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 sptlrpc_enc_pool_put_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 == MDS_CONNECT || opc == MGS_CONNECT) {
1340 /* Suppress timed out reconnect requests */
1341 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1346 * Suppress most unavailable/again reconnect requests, but
1347 * print occasionally so it is clear client is trying to
1348 * connect to a server where no target is running.
1350 if ((err == -ENODEV || err == -EAGAIN) &&
1351 req->rq_import->imp_conn_cnt % 30 != 20)
1355 if (opc == LDLM_ENQUEUE && err == -EAGAIN)
1356 /* -EAGAIN is normal when using POSIX flocks */
1359 if (opc == OBD_PING && (err == -ENODEV || err == -ENOTCONN) &&
1360 (req->rq_xid & 0xf) != 10)
1361 /* Suppress most ping requests, they may fail occasionally */
1368 * Check request processing status.
1369 * Returns the status.
1371 static int ptlrpc_check_status(struct ptlrpc_request *req)
1376 rc = lustre_msg_get_status(req->rq_repmsg);
1377 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1378 struct obd_import *imp = req->rq_import;
1379 struct lnet_nid *nid = &imp->imp_connection->c_peer.nid;
1380 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1382 if (ptlrpc_console_allow(req, opc, rc))
1383 LCONSOLE_ERROR_MSG(0x11,
1384 "%s: operation %s to node %s failed: rc = %d\n",
1385 imp->imp_obd->obd_name,
1387 libcfs_nidstr(nid), rc);
1388 RETURN(rc < 0 ? rc : -EINVAL);
1392 DEBUG_REQ(D_INFO, req, "check status: rc = %d", rc);
1398 * save pre-versions of objects into request for replay.
1399 * Versions are obtained from server reply.
1402 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1404 struct lustre_msg *repmsg = req->rq_repmsg;
1405 struct lustre_msg *reqmsg = req->rq_reqmsg;
1406 __u64 *versions = lustre_msg_get_versions(repmsg);
1409 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1413 lustre_msg_set_versions(reqmsg, versions);
1414 CDEBUG(D_INFO, "Client save versions [%#llx/%#llx]\n",
1415 versions[0], versions[1]);
1420 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1422 struct ptlrpc_request *req;
1424 assert_spin_locked(&imp->imp_lock);
1425 if (list_empty(&imp->imp_unreplied_list))
1428 req = list_first_entry(&imp->imp_unreplied_list, struct ptlrpc_request,
1430 LASSERTF(req->rq_xid >= 1, "XID:%llu\n", req->rq_xid);
1432 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1433 imp->imp_known_replied_xid = req->rq_xid - 1;
1435 return req->rq_xid - 1;
1439 * Callback function called when client receives RPC reply for \a req.
1440 * Returns 0 on success or error code.
1441 * The return alue would be assigned to req->rq_status by the caller
1442 * as request processing status.
1443 * This function also decides if the request needs to be saved for later replay.
1445 static int after_reply(struct ptlrpc_request *req)
1447 struct obd_import *imp = req->rq_import;
1448 struct obd_device *obd = req->rq_import->imp_obd;
1455 LASSERT(obd != NULL);
1456 /* repbuf must be unlinked */
1457 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1459 if (req->rq_reply_truncated) {
1460 if (ptlrpc_no_resend(req)) {
1461 DEBUG_REQ(D_ERROR, req,
1462 "reply buffer overflow, expected=%d, actual size=%d",
1463 req->rq_nob_received, req->rq_repbuf_len);
1467 sptlrpc_cli_free_repbuf(req);
1469 * Pass the required reply buffer size (include
1470 * space for early reply).
1471 * NB: no need to roundup because alloc_repbuf
1474 req->rq_replen = req->rq_nob_received;
1475 req->rq_nob_received = 0;
1476 spin_lock(&req->rq_lock);
1478 spin_unlock(&req->rq_lock);
1482 work_start = ktime_get_real();
1483 timediff = ktime_us_delta(work_start, req->rq_sent_ns);
1486 * NB Until this point, the whole of the incoming message,
1487 * including buflens, status etc is in the sender's byte order.
1489 rc = sptlrpc_cli_unwrap_reply(req);
1491 DEBUG_REQ(D_ERROR, req, "unwrap reply failed: rc = %d", rc);
1496 * Security layer unwrap might ask resend this request.
1501 rc = unpack_reply(req);
1505 /* retry indefinitely on EINPROGRESS */
1506 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1507 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1508 time64_t now = ktime_get_real_seconds();
1510 DEBUG_REQ((req->rq_nr_resend % 8 == 1 ? D_WARNING : 0) |
1511 D_RPCTRACE, req, "resending request on EINPROGRESS");
1512 spin_lock(&req->rq_lock);
1514 spin_unlock(&req->rq_lock);
1515 req->rq_nr_resend++;
1517 /* Readjust the timeout for current conditions */
1518 ptlrpc_at_set_req_timeout(req);
1520 * delay resend to give a chance to the server to get ready.
1521 * The delay is increased by 1s on every resend and is capped to
1522 * the current request timeout (i.e. obd_timeout if AT is off,
1523 * or AT service time x 125% + 5s, see at_est2timeout)
1525 if (req->rq_nr_resend > req->rq_timeout)
1526 req->rq_sent = now + req->rq_timeout;
1528 req->rq_sent = now + req->rq_nr_resend;
1530 /* Resend for EINPROGRESS will use a new XID */
1531 spin_lock(&imp->imp_lock);
1532 list_del_init(&req->rq_unreplied_list);
1533 spin_unlock(&imp->imp_lock);
1538 if (obd->obd_svc_stats) {
1539 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1541 ptlrpc_lprocfs_rpc_sent(req, timediff);
1544 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1545 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1546 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1547 lustre_msg_get_type(req->rq_repmsg));
1551 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1552 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1553 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1554 ptlrpc_at_adj_net_latency(req,
1555 lustre_msg_get_service_timeout(req->rq_repmsg));
1557 rc = ptlrpc_check_status(req);
1561 * Either we've been evicted, or the server has failed for
1562 * some reason. Try to reconnect, and if that fails, punt to
1565 if (ptlrpc_recoverable_error(rc)) {
1566 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1567 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1570 ptlrpc_request_handle_notconn(req);
1575 * Let's look if server sent slv. Do it only for RPC with
1578 ldlm_cli_update_pool(req);
1582 * Store transno in reqmsg for replay.
1584 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1585 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1586 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1589 if (lustre_msg_get_transno(req->rq_repmsg) ||
1590 lustre_msg_get_opc(req->rq_reqmsg) == LDLM_ENQUEUE)
1591 imp->imp_no_cached_data = 0;
1593 if (imp->imp_replayable) {
1594 /* if other threads are waiting for ptlrpc_free_committed()
1595 * they could continue the work of freeing RPCs. That reduces
1596 * lock hold times, and distributes work more fairly across
1597 * waiting threads. We can't use spin_is_contended() since
1598 * there are many other places where imp_lock is held.
1600 atomic_inc(&imp->imp_waiting);
1601 spin_lock(&imp->imp_lock);
1602 atomic_dec(&imp->imp_waiting);
1604 * No point in adding already-committed requests to the replay
1605 * list, we will just remove them immediately. b=9829
1607 if (req->rq_transno != 0 &&
1609 lustre_msg_get_last_committed(req->rq_repmsg) ||
1611 /** version recovery */
1612 ptlrpc_save_versions(req);
1613 ptlrpc_retain_replayable_request(req, imp);
1614 } else if (req->rq_commit_cb &&
1615 list_empty(&req->rq_replay_list)) {
1617 * NB: don't call rq_commit_cb if it's already on
1618 * rq_replay_list, ptlrpc_free_committed() will call
1619 * it later, see LU-3618 for details
1621 spin_unlock(&imp->imp_lock);
1622 req->rq_commit_cb(req);
1623 atomic_inc(&imp->imp_waiting);
1624 spin_lock(&imp->imp_lock);
1625 atomic_dec(&imp->imp_waiting);
1629 * Replay-enabled imports return commit-status information.
1631 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1632 if (likely(committed > imp->imp_peer_committed_transno))
1633 imp->imp_peer_committed_transno = committed;
1635 ptlrpc_free_committed(imp);
1637 if (!list_empty(&imp->imp_replay_list)) {
1638 struct ptlrpc_request *last;
1640 last = list_entry(imp->imp_replay_list.prev,
1641 struct ptlrpc_request,
1644 * Requests with rq_replay stay on the list even if no
1645 * commit is expected.
1647 if (last->rq_transno > imp->imp_peer_committed_transno)
1648 ptlrpc_pinger_commit_expected(imp);
1651 spin_unlock(&imp->imp_lock);
1658 * Helper function to send request \a req over the network for the first time
1659 * Also adjusts request phase.
1660 * Returns 0 on success or error code.
1662 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1664 struct obd_import *imp = req->rq_import;
1669 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1671 /* do not try to go further if there is not enough memory in enc_pool */
1672 if (req->rq_sent && req->rq_bulk)
1673 if (req->rq_bulk->bd_iov_count >
1674 sptlrpc_enc_pool_get_free_pages(PAGES_POOL) &&
1675 pool_is_at_full_capacity())
1678 if (req->rq_sent && (req->rq_sent > ktime_get_real_seconds()) &&
1679 (!req->rq_generation_set ||
1680 req->rq_import_generation == imp->imp_generation))
1683 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1685 spin_lock(&imp->imp_lock);
1687 LASSERT(req->rq_xid != 0);
1688 LASSERT(!list_empty(&req->rq_unreplied_list));
1690 if (!req->rq_generation_set)
1691 req->rq_import_generation = imp->imp_generation;
1693 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1694 spin_lock(&req->rq_lock);
1695 req->rq_waiting = 1;
1696 spin_unlock(&req->rq_lock);
1698 DEBUG_REQ(D_HA, req, "req waiting for recovery: (%s != %s)",
1699 ptlrpc_import_state_name(req->rq_send_state),
1700 ptlrpc_import_state_name(imp->imp_state));
1701 LASSERT(list_empty(&req->rq_list));
1702 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1703 atomic_inc(&req->rq_import->imp_inflight);
1704 spin_unlock(&imp->imp_lock);
1709 spin_unlock(&imp->imp_lock);
1710 req->rq_status = rc;
1711 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1715 LASSERT(list_empty(&req->rq_list));
1716 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1717 atomic_inc(&req->rq_import->imp_inflight);
1720 * find the known replied XID from the unreplied list, CONNECT
1721 * and DISCONNECT requests are skipped to make the sanity check
1722 * on server side happy. see process_req_last_xid().
1724 * For CONNECT: Because replay requests have lower XID, it'll
1725 * break the sanity check if CONNECT bump the exp_last_xid on
1728 * For DISCONNECT: Since client will abort inflight RPC before
1729 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1730 * than the inflight RPC.
1732 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1733 min_xid = ptlrpc_known_replied_xid(imp);
1734 spin_unlock(&imp->imp_lock);
1736 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1738 lustre_msg_set_status(req->rq_reqmsg, current->pid);
1740 /* If the request to be sent is an LDLM callback, do not try to
1742 * An LDLM callback is sent by a server to a client in order to make
1743 * it release a lock, on a communication channel that uses a reverse
1744 * context. It cannot be refreshed on its own, as it is the 'reverse'
1745 * (server-side) representation of a client context.
1746 * We do not care if the reverse context is expired, and want to send
1747 * the LDLM callback anyway. Once the client receives the AST, it is
1748 * its job to refresh its own context if it has expired, hence
1749 * refreshing the associated reverse context on server side, before
1750 * being able to send the LDLM_CANCEL requested by the server.
1752 if (lustre_msg_get_opc(req->rq_reqmsg) != LDLM_BL_CALLBACK &&
1753 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_CP_CALLBACK &&
1754 lustre_msg_get_opc(req->rq_reqmsg) != LDLM_GL_CALLBACK)
1755 rc = sptlrpc_req_refresh_ctx(req, 0);
1758 req->rq_status = rc;
1761 spin_lock(&req->rq_lock);
1762 req->rq_wait_ctx = 1;
1763 spin_unlock(&req->rq_lock);
1769 "Sending RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
1771 imp->imp_obd->obd_uuid.uuid,
1772 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1773 obd_import_nid2str(imp), lustre_msg_get_opc(req->rq_reqmsg),
1774 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
1776 rc = ptl_send_rpc(req, 0);
1777 if (rc == -ENOMEM) {
1778 spin_lock(&imp->imp_lock);
1779 if (!list_empty(&req->rq_list)) {
1780 list_del_init(&req->rq_list);
1781 if (atomic_dec_and_test(&req->rq_import->imp_inflight))
1782 wake_up(&req->rq_import->imp_recovery_waitq);
1784 spin_unlock(&imp->imp_lock);
1785 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1789 DEBUG_REQ(D_HA, req, "send failed, expect timeout: rc = %d",
1791 spin_lock(&req->rq_lock);
1792 req->rq_net_err = 1;
1793 spin_unlock(&req->rq_lock);
1799 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1804 LASSERT(set->set_producer != NULL);
1806 remaining = atomic_read(&set->set_remaining);
1809 * populate the ->set_requests list with requests until we
1810 * reach the maximum number of RPCs in flight for this set
1812 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1813 rc = set->set_producer(set, set->set_producer_arg);
1814 if (rc == -ENOENT) {
1815 /* no more RPC to produce */
1816 set->set_producer = NULL;
1817 set->set_producer_arg = NULL;
1822 RETURN((atomic_read(&set->set_remaining) - remaining));
1826 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1827 * and no more replies are expected.
1828 * (it is possible to get less replies than requests sent e.g. due to timed out
1829 * requests or requests that we had trouble to send out)
1831 * NOTE: This function contains a potential schedule point (cond_resched()).
1833 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1835 struct ptlrpc_request *req, *next;
1836 LIST_HEAD(comp_reqs);
1837 int force_timer_recalc = 0;
1840 if (atomic_read(&set->set_remaining) == 0)
1843 list_for_each_entry_safe(req, next, &set->set_requests,
1845 struct obd_import *imp = req->rq_import;
1846 int unregistered = 0;
1850 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1851 list_move_tail(&req->rq_set_chain, &comp_reqs);
1856 * This schedule point is mainly for the ptlrpcd caller of this
1857 * function. Most ptlrpc sets are not long-lived and unbounded
1858 * in length, but at the least the set used by the ptlrpcd is.
1859 * Since the processing time is unbounded, we need to insert an
1860 * explicit schedule point to make the thread well-behaved.
1865 * If the caller requires to allow to be interpreted by force
1866 * and it has really been interpreted, then move the request
1867 * to RQ_PHASE_INTERPRET phase in spite of what the current
1870 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1871 req->rq_status = -EINTR;
1872 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1875 * Since it is interpreted and we have to wait for
1876 * the reply to be unlinked, then use sync mode.
1880 GOTO(interpret, req->rq_status);
1883 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1884 force_timer_recalc = 1;
1886 /* delayed send - skip */
1887 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1890 /* delayed resend - skip */
1891 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1892 req->rq_sent > ktime_get_real_seconds())
1895 if (!(req->rq_phase == RQ_PHASE_RPC ||
1896 req->rq_phase == RQ_PHASE_BULK ||
1897 req->rq_phase == RQ_PHASE_INTERPRET ||
1898 req->rq_phase == RQ_PHASE_UNREG_RPC ||
1899 req->rq_phase == RQ_PHASE_UNREG_BULK)) {
1900 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1904 if (req->rq_phase == RQ_PHASE_UNREG_RPC ||
1905 req->rq_phase == RQ_PHASE_UNREG_BULK) {
1906 LASSERT(req->rq_next_phase != req->rq_phase);
1907 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1909 if (req->rq_req_deadline &&
1910 !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REQ_UNLINK))
1911 req->rq_req_deadline = 0;
1912 if (req->rq_reply_deadline &&
1913 !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK))
1914 req->rq_reply_deadline = 0;
1915 if (req->rq_bulk_deadline &&
1916 !CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK))
1917 req->rq_bulk_deadline = 0;
1920 * Skip processing until reply is unlinked. We
1921 * can't return to pool before that and we can't
1922 * call interpret before that. We need to make
1923 * sure that all rdma transfers finished and will
1924 * not corrupt any data.
1926 if (req->rq_phase == RQ_PHASE_UNREG_RPC &&
1927 ptlrpc_cli_wait_unlink(req))
1929 if (req->rq_phase == RQ_PHASE_UNREG_BULK &&
1930 ptlrpc_client_bulk_active(req))
1934 * Turn fail_loc off to prevent it from looping
1937 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1938 CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1941 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1942 CFS_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1947 * Move to next phase if reply was successfully
1950 ptlrpc_rqphase_move(req, req->rq_next_phase);
1953 if (req->rq_phase == RQ_PHASE_INTERPRET)
1954 GOTO(interpret, req->rq_status);
1957 * Note that this also will start async reply unlink.
1959 if (req->rq_net_err && !req->rq_timedout) {
1960 ptlrpc_expire_one_request(req, 1);
1963 * Check if we still need to wait for unlink.
1965 if (ptlrpc_cli_wait_unlink(req) ||
1966 ptlrpc_client_bulk_active(req))
1968 /* If there is no need to resend, fail it now. */
1969 if (req->rq_no_resend) {
1970 if (req->rq_status == 0)
1971 req->rq_status = -EIO;
1972 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1973 GOTO(interpret, req->rq_status);
1980 if (!ptlrpc_unregister_reply(req, 1)) {
1981 ptlrpc_unregister_bulk(req, 1);
1985 spin_lock(&req->rq_lock);
1986 req->rq_replied = 0;
1987 spin_unlock(&req->rq_lock);
1988 if (req->rq_status == 0)
1989 req->rq_status = -EIO;
1990 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1991 GOTO(interpret, req->rq_status);
1995 * ptlrpc_set_wait uses l_wait_event_abortable_timeout()
1996 * so it sets rq_intr regardless of individual rpc
1997 * timeouts. The synchronous IO waiting path sets
1998 * rq_intr irrespective of whether ptlrpcd
1999 * has seen a timeout. Our policy is to only interpret
2000 * interrupted rpcs after they have timed out, so we
2001 * need to enforce that here.
2004 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
2005 req->rq_wait_ctx)) {
2006 req->rq_status = -EINTR;
2007 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2008 GOTO(interpret, req->rq_status);
2011 if (req->rq_phase == RQ_PHASE_RPC) {
2012 if (req->rq_timedout || req->rq_resend ||
2013 req->rq_waiting || req->rq_wait_ctx) {
2016 if (!ptlrpc_unregister_reply(req, 1)) {
2017 ptlrpc_unregister_bulk(req, 1);
2021 spin_lock(&imp->imp_lock);
2022 if (ptlrpc_import_delay_req(imp, req,
2025 * put on delay list - only if we wait
2026 * recovery finished - before send
2028 list_move_tail(&req->rq_list,
2029 &imp->imp_delayed_list);
2030 spin_unlock(&imp->imp_lock);
2035 req->rq_status = status;
2036 ptlrpc_rqphase_move(req,
2037 RQ_PHASE_INTERPRET);
2038 spin_unlock(&imp->imp_lock);
2039 GOTO(interpret, req->rq_status);
2041 /* ignore on just initiated connections */
2042 if (ptlrpc_no_resend(req) &&
2043 !req->rq_wait_ctx &&
2044 imp->imp_generation !=
2045 imp->imp_initiated_at) {
2046 req->rq_status = -ENOTCONN;
2047 ptlrpc_rqphase_move(req,
2048 RQ_PHASE_INTERPRET);
2049 spin_unlock(&imp->imp_lock);
2050 GOTO(interpret, req->rq_status);
2053 /* don't resend too fast in case of network
2056 if (ktime_get_real_seconds() < (req->rq_sent + 1)
2057 && req->rq_net_err && req->rq_timedout) {
2059 DEBUG_REQ(D_INFO, req,
2060 "throttle request");
2061 /* Don't try to resend RPC right away
2062 * as it is likely it will fail again
2063 * and ptlrpc_check_set() will be
2064 * called again, keeping this thread
2065 * busy. Instead, wait for the next
2066 * timeout. Flag it as resend to
2067 * ensure we don't wait to long.
2070 spin_unlock(&imp->imp_lock);
2074 list_move_tail(&req->rq_list,
2075 &imp->imp_sending_list);
2077 spin_unlock(&imp->imp_lock);
2079 spin_lock(&req->rq_lock);
2080 req->rq_waiting = 0;
2081 spin_unlock(&req->rq_lock);
2083 if (req->rq_timedout || req->rq_resend) {
2085 * This is re-sending anyways,
2086 * let's mark req as resend.
2088 spin_lock(&req->rq_lock);
2090 spin_unlock(&req->rq_lock);
2093 * rq_wait_ctx is only touched by ptlrpcd,
2094 * so no lock is needed here.
2096 status = sptlrpc_req_refresh_ctx(req, 0);
2099 req->rq_status = status;
2100 spin_lock(&req->rq_lock);
2101 req->rq_wait_ctx = 0;
2102 spin_unlock(&req->rq_lock);
2103 force_timer_recalc = 1;
2105 spin_lock(&req->rq_lock);
2106 req->rq_wait_ctx = 1;
2107 spin_unlock(&req->rq_lock);
2112 spin_lock(&req->rq_lock);
2113 req->rq_wait_ctx = 0;
2114 spin_unlock(&req->rq_lock);
2118 * In any case, the previous bulk should be
2119 * cleaned up to prepare for the new sending
2122 !ptlrpc_unregister_bulk(req, 1))
2125 rc = ptl_send_rpc(req, 0);
2126 if (rc == -ENOMEM) {
2127 spin_lock(&imp->imp_lock);
2128 if (!list_empty(&req->rq_list))
2129 list_del_init(&req->rq_list);
2130 spin_unlock(&imp->imp_lock);
2131 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
2135 DEBUG_REQ(D_HA, req,
2136 "send failed: rc = %d", rc);
2137 force_timer_recalc = 1;
2138 spin_lock(&req->rq_lock);
2139 req->rq_net_err = 1;
2140 spin_unlock(&req->rq_lock);
2143 /* need to reset the timeout */
2144 force_timer_recalc = 1;
2147 spin_lock(&req->rq_lock);
2149 if (ptlrpc_client_early(req)) {
2150 ptlrpc_at_recv_early_reply(req);
2151 spin_unlock(&req->rq_lock);
2155 /* Still waiting for a reply? */
2156 if (ptlrpc_client_recv(req)) {
2157 spin_unlock(&req->rq_lock);
2161 /* Did we actually receive a reply? */
2162 if (!ptlrpc_client_replied(req)) {
2163 spin_unlock(&req->rq_lock);
2167 spin_unlock(&req->rq_lock);
2170 * unlink from net because we are going to
2171 * swab in-place of reply buffer
2173 unregistered = ptlrpc_unregister_reply(req, 1);
2177 req->rq_status = after_reply(req);
2178 if (req->rq_resend) {
2179 force_timer_recalc = 1;
2184 * If there is no bulk associated with this request,
2185 * then we're done and should let the interpreter
2186 * process the reply. Similarly if the RPC returned
2187 * an error, and therefore the bulk will never arrive.
2189 if (!req->rq_bulk || req->rq_status < 0) {
2190 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2191 GOTO(interpret, req->rq_status);
2194 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
2197 LASSERT(req->rq_phase == RQ_PHASE_BULK);
2198 if (ptlrpc_client_bulk_active(req))
2201 if (req->rq_bulk->bd_failure) {
2203 * The RPC reply arrived OK, but the bulk screwed
2204 * up! Dead weird since the server told us the RPC
2205 * was good after getting the REPLY for her GET or
2206 * the ACK for her PUT.
2208 DEBUG_REQ(D_ERROR, req, "bulk transfer failed %d/%d/%d",
2210 req->rq_bulk->bd_nob,
2211 req->rq_bulk->bd_nob_transferred);
2212 req->rq_status = -EIO;
2215 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
2218 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
2221 * This moves to "unregistering" phase we need to wait for
2224 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
2225 /* start async bulk unlink too */
2226 ptlrpc_unregister_bulk(req, 1);
2230 if (!ptlrpc_unregister_bulk(req, async))
2234 * When calling interpret receiving already should be
2237 LASSERT(!req->rq_receiving_reply);
2239 ptlrpc_req_interpret(env, req, req->rq_status);
2241 if (ptlrpcd_check_work(req)) {
2242 atomic_dec(&set->set_remaining);
2245 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2249 "Completed RPC req@%p pname:cluuid:pid:xid:nid:opc:job %s:%s:%d:%llu:%s:%d:%s\n",
2251 imp->imp_obd->obd_uuid.uuid,
2252 lustre_msg_get_status(req->rq_reqmsg),
2254 obd_import_nid2str(imp),
2255 lustre_msg_get_opc(req->rq_reqmsg),
2256 lustre_msg_get_jobid(req->rq_reqmsg) ?: "");
2258 spin_lock(&imp->imp_lock);
2260 * Request already may be not on sending or delaying list. This
2261 * may happen in the case of marking it erroneous for the case
2262 * ptlrpc_import_delay_req(req, status) find it impossible to
2263 * allow sending this rpc and returns *status != 0.
2265 if (!list_empty(&req->rq_list)) {
2266 list_del_init(&req->rq_list);
2267 if (atomic_dec_and_test(&imp->imp_inflight))
2268 wake_up(&imp->imp_recovery_waitq);
2270 list_del_init(&req->rq_unreplied_list);
2271 spin_unlock(&imp->imp_lock);
2273 atomic_dec(&set->set_remaining);
2274 wake_up(&imp->imp_recovery_waitq);
2276 if (set->set_producer) {
2277 /* produce a new request if possible */
2278 if (ptlrpc_set_producer(set) > 0)
2279 force_timer_recalc = 1;
2282 * free the request that has just been completed
2283 * in order not to pollute set->set_requests
2285 list_del_init(&req->rq_set_chain);
2286 spin_lock(&req->rq_lock);
2288 req->rq_invalid_rqset = 0;
2289 spin_unlock(&req->rq_lock);
2291 /* record rq_status to compute the final status later */
2292 if (req->rq_status != 0)
2293 set->set_rc = req->rq_status;
2294 ptlrpc_req_finished(req);
2296 list_move_tail(&req->rq_set_chain, &comp_reqs);
2301 * move completed request at the head of list so it's easier for
2302 * caller to find them
2304 list_splice(&comp_reqs, &set->set_requests);
2306 /* If we hit an error, we want to recover promptly. */
2307 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2309 EXPORT_SYMBOL(ptlrpc_check_set);
2312 * Time out request \a req. is \a async_unlink is set, that means do not wait
2313 * until LNet actually confirms network buffer unlinking.
2314 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2316 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2318 struct obd_import *imp = req->rq_import;
2319 unsigned int debug_mask = D_RPCTRACE;
2324 spin_lock(&req->rq_lock);
2325 req->rq_timedout = 1;
2326 spin_unlock(&req->rq_lock);
2328 opc = lustre_msg_get_opc(req->rq_reqmsg);
2329 if (ptlrpc_console_allow(req, opc,
2330 lustre_msg_get_status(req->rq_reqmsg)))
2331 debug_mask = D_WARNING;
2332 DEBUG_REQ(debug_mask, req, "Request sent has %s: [sent %lld/real %lld]",
2333 req->rq_net_err ? "failed due to network error" :
2334 ((req->rq_real_sent == 0 ||
2335 req->rq_real_sent < req->rq_sent ||
2336 req->rq_real_sent >= req->rq_deadline) ?
2337 "timed out for sent delay" : "timed out for slow reply"),
2338 req->rq_sent, req->rq_real_sent);
2340 if (imp && obd_debug_peer_on_timeout)
2341 LNetDebugPeer(&imp->imp_connection->c_peer);
2343 ptlrpc_unregister_reply(req, async_unlink);
2344 ptlrpc_unregister_bulk(req, async_unlink);
2346 if (obd_dump_on_timeout)
2347 libcfs_debug_dumplog();
2350 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2354 atomic_inc(&imp->imp_timeouts);
2356 /* The DLM server doesn't want recovery run on its imports. */
2357 if (imp->imp_dlm_fake)
2361 * If this request is for recovery or other primordial tasks,
2362 * then error it out here.
2364 if (req->rq_ctx_init || req->rq_ctx_fini ||
2365 req->rq_send_state != LUSTRE_IMP_FULL ||
2366 imp->imp_obd->obd_no_recov) {
2367 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2368 ptlrpc_import_state_name(req->rq_send_state),
2369 ptlrpc_import_state_name(imp->imp_state));
2370 spin_lock(&req->rq_lock);
2371 req->rq_status = -ETIMEDOUT;
2373 spin_unlock(&req->rq_lock);
2378 * if a request can't be resent we can't wait for an answer after
2381 if (ptlrpc_no_resend(req)) {
2382 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2386 if (opc != OBD_PING || req->rq_xid > imp->imp_highest_replied_xid)
2387 ptlrpc_fail_import(imp,
2388 lustre_msg_get_conn_cnt(req->rq_reqmsg));
2394 * Time out all uncompleted requests in request set pointed by \a data
2395 * This is called when a wait times out.
2397 void ptlrpc_expired_set(struct ptlrpc_request_set *set)
2399 struct ptlrpc_request *req;
2400 time64_t now = ktime_get_real_seconds();
2403 LASSERT(set != NULL);
2406 * A timeout expired. See which reqs it applies to...
2408 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2409 /* don't expire request waiting for context */
2410 if (req->rq_wait_ctx)
2413 /* Request in-flight? */
2414 if (!((req->rq_phase == RQ_PHASE_RPC &&
2415 !req->rq_waiting && !req->rq_resend) ||
2416 (req->rq_phase == RQ_PHASE_BULK)))
2419 if (req->rq_timedout || /* already dealt with */
2420 req->rq_deadline > now) /* not expired */
2424 * Deal with this guy. Do it asynchronously to not block
2427 ptlrpc_expire_one_request(req, 1);
2429 * Loops require that we resched once in a while to avoid
2430 * RCU stalls and a few other problems.
2438 * Interrupts (sets interrupted flag) all uncompleted requests in
2439 * a set \a data. This is called when a wait_event is interrupted
2442 static void ptlrpc_interrupted_set(struct ptlrpc_request_set *set)
2444 struct ptlrpc_request *req;
2446 LASSERT(set != NULL);
2447 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2449 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2453 if (req->rq_phase != RQ_PHASE_RPC &&
2454 req->rq_phase != RQ_PHASE_UNREG_RPC &&
2455 !req->rq_allow_intr)
2458 spin_lock(&req->rq_lock);
2460 spin_unlock(&req->rq_lock);
2465 * Get the smallest timeout in the set; this does NOT set a timeout.
2467 time64_t ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2469 time64_t now = ktime_get_real_seconds();
2471 struct ptlrpc_request *req;
2475 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2476 /* Request in-flight? */
2477 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2478 (req->rq_phase == RQ_PHASE_BULK) ||
2479 (req->rq_phase == RQ_PHASE_NEW)))
2482 /* Already timed out. */
2483 if (req->rq_timedout)
2486 /* Waiting for ctx. */
2487 if (req->rq_wait_ctx)
2490 if (req->rq_phase == RQ_PHASE_NEW)
2491 deadline = req->rq_sent;
2492 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2493 deadline = req->rq_sent;
2495 deadline = req->rq_sent + req->rq_timeout;
2497 if (deadline <= now) /* actually expired already */
2498 timeout = 1; /* ASAP */
2499 else if (timeout == 0 || timeout > deadline - now)
2500 timeout = deadline - now;
2506 * Send all unset request from the set and then wait untill all
2507 * requests in the set complete (either get a reply, timeout, get an
2508 * error or otherwise be interrupted).
2509 * Returns 0 on success or error code otherwise.
2511 int ptlrpc_set_wait(const struct lu_env *env, struct ptlrpc_request_set *set)
2513 struct ptlrpc_request *req;
2518 if (set->set_producer)
2519 (void)ptlrpc_set_producer(set);
2521 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2522 if (req->rq_phase == RQ_PHASE_NEW)
2523 (void)ptlrpc_send_new_req(req);
2526 if (list_empty(&set->set_requests))
2530 timeout = ptlrpc_set_next_timeout(set);
2533 * wait until all complete, interrupted, or an in-flight
2536 CDEBUG(D_RPCTRACE, "set %p going to sleep for %lld seconds\n",
2539 if ((timeout == 0 && !signal_pending(current)) ||
2540 set->set_allow_intr) {
2542 * No requests are in-flight (ether timed out
2543 * or delayed), so we can allow interrupts.
2544 * We still want to block for a limited time,
2545 * so we allow interrupts during the timeout.
2547 rc = l_wait_event_abortable_timeout(
2549 ptlrpc_check_set(NULL, set),
2550 cfs_time_seconds(timeout ? timeout : 1));
2553 ptlrpc_expired_set(set);
2554 } else if (rc < 0) {
2556 ptlrpc_interrupted_set(set);
2562 * At least one request is in flight, so no
2563 * interrupts are allowed. Wait until all
2564 * complete, or an in-flight req times out.
2566 rc = wait_event_idle_timeout(
2568 ptlrpc_check_set(NULL, set),
2569 cfs_time_seconds(timeout ? timeout : 1));
2571 ptlrpc_expired_set(set);
2578 * LU-769 - if we ignored the signal because
2579 * it was already pending when we started, we
2580 * need to handle it now or we risk it being
2583 if (rc == -ETIMEDOUT &&
2584 signal_pending(current)) {
2587 siginitset(&new, LUSTRE_FATAL_SIGS);
2588 sigprocmask(SIG_BLOCK, &new, &old);
2590 * In fact we only interrupt for the
2591 * "fatal" signals like SIGINT or
2592 * SIGKILL. We still ignore less
2593 * important signals since ptlrpc set
2594 * is not easily reentrant from
2597 if (signal_pending(current))
2598 ptlrpc_interrupted_set(set);
2599 sigprocmask(SIG_SETMASK, &old, NULL);
2603 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2606 * -EINTR => all requests have been flagged rq_intr so next
2608 * -ETIMEDOUT => someone timed out. When all reqs have
2609 * timed out, signals are enabled allowing completion with
2611 * I don't really care if we go once more round the loop in
2612 * the error cases -eeb.
2614 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2615 list_for_each_entry(req, &set->set_requests,
2617 spin_lock(&req->rq_lock);
2618 req->rq_invalid_rqset = 1;
2619 spin_unlock(&req->rq_lock);
2622 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2624 LASSERT(atomic_read(&set->set_remaining) == 0);
2626 rc = set->set_rc; /* rq_status of already freed requests if any */
2627 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
2628 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2629 if (req->rq_status != 0)
2630 rc = req->rq_status;
2635 EXPORT_SYMBOL(ptlrpc_set_wait);
2638 * Helper fuction for request freeing.
2639 * Called when request count reached zero and request needs to be freed.
2640 * Removes request from all sorts of sending/replay lists it might be on,
2641 * frees network buffers if any are present.
2642 * If \a locked is set, that means caller is already holding import imp_lock
2643 * and so we no longer need to reobtain it (for certain lists manipulations)
2645 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2652 LASSERT(!request->rq_srv_req);
2653 LASSERT(request->rq_export == NULL);
2654 LASSERTF(!request->rq_receiving_reply, "req %px\n", request);
2655 LASSERTF(list_empty(&request->rq_list), "req %px\n", request);
2656 LASSERTF(list_empty(&request->rq_set_chain), "req %px\n", request);
2657 LASSERTF(!request->rq_replay, "req %px\n", request);
2659 req_capsule_fini(&request->rq_pill);
2662 * We must take it off the imp_replay_list first. Otherwise, we'll set
2663 * request->rq_reqmsg to NULL while osc_close is dereferencing it.
2665 if (request->rq_import) {
2667 spin_lock(&request->rq_import->imp_lock);
2668 list_del_init(&request->rq_replay_list);
2669 list_del_init(&request->rq_unreplied_list);
2671 spin_unlock(&request->rq_import->imp_lock);
2673 LASSERTF(list_empty(&request->rq_replay_list), "req %px\n", request);
2675 if (atomic_read(&request->rq_refcount) != 0) {
2676 DEBUG_REQ(D_ERROR, request,
2677 "freeing request with nonzero refcount");
2681 if (request->rq_repbuf)
2682 sptlrpc_cli_free_repbuf(request);
2684 if (request->rq_import) {
2685 if (!ptlrpcd_check_work(request)) {
2686 LASSERT(atomic_read(&request->rq_import->imp_reqs) > 0);
2687 atomic_dec(&request->rq_import->imp_reqs);
2689 class_import_put(request->rq_import);
2690 request->rq_import = NULL;
2692 if (request->rq_bulk)
2693 ptlrpc_free_bulk(request->rq_bulk);
2695 if (request->rq_reqbuf || request->rq_clrbuf)
2696 sptlrpc_cli_free_reqbuf(request);
2698 if (request->rq_cli_ctx)
2699 sptlrpc_req_put_ctx(request, !locked);
2701 if (request->rq_pool)
2702 __ptlrpc_free_req_to_pool(request);
2704 ptlrpc_request_cache_free(request);
2708 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2710 * Drop one request reference. Must be called with import imp_lock held.
2711 * When reference count drops to zero, request is freed.
2713 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2715 assert_spin_locked(&request->rq_import->imp_lock);
2716 (void)__ptlrpc_req_finished(request, 1);
2721 * Drops one reference count for request \a request.
2722 * \a locked set indicates that caller holds import imp_lock.
2723 * Frees the request whe reference count reaches zero.
2725 * \retval 1 the request is freed
2726 * \retval 0 some others still hold references on the request
2728 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2736 LASSERT(request != LP_POISON);
2737 LASSERT(request->rq_reqmsg != LP_POISON);
2739 DEBUG_REQ(D_INFO, request, "refcount now %u",
2740 atomic_read(&request->rq_refcount) - 1);
2742 spin_lock(&request->rq_lock);
2743 count = atomic_dec_return(&request->rq_refcount);
2744 LASSERTF(count >= 0, "Invalid ref count %d\n", count);
2747 * For open RPC, the client does not know the EA size (LOV, ACL, and
2748 * so on) before replied, then the client has to reserve very large
2749 * reply buffer. Such buffer will not be released until the RPC freed.
2750 * Since The open RPC is replayable, we need to keep it in the replay
2751 * list until close. If there are a lot of files opened concurrently,
2752 * then the client may be OOM.
2754 * If fact, it is unnecessary to keep reply buffer for open replay,
2755 * related EAs have already been saved via mdc_save_lovea() before
2756 * coming here. So it is safe to free the reply buffer some earlier
2757 * before releasing the RPC to avoid client OOM. LU-9514
2759 if (count == 1 && request->rq_early_free_repbuf && request->rq_repbuf) {
2760 spin_lock(&request->rq_early_free_lock);
2761 sptlrpc_cli_free_repbuf(request);
2762 request->rq_repbuf = NULL;
2763 request->rq_repbuf_len = 0;
2764 request->rq_repdata = NULL;
2765 request->rq_reqdata_len = 0;
2766 spin_unlock(&request->rq_early_free_lock);
2768 spin_unlock(&request->rq_lock);
2771 __ptlrpc_free_req(request, locked);
2777 * Drops one reference count for a request.
2779 void ptlrpc_req_finished(struct ptlrpc_request *request)
2781 __ptlrpc_req_finished(request, 0);
2783 EXPORT_SYMBOL(ptlrpc_req_finished);
2786 * Returns xid of a \a request
2788 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2790 return request->rq_xid;
2792 EXPORT_SYMBOL(ptlrpc_req_xid);
2795 * Disengage the client's reply buffer from the network
2796 * NB does _NOT_ unregister any client-side bulk.
2797 * IDEMPOTENT, but _not_ safe against concurrent callers.
2798 * The request owner (i.e. the thread doing the I/O) must call...
2799 * Returns 0 on success or 1 if unregistering cannot be made.
2801 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2803 bool discard = false;
2807 LASSERT(!in_interrupt());
2809 /* Let's setup deadline for reply unlink. */
2810 if (CFS_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2811 async && request->rq_reply_deadline == 0 && cfs_fail_val == 0)
2812 request->rq_reply_deadline = ktime_get_real_seconds() +
2813 PTLRPC_REQ_LONG_UNLINK;
2816 * Nothing left to do.
2818 if (!__ptlrpc_cli_wait_unlink(request, &discard))
2821 LNetMDUnlink(request->rq_reply_md_h);
2823 if (discard) /* Discard the request-out callback */
2824 __LNetMDUnlink(request->rq_req_md_h, discard);
2827 * Let's check it once again.
2829 if (!ptlrpc_cli_wait_unlink(request))
2832 /* Move to "Unregistering" phase as reply was not unlinked yet. */
2833 ptlrpc_rqphase_move(request, RQ_PHASE_UNREG_RPC);
2836 * Do not wait for unlink to finish.
2842 * We have to wait_event_idle_timeout() whatever the result, to get
2843 * a chance to run reply_in_callback(), and to make sure we've
2844 * unlinked before returning a req to the pool.
2847 wait_queue_head_t *wq = (request->rq_set) ?
2848 &request->rq_set->set_waitq :
2849 &request->rq_reply_waitq;
2850 int seconds = PTLRPC_REQ_LONG_UNLINK;
2852 * Network access will complete in finite time but the HUGE
2853 * timeout lets us CWARN for visibility of sluggish NALs
2855 while (seconds > 0 &&
2856 wait_event_idle_timeout(
2858 !ptlrpc_cli_wait_unlink(request),
2859 cfs_time_seconds(1)) == 0)
2862 ptlrpc_rqphase_move(request, request->rq_next_phase);
2866 DEBUG_REQ(D_WARNING, request,
2867 "Unexpectedly long timeout receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2868 request->rq_receiving_reply,
2869 request->rq_req_unlinked,
2870 request->rq_reply_unlinked);
2875 static void ptlrpc_free_request(struct ptlrpc_request *req)
2877 spin_lock(&req->rq_lock);
2879 spin_unlock(&req->rq_lock);
2881 if (req->rq_commit_cb)
2882 req->rq_commit_cb(req);
2883 list_del_init(&req->rq_replay_list);
2885 __ptlrpc_req_finished(req, 1);
2889 * the request is committed and dropped from the replay list of its import
2891 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2893 struct obd_import *imp = req->rq_import;
2895 spin_lock(&imp->imp_lock);
2896 if (list_empty(&req->rq_replay_list)) {
2897 spin_unlock(&imp->imp_lock);
2901 if (force || req->rq_transno <= imp->imp_peer_committed_transno) {
2902 if (imp->imp_replay_cursor == &req->rq_replay_list)
2903 imp->imp_replay_cursor = req->rq_replay_list.next;
2904 ptlrpc_free_request(req);
2907 spin_unlock(&imp->imp_lock);
2909 EXPORT_SYMBOL(ptlrpc_request_committed);
2912 * Iterates through replay_list on import and prunes
2913 * all requests have transno smaller than last_committed for the
2914 * import and don't have rq_replay set.
2915 * Since requests are sorted in transno order, stops when meeting first
2916 * transno bigger than last_committed.
2917 * caller must hold imp->imp_lock
2919 void ptlrpc_free_committed(struct obd_import *imp)
2921 struct ptlrpc_request *req, *saved;
2922 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2923 bool skip_committed_list = true;
2924 unsigned int replay_scanned = 0, replay_freed = 0;
2925 unsigned int commit_scanned = 0, commit_freed = 0;
2926 unsigned int debug_level = D_INFO;
2927 __u64 peer_committed_transno;
2929 time64_t start, now;
2932 LASSERT(imp != NULL);
2933 assert_spin_locked(&imp->imp_lock);
2935 start = ktime_get_seconds();
2936 /* save these here, we can potentially drop imp_lock after checking */
2937 peer_committed_transno = imp->imp_peer_committed_transno;
2938 imp_generation = imp->imp_generation;
2940 if (peer_committed_transno == imp->imp_last_transno_checked &&
2941 imp_generation == imp->imp_last_generation_checked) {
2942 CDEBUG(D_INFO, "%s: skip recheck: last_committed %llu\n",
2943 imp->imp_obd->obd_name, peer_committed_transno);
2946 CDEBUG(D_RPCTRACE, "%s: committing for last_committed %llu gen %d\n",
2947 imp->imp_obd->obd_name, peer_committed_transno, imp_generation);
2949 if (imp_generation != imp->imp_last_generation_checked ||
2950 imp->imp_last_transno_checked == 0)
2951 skip_committed_list = false;
2952 /* maybe drop imp_lock here, if another lock protected the lists */
2954 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2956 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2957 LASSERT(req != last_req);
2960 if (req->rq_transno == 0) {
2961 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2965 /* If other threads are waiting on imp_lock, stop processing
2966 * in this thread. Another thread can finish remaining work.
2967 * This may happen if there are huge numbers of open files
2968 * that are closed suddenly or evicted, or if the server
2969 * commit interval is very high vs. RPC rate.
2971 if (++replay_scanned % 2048 == 0) {
2972 now = ktime_get_seconds();
2973 if (now > start + 5)
2974 debug_level = D_WARNING;
2976 if ((replay_freed > 128 && now > start + 3) &&
2977 atomic_read(&imp->imp_waiting)) {
2978 if (debug_level == D_INFO)
2979 debug_level = D_RPCTRACE;
2984 if (req->rq_import_generation < imp_generation) {
2985 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2989 /* not yet committed */
2990 if (req->rq_transno > peer_committed_transno) {
2991 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2995 if (req->rq_replay) {
2996 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2997 list_move_tail(&req->rq_replay_list,
2998 &imp->imp_committed_list);
3002 DEBUG_REQ(D_INFO, req, "commit (last_committed %llu)",
3003 peer_committed_transno);
3006 ptlrpc_free_request(req);
3009 if (skip_committed_list)
3012 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
3014 LASSERT(req->rq_transno != 0);
3016 /* If other threads are waiting on imp_lock, stop processing
3017 * in this thread. Another thread can finish remaining work. */
3018 if (++commit_scanned % 2048 == 0) {
3019 now = ktime_get_seconds();
3020 if (now > start + 6)
3021 debug_level = D_WARNING;
3023 if ((commit_freed > 128 && now > start + 4) &&
3024 atomic_read(&imp->imp_waiting)) {
3025 if (debug_level == D_INFO)
3026 debug_level = D_RPCTRACE;
3031 if (req->rq_import_generation < imp_generation ||
3033 DEBUG_REQ(D_RPCTRACE, req, "free %s open request",
3034 req->rq_import_generation <
3035 imp_generation ? "stale" : "closed");
3037 if (imp->imp_replay_cursor == &req->rq_replay_list)
3038 imp->imp_replay_cursor =
3039 req->rq_replay_list.next;
3042 ptlrpc_free_request(req);
3046 /* if full lists processed without interruption, avoid next scan */
3047 if (debug_level == D_INFO) {
3048 imp->imp_last_transno_checked = peer_committed_transno;
3049 imp->imp_last_generation_checked = imp_generation;
3052 CDEBUG_LIMIT(debug_level,
3053 "%s: %s: skip=%u replay=%u/%u committed=%u/%u\n",
3054 imp->imp_obd->obd_name,
3055 debug_level == D_INFO ? "normal" : "overloaded",
3056 skip_committed_list, replay_freed, replay_scanned,
3057 commit_freed, commit_scanned);
3061 void ptlrpc_cleanup_client(struct obd_import *imp)
3068 * Schedule previously sent request for resend.
3069 * For bulk requests we assign new xid (to avoid problems with
3070 * lost replies and therefore several transfers landing into same buffer
3071 * from different sending attempts).
3073 void ptlrpc_resend_req(struct ptlrpc_request *req)
3075 DEBUG_REQ(D_HA, req, "going to resend");
3076 spin_lock(&req->rq_lock);
3079 * Request got reply but linked to the import list still.
3080 * Let ptlrpc_check_set() process it.
3082 if (ptlrpc_client_replied(req)) {
3083 spin_unlock(&req->rq_lock);
3084 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
3088 req->rq_status = -EAGAIN;
3091 req->rq_net_err = 0;
3092 req->rq_timedout = 0;
3094 ptlrpc_client_wake_req(req);
3095 spin_unlock(&req->rq_lock);
3098 /* XXX: this function and rq_status are currently unused */
3099 void ptlrpc_restart_req(struct ptlrpc_request *req)
3101 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
3102 req->rq_status = -ERESTARTSYS;
3104 spin_lock(&req->rq_lock);
3105 req->rq_restart = 1;
3106 req->rq_timedout = 0;
3107 ptlrpc_client_wake_req(req);
3108 spin_unlock(&req->rq_lock);
3112 * Grab additional reference on a request \a req
3114 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
3117 atomic_inc(&req->rq_refcount);
3120 EXPORT_SYMBOL(ptlrpc_request_addref);
3123 * Add a request to import replay_list.
3124 * Must be called under imp_lock
3126 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
3127 struct obd_import *imp)
3129 struct ptlrpc_request *iter;
3131 assert_spin_locked(&imp->imp_lock);
3133 if (req->rq_transno == 0) {
3134 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
3139 * clear this for new requests that were resent as well
3140 * as resent replayed requests.
3142 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
3144 /* don't re-add requests that have been replayed */
3145 if (!list_empty(&req->rq_replay_list))
3148 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
3150 spin_lock(&req->rq_lock);
3152 spin_unlock(&req->rq_lock);
3154 LASSERT(imp->imp_replayable);
3155 /* Balanced in ptlrpc_free_committed, usually. */
3156 ptlrpc_request_addref(req);
3157 list_for_each_entry_reverse(iter, &imp->imp_replay_list,
3160 * We may have duplicate transnos if we create and then
3161 * open a file, or for closes retained if to match creating
3162 * opens, so use req->rq_xid as a secondary key.
3163 * (See bugs 684, 685, and 428.)
3164 * XXX no longer needed, but all opens need transnos!
3166 if (iter->rq_transno > req->rq_transno)
3169 if (iter->rq_transno == req->rq_transno) {
3170 LASSERT(iter->rq_xid != req->rq_xid);
3171 if (iter->rq_xid > req->rq_xid)
3175 list_add(&req->rq_replay_list, &iter->rq_replay_list);
3179 list_add(&req->rq_replay_list, &imp->imp_replay_list);
3183 * Send request and wait until it completes.
3184 * Returns request processing status.
3186 int ptlrpc_queue_wait(struct ptlrpc_request *req)
3188 struct ptlrpc_request_set *set;
3192 LASSERT(req->rq_set == NULL);
3193 LASSERT(!req->rq_receiving_reply);
3195 set = ptlrpc_prep_set();
3197 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
3201 /* for distributed debugging */
3202 lustre_msg_set_status(req->rq_reqmsg, current->pid);
3204 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
3205 ptlrpc_request_addref(req);
3206 ptlrpc_set_add_req(set, req);
3207 rc = ptlrpc_set_wait(NULL, set);
3208 ptlrpc_set_destroy(set);
3212 EXPORT_SYMBOL(ptlrpc_queue_wait);
3215 * Callback used for replayed requests reply processing.
3216 * In case of successful reply calls registered request replay callback.
3217 * In case of error restart replay process.
3219 static int ptlrpc_replay_interpret(const struct lu_env *env,
3220 struct ptlrpc_request *req,
3223 struct ptlrpc_replay_async_args *aa = args;
3224 struct obd_import *imp = req->rq_import;
3227 atomic_dec(&imp->imp_replay_inflight);
3230 * Note: if it is bulk replay (MDS-MDS replay), then even if
3231 * server got the request, but bulk transfer timeout, let's
3232 * replay the bulk req again
3234 if (!ptlrpc_client_replied(req) ||
3236 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
3237 DEBUG_REQ(D_ERROR, req, "request replay timed out");
3238 GOTO(out, rc = -ETIMEDOUT);
3241 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
3242 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
3243 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
3244 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
3246 /** VBR: check version failure */
3247 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
3248 /** replay was failed due to version mismatch */
3249 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay");
3250 spin_lock(&imp->imp_lock);
3251 imp->imp_vbr_failed = 1;
3252 spin_unlock(&imp->imp_lock);
3253 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3255 /** The transno had better not change over replay. */
3256 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
3257 lustre_msg_get_transno(req->rq_repmsg) ||
3258 lustre_msg_get_transno(req->rq_repmsg) == 0,
3260 lustre_msg_get_transno(req->rq_reqmsg),
3261 lustre_msg_get_transno(req->rq_repmsg));
3264 spin_lock(&imp->imp_lock);
3265 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
3266 spin_unlock(&imp->imp_lock);
3267 LASSERT(imp->imp_last_replay_transno);
3269 /* transaction number shouldn't be bigger than the latest replayed */
3270 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
3271 DEBUG_REQ(D_ERROR, req,
3272 "Reported transno=%llu is bigger than replayed=%llu",
3274 lustre_msg_get_transno(req->rq_reqmsg));
3275 GOTO(out, rc = -EINVAL);
3278 DEBUG_REQ(D_HA, req, "got reply");
3280 /* let the callback do fixups, possibly including in the request */
3281 if (req->rq_replay_cb)
3282 req->rq_replay_cb(req);
3284 if (ptlrpc_client_replied(req) &&
3285 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
3286 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
3287 lustre_msg_get_status(req->rq_repmsg),
3288 aa->praa_old_status);
3291 * Note: If the replay fails for MDT-MDT recovery, let's
3292 * abort all of the following requests in the replay
3293 * and sending list, because MDT-MDT update requests
3294 * are dependent on each other, see LU-7039
3296 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
3297 struct ptlrpc_request *free_req;
3298 struct ptlrpc_request *tmp;
3300 spin_lock(&imp->imp_lock);
3301 list_for_each_entry_safe(free_req, tmp,
3302 &imp->imp_replay_list,
3304 ptlrpc_free_request(free_req);
3307 list_for_each_entry_safe(free_req, tmp,
3308 &imp->imp_committed_list,
3310 ptlrpc_free_request(free_req);
3313 list_for_each_entry_safe(free_req, tmp,
3314 &imp->imp_delayed_list,
3316 spin_lock(&free_req->rq_lock);
3317 free_req->rq_err = 1;
3318 free_req->rq_status = -EIO;
3319 ptlrpc_client_wake_req(free_req);
3320 spin_unlock(&free_req->rq_lock);
3323 list_for_each_entry_safe(free_req, tmp,
3324 &imp->imp_sending_list,
3326 spin_lock(&free_req->rq_lock);
3327 free_req->rq_err = 1;
3328 free_req->rq_status = -EIO;
3329 ptlrpc_client_wake_req(free_req);
3330 spin_unlock(&free_req->rq_lock);
3332 spin_unlock(&imp->imp_lock);
3335 /* Put it back for re-replay. */
3336 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
3340 * Errors while replay can set transno to 0, but
3341 * imp_last_replay_transno shouldn't be set to 0 anyway
3343 if (req->rq_transno == 0)
3344 CERROR("Transno is 0 during replay!\n");
3346 /* continue with recovery */
3347 rc = ptlrpc_import_recovery_state_machine(imp);
3349 req->rq_send_state = aa->praa_old_state;
3352 /* this replay failed, so restart recovery */
3353 ptlrpc_connect_import(imp);
3359 * Prepares and queues request for replay.
3360 * Adds it to ptlrpcd queue for actual sending.
3361 * Returns 0 on success.
3363 int ptlrpc_replay_req(struct ptlrpc_request *req)
3365 struct ptlrpc_replay_async_args *aa;
3369 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3371 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_REPLAY_PAUSE, cfs_fail_val);
3373 aa = ptlrpc_req_async_args(aa, req);
3374 memset(aa, 0, sizeof(*aa));
3376 /* Prepare request to be resent with ptlrpcd */
3377 aa->praa_old_state = req->rq_send_state;
3378 req->rq_send_state = LUSTRE_IMP_REPLAY;
3379 req->rq_phase = RQ_PHASE_NEW;
3380 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3382 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3384 req->rq_interpret_reply = ptlrpc_replay_interpret;
3385 /* Readjust the timeout for current conditions */
3386 ptlrpc_at_set_req_timeout(req);
3388 /* Tell server net_latency to calculate how long to wait for reply. */
3389 lustre_msg_set_service_timeout(req->rq_reqmsg,
3390 ptlrpc_at_get_net_latency(req));
3391 DEBUG_REQ(D_HA, req, "REPLAY");
3393 atomic_inc(&req->rq_import->imp_replay_inflight);
3394 spin_lock(&req->rq_lock);
3395 req->rq_early_free_repbuf = 0;
3396 spin_unlock(&req->rq_lock);
3397 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3399 ptlrpcd_add_req(req);
3404 * Aborts all in-flight request on import \a imp sending and delayed lists
3406 void ptlrpc_abort_inflight(struct obd_import *imp)
3408 struct ptlrpc_request *req;
3412 * Make sure that no new requests get processed for this import.
3413 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3414 * this flag and then putting requests on sending_list or delayed_list.
3416 assert_spin_locked(&imp->imp_lock);
3419 * XXX locking? Maybe we should remove each request with the list
3420 * locked? Also, how do we know if the requests on the list are
3421 * being freed at this time?
3423 list_for_each_entry(req, &imp->imp_sending_list, rq_list) {
3424 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3426 spin_lock(&req->rq_lock);
3427 if (req->rq_import_generation < imp->imp_generation) {
3429 req->rq_status = -EIO;
3430 ptlrpc_client_wake_req(req);
3432 spin_unlock(&req->rq_lock);
3435 list_for_each_entry(req, &imp->imp_delayed_list, rq_list) {
3436 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3438 spin_lock(&req->rq_lock);
3439 if (req->rq_import_generation < imp->imp_generation) {
3441 req->rq_status = -EIO;
3442 ptlrpc_client_wake_req(req);
3444 spin_unlock(&req->rq_lock);
3448 * Last chance to free reqs left on the replay list, but we
3449 * will still leak reqs that haven't committed.
3451 if (imp->imp_replayable)
3452 ptlrpc_free_committed(imp);
3458 * Abort all uncompleted requests in request set \a set
3460 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3462 struct ptlrpc_request *req;
3464 LASSERT(set != NULL);
3466 list_for_each_entry(req, &set->set_requests, rq_set_chain) {
3467 spin_lock(&req->rq_lock);
3468 if (req->rq_phase != RQ_PHASE_RPC) {
3469 spin_unlock(&req->rq_lock);
3474 req->rq_status = -EINTR;
3475 ptlrpc_client_wake_req(req);
3476 spin_unlock(&req->rq_lock);
3481 * Initialize the XID for the node. This is common among all requests on
3482 * this node, and only requires the property that it is monotonically
3483 * increasing. It does not need to be sequential. Since this is also used
3484 * as the RDMA match bits, it is important that a single client NOT have
3485 * the same match bits for two different in-flight requests, hence we do
3486 * NOT want to have an XID per target or similar.
3488 * To avoid an unlikely collision between match bits after a client reboot
3489 * (which would deliver old data into the wrong RDMA buffer) initialize
3490 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3491 * If the time is clearly incorrect, we instead use a 62-bit random number.
3492 * In the worst case the random number will overflow 1M RPCs per second in
3493 * 9133 years, or permutations thereof.
3495 #define YEAR_2004 (1ULL << 30)
3496 void ptlrpc_init_xid(void)
3498 time64_t now = ktime_get_real_seconds();
3501 if (now < YEAR_2004) {
3502 get_random_bytes(&xid, sizeof(xid));
3504 xid |= (1ULL << 61);
3506 xid = (u64)now << 20;
3509 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3510 BUILD_BUG_ON((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) !=
3512 xid &= PTLRPC_BULK_OPS_MASK;
3513 atomic64_set(&ptlrpc_last_xid, xid);
3517 * Increase xid and returns resulting new value to the caller.
3519 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3520 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3521 * itself uses the last bulk xid needed, so the server can determine the
3522 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3523 * xid must align to a power-of-two value.
3525 * This is assumed to be true due to the initial ptlrpc_last_xid
3526 * value also being initialized to a power-of-two value. LU-1431
3528 __u64 ptlrpc_next_xid(void)
3530 return atomic64_add_return(PTLRPC_BULK_OPS_COUNT, &ptlrpc_last_xid);
3534 * If request has a new allocated XID (new request or EINPROGRESS resend),
3535 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3536 * request to ensure previous bulk fails and avoid problems with lost replies
3537 * and therefore several transfers landing into the same buffer from different
3539 * Also, to avoid previous reply landing to a different sending attempt.
3541 void ptlrpc_set_mbits(struct ptlrpc_request *req)
3543 int md_count = req->rq_bulk ? req->rq_bulk->bd_md_count : 1;
3546 * Generate new matchbits for all resend requests, including
3549 if (req->rq_resend) {
3550 __u64 old_mbits = req->rq_mbits;
3553 * First time resend on -EINPROGRESS will generate new xid,
3554 * so we can actually use the rq_xid as rq_mbits in such case,
3555 * however, it's bit hard to distinguish such resend with a
3556 * 'resend for the -EINPROGRESS resend'. To make it simple,
3557 * we opt to generate mbits for all resend cases.
3559 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3561 req->rq_mbits = ptlrpc_next_xid();
3564 * Old version transfers rq_xid to peer as
3567 spin_lock(&req->rq_import->imp_lock);
3568 list_del_init(&req->rq_unreplied_list);
3569 ptlrpc_assign_next_xid_nolock(req);
3570 spin_unlock(&req->rq_import->imp_lock);
3571 req->rq_mbits = req->rq_xid;
3573 CDEBUG(D_HA, "resend with new mbits old x%llu new x%llu\n",
3574 old_mbits, req->rq_mbits);
3575 } else if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
3576 /* Request being sent first time, use xid as matchbits. */
3577 if (OCD_HAS_FLAG(&req->rq_import->imp_connect_data,
3578 BULK_MBITS) || req->rq_mbits == 0)
3580 req->rq_mbits = req->rq_xid;
3582 req->rq_mbits -= md_count - 1;
3586 * Replay request, xid and matchbits have already been
3587 * correctly assigned.
3593 * For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3594 * that server can infer the number of bulks that were prepared,
3597 req->rq_mbits += md_count - 1;
3600 * Set rq_xid as rq_mbits to indicate the final bulk for the old
3601 * server which does not support OBD_CONNECT_BULK_MBITS. LU-6808.
3603 * It's ok to directly set the rq_xid here, since this xid bump
3604 * won't affect the request position in unreplied list.
3606 if (!OCD_HAS_FLAG(&req->rq_import->imp_connect_data, BULK_MBITS))
3607 req->rq_xid = req->rq_mbits;
3611 * Get a glimpse at what next xid value might have been.
3612 * Returns possible next xid.
3614 __u64 ptlrpc_sample_next_xid(void)
3616 return atomic64_read(&ptlrpc_last_xid) + PTLRPC_BULK_OPS_COUNT;
3618 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3621 * Functions for operating ptlrpc workers.
3623 * A ptlrpc work is a function which will be running inside ptlrpc context.
3624 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3626 * 1. after a work is created, it can be used many times, that is:
3627 * handler = ptlrpcd_alloc_work();
3628 * ptlrpcd_queue_work();
3630 * queue it again when necessary:
3631 * ptlrpcd_queue_work();
3632 * ptlrpcd_destroy_work();
3633 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3634 * it will only be queued once in any time. Also as its name implies, it may
3635 * have delay before it really runs by ptlrpcd thread.
3637 struct ptlrpc_work_async_args {
3638 int (*cb)(const struct lu_env *, void *);
3642 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3644 /* re-initialize the req */
3645 req->rq_timeout = obd_timeout;
3646 req->rq_sent = ktime_get_real_seconds();
3647 req->rq_deadline = req->rq_sent + req->rq_timeout;
3648 req->rq_phase = RQ_PHASE_INTERPRET;
3649 req->rq_next_phase = RQ_PHASE_COMPLETE;
3650 req->rq_xid = ptlrpc_next_xid();
3651 req->rq_import_generation = req->rq_import->imp_generation;
3653 ptlrpcd_add_req(req);
3656 static int work_interpreter(const struct lu_env *env,
3657 struct ptlrpc_request *req, void *args, int rc)
3659 struct ptlrpc_work_async_args *arg = args;
3661 LASSERT(ptlrpcd_check_work(req));
3662 LASSERT(arg->cb != NULL);
3664 rc = arg->cb(env, arg->cbdata);
3666 list_del_init(&req->rq_set_chain);
3669 if (atomic_dec_return(&req->rq_refcount) > 1) {
3670 atomic_set(&req->rq_refcount, 2);
3671 ptlrpcd_add_work_req(req);
3676 static int worker_format;
3678 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3680 return req->rq_pill.rc_fmt == (void *)&worker_format;
3684 * Create a work for ptlrpc.
3686 void *ptlrpcd_alloc_work(struct obd_import *imp,
3687 int (*cb)(const struct lu_env *, void *), void *cbdata)
3689 struct ptlrpc_request *req = NULL;
3690 struct ptlrpc_work_async_args *args;
3696 RETURN(ERR_PTR(-EINVAL));
3698 /* copy some code from deprecated fakereq. */
3699 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3701 CERROR("ptlrpc: run out of memory!\n");
3702 RETURN(ERR_PTR(-ENOMEM));
3705 ptlrpc_cli_req_init(req);
3707 req->rq_send_state = LUSTRE_IMP_FULL;
3708 req->rq_type = PTL_RPC_MSG_REQUEST;
3709 req->rq_import = class_import_get(imp);
3710 req->rq_interpret_reply = work_interpreter;
3711 /* don't want reply */
3712 req->rq_no_delay = req->rq_no_resend = 1;
3713 req->rq_pill.rc_fmt = (void *)&worker_format;
3715 args = ptlrpc_req_async_args(args, req);
3717 args->cbdata = cbdata;
3721 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3723 void ptlrpcd_destroy_work(void *handler)
3725 struct ptlrpc_request *req = handler;
3728 ptlrpc_req_finished(req);
3730 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3732 int ptlrpcd_queue_work(void *handler)
3734 struct ptlrpc_request *req = handler;
3737 * Check if the req is already being queued.
3739 * Here comes a trick: it lacks a way of checking if a req is being
3740 * processed reliably in ptlrpc. Here I have to use refcount of req
3741 * for this purpose. This is okay because the caller should use this
3742 * req as opaque data. - Jinshan
3744 LASSERT(atomic_read(&req->rq_refcount) > 0);
3745 if (atomic_inc_return(&req->rq_refcount) == 2)
3746 ptlrpcd_add_work_req(req);
3749 EXPORT_SYMBOL(ptlrpcd_queue_work);