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, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 /** Implementation of client-side PortalRPC interfaces */
39 #define DEBUG_SUBSYSTEM S_RPC
43 #include <liblustre.h>
46 #include <obd_support.h>
47 #include <obd_class.h>
48 #include <lustre_lib.h>
49 #include <lustre_ha.h>
50 #include <lustre_import.h>
51 #include <lustre_req_layout.h>
53 #include "ptlrpc_internal.h"
55 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
58 * Initialize passed in client structure \a cl.
60 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
61 struct ptlrpc_client *cl)
63 cl->cli_request_portal = req_portal;
64 cl->cli_reply_portal = rep_portal;
67 EXPORT_SYMBOL(ptlrpc_init_client);
70 * Return PortalRPC connection for remore uud \a uuid
72 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
74 struct ptlrpc_connection *c;
76 lnet_process_id_t peer;
79 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
80 * before accessing its values. */
81 /* coverity[uninit_use_in_call] */
82 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
84 CNETERR("cannot find peer %s!\n", uuid->uuid);
88 c = ptlrpc_connection_get(peer, self, uuid);
90 memcpy(c->c_remote_uuid.uuid,
91 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
94 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
98 EXPORT_SYMBOL(ptlrpc_uuid_to_connection);
101 * Allocate and initialize new bulk descriptor
102 * Returns pointer to the descriptor or NULL on error.
104 struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
106 struct ptlrpc_bulk_desc *desc;
108 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
112 cfs_spin_lock_init(&desc->bd_lock);
113 cfs_waitq_init(&desc->bd_waitq);
114 desc->bd_max_iov = npages;
115 desc->bd_iov_count = 0;
116 LNetInvalidateHandle(&desc->bd_md_h);
117 desc->bd_portal = portal;
118 desc->bd_type = type;
124 * Prepare bulk descriptor for specified outgoing request \a req that
125 * can fit \a npages * pages. \a type is bulk type. \a portal is where
126 * the bulk to be sent. Used on client-side.
127 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
130 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
131 int npages, int type, int portal)
133 struct obd_import *imp = req->rq_import;
134 struct ptlrpc_bulk_desc *desc;
137 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
138 desc = new_bulk(npages, type, portal);
142 desc->bd_import_generation = req->rq_import_generation;
143 desc->bd_import = class_import_get(imp);
146 desc->bd_cbid.cbid_fn = client_bulk_callback;
147 desc->bd_cbid.cbid_arg = desc;
149 /* This makes req own desc, and free it when she frees herself */
154 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
157 * Add a page \a page to the bulk descriptor \a desc.
158 * Data to transfer in the page starts at offset \a pageoffset and
159 * amount of data to transfer from the page is \a len
161 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
162 cfs_page_t *page, int pageoffset, int len, int pin)
164 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
165 LASSERT(page != NULL);
166 LASSERT(pageoffset >= 0);
168 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
175 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
177 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
180 * Uninitialize and free bulk descriptor \a desc.
181 * Works on bulk descriptors both from server and client side.
183 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin)
188 LASSERT(desc != NULL);
189 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
190 LASSERT(!desc->bd_network_rw); /* network hands off or */
191 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
193 sptlrpc_enc_pool_put_pages(desc);
196 class_export_put(desc->bd_export);
198 class_import_put(desc->bd_import);
201 for (i = 0; i < desc->bd_iov_count ; i++)
202 cfs_page_unpin(desc->bd_iov[i].kiov_page);
205 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
206 bd_iov[desc->bd_max_iov]));
209 EXPORT_SYMBOL(__ptlrpc_free_bulk);
212 * Set server timelimit for this req, i.e. how long are we willing to wait
213 * for reply before timing out this request.
215 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
221 LASSERT(req->rq_import);
224 /* non-AT settings */
226 * \a imp_server_timeout means this is reverse import and
227 * we send (currently only) ASTs to the client and cannot afford
228 * to wait too long for the reply, otherwise the other client
229 * (because of which we are sending this request) would
230 * timeout waiting for us
232 req->rq_timeout = req->rq_import->imp_server_timeout ?
233 obd_timeout / 2 : obd_timeout;
235 at = &req->rq_import->imp_at;
236 idx = import_at_get_index(req->rq_import,
237 req->rq_request_portal);
238 serv_est = at_get(&at->iat_service_estimate[idx]);
239 req->rq_timeout = at_est2timeout(serv_est);
241 /* We could get even fancier here, using history to predict increased
244 /* Let the server know what this RPC timeout is by putting it in the
246 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
248 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
250 /* Adjust max service estimate based on server value */
251 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
252 unsigned int serv_est)
258 LASSERT(req->rq_import);
259 at = &req->rq_import->imp_at;
261 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
262 /* max service estimates are tracked on the server side,
263 so just keep minimal history here */
264 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
266 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
267 "has changed from %d to %d\n",
268 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
269 oldse, at_get(&at->iat_service_estimate[idx]));
272 /* Expected network latency per remote node (secs) */
273 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
275 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
278 /* Adjust expected network latency */
279 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
280 unsigned int service_time)
282 unsigned int nl, oldnl;
284 time_t now = cfs_time_current_sec();
286 LASSERT(req->rq_import);
287 at = &req->rq_import->imp_at;
289 /* Network latency is total time less server processing time */
290 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
291 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
292 CWARN("Reported service time %u > total measured time "
293 CFS_DURATION_T"\n", service_time,
294 cfs_time_sub(now, req->rq_sent));
296 oldnl = at_measured(&at->iat_net_latency, nl);
298 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
299 "has changed from %d to %d\n",
300 req->rq_import->imp_obd->obd_name,
302 &req->rq_import->imp_connection->c_remote_uuid),
303 oldnl, at_get(&at->iat_net_latency));
306 static int unpack_reply(struct ptlrpc_request *req)
310 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
311 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
313 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
318 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
320 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
327 * Handle an early reply message, called with the rq_lock held.
328 * If anything goes wrong just ignore it - same as if it never happened
330 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
332 struct ptlrpc_request *early_req;
338 cfs_spin_unlock(&req->rq_lock);
340 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
342 cfs_spin_lock(&req->rq_lock);
346 rc = unpack_reply(early_req);
348 /* Expecting to increase the service time estimate here */
349 ptlrpc_at_adj_service(req,
350 lustre_msg_get_timeout(early_req->rq_repmsg));
351 ptlrpc_at_adj_net_latency(req,
352 lustre_msg_get_service_time(early_req->rq_repmsg));
355 sptlrpc_cli_finish_early_reply(early_req);
357 cfs_spin_lock(&req->rq_lock);
360 /* Adjust the local timeout for this req */
361 ptlrpc_at_set_req_timeout(req);
363 olddl = req->rq_deadline;
364 /* server assumes it now has rq_timeout from when it sent the
365 early reply, so client should give it at least that long. */
366 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
367 ptlrpc_at_get_net_latency(req);
369 DEBUG_REQ(D_ADAPTTO, req,
370 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
371 "("CFS_DURATION_T"s)", req->rq_early_count,
372 cfs_time_sub(req->rq_deadline,
373 cfs_time_current_sec()),
374 cfs_time_sub(req->rq_deadline, olddl));
381 * Wind down request pool \a pool.
382 * Frees all requests from the pool too
384 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
387 struct ptlrpc_request *req;
389 LASSERT(pool != NULL);
391 cfs_spin_lock(&pool->prp_lock);
392 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
393 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
394 cfs_list_del(&req->rq_list);
395 LASSERT(req->rq_reqbuf);
396 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
397 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
398 OBD_FREE(req, sizeof(*req));
400 cfs_spin_unlock(&pool->prp_lock);
401 OBD_FREE(pool, sizeof(*pool));
403 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
406 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
408 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
413 while (size < pool->prp_rq_size)
416 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
417 size == pool->prp_rq_size,
418 "Trying to change pool size with nonempty pool "
419 "from %d to %d bytes\n", pool->prp_rq_size, size);
421 cfs_spin_lock(&pool->prp_lock);
422 pool->prp_rq_size = size;
423 for (i = 0; i < num_rq; i++) {
424 struct ptlrpc_request *req;
425 struct lustre_msg *msg;
427 cfs_spin_unlock(&pool->prp_lock);
428 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
431 OBD_ALLOC_LARGE(msg, size);
433 OBD_FREE(req, sizeof(struct ptlrpc_request));
436 req->rq_reqbuf = msg;
437 req->rq_reqbuf_len = size;
439 cfs_spin_lock(&pool->prp_lock);
440 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
442 cfs_spin_unlock(&pool->prp_lock);
445 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
448 * Create and initialize new request pool with given attributes:
449 * \a num_rq - initial number of requests to create for the pool
450 * \a msgsize - maximum message size possible for requests in thid pool
451 * \a populate_pool - function to be called when more requests need to be added
453 * Returns pointer to newly created pool or NULL on error.
455 struct ptlrpc_request_pool *
456 ptlrpc_init_rq_pool(int num_rq, int msgsize,
457 void (*populate_pool)(struct ptlrpc_request_pool *, int))
459 struct ptlrpc_request_pool *pool;
461 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
465 /* Request next power of two for the allocation, because internally
466 kernel would do exactly this */
468 cfs_spin_lock_init(&pool->prp_lock);
469 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
470 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
471 pool->prp_populate = populate_pool;
473 populate_pool(pool, num_rq);
475 if (cfs_list_empty(&pool->prp_req_list)) {
476 /* have not allocated a single request for the pool */
477 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
482 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
485 * Fetches one request from pool \a pool
487 static struct ptlrpc_request *
488 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
490 struct ptlrpc_request *request;
491 struct lustre_msg *reqbuf;
496 cfs_spin_lock(&pool->prp_lock);
498 /* See if we have anything in a pool, and bail out if nothing,
499 * in writeout path, where this matters, this is safe to do, because
500 * nothing is lost in this case, and when some in-flight requests
501 * complete, this code will be called again. */
502 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
503 cfs_spin_unlock(&pool->prp_lock);
507 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
509 cfs_list_del_init(&request->rq_list);
510 cfs_spin_unlock(&pool->prp_lock);
512 LASSERT(request->rq_reqbuf);
513 LASSERT(request->rq_pool);
515 reqbuf = request->rq_reqbuf;
516 memset(request, 0, sizeof(*request));
517 request->rq_reqbuf = reqbuf;
518 request->rq_reqbuf_len = pool->prp_rq_size;
519 request->rq_pool = pool;
525 * Returns freed \a request to pool.
527 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
529 struct ptlrpc_request_pool *pool = request->rq_pool;
531 cfs_spin_lock(&pool->prp_lock);
532 LASSERT(cfs_list_empty(&request->rq_list));
533 LASSERT(!request->rq_receiving_reply);
534 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
535 cfs_spin_unlock(&pool->prp_lock);
538 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
539 __u32 version, int opcode,
540 int count, __u32 *lengths, char **bufs,
541 struct ptlrpc_cli_ctx *ctx)
543 struct obd_import *imp = request->rq_import;
548 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
550 rc = sptlrpc_req_get_ctx(request);
555 sptlrpc_req_set_flavor(request, opcode);
557 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
560 LASSERT(!request->rq_pool);
564 lustre_msg_add_version(request->rq_reqmsg, version);
565 request->rq_send_state = LUSTRE_IMP_FULL;
566 request->rq_type = PTL_RPC_MSG_REQUEST;
567 request->rq_export = NULL;
569 request->rq_req_cbid.cbid_fn = request_out_callback;
570 request->rq_req_cbid.cbid_arg = request;
572 request->rq_reply_cbid.cbid_fn = reply_in_callback;
573 request->rq_reply_cbid.cbid_arg = request;
575 request->rq_reply_deadline = 0;
576 request->rq_phase = RQ_PHASE_NEW;
577 request->rq_next_phase = RQ_PHASE_UNDEFINED;
579 request->rq_request_portal = imp->imp_client->cli_request_portal;
580 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
582 ptlrpc_at_set_req_timeout(request);
584 cfs_spin_lock_init(&request->rq_lock);
585 CFS_INIT_LIST_HEAD(&request->rq_list);
586 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
587 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
588 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
589 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
590 CFS_INIT_LIST_HEAD(&request->rq_history_list);
591 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
592 cfs_waitq_init(&request->rq_reply_waitq);
593 cfs_waitq_init(&request->rq_set_waitq);
594 request->rq_xid = ptlrpc_next_xid();
595 cfs_atomic_set(&request->rq_refcount, 1);
597 lustre_msg_set_opc(request->rq_reqmsg, opcode);
601 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
603 class_import_put(imp);
607 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
608 __u32 version, int opcode, char **bufs,
609 struct ptlrpc_cli_ctx *ctx)
613 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
614 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
615 request->rq_pill.rc_area[RCL_CLIENT],
618 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
621 * Pack request buffers for network transfer, performing necessary encryption
622 * steps if necessary.
624 int ptlrpc_request_pack(struct ptlrpc_request *request,
625 __u32 version, int opcode)
628 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
632 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
633 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
634 * have to send old ptlrpc_body to keep interoprability with these
637 * Only three kinds of server->client RPCs so far:
642 * XXX This should be removed whenever we drop the interoprability with
643 * the these old clients.
645 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
646 opcode == LDLM_GL_CALLBACK)
647 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
648 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
652 EXPORT_SYMBOL(ptlrpc_request_pack);
655 * Helper function to allocate new request on import \a imp
656 * and possibly using existing request from pool \a pool if provided.
657 * Returns allocated request structure with import field filled or
661 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
662 struct ptlrpc_request_pool *pool)
664 struct ptlrpc_request *request = NULL;
667 request = ptlrpc_prep_req_from_pool(pool);
670 OBD_ALLOC_PTR(request);
673 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
674 LASSERT(imp != LP_POISON);
675 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
677 LASSERT(imp->imp_client != LP_POISON);
679 request->rq_import = class_import_get(imp);
681 CERROR("request allocation out of memory\n");
688 * Helper function for creating a request.
689 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
690 * buffer structures according to capsule template \a format.
691 * Returns allocated request structure pointer or NULL on error.
693 static struct ptlrpc_request *
694 ptlrpc_request_alloc_internal(struct obd_import *imp,
695 struct ptlrpc_request_pool * pool,
696 const struct req_format *format)
698 struct ptlrpc_request *request;
700 request = __ptlrpc_request_alloc(imp, pool);
704 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
705 req_capsule_set(&request->rq_pill, format);
710 * Allocate new request structure for import \a imp and initialize its
711 * buffer structure according to capsule template \a format.
713 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
714 const struct req_format *format)
716 return ptlrpc_request_alloc_internal(imp, NULL, format);
718 EXPORT_SYMBOL(ptlrpc_request_alloc);
721 * Allocate new request structure for import \a imp from pool \a pool and
722 * initialize its buffer structure according to capsule template \a format.
724 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
725 struct ptlrpc_request_pool * pool,
726 const struct req_format *format)
728 return ptlrpc_request_alloc_internal(imp, pool, format);
730 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
733 * For requests not from pool, free memory of the request structure.
734 * For requests obtained from a pool earlier, return request back to pool.
736 void ptlrpc_request_free(struct ptlrpc_request *request)
738 if (request->rq_pool)
739 __ptlrpc_free_req_to_pool(request);
741 OBD_FREE_PTR(request);
743 EXPORT_SYMBOL(ptlrpc_request_free);
746 * Allocate new request for operatione \a opcode and immediatelly pack it for
748 * Only used for simple requests like OBD_PING where the only important
749 * part of the request is operation itself.
750 * Returns allocated request or NULL on error.
752 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
753 const struct req_format *format,
754 __u32 version, int opcode)
756 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
760 rc = ptlrpc_request_pack(req, version, opcode);
762 ptlrpc_request_free(req);
768 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
771 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
772 * for operation \a opcode. Request would contain \a count buffers.
773 * Sizes of buffers are described in array \a lengths and buffers themselves
774 * are provided by a pointer \a bufs.
775 * Returns prepared request structure pointer or NULL on error.
777 struct ptlrpc_request *
778 ptlrpc_prep_req_pool(struct obd_import *imp,
779 __u32 version, int opcode,
780 int count, __u32 *lengths, char **bufs,
781 struct ptlrpc_request_pool *pool)
783 struct ptlrpc_request *request;
786 request = __ptlrpc_request_alloc(imp, pool);
790 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
791 lengths, bufs, NULL);
793 ptlrpc_request_free(request);
798 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
801 * Same as ptlrpc_prep_req_pool, but without pool
803 struct ptlrpc_request *
804 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
805 __u32 *lengths, char **bufs)
807 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
810 EXPORT_SYMBOL(ptlrpc_prep_req);
813 * Allocate "fake" request that would not be sent anywhere in the end.
814 * Only used as a hack because we have no other way of performing
815 * async actions in lustre between layers.
816 * Used on MDS to request object preallocations from more than one OST at a
819 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
820 unsigned int timeout,
821 ptlrpc_interpterer_t interpreter)
823 struct ptlrpc_request *request = NULL;
826 OBD_ALLOC(request, sizeof(*request));
828 CERROR("request allocation out of memory\n");
832 request->rq_send_state = LUSTRE_IMP_FULL;
833 request->rq_type = PTL_RPC_MSG_REQUEST;
834 request->rq_import = class_import_get(imp);
835 request->rq_export = NULL;
836 request->rq_import_generation = imp->imp_generation;
838 request->rq_timeout = timeout;
839 request->rq_sent = cfs_time_current_sec();
840 request->rq_deadline = request->rq_sent + timeout;
841 request->rq_reply_deadline = request->rq_deadline;
842 request->rq_interpret_reply = interpreter;
843 request->rq_phase = RQ_PHASE_RPC;
844 request->rq_next_phase = RQ_PHASE_INTERPRET;
845 /* don't want reply */
846 request->rq_receiving_reply = 0;
847 request->rq_must_unlink = 0;
848 request->rq_no_delay = request->rq_no_resend = 1;
849 request->rq_fake = 1;
851 cfs_spin_lock_init(&request->rq_lock);
852 CFS_INIT_LIST_HEAD(&request->rq_list);
853 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
854 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
855 CFS_INIT_LIST_HEAD(&request->rq_history_list);
856 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
857 cfs_waitq_init(&request->rq_reply_waitq);
858 cfs_waitq_init(&request->rq_set_waitq);
860 request->rq_xid = ptlrpc_next_xid();
861 cfs_atomic_set(&request->rq_refcount, 1);
865 EXPORT_SYMBOL(ptlrpc_prep_fakereq);
868 * Indicate that processing of "fake" request is finished.
870 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
872 struct ptlrpc_request_set *set = req->rq_set;
875 /* hold ref on the request to prevent others (ptlrpcd) to free it */
876 ptlrpc_request_addref(req);
877 cfs_list_del_init(&req->rq_list);
879 /* if we kill request before timeout - need adjust counter */
880 if (req->rq_phase == RQ_PHASE_RPC && set != NULL &&
881 cfs_atomic_dec_and_test(&set->set_remaining))
884 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
886 /* Only need to call wakeup once when to be empty. */
888 cfs_waitq_signal(&set->set_waitq);
889 ptlrpc_req_finished(req);
891 EXPORT_SYMBOL(ptlrpc_fakereq_finished);
894 * Allocate and initialize new request set structure.
895 * Returns a pointer to the newly allocated set structure or NULL on error.
897 struct ptlrpc_request_set *ptlrpc_prep_set(void)
899 struct ptlrpc_request_set *set;
902 OBD_ALLOC(set, sizeof *set);
905 cfs_atomic_set(&set->set_refcount, 1);
906 CFS_INIT_LIST_HEAD(&set->set_requests);
907 cfs_waitq_init(&set->set_waitq);
908 cfs_atomic_set(&set->set_new_count, 0);
909 cfs_atomic_set(&set->set_remaining, 0);
910 cfs_spin_lock_init(&set->set_new_req_lock);
911 CFS_INIT_LIST_HEAD(&set->set_new_requests);
912 CFS_INIT_LIST_HEAD(&set->set_cblist);
913 set->set_max_inflight = UINT_MAX;
914 set->set_producer = NULL;
915 set->set_producer_arg = NULL;
920 EXPORT_SYMBOL(ptlrpc_prep_set);
923 * Allocate and initialize new request set structure with flow control
924 * extension. This extension allows to control the number of requests in-flight
925 * for the whole set. A callback function to generate requests must be provided
926 * and the request set will keep the number of requests sent over the wire to
928 * Returns a pointer to the newly allocated set structure or NULL on error.
930 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
934 struct ptlrpc_request_set *set;
936 set = ptlrpc_prep_set();
940 set->set_max_inflight = max;
941 set->set_producer = func;
942 set->set_producer_arg = arg;
946 EXPORT_SYMBOL(ptlrpc_prep_fcset);
949 * Wind down and free request set structure previously allocated with
951 * Ensures that all requests on the set have completed and removes
952 * all requests from the request list in a set.
953 * If any unsent request happen to be on the list, pretends that they got
954 * an error in flight and calls their completion handler.
956 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
964 /* Requests on the set should either all be completed, or all be new */
965 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
966 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
967 cfs_list_for_each (tmp, &set->set_requests) {
968 struct ptlrpc_request *req =
969 cfs_list_entry(tmp, struct ptlrpc_request,
972 LASSERT(req->rq_phase == expected_phase);
976 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
977 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
978 cfs_atomic_read(&set->set_remaining), n);
980 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
981 struct ptlrpc_request *req =
982 cfs_list_entry(tmp, struct ptlrpc_request,
984 cfs_list_del_init(&req->rq_set_chain);
986 LASSERT(req->rq_phase == expected_phase);
988 if (req->rq_phase == RQ_PHASE_NEW) {
989 ptlrpc_req_interpret(NULL, req, -EBADR);
990 cfs_atomic_dec(&set->set_remaining);
993 cfs_spin_lock(&req->rq_lock);
995 req->rq_invalid_rqset = 0;
996 cfs_spin_unlock(&req->rq_lock);
998 ptlrpc_req_finished (req);
1001 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
1003 ptlrpc_reqset_put(set);
1006 EXPORT_SYMBOL(ptlrpc_set_destroy);
1009 * Add a callback function \a fn to the set.
1010 * This function would be called when all requests on this set are completed.
1011 * The function will be passed \a data argument.
1013 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
1014 set_interpreter_func fn, void *data)
1016 struct ptlrpc_set_cbdata *cbdata;
1018 OBD_ALLOC_PTR(cbdata);
1022 cbdata->psc_interpret = fn;
1023 cbdata->psc_data = data;
1024 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
1028 EXPORT_SYMBOL(ptlrpc_set_add_cb);
1031 * Add a new request to the general purpose request set.
1032 * Assumes request reference from the caller.
1034 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1035 struct ptlrpc_request *req)
1037 LASSERT(cfs_list_empty(&req->rq_set_chain));
1039 /* The set takes over the caller's request reference */
1040 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
1042 cfs_atomic_inc(&set->set_remaining);
1043 req->rq_queued_time = cfs_time_current();
1045 if (req->rq_reqmsg != NULL)
1046 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1048 if (set->set_producer != NULL)
1049 /* If the request set has a producer callback, the RPC must be
1050 * sent straight away */
1051 ptlrpc_send_new_req(req);
1053 EXPORT_SYMBOL(ptlrpc_set_add_req);
1056 * Add a request to a request with dedicated server thread
1057 * and wake the thread to make any necessary processing.
1058 * Currently only used for ptlrpcd.
1060 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1061 struct ptlrpc_request *req)
1063 struct ptlrpc_request_set *set = pc->pc_set;
1066 LASSERT(req->rq_set == NULL);
1067 LASSERT(cfs_test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1069 cfs_spin_lock(&set->set_new_req_lock);
1071 * The set takes over the caller's request reference.
1074 req->rq_queued_time = cfs_time_current();
1075 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1076 count = cfs_atomic_inc_return(&set->set_new_count);
1077 cfs_spin_unlock(&set->set_new_req_lock);
1079 /* Only need to call wakeup once for the first entry. */
1081 cfs_waitq_signal(&set->set_waitq);
1083 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1084 * guarantee the async RPC can be processed ASAP, we have
1085 * no other better choice. It maybe fixed in future. */
1086 for (i = 0; i < pc->pc_npartners; i++)
1087 cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
1090 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1093 * Based on the current state of the import, determine if the request
1094 * can be sent, is an error, or should be delayed.
1096 * Returns true if this request should be delayed. If false, and
1097 * *status is set, then the request can not be sent and *status is the
1098 * error code. If false and status is 0, then request can be sent.
1100 * The imp->imp_lock must be held.
1102 static int ptlrpc_import_delay_req(struct obd_import *imp,
1103 struct ptlrpc_request *req, int *status)
1108 LASSERT (status != NULL);
1111 if (req->rq_ctx_init || req->rq_ctx_fini) {
1112 /* always allow ctx init/fini rpc go through */
1113 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1114 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1116 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1117 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1119 } else if (ptlrpc_send_limit_expired(req)) {
1120 /* probably doesn't need to be a D_ERROR after initial testing */
1121 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1123 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1124 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1125 /* allow CONNECT even if import is invalid */ ;
1126 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1127 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1130 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1131 if (!imp->imp_deactive)
1132 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1133 *status = -ESHUTDOWN; /* bz 12940 */
1134 } else if (req->rq_import_generation != imp->imp_generation) {
1135 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1137 } else if (req->rq_send_state != imp->imp_state) {
1138 /* invalidate in progress - any requests should be drop */
1139 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1140 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1142 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1143 *status = -EWOULDBLOCK;
1153 * Decide if the eror message regarding provided request \a req
1154 * should be printed to the console or not.
1155 * Makes it's decision on request status and other properties.
1156 * Returns 1 to print error on the system console or 0 if not.
1158 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1163 /* Fake requests include no rq_reqmsg */
1167 LASSERT(req->rq_reqmsg != NULL);
1168 opc = lustre_msg_get_opc(req->rq_reqmsg);
1170 /* Suppress particular reconnect errors which are to be expected. No
1171 * errors are suppressed for the initial connection on an import */
1172 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1173 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1175 /* Suppress timed out reconnect requests */
1176 if (req->rq_timedout)
1179 /* Suppress unavailable/again reconnect requests */
1180 err = lustre_msg_get_status(req->rq_repmsg);
1181 if (err == -ENODEV || err == -EAGAIN)
1189 * Check request processing status.
1190 * Returns the status.
1192 static int ptlrpc_check_status(struct ptlrpc_request *req)
1197 err = lustre_msg_get_status(req->rq_repmsg);
1198 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1199 struct obd_import *imp = req->rq_import;
1200 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1201 if (ptlrpc_console_allow(req))
1202 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1203 "communicating with %s. The %s "
1204 "operation failed with %d\n",
1206 imp->imp_connection->c_peer.nid),
1207 ll_opcode2str(opc), err);
1208 RETURN(err < 0 ? err : -EINVAL);
1212 DEBUG_REQ(D_INFO, req, "status is %d", err);
1213 } else if (err > 0) {
1214 /* XXX: translate this error from net to host */
1215 DEBUG_REQ(D_INFO, req, "status is %d", err);
1222 * save pre-versions of objects into request for replay.
1223 * Versions are obtained from server reply.
1226 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1228 struct lustre_msg *repmsg = req->rq_repmsg;
1229 struct lustre_msg *reqmsg = req->rq_reqmsg;
1230 __u64 *versions = lustre_msg_get_versions(repmsg);
1233 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1237 lustre_msg_set_versions(reqmsg, versions);
1238 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1239 versions[0], versions[1]);
1245 * Callback function called when client receives RPC reply for \a req.
1246 * Returns 0 on success or error code.
1247 * The return alue would be assigned to req->rq_status by the caller
1248 * as request processing status.
1249 * This function also decides if the request needs to be saved for later replay.
1251 static int after_reply(struct ptlrpc_request *req)
1253 struct obd_import *imp = req->rq_import;
1254 struct obd_device *obd = req->rq_import->imp_obd;
1256 struct timeval work_start;
1260 LASSERT(obd != NULL);
1261 /* repbuf must be unlinked */
1262 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1264 if (req->rq_reply_truncate) {
1265 if (ptlrpc_no_resend(req)) {
1266 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1267 " expected: %d, actual size: %d",
1268 req->rq_nob_received, req->rq_repbuf_len);
1272 sptlrpc_cli_free_repbuf(req);
1273 /* Pass the required reply buffer size (include
1274 * space for early reply).
1275 * NB: no need to roundup because alloc_repbuf
1276 * will roundup it */
1277 req->rq_replen = req->rq_nob_received;
1278 req->rq_nob_received = 0;
1284 * NB Until this point, the whole of the incoming message,
1285 * including buflens, status etc is in the sender's byte order.
1287 rc = sptlrpc_cli_unwrap_reply(req);
1289 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1293 /* retry indefinitely on EINPROGRESS */
1294 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1295 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1296 time_t now = cfs_time_current_sec();
1298 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1300 req->rq_nr_resend++;
1302 /* Readjust the timeout for current conditions */
1303 ptlrpc_at_set_req_timeout(req);
1304 /* delay resend to give a chance to the server to get ready.
1305 * The delay is increased by 1s on every resend and is capped to
1306 * the current request timeout (i.e. obd_timeout if AT is off,
1307 * or AT service time x 125% + 5s, see at_est2timeout) */
1308 if (req->rq_nr_resend > req->rq_timeout)
1309 req->rq_sent = now + req->rq_timeout;
1311 req->rq_sent = now + req->rq_nr_resend;
1315 * Security layer unwrap might ask resend this request.
1320 rc = unpack_reply(req);
1324 cfs_gettimeofday(&work_start);
1325 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1326 if (obd->obd_svc_stats != NULL) {
1327 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1329 ptlrpc_lprocfs_rpc_sent(req, timediff);
1332 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1333 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1334 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1335 lustre_msg_get_type(req->rq_repmsg));
1339 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1340 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1341 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1342 ptlrpc_at_adj_net_latency(req,
1343 lustre_msg_get_service_time(req->rq_repmsg));
1345 rc = ptlrpc_check_status(req);
1346 imp->imp_connect_error = rc;
1350 * Either we've been evicted, or the server has failed for
1351 * some reason. Try to reconnect, and if that fails, punt to
1354 if (ll_rpc_recoverable_error(rc)) {
1355 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1356 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1359 ptlrpc_request_handle_notconn(req);
1364 * Let's look if server sent slv. Do it only for RPC with
1367 ldlm_cli_update_pool(req);
1371 * Store transno in reqmsg for replay.
1373 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1374 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1375 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1378 if (imp->imp_replayable) {
1379 cfs_spin_lock(&imp->imp_lock);
1381 * No point in adding already-committed requests to the replay
1382 * list, we will just remove them immediately. b=9829
1384 if (req->rq_transno != 0 &&
1386 lustre_msg_get_last_committed(req->rq_repmsg) ||
1388 /** version recovery */
1389 ptlrpc_save_versions(req);
1390 ptlrpc_retain_replayable_request(req, imp);
1391 } else if (req->rq_commit_cb != NULL) {
1392 cfs_spin_unlock(&imp->imp_lock);
1393 req->rq_commit_cb(req);
1394 cfs_spin_lock(&imp->imp_lock);
1398 * Replay-enabled imports return commit-status information.
1400 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1401 imp->imp_peer_committed_transno =
1402 lustre_msg_get_last_committed(req->rq_repmsg);
1404 ptlrpc_free_committed(imp);
1406 if (req->rq_transno > imp->imp_peer_committed_transno)
1407 ptlrpc_pinger_commit_expected(imp);
1409 cfs_spin_unlock(&imp->imp_lock);
1416 * Helper function to send request \a req over the network for the first time
1417 * Also adjusts request phase.
1418 * Returns 0 on success or error code.
1420 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1422 struct obd_import *imp = req->rq_import;
1426 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1427 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1428 (!req->rq_generation_set ||
1429 req->rq_import_generation == imp->imp_generation))
1432 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1434 cfs_spin_lock(&imp->imp_lock);
1436 if (!req->rq_generation_set)
1437 req->rq_import_generation = imp->imp_generation;
1439 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1440 cfs_spin_lock(&req->rq_lock);
1441 req->rq_waiting = 1;
1442 cfs_spin_unlock(&req->rq_lock);
1444 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1445 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1446 ptlrpc_import_state_name(req->rq_send_state),
1447 ptlrpc_import_state_name(imp->imp_state));
1448 LASSERT(cfs_list_empty(&req->rq_list));
1449 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1450 cfs_atomic_inc(&req->rq_import->imp_inflight);
1451 cfs_spin_unlock(&imp->imp_lock);
1456 cfs_spin_unlock(&imp->imp_lock);
1457 req->rq_status = rc;
1458 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1462 LASSERT(cfs_list_empty(&req->rq_list));
1463 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1464 cfs_atomic_inc(&req->rq_import->imp_inflight);
1465 cfs_spin_unlock(&imp->imp_lock);
1467 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1469 rc = sptlrpc_req_refresh_ctx(req, -1);
1472 req->rq_status = rc;
1475 req->rq_wait_ctx = 1;
1480 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1481 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1482 imp->imp_obd->obd_uuid.uuid,
1483 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1484 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1485 lustre_msg_get_opc(req->rq_reqmsg));
1487 rc = ptl_send_rpc(req, 0);
1489 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1490 req->rq_net_err = 1;
1496 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1501 LASSERT(set->set_producer != NULL);
1503 remaining = cfs_atomic_read(&set->set_remaining);
1505 /* populate the ->set_requests list with requests until we
1506 * reach the maximum number of RPCs in flight for this set */
1507 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1508 rc = set->set_producer(set, set->set_producer_arg);
1509 if (rc == -ENOENT) {
1510 /* no more RPC to produce */
1511 set->set_producer = NULL;
1512 set->set_producer_arg = NULL;
1517 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1521 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1522 * and no more replies are expected.
1523 * (it is possible to get less replies than requests sent e.g. due to timed out
1524 * requests or requests that we had trouble to send out)
1526 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1528 cfs_list_t *tmp, *next;
1529 int force_timer_recalc = 0;
1532 if (cfs_atomic_read(&set->set_remaining) == 0)
1535 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1536 struct ptlrpc_request *req =
1537 cfs_list_entry(tmp, struct ptlrpc_request,
1539 struct obd_import *imp = req->rq_import;
1540 int unregistered = 0;
1543 if (req->rq_phase == RQ_PHASE_NEW &&
1544 ptlrpc_send_new_req(req)) {
1545 force_timer_recalc = 1;
1548 /* delayed send - skip */
1549 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1552 /* delayed resend - skip */
1553 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1554 req->rq_sent > cfs_time_current_sec())
1557 if (!(req->rq_phase == RQ_PHASE_RPC ||
1558 req->rq_phase == RQ_PHASE_BULK ||
1559 req->rq_phase == RQ_PHASE_INTERPRET ||
1560 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1561 req->rq_phase == RQ_PHASE_COMPLETE)) {
1562 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1566 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1567 LASSERT(req->rq_next_phase != req->rq_phase);
1568 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1571 * Skip processing until reply is unlinked. We
1572 * can't return to pool before that and we can't
1573 * call interpret before that. We need to make
1574 * sure that all rdma transfers finished and will
1575 * not corrupt any data.
1577 if (ptlrpc_client_recv_or_unlink(req) ||
1578 ptlrpc_client_bulk_active(req))
1582 * Turn fail_loc off to prevent it from looping
1585 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1586 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1589 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1590 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1595 * Move to next phase if reply was successfully
1598 ptlrpc_rqphase_move(req, req->rq_next_phase);
1601 if (req->rq_phase == RQ_PHASE_COMPLETE)
1604 if (req->rq_phase == RQ_PHASE_INTERPRET)
1605 GOTO(interpret, req->rq_status);
1608 * Note that this also will start async reply unlink.
1610 if (req->rq_net_err && !req->rq_timedout) {
1611 ptlrpc_expire_one_request(req, 1);
1614 * Check if we still need to wait for unlink.
1616 if (ptlrpc_client_recv_or_unlink(req) ||
1617 ptlrpc_client_bulk_active(req))
1619 /* If there is no need to resend, fail it now. */
1620 if (req->rq_no_resend) {
1621 if (req->rq_status == 0)
1622 req->rq_status = -EIO;
1623 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1624 GOTO(interpret, req->rq_status);
1631 cfs_spin_lock(&req->rq_lock);
1632 req->rq_replied = 0;
1633 cfs_spin_unlock(&req->rq_lock);
1634 if (req->rq_status == 0)
1635 req->rq_status = -EIO;
1636 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1637 GOTO(interpret, req->rq_status);
1640 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1641 * so it sets rq_intr regardless of individual rpc
1642 * timeouts. The synchronous IO waiting path sets
1643 * rq_intr irrespective of whether ptlrpcd
1644 * has seen a timeout. Our policy is to only interpret
1645 * interrupted rpcs after they have timed out, so we
1646 * need to enforce that here.
1649 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1650 req->rq_wait_ctx)) {
1651 req->rq_status = -EINTR;
1652 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1653 GOTO(interpret, req->rq_status);
1656 if (req->rq_phase == RQ_PHASE_RPC) {
1657 if (req->rq_timedout || req->rq_resend ||
1658 req->rq_waiting || req->rq_wait_ctx) {
1661 if (!ptlrpc_unregister_reply(req, 1))
1664 cfs_spin_lock(&imp->imp_lock);
1665 if (ptlrpc_import_delay_req(imp, req, &status)){
1666 /* put on delay list - only if we wait
1667 * recovery finished - before send */
1668 cfs_list_del_init(&req->rq_list);
1669 cfs_list_add_tail(&req->rq_list,
1672 cfs_spin_unlock(&imp->imp_lock);
1677 req->rq_status = status;
1678 ptlrpc_rqphase_move(req,
1679 RQ_PHASE_INTERPRET);
1680 cfs_spin_unlock(&imp->imp_lock);
1681 GOTO(interpret, req->rq_status);
1683 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1684 req->rq_status = -ENOTCONN;
1685 ptlrpc_rqphase_move(req,
1686 RQ_PHASE_INTERPRET);
1687 cfs_spin_unlock(&imp->imp_lock);
1688 GOTO(interpret, req->rq_status);
1691 cfs_list_del_init(&req->rq_list);
1692 cfs_list_add_tail(&req->rq_list,
1693 &imp->imp_sending_list);
1695 cfs_spin_unlock(&imp->imp_lock);
1697 cfs_spin_lock(&req->rq_lock);
1698 req->rq_waiting = 0;
1699 cfs_spin_unlock(&req->rq_lock);
1701 if (req->rq_timedout || req->rq_resend) {
1702 /* This is re-sending anyways,
1703 * let's mark req as resend. */
1704 cfs_spin_lock(&req->rq_lock);
1706 cfs_spin_unlock(&req->rq_lock);
1710 if (!ptlrpc_unregister_bulk(req, 1))
1713 /* ensure previous bulk fails */
1714 old_xid = req->rq_xid;
1715 req->rq_xid = ptlrpc_next_xid();
1716 CDEBUG(D_HA, "resend bulk "
1719 old_xid, req->rq_xid);
1723 * rq_wait_ctx is only touched by ptlrpcd,
1724 * so no lock is needed here.
1726 status = sptlrpc_req_refresh_ctx(req, -1);
1729 req->rq_status = status;
1730 cfs_spin_lock(&req->rq_lock);
1731 req->rq_wait_ctx = 0;
1732 cfs_spin_unlock(&req->rq_lock);
1733 force_timer_recalc = 1;
1735 cfs_spin_lock(&req->rq_lock);
1736 req->rq_wait_ctx = 1;
1737 cfs_spin_unlock(&req->rq_lock);
1742 cfs_spin_lock(&req->rq_lock);
1743 req->rq_wait_ctx = 0;
1744 cfs_spin_unlock(&req->rq_lock);
1747 rc = ptl_send_rpc(req, 0);
1749 DEBUG_REQ(D_HA, req, "send failed (%d)",
1751 force_timer_recalc = 1;
1752 cfs_spin_lock(&req->rq_lock);
1753 req->rq_net_err = 1;
1754 cfs_spin_unlock(&req->rq_lock);
1756 /* need to reset the timeout */
1757 force_timer_recalc = 1;
1760 cfs_spin_lock(&req->rq_lock);
1762 if (ptlrpc_client_early(req)) {
1763 ptlrpc_at_recv_early_reply(req);
1764 cfs_spin_unlock(&req->rq_lock);
1768 /* Still waiting for a reply? */
1769 if (ptlrpc_client_recv(req)) {
1770 cfs_spin_unlock(&req->rq_lock);
1774 /* Did we actually receive a reply? */
1775 if (!ptlrpc_client_replied(req)) {
1776 cfs_spin_unlock(&req->rq_lock);
1780 cfs_spin_unlock(&req->rq_lock);
1782 /* unlink from net because we are going to
1783 * swab in-place of reply buffer */
1784 unregistered = ptlrpc_unregister_reply(req, 1);
1788 req->rq_status = after_reply(req);
1792 /* If there is no bulk associated with this request,
1793 * then we're done and should let the interpreter
1794 * process the reply. Similarly if the RPC returned
1795 * an error, and therefore the bulk will never arrive.
1797 if (req->rq_bulk == NULL || req->rq_status < 0) {
1798 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1799 GOTO(interpret, req->rq_status);
1802 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1805 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1806 if (ptlrpc_client_bulk_active(req))
1809 if (!req->rq_bulk->bd_success) {
1810 /* The RPC reply arrived OK, but the bulk screwed
1811 * up! Dead weird since the server told us the RPC
1812 * was good after getting the REPLY for her GET or
1813 * the ACK for her PUT. */
1814 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1815 req->rq_status = -EIO;
1818 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1821 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1823 /* This moves to "unregistering" phase we need to wait for
1825 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1826 /* start async bulk unlink too */
1827 ptlrpc_unregister_bulk(req, 1);
1831 if (!ptlrpc_unregister_bulk(req, 1))
1834 /* When calling interpret receiving already should be
1836 LASSERT(!req->rq_receiving_reply);
1838 ptlrpc_req_interpret(env, req, req->rq_status);
1840 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1842 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1843 "Completed RPC pname:cluuid:pid:xid:nid:"
1844 "opc %s:%s:%d:"LPU64":%s:%d\n",
1845 cfs_curproc_comm(), imp->imp_obd->obd_uuid.uuid,
1846 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1847 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1848 lustre_msg_get_opc(req->rq_reqmsg));
1850 cfs_spin_lock(&imp->imp_lock);
1851 /* Request already may be not on sending or delaying list. This
1852 * may happen in the case of marking it erroneous for the case
1853 * ptlrpc_import_delay_req(req, status) find it impossible to
1854 * allow sending this rpc and returns *status != 0. */
1855 if (!cfs_list_empty(&req->rq_list)) {
1856 cfs_list_del_init(&req->rq_list);
1857 cfs_atomic_dec(&imp->imp_inflight);
1859 cfs_spin_unlock(&imp->imp_lock);
1861 cfs_atomic_dec(&set->set_remaining);
1862 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1864 if (set->set_producer) {
1865 /* produce a new request if possible */
1866 if (ptlrpc_set_producer(set) > 0)
1867 force_timer_recalc = 1;
1869 /* free the request that has just been completed
1870 * in order not to pollute set->set_requests */
1871 cfs_list_del_init(&req->rq_set_chain);
1872 cfs_spin_lock(&req->rq_lock);
1874 req->rq_invalid_rqset = 0;
1875 cfs_spin_unlock(&req->rq_lock);
1877 /* record rq_status to compute the final status later */
1878 if (req->rq_status != 0)
1879 set->set_rc = req->rq_status;
1880 ptlrpc_req_finished(req);
1884 /* If we hit an error, we want to recover promptly. */
1885 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1887 EXPORT_SYMBOL(ptlrpc_check_set);
1890 * Time out request \a req. is \a async_unlink is set, that means do not wait
1891 * until LNet actually confirms network buffer unlinking.
1892 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1894 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1896 struct obd_import *imp = req->rq_import;
1900 cfs_spin_lock(&req->rq_lock);
1901 req->rq_timedout = 1;
1902 cfs_spin_unlock(&req->rq_lock);
1904 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req, "Request "
1905 " sent has %s: [sent "CFS_DURATION_T"/"
1906 "real "CFS_DURATION_T"]",
1907 req->rq_net_err ? "failed due to network error" :
1908 ((req->rq_real_sent == 0 ||
1909 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1910 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1911 "timed out for sent delay" : "timed out for slow reply"),
1912 req->rq_sent, req->rq_real_sent);
1914 if (imp != NULL && obd_debug_peer_on_timeout)
1915 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1917 ptlrpc_unregister_reply(req, async_unlink);
1918 ptlrpc_unregister_bulk(req, async_unlink);
1920 if (obd_dump_on_timeout)
1921 libcfs_debug_dumplog();
1924 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1931 cfs_atomic_inc(&imp->imp_timeouts);
1933 /* The DLM server doesn't want recovery run on its imports. */
1934 if (imp->imp_dlm_fake)
1937 /* If this request is for recovery or other primordial tasks,
1938 * then error it out here. */
1939 if (req->rq_ctx_init || req->rq_ctx_fini ||
1940 req->rq_send_state != LUSTRE_IMP_FULL ||
1941 imp->imp_obd->obd_no_recov) {
1942 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1943 ptlrpc_import_state_name(req->rq_send_state),
1944 ptlrpc_import_state_name(imp->imp_state));
1945 cfs_spin_lock(&req->rq_lock);
1946 req->rq_status = -ETIMEDOUT;
1948 cfs_spin_unlock(&req->rq_lock);
1952 /* if a request can't be resent we can't wait for an answer after
1954 if (ptlrpc_no_resend(req)) {
1955 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1959 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1965 * Time out all uncompleted requests in request set pointed by \a data
1966 * Callback used when waiting on sets with l_wait_event.
1969 int ptlrpc_expired_set(void *data)
1971 struct ptlrpc_request_set *set = data;
1973 time_t now = cfs_time_current_sec();
1976 LASSERT(set != NULL);
1979 * A timeout expired. See which reqs it applies to...
1981 cfs_list_for_each (tmp, &set->set_requests) {
1982 struct ptlrpc_request *req =
1983 cfs_list_entry(tmp, struct ptlrpc_request,
1986 /* don't expire request waiting for context */
1987 if (req->rq_wait_ctx)
1990 /* Request in-flight? */
1991 if (!((req->rq_phase == RQ_PHASE_RPC &&
1992 !req->rq_waiting && !req->rq_resend) ||
1993 (req->rq_phase == RQ_PHASE_BULK)))
1996 if (req->rq_timedout || /* already dealt with */
1997 req->rq_deadline > now) /* not expired */
2000 /* Deal with this guy. Do it asynchronously to not block
2001 * ptlrpcd thread. */
2002 ptlrpc_expire_one_request(req, 1);
2006 * When waiting for a whole set, we always break out of the
2007 * sleep so we can recalculate the timeout, or enable interrupts
2008 * if everyone's timed out.
2012 EXPORT_SYMBOL(ptlrpc_expired_set);
2015 * Sets rq_intr flag in \a req under spinlock.
2017 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2019 cfs_spin_lock(&req->rq_lock);
2021 cfs_spin_unlock(&req->rq_lock);
2023 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2026 * Interrupts (sets interrupted flag) all uncompleted requests in
2027 * a set \a data. Callback for l_wait_event for interruptible waits.
2029 void ptlrpc_interrupted_set(void *data)
2031 struct ptlrpc_request_set *set = data;
2034 LASSERT(set != NULL);
2035 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2037 cfs_list_for_each(tmp, &set->set_requests) {
2038 struct ptlrpc_request *req =
2039 cfs_list_entry(tmp, struct ptlrpc_request,
2042 if (req->rq_phase != RQ_PHASE_RPC &&
2043 req->rq_phase != RQ_PHASE_UNREGISTERING)
2046 ptlrpc_mark_interrupted(req);
2049 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2052 * Get the smallest timeout in the set; this does NOT set a timeout.
2054 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2057 time_t now = cfs_time_current_sec();
2059 struct ptlrpc_request *req;
2063 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
2065 cfs_list_for_each(tmp, &set->set_requests) {
2066 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2069 * Request in-flight?
2071 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2072 (req->rq_phase == RQ_PHASE_BULK) ||
2073 (req->rq_phase == RQ_PHASE_NEW)))
2077 * Already timed out.
2079 if (req->rq_timedout)
2085 if (req->rq_wait_ctx)
2088 if (req->rq_phase == RQ_PHASE_NEW)
2089 deadline = req->rq_sent;
2090 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2091 deadline = req->rq_sent;
2093 deadline = req->rq_sent + req->rq_timeout;
2095 if (deadline <= now) /* actually expired already */
2096 timeout = 1; /* ASAP */
2097 else if (timeout == 0 || timeout > deadline - now)
2098 timeout = deadline - now;
2102 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2105 * Send all unset request from the set and then wait untill all
2106 * requests in the set complete (either get a reply, timeout, get an
2107 * error or otherwise be interrupted).
2108 * Returns 0 on success or error code otherwise.
2110 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2113 struct ptlrpc_request *req;
2114 struct l_wait_info lwi;
2118 if (set->set_producer)
2119 (void)ptlrpc_set_producer(set);
2121 cfs_list_for_each(tmp, &set->set_requests) {
2122 req = cfs_list_entry(tmp, struct ptlrpc_request,
2124 if (req->rq_phase == RQ_PHASE_NEW)
2125 (void)ptlrpc_send_new_req(req);
2128 if (cfs_list_empty(&set->set_requests))
2132 timeout = ptlrpc_set_next_timeout(set);
2134 /* wait until all complete, interrupted, or an in-flight
2136 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2139 if (timeout == 0 && !cfs_signal_pending())
2141 * No requests are in-flight (ether timed out
2142 * or delayed), so we can allow interrupts.
2143 * We still want to block for a limited time,
2144 * so we allow interrupts during the timeout.
2146 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2148 ptlrpc_interrupted_set, set);
2151 * At least one request is in flight, so no
2152 * interrupts are allowed. Wait until all
2153 * complete, or an in-flight req times out.
2155 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2156 ptlrpc_expired_set, set);
2158 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2160 /* LU-769 - if we ignored the signal because it was already
2161 * pending when we started, we need to handle it now or we risk
2162 * it being ignored forever */
2163 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2164 cfs_signal_pending()) {
2165 cfs_sigset_t blocked_sigs =
2166 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2168 /* In fact we only interrupt for the "fatal" signals
2169 * like SIGINT or SIGKILL. We still ignore less
2170 * important signals since ptlrpc set is not easily
2171 * reentrant from userspace again */
2172 if (cfs_signal_pending())
2173 ptlrpc_interrupted_set(set);
2174 cfs_restore_sigs(blocked_sigs);
2177 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2179 /* -EINTR => all requests have been flagged rq_intr so next
2181 * -ETIMEDOUT => someone timed out. When all reqs have
2182 * timed out, signals are enabled allowing completion with
2184 * I don't really care if we go once more round the loop in
2185 * the error cases -eeb. */
2186 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2187 cfs_list_for_each(tmp, &set->set_requests) {
2188 req = cfs_list_entry(tmp, struct ptlrpc_request,
2190 cfs_spin_lock(&req->rq_lock);
2191 req->rq_invalid_rqset = 1;
2192 cfs_spin_unlock(&req->rq_lock);
2195 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2197 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2199 rc = set->set_rc; /* rq_status of already freed requests if any */
2200 cfs_list_for_each(tmp, &set->set_requests) {
2201 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2203 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2204 if (req->rq_status != 0)
2205 rc = req->rq_status;
2208 if (set->set_interpret != NULL) {
2209 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2211 rc = interpreter (set, set->set_arg, rc);
2213 struct ptlrpc_set_cbdata *cbdata, *n;
2216 cfs_list_for_each_entry_safe(cbdata, n,
2217 &set->set_cblist, psc_item) {
2218 cfs_list_del_init(&cbdata->psc_item);
2219 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2222 OBD_FREE_PTR(cbdata);
2228 EXPORT_SYMBOL(ptlrpc_set_wait);
2231 * Helper fuction for request freeing.
2232 * Called when request count reached zero and request needs to be freed.
2233 * Removes request from all sorts of sending/replay lists it might be on,
2234 * frees network buffers if any are present.
2235 * If \a locked is set, that means caller is already holding import imp_lock
2236 * and so we no longer need to reobtain it (for certain lists manipulations)
2238 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2241 if (request == NULL) {
2246 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2247 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2248 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2249 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2250 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2251 LASSERTF(!request->rq_replay, "req %p\n", request);
2253 req_capsule_fini(&request->rq_pill);
2255 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2256 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2257 if (request->rq_import != NULL) {
2259 cfs_spin_lock(&request->rq_import->imp_lock);
2260 cfs_list_del_init(&request->rq_replay_list);
2262 cfs_spin_unlock(&request->rq_import->imp_lock);
2264 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2266 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2267 DEBUG_REQ(D_ERROR, request,
2268 "freeing request with nonzero refcount");
2272 if (request->rq_repbuf != NULL)
2273 sptlrpc_cli_free_repbuf(request);
2274 if (request->rq_export != NULL) {
2275 class_export_put(request->rq_export);
2276 request->rq_export = NULL;
2278 if (request->rq_import != NULL) {
2279 class_import_put(request->rq_import);
2280 request->rq_import = NULL;
2282 if (request->rq_bulk != NULL)
2283 ptlrpc_free_bulk_pin(request->rq_bulk);
2285 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2286 sptlrpc_cli_free_reqbuf(request);
2288 if (request->rq_cli_ctx)
2289 sptlrpc_req_put_ctx(request, !locked);
2291 if (request->rq_pool)
2292 __ptlrpc_free_req_to_pool(request);
2294 OBD_FREE(request, sizeof(*request));
2298 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2300 * Drop one request reference. Must be called with import imp_lock held.
2301 * When reference count drops to zero, reuqest is freed.
2303 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2305 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2306 (void)__ptlrpc_req_finished(request, 1);
2308 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2312 * Drops one reference count for request \a request.
2313 * \a locked set indicates that caller holds import imp_lock.
2314 * Frees the request whe reference count reaches zero.
2316 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2319 if (request == NULL)
2322 if (request == LP_POISON ||
2323 request->rq_reqmsg == LP_POISON) {
2324 CERROR("dereferencing freed request (bug 575)\n");
2329 DEBUG_REQ(D_INFO, request, "refcount now %u",
2330 cfs_atomic_read(&request->rq_refcount) - 1);
2332 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2333 __ptlrpc_free_req(request, locked);
2341 * Drops one reference count for a request.
2343 void ptlrpc_req_finished(struct ptlrpc_request *request)
2345 __ptlrpc_req_finished(request, 0);
2347 EXPORT_SYMBOL(ptlrpc_req_finished);
2350 * Returns xid of a \a request
2352 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2354 return request->rq_xid;
2356 EXPORT_SYMBOL(ptlrpc_req_xid);
2359 * Disengage the client's reply buffer from the network
2360 * NB does _NOT_ unregister any client-side bulk.
2361 * IDEMPOTENT, but _not_ safe against concurrent callers.
2362 * The request owner (i.e. the thread doing the I/O) must call...
2363 * Returns 0 on success or 1 if unregistering cannot be made.
2365 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2369 struct l_wait_info lwi;
2374 LASSERT(!cfs_in_interrupt());
2377 * Let's setup deadline for reply unlink.
2379 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2380 async && request->rq_reply_deadline == 0)
2381 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2384 * Nothing left to do.
2386 if (!ptlrpc_client_recv_or_unlink(request))
2389 LNetMDUnlink(request->rq_reply_md_h);
2392 * Let's check it once again.
2394 if (!ptlrpc_client_recv_or_unlink(request))
2398 * Move to "Unregistering" phase as reply was not unlinked yet.
2400 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2403 * Do not wait for unlink to finish.
2409 * We have to l_wait_event() whatever the result, to give liblustre
2410 * a chance to run reply_in_callback(), and to make sure we've
2411 * unlinked before returning a req to the pool.
2413 if (request->rq_set != NULL)
2414 wq = &request->rq_set->set_waitq;
2416 wq = &request->rq_reply_waitq;
2419 /* Network access will complete in finite time but the HUGE
2420 * timeout lets us CWARN for visibility of sluggish NALs */
2421 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2422 cfs_time_seconds(1), NULL, NULL);
2423 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2426 ptlrpc_rqphase_move(request, request->rq_next_phase);
2430 LASSERT(rc == -ETIMEDOUT);
2431 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2432 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2433 request->rq_must_unlink);
2437 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2440 * Iterates through replay_list on import and prunes
2441 * all requests have transno smaller than last_committed for the
2442 * import and don't have rq_replay set.
2443 * Since requests are sorted in transno order, stops when meetign first
2444 * transno bigger than last_committed.
2445 * caller must hold imp->imp_lock
2447 void ptlrpc_free_committed(struct obd_import *imp)
2449 cfs_list_t *tmp, *saved;
2450 struct ptlrpc_request *req;
2451 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2454 LASSERT(imp != NULL);
2456 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2459 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2460 imp->imp_generation == imp->imp_last_generation_checked) {
2461 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2462 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2466 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2467 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2468 imp->imp_generation);
2469 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2470 imp->imp_last_generation_checked = imp->imp_generation;
2472 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2473 req = cfs_list_entry(tmp, struct ptlrpc_request,
2476 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2477 LASSERT(req != last_req);
2480 if (req->rq_transno == 0) {
2481 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2484 if (req->rq_import_generation < imp->imp_generation) {
2485 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2489 if (req->rq_replay) {
2490 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2494 /* not yet committed */
2495 if (req->rq_transno > imp->imp_peer_committed_transno) {
2496 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2500 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2501 imp->imp_peer_committed_transno);
2503 cfs_spin_lock(&req->rq_lock);
2505 cfs_spin_unlock(&req->rq_lock);
2506 if (req->rq_commit_cb != NULL)
2507 req->rq_commit_cb(req);
2508 cfs_list_del_init(&req->rq_replay_list);
2509 __ptlrpc_req_finished(req, 1);
2516 void ptlrpc_cleanup_client(struct obd_import *imp)
2522 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2525 * Schedule previously sent request for resend.
2526 * For bulk requests we assign new xid (to avoid problems with
2527 * lost replies and therefore several transfers landing into same buffer
2528 * from different sending attempts).
2530 void ptlrpc_resend_req(struct ptlrpc_request *req)
2532 DEBUG_REQ(D_HA, req, "going to resend");
2533 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2534 req->rq_status = -EAGAIN;
2536 cfs_spin_lock(&req->rq_lock);
2538 req->rq_net_err = 0;
2539 req->rq_timedout = 0;
2541 __u64 old_xid = req->rq_xid;
2543 /* ensure previous bulk fails */
2544 req->rq_xid = ptlrpc_next_xid();
2545 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2546 old_xid, req->rq_xid);
2548 ptlrpc_client_wake_req(req);
2549 cfs_spin_unlock(&req->rq_lock);
2551 EXPORT_SYMBOL(ptlrpc_resend_req);
2553 /* XXX: this function and rq_status are currently unused */
2554 void ptlrpc_restart_req(struct ptlrpc_request *req)
2556 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2557 req->rq_status = -ERESTARTSYS;
2559 cfs_spin_lock(&req->rq_lock);
2560 req->rq_restart = 1;
2561 req->rq_timedout = 0;
2562 ptlrpc_client_wake_req(req);
2563 cfs_spin_unlock(&req->rq_lock);
2565 EXPORT_SYMBOL(ptlrpc_restart_req);
2568 * Grab additional reference on a request \a req
2570 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2573 cfs_atomic_inc(&req->rq_refcount);
2576 EXPORT_SYMBOL(ptlrpc_request_addref);
2579 * Add a request to import replay_list.
2580 * Must be called under imp_lock
2582 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2583 struct obd_import *imp)
2587 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2589 if (req->rq_transno == 0) {
2590 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2594 /* clear this for new requests that were resent as well
2595 as resent replayed requests. */
2596 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2598 /* don't re-add requests that have been replayed */
2599 if (!cfs_list_empty(&req->rq_replay_list))
2602 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2604 LASSERT(imp->imp_replayable);
2605 /* Balanced in ptlrpc_free_committed, usually. */
2606 ptlrpc_request_addref(req);
2607 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2608 struct ptlrpc_request *iter =
2609 cfs_list_entry(tmp, struct ptlrpc_request,
2612 /* We may have duplicate transnos if we create and then
2613 * open a file, or for closes retained if to match creating
2614 * opens, so use req->rq_xid as a secondary key.
2615 * (See bugs 684, 685, and 428.)
2616 * XXX no longer needed, but all opens need transnos!
2618 if (iter->rq_transno > req->rq_transno)
2621 if (iter->rq_transno == req->rq_transno) {
2622 LASSERT(iter->rq_xid != req->rq_xid);
2623 if (iter->rq_xid > req->rq_xid)
2627 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2631 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2633 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2636 * Send request and wait until it completes.
2637 * Returns request processing status.
2639 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2641 struct ptlrpc_request_set *set;
2645 LASSERT(req->rq_set == NULL);
2646 LASSERT(!req->rq_receiving_reply);
2648 set = ptlrpc_prep_set();
2650 CERROR("Unable to allocate ptlrpc set.");
2654 /* for distributed debugging */
2655 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2657 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2658 ptlrpc_request_addref(req);
2659 ptlrpc_set_add_req(set, req);
2660 rc = ptlrpc_set_wait(set);
2661 ptlrpc_set_destroy(set);
2665 EXPORT_SYMBOL(ptlrpc_queue_wait);
2667 struct ptlrpc_replay_async_args {
2669 int praa_old_status;
2673 * Callback used for replayed requests reply processing.
2674 * In case of succesful reply calls registeresd request replay callback.
2675 * In case of error restart replay process.
2677 static int ptlrpc_replay_interpret(const struct lu_env *env,
2678 struct ptlrpc_request *req,
2679 void * data, int rc)
2681 struct ptlrpc_replay_async_args *aa = data;
2682 struct obd_import *imp = req->rq_import;
2685 cfs_atomic_dec(&imp->imp_replay_inflight);
2687 if (!ptlrpc_client_replied(req)) {
2688 CERROR("request replay timed out, restarting recovery\n");
2689 GOTO(out, rc = -ETIMEDOUT);
2692 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2693 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2694 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2695 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2697 /** VBR: check version failure */
2698 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2699 /** replay was failed due to version mismatch */
2700 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2701 cfs_spin_lock(&imp->imp_lock);
2702 imp->imp_vbr_failed = 1;
2703 imp->imp_no_lock_replay = 1;
2704 cfs_spin_unlock(&imp->imp_lock);
2705 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2707 /** The transno had better not change over replay. */
2708 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2709 lustre_msg_get_transno(req->rq_repmsg) ||
2710 lustre_msg_get_transno(req->rq_repmsg) == 0,
2712 lustre_msg_get_transno(req->rq_reqmsg),
2713 lustre_msg_get_transno(req->rq_repmsg));
2716 cfs_spin_lock(&imp->imp_lock);
2717 /** if replays by version then gap was occur on server, no trust to locks */
2718 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2719 imp->imp_no_lock_replay = 1;
2720 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2721 cfs_spin_unlock(&imp->imp_lock);
2722 LASSERT(imp->imp_last_replay_transno);
2724 /* transaction number shouldn't be bigger than the latest replayed */
2725 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2726 DEBUG_REQ(D_ERROR, req,
2727 "Reported transno "LPU64" is bigger than the "
2728 "replayed one: "LPU64, req->rq_transno,
2729 lustre_msg_get_transno(req->rq_reqmsg));
2730 GOTO(out, rc = -EINVAL);
2733 DEBUG_REQ(D_HA, req, "got rep");
2735 /* let the callback do fixups, possibly including in the request */
2736 if (req->rq_replay_cb)
2737 req->rq_replay_cb(req);
2739 if (ptlrpc_client_replied(req) &&
2740 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2741 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2742 lustre_msg_get_status(req->rq_repmsg),
2743 aa->praa_old_status);
2745 /* Put it back for re-replay. */
2746 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2750 * Errors while replay can set transno to 0, but
2751 * imp_last_replay_transno shouldn't be set to 0 anyway
2753 if (req->rq_transno == 0)
2754 CERROR("Transno is 0 during replay!\n");
2756 /* continue with recovery */
2757 rc = ptlrpc_import_recovery_state_machine(imp);
2759 req->rq_send_state = aa->praa_old_state;
2762 /* this replay failed, so restart recovery */
2763 ptlrpc_connect_import(imp);
2769 * Prepares and queues request for replay.
2770 * Adds it to ptlrpcd queue for actual sending.
2771 * Returns 0 on success.
2773 int ptlrpc_replay_req(struct ptlrpc_request *req)
2775 struct ptlrpc_replay_async_args *aa;
2778 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2780 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2781 aa = ptlrpc_req_async_args(req);
2782 memset(aa, 0, sizeof *aa);
2784 /* Prepare request to be resent with ptlrpcd */
2785 aa->praa_old_state = req->rq_send_state;
2786 req->rq_send_state = LUSTRE_IMP_REPLAY;
2787 req->rq_phase = RQ_PHASE_NEW;
2788 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2790 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2792 req->rq_interpret_reply = ptlrpc_replay_interpret;
2793 /* Readjust the timeout for current conditions */
2794 ptlrpc_at_set_req_timeout(req);
2796 /* Tell server the net_latency, so the server can calculate how long
2797 * it should wait for next replay */
2798 lustre_msg_set_service_time(req->rq_reqmsg,
2799 ptlrpc_at_get_net_latency(req));
2800 DEBUG_REQ(D_HA, req, "REPLAY");
2802 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2803 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2805 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2808 EXPORT_SYMBOL(ptlrpc_replay_req);
2811 * Aborts all in-flight request on import \a imp sending and delayed lists
2813 void ptlrpc_abort_inflight(struct obd_import *imp)
2815 cfs_list_t *tmp, *n;
2818 /* Make sure that no new requests get processed for this import.
2819 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2820 * this flag and then putting requests on sending_list or delayed_list.
2822 cfs_spin_lock(&imp->imp_lock);
2824 /* XXX locking? Maybe we should remove each request with the list
2825 * locked? Also, how do we know if the requests on the list are
2826 * being freed at this time?
2828 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2829 struct ptlrpc_request *req =
2830 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2832 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2834 cfs_spin_lock (&req->rq_lock);
2835 if (req->rq_import_generation < imp->imp_generation) {
2837 req->rq_status = -EIO;
2838 ptlrpc_client_wake_req(req);
2840 cfs_spin_unlock (&req->rq_lock);
2843 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2844 struct ptlrpc_request *req =
2845 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2847 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2849 cfs_spin_lock (&req->rq_lock);
2850 if (req->rq_import_generation < imp->imp_generation) {
2852 req->rq_status = -EIO;
2853 ptlrpc_client_wake_req(req);
2855 cfs_spin_unlock (&req->rq_lock);
2858 /* Last chance to free reqs left on the replay list, but we
2859 * will still leak reqs that haven't committed. */
2860 if (imp->imp_replayable)
2861 ptlrpc_free_committed(imp);
2863 cfs_spin_unlock(&imp->imp_lock);
2867 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2870 * Abort all uncompleted requests in request set \a set
2872 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2874 cfs_list_t *tmp, *pos;
2876 LASSERT(set != NULL);
2878 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2879 struct ptlrpc_request *req =
2880 cfs_list_entry(pos, struct ptlrpc_request,
2883 cfs_spin_lock(&req->rq_lock);
2884 if (req->rq_phase != RQ_PHASE_RPC) {
2885 cfs_spin_unlock(&req->rq_lock);
2890 req->rq_status = -EINTR;
2891 ptlrpc_client_wake_req(req);
2892 cfs_spin_unlock(&req->rq_lock);
2896 static __u64 ptlrpc_last_xid;
2897 static cfs_spinlock_t ptlrpc_last_xid_lock;
2900 * Initialize the XID for the node. This is common among all requests on
2901 * this node, and only requires the property that it is monotonically
2902 * increasing. It does not need to be sequential. Since this is also used
2903 * as the RDMA match bits, it is important that a single client NOT have
2904 * the same match bits for two different in-flight requests, hence we do
2905 * NOT want to have an XID per target or similar.
2907 * To avoid an unlikely collision between match bits after a client reboot
2908 * (which would deliver old data into the wrong RDMA buffer) initialize
2909 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2910 * If the time is clearly incorrect, we instead use a 62-bit random number.
2911 * In the worst case the random number will overflow 1M RPCs per second in
2912 * 9133 years, or permutations thereof.
2914 #define YEAR_2004 (1ULL << 30)
2915 void ptlrpc_init_xid(void)
2917 time_t now = cfs_time_current_sec();
2919 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2920 if (now < YEAR_2004) {
2921 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2922 ptlrpc_last_xid >>= 2;
2923 ptlrpc_last_xid |= (1ULL << 61);
2925 ptlrpc_last_xid = (__u64)now << 20;
2930 * Increase xid and returns resultng new value to the caller.
2932 __u64 ptlrpc_next_xid(void)
2935 cfs_spin_lock(&ptlrpc_last_xid_lock);
2936 tmp = ++ptlrpc_last_xid;
2937 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2940 EXPORT_SYMBOL(ptlrpc_next_xid);
2943 * Get a glimpse at what next xid value might have been.
2944 * Returns possible next xid.
2946 __u64 ptlrpc_sample_next_xid(void)
2948 #if BITS_PER_LONG == 32
2949 /* need to avoid possible word tearing on 32-bit systems */
2951 cfs_spin_lock(&ptlrpc_last_xid_lock);
2952 tmp = ptlrpc_last_xid + 1;
2953 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2956 /* No need to lock, since returned value is racy anyways */
2957 return ptlrpc_last_xid + 1;
2960 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2963 * Functions for operating ptlrpc workers.
2965 * A ptlrpc work is a function which will be running inside ptlrpc context.
2966 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2968 * 1. after a work is created, it can be used many times, that is:
2969 * handler = ptlrpcd_alloc_work();
2970 * ptlrpcd_queue_work();
2972 * queue it again when necessary:
2973 * ptlrpcd_queue_work();
2974 * ptlrpcd_destroy_work();
2975 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2976 * it will only be queued once in any time. Also as its name implies, it may
2977 * have delay before it really runs by ptlrpcd thread.
2979 struct ptlrpc_work_async_args {
2981 int (*cb)(const struct lu_env *, void *);
2985 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2987 static int work_interpreter(const struct lu_env *env,
2988 struct ptlrpc_request *req, void *data, int rc)
2990 struct ptlrpc_work_async_args *arg = data;
2992 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2993 LASSERT(arg->cb != NULL);
2995 return arg->cb(env, arg->cbdata);
2999 * Create a work for ptlrpc.
3001 void *ptlrpcd_alloc_work(struct obd_import *imp,
3002 int (*cb)(const struct lu_env *, void *), void *cbdata)
3004 struct ptlrpc_request *req = NULL;
3005 struct ptlrpc_work_async_args *args;
3011 RETURN(ERR_PTR(-EINVAL));
3013 /* copy some code from deprecated fakereq. */
3016 CERROR("ptlrpc: run out of memory!\n");
3017 RETURN(ERR_PTR(-ENOMEM));
3020 req->rq_send_state = LUSTRE_IMP_FULL;
3021 req->rq_type = PTL_RPC_MSG_REQUEST;
3022 req->rq_import = class_import_get(imp);
3023 req->rq_export = NULL;
3024 req->rq_interpret_reply = work_interpreter;
3025 /* don't want reply */
3026 req->rq_receiving_reply = 0;
3027 req->rq_must_unlink = 0;
3028 req->rq_no_delay = req->rq_no_resend = 1;
3030 cfs_spin_lock_init(&req->rq_lock);
3031 CFS_INIT_LIST_HEAD(&req->rq_list);
3032 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
3033 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
3034 CFS_INIT_LIST_HEAD(&req->rq_history_list);
3035 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
3036 cfs_waitq_init(&req->rq_reply_waitq);
3037 cfs_waitq_init(&req->rq_set_waitq);
3038 cfs_atomic_set(&req->rq_refcount, 1);
3040 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3041 args = ptlrpc_req_async_args(req);
3042 args->magic = PTLRPC_WORK_MAGIC;
3044 args->cbdata = cbdata;
3048 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3050 void ptlrpcd_destroy_work(void *handler)
3052 struct ptlrpc_request *req = handler;
3055 ptlrpc_req_finished(req);
3057 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3059 int ptlrpcd_queue_work(void *handler)
3061 struct ptlrpc_request *req = handler;
3064 * Check if the req is already being queued.
3066 * Here comes a trick: it lacks a way of checking if a req is being
3067 * processed reliably in ptlrpc. Here I have to use refcount of req
3068 * for this purpose. This is okay because the caller should use this
3069 * req as opaque data. - Jinshan
3071 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
3072 if (cfs_atomic_read(&req->rq_refcount) > 1)
3075 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
3076 cfs_atomic_dec(&req->rq_refcount);
3080 /* re-initialize the req */
3081 req->rq_timeout = obd_timeout;
3082 req->rq_sent = cfs_time_current_sec();
3083 req->rq_deadline = req->rq_sent + req->rq_timeout;
3084 req->rq_reply_deadline = req->rq_deadline;
3085 req->rq_phase = RQ_PHASE_INTERPRET;
3086 req->rq_next_phase = RQ_PHASE_COMPLETE;
3087 req->rq_xid = ptlrpc_next_xid();
3088 req->rq_import_generation = req->rq_import->imp_generation;
3090 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3093 EXPORT_SYMBOL(ptlrpcd_queue_work);