from different
3239 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3241 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3243 LASSERT(bd != NULL);
3245 /* Generate new matchbits for all resend requests, including
3247 if (req->rq_resend) {
3248 __u64 old_mbits = req->rq_mbits;
3250 /* First time resend on -EINPROGRESS will generate new xid,
3251 * so we can actually use the rq_xid as rq_mbits in such case,
3252 * however, it's bit hard to distinguish such resend with a
3253 * 'resend for the -EINPROGRESS resend'. To make it simple,
3254 * we opt to generate mbits for all resend cases. */
3255 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
3256 req->rq_mbits = ptlrpc_next_xid();
3258 /* Old version transfers rq_xid to peer as
3260 spin_lock(&req->rq_import->imp_lock);
3261 list_del_init(&req->rq_unreplied_list);
3262 ptlrpc_assign_next_xid_nolock(req);
3263 spin_unlock(&req->rq_import->imp_lock);
3264 req->rq_mbits = req->rq_xid;
3266 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3267 old_mbits, req->rq_mbits);
3268 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3269 /* Request being sent first time, use xid as matchbits. */
3270 req->rq_mbits = req->rq_xid;
3272 /* Replay request, xid and matchbits have already been
3273 * correctly assigned. */
3277 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3278 * that server can infer the number of bulks that were prepared,
3280 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3283 /* Set rq_xid as rq_mbits to indicate the final bulk for the old
3284 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3286 * It's ok to directly set the rq_xid here, since this xid bump
3287 * won't affect the request position in unreplied list. */
3288 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3289 req->rq_xid = req->rq_mbits;
3293 * Get a glimpse at what next xid value might have been.
3294 * Returns possible next xid.
3296 __u64 ptlrpc_sample_next_xid(void)
3298 #if BITS_PER_LONG == 32
3299 /* need to avoid possible word tearing on 32-bit systems */
3302 spin_lock(&ptlrpc_last_xid_lock);
3303 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3304 spin_unlock(&ptlrpc_last_xid_lock);
3308 /* No need to lock, since returned value is racy anyways */
3309 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3312 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3315 * Functions for operating ptlrpc workers.
3317 * A ptlrpc work is a function which will be running inside ptlrpc context.
3318 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3320 * 1. after a work is created, it can be used many times, that is:
3321 * handler = ptlrpcd_alloc_work();
3322 * ptlrpcd_queue_work();
3324 * queue it again when necessary:
3325 * ptlrpcd_queue_work();
3326 * ptlrpcd_destroy_work();
3327 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3328 * it will only be queued once in any time. Also as its name implies, it may
3329 * have delay before it really runs by ptlrpcd thread.
3331 struct ptlrpc_work_async_args {
3332 int (*cb)(const struct lu_env *, void *);
3336 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3338 /* re-initialize the req */
3339 req->rq_timeout = obd_timeout;
3340 req->rq_sent = ktime_get_real_seconds();
3341 req->rq_deadline = req->rq_sent + req->rq_timeout;
3342 req->rq_phase = RQ_PHASE_INTERPRET;
3343 req->rq_next_phase = RQ_PHASE_COMPLETE;
3344 req->rq_xid = ptlrpc_next_xid();
3345 req->rq_import_generation = req->rq_import->imp_generation;
3347 ptlrpcd_add_req(req);
3350 static int work_interpreter(const struct lu_env *env,
3351 struct ptlrpc_request *req, void *data, int rc)
3353 struct ptlrpc_work_async_args *arg = data;
3355 LASSERT(ptlrpcd_check_work(req));
3356 LASSERT(arg->cb != NULL);
3358 rc = arg->cb(env, arg->cbdata);
3360 list_del_init(&req->rq_set_chain);
3363 if (atomic_dec_return(&req->rq_refcount) > 1) {
3364 atomic_set(&req->rq_refcount, 2);
3365 ptlrpcd_add_work_req(req);
3370 static int worker_format;
3372 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3374 return req->rq_pill.rc_fmt == (void *)&worker_format;
3378 * Create a work for ptlrpc.
3380 void *ptlrpcd_alloc_work(struct obd_import *imp,
3381 int (*cb)(const struct lu_env *, void *), void *cbdata)
3383 struct ptlrpc_request *req = NULL;
3384 struct ptlrpc_work_async_args *args;
3390 RETURN(ERR_PTR(-EINVAL));
3392 /* copy some code from deprecated fakereq. */
3393 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3395 CERROR("ptlrpc: run out of memory!\n");
3396 RETURN(ERR_PTR(-ENOMEM));
3399 ptlrpc_cli_req_init(req);
3401 req->rq_send_state = LUSTRE_IMP_FULL;
3402 req->rq_type = PTL_RPC_MSG_REQUEST;
3403 req->rq_import = class_import_get(imp);
3404 req->rq_interpret_reply = work_interpreter;
3405 /* don't want reply */
3406 req->rq_no_delay = req->rq_no_resend = 1;
3407 req->rq_pill.rc_fmt = (void *)&worker_format;
3409 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3410 args = ptlrpc_req_async_args(req);
3412 args->cbdata = cbdata;
3416 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3418 void ptlrpcd_destroy_work(void *handler)
3420 struct ptlrpc_request *req = handler;
3423 ptlrpc_req_finished(req);
3425 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3427 int ptlrpcd_queue_work(void *handler)
3429 struct ptlrpc_request *req = handler;
3432 * Check if the req is already being queued.
3434 * Here comes a trick: it lacks a way of checking if a req is being
3435 * processed reliably in ptlrpc. Here I have to use refcount of req
3436 * for this purpose. This is okay because the caller should use this
3437 * req as opaque data. - Jinshan
3439 LASSERT(atomic_read(&req->rq_refcount) > 0);
3440 if (atomic_inc_return(&req->rq_refcount) == 2)
3441 ptlrpcd_add_work_req(req);
3444 EXPORT_SYMBOL(ptlrpcd_queue_work);