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 static 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);
292 at = &req->rq_import->imp_at;
294 /* Network latency is total time less server processing time */
295 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
296 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
297 CWARN("Reported service time %u > total measured time "
298 CFS_DURATION_T"\n", service_time,
299 cfs_time_sub(now, req->rq_sent));
301 oldnl = at_measured(&at->iat_net_latency, nl);
303 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
304 "has changed from %d to %d\n",
305 req->rq_import->imp_obd->obd_name,
307 &req->rq_import->imp_connection->c_remote_uuid),
308 oldnl, at_get(&at->iat_net_latency));
311 static int unpack_reply(struct ptlrpc_request *req)
315 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
316 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
318 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
323 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
325 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
332 * Handle an early reply message, called with the rq_lock held.
333 * If anything goes wrong just ignore it - same as if it never happened
335 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
336 __must_hold(&req->rq_lock)
338 struct ptlrpc_request *early_req;
344 spin_unlock(&req->rq_lock);
346 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
348 spin_lock(&req->rq_lock);
352 rc = unpack_reply(early_req);
354 /* Expecting to increase the service time estimate here */
355 ptlrpc_at_adj_service(req,
356 lustre_msg_get_timeout(early_req->rq_repmsg));
357 ptlrpc_at_adj_net_latency(req,
358 lustre_msg_get_service_time(early_req->rq_repmsg));
361 sptlrpc_cli_finish_early_reply(early_req);
364 spin_lock(&req->rq_lock);
368 /* Adjust the local timeout for this req */
369 ptlrpc_at_set_req_timeout(req);
371 spin_lock(&req->rq_lock);
372 olddl = req->rq_deadline;
373 /* server assumes it now has rq_timeout from when the request
374 * arrived, so the client should give it at least that long.
375 * since we don't know the arrival time we'll use the original
377 req->rq_deadline = req->rq_sent + req->rq_timeout +
378 ptlrpc_at_get_net_latency(req);
380 DEBUG_REQ(D_ADAPTTO, req,
381 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
382 "("CFS_DURATION_T"s)", req->rq_early_count,
383 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
384 cfs_time_sub(req->rq_deadline, olddl));
389 struct kmem_cache *request_cache;
391 int ptlrpc_request_cache_init(void)
393 request_cache = kmem_cache_create("ptlrpc_cache",
394 sizeof(struct ptlrpc_request),
395 0, SLAB_HWCACHE_ALIGN, NULL);
396 return request_cache == NULL ? -ENOMEM : 0;
399 void ptlrpc_request_cache_fini(void)
401 kmem_cache_destroy(request_cache);
404 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
406 struct ptlrpc_request *req;
408 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
412 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
414 OBD_SLAB_FREE_PTR(req, request_cache);
418 * Wind down request pool \a pool.
419 * Frees all requests from the pool too
421 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
423 struct list_head *l, *tmp;
424 struct ptlrpc_request *req;
426 LASSERT(pool != NULL);
428 spin_lock(&pool->prp_lock);
429 list_for_each_safe(l, tmp, &pool->prp_req_list) {
430 req = list_entry(l, struct ptlrpc_request, rq_list);
431 list_del(&req->rq_list);
432 LASSERT(req->rq_reqbuf);
433 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
434 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
435 ptlrpc_request_cache_free(req);
437 spin_unlock(&pool->prp_lock);
438 OBD_FREE(pool, sizeof(*pool));
440 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
443 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
445 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
450 while (size < pool->prp_rq_size)
453 LASSERTF(list_empty(&pool->prp_req_list) ||
454 size == pool->prp_rq_size,
455 "Trying to change pool size with nonempty pool "
456 "from %d to %d bytes\n", pool->prp_rq_size, size);
458 spin_lock(&pool->prp_lock);
459 pool->prp_rq_size = size;
460 for (i = 0; i < num_rq; i++) {
461 struct ptlrpc_request *req;
462 struct lustre_msg *msg;
464 spin_unlock(&pool->prp_lock);
465 req = ptlrpc_request_cache_alloc(GFP_NOFS);
468 OBD_ALLOC_LARGE(msg, size);
470 ptlrpc_request_cache_free(req);
473 req->rq_reqbuf = msg;
474 req->rq_reqbuf_len = size;
476 spin_lock(&pool->prp_lock);
477 list_add_tail(&req->rq_list, &pool->prp_req_list);
479 spin_unlock(&pool->prp_lock);
482 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
485 * Create and initialize new request pool with given attributes:
486 * \a num_rq - initial number of requests to create for the pool
487 * \a msgsize - maximum message size possible for requests in thid pool
488 * \a populate_pool - function to be called when more requests need to be added
490 * Returns pointer to newly created pool or NULL on error.
492 struct ptlrpc_request_pool *
493 ptlrpc_init_rq_pool(int num_rq, int msgsize,
494 void (*populate_pool)(struct ptlrpc_request_pool *, int))
496 struct ptlrpc_request_pool *pool;
498 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
502 /* Request next power of two for the allocation, because internally
503 kernel would do exactly this */
505 spin_lock_init(&pool->prp_lock);
506 INIT_LIST_HEAD(&pool->prp_req_list);
507 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
508 pool->prp_populate = populate_pool;
510 populate_pool(pool, num_rq);
512 if (list_empty(&pool->prp_req_list)) {
513 /* have not allocated a single request for the pool */
514 OBD_FREE(pool, sizeof(struct ptlrpc_request_pool));
519 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
522 * Fetches one request from pool \a pool
524 static struct ptlrpc_request *
525 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
527 struct ptlrpc_request *request;
528 struct lustre_msg *reqbuf;
533 spin_lock(&pool->prp_lock);
535 /* See if we have anything in a pool, and bail out if nothing,
536 * in writeout path, where this matters, this is safe to do, because
537 * nothing is lost in this case, and when some in-flight requests
538 * complete, this code will be called again. */
539 if (unlikely(list_empty(&pool->prp_req_list))) {
540 spin_unlock(&pool->prp_lock);
544 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
546 list_del_init(&request->rq_list);
547 spin_unlock(&pool->prp_lock);
549 LASSERT(request->rq_reqbuf);
550 LASSERT(request->rq_pool);
552 reqbuf = request->rq_reqbuf;
553 memset(request, 0, sizeof(*request));
554 request->rq_reqbuf = reqbuf;
555 request->rq_reqbuf_len = pool->prp_rq_size;
556 request->rq_pool = pool;
562 * Returns freed \a request to pool.
564 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
566 struct ptlrpc_request_pool *pool = request->rq_pool;
568 spin_lock(&pool->prp_lock);
569 LASSERT(list_empty(&request->rq_list));
570 LASSERT(!request->rq_receiving_reply);
571 list_add_tail(&request->rq_list, &pool->prp_req_list);
572 spin_unlock(&pool->prp_lock);
575 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
576 __u32 version, int opcode,
577 int count, __u32 *lengths, char **bufs,
578 struct ptlrpc_cli_ctx *ctx)
580 struct obd_import *imp = request->rq_import;
585 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
587 rc = sptlrpc_req_get_ctx(request);
592 sptlrpc_req_set_flavor(request, opcode);
594 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
597 LASSERT(!request->rq_pool);
601 lustre_msg_add_version(request->rq_reqmsg, version);
602 request->rq_send_state = LUSTRE_IMP_FULL;
603 request->rq_type = PTL_RPC_MSG_REQUEST;
604 request->rq_export = NULL;
606 request->rq_req_cbid.cbid_fn = request_out_callback;
607 request->rq_req_cbid.cbid_arg = request;
609 request->rq_reply_cbid.cbid_fn = reply_in_callback;
610 request->rq_reply_cbid.cbid_arg = request;
612 request->rq_reply_deadline = 0;
613 request->rq_phase = RQ_PHASE_NEW;
614 request->rq_next_phase = RQ_PHASE_UNDEFINED;
616 request->rq_request_portal = imp->imp_client->cli_request_portal;
617 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
619 ptlrpc_at_set_req_timeout(request);
621 spin_lock_init(&request->rq_lock);
622 INIT_LIST_HEAD(&request->rq_list);
623 INIT_LIST_HEAD(&request->rq_timed_list);
624 INIT_LIST_HEAD(&request->rq_replay_list);
625 INIT_LIST_HEAD(&request->rq_ctx_chain);
626 INIT_LIST_HEAD(&request->rq_set_chain);
627 INIT_LIST_HEAD(&request->rq_history_list);
628 INIT_LIST_HEAD(&request->rq_exp_list);
629 init_waitqueue_head(&request->rq_reply_waitq);
630 init_waitqueue_head(&request->rq_set_waitq);
631 request->rq_xid = ptlrpc_next_xid();
632 atomic_set(&request->rq_refcount, 1);
634 lustre_msg_set_opc(request->rq_reqmsg, opcode);
638 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
640 class_import_put(imp);
644 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
645 __u32 version, int opcode, char **bufs,
646 struct ptlrpc_cli_ctx *ctx)
650 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
651 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
652 request->rq_pill.rc_area[RCL_CLIENT],
655 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
658 * Pack request buffers for network transfer, performing necessary encryption
659 * steps if necessary.
661 int ptlrpc_request_pack(struct ptlrpc_request *request,
662 __u32 version, int opcode)
665 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
669 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
670 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
671 * have to send old ptlrpc_body to keep interoprability with these
674 * Only three kinds of server->client RPCs so far:
679 * XXX This should be removed whenever we drop the interoprability with
680 * the these old clients.
682 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
683 opcode == LDLM_GL_CALLBACK)
684 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
685 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
689 EXPORT_SYMBOL(ptlrpc_request_pack);
692 * Helper function to allocate new request on import \a imp
693 * and possibly using existing request from pool \a pool if provided.
694 * Returns allocated request structure with import field filled or
698 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
699 struct ptlrpc_request_pool *pool)
701 struct ptlrpc_request *request = NULL;
704 request = ptlrpc_prep_req_from_pool(pool);
707 request = ptlrpc_request_cache_alloc(GFP_NOFS);
710 LASSERTF((unsigned long)imp > 0x1000, "%p\n", imp);
711 LASSERT(imp != LP_POISON);
712 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
714 LASSERT(imp->imp_client != LP_POISON);
716 request->rq_import = class_import_get(imp);
718 CERROR("request allocation out of memory\n");
725 * Helper function for creating a request.
726 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
727 * buffer structures according to capsule template \a format.
728 * Returns allocated request structure pointer or NULL on error.
730 static struct ptlrpc_request *
731 ptlrpc_request_alloc_internal(struct obd_import *imp,
732 struct ptlrpc_request_pool * pool,
733 const struct req_format *format)
735 struct ptlrpc_request *request;
737 request = __ptlrpc_request_alloc(imp, pool);
741 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
742 req_capsule_set(&request->rq_pill, format);
747 * Allocate new request structure for import \a imp and initialize its
748 * buffer structure according to capsule template \a format.
750 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
751 const struct req_format *format)
753 return ptlrpc_request_alloc_internal(imp, NULL, format);
755 EXPORT_SYMBOL(ptlrpc_request_alloc);
758 * Allocate new request structure for import \a imp from pool \a pool and
759 * initialize its buffer structure according to capsule template \a format.
761 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
762 struct ptlrpc_request_pool * pool,
763 const struct req_format *format)
765 return ptlrpc_request_alloc_internal(imp, pool, format);
767 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
770 * For requests not from pool, free memory of the request structure.
771 * For requests obtained from a pool earlier, return request back to pool.
773 void ptlrpc_request_free(struct ptlrpc_request *request)
775 if (request->rq_pool)
776 __ptlrpc_free_req_to_pool(request);
778 ptlrpc_request_cache_free(request);
780 EXPORT_SYMBOL(ptlrpc_request_free);
783 * Allocate new request for operatione \a opcode and immediatelly pack it for
785 * Only used for simple requests like OBD_PING where the only important
786 * part of the request is operation itself.
787 * Returns allocated request or NULL on error.
789 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
790 const struct req_format *format,
791 __u32 version, int opcode)
793 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
797 rc = ptlrpc_request_pack(req, version, opcode);
799 ptlrpc_request_free(req);
805 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
808 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
809 * for operation \a opcode. Request would contain \a count buffers.
810 * Sizes of buffers are described in array \a lengths and buffers themselves
811 * are provided by a pointer \a bufs.
812 * Returns prepared request structure pointer or NULL on error.
814 struct ptlrpc_request *
815 ptlrpc_prep_req_pool(struct obd_import *imp,
816 __u32 version, int opcode,
817 int count, __u32 *lengths, char **bufs,
818 struct ptlrpc_request_pool *pool)
820 struct ptlrpc_request *request;
823 request = __ptlrpc_request_alloc(imp, pool);
827 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
828 lengths, bufs, NULL);
830 ptlrpc_request_free(request);
835 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
838 * Same as ptlrpc_prep_req_pool, but without pool
840 struct ptlrpc_request *
841 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
842 __u32 *lengths, char **bufs)
844 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
847 EXPORT_SYMBOL(ptlrpc_prep_req);
850 * Allocate and initialize new request set structure.
851 * Returns a pointer to the newly allocated set structure or NULL on error.
853 struct ptlrpc_request_set *ptlrpc_prep_set(void)
855 struct ptlrpc_request_set *set;
858 OBD_ALLOC(set, sizeof *set);
861 atomic_set(&set->set_refcount, 1);
862 INIT_LIST_HEAD(&set->set_requests);
863 init_waitqueue_head(&set->set_waitq);
864 atomic_set(&set->set_new_count, 0);
865 atomic_set(&set->set_remaining, 0);
866 spin_lock_init(&set->set_new_req_lock);
867 INIT_LIST_HEAD(&set->set_new_requests);
868 INIT_LIST_HEAD(&set->set_cblist);
869 set->set_max_inflight = UINT_MAX;
870 set->set_producer = NULL;
871 set->set_producer_arg = NULL;
876 EXPORT_SYMBOL(ptlrpc_prep_set);
879 * Allocate and initialize new request set structure with flow control
880 * extension. This extension allows to control the number of requests in-flight
881 * for the whole set. A callback function to generate requests must be provided
882 * and the request set will keep the number of requests sent over the wire to
884 * Returns a pointer to the newly allocated set structure or NULL on error.
886 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
890 struct ptlrpc_request_set *set;
892 set = ptlrpc_prep_set();
896 set->set_max_inflight = max;
897 set->set_producer = func;
898 set->set_producer_arg = arg;
902 EXPORT_SYMBOL(ptlrpc_prep_fcset);
905 * Wind down and free request set structure previously allocated with
907 * Ensures that all requests on the set have completed and removes
908 * all requests from the request list in a set.
909 * If any unsent request happen to be on the list, pretends that they got
910 * an error in flight and calls their completion handler.
912 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
914 struct list_head *tmp;
915 struct list_head *next;
920 /* Requests on the set should either all be completed, or all be new */
921 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
922 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
923 list_for_each(tmp, &set->set_requests) {
924 struct ptlrpc_request *req =
925 list_entry(tmp, struct ptlrpc_request,
928 LASSERT(req->rq_phase == expected_phase);
932 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
933 atomic_read(&set->set_remaining) == n, "%d / %d\n",
934 atomic_read(&set->set_remaining), n);
936 list_for_each_safe(tmp, next, &set->set_requests) {
937 struct ptlrpc_request *req =
938 list_entry(tmp, struct ptlrpc_request,
940 list_del_init(&req->rq_set_chain);
942 LASSERT(req->rq_phase == expected_phase);
944 if (req->rq_phase == RQ_PHASE_NEW) {
945 ptlrpc_req_interpret(NULL, req, -EBADR);
946 atomic_dec(&set->set_remaining);
949 spin_lock(&req->rq_lock);
951 req->rq_invalid_rqset = 0;
952 spin_unlock(&req->rq_lock);
954 ptlrpc_req_finished (req);
957 LASSERT(atomic_read(&set->set_remaining) == 0);
959 ptlrpc_reqset_put(set);
962 EXPORT_SYMBOL(ptlrpc_set_destroy);
965 * Add a callback function \a fn to the set.
966 * This function would be called when all requests on this set are completed.
967 * The function will be passed \a data argument.
969 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
970 set_interpreter_func fn, void *data)
972 struct ptlrpc_set_cbdata *cbdata;
974 OBD_ALLOC_PTR(cbdata);
978 cbdata->psc_interpret = fn;
979 cbdata->psc_data = data;
980 list_add_tail(&cbdata->psc_item, &set->set_cblist);
984 EXPORT_SYMBOL(ptlrpc_set_add_cb);
987 * Add a new request to the general purpose request set.
988 * Assumes request reference from the caller.
990 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
991 struct ptlrpc_request *req)
993 LASSERT(list_empty(&req->rq_set_chain));
995 /* The set takes over the caller's request reference */
996 list_add_tail(&req->rq_set_chain, &set->set_requests);
998 atomic_inc(&set->set_remaining);
999 req->rq_queued_time = cfs_time_current();
1001 if (req->rq_reqmsg != NULL)
1002 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1004 if (set->set_producer != NULL)
1005 /* If the request set has a producer callback, the RPC must be
1006 * sent straight away */
1007 ptlrpc_send_new_req(req);
1009 EXPORT_SYMBOL(ptlrpc_set_add_req);
1012 * Add a request to a request with dedicated server thread
1013 * and wake the thread to make any necessary processing.
1014 * Currently only used for ptlrpcd.
1016 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1017 struct ptlrpc_request *req)
1019 struct ptlrpc_request_set *set = pc->pc_set;
1022 LASSERT(req->rq_set == NULL);
1023 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1025 spin_lock(&set->set_new_req_lock);
1027 * The set takes over the caller's request reference.
1030 req->rq_queued_time = cfs_time_current();
1031 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1032 count = atomic_inc_return(&set->set_new_count);
1033 spin_unlock(&set->set_new_req_lock);
1035 /* Only need to call wakeup once for the first entry. */
1037 wake_up(&set->set_waitq);
1039 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1040 * guarantee the async RPC can be processed ASAP, we have
1041 * no other better choice. It maybe fixed in future. */
1042 for (i = 0; i < pc->pc_npartners; i++)
1043 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1046 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1049 * Based on the current state of the import, determine if the request
1050 * can be sent, is an error, or should be delayed.
1052 * Returns true if this request should be delayed. If false, and
1053 * *status is set, then the request can not be sent and *status is the
1054 * error code. If false and status is 0, then request can be sent.
1056 * The imp->imp_lock must be held.
1058 static int ptlrpc_import_delay_req(struct obd_import *imp,
1059 struct ptlrpc_request *req, int *status)
1064 LASSERT (status != NULL);
1067 if (req->rq_ctx_init || req->rq_ctx_fini) {
1068 /* always allow ctx init/fini rpc go through */
1069 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1070 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1072 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1073 /* pings may safely race with umount */
1074 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1075 D_HA : D_ERROR, req, "IMP_CLOSED ");
1077 } else if (ptlrpc_send_limit_expired(req)) {
1078 /* probably doesn't need to be a D_ERROR after initial testing*/
1079 DEBUG_REQ(D_HA, req, "send limit expired ");
1080 *status = -ETIMEDOUT;
1081 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1082 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1083 /* allow CONNECT even if import is invalid */ ;
1084 if (atomic_read(&imp->imp_inval_count) != 0) {
1085 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1088 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1089 if (!imp->imp_deactive)
1090 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1091 *status = -ESHUTDOWN; /* bz 12940 */
1092 } else if (req->rq_import_generation != imp->imp_generation) {
1093 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1095 } else if (req->rq_send_state != imp->imp_state) {
1096 /* invalidate in progress - any requests should be drop */
1097 if (atomic_read(&imp->imp_inval_count) != 0) {
1098 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1100 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1101 *status = -EWOULDBLOCK;
1102 } else if (req->rq_allow_replay &&
1103 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1104 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1105 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1106 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1107 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1117 * Decide if the eror message regarding provided request \a req
1118 * should be printed to the console or not.
1119 * Makes it's decision on request status and other properties.
1120 * Returns 1 to print error on the system console or 0 if not.
1122 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1127 LASSERT(req->rq_reqmsg != NULL);
1128 opc = lustre_msg_get_opc(req->rq_reqmsg);
1130 /* Suppress particular reconnect errors which are to be expected. No
1131 * errors are suppressed for the initial connection on an import */
1132 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1133 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1135 /* Suppress timed out reconnect requests */
1136 if (req->rq_timedout)
1139 /* Suppress unavailable/again reconnect requests */
1140 err = lustre_msg_get_status(req->rq_repmsg);
1141 if (err == -ENODEV || err == -EAGAIN)
1149 * Check request processing status.
1150 * Returns the status.
1152 static int ptlrpc_check_status(struct ptlrpc_request *req)
1157 err = lustre_msg_get_status(req->rq_repmsg);
1158 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1159 struct obd_import *imp = req->rq_import;
1160 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1161 if (ptlrpc_console_allow(req))
1162 LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s,"
1163 " operation %s failed with %d.\n",
1164 imp->imp_obd->obd_name,
1166 imp->imp_connection->c_peer.nid),
1167 ll_opcode2str(opc), err);
1168 RETURN(err < 0 ? err : -EINVAL);
1172 DEBUG_REQ(D_INFO, req, "status is %d", err);
1173 } else if (err > 0) {
1174 /* XXX: translate this error from net to host */
1175 DEBUG_REQ(D_INFO, req, "status is %d", err);
1182 * save pre-versions of objects into request for replay.
1183 * Versions are obtained from server reply.
1186 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1188 struct lustre_msg *repmsg = req->rq_repmsg;
1189 struct lustre_msg *reqmsg = req->rq_reqmsg;
1190 __u64 *versions = lustre_msg_get_versions(repmsg);
1193 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1197 lustre_msg_set_versions(reqmsg, versions);
1198 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1199 versions[0], versions[1]);
1205 * Callback function called when client receives RPC reply for \a req.
1206 * Returns 0 on success or error code.
1207 * The return alue would be assigned to req->rq_status by the caller
1208 * as request processing status.
1209 * This function also decides if the request needs to be saved for later replay.
1211 static int after_reply(struct ptlrpc_request *req)
1213 struct obd_import *imp = req->rq_import;
1214 struct obd_device *obd = req->rq_import->imp_obd;
1216 struct timeval work_start;
1220 LASSERT(obd != NULL);
1221 /* repbuf must be unlinked */
1222 LASSERT(!req->rq_receiving_reply && !req->rq_reply_unlink);
1224 if (req->rq_reply_truncate) {
1225 if (ptlrpc_no_resend(req)) {
1226 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1227 " expected: %d, actual size: %d",
1228 req->rq_nob_received, req->rq_repbuf_len);
1232 sptlrpc_cli_free_repbuf(req);
1233 /* Pass the required reply buffer size (include
1234 * space for early reply).
1235 * NB: no need to roundup because alloc_repbuf
1236 * will roundup it */
1237 req->rq_replen = req->rq_nob_received;
1238 req->rq_nob_received = 0;
1239 spin_lock(&req->rq_lock);
1241 spin_unlock(&req->rq_lock);
1246 * NB Until this point, the whole of the incoming message,
1247 * including buflens, status etc is in the sender's byte order.
1249 rc = sptlrpc_cli_unwrap_reply(req);
1251 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1256 * Security layer unwrap might ask resend this request.
1261 rc = unpack_reply(req);
1265 /* retry indefinitely on EINPROGRESS */
1266 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1267 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1268 time_t now = cfs_time_current_sec();
1270 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1272 req->rq_nr_resend++;
1274 /* allocate new xid to avoid reply reconstruction */
1275 if (!req->rq_bulk) {
1276 /* new xid is already allocated for bulk in
1277 * ptlrpc_check_set() */
1278 req->rq_xid = ptlrpc_next_xid();
1279 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1280 "resend on EINPROGRESS");
1283 /* Readjust the timeout for current conditions */
1284 ptlrpc_at_set_req_timeout(req);
1285 /* delay resend to give a chance to the server to get ready.
1286 * The delay is increased by 1s on every resend and is capped to
1287 * the current request timeout (i.e. obd_timeout if AT is off,
1288 * or AT service time x 125% + 5s, see at_est2timeout) */
1289 if (req->rq_nr_resend > req->rq_timeout)
1290 req->rq_sent = now + req->rq_timeout;
1292 req->rq_sent = now + req->rq_nr_resend;
1297 do_gettimeofday(&work_start);
1298 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1299 if (obd->obd_svc_stats != NULL) {
1300 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1302 ptlrpc_lprocfs_rpc_sent(req, timediff);
1305 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1306 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1307 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1308 lustre_msg_get_type(req->rq_repmsg));
1312 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1313 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1314 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1315 ptlrpc_at_adj_net_latency(req,
1316 lustre_msg_get_service_time(req->rq_repmsg));
1318 rc = ptlrpc_check_status(req);
1319 imp->imp_connect_error = rc;
1323 * Either we've been evicted, or the server has failed for
1324 * some reason. Try to reconnect, and if that fails, punt to
1327 if (ll_rpc_recoverable_error(rc)) {
1328 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1329 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1332 ptlrpc_request_handle_notconn(req);
1337 * Let's look if server sent slv. Do it only for RPC with
1340 ldlm_cli_update_pool(req);
1344 * Store transno in reqmsg for replay.
1346 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1347 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1348 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1351 if (imp->imp_replayable) {
1352 spin_lock(&imp->imp_lock);
1354 * No point in adding already-committed requests to the replay
1355 * list, we will just remove them immediately. b=9829
1357 if (req->rq_transno != 0 &&
1359 lustre_msg_get_last_committed(req->rq_repmsg) ||
1361 /** version recovery */
1362 ptlrpc_save_versions(req);
1363 ptlrpc_retain_replayable_request(req, imp);
1364 } else if (req->rq_commit_cb != NULL &&
1365 list_empty(&req->rq_replay_list)) {
1366 /* NB: don't call rq_commit_cb if it's already on
1367 * rq_replay_list, ptlrpc_free_committed() will call
1368 * it later, see LU-3618 for details */
1369 spin_unlock(&imp->imp_lock);
1370 req->rq_commit_cb(req);
1371 spin_lock(&imp->imp_lock);
1375 * Replay-enabled imports return commit-status information.
1377 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1378 imp->imp_peer_committed_transno =
1379 lustre_msg_get_last_committed(req->rq_repmsg);
1382 ptlrpc_free_committed(imp);
1384 if (!list_empty(&imp->imp_replay_list)) {
1385 struct ptlrpc_request *last;
1387 last = list_entry(imp->imp_replay_list.prev,
1388 struct ptlrpc_request,
1391 * Requests with rq_replay stay on the list even if no
1392 * commit is expected.
1394 if (last->rq_transno > imp->imp_peer_committed_transno)
1395 ptlrpc_pinger_commit_expected(imp);
1398 spin_unlock(&imp->imp_lock);
1405 * Helper function to send request \a req over the network for the first time
1406 * Also adjusts request phase.
1407 * Returns 0 on success or error code.
1409 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1411 struct obd_import *imp = req->rq_import;
1415 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1416 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1417 (!req->rq_generation_set ||
1418 req->rq_import_generation == imp->imp_generation))
1421 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1423 spin_lock(&imp->imp_lock);
1425 if (!req->rq_generation_set)
1426 req->rq_import_generation = imp->imp_generation;
1428 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1429 spin_lock(&req->rq_lock);
1430 req->rq_waiting = 1;
1431 spin_unlock(&req->rq_lock);
1433 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1434 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1435 ptlrpc_import_state_name(req->rq_send_state),
1436 ptlrpc_import_state_name(imp->imp_state));
1437 LASSERT(list_empty(&req->rq_list));
1438 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1439 atomic_inc(&req->rq_import->imp_inflight);
1440 spin_unlock(&imp->imp_lock);
1445 spin_unlock(&imp->imp_lock);
1446 req->rq_status = rc;
1447 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1451 LASSERT(list_empty(&req->rq_list));
1452 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1453 atomic_inc(&req->rq_import->imp_inflight);
1454 spin_unlock(&imp->imp_lock);
1456 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1458 rc = sptlrpc_req_refresh_ctx(req, -1);
1461 req->rq_status = rc;
1464 spin_lock(&req->rq_lock);
1465 req->rq_wait_ctx = 1;
1466 spin_unlock(&req->rq_lock);
1471 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1472 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1473 imp->imp_obd->obd_uuid.uuid,
1474 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1475 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1476 lustre_msg_get_opc(req->rq_reqmsg));
1478 rc = ptl_send_rpc(req, 0);
1480 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1481 spin_lock(&req->rq_lock);
1482 req->rq_net_err = 1;
1483 spin_unlock(&req->rq_lock);
1489 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1494 LASSERT(set->set_producer != NULL);
1496 remaining = atomic_read(&set->set_remaining);
1498 /* populate the ->set_requests list with requests until we
1499 * reach the maximum number of RPCs in flight for this set */
1500 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1501 rc = set->set_producer(set, set->set_producer_arg);
1502 if (rc == -ENOENT) {
1503 /* no more RPC to produce */
1504 set->set_producer = NULL;
1505 set->set_producer_arg = NULL;
1510 RETURN((atomic_read(&set->set_remaining) - remaining));
1514 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1515 * and no more replies are expected.
1516 * (it is possible to get less replies than requests sent e.g. due to timed out
1517 * requests or requests that we had trouble to send out)
1519 * NOTE: This function contains a potential schedule point (cond_resched()).
1521 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1523 struct list_head *tmp, *next;
1524 struct list_head comp_reqs;
1525 int force_timer_recalc = 0;
1528 if (atomic_read(&set->set_remaining) == 0)
1531 INIT_LIST_HEAD(&comp_reqs);
1532 list_for_each_safe(tmp, next, &set->set_requests) {
1533 struct ptlrpc_request *req =
1534 list_entry(tmp, struct ptlrpc_request,
1536 struct obd_import *imp = req->rq_import;
1537 int unregistered = 0;
1540 /* This schedule point is mainly for the ptlrpcd caller of this
1541 * function. Most ptlrpc sets are not long-lived and unbounded
1542 * in length, but at the least the set used by the ptlrpcd is.
1543 * Since the processing time is unbounded, we need to insert an
1544 * explicit schedule point to make the thread well-behaved.
1548 if (req->rq_phase == RQ_PHASE_NEW &&
1549 ptlrpc_send_new_req(req)) {
1550 force_timer_recalc = 1;
1553 /* delayed send - skip */
1554 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1557 /* delayed resend - skip */
1558 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1559 req->rq_sent > cfs_time_current_sec())
1562 if (!(req->rq_phase == RQ_PHASE_RPC ||
1563 req->rq_phase == RQ_PHASE_BULK ||
1564 req->rq_phase == RQ_PHASE_INTERPRET ||
1565 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1566 req->rq_phase == RQ_PHASE_COMPLETE)) {
1567 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1571 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1572 LASSERT(req->rq_next_phase != req->rq_phase);
1573 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1576 * Skip processing until reply is unlinked. We
1577 * can't return to pool before that and we can't
1578 * call interpret before that. We need to make
1579 * sure that all rdma transfers finished and will
1580 * not corrupt any data.
1582 if (ptlrpc_client_recv_or_unlink(req) ||
1583 ptlrpc_client_bulk_active(req))
1587 * Turn fail_loc off to prevent it from looping
1590 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1591 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1594 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1595 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1600 * Move to next phase if reply was successfully
1603 ptlrpc_rqphase_move(req, req->rq_next_phase);
1606 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1607 list_move_tail(&req->rq_set_chain, &comp_reqs);
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 list_del_init(&req->rq_list);
1678 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 list_del_init(&req->rq_list);
1702 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 (!list_empty(&req->rq_list)) {
1871 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 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);
1897 list_move_tail(&req->rq_set_chain, &comp_reqs);
1901 /* move completed request at the head of list so it's easier for
1902 * caller to find them */
1903 list_splice(&comp_reqs, &set->set_requests);
1905 /* If we hit an error, we want to recover promptly. */
1906 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1908 EXPORT_SYMBOL(ptlrpc_check_set);
1911 * Time out request \a req. is \a async_unlink is set, that means do not wait
1912 * until LNet actually confirms network buffer unlinking.
1913 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1915 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1917 struct obd_import *imp = req->rq_import;
1921 spin_lock(&req->rq_lock);
1922 req->rq_timedout = 1;
1923 spin_unlock(&req->rq_lock);
1925 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1926 "/real "CFS_DURATION_T"]",
1927 req->rq_net_err ? "failed due to network error" :
1928 ((req->rq_real_sent == 0 ||
1929 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1930 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1931 "timed out for sent delay" : "timed out for slow reply"),
1932 req->rq_sent, req->rq_real_sent);
1934 if (imp != NULL && obd_debug_peer_on_timeout)
1935 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1937 ptlrpc_unregister_reply(req, async_unlink);
1938 ptlrpc_unregister_bulk(req, async_unlink);
1940 if (obd_dump_on_timeout)
1941 libcfs_debug_dumplog();
1944 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1948 atomic_inc(&imp->imp_timeouts);
1950 /* The DLM server doesn't want recovery run on its imports. */
1951 if (imp->imp_dlm_fake)
1954 /* If this request is for recovery or other primordial tasks,
1955 * then error it out here. */
1956 if (req->rq_ctx_init || req->rq_ctx_fini ||
1957 req->rq_send_state != LUSTRE_IMP_FULL ||
1958 imp->imp_obd->obd_no_recov) {
1959 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1960 ptlrpc_import_state_name(req->rq_send_state),
1961 ptlrpc_import_state_name(imp->imp_state));
1962 spin_lock(&req->rq_lock);
1963 req->rq_status = -ETIMEDOUT;
1965 spin_unlock(&req->rq_lock);
1969 /* if a request can't be resent we can't wait for an answer after
1971 if (ptlrpc_no_resend(req)) {
1972 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1976 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1982 * Time out all uncompleted requests in request set pointed by \a data
1983 * Callback used when waiting on sets with l_wait_event.
1986 int ptlrpc_expired_set(void *data)
1988 struct ptlrpc_request_set *set = data;
1989 struct list_head *tmp;
1990 time_t now = cfs_time_current_sec();
1993 LASSERT(set != NULL);
1996 * A timeout expired. See which reqs it applies to...
1998 list_for_each(tmp, &set->set_requests) {
1999 struct ptlrpc_request *req =
2000 list_entry(tmp, struct ptlrpc_request,
2003 /* don't expire request waiting for context */
2004 if (req->rq_wait_ctx)
2007 /* Request in-flight? */
2008 if (!((req->rq_phase == RQ_PHASE_RPC &&
2009 !req->rq_waiting && !req->rq_resend) ||
2010 (req->rq_phase == RQ_PHASE_BULK)))
2013 if (req->rq_timedout || /* already dealt with */
2014 req->rq_deadline > now) /* not expired */
2017 /* Deal with this guy. Do it asynchronously to not block
2018 * ptlrpcd thread. */
2019 ptlrpc_expire_one_request(req, 1);
2023 * When waiting for a whole set, we always break out of the
2024 * sleep so we can recalculate the timeout, or enable interrupts
2025 * if everyone's timed out.
2029 EXPORT_SYMBOL(ptlrpc_expired_set);
2032 * Sets rq_intr flag in \a req under spinlock.
2034 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2036 spin_lock(&req->rq_lock);
2038 spin_unlock(&req->rq_lock);
2040 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2043 * Interrupts (sets interrupted flag) all uncompleted requests in
2044 * a set \a data. Callback for l_wait_event for interruptible waits.
2046 void ptlrpc_interrupted_set(void *data)
2048 struct ptlrpc_request_set *set = data;
2049 struct list_head *tmp;
2051 LASSERT(set != NULL);
2052 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2054 list_for_each(tmp, &set->set_requests) {
2055 struct ptlrpc_request *req =
2056 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2058 if (req->rq_phase != RQ_PHASE_RPC &&
2059 req->rq_phase != RQ_PHASE_UNREGISTERING)
2062 ptlrpc_mark_interrupted(req);
2065 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2068 * Get the smallest timeout in the set; this does NOT set a timeout.
2070 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2072 struct list_head *tmp;
2073 time_t now = cfs_time_current_sec();
2075 struct ptlrpc_request *req;
2079 list_for_each(tmp, &set->set_requests) {
2080 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2083 * Request in-flight?
2085 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2086 (req->rq_phase == RQ_PHASE_BULK) ||
2087 (req->rq_phase == RQ_PHASE_NEW)))
2091 * Already timed out.
2093 if (req->rq_timedout)
2099 if (req->rq_wait_ctx)
2102 if (req->rq_phase == RQ_PHASE_NEW)
2103 deadline = req->rq_sent;
2104 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2105 deadline = req->rq_sent;
2107 deadline = req->rq_sent + req->rq_timeout;
2109 if (deadline <= now) /* actually expired already */
2110 timeout = 1; /* ASAP */
2111 else if (timeout == 0 || timeout > deadline - now)
2112 timeout = deadline - now;
2116 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2119 * Send all unset request from the set and then wait untill all
2120 * requests in the set complete (either get a reply, timeout, get an
2121 * error or otherwise be interrupted).
2122 * Returns 0 on success or error code otherwise.
2124 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2126 struct list_head *tmp;
2127 struct ptlrpc_request *req;
2128 struct l_wait_info lwi;
2132 if (set->set_producer)
2133 (void)ptlrpc_set_producer(set);
2135 list_for_each(tmp, &set->set_requests) {
2136 req = list_entry(tmp, struct ptlrpc_request,
2138 if (req->rq_phase == RQ_PHASE_NEW)
2139 (void)ptlrpc_send_new_req(req);
2142 if (list_empty(&set->set_requests))
2146 timeout = ptlrpc_set_next_timeout(set);
2148 /* wait until all complete, interrupted, or an in-flight
2150 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2153 if (timeout == 0 && !cfs_signal_pending())
2155 * No requests are in-flight (ether timed out
2156 * or delayed), so we can allow interrupts.
2157 * We still want to block for a limited time,
2158 * so we allow interrupts during the timeout.
2160 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2162 ptlrpc_interrupted_set, set);
2165 * At least one request is in flight, so no
2166 * interrupts are allowed. Wait until all
2167 * complete, or an in-flight req times out.
2169 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2170 ptlrpc_expired_set, set);
2172 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2174 /* LU-769 - if we ignored the signal because it was already
2175 * pending when we started, we need to handle it now or we risk
2176 * it being ignored forever */
2177 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2178 cfs_signal_pending()) {
2179 sigset_t blocked_sigs =
2180 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2182 /* In fact we only interrupt for the "fatal" signals
2183 * like SIGINT or SIGKILL. We still ignore less
2184 * important signals since ptlrpc set is not easily
2185 * reentrant from userspace again */
2186 if (cfs_signal_pending())
2187 ptlrpc_interrupted_set(set);
2188 cfs_restore_sigs(blocked_sigs);
2191 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2193 /* -EINTR => all requests have been flagged rq_intr so next
2195 * -ETIMEDOUT => someone timed out. When all reqs have
2196 * timed out, signals are enabled allowing completion with
2198 * I don't really care if we go once more round the loop in
2199 * the error cases -eeb. */
2200 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2201 list_for_each(tmp, &set->set_requests) {
2202 req = list_entry(tmp, struct ptlrpc_request,
2204 spin_lock(&req->rq_lock);
2205 req->rq_invalid_rqset = 1;
2206 spin_unlock(&req->rq_lock);
2209 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2211 LASSERT(atomic_read(&set->set_remaining) == 0);
2213 rc = set->set_rc; /* rq_status of already freed requests if any */
2214 list_for_each(tmp, &set->set_requests) {
2215 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2217 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2218 if (req->rq_status != 0)
2219 rc = req->rq_status;
2222 if (set->set_interpret != NULL) {
2223 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2225 rc = interpreter (set, set->set_arg, rc);
2227 struct ptlrpc_set_cbdata *cbdata, *n;
2230 list_for_each_entry_safe(cbdata, n,
2231 &set->set_cblist, psc_item) {
2232 list_del_init(&cbdata->psc_item);
2233 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2236 OBD_FREE_PTR(cbdata);
2242 EXPORT_SYMBOL(ptlrpc_set_wait);
2245 * Helper fuction for request freeing.
2246 * Called when request count reached zero and request needs to be freed.
2247 * Removes request from all sorts of sending/replay lists it might be on,
2248 * frees network buffers if any are present.
2249 * If \a locked is set, that means caller is already holding import imp_lock
2250 * and so we no longer need to reobtain it (for certain lists manipulations)
2252 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2255 if (request == NULL) {
2260 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2261 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2262 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2263 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2264 LASSERTF(list_empty(&request->rq_exp_list), "req %p\n", request);
2265 LASSERTF(!request->rq_replay, "req %p\n", request);
2267 req_capsule_fini(&request->rq_pill);
2269 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2270 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2271 if (request->rq_import != NULL) {
2273 spin_lock(&request->rq_import->imp_lock);
2274 list_del_init(&request->rq_replay_list);
2276 spin_unlock(&request->rq_import->imp_lock);
2278 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2280 if (atomic_read(&request->rq_refcount) != 0) {
2281 DEBUG_REQ(D_ERROR, request,
2282 "freeing request with nonzero refcount");
2286 if (request->rq_repbuf != NULL)
2287 sptlrpc_cli_free_repbuf(request);
2288 if (request->rq_export != NULL) {
2289 class_export_put(request->rq_export);
2290 request->rq_export = NULL;
2292 if (request->rq_import != NULL) {
2293 class_import_put(request->rq_import);
2294 request->rq_import = NULL;
2296 if (request->rq_bulk != NULL)
2297 ptlrpc_free_bulk_pin(request->rq_bulk);
2299 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2300 sptlrpc_cli_free_reqbuf(request);
2302 if (request->rq_cli_ctx)
2303 sptlrpc_req_put_ctx(request, !locked);
2305 if (request->rq_pool)
2306 __ptlrpc_free_req_to_pool(request);
2308 ptlrpc_request_cache_free(request);
2312 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2314 * Drop one request reference. Must be called with import imp_lock held.
2315 * When reference count drops to zero, reuqest is freed.
2317 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2319 assert_spin_locked(&request->rq_import->imp_lock);
2320 (void)__ptlrpc_req_finished(request, 1);
2322 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2326 * Drops one reference count for request \a request.
2327 * \a locked set indicates that caller holds import imp_lock.
2328 * Frees the request whe reference count reaches zero.
2330 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2333 if (request == NULL)
2336 if (request == LP_POISON ||
2337 request->rq_reqmsg == LP_POISON) {
2338 CERROR("dereferencing freed request (bug 575)\n");
2343 DEBUG_REQ(D_INFO, request, "refcount now %u",
2344 atomic_read(&request->rq_refcount) - 1);
2346 if (atomic_dec_and_test(&request->rq_refcount)) {
2347 __ptlrpc_free_req(request, locked);
2355 * Drops one reference count for a request.
2357 void ptlrpc_req_finished(struct ptlrpc_request *request)
2359 __ptlrpc_req_finished(request, 0);
2361 EXPORT_SYMBOL(ptlrpc_req_finished);
2364 * Returns xid of a \a request
2366 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2368 return request->rq_xid;
2370 EXPORT_SYMBOL(ptlrpc_req_xid);
2373 * Disengage the client's reply buffer from the network
2374 * NB does _NOT_ unregister any client-side bulk.
2375 * IDEMPOTENT, but _not_ safe against concurrent callers.
2376 * The request owner (i.e. the thread doing the I/O) must call...
2377 * Returns 0 on success or 1 if unregistering cannot be made.
2379 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2382 struct l_wait_info lwi;
2387 LASSERT(!in_interrupt());
2390 * Let's setup deadline for reply unlink.
2392 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2393 async && request->rq_reply_deadline == 0)
2394 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2397 * Nothing left to do.
2399 if (!ptlrpc_client_recv_or_unlink(request))
2402 LNetMDUnlink(request->rq_reply_md_h);
2405 * Let's check it once again.
2407 if (!ptlrpc_client_recv_or_unlink(request))
2411 * Move to "Unregistering" phase as reply was not unlinked yet.
2413 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2416 * Do not wait for unlink to finish.
2422 * We have to l_wait_event() whatever the result, to give liblustre
2423 * a chance to run reply_in_callback(), and to make sure we've
2424 * unlinked before returning a req to the pool.
2427 /* The wq argument is ignored by user-space wait_event macros */
2428 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2429 &request->rq_set->set_waitq :
2430 &request->rq_reply_waitq;
2431 /* Network access will complete in finite time but the HUGE
2432 * timeout lets us CWARN for visibility of sluggish NALs */
2433 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2434 cfs_time_seconds(1), NULL, NULL);
2435 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2438 ptlrpc_rqphase_move(request, request->rq_next_phase);
2442 LASSERT(rc == -ETIMEDOUT);
2443 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2444 "rvcng=%d unlnk=%d/%d", request->rq_receiving_reply,
2445 request->rq_req_unlink, request->rq_reply_unlink);
2449 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2451 static void ptlrpc_free_request(struct ptlrpc_request *req)
2453 spin_lock(&req->rq_lock);
2455 spin_unlock(&req->rq_lock);
2457 if (req->rq_commit_cb != NULL)
2458 req->rq_commit_cb(req);
2459 list_del_init(&req->rq_replay_list);
2461 __ptlrpc_req_finished(req, 1);
2465 * the request is committed and dropped from the replay list of its import
2467 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2469 struct obd_import *imp = req->rq_import;
2471 spin_lock(&imp->imp_lock);
2472 if (list_empty(&req->rq_replay_list)) {
2473 spin_unlock(&imp->imp_lock);
2477 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2478 ptlrpc_free_request(req);
2480 spin_unlock(&imp->imp_lock);
2482 EXPORT_SYMBOL(ptlrpc_request_committed);
2485 * Iterates through replay_list on import and prunes
2486 * all requests have transno smaller than last_committed for the
2487 * import and don't have rq_replay set.
2488 * Since requests are sorted in transno order, stops when meetign first
2489 * transno bigger than last_committed.
2490 * caller must hold imp->imp_lock
2492 void ptlrpc_free_committed(struct obd_import *imp)
2494 struct ptlrpc_request *req, *saved;
2495 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2496 bool skip_committed_list = true;
2499 LASSERT(imp != NULL);
2500 assert_spin_locked(&imp->imp_lock);
2502 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2503 imp->imp_generation == imp->imp_last_generation_checked) {
2504 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2505 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2508 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2509 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2510 imp->imp_generation);
2512 if (imp->imp_generation != imp->imp_last_generation_checked)
2513 skip_committed_list = false;
2515 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2516 imp->imp_last_generation_checked = imp->imp_generation;
2518 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2520 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2521 LASSERT(req != last_req);
2524 if (req->rq_transno == 0) {
2525 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2528 if (req->rq_import_generation < imp->imp_generation) {
2529 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2533 /* not yet committed */
2534 if (req->rq_transno > imp->imp_peer_committed_transno) {
2535 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2539 if (req->rq_replay) {
2540 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2541 list_move_tail(&req->rq_replay_list,
2542 &imp->imp_committed_list);
2546 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2547 imp->imp_peer_committed_transno);
2549 ptlrpc_free_request(req);
2552 if (skip_committed_list)
2555 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2557 LASSERT(req->rq_transno != 0);
2558 if (req->rq_import_generation < imp->imp_generation) {
2559 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2560 ptlrpc_free_request(req);
2567 void ptlrpc_cleanup_client(struct obd_import *imp)
2572 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2575 * Schedule previously sent request for resend.
2576 * For bulk requests we assign new xid (to avoid problems with
2577 * lost replies and therefore several transfers landing into same buffer
2578 * from different sending attempts).
2580 void ptlrpc_resend_req(struct ptlrpc_request *req)
2582 DEBUG_REQ(D_HA, req, "going to resend");
2583 spin_lock(&req->rq_lock);
2585 /* Request got reply but linked to the import list still.
2586 Let ptlrpc_check_set() to process it. */
2587 if (ptlrpc_client_replied(req)) {
2588 spin_unlock(&req->rq_lock);
2589 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2593 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2594 req->rq_status = -EAGAIN;
2597 req->rq_net_err = 0;
2598 req->rq_timedout = 0;
2600 __u64 old_xid = req->rq_xid;
2602 /* ensure previous bulk fails */
2603 req->rq_xid = ptlrpc_next_xid();
2604 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2605 old_xid, req->rq_xid);
2607 ptlrpc_client_wake_req(req);
2608 spin_unlock(&req->rq_lock);
2610 EXPORT_SYMBOL(ptlrpc_resend_req);
2612 /* XXX: this function and rq_status are currently unused */
2613 void ptlrpc_restart_req(struct ptlrpc_request *req)
2615 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2616 req->rq_status = -ERESTARTSYS;
2618 spin_lock(&req->rq_lock);
2619 req->rq_restart = 1;
2620 req->rq_timedout = 0;
2621 ptlrpc_client_wake_req(req);
2622 spin_unlock(&req->rq_lock);
2624 EXPORT_SYMBOL(ptlrpc_restart_req);
2627 * Grab additional reference on a request \a req
2629 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2632 atomic_inc(&req->rq_refcount);
2635 EXPORT_SYMBOL(ptlrpc_request_addref);
2638 * Add a request to import replay_list.
2639 * Must be called under imp_lock
2641 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2642 struct obd_import *imp)
2644 struct list_head *tmp;
2646 assert_spin_locked(&imp->imp_lock);
2648 if (req->rq_transno == 0) {
2649 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2653 /* clear this for new requests that were resent as well
2654 as resent replayed requests. */
2655 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2657 /* don't re-add requests that have been replayed */
2658 if (!list_empty(&req->rq_replay_list))
2661 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2663 LASSERT(imp->imp_replayable);
2664 /* Balanced in ptlrpc_free_committed, usually. */
2665 ptlrpc_request_addref(req);
2666 list_for_each_prev(tmp, &imp->imp_replay_list) {
2667 struct ptlrpc_request *iter = list_entry(tmp,
2668 struct ptlrpc_request,
2671 /* We may have duplicate transnos if we create and then
2672 * open a file, or for closes retained if to match creating
2673 * opens, so use req->rq_xid as a secondary key.
2674 * (See bugs 684, 685, and 428.)
2675 * XXX no longer needed, but all opens need transnos!
2677 if (iter->rq_transno > req->rq_transno)
2680 if (iter->rq_transno == req->rq_transno) {
2681 LASSERT(iter->rq_xid != req->rq_xid);
2682 if (iter->rq_xid > req->rq_xid)
2686 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2690 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2692 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2695 * Send request and wait until it completes.
2696 * Returns request processing status.
2698 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2700 struct ptlrpc_request_set *set;
2704 LASSERT(req->rq_set == NULL);
2705 LASSERT(!req->rq_receiving_reply);
2707 set = ptlrpc_prep_set();
2709 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2713 /* for distributed debugging */
2714 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2716 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2717 ptlrpc_request_addref(req);
2718 ptlrpc_set_add_req(set, req);
2719 rc = ptlrpc_set_wait(set);
2720 ptlrpc_set_destroy(set);
2724 EXPORT_SYMBOL(ptlrpc_queue_wait);
2726 struct ptlrpc_replay_async_args {
2728 int praa_old_status;
2732 * Callback used for replayed requests reply processing.
2733 * In case of succesful reply calls registeresd request replay callback.
2734 * In case of error restart replay process.
2736 static int ptlrpc_replay_interpret(const struct lu_env *env,
2737 struct ptlrpc_request *req,
2738 void * data, int rc)
2740 struct ptlrpc_replay_async_args *aa = data;
2741 struct obd_import *imp = req->rq_import;
2744 atomic_dec(&imp->imp_replay_inflight);
2746 if (!ptlrpc_client_replied(req)) {
2747 CERROR("request replay timed out, restarting recovery\n");
2748 GOTO(out, rc = -ETIMEDOUT);
2751 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2752 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2753 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2754 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2756 /** VBR: check version failure */
2757 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2758 /** replay was failed due to version mismatch */
2759 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2760 spin_lock(&imp->imp_lock);
2761 imp->imp_vbr_failed = 1;
2762 imp->imp_no_lock_replay = 1;
2763 spin_unlock(&imp->imp_lock);
2764 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2766 /** The transno had better not change over replay. */
2767 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2768 lustre_msg_get_transno(req->rq_repmsg) ||
2769 lustre_msg_get_transno(req->rq_repmsg) == 0,
2771 lustre_msg_get_transno(req->rq_reqmsg),
2772 lustre_msg_get_transno(req->rq_repmsg));
2775 spin_lock(&imp->imp_lock);
2776 /** if replays by version then gap occur on server, no trust to locks */
2777 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2778 imp->imp_no_lock_replay = 1;
2779 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2780 spin_unlock(&imp->imp_lock);
2781 LASSERT(imp->imp_last_replay_transno);
2783 /* transaction number shouldn't be bigger than the latest replayed */
2784 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2785 DEBUG_REQ(D_ERROR, req,
2786 "Reported transno "LPU64" is bigger than the "
2787 "replayed one: "LPU64, req->rq_transno,
2788 lustre_msg_get_transno(req->rq_reqmsg));
2789 GOTO(out, rc = -EINVAL);
2792 DEBUG_REQ(D_HA, req, "got rep");
2794 /* let the callback do fixups, possibly including in the request */
2795 if (req->rq_replay_cb)
2796 req->rq_replay_cb(req);
2798 if (ptlrpc_client_replied(req) &&
2799 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2800 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2801 lustre_msg_get_status(req->rq_repmsg),
2802 aa->praa_old_status);
2804 /* Put it back for re-replay. */
2805 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2809 * Errors while replay can set transno to 0, but
2810 * imp_last_replay_transno shouldn't be set to 0 anyway
2812 if (req->rq_transno == 0)
2813 CERROR("Transno is 0 during replay!\n");
2815 /* continue with recovery */
2816 rc = ptlrpc_import_recovery_state_machine(imp);
2818 req->rq_send_state = aa->praa_old_state;
2821 /* this replay failed, so restart recovery */
2822 ptlrpc_connect_import(imp);
2828 * Prepares and queues request for replay.
2829 * Adds it to ptlrpcd queue for actual sending.
2830 * Returns 0 on success.
2832 int ptlrpc_replay_req(struct ptlrpc_request *req)
2834 struct ptlrpc_replay_async_args *aa;
2837 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2839 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2840 aa = ptlrpc_req_async_args(req);
2841 memset(aa, 0, sizeof *aa);
2843 /* Prepare request to be resent with ptlrpcd */
2844 aa->praa_old_state = req->rq_send_state;
2845 req->rq_send_state = LUSTRE_IMP_REPLAY;
2846 req->rq_phase = RQ_PHASE_NEW;
2847 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2849 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2851 req->rq_interpret_reply = ptlrpc_replay_interpret;
2852 /* Readjust the timeout for current conditions */
2853 ptlrpc_at_set_req_timeout(req);
2855 /* Tell server the net_latency, so the server can calculate how long
2856 * it should wait for next replay */
2857 lustre_msg_set_service_time(req->rq_reqmsg,
2858 ptlrpc_at_get_net_latency(req));
2859 DEBUG_REQ(D_HA, req, "REPLAY");
2861 atomic_inc(&req->rq_import->imp_replay_inflight);
2862 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2864 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2867 EXPORT_SYMBOL(ptlrpc_replay_req);
2870 * Aborts all in-flight request on import \a imp sending and delayed lists
2872 void ptlrpc_abort_inflight(struct obd_import *imp)
2874 struct list_head *tmp, *n;
2877 /* Make sure that no new requests get processed for this import.
2878 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2879 * this flag and then putting requests on sending_list or delayed_list.
2881 spin_lock(&imp->imp_lock);
2883 /* XXX locking? Maybe we should remove each request with the list
2884 * locked? Also, how do we know if the requests on the list are
2885 * being freed at this time?
2887 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2888 struct ptlrpc_request *req = list_entry(tmp,
2889 struct ptlrpc_request,
2892 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2894 spin_lock(&req->rq_lock);
2895 if (req->rq_import_generation < imp->imp_generation) {
2897 req->rq_status = -EIO;
2898 ptlrpc_client_wake_req(req);
2900 spin_unlock(&req->rq_lock);
2903 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2904 struct ptlrpc_request *req =
2905 list_entry(tmp, struct ptlrpc_request, rq_list);
2907 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
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 /* Last chance to free reqs left on the replay list, but we
2919 * will still leak reqs that haven't committed. */
2920 if (imp->imp_replayable)
2921 ptlrpc_free_committed(imp);
2923 spin_unlock(&imp->imp_lock);
2927 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2930 * Abort all uncompleted requests in request set \a set
2932 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2934 struct list_head *tmp, *pos;
2936 LASSERT(set != NULL);
2938 list_for_each_safe(pos, tmp, &set->set_requests) {
2939 struct ptlrpc_request *req =
2940 list_entry(pos, struct ptlrpc_request,
2943 spin_lock(&req->rq_lock);
2944 if (req->rq_phase != RQ_PHASE_RPC) {
2945 spin_unlock(&req->rq_lock);
2950 req->rq_status = -EINTR;
2951 ptlrpc_client_wake_req(req);
2952 spin_unlock(&req->rq_lock);
2956 static __u64 ptlrpc_last_xid;
2957 static spinlock_t ptlrpc_last_xid_lock;
2960 * Initialize the XID for the node. This is common among all requests on
2961 * this node, and only requires the property that it is monotonically
2962 * increasing. It does not need to be sequential. Since this is also used
2963 * as the RDMA match bits, it is important that a single client NOT have
2964 * the same match bits for two different in-flight requests, hence we do
2965 * NOT want to have an XID per target or similar.
2967 * To avoid an unlikely collision between match bits after a client reboot
2968 * (which would deliver old data into the wrong RDMA buffer) initialize
2969 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2970 * If the time is clearly incorrect, we instead use a 62-bit random number.
2971 * In the worst case the random number will overflow 1M RPCs per second in
2972 * 9133 years, or permutations thereof.
2974 #define YEAR_2004 (1ULL << 30)
2975 void ptlrpc_init_xid(void)
2977 time_t now = cfs_time_current_sec();
2979 spin_lock_init(&ptlrpc_last_xid_lock);
2980 if (now < YEAR_2004) {
2981 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2982 ptlrpc_last_xid >>= 2;
2983 ptlrpc_last_xid |= (1ULL << 61);
2985 ptlrpc_last_xid = (__u64)now << 20;
2988 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
2989 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
2990 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
2994 * Increase xid and returns resulting new value to the caller.
2996 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2997 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2998 * itself uses the last bulk xid needed, so the server can determine the
2999 * the number of bulk transfers from the RPC XID and a bitmask. The starting
3000 * xid must align to a power-of-two value.
3002 * This is assumed to be true due to the initial ptlrpc_last_xid
3003 * value also being initialized to a power-of-two value. LU-1431
3005 __u64 ptlrpc_next_xid(void)
3009 spin_lock(&ptlrpc_last_xid_lock);
3010 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3011 ptlrpc_last_xid = next;
3012 spin_unlock(&ptlrpc_last_xid_lock);
3016 EXPORT_SYMBOL(ptlrpc_next_xid);
3019 * Get a glimpse at what next xid value might have been.
3020 * Returns possible next xid.
3022 __u64 ptlrpc_sample_next_xid(void)
3024 #if BITS_PER_LONG == 32
3025 /* need to avoid possible word tearing on 32-bit systems */
3028 spin_lock(&ptlrpc_last_xid_lock);
3029 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3030 spin_unlock(&ptlrpc_last_xid_lock);
3034 /* No need to lock, since returned value is racy anyways */
3035 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3038 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3041 * Functions for operating ptlrpc workers.
3043 * A ptlrpc work is a function which will be running inside ptlrpc context.
3044 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3046 * 1. after a work is created, it can be used many times, that is:
3047 * handler = ptlrpcd_alloc_work();
3048 * ptlrpcd_queue_work();
3050 * queue it again when necessary:
3051 * ptlrpcd_queue_work();
3052 * ptlrpcd_destroy_work();
3053 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3054 * it will only be queued once in any time. Also as its name implies, it may
3055 * have delay before it really runs by ptlrpcd thread.
3057 struct ptlrpc_work_async_args {
3058 int (*cb)(const struct lu_env *, void *);
3062 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3064 /* re-initialize the req */
3065 req->rq_timeout = obd_timeout;
3066 req->rq_sent = cfs_time_current_sec();
3067 req->rq_deadline = req->rq_sent + req->rq_timeout;
3068 req->rq_reply_deadline = req->rq_deadline;
3069 req->rq_phase = RQ_PHASE_INTERPRET;
3070 req->rq_next_phase = RQ_PHASE_COMPLETE;
3071 req->rq_xid = ptlrpc_next_xid();
3072 req->rq_import_generation = req->rq_import->imp_generation;
3074 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3077 static int work_interpreter(const struct lu_env *env,
3078 struct ptlrpc_request *req, void *data, int rc)
3080 struct ptlrpc_work_async_args *arg = data;
3082 LASSERT(ptlrpcd_check_work(req));
3083 LASSERT(arg->cb != NULL);
3085 rc = arg->cb(env, arg->cbdata);
3087 list_del_init(&req->rq_set_chain);
3090 if (atomic_dec_return(&req->rq_refcount) > 1) {
3091 atomic_set(&req->rq_refcount, 2);
3092 ptlrpcd_add_work_req(req);
3097 static int worker_format;
3099 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3101 return req->rq_pill.rc_fmt == (void *)&worker_format;
3105 * Create a work for ptlrpc.
3107 void *ptlrpcd_alloc_work(struct obd_import *imp,
3108 int (*cb)(const struct lu_env *, void *), void *cbdata)
3110 struct ptlrpc_request *req = NULL;
3111 struct ptlrpc_work_async_args *args;
3117 RETURN(ERR_PTR(-EINVAL));
3119 /* copy some code from deprecated fakereq. */
3120 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3122 CERROR("ptlrpc: run out of memory!\n");
3123 RETURN(ERR_PTR(-ENOMEM));
3126 req->rq_send_state = LUSTRE_IMP_FULL;
3127 req->rq_type = PTL_RPC_MSG_REQUEST;
3128 req->rq_import = class_import_get(imp);
3129 req->rq_export = NULL;
3130 req->rq_interpret_reply = work_interpreter;
3131 /* don't want reply */
3132 req->rq_receiving_reply = 0;
3133 req->rq_req_unlink = req->rq_reply_unlink = 0;
3134 req->rq_no_delay = req->rq_no_resend = 1;
3135 req->rq_pill.rc_fmt = (void *)&worker_format;
3137 spin_lock_init(&req->rq_lock);
3138 INIT_LIST_HEAD(&req->rq_list);
3139 INIT_LIST_HEAD(&req->rq_replay_list);
3140 INIT_LIST_HEAD(&req->rq_set_chain);
3141 INIT_LIST_HEAD(&req->rq_history_list);
3142 INIT_LIST_HEAD(&req->rq_exp_list);
3143 init_waitqueue_head(&req->rq_reply_waitq);
3144 init_waitqueue_head(&req->rq_set_waitq);
3145 atomic_set(&req->rq_refcount, 1);
3147 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3148 args = ptlrpc_req_async_args(req);
3150 args->cbdata = cbdata;
3154 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3156 void ptlrpcd_destroy_work(void *handler)
3158 struct ptlrpc_request *req = handler;
3161 ptlrpc_req_finished(req);
3163 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3165 int ptlrpcd_queue_work(void *handler)
3167 struct ptlrpc_request *req = handler;
3170 * Check if the req is already being queued.
3172 * Here comes a trick: it lacks a way of checking if a req is being
3173 * processed reliably in ptlrpc. Here I have to use refcount of req
3174 * for this purpose. This is okay because the caller should use this
3175 * req as opaque data. - Jinshan
3177 LASSERT(atomic_read(&req->rq_refcount) > 0);
3178 if (atomic_inc_return(&req->rq_refcount) == 2)
3179 ptlrpcd_add_work_req(req);
3182 EXPORT_SYMBOL(ptlrpcd_queue_work);