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 and initialize new request set structure.
814 * Returns a pointer to the newly allocated set structure or NULL on error.
816 struct ptlrpc_request_set *ptlrpc_prep_set(void)
818 struct ptlrpc_request_set *set;
821 OBD_ALLOC(set, sizeof *set);
824 cfs_atomic_set(&set->set_refcount, 1);
825 CFS_INIT_LIST_HEAD(&set->set_requests);
826 cfs_waitq_init(&set->set_waitq);
827 cfs_atomic_set(&set->set_new_count, 0);
828 cfs_atomic_set(&set->set_remaining, 0);
829 cfs_spin_lock_init(&set->set_new_req_lock);
830 CFS_INIT_LIST_HEAD(&set->set_new_requests);
831 CFS_INIT_LIST_HEAD(&set->set_cblist);
832 set->set_max_inflight = UINT_MAX;
833 set->set_producer = NULL;
834 set->set_producer_arg = NULL;
839 EXPORT_SYMBOL(ptlrpc_prep_set);
842 * Allocate and initialize new request set structure with flow control
843 * extension. This extension allows to control the number of requests in-flight
844 * for the whole set. A callback function to generate requests must be provided
845 * and the request set will keep the number of requests sent over the wire to
847 * Returns a pointer to the newly allocated set structure or NULL on error.
849 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
853 struct ptlrpc_request_set *set;
855 set = ptlrpc_prep_set();
859 set->set_max_inflight = max;
860 set->set_producer = func;
861 set->set_producer_arg = arg;
865 EXPORT_SYMBOL(ptlrpc_prep_fcset);
868 * Wind down and free request set structure previously allocated with
870 * Ensures that all requests on the set have completed and removes
871 * all requests from the request list in a set.
872 * If any unsent request happen to be on the list, pretends that they got
873 * an error in flight and calls their completion handler.
875 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
883 /* Requests on the set should either all be completed, or all be new */
884 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
885 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
886 cfs_list_for_each (tmp, &set->set_requests) {
887 struct ptlrpc_request *req =
888 cfs_list_entry(tmp, struct ptlrpc_request,
891 LASSERT(req->rq_phase == expected_phase);
895 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
896 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
897 cfs_atomic_read(&set->set_remaining), n);
899 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
900 struct ptlrpc_request *req =
901 cfs_list_entry(tmp, struct ptlrpc_request,
903 cfs_list_del_init(&req->rq_set_chain);
905 LASSERT(req->rq_phase == expected_phase);
907 if (req->rq_phase == RQ_PHASE_NEW) {
908 ptlrpc_req_interpret(NULL, req, -EBADR);
909 cfs_atomic_dec(&set->set_remaining);
912 cfs_spin_lock(&req->rq_lock);
914 req->rq_invalid_rqset = 0;
915 cfs_spin_unlock(&req->rq_lock);
917 ptlrpc_req_finished (req);
920 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
922 ptlrpc_reqset_put(set);
925 EXPORT_SYMBOL(ptlrpc_set_destroy);
928 * Add a callback function \a fn to the set.
929 * This function would be called when all requests on this set are completed.
930 * The function will be passed \a data argument.
932 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
933 set_interpreter_func fn, void *data)
935 struct ptlrpc_set_cbdata *cbdata;
937 OBD_ALLOC_PTR(cbdata);
941 cbdata->psc_interpret = fn;
942 cbdata->psc_data = data;
943 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
947 EXPORT_SYMBOL(ptlrpc_set_add_cb);
950 * Add a new request to the general purpose request set.
951 * Assumes request reference from the caller.
953 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
954 struct ptlrpc_request *req)
956 LASSERT(cfs_list_empty(&req->rq_set_chain));
958 /* The set takes over the caller's request reference */
959 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
961 cfs_atomic_inc(&set->set_remaining);
962 req->rq_queued_time = cfs_time_current();
964 if (req->rq_reqmsg != NULL)
965 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
967 if (set->set_producer != NULL)
968 /* If the request set has a producer callback, the RPC must be
969 * sent straight away */
970 ptlrpc_send_new_req(req);
972 EXPORT_SYMBOL(ptlrpc_set_add_req);
975 * Add a request to a request with dedicated server thread
976 * and wake the thread to make any necessary processing.
977 * Currently only used for ptlrpcd.
979 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
980 struct ptlrpc_request *req)
982 struct ptlrpc_request_set *set = pc->pc_set;
985 LASSERT(req->rq_set == NULL);
986 LASSERT(cfs_test_bit(LIOD_STOP, &pc->pc_flags) == 0);
988 cfs_spin_lock(&set->set_new_req_lock);
990 * The set takes over the caller's request reference.
993 req->rq_queued_time = cfs_time_current();
994 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
995 count = cfs_atomic_inc_return(&set->set_new_count);
996 cfs_spin_unlock(&set->set_new_req_lock);
998 /* Only need to call wakeup once for the first entry. */
1000 cfs_waitq_signal(&set->set_waitq);
1002 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1003 * guarantee the async RPC can be processed ASAP, we have
1004 * no other better choice. It maybe fixed in future. */
1005 for (i = 0; i < pc->pc_npartners; i++)
1006 cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
1009 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1012 * Based on the current state of the import, determine if the request
1013 * can be sent, is an error, or should be delayed.
1015 * Returns true if this request should be delayed. If false, and
1016 * *status is set, then the request can not be sent and *status is the
1017 * error code. If false and status is 0, then request can be sent.
1019 * The imp->imp_lock must be held.
1021 static int ptlrpc_import_delay_req(struct obd_import *imp,
1022 struct ptlrpc_request *req, int *status)
1027 LASSERT (status != NULL);
1030 if (req->rq_ctx_init || req->rq_ctx_fini) {
1031 /* always allow ctx init/fini rpc go through */
1032 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1033 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1035 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1036 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1038 } else if (ptlrpc_send_limit_expired(req)) {
1039 /* probably doesn't need to be a D_ERROR after initial testing */
1040 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1042 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1043 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1044 /* allow CONNECT even if import is invalid */ ;
1045 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1046 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1049 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1050 if (!imp->imp_deactive)
1051 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1052 *status = -ESHUTDOWN; /* bz 12940 */
1053 } else if (req->rq_import_generation != imp->imp_generation) {
1054 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1056 } else if (req->rq_send_state != imp->imp_state) {
1057 /* invalidate in progress - any requests should be drop */
1058 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1059 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1061 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1062 *status = -EWOULDBLOCK;
1072 * Decide if the eror message regarding provided request \a req
1073 * should be printed to the console or not.
1074 * Makes it's decision on request status and other properties.
1075 * Returns 1 to print error on the system console or 0 if not.
1077 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1082 LASSERT(req->rq_reqmsg != NULL);
1083 opc = lustre_msg_get_opc(req->rq_reqmsg);
1085 /* Suppress particular reconnect errors which are to be expected. No
1086 * errors are suppressed for the initial connection on an import */
1087 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1088 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1090 /* Suppress timed out reconnect requests */
1091 if (req->rq_timedout)
1094 /* Suppress unavailable/again reconnect requests */
1095 err = lustre_msg_get_status(req->rq_repmsg);
1096 if (err == -ENODEV || err == -EAGAIN)
1104 * Check request processing status.
1105 * Returns the status.
1107 static int ptlrpc_check_status(struct ptlrpc_request *req)
1112 err = lustre_msg_get_status(req->rq_repmsg);
1113 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1114 struct obd_import *imp = req->rq_import;
1115 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1116 if (ptlrpc_console_allow(req))
1117 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1118 "communicating with %s. The %s "
1119 "operation failed with %d\n",
1121 imp->imp_connection->c_peer.nid),
1122 ll_opcode2str(opc), err);
1123 RETURN(err < 0 ? err : -EINVAL);
1127 DEBUG_REQ(D_INFO, req, "status is %d", err);
1128 } else if (err > 0) {
1129 /* XXX: translate this error from net to host */
1130 DEBUG_REQ(D_INFO, req, "status is %d", err);
1137 * save pre-versions of objects into request for replay.
1138 * Versions are obtained from server reply.
1141 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1143 struct lustre_msg *repmsg = req->rq_repmsg;
1144 struct lustre_msg *reqmsg = req->rq_reqmsg;
1145 __u64 *versions = lustre_msg_get_versions(repmsg);
1148 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1152 lustre_msg_set_versions(reqmsg, versions);
1153 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1154 versions[0], versions[1]);
1160 * Callback function called when client receives RPC reply for \a req.
1161 * Returns 0 on success or error code.
1162 * The return alue would be assigned to req->rq_status by the caller
1163 * as request processing status.
1164 * This function also decides if the request needs to be saved for later replay.
1166 static int after_reply(struct ptlrpc_request *req)
1168 struct obd_import *imp = req->rq_import;
1169 struct obd_device *obd = req->rq_import->imp_obd;
1171 struct timeval work_start;
1175 LASSERT(obd != NULL);
1176 /* repbuf must be unlinked */
1177 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1179 if (req->rq_reply_truncate) {
1180 if (ptlrpc_no_resend(req)) {
1181 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1182 " expected: %d, actual size: %d",
1183 req->rq_nob_received, req->rq_repbuf_len);
1187 sptlrpc_cli_free_repbuf(req);
1188 /* Pass the required reply buffer size (include
1189 * space for early reply).
1190 * NB: no need to roundup because alloc_repbuf
1191 * will roundup it */
1192 req->rq_replen = req->rq_nob_received;
1193 req->rq_nob_received = 0;
1199 * NB Until this point, the whole of the incoming message,
1200 * including buflens, status etc is in the sender's byte order.
1202 rc = sptlrpc_cli_unwrap_reply(req);
1204 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1208 /* retry indefinitely on EINPROGRESS */
1209 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1210 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1211 time_t now = cfs_time_current_sec();
1213 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1215 req->rq_nr_resend++;
1217 /* Readjust the timeout for current conditions */
1218 ptlrpc_at_set_req_timeout(req);
1219 /* delay resend to give a chance to the server to get ready.
1220 * The delay is increased by 1s on every resend and is capped to
1221 * the current request timeout (i.e. obd_timeout if AT is off,
1222 * or AT service time x 125% + 5s, see at_est2timeout) */
1223 if (req->rq_nr_resend > req->rq_timeout)
1224 req->rq_sent = now + req->rq_timeout;
1226 req->rq_sent = now + req->rq_nr_resend;
1230 * Security layer unwrap might ask resend this request.
1235 rc = unpack_reply(req);
1239 cfs_gettimeofday(&work_start);
1240 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1241 if (obd->obd_svc_stats != NULL) {
1242 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1244 ptlrpc_lprocfs_rpc_sent(req, timediff);
1247 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1248 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1249 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1250 lustre_msg_get_type(req->rq_repmsg));
1254 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1255 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1256 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1257 ptlrpc_at_adj_net_latency(req,
1258 lustre_msg_get_service_time(req->rq_repmsg));
1260 rc = ptlrpc_check_status(req);
1261 imp->imp_connect_error = rc;
1265 * Either we've been evicted, or the server has failed for
1266 * some reason. Try to reconnect, and if that fails, punt to
1269 if (ll_rpc_recoverable_error(rc)) {
1270 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1271 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1274 ptlrpc_request_handle_notconn(req);
1279 * Let's look if server sent slv. Do it only for RPC with
1282 ldlm_cli_update_pool(req);
1286 * Store transno in reqmsg for replay.
1288 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1289 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1290 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1293 if (imp->imp_replayable) {
1294 cfs_spin_lock(&imp->imp_lock);
1296 * No point in adding already-committed requests to the replay
1297 * list, we will just remove them immediately. b=9829
1299 if (req->rq_transno != 0 &&
1301 lustre_msg_get_last_committed(req->rq_repmsg) ||
1303 /** version recovery */
1304 ptlrpc_save_versions(req);
1305 ptlrpc_retain_replayable_request(req, imp);
1306 } else if (req->rq_commit_cb != NULL) {
1307 cfs_spin_unlock(&imp->imp_lock);
1308 req->rq_commit_cb(req);
1309 cfs_spin_lock(&imp->imp_lock);
1313 * Replay-enabled imports return commit-status information.
1315 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1316 imp->imp_peer_committed_transno =
1317 lustre_msg_get_last_committed(req->rq_repmsg);
1319 ptlrpc_free_committed(imp);
1321 if (req->rq_transno > imp->imp_peer_committed_transno)
1322 ptlrpc_pinger_commit_expected(imp);
1324 cfs_spin_unlock(&imp->imp_lock);
1331 * Helper function to send request \a req over the network for the first time
1332 * Also adjusts request phase.
1333 * Returns 0 on success or error code.
1335 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1337 struct obd_import *imp = req->rq_import;
1341 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1342 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1343 (!req->rq_generation_set ||
1344 req->rq_import_generation == imp->imp_generation))
1347 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1349 cfs_spin_lock(&imp->imp_lock);
1351 if (!req->rq_generation_set)
1352 req->rq_import_generation = imp->imp_generation;
1354 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1355 cfs_spin_lock(&req->rq_lock);
1356 req->rq_waiting = 1;
1357 cfs_spin_unlock(&req->rq_lock);
1359 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1360 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1361 ptlrpc_import_state_name(req->rq_send_state),
1362 ptlrpc_import_state_name(imp->imp_state));
1363 LASSERT(cfs_list_empty(&req->rq_list));
1364 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1365 cfs_atomic_inc(&req->rq_import->imp_inflight);
1366 cfs_spin_unlock(&imp->imp_lock);
1371 cfs_spin_unlock(&imp->imp_lock);
1372 req->rq_status = rc;
1373 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1377 LASSERT(cfs_list_empty(&req->rq_list));
1378 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1379 cfs_atomic_inc(&req->rq_import->imp_inflight);
1380 cfs_spin_unlock(&imp->imp_lock);
1382 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1384 rc = sptlrpc_req_refresh_ctx(req, -1);
1387 req->rq_status = rc;
1390 req->rq_wait_ctx = 1;
1395 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1396 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1397 imp->imp_obd->obd_uuid.uuid,
1398 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1399 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1400 lustre_msg_get_opc(req->rq_reqmsg));
1402 rc = ptl_send_rpc(req, 0);
1404 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1405 req->rq_net_err = 1;
1411 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1416 LASSERT(set->set_producer != NULL);
1418 remaining = cfs_atomic_read(&set->set_remaining);
1420 /* populate the ->set_requests list with requests until we
1421 * reach the maximum number of RPCs in flight for this set */
1422 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1423 rc = set->set_producer(set, set->set_producer_arg);
1424 if (rc == -ENOENT) {
1425 /* no more RPC to produce */
1426 set->set_producer = NULL;
1427 set->set_producer_arg = NULL;
1432 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1436 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1437 * and no more replies are expected.
1438 * (it is possible to get less replies than requests sent e.g. due to timed out
1439 * requests or requests that we had trouble to send out)
1441 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1443 cfs_list_t *tmp, *next;
1444 int force_timer_recalc = 0;
1447 if (cfs_atomic_read(&set->set_remaining) == 0)
1450 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1451 struct ptlrpc_request *req =
1452 cfs_list_entry(tmp, struct ptlrpc_request,
1454 struct obd_import *imp = req->rq_import;
1455 int unregistered = 0;
1458 if (req->rq_phase == RQ_PHASE_NEW &&
1459 ptlrpc_send_new_req(req)) {
1460 force_timer_recalc = 1;
1463 /* delayed send - skip */
1464 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1467 /* delayed resend - skip */
1468 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1469 req->rq_sent > cfs_time_current_sec())
1472 if (!(req->rq_phase == RQ_PHASE_RPC ||
1473 req->rq_phase == RQ_PHASE_BULK ||
1474 req->rq_phase == RQ_PHASE_INTERPRET ||
1475 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1476 req->rq_phase == RQ_PHASE_COMPLETE)) {
1477 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1481 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1482 LASSERT(req->rq_next_phase != req->rq_phase);
1483 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1486 * Skip processing until reply is unlinked. We
1487 * can't return to pool before that and we can't
1488 * call interpret before that. We need to make
1489 * sure that all rdma transfers finished and will
1490 * not corrupt any data.
1492 if (ptlrpc_client_recv_or_unlink(req) ||
1493 ptlrpc_client_bulk_active(req))
1497 * Turn fail_loc off to prevent it from looping
1500 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1501 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1504 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1505 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1510 * Move to next phase if reply was successfully
1513 ptlrpc_rqphase_move(req, req->rq_next_phase);
1516 if (req->rq_phase == RQ_PHASE_COMPLETE)
1519 if (req->rq_phase == RQ_PHASE_INTERPRET)
1520 GOTO(interpret, req->rq_status);
1523 * Note that this also will start async reply unlink.
1525 if (req->rq_net_err && !req->rq_timedout) {
1526 ptlrpc_expire_one_request(req, 1);
1529 * Check if we still need to wait for unlink.
1531 if (ptlrpc_client_recv_or_unlink(req) ||
1532 ptlrpc_client_bulk_active(req))
1534 /* If there is no need to resend, fail it now. */
1535 if (req->rq_no_resend) {
1536 if (req->rq_status == 0)
1537 req->rq_status = -EIO;
1538 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1539 GOTO(interpret, req->rq_status);
1546 cfs_spin_lock(&req->rq_lock);
1547 req->rq_replied = 0;
1548 cfs_spin_unlock(&req->rq_lock);
1549 if (req->rq_status == 0)
1550 req->rq_status = -EIO;
1551 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1552 GOTO(interpret, req->rq_status);
1555 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1556 * so it sets rq_intr regardless of individual rpc
1557 * timeouts. The synchronous IO waiting path sets
1558 * rq_intr irrespective of whether ptlrpcd
1559 * has seen a timeout. Our policy is to only interpret
1560 * interrupted rpcs after they have timed out, so we
1561 * need to enforce that here.
1564 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1565 req->rq_wait_ctx)) {
1566 req->rq_status = -EINTR;
1567 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1568 GOTO(interpret, req->rq_status);
1571 if (req->rq_phase == RQ_PHASE_RPC) {
1572 if (req->rq_timedout || req->rq_resend ||
1573 req->rq_waiting || req->rq_wait_ctx) {
1576 if (!ptlrpc_unregister_reply(req, 1))
1579 cfs_spin_lock(&imp->imp_lock);
1580 if (ptlrpc_import_delay_req(imp, req, &status)){
1581 /* put on delay list - only if we wait
1582 * recovery finished - before send */
1583 cfs_list_del_init(&req->rq_list);
1584 cfs_list_add_tail(&req->rq_list,
1587 cfs_spin_unlock(&imp->imp_lock);
1592 req->rq_status = status;
1593 ptlrpc_rqphase_move(req,
1594 RQ_PHASE_INTERPRET);
1595 cfs_spin_unlock(&imp->imp_lock);
1596 GOTO(interpret, req->rq_status);
1598 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1599 req->rq_status = -ENOTCONN;
1600 ptlrpc_rqphase_move(req,
1601 RQ_PHASE_INTERPRET);
1602 cfs_spin_unlock(&imp->imp_lock);
1603 GOTO(interpret, req->rq_status);
1606 cfs_list_del_init(&req->rq_list);
1607 cfs_list_add_tail(&req->rq_list,
1608 &imp->imp_sending_list);
1610 cfs_spin_unlock(&imp->imp_lock);
1612 cfs_spin_lock(&req->rq_lock);
1613 req->rq_waiting = 0;
1614 cfs_spin_unlock(&req->rq_lock);
1616 if (req->rq_timedout || req->rq_resend) {
1617 /* This is re-sending anyways,
1618 * let's mark req as resend. */
1619 cfs_spin_lock(&req->rq_lock);
1621 cfs_spin_unlock(&req->rq_lock);
1625 if (!ptlrpc_unregister_bulk(req, 1))
1628 /* ensure previous bulk fails */
1629 old_xid = req->rq_xid;
1630 req->rq_xid = ptlrpc_next_xid();
1631 CDEBUG(D_HA, "resend bulk "
1634 old_xid, req->rq_xid);
1638 * rq_wait_ctx is only touched by ptlrpcd,
1639 * so no lock is needed here.
1641 status = sptlrpc_req_refresh_ctx(req, -1);
1644 req->rq_status = status;
1645 cfs_spin_lock(&req->rq_lock);
1646 req->rq_wait_ctx = 0;
1647 cfs_spin_unlock(&req->rq_lock);
1648 force_timer_recalc = 1;
1650 cfs_spin_lock(&req->rq_lock);
1651 req->rq_wait_ctx = 1;
1652 cfs_spin_unlock(&req->rq_lock);
1657 cfs_spin_lock(&req->rq_lock);
1658 req->rq_wait_ctx = 0;
1659 cfs_spin_unlock(&req->rq_lock);
1662 rc = ptl_send_rpc(req, 0);
1664 DEBUG_REQ(D_HA, req, "send failed (%d)",
1666 force_timer_recalc = 1;
1667 cfs_spin_lock(&req->rq_lock);
1668 req->rq_net_err = 1;
1669 cfs_spin_unlock(&req->rq_lock);
1671 /* need to reset the timeout */
1672 force_timer_recalc = 1;
1675 cfs_spin_lock(&req->rq_lock);
1677 if (ptlrpc_client_early(req)) {
1678 ptlrpc_at_recv_early_reply(req);
1679 cfs_spin_unlock(&req->rq_lock);
1683 /* Still waiting for a reply? */
1684 if (ptlrpc_client_recv(req)) {
1685 cfs_spin_unlock(&req->rq_lock);
1689 /* Did we actually receive a reply? */
1690 if (!ptlrpc_client_replied(req)) {
1691 cfs_spin_unlock(&req->rq_lock);
1695 cfs_spin_unlock(&req->rq_lock);
1697 /* unlink from net because we are going to
1698 * swab in-place of reply buffer */
1699 unregistered = ptlrpc_unregister_reply(req, 1);
1703 req->rq_status = after_reply(req);
1707 /* If there is no bulk associated with this request,
1708 * then we're done and should let the interpreter
1709 * process the reply. Similarly if the RPC returned
1710 * an error, and therefore the bulk will never arrive.
1712 if (req->rq_bulk == NULL || req->rq_status < 0) {
1713 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1714 GOTO(interpret, req->rq_status);
1717 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1720 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1721 if (ptlrpc_client_bulk_active(req))
1724 if (!req->rq_bulk->bd_success) {
1725 /* The RPC reply arrived OK, but the bulk screwed
1726 * up! Dead weird since the server told us the RPC
1727 * was good after getting the REPLY for her GET or
1728 * the ACK for her PUT. */
1729 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1730 req->rq_status = -EIO;
1733 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1736 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1738 /* This moves to "unregistering" phase we need to wait for
1740 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1741 /* start async bulk unlink too */
1742 ptlrpc_unregister_bulk(req, 1);
1746 if (!ptlrpc_unregister_bulk(req, 1))
1749 /* When calling interpret receiving already should be
1751 LASSERT(!req->rq_receiving_reply);
1753 ptlrpc_req_interpret(env, req, req->rq_status);
1755 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1757 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1758 "Completed RPC pname:cluuid:pid:xid:nid:"
1759 "opc %s:%s:%d:"LPU64":%s:%d\n",
1760 cfs_curproc_comm(), imp->imp_obd->obd_uuid.uuid,
1761 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1762 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1763 lustre_msg_get_opc(req->rq_reqmsg));
1765 cfs_spin_lock(&imp->imp_lock);
1766 /* Request already may be not on sending or delaying list. This
1767 * may happen in the case of marking it erroneous for the case
1768 * ptlrpc_import_delay_req(req, status) find it impossible to
1769 * allow sending this rpc and returns *status != 0. */
1770 if (!cfs_list_empty(&req->rq_list)) {
1771 cfs_list_del_init(&req->rq_list);
1772 cfs_atomic_dec(&imp->imp_inflight);
1774 cfs_spin_unlock(&imp->imp_lock);
1776 cfs_atomic_dec(&set->set_remaining);
1777 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1779 if (set->set_producer) {
1780 /* produce a new request if possible */
1781 if (ptlrpc_set_producer(set) > 0)
1782 force_timer_recalc = 1;
1784 /* free the request that has just been completed
1785 * in order not to pollute set->set_requests */
1786 cfs_list_del_init(&req->rq_set_chain);
1787 cfs_spin_lock(&req->rq_lock);
1789 req->rq_invalid_rqset = 0;
1790 cfs_spin_unlock(&req->rq_lock);
1792 /* record rq_status to compute the final status later */
1793 if (req->rq_status != 0)
1794 set->set_rc = req->rq_status;
1795 ptlrpc_req_finished(req);
1799 /* If we hit an error, we want to recover promptly. */
1800 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1802 EXPORT_SYMBOL(ptlrpc_check_set);
1805 * Time out request \a req. is \a async_unlink is set, that means do not wait
1806 * until LNet actually confirms network buffer unlinking.
1807 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1809 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1811 struct obd_import *imp = req->rq_import;
1815 cfs_spin_lock(&req->rq_lock);
1816 req->rq_timedout = 1;
1817 cfs_spin_unlock(&req->rq_lock);
1819 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1820 "/real "CFS_DURATION_T"]",
1821 req->rq_net_err ? "failed due to network error" :
1822 ((req->rq_real_sent == 0 ||
1823 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1824 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1825 "timed out for sent delay" : "timed out for slow reply"),
1826 req->rq_sent, req->rq_real_sent);
1828 if (imp != NULL && obd_debug_peer_on_timeout)
1829 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1831 ptlrpc_unregister_reply(req, async_unlink);
1832 ptlrpc_unregister_bulk(req, async_unlink);
1834 if (obd_dump_on_timeout)
1835 libcfs_debug_dumplog();
1838 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1842 cfs_atomic_inc(&imp->imp_timeouts);
1844 /* The DLM server doesn't want recovery run on its imports. */
1845 if (imp->imp_dlm_fake)
1848 /* If this request is for recovery or other primordial tasks,
1849 * then error it out here. */
1850 if (req->rq_ctx_init || req->rq_ctx_fini ||
1851 req->rq_send_state != LUSTRE_IMP_FULL ||
1852 imp->imp_obd->obd_no_recov) {
1853 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1854 ptlrpc_import_state_name(req->rq_send_state),
1855 ptlrpc_import_state_name(imp->imp_state));
1856 cfs_spin_lock(&req->rq_lock);
1857 req->rq_status = -ETIMEDOUT;
1859 cfs_spin_unlock(&req->rq_lock);
1863 /* if a request can't be resent we can't wait for an answer after
1865 if (ptlrpc_no_resend(req)) {
1866 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1870 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1876 * Time out all uncompleted requests in request set pointed by \a data
1877 * Callback used when waiting on sets with l_wait_event.
1880 int ptlrpc_expired_set(void *data)
1882 struct ptlrpc_request_set *set = data;
1884 time_t now = cfs_time_current_sec();
1887 LASSERT(set != NULL);
1890 * A timeout expired. See which reqs it applies to...
1892 cfs_list_for_each (tmp, &set->set_requests) {
1893 struct ptlrpc_request *req =
1894 cfs_list_entry(tmp, struct ptlrpc_request,
1897 /* don't expire request waiting for context */
1898 if (req->rq_wait_ctx)
1901 /* Request in-flight? */
1902 if (!((req->rq_phase == RQ_PHASE_RPC &&
1903 !req->rq_waiting && !req->rq_resend) ||
1904 (req->rq_phase == RQ_PHASE_BULK)))
1907 if (req->rq_timedout || /* already dealt with */
1908 req->rq_deadline > now) /* not expired */
1911 /* Deal with this guy. Do it asynchronously to not block
1912 * ptlrpcd thread. */
1913 ptlrpc_expire_one_request(req, 1);
1917 * When waiting for a whole set, we always break out of the
1918 * sleep so we can recalculate the timeout, or enable interrupts
1919 * if everyone's timed out.
1923 EXPORT_SYMBOL(ptlrpc_expired_set);
1926 * Sets rq_intr flag in \a req under spinlock.
1928 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1930 cfs_spin_lock(&req->rq_lock);
1932 cfs_spin_unlock(&req->rq_lock);
1934 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
1937 * Interrupts (sets interrupted flag) all uncompleted requests in
1938 * a set \a data. Callback for l_wait_event for interruptible waits.
1940 void ptlrpc_interrupted_set(void *data)
1942 struct ptlrpc_request_set *set = data;
1945 LASSERT(set != NULL);
1946 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
1948 cfs_list_for_each(tmp, &set->set_requests) {
1949 struct ptlrpc_request *req =
1950 cfs_list_entry(tmp, struct ptlrpc_request,
1953 if (req->rq_phase != RQ_PHASE_RPC &&
1954 req->rq_phase != RQ_PHASE_UNREGISTERING)
1957 ptlrpc_mark_interrupted(req);
1960 EXPORT_SYMBOL(ptlrpc_interrupted_set);
1963 * Get the smallest timeout in the set; this does NOT set a timeout.
1965 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1968 time_t now = cfs_time_current_sec();
1970 struct ptlrpc_request *req;
1974 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1976 cfs_list_for_each(tmp, &set->set_requests) {
1977 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1980 * Request in-flight?
1982 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1983 (req->rq_phase == RQ_PHASE_BULK) ||
1984 (req->rq_phase == RQ_PHASE_NEW)))
1988 * Already timed out.
1990 if (req->rq_timedout)
1996 if (req->rq_wait_ctx)
1999 if (req->rq_phase == RQ_PHASE_NEW)
2000 deadline = req->rq_sent;
2001 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2002 deadline = req->rq_sent;
2004 deadline = req->rq_sent + req->rq_timeout;
2006 if (deadline <= now) /* actually expired already */
2007 timeout = 1; /* ASAP */
2008 else if (timeout == 0 || timeout > deadline - now)
2009 timeout = deadline - now;
2013 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2016 * Send all unset request from the set and then wait untill all
2017 * requests in the set complete (either get a reply, timeout, get an
2018 * error or otherwise be interrupted).
2019 * Returns 0 on success or error code otherwise.
2021 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2024 struct ptlrpc_request *req;
2025 struct l_wait_info lwi;
2029 if (set->set_producer)
2030 (void)ptlrpc_set_producer(set);
2032 cfs_list_for_each(tmp, &set->set_requests) {
2033 req = cfs_list_entry(tmp, struct ptlrpc_request,
2035 if (req->rq_phase == RQ_PHASE_NEW)
2036 (void)ptlrpc_send_new_req(req);
2039 if (cfs_list_empty(&set->set_requests))
2043 timeout = ptlrpc_set_next_timeout(set);
2045 /* wait until all complete, interrupted, or an in-flight
2047 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2050 if (timeout == 0 && !cfs_signal_pending())
2052 * No requests are in-flight (ether timed out
2053 * or delayed), so we can allow interrupts.
2054 * We still want to block for a limited time,
2055 * so we allow interrupts during the timeout.
2057 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2059 ptlrpc_interrupted_set, set);
2062 * At least one request is in flight, so no
2063 * interrupts are allowed. Wait until all
2064 * complete, or an in-flight req times out.
2066 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2067 ptlrpc_expired_set, set);
2069 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2071 /* LU-769 - if we ignored the signal because it was already
2072 * pending when we started, we need to handle it now or we risk
2073 * it being ignored forever */
2074 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2075 cfs_signal_pending()) {
2076 cfs_sigset_t blocked_sigs =
2077 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2079 /* In fact we only interrupt for the "fatal" signals
2080 * like SIGINT or SIGKILL. We still ignore less
2081 * important signals since ptlrpc set is not easily
2082 * reentrant from userspace again */
2083 if (cfs_signal_pending())
2084 ptlrpc_interrupted_set(set);
2085 cfs_restore_sigs(blocked_sigs);
2088 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2090 /* -EINTR => all requests have been flagged rq_intr so next
2092 * -ETIMEDOUT => someone timed out. When all reqs have
2093 * timed out, signals are enabled allowing completion with
2095 * I don't really care if we go once more round the loop in
2096 * the error cases -eeb. */
2097 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2098 cfs_list_for_each(tmp, &set->set_requests) {
2099 req = cfs_list_entry(tmp, struct ptlrpc_request,
2101 cfs_spin_lock(&req->rq_lock);
2102 req->rq_invalid_rqset = 1;
2103 cfs_spin_unlock(&req->rq_lock);
2106 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2108 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2110 rc = set->set_rc; /* rq_status of already freed requests if any */
2111 cfs_list_for_each(tmp, &set->set_requests) {
2112 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2114 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2115 if (req->rq_status != 0)
2116 rc = req->rq_status;
2119 if (set->set_interpret != NULL) {
2120 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2122 rc = interpreter (set, set->set_arg, rc);
2124 struct ptlrpc_set_cbdata *cbdata, *n;
2127 cfs_list_for_each_entry_safe(cbdata, n,
2128 &set->set_cblist, psc_item) {
2129 cfs_list_del_init(&cbdata->psc_item);
2130 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2133 OBD_FREE_PTR(cbdata);
2139 EXPORT_SYMBOL(ptlrpc_set_wait);
2142 * Helper fuction for request freeing.
2143 * Called when request count reached zero and request needs to be freed.
2144 * Removes request from all sorts of sending/replay lists it might be on,
2145 * frees network buffers if any are present.
2146 * If \a locked is set, that means caller is already holding import imp_lock
2147 * and so we no longer need to reobtain it (for certain lists manipulations)
2149 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2152 if (request == NULL) {
2157 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2158 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2159 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2160 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2161 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2162 LASSERTF(!request->rq_replay, "req %p\n", request);
2164 req_capsule_fini(&request->rq_pill);
2166 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2167 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2168 if (request->rq_import != NULL) {
2170 cfs_spin_lock(&request->rq_import->imp_lock);
2171 cfs_list_del_init(&request->rq_replay_list);
2173 cfs_spin_unlock(&request->rq_import->imp_lock);
2175 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2177 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2178 DEBUG_REQ(D_ERROR, request,
2179 "freeing request with nonzero refcount");
2183 if (request->rq_repbuf != NULL)
2184 sptlrpc_cli_free_repbuf(request);
2185 if (request->rq_export != NULL) {
2186 class_export_put(request->rq_export);
2187 request->rq_export = NULL;
2189 if (request->rq_import != NULL) {
2190 class_import_put(request->rq_import);
2191 request->rq_import = NULL;
2193 if (request->rq_bulk != NULL)
2194 ptlrpc_free_bulk_pin(request->rq_bulk);
2196 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2197 sptlrpc_cli_free_reqbuf(request);
2199 if (request->rq_cli_ctx)
2200 sptlrpc_req_put_ctx(request, !locked);
2202 if (request->rq_pool)
2203 __ptlrpc_free_req_to_pool(request);
2205 OBD_FREE(request, sizeof(*request));
2209 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2211 * Drop one request reference. Must be called with import imp_lock held.
2212 * When reference count drops to zero, reuqest is freed.
2214 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2216 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2217 (void)__ptlrpc_req_finished(request, 1);
2219 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2223 * Drops one reference count for request \a request.
2224 * \a locked set indicates that caller holds import imp_lock.
2225 * Frees the request whe reference count reaches zero.
2227 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2230 if (request == NULL)
2233 if (request == LP_POISON ||
2234 request->rq_reqmsg == LP_POISON) {
2235 CERROR("dereferencing freed request (bug 575)\n");
2240 DEBUG_REQ(D_INFO, request, "refcount now %u",
2241 cfs_atomic_read(&request->rq_refcount) - 1);
2243 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2244 __ptlrpc_free_req(request, locked);
2252 * Drops one reference count for a request.
2254 void ptlrpc_req_finished(struct ptlrpc_request *request)
2256 __ptlrpc_req_finished(request, 0);
2258 EXPORT_SYMBOL(ptlrpc_req_finished);
2261 * Returns xid of a \a request
2263 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2265 return request->rq_xid;
2267 EXPORT_SYMBOL(ptlrpc_req_xid);
2270 * Disengage the client's reply buffer from the network
2271 * NB does _NOT_ unregister any client-side bulk.
2272 * IDEMPOTENT, but _not_ safe against concurrent callers.
2273 * The request owner (i.e. the thread doing the I/O) must call...
2274 * Returns 0 on success or 1 if unregistering cannot be made.
2276 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2280 struct l_wait_info lwi;
2285 LASSERT(!cfs_in_interrupt());
2288 * Let's setup deadline for reply unlink.
2290 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2291 async && request->rq_reply_deadline == 0)
2292 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2295 * Nothing left to do.
2297 if (!ptlrpc_client_recv_or_unlink(request))
2300 LNetMDUnlink(request->rq_reply_md_h);
2303 * Let's check it once again.
2305 if (!ptlrpc_client_recv_or_unlink(request))
2309 * Move to "Unregistering" phase as reply was not unlinked yet.
2311 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2314 * Do not wait for unlink to finish.
2320 * We have to l_wait_event() whatever the result, to give liblustre
2321 * a chance to run reply_in_callback(), and to make sure we've
2322 * unlinked before returning a req to the pool.
2324 if (request->rq_set != NULL)
2325 wq = &request->rq_set->set_waitq;
2327 wq = &request->rq_reply_waitq;
2330 /* Network access will complete in finite time but the HUGE
2331 * timeout lets us CWARN for visibility of sluggish NALs */
2332 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2333 cfs_time_seconds(1), NULL, NULL);
2334 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2337 ptlrpc_rqphase_move(request, request->rq_next_phase);
2341 LASSERT(rc == -ETIMEDOUT);
2342 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2343 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2344 request->rq_must_unlink);
2348 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2351 * Iterates through replay_list on import and prunes
2352 * all requests have transno smaller than last_committed for the
2353 * import and don't have rq_replay set.
2354 * Since requests are sorted in transno order, stops when meetign first
2355 * transno bigger than last_committed.
2356 * caller must hold imp->imp_lock
2358 void ptlrpc_free_committed(struct obd_import *imp)
2360 cfs_list_t *tmp, *saved;
2361 struct ptlrpc_request *req;
2362 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2365 LASSERT(imp != NULL);
2367 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2370 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2371 imp->imp_generation == imp->imp_last_generation_checked) {
2372 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2373 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2377 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2378 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2379 imp->imp_generation);
2380 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2381 imp->imp_last_generation_checked = imp->imp_generation;
2383 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2384 req = cfs_list_entry(tmp, struct ptlrpc_request,
2387 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2388 LASSERT(req != last_req);
2391 if (req->rq_transno == 0) {
2392 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2395 if (req->rq_import_generation < imp->imp_generation) {
2396 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2400 if (req->rq_replay) {
2401 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2405 /* not yet committed */
2406 if (req->rq_transno > imp->imp_peer_committed_transno) {
2407 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2411 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2412 imp->imp_peer_committed_transno);
2414 cfs_spin_lock(&req->rq_lock);
2416 cfs_spin_unlock(&req->rq_lock);
2417 if (req->rq_commit_cb != NULL)
2418 req->rq_commit_cb(req);
2419 cfs_list_del_init(&req->rq_replay_list);
2420 __ptlrpc_req_finished(req, 1);
2427 void ptlrpc_cleanup_client(struct obd_import *imp)
2433 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2436 * Schedule previously sent request for resend.
2437 * For bulk requests we assign new xid (to avoid problems with
2438 * lost replies and therefore several transfers landing into same buffer
2439 * from different sending attempts).
2441 void ptlrpc_resend_req(struct ptlrpc_request *req)
2443 DEBUG_REQ(D_HA, req, "going to resend");
2444 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2445 req->rq_status = -EAGAIN;
2447 cfs_spin_lock(&req->rq_lock);
2449 req->rq_net_err = 0;
2450 req->rq_timedout = 0;
2452 __u64 old_xid = req->rq_xid;
2454 /* ensure previous bulk fails */
2455 req->rq_xid = ptlrpc_next_xid();
2456 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2457 old_xid, req->rq_xid);
2459 ptlrpc_client_wake_req(req);
2460 cfs_spin_unlock(&req->rq_lock);
2462 EXPORT_SYMBOL(ptlrpc_resend_req);
2464 /* XXX: this function and rq_status are currently unused */
2465 void ptlrpc_restart_req(struct ptlrpc_request *req)
2467 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2468 req->rq_status = -ERESTARTSYS;
2470 cfs_spin_lock(&req->rq_lock);
2471 req->rq_restart = 1;
2472 req->rq_timedout = 0;
2473 ptlrpc_client_wake_req(req);
2474 cfs_spin_unlock(&req->rq_lock);
2476 EXPORT_SYMBOL(ptlrpc_restart_req);
2479 * Grab additional reference on a request \a req
2481 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2484 cfs_atomic_inc(&req->rq_refcount);
2487 EXPORT_SYMBOL(ptlrpc_request_addref);
2490 * Add a request to import replay_list.
2491 * Must be called under imp_lock
2493 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2494 struct obd_import *imp)
2498 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2500 if (req->rq_transno == 0) {
2501 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2505 /* clear this for new requests that were resent as well
2506 as resent replayed requests. */
2507 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2509 /* don't re-add requests that have been replayed */
2510 if (!cfs_list_empty(&req->rq_replay_list))
2513 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2515 LASSERT(imp->imp_replayable);
2516 /* Balanced in ptlrpc_free_committed, usually. */
2517 ptlrpc_request_addref(req);
2518 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2519 struct ptlrpc_request *iter =
2520 cfs_list_entry(tmp, struct ptlrpc_request,
2523 /* We may have duplicate transnos if we create and then
2524 * open a file, or for closes retained if to match creating
2525 * opens, so use req->rq_xid as a secondary key.
2526 * (See bugs 684, 685, and 428.)
2527 * XXX no longer needed, but all opens need transnos!
2529 if (iter->rq_transno > req->rq_transno)
2532 if (iter->rq_transno == req->rq_transno) {
2533 LASSERT(iter->rq_xid != req->rq_xid);
2534 if (iter->rq_xid > req->rq_xid)
2538 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2542 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2544 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2547 * Send request and wait until it completes.
2548 * Returns request processing status.
2550 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2552 struct ptlrpc_request_set *set;
2556 LASSERT(req->rq_set == NULL);
2557 LASSERT(!req->rq_receiving_reply);
2559 set = ptlrpc_prep_set();
2561 CERROR("Unable to allocate ptlrpc set.");
2565 /* for distributed debugging */
2566 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2568 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2569 ptlrpc_request_addref(req);
2570 ptlrpc_set_add_req(set, req);
2571 rc = ptlrpc_set_wait(set);
2572 ptlrpc_set_destroy(set);
2576 EXPORT_SYMBOL(ptlrpc_queue_wait);
2578 struct ptlrpc_replay_async_args {
2580 int praa_old_status;
2584 * Callback used for replayed requests reply processing.
2585 * In case of succesful reply calls registeresd request replay callback.
2586 * In case of error restart replay process.
2588 static int ptlrpc_replay_interpret(const struct lu_env *env,
2589 struct ptlrpc_request *req,
2590 void * data, int rc)
2592 struct ptlrpc_replay_async_args *aa = data;
2593 struct obd_import *imp = req->rq_import;
2596 cfs_atomic_dec(&imp->imp_replay_inflight);
2598 if (!ptlrpc_client_replied(req)) {
2599 CERROR("request replay timed out, restarting recovery\n");
2600 GOTO(out, rc = -ETIMEDOUT);
2603 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2604 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2605 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2606 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2608 /** VBR: check version failure */
2609 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2610 /** replay was failed due to version mismatch */
2611 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2612 cfs_spin_lock(&imp->imp_lock);
2613 imp->imp_vbr_failed = 1;
2614 imp->imp_no_lock_replay = 1;
2615 cfs_spin_unlock(&imp->imp_lock);
2616 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2618 /** The transno had better not change over replay. */
2619 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2620 lustre_msg_get_transno(req->rq_repmsg) ||
2621 lustre_msg_get_transno(req->rq_repmsg) == 0,
2623 lustre_msg_get_transno(req->rq_reqmsg),
2624 lustre_msg_get_transno(req->rq_repmsg));
2627 cfs_spin_lock(&imp->imp_lock);
2628 /** if replays by version then gap was occur on server, no trust to locks */
2629 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2630 imp->imp_no_lock_replay = 1;
2631 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2632 cfs_spin_unlock(&imp->imp_lock);
2633 LASSERT(imp->imp_last_replay_transno);
2635 /* transaction number shouldn't be bigger than the latest replayed */
2636 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2637 DEBUG_REQ(D_ERROR, req,
2638 "Reported transno "LPU64" is bigger than the "
2639 "replayed one: "LPU64, req->rq_transno,
2640 lustre_msg_get_transno(req->rq_reqmsg));
2641 GOTO(out, rc = -EINVAL);
2644 DEBUG_REQ(D_HA, req, "got rep");
2646 /* let the callback do fixups, possibly including in the request */
2647 if (req->rq_replay_cb)
2648 req->rq_replay_cb(req);
2650 if (ptlrpc_client_replied(req) &&
2651 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2652 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2653 lustre_msg_get_status(req->rq_repmsg),
2654 aa->praa_old_status);
2656 /* Put it back for re-replay. */
2657 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2661 * Errors while replay can set transno to 0, but
2662 * imp_last_replay_transno shouldn't be set to 0 anyway
2664 if (req->rq_transno == 0)
2665 CERROR("Transno is 0 during replay!\n");
2667 /* continue with recovery */
2668 rc = ptlrpc_import_recovery_state_machine(imp);
2670 req->rq_send_state = aa->praa_old_state;
2673 /* this replay failed, so restart recovery */
2674 ptlrpc_connect_import(imp);
2680 * Prepares and queues request for replay.
2681 * Adds it to ptlrpcd queue for actual sending.
2682 * Returns 0 on success.
2684 int ptlrpc_replay_req(struct ptlrpc_request *req)
2686 struct ptlrpc_replay_async_args *aa;
2689 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2691 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2692 aa = ptlrpc_req_async_args(req);
2693 memset(aa, 0, sizeof *aa);
2695 /* Prepare request to be resent with ptlrpcd */
2696 aa->praa_old_state = req->rq_send_state;
2697 req->rq_send_state = LUSTRE_IMP_REPLAY;
2698 req->rq_phase = RQ_PHASE_NEW;
2699 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2701 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2703 req->rq_interpret_reply = ptlrpc_replay_interpret;
2704 /* Readjust the timeout for current conditions */
2705 ptlrpc_at_set_req_timeout(req);
2707 /* Tell server the net_latency, so the server can calculate how long
2708 * it should wait for next replay */
2709 lustre_msg_set_service_time(req->rq_reqmsg,
2710 ptlrpc_at_get_net_latency(req));
2711 DEBUG_REQ(D_HA, req, "REPLAY");
2713 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2714 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2716 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2719 EXPORT_SYMBOL(ptlrpc_replay_req);
2722 * Aborts all in-flight request on import \a imp sending and delayed lists
2724 void ptlrpc_abort_inflight(struct obd_import *imp)
2726 cfs_list_t *tmp, *n;
2729 /* Make sure that no new requests get processed for this import.
2730 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2731 * this flag and then putting requests on sending_list or delayed_list.
2733 cfs_spin_lock(&imp->imp_lock);
2735 /* XXX locking? Maybe we should remove each request with the list
2736 * locked? Also, how do we know if the requests on the list are
2737 * being freed at this time?
2739 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2740 struct ptlrpc_request *req =
2741 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2743 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2745 cfs_spin_lock (&req->rq_lock);
2746 if (req->rq_import_generation < imp->imp_generation) {
2748 req->rq_status = -EIO;
2749 ptlrpc_client_wake_req(req);
2751 cfs_spin_unlock (&req->rq_lock);
2754 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2755 struct ptlrpc_request *req =
2756 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2758 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2760 cfs_spin_lock (&req->rq_lock);
2761 if (req->rq_import_generation < imp->imp_generation) {
2763 req->rq_status = -EIO;
2764 ptlrpc_client_wake_req(req);
2766 cfs_spin_unlock (&req->rq_lock);
2769 /* Last chance to free reqs left on the replay list, but we
2770 * will still leak reqs that haven't committed. */
2771 if (imp->imp_replayable)
2772 ptlrpc_free_committed(imp);
2774 cfs_spin_unlock(&imp->imp_lock);
2778 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2781 * Abort all uncompleted requests in request set \a set
2783 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2785 cfs_list_t *tmp, *pos;
2787 LASSERT(set != NULL);
2789 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2790 struct ptlrpc_request *req =
2791 cfs_list_entry(pos, struct ptlrpc_request,
2794 cfs_spin_lock(&req->rq_lock);
2795 if (req->rq_phase != RQ_PHASE_RPC) {
2796 cfs_spin_unlock(&req->rq_lock);
2801 req->rq_status = -EINTR;
2802 ptlrpc_client_wake_req(req);
2803 cfs_spin_unlock(&req->rq_lock);
2807 static __u64 ptlrpc_last_xid;
2808 static cfs_spinlock_t ptlrpc_last_xid_lock;
2811 * Initialize the XID for the node. This is common among all requests on
2812 * this node, and only requires the property that it is monotonically
2813 * increasing. It does not need to be sequential. Since this is also used
2814 * as the RDMA match bits, it is important that a single client NOT have
2815 * the same match bits for two different in-flight requests, hence we do
2816 * NOT want to have an XID per target or similar.
2818 * To avoid an unlikely collision between match bits after a client reboot
2819 * (which would deliver old data into the wrong RDMA buffer) initialize
2820 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2821 * If the time is clearly incorrect, we instead use a 62-bit random number.
2822 * In the worst case the random number will overflow 1M RPCs per second in
2823 * 9133 years, or permutations thereof.
2825 #define YEAR_2004 (1ULL << 30)
2826 void ptlrpc_init_xid(void)
2828 time_t now = cfs_time_current_sec();
2830 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2831 if (now < YEAR_2004) {
2832 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2833 ptlrpc_last_xid >>= 2;
2834 ptlrpc_last_xid |= (1ULL << 61);
2836 ptlrpc_last_xid = (__u64)now << 20;
2841 * Increase xid and returns resultng new value to the caller.
2843 __u64 ptlrpc_next_xid(void)
2846 cfs_spin_lock(&ptlrpc_last_xid_lock);
2847 tmp = ++ptlrpc_last_xid;
2848 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2851 EXPORT_SYMBOL(ptlrpc_next_xid);
2854 * Get a glimpse at what next xid value might have been.
2855 * Returns possible next xid.
2857 __u64 ptlrpc_sample_next_xid(void)
2859 #if BITS_PER_LONG == 32
2860 /* need to avoid possible word tearing on 32-bit systems */
2862 cfs_spin_lock(&ptlrpc_last_xid_lock);
2863 tmp = ptlrpc_last_xid + 1;
2864 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2867 /* No need to lock, since returned value is racy anyways */
2868 return ptlrpc_last_xid + 1;
2871 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2874 * Functions for operating ptlrpc workers.
2876 * A ptlrpc work is a function which will be running inside ptlrpc context.
2877 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2879 * 1. after a work is created, it can be used many times, that is:
2880 * handler = ptlrpcd_alloc_work();
2881 * ptlrpcd_queue_work();
2883 * queue it again when necessary:
2884 * ptlrpcd_queue_work();
2885 * ptlrpcd_destroy_work();
2886 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2887 * it will only be queued once in any time. Also as its name implies, it may
2888 * have delay before it really runs by ptlrpcd thread.
2890 struct ptlrpc_work_async_args {
2892 int (*cb)(const struct lu_env *, void *);
2896 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2898 static int work_interpreter(const struct lu_env *env,
2899 struct ptlrpc_request *req, void *data, int rc)
2901 struct ptlrpc_work_async_args *arg = data;
2903 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2904 LASSERT(arg->cb != NULL);
2906 return arg->cb(env, arg->cbdata);
2910 * Create a work for ptlrpc.
2912 void *ptlrpcd_alloc_work(struct obd_import *imp,
2913 int (*cb)(const struct lu_env *, void *), void *cbdata)
2915 struct ptlrpc_request *req = NULL;
2916 struct ptlrpc_work_async_args *args;
2922 RETURN(ERR_PTR(-EINVAL));
2924 /* copy some code from deprecated fakereq. */
2927 CERROR("ptlrpc: run out of memory!\n");
2928 RETURN(ERR_PTR(-ENOMEM));
2931 req->rq_send_state = LUSTRE_IMP_FULL;
2932 req->rq_type = PTL_RPC_MSG_REQUEST;
2933 req->rq_import = class_import_get(imp);
2934 req->rq_export = NULL;
2935 req->rq_interpret_reply = work_interpreter;
2936 /* don't want reply */
2937 req->rq_receiving_reply = 0;
2938 req->rq_must_unlink = 0;
2939 req->rq_no_delay = req->rq_no_resend = 1;
2941 cfs_spin_lock_init(&req->rq_lock);
2942 CFS_INIT_LIST_HEAD(&req->rq_list);
2943 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
2944 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
2945 CFS_INIT_LIST_HEAD(&req->rq_history_list);
2946 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
2947 cfs_waitq_init(&req->rq_reply_waitq);
2948 cfs_waitq_init(&req->rq_set_waitq);
2949 cfs_atomic_set(&req->rq_refcount, 1);
2951 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
2952 args = ptlrpc_req_async_args(req);
2953 args->magic = PTLRPC_WORK_MAGIC;
2955 args->cbdata = cbdata;
2959 EXPORT_SYMBOL(ptlrpcd_alloc_work);
2961 void ptlrpcd_destroy_work(void *handler)
2963 struct ptlrpc_request *req = handler;
2966 ptlrpc_req_finished(req);
2968 EXPORT_SYMBOL(ptlrpcd_destroy_work);
2970 int ptlrpcd_queue_work(void *handler)
2972 struct ptlrpc_request *req = handler;
2975 * Check if the req is already being queued.
2977 * Here comes a trick: it lacks a way of checking if a req is being
2978 * processed reliably in ptlrpc. Here I have to use refcount of req
2979 * for this purpose. This is okay because the caller should use this
2980 * req as opaque data. - Jinshan
2982 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
2983 if (cfs_atomic_read(&req->rq_refcount) > 1)
2986 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
2987 cfs_atomic_dec(&req->rq_refcount);
2991 /* re-initialize the req */
2992 req->rq_timeout = obd_timeout;
2993 req->rq_sent = cfs_time_current_sec();
2994 req->rq_deadline = req->rq_sent + req->rq_timeout;
2995 req->rq_reply_deadline = req->rq_deadline;
2996 req->rq_phase = RQ_PHASE_INTERPRET;
2997 req->rq_next_phase = RQ_PHASE_COMPLETE;
2998 req->rq_xid = ptlrpc_next_xid();
2999 req->rq_import_generation = req->rq_import->imp_generation;
3001 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3004 EXPORT_SYMBOL(ptlrpcd_queue_work);