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));
134 if (type & PTLRPC_BULK_BUF_KIOV) {
135 OBD_ALLOC_LARGE(GET_KIOV(desc),
136 nfrags * sizeof(*GET_KIOV(desc)));
137 if (GET_KIOV(desc) == NULL)
140 OBD_ALLOC_LARGE(GET_KVEC(desc),
141 nfrags * sizeof(*GET_KVEC(desc)));
142 if (GET_KVEC(desc) == NULL)
146 spin_lock_init(&desc->bd_lock);
147 init_waitqueue_head(&desc->bd_waitq);
148 desc->bd_max_iov = nfrags;
149 desc->bd_iov_count = 0;
150 desc->bd_portal = portal;
151 desc->bd_type = type;
152 desc->bd_md_count = 0;
153 desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
154 LASSERT(max_brw > 0);
155 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
156 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
157 * node. Negotiated ocd_brw_size will always be <= this number. */
158 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
159 LNetInvalidateHandle(&desc->bd_mds[i]);
168 * Prepare bulk descriptor for specified outgoing request \a req that
169 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
170 * the bulk to be sent. Used on client-side.
171 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
174 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
175 unsigned nfrags, unsigned max_brw,
178 const struct ptlrpc_bulk_frag_ops
181 struct obd_import *imp = req->rq_import;
182 struct ptlrpc_bulk_desc *desc;
185 LASSERT(ptlrpc_is_bulk_op_passive(type));
187 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
191 desc->bd_import_generation = req->rq_import_generation;
192 desc->bd_import = class_import_get(imp);
195 desc->bd_cbid.cbid_fn = client_bulk_callback;
196 desc->bd_cbid.cbid_arg = desc;
198 /* This makes req own desc, and free it when she frees herself */
203 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
205 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
206 struct page *page, int pageoffset, int len,
211 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
212 LASSERT(page != NULL);
213 LASSERT(pageoffset >= 0);
215 LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
216 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
218 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
223 page_cache_get(page);
225 kiov->kiov_page = page;
226 kiov->kiov_offset = pageoffset;
227 kiov->kiov_len = len;
229 desc->bd_iov_count++;
231 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
233 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
239 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
240 LASSERT(frag != NULL);
242 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
244 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
248 iovec->iov_base = frag;
249 iovec->iov_len = len;
251 desc->bd_iov_count++;
253 RETURN(desc->bd_nob);
255 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
257 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
261 LASSERT(desc != NULL);
262 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
263 LASSERT(desc->bd_md_count == 0); /* network hands off */
264 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
265 LASSERT(desc->bd_frag_ops != NULL);
267 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
268 sptlrpc_enc_pool_put_pages(desc);
271 class_export_put(desc->bd_export);
273 class_import_put(desc->bd_import);
275 if (desc->bd_frag_ops->release_frags != NULL)
276 desc->bd_frag_ops->release_frags(desc);
278 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
279 OBD_FREE_LARGE(GET_KIOV(desc),
280 desc->bd_max_iov * sizeof(*GET_KIOV(desc)));
282 OBD_FREE_LARGE(GET_KVEC(desc),
283 desc->bd_max_iov * sizeof(*GET_KVEC(desc)));
287 EXPORT_SYMBOL(ptlrpc_free_bulk);
290 * Set server timelimit for this req, i.e. how long are we willing to wait
291 * for reply before timing out this request.
293 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
299 LASSERT(req->rq_import);
302 /* non-AT settings */
304 * \a imp_server_timeout means this is reverse import and
305 * we send (currently only) ASTs to the client and cannot afford
306 * to wait too long for the reply, otherwise the other client
307 * (because of which we are sending this request) would
308 * timeout waiting for us
310 req->rq_timeout = req->rq_import->imp_server_timeout ?
311 obd_timeout / 2 : obd_timeout;
313 at = &req->rq_import->imp_at;
314 idx = import_at_get_index(req->rq_import,
315 req->rq_request_portal);
316 serv_est = at_get(&at->iat_service_estimate[idx]);
317 req->rq_timeout = at_est2timeout(serv_est);
319 /* We could get even fancier here, using history to predict increased
322 /* Let the server know what this RPC timeout is by putting it in the
324 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
326 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
328 /* Adjust max service estimate based on server value */
329 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
330 unsigned int serv_est)
336 LASSERT(req->rq_import);
337 at = &req->rq_import->imp_at;
339 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
340 /* max service estimates are tracked on the server side,
341 so just keep minimal history here */
342 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
344 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
345 "has changed from %d to %d\n",
346 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
347 oldse, at_get(&at->iat_service_estimate[idx]));
350 /* Expected network latency per remote node (secs) */
351 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
353 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
356 /* Adjust expected network latency */
357 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
358 unsigned int service_time)
360 unsigned int nl, oldnl;
362 time_t now = cfs_time_current_sec();
364 LASSERT(req->rq_import);
366 if (service_time > now - req->rq_sent + 3) {
367 /* bz16408, however, this can also happen if early reply
368 * is lost and client RPC is expired and resent, early reply
369 * or reply of original RPC can still be fit in reply buffer
370 * of resent RPC, now client is measuring time from the
371 * resent time, but server sent back service time of original
374 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
375 D_ADAPTTO : D_WARNING,
376 "Reported service time %u > total measured time "
377 CFS_DURATION_T"\n", service_time,
378 cfs_time_sub(now, req->rq_sent));
382 /* Network latency is total time less server processing time */
383 nl = max_t(int, now - req->rq_sent -
384 service_time, 0) + 1; /* st rounding */
385 at = &req->rq_import->imp_at;
387 oldnl = at_measured(&at->iat_net_latency, nl);
389 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
390 "has changed from %d to %d\n",
391 req->rq_import->imp_obd->obd_name,
393 &req->rq_import->imp_connection->c_remote_uuid),
394 oldnl, at_get(&at->iat_net_latency));
397 static int unpack_reply(struct ptlrpc_request *req)
401 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
402 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
404 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
409 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
411 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
418 * Handle an early reply message, called with the rq_lock held.
419 * If anything goes wrong just ignore it - same as if it never happened
421 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
422 __must_hold(&req->rq_lock)
424 struct ptlrpc_request *early_req;
430 spin_unlock(&req->rq_lock);
432 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
434 spin_lock(&req->rq_lock);
438 rc = unpack_reply(early_req);
440 sptlrpc_cli_finish_early_reply(early_req);
441 spin_lock(&req->rq_lock);
445 /* Use new timeout value just to adjust the local value for this
446 * request, don't include it into at_history. It is unclear yet why
447 * service time increased and should it be counted or skipped, e.g.
448 * that can be recovery case or some error or server, the real reply
449 * will add all new data if it is worth to add. */
450 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
451 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
453 /* Network latency can be adjusted, it is pure network delays */
454 ptlrpc_at_adj_net_latency(req,
455 lustre_msg_get_service_time(early_req->rq_repmsg));
457 sptlrpc_cli_finish_early_reply(early_req);
459 spin_lock(&req->rq_lock);
460 olddl = req->rq_deadline;
461 /* server assumes it now has rq_timeout from when the request
462 * arrived, so the client should give it at least that long.
463 * since we don't know the arrival time we'll use the original
465 req->rq_deadline = req->rq_sent + req->rq_timeout +
466 ptlrpc_at_get_net_latency(req);
468 DEBUG_REQ(D_ADAPTTO, req,
469 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
470 "("CFS_DURATION_T"s)", req->rq_early_count,
471 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
472 cfs_time_sub(req->rq_deadline, olddl));
477 static struct kmem_cache *request_cache;
479 int ptlrpc_request_cache_init(void)
481 request_cache = kmem_cache_create("ptlrpc_cache",
482 sizeof(struct ptlrpc_request),
483 0, SLAB_HWCACHE_ALIGN, NULL);
484 return request_cache == NULL ? -ENOMEM : 0;
487 void ptlrpc_request_cache_fini(void)
489 kmem_cache_destroy(request_cache);
492 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
494 struct ptlrpc_request *req;
496 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
500 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
502 OBD_SLAB_FREE_PTR(req, request_cache);
506 * Wind down request pool \a pool.
507 * Frees all requests from the pool too
509 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
511 struct list_head *l, *tmp;
512 struct ptlrpc_request *req;
514 LASSERT(pool != NULL);
516 spin_lock(&pool->prp_lock);
517 list_for_each_safe(l, tmp, &pool->prp_req_list) {
518 req = list_entry(l, struct ptlrpc_request, rq_list);
519 list_del(&req->rq_list);
520 LASSERT(req->rq_reqbuf);
521 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
522 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
523 ptlrpc_request_cache_free(req);
525 spin_unlock(&pool->prp_lock);
526 OBD_FREE(pool, sizeof(*pool));
528 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
531 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
533 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
538 while (size < pool->prp_rq_size)
541 LASSERTF(list_empty(&pool->prp_req_list) ||
542 size == pool->prp_rq_size,
543 "Trying to change pool size with nonempty pool "
544 "from %d to %d bytes\n", pool->prp_rq_size, size);
546 spin_lock(&pool->prp_lock);
547 pool->prp_rq_size = size;
548 for (i = 0; i < num_rq; i++) {
549 struct ptlrpc_request *req;
550 struct lustre_msg *msg;
552 spin_unlock(&pool->prp_lock);
553 req = ptlrpc_request_cache_alloc(GFP_NOFS);
556 OBD_ALLOC_LARGE(msg, size);
558 ptlrpc_request_cache_free(req);
561 req->rq_reqbuf = msg;
562 req->rq_reqbuf_len = size;
564 spin_lock(&pool->prp_lock);
565 list_add_tail(&req->rq_list, &pool->prp_req_list);
567 spin_unlock(&pool->prp_lock);
570 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
573 * Create and initialize new request pool with given attributes:
574 * \a num_rq - initial number of requests to create for the pool
575 * \a msgsize - maximum message size possible for requests in thid pool
576 * \a populate_pool - function to be called when more requests need to be added
578 * Returns pointer to newly created pool or NULL on error.
580 struct ptlrpc_request_pool *
581 ptlrpc_init_rq_pool(int num_rq, int msgsize,
582 int (*populate_pool)(struct ptlrpc_request_pool *, int))
584 struct ptlrpc_request_pool *pool;
586 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
590 /* Request next power of two for the allocation, because internally
591 kernel would do exactly this */
593 spin_lock_init(&pool->prp_lock);
594 INIT_LIST_HEAD(&pool->prp_req_list);
595 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
596 pool->prp_populate = populate_pool;
598 populate_pool(pool, num_rq);
602 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
605 * Fetches one request from pool \a pool
607 static struct ptlrpc_request *
608 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
610 struct ptlrpc_request *request;
611 struct lustre_msg *reqbuf;
616 spin_lock(&pool->prp_lock);
618 /* See if we have anything in a pool, and bail out if nothing,
619 * in writeout path, where this matters, this is safe to do, because
620 * nothing is lost in this case, and when some in-flight requests
621 * complete, this code will be called again. */
622 if (unlikely(list_empty(&pool->prp_req_list))) {
623 spin_unlock(&pool->prp_lock);
627 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
629 list_del_init(&request->rq_list);
630 spin_unlock(&pool->prp_lock);
632 LASSERT(request->rq_reqbuf);
633 LASSERT(request->rq_pool);
635 reqbuf = request->rq_reqbuf;
636 memset(request, 0, sizeof(*request));
637 request->rq_reqbuf = reqbuf;
638 request->rq_reqbuf_len = pool->prp_rq_size;
639 request->rq_pool = pool;
645 * Returns freed \a request to pool.
647 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
649 struct ptlrpc_request_pool *pool = request->rq_pool;
651 spin_lock(&pool->prp_lock);
652 LASSERT(list_empty(&request->rq_list));
653 LASSERT(!request->rq_receiving_reply);
654 list_add_tail(&request->rq_list, &pool->prp_req_list);
655 spin_unlock(&pool->prp_lock);
658 void ptlrpc_add_unreplied(struct ptlrpc_request *req)
660 struct obd_import *imp = req->rq_import;
661 struct list_head *tmp;
662 struct ptlrpc_request *iter;
664 assert_spin_locked(&imp->imp_lock);
665 LASSERT(list_empty(&req->rq_unreplied_list));
667 /* unreplied list is sorted by xid in ascending order */
668 list_for_each_prev(tmp, &imp->imp_unreplied_list) {
669 iter = list_entry(tmp, struct ptlrpc_request,
672 LASSERT(req->rq_xid != iter->rq_xid);
673 if (req->rq_xid < iter->rq_xid)
675 list_add(&req->rq_unreplied_list, &iter->rq_unreplied_list);
678 list_add(&req->rq_unreplied_list, &imp->imp_unreplied_list);
681 void ptlrpc_assign_next_xid_nolock(struct ptlrpc_request *req)
683 req->rq_xid = ptlrpc_next_xid();
684 ptlrpc_add_unreplied(req);
687 static inline void ptlrpc_assign_next_xid(struct ptlrpc_request *req)
689 spin_lock(&req->rq_import->imp_lock);
690 ptlrpc_assign_next_xid_nolock(req);
691 spin_unlock(&req->rq_import->imp_lock);
694 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
695 __u32 version, int opcode, char **bufs,
696 struct ptlrpc_cli_ctx *ctx)
699 struct obd_import *imp;
705 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
706 imp = request->rq_import;
707 lengths = request->rq_pill.rc_area[RCL_CLIENT];
710 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
712 rc = sptlrpc_req_get_ctx(request);
716 sptlrpc_req_set_flavor(request, opcode);
718 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
723 lustre_msg_add_version(request->rq_reqmsg, version);
724 request->rq_send_state = LUSTRE_IMP_FULL;
725 request->rq_type = PTL_RPC_MSG_REQUEST;
727 request->rq_req_cbid.cbid_fn = request_out_callback;
728 request->rq_req_cbid.cbid_arg = request;
730 request->rq_reply_cbid.cbid_fn = reply_in_callback;
731 request->rq_reply_cbid.cbid_arg = request;
733 request->rq_reply_deadline = 0;
734 request->rq_phase = RQ_PHASE_NEW;
735 request->rq_next_phase = RQ_PHASE_UNDEFINED;
737 request->rq_request_portal = imp->imp_client->cli_request_portal;
738 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
740 ptlrpc_at_set_req_timeout(request);
742 lustre_msg_set_opc(request->rq_reqmsg, opcode);
743 ptlrpc_assign_next_xid(request);
748 LASSERT(!request->rq_pool);
749 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
751 class_import_put(imp);
756 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
759 * Pack request buffers for network transfer, performing necessary encryption
760 * steps if necessary.
762 int ptlrpc_request_pack(struct ptlrpc_request *request,
763 __u32 version, int opcode)
766 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
770 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
771 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
772 * have to send old ptlrpc_body to keep interoprability with these
775 * Only three kinds of server->client RPCs so far:
780 * XXX This should be removed whenever we drop the interoprability with
781 * the these old clients.
783 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
784 opcode == LDLM_GL_CALLBACK)
785 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
786 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
790 EXPORT_SYMBOL(ptlrpc_request_pack);
793 * Helper function to allocate new request on import \a imp
794 * and possibly using existing request from pool \a pool if provided.
795 * Returns allocated request structure with import field filled or
799 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
800 struct ptlrpc_request_pool *pool)
802 struct ptlrpc_request *request = NULL;
804 request = ptlrpc_request_cache_alloc(GFP_NOFS);
806 if (!request && pool)
807 request = ptlrpc_prep_req_from_pool(pool);
810 ptlrpc_cli_req_init(request);
812 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
813 LASSERT(imp != LP_POISON);
814 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
816 LASSERT(imp->imp_client != LP_POISON);
818 request->rq_import = class_import_get(imp);
820 CERROR("request allocation out of memory\n");
827 * Helper function for creating a request.
828 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
829 * buffer structures according to capsule template \a format.
830 * Returns allocated request structure pointer or NULL on error.
832 static struct ptlrpc_request *
833 ptlrpc_request_alloc_internal(struct obd_import *imp,
834 struct ptlrpc_request_pool * pool,
835 const struct req_format *format)
837 struct ptlrpc_request *request;
839 request = __ptlrpc_request_alloc(imp, pool);
843 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
844 req_capsule_set(&request->rq_pill, format);
849 * Allocate new request structure for import \a imp and initialize its
850 * buffer structure according to capsule template \a format.
852 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
853 const struct req_format *format)
855 return ptlrpc_request_alloc_internal(imp, NULL, format);
857 EXPORT_SYMBOL(ptlrpc_request_alloc);
860 * Allocate new request structure for import \a imp from pool \a pool and
861 * initialize its buffer structure according to capsule template \a format.
863 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
864 struct ptlrpc_request_pool * pool,
865 const struct req_format *format)
867 return ptlrpc_request_alloc_internal(imp, pool, format);
869 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
872 * For requests not from pool, free memory of the request structure.
873 * For requests obtained from a pool earlier, return request back to pool.
875 void ptlrpc_request_free(struct ptlrpc_request *request)
877 if (request->rq_pool)
878 __ptlrpc_free_req_to_pool(request);
880 ptlrpc_request_cache_free(request);
882 EXPORT_SYMBOL(ptlrpc_request_free);
885 * Allocate new request for operatione \a opcode and immediatelly pack it for
887 * Only used for simple requests like OBD_PING where the only important
888 * part of the request is operation itself.
889 * Returns allocated request or NULL on error.
891 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
892 const struct req_format *format,
893 __u32 version, int opcode)
895 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
899 rc = ptlrpc_request_pack(req, version, opcode);
901 ptlrpc_request_free(req);
907 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
910 * Allocate and initialize new request set structure on the current CPT.
911 * Returns a pointer to the newly allocated set structure or NULL on error.
913 struct ptlrpc_request_set *ptlrpc_prep_set(void)
915 struct ptlrpc_request_set *set;
919 cpt = cfs_cpt_current(cfs_cpt_table, 0);
920 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
923 atomic_set(&set->set_refcount, 1);
924 INIT_LIST_HEAD(&set->set_requests);
925 init_waitqueue_head(&set->set_waitq);
926 atomic_set(&set->set_new_count, 0);
927 atomic_set(&set->set_remaining, 0);
928 spin_lock_init(&set->set_new_req_lock);
929 INIT_LIST_HEAD(&set->set_new_requests);
930 INIT_LIST_HEAD(&set->set_cblist);
931 set->set_max_inflight = UINT_MAX;
932 set->set_producer = NULL;
933 set->set_producer_arg = NULL;
938 EXPORT_SYMBOL(ptlrpc_prep_set);
941 * Allocate and initialize new request set structure with flow control
942 * extension. This extension allows to control the number of requests in-flight
943 * for the whole set. A callback function to generate requests must be provided
944 * and the request set will keep the number of requests sent over the wire to
946 * Returns a pointer to the newly allocated set structure or NULL on error.
948 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
952 struct ptlrpc_request_set *set;
954 set = ptlrpc_prep_set();
958 set->set_max_inflight = max;
959 set->set_producer = func;
960 set->set_producer_arg = arg;
966 * Wind down and free request set structure previously allocated with
968 * Ensures that all requests on the set have completed and removes
969 * all requests from the request list in a set.
970 * If any unsent request happen to be on the list, pretends that they got
971 * an error in flight and calls their completion handler.
973 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
975 struct list_head *tmp;
976 struct list_head *next;
981 /* Requests on the set should either all be completed, or all be new */
982 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
983 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
984 list_for_each(tmp, &set->set_requests) {
985 struct ptlrpc_request *req =
986 list_entry(tmp, struct ptlrpc_request,
989 LASSERT(req->rq_phase == expected_phase);
993 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
994 atomic_read(&set->set_remaining) == n, "%d / %d\n",
995 atomic_read(&set->set_remaining), n);
997 list_for_each_safe(tmp, next, &set->set_requests) {
998 struct ptlrpc_request *req =
999 list_entry(tmp, struct ptlrpc_request,
1001 list_del_init(&req->rq_set_chain);
1003 LASSERT(req->rq_phase == expected_phase);
1005 if (req->rq_phase == RQ_PHASE_NEW) {
1006 ptlrpc_req_interpret(NULL, req, -EBADR);
1007 atomic_dec(&set->set_remaining);
1010 spin_lock(&req->rq_lock);
1012 req->rq_invalid_rqset = 0;
1013 spin_unlock(&req->rq_lock);
1015 ptlrpc_req_finished (req);
1018 LASSERT(atomic_read(&set->set_remaining) == 0);
1020 ptlrpc_reqset_put(set);
1023 EXPORT_SYMBOL(ptlrpc_set_destroy);
1026 * Add a callback function \a fn to the set.
1027 * This function would be called when all requests on this set are completed.
1028 * The function will be passed \a data argument.
1030 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1031 set_interpreter_func fn, void *data)
1033 struct ptlrpc_set_cbdata *cbdata;
1035 OBD_ALLOC_PTR(cbdata);
1039 cbdata->psc_interpret = fn;
1040 cbdata->psc_data = data;
1041 list_add_tail(&cbdata->psc_item, &set->set_cblist);
1047 * Add a new request to the general purpose request set.
1048 * Assumes request reference from the caller.
1050 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1051 struct ptlrpc_request *req)
1053 LASSERT(list_empty(&req->rq_set_chain));
1055 if (req->rq_allow_intr)
1056 set->set_allow_intr = 1;
1058 /* The set takes over the caller's request reference */
1059 list_add_tail(&req->rq_set_chain, &set->set_requests);
1061 atomic_inc(&set->set_remaining);
1062 req->rq_queued_time = cfs_time_current();
1064 if (req->rq_reqmsg != NULL)
1065 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1067 if (set->set_producer != NULL)
1068 /* If the request set has a producer callback, the RPC must be
1069 * sent straight away */
1070 ptlrpc_send_new_req(req);
1072 EXPORT_SYMBOL(ptlrpc_set_add_req);
1075 * Add a request to a request with dedicated server thread
1076 * and wake the thread to make any necessary processing.
1077 * Currently only used for ptlrpcd.
1079 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1080 struct ptlrpc_request *req)
1082 struct ptlrpc_request_set *set = pc->pc_set;
1085 LASSERT(req->rq_set == NULL);
1086 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1088 spin_lock(&set->set_new_req_lock);
1090 * The set takes over the caller's request reference.
1093 req->rq_queued_time = cfs_time_current();
1094 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1095 count = atomic_inc_return(&set->set_new_count);
1096 spin_unlock(&set->set_new_req_lock);
1098 /* Only need to call wakeup once for the first entry. */
1100 wake_up(&set->set_waitq);
1102 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1103 * guarantee the async RPC can be processed ASAP, we have
1104 * no other better choice. It maybe fixed in future. */
1105 for (i = 0; i < pc->pc_npartners; i++)
1106 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1111 * Based on the current state of the import, determine if the request
1112 * can be sent, is an error, or should be delayed.
1114 * Returns true if this request should be delayed. If false, and
1115 * *status is set, then the request can not be sent and *status is the
1116 * error code. If false and status is 0, then request can be sent.
1118 * The imp->imp_lock must be held.
1120 static int ptlrpc_import_delay_req(struct obd_import *imp,
1121 struct ptlrpc_request *req, int *status)
1126 LASSERT (status != NULL);
1129 if (req->rq_ctx_init || req->rq_ctx_fini) {
1130 /* always allow ctx init/fini rpc go through */
1131 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1132 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1134 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1135 /* pings may safely race with umount */
1136 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1137 D_HA : D_ERROR, req, "IMP_CLOSED ");
1139 } else if (ptlrpc_send_limit_expired(req)) {
1140 /* probably doesn't need to be a D_ERROR after initial testing*/
1141 DEBUG_REQ(D_HA, req, "send limit expired ");
1142 *status = -ETIMEDOUT;
1143 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1144 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1145 /* allow CONNECT even if import is invalid */ ;
1146 if (atomic_read(&imp->imp_inval_count) != 0) {
1147 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1150 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1151 if (!imp->imp_deactive)
1152 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1153 *status = -ESHUTDOWN; /* bz 12940 */
1154 } else if (req->rq_import_generation != imp->imp_generation) {
1155 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1157 } else if (req->rq_send_state != imp->imp_state) {
1158 /* invalidate in progress - any requests should be drop */
1159 if (atomic_read(&imp->imp_inval_count) != 0) {
1160 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1162 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1163 *status = -EWOULDBLOCK;
1164 } else if (req->rq_allow_replay &&
1165 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1166 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1167 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1168 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1169 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1179 * Decide if the error message should be printed to the console or not.
1180 * Makes its decision based on request type, status, and failure frequency.
1182 * \param[in] req request that failed and may need a console message
1184 * \retval false if no message should be printed
1185 * \retval true if console message should be printed
1187 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1191 LASSERT(req->rq_reqmsg != NULL);
1192 opc = lustre_msg_get_opc(req->rq_reqmsg);
1194 /* Suppress particular reconnect errors which are to be expected. */
1195 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1198 /* Suppress timed out reconnect requests */
1199 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1203 /* Suppress most unavailable/again reconnect requests, but
1204 * print occasionally so it is clear client is trying to
1205 * connect to a server where no target is running. */
1206 err = lustre_msg_get_status(req->rq_repmsg);
1207 if ((err == -ENODEV || err == -EAGAIN) &&
1208 req->rq_import->imp_conn_cnt % 30 != 20)
1216 * Check request processing status.
1217 * Returns the status.
1219 static int ptlrpc_check_status(struct ptlrpc_request *req)
1224 err = lustre_msg_get_status(req->rq_repmsg);
1225 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1226 struct obd_import *imp = req->rq_import;
1227 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1228 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1230 if (ptlrpc_console_allow(req))
1231 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1232 "failed: rc = %d\n",
1233 imp->imp_obd->obd_name,
1235 libcfs_nid2str(nid), err);
1236 RETURN(err < 0 ? err : -EINVAL);
1240 DEBUG_REQ(D_INFO, req, "status is %d", err);
1241 } else if (err > 0) {
1242 /* XXX: translate this error from net to host */
1243 DEBUG_REQ(D_INFO, req, "status is %d", err);
1250 * save pre-versions of objects into request for replay.
1251 * Versions are obtained from server reply.
1254 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1256 struct lustre_msg *repmsg = req->rq_repmsg;
1257 struct lustre_msg *reqmsg = req->rq_reqmsg;
1258 __u64 *versions = lustre_msg_get_versions(repmsg);
1261 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1265 lustre_msg_set_versions(reqmsg, versions);
1266 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1267 versions[0], versions[1]);
1272 __u64 ptlrpc_known_replied_xid(struct obd_import *imp)
1274 struct ptlrpc_request *req;
1276 assert_spin_locked(&imp->imp_lock);
1277 if (list_empty(&imp->imp_unreplied_list))
1280 req = list_entry(imp->imp_unreplied_list.next, struct ptlrpc_request,
1282 LASSERTF(req->rq_xid >= 1, "XID:"LPU64"\n", req->rq_xid);
1284 if (imp->imp_known_replied_xid < req->rq_xid - 1)
1285 imp->imp_known_replied_xid = req->rq_xid - 1;
1287 return req->rq_xid - 1;
1291 * Callback function called when client receives RPC reply for \a req.
1292 * Returns 0 on success or error code.
1293 * The return alue would be assigned to req->rq_status by the caller
1294 * as request processing status.
1295 * This function also decides if the request needs to be saved for later replay.
1297 static int after_reply(struct ptlrpc_request *req)
1299 struct obd_import *imp = req->rq_import;
1300 struct obd_device *obd = req->rq_import->imp_obd;
1302 struct timeval work_start;
1307 LASSERT(obd != NULL);
1308 /* repbuf must be unlinked */
1309 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1311 if (req->rq_reply_truncated) {
1312 if (ptlrpc_no_resend(req)) {
1313 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1314 " expected: %d, actual size: %d",
1315 req->rq_nob_received, req->rq_repbuf_len);
1319 sptlrpc_cli_free_repbuf(req);
1320 /* Pass the required reply buffer size (include
1321 * space for early reply).
1322 * NB: no need to roundup because alloc_repbuf
1323 * will roundup it */
1324 req->rq_replen = req->rq_nob_received;
1325 req->rq_nob_received = 0;
1326 spin_lock(&req->rq_lock);
1328 spin_unlock(&req->rq_lock);
1332 do_gettimeofday(&work_start);
1333 timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1336 * NB Until this point, the whole of the incoming message,
1337 * including buflens, status etc is in the sender's byte order.
1339 rc = sptlrpc_cli_unwrap_reply(req);
1341 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1346 * Security layer unwrap might ask resend this request.
1351 rc = unpack_reply(req);
1355 /* retry indefinitely on EINPROGRESS */
1356 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1357 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1358 time_t now = cfs_time_current_sec();
1360 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1361 spin_lock(&req->rq_lock);
1363 spin_unlock(&req->rq_lock);
1364 req->rq_nr_resend++;
1366 /* Readjust the timeout for current conditions */
1367 ptlrpc_at_set_req_timeout(req);
1368 /* delay resend to give a chance to the server to get ready.
1369 * The delay is increased by 1s on every resend and is capped to
1370 * the current request timeout (i.e. obd_timeout if AT is off,
1371 * or AT service time x 125% + 5s, see at_est2timeout) */
1372 if (req->rq_nr_resend > req->rq_timeout)
1373 req->rq_sent = now + req->rq_timeout;
1375 req->rq_sent = now + req->rq_nr_resend;
1377 /* Resend for EINPROGRESS will use a new XID */
1378 spin_lock(&imp->imp_lock);
1379 list_del_init(&req->rq_unreplied_list);
1380 spin_unlock(&imp->imp_lock);
1385 if (obd->obd_svc_stats != NULL) {
1386 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1388 ptlrpc_lprocfs_rpc_sent(req, timediff);
1391 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1392 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1393 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1394 lustre_msg_get_type(req->rq_repmsg));
1398 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1399 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1400 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1401 ptlrpc_at_adj_net_latency(req,
1402 lustre_msg_get_service_time(req->rq_repmsg));
1404 rc = ptlrpc_check_status(req);
1405 imp->imp_connect_error = rc;
1409 * Either we've been evicted, or the server has failed for
1410 * some reason. Try to reconnect, and if that fails, punt to
1413 if (ptlrpc_recoverable_error(rc)) {
1414 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1415 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1418 ptlrpc_request_handle_notconn(req);
1423 * Let's look if server sent slv. Do it only for RPC with
1426 ldlm_cli_update_pool(req);
1430 * Store transno in reqmsg for replay.
1432 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1433 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1434 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1437 if (imp->imp_replayable) {
1438 spin_lock(&imp->imp_lock);
1440 * No point in adding already-committed requests to the replay
1441 * list, we will just remove them immediately. b=9829
1443 if (req->rq_transno != 0 &&
1445 lustre_msg_get_last_committed(req->rq_repmsg) ||
1447 /** version recovery */
1448 ptlrpc_save_versions(req);
1449 ptlrpc_retain_replayable_request(req, imp);
1450 } else if (req->rq_commit_cb != NULL &&
1451 list_empty(&req->rq_replay_list)) {
1452 /* NB: don't call rq_commit_cb if it's already on
1453 * rq_replay_list, ptlrpc_free_committed() will call
1454 * it later, see LU-3618 for details */
1455 spin_unlock(&imp->imp_lock);
1456 req->rq_commit_cb(req);
1457 spin_lock(&imp->imp_lock);
1461 * Replay-enabled imports return commit-status information.
1463 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1464 if (likely(committed > imp->imp_peer_committed_transno))
1465 imp->imp_peer_committed_transno = committed;
1467 ptlrpc_free_committed(imp);
1469 if (!list_empty(&imp->imp_replay_list)) {
1470 struct ptlrpc_request *last;
1472 last = list_entry(imp->imp_replay_list.prev,
1473 struct ptlrpc_request,
1476 * Requests with rq_replay stay on the list even if no
1477 * commit is expected.
1479 if (last->rq_transno > imp->imp_peer_committed_transno)
1480 ptlrpc_pinger_commit_expected(imp);
1483 spin_unlock(&imp->imp_lock);
1490 * Helper function to send request \a req over the network for the first time
1491 * Also adjusts request phase.
1492 * Returns 0 on success or error code.
1494 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1496 struct obd_import *imp = req->rq_import;
1501 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1503 /* do not try to go further if there is not enough memory in enc_pool */
1504 if (req->rq_sent && req->rq_bulk != NULL)
1505 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1506 pool_is_at_full_capacity())
1509 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1510 (!req->rq_generation_set ||
1511 req->rq_import_generation == imp->imp_generation))
1514 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1516 spin_lock(&imp->imp_lock);
1518 LASSERT(req->rq_xid != 0);
1519 LASSERT(!list_empty(&req->rq_unreplied_list));
1521 if (!req->rq_generation_set)
1522 req->rq_import_generation = imp->imp_generation;
1524 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1525 spin_lock(&req->rq_lock);
1526 req->rq_waiting = 1;
1527 spin_unlock(&req->rq_lock);
1529 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1530 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1531 ptlrpc_import_state_name(req->rq_send_state),
1532 ptlrpc_import_state_name(imp->imp_state));
1533 LASSERT(list_empty(&req->rq_list));
1534 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1535 atomic_inc(&req->rq_import->imp_inflight);
1536 spin_unlock(&imp->imp_lock);
1541 spin_unlock(&imp->imp_lock);
1542 req->rq_status = rc;
1543 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1547 LASSERT(list_empty(&req->rq_list));
1548 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1549 atomic_inc(&req->rq_import->imp_inflight);
1551 /* find the known replied XID from the unreplied list, CONNECT
1552 * and DISCONNECT requests are skipped to make the sanity check
1553 * on server side happy. see process_req_last_xid().
1555 * For CONNECT: Because replay requests have lower XID, it'll
1556 * break the sanity check if CONNECT bump the exp_last_xid on
1559 * For DISCONNECT: Since client will abort inflight RPC before
1560 * sending DISCONNECT, DISCONNECT may carry an XID which higher
1561 * than the inflight RPC.
1563 if (!ptlrpc_req_is_connect(req) && !ptlrpc_req_is_disconnect(req))
1564 min_xid = ptlrpc_known_replied_xid(imp);
1565 spin_unlock(&imp->imp_lock);
1567 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid);
1569 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1571 rc = sptlrpc_req_refresh_ctx(req, -1);
1574 req->rq_status = rc;
1577 spin_lock(&req->rq_lock);
1578 req->rq_wait_ctx = 1;
1579 spin_unlock(&req->rq_lock);
1584 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1585 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1586 imp->imp_obd->obd_uuid.uuid,
1587 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1588 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1589 lustre_msg_get_opc(req->rq_reqmsg));
1591 rc = ptl_send_rpc(req, 0);
1592 if (rc == -ENOMEM) {
1593 spin_lock(&imp->imp_lock);
1594 if (!list_empty(&req->rq_list)) {
1595 list_del_init(&req->rq_list);
1596 atomic_dec(&req->rq_import->imp_inflight);
1598 spin_unlock(&imp->imp_lock);
1599 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1603 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1604 spin_lock(&req->rq_lock);
1605 req->rq_net_err = 1;
1606 spin_unlock(&req->rq_lock);
1612 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1617 LASSERT(set->set_producer != NULL);
1619 remaining = atomic_read(&set->set_remaining);
1621 /* populate the ->set_requests list with requests until we
1622 * reach the maximum number of RPCs in flight for this set */
1623 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1624 rc = set->set_producer(set, set->set_producer_arg);
1625 if (rc == -ENOENT) {
1626 /* no more RPC to produce */
1627 set->set_producer = NULL;
1628 set->set_producer_arg = NULL;
1633 RETURN((atomic_read(&set->set_remaining) - remaining));
1637 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1638 * and no more replies are expected.
1639 * (it is possible to get less replies than requests sent e.g. due to timed out
1640 * requests or requests that we had trouble to send out)
1642 * NOTE: This function contains a potential schedule point (cond_resched()).
1644 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1646 struct list_head *tmp, *next;
1647 struct list_head comp_reqs;
1648 int force_timer_recalc = 0;
1651 if (atomic_read(&set->set_remaining) == 0)
1654 INIT_LIST_HEAD(&comp_reqs);
1655 list_for_each_safe(tmp, next, &set->set_requests) {
1656 struct ptlrpc_request *req =
1657 list_entry(tmp, struct ptlrpc_request,
1659 struct obd_import *imp = req->rq_import;
1660 int unregistered = 0;
1664 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1665 list_move_tail(&req->rq_set_chain, &comp_reqs);
1669 /* This schedule point is mainly for the ptlrpcd caller of this
1670 * function. Most ptlrpc sets are not long-lived and unbounded
1671 * in length, but at the least the set used by the ptlrpcd is.
1672 * Since the processing time is unbounded, we need to insert an
1673 * explicit schedule point to make the thread well-behaved.
1677 /* If the caller requires to allow to be interpreted by force
1678 * and it has really been interpreted, then move the request
1679 * to RQ_PHASE_INTERPRET phase in spite of what the current
1681 if (unlikely(req->rq_allow_intr && req->rq_intr)) {
1682 req->rq_status = -EINTR;
1683 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1685 /* Since it is interpreted and we have to wait for
1686 * the reply to be unlinked, then use sync mode. */
1689 GOTO(interpret, req->rq_status);
1692 if (req->rq_phase == RQ_PHASE_NEW && ptlrpc_send_new_req(req))
1693 force_timer_recalc = 1;
1695 /* delayed send - skip */
1696 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1699 /* delayed resend - skip */
1700 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1701 req->rq_sent > cfs_time_current_sec())
1704 if (!(req->rq_phase == RQ_PHASE_RPC ||
1705 req->rq_phase == RQ_PHASE_BULK ||
1706 req->rq_phase == RQ_PHASE_INTERPRET ||
1707 req->rq_phase == RQ_PHASE_UNREGISTERING)) {
1708 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1712 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1713 LASSERT(req->rq_next_phase != req->rq_phase);
1714 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1717 * Skip processing until reply is unlinked. We
1718 * can't return to pool before that and we can't
1719 * call interpret before that. We need to make
1720 * sure that all rdma transfers finished and will
1721 * not corrupt any data.
1723 if (ptlrpc_client_recv_or_unlink(req) ||
1724 ptlrpc_client_bulk_active(req))
1728 * Turn fail_loc off to prevent it from looping
1731 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1732 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1735 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1736 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1741 * Move to next phase if reply was successfully
1744 ptlrpc_rqphase_move(req, req->rq_next_phase);
1747 if (req->rq_phase == RQ_PHASE_INTERPRET)
1748 GOTO(interpret, req->rq_status);
1751 * Note that this also will start async reply unlink.
1753 if (req->rq_net_err && !req->rq_timedout) {
1754 ptlrpc_expire_one_request(req, 1);
1757 * Check if we still need to wait for unlink.
1759 if (ptlrpc_client_recv_or_unlink(req) ||
1760 ptlrpc_client_bulk_active(req))
1762 /* If there is no need to resend, fail it now. */
1763 if (req->rq_no_resend) {
1764 if (req->rq_status == 0)
1765 req->rq_status = -EIO;
1766 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1767 GOTO(interpret, req->rq_status);
1774 spin_lock(&req->rq_lock);
1775 req->rq_replied = 0;
1776 spin_unlock(&req->rq_lock);
1777 if (req->rq_status == 0)
1778 req->rq_status = -EIO;
1779 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1780 GOTO(interpret, req->rq_status);
1783 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1784 * so it sets rq_intr regardless of individual rpc
1785 * timeouts. The synchronous IO waiting path sets
1786 * rq_intr irrespective of whether ptlrpcd
1787 * has seen a timeout. Our policy is to only interpret
1788 * interrupted rpcs after they have timed out, so we
1789 * need to enforce that here.
1792 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1793 req->rq_wait_ctx)) {
1794 req->rq_status = -EINTR;
1795 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1796 GOTO(interpret, req->rq_status);
1799 if (req->rq_phase == RQ_PHASE_RPC) {
1800 if (req->rq_timedout || req->rq_resend ||
1801 req->rq_waiting || req->rq_wait_ctx) {
1804 if (!ptlrpc_unregister_reply(req, 1)) {
1805 ptlrpc_unregister_bulk(req, 1);
1809 spin_lock(&imp->imp_lock);
1810 if (ptlrpc_import_delay_req(imp, req, &status)){
1811 /* put on delay list - only if we wait
1812 * recovery finished - before send */
1813 list_del_init(&req->rq_list);
1814 list_add_tail(&req->rq_list,
1817 spin_unlock(&imp->imp_lock);
1822 req->rq_status = status;
1823 ptlrpc_rqphase_move(req,
1824 RQ_PHASE_INTERPRET);
1825 spin_unlock(&imp->imp_lock);
1826 GOTO(interpret, req->rq_status);
1828 if (ptlrpc_no_resend(req) &&
1829 !req->rq_wait_ctx) {
1830 req->rq_status = -ENOTCONN;
1831 ptlrpc_rqphase_move(req,
1832 RQ_PHASE_INTERPRET);
1833 spin_unlock(&imp->imp_lock);
1834 GOTO(interpret, req->rq_status);
1837 list_del_init(&req->rq_list);
1838 list_add_tail(&req->rq_list,
1839 &imp->imp_sending_list);
1841 spin_unlock(&imp->imp_lock);
1843 spin_lock(&req->rq_lock);
1844 req->rq_waiting = 0;
1845 spin_unlock(&req->rq_lock);
1847 if (req->rq_timedout || req->rq_resend) {
1848 /* This is re-sending anyways,
1849 * let's mark req as resend. */
1850 spin_lock(&req->rq_lock);
1852 spin_unlock(&req->rq_lock);
1854 if (req->rq_bulk != NULL &&
1855 !ptlrpc_unregister_bulk(req, 1))
1859 * rq_wait_ctx is only touched by ptlrpcd,
1860 * so no lock is needed here.
1862 status = sptlrpc_req_refresh_ctx(req, -1);
1865 req->rq_status = status;
1866 spin_lock(&req->rq_lock);
1867 req->rq_wait_ctx = 0;
1868 spin_unlock(&req->rq_lock);
1869 force_timer_recalc = 1;
1871 spin_lock(&req->rq_lock);
1872 req->rq_wait_ctx = 1;
1873 spin_unlock(&req->rq_lock);
1878 spin_lock(&req->rq_lock);
1879 req->rq_wait_ctx = 0;
1880 spin_unlock(&req->rq_lock);
1883 rc = ptl_send_rpc(req, 0);
1884 if (rc == -ENOMEM) {
1885 spin_lock(&imp->imp_lock);
1886 if (!list_empty(&req->rq_list))
1887 list_del_init(&req->rq_list);
1888 spin_unlock(&imp->imp_lock);
1889 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1893 DEBUG_REQ(D_HA, req,
1894 "send failed: rc = %d", rc);
1895 force_timer_recalc = 1;
1896 spin_lock(&req->rq_lock);
1897 req->rq_net_err = 1;
1898 spin_unlock(&req->rq_lock);
1901 /* need to reset the timeout */
1902 force_timer_recalc = 1;
1905 spin_lock(&req->rq_lock);
1907 if (ptlrpc_client_early(req)) {
1908 ptlrpc_at_recv_early_reply(req);
1909 spin_unlock(&req->rq_lock);
1913 /* Still waiting for a reply? */
1914 if (ptlrpc_client_recv(req)) {
1915 spin_unlock(&req->rq_lock);
1919 /* Did we actually receive a reply? */
1920 if (!ptlrpc_client_replied(req)) {
1921 spin_unlock(&req->rq_lock);
1925 spin_unlock(&req->rq_lock);
1927 /* unlink from net because we are going to
1928 * swab in-place of reply buffer */
1929 unregistered = ptlrpc_unregister_reply(req, 1);
1933 req->rq_status = after_reply(req);
1937 /* If there is no bulk associated with this request,
1938 * then we're done and should let the interpreter
1939 * process the reply. Similarly if the RPC returned
1940 * an error, and therefore the bulk will never arrive.
1942 if (req->rq_bulk == NULL || req->rq_status < 0) {
1943 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1944 GOTO(interpret, req->rq_status);
1947 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1950 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1951 if (ptlrpc_client_bulk_active(req))
1954 if (req->rq_bulk->bd_failure) {
1955 /* The RPC reply arrived OK, but the bulk screwed
1956 * up! Dead weird since the server told us the RPC
1957 * was good after getting the REPLY for her GET or
1958 * the ACK for her PUT. */
1959 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1960 req->rq_status = -EIO;
1963 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1966 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1968 /* This moves to "unregistering" phase we need to wait for
1970 if (!unregistered && !ptlrpc_unregister_reply(req, async)) {
1971 /* start async bulk unlink too */
1972 ptlrpc_unregister_bulk(req, 1);
1976 if (!ptlrpc_unregister_bulk(req, async))
1979 /* When calling interpret receiving already should be
1981 LASSERT(!req->rq_receiving_reply);
1983 ptlrpc_req_interpret(env, req, req->rq_status);
1985 if (ptlrpcd_check_work(req)) {
1986 atomic_dec(&set->set_remaining);
1989 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1991 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1992 "Completed RPC pname:cluuid:pid:xid:nid:"
1993 "opc %s:%s:%d:"LPU64":%s:%d\n",
1994 current_comm(), imp->imp_obd->obd_uuid.uuid,
1995 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1996 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1997 lustre_msg_get_opc(req->rq_reqmsg));
1999 spin_lock(&imp->imp_lock);
2000 /* Request already may be not on sending or delaying list. This
2001 * may happen in the case of marking it erroneous for the case
2002 * ptlrpc_import_delay_req(req, status) find it impossible to
2003 * allow sending this rpc and returns *status != 0. */
2004 if (!list_empty(&req->rq_list)) {
2005 list_del_init(&req->rq_list);
2006 atomic_dec(&imp->imp_inflight);
2008 list_del_init(&req->rq_unreplied_list);
2009 spin_unlock(&imp->imp_lock);
2011 atomic_dec(&set->set_remaining);
2012 wake_up_all(&imp->imp_recovery_waitq);
2014 if (set->set_producer) {
2015 /* produce a new request if possible */
2016 if (ptlrpc_set_producer(set) > 0)
2017 force_timer_recalc = 1;
2019 /* free the request that has just been completed
2020 * in order not to pollute set->set_requests */
2021 list_del_init(&req->rq_set_chain);
2022 spin_lock(&req->rq_lock);
2024 req->rq_invalid_rqset = 0;
2025 spin_unlock(&req->rq_lock);
2027 /* record rq_status to compute the final status later */
2028 if (req->rq_status != 0)
2029 set->set_rc = req->rq_status;
2030 ptlrpc_req_finished(req);
2032 list_move_tail(&req->rq_set_chain, &comp_reqs);
2036 /* move completed request at the head of list so it's easier for
2037 * caller to find them */
2038 list_splice(&comp_reqs, &set->set_requests);
2040 /* If we hit an error, we want to recover promptly. */
2041 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
2043 EXPORT_SYMBOL(ptlrpc_check_set);
2046 * Time out request \a req. is \a async_unlink is set, that means do not wait
2047 * until LNet actually confirms network buffer unlinking.
2048 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2050 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2052 struct obd_import *imp = req->rq_import;
2056 spin_lock(&req->rq_lock);
2057 req->rq_timedout = 1;
2058 spin_unlock(&req->rq_lock);
2060 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
2061 "/real "CFS_DURATION_T"]",
2062 req->rq_net_err ? "failed due to network error" :
2063 ((req->rq_real_sent == 0 ||
2064 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
2065 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
2066 "timed out for sent delay" : "timed out for slow reply"),
2067 req->rq_sent, req->rq_real_sent);
2069 if (imp != NULL && obd_debug_peer_on_timeout)
2070 LNetDebugPeer(imp->imp_connection->c_peer);
2072 ptlrpc_unregister_reply(req, async_unlink);
2073 ptlrpc_unregister_bulk(req, async_unlink);
2075 if (obd_dump_on_timeout)
2076 libcfs_debug_dumplog();
2079 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2083 atomic_inc(&imp->imp_timeouts);
2085 /* The DLM server doesn't want recovery run on its imports. */
2086 if (imp->imp_dlm_fake)
2089 /* If this request is for recovery or other primordial tasks,
2090 * then error it out here. */
2091 if (req->rq_ctx_init || req->rq_ctx_fini ||
2092 req->rq_send_state != LUSTRE_IMP_FULL ||
2093 imp->imp_obd->obd_no_recov) {
2094 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2095 ptlrpc_import_state_name(req->rq_send_state),
2096 ptlrpc_import_state_name(imp->imp_state));
2097 spin_lock(&req->rq_lock);
2098 req->rq_status = -ETIMEDOUT;
2100 spin_unlock(&req->rq_lock);
2104 /* if a request can't be resent we can't wait for an answer after
2106 if (ptlrpc_no_resend(req)) {
2107 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2111 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2117 * Time out all uncompleted requests in request set pointed by \a data
2118 * Callback used when waiting on sets with l_wait_event.
2121 int ptlrpc_expired_set(void *data)
2123 struct ptlrpc_request_set *set = data;
2124 struct list_head *tmp;
2125 time_t now = cfs_time_current_sec();
2128 LASSERT(set != NULL);
2131 * A timeout expired. See which reqs it applies to...
2133 list_for_each(tmp, &set->set_requests) {
2134 struct ptlrpc_request *req =
2135 list_entry(tmp, struct ptlrpc_request,
2138 /* don't expire request waiting for context */
2139 if (req->rq_wait_ctx)
2142 /* Request in-flight? */
2143 if (!((req->rq_phase == RQ_PHASE_RPC &&
2144 !req->rq_waiting && !req->rq_resend) ||
2145 (req->rq_phase == RQ_PHASE_BULK)))
2148 if (req->rq_timedout || /* already dealt with */
2149 req->rq_deadline > now) /* not expired */
2152 /* Deal with this guy. Do it asynchronously to not block
2153 * ptlrpcd thread. */
2154 ptlrpc_expire_one_request(req, 1);
2158 * When waiting for a whole set, we always break out of the
2159 * sleep so we can recalculate the timeout, or enable interrupts
2160 * if everyone's timed out.
2166 * Sets rq_intr flag in \a req under spinlock.
2168 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2170 spin_lock(&req->rq_lock);
2172 spin_unlock(&req->rq_lock);
2174 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2177 * Interrupts (sets interrupted flag) all uncompleted requests in
2178 * a set \a data. Callback for l_wait_event for interruptible waits.
2180 static void ptlrpc_interrupted_set(void *data)
2182 struct ptlrpc_request_set *set = data;
2183 struct list_head *tmp;
2185 LASSERT(set != NULL);
2186 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2188 list_for_each(tmp, &set->set_requests) {
2189 struct ptlrpc_request *req =
2190 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2195 if (req->rq_phase != RQ_PHASE_RPC &&
2196 req->rq_phase != RQ_PHASE_UNREGISTERING &&
2197 !req->rq_allow_intr)
2200 ptlrpc_mark_interrupted(req);
2205 * Get the smallest timeout in the set; this does NOT set a timeout.
2207 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2209 struct list_head *tmp;
2210 time_t now = cfs_time_current_sec();
2212 struct ptlrpc_request *req;
2216 list_for_each(tmp, &set->set_requests) {
2217 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2220 * Request in-flight?
2222 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2223 (req->rq_phase == RQ_PHASE_BULK) ||
2224 (req->rq_phase == RQ_PHASE_NEW)))
2228 * Already timed out.
2230 if (req->rq_timedout)
2236 if (req->rq_wait_ctx)
2239 if (req->rq_phase == RQ_PHASE_NEW)
2240 deadline = req->rq_sent;
2241 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2242 deadline = req->rq_sent;
2244 deadline = req->rq_sent + req->rq_timeout;
2246 if (deadline <= now) /* actually expired already */
2247 timeout = 1; /* ASAP */
2248 else if (timeout == 0 || timeout > deadline - now)
2249 timeout = deadline - now;
2255 * Send all unset request from the set and then wait untill all
2256 * requests in the set complete (either get a reply, timeout, get an
2257 * error or otherwise be interrupted).
2258 * Returns 0 on success or error code otherwise.
2260 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2262 struct list_head *tmp;
2263 struct ptlrpc_request *req;
2264 struct l_wait_info lwi;
2268 if (set->set_producer)
2269 (void)ptlrpc_set_producer(set);
2271 list_for_each(tmp, &set->set_requests) {
2272 req = list_entry(tmp, struct ptlrpc_request,
2274 if (req->rq_phase == RQ_PHASE_NEW)
2275 (void)ptlrpc_send_new_req(req);
2278 if (list_empty(&set->set_requests))
2282 timeout = ptlrpc_set_next_timeout(set);
2284 /* wait until all complete, interrupted, or an in-flight
2286 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2289 if ((timeout == 0 && !signal_pending(current)) ||
2290 set->set_allow_intr)
2291 /* No requests are in-flight (ether timed out
2292 * or delayed), so we can allow interrupts.
2293 * We still want to block for a limited time,
2294 * so we allow interrupts during the timeout. */
2295 lwi = LWI_TIMEOUT_INTR_ALL(
2296 cfs_time_seconds(timeout ? timeout : 1),
2298 ptlrpc_interrupted_set, set);
2301 * At least one request is in flight, so no
2302 * interrupts are allowed. Wait until all
2303 * complete, or an in-flight req times out.
2305 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2306 ptlrpc_expired_set, set);
2308 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2310 /* LU-769 - if we ignored the signal because it was already
2311 * pending when we started, we need to handle it now or we risk
2312 * it being ignored forever */
2313 if (rc == -ETIMEDOUT &&
2314 (!lwi.lwi_allow_intr || set->set_allow_intr) &&
2315 signal_pending(current)) {
2316 sigset_t blocked_sigs =
2317 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2319 /* In fact we only interrupt for the "fatal" signals
2320 * like SIGINT or SIGKILL. We still ignore less
2321 * important signals since ptlrpc set is not easily
2322 * reentrant from userspace again */
2323 if (signal_pending(current))
2324 ptlrpc_interrupted_set(set);
2325 cfs_restore_sigs(blocked_sigs);
2328 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2330 /* -EINTR => all requests have been flagged rq_intr so next
2332 * -ETIMEDOUT => someone timed out. When all reqs have
2333 * timed out, signals are enabled allowing completion with
2335 * I don't really care if we go once more round the loop in
2336 * the error cases -eeb. */
2337 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2338 list_for_each(tmp, &set->set_requests) {
2339 req = list_entry(tmp, struct ptlrpc_request,
2341 spin_lock(&req->rq_lock);
2342 req->rq_invalid_rqset = 1;
2343 spin_unlock(&req->rq_lock);
2346 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2348 LASSERT(atomic_read(&set->set_remaining) == 0);
2350 rc = set->set_rc; /* rq_status of already freed requests if any */
2351 list_for_each(tmp, &set->set_requests) {
2352 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2354 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2355 if (req->rq_status != 0)
2356 rc = req->rq_status;
2359 if (set->set_interpret != NULL) {
2360 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2362 rc = interpreter (set, set->set_arg, rc);
2364 struct ptlrpc_set_cbdata *cbdata, *n;
2367 list_for_each_entry_safe(cbdata, n,
2368 &set->set_cblist, psc_item) {
2369 list_del_init(&cbdata->psc_item);
2370 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2373 OBD_FREE_PTR(cbdata);
2379 EXPORT_SYMBOL(ptlrpc_set_wait);
2382 * Helper fuction for request freeing.
2383 * Called when request count reached zero and request needs to be freed.
2384 * Removes request from all sorts of sending/replay lists it might be on,
2385 * frees network buffers if any are present.
2386 * If \a locked is set, that means caller is already holding import imp_lock
2387 * and so we no longer need to reobtain it (for certain lists manipulations)
2389 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2393 if (request == NULL)
2396 LASSERT(!request->rq_srv_req);
2397 LASSERT(request->rq_export == NULL);
2398 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2399 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2400 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2401 LASSERTF(!request->rq_replay, "req %p\n", request);
2403 req_capsule_fini(&request->rq_pill);
2405 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2406 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2407 if (request->rq_import != NULL) {
2409 spin_lock(&request->rq_import->imp_lock);
2410 list_del_init(&request->rq_replay_list);
2411 list_del_init(&request->rq_unreplied_list);
2413 spin_unlock(&request->rq_import->imp_lock);
2415 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2417 if (atomic_read(&request->rq_refcount) != 0) {
2418 DEBUG_REQ(D_ERROR, request,
2419 "freeing request with nonzero refcount");
2423 if (request->rq_repbuf != NULL)
2424 sptlrpc_cli_free_repbuf(request);
2426 if (request->rq_import != NULL) {
2427 class_import_put(request->rq_import);
2428 request->rq_import = NULL;
2430 if (request->rq_bulk != NULL)
2431 ptlrpc_free_bulk(request->rq_bulk);
2433 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2434 sptlrpc_cli_free_reqbuf(request);
2436 if (request->rq_cli_ctx)
2437 sptlrpc_req_put_ctx(request, !locked);
2439 if (request->rq_pool)
2440 __ptlrpc_free_req_to_pool(request);
2442 ptlrpc_request_cache_free(request);
2446 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2448 * Drop one request reference. Must be called with import imp_lock held.
2449 * When reference count drops to zero, request is freed.
2451 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2453 assert_spin_locked(&request->rq_import->imp_lock);
2454 (void)__ptlrpc_req_finished(request, 1);
2459 * Drops one reference count for request \a request.
2460 * \a locked set indicates that caller holds import imp_lock.
2461 * Frees the request whe reference count reaches zero.
2463 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2466 if (request == NULL)
2469 if (request == LP_POISON ||
2470 request->rq_reqmsg == LP_POISON) {
2471 CERROR("dereferencing freed request (bug 575)\n");
2476 DEBUG_REQ(D_INFO, request, "refcount now %u",
2477 atomic_read(&request->rq_refcount) - 1);
2479 if (atomic_dec_and_test(&request->rq_refcount)) {
2480 __ptlrpc_free_req(request, locked);
2488 * Drops one reference count for a request.
2490 void ptlrpc_req_finished(struct ptlrpc_request *request)
2492 __ptlrpc_req_finished(request, 0);
2494 EXPORT_SYMBOL(ptlrpc_req_finished);
2497 * Returns xid of a \a request
2499 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2501 return request->rq_xid;
2503 EXPORT_SYMBOL(ptlrpc_req_xid);
2506 * Disengage the client's reply buffer from the network
2507 * NB does _NOT_ unregister any client-side bulk.
2508 * IDEMPOTENT, but _not_ safe against concurrent callers.
2509 * The request owner (i.e. the thread doing the I/O) must call...
2510 * Returns 0 on success or 1 if unregistering cannot be made.
2512 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2515 struct l_wait_info lwi;
2520 LASSERT(!in_interrupt());
2523 * Let's setup deadline for reply unlink.
2525 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2526 async && request->rq_reply_deadline == 0)
2527 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2530 * Nothing left to do.
2532 if (!ptlrpc_client_recv_or_unlink(request))
2535 LNetMDUnlink(request->rq_reply_md_h);
2538 * Let's check it once again.
2540 if (!ptlrpc_client_recv_or_unlink(request))
2544 * Move to "Unregistering" phase as reply was not unlinked yet.
2546 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2549 * Do not wait for unlink to finish.
2555 * We have to l_wait_event() whatever the result, to give liblustre
2556 * a chance to run reply_in_callback(), and to make sure we've
2557 * unlinked before returning a req to the pool.
2560 /* The wq argument is ignored by user-space wait_event macros */
2561 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2562 &request->rq_set->set_waitq :
2563 &request->rq_reply_waitq;
2564 /* Network access will complete in finite time but the HUGE
2565 * timeout lets us CWARN for visibility of sluggish NALs */
2566 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2567 cfs_time_seconds(1), NULL, NULL);
2568 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2571 ptlrpc_rqphase_move(request, request->rq_next_phase);
2575 LASSERT(rc == -ETIMEDOUT);
2576 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2577 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2578 request->rq_receiving_reply,
2579 request->rq_req_unlinked,
2580 request->rq_reply_unlinked);
2585 static void ptlrpc_free_request(struct ptlrpc_request *req)
2587 spin_lock(&req->rq_lock);
2589 spin_unlock(&req->rq_lock);
2591 if (req->rq_commit_cb != NULL)
2592 req->rq_commit_cb(req);
2593 list_del_init(&req->rq_replay_list);
2595 __ptlrpc_req_finished(req, 1);
2599 * the request is committed and dropped from the replay list of its import
2601 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2603 struct obd_import *imp = req->rq_import;
2605 spin_lock(&imp->imp_lock);
2606 if (list_empty(&req->rq_replay_list)) {
2607 spin_unlock(&imp->imp_lock);
2611 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2612 ptlrpc_free_request(req);
2614 spin_unlock(&imp->imp_lock);
2616 EXPORT_SYMBOL(ptlrpc_request_committed);
2619 * Iterates through replay_list on import and prunes
2620 * all requests have transno smaller than last_committed for the
2621 * import and don't have rq_replay set.
2622 * Since requests are sorted in transno order, stops when meetign first
2623 * transno bigger than last_committed.
2624 * caller must hold imp->imp_lock
2626 void ptlrpc_free_committed(struct obd_import *imp)
2628 struct ptlrpc_request *req, *saved;
2629 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2630 bool skip_committed_list = true;
2633 LASSERT(imp != NULL);
2634 assert_spin_locked(&imp->imp_lock);
2636 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2637 imp->imp_generation == imp->imp_last_generation_checked) {
2638 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2639 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2642 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2643 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2644 imp->imp_generation);
2646 if (imp->imp_generation != imp->imp_last_generation_checked ||
2647 imp->imp_last_transno_checked == 0)
2648 skip_committed_list = false;
2650 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2651 imp->imp_last_generation_checked = imp->imp_generation;
2653 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2655 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2656 LASSERT(req != last_req);
2659 if (req->rq_transno == 0) {
2660 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2663 if (req->rq_import_generation < imp->imp_generation) {
2664 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2668 /* not yet committed */
2669 if (req->rq_transno > imp->imp_peer_committed_transno) {
2670 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2674 if (req->rq_replay) {
2675 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2676 list_move_tail(&req->rq_replay_list,
2677 &imp->imp_committed_list);
2681 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2682 imp->imp_peer_committed_transno);
2684 ptlrpc_free_request(req);
2687 if (skip_committed_list)
2690 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2692 LASSERT(req->rq_transno != 0);
2693 if (req->rq_import_generation < imp->imp_generation) {
2694 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2695 ptlrpc_free_request(req);
2696 } else if (!req->rq_replay) {
2697 DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2698 ptlrpc_free_request(req);
2705 void ptlrpc_cleanup_client(struct obd_import *imp)
2712 * Schedule previously sent request for resend.
2713 * For bulk requests we assign new xid (to avoid problems with
2714 * lost replies and therefore several transfers landing into same buffer
2715 * from different sending attempts).
2717 void ptlrpc_resend_req(struct ptlrpc_request *req)
2719 DEBUG_REQ(D_HA, req, "going to resend");
2720 spin_lock(&req->rq_lock);
2722 /* Request got reply but linked to the import list still.
2723 Let ptlrpc_check_set() to process it. */
2724 if (ptlrpc_client_replied(req)) {
2725 spin_unlock(&req->rq_lock);
2726 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2730 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2731 req->rq_status = -EAGAIN;
2734 req->rq_net_err = 0;
2735 req->rq_timedout = 0;
2737 ptlrpc_client_wake_req(req);
2738 spin_unlock(&req->rq_lock);
2741 /* XXX: this function and rq_status are currently unused */
2742 void ptlrpc_restart_req(struct ptlrpc_request *req)
2744 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2745 req->rq_status = -ERESTARTSYS;
2747 spin_lock(&req->rq_lock);
2748 req->rq_restart = 1;
2749 req->rq_timedout = 0;
2750 ptlrpc_client_wake_req(req);
2751 spin_unlock(&req->rq_lock);
2755 * Grab additional reference on a request \a req
2757 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2760 atomic_inc(&req->rq_refcount);
2763 EXPORT_SYMBOL(ptlrpc_request_addref);
2766 * Add a request to import replay_list.
2767 * Must be called under imp_lock
2769 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2770 struct obd_import *imp)
2772 struct list_head *tmp;
2774 assert_spin_locked(&imp->imp_lock);
2776 if (req->rq_transno == 0) {
2777 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2781 /* clear this for new requests that were resent as well
2782 as resent replayed requests. */
2783 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2785 /* don't re-add requests that have been replayed */
2786 if (!list_empty(&req->rq_replay_list))
2789 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2791 spin_lock(&req->rq_lock);
2793 spin_unlock(&req->rq_lock);
2795 LASSERT(imp->imp_replayable);
2796 /* Balanced in ptlrpc_free_committed, usually. */
2797 ptlrpc_request_addref(req);
2798 list_for_each_prev(tmp, &imp->imp_replay_list) {
2799 struct ptlrpc_request *iter = list_entry(tmp,
2800 struct ptlrpc_request,
2803 /* We may have duplicate transnos if we create and then
2804 * open a file, or for closes retained if to match creating
2805 * opens, so use req->rq_xid as a secondary key.
2806 * (See bugs 684, 685, and 428.)
2807 * XXX no longer needed, but all opens need transnos!
2809 if (iter->rq_transno > req->rq_transno)
2812 if (iter->rq_transno == req->rq_transno) {
2813 LASSERT(iter->rq_xid != req->rq_xid);
2814 if (iter->rq_xid > req->rq_xid)
2818 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2822 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2826 * Send request and wait until it completes.
2827 * Returns request processing status.
2829 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2831 struct ptlrpc_request_set *set;
2835 LASSERT(req->rq_set == NULL);
2836 LASSERT(!req->rq_receiving_reply);
2838 set = ptlrpc_prep_set();
2840 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2844 /* for distributed debugging */
2845 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2847 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2848 ptlrpc_request_addref(req);
2849 ptlrpc_set_add_req(set, req);
2850 rc = ptlrpc_set_wait(set);
2851 ptlrpc_set_destroy(set);
2855 EXPORT_SYMBOL(ptlrpc_queue_wait);
2858 * Callback used for replayed requests reply processing.
2859 * In case of successful reply calls registered request replay callback.
2860 * In case of error restart replay process.
2862 static int ptlrpc_replay_interpret(const struct lu_env *env,
2863 struct ptlrpc_request *req,
2864 void * data, int rc)
2866 struct ptlrpc_replay_async_args *aa = data;
2867 struct obd_import *imp = req->rq_import;
2870 atomic_dec(&imp->imp_replay_inflight);
2872 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2873 * server got the request, but bulk transfer timeout, let's
2874 * replay the bulk req again */
2875 if (!ptlrpc_client_replied(req) ||
2876 (req->rq_bulk != NULL &&
2877 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2878 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2879 GOTO(out, rc = -ETIMEDOUT);
2882 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2883 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2884 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2885 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2887 /** VBR: check version failure */
2888 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2889 /** replay was failed due to version mismatch */
2890 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2891 spin_lock(&imp->imp_lock);
2892 imp->imp_vbr_failed = 1;
2893 imp->imp_no_lock_replay = 1;
2894 spin_unlock(&imp->imp_lock);
2895 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2897 /** The transno had better not change over replay. */
2898 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2899 lustre_msg_get_transno(req->rq_repmsg) ||
2900 lustre_msg_get_transno(req->rq_repmsg) == 0,
2902 lustre_msg_get_transno(req->rq_reqmsg),
2903 lustre_msg_get_transno(req->rq_repmsg));
2906 spin_lock(&imp->imp_lock);
2907 /** if replays by version then gap occur on server, no trust to locks */
2908 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2909 imp->imp_no_lock_replay = 1;
2910 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2911 spin_unlock(&imp->imp_lock);
2912 LASSERT(imp->imp_last_replay_transno);
2914 /* transaction number shouldn't be bigger than the latest replayed */
2915 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2916 DEBUG_REQ(D_ERROR, req,
2917 "Reported transno "LPU64" is bigger than the "
2918 "replayed one: "LPU64, req->rq_transno,
2919 lustre_msg_get_transno(req->rq_reqmsg));
2920 GOTO(out, rc = -EINVAL);
2923 DEBUG_REQ(D_HA, req, "got rep");
2925 /* let the callback do fixups, possibly including in the request */
2926 if (req->rq_replay_cb)
2927 req->rq_replay_cb(req);
2929 if (ptlrpc_client_replied(req) &&
2930 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2931 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2932 lustre_msg_get_status(req->rq_repmsg),
2933 aa->praa_old_status);
2935 /* Note: If the replay fails for MDT-MDT recovery, let's
2936 * abort all of the following requests in the replay
2937 * and sending list, because MDT-MDT update requests
2938 * are dependent on each other, see LU-7039 */
2939 if (imp->imp_connect_flags_orig & OBD_CONNECT_MDS_MDS) {
2940 struct ptlrpc_request *free_req;
2941 struct ptlrpc_request *tmp;
2943 spin_lock(&imp->imp_lock);
2944 list_for_each_entry_safe(free_req, tmp,
2945 &imp->imp_replay_list,
2947 ptlrpc_free_request(free_req);
2950 list_for_each_entry_safe(free_req, tmp,
2951 &imp->imp_committed_list,
2953 ptlrpc_free_request(free_req);
2956 list_for_each_entry_safe(free_req, tmp,
2957 &imp->imp_delayed_list,
2959 spin_lock(&free_req->rq_lock);
2960 free_req->rq_err = 1;
2961 free_req->rq_status = -EIO;
2962 ptlrpc_client_wake_req(free_req);
2963 spin_unlock(&free_req->rq_lock);
2966 list_for_each_entry_safe(free_req, tmp,
2967 &imp->imp_sending_list,
2969 spin_lock(&free_req->rq_lock);
2970 free_req->rq_err = 1;
2971 free_req->rq_status = -EIO;
2972 ptlrpc_client_wake_req(free_req);
2973 spin_unlock(&free_req->rq_lock);
2975 spin_unlock(&imp->imp_lock);
2978 /* Put it back for re-replay. */
2979 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2983 * Errors while replay can set transno to 0, but
2984 * imp_last_replay_transno shouldn't be set to 0 anyway
2986 if (req->rq_transno == 0)
2987 CERROR("Transno is 0 during replay!\n");
2989 /* continue with recovery */
2990 rc = ptlrpc_import_recovery_state_machine(imp);
2992 req->rq_send_state = aa->praa_old_state;
2995 /* this replay failed, so restart recovery */
2996 ptlrpc_connect_import(imp);
3002 * Prepares and queues request for replay.
3003 * Adds it to ptlrpcd queue for actual sending.
3004 * Returns 0 on success.
3006 int ptlrpc_replay_req(struct ptlrpc_request *req)
3008 struct ptlrpc_replay_async_args *aa;
3011 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
3013 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
3014 aa = ptlrpc_req_async_args(req);
3015 memset(aa, 0, sizeof *aa);
3017 /* Prepare request to be resent with ptlrpcd */
3018 aa->praa_old_state = req->rq_send_state;
3019 req->rq_send_state = LUSTRE_IMP_REPLAY;
3020 req->rq_phase = RQ_PHASE_NEW;
3021 req->rq_next_phase = RQ_PHASE_UNDEFINED;
3023 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
3025 req->rq_interpret_reply = ptlrpc_replay_interpret;
3026 /* Readjust the timeout for current conditions */
3027 ptlrpc_at_set_req_timeout(req);
3029 /* Tell server the net_latency, so the server can calculate how long
3030 * it should wait for next replay */
3031 lustre_msg_set_service_time(req->rq_reqmsg,
3032 ptlrpc_at_get_net_latency(req));
3033 DEBUG_REQ(D_HA, req, "REPLAY");
3035 atomic_inc(&req->rq_import->imp_replay_inflight);
3036 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
3038 ptlrpcd_add_req(req);
3043 * Aborts all in-flight request on import \a imp sending and delayed lists
3045 void ptlrpc_abort_inflight(struct obd_import *imp)
3047 struct list_head *tmp, *n;
3050 /* Make sure that no new requests get processed for this import.
3051 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
3052 * this flag and then putting requests on sending_list or delayed_list.
3054 spin_lock(&imp->imp_lock);
3056 /* XXX locking? Maybe we should remove each request with the list
3057 * locked? Also, how do we know if the requests on the list are
3058 * being freed at this time?
3060 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
3061 struct ptlrpc_request *req = list_entry(tmp,
3062 struct ptlrpc_request,
3065 DEBUG_REQ(D_RPCTRACE, req, "inflight");
3067 spin_lock(&req->rq_lock);
3068 if (req->rq_import_generation < imp->imp_generation) {
3070 req->rq_status = -EIO;
3071 ptlrpc_client_wake_req(req);
3073 spin_unlock(&req->rq_lock);
3076 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
3077 struct ptlrpc_request *req =
3078 list_entry(tmp, struct ptlrpc_request, rq_list);
3080 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
3082 spin_lock(&req->rq_lock);
3083 if (req->rq_import_generation < imp->imp_generation) {
3085 req->rq_status = -EIO;
3086 ptlrpc_client_wake_req(req);
3088 spin_unlock(&req->rq_lock);
3091 /* Last chance to free reqs left on the replay list, but we
3092 * will still leak reqs that haven't committed. */
3093 if (imp->imp_replayable)
3094 ptlrpc_free_committed(imp);
3096 spin_unlock(&imp->imp_lock);
3102 * Abort all uncompleted requests in request set \a set
3104 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3106 struct list_head *tmp, *pos;
3108 LASSERT(set != NULL);
3110 list_for_each_safe(pos, tmp, &set->set_requests) {
3111 struct ptlrpc_request *req =
3112 list_entry(pos, struct ptlrpc_request,
3115 spin_lock(&req->rq_lock);
3116 if (req->rq_phase != RQ_PHASE_RPC) {
3117 spin_unlock(&req->rq_lock);
3122 req->rq_status = -EINTR;
3123 ptlrpc_client_wake_req(req);
3124 spin_unlock(&req->rq_lock);
3128 static __u64 ptlrpc_last_xid;
3129 static spinlock_t ptlrpc_last_xid_lock;
3132 * Initialize the XID for the node. This is common among all requests on
3133 * this node, and only requires the property that it is monotonically
3134 * increasing. It does not need to be sequential. Since this is also used
3135 * as the RDMA match bits, it is important that a single client NOT have
3136 * the same match bits for two different in-flight requests, hence we do
3137 * NOT want to have an XID per target or similar.
3139 * To avoid an unlikely collision between match bits after a client reboot
3140 * (which would deliver old data into the wrong RDMA buffer) initialize
3141 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3142 * If the time is clearly incorrect, we instead use a 62-bit random number.
3143 * In the worst case the random number will overflow 1M RPCs per second in
3144 * 9133 years, or permutations thereof.
3146 #define YEAR_2004 (1ULL << 30)
3147 void ptlrpc_init_xid(void)
3149 time_t now = cfs_time_current_sec();
3151 spin_lock_init(&ptlrpc_last_xid_lock);
3152 if (now < YEAR_2004) {
3153 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3154 ptlrpc_last_xid >>= 2;
3155 ptlrpc_last_xid |= (1ULL << 61);
3157 ptlrpc_last_xid = (__u64)now << 20;
3160 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3161 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3162 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3166 * Increase xid and returns resulting new value to the caller.
3168 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3169 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3170 * itself uses the last bulk xid needed, so the server can determine the
3171 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3172 * xid must align to a power-of-two value.
3174 * This is assumed to be true due to the initial ptlrpc_last_xid
3175 * value also being initialized to a power-of-two value. LU-1431
3177 __u64 ptlrpc_next_xid(void)
3181 spin_lock(&ptlrpc_last_xid_lock);
3182 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3183 ptlrpc_last_xid = next;
3184 spin_unlock(&ptlrpc_last_xid_lock);
3190 * If request has a new allocated XID (new request or EINPROGRESS resend),
3191 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3192 * request to ensure previous bulk fails and avoid problems with lost replies
3193 * and therefore several transfers landing into the same buffer from different
3196 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3198 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3200 LASSERT(bd != NULL);
3202 if (!req->rq_resend) {
3203 /* this request has a new xid, just use it as bulk matchbits */
3204 req->rq_mbits = req->rq_xid;
3206 } else { /* needs to generate a new matchbits for resend */
3207 __u64 old_mbits = req->rq_mbits;
3209 if ((bd->bd_import->imp_connect_data.ocd_connect_flags &
3210 OBD_CONNECT_BULK_MBITS) != 0)
3211 req->rq_mbits = ptlrpc_next_xid();
3212 else /* old version transfers rq_xid to peer as matchbits */
3213 req->rq_mbits = req->rq_xid = ptlrpc_next_xid();
3215 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
3216 old_mbits, req->rq_mbits);
3219 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3220 * that server can infer the number of bulks that were prepared,
3222 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3227 * Get a glimpse at what next xid value might have been.
3228 * Returns possible next xid.
3230 __u64 ptlrpc_sample_next_xid(void)
3232 #if BITS_PER_LONG == 32
3233 /* need to avoid possible word tearing on 32-bit systems */
3236 spin_lock(&ptlrpc_last_xid_lock);
3237 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3238 spin_unlock(&ptlrpc_last_xid_lock);
3242 /* No need to lock, since returned value is racy anyways */
3243 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3246 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3249 * Functions for operating ptlrpc workers.
3251 * A ptlrpc work is a function which will be running inside ptlrpc context.
3252 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3254 * 1. after a work is created, it can be used many times, that is:
3255 * handler = ptlrpcd_alloc_work();
3256 * ptlrpcd_queue_work();
3258 * queue it again when necessary:
3259 * ptlrpcd_queue_work();
3260 * ptlrpcd_destroy_work();
3261 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3262 * it will only be queued once in any time. Also as its name implies, it may
3263 * have delay before it really runs by ptlrpcd thread.
3265 struct ptlrpc_work_async_args {
3266 int (*cb)(const struct lu_env *, void *);
3270 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3272 /* re-initialize the req */
3273 req->rq_timeout = obd_timeout;
3274 req->rq_sent = cfs_time_current_sec();
3275 req->rq_deadline = req->rq_sent + req->rq_timeout;
3276 req->rq_reply_deadline = req->rq_deadline;
3277 req->rq_phase = RQ_PHASE_INTERPRET;
3278 req->rq_next_phase = RQ_PHASE_COMPLETE;
3279 req->rq_xid = ptlrpc_next_xid();
3280 req->rq_import_generation = req->rq_import->imp_generation;
3282 ptlrpcd_add_req(req);
3285 static int work_interpreter(const struct lu_env *env,
3286 struct ptlrpc_request *req, void *data, int rc)
3288 struct ptlrpc_work_async_args *arg = data;
3290 LASSERT(ptlrpcd_check_work(req));
3291 LASSERT(arg->cb != NULL);
3293 rc = arg->cb(env, arg->cbdata);
3295 list_del_init(&req->rq_set_chain);
3298 if (atomic_dec_return(&req->rq_refcount) > 1) {
3299 atomic_set(&req->rq_refcount, 2);
3300 ptlrpcd_add_work_req(req);
3305 static int worker_format;
3307 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3309 return req->rq_pill.rc_fmt == (void *)&worker_format;
3313 * Create a work for ptlrpc.
3315 void *ptlrpcd_alloc_work(struct obd_import *imp,
3316 int (*cb)(const struct lu_env *, void *), void *cbdata)
3318 struct ptlrpc_request *req = NULL;
3319 struct ptlrpc_work_async_args *args;
3325 RETURN(ERR_PTR(-EINVAL));
3327 /* copy some code from deprecated fakereq. */
3328 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3330 CERROR("ptlrpc: run out of memory!\n");
3331 RETURN(ERR_PTR(-ENOMEM));
3334 ptlrpc_cli_req_init(req);
3336 req->rq_send_state = LUSTRE_IMP_FULL;
3337 req->rq_type = PTL_RPC_MSG_REQUEST;
3338 req->rq_import = class_import_get(imp);
3339 req->rq_interpret_reply = work_interpreter;
3340 /* don't want reply */
3341 req->rq_no_delay = req->rq_no_resend = 1;
3342 req->rq_pill.rc_fmt = (void *)&worker_format;
3344 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3345 args = ptlrpc_req_async_args(req);
3347 args->cbdata = cbdata;
3351 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3353 void ptlrpcd_destroy_work(void *handler)
3355 struct ptlrpc_request *req = handler;
3358 ptlrpc_req_finished(req);
3360 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3362 int ptlrpcd_queue_work(void *handler)
3364 struct ptlrpc_request *req = handler;
3367 * Check if the req is already being queued.
3369 * Here comes a trick: it lacks a way of checking if a req is being
3370 * processed reliably in ptlrpc. Here I have to use refcount of req
3371 * for this purpose. This is okay because the caller should use this
3372 * req as opaque data. - Jinshan
3374 LASSERT(atomic_read(&req->rq_refcount) > 0);
3375 if (atomic_inc_return(&req->rq_refcount) == 2)
3376 ptlrpcd_add_work_req(req);
3379 EXPORT_SYMBOL(ptlrpcd_queue_work);