4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 /** Implementation of client-side PortalRPC interfaces */
39 #define DEBUG_SUBSYSTEM S_RPC
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
48 #include "ptlrpc_internal.h"
50 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_pin_ops = {
51 .add_kiov_frag = ptlrpc_prep_bulk_page_pin,
52 .release_frags = ptlrpc_release_bulk_page_pin,
54 EXPORT_SYMBOL(ptlrpc_bulk_kiov_pin_ops);
56 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kiov_nopin_ops = {
57 .add_kiov_frag = ptlrpc_prep_bulk_page_nopin,
58 .release_frags = ptlrpc_release_bulk_noop,
60 EXPORT_SYMBOL(ptlrpc_bulk_kiov_nopin_ops);
62 const struct ptlrpc_bulk_frag_ops ptlrpc_bulk_kvec_ops = {
63 .add_iov_frag = ptlrpc_prep_bulk_frag,
65 EXPORT_SYMBOL(ptlrpc_bulk_kvec_ops);
67 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
68 static int ptlrpcd_check_work(struct ptlrpc_request *req);
69 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async);
72 * Initialize passed in client structure \a cl.
74 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
75 struct ptlrpc_client *cl)
77 cl->cli_request_portal = req_portal;
78 cl->cli_reply_portal = rep_portal;
81 EXPORT_SYMBOL(ptlrpc_init_client);
84 * Return PortalRPC connection for remore uud \a uuid
86 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
88 struct ptlrpc_connection *c;
90 lnet_process_id_t peer;
93 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
94 * before accessing its values. */
95 /* coverity[uninit_use_in_call] */
96 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
98 CNETERR("cannot find peer %s!\n", uuid->uuid);
102 c = ptlrpc_connection_get(peer, self, uuid);
104 memcpy(c->c_remote_uuid.uuid,
105 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
108 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
114 * Allocate and initialize new bulk descriptor on the sender.
115 * Returns pointer to the descriptor or NULL on error.
117 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned nfrags, unsigned max_brw,
118 enum ptlrpc_bulk_op_type type,
120 const struct ptlrpc_bulk_frag_ops *ops)
122 struct ptlrpc_bulk_desc *desc;
125 /* ensure that only one of KIOV or IOVEC is set but not both */
126 LASSERT((ptlrpc_is_bulk_desc_kiov(type) &&
127 ops->add_kiov_frag != NULL) ||
128 (ptlrpc_is_bulk_desc_kvec(type) &&
129 ops->add_iov_frag != NULL));
131 if (type & PTLRPC_BULK_BUF_KIOV) {
133 offsetof(struct ptlrpc_bulk_desc,
134 bd_u.bd_kiov.bd_vec[nfrags]));
137 offsetof(struct ptlrpc_bulk_desc,
138 bd_u.bd_kvec.bd_kvec[nfrags]));
144 spin_lock_init(&desc->bd_lock);
145 init_waitqueue_head(&desc->bd_waitq);
146 desc->bd_max_iov = nfrags;
147 desc->bd_iov_count = 0;
148 desc->bd_portal = portal;
149 desc->bd_type = type;
150 desc->bd_md_count = 0;
151 desc->bd_frag_ops = (struct ptlrpc_bulk_frag_ops *) ops;
152 LASSERT(max_brw > 0);
153 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
154 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
155 * node. Negotiated ocd_brw_size will always be <= this number. */
156 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
157 LNetInvalidateHandle(&desc->bd_mds[i]);
163 * Prepare bulk descriptor for specified outgoing request \a req that
164 * can fit \a nfrags * pages. \a type is bulk type. \a portal is where
165 * the bulk to be sent. Used on client-side.
166 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
169 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
170 unsigned nfrags, unsigned max_brw,
173 const struct ptlrpc_bulk_frag_ops
176 struct obd_import *imp = req->rq_import;
177 struct ptlrpc_bulk_desc *desc;
180 LASSERT(ptlrpc_is_bulk_op_passive(type));
182 desc = ptlrpc_new_bulk(nfrags, max_brw, type, portal, ops);
186 desc->bd_import_generation = req->rq_import_generation;
187 desc->bd_import = class_import_get(imp);
190 desc->bd_cbid.cbid_fn = client_bulk_callback;
191 desc->bd_cbid.cbid_arg = desc;
193 /* This makes req own desc, and free it when she frees herself */
198 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
200 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
201 struct page *page, int pageoffset, int len,
206 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
207 LASSERT(page != NULL);
208 LASSERT(pageoffset >= 0);
210 LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
211 LASSERT(ptlrpc_is_bulk_desc_kiov(desc->bd_type));
213 kiov = &BD_GET_KIOV(desc, desc->bd_iov_count);
218 page_cache_get(page);
220 kiov->kiov_page = page;
221 kiov->kiov_offset = pageoffset;
222 kiov->kiov_len = len;
224 desc->bd_iov_count++;
226 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
228 int ptlrpc_prep_bulk_frag(struct ptlrpc_bulk_desc *desc,
234 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
235 LASSERT(frag != NULL);
237 LASSERT(ptlrpc_is_bulk_desc_kvec(desc->bd_type));
239 iovec = &BD_GET_KVEC(desc, desc->bd_iov_count);
243 iovec->iov_base = frag;
244 iovec->iov_len = len;
246 desc->bd_iov_count++;
248 RETURN(desc->bd_nob);
250 EXPORT_SYMBOL(ptlrpc_prep_bulk_frag);
252 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
256 LASSERT(desc != NULL);
257 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
258 LASSERT(desc->bd_md_count == 0); /* network hands off */
259 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
260 LASSERT(desc->bd_frag_ops != NULL);
262 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
263 sptlrpc_enc_pool_put_pages(desc);
266 class_export_put(desc->bd_export);
268 class_import_put(desc->bd_import);
270 if (desc->bd_frag_ops->release_frags != NULL)
271 desc->bd_frag_ops->release_frags(desc);
273 if (ptlrpc_is_bulk_desc_kiov(desc->bd_type))
274 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
275 bd_u.bd_kiov.bd_vec[desc->bd_max_iov]));
277 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
278 bd_u.bd_kvec.bd_kvec[desc->
283 EXPORT_SYMBOL(ptlrpc_free_bulk);
286 * Set server timelimit for this req, i.e. how long are we willing to wait
287 * for reply before timing out this request.
289 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
295 LASSERT(req->rq_import);
298 /* non-AT settings */
300 * \a imp_server_timeout means this is reverse import and
301 * we send (currently only) ASTs to the client and cannot afford
302 * to wait too long for the reply, otherwise the other client
303 * (because of which we are sending this request) would
304 * timeout waiting for us
306 req->rq_timeout = req->rq_import->imp_server_timeout ?
307 obd_timeout / 2 : obd_timeout;
309 at = &req->rq_import->imp_at;
310 idx = import_at_get_index(req->rq_import,
311 req->rq_request_portal);
312 serv_est = at_get(&at->iat_service_estimate[idx]);
313 req->rq_timeout = at_est2timeout(serv_est);
315 /* We could get even fancier here, using history to predict increased
318 /* Let the server know what this RPC timeout is by putting it in the
320 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
322 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
324 /* Adjust max service estimate based on server value */
325 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
326 unsigned int serv_est)
332 LASSERT(req->rq_import);
333 at = &req->rq_import->imp_at;
335 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
336 /* max service estimates are tracked on the server side,
337 so just keep minimal history here */
338 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
340 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
341 "has changed from %d to %d\n",
342 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
343 oldse, at_get(&at->iat_service_estimate[idx]));
346 /* Expected network latency per remote node (secs) */
347 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
349 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
352 /* Adjust expected network latency */
353 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
354 unsigned int service_time)
356 unsigned int nl, oldnl;
358 time_t now = cfs_time_current_sec();
360 LASSERT(req->rq_import);
362 if (service_time > now - req->rq_sent + 3) {
363 /* bz16408, however, this can also happen if early reply
364 * is lost and client RPC is expired and resent, early reply
365 * or reply of original RPC can still be fit in reply buffer
366 * of resent RPC, now client is measuring time from the
367 * resent time, but server sent back service time of original
370 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
371 D_ADAPTTO : D_WARNING,
372 "Reported service time %u > total measured time "
373 CFS_DURATION_T"\n", service_time,
374 cfs_time_sub(now, req->rq_sent));
378 /* Network latency is total time less server processing time */
379 nl = max_t(int, now - req->rq_sent -
380 service_time, 0) + 1; /* st rounding */
381 at = &req->rq_import->imp_at;
383 oldnl = at_measured(&at->iat_net_latency, nl);
385 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
386 "has changed from %d to %d\n",
387 req->rq_import->imp_obd->obd_name,
389 &req->rq_import->imp_connection->c_remote_uuid),
390 oldnl, at_get(&at->iat_net_latency));
393 static int unpack_reply(struct ptlrpc_request *req)
397 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
398 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
400 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
405 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
407 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
414 * Handle an early reply message, called with the rq_lock held.
415 * If anything goes wrong just ignore it - same as if it never happened
417 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
418 __must_hold(&req->rq_lock)
420 struct ptlrpc_request *early_req;
426 spin_unlock(&req->rq_lock);
428 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
430 spin_lock(&req->rq_lock);
434 rc = unpack_reply(early_req);
436 sptlrpc_cli_finish_early_reply(early_req);
437 spin_lock(&req->rq_lock);
441 /* Use new timeout value just to adjust the local value for this
442 * request, don't include it into at_history. It is unclear yet why
443 * service time increased and should it be counted or skipped, e.g.
444 * that can be recovery case or some error or server, the real reply
445 * will add all new data if it is worth to add. */
446 req->rq_timeout = lustre_msg_get_timeout(early_req->rq_repmsg);
447 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
449 /* Network latency can be adjusted, it is pure network delays */
450 ptlrpc_at_adj_net_latency(req,
451 lustre_msg_get_service_time(early_req->rq_repmsg));
453 sptlrpc_cli_finish_early_reply(early_req);
455 spin_lock(&req->rq_lock);
456 olddl = req->rq_deadline;
457 /* server assumes it now has rq_timeout from when the request
458 * arrived, so the client should give it at least that long.
459 * since we don't know the arrival time we'll use the original
461 req->rq_deadline = req->rq_sent + req->rq_timeout +
462 ptlrpc_at_get_net_latency(req);
464 DEBUG_REQ(D_ADAPTTO, req,
465 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
466 "("CFS_DURATION_T"s)", req->rq_early_count,
467 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
468 cfs_time_sub(req->rq_deadline, olddl));
473 static struct kmem_cache *request_cache;
475 int ptlrpc_request_cache_init(void)
477 request_cache = kmem_cache_create("ptlrpc_cache",
478 sizeof(struct ptlrpc_request),
479 0, SLAB_HWCACHE_ALIGN, NULL);
480 return request_cache == NULL ? -ENOMEM : 0;
483 void ptlrpc_request_cache_fini(void)
485 kmem_cache_destroy(request_cache);
488 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
490 struct ptlrpc_request *req;
492 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
496 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
498 OBD_SLAB_FREE_PTR(req, request_cache);
502 * Wind down request pool \a pool.
503 * Frees all requests from the pool too
505 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
507 struct list_head *l, *tmp;
508 struct ptlrpc_request *req;
510 LASSERT(pool != NULL);
512 spin_lock(&pool->prp_lock);
513 list_for_each_safe(l, tmp, &pool->prp_req_list) {
514 req = list_entry(l, struct ptlrpc_request, rq_list);
515 list_del(&req->rq_list);
516 LASSERT(req->rq_reqbuf);
517 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
518 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
519 ptlrpc_request_cache_free(req);
521 spin_unlock(&pool->prp_lock);
522 OBD_FREE(pool, sizeof(*pool));
524 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
527 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
529 int ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
534 while (size < pool->prp_rq_size)
537 LASSERTF(list_empty(&pool->prp_req_list) ||
538 size == pool->prp_rq_size,
539 "Trying to change pool size with nonempty pool "
540 "from %d to %d bytes\n", pool->prp_rq_size, size);
542 spin_lock(&pool->prp_lock);
543 pool->prp_rq_size = size;
544 for (i = 0; i < num_rq; i++) {
545 struct ptlrpc_request *req;
546 struct lustre_msg *msg;
548 spin_unlock(&pool->prp_lock);
549 req = ptlrpc_request_cache_alloc(GFP_NOFS);
552 OBD_ALLOC_LARGE(msg, size);
554 ptlrpc_request_cache_free(req);
557 req->rq_reqbuf = msg;
558 req->rq_reqbuf_len = size;
560 spin_lock(&pool->prp_lock);
561 list_add_tail(&req->rq_list, &pool->prp_req_list);
563 spin_unlock(&pool->prp_lock);
566 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
569 * Create and initialize new request pool with given attributes:
570 * \a num_rq - initial number of requests to create for the pool
571 * \a msgsize - maximum message size possible for requests in thid pool
572 * \a populate_pool - function to be called when more requests need to be added
574 * Returns pointer to newly created pool or NULL on error.
576 struct ptlrpc_request_pool *
577 ptlrpc_init_rq_pool(int num_rq, int msgsize,
578 int (*populate_pool)(struct ptlrpc_request_pool *, int))
580 struct ptlrpc_request_pool *pool;
582 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
586 /* Request next power of two for the allocation, because internally
587 kernel would do exactly this */
589 spin_lock_init(&pool->prp_lock);
590 INIT_LIST_HEAD(&pool->prp_req_list);
591 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
592 pool->prp_populate = populate_pool;
594 populate_pool(pool, num_rq);
598 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
601 * Fetches one request from pool \a pool
603 static struct ptlrpc_request *
604 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
606 struct ptlrpc_request *request;
607 struct lustre_msg *reqbuf;
612 spin_lock(&pool->prp_lock);
614 /* See if we have anything in a pool, and bail out if nothing,
615 * in writeout path, where this matters, this is safe to do, because
616 * nothing is lost in this case, and when some in-flight requests
617 * complete, this code will be called again. */
618 if (unlikely(list_empty(&pool->prp_req_list))) {
619 spin_unlock(&pool->prp_lock);
623 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
625 list_del_init(&request->rq_list);
626 spin_unlock(&pool->prp_lock);
628 LASSERT(request->rq_reqbuf);
629 LASSERT(request->rq_pool);
631 reqbuf = request->rq_reqbuf;
632 memset(request, 0, sizeof(*request));
633 request->rq_reqbuf = reqbuf;
634 request->rq_reqbuf_len = pool->prp_rq_size;
635 request->rq_pool = pool;
641 * Returns freed \a request to pool.
643 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
645 struct ptlrpc_request_pool *pool = request->rq_pool;
647 spin_lock(&pool->prp_lock);
648 LASSERT(list_empty(&request->rq_list));
649 LASSERT(!request->rq_receiving_reply);
650 list_add_tail(&request->rq_list, &pool->prp_req_list);
651 spin_unlock(&pool->prp_lock);
654 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
655 __u32 version, int opcode,
656 int count, __u32 *lengths, char **bufs,
657 struct ptlrpc_cli_ctx *ctx)
659 struct obd_import *imp = request->rq_import;
664 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
666 rc = sptlrpc_req_get_ctx(request);
671 sptlrpc_req_set_flavor(request, opcode);
673 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
676 LASSERT(!request->rq_pool);
680 lustre_msg_add_version(request->rq_reqmsg, version);
681 request->rq_send_state = LUSTRE_IMP_FULL;
682 request->rq_type = PTL_RPC_MSG_REQUEST;
684 request->rq_req_cbid.cbid_fn = request_out_callback;
685 request->rq_req_cbid.cbid_arg = request;
687 request->rq_reply_cbid.cbid_fn = reply_in_callback;
688 request->rq_reply_cbid.cbid_arg = request;
690 request->rq_reply_deadline = 0;
691 request->rq_phase = RQ_PHASE_NEW;
692 request->rq_next_phase = RQ_PHASE_UNDEFINED;
694 request->rq_request_portal = imp->imp_client->cli_request_portal;
695 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
697 ptlrpc_at_set_req_timeout(request);
699 lustre_msg_set_opc(request->rq_reqmsg, opcode);
703 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
705 class_import_put(imp);
709 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
710 __u32 version, int opcode, char **bufs,
711 struct ptlrpc_cli_ctx *ctx)
715 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
716 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
717 request->rq_pill.rc_area[RCL_CLIENT],
720 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
723 * Pack request buffers for network transfer, performing necessary encryption
724 * steps if necessary.
726 int ptlrpc_request_pack(struct ptlrpc_request *request,
727 __u32 version, int opcode)
730 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
734 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
735 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
736 * have to send old ptlrpc_body to keep interoprability with these
739 * Only three kinds of server->client RPCs so far:
744 * XXX This should be removed whenever we drop the interoprability with
745 * the these old clients.
747 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
748 opcode == LDLM_GL_CALLBACK)
749 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
750 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
754 EXPORT_SYMBOL(ptlrpc_request_pack);
757 * Helper function to allocate new request on import \a imp
758 * and possibly using existing request from pool \a pool if provided.
759 * Returns allocated request structure with import field filled or
763 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
764 struct ptlrpc_request_pool *pool)
766 struct ptlrpc_request *request = NULL;
768 request = ptlrpc_request_cache_alloc(GFP_NOFS);
770 if (!request && pool)
771 request = ptlrpc_prep_req_from_pool(pool);
774 ptlrpc_cli_req_init(request);
776 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
777 LASSERT(imp != LP_POISON);
778 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
780 LASSERT(imp->imp_client != LP_POISON);
782 request->rq_import = class_import_get(imp);
784 CERROR("request allocation out of memory\n");
791 * Helper function for creating a request.
792 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
793 * buffer structures according to capsule template \a format.
794 * Returns allocated request structure pointer or NULL on error.
796 static struct ptlrpc_request *
797 ptlrpc_request_alloc_internal(struct obd_import *imp,
798 struct ptlrpc_request_pool * pool,
799 const struct req_format *format)
801 struct ptlrpc_request *request;
803 request = __ptlrpc_request_alloc(imp, pool);
807 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
808 req_capsule_set(&request->rq_pill, format);
813 * Allocate new request structure for import \a imp and initialize its
814 * buffer structure according to capsule template \a format.
816 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
817 const struct req_format *format)
819 return ptlrpc_request_alloc_internal(imp, NULL, format);
821 EXPORT_SYMBOL(ptlrpc_request_alloc);
824 * Allocate new request structure for import \a imp from pool \a pool and
825 * initialize its buffer structure according to capsule template \a format.
827 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
828 struct ptlrpc_request_pool * pool,
829 const struct req_format *format)
831 return ptlrpc_request_alloc_internal(imp, pool, format);
833 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
836 * For requests not from pool, free memory of the request structure.
837 * For requests obtained from a pool earlier, return request back to pool.
839 void ptlrpc_request_free(struct ptlrpc_request *request)
841 if (request->rq_pool)
842 __ptlrpc_free_req_to_pool(request);
844 ptlrpc_request_cache_free(request);
846 EXPORT_SYMBOL(ptlrpc_request_free);
849 * Allocate new request for operatione \a opcode and immediatelly pack it for
851 * Only used for simple requests like OBD_PING where the only important
852 * part of the request is operation itself.
853 * Returns allocated request or NULL on error.
855 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
856 const struct req_format *format,
857 __u32 version, int opcode)
859 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
863 rc = ptlrpc_request_pack(req, version, opcode);
865 ptlrpc_request_free(req);
871 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
874 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
875 * for operation \a opcode. Request would contain \a count buffers.
876 * Sizes of buffers are described in array \a lengths and buffers themselves
877 * are provided by a pointer \a bufs.
878 * Returns prepared request structure pointer or NULL on error.
880 struct ptlrpc_request *
881 ptlrpc_prep_req_pool(struct obd_import *imp,
882 __u32 version, int opcode,
883 int count, __u32 *lengths, char **bufs,
884 struct ptlrpc_request_pool *pool)
886 struct ptlrpc_request *request;
889 request = __ptlrpc_request_alloc(imp, pool);
893 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
894 lengths, bufs, NULL);
896 ptlrpc_request_free(request);
903 * Same as ptlrpc_prep_req_pool, but without pool
905 struct ptlrpc_request *
906 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
907 __u32 *lengths, char **bufs)
909 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
914 * Allocate and initialize new request set structure on the current CPT.
915 * Returns a pointer to the newly allocated set structure or NULL on error.
917 struct ptlrpc_request_set *ptlrpc_prep_set(void)
919 struct ptlrpc_request_set *set;
923 cpt = cfs_cpt_current(cfs_cpt_table, 0);
924 OBD_CPT_ALLOC(set, cfs_cpt_table, cpt, sizeof *set);
927 atomic_set(&set->set_refcount, 1);
928 INIT_LIST_HEAD(&set->set_requests);
929 init_waitqueue_head(&set->set_waitq);
930 atomic_set(&set->set_new_count, 0);
931 atomic_set(&set->set_remaining, 0);
932 spin_lock_init(&set->set_new_req_lock);
933 INIT_LIST_HEAD(&set->set_new_requests);
934 INIT_LIST_HEAD(&set->set_cblist);
935 set->set_max_inflight = UINT_MAX;
936 set->set_producer = NULL;
937 set->set_producer_arg = NULL;
942 EXPORT_SYMBOL(ptlrpc_prep_set);
945 * Allocate and initialize new request set structure with flow control
946 * extension. This extension allows to control the number of requests in-flight
947 * for the whole set. A callback function to generate requests must be provided
948 * and the request set will keep the number of requests sent over the wire to
950 * Returns a pointer to the newly allocated set structure or NULL on error.
952 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
956 struct ptlrpc_request_set *set;
958 set = ptlrpc_prep_set();
962 set->set_max_inflight = max;
963 set->set_producer = func;
964 set->set_producer_arg = arg;
970 * Wind down and free request set structure previously allocated with
972 * Ensures that all requests on the set have completed and removes
973 * all requests from the request list in a set.
974 * If any unsent request happen to be on the list, pretends that they got
975 * an error in flight and calls their completion handler.
977 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
979 struct list_head *tmp;
980 struct list_head *next;
985 /* Requests on the set should either all be completed, or all be new */
986 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
987 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
988 list_for_each(tmp, &set->set_requests) {
989 struct ptlrpc_request *req =
990 list_entry(tmp, struct ptlrpc_request,
993 LASSERT(req->rq_phase == expected_phase);
997 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
998 atomic_read(&set->set_remaining) == n, "%d / %d\n",
999 atomic_read(&set->set_remaining), n);
1001 list_for_each_safe(tmp, next, &set->set_requests) {
1002 struct ptlrpc_request *req =
1003 list_entry(tmp, struct ptlrpc_request,
1005 list_del_init(&req->rq_set_chain);
1007 LASSERT(req->rq_phase == expected_phase);
1009 if (req->rq_phase == RQ_PHASE_NEW) {
1010 ptlrpc_req_interpret(NULL, req, -EBADR);
1011 atomic_dec(&set->set_remaining);
1014 spin_lock(&req->rq_lock);
1016 req->rq_invalid_rqset = 0;
1017 spin_unlock(&req->rq_lock);
1019 ptlrpc_req_finished (req);
1022 LASSERT(atomic_read(&set->set_remaining) == 0);
1024 ptlrpc_reqset_put(set);
1027 EXPORT_SYMBOL(ptlrpc_set_destroy);
1030 * Add a callback function \a fn to the set.
1031 * This function would be called when all requests on this set are completed.
1032 * The function will be passed \a data argument.
1034 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1035 set_interpreter_func fn, void *data)
1037 struct ptlrpc_set_cbdata *cbdata;
1039 OBD_ALLOC_PTR(cbdata);
1043 cbdata->psc_interpret = fn;
1044 cbdata->psc_data = data;
1045 list_add_tail(&cbdata->psc_item, &set->set_cblist);
1051 * Add a new request to the general purpose request set.
1052 * Assumes request reference from the caller.
1054 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1055 struct ptlrpc_request *req)
1057 LASSERT(list_empty(&req->rq_set_chain));
1059 /* The set takes over the caller's request reference */
1060 list_add_tail(&req->rq_set_chain, &set->set_requests);
1062 atomic_inc(&set->set_remaining);
1063 req->rq_queued_time = cfs_time_current();
1065 if (req->rq_reqmsg != NULL)
1066 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1068 if (set->set_producer != NULL)
1069 /* If the request set has a producer callback, the RPC must be
1070 * sent straight away */
1071 ptlrpc_send_new_req(req);
1073 EXPORT_SYMBOL(ptlrpc_set_add_req);
1076 * Add a request to a request with dedicated server thread
1077 * and wake the thread to make any necessary processing.
1078 * Currently only used for ptlrpcd.
1080 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1081 struct ptlrpc_request *req)
1083 struct ptlrpc_request_set *set = pc->pc_set;
1086 LASSERT(req->rq_set == NULL);
1087 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1089 spin_lock(&set->set_new_req_lock);
1091 * The set takes over the caller's request reference.
1094 req->rq_queued_time = cfs_time_current();
1095 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1096 count = atomic_inc_return(&set->set_new_count);
1097 spin_unlock(&set->set_new_req_lock);
1099 /* Only need to call wakeup once for the first entry. */
1101 wake_up(&set->set_waitq);
1103 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1104 * guarantee the async RPC can be processed ASAP, we have
1105 * no other better choice. It maybe fixed in future. */
1106 for (i = 0; i < pc->pc_npartners; i++)
1107 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1112 * Based on the current state of the import, determine if the request
1113 * can be sent, is an error, or should be delayed.
1115 * Returns true if this request should be delayed. If false, and
1116 * *status is set, then the request can not be sent and *status is the
1117 * error code. If false and status is 0, then request can be sent.
1119 * The imp->imp_lock must be held.
1121 static int ptlrpc_import_delay_req(struct obd_import *imp,
1122 struct ptlrpc_request *req, int *status)
1127 LASSERT (status != NULL);
1130 if (req->rq_ctx_init || req->rq_ctx_fini) {
1131 /* always allow ctx init/fini rpc go through */
1132 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1133 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1135 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1136 /* pings may safely race with umount */
1137 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1138 D_HA : D_ERROR, req, "IMP_CLOSED ");
1140 } else if (ptlrpc_send_limit_expired(req)) {
1141 /* probably doesn't need to be a D_ERROR after initial testing*/
1142 DEBUG_REQ(D_HA, req, "send limit expired ");
1143 *status = -ETIMEDOUT;
1144 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1145 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1146 /* allow CONNECT even if import is invalid */ ;
1147 if (atomic_read(&imp->imp_inval_count) != 0) {
1148 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1151 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1152 if (!imp->imp_deactive)
1153 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1154 *status = -ESHUTDOWN; /* bz 12940 */
1155 } else if (req->rq_import_generation != imp->imp_generation) {
1156 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1158 } else if (req->rq_send_state != imp->imp_state) {
1159 /* invalidate in progress - any requests should be drop */
1160 if (atomic_read(&imp->imp_inval_count) != 0) {
1161 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1163 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1164 *status = -EWOULDBLOCK;
1165 } else if (req->rq_allow_replay &&
1166 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1167 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1168 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1169 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1170 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1180 * Decide if the error message should be printed to the console or not.
1181 * Makes its decision based on request type, status, and failure frequency.
1183 * \param[in] req request that failed and may need a console message
1185 * \retval false if no message should be printed
1186 * \retval true if console message should be printed
1188 static bool ptlrpc_console_allow(struct ptlrpc_request *req)
1192 LASSERT(req->rq_reqmsg != NULL);
1193 opc = lustre_msg_get_opc(req->rq_reqmsg);
1195 /* Suppress particular reconnect errors which are to be expected. */
1196 if (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT) {
1199 /* Suppress timed out reconnect requests */
1200 if (lustre_handle_is_used(&req->rq_import->imp_remote_handle) ||
1204 /* Suppress most unavailable/again reconnect requests, but
1205 * print occasionally so it is clear client is trying to
1206 * connect to a server where no target is running. */
1207 err = lustre_msg_get_status(req->rq_repmsg);
1208 if ((err == -ENODEV || err == -EAGAIN) &&
1209 req->rq_import->imp_conn_cnt % 30 != 20)
1217 * Check request processing status.
1218 * Returns the status.
1220 static int ptlrpc_check_status(struct ptlrpc_request *req)
1225 err = lustre_msg_get_status(req->rq_repmsg);
1226 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1227 struct obd_import *imp = req->rq_import;
1228 lnet_nid_t nid = imp->imp_connection->c_peer.nid;
1229 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1231 if (ptlrpc_console_allow(req))
1232 LCONSOLE_ERROR_MSG(0x11, "%s: operation %s to node %s "
1233 "failed: rc = %d\n",
1234 imp->imp_obd->obd_name,
1236 libcfs_nid2str(nid), err);
1237 RETURN(err < 0 ? err : -EINVAL);
1241 DEBUG_REQ(D_INFO, req, "status is %d", err);
1242 } else if (err > 0) {
1243 /* XXX: translate this error from net to host */
1244 DEBUG_REQ(D_INFO, req, "status is %d", err);
1251 * save pre-versions of objects into request for replay.
1252 * Versions are obtained from server reply.
1255 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1257 struct lustre_msg *repmsg = req->rq_repmsg;
1258 struct lustre_msg *reqmsg = req->rq_reqmsg;
1259 __u64 *versions = lustre_msg_get_versions(repmsg);
1262 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1266 lustre_msg_set_versions(reqmsg, versions);
1267 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1268 versions[0], versions[1]);
1274 * Callback function called when client receives RPC reply for \a req.
1275 * Returns 0 on success or error code.
1276 * The return alue would be assigned to req->rq_status by the caller
1277 * as request processing status.
1278 * This function also decides if the request needs to be saved for later replay.
1280 static int after_reply(struct ptlrpc_request *req)
1282 struct obd_import *imp = req->rq_import;
1283 struct obd_device *obd = req->rq_import->imp_obd;
1285 struct timeval work_start;
1289 LASSERT(obd != NULL);
1290 /* repbuf must be unlinked */
1291 LASSERT(!req->rq_receiving_reply && req->rq_reply_unlinked);
1293 if (req->rq_reply_truncated) {
1294 if (ptlrpc_no_resend(req)) {
1295 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1296 " expected: %d, actual size: %d",
1297 req->rq_nob_received, req->rq_repbuf_len);
1301 sptlrpc_cli_free_repbuf(req);
1302 /* Pass the required reply buffer size (include
1303 * space for early reply).
1304 * NB: no need to roundup because alloc_repbuf
1305 * will roundup it */
1306 req->rq_replen = req->rq_nob_received;
1307 req->rq_nob_received = 0;
1308 spin_lock(&req->rq_lock);
1310 spin_unlock(&req->rq_lock);
1314 do_gettimeofday(&work_start);
1315 timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1318 * NB Until this point, the whole of the incoming message,
1319 * including buflens, status etc is in the sender's byte order.
1321 rc = sptlrpc_cli_unwrap_reply(req);
1323 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1328 * Security layer unwrap might ask resend this request.
1333 rc = unpack_reply(req);
1337 /* retry indefinitely on EINPROGRESS */
1338 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1339 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1340 time_t now = cfs_time_current_sec();
1342 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1343 spin_lock(&req->rq_lock);
1345 spin_unlock(&req->rq_lock);
1346 req->rq_nr_resend++;
1348 /* Readjust the timeout for current conditions */
1349 ptlrpc_at_set_req_timeout(req);
1350 /* delay resend to give a chance to the server to get ready.
1351 * The delay is increased by 1s on every resend and is capped to
1352 * the current request timeout (i.e. obd_timeout if AT is off,
1353 * or AT service time x 125% + 5s, see at_est2timeout) */
1354 if (req->rq_nr_resend > req->rq_timeout)
1355 req->rq_sent = now + req->rq_timeout;
1357 req->rq_sent = now + req->rq_nr_resend;
1362 if (obd->obd_svc_stats != NULL) {
1363 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1365 ptlrpc_lprocfs_rpc_sent(req, timediff);
1368 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1369 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1370 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1371 lustre_msg_get_type(req->rq_repmsg));
1375 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1376 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1377 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1378 ptlrpc_at_adj_net_latency(req,
1379 lustre_msg_get_service_time(req->rq_repmsg));
1381 rc = ptlrpc_check_status(req);
1382 imp->imp_connect_error = rc;
1386 * Either we've been evicted, or the server has failed for
1387 * some reason. Try to reconnect, and if that fails, punt to
1390 if (ptlrpc_recoverable_error(rc)) {
1391 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1392 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1395 ptlrpc_request_handle_notconn(req);
1400 * Let's look if server sent slv. Do it only for RPC with
1403 ldlm_cli_update_pool(req);
1407 * Store transno in reqmsg for replay.
1409 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1410 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1411 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1414 if (imp->imp_replayable) {
1415 spin_lock(&imp->imp_lock);
1417 * No point in adding already-committed requests to the replay
1418 * list, we will just remove them immediately. b=9829
1420 if (req->rq_transno != 0 &&
1422 lustre_msg_get_last_committed(req->rq_repmsg) ||
1424 /** version recovery */
1425 ptlrpc_save_versions(req);
1426 ptlrpc_retain_replayable_request(req, imp);
1427 } else if (req->rq_commit_cb != NULL &&
1428 list_empty(&req->rq_replay_list)) {
1429 /* NB: don't call rq_commit_cb if it's already on
1430 * rq_replay_list, ptlrpc_free_committed() will call
1431 * it later, see LU-3618 for details */
1432 spin_unlock(&imp->imp_lock);
1433 req->rq_commit_cb(req);
1434 spin_lock(&imp->imp_lock);
1438 * Replay-enabled imports return commit-status information.
1440 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1441 imp->imp_peer_committed_transno =
1442 lustre_msg_get_last_committed(req->rq_repmsg);
1445 ptlrpc_free_committed(imp);
1447 if (!list_empty(&imp->imp_replay_list)) {
1448 struct ptlrpc_request *last;
1450 last = list_entry(imp->imp_replay_list.prev,
1451 struct ptlrpc_request,
1454 * Requests with rq_replay stay on the list even if no
1455 * commit is expected.
1457 if (last->rq_transno > imp->imp_peer_committed_transno)
1458 ptlrpc_pinger_commit_expected(imp);
1461 spin_unlock(&imp->imp_lock);
1468 * Helper function to send request \a req over the network for the first time
1469 * Also adjusts request phase.
1470 * Returns 0 on success or error code.
1472 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1474 struct obd_import *imp = req->rq_import;
1475 struct list_head *tmp;
1476 __u64 min_xid = ~0ULL;
1480 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1481 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1482 (!req->rq_generation_set ||
1483 req->rq_import_generation == imp->imp_generation))
1486 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1488 spin_lock(&imp->imp_lock);
1490 /* the very first time we assign XID. it's important to assign XID
1491 * and put it on the list atomically, so that the lowest assigned
1492 * XID is always known. this is vital for multislot last_rcvd */
1493 if (req->rq_send_state == LUSTRE_IMP_REPLAY) {
1494 LASSERT(req->rq_xid != 0);
1496 LASSERT(req->rq_xid == 0);
1497 req->rq_xid = ptlrpc_next_xid();
1500 if (!req->rq_generation_set)
1501 req->rq_import_generation = imp->imp_generation;
1503 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1504 spin_lock(&req->rq_lock);
1505 req->rq_waiting = 1;
1506 spin_unlock(&req->rq_lock);
1508 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1509 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1510 ptlrpc_import_state_name(req->rq_send_state),
1511 ptlrpc_import_state_name(imp->imp_state));
1512 LASSERT(list_empty(&req->rq_list));
1513 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1514 atomic_inc(&req->rq_import->imp_inflight);
1515 spin_unlock(&imp->imp_lock);
1520 spin_unlock(&imp->imp_lock);
1521 req->rq_status = rc;
1522 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1526 LASSERT(list_empty(&req->rq_list));
1527 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1528 atomic_inc(&req->rq_import->imp_inflight);
1530 /* find the lowest unreplied XID */
1531 list_for_each(tmp, &imp->imp_delayed_list) {
1532 struct ptlrpc_request *r;
1533 r = list_entry(tmp, struct ptlrpc_request, rq_list);
1534 if (r->rq_xid < min_xid)
1535 min_xid = r->rq_xid;
1537 list_for_each(tmp, &imp->imp_sending_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 spin_unlock(&imp->imp_lock);
1545 if (likely(min_xid != ~0ULL))
1546 lustre_msg_set_last_xid(req->rq_reqmsg, min_xid - 1);
1548 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1550 rc = sptlrpc_req_refresh_ctx(req, -1);
1553 req->rq_status = rc;
1556 spin_lock(&req->rq_lock);
1557 req->rq_wait_ctx = 1;
1558 spin_unlock(&req->rq_lock);
1563 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1564 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1565 imp->imp_obd->obd_uuid.uuid,
1566 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1567 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1568 lustre_msg_get_opc(req->rq_reqmsg));
1570 rc = ptl_send_rpc(req, 0);
1572 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1573 spin_lock(&req->rq_lock);
1574 req->rq_net_err = 1;
1575 spin_unlock(&req->rq_lock);
1581 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1586 LASSERT(set->set_producer != NULL);
1588 remaining = atomic_read(&set->set_remaining);
1590 /* populate the ->set_requests list with requests until we
1591 * reach the maximum number of RPCs in flight for this set */
1592 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1593 rc = set->set_producer(set, set->set_producer_arg);
1594 if (rc == -ENOENT) {
1595 /* no more RPC to produce */
1596 set->set_producer = NULL;
1597 set->set_producer_arg = NULL;
1602 RETURN((atomic_read(&set->set_remaining) - remaining));
1606 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1607 * and no more replies are expected.
1608 * (it is possible to get less replies than requests sent e.g. due to timed out
1609 * requests or requests that we had trouble to send out)
1611 * NOTE: This function contains a potential schedule point (cond_resched()).
1613 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1615 struct list_head *tmp, *next;
1616 struct list_head comp_reqs;
1617 int force_timer_recalc = 0;
1620 if (atomic_read(&set->set_remaining) == 0)
1623 INIT_LIST_HEAD(&comp_reqs);
1624 list_for_each_safe(tmp, next, &set->set_requests) {
1625 struct ptlrpc_request *req =
1626 list_entry(tmp, struct ptlrpc_request,
1628 struct obd_import *imp = req->rq_import;
1629 int unregistered = 0;
1632 /* This schedule point is mainly for the ptlrpcd caller of this
1633 * function. Most ptlrpc sets are not long-lived and unbounded
1634 * in length, but at the least the set used by the ptlrpcd is.
1635 * Since the processing time is unbounded, we need to insert an
1636 * explicit schedule point to make the thread well-behaved.
1640 if (req->rq_phase == RQ_PHASE_NEW &&
1641 ptlrpc_send_new_req(req)) {
1642 force_timer_recalc = 1;
1645 /* delayed send - skip */
1646 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1649 /* delayed resend - skip */
1650 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1651 req->rq_sent > cfs_time_current_sec())
1654 if (!(req->rq_phase == RQ_PHASE_RPC ||
1655 req->rq_phase == RQ_PHASE_BULK ||
1656 req->rq_phase == RQ_PHASE_INTERPRET ||
1657 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1658 req->rq_phase == RQ_PHASE_COMPLETE)) {
1659 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1663 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1664 LASSERT(req->rq_next_phase != req->rq_phase);
1665 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1668 * Skip processing until reply is unlinked. We
1669 * can't return to pool before that and we can't
1670 * call interpret before that. We need to make
1671 * sure that all rdma transfers finished and will
1672 * not corrupt any data.
1674 if (ptlrpc_client_recv_or_unlink(req) ||
1675 ptlrpc_client_bulk_active(req))
1679 * Turn fail_loc off to prevent it from looping
1682 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1683 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1686 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1687 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1692 * Move to next phase if reply was successfully
1695 ptlrpc_rqphase_move(req, req->rq_next_phase);
1698 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1699 list_move_tail(&req->rq_set_chain, &comp_reqs);
1703 if (req->rq_phase == RQ_PHASE_INTERPRET)
1704 GOTO(interpret, req->rq_status);
1707 * Note that this also will start async reply unlink.
1709 if (req->rq_net_err && !req->rq_timedout) {
1710 ptlrpc_expire_one_request(req, 1);
1713 * Check if we still need to wait for unlink.
1715 if (ptlrpc_client_recv_or_unlink(req) ||
1716 ptlrpc_client_bulk_active(req))
1718 /* If there is no need to resend, fail it now. */
1719 if (req->rq_no_resend) {
1720 if (req->rq_status == 0)
1721 req->rq_status = -EIO;
1722 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1723 GOTO(interpret, req->rq_status);
1730 spin_lock(&req->rq_lock);
1731 req->rq_replied = 0;
1732 spin_unlock(&req->rq_lock);
1733 if (req->rq_status == 0)
1734 req->rq_status = -EIO;
1735 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1736 GOTO(interpret, req->rq_status);
1739 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1740 * so it sets rq_intr regardless of individual rpc
1741 * timeouts. The synchronous IO waiting path sets
1742 * rq_intr irrespective of whether ptlrpcd
1743 * has seen a timeout. Our policy is to only interpret
1744 * interrupted rpcs after they have timed out, so we
1745 * need to enforce that here.
1748 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1749 req->rq_wait_ctx)) {
1750 req->rq_status = -EINTR;
1751 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1752 GOTO(interpret, req->rq_status);
1755 if (req->rq_phase == RQ_PHASE_RPC) {
1756 if (req->rq_timedout || req->rq_resend ||
1757 req->rq_waiting || req->rq_wait_ctx) {
1760 if (!ptlrpc_unregister_reply(req, 1)) {
1761 ptlrpc_unregister_bulk(req, 1);
1765 spin_lock(&imp->imp_lock);
1766 if (ptlrpc_import_delay_req(imp, req, &status)){
1767 /* put on delay list - only if we wait
1768 * recovery finished - before send */
1769 list_del_init(&req->rq_list);
1770 list_add_tail(&req->rq_list,
1773 spin_unlock(&imp->imp_lock);
1778 req->rq_status = status;
1779 ptlrpc_rqphase_move(req,
1780 RQ_PHASE_INTERPRET);
1781 spin_unlock(&imp->imp_lock);
1782 GOTO(interpret, req->rq_status);
1784 if (ptlrpc_no_resend(req) &&
1785 !req->rq_wait_ctx) {
1786 req->rq_status = -ENOTCONN;
1787 ptlrpc_rqphase_move(req,
1788 RQ_PHASE_INTERPRET);
1789 spin_unlock(&imp->imp_lock);
1790 GOTO(interpret, req->rq_status);
1793 list_del_init(&req->rq_list);
1794 list_add_tail(&req->rq_list,
1795 &imp->imp_sending_list);
1797 spin_unlock(&imp->imp_lock);
1799 spin_lock(&req->rq_lock);
1800 req->rq_waiting = 0;
1801 spin_unlock(&req->rq_lock);
1803 if (req->rq_timedout || req->rq_resend) {
1804 /* This is re-sending anyways,
1805 * let's mark req as resend. */
1806 spin_lock(&req->rq_lock);
1808 spin_unlock(&req->rq_lock);
1810 if (req->rq_bulk != NULL &&
1811 !ptlrpc_unregister_bulk(req, 1))
1815 * rq_wait_ctx is only touched by ptlrpcd,
1816 * so no lock is needed here.
1818 status = sptlrpc_req_refresh_ctx(req, -1);
1821 req->rq_status = status;
1822 spin_lock(&req->rq_lock);
1823 req->rq_wait_ctx = 0;
1824 spin_unlock(&req->rq_lock);
1825 force_timer_recalc = 1;
1827 spin_lock(&req->rq_lock);
1828 req->rq_wait_ctx = 1;
1829 spin_unlock(&req->rq_lock);
1834 spin_lock(&req->rq_lock);
1835 req->rq_wait_ctx = 0;
1836 spin_unlock(&req->rq_lock);
1839 rc = ptl_send_rpc(req, 0);
1841 DEBUG_REQ(D_HA, req,
1842 "send failed: rc = %d", rc);
1843 force_timer_recalc = 1;
1844 spin_lock(&req->rq_lock);
1845 req->rq_net_err = 1;
1846 spin_unlock(&req->rq_lock);
1849 /* need to reset the timeout */
1850 force_timer_recalc = 1;
1853 spin_lock(&req->rq_lock);
1855 if (ptlrpc_client_early(req)) {
1856 ptlrpc_at_recv_early_reply(req);
1857 spin_unlock(&req->rq_lock);
1861 /* Still waiting for a reply? */
1862 if (ptlrpc_client_recv(req)) {
1863 spin_unlock(&req->rq_lock);
1867 /* Did we actually receive a reply? */
1868 if (!ptlrpc_client_replied(req)) {
1869 spin_unlock(&req->rq_lock);
1873 spin_unlock(&req->rq_lock);
1875 /* unlink from net because we are going to
1876 * swab in-place of reply buffer */
1877 unregistered = ptlrpc_unregister_reply(req, 1);
1881 req->rq_status = after_reply(req);
1885 /* If there is no bulk associated with this request,
1886 * then we're done and should let the interpreter
1887 * process the reply. Similarly if the RPC returned
1888 * an error, and therefore the bulk will never arrive.
1890 if (req->rq_bulk == NULL || req->rq_status < 0) {
1891 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1892 GOTO(interpret, req->rq_status);
1895 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1898 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1899 if (ptlrpc_client_bulk_active(req))
1902 if (req->rq_bulk->bd_failure) {
1903 /* The RPC reply arrived OK, but the bulk screwed
1904 * up! Dead weird since the server told us the RPC
1905 * was good after getting the REPLY for her GET or
1906 * the ACK for her PUT. */
1907 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1908 req->rq_status = -EIO;
1911 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1914 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1916 /* This moves to "unregistering" phase we need to wait for
1918 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1919 /* start async bulk unlink too */
1920 ptlrpc_unregister_bulk(req, 1);
1924 if (!ptlrpc_unregister_bulk(req, 1))
1927 /* When calling interpret receiving already should be
1929 LASSERT(!req->rq_receiving_reply);
1931 ptlrpc_req_interpret(env, req, req->rq_status);
1933 if (ptlrpcd_check_work(req)) {
1934 atomic_dec(&set->set_remaining);
1937 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1939 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1940 "Completed RPC pname:cluuid:pid:xid:nid:"
1941 "opc %s:%s:%d:"LPU64":%s:%d\n",
1942 current_comm(), imp->imp_obd->obd_uuid.uuid,
1943 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1944 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1945 lustre_msg_get_opc(req->rq_reqmsg));
1947 spin_lock(&imp->imp_lock);
1948 /* Request already may be not on sending or delaying list. This
1949 * may happen in the case of marking it erroneous for the case
1950 * ptlrpc_import_delay_req(req, status) find it impossible to
1951 * allow sending this rpc and returns *status != 0. */
1952 if (!list_empty(&req->rq_list)) {
1953 list_del_init(&req->rq_list);
1954 atomic_dec(&imp->imp_inflight);
1956 spin_unlock(&imp->imp_lock);
1958 atomic_dec(&set->set_remaining);
1959 wake_up_all(&imp->imp_recovery_waitq);
1961 if (set->set_producer) {
1962 /* produce a new request if possible */
1963 if (ptlrpc_set_producer(set) > 0)
1964 force_timer_recalc = 1;
1966 /* free the request that has just been completed
1967 * in order not to pollute set->set_requests */
1968 list_del_init(&req->rq_set_chain);
1969 spin_lock(&req->rq_lock);
1971 req->rq_invalid_rqset = 0;
1972 spin_unlock(&req->rq_lock);
1974 /* record rq_status to compute the final status later */
1975 if (req->rq_status != 0)
1976 set->set_rc = req->rq_status;
1977 ptlrpc_req_finished(req);
1979 list_move_tail(&req->rq_set_chain, &comp_reqs);
1983 /* move completed request at the head of list so it's easier for
1984 * caller to find them */
1985 list_splice(&comp_reqs, &set->set_requests);
1987 /* If we hit an error, we want to recover promptly. */
1988 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1990 EXPORT_SYMBOL(ptlrpc_check_set);
1993 * Time out request \a req. is \a async_unlink is set, that means do not wait
1994 * until LNet actually confirms network buffer unlinking.
1995 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1997 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1999 struct obd_import *imp = req->rq_import;
2003 spin_lock(&req->rq_lock);
2004 req->rq_timedout = 1;
2005 spin_unlock(&req->rq_lock);
2007 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
2008 "/real "CFS_DURATION_T"]",
2009 req->rq_net_err ? "failed due to network error" :
2010 ((req->rq_real_sent == 0 ||
2011 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
2012 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
2013 "timed out for sent delay" : "timed out for slow reply"),
2014 req->rq_sent, req->rq_real_sent);
2016 if (imp != NULL && obd_debug_peer_on_timeout)
2017 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
2019 ptlrpc_unregister_reply(req, async_unlink);
2020 ptlrpc_unregister_bulk(req, async_unlink);
2022 if (obd_dump_on_timeout)
2023 libcfs_debug_dumplog();
2026 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
2030 atomic_inc(&imp->imp_timeouts);
2032 /* The DLM server doesn't want recovery run on its imports. */
2033 if (imp->imp_dlm_fake)
2036 /* If this request is for recovery or other primordial tasks,
2037 * then error it out here. */
2038 if (req->rq_ctx_init || req->rq_ctx_fini ||
2039 req->rq_send_state != LUSTRE_IMP_FULL ||
2040 imp->imp_obd->obd_no_recov) {
2041 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
2042 ptlrpc_import_state_name(req->rq_send_state),
2043 ptlrpc_import_state_name(imp->imp_state));
2044 spin_lock(&req->rq_lock);
2045 req->rq_status = -ETIMEDOUT;
2047 spin_unlock(&req->rq_lock);
2051 /* if a request can't be resent we can't wait for an answer after
2053 if (ptlrpc_no_resend(req)) {
2054 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
2058 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
2064 * Time out all uncompleted requests in request set pointed by \a data
2065 * Callback used when waiting on sets with l_wait_event.
2068 int ptlrpc_expired_set(void *data)
2070 struct ptlrpc_request_set *set = data;
2071 struct list_head *tmp;
2072 time_t now = cfs_time_current_sec();
2075 LASSERT(set != NULL);
2078 * A timeout expired. See which reqs it applies to...
2080 list_for_each(tmp, &set->set_requests) {
2081 struct ptlrpc_request *req =
2082 list_entry(tmp, struct ptlrpc_request,
2085 /* don't expire request waiting for context */
2086 if (req->rq_wait_ctx)
2089 /* Request in-flight? */
2090 if (!((req->rq_phase == RQ_PHASE_RPC &&
2091 !req->rq_waiting && !req->rq_resend) ||
2092 (req->rq_phase == RQ_PHASE_BULK)))
2095 if (req->rq_timedout || /* already dealt with */
2096 req->rq_deadline > now) /* not expired */
2099 /* Deal with this guy. Do it asynchronously to not block
2100 * ptlrpcd thread. */
2101 ptlrpc_expire_one_request(req, 1);
2105 * When waiting for a whole set, we always break out of the
2106 * sleep so we can recalculate the timeout, or enable interrupts
2107 * if everyone's timed out.
2113 * Sets rq_intr flag in \a req under spinlock.
2115 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2117 spin_lock(&req->rq_lock);
2119 spin_unlock(&req->rq_lock);
2121 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2124 * Interrupts (sets interrupted flag) all uncompleted requests in
2125 * a set \a data. Callback for l_wait_event for interruptible waits.
2127 static void ptlrpc_interrupted_set(void *data)
2129 struct ptlrpc_request_set *set = data;
2130 struct list_head *tmp;
2132 LASSERT(set != NULL);
2133 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2135 list_for_each(tmp, &set->set_requests) {
2136 struct ptlrpc_request *req =
2137 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2139 if (req->rq_phase != RQ_PHASE_RPC &&
2140 req->rq_phase != RQ_PHASE_UNREGISTERING)
2143 ptlrpc_mark_interrupted(req);
2148 * Get the smallest timeout in the set; this does NOT set a timeout.
2150 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2152 struct list_head *tmp;
2153 time_t now = cfs_time_current_sec();
2155 struct ptlrpc_request *req;
2159 list_for_each(tmp, &set->set_requests) {
2160 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2163 * Request in-flight?
2165 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2166 (req->rq_phase == RQ_PHASE_BULK) ||
2167 (req->rq_phase == RQ_PHASE_NEW)))
2171 * Already timed out.
2173 if (req->rq_timedout)
2179 if (req->rq_wait_ctx)
2182 if (req->rq_phase == RQ_PHASE_NEW)
2183 deadline = req->rq_sent;
2184 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2185 deadline = req->rq_sent;
2187 deadline = req->rq_sent + req->rq_timeout;
2189 if (deadline <= now) /* actually expired already */
2190 timeout = 1; /* ASAP */
2191 else if (timeout == 0 || timeout > deadline - now)
2192 timeout = deadline - now;
2198 * Send all unset request from the set and then wait untill all
2199 * requests in the set complete (either get a reply, timeout, get an
2200 * error or otherwise be interrupted).
2201 * Returns 0 on success or error code otherwise.
2203 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2205 struct list_head *tmp;
2206 struct ptlrpc_request *req;
2207 struct l_wait_info lwi;
2211 if (set->set_producer)
2212 (void)ptlrpc_set_producer(set);
2214 list_for_each(tmp, &set->set_requests) {
2215 req = list_entry(tmp, struct ptlrpc_request,
2217 if (req->rq_phase == RQ_PHASE_NEW)
2218 (void)ptlrpc_send_new_req(req);
2221 if (list_empty(&set->set_requests))
2225 timeout = ptlrpc_set_next_timeout(set);
2227 /* wait until all complete, interrupted, or an in-flight
2229 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2232 if (timeout == 0 && !signal_pending(current))
2234 * No requests are in-flight (ether timed out
2235 * or delayed), so we can allow interrupts.
2236 * We still want to block for a limited time,
2237 * so we allow interrupts during the timeout.
2239 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2241 ptlrpc_interrupted_set, set);
2244 * At least one request is in flight, so no
2245 * interrupts are allowed. Wait until all
2246 * complete, or an in-flight req times out.
2248 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2249 ptlrpc_expired_set, set);
2251 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2253 /* LU-769 - if we ignored the signal because it was already
2254 * pending when we started, we need to handle it now or we risk
2255 * it being ignored forever */
2256 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2257 signal_pending(current)) {
2258 sigset_t blocked_sigs =
2259 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2261 /* In fact we only interrupt for the "fatal" signals
2262 * like SIGINT or SIGKILL. We still ignore less
2263 * important signals since ptlrpc set is not easily
2264 * reentrant from userspace again */
2265 if (signal_pending(current))
2266 ptlrpc_interrupted_set(set);
2267 cfs_restore_sigs(blocked_sigs);
2270 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2272 /* -EINTR => all requests have been flagged rq_intr so next
2274 * -ETIMEDOUT => someone timed out. When all reqs have
2275 * timed out, signals are enabled allowing completion with
2277 * I don't really care if we go once more round the loop in
2278 * the error cases -eeb. */
2279 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2280 list_for_each(tmp, &set->set_requests) {
2281 req = list_entry(tmp, struct ptlrpc_request,
2283 spin_lock(&req->rq_lock);
2284 req->rq_invalid_rqset = 1;
2285 spin_unlock(&req->rq_lock);
2288 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2290 LASSERT(atomic_read(&set->set_remaining) == 0);
2292 rc = set->set_rc; /* rq_status of already freed requests if any */
2293 list_for_each(tmp, &set->set_requests) {
2294 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2296 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2297 if (req->rq_status != 0)
2298 rc = req->rq_status;
2301 if (set->set_interpret != NULL) {
2302 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2304 rc = interpreter (set, set->set_arg, rc);
2306 struct ptlrpc_set_cbdata *cbdata, *n;
2309 list_for_each_entry_safe(cbdata, n,
2310 &set->set_cblist, psc_item) {
2311 list_del_init(&cbdata->psc_item);
2312 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2315 OBD_FREE_PTR(cbdata);
2321 EXPORT_SYMBOL(ptlrpc_set_wait);
2324 * Helper fuction for request freeing.
2325 * Called when request count reached zero and request needs to be freed.
2326 * Removes request from all sorts of sending/replay lists it might be on,
2327 * frees network buffers if any are present.
2328 * If \a locked is set, that means caller is already holding import imp_lock
2329 * and so we no longer need to reobtain it (for certain lists manipulations)
2331 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2335 if (request == NULL)
2338 LASSERT(!request->rq_srv_req);
2339 LASSERT(request->rq_export == NULL);
2340 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2341 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2342 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2343 LASSERTF(!request->rq_replay, "req %p\n", request);
2345 req_capsule_fini(&request->rq_pill);
2347 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2348 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2349 if (request->rq_import != NULL) {
2351 spin_lock(&request->rq_import->imp_lock);
2352 list_del_init(&request->rq_replay_list);
2354 spin_unlock(&request->rq_import->imp_lock);
2356 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2358 if (atomic_read(&request->rq_refcount) != 0) {
2359 DEBUG_REQ(D_ERROR, request,
2360 "freeing request with nonzero refcount");
2364 if (request->rq_repbuf != NULL)
2365 sptlrpc_cli_free_repbuf(request);
2367 if (request->rq_import != NULL) {
2368 class_import_put(request->rq_import);
2369 request->rq_import = NULL;
2371 if (request->rq_bulk != NULL)
2372 ptlrpc_free_bulk(request->rq_bulk);
2374 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2375 sptlrpc_cli_free_reqbuf(request);
2377 if (request->rq_cli_ctx)
2378 sptlrpc_req_put_ctx(request, !locked);
2380 if (request->rq_pool)
2381 __ptlrpc_free_req_to_pool(request);
2383 ptlrpc_request_cache_free(request);
2387 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2389 * Drop one request reference. Must be called with import imp_lock held.
2390 * When reference count drops to zero, request is freed.
2392 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2394 assert_spin_locked(&request->rq_import->imp_lock);
2395 (void)__ptlrpc_req_finished(request, 1);
2400 * Drops one reference count for request \a request.
2401 * \a locked set indicates that caller holds import imp_lock.
2402 * Frees the request whe reference count reaches zero.
2404 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2407 if (request == NULL)
2410 if (request == LP_POISON ||
2411 request->rq_reqmsg == LP_POISON) {
2412 CERROR("dereferencing freed request (bug 575)\n");
2417 DEBUG_REQ(D_INFO, request, "refcount now %u",
2418 atomic_read(&request->rq_refcount) - 1);
2420 if (atomic_dec_and_test(&request->rq_refcount)) {
2421 __ptlrpc_free_req(request, locked);
2429 * Drops one reference count for a request.
2431 void ptlrpc_req_finished(struct ptlrpc_request *request)
2433 __ptlrpc_req_finished(request, 0);
2435 EXPORT_SYMBOL(ptlrpc_req_finished);
2438 * Returns xid of a \a request
2440 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2442 return request->rq_xid;
2444 EXPORT_SYMBOL(ptlrpc_req_xid);
2447 * Disengage the client's reply buffer from the network
2448 * NB does _NOT_ unregister any client-side bulk.
2449 * IDEMPOTENT, but _not_ safe against concurrent callers.
2450 * The request owner (i.e. the thread doing the I/O) must call...
2451 * Returns 0 on success or 1 if unregistering cannot be made.
2453 static int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2456 struct l_wait_info lwi;
2461 LASSERT(!in_interrupt());
2464 * Let's setup deadline for reply unlink.
2466 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2467 async && request->rq_reply_deadline == 0)
2468 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2471 * Nothing left to do.
2473 if (!ptlrpc_client_recv_or_unlink(request))
2476 LNetMDUnlink(request->rq_reply_md_h);
2479 * Let's check it once again.
2481 if (!ptlrpc_client_recv_or_unlink(request))
2485 * Move to "Unregistering" phase as reply was not unlinked yet.
2487 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2490 * Do not wait for unlink to finish.
2496 * We have to l_wait_event() whatever the result, to give liblustre
2497 * a chance to run reply_in_callback(), and to make sure we've
2498 * unlinked before returning a req to the pool.
2501 /* The wq argument is ignored by user-space wait_event macros */
2502 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2503 &request->rq_set->set_waitq :
2504 &request->rq_reply_waitq;
2505 /* Network access will complete in finite time but the HUGE
2506 * timeout lets us CWARN for visibility of sluggish NALs */
2507 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2508 cfs_time_seconds(1), NULL, NULL);
2509 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2512 ptlrpc_rqphase_move(request, request->rq_next_phase);
2516 LASSERT(rc == -ETIMEDOUT);
2517 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2518 "receiving_reply=%d req_ulinked=%d reply_unlinked=%d",
2519 request->rq_receiving_reply,
2520 request->rq_req_unlinked,
2521 request->rq_reply_unlinked);
2526 static void ptlrpc_free_request(struct ptlrpc_request *req)
2528 spin_lock(&req->rq_lock);
2530 spin_unlock(&req->rq_lock);
2532 if (req->rq_commit_cb != NULL)
2533 req->rq_commit_cb(req);
2534 list_del_init(&req->rq_replay_list);
2536 __ptlrpc_req_finished(req, 1);
2540 * the request is committed and dropped from the replay list of its import
2542 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2544 struct obd_import *imp = req->rq_import;
2546 spin_lock(&imp->imp_lock);
2547 if (list_empty(&req->rq_replay_list)) {
2548 spin_unlock(&imp->imp_lock);
2552 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2553 ptlrpc_free_request(req);
2555 spin_unlock(&imp->imp_lock);
2557 EXPORT_SYMBOL(ptlrpc_request_committed);
2560 * Iterates through replay_list on import and prunes
2561 * all requests have transno smaller than last_committed for the
2562 * import and don't have rq_replay set.
2563 * Since requests are sorted in transno order, stops when meetign first
2564 * transno bigger than last_committed.
2565 * caller must hold imp->imp_lock
2567 void ptlrpc_free_committed(struct obd_import *imp)
2569 struct ptlrpc_request *req, *saved;
2570 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2571 bool skip_committed_list = true;
2574 LASSERT(imp != NULL);
2575 assert_spin_locked(&imp->imp_lock);
2577 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2578 imp->imp_generation == imp->imp_last_generation_checked) {
2579 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2580 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2583 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2584 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2585 imp->imp_generation);
2587 if (imp->imp_generation != imp->imp_last_generation_checked ||
2588 imp->imp_last_transno_checked == 0)
2589 skip_committed_list = false;
2591 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2592 imp->imp_last_generation_checked = imp->imp_generation;
2594 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2596 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2597 LASSERT(req != last_req);
2600 if (req->rq_transno == 0) {
2601 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2604 if (req->rq_import_generation < imp->imp_generation) {
2605 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2609 /* not yet committed */
2610 if (req->rq_transno > imp->imp_peer_committed_transno) {
2611 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2615 if (req->rq_replay) {
2616 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2617 list_move_tail(&req->rq_replay_list,
2618 &imp->imp_committed_list);
2622 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2623 imp->imp_peer_committed_transno);
2625 ptlrpc_free_request(req);
2628 if (skip_committed_list)
2631 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2633 LASSERT(req->rq_transno != 0);
2634 if (req->rq_import_generation < imp->imp_generation) {
2635 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2636 ptlrpc_free_request(req);
2637 } else if (!req->rq_replay) {
2638 DEBUG_REQ(D_RPCTRACE, req, "free closed open request");
2639 ptlrpc_free_request(req);
2646 void ptlrpc_cleanup_client(struct obd_import *imp)
2653 * Schedule previously sent request for resend.
2654 * For bulk requests we assign new xid (to avoid problems with
2655 * lost replies and therefore several transfers landing into same buffer
2656 * from different sending attempts).
2658 void ptlrpc_resend_req(struct ptlrpc_request *req)
2660 DEBUG_REQ(D_HA, req, "going to resend");
2661 spin_lock(&req->rq_lock);
2663 /* Request got reply but linked to the import list still.
2664 Let ptlrpc_check_set() to process it. */
2665 if (ptlrpc_client_replied(req)) {
2666 spin_unlock(&req->rq_lock);
2667 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2671 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2672 req->rq_status = -EAGAIN;
2675 req->rq_net_err = 0;
2676 req->rq_timedout = 0;
2678 ptlrpc_client_wake_req(req);
2679 spin_unlock(&req->rq_lock);
2682 /* XXX: this function and rq_status are currently unused */
2683 void ptlrpc_restart_req(struct ptlrpc_request *req)
2685 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2686 req->rq_status = -ERESTARTSYS;
2688 spin_lock(&req->rq_lock);
2689 req->rq_restart = 1;
2690 req->rq_timedout = 0;
2691 ptlrpc_client_wake_req(req);
2692 spin_unlock(&req->rq_lock);
2696 * Grab additional reference on a request \a req
2698 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2701 atomic_inc(&req->rq_refcount);
2704 EXPORT_SYMBOL(ptlrpc_request_addref);
2707 * Add a request to import replay_list.
2708 * Must be called under imp_lock
2710 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2711 struct obd_import *imp)
2713 struct list_head *tmp;
2715 assert_spin_locked(&imp->imp_lock);
2717 if (req->rq_transno == 0) {
2718 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2722 /* clear this for new requests that were resent as well
2723 as resent replayed requests. */
2724 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2726 /* don't re-add requests that have been replayed */
2727 if (!list_empty(&req->rq_replay_list))
2730 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2732 spin_lock(&req->rq_lock);
2734 spin_unlock(&req->rq_lock);
2736 LASSERT(imp->imp_replayable);
2737 /* Balanced in ptlrpc_free_committed, usually. */
2738 ptlrpc_request_addref(req);
2739 list_for_each_prev(tmp, &imp->imp_replay_list) {
2740 struct ptlrpc_request *iter = list_entry(tmp,
2741 struct ptlrpc_request,
2744 /* We may have duplicate transnos if we create and then
2745 * open a file, or for closes retained if to match creating
2746 * opens, so use req->rq_xid as a secondary key.
2747 * (See bugs 684, 685, and 428.)
2748 * XXX no longer needed, but all opens need transnos!
2750 if (iter->rq_transno > req->rq_transno)
2753 if (iter->rq_transno == req->rq_transno) {
2754 LASSERT(iter->rq_xid != req->rq_xid);
2755 if (iter->rq_xid > req->rq_xid)
2759 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2763 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2767 * Send request and wait until it completes.
2768 * Returns request processing status.
2770 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2772 struct ptlrpc_request_set *set;
2776 LASSERT(req->rq_set == NULL);
2777 LASSERT(!req->rq_receiving_reply);
2779 set = ptlrpc_prep_set();
2781 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2785 /* for distributed debugging */
2786 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2788 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2789 ptlrpc_request_addref(req);
2790 ptlrpc_set_add_req(set, req);
2791 rc = ptlrpc_set_wait(set);
2792 ptlrpc_set_destroy(set);
2796 EXPORT_SYMBOL(ptlrpc_queue_wait);
2799 * Callback used for replayed requests reply processing.
2800 * In case of successful reply calls registered request replay callback.
2801 * In case of error restart replay process.
2803 static int ptlrpc_replay_interpret(const struct lu_env *env,
2804 struct ptlrpc_request *req,
2805 void * data, int rc)
2807 struct ptlrpc_replay_async_args *aa = data;
2808 struct obd_import *imp = req->rq_import;
2811 atomic_dec(&imp->imp_replay_inflight);
2813 /* Note: if it is bulk replay (MDS-MDS replay), then even if
2814 * server got the request, but bulk transfer timeout, let's
2815 * replay the bulk req again */
2816 if (!ptlrpc_client_replied(req) ||
2817 (req->rq_bulk != NULL &&
2818 lustre_msg_get_status(req->rq_repmsg) == -ETIMEDOUT)) {
2819 DEBUG_REQ(D_ERROR, req, "request replay timed out.\n");
2820 GOTO(out, rc = -ETIMEDOUT);
2823 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2824 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2825 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2826 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2828 /** VBR: check version failure */
2829 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2830 /** replay was failed due to version mismatch */
2831 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2832 spin_lock(&imp->imp_lock);
2833 imp->imp_vbr_failed = 1;
2834 imp->imp_no_lock_replay = 1;
2835 spin_unlock(&imp->imp_lock);
2836 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2838 /** The transno had better not change over replay. */
2839 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2840 lustre_msg_get_transno(req->rq_repmsg) ||
2841 lustre_msg_get_transno(req->rq_repmsg) == 0,
2843 lustre_msg_get_transno(req->rq_reqmsg),
2844 lustre_msg_get_transno(req->rq_repmsg));
2847 spin_lock(&imp->imp_lock);
2848 /** if replays by version then gap occur on server, no trust to locks */
2849 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2850 imp->imp_no_lock_replay = 1;
2851 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2852 spin_unlock(&imp->imp_lock);
2853 LASSERT(imp->imp_last_replay_transno);
2855 /* transaction number shouldn't be bigger than the latest replayed */
2856 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2857 DEBUG_REQ(D_ERROR, req,
2858 "Reported transno "LPU64" is bigger than the "
2859 "replayed one: "LPU64, req->rq_transno,
2860 lustre_msg_get_transno(req->rq_reqmsg));
2861 GOTO(out, rc = -EINVAL);
2864 DEBUG_REQ(D_HA, req, "got rep");
2866 /* let the callback do fixups, possibly including in the request */
2867 if (req->rq_replay_cb)
2868 req->rq_replay_cb(req);
2870 if (ptlrpc_client_replied(req) &&
2871 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2872 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2873 lustre_msg_get_status(req->rq_repmsg),
2874 aa->praa_old_status);
2876 /* Put it back for re-replay. */
2877 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2881 * Errors while replay can set transno to 0, but
2882 * imp_last_replay_transno shouldn't be set to 0 anyway
2884 if (req->rq_transno == 0)
2885 CERROR("Transno is 0 during replay!\n");
2887 /* continue with recovery */
2888 rc = ptlrpc_import_recovery_state_machine(imp);
2890 req->rq_send_state = aa->praa_old_state;
2893 /* this replay failed, so restart recovery */
2894 ptlrpc_connect_import(imp);
2900 * Prepares and queues request for replay.
2901 * Adds it to ptlrpcd queue for actual sending.
2902 * Returns 0 on success.
2904 int ptlrpc_replay_req(struct ptlrpc_request *req)
2906 struct ptlrpc_replay_async_args *aa;
2909 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2911 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2912 aa = ptlrpc_req_async_args(req);
2913 memset(aa, 0, sizeof *aa);
2915 /* Prepare request to be resent with ptlrpcd */
2916 aa->praa_old_state = req->rq_send_state;
2917 req->rq_send_state = LUSTRE_IMP_REPLAY;
2918 req->rq_phase = RQ_PHASE_NEW;
2919 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2921 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2923 req->rq_interpret_reply = ptlrpc_replay_interpret;
2924 /* Readjust the timeout for current conditions */
2925 ptlrpc_at_set_req_timeout(req);
2927 /* Tell server the net_latency, so the server can calculate how long
2928 * it should wait for next replay */
2929 lustre_msg_set_service_time(req->rq_reqmsg,
2930 ptlrpc_at_get_net_latency(req));
2931 DEBUG_REQ(D_HA, req, "REPLAY");
2933 atomic_inc(&req->rq_import->imp_replay_inflight);
2934 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2936 ptlrpcd_add_req(req);
2941 * Aborts all in-flight request on import \a imp sending and delayed lists
2943 void ptlrpc_abort_inflight(struct obd_import *imp)
2945 struct list_head *tmp, *n;
2948 /* Make sure that no new requests get processed for this import.
2949 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2950 * this flag and then putting requests on sending_list or delayed_list.
2952 spin_lock(&imp->imp_lock);
2954 /* XXX locking? Maybe we should remove each request with the list
2955 * locked? Also, how do we know if the requests on the list are
2956 * being freed at this time?
2958 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2959 struct ptlrpc_request *req = list_entry(tmp,
2960 struct ptlrpc_request,
2963 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2965 spin_lock(&req->rq_lock);
2966 if (req->rq_import_generation < imp->imp_generation) {
2968 req->rq_status = -EIO;
2969 ptlrpc_client_wake_req(req);
2971 spin_unlock(&req->rq_lock);
2974 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2975 struct ptlrpc_request *req =
2976 list_entry(tmp, struct ptlrpc_request, rq_list);
2978 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2980 spin_lock(&req->rq_lock);
2981 if (req->rq_import_generation < imp->imp_generation) {
2983 req->rq_status = -EIO;
2984 ptlrpc_client_wake_req(req);
2986 spin_unlock(&req->rq_lock);
2989 /* Last chance to free reqs left on the replay list, but we
2990 * will still leak reqs that haven't committed. */
2991 if (imp->imp_replayable)
2992 ptlrpc_free_committed(imp);
2994 spin_unlock(&imp->imp_lock);
3000 * Abort all uncompleted requests in request set \a set
3002 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
3004 struct list_head *tmp, *pos;
3006 LASSERT(set != NULL);
3008 list_for_each_safe(pos, tmp, &set->set_requests) {
3009 struct ptlrpc_request *req =
3010 list_entry(pos, struct ptlrpc_request,
3013 spin_lock(&req->rq_lock);
3014 if (req->rq_phase != RQ_PHASE_RPC) {
3015 spin_unlock(&req->rq_lock);
3020 req->rq_status = -EINTR;
3021 ptlrpc_client_wake_req(req);
3022 spin_unlock(&req->rq_lock);
3026 static __u64 ptlrpc_last_xid;
3027 static spinlock_t ptlrpc_last_xid_lock;
3030 * Initialize the XID for the node. This is common among all requests on
3031 * this node, and only requires the property that it is monotonically
3032 * increasing. It does not need to be sequential. Since this is also used
3033 * as the RDMA match bits, it is important that a single client NOT have
3034 * the same match bits for two different in-flight requests, hence we do
3035 * NOT want to have an XID per target or similar.
3037 * To avoid an unlikely collision between match bits after a client reboot
3038 * (which would deliver old data into the wrong RDMA buffer) initialize
3039 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
3040 * If the time is clearly incorrect, we instead use a 62-bit random number.
3041 * In the worst case the random number will overflow 1M RPCs per second in
3042 * 9133 years, or permutations thereof.
3044 #define YEAR_2004 (1ULL << 30)
3045 void ptlrpc_init_xid(void)
3047 time_t now = cfs_time_current_sec();
3049 spin_lock_init(&ptlrpc_last_xid_lock);
3050 if (now < YEAR_2004) {
3051 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
3052 ptlrpc_last_xid >>= 2;
3053 ptlrpc_last_xid |= (1ULL << 61);
3055 ptlrpc_last_xid = (__u64)now << 20;
3058 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3059 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3060 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3064 * Increase xid and returns resulting new value to the caller.
3066 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3067 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3068 * itself uses the last bulk xid needed, so the server can determine the
3069 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3070 * xid must align to a power-of-two value.
3072 * This is assumed to be true due to the initial ptlrpc_last_xid
3073 * value also being initialized to a power-of-two value. LU-1431
3075 __u64 ptlrpc_next_xid(void)
3079 spin_lock(&ptlrpc_last_xid_lock);
3080 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3081 ptlrpc_last_xid = next;
3082 spin_unlock(&ptlrpc_last_xid_lock);
3088 * If request has a new allocated XID (new request or EINPROGRESS resend),
3089 * use this XID as matchbits of bulk, otherwise allocate a new matchbits for
3090 * request to ensure previous bulk fails and avoid problems with lost replies
3091 * and therefore several transfers landing into the same buffer from different
3094 void ptlrpc_set_bulk_mbits(struct ptlrpc_request *req)
3096 struct ptlrpc_bulk_desc *bd = req->rq_bulk;
3098 LASSERT(bd != NULL);
3100 if (!req->rq_resend || req->rq_nr_resend != 0) {
3101 /* this request has a new xid, just use it as bulk matchbits */
3102 req->rq_mbits = req->rq_xid;
3104 } else { /* needs to generate a new matchbits for resend */
3105 __u64 old_mbits = req->rq_mbits;
3107 if ((bd->bd_import->imp_connect_data.ocd_connect_flags &
3108 OBD_CONNECT_BULK_MBITS) != 0)
3109 req->rq_mbits = ptlrpc_next_xid();
3110 else /* old version transfers rq_xid to peer as matchbits */
3111 req->rq_mbits = req->rq_xid = ptlrpc_next_xid();
3113 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
3114 old_mbits, req->rq_mbits);
3117 /* For multi-bulk RPCs, rq_mbits is the last mbits needed for bulks so
3118 * that server can infer the number of bulks that were prepared,
3120 req->rq_mbits += ((bd->bd_iov_count + LNET_MAX_IOV - 1) /
3125 * Get a glimpse at what next xid value might have been.
3126 * Returns possible next xid.
3128 __u64 ptlrpc_sample_next_xid(void)
3130 #if BITS_PER_LONG == 32
3131 /* need to avoid possible word tearing on 32-bit systems */
3134 spin_lock(&ptlrpc_last_xid_lock);
3135 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3136 spin_unlock(&ptlrpc_last_xid_lock);
3140 /* No need to lock, since returned value is racy anyways */
3141 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3144 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3147 * Functions for operating ptlrpc workers.
3149 * A ptlrpc work is a function which will be running inside ptlrpc context.
3150 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3152 * 1. after a work is created, it can be used many times, that is:
3153 * handler = ptlrpcd_alloc_work();
3154 * ptlrpcd_queue_work();
3156 * queue it again when necessary:
3157 * ptlrpcd_queue_work();
3158 * ptlrpcd_destroy_work();
3159 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3160 * it will only be queued once in any time. Also as its name implies, it may
3161 * have delay before it really runs by ptlrpcd thread.
3163 struct ptlrpc_work_async_args {
3164 int (*cb)(const struct lu_env *, void *);
3168 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3170 /* re-initialize the req */
3171 req->rq_timeout = obd_timeout;
3172 req->rq_sent = cfs_time_current_sec();
3173 req->rq_deadline = req->rq_sent + req->rq_timeout;
3174 req->rq_reply_deadline = req->rq_deadline;
3175 req->rq_phase = RQ_PHASE_INTERPRET;
3176 req->rq_next_phase = RQ_PHASE_COMPLETE;
3177 req->rq_xid = ptlrpc_next_xid();
3178 req->rq_import_generation = req->rq_import->imp_generation;
3180 ptlrpcd_add_req(req);
3183 static int work_interpreter(const struct lu_env *env,
3184 struct ptlrpc_request *req, void *data, int rc)
3186 struct ptlrpc_work_async_args *arg = data;
3188 LASSERT(ptlrpcd_check_work(req));
3189 LASSERT(arg->cb != NULL);
3191 rc = arg->cb(env, arg->cbdata);
3193 list_del_init(&req->rq_set_chain);
3196 if (atomic_dec_return(&req->rq_refcount) > 1) {
3197 atomic_set(&req->rq_refcount, 2);
3198 ptlrpcd_add_work_req(req);
3203 static int worker_format;
3205 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3207 return req->rq_pill.rc_fmt == (void *)&worker_format;
3211 * Create a work for ptlrpc.
3213 void *ptlrpcd_alloc_work(struct obd_import *imp,
3214 int (*cb)(const struct lu_env *, void *), void *cbdata)
3216 struct ptlrpc_request *req = NULL;
3217 struct ptlrpc_work_async_args *args;
3223 RETURN(ERR_PTR(-EINVAL));
3225 /* copy some code from deprecated fakereq. */
3226 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3228 CERROR("ptlrpc: run out of memory!\n");
3229 RETURN(ERR_PTR(-ENOMEM));
3232 ptlrpc_cli_req_init(req);
3234 req->rq_send_state = LUSTRE_IMP_FULL;
3235 req->rq_type = PTL_RPC_MSG_REQUEST;
3236 req->rq_import = class_import_get(imp);
3237 req->rq_interpret_reply = work_interpreter;
3238 /* don't want reply */
3239 req->rq_no_delay = req->rq_no_resend = 1;
3240 req->rq_pill.rc_fmt = (void *)&worker_format;
3242 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3243 args = ptlrpc_req_async_args(req);
3245 args->cbdata = cbdata;
3249 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3251 void ptlrpcd_destroy_work(void *handler)
3253 struct ptlrpc_request *req = handler;
3256 ptlrpc_req_finished(req);
3258 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3260 int ptlrpcd_queue_work(void *handler)
3262 struct ptlrpc_request *req = handler;
3265 * Check if the req is already being queued.
3267 * Here comes a trick: it lacks a way of checking if a req is being
3268 * processed reliably in ptlrpc. Here I have to use refcount of req
3269 * for this purpose. This is okay because the caller should use this
3270 * req as opaque data. - Jinshan
3272 LASSERT(atomic_read(&req->rq_refcount) > 0);
3273 if (atomic_inc_return(&req->rq_refcount) == 2)
3274 ptlrpcd_add_work_req(req);
3277 EXPORT_SYMBOL(ptlrpcd_queue_work);