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/
32 /** Implementation of client-side PortalRPC interfaces */
34 #define DEBUG_SUBSYSTEM S_RPC
36 #include <linux/delay.h>
37 #include <linux/random.h>
39 #include <lnet/lib-lnet.h>
40 #include <obd_support.h>
41 #include <obd_class.h>
42 #include <lustre_lib.h>
43 #include <lustre_ha.h>
44 #include <lustre_import.h>
45 #include <lustre_req_layout.h>
47 #include "ptlrpc_internal.h"
49 static void ptlrpc_prep_bulk_page_pin(struct ptlrpc_bulk_desc *desc,
50 struct page *page, int pageoffset,
53 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 1);
56 static void ptlrpc_prep_bulk_page_nopin(struct ptlrpc_bulk_desc *desc,
57 struct page *page, int pageoffset,
60 __ptlrpc_prep_bulk_page(desc, page, pageoffset, len, 0);
63 static void ptlrpc_release_bulk_page_pin(struct ptlrpc_bulk_desc *desc)
67 for (i = 0; i < desc->bd_iov_count ; i++)
68 put_page(desc->bd_vec[i].bv_page);
71 static int ptlrpc_prep_bulk_frag_pages(struct ptlrpc_bulk_desc *desc,
74 unsigned int offset = (unsigned long)frag & ~PAGE_MASK;
78 int page_len = min_t(unsigned int, PAGE_SIZE - offset,
80 unsigned long vaddr = (unsigned long)frag;
82 ptlrpc_prep_bulk_page_nopin(desc,
83 lnet_kvaddr_to_page(vaddr),
93 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
94 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
95 .release_frags = ptlrpc_release_bulk_page_pin,
97 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
99 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
100 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
101 .release_frags = ptlrpc_release_bulk_noop,
102 .add_iov_frag = ptlrpc_prep_bulk_frag_pages,
104 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
106 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
107 static int ptlrpcd_check_work(struct ptlrpc_request *req);
108 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
111 * Initialize passed in client structure \a cl.
113 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
114 struct ptlrpc_client *cl)
116 cl->cli_request_portal = req_portal;
117 cl->cli_reply_portal = rep_portal;
120 EXPORT_SYMBOL(ptlrpc_init_client);
123 * Return PortalRPC connection for remore uud \a uuid
125 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
128 struct ptlrpc_connection *c;
129 struct lnet_nid self;
130 struct lnet_processid peer;
134 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
135 * before accessing its values.
137 /* coverity[uninit_use_in_call] */
138 err = ptlrpc_uuid_to_peer(uuid, &peer, &self, refnet);
140 CNETERR("cannot find peer %s!\n", uuid->uuid);
144 c = ptlrpc_connection_get(&peer, &self, uuid);
146 memcpy(c->c_remote_uuid.uuid,
147 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
150 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
156 * Allocate and initialize new bulk descriptor on the sender.
157 * Returns pointer to the descriptor or NULL on error.
159 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
160 unsigned int max_brw,
161 enum ptlrpc_bulk_op_type type,
163 const struct ptlrpc_bulk_frag_ops *ops)
165 struct ptlrpc_bulk_desc *desc;
168 LASSERT(ops->add_kiov_frag != NULL);
170 if (max_brw > PTLRPC_BULK_OPS_COUNT)
173 if (nfrags > LNET_MAX_IOV * max_brw)
180 OBD_ALLOC_LARGE(desc->bd_vec,
181 nfrags * sizeof(*desc->bd_vec));
185 spin_lock_init(&desc->bd_lock);
186 init_waitqueue_head(&desc->bd_waitq);
187 desc->bd_max_iov = nfrags;
188 desc->bd_iov_count = 0;
189 desc->bd_portal = portal;
190 desc->bd_type = type;
191 desc->bd_md_count = 0;
192 desc->bd_nob_last = LNET_MTU;
193 desc->bd_frag_ops = ops;
194 LASSERT(max_brw > 0);
195 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
197 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
198 * node. Negotiated ocd_brw_size will always be <= this number.
200 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
201 LNetInvalidateMDHandle(&desc->bd_mds[i]);
210 * Prepare bulk descriptor for specified outgoing request \a req that
211 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
212 * the bulk to be sent. Used on client-side.
213 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
216 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
218 unsigned int max_brw,
221 const struct ptlrpc_bulk_frag_ops
224 struct obd_import *imp = req->rq_import;
225 struct ptlrpc_bulk_desc *desc;
228 LASSERT(ptlrpc_is_bulk_op_passive(type));
230 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
234 desc->bd_import = class_import_get(imp);
237 desc->bd_cbid.cbid_fn = client_bulk_callback;
238 desc->bd_cbid.cbid_arg = desc;
240 /* This makes req own desc, and free it when she frees herself */
245 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
247 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
248 struct page *page, int pageoffset, int len,
251 struct bio_vec *kiov;
253 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
254 LASSERT(page != NULL);
255 LASSERT(pageoffset >= 0);
257 LASSERT(pageoffset + len <= PAGE_SIZE);
259 kiov = &desc->bd_vec[desc->bd_iov_count];
261 if (((desc->bd_iov_count % LNET_MAX_IOV) == 0) ||
262 ((desc->bd_nob_last + len) > LNET_MTU)) {
263 desc->bd_mds_off[desc->bd_md_count] = desc->bd_iov_count;
265 desc->bd_nob_last = 0;
266 LASSERT(desc->bd_md_count <= PTLRPC_BULK_OPS_COUNT);
269 desc->bd_nob_last += len;
275 kiov->bv_page = page;
276 kiov->bv_offset = pageoffset;
279 desc->bd_iov_count++;
281 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
283 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
287 LASSERT(desc != NULL);
288 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
289 LASSERT(desc->bd_refs == 0); /* network hands off */
290 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
291 LASSERT(desc->bd_frag_ops != NULL);
293 sptlrpc_enc_pool_put_pages(desc);
296 class_export_put(desc->bd_export);
298 class_import_put(desc->bd_import);
300 if (desc->bd_frag_ops->release_frags != NULL)
301 desc->bd_frag_ops->release_frags(desc);
303 OBD_FREE_LARGE(desc->bd_vec,
304 desc->bd_max_iov * sizeof(*desc->bd_vec));
308 EXPORT_SYMBOL(ptlrpc_free_bulk);
311 * Set server timelimit for this req, i.e. how long are we willing to wait
312 * for reply before timing out this request.
314 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
316 LASSERT(req->rq_import);
319 /* non-AT settings */
321 * \a imp_server_timeout means this is reverse import and
322 * we send (currently only) ASTs to the client and cannot afford
323 * to wait too long for the reply, otherwise the other client
324 * (because of which we are sending this request) would
325 * timeout waiting for us
327 req->rq_timeout = req->rq_import->imp_server_timeout ?
328 obd_timeout / 2 : obd_timeout;
330 struct imp_at *at = &req->rq_import->imp_at;
334 idx = import_at_get_index(req->rq_import,
335 req->rq_request_portal);
336 serv_est = at_get(&at->iat_service_estimate[idx]);
338 * Currently a 32 bit value is sent over the
339 * wire for rq_timeout so please don't change this
340 * to time64_t. The work for LU-1158 will in time
341 * replace rq_timeout with a 64 bit nanosecond value
343 req->rq_timeout = at_est2timeout(serv_est);
346 * We could get even fancier here, using history to predict increased
349 * Let the server know what this RPC timeout is by putting it in the
352 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
354 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
356 /* Adjust max service estimate based on server value */
357 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
364 LASSERT(req->rq_import);
365 at = &req->rq_import->imp_at;
367 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
369 * max service estimates are tracked on the server side,
370 * so just keep minimal history here
372 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
375 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
376 req->rq_import->imp_obd->obd_name,
377 req->rq_request_portal,
378 oldse, at_get(&at->iat_service_estimate[idx]));
381 /* Expected network latency per remote node (secs) */
382 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
384 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
387 /* Adjust expected network latency */
388 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
389 timeout_t service_timeout)
391 time64_t now = ktime_get_real_seconds();
396 LASSERT(req->rq_import);
398 if (service_timeout > now - req->rq_sent + 3) {
400 * b=16408, however, this can also happen if early reply
401 * is lost and client RPC is expired and resent, early reply
402 * or reply of original RPC can still be fit in reply buffer
403 * of resent RPC, now client is measuring time from the
404 * resent time, but server sent back service time of original
407 CDEBUG_LIMIT((lustre_msg_get_flags(req->rq_reqmsg) &
408 MSG_RESENT) ? D_ADAPTTO : D_WARNING,
409 "Reported service time %u > total measured time %lld\n",
410 service_timeout, now - req->rq_sent);
414 /* Network latency is total time less server processing time,
417 nl = max_t(timeout_t, now - req->rq_sent - service_timeout, 0) + 1;
418 at = &req->rq_import->imp_at;
420 oldnl = at_measured(&at->iat_net_latency, nl);
423 "The network latency for %s (nid %s) has changed from %d to %d\n",
424 req->rq_import->imp_obd->obd_name,
425 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
426 oldnl, at_get(&at->iat_net_latency));
429 static int unpack_reply(struct ptlrpc_request *req)
433 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
434 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
436 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
442 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
444 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
452 * Handle an early reply message, called with the rq_lock held.
453 * If anything goes wrong just ignore it - same as if it never happened
455 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
456 __must_hold(&req->rq_lock)
458 struct ptlrpc_request *early_req;
459 timeout_t service_timeout;
465 spin_unlock(&req->rq_lock);
467 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
469 spin_lock(&req->rq_lock);
473 rc = unpack_reply(early_req);
475 sptlrpc_cli_finish_early_reply(early_req);
476 spin_lock(&req->rq_lock);
481 * Use new timeout value just to adjust the local value for this
482 * request, don't include it into at_history. It is unclear yet why
483 * service time increased and should it be counted or skipped, e.g.
484 * that can be recovery case or some error or server, the real reply
485 * will add all new data if it is worth to add.
487 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
488 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
490 /* Network latency can be adjusted, it is pure network delays */
491 service_timeout = lustre_msg_get_service_timeout(early_req->rq_repmsg);
492 ptlrpc_at_adj_net_latency(req, service_timeout);
494 sptlrpc_cli_finish_early_reply(early_req);
496 spin_lock(&req->rq_lock);
497 olddl = req->rq_deadline;
499 * server assumes it now has rq_timeout from when the request
500 * arrived, so the client should give it at least that long.
501 * since we don't know the arrival time we'll use the original
504 req->rq_deadline = req->rq_sent + req->rq_timeout +
505 ptlrpc_at_get_net_latency(req);
507 /* The below message is checked in replay-single.sh test_65{a,b} */
508 /* The below message is checked in sanity-{gss,krb5} test_8 */
509 DEBUG_REQ(D_ADAPTTO, req,
510 "Early reply #%d, new deadline in %llds (%llds)",
512 req->rq_deadline - ktime_get_real_seconds(),
513 req->rq_deadline - olddl);
518 static struct kmem_cache *request_cache;
520 int ptlrpc_request_cache_init(void)
522 request_cache = kmem_cache_create("ptlrpc_cache",
523 sizeof(struct ptlrpc_request),
524 0, SLAB_HWCACHE_ALIGN, NULL);
525 return request_cache ? 0 : -ENOMEM;
528 void ptlrpc_request_cache_fini(void)
530 kmem_cache_destroy(request_cache);
533 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
535 struct ptlrpc_request *req;
537 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
541 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
543 OBD_SLAB_FREE_PTR(req, request_cache);
547 * Wind down request pool \a pool.
548 * Frees all requests from the pool too
550 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
552 struct ptlrpc_request *req;
554 LASSERT(pool != NULL);
556 spin_lock(&pool->prp_lock);
557 while ((req = list_first_entry_or_null(&pool->prp_req_list,
558 struct ptlrpc_request,
560 list_del(&req->rq_list);
561 LASSERT(req->rq_reqbuf);
562 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
563 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
564 ptlrpc_request_cache_free(req);
566 spin_unlock(&pool->prp_lock);
567 OBD_FREE(pool, sizeof(*pool));
569 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
572 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
574 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
579 while (size < pool->prp_rq_size)
582 LASSERTF(list_empty(&pool->prp_req_list) ||
583 size == pool->prp_rq_size,
584 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
585 pool->prp_rq_size, size);
587 pool->prp_rq_size = size;
588 for (i = 0; i < num_rq; i++) {
589 struct ptlrpc_request *req;
590 struct lustre_msg *msg;
592 req = ptlrpc_request_cache_alloc(GFP_NOFS);
595 OBD_ALLOC_LARGE(msg, size);
597 ptlrpc_request_cache_free(req);
600 req->rq_reqbuf = msg;
601 req->rq_reqbuf_len = size;
603 spin_lock(&pool->prp_lock);
604 list_add_tail(&req->rq_list, &pool->prp_req_list);
605 spin_unlock(&pool->prp_lock);
609 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
612 * Create and initialize new request pool with given attributes:
613 * \a num_rq - initial number of requests to create for the pool
614 * \a msgsize - maximum message size possible for requests in thid pool
615 * \a populate_pool - function to be called when more requests need to be added
617 * Returns pointer to newly created pool or NULL on error.
619 struct ptlrpc_request_pool *
620 ptlrpc_init_rq_pool(int num_rq, int msgsize,
621 int (*populate_pool)(struct ptlrpc_request_pool *, int))
623 struct ptlrpc_request_pool *pool;
630 * Request next power of two for the allocation, because internally
631 * kernel would do exactly this
633 spin_lock_init(&pool->prp_lock);
634 INIT_LIST_HEAD(&pool->prp_req_list);
635 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
636 pool->prp_populate = populate_pool;
638 populate_pool(pool, num_rq);
642 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
645 * Fetches one request from pool \a pool
647 static struct ptlrpc_request *
648 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
650 struct ptlrpc_request *request;
651 struct lustre_msg *reqbuf;
656 spin_lock(&pool->prp_lock);
659 * See if we have anything in a pool, and bail out if nothing,
660 * in writeout path, where this matters, this is safe to do, because
661 * nothing is lost in this case, and when some in-flight requests
662 * complete, this code will be called again.
664 if (unlikely(list_empty(&pool->prp_req_list))) {
665 spin_unlock(&pool->prp_lock);
669 request = list_first_entry(&pool->prp_req_list, struct ptlrpc_request,
671 list_del_init(&request->rq_list);
672 spin_unlock(&pool->prp_lock);
674 LASSERT(request->rq_reqbuf);
675 LASSERT(request->rq_pool);
677 reqbuf = request->rq_reqbuf;
678 memset(request, 0, sizeof(*request));
679 request->rq_reqbuf = reqbuf;
680 request->rq_reqbuf_len = pool->prp_rq_size;
681 request->rq_pool = pool;
687 * Returns freed \a request to pool.
689 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
691 struct ptlrpc_request_pool *pool = request->rq_pool;
693 spin_lock(&pool->prp_lock);
694 LASSERT(list_empty(&request->rq_list));
695 LASSERT(!request->rq_receiving_reply);
696 list_add_tail(&request->rq_list, &pool->prp_req_list);
697 spin_unlock(&pool->prp_lock);
700 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
702 struct obd_import *imp = req->rq_import;
703 struct ptlrpc_request *iter;
705 assert_spin_locked(&imp->imp_lock);
706 LASSERT(list_empty(&req->rq_unreplied_list));
708 /* unreplied list is sorted by xid in ascending order */
709 list_for_each_entry_reverse(iter, &imp->imp_unreplied_list,
711 LASSERT(req->rq_xid != iter->rq_xid);
712 if (req->rq_xid < iter->rq_xid)
714 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
717 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
720 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
722 req->rq_xid = ptlrpc_next_xid();
723 ptlrpc_add_unreplied(req);
726 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
728 spin_lock(&req->rq_import->imp_lock);
729 ptlrpc_assign_next_xid_nolock(req);
730 spin_unlock(&req->rq_import->imp_lock);
733 static atomic64_t ptlrpc_last_xid;
735 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
737 spin_lock(&req->rq_import->imp_lock);
738 list_del_init(&req->rq_unreplied_list);
739 ptlrpc_assign_next_xid_nolock(req);
740 spin_unlock(&req->rq_import->imp_lock);
741 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
744 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
746 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
750 opc = lustre_msg_get_opc(req->rq_reqmsg);
751 tag = obd_get_mod_rpc_slot(cli, opc);
752 lustre_msg_set_tag(req->rq_reqmsg, tag);
753 ptlrpc_reassign_next_xid(req);
755 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
757 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
759 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
762 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
763 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
765 obd_put_mod_rpc_slot(cli, opc, tag);
768 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
770 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
771 __u32 version, int opcode, char **bufs,
772 struct ptlrpc_cli_ctx *ctx)
775 struct obd_import *imp;
781 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
782 imp = request->rq_import;
783 lengths = request->rq_pill.rc_area[RCL_CLIENT];
786 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
788 rc = sptlrpc_req_get_ctx(request);
792 sptlrpc_req_set_flavor(request, opcode);
794 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
799 lustre_msg_add_version(request->rq_reqmsg, version);
800 request->rq_send_state = LUSTRE_IMP_FULL;
801 request->rq_type = PTL_RPC_MSG_REQUEST;
803 request->rq_req_cbid.cbid_fn = request_out_callback;
804 request->rq_req_cbid.cbid_arg = request;
806 request->rq_reply_cbid.cbid_fn = reply_in_callback;
807 request->rq_reply_cbid.cbid_arg = request;
809 request->rq_reply_deadline = 0;
810 request->rq_bulk_deadline = 0;
811 request->rq_req_deadline = 0;
812 request->rq_phase = RQ_PHASE_NEW;
813 request->rq_next_phase = RQ_PHASE_UNDEFINED;
815 request->rq_request_portal = imp->imp_client->cli_request_portal;
816 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
818 ptlrpc_at_set_req_timeout(request);
820 lustre_msg_set_opc(request->rq_reqmsg, opcode);
822 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
823 if (cfs_fail_val == opcode) {
824 time64_t *fail_t = NULL, *fail2_t = NULL;
826 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
827 fail_t = &request->rq_bulk_deadline;
828 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
829 fail_t = &request->rq_reply_deadline;
830 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
831 fail_t = &request->rq_req_deadline;
832 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
833 fail_t = &request->rq_reply_deadline;
834 fail2_t = &request->rq_bulk_deadline;
835 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
836 time64_t now = ktime_get_real_seconds();
837 u64 xid = ((u64)now >> 4) << 24;
839 atomic64_set(&ptlrpc_last_xid, xid);
843 *fail_t = ktime_get_real_seconds() +
844 PTLRPC_REQ_LONG_UNLINK;
847 *fail2_t = ktime_get_real_seconds() +
848 PTLRPC_REQ_LONG_UNLINK;
851 * The RPC is infected, let the test to change the
854 msleep(4 * MSEC_PER_SEC);
857 ptlrpc_assign_next_xid(request);
862 LASSERT(!request->rq_pool);
863 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
865 atomic_dec(&imp->imp_reqs);
866 class_import_put(imp);
870 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
873 * Pack request buffers for network transfer, performing necessary encryption
874 * steps if necessary.
876 int ptlrpc_request_pack(struct ptlrpc_request *request,
877 __u32 version, int opcode)
879 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
881 EXPORT_SYMBOL(ptlrpc_request_pack);
884 * Helper function to allocate new request on import \a imp
885 * and possibly using existing request from pool \a pool if provided.
886 * Returns allocated request structure with import field filled or
890 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
891 struct ptlrpc_request_pool *pool)
893 struct ptlrpc_request *request = NULL;
895 request = ptlrpc_request_cache_alloc(GFP_NOFS);
897 if (!request && pool)
898 request = ptlrpc_prep_req_from_pool(pool);
901 ptlrpc_cli_req_init(request);
903 LASSERTF((unsigned long)imp > 0x1000, "%p\n", imp);
904 LASSERT(imp != LP_POISON);
905 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
907 LASSERT(imp->imp_client != LP_POISON);
909 request->rq_import = class_import_get(imp);
910 atomic_inc(&imp->imp_reqs);
912 CERROR("request allocation out of memory\n");
918 static int ptlrpc_reconnect_if_idle(struct obd_import *imp)
923 * initiate connection if needed when the import has been
924 * referenced by the new request to avoid races with disconnect.
925 * serialize this check against conditional state=IDLE
926 * in ptlrpc_disconnect_idle_interpret()
928 spin_lock(&imp->imp_lock);
929 if (imp->imp_state == LUSTRE_IMP_IDLE) {
930 imp->imp_generation++;
931 imp->imp_initiated_at = imp->imp_generation;
932 imp->imp_state = LUSTRE_IMP_NEW;
934 /* connect_import_locked releases imp_lock */
935 rc = ptlrpc_connect_import_locked(imp);
938 ptlrpc_pinger_add_import(imp);
940 spin_unlock(&imp->imp_lock);
946 * Helper function for creating a request.
947 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
948 * buffer structures according to capsule template \a format.
949 * Returns allocated request structure pointer or NULL on error.
951 static struct ptlrpc_request *
952 ptlrpc_request_alloc_internal(struct obd_import *imp,
953 struct ptlrpc_request_pool *pool,
954 const struct req_format *format)
956 struct ptlrpc_request *request;
958 request = __ptlrpc_request_alloc(imp, pool);
962 /* don't make expensive check for idling connection
963 * if it's already connected */
964 if (unlikely(imp->imp_state != LUSTRE_IMP_FULL)) {
965 if (ptlrpc_reconnect_if_idle(imp) < 0) {
966 atomic_dec(&imp->imp_reqs);
967 ptlrpc_request_free(request);
972 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
973 req_capsule_set(&request->rq_pill, format);
978 * Allocate new request structure for import \a imp and initialize its
979 * buffer structure according to capsule template \a format.
981 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
982 const struct req_format *format)
984 return ptlrpc_request_alloc_internal(imp, NULL, format);
986 EXPORT_SYMBOL(ptlrpc_request_alloc);
989 * Allocate new request structure for import \a imp from pool \a pool and
990 * initialize its buffer structure according to capsule template \a format.
992 struct ptlrpc_request *
993 ptlrpc_request_alloc_pool(struct obd_import *imp,
994 struct ptlrpc_request_pool *pool,
995 const struct req_format *format)
997 return ptlrpc_request_alloc_internal(imp, pool, format);
999 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
1002 * For requests not from pool, free memory of the request structure.
1003 * For requests obtained from a pool earlier, return request back to pool.
1005 void ptlrpc_request_free(struct ptlrpc_request *request)
1007 if (request->rq_pool)
1008 __ptlrpc_free_req_to_pool(request);
1010 ptlrpc_request_cache_free(request);
1012 EXPORT_SYMBOL(ptlrpc_request_free);
1015 * Allocate new request for operatione \a opcode and immediatelly pack it for
1017 * Only used for simple requests like OBD_PING where the only important
1018 * part of the request is operation itself.
1019 * Returns allocated request or NULL on error.
1021 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1022 const struct req_format *format,
1023 __u32 version, int opcode)
1025 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1029 rc = ptlrpc_request_pack(req, version, opcode);
1031 ptlrpc_request_free(req);
1037 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1040 * Allocate and initialize new request set structure on the current CPT.
1041 * Returns a pointer to the newly allocated set structure or NULL on error.
1043 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1045 struct ptlrpc_request_set *set;
1049 cpt = cfs_cpt_current(cfs_cpt_tab, 0);
1050 OBD_CPT_ALLOC(set, cfs_cpt_tab, cpt, sizeof(*set));
1053 atomic_set(&set->set_refcount, 1);
1054 INIT_LIST_HEAD(&set->set_requests);
1055 init_waitqueue_head(&set->set_waitq);
1056 atomic_set(&set->set_new_count, 0);
1057 atomic_set(&set->set_remaining, 0);
1058 spin_lock_init(&set->set_new_req_lock);
1059 INIT_LIST_HEAD(&set->set_new_requests);
1060 set->set_max_inflight = UINT_MAX;
1061 set->set_producer = NULL;
1062 set->set_producer_arg = NULL;
1067 EXPORT_SYMBOL(ptlrpc_prep_set);
1070 * Allocate and initialize new request set structure with flow control
1071 * extension. This extension allows to control the number of requests in-flight
1072 * for the whole set. A callback function to generate requests must be provided
1073 * and the request set will keep the number of requests sent over the wire to
1075 * Returns a pointer to the newly allocated set structure or NULL on error.
1077 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1081 struct ptlrpc_request_set *set;
1083 set = ptlrpc_prep_set();
1087 set->set_max_inflight = max;
1088 set->set_producer = func;
1089 set->set_producer_arg = arg;
1095 * Wind down and free request set structure previously allocated with
1097 * Ensures that all requests on the set have completed and removes
1098 * all requests from the request list in a set.
1099 * If any unsent request happen to be on the list, pretends that they got
1100 * an error in flight and calls their completion handler.
1102 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1104 struct ptlrpc_request *req;
1110 /* Requests on the set should either all be completed, or all be new */
1111 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1112 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1113 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
1114 LASSERT(req->rq_phase == expected_phase);
1118 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1119 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1120 atomic_read(&set->set_remaining), n);
1122 while ((req = list_first_entry_or_null(&set->set_requests,
1123 struct ptlrpc_request,
1125 list_del_init(&req->rq_set_chain);
1127 LASSERT(req->rq_phase == expected_phase);
1129 if (req->rq_phase == RQ_PHASE_NEW) {
1130 ptlrpc_req_interpret(NULL, req, -EBADR);
1131 atomic_dec(&set->set_remaining);
1134 spin_lock(&req->rq_lock);
1136 req->rq_invalid_rqset = 0;
1137 spin_unlock(&req->rq_lock);
1139 ptlrpc_req_finished(req);
1142 LASSERT(atomic_read(&set->set_remaining) == 0);
1144 ptlrpc_reqset_put(set);
1147 EXPORT_SYMBOL(ptlrpc_set_destroy);
1150 * Add a new request to the general purpose request set.
1151 * Assumes request reference from the caller.
1153 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1154 struct ptlrpc_request *req)
1156 if (set == PTLRPCD_SET) {
1157 ptlrpcd_add_req(req);
1161 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1162 LASSERT(list_empty(&req->rq_set_chain));
1164 if (req->rq_allow_intr)
1165 set->set_allow_intr = 1;
1167 /* The set takes over the caller's request reference */
1168 list_add_tail(&req->rq_set_chain, &set->set_requests);
1170 atomic_inc(&set->set_remaining);
1171 req->rq_queued_time = ktime_get_seconds();
1174 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1176 if (set->set_producer)
1178 * If the request set has a producer callback, the RPC must be
1179 * sent straight away
1181 ptlrpc_send_new_req(req);
1183 EXPORT_SYMBOL(ptlrpc_set_add_req);
1186 * Add a request to a request with dedicated server thread
1187 * and wake the thread to make any necessary processing.
1188 * Currently only used for ptlrpcd.
1190 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1191 struct ptlrpc_request *req)
1193 struct ptlrpc_request_set *set = pc->pc_set;
1196 LASSERT(req->rq_set == NULL);
1197 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1199 spin_lock(&set->set_new_req_lock);
1201 * The set takes over the caller's request reference.
1204 req->rq_queued_time = ktime_get_seconds();
1205 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1206 count = atomic_inc_return(&set->set_new_count);
1207 spin_unlock(&set->set_new_req_lock);
1209 /* Only need to call wakeup once for the first entry. */
1211 wake_up(&set->set_waitq);
1214 * XXX: It maybe unnecessary to wakeup all the partners. But to
1215 * guarantee the async RPC can be processed ASAP, we have
1216 * no other better choice. It maybe fixed in future.
1218 for (i = 0; i < pc->pc_npartners; i++)
1219 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1224 * Based on the current state of the import, determine if the request
1225 * can be sent, is an error, or should be delayed.
1227 * Returns true if this request should be delayed. If false, and
1228 * *status is set, then the request can not be sent and *status is the
1229 * error code. If false and status is 0, then request can be sent.
1231 * The imp->imp_lock must be held.
1233 static int ptlrpc_import_delay_req(struct obd_import *imp,
1234 struct ptlrpc_request *req, int *status)
1242 if (req->rq_ctx_init || req->rq_ctx_fini) {
1243 /* always allow ctx init/fini rpc go through */
1244 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1245 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1247 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1248 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1251 * pings or MDS-equivalent STATFS may safely
1254 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1255 D_HA : D_ERROR, req, "IMP_CLOSED");
1257 } else if (ptlrpc_send_limit_expired(req)) {
1258 /* probably doesn't need to be a D_ERROR afterinitial testing */
1259 DEBUG_REQ(D_HA, req, "send limit expired");
1260 *status = -ETIMEDOUT;
1261 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1262 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1263 ;/* allow CONNECT even if import is invalid */
1264 if (atomic_read(&imp->imp_inval_count) != 0) {
1265 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1268 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1269 if (!imp->imp_deactive)
1270 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1271 *status = -ESHUTDOWN; /* b=12940 */
1272 } else if (req->rq_import_generation != imp->imp_generation) {
1273 DEBUG_REQ(req->rq_no_resend ? D_INFO : D_ERROR,
1274 req, "req wrong generation:");
1276 } else if (req->rq_send_state != imp->imp_state) {
1277 /* invalidate in progress - any requests should be drop */
1278 if (atomic_read(&imp->imp_inval_count) != 0) {
1279 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1281 } else if (req->rq_no_delay &&
1282 imp->imp_generation != imp->imp_initiated_at) {
1283 /* ignore nodelay for requests initiating connections */
1285 } else if (req->rq_allow_replay &&
1286 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1287 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1288 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1289 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1290 DEBUG_REQ(D_HA, req, "allow during recovery");
1300 * Decide if the error message should be printed to the console or not.
1301 * Makes its decision based on request type, status, and failure frequency.
1303 * \param[in] req request that failed and may need a console message
1305 * \retval false if no message should be printed
1306 * \retval true if console message should be printed
1308 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1310 LASSERT(req->rq_reqmsg != NULL);
1312 /* Suppress particular reconnect errors which are to be expected. */
1313 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1314 /* Suppress timed out reconnect requests */
1315 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1320 * Suppress most unavailable/again reconnect requests, but
1321 * print occasionally so it is clear client is trying to
1322 * connect to a server where no target is running.
1324 if ((err == -ENODEV || err == -EAGAIN) &&
1325 req->rq_import->imp_conn_cnt % 30 != 20)
1329 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1330 /* -EAGAIN is normal when using POSIX flocks */
1333 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1334 (req->rq_xid & 0xf) != 10)
1335 /* Suppress most ping requests, they may fail occasionally */
1342 * Check request processing status.
1343 * Returns the status.
1345 static int ptlrpc_check_status(struct ptlrpc_request *req)
1350 rc = lustre_msg_get_status(req->rq_repmsg);
1351 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1352 struct obd_import *imp = req->rq_import;
1353 struct lnet_nid *nid = &imp->imp_connection->c_peer.nid;
1354 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1356 if (ptlrpc_console_allow(req, opc, rc))
1357 LCONSOLE_ERROR_MSG(0x11,
1358 "%s: operation %s to node %s failed: rc = %d\n",
1359 imp->imp_obd->obd_name,
1361 libcfs_nidstr(nid), rc);
1362 RETURN(rc < 0 ? rc : -EINVAL);
1366 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1372 * save pre-versions of objects into request for replay.
1373 * Versions are obtained from server reply.
1376 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1378 struct lustre_msg *repmsg = req->rq_repmsg;
1379 struct lustre_msg *reqmsg = req->rq_reqmsg;
1380 __u64 *versions = lustre_msg_get_versions(repmsg);
1383 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1387 lustre_msg_set_versions(reqmsg, versions);
1388 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1389 versions[0], versions[1]);
1394 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1396 struct ptlrpc_request *req;
1398 assert_spin_locked(&imp->imp_lock);
1399 if (list_empty(&imp->imp_unreplied_list))
1402 req = list_first_entry(&imp->imp_unreplied_list, struct ptlrpc_request,
1404 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1406 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1407 imp->imp_known_replied_xid = req->rq_xid - 1;
1409 return req->rq_xid - 1;
1413 * Callback function called when client receives RPC reply for \a req.
1414 * Returns 0 on success or error code.
1415 * The return alue would be assigned to req->rq_status by the caller
1416 * as request processing status.
1417 * This function also decides if the request needs to be saved for later replay.
1419 static int after_reply(struct ptlrpc_request *req)
1421 struct obd_import *imp = req->rq_import;
1422 struct obd_device *obd = req->rq_import->imp_obd;
1429 LASSERT(obd != NULL);
1430 /* repbuf must be unlinked */
1431 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1433 if (req->rq_reply_truncated) {
1434 if (ptlrpc_no_resend(req)) {
1435 DEBUG_REQ(D_ERROR, req,
1436 "reply buffer overflow, expected=%d, actual size=%d",
1437 req->rq_nob_received, req->rq_repbuf_len);
1441 sptlrpc_cli_free_repbuf(req);
1443 * Pass the required reply buffer size (include
1444 * space for early reply).
1445 * NB: no need to roundup because alloc_repbuf
1448 req->rq_replen = req->rq_nob_received;
1449 req->rq_nob_received = 0;
1450 spin_lock(&req->rq_lock);
1452 spin_unlock(&req->rq_lock);
1456 work_start = ktime_get_real();
1457 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1460 * NB Until this point, the whole of the incoming message,
1461 * including buflens, status etc is in the sender's byte order.
1463 rc = sptlrpc_cli_unwrap_reply(req);
1465 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1470 * Security layer unwrap might ask resend this request.
1475 rc = unpack_reply(req);
1479 /* retry indefinitely on EINPROGRESS */
1480 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1481 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1482 time64_t now = ktime_get_real_seconds();
1484 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1485 D_RPCTRACE, req, "resending request on EINPROGRESS");
1486 spin_lock(&req->rq_lock);
1488 spin_unlock(&req->rq_lock);
1489 req->rq_nr_resend++;
1491 /* Readjust the timeout for current conditions */
1492 ptlrpc_at_set_req_timeout(req);
1494 * delay resend to give a chance to the server to get ready.
1495 * The delay is increased by 1s on every resend and is capped to
1496 * the current request timeout (i.e. obd_timeout if AT is off,
1497 * or AT service time x 125% + 5s, see at_est2timeout)
1499 if (req->rq_nr_resend > req->rq_timeout)
1500 req->rq_sent = now + req->rq_timeout;
1502 req->rq_sent = now + req->rq_nr_resend;
1504 /* Resend for EINPROGRESS will use a new XID */
1505 spin_lock(&imp->imp_lock);
1506 list_del_init(&req->rq_unreplied_list);
1507 spin_unlock(&imp->imp_lock);
1512 if (obd->obd_svc_stats) {
1513 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1515 ptlrpc_lprocfs_rpc_sent(req, timediff);
1518 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1519 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1520 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1521 lustre_msg_get_type(req->rq_repmsg));
1525 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1526 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1527 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1528 ptlrpc_at_adj_net_latency(req,
1529 lustre_msg_get_service_timeout(req->rq_repmsg));
1531 rc = ptlrpc_check_status(req);
1535 * Either we've been evicted, or the server has failed for
1536 * some reason. Try to reconnect, and if that fails, punt to
1539 if (ptlrpc_recoverable_error(rc)) {
1540 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1541 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1544 ptlrpc_request_handle_notconn(req);
1549 * Let's look if server sent slv. Do it only for RPC with
1552 ldlm_cli_update_pool(req);
1556 * Store transno in reqmsg for replay.
1558 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1559 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1560 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1563 if (imp->imp_replayable) {
1564 /* if other threads are waiting for ptlrpc_free_committed()
1565 * they could continue the work of freeing RPCs. That reduces
1566 * lock hold times, and distributes work more fairly across
1567 * waiting threads. We can't use spin_is_contended() since
1568 * there are many other places where imp_lock is held.
1570 atomic_inc(&imp->imp_waiting);
1571 spin_lock(&imp->imp_lock);
1572 atomic_dec(&imp->imp_waiting);
1574 * No point in adding already-committed requests to the replay
1575 * list, we will just remove them immediately. b=9829
1577 if (req->rq_transno != 0 &&
1579 lustre_msg_get_last_committed(req->rq_repmsg) ||
1581 /** version recovery */
1582 ptlrpc_save_versions(req);
1583 ptlrpc_retain_replayable_request(req, imp);
1584 } else if (req->rq_commit_cb &&
1585 list_empty(&req->rq_replay_list)) {
1587 * NB: don't call rq_commit_cb if it's already on
1588 * rq_replay_list, ptlrpc_free_committed() will call
1589 * it later, see LU-3618 for details
1591 spin_unlock(&imp->imp_lock);
1592 req->rq_commit_cb(req);
1593 atomic_inc(&imp->imp_waiting);
1594 spin_lock(&imp->imp_lock);
1595 atomic_dec(&imp->imp_waiting);
1599 * Replay-enabled imports return commit-status information.
1601 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1602 if (likely(committed > imp->imp_peer_committed_transno))
1603 imp->imp_peer_committed_transno = committed;
1605 ptlrpc_free_committed(imp);
1607 if (!list_empty(&imp->imp_replay_list)) {
1608 struct ptlrpc_request *last;
1610 last = list_entry(imp->imp_replay_list.prev,
1611 struct ptlrpc_request,
1614 * Requests with rq_replay stay on the list even if no
1615 * commit is expected.
1617 if (last->rq_transno > imp->imp_peer_committed_transno)
1618 ptlrpc_pinger_commit_expected(imp);
1621 spin_unlock(&imp->imp_lock);
1628 * Helper function to send request \a req over the network for the first time
1629 * Also adjusts request phase.
1630 * Returns 0 on success or error code.
1632 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1634 struct obd_import *imp = req->rq_import;
1639 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1641 /* do not try to go further if there is not enough memory in enc_pool */
1642 if (req->rq_sent && req->rq_bulk)
1643 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1644 pool_is_at_full_capacity())
1647 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1648 (!req->rq_generation_set ||
1649 req->rq_import_generation == imp->imp_generation))
1652 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1654 spin_lock(&imp->imp_lock);
1656 LASSERT(req->rq_xid != 0);
1657 LASSERT(!list_empty(&req->rq_unreplied_list));
1659 if (!req->rq_generation_set)
1660 req->rq_import_generation = imp->imp_generation;
1662 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1663 spin_lock(&req->rq_lock);
1664 req->rq_waiting = 1;
1665 spin_unlock(&req->rq_lock);
1667 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1668 ptlrpc_import_state_name(req->rq_send_state),
1669 ptlrpc_import_state_name(imp->imp_state));
1670 LASSERT(list_empty(&req->rq_list));
1671 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1672 atomic_inc(&req->rq_import->imp_inflight);
1673 spin_unlock(&imp->imp_lock);
1678 spin_unlock(&imp->imp_lock);
1679 req->rq_status = rc;
1680 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1684 LASSERT(list_empty(&req->rq_list));
1685 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1686 atomic_inc(&req->rq_import->imp_inflight);
1689 * find the known replied XID from the unreplied list, CONNECT
1690 * and DISCONNECT requests are skipped to make the sanity check
1691 * on server side happy. see process_req_last_xid().
1693 * For CONNECT: Because replay requests have lower XID, it'll
1694 * break the sanity check if CONNECT bump the exp_last_xid on
1697 * For DISCONNECT: Since client will abort inflight RPC before
1698 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1699 * than the inflight RPC.
1701 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1702 min_xid = ptlrpc_known_replied_xid(imp);
1703 spin_unlock(&imp->imp_lock);
1705 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1707 lustre_msg_set_status(req->rq_reqmsg, current->pid);
1709 /* If the request to be sent is an LDLM callback, do not try to
1711 * An LDLM callback is sent by a server to a client in order to make
1712 * it release a lock, on a communication channel that uses a reverse
1713 * context. It cannot be refreshed on its own, as it is the 'reverse'
1714 * (server-side) representation of a client context.
1715 * We do not care if the reverse context is expired, and want to send
1716 * the LDLM callback anyway. Once the client receives the AST, it is
1717 * its job to refresh its own context if it has expired, hence
1718 * refreshing the associated reverse context on server side, before
1719 * being able to send the LDLM_CANCEL requested by the server.
1721 if (lustre_msg_get_opc(req->rq_reqmsg) != LDLM_BL_CALLBACK &&
1722 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_CP_CALLBACK &&
1723 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_GL_CALLBACK)
1724 rc = sptlrpc_req_refresh_ctx(req, 0);
1727 req->rq_status = rc;
1730 spin_lock(&req->rq_lock);
1731 req->rq_wait_ctx = 1;
1732 spin_unlock(&req->rq_lock);
1738 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1740 imp->imp_obd->obd_uuid.uuid,
1741 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1742 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1743 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1745 rc = ptl_send_rpc(req, 0);
1746 if (rc == -ENOMEM) {
1747 spin_lock(&imp->imp_lock);
1748 if (!list_empty(&req->rq_list)) {
1749 list_del_init(&req->rq_list);
1750 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1751 wake_up(&req->rq_import->imp_recovery_waitq);
1753 spin_unlock(&imp->imp_lock);
1754 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1758 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1760 spin_lock(&req->rq_lock);
1761 req->rq_net_err = 1;
1762 spin_unlock(&req->rq_lock);
1768 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1773 LASSERT(set->set_producer != NULL);
1775 remaining = atomic_read(&set->set_remaining);
1778 * populate the ->set_requests list with requests until we
1779 * reach the maximum number of RPCs in flight for this set
1781 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1782 rc = set->set_producer(set, set->set_producer_arg);
1783 if (rc == -ENOENT) {
1784 /* no more RPC to produce */
1785 set->set_producer = NULL;
1786 set->set_producer_arg = NULL;
1791 RETURN((atomic_read(&set->set_remaining) - remaining));
1795 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1796 * and no more replies are expected.
1797 * (it is possible to get less replies than requests sent e.g. due to timed out
1798 * requests or requests that we had trouble to send out)
1800 * NOTE: This function contains a potential schedule point (cond_resched()).
1802 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1804 struct ptlrpc_request *req, *next;
1805 LIST_HEAD(comp_reqs);
1806 int force_timer_recalc = 0;
1809 if (atomic_read(&set->set_remaining) == 0)
1812 list_for_each_entry_safe(req, next, &set->set_requests,
1814 struct obd_import *imp = req->rq_import;
1815 int unregistered = 0;
1819 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1820 list_move_tail(&req->rq_set_chain, &comp_reqs);
1825 * This schedule point is mainly for the ptlrpcd caller of this
1826 * function. Most ptlrpc sets are not long-lived and unbounded
1827 * in length, but at the least the set used by the ptlrpcd is.
1828 * Since the processing time is unbounded, we need to insert an
1829 * explicit schedule point to make the thread well-behaved.
1834 * If the caller requires to allow to be interpreted by force
1835 * and it has really been interpreted, then move the request
1836 * to RQ_PHASE_INTERPRET phase in spite of what the current
1839 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1840 req->rq_status = -EINTR;
1841 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1844 * Since it is interpreted and we have to wait for
1845 * the reply to be unlinked, then use sync mode.
1849 GOTO(interpret, req->rq_status);
1852 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1853 force_timer_recalc = 1;
1855 /* delayed send - skip */
1856 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1859 /* delayed resend - skip */
1860 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1861 req->rq_sent > ktime_get_real_seconds())
1864 if (!(req->rq_phase == RQ_PHASE_RPC ||
1865 req->rq_phase == RQ_PHASE_BULK ||
1866 req->rq_phase == RQ_PHASE_INTERPRET ||
1867 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1868 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1869 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1873 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1874 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1875 LASSERT(req->rq_next_phase != req->rq_phase);
1876 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1878 if (req->rq_req_deadline &&
1879 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1880 req->rq_req_deadline = 0;
1881 if (req->rq_reply_deadline &&
1882 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1883 req->rq_reply_deadline = 0;
1884 if (req->rq_bulk_deadline &&
1885 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1886 req->rq_bulk_deadline = 0;
1889 * Skip processing until reply is unlinked. We
1890 * can't return to pool before that and we can't
1891 * call interpret before that. We need to make
1892 * sure that all rdma transfers finished and will
1893 * not corrupt any data.
1895 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1896 ptlrpc_cli_wait_unlink(req))
1898 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1899 ptlrpc_client_bulk_active(req))
1903 * Turn fail_loc off to prevent it from looping
1906 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1907 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1910 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1911 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1916 * Move to next phase if reply was successfully
1919 ptlrpc_rqphase_move(req, req->rq_next_phase);
1922 if (req->rq_phase == RQ_PHASE_INTERPRET)
1923 GOTO(interpret, req->rq_status);
1926 * Note that this also will start async reply unlink.
1928 if (req->rq_net_err && !req->rq_timedout) {
1929 ptlrpc_expire_one_request(req, 1);
1932 * Check if we still need to wait for unlink.
1934 if (ptlrpc_cli_wait_unlink(req) ||
1935 ptlrpc_client_bulk_active(req))
1937 /* If there is no need to resend, fail it now. */
1938 if (req->rq_no_resend) {
1939 if (req->rq_status == 0)
1940 req->rq_status = -EIO;
1941 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1942 GOTO(interpret, req->rq_status);
1949 if (!ptlrpc_unregister_reply(req, 1)) {
1950 ptlrpc_unregister_bulk(req, 1);
1954 spin_lock(&req->rq_lock);
1955 req->rq_replied = 0;
1956 spin_unlock(&req->rq_lock);
1957 if (req->rq_status == 0)
1958 req->rq_status = -EIO;
1959 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1960 GOTO(interpret, req->rq_status);
1964 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1965 * so it sets rq_intr regardless of individual rpc
1966 * timeouts. The synchronous IO waiting path sets
1967 * rq_intr irrespective of whether ptlrpcd
1968 * has seen a timeout. Our policy is to only interpret
1969 * interrupted rpcs after they have timed out, so we
1970 * need to enforce that here.
1973 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1974 req->rq_wait_ctx)) {
1975 req->rq_status = -EINTR;
1976 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1977 GOTO(interpret, req->rq_status);
1980 if (req->rq_phase == RQ_PHASE_RPC) {
1981 if (req->rq_timedout || req->rq_resend ||
1982 req->rq_waiting || req->rq_wait_ctx) {
1985 if (!ptlrpc_unregister_reply(req, 1)) {
1986 ptlrpc_unregister_bulk(req, 1);
1990 spin_lock(&imp->imp_lock);
1991 if (ptlrpc_import_delay_req(imp, req,
1994 * put on delay list - only if we wait
1995 * recovery finished - before send
1997 list_move_tail(&req->rq_list,
1998 &imp->imp_delayed_list);
1999 spin_unlock(&imp->imp_lock);
2004 req->rq_status = status;
2005 ptlrpc_rqphase_move(req,
2006 RQ_PHASE_INTERPRET);
2007 spin_unlock(&imp->imp_lock);
2008 GOTO(interpret, req->rq_status);
2010 /* ignore on just initiated connections */
2011 if (ptlrpc_no_resend(req) &&
2012 !req->rq_wait_ctx &&
2013 imp->imp_generation !=
2014 imp->imp_initiated_at) {
2015 req->rq_status = -ENOTCONN;
2016 ptlrpc_rqphase_move(req,
2017 RQ_PHASE_INTERPRET);
2018 spin_unlock(&imp->imp_lock);
2019 GOTO(interpret, req->rq_status);
2022 /* don't resend too fast in case of network
2025 if (ktime_get_real_seconds() < (req->rq_sent + 1)
2026 && req->rq_net_err && req->rq_timedout) {
2028 DEBUG_REQ(D_INFO, req,
2029 "throttle request");
2030 /* Don't try to resend RPC right away
2031 * as it is likely it will fail again
2032 * and ptlrpc_check_set() will be
2033 * called again, keeping this thread
2034 * busy. Instead, wait for the next
2035 * timeout. Flag it as resend to
2036 * ensure we don't wait to long.
2039 spin_unlock(&imp->imp_lock);
2043 list_move_tail(&req->rq_list,
2044 &imp->imp_sending_list);
2046 spin_unlock(&imp->imp_lock);
2048 spin_lock(&req->rq_lock);
2049 req->rq_waiting = 0;
2050 spin_unlock(&req->rq_lock);
2052 if (req->rq_timedout || req->rq_resend) {
2054 * This is re-sending anyways,
2055 * let's mark req as resend.
2057 spin_lock(&req->rq_lock);
2059 spin_unlock(&req->rq_lock);
2062 * rq_wait_ctx is only touched by ptlrpcd,
2063 * so no lock is needed here.
2065 status = sptlrpc_req_refresh_ctx(req, 0);
2068 req->rq_status = status;
2069 spin_lock(&req->rq_lock);
2070 req->rq_wait_ctx = 0;
2071 spin_unlock(&req->rq_lock);
2072 force_timer_recalc = 1;
2074 spin_lock(&req->rq_lock);
2075 req->rq_wait_ctx = 1;
2076 spin_unlock(&req->rq_lock);
2081 spin_lock(&req->rq_lock);
2082 req->rq_wait_ctx = 0;
2083 spin_unlock(&req->rq_lock);
2087 * In any case, the previous bulk should be
2088 * cleaned up to prepare for the new sending
2091 !ptlrpc_unregister_bulk(req, 1))
2094 rc = ptl_send_rpc(req, 0);
2095 if (rc == -ENOMEM) {
2096 spin_lock(&imp->imp_lock);
2097 if (!list_empty(&req->rq_list))
2098 list_del_init(&req->rq_list);
2099 spin_unlock(&imp->imp_lock);
2100 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2104 DEBUG_REQ(D_HA, req,
2105 "send failed: rc = %d", rc);
2106 force_timer_recalc = 1;
2107 spin_lock(&req->rq_lock);
2108 req->rq_net_err = 1;
2109 spin_unlock(&req->rq_lock);
2112 /* need to reset the timeout */
2113 force_timer_recalc = 1;
2116 spin_lock(&req->rq_lock);
2118 if (ptlrpc_client_early(req)) {
2119 ptlrpc_at_recv_early_reply(req);
2120 spin_unlock(&req->rq_lock);
2124 /* Still waiting for a reply? */
2125 if (ptlrpc_client_recv(req)) {
2126 spin_unlock(&req->rq_lock);
2130 /* Did we actually receive a reply? */
2131 if (!ptlrpc_client_replied(req)) {
2132 spin_unlock(&req->rq_lock);
2136 spin_unlock(&req->rq_lock);
2139 * unlink from net because we are going to
2140 * swab in-place of reply buffer
2142 unregistered = ptlrpc_unregister_reply(req, 1);
2146 req->rq_status = after_reply(req);
2147 if (req->rq_resend) {
2148 force_timer_recalc = 1;
2153 * If there is no bulk associated with this request,
2154 * then we're done and should let the interpreter
2155 * process the reply. Similarly if the RPC returned
2156 * an error, and therefore the bulk will never arrive.
2158 if (!req->rq_bulk || req->rq_status < 0) {
2159 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2160 GOTO(interpret, req->rq_status);
2163 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2166 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2167 if (ptlrpc_client_bulk_active(req))
2170 if (req->rq_bulk->bd_failure) {
2172 * The RPC reply arrived OK, but the bulk screwed
2173 * up! Dead weird since the server told us the RPC
2174 * was good after getting the REPLY for her GET or
2175 * the ACK for her PUT.
2177 DEBUG_REQ(D_ERROR, req, "bulk transfer failed %d/%d/%d",
2179 req->rq_bulk->bd_nob,
2180 req->rq_bulk->bd_nob_transferred);
2181 req->rq_status = -EIO;
2184 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2187 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2190 * This moves to "unregistering" phase we need to wait for
2193 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2194 /* start async bulk unlink too */
2195 ptlrpc_unregister_bulk(req, 1);
2199 if (!ptlrpc_unregister_bulk(req, async))
2203 * When calling interpret receiving already should be
2206 LASSERT(!req->rq_receiving_reply);
2208 ptlrpc_req_interpret(env, req, req->rq_status);
2210 if (ptlrpcd_check_work(req)) {
2211 atomic_dec(&set->set_remaining);
2214 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2218 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2220 imp->imp_obd->obd_uuid.uuid,
2221 lustre_msg_get_status(req->rq_reqmsg),
2223 obd_import_nid2str(imp),
2224 lustre_msg_get_opc(req->rq_reqmsg),
2225 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2227 spin_lock(&imp->imp_lock);
2229 * Request already may be not on sending or delaying list. This
2230 * may happen in the case of marking it erroneous for the case
2231 * ptlrpc_import_delay_req(req, status) find it impossible to
2232 * allow sending this rpc and returns *status != 0.
2234 if (!list_empty(&req->rq_list)) {
2235 list_del_init(&req->rq_list);
2236 if (atomic_dec_and_test(&imp->imp_inflight))
2237 wake_up(&imp->imp_recovery_waitq);
2239 list_del_init(&req->rq_unreplied_list);
2240 spin_unlock(&imp->imp_lock);
2242 atomic_dec(&set->set_remaining);
2243 wake_up(&imp->imp_recovery_waitq);
2245 if (set->set_producer) {
2246 /* produce a new request if possible */
2247 if (ptlrpc_set_producer(set) > 0)
2248 force_timer_recalc = 1;
2251 * free the request that has just been completed
2252 * in order not to pollute set->set_requests
2254 list_del_init(&req->rq_set_chain);
2255 spin_lock(&req->rq_lock);
2257 req->rq_invalid_rqset = 0;
2258 spin_unlock(&req->rq_lock);
2260 /* record rq_status to compute the final status later */
2261 if (req->rq_status != 0)
2262 set->set_rc = req->rq_status;
2263 ptlrpc_req_finished(req);
2265 list_move_tail(&req->rq_set_chain, &comp_reqs);
2270 * move completed request at the head of list so it's easier for
2271 * caller to find them
2273 list_splice(&comp_reqs, &set->set_requests);
2275 /* If we hit an error, we want to recover promptly. */
2276 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2278 EXPORT_SYMBOL(ptlrpc_check_set);
2281 * Time out request \a req. is \a async_unlink is set, that means do not wait
2282 * until LNet actually confirms network buffer unlinking.
2283 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2285 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2287 struct obd_import *imp = req->rq_import;
2288 unsigned int debug_mask = D_RPCTRACE;
2292 spin_lock(&req->rq_lock);
2293 req->rq_timedout = 1;
2294 spin_unlock(&req->rq_lock);
2296 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2297 lustre_msg_get_status(req->rq_reqmsg)))
2298 debug_mask = D_WARNING;
2299 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2300 req->rq_net_err ? "failed due to network error" :
2301 ((req->rq_real_sent == 0 ||
2302 req->rq_real_sent < req->rq_sent ||
2303 req->rq_real_sent >= req->rq_deadline) ?
2304 "timed out for sent delay" : "timed out for slow reply"),
2305 req->rq_sent, req->rq_real_sent);
2307 if (imp && obd_debug_peer_on_timeout)
2308 LNetDebugPeer(&imp->imp_connection->c_peer);
2310 ptlrpc_unregister_reply(req, async_unlink);
2311 ptlrpc_unregister_bulk(req, async_unlink);
2313 if (obd_dump_on_timeout)
2314 libcfs_debug_dumplog();
2317 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2321 atomic_inc(&imp->imp_timeouts);
2323 /* The DLM server doesn't want recovery run on its imports. */
2324 if (imp->imp_dlm_fake)
2328 * If this request is for recovery or other primordial tasks,
2329 * then error it out here.
2331 if (req->rq_ctx_init || req->rq_ctx_fini ||
2332 req->rq_send_state != LUSTRE_IMP_FULL ||
2333 imp->imp_obd->obd_no_recov) {
2334 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2335 ptlrpc_import_state_name(req->rq_send_state),
2336 ptlrpc_import_state_name(imp->imp_state));
2337 spin_lock(&req->rq_lock);
2338 req->rq_status = -ETIMEDOUT;
2340 spin_unlock(&req->rq_lock);
2345 * if a request can't be resent we can't wait for an answer after
2348 if (ptlrpc_no_resend(req)) {
2349 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2353 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2359 * Time out all uncompleted requests in request set pointed by \a data
2360 * This is called when a wait times out.
2362 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2364 struct ptlrpc_request *req;
2365 time64_t now = ktime_get_real_seconds();
2368 LASSERT(set != NULL);
2371 * A timeout expired. See which reqs it applies to...
2373 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2374 /* don't expire request waiting for context */
2375 if (req->rq_wait_ctx)
2378 /* Request in-flight? */
2379 if (!((req->rq_phase == RQ_PHASE_RPC &&
2380 !req->rq_waiting && !req->rq_resend) ||
2381 (req->rq_phase == RQ_PHASE_BULK)))
2384 if (req->rq_timedout || /* already dealt with */
2385 req->rq_deadline > now) /* not expired */
2389 * Deal with this guy. Do it asynchronously to not block
2392 ptlrpc_expire_one_request(req, 1);
2394 * Loops require that we resched once in a while to avoid
2395 * RCU stalls and a few other problems.
2403 * Interrupts (sets interrupted flag) all uncompleted requests in
2404 * a set \a data. This is called when a wait_event is interrupted
2407 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2409 struct ptlrpc_request *req;
2411 LASSERT(set != NULL);
2412 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2414 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2418 if (req->rq_phase != RQ_PHASE_RPC &&
2419 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2420 !req->rq_allow_intr)
2423 spin_lock(&req->rq_lock);
2425 spin_unlock(&req->rq_lock);
2430 * Get the smallest timeout in the set; this does NOT set a timeout.
2432 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2434 time64_t now = ktime_get_real_seconds();
2436 struct ptlrpc_request *req;
2440 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2441 /* Request in-flight? */
2442 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2443 (req->rq_phase == RQ_PHASE_BULK) ||
2444 (req->rq_phase == RQ_PHASE_NEW)))
2447 /* Already timed out. */
2448 if (req->rq_timedout)
2451 /* Waiting for ctx. */
2452 if (req->rq_wait_ctx)
2455 if (req->rq_phase == RQ_PHASE_NEW)
2456 deadline = req->rq_sent;
2457 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2458 deadline = req->rq_sent;
2460 deadline = req->rq_sent + req->rq_timeout;
2462 if (deadline <= now) /* actually expired already */
2463 timeout = 1; /* ASAP */
2464 else if (timeout == 0 || timeout > deadline - now)
2465 timeout = deadline - now;
2471 * Send all unset request from the set and then wait untill all
2472 * requests in the set complete (either get a reply, timeout, get an
2473 * error or otherwise be interrupted).
2474 * Returns 0 on success or error code otherwise.
2476 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2478 struct ptlrpc_request *req;
2483 if (set->set_producer)
2484 (void)ptlrpc_set_producer(set);
2486 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2487 if (req->rq_phase == RQ_PHASE_NEW)
2488 (void)ptlrpc_send_new_req(req);
2491 if (list_empty(&set->set_requests))
2495 timeout = ptlrpc_set_next_timeout(set);
2498 * wait until all complete, interrupted, or an in-flight
2501 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2504 if ((timeout == 0 && !signal_pending(current)) ||
2505 set->set_allow_intr) {
2507 * No requests are in-flight (ether timed out
2508 * or delayed), so we can allow interrupts.
2509 * We still want to block for a limited time,
2510 * so we allow interrupts during the timeout.
2512 rc = l_wait_event_abortable_timeout(
2514 ptlrpc_check_set(NULL, set),
2515 cfs_time_seconds(timeout ? timeout : 1));
2518 ptlrpc_expired_set(set);
2519 } else if (rc < 0) {
2521 ptlrpc_interrupted_set(set);
2527 * At least one request is in flight, so no
2528 * interrupts are allowed. Wait until all
2529 * complete, or an in-flight req times out.
2531 rc = wait_event_idle_timeout(
2533 ptlrpc_check_set(NULL, set),
2534 cfs_time_seconds(timeout ? timeout : 1));
2536 ptlrpc_expired_set(set);
2543 * LU-769 - if we ignored the signal because
2544 * it was already pending when we started, we
2545 * need to handle it now or we risk it being
2548 if (rc == -ETIMEDOUT &&
2549 signal_pending(current)) {
2552 siginitset(&new, LUSTRE_FATAL_SIGS);
2553 sigprocmask(SIG_BLOCK, &new, &old);
2555 * In fact we only interrupt for the
2556 * "fatal" signals like SIGINT or
2557 * SIGKILL. We still ignore less
2558 * important signals since ptlrpc set
2559 * is not easily reentrant from
2562 if (signal_pending(current))
2563 ptlrpc_interrupted_set(set);
2564 sigprocmask(SIG_SETMASK, &old, NULL);
2568 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2571 * -EINTR => all requests have been flagged rq_intr so next
2573 * -ETIMEDOUT => someone timed out. When all reqs have
2574 * timed out, signals are enabled allowing completion with
2576 * I don't really care if we go once more round the loop in
2577 * the error cases -eeb.
2579 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2580 list_for_each_entry(req, &set->set_requests,
2582 spin_lock(&req->rq_lock);
2583 req->rq_invalid_rqset = 1;
2584 spin_unlock(&req->rq_lock);
2587 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2589 LASSERT(atomic_read(&set->set_remaining) == 0);
2591 rc = set->set_rc; /* rq_status of already freed requests if any */
2592 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2593 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2594 if (req->rq_status != 0)
2595 rc = req->rq_status;
2600 EXPORT_SYMBOL(ptlrpc_set_wait);
2603 * Helper fuction for request freeing.
2604 * Called when request count reached zero and request needs to be freed.
2605 * Removes request from all sorts of sending/replay lists it might be on,
2606 * frees network buffers if any are present.
2607 * If \a locked is set, that means caller is already holding import imp_lock
2608 * and so we no longer need to reobtain it (for certain lists manipulations)
2610 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2617 LASSERT(!request->rq_srv_req);
2618 LASSERT(request->rq_export == NULL);
2619 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2620 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2621 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2622 LASSERTF(!request->rq_replay, "req %p\n", request);
2624 req_capsule_fini(&request->rq_pill);
2627 * We must take it off the imp_replay_list first. Otherwise, we'll set
2628 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2630 if (request->rq_import) {
2632 spin_lock(&request->rq_import->imp_lock);
2633 list_del_init(&request->rq_replay_list);
2634 list_del_init(&request->rq_unreplied_list);
2636 spin_unlock(&request->rq_import->imp_lock);
2638 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2640 if (atomic_read(&request->rq_refcount) != 0) {
2641 DEBUG_REQ(D_ERROR, request,
2642 "freeing request with nonzero refcount");
2646 if (request->rq_repbuf)
2647 sptlrpc_cli_free_repbuf(request);
2649 if (request->rq_import) {
2650 if (!ptlrpcd_check_work(request)) {
2651 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2652 atomic_dec(&request->rq_import->imp_reqs);
2654 class_import_put(request->rq_import);
2655 request->rq_import = NULL;
2657 if (request->rq_bulk)
2658 ptlrpc_free_bulk(request->rq_bulk);
2660 if (request->rq_reqbuf || request->rq_clrbuf)
2661 sptlrpc_cli_free_reqbuf(request);
2663 if (request->rq_cli_ctx)
2664 sptlrpc_req_put_ctx(request, !locked);
2666 if (request->rq_pool)
2667 __ptlrpc_free_req_to_pool(request);
2669 ptlrpc_request_cache_free(request);
2673 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2675 * Drop one request reference. Must be called with import imp_lock held.
2676 * When reference count drops to zero, request is freed.
2678 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2680 assert_spin_locked(&request->rq_import->imp_lock);
2681 (void)__ptlrpc_req_finished(request, 1);
2686 * Drops one reference count for request \a request.
2687 * \a locked set indicates that caller holds import imp_lock.
2688 * Frees the request whe reference count reaches zero.
2690 * \retval 1 the request is freed
2691 * \retval 0 some others still hold references on the request
2693 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2701 LASSERT(request != LP_POISON);
2702 LASSERT(request->rq_reqmsg != LP_POISON);
2704 DEBUG_REQ(D_INFO, request, "refcount now %u",
2705 atomic_read(&request->rq_refcount) - 1);
2707 spin_lock(&request->rq_lock);
2708 count = atomic_dec_return(&request->rq_refcount);
2709 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2712 * For open RPC, the client does not know the EA size (LOV, ACL, and
2713 * so on) before replied, then the client has to reserve very large
2714 * reply buffer. Such buffer will not be released until the RPC freed.
2715 * Since The open RPC is replayable, we need to keep it in the replay
2716 * list until close. If there are a lot of files opened concurrently,
2717 * then the client may be OOM.
2719 * If fact, it is unnecessary to keep reply buffer for open replay,
2720 * related EAs have already been saved via mdc_save_lovea() before
2721 * coming here. So it is safe to free the reply buffer some earlier
2722 * before releasing the RPC to avoid client OOM. LU-9514
2724 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2725 spin_lock(&request->rq_early_free_lock);
2726 sptlrpc_cli_free_repbuf(request);
2727 request->rq_repbuf = NULL;
2728 request->rq_repbuf_len = 0;
2729 request->rq_repdata = NULL;
2730 request->rq_reqdata_len = 0;
2731 spin_unlock(&request->rq_early_free_lock);
2733 spin_unlock(&request->rq_lock);
2736 __ptlrpc_free_req(request, locked);
2742 * Drops one reference count for a request.
2744 void ptlrpc_req_finished(struct ptlrpc_request *request)
2746 __ptlrpc_req_finished(request, 0);
2748 EXPORT_SYMBOL(ptlrpc_req_finished);
2751 * Returns xid of a \a request
2753 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2755 return request->rq_xid;
2757 EXPORT_SYMBOL(ptlrpc_req_xid);
2760 * Disengage the client's reply buffer from the network
2761 * NB does _NOT_ unregister any client-side bulk.
2762 * IDEMPOTENT, but _not_ safe against concurrent callers.
2763 * The request owner (i.e. the thread doing the I/O) must call...
2764 * Returns 0 on success or 1 if unregistering cannot be made.
2766 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2768 bool discard = false;
2772 LASSERT(!in_interrupt());
2774 /* Let's setup deadline for reply unlink. */
2775 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2776 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2777 request->rq_reply_deadline = ktime_get_real_seconds() +
2778 PTLRPC_REQ_LONG_UNLINK;
2781 * Nothing left to do.
2783 if (!__ptlrpc_cli_wait_unlink(request, &discard))
2786 LNetMDUnlink(request->rq_reply_md_h);
2788 if (discard) /* Discard the request-out callback */
2789 __LNetMDUnlink(request->rq_req_md_h, discard);
2792 * Let's check it once again.
2794 if (!ptlrpc_cli_wait_unlink(request))
2797 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2798 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2801 * Do not wait for unlink to finish.
2807 * We have to wait_event_idle_timeout() whatever the result, to get
2808 * a chance to run reply_in_callback(), and to make sure we've
2809 * unlinked before returning a req to the pool.
2812 wait_queue_head_t *wq = (request->rq_set) ?
2813 &request->rq_set->set_waitq :
2814 &request->rq_reply_waitq;
2815 int seconds = PTLRPC_REQ_LONG_UNLINK;
2817 * Network access will complete in finite time but the HUGE
2818 * timeout lets us CWARN for visibility of sluggish NALs
2820 while (seconds > 0 &&
2821 wait_event_idle_timeout(
2823 !ptlrpc_cli_wait_unlink(request),
2824 cfs_time_seconds(1)) == 0)
2827 ptlrpc_rqphase_move(request, request->rq_next_phase);
2831 DEBUG_REQ(D_WARNING, request,
2832 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2833 request->rq_receiving_reply,
2834 request->rq_req_unlinked,
2835 request->rq_reply_unlinked);
2840 static void ptlrpc_free_request(struct ptlrpc_request *req)
2842 spin_lock(&req->rq_lock);
2844 spin_unlock(&req->rq_lock);
2846 if (req->rq_commit_cb)
2847 req->rq_commit_cb(req);
2848 list_del_init(&req->rq_replay_list);
2850 __ptlrpc_req_finished(req, 1);
2854 * the request is committed and dropped from the replay list of its import
2856 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2858 struct obd_import *imp = req->rq_import;
2860 spin_lock(&imp->imp_lock);
2861 if (list_empty(&req->rq_replay_list)) {
2862 spin_unlock(&imp->imp_lock);
2866 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2867 if (imp->imp_replay_cursor == &req->rq_replay_list)
2868 imp->imp_replay_cursor = req->rq_replay_list.next;
2869 ptlrpc_free_request(req);
2872 spin_unlock(&imp->imp_lock);
2874 EXPORT_SYMBOL(ptlrpc_request_committed);
2877 * Iterates through replay_list on import and prunes
2878 * all requests have transno smaller than last_committed for the
2879 * import and don't have rq_replay set.
2880 * Since requests are sorted in transno order, stops when meeting first
2881 * transno bigger than last_committed.
2882 * caller must hold imp->imp_lock
2884 void ptlrpc_free_committed(struct obd_import *imp)
2886 struct ptlrpc_request *req, *saved;
2887 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2888 bool skip_committed_list = true;
2889 unsigned int replay_scanned = 0, replay_freed = 0;
2890 unsigned int commit_scanned = 0, commit_freed = 0;
2891 unsigned int debug_level = D_INFO;
2892 __u64 peer_committed_transno;
2894 time64_t start, now;
2897 LASSERT(imp != NULL);
2898 assert_spin_locked(&imp->imp_lock);
2900 start = ktime_get_seconds();
2901 /* save these here, we can potentially drop imp_lock after checking */
2902 peer_committed_transno = imp->imp_peer_committed_transno;
2903 imp_generation = imp->imp_generation;
2905 if (peer_committed_transno == imp->imp_last_transno_checked &&
2906 imp_generation == imp->imp_last_generation_checked) {
2907 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2908 imp->imp_obd->obd_name, peer_committed_transno);
2911 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2912 imp->imp_obd->obd_name, peer_committed_transno, imp_generation);
2914 if (imp_generation != imp->imp_last_generation_checked ||
2915 imp->imp_last_transno_checked == 0)
2916 skip_committed_list = false;
2917 /* maybe drop imp_lock here, if another lock protected the lists */
2919 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2921 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2922 LASSERT(req != last_req);
2925 if (req->rq_transno == 0) {
2926 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2930 /* If other threads are waiting on imp_lock, stop processing
2931 * in this thread. Another thread can finish remaining work.
2932 * This may happen if there are huge numbers of open files
2933 * that are closed suddenly or evicted, or if the server
2934 * commit interval is very high vs. RPC rate.
2936 if (++replay_scanned % 2048 == 0) {
2937 now = ktime_get_seconds();
2938 if (now > start + 5)
2939 debug_level = D_WARNING;
2941 if ((replay_freed > 128 && now > start + 3) &&
2942 atomic_read(&imp->imp_waiting)) {
2943 if (debug_level == D_INFO)
2944 debug_level = D_RPCTRACE;
2949 if (req->rq_import_generation < imp_generation) {
2950 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2954 /* not yet committed */
2955 if (req->rq_transno > peer_committed_transno) {
2956 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2960 if (req->rq_replay) {
2961 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2962 list_move_tail(&req->rq_replay_list,
2963 &imp->imp_committed_list);
2967 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2968 peer_committed_transno);
2971 ptlrpc_free_request(req);
2974 if (skip_committed_list)
2977 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2979 LASSERT(req->rq_transno != 0);
2981 /* If other threads are waiting on imp_lock, stop processing
2982 * in this thread. Another thread can finish remaining work. */
2983 if (++commit_scanned % 2048 == 0) {
2984 now = ktime_get_seconds();
2985 if (now > start + 6)
2986 debug_level = D_WARNING;
2988 if ((commit_freed > 128 && now > start + 4) &&
2989 atomic_read(&imp->imp_waiting)) {
2990 if (debug_level == D_INFO)
2991 debug_level = D_RPCTRACE;
2996 if (req->rq_import_generation < imp_generation ||
2998 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
2999 req->rq_import_generation <
3000 imp_generation ? "stale" : "closed");
3002 if (imp->imp_replay_cursor == &req->rq_replay_list)
3003 imp->imp_replay_cursor =
3004 req->rq_replay_list.next;
3007 ptlrpc_free_request(req);
3011 /* if full lists processed without interruption, avoid next scan */
3012 if (debug_level == D_INFO) {
3013 imp->imp_last_transno_checked = peer_committed_transno;
3014 imp->imp_last_generation_checked = imp_generation;
3017 CDEBUG_LIMIT(debug_level,
3018 "%s: %s: skip=%u replay=%u/%u committed=%u/%u\n",
3019 imp->imp_obd->obd_name,
3020 debug_level == D_INFO ? "normal" : "overloaded",
3021 skip_committed_list, replay_freed, replay_scanned,
3022 commit_freed, commit_scanned);
3026 void ptlrpc_cleanup_client(struct obd_import *imp)
3033 * Schedule previously sent request for resend.
3034 * For bulk requests we assign new xid (to avoid problems with
3035 * lost replies and therefore several transfers landing into same buffer
3036 * from different sending attempts).
3038 void ptlrpc_resend_req(struct ptlrpc_request *req)
3040 DEBUG_REQ(D_HA, req, "going to resend");
3041 spin_lock(&req->rq_lock);
3044 * Request got reply but linked to the import list still.
3045 * Let ptlrpc_check_set() process it.
3047 if (ptlrpc_client_replied(req)) {
3048 spin_unlock(&req->rq_lock);
3049 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
3053 req->rq_status = -EAGAIN;
3056 req->rq_net_err = 0;
3057 req->rq_timedout = 0;
3059 ptlrpc_client_wake_req(req);
3060 spin_unlock(&req->rq_lock);
3063 /* XXX: this function and rq_status are currently unused */
3064 void ptlrpc_restart_req(struct ptlrpc_request *req)
3066 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
3067 req->rq_status = -ERESTARTSYS;
3069 spin_lock(&req->rq_lock);
3070 req->rq_restart = 1;
3071 req->rq_timedout = 0;
3072 ptlrpc_client_wake_req(req);
3073 spin_unlock(&req->rq_lock);
3077 * Grab additional reference on a request \a req
3079 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
3082 atomic_inc(&req->rq_refcount);
3085 EXPORT_SYMBOL(ptlrpc_request_addref);
3088 * Add a request to import replay_list.
3089 * Must be called under imp_lock
3091 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3092 struct obd_import *imp)
3094 struct ptlrpc_request *iter;
3096 assert_spin_locked(&imp->imp_lock);
3098 if (req->rq_transno == 0) {
3099 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3104 * clear this for new requests that were resent as well
3105 * as resent replayed requests.
3107 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3109 /* don't re-add requests that have been replayed */
3110 if (!list_empty(&req->rq_replay_list))
3113 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3115 spin_lock(&req->rq_lock);
3117 spin_unlock(&req->rq_lock);
3119 LASSERT(imp->imp_replayable);
3120 /* Balanced in ptlrpc_free_committed, usually. */
3121 ptlrpc_request_addref(req);
3122 list_for_each_entry_reverse(iter, &imp->imp_replay_list,
3125 * We may have duplicate transnos if we create and then
3126 * open a file, or for closes retained if to match creating
3127 * opens, so use req->rq_xid as a secondary key.
3128 * (See bugs 684, 685, and 428.)
3129 * XXX no longer needed, but all opens need transnos!
3131 if (iter->rq_transno > req->rq_transno)
3134 if (iter->rq_transno == req->rq_transno) {
3135 LASSERT(iter->rq_xid != req->rq_xid);
3136 if (iter->rq_xid > req->rq_xid)
3140 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3144 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3148 * Send request and wait until it completes.
3149 * Returns request processing status.
3151 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3153 struct ptlrpc_request_set *set;
3157 LASSERT(req->rq_set == NULL);
3158 LASSERT(!req->rq_receiving_reply);
3160 set = ptlrpc_prep_set();
3162 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3166 /* for distributed debugging */
3167 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3169 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3170 ptlrpc_request_addref(req);
3171 ptlrpc_set_add_req(set, req);
3172 rc = ptlrpc_set_wait(NULL, set);
3173 ptlrpc_set_destroy(set);
3177 EXPORT_SYMBOL(ptlrpc_queue_wait);
3180 * Callback used for replayed requests reply processing.
3181 * In case of successful reply calls registered request replay callback.
3182 * In case of error restart replay process.
3184 static int ptlrpc_replay_interpret(const struct lu_env *env,
3185 struct ptlrpc_request *req,
3188 struct ptlrpc_replay_async_args *aa = args;
3189 struct obd_import *imp = req->rq_import;
3192 atomic_dec(&imp->imp_replay_inflight);
3195 * Note: if it is bulk replay (MDS-MDS replay), then even if
3196 * server got the request, but bulk transfer timeout, let's
3197 * replay the bulk req again
3199 if (!ptlrpc_client_replied(req) ||
3201 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3202 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3203 GOTO(out, rc = -ETIMEDOUT);
3206 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3207 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3208 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3209 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3211 /** VBR: check version failure */
3212 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3213 /** replay was failed due to version mismatch */
3214 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3215 spin_lock(&imp->imp_lock);
3216 imp->imp_vbr_failed = 1;
3217 spin_unlock(&imp->imp_lock);
3218 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3220 /** The transno had better not change over replay. */
3221 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3222 lustre_msg_get_transno(req->rq_repmsg) ||
3223 lustre_msg_get_transno(req->rq_repmsg) == 0,
3225 lustre_msg_get_transno(req->rq_reqmsg),
3226 lustre_msg_get_transno(req->rq_repmsg));
3229 spin_lock(&imp->imp_lock);
3230 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3231 spin_unlock(&imp->imp_lock);
3232 LASSERT(imp->imp_last_replay_transno);
3234 /* transaction number shouldn't be bigger than the latest replayed */
3235 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3236 DEBUG_REQ(D_ERROR, req,
3237 "Reported transno=%llu is bigger than replayed=%llu",
3239 lustre_msg_get_transno(req->rq_reqmsg));
3240 GOTO(out, rc = -EINVAL);
3243 DEBUG_REQ(D_HA, req, "got reply");
3245 /* let the callback do fixups, possibly including in the request */
3246 if (req->rq_replay_cb)
3247 req->rq_replay_cb(req);
3249 if (ptlrpc_client_replied(req) &&
3250 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3251 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3252 lustre_msg_get_status(req->rq_repmsg),
3253 aa->praa_old_status);
3256 * Note: If the replay fails for MDT-MDT recovery, let's
3257 * abort all of the following requests in the replay
3258 * and sending list, because MDT-MDT update requests
3259 * are dependent on each other, see LU-7039
3261 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3262 struct ptlrpc_request *free_req;
3263 struct ptlrpc_request *tmp;
3265 spin_lock(&imp->imp_lock);
3266 list_for_each_entry_safe(free_req, tmp,
3267 &imp->imp_replay_list,
3269 ptlrpc_free_request(free_req);
3272 list_for_each_entry_safe(free_req, tmp,
3273 &imp->imp_committed_list,
3275 ptlrpc_free_request(free_req);
3278 list_for_each_entry_safe(free_req, tmp,
3279 &imp->imp_delayed_list,
3281 spin_lock(&free_req->rq_lock);
3282 free_req->rq_err = 1;
3283 free_req->rq_status = -EIO;
3284 ptlrpc_client_wake_req(free_req);
3285 spin_unlock(&free_req->rq_lock);
3288 list_for_each_entry_safe(free_req, tmp,
3289 &imp->imp_sending_list,
3291 spin_lock(&free_req->rq_lock);
3292 free_req->rq_err = 1;
3293 free_req->rq_status = -EIO;
3294 ptlrpc_client_wake_req(free_req);
3295 spin_unlock(&free_req->rq_lock);
3297 spin_unlock(&imp->imp_lock);
3300 /* Put it back for re-replay. */
3301 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3305 * Errors while replay can set transno to 0, but
3306 * imp_last_replay_transno shouldn't be set to 0 anyway
3308 if (req->rq_transno == 0)
3309 CERROR("Transno is 0 during replay!\n");
3311 /* continue with recovery */
3312 rc = ptlrpc_import_recovery_state_machine(imp);
3314 req->rq_send_state = aa->praa_old_state;
3317 /* this replay failed, so restart recovery */
3318 ptlrpc_connect_import(imp);
3324 * Prepares and queues request for replay.
3325 * Adds it to ptlrpcd queue for actual sending.
3326 * Returns 0 on success.
3328 int ptlrpc_replay_req(struct ptlrpc_request *req)
3330 struct ptlrpc_replay_async_args *aa;
3334 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3336 aa = ptlrpc_req_async_args(aa, req);
3337 memset(aa, 0, sizeof(*aa));
3339 /* Prepare request to be resent with ptlrpcd */
3340 aa->praa_old_state = req->rq_send_state;
3341 req->rq_send_state = LUSTRE_IMP_REPLAY;
3342 req->rq_phase = RQ_PHASE_NEW;
3343 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3345 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3347 req->rq_interpret_reply = ptlrpc_replay_interpret;
3348 /* Readjust the timeout for current conditions */
3349 ptlrpc_at_set_req_timeout(req);
3351 /* Tell server net_latency to calculate how long to wait for reply. */
3352 lustre_msg_set_service_timeout(req->rq_reqmsg,
3353 ptlrpc_at_get_net_latency(req));
3354 DEBUG_REQ(D_HA, req, "REPLAY");
3356 atomic_inc(&req->rq_import->imp_replay_inflight);
3357 spin_lock(&req->rq_lock);
3358 req->rq_early_free_repbuf = 0;
3359 spin_unlock(&req->rq_lock);
3360 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3362 ptlrpcd_add_req(req);
3367 * Aborts all in-flight request on import \a imp sending and delayed lists
3369 void ptlrpc_abort_inflight(struct obd_import *imp)
3371 struct ptlrpc_request *req;
3375 * Make sure that no new requests get processed for this import.
3376 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3377 * this flag and then putting requests on sending_list or delayed_list.
3379 assert_spin_locked(&imp->imp_lock);
3382 * XXX locking? Maybe we should remove each request with the list
3383 * locked? Also, how do we know if the requests on the list are
3384 * being freed at this time?
3386 list_for_each_entry(req, &imp->imp_sending_list, rq_list) {
3387 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3389 spin_lock(&req->rq_lock);
3390 if (req->rq_import_generation < imp->imp_generation) {
3392 req->rq_status = -EIO;
3393 ptlrpc_client_wake_req(req);
3395 spin_unlock(&req->rq_lock);
3398 list_for_each_entry(req, &imp->imp_delayed_list, rq_list) {
3399 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3401 spin_lock(&req->rq_lock);
3402 if (req->rq_import_generation < imp->imp_generation) {
3404 req->rq_status = -EIO;
3405 ptlrpc_client_wake_req(req);
3407 spin_unlock(&req->rq_lock);
3411 * Last chance to free reqs left on the replay list, but we
3412 * will still leak reqs that haven't committed.
3414 if (imp->imp_replayable)
3415 ptlrpc_free_committed(imp);
3421 * Abort all uncompleted requests in request set \a set
3423 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3425 struct ptlrpc_request *req;
3427 LASSERT(set != NULL);
3429 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
3430 spin_lock(&req->rq_lock);
3431 if (req->rq_phase != RQ_PHASE_RPC) {
3432 spin_unlock(&req->rq_lock);
3437 req->rq_status = -EINTR;
3438 ptlrpc_client_wake_req(req);
3439 spin_unlock(&req->rq_lock);
3444 * Initialize the XID for the node. This is common among all requests on
3445 * this node, and only requires the property that it is monotonically
3446 * increasing. It does not need to be sequential. Since this is also used
3447 * as the RDMA match bits, it is important that a single client NOT have
3448 * the same match bits for two different in-flight requests, hence we do
3449 * NOT want to have an XID per target or similar.
3451 * To avoid an unlikely collision between match bits after a client reboot
3452 * (which would deliver old data into the wrong RDMA buffer) initialize
3453 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3454 * If the time is clearly incorrect, we instead use a 62-bit random number.
3455 * In the worst case the random number will overflow 1M RPCs per second in
3456 * 9133 years, or permutations thereof.
3458 #define YEAR_2004 (1ULL << 30)
3459 void ptlrpc_init_xid(void)
3461 time64_t now = ktime_get_real_seconds();
3464 if (now < YEAR_2004) {
3465 get_random_bytes(&xid, sizeof(xid));
3467 xid |= (1ULL << 61);
3469 xid = (u64)now << 20;
3472 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3473 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3475 xid &= PTLRPC_BULK_OPS_MASK;
3476 atomic64_set(&ptlrpc_last_xid, xid);
3480 * Increase xid and returns resulting new value to the caller.
3482 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3483 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3484 * itself uses the last bulk xid needed, so the server can determine the
3485 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3486 * xid must align to a power-of-two value.
3488 * This is assumed to be true due to the initial ptlrpc_last_xid
3489 * value also being initialized to a power-of-two value. LU-1431
3491 __u64 ptlrpc_next_xid(void)
3493 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3497 * If request has a new allocated XID (new request or EINPROGRESS resend),
3498 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3499 * request to ensure previous bulk fails and avoid problems with lost replies
3500 * and therefore several transfers landing into the same buffer from different
3502 * Also, to avoid previous reply landing to a different sending attempt.
3504 void ptlrpc_set_mbits(struct ptlrpc_request *req)
3506 int md_count = req->rq_bulk ? req->rq_bulk->bd_md_count : 1;
3509 * Generate new matchbits for all resend requests, including
3512 if (req->rq_resend) {
3513 __u64 old_mbits = req->rq_mbits;
3516 * First time resend on -EINPROGRESS will generate new xid,
3517 * so we can actually use the rq_xid as rq_mbits in such case,
3518 * however, it's bit hard to distinguish such resend with a
3519 * 'resend for the -EINPROGRESS resend'. To make it simple,
3520 * we opt to generate mbits for all resend cases.
3522 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3524 req->rq_mbits = ptlrpc_next_xid();
3527 * Old version transfers rq_xid to peer as
3530 spin_lock(&req->rq_import->imp_lock);
3531 list_del_init(&req->rq_unreplied_list);
3532 ptlrpc_assign_next_xid_nolock(req);
3533 spin_unlock(&req->rq_import->imp_lock);
3534 req->rq_mbits = req->rq_xid;
3536 CDEBUG(D_HA, "resend with new mbits old x%llu new x%llu\n",
3537 old_mbits, req->rq_mbits);
3538 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3539 /* Request being sent first time, use xid as matchbits. */
3540 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3541 BULK_MBITS) || req->rq_mbits == 0)
3543 req->rq_mbits = req->rq_xid;
3545 req->rq_mbits -= md_count - 1;
3549 * Replay request, xid and matchbits have already been
3550 * correctly assigned.
3556 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3557 * that server can infer the number of bulks that were prepared,
3560 req->rq_mbits += md_count - 1;
3563 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3564 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3566 * It's ok to directly set the rq_xid here, since this xid bump
3567 * won't affect the request position in unreplied list.
3569 if (!OCD_HAS_FLAG(&req->rq_import->imp_connect_data, BULK_MBITS))
3570 req->rq_xid = req->rq_mbits;
3574 * Get a glimpse at what next xid value might have been.
3575 * Returns possible next xid.
3577 __u64 ptlrpc_sample_next_xid(void)
3579 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3581 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3584 * Functions for operating ptlrpc workers.
3586 * A ptlrpc work is a function which will be running inside ptlrpc context.
3587 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3589 * 1. after a work is created, it can be used many times, that is:
3590 * handler = ptlrpcd_alloc_work();
3591 * ptlrpcd_queue_work();
3593 * queue it again when necessary:
3594 * ptlrpcd_queue_work();
3595 * ptlrpcd_destroy_work();
3596 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3597 * it will only be queued once in any time. Also as its name implies, it may
3598 * have delay before it really runs by ptlrpcd thread.
3600 struct ptlrpc_work_async_args {
3601 int (*cb)(const struct lu_env *, void *);
3605 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3607 /* re-initialize the req */
3608 req->rq_timeout = obd_timeout;
3609 req->rq_sent = ktime_get_real_seconds();
3610 req->rq_deadline = req->rq_sent + req->rq_timeout;
3611 req->rq_phase = RQ_PHASE_INTERPRET;
3612 req->rq_next_phase = RQ_PHASE_COMPLETE;
3613 req->rq_xid = ptlrpc_next_xid();
3614 req->rq_import_generation = req->rq_import->imp_generation;
3616 ptlrpcd_add_req(req);
3619 static int work_interpreter(const struct lu_env *env,
3620 struct ptlrpc_request *req, void *args, int rc)
3622 struct ptlrpc_work_async_args *arg = args;
3624 LASSERT(ptlrpcd_check_work(req));
3625 LASSERT(arg->cb != NULL);
3627 rc = arg->cb(env, arg->cbdata);
3629 list_del_init(&req->rq_set_chain);
3632 if (atomic_dec_return(&req->rq_refcount) > 1) {
3633 atomic_set(&req->rq_refcount, 2);
3634 ptlrpcd_add_work_req(req);
3639 static int worker_format;
3641 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3643 return req->rq_pill.rc_fmt == (void *)&worker_format;
3647 * Create a work for ptlrpc.
3649 void *ptlrpcd_alloc_work(struct obd_import *imp,
3650 int (*cb)(const struct lu_env *, void *), void *cbdata)
3652 struct ptlrpc_request *req = NULL;
3653 struct ptlrpc_work_async_args *args;
3659 RETURN(ERR_PTR(-EINVAL));
3661 /* copy some code from deprecated fakereq. */
3662 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3664 CERROR("ptlrpc: run out of memory!\n");
3665 RETURN(ERR_PTR(-ENOMEM));
3668 ptlrpc_cli_req_init(req);
3670 req->rq_send_state = LUSTRE_IMP_FULL;
3671 req->rq_type = PTL_RPC_MSG_REQUEST;
3672 req->rq_import = class_import_get(imp);
3673 req->rq_interpret_reply = work_interpreter;
3674 /* don't want reply */
3675 req->rq_no_delay = req->rq_no_resend = 1;
3676 req->rq_pill.rc_fmt = (void *)&worker_format;
3678 args = ptlrpc_req_async_args(args, req);
3680 args->cbdata = cbdata;
3684 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3686 void ptlrpcd_destroy_work(void *handler)
3688 struct ptlrpc_request *req = handler;
3691 ptlrpc_req_finished(req);
3693 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3695 int ptlrpcd_queue_work(void *handler)
3697 struct ptlrpc_request *req = handler;
3700 * Check if the req is already being queued.
3702 * Here comes a trick: it lacks a way of checking if a req is being
3703 * processed reliably in ptlrpc. Here I have to use refcount of req
3704 * for this purpose. This is okay because the caller should use this
3705 * req as opaque data. - Jinshan
3707 LASSERT(atomic_read(&req->rq_refcount) > 0);
3708 if (atomic_inc_return(&req->rq_refcount) == 2)
3709 ptlrpcd_add_work_req(req);
3712 EXPORT_SYMBOL(ptlrpcd_queue_work);