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 static __u64 ptlrpc_last_xid;
693 static spinlock_t ptlrpc_last_xid_lock;
695 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
696 __u32 version, int opcode, char **bufs,
697 struct ptlrpc_cli_ctx *ctx)
700 struct obd_import *imp;
706 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
707 imp = request->rq_import;
708 lengths = request->rq_pill.rc_area[RCL_CLIENT];
711 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
713 rc = sptlrpc_req_get_ctx(request);
717 sptlrpc_req_set_flavor(request, opcode);
719 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
724 lustre_msg_add_version(request->rq_reqmsg, version);
725 request->rq_send_state = LUSTRE_IMP_FULL;
726 request->rq_type = PTL_RPC_MSG_REQUEST;
728 request->rq_req_cbid.cbid_fn = request_out_callback;
729 request->rq_req_cbid.cbid_arg = request;
731 request->rq_reply_cbid.cbid_fn = reply_in_callback;
732 request->rq_reply_cbid.cbid_arg = request;
734 request->rq_reply_deadline = 0;
735 request->rq_bulk_deadline = 0;
736 request->rq_req_deadline = 0;
737 request->rq_phase = RQ_PHASE_NEW;
738 request->rq_next_phase = RQ_PHASE_UNDEFINED;
740 request->rq_request_portal = imp->imp_client->cli_request_portal;
741 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
743 ptlrpc_at_set_req_timeout(request);
745 lustre_msg_set_opc(request->rq_reqmsg, opcode);
747 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
748 if (cfs_fail_val == opcode) {
749 time64_t *fail_t = NULL, *fail2_t = NULL;
751 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
752 fail_t = &request->rq_bulk_deadline;
753 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
754 fail_t = &request->rq_reply_deadline;
755 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
756 fail_t = &request->rq_req_deadline;
757 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
758 fail_t = &request->rq_reply_deadline;
759 fail2_t = &request->rq_bulk_deadline;
760 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
761 time64_t now = ktime_get_real_seconds();
762 spin_lock(&ptlrpc_last_xid_lock);
763 ptlrpc_last_xid = ((__u64)now >> 4) << 24;
764 spin_unlock(&ptlrpc_last_xid_lock);
768 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
771 *fail2_t = ktime_get_real_seconds() +
775 * The RPC is infected, let the test to change the
778 msleep(4 * MSEC_PER_SEC);
781 ptlrpc_assign_next_xid(request);
786 LASSERT(!request->rq_pool);
787 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
789 class_import_put(imp);
794 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
797 * Pack request buffers for network transfer, performing necessary encryption
798 * steps if necessary.
800 int ptlrpc_request_pack(struct ptlrpc_request *request,
801 __u32 version, int opcode)
804 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
808 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
809 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
810 * have to send old ptlrpc_body to keep interoprability with these
813 * Only three kinds of server->client RPCs so far:
818 * XXX This should be removed whenever we drop the interoprability with
819 * the these old clients.
821 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
822 opcode == LDLM_GL_CALLBACK)
823 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
824 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
828 EXPORT_SYMBOL(ptlrpc_request_pack);
831 * Helper function to allocate new request on import \a imp
832 * and possibly using existing request from pool \a pool if provided.
833 * Returns allocated request structure with import field filled or
837 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
838 struct ptlrpc_request_pool *pool)
840 struct ptlrpc_request *request = NULL;
842 request = ptlrpc_request_cache_alloc(GFP_NOFS);
844 if (!request && pool)
845 request = ptlrpc_prep_req_from_pool(pool);
848 ptlrpc_cli_req_init(request);
850 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
851 LASSERT(imp != LP_POISON);
852 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
854 LASSERT(imp->imp_client != LP_POISON);
856 request->rq_import = class_import_get(imp);
858 CERROR("request allocation out of memory\n");
865 * Helper function for creating a request.
866 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
867 * buffer structures according to capsule template \a format.
868 * Returns allocated request structure pointer or NULL on error.
870 static struct ptlrpc_request *
871 ptlrpc_request_alloc_internal(struct obd_import *imp,
872 struct ptlrpc_request_pool * pool,
873 const struct req_format *format)
875 struct ptlrpc_request *request;
878 request = __ptlrpc_request_alloc(imp, pool);
882 /* initiate connection if needed when the import has been
883 * referenced by the new request to avoid races with disconnect */
884 if (unlikely(imp->imp_state == LUSTRE_IMP_IDLE)) {
886 CDEBUG_LIMIT(imp->imp_idle_debug,
887 "%s: reconnect after %llds idle\n",
888 imp->imp_obd->obd_name, ktime_get_real_seconds() -
889 imp->imp_last_reply_time);
890 spin_lock(&imp->imp_lock);
891 if (imp->imp_state == LUSTRE_IMP_IDLE) {
892 imp->imp_generation++;
893 imp->imp_initiated_at = imp->imp_generation;
894 imp->imp_state = LUSTRE_IMP_NEW;
897 spin_unlock(&imp->imp_lock);
899 rc = ptlrpc_connect_import(imp);
901 ptlrpc_request_free(request);
904 ptlrpc_pinger_add_import(imp);
908 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
909 req_capsule_set(&request->rq_pill, format);
914 * Allocate new request structure for import \a imp and initialize its
915 * buffer structure according to capsule template \a format.
917 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
918 const struct req_format *format)
920 return ptlrpc_request_alloc_internal(imp, NULL, format);
922 EXPORT_SYMBOL(ptlrpc_request_alloc);
925 * Allocate new request structure for import \a imp from pool \a pool and
926 * initialize its buffer structure according to capsule template \a format.
928 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
929 struct ptlrpc_request_pool * pool,
930 const struct req_format *format)
932 return ptlrpc_request_alloc_internal(imp, pool, format);
934 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
937 * For requests not from pool, free memory of the request structure.
938 * For requests obtained from a pool earlier, return request back to pool.
940 void ptlrpc_request_free(struct ptlrpc_request *request)
942 if (request->rq_pool)
943 __ptlrpc_free_req_to_pool(request);
945 ptlrpc_request_cache_free(request);
947 EXPORT_SYMBOL(ptlrpc_request_free);
950 * Allocate new request for operatione \a opcode and immediatelly pack it for
952 * Only used for simple requests like OBD_PING where the only important
953 * part of the request is operation itself.
954 * Returns allocated request or NULL on error.
956 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
957 const struct req_format *format,
958 __u32 version, int opcode)
960 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
964 rc = ptlrpc_request_pack(req, version, opcode);
966 ptlrpc_request_free(req);
972 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
975 * Allocate and initialize new request set structure on the current CPT.
976 * Returns a pointer to the newly allocated set structure or NULL on error.
978 struct ptlrpc_request_set *ptlrpc_prep_set(void)
980 struct ptlrpc_request_set *set;
984 cpt = cfs_cpt_current(cfs_cpt_table, 0);
985 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
988 atomic_set(&set->set_refcount, 1);
989 INIT_LIST_HEAD(&set->set_requests);
990 init_waitqueue_head(&set->set_waitq);
991 atomic_set(&set->set_new_count, 0);
992 atomic_set(&set->set_remaining, 0);
993 spin_lock_init(&set->set_new_req_lock);
994 INIT_LIST_HEAD(&set->set_new_requests);
995 set->set_max_inflight = UINT_MAX;
996 set->set_producer = NULL;
997 set->set_producer_arg = NULL;
1002 EXPORT_SYMBOL(ptlrpc_prep_set);
1005 * Allocate and initialize new request set structure with flow control
1006 * extension. This extension allows to control the number of requests in-flight
1007 * for the whole set. A callback function to generate requests must be provided
1008 * and the request set will keep the number of requests sent over the wire to
1010 * Returns a pointer to the newly allocated set structure or NULL on error.
1012 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1016 struct ptlrpc_request_set *set;
1018 set = ptlrpc_prep_set();
1022 set->set_max_inflight = max;
1023 set->set_producer = func;
1024 set->set_producer_arg = arg;
1030 * Wind down and free request set structure previously allocated with
1032 * Ensures that all requests on the set have completed and removes
1033 * all requests from the request list in a set.
1034 * If any unsent request happen to be on the list, pretends that they got
1035 * an error in flight and calls their completion handler.
1037 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1039 struct list_head *tmp;
1040 struct list_head *next;
1045 /* Requests on the set should either all be completed, or all be new */
1046 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1047 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1048 list_for_each(tmp, &set->set_requests) {
1049 struct ptlrpc_request *req =
1050 list_entry(tmp, struct ptlrpc_request,
1053 LASSERT(req->rq_phase == expected_phase);
1057 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1058 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1059 atomic_read(&set->set_remaining), n);
1061 list_for_each_safe(tmp, next, &set->set_requests) {
1062 struct ptlrpc_request *req =
1063 list_entry(tmp, struct ptlrpc_request,
1065 list_del_init(&req->rq_set_chain);
1067 LASSERT(req->rq_phase == expected_phase);
1069 if (req->rq_phase == RQ_PHASE_NEW) {
1070 ptlrpc_req_interpret(NULL, req, -EBADR);
1071 atomic_dec(&set->set_remaining);
1074 spin_lock(&req->rq_lock);
1076 req->rq_invalid_rqset = 0;
1077 spin_unlock(&req->rq_lock);
1079 ptlrpc_req_finished (req);
1082 LASSERT(atomic_read(&set->set_remaining) == 0);
1084 ptlrpc_reqset_put(set);
1087 EXPORT_SYMBOL(ptlrpc_set_destroy);
1090 * Add a new request to the general purpose request set.
1091 * Assumes request reference from the caller.
1093 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1094 struct ptlrpc_request *req)
1096 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1097 LASSERT(list_empty(&req->rq_set_chain));
1099 if (req->rq_allow_intr)
1100 set->set_allow_intr = 1;
1102 /* The set takes over the caller's request reference */
1103 list_add_tail(&req->rq_set_chain, &set->set_requests);
1105 atomic_inc(&set->set_remaining);
1106 req->rq_queued_time = ktime_get_seconds();
1108 if (req->rq_reqmsg != NULL)
1109 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1111 if (set->set_producer != NULL)
1112 /* If the request set has a producer callback, the RPC must be
1113 * sent straight away */
1114 ptlrpc_send_new_req(req);
1116 EXPORT_SYMBOL(ptlrpc_set_add_req);
1119 * Add a request to a request with dedicated server thread
1120 * and wake the thread to make any necessary processing.
1121 * Currently only used for ptlrpcd.
1123 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1124 struct ptlrpc_request *req)
1126 struct ptlrpc_request_set *set = pc->pc_set;
1129 LASSERT(req->rq_set == NULL);
1130 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1132 spin_lock(&set->set_new_req_lock);
1134 * The set takes over the caller's request reference.
1137 req->rq_queued_time = ktime_get_seconds();
1138 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1139 count = atomic_inc_return(&set->set_new_count);
1140 spin_unlock(&set->set_new_req_lock);
1142 /* Only need to call wakeup once for the first entry. */
1144 wake_up(&set->set_waitq);
1146 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1147 * guarantee the async RPC can be processed ASAP, we have
1148 * no other better choice. It maybe fixed in future. */
1149 for (i = 0; i < pc->pc_npartners; i++)
1150 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1155 * Based on the current state of the import, determine if the request
1156 * can be sent, is an error, or should be delayed.
1158 * Returns true if this request should be delayed. If false, and
1159 * *status is set, then the request can not be sent and *status is the
1160 * error code. If false and status is 0, then request can be sent.
1162 * The imp->imp_lock must be held.
1164 static int ptlrpc_import_delay_req(struct obd_import *imp,
1165 struct ptlrpc_request *req, int *status)
1170 LASSERT (status != NULL);
1173 if (req->rq_ctx_init || req->rq_ctx_fini) {
1174 /* always allow ctx init/fini rpc go through */
1175 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1176 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1178 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1179 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1181 /* pings or MDS-equivalent STATFS may safely race with umount */
1182 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1183 D_HA : D_ERROR, req, "IMP_CLOSED ");
1185 } else if (ptlrpc_send_limit_expired(req)) {
1186 /* probably doesn't need to be a D_ERROR after initial testing*/
1187 DEBUG_REQ(D_HA, req, "send limit expired ");
1188 *status = -ETIMEDOUT;
1189 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1190 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1191 /* allow CONNECT even if import is invalid */ ;
1192 if (atomic_read(&imp->imp_inval_count) != 0) {
1193 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1196 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1197 if (!imp->imp_deactive)
1198 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1199 *status = -ESHUTDOWN; /* bz 12940 */
1200 } else if (req->rq_import_generation != imp->imp_generation) {
1201 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1203 } else if (req->rq_send_state != imp->imp_state) {
1204 /* invalidate in progress - any requests should be drop */
1205 if (atomic_read(&imp->imp_inval_count) != 0) {
1206 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1208 } else if (req->rq_no_delay &&
1209 imp->imp_generation != imp->imp_initiated_at) {
1210 /* ignore nodelay for requests initiating connections */
1211 *status = -EWOULDBLOCK;
1212 } else if (req->rq_allow_replay &&
1213 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1214 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1215 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1216 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1217 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1227 * Decide if the error message should be printed to the console or not.
1228 * Makes its decision based on request type, status, and failure frequency.
1230 * \param[in] req request that failed and may need a console message
1232 * \retval false if no message should be printed
1233 * \retval true if console message should be printed
1235 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1237 LASSERT(req->rq_reqmsg != NULL);
1239 /* Suppress particular reconnect errors which are to be expected. */
1240 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1242 /* Suppress timed out reconnect requests */
1243 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1247 /* Suppress most unavailable/again reconnect requests, but
1248 * print occasionally so it is clear client is trying to
1249 * connect to a server where no target is running. */
1250 if ((err == -ENODEV || err == -EAGAIN) &&
1251 req->rq_import->imp_conn_cnt % 30 != 20)
1255 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1256 /* -EAGAIN is normal when using POSIX flocks */
1259 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1260 (req->rq_xid & 0xf) != 10)
1261 /* Suppress most ping requests, they may fail occasionally */
1268 * Check request processing status.
1269 * Returns the status.
1271 static int ptlrpc_check_status(struct ptlrpc_request *req)
1276 err = lustre_msg_get_status(req->rq_repmsg);
1277 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1278 struct obd_import *imp = req->rq_import;
1279 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1280 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1282 if (ptlrpc_console_allow(req, opc, err))
1283 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1284 "failed: rc = %d\n",
1285 imp->imp_obd->obd_name,
1287 libcfs_nid2str(nid), err);
1288 RETURN(err < 0 ? err : -EINVAL);
1292 DEBUG_REQ(D_INFO, req, "status is %d", err);
1293 } else if (err > 0) {
1294 /* XXX: translate this error from net to host */
1295 DEBUG_REQ(D_INFO, req, "status is %d", err);
1302 * save pre-versions of objects into request for replay.
1303 * Versions are obtained from server reply.
1306 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1308 struct lustre_msg *repmsg = req->rq_repmsg;
1309 struct lustre_msg *reqmsg = req->rq_reqmsg;
1310 __u64 *versions = lustre_msg_get_versions(repmsg);
1313 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1317 lustre_msg_set_versions(reqmsg, versions);
1318 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1319 versions[0], versions[1]);
1324 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1326 struct ptlrpc_request *req;
1328 assert_spin_locked(&imp->imp_lock);
1329 if (list_empty(&imp->imp_unreplied_list))
1332 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1334 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1336 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1337 imp->imp_known_replied_xid = req->rq_xid - 1;
1339 return req->rq_xid - 1;
1343 * Callback function called when client receives RPC reply for \a req.
1344 * Returns 0 on success or error code.
1345 * The return alue would be assigned to req->rq_status by the caller
1346 * as request processing status.
1347 * This function also decides if the request needs to be saved for later replay.
1349 static int after_reply(struct ptlrpc_request *req)
1351 struct obd_import *imp = req->rq_import;
1352 struct obd_device *obd = req->rq_import->imp_obd;
1359 LASSERT(obd != NULL);
1360 /* repbuf must be unlinked */
1361 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1363 if (req->rq_reply_truncated) {
1364 if (ptlrpc_no_resend(req)) {
1365 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1366 " expected: %d, actual size: %d",
1367 req->rq_nob_received, req->rq_repbuf_len);
1371 sptlrpc_cli_free_repbuf(req);
1372 /* Pass the required reply buffer size (include
1373 * space for early reply).
1374 * NB: no need to roundup because alloc_repbuf
1375 * will roundup it */
1376 req->rq_replen = req->rq_nob_received;
1377 req->rq_nob_received = 0;
1378 spin_lock(&req->rq_lock);
1380 spin_unlock(&req->rq_lock);
1384 work_start = ktime_get_real();
1385 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1388 * NB Until this point, the whole of the incoming message,
1389 * including buflens, status etc is in the sender's byte order.
1391 rc = sptlrpc_cli_unwrap_reply(req);
1393 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1398 * Security layer unwrap might ask resend this request.
1403 rc = unpack_reply(req);
1407 /* retry indefinitely on EINPROGRESS */
1408 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1409 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1410 time64_t now = ktime_get_real_seconds();
1412 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1413 spin_lock(&req->rq_lock);
1415 spin_unlock(&req->rq_lock);
1416 req->rq_nr_resend++;
1418 /* Readjust the timeout for current conditions */
1419 ptlrpc_at_set_req_timeout(req);
1420 /* delay resend to give a chance to the server to get ready.
1421 * The delay is increased by 1s on every resend and is capped to
1422 * the current request timeout (i.e. obd_timeout if AT is off,
1423 * or AT service time x 125% + 5s, see at_est2timeout) */
1424 if (req->rq_nr_resend > req->rq_timeout)
1425 req->rq_sent = now + req->rq_timeout;
1427 req->rq_sent = now + req->rq_nr_resend;
1429 /* Resend for EINPROGRESS will use a new XID */
1430 spin_lock(&imp->imp_lock);
1431 list_del_init(&req->rq_unreplied_list);
1432 spin_unlock(&imp->imp_lock);
1437 if (obd->obd_svc_stats != NULL) {
1438 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1440 ptlrpc_lprocfs_rpc_sent(req, timediff);
1443 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1444 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1445 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1446 lustre_msg_get_type(req->rq_repmsg));
1450 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1451 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1452 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1453 ptlrpc_at_adj_net_latency(req,
1454 lustre_msg_get_service_time(req->rq_repmsg));
1456 rc = ptlrpc_check_status(req);
1460 * Either we've been evicted, or the server has failed for
1461 * some reason. Try to reconnect, and if that fails, punt to
1464 if (ptlrpc_recoverable_error(rc)) {
1465 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1466 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1469 ptlrpc_request_handle_notconn(req);
1474 * Let's look if server sent slv. Do it only for RPC with
1477 ldlm_cli_update_pool(req);
1481 * Store transno in reqmsg for replay.
1483 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1484 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1485 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1488 if (imp->imp_replayable) {
1489 spin_lock(&imp->imp_lock);
1491 * No point in adding already-committed requests to the replay
1492 * list, we will just remove them immediately. b=9829
1494 if (req->rq_transno != 0 &&
1496 lustre_msg_get_last_committed(req->rq_repmsg) ||
1498 /** version recovery */
1499 ptlrpc_save_versions(req);
1500 ptlrpc_retain_replayable_request(req, imp);
1501 } else if (req->rq_commit_cb != NULL &&
1502 list_empty(&req->rq_replay_list)) {
1503 /* NB: don't call rq_commit_cb if it's already on
1504 * rq_replay_list, ptlrpc_free_committed() will call
1505 * it later, see LU-3618 for details */
1506 spin_unlock(&imp->imp_lock);
1507 req->rq_commit_cb(req);
1508 spin_lock(&imp->imp_lock);
1512 * Replay-enabled imports return commit-status information.
1514 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1515 if (likely(committed > imp->imp_peer_committed_transno))
1516 imp->imp_peer_committed_transno = committed;
1518 ptlrpc_free_committed(imp);
1520 if (!list_empty(&imp->imp_replay_list)) {
1521 struct ptlrpc_request *last;
1523 last = list_entry(imp->imp_replay_list.prev,
1524 struct ptlrpc_request,
1527 * Requests with rq_replay stay on the list even if no
1528 * commit is expected.
1530 if (last->rq_transno > imp->imp_peer_committed_transno)
1531 ptlrpc_pinger_commit_expected(imp);
1534 spin_unlock(&imp->imp_lock);
1541 * Helper function to send request \a req over the network for the first time
1542 * Also adjusts request phase.
1543 * Returns 0 on success or error code.
1545 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1547 struct obd_import *imp = req->rq_import;
1552 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1554 /* do not try to go further if there is not enough memory in enc_pool */
1555 if (req->rq_sent && req->rq_bulk != NULL)
1556 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1557 pool_is_at_full_capacity())
1560 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1561 (!req->rq_generation_set ||
1562 req->rq_import_generation == imp->imp_generation))
1565 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1567 spin_lock(&imp->imp_lock);
1569 LASSERT(req->rq_xid != 0);
1570 LASSERT(!list_empty(&req->rq_unreplied_list));
1572 if (!req->rq_generation_set)
1573 req->rq_import_generation = imp->imp_generation;
1575 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1576 spin_lock(&req->rq_lock);
1577 req->rq_waiting = 1;
1578 spin_unlock(&req->rq_lock);
1580 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1581 ptlrpc_import_state_name(req->rq_send_state),
1582 ptlrpc_import_state_name(imp->imp_state));
1583 LASSERT(list_empty(&req->rq_list));
1584 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1585 atomic_inc(&req->rq_import->imp_inflight);
1586 spin_unlock(&imp->imp_lock);
1591 spin_unlock(&imp->imp_lock);
1592 req->rq_status = rc;
1593 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1597 LASSERT(list_empty(&req->rq_list));
1598 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1599 atomic_inc(&req->rq_import->imp_inflight);
1601 /* find the known replied XID from the unreplied list, CONNECT
1602 * and DISCONNECT requests are skipped to make the sanity check
1603 * on server side happy. see process_req_last_xid().
1605 * For CONNECT: Because replay requests have lower XID, it'll
1606 * break the sanity check if CONNECT bump the exp_last_xid on
1609 * For DISCONNECT: Since client will abort inflight RPC before
1610 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1611 * than the inflight RPC.
1613 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1614 min_xid = ptlrpc_known_replied_xid(imp);
1615 spin_unlock(&imp->imp_lock);
1617 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1619 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1621 rc = sptlrpc_req_refresh_ctx(req, -1);
1624 req->rq_status = rc;
1627 spin_lock(&req->rq_lock);
1628 req->rq_wait_ctx = 1;
1629 spin_unlock(&req->rq_lock);
1634 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1635 " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1636 imp->imp_obd->obd_uuid.uuid,
1637 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1638 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg));
1640 rc = ptl_send_rpc(req, 0);
1641 if (rc == -ENOMEM) {
1642 spin_lock(&imp->imp_lock);
1643 if (!list_empty(&req->rq_list)) {
1644 list_del_init(&req->rq_list);
1645 atomic_dec(&req->rq_import->imp_inflight);
1647 spin_unlock(&imp->imp_lock);
1648 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1652 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1653 spin_lock(&req->rq_lock);
1654 req->rq_net_err = 1;
1655 spin_unlock(&req->rq_lock);
1661 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1666 LASSERT(set->set_producer != NULL);
1668 remaining = atomic_read(&set->set_remaining);
1670 /* populate the ->set_requests list with requests until we
1671 * reach the maximum number of RPCs in flight for this set */
1672 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1673 rc = set->set_producer(set, set->set_producer_arg);
1674 if (rc == -ENOENT) {
1675 /* no more RPC to produce */
1676 set->set_producer = NULL;
1677 set->set_producer_arg = NULL;
1682 RETURN((atomic_read(&set->set_remaining) - remaining));
1686 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1687 * and no more replies are expected.
1688 * (it is possible to get less replies than requests sent e.g. due to timed out
1689 * requests or requests that we had trouble to send out)
1691 * NOTE: This function contains a potential schedule point (cond_resched()).
1693 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1695 struct list_head *tmp, *next;
1696 struct list_head comp_reqs;
1697 int force_timer_recalc = 0;
1700 if (atomic_read(&set->set_remaining) == 0)
1703 INIT_LIST_HEAD(&comp_reqs);
1704 list_for_each_safe(tmp, next, &set->set_requests) {
1705 struct ptlrpc_request *req =
1706 list_entry(tmp, struct ptlrpc_request,
1708 struct obd_import *imp = req->rq_import;
1709 int unregistered = 0;
1713 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1714 list_move_tail(&req->rq_set_chain, &comp_reqs);
1718 /* This schedule point is mainly for the ptlrpcd caller of this
1719 * function. Most ptlrpc sets are not long-lived and unbounded
1720 * in length, but at the least the set used by the ptlrpcd is.
1721 * Since the processing time is unbounded, we need to insert an
1722 * explicit schedule point to make the thread well-behaved.
1726 /* If the caller requires to allow to be interpreted by force
1727 * and it has really been interpreted, then move the request
1728 * to RQ_PHASE_INTERPRET phase in spite of what the current
1730 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1731 req->rq_status = -EINTR;
1732 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1734 /* Since it is interpreted and we have to wait for
1735 * the reply to be unlinked, then use sync mode. */
1738 GOTO(interpret, req->rq_status);
1741 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1742 force_timer_recalc = 1;
1744 /* delayed send - skip */
1745 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1748 /* delayed resend - skip */
1749 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1750 req->rq_sent > ktime_get_real_seconds())
1753 if (!(req->rq_phase == RQ_PHASE_RPC ||
1754 req->rq_phase == RQ_PHASE_BULK ||
1755 req->rq_phase == RQ_PHASE_INTERPRET ||
1756 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1757 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1758 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1762 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1763 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1764 LASSERT(req->rq_next_phase != req->rq_phase);
1765 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1767 if (req->rq_req_deadline &&
1768 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1769 req->rq_req_deadline = 0;
1770 if (req->rq_reply_deadline &&
1771 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1772 req->rq_reply_deadline = 0;
1773 if (req->rq_bulk_deadline &&
1774 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1775 req->rq_bulk_deadline = 0;
1778 * Skip processing until reply is unlinked. We
1779 * can't return to pool before that and we can't
1780 * call interpret before that. We need to make
1781 * sure that all rdma transfers finished and will
1782 * not corrupt any data.
1784 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1785 ptlrpc_client_recv_or_unlink(req))
1787 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1788 ptlrpc_client_bulk_active(req))
1792 * Turn fail_loc off to prevent it from looping
1795 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1796 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1799 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1800 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1805 * Move to next phase if reply was successfully
1808 ptlrpc_rqphase_move(req, req->rq_next_phase);
1811 if (req->rq_phase == RQ_PHASE_INTERPRET)
1812 GOTO(interpret, req->rq_status);
1815 * Note that this also will start async reply unlink.
1817 if (req->rq_net_err && !req->rq_timedout) {
1818 ptlrpc_expire_one_request(req, 1);
1821 * Check if we still need to wait for unlink.
1823 if (ptlrpc_client_recv_or_unlink(req) ||
1824 ptlrpc_client_bulk_active(req))
1826 /* If there is no need to resend, fail it now. */
1827 if (req->rq_no_resend) {
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);
1838 spin_lock(&req->rq_lock);
1839 req->rq_replied = 0;
1840 spin_unlock(&req->rq_lock);
1841 if (req->rq_status == 0)
1842 req->rq_status = -EIO;
1843 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1844 GOTO(interpret, req->rq_status);
1847 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1848 * so it sets rq_intr regardless of individual rpc
1849 * timeouts. The synchronous IO waiting path sets
1850 * rq_intr irrespective of whether ptlrpcd
1851 * has seen a timeout. Our policy is to only interpret
1852 * interrupted rpcs after they have timed out, so we
1853 * need to enforce that here.
1856 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1857 req->rq_wait_ctx)) {
1858 req->rq_status = -EINTR;
1859 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1860 GOTO(interpret, req->rq_status);
1863 if (req->rq_phase == RQ_PHASE_RPC) {
1864 if (req->rq_timedout || req->rq_resend ||
1865 req->rq_waiting || req->rq_wait_ctx) {
1868 if (!ptlrpc_unregister_reply(req, 1)) {
1869 ptlrpc_unregister_bulk(req, 1);
1873 spin_lock(&imp->imp_lock);
1874 if (ptlrpc_import_delay_req(imp, req, &status)){
1875 /* put on delay list - only if we wait
1876 * recovery finished - before send */
1877 list_del_init(&req->rq_list);
1878 list_add_tail(&req->rq_list,
1881 spin_unlock(&imp->imp_lock);
1886 req->rq_status = status;
1887 ptlrpc_rqphase_move(req,
1888 RQ_PHASE_INTERPRET);
1889 spin_unlock(&imp->imp_lock);
1890 GOTO(interpret, req->rq_status);
1892 /* ignore on just initiated connections */
1893 if (ptlrpc_no_resend(req) &&
1894 !req->rq_wait_ctx &&
1895 imp->imp_generation !=
1896 imp->imp_initiated_at) {
1897 req->rq_status = -ENOTCONN;
1898 ptlrpc_rqphase_move(req,
1899 RQ_PHASE_INTERPRET);
1900 spin_unlock(&imp->imp_lock);
1901 GOTO(interpret, req->rq_status);
1904 list_del_init(&req->rq_list);
1905 list_add_tail(&req->rq_list,
1906 &imp->imp_sending_list);
1908 spin_unlock(&imp->imp_lock);
1910 spin_lock(&req->rq_lock);
1911 req->rq_waiting = 0;
1912 spin_unlock(&req->rq_lock);
1914 if (req->rq_timedout || req->rq_resend) {
1915 /* This is re-sending anyways,
1916 * let's mark req as resend. */
1917 spin_lock(&req->rq_lock);
1919 spin_unlock(&req->rq_lock);
1922 * rq_wait_ctx is only touched by ptlrpcd,
1923 * so no lock is needed here.
1925 status = sptlrpc_req_refresh_ctx(req, -1);
1928 req->rq_status = status;
1929 spin_lock(&req->rq_lock);
1930 req->rq_wait_ctx = 0;
1931 spin_unlock(&req->rq_lock);
1932 force_timer_recalc = 1;
1934 spin_lock(&req->rq_lock);
1935 req->rq_wait_ctx = 1;
1936 spin_unlock(&req->rq_lock);
1941 spin_lock(&req->rq_lock);
1942 req->rq_wait_ctx = 0;
1943 spin_unlock(&req->rq_lock);
1946 /* In any case, the previous bulk should be
1947 * cleaned up to prepare for the new sending */
1948 if (req->rq_bulk != NULL &&
1949 !ptlrpc_unregister_bulk(req, 1))
1952 rc = ptl_send_rpc(req, 0);
1953 if (rc == -ENOMEM) {
1954 spin_lock(&imp->imp_lock);
1955 if (!list_empty(&req->rq_list))
1956 list_del_init(&req->rq_list);
1957 spin_unlock(&imp->imp_lock);
1958 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1962 DEBUG_REQ(D_HA, req,
1963 "send failed: rc = %d", rc);
1964 force_timer_recalc = 1;
1965 spin_lock(&req->rq_lock);
1966 req->rq_net_err = 1;
1967 spin_unlock(&req->rq_lock);
1970 /* need to reset the timeout */
1971 force_timer_recalc = 1;
1974 spin_lock(&req->rq_lock);
1976 if (ptlrpc_client_early(req)) {
1977 ptlrpc_at_recv_early_reply(req);
1978 spin_unlock(&req->rq_lock);
1982 /* Still waiting for a reply? */
1983 if (ptlrpc_client_recv(req)) {
1984 spin_unlock(&req->rq_lock);
1988 /* Did we actually receive a reply? */
1989 if (!ptlrpc_client_replied(req)) {
1990 spin_unlock(&req->rq_lock);
1994 spin_unlock(&req->rq_lock);
1996 /* unlink from net because we are going to
1997 * swab in-place of reply buffer */
1998 unregistered = ptlrpc_unregister_reply(req, 1);
2002 req->rq_status = after_reply(req);
2006 /* If there is no bulk associated with this request,
2007 * then we're done and should let the interpreter
2008 * process the reply. Similarly if the RPC returned
2009 * an error, and therefore the bulk will never arrive.
2011 if (req->rq_bulk == NULL || req->rq_status < 0) {
2012 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2013 GOTO(interpret, req->rq_status);
2016 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2019 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2020 if (ptlrpc_client_bulk_active(req))
2023 if (req->rq_bulk->bd_failure) {
2024 /* The RPC reply arrived OK, but the bulk screwed
2025 * up! Dead weird since the server told us the RPC
2026 * was good after getting the REPLY for her GET or
2027 * the ACK for her PUT. */
2028 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2029 req->rq_status = -EIO;
2032 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2035 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2037 /* This moves to "unregistering" phase we need to wait for
2039 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2040 /* start async bulk unlink too */
2041 ptlrpc_unregister_bulk(req, 1);
2045 if (!ptlrpc_unregister_bulk(req, async))
2048 /* When calling interpret receiving already should be
2050 LASSERT(!req->rq_receiving_reply);
2052 ptlrpc_req_interpret(env, req, req->rq_status);
2054 if (ptlrpcd_check_work(req)) {
2055 atomic_dec(&set->set_remaining);
2058 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2060 if (req->rq_reqmsg != NULL)
2062 "Completed RPC pname:cluuid:pid:xid:nid:"
2063 "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
2064 imp->imp_obd->obd_uuid.uuid,
2065 lustre_msg_get_status(req->rq_reqmsg),
2067 obd_import_nid2str(imp),
2068 lustre_msg_get_opc(req->rq_reqmsg));
2070 spin_lock(&imp->imp_lock);
2071 /* Request already may be not on sending or delaying list. This
2072 * may happen in the case of marking it erroneous for the case
2073 * ptlrpc_import_delay_req(req, status) find it impossible to
2074 * allow sending this rpc and returns *status != 0. */
2075 if (!list_empty(&req->rq_list)) {
2076 list_del_init(&req->rq_list);
2077 atomic_dec(&imp->imp_inflight);
2079 list_del_init(&req->rq_unreplied_list);
2080 spin_unlock(&imp->imp_lock);
2082 atomic_dec(&set->set_remaining);
2083 wake_up_all(&imp->imp_recovery_waitq);
2085 if (set->set_producer) {
2086 /* produce a new request if possible */
2087 if (ptlrpc_set_producer(set) > 0)
2088 force_timer_recalc = 1;
2090 /* free the request that has just been completed
2091 * in order not to pollute set->set_requests */
2092 list_del_init(&req->rq_set_chain);
2093 spin_lock(&req->rq_lock);
2095 req->rq_invalid_rqset = 0;
2096 spin_unlock(&req->rq_lock);
2098 /* record rq_status to compute the final status later */
2099 if (req->rq_status != 0)
2100 set->set_rc = req->rq_status;
2101 ptlrpc_req_finished(req);
2103 list_move_tail(&req->rq_set_chain, &comp_reqs);
2107 /* move completed request at the head of list so it's easier for
2108 * caller to find them */
2109 list_splice(&comp_reqs, &set->set_requests);
2111 /* If we hit an error, we want to recover promptly. */
2112 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2114 EXPORT_SYMBOL(ptlrpc_check_set);
2117 * Time out request \a req. is \a async_unlink is set, that means do not wait
2118 * until LNet actually confirms network buffer unlinking.
2119 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2121 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2123 struct obd_import *imp = req->rq_import;
2124 unsigned int debug_mask = D_RPCTRACE;
2128 spin_lock(&req->rq_lock);
2129 req->rq_timedout = 1;
2130 spin_unlock(&req->rq_lock);
2132 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2133 lustre_msg_get_status(req->rq_reqmsg)))
2134 debug_mask = D_WARNING;
2135 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2136 req->rq_net_err ? "failed due to network error" :
2137 ((req->rq_real_sent == 0 ||
2138 req->rq_real_sent < req->rq_sent ||
2139 req->rq_real_sent >= req->rq_deadline) ?
2140 "timed out for sent delay" : "timed out for slow reply"),
2141 (s64)req->rq_sent, (s64)req->rq_real_sent);
2143 if (imp != NULL && obd_debug_peer_on_timeout)
2144 LNetDebugPeer(imp->imp_connection->c_peer);
2146 ptlrpc_unregister_reply(req, async_unlink);
2147 ptlrpc_unregister_bulk(req, async_unlink);
2149 if (obd_dump_on_timeout)
2150 libcfs_debug_dumplog();
2153 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2157 atomic_inc(&imp->imp_timeouts);
2159 /* The DLM server doesn't want recovery run on its imports. */
2160 if (imp->imp_dlm_fake)
2163 /* If this request is for recovery or other primordial tasks,
2164 * then error it out here. */
2165 if (req->rq_ctx_init || req->rq_ctx_fini ||
2166 req->rq_send_state != LUSTRE_IMP_FULL ||
2167 imp->imp_obd->obd_no_recov) {
2168 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2169 ptlrpc_import_state_name(req->rq_send_state),
2170 ptlrpc_import_state_name(imp->imp_state));
2171 spin_lock(&req->rq_lock);
2172 req->rq_status = -ETIMEDOUT;
2174 spin_unlock(&req->rq_lock);
2178 /* if a request can't be resent we can't wait for an answer after
2180 if (ptlrpc_no_resend(req)) {
2181 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2185 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2191 * Time out all uncompleted requests in request set pointed by \a data
2192 * Callback used when waiting on sets with l_wait_event.
2195 int ptlrpc_expired_set(void *data)
2197 struct ptlrpc_request_set *set = data;
2198 struct list_head *tmp;
2199 time64_t now = ktime_get_real_seconds();
2202 LASSERT(set != NULL);
2205 * A timeout expired. See which reqs it applies to...
2207 list_for_each(tmp, &set->set_requests) {
2208 struct ptlrpc_request *req =
2209 list_entry(tmp, struct ptlrpc_request,
2212 /* don't expire request waiting for context */
2213 if (req->rq_wait_ctx)
2216 /* Request in-flight? */
2217 if (!((req->rq_phase == RQ_PHASE_RPC &&
2218 !req->rq_waiting && !req->rq_resend) ||
2219 (req->rq_phase == RQ_PHASE_BULK)))
2222 if (req->rq_timedout || /* already dealt with */
2223 req->rq_deadline > now) /* not expired */
2226 /* Deal with this guy. Do it asynchronously to not block
2227 * ptlrpcd thread. */
2228 ptlrpc_expire_one_request(req, 1);
2232 * When waiting for a whole set, we always break out of the
2233 * sleep so we can recalculate the timeout, or enable interrupts
2234 * if everyone's timed out.
2240 * Sets rq_intr flag in \a req under spinlock.
2242 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2244 spin_lock(&req->rq_lock);
2246 spin_unlock(&req->rq_lock);
2248 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2251 * Interrupts (sets interrupted flag) all uncompleted requests in
2252 * a set \a data. Callback for l_wait_event for interruptible waits.
2254 static void ptlrpc_interrupted_set(void *data)
2256 struct ptlrpc_request_set *set = data;
2257 struct list_head *tmp;
2259 LASSERT(set != NULL);
2260 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2262 list_for_each(tmp, &set->set_requests) {
2263 struct ptlrpc_request *req =
2264 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2269 if (req->rq_phase != RQ_PHASE_RPC &&
2270 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2271 !req->rq_allow_intr)
2274 ptlrpc_mark_interrupted(req);
2279 * Get the smallest timeout in the set; this does NOT set a timeout.
2281 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2283 struct list_head *tmp;
2284 time64_t now = ktime_get_real_seconds();
2286 struct ptlrpc_request *req;
2290 list_for_each(tmp, &set->set_requests) {
2291 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2294 * Request in-flight?
2296 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2297 (req->rq_phase == RQ_PHASE_BULK) ||
2298 (req->rq_phase == RQ_PHASE_NEW)))
2302 * Already timed out.
2304 if (req->rq_timedout)
2310 if (req->rq_wait_ctx)
2313 if (req->rq_phase == RQ_PHASE_NEW)
2314 deadline = req->rq_sent;
2315 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2316 deadline = req->rq_sent;
2318 deadline = req->rq_sent + req->rq_timeout;
2320 if (deadline <= now) /* actually expired already */
2321 timeout = 1; /* ASAP */
2322 else if (timeout == 0 || timeout > deadline - now)
2323 timeout = deadline - now;
2329 * Send all unset request from the set and then wait untill all
2330 * requests in the set complete (either get a reply, timeout, get an
2331 * error or otherwise be interrupted).
2332 * Returns 0 on success or error code otherwise.
2334 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2336 struct list_head *tmp;
2337 struct ptlrpc_request *req;
2338 struct l_wait_info lwi;
2344 if (set->set_producer)
2345 (void)ptlrpc_set_producer(set);
2347 list_for_each(tmp, &set->set_requests) {
2348 req = list_entry(tmp, struct ptlrpc_request,
2350 if (req->rq_phase == RQ_PHASE_NEW)
2351 (void)ptlrpc_send_new_req(req);
2354 if (list_empty(&set->set_requests))
2357 /* ideally we want env provide by the caller all the time,
2358 * but at the moment that would mean a massive change in
2359 * LDLM while benefits would be close to zero, so just
2360 * initialize env here for those rare cases */
2362 /* XXX: skip on the client side? */
2363 rc = lu_env_init(&_env, LCT_DT_THREAD);
2370 timeout = ptlrpc_set_next_timeout(set);
2372 /* wait until all complete, interrupted, or an in-flight
2374 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2377 if ((timeout == 0 && !signal_pending(current)) ||
2378 set->set_allow_intr)
2379 /* No requests are in-flight (ether timed out
2380 * or delayed), so we can allow interrupts.
2381 * We still want to block for a limited time,
2382 * so we allow interrupts during the timeout. */
2383 lwi = LWI_TIMEOUT_INTR_ALL(
2384 cfs_time_seconds(timeout ? timeout : 1),
2386 ptlrpc_interrupted_set, set);
2389 * At least one request is in flight, so no
2390 * interrupts are allowed. Wait until all
2391 * complete, or an in-flight req times out.
2393 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2394 ptlrpc_expired_set, set);
2396 rc = l_wait_event(set->set_waitq,
2397 ptlrpc_check_set(env, set), &lwi);
2399 /* LU-769 - if we ignored the signal because it was already
2400 * pending when we started, we need to handle it now or we risk
2401 * it being ignored forever */
2402 if (rc == -ETIMEDOUT &&
2403 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2404 signal_pending(current)) {
2405 sigset_t blocked_sigs =
2406 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2408 /* In fact we only interrupt for the "fatal" signals
2409 * like SIGINT or SIGKILL. We still ignore less
2410 * important signals since ptlrpc set is not easily
2411 * reentrant from userspace again */
2412 if (signal_pending(current))
2413 ptlrpc_interrupted_set(set);
2414 cfs_restore_sigs(blocked_sigs);
2417 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2419 /* -EINTR => all requests have been flagged rq_intr so next
2421 * -ETIMEDOUT => someone timed out. When all reqs have
2422 * timed out, signals are enabled allowing completion with
2424 * I don't really care if we go once more round the loop in
2425 * the error cases -eeb. */
2426 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2427 list_for_each(tmp, &set->set_requests) {
2428 req = list_entry(tmp, struct ptlrpc_request,
2430 spin_lock(&req->rq_lock);
2431 req->rq_invalid_rqset = 1;
2432 spin_unlock(&req->rq_lock);
2435 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2437 LASSERT(atomic_read(&set->set_remaining) == 0);
2439 rc = set->set_rc; /* rq_status of already freed requests if any */
2440 list_for_each(tmp, &set->set_requests) {
2441 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2443 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2444 if (req->rq_status != 0)
2445 rc = req->rq_status;
2448 if (env && env == &_env)
2453 EXPORT_SYMBOL(ptlrpc_set_wait);
2456 * Helper fuction for request freeing.
2457 * Called when request count reached zero and request needs to be freed.
2458 * Removes request from all sorts of sending/replay lists it might be on,
2459 * frees network buffers if any are present.
2460 * If \a locked is set, that means caller is already holding import imp_lock
2461 * and so we no longer need to reobtain it (for certain lists manipulations)
2463 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2467 if (request == NULL)
2470 LASSERT(!request->rq_srv_req);
2471 LASSERT(request->rq_export == NULL);
2472 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2473 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2474 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2475 LASSERTF(!request->rq_replay, "req %p\n", request);
2477 req_capsule_fini(&request->rq_pill);
2479 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2480 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2481 if (request->rq_import != NULL) {
2483 spin_lock(&request->rq_import->imp_lock);
2484 list_del_init(&request->rq_replay_list);
2485 list_del_init(&request->rq_unreplied_list);
2487 spin_unlock(&request->rq_import->imp_lock);
2489 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2491 if (atomic_read(&request->rq_refcount) != 0) {
2492 DEBUG_REQ(D_ERROR, request,
2493 "freeing request with nonzero refcount");
2497 if (request->rq_repbuf != NULL)
2498 sptlrpc_cli_free_repbuf(request);
2500 if (request->rq_import != NULL) {
2501 class_import_put(request->rq_import);
2502 request->rq_import = NULL;
2504 if (request->rq_bulk != NULL)
2505 ptlrpc_free_bulk(request->rq_bulk);
2507 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2508 sptlrpc_cli_free_reqbuf(request);
2510 if (request->rq_cli_ctx)
2511 sptlrpc_req_put_ctx(request, !locked);
2513 if (request->rq_pool)
2514 __ptlrpc_free_req_to_pool(request);
2516 ptlrpc_request_cache_free(request);
2520 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2522 * Drop one request reference. Must be called with import imp_lock held.
2523 * When reference count drops to zero, request is freed.
2525 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2527 assert_spin_locked(&request->rq_import->imp_lock);
2528 (void)__ptlrpc_req_finished(request, 1);
2533 * Drops one reference count for request \a request.
2534 * \a locked set indicates that caller holds import imp_lock.
2535 * Frees the request whe reference count reaches zero.
2537 * \retval 1 the request is freed
2538 * \retval 0 some others still hold references on the request
2540 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2548 LASSERT(request != LP_POISON);
2549 LASSERT(request->rq_reqmsg != LP_POISON);
2551 DEBUG_REQ(D_INFO, request, "refcount now %u",
2552 atomic_read(&request->rq_refcount) - 1);
2554 spin_lock(&request->rq_lock);
2555 count = atomic_dec_return(&request->rq_refcount);
2556 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2558 /* For open RPC, the client does not know the EA size (LOV, ACL, and
2559 * so on) before replied, then the client has to reserve very large
2560 * reply buffer. Such buffer will not be released until the RPC freed.
2561 * Since The open RPC is replayable, we need to keep it in the replay
2562 * list until close. If there are a lot of files opened concurrently,
2563 * then the client may be OOM.
2565 * If fact, it is unnecessary to keep reply buffer for open replay,
2566 * related EAs have already been saved via mdc_save_lovea() before
2567 * coming here. So it is safe to free the reply buffer some earlier
2568 * before releasing the RPC to avoid client OOM. LU-9514 */
2569 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2570 spin_lock(&request->rq_early_free_lock);
2571 sptlrpc_cli_free_repbuf(request);
2572 request->rq_repbuf = NULL;
2573 request->rq_repbuf_len = 0;
2574 request->rq_repdata = NULL;
2575 request->rq_reqdata_len = 0;
2576 spin_unlock(&request->rq_early_free_lock);
2578 spin_unlock(&request->rq_lock);
2581 __ptlrpc_free_req(request, locked);
2587 * Drops one reference count for a request.
2589 void ptlrpc_req_finished(struct ptlrpc_request *request)
2591 __ptlrpc_req_finished(request, 0);
2593 EXPORT_SYMBOL(ptlrpc_req_finished);
2596 * Returns xid of a \a request
2598 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2600 return request->rq_xid;
2602 EXPORT_SYMBOL(ptlrpc_req_xid);
2605 * Disengage the client's reply buffer from the network
2606 * NB does _NOT_ unregister any client-side bulk.
2607 * IDEMPOTENT, but _not_ safe against concurrent callers.
2608 * The request owner (i.e. the thread doing the I/O) must call...
2609 * Returns 0 on success or 1 if unregistering cannot be made.
2611 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2614 struct l_wait_info lwi;
2619 LASSERT(!in_interrupt());
2621 /* Let's setup deadline for reply unlink. */
2622 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2623 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2624 request->rq_reply_deadline = ktime_get_real_seconds() +
2628 * Nothing left to do.
2630 if (!ptlrpc_client_recv_or_unlink(request))
2633 LNetMDUnlink(request->rq_reply_md_h);
2636 * Let's check it once again.
2638 if (!ptlrpc_client_recv_or_unlink(request))
2641 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2642 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2645 * Do not wait for unlink to finish.
2651 * We have to l_wait_event() whatever the result, to give liblustre
2652 * a chance to run reply_in_callback(), and to make sure we've
2653 * unlinked before returning a req to the pool.
2656 /* The wq argument is ignored by user-space wait_event macros */
2657 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2658 &request->rq_set->set_waitq :
2659 &request->rq_reply_waitq;
2660 /* Network access will complete in finite time but the HUGE
2661 * timeout lets us CWARN for visibility of sluggish NALs */
2662 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2663 cfs_time_seconds(1), NULL, NULL);
2664 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2667 ptlrpc_rqphase_move(request, request->rq_next_phase);
2671 LASSERT(rc == -ETIMEDOUT);
2672 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2673 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2674 request->rq_receiving_reply,
2675 request->rq_req_unlinked,
2676 request->rq_reply_unlinked);
2681 static void ptlrpc_free_request(struct ptlrpc_request *req)
2683 spin_lock(&req->rq_lock);
2685 spin_unlock(&req->rq_lock);
2687 if (req->rq_commit_cb != NULL)
2688 req->rq_commit_cb(req);
2689 list_del_init(&req->rq_replay_list);
2691 __ptlrpc_req_finished(req, 1);
2695 * the request is committed and dropped from the replay list of its import
2697 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2699 struct obd_import *imp = req->rq_import;
2701 spin_lock(&imp->imp_lock);
2702 if (list_empty(&req->rq_replay_list)) {
2703 spin_unlock(&imp->imp_lock);
2707 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2708 if (imp->imp_replay_cursor == &req->rq_replay_list)
2709 imp->imp_replay_cursor = req->rq_replay_list.next;
2710 ptlrpc_free_request(req);
2713 spin_unlock(&imp->imp_lock);
2715 EXPORT_SYMBOL(ptlrpc_request_committed);
2718 * Iterates through replay_list on import and prunes
2719 * all requests have transno smaller than last_committed for the
2720 * import and don't have rq_replay set.
2721 * Since requests are sorted in transno order, stops when meetign first
2722 * transno bigger than last_committed.
2723 * caller must hold imp->imp_lock
2725 void ptlrpc_free_committed(struct obd_import *imp)
2727 struct ptlrpc_request *req, *saved;
2728 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2729 bool skip_committed_list = true;
2732 LASSERT(imp != NULL);
2733 assert_spin_locked(&imp->imp_lock);
2735 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2736 imp->imp_generation == imp->imp_last_generation_checked) {
2737 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2738 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2741 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2742 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2743 imp->imp_generation);
2745 if (imp->imp_generation != imp->imp_last_generation_checked ||
2746 imp->imp_last_transno_checked == 0)
2747 skip_committed_list = false;
2749 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2750 imp->imp_last_generation_checked = imp->imp_generation;
2752 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2754 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2755 LASSERT(req != last_req);
2758 if (req->rq_transno == 0) {
2759 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2762 if (req->rq_import_generation < imp->imp_generation) {
2763 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2767 /* not yet committed */
2768 if (req->rq_transno > imp->imp_peer_committed_transno) {
2769 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2773 if (req->rq_replay) {
2774 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2775 list_move_tail(&req->rq_replay_list,
2776 &imp->imp_committed_list);
2780 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2781 imp->imp_peer_committed_transno);
2783 ptlrpc_free_request(req);
2786 if (skip_committed_list)
2789 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2791 LASSERT(req->rq_transno != 0);
2792 if (req->rq_import_generation < imp->imp_generation ||
2794 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2795 req->rq_import_generation <
2796 imp->imp_generation ? "stale" : "closed");
2798 if (imp->imp_replay_cursor == &req->rq_replay_list)
2799 imp->imp_replay_cursor =
2800 req->rq_replay_list.next;
2802 ptlrpc_free_request(req);
2809 void ptlrpc_cleanup_client(struct obd_import *imp)
2816 * Schedule previously sent request for resend.
2817 * For bulk requests we assign new xid (to avoid problems with
2818 * lost replies and therefore several transfers landing into same buffer
2819 * from different sending attempts).
2821 void ptlrpc_resend_req(struct ptlrpc_request *req)
2823 DEBUG_REQ(D_HA, req, "going to resend");
2824 spin_lock(&req->rq_lock);
2826 /* Request got reply but linked to the import list still.
2827 Let ptlrpc_check_set() to process it. */
2828 if (ptlrpc_client_replied(req)) {
2829 spin_unlock(&req->rq_lock);
2830 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2834 req->rq_status = -EAGAIN;
2837 req->rq_net_err = 0;
2838 req->rq_timedout = 0;
2840 ptlrpc_client_wake_req(req);
2841 spin_unlock(&req->rq_lock);
2844 /* XXX: this function and rq_status are currently unused */
2845 void ptlrpc_restart_req(struct ptlrpc_request *req)
2847 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2848 req->rq_status = -ERESTARTSYS;
2850 spin_lock(&req->rq_lock);
2851 req->rq_restart = 1;
2852 req->rq_timedout = 0;
2853 ptlrpc_client_wake_req(req);
2854 spin_unlock(&req->rq_lock);
2858 * Grab additional reference on a request \a req
2860 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2863 atomic_inc(&req->rq_refcount);
2866 EXPORT_SYMBOL(ptlrpc_request_addref);
2869 * Add a request to import replay_list.
2870 * Must be called under imp_lock
2872 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2873 struct obd_import *imp)
2875 struct list_head *tmp;
2877 assert_spin_locked(&imp->imp_lock);
2879 if (req->rq_transno == 0) {
2880 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2884 /* clear this for new requests that were resent as well
2885 as resent replayed requests. */
2886 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2888 /* don't re-add requests that have been replayed */
2889 if (!list_empty(&req->rq_replay_list))
2892 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2894 spin_lock(&req->rq_lock);
2896 spin_unlock(&req->rq_lock);
2898 LASSERT(imp->imp_replayable);
2899 /* Balanced in ptlrpc_free_committed, usually. */
2900 ptlrpc_request_addref(req);
2901 list_for_each_prev(tmp, &imp->imp_replay_list) {
2902 struct ptlrpc_request *iter = list_entry(tmp,
2903 struct ptlrpc_request,
2906 /* We may have duplicate transnos if we create and then
2907 * open a file, or for closes retained if to match creating
2908 * opens, so use req->rq_xid as a secondary key.
2909 * (See bugs 684, 685, and 428.)
2910 * XXX no longer needed, but all opens need transnos!
2912 if (iter->rq_transno > req->rq_transno)
2915 if (iter->rq_transno == req->rq_transno) {
2916 LASSERT(iter->rq_xid != req->rq_xid);
2917 if (iter->rq_xid > req->rq_xid)
2921 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2925 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2929 * Send request and wait until it completes.
2930 * Returns request processing status.
2932 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2934 struct ptlrpc_request_set *set;
2938 LASSERT(req->rq_set == NULL);
2939 LASSERT(!req->rq_receiving_reply);
2941 set = ptlrpc_prep_set();
2943 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2947 /* for distributed debugging */
2948 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2950 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2951 ptlrpc_request_addref(req);
2952 ptlrpc_set_add_req(set, req);
2953 rc = ptlrpc_set_wait(NULL, set);
2954 ptlrpc_set_destroy(set);
2958 EXPORT_SYMBOL(ptlrpc_queue_wait);
2961 * Callback used for replayed requests reply processing.
2962 * In case of successful reply calls registered request replay callback.
2963 * In case of error restart replay process.
2965 static int ptlrpc_replay_interpret(const struct lu_env *env,
2966 struct ptlrpc_request *req,
2969 struct ptlrpc_replay_async_args *aa = args;
2970 struct obd_import *imp = req->rq_import;
2973 atomic_dec(&imp->imp_replay_inflight);
2975 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2976 * server got the request, but bulk transfer timeout, let's
2977 * replay the bulk req again */
2978 if (!ptlrpc_client_replied(req) ||
2979 (req->rq_bulk != NULL &&
2980 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2981 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2982 GOTO(out, rc = -ETIMEDOUT);
2985 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2986 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2987 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2988 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2990 /** VBR: check version failure */
2991 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2992 /** replay was failed due to version mismatch */
2993 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2994 spin_lock(&imp->imp_lock);
2995 imp->imp_vbr_failed = 1;
2996 spin_unlock(&imp->imp_lock);
2997 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2999 /** The transno had better not change over replay. */
3000 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3001 lustre_msg_get_transno(req->rq_repmsg) ||
3002 lustre_msg_get_transno(req->rq_repmsg) == 0,
3004 lustre_msg_get_transno(req->rq_reqmsg),
3005 lustre_msg_get_transno(req->rq_repmsg));
3008 spin_lock(&imp->imp_lock);
3009 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3010 spin_unlock(&imp->imp_lock);
3011 LASSERT(imp->imp_last_replay_transno);
3013 /* transaction number shouldn't be bigger than the latest replayed */
3014 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3015 DEBUG_REQ(D_ERROR, req,
3016 "Reported transno %llu is bigger than the "
3017 "replayed one: %llu", req->rq_transno,
3018 lustre_msg_get_transno(req->rq_reqmsg));
3019 GOTO(out, rc = -EINVAL);
3022 DEBUG_REQ(D_HA, req, "got rep");
3024 /* let the callback do fixups, possibly including in the request */
3025 if (req->rq_replay_cb)
3026 req->rq_replay_cb(req);
3028 if (ptlrpc_client_replied(req) &&
3029 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3030 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3031 lustre_msg_get_status(req->rq_repmsg),
3032 aa->praa_old_status);
3034 /* Note: If the replay fails for MDT-MDT recovery, let's
3035 * abort all of the following requests in the replay
3036 * and sending list, because MDT-MDT update requests
3037 * are dependent on each other, see LU-7039 */
3038 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3039 struct ptlrpc_request *free_req;
3040 struct ptlrpc_request *tmp;
3042 spin_lock(&imp->imp_lock);
3043 list_for_each_entry_safe(free_req, tmp,
3044 &imp->imp_replay_list,
3046 ptlrpc_free_request(free_req);
3049 list_for_each_entry_safe(free_req, tmp,
3050 &imp->imp_committed_list,
3052 ptlrpc_free_request(free_req);
3055 list_for_each_entry_safe(free_req, tmp,
3056 &imp->imp_delayed_list,
3058 spin_lock(&free_req->rq_lock);
3059 free_req->rq_err = 1;
3060 free_req->rq_status = -EIO;
3061 ptlrpc_client_wake_req(free_req);
3062 spin_unlock(&free_req->rq_lock);
3065 list_for_each_entry_safe(free_req, tmp,
3066 &imp->imp_sending_list,
3068 spin_lock(&free_req->rq_lock);
3069 free_req->rq_err = 1;
3070 free_req->rq_status = -EIO;
3071 ptlrpc_client_wake_req(free_req);
3072 spin_unlock(&free_req->rq_lock);
3074 spin_unlock(&imp->imp_lock);
3077 /* Put it back for re-replay. */
3078 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3082 * Errors while replay can set transno to 0, but
3083 * imp_last_replay_transno shouldn't be set to 0 anyway
3085 if (req->rq_transno == 0)
3086 CERROR("Transno is 0 during replay!\n");
3088 /* continue with recovery */
3089 rc = ptlrpc_import_recovery_state_machine(imp);
3091 req->rq_send_state = aa->praa_old_state;
3094 /* this replay failed, so restart recovery */
3095 ptlrpc_connect_import(imp);
3101 * Prepares and queues request for replay.
3102 * Adds it to ptlrpcd queue for actual sending.
3103 * Returns 0 on success.
3105 int ptlrpc_replay_req(struct ptlrpc_request *req)
3107 struct ptlrpc_replay_async_args *aa;
3111 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3113 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3114 aa = ptlrpc_req_async_args(req);
3115 memset(aa, 0, sizeof(*aa));
3117 /* Prepare request to be resent with ptlrpcd */
3118 aa->praa_old_state = req->rq_send_state;
3119 req->rq_send_state = LUSTRE_IMP_REPLAY;
3120 req->rq_phase = RQ_PHASE_NEW;
3121 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3123 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3125 req->rq_interpret_reply = ptlrpc_replay_interpret;
3126 /* Readjust the timeout for current conditions */
3127 ptlrpc_at_set_req_timeout(req);
3129 /* Tell server net_latency to calculate how long to wait for reply. */
3130 lustre_msg_set_service_time(req->rq_reqmsg,
3131 ptlrpc_at_get_net_latency(req));
3132 DEBUG_REQ(D_HA, req, "REPLAY");
3134 atomic_inc(&req->rq_import->imp_replay_inflight);
3135 spin_lock(&req->rq_lock);
3136 req->rq_early_free_repbuf = 0;
3137 spin_unlock(&req->rq_lock);
3138 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3140 ptlrpcd_add_req(req);
3145 * Aborts all in-flight request on import \a imp sending and delayed lists
3147 void ptlrpc_abort_inflight(struct obd_import *imp)
3149 struct list_head *tmp, *n;
3152 /* Make sure that no new requests get processed for this import.
3153 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3154 * this flag and then putting requests on sending_list or delayed_list.
3156 spin_lock(&imp->imp_lock);
3158 /* XXX locking? Maybe we should remove each request with the list
3159 * locked? Also, how do we know if the requests on the list are
3160 * being freed at this time?
3162 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3163 struct ptlrpc_request *req = list_entry(tmp,
3164 struct ptlrpc_request,
3167 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3169 spin_lock(&req->rq_lock);
3170 if (req->rq_import_generation < imp->imp_generation) {
3172 req->rq_status = -EIO;
3173 ptlrpc_client_wake_req(req);
3175 spin_unlock(&req->rq_lock);
3178 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3179 struct ptlrpc_request *req =
3180 list_entry(tmp, struct ptlrpc_request, rq_list);
3182 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3184 spin_lock(&req->rq_lock);
3185 if (req->rq_import_generation < imp->imp_generation) {
3187 req->rq_status = -EIO;
3188 ptlrpc_client_wake_req(req);
3190 spin_unlock(&req->rq_lock);
3193 /* Last chance to free reqs left on the replay list, but we
3194 * will still leak reqs that haven't committed. */
3195 if (imp->imp_replayable)
3196 ptlrpc_free_committed(imp);
3198 spin_unlock(&imp->imp_lock);
3204 * Abort all uncompleted requests in request set \a set
3206 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3208 struct list_head *tmp, *pos;
3210 LASSERT(set != NULL);
3212 list_for_each_safe(pos, tmp, &set->set_requests) {
3213 struct ptlrpc_request *req =
3214 list_entry(pos, struct ptlrpc_request,
3217 spin_lock(&req->rq_lock);
3218 if (req->rq_phase != RQ_PHASE_RPC) {
3219 spin_unlock(&req->rq_lock);
3224 req->rq_status = -EINTR;
3225 ptlrpc_client_wake_req(req);
3226 spin_unlock(&req->rq_lock);
3231 * Initialize the XID for the node. This is common among all requests on
3232 * this node, and only requires the property that it is monotonically
3233 * increasing. It does not need to be sequential. Since this is also used
3234 * as the RDMA match bits, it is important that a single client NOT have
3235 * the same match bits for two different in-flight requests, hence we do
3236 * NOT want to have an XID per target or similar.
3238 * To avoid an unlikely collision between match bits after a client reboot
3239 * (which would deliver old data into the wrong RDMA buffer) initialize
3240 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3241 * If the time is clearly incorrect, we instead use a 62-bit random number.
3242 * In the worst case the random number will overflow 1M RPCs per second in
3243 * 9133 years, or permutations thereof.
3245 #define YEAR_2004 (1ULL << 30)
3246 void ptlrpc_init_xid(void)
3248 time64_t now = ktime_get_real_seconds();
3250 spin_lock_init(&ptlrpc_last_xid_lock);
3251 if (now < YEAR_2004) {
3252 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3253 ptlrpc_last_xid >>= 2;
3254 ptlrpc_last_xid |= (1ULL << 61);
3256 ptlrpc_last_xid = (__u64)now << 20;
3259 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3260 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3261 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3265 * Increase xid and returns resulting new value to the caller.
3267 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3268 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3269 * itself uses the last bulk xid needed, so the server can determine the
3270 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3271 * xid must align to a power-of-two value.
3273 * This is assumed to be true due to the initial ptlrpc_last_xid
3274 * value also being initialized to a power-of-two value. LU-1431
3276 __u64 ptlrpc_next_xid(void)
3280 spin_lock(&ptlrpc_last_xid_lock);
3281 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3282 ptlrpc_last_xid = next;
3283 spin_unlock(&ptlrpc_last_xid_lock);
3289 * If request has a new allocated XID (new request or EINPROGRESS resend),
3290 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3291 * request to ensure previous bulk fails and avoid problems with lost replies
3292 * and therefore several transfers landing into the same buffer from different
3295 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3297 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3299 LASSERT(bd != NULL);
3301 /* Generate new matchbits for all resend requests, including
3303 if (req->rq_resend) {
3304 __u64 old_mbits = req->rq_mbits;
3306 /* First time resend on -EINPROGRESS will generate new xid,
3307 * so we can actually use the rq_xid as rq_mbits in such case,
3308 * however, it's bit hard to distinguish such resend with a
3309 * 'resend for the -EINPROGRESS resend'. To make it simple,
3310 * we opt to generate mbits for all resend cases. */
3311 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
3312 req->rq_mbits = ptlrpc_next_xid();
3314 /* Old version transfers rq_xid to peer as
3316 spin_lock(&req->rq_import->imp_lock);
3317 list_del_init(&req->rq_unreplied_list);
3318 ptlrpc_assign_next_xid_nolock(req);
3319 spin_unlock(&req->rq_import->imp_lock);
3320 req->rq_mbits = req->rq_xid;
3322 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3323 old_mbits, req->rq_mbits);
3324 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3325 /* Request being sent first time, use xid as matchbits. */
3326 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)
3327 || req->rq_mbits == 0) {
3328 req->rq_mbits = req->rq_xid;
3330 int total_md = (bd->bd_iov_count + LNET_MAX_IOV - 1) /
3332 req->rq_mbits -= total_md - 1;
3335 /* Replay request, xid and matchbits have already been
3336 * correctly assigned. */
3340 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3341 * that server can infer the number of bulks that were prepared,
3343 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3346 /* Set rq_xid as rq_mbits to indicate the final bulk for the old
3347 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3349 * It's ok to directly set the rq_xid here, since this xid bump
3350 * won't affect the request position in unreplied list. */
3351 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3352 req->rq_xid = req->rq_mbits;
3356 * Get a glimpse at what next xid value might have been.
3357 * Returns possible next xid.
3359 __u64 ptlrpc_sample_next_xid(void)
3361 #if BITS_PER_LONG == 32
3362 /* need to avoid possible word tearing on 32-bit systems */
3365 spin_lock(&ptlrpc_last_xid_lock);
3366 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3367 spin_unlock(&ptlrpc_last_xid_lock);
3371 /* No need to lock, since returned value is racy anyways */
3372 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3375 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3378 * Functions for operating ptlrpc workers.
3380 * A ptlrpc work is a function which will be running inside ptlrpc context.
3381 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3383 * 1. after a work is created, it can be used many times, that is:
3384 * handler = ptlrpcd_alloc_work();
3385 * ptlrpcd_queue_work();
3387 * queue it again when necessary:
3388 * ptlrpcd_queue_work();
3389 * ptlrpcd_destroy_work();
3390 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3391 * it will only be queued once in any time. Also as its name implies, it may
3392 * have delay before it really runs by ptlrpcd thread.
3394 struct ptlrpc_work_async_args {
3395 int (*cb)(const struct lu_env *, void *);
3399 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3401 /* re-initialize the req */
3402 req->rq_timeout = obd_timeout;
3403 req->rq_sent = ktime_get_real_seconds();
3404 req->rq_deadline = req->rq_sent + req->rq_timeout;
3405 req->rq_phase = RQ_PHASE_INTERPRET;
3406 req->rq_next_phase = RQ_PHASE_COMPLETE;
3407 req->rq_xid = ptlrpc_next_xid();
3408 req->rq_import_generation = req->rq_import->imp_generation;
3410 ptlrpcd_add_req(req);
3413 static int work_interpreter(const struct lu_env *env,
3414 struct ptlrpc_request *req, void *args, int rc)
3416 struct ptlrpc_work_async_args *arg = args;
3418 LASSERT(ptlrpcd_check_work(req));
3419 LASSERT(arg->cb != NULL);
3421 rc = arg->cb(env, arg->cbdata);
3423 list_del_init(&req->rq_set_chain);
3426 if (atomic_dec_return(&req->rq_refcount) > 1) {
3427 atomic_set(&req->rq_refcount, 2);
3428 ptlrpcd_add_work_req(req);
3433 static int worker_format;
3435 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3437 return req->rq_pill.rc_fmt == (void *)&worker_format;
3441 * Create a work for ptlrpc.
3443 void *ptlrpcd_alloc_work(struct obd_import *imp,
3444 int (*cb)(const struct lu_env *, void *), void *cbdata)
3446 struct ptlrpc_request *req = NULL;
3447 struct ptlrpc_work_async_args *args;
3453 RETURN(ERR_PTR(-EINVAL));
3455 /* copy some code from deprecated fakereq. */
3456 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3458 CERROR("ptlrpc: run out of memory!\n");
3459 RETURN(ERR_PTR(-ENOMEM));
3462 ptlrpc_cli_req_init(req);
3464 req->rq_send_state = LUSTRE_IMP_FULL;
3465 req->rq_type = PTL_RPC_MSG_REQUEST;
3466 req->rq_import = class_import_get(imp);
3467 req->rq_interpret_reply = work_interpreter;
3468 /* don't want reply */
3469 req->rq_no_delay = req->rq_no_resend = 1;
3470 req->rq_pill.rc_fmt = (void *)&worker_format;
3472 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3473 args = ptlrpc_req_async_args(req);
3475 args->cbdata = cbdata;
3479 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3481 void ptlrpcd_destroy_work(void *handler)
3483 struct ptlrpc_request *req = handler;
3486 ptlrpc_req_finished(req);
3488 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3490 int ptlrpcd_queue_work(void *handler)
3492 struct ptlrpc_request *req = handler;
3495 * Check if the req is already being queued.
3497 * Here comes a trick: it lacks a way of checking if a req is being
3498 * processed reliably in ptlrpc. Here I have to use refcount of req
3499 * for this purpose. This is okay because the caller should use this
3500 * req as opaque data. - Jinshan
3502 LASSERT(atomic_read(&req->rq_refcount) > 0);
3503 if (atomic_inc_return(&req->rq_refcount) == 2)
3504 ptlrpcd_add_work_req(req);
3507 EXPORT_SYMBOL(ptlrpcd_queue_work);