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, 2013, 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
43 #include <liblustre.h>
46 #include <obd_support.h>
47 #include <obd_class.h>
48 #include <lustre_lib.h>
49 #include <lustre_ha.h>
50 #include <lustre_import.h>
51 #include <lustre_req_layout.h>
53 #include "ptlrpc_internal.h"
55 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
56 static int ptlrpcd_check_work(struct ptlrpc_request *req);
59 * Initialize passed in client structure \a cl.
61 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
62 struct ptlrpc_client *cl)
64 cl->cli_request_portal = req_portal;
65 cl->cli_reply_portal = rep_portal;
68 EXPORT_SYMBOL(ptlrpc_init_client);
71 * Return PortalRPC connection for remore uud \a uuid
73 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
75 struct ptlrpc_connection *c;
77 lnet_process_id_t peer;
80 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
81 * before accessing its values. */
82 /* coverity[uninit_use_in_call] */
83 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
85 CNETERR("cannot find peer %s!\n", uuid->uuid);
89 c = ptlrpc_connection_get(peer, self, uuid);
91 memcpy(c->c_remote_uuid.uuid,
92 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
95 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
99 EXPORT_SYMBOL(ptlrpc_uuid_to_connection);
102 * Allocate and initialize new bulk descriptor on the sender.
103 * Returns pointer to the descriptor or NULL on error.
105 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned npages, unsigned max_brw,
106 unsigned type, unsigned portal)
108 struct ptlrpc_bulk_desc *desc;
111 OBD_ALLOC(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[npages]));
115 spin_lock_init(&desc->bd_lock);
116 init_waitqueue_head(&desc->bd_waitq);
117 desc->bd_max_iov = npages;
118 desc->bd_iov_count = 0;
119 desc->bd_portal = portal;
120 desc->bd_type = type;
121 desc->bd_md_count = 0;
122 LASSERT(max_brw > 0);
123 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
124 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
125 * node. Negotiated ocd_brw_size will always be <= this number. */
126 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
127 LNetInvalidateHandle(&desc->bd_mds[i]);
133 * Prepare bulk descriptor for specified outgoing request \a req that
134 * can fit \a npages * pages. \a type is bulk type. \a portal is where
135 * the bulk to be sent. Used on client-side.
136 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
139 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
140 unsigned npages, unsigned max_brw,
141 unsigned type, unsigned portal)
143 struct obd_import *imp = req->rq_import;
144 struct ptlrpc_bulk_desc *desc;
147 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
148 desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
152 desc->bd_import_generation = req->rq_import_generation;
153 desc->bd_import = class_import_get(imp);
156 desc->bd_cbid.cbid_fn = client_bulk_callback;
157 desc->bd_cbid.cbid_arg = desc;
159 /* This makes req own desc, and free it when she frees herself */
164 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
167 * Add a page \a page to the bulk descriptor \a desc.
168 * Data to transfer in the page starts at offset \a pageoffset and
169 * amount of data to transfer from the page is \a len
171 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
172 struct page *page, int pageoffset, int len, int pin)
174 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
175 LASSERT(page != NULL);
176 LASSERT(pageoffset >= 0);
178 LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
183 page_cache_get(page);
185 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
187 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
190 * Uninitialize and free bulk descriptor \a desc.
191 * Works on bulk descriptors both from server and client side.
193 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin)
198 LASSERT(desc != NULL);
199 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
200 LASSERT(desc->bd_md_count == 0); /* network hands off */
201 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
203 sptlrpc_enc_pool_put_pages(desc);
206 class_export_put(desc->bd_export);
208 class_import_put(desc->bd_import);
211 for (i = 0; i < desc->bd_iov_count ; i++)
212 page_cache_release(desc->bd_iov[i].kiov_page);
215 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
216 bd_iov[desc->bd_max_iov]));
219 EXPORT_SYMBOL(__ptlrpc_free_bulk);
222 * Set server timelimit for this req, i.e. how long are we willing to wait
223 * for reply before timing out this request.
225 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
231 LASSERT(req->rq_import);
234 /* non-AT settings */
236 * \a imp_server_timeout means this is reverse import and
237 * we send (currently only) ASTs to the client and cannot afford
238 * to wait too long for the reply, otherwise the other client
239 * (because of which we are sending this request) would
240 * timeout waiting for us
242 req->rq_timeout = req->rq_import->imp_server_timeout ?
243 obd_timeout / 2 : obd_timeout;
245 at = &req->rq_import->imp_at;
246 idx = import_at_get_index(req->rq_import,
247 req->rq_request_portal);
248 serv_est = at_get(&at->iat_service_estimate[idx]);
249 req->rq_timeout = at_est2timeout(serv_est);
251 /* We could get even fancier here, using history to predict increased
254 /* Let the server know what this RPC timeout is by putting it in the
256 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
258 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
260 /* Adjust max service estimate based on server value */
261 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
262 unsigned int serv_est)
268 LASSERT(req->rq_import);
269 at = &req->rq_import->imp_at;
271 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
272 /* max service estimates are tracked on the server side,
273 so just keep minimal history here */
274 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
276 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
277 "has changed from %d to %d\n",
278 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
279 oldse, at_get(&at->iat_service_estimate[idx]));
282 /* Expected network latency per remote node (secs) */
283 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
285 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
288 /* Adjust expected network latency */
289 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
290 unsigned int service_time)
292 unsigned int nl, oldnl;
294 time_t now = cfs_time_current_sec();
296 LASSERT(req->rq_import);
297 at = &req->rq_import->imp_at;
299 /* Network latency is total time less server processing time */
300 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
301 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
302 CWARN("Reported service time %u > total measured time "
303 CFS_DURATION_T"\n", service_time,
304 cfs_time_sub(now, req->rq_sent));
306 oldnl = at_measured(&at->iat_net_latency, nl);
308 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
309 "has changed from %d to %d\n",
310 req->rq_import->imp_obd->obd_name,
312 &req->rq_import->imp_connection->c_remote_uuid),
313 oldnl, at_get(&at->iat_net_latency));
316 static int unpack_reply(struct ptlrpc_request *req)
320 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
321 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
323 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
328 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
330 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
337 * Handle an early reply message, called with the rq_lock held.
338 * If anything goes wrong just ignore it - same as if it never happened
340 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
342 struct ptlrpc_request *early_req;
348 spin_unlock(&req->rq_lock);
350 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
352 spin_lock(&req->rq_lock);
356 rc = unpack_reply(early_req);
358 /* Expecting to increase the service time estimate here */
359 ptlrpc_at_adj_service(req,
360 lustre_msg_get_timeout(early_req->rq_repmsg));
361 ptlrpc_at_adj_net_latency(req,
362 lustre_msg_get_service_time(early_req->rq_repmsg));
365 sptlrpc_cli_finish_early_reply(early_req);
368 spin_lock(&req->rq_lock);
372 /* Adjust the local timeout for this req */
373 ptlrpc_at_set_req_timeout(req);
375 spin_lock(&req->rq_lock);
376 olddl = req->rq_deadline;
377 /* server assumes it now has rq_timeout from when the request
378 * arrived, so the client should give it at least that long.
379 * since we don't know the arrival time we'll use the original
381 req->rq_deadline = req->rq_sent + req->rq_timeout +
382 ptlrpc_at_get_net_latency(req);
384 DEBUG_REQ(D_ADAPTTO, req,
385 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
386 "("CFS_DURATION_T"s)", req->rq_early_count,
387 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
388 cfs_time_sub(req->rq_deadline, olddl));
393 struct kmem_cache *request_cache;
395 int ptlrpc_request_cache_init(void)
397 request_cache = kmem_cache_create("ptlrpc_cache",
398 sizeof(struct ptlrpc_request),
399 0, SLAB_HWCACHE_ALIGN, NULL);
400 return request_cache == NULL ? -ENOMEM : 0;
403 void ptlrpc_request_cache_fini(void)
405 kmem_cache_destroy(request_cache);
408 struct ptlrpc_request *ptlrpc_request_cache_alloc(int flags)
410 struct ptlrpc_request *req;
412 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
416 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
418 OBD_SLAB_FREE_PTR(req, request_cache);
422 * Wind down request pool \a pool.
423 * Frees all requests from the pool too
425 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
428 struct ptlrpc_request *req;
430 LASSERT(pool != NULL);
432 spin_lock(&pool->prp_lock);
433 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
434 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
435 cfs_list_del(&req->rq_list);
436 LASSERT(req->rq_reqbuf);
437 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
438 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
439 ptlrpc_request_cache_free(req);
441 spin_unlock(&pool->prp_lock);
442 OBD_FREE(pool, sizeof(*pool));
444 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
447 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
449 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
454 while (size < pool->prp_rq_size)
457 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
458 size == pool->prp_rq_size,
459 "Trying to change pool size with nonempty pool "
460 "from %d to %d bytes\n", pool->prp_rq_size, size);
462 spin_lock(&pool->prp_lock);
463 pool->prp_rq_size = size;
464 for (i = 0; i < num_rq; i++) {
465 struct ptlrpc_request *req;
466 struct lustre_msg *msg;
468 spin_unlock(&pool->prp_lock);
469 req = ptlrpc_request_cache_alloc(GFP_NOFS);
472 OBD_ALLOC_LARGE(msg, size);
474 ptlrpc_request_cache_free(req);
477 req->rq_reqbuf = msg;
478 req->rq_reqbuf_len = size;
480 spin_lock(&pool->prp_lock);
481 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
483 spin_unlock(&pool->prp_lock);
486 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
489 * Create and initialize new request pool with given attributes:
490 * \a num_rq - initial number of requests to create for the pool
491 * \a msgsize - maximum message size possible for requests in thid pool
492 * \a populate_pool - function to be called when more requests need to be added
494 * Returns pointer to newly created pool or NULL on error.
496 struct ptlrpc_request_pool *
497 ptlrpc_init_rq_pool(int num_rq, int msgsize,
498 void (*populate_pool)(struct ptlrpc_request_pool *, int))
500 struct ptlrpc_request_pool *pool;
502 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
506 /* Request next power of two for the allocation, because internally
507 kernel would do exactly this */
509 spin_lock_init(&pool->prp_lock);
510 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
511 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
512 pool->prp_populate = populate_pool;
514 populate_pool(pool, num_rq);
516 if (cfs_list_empty(&pool->prp_req_list)) {
517 /* have not allocated a single request for the pool */
518 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
523 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
526 * Fetches one request from pool \a pool
528 static struct ptlrpc_request *
529 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
531 struct ptlrpc_request *request;
532 struct lustre_msg *reqbuf;
537 spin_lock(&pool->prp_lock);
539 /* See if we have anything in a pool, and bail out if nothing,
540 * in writeout path, where this matters, this is safe to do, because
541 * nothing is lost in this case, and when some in-flight requests
542 * complete, this code will be called again. */
543 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
544 spin_unlock(&pool->prp_lock);
548 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
550 cfs_list_del_init(&request->rq_list);
551 spin_unlock(&pool->prp_lock);
553 LASSERT(request->rq_reqbuf);
554 LASSERT(request->rq_pool);
556 reqbuf = request->rq_reqbuf;
557 memset(request, 0, sizeof(*request));
558 request->rq_reqbuf = reqbuf;
559 request->rq_reqbuf_len = pool->prp_rq_size;
560 request->rq_pool = pool;
566 * Returns freed \a request to pool.
568 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
570 struct ptlrpc_request_pool *pool = request->rq_pool;
572 spin_lock(&pool->prp_lock);
573 LASSERT(cfs_list_empty(&request->rq_list));
574 LASSERT(!request->rq_receiving_reply);
575 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
576 spin_unlock(&pool->prp_lock);
579 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
580 __u32 version, int opcode,
581 int count, __u32 *lengths, char **bufs,
582 struct ptlrpc_cli_ctx *ctx)
584 struct obd_import *imp = request->rq_import;
589 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
591 rc = sptlrpc_req_get_ctx(request);
596 sptlrpc_req_set_flavor(request, opcode);
598 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
601 LASSERT(!request->rq_pool);
605 lustre_msg_add_version(request->rq_reqmsg, version);
606 request->rq_send_state = LUSTRE_IMP_FULL;
607 request->rq_type = PTL_RPC_MSG_REQUEST;
608 request->rq_export = NULL;
610 request->rq_req_cbid.cbid_fn = request_out_callback;
611 request->rq_req_cbid.cbid_arg = request;
613 request->rq_reply_cbid.cbid_fn = reply_in_callback;
614 request->rq_reply_cbid.cbid_arg = request;
616 request->rq_reply_deadline = 0;
617 request->rq_phase = RQ_PHASE_NEW;
618 request->rq_next_phase = RQ_PHASE_UNDEFINED;
620 request->rq_request_portal = imp->imp_client->cli_request_portal;
621 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
623 ptlrpc_at_set_req_timeout(request);
625 spin_lock_init(&request->rq_lock);
626 CFS_INIT_LIST_HEAD(&request->rq_list);
627 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
628 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
629 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
630 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
631 CFS_INIT_LIST_HEAD(&request->rq_history_list);
632 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
633 init_waitqueue_head(&request->rq_reply_waitq);
634 init_waitqueue_head(&request->rq_set_waitq);
635 request->rq_xid = ptlrpc_next_xid();
636 atomic_set(&request->rq_refcount, 1);
638 lustre_msg_set_opc(request->rq_reqmsg, opcode);
642 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
644 class_import_put(imp);
648 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
649 __u32 version, int opcode, char **bufs,
650 struct ptlrpc_cli_ctx *ctx)
654 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
655 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
656 request->rq_pill.rc_area[RCL_CLIENT],
659 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
662 * Pack request buffers for network transfer, performing necessary encryption
663 * steps if necessary.
665 int ptlrpc_request_pack(struct ptlrpc_request *request,
666 __u32 version, int opcode)
669 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
673 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
674 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
675 * have to send old ptlrpc_body to keep interoprability with these
678 * Only three kinds of server->client RPCs so far:
683 * XXX This should be removed whenever we drop the interoprability with
684 * the these old clients.
686 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
687 opcode == LDLM_GL_CALLBACK)
688 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
689 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
693 EXPORT_SYMBOL(ptlrpc_request_pack);
696 * Helper function to allocate new request on import \a imp
697 * and possibly using existing request from pool \a pool if provided.
698 * Returns allocated request structure with import field filled or
702 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
703 struct ptlrpc_request_pool *pool)
705 struct ptlrpc_request *request = NULL;
708 request = ptlrpc_prep_req_from_pool(pool);
711 request = ptlrpc_request_cache_alloc(GFP_NOFS);
714 LASSERTF((unsigned long)imp > 0x1000, "%p\n", imp);
715 LASSERT(imp != LP_POISON);
716 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
718 LASSERT(imp->imp_client != LP_POISON);
720 request->rq_import = class_import_get(imp);
722 CERROR("request allocation out of memory\n");
729 * Helper function for creating a request.
730 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
731 * buffer structures according to capsule template \a format.
732 * Returns allocated request structure pointer or NULL on error.
734 static struct ptlrpc_request *
735 ptlrpc_request_alloc_internal(struct obd_import *imp,
736 struct ptlrpc_request_pool * pool,
737 const struct req_format *format)
739 struct ptlrpc_request *request;
741 request = __ptlrpc_request_alloc(imp, pool);
745 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
746 req_capsule_set(&request->rq_pill, format);
751 * Allocate new request structure for import \a imp and initialize its
752 * buffer structure according to capsule template \a format.
754 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
755 const struct req_format *format)
757 return ptlrpc_request_alloc_internal(imp, NULL, format);
759 EXPORT_SYMBOL(ptlrpc_request_alloc);
762 * Allocate new request structure for import \a imp from pool \a pool and
763 * initialize its buffer structure according to capsule template \a format.
765 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
766 struct ptlrpc_request_pool * pool,
767 const struct req_format *format)
769 return ptlrpc_request_alloc_internal(imp, pool, format);
771 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
774 * For requests not from pool, free memory of the request structure.
775 * For requests obtained from a pool earlier, return request back to pool.
777 void ptlrpc_request_free(struct ptlrpc_request *request)
779 if (request->rq_pool)
780 __ptlrpc_free_req_to_pool(request);
782 ptlrpc_request_cache_free(request);
784 EXPORT_SYMBOL(ptlrpc_request_free);
787 * Allocate new request for operatione \a opcode and immediatelly pack it for
789 * Only used for simple requests like OBD_PING where the only important
790 * part of the request is operation itself.
791 * Returns allocated request or NULL on error.
793 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
794 const struct req_format *format,
795 __u32 version, int opcode)
797 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
801 rc = ptlrpc_request_pack(req, version, opcode);
803 ptlrpc_request_free(req);
809 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
812 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
813 * for operation \a opcode. Request would contain \a count buffers.
814 * Sizes of buffers are described in array \a lengths and buffers themselves
815 * are provided by a pointer \a bufs.
816 * Returns prepared request structure pointer or NULL on error.
818 struct ptlrpc_request *
819 ptlrpc_prep_req_pool(struct obd_import *imp,
820 __u32 version, int opcode,
821 int count, __u32 *lengths, char **bufs,
822 struct ptlrpc_request_pool *pool)
824 struct ptlrpc_request *request;
827 request = __ptlrpc_request_alloc(imp, pool);
831 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
832 lengths, bufs, NULL);
834 ptlrpc_request_free(request);
839 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
842 * Same as ptlrpc_prep_req_pool, but without pool
844 struct ptlrpc_request *
845 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
846 __u32 *lengths, char **bufs)
848 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
851 EXPORT_SYMBOL(ptlrpc_prep_req);
854 * Allocate and initialize new request set structure.
855 * Returns a pointer to the newly allocated set structure or NULL on error.
857 struct ptlrpc_request_set *ptlrpc_prep_set(void)
859 struct ptlrpc_request_set *set;
862 OBD_ALLOC(set, sizeof *set);
865 atomic_set(&set->set_refcount, 1);
866 CFS_INIT_LIST_HEAD(&set->set_requests);
867 init_waitqueue_head(&set->set_waitq);
868 atomic_set(&set->set_new_count, 0);
869 atomic_set(&set->set_remaining, 0);
870 spin_lock_init(&set->set_new_req_lock);
871 CFS_INIT_LIST_HEAD(&set->set_new_requests);
872 CFS_INIT_LIST_HEAD(&set->set_cblist);
873 set->set_max_inflight = UINT_MAX;
874 set->set_producer = NULL;
875 set->set_producer_arg = NULL;
880 EXPORT_SYMBOL(ptlrpc_prep_set);
883 * Allocate and initialize new request set structure with flow control
884 * extension. This extension allows to control the number of requests in-flight
885 * for the whole set. A callback function to generate requests must be provided
886 * and the request set will keep the number of requests sent over the wire to
888 * Returns a pointer to the newly allocated set structure or NULL on error.
890 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
894 struct ptlrpc_request_set *set;
896 set = ptlrpc_prep_set();
900 set->set_max_inflight = max;
901 set->set_producer = func;
902 set->set_producer_arg = arg;
906 EXPORT_SYMBOL(ptlrpc_prep_fcset);
909 * Wind down and free request set structure previously allocated with
911 * Ensures that all requests on the set have completed and removes
912 * all requests from the request list in a set.
913 * If any unsent request happen to be on the list, pretends that they got
914 * an error in flight and calls their completion handler.
916 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
924 /* Requests on the set should either all be completed, or all be new */
925 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
926 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
927 cfs_list_for_each (tmp, &set->set_requests) {
928 struct ptlrpc_request *req =
929 cfs_list_entry(tmp, struct ptlrpc_request,
932 LASSERT(req->rq_phase == expected_phase);
936 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
937 atomic_read(&set->set_remaining) == n, "%d / %d\n",
938 atomic_read(&set->set_remaining), n);
940 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
941 struct ptlrpc_request *req =
942 cfs_list_entry(tmp, struct ptlrpc_request,
944 cfs_list_del_init(&req->rq_set_chain);
946 LASSERT(req->rq_phase == expected_phase);
948 if (req->rq_phase == RQ_PHASE_NEW) {
949 ptlrpc_req_interpret(NULL, req, -EBADR);
950 atomic_dec(&set->set_remaining);
953 spin_lock(&req->rq_lock);
955 req->rq_invalid_rqset = 0;
956 spin_unlock(&req->rq_lock);
958 ptlrpc_req_finished (req);
961 LASSERT(atomic_read(&set->set_remaining) == 0);
963 ptlrpc_reqset_put(set);
966 EXPORT_SYMBOL(ptlrpc_set_destroy);
969 * Add a callback function \a fn to the set.
970 * This function would be called when all requests on this set are completed.
971 * The function will be passed \a data argument.
973 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
974 set_interpreter_func fn, void *data)
976 struct ptlrpc_set_cbdata *cbdata;
978 OBD_ALLOC_PTR(cbdata);
982 cbdata->psc_interpret = fn;
983 cbdata->psc_data = data;
984 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
988 EXPORT_SYMBOL(ptlrpc_set_add_cb);
991 * Add a new request to the general purpose request set.
992 * Assumes request reference from the caller.
994 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
995 struct ptlrpc_request *req)
997 LASSERT(cfs_list_empty(&req->rq_set_chain));
999 /* The set takes over the caller's request reference */
1000 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
1002 atomic_inc(&set->set_remaining);
1003 req->rq_queued_time = cfs_time_current();
1005 if (req->rq_reqmsg != NULL)
1006 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1008 if (set->set_producer != NULL)
1009 /* If the request set has a producer callback, the RPC must be
1010 * sent straight away */
1011 ptlrpc_send_new_req(req);
1013 EXPORT_SYMBOL(ptlrpc_set_add_req);
1016 * Add a request to a request with dedicated server thread
1017 * and wake the thread to make any necessary processing.
1018 * Currently only used for ptlrpcd.
1020 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1021 struct ptlrpc_request *req)
1023 struct ptlrpc_request_set *set = pc->pc_set;
1026 LASSERT(req->rq_set == NULL);
1027 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1029 spin_lock(&set->set_new_req_lock);
1031 * The set takes over the caller's request reference.
1034 req->rq_queued_time = cfs_time_current();
1035 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1036 count = atomic_inc_return(&set->set_new_count);
1037 spin_unlock(&set->set_new_req_lock);
1039 /* Only need to call wakeup once for the first entry. */
1041 wake_up(&set->set_waitq);
1043 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1044 * guarantee the async RPC can be processed ASAP, we have
1045 * no other better choice. It maybe fixed in future. */
1046 for (i = 0; i < pc->pc_npartners; i++)
1047 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1050 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1053 * Based on the current state of the import, determine if the request
1054 * can be sent, is an error, or should be delayed.
1056 * Returns true if this request should be delayed. If false, and
1057 * *status is set, then the request can not be sent and *status is the
1058 * error code. If false and status is 0, then request can be sent.
1060 * The imp->imp_lock must be held.
1062 static int ptlrpc_import_delay_req(struct obd_import *imp,
1063 struct ptlrpc_request *req, int *status)
1068 LASSERT (status != NULL);
1071 if (req->rq_ctx_init || req->rq_ctx_fini) {
1072 /* always allow ctx init/fini rpc go through */
1073 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1074 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1076 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1077 /* pings may safely race with umount */
1078 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1079 D_HA : D_ERROR, req, "IMP_CLOSED ");
1081 } else if (ptlrpc_send_limit_expired(req)) {
1082 /* probably doesn't need to be a D_ERROR after initial testing */
1083 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1085 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1086 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1087 /* allow CONNECT even if import is invalid */ ;
1088 if (atomic_read(&imp->imp_inval_count) != 0) {
1089 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1092 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1093 if (!imp->imp_deactive)
1094 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1095 *status = -ESHUTDOWN; /* bz 12940 */
1096 } else if (req->rq_import_generation != imp->imp_generation) {
1097 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1099 } else if (req->rq_send_state != imp->imp_state) {
1100 /* invalidate in progress - any requests should be drop */
1101 if (atomic_read(&imp->imp_inval_count) != 0) {
1102 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1104 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1105 *status = -EWOULDBLOCK;
1106 } else if (req->rq_allow_replay &&
1107 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1108 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1109 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1110 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1111 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1121 * Decide if the eror message regarding provided request \a req
1122 * should be printed to the console or not.
1123 * Makes it's decision on request status and other properties.
1124 * Returns 1 to print error on the system console or 0 if not.
1126 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1131 LASSERT(req->rq_reqmsg != NULL);
1132 opc = lustre_msg_get_opc(req->rq_reqmsg);
1134 /* Suppress particular reconnect errors which are to be expected. No
1135 * errors are suppressed for the initial connection on an import */
1136 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1137 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1139 /* Suppress timed out reconnect requests */
1140 if (req->rq_timedout)
1143 /* Suppress unavailable/again reconnect requests */
1144 err = lustre_msg_get_status(req->rq_repmsg);
1145 if (err == -ENODEV || err == -EAGAIN)
1153 * Check request processing status.
1154 * Returns the status.
1156 static int ptlrpc_check_status(struct ptlrpc_request *req)
1161 err = lustre_msg_get_status(req->rq_repmsg);
1162 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1163 struct obd_import *imp = req->rq_import;
1164 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1165 if (ptlrpc_console_allow(req))
1166 LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s,"
1167 " operation %s failed with %d.\n",
1168 imp->imp_obd->obd_name,
1170 imp->imp_connection->c_peer.nid),
1171 ll_opcode2str(opc), err);
1172 RETURN(err < 0 ? err : -EINVAL);
1176 DEBUG_REQ(D_INFO, req, "status is %d", err);
1177 } else if (err > 0) {
1178 /* XXX: translate this error from net to host */
1179 DEBUG_REQ(D_INFO, req, "status is %d", err);
1186 * save pre-versions of objects into request for replay.
1187 * Versions are obtained from server reply.
1190 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1192 struct lustre_msg *repmsg = req->rq_repmsg;
1193 struct lustre_msg *reqmsg = req->rq_reqmsg;
1194 __u64 *versions = lustre_msg_get_versions(repmsg);
1197 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1201 lustre_msg_set_versions(reqmsg, versions);
1202 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1203 versions[0], versions[1]);
1209 * Callback function called when client receives RPC reply for \a req.
1210 * Returns 0 on success or error code.
1211 * The return alue would be assigned to req->rq_status by the caller
1212 * as request processing status.
1213 * This function also decides if the request needs to be saved for later replay.
1215 static int after_reply(struct ptlrpc_request *req)
1217 struct obd_import *imp = req->rq_import;
1218 struct obd_device *obd = req->rq_import->imp_obd;
1220 struct timeval work_start;
1224 LASSERT(obd != NULL);
1225 /* repbuf must be unlinked */
1226 LASSERT(!req->rq_receiving_reply && !req->rq_reply_unlink);
1228 if (req->rq_reply_truncate) {
1229 if (ptlrpc_no_resend(req)) {
1230 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1231 " expected: %d, actual size: %d",
1232 req->rq_nob_received, req->rq_repbuf_len);
1236 sptlrpc_cli_free_repbuf(req);
1237 /* Pass the required reply buffer size (include
1238 * space for early reply).
1239 * NB: no need to roundup because alloc_repbuf
1240 * will roundup it */
1241 req->rq_replen = req->rq_nob_received;
1242 req->rq_nob_received = 0;
1243 spin_lock(&req->rq_lock);
1245 spin_unlock(&req->rq_lock);
1250 * NB Until this point, the whole of the incoming message,
1251 * including buflens, status etc is in the sender's byte order.
1253 rc = sptlrpc_cli_unwrap_reply(req);
1255 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1260 * Security layer unwrap might ask resend this request.
1265 rc = unpack_reply(req);
1269 /* retry indefinitely on EINPROGRESS */
1270 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1271 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1272 time_t now = cfs_time_current_sec();
1274 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1276 req->rq_nr_resend++;
1278 /* allocate new xid to avoid reply reconstruction */
1279 if (!req->rq_bulk) {
1280 /* new xid is already allocated for bulk in
1281 * ptlrpc_check_set() */
1282 req->rq_xid = ptlrpc_next_xid();
1283 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1284 "resend on EINPROGRESS");
1287 /* Readjust the timeout for current conditions */
1288 ptlrpc_at_set_req_timeout(req);
1289 /* delay resend to give a chance to the server to get ready.
1290 * The delay is increased by 1s on every resend and is capped to
1291 * the current request timeout (i.e. obd_timeout if AT is off,
1292 * or AT service time x 125% + 5s, see at_est2timeout) */
1293 if (req->rq_nr_resend > req->rq_timeout)
1294 req->rq_sent = now + req->rq_timeout;
1296 req->rq_sent = now + req->rq_nr_resend;
1301 do_gettimeofday(&work_start);
1302 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1303 if (obd->obd_svc_stats != NULL) {
1304 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1306 ptlrpc_lprocfs_rpc_sent(req, timediff);
1309 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1310 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1311 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1312 lustre_msg_get_type(req->rq_repmsg));
1316 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1317 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1318 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1319 ptlrpc_at_adj_net_latency(req,
1320 lustre_msg_get_service_time(req->rq_repmsg));
1322 rc = ptlrpc_check_status(req);
1323 imp->imp_connect_error = rc;
1327 * Either we've been evicted, or the server has failed for
1328 * some reason. Try to reconnect, and if that fails, punt to
1331 if (ll_rpc_recoverable_error(rc)) {
1332 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1333 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1336 ptlrpc_request_handle_notconn(req);
1341 * Let's look if server sent slv. Do it only for RPC with
1344 ldlm_cli_update_pool(req);
1348 * Store transno in reqmsg for replay.
1350 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1351 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1352 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1355 if (imp->imp_replayable) {
1356 spin_lock(&imp->imp_lock);
1358 * No point in adding already-committed requests to the replay
1359 * list, we will just remove them immediately. b=9829
1361 if (req->rq_transno != 0 &&
1363 lustre_msg_get_last_committed(req->rq_repmsg) ||
1365 /** version recovery */
1366 ptlrpc_save_versions(req);
1367 ptlrpc_retain_replayable_request(req, imp);
1368 } else if (req->rq_commit_cb != NULL &&
1369 list_empty(&req->rq_replay_list)) {
1370 /* NB: don't call rq_commit_cb if it's already on
1371 * rq_replay_list, ptlrpc_free_committed() will call
1372 * it later, see LU-3618 for details */
1373 spin_unlock(&imp->imp_lock);
1374 req->rq_commit_cb(req);
1375 spin_lock(&imp->imp_lock);
1379 * Replay-enabled imports return commit-status information.
1381 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1382 imp->imp_peer_committed_transno =
1383 lustre_msg_get_last_committed(req->rq_repmsg);
1386 ptlrpc_free_committed(imp);
1388 if (!cfs_list_empty(&imp->imp_replay_list)) {
1389 struct ptlrpc_request *last;
1391 last = cfs_list_entry(imp->imp_replay_list.prev,
1392 struct ptlrpc_request,
1395 * Requests with rq_replay stay on the list even if no
1396 * commit is expected.
1398 if (last->rq_transno > imp->imp_peer_committed_transno)
1399 ptlrpc_pinger_commit_expected(imp);
1402 spin_unlock(&imp->imp_lock);
1409 * Helper function to send request \a req over the network for the first time
1410 * Also adjusts request phase.
1411 * Returns 0 on success or error code.
1413 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1415 struct obd_import *imp = req->rq_import;
1419 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1420 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1421 (!req->rq_generation_set ||
1422 req->rq_import_generation == imp->imp_generation))
1425 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1427 spin_lock(&imp->imp_lock);
1429 if (!req->rq_generation_set)
1430 req->rq_import_generation = imp->imp_generation;
1432 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1433 spin_lock(&req->rq_lock);
1434 req->rq_waiting = 1;
1435 spin_unlock(&req->rq_lock);
1437 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1438 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1439 ptlrpc_import_state_name(req->rq_send_state),
1440 ptlrpc_import_state_name(imp->imp_state));
1441 LASSERT(cfs_list_empty(&req->rq_list));
1442 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1443 atomic_inc(&req->rq_import->imp_inflight);
1444 spin_unlock(&imp->imp_lock);
1449 spin_unlock(&imp->imp_lock);
1450 req->rq_status = rc;
1451 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1455 LASSERT(cfs_list_empty(&req->rq_list));
1456 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1457 atomic_inc(&req->rq_import->imp_inflight);
1458 spin_unlock(&imp->imp_lock);
1460 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1462 rc = sptlrpc_req_refresh_ctx(req, -1);
1465 req->rq_status = rc;
1468 spin_lock(&req->rq_lock);
1469 req->rq_wait_ctx = 1;
1470 spin_unlock(&req->rq_lock);
1475 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1476 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1477 imp->imp_obd->obd_uuid.uuid,
1478 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1479 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1480 lustre_msg_get_opc(req->rq_reqmsg));
1482 rc = ptl_send_rpc(req, 0);
1484 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1485 spin_lock(&req->rq_lock);
1486 req->rq_net_err = 1;
1487 spin_unlock(&req->rq_lock);
1493 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1498 LASSERT(set->set_producer != NULL);
1500 remaining = atomic_read(&set->set_remaining);
1502 /* populate the ->set_requests list with requests until we
1503 * reach the maximum number of RPCs in flight for this set */
1504 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1505 rc = set->set_producer(set, set->set_producer_arg);
1506 if (rc == -ENOENT) {
1507 /* no more RPC to produce */
1508 set->set_producer = NULL;
1509 set->set_producer_arg = NULL;
1514 RETURN((atomic_read(&set->set_remaining) - remaining));
1518 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1519 * and no more replies are expected.
1520 * (it is possible to get less replies than requests sent e.g. due to timed out
1521 * requests or requests that we had trouble to send out)
1523 * NOTE: This function contains a potential schedule point (cond_resched()).
1525 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1527 cfs_list_t *tmp, *next;
1528 int force_timer_recalc = 0;
1531 if (atomic_read(&set->set_remaining) == 0)
1534 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1535 struct ptlrpc_request *req =
1536 cfs_list_entry(tmp, struct ptlrpc_request,
1538 struct obd_import *imp = req->rq_import;
1539 int unregistered = 0;
1542 /* This schedule point is mainly for the ptlrpcd caller of this
1543 * function. Most ptlrpc sets are not long-lived and unbounded
1544 * in length, but at the least the set used by the ptlrpcd is.
1545 * Since the processing time is unbounded, we need to insert an
1546 * explicit schedule point to make the thread well-behaved.
1550 if (req->rq_phase == RQ_PHASE_NEW &&
1551 ptlrpc_send_new_req(req)) {
1552 force_timer_recalc = 1;
1555 /* delayed send - skip */
1556 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1559 /* delayed resend - skip */
1560 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1561 req->rq_sent > cfs_time_current_sec())
1564 if (!(req->rq_phase == RQ_PHASE_RPC ||
1565 req->rq_phase == RQ_PHASE_BULK ||
1566 req->rq_phase == RQ_PHASE_INTERPRET ||
1567 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1568 req->rq_phase == RQ_PHASE_COMPLETE)) {
1569 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1573 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1574 LASSERT(req->rq_next_phase != req->rq_phase);
1575 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1578 * Skip processing until reply is unlinked. We
1579 * can't return to pool before that and we can't
1580 * call interpret before that. We need to make
1581 * sure that all rdma transfers finished and will
1582 * not corrupt any data.
1584 if (ptlrpc_client_recv_or_unlink(req) ||
1585 ptlrpc_client_bulk_active(req))
1589 * Turn fail_loc off to prevent it from looping
1592 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1593 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1596 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1597 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1602 * Move to next phase if reply was successfully
1605 ptlrpc_rqphase_move(req, req->rq_next_phase);
1608 if (req->rq_phase == RQ_PHASE_COMPLETE)
1611 if (req->rq_phase == RQ_PHASE_INTERPRET)
1612 GOTO(interpret, req->rq_status);
1615 * Note that this also will start async reply unlink.
1617 if (req->rq_net_err && !req->rq_timedout) {
1618 ptlrpc_expire_one_request(req, 1);
1621 * Check if we still need to wait for unlink.
1623 if (ptlrpc_client_recv_or_unlink(req) ||
1624 ptlrpc_client_bulk_active(req))
1626 /* If there is no need to resend, fail it now. */
1627 if (req->rq_no_resend) {
1628 if (req->rq_status == 0)
1629 req->rq_status = -EIO;
1630 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1631 GOTO(interpret, req->rq_status);
1638 spin_lock(&req->rq_lock);
1639 req->rq_replied = 0;
1640 spin_unlock(&req->rq_lock);
1641 if (req->rq_status == 0)
1642 req->rq_status = -EIO;
1643 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1644 GOTO(interpret, req->rq_status);
1647 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1648 * so it sets rq_intr regardless of individual rpc
1649 * timeouts. The synchronous IO waiting path sets
1650 * rq_intr irrespective of whether ptlrpcd
1651 * has seen a timeout. Our policy is to only interpret
1652 * interrupted rpcs after they have timed out, so we
1653 * need to enforce that here.
1656 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1657 req->rq_wait_ctx)) {
1658 req->rq_status = -EINTR;
1659 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1660 GOTO(interpret, req->rq_status);
1663 if (req->rq_phase == RQ_PHASE_RPC) {
1664 if (req->rq_timedout || req->rq_resend ||
1665 req->rq_waiting || req->rq_wait_ctx) {
1668 if (!ptlrpc_unregister_reply(req, 1)) {
1669 ptlrpc_unregister_bulk(req, 1);
1673 spin_lock(&imp->imp_lock);
1674 if (ptlrpc_import_delay_req(imp, req, &status)){
1675 /* put on delay list - only if we wait
1676 * recovery finished - before send */
1677 cfs_list_del_init(&req->rq_list);
1678 cfs_list_add_tail(&req->rq_list,
1681 spin_unlock(&imp->imp_lock);
1686 req->rq_status = status;
1687 ptlrpc_rqphase_move(req,
1688 RQ_PHASE_INTERPRET);
1689 spin_unlock(&imp->imp_lock);
1690 GOTO(interpret, req->rq_status);
1692 if (ptlrpc_no_resend(req) &&
1693 !req->rq_wait_ctx) {
1694 req->rq_status = -ENOTCONN;
1695 ptlrpc_rqphase_move(req,
1696 RQ_PHASE_INTERPRET);
1697 spin_unlock(&imp->imp_lock);
1698 GOTO(interpret, req->rq_status);
1701 cfs_list_del_init(&req->rq_list);
1702 cfs_list_add_tail(&req->rq_list,
1703 &imp->imp_sending_list);
1705 spin_unlock(&imp->imp_lock);
1707 spin_lock(&req->rq_lock);
1708 req->rq_waiting = 0;
1709 spin_unlock(&req->rq_lock);
1711 if (req->rq_timedout || req->rq_resend) {
1712 /* This is re-sending anyways,
1713 * let's mark req as resend. */
1714 spin_lock(&req->rq_lock);
1716 spin_unlock(&req->rq_lock);
1720 if (!ptlrpc_unregister_bulk(req, 1))
1723 /* ensure previous bulk fails */
1724 old_xid = req->rq_xid;
1725 req->rq_xid = ptlrpc_next_xid();
1726 CDEBUG(D_HA, "resend bulk "
1729 old_xid, req->rq_xid);
1733 * rq_wait_ctx is only touched by ptlrpcd,
1734 * so no lock is needed here.
1736 status = sptlrpc_req_refresh_ctx(req, -1);
1739 req->rq_status = status;
1740 spin_lock(&req->rq_lock);
1741 req->rq_wait_ctx = 0;
1742 spin_unlock(&req->rq_lock);
1743 force_timer_recalc = 1;
1745 spin_lock(&req->rq_lock);
1746 req->rq_wait_ctx = 1;
1747 spin_unlock(&req->rq_lock);
1752 spin_lock(&req->rq_lock);
1753 req->rq_wait_ctx = 0;
1754 spin_unlock(&req->rq_lock);
1757 rc = ptl_send_rpc(req, 0);
1759 DEBUG_REQ(D_HA, req,
1760 "send failed: rc = %d", rc);
1761 force_timer_recalc = 1;
1762 spin_lock(&req->rq_lock);
1763 req->rq_net_err = 1;
1764 spin_unlock(&req->rq_lock);
1767 /* need to reset the timeout */
1768 force_timer_recalc = 1;
1771 spin_lock(&req->rq_lock);
1773 if (ptlrpc_client_early(req)) {
1774 ptlrpc_at_recv_early_reply(req);
1775 spin_unlock(&req->rq_lock);
1779 /* Still waiting for a reply? */
1780 if (ptlrpc_client_recv(req)) {
1781 spin_unlock(&req->rq_lock);
1785 /* Did we actually receive a reply? */
1786 if (!ptlrpc_client_replied(req)) {
1787 spin_unlock(&req->rq_lock);
1791 spin_unlock(&req->rq_lock);
1793 /* unlink from net because we are going to
1794 * swab in-place of reply buffer */
1795 unregistered = ptlrpc_unregister_reply(req, 1);
1799 req->rq_status = after_reply(req);
1803 /* If there is no bulk associated with this request,
1804 * then we're done and should let the interpreter
1805 * process the reply. Similarly if the RPC returned
1806 * an error, and therefore the bulk will never arrive.
1808 if (req->rq_bulk == NULL || req->rq_status < 0) {
1809 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1810 GOTO(interpret, req->rq_status);
1813 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1816 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1817 if (ptlrpc_client_bulk_active(req))
1820 if (req->rq_bulk->bd_failure) {
1821 /* The RPC reply arrived OK, but the bulk screwed
1822 * up! Dead weird since the server told us the RPC
1823 * was good after getting the REPLY for her GET or
1824 * the ACK for her PUT. */
1825 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1826 req->rq_status = -EIO;
1829 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1832 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1834 /* This moves to "unregistering" phase we need to wait for
1836 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1837 /* start async bulk unlink too */
1838 ptlrpc_unregister_bulk(req, 1);
1842 if (!ptlrpc_unregister_bulk(req, 1))
1845 /* When calling interpret receiving already should be
1847 LASSERT(!req->rq_receiving_reply);
1849 ptlrpc_req_interpret(env, req, req->rq_status);
1851 if (ptlrpcd_check_work(req)) {
1852 atomic_dec(&set->set_remaining);
1855 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1857 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1858 "Completed RPC pname:cluuid:pid:xid:nid:"
1859 "opc %s:%s:%d:"LPU64":%s:%d\n",
1860 current_comm(), imp->imp_obd->obd_uuid.uuid,
1861 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1862 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1863 lustre_msg_get_opc(req->rq_reqmsg));
1865 spin_lock(&imp->imp_lock);
1866 /* Request already may be not on sending or delaying list. This
1867 * may happen in the case of marking it erroneous for the case
1868 * ptlrpc_import_delay_req(req, status) find it impossible to
1869 * allow sending this rpc and returns *status != 0. */
1870 if (!cfs_list_empty(&req->rq_list)) {
1871 cfs_list_del_init(&req->rq_list);
1872 atomic_dec(&imp->imp_inflight);
1874 spin_unlock(&imp->imp_lock);
1876 atomic_dec(&set->set_remaining);
1877 wake_up_all(&imp->imp_recovery_waitq);
1879 if (set->set_producer) {
1880 /* produce a new request if possible */
1881 if (ptlrpc_set_producer(set) > 0)
1882 force_timer_recalc = 1;
1884 /* free the request that has just been completed
1885 * in order not to pollute set->set_requests */
1886 cfs_list_del_init(&req->rq_set_chain);
1887 spin_lock(&req->rq_lock);
1889 req->rq_invalid_rqset = 0;
1890 spin_unlock(&req->rq_lock);
1892 /* record rq_status to compute the final status later */
1893 if (req->rq_status != 0)
1894 set->set_rc = req->rq_status;
1895 ptlrpc_req_finished(req);
1899 /* If we hit an error, we want to recover promptly. */
1900 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1902 EXPORT_SYMBOL(ptlrpc_check_set);
1905 * Time out request \a req. is \a async_unlink is set, that means do not wait
1906 * until LNet actually confirms network buffer unlinking.
1907 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1909 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1911 struct obd_import *imp = req->rq_import;
1915 spin_lock(&req->rq_lock);
1916 req->rq_timedout = 1;
1917 spin_unlock(&req->rq_lock);
1919 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1920 "/real "CFS_DURATION_T"]",
1921 req->rq_net_err ? "failed due to network error" :
1922 ((req->rq_real_sent == 0 ||
1923 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1924 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1925 "timed out for sent delay" : "timed out for slow reply"),
1926 req->rq_sent, req->rq_real_sent);
1928 if (imp != NULL && obd_debug_peer_on_timeout)
1929 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1931 ptlrpc_unregister_reply(req, async_unlink);
1932 ptlrpc_unregister_bulk(req, async_unlink);
1934 if (obd_dump_on_timeout)
1935 libcfs_debug_dumplog();
1938 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1942 atomic_inc(&imp->imp_timeouts);
1944 /* The DLM server doesn't want recovery run on its imports. */
1945 if (imp->imp_dlm_fake)
1948 /* If this request is for recovery or other primordial tasks,
1949 * then error it out here. */
1950 if (req->rq_ctx_init || req->rq_ctx_fini ||
1951 req->rq_send_state != LUSTRE_IMP_FULL ||
1952 imp->imp_obd->obd_no_recov) {
1953 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1954 ptlrpc_import_state_name(req->rq_send_state),
1955 ptlrpc_import_state_name(imp->imp_state));
1956 spin_lock(&req->rq_lock);
1957 req->rq_status = -ETIMEDOUT;
1959 spin_unlock(&req->rq_lock);
1963 /* if a request can't be resent we can't wait for an answer after
1965 if (ptlrpc_no_resend(req)) {
1966 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1970 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1976 * Time out all uncompleted requests in request set pointed by \a data
1977 * Callback used when waiting on sets with l_wait_event.
1980 int ptlrpc_expired_set(void *data)
1982 struct ptlrpc_request_set *set = data;
1984 time_t now = cfs_time_current_sec();
1987 LASSERT(set != NULL);
1990 * A timeout expired. See which reqs it applies to...
1992 cfs_list_for_each (tmp, &set->set_requests) {
1993 struct ptlrpc_request *req =
1994 cfs_list_entry(tmp, struct ptlrpc_request,
1997 /* don't expire request waiting for context */
1998 if (req->rq_wait_ctx)
2001 /* Request in-flight? */
2002 if (!((req->rq_phase == RQ_PHASE_RPC &&
2003 !req->rq_waiting && !req->rq_resend) ||
2004 (req->rq_phase == RQ_PHASE_BULK)))
2007 if (req->rq_timedout || /* already dealt with */
2008 req->rq_deadline > now) /* not expired */
2011 /* Deal with this guy. Do it asynchronously to not block
2012 * ptlrpcd thread. */
2013 ptlrpc_expire_one_request(req, 1);
2017 * When waiting for a whole set, we always break out of the
2018 * sleep so we can recalculate the timeout, or enable interrupts
2019 * if everyone's timed out.
2023 EXPORT_SYMBOL(ptlrpc_expired_set);
2026 * Sets rq_intr flag in \a req under spinlock.
2028 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2030 spin_lock(&req->rq_lock);
2032 spin_unlock(&req->rq_lock);
2034 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2037 * Interrupts (sets interrupted flag) all uncompleted requests in
2038 * a set \a data. Callback for l_wait_event for interruptible waits.
2040 void ptlrpc_interrupted_set(void *data)
2042 struct ptlrpc_request_set *set = data;
2045 LASSERT(set != NULL);
2046 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2048 cfs_list_for_each(tmp, &set->set_requests) {
2049 struct ptlrpc_request *req =
2050 cfs_list_entry(tmp, struct ptlrpc_request,
2053 if (req->rq_phase != RQ_PHASE_RPC &&
2054 req->rq_phase != RQ_PHASE_UNREGISTERING)
2057 ptlrpc_mark_interrupted(req);
2060 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2063 * Get the smallest timeout in the set; this does NOT set a timeout.
2065 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2068 time_t now = cfs_time_current_sec();
2070 struct ptlrpc_request *req;
2074 cfs_list_for_each(tmp, &set->set_requests) {
2075 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2078 * Request in-flight?
2080 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2081 (req->rq_phase == RQ_PHASE_BULK) ||
2082 (req->rq_phase == RQ_PHASE_NEW)))
2086 * Already timed out.
2088 if (req->rq_timedout)
2094 if (req->rq_wait_ctx)
2097 if (req->rq_phase == RQ_PHASE_NEW)
2098 deadline = req->rq_sent;
2099 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2100 deadline = req->rq_sent;
2102 deadline = req->rq_sent + req->rq_timeout;
2104 if (deadline <= now) /* actually expired already */
2105 timeout = 1; /* ASAP */
2106 else if (timeout == 0 || timeout > deadline - now)
2107 timeout = deadline - now;
2111 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2114 * Send all unset request from the set and then wait untill all
2115 * requests in the set complete (either get a reply, timeout, get an
2116 * error or otherwise be interrupted).
2117 * Returns 0 on success or error code otherwise.
2119 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2122 struct ptlrpc_request *req;
2123 struct l_wait_info lwi;
2127 if (set->set_producer)
2128 (void)ptlrpc_set_producer(set);
2130 cfs_list_for_each(tmp, &set->set_requests) {
2131 req = cfs_list_entry(tmp, struct ptlrpc_request,
2133 if (req->rq_phase == RQ_PHASE_NEW)
2134 (void)ptlrpc_send_new_req(req);
2137 if (cfs_list_empty(&set->set_requests))
2141 timeout = ptlrpc_set_next_timeout(set);
2143 /* wait until all complete, interrupted, or an in-flight
2145 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2148 if (timeout == 0 && !cfs_signal_pending())
2150 * No requests are in-flight (ether timed out
2151 * or delayed), so we can allow interrupts.
2152 * We still want to block for a limited time,
2153 * so we allow interrupts during the timeout.
2155 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2157 ptlrpc_interrupted_set, set);
2160 * At least one request is in flight, so no
2161 * interrupts are allowed. Wait until all
2162 * complete, or an in-flight req times out.
2164 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2165 ptlrpc_expired_set, set);
2167 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2169 /* LU-769 - if we ignored the signal because it was already
2170 * pending when we started, we need to handle it now or we risk
2171 * it being ignored forever */
2172 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2173 cfs_signal_pending()) {
2174 sigset_t blocked_sigs =
2175 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2177 /* In fact we only interrupt for the "fatal" signals
2178 * like SIGINT or SIGKILL. We still ignore less
2179 * important signals since ptlrpc set is not easily
2180 * reentrant from userspace again */
2181 if (cfs_signal_pending())
2182 ptlrpc_interrupted_set(set);
2183 cfs_restore_sigs(blocked_sigs);
2186 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2188 /* -EINTR => all requests have been flagged rq_intr so next
2190 * -ETIMEDOUT => someone timed out. When all reqs have
2191 * timed out, signals are enabled allowing completion with
2193 * I don't really care if we go once more round the loop in
2194 * the error cases -eeb. */
2195 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2196 cfs_list_for_each(tmp, &set->set_requests) {
2197 req = cfs_list_entry(tmp, struct ptlrpc_request,
2199 spin_lock(&req->rq_lock);
2200 req->rq_invalid_rqset = 1;
2201 spin_unlock(&req->rq_lock);
2204 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2206 LASSERT(atomic_read(&set->set_remaining) == 0);
2208 rc = set->set_rc; /* rq_status of already freed requests if any */
2209 cfs_list_for_each(tmp, &set->set_requests) {
2210 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2212 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2213 if (req->rq_status != 0)
2214 rc = req->rq_status;
2217 if (set->set_interpret != NULL) {
2218 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2220 rc = interpreter (set, set->set_arg, rc);
2222 struct ptlrpc_set_cbdata *cbdata, *n;
2225 cfs_list_for_each_entry_safe(cbdata, n,
2226 &set->set_cblist, psc_item) {
2227 cfs_list_del_init(&cbdata->psc_item);
2228 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2231 OBD_FREE_PTR(cbdata);
2237 EXPORT_SYMBOL(ptlrpc_set_wait);
2240 * Helper fuction for request freeing.
2241 * Called when request count reached zero and request needs to be freed.
2242 * Removes request from all sorts of sending/replay lists it might be on,
2243 * frees network buffers if any are present.
2244 * If \a locked is set, that means caller is already holding import imp_lock
2245 * and so we no longer need to reobtain it (for certain lists manipulations)
2247 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2250 if (request == NULL) {
2255 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2256 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2257 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2258 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2259 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2260 LASSERTF(!request->rq_replay, "req %p\n", request);
2262 req_capsule_fini(&request->rq_pill);
2264 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2265 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2266 if (request->rq_import != NULL) {
2268 spin_lock(&request->rq_import->imp_lock);
2269 cfs_list_del_init(&request->rq_replay_list);
2271 spin_unlock(&request->rq_import->imp_lock);
2273 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2275 if (atomic_read(&request->rq_refcount) != 0) {
2276 DEBUG_REQ(D_ERROR, request,
2277 "freeing request with nonzero refcount");
2281 if (request->rq_repbuf != NULL)
2282 sptlrpc_cli_free_repbuf(request);
2283 if (request->rq_export != NULL) {
2284 class_export_put(request->rq_export);
2285 request->rq_export = NULL;
2287 if (request->rq_import != NULL) {
2288 class_import_put(request->rq_import);
2289 request->rq_import = NULL;
2291 if (request->rq_bulk != NULL)
2292 ptlrpc_free_bulk_pin(request->rq_bulk);
2294 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2295 sptlrpc_cli_free_reqbuf(request);
2297 if (request->rq_cli_ctx)
2298 sptlrpc_req_put_ctx(request, !locked);
2300 if (request->rq_pool)
2301 __ptlrpc_free_req_to_pool(request);
2303 ptlrpc_request_cache_free(request);
2307 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2309 * Drop one request reference. Must be called with import imp_lock held.
2310 * When reference count drops to zero, reuqest is freed.
2312 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2314 assert_spin_locked(&request->rq_import->imp_lock);
2315 (void)__ptlrpc_req_finished(request, 1);
2317 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2321 * Drops one reference count for request \a request.
2322 * \a locked set indicates that caller holds import imp_lock.
2323 * Frees the request whe reference count reaches zero.
2325 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2328 if (request == NULL)
2331 if (request == LP_POISON ||
2332 request->rq_reqmsg == LP_POISON) {
2333 CERROR("dereferencing freed request (bug 575)\n");
2338 DEBUG_REQ(D_INFO, request, "refcount now %u",
2339 atomic_read(&request->rq_refcount) - 1);
2341 if (atomic_dec_and_test(&request->rq_refcount)) {
2342 __ptlrpc_free_req(request, locked);
2350 * Drops one reference count for a request.
2352 void ptlrpc_req_finished(struct ptlrpc_request *request)
2354 __ptlrpc_req_finished(request, 0);
2356 EXPORT_SYMBOL(ptlrpc_req_finished);
2359 * Returns xid of a \a request
2361 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2363 return request->rq_xid;
2365 EXPORT_SYMBOL(ptlrpc_req_xid);
2368 * Disengage the client's reply buffer from the network
2369 * NB does _NOT_ unregister any client-side bulk.
2370 * IDEMPOTENT, but _not_ safe against concurrent callers.
2371 * The request owner (i.e. the thread doing the I/O) must call...
2372 * Returns 0 on success or 1 if unregistering cannot be made.
2374 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2377 struct l_wait_info lwi;
2382 LASSERT(!in_interrupt());
2385 * Let's setup deadline for reply unlink.
2387 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2388 async && request->rq_reply_deadline == 0)
2389 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2392 * Nothing left to do.
2394 if (!ptlrpc_client_recv_or_unlink(request))
2397 LNetMDUnlink(request->rq_reply_md_h);
2400 * Let's check it once again.
2402 if (!ptlrpc_client_recv_or_unlink(request))
2406 * Move to "Unregistering" phase as reply was not unlinked yet.
2408 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2411 * Do not wait for unlink to finish.
2417 * We have to l_wait_event() whatever the result, to give liblustre
2418 * a chance to run reply_in_callback(), and to make sure we've
2419 * unlinked before returning a req to the pool.
2423 /* The wq argument is ignored by user-space wait_event macros */
2424 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2425 &request->rq_set->set_waitq :
2426 &request->rq_reply_waitq;
2428 /* Network access will complete in finite time but the HUGE
2429 * timeout lets us CWARN for visibility of sluggish NALs */
2430 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2431 cfs_time_seconds(1), NULL, NULL);
2432 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2435 ptlrpc_rqphase_move(request, request->rq_next_phase);
2439 LASSERT(rc == -ETIMEDOUT);
2440 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2441 "rvcng=%d unlnk=%d/%d", request->rq_receiving_reply,
2442 request->rq_req_unlink, request->rq_reply_unlink);
2446 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2448 static void ptlrpc_free_request(struct ptlrpc_request *req)
2450 spin_lock(&req->rq_lock);
2452 spin_unlock(&req->rq_lock);
2454 if (req->rq_commit_cb != NULL)
2455 req->rq_commit_cb(req);
2456 cfs_list_del_init(&req->rq_replay_list);
2458 __ptlrpc_req_finished(req, 1);
2462 * the request is committed and dropped from the replay list of its import
2464 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2466 struct obd_import *imp = req->rq_import;
2468 spin_lock(&imp->imp_lock);
2469 if (cfs_list_empty(&req->rq_replay_list)) {
2470 spin_unlock(&imp->imp_lock);
2474 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2475 ptlrpc_free_request(req);
2477 spin_unlock(&imp->imp_lock);
2479 EXPORT_SYMBOL(ptlrpc_request_committed);
2482 * Iterates through replay_list on import and prunes
2483 * all requests have transno smaller than last_committed for the
2484 * import and don't have rq_replay set.
2485 * Since requests are sorted in transno order, stops when meetign first
2486 * transno bigger than last_committed.
2487 * caller must hold imp->imp_lock
2489 void ptlrpc_free_committed(struct obd_import *imp)
2491 struct ptlrpc_request *req, *saved;
2492 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2493 bool skip_committed_list = true;
2496 LASSERT(imp != NULL);
2497 assert_spin_locked(&imp->imp_lock);
2499 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2500 imp->imp_generation == imp->imp_last_generation_checked) {
2501 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2502 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2505 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2506 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2507 imp->imp_generation);
2509 if (imp->imp_generation != imp->imp_last_generation_checked)
2510 skip_committed_list = false;
2512 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2513 imp->imp_last_generation_checked = imp->imp_generation;
2515 cfs_list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2517 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2518 LASSERT(req != last_req);
2521 if (req->rq_transno == 0) {
2522 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2525 if (req->rq_import_generation < imp->imp_generation) {
2526 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2530 /* not yet committed */
2531 if (req->rq_transno > imp->imp_peer_committed_transno) {
2532 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2536 if (req->rq_replay) {
2537 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2538 cfs_list_move_tail(&req->rq_replay_list,
2539 &imp->imp_committed_list);
2543 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2544 imp->imp_peer_committed_transno);
2546 ptlrpc_free_request(req);
2549 if (skip_committed_list)
2552 cfs_list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2554 LASSERT(req->rq_transno != 0);
2555 if (req->rq_import_generation < imp->imp_generation) {
2556 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2557 ptlrpc_free_request(req);
2564 void ptlrpc_cleanup_client(struct obd_import *imp)
2569 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2572 * Schedule previously sent request for resend.
2573 * For bulk requests we assign new xid (to avoid problems with
2574 * lost replies and therefore several transfers landing into same buffer
2575 * from different sending attempts).
2577 void ptlrpc_resend_req(struct ptlrpc_request *req)
2579 DEBUG_REQ(D_HA, req, "going to resend");
2580 spin_lock(&req->rq_lock);
2582 /* Request got reply but linked to the import list still.
2583 Let ptlrpc_check_set() to process it. */
2584 if (ptlrpc_client_replied(req)) {
2585 spin_unlock(&req->rq_lock);
2586 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2590 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2591 req->rq_status = -EAGAIN;
2594 req->rq_net_err = 0;
2595 req->rq_timedout = 0;
2597 __u64 old_xid = req->rq_xid;
2599 /* ensure previous bulk fails */
2600 req->rq_xid = ptlrpc_next_xid();
2601 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2602 old_xid, req->rq_xid);
2604 ptlrpc_client_wake_req(req);
2605 spin_unlock(&req->rq_lock);
2607 EXPORT_SYMBOL(ptlrpc_resend_req);
2609 /* XXX: this function and rq_status are currently unused */
2610 void ptlrpc_restart_req(struct ptlrpc_request *req)
2612 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2613 req->rq_status = -ERESTARTSYS;
2615 spin_lock(&req->rq_lock);
2616 req->rq_restart = 1;
2617 req->rq_timedout = 0;
2618 ptlrpc_client_wake_req(req);
2619 spin_unlock(&req->rq_lock);
2621 EXPORT_SYMBOL(ptlrpc_restart_req);
2624 * Grab additional reference on a request \a req
2626 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2629 atomic_inc(&req->rq_refcount);
2632 EXPORT_SYMBOL(ptlrpc_request_addref);
2635 * Add a request to import replay_list.
2636 * Must be called under imp_lock
2638 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2639 struct obd_import *imp)
2643 assert_spin_locked(&imp->imp_lock);
2645 if (req->rq_transno == 0) {
2646 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2650 /* clear this for new requests that were resent as well
2651 as resent replayed requests. */
2652 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2654 /* don't re-add requests that have been replayed */
2655 if (!cfs_list_empty(&req->rq_replay_list))
2658 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2660 LASSERT(imp->imp_replayable);
2661 /* Balanced in ptlrpc_free_committed, usually. */
2662 ptlrpc_request_addref(req);
2663 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2664 struct ptlrpc_request *iter =
2665 cfs_list_entry(tmp, struct ptlrpc_request,
2668 /* We may have duplicate transnos if we create and then
2669 * open a file, or for closes retained if to match creating
2670 * opens, so use req->rq_xid as a secondary key.
2671 * (See bugs 684, 685, and 428.)
2672 * XXX no longer needed, but all opens need transnos!
2674 if (iter->rq_transno > req->rq_transno)
2677 if (iter->rq_transno == req->rq_transno) {
2678 LASSERT(iter->rq_xid != req->rq_xid);
2679 if (iter->rq_xid > req->rq_xid)
2683 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2687 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2689 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2692 * Send request and wait until it completes.
2693 * Returns request processing status.
2695 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2697 struct ptlrpc_request_set *set;
2701 LASSERT(req->rq_set == NULL);
2702 LASSERT(!req->rq_receiving_reply);
2704 set = ptlrpc_prep_set();
2706 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2710 /* for distributed debugging */
2711 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2713 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2714 ptlrpc_request_addref(req);
2715 ptlrpc_set_add_req(set, req);
2716 rc = ptlrpc_set_wait(set);
2717 ptlrpc_set_destroy(set);
2721 EXPORT_SYMBOL(ptlrpc_queue_wait);
2723 struct ptlrpc_replay_async_args {
2725 int praa_old_status;
2729 * Callback used for replayed requests reply processing.
2730 * In case of succesful reply calls registeresd request replay callback.
2731 * In case of error restart replay process.
2733 static int ptlrpc_replay_interpret(const struct lu_env *env,
2734 struct ptlrpc_request *req,
2735 void * data, int rc)
2737 struct ptlrpc_replay_async_args *aa = data;
2738 struct obd_import *imp = req->rq_import;
2741 atomic_dec(&imp->imp_replay_inflight);
2743 if (!ptlrpc_client_replied(req)) {
2744 CERROR("request replay timed out, restarting recovery\n");
2745 GOTO(out, rc = -ETIMEDOUT);
2748 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2749 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2750 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2751 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2753 /** VBR: check version failure */
2754 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2755 /** replay was failed due to version mismatch */
2756 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2757 spin_lock(&imp->imp_lock);
2758 imp->imp_vbr_failed = 1;
2759 imp->imp_no_lock_replay = 1;
2760 spin_unlock(&imp->imp_lock);
2761 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2763 /** The transno had better not change over replay. */
2764 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2765 lustre_msg_get_transno(req->rq_repmsg) ||
2766 lustre_msg_get_transno(req->rq_repmsg) == 0,
2768 lustre_msg_get_transno(req->rq_reqmsg),
2769 lustre_msg_get_transno(req->rq_repmsg));
2772 spin_lock(&imp->imp_lock);
2773 /** if replays by version then gap occur on server, no trust to locks */
2774 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2775 imp->imp_no_lock_replay = 1;
2776 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2777 spin_unlock(&imp->imp_lock);
2778 LASSERT(imp->imp_last_replay_transno);
2780 /* transaction number shouldn't be bigger than the latest replayed */
2781 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2782 DEBUG_REQ(D_ERROR, req,
2783 "Reported transno "LPU64" is bigger than the "
2784 "replayed one: "LPU64, req->rq_transno,
2785 lustre_msg_get_transno(req->rq_reqmsg));
2786 GOTO(out, rc = -EINVAL);
2789 DEBUG_REQ(D_HA, req, "got rep");
2791 /* let the callback do fixups, possibly including in the request */
2792 if (req->rq_replay_cb)
2793 req->rq_replay_cb(req);
2795 if (ptlrpc_client_replied(req) &&
2796 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2797 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2798 lustre_msg_get_status(req->rq_repmsg),
2799 aa->praa_old_status);
2801 /* Put it back for re-replay. */
2802 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2806 * Errors while replay can set transno to 0, but
2807 * imp_last_replay_transno shouldn't be set to 0 anyway
2809 if (req->rq_transno == 0)
2810 CERROR("Transno is 0 during replay!\n");
2812 /* continue with recovery */
2813 rc = ptlrpc_import_recovery_state_machine(imp);
2815 req->rq_send_state = aa->praa_old_state;
2818 /* this replay failed, so restart recovery */
2819 ptlrpc_connect_import(imp);
2825 * Prepares and queues request for replay.
2826 * Adds it to ptlrpcd queue for actual sending.
2827 * Returns 0 on success.
2829 int ptlrpc_replay_req(struct ptlrpc_request *req)
2831 struct ptlrpc_replay_async_args *aa;
2834 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2836 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2837 aa = ptlrpc_req_async_args(req);
2838 memset(aa, 0, sizeof *aa);
2840 /* Prepare request to be resent with ptlrpcd */
2841 aa->praa_old_state = req->rq_send_state;
2842 req->rq_send_state = LUSTRE_IMP_REPLAY;
2843 req->rq_phase = RQ_PHASE_NEW;
2844 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2846 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2848 req->rq_interpret_reply = ptlrpc_replay_interpret;
2849 /* Readjust the timeout for current conditions */
2850 ptlrpc_at_set_req_timeout(req);
2852 /* Tell server the net_latency, so the server can calculate how long
2853 * it should wait for next replay */
2854 lustre_msg_set_service_time(req->rq_reqmsg,
2855 ptlrpc_at_get_net_latency(req));
2856 DEBUG_REQ(D_HA, req, "REPLAY");
2858 atomic_inc(&req->rq_import->imp_replay_inflight);
2859 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2861 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2864 EXPORT_SYMBOL(ptlrpc_replay_req);
2867 * Aborts all in-flight request on import \a imp sending and delayed lists
2869 void ptlrpc_abort_inflight(struct obd_import *imp)
2871 cfs_list_t *tmp, *n;
2874 /* Make sure that no new requests get processed for this import.
2875 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2876 * this flag and then putting requests on sending_list or delayed_list.
2878 spin_lock(&imp->imp_lock);
2880 /* XXX locking? Maybe we should remove each request with the list
2881 * locked? Also, how do we know if the requests on the list are
2882 * being freed at this time?
2884 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2885 struct ptlrpc_request *req =
2886 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2888 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2890 spin_lock(&req->rq_lock);
2891 if (req->rq_import_generation < imp->imp_generation) {
2893 req->rq_status = -EIO;
2894 ptlrpc_client_wake_req(req);
2896 spin_unlock(&req->rq_lock);
2899 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2900 struct ptlrpc_request *req =
2901 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2903 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2905 spin_lock(&req->rq_lock);
2906 if (req->rq_import_generation < imp->imp_generation) {
2908 req->rq_status = -EIO;
2909 ptlrpc_client_wake_req(req);
2911 spin_unlock(&req->rq_lock);
2914 /* Last chance to free reqs left on the replay list, but we
2915 * will still leak reqs that haven't committed. */
2916 if (imp->imp_replayable)
2917 ptlrpc_free_committed(imp);
2919 spin_unlock(&imp->imp_lock);
2923 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2926 * Abort all uncompleted requests in request set \a set
2928 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2930 cfs_list_t *tmp, *pos;
2932 LASSERT(set != NULL);
2934 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2935 struct ptlrpc_request *req =
2936 cfs_list_entry(pos, struct ptlrpc_request,
2939 spin_lock(&req->rq_lock);
2940 if (req->rq_phase != RQ_PHASE_RPC) {
2941 spin_unlock(&req->rq_lock);
2946 req->rq_status = -EINTR;
2947 ptlrpc_client_wake_req(req);
2948 spin_unlock(&req->rq_lock);
2952 static __u64 ptlrpc_last_xid;
2953 static spinlock_t ptlrpc_last_xid_lock;
2956 * Initialize the XID for the node. This is common among all requests on
2957 * this node, and only requires the property that it is monotonically
2958 * increasing. It does not need to be sequential. Since this is also used
2959 * as the RDMA match bits, it is important that a single client NOT have
2960 * the same match bits for two different in-flight requests, hence we do
2961 * NOT want to have an XID per target or similar.
2963 * To avoid an unlikely collision between match bits after a client reboot
2964 * (which would deliver old data into the wrong RDMA buffer) initialize
2965 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2966 * If the time is clearly incorrect, we instead use a 62-bit random number.
2967 * In the worst case the random number will overflow 1M RPCs per second in
2968 * 9133 years, or permutations thereof.
2970 #define YEAR_2004 (1ULL << 30)
2971 void ptlrpc_init_xid(void)
2973 time_t now = cfs_time_current_sec();
2975 spin_lock_init(&ptlrpc_last_xid_lock);
2976 if (now < YEAR_2004) {
2977 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2978 ptlrpc_last_xid >>= 2;
2979 ptlrpc_last_xid |= (1ULL << 61);
2981 ptlrpc_last_xid = (__u64)now << 20;
2984 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
2985 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
2986 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
2990 * Increase xid and returns resulting new value to the caller.
2992 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2993 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2994 * itself uses the last bulk xid needed, so the server can determine the
2995 * the number of bulk transfers from the RPC XID and a bitmask. The starting
2996 * xid must align to a power-of-two value.
2998 * This is assumed to be true due to the initial ptlrpc_last_xid
2999 * value also being initialized to a power-of-two value. LU-1431
3001 __u64 ptlrpc_next_xid(void)
3005 spin_lock(&ptlrpc_last_xid_lock);
3006 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3007 ptlrpc_last_xid = next;
3008 spin_unlock(&ptlrpc_last_xid_lock);
3012 EXPORT_SYMBOL(ptlrpc_next_xid);
3015 * Get a glimpse at what next xid value might have been.
3016 * Returns possible next xid.
3018 __u64 ptlrpc_sample_next_xid(void)
3020 #if BITS_PER_LONG == 32
3021 /* need to avoid possible word tearing on 32-bit systems */
3024 spin_lock(&ptlrpc_last_xid_lock);
3025 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3026 spin_unlock(&ptlrpc_last_xid_lock);
3030 /* No need to lock, since returned value is racy anyways */
3031 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3034 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3037 * Functions for operating ptlrpc workers.
3039 * A ptlrpc work is a function which will be running inside ptlrpc context.
3040 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3042 * 1. after a work is created, it can be used many times, that is:
3043 * handler = ptlrpcd_alloc_work();
3044 * ptlrpcd_queue_work();
3046 * queue it again when necessary:
3047 * ptlrpcd_queue_work();
3048 * ptlrpcd_destroy_work();
3049 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3050 * it will only be queued once in any time. Also as its name implies, it may
3051 * have delay before it really runs by ptlrpcd thread.
3053 struct ptlrpc_work_async_args {
3054 int (*cb)(const struct lu_env *, void *);
3058 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3060 /* re-initialize the req */
3061 req->rq_timeout = obd_timeout;
3062 req->rq_sent = cfs_time_current_sec();
3063 req->rq_deadline = req->rq_sent + req->rq_timeout;
3064 req->rq_reply_deadline = req->rq_deadline;
3065 req->rq_phase = RQ_PHASE_INTERPRET;
3066 req->rq_next_phase = RQ_PHASE_COMPLETE;
3067 req->rq_xid = ptlrpc_next_xid();
3068 req->rq_import_generation = req->rq_import->imp_generation;
3070 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3073 static int work_interpreter(const struct lu_env *env,
3074 struct ptlrpc_request *req, void *data, int rc)
3076 struct ptlrpc_work_async_args *arg = data;
3078 LASSERT(ptlrpcd_check_work(req));
3079 LASSERT(arg->cb != NULL);
3081 rc = arg->cb(env, arg->cbdata);
3083 list_del_init(&req->rq_set_chain);
3086 if (atomic_dec_return(&req->rq_refcount) > 1) {
3087 atomic_set(&req->rq_refcount, 2);
3088 ptlrpcd_add_work_req(req);
3093 static int worker_format;
3095 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3097 return req->rq_pill.rc_fmt == (void *)&worker_format;
3101 * Create a work for ptlrpc.
3103 void *ptlrpcd_alloc_work(struct obd_import *imp,
3104 int (*cb)(const struct lu_env *, void *), void *cbdata)
3106 struct ptlrpc_request *req = NULL;
3107 struct ptlrpc_work_async_args *args;
3113 RETURN(ERR_PTR(-EINVAL));
3115 /* copy some code from deprecated fakereq. */
3116 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3118 CERROR("ptlrpc: run out of memory!\n");
3119 RETURN(ERR_PTR(-ENOMEM));
3122 req->rq_send_state = LUSTRE_IMP_FULL;
3123 req->rq_type = PTL_RPC_MSG_REQUEST;
3124 req->rq_import = class_import_get(imp);
3125 req->rq_export = NULL;
3126 req->rq_interpret_reply = work_interpreter;
3127 /* don't want reply */
3128 req->rq_receiving_reply = 0;
3129 req->rq_req_unlink = req->rq_reply_unlink = 0;
3130 req->rq_no_delay = req->rq_no_resend = 1;
3131 req->rq_pill.rc_fmt = (void *)&worker_format;
3133 spin_lock_init(&req->rq_lock);
3134 CFS_INIT_LIST_HEAD(&req->rq_list);
3135 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
3136 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
3137 CFS_INIT_LIST_HEAD(&req->rq_history_list);
3138 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
3139 init_waitqueue_head(&req->rq_reply_waitq);
3140 init_waitqueue_head(&req->rq_set_waitq);
3141 atomic_set(&req->rq_refcount, 1);
3143 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3144 args = ptlrpc_req_async_args(req);
3146 args->cbdata = cbdata;
3150 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3152 void ptlrpcd_destroy_work(void *handler)
3154 struct ptlrpc_request *req = handler;
3157 ptlrpc_req_finished(req);
3159 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3161 int ptlrpcd_queue_work(void *handler)
3163 struct ptlrpc_request *req = handler;
3166 * Check if the req is already being queued.
3168 * Here comes a trick: it lacks a way of checking if a req is being
3169 * processed reliably in ptlrpc. Here I have to use refcount of req
3170 * for this purpose. This is okay because the caller should use this
3171 * req as opaque data. - Jinshan
3173 LASSERT(atomic_read(&req->rq_refcount) > 0);
3174 if (atomic_inc_return(&req->rq_refcount) == 2)
3175 ptlrpcd_add_work_req(req);
3178 EXPORT_SYMBOL(ptlrpcd_queue_work);