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();
1173 if (req->rq_reqmsg) {
1174 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1175 lustre_msg_set_uid_gid(req->rq_reqmsg, NULL, NULL);
1178 if (set->set_producer)
1180 * If the request set has a producer callback, the RPC must be
1181 * sent straight away
1183 ptlrpc_send_new_req(req);
1185 EXPORT_SYMBOL(ptlrpc_set_add_req);
1188 * Add a request to a request with dedicated server thread
1189 * and wake the thread to make any necessary processing.
1190 * Currently only used for ptlrpcd.
1192 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1193 struct ptlrpc_request *req)
1195 struct ptlrpc_request_set *set = pc->pc_set;
1198 LASSERT(req->rq_set == NULL);
1199 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1201 spin_lock(&set->set_new_req_lock);
1203 * The set takes over the caller's request reference.
1206 req->rq_queued_time = ktime_get_seconds();
1207 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1208 count = atomic_inc_return(&set->set_new_count);
1209 spin_unlock(&set->set_new_req_lock);
1211 /* Only need to call wakeup once for the first entry. */
1213 wake_up(&set->set_waitq);
1216 * XXX: It maybe unnecessary to wakeup all the partners. But to
1217 * guarantee the async RPC can be processed ASAP, we have
1218 * no other better choice. It maybe fixed in future.
1220 for (i = 0; i < pc->pc_npartners; i++)
1221 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1226 * Based on the current state of the import, determine if the request
1227 * can be sent, is an error, or should be delayed.
1229 * Returns true if this request should be delayed. If false, and
1230 * *status is set, then the request can not be sent and *status is the
1231 * error code. If false and status is 0, then request can be sent.
1233 * The imp->imp_lock must be held.
1235 static int ptlrpc_import_delay_req(struct obd_import *imp,
1236 struct ptlrpc_request *req, int *status)
1244 if (req->rq_ctx_init || req->rq_ctx_fini) {
1245 /* always allow ctx init/fini rpc go through */
1246 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1247 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1249 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1250 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1253 * pings or MDS-equivalent STATFS may safely
1256 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1257 D_HA : D_ERROR, req, "IMP_CLOSED");
1259 } else if (ptlrpc_send_limit_expired(req)) {
1260 /* probably doesn't need to be a D_ERROR afterinitial testing */
1261 DEBUG_REQ(D_HA, req, "send limit expired");
1262 *status = -ETIMEDOUT;
1263 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1264 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1265 ;/* allow CONNECT even if import is invalid */
1266 if (atomic_read(&imp->imp_inval_count) != 0) {
1267 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1270 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1271 if (!imp->imp_deactive)
1272 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1273 *status = -ESHUTDOWN; /* b=12940 */
1274 } else if (req->rq_import_generation != imp->imp_generation) {
1275 DEBUG_REQ(req->rq_no_resend ? D_INFO : D_ERROR,
1276 req, "req wrong generation:");
1278 } else if (req->rq_send_state != imp->imp_state) {
1279 /* invalidate in progress - any requests should be drop */
1280 if (atomic_read(&imp->imp_inval_count) != 0) {
1281 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1283 } else if (req->rq_no_delay &&
1284 imp->imp_generation != imp->imp_initiated_at) {
1285 /* ignore nodelay for requests initiating connections */
1287 } else if (req->rq_allow_replay &&
1288 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1289 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1290 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1291 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1292 DEBUG_REQ(D_HA, req, "allow during recovery");
1302 * Decide if the error message should be printed to the console or not.
1303 * Makes its decision based on request type, status, and failure frequency.
1305 * \param[in] req request that failed and may need a console message
1307 * \retval false if no message should be printed
1308 * \retval true if console message should be printed
1310 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1312 LASSERT(req->rq_reqmsg != NULL);
1314 /* Suppress particular reconnect errors which are to be expected. */
1315 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1316 /* Suppress timed out reconnect requests */
1317 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1322 * Suppress most unavailable/again reconnect requests, but
1323 * print occasionally so it is clear client is trying to
1324 * connect to a server where no target is running.
1326 if ((err == -ENODEV || err == -EAGAIN) &&
1327 req->rq_import->imp_conn_cnt % 30 != 20)
1331 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1332 /* -EAGAIN is normal when using POSIX flocks */
1335 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1336 (req->rq_xid & 0xf) != 10)
1337 /* Suppress most ping requests, they may fail occasionally */
1344 * Check request processing status.
1345 * Returns the status.
1347 static int ptlrpc_check_status(struct ptlrpc_request *req)
1352 rc = lustre_msg_get_status(req->rq_repmsg);
1353 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1354 struct obd_import *imp = req->rq_import;
1355 struct lnet_nid *nid = &imp->imp_connection->c_peer.nid;
1356 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1358 if (ptlrpc_console_allow(req, opc, rc))
1359 LCONSOLE_ERROR_MSG(0x11,
1360 "%s: operation %s to node %s failed: rc = %d\n",
1361 imp->imp_obd->obd_name,
1363 libcfs_nidstr(nid), rc);
1364 RETURN(rc < 0 ? rc : -EINVAL);
1368 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1374 * save pre-versions of objects into request for replay.
1375 * Versions are obtained from server reply.
1378 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1380 struct lustre_msg *repmsg = req->rq_repmsg;
1381 struct lustre_msg *reqmsg = req->rq_reqmsg;
1382 __u64 *versions = lustre_msg_get_versions(repmsg);
1385 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1389 lustre_msg_set_versions(reqmsg, versions);
1390 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1391 versions[0], versions[1]);
1396 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1398 struct ptlrpc_request *req;
1400 assert_spin_locked(&imp->imp_lock);
1401 if (list_empty(&imp->imp_unreplied_list))
1404 req = list_first_entry(&imp->imp_unreplied_list, struct ptlrpc_request,
1406 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1408 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1409 imp->imp_known_replied_xid = req->rq_xid - 1;
1411 return req->rq_xid - 1;
1415 * Callback function called when client receives RPC reply for \a req.
1416 * Returns 0 on success or error code.
1417 * The return alue would be assigned to req->rq_status by the caller
1418 * as request processing status.
1419 * This function also decides if the request needs to be saved for later replay.
1421 static int after_reply(struct ptlrpc_request *req)
1423 struct obd_import *imp = req->rq_import;
1424 struct obd_device *obd = req->rq_import->imp_obd;
1431 LASSERT(obd != NULL);
1432 /* repbuf must be unlinked */
1433 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1435 if (req->rq_reply_truncated) {
1436 if (ptlrpc_no_resend(req)) {
1437 DEBUG_REQ(D_ERROR, req,
1438 "reply buffer overflow, expected=%d, actual size=%d",
1439 req->rq_nob_received, req->rq_repbuf_len);
1443 sptlrpc_cli_free_repbuf(req);
1445 * Pass the required reply buffer size (include
1446 * space for early reply).
1447 * NB: no need to roundup because alloc_repbuf
1450 req->rq_replen = req->rq_nob_received;
1451 req->rq_nob_received = 0;
1452 spin_lock(&req->rq_lock);
1454 spin_unlock(&req->rq_lock);
1458 work_start = ktime_get_real();
1459 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1462 * NB Until this point, the whole of the incoming message,
1463 * including buflens, status etc is in the sender's byte order.
1465 rc = sptlrpc_cli_unwrap_reply(req);
1467 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1472 * Security layer unwrap might ask resend this request.
1477 rc = unpack_reply(req);
1481 /* retry indefinitely on EINPROGRESS */
1482 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1483 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1484 time64_t now = ktime_get_real_seconds();
1486 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1487 D_RPCTRACE, req, "resending request on EINPROGRESS");
1488 spin_lock(&req->rq_lock);
1490 spin_unlock(&req->rq_lock);
1491 req->rq_nr_resend++;
1493 /* Readjust the timeout for current conditions */
1494 ptlrpc_at_set_req_timeout(req);
1496 * delay resend to give a chance to the server to get ready.
1497 * The delay is increased by 1s on every resend and is capped to
1498 * the current request timeout (i.e. obd_timeout if AT is off,
1499 * or AT service time x 125% + 5s, see at_est2timeout)
1501 if (req->rq_nr_resend > req->rq_timeout)
1502 req->rq_sent = now + req->rq_timeout;
1504 req->rq_sent = now + req->rq_nr_resend;
1506 /* Resend for EINPROGRESS will use a new XID */
1507 spin_lock(&imp->imp_lock);
1508 list_del_init(&req->rq_unreplied_list);
1509 spin_unlock(&imp->imp_lock);
1514 if (obd->obd_svc_stats) {
1515 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1517 ptlrpc_lprocfs_rpc_sent(req, timediff);
1520 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1521 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1522 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1523 lustre_msg_get_type(req->rq_repmsg));
1527 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1528 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1529 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1530 ptlrpc_at_adj_net_latency(req,
1531 lustre_msg_get_service_timeout(req->rq_repmsg));
1533 rc = ptlrpc_check_status(req);
1537 * Either we've been evicted, or the server has failed for
1538 * some reason. Try to reconnect, and if that fails, punt to
1541 if (ptlrpc_recoverable_error(rc)) {
1542 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1543 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1546 ptlrpc_request_handle_notconn(req);
1551 * Let's look if server sent slv. Do it only for RPC with
1554 ldlm_cli_update_pool(req);
1558 * Store transno in reqmsg for replay.
1560 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1561 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1562 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1565 if (imp->imp_replayable) {
1566 /* if other threads are waiting for ptlrpc_free_committed()
1567 * they could continue the work of freeing RPCs. That reduces
1568 * lock hold times, and distributes work more fairly across
1569 * waiting threads. We can't use spin_is_contended() since
1570 * there are many other places where imp_lock is held.
1572 atomic_inc(&imp->imp_waiting);
1573 spin_lock(&imp->imp_lock);
1574 atomic_dec(&imp->imp_waiting);
1576 * No point in adding already-committed requests to the replay
1577 * list, we will just remove them immediately. b=9829
1579 if (req->rq_transno != 0 &&
1581 lustre_msg_get_last_committed(req->rq_repmsg) ||
1583 /** version recovery */
1584 ptlrpc_save_versions(req);
1585 ptlrpc_retain_replayable_request(req, imp);
1586 } else if (req->rq_commit_cb &&
1587 list_empty(&req->rq_replay_list)) {
1589 * NB: don't call rq_commit_cb if it's already on
1590 * rq_replay_list, ptlrpc_free_committed() will call
1591 * it later, see LU-3618 for details
1593 spin_unlock(&imp->imp_lock);
1594 req->rq_commit_cb(req);
1595 atomic_inc(&imp->imp_waiting);
1596 spin_lock(&imp->imp_lock);
1597 atomic_dec(&imp->imp_waiting);
1601 * Replay-enabled imports return commit-status information.
1603 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1604 if (likely(committed > imp->imp_peer_committed_transno))
1605 imp->imp_peer_committed_transno = committed;
1607 ptlrpc_free_committed(imp);
1609 if (!list_empty(&imp->imp_replay_list)) {
1610 struct ptlrpc_request *last;
1612 last = list_entry(imp->imp_replay_list.prev,
1613 struct ptlrpc_request,
1616 * Requests with rq_replay stay on the list even if no
1617 * commit is expected.
1619 if (last->rq_transno > imp->imp_peer_committed_transno)
1620 ptlrpc_pinger_commit_expected(imp);
1623 spin_unlock(&imp->imp_lock);
1630 * Helper function to send request \a req over the network for the first time
1631 * Also adjusts request phase.
1632 * Returns 0 on success or error code.
1634 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1636 struct obd_import *imp = req->rq_import;
1641 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1643 /* do not try to go further if there is not enough memory in enc_pool */
1644 if (req->rq_sent && req->rq_bulk)
1645 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1646 pool_is_at_full_capacity())
1649 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1650 (!req->rq_generation_set ||
1651 req->rq_import_generation == imp->imp_generation))
1654 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1656 spin_lock(&imp->imp_lock);
1658 LASSERT(req->rq_xid != 0);
1659 LASSERT(!list_empty(&req->rq_unreplied_list));
1661 if (!req->rq_generation_set)
1662 req->rq_import_generation = imp->imp_generation;
1664 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1665 spin_lock(&req->rq_lock);
1666 req->rq_waiting = 1;
1667 spin_unlock(&req->rq_lock);
1669 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1670 ptlrpc_import_state_name(req->rq_send_state),
1671 ptlrpc_import_state_name(imp->imp_state));
1672 LASSERT(list_empty(&req->rq_list));
1673 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1674 atomic_inc(&req->rq_import->imp_inflight);
1675 spin_unlock(&imp->imp_lock);
1680 spin_unlock(&imp->imp_lock);
1681 req->rq_status = rc;
1682 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1686 LASSERT(list_empty(&req->rq_list));
1687 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1688 atomic_inc(&req->rq_import->imp_inflight);
1691 * find the known replied XID from the unreplied list, CONNECT
1692 * and DISCONNECT requests are skipped to make the sanity check
1693 * on server side happy. see process_req_last_xid().
1695 * For CONNECT: Because replay requests have lower XID, it'll
1696 * break the sanity check if CONNECT bump the exp_last_xid on
1699 * For DISCONNECT: Since client will abort inflight RPC before
1700 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1701 * than the inflight RPC.
1703 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1704 min_xid = ptlrpc_known_replied_xid(imp);
1705 spin_unlock(&imp->imp_lock);
1707 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1709 lustre_msg_set_status(req->rq_reqmsg, current->pid);
1711 /* If the request to be sent is an LDLM callback, do not try to
1713 * An LDLM callback is sent by a server to a client in order to make
1714 * it release a lock, on a communication channel that uses a reverse
1715 * context. It cannot be refreshed on its own, as it is the 'reverse'
1716 * (server-side) representation of a client context.
1717 * We do not care if the reverse context is expired, and want to send
1718 * the LDLM callback anyway. Once the client receives the AST, it is
1719 * its job to refresh its own context if it has expired, hence
1720 * refreshing the associated reverse context on server side, before
1721 * being able to send the LDLM_CANCEL requested by the server.
1723 if (lustre_msg_get_opc(req->rq_reqmsg) != LDLM_BL_CALLBACK &&
1724 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_CP_CALLBACK &&
1725 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_GL_CALLBACK)
1726 rc = sptlrpc_req_refresh_ctx(req, 0);
1729 req->rq_status = rc;
1732 spin_lock(&req->rq_lock);
1733 req->rq_wait_ctx = 1;
1734 spin_unlock(&req->rq_lock);
1740 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1742 imp->imp_obd->obd_uuid.uuid,
1743 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1744 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1745 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1747 rc = ptl_send_rpc(req, 0);
1748 if (rc == -ENOMEM) {
1749 spin_lock(&imp->imp_lock);
1750 if (!list_empty(&req->rq_list)) {
1751 list_del_init(&req->rq_list);
1752 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1753 wake_up(&req->rq_import->imp_recovery_waitq);
1755 spin_unlock(&imp->imp_lock);
1756 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1760 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1762 spin_lock(&req->rq_lock);
1763 req->rq_net_err = 1;
1764 spin_unlock(&req->rq_lock);
1770 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1775 LASSERT(set->set_producer != NULL);
1777 remaining = atomic_read(&set->set_remaining);
1780 * populate the ->set_requests list with requests until we
1781 * reach the maximum number of RPCs in flight for this set
1783 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1784 rc = set->set_producer(set, set->set_producer_arg);
1785 if (rc == -ENOENT) {
1786 /* no more RPC to produce */
1787 set->set_producer = NULL;
1788 set->set_producer_arg = NULL;
1793 RETURN((atomic_read(&set->set_remaining) - remaining));
1797 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1798 * and no more replies are expected.
1799 * (it is possible to get less replies than requests sent e.g. due to timed out
1800 * requests or requests that we had trouble to send out)
1802 * NOTE: This function contains a potential schedule point (cond_resched()).
1804 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1806 struct ptlrpc_request *req, *next;
1807 LIST_HEAD(comp_reqs);
1808 int force_timer_recalc = 0;
1811 if (atomic_read(&set->set_remaining) == 0)
1814 list_for_each_entry_safe(req, next, &set->set_requests,
1816 struct obd_import *imp = req->rq_import;
1817 int unregistered = 0;
1821 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1822 list_move_tail(&req->rq_set_chain, &comp_reqs);
1827 * This schedule point is mainly for the ptlrpcd caller of this
1828 * function. Most ptlrpc sets are not long-lived and unbounded
1829 * in length, but at the least the set used by the ptlrpcd is.
1830 * Since the processing time is unbounded, we need to insert an
1831 * explicit schedule point to make the thread well-behaved.
1836 * If the caller requires to allow to be interpreted by force
1837 * and it has really been interpreted, then move the request
1838 * to RQ_PHASE_INTERPRET phase in spite of what the current
1841 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1842 req->rq_status = -EINTR;
1843 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1846 * Since it is interpreted and we have to wait for
1847 * the reply to be unlinked, then use sync mode.
1851 GOTO(interpret, req->rq_status);
1854 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1855 force_timer_recalc = 1;
1857 /* delayed send - skip */
1858 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1861 /* delayed resend - skip */
1862 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1863 req->rq_sent > ktime_get_real_seconds())
1866 if (!(req->rq_phase == RQ_PHASE_RPC ||
1867 req->rq_phase == RQ_PHASE_BULK ||
1868 req->rq_phase == RQ_PHASE_INTERPRET ||
1869 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1870 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1871 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1875 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1876 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1877 LASSERT(req->rq_next_phase != req->rq_phase);
1878 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1880 if (req->rq_req_deadline &&
1881 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1882 req->rq_req_deadline = 0;
1883 if (req->rq_reply_deadline &&
1884 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1885 req->rq_reply_deadline = 0;
1886 if (req->rq_bulk_deadline &&
1887 !OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1888 req->rq_bulk_deadline = 0;
1891 * Skip processing until reply is unlinked. We
1892 * can't return to pool before that and we can't
1893 * call interpret before that. We need to make
1894 * sure that all rdma transfers finished and will
1895 * not corrupt any data.
1897 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1898 ptlrpc_cli_wait_unlink(req))
1900 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1901 ptlrpc_client_bulk_active(req))
1905 * Turn fail_loc off to prevent it from looping
1908 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1909 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1912 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1913 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1918 * Move to next phase if reply was successfully
1921 ptlrpc_rqphase_move(req, req->rq_next_phase);
1924 if (req->rq_phase == RQ_PHASE_INTERPRET)
1925 GOTO(interpret, req->rq_status);
1928 * Note that this also will start async reply unlink.
1930 if (req->rq_net_err && !req->rq_timedout) {
1931 ptlrpc_expire_one_request(req, 1);
1934 * Check if we still need to wait for unlink.
1936 if (ptlrpc_cli_wait_unlink(req) ||
1937 ptlrpc_client_bulk_active(req))
1939 /* If there is no need to resend, fail it now. */
1940 if (req->rq_no_resend) {
1941 if (req->rq_status == 0)
1942 req->rq_status = -EIO;
1943 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1944 GOTO(interpret, req->rq_status);
1951 if (!ptlrpc_unregister_reply(req, 1)) {
1952 ptlrpc_unregister_bulk(req, 1);
1956 spin_lock(&req->rq_lock);
1957 req->rq_replied = 0;
1958 spin_unlock(&req->rq_lock);
1959 if (req->rq_status == 0)
1960 req->rq_status = -EIO;
1961 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1962 GOTO(interpret, req->rq_status);
1966 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1967 * so it sets rq_intr regardless of individual rpc
1968 * timeouts. The synchronous IO waiting path sets
1969 * rq_intr irrespective of whether ptlrpcd
1970 * has seen a timeout. Our policy is to only interpret
1971 * interrupted rpcs after they have timed out, so we
1972 * need to enforce that here.
1975 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1976 req->rq_wait_ctx)) {
1977 req->rq_status = -EINTR;
1978 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1979 GOTO(interpret, req->rq_status);
1982 if (req->rq_phase == RQ_PHASE_RPC) {
1983 if (req->rq_timedout || req->rq_resend ||
1984 req->rq_waiting || req->rq_wait_ctx) {
1987 if (!ptlrpc_unregister_reply(req, 1)) {
1988 ptlrpc_unregister_bulk(req, 1);
1992 spin_lock(&imp->imp_lock);
1993 if (ptlrpc_import_delay_req(imp, req,
1996 * put on delay list - only if we wait
1997 * recovery finished - before send
1999 list_move_tail(&req->rq_list,
2000 &imp->imp_delayed_list);
2001 spin_unlock(&imp->imp_lock);
2006 req->rq_status = status;
2007 ptlrpc_rqphase_move(req,
2008 RQ_PHASE_INTERPRET);
2009 spin_unlock(&imp->imp_lock);
2010 GOTO(interpret, req->rq_status);
2012 /* ignore on just initiated connections */
2013 if (ptlrpc_no_resend(req) &&
2014 !req->rq_wait_ctx &&
2015 imp->imp_generation !=
2016 imp->imp_initiated_at) {
2017 req->rq_status = -ENOTCONN;
2018 ptlrpc_rqphase_move(req,
2019 RQ_PHASE_INTERPRET);
2020 spin_unlock(&imp->imp_lock);
2021 GOTO(interpret, req->rq_status);
2024 /* don't resend too fast in case of network
2027 if (ktime_get_real_seconds() < (req->rq_sent + 1)
2028 && req->rq_net_err && req->rq_timedout) {
2030 DEBUG_REQ(D_INFO, req,
2031 "throttle request");
2032 /* Don't try to resend RPC right away
2033 * as it is likely it will fail again
2034 * and ptlrpc_check_set() will be
2035 * called again, keeping this thread
2036 * busy. Instead, wait for the next
2037 * timeout. Flag it as resend to
2038 * ensure we don't wait to long.
2041 spin_unlock(&imp->imp_lock);
2045 list_move_tail(&req->rq_list,
2046 &imp->imp_sending_list);
2048 spin_unlock(&imp->imp_lock);
2050 spin_lock(&req->rq_lock);
2051 req->rq_waiting = 0;
2052 spin_unlock(&req->rq_lock);
2054 if (req->rq_timedout || req->rq_resend) {
2056 * This is re-sending anyways,
2057 * let's mark req as resend.
2059 spin_lock(&req->rq_lock);
2061 spin_unlock(&req->rq_lock);
2064 * rq_wait_ctx is only touched by ptlrpcd,
2065 * so no lock is needed here.
2067 status = sptlrpc_req_refresh_ctx(req, 0);
2070 req->rq_status = status;
2071 spin_lock(&req->rq_lock);
2072 req->rq_wait_ctx = 0;
2073 spin_unlock(&req->rq_lock);
2074 force_timer_recalc = 1;
2076 spin_lock(&req->rq_lock);
2077 req->rq_wait_ctx = 1;
2078 spin_unlock(&req->rq_lock);
2083 spin_lock(&req->rq_lock);
2084 req->rq_wait_ctx = 0;
2085 spin_unlock(&req->rq_lock);
2089 * In any case, the previous bulk should be
2090 * cleaned up to prepare for the new sending
2093 !ptlrpc_unregister_bulk(req, 1))
2096 rc = ptl_send_rpc(req, 0);
2097 if (rc == -ENOMEM) {
2098 spin_lock(&imp->imp_lock);
2099 if (!list_empty(&req->rq_list))
2100 list_del_init(&req->rq_list);
2101 spin_unlock(&imp->imp_lock);
2102 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2106 DEBUG_REQ(D_HA, req,
2107 "send failed: rc = %d", rc);
2108 force_timer_recalc = 1;
2109 spin_lock(&req->rq_lock);
2110 req->rq_net_err = 1;
2111 spin_unlock(&req->rq_lock);
2114 /* need to reset the timeout */
2115 force_timer_recalc = 1;
2118 spin_lock(&req->rq_lock);
2120 if (ptlrpc_client_early(req)) {
2121 ptlrpc_at_recv_early_reply(req);
2122 spin_unlock(&req->rq_lock);
2126 /* Still waiting for a reply? */
2127 if (ptlrpc_client_recv(req)) {
2128 spin_unlock(&req->rq_lock);
2132 /* Did we actually receive a reply? */
2133 if (!ptlrpc_client_replied(req)) {
2134 spin_unlock(&req->rq_lock);
2138 spin_unlock(&req->rq_lock);
2141 * unlink from net because we are going to
2142 * swab in-place of reply buffer
2144 unregistered = ptlrpc_unregister_reply(req, 1);
2148 req->rq_status = after_reply(req);
2149 if (req->rq_resend) {
2150 force_timer_recalc = 1;
2155 * If there is no bulk associated with this request,
2156 * then we're done and should let the interpreter
2157 * process the reply. Similarly if the RPC returned
2158 * an error, and therefore the bulk will never arrive.
2160 if (!req->rq_bulk || req->rq_status < 0) {
2161 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2162 GOTO(interpret, req->rq_status);
2165 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2168 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2169 if (ptlrpc_client_bulk_active(req))
2172 if (req->rq_bulk->bd_failure) {
2174 * The RPC reply arrived OK, but the bulk screwed
2175 * up! Dead weird since the server told us the RPC
2176 * was good after getting the REPLY for her GET or
2177 * the ACK for her PUT.
2179 DEBUG_REQ(D_ERROR, req, "bulk transfer failed %d/%d/%d",
2181 req->rq_bulk->bd_nob,
2182 req->rq_bulk->bd_nob_transferred);
2183 req->rq_status = -EIO;
2186 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2189 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2192 * This moves to "unregistering" phase we need to wait for
2195 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2196 /* start async bulk unlink too */
2197 ptlrpc_unregister_bulk(req, 1);
2201 if (!ptlrpc_unregister_bulk(req, async))
2205 * When calling interpret receiving already should be
2208 LASSERT(!req->rq_receiving_reply);
2210 ptlrpc_req_interpret(env, req, req->rq_status);
2212 if (ptlrpcd_check_work(req)) {
2213 atomic_dec(&set->set_remaining);
2216 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2220 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2222 imp->imp_obd->obd_uuid.uuid,
2223 lustre_msg_get_status(req->rq_reqmsg),
2225 obd_import_nid2str(imp),
2226 lustre_msg_get_opc(req->rq_reqmsg),
2227 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2229 spin_lock(&imp->imp_lock);
2231 * Request already may be not on sending or delaying list. This
2232 * may happen in the case of marking it erroneous for the case
2233 * ptlrpc_import_delay_req(req, status) find it impossible to
2234 * allow sending this rpc and returns *status != 0.
2236 if (!list_empty(&req->rq_list)) {
2237 list_del_init(&req->rq_list);
2238 if (atomic_dec_and_test(&imp->imp_inflight))
2239 wake_up(&imp->imp_recovery_waitq);
2241 list_del_init(&req->rq_unreplied_list);
2242 spin_unlock(&imp->imp_lock);
2244 atomic_dec(&set->set_remaining);
2245 wake_up(&imp->imp_recovery_waitq);
2247 if (set->set_producer) {
2248 /* produce a new request if possible */
2249 if (ptlrpc_set_producer(set) > 0)
2250 force_timer_recalc = 1;
2253 * free the request that has just been completed
2254 * in order not to pollute set->set_requests
2256 list_del_init(&req->rq_set_chain);
2257 spin_lock(&req->rq_lock);
2259 req->rq_invalid_rqset = 0;
2260 spin_unlock(&req->rq_lock);
2262 /* record rq_status to compute the final status later */
2263 if (req->rq_status != 0)
2264 set->set_rc = req->rq_status;
2265 ptlrpc_req_finished(req);
2267 list_move_tail(&req->rq_set_chain, &comp_reqs);
2272 * move completed request at the head of list so it's easier for
2273 * caller to find them
2275 list_splice(&comp_reqs, &set->set_requests);
2277 /* If we hit an error, we want to recover promptly. */
2278 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2280 EXPORT_SYMBOL(ptlrpc_check_set);
2283 * Time out request \a req. is \a async_unlink is set, that means do not wait
2284 * until LNet actually confirms network buffer unlinking.
2285 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2287 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2289 struct obd_import *imp = req->rq_import;
2290 unsigned int debug_mask = D_RPCTRACE;
2294 spin_lock(&req->rq_lock);
2295 req->rq_timedout = 1;
2296 spin_unlock(&req->rq_lock);
2298 if (ptlrpc_console_allow(req, lustre_msg_get_opc(req->rq_reqmsg),
2299 lustre_msg_get_status(req->rq_reqmsg)))
2300 debug_mask = D_WARNING;
2301 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2302 req->rq_net_err ? "failed due to network error" :
2303 ((req->rq_real_sent == 0 ||
2304 req->rq_real_sent < req->rq_sent ||
2305 req->rq_real_sent >= req->rq_deadline) ?
2306 "timed out for sent delay" : "timed out for slow reply"),
2307 req->rq_sent, req->rq_real_sent);
2309 if (imp && obd_debug_peer_on_timeout)
2310 LNetDebugPeer(&imp->imp_connection->c_peer);
2312 ptlrpc_unregister_reply(req, async_unlink);
2313 ptlrpc_unregister_bulk(req, async_unlink);
2315 if (obd_dump_on_timeout)
2316 libcfs_debug_dumplog();
2319 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2323 atomic_inc(&imp->imp_timeouts);
2325 /* The DLM server doesn't want recovery run on its imports. */
2326 if (imp->imp_dlm_fake)
2330 * If this request is for recovery or other primordial tasks,
2331 * then error it out here.
2333 if (req->rq_ctx_init || req->rq_ctx_fini ||
2334 req->rq_send_state != LUSTRE_IMP_FULL ||
2335 imp->imp_obd->obd_no_recov) {
2336 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2337 ptlrpc_import_state_name(req->rq_send_state),
2338 ptlrpc_import_state_name(imp->imp_state));
2339 spin_lock(&req->rq_lock);
2340 req->rq_status = -ETIMEDOUT;
2342 spin_unlock(&req->rq_lock);
2347 * if a request can't be resent we can't wait for an answer after
2350 if (ptlrpc_no_resend(req)) {
2351 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2355 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2361 * Time out all uncompleted requests in request set pointed by \a data
2362 * This is called when a wait times out.
2364 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2366 struct ptlrpc_request *req;
2367 time64_t now = ktime_get_real_seconds();
2370 LASSERT(set != NULL);
2373 * A timeout expired. See which reqs it applies to...
2375 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2376 /* don't expire request waiting for context */
2377 if (req->rq_wait_ctx)
2380 /* Request in-flight? */
2381 if (!((req->rq_phase == RQ_PHASE_RPC &&
2382 !req->rq_waiting && !req->rq_resend) ||
2383 (req->rq_phase == RQ_PHASE_BULK)))
2386 if (req->rq_timedout || /* already dealt with */
2387 req->rq_deadline > now) /* not expired */
2391 * Deal with this guy. Do it asynchronously to not block
2394 ptlrpc_expire_one_request(req, 1);
2396 * Loops require that we resched once in a while to avoid
2397 * RCU stalls and a few other problems.
2405 * Interrupts (sets interrupted flag) all uncompleted requests in
2406 * a set \a data. This is called when a wait_event is interrupted
2409 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2411 struct ptlrpc_request *req;
2413 LASSERT(set != NULL);
2414 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2416 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2420 if (req->rq_phase != RQ_PHASE_RPC &&
2421 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2422 !req->rq_allow_intr)
2425 spin_lock(&req->rq_lock);
2427 spin_unlock(&req->rq_lock);
2432 * Get the smallest timeout in the set; this does NOT set a timeout.
2434 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2436 time64_t now = ktime_get_real_seconds();
2438 struct ptlrpc_request *req;
2442 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2443 /* Request in-flight? */
2444 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2445 (req->rq_phase == RQ_PHASE_BULK) ||
2446 (req->rq_phase == RQ_PHASE_NEW)))
2449 /* Already timed out. */
2450 if (req->rq_timedout)
2453 /* Waiting for ctx. */
2454 if (req->rq_wait_ctx)
2457 if (req->rq_phase == RQ_PHASE_NEW)
2458 deadline = req->rq_sent;
2459 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2460 deadline = req->rq_sent;
2462 deadline = req->rq_sent + req->rq_timeout;
2464 if (deadline <= now) /* actually expired already */
2465 timeout = 1; /* ASAP */
2466 else if (timeout == 0 || timeout > deadline - now)
2467 timeout = deadline - now;
2473 * Send all unset request from the set and then wait untill all
2474 * requests in the set complete (either get a reply, timeout, get an
2475 * error or otherwise be interrupted).
2476 * Returns 0 on success or error code otherwise.
2478 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2480 struct ptlrpc_request *req;
2485 if (set->set_producer)
2486 (void)ptlrpc_set_producer(set);
2488 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2489 if (req->rq_phase == RQ_PHASE_NEW)
2490 (void)ptlrpc_send_new_req(req);
2493 if (list_empty(&set->set_requests))
2497 timeout = ptlrpc_set_next_timeout(set);
2500 * wait until all complete, interrupted, or an in-flight
2503 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2506 if ((timeout == 0 && !signal_pending(current)) ||
2507 set->set_allow_intr) {
2509 * No requests are in-flight (ether timed out
2510 * or delayed), so we can allow interrupts.
2511 * We still want to block for a limited time,
2512 * so we allow interrupts during the timeout.
2514 rc = l_wait_event_abortable_timeout(
2516 ptlrpc_check_set(NULL, set),
2517 cfs_time_seconds(timeout ? timeout : 1));
2520 ptlrpc_expired_set(set);
2521 } else if (rc < 0) {
2523 ptlrpc_interrupted_set(set);
2529 * At least one request is in flight, so no
2530 * interrupts are allowed. Wait until all
2531 * complete, or an in-flight req times out.
2533 rc = wait_event_idle_timeout(
2535 ptlrpc_check_set(NULL, set),
2536 cfs_time_seconds(timeout ? timeout : 1));
2538 ptlrpc_expired_set(set);
2545 * LU-769 - if we ignored the signal because
2546 * it was already pending when we started, we
2547 * need to handle it now or we risk it being
2550 if (rc == -ETIMEDOUT &&
2551 signal_pending(current)) {
2554 siginitset(&new, LUSTRE_FATAL_SIGS);
2555 sigprocmask(SIG_BLOCK, &new, &old);
2557 * In fact we only interrupt for the
2558 * "fatal" signals like SIGINT or
2559 * SIGKILL. We still ignore less
2560 * important signals since ptlrpc set
2561 * is not easily reentrant from
2564 if (signal_pending(current))
2565 ptlrpc_interrupted_set(set);
2566 sigprocmask(SIG_SETMASK, &old, NULL);
2570 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2573 * -EINTR => all requests have been flagged rq_intr so next
2575 * -ETIMEDOUT => someone timed out. When all reqs have
2576 * timed out, signals are enabled allowing completion with
2578 * I don't really care if we go once more round the loop in
2579 * the error cases -eeb.
2581 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2582 list_for_each_entry(req, &set->set_requests,
2584 spin_lock(&req->rq_lock);
2585 req->rq_invalid_rqset = 1;
2586 spin_unlock(&req->rq_lock);
2589 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2591 LASSERT(atomic_read(&set->set_remaining) == 0);
2593 rc = set->set_rc; /* rq_status of already freed requests if any */
2594 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2595 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2596 if (req->rq_status != 0)
2597 rc = req->rq_status;
2602 EXPORT_SYMBOL(ptlrpc_set_wait);
2605 * Helper fuction for request freeing.
2606 * Called when request count reached zero and request needs to be freed.
2607 * Removes request from all sorts of sending/replay lists it might be on,
2608 * frees network buffers if any are present.
2609 * If \a locked is set, that means caller is already holding import imp_lock
2610 * and so we no longer need to reobtain it (for certain lists manipulations)
2612 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2619 LASSERT(!request->rq_srv_req);
2620 LASSERT(request->rq_export == NULL);
2621 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2622 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2623 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2624 LASSERTF(!request->rq_replay, "req %p\n", request);
2626 req_capsule_fini(&request->rq_pill);
2629 * We must take it off the imp_replay_list first. Otherwise, we'll set
2630 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2632 if (request->rq_import) {
2634 spin_lock(&request->rq_import->imp_lock);
2635 list_del_init(&request->rq_replay_list);
2636 list_del_init(&request->rq_unreplied_list);
2638 spin_unlock(&request->rq_import->imp_lock);
2640 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2642 if (atomic_read(&request->rq_refcount) != 0) {
2643 DEBUG_REQ(D_ERROR, request,
2644 "freeing request with nonzero refcount");
2648 if (request->rq_repbuf)
2649 sptlrpc_cli_free_repbuf(request);
2651 if (request->rq_import) {
2652 if (!ptlrpcd_check_work(request)) {
2653 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2654 atomic_dec(&request->rq_import->imp_reqs);
2656 class_import_put(request->rq_import);
2657 request->rq_import = NULL;
2659 if (request->rq_bulk)
2660 ptlrpc_free_bulk(request->rq_bulk);
2662 if (request->rq_reqbuf || request->rq_clrbuf)
2663 sptlrpc_cli_free_reqbuf(request);
2665 if (request->rq_cli_ctx)
2666 sptlrpc_req_put_ctx(request, !locked);
2668 if (request->rq_pool)
2669 __ptlrpc_free_req_to_pool(request);
2671 ptlrpc_request_cache_free(request);
2675 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2677 * Drop one request reference. Must be called with import imp_lock held.
2678 * When reference count drops to zero, request is freed.
2680 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2682 assert_spin_locked(&request->rq_import->imp_lock);
2683 (void)__ptlrpc_req_finished(request, 1);
2688 * Drops one reference count for request \a request.
2689 * \a locked set indicates that caller holds import imp_lock.
2690 * Frees the request whe reference count reaches zero.
2692 * \retval 1 the request is freed
2693 * \retval 0 some others still hold references on the request
2695 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2703 LASSERT(request != LP_POISON);
2704 LASSERT(request->rq_reqmsg != LP_POISON);
2706 DEBUG_REQ(D_INFO, request, "refcount now %u",
2707 atomic_read(&request->rq_refcount) - 1);
2709 spin_lock(&request->rq_lock);
2710 count = atomic_dec_return(&request->rq_refcount);
2711 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2714 * For open RPC, the client does not know the EA size (LOV, ACL, and
2715 * so on) before replied, then the client has to reserve very large
2716 * reply buffer. Such buffer will not be released until the RPC freed.
2717 * Since The open RPC is replayable, we need to keep it in the replay
2718 * list until close. If there are a lot of files opened concurrently,
2719 * then the client may be OOM.
2721 * If fact, it is unnecessary to keep reply buffer for open replay,
2722 * related EAs have already been saved via mdc_save_lovea() before
2723 * coming here. So it is safe to free the reply buffer some earlier
2724 * before releasing the RPC to avoid client OOM. LU-9514
2726 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2727 spin_lock(&request->rq_early_free_lock);
2728 sptlrpc_cli_free_repbuf(request);
2729 request->rq_repbuf = NULL;
2730 request->rq_repbuf_len = 0;
2731 request->rq_repdata = NULL;
2732 request->rq_reqdata_len = 0;
2733 spin_unlock(&request->rq_early_free_lock);
2735 spin_unlock(&request->rq_lock);
2738 __ptlrpc_free_req(request, locked);
2744 * Drops one reference count for a request.
2746 void ptlrpc_req_finished(struct ptlrpc_request *request)
2748 __ptlrpc_req_finished(request, 0);
2750 EXPORT_SYMBOL(ptlrpc_req_finished);
2753 * Returns xid of a \a request
2755 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2757 return request->rq_xid;
2759 EXPORT_SYMBOL(ptlrpc_req_xid);
2762 * Disengage the client's reply buffer from the network
2763 * NB does _NOT_ unregister any client-side bulk.
2764 * IDEMPOTENT, but _not_ safe against concurrent callers.
2765 * The request owner (i.e. the thread doing the I/O) must call...
2766 * Returns 0 on success or 1 if unregistering cannot be made.
2768 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2770 bool discard = false;
2774 LASSERT(!in_interrupt());
2776 /* Let's setup deadline for reply unlink. */
2777 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2778 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2779 request->rq_reply_deadline = ktime_get_real_seconds() +
2780 PTLRPC_REQ_LONG_UNLINK;
2783 * Nothing left to do.
2785 if (!__ptlrpc_cli_wait_unlink(request, &discard))
2788 LNetMDUnlink(request->rq_reply_md_h);
2790 if (discard) /* Discard the request-out callback */
2791 __LNetMDUnlink(request->rq_req_md_h, discard);
2794 * Let's check it once again.
2796 if (!ptlrpc_cli_wait_unlink(request))
2799 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2800 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2803 * Do not wait for unlink to finish.
2809 * We have to wait_event_idle_timeout() whatever the result, to get
2810 * a chance to run reply_in_callback(), and to make sure we've
2811 * unlinked before returning a req to the pool.
2814 wait_queue_head_t *wq = (request->rq_set) ?
2815 &request->rq_set->set_waitq :
2816 &request->rq_reply_waitq;
2817 int seconds = PTLRPC_REQ_LONG_UNLINK;
2819 * Network access will complete in finite time but the HUGE
2820 * timeout lets us CWARN for visibility of sluggish NALs
2822 while (seconds > 0 &&
2823 wait_event_idle_timeout(
2825 !ptlrpc_cli_wait_unlink(request),
2826 cfs_time_seconds(1)) == 0)
2829 ptlrpc_rqphase_move(request, request->rq_next_phase);
2833 DEBUG_REQ(D_WARNING, request,
2834 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2835 request->rq_receiving_reply,
2836 request->rq_req_unlinked,
2837 request->rq_reply_unlinked);
2842 static void ptlrpc_free_request(struct ptlrpc_request *req)
2844 spin_lock(&req->rq_lock);
2846 spin_unlock(&req->rq_lock);
2848 if (req->rq_commit_cb)
2849 req->rq_commit_cb(req);
2850 list_del_init(&req->rq_replay_list);
2852 __ptlrpc_req_finished(req, 1);
2856 * the request is committed and dropped from the replay list of its import
2858 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2860 struct obd_import *imp = req->rq_import;
2862 spin_lock(&imp->imp_lock);
2863 if (list_empty(&req->rq_replay_list)) {
2864 spin_unlock(&imp->imp_lock);
2868 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2869 if (imp->imp_replay_cursor == &req->rq_replay_list)
2870 imp->imp_replay_cursor = req->rq_replay_list.next;
2871 ptlrpc_free_request(req);
2874 spin_unlock(&imp->imp_lock);
2876 EXPORT_SYMBOL(ptlrpc_request_committed);
2879 * Iterates through replay_list on import and prunes
2880 * all requests have transno smaller than last_committed for the
2881 * import and don't have rq_replay set.
2882 * Since requests are sorted in transno order, stops when meeting first
2883 * transno bigger than last_committed.
2884 * caller must hold imp->imp_lock
2886 void ptlrpc_free_committed(struct obd_import *imp)
2888 struct ptlrpc_request *req, *saved;
2889 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2890 bool skip_committed_list = true;
2891 unsigned int replay_scanned = 0, replay_freed = 0;
2892 unsigned int commit_scanned = 0, commit_freed = 0;
2893 unsigned int debug_level = D_INFO;
2894 __u64 peer_committed_transno;
2896 time64_t start, now;
2899 LASSERT(imp != NULL);
2900 assert_spin_locked(&imp->imp_lock);
2902 start = ktime_get_seconds();
2903 /* save these here, we can potentially drop imp_lock after checking */
2904 peer_committed_transno = imp->imp_peer_committed_transno;
2905 imp_generation = imp->imp_generation;
2907 if (peer_committed_transno == imp->imp_last_transno_checked &&
2908 imp_generation == imp->imp_last_generation_checked) {
2909 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2910 imp->imp_obd->obd_name, peer_committed_transno);
2913 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2914 imp->imp_obd->obd_name, peer_committed_transno, imp_generation);
2916 if (imp_generation != imp->imp_last_generation_checked ||
2917 imp->imp_last_transno_checked == 0)
2918 skip_committed_list = false;
2919 /* maybe drop imp_lock here, if another lock protected the lists */
2921 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2923 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2924 LASSERT(req != last_req);
2927 if (req->rq_transno == 0) {
2928 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2932 /* If other threads are waiting on imp_lock, stop processing
2933 * in this thread. Another thread can finish remaining work.
2934 * This may happen if there are huge numbers of open files
2935 * that are closed suddenly or evicted, or if the server
2936 * commit interval is very high vs. RPC rate.
2938 if (++replay_scanned % 2048 == 0) {
2939 now = ktime_get_seconds();
2940 if (now > start + 5)
2941 debug_level = D_WARNING;
2943 if ((replay_freed > 128 && now > start + 3) &&
2944 atomic_read(&imp->imp_waiting)) {
2945 if (debug_level == D_INFO)
2946 debug_level = D_RPCTRACE;
2951 if (req->rq_import_generation < imp_generation) {
2952 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2956 /* not yet committed */
2957 if (req->rq_transno > peer_committed_transno) {
2958 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2962 if (req->rq_replay) {
2963 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2964 list_move_tail(&req->rq_replay_list,
2965 &imp->imp_committed_list);
2969 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
2970 peer_committed_transno);
2973 ptlrpc_free_request(req);
2976 if (skip_committed_list)
2979 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2981 LASSERT(req->rq_transno != 0);
2983 /* If other threads are waiting on imp_lock, stop processing
2984 * in this thread. Another thread can finish remaining work. */
2985 if (++commit_scanned % 2048 == 0) {
2986 now = ktime_get_seconds();
2987 if (now > start + 6)
2988 debug_level = D_WARNING;
2990 if ((commit_freed > 128 && now > start + 4) &&
2991 atomic_read(&imp->imp_waiting)) {
2992 if (debug_level == D_INFO)
2993 debug_level = D_RPCTRACE;
2998 if (req->rq_import_generation < imp_generation ||
3000 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
3001 req->rq_import_generation <
3002 imp_generation ? "stale" : "closed");
3004 if (imp->imp_replay_cursor == &req->rq_replay_list)
3005 imp->imp_replay_cursor =
3006 req->rq_replay_list.next;
3009 ptlrpc_free_request(req);
3013 /* if full lists processed without interruption, avoid next scan */
3014 if (debug_level == D_INFO) {
3015 imp->imp_last_transno_checked = peer_committed_transno;
3016 imp->imp_last_generation_checked = imp_generation;
3019 CDEBUG_LIMIT(debug_level,
3020 "%s: %s: skip=%u replay=%u/%u committed=%u/%u\n",
3021 imp->imp_obd->obd_name,
3022 debug_level == D_INFO ? "normal" : "overloaded",
3023 skip_committed_list, replay_freed, replay_scanned,
3024 commit_freed, commit_scanned);
3028 void ptlrpc_cleanup_client(struct obd_import *imp)
3035 * Schedule previously sent request for resend.
3036 * For bulk requests we assign new xid (to avoid problems with
3037 * lost replies and therefore several transfers landing into same buffer
3038 * from different sending attempts).
3040 void ptlrpc_resend_req(struct ptlrpc_request *req)
3042 DEBUG_REQ(D_HA, req, "going to resend");
3043 spin_lock(&req->rq_lock);
3046 * Request got reply but linked to the import list still.
3047 * Let ptlrpc_check_set() process it.
3049 if (ptlrpc_client_replied(req)) {
3050 spin_unlock(&req->rq_lock);
3051 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
3055 req->rq_status = -EAGAIN;
3058 req->rq_net_err = 0;
3059 req->rq_timedout = 0;
3061 ptlrpc_client_wake_req(req);
3062 spin_unlock(&req->rq_lock);
3065 /* XXX: this function and rq_status are currently unused */
3066 void ptlrpc_restart_req(struct ptlrpc_request *req)
3068 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
3069 req->rq_status = -ERESTARTSYS;
3071 spin_lock(&req->rq_lock);
3072 req->rq_restart = 1;
3073 req->rq_timedout = 0;
3074 ptlrpc_client_wake_req(req);
3075 spin_unlock(&req->rq_lock);
3079 * Grab additional reference on a request \a req
3081 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
3084 atomic_inc(&req->rq_refcount);
3087 EXPORT_SYMBOL(ptlrpc_request_addref);
3090 * Add a request to import replay_list.
3091 * Must be called under imp_lock
3093 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3094 struct obd_import *imp)
3096 struct ptlrpc_request *iter;
3098 assert_spin_locked(&imp->imp_lock);
3100 if (req->rq_transno == 0) {
3101 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3106 * clear this for new requests that were resent as well
3107 * as resent replayed requests.
3109 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3111 /* don't re-add requests that have been replayed */
3112 if (!list_empty(&req->rq_replay_list))
3115 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3117 spin_lock(&req->rq_lock);
3119 spin_unlock(&req->rq_lock);
3121 LASSERT(imp->imp_replayable);
3122 /* Balanced in ptlrpc_free_committed, usually. */
3123 ptlrpc_request_addref(req);
3124 list_for_each_entry_reverse(iter, &imp->imp_replay_list,
3127 * We may have duplicate transnos if we create and then
3128 * open a file, or for closes retained if to match creating
3129 * opens, so use req->rq_xid as a secondary key.
3130 * (See bugs 684, 685, and 428.)
3131 * XXX no longer needed, but all opens need transnos!
3133 if (iter->rq_transno > req->rq_transno)
3136 if (iter->rq_transno == req->rq_transno) {
3137 LASSERT(iter->rq_xid != req->rq_xid);
3138 if (iter->rq_xid > req->rq_xid)
3142 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3146 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3150 * Send request and wait until it completes.
3151 * Returns request processing status.
3153 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3155 struct ptlrpc_request_set *set;
3159 LASSERT(req->rq_set == NULL);
3160 LASSERT(!req->rq_receiving_reply);
3162 set = ptlrpc_prep_set();
3164 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3168 /* for distributed debugging */
3169 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3171 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3172 ptlrpc_request_addref(req);
3173 ptlrpc_set_add_req(set, req);
3174 rc = ptlrpc_set_wait(NULL, set);
3175 ptlrpc_set_destroy(set);
3179 EXPORT_SYMBOL(ptlrpc_queue_wait);
3182 * Callback used for replayed requests reply processing.
3183 * In case of successful reply calls registered request replay callback.
3184 * In case of error restart replay process.
3186 static int ptlrpc_replay_interpret(const struct lu_env *env,
3187 struct ptlrpc_request *req,
3190 struct ptlrpc_replay_async_args *aa = args;
3191 struct obd_import *imp = req->rq_import;
3194 atomic_dec(&imp->imp_replay_inflight);
3197 * Note: if it is bulk replay (MDS-MDS replay), then even if
3198 * server got the request, but bulk transfer timeout, let's
3199 * replay the bulk req again
3201 if (!ptlrpc_client_replied(req) ||
3203 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3204 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3205 GOTO(out, rc = -ETIMEDOUT);
3208 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3209 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3210 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3211 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3213 /** VBR: check version failure */
3214 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3215 /** replay was failed due to version mismatch */
3216 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3217 spin_lock(&imp->imp_lock);
3218 imp->imp_vbr_failed = 1;
3219 spin_unlock(&imp->imp_lock);
3220 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3222 /** The transno had better not change over replay. */
3223 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3224 lustre_msg_get_transno(req->rq_repmsg) ||
3225 lustre_msg_get_transno(req->rq_repmsg) == 0,
3227 lustre_msg_get_transno(req->rq_reqmsg),
3228 lustre_msg_get_transno(req->rq_repmsg));
3231 spin_lock(&imp->imp_lock);
3232 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3233 spin_unlock(&imp->imp_lock);
3234 LASSERT(imp->imp_last_replay_transno);
3236 /* transaction number shouldn't be bigger than the latest replayed */
3237 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3238 DEBUG_REQ(D_ERROR, req,
3239 "Reported transno=%llu is bigger than replayed=%llu",
3241 lustre_msg_get_transno(req->rq_reqmsg));
3242 GOTO(out, rc = -EINVAL);
3245 DEBUG_REQ(D_HA, req, "got reply");
3247 /* let the callback do fixups, possibly including in the request */
3248 if (req->rq_replay_cb)
3249 req->rq_replay_cb(req);
3251 if (ptlrpc_client_replied(req) &&
3252 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3253 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3254 lustre_msg_get_status(req->rq_repmsg),
3255 aa->praa_old_status);
3258 * Note: If the replay fails for MDT-MDT recovery, let's
3259 * abort all of the following requests in the replay
3260 * and sending list, because MDT-MDT update requests
3261 * are dependent on each other, see LU-7039
3263 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3264 struct ptlrpc_request *free_req;
3265 struct ptlrpc_request *tmp;
3267 spin_lock(&imp->imp_lock);
3268 list_for_each_entry_safe(free_req, tmp,
3269 &imp->imp_replay_list,
3271 ptlrpc_free_request(free_req);
3274 list_for_each_entry_safe(free_req, tmp,
3275 &imp->imp_committed_list,
3277 ptlrpc_free_request(free_req);
3280 list_for_each_entry_safe(free_req, tmp,
3281 &imp->imp_delayed_list,
3283 spin_lock(&free_req->rq_lock);
3284 free_req->rq_err = 1;
3285 free_req->rq_status = -EIO;
3286 ptlrpc_client_wake_req(free_req);
3287 spin_unlock(&free_req->rq_lock);
3290 list_for_each_entry_safe(free_req, tmp,
3291 &imp->imp_sending_list,
3293 spin_lock(&free_req->rq_lock);
3294 free_req->rq_err = 1;
3295 free_req->rq_status = -EIO;
3296 ptlrpc_client_wake_req(free_req);
3297 spin_unlock(&free_req->rq_lock);
3299 spin_unlock(&imp->imp_lock);
3302 /* Put it back for re-replay. */
3303 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3307 * Errors while replay can set transno to 0, but
3308 * imp_last_replay_transno shouldn't be set to 0 anyway
3310 if (req->rq_transno == 0)
3311 CERROR("Transno is 0 during replay!\n");
3313 /* continue with recovery */
3314 rc = ptlrpc_import_recovery_state_machine(imp);
3316 req->rq_send_state = aa->praa_old_state;
3319 /* this replay failed, so restart recovery */
3320 ptlrpc_connect_import(imp);
3326 * Prepares and queues request for replay.
3327 * Adds it to ptlrpcd queue for actual sending.
3328 * Returns 0 on success.
3330 int ptlrpc_replay_req(struct ptlrpc_request *req)
3332 struct ptlrpc_replay_async_args *aa;
3336 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3338 aa = ptlrpc_req_async_args(aa, req);
3339 memset(aa, 0, sizeof(*aa));
3341 /* Prepare request to be resent with ptlrpcd */
3342 aa->praa_old_state = req->rq_send_state;
3343 req->rq_send_state = LUSTRE_IMP_REPLAY;
3344 req->rq_phase = RQ_PHASE_NEW;
3345 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3347 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3349 req->rq_interpret_reply = ptlrpc_replay_interpret;
3350 /* Readjust the timeout for current conditions */
3351 ptlrpc_at_set_req_timeout(req);
3353 /* Tell server net_latency to calculate how long to wait for reply. */
3354 lustre_msg_set_service_timeout(req->rq_reqmsg,
3355 ptlrpc_at_get_net_latency(req));
3356 DEBUG_REQ(D_HA, req, "REPLAY");
3358 atomic_inc(&req->rq_import->imp_replay_inflight);
3359 spin_lock(&req->rq_lock);
3360 req->rq_early_free_repbuf = 0;
3361 spin_unlock(&req->rq_lock);
3362 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3364 ptlrpcd_add_req(req);
3369 * Aborts all in-flight request on import \a imp sending and delayed lists
3371 void ptlrpc_abort_inflight(struct obd_import *imp)
3373 struct ptlrpc_request *req;
3377 * Make sure that no new requests get processed for this import.
3378 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3379 * this flag and then putting requests on sending_list or delayed_list.
3381 assert_spin_locked(&imp->imp_lock);
3384 * XXX locking? Maybe we should remove each request with the list
3385 * locked? Also, how do we know if the requests on the list are
3386 * being freed at this time?
3388 list_for_each_entry(req, &imp->imp_sending_list, rq_list) {
3389 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3391 spin_lock(&req->rq_lock);
3392 if (req->rq_import_generation < imp->imp_generation) {
3394 req->rq_status = -EIO;
3395 ptlrpc_client_wake_req(req);
3397 spin_unlock(&req->rq_lock);
3400 list_for_each_entry(req, &imp->imp_delayed_list, rq_list) {
3401 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3403 spin_lock(&req->rq_lock);
3404 if (req->rq_import_generation < imp->imp_generation) {
3406 req->rq_status = -EIO;
3407 ptlrpc_client_wake_req(req);
3409 spin_unlock(&req->rq_lock);
3413 * Last chance to free reqs left on the replay list, but we
3414 * will still leak reqs that haven't committed.
3416 if (imp->imp_replayable)
3417 ptlrpc_free_committed(imp);
3423 * Abort all uncompleted requests in request set \a set
3425 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3427 struct ptlrpc_request *req;
3429 LASSERT(set != NULL);
3431 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
3432 spin_lock(&req->rq_lock);
3433 if (req->rq_phase != RQ_PHASE_RPC) {
3434 spin_unlock(&req->rq_lock);
3439 req->rq_status = -EINTR;
3440 ptlrpc_client_wake_req(req);
3441 spin_unlock(&req->rq_lock);
3446 * Initialize the XID for the node. This is common among all requests on
3447 * this node, and only requires the property that it is monotonically
3448 * increasing. It does not need to be sequential. Since this is also used
3449 * as the RDMA match bits, it is important that a single client NOT have
3450 * the same match bits for two different in-flight requests, hence we do
3451 * NOT want to have an XID per target or similar.
3453 * To avoid an unlikely collision between match bits after a client reboot
3454 * (which would deliver old data into the wrong RDMA buffer) initialize
3455 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3456 * If the time is clearly incorrect, we instead use a 62-bit random number.
3457 * In the worst case the random number will overflow 1M RPCs per second in
3458 * 9133 years, or permutations thereof.
3460 #define YEAR_2004 (1ULL << 30)
3461 void ptlrpc_init_xid(void)
3463 time64_t now = ktime_get_real_seconds();
3466 if (now < YEAR_2004) {
3467 get_random_bytes(&xid, sizeof(xid));
3469 xid |= (1ULL << 61);
3471 xid = (u64)now << 20;
3474 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3475 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3477 xid &= PTLRPC_BULK_OPS_MASK;
3478 atomic64_set(&ptlrpc_last_xid, xid);
3482 * Increase xid and returns resulting new value to the caller.
3484 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3485 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3486 * itself uses the last bulk xid needed, so the server can determine the
3487 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3488 * xid must align to a power-of-two value.
3490 * This is assumed to be true due to the initial ptlrpc_last_xid
3491 * value also being initialized to a power-of-two value. LU-1431
3493 __u64 ptlrpc_next_xid(void)
3495 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3499 * If request has a new allocated XID (new request or EINPROGRESS resend),
3500 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3501 * request to ensure previous bulk fails and avoid problems with lost replies
3502 * and therefore several transfers landing into the same buffer from different
3504 * Also, to avoid previous reply landing to a different sending attempt.
3506 void ptlrpc_set_mbits(struct ptlrpc_request *req)
3508 int md_count = req->rq_bulk ? req->rq_bulk->bd_md_count : 1;
3511 * Generate new matchbits for all resend requests, including
3514 if (req->rq_resend) {
3515 __u64 old_mbits = req->rq_mbits;
3518 * First time resend on -EINPROGRESS will generate new xid,
3519 * so we can actually use the rq_xid as rq_mbits in such case,
3520 * however, it's bit hard to distinguish such resend with a
3521 * 'resend for the -EINPROGRESS resend'. To make it simple,
3522 * we opt to generate mbits for all resend cases.
3524 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3526 req->rq_mbits = ptlrpc_next_xid();
3529 * Old version transfers rq_xid to peer as
3532 spin_lock(&req->rq_import->imp_lock);
3533 list_del_init(&req->rq_unreplied_list);
3534 ptlrpc_assign_next_xid_nolock(req);
3535 spin_unlock(&req->rq_import->imp_lock);
3536 req->rq_mbits = req->rq_xid;
3538 CDEBUG(D_HA, "resend with new mbits old x%llu new x%llu\n",
3539 old_mbits, req->rq_mbits);
3540 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3541 /* Request being sent first time, use xid as matchbits. */
3542 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3543 BULK_MBITS) || req->rq_mbits == 0)
3545 req->rq_mbits = req->rq_xid;
3547 req->rq_mbits -= md_count - 1;
3551 * Replay request, xid and matchbits have already been
3552 * correctly assigned.
3558 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3559 * that server can infer the number of bulks that were prepared,
3562 req->rq_mbits += md_count - 1;
3565 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3566 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3568 * It's ok to directly set the rq_xid here, since this xid bump
3569 * won't affect the request position in unreplied list.
3571 if (!OCD_HAS_FLAG(&req->rq_import->imp_connect_data, BULK_MBITS))
3572 req->rq_xid = req->rq_mbits;
3576 * Get a glimpse at what next xid value might have been.
3577 * Returns possible next xid.
3579 __u64 ptlrpc_sample_next_xid(void)
3581 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3583 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3586 * Functions for operating ptlrpc workers.
3588 * A ptlrpc work is a function which will be running inside ptlrpc context.
3589 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3591 * 1. after a work is created, it can be used many times, that is:
3592 * handler = ptlrpcd_alloc_work();
3593 * ptlrpcd_queue_work();
3595 * queue it again when necessary:
3596 * ptlrpcd_queue_work();
3597 * ptlrpcd_destroy_work();
3598 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3599 * it will only be queued once in any time. Also as its name implies, it may
3600 * have delay before it really runs by ptlrpcd thread.
3602 struct ptlrpc_work_async_args {
3603 int (*cb)(const struct lu_env *, void *);
3607 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3609 /* re-initialize the req */
3610 req->rq_timeout = obd_timeout;
3611 req->rq_sent = ktime_get_real_seconds();
3612 req->rq_deadline = req->rq_sent + req->rq_timeout;
3613 req->rq_phase = RQ_PHASE_INTERPRET;
3614 req->rq_next_phase = RQ_PHASE_COMPLETE;
3615 req->rq_xid = ptlrpc_next_xid();
3616 req->rq_import_generation = req->rq_import->imp_generation;
3618 ptlrpcd_add_req(req);
3621 static int work_interpreter(const struct lu_env *env,
3622 struct ptlrpc_request *req, void *args, int rc)
3624 struct ptlrpc_work_async_args *arg = args;
3626 LASSERT(ptlrpcd_check_work(req));
3627 LASSERT(arg->cb != NULL);
3629 rc = arg->cb(env, arg->cbdata);
3631 list_del_init(&req->rq_set_chain);
3634 if (atomic_dec_return(&req->rq_refcount) > 1) {
3635 atomic_set(&req->rq_refcount, 2);
3636 ptlrpcd_add_work_req(req);
3641 static int worker_format;
3643 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3645 return req->rq_pill.rc_fmt == (void *)&worker_format;
3649 * Create a work for ptlrpc.
3651 void *ptlrpcd_alloc_work(struct obd_import *imp,
3652 int (*cb)(const struct lu_env *, void *), void *cbdata)
3654 struct ptlrpc_request *req = NULL;
3655 struct ptlrpc_work_async_args *args;
3661 RETURN(ERR_PTR(-EINVAL));
3663 /* copy some code from deprecated fakereq. */
3664 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3666 CERROR("ptlrpc: run out of memory!\n");
3667 RETURN(ERR_PTR(-ENOMEM));
3670 ptlrpc_cli_req_init(req);
3672 req->rq_send_state = LUSTRE_IMP_FULL;
3673 req->rq_type = PTL_RPC_MSG_REQUEST;
3674 req->rq_import = class_import_get(imp);
3675 req->rq_interpret_reply = work_interpreter;
3676 /* don't want reply */
3677 req->rq_no_delay = req->rq_no_resend = 1;
3678 req->rq_pill.rc_fmt = (void *)&worker_format;
3680 args = ptlrpc_req_async_args(args, req);
3682 args->cbdata = cbdata;
3686 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3688 void ptlrpcd_destroy_work(void *handler)
3690 struct ptlrpc_request *req = handler;
3693 ptlrpc_req_finished(req);
3695 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3697 int ptlrpcd_queue_work(void *handler)
3699 struct ptlrpc_request *req = handler;
3702 * Check if the req is already being queued.
3704 * Here comes a trick: it lacks a way of checking if a req is being
3705 * processed reliably in ptlrpc. Here I have to use refcount of req
3706 * for this purpose. This is okay because the caller should use this
3707 * req as opaque data. - Jinshan
3709 LASSERT(atomic_read(&req->rq_refcount) > 0);
3710 if (atomic_inc_return(&req->rq_refcount) == 2)
3711 ptlrpcd_add_work_req(req);
3714 EXPORT_SYMBOL(ptlrpcd_queue_work);