4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2015, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 /** Implementation of client-side PortalRPC interfaces */
39 #define DEBUG_SUBSYSTEM S_RPC
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
48 #include "ptlrpc_internal.h"
50 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
51 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
52 .release_frags = ptlrpc_release_bulk_page_pin,
54 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
56 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
57 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
58 .release_frags = ptlrpc_release_bulk_noop,
60 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
62 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
63 .add_iov_frag = ptlrpc_prep_bulk_frag,
65 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
67 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
68 static int ptlrpcd_check_work(struct ptlrpc_request *req);
69 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
72 * Initialize passed in client structure \a cl.
74 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
75 struct ptlrpc_client *cl)
77 cl->cli_request_portal = req_portal;
78 cl->cli_reply_portal = rep_portal;
81 EXPORT_SYMBOL(ptlrpc_init_client);
84 * Return PortalRPC connection for remore uud \a uuid
86 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
88 struct ptlrpc_connection *c;
90 lnet_process_id_t peer;
93 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
94 * before accessing its values. */
95 /* coverity[uninit_use_in_call] */
96 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
98 CNETERR("cannot find peer %s!\n", uuid->uuid);
102 c = ptlrpc_connection_get(peer, self, uuid);
104 memcpy(c->c_remote_uuid.uuid,
105 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
108 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
114 * Allocate and initialize new bulk descriptor on the sender.
115 * Returns pointer to the descriptor or NULL on error.
117 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
118 enum ptlrpc_bulk_op_type type,
120 const struct ptlrpc_bulk_frag_ops *ops)
122 struct ptlrpc_bulk_desc *desc;
125 /* ensure that only one of KIOV or IOVEC is set but not both */
126 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
127 ops->add_kiov_frag != NULL) ||
128 (ptlrpc_is_bulk_desc_kvec(type) &&
129 ops->add_iov_frag != NULL));
131 if (type & PTLRPC_BULK_BUF_KIOV) {
133 offsetof(struct ptlrpc_bulk_desc,
134 bd_u.bd_kiov.bd_vec[nfrags]));
137 offsetof(struct ptlrpc_bulk_desc,
138 bd_u.bd_kvec.bd_kvec[nfrags]));
144 spin_lock_init(&desc->bd_lock);
145 init_waitqueue_head(&desc->bd_waitq);
146 desc->bd_max_iov = nfrags;
147 desc->bd_iov_count = 0;
148 desc->bd_portal = portal;
149 desc->bd_type = type;
150 desc->bd_md_count = 0;
151 desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
152 LASSERT(max_brw > 0);
153 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
154 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
155 * node. Negotiated ocd_brw_size will always be <= this number. */
156 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
157 LNetInvalidateHandle(&desc->bd_mds[i]);
163 * Prepare bulk descriptor for specified outgoing request \a req that
164 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
165 * the bulk to be sent. Used on client-side.
166 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
169 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
170 unsigned nfrags, unsigned max_brw,
173 const struct ptlrpc_bulk_frag_ops
176 struct obd_import *imp = req->rq_import;
177 struct ptlrpc_bulk_desc *desc;
180 LASSERT(ptlrpc_is_bulk_op_passive(type));
182 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
186 desc->bd_import_generation = req->rq_import_generation;
187 desc->bd_import = class_import_get(imp);
190 desc->bd_cbid.cbid_fn = client_bulk_callback;
191 desc->bd_cbid.cbid_arg = desc;
193 /* This makes req own desc, and free it when she frees herself */
198 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
200 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
201 struct page *page, int pageoffset, int len,
206 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
207 LASSERT(page != NULL);
208 LASSERT(pageoffset >= 0);
210 LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
211 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
213 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
218 page_cache_get(page);
220 kiov->kiov_page = page;
221 kiov->kiov_offset = pageoffset;
222 kiov->kiov_len = len;
224 desc->bd_iov_count++;
226 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
228 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
234 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
235 LASSERT(frag != NULL);
237 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
239 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
243 iovec->iov_base = frag;
244 iovec->iov_len = len;
246 desc->bd_iov_count++;
248 RETURN(desc->bd_nob);
250 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
252 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
256 LASSERT(desc != NULL);
257 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
258 LASSERT(desc->bd_md_count == 0); /* network hands off */
259 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
260 LASSERT(desc->bd_frag_ops != NULL);
262 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
263 sptlrpc_enc_pool_put_pages(desc);
266 class_export_put(desc->bd_export);
268 class_import_put(desc->bd_import);
270 if (desc->bd_frag_ops->release_frags != NULL)
271 desc->bd_frag_ops->release_frags(desc);
273 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
274 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
275 bd_u.bd_kiov.bd_vec[desc->bd_max_iov]));
277 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
278 bd_u.bd_kvec.bd_kvec[desc->
283 EXPORT_SYMBOL(ptlrpc_free_bulk);
286 * Set server timelimit for this req, i.e. how long are we willing to wait
287 * for reply before timing out this request.
289 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
295 LASSERT(req->rq_import);
298 /* non-AT settings */
300 * \a imp_server_timeout means this is reverse import and
301 * we send (currently only) ASTs to the client and cannot afford
302 * to wait too long for the reply, otherwise the other client
303 * (because of which we are sending this request) would
304 * timeout waiting for us
306 req->rq_timeout = req->rq_import->imp_server_timeout ?
307 obd_timeout / 2 : obd_timeout;
309 at = &req->rq_import->imp_at;
310 idx = import_at_get_index(req->rq_import,
311 req->rq_request_portal);
312 serv_est = at_get(&at->iat_service_estimate[idx]);
313 req->rq_timeout = at_est2timeout(serv_est);
315 /* We could get even fancier here, using history to predict increased
318 /* Let the server know what this RPC timeout is by putting it in the
320 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
322 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
324 /* Adjust max service estimate based on server value */
325 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
326 unsigned int serv_est)
332 LASSERT(req->rq_import);
333 at = &req->rq_import->imp_at;
335 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
336 /* max service estimates are tracked on the server side,
337 so just keep minimal history here */
338 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
340 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
341 "has changed from %d to %d\n",
342 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
343 oldse, at_get(&at->iat_service_estimate[idx]));
346 /* Expected network latency per remote node (secs) */
347 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
349 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
352 /* Adjust expected network latency */
353 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
354 unsigned int service_time)
356 unsigned int nl, oldnl;
358 time_t now = cfs_time_current_sec();
360 LASSERT(req->rq_import);
362 if (service_time > now - req->rq_sent + 3) {
363 /* bz16408, however, this can also happen if early reply
364 * is lost and client RPC is expired and resent, early reply
365 * or reply of original RPC can still be fit in reply buffer
366 * of resent RPC, now client is measuring time from the
367 * resent time, but server sent back service time of original
370 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
371 D_ADAPTTO : D_WARNING,
372 "Reported service time %u > total measured time "
373 CFS_DURATION_T"\n", service_time,
374 cfs_time_sub(now, req->rq_sent));
378 /* Network latency is total time less server processing time */
379 nl = max_t(int, now - req->rq_sent -
380 service_time, 0) + 1; /* st rounding */
381 at = &req->rq_import->imp_at;
383 oldnl = at_measured(&at->iat_net_latency, nl);
385 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
386 "has changed from %d to %d\n",
387 req->rq_import->imp_obd->obd_name,
389 &req->rq_import->imp_connection->c_remote_uuid),
390 oldnl, at_get(&at->iat_net_latency));
393 static int unpack_reply(struct ptlrpc_request *req)
397 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
398 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
400 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
405 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
407 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
414 * Handle an early reply message, called with the rq_lock held.
415 * If anything goes wrong just ignore it - same as if it never happened
417 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
418 __must_hold(&req->rq_lock)
420 struct ptlrpc_request *early_req;
426 spin_unlock(&req->rq_lock);
428 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
430 spin_lock(&req->rq_lock);
434 rc = unpack_reply(early_req);
436 sptlrpc_cli_finish_early_reply(early_req);
437 spin_lock(&req->rq_lock);
441 /* Use new timeout value just to adjust the local value for this
442 * request, don't include it into at_history. It is unclear yet why
443 * service time increased and should it be counted or skipped, e.g.
444 * that can be recovery case or some error or server, the real reply
445 * will add all new data if it is worth to add. */
446 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
447 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
449 /* Network latency can be adjusted, it is pure network delays */
450 ptlrpc_at_adj_net_latency(req,
451 lustre_msg_get_service_time(early_req->rq_repmsg));
453 sptlrpc_cli_finish_early_reply(early_req);
455 spin_lock(&req->rq_lock);
456 olddl = req->rq_deadline;
457 /* server assumes it now has rq_timeout from when the request
458 * arrived, so the client should give it at least that long.
459 * since we don't know the arrival time we'll use the original
461 req->rq_deadline = req->rq_sent + req->rq_timeout +
462 ptlrpc_at_get_net_latency(req);
464 DEBUG_REQ(D_ADAPTTO, req,
465 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
466 "("CFS_DURATION_T"s)", req->rq_early_count,
467 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
468 cfs_time_sub(req->rq_deadline, olddl));
473 static struct kmem_cache *request_cache;
475 int ptlrpc_request_cache_init(void)
477 request_cache = kmem_cache_create("ptlrpc_cache",
478 sizeof(struct ptlrpc_request),
479 0, SLAB_HWCACHE_ALIGN, NULL);
480 return request_cache == NULL ? -ENOMEM : 0;
483 void ptlrpc_request_cache_fini(void)
485 kmem_cache_destroy(request_cache);
488 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
490 struct ptlrpc_request *req;
492 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
496 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
498 OBD_SLAB_FREE_PTR(req, request_cache);
502 * Wind down request pool \a pool.
503 * Frees all requests from the pool too
505 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
507 struct list_head *l, *tmp;
508 struct ptlrpc_request *req;
510 LASSERT(pool != NULL);
512 spin_lock(&pool->prp_lock);
513 list_for_each_safe(l, tmp, &pool->prp_req_list) {
514 req = list_entry(l, struct ptlrpc_request, rq_list);
515 list_del(&req->rq_list);
516 LASSERT(req->rq_reqbuf);
517 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
518 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
519 ptlrpc_request_cache_free(req);
521 spin_unlock(&pool->prp_lock);
522 OBD_FREE(pool, sizeof(*pool));
524 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
527 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
529 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
534 while (size < pool->prp_rq_size)
537 LASSERTF(list_empty(&pool->prp_req_list) ||
538 size == pool->prp_rq_size,
539 "Trying to change pool size with nonempty pool "
540 "from %d to %d bytes\n", pool->prp_rq_size, size);
542 spin_lock(&pool->prp_lock);
543 pool->prp_rq_size = size;
544 for (i = 0; i < num_rq; i++) {
545 struct ptlrpc_request *req;
546 struct lustre_msg *msg;
548 spin_unlock(&pool->prp_lock);
549 req = ptlrpc_request_cache_alloc(GFP_NOFS);
552 OBD_ALLOC_LARGE(msg, size);
554 ptlrpc_request_cache_free(req);
557 req->rq_reqbuf = msg;
558 req->rq_reqbuf_len = size;
560 spin_lock(&pool->prp_lock);
561 list_add_tail(&req->rq_list, &pool->prp_req_list);
563 spin_unlock(&pool->prp_lock);
566 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
569 * Create and initialize new request pool with given attributes:
570 * \a num_rq - initial number of requests to create for the pool
571 * \a msgsize - maximum message size possible for requests in thid pool
572 * \a populate_pool - function to be called when more requests need to be added
574 * Returns pointer to newly created pool or NULL on error.
576 struct ptlrpc_request_pool *
577 ptlrpc_init_rq_pool(int num_rq, int msgsize,
578 int (*populate_pool)(struct ptlrpc_request_pool *, int))
580 struct ptlrpc_request_pool *pool;
582 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
586 /* Request next power of two for the allocation, because internally
587 kernel would do exactly this */
589 spin_lock_init(&pool->prp_lock);
590 INIT_LIST_HEAD(&pool->prp_req_list);
591 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
592 pool->prp_populate = populate_pool;
594 populate_pool(pool, num_rq);
598 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
601 * Fetches one request from pool \a pool
603 static struct ptlrpc_request *
604 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
606 struct ptlrpc_request *request;
607 struct lustre_msg *reqbuf;
612 spin_lock(&pool->prp_lock);
614 /* See if we have anything in a pool, and bail out if nothing,
615 * in writeout path, where this matters, this is safe to do, because
616 * nothing is lost in this case, and when some in-flight requests
617 * complete, this code will be called again. */
618 if (unlikely(list_empty(&pool->prp_req_list))) {
619 spin_unlock(&pool->prp_lock);
623 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
625 list_del_init(&request->rq_list);
626 spin_unlock(&pool->prp_lock);
628 LASSERT(request->rq_reqbuf);
629 LASSERT(request->rq_pool);
631 reqbuf = request->rq_reqbuf;
632 memset(request, 0, sizeof(*request));
633 request->rq_reqbuf = reqbuf;
634 request->rq_reqbuf_len = pool->prp_rq_size;
635 request->rq_pool = pool;
641 * Returns freed \a request to pool.
643 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
645 struct ptlrpc_request_pool *pool = request->rq_pool;
647 spin_lock(&pool->prp_lock);
648 LASSERT(list_empty(&request->rq_list));
649 LASSERT(!request->rq_receiving_reply);
650 list_add_tail(&request->rq_list, &pool->prp_req_list);
651 spin_unlock(&pool->prp_lock);
654 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
655 __u32 version, int opcode, char **bufs,
656 struct ptlrpc_cli_ctx *ctx)
659 struct obd_import *imp;
665 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
666 imp = request->rq_import;
667 lengths = request->rq_pill.rc_area[RCL_CLIENT];
670 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
672 rc = sptlrpc_req_get_ctx(request);
676 sptlrpc_req_set_flavor(request, opcode);
678 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
683 lustre_msg_add_version(request->rq_reqmsg, version);
684 request->rq_send_state = LUSTRE_IMP_FULL;
685 request->rq_type = PTL_RPC_MSG_REQUEST;
687 request->rq_req_cbid.cbid_fn = request_out_callback;
688 request->rq_req_cbid.cbid_arg = request;
690 request->rq_reply_cbid.cbid_fn = reply_in_callback;
691 request->rq_reply_cbid.cbid_arg = request;
693 request->rq_reply_deadline = 0;
694 request->rq_phase = RQ_PHASE_NEW;
695 request->rq_next_phase = RQ_PHASE_UNDEFINED;
697 request->rq_request_portal = imp->imp_client->cli_request_portal;
698 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
700 ptlrpc_at_set_req_timeout(request);
702 lustre_msg_set_opc(request->rq_reqmsg, opcode);
707 LASSERT(!request->rq_pool);
708 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
710 class_import_put(imp);
715 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
718 * Pack request buffers for network transfer, performing necessary encryption
719 * steps if necessary.
721 int ptlrpc_request_pack(struct ptlrpc_request *request,
722 __u32 version, int opcode)
725 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
729 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
730 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
731 * have to send old ptlrpc_body to keep interoprability with these
734 * Only three kinds of server->client RPCs so far:
739 * XXX This should be removed whenever we drop the interoprability with
740 * the these old clients.
742 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
743 opcode == LDLM_GL_CALLBACK)
744 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
745 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
749 EXPORT_SYMBOL(ptlrpc_request_pack);
752 * Helper function to allocate new request on import \a imp
753 * and possibly using existing request from pool \a pool if provided.
754 * Returns allocated request structure with import field filled or
758 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
759 struct ptlrpc_request_pool *pool)
761 struct ptlrpc_request *request = NULL;
763 request = ptlrpc_request_cache_alloc(GFP_NOFS);
765 if (!request && pool)
766 request = ptlrpc_prep_req_from_pool(pool);
769 ptlrpc_cli_req_init(request);
771 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
772 LASSERT(imp != LP_POISON);
773 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
775 LASSERT(imp->imp_client != LP_POISON);
777 request->rq_import = class_import_get(imp);
779 CERROR("request allocation out of memory\n");
786 * Helper function for creating a request.
787 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
788 * buffer structures according to capsule template \a format.
789 * Returns allocated request structure pointer or NULL on error.
791 static struct ptlrpc_request *
792 ptlrpc_request_alloc_internal(struct obd_import *imp,
793 struct ptlrpc_request_pool * pool,
794 const struct req_format *format)
796 struct ptlrpc_request *request;
798 request = __ptlrpc_request_alloc(imp, pool);
802 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
803 req_capsule_set(&request->rq_pill, format);
808 * Allocate new request structure for import \a imp and initialize its
809 * buffer structure according to capsule template \a format.
811 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
812 const struct req_format *format)
814 return ptlrpc_request_alloc_internal(imp, NULL, format);
816 EXPORT_SYMBOL(ptlrpc_request_alloc);
819 * Allocate new request structure for import \a imp from pool \a pool and
820 * initialize its buffer structure according to capsule template \a format.
822 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
823 struct ptlrpc_request_pool * pool,
824 const struct req_format *format)
826 return ptlrpc_request_alloc_internal(imp, pool, format);
828 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
831 * For requests not from pool, free memory of the request structure.
832 * For requests obtained from a pool earlier, return request back to pool.
834 void ptlrpc_request_free(struct ptlrpc_request *request)
836 if (request->rq_pool)
837 __ptlrpc_free_req_to_pool(request);
839 ptlrpc_request_cache_free(request);
841 EXPORT_SYMBOL(ptlrpc_request_free);
844 * Allocate new request for operatione \a opcode and immediatelly pack it for
846 * Only used for simple requests like OBD_PING where the only important
847 * part of the request is operation itself.
848 * Returns allocated request or NULL on error.
850 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
851 const struct req_format *format,
852 __u32 version, int opcode)
854 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
858 rc = ptlrpc_request_pack(req, version, opcode);
860 ptlrpc_request_free(req);
866 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
869 * Allocate and initialize new request set structure on the current CPT.
870 * Returns a pointer to the newly allocated set structure or NULL on error.
872 struct ptlrpc_request_set *ptlrpc_prep_set(void)
874 struct ptlrpc_request_set *set;
878 cpt = cfs_cpt_current(cfs_cpt_table, 0);
879 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
882 atomic_set(&set->set_refcount, 1);
883 INIT_LIST_HEAD(&set->set_requests);
884 init_waitqueue_head(&set->set_waitq);
885 atomic_set(&set->set_new_count, 0);
886 atomic_set(&set->set_remaining, 0);
887 spin_lock_init(&set->set_new_req_lock);
888 INIT_LIST_HEAD(&set->set_new_requests);
889 INIT_LIST_HEAD(&set->set_cblist);
890 set->set_max_inflight = UINT_MAX;
891 set->set_producer = NULL;
892 set->set_producer_arg = NULL;
897 EXPORT_SYMBOL(ptlrpc_prep_set);
900 * Allocate and initialize new request set structure with flow control
901 * extension. This extension allows to control the number of requests in-flight
902 * for the whole set. A callback function to generate requests must be provided
903 * and the request set will keep the number of requests sent over the wire to
905 * Returns a pointer to the newly allocated set structure or NULL on error.
907 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
911 struct ptlrpc_request_set *set;
913 set = ptlrpc_prep_set();
917 set->set_max_inflight = max;
918 set->set_producer = func;
919 set->set_producer_arg = arg;
925 * Wind down and free request set structure previously allocated with
927 * Ensures that all requests on the set have completed and removes
928 * all requests from the request list in a set.
929 * If any unsent request happen to be on the list, pretends that they got
930 * an error in flight and calls their completion handler.
932 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
934 struct list_head *tmp;
935 struct list_head *next;
940 /* Requests on the set should either all be completed, or all be new */
941 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
942 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
943 list_for_each(tmp, &set->set_requests) {
944 struct ptlrpc_request *req =
945 list_entry(tmp, struct ptlrpc_request,
948 LASSERT(req->rq_phase == expected_phase);
952 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
953 atomic_read(&set->set_remaining) == n, "%d / %d\n",
954 atomic_read(&set->set_remaining), n);
956 list_for_each_safe(tmp, next, &set->set_requests) {
957 struct ptlrpc_request *req =
958 list_entry(tmp, struct ptlrpc_request,
960 list_del_init(&req->rq_set_chain);
962 LASSERT(req->rq_phase == expected_phase);
964 if (req->rq_phase == RQ_PHASE_NEW) {
965 ptlrpc_req_interpret(NULL, req, -EBADR);
966 atomic_dec(&set->set_remaining);
969 spin_lock(&req->rq_lock);
971 req->rq_invalid_rqset = 0;
972 spin_unlock(&req->rq_lock);
974 ptlrpc_req_finished (req);
977 LASSERT(atomic_read(&set->set_remaining) == 0);
979 ptlrpc_reqset_put(set);
982 EXPORT_SYMBOL(ptlrpc_set_destroy);
985 * Add a callback function \a fn to the set.
986 * This function would be called when all requests on this set are completed.
987 * The function will be passed \a data argument.
989 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
990 set_interpreter_func fn, void *data)
992 struct ptlrpc_set_cbdata *cbdata;
994 OBD_ALLOC_PTR(cbdata);
998 cbdata->psc_interpret = fn;
999 cbdata->psc_data = data;
1000 list_add_tail(&cbdata->psc_item, &set->set_cblist);
1006 * Add a new request to the general purpose request set.
1007 * Assumes request reference from the caller.
1009 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1010 struct ptlrpc_request *req)
1012 LASSERT(list_empty(&req->rq_set_chain));
1014 /* The set takes over the caller's request reference */
1015 list_add_tail(&req->rq_set_chain, &set->set_requests);
1017 atomic_inc(&set->set_remaining);
1018 req->rq_queued_time = cfs_time_current();
1020 if (req->rq_reqmsg != NULL)
1021 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1023 if (set->set_producer != NULL)
1024 /* If the request set has a producer callback, the RPC must be
1025 * sent straight away */
1026 ptlrpc_send_new_req(req);
1028 EXPORT_SYMBOL(ptlrpc_set_add_req);
1031 * Add a request to a request with dedicated server thread
1032 * and wake the thread to make any necessary processing.
1033 * Currently only used for ptlrpcd.
1035 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1036 struct ptlrpc_request *req)
1038 struct ptlrpc_request_set *set = pc->pc_set;
1041 LASSERT(req->rq_set == NULL);
1042 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1044 spin_lock(&set->set_new_req_lock);
1046 * The set takes over the caller's request reference.
1049 req->rq_queued_time = cfs_time_current();
1050 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1051 count = atomic_inc_return(&set->set_new_count);
1052 spin_unlock(&set->set_new_req_lock);
1054 /* Only need to call wakeup once for the first entry. */
1056 wake_up(&set->set_waitq);
1058 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1059 * guarantee the async RPC can be processed ASAP, we have
1060 * no other better choice. It maybe fixed in future. */
1061 for (i = 0; i < pc->pc_npartners; i++)
1062 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1067 * Based on the current state of the import, determine if the request
1068 * can be sent, is an error, or should be delayed.
1070 * Returns true if this request should be delayed. If false, and
1071 * *status is set, then the request can not be sent and *status is the
1072 * error code. If false and status is 0, then request can be sent.
1074 * The imp->imp_lock must be held.
1076 static int ptlrpc_import_delay_req(struct obd_import *imp,
1077 struct ptlrpc_request *req, int *status)
1082 LASSERT (status != NULL);
1085 if (req->rq_ctx_init || req->rq_ctx_fini) {
1086 /* always allow ctx init/fini rpc go through */
1087 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1088 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1090 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1091 /* pings may safely race with umount */
1092 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1093 D_HA : D_ERROR, req, "IMP_CLOSED ");
1095 } else if (ptlrpc_send_limit_expired(req)) {
1096 /* probably doesn't need to be a D_ERROR after initial testing*/
1097 DEBUG_REQ(D_HA, req, "send limit expired ");
1098 *status = -ETIMEDOUT;
1099 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1100 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1101 /* allow CONNECT even if import is invalid */ ;
1102 if (atomic_read(&imp->imp_inval_count) != 0) {
1103 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1106 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1107 if (!imp->imp_deactive)
1108 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1109 *status = -ESHUTDOWN; /* bz 12940 */
1110 } else if (req->rq_import_generation != imp->imp_generation) {
1111 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1113 } else if (req->rq_send_state != imp->imp_state) {
1114 /* invalidate in progress - any requests should be drop */
1115 if (atomic_read(&imp->imp_inval_count) != 0) {
1116 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1118 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1119 *status = -EWOULDBLOCK;
1120 } else if (req->rq_allow_replay &&
1121 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1122 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1123 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1124 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1125 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1135 * Decide if the error message should be printed to the console or not.
1136 * Makes its decision based on request type, status, and failure frequency.
1138 * \param[in] req request that failed and may need a console message
1140 * \retval false if no message should be printed
1141 * \retval true if console message should be printed
1143 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1147 LASSERT(req->rq_reqmsg != NULL);
1148 opc = lustre_msg_get_opc(req->rq_reqmsg);
1150 /* Suppress particular reconnect errors which are to be expected. */
1151 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1154 /* Suppress timed out reconnect requests */
1155 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1159 /* Suppress most unavailable/again reconnect requests, but
1160 * print occasionally so it is clear client is trying to
1161 * connect to a server where no target is running. */
1162 err = lustre_msg_get_status(req->rq_repmsg);
1163 if ((err == -ENODEV || err == -EAGAIN) &&
1164 req->rq_import->imp_conn_cnt % 30 != 20)
1172 * Check request processing status.
1173 * Returns the status.
1175 static int ptlrpc_check_status(struct ptlrpc_request *req)
1180 err = lustre_msg_get_status(req->rq_repmsg);
1181 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1182 struct obd_import *imp = req->rq_import;
1183 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1184 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1186 if (ptlrpc_console_allow(req))
1187 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1188 "failed: rc = %d\n",
1189 imp->imp_obd->obd_name,
1191 libcfs_nid2str(nid), err);
1192 RETURN(err < 0 ? err : -EINVAL);
1196 DEBUG_REQ(D_INFO, req, "status is %d", err);
1197 } else if (err > 0) {
1198 /* XXX: translate this error from net to host */
1199 DEBUG_REQ(D_INFO, req, "status is %d", err);
1206 * save pre-versions of objects into request for replay.
1207 * Versions are obtained from server reply.
1210 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1212 struct lustre_msg *repmsg = req->rq_repmsg;
1213 struct lustre_msg *reqmsg = req->rq_reqmsg;
1214 __u64 *versions = lustre_msg_get_versions(repmsg);
1217 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1221 lustre_msg_set_versions(reqmsg, versions);
1222 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1223 versions[0], versions[1]);
1229 * Callback function called when client receives RPC reply for \a req.
1230 * Returns 0 on success or error code.
1231 * The return alue would be assigned to req->rq_status by the caller
1232 * as request processing status.
1233 * This function also decides if the request needs to be saved for later replay.
1235 static int after_reply(struct ptlrpc_request *req)
1237 struct obd_import *imp = req->rq_import;
1238 struct obd_device *obd = req->rq_import->imp_obd;
1240 struct timeval work_start;
1245 LASSERT(obd != NULL);
1246 /* repbuf must be unlinked */
1247 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1249 if (req->rq_reply_truncated) {
1250 if (ptlrpc_no_resend(req)) {
1251 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1252 " expected: %d, actual size: %d",
1253 req->rq_nob_received, req->rq_repbuf_len);
1257 sptlrpc_cli_free_repbuf(req);
1258 /* Pass the required reply buffer size (include
1259 * space for early reply).
1260 * NB: no need to roundup because alloc_repbuf
1261 * will roundup it */
1262 req->rq_replen = req->rq_nob_received;
1263 req->rq_nob_received = 0;
1264 spin_lock(&req->rq_lock);
1266 spin_unlock(&req->rq_lock);
1270 do_gettimeofday(&work_start);
1271 timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1274 * NB Until this point, the whole of the incoming message,
1275 * including buflens, status etc is in the sender's byte order.
1277 rc = sptlrpc_cli_unwrap_reply(req);
1279 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1284 * Security layer unwrap might ask resend this request.
1289 rc = unpack_reply(req);
1293 /* retry indefinitely on EINPROGRESS */
1294 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1295 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1296 time_t now = cfs_time_current_sec();
1298 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1299 spin_lock(&req->rq_lock);
1301 spin_unlock(&req->rq_lock);
1302 req->rq_nr_resend++;
1304 /* Readjust the timeout for current conditions */
1305 ptlrpc_at_set_req_timeout(req);
1306 /* delay resend to give a chance to the server to get ready.
1307 * The delay is increased by 1s on every resend and is capped to
1308 * the current request timeout (i.e. obd_timeout if AT is off,
1309 * or AT service time x 125% + 5s, see at_est2timeout) */
1310 if (req->rq_nr_resend > req->rq_timeout)
1311 req->rq_sent = now + req->rq_timeout;
1313 req->rq_sent = now + req->rq_nr_resend;
1318 if (obd->obd_svc_stats != NULL) {
1319 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1321 ptlrpc_lprocfs_rpc_sent(req, timediff);
1324 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1325 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1326 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1327 lustre_msg_get_type(req->rq_repmsg));
1331 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1332 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1333 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1334 ptlrpc_at_adj_net_latency(req,
1335 lustre_msg_get_service_time(req->rq_repmsg));
1337 rc = ptlrpc_check_status(req);
1338 imp->imp_connect_error = rc;
1342 * Either we've been evicted, or the server has failed for
1343 * some reason. Try to reconnect, and if that fails, punt to
1346 if (ptlrpc_recoverable_error(rc)) {
1347 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1348 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1351 ptlrpc_request_handle_notconn(req);
1356 * Let's look if server sent slv. Do it only for RPC with
1359 ldlm_cli_update_pool(req);
1363 * Store transno in reqmsg for replay.
1365 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1366 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1367 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1370 if (imp->imp_replayable) {
1371 spin_lock(&imp->imp_lock);
1373 * No point in adding already-committed requests to the replay
1374 * list, we will just remove them immediately. b=9829
1376 if (req->rq_transno != 0 &&
1378 lustre_msg_get_last_committed(req->rq_repmsg) ||
1380 /** version recovery */
1381 ptlrpc_save_versions(req);
1382 ptlrpc_retain_replayable_request(req, imp);
1383 } else if (req->rq_commit_cb != NULL &&
1384 list_empty(&req->rq_replay_list)) {
1385 /* NB: don't call rq_commit_cb if it's already on
1386 * rq_replay_list, ptlrpc_free_committed() will call
1387 * it later, see LU-3618 for details */
1388 spin_unlock(&imp->imp_lock);
1389 req->rq_commit_cb(req);
1390 spin_lock(&imp->imp_lock);
1394 * Replay-enabled imports return commit-status information.
1396 committed = lustre_msg_get_last_committed(req->rq_repmsg);
1397 if (likely(committed > imp->imp_peer_committed_transno))
1398 imp->imp_peer_committed_transno = committed;
1400 ptlrpc_free_committed(imp);
1402 if (!list_empty(&imp->imp_replay_list)) {
1403 struct ptlrpc_request *last;
1405 last = list_entry(imp->imp_replay_list.prev,
1406 struct ptlrpc_request,
1409 * Requests with rq_replay stay on the list even if no
1410 * commit is expected.
1412 if (last->rq_transno > imp->imp_peer_committed_transno)
1413 ptlrpc_pinger_commit_expected(imp);
1416 spin_unlock(&imp->imp_lock);
1423 * Helper function to send request \a req over the network for the first time
1424 * Also adjusts request phase.
1425 * Returns 0 on success or error code.
1427 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1429 struct obd_import *imp = req->rq_import;
1430 struct list_head *tmp;
1431 __u64 min_xid = ~0ULL;
1435 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1437 /* do not try to go further if there is not enough memory in enc_pool */
1438 if (req->rq_sent && req->rq_bulk != NULL)
1439 if (req->rq_bulk->bd_iov_count > get_free_pages_in_pool() &&
1440 pool_is_at_full_capacity())
1443 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1444 (!req->rq_generation_set ||
1445 req->rq_import_generation == imp->imp_generation))
1448 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1450 spin_lock(&imp->imp_lock);
1452 /* the very first time we assign XID. it's important to assign XID
1453 * and put it on the list atomically, so that the lowest assigned
1454 * XID is always known. this is vital for multislot last_rcvd */
1455 if (req->rq_send_state == LUSTRE_IMP_REPLAY) {
1456 LASSERT(req->rq_xid != 0);
1458 LASSERT(req->rq_xid == 0);
1459 req->rq_xid = ptlrpc_next_xid();
1462 if (!req->rq_generation_set)
1463 req->rq_import_generation = imp->imp_generation;
1465 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1466 spin_lock(&req->rq_lock);
1467 req->rq_waiting = 1;
1468 spin_unlock(&req->rq_lock);
1470 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1471 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1472 ptlrpc_import_state_name(req->rq_send_state),
1473 ptlrpc_import_state_name(imp->imp_state));
1474 LASSERT(list_empty(&req->rq_list));
1475 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1476 atomic_inc(&req->rq_import->imp_inflight);
1477 spin_unlock(&imp->imp_lock);
1482 spin_unlock(&imp->imp_lock);
1483 req->rq_status = rc;
1484 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1488 LASSERT(list_empty(&req->rq_list));
1489 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1490 atomic_inc(&req->rq_import->imp_inflight);
1492 /* find the lowest unreplied XID */
1493 list_for_each(tmp, &imp->imp_delayed_list) {
1494 struct ptlrpc_request *r;
1495 r = list_entry(tmp, struct ptlrpc_request, rq_list);
1496 if (r->rq_xid < min_xid)
1497 min_xid = r->rq_xid;
1499 list_for_each(tmp, &imp->imp_sending_list) {
1500 struct ptlrpc_request *r;
1501 r = list_entry(tmp, struct ptlrpc_request, rq_list);
1502 if (r->rq_xid < min_xid)
1503 min_xid = r->rq_xid;
1505 spin_unlock(&imp->imp_lock);
1507 if (likely(min_xid != ~0ULL))
1508 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid - 1);
1510 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1512 rc = sptlrpc_req_refresh_ctx(req, -1);
1515 req->rq_status = rc;
1518 spin_lock(&req->rq_lock);
1519 req->rq_wait_ctx = 1;
1520 spin_unlock(&req->rq_lock);
1525 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1526 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1527 imp->imp_obd->obd_uuid.uuid,
1528 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1529 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1530 lustre_msg_get_opc(req->rq_reqmsg));
1532 rc = ptl_send_rpc(req, 0);
1533 if (rc == -ENOMEM) {
1534 spin_lock(&imp->imp_lock);
1535 if (!list_empty(&req->rq_list)) {
1536 list_del_init(&req->rq_list);
1537 atomic_dec(&req->rq_import->imp_inflight);
1539 spin_unlock(&imp->imp_lock);
1540 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1544 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1545 spin_lock(&req->rq_lock);
1546 req->rq_net_err = 1;
1547 spin_unlock(&req->rq_lock);
1553 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1558 LASSERT(set->set_producer != NULL);
1560 remaining = atomic_read(&set->set_remaining);
1562 /* populate the ->set_requests list with requests until we
1563 * reach the maximum number of RPCs in flight for this set */
1564 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1565 rc = set->set_producer(set, set->set_producer_arg);
1566 if (rc == -ENOENT) {
1567 /* no more RPC to produce */
1568 set->set_producer = NULL;
1569 set->set_producer_arg = NULL;
1574 RETURN((atomic_read(&set->set_remaining) - remaining));
1578 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1579 * and no more replies are expected.
1580 * (it is possible to get less replies than requests sent e.g. due to timed out
1581 * requests or requests that we had trouble to send out)
1583 * NOTE: This function contains a potential schedule point (cond_resched()).
1585 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1587 struct list_head *tmp, *next;
1588 struct list_head comp_reqs;
1589 int force_timer_recalc = 0;
1592 if (atomic_read(&set->set_remaining) == 0)
1595 INIT_LIST_HEAD(&comp_reqs);
1596 list_for_each_safe(tmp, next, &set->set_requests) {
1597 struct ptlrpc_request *req =
1598 list_entry(tmp, struct ptlrpc_request,
1600 struct obd_import *imp = req->rq_import;
1601 int unregistered = 0;
1604 /* This schedule point is mainly for the ptlrpcd caller of this
1605 * function. Most ptlrpc sets are not long-lived and unbounded
1606 * in length, but at the least the set used by the ptlrpcd is.
1607 * Since the processing time is unbounded, we need to insert an
1608 * explicit schedule point to make the thread well-behaved.
1612 if (req->rq_phase == RQ_PHASE_NEW &&
1613 ptlrpc_send_new_req(req)) {
1614 force_timer_recalc = 1;
1617 /* delayed send - skip */
1618 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1621 /* delayed resend - skip */
1622 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1623 req->rq_sent > cfs_time_current_sec())
1626 if (!(req->rq_phase == RQ_PHASE_RPC ||
1627 req->rq_phase == RQ_PHASE_BULK ||
1628 req->rq_phase == RQ_PHASE_INTERPRET ||
1629 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1630 req->rq_phase == RQ_PHASE_COMPLETE)) {
1631 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1635 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1636 LASSERT(req->rq_next_phase != req->rq_phase);
1637 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1640 * Skip processing until reply is unlinked. We
1641 * can't return to pool before that and we can't
1642 * call interpret before that. We need to make
1643 * sure that all rdma transfers finished and will
1644 * not corrupt any data.
1646 if (ptlrpc_client_recv_or_unlink(req) ||
1647 ptlrpc_client_bulk_active(req))
1651 * Turn fail_loc off to prevent it from looping
1654 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1655 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1658 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1659 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1664 * Move to next phase if reply was successfully
1667 ptlrpc_rqphase_move(req, req->rq_next_phase);
1670 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1671 list_move_tail(&req->rq_set_chain, &comp_reqs);
1675 if (req->rq_phase == RQ_PHASE_INTERPRET)
1676 GOTO(interpret, req->rq_status);
1679 * Note that this also will start async reply unlink.
1681 if (req->rq_net_err && !req->rq_timedout) {
1682 ptlrpc_expire_one_request(req, 1);
1685 * Check if we still need to wait for unlink.
1687 if (ptlrpc_client_recv_or_unlink(req) ||
1688 ptlrpc_client_bulk_active(req))
1690 /* If there is no need to resend, fail it now. */
1691 if (req->rq_no_resend) {
1692 if (req->rq_status == 0)
1693 req->rq_status = -EIO;
1694 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1695 GOTO(interpret, req->rq_status);
1702 spin_lock(&req->rq_lock);
1703 req->rq_replied = 0;
1704 spin_unlock(&req->rq_lock);
1705 if (req->rq_status == 0)
1706 req->rq_status = -EIO;
1707 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1708 GOTO(interpret, req->rq_status);
1711 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1712 * so it sets rq_intr regardless of individual rpc
1713 * timeouts. The synchronous IO waiting path sets
1714 * rq_intr irrespective of whether ptlrpcd
1715 * has seen a timeout. Our policy is to only interpret
1716 * interrupted rpcs after they have timed out, so we
1717 * need to enforce that here.
1720 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1721 req->rq_wait_ctx)) {
1722 req->rq_status = -EINTR;
1723 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1724 GOTO(interpret, req->rq_status);
1727 if (req->rq_phase == RQ_PHASE_RPC) {
1728 if (req->rq_timedout || req->rq_resend ||
1729 req->rq_waiting || req->rq_wait_ctx) {
1732 if (!ptlrpc_unregister_reply(req, 1)) {
1733 ptlrpc_unregister_bulk(req, 1);
1737 spin_lock(&imp->imp_lock);
1738 if (ptlrpc_import_delay_req(imp, req, &status)){
1739 /* put on delay list - only if we wait
1740 * recovery finished - before send */
1741 list_del_init(&req->rq_list);
1742 list_add_tail(&req->rq_list,
1745 spin_unlock(&imp->imp_lock);
1750 req->rq_status = status;
1751 ptlrpc_rqphase_move(req,
1752 RQ_PHASE_INTERPRET);
1753 spin_unlock(&imp->imp_lock);
1754 GOTO(interpret, req->rq_status);
1756 if (ptlrpc_no_resend(req) &&
1757 !req->rq_wait_ctx) {
1758 req->rq_status = -ENOTCONN;
1759 ptlrpc_rqphase_move(req,
1760 RQ_PHASE_INTERPRET);
1761 spin_unlock(&imp->imp_lock);
1762 GOTO(interpret, req->rq_status);
1765 list_del_init(&req->rq_list);
1766 list_add_tail(&req->rq_list,
1767 &imp->imp_sending_list);
1769 spin_unlock(&imp->imp_lock);
1771 spin_lock(&req->rq_lock);
1772 req->rq_waiting = 0;
1773 spin_unlock(&req->rq_lock);
1775 if (req->rq_timedout || req->rq_resend) {
1776 /* This is re-sending anyways,
1777 * let's mark req as resend. */
1778 spin_lock(&req->rq_lock);
1780 spin_unlock(&req->rq_lock);
1782 if (req->rq_bulk != NULL &&
1783 !ptlrpc_unregister_bulk(req, 1))
1787 * rq_wait_ctx is only touched by ptlrpcd,
1788 * so no lock is needed here.
1790 status = sptlrpc_req_refresh_ctx(req, -1);
1793 req->rq_status = status;
1794 spin_lock(&req->rq_lock);
1795 req->rq_wait_ctx = 0;
1796 spin_unlock(&req->rq_lock);
1797 force_timer_recalc = 1;
1799 spin_lock(&req->rq_lock);
1800 req->rq_wait_ctx = 1;
1801 spin_unlock(&req->rq_lock);
1806 spin_lock(&req->rq_lock);
1807 req->rq_wait_ctx = 0;
1808 spin_unlock(&req->rq_lock);
1811 rc = ptl_send_rpc(req, 0);
1812 if (rc == -ENOMEM) {
1813 spin_lock(&imp->imp_lock);
1814 if (!list_empty(&req->rq_list))
1815 list_del_init(&req->rq_list);
1816 spin_unlock(&imp->imp_lock);
1817 ptlrpc_rqphase_move(req, RQ_PHASE_NEW);
1821 DEBUG_REQ(D_HA, req,
1822 "send failed: rc = %d", rc);
1823 force_timer_recalc = 1;
1824 spin_lock(&req->rq_lock);
1825 req->rq_net_err = 1;
1826 spin_unlock(&req->rq_lock);
1829 /* need to reset the timeout */
1830 force_timer_recalc = 1;
1833 spin_lock(&req->rq_lock);
1835 if (ptlrpc_client_early(req)) {
1836 ptlrpc_at_recv_early_reply(req);
1837 spin_unlock(&req->rq_lock);
1841 /* Still waiting for a reply? */
1842 if (ptlrpc_client_recv(req)) {
1843 spin_unlock(&req->rq_lock);
1847 /* Did we actually receive a reply? */
1848 if (!ptlrpc_client_replied(req)) {
1849 spin_unlock(&req->rq_lock);
1853 spin_unlock(&req->rq_lock);
1855 /* unlink from net because we are going to
1856 * swab in-place of reply buffer */
1857 unregistered = ptlrpc_unregister_reply(req, 1);
1861 req->rq_status = after_reply(req);
1865 /* If there is no bulk associated with this request,
1866 * then we're done and should let the interpreter
1867 * process the reply. Similarly if the RPC returned
1868 * an error, and therefore the bulk will never arrive.
1870 if (req->rq_bulk == NULL || req->rq_status < 0) {
1871 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1872 GOTO(interpret, req->rq_status);
1875 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1878 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1879 if (ptlrpc_client_bulk_active(req))
1882 if (req->rq_bulk->bd_failure) {
1883 /* The RPC reply arrived OK, but the bulk screwed
1884 * up! Dead weird since the server told us the RPC
1885 * was good after getting the REPLY for her GET or
1886 * the ACK for her PUT. */
1887 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1888 req->rq_status = -EIO;
1891 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1894 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1896 /* This moves to "unregistering" phase we need to wait for
1898 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1899 /* start async bulk unlink too */
1900 ptlrpc_unregister_bulk(req, 1);
1904 if (!ptlrpc_unregister_bulk(req, 1))
1907 /* When calling interpret receiving already should be
1909 LASSERT(!req->rq_receiving_reply);
1911 ptlrpc_req_interpret(env, req, req->rq_status);
1913 if (ptlrpcd_check_work(req)) {
1914 atomic_dec(&set->set_remaining);
1917 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1919 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1920 "Completed RPC pname:cluuid:pid:xid:nid:"
1921 "opc %s:%s:%d:"LPU64":%s:%d\n",
1922 current_comm(), imp->imp_obd->obd_uuid.uuid,
1923 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1924 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1925 lustre_msg_get_opc(req->rq_reqmsg));
1927 spin_lock(&imp->imp_lock);
1928 /* Request already may be not on sending or delaying list. This
1929 * may happen in the case of marking it erroneous for the case
1930 * ptlrpc_import_delay_req(req, status) find it impossible to
1931 * allow sending this rpc and returns *status != 0. */
1932 if (!list_empty(&req->rq_list)) {
1933 list_del_init(&req->rq_list);
1934 atomic_dec(&imp->imp_inflight);
1936 spin_unlock(&imp->imp_lock);
1938 atomic_dec(&set->set_remaining);
1939 wake_up_all(&imp->imp_recovery_waitq);
1941 if (set->set_producer) {
1942 /* produce a new request if possible */
1943 if (ptlrpc_set_producer(set) > 0)
1944 force_timer_recalc = 1;
1946 /* free the request that has just been completed
1947 * in order not to pollute set->set_requests */
1948 list_del_init(&req->rq_set_chain);
1949 spin_lock(&req->rq_lock);
1951 req->rq_invalid_rqset = 0;
1952 spin_unlock(&req->rq_lock);
1954 /* record rq_status to compute the final status later */
1955 if (req->rq_status != 0)
1956 set->set_rc = req->rq_status;
1957 ptlrpc_req_finished(req);
1959 list_move_tail(&req->rq_set_chain, &comp_reqs);
1963 /* move completed request at the head of list so it's easier for
1964 * caller to find them */
1965 list_splice(&comp_reqs, &set->set_requests);
1967 /* If we hit an error, we want to recover promptly. */
1968 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1970 EXPORT_SYMBOL(ptlrpc_check_set);
1973 * Time out request \a req. is \a async_unlink is set, that means do not wait
1974 * until LNet actually confirms network buffer unlinking.
1975 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1977 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1979 struct obd_import *imp = req->rq_import;
1983 spin_lock(&req->rq_lock);
1984 req->rq_timedout = 1;
1985 spin_unlock(&req->rq_lock);
1987 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1988 "/real "CFS_DURATION_T"]",
1989 req->rq_net_err ? "failed due to network error" :
1990 ((req->rq_real_sent == 0 ||
1991 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1992 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1993 "timed out for sent delay" : "timed out for slow reply"),
1994 req->rq_sent, req->rq_real_sent);
1996 if (imp != NULL && obd_debug_peer_on_timeout)
1997 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1999 ptlrpc_unregister_reply(req, async_unlink);
2000 ptlrpc_unregister_bulk(req, async_unlink);
2002 if (obd_dump_on_timeout)
2003 libcfs_debug_dumplog();
2006 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2010 atomic_inc(&imp->imp_timeouts);
2012 /* The DLM server doesn't want recovery run on its imports. */
2013 if (imp->imp_dlm_fake)
2016 /* If this request is for recovery or other primordial tasks,
2017 * then error it out here. */
2018 if (req->rq_ctx_init || req->rq_ctx_fini ||
2019 req->rq_send_state != LUSTRE_IMP_FULL ||
2020 imp->imp_obd->obd_no_recov) {
2021 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2022 ptlrpc_import_state_name(req->rq_send_state),
2023 ptlrpc_import_state_name(imp->imp_state));
2024 spin_lock(&req->rq_lock);
2025 req->rq_status = -ETIMEDOUT;
2027 spin_unlock(&req->rq_lock);
2031 /* if a request can't be resent we can't wait for an answer after
2033 if (ptlrpc_no_resend(req)) {
2034 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2038 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2044 * Time out all uncompleted requests in request set pointed by \a data
2045 * Callback used when waiting on sets with l_wait_event.
2048 int ptlrpc_expired_set(void *data)
2050 struct ptlrpc_request_set *set = data;
2051 struct list_head *tmp;
2052 time_t now = cfs_time_current_sec();
2055 LASSERT(set != NULL);
2058 * A timeout expired. See which reqs it applies to...
2060 list_for_each(tmp, &set->set_requests) {
2061 struct ptlrpc_request *req =
2062 list_entry(tmp, struct ptlrpc_request,
2065 /* don't expire request waiting for context */
2066 if (req->rq_wait_ctx)
2069 /* Request in-flight? */
2070 if (!((req->rq_phase == RQ_PHASE_RPC &&
2071 !req->rq_waiting && !req->rq_resend) ||
2072 (req->rq_phase == RQ_PHASE_BULK)))
2075 if (req->rq_timedout || /* already dealt with */
2076 req->rq_deadline > now) /* not expired */
2079 /* Deal with this guy. Do it asynchronously to not block
2080 * ptlrpcd thread. */
2081 ptlrpc_expire_one_request(req, 1);
2085 * When waiting for a whole set, we always break out of the
2086 * sleep so we can recalculate the timeout, or enable interrupts
2087 * if everyone's timed out.
2093 * Sets rq_intr flag in \a req under spinlock.
2095 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2097 spin_lock(&req->rq_lock);
2099 spin_unlock(&req->rq_lock);
2101 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2104 * Interrupts (sets interrupted flag) all uncompleted requests in
2105 * a set \a data. Callback for l_wait_event for interruptible waits.
2107 static void ptlrpc_interrupted_set(void *data)
2109 struct ptlrpc_request_set *set = data;
2110 struct list_head *tmp;
2112 LASSERT(set != NULL);
2113 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2115 list_for_each(tmp, &set->set_requests) {
2116 struct ptlrpc_request *req =
2117 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2119 if (req->rq_phase != RQ_PHASE_RPC &&
2120 req->rq_phase != RQ_PHASE_UNREGISTERING)
2123 ptlrpc_mark_interrupted(req);
2128 * Get the smallest timeout in the set; this does NOT set a timeout.
2130 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2132 struct list_head *tmp;
2133 time_t now = cfs_time_current_sec();
2135 struct ptlrpc_request *req;
2139 list_for_each(tmp, &set->set_requests) {
2140 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2143 * Request in-flight?
2145 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2146 (req->rq_phase == RQ_PHASE_BULK) ||
2147 (req->rq_phase == RQ_PHASE_NEW)))
2151 * Already timed out.
2153 if (req->rq_timedout)
2159 if (req->rq_wait_ctx)
2162 if (req->rq_phase == RQ_PHASE_NEW)
2163 deadline = req->rq_sent;
2164 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2165 deadline = req->rq_sent;
2167 deadline = req->rq_sent + req->rq_timeout;
2169 if (deadline <= now) /* actually expired already */
2170 timeout = 1; /* ASAP */
2171 else if (timeout == 0 || timeout > deadline - now)
2172 timeout = deadline - now;
2178 * Send all unset request from the set and then wait untill all
2179 * requests in the set complete (either get a reply, timeout, get an
2180 * error or otherwise be interrupted).
2181 * Returns 0 on success or error code otherwise.
2183 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2185 struct list_head *tmp;
2186 struct ptlrpc_request *req;
2187 struct l_wait_info lwi;
2191 if (set->set_producer)
2192 (void)ptlrpc_set_producer(set);
2194 list_for_each(tmp, &set->set_requests) {
2195 req = list_entry(tmp, struct ptlrpc_request,
2197 if (req->rq_phase == RQ_PHASE_NEW)
2198 (void)ptlrpc_send_new_req(req);
2201 if (list_empty(&set->set_requests))
2205 timeout = ptlrpc_set_next_timeout(set);
2207 /* wait until all complete, interrupted, or an in-flight
2209 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2212 if (timeout == 0 && !signal_pending(current))
2214 * No requests are in-flight (ether timed out
2215 * or delayed), so we can allow interrupts.
2216 * We still want to block for a limited time,
2217 * so we allow interrupts during the timeout.
2219 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2221 ptlrpc_interrupted_set, set);
2224 * At least one request is in flight, so no
2225 * interrupts are allowed. Wait until all
2226 * complete, or an in-flight req times out.
2228 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2229 ptlrpc_expired_set, set);
2231 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2233 /* LU-769 - if we ignored the signal because it was already
2234 * pending when we started, we need to handle it now or we risk
2235 * it being ignored forever */
2236 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2237 signal_pending(current)) {
2238 sigset_t blocked_sigs =
2239 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2241 /* In fact we only interrupt for the "fatal" signals
2242 * like SIGINT or SIGKILL. We still ignore less
2243 * important signals since ptlrpc set is not easily
2244 * reentrant from userspace again */
2245 if (signal_pending(current))
2246 ptlrpc_interrupted_set(set);
2247 cfs_restore_sigs(blocked_sigs);
2250 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2252 /* -EINTR => all requests have been flagged rq_intr so next
2254 * -ETIMEDOUT => someone timed out. When all reqs have
2255 * timed out, signals are enabled allowing completion with
2257 * I don't really care if we go once more round the loop in
2258 * the error cases -eeb. */
2259 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2260 list_for_each(tmp, &set->set_requests) {
2261 req = list_entry(tmp, struct ptlrpc_request,
2263 spin_lock(&req->rq_lock);
2264 req->rq_invalid_rqset = 1;
2265 spin_unlock(&req->rq_lock);
2268 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2270 LASSERT(atomic_read(&set->set_remaining) == 0);
2272 rc = set->set_rc; /* rq_status of already freed requests if any */
2273 list_for_each(tmp, &set->set_requests) {
2274 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2276 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2277 if (req->rq_status != 0)
2278 rc = req->rq_status;
2281 if (set->set_interpret != NULL) {
2282 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2284 rc = interpreter (set, set->set_arg, rc);
2286 struct ptlrpc_set_cbdata *cbdata, *n;
2289 list_for_each_entry_safe(cbdata, n,
2290 &set->set_cblist, psc_item) {
2291 list_del_init(&cbdata->psc_item);
2292 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2295 OBD_FREE_PTR(cbdata);
2301 EXPORT_SYMBOL(ptlrpc_set_wait);
2304 * Helper fuction for request freeing.
2305 * Called when request count reached zero and request needs to be freed.
2306 * Removes request from all sorts of sending/replay lists it might be on,
2307 * frees network buffers if any are present.
2308 * If \a locked is set, that means caller is already holding import imp_lock
2309 * and so we no longer need to reobtain it (for certain lists manipulations)
2311 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2315 if (request == NULL)
2318 LASSERT(!request->rq_srv_req);
2319 LASSERT(request->rq_export == NULL);
2320 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2321 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2322 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2323 LASSERTF(!request->rq_replay, "req %p\n", request);
2325 req_capsule_fini(&request->rq_pill);
2327 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2328 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2329 if (request->rq_import != NULL) {
2331 spin_lock(&request->rq_import->imp_lock);
2332 list_del_init(&request->rq_replay_list);
2334 spin_unlock(&request->rq_import->imp_lock);
2336 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2338 if (atomic_read(&request->rq_refcount) != 0) {
2339 DEBUG_REQ(D_ERROR, request,
2340 "freeing request with nonzero refcount");
2344 if (request->rq_repbuf != NULL)
2345 sptlrpc_cli_free_repbuf(request);
2347 if (request->rq_import != NULL) {
2348 class_import_put(request->rq_import);
2349 request->rq_import = NULL;
2351 if (request->rq_bulk != NULL)
2352 ptlrpc_free_bulk(request->rq_bulk);
2354 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2355 sptlrpc_cli_free_reqbuf(request);
2357 if (request->rq_cli_ctx)
2358 sptlrpc_req_put_ctx(request, !locked);
2360 if (request->rq_pool)
2361 __ptlrpc_free_req_to_pool(request);
2363 ptlrpc_request_cache_free(request);
2367 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2369 * Drop one request reference. Must be called with import imp_lock held.
2370 * When reference count drops to zero, request is freed.
2372 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2374 assert_spin_locked(&request->rq_import->imp_lock);
2375 (void)__ptlrpc_req_finished(request, 1);
2380 * Drops one reference count for request \a request.
2381 * \a locked set indicates that caller holds import imp_lock.
2382 * Frees the request whe reference count reaches zero.
2384 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2387 if (request == NULL)
2390 if (request == LP_POISON ||
2391 request->rq_reqmsg == LP_POISON) {
2392 CERROR("dereferencing freed request (bug 575)\n");
2397 DEBUG_REQ(D_INFO, request, "refcount now %u",
2398 atomic_read(&request->rq_refcount) - 1);
2400 if (atomic_dec_and_test(&request->rq_refcount)) {
2401 __ptlrpc_free_req(request, locked);
2409 * Drops one reference count for a request.
2411 void ptlrpc_req_finished(struct ptlrpc_request *request)
2413 __ptlrpc_req_finished(request, 0);
2415 EXPORT_SYMBOL(ptlrpc_req_finished);
2418 * Returns xid of a \a request
2420 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2422 return request->rq_xid;
2424 EXPORT_SYMBOL(ptlrpc_req_xid);
2427 * Disengage the client's reply buffer from the network
2428 * NB does _NOT_ unregister any client-side bulk.
2429 * IDEMPOTENT, but _not_ safe against concurrent callers.
2430 * The request owner (i.e. the thread doing the I/O) must call...
2431 * Returns 0 on success or 1 if unregistering cannot be made.
2433 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2436 struct l_wait_info lwi;
2441 LASSERT(!in_interrupt());
2444 * Let's setup deadline for reply unlink.
2446 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2447 async && request->rq_reply_deadline == 0)
2448 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2451 * Nothing left to do.
2453 if (!ptlrpc_client_recv_or_unlink(request))
2456 LNetMDUnlink(request->rq_reply_md_h);
2459 * Let's check it once again.
2461 if (!ptlrpc_client_recv_or_unlink(request))
2465 * Move to "Unregistering" phase as reply was not unlinked yet.
2467 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2470 * Do not wait for unlink to finish.
2476 * We have to l_wait_event() whatever the result, to give liblustre
2477 * a chance to run reply_in_callback(), and to make sure we've
2478 * unlinked before returning a req to the pool.
2481 /* The wq argument is ignored by user-space wait_event macros */
2482 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2483 &request->rq_set->set_waitq :
2484 &request->rq_reply_waitq;
2485 /* Network access will complete in finite time but the HUGE
2486 * timeout lets us CWARN for visibility of sluggish NALs */
2487 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2488 cfs_time_seconds(1), NULL, NULL);
2489 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2492 ptlrpc_rqphase_move(request, request->rq_next_phase);
2496 LASSERT(rc == -ETIMEDOUT);
2497 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2498 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2499 request->rq_receiving_reply,
2500 request->rq_req_unlinked,
2501 request->rq_reply_unlinked);
2506 static void ptlrpc_free_request(struct ptlrpc_request *req)
2508 spin_lock(&req->rq_lock);
2510 spin_unlock(&req->rq_lock);
2512 if (req->rq_commit_cb != NULL)
2513 req->rq_commit_cb(req);
2514 list_del_init(&req->rq_replay_list);
2516 __ptlrpc_req_finished(req, 1);
2520 * the request is committed and dropped from the replay list of its import
2522 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2524 struct obd_import *imp = req->rq_import;
2526 spin_lock(&imp->imp_lock);
2527 if (list_empty(&req->rq_replay_list)) {
2528 spin_unlock(&imp->imp_lock);
2532 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2533 ptlrpc_free_request(req);
2535 spin_unlock(&imp->imp_lock);
2537 EXPORT_SYMBOL(ptlrpc_request_committed);
2540 * Iterates through replay_list on import and prunes
2541 * all requests have transno smaller than last_committed for the
2542 * import and don't have rq_replay set.
2543 * Since requests are sorted in transno order, stops when meetign first
2544 * transno bigger than last_committed.
2545 * caller must hold imp->imp_lock
2547 void ptlrpc_free_committed(struct obd_import *imp)
2549 struct ptlrpc_request *req, *saved;
2550 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2551 bool skip_committed_list = true;
2554 LASSERT(imp != NULL);
2555 assert_spin_locked(&imp->imp_lock);
2557 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2558 imp->imp_generation == imp->imp_last_generation_checked) {
2559 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2560 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2563 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2564 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2565 imp->imp_generation);
2567 if (imp->imp_generation != imp->imp_last_generation_checked ||
2568 imp->imp_last_transno_checked == 0)
2569 skip_committed_list = false;
2571 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2572 imp->imp_last_generation_checked = imp->imp_generation;
2574 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2576 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2577 LASSERT(req != last_req);
2580 if (req->rq_transno == 0) {
2581 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2584 if (req->rq_import_generation < imp->imp_generation) {
2585 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2589 /* not yet committed */
2590 if (req->rq_transno > imp->imp_peer_committed_transno) {
2591 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2595 if (req->rq_replay) {
2596 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2597 list_move_tail(&req->rq_replay_list,
2598 &imp->imp_committed_list);
2602 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2603 imp->imp_peer_committed_transno);
2605 ptlrpc_free_request(req);
2608 if (skip_committed_list)
2611 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2613 LASSERT(req->rq_transno != 0);
2614 if (req->rq_import_generation < imp->imp_generation) {
2615 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2616 ptlrpc_free_request(req);
2617 } else if (!req->rq_replay) {
2618 DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2619 ptlrpc_free_request(req);
2626 void ptlrpc_cleanup_client(struct obd_import *imp)
2633 * Schedule previously sent request for resend.
2634 * For bulk requests we assign new xid (to avoid problems with
2635 * lost replies and therefore several transfers landing into same buffer
2636 * from different sending attempts).
2638 void ptlrpc_resend_req(struct ptlrpc_request *req)
2640 DEBUG_REQ(D_HA, req, "going to resend");
2641 spin_lock(&req->rq_lock);
2643 /* Request got reply but linked to the import list still.
2644 Let ptlrpc_check_set() to process it. */
2645 if (ptlrpc_client_replied(req)) {
2646 spin_unlock(&req->rq_lock);
2647 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2651 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2652 req->rq_status = -EAGAIN;
2655 req->rq_net_err = 0;
2656 req->rq_timedout = 0;
2658 ptlrpc_client_wake_req(req);
2659 spin_unlock(&req->rq_lock);
2662 /* XXX: this function and rq_status are currently unused */
2663 void ptlrpc_restart_req(struct ptlrpc_request *req)
2665 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2666 req->rq_status = -ERESTARTSYS;
2668 spin_lock(&req->rq_lock);
2669 req->rq_restart = 1;
2670 req->rq_timedout = 0;
2671 ptlrpc_client_wake_req(req);
2672 spin_unlock(&req->rq_lock);
2676 * Grab additional reference on a request \a req
2678 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2681 atomic_inc(&req->rq_refcount);
2684 EXPORT_SYMBOL(ptlrpc_request_addref);
2687 * Add a request to import replay_list.
2688 * Must be called under imp_lock
2690 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2691 struct obd_import *imp)
2693 struct list_head *tmp;
2695 assert_spin_locked(&imp->imp_lock);
2697 if (req->rq_transno == 0) {
2698 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2702 /* clear this for new requests that were resent as well
2703 as resent replayed requests. */
2704 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2706 /* don't re-add requests that have been replayed */
2707 if (!list_empty(&req->rq_replay_list))
2710 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2712 spin_lock(&req->rq_lock);
2714 spin_unlock(&req->rq_lock);
2716 LASSERT(imp->imp_replayable);
2717 /* Balanced in ptlrpc_free_committed, usually. */
2718 ptlrpc_request_addref(req);
2719 list_for_each_prev(tmp, &imp->imp_replay_list) {
2720 struct ptlrpc_request *iter = list_entry(tmp,
2721 struct ptlrpc_request,
2724 /* We may have duplicate transnos if we create and then
2725 * open a file, or for closes retained if to match creating
2726 * opens, so use req->rq_xid as a secondary key.
2727 * (See bugs 684, 685, and 428.)
2728 * XXX no longer needed, but all opens need transnos!
2730 if (iter->rq_transno > req->rq_transno)
2733 if (iter->rq_transno == req->rq_transno) {
2734 LASSERT(iter->rq_xid != req->rq_xid);
2735 if (iter->rq_xid > req->rq_xid)
2739 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2743 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2747 * Send request and wait until it completes.
2748 * Returns request processing status.
2750 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2752 struct ptlrpc_request_set *set;
2756 LASSERT(req->rq_set == NULL);
2757 LASSERT(!req->rq_receiving_reply);
2759 set = ptlrpc_prep_set();
2761 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2765 /* for distributed debugging */
2766 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2768 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2769 ptlrpc_request_addref(req);
2770 ptlrpc_set_add_req(set, req);
2771 rc = ptlrpc_set_wait(set);
2772 ptlrpc_set_destroy(set);
2776 EXPORT_SYMBOL(ptlrpc_queue_wait);
2779 * Callback used for replayed requests reply processing.
2780 * In case of successful reply calls registered request replay callback.
2781 * In case of error restart replay process.
2783 static int ptlrpc_replay_interpret(const struct lu_env *env,
2784 struct ptlrpc_request *req,
2785 void * data, int rc)
2787 struct ptlrpc_replay_async_args *aa = data;
2788 struct obd_import *imp = req->rq_import;
2791 atomic_dec(&imp->imp_replay_inflight);
2793 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2794 * server got the request, but bulk transfer timeout, let's
2795 * replay the bulk req again */
2796 if (!ptlrpc_client_replied(req) ||
2797 (req->rq_bulk != NULL &&
2798 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2799 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2800 GOTO(out, rc = -ETIMEDOUT);
2803 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2804 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2805 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2806 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2808 /** VBR: check version failure */
2809 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2810 /** replay was failed due to version mismatch */
2811 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2812 spin_lock(&imp->imp_lock);
2813 imp->imp_vbr_failed = 1;
2814 imp->imp_no_lock_replay = 1;
2815 spin_unlock(&imp->imp_lock);
2816 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2818 /** The transno had better not change over replay. */
2819 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2820 lustre_msg_get_transno(req->rq_repmsg) ||
2821 lustre_msg_get_transno(req->rq_repmsg) == 0,
2823 lustre_msg_get_transno(req->rq_reqmsg),
2824 lustre_msg_get_transno(req->rq_repmsg));
2827 spin_lock(&imp->imp_lock);
2828 /** if replays by version then gap occur on server, no trust to locks */
2829 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2830 imp->imp_no_lock_replay = 1;
2831 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2832 spin_unlock(&imp->imp_lock);
2833 LASSERT(imp->imp_last_replay_transno);
2835 /* transaction number shouldn't be bigger than the latest replayed */
2836 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2837 DEBUG_REQ(D_ERROR, req,
2838 "Reported transno "LPU64" is bigger than the "
2839 "replayed one: "LPU64, req->rq_transno,
2840 lustre_msg_get_transno(req->rq_reqmsg));
2841 GOTO(out, rc = -EINVAL);
2844 DEBUG_REQ(D_HA, req, "got rep");
2846 /* let the callback do fixups, possibly including in the request */
2847 if (req->rq_replay_cb)
2848 req->rq_replay_cb(req);
2850 if (ptlrpc_client_replied(req) &&
2851 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2852 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2853 lustre_msg_get_status(req->rq_repmsg),
2854 aa->praa_old_status);
2856 /* Put it back for re-replay. */
2857 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2861 * Errors while replay can set transno to 0, but
2862 * imp_last_replay_transno shouldn't be set to 0 anyway
2864 if (req->rq_transno == 0)
2865 CERROR("Transno is 0 during replay!\n");
2867 /* continue with recovery */
2868 rc = ptlrpc_import_recovery_state_machine(imp);
2870 req->rq_send_state = aa->praa_old_state;
2873 /* this replay failed, so restart recovery */
2874 ptlrpc_connect_import(imp);
2880 * Prepares and queues request for replay.
2881 * Adds it to ptlrpcd queue for actual sending.
2882 * Returns 0 on success.
2884 int ptlrpc_replay_req(struct ptlrpc_request *req)
2886 struct ptlrpc_replay_async_args *aa;
2889 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2891 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2892 aa = ptlrpc_req_async_args(req);
2893 memset(aa, 0, sizeof *aa);
2895 /* Prepare request to be resent with ptlrpcd */
2896 aa->praa_old_state = req->rq_send_state;
2897 req->rq_send_state = LUSTRE_IMP_REPLAY;
2898 req->rq_phase = RQ_PHASE_NEW;
2899 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2901 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2903 req->rq_interpret_reply = ptlrpc_replay_interpret;
2904 /* Readjust the timeout for current conditions */
2905 ptlrpc_at_set_req_timeout(req);
2907 /* Tell server the net_latency, so the server can calculate how long
2908 * it should wait for next replay */
2909 lustre_msg_set_service_time(req->rq_reqmsg,
2910 ptlrpc_at_get_net_latency(req));
2911 DEBUG_REQ(D_HA, req, "REPLAY");
2913 atomic_inc(&req->rq_import->imp_replay_inflight);
2914 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2916 ptlrpcd_add_req(req);
2921 * Aborts all in-flight request on import \a imp sending and delayed lists
2923 void ptlrpc_abort_inflight(struct obd_import *imp)
2925 struct list_head *tmp, *n;
2928 /* Make sure that no new requests get processed for this import.
2929 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2930 * this flag and then putting requests on sending_list or delayed_list.
2932 spin_lock(&imp->imp_lock);
2934 /* XXX locking? Maybe we should remove each request with the list
2935 * locked? Also, how do we know if the requests on the list are
2936 * being freed at this time?
2938 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2939 struct ptlrpc_request *req = list_entry(tmp,
2940 struct ptlrpc_request,
2943 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2945 spin_lock(&req->rq_lock);
2946 if (req->rq_import_generation < imp->imp_generation) {
2948 req->rq_status = -EIO;
2949 ptlrpc_client_wake_req(req);
2951 spin_unlock(&req->rq_lock);
2954 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2955 struct ptlrpc_request *req =
2956 list_entry(tmp, struct ptlrpc_request, rq_list);
2958 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2960 spin_lock(&req->rq_lock);
2961 if (req->rq_import_generation < imp->imp_generation) {
2963 req->rq_status = -EIO;
2964 ptlrpc_client_wake_req(req);
2966 spin_unlock(&req->rq_lock);
2969 /* Last chance to free reqs left on the replay list, but we
2970 * will still leak reqs that haven't committed. */
2971 if (imp->imp_replayable)
2972 ptlrpc_free_committed(imp);
2974 spin_unlock(&imp->imp_lock);
2980 * Abort all uncompleted requests in request set \a set
2982 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2984 struct list_head *tmp, *pos;
2986 LASSERT(set != NULL);
2988 list_for_each_safe(pos, tmp, &set->set_requests) {
2989 struct ptlrpc_request *req =
2990 list_entry(pos, struct ptlrpc_request,
2993 spin_lock(&req->rq_lock);
2994 if (req->rq_phase != RQ_PHASE_RPC) {
2995 spin_unlock(&req->rq_lock);
3000 req->rq_status = -EINTR;
3001 ptlrpc_client_wake_req(req);
3002 spin_unlock(&req->rq_lock);
3006 static __u64 ptlrpc_last_xid;
3007 static spinlock_t ptlrpc_last_xid_lock;
3010 * Initialize the XID for the node. This is common among all requests on
3011 * this node, and only requires the property that it is monotonically
3012 * increasing. It does not need to be sequential. Since this is also used
3013 * as the RDMA match bits, it is important that a single client NOT have
3014 * the same match bits for two different in-flight requests, hence we do
3015 * NOT want to have an XID per target or similar.
3017 * To avoid an unlikely collision between match bits after a client reboot
3018 * (which would deliver old data into the wrong RDMA buffer) initialize
3019 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3020 * If the time is clearly incorrect, we instead use a 62-bit random number.
3021 * In the worst case the random number will overflow 1M RPCs per second in
3022 * 9133 years, or permutations thereof.
3024 #define YEAR_2004 (1ULL << 30)
3025 void ptlrpc_init_xid(void)
3027 time_t now = cfs_time_current_sec();
3029 spin_lock_init(&ptlrpc_last_xid_lock);
3030 if (now < YEAR_2004) {
3031 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3032 ptlrpc_last_xid >>= 2;
3033 ptlrpc_last_xid |= (1ULL << 61);
3035 ptlrpc_last_xid = (__u64)now << 20;
3038 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3039 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3040 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3044 * Increase xid and returns resulting new value to the caller.
3046 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3047 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3048 * itself uses the last bulk xid needed, so the server can determine the
3049 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3050 * xid must align to a power-of-two value.
3052 * This is assumed to be true due to the initial ptlrpc_last_xid
3053 * value also being initialized to a power-of-two value. LU-1431
3055 __u64 ptlrpc_next_xid(void)
3059 spin_lock(&ptlrpc_last_xid_lock);
3060 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3061 ptlrpc_last_xid = next;
3062 spin_unlock(&ptlrpc_last_xid_lock);
3068 * If request has a new allocated XID (new request or EINPROGRESS resend),
3069 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3070 * request to ensure previous bulk fails and avoid problems with lost replies
3071 * and therefore several transfers landing into the same buffer from different
3074 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3076 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3078 LASSERT(bd != NULL);
3080 if (!req->rq_resend || req->rq_nr_resend != 0) {
3081 /* this request has a new xid, just use it as bulk matchbits */
3082 req->rq_mbits = req->rq_xid;
3084 } else { /* needs to generate a new matchbits for resend */
3085 __u64 old_mbits = req->rq_mbits;
3087 if ((bd->bd_import->imp_connect_data.ocd_connect_flags &
3088 OBD_CONNECT_BULK_MBITS) != 0)
3089 req->rq_mbits = ptlrpc_next_xid();
3090 else /* old version transfers rq_xid to peer as matchbits */
3091 req->rq_mbits = req->rq_xid = ptlrpc_next_xid();
3093 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
3094 old_mbits, req->rq_mbits);
3097 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3098 * that server can infer the number of bulks that were prepared,
3100 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3105 * Get a glimpse at what next xid value might have been.
3106 * Returns possible next xid.
3108 __u64 ptlrpc_sample_next_xid(void)
3110 #if BITS_PER_LONG == 32
3111 /* need to avoid possible word tearing on 32-bit systems */
3114 spin_lock(&ptlrpc_last_xid_lock);
3115 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3116 spin_unlock(&ptlrpc_last_xid_lock);
3120 /* No need to lock, since returned value is racy anyways */
3121 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3124 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3127 * Functions for operating ptlrpc workers.
3129 * A ptlrpc work is a function which will be running inside ptlrpc context.
3130 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3132 * 1. after a work is created, it can be used many times, that is:
3133 * handler = ptlrpcd_alloc_work();
3134 * ptlrpcd_queue_work();
3136 * queue it again when necessary:
3137 * ptlrpcd_queue_work();
3138 * ptlrpcd_destroy_work();
3139 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3140 * it will only be queued once in any time. Also as its name implies, it may
3141 * have delay before it really runs by ptlrpcd thread.
3143 struct ptlrpc_work_async_args {
3144 int (*cb)(const struct lu_env *, void *);
3148 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3150 /* re-initialize the req */
3151 req->rq_timeout = obd_timeout;
3152 req->rq_sent = cfs_time_current_sec();
3153 req->rq_deadline = req->rq_sent + req->rq_timeout;
3154 req->rq_reply_deadline = req->rq_deadline;
3155 req->rq_phase = RQ_PHASE_INTERPRET;
3156 req->rq_next_phase = RQ_PHASE_COMPLETE;
3157 req->rq_xid = ptlrpc_next_xid();
3158 req->rq_import_generation = req->rq_import->imp_generation;
3160 ptlrpcd_add_req(req);
3163 static int work_interpreter(const struct lu_env *env,
3164 struct ptlrpc_request *req, void *data, int rc)
3166 struct ptlrpc_work_async_args *arg = data;
3168 LASSERT(ptlrpcd_check_work(req));
3169 LASSERT(arg->cb != NULL);
3171 rc = arg->cb(env, arg->cbdata);
3173 list_del_init(&req->rq_set_chain);
3176 if (atomic_dec_return(&req->rq_refcount) > 1) {
3177 atomic_set(&req->rq_refcount, 2);
3178 ptlrpcd_add_work_req(req);
3183 static int worker_format;
3185 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3187 return req->rq_pill.rc_fmt == (void *)&worker_format;
3191 * Create a work for ptlrpc.
3193 void *ptlrpcd_alloc_work(struct obd_import *imp,
3194 int (*cb)(const struct lu_env *, void *), void *cbdata)
3196 struct ptlrpc_request *req = NULL;
3197 struct ptlrpc_work_async_args *args;
3203 RETURN(ERR_PTR(-EINVAL));
3205 /* copy some code from deprecated fakereq. */
3206 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3208 CERROR("ptlrpc: run out of memory!\n");
3209 RETURN(ERR_PTR(-ENOMEM));
3212 ptlrpc_cli_req_init(req);
3214 req->rq_send_state = LUSTRE_IMP_FULL;
3215 req->rq_type = PTL_RPC_MSG_REQUEST;
3216 req->rq_import = class_import_get(imp);
3217 req->rq_interpret_reply = work_interpreter;
3218 /* don't want reply */
3219 req->rq_no_delay = req->rq_no_resend = 1;
3220 req->rq_pill.rc_fmt = (void *)&worker_format;
3222 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3223 args = ptlrpc_req_async_args(req);
3225 args->cbdata = cbdata;
3229 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3231 void ptlrpcd_destroy_work(void *handler)
3233 struct ptlrpc_request *req = handler;
3236 ptlrpc_req_finished(req);
3238 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3240 int ptlrpcd_queue_work(void *handler)
3242 struct ptlrpc_request *req = handler;
3245 * Check if the req is already being queued.
3247 * Here comes a trick: it lacks a way of checking if a req is being
3248 * processed reliably in ptlrpc. Here I have to use refcount of req
3249 * for this purpose. This is okay because the caller should use this
3250 * req as opaque data. - Jinshan
3252 LASSERT(atomic_read(&req->rq_refcount) > 0);
3253 if (atomic_inc_return(&req->rq_refcount) == 2)
3254 ptlrpcd_add_work_req(req);
3257 EXPORT_SYMBOL(ptlrpcd_queue_work);