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, 2014, 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 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
691 __u32 version, int opcode,
692 int count, __u32 *lengths, char **bufs,
693 struct ptlrpc_cli_ctx *ctx)
695 struct obd_import *imp = request->rq_import;
700 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
702 rc = sptlrpc_req_get_ctx(request);
707 sptlrpc_req_set_flavor(request, opcode);
709 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
712 LASSERT(!request->rq_pool);
716 lustre_msg_add_version(request->rq_reqmsg, version);
717 request->rq_send_state = LUSTRE_IMP_FULL;
718 request->rq_type = PTL_RPC_MSG_REQUEST;
720 request->rq_req_cbid.cbid_fn = request_out_callback;
721 request->rq_req_cbid.cbid_arg = request;
723 request->rq_reply_cbid.cbid_fn = reply_in_callback;
724 request->rq_reply_cbid.cbid_arg = request;
726 request->rq_reply_deadline = 0;
727 request->rq_phase = RQ_PHASE_NEW;
728 request->rq_next_phase = RQ_PHASE_UNDEFINED;
730 request->rq_request_portal = imp->imp_client->cli_request_portal;
731 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
733 ptlrpc_at_set_req_timeout(request);
735 lustre_msg_set_opc(request->rq_reqmsg, opcode);
736 ptlrpc_assign_next_xid(request);
740 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
742 class_import_put(imp);
746 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
747 __u32 version, int opcode, char **bufs,
748 struct ptlrpc_cli_ctx *ctx)
752 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
753 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
754 request->rq_pill.rc_area[RCL_CLIENT],
757 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
760 * Pack request buffers for network transfer, performing necessary encryption
761 * steps if necessary.
763 int ptlrpc_request_pack(struct ptlrpc_request *request,
764 __u32 version, int opcode)
767 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
771 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
772 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
773 * have to send old ptlrpc_body to keep interoprability with these
776 * Only three kinds of server->client RPCs so far:
781 * XXX This should be removed whenever we drop the interoprability with
782 * the these old clients.
784 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
785 opcode == LDLM_GL_CALLBACK)
786 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
787 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
791 EXPORT_SYMBOL(ptlrpc_request_pack);
794 * Helper function to allocate new request on import \a imp
795 * and possibly using existing request from pool \a pool if provided.
796 * Returns allocated request structure with import field filled or
800 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
801 struct ptlrpc_request_pool *pool)
803 struct ptlrpc_request *request = NULL;
805 request = ptlrpc_request_cache_alloc(GFP_NOFS);
807 if (!request && pool)
808 request = ptlrpc_prep_req_from_pool(pool);
811 ptlrpc_cli_req_init(request);
813 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
814 LASSERT(imp != LP_POISON);
815 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
817 LASSERT(imp->imp_client != LP_POISON);
819 request->rq_import = class_import_get(imp);
821 CERROR("request allocation out of memory\n");
828 * Helper function for creating a request.
829 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
830 * buffer structures according to capsule template \a format.
831 * Returns allocated request structure pointer or NULL on error.
833 static struct ptlrpc_request *
834 ptlrpc_request_alloc_internal(struct obd_import *imp,
835 struct ptlrpc_request_pool * pool,
836 const struct req_format *format)
838 struct ptlrpc_request *request;
840 request = __ptlrpc_request_alloc(imp, pool);
844 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
845 req_capsule_set(&request->rq_pill, format);
850 * Allocate new request structure for import \a imp and initialize its
851 * buffer structure according to capsule template \a format.
853 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
854 const struct req_format *format)
856 return ptlrpc_request_alloc_internal(imp, NULL, format);
858 EXPORT_SYMBOL(ptlrpc_request_alloc);
861 * Allocate new request structure for import \a imp from pool \a pool and
862 * initialize its buffer structure according to capsule template \a format.
864 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
865 struct ptlrpc_request_pool * pool,
866 const struct req_format *format)
868 return ptlrpc_request_alloc_internal(imp, pool, format);
870 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
873 * For requests not from pool, free memory of the request structure.
874 * For requests obtained from a pool earlier, return request back to pool.
876 void ptlrpc_request_free(struct ptlrpc_request *request)
878 if (request->rq_pool)
879 __ptlrpc_free_req_to_pool(request);
881 ptlrpc_request_cache_free(request);
883 EXPORT_SYMBOL(ptlrpc_request_free);
886 * Allocate new request for operatione \a opcode and immediatelly pack it for
888 * Only used for simple requests like OBD_PING where the only important
889 * part of the request is operation itself.
890 * Returns allocated request or NULL on error.
892 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
893 const struct req_format *format,
894 __u32 version, int opcode)
896 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
900 rc = ptlrpc_request_pack(req, version, opcode);
902 ptlrpc_request_free(req);
908 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
911 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
912 * for operation \a opcode. Request would contain \a count buffers.
913 * Sizes of buffers are described in array \a lengths and buffers themselves
914 * are provided by a pointer \a bufs.
915 * Returns prepared request structure pointer or NULL on error.
917 struct ptlrpc_request *
918 ptlrpc_prep_req_pool(struct obd_import *imp,
919 __u32 version, int opcode,
920 int count, __u32 *lengths, char **bufs,
921 struct ptlrpc_request_pool *pool)
923 struct ptlrpc_request *request;
926 request = __ptlrpc_request_alloc(imp, pool);
930 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
931 lengths, bufs, NULL);
933 ptlrpc_request_free(request);
940 * Same as ptlrpc_prep_req_pool, but without pool
942 struct ptlrpc_request *
943 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
944 __u32 *lengths, char **bufs)
946 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
951 * Allocate and initialize new request set structure on the current CPT.
952 * Returns a pointer to the newly allocated set structure or NULL on error.
954 struct ptlrpc_request_set *ptlrpc_prep_set(void)
956 struct ptlrpc_request_set *set;
960 cpt = cfs_cpt_current(cfs_cpt_table, 0);
961 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
964 atomic_set(&set->set_refcount, 1);
965 INIT_LIST_HEAD(&set->set_requests);
966 init_waitqueue_head(&set->set_waitq);
967 atomic_set(&set->set_new_count, 0);
968 atomic_set(&set->set_remaining, 0);
969 spin_lock_init(&set->set_new_req_lock);
970 INIT_LIST_HEAD(&set->set_new_requests);
971 INIT_LIST_HEAD(&set->set_cblist);
972 set->set_max_inflight = UINT_MAX;
973 set->set_producer = NULL;
974 set->set_producer_arg = NULL;
979 EXPORT_SYMBOL(ptlrpc_prep_set);
982 * Allocate and initialize new request set structure with flow control
983 * extension. This extension allows to control the number of requests in-flight
984 * for the whole set. A callback function to generate requests must be provided
985 * and the request set will keep the number of requests sent over the wire to
987 * Returns a pointer to the newly allocated set structure or NULL on error.
989 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
993 struct ptlrpc_request_set *set;
995 set = ptlrpc_prep_set();
999 set->set_max_inflight = max;
1000 set->set_producer = func;
1001 set->set_producer_arg = arg;
1007 * Wind down and free request set structure previously allocated with
1009 * Ensures that all requests on the set have completed and removes
1010 * all requests from the request list in a set.
1011 * If any unsent request happen to be on the list, pretends that they got
1012 * an error in flight and calls their completion handler.
1014 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
1016 struct list_head *tmp;
1017 struct list_head *next;
1022 /* Requests on the set should either all be completed, or all be new */
1023 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
1024 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
1025 list_for_each(tmp, &set->set_requests) {
1026 struct ptlrpc_request *req =
1027 list_entry(tmp, struct ptlrpc_request,
1030 LASSERT(req->rq_phase == expected_phase);
1034 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1035 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1036 atomic_read(&set->set_remaining), n);
1038 list_for_each_safe(tmp, next, &set->set_requests) {
1039 struct ptlrpc_request *req =
1040 list_entry(tmp, struct ptlrpc_request,
1042 list_del_init(&req->rq_set_chain);
1044 LASSERT(req->rq_phase == expected_phase);
1046 if (req->rq_phase == RQ_PHASE_NEW) {
1047 ptlrpc_req_interpret(NULL, req, -EBADR);
1048 atomic_dec(&set->set_remaining);
1051 spin_lock(&req->rq_lock);
1053 req->rq_invalid_rqset = 0;
1054 spin_unlock(&req->rq_lock);
1056 ptlrpc_req_finished (req);
1059 LASSERT(atomic_read(&set->set_remaining) == 0);
1061 ptlrpc_reqset_put(set);
1064 EXPORT_SYMBOL(ptlrpc_set_destroy);
1067 * Add a callback function \a fn to the set.
1068 * This function would be called when all requests on this set are completed.
1069 * The function will be passed \a data argument.
1071 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1072 set_interpreter_func fn, void *data)
1074 struct ptlrpc_set_cbdata *cbdata;
1076 OBD_ALLOC_PTR(cbdata);
1080 cbdata->psc_interpret = fn;
1081 cbdata->psc_data = data;
1082 list_add_tail(&cbdata->psc_item, &set->set_cblist);
1088 * Add a new request to the general purpose request set.
1089 * Assumes request reference from the caller.
1091 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1092 struct ptlrpc_request *req)
1094 LASSERT(list_empty(&req->rq_set_chain));
1096 /* The set takes over the caller's request reference */
1097 list_add_tail(&req->rq_set_chain, &set->set_requests);
1099 atomic_inc(&set->set_remaining);
1100 req->rq_queued_time = cfs_time_current();
1102 if (req->rq_reqmsg != NULL)
1103 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1105 if (set->set_producer != NULL)
1106 /* If the request set has a producer callback, the RPC must be
1107 * sent straight away */
1108 ptlrpc_send_new_req(req);
1110 EXPORT_SYMBOL(ptlrpc_set_add_req);
1113 * Add a request to a request with dedicated server thread
1114 * and wake the thread to make any necessary processing.
1115 * Currently only used for ptlrpcd.
1117 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1118 struct ptlrpc_request *req)
1120 struct ptlrpc_request_set *set = pc->pc_set;
1123 LASSERT(req->rq_set == NULL);
1124 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1126 spin_lock(&set->set_new_req_lock);
1128 * The set takes over the caller's request reference.
1131 req->rq_queued_time = cfs_time_current();
1132 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1133 count = atomic_inc_return(&set->set_new_count);
1134 spin_unlock(&set->set_new_req_lock);
1136 /* Only need to call wakeup once for the first entry. */
1138 wake_up(&set->set_waitq);
1140 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1141 * guarantee the async RPC can be processed ASAP, we have
1142 * no other better choice. It maybe fixed in future. */
1143 for (i = 0; i < pc->pc_npartners; i++)
1144 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1149 * Based on the current state of the import, determine if the request
1150 * can be sent, is an error, or should be delayed.
1152 * Returns true if this request should be delayed. If false, and
1153 * *status is set, then the request can not be sent and *status is the
1154 * error code. If false and status is 0, then request can be sent.
1156 * The imp->imp_lock must be held.
1158 static int ptlrpc_import_delay_req(struct obd_import *imp,
1159 struct ptlrpc_request *req, int *status)
1164 LASSERT (status != NULL);
1167 if (req->rq_ctx_init || req->rq_ctx_fini) {
1168 /* always allow ctx init/fini rpc go through */
1169 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1170 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1172 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1173 /* pings may safely race with umount */
1174 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1175 D_HA : D_ERROR, req, "IMP_CLOSED ");
1177 } else if (ptlrpc_send_limit_expired(req)) {
1178 /* probably doesn't need to be a D_ERROR after initial testing*/
1179 DEBUG_REQ(D_HA, req, "send limit expired ");
1180 *status = -ETIMEDOUT;
1181 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1182 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1183 /* allow CONNECT even if import is invalid */ ;
1184 if (atomic_read(&imp->imp_inval_count) != 0) {
1185 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1188 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1189 if (!imp->imp_deactive)
1190 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1191 *status = -ESHUTDOWN; /* bz 12940 */
1192 } else if (req->rq_import_generation != imp->imp_generation) {
1193 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1195 } else if (req->rq_send_state != imp->imp_state) {
1196 /* invalidate in progress - any requests should be drop */
1197 if (atomic_read(&imp->imp_inval_count) != 0) {
1198 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1200 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1201 *status = -EWOULDBLOCK;
1202 } else if (req->rq_allow_replay &&
1203 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1204 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1205 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1206 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1207 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1217 * Decide if the error message should be printed to the console or not.
1218 * Makes its decision based on request type, status, and failure frequency.
1220 * \param[in] req request that failed and may need a console message
1222 * \retval false if no message should be printed
1223 * \retval true if console message should be printed
1225 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1229 LASSERT(req->rq_reqmsg != NULL);
1230 opc = lustre_msg_get_opc(req->rq_reqmsg);
1232 /* Suppress particular reconnect errors which are to be expected. */
1233 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1236 /* Suppress timed out reconnect requests */
1237 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1241 /* Suppress most unavailable/again reconnect requests, but
1242 * print occasionally so it is clear client is trying to
1243 * connect to a server where no target is running. */
1244 err = lustre_msg_get_status(req->rq_repmsg);
1245 if ((err == -ENODEV || err == -EAGAIN) &&
1246 req->rq_import->imp_conn_cnt % 30 != 20)
1254 * Check request processing status.
1255 * Returns the status.
1257 static int ptlrpc_check_status(struct ptlrpc_request *req)
1262 err = lustre_msg_get_status(req->rq_repmsg);
1263 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1264 struct obd_import *imp = req->rq_import;
1265 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1266 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1268 if (ptlrpc_console_allow(req))
1269 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1270 "failed: rc = %d\n",
1271 imp->imp_obd->obd_name,
1273 libcfs_nid2str(nid), err);
1274 RETURN(err < 0 ? err : -EINVAL);
1278 DEBUG_REQ(D_INFO, req, "status is %d", err);
1279 } else if (err > 0) {
1280 /* XXX: translate this error from net to host */
1281 DEBUG_REQ(D_INFO, req, "status is %d", err);
1288 * save pre-versions of objects into request for replay.
1289 * Versions are obtained from server reply.
1292 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1294 struct lustre_msg *repmsg = req->rq_repmsg;
1295 struct lustre_msg *reqmsg = req->rq_reqmsg;
1296 __u64 *versions = lustre_msg_get_versions(repmsg);
1299 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1303 lustre_msg_set_versions(reqmsg, versions);
1304 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1305 versions[0], versions[1]);
1310 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1312 struct ptlrpc_request *req;
1314 assert_spin_locked(&imp->imp_lock);
1315 if (list_empty(&imp->imp_unreplied_list))
1318 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1320 LASSERTF(req->rq_xid >= 1, "XID:"LPU64"\n", req->rq_xid);
1322 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1323 imp->imp_known_replied_xid = req->rq_xid - 1;
1325 return req->rq_xid - 1;
1329 * Callback function called when client receives RPC reply for \a req.
1330 * Returns 0 on success or error code.
1331 * The return alue would be assigned to req->rq_status by the caller
1332 * as request processing status.
1333 * This function also decides if the request needs to be saved for later replay.
1335 static int after_reply(struct ptlrpc_request *req)
1337 struct obd_import *imp = req->rq_import;
1338 struct obd_device *obd = req->rq_import->imp_obd;
1340 struct timeval work_start;
1345 LASSERT(obd != NULL);
1346 /* repbuf must be unlinked */
1347 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1349 if (req->rq_reply_truncated) {
1350 if (ptlrpc_no_resend(req)) {
1351 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1352 " expected: %d, actual size: %d",
1353 req->rq_nob_received, req->rq_repbuf_len);
1357 sptlrpc_cli_free_repbuf(req);
1358 /* Pass the required reply buffer size (include
1359 * space for early reply).
1360 * NB: no need to roundup because alloc_repbuf
1361 * will roundup it */
1362 req->rq_replen = req->rq_nob_received;
1363 req->rq_nob_received = 0;
1364 spin_lock(&req->rq_lock);
1366 spin_unlock(&req->rq_lock);
1370 do_gettimeofday(&work_start);
1371 timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1374 * NB Until this point, the whole of the incoming message,
1375 * including buflens, status etc is in the sender's byte order.
1377 rc = sptlrpc_cli_unwrap_reply(req);
1379 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1384 * Security layer unwrap might ask resend this request.
1389 rc = unpack_reply(req);
1393 /* retry indefinitely on EINPROGRESS */
1394 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1395 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1396 time_t now = cfs_time_current_sec();
1398 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1399 spin_lock(&req->rq_lock);
1401 spin_unlock(&req->rq_lock);
1402 req->rq_nr_resend++;
1404 /* Readjust the timeout for current conditions */
1405 ptlrpc_at_set_req_timeout(req);
1406 /* delay resend to give a chance to the server to get ready.
1407 * The delay is increased by 1s on every resend and is capped to
1408 * the current request timeout (i.e. obd_timeout if AT is off,
1409 * or AT service time x 125% + 5s, see at_est2timeout) */
1410 if (req->rq_nr_resend > req->rq_timeout)
1411 req->rq_sent = now + req->rq_timeout;
1413 req->rq_sent = now + req->rq_nr_resend;
1415 /* Resend for EINPROGRESS will use a new XID */
1416 spin_lock(&imp->imp_lock);
1417 list_del_init(&req->rq_unreplied_list);
1418 spin_unlock(&imp->imp_lock);
1423 if (obd->obd_svc_stats != NULL) {
1424 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1426 ptlrpc_lprocfs_rpc_sent(req, timediff);
1429 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1430 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1431 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1432 lustre_msg_get_type(req->rq_repmsg));
1436 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1437 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1438 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1439 ptlrpc_at_adj_net_latency(req,
1440 lustre_msg_get_service_time(req->rq_repmsg));
1442 rc = ptlrpc_check_status(req);
1443 imp->imp_connect_error = rc;
1447 * Either we've been evicted, or the server has failed for
1448 * some reason. Try to reconnect, and if that fails, punt to
1451 if (ptlrpc_recoverable_error(rc)) {
1452 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1453 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1456 ptlrpc_request_handle_notconn(req);
1461 * Let's look if server sent slv. Do it only for RPC with
1464 ldlm_cli_update_pool(req);
1468 * Store transno in reqmsg for replay.
1470 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1471 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1472 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1475 if (imp->imp_replayable) {
1476 spin_lock(&imp->imp_lock);
1478 * No point in adding already-committed requests to the replay
1479 * list, we will just remove them immediately. b=9829
1481 if (req->rq_transno != 0 &&
1483 lustre_msg_get_last_committed(req->rq_repmsg) ||
1485 /** version recovery */
1486 ptlrpc_save_versions(req);
1487 ptlrpc_retain_replayable_request(req, imp);
1488 } else if (req->rq_commit_cb != NULL &&
1489 list_empty(&req->rq_replay_list)) {
1490 /* NB: don't call rq_commit_cb if it's already on
1491 * rq_replay_list, ptlrpc_free_committed() will call
1492 * it later, see LU-3618 for details */
1493 spin_unlock(&imp->imp_lock);
1494 req->rq_commit_cb(req);
1495 spin_lock(&imp->imp_lock);
1499 * Replay-enabled imports return commit-status information.
1501 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1502 if (likely(committed > imp->imp_peer_committed_transno))
1503 imp->imp_peer_committed_transno = committed;
1505 ptlrpc_free_committed(imp);
1507 if (!list_empty(&imp->imp_replay_list)) {
1508 struct ptlrpc_request *last;
1510 last = list_entry(imp->imp_replay_list.prev,
1511 struct ptlrpc_request,
1514 * Requests with rq_replay stay on the list even if no
1515 * commit is expected.
1517 if (last->rq_transno > imp->imp_peer_committed_transno)
1518 ptlrpc_pinger_commit_expected(imp);
1521 spin_unlock(&imp->imp_lock);
1528 * Helper function to send request \a req over the network for the first time
1529 * Also adjusts request phase.
1530 * Returns 0 on success or error code.
1532 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1534 struct obd_import *imp = req->rq_import;
1539 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1541 /* do not try to go further if there is not enough memory in enc_pool */
1542 if (req->rq_sent && req->rq_bulk != NULL)
1543 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1544 pool_is_at_full_capacity())
1547 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1548 (!req->rq_generation_set ||
1549 req->rq_import_generation == imp->imp_generation))
1552 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1554 spin_lock(&imp->imp_lock);
1556 LASSERT(req->rq_xid != 0);
1557 LASSERT(!list_empty(&req->rq_unreplied_list));
1559 if (!req->rq_generation_set)
1560 req->rq_import_generation = imp->imp_generation;
1562 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1563 spin_lock(&req->rq_lock);
1564 req->rq_waiting = 1;
1565 spin_unlock(&req->rq_lock);
1567 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1568 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1569 ptlrpc_import_state_name(req->rq_send_state),
1570 ptlrpc_import_state_name(imp->imp_state));
1571 LASSERT(list_empty(&req->rq_list));
1572 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1573 atomic_inc(&req->rq_import->imp_inflight);
1574 spin_unlock(&imp->imp_lock);
1579 spin_unlock(&imp->imp_lock);
1580 req->rq_status = rc;
1581 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1585 LASSERT(list_empty(&req->rq_list));
1586 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1587 atomic_inc(&req->rq_import->imp_inflight);
1589 /* find the known replied XID from the unreplied list, CONNECT
1590 * and DISCONNECT requests are skipped to make the sanity check
1591 * on server side happy. see process_req_last_xid().
1593 * For CONNECT: Because replay requests have lower XID, it'll
1594 * break the sanity check if CONNECT bump the exp_last_xid on
1597 * For DISCONNECT: Since client will abort inflight RPC before
1598 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1599 * than the inflight RPC.
1601 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1602 min_xid = ptlrpc_known_replied_xid(imp);
1603 spin_unlock(&imp->imp_lock);
1605 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1607 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1609 rc = sptlrpc_req_refresh_ctx(req, -1);
1612 req->rq_status = rc;
1615 spin_lock(&req->rq_lock);
1616 req->rq_wait_ctx = 1;
1617 spin_unlock(&req->rq_lock);
1622 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1623 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1624 imp->imp_obd->obd_uuid.uuid,
1625 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1626 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1627 lustre_msg_get_opc(req->rq_reqmsg));
1629 rc = ptl_send_rpc(req, 0);
1630 if (rc == -ENOMEM) {
1631 spin_lock(&imp->imp_lock);
1632 if (!list_empty(&req->rq_list)) {
1633 list_del_init(&req->rq_list);
1634 atomic_dec(&req->rq_import->imp_inflight);
1636 spin_unlock(&imp->imp_lock);
1637 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1641 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1642 spin_lock(&req->rq_lock);
1643 req->rq_net_err = 1;
1644 spin_unlock(&req->rq_lock);
1650 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1655 LASSERT(set->set_producer != NULL);
1657 remaining = atomic_read(&set->set_remaining);
1659 /* populate the ->set_requests list with requests until we
1660 * reach the maximum number of RPCs in flight for this set */
1661 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1662 rc = set->set_producer(set, set->set_producer_arg);
1663 if (rc == -ENOENT) {
1664 /* no more RPC to produce */
1665 set->set_producer = NULL;
1666 set->set_producer_arg = NULL;
1671 RETURN((atomic_read(&set->set_remaining) - remaining));
1675 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1676 * and no more replies are expected.
1677 * (it is possible to get less replies than requests sent e.g. due to timed out
1678 * requests or requests that we had trouble to send out)
1680 * NOTE: This function contains a potential schedule point (cond_resched()).
1682 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1684 struct list_head *tmp, *next;
1685 struct list_head comp_reqs;
1686 int force_timer_recalc = 0;
1689 if (atomic_read(&set->set_remaining) == 0)
1692 INIT_LIST_HEAD(&comp_reqs);
1693 list_for_each_safe(tmp, next, &set->set_requests) {
1694 struct ptlrpc_request *req =
1695 list_entry(tmp, struct ptlrpc_request,
1697 struct obd_import *imp = req->rq_import;
1698 int unregistered = 0;
1701 /* This schedule point is mainly for the ptlrpcd caller of this
1702 * function. Most ptlrpc sets are not long-lived and unbounded
1703 * in length, but at the least the set used by the ptlrpcd is.
1704 * Since the processing time is unbounded, we need to insert an
1705 * explicit schedule point to make the thread well-behaved.
1709 if (req->rq_phase == RQ_PHASE_NEW &&
1710 ptlrpc_send_new_req(req)) {
1711 force_timer_recalc = 1;
1714 /* delayed send - skip */
1715 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1718 /* delayed resend - skip */
1719 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1720 req->rq_sent > cfs_time_current_sec())
1723 if (!(req->rq_phase == RQ_PHASE_RPC ||
1724 req->rq_phase == RQ_PHASE_BULK ||
1725 req->rq_phase == RQ_PHASE_INTERPRET ||
1726 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1727 req->rq_phase == RQ_PHASE_COMPLETE)) {
1728 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1732 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1733 LASSERT(req->rq_next_phase != req->rq_phase);
1734 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1737 * Skip processing until reply is unlinked. We
1738 * can't return to pool before that and we can't
1739 * call interpret before that. We need to make
1740 * sure that all rdma transfers finished and will
1741 * not corrupt any data.
1743 if (ptlrpc_client_recv_or_unlink(req) ||
1744 ptlrpc_client_bulk_active(req))
1748 * Turn fail_loc off to prevent it from looping
1751 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1752 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1755 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1756 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1761 * Move to next phase if reply was successfully
1764 ptlrpc_rqphase_move(req, req->rq_next_phase);
1767 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1768 list_move_tail(&req->rq_set_chain, &comp_reqs);
1772 if (req->rq_phase == RQ_PHASE_INTERPRET)
1773 GOTO(interpret, req->rq_status);
1776 * Note that this also will start async reply unlink.
1778 if (req->rq_net_err && !req->rq_timedout) {
1779 ptlrpc_expire_one_request(req, 1);
1782 * Check if we still need to wait for unlink.
1784 if (ptlrpc_client_recv_or_unlink(req) ||
1785 ptlrpc_client_bulk_active(req))
1787 /* If there is no need to resend, fail it now. */
1788 if (req->rq_no_resend) {
1789 if (req->rq_status == 0)
1790 req->rq_status = -EIO;
1791 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1792 GOTO(interpret, req->rq_status);
1799 spin_lock(&req->rq_lock);
1800 req->rq_replied = 0;
1801 spin_unlock(&req->rq_lock);
1802 if (req->rq_status == 0)
1803 req->rq_status = -EIO;
1804 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1805 GOTO(interpret, req->rq_status);
1808 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1809 * so it sets rq_intr regardless of individual rpc
1810 * timeouts. The synchronous IO waiting path sets
1811 * rq_intr irrespective of whether ptlrpcd
1812 * has seen a timeout. Our policy is to only interpret
1813 * interrupted rpcs after they have timed out, so we
1814 * need to enforce that here.
1817 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1818 req->rq_wait_ctx)) {
1819 req->rq_status = -EINTR;
1820 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1821 GOTO(interpret, req->rq_status);
1824 if (req->rq_phase == RQ_PHASE_RPC) {
1825 if (req->rq_timedout || req->rq_resend ||
1826 req->rq_waiting || req->rq_wait_ctx) {
1829 if (!ptlrpc_unregister_reply(req, 1)) {
1830 ptlrpc_unregister_bulk(req, 1);
1834 spin_lock(&imp->imp_lock);
1835 if (ptlrpc_import_delay_req(imp, req, &status)){
1836 /* put on delay list - only if we wait
1837 * recovery finished - before send */
1838 list_del_init(&req->rq_list);
1839 list_add_tail(&req->rq_list,
1842 spin_unlock(&imp->imp_lock);
1847 req->rq_status = status;
1848 ptlrpc_rqphase_move(req,
1849 RQ_PHASE_INTERPRET);
1850 spin_unlock(&imp->imp_lock);
1851 GOTO(interpret, req->rq_status);
1853 if (ptlrpc_no_resend(req) &&
1854 !req->rq_wait_ctx) {
1855 req->rq_status = -ENOTCONN;
1856 ptlrpc_rqphase_move(req,
1857 RQ_PHASE_INTERPRET);
1858 spin_unlock(&imp->imp_lock);
1859 GOTO(interpret, req->rq_status);
1862 list_del_init(&req->rq_list);
1863 list_add_tail(&req->rq_list,
1864 &imp->imp_sending_list);
1866 spin_unlock(&imp->imp_lock);
1868 spin_lock(&req->rq_lock);
1869 req->rq_waiting = 0;
1870 spin_unlock(&req->rq_lock);
1872 if (req->rq_timedout || req->rq_resend) {
1873 /* This is re-sending anyways,
1874 * let's mark req as resend. */
1875 spin_lock(&req->rq_lock);
1877 spin_unlock(&req->rq_lock);
1879 if (req->rq_bulk != NULL &&
1880 !ptlrpc_unregister_bulk(req, 1))
1884 * rq_wait_ctx is only touched by ptlrpcd,
1885 * so no lock is needed here.
1887 status = sptlrpc_req_refresh_ctx(req, -1);
1890 req->rq_status = status;
1891 spin_lock(&req->rq_lock);
1892 req->rq_wait_ctx = 0;
1893 spin_unlock(&req->rq_lock);
1894 force_timer_recalc = 1;
1896 spin_lock(&req->rq_lock);
1897 req->rq_wait_ctx = 1;
1898 spin_unlock(&req->rq_lock);
1903 spin_lock(&req->rq_lock);
1904 req->rq_wait_ctx = 0;
1905 spin_unlock(&req->rq_lock);
1908 rc = ptl_send_rpc(req, 0);
1909 if (rc == -ENOMEM) {
1910 spin_lock(&imp->imp_lock);
1911 if (!list_empty(&req->rq_list))
1912 list_del_init(&req->rq_list);
1913 spin_unlock(&imp->imp_lock);
1914 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1918 DEBUG_REQ(D_HA, req,
1919 "send failed: rc = %d", rc);
1920 force_timer_recalc = 1;
1921 spin_lock(&req->rq_lock);
1922 req->rq_net_err = 1;
1923 spin_unlock(&req->rq_lock);
1926 /* need to reset the timeout */
1927 force_timer_recalc = 1;
1930 spin_lock(&req->rq_lock);
1932 if (ptlrpc_client_early(req)) {
1933 ptlrpc_at_recv_early_reply(req);
1934 spin_unlock(&req->rq_lock);
1938 /* Still waiting for a reply? */
1939 if (ptlrpc_client_recv(req)) {
1940 spin_unlock(&req->rq_lock);
1944 /* Did we actually receive a reply? */
1945 if (!ptlrpc_client_replied(req)) {
1946 spin_unlock(&req->rq_lock);
1950 spin_unlock(&req->rq_lock);
1952 /* unlink from net because we are going to
1953 * swab in-place of reply buffer */
1954 unregistered = ptlrpc_unregister_reply(req, 1);
1958 req->rq_status = after_reply(req);
1962 /* If there is no bulk associated with this request,
1963 * then we're done and should let the interpreter
1964 * process the reply. Similarly if the RPC returned
1965 * an error, and therefore the bulk will never arrive.
1967 if (req->rq_bulk == NULL || req->rq_status < 0) {
1968 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1969 GOTO(interpret, req->rq_status);
1972 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1975 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1976 if (ptlrpc_client_bulk_active(req))
1979 if (req->rq_bulk->bd_failure) {
1980 /* The RPC reply arrived OK, but the bulk screwed
1981 * up! Dead weird since the server told us the RPC
1982 * was good after getting the REPLY for her GET or
1983 * the ACK for her PUT. */
1984 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1985 req->rq_status = -EIO;
1988 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1991 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1993 /* This moves to "unregistering" phase we need to wait for
1995 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1996 /* start async bulk unlink too */
1997 ptlrpc_unregister_bulk(req, 1);
2001 if (!ptlrpc_unregister_bulk(req, 1))
2004 /* When calling interpret receiving already should be
2006 LASSERT(!req->rq_receiving_reply);
2008 ptlrpc_req_interpret(env, req, req->rq_status);
2010 if (ptlrpcd_check_work(req)) {
2011 atomic_dec(&set->set_remaining);
2014 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
2016 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
2017 "Completed RPC pname:cluuid:pid:xid:nid:"
2018 "opc %s:%s:%d:"LPU64":%s:%d\n",
2019 current_comm(), imp->imp_obd->obd_uuid.uuid,
2020 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
2021 libcfs_nid2str(imp->imp_connection->c_peer.nid),
2022 lustre_msg_get_opc(req->rq_reqmsg));
2024 spin_lock(&imp->imp_lock);
2025 /* Request already may be not on sending or delaying list. This
2026 * may happen in the case of marking it erroneous for the case
2027 * ptlrpc_import_delay_req(req, status) find it impossible to
2028 * allow sending this rpc and returns *status != 0. */
2029 if (!list_empty(&req->rq_list)) {
2030 list_del_init(&req->rq_list);
2031 atomic_dec(&imp->imp_inflight);
2033 list_del_init(&req->rq_unreplied_list);
2034 spin_unlock(&imp->imp_lock);
2036 atomic_dec(&set->set_remaining);
2037 wake_up_all(&imp->imp_recovery_waitq);
2039 if (set->set_producer) {
2040 /* produce a new request if possible */
2041 if (ptlrpc_set_producer(set) > 0)
2042 force_timer_recalc = 1;
2044 /* free the request that has just been completed
2045 * in order not to pollute set->set_requests */
2046 list_del_init(&req->rq_set_chain);
2047 spin_lock(&req->rq_lock);
2049 req->rq_invalid_rqset = 0;
2050 spin_unlock(&req->rq_lock);
2052 /* record rq_status to compute the final status later */
2053 if (req->rq_status != 0)
2054 set->set_rc = req->rq_status;
2055 ptlrpc_req_finished(req);
2057 list_move_tail(&req->rq_set_chain, &comp_reqs);
2061 /* move completed request at the head of list so it's easier for
2062 * caller to find them */
2063 list_splice(&comp_reqs, &set->set_requests);
2065 /* If we hit an error, we want to recover promptly. */
2066 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2068 EXPORT_SYMBOL(ptlrpc_check_set);
2071 * Time out request \a req. is \a async_unlink is set, that means do not wait
2072 * until LNet actually confirms network buffer unlinking.
2073 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2075 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2077 struct obd_import *imp = req->rq_import;
2081 spin_lock(&req->rq_lock);
2082 req->rq_timedout = 1;
2083 spin_unlock(&req->rq_lock);
2085 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
2086 "/real "CFS_DURATION_T"]",
2087 req->rq_net_err ? "failed due to network error" :
2088 ((req->rq_real_sent == 0 ||
2089 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
2090 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
2091 "timed out for sent delay" : "timed out for slow reply"),
2092 req->rq_sent, req->rq_real_sent);
2094 if (imp != NULL && obd_debug_peer_on_timeout)
2095 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
2097 ptlrpc_unregister_reply(req, async_unlink);
2098 ptlrpc_unregister_bulk(req, async_unlink);
2100 if (obd_dump_on_timeout)
2101 libcfs_debug_dumplog();
2104 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2108 atomic_inc(&imp->imp_timeouts);
2110 /* The DLM server doesn't want recovery run on its imports. */
2111 if (imp->imp_dlm_fake)
2114 /* If this request is for recovery or other primordial tasks,
2115 * then error it out here. */
2116 if (req->rq_ctx_init || req->rq_ctx_fini ||
2117 req->rq_send_state != LUSTRE_IMP_FULL ||
2118 imp->imp_obd->obd_no_recov) {
2119 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2120 ptlrpc_import_state_name(req->rq_send_state),
2121 ptlrpc_import_state_name(imp->imp_state));
2122 spin_lock(&req->rq_lock);
2123 req->rq_status = -ETIMEDOUT;
2125 spin_unlock(&req->rq_lock);
2129 /* if a request can't be resent we can't wait for an answer after
2131 if (ptlrpc_no_resend(req)) {
2132 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2136 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2142 * Time out all uncompleted requests in request set pointed by \a data
2143 * Callback used when waiting on sets with l_wait_event.
2146 int ptlrpc_expired_set(void *data)
2148 struct ptlrpc_request_set *set = data;
2149 struct list_head *tmp;
2150 time_t now = cfs_time_current_sec();
2153 LASSERT(set != NULL);
2156 * A timeout expired. See which reqs it applies to...
2158 list_for_each(tmp, &set->set_requests) {
2159 struct ptlrpc_request *req =
2160 list_entry(tmp, struct ptlrpc_request,
2163 /* don't expire request waiting for context */
2164 if (req->rq_wait_ctx)
2167 /* Request in-flight? */
2168 if (!((req->rq_phase == RQ_PHASE_RPC &&
2169 !req->rq_waiting && !req->rq_resend) ||
2170 (req->rq_phase == RQ_PHASE_BULK)))
2173 if (req->rq_timedout || /* already dealt with */
2174 req->rq_deadline > now) /* not expired */
2177 /* Deal with this guy. Do it asynchronously to not block
2178 * ptlrpcd thread. */
2179 ptlrpc_expire_one_request(req, 1);
2183 * When waiting for a whole set, we always break out of the
2184 * sleep so we can recalculate the timeout, or enable interrupts
2185 * if everyone's timed out.
2191 * Sets rq_intr flag in \a req under spinlock.
2193 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2195 spin_lock(&req->rq_lock);
2197 spin_unlock(&req->rq_lock);
2199 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2202 * Interrupts (sets interrupted flag) all uncompleted requests in
2203 * a set \a data. Callback for l_wait_event for interruptible waits.
2205 static void ptlrpc_interrupted_set(void *data)
2207 struct ptlrpc_request_set *set = data;
2208 struct list_head *tmp;
2210 LASSERT(set != NULL);
2211 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2213 list_for_each(tmp, &set->set_requests) {
2214 struct ptlrpc_request *req =
2215 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2217 if (req->rq_phase != RQ_PHASE_RPC &&
2218 req->rq_phase != RQ_PHASE_UNREGISTERING)
2221 ptlrpc_mark_interrupted(req);
2226 * Get the smallest timeout in the set; this does NOT set a timeout.
2228 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2230 struct list_head *tmp;
2231 time_t now = cfs_time_current_sec();
2233 struct ptlrpc_request *req;
2237 list_for_each(tmp, &set->set_requests) {
2238 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2241 * Request in-flight?
2243 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2244 (req->rq_phase == RQ_PHASE_BULK) ||
2245 (req->rq_phase == RQ_PHASE_NEW)))
2249 * Already timed out.
2251 if (req->rq_timedout)
2257 if (req->rq_wait_ctx)
2260 if (req->rq_phase == RQ_PHASE_NEW)
2261 deadline = req->rq_sent;
2262 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2263 deadline = req->rq_sent;
2265 deadline = req->rq_sent + req->rq_timeout;
2267 if (deadline <= now) /* actually expired already */
2268 timeout = 1; /* ASAP */
2269 else if (timeout == 0 || timeout > deadline - now)
2270 timeout = deadline - now;
2276 * Send all unset request from the set and then wait untill all
2277 * requests in the set complete (either get a reply, timeout, get an
2278 * error or otherwise be interrupted).
2279 * Returns 0 on success or error code otherwise.
2281 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2283 struct list_head *tmp;
2284 struct ptlrpc_request *req;
2285 struct l_wait_info lwi;
2289 if (set->set_producer)
2290 (void)ptlrpc_set_producer(set);
2292 list_for_each(tmp, &set->set_requests) {
2293 req = list_entry(tmp, struct ptlrpc_request,
2295 if (req->rq_phase == RQ_PHASE_NEW)
2296 (void)ptlrpc_send_new_req(req);
2299 if (list_empty(&set->set_requests))
2303 timeout = ptlrpc_set_next_timeout(set);
2305 /* wait until all complete, interrupted, or an in-flight
2307 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2310 if (timeout == 0 && !signal_pending(current))
2312 * No requests are in-flight (ether timed out
2313 * or delayed), so we can allow interrupts.
2314 * We still want to block for a limited time,
2315 * so we allow interrupts during the timeout.
2317 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2319 ptlrpc_interrupted_set, set);
2322 * At least one request is in flight, so no
2323 * interrupts are allowed. Wait until all
2324 * complete, or an in-flight req times out.
2326 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2327 ptlrpc_expired_set, set);
2329 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2331 /* LU-769 - if we ignored the signal because it was already
2332 * pending when we started, we need to handle it now or we risk
2333 * it being ignored forever */
2334 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2335 signal_pending(current)) {
2336 sigset_t blocked_sigs =
2337 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2339 /* In fact we only interrupt for the "fatal" signals
2340 * like SIGINT or SIGKILL. We still ignore less
2341 * important signals since ptlrpc set is not easily
2342 * reentrant from userspace again */
2343 if (signal_pending(current))
2344 ptlrpc_interrupted_set(set);
2345 cfs_restore_sigs(blocked_sigs);
2348 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2350 /* -EINTR => all requests have been flagged rq_intr so next
2352 * -ETIMEDOUT => someone timed out. When all reqs have
2353 * timed out, signals are enabled allowing completion with
2355 * I don't really care if we go once more round the loop in
2356 * the error cases -eeb. */
2357 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2358 list_for_each(tmp, &set->set_requests) {
2359 req = list_entry(tmp, struct ptlrpc_request,
2361 spin_lock(&req->rq_lock);
2362 req->rq_invalid_rqset = 1;
2363 spin_unlock(&req->rq_lock);
2366 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2368 LASSERT(atomic_read(&set->set_remaining) == 0);
2370 rc = set->set_rc; /* rq_status of already freed requests if any */
2371 list_for_each(tmp, &set->set_requests) {
2372 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2374 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2375 if (req->rq_status != 0)
2376 rc = req->rq_status;
2379 if (set->set_interpret != NULL) {
2380 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2382 rc = interpreter (set, set->set_arg, rc);
2384 struct ptlrpc_set_cbdata *cbdata, *n;
2387 list_for_each_entry_safe(cbdata, n,
2388 &set->set_cblist, psc_item) {
2389 list_del_init(&cbdata->psc_item);
2390 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2393 OBD_FREE_PTR(cbdata);
2399 EXPORT_SYMBOL(ptlrpc_set_wait);
2402 * Helper fuction for request freeing.
2403 * Called when request count reached zero and request needs to be freed.
2404 * Removes request from all sorts of sending/replay lists it might be on,
2405 * frees network buffers if any are present.
2406 * If \a locked is set, that means caller is already holding import imp_lock
2407 * and so we no longer need to reobtain it (for certain lists manipulations)
2409 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2413 if (request == NULL)
2416 LASSERT(!request->rq_srv_req);
2417 LASSERT(request->rq_export == NULL);
2418 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2419 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2420 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2421 LASSERTF(!request->rq_replay, "req %p\n", request);
2423 req_capsule_fini(&request->rq_pill);
2425 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2426 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2427 if (request->rq_import != NULL) {
2429 spin_lock(&request->rq_import->imp_lock);
2430 list_del_init(&request->rq_replay_list);
2431 list_del_init(&request->rq_unreplied_list);
2433 spin_unlock(&request->rq_import->imp_lock);
2435 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2437 if (atomic_read(&request->rq_refcount) != 0) {
2438 DEBUG_REQ(D_ERROR, request,
2439 "freeing request with nonzero refcount");
2443 if (request->rq_repbuf != NULL)
2444 sptlrpc_cli_free_repbuf(request);
2446 if (request->rq_import != NULL) {
2447 class_import_put(request->rq_import);
2448 request->rq_import = NULL;
2450 if (request->rq_bulk != NULL)
2451 ptlrpc_free_bulk(request->rq_bulk);
2453 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2454 sptlrpc_cli_free_reqbuf(request);
2456 if (request->rq_cli_ctx)
2457 sptlrpc_req_put_ctx(request, !locked);
2459 if (request->rq_pool)
2460 __ptlrpc_free_req_to_pool(request);
2462 ptlrpc_request_cache_free(request);
2466 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2468 * Drop one request reference. Must be called with import imp_lock held.
2469 * When reference count drops to zero, request is freed.
2471 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2473 assert_spin_locked(&request->rq_import->imp_lock);
2474 (void)__ptlrpc_req_finished(request, 1);
2479 * Drops one reference count for request \a request.
2480 * \a locked set indicates that caller holds import imp_lock.
2481 * Frees the request whe reference count reaches zero.
2483 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2486 if (request == NULL)
2489 if (request == LP_POISON ||
2490 request->rq_reqmsg == LP_POISON) {
2491 CERROR("dereferencing freed request (bug 575)\n");
2496 DEBUG_REQ(D_INFO, request, "refcount now %u",
2497 atomic_read(&request->rq_refcount) - 1);
2499 if (atomic_dec_and_test(&request->rq_refcount)) {
2500 __ptlrpc_free_req(request, locked);
2508 * Drops one reference count for a request.
2510 void ptlrpc_req_finished(struct ptlrpc_request *request)
2512 __ptlrpc_req_finished(request, 0);
2514 EXPORT_SYMBOL(ptlrpc_req_finished);
2517 * Returns xid of a \a request
2519 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2521 return request->rq_xid;
2523 EXPORT_SYMBOL(ptlrpc_req_xid);
2526 * Disengage the client's reply buffer from the network
2527 * NB does _NOT_ unregister any client-side bulk.
2528 * IDEMPOTENT, but _not_ safe against concurrent callers.
2529 * The request owner (i.e. the thread doing the I/O) must call...
2530 * Returns 0 on success or 1 if unregistering cannot be made.
2532 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2535 struct l_wait_info lwi;
2540 LASSERT(!in_interrupt());
2543 * Let's setup deadline for reply unlink.
2545 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2546 async && request->rq_reply_deadline == 0)
2547 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2550 * Nothing left to do.
2552 if (!ptlrpc_client_recv_or_unlink(request))
2555 LNetMDUnlink(request->rq_reply_md_h);
2558 * Let's check it once again.
2560 if (!ptlrpc_client_recv_or_unlink(request))
2564 * Move to "Unregistering" phase as reply was not unlinked yet.
2566 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2569 * Do not wait for unlink to finish.
2575 * We have to l_wait_event() whatever the result, to give liblustre
2576 * a chance to run reply_in_callback(), and to make sure we've
2577 * unlinked before returning a req to the pool.
2580 /* The wq argument is ignored by user-space wait_event macros */
2581 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2582 &request->rq_set->set_waitq :
2583 &request->rq_reply_waitq;
2584 /* Network access will complete in finite time but the HUGE
2585 * timeout lets us CWARN for visibility of sluggish NALs */
2586 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2587 cfs_time_seconds(1), NULL, NULL);
2588 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2591 ptlrpc_rqphase_move(request, request->rq_next_phase);
2595 LASSERT(rc == -ETIMEDOUT);
2596 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2597 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2598 request->rq_receiving_reply,
2599 request->rq_req_unlinked,
2600 request->rq_reply_unlinked);
2605 static void ptlrpc_free_request(struct ptlrpc_request *req)
2607 spin_lock(&req->rq_lock);
2609 spin_unlock(&req->rq_lock);
2611 if (req->rq_commit_cb != NULL)
2612 req->rq_commit_cb(req);
2613 list_del_init(&req->rq_replay_list);
2615 __ptlrpc_req_finished(req, 1);
2619 * the request is committed and dropped from the replay list of its import
2621 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2623 struct obd_import *imp = req->rq_import;
2625 spin_lock(&imp->imp_lock);
2626 if (list_empty(&req->rq_replay_list)) {
2627 spin_unlock(&imp->imp_lock);
2631 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2632 ptlrpc_free_request(req);
2634 spin_unlock(&imp->imp_lock);
2636 EXPORT_SYMBOL(ptlrpc_request_committed);
2639 * Iterates through replay_list on import and prunes
2640 * all requests have transno smaller than last_committed for the
2641 * import and don't have rq_replay set.
2642 * Since requests are sorted in transno order, stops when meetign first
2643 * transno bigger than last_committed.
2644 * caller must hold imp->imp_lock
2646 void ptlrpc_free_committed(struct obd_import *imp)
2648 struct ptlrpc_request *req, *saved;
2649 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2650 bool skip_committed_list = true;
2653 LASSERT(imp != NULL);
2654 assert_spin_locked(&imp->imp_lock);
2656 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2657 imp->imp_generation == imp->imp_last_generation_checked) {
2658 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2659 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2662 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2663 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2664 imp->imp_generation);
2666 if (imp->imp_generation != imp->imp_last_generation_checked ||
2667 imp->imp_last_transno_checked == 0)
2668 skip_committed_list = false;
2670 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2671 imp->imp_last_generation_checked = imp->imp_generation;
2673 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2675 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2676 LASSERT(req != last_req);
2679 if (req->rq_transno == 0) {
2680 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2683 if (req->rq_import_generation < imp->imp_generation) {
2684 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2688 /* not yet committed */
2689 if (req->rq_transno > imp->imp_peer_committed_transno) {
2690 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2694 if (req->rq_replay) {
2695 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2696 list_move_tail(&req->rq_replay_list,
2697 &imp->imp_committed_list);
2701 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2702 imp->imp_peer_committed_transno);
2704 ptlrpc_free_request(req);
2707 if (skip_committed_list)
2710 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2712 LASSERT(req->rq_transno != 0);
2713 if (req->rq_import_generation < imp->imp_generation) {
2714 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2715 ptlrpc_free_request(req);
2716 } else if (!req->rq_replay) {
2717 DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2718 ptlrpc_free_request(req);
2725 void ptlrpc_cleanup_client(struct obd_import *imp)
2732 * Schedule previously sent request for resend.
2733 * For bulk requests we assign new xid (to avoid problems with
2734 * lost replies and therefore several transfers landing into same buffer
2735 * from different sending attempts).
2737 void ptlrpc_resend_req(struct ptlrpc_request *req)
2739 DEBUG_REQ(D_HA, req, "going to resend");
2740 spin_lock(&req->rq_lock);
2742 /* Request got reply but linked to the import list still.
2743 Let ptlrpc_check_set() to process it. */
2744 if (ptlrpc_client_replied(req)) {
2745 spin_unlock(&req->rq_lock);
2746 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2750 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2751 req->rq_status = -EAGAIN;
2754 req->rq_net_err = 0;
2755 req->rq_timedout = 0;
2757 ptlrpc_client_wake_req(req);
2758 spin_unlock(&req->rq_lock);
2761 /* XXX: this function and rq_status are currently unused */
2762 void ptlrpc_restart_req(struct ptlrpc_request *req)
2764 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2765 req->rq_status = -ERESTARTSYS;
2767 spin_lock(&req->rq_lock);
2768 req->rq_restart = 1;
2769 req->rq_timedout = 0;
2770 ptlrpc_client_wake_req(req);
2771 spin_unlock(&req->rq_lock);
2775 * Grab additional reference on a request \a req
2777 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2780 atomic_inc(&req->rq_refcount);
2783 EXPORT_SYMBOL(ptlrpc_request_addref);
2786 * Add a request to import replay_list.
2787 * Must be called under imp_lock
2789 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2790 struct obd_import *imp)
2792 struct list_head *tmp;
2794 assert_spin_locked(&imp->imp_lock);
2796 if (req->rq_transno == 0) {
2797 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2801 /* clear this for new requests that were resent as well
2802 as resent replayed requests. */
2803 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2805 /* don't re-add requests that have been replayed */
2806 if (!list_empty(&req->rq_replay_list))
2809 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2811 spin_lock(&req->rq_lock);
2813 spin_unlock(&req->rq_lock);
2815 LASSERT(imp->imp_replayable);
2816 /* Balanced in ptlrpc_free_committed, usually. */
2817 ptlrpc_request_addref(req);
2818 list_for_each_prev(tmp, &imp->imp_replay_list) {
2819 struct ptlrpc_request *iter = list_entry(tmp,
2820 struct ptlrpc_request,
2823 /* We may have duplicate transnos if we create and then
2824 * open a file, or for closes retained if to match creating
2825 * opens, so use req->rq_xid as a secondary key.
2826 * (See bugs 684, 685, and 428.)
2827 * XXX no longer needed, but all opens need transnos!
2829 if (iter->rq_transno > req->rq_transno)
2832 if (iter->rq_transno == req->rq_transno) {
2833 LASSERT(iter->rq_xid != req->rq_xid);
2834 if (iter->rq_xid > req->rq_xid)
2838 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2842 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2846 * Send request and wait until it completes.
2847 * Returns request processing status.
2849 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2851 struct ptlrpc_request_set *set;
2855 LASSERT(req->rq_set == NULL);
2856 LASSERT(!req->rq_receiving_reply);
2858 set = ptlrpc_prep_set();
2860 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2864 /* for distributed debugging */
2865 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2867 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2868 ptlrpc_request_addref(req);
2869 ptlrpc_set_add_req(set, req);
2870 rc = ptlrpc_set_wait(set);
2871 ptlrpc_set_destroy(set);
2875 EXPORT_SYMBOL(ptlrpc_queue_wait);
2878 * Callback used for replayed requests reply processing.
2879 * In case of successful reply calls registered request replay callback.
2880 * In case of error restart replay process.
2882 static int ptlrpc_replay_interpret(const struct lu_env *env,
2883 struct ptlrpc_request *req,
2884 void * data, int rc)
2886 struct ptlrpc_replay_async_args *aa = data;
2887 struct obd_import *imp = req->rq_import;
2890 atomic_dec(&imp->imp_replay_inflight);
2892 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2893 * server got the request, but bulk transfer timeout, let's
2894 * replay the bulk req again */
2895 if (!ptlrpc_client_replied(req) ||
2896 (req->rq_bulk != NULL &&
2897 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2898 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2899 GOTO(out, rc = -ETIMEDOUT);
2902 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2903 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2904 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2905 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2907 /** VBR: check version failure */
2908 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2909 /** replay was failed due to version mismatch */
2910 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2911 spin_lock(&imp->imp_lock);
2912 imp->imp_vbr_failed = 1;
2913 imp->imp_no_lock_replay = 1;
2914 spin_unlock(&imp->imp_lock);
2915 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2917 /** The transno had better not change over replay. */
2918 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2919 lustre_msg_get_transno(req->rq_repmsg) ||
2920 lustre_msg_get_transno(req->rq_repmsg) == 0,
2922 lustre_msg_get_transno(req->rq_reqmsg),
2923 lustre_msg_get_transno(req->rq_repmsg));
2926 spin_lock(&imp->imp_lock);
2927 /** if replays by version then gap occur on server, no trust to locks */
2928 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2929 imp->imp_no_lock_replay = 1;
2930 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2931 spin_unlock(&imp->imp_lock);
2932 LASSERT(imp->imp_last_replay_transno);
2934 /* transaction number shouldn't be bigger than the latest replayed */
2935 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2936 DEBUG_REQ(D_ERROR, req,
2937 "Reported transno "LPU64" is bigger than the "
2938 "replayed one: "LPU64, req->rq_transno,
2939 lustre_msg_get_transno(req->rq_reqmsg));
2940 GOTO(out, rc = -EINVAL);
2943 DEBUG_REQ(D_HA, req, "got rep");
2945 /* let the callback do fixups, possibly including in the request */
2946 if (req->rq_replay_cb)
2947 req->rq_replay_cb(req);
2949 if (ptlrpc_client_replied(req) &&
2950 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2951 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2952 lustre_msg_get_status(req->rq_repmsg),
2953 aa->praa_old_status);
2955 /* Put it back for re-replay. */
2956 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2960 * Errors while replay can set transno to 0, but
2961 * imp_last_replay_transno shouldn't be set to 0 anyway
2963 if (req->rq_transno == 0)
2964 CERROR("Transno is 0 during replay!\n");
2966 /* continue with recovery */
2967 rc = ptlrpc_import_recovery_state_machine(imp);
2969 req->rq_send_state = aa->praa_old_state;
2972 /* this replay failed, so restart recovery */
2973 ptlrpc_connect_import(imp);
2979 * Prepares and queues request for replay.
2980 * Adds it to ptlrpcd queue for actual sending.
2981 * Returns 0 on success.
2983 int ptlrpc_replay_req(struct ptlrpc_request *req)
2985 struct ptlrpc_replay_async_args *aa;
2988 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2990 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2991 aa = ptlrpc_req_async_args(req);
2992 memset(aa, 0, sizeof *aa);
2994 /* Prepare request to be resent with ptlrpcd */
2995 aa->praa_old_state = req->rq_send_state;
2996 req->rq_send_state = LUSTRE_IMP_REPLAY;
2997 req->rq_phase = RQ_PHASE_NEW;
2998 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3000 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3002 req->rq_interpret_reply = ptlrpc_replay_interpret;
3003 /* Readjust the timeout for current conditions */
3004 ptlrpc_at_set_req_timeout(req);
3006 /* Tell server the net_latency, so the server can calculate how long
3007 * it should wait for next replay */
3008 lustre_msg_set_service_time(req->rq_reqmsg,
3009 ptlrpc_at_get_net_latency(req));
3010 DEBUG_REQ(D_HA, req, "REPLAY");
3012 atomic_inc(&req->rq_import->imp_replay_inflight);
3013 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3015 ptlrpcd_add_req(req);
3020 * Aborts all in-flight request on import \a imp sending and delayed lists
3022 void ptlrpc_abort_inflight(struct obd_import *imp)
3024 struct list_head *tmp, *n;
3027 /* Make sure that no new requests get processed for this import.
3028 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3029 * this flag and then putting requests on sending_list or delayed_list.
3031 spin_lock(&imp->imp_lock);
3033 /* XXX locking? Maybe we should remove each request with the list
3034 * locked? Also, how do we know if the requests on the list are
3035 * being freed at this time?
3037 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3038 struct ptlrpc_request *req = list_entry(tmp,
3039 struct ptlrpc_request,
3042 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3044 spin_lock(&req->rq_lock);
3045 if (req->rq_import_generation < imp->imp_generation) {
3047 req->rq_status = -EIO;
3048 ptlrpc_client_wake_req(req);
3050 spin_unlock(&req->rq_lock);
3053 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3054 struct ptlrpc_request *req =
3055 list_entry(tmp, struct ptlrpc_request, rq_list);
3057 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3059 spin_lock(&req->rq_lock);
3060 if (req->rq_import_generation < imp->imp_generation) {
3062 req->rq_status = -EIO;
3063 ptlrpc_client_wake_req(req);
3065 spin_unlock(&req->rq_lock);
3068 /* Last chance to free reqs left on the replay list, but we
3069 * will still leak reqs that haven't committed. */
3070 if (imp->imp_replayable)
3071 ptlrpc_free_committed(imp);
3073 spin_unlock(&imp->imp_lock);
3079 * Abort all uncompleted requests in request set \a set
3081 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3083 struct list_head *tmp, *pos;
3085 LASSERT(set != NULL);
3087 list_for_each_safe(pos, tmp, &set->set_requests) {
3088 struct ptlrpc_request *req =
3089 list_entry(pos, struct ptlrpc_request,
3092 spin_lock(&req->rq_lock);
3093 if (req->rq_phase != RQ_PHASE_RPC) {
3094 spin_unlock(&req->rq_lock);
3099 req->rq_status = -EINTR;
3100 ptlrpc_client_wake_req(req);
3101 spin_unlock(&req->rq_lock);
3105 static __u64 ptlrpc_last_xid;
3106 static spinlock_t ptlrpc_last_xid_lock;
3109 * Initialize the XID for the node. This is common among all requests on
3110 * this node, and only requires the property that it is monotonically
3111 * increasing. It does not need to be sequential. Since this is also used
3112 * as the RDMA match bits, it is important that a single client NOT have
3113 * the same match bits for two different in-flight requests, hence we do
3114 * NOT want to have an XID per target or similar.
3116 * To avoid an unlikely collision between match bits after a client reboot
3117 * (which would deliver old data into the wrong RDMA buffer) initialize
3118 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3119 * If the time is clearly incorrect, we instead use a 62-bit random number.
3120 * In the worst case the random number will overflow 1M RPCs per second in
3121 * 9133 years, or permutations thereof.
3123 #define YEAR_2004 (1ULL << 30)
3124 void ptlrpc_init_xid(void)
3126 time_t now = cfs_time_current_sec();
3128 spin_lock_init(&ptlrpc_last_xid_lock);
3129 if (now < YEAR_2004) {
3130 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3131 ptlrpc_last_xid >>= 2;
3132 ptlrpc_last_xid |= (1ULL << 61);
3134 ptlrpc_last_xid = (__u64)now << 20;
3137 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3138 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3139 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3143 * Increase xid and returns resulting new value to the caller.
3145 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3146 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3147 * itself uses the last bulk xid needed, so the server can determine the
3148 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3149 * xid must align to a power-of-two value.
3151 * This is assumed to be true due to the initial ptlrpc_last_xid
3152 * value also being initialized to a power-of-two value. LU-1431
3154 __u64 ptlrpc_next_xid(void)
3158 spin_lock(&ptlrpc_last_xid_lock);
3159 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3160 ptlrpc_last_xid = next;
3161 spin_unlock(&ptlrpc_last_xid_lock);
3167 * If request has a new allocated XID (new request or EINPROGRESS resend),
3168 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3169 * request to ensure previous bulk fails and avoid problems with lost replies
3170 * and therefore several transfers landing into the same buffer from different
3173 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3175 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3177 LASSERT(bd != NULL);
3179 if (!req->rq_resend) {
3180 /* this request has a new xid, just use it as bulk matchbits */
3181 req->rq_mbits = req->rq_xid;
3183 } else { /* needs to generate a new matchbits for resend */
3184 __u64 old_mbits = req->rq_mbits;
3186 if ((bd->bd_import->imp_connect_data.ocd_connect_flags &
3187 OBD_CONNECT_BULK_MBITS) != 0)
3188 req->rq_mbits = ptlrpc_next_xid();
3189 else /* old version transfers rq_xid to peer as matchbits */
3190 req->rq_mbits = req->rq_xid = ptlrpc_next_xid();
3192 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
3193 old_mbits, req->rq_mbits);
3196 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3197 * that server can infer the number of bulks that were prepared,
3199 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3204 * Get a glimpse at what next xid value might have been.
3205 * Returns possible next xid.
3207 __u64 ptlrpc_sample_next_xid(void)
3209 #if BITS_PER_LONG == 32
3210 /* need to avoid possible word tearing on 32-bit systems */
3213 spin_lock(&ptlrpc_last_xid_lock);
3214 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3215 spin_unlock(&ptlrpc_last_xid_lock);
3219 /* No need to lock, since returned value is racy anyways */
3220 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3223 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3226 * Functions for operating ptlrpc workers.
3228 * A ptlrpc work is a function which will be running inside ptlrpc context.
3229 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3231 * 1. after a work is created, it can be used many times, that is:
3232 * handler = ptlrpcd_alloc_work();
3233 * ptlrpcd_queue_work();
3235 * queue it again when necessary:
3236 * ptlrpcd_queue_work();
3237 * ptlrpcd_destroy_work();
3238 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3239 * it will only be queued once in any time. Also as its name implies, it may
3240 * have delay before it really runs by ptlrpcd thread.
3242 struct ptlrpc_work_async_args {
3243 int (*cb)(const struct lu_env *, void *);
3247 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3249 /* re-initialize the req */
3250 req->rq_timeout = obd_timeout;
3251 req->rq_sent = cfs_time_current_sec();
3252 req->rq_deadline = req->rq_sent + req->rq_timeout;
3253 req->rq_reply_deadline = req->rq_deadline;
3254 req->rq_phase = RQ_PHASE_INTERPRET;
3255 req->rq_next_phase = RQ_PHASE_COMPLETE;
3256 req->rq_xid = ptlrpc_next_xid();
3257 req->rq_import_generation = req->rq_import->imp_generation;
3259 ptlrpcd_add_req(req);
3262 static int work_interpreter(const struct lu_env *env,
3263 struct ptlrpc_request *req, void *data, int rc)
3265 struct ptlrpc_work_async_args *arg = data;
3267 LASSERT(ptlrpcd_check_work(req));
3268 LASSERT(arg->cb != NULL);
3270 rc = arg->cb(env, arg->cbdata);
3272 list_del_init(&req->rq_set_chain);
3275 if (atomic_dec_return(&req->rq_refcount) > 1) {
3276 atomic_set(&req->rq_refcount, 2);
3277 ptlrpcd_add_work_req(req);
3282 static int worker_format;
3284 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3286 return req->rq_pill.rc_fmt == (void *)&worker_format;
3290 * Create a work for ptlrpc.
3292 void *ptlrpcd_alloc_work(struct obd_import *imp,
3293 int (*cb)(const struct lu_env *, void *), void *cbdata)
3295 struct ptlrpc_request *req = NULL;
3296 struct ptlrpc_work_async_args *args;
3302 RETURN(ERR_PTR(-EINVAL));
3304 /* copy some code from deprecated fakereq. */
3305 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3307 CERROR("ptlrpc: run out of memory!\n");
3308 RETURN(ERR_PTR(-ENOMEM));
3311 ptlrpc_cli_req_init(req);
3313 req->rq_send_state = LUSTRE_IMP_FULL;
3314 req->rq_type = PTL_RPC_MSG_REQUEST;
3315 req->rq_import = class_import_get(imp);
3316 req->rq_interpret_reply = work_interpreter;
3317 /* don't want reply */
3318 req->rq_no_delay = req->rq_no_resend = 1;
3319 req->rq_pill.rc_fmt = (void *)&worker_format;
3321 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3322 args = ptlrpc_req_async_args(req);
3324 args->cbdata = cbdata;
3328 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3330 void ptlrpcd_destroy_work(void *handler)
3332 struct ptlrpc_request *req = handler;
3335 ptlrpc_req_finished(req);
3337 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3339 int ptlrpcd_queue_work(void *handler)
3341 struct ptlrpc_request *req = handler;
3344 * Check if the req is already being queued.
3346 * Here comes a trick: it lacks a way of checking if a req is being
3347 * processed reliably in ptlrpc. Here I have to use refcount of req
3348 * for this purpose. This is okay because the caller should use this
3349 * req as opaque data. - Jinshan
3351 LASSERT(atomic_read(&req->rq_refcount) > 0);
3352 if (atomic_inc_return(&req->rq_refcount) == 2)
3353 ptlrpcd_add_work_req(req);
3356 EXPORT_SYMBOL(ptlrpcd_queue_work);