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,
82 if (!is_vmalloc_addr(frag))
83 p = virt_to_page((unsigned long)frag);
85 p = vmalloc_to_page(frag);
86 ptlrpc_prep_bulk_page_nopin(desc, p, offset, page_len);
95 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
96 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
97 .release_frags = ptlrpc_release_bulk_page_pin,
99 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
101 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
102 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
103 .release_frags = ptlrpc_release_bulk_noop,
104 .add_iov_frag = ptlrpc_prep_bulk_frag_pages,
106 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
108 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
109 static int ptlrpcd_check_work(struct ptlrpc_request *req);
110 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
113 * Initialize passed in client structure \a cl.
115 void ptlrpc_init_client(int req_portal, int rep_portal, const char *name,
116 struct ptlrpc_client *cl)
118 cl->cli_request_portal = req_portal;
119 cl->cli_reply_portal = rep_portal;
122 EXPORT_SYMBOL(ptlrpc_init_client);
125 * Return PortalRPC connection for remore uud \a uuid
127 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid,
130 struct ptlrpc_connection *c;
131 struct lnet_nid self;
132 struct lnet_processid peer;
136 * ptlrpc_uuid_to_peer() initializes its 2nd parameter
137 * before accessing its values.
139 err = ptlrpc_uuid_to_peer(uuid, &peer, &self, refnet);
141 CNETERR("cannot find peer %s!\n", uuid->uuid);
145 c = ptlrpc_connection_get(&peer, &self, uuid);
147 memcpy(c->c_remote_uuid.uuid,
148 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
151 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
157 * Allocate and initialize new bulk descriptor on the sender.
158 * Returns pointer to the descriptor or NULL on error.
160 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned int nfrags,
161 unsigned int max_brw,
162 enum ptlrpc_bulk_op_type type,
164 const struct ptlrpc_bulk_frag_ops *ops)
166 struct ptlrpc_bulk_desc *desc;
169 LASSERT(ops->add_kiov_frag != NULL);
171 if (max_brw > PTLRPC_BULK_OPS_COUNT)
174 if (nfrags > LNET_MAX_IOV * max_brw)
181 OBD_ALLOC_LARGE(desc->bd_vec,
182 nfrags * sizeof(*desc->bd_vec));
186 spin_lock_init(&desc->bd_lock);
187 init_waitqueue_head(&desc->bd_waitq);
188 desc->bd_max_iov = nfrags;
189 desc->bd_iov_count = 0;
190 desc->bd_portal = portal;
191 desc->bd_type = type;
192 desc->bd_md_count = 0;
193 desc->bd_nob_last = LNET_MTU;
194 desc->bd_frag_ops = ops;
195 LASSERT(max_brw > 0);
196 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
198 * PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
199 * node. Negotiated ocd_brw_size will always be <= this number.
201 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
202 LNetInvalidateMDHandle(&desc->bd_mds[i]);
211 * Prepare bulk descriptor for specified outgoing request \a req that
212 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
213 * the bulk to be sent. Used on client-side.
214 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
217 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
219 unsigned int max_brw,
222 const struct ptlrpc_bulk_frag_ops
225 struct obd_import *imp = req->rq_import;
226 struct ptlrpc_bulk_desc *desc;
229 LASSERT(ptlrpc_is_bulk_op_passive(type));
231 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
235 desc->bd_import = class_import_get(imp);
238 desc->bd_cbid.cbid_fn = client_bulk_callback;
239 desc->bd_cbid.cbid_arg = desc;
241 /* This makes req own desc, and free it when she frees herself */
246 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
248 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
249 struct page *page, int pageoffset, int len,
252 struct bio_vec *kiov;
254 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
255 LASSERT(page != NULL);
256 LASSERT(pageoffset >= 0);
258 LASSERT(pageoffset + len <= PAGE_SIZE);
260 kiov = &desc->bd_vec[desc->bd_iov_count];
262 if (((desc->bd_iov_count % LNET_MAX_IOV) == 0) ||
263 ((desc->bd_nob_last + len) > LNET_MTU)) {
264 desc->bd_mds_off[desc->bd_md_count] = desc->bd_iov_count;
266 desc->bd_nob_last = 0;
267 LASSERT(desc->bd_md_count <= PTLRPC_BULK_OPS_COUNT);
270 desc->bd_nob_last += len;
276 kiov->bv_page = page;
277 kiov->bv_offset = pageoffset;
280 desc->bd_iov_count++;
282 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
284 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
288 LASSERT(desc != NULL);
289 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
290 LASSERT(desc->bd_refs == 0); /* network hands off */
291 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
292 LASSERT(desc->bd_frag_ops != NULL);
294 obd_pool_put_desc_pages(desc);
297 class_export_put(desc->bd_export);
299 class_import_put(desc->bd_import);
301 if (desc->bd_frag_ops->release_frags != NULL)
302 desc->bd_frag_ops->release_frags(desc);
304 OBD_FREE_LARGE(desc->bd_vec,
305 desc->bd_max_iov * sizeof(*desc->bd_vec));
309 EXPORT_SYMBOL(ptlrpc_free_bulk);
312 * Set server timelimit for this req, i.e. how long are we willing to wait
313 * for reply before timing out this request.
315 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
317 struct obd_device *obd;
319 LASSERT(req->rq_import);
320 obd = req->rq_import->imp_obd;
322 if (obd_at_off(obd)) {
323 /* non-AT settings */
325 * \a imp_server_timeout means this is reverse import and
326 * we send (currently only) ASTs to the client and cannot afford
327 * to wait too long for the reply, otherwise the other client
328 * (because of which we are sending this request) would
329 * timeout waiting for us
331 req->rq_timeout = req->rq_import->imp_server_timeout ?
332 obd_timeout / 2 : obd_timeout;
334 struct imp_at *at = &req->rq_import->imp_at;
338 idx = import_at_get_index(req->rq_import,
339 req->rq_request_portal);
340 serv_est = obd_at_get(obd, &at->iat_service_estimate[idx]);
342 * Currently a 32 bit value is sent over the
343 * wire for rq_timeout so please don't change this
344 * to time64_t. The work for LU-1158 will in time
345 * replace rq_timeout with a 64 bit nanosecond value
347 req->rq_timeout = at_est2timeout(serv_est);
350 * We could get even fancier here, using history to predict increased
353 * Let the server know what this RPC timeout is by putting it in the
356 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
358 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
360 /* Adjust max service estimate based on server value */
361 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
367 struct obd_device *obd;
369 LASSERT(req->rq_import);
370 obd = req->rq_import->imp_obd;
371 at = &req->rq_import->imp_at;
373 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
375 * max service estimates are tracked on the server side,
376 * so just keep minimal history here
378 oldse = obd_at_measure(obd, &at->iat_service_estimate[idx], serv_est);
380 unsigned int at_est = obd_at_get(obd,
381 &at->iat_service_estimate[idx]);
383 "The RPC service estimate for %s ptl %d has changed from %d to %d\n",
384 req->rq_import->imp_obd->obd_name,
385 req->rq_request_portal,
391 * Returns Expected network latency per remote node (secs).
393 * \param[in] req ptlrpc request
395 * \retval 0 if AT(Adaptive Timeout) is off
396 * \retval >0 (iat_net_latency) latency per node
398 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
400 struct obd_device *obd = req->rq_import->imp_obd;
402 return obd_at_off(obd) ?
403 0 : obd_at_get(obd, &req->rq_import->imp_at.iat_net_latency);
406 /* Adjust expected network latency */
407 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
408 timeout_t service_timeout)
410 time64_t now = ktime_get_real_seconds();
414 struct obd_device *obd;
416 LASSERT(req->rq_import);
417 obd = req->rq_import->imp_obd;
419 if (service_timeout > now - req->rq_sent + 3) {
421 * b=16408, however, this can also happen if early reply
422 * is lost and client RPC is expired and resent, early reply
423 * or reply of original RPC can still be fit in reply buffer
424 * of resent RPC, now client is measuring time from the
425 * resent time, but server sent back service time of original
428 CDEBUG_LIMIT((lustre_msg_get_flags(req->rq_reqmsg) &
429 MSG_RESENT) ? D_ADAPTTO : D_WARNING,
430 "Reported service time %u > total measured time %lld\n",
431 service_timeout, now - req->rq_sent);
435 /* Network latency is total time less server processing time,
438 nl = max_t(timeout_t, now - req->rq_sent - service_timeout, 0) + 1;
439 at = &req->rq_import->imp_at;
441 oldnl = obd_at_measure(obd, &at->iat_net_latency, nl);
443 timeout_t timeout = obd_at_get(obd, &at->iat_net_latency);
446 "The network latency for %s (nid %s) has changed from %d to %d\n",
447 req->rq_import->imp_obd->obd_name,
448 obd_uuid2str(&req->rq_import->imp_connection->c_remote_uuid),
453 static int unpack_reply(struct ptlrpc_request *req)
457 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
458 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
460 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: rc = %d",
466 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
468 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: rc = %d",
476 * Handle an early reply message, called with the rq_lock held.
477 * If anything goes wrong just ignore it - same as if it never happened
479 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
480 __must_hold(&req->rq_lock)
482 struct ptlrpc_request *early_req;
483 timeout_t service_timeout;
489 spin_unlock(&req->rq_lock);
491 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
493 spin_lock(&req->rq_lock);
497 rc = unpack_reply(early_req);
499 sptlrpc_cli_finish_early_reply(early_req);
500 spin_lock(&req->rq_lock);
505 * Use new timeout value just to adjust the local value for this
506 * request, don't include it into at_history. It is unclear yet why
507 * service time increased and should it be counted or skipped, e.g.
508 * that can be recovery case or some error or server, the real reply
509 * will add all new data if it is worth to add.
511 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
512 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
514 /* Network latency can be adjusted, it is pure network delays */
515 service_timeout = lustre_msg_get_service_timeout(early_req->rq_repmsg);
516 ptlrpc_at_adj_net_latency(req, service_timeout);
518 sptlrpc_cli_finish_early_reply(early_req);
520 spin_lock(&req->rq_lock);
521 olddl = req->rq_deadline;
523 * server assumes it now has rq_timeout from when the request
524 * arrived, so the client should give it at least that long.
525 * since we don't know the arrival time we'll use the original
528 req->rq_deadline = req->rq_sent + req->rq_timeout +
529 ptlrpc_at_get_net_latency(req);
531 /* The below message is checked in replay-single.sh test_65{a,b} */
532 /* The below message is checked in sanity-{gss,krb5} test_8 */
533 DEBUG_REQ(D_ADAPTTO, req,
534 "Early reply #%d, new deadline in %llds (%llds)",
536 req->rq_deadline - ktime_get_real_seconds(),
537 req->rq_deadline - olddl);
542 static struct kmem_cache *request_cache;
544 int ptlrpc_request_cache_init(void)
546 request_cache = kmem_cache_create("ptlrpc_cache",
547 sizeof(struct ptlrpc_request),
548 0, SLAB_HWCACHE_ALIGN, NULL);
549 return request_cache ? 0 : -ENOMEM;
552 void ptlrpc_request_cache_fini(void)
554 kmem_cache_destroy(request_cache);
557 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
559 struct ptlrpc_request *req;
561 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
565 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
567 OBD_SLAB_FREE_PTR(req, request_cache);
571 * Wind down request pool \a pool.
572 * Frees all requests from the pool too
574 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
576 struct ptlrpc_request *req;
578 LASSERT(pool != NULL);
580 spin_lock(&pool->prp_lock);
581 while ((req = list_first_entry_or_null(&pool->prp_req_list,
582 struct ptlrpc_request,
584 list_del(&req->rq_list);
585 LASSERT(req->rq_reqbuf);
586 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
587 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
588 ptlrpc_request_cache_free(req);
590 spin_unlock(&pool->prp_lock);
591 OBD_FREE(pool, sizeof(*pool));
593 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
596 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
598 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
603 while (size < pool->prp_rq_size)
606 LASSERTF(list_empty(&pool->prp_req_list) ||
607 size == pool->prp_rq_size,
608 "Trying to change pool size with nonempty pool from %d to %d bytes\n",
609 pool->prp_rq_size, size);
611 pool->prp_rq_size = size;
612 for (i = 0; i < num_rq; i++) {
613 struct ptlrpc_request *req;
614 struct lustre_msg *msg;
616 req = ptlrpc_request_cache_alloc(GFP_NOFS);
619 OBD_ALLOC_LARGE(msg, size);
621 ptlrpc_request_cache_free(req);
624 req->rq_reqbuf = msg;
625 req->rq_reqbuf_len = size;
627 spin_lock(&pool->prp_lock);
628 list_add_tail(&req->rq_list, &pool->prp_req_list);
629 spin_unlock(&pool->prp_lock);
633 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
636 * Create and initialize new request pool with given attributes:
637 * \a num_rq - initial number of requests to create for the pool
638 * \a msgsize - maximum message size possible for requests in thid pool
639 * \a populate_pool - function to be called when more requests need to be added
641 * Returns pointer to newly created pool or NULL on error.
643 struct ptlrpc_request_pool *
644 ptlrpc_init_rq_pool(int num_rq, int msgsize,
645 int (*populate_pool)(struct ptlrpc_request_pool *, int))
647 struct ptlrpc_request_pool *pool;
654 * Request next power of two for the allocation, because internally
655 * kernel would do exactly this
657 spin_lock_init(&pool->prp_lock);
658 INIT_LIST_HEAD(&pool->prp_req_list);
659 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
660 pool->prp_populate = populate_pool;
662 populate_pool(pool, num_rq);
666 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
669 * Fetches one request from pool \a pool
671 static struct ptlrpc_request *
672 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
674 struct ptlrpc_request *request;
675 struct lustre_msg *reqbuf;
680 spin_lock(&pool->prp_lock);
683 * See if we have anything in a pool, and bail out if nothing,
684 * in writeout path, where this matters, this is safe to do, because
685 * nothing is lost in this case, and when some in-flight requests
686 * complete, this code will be called again.
688 if (unlikely(list_empty(&pool->prp_req_list))) {
689 spin_unlock(&pool->prp_lock);
693 request = list_first_entry(&pool->prp_req_list, struct ptlrpc_request,
695 list_del_init(&request->rq_list);
696 spin_unlock(&pool->prp_lock);
698 LASSERT(request->rq_reqbuf);
699 LASSERT(request->rq_pool);
701 reqbuf = request->rq_reqbuf;
702 memset(request, 0, sizeof(*request));
703 request->rq_reqbuf = reqbuf;
704 request->rq_reqbuf_len = pool->prp_rq_size;
705 request->rq_pool = pool;
711 * Returns freed \a request to pool.
713 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
715 struct ptlrpc_request_pool *pool = request->rq_pool;
717 spin_lock(&pool->prp_lock);
718 LASSERT(list_empty(&request->rq_list));
719 LASSERT(!request->rq_receiving_reply);
720 list_add_tail(&request->rq_list, &pool->prp_req_list);
721 spin_unlock(&pool->prp_lock);
724 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
726 struct obd_import *imp = req->rq_import;
727 struct ptlrpc_request *iter;
729 assert_spin_locked(&imp->imp_lock);
730 LASSERT(list_empty(&req->rq_unreplied_list));
732 /* unreplied list is sorted by xid in ascending order */
733 list_for_each_entry_reverse(iter, &imp->imp_unreplied_list,
735 LASSERT(req->rq_xid != iter->rq_xid);
736 if (req->rq_xid < iter->rq_xid)
738 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
741 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
744 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
746 req->rq_xid = ptlrpc_next_xid();
747 ptlrpc_add_unreplied(req);
750 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
752 spin_lock(&req->rq_import->imp_lock);
753 ptlrpc_assign_next_xid_nolock(req);
754 spin_unlock(&req->rq_import->imp_lock);
757 static atomic64_t ptlrpc_last_xid;
759 static void ptlrpc_reassign_next_xid(struct ptlrpc_request *req)
761 spin_lock(&req->rq_import->imp_lock);
762 list_del_init(&req->rq_unreplied_list);
763 ptlrpc_assign_next_xid_nolock(req);
764 spin_unlock(&req->rq_import->imp_lock);
765 DEBUG_REQ(D_RPCTRACE, req, "reassign xid");
768 void ptlrpc_get_mod_rpc_slot(struct ptlrpc_request *req)
770 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
774 opc = lustre_msg_get_opc(req->rq_reqmsg);
775 tag = obd_get_mod_rpc_slot(cli, opc);
776 lustre_msg_set_tag(req->rq_reqmsg, tag);
777 ptlrpc_reassign_next_xid(req);
779 EXPORT_SYMBOL(ptlrpc_get_mod_rpc_slot);
781 void ptlrpc_put_mod_rpc_slot(struct ptlrpc_request *req)
783 __u16 tag = lustre_msg_get_tag(req->rq_reqmsg);
786 struct client_obd *cli = &req->rq_import->imp_obd->u.cli;
787 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
789 obd_put_mod_rpc_slot(cli, opc, tag);
792 EXPORT_SYMBOL(ptlrpc_put_mod_rpc_slot);
794 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
795 __u32 version, int opcode, char **bufs,
796 struct ptlrpc_cli_ctx *ctx)
799 struct obd_import *imp;
805 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
806 imp = request->rq_import;
807 lengths = request->rq_pill.rc_area[RCL_CLIENT];
810 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
812 rc = sptlrpc_req_get_ctx(request);
816 sptlrpc_req_set_flavor(request, opcode);
818 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
823 lustre_msg_add_version(request->rq_reqmsg, version);
824 request->rq_send_state = LUSTRE_IMP_FULL;
825 request->rq_type = PTL_RPC_MSG_REQUEST;
827 request->rq_req_cbid.cbid_fn = request_out_callback;
828 request->rq_req_cbid.cbid_arg = request;
830 request->rq_reply_cbid.cbid_fn = reply_in_callback;
831 request->rq_reply_cbid.cbid_arg = request;
833 request->rq_reply_deadline = 0;
834 request->rq_bulk_deadline = 0;
835 request->rq_req_deadline = 0;
836 request->rq_phase = RQ_PHASE_NEW;
837 request->rq_next_phase = RQ_PHASE_UNDEFINED;
839 request->rq_request_portal = imp->imp_client->cli_request_portal;
840 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
842 ptlrpc_at_set_req_timeout(request);
844 lustre_msg_set_opc(request->rq_reqmsg, opcode);
846 /* Let's setup deadline for req/reply/bulk unlink for opcode. */
847 if (cfs_fail_val == opcode) {
848 time64_t *fail_t = NULL, *fail2_t = NULL;
850 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
851 fail_t = &request->rq_bulk_deadline;
852 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
853 fail_t = &request->rq_reply_deadline;
854 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK)) {
855 fail_t = &request->rq_req_deadline;
856 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BOTH_UNLINK)) {
857 fail_t = &request->rq_reply_deadline;
858 fail2_t = &request->rq_bulk_deadline;
859 } else if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_ROUND_XID)) {
860 time64_t now = ktime_get_real_seconds();
861 u64 xid = ((u64)now >> 4) << 24;
863 atomic64_set(&ptlrpc_last_xid, xid);
867 *fail_t = ktime_get_real_seconds() +
868 PTLRPC_REQ_LONG_UNLINK;
871 *fail2_t = ktime_get_real_seconds() +
872 PTLRPC_REQ_LONG_UNLINK;
875 * The RPC is infected, let the test to change the
878 msleep(4 * MSEC_PER_SEC);
881 ptlrpc_assign_next_xid(request);
886 LASSERT(!request->rq_pool);
887 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
889 atomic_dec(&imp->imp_reqs);
890 class_import_put(imp);
894 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
897 * Pack request buffers for network transfer, performing necessary encryption
898 * steps if necessary.
900 int ptlrpc_request_pack(struct ptlrpc_request *request,
901 __u32 version, int opcode)
903 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
905 EXPORT_SYMBOL(ptlrpc_request_pack);
908 * Helper function to allocate new request on import \a imp
909 * and possibly using existing request from pool \a pool if provided.
910 * Returns allocated request structure with import field filled or
914 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
915 struct ptlrpc_request_pool *pool)
917 struct ptlrpc_request *request = NULL;
919 request = ptlrpc_request_cache_alloc(GFP_NOFS);
921 if (!request && pool)
922 request = ptlrpc_prep_req_from_pool(pool);
925 ptlrpc_cli_req_init(request);
927 LASSERTF((unsigned long)imp > 0x1000, "%px\n", imp);
928 LASSERT(imp != LP_POISON);
929 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%px\n",
931 LASSERT(imp->imp_client != LP_POISON);
933 request->rq_import = class_import_get(imp);
934 atomic_inc(&imp->imp_reqs);
936 CERROR("request allocation out of memory\n");
942 static int ptlrpc_reconnect_if_idle(struct obd_import *imp)
947 * initiate connection if needed when the import has been
948 * referenced by the new request to avoid races with disconnect.
949 * serialize this check against conditional state=IDLE
950 * in ptlrpc_disconnect_idle_interpret()
952 spin_lock(&imp->imp_lock);
953 if (imp->imp_state == LUSTRE_IMP_IDLE) {
954 imp->imp_generation++;
955 imp->imp_initiated_at = imp->imp_generation;
956 imp->imp_state = LUSTRE_IMP_NEW;
958 /* connect_import_locked releases imp_lock */
959 rc = ptlrpc_connect_import_locked(imp);
962 ptlrpc_pinger_add_import(imp);
964 spin_unlock(&imp->imp_lock);
970 * Helper function for creating a request.
971 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
972 * buffer structures according to capsule template \a format.
973 * Returns allocated request structure pointer or NULL on error.
975 static struct ptlrpc_request *
976 ptlrpc_request_alloc_internal(struct obd_import *imp,
977 struct ptlrpc_request_pool *pool,
978 const struct req_format *format)
980 struct ptlrpc_request *request;
982 request = __ptlrpc_request_alloc(imp, pool);
986 /* don't make expensive check for idling connection
987 * if it's already connected */
988 if (unlikely(imp->imp_state != LUSTRE_IMP_FULL)) {
989 if (ptlrpc_reconnect_if_idle(imp) < 0) {
990 atomic_dec(&imp->imp_reqs);
991 ptlrpc_request_free(request);
996 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
997 req_capsule_set(&request->rq_pill, format);
1002 * Allocate new request structure for import \a imp and initialize its
1003 * buffer structure according to capsule template \a format.
1005 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
1006 const struct req_format *format)
1008 return ptlrpc_request_alloc_internal(imp, NULL, format);
1010 EXPORT_SYMBOL(ptlrpc_request_alloc);
1013 * Allocate new request structure for import \a imp from pool \a pool and
1014 * initialize its buffer structure according to capsule template \a format.
1016 struct ptlrpc_request *
1017 ptlrpc_request_alloc_pool(struct obd_import *imp,
1018 struct ptlrpc_request_pool *pool,
1019 const struct req_format *format)
1021 return ptlrpc_request_alloc_internal(imp, pool, format);
1023 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
1026 * For requests not from pool, free memory of the request structure.
1027 * For requests obtained from a pool earlier, return request back to pool.
1029 void ptlrpc_request_free(struct ptlrpc_request *request)
1031 if (request->rq_pool)
1032 __ptlrpc_free_req_to_pool(request);
1034 ptlrpc_request_cache_free(request);
1036 EXPORT_SYMBOL(ptlrpc_request_free);
1039 * Allocate new request for operatione \a opcode and immediatelly pack it for
1041 * Only used for simple requests like OBD_PING where the only important
1042 * part of the request is operation itself.
1043 * Returns allocated request or NULL on error.
1045 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
1046 const struct req_format *format,
1047 __u32 version, int opcode)
1049 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
1053 rc = ptlrpc_request_pack(req, version, opcode);
1055 ptlrpc_request_free(req);
1061 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
1064 * Allocate and initialize new request set structure on the current CPT.
1065 * Returns a pointer to the newly allocated set structure or NULL on error.
1067 struct ptlrpc_request_set *ptlrpc_prep_set(void)
1069 struct ptlrpc_request_set *set;
1073 cpt = cfs_cpt_current(cfs_cpt_tab, 0);
1074 OBD_CPT_ALLOC(set, cfs_cpt_tab, cpt, sizeof(*set));
1077 atomic_set(&set->set_refcount, 1);
1078 INIT_LIST_HEAD(&set->set_requests);
1079 init_waitqueue_head(&set->set_waitq);
1080 atomic_set(&set->set_new_count, 0);
1081 atomic_set(&set->set_remaining, 0);
1082 spin_lock_init(&set->set_new_req_lock);
1083 INIT_LIST_HEAD(&set->set_new_requests);
1084 set->set_max_inflight = UINT_MAX;
1085 set->set_producer = NULL;
1086 set->set_producer_arg = NULL;
1091 EXPORT_SYMBOL(ptlrpc_prep_set);
1094 * Allocate and initialize new request set structure with flow control
1095 * extension. This extension allows to control the number of requests in-flight
1096 * for the whole set. A callback function to generate requests must be provided
1097 * and the request set will keep the number of requests sent over the wire to
1099 * Returns a pointer to the newly allocated set structure or NULL on error.
1101 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
1105 struct ptlrpc_request_set *set;
1107 set = ptlrpc_prep_set();
1111 set->set_max_inflight = max;
1112 set->set_producer = func;
1113 set->set_producer_arg = arg;
1119 * Wind down and free request set structure previously allocated with
1121 * Ensures that all requests on the set have completed and removes
1122 * all requests from the request list in a set.
1123 * If any unsent request happen to be on the list, pretends that they got
1124 * an error in flight and calls their completion handler.
1126 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1128 struct ptlrpc_request *req;
1134 /* Requests on the set should either all be completed, or all be new */
1135 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1136 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1137 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
1138 LASSERT(req->rq_phase == expected_phase);
1142 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1143 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1144 atomic_read(&set->set_remaining), n);
1146 while ((req = list_first_entry_or_null(&set->set_requests,
1147 struct ptlrpc_request,
1149 list_del_init(&req->rq_set_chain);
1151 LASSERT(req->rq_phase == expected_phase);
1153 if (req->rq_phase == RQ_PHASE_NEW) {
1154 ptlrpc_req_interpret(NULL, req, -EBADR);
1155 atomic_dec(&set->set_remaining);
1158 spin_lock(&req->rq_lock);
1160 req->rq_invalid_rqset = 0;
1161 spin_unlock(&req->rq_lock);
1163 ptlrpc_req_finished(req);
1166 LASSERT(atomic_read(&set->set_remaining) == 0);
1168 ptlrpc_reqset_put(set);
1171 EXPORT_SYMBOL(ptlrpc_set_destroy);
1174 * Add a new request to the general purpose request set.
1175 * Assumes request reference from the caller.
1177 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1178 struct ptlrpc_request *req)
1180 if (set == PTLRPCD_SET) {
1181 ptlrpcd_add_req(req);
1185 LASSERT(req->rq_import->imp_state != LUSTRE_IMP_IDLE);
1186 LASSERT(list_empty(&req->rq_set_chain));
1188 if (req->rq_allow_intr)
1189 set->set_allow_intr = 1;
1191 /* The set takes over the caller's request reference */
1192 list_add_tail(&req->rq_set_chain, &set->set_requests);
1194 atomic_inc(&set->set_remaining);
1195 req->rq_queued_time = ktime_get_seconds();
1197 if (req->rq_reqmsg) {
1198 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1199 lustre_msg_set_uid_gid(req->rq_reqmsg, NULL, NULL);
1202 if (set->set_producer)
1204 * If the request set has a producer callback, the RPC must be
1205 * sent straight away
1207 ptlrpc_send_new_req(req);
1209 EXPORT_SYMBOL(ptlrpc_set_add_req);
1212 * Add a request to a request with dedicated server thread
1213 * and wake the thread to make any necessary processing.
1214 * Currently only used for ptlrpcd.
1216 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1217 struct ptlrpc_request *req)
1219 struct ptlrpc_request_set *set = pc->pc_set;
1222 LASSERT(req->rq_set == NULL);
1223 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1225 spin_lock(&set->set_new_req_lock);
1227 * The set takes over the caller's request reference.
1230 req->rq_queued_time = ktime_get_seconds();
1231 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1232 count = atomic_inc_return(&set->set_new_count);
1233 spin_unlock(&set->set_new_req_lock);
1235 /* Only need to call wakeup once for the first entry. */
1237 wake_up(&set->set_waitq);
1240 * XXX: It maybe unnecessary to wakeup all the partners. But to
1241 * guarantee the async RPC can be processed ASAP, we have
1242 * no other better choice. It maybe fixed in future.
1244 for (i = 0; i < pc->pc_npartners; i++)
1245 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1250 * Based on the current state of the import, determine if the request
1251 * can be sent, is an error, or should be delayed.
1253 * Returns true if this request should be delayed. If false, and
1254 * *status is set, then the request can not be sent and *status is the
1255 * error code. If false and status is 0, then request can be sent.
1257 * The imp->imp_lock must be held.
1259 static int ptlrpc_import_delay_req(struct obd_import *imp,
1260 struct ptlrpc_request *req, int *status)
1268 if (req->rq_ctx_init || req->rq_ctx_fini) {
1269 /* always allow ctx init/fini rpc go through */
1270 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1271 DEBUG_REQ(D_ERROR, req, "Uninitialized import");
1273 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1274 unsigned int opc = lustre_msg_get_opc(req->rq_reqmsg);
1277 * pings or MDS-equivalent STATFS may safely
1280 DEBUG_REQ((opc == OBD_PING || opc == OST_STATFS) ?
1281 D_HA : D_ERROR, req, "IMP_CLOSED");
1283 } else if (ptlrpc_send_limit_expired(req)) {
1284 /* probably doesn't need to be a D_ERROR afterinitial testing */
1285 DEBUG_REQ(D_HA, req, "send limit expired");
1286 *status = -ETIMEDOUT;
1287 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1288 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1289 ;/* allow CONNECT even if import is invalid */
1290 if (atomic_read(&imp->imp_inval_count) != 0) {
1291 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1294 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1295 if (!imp->imp_deactive)
1296 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1297 *status = -ESHUTDOWN; /* b=12940 */
1298 } else if (req->rq_import_generation != imp->imp_generation) {
1299 DEBUG_REQ(req->rq_no_resend ? D_INFO : D_ERROR,
1300 req, "req wrong generation:");
1302 } else if (req->rq_send_state != imp->imp_state) {
1303 /* invalidate in progress - any requests should be drop */
1304 if (atomic_read(&imp->imp_inval_count) != 0) {
1305 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1307 } else if (req->rq_no_delay &&
1308 imp->imp_generation != imp->imp_initiated_at) {
1309 /* ignore nodelay for requests initiating connections */
1311 } else if (req->rq_allow_replay &&
1312 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1313 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1314 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1315 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1316 DEBUG_REQ(D_HA, req, "allow during recovery");
1326 * Decide if the error message should be printed to the console or not.
1327 * Makes its decision based on request type, status, and failure frequency.
1329 * \param[in] req request that failed and may need a console message
1331 * \retval false if no message should be printed
1332 * \retval true if console message should be printed
1334 static bool ptlrpc_console_allow(struct ptlrpc_request *req, __u32 opc, int err)
1336 LASSERT(req->rq_reqmsg != NULL);
1338 /* Suppress particular reconnect errors which are to be expected. */
1339 if (opc == OST_CONNECT || opc == OST_DISCONNECT ||
1340 opc == MDS_CONNECT || opc == MDS_DISCONNECT ||
1341 opc == MGS_CONNECT || opc == MGS_DISCONNECT) {
1342 /* Suppress timed out reconnect/disconnect requests */
1343 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1348 * Suppress most unavailable/again reconnect requests, but
1349 * print occasionally so it is clear client is trying to
1350 * connect to a server where no target is running.
1352 if ((err == -ENODEV || err == -EAGAIN) &&
1353 req->rq_import->imp_conn_cnt % 30 != 20)
1357 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1358 /* -EAGAIN is normal when using POSIX flocks */
1361 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1362 (req->rq_xid & 0xf) != 10)
1363 /* Suppress most ping requests, they may fail occasionally */
1370 * Check request processing status.
1371 * Returns the status.
1373 static int ptlrpc_check_status(struct ptlrpc_request *req)
1378 rc = lustre_msg_get_status(req->rq_repmsg);
1379 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1380 struct obd_import *imp = req->rq_import;
1381 struct lnet_nid *nid = &imp->imp_connection->c_peer.nid;
1382 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1384 if (ptlrpc_console_allow(req, opc, rc))
1385 LCONSOLE_ERROR_MSG(0x11,
1386 "%s: operation %s to node %s failed: rc = %d\n",
1387 imp->imp_obd->obd_name,
1389 libcfs_nidstr(nid), rc);
1390 RETURN(rc < 0 ? rc : -EINVAL);
1394 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1400 * save pre-versions of objects into request for replay.
1401 * Versions are obtained from server reply.
1404 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1406 struct lustre_msg *repmsg = req->rq_repmsg;
1407 struct lustre_msg *reqmsg = req->rq_reqmsg;
1408 __u64 *versions = lustre_msg_get_versions(repmsg);
1411 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1415 lustre_msg_set_versions(reqmsg, versions);
1416 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1417 versions[0], versions[1]);
1422 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1424 struct ptlrpc_request *req;
1426 assert_spin_locked(&imp->imp_lock);
1427 if (list_empty(&imp->imp_unreplied_list))
1430 req = list_first_entry(&imp->imp_unreplied_list, struct ptlrpc_request,
1432 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1434 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1435 imp->imp_known_replied_xid = req->rq_xid - 1;
1437 return req->rq_xid - 1;
1441 * Callback function called when client receives RPC reply for \a req.
1442 * Returns 0 on success or error code.
1443 * The return alue would be assigned to req->rq_status by the caller
1444 * as request processing status.
1445 * This function also decides if the request needs to be saved for later replay.
1447 static int after_reply(struct ptlrpc_request *req)
1449 struct obd_import *imp = req->rq_import;
1450 struct obd_device *obd = req->rq_import->imp_obd;
1457 LASSERT(obd != NULL);
1458 /* repbuf must be unlinked */
1459 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1461 if (req->rq_reply_truncated) {
1462 if (ptlrpc_no_resend(req)) {
1463 DEBUG_REQ(D_ERROR, req,
1464 "reply buffer overflow, expected=%d, actual size=%d",
1465 req->rq_nob_received, req->rq_repbuf_len);
1469 sptlrpc_cli_free_repbuf(req);
1471 * Pass the required reply buffer size (include
1472 * space for early reply).
1473 * NB: no need to roundup because alloc_repbuf
1476 req->rq_replen = req->rq_nob_received;
1477 req->rq_nob_received = 0;
1478 spin_lock(&req->rq_lock);
1480 spin_unlock(&req->rq_lock);
1484 work_start = ktime_get_real();
1485 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1488 * NB Until this point, the whole of the incoming message,
1489 * including buflens, status etc is in the sender's byte order.
1491 rc = sptlrpc_cli_unwrap_reply(req);
1493 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1498 * Security layer unwrap might ask resend this request.
1503 rc = unpack_reply(req);
1507 /* retry indefinitely on EINPROGRESS */
1508 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1509 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1510 time64_t now = ktime_get_real_seconds();
1512 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1513 D_RPCTRACE, req, "resending request on EINPROGRESS");
1514 spin_lock(&req->rq_lock);
1516 spin_unlock(&req->rq_lock);
1517 req->rq_nr_resend++;
1519 /* Readjust the timeout for current conditions */
1520 ptlrpc_at_set_req_timeout(req);
1522 * delay resend to give a chance to the server to get ready.
1523 * The delay is increased by 1s on every resend and is capped to
1524 * the current request timeout (i.e. obd_timeout if AT is off,
1525 * or AT service time x 125% + 5s, see at_est2timeout)
1527 if (req->rq_nr_resend > req->rq_timeout)
1528 req->rq_sent = now + req->rq_timeout;
1530 req->rq_sent = now + req->rq_nr_resend;
1532 /* Resend for EINPROGRESS will use a new XID */
1533 spin_lock(&imp->imp_lock);
1534 list_del_init(&req->rq_unreplied_list);
1535 spin_unlock(&imp->imp_lock);
1540 if (obd->obd_svc_stats) {
1541 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1543 ptlrpc_lprocfs_rpc_sent(req, timediff);
1546 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1547 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1548 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1549 lustre_msg_get_type(req->rq_repmsg));
1553 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1554 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1555 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1556 ptlrpc_at_adj_net_latency(req,
1557 lustre_msg_get_service_timeout(req->rq_repmsg));
1559 rc = ptlrpc_check_status(req);
1563 * Either we've been evicted, or the server has failed for
1564 * some reason. Try to reconnect, and if that fails, punt to
1567 if (ptlrpc_recoverable_error(rc)) {
1568 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1569 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1572 ptlrpc_request_handle_notconn(req);
1577 * Let's look if server sent slv. Do it only for RPC with
1580 ldlm_cli_update_pool(req);
1584 * Store transno in reqmsg for replay.
1586 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1587 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1588 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1591 if (lustre_msg_get_transno(req->rq_repmsg) ||
1592 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_ENQUEUE)
1593 imp->imp_no_cached_data = 0;
1595 if (imp->imp_replayable) {
1596 /* if other threads are waiting for ptlrpc_free_committed()
1597 * they could continue the work of freeing RPCs. That reduces
1598 * lock hold times, and distributes work more fairly across
1599 * waiting threads. We can't use spin_is_contended() since
1600 * there are many other places where imp_lock is held.
1602 atomic_inc(&imp->imp_waiting);
1603 spin_lock(&imp->imp_lock);
1604 atomic_dec(&imp->imp_waiting);
1606 * No point in adding already-committed requests to the replay
1607 * list, we will just remove them immediately. b=9829
1609 if (req->rq_transno != 0 &&
1611 lustre_msg_get_last_committed(req->rq_repmsg) ||
1613 /** version recovery */
1614 ptlrpc_save_versions(req);
1615 ptlrpc_retain_replayable_request(req, imp);
1616 } else if (req->rq_commit_cb &&
1617 list_empty(&req->rq_replay_list)) {
1619 * NB: don't call rq_commit_cb if it's already on
1620 * rq_replay_list, ptlrpc_free_committed() will call
1621 * it later, see LU-3618 for details
1623 spin_unlock(&imp->imp_lock);
1624 req->rq_commit_cb(req);
1625 atomic_inc(&imp->imp_waiting);
1626 spin_lock(&imp->imp_lock);
1627 atomic_dec(&imp->imp_waiting);
1631 * Replay-enabled imports return commit-status information.
1633 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1634 if (likely(committed > imp->imp_peer_committed_transno))
1635 imp->imp_peer_committed_transno = committed;
1637 ptlrpc_free_committed(imp);
1639 if (!list_empty(&imp->imp_replay_list)) {
1640 struct ptlrpc_request *last;
1642 last = list_entry(imp->imp_replay_list.prev,
1643 struct ptlrpc_request,
1646 * Requests with rq_replay stay on the list even if no
1647 * commit is expected.
1649 if (last->rq_transno > imp->imp_peer_committed_transno)
1650 ptlrpc_pinger_commit_expected(imp);
1653 spin_unlock(&imp->imp_lock);
1660 * Helper function to send request \a req over the network for the first time
1661 * Also adjusts request phase.
1662 * Returns 0 on success or error code.
1664 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1666 struct obd_import *imp = req->rq_import;
1671 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1673 /* do not try to go further if there is not enough memory in pool */
1674 if (req->rq_sent && req->rq_bulk)
1675 if (req->rq_bulk->bd_iov_count >
1676 obd_pool_get_free_pages(0) &&
1677 pool_is_at_full_capacity(0))
1680 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1681 (!req->rq_generation_set ||
1682 req->rq_import_generation == imp->imp_generation))
1685 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1687 spin_lock(&imp->imp_lock);
1689 LASSERT(req->rq_xid != 0);
1690 LASSERT(!list_empty(&req->rq_unreplied_list));
1692 if (!req->rq_generation_set)
1693 req->rq_import_generation = imp->imp_generation;
1695 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1696 spin_lock(&req->rq_lock);
1697 req->rq_waiting = 1;
1698 spin_unlock(&req->rq_lock);
1700 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1701 ptlrpc_import_state_name(req->rq_send_state),
1702 ptlrpc_import_state_name(imp->imp_state));
1703 LASSERT(list_empty(&req->rq_list));
1704 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1705 atomic_inc(&req->rq_import->imp_inflight);
1706 spin_unlock(&imp->imp_lock);
1711 spin_unlock(&imp->imp_lock);
1712 req->rq_status = rc;
1713 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1717 LASSERT(list_empty(&req->rq_list));
1718 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1719 atomic_inc(&req->rq_import->imp_inflight);
1722 * find the known replied XID from the unreplied list, CONNECT
1723 * and DISCONNECT requests are skipped to make the sanity check
1724 * on server side happy. see process_req_last_xid().
1726 * For CONNECT: Because replay requests have lower XID, it'll
1727 * break the sanity check if CONNECT bump the exp_last_xid on
1730 * For DISCONNECT: Since client will abort inflight RPC before
1731 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1732 * than the inflight RPC.
1734 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1735 min_xid = ptlrpc_known_replied_xid(imp);
1736 spin_unlock(&imp->imp_lock);
1738 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1740 lustre_msg_set_status(req->rq_reqmsg, current->pid);
1742 /* If the request to be sent is an LDLM callback, do not try to
1744 * An LDLM callback is sent by a server to a client in order to make
1745 * it release a lock, on a communication channel that uses a reverse
1746 * context. It cannot be refreshed on its own, as it is the 'reverse'
1747 * (server-side) representation of a client context.
1748 * We do not care if the reverse context is expired, and want to send
1749 * the LDLM callback anyway. Once the client receives the AST, it is
1750 * its job to refresh its own context if it has expired, hence
1751 * refreshing the associated reverse context on server side, before
1752 * being able to send the LDLM_CANCEL requested by the server.
1754 if (lustre_msg_get_opc(req->rq_reqmsg) != LDLM_BL_CALLBACK &&
1755 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_CP_CALLBACK &&
1756 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_GL_CALLBACK)
1757 rc = sptlrpc_req_refresh_ctx(req, 0);
1760 req->rq_status = rc;
1763 spin_lock(&req->rq_lock);
1764 req->rq_wait_ctx = 1;
1765 spin_unlock(&req->rq_lock);
1771 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1773 imp->imp_obd->obd_uuid.uuid,
1774 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1775 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1776 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1778 rc = ptl_send_rpc(req, 0);
1779 if (rc == -ENOMEM) {
1780 spin_lock(&imp->imp_lock);
1781 if (!list_empty(&req->rq_list)) {
1782 list_del_init(&req->rq_list);
1783 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1784 wake_up(&req->rq_import->imp_recovery_waitq);
1786 spin_unlock(&imp->imp_lock);
1787 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1791 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1793 spin_lock(&req->rq_lock);
1794 req->rq_net_err = 1;
1795 spin_unlock(&req->rq_lock);
1801 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1806 LASSERT(set->set_producer != NULL);
1808 remaining = atomic_read(&set->set_remaining);
1811 * populate the ->set_requests list with requests until we
1812 * reach the maximum number of RPCs in flight for this set
1814 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1815 rc = set->set_producer(set, set->set_producer_arg);
1816 if (rc == -ENOENT) {
1817 /* no more RPC to produce */
1818 set->set_producer = NULL;
1819 set->set_producer_arg = NULL;
1824 RETURN((atomic_read(&set->set_remaining) - remaining));
1828 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1829 * and no more replies are expected.
1830 * (it is possible to get less replies than requests sent e.g. due to timed out
1831 * requests or requests that we had trouble to send out)
1833 * NOTE: This function contains a potential schedule point (cond_resched()).
1835 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1837 struct ptlrpc_request *req, *next;
1838 LIST_HEAD(comp_reqs);
1839 int force_timer_recalc = 0;
1842 if (atomic_read(&set->set_remaining) == 0)
1845 list_for_each_entry_safe(req, next, &set->set_requests,
1847 struct obd_import *imp = req->rq_import;
1848 int unregistered = 0;
1852 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1853 list_move_tail(&req->rq_set_chain, &comp_reqs);
1858 * This schedule point is mainly for the ptlrpcd caller of this
1859 * function. Most ptlrpc sets are not long-lived and unbounded
1860 * in length, but at the least the set used by the ptlrpcd is.
1861 * Since the processing time is unbounded, we need to insert an
1862 * explicit schedule point to make the thread well-behaved.
1867 * If the caller requires to allow to be interpreted by force
1868 * and it has really been interpreted, then move the request
1869 * to RQ_PHASE_INTERPRET phase in spite of what the current
1872 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1873 req->rq_status = -EINTR;
1874 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1877 * Since it is interpreted and we have to wait for
1878 * the reply to be unlinked, then use sync mode.
1882 GOTO(interpret, req->rq_status);
1885 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1886 force_timer_recalc = 1;
1888 /* delayed send - skip */
1889 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1892 /* delayed resend - skip */
1893 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1894 req->rq_sent > ktime_get_real_seconds())
1897 if (!(req->rq_phase == RQ_PHASE_RPC ||
1898 req->rq_phase == RQ_PHASE_BULK ||
1899 req->rq_phase == RQ_PHASE_INTERPRET ||
1900 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1901 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1902 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1906 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1907 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1908 LASSERT(req->rq_next_phase != req->rq_phase);
1909 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1911 if (req->rq_req_deadline &&
1912 !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1913 req->rq_req_deadline = 0;
1914 if (req->rq_reply_deadline &&
1915 !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1916 req->rq_reply_deadline = 0;
1917 if (req->rq_bulk_deadline &&
1918 !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1919 req->rq_bulk_deadline = 0;
1922 * Skip processing until reply is unlinked. We
1923 * can't return to pool before that and we can't
1924 * call interpret before that. We need to make
1925 * sure that all rdma transfers finished and will
1926 * not corrupt any data.
1928 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1929 ptlrpc_cli_wait_unlink(req))
1931 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1932 ptlrpc_client_bulk_active(req))
1936 * Turn fail_loc off to prevent it from looping
1939 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1940 CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1943 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1944 CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1949 * Move to next phase if reply was successfully
1952 ptlrpc_rqphase_move(req, req->rq_next_phase);
1955 if (req->rq_phase == RQ_PHASE_INTERPRET)
1956 GOTO(interpret, req->rq_status);
1959 * Note that this also will start async reply unlink.
1961 if (req->rq_net_err && !req->rq_timedout) {
1962 ptlrpc_expire_one_request(req, 1);
1965 * Check if we still need to wait for unlink.
1967 if (ptlrpc_cli_wait_unlink(req) ||
1968 ptlrpc_client_bulk_active(req))
1970 /* If there is no need to resend, fail it now. */
1971 if (req->rq_no_resend) {
1972 if (req->rq_status == 0)
1973 req->rq_status = -EIO;
1974 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1975 GOTO(interpret, req->rq_status);
1982 if (!ptlrpc_unregister_reply(req, 1)) {
1983 ptlrpc_unregister_bulk(req, 1);
1987 spin_lock(&req->rq_lock);
1988 req->rq_replied = 0;
1989 spin_unlock(&req->rq_lock);
1990 if (req->rq_status == 0)
1991 req->rq_status = -EIO;
1992 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1993 GOTO(interpret, req->rq_status);
1997 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1998 * so it sets rq_intr regardless of individual rpc
1999 * timeouts. The synchronous IO waiting path sets
2000 * rq_intr irrespective of whether ptlrpcd
2001 * has seen a timeout. Our policy is to only interpret
2002 * interrupted rpcs after they have timed out, so we
2003 * need to enforce that here.
2006 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
2007 req->rq_wait_ctx)) {
2008 req->rq_status = -EINTR;
2009 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2010 GOTO(interpret, req->rq_status);
2013 if (req->rq_phase == RQ_PHASE_RPC) {
2014 if (req->rq_timedout || req->rq_resend ||
2015 req->rq_waiting || req->rq_wait_ctx) {
2018 if (!ptlrpc_unregister_reply(req, 1)) {
2019 ptlrpc_unregister_bulk(req, 1);
2023 spin_lock(&imp->imp_lock);
2024 if (ptlrpc_import_delay_req(imp, req,
2027 * put on delay list - only if we wait
2028 * recovery finished - before send
2030 list_move_tail(&req->rq_list,
2031 &imp->imp_delayed_list);
2032 spin_unlock(&imp->imp_lock);
2037 req->rq_status = status;
2038 ptlrpc_rqphase_move(req,
2039 RQ_PHASE_INTERPRET);
2040 spin_unlock(&imp->imp_lock);
2041 GOTO(interpret, req->rq_status);
2043 /* ignore on just initiated connections */
2044 if (ptlrpc_no_resend(req) &&
2045 !req->rq_wait_ctx &&
2046 imp->imp_generation !=
2047 imp->imp_initiated_at) {
2048 req->rq_status = -ENOTCONN;
2049 ptlrpc_rqphase_move(req,
2050 RQ_PHASE_INTERPRET);
2051 spin_unlock(&imp->imp_lock);
2052 GOTO(interpret, req->rq_status);
2055 /* don't resend too fast in case of network
2058 if (ktime_get_real_seconds() < (req->rq_sent + 1)
2059 && req->rq_net_err && req->rq_timedout) {
2061 DEBUG_REQ(D_INFO, req,
2062 "throttle request");
2063 /* Don't try to resend RPC right away
2064 * as it is likely it will fail again
2065 * and ptlrpc_check_set() will be
2066 * called again, keeping this thread
2067 * busy. Instead, wait for the next
2068 * timeout. Flag it as resend to
2069 * ensure we don't wait to long.
2072 spin_unlock(&imp->imp_lock);
2076 list_move_tail(&req->rq_list,
2077 &imp->imp_sending_list);
2079 spin_unlock(&imp->imp_lock);
2081 spin_lock(&req->rq_lock);
2082 req->rq_waiting = 0;
2083 spin_unlock(&req->rq_lock);
2085 if (req->rq_timedout || req->rq_resend) {
2087 * This is re-sending anyways,
2088 * let's mark req as resend.
2090 spin_lock(&req->rq_lock);
2092 spin_unlock(&req->rq_lock);
2095 * rq_wait_ctx is only touched by ptlrpcd,
2096 * so no lock is needed here.
2098 status = sptlrpc_req_refresh_ctx(req, 0);
2101 req->rq_status = status;
2102 spin_lock(&req->rq_lock);
2103 req->rq_wait_ctx = 0;
2104 spin_unlock(&req->rq_lock);
2105 force_timer_recalc = 1;
2107 spin_lock(&req->rq_lock);
2108 req->rq_wait_ctx = 1;
2109 spin_unlock(&req->rq_lock);
2114 spin_lock(&req->rq_lock);
2115 req->rq_wait_ctx = 0;
2116 spin_unlock(&req->rq_lock);
2120 * In any case, the previous bulk should be
2121 * cleaned up to prepare for the new sending
2124 !ptlrpc_unregister_bulk(req, 1))
2127 rc = ptl_send_rpc(req, 0);
2128 if (rc == -ENOMEM) {
2129 spin_lock(&imp->imp_lock);
2130 if (!list_empty(&req->rq_list)) {
2131 list_del_init(&req->rq_list);
2132 if (atomic_dec_and_test(&imp->imp_inflight))
2133 wake_up(&imp->imp_recovery_waitq);
2135 spin_unlock(&imp->imp_lock);
2136 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2140 DEBUG_REQ(D_HA, req,
2141 "send failed: rc = %d", rc);
2142 force_timer_recalc = 1;
2143 spin_lock(&req->rq_lock);
2144 req->rq_net_err = 1;
2145 spin_unlock(&req->rq_lock);
2148 /* need to reset the timeout */
2149 force_timer_recalc = 1;
2152 spin_lock(&req->rq_lock);
2154 if (ptlrpc_client_early(req)) {
2155 ptlrpc_at_recv_early_reply(req);
2156 spin_unlock(&req->rq_lock);
2160 /* Still waiting for a reply? */
2161 if (ptlrpc_client_recv(req)) {
2162 spin_unlock(&req->rq_lock);
2166 /* Did we actually receive a reply? */
2167 if (!ptlrpc_client_replied(req)) {
2168 spin_unlock(&req->rq_lock);
2172 spin_unlock(&req->rq_lock);
2175 * unlink from net because we are going to
2176 * swab in-place of reply buffer
2178 unregistered = ptlrpc_unregister_reply(req, 1);
2182 req->rq_status = after_reply(req);
2183 if (req->rq_resend) {
2184 force_timer_recalc = 1;
2189 * If there is no bulk associated with this request,
2190 * then we're done and should let the interpreter
2191 * process the reply. Similarly if the RPC returned
2192 * an error, and therefore the bulk will never arrive.
2194 if (!req->rq_bulk || req->rq_status < 0) {
2195 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2196 GOTO(interpret, req->rq_status);
2199 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2202 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2203 if (ptlrpc_client_bulk_active(req))
2206 if (req->rq_bulk->bd_failure) {
2208 * The RPC reply arrived OK, but the bulk screwed
2209 * up! Dead weird since the server told us the RPC
2210 * was good after getting the REPLY for her GET or
2211 * the ACK for her PUT.
2213 DEBUG_REQ(D_ERROR, req, "bulk transfer failed %d/%d/%d",
2215 req->rq_bulk->bd_nob,
2216 req->rq_bulk->bd_nob_transferred);
2217 req->rq_status = -EIO;
2220 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2223 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2226 * This moves to "unregistering" phase we need to wait for
2229 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2230 /* start async bulk unlink too */
2231 ptlrpc_unregister_bulk(req, 1);
2235 if (!ptlrpc_unregister_bulk(req, async))
2239 * When calling interpret receiving already should be
2242 LASSERT(!req->rq_receiving_reply);
2244 ptlrpc_req_interpret(env, req, req->rq_status);
2246 if (ptlrpcd_check_work(req)) {
2247 atomic_dec(&set->set_remaining);
2250 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2254 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2256 imp->imp_obd->obd_uuid.uuid,
2257 lustre_msg_get_status(req->rq_reqmsg),
2259 obd_import_nid2str(imp),
2260 lustre_msg_get_opc(req->rq_reqmsg),
2261 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2263 spin_lock(&imp->imp_lock);
2265 * Request already may be not on sending or delaying list. This
2266 * may happen in the case of marking it erroneous for the case
2267 * ptlrpc_import_delay_req(req, status) find it impossible to
2268 * allow sending this rpc and returns *status != 0.
2270 if (!list_empty(&req->rq_list)) {
2271 list_del_init(&req->rq_list);
2272 if (atomic_dec_and_test(&imp->imp_inflight))
2273 wake_up(&imp->imp_recovery_waitq);
2275 list_del_init(&req->rq_unreplied_list);
2276 spin_unlock(&imp->imp_lock);
2278 atomic_dec(&set->set_remaining);
2279 wake_up(&imp->imp_recovery_waitq);
2281 if (set->set_producer) {
2282 /* produce a new request if possible */
2283 if (ptlrpc_set_producer(set) > 0)
2284 force_timer_recalc = 1;
2287 * free the request that has just been completed
2288 * in order not to pollute set->set_requests
2290 list_del_init(&req->rq_set_chain);
2291 spin_lock(&req->rq_lock);
2293 req->rq_invalid_rqset = 0;
2294 spin_unlock(&req->rq_lock);
2296 /* record rq_status to compute the final status later */
2297 if (req->rq_status != 0)
2298 set->set_rc = req->rq_status;
2299 ptlrpc_req_finished(req);
2301 list_move_tail(&req->rq_set_chain, &comp_reqs);
2306 * move completed request at the head of list so it's easier for
2307 * caller to find them
2309 list_splice(&comp_reqs, &set->set_requests);
2311 /* If we hit an error, we want to recover promptly. */
2312 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2314 EXPORT_SYMBOL(ptlrpc_check_set);
2317 * Time out request \a req. is \a async_unlink is set, that means do not wait
2318 * until LNet actually confirms network buffer unlinking.
2319 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2321 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2323 struct obd_import *imp = req->rq_import;
2324 unsigned int debug_mask = D_RPCTRACE;
2329 spin_lock(&req->rq_lock);
2330 req->rq_timedout = 1;
2331 spin_unlock(&req->rq_lock);
2333 opc = lustre_msg_get_opc(req->rq_reqmsg);
2334 if (ptlrpc_console_allow(req, opc,
2335 lustre_msg_get_status(req->rq_reqmsg)))
2336 debug_mask = D_WARNING;
2337 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2338 req->rq_net_err ? "failed due to network error" :
2339 ((req->rq_real_sent == 0 ||
2340 req->rq_real_sent < req->rq_sent ||
2341 req->rq_real_sent >= req->rq_deadline) ?
2342 "timed out for sent delay" : "timed out for slow reply"),
2343 req->rq_sent, req->rq_real_sent);
2345 if (imp && obd_debug_peer_on_timeout)
2346 LNetDebugPeer(&imp->imp_connection->c_peer);
2348 ptlrpc_unregister_reply(req, async_unlink);
2349 ptlrpc_unregister_bulk(req, async_unlink);
2351 if (obd_dump_on_timeout)
2352 libcfs_debug_dumplog();
2355 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2359 atomic_inc(&imp->imp_timeouts);
2361 /* The DLM server doesn't want recovery run on its imports. */
2362 if (imp->imp_dlm_fake)
2366 * If this request is for recovery or other primordial tasks,
2367 * then error it out here.
2369 if (req->rq_ctx_init || req->rq_ctx_fini ||
2370 req->rq_send_state != LUSTRE_IMP_FULL ||
2371 imp->imp_obd->obd_no_recov) {
2372 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2373 ptlrpc_import_state_name(req->rq_send_state),
2374 ptlrpc_import_state_name(imp->imp_state));
2375 spin_lock(&req->rq_lock);
2376 req->rq_status = -ETIMEDOUT;
2378 spin_unlock(&req->rq_lock);
2383 * if a request can't be resent we can't wait for an answer after
2386 if (ptlrpc_no_resend(req)) {
2387 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2391 if (opc != OBD_PING || req->rq_xid > imp->imp_highest_replied_xid)
2392 ptlrpc_fail_import(imp,
2393 lustre_msg_get_conn_cnt(req->rq_reqmsg));
2399 * Time out all uncompleted requests in request set pointed by \a data
2400 * This is called when a wait times out.
2402 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2404 struct ptlrpc_request *req;
2405 time64_t now = ktime_get_real_seconds();
2408 LASSERT(set != NULL);
2411 * A timeout expired. See which reqs it applies to...
2413 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2414 /* don't expire request waiting for context */
2415 if (req->rq_wait_ctx)
2418 /* Request in-flight? */
2419 if (!((req->rq_phase == RQ_PHASE_RPC &&
2420 !req->rq_waiting && !req->rq_resend) ||
2421 (req->rq_phase == RQ_PHASE_BULK)))
2424 if (req->rq_timedout || /* already dealt with */
2425 req->rq_deadline > now) /* not expired */
2429 * Deal with this guy. Do it asynchronously to not block
2432 ptlrpc_expire_one_request(req, 1);
2434 * Loops require that we resched once in a while to avoid
2435 * RCU stalls and a few other problems.
2443 * Interrupts (sets interrupted flag) all uncompleted requests in
2444 * a set \a data. This is called when a wait_event is interrupted
2447 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2449 struct ptlrpc_request *req;
2451 LASSERT(set != NULL);
2452 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2454 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2458 if (req->rq_phase != RQ_PHASE_RPC &&
2459 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2460 !req->rq_allow_intr)
2463 spin_lock(&req->rq_lock);
2465 spin_unlock(&req->rq_lock);
2470 * Get the smallest timeout in the set; this does NOT set a timeout.
2472 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2474 time64_t now = ktime_get_real_seconds();
2476 struct ptlrpc_request *req;
2480 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2481 /* Request in-flight? */
2482 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2483 (req->rq_phase == RQ_PHASE_BULK) ||
2484 (req->rq_phase == RQ_PHASE_NEW)))
2487 /* Already timed out. */
2488 if (req->rq_timedout)
2491 /* Waiting for ctx. */
2492 if (req->rq_wait_ctx)
2495 if (req->rq_phase == RQ_PHASE_NEW)
2496 deadline = req->rq_sent;
2497 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2498 deadline = req->rq_sent;
2500 deadline = req->rq_sent + req->rq_timeout;
2502 if (deadline <= now) /* actually expired already */
2503 timeout = 1; /* ASAP */
2504 else if (timeout == 0 || timeout > deadline - now)
2505 timeout = deadline - now;
2511 * Send all unset request from the set and then wait untill all
2512 * requests in the set complete (either get a reply, timeout, get an
2513 * error or otherwise be interrupted).
2514 * Returns 0 on success or error code otherwise.
2516 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2518 struct ptlrpc_request *req;
2523 if (set->set_producer)
2524 (void)ptlrpc_set_producer(set);
2526 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2527 if (req->rq_phase == RQ_PHASE_NEW)
2528 (void)ptlrpc_send_new_req(req);
2531 if (list_empty(&set->set_requests))
2535 timeout = ptlrpc_set_next_timeout(set);
2538 * wait until all complete, interrupted, or an in-flight
2541 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2544 if ((timeout == 0 && !signal_pending(current)) ||
2545 set->set_allow_intr) {
2547 * No requests are in-flight (ether timed out
2548 * or delayed), so we can allow interrupts.
2549 * We still want to block for a limited time,
2550 * so we allow interrupts during the timeout.
2552 rc = l_wait_event_abortable_timeout(
2554 ptlrpc_check_set(NULL, set),
2555 cfs_time_seconds(timeout ? timeout : 1));
2558 ptlrpc_expired_set(set);
2559 } else if (rc < 0) {
2561 ptlrpc_interrupted_set(set);
2567 * At least one request is in flight, so no
2568 * interrupts are allowed. Wait until all
2569 * complete, or an in-flight req times out.
2571 rc = wait_event_idle_timeout(
2573 ptlrpc_check_set(NULL, set),
2574 cfs_time_seconds(timeout ? timeout : 1));
2576 ptlrpc_expired_set(set);
2583 * LU-769 - if we ignored the signal because
2584 * it was already pending when we started, we
2585 * need to handle it now or we risk it being
2588 if (rc == -ETIMEDOUT &&
2589 signal_pending(current)) {
2592 siginitset(&new, LUSTRE_FATAL_SIGS);
2593 sigprocmask(SIG_BLOCK, &new, &old);
2595 * In fact we only interrupt for the
2596 * "fatal" signals like SIGINT or
2597 * SIGKILL. We still ignore less
2598 * important signals since ptlrpc set
2599 * is not easily reentrant from
2602 if (signal_pending(current))
2603 ptlrpc_interrupted_set(set);
2604 sigprocmask(SIG_SETMASK, &old, NULL);
2608 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2611 * -EINTR => all requests have been flagged rq_intr so next
2613 * -ETIMEDOUT => someone timed out. When all reqs have
2614 * timed out, signals are enabled allowing completion with
2616 * I don't really care if we go once more round the loop in
2617 * the error cases -eeb.
2619 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2620 list_for_each_entry(req, &set->set_requests,
2622 spin_lock(&req->rq_lock);
2623 req->rq_invalid_rqset = 1;
2624 spin_unlock(&req->rq_lock);
2627 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2629 LASSERT(atomic_read(&set->set_remaining) == 0);
2631 rc = set->set_rc; /* rq_status of already freed requests if any */
2632 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2633 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2634 if (req->rq_status != 0)
2635 rc = req->rq_status;
2640 EXPORT_SYMBOL(ptlrpc_set_wait);
2643 * Helper fuction for request freeing.
2644 * Called when request count reached zero and request needs to be freed.
2645 * Removes request from all sorts of sending/replay lists it might be on,
2646 * frees network buffers if any are present.
2647 * If \a locked is set, that means caller is already holding import imp_lock
2648 * and so we no longer need to reobtain it (for certain lists manipulations)
2650 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2657 LASSERT(!request->rq_srv_req);
2658 LASSERT(request->rq_export == NULL);
2659 LASSERTF(!request->rq_receiving_reply, "req %px\n", request);
2660 LASSERTF(list_empty(&request->rq_list), "req %px\n", request);
2661 LASSERTF(list_empty(&request->rq_set_chain), "req %px\n", request);
2662 LASSERTF(!request->rq_replay, "req %px\n", request);
2664 req_capsule_fini(&request->rq_pill);
2667 * We must take it off the imp_replay_list first. Otherwise, we'll set
2668 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2670 if (request->rq_import) {
2672 spin_lock(&request->rq_import->imp_lock);
2673 list_del_init(&request->rq_replay_list);
2674 list_del_init(&request->rq_unreplied_list);
2676 spin_unlock(&request->rq_import->imp_lock);
2678 LASSERTF(list_empty(&request->rq_replay_list), "req %px\n", request);
2680 if (atomic_read(&request->rq_refcount) != 0) {
2681 DEBUG_REQ(D_ERROR, request,
2682 "freeing request with nonzero refcount");
2686 if (request->rq_repbuf)
2687 sptlrpc_cli_free_repbuf(request);
2689 if (request->rq_import) {
2690 if (!ptlrpcd_check_work(request)) {
2691 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2692 atomic_dec(&request->rq_import->imp_reqs);
2694 class_import_put(request->rq_import);
2695 request->rq_import = NULL;
2697 if (request->rq_bulk)
2698 ptlrpc_free_bulk(request->rq_bulk);
2700 if (request->rq_reqbuf || request->rq_clrbuf)
2701 sptlrpc_cli_free_reqbuf(request);
2703 if (request->rq_cli_ctx)
2704 sptlrpc_req_put_ctx(request, !locked);
2706 if (request->rq_pool)
2707 __ptlrpc_free_req_to_pool(request);
2709 ptlrpc_request_cache_free(request);
2715 * Drops one reference count for request \a request.
2716 * \a locked set indicates that caller holds import imp_lock.
2717 * Frees the request whe reference count reaches zero.
2719 * \retval 1 the request is freed
2720 * \retval 0 some others still hold references on the request
2722 static int __ptlrpc_req_put(struct ptlrpc_request *request, int locked)
2730 LASSERT(request != LP_POISON);
2731 LASSERT(request->rq_reqmsg != LP_POISON);
2733 DEBUG_REQ(D_INFO, request, "refcount now %u",
2734 atomic_read(&request->rq_refcount) - 1);
2736 spin_lock(&request->rq_lock);
2737 count = atomic_dec_return(&request->rq_refcount);
2738 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2741 * For open RPC, the client does not know the EA size (LOV, ACL, and
2742 * so on) before replied, then the client has to reserve very large
2743 * reply buffer. Such buffer will not be released until the RPC freed.
2744 * Since The open RPC is replayable, we need to keep it in the replay
2745 * list until close. If there are a lot of files opened concurrently,
2746 * then the client may be OOM.
2748 * If fact, it is unnecessary to keep reply buffer for open replay,
2749 * related EAs have already been saved via mdc_save_lovea() before
2750 * coming here. So it is safe to free the reply buffer some earlier
2751 * before releasing the RPC to avoid client OOM. LU-9514
2753 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2754 spin_lock(&request->rq_early_free_lock);
2755 sptlrpc_cli_free_repbuf(request);
2756 request->rq_repbuf = NULL;
2757 request->rq_repbuf_len = 0;
2758 request->rq_repdata = NULL;
2759 request->rq_reqdata_len = 0;
2760 spin_unlock(&request->rq_early_free_lock);
2762 spin_unlock(&request->rq_lock);
2765 __ptlrpc_free_req(request, locked);
2771 * Drop one request reference. Must be called with import imp_lock held.
2772 * When reference count drops to zero, request is freed.
2774 void ptlrpc_req_put_with_imp_lock(struct ptlrpc_request *request)
2776 assert_spin_locked(&request->rq_import->imp_lock);
2777 (void)__ptlrpc_req_put(request, 1);
2781 * Drops one reference count for a request.
2783 void ptlrpc_req_put(struct ptlrpc_request *request)
2785 __ptlrpc_req_put(request, 0);
2787 EXPORT_SYMBOL(ptlrpc_req_put);
2791 * Returns xid of a \a request
2793 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2795 return request->rq_xid;
2797 EXPORT_SYMBOL(ptlrpc_req_xid);
2800 * Disengage the client's reply buffer from the network
2801 * NB does _NOT_ unregister any client-side bulk.
2802 * IDEMPOTENT, but _not_ safe against concurrent callers.
2803 * The request owner (i.e. the thread doing the I/O) must call...
2804 * Returns 0 on success or 1 if unregistering cannot be made.
2806 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2812 LASSERT(!in_interrupt());
2814 /* Let's setup deadline for reply unlink. */
2815 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2816 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2817 request->rq_reply_deadline = ktime_get_real_seconds() +
2818 PTLRPC_REQ_LONG_UNLINK;
2821 * Nothing left to do.
2823 if (!ptlrpc_cli_wait_unlink(request))
2826 LNetMDUnlink(request->rq_reply_md_h);
2828 spin_lock(&request->rq_lock);
2829 discard = request->rq_reply_unlinked && !request->rq_req_unlinked;
2830 spin_unlock(&request->rq_lock);
2832 if (discard) /* Discard the request-out callback */
2833 __LNetMDUnlink(request->rq_req_md_h, discard);
2836 * Let's check it once again.
2838 if (!ptlrpc_cli_wait_unlink(request))
2841 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2842 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2845 * Do not wait for unlink to finish.
2851 * We have to wait_event_idle_timeout() whatever the result, to get
2852 * a chance to run reply_in_callback(), and to make sure we've
2853 * unlinked before returning a req to the pool.
2856 wait_queue_head_t *wq = (request->rq_set) ?
2857 &request->rq_set->set_waitq :
2858 &request->rq_reply_waitq;
2859 int seconds = PTLRPC_REQ_LONG_UNLINK;
2861 * Network access will complete in finite time but the HUGE
2862 * timeout lets us CWARN for visibility of sluggish NALs
2864 while (seconds > 0 &&
2865 wait_event_idle_timeout(
2867 !ptlrpc_cli_wait_unlink(request),
2868 cfs_time_seconds(1)) == 0)
2871 ptlrpc_rqphase_move(request, request->rq_next_phase);
2875 DEBUG_REQ(D_WARNING, request,
2876 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2877 request->rq_receiving_reply,
2878 request->rq_req_unlinked,
2879 request->rq_reply_unlinked);
2884 static void ptlrpc_free_request(struct ptlrpc_request *req)
2886 spin_lock(&req->rq_lock);
2888 spin_unlock(&req->rq_lock);
2890 if (req->rq_commit_cb)
2891 req->rq_commit_cb(req);
2892 list_del_init(&req->rq_replay_list);
2894 __ptlrpc_req_put(req, 1);
2898 * the request is committed and dropped from the replay list of its import
2900 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2902 struct obd_import *imp = req->rq_import;
2904 spin_lock(&imp->imp_lock);
2905 if (list_empty(&req->rq_replay_list)) {
2906 spin_unlock(&imp->imp_lock);
2910 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2911 if (imp->imp_replay_cursor == &req->rq_replay_list)
2912 imp->imp_replay_cursor = req->rq_replay_list.next;
2913 ptlrpc_free_request(req);
2916 spin_unlock(&imp->imp_lock);
2918 EXPORT_SYMBOL(ptlrpc_request_committed);
2921 * Iterates through replay_list on import and prunes
2922 * all requests have transno smaller than last_committed for the
2923 * import and don't have rq_replay set.
2924 * Since requests are sorted in transno order, stops when meeting first
2925 * transno bigger than last_committed.
2926 * caller must hold imp->imp_lock
2928 void ptlrpc_free_committed(struct obd_import *imp)
2930 struct ptlrpc_request *req, *saved;
2931 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2932 bool skip_committed_list = true;
2933 unsigned int replay_scanned = 0, replay_freed = 0;
2934 unsigned int commit_scanned = 0, commit_freed = 0;
2935 unsigned int debug_level = D_INFO;
2936 __u64 peer_committed_transno;
2938 time64_t start, now;
2941 LASSERT(imp != NULL);
2942 assert_spin_locked(&imp->imp_lock);
2944 start = ktime_get_seconds();
2945 /* save these here, we can potentially drop imp_lock after checking */
2946 peer_committed_transno = imp->imp_peer_committed_transno;
2947 imp_generation = imp->imp_generation;
2949 if (peer_committed_transno == imp->imp_last_transno_checked &&
2950 imp_generation == imp->imp_last_generation_checked) {
2951 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2952 imp->imp_obd->obd_name, peer_committed_transno);
2955 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2956 imp->imp_obd->obd_name, peer_committed_transno, imp_generation);
2958 if (imp_generation != imp->imp_last_generation_checked ||
2959 imp->imp_last_transno_checked == 0)
2960 skip_committed_list = false;
2961 /* maybe drop imp_lock here, if another lock protected the lists */
2963 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2965 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2966 LASSERT(req != last_req);
2969 if (req->rq_transno == 0) {
2970 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2974 /* If other threads are waiting on imp_lock, stop processing
2975 * in this thread. Another thread can finish remaining work.
2976 * This may happen if there are huge numbers of open files
2977 * that are closed suddenly or evicted, or if the server
2978 * commit interval is very high vs. RPC rate.
2980 if (++replay_scanned % 2048 == 0) {
2981 now = ktime_get_seconds();
2982 if (now > start + 5)
2983 debug_level = D_WARNING;
2985 if ((replay_freed > 128 && now > start + 3) &&
2986 atomic_read(&imp->imp_waiting)) {
2987 if (debug_level == D_INFO)
2988 debug_level = D_RPCTRACE;
2993 if (req->rq_import_generation < imp_generation) {
2994 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2998 /* not yet committed */
2999 if (req->rq_transno > peer_committed_transno) {
3000 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
3004 if (req->rq_replay) {
3005 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
3006 list_move_tail(&req->rq_replay_list,
3007 &imp->imp_committed_list);
3011 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
3012 peer_committed_transno);
3015 ptlrpc_free_request(req);
3018 if (skip_committed_list)
3021 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
3023 LASSERT(req->rq_transno != 0);
3025 /* If other threads are waiting on imp_lock, stop processing
3026 * in this thread. Another thread can finish remaining work. */
3027 if (++commit_scanned % 2048 == 0) {
3028 now = ktime_get_seconds();
3029 if (now > start + 6)
3030 debug_level = D_WARNING;
3032 if ((commit_freed > 128 && now > start + 4) &&
3033 atomic_read(&imp->imp_waiting)) {
3034 if (debug_level == D_INFO)
3035 debug_level = D_RPCTRACE;
3040 if (req->rq_import_generation < imp_generation ||
3042 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
3043 req->rq_import_generation <
3044 imp_generation ? "stale" : "closed");
3046 if (imp->imp_replay_cursor == &req->rq_replay_list)
3047 imp->imp_replay_cursor =
3048 req->rq_replay_list.next;
3051 ptlrpc_free_request(req);
3055 /* if full lists processed without interruption, avoid next scan */
3056 if (debug_level == D_INFO) {
3057 imp->imp_last_transno_checked = peer_committed_transno;
3058 imp->imp_last_generation_checked = imp_generation;
3061 CDEBUG_LIMIT(debug_level,
3062 "%s: %s: skip=%u replay=%u/%u committed=%u/%u\n",
3063 imp->imp_obd->obd_name,
3064 debug_level == D_INFO ? "normal" : "overloaded",
3065 skip_committed_list, replay_freed, replay_scanned,
3066 commit_freed, commit_scanned);
3070 void ptlrpc_cleanup_client(struct obd_import *imp)
3077 * Schedule previously sent request for resend.
3078 * For bulk requests we assign new xid (to avoid problems with
3079 * lost replies and therefore several transfers landing into same buffer
3080 * from different sending attempts).
3082 void ptlrpc_resend_req(struct ptlrpc_request *req)
3084 DEBUG_REQ(D_HA, req, "going to resend");
3085 spin_lock(&req->rq_lock);
3088 * Request got reply but linked to the import list still.
3089 * Let ptlrpc_check_set() process it.
3091 if (ptlrpc_client_replied(req)) {
3092 spin_unlock(&req->rq_lock);
3093 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
3097 req->rq_status = -EAGAIN;
3100 req->rq_net_err = 0;
3101 req->rq_timedout = 0;
3103 ptlrpc_client_wake_req(req);
3104 spin_unlock(&req->rq_lock);
3107 /* XXX: this function and rq_status are currently unused */
3108 void ptlrpc_restart_req(struct ptlrpc_request *req)
3110 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
3111 req->rq_status = -ERESTARTSYS;
3113 spin_lock(&req->rq_lock);
3114 req->rq_restart = 1;
3115 req->rq_timedout = 0;
3116 ptlrpc_client_wake_req(req);
3117 spin_unlock(&req->rq_lock);
3121 * Grab additional reference on a request \a req
3123 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
3126 atomic_inc(&req->rq_refcount);
3129 EXPORT_SYMBOL(ptlrpc_request_addref);
3132 * Add a request to import replay_list.
3133 * Must be called under imp_lock
3135 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3136 struct obd_import *imp)
3138 struct ptlrpc_request *iter;
3140 assert_spin_locked(&imp->imp_lock);
3142 if (req->rq_transno == 0) {
3143 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3148 * clear this for new requests that were resent as well
3149 * as resent replayed requests.
3151 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3153 /* don't re-add requests that have been replayed */
3154 if (!list_empty(&req->rq_replay_list))
3157 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3159 spin_lock(&req->rq_lock);
3161 spin_unlock(&req->rq_lock);
3163 LASSERT(imp->imp_replayable);
3164 /* Balanced in ptlrpc_free_committed, usually. */
3165 ptlrpc_request_addref(req);
3166 list_for_each_entry_reverse(iter, &imp->imp_replay_list,
3169 * We may have duplicate transnos if we create and then
3170 * open a file, or for closes retained if to match creating
3171 * opens, so use req->rq_xid as a secondary key.
3172 * (See bugs 684, 685, and 428.)
3173 * XXX no longer needed, but all opens need transnos!
3175 if (iter->rq_transno > req->rq_transno)
3178 if (iter->rq_transno == req->rq_transno) {
3179 LASSERT(iter->rq_xid != req->rq_xid);
3180 if (iter->rq_xid > req->rq_xid)
3184 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3188 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3192 * Send request and wait until it completes.
3193 * Returns request processing status.
3195 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3197 struct ptlrpc_request_set *set;
3201 LASSERT(req->rq_set == NULL);
3202 LASSERT(!req->rq_receiving_reply);
3204 set = ptlrpc_prep_set();
3206 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3210 /* for distributed debugging */
3211 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3213 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3214 ptlrpc_request_addref(req);
3215 ptlrpc_set_add_req(set, req);
3216 rc = ptlrpc_set_wait(NULL, set);
3217 ptlrpc_set_destroy(set);
3221 EXPORT_SYMBOL(ptlrpc_queue_wait);
3224 * Callback used for replayed requests reply processing.
3225 * In case of successful reply calls registered request replay callback.
3226 * In case of error restart replay process.
3228 static int ptlrpc_replay_interpret(const struct lu_env *env,
3229 struct ptlrpc_request *req,
3232 struct ptlrpc_replay_async_args *aa = args;
3233 struct obd_import *imp = req->rq_import;
3236 atomic_dec(&imp->imp_replay_inflight);
3239 * Note: if it is bulk replay (MDS-MDS replay), then even if
3240 * server got the request, but bulk transfer timeout, let's
3241 * replay the bulk req again
3243 if (!ptlrpc_client_replied(req) ||
3245 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3246 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3247 GOTO(out, rc = -ETIMEDOUT);
3250 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3251 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3252 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3253 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3255 /** VBR: check version failure */
3256 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3257 /** replay was failed due to version mismatch */
3258 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3259 spin_lock(&imp->imp_lock);
3260 imp->imp_vbr_failed = 1;
3261 spin_unlock(&imp->imp_lock);
3262 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3264 /** The transno had better not change over replay. */
3265 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3266 lustre_msg_get_transno(req->rq_repmsg) ||
3267 lustre_msg_get_transno(req->rq_repmsg) == 0,
3269 lustre_msg_get_transno(req->rq_reqmsg),
3270 lustre_msg_get_transno(req->rq_repmsg));
3273 spin_lock(&imp->imp_lock);
3274 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3275 spin_unlock(&imp->imp_lock);
3276 LASSERT(imp->imp_last_replay_transno);
3278 /* transaction number shouldn't be bigger than the latest replayed */
3279 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3280 DEBUG_REQ(D_ERROR, req,
3281 "Reported transno=%llu is bigger than replayed=%llu",
3283 lustre_msg_get_transno(req->rq_reqmsg));
3284 GOTO(out, rc = -EINVAL);
3287 DEBUG_REQ(D_HA, req, "got reply");
3289 /* let the callback do fixups, possibly including in the request */
3290 if (req->rq_replay_cb)
3291 req->rq_replay_cb(req);
3293 if (ptlrpc_client_replied(req) &&
3294 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3295 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3296 lustre_msg_get_status(req->rq_repmsg),
3297 aa->praa_old_status);
3300 * Note: If the replay fails for MDT-MDT recovery, let's
3301 * abort all of the following requests in the replay
3302 * and sending list, because MDT-MDT update requests
3303 * are dependent on each other, see LU-7039
3305 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3306 struct ptlrpc_request *free_req;
3307 struct ptlrpc_request *tmp;
3309 spin_lock(&imp->imp_lock);
3310 list_for_each_entry_safe(free_req, tmp,
3311 &imp->imp_replay_list,
3313 ptlrpc_free_request(free_req);
3316 list_for_each_entry_safe(free_req, tmp,
3317 &imp->imp_committed_list,
3319 ptlrpc_free_request(free_req);
3322 list_for_each_entry_safe(free_req, tmp,
3323 &imp->imp_delayed_list,
3325 spin_lock(&free_req->rq_lock);
3326 free_req->rq_err = 1;
3327 free_req->rq_status = -EIO;
3328 ptlrpc_client_wake_req(free_req);
3329 spin_unlock(&free_req->rq_lock);
3332 list_for_each_entry_safe(free_req, tmp,
3333 &imp->imp_sending_list,
3335 spin_lock(&free_req->rq_lock);
3336 free_req->rq_err = 1;
3337 free_req->rq_status = -EIO;
3338 ptlrpc_client_wake_req(free_req);
3339 spin_unlock(&free_req->rq_lock);
3341 spin_unlock(&imp->imp_lock);
3344 /* Put it back for re-replay. */
3345 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3349 * Errors while replay can set transno to 0, but
3350 * imp_last_replay_transno shouldn't be set to 0 anyway
3352 if (req->rq_transno == 0)
3353 CERROR("Transno is 0 during replay!\n");
3355 /* continue with recovery */
3356 rc = ptlrpc_import_recovery_state_machine(imp);
3358 req->rq_send_state = aa->praa_old_state;
3361 /* this replay failed, so restart recovery */
3362 ptlrpc_connect_import(imp);
3368 * Prepares and queues request for replay.
3369 * Adds it to ptlrpcd queue for actual sending.
3370 * Returns 0 on success.
3372 int ptlrpc_replay_req(struct ptlrpc_request *req)
3374 struct ptlrpc_replay_async_args *aa;
3378 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3380 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_REPLAY_PAUSE, cfs_fail_val);
3382 aa = ptlrpc_req_async_args(aa, req);
3383 memset(aa, 0, sizeof(*aa));
3385 /* Prepare request to be resent with ptlrpcd */
3386 aa->praa_old_state = req->rq_send_state;
3387 req->rq_send_state = LUSTRE_IMP_REPLAY;
3388 req->rq_phase = RQ_PHASE_NEW;
3389 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3391 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3393 req->rq_interpret_reply = ptlrpc_replay_interpret;
3394 /* Readjust the timeout for current conditions */
3395 ptlrpc_at_set_req_timeout(req);
3397 /* Tell server net_latency to calculate how long to wait for reply. */
3398 lustre_msg_set_service_timeout(req->rq_reqmsg,
3399 ptlrpc_at_get_net_latency(req));
3400 DEBUG_REQ(D_HA, req, "REPLAY");
3402 atomic_inc(&req->rq_import->imp_replay_inflight);
3403 spin_lock(&req->rq_lock);
3404 req->rq_early_free_repbuf = 0;
3405 spin_unlock(&req->rq_lock);
3406 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3408 ptlrpcd_add_req(req);
3413 * Aborts all in-flight request on import \a imp sending and delayed lists
3415 void ptlrpc_abort_inflight(struct obd_import *imp)
3417 struct ptlrpc_request *req;
3421 * Make sure that no new requests get processed for this import.
3422 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3423 * this flag and then putting requests on sending_list or delayed_list.
3425 assert_spin_locked(&imp->imp_lock);
3428 * XXX locking? Maybe we should remove each request with the list
3429 * locked? Also, how do we know if the requests on the list are
3430 * being freed at this time?
3432 list_for_each_entry(req, &imp->imp_sending_list, rq_list) {
3433 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3435 spin_lock(&req->rq_lock);
3436 if (req->rq_import_generation < imp->imp_generation) {
3438 req->rq_status = -EIO;
3439 ptlrpc_client_wake_req(req);
3441 spin_unlock(&req->rq_lock);
3444 list_for_each_entry(req, &imp->imp_delayed_list, rq_list) {
3445 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3447 spin_lock(&req->rq_lock);
3448 if (req->rq_import_generation < imp->imp_generation) {
3450 req->rq_status = -EIO;
3451 ptlrpc_client_wake_req(req);
3453 spin_unlock(&req->rq_lock);
3457 * Last chance to free reqs left on the replay list, but we
3458 * will still leak reqs that haven't committed.
3460 if (imp->imp_replayable)
3461 ptlrpc_free_committed(imp);
3467 * Abort all uncompleted requests in request set \a set
3469 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3471 struct ptlrpc_request *req;
3473 LASSERT(set != NULL);
3475 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
3476 spin_lock(&req->rq_lock);
3477 if (req->rq_phase != RQ_PHASE_RPC) {
3478 spin_unlock(&req->rq_lock);
3483 req->rq_status = -EINTR;
3484 ptlrpc_client_wake_req(req);
3485 spin_unlock(&req->rq_lock);
3490 * Initialize the XID for the node. This is common among all requests on
3491 * this node, and only requires the property that it is monotonically
3492 * increasing. It does not need to be sequential. Since this is also used
3493 * as the RDMA match bits, it is important that a single client NOT have
3494 * the same match bits for two different in-flight requests, hence we do
3495 * NOT want to have an XID per target or similar.
3497 * To avoid an unlikely collision between match bits after a client reboot
3498 * (which would deliver old data into the wrong RDMA buffer) initialize
3499 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3500 * If the time is clearly incorrect, we instead use a 62-bit random number.
3501 * In the worst case the random number will overflow 1M RPCs per second in
3502 * 9133 years, or permutations thereof.
3504 #define YEAR_2004 (1ULL << 30)
3505 void ptlrpc_init_xid(void)
3507 time64_t now = ktime_get_real_seconds();
3510 if (now < YEAR_2004) {
3511 get_random_bytes(&xid, sizeof(xid));
3513 xid |= (1ULL << 61);
3515 xid = (u64)now << 20;
3518 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3519 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3521 xid &= PTLRPC_BULK_OPS_MASK;
3522 atomic64_set(&ptlrpc_last_xid, xid);
3526 * Increase xid and returns resulting new value to the caller.
3528 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3529 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3530 * itself uses the last bulk xid needed, so the server can determine the
3531 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3532 * xid must align to a power-of-two value.
3534 * This is assumed to be true due to the initial ptlrpc_last_xid
3535 * value also being initialized to a power-of-two value. LU-1431
3537 __u64 ptlrpc_next_xid(void)
3539 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3543 * If request has a new allocated XID (new request or EINPROGRESS resend),
3544 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3545 * request to ensure previous bulk fails and avoid problems with lost replies
3546 * and therefore several transfers landing into the same buffer from different
3548 * Also, to avoid previous reply landing to a different sending attempt.
3550 void ptlrpc_set_mbits(struct ptlrpc_request *req)
3552 int md_count = req->rq_bulk ? req->rq_bulk->bd_md_count : 1;
3555 * Generate new matchbits for all resend requests, including
3558 if (req->rq_resend) {
3559 __u64 old_mbits = req->rq_mbits;
3562 * First time resend on -EINPROGRESS will generate new xid,
3563 * so we can actually use the rq_xid as rq_mbits in such case,
3564 * however, it's bit hard to distinguish such resend with a
3565 * 'resend for the -EINPROGRESS resend'. To make it simple,
3566 * we opt to generate mbits for all resend cases.
3568 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3570 req->rq_mbits = ptlrpc_next_xid();
3573 * Old version transfers rq_xid to peer as
3576 spin_lock(&req->rq_import->imp_lock);
3577 list_del_init(&req->rq_unreplied_list);
3578 ptlrpc_assign_next_xid_nolock(req);
3579 spin_unlock(&req->rq_import->imp_lock);
3580 req->rq_mbits = req->rq_xid;
3582 CDEBUG(D_HA, "resend with new mbits old x%llu new x%llu\n",
3583 old_mbits, req->rq_mbits);
3584 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3585 /* Request being sent first time, use xid as matchbits. */
3586 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3587 BULK_MBITS) || req->rq_mbits == 0)
3589 req->rq_mbits = req->rq_xid;
3591 req->rq_mbits -= md_count - 1;
3595 * Replay request, xid and matchbits have already been
3596 * correctly assigned.
3602 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3603 * that server can infer the number of bulks that were prepared,
3606 req->rq_mbits += md_count - 1;
3609 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3610 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3612 * It's ok to directly set the rq_xid here, since this xid bump
3613 * won't affect the request position in unreplied list.
3615 if (!OCD_HAS_FLAG(&req->rq_import->imp_connect_data, BULK_MBITS))
3616 req->rq_xid = req->rq_mbits;
3620 * Get a glimpse at what next xid value might have been.
3621 * Returns possible next xid.
3623 __u64 ptlrpc_sample_next_xid(void)
3625 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3627 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3630 * Functions for operating ptlrpc workers.
3632 * A ptlrpc work is a function which will be running inside ptlrpc context.
3633 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3635 * 1. after a work is created, it can be used many times, that is:
3636 * handler = ptlrpcd_alloc_work();
3637 * ptlrpcd_queue_work();
3639 * queue it again when necessary:
3640 * ptlrpcd_queue_work();
3641 * ptlrpcd_destroy_work();
3642 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3643 * it will only be queued once in any time. Also as its name implies, it may
3644 * have delay before it really runs by ptlrpcd thread.
3646 struct ptlrpc_work_async_args {
3647 int (*cb)(const struct lu_env *, void *);
3651 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3653 /* re-initialize the req */
3654 req->rq_timeout = obd_timeout;
3655 req->rq_sent = ktime_get_real_seconds();
3656 req->rq_deadline = req->rq_sent + req->rq_timeout;
3657 req->rq_phase = RQ_PHASE_INTERPRET;
3658 req->rq_next_phase = RQ_PHASE_COMPLETE;
3659 req->rq_xid = ptlrpc_next_xid();
3660 req->rq_import_generation = req->rq_import->imp_generation;
3662 ptlrpcd_add_req(req);
3665 static int work_interpreter(const struct lu_env *env,
3666 struct ptlrpc_request *req, void *args, int rc)
3668 struct ptlrpc_work_async_args *arg = args;
3670 LASSERT(ptlrpcd_check_work(req));
3671 LASSERT(arg->cb != NULL);
3673 rc = arg->cb(env, arg->cbdata);
3675 list_del_init(&req->rq_set_chain);
3678 if (atomic_dec_return(&req->rq_refcount) > 1) {
3679 atomic_set(&req->rq_refcount, 2);
3680 ptlrpcd_add_work_req(req);
3685 static int worker_format;
3687 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3689 return req->rq_pill.rc_fmt == (void *)&worker_format;
3693 * Create a work for ptlrpc.
3695 void *ptlrpcd_alloc_work(struct obd_import *imp,
3696 int (*cb)(const struct lu_env *, void *), void *cbdata)
3698 struct ptlrpc_request *req = NULL;
3699 struct ptlrpc_work_async_args *args;
3705 RETURN(ERR_PTR(-EINVAL));
3707 /* copy some code from deprecated fakereq. */
3708 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3710 CERROR("ptlrpc: run out of memory!\n");
3711 RETURN(ERR_PTR(-ENOMEM));
3714 ptlrpc_cli_req_init(req);
3716 req->rq_send_state = LUSTRE_IMP_FULL;
3717 req->rq_type = PTL_RPC_MSG_REQUEST;
3718 req->rq_import = class_import_get(imp);
3719 req->rq_interpret_reply = work_interpreter;
3720 /* don't want reply */
3721 req->rq_no_delay = req->rq_no_resend = 1;
3722 req->rq_pill.rc_fmt = (void *)&worker_format;
3724 args = ptlrpc_req_async_args(args, req);
3726 args->cbdata = cbdata;
3730 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3732 void ptlrpcd_destroy_work(void *handler)
3734 struct ptlrpc_request *req = handler;
3737 ptlrpc_req_finished(req);
3739 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3741 int ptlrpcd_queue_work(void *handler)
3743 struct ptlrpc_request *req = handler;
3746 * Check if the req is already being queued.
3748 * Here comes a trick: it lacks a way of checking if a req is being
3749 * processed reliably in ptlrpc. Here I have to use refcount of req
3750 * for this purpose. This is okay because the caller should use this
3751 * req as opaque data. - Jinshan
3753 LASSERT(atomic_read(&req->rq_refcount) > 0);
3754 if (atomic_inc_return(&req->rq_refcount) == 2)
3755 ptlrpcd_add_work_req(req);
3758 EXPORT_SYMBOL(ptlrpcd_queue_work);