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;
870 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
872 CDEBUG_LIMIT(imp->imp_idle_debug,
873 "%s: reconnect after %llds idle\n",
874 imp->imp_obd->obd_name, ktime_get_real_seconds() -
875 imp->imp_last_reply_time);
876 spin_lock(&imp->imp_lock);
877 if (imp->imp_state == LUSTRE_IMP_IDLE) {
878 imp->imp_generation++;
879 imp->imp_initiated_at = imp->imp_generation;
880 imp->imp_state = LUSTRE_IMP_NEW;
883 spin_unlock(&imp->imp_lock);
885 rc = ptlrpc_connect_import(imp);
888 ptlrpc_pinger_add_import(imp);
892 request = __ptlrpc_request_alloc(imp, pool);
896 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
897 req_capsule_set(&request->rq_pill, format);
902 * Allocate new request structure for import \a imp and initialize its
903 * buffer structure according to capsule template \a format.
905 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
906 const struct req_format *format)
908 return ptlrpc_request_alloc_internal(imp, NULL, format);
910 EXPORT_SYMBOL(ptlrpc_request_alloc);
913 * Allocate new request structure for import \a imp from pool \a pool and
914 * initialize its buffer structure according to capsule template \a format.
916 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
917 struct ptlrpc_request_pool * pool,
918 const struct req_format *format)
920 return ptlrpc_request_alloc_internal(imp, pool, format);
922 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
925 * For requests not from pool, free memory of the request structure.
926 * For requests obtained from a pool earlier, return request back to pool.
928 void ptlrpc_request_free(struct ptlrpc_request *request)
930 if (request->rq_pool)
931 __ptlrpc_free_req_to_pool(request);
933 ptlrpc_request_cache_free(request);
935 EXPORT_SYMBOL(ptlrpc_request_free);
938 * Allocate new request for operatione \a opcode and immediatelly pack it for
940 * Only used for simple requests like OBD_PING where the only important
941 * part of the request is operation itself.
942 * Returns allocated request or NULL on error.
944 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
945 const struct req_format *format,
946 __u32 version, int opcode)
948 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
952 rc = ptlrpc_request_pack(req, version, opcode);
954 ptlrpc_request_free(req);
960 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
963 * Allocate and initialize new request set structure on the current CPT.
964 * Returns a pointer to the newly allocated set structure or NULL on error.
966 struct ptlrpc_request_set *ptlrpc_prep_set(void)
968 struct ptlrpc_request_set *set;
972 cpt = cfs_cpt_current(cfs_cpt_table, 0);
973 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
976 atomic_set(&set->set_refcount, 1);
977 INIT_LIST_HEAD(&set->set_requests);
978 init_waitqueue_head(&set->set_waitq);
979 atomic_set(&set->set_new_count, 0);
980 atomic_set(&set->set_remaining, 0);
981 spin_lock_init(&set->set_new_req_lock);
982 INIT_LIST_HEAD(&set->set_new_requests);
983 set->set_max_inflight = UINT_MAX;
984 set->set_producer = NULL;
985 set->set_producer_arg = NULL;
990 EXPORT_SYMBOL(ptlrpc_prep_set);
993 * Allocate and initialize new request set structure with flow control
994 * extension. This extension allows to control the number of requests in-flight
995 * for the whole set. A callback function to generate requests must be provided
996 * and the request set will keep the number of requests sent over the wire to
998 * Returns a pointer to the newly allocated set structure or NULL on error.
1000 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1004 struct ptlrpc_request_set *set;
1006 set = ptlrpc_prep_set();
1010 set->set_max_inflight = max;
1011 set->set_producer = func;
1012 set->set_producer_arg = arg;
1018 * Wind down and free request set structure previously allocated with
1020 * Ensures that all requests on the set have completed and removes
1021 * all requests from the request list in a set.
1022 * If any unsent request happen to be on the list, pretends that they got
1023 * an error in flight and calls their completion handler.
1025 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1027 struct list_head *tmp;
1028 struct list_head *next;
1033 /* Requests on the set should either all be completed, or all be new */
1034 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1035 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1036 list_for_each(tmp, &set->set_requests) {
1037 struct ptlrpc_request *req =
1038 list_entry(tmp, struct ptlrpc_request,
1041 LASSERT(req->rq_phase == expected_phase);
1045 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1046 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1047 atomic_read(&set->set_remaining), n);
1049 list_for_each_safe(tmp, next, &set->set_requests) {
1050 struct ptlrpc_request *req =
1051 list_entry(tmp, struct ptlrpc_request,
1053 list_del_init(&req->rq_set_chain);
1055 LASSERT(req->rq_phase == expected_phase);
1057 if (req->rq_phase == RQ_PHASE_NEW) {
1058 ptlrpc_req_interpret(NULL, req, -EBADR);
1059 atomic_dec(&set->set_remaining);
1062 spin_lock(&req->rq_lock);
1064 req->rq_invalid_rqset = 0;
1065 spin_unlock(&req->rq_lock);
1067 ptlrpc_req_finished (req);
1070 LASSERT(atomic_read(&set->set_remaining) == 0);
1072 ptlrpc_reqset_put(set);
1075 EXPORT_SYMBOL(ptlrpc_set_destroy);
1078 * Add a new request to the general purpose request set.
1079 * Assumes request reference from the caller.
1081 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1082 struct ptlrpc_request *req)
1084 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1085 LASSERT(list_empty(&req->rq_set_chain));
1087 if (req->rq_allow_intr)
1088 set->set_allow_intr = 1;
1090 /* The set takes over the caller's request reference */
1091 list_add_tail(&req->rq_set_chain, &set->set_requests);
1093 atomic_inc(&set->set_remaining);
1094 req->rq_queued_time = ktime_get_seconds();
1096 if (req->rq_reqmsg != NULL)
1097 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1099 if (set->set_producer != NULL)
1100 /* If the request set has a producer callback, the RPC must be
1101 * sent straight away */
1102 ptlrpc_send_new_req(req);
1104 EXPORT_SYMBOL(ptlrpc_set_add_req);
1107 * Add a request to a request with dedicated server thread
1108 * and wake the thread to make any necessary processing.
1109 * Currently only used for ptlrpcd.
1111 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1112 struct ptlrpc_request *req)
1114 struct ptlrpc_request_set *set = pc->pc_set;
1117 LASSERT(req->rq_set == NULL);
1118 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1120 spin_lock(&set->set_new_req_lock);
1122 * The set takes over the caller's request reference.
1125 req->rq_queued_time = ktime_get_seconds();
1126 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1127 count = atomic_inc_return(&set->set_new_count);
1128 spin_unlock(&set->set_new_req_lock);
1130 /* Only need to call wakeup once for the first entry. */
1132 wake_up(&set->set_waitq);
1134 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1135 * guarantee the async RPC can be processed ASAP, we have
1136 * no other better choice. It maybe fixed in future. */
1137 for (i = 0; i < pc->pc_npartners; i++)
1138 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1143 * Based on the current state of the import, determine if the request
1144 * can be sent, is an error, or should be delayed.
1146 * Returns true if this request should be delayed. If false, and
1147 * *status is set, then the request can not be sent and *status is the
1148 * error code. If false and status is 0, then request can be sent.
1150 * The imp->imp_lock must be held.
1152 static int ptlrpc_import_delay_req(struct obd_import *imp,
1153 struct ptlrpc_request *req, int *status)
1158 LASSERT (status != NULL);
1161 if (req->rq_ctx_init || req->rq_ctx_fini) {
1162 /* always allow ctx init/fini rpc go through */
1163 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1164 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1166 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1167 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1169 /* pings or MDS-equivalent STATFS may safely race with umount */
1170 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1171 D_HA : D_ERROR, req, "IMP_CLOSED ");
1173 } else if (ptlrpc_send_limit_expired(req)) {
1174 /* probably doesn't need to be a D_ERROR after initial testing*/
1175 DEBUG_REQ(D_HA, req, "send limit expired ");
1176 *status = -ETIMEDOUT;
1177 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1178 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1179 /* allow CONNECT even if import is invalid */ ;
1180 if (atomic_read(&imp->imp_inval_count) != 0) {
1181 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1184 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1185 if (!imp->imp_deactive)
1186 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1187 *status = -ESHUTDOWN; /* bz 12940 */
1188 } else if (req->rq_import_generation != imp->imp_generation) {
1189 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1191 } else if (req->rq_send_state != imp->imp_state) {
1192 /* invalidate in progress - any requests should be drop */
1193 if (atomic_read(&imp->imp_inval_count) != 0) {
1194 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1196 } else if (req->rq_no_delay &&
1197 imp->imp_generation != imp->imp_initiated_at) {
1198 /* ignore nodelay for requests initiating connections */
1199 *status = -EWOULDBLOCK;
1200 } else if (req->rq_allow_replay &&
1201 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1202 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1203 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1204 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1205 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1215 * Decide if the error message should be printed to the console or not.
1216 * Makes its decision based on request type, status, and failure frequency.
1218 * \param[in] req request that failed and may need a console message
1220 * \retval false if no message should be printed
1221 * \retval true if console message should be printed
1223 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1225 LASSERT(req->rq_reqmsg != NULL);
1227 /* Suppress particular reconnect errors which are to be expected. */
1228 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1230 /* Suppress timed out reconnect requests */
1231 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1235 /* Suppress most unavailable/again reconnect requests, but
1236 * print occasionally so it is clear client is trying to
1237 * connect to a server where no target is running. */
1238 if ((err == -ENODEV || err == -EAGAIN) &&
1239 req->rq_import->imp_conn_cnt % 30 != 20)
1243 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1244 /* -EAGAIN is normal when using POSIX flocks */
1247 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1248 (req->rq_xid & 0xf) != 10)
1249 /* Suppress most ping requests, they may fail occasionally */
1256 * Check request processing status.
1257 * Returns the status.
1259 static int ptlrpc_check_status(struct ptlrpc_request *req)
1264 err = lustre_msg_get_status(req->rq_repmsg);
1265 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1266 struct obd_import *imp = req->rq_import;
1267 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1268 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1270 if (ptlrpc_console_allow(req, opc, err))
1271 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1272 "failed: rc = %d\n",
1273 imp->imp_obd->obd_name,
1275 libcfs_nid2str(nid), err);
1276 RETURN(err < 0 ? err : -EINVAL);
1280 DEBUG_REQ(D_INFO, req, "status is %d", err);
1281 } else if (err > 0) {
1282 /* XXX: translate this error from net to host */
1283 DEBUG_REQ(D_INFO, req, "status is %d", err);
1290 * save pre-versions of objects into request for replay.
1291 * Versions are obtained from server reply.
1294 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1296 struct lustre_msg *repmsg = req->rq_repmsg;
1297 struct lustre_msg *reqmsg = req->rq_reqmsg;
1298 __u64 *versions = lustre_msg_get_versions(repmsg);
1301 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1305 lustre_msg_set_versions(reqmsg, versions);
1306 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1307 versions[0], versions[1]);
1312 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1314 struct ptlrpc_request *req;
1316 assert_spin_locked(&imp->imp_lock);
1317 if (list_empty(&imp->imp_unreplied_list))
1320 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1322 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1324 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1325 imp->imp_known_replied_xid = req->rq_xid - 1;
1327 return req->rq_xid - 1;
1331 * Callback function called when client receives RPC reply for \a req.
1332 * Returns 0 on success or error code.
1333 * The return alue would be assigned to req->rq_status by the caller
1334 * as request processing status.
1335 * This function also decides if the request needs to be saved for later replay.
1337 static int after_reply(struct ptlrpc_request *req)
1339 struct obd_import *imp = req->rq_import;
1340 struct obd_device *obd = req->rq_import->imp_obd;
1347 LASSERT(obd != NULL);
1348 /* repbuf must be unlinked */
1349 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1351 if (req->rq_reply_truncated) {
1352 if (ptlrpc_no_resend(req)) {
1353 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1354 " expected: %d, actual size: %d",
1355 req->rq_nob_received, req->rq_repbuf_len);
1359 sptlrpc_cli_free_repbuf(req);
1360 /* Pass the required reply buffer size (include
1361 * space for early reply).
1362 * NB: no need to roundup because alloc_repbuf
1363 * will roundup it */
1364 req->rq_replen = req->rq_nob_received;
1365 req->rq_nob_received = 0;
1366 spin_lock(&req->rq_lock);
1368 spin_unlock(&req->rq_lock);
1372 work_start = ktime_get_real();
1373 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1376 * NB Until this point, the whole of the incoming message,
1377 * including buflens, status etc is in the sender's byte order.
1379 rc = sptlrpc_cli_unwrap_reply(req);
1381 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1386 * Security layer unwrap might ask resend this request.
1391 rc = unpack_reply(req);
1395 /* retry indefinitely on EINPROGRESS */
1396 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1397 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1398 time64_t now = ktime_get_real_seconds();
1400 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1401 spin_lock(&req->rq_lock);
1403 spin_unlock(&req->rq_lock);
1404 req->rq_nr_resend++;
1406 /* Readjust the timeout for current conditions */
1407 ptlrpc_at_set_req_timeout(req);
1408 /* delay resend to give a chance to the server to get ready.
1409 * The delay is increased by 1s on every resend and is capped to
1410 * the current request timeout (i.e. obd_timeout if AT is off,
1411 * or AT service time x 125% + 5s, see at_est2timeout) */
1412 if (req->rq_nr_resend > req->rq_timeout)
1413 req->rq_sent = now + req->rq_timeout;
1415 req->rq_sent = now + req->rq_nr_resend;
1417 /* Resend for EINPROGRESS will use a new XID */
1418 spin_lock(&imp->imp_lock);
1419 list_del_init(&req->rq_unreplied_list);
1420 spin_unlock(&imp->imp_lock);
1425 if (obd->obd_svc_stats != NULL) {
1426 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1428 ptlrpc_lprocfs_rpc_sent(req, timediff);
1431 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1432 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1433 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1434 lustre_msg_get_type(req->rq_repmsg));
1438 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1439 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1440 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1441 ptlrpc_at_adj_net_latency(req,
1442 lustre_msg_get_service_time(req->rq_repmsg));
1444 rc = ptlrpc_check_status(req);
1445 imp->imp_connect_error = rc;
1449 * Either we've been evicted, or the server has failed for
1450 * some reason. Try to reconnect, and if that fails, punt to
1453 if (ptlrpc_recoverable_error(rc)) {
1454 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1455 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1458 ptlrpc_request_handle_notconn(req);
1463 * Let's look if server sent slv. Do it only for RPC with
1466 ldlm_cli_update_pool(req);
1470 * Store transno in reqmsg for replay.
1472 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1473 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1474 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1477 if (imp->imp_replayable) {
1478 spin_lock(&imp->imp_lock);
1480 * No point in adding already-committed requests to the replay
1481 * list, we will just remove them immediately. b=9829
1483 if (req->rq_transno != 0 &&
1485 lustre_msg_get_last_committed(req->rq_repmsg) ||
1487 /** version recovery */
1488 ptlrpc_save_versions(req);
1489 ptlrpc_retain_replayable_request(req, imp);
1490 } else if (req->rq_commit_cb != NULL &&
1491 list_empty(&req->rq_replay_list)) {
1492 /* NB: don't call rq_commit_cb if it's already on
1493 * rq_replay_list, ptlrpc_free_committed() will call
1494 * it later, see LU-3618 for details */
1495 spin_unlock(&imp->imp_lock);
1496 req->rq_commit_cb(req);
1497 spin_lock(&imp->imp_lock);
1501 * Replay-enabled imports return commit-status information.
1503 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1504 if (likely(committed > imp->imp_peer_committed_transno))
1505 imp->imp_peer_committed_transno = committed;
1507 ptlrpc_free_committed(imp);
1509 if (!list_empty(&imp->imp_replay_list)) {
1510 struct ptlrpc_request *last;
1512 last = list_entry(imp->imp_replay_list.prev,
1513 struct ptlrpc_request,
1516 * Requests with rq_replay stay on the list even if no
1517 * commit is expected.
1519 if (last->rq_transno > imp->imp_peer_committed_transno)
1520 ptlrpc_pinger_commit_expected(imp);
1523 spin_unlock(&imp->imp_lock);
1530 * Helper function to send request \a req over the network for the first time
1531 * Also adjusts request phase.
1532 * Returns 0 on success or error code.
1534 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1536 struct obd_import *imp = req->rq_import;
1541 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1543 /* do not try to go further if there is not enough memory in enc_pool */
1544 if (req->rq_sent && req->rq_bulk != NULL)
1545 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1546 pool_is_at_full_capacity())
1549 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1550 (!req->rq_generation_set ||
1551 req->rq_import_generation == imp->imp_generation))
1554 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1556 spin_lock(&imp->imp_lock);
1558 LASSERT(req->rq_xid != 0);
1559 LASSERT(!list_empty(&req->rq_unreplied_list));
1561 if (!req->rq_generation_set)
1562 req->rq_import_generation = imp->imp_generation;
1564 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1565 spin_lock(&req->rq_lock);
1566 req->rq_waiting = 1;
1567 spin_unlock(&req->rq_lock);
1569 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1570 ptlrpc_import_state_name(req->rq_send_state),
1571 ptlrpc_import_state_name(imp->imp_state));
1572 LASSERT(list_empty(&req->rq_list));
1573 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1574 atomic_inc(&req->rq_import->imp_inflight);
1575 spin_unlock(&imp->imp_lock);
1580 spin_unlock(&imp->imp_lock);
1581 req->rq_status = rc;
1582 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1586 LASSERT(list_empty(&req->rq_list));
1587 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1588 atomic_inc(&req->rq_import->imp_inflight);
1590 /* find the known replied XID from the unreplied list, CONNECT
1591 * and DISCONNECT requests are skipped to make the sanity check
1592 * on server side happy. see process_req_last_xid().
1594 * For CONNECT: Because replay requests have lower XID, it'll
1595 * break the sanity check if CONNECT bump the exp_last_xid on
1598 * For DISCONNECT: Since client will abort inflight RPC before
1599 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1600 * than the inflight RPC.
1602 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1603 min_xid = ptlrpc_known_replied_xid(imp);
1604 spin_unlock(&imp->imp_lock);
1606 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1608 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1610 rc = sptlrpc_req_refresh_ctx(req, -1);
1613 req->rq_status = rc;
1616 spin_lock(&req->rq_lock);
1617 req->rq_wait_ctx = 1;
1618 spin_unlock(&req->rq_lock);
1623 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1624 " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1625 imp->imp_obd->obd_uuid.uuid,
1626 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1627 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg));
1629 rc = ptl_send_rpc(req, 0);
1630 if (rc == -ENOMEM) {
1631 spin_lock(&imp->imp_lock);
1632 if (!list_empty(&req->rq_list)) {
1633 list_del_init(&req->rq_list);
1634 atomic_dec(&req->rq_import->imp_inflight);
1636 spin_unlock(&imp->imp_lock);
1637 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1641 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1642 spin_lock(&req->rq_lock);
1643 req->rq_net_err = 1;
1644 spin_unlock(&req->rq_lock);
1650 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1655 LASSERT(set->set_producer != NULL);
1657 remaining = atomic_read(&set->set_remaining);
1659 /* populate the ->set_requests list with requests until we
1660 * reach the maximum number of RPCs in flight for this set */
1661 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1662 rc = set->set_producer(set, set->set_producer_arg);
1663 if (rc == -ENOENT) {
1664 /* no more RPC to produce */
1665 set->set_producer = NULL;
1666 set->set_producer_arg = NULL;
1671 RETURN((atomic_read(&set->set_remaining) - remaining));
1675 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1676 * and no more replies are expected.
1677 * (it is possible to get less replies than requests sent e.g. due to timed out
1678 * requests or requests that we had trouble to send out)
1680 * NOTE: This function contains a potential schedule point (cond_resched()).
1682 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1684 struct list_head *tmp, *next;
1685 struct list_head comp_reqs;
1686 int force_timer_recalc = 0;
1689 if (atomic_read(&set->set_remaining) == 0)
1692 INIT_LIST_HEAD(&comp_reqs);
1693 list_for_each_safe(tmp, next, &set->set_requests) {
1694 struct ptlrpc_request *req =
1695 list_entry(tmp, struct ptlrpc_request,
1697 struct obd_import *imp = req->rq_import;
1698 int unregistered = 0;
1702 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1703 list_move_tail(&req->rq_set_chain, &comp_reqs);
1707 /* This schedule point is mainly for the ptlrpcd caller of this
1708 * function. Most ptlrpc sets are not long-lived and unbounded
1709 * in length, but at the least the set used by the ptlrpcd is.
1710 * Since the processing time is unbounded, we need to insert an
1711 * explicit schedule point to make the thread well-behaved.
1715 /* If the caller requires to allow to be interpreted by force
1716 * and it has really been interpreted, then move the request
1717 * to RQ_PHASE_INTERPRET phase in spite of what the current
1719 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1720 req->rq_status = -EINTR;
1721 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1723 /* Since it is interpreted and we have to wait for
1724 * the reply to be unlinked, then use sync mode. */
1727 GOTO(interpret, req->rq_status);
1730 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1731 force_timer_recalc = 1;
1733 /* delayed send - skip */
1734 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1737 /* delayed resend - skip */
1738 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1739 req->rq_sent > ktime_get_real_seconds())
1742 if (!(req->rq_phase == RQ_PHASE_RPC ||
1743 req->rq_phase == RQ_PHASE_BULK ||
1744 req->rq_phase == RQ_PHASE_INTERPRET ||
1745 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1746 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1747 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1751 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1752 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1753 LASSERT(req->rq_next_phase != req->rq_phase);
1754 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1756 if (req->rq_req_deadline &&
1757 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1758 req->rq_req_deadline = 0;
1759 if (req->rq_reply_deadline &&
1760 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1761 req->rq_reply_deadline = 0;
1762 if (req->rq_bulk_deadline &&
1763 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1764 req->rq_bulk_deadline = 0;
1767 * Skip processing until reply is unlinked. We
1768 * can't return to pool before that and we can't
1769 * call interpret before that. We need to make
1770 * sure that all rdma transfers finished and will
1771 * not corrupt any data.
1773 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1774 ptlrpc_client_recv_or_unlink(req))
1776 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1777 ptlrpc_client_bulk_active(req))
1781 * Turn fail_loc off to prevent it from looping
1784 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1785 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1788 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1789 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1794 * Move to next phase if reply was successfully
1797 ptlrpc_rqphase_move(req, req->rq_next_phase);
1800 if (req->rq_phase == RQ_PHASE_INTERPRET)
1801 GOTO(interpret, req->rq_status);
1804 * Note that this also will start async reply unlink.
1806 if (req->rq_net_err && !req->rq_timedout) {
1807 ptlrpc_expire_one_request(req, 1);
1810 * Check if we still need to wait for unlink.
1812 if (ptlrpc_client_recv_or_unlink(req) ||
1813 ptlrpc_client_bulk_active(req))
1815 /* If there is no need to resend, fail it now. */
1816 if (req->rq_no_resend) {
1817 if (req->rq_status == 0)
1818 req->rq_status = -EIO;
1819 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1820 GOTO(interpret, req->rq_status);
1827 spin_lock(&req->rq_lock);
1828 req->rq_replied = 0;
1829 spin_unlock(&req->rq_lock);
1830 if (req->rq_status == 0)
1831 req->rq_status = -EIO;
1832 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1833 GOTO(interpret, req->rq_status);
1836 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1837 * so it sets rq_intr regardless of individual rpc
1838 * timeouts. The synchronous IO waiting path sets
1839 * rq_intr irrespective of whether ptlrpcd
1840 * has seen a timeout. Our policy is to only interpret
1841 * interrupted rpcs after they have timed out, so we
1842 * need to enforce that here.
1845 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1846 req->rq_wait_ctx)) {
1847 req->rq_status = -EINTR;
1848 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1849 GOTO(interpret, req->rq_status);
1852 if (req->rq_phase == RQ_PHASE_RPC) {
1853 if (req->rq_timedout || req->rq_resend ||
1854 req->rq_waiting || req->rq_wait_ctx) {
1857 if (!ptlrpc_unregister_reply(req, 1)) {
1858 ptlrpc_unregister_bulk(req, 1);
1862 spin_lock(&imp->imp_lock);
1863 if (ptlrpc_import_delay_req(imp, req, &status)){
1864 /* put on delay list - only if we wait
1865 * recovery finished - before send */
1866 list_del_init(&req->rq_list);
1867 list_add_tail(&req->rq_list,
1870 spin_unlock(&imp->imp_lock);
1875 req->rq_status = status;
1876 ptlrpc_rqphase_move(req,
1877 RQ_PHASE_INTERPRET);
1878 spin_unlock(&imp->imp_lock);
1879 GOTO(interpret, req->rq_status);
1881 /* ignore on just initiated connections */
1882 if (ptlrpc_no_resend(req) &&
1883 !req->rq_wait_ctx &&
1884 imp->imp_generation !=
1885 imp->imp_initiated_at) {
1886 req->rq_status = -ENOTCONN;
1887 ptlrpc_rqphase_move(req,
1888 RQ_PHASE_INTERPRET);
1889 spin_unlock(&imp->imp_lock);
1890 GOTO(interpret, req->rq_status);
1893 list_del_init(&req->rq_list);
1894 list_add_tail(&req->rq_list,
1895 &imp->imp_sending_list);
1897 spin_unlock(&imp->imp_lock);
1899 spin_lock(&req->rq_lock);
1900 req->rq_waiting = 0;
1901 spin_unlock(&req->rq_lock);
1903 if (req->rq_timedout || req->rq_resend) {
1904 /* This is re-sending anyways,
1905 * let's mark req as resend. */
1906 spin_lock(&req->rq_lock);
1908 spin_unlock(&req->rq_lock);
1910 if (req->rq_bulk != NULL &&
1911 !ptlrpc_unregister_bulk(req, 1))
1915 * rq_wait_ctx is only touched by ptlrpcd,
1916 * so no lock is needed here.
1918 status = sptlrpc_req_refresh_ctx(req, -1);
1921 req->rq_status = status;
1922 spin_lock(&req->rq_lock);
1923 req->rq_wait_ctx = 0;
1924 spin_unlock(&req->rq_lock);
1925 force_timer_recalc = 1;
1927 spin_lock(&req->rq_lock);
1928 req->rq_wait_ctx = 1;
1929 spin_unlock(&req->rq_lock);
1934 spin_lock(&req->rq_lock);
1935 req->rq_wait_ctx = 0;
1936 spin_unlock(&req->rq_lock);
1939 rc = ptl_send_rpc(req, 0);
1940 if (rc == -ENOMEM) {
1941 spin_lock(&imp->imp_lock);
1942 if (!list_empty(&req->rq_list))
1943 list_del_init(&req->rq_list);
1944 spin_unlock(&imp->imp_lock);
1945 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1949 DEBUG_REQ(D_HA, req,
1950 "send failed: rc = %d", rc);
1951 force_timer_recalc = 1;
1952 spin_lock(&req->rq_lock);
1953 req->rq_net_err = 1;
1954 spin_unlock(&req->rq_lock);
1957 /* need to reset the timeout */
1958 force_timer_recalc = 1;
1961 spin_lock(&req->rq_lock);
1963 if (ptlrpc_client_early(req)) {
1964 ptlrpc_at_recv_early_reply(req);
1965 spin_unlock(&req->rq_lock);
1969 /* Still waiting for a reply? */
1970 if (ptlrpc_client_recv(req)) {
1971 spin_unlock(&req->rq_lock);
1975 /* Did we actually receive a reply? */
1976 if (!ptlrpc_client_replied(req)) {
1977 spin_unlock(&req->rq_lock);
1981 spin_unlock(&req->rq_lock);
1983 /* unlink from net because we are going to
1984 * swab in-place of reply buffer */
1985 unregistered = ptlrpc_unregister_reply(req, 1);
1989 req->rq_status = after_reply(req);
1993 /* If there is no bulk associated with this request,
1994 * then we're done and should let the interpreter
1995 * process the reply. Similarly if the RPC returned
1996 * an error, and therefore the bulk will never arrive.
1998 if (req->rq_bulk == NULL || req->rq_status < 0) {
1999 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2000 GOTO(interpret, req->rq_status);
2003 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2006 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2007 if (ptlrpc_client_bulk_active(req))
2010 if (req->rq_bulk->bd_failure) {
2011 /* The RPC reply arrived OK, but the bulk screwed
2012 * up! Dead weird since the server told us the RPC
2013 * was good after getting the REPLY for her GET or
2014 * the ACK for her PUT. */
2015 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2016 req->rq_status = -EIO;
2019 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2022 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2024 /* This moves to "unregistering" phase we need to wait for
2026 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2027 /* start async bulk unlink too */
2028 ptlrpc_unregister_bulk(req, 1);
2032 if (!ptlrpc_unregister_bulk(req, async))
2035 /* When calling interpret receiving already should be
2037 LASSERT(!req->rq_receiving_reply);
2039 ptlrpc_req_interpret(env, req, req->rq_status);
2041 if (ptlrpcd_check_work(req)) {
2042 atomic_dec(&set->set_remaining);
2045 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2047 if (req->rq_reqmsg != NULL)
2049 "Completed RPC pname:cluuid:pid:xid:nid:"
2050 "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
2051 imp->imp_obd->obd_uuid.uuid,
2052 lustre_msg_get_status(req->rq_reqmsg),
2054 obd_import_nid2str(imp),
2055 lustre_msg_get_opc(req->rq_reqmsg));
2057 spin_lock(&imp->imp_lock);
2058 /* Request already may be not on sending or delaying list. This
2059 * may happen in the case of marking it erroneous for the case
2060 * ptlrpc_import_delay_req(req, status) find it impossible to
2061 * allow sending this rpc and returns *status != 0. */
2062 if (!list_empty(&req->rq_list)) {
2063 list_del_init(&req->rq_list);
2064 atomic_dec(&imp->imp_inflight);
2066 list_del_init(&req->rq_unreplied_list);
2067 spin_unlock(&imp->imp_lock);
2069 atomic_dec(&set->set_remaining);
2070 wake_up_all(&imp->imp_recovery_waitq);
2072 if (set->set_producer) {
2073 /* produce a new request if possible */
2074 if (ptlrpc_set_producer(set) > 0)
2075 force_timer_recalc = 1;
2077 /* free the request that has just been completed
2078 * in order not to pollute set->set_requests */
2079 list_del_init(&req->rq_set_chain);
2080 spin_lock(&req->rq_lock);
2082 req->rq_invalid_rqset = 0;
2083 spin_unlock(&req->rq_lock);
2085 /* record rq_status to compute the final status later */
2086 if (req->rq_status != 0)
2087 set->set_rc = req->rq_status;
2088 ptlrpc_req_finished(req);
2090 list_move_tail(&req->rq_set_chain, &comp_reqs);
2094 /* move completed request at the head of list so it's easier for
2095 * caller to find them */
2096 list_splice(&comp_reqs, &set->set_requests);
2098 /* If we hit an error, we want to recover promptly. */
2099 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2101 EXPORT_SYMBOL(ptlrpc_check_set);
2104 * Time out request \a req. is \a async_unlink is set, that means do not wait
2105 * until LNet actually confirms network buffer unlinking.
2106 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2108 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2110 struct obd_import *imp = req->rq_import;
2111 unsigned int debug_mask = D_RPCTRACE;
2115 spin_lock(&req->rq_lock);
2116 req->rq_timedout = 1;
2117 spin_unlock(&req->rq_lock);
2119 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2120 lustre_msg_get_status(req->rq_reqmsg)))
2121 debug_mask = D_WARNING;
2122 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2123 req->rq_net_err ? "failed due to network error" :
2124 ((req->rq_real_sent == 0 ||
2125 req->rq_real_sent < req->rq_sent ||
2126 req->rq_real_sent >= req->rq_deadline) ?
2127 "timed out for sent delay" : "timed out for slow reply"),
2128 (s64)req->rq_sent, (s64)req->rq_real_sent);
2130 if (imp != NULL && obd_debug_peer_on_timeout)
2131 LNetDebugPeer(imp->imp_connection->c_peer);
2133 ptlrpc_unregister_reply(req, async_unlink);
2134 ptlrpc_unregister_bulk(req, async_unlink);
2136 if (obd_dump_on_timeout)
2137 libcfs_debug_dumplog();
2140 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2144 atomic_inc(&imp->imp_timeouts);
2146 /* The DLM server doesn't want recovery run on its imports. */
2147 if (imp->imp_dlm_fake)
2150 /* If this request is for recovery or other primordial tasks,
2151 * then error it out here. */
2152 if (req->rq_ctx_init || req->rq_ctx_fini ||
2153 req->rq_send_state != LUSTRE_IMP_FULL ||
2154 imp->imp_obd->obd_no_recov) {
2155 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2156 ptlrpc_import_state_name(req->rq_send_state),
2157 ptlrpc_import_state_name(imp->imp_state));
2158 spin_lock(&req->rq_lock);
2159 req->rq_status = -ETIMEDOUT;
2161 spin_unlock(&req->rq_lock);
2165 /* if a request can't be resent we can't wait for an answer after
2167 if (ptlrpc_no_resend(req)) {
2168 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2172 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2178 * Time out all uncompleted requests in request set pointed by \a data
2179 * Callback used when waiting on sets with l_wait_event.
2182 int ptlrpc_expired_set(void *data)
2184 struct ptlrpc_request_set *set = data;
2185 struct list_head *tmp;
2186 time64_t now = ktime_get_real_seconds();
2189 LASSERT(set != NULL);
2192 * A timeout expired. See which reqs it applies to...
2194 list_for_each(tmp, &set->set_requests) {
2195 struct ptlrpc_request *req =
2196 list_entry(tmp, struct ptlrpc_request,
2199 /* don't expire request waiting for context */
2200 if (req->rq_wait_ctx)
2203 /* Request in-flight? */
2204 if (!((req->rq_phase == RQ_PHASE_RPC &&
2205 !req->rq_waiting && !req->rq_resend) ||
2206 (req->rq_phase == RQ_PHASE_BULK)))
2209 if (req->rq_timedout || /* already dealt with */
2210 req->rq_deadline > now) /* not expired */
2213 /* Deal with this guy. Do it asynchronously to not block
2214 * ptlrpcd thread. */
2215 ptlrpc_expire_one_request(req, 1);
2219 * When waiting for a whole set, we always break out of the
2220 * sleep so we can recalculate the timeout, or enable interrupts
2221 * if everyone's timed out.
2227 * Sets rq_intr flag in \a req under spinlock.
2229 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2231 spin_lock(&req->rq_lock);
2233 spin_unlock(&req->rq_lock);
2235 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2238 * Interrupts (sets interrupted flag) all uncompleted requests in
2239 * a set \a data. Callback for l_wait_event for interruptible waits.
2241 static void ptlrpc_interrupted_set(void *data)
2243 struct ptlrpc_request_set *set = data;
2244 struct list_head *tmp;
2246 LASSERT(set != NULL);
2247 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2249 list_for_each(tmp, &set->set_requests) {
2250 struct ptlrpc_request *req =
2251 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2256 if (req->rq_phase != RQ_PHASE_RPC &&
2257 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2258 !req->rq_allow_intr)
2261 ptlrpc_mark_interrupted(req);
2266 * Get the smallest timeout in the set; this does NOT set a timeout.
2268 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2270 struct list_head *tmp;
2271 time64_t now = ktime_get_real_seconds();
2273 struct ptlrpc_request *req;
2277 list_for_each(tmp, &set->set_requests) {
2278 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2281 * Request in-flight?
2283 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2284 (req->rq_phase == RQ_PHASE_BULK) ||
2285 (req->rq_phase == RQ_PHASE_NEW)))
2289 * Already timed out.
2291 if (req->rq_timedout)
2297 if (req->rq_wait_ctx)
2300 if (req->rq_phase == RQ_PHASE_NEW)
2301 deadline = req->rq_sent;
2302 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2303 deadline = req->rq_sent;
2305 deadline = req->rq_sent + req->rq_timeout;
2307 if (deadline <= now) /* actually expired already */
2308 timeout = 1; /* ASAP */
2309 else if (timeout == 0 || timeout > deadline - now)
2310 timeout = deadline - now;
2316 * Send all unset request from the set and then wait untill all
2317 * requests in the set complete (either get a reply, timeout, get an
2318 * error or otherwise be interrupted).
2319 * Returns 0 on success or error code otherwise.
2321 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2323 struct list_head *tmp;
2324 struct ptlrpc_request *req;
2325 struct l_wait_info lwi;
2330 if (set->set_producer)
2331 (void)ptlrpc_set_producer(set);
2333 list_for_each(tmp, &set->set_requests) {
2334 req = list_entry(tmp, struct ptlrpc_request,
2336 if (req->rq_phase == RQ_PHASE_NEW)
2337 (void)ptlrpc_send_new_req(req);
2340 if (list_empty(&set->set_requests))
2344 timeout = ptlrpc_set_next_timeout(set);
2346 /* wait until all complete, interrupted, or an in-flight
2348 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2351 if ((timeout == 0 && !signal_pending(current)) ||
2352 set->set_allow_intr)
2353 /* No requests are in-flight (ether timed out
2354 * or delayed), so we can allow interrupts.
2355 * We still want to block for a limited time,
2356 * so we allow interrupts during the timeout. */
2357 lwi = LWI_TIMEOUT_INTR_ALL(
2358 cfs_time_seconds(timeout ? timeout : 1),
2360 ptlrpc_interrupted_set, set);
2363 * At least one request is in flight, so no
2364 * interrupts are allowed. Wait until all
2365 * complete, or an in-flight req times out.
2367 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2368 ptlrpc_expired_set, set);
2370 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2372 /* LU-769 - if we ignored the signal because it was already
2373 * pending when we started, we need to handle it now or we risk
2374 * it being ignored forever */
2375 if (rc == -ETIMEDOUT &&
2376 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2377 signal_pending(current)) {
2378 sigset_t blocked_sigs =
2379 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2381 /* In fact we only interrupt for the "fatal" signals
2382 * like SIGINT or SIGKILL. We still ignore less
2383 * important signals since ptlrpc set is not easily
2384 * reentrant from userspace again */
2385 if (signal_pending(current))
2386 ptlrpc_interrupted_set(set);
2387 cfs_restore_sigs(blocked_sigs);
2390 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2392 /* -EINTR => all requests have been flagged rq_intr so next
2394 * -ETIMEDOUT => someone timed out. When all reqs have
2395 * timed out, signals are enabled allowing completion with
2397 * I don't really care if we go once more round the loop in
2398 * the error cases -eeb. */
2399 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2400 list_for_each(tmp, &set->set_requests) {
2401 req = list_entry(tmp, struct ptlrpc_request,
2403 spin_lock(&req->rq_lock);
2404 req->rq_invalid_rqset = 1;
2405 spin_unlock(&req->rq_lock);
2408 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2410 LASSERT(atomic_read(&set->set_remaining) == 0);
2412 rc = set->set_rc; /* rq_status of already freed requests if any */
2413 list_for_each(tmp, &set->set_requests) {
2414 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2416 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2417 if (req->rq_status != 0)
2418 rc = req->rq_status;
2423 EXPORT_SYMBOL(ptlrpc_set_wait);
2426 * Helper fuction for request freeing.
2427 * Called when request count reached zero and request needs to be freed.
2428 * Removes request from all sorts of sending/replay lists it might be on,
2429 * frees network buffers if any are present.
2430 * If \a locked is set, that means caller is already holding import imp_lock
2431 * and so we no longer need to reobtain it (for certain lists manipulations)
2433 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2437 if (request == NULL)
2440 LASSERT(!request->rq_srv_req);
2441 LASSERT(request->rq_export == NULL);
2442 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2443 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2444 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2445 LASSERTF(!request->rq_replay, "req %p\n", request);
2447 req_capsule_fini(&request->rq_pill);
2449 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2450 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2451 if (request->rq_import != NULL) {
2453 spin_lock(&request->rq_import->imp_lock);
2454 list_del_init(&request->rq_replay_list);
2455 list_del_init(&request->rq_unreplied_list);
2457 spin_unlock(&request->rq_import->imp_lock);
2459 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2461 if (atomic_read(&request->rq_refcount) != 0) {
2462 DEBUG_REQ(D_ERROR, request,
2463 "freeing request with nonzero refcount");
2467 if (request->rq_repbuf != NULL)
2468 sptlrpc_cli_free_repbuf(request);
2470 if (request->rq_import != NULL) {
2471 class_import_put(request->rq_import);
2472 request->rq_import = NULL;
2474 if (request->rq_bulk != NULL)
2475 ptlrpc_free_bulk(request->rq_bulk);
2477 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2478 sptlrpc_cli_free_reqbuf(request);
2480 if (request->rq_cli_ctx)
2481 sptlrpc_req_put_ctx(request, !locked);
2483 if (request->rq_pool)
2484 __ptlrpc_free_req_to_pool(request);
2486 ptlrpc_request_cache_free(request);
2490 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2492 * Drop one request reference. Must be called with import imp_lock held.
2493 * When reference count drops to zero, request is freed.
2495 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2497 assert_spin_locked(&request->rq_import->imp_lock);
2498 (void)__ptlrpc_req_finished(request, 1);
2503 * Drops one reference count for request \a request.
2504 * \a locked set indicates that caller holds import imp_lock.
2505 * Frees the request whe reference count reaches zero.
2507 * \retval 1 the request is freed
2508 * \retval 0 some others still hold references on the request
2510 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2518 LASSERT(request != LP_POISON);
2519 LASSERT(request->rq_reqmsg != LP_POISON);
2521 DEBUG_REQ(D_INFO, request, "refcount now %u",
2522 atomic_read(&request->rq_refcount) - 1);
2524 spin_lock(&request->rq_lock);
2525 count = atomic_dec_return(&request->rq_refcount);
2526 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2528 /* For open RPC, the client does not know the EA size (LOV, ACL, and
2529 * so on) before replied, then the client has to reserve very large
2530 * reply buffer. Such buffer will not be released until the RPC freed.
2531 * Since The open RPC is replayable, we need to keep it in the replay
2532 * list until close. If there are a lot of files opened concurrently,
2533 * then the client may be OOM.
2535 * If fact, it is unnecessary to keep reply buffer for open replay,
2536 * related EAs have already been saved via mdc_save_lovea() before
2537 * coming here. So it is safe to free the reply buffer some earlier
2538 * before releasing the RPC to avoid client OOM. LU-9514 */
2539 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2540 spin_lock(&request->rq_early_free_lock);
2541 sptlrpc_cli_free_repbuf(request);
2542 request->rq_repbuf = NULL;
2543 request->rq_repbuf_len = 0;
2544 request->rq_repdata = NULL;
2545 request->rq_reqdata_len = 0;
2546 spin_unlock(&request->rq_early_free_lock);
2548 spin_unlock(&request->rq_lock);
2551 __ptlrpc_free_req(request, locked);
2557 * Drops one reference count for a request.
2559 void ptlrpc_req_finished(struct ptlrpc_request *request)
2561 __ptlrpc_req_finished(request, 0);
2563 EXPORT_SYMBOL(ptlrpc_req_finished);
2566 * Returns xid of a \a request
2568 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2570 return request->rq_xid;
2572 EXPORT_SYMBOL(ptlrpc_req_xid);
2575 * Disengage the client's reply buffer from the network
2576 * NB does _NOT_ unregister any client-side bulk.
2577 * IDEMPOTENT, but _not_ safe against concurrent callers.
2578 * The request owner (i.e. the thread doing the I/O) must call...
2579 * Returns 0 on success or 1 if unregistering cannot be made.
2581 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2584 struct l_wait_info lwi;
2589 LASSERT(!in_interrupt());
2591 /* Let's setup deadline for reply unlink. */
2592 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2593 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2594 request->rq_reply_deadline = ktime_get_real_seconds() +
2598 * Nothing left to do.
2600 if (!ptlrpc_client_recv_or_unlink(request))
2603 LNetMDUnlink(request->rq_reply_md_h);
2606 * Let's check it once again.
2608 if (!ptlrpc_client_recv_or_unlink(request))
2611 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2612 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2615 * Do not wait for unlink to finish.
2621 * We have to l_wait_event() whatever the result, to give liblustre
2622 * a chance to run reply_in_callback(), and to make sure we've
2623 * unlinked before returning a req to the pool.
2626 /* The wq argument is ignored by user-space wait_event macros */
2627 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2628 &request->rq_set->set_waitq :
2629 &request->rq_reply_waitq;
2630 /* Network access will complete in finite time but the HUGE
2631 * timeout lets us CWARN for visibility of sluggish NALs */
2632 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2633 cfs_time_seconds(1), NULL, NULL);
2634 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2637 ptlrpc_rqphase_move(request, request->rq_next_phase);
2641 LASSERT(rc == -ETIMEDOUT);
2642 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2643 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2644 request->rq_receiving_reply,
2645 request->rq_req_unlinked,
2646 request->rq_reply_unlinked);
2651 static void ptlrpc_free_request(struct ptlrpc_request *req)
2653 spin_lock(&req->rq_lock);
2655 spin_unlock(&req->rq_lock);
2657 if (req->rq_commit_cb != NULL)
2658 req->rq_commit_cb(req);
2659 list_del_init(&req->rq_replay_list);
2661 __ptlrpc_req_finished(req, 1);
2665 * the request is committed and dropped from the replay list of its import
2667 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2669 struct obd_import *imp = req->rq_import;
2671 spin_lock(&imp->imp_lock);
2672 if (list_empty(&req->rq_replay_list)) {
2673 spin_unlock(&imp->imp_lock);
2677 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2678 if (imp->imp_replay_cursor == &req->rq_replay_list)
2679 imp->imp_replay_cursor = req->rq_replay_list.next;
2680 ptlrpc_free_request(req);
2683 spin_unlock(&imp->imp_lock);
2685 EXPORT_SYMBOL(ptlrpc_request_committed);
2688 * Iterates through replay_list on import and prunes
2689 * all requests have transno smaller than last_committed for the
2690 * import and don't have rq_replay set.
2691 * Since requests are sorted in transno order, stops when meetign first
2692 * transno bigger than last_committed.
2693 * caller must hold imp->imp_lock
2695 void ptlrpc_free_committed(struct obd_import *imp)
2697 struct ptlrpc_request *req, *saved;
2698 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2699 bool skip_committed_list = true;
2702 LASSERT(imp != NULL);
2703 assert_spin_locked(&imp->imp_lock);
2705 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2706 imp->imp_generation == imp->imp_last_generation_checked) {
2707 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2708 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2711 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2712 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2713 imp->imp_generation);
2715 if (imp->imp_generation != imp->imp_last_generation_checked ||
2716 imp->imp_last_transno_checked == 0)
2717 skip_committed_list = false;
2719 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2720 imp->imp_last_generation_checked = imp->imp_generation;
2722 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2724 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2725 LASSERT(req != last_req);
2728 if (req->rq_transno == 0) {
2729 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2732 if (req->rq_import_generation < imp->imp_generation) {
2733 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2737 /* not yet committed */
2738 if (req->rq_transno > imp->imp_peer_committed_transno) {
2739 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2743 if (req->rq_replay) {
2744 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2745 list_move_tail(&req->rq_replay_list,
2746 &imp->imp_committed_list);
2750 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2751 imp->imp_peer_committed_transno);
2753 ptlrpc_free_request(req);
2756 if (skip_committed_list)
2759 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2761 LASSERT(req->rq_transno != 0);
2762 if (req->rq_import_generation < imp->imp_generation ||
2764 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2765 req->rq_import_generation <
2766 imp->imp_generation ? "stale" : "closed");
2768 if (imp->imp_replay_cursor == &req->rq_replay_list)
2769 imp->imp_replay_cursor =
2770 req->rq_replay_list.next;
2772 ptlrpc_free_request(req);
2779 void ptlrpc_cleanup_client(struct obd_import *imp)
2786 * Schedule previously sent request for resend.
2787 * For bulk requests we assign new xid (to avoid problems with
2788 * lost replies and therefore several transfers landing into same buffer
2789 * from different sending attempts).
2791 void ptlrpc_resend_req(struct ptlrpc_request *req)
2793 DEBUG_REQ(D_HA, req, "going to resend");
2794 spin_lock(&req->rq_lock);
2796 /* Request got reply but linked to the import list still.
2797 Let ptlrpc_check_set() to process it. */
2798 if (ptlrpc_client_replied(req)) {
2799 spin_unlock(&req->rq_lock);
2800 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2804 req->rq_status = -EAGAIN;
2807 req->rq_net_err = 0;
2808 req->rq_timedout = 0;
2810 ptlrpc_client_wake_req(req);
2811 spin_unlock(&req->rq_lock);
2814 /* XXX: this function and rq_status are currently unused */
2815 void ptlrpc_restart_req(struct ptlrpc_request *req)
2817 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2818 req->rq_status = -ERESTARTSYS;
2820 spin_lock(&req->rq_lock);
2821 req->rq_restart = 1;
2822 req->rq_timedout = 0;
2823 ptlrpc_client_wake_req(req);
2824 spin_unlock(&req->rq_lock);
2828 * Grab additional reference on a request \a req
2830 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2833 atomic_inc(&req->rq_refcount);
2836 EXPORT_SYMBOL(ptlrpc_request_addref);
2839 * Add a request to import replay_list.
2840 * Must be called under imp_lock
2842 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2843 struct obd_import *imp)
2845 struct list_head *tmp;
2847 assert_spin_locked(&imp->imp_lock);
2849 if (req->rq_transno == 0) {
2850 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2854 /* clear this for new requests that were resent as well
2855 as resent replayed requests. */
2856 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2858 /* don't re-add requests that have been replayed */
2859 if (!list_empty(&req->rq_replay_list))
2862 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2864 spin_lock(&req->rq_lock);
2866 spin_unlock(&req->rq_lock);
2868 LASSERT(imp->imp_replayable);
2869 /* Balanced in ptlrpc_free_committed, usually. */
2870 ptlrpc_request_addref(req);
2871 list_for_each_prev(tmp, &imp->imp_replay_list) {
2872 struct ptlrpc_request *iter = list_entry(tmp,
2873 struct ptlrpc_request,
2876 /* We may have duplicate transnos if we create and then
2877 * open a file, or for closes retained if to match creating
2878 * opens, so use req->rq_xid as a secondary key.
2879 * (See bugs 684, 685, and 428.)
2880 * XXX no longer needed, but all opens need transnos!
2882 if (iter->rq_transno > req->rq_transno)
2885 if (iter->rq_transno == req->rq_transno) {
2886 LASSERT(iter->rq_xid != req->rq_xid);
2887 if (iter->rq_xid > req->rq_xid)
2891 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2895 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2899 * Send request and wait until it completes.
2900 * Returns request processing status.
2902 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2904 struct ptlrpc_request_set *set;
2908 LASSERT(req->rq_set == NULL);
2909 LASSERT(!req->rq_receiving_reply);
2911 set = ptlrpc_prep_set();
2913 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2917 /* for distributed debugging */
2918 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2920 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2921 ptlrpc_request_addref(req);
2922 ptlrpc_set_add_req(set, req);
2923 rc = ptlrpc_set_wait(set);
2924 ptlrpc_set_destroy(set);
2928 EXPORT_SYMBOL(ptlrpc_queue_wait);
2931 * Callback used for replayed requests reply processing.
2932 * In case of successful reply calls registered request replay callback.
2933 * In case of error restart replay process.
2935 static int ptlrpc_replay_interpret(const struct lu_env *env,
2936 struct ptlrpc_request *req,
2937 void * data, int rc)
2939 struct ptlrpc_replay_async_args *aa = data;
2940 struct obd_import *imp = req->rq_import;
2943 atomic_dec(&imp->imp_replay_inflight);
2945 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2946 * server got the request, but bulk transfer timeout, let's
2947 * replay the bulk req again */
2948 if (!ptlrpc_client_replied(req) ||
2949 (req->rq_bulk != NULL &&
2950 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2951 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2952 GOTO(out, rc = -ETIMEDOUT);
2955 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2956 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2957 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2958 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2960 /** VBR: check version failure */
2961 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2962 /** replay was failed due to version mismatch */
2963 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2964 spin_lock(&imp->imp_lock);
2965 imp->imp_vbr_failed = 1;
2966 spin_unlock(&imp->imp_lock);
2967 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2969 /** The transno had better not change over replay. */
2970 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2971 lustre_msg_get_transno(req->rq_repmsg) ||
2972 lustre_msg_get_transno(req->rq_repmsg) == 0,
2974 lustre_msg_get_transno(req->rq_reqmsg),
2975 lustre_msg_get_transno(req->rq_repmsg));
2978 spin_lock(&imp->imp_lock);
2979 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2980 spin_unlock(&imp->imp_lock);
2981 LASSERT(imp->imp_last_replay_transno);
2983 /* transaction number shouldn't be bigger than the latest replayed */
2984 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2985 DEBUG_REQ(D_ERROR, req,
2986 "Reported transno %llu is bigger than the "
2987 "replayed one: %llu", req->rq_transno,
2988 lustre_msg_get_transno(req->rq_reqmsg));
2989 GOTO(out, rc = -EINVAL);
2992 DEBUG_REQ(D_HA, req, "got rep");
2994 /* let the callback do fixups, possibly including in the request */
2995 if (req->rq_replay_cb)
2996 req->rq_replay_cb(req);
2998 if (ptlrpc_client_replied(req) &&
2999 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3000 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3001 lustre_msg_get_status(req->rq_repmsg),
3002 aa->praa_old_status);
3004 /* Note: If the replay fails for MDT-MDT recovery, let's
3005 * abort all of the following requests in the replay
3006 * and sending list, because MDT-MDT update requests
3007 * are dependent on each other, see LU-7039 */
3008 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3009 struct ptlrpc_request *free_req;
3010 struct ptlrpc_request *tmp;
3012 spin_lock(&imp->imp_lock);
3013 list_for_each_entry_safe(free_req, tmp,
3014 &imp->imp_replay_list,
3016 ptlrpc_free_request(free_req);
3019 list_for_each_entry_safe(free_req, tmp,
3020 &imp->imp_committed_list,
3022 ptlrpc_free_request(free_req);
3025 list_for_each_entry_safe(free_req, tmp,
3026 &imp->imp_delayed_list,
3028 spin_lock(&free_req->rq_lock);
3029 free_req->rq_err = 1;
3030 free_req->rq_status = -EIO;
3031 ptlrpc_client_wake_req(free_req);
3032 spin_unlock(&free_req->rq_lock);
3035 list_for_each_entry_safe(free_req, tmp,
3036 &imp->imp_sending_list,
3038 spin_lock(&free_req->rq_lock);
3039 free_req->rq_err = 1;
3040 free_req->rq_status = -EIO;
3041 ptlrpc_client_wake_req(free_req);
3042 spin_unlock(&free_req->rq_lock);
3044 spin_unlock(&imp->imp_lock);
3047 /* Put it back for re-replay. */
3048 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3052 * Errors while replay can set transno to 0, but
3053 * imp_last_replay_transno shouldn't be set to 0 anyway
3055 if (req->rq_transno == 0)
3056 CERROR("Transno is 0 during replay!\n");
3058 /* continue with recovery */
3059 rc = ptlrpc_import_recovery_state_machine(imp);
3061 req->rq_send_state = aa->praa_old_state;
3064 /* this replay failed, so restart recovery */
3065 ptlrpc_connect_import(imp);
3071 * Prepares and queues request for replay.
3072 * Adds it to ptlrpcd queue for actual sending.
3073 * Returns 0 on success.
3075 int ptlrpc_replay_req(struct ptlrpc_request *req)
3077 struct ptlrpc_replay_async_args *aa;
3081 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3083 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3084 aa = ptlrpc_req_async_args(req);
3085 memset(aa, 0, sizeof(*aa));
3087 /* Prepare request to be resent with ptlrpcd */
3088 aa->praa_old_state = req->rq_send_state;
3089 req->rq_send_state = LUSTRE_IMP_REPLAY;
3090 req->rq_phase = RQ_PHASE_NEW;
3091 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3093 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3095 req->rq_interpret_reply = ptlrpc_replay_interpret;
3096 /* Readjust the timeout for current conditions */
3097 ptlrpc_at_set_req_timeout(req);
3099 /* Tell server the net_latency, so the server can calculate how long
3100 * it should wait for next replay */
3101 lustre_msg_set_service_time(req->rq_reqmsg,
3102 ptlrpc_at_get_net_latency(req));
3103 DEBUG_REQ(D_HA, req, "REPLAY");
3105 atomic_inc(&req->rq_import->imp_replay_inflight);
3106 spin_lock(&req->rq_lock);
3107 req->rq_early_free_repbuf = 0;
3108 spin_unlock(&req->rq_lock);
3109 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3111 ptlrpcd_add_req(req);
3116 * Aborts all in-flight request on import \a imp sending and delayed lists
3118 void ptlrpc_abort_inflight(struct obd_import *imp)
3120 struct list_head *tmp, *n;
3123 /* Make sure that no new requests get processed for this import.
3124 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3125 * this flag and then putting requests on sending_list or delayed_list.
3127 spin_lock(&imp->imp_lock);
3129 /* XXX locking? Maybe we should remove each request with the list
3130 * locked? Also, how do we know if the requests on the list are
3131 * being freed at this time?
3133 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3134 struct ptlrpc_request *req = list_entry(tmp,
3135 struct ptlrpc_request,
3138 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3140 spin_lock(&req->rq_lock);
3141 if (req->rq_import_generation < imp->imp_generation) {
3143 req->rq_status = -EIO;
3144 ptlrpc_client_wake_req(req);
3146 spin_unlock(&req->rq_lock);
3149 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3150 struct ptlrpc_request *req =
3151 list_entry(tmp, struct ptlrpc_request, rq_list);
3153 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3155 spin_lock(&req->rq_lock);
3156 if (req->rq_import_generation < imp->imp_generation) {
3158 req->rq_status = -EIO;
3159 ptlrpc_client_wake_req(req);
3161 spin_unlock(&req->rq_lock);
3164 /* Last chance to free reqs left on the replay list, but we
3165 * will still leak reqs that haven't committed. */
3166 if (imp->imp_replayable)
3167 ptlrpc_free_committed(imp);
3169 spin_unlock(&imp->imp_lock);
3175 * Abort all uncompleted requests in request set \a set
3177 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3179 struct list_head *tmp, *pos;
3181 LASSERT(set != NULL);
3183 list_for_each_safe(pos, tmp, &set->set_requests) {
3184 struct ptlrpc_request *req =
3185 list_entry(pos, struct ptlrpc_request,
3188 spin_lock(&req->rq_lock);
3189 if (req->rq_phase != RQ_PHASE_RPC) {
3190 spin_unlock(&req->rq_lock);
3195 req->rq_status = -EINTR;
3196 ptlrpc_client_wake_req(req);
3197 spin_unlock(&req->rq_lock);
3201 static __u64 ptlrpc_last_xid;
3202 static spinlock_t ptlrpc_last_xid_lock;
3205 * Initialize the XID for the node. This is common among all requests on
3206 * this node, and only requires the property that it is monotonically
3207 * increasing. It does not need to be sequential. Since this is also used
3208 * as the RDMA match bits, it is important that a single client NOT have
3209 * the same match bits for two different in-flight requests, hence we do
3210 * NOT want to have an XID per target or similar.
3212 * To avoid an unlikely collision between match bits after a client reboot
3213 * (which would deliver old data into the wrong RDMA buffer) initialize
3214 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3215 * If the time is clearly incorrect, we instead use a 62-bit random number.
3216 * In the worst case the random number will overflow 1M RPCs per second in
3217 * 9133 years, or permutations thereof.
3219 #define YEAR_2004 (1ULL << 30)
3220 void ptlrpc_init_xid(void)
3222 time64_t now = ktime_get_real_seconds();
3224 spin_lock_init(&ptlrpc_last_xid_lock);
3225 if (now < YEAR_2004) {
3226 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3227 ptlrpc_last_xid >>= 2;
3228 ptlrpc_last_xid |= (1ULL << 61);
3230 ptlrpc_last_xid = (__u64)now << 20;
3233 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3234 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3235 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3239 * Increase xid and returns resulting new value to the caller.
3241 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3242 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3243 * itself uses the last bulk xid needed, so the server can determine the
3244 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3245 * xid must align to a power-of-two value.
3247 * This is assumed to be true due to the initial ptlrpc_last_xid
3248 * value also being initialized to a power-of-two value. LU-1431
3250 __u64 ptlrpc_next_xid(void)
3254 spin_lock(&ptlrpc_last_xid_lock);
3255 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3256 ptlrpc_last_xid = next;
3257 spin_unlock(&ptlrpc_last_xid_lock);
3263 * If request has a new allocated XID (new request or EINPROGRESS resend),
3264 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3265 * request to ensure previous bulk fails and avoid problems with lost replies
3266 * and therefore several transfers landing into the same buffer from different
3269 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3271 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3273 LASSERT(bd != NULL);
3275 /* Generate new matchbits for all resend requests, including
3277 if (req->rq_resend) {
3278 __u64 old_mbits = req->rq_mbits;
3280 /* First time resend on -EINPROGRESS will generate new xid,
3281 * so we can actually use the rq_xid as rq_mbits in such case,
3282 * however, it's bit hard to distinguish such resend with a
3283 * 'resend for the -EINPROGRESS resend'. To make it simple,
3284 * we opt to generate mbits for all resend cases. */
3285 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
3286 req->rq_mbits = ptlrpc_next_xid();
3288 /* Old version transfers rq_xid to peer as
3290 spin_lock(&req->rq_import->imp_lock);
3291 list_del_init(&req->rq_unreplied_list);
3292 ptlrpc_assign_next_xid_nolock(req);
3293 spin_unlock(&req->rq_import->imp_lock);
3294 req->rq_mbits = req->rq_xid;
3296 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3297 old_mbits, req->rq_mbits);
3298 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3299 /* Request being sent first time, use xid as matchbits. */
3300 req->rq_mbits = req->rq_xid;
3302 /* Replay request, xid and matchbits have already been
3303 * correctly assigned. */
3307 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3308 * that server can infer the number of bulks that were prepared,
3310 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3313 /* Set rq_xid as rq_mbits to indicate the final bulk for the old
3314 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3316 * It's ok to directly set the rq_xid here, since this xid bump
3317 * won't affect the request position in unreplied list. */
3318 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3319 req->rq_xid = req->rq_mbits;
3323 * Get a glimpse at what next xid value might have been.
3324 * Returns possible next xid.
3326 __u64 ptlrpc_sample_next_xid(void)
3328 #if BITS_PER_LONG == 32
3329 /* need to avoid possible word tearing on 32-bit systems */
3332 spin_lock(&ptlrpc_last_xid_lock);
3333 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3334 spin_unlock(&ptlrpc_last_xid_lock);
3338 /* No need to lock, since returned value is racy anyways */
3339 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3342 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3345 * Functions for operating ptlrpc workers.
3347 * A ptlrpc work is a function which will be running inside ptlrpc context.
3348 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3350 * 1. after a work is created, it can be used many times, that is:
3351 * handler = ptlrpcd_alloc_work();
3352 * ptlrpcd_queue_work();
3354 * queue it again when necessary:
3355 * ptlrpcd_queue_work();
3356 * ptlrpcd_destroy_work();
3357 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3358 * it will only be queued once in any time. Also as its name implies, it may
3359 * have delay before it really runs by ptlrpcd thread.
3361 struct ptlrpc_work_async_args {
3362 int (*cb)(const struct lu_env *, void *);
3366 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3368 /* re-initialize the req */
3369 req->rq_timeout = obd_timeout;
3370 req->rq_sent = ktime_get_real_seconds();
3371 req->rq_deadline = req->rq_sent + req->rq_timeout;
3372 req->rq_phase = RQ_PHASE_INTERPRET;
3373 req->rq_next_phase = RQ_PHASE_COMPLETE;
3374 req->rq_xid = ptlrpc_next_xid();
3375 req->rq_import_generation = req->rq_import->imp_generation;
3377 ptlrpcd_add_req(req);
3380 static int work_interpreter(const struct lu_env *env,
3381 struct ptlrpc_request *req, void *data, int rc)
3383 struct ptlrpc_work_async_args *arg = data;
3385 LASSERT(ptlrpcd_check_work(req));
3386 LASSERT(arg->cb != NULL);
3388 rc = arg->cb(env, arg->cbdata);
3390 list_del_init(&req->rq_set_chain);
3393 if (atomic_dec_return(&req->rq_refcount) > 1) {
3394 atomic_set(&req->rq_refcount, 2);
3395 ptlrpcd_add_work_req(req);
3400 static int worker_format;
3402 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3404 return req->rq_pill.rc_fmt == (void *)&worker_format;
3408 * Create a work for ptlrpc.
3410 void *ptlrpcd_alloc_work(struct obd_import *imp,
3411 int (*cb)(const struct lu_env *, void *), void *cbdata)
3413 struct ptlrpc_request *req = NULL;
3414 struct ptlrpc_work_async_args *args;
3420 RETURN(ERR_PTR(-EINVAL));
3422 /* copy some code from deprecated fakereq. */
3423 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3425 CERROR("ptlrpc: run out of memory!\n");
3426 RETURN(ERR_PTR(-ENOMEM));
3429 ptlrpc_cli_req_init(req);
3431 req->rq_send_state = LUSTRE_IMP_FULL;
3432 req->rq_type = PTL_RPC_MSG_REQUEST;
3433 req->rq_import = class_import_get(imp);
3434 req->rq_interpret_reply = work_interpreter;
3435 /* don't want reply */
3436 req->rq_no_delay = req->rq_no_resend = 1;
3437 req->rq_pill.rc_fmt = (void *)&worker_format;
3439 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3440 args = ptlrpc_req_async_args(req);
3442 args->cbdata = cbdata;
3446 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3448 void ptlrpcd_destroy_work(void *handler)
3450 struct ptlrpc_request *req = handler;
3453 ptlrpc_req_finished(req);
3455 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3457 int ptlrpcd_queue_work(void *handler)
3459 struct ptlrpc_request *req = handler;
3462 * Check if the req is already being queued.
3464 * Here comes a trick: it lacks a way of checking if a req is being
3465 * processed reliably in ptlrpc. Here I have to use refcount of req
3466 * for this purpose. This is okay because the caller should use this
3467 * req as opaque data. - Jinshan
3469 LASSERT(atomic_read(&req->rq_refcount) > 0);
3470 if (atomic_inc_return(&req->rq_refcount) == 2)
3471 ptlrpcd_add_work_req(req);
3474 EXPORT_SYMBOL(ptlrpcd_queue_work);