1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
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"
56 * Initialize passed in client structure \a cl.
58 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
59 struct ptlrpc_client *cl)
61 cl->cli_request_portal = req_portal;
62 cl->cli_reply_portal = rep_portal;
67 * Return PortalRPC connection for remore uud \a uuid
69 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
71 struct ptlrpc_connection *c;
73 lnet_process_id_t peer;
76 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
78 CDEBUG(D_NETERROR, "cannot find peer %s!\n", uuid->uuid);
82 c = ptlrpc_connection_get(peer, self, uuid);
84 memcpy(c->c_remote_uuid.uuid,
85 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
88 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
94 * Allocate and initialize new bulk descriptor
95 * Returns pointer to the descriptor or NULL on error.
97 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
99 struct ptlrpc_bulk_desc *desc;
101 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
105 cfs_spin_lock_init(&desc->bd_lock);
106 cfs_waitq_init(&desc->bd_waitq);
107 desc->bd_max_iov = npages;
108 desc->bd_iov_count = 0;
109 LNetInvalidateHandle(&desc->bd_md_h);
110 desc->bd_portal = portal;
111 desc->bd_type = type;
117 * Prepare bulk descriptor for specified outgoing request \a req that
118 * can fit \a npages * pages. \a type is bulk type. \a portal is where
119 * the bulk to be sent. Used on client-side.
120 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
123 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
124 int npages, int type, int portal)
126 struct obd_import *imp = req->rq_import;
127 struct ptlrpc_bulk_desc *desc;
130 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
131 desc = new_bulk(npages, type, portal);
135 desc->bd_import_generation = req->rq_import_generation;
136 desc->bd_import = class_import_get(imp);
139 desc->bd_cbid.cbid_fn = client_bulk_callback;
140 desc->bd_cbid.cbid_arg = desc;
142 /* This makes req own desc, and free it when she frees herself */
149 * Prepare bulk descriptor for specified incoming request \a req that
150 * can fit \a npages * pages. \a type is bulk type. \a portal is where
151 * the bulk to be sent. Used on server-side after request was already
153 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
156 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
157 int npages, int type, int portal)
159 struct obd_export *exp = req->rq_export;
160 struct ptlrpc_bulk_desc *desc;
163 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
165 desc = new_bulk(npages, type, portal);
169 desc->bd_export = class_export_get(exp);
172 desc->bd_cbid.cbid_fn = server_bulk_callback;
173 desc->bd_cbid.cbid_arg = desc;
175 /* NB we don't assign rq_bulk here; server-side requests are
176 * re-used, and the handler frees the bulk desc explicitly. */
182 * Add a page \a page to the bulk descriptor \a desc.
183 * Data to transfer in the page starts at offset \a pageoffset and
184 * amount of data to transfer from the page is \a len
186 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
187 cfs_page_t *page, int pageoffset, int len)
189 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
190 LASSERT(page != NULL);
191 LASSERT(pageoffset >= 0);
193 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
197 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
201 * Uninitialize and free bulk descriptor \a desc.
202 * Works on bulk descriptors both from server and client side.
204 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
208 LASSERT(desc != NULL);
209 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
210 LASSERT(!desc->bd_network_rw); /* network hands off or */
211 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
213 sptlrpc_enc_pool_put_pages(desc);
216 class_export_put(desc->bd_export);
218 class_import_put(desc->bd_import);
220 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
221 bd_iov[desc->bd_max_iov]));
226 * Set server timelimit for this req, i.e. how long are we willing to wait
227 * for reply before timing out this request.
229 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
235 LASSERT(req->rq_import);
238 /* non-AT settings */
240 * \a imp_server_timeout means this is reverse import and
241 * we send (currently only) ASTs to the client and cannot afford
242 * to wait too long for the reply, otherwise the other client
243 * (because of which we are sending this request) would
244 * timeout waiting for us
246 req->rq_timeout = req->rq_import->imp_server_timeout ?
247 obd_timeout / 2 : obd_timeout;
249 at = &req->rq_import->imp_at;
250 idx = import_at_get_index(req->rq_import,
251 req->rq_request_portal);
252 serv_est = at_get(&at->iat_service_estimate[idx]);
253 req->rq_timeout = at_est2timeout(serv_est);
255 /* We could get even fancier here, using history to predict increased
258 /* Let the server know what this RPC timeout is by putting it in the
260 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
263 /* Adjust max service estimate based on server value */
264 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
265 unsigned int serv_est)
271 LASSERT(req->rq_import);
272 at = &req->rq_import->imp_at;
274 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
275 /* max service estimates are tracked on the server side,
276 so just keep minimal history here */
277 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
279 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
280 "has changed from %d to %d\n",
281 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
282 oldse, at_get(&at->iat_service_estimate[idx]));
285 /* Expected network latency per remote node (secs) */
286 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
288 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
291 /* Adjust expected network latency */
292 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
293 unsigned int service_time)
295 unsigned int nl, oldnl;
297 time_t now = cfs_time_current_sec();
299 LASSERT(req->rq_import);
300 at = &req->rq_import->imp_at;
302 /* Network latency is total time less server processing time */
303 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
304 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
305 CWARN("Reported service time %u > total measured time "
306 CFS_DURATION_T"\n", service_time,
307 cfs_time_sub(now, req->rq_sent));
309 oldnl = at_measured(&at->iat_net_latency, nl);
311 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
312 "has changed from %d to %d\n",
313 req->rq_import->imp_obd->obd_name,
315 &req->rq_import->imp_connection->c_remote_uuid),
316 oldnl, at_get(&at->iat_net_latency));
319 static int unpack_reply(struct ptlrpc_request *req)
323 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
324 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
326 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
331 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
333 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
340 * Handle an early reply message, called with the rq_lock held.
341 * If anything goes wrong just ignore it - same as if it never happened
343 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
345 struct ptlrpc_request *early_req;
351 cfs_spin_unlock(&req->rq_lock);
353 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
355 cfs_spin_lock(&req->rq_lock);
359 rc = unpack_reply(early_req);
361 /* Expecting to increase the service time estimate here */
362 ptlrpc_at_adj_service(req,
363 lustre_msg_get_timeout(early_req->rq_repmsg));
364 ptlrpc_at_adj_net_latency(req,
365 lustre_msg_get_service_time(early_req->rq_repmsg));
368 sptlrpc_cli_finish_early_reply(early_req);
370 cfs_spin_lock(&req->rq_lock);
373 /* Adjust the local timeout for this req */
374 ptlrpc_at_set_req_timeout(req);
376 olddl = req->rq_deadline;
377 /* server assumes it now has rq_timeout from when it sent the
378 early reply, so client should give it at least that long. */
379 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
380 ptlrpc_at_get_net_latency(req);
382 DEBUG_REQ(D_ADAPTTO, req,
383 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
384 "("CFS_DURATION_T"s)", req->rq_early_count,
385 cfs_time_sub(req->rq_deadline,
386 cfs_time_current_sec()),
387 cfs_time_sub(req->rq_deadline, olddl));
394 * Wind down request pool \a pool.
395 * Frees all requests from the pool too
397 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
400 struct ptlrpc_request *req;
402 LASSERT(pool != NULL);
404 cfs_spin_lock(&pool->prp_lock);
405 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
406 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
407 cfs_list_del(&req->rq_list);
408 LASSERT(req->rq_reqbuf);
409 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
410 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
411 OBD_FREE(req, sizeof(*req));
413 cfs_spin_unlock(&pool->prp_lock);
414 OBD_FREE(pool, sizeof(*pool));
418 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
420 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
425 while (size < pool->prp_rq_size)
428 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
429 size == pool->prp_rq_size,
430 "Trying to change pool size with nonempty pool "
431 "from %d to %d bytes\n", pool->prp_rq_size, size);
433 cfs_spin_lock(&pool->prp_lock);
434 pool->prp_rq_size = size;
435 for (i = 0; i < num_rq; i++) {
436 struct ptlrpc_request *req;
437 struct lustre_msg *msg;
439 cfs_spin_unlock(&pool->prp_lock);
440 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
443 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
445 OBD_FREE(req, sizeof(struct ptlrpc_request));
448 req->rq_reqbuf = msg;
449 req->rq_reqbuf_len = size;
451 cfs_spin_lock(&pool->prp_lock);
452 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
454 cfs_spin_unlock(&pool->prp_lock);
459 * Create and initialize new request pool with given attributes:
460 * \a num_rq - initial number of requests to create for the pool
461 * \a msgsize - maximum message size possible for requests in thid pool
462 * \a populate_pool - function to be called when more requests need to be added
464 * Returns pointer to newly created pool or NULL on error.
466 struct ptlrpc_request_pool *
467 ptlrpc_init_rq_pool(int num_rq, int msgsize,
468 void (*populate_pool)(struct ptlrpc_request_pool *, int))
470 struct ptlrpc_request_pool *pool;
472 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
476 /* Request next power of two for the allocation, because internally
477 kernel would do exactly this */
479 cfs_spin_lock_init(&pool->prp_lock);
480 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
481 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
482 pool->prp_populate = populate_pool;
484 populate_pool(pool, num_rq);
486 if (cfs_list_empty(&pool->prp_req_list)) {
487 /* have not allocated a single request for the pool */
488 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
495 * Fetches one request from pool \a pool
497 static struct ptlrpc_request *
498 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
500 struct ptlrpc_request *request;
501 struct lustre_msg *reqbuf;
506 cfs_spin_lock(&pool->prp_lock);
508 /* See if we have anything in a pool, and bail out if nothing,
509 * in writeout path, where this matters, this is safe to do, because
510 * nothing is lost in this case, and when some in-flight requests
511 * complete, this code will be called again. */
512 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
513 cfs_spin_unlock(&pool->prp_lock);
517 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
519 cfs_list_del_init(&request->rq_list);
520 cfs_spin_unlock(&pool->prp_lock);
522 LASSERT(request->rq_reqbuf);
523 LASSERT(request->rq_pool);
525 reqbuf = request->rq_reqbuf;
526 memset(request, 0, sizeof(*request));
527 request->rq_reqbuf = reqbuf;
528 request->rq_reqbuf_len = pool->prp_rq_size;
529 request->rq_pool = pool;
535 * Returns freed \a request to pool.
537 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
539 struct ptlrpc_request_pool *pool = request->rq_pool;
541 cfs_spin_lock(&pool->prp_lock);
542 LASSERT(cfs_list_empty(&request->rq_list));
543 LASSERT(!request->rq_receiving_reply);
544 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
545 cfs_spin_unlock(&pool->prp_lock);
548 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
549 __u32 version, int opcode,
550 int count, __u32 *lengths, char **bufs,
551 struct ptlrpc_cli_ctx *ctx)
553 struct obd_import *imp = request->rq_import;
558 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
560 rc = sptlrpc_req_get_ctx(request);
565 sptlrpc_req_set_flavor(request, opcode);
567 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
570 LASSERT(!request->rq_pool);
574 lustre_msg_add_version(request->rq_reqmsg, version);
575 request->rq_send_state = LUSTRE_IMP_FULL;
576 request->rq_type = PTL_RPC_MSG_REQUEST;
577 request->rq_export = NULL;
579 request->rq_req_cbid.cbid_fn = request_out_callback;
580 request->rq_req_cbid.cbid_arg = request;
582 request->rq_reply_cbid.cbid_fn = reply_in_callback;
583 request->rq_reply_cbid.cbid_arg = request;
585 request->rq_reply_deadline = 0;
586 request->rq_phase = RQ_PHASE_NEW;
587 request->rq_next_phase = RQ_PHASE_UNDEFINED;
589 request->rq_request_portal = imp->imp_client->cli_request_portal;
590 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
592 ptlrpc_at_set_req_timeout(request);
594 cfs_spin_lock_init(&request->rq_lock);
595 CFS_INIT_LIST_HEAD(&request->rq_list);
596 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
597 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
598 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
599 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
600 CFS_INIT_LIST_HEAD(&request->rq_history_list);
601 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
602 cfs_waitq_init(&request->rq_reply_waitq);
603 cfs_waitq_init(&request->rq_set_waitq);
604 request->rq_xid = ptlrpc_next_xid();
605 cfs_atomic_set(&request->rq_refcount, 1);
607 lustre_msg_set_opc(request->rq_reqmsg, opcode);
611 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
613 class_import_put(imp);
617 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
618 __u32 version, int opcode, char **bufs,
619 struct ptlrpc_cli_ctx *ctx)
623 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
624 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
625 request->rq_pill.rc_area[RCL_CLIENT],
628 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
631 * Pack request buffers for network transfer, performing necessary encryption
632 * steps if necessary.
634 int ptlrpc_request_pack(struct ptlrpc_request *request,
635 __u32 version, int opcode)
637 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
641 * Helper function to allocate new request on import \a imp
642 * and possibly using existing request from pool \a pool if provided.
643 * Returns allocated request structure with import field filled or
647 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
648 struct ptlrpc_request_pool *pool)
650 struct ptlrpc_request *request = NULL;
653 request = ptlrpc_prep_req_from_pool(pool);
656 OBD_ALLOC_PTR(request);
659 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
660 LASSERT(imp != LP_POISON);
661 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
663 LASSERT(imp->imp_client != LP_POISON);
665 request->rq_import = class_import_get(imp);
667 CERROR("request allocation out of memory\n");
674 * Helper function for creating a request.
675 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
676 * buffer structures according to capsule template \a format.
677 * Returns allocated request structure pointer or NULL on error.
679 static struct ptlrpc_request *
680 ptlrpc_request_alloc_internal(struct obd_import *imp,
681 struct ptlrpc_request_pool * pool,
682 const struct req_format *format)
684 struct ptlrpc_request *request;
686 request = __ptlrpc_request_alloc(imp, pool);
690 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
691 req_capsule_set(&request->rq_pill, format);
696 * Allocate new request structure for import \a imp and initialize its
697 * buffer structure according to capsule template \a format.
699 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
700 const struct req_format *format)
702 return ptlrpc_request_alloc_internal(imp, NULL, format);
706 * Allocate new request structure for import \a imp from pool \a pool and
707 * initialize its buffer structure according to capsule template \a format.
709 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
710 struct ptlrpc_request_pool * pool,
711 const struct req_format *format)
713 return ptlrpc_request_alloc_internal(imp, pool, format);
717 * For requests not from pool, free memory of the request structure.
718 * For requests obtained from a pool earlier, return request back to pool.
720 void ptlrpc_request_free(struct ptlrpc_request *request)
722 if (request->rq_pool)
723 __ptlrpc_free_req_to_pool(request);
725 OBD_FREE_PTR(request);
729 * Allocate new request for operatione \a opcode and immediatelly pack it for
731 * Only used for simple requests like OBD_PING where the only important
732 * part of the request is operation itself.
733 * Returns allocated request or NULL on error.
735 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
736 const struct req_format *format,
737 __u32 version, int opcode)
739 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
743 rc = ptlrpc_request_pack(req, version, opcode);
745 ptlrpc_request_free(req);
753 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
754 * for operation \a opcode. Request would contain \a count buffers.
755 * Sizes of buffers are described in array \a lengths and buffers themselves
756 * are provided by a pointer \a bufs.
757 * Returns prepared request structure pointer or NULL on error.
759 struct ptlrpc_request *
760 ptlrpc_prep_req_pool(struct obd_import *imp,
761 __u32 version, int opcode,
762 int count, __u32 *lengths, char **bufs,
763 struct ptlrpc_request_pool *pool)
765 struct ptlrpc_request *request;
768 request = __ptlrpc_request_alloc(imp, pool);
772 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
773 lengths, bufs, NULL);
775 ptlrpc_request_free(request);
782 * Same as ptlrpc_prep_req_pool, but without pool
784 struct ptlrpc_request *
785 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
786 __u32 *lengths, char **bufs)
788 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
793 * Allocate "fake" request that would not be sent anywhere in the end.
794 * Only used as a hack because we have no other way of performing
795 * async actions in lustre between layers.
796 * Used on MDS to request object preallocations from more than one OST at a
799 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
800 unsigned int timeout,
801 ptlrpc_interpterer_t interpreter)
803 struct ptlrpc_request *request = NULL;
806 OBD_ALLOC(request, sizeof(*request));
808 CERROR("request allocation out of memory\n");
812 request->rq_send_state = LUSTRE_IMP_FULL;
813 request->rq_type = PTL_RPC_MSG_REQUEST;
814 request->rq_import = class_import_get(imp);
815 request->rq_export = NULL;
816 request->rq_import_generation = imp->imp_generation;
818 request->rq_timeout = timeout;
819 request->rq_sent = cfs_time_current_sec();
820 request->rq_deadline = request->rq_sent + timeout;
821 request->rq_reply_deadline = request->rq_deadline;
822 request->rq_interpret_reply = interpreter;
823 request->rq_phase = RQ_PHASE_RPC;
824 request->rq_next_phase = RQ_PHASE_INTERPRET;
825 /* don't want reply */
826 request->rq_receiving_reply = 0;
827 request->rq_must_unlink = 0;
828 request->rq_no_delay = request->rq_no_resend = 1;
829 request->rq_fake = 1;
831 cfs_spin_lock_init(&request->rq_lock);
832 CFS_INIT_LIST_HEAD(&request->rq_list);
833 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
834 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
835 CFS_INIT_LIST_HEAD(&request->rq_history_list);
836 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
837 cfs_waitq_init(&request->rq_reply_waitq);
838 cfs_waitq_init(&request->rq_set_waitq);
840 request->rq_xid = ptlrpc_next_xid();
841 cfs_atomic_set(&request->rq_refcount, 1);
847 * Indicate that processing of "fake" request is finished.
849 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
851 /* if we kill request before timeout - need adjust counter */
852 if (req->rq_phase == RQ_PHASE_RPC) {
853 struct ptlrpc_request_set *set = req->rq_set;
856 cfs_atomic_dec(&set->set_remaining);
859 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
860 cfs_list_del_init(&req->rq_list);
864 * Allocate and initialize new request set structure.
865 * Returns a pointer to the newly allocated set structure or NULL on error.
867 struct ptlrpc_request_set *ptlrpc_prep_set(void)
869 struct ptlrpc_request_set *set;
872 OBD_ALLOC(set, sizeof *set);
875 CFS_INIT_LIST_HEAD(&set->set_requests);
876 cfs_waitq_init(&set->set_waitq);
877 cfs_atomic_set(&set->set_remaining, 0);
878 cfs_spin_lock_init(&set->set_new_req_lock);
879 CFS_INIT_LIST_HEAD(&set->set_new_requests);
880 CFS_INIT_LIST_HEAD(&set->set_cblist);
886 * Wind down and free request set structure previously allocated with
888 * Ensures that all requests on the set have completed and removes
889 * all requests from the request list in a set.
890 * If any unsent request happen to be on the list, pretends that they got
891 * an error in flight and calls their completion handler.
893 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
901 /* Requests on the set should either all be completed, or all be new */
902 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
903 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
904 cfs_list_for_each (tmp, &set->set_requests) {
905 struct ptlrpc_request *req =
906 cfs_list_entry(tmp, struct ptlrpc_request,
909 LASSERT(req->rq_phase == expected_phase);
913 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
914 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
915 cfs_atomic_read(&set->set_remaining), n);
917 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
918 struct ptlrpc_request *req =
919 cfs_list_entry(tmp, struct ptlrpc_request,
921 cfs_list_del_init(&req->rq_set_chain);
923 LASSERT(req->rq_phase == expected_phase);
925 if (req->rq_phase == RQ_PHASE_NEW) {
926 ptlrpc_req_interpret(NULL, req, -EBADR);
927 cfs_atomic_dec(&set->set_remaining);
930 cfs_spin_lock(&req->rq_lock);
932 req->rq_invalid_rqset = 0;
933 cfs_spin_unlock(&req->rq_lock);
935 ptlrpc_req_finished (req);
938 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
940 OBD_FREE(set, sizeof(*set));
945 * Add a callback function \a fn to the set.
946 * This function would be called when all requests on this set are completed.
947 * The function will be passed \a data argument.
949 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
950 set_interpreter_func fn, void *data)
952 struct ptlrpc_set_cbdata *cbdata;
954 OBD_ALLOC_PTR(cbdata);
958 cbdata->psc_interpret = fn;
959 cbdata->psc_data = data;
960 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
966 * Add a new request to the general purpose request set.
967 * Assumes request reference from the caller.
969 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
970 struct ptlrpc_request *req)
972 /* The set takes over the caller's request reference */
973 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
975 cfs_atomic_inc(&set->set_remaining);
976 req->rq_queued_time = cfs_time_current(); /* Where is the best place to set this? */
980 * Add a request to a request with dedicated server thread
981 * and wake the thread to make any necessary processing.
982 * Currently only used for ptlrpcd.
983 * Returns 0 if succesful or non zero error code on error.
984 * (the only possible error for now is if the dedicated server thread
987 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
988 struct ptlrpc_request *req)
990 struct ptlrpc_request_set *set = pc->pc_set;
993 * Let caller know that we stopped and will not handle this request.
994 * It needs to take care itself of request.
996 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags))
999 cfs_spin_lock(&set->set_new_req_lock);
1001 * The set takes over the caller's request reference.
1003 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1005 cfs_spin_unlock(&set->set_new_req_lock);
1007 cfs_waitq_signal(&set->set_waitq);
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.");
1036 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1037 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1039 } else if (ptlrpc_send_limit_expired(req)) {
1040 /* probably doesn't need to be a D_ERROR after initial testing */
1041 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1043 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1044 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1045 /* allow CONNECT even if import is invalid */ ;
1046 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1047 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1050 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1051 if (!imp->imp_deactive)
1052 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1053 *status = -ESHUTDOWN; /* bz 12940 */
1054 } else if (req->rq_import_generation != imp->imp_generation) {
1055 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1057 } else if (req->rq_send_state != imp->imp_state) {
1058 /* invalidate in progress - any requests should be drop */
1059 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1060 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1062 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1063 *status = -EWOULDBLOCK;
1073 * Decide if the eror message regarding provided request \a req
1074 * should be printed to the console or not.
1075 * Makes it's decision on request status and other properties.
1076 * Returns 1 to print error on the system console or 0 if not.
1078 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1080 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1083 /* Suppress particular reconnect errors which are to be expected. No
1084 * errors are suppressed for the initial connection on an import */
1085 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1086 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1088 /* Suppress timed out reconnect requests */
1089 if (req->rq_timedout)
1092 /* Suppress unavailable/again reconnect requests */
1093 err = lustre_msg_get_status(req->rq_repmsg);
1094 if (err == -ENODEV || err == -EAGAIN)
1102 * Check request processing status.
1103 * Returns the status.
1105 static int ptlrpc_check_status(struct ptlrpc_request *req)
1110 err = lustre_msg_get_status(req->rq_repmsg);
1111 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1112 struct obd_import *imp = req->rq_import;
1113 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1114 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
1115 " with %s. The %s operation failed with %d\n",
1116 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1117 ll_opcode2str(opc), err);
1118 RETURN(err < 0 ? err : -EINVAL);
1122 DEBUG_REQ(D_INFO, req, "status is %d", err);
1123 } else if (err > 0) {
1124 /* XXX: translate this error from net to host */
1125 DEBUG_REQ(D_INFO, req, "status is %d", err);
1128 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1129 struct obd_import *imp = req->rq_import;
1130 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1132 if (ptlrpc_console_allow(req))
1133 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1134 "communicating with %s. The %s "
1135 "operation failed with %d\n",
1137 imp->imp_connection->c_peer.nid),
1138 ll_opcode2str(opc), err);
1140 RETURN(err < 0 ? err : -EINVAL);
1147 * save pre-versions of objects into request for replay.
1148 * Versions are obtained from server reply.
1151 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1153 struct lustre_msg *repmsg = req->rq_repmsg;
1154 struct lustre_msg *reqmsg = req->rq_reqmsg;
1155 __u64 *versions = lustre_msg_get_versions(repmsg);
1158 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1162 lustre_msg_set_versions(reqmsg, versions);
1163 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1164 versions[0], versions[1]);
1170 * Callback function called when client receives RPC reply for \a req.
1171 * Returns 0 on success or error code.
1172 * The return alue would be assigned to req->rq_status by the caller
1173 * as request processing status.
1174 * This function also decides if the request needs to be saved for later replay.
1176 static int after_reply(struct ptlrpc_request *req)
1178 struct obd_import *imp = req->rq_import;
1179 struct obd_device *obd = req->rq_import->imp_obd;
1181 struct timeval work_start;
1185 LASSERT(obd != NULL);
1186 /* repbuf must be unlinked */
1187 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1189 if (req->rq_reply_truncate) {
1190 if (ptlrpc_no_resend(req)) {
1191 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1192 " expected: %d, actual size: %d",
1193 req->rq_nob_received, req->rq_repbuf_len);
1197 sptlrpc_cli_free_repbuf(req);
1198 /* Pass the required reply buffer size (include
1199 * space for early reply).
1200 * NB: no need to roundup because alloc_repbuf
1201 * will roundup it */
1202 req->rq_replen = req->rq_nob_received;
1203 req->rq_nob_received = 0;
1209 * NB Until this point, the whole of the incoming message,
1210 * including buflens, status etc is in the sender's byte order.
1212 rc = sptlrpc_cli_unwrap_reply(req);
1214 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1219 * Security layer unwrap might ask resend this request.
1224 rc = unpack_reply(req);
1228 cfs_gettimeofday(&work_start);
1229 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1230 if (obd->obd_svc_stats != NULL) {
1231 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1233 ptlrpc_lprocfs_rpc_sent(req, timediff);
1236 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1237 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1238 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1239 lustre_msg_get_type(req->rq_repmsg));
1243 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1244 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1245 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1246 ptlrpc_at_adj_net_latency(req,
1247 lustre_msg_get_service_time(req->rq_repmsg));
1249 rc = ptlrpc_check_status(req);
1250 imp->imp_connect_error = rc;
1254 * Either we've been evicted, or the server has failed for
1255 * some reason. Try to reconnect, and if that fails, punt to
1258 if (ll_rpc_recoverable_error(rc)) {
1259 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1260 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1263 ptlrpc_request_handle_notconn(req);
1268 * Let's look if server sent slv. Do it only for RPC with
1271 ldlm_cli_update_pool(req);
1275 * Store transno in reqmsg for replay.
1277 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1278 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1279 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1282 if (imp->imp_replayable) {
1283 cfs_spin_lock(&imp->imp_lock);
1285 * No point in adding already-committed requests to the replay
1286 * list, we will just remove them immediately. b=9829
1288 if (req->rq_transno != 0 &&
1290 lustre_msg_get_last_committed(req->rq_repmsg) ||
1292 /** version recovery */
1293 ptlrpc_save_versions(req);
1294 ptlrpc_retain_replayable_request(req, imp);
1295 } else if (req->rq_commit_cb != NULL) {
1296 cfs_spin_unlock(&imp->imp_lock);
1297 req->rq_commit_cb(req);
1298 cfs_spin_lock(&imp->imp_lock);
1302 * Replay-enabled imports return commit-status information.
1304 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1305 imp->imp_peer_committed_transno =
1306 lustre_msg_get_last_committed(req->rq_repmsg);
1308 ptlrpc_free_committed(imp);
1309 cfs_spin_unlock(&imp->imp_lock);
1316 * Helper function to send request \a req over the network for the first time
1317 * Also adjusts request phase.
1318 * Returns 0 on success or error code.
1320 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1322 struct obd_import *imp;
1326 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1327 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1330 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1332 imp = req->rq_import;
1333 cfs_spin_lock(&imp->imp_lock);
1335 req->rq_import_generation = imp->imp_generation;
1337 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1338 cfs_spin_lock(&req->rq_lock);
1339 req->rq_waiting = 1;
1340 cfs_spin_unlock(&req->rq_lock);
1342 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1343 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1344 ptlrpc_import_state_name(req->rq_send_state),
1345 ptlrpc_import_state_name(imp->imp_state));
1346 LASSERT(cfs_list_empty(&req->rq_list));
1347 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1348 cfs_atomic_inc(&req->rq_import->imp_inflight);
1349 cfs_spin_unlock(&imp->imp_lock);
1354 cfs_spin_unlock(&imp->imp_lock);
1355 req->rq_status = rc;
1356 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1360 LASSERT(cfs_list_empty(&req->rq_list));
1361 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1362 cfs_atomic_inc(&req->rq_import->imp_inflight);
1363 cfs_spin_unlock(&imp->imp_lock);
1365 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1367 rc = sptlrpc_req_refresh_ctx(req, -1);
1370 req->rq_status = rc;
1373 req->rq_wait_ctx = 1;
1378 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1379 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1380 imp->imp_obd->obd_uuid.uuid,
1381 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1382 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1383 lustre_msg_get_opc(req->rq_reqmsg));
1385 rc = ptl_send_rpc(req, 0);
1387 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1388 req->rq_net_err = 1;
1395 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1396 * and no more replies are expected.
1397 * (it is possible to get less replies than requests sent e.g. due to timed out
1398 * requests or requests that we had trouble to send out)
1400 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1403 int force_timer_recalc = 0;
1406 if (cfs_atomic_read(&set->set_remaining) == 0)
1409 cfs_list_for_each(tmp, &set->set_requests) {
1410 struct ptlrpc_request *req =
1411 cfs_list_entry(tmp, struct ptlrpc_request,
1413 struct obd_import *imp = req->rq_import;
1414 int unregistered = 0;
1417 if (req->rq_phase == RQ_PHASE_NEW &&
1418 ptlrpc_send_new_req(req)) {
1419 force_timer_recalc = 1;
1422 /* delayed send - skip */
1423 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1426 if (!(req->rq_phase == RQ_PHASE_RPC ||
1427 req->rq_phase == RQ_PHASE_BULK ||
1428 req->rq_phase == RQ_PHASE_INTERPRET ||
1429 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1430 req->rq_phase == RQ_PHASE_COMPLETE)) {
1431 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1435 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1436 LASSERT(req->rq_next_phase != req->rq_phase);
1437 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1440 * Skip processing until reply is unlinked. We
1441 * can't return to pool before that and we can't
1442 * call interpret before that. We need to make
1443 * sure that all rdma transfers finished and will
1444 * not corrupt any data.
1446 if (ptlrpc_client_recv_or_unlink(req) ||
1447 ptlrpc_client_bulk_active(req))
1451 * Turn fail_loc off to prevent it from looping
1454 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1455 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1458 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1459 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1464 * Move to next phase if reply was successfully
1467 ptlrpc_rqphase_move(req, req->rq_next_phase);
1470 if (req->rq_phase == RQ_PHASE_COMPLETE)
1473 if (req->rq_phase == RQ_PHASE_INTERPRET)
1474 GOTO(interpret, req->rq_status);
1477 * Note that this also will start async reply unlink.
1479 if (req->rq_net_err && !req->rq_timedout) {
1480 ptlrpc_expire_one_request(req, 1);
1483 * Check if we still need to wait for unlink.
1485 if (ptlrpc_client_recv_or_unlink(req) ||
1486 ptlrpc_client_bulk_active(req))
1491 cfs_spin_lock(&req->rq_lock);
1492 req->rq_replied = 0;
1493 cfs_spin_unlock(&req->rq_lock);
1494 if (req->rq_status == 0)
1495 req->rq_status = -EIO;
1496 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1497 GOTO(interpret, req->rq_status);
1500 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1501 * so it sets rq_intr regardless of individual rpc
1502 * timeouts. The synchronous IO waiting path sets
1503 * rq_intr irrespective of whether ptlrpcd
1504 * has seen a timeout. Our policy is to only interpret
1505 * interrupted rpcs after they have timed out, so we
1506 * need to enforce that here.
1509 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1510 req->rq_wait_ctx)) {
1511 req->rq_status = -EINTR;
1512 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1513 GOTO(interpret, req->rq_status);
1516 if (req->rq_phase == RQ_PHASE_RPC) {
1517 if (req->rq_timedout || req->rq_resend ||
1518 req->rq_waiting || req->rq_wait_ctx) {
1521 if (!ptlrpc_unregister_reply(req, 1))
1524 cfs_spin_lock(&imp->imp_lock);
1525 if (ptlrpc_import_delay_req(imp, req, &status)){
1526 /* put on delay list - only if we wait
1527 * recovery finished - before send */
1528 cfs_list_del_init(&req->rq_list);
1529 cfs_list_add_tail(&req->rq_list,
1532 cfs_spin_unlock(&imp->imp_lock);
1537 req->rq_status = status;
1538 ptlrpc_rqphase_move(req,
1539 RQ_PHASE_INTERPRET);
1540 cfs_spin_unlock(&imp->imp_lock);
1541 GOTO(interpret, req->rq_status);
1543 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1544 req->rq_status = -ENOTCONN;
1545 ptlrpc_rqphase_move(req,
1546 RQ_PHASE_INTERPRET);
1547 cfs_spin_unlock(&imp->imp_lock);
1548 GOTO(interpret, req->rq_status);
1551 cfs_list_del_init(&req->rq_list);
1552 cfs_list_add_tail(&req->rq_list,
1553 &imp->imp_sending_list);
1555 cfs_spin_unlock(&imp->imp_lock);
1557 cfs_spin_lock(&req->rq_lock);
1558 req->rq_waiting = 0;
1559 cfs_spin_unlock(&req->rq_lock);
1561 if (req->rq_timedout || req->rq_resend) {
1562 /* This is re-sending anyways,
1563 * let's mark req as resend. */
1564 cfs_spin_lock(&req->rq_lock);
1566 cfs_spin_unlock(&req->rq_lock);
1570 if (!ptlrpc_unregister_bulk(req, 1))
1573 /* ensure previous bulk fails */
1574 old_xid = req->rq_xid;
1575 req->rq_xid = ptlrpc_next_xid();
1576 CDEBUG(D_HA, "resend bulk "
1579 old_xid, req->rq_xid);
1583 * rq_wait_ctx is only touched by ptlrpcd,
1584 * so no lock is needed here.
1586 status = sptlrpc_req_refresh_ctx(req, -1);
1589 req->rq_status = status;
1590 cfs_spin_lock(&req->rq_lock);
1591 req->rq_wait_ctx = 0;
1592 cfs_spin_unlock(&req->rq_lock);
1593 force_timer_recalc = 1;
1595 cfs_spin_lock(&req->rq_lock);
1596 req->rq_wait_ctx = 1;
1597 cfs_spin_unlock(&req->rq_lock);
1602 cfs_spin_lock(&req->rq_lock);
1603 req->rq_wait_ctx = 0;
1604 cfs_spin_unlock(&req->rq_lock);
1607 rc = ptl_send_rpc(req, 0);
1609 DEBUG_REQ(D_HA, req, "send failed (%d)",
1611 force_timer_recalc = 1;
1612 cfs_spin_lock(&req->rq_lock);
1613 req->rq_net_err = 1;
1614 cfs_spin_unlock(&req->rq_lock);
1616 /* need to reset the timeout */
1617 force_timer_recalc = 1;
1620 cfs_spin_lock(&req->rq_lock);
1622 if (ptlrpc_client_early(req)) {
1623 ptlrpc_at_recv_early_reply(req);
1624 cfs_spin_unlock(&req->rq_lock);
1628 /* Still waiting for a reply? */
1629 if (ptlrpc_client_recv(req)) {
1630 cfs_spin_unlock(&req->rq_lock);
1634 /* Did we actually receive a reply? */
1635 if (!ptlrpc_client_replied(req)) {
1636 cfs_spin_unlock(&req->rq_lock);
1640 cfs_spin_unlock(&req->rq_lock);
1642 /* unlink from net because we are going to
1643 * swab in-place of reply buffer */
1644 unregistered = ptlrpc_unregister_reply(req, 1);
1648 req->rq_status = after_reply(req);
1652 /* If there is no bulk associated with this request,
1653 * then we're done and should let the interpreter
1654 * process the reply. Similarly if the RPC returned
1655 * an error, and therefore the bulk will never arrive.
1657 if (req->rq_bulk == NULL || req->rq_status != 0) {
1658 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1659 GOTO(interpret, req->rq_status);
1662 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1665 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1666 if (ptlrpc_client_bulk_active(req))
1669 if (!req->rq_bulk->bd_success) {
1670 /* The RPC reply arrived OK, but the bulk screwed
1671 * up! Dead weird since the server told us the RPC
1672 * was good after getting the REPLY for her GET or
1673 * the ACK for her PUT. */
1674 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1678 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1681 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1683 /* This moves to "unregistering" phase we need to wait for
1685 if (!unregistered && !ptlrpc_unregister_reply(req, 1))
1688 if (!ptlrpc_unregister_bulk(req, 1))
1691 /* When calling interpret receiving already should be
1693 LASSERT(!req->rq_receiving_reply);
1695 ptlrpc_req_interpret(env, req, req->rq_status);
1697 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1699 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1700 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1701 imp->imp_obd->obd_uuid.uuid,
1702 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1704 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1705 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1707 cfs_spin_lock(&imp->imp_lock);
1708 /* Request already may be not on sending or delaying list. This
1709 * may happen in the case of marking it erroneous for the case
1710 * ptlrpc_import_delay_req(req, status) find it impossible to
1711 * allow sending this rpc and returns *status != 0. */
1712 if (!cfs_list_empty(&req->rq_list)) {
1713 cfs_list_del_init(&req->rq_list);
1714 cfs_atomic_dec(&imp->imp_inflight);
1716 cfs_spin_unlock(&imp->imp_lock);
1718 cfs_atomic_dec(&set->set_remaining);
1719 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1722 /* If we hit an error, we want to recover promptly. */
1723 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1727 * Time out request \a req. is \a async_unlink is set, that means do not wait
1728 * until LNet actually confirms network buffer unlinking.
1729 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1731 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1733 struct obd_import *imp = req->rq_import;
1737 cfs_spin_lock(&req->rq_lock);
1738 req->rq_timedout = 1;
1739 cfs_spin_unlock(&req->rq_lock);
1741 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req, "Request x"LPU64
1742 " sent from %s to NID %s has %s: [sent "CFS_DURATION_T"] "
1743 "[real_sent "CFS_DURATION_T"] [current "CFS_DURATION_T"] "
1744 "[deadline "CFS_DURATION_T"s] [delay "CFS_DURATION_T"s]",
1745 req->rq_xid, imp ? imp->imp_obd->obd_name : "<?>",
1746 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1747 req->rq_net_err ? "failed due to network error" :
1748 ((req->rq_real_sent == 0 ||
1749 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1750 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1751 "timed out for sent delay" : "timed out for slow reply"),
1752 req->rq_sent, req->rq_real_sent, cfs_time_current_sec(),
1753 cfs_time_sub(req->rq_deadline, req->rq_sent),
1754 cfs_time_sub(cfs_time_current_sec(), req->rq_deadline));
1756 if (imp != NULL && obd_debug_peer_on_timeout)
1757 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1759 ptlrpc_unregister_reply(req, async_unlink);
1760 ptlrpc_unregister_bulk(req, async_unlink);
1762 if (obd_dump_on_timeout)
1763 libcfs_debug_dumplog();
1766 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1773 cfs_atomic_inc(&imp->imp_timeouts);
1775 /* The DLM server doesn't want recovery run on its imports. */
1776 if (imp->imp_dlm_fake)
1779 /* If this request is for recovery or other primordial tasks,
1780 * then error it out here. */
1781 if (req->rq_ctx_init || req->rq_ctx_fini ||
1782 req->rq_send_state != LUSTRE_IMP_FULL ||
1783 imp->imp_obd->obd_no_recov) {
1784 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1785 ptlrpc_import_state_name(req->rq_send_state),
1786 ptlrpc_import_state_name(imp->imp_state));
1787 cfs_spin_lock(&req->rq_lock);
1788 req->rq_status = -ETIMEDOUT;
1790 cfs_spin_unlock(&req->rq_lock);
1794 /* if a request can't be resent we can't wait for an answer after
1796 if (ptlrpc_no_resend(req)) {
1797 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1801 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1807 * Time out all uncompleted requests in request set pointed by \a data
1808 * Callback used when waiting on sets with l_wait_event.
1811 int ptlrpc_expired_set(void *data)
1813 struct ptlrpc_request_set *set = data;
1815 time_t now = cfs_time_current_sec();
1818 LASSERT(set != NULL);
1821 * A timeout expired. See which reqs it applies to...
1823 cfs_list_for_each (tmp, &set->set_requests) {
1824 struct ptlrpc_request *req =
1825 cfs_list_entry(tmp, struct ptlrpc_request,
1828 /* don't expire request waiting for context */
1829 if (req->rq_wait_ctx)
1832 /* Request in-flight? */
1833 if (!((req->rq_phase == RQ_PHASE_RPC &&
1834 !req->rq_waiting && !req->rq_resend) ||
1835 (req->rq_phase == RQ_PHASE_BULK)))
1838 if (req->rq_timedout || /* already dealt with */
1839 req->rq_deadline > now) /* not expired */
1842 /* Deal with this guy. Do it asynchronously to not block
1843 * ptlrpcd thread. */
1844 ptlrpc_expire_one_request(req, 1);
1848 * When waiting for a whole set, we always break out of the
1849 * sleep so we can recalculate the timeout, or enable interrupts
1850 * if everyone's timed out.
1856 * Sets rq_intr flag in \a req under spinlock.
1858 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1860 cfs_spin_lock(&req->rq_lock);
1862 cfs_spin_unlock(&req->rq_lock);
1866 * Interrupts (sets interrupted flag) all uncompleted requests in
1867 * a set \a data. Callback for l_wait_event for interruptible waits.
1869 void ptlrpc_interrupted_set(void *data)
1871 struct ptlrpc_request_set *set = data;
1874 LASSERT(set != NULL);
1875 CERROR("INTERRUPTED SET %p\n", set);
1877 cfs_list_for_each(tmp, &set->set_requests) {
1878 struct ptlrpc_request *req =
1879 cfs_list_entry(tmp, struct ptlrpc_request,
1882 if (req->rq_phase != RQ_PHASE_RPC &&
1883 req->rq_phase != RQ_PHASE_UNREGISTERING)
1886 ptlrpc_mark_interrupted(req);
1891 * Get the smallest timeout in the set; this does NOT set a timeout.
1893 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1896 time_t now = cfs_time_current_sec();
1898 struct ptlrpc_request *req;
1902 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1904 cfs_list_for_each(tmp, &set->set_requests) {
1905 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1908 * Request in-flight?
1910 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1911 (req->rq_phase == RQ_PHASE_BULK) ||
1912 (req->rq_phase == RQ_PHASE_NEW)))
1916 * Already timed out.
1918 if (req->rq_timedout)
1924 if (req->rq_wait_ctx)
1927 if (req->rq_phase == RQ_PHASE_NEW)
1928 deadline = req->rq_sent;
1930 deadline = req->rq_sent + req->rq_timeout;
1932 if (deadline <= now) /* actually expired already */
1933 timeout = 1; /* ASAP */
1934 else if (timeout == 0 || timeout > deadline - now)
1935 timeout = deadline - now;
1941 * Send all unset request from the set and then wait untill all
1942 * requests in the set complete (either get a reply, timeout, get an
1943 * error or otherwise be interrupted).
1944 * Returns 0 on success or error code otherwise.
1946 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1949 struct ptlrpc_request *req;
1950 struct l_wait_info lwi;
1954 if (cfs_list_empty(&set->set_requests))
1957 cfs_list_for_each(tmp, &set->set_requests) {
1958 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1959 if (req->rq_phase == RQ_PHASE_NEW)
1960 (void)ptlrpc_send_new_req(req);
1964 timeout = ptlrpc_set_next_timeout(set);
1966 /* wait until all complete, interrupted, or an in-flight
1968 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1971 if (timeout == 0 && !cfs_signal_pending())
1973 * No requests are in-flight (ether timed out
1974 * or delayed), so we can allow interrupts.
1975 * We still want to block for a limited time,
1976 * so we allow interrupts during the timeout.
1978 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
1980 ptlrpc_interrupted_set, set);
1983 * At least one request is in flight, so no
1984 * interrupts are allowed. Wait until all
1985 * complete, or an in-flight req times out.
1987 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
1988 ptlrpc_expired_set, set);
1990 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
1992 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1994 /* -EINTR => all requests have been flagged rq_intr so next
1996 * -ETIMEDOUT => someone timed out. When all reqs have
1997 * timed out, signals are enabled allowing completion with
1999 * I don't really care if we go once more round the loop in
2000 * the error cases -eeb. */
2001 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2002 cfs_list_for_each(tmp, &set->set_requests) {
2003 req = cfs_list_entry(tmp, struct ptlrpc_request,
2005 cfs_spin_lock(&req->rq_lock);
2006 req->rq_invalid_rqset = 1;
2007 cfs_spin_unlock(&req->rq_lock);
2010 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2012 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2015 cfs_list_for_each(tmp, &set->set_requests) {
2016 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2018 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2019 if (req->rq_status != 0)
2020 rc = req->rq_status;
2023 if (set->set_interpret != NULL) {
2024 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2026 rc = interpreter (set, set->set_arg, rc);
2028 struct ptlrpc_set_cbdata *cbdata, *n;
2031 cfs_list_for_each_entry_safe(cbdata, n,
2032 &set->set_cblist, psc_item) {
2033 cfs_list_del_init(&cbdata->psc_item);
2034 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2037 OBD_FREE_PTR(cbdata);
2045 * Helper fuction for request freeing.
2046 * Called when request count reached zero and request needs to be freed.
2047 * Removes request from all sorts of sending/replay lists it might be on,
2048 * frees network buffers if any are present.
2049 * If \a locked is set, that means caller is already holding import imp_lock
2050 * and so we no longer need to reobtain it (for certain lists manipulations)
2052 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2055 if (request == NULL) {
2060 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2061 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2062 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2063 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2064 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2065 LASSERTF(!request->rq_replay, "req %p\n", request);
2066 LASSERT(request->rq_cli_ctx || request->rq_fake);
2068 req_capsule_fini(&request->rq_pill);
2070 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2071 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2072 if (request->rq_import != NULL) {
2074 cfs_spin_lock(&request->rq_import->imp_lock);
2075 cfs_list_del_init(&request->rq_replay_list);
2077 cfs_spin_unlock(&request->rq_import->imp_lock);
2079 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2081 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2082 DEBUG_REQ(D_ERROR, request,
2083 "freeing request with nonzero refcount");
2087 if (request->rq_repbuf != NULL)
2088 sptlrpc_cli_free_repbuf(request);
2089 if (request->rq_export != NULL) {
2090 class_export_put(request->rq_export);
2091 request->rq_export = NULL;
2093 if (request->rq_import != NULL) {
2094 class_import_put(request->rq_import);
2095 request->rq_import = NULL;
2097 if (request->rq_bulk != NULL)
2098 ptlrpc_free_bulk(request->rq_bulk);
2100 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2101 sptlrpc_cli_free_reqbuf(request);
2103 if (request->rq_cli_ctx)
2104 sptlrpc_req_put_ctx(request, !locked);
2106 if (request->rq_pool)
2107 __ptlrpc_free_req_to_pool(request);
2109 OBD_FREE(request, sizeof(*request));
2113 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2115 * Drop one request reference. Must be called with import imp_lock held.
2116 * When reference count drops to zero, reuqest is freed.
2118 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2120 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2121 (void)__ptlrpc_req_finished(request, 1);
2126 * Drops one reference count for request \a request.
2127 * \a locked set indicates that caller holds import imp_lock.
2128 * Frees the request whe reference count reaches zero.
2130 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2133 if (request == NULL)
2136 if (request == LP_POISON ||
2137 request->rq_reqmsg == LP_POISON) {
2138 CERROR("dereferencing freed request (bug 575)\n");
2143 DEBUG_REQ(D_INFO, request, "refcount now %u",
2144 cfs_atomic_read(&request->rq_refcount) - 1);
2146 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2147 __ptlrpc_free_req(request, locked);
2155 * Drops one reference count for a request.
2157 void ptlrpc_req_finished(struct ptlrpc_request *request)
2159 __ptlrpc_req_finished(request, 0);
2163 * Returns xid of a \a request
2165 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2167 return request->rq_xid;
2169 EXPORT_SYMBOL(ptlrpc_req_xid);
2172 * Disengage the client's reply buffer from the network
2173 * NB does _NOT_ unregister any client-side bulk.
2174 * IDEMPOTENT, but _not_ safe against concurrent callers.
2175 * The request owner (i.e. the thread doing the I/O) must call...
2176 * Returns 0 on success or 1 if unregistering cannot be made.
2178 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2182 struct l_wait_info lwi;
2187 LASSERT(!cfs_in_interrupt());
2190 * Let's setup deadline for reply unlink.
2192 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2193 async && request->rq_reply_deadline == 0)
2194 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2197 * Nothing left to do.
2199 if (!ptlrpc_client_recv_or_unlink(request))
2202 LNetMDUnlink(request->rq_reply_md_h);
2205 * Let's check it once again.
2207 if (!ptlrpc_client_recv_or_unlink(request))
2211 * Move to "Unregistering" phase as reply was not unlinked yet.
2213 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2216 * Do not wait for unlink to finish.
2222 * We have to l_wait_event() whatever the result, to give liblustre
2223 * a chance to run reply_in_callback(), and to make sure we've
2224 * unlinked before returning a req to the pool.
2226 if (request->rq_set != NULL)
2227 wq = &request->rq_set->set_waitq;
2229 wq = &request->rq_reply_waitq;
2232 /* Network access will complete in finite time but the HUGE
2233 * timeout lets us CWARN for visibility of sluggish NALs */
2234 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2235 cfs_time_seconds(1), NULL, NULL);
2236 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2239 ptlrpc_rqphase_move(request, request->rq_next_phase);
2243 LASSERT(rc == -ETIMEDOUT);
2244 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2245 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2246 request->rq_must_unlink);
2252 * Iterates through replay_list on import and prunes
2253 * all requests have transno smaller than last_committed for the
2254 * import and don't have rq_replay set.
2255 * Since requests are sorted in transno order, stops when meetign first
2256 * transno bigger than last_committed.
2257 * caller must hold imp->imp_lock
2259 void ptlrpc_free_committed(struct obd_import *imp)
2261 cfs_list_t *tmp, *saved;
2262 struct ptlrpc_request *req;
2263 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2266 LASSERT(imp != NULL);
2268 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2271 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2272 imp->imp_generation == imp->imp_last_generation_checked) {
2273 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
2274 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2278 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2279 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2280 imp->imp_generation);
2281 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2282 imp->imp_last_generation_checked = imp->imp_generation;
2284 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2285 req = cfs_list_entry(tmp, struct ptlrpc_request,
2288 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2289 LASSERT(req != last_req);
2292 if (req->rq_transno == 0) {
2293 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2296 if (req->rq_import_generation < imp->imp_generation) {
2297 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2301 if (req->rq_replay) {
2302 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2306 /* not yet committed */
2307 if (req->rq_transno > imp->imp_peer_committed_transno) {
2308 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2312 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
2313 imp->imp_peer_committed_transno);
2315 cfs_spin_lock(&req->rq_lock);
2317 cfs_spin_unlock(&req->rq_lock);
2318 if (req->rq_commit_cb != NULL)
2319 req->rq_commit_cb(req);
2320 cfs_list_del_init(&req->rq_replay_list);
2321 __ptlrpc_req_finished(req, 1);
2328 void ptlrpc_cleanup_client(struct obd_import *imp)
2336 * Schedule previously sent request for resend.
2337 * For bulk requests we assign new xid (to avoid problems with
2338 * lost replies and therefore several transfers landing into same buffer
2339 * from different sending attempts).
2341 void ptlrpc_resend_req(struct ptlrpc_request *req)
2343 DEBUG_REQ(D_HA, req, "going to resend");
2344 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2345 req->rq_status = -EAGAIN;
2347 cfs_spin_lock(&req->rq_lock);
2349 req->rq_net_err = 0;
2350 req->rq_timedout = 0;
2352 __u64 old_xid = req->rq_xid;
2354 /* ensure previous bulk fails */
2355 req->rq_xid = ptlrpc_next_xid();
2356 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2357 old_xid, req->rq_xid);
2359 ptlrpc_client_wake_req(req);
2360 cfs_spin_unlock(&req->rq_lock);
2363 /* XXX: this function and rq_status are currently unused */
2364 void ptlrpc_restart_req(struct ptlrpc_request *req)
2366 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2367 req->rq_status = -ERESTARTSYS;
2369 cfs_spin_lock(&req->rq_lock);
2370 req->rq_restart = 1;
2371 req->rq_timedout = 0;
2372 ptlrpc_client_wake_req(req);
2373 cfs_spin_unlock(&req->rq_lock);
2377 * Grab additional reference on a request \a req
2379 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2382 cfs_atomic_inc(&req->rq_refcount);
2387 * Add a request to import replay_list.
2388 * Must be called under imp_lock
2390 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2391 struct obd_import *imp)
2395 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2397 if (req->rq_transno == 0) {
2398 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2402 /* clear this for new requests that were resent as well
2403 as resent replayed requests. */
2404 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2406 /* don't re-add requests that have been replayed */
2407 if (!cfs_list_empty(&req->rq_replay_list))
2410 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2412 LASSERT(imp->imp_replayable);
2413 /* Balanced in ptlrpc_free_committed, usually. */
2414 ptlrpc_request_addref(req);
2415 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2416 struct ptlrpc_request *iter =
2417 cfs_list_entry(tmp, struct ptlrpc_request,
2420 /* We may have duplicate transnos if we create and then
2421 * open a file, or for closes retained if to match creating
2422 * opens, so use req->rq_xid as a secondary key.
2423 * (See bugs 684, 685, and 428.)
2424 * XXX no longer needed, but all opens need transnos!
2426 if (iter->rq_transno > req->rq_transno)
2429 if (iter->rq_transno == req->rq_transno) {
2430 LASSERT(iter->rq_xid != req->rq_xid);
2431 if (iter->rq_xid > req->rq_xid)
2435 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2439 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2443 * Send request and wait until it completes.
2444 * Returns request processing status.
2446 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2448 struct ptlrpc_request_set *set;
2452 LASSERT(req->rq_set == NULL);
2453 LASSERT(!req->rq_receiving_reply);
2455 set = ptlrpc_prep_set();
2457 CERROR("Unable to allocate ptlrpc set.");
2461 /* for distributed debugging */
2462 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2464 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2465 ptlrpc_request_addref(req);
2466 ptlrpc_set_add_req(set, req);
2467 rc = ptlrpc_set_wait(set);
2468 ptlrpc_set_destroy(set);
2473 struct ptlrpc_replay_async_args {
2475 int praa_old_status;
2479 * Callback used for replayed requests reply processing.
2480 * In case of succesful reply calls registeresd request replay callback.
2481 * In case of error restart replay process.
2483 static int ptlrpc_replay_interpret(const struct lu_env *env,
2484 struct ptlrpc_request *req,
2485 void * data, int rc)
2487 struct ptlrpc_replay_async_args *aa = data;
2488 struct obd_import *imp = req->rq_import;
2491 cfs_atomic_dec(&imp->imp_replay_inflight);
2493 if (!ptlrpc_client_replied(req)) {
2494 CERROR("request replay timed out, restarting recovery\n");
2495 GOTO(out, rc = -ETIMEDOUT);
2498 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2499 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2500 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2501 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2503 /** VBR: check version failure */
2504 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2505 /** replay was failed due to version mismatch */
2506 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2507 cfs_spin_lock(&imp->imp_lock);
2508 imp->imp_vbr_failed = 1;
2509 imp->imp_no_lock_replay = 1;
2510 cfs_spin_unlock(&imp->imp_lock);
2512 /** The transno had better not change over replay. */
2513 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2514 lustre_msg_get_transno(req->rq_repmsg) ||
2515 lustre_msg_get_transno(req->rq_repmsg) == 0,
2517 lustre_msg_get_transno(req->rq_reqmsg),
2518 lustre_msg_get_transno(req->rq_repmsg));
2521 cfs_spin_lock(&imp->imp_lock);
2522 /** if replays by version then gap was occur on server, no trust to locks */
2523 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2524 imp->imp_no_lock_replay = 1;
2525 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2526 cfs_spin_unlock(&imp->imp_lock);
2527 LASSERT(imp->imp_last_replay_transno);
2529 DEBUG_REQ(D_HA, req, "got rep");
2531 /* let the callback do fixups, possibly including in the request */
2532 if (req->rq_replay_cb)
2533 req->rq_replay_cb(req);
2535 if (ptlrpc_client_replied(req) &&
2536 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2537 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2538 lustre_msg_get_status(req->rq_repmsg),
2539 aa->praa_old_status);
2541 /* Put it back for re-replay. */
2542 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2546 * Errors while replay can set transno to 0, but
2547 * imp_last_replay_transno shouldn't be set to 0 anyway
2549 if (req->rq_transno > 0) {
2550 cfs_spin_lock(&imp->imp_lock);
2551 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2552 imp->imp_last_replay_transno = req->rq_transno;
2553 cfs_spin_unlock(&imp->imp_lock);
2555 CERROR("Transno is 0 during replay!\n");
2556 /* continue with recovery */
2557 rc = ptlrpc_import_recovery_state_machine(imp);
2559 req->rq_send_state = aa->praa_old_state;
2562 /* this replay failed, so restart recovery */
2563 ptlrpc_connect_import(imp, NULL);
2569 * Prepares and queues request for replay.
2570 * Adds it to ptlrpcd queue for actual sending.
2571 * Returns 0 on success.
2573 int ptlrpc_replay_req(struct ptlrpc_request *req)
2575 struct ptlrpc_replay_async_args *aa;
2578 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2579 /* Not handling automatic bulk replay yet (or ever?) */
2580 LASSERT(req->rq_bulk == NULL);
2582 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2583 aa = ptlrpc_req_async_args(req);
2584 memset(aa, 0, sizeof *aa);
2586 /* Prepare request to be resent with ptlrpcd */
2587 aa->praa_old_state = req->rq_send_state;
2588 req->rq_send_state = LUSTRE_IMP_REPLAY;
2589 req->rq_phase = RQ_PHASE_NEW;
2590 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2592 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2594 req->rq_interpret_reply = ptlrpc_replay_interpret;
2595 /* Readjust the timeout for current conditions */
2596 ptlrpc_at_set_req_timeout(req);
2598 DEBUG_REQ(D_HA, req, "REPLAY");
2600 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2601 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2603 ptlrpcd_add_req(req, PSCOPE_OTHER);
2608 * Aborts all in-flight request on import \a imp sending and delayed lists
2610 void ptlrpc_abort_inflight(struct obd_import *imp)
2612 cfs_list_t *tmp, *n;
2615 /* Make sure that no new requests get processed for this import.
2616 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2617 * this flag and then putting requests on sending_list or delayed_list.
2619 cfs_spin_lock(&imp->imp_lock);
2621 /* XXX locking? Maybe we should remove each request with the list
2622 * locked? Also, how do we know if the requests on the list are
2623 * being freed at this time?
2625 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2626 struct ptlrpc_request *req =
2627 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2629 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2631 cfs_spin_lock (&req->rq_lock);
2632 if (req->rq_import_generation < imp->imp_generation) {
2634 req->rq_status = -EINTR;
2635 ptlrpc_client_wake_req(req);
2637 cfs_spin_unlock (&req->rq_lock);
2640 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2641 struct ptlrpc_request *req =
2642 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2644 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2646 cfs_spin_lock (&req->rq_lock);
2647 if (req->rq_import_generation < imp->imp_generation) {
2649 req->rq_status = -EINTR;
2650 ptlrpc_client_wake_req(req);
2652 cfs_spin_unlock (&req->rq_lock);
2655 /* Last chance to free reqs left on the replay list, but we
2656 * will still leak reqs that haven't committed. */
2657 if (imp->imp_replayable)
2658 ptlrpc_free_committed(imp);
2660 cfs_spin_unlock(&imp->imp_lock);
2666 * Abort all uncompleted requests in request set \a set
2668 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2670 cfs_list_t *tmp, *pos;
2672 LASSERT(set != NULL);
2674 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2675 struct ptlrpc_request *req =
2676 cfs_list_entry(pos, struct ptlrpc_request,
2679 cfs_spin_lock(&req->rq_lock);
2680 if (req->rq_phase != RQ_PHASE_RPC) {
2681 cfs_spin_unlock(&req->rq_lock);
2686 req->rq_status = -EINTR;
2687 ptlrpc_client_wake_req(req);
2688 cfs_spin_unlock(&req->rq_lock);
2692 static __u64 ptlrpc_last_xid;
2693 static cfs_spinlock_t ptlrpc_last_xid_lock;
2696 * Initialize the XID for the node. This is common among all requests on
2697 * this node, and only requires the property that it is monotonically
2698 * increasing. It does not need to be sequential. Since this is also used
2699 * as the RDMA match bits, it is important that a single client NOT have
2700 * the same match bits for two different in-flight requests, hence we do
2701 * NOT want to have an XID per target or similar.
2703 * To avoid an unlikely collision between match bits after a client reboot
2704 * (which would deliver old data into the wrong RDMA buffer) initialize
2705 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2706 * If the time is clearly incorrect, we instead use a 62-bit random number.
2707 * In the worst case the random number will overflow 1M RPCs per second in
2708 * 9133 years, or permutations thereof.
2710 #define YEAR_2004 (1ULL << 30)
2711 void ptlrpc_init_xid(void)
2713 time_t now = cfs_time_current_sec();
2715 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2716 if (now < YEAR_2004) {
2717 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2718 ptlrpc_last_xid >>= 2;
2719 ptlrpc_last_xid |= (1ULL << 61);
2721 ptlrpc_last_xid = (__u64)now << 20;
2726 * Increase xid and returns resultng new value to the caller.
2728 __u64 ptlrpc_next_xid(void)
2731 cfs_spin_lock(&ptlrpc_last_xid_lock);
2732 tmp = ++ptlrpc_last_xid;
2733 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2738 * Get a glimpse at what next xid value might have been.
2739 * Returns possible next xid.
2741 __u64 ptlrpc_sample_next_xid(void)
2743 #if BITS_PER_LONG == 32
2744 /* need to avoid possible word tearing on 32-bit systems */
2746 cfs_spin_lock(&ptlrpc_last_xid_lock);
2747 tmp = ptlrpc_last_xid + 1;
2748 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2751 /* No need to lock, since returned value is racy anyways */
2752 return ptlrpc_last_xid + 1;
2755 EXPORT_SYMBOL(ptlrpc_sample_next_xid);