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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 /** Implementation of client-side PortalRPC interfaces */
39 #define DEBUG_SUBSYSTEM S_RPC
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
48 #include "ptlrpc_internal.h"
50 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
51 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
52 .release_frags = ptlrpc_release_bulk_page_pin,
54 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
56 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
57 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
58 .release_frags = ptlrpc_release_bulk_noop,
60 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
62 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
63 .add_iov_frag = ptlrpc_prep_bulk_frag,
65 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
67 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
68 static int ptlrpcd_check_work(struct ptlrpc_request *req);
69 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
72 * Initialize passed in client structure \a cl.
74 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
75 struct ptlrpc_client *cl)
77 cl->cli_request_portal = req_portal;
78 cl->cli_reply_portal = rep_portal;
81 EXPORT_SYMBOL(ptlrpc_init_client);
84 * Return PortalRPC connection for remore uud \a uuid
86 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
88 struct ptlrpc_connection *c;
90 lnet_process_id_t peer;
93 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
94 * before accessing its values. */
95 /* coverity[uninit_use_in_call] */
96 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
98 CNETERR("cannot find peer %s!\n", uuid->uuid);
102 c = ptlrpc_connection_get(peer, self, uuid);
104 memcpy(c->c_remote_uuid.uuid,
105 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
108 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
114 * Allocate and initialize new bulk descriptor on the sender.
115 * Returns pointer to the descriptor or NULL on error.
117 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
118 enum ptlrpc_bulk_op_type type,
120 const struct ptlrpc_bulk_frag_ops *ops)
122 struct ptlrpc_bulk_desc *desc;
125 /* ensure that only one of KIOV or IOVEC is set but not both */
126 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
127 ops->add_kiov_frag != NULL) ||
128 (ptlrpc_is_bulk_desc_kvec(type) &&
129 ops->add_iov_frag != NULL));
131 if (type & PTLRPC_BULK_BUF_KIOV) {
133 offsetof(struct ptlrpc_bulk_desc,
134 bd_u.bd_kiov.bd_vec[nfrags]));
137 offsetof(struct ptlrpc_bulk_desc,
138 bd_u.bd_kvec.bd_kvec[nfrags]));
144 spin_lock_init(&desc->bd_lock);
145 init_waitqueue_head(&desc->bd_waitq);
146 desc->bd_max_iov = nfrags;
147 desc->bd_iov_count = 0;
148 desc->bd_portal = portal;
149 desc->bd_type = type;
150 desc->bd_md_count = 0;
151 desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
152 LASSERT(max_brw > 0);
153 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
154 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
155 * node. Negotiated ocd_brw_size will always be <= this number. */
156 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
157 LNetInvalidateHandle(&desc->bd_mds[i]);
163 * Prepare bulk descriptor for specified outgoing request \a req that
164 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
165 * the bulk to be sent. Used on client-side.
166 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
169 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
170 unsigned nfrags, unsigned max_brw,
173 const struct ptlrpc_bulk_frag_ops
176 struct obd_import *imp = req->rq_import;
177 struct ptlrpc_bulk_desc *desc;
180 LASSERT(ptlrpc_is_bulk_op_passive(type));
182 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
186 desc->bd_import_generation = req->rq_import_generation;
187 desc->bd_import = class_import_get(imp);
190 desc->bd_cbid.cbid_fn = client_bulk_callback;
191 desc->bd_cbid.cbid_arg = desc;
193 /* This makes req own desc, and free it when she frees herself */
198 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
200 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
201 struct page *page, int pageoffset, int len,
206 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
207 LASSERT(page != NULL);
208 LASSERT(pageoffset >= 0);
210 LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
211 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
213 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
218 page_cache_get(page);
220 kiov->kiov_page = page;
221 kiov->kiov_offset = pageoffset;
222 kiov->kiov_len = len;
224 desc->bd_iov_count++;
226 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
228 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
234 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
235 LASSERT(frag != NULL);
237 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
239 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
243 iovec->iov_base = frag;
244 iovec->iov_len = len;
246 desc->bd_iov_count++;
248 RETURN(desc->bd_nob);
250 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
252 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
256 LASSERT(desc != NULL);
257 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
258 LASSERT(desc->bd_md_count == 0); /* network hands off */
259 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
260 LASSERT(desc->bd_frag_ops != NULL);
262 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
263 sptlrpc_enc_pool_put_pages(desc);
266 class_export_put(desc->bd_export);
268 class_import_put(desc->bd_import);
270 if (desc->bd_frag_ops->release_frags != NULL)
271 desc->bd_frag_ops->release_frags(desc);
273 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
274 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
275 bd_u.bd_kiov.bd_vec[desc->bd_max_iov]));
277 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
278 bd_u.bd_kvec.bd_kvec[desc->
283 EXPORT_SYMBOL(ptlrpc_free_bulk);
286 * Set server timelimit for this req, i.e. how long are we willing to wait
287 * for reply before timing out this request.
289 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
295 LASSERT(req->rq_import);
298 /* non-AT settings */
300 * \a imp_server_timeout means this is reverse import and
301 * we send (currently only) ASTs to the client and cannot afford
302 * to wait too long for the reply, otherwise the other client
303 * (because of which we are sending this request) would
304 * timeout waiting for us
306 req->rq_timeout = req->rq_import->imp_server_timeout ?
307 obd_timeout / 2 : obd_timeout;
309 at = &req->rq_import->imp_at;
310 idx = import_at_get_index(req->rq_import,
311 req->rq_request_portal);
312 serv_est = at_get(&at->iat_service_estimate[idx]);
313 req->rq_timeout = at_est2timeout(serv_est);
315 /* We could get even fancier here, using history to predict increased
318 /* Let the server know what this RPC timeout is by putting it in the
320 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
322 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
324 /* Adjust max service estimate based on server value */
325 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
326 unsigned int serv_est)
332 LASSERT(req->rq_import);
333 at = &req->rq_import->imp_at;
335 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
336 /* max service estimates are tracked on the server side,
337 so just keep minimal history here */
338 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
340 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
341 "has changed from %d to %d\n",
342 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
343 oldse, at_get(&at->iat_service_estimate[idx]));
346 /* Expected network latency per remote node (secs) */
347 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
349 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
352 /* Adjust expected network latency */
353 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
354 unsigned int service_time)
356 unsigned int nl, oldnl;
358 time_t now = cfs_time_current_sec();
360 LASSERT(req->rq_import);
362 if (service_time > now - req->rq_sent + 3) {
363 /* bz16408, however, this can also happen if early reply
364 * is lost and client RPC is expired and resent, early reply
365 * or reply of original RPC can still be fit in reply buffer
366 * of resent RPC, now client is measuring time from the
367 * resent time, but server sent back service time of original
370 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
371 D_ADAPTTO : D_WARNING,
372 "Reported service time %u > total measured time "
373 CFS_DURATION_T"\n", service_time,
374 cfs_time_sub(now, req->rq_sent));
378 /* Network latency is total time less server processing time */
379 nl = max_t(int, now - req->rq_sent -
380 service_time, 0) + 1; /* st rounding */
381 at = &req->rq_import->imp_at;
383 oldnl = at_measured(&at->iat_net_latency, nl);
385 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
386 "has changed from %d to %d\n",
387 req->rq_import->imp_obd->obd_name,
389 &req->rq_import->imp_connection->c_remote_uuid),
390 oldnl, at_get(&at->iat_net_latency));
393 static int unpack_reply(struct ptlrpc_request *req)
397 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
398 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
400 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
405 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
407 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
414 * Handle an early reply message, called with the rq_lock held.
415 * If anything goes wrong just ignore it - same as if it never happened
417 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
418 __must_hold(&req->rq_lock)
420 struct ptlrpc_request *early_req;
426 spin_unlock(&req->rq_lock);
428 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
430 spin_lock(&req->rq_lock);
434 rc = unpack_reply(early_req);
436 sptlrpc_cli_finish_early_reply(early_req);
437 spin_lock(&req->rq_lock);
441 /* Use new timeout value just to adjust the local value for this
442 * request, don't include it into at_history. It is unclear yet why
443 * service time increased and should it be counted or skipped, e.g.
444 * that can be recovery case or some error or server, the real reply
445 * will add all new data if it is worth to add. */
446 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
447 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
449 /* Network latency can be adjusted, it is pure network delays */
450 ptlrpc_at_adj_net_latency(req,
451 lustre_msg_get_service_time(early_req->rq_repmsg));
453 sptlrpc_cli_finish_early_reply(early_req);
455 spin_lock(&req->rq_lock);
456 olddl = req->rq_deadline;
457 /* server assumes it now has rq_timeout from when the request
458 * arrived, so the client should give it at least that long.
459 * since we don't know the arrival time we'll use the original
461 req->rq_deadline = req->rq_sent + req->rq_timeout +
462 ptlrpc_at_get_net_latency(req);
464 DEBUG_REQ(D_ADAPTTO, req,
465 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
466 "("CFS_DURATION_T"s)", req->rq_early_count,
467 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
468 cfs_time_sub(req->rq_deadline, olddl));
473 static struct kmem_cache *request_cache;
475 int ptlrpc_request_cache_init(void)
477 request_cache = kmem_cache_create("ptlrpc_cache",
478 sizeof(struct ptlrpc_request),
479 0, SLAB_HWCACHE_ALIGN, NULL);
480 return request_cache == NULL ? -ENOMEM : 0;
483 void ptlrpc_request_cache_fini(void)
485 kmem_cache_destroy(request_cache);
488 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
490 struct ptlrpc_request *req;
492 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
496 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
498 OBD_SLAB_FREE_PTR(req, request_cache);
502 * Wind down request pool \a pool.
503 * Frees all requests from the pool too
505 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
507 struct list_head *l, *tmp;
508 struct ptlrpc_request *req;
510 LASSERT(pool != NULL);
512 spin_lock(&pool->prp_lock);
513 list_for_each_safe(l, tmp, &pool->prp_req_list) {
514 req = list_entry(l, struct ptlrpc_request, rq_list);
515 list_del(&req->rq_list);
516 LASSERT(req->rq_reqbuf);
517 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
518 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
519 ptlrpc_request_cache_free(req);
521 spin_unlock(&pool->prp_lock);
522 OBD_FREE(pool, sizeof(*pool));
524 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
527 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
529 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
534 while (size < pool->prp_rq_size)
537 LASSERTF(list_empty(&pool->prp_req_list) ||
538 size == pool->prp_rq_size,
539 "Trying to change pool size with nonempty pool "
540 "from %d to %d bytes\n", pool->prp_rq_size, size);
542 spin_lock(&pool->prp_lock);
543 pool->prp_rq_size = size;
544 for (i = 0; i < num_rq; i++) {
545 struct ptlrpc_request *req;
546 struct lustre_msg *msg;
548 spin_unlock(&pool->prp_lock);
549 req = ptlrpc_request_cache_alloc(GFP_NOFS);
552 OBD_ALLOC_LARGE(msg, size);
554 ptlrpc_request_cache_free(req);
557 req->rq_reqbuf = msg;
558 req->rq_reqbuf_len = size;
560 spin_lock(&pool->prp_lock);
561 list_add_tail(&req->rq_list, &pool->prp_req_list);
563 spin_unlock(&pool->prp_lock);
566 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
569 * Create and initialize new request pool with given attributes:
570 * \a num_rq - initial number of requests to create for the pool
571 * \a msgsize - maximum message size possible for requests in thid pool
572 * \a populate_pool - function to be called when more requests need to be added
574 * Returns pointer to newly created pool or NULL on error.
576 struct ptlrpc_request_pool *
577 ptlrpc_init_rq_pool(int num_rq, int msgsize,
578 int (*populate_pool)(struct ptlrpc_request_pool *, int))
580 struct ptlrpc_request_pool *pool;
582 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
586 /* Request next power of two for the allocation, because internally
587 kernel would do exactly this */
589 spin_lock_init(&pool->prp_lock);
590 INIT_LIST_HEAD(&pool->prp_req_list);
591 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
592 pool->prp_populate = populate_pool;
594 populate_pool(pool, num_rq);
598 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
601 * Fetches one request from pool \a pool
603 static struct ptlrpc_request *
604 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
606 struct ptlrpc_request *request;
607 struct lustre_msg *reqbuf;
612 spin_lock(&pool->prp_lock);
614 /* See if we have anything in a pool, and bail out if nothing,
615 * in writeout path, where this matters, this is safe to do, because
616 * nothing is lost in this case, and when some in-flight requests
617 * complete, this code will be called again. */
618 if (unlikely(list_empty(&pool->prp_req_list))) {
619 spin_unlock(&pool->prp_lock);
623 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
625 list_del_init(&request->rq_list);
626 spin_unlock(&pool->prp_lock);
628 LASSERT(request->rq_reqbuf);
629 LASSERT(request->rq_pool);
631 reqbuf = request->rq_reqbuf;
632 memset(request, 0, sizeof(*request));
633 request->rq_reqbuf = reqbuf;
634 request->rq_reqbuf_len = pool->prp_rq_size;
635 request->rq_pool = pool;
641 * Returns freed \a request to pool.
643 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
645 struct ptlrpc_request_pool *pool = request->rq_pool;
647 spin_lock(&pool->prp_lock);
648 LASSERT(list_empty(&request->rq_list));
649 LASSERT(!request->rq_receiving_reply);
650 list_add_tail(&request->rq_list, &pool->prp_req_list);
651 spin_unlock(&pool->prp_lock);
654 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
656 struct obd_import *imp = req->rq_import;
657 struct list_head *tmp;
658 struct ptlrpc_request *iter;
660 assert_spin_locked(&imp->imp_lock);
661 LASSERT(list_empty(&req->rq_unreplied_list));
663 /* unreplied list is sorted by xid in ascending order */
664 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
665 iter = list_entry(tmp, struct ptlrpc_request,
668 LASSERT(req->rq_xid != iter->rq_xid);
669 if (req->rq_xid < iter->rq_xid)
671 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
674 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
677 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
679 req->rq_xid = ptlrpc_next_xid();
680 ptlrpc_add_unreplied(req);
683 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
685 spin_lock(&req->rq_import->imp_lock);
686 ptlrpc_assign_next_xid_nolock(req);
687 spin_unlock(&req->rq_import->imp_lock);
690 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
691 __u32 version, int opcode, char **bufs,
692 struct ptlrpc_cli_ctx *ctx)
695 struct obd_import *imp;
701 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
702 imp = request->rq_import;
703 lengths = request->rq_pill.rc_area[RCL_CLIENT];
706 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
708 rc = sptlrpc_req_get_ctx(request);
712 sptlrpc_req_set_flavor(request, opcode);
714 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
719 lustre_msg_add_version(request->rq_reqmsg, version);
720 request->rq_send_state = LUSTRE_IMP_FULL;
721 request->rq_type = PTL_RPC_MSG_REQUEST;
723 request->rq_req_cbid.cbid_fn = request_out_callback;
724 request->rq_req_cbid.cbid_arg = request;
726 request->rq_reply_cbid.cbid_fn = reply_in_callback;
727 request->rq_reply_cbid.cbid_arg = request;
729 request->rq_reply_deadline = 0;
730 request->rq_phase = RQ_PHASE_NEW;
731 request->rq_next_phase = RQ_PHASE_UNDEFINED;
733 request->rq_request_portal = imp->imp_client->cli_request_portal;
734 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
736 ptlrpc_at_set_req_timeout(request);
738 lustre_msg_set_opc(request->rq_reqmsg, opcode);
739 ptlrpc_assign_next_xid(request);
744 LASSERT(!request->rq_pool);
745 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
747 class_import_put(imp);
752 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
755 * Pack request buffers for network transfer, performing necessary encryption
756 * steps if necessary.
758 int ptlrpc_request_pack(struct ptlrpc_request *request,
759 __u32 version, int opcode)
762 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
766 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
767 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
768 * have to send old ptlrpc_body to keep interoprability with these
771 * Only three kinds of server->client RPCs so far:
776 * XXX This should be removed whenever we drop the interoprability with
777 * the these old clients.
779 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
780 opcode == LDLM_GL_CALLBACK)
781 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
782 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
786 EXPORT_SYMBOL(ptlrpc_request_pack);
789 * Helper function to allocate new request on import \a imp
790 * and possibly using existing request from pool \a pool if provided.
791 * Returns allocated request structure with import field filled or
795 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
796 struct ptlrpc_request_pool *pool)
798 struct ptlrpc_request *request = NULL;
800 request = ptlrpc_request_cache_alloc(GFP_NOFS);
802 if (!request && pool)
803 request = ptlrpc_prep_req_from_pool(pool);
806 ptlrpc_cli_req_init(request);
808 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
809 LASSERT(imp != LP_POISON);
810 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
812 LASSERT(imp->imp_client != LP_POISON);
814 request->rq_import = class_import_get(imp);
816 CERROR("request allocation out of memory\n");
823 * Helper function for creating a request.
824 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
825 * buffer structures according to capsule template \a format.
826 * Returns allocated request structure pointer or NULL on error.
828 static struct ptlrpc_request *
829 ptlrpc_request_alloc_internal(struct obd_import *imp,
830 struct ptlrpc_request_pool * pool,
831 const struct req_format *format)
833 struct ptlrpc_request *request;
835 request = __ptlrpc_request_alloc(imp, pool);
839 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
840 req_capsule_set(&request->rq_pill, format);
845 * Allocate new request structure for import \a imp and initialize its
846 * buffer structure according to capsule template \a format.
848 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
849 const struct req_format *format)
851 return ptlrpc_request_alloc_internal(imp, NULL, format);
853 EXPORT_SYMBOL(ptlrpc_request_alloc);
856 * Allocate new request structure for import \a imp from pool \a pool and
857 * initialize its buffer structure according to capsule template \a format.
859 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
860 struct ptlrpc_request_pool * pool,
861 const struct req_format *format)
863 return ptlrpc_request_alloc_internal(imp, pool, format);
865 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
868 * For requests not from pool, free memory of the request structure.
869 * For requests obtained from a pool earlier, return request back to pool.
871 void ptlrpc_request_free(struct ptlrpc_request *request)
873 if (request->rq_pool)
874 __ptlrpc_free_req_to_pool(request);
876 ptlrpc_request_cache_free(request);
878 EXPORT_SYMBOL(ptlrpc_request_free);
881 * Allocate new request for operatione \a opcode and immediatelly pack it for
883 * Only used for simple requests like OBD_PING where the only important
884 * part of the request is operation itself.
885 * Returns allocated request or NULL on error.
887 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
888 const struct req_format *format,
889 __u32 version, int opcode)
891 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
895 rc = ptlrpc_request_pack(req, version, opcode);
897 ptlrpc_request_free(req);
903 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
906 * Allocate and initialize new request set structure on the current CPT.
907 * Returns a pointer to the newly allocated set structure or NULL on error.
909 struct ptlrpc_request_set *ptlrpc_prep_set(void)
911 struct ptlrpc_request_set *set;
915 cpt = cfs_cpt_current(cfs_cpt_table, 0);
916 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
919 atomic_set(&set->set_refcount, 1);
920 INIT_LIST_HEAD(&set->set_requests);
921 init_waitqueue_head(&set->set_waitq);
922 atomic_set(&set->set_new_count, 0);
923 atomic_set(&set->set_remaining, 0);
924 spin_lock_init(&set->set_new_req_lock);
925 INIT_LIST_HEAD(&set->set_new_requests);
926 INIT_LIST_HEAD(&set->set_cblist);
927 set->set_max_inflight = UINT_MAX;
928 set->set_producer = NULL;
929 set->set_producer_arg = NULL;
934 EXPORT_SYMBOL(ptlrpc_prep_set);
937 * Allocate and initialize new request set structure with flow control
938 * extension. This extension allows to control the number of requests in-flight
939 * for the whole set. A callback function to generate requests must be provided
940 * and the request set will keep the number of requests sent over the wire to
942 * Returns a pointer to the newly allocated set structure or NULL on error.
944 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
948 struct ptlrpc_request_set *set;
950 set = ptlrpc_prep_set();
954 set->set_max_inflight = max;
955 set->set_producer = func;
956 set->set_producer_arg = arg;
962 * Wind down and free request set structure previously allocated with
964 * Ensures that all requests on the set have completed and removes
965 * all requests from the request list in a set.
966 * If any unsent request happen to be on the list, pretends that they got
967 * an error in flight and calls their completion handler.
969 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
971 struct list_head *tmp;
972 struct list_head *next;
977 /* Requests on the set should either all be completed, or all be new */
978 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
979 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
980 list_for_each(tmp, &set->set_requests) {
981 struct ptlrpc_request *req =
982 list_entry(tmp, struct ptlrpc_request,
985 LASSERT(req->rq_phase == expected_phase);
989 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
990 atomic_read(&set->set_remaining) == n, "%d / %d\n",
991 atomic_read(&set->set_remaining), n);
993 list_for_each_safe(tmp, next, &set->set_requests) {
994 struct ptlrpc_request *req =
995 list_entry(tmp, struct ptlrpc_request,
997 list_del_init(&req->rq_set_chain);
999 LASSERT(req->rq_phase == expected_phase);
1001 if (req->rq_phase == RQ_PHASE_NEW) {
1002 ptlrpc_req_interpret(NULL, req, -EBADR);
1003 atomic_dec(&set->set_remaining);
1006 spin_lock(&req->rq_lock);
1008 req->rq_invalid_rqset = 0;
1009 spin_unlock(&req->rq_lock);
1011 ptlrpc_req_finished (req);
1014 LASSERT(atomic_read(&set->set_remaining) == 0);
1016 ptlrpc_reqset_put(set);
1019 EXPORT_SYMBOL(ptlrpc_set_destroy);
1022 * Add a callback function \a fn to the set.
1023 * This function would be called when all requests on this set are completed.
1024 * The function will be passed \a data argument.
1026 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1027 set_interpreter_func fn, void *data)
1029 struct ptlrpc_set_cbdata *cbdata;
1031 OBD_ALLOC_PTR(cbdata);
1035 cbdata->psc_interpret = fn;
1036 cbdata->psc_data = data;
1037 list_add_tail(&cbdata->psc_item, &set->set_cblist);
1043 * Add a new request to the general purpose request set.
1044 * Assumes request reference from the caller.
1046 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1047 struct ptlrpc_request *req)
1049 LASSERT(list_empty(&req->rq_set_chain));
1051 if (req->rq_allow_intr)
1052 set->set_allow_intr = 1;
1054 /* The set takes over the caller's request reference */
1055 list_add_tail(&req->rq_set_chain, &set->set_requests);
1057 atomic_inc(&set->set_remaining);
1058 req->rq_queued_time = cfs_time_current();
1060 if (req->rq_reqmsg != NULL)
1061 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1063 if (set->set_producer != NULL)
1064 /* If the request set has a producer callback, the RPC must be
1065 * sent straight away */
1066 ptlrpc_send_new_req(req);
1068 EXPORT_SYMBOL(ptlrpc_set_add_req);
1071 * Add a request to a request with dedicated server thread
1072 * and wake the thread to make any necessary processing.
1073 * Currently only used for ptlrpcd.
1075 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1076 struct ptlrpc_request *req)
1078 struct ptlrpc_request_set *set = pc->pc_set;
1081 LASSERT(req->rq_set == NULL);
1082 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1084 spin_lock(&set->set_new_req_lock);
1086 * The set takes over the caller's request reference.
1089 req->rq_queued_time = cfs_time_current();
1090 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1091 count = atomic_inc_return(&set->set_new_count);
1092 spin_unlock(&set->set_new_req_lock);
1094 /* Only need to call wakeup once for the first entry. */
1096 wake_up(&set->set_waitq);
1098 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1099 * guarantee the async RPC can be processed ASAP, we have
1100 * no other better choice. It maybe fixed in future. */
1101 for (i = 0; i < pc->pc_npartners; i++)
1102 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1107 * Based on the current state of the import, determine if the request
1108 * can be sent, is an error, or should be delayed.
1110 * Returns true if this request should be delayed. If false, and
1111 * *status is set, then the request can not be sent and *status is the
1112 * error code. If false and status is 0, then request can be sent.
1114 * The imp->imp_lock must be held.
1116 static int ptlrpc_import_delay_req(struct obd_import *imp,
1117 struct ptlrpc_request *req, int *status)
1122 LASSERT (status != NULL);
1125 if (req->rq_ctx_init || req->rq_ctx_fini) {
1126 /* always allow ctx init/fini rpc go through */
1127 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1128 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1130 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1131 /* pings may safely race with umount */
1132 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1133 D_HA : D_ERROR, req, "IMP_CLOSED ");
1135 } else if (ptlrpc_send_limit_expired(req)) {
1136 /* probably doesn't need to be a D_ERROR after initial testing*/
1137 DEBUG_REQ(D_HA, req, "send limit expired ");
1138 *status = -ETIMEDOUT;
1139 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1140 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1141 /* allow CONNECT even if import is invalid */ ;
1142 if (atomic_read(&imp->imp_inval_count) != 0) {
1143 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1146 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1147 if (!imp->imp_deactive)
1148 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1149 *status = -ESHUTDOWN; /* bz 12940 */
1150 } else if (req->rq_import_generation != imp->imp_generation) {
1151 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1153 } else if (req->rq_send_state != imp->imp_state) {
1154 /* invalidate in progress - any requests should be drop */
1155 if (atomic_read(&imp->imp_inval_count) != 0) {
1156 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1158 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1159 *status = -EWOULDBLOCK;
1160 } else if (req->rq_allow_replay &&
1161 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1162 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1163 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1164 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1165 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1175 * Decide if the error message should be printed to the console or not.
1176 * Makes its decision based on request type, status, and failure frequency.
1178 * \param[in] req request that failed and may need a console message
1180 * \retval false if no message should be printed
1181 * \retval true if console message should be printed
1183 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1187 LASSERT(req->rq_reqmsg != NULL);
1188 opc = lustre_msg_get_opc(req->rq_reqmsg);
1190 /* Suppress particular reconnect errors which are to be expected. */
1191 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1194 /* Suppress timed out reconnect requests */
1195 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1199 /* Suppress most unavailable/again reconnect requests, but
1200 * print occasionally so it is clear client is trying to
1201 * connect to a server where no target is running. */
1202 err = lustre_msg_get_status(req->rq_repmsg);
1203 if ((err == -ENODEV || err == -EAGAIN) &&
1204 req->rq_import->imp_conn_cnt % 30 != 20)
1212 * Check request processing status.
1213 * Returns the status.
1215 static int ptlrpc_check_status(struct ptlrpc_request *req)
1220 err = lustre_msg_get_status(req->rq_repmsg);
1221 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1222 struct obd_import *imp = req->rq_import;
1223 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1224 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1226 if (ptlrpc_console_allow(req))
1227 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1228 "failed: rc = %d\n",
1229 imp->imp_obd->obd_name,
1231 libcfs_nid2str(nid), err);
1232 RETURN(err < 0 ? err : -EINVAL);
1236 DEBUG_REQ(D_INFO, req, "status is %d", err);
1237 } else if (err > 0) {
1238 /* XXX: translate this error from net to host */
1239 DEBUG_REQ(D_INFO, req, "status is %d", err);
1246 * save pre-versions of objects into request for replay.
1247 * Versions are obtained from server reply.
1250 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1252 struct lustre_msg *repmsg = req->rq_repmsg;
1253 struct lustre_msg *reqmsg = req->rq_reqmsg;
1254 __u64 *versions = lustre_msg_get_versions(repmsg);
1257 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1261 lustre_msg_set_versions(reqmsg, versions);
1262 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1263 versions[0], versions[1]);
1268 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1270 struct ptlrpc_request *req;
1272 assert_spin_locked(&imp->imp_lock);
1273 if (list_empty(&imp->imp_unreplied_list))
1276 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1278 LASSERTF(req->rq_xid >= 1, "XID:"LPU64"\n", req->rq_xid);
1280 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1281 imp->imp_known_replied_xid = req->rq_xid - 1;
1283 return req->rq_xid - 1;
1287 * Callback function called when client receives RPC reply for \a req.
1288 * Returns 0 on success or error code.
1289 * The return alue would be assigned to req->rq_status by the caller
1290 * as request processing status.
1291 * This function also decides if the request needs to be saved for later replay.
1293 static int after_reply(struct ptlrpc_request *req)
1295 struct obd_import *imp = req->rq_import;
1296 struct obd_device *obd = req->rq_import->imp_obd;
1298 struct timeval work_start;
1303 LASSERT(obd != NULL);
1304 /* repbuf must be unlinked */
1305 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1307 if (req->rq_reply_truncated) {
1308 if (ptlrpc_no_resend(req)) {
1309 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1310 " expected: %d, actual size: %d",
1311 req->rq_nob_received, req->rq_repbuf_len);
1315 sptlrpc_cli_free_repbuf(req);
1316 /* Pass the required reply buffer size (include
1317 * space for early reply).
1318 * NB: no need to roundup because alloc_repbuf
1319 * will roundup it */
1320 req->rq_replen = req->rq_nob_received;
1321 req->rq_nob_received = 0;
1322 spin_lock(&req->rq_lock);
1324 spin_unlock(&req->rq_lock);
1328 do_gettimeofday(&work_start);
1329 timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1332 * NB Until this point, the whole of the incoming message,
1333 * including buflens, status etc is in the sender's byte order.
1335 rc = sptlrpc_cli_unwrap_reply(req);
1337 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1342 * Security layer unwrap might ask resend this request.
1347 rc = unpack_reply(req);
1351 /* retry indefinitely on EINPROGRESS */
1352 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1353 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1354 time_t now = cfs_time_current_sec();
1356 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1357 spin_lock(&req->rq_lock);
1359 spin_unlock(&req->rq_lock);
1360 req->rq_nr_resend++;
1362 /* Readjust the timeout for current conditions */
1363 ptlrpc_at_set_req_timeout(req);
1364 /* delay resend to give a chance to the server to get ready.
1365 * The delay is increased by 1s on every resend and is capped to
1366 * the current request timeout (i.e. obd_timeout if AT is off,
1367 * or AT service time x 125% + 5s, see at_est2timeout) */
1368 if (req->rq_nr_resend > req->rq_timeout)
1369 req->rq_sent = now + req->rq_timeout;
1371 req->rq_sent = now + req->rq_nr_resend;
1373 /* Resend for EINPROGRESS will use a new XID */
1374 spin_lock(&imp->imp_lock);
1375 list_del_init(&req->rq_unreplied_list);
1376 spin_unlock(&imp->imp_lock);
1381 if (obd->obd_svc_stats != NULL) {
1382 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1384 ptlrpc_lprocfs_rpc_sent(req, timediff);
1387 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1388 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1389 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1390 lustre_msg_get_type(req->rq_repmsg));
1394 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1395 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1396 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1397 ptlrpc_at_adj_net_latency(req,
1398 lustre_msg_get_service_time(req->rq_repmsg));
1400 rc = ptlrpc_check_status(req);
1401 imp->imp_connect_error = rc;
1405 * Either we've been evicted, or the server has failed for
1406 * some reason. Try to reconnect, and if that fails, punt to
1409 if (ptlrpc_recoverable_error(rc)) {
1410 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1411 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1414 ptlrpc_request_handle_notconn(req);
1419 * Let's look if server sent slv. Do it only for RPC with
1422 ldlm_cli_update_pool(req);
1426 * Store transno in reqmsg for replay.
1428 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1429 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1430 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1433 if (imp->imp_replayable) {
1434 spin_lock(&imp->imp_lock);
1436 * No point in adding already-committed requests to the replay
1437 * list, we will just remove them immediately. b=9829
1439 if (req->rq_transno != 0 &&
1441 lustre_msg_get_last_committed(req->rq_repmsg) ||
1443 /** version recovery */
1444 ptlrpc_save_versions(req);
1445 ptlrpc_retain_replayable_request(req, imp);
1446 } else if (req->rq_commit_cb != NULL &&
1447 list_empty(&req->rq_replay_list)) {
1448 /* NB: don't call rq_commit_cb if it's already on
1449 * rq_replay_list, ptlrpc_free_committed() will call
1450 * it later, see LU-3618 for details */
1451 spin_unlock(&imp->imp_lock);
1452 req->rq_commit_cb(req);
1453 spin_lock(&imp->imp_lock);
1457 * Replay-enabled imports return commit-status information.
1459 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1460 if (likely(committed > imp->imp_peer_committed_transno))
1461 imp->imp_peer_committed_transno = committed;
1463 ptlrpc_free_committed(imp);
1465 if (!list_empty(&imp->imp_replay_list)) {
1466 struct ptlrpc_request *last;
1468 last = list_entry(imp->imp_replay_list.prev,
1469 struct ptlrpc_request,
1472 * Requests with rq_replay stay on the list even if no
1473 * commit is expected.
1475 if (last->rq_transno > imp->imp_peer_committed_transno)
1476 ptlrpc_pinger_commit_expected(imp);
1479 spin_unlock(&imp->imp_lock);
1486 * Helper function to send request \a req over the network for the first time
1487 * Also adjusts request phase.
1488 * Returns 0 on success or error code.
1490 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1492 struct obd_import *imp = req->rq_import;
1497 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1499 /* do not try to go further if there is not enough memory in enc_pool */
1500 if (req->rq_sent && req->rq_bulk != NULL)
1501 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1502 pool_is_at_full_capacity())
1505 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1506 (!req->rq_generation_set ||
1507 req->rq_import_generation == imp->imp_generation))
1510 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1512 spin_lock(&imp->imp_lock);
1514 LASSERT(req->rq_xid != 0);
1515 LASSERT(!list_empty(&req->rq_unreplied_list));
1517 if (!req->rq_generation_set)
1518 req->rq_import_generation = imp->imp_generation;
1520 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1521 spin_lock(&req->rq_lock);
1522 req->rq_waiting = 1;
1523 spin_unlock(&req->rq_lock);
1525 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1526 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1527 ptlrpc_import_state_name(req->rq_send_state),
1528 ptlrpc_import_state_name(imp->imp_state));
1529 LASSERT(list_empty(&req->rq_list));
1530 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1531 atomic_inc(&req->rq_import->imp_inflight);
1532 spin_unlock(&imp->imp_lock);
1537 spin_unlock(&imp->imp_lock);
1538 req->rq_status = rc;
1539 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1543 LASSERT(list_empty(&req->rq_list));
1544 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1545 atomic_inc(&req->rq_import->imp_inflight);
1547 /* find the known replied XID from the unreplied list, CONNECT
1548 * and DISCONNECT requests are skipped to make the sanity check
1549 * on server side happy. see process_req_last_xid().
1551 * For CONNECT: Because replay requests have lower XID, it'll
1552 * break the sanity check if CONNECT bump the exp_last_xid on
1555 * For DISCONNECT: Since client will abort inflight RPC before
1556 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1557 * than the inflight RPC.
1559 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1560 min_xid = ptlrpc_known_replied_xid(imp);
1561 spin_unlock(&imp->imp_lock);
1563 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1565 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1567 rc = sptlrpc_req_refresh_ctx(req, -1);
1570 req->rq_status = rc;
1573 spin_lock(&req->rq_lock);
1574 req->rq_wait_ctx = 1;
1575 spin_unlock(&req->rq_lock);
1580 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1581 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1582 imp->imp_obd->obd_uuid.uuid,
1583 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1584 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1585 lustre_msg_get_opc(req->rq_reqmsg));
1587 rc = ptl_send_rpc(req, 0);
1588 if (rc == -ENOMEM) {
1589 spin_lock(&imp->imp_lock);
1590 if (!list_empty(&req->rq_list)) {
1591 list_del_init(&req->rq_list);
1592 atomic_dec(&req->rq_import->imp_inflight);
1594 spin_unlock(&imp->imp_lock);
1595 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1599 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1600 spin_lock(&req->rq_lock);
1601 req->rq_net_err = 1;
1602 spin_unlock(&req->rq_lock);
1608 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1613 LASSERT(set->set_producer != NULL);
1615 remaining = atomic_read(&set->set_remaining);
1617 /* populate the ->set_requests list with requests until we
1618 * reach the maximum number of RPCs in flight for this set */
1619 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1620 rc = set->set_producer(set, set->set_producer_arg);
1621 if (rc == -ENOENT) {
1622 /* no more RPC to produce */
1623 set->set_producer = NULL;
1624 set->set_producer_arg = NULL;
1629 RETURN((atomic_read(&set->set_remaining) - remaining));
1633 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1634 * and no more replies are expected.
1635 * (it is possible to get less replies than requests sent e.g. due to timed out
1636 * requests or requests that we had trouble to send out)
1638 * NOTE: This function contains a potential schedule point (cond_resched()).
1640 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1642 struct list_head *tmp, *next;
1643 struct list_head comp_reqs;
1644 int force_timer_recalc = 0;
1647 if (atomic_read(&set->set_remaining) == 0)
1650 INIT_LIST_HEAD(&comp_reqs);
1651 list_for_each_safe(tmp, next, &set->set_requests) {
1652 struct ptlrpc_request *req =
1653 list_entry(tmp, struct ptlrpc_request,
1655 struct obd_import *imp = req->rq_import;
1656 int unregistered = 0;
1660 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1661 list_move_tail(&req->rq_set_chain, &comp_reqs);
1665 /* This schedule point is mainly for the ptlrpcd caller of this
1666 * function. Most ptlrpc sets are not long-lived and unbounded
1667 * in length, but at the least the set used by the ptlrpcd is.
1668 * Since the processing time is unbounded, we need to insert an
1669 * explicit schedule point to make the thread well-behaved.
1673 /* If the caller requires to allow to be interpreted by force
1674 * and it has really been interpreted, then move the request
1675 * to RQ_PHASE_INTERPRET phase in spite of what the current
1677 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1678 req->rq_status = -EINTR;
1679 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1681 /* Since it is interpreted and we have to wait for
1682 * the reply to be unlinked, then use sync mode. */
1685 GOTO(interpret, req->rq_status);
1688 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1689 force_timer_recalc = 1;
1691 /* delayed send - skip */
1692 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1695 /* delayed resend - skip */
1696 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1697 req->rq_sent > cfs_time_current_sec())
1700 if (!(req->rq_phase == RQ_PHASE_RPC ||
1701 req->rq_phase == RQ_PHASE_BULK ||
1702 req->rq_phase == RQ_PHASE_INTERPRET ||
1703 req->rq_phase == RQ_PHASE_UNREGISTERING)) {
1704 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1708 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1709 LASSERT(req->rq_next_phase != req->rq_phase);
1710 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1713 * Skip processing until reply is unlinked. We
1714 * can't return to pool before that and we can't
1715 * call interpret before that. We need to make
1716 * sure that all rdma transfers finished and will
1717 * not corrupt any data.
1719 if (ptlrpc_client_recv_or_unlink(req) ||
1720 ptlrpc_client_bulk_active(req))
1724 * Turn fail_loc off to prevent it from looping
1727 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1728 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1731 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1732 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1737 * Move to next phase if reply was successfully
1740 ptlrpc_rqphase_move(req, req->rq_next_phase);
1743 if (req->rq_phase == RQ_PHASE_INTERPRET)
1744 GOTO(interpret, req->rq_status);
1747 * Note that this also will start async reply unlink.
1749 if (req->rq_net_err && !req->rq_timedout) {
1750 ptlrpc_expire_one_request(req, 1);
1753 * Check if we still need to wait for unlink.
1755 if (ptlrpc_client_recv_or_unlink(req) ||
1756 ptlrpc_client_bulk_active(req))
1758 /* If there is no need to resend, fail it now. */
1759 if (req->rq_no_resend) {
1760 if (req->rq_status == 0)
1761 req->rq_status = -EIO;
1762 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1763 GOTO(interpret, req->rq_status);
1770 spin_lock(&req->rq_lock);
1771 req->rq_replied = 0;
1772 spin_unlock(&req->rq_lock);
1773 if (req->rq_status == 0)
1774 req->rq_status = -EIO;
1775 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1776 GOTO(interpret, req->rq_status);
1779 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1780 * so it sets rq_intr regardless of individual rpc
1781 * timeouts. The synchronous IO waiting path sets
1782 * rq_intr irrespective of whether ptlrpcd
1783 * has seen a timeout. Our policy is to only interpret
1784 * interrupted rpcs after they have timed out, so we
1785 * need to enforce that here.
1788 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1789 req->rq_wait_ctx)) {
1790 req->rq_status = -EINTR;
1791 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1792 GOTO(interpret, req->rq_status);
1795 if (req->rq_phase == RQ_PHASE_RPC) {
1796 if (req->rq_timedout || req->rq_resend ||
1797 req->rq_waiting || req->rq_wait_ctx) {
1800 if (!ptlrpc_unregister_reply(req, 1)) {
1801 ptlrpc_unregister_bulk(req, 1);
1805 spin_lock(&imp->imp_lock);
1806 if (ptlrpc_import_delay_req(imp, req, &status)){
1807 /* put on delay list - only if we wait
1808 * recovery finished - before send */
1809 list_del_init(&req->rq_list);
1810 list_add_tail(&req->rq_list,
1813 spin_unlock(&imp->imp_lock);
1818 req->rq_status = status;
1819 ptlrpc_rqphase_move(req,
1820 RQ_PHASE_INTERPRET);
1821 spin_unlock(&imp->imp_lock);
1822 GOTO(interpret, req->rq_status);
1824 if (ptlrpc_no_resend(req) &&
1825 !req->rq_wait_ctx) {
1826 req->rq_status = -ENOTCONN;
1827 ptlrpc_rqphase_move(req,
1828 RQ_PHASE_INTERPRET);
1829 spin_unlock(&imp->imp_lock);
1830 GOTO(interpret, req->rq_status);
1833 list_del_init(&req->rq_list);
1834 list_add_tail(&req->rq_list,
1835 &imp->imp_sending_list);
1837 spin_unlock(&imp->imp_lock);
1839 spin_lock(&req->rq_lock);
1840 req->rq_waiting = 0;
1841 spin_unlock(&req->rq_lock);
1843 if (req->rq_timedout || req->rq_resend) {
1844 /* This is re-sending anyways,
1845 * let's mark req as resend. */
1846 spin_lock(&req->rq_lock);
1848 spin_unlock(&req->rq_lock);
1850 if (req->rq_bulk != NULL &&
1851 !ptlrpc_unregister_bulk(req, 1))
1855 * rq_wait_ctx is only touched by ptlrpcd,
1856 * so no lock is needed here.
1858 status = sptlrpc_req_refresh_ctx(req, -1);
1861 req->rq_status = status;
1862 spin_lock(&req->rq_lock);
1863 req->rq_wait_ctx = 0;
1864 spin_unlock(&req->rq_lock);
1865 force_timer_recalc = 1;
1867 spin_lock(&req->rq_lock);
1868 req->rq_wait_ctx = 1;
1869 spin_unlock(&req->rq_lock);
1874 spin_lock(&req->rq_lock);
1875 req->rq_wait_ctx = 0;
1876 spin_unlock(&req->rq_lock);
1879 rc = ptl_send_rpc(req, 0);
1880 if (rc == -ENOMEM) {
1881 spin_lock(&imp->imp_lock);
1882 if (!list_empty(&req->rq_list))
1883 list_del_init(&req->rq_list);
1884 spin_unlock(&imp->imp_lock);
1885 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1889 DEBUG_REQ(D_HA, req,
1890 "send failed: rc = %d", rc);
1891 force_timer_recalc = 1;
1892 spin_lock(&req->rq_lock);
1893 req->rq_net_err = 1;
1894 spin_unlock(&req->rq_lock);
1897 /* need to reset the timeout */
1898 force_timer_recalc = 1;
1901 spin_lock(&req->rq_lock);
1903 if (ptlrpc_client_early(req)) {
1904 ptlrpc_at_recv_early_reply(req);
1905 spin_unlock(&req->rq_lock);
1909 /* Still waiting for a reply? */
1910 if (ptlrpc_client_recv(req)) {
1911 spin_unlock(&req->rq_lock);
1915 /* Did we actually receive a reply? */
1916 if (!ptlrpc_client_replied(req)) {
1917 spin_unlock(&req->rq_lock);
1921 spin_unlock(&req->rq_lock);
1923 /* unlink from net because we are going to
1924 * swab in-place of reply buffer */
1925 unregistered = ptlrpc_unregister_reply(req, 1);
1929 req->rq_status = after_reply(req);
1933 /* If there is no bulk associated with this request,
1934 * then we're done and should let the interpreter
1935 * process the reply. Similarly if the RPC returned
1936 * an error, and therefore the bulk will never arrive.
1938 if (req->rq_bulk == NULL || req->rq_status < 0) {
1939 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1940 GOTO(interpret, req->rq_status);
1943 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1946 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1947 if (ptlrpc_client_bulk_active(req))
1950 if (req->rq_bulk->bd_failure) {
1951 /* The RPC reply arrived OK, but the bulk screwed
1952 * up! Dead weird since the server told us the RPC
1953 * was good after getting the REPLY for her GET or
1954 * the ACK for her PUT. */
1955 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1956 req->rq_status = -EIO;
1959 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1962 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1964 /* This moves to "unregistering" phase we need to wait for
1966 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
1967 /* start async bulk unlink too */
1968 ptlrpc_unregister_bulk(req, 1);
1972 if (!ptlrpc_unregister_bulk(req, async))
1975 /* When calling interpret receiving already should be
1977 LASSERT(!req->rq_receiving_reply);
1979 ptlrpc_req_interpret(env, req, req->rq_status);
1981 if (ptlrpcd_check_work(req)) {
1982 atomic_dec(&set->set_remaining);
1985 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1987 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1988 "Completed RPC pname:cluuid:pid:xid:nid:"
1989 "opc %s:%s:%d:"LPU64":%s:%d\n",
1990 current_comm(), imp->imp_obd->obd_uuid.uuid,
1991 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1992 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1993 lustre_msg_get_opc(req->rq_reqmsg));
1995 spin_lock(&imp->imp_lock);
1996 /* Request already may be not on sending or delaying list. This
1997 * may happen in the case of marking it erroneous for the case
1998 * ptlrpc_import_delay_req(req, status) find it impossible to
1999 * allow sending this rpc and returns *status != 0. */
2000 if (!list_empty(&req->rq_list)) {
2001 list_del_init(&req->rq_list);
2002 atomic_dec(&imp->imp_inflight);
2004 list_del_init(&req->rq_unreplied_list);
2005 spin_unlock(&imp->imp_lock);
2007 atomic_dec(&set->set_remaining);
2008 wake_up_all(&imp->imp_recovery_waitq);
2010 if (set->set_producer) {
2011 /* produce a new request if possible */
2012 if (ptlrpc_set_producer(set) > 0)
2013 force_timer_recalc = 1;
2015 /* free the request that has just been completed
2016 * in order not to pollute set->set_requests */
2017 list_del_init(&req->rq_set_chain);
2018 spin_lock(&req->rq_lock);
2020 req->rq_invalid_rqset = 0;
2021 spin_unlock(&req->rq_lock);
2023 /* record rq_status to compute the final status later */
2024 if (req->rq_status != 0)
2025 set->set_rc = req->rq_status;
2026 ptlrpc_req_finished(req);
2028 list_move_tail(&req->rq_set_chain, &comp_reqs);
2032 /* move completed request at the head of list so it's easier for
2033 * caller to find them */
2034 list_splice(&comp_reqs, &set->set_requests);
2036 /* If we hit an error, we want to recover promptly. */
2037 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2039 EXPORT_SYMBOL(ptlrpc_check_set);
2042 * Time out request \a req. is \a async_unlink is set, that means do not wait
2043 * until LNet actually confirms network buffer unlinking.
2044 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2046 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2048 struct obd_import *imp = req->rq_import;
2052 spin_lock(&req->rq_lock);
2053 req->rq_timedout = 1;
2054 spin_unlock(&req->rq_lock);
2056 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
2057 "/real "CFS_DURATION_T"]",
2058 req->rq_net_err ? "failed due to network error" :
2059 ((req->rq_real_sent == 0 ||
2060 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
2061 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
2062 "timed out for sent delay" : "timed out for slow reply"),
2063 req->rq_sent, req->rq_real_sent);
2065 if (imp != NULL && obd_debug_peer_on_timeout)
2066 LNetDebugPeer(imp->imp_connection->c_peer);
2068 ptlrpc_unregister_reply(req, async_unlink);
2069 ptlrpc_unregister_bulk(req, async_unlink);
2071 if (obd_dump_on_timeout)
2072 libcfs_debug_dumplog();
2075 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2079 atomic_inc(&imp->imp_timeouts);
2081 /* The DLM server doesn't want recovery run on its imports. */
2082 if (imp->imp_dlm_fake)
2085 /* If this request is for recovery or other primordial tasks,
2086 * then error it out here. */
2087 if (req->rq_ctx_init || req->rq_ctx_fini ||
2088 req->rq_send_state != LUSTRE_IMP_FULL ||
2089 imp->imp_obd->obd_no_recov) {
2090 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2091 ptlrpc_import_state_name(req->rq_send_state),
2092 ptlrpc_import_state_name(imp->imp_state));
2093 spin_lock(&req->rq_lock);
2094 req->rq_status = -ETIMEDOUT;
2096 spin_unlock(&req->rq_lock);
2100 /* if a request can't be resent we can't wait for an answer after
2102 if (ptlrpc_no_resend(req)) {
2103 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2107 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2113 * Time out all uncompleted requests in request set pointed by \a data
2114 * Callback used when waiting on sets with l_wait_event.
2117 int ptlrpc_expired_set(void *data)
2119 struct ptlrpc_request_set *set = data;
2120 struct list_head *tmp;
2121 time_t now = cfs_time_current_sec();
2124 LASSERT(set != NULL);
2127 * A timeout expired. See which reqs it applies to...
2129 list_for_each(tmp, &set->set_requests) {
2130 struct ptlrpc_request *req =
2131 list_entry(tmp, struct ptlrpc_request,
2134 /* don't expire request waiting for context */
2135 if (req->rq_wait_ctx)
2138 /* Request in-flight? */
2139 if (!((req->rq_phase == RQ_PHASE_RPC &&
2140 !req->rq_waiting && !req->rq_resend) ||
2141 (req->rq_phase == RQ_PHASE_BULK)))
2144 if (req->rq_timedout || /* already dealt with */
2145 req->rq_deadline > now) /* not expired */
2148 /* Deal with this guy. Do it asynchronously to not block
2149 * ptlrpcd thread. */
2150 ptlrpc_expire_one_request(req, 1);
2154 * When waiting for a whole set, we always break out of the
2155 * sleep so we can recalculate the timeout, or enable interrupts
2156 * if everyone's timed out.
2162 * Sets rq_intr flag in \a req under spinlock.
2164 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2166 spin_lock(&req->rq_lock);
2168 spin_unlock(&req->rq_lock);
2170 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2173 * Interrupts (sets interrupted flag) all uncompleted requests in
2174 * a set \a data. Callback for l_wait_event for interruptible waits.
2176 static void ptlrpc_interrupted_set(void *data)
2178 struct ptlrpc_request_set *set = data;
2179 struct list_head *tmp;
2181 LASSERT(set != NULL);
2182 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2184 list_for_each(tmp, &set->set_requests) {
2185 struct ptlrpc_request *req =
2186 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2191 if (req->rq_phase != RQ_PHASE_RPC &&
2192 req->rq_phase != RQ_PHASE_UNREGISTERING &&
2193 !req->rq_allow_intr)
2196 ptlrpc_mark_interrupted(req);
2201 * Get the smallest timeout in the set; this does NOT set a timeout.
2203 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2205 struct list_head *tmp;
2206 time_t now = cfs_time_current_sec();
2208 struct ptlrpc_request *req;
2212 list_for_each(tmp, &set->set_requests) {
2213 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2216 * Request in-flight?
2218 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2219 (req->rq_phase == RQ_PHASE_BULK) ||
2220 (req->rq_phase == RQ_PHASE_NEW)))
2224 * Already timed out.
2226 if (req->rq_timedout)
2232 if (req->rq_wait_ctx)
2235 if (req->rq_phase == RQ_PHASE_NEW)
2236 deadline = req->rq_sent;
2237 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2238 deadline = req->rq_sent;
2240 deadline = req->rq_sent + req->rq_timeout;
2242 if (deadline <= now) /* actually expired already */
2243 timeout = 1; /* ASAP */
2244 else if (timeout == 0 || timeout > deadline - now)
2245 timeout = deadline - now;
2251 * Send all unset request from the set and then wait untill all
2252 * requests in the set complete (either get a reply, timeout, get an
2253 * error or otherwise be interrupted).
2254 * Returns 0 on success or error code otherwise.
2256 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2258 struct list_head *tmp;
2259 struct ptlrpc_request *req;
2260 struct l_wait_info lwi;
2264 if (set->set_producer)
2265 (void)ptlrpc_set_producer(set);
2267 list_for_each(tmp, &set->set_requests) {
2268 req = list_entry(tmp, struct ptlrpc_request,
2270 if (req->rq_phase == RQ_PHASE_NEW)
2271 (void)ptlrpc_send_new_req(req);
2274 if (list_empty(&set->set_requests))
2278 timeout = ptlrpc_set_next_timeout(set);
2280 /* wait until all complete, interrupted, or an in-flight
2282 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2285 if ((timeout == 0 && !signal_pending(current)) ||
2286 set->set_allow_intr)
2287 /* No requests are in-flight (ether timed out
2288 * or delayed), so we can allow interrupts.
2289 * We still want to block for a limited time,
2290 * so we allow interrupts during the timeout. */
2291 lwi = LWI_TIMEOUT_INTR_ALL(
2292 cfs_time_seconds(timeout ? timeout : 1),
2294 ptlrpc_interrupted_set, set);
2297 * At least one request is in flight, so no
2298 * interrupts are allowed. Wait until all
2299 * complete, or an in-flight req times out.
2301 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2302 ptlrpc_expired_set, set);
2304 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2306 /* LU-769 - if we ignored the signal because it was already
2307 * pending when we started, we need to handle it now or we risk
2308 * it being ignored forever */
2309 if (rc == -ETIMEDOUT &&
2310 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2311 signal_pending(current)) {
2312 sigset_t blocked_sigs =
2313 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2315 /* In fact we only interrupt for the "fatal" signals
2316 * like SIGINT or SIGKILL. We still ignore less
2317 * important signals since ptlrpc set is not easily
2318 * reentrant from userspace again */
2319 if (signal_pending(current))
2320 ptlrpc_interrupted_set(set);
2321 cfs_restore_sigs(blocked_sigs);
2324 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2326 /* -EINTR => all requests have been flagged rq_intr so next
2328 * -ETIMEDOUT => someone timed out. When all reqs have
2329 * timed out, signals are enabled allowing completion with
2331 * I don't really care if we go once more round the loop in
2332 * the error cases -eeb. */
2333 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2334 list_for_each(tmp, &set->set_requests) {
2335 req = list_entry(tmp, struct ptlrpc_request,
2337 spin_lock(&req->rq_lock);
2338 req->rq_invalid_rqset = 1;
2339 spin_unlock(&req->rq_lock);
2342 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2344 LASSERT(atomic_read(&set->set_remaining) == 0);
2346 rc = set->set_rc; /* rq_status of already freed requests if any */
2347 list_for_each(tmp, &set->set_requests) {
2348 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2350 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2351 if (req->rq_status != 0)
2352 rc = req->rq_status;
2355 if (set->set_interpret != NULL) {
2356 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2358 rc = interpreter (set, set->set_arg, rc);
2360 struct ptlrpc_set_cbdata *cbdata, *n;
2363 list_for_each_entry_safe(cbdata, n,
2364 &set->set_cblist, psc_item) {
2365 list_del_init(&cbdata->psc_item);
2366 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2369 OBD_FREE_PTR(cbdata);
2375 EXPORT_SYMBOL(ptlrpc_set_wait);
2378 * Helper fuction for request freeing.
2379 * Called when request count reached zero and request needs to be freed.
2380 * Removes request from all sorts of sending/replay lists it might be on,
2381 * frees network buffers if any are present.
2382 * If \a locked is set, that means caller is already holding import imp_lock
2383 * and so we no longer need to reobtain it (for certain lists manipulations)
2385 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2389 if (request == NULL)
2392 LASSERT(!request->rq_srv_req);
2393 LASSERT(request->rq_export == NULL);
2394 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2395 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2396 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2397 LASSERTF(!request->rq_replay, "req %p\n", request);
2399 req_capsule_fini(&request->rq_pill);
2401 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2402 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2403 if (request->rq_import != NULL) {
2405 spin_lock(&request->rq_import->imp_lock);
2406 list_del_init(&request->rq_replay_list);
2407 list_del_init(&request->rq_unreplied_list);
2409 spin_unlock(&request->rq_import->imp_lock);
2411 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2413 if (atomic_read(&request->rq_refcount) != 0) {
2414 DEBUG_REQ(D_ERROR, request,
2415 "freeing request with nonzero refcount");
2419 if (request->rq_repbuf != NULL)
2420 sptlrpc_cli_free_repbuf(request);
2422 if (request->rq_import != NULL) {
2423 class_import_put(request->rq_import);
2424 request->rq_import = NULL;
2426 if (request->rq_bulk != NULL)
2427 ptlrpc_free_bulk(request->rq_bulk);
2429 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2430 sptlrpc_cli_free_reqbuf(request);
2432 if (request->rq_cli_ctx)
2433 sptlrpc_req_put_ctx(request, !locked);
2435 if (request->rq_pool)
2436 __ptlrpc_free_req_to_pool(request);
2438 ptlrpc_request_cache_free(request);
2442 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2444 * Drop one request reference. Must be called with import imp_lock held.
2445 * When reference count drops to zero, request is freed.
2447 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2449 assert_spin_locked(&request->rq_import->imp_lock);
2450 (void)__ptlrpc_req_finished(request, 1);
2455 * Drops one reference count for request \a request.
2456 * \a locked set indicates that caller holds import imp_lock.
2457 * Frees the request whe reference count reaches zero.
2459 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2462 if (request == NULL)
2465 if (request == LP_POISON ||
2466 request->rq_reqmsg == LP_POISON) {
2467 CERROR("dereferencing freed request (bug 575)\n");
2472 DEBUG_REQ(D_INFO, request, "refcount now %u",
2473 atomic_read(&request->rq_refcount) - 1);
2475 if (atomic_dec_and_test(&request->rq_refcount)) {
2476 __ptlrpc_free_req(request, locked);
2484 * Drops one reference count for a request.
2486 void ptlrpc_req_finished(struct ptlrpc_request *request)
2488 __ptlrpc_req_finished(request, 0);
2490 EXPORT_SYMBOL(ptlrpc_req_finished);
2493 * Returns xid of a \a request
2495 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2497 return request->rq_xid;
2499 EXPORT_SYMBOL(ptlrpc_req_xid);
2502 * Disengage the client's reply buffer from the network
2503 * NB does _NOT_ unregister any client-side bulk.
2504 * IDEMPOTENT, but _not_ safe against concurrent callers.
2505 * The request owner (i.e. the thread doing the I/O) must call...
2506 * Returns 0 on success or 1 if unregistering cannot be made.
2508 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2511 struct l_wait_info lwi;
2516 LASSERT(!in_interrupt());
2519 * Let's setup deadline for reply unlink.
2521 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2522 async && request->rq_reply_deadline == 0)
2523 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2526 * Nothing left to do.
2528 if (!ptlrpc_client_recv_or_unlink(request))
2531 LNetMDUnlink(request->rq_reply_md_h);
2534 * Let's check it once again.
2536 if (!ptlrpc_client_recv_or_unlink(request))
2540 * Move to "Unregistering" phase as reply was not unlinked yet.
2542 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2545 * Do not wait for unlink to finish.
2551 * We have to l_wait_event() whatever the result, to give liblustre
2552 * a chance to run reply_in_callback(), and to make sure we've
2553 * unlinked before returning a req to the pool.
2556 /* The wq argument is ignored by user-space wait_event macros */
2557 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2558 &request->rq_set->set_waitq :
2559 &request->rq_reply_waitq;
2560 /* Network access will complete in finite time but the HUGE
2561 * timeout lets us CWARN for visibility of sluggish NALs */
2562 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2563 cfs_time_seconds(1), NULL, NULL);
2564 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2567 ptlrpc_rqphase_move(request, request->rq_next_phase);
2571 LASSERT(rc == -ETIMEDOUT);
2572 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2573 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2574 request->rq_receiving_reply,
2575 request->rq_req_unlinked,
2576 request->rq_reply_unlinked);
2581 static void ptlrpc_free_request(struct ptlrpc_request *req)
2583 spin_lock(&req->rq_lock);
2585 spin_unlock(&req->rq_lock);
2587 if (req->rq_commit_cb != NULL)
2588 req->rq_commit_cb(req);
2589 list_del_init(&req->rq_replay_list);
2591 __ptlrpc_req_finished(req, 1);
2595 * the request is committed and dropped from the replay list of its import
2597 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2599 struct obd_import *imp = req->rq_import;
2601 spin_lock(&imp->imp_lock);
2602 if (list_empty(&req->rq_replay_list)) {
2603 spin_unlock(&imp->imp_lock);
2607 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2608 ptlrpc_free_request(req);
2610 spin_unlock(&imp->imp_lock);
2612 EXPORT_SYMBOL(ptlrpc_request_committed);
2615 * Iterates through replay_list on import and prunes
2616 * all requests have transno smaller than last_committed for the
2617 * import and don't have rq_replay set.
2618 * Since requests are sorted in transno order, stops when meetign first
2619 * transno bigger than last_committed.
2620 * caller must hold imp->imp_lock
2622 void ptlrpc_free_committed(struct obd_import *imp)
2624 struct ptlrpc_request *req, *saved;
2625 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2626 bool skip_committed_list = true;
2629 LASSERT(imp != NULL);
2630 assert_spin_locked(&imp->imp_lock);
2632 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2633 imp->imp_generation == imp->imp_last_generation_checked) {
2634 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2635 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2638 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2639 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2640 imp->imp_generation);
2642 if (imp->imp_generation != imp->imp_last_generation_checked ||
2643 imp->imp_last_transno_checked == 0)
2644 skip_committed_list = false;
2646 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2647 imp->imp_last_generation_checked = imp->imp_generation;
2649 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2651 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2652 LASSERT(req != last_req);
2655 if (req->rq_transno == 0) {
2656 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2659 if (req->rq_import_generation < imp->imp_generation) {
2660 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2664 /* not yet committed */
2665 if (req->rq_transno > imp->imp_peer_committed_transno) {
2666 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2670 if (req->rq_replay) {
2671 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2672 list_move_tail(&req->rq_replay_list,
2673 &imp->imp_committed_list);
2677 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2678 imp->imp_peer_committed_transno);
2680 ptlrpc_free_request(req);
2683 if (skip_committed_list)
2686 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2688 LASSERT(req->rq_transno != 0);
2689 if (req->rq_import_generation < imp->imp_generation) {
2690 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2691 ptlrpc_free_request(req);
2692 } else if (!req->rq_replay) {
2693 DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2694 ptlrpc_free_request(req);
2701 void ptlrpc_cleanup_client(struct obd_import *imp)
2708 * Schedule previously sent request for resend.
2709 * For bulk requests we assign new xid (to avoid problems with
2710 * lost replies and therefore several transfers landing into same buffer
2711 * from different sending attempts).
2713 void ptlrpc_resend_req(struct ptlrpc_request *req)
2715 DEBUG_REQ(D_HA, req, "going to resend");
2716 spin_lock(&req->rq_lock);
2718 /* Request got reply but linked to the import list still.
2719 Let ptlrpc_check_set() to process it. */
2720 if (ptlrpc_client_replied(req)) {
2721 spin_unlock(&req->rq_lock);
2722 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2726 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2727 req->rq_status = -EAGAIN;
2730 req->rq_net_err = 0;
2731 req->rq_timedout = 0;
2733 ptlrpc_client_wake_req(req);
2734 spin_unlock(&req->rq_lock);
2737 /* XXX: this function and rq_status are currently unused */
2738 void ptlrpc_restart_req(struct ptlrpc_request *req)
2740 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2741 req->rq_status = -ERESTARTSYS;
2743 spin_lock(&req->rq_lock);
2744 req->rq_restart = 1;
2745 req->rq_timedout = 0;
2746 ptlrpc_client_wake_req(req);
2747 spin_unlock(&req->rq_lock);
2751 * Grab additional reference on a request \a req
2753 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2756 atomic_inc(&req->rq_refcount);
2759 EXPORT_SYMBOL(ptlrpc_request_addref);
2762 * Add a request to import replay_list.
2763 * Must be called under imp_lock
2765 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2766 struct obd_import *imp)
2768 struct list_head *tmp;
2770 assert_spin_locked(&imp->imp_lock);
2772 if (req->rq_transno == 0) {
2773 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2777 /* clear this for new requests that were resent as well
2778 as resent replayed requests. */
2779 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2781 /* don't re-add requests that have been replayed */
2782 if (!list_empty(&req->rq_replay_list))
2785 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2787 spin_lock(&req->rq_lock);
2789 spin_unlock(&req->rq_lock);
2791 LASSERT(imp->imp_replayable);
2792 /* Balanced in ptlrpc_free_committed, usually. */
2793 ptlrpc_request_addref(req);
2794 list_for_each_prev(tmp, &imp->imp_replay_list) {
2795 struct ptlrpc_request *iter = list_entry(tmp,
2796 struct ptlrpc_request,
2799 /* We may have duplicate transnos if we create and then
2800 * open a file, or for closes retained if to match creating
2801 * opens, so use req->rq_xid as a secondary key.
2802 * (See bugs 684, 685, and 428.)
2803 * XXX no longer needed, but all opens need transnos!
2805 if (iter->rq_transno > req->rq_transno)
2808 if (iter->rq_transno == req->rq_transno) {
2809 LASSERT(iter->rq_xid != req->rq_xid);
2810 if (iter->rq_xid > req->rq_xid)
2814 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2818 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2822 * Send request and wait until it completes.
2823 * Returns request processing status.
2825 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2827 struct ptlrpc_request_set *set;
2831 LASSERT(req->rq_set == NULL);
2832 LASSERT(!req->rq_receiving_reply);
2834 set = ptlrpc_prep_set();
2836 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2840 /* for distributed debugging */
2841 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2843 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2844 ptlrpc_request_addref(req);
2845 ptlrpc_set_add_req(set, req);
2846 rc = ptlrpc_set_wait(set);
2847 ptlrpc_set_destroy(set);
2851 EXPORT_SYMBOL(ptlrpc_queue_wait);
2854 * Callback used for replayed requests reply processing.
2855 * In case of successful reply calls registered request replay callback.
2856 * In case of error restart replay process.
2858 static int ptlrpc_replay_interpret(const struct lu_env *env,
2859 struct ptlrpc_request *req,
2860 void * data, int rc)
2862 struct ptlrpc_replay_async_args *aa = data;
2863 struct obd_import *imp = req->rq_import;
2866 atomic_dec(&imp->imp_replay_inflight);
2868 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2869 * server got the request, but bulk transfer timeout, let's
2870 * replay the bulk req again */
2871 if (!ptlrpc_client_replied(req) ||
2872 (req->rq_bulk != NULL &&
2873 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2874 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2875 GOTO(out, rc = -ETIMEDOUT);
2878 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2879 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2880 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2881 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2883 /** VBR: check version failure */
2884 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2885 /** replay was failed due to version mismatch */
2886 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2887 spin_lock(&imp->imp_lock);
2888 imp->imp_vbr_failed = 1;
2889 imp->imp_no_lock_replay = 1;
2890 spin_unlock(&imp->imp_lock);
2891 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2893 /** The transno had better not change over replay. */
2894 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2895 lustre_msg_get_transno(req->rq_repmsg) ||
2896 lustre_msg_get_transno(req->rq_repmsg) == 0,
2898 lustre_msg_get_transno(req->rq_reqmsg),
2899 lustre_msg_get_transno(req->rq_repmsg));
2902 spin_lock(&imp->imp_lock);
2903 /** if replays by version then gap occur on server, no trust to locks */
2904 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2905 imp->imp_no_lock_replay = 1;
2906 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2907 spin_unlock(&imp->imp_lock);
2908 LASSERT(imp->imp_last_replay_transno);
2910 /* transaction number shouldn't be bigger than the latest replayed */
2911 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2912 DEBUG_REQ(D_ERROR, req,
2913 "Reported transno "LPU64" is bigger than the "
2914 "replayed one: "LPU64, req->rq_transno,
2915 lustre_msg_get_transno(req->rq_reqmsg));
2916 GOTO(out, rc = -EINVAL);
2919 DEBUG_REQ(D_HA, req, "got rep");
2921 /* let the callback do fixups, possibly including in the request */
2922 if (req->rq_replay_cb)
2923 req->rq_replay_cb(req);
2925 if (ptlrpc_client_replied(req) &&
2926 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2927 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2928 lustre_msg_get_status(req->rq_repmsg),
2929 aa->praa_old_status);
2931 /* Note: If the replay fails for MDT-MDT recovery, let's
2932 * abort all of the following requests in the replay
2933 * and sending list, because MDT-MDT update requests
2934 * are dependent on each other, see LU-7039 */
2935 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
2936 struct ptlrpc_request *free_req;
2937 struct ptlrpc_request *tmp;
2939 spin_lock(&imp->imp_lock);
2940 list_for_each_entry_safe(free_req, tmp,
2941 &imp->imp_replay_list,
2943 ptlrpc_free_request(free_req);
2946 list_for_each_entry_safe(free_req, tmp,
2947 &imp->imp_committed_list,
2949 ptlrpc_free_request(free_req);
2952 list_for_each_entry_safe(free_req, tmp,
2953 &imp->imp_delayed_list,
2955 spin_lock(&free_req->rq_lock);
2956 free_req->rq_err = 1;
2957 free_req->rq_status = -EIO;
2958 ptlrpc_client_wake_req(free_req);
2959 spin_unlock(&free_req->rq_lock);
2962 list_for_each_entry_safe(free_req, tmp,
2963 &imp->imp_sending_list,
2965 spin_lock(&free_req->rq_lock);
2966 free_req->rq_err = 1;
2967 free_req->rq_status = -EIO;
2968 ptlrpc_client_wake_req(free_req);
2969 spin_unlock(&free_req->rq_lock);
2971 spin_unlock(&imp->imp_lock);
2974 /* Put it back for re-replay. */
2975 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2979 * Errors while replay can set transno to 0, but
2980 * imp_last_replay_transno shouldn't be set to 0 anyway
2982 if (req->rq_transno == 0)
2983 CERROR("Transno is 0 during replay!\n");
2985 /* continue with recovery */
2986 rc = ptlrpc_import_recovery_state_machine(imp);
2988 req->rq_send_state = aa->praa_old_state;
2991 /* this replay failed, so restart recovery */
2992 ptlrpc_connect_import(imp);
2998 * Prepares and queues request for replay.
2999 * Adds it to ptlrpcd queue for actual sending.
3000 * Returns 0 on success.
3002 int ptlrpc_replay_req(struct ptlrpc_request *req)
3004 struct ptlrpc_replay_async_args *aa;
3007 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3009 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
3010 aa = ptlrpc_req_async_args(req);
3011 memset(aa, 0, sizeof *aa);
3013 /* Prepare request to be resent with ptlrpcd */
3014 aa->praa_old_state = req->rq_send_state;
3015 req->rq_send_state = LUSTRE_IMP_REPLAY;
3016 req->rq_phase = RQ_PHASE_NEW;
3017 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3019 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3021 req->rq_interpret_reply = ptlrpc_replay_interpret;
3022 /* Readjust the timeout for current conditions */
3023 ptlrpc_at_set_req_timeout(req);
3025 /* Tell server the net_latency, so the server can calculate how long
3026 * it should wait for next replay */
3027 lustre_msg_set_service_time(req->rq_reqmsg,
3028 ptlrpc_at_get_net_latency(req));
3029 DEBUG_REQ(D_HA, req, "REPLAY");
3031 atomic_inc(&req->rq_import->imp_replay_inflight);
3032 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3034 ptlrpcd_add_req(req);
3039 * Aborts all in-flight request on import \a imp sending and delayed lists
3041 void ptlrpc_abort_inflight(struct obd_import *imp)
3043 struct list_head *tmp, *n;
3046 /* Make sure that no new requests get processed for this import.
3047 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3048 * this flag and then putting requests on sending_list or delayed_list.
3050 spin_lock(&imp->imp_lock);
3052 /* XXX locking? Maybe we should remove each request with the list
3053 * locked? Also, how do we know if the requests on the list are
3054 * being freed at this time?
3056 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3057 struct ptlrpc_request *req = list_entry(tmp,
3058 struct ptlrpc_request,
3061 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3063 spin_lock(&req->rq_lock);
3064 if (req->rq_import_generation < imp->imp_generation) {
3066 req->rq_status = -EIO;
3067 ptlrpc_client_wake_req(req);
3069 spin_unlock(&req->rq_lock);
3072 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3073 struct ptlrpc_request *req =
3074 list_entry(tmp, struct ptlrpc_request, rq_list);
3076 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3078 spin_lock(&req->rq_lock);
3079 if (req->rq_import_generation < imp->imp_generation) {
3081 req->rq_status = -EIO;
3082 ptlrpc_client_wake_req(req);
3084 spin_unlock(&req->rq_lock);
3087 /* Last chance to free reqs left on the replay list, but we
3088 * will still leak reqs that haven't committed. */
3089 if (imp->imp_replayable)
3090 ptlrpc_free_committed(imp);
3092 spin_unlock(&imp->imp_lock);
3098 * Abort all uncompleted requests in request set \a set
3100 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3102 struct list_head *tmp, *pos;
3104 LASSERT(set != NULL);
3106 list_for_each_safe(pos, tmp, &set->set_requests) {
3107 struct ptlrpc_request *req =
3108 list_entry(pos, struct ptlrpc_request,
3111 spin_lock(&req->rq_lock);
3112 if (req->rq_phase != RQ_PHASE_RPC) {
3113 spin_unlock(&req->rq_lock);
3118 req->rq_status = -EINTR;
3119 ptlrpc_client_wake_req(req);
3120 spin_unlock(&req->rq_lock);
3124 static __u64 ptlrpc_last_xid;
3125 static spinlock_t ptlrpc_last_xid_lock;
3128 * Initialize the XID for the node. This is common among all requests on
3129 * this node, and only requires the property that it is monotonically
3130 * increasing. It does not need to be sequential. Since this is also used
3131 * as the RDMA match bits, it is important that a single client NOT have
3132 * the same match bits for two different in-flight requests, hence we do
3133 * NOT want to have an XID per target or similar.
3135 * To avoid an unlikely collision between match bits after a client reboot
3136 * (which would deliver old data into the wrong RDMA buffer) initialize
3137 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3138 * If the time is clearly incorrect, we instead use a 62-bit random number.
3139 * In the worst case the random number will overflow 1M RPCs per second in
3140 * 9133 years, or permutations thereof.
3142 #define YEAR_2004 (1ULL << 30)
3143 void ptlrpc_init_xid(void)
3145 time_t now = cfs_time_current_sec();
3147 spin_lock_init(&ptlrpc_last_xid_lock);
3148 if (now < YEAR_2004) {
3149 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3150 ptlrpc_last_xid >>= 2;
3151 ptlrpc_last_xid |= (1ULL << 61);
3153 ptlrpc_last_xid = (__u64)now << 20;
3156 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3157 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3158 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3162 * Increase xid and returns resulting new value to the caller.
3164 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3165 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3166 * itself uses the last bulk xid needed, so the server can determine the
3167 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3168 * xid must align to a power-of-two value.
3170 * This is assumed to be true due to the initial ptlrpc_last_xid
3171 * value also being initialized to a power-of-two value. LU-1431
3173 __u64 ptlrpc_next_xid(void)
3177 spin_lock(&ptlrpc_last_xid_lock);
3178 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3179 ptlrpc_last_xid = next;
3180 spin_unlock(&ptlrpc_last_xid_lock);
3186 * If request has a new allocated XID (new request or EINPROGRESS resend),
3187 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3188 * request to ensure previous bulk fails and avoid problems with lost replies
3189 * and therefore several transfers landing into the same buffer from different
3192 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3194 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3196 LASSERT(bd != NULL);
3198 if (!req->rq_resend) {
3199 /* this request has a new xid, just use it as bulk matchbits */
3200 req->rq_mbits = req->rq_xid;
3202 } else { /* needs to generate a new matchbits for resend */
3203 __u64 old_mbits = req->rq_mbits;
3205 if ((bd->bd_import->imp_connect_data.ocd_connect_flags &
3206 OBD_CONNECT_BULK_MBITS) != 0)
3207 req->rq_mbits = ptlrpc_next_xid();
3208 else /* old version transfers rq_xid to peer as matchbits */
3209 req->rq_mbits = req->rq_xid = ptlrpc_next_xid();
3211 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
3212 old_mbits, req->rq_mbits);
3215 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3216 * that server can infer the number of bulks that were prepared,
3218 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3223 * Get a glimpse at what next xid value might have been.
3224 * Returns possible next xid.
3226 __u64 ptlrpc_sample_next_xid(void)
3228 #if BITS_PER_LONG == 32
3229 /* need to avoid possible word tearing on 32-bit systems */
3232 spin_lock(&ptlrpc_last_xid_lock);
3233 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3234 spin_unlock(&ptlrpc_last_xid_lock);
3238 /* No need to lock, since returned value is racy anyways */
3239 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3242 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3245 * Functions for operating ptlrpc workers.
3247 * A ptlrpc work is a function which will be running inside ptlrpc context.
3248 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3250 * 1. after a work is created, it can be used many times, that is:
3251 * handler = ptlrpcd_alloc_work();
3252 * ptlrpcd_queue_work();
3254 * queue it again when necessary:
3255 * ptlrpcd_queue_work();
3256 * ptlrpcd_destroy_work();
3257 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3258 * it will only be queued once in any time. Also as its name implies, it may
3259 * have delay before it really runs by ptlrpcd thread.
3261 struct ptlrpc_work_async_args {
3262 int (*cb)(const struct lu_env *, void *);
3266 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3268 /* re-initialize the req */
3269 req->rq_timeout = obd_timeout;
3270 req->rq_sent = cfs_time_current_sec();
3271 req->rq_deadline = req->rq_sent + req->rq_timeout;
3272 req->rq_reply_deadline = req->rq_deadline;
3273 req->rq_phase = RQ_PHASE_INTERPRET;
3274 req->rq_next_phase = RQ_PHASE_COMPLETE;
3275 req->rq_xid = ptlrpc_next_xid();
3276 req->rq_import_generation = req->rq_import->imp_generation;
3278 ptlrpcd_add_req(req);
3281 static int work_interpreter(const struct lu_env *env,
3282 struct ptlrpc_request *req, void *data, int rc)
3284 struct ptlrpc_work_async_args *arg = data;
3286 LASSERT(ptlrpcd_check_work(req));
3287 LASSERT(arg->cb != NULL);
3289 rc = arg->cb(env, arg->cbdata);
3291 list_del_init(&req->rq_set_chain);
3294 if (atomic_dec_return(&req->rq_refcount) > 1) {
3295 atomic_set(&req->rq_refcount, 2);
3296 ptlrpcd_add_work_req(req);
3301 static int worker_format;
3303 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3305 return req->rq_pill.rc_fmt == (void *)&worker_format;
3309 * Create a work for ptlrpc.
3311 void *ptlrpcd_alloc_work(struct obd_import *imp,
3312 int (*cb)(const struct lu_env *, void *), void *cbdata)
3314 struct ptlrpc_request *req = NULL;
3315 struct ptlrpc_work_async_args *args;
3321 RETURN(ERR_PTR(-EINVAL));
3323 /* copy some code from deprecated fakereq. */
3324 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3326 CERROR("ptlrpc: run out of memory!\n");
3327 RETURN(ERR_PTR(-ENOMEM));
3330 ptlrpc_cli_req_init(req);
3332 req->rq_send_state = LUSTRE_IMP_FULL;
3333 req->rq_type = PTL_RPC_MSG_REQUEST;
3334 req->rq_import = class_import_get(imp);
3335 req->rq_interpret_reply = work_interpreter;
3336 /* don't want reply */
3337 req->rq_no_delay = req->rq_no_resend = 1;
3338 req->rq_pill.rc_fmt = (void *)&worker_format;
3340 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3341 args = ptlrpc_req_async_args(req);
3343 args->cbdata = cbdata;
3347 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3349 void ptlrpcd_destroy_work(void *handler)
3351 struct ptlrpc_request *req = handler;
3354 ptlrpc_req_finished(req);
3356 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3358 int ptlrpcd_queue_work(void *handler)
3360 struct ptlrpc_request *req = handler;
3363 * Check if the req is already being queued.
3365 * Here comes a trick: it lacks a way of checking if a req is being
3366 * processed reliably in ptlrpc. Here I have to use refcount of req
3367 * for this purpose. This is okay because the caller should use this
3368 * req as opaque data. - Jinshan
3370 LASSERT(atomic_read(&req->rq_refcount) > 0);
3371 if (atomic_inc_return(&req->rq_refcount) == 2)
3372 ptlrpcd_add_work_req(req);
3375 EXPORT_SYMBOL(ptlrpcd_queue_work);