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, 2015, 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 <obd_support.h>
38 #include <obd_class.h>
39 #include <lustre_lib.h>
40 #include <lustre_ha.h>
41 #include <lustre_import.h>
42 #include <lustre_req_layout.h>
44 #include "ptlrpc_internal.h"
46 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
47 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
48 .release_frags = ptlrpc_release_bulk_page_pin,
50 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
52 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
53 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
54 .release_frags = NULL,
56 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
58 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
59 static int ptlrpcd_check_work(struct ptlrpc_request *req);
60 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
63 * Initialize passed in client structure \a cl.
65 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
66 struct ptlrpc_client *cl)
68 cl->cli_request_portal = req_portal;
69 cl->cli_reply_portal = rep_portal;
72 EXPORT_SYMBOL(ptlrpc_init_client);
75 * Return PortalRPC connection for remote uud \a uuid
77 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
79 struct ptlrpc_connection *c;
81 struct lnet_process_id peer;
85 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
86 * before accessing its values.
87 * coverity[uninit_use_in_call]
89 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
91 CNETERR("cannot find peer %s!\n", uuid->uuid);
95 c = ptlrpc_connection_get(peer, self, uuid);
97 memcpy(c->c_remote_uuid.uuid,
98 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
101 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
107 * Allocate and initialize new bulk descriptor on the sender.
108 * Returns pointer to the descriptor or NULL on error.
110 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
111 unsigned int max_brw,
112 enum ptlrpc_bulk_op_type type,
114 const struct ptlrpc_bulk_frag_ops *ops)
116 struct ptlrpc_bulk_desc *desc;
119 /* ensure that only one of KIOV or IOVEC is set but not both */
120 LASSERT((ptlrpc_is_bulk_desc_kiov(type) && ops->add_kiov_frag) ||
121 (ptlrpc_is_bulk_desc_kvec(type) && ops->add_iov_frag));
123 desc = kzalloc(sizeof(*desc), GFP_NOFS);
127 if (type & PTLRPC_BULK_BUF_KIOV) {
128 GET_KIOV(desc) = kcalloc(nfrags, sizeof(*GET_KIOV(desc)),
133 GET_KVEC(desc) = kcalloc(nfrags, sizeof(*GET_KVEC(desc)),
139 spin_lock_init(&desc->bd_lock);
140 init_waitqueue_head(&desc->bd_waitq);
141 desc->bd_max_iov = nfrags;
142 desc->bd_iov_count = 0;
143 desc->bd_portal = portal;
144 desc->bd_type = type;
145 desc->bd_md_count = 0;
146 desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *)ops;
147 LASSERT(max_brw > 0);
148 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
150 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
151 * node. Negotiated ocd_brw_size will always be <= this number.
153 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
154 LNetInvalidateMDHandle(&desc->bd_mds[i]);
163 * Prepare bulk descriptor for specified outgoing request \a req that
164 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
165 * the bulk to be sent. Used on client-side.
166 * Returns pointer to newly allocated initialized bulk descriptor or NULL on
169 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
171 unsigned int max_brw,
174 const struct ptlrpc_bulk_frag_ops *ops)
176 struct obd_import *imp = req->rq_import;
177 struct ptlrpc_bulk_desc *desc;
179 LASSERT(ptlrpc_is_bulk_op_passive(type));
181 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
185 desc->bd_import_generation = req->rq_import_generation;
186 desc->bd_import = class_import_get(imp);
189 desc->bd_cbid.cbid_fn = client_bulk_callback;
190 desc->bd_cbid.cbid_arg = desc;
192 /* This makes req own desc, and free it when she frees herself */
197 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
199 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
200 struct page *page, int pageoffset, int len, int pin)
202 struct bio_vec *kiov;
204 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
206 LASSERT(pageoffset >= 0);
208 LASSERT(pageoffset + len <= PAGE_SIZE);
209 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
211 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
218 kiov->bv_page = page;
219 kiov->bv_offset = pageoffset;
222 desc->bd_iov_count++;
224 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
226 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
231 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
234 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
236 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
240 iovec->iov_base = frag;
241 iovec->iov_len = len;
243 desc->bd_iov_count++;
247 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
249 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
251 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
252 LASSERT(desc->bd_md_count == 0); /* network hands off */
253 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
254 LASSERT(desc->bd_frag_ops);
256 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
257 sptlrpc_enc_pool_put_pages(desc);
260 class_export_put(desc->bd_export);
262 class_import_put(desc->bd_import);
264 if (desc->bd_frag_ops->release_frags)
265 desc->bd_frag_ops->release_frags(desc);
267 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
268 kfree(GET_KIOV(desc));
270 kfree(GET_KVEC(desc));
274 EXPORT_SYMBOL(ptlrpc_free_bulk);
277 * Set server timelimit for this req, i.e. how long are we willing to wait
278 * for reply before timing out this request.
280 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
286 LASSERT(req->rq_import);
292 * \a imp_server_timeout means this is reverse import and
293 * we send (currently only) ASTs to the client and cannot afford
294 * to wait too long for the reply, otherwise the other client
295 * (because of which we are sending this request) would
296 * timeout waiting for us
298 req->rq_timeout = req->rq_import->imp_server_timeout ?
299 obd_timeout / 2 : obd_timeout;
301 at = &req->rq_import->imp_at;
302 idx = import_at_get_index(req->rq_import,
303 req->rq_request_portal);
304 serv_est = at_get(&at->iat_service_estimate[idx]);
305 req->rq_timeout = at_est2timeout(serv_est);
308 * We could get even fancier here, using history to predict increased
313 * Let the server know what this RPC timeout is by putting it in the
316 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
318 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
320 /* Adjust max service estimate based on server value */
321 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
322 unsigned int serv_est)
328 LASSERT(req->rq_import);
329 at = &req->rq_import->imp_at;
331 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
333 * max service estimates are tracked on the server side,
334 * so just keep minimal history here
336 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
338 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
339 req->rq_import->imp_obd->obd_name, req->rq_request_portal,
340 oldse, at_get(&at->iat_service_estimate[idx]));
343 /* Expected network latency per remote node (secs) */
344 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
346 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
349 /* Adjust expected network latency */
350 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
351 unsigned int service_time)
353 unsigned int nl, oldnl;
355 time64_t now = ktime_get_real_seconds();
357 LASSERT(req->rq_import);
359 if (service_time > now - req->rq_sent + 3) {
361 * bz16408, however, this can also happen if early reply
362 * is lost and client RPC is expired and resent, early reply
363 * or reply of original RPC can still be fit in reply buffer
364 * of resent RPC, now client is measuring time from the
365 * resent time, but server sent back service time of original
368 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
369 D_ADAPTTO : D_WARNING,
370 "Reported service time %u > total measured time %lld\n",
371 service_time, now - req->rq_sent);
375 /* Network latency is total time less server processing time */
376 nl = max_t(int, now - req->rq_sent -
377 service_time, 0) + 1; /* st rounding */
378 at = &req->rq_import->imp_at;
380 oldnl = at_measured(&at->iat_net_latency, nl);
382 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) has changed from %d to %d\n",
383 req->rq_import->imp_obd->obd_name,
385 &req->rq_import->imp_connection->c_remote_uuid),
386 oldnl, at_get(&at->iat_net_latency));
389 static int unpack_reply(struct ptlrpc_request *req)
393 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
394 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
396 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
401 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
403 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
410 * Handle an early reply message, called with the rq_lock held.
411 * If anything goes wrong just ignore it - same as if it never happened
413 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
414 __must_hold(&req->rq_lock)
416 struct ptlrpc_request *early_req;
421 spin_unlock(&req->rq_lock);
423 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
425 spin_lock(&req->rq_lock);
429 rc = unpack_reply(early_req);
431 sptlrpc_cli_finish_early_reply(early_req);
432 spin_lock(&req->rq_lock);
437 * Use new timeout value just to adjust the local value for this
438 * request, don't include it into at_history. It is unclear yet why
439 * service time increased and should it be counted or skipped, e.g.
440 * that can be recovery case or some error or server, the real reply
441 * will add all new data if it is worth to add.
443 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
444 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
446 /* Network latency can be adjusted, it is pure network delays */
447 ptlrpc_at_adj_net_latency(req,
448 lustre_msg_get_service_time(early_req->rq_repmsg));
450 sptlrpc_cli_finish_early_reply(early_req);
452 spin_lock(&req->rq_lock);
453 olddl = req->rq_deadline;
455 * server assumes it now has rq_timeout from when the request
456 * arrived, so the client should give it at least that long.
457 * since we don't know the arrival time we'll use the original
460 req->rq_deadline = req->rq_sent + req->rq_timeout +
461 ptlrpc_at_get_net_latency(req);
463 DEBUG_REQ(D_ADAPTTO, req,
464 "Early reply #%d, new deadline in %lds (%lds)",
466 (long)(req->rq_deadline - ktime_get_real_seconds()),
467 (long)(req->rq_deadline - olddl));
472 static struct kmem_cache *request_cache;
474 int ptlrpc_request_cache_init(void)
476 request_cache = kmem_cache_create("ptlrpc_cache",
477 sizeof(struct ptlrpc_request),
478 0, SLAB_HWCACHE_ALIGN, NULL);
479 return !request_cache ? -ENOMEM : 0;
482 void ptlrpc_request_cache_fini(void)
484 kmem_cache_destroy(request_cache);
487 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
489 struct ptlrpc_request *req;
491 req = kmem_cache_zalloc(request_cache, flags);
495 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
497 kmem_cache_free(request_cache, req);
501 * Wind down request pool \a pool.
502 * Frees all requests from the pool too
504 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
506 struct list_head *l, *tmp;
507 struct ptlrpc_request *req;
509 spin_lock(&pool->prp_lock);
510 list_for_each_safe(l, tmp, &pool->prp_req_list) {
511 req = list_entry(l, struct ptlrpc_request, rq_list);
512 list_del(&req->rq_list);
513 LASSERT(req->rq_reqbuf);
514 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
515 kvfree(req->rq_reqbuf);
516 ptlrpc_request_cache_free(req);
518 spin_unlock(&pool->prp_lock);
521 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
524 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
526 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
531 while (size < pool->prp_rq_size)
534 LASSERTF(list_empty(&pool->prp_req_list) ||
535 size == pool->prp_rq_size,
536 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
537 pool->prp_rq_size, size);
539 spin_lock(&pool->prp_lock);
540 pool->prp_rq_size = size;
541 for (i = 0; i < num_rq; i++) {
542 struct ptlrpc_request *req;
543 struct lustre_msg *msg;
545 spin_unlock(&pool->prp_lock);
546 req = ptlrpc_request_cache_alloc(GFP_NOFS);
549 msg = libcfs_kvzalloc(size, GFP_NOFS);
551 ptlrpc_request_cache_free(req);
554 req->rq_reqbuf = msg;
555 req->rq_reqbuf_len = size;
557 spin_lock(&pool->prp_lock);
558 list_add_tail(&req->rq_list, &pool->prp_req_list);
560 spin_unlock(&pool->prp_lock);
563 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
566 * Create and initialize new request pool with given attributes:
567 * \a num_rq - initial number of requests to create for the pool
568 * \a msgsize - maximum message size possible for requests in thid pool
569 * \a populate_pool - function to be called when more requests need to be added
571 * Returns pointer to newly created pool or NULL on error.
573 struct ptlrpc_request_pool *
574 ptlrpc_init_rq_pool(int num_rq, int msgsize,
575 int (*populate_pool)(struct ptlrpc_request_pool *, int))
577 struct ptlrpc_request_pool *pool;
579 pool = kzalloc(sizeof(struct ptlrpc_request_pool), GFP_NOFS);
584 * Request next power of two for the allocation, because internally
585 * kernel would do exactly this
588 spin_lock_init(&pool->prp_lock);
589 INIT_LIST_HEAD(&pool->prp_req_list);
590 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
591 pool->prp_populate = populate_pool;
593 populate_pool(pool, num_rq);
597 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
600 * Fetches one request from pool \a pool
602 static struct ptlrpc_request *
603 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
605 struct ptlrpc_request *request;
606 struct lustre_msg *reqbuf;
611 spin_lock(&pool->prp_lock);
614 * See if we have anything in a pool, and bail out if nothing,
615 * in writeout path, where this matters, this is safe to do, because
616 * nothing is lost in this case, and when some in-flight requests
617 * complete, this code will be called again.
619 if (unlikely(list_empty(&pool->prp_req_list))) {
620 spin_unlock(&pool->prp_lock);
624 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
626 list_del_init(&request->rq_list);
627 spin_unlock(&pool->prp_lock);
629 LASSERT(request->rq_reqbuf);
630 LASSERT(request->rq_pool);
632 reqbuf = request->rq_reqbuf;
633 memset(request, 0, sizeof(*request));
634 request->rq_reqbuf = reqbuf;
635 request->rq_reqbuf_len = pool->prp_rq_size;
636 request->rq_pool = pool;
642 * Returns freed \a request to pool.
644 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
646 struct ptlrpc_request_pool *pool = request->rq_pool;
648 spin_lock(&pool->prp_lock);
649 LASSERT(list_empty(&request->rq_list));
650 LASSERT(!request->rq_receiving_reply);
651 list_add_tail(&request->rq_list, &pool->prp_req_list);
652 spin_unlock(&pool->prp_lock);
655 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
657 struct obd_import *imp = req->rq_import;
658 struct list_head *tmp;
659 struct ptlrpc_request *iter;
661 assert_spin_locked(&imp->imp_lock);
662 LASSERT(list_empty(&req->rq_unreplied_list));
664 /* unreplied list is sorted by xid in ascending order */
665 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
666 iter = list_entry(tmp, struct ptlrpc_request,
669 LASSERT(req->rq_xid != iter->rq_xid);
670 if (req->rq_xid < iter->rq_xid)
672 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
675 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
678 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
680 req->rq_xid = ptlrpc_next_xid();
681 ptlrpc_add_unreplied(req);
684 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
686 spin_lock(&req->rq_import->imp_lock);
687 ptlrpc_assign_next_xid_nolock(req);
688 spin_unlock(&req->rq_import->imp_lock);
691 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
692 __u32 version, int opcode, char **bufs,
693 struct ptlrpc_cli_ctx *ctx)
696 struct obd_import *imp;
700 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
701 imp = request->rq_import;
702 lengths = request->rq_pill.rc_area[RCL_CLIENT];
705 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
707 rc = sptlrpc_req_get_ctx(request);
711 sptlrpc_req_set_flavor(request, opcode);
713 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
718 lustre_msg_add_version(request->rq_reqmsg, version);
719 request->rq_send_state = LUSTRE_IMP_FULL;
720 request->rq_type = PTL_RPC_MSG_REQUEST;
722 request->rq_req_cbid.cbid_fn = request_out_callback;
723 request->rq_req_cbid.cbid_arg = request;
725 request->rq_reply_cbid.cbid_fn = reply_in_callback;
726 request->rq_reply_cbid.cbid_arg = request;
728 request->rq_reply_deadline = 0;
729 request->rq_bulk_deadline = 0;
730 request->rq_req_deadline = 0;
731 request->rq_phase = RQ_PHASE_NEW;
732 request->rq_next_phase = RQ_PHASE_UNDEFINED;
734 request->rq_request_portal = imp->imp_client->cli_request_portal;
735 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
737 ptlrpc_at_set_req_timeout(request);
739 lustre_msg_set_opc(request->rq_reqmsg, opcode);
740 ptlrpc_assign_next_xid(request);
742 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
743 if (cfs_fail_val == opcode) {
744 time64_t *fail_t = NULL, *fail2_t = NULL;
746 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
747 fail_t = &request->rq_bulk_deadline;
748 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
749 fail_t = &request->rq_reply_deadline;
750 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
751 fail_t = &request->rq_req_deadline;
752 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
753 fail_t = &request->rq_reply_deadline;
754 fail2_t = &request->rq_bulk_deadline;
758 *fail_t = ktime_get_real_seconds() + LONG_UNLINK;
761 *fail2_t = ktime_get_real_seconds() +
764 /* The RPC is infected, let the test change the
767 set_current_state(TASK_UNINTERRUPTIBLE);
768 schedule_timeout(cfs_time_seconds(2));
769 set_current_state(TASK_RUNNING);
776 LASSERT(!request->rq_pool);
777 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
779 class_import_put(imp);
782 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
785 * Pack request buffers for network transfer, performing necessary encryption
786 * steps if necessary.
788 int ptlrpc_request_pack(struct ptlrpc_request *request,
789 __u32 version, int opcode)
793 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
798 * For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
799 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
800 * have to send old ptlrpc_body to keep interoperability with these
803 * Only three kinds of server->client RPCs so far:
808 * XXX This should be removed whenever we drop the interoperability with
809 * the these old clients.
811 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
812 opcode == LDLM_GL_CALLBACK)
813 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
814 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
818 EXPORT_SYMBOL(ptlrpc_request_pack);
821 * Helper function to allocate new request on import \a imp
822 * and possibly using existing request from pool \a pool if provided.
823 * Returns allocated request structure with import field filled or
827 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
828 struct ptlrpc_request_pool *pool)
830 struct ptlrpc_request *request;
832 request = ptlrpc_request_cache_alloc(GFP_NOFS);
834 if (!request && pool)
835 request = ptlrpc_prep_req_from_pool(pool);
838 ptlrpc_cli_req_init(request);
840 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
841 LASSERT(imp != LP_POISON);
842 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
844 LASSERT(imp->imp_client != LP_POISON);
846 request->rq_import = class_import_get(imp);
848 CERROR("request allocation out of memory\n");
855 * Helper function for creating a request.
856 * Calls __ptlrpc_request_alloc to allocate new request structure and inits
857 * buffer structures according to capsule template \a format.
858 * Returns allocated request structure pointer or NULL on error.
860 static struct ptlrpc_request *
861 ptlrpc_request_alloc_internal(struct obd_import *imp,
862 struct ptlrpc_request_pool *pool,
863 const struct req_format *format)
865 struct ptlrpc_request *request;
867 request = __ptlrpc_request_alloc(imp, pool);
871 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
872 req_capsule_set(&request->rq_pill, format);
877 * Allocate new request structure for import \a imp and initialize its
878 * buffer structure according to capsule template \a format.
880 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
881 const struct req_format *format)
883 return ptlrpc_request_alloc_internal(imp, NULL, format);
885 EXPORT_SYMBOL(ptlrpc_request_alloc);
888 * Allocate new request structure for import \a imp from pool \a pool and
889 * initialize its buffer structure according to capsule template \a format.
891 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
892 struct ptlrpc_request_pool *pool,
893 const struct req_format *format)
895 return ptlrpc_request_alloc_internal(imp, pool, format);
897 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
900 * For requests not from pool, free memory of the request structure.
901 * For requests obtained from a pool earlier, return request back to pool.
903 void ptlrpc_request_free(struct ptlrpc_request *request)
905 if (request->rq_pool)
906 __ptlrpc_free_req_to_pool(request);
908 ptlrpc_request_cache_free(request);
910 EXPORT_SYMBOL(ptlrpc_request_free);
913 * Allocate new request for operation \a opcode and immediately pack it for
915 * Only used for simple requests like OBD_PING where the only important
916 * part of the request is operation itself.
917 * Returns allocated request or NULL on error.
919 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
920 const struct req_format *format,
921 __u32 version, int opcode)
923 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
927 rc = ptlrpc_request_pack(req, version, opcode);
929 ptlrpc_request_free(req);
935 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
938 * Allocate and initialize new request set structure on the current CPT.
939 * Returns a pointer to the newly allocated set structure or NULL on error.
941 struct ptlrpc_request_set *ptlrpc_prep_set(void)
943 struct ptlrpc_request_set *set;
946 cpt = cfs_cpt_current(cfs_cpt_table, 0);
947 set = kzalloc_node(sizeof(*set), GFP_NOFS,
948 cfs_cpt_spread_node(cfs_cpt_table, cpt));
951 atomic_set(&set->set_refcount, 1);
952 INIT_LIST_HEAD(&set->set_requests);
953 init_waitqueue_head(&set->set_waitq);
954 atomic_set(&set->set_new_count, 0);
955 atomic_set(&set->set_remaining, 0);
956 spin_lock_init(&set->set_new_req_lock);
957 INIT_LIST_HEAD(&set->set_new_requests);
958 INIT_LIST_HEAD(&set->set_cblist);
959 set->set_max_inflight = UINT_MAX;
960 set->set_producer = NULL;
961 set->set_producer_arg = NULL;
966 EXPORT_SYMBOL(ptlrpc_prep_set);
969 * Allocate and initialize new request set structure with flow control
970 * extension. This extension allows to control the number of requests in-flight
971 * for the whole set. A callback function to generate requests must be provided
972 * and the request set will keep the number of requests sent over the wire to
974 * Returns a pointer to the newly allocated set structure or NULL on error.
976 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
980 struct ptlrpc_request_set *set;
982 set = ptlrpc_prep_set();
986 set->set_max_inflight = max;
987 set->set_producer = func;
988 set->set_producer_arg = arg;
994 * Wind down and free request set structure previously allocated with
996 * Ensures that all requests on the set have completed and removes
997 * all requests from the request list in a set.
998 * If any unsent request happen to be on the list, pretends that they got
999 * an error in flight and calls their completion handler.
1001 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1003 struct list_head *tmp;
1004 struct list_head *next;
1008 /* Requests on the set should either all be completed, or all be new */
1009 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1010 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1011 list_for_each(tmp, &set->set_requests) {
1012 struct ptlrpc_request *req =
1013 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1015 LASSERT(req->rq_phase == expected_phase);
1019 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1020 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1021 atomic_read(&set->set_remaining), n);
1023 list_for_each_safe(tmp, next, &set->set_requests) {
1024 struct ptlrpc_request *req =
1025 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1026 list_del_init(&req->rq_set_chain);
1028 LASSERT(req->rq_phase == expected_phase);
1030 if (req->rq_phase == RQ_PHASE_NEW) {
1031 ptlrpc_req_interpret(NULL, req, -EBADR);
1032 atomic_dec(&set->set_remaining);
1035 spin_lock(&req->rq_lock);
1037 req->rq_invalid_rqset = 0;
1038 spin_unlock(&req->rq_lock);
1040 ptlrpc_req_finished(req);
1043 LASSERT(atomic_read(&set->set_remaining) == 0);
1045 ptlrpc_reqset_put(set);
1047 EXPORT_SYMBOL(ptlrpc_set_destroy);
1050 * Add a new request to the general purpose request set.
1051 * Assumes request reference from the caller.
1053 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1054 struct ptlrpc_request *req)
1056 LASSERT(list_empty(&req->rq_set_chain));
1058 /* The set takes over the caller's request reference */
1059 list_add_tail(&req->rq_set_chain, &set->set_requests);
1061 atomic_inc(&set->set_remaining);
1062 req->rq_queued_time = cfs_time_current();
1065 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1067 if (set->set_producer)
1069 * If the request set has a producer callback, the RPC must be
1070 * sent straight away
1072 ptlrpc_send_new_req(req);
1074 EXPORT_SYMBOL(ptlrpc_set_add_req);
1077 * Add a request to a request with dedicated server thread
1078 * and wake the thread to make any necessary processing.
1079 * Currently only used for ptlrpcd.
1081 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1082 struct ptlrpc_request *req)
1084 struct ptlrpc_request_set *set = pc->pc_set;
1087 LASSERT(!req->rq_set);
1088 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1090 spin_lock(&set->set_new_req_lock);
1091 /* The set takes over the caller's request reference. */
1093 req->rq_queued_time = cfs_time_current();
1094 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1095 count = atomic_inc_return(&set->set_new_count);
1096 spin_unlock(&set->set_new_req_lock);
1098 /* Only need to call wakeup once for the first entry. */
1100 wake_up(&set->set_waitq);
1103 * XXX: It maybe unnecessary to wakeup all the partners. But to
1104 * guarantee the async RPC can be processed ASAP, we have
1105 * no other better choice. It maybe fixed in future.
1107 for (i = 0; i < pc->pc_npartners; i++)
1108 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1113 * Based on the current state of the import, determine if the request
1114 * can be sent, is an error, or should be delayed.
1116 * Returns true if this request should be delayed. If false, and
1117 * *status is set, then the request can not be sent and *status is the
1118 * error code. If false and status is 0, then request can be sent.
1120 * The imp->imp_lock must be held.
1122 static int ptlrpc_import_delay_req(struct obd_import *imp,
1123 struct ptlrpc_request *req, int *status)
1129 if (req->rq_ctx_init || req->rq_ctx_fini) {
1130 /* always allow ctx init/fini rpc go through */
1131 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1132 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1134 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1135 /* pings may safely race with umount */
1136 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1137 D_HA : D_ERROR, req, "IMP_CLOSED ");
1139 } else if (ptlrpc_send_limit_expired(req)) {
1140 /* probably doesn't need to be a D_ERROR after initial testing */
1141 DEBUG_REQ(D_HA, req, "send limit expired ");
1142 *status = -ETIMEDOUT;
1143 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1144 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1145 /* allow CONNECT even if import is invalid */
1146 if (atomic_read(&imp->imp_inval_count) != 0) {
1147 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1150 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1151 if (!imp->imp_deactive)
1152 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1153 *status = -ESHUTDOWN; /* bz 12940 */
1154 } else if (req->rq_import_generation != imp->imp_generation) {
1155 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1157 } else if (req->rq_send_state != imp->imp_state) {
1158 /* invalidate in progress - any requests should be drop */
1159 if (atomic_read(&imp->imp_inval_count) != 0) {
1160 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1162 } else if (req->rq_no_delay) {
1163 *status = -EWOULDBLOCK;
1164 } else if (req->rq_allow_replay &&
1165 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1166 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1167 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1168 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1169 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1179 * Decide if the error message should be printed to the console or not.
1180 * Makes its decision based on request type, status, and failure frequency.
1182 * \param[in] req request that failed and may need a console message
1184 * \retval false if no message should be printed
1185 * \retval true if console message should be printed
1187 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1191 LASSERT(req->rq_reqmsg);
1192 opc = lustre_msg_get_opc(req->rq_reqmsg);
1194 /* Suppress particular reconnect errors which are to be expected. */
1195 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1198 /* Suppress timed out reconnect requests */
1199 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1204 * Suppress most unavailable/again reconnect requests, but
1205 * print occasionally so it is clear client is trying to
1206 * connect to a server where no target is running.
1208 err = lustre_msg_get_status(req->rq_repmsg);
1209 if ((err == -ENODEV || err == -EAGAIN) &&
1210 req->rq_import->imp_conn_cnt % 30 != 20)
1218 * Check request processing status.
1219 * Returns the status.
1221 static int ptlrpc_check_status(struct ptlrpc_request *req)
1225 err = lustre_msg_get_status(req->rq_repmsg);
1226 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1227 struct obd_import *imp = req->rq_import;
1228 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1229 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1231 /* -EAGAIN is normal when using POSIX flocks */
1232 if (ptlrpc_console_allow(req) &&
1233 !(opc == LDLM_ENQUEUE && err == -EAGAIN))
1234 LCONSOLE_ERROR_MSG(0x011, "%s: operation %s to node %s failed: rc = %d\n",
1235 imp->imp_obd->obd_name,
1237 libcfs_nid2str(nid), err);
1238 return err < 0 ? err : -EINVAL;
1242 DEBUG_REQ(D_INFO, req, "status is %d", err);
1244 /* XXX: translate this error from net to host */
1245 DEBUG_REQ(D_INFO, req, "status is %d", err);
1251 * save pre-versions of objects into request for replay.
1252 * Versions are obtained from server reply.
1255 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1257 struct lustre_msg *repmsg = req->rq_repmsg;
1258 struct lustre_msg *reqmsg = req->rq_reqmsg;
1259 __u64 *versions = lustre_msg_get_versions(repmsg);
1261 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1265 lustre_msg_set_versions(reqmsg, versions);
1266 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1267 versions[0], versions[1]);
1270 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1272 struct ptlrpc_request *req;
1274 assert_spin_locked(&imp->imp_lock);
1275 if (list_empty(&imp->imp_unreplied_list))
1278 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1280 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1282 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1283 imp->imp_known_replied_xid = req->rq_xid - 1;
1285 return req->rq_xid - 1;
1289 * Callback function called when client receives RPC reply for \a req.
1290 * Returns 0 on success or error code.
1291 * The return value would be assigned to req->rq_status by the caller
1292 * as request processing status.
1293 * This function also decides if the request needs to be saved for later replay.
1295 static int after_reply(struct ptlrpc_request *req)
1297 struct obd_import *imp = req->rq_import;
1298 struct obd_device *obd = req->rq_import->imp_obd;
1300 struct timespec64 work_start;
1305 /* repbuf must be unlinked */
1306 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1308 if (req->rq_reply_truncated) {
1309 if (ptlrpc_no_resend(req)) {
1310 DEBUG_REQ(D_ERROR, req, "reply buffer overflow, expected: %d, actual size: %d",
1311 req->rq_nob_received, req->rq_repbuf_len);
1315 sptlrpc_cli_free_repbuf(req);
1317 * Pass the required reply buffer size (include space for early
1318 * reply). NB: no need to round up because alloc_repbuf will
1321 req->rq_replen = req->rq_nob_received;
1322 req->rq_nob_received = 0;
1323 spin_lock(&req->rq_lock);
1325 spin_unlock(&req->rq_lock);
1329 ktime_get_real_ts64(&work_start);
1330 timediff = (work_start.tv_sec - req->rq_sent_tv.tv_sec) * USEC_PER_SEC +
1331 (work_start.tv_nsec - req->rq_sent_tv.tv_nsec) /
1334 * NB Until this point, the whole of the incoming message,
1335 * including buflens, status etc is in the sender's byte order.
1337 rc = sptlrpc_cli_unwrap_reply(req);
1339 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1343 /* Security layer unwrap might ask resend this request. */
1347 rc = unpack_reply(req);
1351 /* retry indefinitely on EINPROGRESS */
1352 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1353 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1354 time64_t now = ktime_get_real_seconds();
1356 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1357 spin_lock(&req->rq_lock);
1359 spin_unlock(&req->rq_lock);
1360 req->rq_nr_resend++;
1362 /* Readjust the timeout for current conditions */
1363 ptlrpc_at_set_req_timeout(req);
1365 * delay resend to give a chance to the server to get ready.
1366 * The delay is increased by 1s on every resend and is capped to
1367 * the current request timeout (i.e. obd_timeout if AT is off,
1368 * or AT service time x 125% + 5s, see at_est2timeout)
1370 if (req->rq_nr_resend > req->rq_timeout)
1371 req->rq_sent = now + req->rq_timeout;
1373 req->rq_sent = now + req->rq_nr_resend;
1375 /* Resend for EINPROGRESS will use a new XID */
1376 spin_lock(&imp->imp_lock);
1377 list_del_init(&req->rq_unreplied_list);
1378 spin_unlock(&imp->imp_lock);
1383 if (obd->obd_svc_stats) {
1384 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1386 ptlrpc_lprocfs_rpc_sent(req, timediff);
1389 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1390 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1391 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1392 lustre_msg_get_type(req->rq_repmsg));
1396 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1397 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1398 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1399 ptlrpc_at_adj_net_latency(req,
1400 lustre_msg_get_service_time(req->rq_repmsg));
1402 rc = ptlrpc_check_status(req);
1403 imp->imp_connect_error = rc;
1407 * Either we've been evicted, or the server has failed for
1408 * some reason. Try to reconnect, and if that fails, punt to
1411 if (ptlrpc_recoverable_error(rc)) {
1412 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1413 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1416 ptlrpc_request_handle_notconn(req);
1421 * Let's look if server sent slv. Do it only for RPC with
1424 ldlm_cli_update_pool(req);
1427 /* Store transno in reqmsg for replay. */
1428 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1429 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1430 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1433 if (imp->imp_replayable) {
1434 spin_lock(&imp->imp_lock);
1436 * No point in adding already-committed requests to the replay
1437 * list, we will just remove them immediately. b=9829
1439 if (req->rq_transno != 0 &&
1441 lustre_msg_get_last_committed(req->rq_repmsg) ||
1443 /* version recovery */
1444 ptlrpc_save_versions(req);
1445 ptlrpc_retain_replayable_request(req, imp);
1446 } else if (req->rq_commit_cb &&
1447 list_empty(&req->rq_replay_list)) {
1449 * NB: don't call rq_commit_cb if it's already on
1450 * rq_replay_list, ptlrpc_free_committed() will call
1451 * it later, see LU-3618 for details
1453 spin_unlock(&imp->imp_lock);
1454 req->rq_commit_cb(req);
1455 spin_lock(&imp->imp_lock);
1458 /* Replay-enabled imports return commit-status information. */
1459 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1460 if (likely(committed > imp->imp_peer_committed_transno))
1461 imp->imp_peer_committed_transno = committed;
1463 ptlrpc_free_committed(imp);
1465 if (!list_empty(&imp->imp_replay_list)) {
1466 struct ptlrpc_request *last;
1468 last = list_entry(imp->imp_replay_list.prev,
1469 struct ptlrpc_request,
1472 * Requests with rq_replay stay on the list even if no
1473 * commit is expected.
1475 if (last->rq_transno > imp->imp_peer_committed_transno)
1476 ptlrpc_pinger_commit_expected(imp);
1479 spin_unlock(&imp->imp_lock);
1486 * Helper function to send request \a req over the network for the first time
1487 * Also adjusts request phase.
1488 * Returns 0 on success or error code.
1490 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1492 struct obd_import *imp = req->rq_import;
1496 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1498 /* do not try to go further if there is not enough memory in enc_pool */
1499 if (req->rq_sent && req->rq_bulk)
1500 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1501 pool_is_at_full_capacity())
1504 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1505 (!req->rq_generation_set ||
1506 req->rq_import_generation == imp->imp_generation))
1509 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1511 spin_lock(&imp->imp_lock);
1513 LASSERT(req->rq_xid);
1514 LASSERT(!list_empty(&req->rq_unreplied_list));
1516 if (!req->rq_generation_set)
1517 req->rq_import_generation = imp->imp_generation;
1519 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1520 spin_lock(&req->rq_lock);
1521 req->rq_waiting = 1;
1522 spin_unlock(&req->rq_lock);
1524 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: (%s != %s)",
1525 lustre_msg_get_status(req->rq_reqmsg),
1526 ptlrpc_import_state_name(req->rq_send_state),
1527 ptlrpc_import_state_name(imp->imp_state));
1528 LASSERT(list_empty(&req->rq_list));
1529 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1530 atomic_inc(&req->rq_import->imp_inflight);
1531 spin_unlock(&imp->imp_lock);
1536 spin_unlock(&imp->imp_lock);
1537 req->rq_status = rc;
1538 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1542 LASSERT(list_empty(&req->rq_list));
1543 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1544 atomic_inc(&req->rq_import->imp_inflight);
1546 /* find the known replied XID from the unreplied list, CONNECT
1547 * and DISCONNECT requests are skipped to make the sanity check
1548 * on server side happy. see process_req_last_xid().
1550 * For CONNECT: Because replay requests have lower XID, it'll
1551 * break the sanity check if CONNECT bump the exp_last_xid on
1554 * For DISCONNECT: Since client will abort inflight RPC before
1555 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1556 * than the inflight RPC.
1558 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1559 min_xid = ptlrpc_known_replied_xid(imp);
1560 spin_unlock(&imp->imp_lock);
1562 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1564 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1566 rc = sptlrpc_req_refresh_ctx(req, -1);
1569 req->rq_status = rc;
1572 spin_lock(&req->rq_lock);
1573 req->rq_wait_ctx = 1;
1574 spin_unlock(&req->rq_lock);
1578 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc %s:%s:%d:%llu:%s:%d\n",
1580 imp->imp_obd->obd_uuid.uuid,
1581 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1582 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1583 lustre_msg_get_opc(req->rq_reqmsg));
1585 rc = ptl_send_rpc(req, 0);
1586 if (rc == -ENOMEM) {
1587 spin_lock(&imp->imp_lock);
1588 if (!list_empty(&req->rq_list)) {
1589 list_del_init(&req->rq_list);
1590 atomic_dec(&req->rq_import->imp_inflight);
1592 spin_unlock(&imp->imp_lock);
1593 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1597 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1598 spin_lock(&req->rq_lock);
1599 req->rq_net_err = 1;
1600 spin_unlock(&req->rq_lock);
1606 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1610 LASSERT(set->set_producer);
1612 remaining = atomic_read(&set->set_remaining);
1615 * populate the ->set_requests list with requests until we
1616 * reach the maximum number of RPCs in flight for this set
1618 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1619 rc = set->set_producer(set, set->set_producer_arg);
1620 if (rc == -ENOENT) {
1621 /* no more RPC to produce */
1622 set->set_producer = NULL;
1623 set->set_producer_arg = NULL;
1628 return (atomic_read(&set->set_remaining) - remaining);
1632 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1633 * and no more replies are expected.
1634 * (it is possible to get less replies than requests sent e.g. due to timed out
1635 * requests or requests that we had trouble to send out)
1637 * NOTE: This function contains a potential schedule point (cond_resched()).
1639 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1641 struct list_head *tmp, *next;
1642 struct list_head comp_reqs;
1643 int force_timer_recalc = 0;
1645 if (atomic_read(&set->set_remaining) == 0)
1648 INIT_LIST_HEAD(&comp_reqs);
1649 list_for_each_safe(tmp, next, &set->set_requests) {
1650 struct ptlrpc_request *req =
1651 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1652 struct obd_import *imp = req->rq_import;
1653 int unregistered = 0;
1657 * This schedule point is mainly for the ptlrpcd caller of this
1658 * function. Most ptlrpc sets are not long-lived and unbounded
1659 * in length, but at the least the set used by the ptlrpcd is.
1660 * Since the processing time is unbounded, we need to insert an
1661 * explicit schedule point to make the thread well-behaved.
1665 if (req->rq_phase == RQ_PHASE_NEW &&
1666 ptlrpc_send_new_req(req)) {
1667 force_timer_recalc = 1;
1670 /* delayed send - skip */
1671 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1674 /* delayed resend - skip */
1675 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1676 req->rq_sent > ktime_get_real_seconds())
1679 if (!(req->rq_phase == RQ_PHASE_RPC ||
1680 req->rq_phase == RQ_PHASE_BULK ||
1681 req->rq_phase == RQ_PHASE_INTERPRET ||
1682 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1683 req->rq_phase == RQ_PHASE_UNREG_BULK ||
1684 req->rq_phase == RQ_PHASE_COMPLETE)) {
1685 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1689 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1690 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1691 LASSERT(req->rq_next_phase != req->rq_phase);
1692 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1694 if (req->rq_req_deadline &&
1695 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1696 req->rq_req_deadline = 0;
1697 if (req->rq_reply_deadline &&
1698 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1699 req->rq_reply_deadline = 0;
1700 if (req->rq_bulk_deadline &&
1701 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1702 req->rq_bulk_deadline = 0;
1705 * Skip processing until reply is unlinked. We
1706 * can't return to pool before that and we can't
1707 * call interpret before that. We need to make
1708 * sure that all rdma transfers finished and will
1709 * not corrupt any data.
1711 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1712 ptlrpc_client_recv_or_unlink(req))
1714 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1715 ptlrpc_client_bulk_active(req))
1719 * Turn fail_loc off to prevent it from looping
1722 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1723 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1726 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1727 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1731 /* Move to next phase if reply was successfully
1734 ptlrpc_rqphase_move(req, req->rq_next_phase);
1737 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1738 list_move_tail(&req->rq_set_chain, &comp_reqs);
1742 if (req->rq_phase == RQ_PHASE_INTERPRET)
1745 /* Note that this also will start async reply unlink. */
1746 if (req->rq_net_err && !req->rq_timedout) {
1747 ptlrpc_expire_one_request(req, 1);
1749 /* Check if we still need to wait for unlink. */
1750 if (ptlrpc_client_recv_or_unlink(req) ||
1751 ptlrpc_client_bulk_active(req))
1753 /* If there is no need to resend, fail it now. */
1754 if (req->rq_no_resend) {
1755 if (req->rq_status == 0)
1756 req->rq_status = -EIO;
1757 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1765 spin_lock(&req->rq_lock);
1766 req->rq_replied = 0;
1767 spin_unlock(&req->rq_lock);
1768 if (req->rq_status == 0)
1769 req->rq_status = -EIO;
1770 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1775 * ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1776 * so it sets rq_intr regardless of individual rpc
1777 * timeouts. The synchronous IO waiting path sets
1778 * rq_intr irrespective of whether ptlrpcd
1779 * has seen a timeout. Our policy is to only interpret
1780 * interrupted rpcs after they have timed out, so we
1781 * need to enforce that here.
1784 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1785 req->rq_wait_ctx)) {
1786 req->rq_status = -EINTR;
1787 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1791 if (req->rq_phase == RQ_PHASE_RPC) {
1792 if (req->rq_timedout || req->rq_resend ||
1793 req->rq_waiting || req->rq_wait_ctx) {
1796 if (!ptlrpc_unregister_reply(req, 1)) {
1797 ptlrpc_unregister_bulk(req, 1);
1801 spin_lock(&imp->imp_lock);
1802 if (ptlrpc_import_delay_req(imp, req,
1805 * put on delay list - only if we wait
1806 * recovery finished - before send
1808 list_del_init(&req->rq_list);
1809 list_add_tail(&req->rq_list,
1810 &imp->imp_delayed_list);
1811 spin_unlock(&imp->imp_lock);
1816 req->rq_status = status;
1817 ptlrpc_rqphase_move(req,
1818 RQ_PHASE_INTERPRET);
1819 spin_unlock(&imp->imp_lock);
1822 if (ptlrpc_no_resend(req) &&
1823 !req->rq_wait_ctx) {
1824 req->rq_status = -ENOTCONN;
1825 ptlrpc_rqphase_move(req,
1826 RQ_PHASE_INTERPRET);
1827 spin_unlock(&imp->imp_lock);
1831 list_del_init(&req->rq_list);
1832 list_add_tail(&req->rq_list,
1833 &imp->imp_sending_list);
1835 spin_unlock(&imp->imp_lock);
1837 spin_lock(&req->rq_lock);
1838 req->rq_waiting = 0;
1839 spin_unlock(&req->rq_lock);
1841 if (req->rq_timedout || req->rq_resend) {
1842 /* This is re-sending anyway, let's mark req as resend. */
1843 spin_lock(&req->rq_lock);
1845 spin_unlock(&req->rq_lock);
1847 !ptlrpc_unregister_bulk(req, 1))
1851 * rq_wait_ctx is only touched by ptlrpcd,
1852 * so no lock is needed here.
1854 status = sptlrpc_req_refresh_ctx(req, -1);
1857 req->rq_status = status;
1858 spin_lock(&req->rq_lock);
1859 req->rq_wait_ctx = 0;
1860 spin_unlock(&req->rq_lock);
1861 force_timer_recalc = 1;
1863 spin_lock(&req->rq_lock);
1864 req->rq_wait_ctx = 1;
1865 spin_unlock(&req->rq_lock);
1870 spin_lock(&req->rq_lock);
1871 req->rq_wait_ctx = 0;
1872 spin_unlock(&req->rq_lock);
1875 rc = ptl_send_rpc(req, 0);
1876 if (rc == -ENOMEM) {
1877 spin_lock(&imp->imp_lock);
1878 if (!list_empty(&req->rq_list))
1879 list_del_init(&req->rq_list);
1880 spin_unlock(&imp->imp_lock);
1881 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1885 DEBUG_REQ(D_HA, req,
1886 "send failed: rc = %d", rc);
1887 force_timer_recalc = 1;
1888 spin_lock(&req->rq_lock);
1889 req->rq_net_err = 1;
1890 spin_unlock(&req->rq_lock);
1893 /* need to reset the timeout */
1894 force_timer_recalc = 1;
1897 spin_lock(&req->rq_lock);
1899 if (ptlrpc_client_early(req)) {
1900 ptlrpc_at_recv_early_reply(req);
1901 spin_unlock(&req->rq_lock);
1905 /* Still waiting for a reply? */
1906 if (ptlrpc_client_recv(req)) {
1907 spin_unlock(&req->rq_lock);
1911 /* Did we actually receive a reply? */
1912 if (!ptlrpc_client_replied(req)) {
1913 spin_unlock(&req->rq_lock);
1917 spin_unlock(&req->rq_lock);
1920 * unlink from net because we are going to
1921 * swab in-place of reply buffer
1923 unregistered = ptlrpc_unregister_reply(req, 1);
1927 req->rq_status = after_reply(req);
1932 * If there is no bulk associated with this request,
1933 * then we're done and should let the interpreter
1934 * process the reply. Similarly if the RPC returned
1935 * an error, and therefore the bulk will never arrive.
1937 if (!req->rq_bulk || req->rq_status < 0) {
1938 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1942 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1945 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1946 if (ptlrpc_client_bulk_active(req))
1949 if (req->rq_bulk->bd_failure) {
1951 * The RPC reply arrived OK, but the bulk screwed
1952 * up! Dead weird since the server told us the RPC
1953 * was good after getting the REPLY for her GET or
1954 * the ACK for her PUT.
1956 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1957 req->rq_status = -EIO;
1960 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1963 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1966 * This moves to "unregistering" phase we need to wait for
1969 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1970 /* start async bulk unlink too */
1971 ptlrpc_unregister_bulk(req, 1);
1975 if (!ptlrpc_unregister_bulk(req, 1))
1978 /* When calling interpret receive should already be finished. */
1979 LASSERT(!req->rq_receiving_reply);
1981 ptlrpc_req_interpret(env, req, req->rq_status);
1983 if (ptlrpcd_check_work(req)) {
1984 atomic_dec(&set->set_remaining);
1987 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1989 CDEBUG(req->rq_reqmsg ? D_RPCTRACE : 0,
1990 "Completed RPC pname:cluuid:pid:xid:nid:opc %s:%s:%d:%llu:%s:%d\n",
1991 current_comm(), imp->imp_obd->obd_uuid.uuid,
1992 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1993 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1994 lustre_msg_get_opc(req->rq_reqmsg));
1996 spin_lock(&imp->imp_lock);
1998 * Request already may be not on sending or delaying list. This
1999 * may happen in the case of marking it erroneous for the case
2000 * ptlrpc_import_delay_req(req, status) find it impossible to
2001 * allow sending this rpc and returns *status != 0.
2003 if (!list_empty(&req->rq_list)) {
2004 list_del_init(&req->rq_list);
2005 atomic_dec(&imp->imp_inflight);
2007 list_del_init(&req->rq_unreplied_list);
2008 spin_unlock(&imp->imp_lock);
2010 atomic_dec(&set->set_remaining);
2011 wake_up_all(&imp->imp_recovery_waitq);
2013 if (set->set_producer) {
2014 /* produce a new request if possible */
2015 if (ptlrpc_set_producer(set) > 0)
2016 force_timer_recalc = 1;
2019 * free the request that has just been completed
2020 * in order not to pollute set->set_requests
2022 list_del_init(&req->rq_set_chain);
2023 spin_lock(&req->rq_lock);
2025 req->rq_invalid_rqset = 0;
2026 spin_unlock(&req->rq_lock);
2028 /* record rq_status to compute the final status later */
2029 if (req->rq_status != 0)
2030 set->set_rc = req->rq_status;
2031 ptlrpc_req_finished(req);
2033 list_move_tail(&req->rq_set_chain, &comp_reqs);
2038 * move completed request at the head of list so it's easier for
2039 * caller to find them
2041 list_splice(&comp_reqs, &set->set_requests);
2043 /* If we hit an error, we want to recover promptly. */
2044 return atomic_read(&set->set_remaining) == 0 || force_timer_recalc;
2046 EXPORT_SYMBOL(ptlrpc_check_set);
2049 * Time out request \a req. is \a async_unlink is set, that means do not wait
2050 * until LNet actually confirms network buffer unlinking.
2051 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2053 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2055 struct obd_import *imp = req->rq_import;
2058 spin_lock(&req->rq_lock);
2059 req->rq_timedout = 1;
2060 spin_unlock(&req->rq_lock);
2062 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent %lld/real %lld]",
2063 req->rq_net_err ? "failed due to network error" :
2064 ((req->rq_real_sent == 0 ||
2065 req->rq_real_sent < req->rq_sent ||
2066 req->rq_real_sent >= req->rq_deadline) ?
2067 "timed out for sent delay" : "timed out for slow reply"),
2068 (s64)req->rq_sent, (s64)req->rq_real_sent);
2070 if (imp && obd_debug_peer_on_timeout)
2071 LNetDebugPeer(imp->imp_connection->c_peer);
2073 ptlrpc_unregister_reply(req, async_unlink);
2074 ptlrpc_unregister_bulk(req, async_unlink);
2076 if (obd_dump_on_timeout)
2077 libcfs_debug_dumplog();
2080 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2084 atomic_inc(&imp->imp_timeouts);
2086 /* The DLM server doesn't want recovery run on its imports. */
2087 if (imp->imp_dlm_fake)
2091 * If this request is for recovery or other primordial tasks,
2092 * then error it out here.
2094 if (req->rq_ctx_init || req->rq_ctx_fini ||
2095 req->rq_send_state != LUSTRE_IMP_FULL ||
2096 imp->imp_obd->obd_no_recov) {
2097 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2098 ptlrpc_import_state_name(req->rq_send_state),
2099 ptlrpc_import_state_name(imp->imp_state));
2100 spin_lock(&req->rq_lock);
2101 req->rq_status = -ETIMEDOUT;
2103 spin_unlock(&req->rq_lock);
2108 * if a request can't be resent we can't wait for an answer after
2111 if (ptlrpc_no_resend(req)) {
2112 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2116 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2122 * Time out all uncompleted requests in request set pointed by \a data
2123 * Callback used when waiting on sets with l_wait_event.
2126 int ptlrpc_expired_set(void *data)
2128 struct ptlrpc_request_set *set = data;
2129 struct list_head *tmp;
2130 time64_t now = ktime_get_real_seconds();
2132 /* A timeout expired. See which reqs it applies to... */
2133 list_for_each(tmp, &set->set_requests) {
2134 struct ptlrpc_request *req =
2135 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2137 /* don't expire request waiting for context */
2138 if (req->rq_wait_ctx)
2141 /* Request in-flight? */
2142 if (!((req->rq_phase == RQ_PHASE_RPC &&
2143 !req->rq_waiting && !req->rq_resend) ||
2144 (req->rq_phase == RQ_PHASE_BULK)))
2147 if (req->rq_timedout || /* already dealt with */
2148 req->rq_deadline > now) /* not expired */
2152 * Deal with this guy. Do it asynchronously to not block
2155 ptlrpc_expire_one_request(req, 1);
2159 * When waiting for a whole set, we always break out of the
2160 * sleep so we can recalculate the timeout, or enable interrupts
2161 * if everyone's timed out.
2167 * Sets rq_intr flag in \a req under spinlock.
2169 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2171 spin_lock(&req->rq_lock);
2173 spin_unlock(&req->rq_lock);
2175 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2178 * Interrupts (sets interrupted flag) all uncompleted requests in
2179 * a set \a data. Callback for l_wait_event for interruptible waits.
2181 static void ptlrpc_interrupted_set(void *data)
2183 struct ptlrpc_request_set *set = data;
2184 struct list_head *tmp;
2186 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2188 list_for_each(tmp, &set->set_requests) {
2189 struct ptlrpc_request *req =
2190 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2192 if (req->rq_phase != RQ_PHASE_RPC &&
2193 req->rq_phase != RQ_PHASE_UNREG_RPC)
2196 ptlrpc_mark_interrupted(req);
2201 * Get the smallest timeout in the set; this does NOT set a timeout.
2203 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2205 struct list_head *tmp;
2206 time64_t now = ktime_get_real_seconds();
2208 struct ptlrpc_request *req;
2211 list_for_each(tmp, &set->set_requests) {
2212 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2214 /* Request in-flight? */
2215 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2216 (req->rq_phase == RQ_PHASE_BULK) ||
2217 (req->rq_phase == RQ_PHASE_NEW)))
2220 /* Already timed out. */
2221 if (req->rq_timedout)
2224 /* Waiting for ctx. */
2225 if (req->rq_wait_ctx)
2228 if (req->rq_phase == RQ_PHASE_NEW)
2229 deadline = req->rq_sent;
2230 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2231 deadline = req->rq_sent;
2233 deadline = req->rq_sent + req->rq_timeout;
2235 if (deadline <= now) /* actually expired already */
2236 timeout = 1; /* ASAP */
2237 else if (timeout == 0 || timeout > deadline - now)
2238 timeout = deadline - now;
2244 * Send all unset request from the set and then wait until all
2245 * requests in the set complete (either get a reply, timeout, get an
2246 * error or otherwise be interrupted).
2247 * Returns 0 on success or error code otherwise.
2249 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2251 struct list_head *tmp;
2252 struct ptlrpc_request *req;
2253 struct l_wait_info lwi;
2256 if (set->set_producer)
2257 (void)ptlrpc_set_producer(set);
2259 list_for_each(tmp, &set->set_requests) {
2260 req = list_entry(tmp, struct ptlrpc_request,
2262 if (req->rq_phase == RQ_PHASE_NEW)
2263 (void)ptlrpc_send_new_req(req);
2266 if (list_empty(&set->set_requests))
2270 timeout = ptlrpc_set_next_timeout(set);
2273 * wait until all complete, interrupted, or an in-flight
2276 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2279 if (timeout == 0 && !signal_pending(current))
2281 * No requests are in-flight (ether timed out
2282 * or delayed), so we can allow interrupts.
2283 * We still want to block for a limited time,
2284 * so we allow interrupts during the timeout.
2286 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2288 ptlrpc_interrupted_set, set);
2291 * At least one request is in flight, so no
2292 * interrupts are allowed. Wait until all
2293 * complete, or an in-flight req times out.
2295 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout ? timeout : 1),
2296 ptlrpc_expired_set, set);
2298 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2301 * LU-769 - if we ignored the signal because it was already
2302 * pending when we started, we need to handle it now or we risk
2303 * it being ignored forever
2305 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2306 signal_pending(current)) {
2307 sigset_t blocked_sigs =
2308 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2311 * In fact we only interrupt for the "fatal" signals
2312 * like SIGINT or SIGKILL. We still ignore less
2313 * important signals since ptlrpc set is not easily
2314 * reentrant from userspace again
2316 if (signal_pending(current))
2317 ptlrpc_interrupted_set(set);
2318 cfs_restore_sigs(blocked_sigs);
2321 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2324 * -EINTR => all requests have been flagged rq_intr so next
2326 * -ETIMEDOUT => someone timed out. When all reqs have
2327 * timed out, signals are enabled allowing completion with
2329 * I don't really care if we go once more round the loop in
2330 * the error cases -eeb.
2332 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2333 list_for_each(tmp, &set->set_requests) {
2334 req = list_entry(tmp, struct ptlrpc_request,
2336 spin_lock(&req->rq_lock);
2337 req->rq_invalid_rqset = 1;
2338 spin_unlock(&req->rq_lock);
2341 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2343 LASSERT(atomic_read(&set->set_remaining) == 0);
2345 rc = set->set_rc; /* rq_status of already freed requests if any */
2346 list_for_each(tmp, &set->set_requests) {
2347 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2349 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2350 if (req->rq_status != 0)
2351 rc = req->rq_status;
2354 if (set->set_interpret) {
2355 int (*interpreter)(struct ptlrpc_request_set *set, void *, int) =
2357 rc = interpreter(set, set->set_arg, rc);
2359 struct ptlrpc_set_cbdata *cbdata, *n;
2362 list_for_each_entry_safe(cbdata, n,
2363 &set->set_cblist, psc_item) {
2364 list_del_init(&cbdata->psc_item);
2365 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2374 EXPORT_SYMBOL(ptlrpc_set_wait);
2377 * Helper function for request freeing.
2378 * Called when request count reached zero and request needs to be freed.
2379 * Removes request from all sorts of sending/replay lists it might be on,
2380 * frees network buffers if any are present.
2381 * If \a locked is set, that means caller is already holding import imp_lock
2382 * and so we no longer need to reobtain it (for certain lists manipulations)
2384 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2388 LASSERT(!request->rq_srv_req);
2389 LASSERT(!request->rq_export);
2390 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2391 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2392 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2393 LASSERTF(!request->rq_replay, "req %p\n", request);
2395 req_capsule_fini(&request->rq_pill);
2398 * We must take it off the imp_replay_list first. Otherwise, we'll set
2399 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2401 if (request->rq_import) {
2403 spin_lock(&request->rq_import->imp_lock);
2404 list_del_init(&request->rq_replay_list);
2405 list_del_init(&request->rq_unreplied_list);
2407 spin_unlock(&request->rq_import->imp_lock);
2409 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2411 if (atomic_read(&request->rq_refcount) != 0) {
2412 DEBUG_REQ(D_ERROR, request,
2413 "freeing request with nonzero refcount");
2417 if (request->rq_repbuf)
2418 sptlrpc_cli_free_repbuf(request);
2420 if (request->rq_import) {
2421 class_import_put(request->rq_import);
2422 request->rq_import = NULL;
2424 if (request->rq_bulk)
2425 ptlrpc_free_bulk(request->rq_bulk);
2427 if (request->rq_reqbuf || request->rq_clrbuf)
2428 sptlrpc_cli_free_reqbuf(request);
2430 if (request->rq_cli_ctx)
2431 sptlrpc_req_put_ctx(request, !locked);
2433 if (request->rq_pool)
2434 __ptlrpc_free_req_to_pool(request);
2436 ptlrpc_request_cache_free(request);
2441 * Drops one reference count for request \a request.
2442 * \a locked set indicates that caller holds import imp_lock.
2443 * Frees the request when reference count reaches zero.
2445 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2450 if (request == LP_POISON ||
2451 request->rq_reqmsg == LP_POISON) {
2452 CERROR("dereferencing freed request (bug 575)\n");
2457 DEBUG_REQ(D_INFO, request, "refcount now %u",
2458 atomic_read(&request->rq_refcount) - 1);
2460 if (atomic_dec_and_test(&request->rq_refcount)) {
2461 __ptlrpc_free_req(request, locked);
2469 * Drops one reference count for a request.
2471 void ptlrpc_req_finished(struct ptlrpc_request *request)
2473 __ptlrpc_req_finished(request, 0);
2475 EXPORT_SYMBOL(ptlrpc_req_finished);
2478 * Returns xid of a \a request
2480 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2482 return request->rq_xid;
2484 EXPORT_SYMBOL(ptlrpc_req_xid);
2487 * Disengage the client's reply buffer from the network
2488 * NB does _NOT_ unregister any client-side bulk.
2489 * IDEMPOTENT, but _not_ safe against concurrent callers.
2490 * The request owner (i.e. the thread doing the I/O) must call...
2491 * Returns 0 on success or 1 if unregistering cannot be made.
2493 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2496 wait_queue_head_t *wq;
2497 struct l_wait_info lwi;
2500 LASSERT(!in_interrupt());
2502 /* Let's setup deadline for reply unlink. */
2503 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2504 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2505 request->rq_reply_deadline =
2506 ktime_get_real_seconds() + LONG_UNLINK;
2508 /* Nothing left to do. */
2509 if (!ptlrpc_client_recv_or_unlink(request))
2512 LNetMDUnlink(request->rq_reply_md_h);
2514 /* Let's check it once again. */
2515 if (!ptlrpc_client_recv_or_unlink(request))
2518 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2519 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2521 /* Do not wait for unlink to finish. */
2526 * We have to l_wait_event() whatever the result, to give liblustre
2527 * a chance to run reply_in_callback(), and to make sure we've
2528 * unlinked before returning a req to the pool.
2530 if (request->rq_set)
2531 wq = &request->rq_set->set_waitq;
2533 wq = &request->rq_reply_waitq;
2537 * Network access will complete in finite time but the HUGE
2538 * timeout lets us CWARN for visibility of sluggish NALs
2540 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2541 cfs_time_seconds(1), NULL, NULL);
2542 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2545 ptlrpc_rqphase_move(request, request->rq_next_phase);
2549 LASSERT(rc == -ETIMEDOUT);
2550 DEBUG_REQ(D_WARNING, request,
2551 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2552 request->rq_receiving_reply,
2553 request->rq_req_unlinked,
2554 request->rq_reply_unlinked);
2559 static void ptlrpc_free_request(struct ptlrpc_request *req)
2561 spin_lock(&req->rq_lock);
2563 spin_unlock(&req->rq_lock);
2565 if (req->rq_commit_cb)
2566 req->rq_commit_cb(req);
2567 list_del_init(&req->rq_replay_list);
2569 __ptlrpc_req_finished(req, 1);
2573 * the request is committed and dropped from the replay list of its import
2575 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2577 struct obd_import *imp = req->rq_import;
2579 spin_lock(&imp->imp_lock);
2580 if (list_empty(&req->rq_replay_list)) {
2581 spin_unlock(&imp->imp_lock);
2585 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2586 ptlrpc_free_request(req);
2588 spin_unlock(&imp->imp_lock);
2590 EXPORT_SYMBOL(ptlrpc_request_committed);
2593 * Iterates through replay_list on import and prunes
2594 * all requests have transno smaller than last_committed for the
2595 * import and don't have rq_replay set.
2596 * Since requests are sorted in transno order, stops when meeting first
2597 * transno bigger than last_committed.
2598 * caller must hold imp->imp_lock
2600 void ptlrpc_free_committed(struct obd_import *imp)
2602 struct ptlrpc_request *req, *saved;
2603 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2604 bool skip_committed_list = true;
2606 assert_spin_locked(&imp->imp_lock);
2608 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2609 imp->imp_generation == imp->imp_last_generation_checked) {
2610 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2611 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2614 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2615 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2616 imp->imp_generation);
2618 if (imp->imp_generation != imp->imp_last_generation_checked ||
2619 !imp->imp_last_transno_checked)
2620 skip_committed_list = false;
2622 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2623 imp->imp_last_generation_checked = imp->imp_generation;
2625 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2627 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2628 LASSERT(req != last_req);
2631 if (req->rq_transno == 0) {
2632 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2635 if (req->rq_import_generation < imp->imp_generation) {
2636 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2640 /* not yet committed */
2641 if (req->rq_transno > imp->imp_peer_committed_transno) {
2642 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2646 if (req->rq_replay) {
2647 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2648 list_move_tail(&req->rq_replay_list,
2649 &imp->imp_committed_list);
2653 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2654 imp->imp_peer_committed_transno);
2656 ptlrpc_free_request(req);
2658 if (skip_committed_list)
2661 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2663 LASSERT(req->rq_transno != 0);
2664 if (req->rq_import_generation < imp->imp_generation ||
2666 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2667 req->rq_import_generation <
2668 imp->imp_generation ? "stale" : "closed");
2670 if (imp->imp_replay_cursor == &req->rq_replay_list)
2671 imp->imp_replay_cursor =
2672 req->rq_replay_list.next;
2674 ptlrpc_free_request(req);
2680 * Schedule previously sent request for resend.
2681 * For bulk requests we assign new xid (to avoid problems with
2682 * lost replies and therefore several transfers landing into same buffer
2683 * from different sending attempts).
2685 void ptlrpc_resend_req(struct ptlrpc_request *req)
2687 DEBUG_REQ(D_HA, req, "going to resend");
2688 spin_lock(&req->rq_lock);
2691 * Request got reply but linked to the import list still.
2692 * Let ptlrpc_check_set() to process it.
2694 if (ptlrpc_client_replied(req)) {
2695 spin_unlock(&req->rq_lock);
2696 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2700 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2701 req->rq_status = -EAGAIN;
2704 req->rq_net_err = 0;
2705 req->rq_timedout = 0;
2706 ptlrpc_client_wake_req(req);
2707 spin_unlock(&req->rq_lock);
2711 * Grab additional reference on a request \a req
2713 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2715 atomic_inc(&req->rq_refcount);
2718 EXPORT_SYMBOL(ptlrpc_request_addref);
2721 * Add a request to import replay_list.
2722 * Must be called under imp_lock
2724 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2725 struct obd_import *imp)
2727 struct list_head *tmp;
2729 assert_spin_locked(&imp->imp_lock);
2731 if (req->rq_transno == 0) {
2732 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2737 * clear this for new requests that were resent as well
2738 * as resent replayed requests.
2740 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2742 /* don't re-add requests that have been replayed */
2743 if (!list_empty(&req->rq_replay_list))
2746 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2748 spin_lock(&req->rq_lock);
2750 spin_unlock(&req->rq_lock);
2752 LASSERT(imp->imp_replayable);
2753 /* Balanced in ptlrpc_free_committed, usually. */
2754 ptlrpc_request_addref(req);
2755 list_for_each_prev(tmp, &imp->imp_replay_list) {
2756 struct ptlrpc_request *iter =
2757 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
2760 * We may have duplicate transnos if we create and then
2761 * open a file, or for closes retained if to match creating
2762 * opens, so use req->rq_xid as a secondary key.
2763 * (See bugs 684, 685, and 428.)
2764 * XXX no longer needed, but all opens need transnos!
2766 if (iter->rq_transno > req->rq_transno)
2769 if (iter->rq_transno == req->rq_transno) {
2770 LASSERT(iter->rq_xid != req->rq_xid);
2771 if (iter->rq_xid > req->rq_xid)
2775 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2779 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2783 * Send request and wait until it completes.
2784 * Returns request processing status.
2786 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2788 struct ptlrpc_request_set *set;
2791 LASSERT(!req->rq_set);
2792 LASSERT(!req->rq_receiving_reply);
2794 set = ptlrpc_prep_set();
2796 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2800 /* for distributed debugging */
2801 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2803 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2804 ptlrpc_request_addref(req);
2805 ptlrpc_set_add_req(set, req);
2806 rc = ptlrpc_set_wait(set);
2807 ptlrpc_set_destroy(set);
2811 EXPORT_SYMBOL(ptlrpc_queue_wait);
2814 * Callback used for replayed requests reply processing.
2815 * In case of successful reply calls registered request replay callback.
2816 * In case of error restart replay process.
2818 static int ptlrpc_replay_interpret(const struct lu_env *env,
2819 struct ptlrpc_request *req,
2822 struct ptlrpc_replay_async_args *aa = data;
2823 struct obd_import *imp = req->rq_import;
2825 atomic_dec(&imp->imp_replay_inflight);
2828 * Note: if it is bulk replay (MDS-MDS replay), then even if
2829 * server got the request, but bulk transfer timeout, let's
2830 * replay the bulk req again
2832 if (!ptlrpc_client_replied(req) ||
2834 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2835 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2840 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2841 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2842 lustre_msg_get_status(req->rq_repmsg) == -ENODEV)) {
2843 rc = lustre_msg_get_status(req->rq_repmsg);
2847 /** VBR: check version failure */
2848 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2849 /** replay was failed due to version mismatch */
2850 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2851 spin_lock(&imp->imp_lock);
2852 imp->imp_vbr_failed = 1;
2853 imp->imp_no_lock_replay = 1;
2854 spin_unlock(&imp->imp_lock);
2855 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2857 /** The transno had better not change over replay. */
2858 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2859 lustre_msg_get_transno(req->rq_repmsg) ||
2860 lustre_msg_get_transno(req->rq_repmsg) == 0,
2862 lustre_msg_get_transno(req->rq_reqmsg),
2863 lustre_msg_get_transno(req->rq_repmsg));
2866 spin_lock(&imp->imp_lock);
2867 /** if replays by version then gap occur on server, no trust to locks */
2868 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2869 imp->imp_no_lock_replay = 1;
2870 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2871 spin_unlock(&imp->imp_lock);
2872 LASSERT(imp->imp_last_replay_transno);
2874 /* transaction number shouldn't be bigger than the latest replayed */
2875 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2876 DEBUG_REQ(D_ERROR, req,
2877 "Reported transno %llu is bigger than the replayed one: %llu",
2879 lustre_msg_get_transno(req->rq_reqmsg));
2884 DEBUG_REQ(D_HA, req, "got rep");
2886 /* let the callback do fixups, possibly including in the request */
2887 if (req->rq_replay_cb)
2888 req->rq_replay_cb(req);
2890 if (ptlrpc_client_replied(req) &&
2891 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2892 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2893 lustre_msg_get_status(req->rq_repmsg),
2894 aa->praa_old_status);
2896 /* Put it back for re-replay. */
2897 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2901 * Errors while replay can set transno to 0, but
2902 * imp_last_replay_transno shouldn't be set to 0 anyway
2904 if (req->rq_transno == 0)
2905 CERROR("Transno is 0 during replay!\n");
2907 /* continue with recovery */
2908 rc = ptlrpc_import_recovery_state_machine(imp);
2910 req->rq_send_state = aa->praa_old_state;
2913 /* this replay failed, so restart recovery */
2914 ptlrpc_connect_import(imp);
2920 * Prepares and queues request for replay.
2921 * Adds it to ptlrpcd queue for actual sending.
2922 * Returns 0 on success.
2924 int ptlrpc_replay_req(struct ptlrpc_request *req)
2926 struct ptlrpc_replay_async_args *aa;
2928 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2930 LASSERT(sizeof(*aa) <= sizeof(req->rq_async_args));
2931 aa = ptlrpc_req_async_args(req);
2932 memset(aa, 0, sizeof(*aa));
2934 /* Prepare request to be resent with ptlrpcd */
2935 aa->praa_old_state = req->rq_send_state;
2936 req->rq_send_state = LUSTRE_IMP_REPLAY;
2937 req->rq_phase = RQ_PHASE_NEW;
2938 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2940 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2942 req->rq_interpret_reply = ptlrpc_replay_interpret;
2943 /* Readjust the timeout for current conditions */
2944 ptlrpc_at_set_req_timeout(req);
2947 * Tell server the net_latency, so the server can calculate how long
2948 * it should wait for next replay
2950 lustre_msg_set_service_time(req->rq_reqmsg,
2951 ptlrpc_at_get_net_latency(req));
2952 DEBUG_REQ(D_HA, req, "REPLAY");
2954 atomic_inc(&req->rq_import->imp_replay_inflight);
2955 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2957 ptlrpcd_add_req(req);
2962 * Aborts all in-flight request on import \a imp sending and delayed lists
2964 void ptlrpc_abort_inflight(struct obd_import *imp)
2966 struct list_head *tmp, *n;
2969 * Make sure that no new requests get processed for this import.
2970 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2971 * this flag and then putting requests on sending_list or delayed_list.
2973 spin_lock(&imp->imp_lock);
2976 * XXX locking? Maybe we should remove each request with the list
2977 * locked? Also, how do we know if the requests on the list are
2978 * being freed at this time?
2980 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2981 struct ptlrpc_request *req =
2982 list_entry(tmp, struct ptlrpc_request, rq_list);
2984 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2986 spin_lock(&req->rq_lock);
2987 if (req->rq_import_generation < imp->imp_generation) {
2989 req->rq_status = -EIO;
2990 ptlrpc_client_wake_req(req);
2992 spin_unlock(&req->rq_lock);
2995 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2996 struct ptlrpc_request *req =
2997 list_entry(tmp, struct ptlrpc_request, rq_list);
2999 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3001 spin_lock(&req->rq_lock);
3002 if (req->rq_import_generation < imp->imp_generation) {
3004 req->rq_status = -EIO;
3005 ptlrpc_client_wake_req(req);
3007 spin_unlock(&req->rq_lock);
3011 * Last chance to free reqs left on the replay list, but we
3012 * will still leak reqs that haven't committed.
3014 if (imp->imp_replayable)
3015 ptlrpc_free_committed(imp);
3017 spin_unlock(&imp->imp_lock);
3021 * Abort all uncompleted requests in request set \a set
3023 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3025 struct list_head *tmp, *pos;
3027 list_for_each_safe(pos, tmp, &set->set_requests) {
3028 struct ptlrpc_request *req =
3029 list_entry(pos, struct ptlrpc_request, rq_set_chain);
3031 spin_lock(&req->rq_lock);
3032 if (req->rq_phase != RQ_PHASE_RPC) {
3033 spin_unlock(&req->rq_lock);
3038 req->rq_status = -EINTR;
3039 ptlrpc_client_wake_req(req);
3040 spin_unlock(&req->rq_lock);
3044 static __u64 ptlrpc_last_xid;
3045 static spinlock_t ptlrpc_last_xid_lock;
3048 * Initialize the XID for the node. This is common among all requests on
3049 * this node, and only requires the property that it is monotonically
3050 * increasing. It does not need to be sequential. Since this is also used
3051 * as the RDMA match bits, it is important that a single client NOT have
3052 * the same match bits for two different in-flight requests, hence we do
3053 * NOT want to have an XID per target or similar.
3055 * To avoid an unlikely collision between match bits after a client reboot
3056 * (which would deliver old data into the wrong RDMA buffer) initialize
3057 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3058 * If the time is clearly incorrect, we instead use a 62-bit random number.
3059 * In the worst case the random number will overflow 1M RPCs per second in
3060 * 9133 years, or permutations thereof.
3062 #define YEAR_2004 (1ULL << 30)
3063 void ptlrpc_init_xid(void)
3065 time64_t now = ktime_get_real_seconds();
3067 spin_lock_init(&ptlrpc_last_xid_lock);
3068 if (now < YEAR_2004) {
3069 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3070 ptlrpc_last_xid >>= 2;
3071 ptlrpc_last_xid |= (1ULL << 61);
3073 ptlrpc_last_xid = (__u64)now << 20;
3076 /* Always need to be aligned to a power-of-two for multi-bulk BRW */
3077 BUILD_BUG_ON(((PTLRPC_BULK_OPS_COUNT - 1) & PTLRPC_BULK_OPS_COUNT) != 0);
3078 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3082 * Increase xid and returns resulting new value to the caller.
3084 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3085 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3086 * itself uses the last bulk xid needed, so the server can determine the
3087 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3088 * xid must align to a power-of-two value.
3090 * This is assumed to be true due to the initial ptlrpc_last_xid
3091 * value also being initialized to a power-of-two value. LU-1431
3093 __u64 ptlrpc_next_xid(void)
3097 spin_lock(&ptlrpc_last_xid_lock);
3098 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3099 ptlrpc_last_xid = next;
3100 spin_unlock(&ptlrpc_last_xid_lock);
3106 * If request has a new allocated XID (new request or EINPROGRESS resend),
3107 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3108 * request to ensure previous bulk fails and avoid problems with lost replies
3109 * and therefore several transfers landing into the same buffer from different
3112 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3114 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3119 * Generate new matchbits for all resend requests, including
3122 if (req->rq_resend) {
3123 u64 old_mbits = req->rq_mbits;
3126 * First time resend on -EINPROGRESS will generate new xid,
3127 * so we can actually use the rq_xid as rq_mbits in such case,
3128 * however, it's bit hard to distinguish such resend with a
3129 * 'resend for the -EINPROGRESS resend'. To make it simple,
3130 * we opt to generate mbits for all resend cases.
3132 if ((bd->bd_import->imp_connect_data.ocd_connect_flags &
3133 OBD_CONNECT_BULK_MBITS)) {
3134 req->rq_mbits = ptlrpc_next_xid();
3136 /* old version transfers rq_xid to peer as matchbits */
3137 spin_lock(&req->rq_import->imp_lock);
3138 list_del_init(&req->rq_unreplied_list);
3139 ptlrpc_assign_next_xid_nolock(req);
3140 spin_unlock(&req->rq_import->imp_lock);
3141 req->rq_mbits = req->rq_xid;
3144 CDEBUG(D_HA, "resend bulk old x%llu new x%llu\n",
3145 old_mbits, req->rq_mbits);
3146 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3147 /* Request being sent first time, use xid as matchbits. */
3148 req->rq_mbits = req->rq_xid;
3151 * Replay request, xid and matchbits have already been
3152 * correctly assigned.
3158 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3159 * that server can infer the number of bulks that were prepared,
3162 req->rq_mbits += DIV_ROUND_UP(bd->bd_iov_count, LNET_MAX_IOV) - 1;
3166 * Get a glimpse at what next xid value might have been.
3167 * Returns possible next xid.
3169 __u64 ptlrpc_sample_next_xid(void)
3171 #if BITS_PER_LONG == 32
3172 /* need to avoid possible word tearing on 32-bit systems */
3175 spin_lock(&ptlrpc_last_xid_lock);
3176 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3177 spin_unlock(&ptlrpc_last_xid_lock);
3181 /* No need to lock, since returned value is racy anyways */
3182 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3185 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3188 * Functions for operating ptlrpc workers.
3190 * A ptlrpc work is a function which will be running inside ptlrpc context.
3191 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3193 * 1. after a work is created, it can be used many times, that is:
3194 * handler = ptlrpcd_alloc_work();
3195 * ptlrpcd_queue_work();
3197 * queue it again when necessary:
3198 * ptlrpcd_queue_work();
3199 * ptlrpcd_destroy_work();
3200 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3201 * it will only be queued once in any time. Also as its name implies, it may
3202 * have delay before it really runs by ptlrpcd thread.
3204 struct ptlrpc_work_async_args {
3205 int (*cb)(const struct lu_env *, void *);
3209 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3211 /* re-initialize the req */
3212 req->rq_timeout = obd_timeout;
3213 req->rq_sent = ktime_get_real_seconds();
3214 req->rq_deadline = req->rq_sent + req->rq_timeout;
3215 req->rq_phase = RQ_PHASE_INTERPRET;
3216 req->rq_next_phase = RQ_PHASE_COMPLETE;
3217 req->rq_xid = ptlrpc_next_xid();
3218 req->rq_import_generation = req->rq_import->imp_generation;
3220 ptlrpcd_add_req(req);
3223 static int work_interpreter(const struct lu_env *env,
3224 struct ptlrpc_request *req, void *data, int rc)
3226 struct ptlrpc_work_async_args *arg = data;
3228 LASSERT(ptlrpcd_check_work(req));
3230 rc = arg->cb(env, arg->cbdata);
3232 list_del_init(&req->rq_set_chain);
3235 if (atomic_dec_return(&req->rq_refcount) > 1) {
3236 atomic_set(&req->rq_refcount, 2);
3237 ptlrpcd_add_work_req(req);
3242 static int worker_format;
3244 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3246 return req->rq_pill.rc_fmt == (void *)&worker_format;
3250 * Create a work for ptlrpc.
3252 void *ptlrpcd_alloc_work(struct obd_import *imp,
3253 int (*cb)(const struct lu_env *, void *), void *cbdata)
3255 struct ptlrpc_request *req = NULL;
3256 struct ptlrpc_work_async_args *args;
3261 return ERR_PTR(-EINVAL);
3263 /* copy some code from deprecated fakereq. */
3264 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3266 CERROR("ptlrpc: run out of memory!\n");
3267 return ERR_PTR(-ENOMEM);
3270 ptlrpc_cli_req_init(req);
3272 req->rq_send_state = LUSTRE_IMP_FULL;
3273 req->rq_type = PTL_RPC_MSG_REQUEST;
3274 req->rq_import = class_import_get(imp);
3275 req->rq_interpret_reply = work_interpreter;
3276 /* don't want reply */
3277 req->rq_no_delay = 1;
3278 req->rq_no_resend = 1;
3279 req->rq_pill.rc_fmt = (void *)&worker_format;
3281 BUILD_BUG_ON(sizeof(*args) > sizeof(req->rq_async_args));
3282 args = ptlrpc_req_async_args(req);
3284 args->cbdata = cbdata;
3288 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3290 void ptlrpcd_destroy_work(void *handler)
3292 struct ptlrpc_request *req = handler;
3295 ptlrpc_req_finished(req);
3297 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3299 int ptlrpcd_queue_work(void *handler)
3301 struct ptlrpc_request *req = handler;
3304 * Check if the req is already being queued.
3306 * Here comes a trick: it lacks a way of checking if a req is being
3307 * processed reliably in ptlrpc. Here I have to use refcount of req
3308 * for this purpose. This is okay because the caller should use this
3309 * req as opaque data. - Jinshan
3311 LASSERT(atomic_read(&req->rq_refcount) > 0);
3312 if (atomic_inc_return(&req->rq_refcount) == 2)
3313 ptlrpcd_add_work_req(req);
3316 EXPORT_SYMBOL(ptlrpcd_queue_work);