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_refs == 0); /* network hands off */
263 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
264 LASSERT(desc->bd_frag_ops != NULL);
266 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
267 sptlrpc_enc_pool_put_pages(desc);
270 class_export_put(desc->bd_export);
272 class_import_put(desc->bd_import);
274 if (desc->bd_frag_ops->release_frags != NULL)
275 desc->bd_frag_ops->release_frags(desc);
277 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
278 OBD_FREE_LARGE(GET_KIOV(desc),
279 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
281 OBD_FREE_LARGE(GET_KVEC(desc),
282 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
286 EXPORT_SYMBOL(ptlrpc_free_bulk);
289 * Set server timelimit for this req, i.e. how long are we willing to wait
290 * for reply before timing out this request.
292 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
298 LASSERT(req->rq_import);
301 /* non-AT settings */
303 * \a imp_server_timeout means this is reverse import and
304 * we send (currently only) ASTs to the client and cannot afford
305 * to wait too long for the reply, otherwise the other client
306 * (because of which we are sending this request) would
307 * timeout waiting for us
309 req->rq_timeout = req->rq_import->imp_server_timeout ?
310 obd_timeout / 2 : obd_timeout;
312 at = &req->rq_import->imp_at;
313 idx = import_at_get_index(req->rq_import,
314 req->rq_request_portal);
315 serv_est = at_get(&at->iat_service_estimate[idx]);
316 req->rq_timeout = at_est2timeout(serv_est);
318 /* We could get even fancier here, using history to predict increased
321 /* Let the server know what this RPC timeout is by putting it in the
323 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
325 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
327 /* Adjust max service estimate based on server value */
328 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
329 unsigned int serv_est)
335 LASSERT(req->rq_import);
336 at = &req->rq_import->imp_at;
338 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
339 /* max service estimates are tracked on the server side,
340 so just keep minimal history here */
341 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
343 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
344 "has changed from %d to %d\n",
345 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
346 oldse, at_get(&at->iat_service_estimate[idx]));
349 /* Expected network latency per remote node (secs) */
350 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
352 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
355 /* Adjust expected network latency */
356 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
357 unsigned int service_time)
359 unsigned int nl, oldnl;
361 time64_t now = ktime_get_real_seconds();
363 LASSERT(req->rq_import);
365 if (service_time > now - req->rq_sent + 3) {
366 /* bz16408, however, this can also happen if early reply
367 * is lost and client RPC is expired and resent, early reply
368 * or reply of original RPC can still be fit in reply buffer
369 * of resent RPC, now client is measuring time from the
370 * resent time, but server sent back service time of original
373 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
374 D_ADAPTTO : D_WARNING,
375 "Reported service time %u > total measured time %lld\n",
376 service_time, now - req->rq_sent);
380 /* Network latency is total time less server processing time */
381 nl = max_t(int, now - req->rq_sent -
382 service_time, 0) + 1; /* st rounding */
383 at = &req->rq_import->imp_at;
385 oldnl = at_measured(&at->iat_net_latency, nl);
387 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
388 "has changed from %d to %d\n",
389 req->rq_import->imp_obd->obd_name,
391 &req->rq_import->imp_connection->c_remote_uuid),
392 oldnl, at_get(&at->iat_net_latency));
395 static int unpack_reply(struct ptlrpc_request *req)
399 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
400 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
402 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
407 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
409 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
416 * Handle an early reply message, called with the rq_lock held.
417 * If anything goes wrong just ignore it - same as if it never happened
419 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
420 __must_hold(&req->rq_lock)
422 struct ptlrpc_request *early_req;
428 spin_unlock(&req->rq_lock);
430 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
432 spin_lock(&req->rq_lock);
436 rc = unpack_reply(early_req);
438 sptlrpc_cli_finish_early_reply(early_req);
439 spin_lock(&req->rq_lock);
443 /* Use new timeout value just to adjust the local value for this
444 * request, don't include it into at_history. It is unclear yet why
445 * service time increased and should it be counted or skipped, e.g.
446 * that can be recovery case or some error or server, the real reply
447 * will add all new data if it is worth to add. */
448 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
449 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
451 /* Network latency can be adjusted, it is pure network delays */
452 ptlrpc_at_adj_net_latency(req,
453 lustre_msg_get_service_time(early_req->rq_repmsg));
455 sptlrpc_cli_finish_early_reply(early_req);
457 spin_lock(&req->rq_lock);
458 olddl = req->rq_deadline;
459 /* server assumes it now has rq_timeout from when the request
460 * arrived, so the client should give it at least that long.
461 * since we don't know the arrival time we'll use the original
463 req->rq_deadline = req->rq_sent + req->rq_timeout +
464 ptlrpc_at_get_net_latency(req);
466 DEBUG_REQ(D_ADAPTTO, req,
467 "Early reply #%d, new deadline in %llds (%llds)",
469 req->rq_deadline - ktime_get_real_seconds(),
470 req->rq_deadline - olddl);
475 static struct kmem_cache *request_cache;
477 int ptlrpc_request_cache_init(void)
479 request_cache = kmem_cache_create("ptlrpc_cache",
480 sizeof(struct ptlrpc_request),
481 0, SLAB_HWCACHE_ALIGN, NULL);
482 return request_cache == NULL ? -ENOMEM : 0;
485 void ptlrpc_request_cache_fini(void)
487 kmem_cache_destroy(request_cache);
490 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
492 struct ptlrpc_request *req;
494 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
498 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
500 OBD_SLAB_FREE_PTR(req, request_cache);
504 * Wind down request pool \a pool.
505 * Frees all requests from the pool too
507 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
509 struct list_head *l, *tmp;
510 struct ptlrpc_request *req;
512 LASSERT(pool != NULL);
514 spin_lock(&pool->prp_lock);
515 list_for_each_safe(l, tmp, &pool->prp_req_list) {
516 req = list_entry(l, struct ptlrpc_request, rq_list);
517 list_del(&req->rq_list);
518 LASSERT(req->rq_reqbuf);
519 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
520 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
521 ptlrpc_request_cache_free(req);
523 spin_unlock(&pool->prp_lock);
524 OBD_FREE(pool, sizeof(*pool));
526 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
529 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
531 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
536 while (size < pool->prp_rq_size)
539 LASSERTF(list_empty(&pool->prp_req_list) ||
540 size == pool->prp_rq_size,
541 "Trying to change pool size with nonempty pool "
542 "from %d to %d bytes\n", pool->prp_rq_size, size);
544 spin_lock(&pool->prp_lock);
545 pool->prp_rq_size = size;
546 for (i = 0; i < num_rq; i++) {
547 struct ptlrpc_request *req;
548 struct lustre_msg *msg;
550 spin_unlock(&pool->prp_lock);
551 req = ptlrpc_request_cache_alloc(GFP_NOFS);
554 OBD_ALLOC_LARGE(msg, size);
556 ptlrpc_request_cache_free(req);
559 req->rq_reqbuf = msg;
560 req->rq_reqbuf_len = size;
562 spin_lock(&pool->prp_lock);
563 list_add_tail(&req->rq_list, &pool->prp_req_list);
565 spin_unlock(&pool->prp_lock);
568 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
571 * Create and initialize new request pool with given attributes:
572 * \a num_rq - initial number of requests to create for the pool
573 * \a msgsize - maximum message size possible for requests in thid pool
574 * \a populate_pool - function to be called when more requests need to be added
576 * Returns pointer to newly created pool or NULL on error.
578 struct ptlrpc_request_pool *
579 ptlrpc_init_rq_pool(int num_rq, int msgsize,
580 int (*populate_pool)(struct ptlrpc_request_pool *, int))
582 struct ptlrpc_request_pool *pool;
584 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
588 /* Request next power of two for the allocation, because internally
589 kernel would do exactly this */
591 spin_lock_init(&pool->prp_lock);
592 INIT_LIST_HEAD(&pool->prp_req_list);
593 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
594 pool->prp_populate = populate_pool;
596 populate_pool(pool, num_rq);
600 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
603 * Fetches one request from pool \a pool
605 static struct ptlrpc_request *
606 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
608 struct ptlrpc_request *request;
609 struct lustre_msg *reqbuf;
614 spin_lock(&pool->prp_lock);
616 /* See if we have anything in a pool, and bail out if nothing,
617 * in writeout path, where this matters, this is safe to do, because
618 * nothing is lost in this case, and when some in-flight requests
619 * complete, this code will be called again. */
620 if (unlikely(list_empty(&pool->prp_req_list))) {
621 spin_unlock(&pool->prp_lock);
625 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
627 list_del_init(&request->rq_list);
628 spin_unlock(&pool->prp_lock);
630 LASSERT(request->rq_reqbuf);
631 LASSERT(request->rq_pool);
633 reqbuf = request->rq_reqbuf;
634 memset(request, 0, sizeof(*request));
635 request->rq_reqbuf = reqbuf;
636 request->rq_reqbuf_len = pool->prp_rq_size;
637 request->rq_pool = pool;
643 * Returns freed \a request to pool.
645 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
647 struct ptlrpc_request_pool *pool = request->rq_pool;
649 spin_lock(&pool->prp_lock);
650 LASSERT(list_empty(&request->rq_list));
651 LASSERT(!request->rq_receiving_reply);
652 list_add_tail(&request->rq_list, &pool->prp_req_list);
653 spin_unlock(&pool->prp_lock);
656 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
658 struct obd_import *imp = req->rq_import;
659 struct list_head *tmp;
660 struct ptlrpc_request *iter;
662 assert_spin_locked(&imp->imp_lock);
663 LASSERT(list_empty(&req->rq_unreplied_list));
665 /* unreplied list is sorted by xid in ascending order */
666 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
667 iter = list_entry(tmp, struct ptlrpc_request,
670 LASSERT(req->rq_xid != iter->rq_xid);
671 if (req->rq_xid < iter->rq_xid)
673 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
676 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
679 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
681 req->rq_xid = ptlrpc_next_xid();
682 ptlrpc_add_unreplied(req);
685 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
687 spin_lock(&req->rq_import->imp_lock);
688 ptlrpc_assign_next_xid_nolock(req);
689 spin_unlock(&req->rq_import->imp_lock);
692 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
693 __u32 version, int opcode, char **bufs,
694 struct ptlrpc_cli_ctx *ctx)
697 struct obd_import *imp;
703 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
704 imp = request->rq_import;
705 lengths = request->rq_pill.rc_area[RCL_CLIENT];
708 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
710 rc = sptlrpc_req_get_ctx(request);
714 sptlrpc_req_set_flavor(request, opcode);
716 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
721 lustre_msg_add_version(request->rq_reqmsg, version);
722 request->rq_send_state = LUSTRE_IMP_FULL;
723 request->rq_type = PTL_RPC_MSG_REQUEST;
725 request->rq_req_cbid.cbid_fn = request_out_callback;
726 request->rq_req_cbid.cbid_arg = request;
728 request->rq_reply_cbid.cbid_fn = reply_in_callback;
729 request->rq_reply_cbid.cbid_arg = request;
731 request->rq_reply_deadline = 0;
732 request->rq_bulk_deadline = 0;
733 request->rq_req_deadline = 0;
734 request->rq_phase = RQ_PHASE_NEW;
735 request->rq_next_phase = RQ_PHASE_UNDEFINED;
737 request->rq_request_portal = imp->imp_client->cli_request_portal;
738 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
740 ptlrpc_at_set_req_timeout(request);
742 lustre_msg_set_opc(request->rq_reqmsg, opcode);
743 ptlrpc_assign_next_xid(request);
745 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
746 if (cfs_fail_val == opcode) {
747 time64_t *fail_t = NULL, *fail2_t = NULL;
749 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
750 fail_t = &request->rq_bulk_deadline;
751 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
752 fail_t = &request->rq_reply_deadline;
753 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
754 fail_t = &request->rq_req_deadline;
755 else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
756 fail_t = &request->rq_reply_deadline;
757 fail2_t = &request->rq_bulk_deadline;
761 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
764 *fail2_t = ktime_get_real_seconds() +
768 * The RPC is infected, let the test to change the
771 msleep(4 * MSEC_PER_SEC);
778 LASSERT(!request->rq_pool);
779 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
781 atomic_dec(&imp->imp_reqs);
782 class_import_put(imp);
787 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
790 * Pack request buffers for network transfer, performing necessary encryption
791 * steps if necessary.
793 int ptlrpc_request_pack(struct ptlrpc_request *request,
794 __u32 version, int opcode)
797 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
801 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
802 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
803 * have to send old ptlrpc_body to keep interoprability with these
806 * Only three kinds of server->client RPCs so far:
811 * XXX This should be removed whenever we drop the interoprability with
812 * the these old clients.
814 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
815 opcode == LDLM_GL_CALLBACK)
816 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
817 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
821 EXPORT_SYMBOL(ptlrpc_request_pack);
824 * Helper function to allocate new request on import \a imp
825 * and possibly using existing request from pool \a pool if provided.
826 * Returns allocated request structure with import field filled or
830 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
831 struct ptlrpc_request_pool *pool)
833 struct ptlrpc_request *request = NULL;
835 request = ptlrpc_request_cache_alloc(GFP_NOFS);
837 if (!request && pool)
838 request = ptlrpc_prep_req_from_pool(pool);
841 ptlrpc_cli_req_init(request);
843 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
844 LASSERT(imp != LP_POISON);
845 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
847 LASSERT(imp->imp_client != LP_POISON);
849 request->rq_import = class_import_get(imp);
850 atomic_inc(&imp->imp_reqs);
852 CERROR("request allocation out of memory\n");
858 static int ptlrpc_reconnect_if_idle(struct obd_import *imp)
863 * initiate connection if needed when the import has been
864 * referenced by the new request to avoid races with disconnect.
865 * serialize this check against conditional state=IDLE
866 * in ptlrpc_disconnect_idle_interpret()
868 spin_lock(&imp->imp_lock);
869 if (imp->imp_state == LUSTRE_IMP_IDLE) {
870 imp->imp_generation++;
871 imp->imp_initiated_at = imp->imp_generation;
872 imp->imp_state = LUSTRE_IMP_NEW;
874 /* connect_import_locked releases imp_lock */
875 rc = ptlrpc_connect_import_locked(imp);
878 ptlrpc_pinger_add_import(imp);
880 spin_unlock(&imp->imp_lock);
886 * Helper function for creating a request.
887 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
888 * buffer structures according to capsule template \a format.
889 * Returns allocated request structure pointer or NULL on error.
891 static struct ptlrpc_request *
892 ptlrpc_request_alloc_internal(struct obd_import *imp,
893 struct ptlrpc_request_pool * pool,
894 const struct req_format *format)
896 struct ptlrpc_request *request;
898 request = __ptlrpc_request_alloc(imp, pool);
902 /* don't make expensive check for idling connection
903 * if it's already connected */
904 if (unlikely(imp->imp_state != LUSTRE_IMP_FULL)) {
905 if (ptlrpc_reconnect_if_idle(imp) < 0) {
906 atomic_dec(&imp->imp_reqs);
907 ptlrpc_request_free(request);
912 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
913 req_capsule_set(&request->rq_pill, format);
918 * Allocate new request structure for import \a imp and initialize its
919 * buffer structure according to capsule template \a format.
921 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
922 const struct req_format *format)
924 return ptlrpc_request_alloc_internal(imp, NULL, format);
926 EXPORT_SYMBOL(ptlrpc_request_alloc);
929 * Allocate new request structure for import \a imp from pool \a pool and
930 * initialize its buffer structure according to capsule template \a format.
932 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
933 struct ptlrpc_request_pool * pool,
934 const struct req_format *format)
936 return ptlrpc_request_alloc_internal(imp, pool, format);
938 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
941 * For requests not from pool, free memory of the request structure.
942 * For requests obtained from a pool earlier, return request back to pool.
944 void ptlrpc_request_free(struct ptlrpc_request *request)
946 if (request->rq_pool)
947 __ptlrpc_free_req_to_pool(request);
949 ptlrpc_request_cache_free(request);
951 EXPORT_SYMBOL(ptlrpc_request_free);
954 * Allocate new request for operatione \a opcode and immediatelly pack it for
956 * Only used for simple requests like OBD_PING where the only important
957 * part of the request is operation itself.
958 * Returns allocated request or NULL on error.
960 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
961 const struct req_format *format,
962 __u32 version, int opcode)
964 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
968 rc = ptlrpc_request_pack(req, version, opcode);
970 ptlrpc_request_free(req);
976 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
979 * Allocate and initialize new request set structure on the current CPT.
980 * Returns a pointer to the newly allocated set structure or NULL on error.
982 struct ptlrpc_request_set *ptlrpc_prep_set(void)
984 struct ptlrpc_request_set *set;
988 cpt = cfs_cpt_current(cfs_cpt_table, 0);
989 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
992 atomic_set(&set->set_refcount, 1);
993 INIT_LIST_HEAD(&set->set_requests);
994 init_waitqueue_head(&set->set_waitq);
995 atomic_set(&set->set_new_count, 0);
996 atomic_set(&set->set_remaining, 0);
997 spin_lock_init(&set->set_new_req_lock);
998 INIT_LIST_HEAD(&set->set_new_requests);
999 set->set_max_inflight = UINT_MAX;
1000 set->set_producer = NULL;
1001 set->set_producer_arg = NULL;
1006 EXPORT_SYMBOL(ptlrpc_prep_set);
1009 * Allocate and initialize new request set structure with flow control
1010 * extension. This extension allows to control the number of requests in-flight
1011 * for the whole set. A callback function to generate requests must be provided
1012 * and the request set will keep the number of requests sent over the wire to
1014 * Returns a pointer to the newly allocated set structure or NULL on error.
1016 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1020 struct ptlrpc_request_set *set;
1022 set = ptlrpc_prep_set();
1026 set->set_max_inflight = max;
1027 set->set_producer = func;
1028 set->set_producer_arg = arg;
1034 * Wind down and free request set structure previously allocated with
1036 * Ensures that all requests on the set have completed and removes
1037 * all requests from the request list in a set.
1038 * If any unsent request happen to be on the list, pretends that they got
1039 * an error in flight and calls their completion handler.
1041 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1043 struct list_head *tmp;
1044 struct list_head *next;
1049 /* Requests on the set should either all be completed, or all be new */
1050 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1051 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1052 list_for_each(tmp, &set->set_requests) {
1053 struct ptlrpc_request *req =
1054 list_entry(tmp, struct ptlrpc_request,
1057 LASSERT(req->rq_phase == expected_phase);
1061 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1062 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1063 atomic_read(&set->set_remaining), n);
1065 list_for_each_safe(tmp, next, &set->set_requests) {
1066 struct ptlrpc_request *req =
1067 list_entry(tmp, struct ptlrpc_request,
1069 list_del_init(&req->rq_set_chain);
1071 LASSERT(req->rq_phase == expected_phase);
1073 if (req->rq_phase == RQ_PHASE_NEW) {
1074 ptlrpc_req_interpret(NULL, req, -EBADR);
1075 atomic_dec(&set->set_remaining);
1078 spin_lock(&req->rq_lock);
1080 req->rq_invalid_rqset = 0;
1081 spin_unlock(&req->rq_lock);
1083 ptlrpc_req_finished (req);
1086 LASSERT(atomic_read(&set->set_remaining) == 0);
1088 ptlrpc_reqset_put(set);
1091 EXPORT_SYMBOL(ptlrpc_set_destroy);
1094 * Add a new request to the general purpose request set.
1095 * Assumes request reference from the caller.
1097 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1098 struct ptlrpc_request *req)
1100 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1101 LASSERT(list_empty(&req->rq_set_chain));
1103 if (req->rq_allow_intr)
1104 set->set_allow_intr = 1;
1106 /* The set takes over the caller's request reference */
1107 list_add_tail(&req->rq_set_chain, &set->set_requests);
1109 atomic_inc(&set->set_remaining);
1110 req->rq_queued_time = ktime_get_seconds();
1112 if (req->rq_reqmsg != NULL)
1113 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1115 if (set->set_producer != NULL)
1116 /* If the request set has a producer callback, the RPC must be
1117 * sent straight away */
1118 ptlrpc_send_new_req(req);
1120 EXPORT_SYMBOL(ptlrpc_set_add_req);
1123 * Add a request to a request with dedicated server thread
1124 * and wake the thread to make any necessary processing.
1125 * Currently only used for ptlrpcd.
1127 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1128 struct ptlrpc_request *req)
1130 struct ptlrpc_request_set *set = pc->pc_set;
1133 LASSERT(req->rq_set == NULL);
1134 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1136 spin_lock(&set->set_new_req_lock);
1138 * The set takes over the caller's request reference.
1141 req->rq_queued_time = ktime_get_seconds();
1142 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1143 count = atomic_inc_return(&set->set_new_count);
1144 spin_unlock(&set->set_new_req_lock);
1146 /* Only need to call wakeup once for the first entry. */
1148 wake_up(&set->set_waitq);
1150 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1151 * guarantee the async RPC can be processed ASAP, we have
1152 * no other better choice. It maybe fixed in future. */
1153 for (i = 0; i < pc->pc_npartners; i++)
1154 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1159 * Based on the current state of the import, determine if the request
1160 * can be sent, is an error, or should be delayed.
1162 * Returns true if this request should be delayed. If false, and
1163 * *status is set, then the request can not be sent and *status is the
1164 * error code. If false and status is 0, then request can be sent.
1166 * The imp->imp_lock must be held.
1168 static int ptlrpc_import_delay_req(struct obd_import *imp,
1169 struct ptlrpc_request *req, int *status)
1174 LASSERT (status != NULL);
1177 if (req->rq_ctx_init || req->rq_ctx_fini) {
1178 /* always allow ctx init/fini rpc go through */
1179 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1180 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1182 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1183 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1185 /* pings or MDS-equivalent STATFS may safely race with umount */
1186 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1187 D_HA : D_ERROR, req, "IMP_CLOSED ");
1189 } else if (ptlrpc_send_limit_expired(req)) {
1190 /* probably doesn't need to be a D_ERROR after initial testing*/
1191 DEBUG_REQ(D_HA, req, "send limit expired ");
1192 *status = -ETIMEDOUT;
1193 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1194 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1195 /* allow CONNECT even if import is invalid */ ;
1196 if (atomic_read(&imp->imp_inval_count) != 0) {
1197 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1200 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1201 if (!imp->imp_deactive)
1202 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1203 *status = -ESHUTDOWN; /* bz 12940 */
1204 } else if (req->rq_import_generation != imp->imp_generation) {
1205 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1207 } else if (req->rq_send_state != imp->imp_state) {
1208 /* invalidate in progress - any requests should be drop */
1209 if (atomic_read(&imp->imp_inval_count) != 0) {
1210 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1212 } else if (req->rq_no_delay &&
1213 imp->imp_generation != imp->imp_initiated_at) {
1214 /* ignore nodelay for requests initiating connections */
1215 *status = -EWOULDBLOCK;
1216 } else if (req->rq_allow_replay &&
1217 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1218 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1219 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1220 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1221 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1231 * Decide if the error message should be printed to the console or not.
1232 * Makes its decision based on request type, status, and failure frequency.
1234 * \param[in] req request that failed and may need a console message
1236 * \retval false if no message should be printed
1237 * \retval true if console message should be printed
1239 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1241 LASSERT(req->rq_reqmsg != NULL);
1243 /* Suppress particular reconnect errors which are to be expected. */
1244 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1246 /* Suppress timed out reconnect requests */
1247 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1251 /* Suppress most unavailable/again reconnect requests, but
1252 * print occasionally so it is clear client is trying to
1253 * connect to a server where no target is running. */
1254 if ((err == -ENODEV || err == -EAGAIN) &&
1255 req->rq_import->imp_conn_cnt % 30 != 20)
1259 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1260 /* -EAGAIN is normal when using POSIX flocks */
1263 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1264 (req->rq_xid & 0xf) != 10)
1265 /* Suppress most ping requests, they may fail occasionally */
1272 * Check request processing status.
1273 * Returns the status.
1275 static int ptlrpc_check_status(struct ptlrpc_request *req)
1280 err = lustre_msg_get_status(req->rq_repmsg);
1281 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1282 struct obd_import *imp = req->rq_import;
1283 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1284 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1286 if (ptlrpc_console_allow(req, opc, err))
1287 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1288 "failed: rc = %d\n",
1289 imp->imp_obd->obd_name,
1291 libcfs_nid2str(nid), err);
1292 RETURN(err < 0 ? err : -EINVAL);
1296 DEBUG_REQ(D_INFO, req, "status is %d", err);
1297 } else if (err > 0) {
1298 /* XXX: translate this error from net to host */
1299 DEBUG_REQ(D_INFO, req, "status is %d", err);
1306 * save pre-versions of objects into request for replay.
1307 * Versions are obtained from server reply.
1310 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1312 struct lustre_msg *repmsg = req->rq_repmsg;
1313 struct lustre_msg *reqmsg = req->rq_reqmsg;
1314 __u64 *versions = lustre_msg_get_versions(repmsg);
1317 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1321 lustre_msg_set_versions(reqmsg, versions);
1322 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1323 versions[0], versions[1]);
1328 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1330 struct ptlrpc_request *req;
1332 assert_spin_locked(&imp->imp_lock);
1333 if (list_empty(&imp->imp_unreplied_list))
1336 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1338 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1340 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1341 imp->imp_known_replied_xid = req->rq_xid - 1;
1343 return req->rq_xid - 1;
1347 * Callback function called when client receives RPC reply for \a req.
1348 * Returns 0 on success or error code.
1349 * The return alue would be assigned to req->rq_status by the caller
1350 * as request processing status.
1351 * This function also decides if the request needs to be saved for later replay.
1353 static int after_reply(struct ptlrpc_request *req)
1355 struct obd_import *imp = req->rq_import;
1356 struct obd_device *obd = req->rq_import->imp_obd;
1363 LASSERT(obd != NULL);
1364 /* repbuf must be unlinked */
1365 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1367 if (req->rq_reply_truncated) {
1368 if (ptlrpc_no_resend(req)) {
1369 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1370 " expected: %d, actual size: %d",
1371 req->rq_nob_received, req->rq_repbuf_len);
1375 sptlrpc_cli_free_repbuf(req);
1376 /* Pass the required reply buffer size (include
1377 * space for early reply).
1378 * NB: no need to roundup because alloc_repbuf
1379 * will roundup it */
1380 req->rq_replen = req->rq_nob_received;
1381 req->rq_nob_received = 0;
1382 spin_lock(&req->rq_lock);
1384 spin_unlock(&req->rq_lock);
1388 work_start = ktime_get_real();
1389 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1392 * NB Until this point, the whole of the incoming message,
1393 * including buflens, status etc is in the sender's byte order.
1395 rc = sptlrpc_cli_unwrap_reply(req);
1397 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1402 * Security layer unwrap might ask resend this request.
1407 rc = unpack_reply(req);
1411 /* retry indefinitely on EINPROGRESS */
1412 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1413 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1414 time64_t now = ktime_get_real_seconds();
1416 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1417 spin_lock(&req->rq_lock);
1419 spin_unlock(&req->rq_lock);
1420 req->rq_nr_resend++;
1422 /* Readjust the timeout for current conditions */
1423 ptlrpc_at_set_req_timeout(req);
1424 /* delay resend to give a chance to the server to get ready.
1425 * The delay is increased by 1s on every resend and is capped to
1426 * the current request timeout (i.e. obd_timeout if AT is off,
1427 * or AT service time x 125% + 5s, see at_est2timeout) */
1428 if (req->rq_nr_resend > req->rq_timeout)
1429 req->rq_sent = now + req->rq_timeout;
1431 req->rq_sent = now + req->rq_nr_resend;
1433 /* Resend for EINPROGRESS will use a new XID */
1434 spin_lock(&imp->imp_lock);
1435 list_del_init(&req->rq_unreplied_list);
1436 spin_unlock(&imp->imp_lock);
1441 if (obd->obd_svc_stats != NULL) {
1442 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1444 ptlrpc_lprocfs_rpc_sent(req, timediff);
1447 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1448 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1449 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1450 lustre_msg_get_type(req->rq_repmsg));
1454 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1455 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1456 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1457 ptlrpc_at_adj_net_latency(req,
1458 lustre_msg_get_service_time(req->rq_repmsg));
1460 rc = ptlrpc_check_status(req);
1461 imp->imp_connect_error = rc;
1465 * Either we've been evicted, or the server has failed for
1466 * some reason. Try to reconnect, and if that fails, punt to
1469 if (ptlrpc_recoverable_error(rc)) {
1470 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1471 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1474 ptlrpc_request_handle_notconn(req);
1479 * Let's look if server sent slv. Do it only for RPC with
1482 ldlm_cli_update_pool(req);
1486 * Store transno in reqmsg for replay.
1488 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1489 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1490 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1493 if (imp->imp_replayable) {
1494 spin_lock(&imp->imp_lock);
1496 * No point in adding already-committed requests to the replay
1497 * list, we will just remove them immediately. b=9829
1499 if (req->rq_transno != 0 &&
1501 lustre_msg_get_last_committed(req->rq_repmsg) ||
1503 /** version recovery */
1504 ptlrpc_save_versions(req);
1505 ptlrpc_retain_replayable_request(req, imp);
1506 } else if (req->rq_commit_cb != NULL &&
1507 list_empty(&req->rq_replay_list)) {
1508 /* NB: don't call rq_commit_cb if it's already on
1509 * rq_replay_list, ptlrpc_free_committed() will call
1510 * it later, see LU-3618 for details */
1511 spin_unlock(&imp->imp_lock);
1512 req->rq_commit_cb(req);
1513 spin_lock(&imp->imp_lock);
1517 * Replay-enabled imports return commit-status information.
1519 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1520 if (likely(committed > imp->imp_peer_committed_transno))
1521 imp->imp_peer_committed_transno = committed;
1523 ptlrpc_free_committed(imp);
1525 if (!list_empty(&imp->imp_replay_list)) {
1526 struct ptlrpc_request *last;
1528 last = list_entry(imp->imp_replay_list.prev,
1529 struct ptlrpc_request,
1532 * Requests with rq_replay stay on the list even if no
1533 * commit is expected.
1535 if (last->rq_transno > imp->imp_peer_committed_transno)
1536 ptlrpc_pinger_commit_expected(imp);
1539 spin_unlock(&imp->imp_lock);
1546 * Helper function to send request \a req over the network for the first time
1547 * Also adjusts request phase.
1548 * Returns 0 on success or error code.
1550 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1552 struct obd_import *imp = req->rq_import;
1557 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1559 /* do not try to go further if there is not enough memory in enc_pool */
1560 if (req->rq_sent && req->rq_bulk != NULL)
1561 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1562 pool_is_at_full_capacity())
1565 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1566 (!req->rq_generation_set ||
1567 req->rq_import_generation == imp->imp_generation))
1570 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1572 spin_lock(&imp->imp_lock);
1574 LASSERT(req->rq_xid != 0);
1575 LASSERT(!list_empty(&req->rq_unreplied_list));
1577 if (!req->rq_generation_set)
1578 req->rq_import_generation = imp->imp_generation;
1580 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1581 spin_lock(&req->rq_lock);
1582 req->rq_waiting = 1;
1583 spin_unlock(&req->rq_lock);
1585 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1586 ptlrpc_import_state_name(req->rq_send_state),
1587 ptlrpc_import_state_name(imp->imp_state));
1588 LASSERT(list_empty(&req->rq_list));
1589 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1590 atomic_inc(&req->rq_import->imp_inflight);
1591 spin_unlock(&imp->imp_lock);
1596 spin_unlock(&imp->imp_lock);
1597 req->rq_status = rc;
1598 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1602 LASSERT(list_empty(&req->rq_list));
1603 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1604 atomic_inc(&req->rq_import->imp_inflight);
1606 /* find the known replied XID from the unreplied list, CONNECT
1607 * and DISCONNECT requests are skipped to make the sanity check
1608 * on server side happy. see process_req_last_xid().
1610 * For CONNECT: Because replay requests have lower XID, it'll
1611 * break the sanity check if CONNECT bump the exp_last_xid on
1614 * For DISCONNECT: Since client will abort inflight RPC before
1615 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1616 * than the inflight RPC.
1618 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1619 min_xid = ptlrpc_known_replied_xid(imp);
1620 spin_unlock(&imp->imp_lock);
1622 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1624 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1626 rc = sptlrpc_req_refresh_ctx(req, -1);
1629 req->rq_status = rc;
1632 spin_lock(&req->rq_lock);
1633 req->rq_wait_ctx = 1;
1634 spin_unlock(&req->rq_lock);
1639 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1640 " %s:%s:%d:%llu:%s:%d\n", current_comm(),
1641 imp->imp_obd->obd_uuid.uuid,
1642 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1643 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg));
1645 rc = ptl_send_rpc(req, 0);
1646 if (rc == -ENOMEM) {
1647 spin_lock(&imp->imp_lock);
1648 if (!list_empty(&req->rq_list)) {
1649 list_del_init(&req->rq_list);
1650 atomic_dec(&req->rq_import->imp_inflight);
1652 spin_unlock(&imp->imp_lock);
1653 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1657 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1658 spin_lock(&req->rq_lock);
1659 req->rq_net_err = 1;
1660 spin_unlock(&req->rq_lock);
1666 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1671 LASSERT(set->set_producer != NULL);
1673 remaining = atomic_read(&set->set_remaining);
1675 /* populate the ->set_requests list with requests until we
1676 * reach the maximum number of RPCs in flight for this set */
1677 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1678 rc = set->set_producer(set, set->set_producer_arg);
1679 if (rc == -ENOENT) {
1680 /* no more RPC to produce */
1681 set->set_producer = NULL;
1682 set->set_producer_arg = NULL;
1687 RETURN((atomic_read(&set->set_remaining) - remaining));
1691 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1692 * and no more replies are expected.
1693 * (it is possible to get less replies than requests sent e.g. due to timed out
1694 * requests or requests that we had trouble to send out)
1696 * NOTE: This function contains a potential schedule point (cond_resched()).
1698 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1700 struct list_head *tmp, *next;
1701 struct list_head comp_reqs;
1702 int force_timer_recalc = 0;
1705 if (atomic_read(&set->set_remaining) == 0)
1708 INIT_LIST_HEAD(&comp_reqs);
1709 list_for_each_safe(tmp, next, &set->set_requests) {
1710 struct ptlrpc_request *req =
1711 list_entry(tmp, struct ptlrpc_request,
1713 struct obd_import *imp = req->rq_import;
1714 int unregistered = 0;
1718 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1719 list_move_tail(&req->rq_set_chain, &comp_reqs);
1723 /* This schedule point is mainly for the ptlrpcd caller of this
1724 * function. Most ptlrpc sets are not long-lived and unbounded
1725 * in length, but at the least the set used by the ptlrpcd is.
1726 * Since the processing time is unbounded, we need to insert an
1727 * explicit schedule point to make the thread well-behaved.
1731 /* If the caller requires to allow to be interpreted by force
1732 * and it has really been interpreted, then move the request
1733 * to RQ_PHASE_INTERPRET phase in spite of what the current
1735 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1736 req->rq_status = -EINTR;
1737 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1739 /* Since it is interpreted and we have to wait for
1740 * the reply to be unlinked, then use sync mode. */
1743 GOTO(interpret, req->rq_status);
1746 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1747 force_timer_recalc = 1;
1749 /* delayed send - skip */
1750 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1753 /* delayed resend - skip */
1754 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1755 req->rq_sent > ktime_get_real_seconds())
1758 if (!(req->rq_phase == RQ_PHASE_RPC ||
1759 req->rq_phase == RQ_PHASE_BULK ||
1760 req->rq_phase == RQ_PHASE_INTERPRET ||
1761 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1762 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1763 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1767 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1768 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1769 LASSERT(req->rq_next_phase != req->rq_phase);
1770 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1772 if (req->rq_req_deadline &&
1773 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1774 req->rq_req_deadline = 0;
1775 if (req->rq_reply_deadline &&
1776 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1777 req->rq_reply_deadline = 0;
1778 if (req->rq_bulk_deadline &&
1779 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1780 req->rq_bulk_deadline = 0;
1783 * Skip processing until reply is unlinked. We
1784 * can't return to pool before that and we can't
1785 * call interpret before that. We need to make
1786 * sure that all rdma transfers finished and will
1787 * not corrupt any data.
1789 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1790 ptlrpc_client_recv_or_unlink(req))
1792 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1793 ptlrpc_client_bulk_active(req))
1797 * Turn fail_loc off to prevent it from looping
1800 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1801 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1804 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1805 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1810 * Move to next phase if reply was successfully
1813 ptlrpc_rqphase_move(req, req->rq_next_phase);
1816 if (req->rq_phase == RQ_PHASE_INTERPRET)
1817 GOTO(interpret, req->rq_status);
1820 * Note that this also will start async reply unlink.
1822 if (req->rq_net_err && !req->rq_timedout) {
1823 ptlrpc_expire_one_request(req, 1);
1826 * Check if we still need to wait for unlink.
1828 if (ptlrpc_client_recv_or_unlink(req) ||
1829 ptlrpc_client_bulk_active(req))
1831 /* If there is no need to resend, fail it now. */
1832 if (req->rq_no_resend) {
1833 if (req->rq_status == 0)
1834 req->rq_status = -EIO;
1835 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1836 GOTO(interpret, req->rq_status);
1843 spin_lock(&req->rq_lock);
1844 req->rq_replied = 0;
1845 spin_unlock(&req->rq_lock);
1846 if (req->rq_status == 0)
1847 req->rq_status = -EIO;
1848 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1849 GOTO(interpret, req->rq_status);
1852 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1853 * so it sets rq_intr regardless of individual rpc
1854 * timeouts. The synchronous IO waiting path sets
1855 * rq_intr irrespective of whether ptlrpcd
1856 * has seen a timeout. Our policy is to only interpret
1857 * interrupted rpcs after they have timed out, so we
1858 * need to enforce that here.
1861 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1862 req->rq_wait_ctx)) {
1863 req->rq_status = -EINTR;
1864 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1865 GOTO(interpret, req->rq_status);
1868 if (req->rq_phase == RQ_PHASE_RPC) {
1869 if (req->rq_timedout || req->rq_resend ||
1870 req->rq_waiting || req->rq_wait_ctx) {
1873 if (!ptlrpc_unregister_reply(req, 1)) {
1874 ptlrpc_unregister_bulk(req, 1);
1878 spin_lock(&imp->imp_lock);
1879 if (ptlrpc_import_delay_req(imp, req, &status)){
1880 /* put on delay list - only if we wait
1881 * recovery finished - before send */
1882 list_del_init(&req->rq_list);
1883 list_add_tail(&req->rq_list,
1886 spin_unlock(&imp->imp_lock);
1891 req->rq_status = status;
1892 ptlrpc_rqphase_move(req,
1893 RQ_PHASE_INTERPRET);
1894 spin_unlock(&imp->imp_lock);
1895 GOTO(interpret, req->rq_status);
1897 /* ignore on just initiated connections */
1898 if (ptlrpc_no_resend(req) &&
1899 !req->rq_wait_ctx &&
1900 imp->imp_generation !=
1901 imp->imp_initiated_at) {
1902 req->rq_status = -ENOTCONN;
1903 ptlrpc_rqphase_move(req,
1904 RQ_PHASE_INTERPRET);
1905 spin_unlock(&imp->imp_lock);
1906 GOTO(interpret, req->rq_status);
1909 list_del_init(&req->rq_list);
1910 list_add_tail(&req->rq_list,
1911 &imp->imp_sending_list);
1913 spin_unlock(&imp->imp_lock);
1915 spin_lock(&req->rq_lock);
1916 req->rq_waiting = 0;
1917 spin_unlock(&req->rq_lock);
1919 if (req->rq_timedout || req->rq_resend) {
1920 /* This is re-sending anyways,
1921 * let's mark req as resend. */
1922 spin_lock(&req->rq_lock);
1924 spin_unlock(&req->rq_lock);
1927 * rq_wait_ctx is only touched by ptlrpcd,
1928 * so no lock is needed here.
1930 status = sptlrpc_req_refresh_ctx(req, -1);
1933 req->rq_status = status;
1934 spin_lock(&req->rq_lock);
1935 req->rq_wait_ctx = 0;
1936 spin_unlock(&req->rq_lock);
1937 force_timer_recalc = 1;
1939 spin_lock(&req->rq_lock);
1940 req->rq_wait_ctx = 1;
1941 spin_unlock(&req->rq_lock);
1946 spin_lock(&req->rq_lock);
1947 req->rq_wait_ctx = 0;
1948 spin_unlock(&req->rq_lock);
1951 /* In any case, the previous bulk should be
1952 * cleaned up to prepare for the new sending */
1953 if (req->rq_bulk != NULL &&
1954 !ptlrpc_unregister_bulk(req, 1))
1957 rc = ptl_send_rpc(req, 0);
1958 if (rc == -ENOMEM) {
1959 spin_lock(&imp->imp_lock);
1960 if (!list_empty(&req->rq_list))
1961 list_del_init(&req->rq_list);
1962 spin_unlock(&imp->imp_lock);
1963 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1967 DEBUG_REQ(D_HA, req,
1968 "send failed: rc = %d", rc);
1969 force_timer_recalc = 1;
1970 spin_lock(&req->rq_lock);
1971 req->rq_net_err = 1;
1972 spin_unlock(&req->rq_lock);
1975 /* need to reset the timeout */
1976 force_timer_recalc = 1;
1979 spin_lock(&req->rq_lock);
1981 if (ptlrpc_client_early(req)) {
1982 ptlrpc_at_recv_early_reply(req);
1983 spin_unlock(&req->rq_lock);
1987 /* Still waiting for a reply? */
1988 if (ptlrpc_client_recv(req)) {
1989 spin_unlock(&req->rq_lock);
1993 /* Did we actually receive a reply? */
1994 if (!ptlrpc_client_replied(req)) {
1995 spin_unlock(&req->rq_lock);
1999 spin_unlock(&req->rq_lock);
2001 /* unlink from net because we are going to
2002 * swab in-place of reply buffer */
2003 unregistered = ptlrpc_unregister_reply(req, 1);
2007 req->rq_status = after_reply(req);
2011 /* If there is no bulk associated with this request,
2012 * then we're done and should let the interpreter
2013 * process the reply. Similarly if the RPC returned
2014 * an error, and therefore the bulk will never arrive.
2016 if (req->rq_bulk == NULL || req->rq_status < 0) {
2017 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2018 GOTO(interpret, req->rq_status);
2021 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2024 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2025 if (ptlrpc_client_bulk_active(req))
2028 if (req->rq_bulk->bd_failure) {
2029 /* The RPC reply arrived OK, but the bulk screwed
2030 * up! Dead weird since the server told us the RPC
2031 * was good after getting the REPLY for her GET or
2032 * the ACK for her PUT. */
2033 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
2034 req->rq_status = -EIO;
2037 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2040 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2042 /* This moves to "unregistering" phase we need to wait for
2044 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2045 /* start async bulk unlink too */
2046 ptlrpc_unregister_bulk(req, 1);
2050 if (!ptlrpc_unregister_bulk(req, async))
2053 /* When calling interpret receiving already should be
2055 LASSERT(!req->rq_receiving_reply);
2057 ptlrpc_req_interpret(env, req, req->rq_status);
2059 if (ptlrpcd_check_work(req)) {
2060 atomic_dec(&set->set_remaining);
2063 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2065 if (req->rq_reqmsg != NULL)
2067 "Completed RPC pname:cluuid:pid:xid:nid:"
2068 "opc %s:%s:%d:%llu:%s:%d\n", current_comm(),
2069 imp->imp_obd->obd_uuid.uuid,
2070 lustre_msg_get_status(req->rq_reqmsg),
2072 obd_import_nid2str(imp),
2073 lustre_msg_get_opc(req->rq_reqmsg));
2075 spin_lock(&imp->imp_lock);
2076 /* Request already may be not on sending or delaying list. This
2077 * may happen in the case of marking it erroneous for the case
2078 * ptlrpc_import_delay_req(req, status) find it impossible to
2079 * allow sending this rpc and returns *status != 0. */
2080 if (!list_empty(&req->rq_list)) {
2081 list_del_init(&req->rq_list);
2082 atomic_dec(&imp->imp_inflight);
2084 list_del_init(&req->rq_unreplied_list);
2085 spin_unlock(&imp->imp_lock);
2087 atomic_dec(&set->set_remaining);
2088 wake_up_all(&imp->imp_recovery_waitq);
2090 if (set->set_producer) {
2091 /* produce a new request if possible */
2092 if (ptlrpc_set_producer(set) > 0)
2093 force_timer_recalc = 1;
2095 /* free the request that has just been completed
2096 * in order not to pollute set->set_requests */
2097 list_del_init(&req->rq_set_chain);
2098 spin_lock(&req->rq_lock);
2100 req->rq_invalid_rqset = 0;
2101 spin_unlock(&req->rq_lock);
2103 /* record rq_status to compute the final status later */
2104 if (req->rq_status != 0)
2105 set->set_rc = req->rq_status;
2106 ptlrpc_req_finished(req);
2108 list_move_tail(&req->rq_set_chain, &comp_reqs);
2112 /* move completed request at the head of list so it's easier for
2113 * caller to find them */
2114 list_splice(&comp_reqs, &set->set_requests);
2116 /* If we hit an error, we want to recover promptly. */
2117 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2119 EXPORT_SYMBOL(ptlrpc_check_set);
2122 * Time out request \a req. is \a async_unlink is set, that means do not wait
2123 * until LNet actually confirms network buffer unlinking.
2124 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2126 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2128 struct obd_import *imp = req->rq_import;
2129 unsigned int debug_mask = D_RPCTRACE;
2133 spin_lock(&req->rq_lock);
2134 req->rq_timedout = 1;
2135 spin_unlock(&req->rq_lock);
2137 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2138 lustre_msg_get_status(req->rq_reqmsg)))
2139 debug_mask = D_WARNING;
2140 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2141 req->rq_net_err ? "failed due to network error" :
2142 ((req->rq_real_sent == 0 ||
2143 req->rq_real_sent < req->rq_sent ||
2144 req->rq_real_sent >= req->rq_deadline) ?
2145 "timed out for sent delay" : "timed out for slow reply"),
2146 (s64)req->rq_sent, (s64)req->rq_real_sent);
2148 if (imp != NULL && obd_debug_peer_on_timeout)
2149 LNetDebugPeer(imp->imp_connection->c_peer);
2151 ptlrpc_unregister_reply(req, async_unlink);
2152 ptlrpc_unregister_bulk(req, async_unlink);
2154 if (obd_dump_on_timeout)
2155 libcfs_debug_dumplog();
2158 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2162 atomic_inc(&imp->imp_timeouts);
2164 /* The DLM server doesn't want recovery run on its imports. */
2165 if (imp->imp_dlm_fake)
2168 /* If this request is for recovery or other primordial tasks,
2169 * then error it out here. */
2170 if (req->rq_ctx_init || req->rq_ctx_fini ||
2171 req->rq_send_state != LUSTRE_IMP_FULL ||
2172 imp->imp_obd->obd_no_recov) {
2173 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2174 ptlrpc_import_state_name(req->rq_send_state),
2175 ptlrpc_import_state_name(imp->imp_state));
2176 spin_lock(&req->rq_lock);
2177 req->rq_status = -ETIMEDOUT;
2179 spin_unlock(&req->rq_lock);
2183 /* if a request can't be resent we can't wait for an answer after
2185 if (ptlrpc_no_resend(req)) {
2186 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2190 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2196 * Time out all uncompleted requests in request set pointed by \a data
2197 * Callback used when waiting on sets with l_wait_event.
2200 int ptlrpc_expired_set(void *data)
2202 struct ptlrpc_request_set *set = data;
2203 struct list_head *tmp;
2204 time64_t now = ktime_get_real_seconds();
2207 LASSERT(set != NULL);
2210 * A timeout expired. See which reqs it applies to...
2212 list_for_each(tmp, &set->set_requests) {
2213 struct ptlrpc_request *req =
2214 list_entry(tmp, struct ptlrpc_request,
2217 /* don't expire request waiting for context */
2218 if (req->rq_wait_ctx)
2221 /* Request in-flight? */
2222 if (!((req->rq_phase == RQ_PHASE_RPC &&
2223 !req->rq_waiting && !req->rq_resend) ||
2224 (req->rq_phase == RQ_PHASE_BULK)))
2227 if (req->rq_timedout || /* already dealt with */
2228 req->rq_deadline > now) /* not expired */
2231 /* Deal with this guy. Do it asynchronously to not block
2232 * ptlrpcd thread. */
2233 ptlrpc_expire_one_request(req, 1);
2237 * When waiting for a whole set, we always break out of the
2238 * sleep so we can recalculate the timeout, or enable interrupts
2239 * if everyone's timed out.
2245 * Sets rq_intr flag in \a req under spinlock.
2247 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2249 spin_lock(&req->rq_lock);
2251 spin_unlock(&req->rq_lock);
2253 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2256 * Interrupts (sets interrupted flag) all uncompleted requests in
2257 * a set \a data. Callback for l_wait_event for interruptible waits.
2259 static void ptlrpc_interrupted_set(void *data)
2261 struct ptlrpc_request_set *set = data;
2262 struct list_head *tmp;
2264 LASSERT(set != NULL);
2265 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2267 list_for_each(tmp, &set->set_requests) {
2268 struct ptlrpc_request *req =
2269 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2274 if (req->rq_phase != RQ_PHASE_RPC &&
2275 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2276 !req->rq_allow_intr)
2279 ptlrpc_mark_interrupted(req);
2284 * Get the smallest timeout in the set; this does NOT set a timeout.
2286 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2288 struct list_head *tmp;
2289 time64_t now = ktime_get_real_seconds();
2291 struct ptlrpc_request *req;
2295 list_for_each(tmp, &set->set_requests) {
2296 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2299 * Request in-flight?
2301 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2302 (req->rq_phase == RQ_PHASE_BULK) ||
2303 (req->rq_phase == RQ_PHASE_NEW)))
2307 * Already timed out.
2309 if (req->rq_timedout)
2315 if (req->rq_wait_ctx)
2318 if (req->rq_phase == RQ_PHASE_NEW)
2319 deadline = req->rq_sent;
2320 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2321 deadline = req->rq_sent;
2323 deadline = req->rq_sent + req->rq_timeout;
2325 if (deadline <= now) /* actually expired already */
2326 timeout = 1; /* ASAP */
2327 else if (timeout == 0 || timeout > deadline - now)
2328 timeout = deadline - now;
2334 * Send all unset request from the set and then wait untill all
2335 * requests in the set complete (either get a reply, timeout, get an
2336 * error or otherwise be interrupted).
2337 * Returns 0 on success or error code otherwise.
2339 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2341 struct list_head *tmp;
2342 struct ptlrpc_request *req;
2343 struct l_wait_info lwi;
2348 if (set->set_producer)
2349 (void)ptlrpc_set_producer(set);
2351 list_for_each(tmp, &set->set_requests) {
2352 req = list_entry(tmp, struct ptlrpc_request,
2354 if (req->rq_phase == RQ_PHASE_NEW)
2355 (void)ptlrpc_send_new_req(req);
2358 if (list_empty(&set->set_requests))
2362 timeout = ptlrpc_set_next_timeout(set);
2364 /* wait until all complete, interrupted, or an in-flight
2366 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2369 if ((timeout == 0 && !signal_pending(current)) ||
2370 set->set_allow_intr)
2371 /* No requests are in-flight (ether timed out
2372 * or delayed), so we can allow interrupts.
2373 * We still want to block for a limited time,
2374 * so we allow interrupts during the timeout. */
2375 lwi = LWI_TIMEOUT_INTR_ALL(
2376 cfs_time_seconds(timeout ? timeout : 1),
2378 ptlrpc_interrupted_set, set);
2381 * At least one request is in flight, so no
2382 * interrupts are allowed. Wait until all
2383 * complete, or an in-flight req times out.
2385 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2386 ptlrpc_expired_set, set);
2388 rc = l_wait_event(set->set_waitq,
2389 ptlrpc_check_set(NULL, set), &lwi);
2391 /* LU-769 - if we ignored the signal because it was already
2392 * pending when we started, we need to handle it now or we risk
2393 * it being ignored forever */
2394 if (rc == -ETIMEDOUT &&
2395 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2396 signal_pending(current)) {
2397 sigset_t blocked_sigs =
2398 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2400 /* In fact we only interrupt for the "fatal" signals
2401 * like SIGINT or SIGKILL. We still ignore less
2402 * important signals since ptlrpc set is not easily
2403 * reentrant from userspace again */
2404 if (signal_pending(current))
2405 ptlrpc_interrupted_set(set);
2406 cfs_restore_sigs(blocked_sigs);
2409 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2411 /* -EINTR => all requests have been flagged rq_intr so next
2413 * -ETIMEDOUT => someone timed out. When all reqs have
2414 * timed out, signals are enabled allowing completion with
2416 * I don't really care if we go once more round the loop in
2417 * the error cases -eeb. */
2418 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2419 list_for_each(tmp, &set->set_requests) {
2420 req = list_entry(tmp, struct ptlrpc_request,
2422 spin_lock(&req->rq_lock);
2423 req->rq_invalid_rqset = 1;
2424 spin_unlock(&req->rq_lock);
2427 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2429 LASSERT(atomic_read(&set->set_remaining) == 0);
2431 rc = set->set_rc; /* rq_status of already freed requests if any */
2432 list_for_each(tmp, &set->set_requests) {
2433 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2435 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2436 if (req->rq_status != 0)
2437 rc = req->rq_status;
2442 EXPORT_SYMBOL(ptlrpc_set_wait);
2445 * Helper fuction for request freeing.
2446 * Called when request count reached zero and request needs to be freed.
2447 * Removes request from all sorts of sending/replay lists it might be on,
2448 * frees network buffers if any are present.
2449 * If \a locked is set, that means caller is already holding import imp_lock
2450 * and so we no longer need to reobtain it (for certain lists manipulations)
2452 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2456 if (request == NULL)
2459 LASSERT(!request->rq_srv_req);
2460 LASSERT(request->rq_export == NULL);
2461 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2462 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2463 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2464 LASSERTF(!request->rq_replay, "req %p\n", request);
2466 req_capsule_fini(&request->rq_pill);
2468 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2469 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2470 if (request->rq_import != NULL) {
2472 spin_lock(&request->rq_import->imp_lock);
2473 list_del_init(&request->rq_replay_list);
2474 list_del_init(&request->rq_unreplied_list);
2476 spin_unlock(&request->rq_import->imp_lock);
2478 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2480 if (atomic_read(&request->rq_refcount) != 0) {
2481 DEBUG_REQ(D_ERROR, request,
2482 "freeing request with nonzero refcount");
2486 if (request->rq_repbuf != NULL)
2487 sptlrpc_cli_free_repbuf(request);
2489 if (request->rq_import != NULL) {
2490 if (!ptlrpcd_check_work(request)) {
2491 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2492 atomic_dec(&request->rq_import->imp_reqs);
2494 class_import_put(request->rq_import);
2495 request->rq_import = NULL;
2497 if (request->rq_bulk != NULL)
2498 ptlrpc_free_bulk(request->rq_bulk);
2500 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2501 sptlrpc_cli_free_reqbuf(request);
2503 if (request->rq_cli_ctx)
2504 sptlrpc_req_put_ctx(request, !locked);
2506 if (request->rq_pool)
2507 __ptlrpc_free_req_to_pool(request);
2509 ptlrpc_request_cache_free(request);
2513 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2515 * Drop one request reference. Must be called with import imp_lock held.
2516 * When reference count drops to zero, request is freed.
2518 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2520 assert_spin_locked(&request->rq_import->imp_lock);
2521 (void)__ptlrpc_req_finished(request, 1);
2526 * Drops one reference count for request \a request.
2527 * \a locked set indicates that caller holds import imp_lock.
2528 * Frees the request whe reference count reaches zero.
2530 * \retval 1 the request is freed
2531 * \retval 0 some others still hold references on the request
2533 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2541 LASSERT(request != LP_POISON);
2542 LASSERT(request->rq_reqmsg != LP_POISON);
2544 DEBUG_REQ(D_INFO, request, "refcount now %u",
2545 atomic_read(&request->rq_refcount) - 1);
2547 spin_lock(&request->rq_lock);
2548 count = atomic_dec_return(&request->rq_refcount);
2549 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2551 /* For open RPC, the client does not know the EA size (LOV, ACL, and
2552 * so on) before replied, then the client has to reserve very large
2553 * reply buffer. Such buffer will not be released until the RPC freed.
2554 * Since The open RPC is replayable, we need to keep it in the replay
2555 * list until close. If there are a lot of files opened concurrently,
2556 * then the client may be OOM.
2558 * If fact, it is unnecessary to keep reply buffer for open replay,
2559 * related EAs have already been saved via mdc_save_lovea() before
2560 * coming here. So it is safe to free the reply buffer some earlier
2561 * before releasing the RPC to avoid client OOM. LU-9514 */
2562 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2563 spin_lock(&request->rq_early_free_lock);
2564 sptlrpc_cli_free_repbuf(request);
2565 request->rq_repbuf = NULL;
2566 request->rq_repbuf_len = 0;
2567 request->rq_repdata = NULL;
2568 request->rq_reqdata_len = 0;
2569 spin_unlock(&request->rq_early_free_lock);
2571 spin_unlock(&request->rq_lock);
2574 __ptlrpc_free_req(request, locked);
2580 * Drops one reference count for a request.
2582 void ptlrpc_req_finished(struct ptlrpc_request *request)
2584 __ptlrpc_req_finished(request, 0);
2586 EXPORT_SYMBOL(ptlrpc_req_finished);
2589 * Returns xid of a \a request
2591 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2593 return request->rq_xid;
2595 EXPORT_SYMBOL(ptlrpc_req_xid);
2598 * Disengage the client's reply buffer from the network
2599 * NB does _NOT_ unregister any client-side bulk.
2600 * IDEMPOTENT, but _not_ safe against concurrent callers.
2601 * The request owner (i.e. the thread doing the I/O) must call...
2602 * Returns 0 on success or 1 if unregistering cannot be made.
2604 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2607 struct l_wait_info lwi;
2612 LASSERT(!in_interrupt());
2614 /* Let's setup deadline for reply unlink. */
2615 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2616 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2617 request->rq_reply_deadline = ktime_get_real_seconds() +
2621 * Nothing left to do.
2623 if (!ptlrpc_client_recv_or_unlink(request))
2626 LNetMDUnlink(request->rq_reply_md_h);
2629 * Let's check it once again.
2631 if (!ptlrpc_client_recv_or_unlink(request))
2634 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2635 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2638 * Do not wait for unlink to finish.
2644 * We have to l_wait_event() whatever the result, to give liblustre
2645 * a chance to run reply_in_callback(), and to make sure we've
2646 * unlinked before returning a req to the pool.
2649 /* The wq argument is ignored by user-space wait_event macros */
2650 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2651 &request->rq_set->set_waitq :
2652 &request->rq_reply_waitq;
2653 /* Network access will complete in finite time but the HUGE
2654 * timeout lets us CWARN for visibility of sluggish NALs */
2655 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2656 cfs_time_seconds(1), NULL, NULL);
2657 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2660 ptlrpc_rqphase_move(request, request->rq_next_phase);
2664 LASSERT(rc == -ETIMEDOUT);
2665 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2666 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2667 request->rq_receiving_reply,
2668 request->rq_req_unlinked,
2669 request->rq_reply_unlinked);
2674 static void ptlrpc_free_request(struct ptlrpc_request *req)
2676 spin_lock(&req->rq_lock);
2678 spin_unlock(&req->rq_lock);
2680 if (req->rq_commit_cb != NULL)
2681 req->rq_commit_cb(req);
2682 list_del_init(&req->rq_replay_list);
2684 __ptlrpc_req_finished(req, 1);
2688 * the request is committed and dropped from the replay list of its import
2690 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2692 struct obd_import *imp = req->rq_import;
2694 spin_lock(&imp->imp_lock);
2695 if (list_empty(&req->rq_replay_list)) {
2696 spin_unlock(&imp->imp_lock);
2700 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2701 if (imp->imp_replay_cursor == &req->rq_replay_list)
2702 imp->imp_replay_cursor = req->rq_replay_list.next;
2703 ptlrpc_free_request(req);
2706 spin_unlock(&imp->imp_lock);
2708 EXPORT_SYMBOL(ptlrpc_request_committed);
2711 * Iterates through replay_list on import and prunes
2712 * all requests have transno smaller than last_committed for the
2713 * import and don't have rq_replay set.
2714 * Since requests are sorted in transno order, stops when meetign first
2715 * transno bigger than last_committed.
2716 * caller must hold imp->imp_lock
2718 void ptlrpc_free_committed(struct obd_import *imp)
2720 struct ptlrpc_request *req, *saved;
2721 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2722 bool skip_committed_list = true;
2725 LASSERT(imp != NULL);
2726 assert_spin_locked(&imp->imp_lock);
2728 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2729 imp->imp_generation == imp->imp_last_generation_checked) {
2730 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2731 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2734 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2735 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2736 imp->imp_generation);
2738 if (imp->imp_generation != imp->imp_last_generation_checked ||
2739 imp->imp_last_transno_checked == 0)
2740 skip_committed_list = false;
2742 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2743 imp->imp_last_generation_checked = imp->imp_generation;
2745 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2747 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2748 LASSERT(req != last_req);
2751 if (req->rq_transno == 0) {
2752 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2755 if (req->rq_import_generation < imp->imp_generation) {
2756 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2760 /* not yet committed */
2761 if (req->rq_transno > imp->imp_peer_committed_transno) {
2762 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2766 if (req->rq_replay) {
2767 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2768 list_move_tail(&req->rq_replay_list,
2769 &imp->imp_committed_list);
2773 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2774 imp->imp_peer_committed_transno);
2776 ptlrpc_free_request(req);
2779 if (skip_committed_list)
2782 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2784 LASSERT(req->rq_transno != 0);
2785 if (req->rq_import_generation < imp->imp_generation ||
2787 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2788 req->rq_import_generation <
2789 imp->imp_generation ? "stale" : "closed");
2791 if (imp->imp_replay_cursor == &req->rq_replay_list)
2792 imp->imp_replay_cursor =
2793 req->rq_replay_list.next;
2795 ptlrpc_free_request(req);
2802 void ptlrpc_cleanup_client(struct obd_import *imp)
2809 * Schedule previously sent request for resend.
2810 * For bulk requests we assign new xid (to avoid problems with
2811 * lost replies and therefore several transfers landing into same buffer
2812 * from different sending attempts).
2814 void ptlrpc_resend_req(struct ptlrpc_request *req)
2816 DEBUG_REQ(D_HA, req, "going to resend");
2817 spin_lock(&req->rq_lock);
2819 /* Request got reply but linked to the import list still.
2820 Let ptlrpc_check_set() to process it. */
2821 if (ptlrpc_client_replied(req)) {
2822 spin_unlock(&req->rq_lock);
2823 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2827 req->rq_status = -EAGAIN;
2830 req->rq_net_err = 0;
2831 req->rq_timedout = 0;
2833 ptlrpc_client_wake_req(req);
2834 spin_unlock(&req->rq_lock);
2837 /* XXX: this function and rq_status are currently unused */
2838 void ptlrpc_restart_req(struct ptlrpc_request *req)
2840 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2841 req->rq_status = -ERESTARTSYS;
2843 spin_lock(&req->rq_lock);
2844 req->rq_restart = 1;
2845 req->rq_timedout = 0;
2846 ptlrpc_client_wake_req(req);
2847 spin_unlock(&req->rq_lock);
2851 * Grab additional reference on a request \a req
2853 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2856 atomic_inc(&req->rq_refcount);
2859 EXPORT_SYMBOL(ptlrpc_request_addref);
2862 * Add a request to import replay_list.
2863 * Must be called under imp_lock
2865 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2866 struct obd_import *imp)
2868 struct list_head *tmp;
2870 assert_spin_locked(&imp->imp_lock);
2872 if (req->rq_transno == 0) {
2873 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2877 /* clear this for new requests that were resent as well
2878 as resent replayed requests. */
2879 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2881 /* don't re-add requests that have been replayed */
2882 if (!list_empty(&req->rq_replay_list))
2885 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2887 spin_lock(&req->rq_lock);
2889 spin_unlock(&req->rq_lock);
2891 LASSERT(imp->imp_replayable);
2892 /* Balanced in ptlrpc_free_committed, usually. */
2893 ptlrpc_request_addref(req);
2894 list_for_each_prev(tmp, &imp->imp_replay_list) {
2895 struct ptlrpc_request *iter = list_entry(tmp,
2896 struct ptlrpc_request,
2899 /* We may have duplicate transnos if we create and then
2900 * open a file, or for closes retained if to match creating
2901 * opens, so use req->rq_xid as a secondary key.
2902 * (See bugs 684, 685, and 428.)
2903 * XXX no longer needed, but all opens need transnos!
2905 if (iter->rq_transno > req->rq_transno)
2908 if (iter->rq_transno == req->rq_transno) {
2909 LASSERT(iter->rq_xid != req->rq_xid);
2910 if (iter->rq_xid > req->rq_xid)
2914 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2918 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2922 * Send request and wait until it completes.
2923 * Returns request processing status.
2925 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2927 struct ptlrpc_request_set *set;
2931 LASSERT(req->rq_set == NULL);
2932 LASSERT(!req->rq_receiving_reply);
2934 set = ptlrpc_prep_set();
2936 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2940 /* for distributed debugging */
2941 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2943 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2944 ptlrpc_request_addref(req);
2945 ptlrpc_set_add_req(set, req);
2946 rc = ptlrpc_set_wait(NULL, set);
2947 ptlrpc_set_destroy(set);
2951 EXPORT_SYMBOL(ptlrpc_queue_wait);
2954 * Callback used for replayed requests reply processing.
2955 * In case of successful reply calls registered request replay callback.
2956 * In case of error restart replay process.
2958 static int ptlrpc_replay_interpret(const struct lu_env *env,
2959 struct ptlrpc_request *req,
2960 void * data, int rc)
2962 struct ptlrpc_replay_async_args *aa = data;
2963 struct obd_import *imp = req->rq_import;
2966 atomic_dec(&imp->imp_replay_inflight);
2968 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2969 * server got the request, but bulk transfer timeout, let's
2970 * replay the bulk req again */
2971 if (!ptlrpc_client_replied(req) ||
2972 (req->rq_bulk != NULL &&
2973 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2974 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2975 GOTO(out, rc = -ETIMEDOUT);
2978 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2979 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2980 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2981 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2983 /** VBR: check version failure */
2984 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2985 /** replay was failed due to version mismatch */
2986 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2987 spin_lock(&imp->imp_lock);
2988 imp->imp_vbr_failed = 1;
2989 spin_unlock(&imp->imp_lock);
2990 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2992 /** The transno had better not change over replay. */
2993 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2994 lustre_msg_get_transno(req->rq_repmsg) ||
2995 lustre_msg_get_transno(req->rq_repmsg) == 0,
2997 lustre_msg_get_transno(req->rq_reqmsg),
2998 lustre_msg_get_transno(req->rq_repmsg));
3001 spin_lock(&imp->imp_lock);
3002 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3003 spin_unlock(&imp->imp_lock);
3004 LASSERT(imp->imp_last_replay_transno);
3006 /* transaction number shouldn't be bigger than the latest replayed */
3007 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3008 DEBUG_REQ(D_ERROR, req,
3009 "Reported transno %llu is bigger than the "
3010 "replayed one: %llu", req->rq_transno,
3011 lustre_msg_get_transno(req->rq_reqmsg));
3012 GOTO(out, rc = -EINVAL);
3015 DEBUG_REQ(D_HA, req, "got rep");
3017 /* let the callback do fixups, possibly including in the request */
3018 if (req->rq_replay_cb)
3019 req->rq_replay_cb(req);
3021 if (ptlrpc_client_replied(req) &&
3022 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3023 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3024 lustre_msg_get_status(req->rq_repmsg),
3025 aa->praa_old_status);
3027 /* Note: If the replay fails for MDT-MDT recovery, let's
3028 * abort all of the following requests in the replay
3029 * and sending list, because MDT-MDT update requests
3030 * are dependent on each other, see LU-7039 */
3031 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3032 struct ptlrpc_request *free_req;
3033 struct ptlrpc_request *tmp;
3035 spin_lock(&imp->imp_lock);
3036 list_for_each_entry_safe(free_req, tmp,
3037 &imp->imp_replay_list,
3039 ptlrpc_free_request(free_req);
3042 list_for_each_entry_safe(free_req, tmp,
3043 &imp->imp_committed_list,
3045 ptlrpc_free_request(free_req);
3048 list_for_each_entry_safe(free_req, tmp,
3049 &imp->imp_delayed_list,
3051 spin_lock(&free_req->rq_lock);
3052 free_req->rq_err = 1;
3053 free_req->rq_status = -EIO;
3054 ptlrpc_client_wake_req(free_req);
3055 spin_unlock(&free_req->rq_lock);
3058 list_for_each_entry_safe(free_req, tmp,
3059 &imp->imp_sending_list,
3061 spin_lock(&free_req->rq_lock);
3062 free_req->rq_err = 1;
3063 free_req->rq_status = -EIO;
3064 ptlrpc_client_wake_req(free_req);
3065 spin_unlock(&free_req->rq_lock);
3067 spin_unlock(&imp->imp_lock);
3070 /* Put it back for re-replay. */
3071 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3075 * Errors while replay can set transno to 0, but
3076 * imp_last_replay_transno shouldn't be set to 0 anyway
3078 if (req->rq_transno == 0)
3079 CERROR("Transno is 0 during replay!\n");
3081 /* continue with recovery */
3082 rc = ptlrpc_import_recovery_state_machine(imp);
3084 req->rq_send_state = aa->praa_old_state;
3087 /* this replay failed, so restart recovery */
3088 ptlrpc_connect_import(imp);
3094 * Prepares and queues request for replay.
3095 * Adds it to ptlrpcd queue for actual sending.
3096 * Returns 0 on success.
3098 int ptlrpc_replay_req(struct ptlrpc_request *req)
3100 struct ptlrpc_replay_async_args *aa;
3104 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3106 CLASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
3107 aa = ptlrpc_req_async_args(req);
3108 memset(aa, 0, sizeof(*aa));
3110 /* Prepare request to be resent with ptlrpcd */
3111 aa->praa_old_state = req->rq_send_state;
3112 req->rq_send_state = LUSTRE_IMP_REPLAY;
3113 req->rq_phase = RQ_PHASE_NEW;
3114 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3116 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3118 req->rq_interpret_reply = ptlrpc_replay_interpret;
3119 /* Readjust the timeout for current conditions */
3120 ptlrpc_at_set_req_timeout(req);
3122 /* Tell server the net_latency, so the server can calculate how long
3123 * it should wait for next replay */
3124 lustre_msg_set_service_time(req->rq_reqmsg,
3125 ptlrpc_at_get_net_latency(req));
3126 DEBUG_REQ(D_HA, req, "REPLAY");
3128 atomic_inc(&req->rq_import->imp_replay_inflight);
3129 spin_lock(&req->rq_lock);
3130 req->rq_early_free_repbuf = 0;
3131 spin_unlock(&req->rq_lock);
3132 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3134 ptlrpcd_add_req(req);
3139 * Aborts all in-flight request on import \a imp sending and delayed lists
3141 void ptlrpc_abort_inflight(struct obd_import *imp)
3143 struct list_head *tmp, *n;
3147 * Make sure that no new requests get processed for this import.
3148 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3149 * this flag and then putting requests on sending_list or delayed_list.
3151 assert_spin_locked(&imp->imp_lock);
3153 /* XXX locking? Maybe we should remove each request with the list
3154 * locked? Also, how do we know if the requests on the list are
3155 * being freed at this time?
3157 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3158 struct ptlrpc_request *req = list_entry(tmp,
3159 struct ptlrpc_request,
3162 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3164 spin_lock(&req->rq_lock);
3165 if (req->rq_import_generation < imp->imp_generation) {
3167 req->rq_status = -EIO;
3168 ptlrpc_client_wake_req(req);
3170 spin_unlock(&req->rq_lock);
3173 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3174 struct ptlrpc_request *req =
3175 list_entry(tmp, struct ptlrpc_request, rq_list);
3177 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3179 spin_lock(&req->rq_lock);
3180 if (req->rq_import_generation < imp->imp_generation) {
3182 req->rq_status = -EIO;
3183 ptlrpc_client_wake_req(req);
3185 spin_unlock(&req->rq_lock);
3188 /* Last chance to free reqs left on the replay list, but we
3189 * will still leak reqs that haven't committed. */
3190 if (imp->imp_replayable)
3191 ptlrpc_free_committed(imp);
3197 * Abort all uncompleted requests in request set \a set
3199 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3201 struct list_head *tmp, *pos;
3203 LASSERT(set != NULL);
3205 list_for_each_safe(pos, tmp, &set->set_requests) {
3206 struct ptlrpc_request *req =
3207 list_entry(pos, struct ptlrpc_request,
3210 spin_lock(&req->rq_lock);
3211 if (req->rq_phase != RQ_PHASE_RPC) {
3212 spin_unlock(&req->rq_lock);
3217 req->rq_status = -EINTR;
3218 ptlrpc_client_wake_req(req);
3219 spin_unlock(&req->rq_lock);
3223 static __u64 ptlrpc_last_xid;
3224 static spinlock_t ptlrpc_last_xid_lock;
3227 * Initialize the XID for the node. This is common among all requests on
3228 * this node, and only requires the property that it is monotonically
3229 * increasing. It does not need to be sequential. Since this is also used
3230 * as the RDMA match bits, it is important that a single client NOT have
3231 * the same match bits for two different in-flight requests, hence we do
3232 * NOT want to have an XID per target or similar.
3234 * To avoid an unlikely collision between match bits after a client reboot
3235 * (which would deliver old data into the wrong RDMA buffer) initialize
3236 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3237 * If the time is clearly incorrect, we instead use a 62-bit random number.
3238 * In the worst case the random number will overflow 1M RPCs per second in
3239 * 9133 years, or permutations thereof.
3241 #define YEAR_2004 (1ULL << 30)
3242 void ptlrpc_init_xid(void)
3244 time64_t now = ktime_get_real_seconds();
3246 spin_lock_init(&ptlrpc_last_xid_lock);
3247 if (now < YEAR_2004) {
3248 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3249 ptlrpc_last_xid >>= 2;
3250 ptlrpc_last_xid |= (1ULL << 61);
3252 ptlrpc_last_xid = (__u64)now << 20;
3255 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3256 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3257 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3261 * Increase xid and returns resulting new value to the caller.
3263 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3264 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3265 * itself uses the last bulk xid needed, so the server can determine the
3266 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3267 * xid must align to a power-of-two value.
3269 * This is assumed to be true due to the initial ptlrpc_last_xid
3270 * value also being initialized to a power-of-two value. LU-1431
3272 __u64 ptlrpc_next_xid(void)
3276 spin_lock(&ptlrpc_last_xid_lock);
3277 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3278 ptlrpc_last_xid = next;
3279 spin_unlock(&ptlrpc_last_xid_lock);
3285 * If request has a new allocated XID (new request or EINPROGRESS resend),
3286 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3287 * request to ensure previous bulk fails and avoid problems with lost replies
3288 * and therefore several transfers landing into the same buffer from different
3291 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3293 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3295 LASSERT(bd != NULL);
3297 /* Generate new matchbits for all resend requests, including
3299 if (req->rq_resend) {
3300 __u64 old_mbits = req->rq_mbits;
3302 /* First time resend on -EINPROGRESS will generate new xid,
3303 * so we can actually use the rq_xid as rq_mbits in such case,
3304 * however, it's bit hard to distinguish such resend with a
3305 * 'resend for the -EINPROGRESS resend'. To make it simple,
3306 * we opt to generate mbits for all resend cases. */
3307 if (OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS)){
3308 req->rq_mbits = ptlrpc_next_xid();
3310 /* Old version transfers rq_xid to peer as
3312 spin_lock(&req->rq_import->imp_lock);
3313 list_del_init(&req->rq_unreplied_list);
3314 ptlrpc_assign_next_xid_nolock(req);
3315 spin_unlock(&req->rq_import->imp_lock);
3316 req->rq_mbits = req->rq_xid;
3318 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3319 old_mbits, req->rq_mbits);
3320 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3321 /* Request being sent first time, use xid as matchbits. */
3322 req->rq_mbits = req->rq_xid;
3324 /* Replay request, xid and matchbits have already been
3325 * correctly assigned. */
3329 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3330 * that server can infer the number of bulks that were prepared,
3332 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3335 /* Set rq_xid as rq_mbits to indicate the final bulk for the old
3336 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3338 * It's ok to directly set the rq_xid here, since this xid bump
3339 * won't affect the request position in unreplied list. */
3340 if (!OCD_HAS_FLAG(&bd->bd_import->imp_connect_data, BULK_MBITS))
3341 req->rq_xid = req->rq_mbits;
3345 * Get a glimpse at what next xid value might have been.
3346 * Returns possible next xid.
3348 __u64 ptlrpc_sample_next_xid(void)
3350 #if BITS_PER_LONG == 32
3351 /* need to avoid possible word tearing on 32-bit systems */
3354 spin_lock(&ptlrpc_last_xid_lock);
3355 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3356 spin_unlock(&ptlrpc_last_xid_lock);
3360 /* No need to lock, since returned value is racy anyways */
3361 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3364 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3367 * Functions for operating ptlrpc workers.
3369 * A ptlrpc work is a function which will be running inside ptlrpc context.
3370 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3372 * 1. after a work is created, it can be used many times, that is:
3373 * handler = ptlrpcd_alloc_work();
3374 * ptlrpcd_queue_work();
3376 * queue it again when necessary:
3377 * ptlrpcd_queue_work();
3378 * ptlrpcd_destroy_work();
3379 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3380 * it will only be queued once in any time. Also as its name implies, it may
3381 * have delay before it really runs by ptlrpcd thread.
3383 struct ptlrpc_work_async_args {
3384 int (*cb)(const struct lu_env *, void *);
3388 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3390 /* re-initialize the req */
3391 req->rq_timeout = obd_timeout;
3392 req->rq_sent = ktime_get_real_seconds();
3393 req->rq_deadline = req->rq_sent + req->rq_timeout;
3394 req->rq_phase = RQ_PHASE_INTERPRET;
3395 req->rq_next_phase = RQ_PHASE_COMPLETE;
3396 req->rq_xid = ptlrpc_next_xid();
3397 req->rq_import_generation = req->rq_import->imp_generation;
3399 ptlrpcd_add_req(req);
3402 static int work_interpreter(const struct lu_env *env,
3403 struct ptlrpc_request *req, void *data, int rc)
3405 struct ptlrpc_work_async_args *arg = data;
3407 LASSERT(ptlrpcd_check_work(req));
3408 LASSERT(arg->cb != NULL);
3410 rc = arg->cb(env, arg->cbdata);
3412 list_del_init(&req->rq_set_chain);
3415 if (atomic_dec_return(&req->rq_refcount) > 1) {
3416 atomic_set(&req->rq_refcount, 2);
3417 ptlrpcd_add_work_req(req);
3422 static int worker_format;
3424 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3426 return req->rq_pill.rc_fmt == (void *)&worker_format;
3430 * Create a work for ptlrpc.
3432 void *ptlrpcd_alloc_work(struct obd_import *imp,
3433 int (*cb)(const struct lu_env *, void *), void *cbdata)
3435 struct ptlrpc_request *req = NULL;
3436 struct ptlrpc_work_async_args *args;
3442 RETURN(ERR_PTR(-EINVAL));
3444 /* copy some code from deprecated fakereq. */
3445 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3447 CERROR("ptlrpc: run out of memory!\n");
3448 RETURN(ERR_PTR(-ENOMEM));
3451 ptlrpc_cli_req_init(req);
3453 req->rq_send_state = LUSTRE_IMP_FULL;
3454 req->rq_type = PTL_RPC_MSG_REQUEST;
3455 req->rq_import = class_import_get(imp);
3456 req->rq_interpret_reply = work_interpreter;
3457 /* don't want reply */
3458 req->rq_no_delay = req->rq_no_resend = 1;
3459 req->rq_pill.rc_fmt = (void *)&worker_format;
3461 CLASSERT(sizeof(*args) <= sizeof(req->rq_async_args));
3462 args = ptlrpc_req_async_args(req);
3464 args->cbdata = cbdata;
3468 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3470 void ptlrpcd_destroy_work(void *handler)
3472 struct ptlrpc_request *req = handler;
3475 ptlrpc_req_finished(req);
3477 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3479 int ptlrpcd_queue_work(void *handler)
3481 struct ptlrpc_request *req = handler;
3484 * Check if the req is already being queued.
3486 * Here comes a trick: it lacks a way of checking if a req is being
3487 * processed reliably in ptlrpc. Here I have to use refcount of req
3488 * for this purpose. This is okay because the caller should use this
3489 * req as opaque data. - Jinshan
3491 LASSERT(atomic_read(&req->rq_refcount) > 0);
3492 if (atomic_inc_return(&req->rq_refcount) == 2)
3493 ptlrpcd_add_work_req(req);
3496 EXPORT_SYMBOL(ptlrpcd_queue_work);