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 void 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 void (*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);
596 if (list_empty(&pool->prp_req_list)) {
597 /* have not allocated a single request for the pool */
598 OBD_FREE(pool, sizeof(struct ptlrpc_request_pool));
603 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
606 * Fetches one request from pool \a pool
608 static struct ptlrpc_request *
609 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
611 struct ptlrpc_request *request;
612 struct lustre_msg *reqbuf;
617 spin_lock(&pool->prp_lock);
619 /* See if we have anything in a pool, and bail out if nothing,
620 * in writeout path, where this matters, this is safe to do, because
621 * nothing is lost in this case, and when some in-flight requests
622 * complete, this code will be called again. */
623 if (unlikely(list_empty(&pool->prp_req_list))) {
624 spin_unlock(&pool->prp_lock);
628 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
630 list_del_init(&request->rq_list);
631 spin_unlock(&pool->prp_lock);
633 LASSERT(request->rq_reqbuf);
634 LASSERT(request->rq_pool);
636 reqbuf = request->rq_reqbuf;
637 memset(request, 0, sizeof(*request));
638 request->rq_reqbuf = reqbuf;
639 request->rq_reqbuf_len = pool->prp_rq_size;
640 request->rq_pool = pool;
646 * Returns freed \a request to pool.
648 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
650 struct ptlrpc_request_pool *pool = request->rq_pool;
652 spin_lock(&pool->prp_lock);
653 LASSERT(list_empty(&request->rq_list));
654 LASSERT(!request->rq_receiving_reply);
655 list_add_tail(&request->rq_list, &pool->prp_req_list);
656 spin_unlock(&pool->prp_lock);
659 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
660 __u32 version, int opcode,
661 int count, __u32 *lengths, char **bufs,
662 struct ptlrpc_cli_ctx *ctx)
664 struct obd_import *imp = request->rq_import;
669 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
671 rc = sptlrpc_req_get_ctx(request);
676 sptlrpc_req_set_flavor(request, opcode);
678 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
681 LASSERT(!request->rq_pool);
685 lustre_msg_add_version(request->rq_reqmsg, version);
686 request->rq_send_state = LUSTRE_IMP_FULL;
687 request->rq_type = PTL_RPC_MSG_REQUEST;
689 request->rq_req_cbid.cbid_fn = request_out_callback;
690 request->rq_req_cbid.cbid_arg = request;
692 request->rq_reply_cbid.cbid_fn = reply_in_callback;
693 request->rq_reply_cbid.cbid_arg = request;
695 request->rq_reply_deadline = 0;
696 request->rq_phase = RQ_PHASE_NEW;
697 request->rq_next_phase = RQ_PHASE_UNDEFINED;
699 request->rq_request_portal = imp->imp_client->cli_request_portal;
700 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
702 ptlrpc_at_set_req_timeout(request);
704 lustre_msg_set_opc(request->rq_reqmsg, opcode);
708 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
710 class_import_put(imp);
714 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
715 __u32 version, int opcode, char **bufs,
716 struct ptlrpc_cli_ctx *ctx)
720 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
721 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
722 request->rq_pill.rc_area[RCL_CLIENT],
725 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
728 * Pack request buffers for network transfer, performing necessary encryption
729 * steps if necessary.
731 int ptlrpc_request_pack(struct ptlrpc_request *request,
732 __u32 version, int opcode)
735 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
739 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
740 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
741 * have to send old ptlrpc_body to keep interoprability with these
744 * Only three kinds of server->client RPCs so far:
749 * XXX This should be removed whenever we drop the interoprability with
750 * the these old clients.
752 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
753 opcode == LDLM_GL_CALLBACK)
754 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
755 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
759 EXPORT_SYMBOL(ptlrpc_request_pack);
762 * Helper function to allocate new request on import \a imp
763 * and possibly using existing request from pool \a pool if provided.
764 * Returns allocated request structure with import field filled or
768 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
769 struct ptlrpc_request_pool *pool)
771 struct ptlrpc_request *request = NULL;
774 request = ptlrpc_prep_req_from_pool(pool);
777 request = ptlrpc_request_cache_alloc(GFP_NOFS);
780 ptlrpc_cli_req_init(request);
782 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
783 LASSERT(imp != LP_POISON);
784 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
786 LASSERT(imp->imp_client != LP_POISON);
788 request->rq_import = class_import_get(imp);
790 CERROR("request allocation out of memory\n");
797 * Helper function for creating a request.
798 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
799 * buffer structures according to capsule template \a format.
800 * Returns allocated request structure pointer or NULL on error.
802 static struct ptlrpc_request *
803 ptlrpc_request_alloc_internal(struct obd_import *imp,
804 struct ptlrpc_request_pool * pool,
805 const struct req_format *format)
807 struct ptlrpc_request *request;
809 request = __ptlrpc_request_alloc(imp, pool);
813 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
814 req_capsule_set(&request->rq_pill, format);
819 * Allocate new request structure for import \a imp and initialize its
820 * buffer structure according to capsule template \a format.
822 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
823 const struct req_format *format)
825 return ptlrpc_request_alloc_internal(imp, NULL, format);
827 EXPORT_SYMBOL(ptlrpc_request_alloc);
830 * Allocate new request structure for import \a imp from pool \a pool and
831 * initialize its buffer structure according to capsule template \a format.
833 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
834 struct ptlrpc_request_pool * pool,
835 const struct req_format *format)
837 return ptlrpc_request_alloc_internal(imp, pool, format);
839 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
842 * For requests not from pool, free memory of the request structure.
843 * For requests obtained from a pool earlier, return request back to pool.
845 void ptlrpc_request_free(struct ptlrpc_request *request)
847 if (request->rq_pool)
848 __ptlrpc_free_req_to_pool(request);
850 ptlrpc_request_cache_free(request);
852 EXPORT_SYMBOL(ptlrpc_request_free);
855 * Allocate new request for operatione \a opcode and immediatelly pack it for
857 * Only used for simple requests like OBD_PING where the only important
858 * part of the request is operation itself.
859 * Returns allocated request or NULL on error.
861 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
862 const struct req_format *format,
863 __u32 version, int opcode)
865 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
869 rc = ptlrpc_request_pack(req, version, opcode);
871 ptlrpc_request_free(req);
877 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
880 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
881 * for operation \a opcode. Request would contain \a count buffers.
882 * Sizes of buffers are described in array \a lengths and buffers themselves
883 * are provided by a pointer \a bufs.
884 * Returns prepared request structure pointer or NULL on error.
886 struct ptlrpc_request *
887 ptlrpc_prep_req_pool(struct obd_import *imp,
888 __u32 version, int opcode,
889 int count, __u32 *lengths, char **bufs,
890 struct ptlrpc_request_pool *pool)
892 struct ptlrpc_request *request;
895 request = __ptlrpc_request_alloc(imp, pool);
899 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
900 lengths, bufs, NULL);
902 ptlrpc_request_free(request);
909 * Same as ptlrpc_prep_req_pool, but without pool
911 struct ptlrpc_request *
912 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
913 __u32 *lengths, char **bufs)
915 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
920 * Allocate and initialize new request set structure on the current CPT.
921 * Returns a pointer to the newly allocated set structure or NULL on error.
923 struct ptlrpc_request_set *ptlrpc_prep_set(void)
925 struct ptlrpc_request_set *set;
929 cpt = cfs_cpt_current(cfs_cpt_table, 0);
930 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
933 atomic_set(&set->set_refcount, 1);
934 INIT_LIST_HEAD(&set->set_requests);
935 init_waitqueue_head(&set->set_waitq);
936 atomic_set(&set->set_new_count, 0);
937 atomic_set(&set->set_remaining, 0);
938 spin_lock_init(&set->set_new_req_lock);
939 INIT_LIST_HEAD(&set->set_new_requests);
940 INIT_LIST_HEAD(&set->set_cblist);
941 set->set_max_inflight = UINT_MAX;
942 set->set_producer = NULL;
943 set->set_producer_arg = NULL;
948 EXPORT_SYMBOL(ptlrpc_prep_set);
951 * Allocate and initialize new request set structure with flow control
952 * extension. This extension allows to control the number of requests in-flight
953 * for the whole set. A callback function to generate requests must be provided
954 * and the request set will keep the number of requests sent over the wire to
956 * Returns a pointer to the newly allocated set structure or NULL on error.
958 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
962 struct ptlrpc_request_set *set;
964 set = ptlrpc_prep_set();
968 set->set_max_inflight = max;
969 set->set_producer = func;
970 set->set_producer_arg = arg;
976 * Wind down and free request set structure previously allocated with
978 * Ensures that all requests on the set have completed and removes
979 * all requests from the request list in a set.
980 * If any unsent request happen to be on the list, pretends that they got
981 * an error in flight and calls their completion handler.
983 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
985 struct list_head *tmp;
986 struct list_head *next;
991 /* Requests on the set should either all be completed, or all be new */
992 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
993 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
994 list_for_each(tmp, &set->set_requests) {
995 struct ptlrpc_request *req =
996 list_entry(tmp, struct ptlrpc_request,
999 LASSERT(req->rq_phase == expected_phase);
1003 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
1004 atomic_read(&set->set_remaining) == n, "%d / %d\n",
1005 atomic_read(&set->set_remaining), n);
1007 list_for_each_safe(tmp, next, &set->set_requests) {
1008 struct ptlrpc_request *req =
1009 list_entry(tmp, struct ptlrpc_request,
1011 list_del_init(&req->rq_set_chain);
1013 LASSERT(req->rq_phase == expected_phase);
1015 if (req->rq_phase == RQ_PHASE_NEW) {
1016 ptlrpc_req_interpret(NULL, req, -EBADR);
1017 atomic_dec(&set->set_remaining);
1020 spin_lock(&req->rq_lock);
1022 req->rq_invalid_rqset = 0;
1023 spin_unlock(&req->rq_lock);
1025 ptlrpc_req_finished (req);
1028 LASSERT(atomic_read(&set->set_remaining) == 0);
1030 ptlrpc_reqset_put(set);
1033 EXPORT_SYMBOL(ptlrpc_set_destroy);
1036 * Add a callback function \a fn to the set.
1037 * This function would be called when all requests on this set are completed.
1038 * The function will be passed \a data argument.
1040 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1041 set_interpreter_func fn, void *data)
1043 struct ptlrpc_set_cbdata *cbdata;
1045 OBD_ALLOC_PTR(cbdata);
1049 cbdata->psc_interpret = fn;
1050 cbdata->psc_data = data;
1051 list_add_tail(&cbdata->psc_item, &set->set_cblist);
1057 * Add a new request to the general purpose request set.
1058 * Assumes request reference from the caller.
1060 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1061 struct ptlrpc_request *req)
1063 LASSERT(list_empty(&req->rq_set_chain));
1065 /* The set takes over the caller's request reference */
1066 list_add_tail(&req->rq_set_chain, &set->set_requests);
1068 atomic_inc(&set->set_remaining);
1069 req->rq_queued_time = cfs_time_current();
1071 if (req->rq_reqmsg != NULL)
1072 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1074 if (set->set_producer != NULL)
1075 /* If the request set has a producer callback, the RPC must be
1076 * sent straight away */
1077 ptlrpc_send_new_req(req);
1079 EXPORT_SYMBOL(ptlrpc_set_add_req);
1082 * Add a request to a request with dedicated server thread
1083 * and wake the thread to make any necessary processing.
1084 * Currently only used for ptlrpcd.
1086 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1087 struct ptlrpc_request *req)
1089 struct ptlrpc_request_set *set = pc->pc_set;
1092 LASSERT(req->rq_set == NULL);
1093 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1095 spin_lock(&set->set_new_req_lock);
1097 * The set takes over the caller's request reference.
1100 req->rq_queued_time = cfs_time_current();
1101 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1102 count = atomic_inc_return(&set->set_new_count);
1103 spin_unlock(&set->set_new_req_lock);
1105 /* Only need to call wakeup once for the first entry. */
1107 wake_up(&set->set_waitq);
1109 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1110 * guarantee the async RPC can be processed ASAP, we have
1111 * no other better choice. It maybe fixed in future. */
1112 for (i = 0; i < pc->pc_npartners; i++)
1113 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1118 * Based on the current state of the import, determine if the request
1119 * can be sent, is an error, or should be delayed.
1121 * Returns true if this request should be delayed. If false, and
1122 * *status is set, then the request can not be sent and *status is the
1123 * error code. If false and status is 0, then request can be sent.
1125 * The imp->imp_lock must be held.
1127 static int ptlrpc_import_delay_req(struct obd_import *imp,
1128 struct ptlrpc_request *req, int *status)
1133 LASSERT (status != NULL);
1136 if (req->rq_ctx_init || req->rq_ctx_fini) {
1137 /* always allow ctx init/fini rpc go through */
1138 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1139 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1141 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1142 /* pings may safely race with umount */
1143 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1144 D_HA : D_ERROR, req, "IMP_CLOSED ");
1146 } else if (ptlrpc_send_limit_expired(req)) {
1147 /* probably doesn't need to be a D_ERROR after initial testing*/
1148 DEBUG_REQ(D_HA, req, "send limit expired ");
1149 *status = -ETIMEDOUT;
1150 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1151 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1152 /* allow CONNECT even if import is invalid */ ;
1153 if (atomic_read(&imp->imp_inval_count) != 0) {
1154 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1157 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1158 if (!imp->imp_deactive)
1159 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1160 *status = -ESHUTDOWN; /* bz 12940 */
1161 } else if (req->rq_import_generation != imp->imp_generation) {
1162 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1164 } else if (req->rq_send_state != imp->imp_state) {
1165 /* invalidate in progress - any requests should be drop */
1166 if (atomic_read(&imp->imp_inval_count) != 0) {
1167 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1169 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1170 *status = -EWOULDBLOCK;
1171 } else if (req->rq_allow_replay &&
1172 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1173 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1174 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1175 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1176 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1186 * Decide if the error message should be printed to the console or not.
1187 * Makes its decision based on request type, status, and failure frequency.
1189 * \param[in] req request that failed and may need a console message
1191 * \retval false if no message should be printed
1192 * \retval true if console message should be printed
1194 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1198 LASSERT(req->rq_reqmsg != NULL);
1199 opc = lustre_msg_get_opc(req->rq_reqmsg);
1201 /* Suppress particular reconnect errors which are to be expected. */
1202 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1205 /* Suppress timed out reconnect requests */
1206 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1210 /* Suppress most unavailable/again reconnect requests, but
1211 * print occasionally so it is clear client is trying to
1212 * connect to a server where no target is running. */
1213 err = lustre_msg_get_status(req->rq_repmsg);
1214 if ((err == -ENODEV || err == -EAGAIN) &&
1215 req->rq_import->imp_conn_cnt % 30 != 20)
1223 * Check request processing status.
1224 * Returns the status.
1226 static int ptlrpc_check_status(struct ptlrpc_request *req)
1231 err = lustre_msg_get_status(req->rq_repmsg);
1232 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1233 struct obd_import *imp = req->rq_import;
1234 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1235 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1237 if (ptlrpc_console_allow(req))
1238 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1239 "failed: rc = %d\n",
1240 imp->imp_obd->obd_name,
1242 libcfs_nid2str(nid), err);
1243 RETURN(err < 0 ? err : -EINVAL);
1247 DEBUG_REQ(D_INFO, req, "status is %d", err);
1248 } else if (err > 0) {
1249 /* XXX: translate this error from net to host */
1250 DEBUG_REQ(D_INFO, req, "status is %d", err);
1257 * save pre-versions of objects into request for replay.
1258 * Versions are obtained from server reply.
1261 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1263 struct lustre_msg *repmsg = req->rq_repmsg;
1264 struct lustre_msg *reqmsg = req->rq_reqmsg;
1265 __u64 *versions = lustre_msg_get_versions(repmsg);
1268 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1272 lustre_msg_set_versions(reqmsg, versions);
1273 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1274 versions[0], versions[1]);
1280 * Callback function called when client receives RPC reply for \a req.
1281 * Returns 0 on success or error code.
1282 * The return alue would be assigned to req->rq_status by the caller
1283 * as request processing status.
1284 * This function also decides if the request needs to be saved for later replay.
1286 static int after_reply(struct ptlrpc_request *req)
1288 struct obd_import *imp = req->rq_import;
1289 struct obd_device *obd = req->rq_import->imp_obd;
1291 struct timeval work_start;
1295 LASSERT(obd != NULL);
1296 /* repbuf must be unlinked */
1297 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1299 if (req->rq_reply_truncated) {
1300 if (ptlrpc_no_resend(req)) {
1301 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1302 " expected: %d, actual size: %d",
1303 req->rq_nob_received, req->rq_repbuf_len);
1307 sptlrpc_cli_free_repbuf(req);
1308 /* Pass the required reply buffer size (include
1309 * space for early reply).
1310 * NB: no need to roundup because alloc_repbuf
1311 * will roundup it */
1312 req->rq_replen = req->rq_nob_received;
1313 req->rq_nob_received = 0;
1314 spin_lock(&req->rq_lock);
1316 spin_unlock(&req->rq_lock);
1320 do_gettimeofday(&work_start);
1321 timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1324 * NB Until this point, the whole of the incoming message,
1325 * including buflens, status etc is in the sender's byte order.
1327 rc = sptlrpc_cli_unwrap_reply(req);
1329 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1334 * Security layer unwrap might ask resend this request.
1339 rc = unpack_reply(req);
1343 /* retry indefinitely on EINPROGRESS */
1344 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1345 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1346 time_t now = cfs_time_current_sec();
1348 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1349 spin_lock(&req->rq_lock);
1351 spin_unlock(&req->rq_lock);
1352 req->rq_nr_resend++;
1354 /* Readjust the timeout for current conditions */
1355 ptlrpc_at_set_req_timeout(req);
1356 /* delay resend to give a chance to the server to get ready.
1357 * The delay is increased by 1s on every resend and is capped to
1358 * the current request timeout (i.e. obd_timeout if AT is off,
1359 * or AT service time x 125% + 5s, see at_est2timeout) */
1360 if (req->rq_nr_resend > req->rq_timeout)
1361 req->rq_sent = now + req->rq_timeout;
1363 req->rq_sent = now + req->rq_nr_resend;
1368 if (obd->obd_svc_stats != NULL) {
1369 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1371 ptlrpc_lprocfs_rpc_sent(req, timediff);
1374 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1375 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1376 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1377 lustre_msg_get_type(req->rq_repmsg));
1381 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1382 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1383 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1384 ptlrpc_at_adj_net_latency(req,
1385 lustre_msg_get_service_time(req->rq_repmsg));
1387 rc = ptlrpc_check_status(req);
1388 imp->imp_connect_error = rc;
1392 * Either we've been evicted, or the server has failed for
1393 * some reason. Try to reconnect, and if that fails, punt to
1396 if (ptlrpc_recoverable_error(rc)) {
1397 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1398 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1401 ptlrpc_request_handle_notconn(req);
1406 * Let's look if server sent slv. Do it only for RPC with
1409 ldlm_cli_update_pool(req);
1413 * Store transno in reqmsg for replay.
1415 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1416 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1417 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1420 if (imp->imp_replayable) {
1421 spin_lock(&imp->imp_lock);
1423 * No point in adding already-committed requests to the replay
1424 * list, we will just remove them immediately. b=9829
1426 if (req->rq_transno != 0 &&
1428 lustre_msg_get_last_committed(req->rq_repmsg) ||
1430 /** version recovery */
1431 ptlrpc_save_versions(req);
1432 ptlrpc_retain_replayable_request(req, imp);
1433 } else if (req->rq_commit_cb != NULL &&
1434 list_empty(&req->rq_replay_list)) {
1435 /* NB: don't call rq_commit_cb if it's already on
1436 * rq_replay_list, ptlrpc_free_committed() will call
1437 * it later, see LU-3618 for details */
1438 spin_unlock(&imp->imp_lock);
1439 req->rq_commit_cb(req);
1440 spin_lock(&imp->imp_lock);
1444 * Replay-enabled imports return commit-status information.
1446 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1447 imp->imp_peer_committed_transno =
1448 lustre_msg_get_last_committed(req->rq_repmsg);
1451 ptlrpc_free_committed(imp);
1453 if (!list_empty(&imp->imp_replay_list)) {
1454 struct ptlrpc_request *last;
1456 last = list_entry(imp->imp_replay_list.prev,
1457 struct ptlrpc_request,
1460 * Requests with rq_replay stay on the list even if no
1461 * commit is expected.
1463 if (last->rq_transno > imp->imp_peer_committed_transno)
1464 ptlrpc_pinger_commit_expected(imp);
1467 spin_unlock(&imp->imp_lock);
1474 * Helper function to send request \a req over the network for the first time
1475 * Also adjusts request phase.
1476 * Returns 0 on success or error code.
1478 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1480 struct obd_import *imp = req->rq_import;
1481 struct list_head *tmp;
1482 __u64 min_xid = ~0ULL;
1486 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1487 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1488 (!req->rq_generation_set ||
1489 req->rq_import_generation == imp->imp_generation))
1492 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1494 spin_lock(&imp->imp_lock);
1496 /* the very first time we assign XID. it's important to assign XID
1497 * and put it on the list atomically, so that the lowest assigned
1498 * XID is always known. this is vital for multislot last_rcvd */
1499 if (req->rq_send_state == LUSTRE_IMP_REPLAY) {
1500 LASSERT(req->rq_xid != 0);
1502 LASSERT(req->rq_xid == 0);
1503 req->rq_xid = ptlrpc_next_xid();
1506 if (!req->rq_generation_set)
1507 req->rq_import_generation = imp->imp_generation;
1509 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1510 spin_lock(&req->rq_lock);
1511 req->rq_waiting = 1;
1512 spin_unlock(&req->rq_lock);
1514 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1515 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1516 ptlrpc_import_state_name(req->rq_send_state),
1517 ptlrpc_import_state_name(imp->imp_state));
1518 LASSERT(list_empty(&req->rq_list));
1519 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1520 atomic_inc(&req->rq_import->imp_inflight);
1521 spin_unlock(&imp->imp_lock);
1526 spin_unlock(&imp->imp_lock);
1527 req->rq_status = rc;
1528 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1532 LASSERT(list_empty(&req->rq_list));
1533 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1534 atomic_inc(&req->rq_import->imp_inflight);
1536 /* find the lowest unreplied XID */
1537 list_for_each(tmp, &imp->imp_delayed_list) {
1538 struct ptlrpc_request *r;
1539 r = list_entry(tmp, struct ptlrpc_request, rq_list);
1540 if (r->rq_xid < min_xid)
1541 min_xid = r->rq_xid;
1543 list_for_each(tmp, &imp->imp_sending_list) {
1544 struct ptlrpc_request *r;
1545 r = list_entry(tmp, struct ptlrpc_request, rq_list);
1546 if (r->rq_xid < min_xid)
1547 min_xid = r->rq_xid;
1549 spin_unlock(&imp->imp_lock);
1551 if (likely(min_xid != ~0ULL))
1552 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid - 1);
1554 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1556 rc = sptlrpc_req_refresh_ctx(req, -1);
1559 req->rq_status = rc;
1562 spin_lock(&req->rq_lock);
1563 req->rq_wait_ctx = 1;
1564 spin_unlock(&req->rq_lock);
1569 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1570 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1571 imp->imp_obd->obd_uuid.uuid,
1572 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1573 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1574 lustre_msg_get_opc(req->rq_reqmsg));
1576 rc = ptl_send_rpc(req, 0);
1578 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1579 spin_lock(&req->rq_lock);
1580 req->rq_net_err = 1;
1581 spin_unlock(&req->rq_lock);
1587 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1592 LASSERT(set->set_producer != NULL);
1594 remaining = atomic_read(&set->set_remaining);
1596 /* populate the ->set_requests list with requests until we
1597 * reach the maximum number of RPCs in flight for this set */
1598 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1599 rc = set->set_producer(set, set->set_producer_arg);
1600 if (rc == -ENOENT) {
1601 /* no more RPC to produce */
1602 set->set_producer = NULL;
1603 set->set_producer_arg = NULL;
1608 RETURN((atomic_read(&set->set_remaining) - remaining));
1612 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1613 * and no more replies are expected.
1614 * (it is possible to get less replies than requests sent e.g. due to timed out
1615 * requests or requests that we had trouble to send out)
1617 * NOTE: This function contains a potential schedule point (cond_resched()).
1619 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1621 struct list_head *tmp, *next;
1622 struct list_head comp_reqs;
1623 int force_timer_recalc = 0;
1626 if (atomic_read(&set->set_remaining) == 0)
1629 INIT_LIST_HEAD(&comp_reqs);
1630 list_for_each_safe(tmp, next, &set->set_requests) {
1631 struct ptlrpc_request *req =
1632 list_entry(tmp, struct ptlrpc_request,
1634 struct obd_import *imp = req->rq_import;
1635 int unregistered = 0;
1638 /* This schedule point is mainly for the ptlrpcd caller of this
1639 * function. Most ptlrpc sets are not long-lived and unbounded
1640 * in length, but at the least the set used by the ptlrpcd is.
1641 * Since the processing time is unbounded, we need to insert an
1642 * explicit schedule point to make the thread well-behaved.
1646 if (req->rq_phase == RQ_PHASE_NEW &&
1647 ptlrpc_send_new_req(req)) {
1648 force_timer_recalc = 1;
1651 /* delayed send - skip */
1652 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1655 /* delayed resend - skip */
1656 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1657 req->rq_sent > cfs_time_current_sec())
1660 if (!(req->rq_phase == RQ_PHASE_RPC ||
1661 req->rq_phase == RQ_PHASE_BULK ||
1662 req->rq_phase == RQ_PHASE_INTERPRET ||
1663 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1664 req->rq_phase == RQ_PHASE_COMPLETE)) {
1665 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1669 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1670 LASSERT(req->rq_next_phase != req->rq_phase);
1671 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1674 * Skip processing until reply is unlinked. We
1675 * can't return to pool before that and we can't
1676 * call interpret before that. We need to make
1677 * sure that all rdma transfers finished and will
1678 * not corrupt any data.
1680 if (ptlrpc_client_recv_or_unlink(req) ||
1681 ptlrpc_client_bulk_active(req))
1685 * Turn fail_loc off to prevent it from looping
1688 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1689 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1692 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1693 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1698 * Move to next phase if reply was successfully
1701 ptlrpc_rqphase_move(req, req->rq_next_phase);
1704 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1705 list_move_tail(&req->rq_set_chain, &comp_reqs);
1709 if (req->rq_phase == RQ_PHASE_INTERPRET)
1710 GOTO(interpret, req->rq_status);
1713 * Note that this also will start async reply unlink.
1715 if (req->rq_net_err && !req->rq_timedout) {
1716 ptlrpc_expire_one_request(req, 1);
1719 * Check if we still need to wait for unlink.
1721 if (ptlrpc_client_recv_or_unlink(req) ||
1722 ptlrpc_client_bulk_active(req))
1724 /* If there is no need to resend, fail it now. */
1725 if (req->rq_no_resend) {
1726 if (req->rq_status == 0)
1727 req->rq_status = -EIO;
1728 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1729 GOTO(interpret, req->rq_status);
1736 spin_lock(&req->rq_lock);
1737 req->rq_replied = 0;
1738 spin_unlock(&req->rq_lock);
1739 if (req->rq_status == 0)
1740 req->rq_status = -EIO;
1741 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1742 GOTO(interpret, req->rq_status);
1745 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1746 * so it sets rq_intr regardless of individual rpc
1747 * timeouts. The synchronous IO waiting path sets
1748 * rq_intr irrespective of whether ptlrpcd
1749 * has seen a timeout. Our policy is to only interpret
1750 * interrupted rpcs after they have timed out, so we
1751 * need to enforce that here.
1754 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1755 req->rq_wait_ctx)) {
1756 req->rq_status = -EINTR;
1757 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1758 GOTO(interpret, req->rq_status);
1761 if (req->rq_phase == RQ_PHASE_RPC) {
1762 if (req->rq_timedout || req->rq_resend ||
1763 req->rq_waiting || req->rq_wait_ctx) {
1766 if (!ptlrpc_unregister_reply(req, 1)) {
1767 ptlrpc_unregister_bulk(req, 1);
1771 spin_lock(&imp->imp_lock);
1772 if (ptlrpc_import_delay_req(imp, req, &status)){
1773 /* put on delay list - only if we wait
1774 * recovery finished - before send */
1775 list_del_init(&req->rq_list);
1776 list_add_tail(&req->rq_list,
1779 spin_unlock(&imp->imp_lock);
1784 req->rq_status = status;
1785 ptlrpc_rqphase_move(req,
1786 RQ_PHASE_INTERPRET);
1787 spin_unlock(&imp->imp_lock);
1788 GOTO(interpret, req->rq_status);
1790 if (ptlrpc_no_resend(req) &&
1791 !req->rq_wait_ctx) {
1792 req->rq_status = -ENOTCONN;
1793 ptlrpc_rqphase_move(req,
1794 RQ_PHASE_INTERPRET);
1795 spin_unlock(&imp->imp_lock);
1796 GOTO(interpret, req->rq_status);
1799 list_del_init(&req->rq_list);
1800 list_add_tail(&req->rq_list,
1801 &imp->imp_sending_list);
1803 spin_unlock(&imp->imp_lock);
1805 spin_lock(&req->rq_lock);
1806 req->rq_waiting = 0;
1807 spin_unlock(&req->rq_lock);
1809 if (req->rq_timedout || req->rq_resend) {
1810 /* This is re-sending anyways,
1811 * let's mark req as resend. */
1812 spin_lock(&req->rq_lock);
1814 spin_unlock(&req->rq_lock);
1816 if (req->rq_bulk != NULL &&
1817 !ptlrpc_unregister_bulk(req, 1))
1821 * rq_wait_ctx is only touched by ptlrpcd,
1822 * so no lock is needed here.
1824 status = sptlrpc_req_refresh_ctx(req, -1);
1827 req->rq_status = status;
1828 spin_lock(&req->rq_lock);
1829 req->rq_wait_ctx = 0;
1830 spin_unlock(&req->rq_lock);
1831 force_timer_recalc = 1;
1833 spin_lock(&req->rq_lock);
1834 req->rq_wait_ctx = 1;
1835 spin_unlock(&req->rq_lock);
1840 spin_lock(&req->rq_lock);
1841 req->rq_wait_ctx = 0;
1842 spin_unlock(&req->rq_lock);
1845 rc = ptl_send_rpc(req, 0);
1847 DEBUG_REQ(D_HA, req,
1848 "send failed: rc = %d", rc);
1849 force_timer_recalc = 1;
1850 spin_lock(&req->rq_lock);
1851 req->rq_net_err = 1;
1852 spin_unlock(&req->rq_lock);
1855 /* need to reset the timeout */
1856 force_timer_recalc = 1;
1859 spin_lock(&req->rq_lock);
1861 if (ptlrpc_client_early(req)) {
1862 ptlrpc_at_recv_early_reply(req);
1863 spin_unlock(&req->rq_lock);
1867 /* Still waiting for a reply? */
1868 if (ptlrpc_client_recv(req)) {
1869 spin_unlock(&req->rq_lock);
1873 /* Did we actually receive a reply? */
1874 if (!ptlrpc_client_replied(req)) {
1875 spin_unlock(&req->rq_lock);
1879 spin_unlock(&req->rq_lock);
1881 /* unlink from net because we are going to
1882 * swab in-place of reply buffer */
1883 unregistered = ptlrpc_unregister_reply(req, 1);
1887 req->rq_status = after_reply(req);
1891 /* If there is no bulk associated with this request,
1892 * then we're done and should let the interpreter
1893 * process the reply. Similarly if the RPC returned
1894 * an error, and therefore the bulk will never arrive.
1896 if (req->rq_bulk == NULL || req->rq_status < 0) {
1897 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1898 GOTO(interpret, req->rq_status);
1901 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1904 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1905 if (ptlrpc_client_bulk_active(req))
1908 if (req->rq_bulk->bd_failure) {
1909 /* The RPC reply arrived OK, but the bulk screwed
1910 * up! Dead weird since the server told us the RPC
1911 * was good after getting the REPLY for her GET or
1912 * the ACK for her PUT. */
1913 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1914 req->rq_status = -EIO;
1917 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1920 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1922 /* This moves to "unregistering" phase we need to wait for
1924 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1925 /* start async bulk unlink too */
1926 ptlrpc_unregister_bulk(req, 1);
1930 if (!ptlrpc_unregister_bulk(req, 1))
1933 /* When calling interpret receiving already should be
1935 LASSERT(!req->rq_receiving_reply);
1937 ptlrpc_req_interpret(env, req, req->rq_status);
1939 if (ptlrpcd_check_work(req)) {
1940 atomic_dec(&set->set_remaining);
1943 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1945 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1946 "Completed RPC pname:cluuid:pid:xid:nid:"
1947 "opc %s:%s:%d:"LPU64":%s:%d\n",
1948 current_comm(), imp->imp_obd->obd_uuid.uuid,
1949 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1950 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1951 lustre_msg_get_opc(req->rq_reqmsg));
1953 spin_lock(&imp->imp_lock);
1954 /* Request already may be not on sending or delaying list. This
1955 * may happen in the case of marking it erroneous for the case
1956 * ptlrpc_import_delay_req(req, status) find it impossible to
1957 * allow sending this rpc and returns *status != 0. */
1958 if (!list_empty(&req->rq_list)) {
1959 list_del_init(&req->rq_list);
1960 atomic_dec(&imp->imp_inflight);
1962 spin_unlock(&imp->imp_lock);
1964 atomic_dec(&set->set_remaining);
1965 wake_up_all(&imp->imp_recovery_waitq);
1967 if (set->set_producer) {
1968 /* produce a new request if possible */
1969 if (ptlrpc_set_producer(set) > 0)
1970 force_timer_recalc = 1;
1972 /* free the request that has just been completed
1973 * in order not to pollute set->set_requests */
1974 list_del_init(&req->rq_set_chain);
1975 spin_lock(&req->rq_lock);
1977 req->rq_invalid_rqset = 0;
1978 spin_unlock(&req->rq_lock);
1980 /* record rq_status to compute the final status later */
1981 if (req->rq_status != 0)
1982 set->set_rc = req->rq_status;
1983 ptlrpc_req_finished(req);
1985 list_move_tail(&req->rq_set_chain, &comp_reqs);
1989 /* move completed request at the head of list so it's easier for
1990 * caller to find them */
1991 list_splice(&comp_reqs, &set->set_requests);
1993 /* If we hit an error, we want to recover promptly. */
1994 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1996 EXPORT_SYMBOL(ptlrpc_check_set);
1999 * Time out request \a req. is \a async_unlink is set, that means do not wait
2000 * until LNet actually confirms network buffer unlinking.
2001 * Return 1 if we should give up further retrying attempts or 0 otherwise.
2003 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
2005 struct obd_import *imp = req->rq_import;
2009 spin_lock(&req->rq_lock);
2010 req->rq_timedout = 1;
2011 spin_unlock(&req->rq_lock);
2013 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
2014 "/real "CFS_DURATION_T"]",
2015 req->rq_net_err ? "failed due to network error" :
2016 ((req->rq_real_sent == 0 ||
2017 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
2018 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
2019 "timed out for sent delay" : "timed out for slow reply"),
2020 req->rq_sent, req->rq_real_sent);
2022 if (imp != NULL && obd_debug_peer_on_timeout)
2023 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
2025 ptlrpc_unregister_reply(req, async_unlink);
2026 ptlrpc_unregister_bulk(req, async_unlink);
2028 if (obd_dump_on_timeout)
2029 libcfs_debug_dumplog();
2032 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2036 atomic_inc(&imp->imp_timeouts);
2038 /* The DLM server doesn't want recovery run on its imports. */
2039 if (imp->imp_dlm_fake)
2042 /* If this request is for recovery or other primordial tasks,
2043 * then error it out here. */
2044 if (req->rq_ctx_init || req->rq_ctx_fini ||
2045 req->rq_send_state != LUSTRE_IMP_FULL ||
2046 imp->imp_obd->obd_no_recov) {
2047 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2048 ptlrpc_import_state_name(req->rq_send_state),
2049 ptlrpc_import_state_name(imp->imp_state));
2050 spin_lock(&req->rq_lock);
2051 req->rq_status = -ETIMEDOUT;
2053 spin_unlock(&req->rq_lock);
2057 /* if a request can't be resent we can't wait for an answer after
2059 if (ptlrpc_no_resend(req)) {
2060 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2064 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2070 * Time out all uncompleted requests in request set pointed by \a data
2071 * Callback used when waiting on sets with l_wait_event.
2074 int ptlrpc_expired_set(void *data)
2076 struct ptlrpc_request_set *set = data;
2077 struct list_head *tmp;
2078 time_t now = cfs_time_current_sec();
2081 LASSERT(set != NULL);
2084 * A timeout expired. See which reqs it applies to...
2086 list_for_each(tmp, &set->set_requests) {
2087 struct ptlrpc_request *req =
2088 list_entry(tmp, struct ptlrpc_request,
2091 /* don't expire request waiting for context */
2092 if (req->rq_wait_ctx)
2095 /* Request in-flight? */
2096 if (!((req->rq_phase == RQ_PHASE_RPC &&
2097 !req->rq_waiting && !req->rq_resend) ||
2098 (req->rq_phase == RQ_PHASE_BULK)))
2101 if (req->rq_timedout || /* already dealt with */
2102 req->rq_deadline > now) /* not expired */
2105 /* Deal with this guy. Do it asynchronously to not block
2106 * ptlrpcd thread. */
2107 ptlrpc_expire_one_request(req, 1);
2111 * When waiting for a whole set, we always break out of the
2112 * sleep so we can recalculate the timeout, or enable interrupts
2113 * if everyone's timed out.
2119 * Sets rq_intr flag in \a req under spinlock.
2121 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2123 spin_lock(&req->rq_lock);
2125 spin_unlock(&req->rq_lock);
2127 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2130 * Interrupts (sets interrupted flag) all uncompleted requests in
2131 * a set \a data. Callback for l_wait_event for interruptible waits.
2133 static void ptlrpc_interrupted_set(void *data)
2135 struct ptlrpc_request_set *set = data;
2136 struct list_head *tmp;
2138 LASSERT(set != NULL);
2139 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2141 list_for_each(tmp, &set->set_requests) {
2142 struct ptlrpc_request *req =
2143 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2145 if (req->rq_phase != RQ_PHASE_RPC &&
2146 req->rq_phase != RQ_PHASE_UNREGISTERING)
2149 ptlrpc_mark_interrupted(req);
2154 * Get the smallest timeout in the set; this does NOT set a timeout.
2156 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2158 struct list_head *tmp;
2159 time_t now = cfs_time_current_sec();
2161 struct ptlrpc_request *req;
2165 list_for_each(tmp, &set->set_requests) {
2166 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2169 * Request in-flight?
2171 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2172 (req->rq_phase == RQ_PHASE_BULK) ||
2173 (req->rq_phase == RQ_PHASE_NEW)))
2177 * Already timed out.
2179 if (req->rq_timedout)
2185 if (req->rq_wait_ctx)
2188 if (req->rq_phase == RQ_PHASE_NEW)
2189 deadline = req->rq_sent;
2190 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2191 deadline = req->rq_sent;
2193 deadline = req->rq_sent + req->rq_timeout;
2195 if (deadline <= now) /* actually expired already */
2196 timeout = 1; /* ASAP */
2197 else if (timeout == 0 || timeout > deadline - now)
2198 timeout = deadline - now;
2204 * Send all unset request from the set and then wait untill all
2205 * requests in the set complete (either get a reply, timeout, get an
2206 * error or otherwise be interrupted).
2207 * Returns 0 on success or error code otherwise.
2209 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2211 struct list_head *tmp;
2212 struct ptlrpc_request *req;
2213 struct l_wait_info lwi;
2217 if (set->set_producer)
2218 (void)ptlrpc_set_producer(set);
2220 list_for_each(tmp, &set->set_requests) {
2221 req = list_entry(tmp, struct ptlrpc_request,
2223 if (req->rq_phase == RQ_PHASE_NEW)
2224 (void)ptlrpc_send_new_req(req);
2227 if (list_empty(&set->set_requests))
2231 timeout = ptlrpc_set_next_timeout(set);
2233 /* wait until all complete, interrupted, or an in-flight
2235 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2238 if (timeout == 0 && !signal_pending(current))
2240 * No requests are in-flight (ether timed out
2241 * or delayed), so we can allow interrupts.
2242 * We still want to block for a limited time,
2243 * so we allow interrupts during the timeout.
2245 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2247 ptlrpc_interrupted_set, set);
2250 * At least one request is in flight, so no
2251 * interrupts are allowed. Wait until all
2252 * complete, or an in-flight req times out.
2254 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2255 ptlrpc_expired_set, set);
2257 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2259 /* LU-769 - if we ignored the signal because it was already
2260 * pending when we started, we need to handle it now or we risk
2261 * it being ignored forever */
2262 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2263 signal_pending(current)) {
2264 sigset_t blocked_sigs =
2265 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2267 /* In fact we only interrupt for the "fatal" signals
2268 * like SIGINT or SIGKILL. We still ignore less
2269 * important signals since ptlrpc set is not easily
2270 * reentrant from userspace again */
2271 if (signal_pending(current))
2272 ptlrpc_interrupted_set(set);
2273 cfs_restore_sigs(blocked_sigs);
2276 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2278 /* -EINTR => all requests have been flagged rq_intr so next
2280 * -ETIMEDOUT => someone timed out. When all reqs have
2281 * timed out, signals are enabled allowing completion with
2283 * I don't really care if we go once more round the loop in
2284 * the error cases -eeb. */
2285 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2286 list_for_each(tmp, &set->set_requests) {
2287 req = list_entry(tmp, struct ptlrpc_request,
2289 spin_lock(&req->rq_lock);
2290 req->rq_invalid_rqset = 1;
2291 spin_unlock(&req->rq_lock);
2294 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2296 LASSERT(atomic_read(&set->set_remaining) == 0);
2298 rc = set->set_rc; /* rq_status of already freed requests if any */
2299 list_for_each(tmp, &set->set_requests) {
2300 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2302 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2303 if (req->rq_status != 0)
2304 rc = req->rq_status;
2307 if (set->set_interpret != NULL) {
2308 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2310 rc = interpreter (set, set->set_arg, rc);
2312 struct ptlrpc_set_cbdata *cbdata, *n;
2315 list_for_each_entry_safe(cbdata, n,
2316 &set->set_cblist, psc_item) {
2317 list_del_init(&cbdata->psc_item);
2318 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2321 OBD_FREE_PTR(cbdata);
2327 EXPORT_SYMBOL(ptlrpc_set_wait);
2330 * Helper fuction for request freeing.
2331 * Called when request count reached zero and request needs to be freed.
2332 * Removes request from all sorts of sending/replay lists it might be on,
2333 * frees network buffers if any are present.
2334 * If \a locked is set, that means caller is already holding import imp_lock
2335 * and so we no longer need to reobtain it (for certain lists manipulations)
2337 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2341 if (request == NULL)
2344 LASSERT(!request->rq_srv_req);
2345 LASSERT(request->rq_export == NULL);
2346 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2347 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2348 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2349 LASSERTF(!request->rq_replay, "req %p\n", request);
2351 req_capsule_fini(&request->rq_pill);
2353 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2354 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2355 if (request->rq_import != NULL) {
2357 spin_lock(&request->rq_import->imp_lock);
2358 list_del_init(&request->rq_replay_list);
2360 spin_unlock(&request->rq_import->imp_lock);
2362 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2364 if (atomic_read(&request->rq_refcount) != 0) {
2365 DEBUG_REQ(D_ERROR, request,
2366 "freeing request with nonzero refcount");
2370 if (request->rq_repbuf != NULL)
2371 sptlrpc_cli_free_repbuf(request);
2373 if (request->rq_import != NULL) {
2374 class_import_put(request->rq_import);
2375 request->rq_import = NULL;
2377 if (request->rq_bulk != NULL)
2378 ptlrpc_free_bulk(request->rq_bulk);
2380 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2381 sptlrpc_cli_free_reqbuf(request);
2383 if (request->rq_cli_ctx)
2384 sptlrpc_req_put_ctx(request, !locked);
2386 if (request->rq_pool)
2387 __ptlrpc_free_req_to_pool(request);
2389 ptlrpc_request_cache_free(request);
2393 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2395 * Drop one request reference. Must be called with import imp_lock held.
2396 * When reference count drops to zero, request is freed.
2398 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2400 assert_spin_locked(&request->rq_import->imp_lock);
2401 (void)__ptlrpc_req_finished(request, 1);
2406 * Drops one reference count for request \a request.
2407 * \a locked set indicates that caller holds import imp_lock.
2408 * Frees the request whe reference count reaches zero.
2410 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2413 if (request == NULL)
2416 if (request == LP_POISON ||
2417 request->rq_reqmsg == LP_POISON) {
2418 CERROR("dereferencing freed request (bug 575)\n");
2423 DEBUG_REQ(D_INFO, request, "refcount now %u",
2424 atomic_read(&request->rq_refcount) - 1);
2426 if (atomic_dec_and_test(&request->rq_refcount)) {
2427 __ptlrpc_free_req(request, locked);
2435 * Drops one reference count for a request.
2437 void ptlrpc_req_finished(struct ptlrpc_request *request)
2439 __ptlrpc_req_finished(request, 0);
2441 EXPORT_SYMBOL(ptlrpc_req_finished);
2444 * Returns xid of a \a request
2446 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2448 return request->rq_xid;
2450 EXPORT_SYMBOL(ptlrpc_req_xid);
2453 * Disengage the client's reply buffer from the network
2454 * NB does _NOT_ unregister any client-side bulk.
2455 * IDEMPOTENT, but _not_ safe against concurrent callers.
2456 * The request owner (i.e. the thread doing the I/O) must call...
2457 * Returns 0 on success or 1 if unregistering cannot be made.
2459 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2462 struct l_wait_info lwi;
2467 LASSERT(!in_interrupt());
2470 * Let's setup deadline for reply unlink.
2472 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2473 async && request->rq_reply_deadline == 0)
2474 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2477 * Nothing left to do.
2479 if (!ptlrpc_client_recv_or_unlink(request))
2482 LNetMDUnlink(request->rq_reply_md_h);
2485 * Let's check it once again.
2487 if (!ptlrpc_client_recv_or_unlink(request))
2491 * Move to "Unregistering" phase as reply was not unlinked yet.
2493 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2496 * Do not wait for unlink to finish.
2502 * We have to l_wait_event() whatever the result, to give liblustre
2503 * a chance to run reply_in_callback(), and to make sure we've
2504 * unlinked before returning a req to the pool.
2507 /* The wq argument is ignored by user-space wait_event macros */
2508 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2509 &request->rq_set->set_waitq :
2510 &request->rq_reply_waitq;
2511 /* Network access will complete in finite time but the HUGE
2512 * timeout lets us CWARN for visibility of sluggish NALs */
2513 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2514 cfs_time_seconds(1), NULL, NULL);
2515 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2518 ptlrpc_rqphase_move(request, request->rq_next_phase);
2522 LASSERT(rc == -ETIMEDOUT);
2523 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2524 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2525 request->rq_receiving_reply,
2526 request->rq_req_unlinked,
2527 request->rq_reply_unlinked);
2532 static void ptlrpc_free_request(struct ptlrpc_request *req)
2534 spin_lock(&req->rq_lock);
2536 spin_unlock(&req->rq_lock);
2538 if (req->rq_commit_cb != NULL)
2539 req->rq_commit_cb(req);
2540 list_del_init(&req->rq_replay_list);
2542 __ptlrpc_req_finished(req, 1);
2546 * the request is committed and dropped from the replay list of its import
2548 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2550 struct obd_import *imp = req->rq_import;
2552 spin_lock(&imp->imp_lock);
2553 if (list_empty(&req->rq_replay_list)) {
2554 spin_unlock(&imp->imp_lock);
2558 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2559 ptlrpc_free_request(req);
2561 spin_unlock(&imp->imp_lock);
2563 EXPORT_SYMBOL(ptlrpc_request_committed);
2566 * Iterates through replay_list on import and prunes
2567 * all requests have transno smaller than last_committed for the
2568 * import and don't have rq_replay set.
2569 * Since requests are sorted in transno order, stops when meetign first
2570 * transno bigger than last_committed.
2571 * caller must hold imp->imp_lock
2573 void ptlrpc_free_committed(struct obd_import *imp)
2575 struct ptlrpc_request *req, *saved;
2576 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2577 bool skip_committed_list = true;
2580 LASSERT(imp != NULL);
2581 assert_spin_locked(&imp->imp_lock);
2583 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2584 imp->imp_generation == imp->imp_last_generation_checked) {
2585 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2586 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2589 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2590 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2591 imp->imp_generation);
2593 if (imp->imp_generation != imp->imp_last_generation_checked ||
2594 imp->imp_last_transno_checked == 0)
2595 skip_committed_list = false;
2597 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2598 imp->imp_last_generation_checked = imp->imp_generation;
2600 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2602 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2603 LASSERT(req != last_req);
2606 if (req->rq_transno == 0) {
2607 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2610 if (req->rq_import_generation < imp->imp_generation) {
2611 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2615 /* not yet committed */
2616 if (req->rq_transno > imp->imp_peer_committed_transno) {
2617 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2621 if (req->rq_replay) {
2622 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2623 list_move_tail(&req->rq_replay_list,
2624 &imp->imp_committed_list);
2628 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2629 imp->imp_peer_committed_transno);
2631 ptlrpc_free_request(req);
2634 if (skip_committed_list)
2637 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2639 LASSERT(req->rq_transno != 0);
2640 if (req->rq_import_generation < imp->imp_generation) {
2641 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2642 ptlrpc_free_request(req);
2643 } else if (!req->rq_replay) {
2644 DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2645 ptlrpc_free_request(req);
2652 void ptlrpc_cleanup_client(struct obd_import *imp)
2659 * Schedule previously sent request for resend.
2660 * For bulk requests we assign new xid (to avoid problems with
2661 * lost replies and therefore several transfers landing into same buffer
2662 * from different sending attempts).
2664 void ptlrpc_resend_req(struct ptlrpc_request *req)
2666 DEBUG_REQ(D_HA, req, "going to resend");
2667 spin_lock(&req->rq_lock);
2669 /* Request got reply but linked to the import list still.
2670 Let ptlrpc_check_set() to process it. */
2671 if (ptlrpc_client_replied(req)) {
2672 spin_unlock(&req->rq_lock);
2673 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2677 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2678 req->rq_status = -EAGAIN;
2681 req->rq_net_err = 0;
2682 req->rq_timedout = 0;
2684 ptlrpc_client_wake_req(req);
2685 spin_unlock(&req->rq_lock);
2688 /* XXX: this function and rq_status are currently unused */
2689 void ptlrpc_restart_req(struct ptlrpc_request *req)
2691 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2692 req->rq_status = -ERESTARTSYS;
2694 spin_lock(&req->rq_lock);
2695 req->rq_restart = 1;
2696 req->rq_timedout = 0;
2697 ptlrpc_client_wake_req(req);
2698 spin_unlock(&req->rq_lock);
2702 * Grab additional reference on a request \a req
2704 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2707 atomic_inc(&req->rq_refcount);
2710 EXPORT_SYMBOL(ptlrpc_request_addref);
2713 * Add a request to import replay_list.
2714 * Must be called under imp_lock
2716 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2717 struct obd_import *imp)
2719 struct list_head *tmp;
2721 assert_spin_locked(&imp->imp_lock);
2723 if (req->rq_transno == 0) {
2724 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2728 /* clear this for new requests that were resent as well
2729 as resent replayed requests. */
2730 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2732 /* don't re-add requests that have been replayed */
2733 if (!list_empty(&req->rq_replay_list))
2736 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2738 LASSERT(imp->imp_replayable);
2739 /* Balanced in ptlrpc_free_committed, usually. */
2740 ptlrpc_request_addref(req);
2741 list_for_each_prev(tmp, &imp->imp_replay_list) {
2742 struct ptlrpc_request *iter = list_entry(tmp,
2743 struct ptlrpc_request,
2746 /* We may have duplicate transnos if we create and then
2747 * open a file, or for closes retained if to match creating
2748 * opens, so use req->rq_xid as a secondary key.
2749 * (See bugs 684, 685, and 428.)
2750 * XXX no longer needed, but all opens need transnos!
2752 if (iter->rq_transno > req->rq_transno)
2755 if (iter->rq_transno == req->rq_transno) {
2756 LASSERT(iter->rq_xid != req->rq_xid);
2757 if (iter->rq_xid > req->rq_xid)
2761 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2765 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2769 * Send request and wait until it completes.
2770 * Returns request processing status.
2772 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2774 struct ptlrpc_request_set *set;
2778 LASSERT(req->rq_set == NULL);
2779 LASSERT(!req->rq_receiving_reply);
2781 set = ptlrpc_prep_set();
2783 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2787 /* for distributed debugging */
2788 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2790 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2791 ptlrpc_request_addref(req);
2792 ptlrpc_set_add_req(set, req);
2793 rc = ptlrpc_set_wait(set);
2794 ptlrpc_set_destroy(set);
2798 EXPORT_SYMBOL(ptlrpc_queue_wait);
2801 * Callback used for replayed requests reply processing.
2802 * In case of successful reply calls registered request replay callback.
2803 * In case of error restart replay process.
2805 static int ptlrpc_replay_interpret(const struct lu_env *env,
2806 struct ptlrpc_request *req,
2807 void * data, int rc)
2809 struct ptlrpc_replay_async_args *aa = data;
2810 struct obd_import *imp = req->rq_import;
2813 atomic_dec(&imp->imp_replay_inflight);
2815 if (!ptlrpc_client_replied(req)) {
2816 CERROR("request replay timed out, restarting recovery\n");
2817 GOTO(out, rc = -ETIMEDOUT);
2820 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2821 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2822 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2823 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2825 /** VBR: check version failure */
2826 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2827 /** replay was failed due to version mismatch */
2828 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2829 spin_lock(&imp->imp_lock);
2830 imp->imp_vbr_failed = 1;
2831 imp->imp_no_lock_replay = 1;
2832 spin_unlock(&imp->imp_lock);
2833 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2835 /** The transno had better not change over replay. */
2836 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2837 lustre_msg_get_transno(req->rq_repmsg) ||
2838 lustre_msg_get_transno(req->rq_repmsg) == 0,
2840 lustre_msg_get_transno(req->rq_reqmsg),
2841 lustre_msg_get_transno(req->rq_repmsg));
2844 spin_lock(&imp->imp_lock);
2845 /** if replays by version then gap occur on server, no trust to locks */
2846 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2847 imp->imp_no_lock_replay = 1;
2848 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2849 spin_unlock(&imp->imp_lock);
2850 LASSERT(imp->imp_last_replay_transno);
2852 /* transaction number shouldn't be bigger than the latest replayed */
2853 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2854 DEBUG_REQ(D_ERROR, req,
2855 "Reported transno "LPU64" is bigger than the "
2856 "replayed one: "LPU64, req->rq_transno,
2857 lustre_msg_get_transno(req->rq_reqmsg));
2858 GOTO(out, rc = -EINVAL);
2861 DEBUG_REQ(D_HA, req, "got rep");
2863 /* let the callback do fixups, possibly including in the request */
2864 if (req->rq_replay_cb)
2865 req->rq_replay_cb(req);
2867 if (ptlrpc_client_replied(req) &&
2868 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2869 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2870 lustre_msg_get_status(req->rq_repmsg),
2871 aa->praa_old_status);
2873 /* Put it back for re-replay. */
2874 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2878 * Errors while replay can set transno to 0, but
2879 * imp_last_replay_transno shouldn't be set to 0 anyway
2881 if (req->rq_transno == 0)
2882 CERROR("Transno is 0 during replay!\n");
2884 /* continue with recovery */
2885 rc = ptlrpc_import_recovery_state_machine(imp);
2887 req->rq_send_state = aa->praa_old_state;
2890 /* this replay failed, so restart recovery */
2891 ptlrpc_connect_import(imp);
2897 * Prepares and queues request for replay.
2898 * Adds it to ptlrpcd queue for actual sending.
2899 * Returns 0 on success.
2901 int ptlrpc_replay_req(struct ptlrpc_request *req)
2903 struct ptlrpc_replay_async_args *aa;
2906 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2908 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2909 aa = ptlrpc_req_async_args(req);
2910 memset(aa, 0, sizeof *aa);
2912 /* Prepare request to be resent with ptlrpcd */
2913 aa->praa_old_state = req->rq_send_state;
2914 req->rq_send_state = LUSTRE_IMP_REPLAY;
2915 req->rq_phase = RQ_PHASE_NEW;
2916 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2918 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2920 req->rq_interpret_reply = ptlrpc_replay_interpret;
2921 /* Readjust the timeout for current conditions */
2922 ptlrpc_at_set_req_timeout(req);
2924 /* Tell server the net_latency, so the server can calculate how long
2925 * it should wait for next replay */
2926 lustre_msg_set_service_time(req->rq_reqmsg,
2927 ptlrpc_at_get_net_latency(req));
2928 DEBUG_REQ(D_HA, req, "REPLAY");
2930 atomic_inc(&req->rq_import->imp_replay_inflight);
2931 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2933 ptlrpcd_add_req(req);
2938 * Aborts all in-flight request on import \a imp sending and delayed lists
2940 void ptlrpc_abort_inflight(struct obd_import *imp)
2942 struct list_head *tmp, *n;
2945 /* Make sure that no new requests get processed for this import.
2946 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2947 * this flag and then putting requests on sending_list or delayed_list.
2949 spin_lock(&imp->imp_lock);
2951 /* XXX locking? Maybe we should remove each request with the list
2952 * locked? Also, how do we know if the requests on the list are
2953 * being freed at this time?
2955 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2956 struct ptlrpc_request *req = list_entry(tmp,
2957 struct ptlrpc_request,
2960 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2962 spin_lock(&req->rq_lock);
2963 if (req->rq_import_generation < imp->imp_generation) {
2965 req->rq_status = -EIO;
2966 ptlrpc_client_wake_req(req);
2968 spin_unlock(&req->rq_lock);
2971 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2972 struct ptlrpc_request *req =
2973 list_entry(tmp, struct ptlrpc_request, rq_list);
2975 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2977 spin_lock(&req->rq_lock);
2978 if (req->rq_import_generation < imp->imp_generation) {
2980 req->rq_status = -EIO;
2981 ptlrpc_client_wake_req(req);
2983 spin_unlock(&req->rq_lock);
2986 /* Last chance to free reqs left on the replay list, but we
2987 * will still leak reqs that haven't committed. */
2988 if (imp->imp_replayable)
2989 ptlrpc_free_committed(imp);
2991 spin_unlock(&imp->imp_lock);
2997 * Abort all uncompleted requests in request set \a set
2999 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3001 struct list_head *tmp, *pos;
3003 LASSERT(set != NULL);
3005 list_for_each_safe(pos, tmp, &set->set_requests) {
3006 struct ptlrpc_request *req =
3007 list_entry(pos, struct ptlrpc_request,
3010 spin_lock(&req->rq_lock);
3011 if (req->rq_phase != RQ_PHASE_RPC) {
3012 spin_unlock(&req->rq_lock);
3017 req->rq_status = -EINTR;
3018 ptlrpc_client_wake_req(req);
3019 spin_unlock(&req->rq_lock);
3023 static __u64 ptlrpc_last_xid;
3024 static spinlock_t ptlrpc_last_xid_lock;
3027 * Initialize the XID for the node. This is common among all requests on
3028 * this node, and only requires the property that it is monotonically
3029 * increasing. It does not need to be sequential. Since this is also used
3030 * as the RDMA match bits, it is important that a single client NOT have
3031 * the same match bits for two different in-flight requests, hence we do
3032 * NOT want to have an XID per target or similar.
3034 * To avoid an unlikely collision between match bits after a client reboot
3035 * (which would deliver old data into the wrong RDMA buffer) initialize
3036 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3037 * If the time is clearly incorrect, we instead use a 62-bit random number.
3038 * In the worst case the random number will overflow 1M RPCs per second in
3039 * 9133 years, or permutations thereof.
3041 #define YEAR_2004 (1ULL << 30)
3042 void ptlrpc_init_xid(void)
3044 time_t now = cfs_time_current_sec();
3046 spin_lock_init(&ptlrpc_last_xid_lock);
3047 if (now < YEAR_2004) {
3048 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3049 ptlrpc_last_xid >>= 2;
3050 ptlrpc_last_xid |= (1ULL << 61);
3052 ptlrpc_last_xid = (__u64)now << 20;
3055 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3056 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3057 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3061 * Increase xid and returns resulting new value to the caller.
3063 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3064 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3065 * itself uses the last bulk xid needed, so the server can determine the
3066 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3067 * xid must align to a power-of-two value.
3069 * This is assumed to be true due to the initial ptlrpc_last_xid
3070 * value also being initialized to a power-of-two value. LU-1431
3072 __u64 ptlrpc_next_xid(void)
3076 spin_lock(&ptlrpc_last_xid_lock);
3077 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3078 ptlrpc_last_xid = next;
3079 spin_unlock(&ptlrpc_last_xid_lock);
3085 * If request has a new allocated XID (new request or EINPROGRESS resend),
3086 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3087 * request to ensure previous bulk fails and avoid problems with lost replies
3088 * and therefore several transfers landing into the same buffer from different
3091 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3093 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3095 LASSERT(bd != NULL);
3097 if (!req->rq_resend || req->rq_nr_resend != 0) {
3098 /* this request has a new xid, just use it as bulk matchbits */
3099 req->rq_mbits = req->rq_xid;
3101 } else { /* needs to generate a new matchbits for resend */
3102 __u64 old_mbits = req->rq_mbits;
3104 if ((bd->bd_import->imp_connect_data.ocd_connect_flags &
3105 OBD_CONNECT_BULK_MBITS) != 0)
3106 req->rq_mbits = ptlrpc_next_xid();
3107 else /* old version transfers rq_xid to peer as matchbits */
3108 req->rq_mbits = req->rq_xid = ptlrpc_next_xid();
3110 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
3111 old_mbits, req->rq_mbits);
3114 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3115 * that server can infer the number of bulks that were prepared,
3117 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3122 * Get a glimpse at what next xid value might have been.
3123 * Returns possible next xid.
3125 __u64 ptlrpc_sample_next_xid(void)
3127 #if BITS_PER_LONG == 32
3128 /* need to avoid possible word tearing on 32-bit systems */
3131 spin_lock(&ptlrpc_last_xid_lock);
3132 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3133 spin_unlock(&ptlrpc_last_xid_lock);
3137 /* No need to lock, since returned value is racy anyways */
3138 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3141 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3144 * Functions for operating ptlrpc workers.
3146 * A ptlrpc work is a function which will be running inside ptlrpc context.
3147 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3149 * 1. after a work is created, it can be used many times, that is:
3150 * handler = ptlrpcd_alloc_work();
3151 * ptlrpcd_queue_work();
3153 * queue it again when necessary:
3154 * ptlrpcd_queue_work();
3155 * ptlrpcd_destroy_work();
3156 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3157 * it will only be queued once in any time. Also as its name implies, it may
3158 * have delay before it really runs by ptlrpcd thread.
3160 struct ptlrpc_work_async_args {
3161 int (*cb)(const struct lu_env *, void *);
3165 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3167 /* re-initialize the req */
3168 req->rq_timeout = obd_timeout;
3169 req->rq_sent = cfs_time_current_sec();
3170 req->rq_deadline = req->rq_sent + req->rq_timeout;
3171 req->rq_reply_deadline = req->rq_deadline;
3172 req->rq_phase = RQ_PHASE_INTERPRET;
3173 req->rq_next_phase = RQ_PHASE_COMPLETE;
3174 req->rq_xid = ptlrpc_next_xid();
3175 req->rq_import_generation = req->rq_import->imp_generation;
3177 ptlrpcd_add_req(req);
3180 static int work_interpreter(const struct lu_env *env,
3181 struct ptlrpc_request *req, void *data, int rc)
3183 struct ptlrpc_work_async_args *arg = data;
3185 LASSERT(ptlrpcd_check_work(req));
3186 LASSERT(arg->cb != NULL);
3188 rc = arg->cb(env, arg->cbdata);
3190 list_del_init(&req->rq_set_chain);
3193 if (atomic_dec_return(&req->rq_refcount) > 1) {
3194 atomic_set(&req->rq_refcount, 2);
3195 ptlrpcd_add_work_req(req);
3200 static int worker_format;
3202 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3204 return req->rq_pill.rc_fmt == (void *)&worker_format;
3208 * Create a work for ptlrpc.
3210 void *ptlrpcd_alloc_work(struct obd_import *imp,
3211 int (*cb)(const struct lu_env *, void *), void *cbdata)
3213 struct ptlrpc_request *req = NULL;
3214 struct ptlrpc_work_async_args *args;
3220 RETURN(ERR_PTR(-EINVAL));
3222 /* copy some code from deprecated fakereq. */
3223 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3225 CERROR("ptlrpc: run out of memory!\n");
3226 RETURN(ERR_PTR(-ENOMEM));
3229 ptlrpc_cli_req_init(req);
3231 req->rq_send_state = LUSTRE_IMP_FULL;
3232 req->rq_type = PTL_RPC_MSG_REQUEST;
3233 req->rq_import = class_import_get(imp);
3234 req->rq_interpret_reply = work_interpreter;
3235 /* don't want reply */
3236 req->rq_no_delay = req->rq_no_resend = 1;
3237 req->rq_pill.rc_fmt = (void *)&worker_format;
3239 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3240 args = ptlrpc_req_async_args(req);
3242 args->cbdata = cbdata;
3246 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3248 void ptlrpcd_destroy_work(void *handler)
3250 struct ptlrpc_request *req = handler;
3253 ptlrpc_req_finished(req);
3255 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3257 int ptlrpcd_queue_work(void *handler)
3259 struct ptlrpc_request *req = handler;
3262 * Check if the req is already being queued.
3264 * Here comes a trick: it lacks a way of checking if a req is being
3265 * processed reliably in ptlrpc. Here I have to use refcount of req
3266 * for this purpose. This is okay because the caller should use this
3267 * req as opaque data. - Jinshan
3269 LASSERT(atomic_read(&req->rq_refcount) > 0);
3270 if (atomic_inc_return(&req->rq_refcount) == 2)
3271 ptlrpcd_add_work_req(req);
3274 EXPORT_SYMBOL(ptlrpcd_queue_work);