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
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 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
51 static int ptlrpcd_check_work(struct ptlrpc_request *req);
54 * Initialize passed in client structure \a cl.
56 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
57 struct ptlrpc_client *cl)
59 cl->cli_request_portal = req_portal;
60 cl->cli_reply_portal = rep_portal;
63 EXPORT_SYMBOL(ptlrpc_init_client);
66 * Return PortalRPC connection for remore uud \a uuid
68 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
70 struct ptlrpc_connection *c;
72 lnet_process_id_t peer;
75 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
76 * before accessing its values. */
77 /* coverity[uninit_use_in_call] */
78 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
80 CNETERR("cannot find peer %s!\n", uuid->uuid);
84 c = ptlrpc_connection_get(peer, self, uuid);
86 memcpy(c->c_remote_uuid.uuid,
87 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
90 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
94 EXPORT_SYMBOL(ptlrpc_uuid_to_connection);
97 * Allocate and initialize new bulk descriptor on the sender.
98 * Returns pointer to the descriptor or NULL on error.
100 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned npages, unsigned max_brw,
101 unsigned type, unsigned portal)
103 struct ptlrpc_bulk_desc *desc;
106 OBD_ALLOC(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[npages]));
110 spin_lock_init(&desc->bd_lock);
111 init_waitqueue_head(&desc->bd_waitq);
112 desc->bd_max_iov = npages;
113 desc->bd_iov_count = 0;
114 desc->bd_portal = portal;
115 desc->bd_type = type;
116 desc->bd_md_count = 0;
117 LASSERT(max_brw > 0);
118 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
119 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
120 * node. Negotiated ocd_brw_size will always be <= this number. */
121 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
122 LNetInvalidateHandle(&desc->bd_mds[i]);
128 * Prepare bulk descriptor for specified outgoing request \a req that
129 * can fit \a npages * pages. \a type is bulk type. \a portal is where
130 * the bulk to be sent. Used on client-side.
131 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
134 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
135 unsigned npages, unsigned max_brw,
136 unsigned type, unsigned portal)
138 struct obd_import *imp = req->rq_import;
139 struct ptlrpc_bulk_desc *desc;
142 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
143 desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
147 desc->bd_import_generation = req->rq_import_generation;
148 desc->bd_import = class_import_get(imp);
151 desc->bd_cbid.cbid_fn = client_bulk_callback;
152 desc->bd_cbid.cbid_arg = desc;
154 /* This makes req own desc, and free it when she frees herself */
159 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
162 * Add a page \a page to the bulk descriptor \a desc.
163 * Data to transfer in the page starts at offset \a pageoffset and
164 * amount of data to transfer from the page is \a len
166 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
167 struct page *page, int pageoffset, int len, int pin)
169 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
170 LASSERT(page != NULL);
171 LASSERT(pageoffset >= 0);
173 LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
178 page_cache_get(page);
180 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
182 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
185 * Uninitialize and free bulk descriptor \a desc.
186 * Works on bulk descriptors both from server and client side.
188 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin)
193 LASSERT(desc != NULL);
194 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
195 LASSERT(desc->bd_md_count == 0); /* network hands off */
196 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
198 sptlrpc_enc_pool_put_pages(desc);
201 class_export_put(desc->bd_export);
203 class_import_put(desc->bd_import);
206 for (i = 0; i < desc->bd_iov_count ; i++)
207 page_cache_release(desc->bd_iov[i].kiov_page);
210 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
211 bd_iov[desc->bd_max_iov]));
214 EXPORT_SYMBOL(__ptlrpc_free_bulk);
217 * Set server timelimit for this req, i.e. how long are we willing to wait
218 * for reply before timing out this request.
220 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
226 LASSERT(req->rq_import);
229 /* non-AT settings */
231 * \a imp_server_timeout means this is reverse import and
232 * we send (currently only) ASTs to the client and cannot afford
233 * to wait too long for the reply, otherwise the other client
234 * (because of which we are sending this request) would
235 * timeout waiting for us
237 req->rq_timeout = req->rq_import->imp_server_timeout ?
238 obd_timeout / 2 : obd_timeout;
240 at = &req->rq_import->imp_at;
241 idx = import_at_get_index(req->rq_import,
242 req->rq_request_portal);
243 serv_est = at_get(&at->iat_service_estimate[idx]);
244 req->rq_timeout = at_est2timeout(serv_est);
246 /* We could get even fancier here, using history to predict increased
249 /* Let the server know what this RPC timeout is by putting it in the
251 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
253 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
255 /* Adjust max service estimate based on server value */
256 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
257 unsigned int serv_est)
263 LASSERT(req->rq_import);
264 at = &req->rq_import->imp_at;
266 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
267 /* max service estimates are tracked on the server side,
268 so just keep minimal history here */
269 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
271 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
272 "has changed from %d to %d\n",
273 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
274 oldse, at_get(&at->iat_service_estimate[idx]));
277 /* Expected network latency per remote node (secs) */
278 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
280 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
283 /* Adjust expected network latency */
284 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
285 unsigned int service_time)
287 unsigned int nl, oldnl;
289 time_t now = cfs_time_current_sec();
291 LASSERT(req->rq_import);
293 if (service_time > now - req->rq_sent + 3) {
294 /* bz16408, however, this can also happen if early reply
295 * is lost and client RPC is expired and resent, early reply
296 * or reply of original RPC can still be fit in reply buffer
297 * of resent RPC, now client is measuring time from the
298 * resent time, but server sent back service time of original
301 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
302 D_ADAPTTO : D_WARNING,
303 "Reported service time %u > total measured time "
304 CFS_DURATION_T"\n", service_time,
305 cfs_time_sub(now, req->rq_sent));
309 /* Network latency is total time less server processing time */
310 nl = max_t(int, now - req->rq_sent -
311 service_time, 0) + 1; /* st rounding */
312 at = &req->rq_import->imp_at;
314 oldnl = at_measured(&at->iat_net_latency, nl);
316 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
317 "has changed from %d to %d\n",
318 req->rq_import->imp_obd->obd_name,
320 &req->rq_import->imp_connection->c_remote_uuid),
321 oldnl, at_get(&at->iat_net_latency));
324 static int unpack_reply(struct ptlrpc_request *req)
328 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
329 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
331 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
336 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
338 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
345 * Handle an early reply message, called with the rq_lock held.
346 * If anything goes wrong just ignore it - same as if it never happened
348 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
349 __must_hold(&req->rq_lock)
351 struct ptlrpc_request *early_req;
357 spin_unlock(&req->rq_lock);
359 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
361 spin_lock(&req->rq_lock);
365 rc = unpack_reply(early_req);
367 /* Expecting to increase the service time estimate here */
368 ptlrpc_at_adj_service(req,
369 lustre_msg_get_timeout(early_req->rq_repmsg));
370 ptlrpc_at_adj_net_latency(req,
371 lustre_msg_get_service_time(early_req->rq_repmsg));
374 sptlrpc_cli_finish_early_reply(early_req);
377 spin_lock(&req->rq_lock);
381 /* Adjust the local timeout for this req */
382 ptlrpc_at_set_req_timeout(req);
384 spin_lock(&req->rq_lock);
385 olddl = req->rq_deadline;
386 /* server assumes it now has rq_timeout from when the request
387 * arrived, so the client should give it at least that long.
388 * since we don't know the arrival time we'll use the original
390 req->rq_deadline = req->rq_sent + req->rq_timeout +
391 ptlrpc_at_get_net_latency(req);
393 DEBUG_REQ(D_ADAPTTO, req,
394 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
395 "("CFS_DURATION_T"s)", req->rq_early_count,
396 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
397 cfs_time_sub(req->rq_deadline, olddl));
402 struct kmem_cache *request_cache;
404 int ptlrpc_request_cache_init(void)
406 request_cache = kmem_cache_create("ptlrpc_cache",
407 sizeof(struct ptlrpc_request),
408 0, SLAB_HWCACHE_ALIGN, NULL);
409 return request_cache == NULL ? -ENOMEM : 0;
412 void ptlrpc_request_cache_fini(void)
414 kmem_cache_destroy(request_cache);
417 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
419 struct ptlrpc_request *req;
421 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
425 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
427 OBD_SLAB_FREE_PTR(req, request_cache);
431 * Wind down request pool \a pool.
432 * Frees all requests from the pool too
434 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
436 struct list_head *l, *tmp;
437 struct ptlrpc_request *req;
439 LASSERT(pool != NULL);
441 spin_lock(&pool->prp_lock);
442 list_for_each_safe(l, tmp, &pool->prp_req_list) {
443 req = list_entry(l, struct ptlrpc_request, rq_list);
444 list_del(&req->rq_list);
445 LASSERT(req->rq_reqbuf);
446 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
447 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
448 ptlrpc_request_cache_free(req);
450 spin_unlock(&pool->prp_lock);
451 OBD_FREE(pool, sizeof(*pool));
453 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
456 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
458 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
463 while (size < pool->prp_rq_size)
466 LASSERTF(list_empty(&pool->prp_req_list) ||
467 size == pool->prp_rq_size,
468 "Trying to change pool size with nonempty pool "
469 "from %d to %d bytes\n", pool->prp_rq_size, size);
471 spin_lock(&pool->prp_lock);
472 pool->prp_rq_size = size;
473 for (i = 0; i < num_rq; i++) {
474 struct ptlrpc_request *req;
475 struct lustre_msg *msg;
477 spin_unlock(&pool->prp_lock);
478 req = ptlrpc_request_cache_alloc(GFP_NOFS);
481 OBD_ALLOC_LARGE(msg, size);
483 ptlrpc_request_cache_free(req);
486 req->rq_reqbuf = msg;
487 req->rq_reqbuf_len = size;
489 spin_lock(&pool->prp_lock);
490 list_add_tail(&req->rq_list, &pool->prp_req_list);
492 spin_unlock(&pool->prp_lock);
495 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
498 * Create and initialize new request pool with given attributes:
499 * \a num_rq - initial number of requests to create for the pool
500 * \a msgsize - maximum message size possible for requests in thid pool
501 * \a populate_pool - function to be called when more requests need to be added
503 * Returns pointer to newly created pool or NULL on error.
505 struct ptlrpc_request_pool *
506 ptlrpc_init_rq_pool(int num_rq, int msgsize,
507 void (*populate_pool)(struct ptlrpc_request_pool *, int))
509 struct ptlrpc_request_pool *pool;
511 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
515 /* Request next power of two for the allocation, because internally
516 kernel would do exactly this */
518 spin_lock_init(&pool->prp_lock);
519 INIT_LIST_HEAD(&pool->prp_req_list);
520 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
521 pool->prp_populate = populate_pool;
523 populate_pool(pool, num_rq);
525 if (list_empty(&pool->prp_req_list)) {
526 /* have not allocated a single request for the pool */
527 OBD_FREE(pool, sizeof(struct ptlrpc_request_pool));
532 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
535 * Fetches one request from pool \a pool
537 static struct ptlrpc_request *
538 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
540 struct ptlrpc_request *request;
541 struct lustre_msg *reqbuf;
546 spin_lock(&pool->prp_lock);
548 /* See if we have anything in a pool, and bail out if nothing,
549 * in writeout path, where this matters, this is safe to do, because
550 * nothing is lost in this case, and when some in-flight requests
551 * complete, this code will be called again. */
552 if (unlikely(list_empty(&pool->prp_req_list))) {
553 spin_unlock(&pool->prp_lock);
557 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
559 list_del_init(&request->rq_list);
560 spin_unlock(&pool->prp_lock);
562 LASSERT(request->rq_reqbuf);
563 LASSERT(request->rq_pool);
565 reqbuf = request->rq_reqbuf;
566 memset(request, 0, sizeof(*request));
567 request->rq_reqbuf = reqbuf;
568 request->rq_reqbuf_len = pool->prp_rq_size;
569 request->rq_pool = pool;
575 * Returns freed \a request to pool.
577 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
579 struct ptlrpc_request_pool *pool = request->rq_pool;
581 spin_lock(&pool->prp_lock);
582 LASSERT(list_empty(&request->rq_list));
583 LASSERT(!request->rq_receiving_reply);
584 list_add_tail(&request->rq_list, &pool->prp_req_list);
585 spin_unlock(&pool->prp_lock);
588 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
589 __u32 version, int opcode,
590 int count, __u32 *lengths, char **bufs,
591 struct ptlrpc_cli_ctx *ctx)
593 struct obd_import *imp = request->rq_import;
598 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
600 rc = sptlrpc_req_get_ctx(request);
605 sptlrpc_req_set_flavor(request, opcode);
607 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
610 LASSERT(!request->rq_pool);
614 lustre_msg_add_version(request->rq_reqmsg, version);
615 request->rq_send_state = LUSTRE_IMP_FULL;
616 request->rq_type = PTL_RPC_MSG_REQUEST;
617 request->rq_export = NULL;
619 request->rq_req_cbid.cbid_fn = request_out_callback;
620 request->rq_req_cbid.cbid_arg = request;
622 request->rq_reply_cbid.cbid_fn = reply_in_callback;
623 request->rq_reply_cbid.cbid_arg = request;
625 request->rq_reply_deadline = 0;
626 request->rq_phase = RQ_PHASE_NEW;
627 request->rq_next_phase = RQ_PHASE_UNDEFINED;
629 request->rq_request_portal = imp->imp_client->cli_request_portal;
630 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
632 ptlrpc_at_set_req_timeout(request);
634 spin_lock_init(&request->rq_lock);
635 INIT_LIST_HEAD(&request->rq_list);
636 INIT_LIST_HEAD(&request->rq_timed_list);
637 INIT_LIST_HEAD(&request->rq_replay_list);
638 INIT_LIST_HEAD(&request->rq_ctx_chain);
639 INIT_LIST_HEAD(&request->rq_set_chain);
640 INIT_LIST_HEAD(&request->rq_history_list);
641 INIT_LIST_HEAD(&request->rq_exp_list);
642 init_waitqueue_head(&request->rq_reply_waitq);
643 init_waitqueue_head(&request->rq_set_waitq);
644 request->rq_xid = ptlrpc_next_xid();
645 atomic_set(&request->rq_refcount, 1);
647 lustre_msg_set_opc(request->rq_reqmsg, opcode);
651 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
653 class_import_put(imp);
657 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
658 __u32 version, int opcode, char **bufs,
659 struct ptlrpc_cli_ctx *ctx)
663 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
664 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
665 request->rq_pill.rc_area[RCL_CLIENT],
668 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
671 * Pack request buffers for network transfer, performing necessary encryption
672 * steps if necessary.
674 int ptlrpc_request_pack(struct ptlrpc_request *request,
675 __u32 version, int opcode)
678 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
682 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
683 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
684 * have to send old ptlrpc_body to keep interoprability with these
687 * Only three kinds of server->client RPCs so far:
692 * XXX This should be removed whenever we drop the interoprability with
693 * the these old clients.
695 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
696 opcode == LDLM_GL_CALLBACK)
697 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
698 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
702 EXPORT_SYMBOL(ptlrpc_request_pack);
705 * Helper function to allocate new request on import \a imp
706 * and possibly using existing request from pool \a pool if provided.
707 * Returns allocated request structure with import field filled or
711 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
712 struct ptlrpc_request_pool *pool)
714 struct ptlrpc_request *request = NULL;
717 request = ptlrpc_prep_req_from_pool(pool);
720 request = ptlrpc_request_cache_alloc(GFP_NOFS);
723 LASSERTF((unsigned long)imp > 0x1000, "%p\n", imp);
724 LASSERT(imp != LP_POISON);
725 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
727 LASSERT(imp->imp_client != LP_POISON);
729 request->rq_import = class_import_get(imp);
731 CERROR("request allocation out of memory\n");
738 * Helper function for creating a request.
739 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
740 * buffer structures according to capsule template \a format.
741 * Returns allocated request structure pointer or NULL on error.
743 static struct ptlrpc_request *
744 ptlrpc_request_alloc_internal(struct obd_import *imp,
745 struct ptlrpc_request_pool * pool,
746 const struct req_format *format)
748 struct ptlrpc_request *request;
750 request = __ptlrpc_request_alloc(imp, pool);
754 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
755 req_capsule_set(&request->rq_pill, format);
760 * Allocate new request structure for import \a imp and initialize its
761 * buffer structure according to capsule template \a format.
763 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
764 const struct req_format *format)
766 return ptlrpc_request_alloc_internal(imp, NULL, format);
768 EXPORT_SYMBOL(ptlrpc_request_alloc);
771 * Allocate new request structure for import \a imp from pool \a pool and
772 * initialize its buffer structure according to capsule template \a format.
774 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
775 struct ptlrpc_request_pool * pool,
776 const struct req_format *format)
778 return ptlrpc_request_alloc_internal(imp, pool, format);
780 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
783 * For requests not from pool, free memory of the request structure.
784 * For requests obtained from a pool earlier, return request back to pool.
786 void ptlrpc_request_free(struct ptlrpc_request *request)
788 if (request->rq_pool)
789 __ptlrpc_free_req_to_pool(request);
791 ptlrpc_request_cache_free(request);
793 EXPORT_SYMBOL(ptlrpc_request_free);
796 * Allocate new request for operatione \a opcode and immediatelly pack it for
798 * Only used for simple requests like OBD_PING where the only important
799 * part of the request is operation itself.
800 * Returns allocated request or NULL on error.
802 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
803 const struct req_format *format,
804 __u32 version, int opcode)
806 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
810 rc = ptlrpc_request_pack(req, version, opcode);
812 ptlrpc_request_free(req);
818 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
821 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
822 * for operation \a opcode. Request would contain \a count buffers.
823 * Sizes of buffers are described in array \a lengths and buffers themselves
824 * are provided by a pointer \a bufs.
825 * Returns prepared request structure pointer or NULL on error.
827 struct ptlrpc_request *
828 ptlrpc_prep_req_pool(struct obd_import *imp,
829 __u32 version, int opcode,
830 int count, __u32 *lengths, char **bufs,
831 struct ptlrpc_request_pool *pool)
833 struct ptlrpc_request *request;
836 request = __ptlrpc_request_alloc(imp, pool);
840 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
841 lengths, bufs, NULL);
843 ptlrpc_request_free(request);
848 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
851 * Same as ptlrpc_prep_req_pool, but without pool
853 struct ptlrpc_request *
854 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
855 __u32 *lengths, char **bufs)
857 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
860 EXPORT_SYMBOL(ptlrpc_prep_req);
863 * Allocate and initialize new request set structure.
864 * Returns a pointer to the newly allocated set structure or NULL on error.
866 struct ptlrpc_request_set *ptlrpc_prep_set(void)
868 struct ptlrpc_request_set *set;
871 OBD_ALLOC(set, sizeof *set);
874 atomic_set(&set->set_refcount, 1);
875 INIT_LIST_HEAD(&set->set_requests);
876 init_waitqueue_head(&set->set_waitq);
877 atomic_set(&set->set_new_count, 0);
878 atomic_set(&set->set_remaining, 0);
879 spin_lock_init(&set->set_new_req_lock);
880 INIT_LIST_HEAD(&set->set_new_requests);
881 INIT_LIST_HEAD(&set->set_cblist);
882 set->set_max_inflight = UINT_MAX;
883 set->set_producer = NULL;
884 set->set_producer_arg = NULL;
889 EXPORT_SYMBOL(ptlrpc_prep_set);
892 * Allocate and initialize new request set structure with flow control
893 * extension. This extension allows to control the number of requests in-flight
894 * for the whole set. A callback function to generate requests must be provided
895 * and the request set will keep the number of requests sent over the wire to
897 * Returns a pointer to the newly allocated set structure or NULL on error.
899 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
903 struct ptlrpc_request_set *set;
905 set = ptlrpc_prep_set();
909 set->set_max_inflight = max;
910 set->set_producer = func;
911 set->set_producer_arg = arg;
915 EXPORT_SYMBOL(ptlrpc_prep_fcset);
918 * Wind down and free request set structure previously allocated with
920 * Ensures that all requests on the set have completed and removes
921 * all requests from the request list in a set.
922 * If any unsent request happen to be on the list, pretends that they got
923 * an error in flight and calls their completion handler.
925 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
927 struct list_head *tmp;
928 struct list_head *next;
933 /* Requests on the set should either all be completed, or all be new */
934 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
935 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
936 list_for_each(tmp, &set->set_requests) {
937 struct ptlrpc_request *req =
938 list_entry(tmp, struct ptlrpc_request,
941 LASSERT(req->rq_phase == expected_phase);
945 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
946 atomic_read(&set->set_remaining) == n, "%d / %d\n",
947 atomic_read(&set->set_remaining), n);
949 list_for_each_safe(tmp, next, &set->set_requests) {
950 struct ptlrpc_request *req =
951 list_entry(tmp, struct ptlrpc_request,
953 list_del_init(&req->rq_set_chain);
955 LASSERT(req->rq_phase == expected_phase);
957 if (req->rq_phase == RQ_PHASE_NEW) {
958 ptlrpc_req_interpret(NULL, req, -EBADR);
959 atomic_dec(&set->set_remaining);
962 spin_lock(&req->rq_lock);
964 req->rq_invalid_rqset = 0;
965 spin_unlock(&req->rq_lock);
967 ptlrpc_req_finished (req);
970 LASSERT(atomic_read(&set->set_remaining) == 0);
972 ptlrpc_reqset_put(set);
975 EXPORT_SYMBOL(ptlrpc_set_destroy);
978 * Add a callback function \a fn to the set.
979 * This function would be called when all requests on this set are completed.
980 * The function will be passed \a data argument.
982 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
983 set_interpreter_func fn, void *data)
985 struct ptlrpc_set_cbdata *cbdata;
987 OBD_ALLOC_PTR(cbdata);
991 cbdata->psc_interpret = fn;
992 cbdata->psc_data = data;
993 list_add_tail(&cbdata->psc_item, &set->set_cblist);
997 EXPORT_SYMBOL(ptlrpc_set_add_cb);
1000 * Add a new request to the general purpose request set.
1001 * Assumes request reference from the caller.
1003 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1004 struct ptlrpc_request *req)
1006 LASSERT(list_empty(&req->rq_set_chain));
1008 /* The set takes over the caller's request reference */
1009 list_add_tail(&req->rq_set_chain, &set->set_requests);
1011 atomic_inc(&set->set_remaining);
1012 req->rq_queued_time = cfs_time_current();
1014 if (req->rq_reqmsg != NULL)
1015 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1017 if (set->set_producer != NULL)
1018 /* If the request set has a producer callback, the RPC must be
1019 * sent straight away */
1020 ptlrpc_send_new_req(req);
1022 EXPORT_SYMBOL(ptlrpc_set_add_req);
1025 * Add a request to a request with dedicated server thread
1026 * and wake the thread to make any necessary processing.
1027 * Currently only used for ptlrpcd.
1029 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1030 struct ptlrpc_request *req)
1032 struct ptlrpc_request_set *set = pc->pc_set;
1035 LASSERT(req->rq_set == NULL);
1036 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1038 spin_lock(&set->set_new_req_lock);
1040 * The set takes over the caller's request reference.
1043 req->rq_queued_time = cfs_time_current();
1044 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1045 count = atomic_inc_return(&set->set_new_count);
1046 spin_unlock(&set->set_new_req_lock);
1048 /* Only need to call wakeup once for the first entry. */
1050 wake_up(&set->set_waitq);
1052 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1053 * guarantee the async RPC can be processed ASAP, we have
1054 * no other better choice. It maybe fixed in future. */
1055 for (i = 0; i < pc->pc_npartners; i++)
1056 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1059 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1062 * Based on the current state of the import, determine if the request
1063 * can be sent, is an error, or should be delayed.
1065 * Returns true if this request should be delayed. If false, and
1066 * *status is set, then the request can not be sent and *status is the
1067 * error code. If false and status is 0, then request can be sent.
1069 * The imp->imp_lock must be held.
1071 static int ptlrpc_import_delay_req(struct obd_import *imp,
1072 struct ptlrpc_request *req, int *status)
1077 LASSERT (status != NULL);
1080 if (req->rq_ctx_init || req->rq_ctx_fini) {
1081 /* always allow ctx init/fini rpc go through */
1082 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1083 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1085 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1086 /* pings may safely race with umount */
1087 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1088 D_HA : D_ERROR, req, "IMP_CLOSED ");
1090 } else if (ptlrpc_send_limit_expired(req)) {
1091 /* probably doesn't need to be a D_ERROR after initial testing*/
1092 DEBUG_REQ(D_HA, req, "send limit expired ");
1093 *status = -ETIMEDOUT;
1094 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1095 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1096 /* allow CONNECT even if import is invalid */ ;
1097 if (atomic_read(&imp->imp_inval_count) != 0) {
1098 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1101 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1102 if (!imp->imp_deactive)
1103 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1104 *status = -ESHUTDOWN; /* bz 12940 */
1105 } else if (req->rq_import_generation != imp->imp_generation) {
1106 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1108 } else if (req->rq_send_state != imp->imp_state) {
1109 /* invalidate in progress - any requests should be drop */
1110 if (atomic_read(&imp->imp_inval_count) != 0) {
1111 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1113 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1114 *status = -EWOULDBLOCK;
1115 } else if (req->rq_allow_replay &&
1116 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1117 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1118 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1119 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1120 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1130 * Decide if the eror message regarding provided request \a req
1131 * should be printed to the console or not.
1132 * Makes it's decision on request status and other properties.
1133 * Returns 1 to print error on the system console or 0 if not.
1135 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1140 LASSERT(req->rq_reqmsg != NULL);
1141 opc = lustre_msg_get_opc(req->rq_reqmsg);
1143 /* Suppress particular reconnect errors which are to be expected. No
1144 * errors are suppressed for the initial connection on an import */
1145 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1146 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1148 /* Suppress timed out reconnect requests */
1149 if (req->rq_timedout)
1152 /* Suppress unavailable/again reconnect requests */
1153 err = lustre_msg_get_status(req->rq_repmsg);
1154 if (err == -ENODEV || err == -EAGAIN)
1162 * Check request processing status.
1163 * Returns the status.
1165 static int ptlrpc_check_status(struct ptlrpc_request *req)
1170 err = lustre_msg_get_status(req->rq_repmsg);
1171 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1172 struct obd_import *imp = req->rq_import;
1173 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1174 if (ptlrpc_console_allow(req))
1175 LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s,"
1176 " operation %s failed with %d.\n",
1177 imp->imp_obd->obd_name,
1179 imp->imp_connection->c_peer.nid),
1180 ll_opcode2str(opc), err);
1181 RETURN(err < 0 ? err : -EINVAL);
1185 DEBUG_REQ(D_INFO, req, "status is %d", err);
1186 } else if (err > 0) {
1187 /* XXX: translate this error from net to host */
1188 DEBUG_REQ(D_INFO, req, "status is %d", err);
1195 * save pre-versions of objects into request for replay.
1196 * Versions are obtained from server reply.
1199 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1201 struct lustre_msg *repmsg = req->rq_repmsg;
1202 struct lustre_msg *reqmsg = req->rq_reqmsg;
1203 __u64 *versions = lustre_msg_get_versions(repmsg);
1206 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1210 lustre_msg_set_versions(reqmsg, versions);
1211 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1212 versions[0], versions[1]);
1218 * Callback function called when client receives RPC reply for \a req.
1219 * Returns 0 on success or error code.
1220 * The return alue would be assigned to req->rq_status by the caller
1221 * as request processing status.
1222 * This function also decides if the request needs to be saved for later replay.
1224 static int after_reply(struct ptlrpc_request *req)
1226 struct obd_import *imp = req->rq_import;
1227 struct obd_device *obd = req->rq_import->imp_obd;
1229 struct timeval work_start;
1233 LASSERT(obd != NULL);
1234 /* repbuf must be unlinked */
1235 LASSERT(!req->rq_receiving_reply && !req->rq_reply_unlink);
1237 if (req->rq_reply_truncate) {
1238 if (ptlrpc_no_resend(req)) {
1239 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1240 " expected: %d, actual size: %d",
1241 req->rq_nob_received, req->rq_repbuf_len);
1245 sptlrpc_cli_free_repbuf(req);
1246 /* Pass the required reply buffer size (include
1247 * space for early reply).
1248 * NB: no need to roundup because alloc_repbuf
1249 * will roundup it */
1250 req->rq_replen = req->rq_nob_received;
1251 req->rq_nob_received = 0;
1252 spin_lock(&req->rq_lock);
1254 spin_unlock(&req->rq_lock);
1259 * NB Until this point, the whole of the incoming message,
1260 * including buflens, status etc is in the sender's byte order.
1262 rc = sptlrpc_cli_unwrap_reply(req);
1264 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1269 * Security layer unwrap might ask resend this request.
1274 rc = unpack_reply(req);
1278 /* retry indefinitely on EINPROGRESS */
1279 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1280 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1281 time_t now = cfs_time_current_sec();
1283 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1284 spin_lock(&req->rq_lock);
1286 spin_unlock(&req->rq_lock);
1287 req->rq_nr_resend++;
1289 /* allocate new xid to avoid reply reconstruction */
1290 if (!req->rq_bulk) {
1291 /* new xid is already allocated for bulk in
1292 * ptlrpc_check_set() */
1293 req->rq_xid = ptlrpc_next_xid();
1294 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1295 "resend on EINPROGRESS");
1298 /* Readjust the timeout for current conditions */
1299 ptlrpc_at_set_req_timeout(req);
1300 /* delay resend to give a chance to the server to get ready.
1301 * The delay is increased by 1s on every resend and is capped to
1302 * the current request timeout (i.e. obd_timeout if AT is off,
1303 * or AT service time x 125% + 5s, see at_est2timeout) */
1304 if (req->rq_nr_resend > req->rq_timeout)
1305 req->rq_sent = now + req->rq_timeout;
1307 req->rq_sent = now + req->rq_nr_resend;
1312 do_gettimeofday(&work_start);
1313 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1314 if (obd->obd_svc_stats != NULL) {
1315 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1317 ptlrpc_lprocfs_rpc_sent(req, timediff);
1320 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1321 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1322 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1323 lustre_msg_get_type(req->rq_repmsg));
1327 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1328 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1329 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1330 ptlrpc_at_adj_net_latency(req,
1331 lustre_msg_get_service_time(req->rq_repmsg));
1333 rc = ptlrpc_check_status(req);
1334 imp->imp_connect_error = rc;
1338 * Either we've been evicted, or the server has failed for
1339 * some reason. Try to reconnect, and if that fails, punt to
1342 if (ll_rpc_recoverable_error(rc)) {
1343 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1344 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1347 ptlrpc_request_handle_notconn(req);
1352 * Let's look if server sent slv. Do it only for RPC with
1355 ldlm_cli_update_pool(req);
1359 * Store transno in reqmsg for replay.
1361 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1362 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1363 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1366 if (imp->imp_replayable) {
1367 spin_lock(&imp->imp_lock);
1369 * No point in adding already-committed requests to the replay
1370 * list, we will just remove them immediately. b=9829
1372 if (req->rq_transno != 0 &&
1374 lustre_msg_get_last_committed(req->rq_repmsg) ||
1376 /** version recovery */
1377 ptlrpc_save_versions(req);
1378 ptlrpc_retain_replayable_request(req, imp);
1379 } else if (req->rq_commit_cb != NULL &&
1380 list_empty(&req->rq_replay_list)) {
1381 /* NB: don't call rq_commit_cb if it's already on
1382 * rq_replay_list, ptlrpc_free_committed() will call
1383 * it later, see LU-3618 for details */
1384 spin_unlock(&imp->imp_lock);
1385 req->rq_commit_cb(req);
1386 spin_lock(&imp->imp_lock);
1390 * Replay-enabled imports return commit-status information.
1392 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1393 imp->imp_peer_committed_transno =
1394 lustre_msg_get_last_committed(req->rq_repmsg);
1397 ptlrpc_free_committed(imp);
1399 if (!list_empty(&imp->imp_replay_list)) {
1400 struct ptlrpc_request *last;
1402 last = list_entry(imp->imp_replay_list.prev,
1403 struct ptlrpc_request,
1406 * Requests with rq_replay stay on the list even if no
1407 * commit is expected.
1409 if (last->rq_transno > imp->imp_peer_committed_transno)
1410 ptlrpc_pinger_commit_expected(imp);
1413 spin_unlock(&imp->imp_lock);
1420 * Helper function to send request \a req over the network for the first time
1421 * Also adjusts request phase.
1422 * Returns 0 on success or error code.
1424 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1426 struct obd_import *imp = req->rq_import;
1430 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1431 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1432 (!req->rq_generation_set ||
1433 req->rq_import_generation == imp->imp_generation))
1436 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1438 spin_lock(&imp->imp_lock);
1440 if (!req->rq_generation_set)
1441 req->rq_import_generation = imp->imp_generation;
1443 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1444 spin_lock(&req->rq_lock);
1445 req->rq_waiting = 1;
1446 spin_unlock(&req->rq_lock);
1448 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1449 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1450 ptlrpc_import_state_name(req->rq_send_state),
1451 ptlrpc_import_state_name(imp->imp_state));
1452 LASSERT(list_empty(&req->rq_list));
1453 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1454 atomic_inc(&req->rq_import->imp_inflight);
1455 spin_unlock(&imp->imp_lock);
1460 spin_unlock(&imp->imp_lock);
1461 req->rq_status = rc;
1462 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1466 LASSERT(list_empty(&req->rq_list));
1467 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1468 atomic_inc(&req->rq_import->imp_inflight);
1469 spin_unlock(&imp->imp_lock);
1471 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1473 rc = sptlrpc_req_refresh_ctx(req, -1);
1476 req->rq_status = rc;
1479 spin_lock(&req->rq_lock);
1480 req->rq_wait_ctx = 1;
1481 spin_unlock(&req->rq_lock);
1486 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1487 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1488 imp->imp_obd->obd_uuid.uuid,
1489 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1490 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1491 lustre_msg_get_opc(req->rq_reqmsg));
1493 rc = ptl_send_rpc(req, 0);
1495 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1496 spin_lock(&req->rq_lock);
1497 req->rq_net_err = 1;
1498 spin_unlock(&req->rq_lock);
1504 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1509 LASSERT(set->set_producer != NULL);
1511 remaining = atomic_read(&set->set_remaining);
1513 /* populate the ->set_requests list with requests until we
1514 * reach the maximum number of RPCs in flight for this set */
1515 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1516 rc = set->set_producer(set, set->set_producer_arg);
1517 if (rc == -ENOENT) {
1518 /* no more RPC to produce */
1519 set->set_producer = NULL;
1520 set->set_producer_arg = NULL;
1525 RETURN((atomic_read(&set->set_remaining) - remaining));
1529 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1530 * and no more replies are expected.
1531 * (it is possible to get less replies than requests sent e.g. due to timed out
1532 * requests or requests that we had trouble to send out)
1534 * NOTE: This function contains a potential schedule point (cond_resched()).
1536 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1538 struct list_head *tmp, *next;
1539 struct list_head comp_reqs;
1540 int force_timer_recalc = 0;
1543 if (atomic_read(&set->set_remaining) == 0)
1546 INIT_LIST_HEAD(&comp_reqs);
1547 list_for_each_safe(tmp, next, &set->set_requests) {
1548 struct ptlrpc_request *req =
1549 list_entry(tmp, struct ptlrpc_request,
1551 struct obd_import *imp = req->rq_import;
1552 int unregistered = 0;
1555 /* This schedule point is mainly for the ptlrpcd caller of this
1556 * function. Most ptlrpc sets are not long-lived and unbounded
1557 * in length, but at the least the set used by the ptlrpcd is.
1558 * Since the processing time is unbounded, we need to insert an
1559 * explicit schedule point to make the thread well-behaved.
1563 if (req->rq_phase == RQ_PHASE_NEW &&
1564 ptlrpc_send_new_req(req)) {
1565 force_timer_recalc = 1;
1568 /* delayed send - skip */
1569 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1572 /* delayed resend - skip */
1573 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1574 req->rq_sent > cfs_time_current_sec())
1577 if (!(req->rq_phase == RQ_PHASE_RPC ||
1578 req->rq_phase == RQ_PHASE_BULK ||
1579 req->rq_phase == RQ_PHASE_INTERPRET ||
1580 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1581 req->rq_phase == RQ_PHASE_COMPLETE)) {
1582 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1586 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1587 LASSERT(req->rq_next_phase != req->rq_phase);
1588 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1591 * Skip processing until reply is unlinked. We
1592 * can't return to pool before that and we can't
1593 * call interpret before that. We need to make
1594 * sure that all rdma transfers finished and will
1595 * not corrupt any data.
1597 if (ptlrpc_client_recv_or_unlink(req) ||
1598 ptlrpc_client_bulk_active(req))
1602 * Turn fail_loc off to prevent it from looping
1605 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1606 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1609 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1610 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1615 * Move to next phase if reply was successfully
1618 ptlrpc_rqphase_move(req, req->rq_next_phase);
1621 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1622 list_move_tail(&req->rq_set_chain, &comp_reqs);
1626 if (req->rq_phase == RQ_PHASE_INTERPRET)
1627 GOTO(interpret, req->rq_status);
1630 * Note that this also will start async reply unlink.
1632 if (req->rq_net_err && !req->rq_timedout) {
1633 ptlrpc_expire_one_request(req, 1);
1636 * Check if we still need to wait for unlink.
1638 if (ptlrpc_client_recv_or_unlink(req) ||
1639 ptlrpc_client_bulk_active(req))
1641 /* If there is no need to resend, fail it now. */
1642 if (req->rq_no_resend) {
1643 if (req->rq_status == 0)
1644 req->rq_status = -EIO;
1645 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1646 GOTO(interpret, req->rq_status);
1653 spin_lock(&req->rq_lock);
1654 req->rq_replied = 0;
1655 spin_unlock(&req->rq_lock);
1656 if (req->rq_status == 0)
1657 req->rq_status = -EIO;
1658 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1659 GOTO(interpret, req->rq_status);
1662 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1663 * so it sets rq_intr regardless of individual rpc
1664 * timeouts. The synchronous IO waiting path sets
1665 * rq_intr irrespective of whether ptlrpcd
1666 * has seen a timeout. Our policy is to only interpret
1667 * interrupted rpcs after they have timed out, so we
1668 * need to enforce that here.
1671 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1672 req->rq_wait_ctx)) {
1673 req->rq_status = -EINTR;
1674 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1675 GOTO(interpret, req->rq_status);
1678 if (req->rq_phase == RQ_PHASE_RPC) {
1679 if (req->rq_timedout || req->rq_resend ||
1680 req->rq_waiting || req->rq_wait_ctx) {
1683 if (!ptlrpc_unregister_reply(req, 1)) {
1684 ptlrpc_unregister_bulk(req, 1);
1688 spin_lock(&imp->imp_lock);
1689 if (ptlrpc_import_delay_req(imp, req, &status)){
1690 /* put on delay list - only if we wait
1691 * recovery finished - before send */
1692 list_del_init(&req->rq_list);
1693 list_add_tail(&req->rq_list,
1696 spin_unlock(&imp->imp_lock);
1701 req->rq_status = status;
1702 ptlrpc_rqphase_move(req,
1703 RQ_PHASE_INTERPRET);
1704 spin_unlock(&imp->imp_lock);
1705 GOTO(interpret, req->rq_status);
1707 if (ptlrpc_no_resend(req) &&
1708 !req->rq_wait_ctx) {
1709 req->rq_status = -ENOTCONN;
1710 ptlrpc_rqphase_move(req,
1711 RQ_PHASE_INTERPRET);
1712 spin_unlock(&imp->imp_lock);
1713 GOTO(interpret, req->rq_status);
1716 list_del_init(&req->rq_list);
1717 list_add_tail(&req->rq_list,
1718 &imp->imp_sending_list);
1720 spin_unlock(&imp->imp_lock);
1722 spin_lock(&req->rq_lock);
1723 req->rq_waiting = 0;
1724 spin_unlock(&req->rq_lock);
1726 if (req->rq_timedout || req->rq_resend) {
1727 /* This is re-sending anyways,
1728 * let's mark req as resend. */
1729 spin_lock(&req->rq_lock);
1731 spin_unlock(&req->rq_lock);
1735 if (!ptlrpc_unregister_bulk(req, 1))
1738 /* ensure previous bulk fails */
1739 old_xid = req->rq_xid;
1740 req->rq_xid = ptlrpc_next_xid();
1741 CDEBUG(D_HA, "resend bulk "
1744 old_xid, req->rq_xid);
1748 * rq_wait_ctx is only touched by ptlrpcd,
1749 * so no lock is needed here.
1751 status = sptlrpc_req_refresh_ctx(req, -1);
1754 req->rq_status = status;
1755 spin_lock(&req->rq_lock);
1756 req->rq_wait_ctx = 0;
1757 spin_unlock(&req->rq_lock);
1758 force_timer_recalc = 1;
1760 spin_lock(&req->rq_lock);
1761 req->rq_wait_ctx = 1;
1762 spin_unlock(&req->rq_lock);
1767 spin_lock(&req->rq_lock);
1768 req->rq_wait_ctx = 0;
1769 spin_unlock(&req->rq_lock);
1772 rc = ptl_send_rpc(req, 0);
1774 DEBUG_REQ(D_HA, req,
1775 "send failed: rc = %d", rc);
1776 force_timer_recalc = 1;
1777 spin_lock(&req->rq_lock);
1778 req->rq_net_err = 1;
1779 spin_unlock(&req->rq_lock);
1782 /* need to reset the timeout */
1783 force_timer_recalc = 1;
1786 spin_lock(&req->rq_lock);
1788 if (ptlrpc_client_early(req)) {
1789 ptlrpc_at_recv_early_reply(req);
1790 spin_unlock(&req->rq_lock);
1794 /* Still waiting for a reply? */
1795 if (ptlrpc_client_recv(req)) {
1796 spin_unlock(&req->rq_lock);
1800 /* Did we actually receive a reply? */
1801 if (!ptlrpc_client_replied(req)) {
1802 spin_unlock(&req->rq_lock);
1806 spin_unlock(&req->rq_lock);
1808 /* unlink from net because we are going to
1809 * swab in-place of reply buffer */
1810 unregistered = ptlrpc_unregister_reply(req, 1);
1814 req->rq_status = after_reply(req);
1818 /* If there is no bulk associated with this request,
1819 * then we're done and should let the interpreter
1820 * process the reply. Similarly if the RPC returned
1821 * an error, and therefore the bulk will never arrive.
1823 if (req->rq_bulk == NULL || req->rq_status < 0) {
1824 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1825 GOTO(interpret, req->rq_status);
1828 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1831 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1832 if (ptlrpc_client_bulk_active(req))
1835 if (req->rq_bulk->bd_failure) {
1836 /* The RPC reply arrived OK, but the bulk screwed
1837 * up! Dead weird since the server told us the RPC
1838 * was good after getting the REPLY for her GET or
1839 * the ACK for her PUT. */
1840 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1841 req->rq_status = -EIO;
1844 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1847 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1849 /* This moves to "unregistering" phase we need to wait for
1851 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1852 /* start async bulk unlink too */
1853 ptlrpc_unregister_bulk(req, 1);
1857 if (!ptlrpc_unregister_bulk(req, 1))
1860 /* When calling interpret receiving already should be
1862 LASSERT(!req->rq_receiving_reply);
1864 ptlrpc_req_interpret(env, req, req->rq_status);
1866 if (ptlrpcd_check_work(req)) {
1867 atomic_dec(&set->set_remaining);
1870 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1872 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1873 "Completed RPC pname:cluuid:pid:xid:nid:"
1874 "opc %s:%s:%d:"LPU64":%s:%d\n",
1875 current_comm(), imp->imp_obd->obd_uuid.uuid,
1876 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1877 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1878 lustre_msg_get_opc(req->rq_reqmsg));
1880 spin_lock(&imp->imp_lock);
1881 /* Request already may be not on sending or delaying list. This
1882 * may happen in the case of marking it erroneous for the case
1883 * ptlrpc_import_delay_req(req, status) find it impossible to
1884 * allow sending this rpc and returns *status != 0. */
1885 if (!list_empty(&req->rq_list)) {
1886 list_del_init(&req->rq_list);
1887 atomic_dec(&imp->imp_inflight);
1889 spin_unlock(&imp->imp_lock);
1891 atomic_dec(&set->set_remaining);
1892 wake_up_all(&imp->imp_recovery_waitq);
1894 if (set->set_producer) {
1895 /* produce a new request if possible */
1896 if (ptlrpc_set_producer(set) > 0)
1897 force_timer_recalc = 1;
1899 /* free the request that has just been completed
1900 * in order not to pollute set->set_requests */
1901 list_del_init(&req->rq_set_chain);
1902 spin_lock(&req->rq_lock);
1904 req->rq_invalid_rqset = 0;
1905 spin_unlock(&req->rq_lock);
1907 /* record rq_status to compute the final status later */
1908 if (req->rq_status != 0)
1909 set->set_rc = req->rq_status;
1910 ptlrpc_req_finished(req);
1912 list_move_tail(&req->rq_set_chain, &comp_reqs);
1916 /* move completed request at the head of list so it's easier for
1917 * caller to find them */
1918 list_splice(&comp_reqs, &set->set_requests);
1920 /* If we hit an error, we want to recover promptly. */
1921 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1923 EXPORT_SYMBOL(ptlrpc_check_set);
1926 * Time out request \a req. is \a async_unlink is set, that means do not wait
1927 * until LNet actually confirms network buffer unlinking.
1928 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1930 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1932 struct obd_import *imp = req->rq_import;
1936 spin_lock(&req->rq_lock);
1937 req->rq_timedout = 1;
1938 spin_unlock(&req->rq_lock);
1940 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1941 "/real "CFS_DURATION_T"]",
1942 req->rq_net_err ? "failed due to network error" :
1943 ((req->rq_real_sent == 0 ||
1944 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1945 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1946 "timed out for sent delay" : "timed out for slow reply"),
1947 req->rq_sent, req->rq_real_sent);
1949 if (imp != NULL && obd_debug_peer_on_timeout)
1950 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1952 ptlrpc_unregister_reply(req, async_unlink);
1953 ptlrpc_unregister_bulk(req, async_unlink);
1955 if (obd_dump_on_timeout)
1956 libcfs_debug_dumplog();
1959 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1963 atomic_inc(&imp->imp_timeouts);
1965 /* The DLM server doesn't want recovery run on its imports. */
1966 if (imp->imp_dlm_fake)
1969 /* If this request is for recovery or other primordial tasks,
1970 * then error it out here. */
1971 if (req->rq_ctx_init || req->rq_ctx_fini ||
1972 req->rq_send_state != LUSTRE_IMP_FULL ||
1973 imp->imp_obd->obd_no_recov) {
1974 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1975 ptlrpc_import_state_name(req->rq_send_state),
1976 ptlrpc_import_state_name(imp->imp_state));
1977 spin_lock(&req->rq_lock);
1978 req->rq_status = -ETIMEDOUT;
1980 spin_unlock(&req->rq_lock);
1984 /* if a request can't be resent we can't wait for an answer after
1986 if (ptlrpc_no_resend(req)) {
1987 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1991 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1997 * Time out all uncompleted requests in request set pointed by \a data
1998 * Callback used when waiting on sets with l_wait_event.
2001 int ptlrpc_expired_set(void *data)
2003 struct ptlrpc_request_set *set = data;
2004 struct list_head *tmp;
2005 time_t now = cfs_time_current_sec();
2008 LASSERT(set != NULL);
2011 * A timeout expired. See which reqs it applies to...
2013 list_for_each(tmp, &set->set_requests) {
2014 struct ptlrpc_request *req =
2015 list_entry(tmp, struct ptlrpc_request,
2018 /* don't expire request waiting for context */
2019 if (req->rq_wait_ctx)
2022 /* Request in-flight? */
2023 if (!((req->rq_phase == RQ_PHASE_RPC &&
2024 !req->rq_waiting && !req->rq_resend) ||
2025 (req->rq_phase == RQ_PHASE_BULK)))
2028 if (req->rq_timedout || /* already dealt with */
2029 req->rq_deadline > now) /* not expired */
2032 /* Deal with this guy. Do it asynchronously to not block
2033 * ptlrpcd thread. */
2034 ptlrpc_expire_one_request(req, 1);
2038 * When waiting for a whole set, we always break out of the
2039 * sleep so we can recalculate the timeout, or enable interrupts
2040 * if everyone's timed out.
2044 EXPORT_SYMBOL(ptlrpc_expired_set);
2047 * Sets rq_intr flag in \a req under spinlock.
2049 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2051 spin_lock(&req->rq_lock);
2053 spin_unlock(&req->rq_lock);
2055 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2058 * Interrupts (sets interrupted flag) all uncompleted requests in
2059 * a set \a data. Callback for l_wait_event for interruptible waits.
2061 void ptlrpc_interrupted_set(void *data)
2063 struct ptlrpc_request_set *set = data;
2064 struct list_head *tmp;
2066 LASSERT(set != NULL);
2067 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2069 list_for_each(tmp, &set->set_requests) {
2070 struct ptlrpc_request *req =
2071 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2073 if (req->rq_phase != RQ_PHASE_RPC &&
2074 req->rq_phase != RQ_PHASE_UNREGISTERING)
2077 ptlrpc_mark_interrupted(req);
2080 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2083 * Get the smallest timeout in the set; this does NOT set a timeout.
2085 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2087 struct list_head *tmp;
2088 time_t now = cfs_time_current_sec();
2090 struct ptlrpc_request *req;
2094 list_for_each(tmp, &set->set_requests) {
2095 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2098 * Request in-flight?
2100 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2101 (req->rq_phase == RQ_PHASE_BULK) ||
2102 (req->rq_phase == RQ_PHASE_NEW)))
2106 * Already timed out.
2108 if (req->rq_timedout)
2114 if (req->rq_wait_ctx)
2117 if (req->rq_phase == RQ_PHASE_NEW)
2118 deadline = req->rq_sent;
2119 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2120 deadline = req->rq_sent;
2122 deadline = req->rq_sent + req->rq_timeout;
2124 if (deadline <= now) /* actually expired already */
2125 timeout = 1; /* ASAP */
2126 else if (timeout == 0 || timeout > deadline - now)
2127 timeout = deadline - now;
2131 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2134 * Send all unset request from the set and then wait untill all
2135 * requests in the set complete (either get a reply, timeout, get an
2136 * error or otherwise be interrupted).
2137 * Returns 0 on success or error code otherwise.
2139 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2141 struct list_head *tmp;
2142 struct ptlrpc_request *req;
2143 struct l_wait_info lwi;
2147 if (set->set_producer)
2148 (void)ptlrpc_set_producer(set);
2150 list_for_each(tmp, &set->set_requests) {
2151 req = list_entry(tmp, struct ptlrpc_request,
2153 if (req->rq_phase == RQ_PHASE_NEW)
2154 (void)ptlrpc_send_new_req(req);
2157 if (list_empty(&set->set_requests))
2161 timeout = ptlrpc_set_next_timeout(set);
2163 /* wait until all complete, interrupted, or an in-flight
2165 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2168 if (timeout == 0 && !cfs_signal_pending())
2170 * No requests are in-flight (ether timed out
2171 * or delayed), so we can allow interrupts.
2172 * We still want to block for a limited time,
2173 * so we allow interrupts during the timeout.
2175 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2177 ptlrpc_interrupted_set, set);
2180 * At least one request is in flight, so no
2181 * interrupts are allowed. Wait until all
2182 * complete, or an in-flight req times out.
2184 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2185 ptlrpc_expired_set, set);
2187 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2189 /* LU-769 - if we ignored the signal because it was already
2190 * pending when we started, we need to handle it now or we risk
2191 * it being ignored forever */
2192 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2193 cfs_signal_pending()) {
2194 sigset_t blocked_sigs =
2195 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2197 /* In fact we only interrupt for the "fatal" signals
2198 * like SIGINT or SIGKILL. We still ignore less
2199 * important signals since ptlrpc set is not easily
2200 * reentrant from userspace again */
2201 if (cfs_signal_pending())
2202 ptlrpc_interrupted_set(set);
2203 cfs_restore_sigs(blocked_sigs);
2206 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2208 /* -EINTR => all requests have been flagged rq_intr so next
2210 * -ETIMEDOUT => someone timed out. When all reqs have
2211 * timed out, signals are enabled allowing completion with
2213 * I don't really care if we go once more round the loop in
2214 * the error cases -eeb. */
2215 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2216 list_for_each(tmp, &set->set_requests) {
2217 req = list_entry(tmp, struct ptlrpc_request,
2219 spin_lock(&req->rq_lock);
2220 req->rq_invalid_rqset = 1;
2221 spin_unlock(&req->rq_lock);
2224 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2226 LASSERT(atomic_read(&set->set_remaining) == 0);
2228 rc = set->set_rc; /* rq_status of already freed requests if any */
2229 list_for_each(tmp, &set->set_requests) {
2230 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2232 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2233 if (req->rq_status != 0)
2234 rc = req->rq_status;
2237 if (set->set_interpret != NULL) {
2238 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2240 rc = interpreter (set, set->set_arg, rc);
2242 struct ptlrpc_set_cbdata *cbdata, *n;
2245 list_for_each_entry_safe(cbdata, n,
2246 &set->set_cblist, psc_item) {
2247 list_del_init(&cbdata->psc_item);
2248 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2251 OBD_FREE_PTR(cbdata);
2257 EXPORT_SYMBOL(ptlrpc_set_wait);
2260 * Helper fuction for request freeing.
2261 * Called when request count reached zero and request needs to be freed.
2262 * Removes request from all sorts of sending/replay lists it might be on,
2263 * frees network buffers if any are present.
2264 * If \a locked is set, that means caller is already holding import imp_lock
2265 * and so we no longer need to reobtain it (for certain lists manipulations)
2267 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2270 if (request == NULL) {
2275 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2276 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2277 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2278 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2279 LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request);
2280 LASSERTF(!request->rq_replay, "req %p\n", request);
2282 req_capsule_fini(&request->rq_pill);
2284 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2285 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2286 if (request->rq_import != NULL) {
2288 spin_lock(&request->rq_import->imp_lock);
2289 list_del_init(&request->rq_replay_list);
2291 spin_unlock(&request->rq_import->imp_lock);
2293 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2295 if (atomic_read(&request->rq_refcount) != 0) {
2296 DEBUG_REQ(D_ERROR, request,
2297 "freeing request with nonzero refcount");
2301 if (request->rq_repbuf != NULL)
2302 sptlrpc_cli_free_repbuf(request);
2303 if (request->rq_export != NULL) {
2304 class_export_put(request->rq_export);
2305 request->rq_export = NULL;
2307 if (request->rq_import != NULL) {
2308 class_import_put(request->rq_import);
2309 request->rq_import = NULL;
2311 if (request->rq_bulk != NULL)
2312 ptlrpc_free_bulk_pin(request->rq_bulk);
2314 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2315 sptlrpc_cli_free_reqbuf(request);
2317 if (request->rq_cli_ctx)
2318 sptlrpc_req_put_ctx(request, !locked);
2320 if (request->rq_pool)
2321 __ptlrpc_free_req_to_pool(request);
2323 ptlrpc_request_cache_free(request);
2327 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2329 * Drop one request reference. Must be called with import imp_lock held.
2330 * When reference count drops to zero, reuqest is freed.
2332 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2334 assert_spin_locked(&request->rq_import->imp_lock);
2335 (void)__ptlrpc_req_finished(request, 1);
2337 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2341 * Drops one reference count for request \a request.
2342 * \a locked set indicates that caller holds import imp_lock.
2343 * Frees the request whe reference count reaches zero.
2345 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2348 if (request == NULL)
2351 if (request == LP_POISON ||
2352 request->rq_reqmsg == LP_POISON) {
2353 CERROR("dereferencing freed request (bug 575)\n");
2358 DEBUG_REQ(D_INFO, request, "refcount now %u",
2359 atomic_read(&request->rq_refcount) - 1);
2361 if (atomic_dec_and_test(&request->rq_refcount)) {
2362 __ptlrpc_free_req(request, locked);
2370 * Drops one reference count for a request.
2372 void ptlrpc_req_finished(struct ptlrpc_request *request)
2374 __ptlrpc_req_finished(request, 0);
2376 EXPORT_SYMBOL(ptlrpc_req_finished);
2379 * Returns xid of a \a request
2381 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2383 return request->rq_xid;
2385 EXPORT_SYMBOL(ptlrpc_req_xid);
2388 * Disengage the client's reply buffer from the network
2389 * NB does _NOT_ unregister any client-side bulk.
2390 * IDEMPOTENT, but _not_ safe against concurrent callers.
2391 * The request owner (i.e. the thread doing the I/O) must call...
2392 * Returns 0 on success or 1 if unregistering cannot be made.
2394 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2397 struct l_wait_info lwi;
2402 LASSERT(!in_interrupt());
2405 * Let's setup deadline for reply unlink.
2407 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2408 async && request->rq_reply_deadline == 0)
2409 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2412 * Nothing left to do.
2414 if (!ptlrpc_client_recv_or_unlink(request))
2417 LNetMDUnlink(request->rq_reply_md_h);
2420 * Let's check it once again.
2422 if (!ptlrpc_client_recv_or_unlink(request))
2426 * Move to "Unregistering" phase as reply was not unlinked yet.
2428 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2431 * Do not wait for unlink to finish.
2437 * We have to l_wait_event() whatever the result, to give liblustre
2438 * a chance to run reply_in_callback(), and to make sure we've
2439 * unlinked before returning a req to the pool.
2442 /* The wq argument is ignored by user-space wait_event macros */
2443 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2444 &request->rq_set->set_waitq :
2445 &request->rq_reply_waitq;
2446 /* Network access will complete in finite time but the HUGE
2447 * timeout lets us CWARN for visibility of sluggish NALs */
2448 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2449 cfs_time_seconds(1), NULL, NULL);
2450 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2453 ptlrpc_rqphase_move(request, request->rq_next_phase);
2457 LASSERT(rc == -ETIMEDOUT);
2458 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2459 "rvcng=%d unlnk=%d/%d", request->rq_receiving_reply,
2460 request->rq_req_unlink, request->rq_reply_unlink);
2464 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2466 static void ptlrpc_free_request(struct ptlrpc_request *req)
2468 spin_lock(&req->rq_lock);
2470 spin_unlock(&req->rq_lock);
2472 if (req->rq_commit_cb != NULL)
2473 req->rq_commit_cb(req);
2474 list_del_init(&req->rq_replay_list);
2476 __ptlrpc_req_finished(req, 1);
2480 * the request is committed and dropped from the replay list of its import
2482 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2484 struct obd_import *imp = req->rq_import;
2486 spin_lock(&imp->imp_lock);
2487 if (list_empty(&req->rq_replay_list)) {
2488 spin_unlock(&imp->imp_lock);
2492 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2493 ptlrpc_free_request(req);
2495 spin_unlock(&imp->imp_lock);
2497 EXPORT_SYMBOL(ptlrpc_request_committed);
2500 * Iterates through replay_list on import and prunes
2501 * all requests have transno smaller than last_committed for the
2502 * import and don't have rq_replay set.
2503 * Since requests are sorted in transno order, stops when meetign first
2504 * transno bigger than last_committed.
2505 * caller must hold imp->imp_lock
2507 void ptlrpc_free_committed(struct obd_import *imp)
2509 struct ptlrpc_request *req, *saved;
2510 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2511 bool skip_committed_list = true;
2514 LASSERT(imp != NULL);
2515 assert_spin_locked(&imp->imp_lock);
2517 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2518 imp->imp_generation == imp->imp_last_generation_checked) {
2519 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2520 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2523 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2524 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2525 imp->imp_generation);
2527 if (imp->imp_generation != imp->imp_last_generation_checked)
2528 skip_committed_list = false;
2530 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2531 imp->imp_last_generation_checked = imp->imp_generation;
2533 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2535 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2536 LASSERT(req != last_req);
2539 if (req->rq_transno == 0) {
2540 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2543 if (req->rq_import_generation < imp->imp_generation) {
2544 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2548 /* not yet committed */
2549 if (req->rq_transno > imp->imp_peer_committed_transno) {
2550 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2554 if (req->rq_replay) {
2555 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2556 list_move_tail(&req->rq_replay_list,
2557 &imp->imp_committed_list);
2561 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2562 imp->imp_peer_committed_transno);
2564 ptlrpc_free_request(req);
2567 if (skip_committed_list)
2570 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2572 LASSERT(req->rq_transno != 0);
2573 if (req->rq_import_generation < imp->imp_generation) {
2574 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2575 ptlrpc_free_request(req);
2582 void ptlrpc_cleanup_client(struct obd_import *imp)
2587 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2590 * Schedule previously sent request for resend.
2591 * For bulk requests we assign new xid (to avoid problems with
2592 * lost replies and therefore several transfers landing into same buffer
2593 * from different sending attempts).
2595 void ptlrpc_resend_req(struct ptlrpc_request *req)
2597 DEBUG_REQ(D_HA, req, "going to resend");
2598 spin_lock(&req->rq_lock);
2600 /* Request got reply but linked to the import list still.
2601 Let ptlrpc_check_set() to process it. */
2602 if (ptlrpc_client_replied(req)) {
2603 spin_unlock(&req->rq_lock);
2604 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2608 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2609 req->rq_status = -EAGAIN;
2612 req->rq_net_err = 0;
2613 req->rq_timedout = 0;
2615 __u64 old_xid = req->rq_xid;
2617 /* ensure previous bulk fails */
2618 req->rq_xid = ptlrpc_next_xid();
2619 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2620 old_xid, req->rq_xid);
2622 ptlrpc_client_wake_req(req);
2623 spin_unlock(&req->rq_lock);
2625 EXPORT_SYMBOL(ptlrpc_resend_req);
2627 /* XXX: this function and rq_status are currently unused */
2628 void ptlrpc_restart_req(struct ptlrpc_request *req)
2630 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2631 req->rq_status = -ERESTARTSYS;
2633 spin_lock(&req->rq_lock);
2634 req->rq_restart = 1;
2635 req->rq_timedout = 0;
2636 ptlrpc_client_wake_req(req);
2637 spin_unlock(&req->rq_lock);
2639 EXPORT_SYMBOL(ptlrpc_restart_req);
2642 * Grab additional reference on a request \a req
2644 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2647 atomic_inc(&req->rq_refcount);
2650 EXPORT_SYMBOL(ptlrpc_request_addref);
2653 * Add a request to import replay_list.
2654 * Must be called under imp_lock
2656 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2657 struct obd_import *imp)
2659 struct list_head *tmp;
2661 assert_spin_locked(&imp->imp_lock);
2663 if (req->rq_transno == 0) {
2664 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2668 /* clear this for new requests that were resent as well
2669 as resent replayed requests. */
2670 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2672 /* don't re-add requests that have been replayed */
2673 if (!list_empty(&req->rq_replay_list))
2676 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2678 LASSERT(imp->imp_replayable);
2679 /* Balanced in ptlrpc_free_committed, usually. */
2680 ptlrpc_request_addref(req);
2681 list_for_each_prev(tmp, &imp->imp_replay_list) {
2682 struct ptlrpc_request *iter = list_entry(tmp,
2683 struct ptlrpc_request,
2686 /* We may have duplicate transnos if we create and then
2687 * open a file, or for closes retained if to match creating
2688 * opens, so use req->rq_xid as a secondary key.
2689 * (See bugs 684, 685, and 428.)
2690 * XXX no longer needed, but all opens need transnos!
2692 if (iter->rq_transno > req->rq_transno)
2695 if (iter->rq_transno == req->rq_transno) {
2696 LASSERT(iter->rq_xid != req->rq_xid);
2697 if (iter->rq_xid > req->rq_xid)
2701 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2705 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2707 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2710 * Send request and wait until it completes.
2711 * Returns request processing status.
2713 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2715 struct ptlrpc_request_set *set;
2719 LASSERT(req->rq_set == NULL);
2720 LASSERT(!req->rq_receiving_reply);
2722 set = ptlrpc_prep_set();
2724 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2728 /* for distributed debugging */
2729 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2731 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2732 ptlrpc_request_addref(req);
2733 ptlrpc_set_add_req(set, req);
2734 rc = ptlrpc_set_wait(set);
2735 ptlrpc_set_destroy(set);
2739 EXPORT_SYMBOL(ptlrpc_queue_wait);
2741 struct ptlrpc_replay_async_args {
2743 int praa_old_status;
2747 * Callback used for replayed requests reply processing.
2748 * In case of succesful reply calls registeresd request replay callback.
2749 * In case of error restart replay process.
2751 static int ptlrpc_replay_interpret(const struct lu_env *env,
2752 struct ptlrpc_request *req,
2753 void * data, int rc)
2755 struct ptlrpc_replay_async_args *aa = data;
2756 struct obd_import *imp = req->rq_import;
2759 atomic_dec(&imp->imp_replay_inflight);
2761 if (!ptlrpc_client_replied(req)) {
2762 CERROR("request replay timed out, restarting recovery\n");
2763 GOTO(out, rc = -ETIMEDOUT);
2766 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2767 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2768 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2769 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2771 /** VBR: check version failure */
2772 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2773 /** replay was failed due to version mismatch */
2774 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2775 spin_lock(&imp->imp_lock);
2776 imp->imp_vbr_failed = 1;
2777 imp->imp_no_lock_replay = 1;
2778 spin_unlock(&imp->imp_lock);
2779 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2781 /** The transno had better not change over replay. */
2782 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2783 lustre_msg_get_transno(req->rq_repmsg) ||
2784 lustre_msg_get_transno(req->rq_repmsg) == 0,
2786 lustre_msg_get_transno(req->rq_reqmsg),
2787 lustre_msg_get_transno(req->rq_repmsg));
2790 spin_lock(&imp->imp_lock);
2791 /** if replays by version then gap occur on server, no trust to locks */
2792 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2793 imp->imp_no_lock_replay = 1;
2794 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2795 spin_unlock(&imp->imp_lock);
2796 LASSERT(imp->imp_last_replay_transno);
2798 /* transaction number shouldn't be bigger than the latest replayed */
2799 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2800 DEBUG_REQ(D_ERROR, req,
2801 "Reported transno "LPU64" is bigger than the "
2802 "replayed one: "LPU64, req->rq_transno,
2803 lustre_msg_get_transno(req->rq_reqmsg));
2804 GOTO(out, rc = -EINVAL);
2807 DEBUG_REQ(D_HA, req, "got rep");
2809 /* let the callback do fixups, possibly including in the request */
2810 if (req->rq_replay_cb)
2811 req->rq_replay_cb(req);
2813 if (ptlrpc_client_replied(req) &&
2814 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2815 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2816 lustre_msg_get_status(req->rq_repmsg),
2817 aa->praa_old_status);
2819 /* Put it back for re-replay. */
2820 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2824 * Errors while replay can set transno to 0, but
2825 * imp_last_replay_transno shouldn't be set to 0 anyway
2827 if (req->rq_transno == 0)
2828 CERROR("Transno is 0 during replay!\n");
2830 /* continue with recovery */
2831 rc = ptlrpc_import_recovery_state_machine(imp);
2833 req->rq_send_state = aa->praa_old_state;
2836 /* this replay failed, so restart recovery */
2837 ptlrpc_connect_import(imp);
2843 * Prepares and queues request for replay.
2844 * Adds it to ptlrpcd queue for actual sending.
2845 * Returns 0 on success.
2847 int ptlrpc_replay_req(struct ptlrpc_request *req)
2849 struct ptlrpc_replay_async_args *aa;
2852 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2854 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2855 aa = ptlrpc_req_async_args(req);
2856 memset(aa, 0, sizeof *aa);
2858 /* Prepare request to be resent with ptlrpcd */
2859 aa->praa_old_state = req->rq_send_state;
2860 req->rq_send_state = LUSTRE_IMP_REPLAY;
2861 req->rq_phase = RQ_PHASE_NEW;
2862 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2864 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2866 req->rq_interpret_reply = ptlrpc_replay_interpret;
2867 /* Readjust the timeout for current conditions */
2868 ptlrpc_at_set_req_timeout(req);
2870 /* Tell server the net_latency, so the server can calculate how long
2871 * it should wait for next replay */
2872 lustre_msg_set_service_time(req->rq_reqmsg,
2873 ptlrpc_at_get_net_latency(req));
2874 DEBUG_REQ(D_HA, req, "REPLAY");
2876 atomic_inc(&req->rq_import->imp_replay_inflight);
2877 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2879 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2882 EXPORT_SYMBOL(ptlrpc_replay_req);
2885 * Aborts all in-flight request on import \a imp sending and delayed lists
2887 void ptlrpc_abort_inflight(struct obd_import *imp)
2889 struct list_head *tmp, *n;
2892 /* Make sure that no new requests get processed for this import.
2893 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2894 * this flag and then putting requests on sending_list or delayed_list.
2896 spin_lock(&imp->imp_lock);
2898 /* XXX locking? Maybe we should remove each request with the list
2899 * locked? Also, how do we know if the requests on the list are
2900 * being freed at this time?
2902 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2903 struct ptlrpc_request *req = list_entry(tmp,
2904 struct ptlrpc_request,
2907 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2909 spin_lock(&req->rq_lock);
2910 if (req->rq_import_generation < imp->imp_generation) {
2912 req->rq_status = -EIO;
2913 ptlrpc_client_wake_req(req);
2915 spin_unlock(&req->rq_lock);
2918 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2919 struct ptlrpc_request *req =
2920 list_entry(tmp, struct ptlrpc_request, rq_list);
2922 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2924 spin_lock(&req->rq_lock);
2925 if (req->rq_import_generation < imp->imp_generation) {
2927 req->rq_status = -EIO;
2928 ptlrpc_client_wake_req(req);
2930 spin_unlock(&req->rq_lock);
2933 /* Last chance to free reqs left on the replay list, but we
2934 * will still leak reqs that haven't committed. */
2935 if (imp->imp_replayable)
2936 ptlrpc_free_committed(imp);
2938 spin_unlock(&imp->imp_lock);
2942 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2945 * Abort all uncompleted requests in request set \a set
2947 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2949 struct list_head *tmp, *pos;
2951 LASSERT(set != NULL);
2953 list_for_each_safe(pos, tmp, &set->set_requests) {
2954 struct ptlrpc_request *req =
2955 list_entry(pos, struct ptlrpc_request,
2958 spin_lock(&req->rq_lock);
2959 if (req->rq_phase != RQ_PHASE_RPC) {
2960 spin_unlock(&req->rq_lock);
2965 req->rq_status = -EINTR;
2966 ptlrpc_client_wake_req(req);
2967 spin_unlock(&req->rq_lock);
2971 static __u64 ptlrpc_last_xid;
2972 static spinlock_t ptlrpc_last_xid_lock;
2975 * Initialize the XID for the node. This is common among all requests on
2976 * this node, and only requires the property that it is monotonically
2977 * increasing. It does not need to be sequential. Since this is also used
2978 * as the RDMA match bits, it is important that a single client NOT have
2979 * the same match bits for two different in-flight requests, hence we do
2980 * NOT want to have an XID per target or similar.
2982 * To avoid an unlikely collision between match bits after a client reboot
2983 * (which would deliver old data into the wrong RDMA buffer) initialize
2984 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2985 * If the time is clearly incorrect, we instead use a 62-bit random number.
2986 * In the worst case the random number will overflow 1M RPCs per second in
2987 * 9133 years, or permutations thereof.
2989 #define YEAR_2004 (1ULL << 30)
2990 void ptlrpc_init_xid(void)
2992 time_t now = cfs_time_current_sec();
2994 spin_lock_init(&ptlrpc_last_xid_lock);
2995 if (now < YEAR_2004) {
2996 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2997 ptlrpc_last_xid >>= 2;
2998 ptlrpc_last_xid |= (1ULL << 61);
3000 ptlrpc_last_xid = (__u64)now << 20;
3003 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
3004 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
3005 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
3009 * Increase xid and returns resulting new value to the caller.
3011 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
3012 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
3013 * itself uses the last bulk xid needed, so the server can determine the
3014 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3015 * xid must align to a power-of-two value.
3017 * This is assumed to be true due to the initial ptlrpc_last_xid
3018 * value also being initialized to a power-of-two value. LU-1431
3020 __u64 ptlrpc_next_xid(void)
3024 spin_lock(&ptlrpc_last_xid_lock);
3025 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3026 ptlrpc_last_xid = next;
3027 spin_unlock(&ptlrpc_last_xid_lock);
3031 EXPORT_SYMBOL(ptlrpc_next_xid);
3034 * Get a glimpse at what next xid value might have been.
3035 * Returns possible next xid.
3037 __u64 ptlrpc_sample_next_xid(void)
3039 #if BITS_PER_LONG == 32
3040 /* need to avoid possible word tearing on 32-bit systems */
3043 spin_lock(&ptlrpc_last_xid_lock);
3044 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3045 spin_unlock(&ptlrpc_last_xid_lock);
3049 /* No need to lock, since returned value is racy anyways */
3050 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3053 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3056 * Functions for operating ptlrpc workers.
3058 * A ptlrpc work is a function which will be running inside ptlrpc context.
3059 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3061 * 1. after a work is created, it can be used many times, that is:
3062 * handler = ptlrpcd_alloc_work();
3063 * ptlrpcd_queue_work();
3065 * queue it again when necessary:
3066 * ptlrpcd_queue_work();
3067 * ptlrpcd_destroy_work();
3068 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3069 * it will only be queued once in any time. Also as its name implies, it may
3070 * have delay before it really runs by ptlrpcd thread.
3072 struct ptlrpc_work_async_args {
3073 int (*cb)(const struct lu_env *, void *);
3077 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3079 /* re-initialize the req */
3080 req->rq_timeout = obd_timeout;
3081 req->rq_sent = cfs_time_current_sec();
3082 req->rq_deadline = req->rq_sent + req->rq_timeout;
3083 req->rq_reply_deadline = req->rq_deadline;
3084 req->rq_phase = RQ_PHASE_INTERPRET;
3085 req->rq_next_phase = RQ_PHASE_COMPLETE;
3086 req->rq_xid = ptlrpc_next_xid();
3087 req->rq_import_generation = req->rq_import->imp_generation;
3089 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3092 static int work_interpreter(const struct lu_env *env,
3093 struct ptlrpc_request *req, void *data, int rc)
3095 struct ptlrpc_work_async_args *arg = data;
3097 LASSERT(ptlrpcd_check_work(req));
3098 LASSERT(arg->cb != NULL);
3100 rc = arg->cb(env, arg->cbdata);
3102 list_del_init(&req->rq_set_chain);
3105 if (atomic_dec_return(&req->rq_refcount) > 1) {
3106 atomic_set(&req->rq_refcount, 2);
3107 ptlrpcd_add_work_req(req);
3112 static int worker_format;
3114 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3116 return req->rq_pill.rc_fmt == (void *)&worker_format;
3120 * Create a work for ptlrpc.
3122 void *ptlrpcd_alloc_work(struct obd_import *imp,
3123 int (*cb)(const struct lu_env *, void *), void *cbdata)
3125 struct ptlrpc_request *req = NULL;
3126 struct ptlrpc_work_async_args *args;
3132 RETURN(ERR_PTR(-EINVAL));
3134 /* copy some code from deprecated fakereq. */
3135 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3137 CERROR("ptlrpc: run out of memory!\n");
3138 RETURN(ERR_PTR(-ENOMEM));
3141 req->rq_send_state = LUSTRE_IMP_FULL;
3142 req->rq_type = PTL_RPC_MSG_REQUEST;
3143 req->rq_import = class_import_get(imp);
3144 req->rq_export = NULL;
3145 req->rq_interpret_reply = work_interpreter;
3146 /* don't want reply */
3147 req->rq_receiving_reply = 0;
3148 req->rq_req_unlink = req->rq_reply_unlink = 0;
3149 req->rq_no_delay = req->rq_no_resend = 1;
3150 req->rq_pill.rc_fmt = (void *)&worker_format;
3152 spin_lock_init(&req->rq_lock);
3153 INIT_LIST_HEAD(&req->rq_list);
3154 INIT_LIST_HEAD(&req->rq_replay_list);
3155 INIT_LIST_HEAD(&req->rq_set_chain);
3156 INIT_LIST_HEAD(&req->rq_history_list);
3157 INIT_LIST_HEAD(&req->rq_exp_list);
3158 init_waitqueue_head(&req->rq_reply_waitq);
3159 init_waitqueue_head(&req->rq_set_waitq);
3160 atomic_set(&req->rq_refcount, 1);
3162 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3163 args = ptlrpc_req_async_args(req);
3165 args->cbdata = cbdata;
3169 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3171 void ptlrpcd_destroy_work(void *handler)
3173 struct ptlrpc_request *req = handler;
3176 ptlrpc_req_finished(req);
3178 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3180 int ptlrpcd_queue_work(void *handler)
3182 struct ptlrpc_request *req = handler;
3185 * Check if the req is already being queued.
3187 * Here comes a trick: it lacks a way of checking if a req is being
3188 * processed reliably in ptlrpc. Here I have to use refcount of req
3189 * for this purpose. This is okay because the caller should use this
3190 * req as opaque data. - Jinshan
3192 LASSERT(atomic_read(&req->rq_refcount) > 0);
3193 if (atomic_inc_return(&req->rq_refcount) == 2)
3194 ptlrpcd_add_work_req(req);
3197 EXPORT_SYMBOL(ptlrpcd_queue_work);