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(D_INFO, "%s: connect at new req\n",
873 imp->imp_obd->obd_name);
874 spin_lock(&imp->imp_lock);
875 if (imp->imp_state == LUSTRE_IMP_IDLE) {
876 imp->imp_generation++;
877 imp->imp_initiated_at = imp->imp_generation;
878 imp->imp_state = LUSTRE_IMP_NEW;
881 spin_unlock(&imp->imp_lock);
883 rc = ptlrpc_connect_import(imp);
886 ptlrpc_pinger_add_import(imp);
890 request = __ptlrpc_request_alloc(imp, pool);
894 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
895 req_capsule_set(&request->rq_pill, format);
900 * Allocate new request structure for import \a imp and initialize its
901 * buffer structure according to capsule template \a format.
903 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
904 const struct req_format *format)
906 return ptlrpc_request_alloc_internal(imp, NULL, format);
908 EXPORT_SYMBOL(ptlrpc_request_alloc);
911 * Allocate new request structure for import \a imp from pool \a pool and
912 * initialize its buffer structure according to capsule template \a format.
914 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
915 struct ptlrpc_request_pool * pool,
916 const struct req_format *format)
918 return ptlrpc_request_alloc_internal(imp, pool, format);
920 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
923 * For requests not from pool, free memory of the request structure.
924 * For requests obtained from a pool earlier, return request back to pool.
926 void ptlrpc_request_free(struct ptlrpc_request *request)
928 if (request->rq_pool)
929 __ptlrpc_free_req_to_pool(request);
931 ptlrpc_request_cache_free(request);
933 EXPORT_SYMBOL(ptlrpc_request_free);
936 * Allocate new request for operatione \a opcode and immediatelly pack it for
938 * Only used for simple requests like OBD_PING where the only important
939 * part of the request is operation itself.
940 * Returns allocated request or NULL on error.
942 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
943 const struct req_format *format,
944 __u32 version, int opcode)
946 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
950 rc = ptlrpc_request_pack(req, version, opcode);
952 ptlrpc_request_free(req);
958 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
961 * Allocate and initialize new request set structure on the current CPT.
962 * Returns a pointer to the newly allocated set structure or NULL on error.
964 struct ptlrpc_request_set *ptlrpc_prep_set(void)
966 struct ptlrpc_request_set *set;
970 cpt = cfs_cpt_current(cfs_cpt_table, 0);
971 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
974 atomic_set(&set->set_refcount, 1);
975 INIT_LIST_HEAD(&set->set_requests);
976 init_waitqueue_head(&set->set_waitq);
977 atomic_set(&set->set_new_count, 0);
978 atomic_set(&set->set_remaining, 0);
979 spin_lock_init(&set->set_new_req_lock);
980 INIT_LIST_HEAD(&set->set_new_requests);
981 set->set_max_inflight = UINT_MAX;
982 set->set_producer = NULL;
983 set->set_producer_arg = NULL;
988 EXPORT_SYMBOL(ptlrpc_prep_set);
991 * Allocate and initialize new request set structure with flow control
992 * extension. This extension allows to control the number of requests in-flight
993 * for the whole set. A callback function to generate requests must be provided
994 * and the request set will keep the number of requests sent over the wire to
996 * Returns a pointer to the newly allocated set structure or NULL on error.
998 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1002 struct ptlrpc_request_set *set;
1004 set = ptlrpc_prep_set();
1008 set->set_max_inflight = max;
1009 set->set_producer = func;
1010 set->set_producer_arg = arg;
1016 * Wind down and free request set structure previously allocated with
1018 * Ensures that all requests on the set have completed and removes
1019 * all requests from the request list in a set.
1020 * If any unsent request happen to be on the list, pretends that they got
1021 * an error in flight and calls their completion handler.
1023 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1025 struct list_head *tmp;
1026 struct list_head *next;
1031 /* Requests on the set should either all be completed, or all be new */
1032 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1033 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1034 list_for_each(tmp, &set->set_requests) {
1035 struct ptlrpc_request *req =
1036 list_entry(tmp, struct ptlrpc_request,
1039 LASSERT(req->rq_phase == expected_phase);
1043 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1044 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1045 atomic_read(&set->set_remaining), n);
1047 list_for_each_safe(tmp, next, &set->set_requests) {
1048 struct ptlrpc_request *req =
1049 list_entry(tmp, struct ptlrpc_request,
1051 list_del_init(&req->rq_set_chain);
1053 LASSERT(req->rq_phase == expected_phase);
1055 if (req->rq_phase == RQ_PHASE_NEW) {
1056 ptlrpc_req_interpret(NULL, req, -EBADR);
1057 atomic_dec(&set->set_remaining);
1060 spin_lock(&req->rq_lock);
1062 req->rq_invalid_rqset = 0;
1063 spin_unlock(&req->rq_lock);
1065 ptlrpc_req_finished (req);
1068 LASSERT(atomic_read(&set->set_remaining) == 0);
1070 ptlrpc_reqset_put(set);
1073 EXPORT_SYMBOL(ptlrpc_set_destroy);
1076 * Add a new request to the general purpose request set.
1077 * Assumes request reference from the caller.
1079 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1080 struct ptlrpc_request *req)
1082 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1083 LASSERT(list_empty(&req->rq_set_chain));
1085 if (req->rq_allow_intr)
1086 set->set_allow_intr = 1;
1088 /* The set takes over the caller's request reference */
1089 list_add_tail(&req->rq_set_chain, &set->set_requests);
1091 atomic_inc(&set->set_remaining);
1092 req->rq_queued_time = ktime_get_seconds();
1094 if (req->rq_reqmsg != NULL)
1095 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1097 if (set->set_producer != NULL)
1098 /* If the request set has a producer callback, the RPC must be
1099 * sent straight away */
1100 ptlrpc_send_new_req(req);
1102 EXPORT_SYMBOL(ptlrpc_set_add_req);
1105 * Add a request to a request with dedicated server thread
1106 * and wake the thread to make any necessary processing.
1107 * Currently only used for ptlrpcd.
1109 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1110 struct ptlrpc_request *req)
1112 struct ptlrpc_request_set *set = pc->pc_set;
1115 LASSERT(req->rq_set == NULL);
1116 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1118 spin_lock(&set->set_new_req_lock);
1120 * The set takes over the caller's request reference.
1123 req->rq_queued_time = ktime_get_seconds();
1124 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1125 count = atomic_inc_return(&set->set_new_count);
1126 spin_unlock(&set->set_new_req_lock);
1128 /* Only need to call wakeup once for the first entry. */
1130 wake_up(&set->set_waitq);
1132 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1133 * guarantee the async RPC can be processed ASAP, we have
1134 * no other better choice. It maybe fixed in future. */
1135 for (i = 0; i < pc->pc_npartners; i++)
1136 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1141 * Based on the current state of the import, determine if the request
1142 * can be sent, is an error, or should be delayed.
1144 * Returns true if this request should be delayed. If false, and
1145 * *status is set, then the request can not be sent and *status is the
1146 * error code. If false and status is 0, then request can be sent.
1148 * The imp->imp_lock must be held.
1150 static int ptlrpc_import_delay_req(struct obd_import *imp,
1151 struct ptlrpc_request *req, int *status)
1156 LASSERT (status != NULL);
1159 if (req->rq_ctx_init || req->rq_ctx_fini) {
1160 /* always allow ctx init/fini rpc go through */
1161 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1162 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1164 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1165 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1167 /* pings or MDS-equivalent STATFS may safely race with umount */
1168 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1169 D_HA : D_ERROR, req, "IMP_CLOSED ");
1171 } else if (ptlrpc_send_limit_expired(req)) {
1172 /* probably doesn't need to be a D_ERROR after initial testing*/
1173 DEBUG_REQ(D_HA, req, "send limit expired ");
1174 *status = -ETIMEDOUT;
1175 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1176 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1177 /* allow CONNECT even if import is invalid */ ;
1178 if (atomic_read(&imp->imp_inval_count) != 0) {
1179 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1182 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1183 if (!imp->imp_deactive)
1184 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1185 *status = -ESHUTDOWN; /* bz 12940 */
1186 } else if (req->rq_import_generation != imp->imp_generation) {
1187 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1189 } else if (req->rq_send_state != imp->imp_state) {
1190 /* invalidate in progress - any requests should be drop */
1191 if (atomic_read(&imp->imp_inval_count) != 0) {
1192 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1194 } else if (req->rq_no_delay &&
1195 imp->imp_generation != imp->imp_initiated_at) {
1196 /* ignore nodelay for requests initiating connections */
1197 *status = -EWOULDBLOCK;
1198 } else if (req->rq_allow_replay &&
1199 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1200 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1201 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1202 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1203 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1213 * Decide if the error message should be printed to the console or not.
1214 * Makes its decision based on request type, status, and failure frequency.
1216 * \param[in] req request that failed and may need a console message
1218 * \retval false if no message should be printed
1219 * \retval true if console message should be printed
1221 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1223 LASSERT(req->rq_reqmsg != NULL);
1225 /* Suppress particular reconnect errors which are to be expected. */
1226 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1228 /* Suppress timed out reconnect requests */
1229 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1233 /* Suppress most unavailable/again reconnect requests, but
1234 * print occasionally so it is clear client is trying to
1235 * connect to a server where no target is running. */
1236 if ((err == -ENODEV || err == -EAGAIN) &&
1237 req->rq_import->imp_conn_cnt % 30 != 20)
1241 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1242 /* -EAGAIN is normal when using POSIX flocks */
1245 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1246 (req->rq_xid & 0xf) != 10)
1247 /* Suppress most ping requests, they may fail occasionally */
1254 * Check request processing status.
1255 * Returns the status.
1257 static int ptlrpc_check_status(struct ptlrpc_request *req)
1262 err = lustre_msg_get_status(req->rq_repmsg);
1263 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1264 struct obd_import *imp = req->rq_import;
1265 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1266 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1268 if (ptlrpc_console_allow(req, opc, err))
1269 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1270 "failed: rc = %d\n",
1271 imp->imp_obd->obd_name,
1273 libcfs_nid2str(nid), err);
1274 RETURN(err < 0 ? err : -EINVAL);
1278 DEBUG_REQ(D_INFO, req, "status is %d", err);
1279 } else if (err > 0) {
1280 /* XXX: translate this error from net to host */
1281 DEBUG_REQ(D_INFO, req, "status is %d", err);
1288 * save pre-versions of objects into request for replay.
1289 * Versions are obtained from server reply.
1292 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1294 struct lustre_msg *repmsg = req->rq_repmsg;
1295 struct lustre_msg *reqmsg = req->rq_reqmsg;
1296 __u64 *versions = lustre_msg_get_versions(repmsg);
1299 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1303 lustre_msg_set_versions(reqmsg, versions);
1304 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1305 versions[0], versions[1]);
1310 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1312 struct ptlrpc_request *req;
1314 assert_spin_locked(&imp->imp_lock);
1315 if (list_empty(&imp->imp_unreplied_list))
1318 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1320 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1322 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1323 imp->imp_known_replied_xid = req->rq_xid - 1;
1325 return req->rq_xid - 1;
1329 * Callback function called when client receives RPC reply for \a req.
1330 * Returns 0 on success or error code.
1331 * The return alue would be assigned to req->rq_status by the caller
1332 * as request processing status.
1333 * This function also decides if the request needs to be saved for later replay.
1335 static int after_reply(struct ptlrpc_request *req)
1337 struct obd_import *imp = req->rq_import;
1338 struct obd_device *obd = req->rq_import->imp_obd;
1345 LASSERT(obd != NULL);
1346 /* repbuf must be unlinked */
1347 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1349 if (req->rq_reply_truncated) {
1350 if (ptlrpc_no_resend(req)) {
1351 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1352 " expected: %d, actual size: %d",
1353 req->rq_nob_received, req->rq_repbuf_len);
1357 sptlrpc_cli_free_repbuf(req);
1358 /* Pass the required reply buffer size (include
1359 * space for early reply).
1360 * NB: no need to roundup because alloc_repbuf
1361 * will roundup it */
1362 req->rq_replen = req->rq_nob_received;
1363 req->rq_nob_received = 0;
1364 spin_lock(&req->rq_lock);
1366 spin_unlock(&req->rq_lock);
1370 work_start = ktime_get_real();
1371 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1374 * NB Until this point, the whole of the incoming message,
1375 * including buflens, status etc is in the sender's byte order.
1377 rc = sptlrpc_cli_unwrap_reply(req);
1379 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1384 * Security layer unwrap might ask resend this request.
1389 rc = unpack_reply(req);
1393 /* retry indefinitely on EINPROGRESS */
1394 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1395 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1396 time64_t now = ktime_get_real_seconds();
1398 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1399 spin_lock(&req->rq_lock);
1401 spin_unlock(&req->rq_lock);
1402 req->rq_nr_resend++;
1404 /* Readjust the timeout for current conditions */
1405 ptlrpc_at_set_req_timeout(req);
1406 /* delay resend to give a chance to the server to get ready.
1407 * The delay is increased by 1s on every resend and is capped to
1408 * the current request timeout (i.e. obd_timeout if AT is off,
1409 * or AT service time x 125% + 5s, see at_est2timeout) */
1410 if (req->rq_nr_resend > req->rq_timeout)
1411 req->rq_sent = now + req->rq_timeout;
1413 req->rq_sent = now + req->rq_nr_resend;
1415 /* Resend for EINPROGRESS will use a new XID */
1416 spin_lock(&imp->imp_lock);
1417 list_del_init(&req->rq_unreplied_list);
1418 spin_unlock(&imp->imp_lock);
1423 if (obd->obd_svc_stats != NULL) {
1424 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1426 ptlrpc_lprocfs_rpc_sent(req, timediff);
1429 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1430 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1431 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1432 lustre_msg_get_type(req->rq_repmsg));
1436 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1437 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1438 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1439 ptlrpc_at_adj_net_latency(req,
1440 lustre_msg_get_service_time(req->rq_repmsg));
1442 rc = ptlrpc_check_status(req);
1443 imp->imp_connect_error = rc;
1447 * Either we've been evicted, or the server has failed for
1448 * some reason. Try to reconnect, and if that fails, punt to
1451 if (ptlrpc_recoverable_error(rc)) {
1452 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1453 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1456 ptlrpc_request_handle_notconn(req);
1461 * Let's look if server sent slv. Do it only for RPC with
1464 ldlm_cli_update_pool(req);
1468 * Store transno in reqmsg for replay.
1470 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1471 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1472 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1475 if (imp->imp_replayable) {
1476 spin_lock(&imp->imp_lock);
1478 * No point in adding already-committed requests to the replay
1479 * list, we will just remove them immediately. b=9829
1481 if (req->rq_transno != 0 &&
1483 lustre_msg_get_last_committed(req->rq_repmsg) ||
1485 /** version recovery */
1486 ptlrpc_save_versions(req);
1487 ptlrpc_retain_replayable_request(req, imp);
1488 } else if (req->rq_commit_cb != NULL &&
1489 list_empty(&req->rq_replay_list)) {
1490 /* NB: don't call rq_commit_cb if it's already on
1491 * rq_replay_list, ptlrpc_free_committed() will call
1492 * it later, see LU-3618 for details */
1493 spin_unlock(&imp->imp_lock);
1494 req->rq_commit_cb(req);
1495 spin_lock(&imp->imp_lock);
1499 * Replay-enabled imports return commit-status information.
1501 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1502 if (likely(committed > imp->imp_peer_committed_transno))
1503 imp->imp_peer_committed_transno = committed;
1505 ptlrpc_free_committed(imp);
1507 if (!list_empty(&imp->imp_replay_list)) {
1508 struct ptlrpc_request *last;
1510 last = list_entry(imp->imp_replay_list.prev,
1511 struct ptlrpc_request,
1514 * Requests with rq_replay stay on the list even if no
1515 * commit is expected.
1517 if (last->rq_transno > imp->imp_peer_committed_transno)
1518 ptlrpc_pinger_commit_expected(imp);
1521 spin_unlock(&imp->imp_lock);
1528 * Helper function to send request \a req over the network for the first time
1529 * Also adjusts request phase.
1530 * Returns 0 on success or error code.
1532 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1534 struct obd_import *imp = req->rq_import;
1539 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1541 /* do not try to go further if there is not enough memory in enc_pool */
1542 if (req->rq_sent && req->rq_bulk != NULL)
1543 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1544 pool_is_at_full_capacity())
1547 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1548 (!req->rq_generation_set ||
1549 req->rq_import_generation == imp->imp_generation))
1552 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1554 spin_lock(&imp->imp_lock);
1556 LASSERT(req->rq_xid != 0);
1557 LASSERT(!list_empty(&req->rq_unreplied_list));
1559 if (!req->rq_generation_set)
1560 req->rq_import_generation = imp->imp_generation;
1562 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1563 spin_lock(&req->rq_lock);
1564 req->rq_waiting = 1;
1565 spin_unlock(&req->rq_lock);
1567 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1568 ptlrpc_import_state_name(req->rq_send_state),
1569 ptlrpc_import_state_name(imp->imp_state));
1570 LASSERT(list_empty(&req->rq_list));
1571 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1572 atomic_inc(&req->rq_import->imp_inflight);
1573 spin_unlock(&imp->imp_lock);
1578 spin_unlock(&imp->imp_lock);
1579 req->rq_status = rc;
1580 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1584 LASSERT(list_empty(&req->rq_list));
1585 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1586 atomic_inc(&req->rq_import->imp_inflight);
1588 /* find the known replied XID from the unreplied list, CONNECT
1589 * and DISCONNECT requests are skipped to make the sanity check
1590 * on server side happy. see process_req_last_xid().
1592 * For CONNECT: Because replay requests have lower XID, it'll
1593 * break the sanity check if CONNECT bump the exp_last_xid on
1596 * For DISCONNECT: Since client will abort inflight RPC before
1597 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1598 * than the inflight RPC.
1600 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1601 min_xid = ptlrpc_known_replied_xid(imp);
1602 spin_unlock(&imp->imp_lock);
1604 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1606 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1608 rc = sptlrpc_req_refresh_ctx(req, -1);
1611 req->rq_status = rc;
1614 spin_lock(&req->rq_lock);
1615 req->rq_wait_ctx = 1;
1616 spin_unlock(&req->rq_lock);
1621 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1622 " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1623 imp->imp_obd->obd_uuid.uuid,
1624 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1625 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg));
1627 rc = ptl_send_rpc(req, 0);
1628 if (rc == -ENOMEM) {
1629 spin_lock(&imp->imp_lock);
1630 if (!list_empty(&req->rq_list)) {
1631 list_del_init(&req->rq_list);
1632 atomic_dec(&req->rq_import->imp_inflight);
1634 spin_unlock(&imp->imp_lock);
1635 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1639 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1640 spin_lock(&req->rq_lock);
1641 req->rq_net_err = 1;
1642 spin_unlock(&req->rq_lock);
1648 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1653 LASSERT(set->set_producer != NULL);
1655 remaining = atomic_read(&set->set_remaining);
1657 /* populate the ->set_requests list with requests until we
1658 * reach the maximum number of RPCs in flight for this set */
1659 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1660 rc = set->set_producer(set, set->set_producer_arg);
1661 if (rc == -ENOENT) {
1662 /* no more RPC to produce */
1663 set->set_producer = NULL;
1664 set->set_producer_arg = NULL;
1669 RETURN((atomic_read(&set->set_remaining) - remaining));
1673 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1674 * and no more replies are expected.
1675 * (it is possible to get less replies than requests sent e.g. due to timed out
1676 * requests or requests that we had trouble to send out)
1678 * NOTE: This function contains a potential schedule point (cond_resched()).
1680 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1682 struct list_head *tmp, *next;
1683 struct list_head comp_reqs;
1684 int force_timer_recalc = 0;
1687 if (atomic_read(&set->set_remaining) == 0)
1690 INIT_LIST_HEAD(&comp_reqs);
1691 list_for_each_safe(tmp, next, &set->set_requests) {
1692 struct ptlrpc_request *req =
1693 list_entry(tmp, struct ptlrpc_request,
1695 struct obd_import *imp = req->rq_import;
1696 int unregistered = 0;
1700 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1701 list_move_tail(&req->rq_set_chain, &comp_reqs);
1705 /* This schedule point is mainly for the ptlrpcd caller of this
1706 * function. Most ptlrpc sets are not long-lived and unbounded
1707 * in length, but at the least the set used by the ptlrpcd is.
1708 * Since the processing time is unbounded, we need to insert an
1709 * explicit schedule point to make the thread well-behaved.
1713 /* If the caller requires to allow to be interpreted by force
1714 * and it has really been interpreted, then move the request
1715 * to RQ_PHASE_INTERPRET phase in spite of what the current
1717 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1718 req->rq_status = -EINTR;
1719 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1721 /* Since it is interpreted and we have to wait for
1722 * the reply to be unlinked, then use sync mode. */
1725 GOTO(interpret, req->rq_status);
1728 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1729 force_timer_recalc = 1;
1731 /* delayed send - skip */
1732 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1735 /* delayed resend - skip */
1736 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1737 req->rq_sent > ktime_get_real_seconds())
1740 if (!(req->rq_phase == RQ_PHASE_RPC ||
1741 req->rq_phase == RQ_PHASE_BULK ||
1742 req->rq_phase == RQ_PHASE_INTERPRET ||
1743 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1744 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1745 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1749 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1750 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1751 LASSERT(req->rq_next_phase != req->rq_phase);
1752 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1754 if (req->rq_req_deadline &&
1755 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1756 req->rq_req_deadline = 0;
1757 if (req->rq_reply_deadline &&
1758 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1759 req->rq_reply_deadline = 0;
1760 if (req->rq_bulk_deadline &&
1761 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1762 req->rq_bulk_deadline = 0;
1765 * Skip processing until reply is unlinked. We
1766 * can't return to pool before that and we can't
1767 * call interpret before that. We need to make
1768 * sure that all rdma transfers finished and will
1769 * not corrupt any data.
1771 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1772 ptlrpc_client_recv_or_unlink(req))
1774 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1775 ptlrpc_client_bulk_active(req))
1779 * Turn fail_loc off to prevent it from looping
1782 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1783 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1786 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1787 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1792 * Move to next phase if reply was successfully
1795 ptlrpc_rqphase_move(req, req->rq_next_phase);
1798 if (req->rq_phase == RQ_PHASE_INTERPRET)
1799 GOTO(interpret, req->rq_status);
1802 * Note that this also will start async reply unlink.
1804 if (req->rq_net_err && !req->rq_timedout) {
1805 ptlrpc_expire_one_request(req, 1);
1808 * Check if we still need to wait for unlink.
1810 if (ptlrpc_client_recv_or_unlink(req) ||
1811 ptlrpc_client_bulk_active(req))
1813 /* If there is no need to resend, fail it now. */
1814 if (req->rq_no_resend) {
1815 if (req->rq_status == 0)
1816 req->rq_status = -EIO;
1817 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1818 GOTO(interpret, req->rq_status);
1825 spin_lock(&req->rq_lock);
1826 req->rq_replied = 0;
1827 spin_unlock(&req->rq_lock);
1828 if (req->rq_status == 0)
1829 req->rq_status = -EIO;
1830 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1831 GOTO(interpret, req->rq_status);
1834 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1835 * so it sets rq_intr regardless of individual rpc
1836 * timeouts. The synchronous IO waiting path sets
1837 * rq_intr irrespective of whether ptlrpcd
1838 * has seen a timeout. Our policy is to only interpret
1839 * interrupted rpcs after they have timed out, so we
1840 * need to enforce that here.
1843 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1844 req->rq_wait_ctx)) {
1845 req->rq_status = -EINTR;
1846 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1847 GOTO(interpret, req->rq_status);
1850 if (req->rq_phase == RQ_PHASE_RPC) {
1851 if (req->rq_timedout || req->rq_resend ||
1852 req->rq_waiting || req->rq_wait_ctx) {
1855 if (!ptlrpc_unregister_reply(req, 1)) {
1856 ptlrpc_unregister_bulk(req, 1);
1860 spin_lock(&imp->imp_lock);
1861 if (ptlrpc_import_delay_req(imp, req, &status)){
1862 /* put on delay list - only if we wait
1863 * recovery finished - before send */
1864 list_del_init(&req->rq_list);
1865 list_add_tail(&req->rq_list,
1868 spin_unlock(&imp->imp_lock);
1873 req->rq_status = status;
1874 ptlrpc_rqphase_move(req,
1875 RQ_PHASE_INTERPRET);
1876 spin_unlock(&imp->imp_lock);
1877 GOTO(interpret, req->rq_status);
1879 /* ignore on just initiated connections */
1880 if (ptlrpc_no_resend(req) &&
1881 !req->rq_wait_ctx &&
1882 imp->imp_generation !=
1883 imp->imp_initiated_at) {
1884 req->rq_status = -ENOTCONN;
1885 ptlrpc_rqphase_move(req,
1886 RQ_PHASE_INTERPRET);
1887 spin_unlock(&imp->imp_lock);
1888 GOTO(interpret, req->rq_status);
1891 list_del_init(&req->rq_list);
1892 list_add_tail(&req->rq_list,
1893 &imp->imp_sending_list);
1895 spin_unlock(&imp->imp_lock);
1897 spin_lock(&req->rq_lock);
1898 req->rq_waiting = 0;
1899 spin_unlock(&req->rq_lock);
1901 if (req->rq_timedout || req->rq_resend) {
1902 /* This is re-sending anyways,
1903 * let's mark req as resend. */
1904 spin_lock(&req->rq_lock);
1906 spin_unlock(&req->rq_lock);
1908 if (req->rq_bulk != NULL &&
1909 !ptlrpc_unregister_bulk(req, 1))
1913 * rq_wait_ctx is only touched by ptlrpcd,
1914 * so no lock is needed here.
1916 status = sptlrpc_req_refresh_ctx(req, -1);
1919 req->rq_status = status;
1920 spin_lock(&req->rq_lock);
1921 req->rq_wait_ctx = 0;
1922 spin_unlock(&req->rq_lock);
1923 force_timer_recalc = 1;
1925 spin_lock(&req->rq_lock);
1926 req->rq_wait_ctx = 1;
1927 spin_unlock(&req->rq_lock);
1932 spin_lock(&req->rq_lock);
1933 req->rq_wait_ctx = 0;
1934 spin_unlock(&req->rq_lock);
1937 rc = ptl_send_rpc(req, 0);
1938 if (rc == -ENOMEM) {
1939 spin_lock(&imp->imp_lock);
1940 if (!list_empty(&req->rq_list))
1941 list_del_init(&req->rq_list);
1942 spin_unlock(&imp->imp_lock);
1943 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1947 DEBUG_REQ(D_HA, req,
1948 "send failed: rc = %d", rc);
1949 force_timer_recalc = 1;
1950 spin_lock(&req->rq_lock);
1951 req->rq_net_err = 1;
1952 spin_unlock(&req->rq_lock);
1955 /* need to reset the timeout */
1956 force_timer_recalc = 1;
1959 spin_lock(&req->rq_lock);
1961 if (ptlrpc_client_early(req)) {
1962 ptlrpc_at_recv_early_reply(req);
1963 spin_unlock(&req->rq_lock);
1967 /* Still waiting for a reply? */
1968 if (ptlrpc_client_recv(req)) {
1969 spin_unlock(&req->rq_lock);
1973 /* Did we actually receive a reply? */
1974 if (!ptlrpc_client_replied(req)) {
1975 spin_unlock(&req->rq_lock);
1979 spin_unlock(&req->rq_lock);
1981 /* unlink from net because we are going to
1982 * swab in-place of reply buffer */
1983 unregistered = ptlrpc_unregister_reply(req, 1);
1987 req->rq_status = after_reply(req);
1991 /* If there is no bulk associated with this request,
1992 * then we're done and should let the interpreter
1993 * process the reply. Similarly if the RPC returned
1994 * an error, and therefore the bulk will never arrive.
1996 if (req->rq_bulk == NULL || req->rq_status < 0) {
1997 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1998 GOTO(interpret, req->rq_status);
2001 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2004 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2005 if (ptlrpc_client_bulk_active(req))
2008 if (req->rq_bulk->bd_failure) {
2009 /* The RPC reply arrived OK, but the bulk screwed
2010 * up! Dead weird since the server told us the RPC
2011 * was good after getting the REPLY for her GET or
2012 * the ACK for her PUT. */
2013 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2014 req->rq_status = -EIO;
2017 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2020 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2022 /* This moves to "unregistering" phase we need to wait for
2024 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2025 /* start async bulk unlink too */
2026 ptlrpc_unregister_bulk(req, 1);
2030 if (!ptlrpc_unregister_bulk(req, async))
2033 /* When calling interpret receiving already should be
2035 LASSERT(!req->rq_receiving_reply);
2037 ptlrpc_req_interpret(env, req, req->rq_status);
2039 if (ptlrpcd_check_work(req)) {
2040 atomic_dec(&set->set_remaining);
2043 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2045 if (req->rq_reqmsg != NULL)
2047 "Completed RPC pname:cluuid:pid:xid:nid:"
2048 "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
2049 imp->imp_obd->obd_uuid.uuid,
2050 lustre_msg_get_status(req->rq_reqmsg),
2052 obd_import_nid2str(imp),
2053 lustre_msg_get_opc(req->rq_reqmsg));
2055 spin_lock(&imp->imp_lock);
2056 /* Request already may be not on sending or delaying list. This
2057 * may happen in the case of marking it erroneous for the case
2058 * ptlrpc_import_delay_req(req, status) find it impossible to
2059 * allow sending this rpc and returns *status != 0. */
2060 if (!list_empty(&req->rq_list)) {
2061 list_del_init(&req->rq_list);
2062 atomic_dec(&imp->imp_inflight);
2064 list_del_init(&req->rq_unreplied_list);
2065 spin_unlock(&imp->imp_lock);
2067 atomic_dec(&set->set_remaining);
2068 wake_up_all(&imp->imp_recovery_waitq);
2070 if (set->set_producer) {
2071 /* produce a new request if possible */
2072 if (ptlrpc_set_producer(set) > 0)
2073 force_timer_recalc = 1;
2075 /* free the request that has just been completed
2076 * in order not to pollute set->set_requests */
2077 list_del_init(&req->rq_set_chain);
2078 spin_lock(&req->rq_lock);
2080 req->rq_invalid_rqset = 0;
2081 spin_unlock(&req->rq_lock);
2083 /* record rq_status to compute the final status later */
2084 if (req->rq_status != 0)
2085 set->set_rc = req->rq_status;
2086 ptlrpc_req_finished(req);
2088 list_move_tail(&req->rq_set_chain, &comp_reqs);
2092 /* move completed request at the head of list so it's easier for
2093 * caller to find them */
2094 list_splice(&comp_reqs, &set->set_requests);
2096 /* If we hit an error, we want to recover promptly. */
2097 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2099 EXPORT_SYMBOL(ptlrpc_check_set);
2102 * Time out request \a req. is \a async_unlink is set, that means do not wait
2103 * until LNet actually confirms network buffer unlinking.
2104 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2106 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2108 struct obd_import *imp = req->rq_import;
2109 unsigned int debug_mask = D_RPCTRACE;
2113 spin_lock(&req->rq_lock);
2114 req->rq_timedout = 1;
2115 spin_unlock(&req->rq_lock);
2117 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2118 lustre_msg_get_status(req->rq_reqmsg)))
2119 debug_mask = D_WARNING;
2120 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2121 req->rq_net_err ? "failed due to network error" :
2122 ((req->rq_real_sent == 0 ||
2123 req->rq_real_sent < req->rq_sent ||
2124 req->rq_real_sent >= req->rq_deadline) ?
2125 "timed out for sent delay" : "timed out for slow reply"),
2126 (s64)req->rq_sent, (s64)req->rq_real_sent);
2128 if (imp != NULL && obd_debug_peer_on_timeout)
2129 LNetDebugPeer(imp->imp_connection->c_peer);
2131 ptlrpc_unregister_reply(req, async_unlink);
2132 ptlrpc_unregister_bulk(req, async_unlink);
2134 if (obd_dump_on_timeout)
2135 libcfs_debug_dumplog();
2138 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2142 atomic_inc(&imp->imp_timeouts);
2144 /* The DLM server doesn't want recovery run on its imports. */
2145 if (imp->imp_dlm_fake)
2148 /* If this request is for recovery or other primordial tasks,
2149 * then error it out here. */
2150 if (req->rq_ctx_init || req->rq_ctx_fini ||
2151 req->rq_send_state != LUSTRE_IMP_FULL ||
2152 imp->imp_obd->obd_no_recov) {
2153 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2154 ptlrpc_import_state_name(req->rq_send_state),
2155 ptlrpc_import_state_name(imp->imp_state));
2156 spin_lock(&req->rq_lock);
2157 req->rq_status = -ETIMEDOUT;
2159 spin_unlock(&req->rq_lock);
2163 /* if a request can't be resent we can't wait for an answer after
2165 if (ptlrpc_no_resend(req)) {
2166 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2170 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2176 * Time out all uncompleted requests in request set pointed by \a data
2177 * Callback used when waiting on sets with l_wait_event.
2180 int ptlrpc_expired_set(void *data)
2182 struct ptlrpc_request_set *set = data;
2183 struct list_head *tmp;
2184 time64_t now = ktime_get_real_seconds();
2187 LASSERT(set != NULL);
2190 * A timeout expired. See which reqs it applies to...
2192 list_for_each(tmp, &set->set_requests) {
2193 struct ptlrpc_request *req =
2194 list_entry(tmp, struct ptlrpc_request,
2197 /* don't expire request waiting for context */
2198 if (req->rq_wait_ctx)
2201 /* Request in-flight? */
2202 if (!((req->rq_phase == RQ_PHASE_RPC &&
2203 !req->rq_waiting && !req->rq_resend) ||
2204 (req->rq_phase == RQ_PHASE_BULK)))
2207 if (req->rq_timedout || /* already dealt with */
2208 req->rq_deadline > now) /* not expired */
2211 /* Deal with this guy. Do it asynchronously to not block
2212 * ptlrpcd thread. */
2213 ptlrpc_expire_one_request(req, 1);
2217 * When waiting for a whole set, we always break out of the
2218 * sleep so we can recalculate the timeout, or enable interrupts
2219 * if everyone's timed out.
2225 * Sets rq_intr flag in \a req under spinlock.
2227 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2229 spin_lock(&req->rq_lock);
2231 spin_unlock(&req->rq_lock);
2233 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2236 * Interrupts (sets interrupted flag) all uncompleted requests in
2237 * a set \a data. Callback for l_wait_event for interruptible waits.
2239 static void ptlrpc_interrupted_set(void *data)
2241 struct ptlrpc_request_set *set = data;
2242 struct list_head *tmp;
2244 LASSERT(set != NULL);
2245 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2247 list_for_each(tmp, &set->set_requests) {
2248 struct ptlrpc_request *req =
2249 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2254 if (req->rq_phase != RQ_PHASE_RPC &&
2255 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2256 !req->rq_allow_intr)
2259 ptlrpc_mark_interrupted(req);
2264 * Get the smallest timeout in the set; this does NOT set a timeout.
2266 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2268 struct list_head *tmp;
2269 time64_t now = ktime_get_real_seconds();
2271 struct ptlrpc_request *req;
2275 list_for_each(tmp, &set->set_requests) {
2276 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2279 * Request in-flight?
2281 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2282 (req->rq_phase == RQ_PHASE_BULK) ||
2283 (req->rq_phase == RQ_PHASE_NEW)))
2287 * Already timed out.
2289 if (req->rq_timedout)
2295 if (req->rq_wait_ctx)
2298 if (req->rq_phase == RQ_PHASE_NEW)
2299 deadline = req->rq_sent;
2300 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2301 deadline = req->rq_sent;
2303 deadline = req->rq_sent + req->rq_timeout;
2305 if (deadline <= now) /* actually expired already */
2306 timeout = 1; /* ASAP */
2307 else if (timeout == 0 || timeout > deadline - now)
2308 timeout = deadline - now;
2314 * Send all unset request from the set and then wait untill all
2315 * requests in the set complete (either get a reply, timeout, get an
2316 * error or otherwise be interrupted).
2317 * Returns 0 on success or error code otherwise.
2319 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2321 struct list_head *tmp;
2322 struct ptlrpc_request *req;
2323 struct l_wait_info lwi;
2328 if (set->set_producer)
2329 (void)ptlrpc_set_producer(set);
2331 list_for_each(tmp, &set->set_requests) {
2332 req = list_entry(tmp, struct ptlrpc_request,
2334 if (req->rq_phase == RQ_PHASE_NEW)
2335 (void)ptlrpc_send_new_req(req);
2338 if (list_empty(&set->set_requests))
2342 timeout = ptlrpc_set_next_timeout(set);
2344 /* wait until all complete, interrupted, or an in-flight
2346 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2349 if ((timeout == 0 && !signal_pending(current)) ||
2350 set->set_allow_intr)
2351 /* No requests are in-flight (ether timed out
2352 * or delayed), so we can allow interrupts.
2353 * We still want to block for a limited time,
2354 * so we allow interrupts during the timeout. */
2355 lwi = LWI_TIMEOUT_INTR_ALL(
2356 cfs_time_seconds(timeout ? timeout : 1),
2358 ptlrpc_interrupted_set, set);
2361 * At least one request is in flight, so no
2362 * interrupts are allowed. Wait until all
2363 * complete, or an in-flight req times out.
2365 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2366 ptlrpc_expired_set, set);
2368 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2370 /* LU-769 - if we ignored the signal because it was already
2371 * pending when we started, we need to handle it now or we risk
2372 * it being ignored forever */
2373 if (rc == -ETIMEDOUT &&
2374 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2375 signal_pending(current)) {
2376 sigset_t blocked_sigs =
2377 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2379 /* In fact we only interrupt for the "fatal" signals
2380 * like SIGINT or SIGKILL. We still ignore less
2381 * important signals since ptlrpc set is not easily
2382 * reentrant from userspace again */
2383 if (signal_pending(current))
2384 ptlrpc_interrupted_set(set);
2385 cfs_restore_sigs(blocked_sigs);
2388 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2390 /* -EINTR => all requests have been flagged rq_intr so next
2392 * -ETIMEDOUT => someone timed out. When all reqs have
2393 * timed out, signals are enabled allowing completion with
2395 * I don't really care if we go once more round the loop in
2396 * the error cases -eeb. */
2397 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2398 list_for_each(tmp, &set->set_requests) {
2399 req = list_entry(tmp, struct ptlrpc_request,
2401 spin_lock(&req->rq_lock);
2402 req->rq_invalid_rqset = 1;
2403 spin_unlock(&req->rq_lock);
2406 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2408 LASSERT(atomic_read(&set->set_remaining) == 0);
2410 rc = set->set_rc; /* rq_status of already freed requests if any */
2411 list_for_each(tmp, &set->set_requests) {
2412 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2414 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2415 if (req->rq_status != 0)
2416 rc = req->rq_status;
2421 EXPORT_SYMBOL(ptlrpc_set_wait);
2424 * Helper fuction for request freeing.
2425 * Called when request count reached zero and request needs to be freed.
2426 * Removes request from all sorts of sending/replay lists it might be on,
2427 * frees network buffers if any are present.
2428 * If \a locked is set, that means caller is already holding import imp_lock
2429 * and so we no longer need to reobtain it (for certain lists manipulations)
2431 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2435 if (request == NULL)
2438 LASSERT(!request->rq_srv_req);
2439 LASSERT(request->rq_export == NULL);
2440 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2441 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2442 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2443 LASSERTF(!request->rq_replay, "req %p\n", request);
2445 req_capsule_fini(&request->rq_pill);
2447 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2448 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2449 if (request->rq_import != NULL) {
2451 spin_lock(&request->rq_import->imp_lock);
2452 list_del_init(&request->rq_replay_list);
2453 list_del_init(&request->rq_unreplied_list);
2455 spin_unlock(&request->rq_import->imp_lock);
2457 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2459 if (atomic_read(&request->rq_refcount) != 0) {
2460 DEBUG_REQ(D_ERROR, request,
2461 "freeing request with nonzero refcount");
2465 if (request->rq_repbuf != NULL)
2466 sptlrpc_cli_free_repbuf(request);
2468 if (request->rq_import != NULL) {
2469 class_import_put(request->rq_import);
2470 request->rq_import = NULL;
2472 if (request->rq_bulk != NULL)
2473 ptlrpc_free_bulk(request->rq_bulk);
2475 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2476 sptlrpc_cli_free_reqbuf(request);
2478 if (request->rq_cli_ctx)
2479 sptlrpc_req_put_ctx(request, !locked);
2481 if (request->rq_pool)
2482 __ptlrpc_free_req_to_pool(request);
2484 ptlrpc_request_cache_free(request);
2488 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2490 * Drop one request reference. Must be called with import imp_lock held.
2491 * When reference count drops to zero, request is freed.
2493 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2495 assert_spin_locked(&request->rq_import->imp_lock);
2496 (void)__ptlrpc_req_finished(request, 1);
2501 * Drops one reference count for request \a request.
2502 * \a locked set indicates that caller holds import imp_lock.
2503 * Frees the request whe reference count reaches zero.
2505 * \retval 1 the request is freed
2506 * \retval 0 some others still hold references on the request
2508 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2516 LASSERT(request != LP_POISON);
2517 LASSERT(request->rq_reqmsg != LP_POISON);
2519 DEBUG_REQ(D_INFO, request, "refcount now %u",
2520 atomic_read(&request->rq_refcount) - 1);
2522 spin_lock(&request->rq_lock);
2523 count = atomic_dec_return(&request->rq_refcount);
2524 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2526 /* For open RPC, the client does not know the EA size (LOV, ACL, and
2527 * so on) before replied, then the client has to reserve very large
2528 * reply buffer. Such buffer will not be released until the RPC freed.
2529 * Since The open RPC is replayable, we need to keep it in the replay
2530 * list until close. If there are a lot of files opened concurrently,
2531 * then the client may be OOM.
2533 * If fact, it is unnecessary to keep reply buffer for open replay,
2534 * related EAs have already been saved via mdc_save_lovea() before
2535 * coming here. So it is safe to free the reply buffer some earlier
2536 * before releasing the RPC to avoid client OOM. LU-9514 */
2537 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2538 spin_lock(&request->rq_early_free_lock);
2539 sptlrpc_cli_free_repbuf(request);
2540 request->rq_repbuf = NULL;
2541 request->rq_repbuf_len = 0;
2542 request->rq_repdata = NULL;
2543 request->rq_reqdata_len = 0;
2544 spin_unlock(&request->rq_early_free_lock);
2546 spin_unlock(&request->rq_lock);
2549 __ptlrpc_free_req(request, locked);
2555 * Drops one reference count for a request.
2557 void ptlrpc_req_finished(struct ptlrpc_request *request)
2559 __ptlrpc_req_finished(request, 0);
2561 EXPORT_SYMBOL(ptlrpc_req_finished);
2564 * Returns xid of a \a request
2566 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2568 return request->rq_xid;
2570 EXPORT_SYMBOL(ptlrpc_req_xid);
2573 * Disengage the client's reply buffer from the network
2574 * NB does _NOT_ unregister any client-side bulk.
2575 * IDEMPOTENT, but _not_ safe against concurrent callers.
2576 * The request owner (i.e. the thread doing the I/O) must call...
2577 * Returns 0 on success or 1 if unregistering cannot be made.
2579 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2582 struct l_wait_info lwi;
2587 LASSERT(!in_interrupt());
2589 /* Let's setup deadline for reply unlink. */
2590 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2591 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2592 request->rq_reply_deadline = ktime_get_real_seconds() +
2596 * Nothing left to do.
2598 if (!ptlrpc_client_recv_or_unlink(request))
2601 LNetMDUnlink(request->rq_reply_md_h);
2604 * Let's check it once again.
2606 if (!ptlrpc_client_recv_or_unlink(request))
2609 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2610 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2613 * Do not wait for unlink to finish.
2619 * We have to l_wait_event() whatever the result, to give liblustre
2620 * a chance to run reply_in_callback(), and to make sure we've
2621 * unlinked before returning a req to the pool.
2624 /* The wq argument is ignored by user-space wait_event macros */
2625 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2626 &request->rq_set->set_waitq :
2627 &request->rq_reply_waitq;
2628 /* Network access will complete in finite time but the HUGE
2629 * timeout lets us CWARN for visibility of sluggish NALs */
2630 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2631 cfs_time_seconds(1), NULL, NULL);
2632 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2635 ptlrpc_rqphase_move(request, request->rq_next_phase);
2639 LASSERT(rc == -ETIMEDOUT);
2640 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2641 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2642 request->rq_receiving_reply,
2643 request->rq_req_unlinked,
2644 request->rq_reply_unlinked);
2649 static void ptlrpc_free_request(struct ptlrpc_request *req)
2651 spin_lock(&req->rq_lock);
2653 spin_unlock(&req->rq_lock);
2655 if (req->rq_commit_cb != NULL)
2656 req->rq_commit_cb(req);
2657 list_del_init(&req->rq_replay_list);
2659 __ptlrpc_req_finished(req, 1);
2663 * the request is committed and dropped from the replay list of its import
2665 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2667 struct obd_import *imp = req->rq_import;
2669 spin_lock(&imp->imp_lock);
2670 if (list_empty(&req->rq_replay_list)) {
2671 spin_unlock(&imp->imp_lock);
2675 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2676 ptlrpc_free_request(req);
2678 spin_unlock(&imp->imp_lock);
2680 EXPORT_SYMBOL(ptlrpc_request_committed);
2683 * Iterates through replay_list on import and prunes
2684 * all requests have transno smaller than last_committed for the
2685 * import and don't have rq_replay set.
2686 * Since requests are sorted in transno order, stops when meetign first
2687 * transno bigger than last_committed.
2688 * caller must hold imp->imp_lock
2690 void ptlrpc_free_committed(struct obd_import *imp)
2692 struct ptlrpc_request *req, *saved;
2693 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2694 bool skip_committed_list = true;
2697 LASSERT(imp != NULL);
2698 assert_spin_locked(&imp->imp_lock);
2700 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2701 imp->imp_generation == imp->imp_last_generation_checked) {
2702 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2703 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2706 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2707 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2708 imp->imp_generation);
2710 if (imp->imp_generation != imp->imp_last_generation_checked ||
2711 imp->imp_last_transno_checked == 0)
2712 skip_committed_list = false;
2714 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2715 imp->imp_last_generation_checked = imp->imp_generation;
2717 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2719 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2720 LASSERT(req != last_req);
2723 if (req->rq_transno == 0) {
2724 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2727 if (req->rq_import_generation < imp->imp_generation) {
2728 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2732 /* not yet committed */
2733 if (req->rq_transno > imp->imp_peer_committed_transno) {
2734 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2738 if (req->rq_replay) {
2739 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2740 list_move_tail(&req->rq_replay_list,
2741 &imp->imp_committed_list);
2745 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2746 imp->imp_peer_committed_transno);
2748 ptlrpc_free_request(req);
2751 if (skip_committed_list)
2754 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2756 LASSERT(req->rq_transno != 0);
2757 if (req->rq_import_generation < imp->imp_generation ||
2759 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2760 req->rq_import_generation <
2761 imp->imp_generation ? "stale" : "closed");
2763 if (imp->imp_replay_cursor == &req->rq_replay_list)
2764 imp->imp_replay_cursor =
2765 req->rq_replay_list.next;
2767 ptlrpc_free_request(req);
2774 void ptlrpc_cleanup_client(struct obd_import *imp)
2781 * Schedule previously sent request for resend.
2782 * For bulk requests we assign new xid (to avoid problems with
2783 * lost replies and therefore several transfers landing into same buffer
2784 * from different sending attempts).
2786 void ptlrpc_resend_req(struct ptlrpc_request *req)
2788 DEBUG_REQ(D_HA, req, "going to resend");
2789 spin_lock(&req->rq_lock);
2791 /* Request got reply but linked to the import list still.
2792 Let ptlrpc_check_set() to process it. */
2793 if (ptlrpc_client_replied(req)) {
2794 spin_unlock(&req->rq_lock);
2795 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2799 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2800 req->rq_status = -EAGAIN;
2803 req->rq_net_err = 0;
2804 req->rq_timedout = 0;
2806 ptlrpc_client_wake_req(req);
2807 spin_unlock(&req->rq_lock);
2810 /* XXX: this function and rq_status are currently unused */
2811 void ptlrpc_restart_req(struct ptlrpc_request *req)
2813 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2814 req->rq_status = -ERESTARTSYS;
2816 spin_lock(&req->rq_lock);
2817 req->rq_restart = 1;
2818 req->rq_timedout = 0;
2819 ptlrpc_client_wake_req(req);
2820 spin_unlock(&req->rq_lock);
2824 * Grab additional reference on a request \a req
2826 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2829 atomic_inc(&req->rq_refcount);
2832 EXPORT_SYMBOL(ptlrpc_request_addref);
2835 * Add a request to import replay_list.
2836 * Must be called under imp_lock
2838 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2839 struct obd_import *imp)
2841 struct list_head *tmp;
2843 assert_spin_locked(&imp->imp_lock);
2845 if (req->rq_transno == 0) {
2846 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2850 /* clear this for new requests that were resent as well
2851 as resent replayed requests. */
2852 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2854 /* don't re-add requests that have been replayed */
2855 if (!list_empty(&req->rq_replay_list))
2858 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2860 spin_lock(&req->rq_lock);
2862 spin_unlock(&req->rq_lock);
2864 LASSERT(imp->imp_replayable);
2865 /* Balanced in ptlrpc_free_committed, usually. */
2866 ptlrpc_request_addref(req);
2867 list_for_each_prev(tmp, &imp->imp_replay_list) {
2868 struct ptlrpc_request *iter = list_entry(tmp,
2869 struct ptlrpc_request,
2872 /* We may have duplicate transnos if we create and then
2873 * open a file, or for closes retained if to match creating
2874 * opens, so use req->rq_xid as a secondary key.
2875 * (See bugs 684, 685, and 428.)
2876 * XXX no longer needed, but all opens need transnos!
2878 if (iter->rq_transno > req->rq_transno)
2881 if (iter->rq_transno == req->rq_transno) {
2882 LASSERT(iter->rq_xid != req->rq_xid);
2883 if (iter->rq_xid > req->rq_xid)
2887 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2891 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2895 * Send request and wait until it completes.
2896 * Returns request processing status.
2898 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2900 struct ptlrpc_request_set *set;
2904 LASSERT(req->rq_set == NULL);
2905 LASSERT(!req->rq_receiving_reply);
2907 set = ptlrpc_prep_set();
2909 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2913 /* for distributed debugging */
2914 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2916 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2917 ptlrpc_request_addref(req);
2918 ptlrpc_set_add_req(set, req);
2919 rc = ptlrpc_set_wait(set);
2920 ptlrpc_set_destroy(set);
2924 EXPORT_SYMBOL(ptlrpc_queue_wait);
2927 * Callback used for replayed requests reply processing.
2928 * In case of successful reply calls registered request replay callback.
2929 * In case of error restart replay process.
2931 static int ptlrpc_replay_interpret(const struct lu_env *env,
2932 struct ptlrpc_request *req,
2933 void * data, int rc)
2935 struct ptlrpc_replay_async_args *aa = data;
2936 struct obd_import *imp = req->rq_import;
2939 atomic_dec(&imp->imp_replay_inflight);
2941 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2942 * server got the request, but bulk transfer timeout, let's
2943 * replay the bulk req again */
2944 if (!ptlrpc_client_replied(req) ||
2945 (req->rq_bulk != NULL &&
2946 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2947 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2948 GOTO(out, rc = -ETIMEDOUT);
2951 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2952 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2953 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2954 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2956 /** VBR: check version failure */
2957 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2958 /** replay was failed due to version mismatch */
2959 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2960 spin_lock(&imp->imp_lock);
2961 imp->imp_vbr_failed = 1;
2962 spin_unlock(&imp->imp_lock);
2963 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2965 /** The transno had better not change over replay. */
2966 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2967 lustre_msg_get_transno(req->rq_repmsg) ||
2968 lustre_msg_get_transno(req->rq_repmsg) == 0,
2970 lustre_msg_get_transno(req->rq_reqmsg),
2971 lustre_msg_get_transno(req->rq_repmsg));
2974 spin_lock(&imp->imp_lock);
2975 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2976 spin_unlock(&imp->imp_lock);
2977 LASSERT(imp->imp_last_replay_transno);
2979 /* transaction number shouldn't be bigger than the latest replayed */
2980 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2981 DEBUG_REQ(D_ERROR, req,
2982 "Reported transno %llu is bigger than the "
2983 "replayed one: %llu", req->rq_transno,
2984 lustre_msg_get_transno(req->rq_reqmsg));
2985 GOTO(out, rc = -EINVAL);
2988 DEBUG_REQ(D_HA, req, "got rep");
2990 /* let the callback do fixups, possibly including in the request */
2991 if (req->rq_replay_cb)
2992 req->rq_replay_cb(req);
2994 if (ptlrpc_client_replied(req) &&
2995 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2996 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2997 lustre_msg_get_status(req->rq_repmsg),
2998 aa->praa_old_status);
3000 /* Note: If the replay fails for MDT-MDT recovery, let's
3001 * abort all of the following requests in the replay
3002 * and sending list, because MDT-MDT update requests
3003 * are dependent on each other, see LU-7039 */
3004 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3005 struct ptlrpc_request *free_req;
3006 struct ptlrpc_request *tmp;
3008 spin_lock(&imp->imp_lock);
3009 list_for_each_entry_safe(free_req, tmp,
3010 &imp->imp_replay_list,
3012 ptlrpc_free_request(free_req);
3015 list_for_each_entry_safe(free_req, tmp,
3016 &imp->imp_committed_list,
3018 ptlrpc_free_request(free_req);
3021 list_for_each_entry_safe(free_req, tmp,
3022 &imp->imp_delayed_list,
3024 spin_lock(&free_req->rq_lock);
3025 free_req->rq_err = 1;
3026 free_req->rq_status = -EIO;
3027 ptlrpc_client_wake_req(free_req);
3028 spin_unlock(&free_req->rq_lock);
3031 list_for_each_entry_safe(free_req, tmp,
3032 &imp->imp_sending_list,
3034 spin_lock(&free_req->rq_lock);
3035 free_req->rq_err = 1;
3036 free_req->rq_status = -EIO;
3037 ptlrpc_client_wake_req(free_req);
3038 spin_unlock(&free_req->rq_lock);
3040 spin_unlock(&imp->imp_lock);
3043 /* Put it back for re-replay. */
3044 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3048 * Errors while replay can set transno to 0, but
3049 * imp_last_replay_transno shouldn't be set to 0 anyway
3051 if (req->rq_transno == 0)
3052 CERROR("Transno is 0 during replay!\n");
3054 /* continue with recovery */
3055 rc = ptlrpc_import_recovery_state_machine(imp);
3057 req->rq_send_state = aa->praa_old_state;
3060 /* this replay failed, so restart recovery */
3061 ptlrpc_connect_import(imp);
3067 * Prepares and queues request for replay.
3068 * Adds it to ptlrpcd queue for actual sending.
3069 * Returns 0 on success.
3071 int ptlrpc_replay_req(struct ptlrpc_request *req)
3073 struct ptlrpc_replay_async_args *aa;
3077 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3079 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3080 aa = ptlrpc_req_async_args(req);
3081 memset(aa, 0, sizeof(*aa));
3083 /* Prepare request to be resent with ptlrpcd */
3084 aa->praa_old_state = req->rq_send_state;
3085 req->rq_send_state = LUSTRE_IMP_REPLAY;
3086 req->rq_phase = RQ_PHASE_NEW;
3087 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3089 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3091 req->rq_interpret_reply = ptlrpc_replay_interpret;
3092 /* Readjust the timeout for current conditions */
3093 ptlrpc_at_set_req_timeout(req);
3095 /* Tell server the net_latency, so the server can calculate how long
3096 * it should wait for next replay */
3097 lustre_msg_set_service_time(req->rq_reqmsg,
3098 ptlrpc_at_get_net_latency(req));
3099 DEBUG_REQ(D_HA, req, "REPLAY");
3101 atomic_inc(&req->rq_import->imp_replay_inflight);
3102 spin_lock(&req->rq_lock);
3103 req->rq_early_free_repbuf = 0;
3104 spin_unlock(&req->rq_lock);
3105 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3107 ptlrpcd_add_req(req);
3112 * Aborts all in-flight request on import \a imp sending and delayed lists
3114 void ptlrpc_abort_inflight(struct obd_import *imp)
3116 struct list_head *tmp, *n;
3119 /* Make sure that no new requests get processed for this import.
3120 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3121 * this flag and then putting requests on sending_list or delayed_list.
3123 spin_lock(&imp->imp_lock);
3125 /* XXX locking? Maybe we should remove each request with the list
3126 * locked? Also, how do we know if the requests on the list are
3127 * being freed at this time?
3129 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3130 struct ptlrpc_request *req = list_entry(tmp,
3131 struct ptlrpc_request,
3134 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3136 spin_lock(&req->rq_lock);
3137 if (req->rq_import_generation < imp->imp_generation) {
3139 req->rq_status = -EIO;
3140 ptlrpc_client_wake_req(req);
3142 spin_unlock(&req->rq_lock);
3145 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3146 struct ptlrpc_request *req =
3147 list_entry(tmp, struct ptlrpc_request, rq_list);
3149 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3151 spin_lock(&req->rq_lock);
3152 if (req->rq_import_generation < imp->imp_generation) {
3154 req->rq_status = -EIO;
3155 ptlrpc_client_wake_req(req);
3157 spin_unlock(&req->rq_lock);
3160 /* Last chance to free reqs left on the replay list, but we
3161 * will still leak reqs that haven't committed. */
3162 if (imp->imp_replayable)
3163 ptlrpc_free_committed(imp);
3165 spin_unlock(&imp->imp_lock);
3171 * Abort all uncompleted requests in request set \a set
3173 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3175 struct list_head *tmp, *pos;
3177 LASSERT(set != NULL);
3179 list_for_each_safe(pos, tmp, &set->set_requests) {
3180 struct ptlrpc_request *req =
3181 list_entry(pos, struct ptlrpc_request,
3184 spin_lock(&req->rq_lock);
3185 if (req->rq_phase != RQ_PHASE_RPC) {
3186 spin_unlock(&req->rq_lock);
3191 req->rq_status = -EINTR;
3192 ptlrpc_client_wake_req(req);
3193 spin_unlock(&req->rq_lock);
3197 static __u64 ptlrpc_last_xid;
3198 static spinlock_t ptlrpc_last_xid_lock;
3201 * Initialize the XID for the node. This is common among all requests on
3202 * this node, and only requires the property that it is monotonically
3203 * increasing. It does not need to be sequential. Since this is also used
3204 * as the RDMA match bits, it is important that a single client NOT have
3205 * the same match bits for two different in-flight requests, hence we do
3206 * NOT want to have an XID per target or similar.
3208 * To avoid an unlikely collision between match bits after a client reboot
3209 * (which would deliver old data into the wrong RDMA buffer) initialize
3210 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3211 * If the time is clearly incorrect, we instead use a 62-bit random number.
3212 * In the worst case the random number will overflow 1M RPCs per second in
3213 * 9133 years, or permutations thereof.
3215 #define YEAR_2004 (1ULL << 30)
3216 void ptlrpc_init_xid(void)
3218 time64_t now = ktime_get_real_seconds();
3220 spin_lock_init(&ptlrpc_last_xid_lock);
3221 if (now < YEAR_2004) {
3222 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3223 ptlrpc_last_xid >>= 2;
3224 ptlrpc_last_xid |= (1ULL << 61);
3226 ptlrpc_last_xid = (__u64)now << 20;
3229 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3230 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3231 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3235 * Increase xid and returns resulting new value to the caller.
3237 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3238 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3239 * itself uses the last bulk xid needed, so the server can determine the
3240 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3241 * xid must align to a power-of-two value.
3243 * This is assumed to be true due to the initial ptlrpc_last_xid
3244 * value also being initialized to a power-of-two value. LU-1431
3246 __u64 ptlrpc_next_xid(void)
3250 spin_lock(&ptlrpc_last_xid_lock);
3251 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3252 ptlrpc_last_xid = next;
3253 spin_unlock(&ptlrpc_last_xid_lock);
3259 * If request has a new allocated XID (new request or EINPROGRESS resend),
3260 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3261 * request to ensure previous bulk fails and avoid problems with lost replies
3262 * and therefore several transfers landing into the same buffer from different
3265 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3267 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3269 LASSERT(bd != NULL);
3271 /* Generate new matchbits for all resend requests, including
3273 if (req->rq_resend) {
3274 __u64 old_mbits = req->rq_mbits;
3276 /* First time resend on -EINPROGRESS will generate new xid,
3277 * so we can actually use the rq_xid as rq_mbits in such case,
3278 * however, it's bit hard to distinguish such resend with a
3279 * 'resend for the -EINPROGRESS resend'. To make it simple,
3280 * we opt to generate mbits for all resend cases. */
3281 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
3282 req->rq_mbits = ptlrpc_next_xid();
3284 /* Old version transfers rq_xid to peer as
3286 spin_lock(&req->rq_import->imp_lock);
3287 list_del_init(&req->rq_unreplied_list);
3288 ptlrpc_assign_next_xid_nolock(req);
3289 spin_unlock(&req->rq_import->imp_lock);
3290 req->rq_mbits = req->rq_xid;
3292 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3293 old_mbits, req->rq_mbits);
3294 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3295 /* Request being sent first time, use xid as matchbits. */
3296 req->rq_mbits = req->rq_xid;
3298 /* Replay request, xid and matchbits have already been
3299 * correctly assigned. */
3303 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3304 * that server can infer the number of bulks that were prepared,
3306 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3309 /* Set rq_xid as rq_mbits to indicate the final bulk for the old
3310 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3312 * It's ok to directly set the rq_xid here, since this xid bump
3313 * won't affect the request position in unreplied list. */
3314 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3315 req->rq_xid = req->rq_mbits;
3319 * Get a glimpse at what next xid value might have been.
3320 * Returns possible next xid.
3322 __u64 ptlrpc_sample_next_xid(void)
3324 #if BITS_PER_LONG == 32
3325 /* need to avoid possible word tearing on 32-bit systems */
3328 spin_lock(&ptlrpc_last_xid_lock);
3329 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3330 spin_unlock(&ptlrpc_last_xid_lock);
3334 /* No need to lock, since returned value is racy anyways */
3335 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3338 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3341 * Functions for operating ptlrpc workers.
3343 * A ptlrpc work is a function which will be running inside ptlrpc context.
3344 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3346 * 1. after a work is created, it can be used many times, that is:
3347 * handler = ptlrpcd_alloc_work();
3348 * ptlrpcd_queue_work();
3350 * queue it again when necessary:
3351 * ptlrpcd_queue_work();
3352 * ptlrpcd_destroy_work();
3353 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3354 * it will only be queued once in any time. Also as its name implies, it may
3355 * have delay before it really runs by ptlrpcd thread.
3357 struct ptlrpc_work_async_args {
3358 int (*cb)(const struct lu_env *, void *);
3362 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3364 /* re-initialize the req */
3365 req->rq_timeout = obd_timeout;
3366 req->rq_sent = ktime_get_real_seconds();
3367 req->rq_deadline = req->rq_sent + req->rq_timeout;
3368 req->rq_phase = RQ_PHASE_INTERPRET;
3369 req->rq_next_phase = RQ_PHASE_COMPLETE;
3370 req->rq_xid = ptlrpc_next_xid();
3371 req->rq_import_generation = req->rq_import->imp_generation;
3373 ptlrpcd_add_req(req);
3376 static int work_interpreter(const struct lu_env *env,
3377 struct ptlrpc_request *req, void *data, int rc)
3379 struct ptlrpc_work_async_args *arg = data;
3381 LASSERT(ptlrpcd_check_work(req));
3382 LASSERT(arg->cb != NULL);
3384 rc = arg->cb(env, arg->cbdata);
3386 list_del_init(&req->rq_set_chain);
3389 if (atomic_dec_return(&req->rq_refcount) > 1) {
3390 atomic_set(&req->rq_refcount, 2);
3391 ptlrpcd_add_work_req(req);
3396 static int worker_format;
3398 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3400 return req->rq_pill.rc_fmt == (void *)&worker_format;
3404 * Create a work for ptlrpc.
3406 void *ptlrpcd_alloc_work(struct obd_import *imp,
3407 int (*cb)(const struct lu_env *, void *), void *cbdata)
3409 struct ptlrpc_request *req = NULL;
3410 struct ptlrpc_work_async_args *args;
3416 RETURN(ERR_PTR(-EINVAL));
3418 /* copy some code from deprecated fakereq. */
3419 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3421 CERROR("ptlrpc: run out of memory!\n");
3422 RETURN(ERR_PTR(-ENOMEM));
3425 ptlrpc_cli_req_init(req);
3427 req->rq_send_state = LUSTRE_IMP_FULL;
3428 req->rq_type = PTL_RPC_MSG_REQUEST;
3429 req->rq_import = class_import_get(imp);
3430 req->rq_interpret_reply = work_interpreter;
3431 /* don't want reply */
3432 req->rq_no_delay = req->rq_no_resend = 1;
3433 req->rq_pill.rc_fmt = (void *)&worker_format;
3435 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3436 args = ptlrpc_req_async_args(req);
3438 args->cbdata = cbdata;
3442 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3444 void ptlrpcd_destroy_work(void *handler)
3446 struct ptlrpc_request *req = handler;
3449 ptlrpc_req_finished(req);
3451 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3453 int ptlrpcd_queue_work(void *handler)
3455 struct ptlrpc_request *req = handler;
3458 * Check if the req is already being queued.
3460 * Here comes a trick: it lacks a way of checking if a req is being
3461 * processed reliably in ptlrpc. Here I have to use refcount of req
3462 * for this purpose. This is okay because the caller should use this
3463 * req as opaque data. - Jinshan
3465 LASSERT(atomic_read(&req->rq_refcount) > 0);
3466 if (atomic_inc_return(&req->rq_refcount) == 2)
3467 ptlrpcd_add_work_req(req);
3470 EXPORT_SYMBOL(ptlrpcd_queue_work);