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 CNETERR("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);
198 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
202 * Uninitialize and free bulk descriptor \a desc.
203 * Works on bulk descriptors both from server and client side.
205 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
210 LASSERT(desc != NULL);
211 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
212 LASSERT(!desc->bd_network_rw); /* network hands off or */
213 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
215 sptlrpc_enc_pool_put_pages(desc);
218 class_export_put(desc->bd_export);
220 class_import_put(desc->bd_import);
222 for (i = 0; i < desc->bd_iov_count ; i++)
223 cfs_page_unpin(desc->bd_iov[i].kiov_page);
225 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
226 bd_iov[desc->bd_max_iov]));
231 * Set server timelimit for this req, i.e. how long are we willing to wait
232 * for reply before timing out this request.
234 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
240 LASSERT(req->rq_import);
243 /* non-AT settings */
245 * \a imp_server_timeout means this is reverse import and
246 * we send (currently only) ASTs to the client and cannot afford
247 * to wait too long for the reply, otherwise the other client
248 * (because of which we are sending this request) would
249 * timeout waiting for us
251 req->rq_timeout = req->rq_import->imp_server_timeout ?
252 obd_timeout / 2 : obd_timeout;
254 at = &req->rq_import->imp_at;
255 idx = import_at_get_index(req->rq_import,
256 req->rq_request_portal);
257 serv_est = at_get(&at->iat_service_estimate[idx]);
258 req->rq_timeout = at_est2timeout(serv_est);
260 /* We could get even fancier here, using history to predict increased
263 /* Let the server know what this RPC timeout is by putting it in the
265 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
268 /* Adjust max service estimate based on server value */
269 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
270 unsigned int serv_est)
276 LASSERT(req->rq_import);
277 at = &req->rq_import->imp_at;
279 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
280 /* max service estimates are tracked on the server side,
281 so just keep minimal history here */
282 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
284 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
285 "has changed from %d to %d\n",
286 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
287 oldse, at_get(&at->iat_service_estimate[idx]));
290 /* Expected network latency per remote node (secs) */
291 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
293 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
296 /* Adjust expected network latency */
297 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
298 unsigned int service_time)
300 unsigned int nl, oldnl;
302 time_t now = cfs_time_current_sec();
304 LASSERT(req->rq_import);
305 at = &req->rq_import->imp_at;
307 /* Network latency is total time less server processing time */
308 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
309 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
310 CWARN("Reported service time %u > total measured time "
311 CFS_DURATION_T"\n", service_time,
312 cfs_time_sub(now, req->rq_sent));
314 oldnl = at_measured(&at->iat_net_latency, nl);
316 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
317 "has changed from %d to %d\n",
318 req->rq_import->imp_obd->obd_name,
320 &req->rq_import->imp_connection->c_remote_uuid),
321 oldnl, at_get(&at->iat_net_latency));
324 static int unpack_reply(struct ptlrpc_request *req)
328 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
329 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
331 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
336 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
338 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
345 * Handle an early reply message, called with the rq_lock held.
346 * If anything goes wrong just ignore it - same as if it never happened
348 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
350 struct ptlrpc_request *early_req;
356 cfs_spin_unlock(&req->rq_lock);
358 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
360 cfs_spin_lock(&req->rq_lock);
364 rc = unpack_reply(early_req);
366 /* Expecting to increase the service time estimate here */
367 ptlrpc_at_adj_service(req,
368 lustre_msg_get_timeout(early_req->rq_repmsg));
369 ptlrpc_at_adj_net_latency(req,
370 lustre_msg_get_service_time(early_req->rq_repmsg));
373 sptlrpc_cli_finish_early_reply(early_req);
375 cfs_spin_lock(&req->rq_lock);
378 /* Adjust the local timeout for this req */
379 ptlrpc_at_set_req_timeout(req);
381 olddl = req->rq_deadline;
382 /* server assumes it now has rq_timeout from when it sent the
383 early reply, so client should give it at least that long. */
384 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
385 ptlrpc_at_get_net_latency(req);
387 DEBUG_REQ(D_ADAPTTO, req,
388 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
389 "("CFS_DURATION_T"s)", req->rq_early_count,
390 cfs_time_sub(req->rq_deadline,
391 cfs_time_current_sec()),
392 cfs_time_sub(req->rq_deadline, olddl));
399 * Wind down request pool \a pool.
400 * Frees all requests from the pool too
402 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
405 struct ptlrpc_request *req;
407 LASSERT(pool != NULL);
409 cfs_spin_lock(&pool->prp_lock);
410 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
411 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
412 cfs_list_del(&req->rq_list);
413 LASSERT(req->rq_reqbuf);
414 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
415 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
416 OBD_FREE(req, sizeof(*req));
418 cfs_spin_unlock(&pool->prp_lock);
419 OBD_FREE(pool, sizeof(*pool));
423 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
425 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
430 while (size < pool->prp_rq_size)
433 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
434 size == pool->prp_rq_size,
435 "Trying to change pool size with nonempty pool "
436 "from %d to %d bytes\n", pool->prp_rq_size, size);
438 cfs_spin_lock(&pool->prp_lock);
439 pool->prp_rq_size = size;
440 for (i = 0; i < num_rq; i++) {
441 struct ptlrpc_request *req;
442 struct lustre_msg *msg;
444 cfs_spin_unlock(&pool->prp_lock);
445 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
448 OBD_ALLOC_LARGE(msg, size);
450 OBD_FREE(req, sizeof(struct ptlrpc_request));
453 req->rq_reqbuf = msg;
454 req->rq_reqbuf_len = size;
456 cfs_spin_lock(&pool->prp_lock);
457 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
459 cfs_spin_unlock(&pool->prp_lock);
464 * Create and initialize new request pool with given attributes:
465 * \a num_rq - initial number of requests to create for the pool
466 * \a msgsize - maximum message size possible for requests in thid pool
467 * \a populate_pool - function to be called when more requests need to be added
469 * Returns pointer to newly created pool or NULL on error.
471 struct ptlrpc_request_pool *
472 ptlrpc_init_rq_pool(int num_rq, int msgsize,
473 void (*populate_pool)(struct ptlrpc_request_pool *, int))
475 struct ptlrpc_request_pool *pool;
477 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
481 /* Request next power of two for the allocation, because internally
482 kernel would do exactly this */
484 cfs_spin_lock_init(&pool->prp_lock);
485 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
486 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
487 pool->prp_populate = populate_pool;
489 populate_pool(pool, num_rq);
491 if (cfs_list_empty(&pool->prp_req_list)) {
492 /* have not allocated a single request for the pool */
493 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
500 * Fetches one request from pool \a pool
502 static struct ptlrpc_request *
503 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
505 struct ptlrpc_request *request;
506 struct lustre_msg *reqbuf;
511 cfs_spin_lock(&pool->prp_lock);
513 /* See if we have anything in a pool, and bail out if nothing,
514 * in writeout path, where this matters, this is safe to do, because
515 * nothing is lost in this case, and when some in-flight requests
516 * complete, this code will be called again. */
517 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
518 cfs_spin_unlock(&pool->prp_lock);
522 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
524 cfs_list_del_init(&request->rq_list);
525 cfs_spin_unlock(&pool->prp_lock);
527 LASSERT(request->rq_reqbuf);
528 LASSERT(request->rq_pool);
530 reqbuf = request->rq_reqbuf;
531 memset(request, 0, sizeof(*request));
532 request->rq_reqbuf = reqbuf;
533 request->rq_reqbuf_len = pool->prp_rq_size;
534 request->rq_pool = pool;
540 * Returns freed \a request to pool.
542 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
544 struct ptlrpc_request_pool *pool = request->rq_pool;
546 cfs_spin_lock(&pool->prp_lock);
547 LASSERT(cfs_list_empty(&request->rq_list));
548 LASSERT(!request->rq_receiving_reply);
549 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
550 cfs_spin_unlock(&pool->prp_lock);
553 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
554 __u32 version, int opcode,
555 int count, __u32 *lengths, char **bufs,
556 struct ptlrpc_cli_ctx *ctx)
558 struct obd_import *imp = request->rq_import;
563 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
565 rc = sptlrpc_req_get_ctx(request);
570 sptlrpc_req_set_flavor(request, opcode);
572 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
575 LASSERT(!request->rq_pool);
579 lustre_msg_add_version(request->rq_reqmsg, version);
580 request->rq_send_state = LUSTRE_IMP_FULL;
581 request->rq_type = PTL_RPC_MSG_REQUEST;
582 request->rq_export = NULL;
584 request->rq_req_cbid.cbid_fn = request_out_callback;
585 request->rq_req_cbid.cbid_arg = request;
587 request->rq_reply_cbid.cbid_fn = reply_in_callback;
588 request->rq_reply_cbid.cbid_arg = request;
590 request->rq_reply_deadline = 0;
591 request->rq_phase = RQ_PHASE_NEW;
592 request->rq_next_phase = RQ_PHASE_UNDEFINED;
594 request->rq_request_portal = imp->imp_client->cli_request_portal;
595 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
597 ptlrpc_at_set_req_timeout(request);
599 cfs_spin_lock_init(&request->rq_lock);
600 CFS_INIT_LIST_HEAD(&request->rq_list);
601 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
602 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
603 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
604 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
605 CFS_INIT_LIST_HEAD(&request->rq_history_list);
606 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
607 cfs_waitq_init(&request->rq_reply_waitq);
608 cfs_waitq_init(&request->rq_set_waitq);
609 request->rq_xid = ptlrpc_next_xid();
610 cfs_atomic_set(&request->rq_refcount, 1);
612 lustre_msg_set_opc(request->rq_reqmsg, opcode);
616 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
618 class_import_put(imp);
622 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
623 __u32 version, int opcode, char **bufs,
624 struct ptlrpc_cli_ctx *ctx)
628 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
629 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
630 request->rq_pill.rc_area[RCL_CLIENT],
633 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
636 * Pack request buffers for network transfer, performing necessary encryption
637 * steps if necessary.
639 int ptlrpc_request_pack(struct ptlrpc_request *request,
640 __u32 version, int opcode)
642 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
646 * Helper function to allocate new request on import \a imp
647 * and possibly using existing request from pool \a pool if provided.
648 * Returns allocated request structure with import field filled or
652 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
653 struct ptlrpc_request_pool *pool)
655 struct ptlrpc_request *request = NULL;
658 request = ptlrpc_prep_req_from_pool(pool);
661 OBD_ALLOC_PTR(request);
664 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
665 LASSERT(imp != LP_POISON);
666 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
668 LASSERT(imp->imp_client != LP_POISON);
670 request->rq_import = class_import_get(imp);
672 CERROR("request allocation out of memory\n");
679 * Helper function for creating a request.
680 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
681 * buffer structures according to capsule template \a format.
682 * Returns allocated request structure pointer or NULL on error.
684 static struct ptlrpc_request *
685 ptlrpc_request_alloc_internal(struct obd_import *imp,
686 struct ptlrpc_request_pool * pool,
687 const struct req_format *format)
689 struct ptlrpc_request *request;
691 request = __ptlrpc_request_alloc(imp, pool);
695 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
696 req_capsule_set(&request->rq_pill, format);
701 * Allocate new request structure for import \a imp and initialize its
702 * buffer structure according to capsule template \a format.
704 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
705 const struct req_format *format)
707 return ptlrpc_request_alloc_internal(imp, NULL, format);
711 * Allocate new request structure for import \a imp from pool \a pool and
712 * initialize its buffer structure according to capsule template \a format.
714 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
715 struct ptlrpc_request_pool * pool,
716 const struct req_format *format)
718 return ptlrpc_request_alloc_internal(imp, pool, format);
722 * For requests not from pool, free memory of the request structure.
723 * For requests obtained from a pool earlier, return request back to pool.
725 void ptlrpc_request_free(struct ptlrpc_request *request)
727 if (request->rq_pool)
728 __ptlrpc_free_req_to_pool(request);
730 OBD_FREE_PTR(request);
734 * Allocate new request for operatione \a opcode and immediatelly pack it for
736 * Only used for simple requests like OBD_PING where the only important
737 * part of the request is operation itself.
738 * Returns allocated request or NULL on error.
740 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
741 const struct req_format *format,
742 __u32 version, int opcode)
744 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
748 rc = ptlrpc_request_pack(req, version, opcode);
750 ptlrpc_request_free(req);
758 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
759 * for operation \a opcode. Request would contain \a count buffers.
760 * Sizes of buffers are described in array \a lengths and buffers themselves
761 * are provided by a pointer \a bufs.
762 * Returns prepared request structure pointer or NULL on error.
764 struct ptlrpc_request *
765 ptlrpc_prep_req_pool(struct obd_import *imp,
766 __u32 version, int opcode,
767 int count, __u32 *lengths, char **bufs,
768 struct ptlrpc_request_pool *pool)
770 struct ptlrpc_request *request;
773 request = __ptlrpc_request_alloc(imp, pool);
777 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
778 lengths, bufs, NULL);
780 ptlrpc_request_free(request);
787 * Same as ptlrpc_prep_req_pool, but without pool
789 struct ptlrpc_request *
790 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
791 __u32 *lengths, char **bufs)
793 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
798 * Allocate "fake" request that would not be sent anywhere in the end.
799 * Only used as a hack because we have no other way of performing
800 * async actions in lustre between layers.
801 * Used on MDS to request object preallocations from more than one OST at a
804 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
805 unsigned int timeout,
806 ptlrpc_interpterer_t interpreter)
808 struct ptlrpc_request *request = NULL;
811 OBD_ALLOC(request, sizeof(*request));
813 CERROR("request allocation out of memory\n");
817 request->rq_send_state = LUSTRE_IMP_FULL;
818 request->rq_type = PTL_RPC_MSG_REQUEST;
819 request->rq_import = class_import_get(imp);
820 request->rq_export = NULL;
821 request->rq_import_generation = imp->imp_generation;
823 request->rq_timeout = timeout;
824 request->rq_sent = cfs_time_current_sec();
825 request->rq_deadline = request->rq_sent + timeout;
826 request->rq_reply_deadline = request->rq_deadline;
827 request->rq_interpret_reply = interpreter;
828 request->rq_phase = RQ_PHASE_RPC;
829 request->rq_next_phase = RQ_PHASE_INTERPRET;
830 /* don't want reply */
831 request->rq_receiving_reply = 0;
832 request->rq_must_unlink = 0;
833 request->rq_no_delay = request->rq_no_resend = 1;
834 request->rq_fake = 1;
836 cfs_spin_lock_init(&request->rq_lock);
837 CFS_INIT_LIST_HEAD(&request->rq_list);
838 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
839 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
840 CFS_INIT_LIST_HEAD(&request->rq_history_list);
841 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
842 cfs_waitq_init(&request->rq_reply_waitq);
843 cfs_waitq_init(&request->rq_set_waitq);
845 request->rq_xid = ptlrpc_next_xid();
846 cfs_atomic_set(&request->rq_refcount, 1);
852 * Indicate that processing of "fake" request is finished.
854 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
856 /* if we kill request before timeout - need adjust counter */
857 if (req->rq_phase == RQ_PHASE_RPC) {
858 struct ptlrpc_request_set *set = req->rq_set;
861 cfs_atomic_dec(&set->set_remaining);
864 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
865 cfs_list_del_init(&req->rq_list);
869 * Allocate and initialize new request set structure.
870 * Returns a pointer to the newly allocated set structure or NULL on error.
872 struct ptlrpc_request_set *ptlrpc_prep_set(void)
874 struct ptlrpc_request_set *set;
877 OBD_ALLOC(set, sizeof *set);
880 CFS_INIT_LIST_HEAD(&set->set_requests);
881 cfs_waitq_init(&set->set_waitq);
882 cfs_atomic_set(&set->set_remaining, 0);
883 cfs_spin_lock_init(&set->set_new_req_lock);
884 CFS_INIT_LIST_HEAD(&set->set_new_requests);
885 CFS_INIT_LIST_HEAD(&set->set_cblist);
891 * Wind down and free request set structure previously allocated with
893 * Ensures that all requests on the set have completed and removes
894 * all requests from the request list in a set.
895 * If any unsent request happen to be on the list, pretends that they got
896 * an error in flight and calls their completion handler.
898 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
906 /* Requests on the set should either all be completed, or all be new */
907 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
908 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
909 cfs_list_for_each (tmp, &set->set_requests) {
910 struct ptlrpc_request *req =
911 cfs_list_entry(tmp, struct ptlrpc_request,
914 LASSERT(req->rq_phase == expected_phase);
918 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
919 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
920 cfs_atomic_read(&set->set_remaining), n);
922 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
923 struct ptlrpc_request *req =
924 cfs_list_entry(tmp, struct ptlrpc_request,
926 cfs_list_del_init(&req->rq_set_chain);
928 LASSERT(req->rq_phase == expected_phase);
930 if (req->rq_phase == RQ_PHASE_NEW) {
931 ptlrpc_req_interpret(NULL, req, -EBADR);
932 cfs_atomic_dec(&set->set_remaining);
935 cfs_spin_lock(&req->rq_lock);
937 req->rq_invalid_rqset = 0;
938 cfs_spin_unlock(&req->rq_lock);
940 ptlrpc_req_finished (req);
943 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
945 OBD_FREE(set, sizeof(*set));
950 * Add a callback function \a fn to the set.
951 * This function would be called when all requests on this set are completed.
952 * The function will be passed \a data argument.
954 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
955 set_interpreter_func fn, void *data)
957 struct ptlrpc_set_cbdata *cbdata;
959 OBD_ALLOC_PTR(cbdata);
963 cbdata->psc_interpret = fn;
964 cbdata->psc_data = data;
965 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
971 * Add a new request to the general purpose request set.
972 * Assumes request reference from the caller.
974 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
975 struct ptlrpc_request *req)
977 /* The set takes over the caller's request reference */
978 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
980 cfs_atomic_inc(&set->set_remaining);
981 req->rq_queued_time = cfs_time_current(); /* Where is the best place to set this? */
985 * Add a request to a request with dedicated server thread
986 * and wake the thread to make any necessary processing.
987 * Currently only used for ptlrpcd.
988 * Returns 0 if succesful or non zero error code on error.
989 * (the only possible error for now is if the dedicated server thread
992 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
993 struct ptlrpc_request *req)
995 struct ptlrpc_request_set *set = pc->pc_set;
998 * Let caller know that we stopped and will not handle this request.
999 * It needs to take care itself of request.
1001 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags))
1004 cfs_spin_lock(&set->set_new_req_lock);
1006 * The set takes over the caller's request reference.
1008 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1010 cfs_spin_unlock(&set->set_new_req_lock);
1012 cfs_waitq_signal(&set->set_waitq);
1017 * Based on the current state of the import, determine if the request
1018 * can be sent, is an error, or should be delayed.
1020 * Returns true if this request should be delayed. If false, and
1021 * *status is set, then the request can not be sent and *status is the
1022 * error code. If false and status is 0, then request can be sent.
1024 * The imp->imp_lock must be held.
1026 static int ptlrpc_import_delay_req(struct obd_import *imp,
1027 struct ptlrpc_request *req, int *status)
1032 LASSERT (status != NULL);
1035 if (req->rq_ctx_init || req->rq_ctx_fini) {
1036 /* always allow ctx init/fini rpc go through */
1037 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1038 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1041 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1042 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1044 } else if (ptlrpc_send_limit_expired(req)) {
1045 /* probably doesn't need to be a D_ERROR after initial testing */
1046 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1048 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1049 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1050 /* allow CONNECT even if import is invalid */ ;
1051 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1052 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1055 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1056 if (!imp->imp_deactive)
1057 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1058 *status = -ESHUTDOWN; /* bz 12940 */
1059 } else if (req->rq_import_generation != imp->imp_generation) {
1060 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1062 } else if (req->rq_send_state != imp->imp_state) {
1063 /* invalidate in progress - any requests should be drop */
1064 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1065 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1067 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1068 *status = -EWOULDBLOCK;
1078 * Decide if the eror message regarding provided request \a req
1079 * should be printed to the console or not.
1080 * Makes it's decision on request status and other properties.
1081 * Returns 1 to print error on the system console or 0 if not.
1083 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1085 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1088 /* Suppress particular reconnect errors which are to be expected. No
1089 * errors are suppressed for the initial connection on an import */
1090 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1091 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1093 /* Suppress timed out reconnect requests */
1094 if (req->rq_timedout)
1097 /* Suppress unavailable/again reconnect requests */
1098 err = lustre_msg_get_status(req->rq_repmsg);
1099 if (err == -ENODEV || err == -EAGAIN)
1107 * Check request processing status.
1108 * Returns the status.
1110 static int ptlrpc_check_status(struct ptlrpc_request *req)
1115 err = lustre_msg_get_status(req->rq_repmsg);
1116 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1117 struct obd_import *imp = req->rq_import;
1118 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1119 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
1120 " with %s. The %s operation failed with %d\n",
1121 libcfs_nid2str(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);
1133 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1134 struct obd_import *imp = req->rq_import;
1135 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1137 if (ptlrpc_console_allow(req))
1138 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1139 "communicating with %s. The %s "
1140 "operation failed with %d\n",
1142 imp->imp_connection->c_peer.nid),
1143 ll_opcode2str(opc), err);
1145 RETURN(err < 0 ? err : -EINVAL);
1152 * save pre-versions of objects into request for replay.
1153 * Versions are obtained from server reply.
1156 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1158 struct lustre_msg *repmsg = req->rq_repmsg;
1159 struct lustre_msg *reqmsg = req->rq_reqmsg;
1160 __u64 *versions = lustre_msg_get_versions(repmsg);
1163 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1167 lustre_msg_set_versions(reqmsg, versions);
1168 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1169 versions[0], versions[1]);
1175 * Callback function called when client receives RPC reply for \a req.
1176 * Returns 0 on success or error code.
1177 * The return alue would be assigned to req->rq_status by the caller
1178 * as request processing status.
1179 * This function also decides if the request needs to be saved for later replay.
1181 static int after_reply(struct ptlrpc_request *req)
1183 struct obd_import *imp = req->rq_import;
1184 struct obd_device *obd = req->rq_import->imp_obd;
1186 struct timeval work_start;
1190 LASSERT(obd != NULL);
1191 /* repbuf must be unlinked */
1192 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1194 if (req->rq_reply_truncate) {
1195 if (ptlrpc_no_resend(req)) {
1196 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1197 " expected: %d, actual size: %d",
1198 req->rq_nob_received, req->rq_repbuf_len);
1202 sptlrpc_cli_free_repbuf(req);
1203 /* Pass the required reply buffer size (include
1204 * space for early reply).
1205 * NB: no need to roundup because alloc_repbuf
1206 * will roundup it */
1207 req->rq_replen = req->rq_nob_received;
1208 req->rq_nob_received = 0;
1214 * NB Until this point, the whole of the incoming message,
1215 * including buflens, status etc is in the sender's byte order.
1217 rc = sptlrpc_cli_unwrap_reply(req);
1219 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1224 * Security layer unwrap might ask resend this request.
1229 rc = unpack_reply(req);
1233 cfs_gettimeofday(&work_start);
1234 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1235 if (obd->obd_svc_stats != NULL) {
1236 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1238 ptlrpc_lprocfs_rpc_sent(req, timediff);
1241 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1242 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1243 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1244 lustre_msg_get_type(req->rq_repmsg));
1248 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1249 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1250 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1251 ptlrpc_at_adj_net_latency(req,
1252 lustre_msg_get_service_time(req->rq_repmsg));
1254 rc = ptlrpc_check_status(req);
1255 imp->imp_connect_error = rc;
1259 * Either we've been evicted, or the server has failed for
1260 * some reason. Try to reconnect, and if that fails, punt to
1263 if (ll_rpc_recoverable_error(rc)) {
1264 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1265 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1268 ptlrpc_request_handle_notconn(req);
1273 * Let's look if server sent slv. Do it only for RPC with
1276 ldlm_cli_update_pool(req);
1280 * Store transno in reqmsg for replay.
1282 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1283 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1284 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1287 if (imp->imp_replayable) {
1288 cfs_spin_lock(&imp->imp_lock);
1290 * No point in adding already-committed requests to the replay
1291 * list, we will just remove them immediately. b=9829
1293 if (req->rq_transno != 0 &&
1295 lustre_msg_get_last_committed(req->rq_repmsg) ||
1297 /** version recovery */
1298 ptlrpc_save_versions(req);
1299 ptlrpc_retain_replayable_request(req, imp);
1300 } else if (req->rq_commit_cb != NULL) {
1301 cfs_spin_unlock(&imp->imp_lock);
1302 req->rq_commit_cb(req);
1303 cfs_spin_lock(&imp->imp_lock);
1307 * Replay-enabled imports return commit-status information.
1309 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1310 imp->imp_peer_committed_transno =
1311 lustre_msg_get_last_committed(req->rq_repmsg);
1313 ptlrpc_free_committed(imp);
1315 if (req->rq_transno > imp->imp_peer_committed_transno)
1316 ptlrpc_pinger_commit_expected(imp);
1318 cfs_spin_unlock(&imp->imp_lock);
1325 * Helper function to send request \a req over the network for the first time
1326 * Also adjusts request phase.
1327 * Returns 0 on success or error code.
1329 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1331 struct obd_import *imp;
1335 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1336 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1339 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1341 imp = req->rq_import;
1342 cfs_spin_lock(&imp->imp_lock);
1344 req->rq_import_generation = imp->imp_generation;
1346 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1347 cfs_spin_lock(&req->rq_lock);
1348 req->rq_waiting = 1;
1349 cfs_spin_unlock(&req->rq_lock);
1351 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1352 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1353 ptlrpc_import_state_name(req->rq_send_state),
1354 ptlrpc_import_state_name(imp->imp_state));
1355 LASSERT(cfs_list_empty(&req->rq_list));
1356 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1357 cfs_atomic_inc(&req->rq_import->imp_inflight);
1358 cfs_spin_unlock(&imp->imp_lock);
1363 cfs_spin_unlock(&imp->imp_lock);
1364 req->rq_status = rc;
1365 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1369 LASSERT(cfs_list_empty(&req->rq_list));
1370 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1371 cfs_atomic_inc(&req->rq_import->imp_inflight);
1372 cfs_spin_unlock(&imp->imp_lock);
1374 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1376 rc = sptlrpc_req_refresh_ctx(req, -1);
1379 req->rq_status = rc;
1382 req->rq_wait_ctx = 1;
1387 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1388 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1389 imp->imp_obd->obd_uuid.uuid,
1390 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1391 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1392 lustre_msg_get_opc(req->rq_reqmsg));
1394 rc = ptl_send_rpc(req, 0);
1396 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1397 req->rq_net_err = 1;
1404 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1405 * and no more replies are expected.
1406 * (it is possible to get less replies than requests sent e.g. due to timed out
1407 * requests or requests that we had trouble to send out)
1409 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1412 int force_timer_recalc = 0;
1415 if (cfs_atomic_read(&set->set_remaining) == 0)
1418 cfs_list_for_each(tmp, &set->set_requests) {
1419 struct ptlrpc_request *req =
1420 cfs_list_entry(tmp, struct ptlrpc_request,
1422 struct obd_import *imp = req->rq_import;
1423 int unregistered = 0;
1426 if (req->rq_phase == RQ_PHASE_NEW &&
1427 ptlrpc_send_new_req(req)) {
1428 force_timer_recalc = 1;
1431 /* delayed send - skip */
1432 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1435 if (!(req->rq_phase == RQ_PHASE_RPC ||
1436 req->rq_phase == RQ_PHASE_BULK ||
1437 req->rq_phase == RQ_PHASE_INTERPRET ||
1438 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1439 req->rq_phase == RQ_PHASE_COMPLETE)) {
1440 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1444 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1445 LASSERT(req->rq_next_phase != req->rq_phase);
1446 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1449 * Skip processing until reply is unlinked. We
1450 * can't return to pool before that and we can't
1451 * call interpret before that. We need to make
1452 * sure that all rdma transfers finished and will
1453 * not corrupt any data.
1455 if (ptlrpc_client_recv_or_unlink(req) ||
1456 ptlrpc_client_bulk_active(req))
1460 * Turn fail_loc off to prevent it from looping
1463 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1464 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1467 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1468 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1473 * Move to next phase if reply was successfully
1476 ptlrpc_rqphase_move(req, req->rq_next_phase);
1479 if (req->rq_phase == RQ_PHASE_COMPLETE)
1482 if (req->rq_phase == RQ_PHASE_INTERPRET)
1483 GOTO(interpret, req->rq_status);
1486 * Note that this also will start async reply unlink.
1488 if (req->rq_net_err && !req->rq_timedout) {
1489 ptlrpc_expire_one_request(req, 1);
1492 * Check if we still need to wait for unlink.
1494 if (ptlrpc_client_recv_or_unlink(req) ||
1495 ptlrpc_client_bulk_active(req))
1497 /* If there is no need to resend, fail it now. */
1498 if (req->rq_no_resend) {
1499 if (req->rq_status == 0)
1500 req->rq_status = -EIO;
1501 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1502 GOTO(interpret, req->rq_status);
1509 cfs_spin_lock(&req->rq_lock);
1510 req->rq_replied = 0;
1511 cfs_spin_unlock(&req->rq_lock);
1512 if (req->rq_status == 0)
1513 req->rq_status = -EIO;
1514 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1515 GOTO(interpret, req->rq_status);
1518 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1519 * so it sets rq_intr regardless of individual rpc
1520 * timeouts. The synchronous IO waiting path sets
1521 * rq_intr irrespective of whether ptlrpcd
1522 * has seen a timeout. Our policy is to only interpret
1523 * interrupted rpcs after they have timed out, so we
1524 * need to enforce that here.
1527 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1528 req->rq_wait_ctx)) {
1529 req->rq_status = -EINTR;
1530 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1531 GOTO(interpret, req->rq_status);
1534 if (req->rq_phase == RQ_PHASE_RPC) {
1535 if (req->rq_timedout || req->rq_resend ||
1536 req->rq_waiting || req->rq_wait_ctx) {
1539 if (!ptlrpc_unregister_reply(req, 1))
1542 cfs_spin_lock(&imp->imp_lock);
1543 if (ptlrpc_import_delay_req(imp, req, &status)){
1544 /* put on delay list - only if we wait
1545 * recovery finished - before send */
1546 cfs_list_del_init(&req->rq_list);
1547 cfs_list_add_tail(&req->rq_list,
1550 cfs_spin_unlock(&imp->imp_lock);
1555 req->rq_status = status;
1556 ptlrpc_rqphase_move(req,
1557 RQ_PHASE_INTERPRET);
1558 cfs_spin_unlock(&imp->imp_lock);
1559 GOTO(interpret, req->rq_status);
1561 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1562 req->rq_status = -ENOTCONN;
1563 ptlrpc_rqphase_move(req,
1564 RQ_PHASE_INTERPRET);
1565 cfs_spin_unlock(&imp->imp_lock);
1566 GOTO(interpret, req->rq_status);
1569 cfs_list_del_init(&req->rq_list);
1570 cfs_list_add_tail(&req->rq_list,
1571 &imp->imp_sending_list);
1573 cfs_spin_unlock(&imp->imp_lock);
1575 cfs_spin_lock(&req->rq_lock);
1576 req->rq_waiting = 0;
1577 cfs_spin_unlock(&req->rq_lock);
1579 if (req->rq_timedout || req->rq_resend) {
1580 /* This is re-sending anyways,
1581 * let's mark req as resend. */
1582 cfs_spin_lock(&req->rq_lock);
1584 cfs_spin_unlock(&req->rq_lock);
1588 if (!ptlrpc_unregister_bulk(req, 1))
1591 /* ensure previous bulk fails */
1592 old_xid = req->rq_xid;
1593 req->rq_xid = ptlrpc_next_xid();
1594 CDEBUG(D_HA, "resend bulk "
1597 old_xid, req->rq_xid);
1601 * rq_wait_ctx is only touched by ptlrpcd,
1602 * so no lock is needed here.
1604 status = sptlrpc_req_refresh_ctx(req, -1);
1607 req->rq_status = status;
1608 cfs_spin_lock(&req->rq_lock);
1609 req->rq_wait_ctx = 0;
1610 cfs_spin_unlock(&req->rq_lock);
1611 force_timer_recalc = 1;
1613 cfs_spin_lock(&req->rq_lock);
1614 req->rq_wait_ctx = 1;
1615 cfs_spin_unlock(&req->rq_lock);
1620 cfs_spin_lock(&req->rq_lock);
1621 req->rq_wait_ctx = 0;
1622 cfs_spin_unlock(&req->rq_lock);
1625 rc = ptl_send_rpc(req, 0);
1627 DEBUG_REQ(D_HA, req, "send failed (%d)",
1629 force_timer_recalc = 1;
1630 cfs_spin_lock(&req->rq_lock);
1631 req->rq_net_err = 1;
1632 cfs_spin_unlock(&req->rq_lock);
1634 /* need to reset the timeout */
1635 force_timer_recalc = 1;
1638 cfs_spin_lock(&req->rq_lock);
1640 if (ptlrpc_client_early(req)) {
1641 ptlrpc_at_recv_early_reply(req);
1642 cfs_spin_unlock(&req->rq_lock);
1646 /* Still waiting for a reply? */
1647 if (ptlrpc_client_recv(req)) {
1648 cfs_spin_unlock(&req->rq_lock);
1652 /* Did we actually receive a reply? */
1653 if (!ptlrpc_client_replied(req)) {
1654 cfs_spin_unlock(&req->rq_lock);
1658 cfs_spin_unlock(&req->rq_lock);
1660 /* unlink from net because we are going to
1661 * swab in-place of reply buffer */
1662 unregistered = ptlrpc_unregister_reply(req, 1);
1666 req->rq_status = after_reply(req);
1670 /* If there is no bulk associated with this request,
1671 * then we're done and should let the interpreter
1672 * process the reply. Similarly if the RPC returned
1673 * an error, and therefore the bulk will never arrive.
1675 if (req->rq_bulk == NULL || req->rq_status != 0) {
1676 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1677 GOTO(interpret, req->rq_status);
1680 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1683 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1684 if (ptlrpc_client_bulk_active(req))
1687 if (!req->rq_bulk->bd_success) {
1688 /* The RPC reply arrived OK, but the bulk screwed
1689 * up! Dead weird since the server told us the RPC
1690 * was good after getting the REPLY for her GET or
1691 * the ACK for her PUT. */
1692 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1696 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1699 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1701 /* This moves to "unregistering" phase we need to wait for
1703 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1704 /* start async bulk unlink too */
1705 ptlrpc_unregister_bulk(req, 1);
1709 if (!ptlrpc_unregister_bulk(req, 1))
1712 /* When calling interpret receiving already should be
1714 LASSERT(!req->rq_receiving_reply);
1716 ptlrpc_req_interpret(env, req, req->rq_status);
1718 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1720 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1721 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1722 imp->imp_obd->obd_uuid.uuid,
1723 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1725 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1726 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1728 cfs_spin_lock(&imp->imp_lock);
1729 /* Request already may be not on sending or delaying list. This
1730 * may happen in the case of marking it erroneous for the case
1731 * ptlrpc_import_delay_req(req, status) find it impossible to
1732 * allow sending this rpc and returns *status != 0. */
1733 if (!cfs_list_empty(&req->rq_list)) {
1734 cfs_list_del_init(&req->rq_list);
1735 cfs_atomic_dec(&imp->imp_inflight);
1737 cfs_spin_unlock(&imp->imp_lock);
1739 cfs_atomic_dec(&set->set_remaining);
1740 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1743 /* If we hit an error, we want to recover promptly. */
1744 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1748 * Time out request \a req. is \a async_unlink is set, that means do not wait
1749 * until LNet actually confirms network buffer unlinking.
1750 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1752 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1754 struct obd_import *imp = req->rq_import;
1758 cfs_spin_lock(&req->rq_lock);
1759 req->rq_timedout = 1;
1760 cfs_spin_unlock(&req->rq_lock);
1762 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req, "Request x"LPU64
1763 " sent from %s to NID %s has %s: [sent "CFS_DURATION_T"] "
1764 "[real_sent "CFS_DURATION_T"] [current "CFS_DURATION_T"] "
1765 "[deadline "CFS_DURATION_T"s] [delay "CFS_DURATION_T"s]",
1766 req->rq_xid, imp ? imp->imp_obd->obd_name : "<?>",
1767 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1768 req->rq_net_err ? "failed due to network error" :
1769 ((req->rq_real_sent == 0 ||
1770 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1771 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1772 "timed out for sent delay" : "timed out for slow reply"),
1773 req->rq_sent, req->rq_real_sent, cfs_time_current_sec(),
1774 cfs_time_sub(req->rq_deadline, req->rq_sent),
1775 cfs_time_sub(cfs_time_current_sec(), req->rq_deadline));
1777 if (imp != NULL && obd_debug_peer_on_timeout)
1778 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1780 ptlrpc_unregister_reply(req, async_unlink);
1781 ptlrpc_unregister_bulk(req, async_unlink);
1783 if (obd_dump_on_timeout)
1784 libcfs_debug_dumplog();
1787 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1794 cfs_atomic_inc(&imp->imp_timeouts);
1796 /* The DLM server doesn't want recovery run on its imports. */
1797 if (imp->imp_dlm_fake)
1800 /* If this request is for recovery or other primordial tasks,
1801 * then error it out here. */
1802 if (req->rq_ctx_init || req->rq_ctx_fini ||
1803 req->rq_send_state != LUSTRE_IMP_FULL ||
1804 imp->imp_obd->obd_no_recov) {
1805 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1806 ptlrpc_import_state_name(req->rq_send_state),
1807 ptlrpc_import_state_name(imp->imp_state));
1808 cfs_spin_lock(&req->rq_lock);
1809 req->rq_status = -ETIMEDOUT;
1811 cfs_spin_unlock(&req->rq_lock);
1815 /* if a request can't be resent we can't wait for an answer after
1817 if (ptlrpc_no_resend(req)) {
1818 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1822 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1828 * Time out all uncompleted requests in request set pointed by \a data
1829 * Callback used when waiting on sets with l_wait_event.
1832 int ptlrpc_expired_set(void *data)
1834 struct ptlrpc_request_set *set = data;
1836 time_t now = cfs_time_current_sec();
1839 LASSERT(set != NULL);
1842 * A timeout expired. See which reqs it applies to...
1844 cfs_list_for_each (tmp, &set->set_requests) {
1845 struct ptlrpc_request *req =
1846 cfs_list_entry(tmp, struct ptlrpc_request,
1849 /* don't expire request waiting for context */
1850 if (req->rq_wait_ctx)
1853 /* Request in-flight? */
1854 if (!((req->rq_phase == RQ_PHASE_RPC &&
1855 !req->rq_waiting && !req->rq_resend) ||
1856 (req->rq_phase == RQ_PHASE_BULK)))
1859 if (req->rq_timedout || /* already dealt with */
1860 req->rq_deadline > now) /* not expired */
1863 /* Deal with this guy. Do it asynchronously to not block
1864 * ptlrpcd thread. */
1865 ptlrpc_expire_one_request(req, 1);
1869 * When waiting for a whole set, we always break out of the
1870 * sleep so we can recalculate the timeout, or enable interrupts
1871 * if everyone's timed out.
1877 * Sets rq_intr flag in \a req under spinlock.
1879 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1881 cfs_spin_lock(&req->rq_lock);
1883 cfs_spin_unlock(&req->rq_lock);
1887 * Interrupts (sets interrupted flag) all uncompleted requests in
1888 * a set \a data. Callback for l_wait_event for interruptible waits.
1890 void ptlrpc_interrupted_set(void *data)
1892 struct ptlrpc_request_set *set = data;
1895 LASSERT(set != NULL);
1896 CERROR("INTERRUPTED SET %p\n", set);
1898 cfs_list_for_each(tmp, &set->set_requests) {
1899 struct ptlrpc_request *req =
1900 cfs_list_entry(tmp, struct ptlrpc_request,
1903 if (req->rq_phase != RQ_PHASE_RPC &&
1904 req->rq_phase != RQ_PHASE_UNREGISTERING)
1907 ptlrpc_mark_interrupted(req);
1912 * Get the smallest timeout in the set; this does NOT set a timeout.
1914 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1917 time_t now = cfs_time_current_sec();
1919 struct ptlrpc_request *req;
1923 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1925 cfs_list_for_each(tmp, &set->set_requests) {
1926 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1929 * Request in-flight?
1931 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1932 (req->rq_phase == RQ_PHASE_BULK) ||
1933 (req->rq_phase == RQ_PHASE_NEW)))
1937 * Already timed out.
1939 if (req->rq_timedout)
1945 if (req->rq_wait_ctx)
1948 if (req->rq_phase == RQ_PHASE_NEW)
1949 deadline = req->rq_sent;
1951 deadline = req->rq_sent + req->rq_timeout;
1953 if (deadline <= now) /* actually expired already */
1954 timeout = 1; /* ASAP */
1955 else if (timeout == 0 || timeout > deadline - now)
1956 timeout = deadline - now;
1962 * Send all unset request from the set and then wait untill all
1963 * requests in the set complete (either get a reply, timeout, get an
1964 * error or otherwise be interrupted).
1965 * Returns 0 on success or error code otherwise.
1967 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1970 struct ptlrpc_request *req;
1971 struct l_wait_info lwi;
1975 if (cfs_list_empty(&set->set_requests))
1978 cfs_list_for_each(tmp, &set->set_requests) {
1979 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1980 if (req->rq_phase == RQ_PHASE_NEW)
1981 (void)ptlrpc_send_new_req(req);
1985 timeout = ptlrpc_set_next_timeout(set);
1987 /* wait until all complete, interrupted, or an in-flight
1989 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1992 if (timeout == 0 && !cfs_signal_pending())
1994 * No requests are in-flight (ether timed out
1995 * or delayed), so we can allow interrupts.
1996 * We still want to block for a limited time,
1997 * so we allow interrupts during the timeout.
1999 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2001 ptlrpc_interrupted_set, set);
2004 * At least one request is in flight, so no
2005 * interrupts are allowed. Wait until all
2006 * complete, or an in-flight req times out.
2008 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2009 ptlrpc_expired_set, set);
2011 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2013 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2015 /* -EINTR => all requests have been flagged rq_intr so next
2017 * -ETIMEDOUT => someone timed out. When all reqs have
2018 * timed out, signals are enabled allowing completion with
2020 * I don't really care if we go once more round the loop in
2021 * the error cases -eeb. */
2022 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2023 cfs_list_for_each(tmp, &set->set_requests) {
2024 req = cfs_list_entry(tmp, struct ptlrpc_request,
2026 cfs_spin_lock(&req->rq_lock);
2027 req->rq_invalid_rqset = 1;
2028 cfs_spin_unlock(&req->rq_lock);
2031 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2033 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2036 cfs_list_for_each(tmp, &set->set_requests) {
2037 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2039 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2040 if (req->rq_status != 0)
2041 rc = req->rq_status;
2044 if (set->set_interpret != NULL) {
2045 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2047 rc = interpreter (set, set->set_arg, rc);
2049 struct ptlrpc_set_cbdata *cbdata, *n;
2052 cfs_list_for_each_entry_safe(cbdata, n,
2053 &set->set_cblist, psc_item) {
2054 cfs_list_del_init(&cbdata->psc_item);
2055 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2058 OBD_FREE_PTR(cbdata);
2066 * Helper fuction for request freeing.
2067 * Called when request count reached zero and request needs to be freed.
2068 * Removes request from all sorts of sending/replay lists it might be on,
2069 * frees network buffers if any are present.
2070 * If \a locked is set, that means caller is already holding import imp_lock
2071 * and so we no longer need to reobtain it (for certain lists manipulations)
2073 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2076 if (request == NULL) {
2081 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2082 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2083 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2084 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2085 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2086 LASSERTF(!request->rq_replay, "req %p\n", request);
2087 LASSERT(request->rq_cli_ctx || request->rq_fake);
2089 req_capsule_fini(&request->rq_pill);
2091 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2092 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2093 if (request->rq_import != NULL) {
2095 cfs_spin_lock(&request->rq_import->imp_lock);
2096 cfs_list_del_init(&request->rq_replay_list);
2098 cfs_spin_unlock(&request->rq_import->imp_lock);
2100 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2102 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2103 DEBUG_REQ(D_ERROR, request,
2104 "freeing request with nonzero refcount");
2108 if (request->rq_repbuf != NULL)
2109 sptlrpc_cli_free_repbuf(request);
2110 if (request->rq_export != NULL) {
2111 class_export_put(request->rq_export);
2112 request->rq_export = NULL;
2114 if (request->rq_import != NULL) {
2115 class_import_put(request->rq_import);
2116 request->rq_import = NULL;
2118 if (request->rq_bulk != NULL)
2119 ptlrpc_free_bulk(request->rq_bulk);
2121 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2122 sptlrpc_cli_free_reqbuf(request);
2124 if (request->rq_cli_ctx)
2125 sptlrpc_req_put_ctx(request, !locked);
2127 if (request->rq_pool)
2128 __ptlrpc_free_req_to_pool(request);
2130 OBD_FREE(request, sizeof(*request));
2134 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2136 * Drop one request reference. Must be called with import imp_lock held.
2137 * When reference count drops to zero, reuqest is freed.
2139 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2141 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2142 (void)__ptlrpc_req_finished(request, 1);
2147 * Drops one reference count for request \a request.
2148 * \a locked set indicates that caller holds import imp_lock.
2149 * Frees the request whe reference count reaches zero.
2151 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2154 if (request == NULL)
2157 if (request == LP_POISON ||
2158 request->rq_reqmsg == LP_POISON) {
2159 CERROR("dereferencing freed request (bug 575)\n");
2164 DEBUG_REQ(D_INFO, request, "refcount now %u",
2165 cfs_atomic_read(&request->rq_refcount) - 1);
2167 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2168 __ptlrpc_free_req(request, locked);
2176 * Drops one reference count for a request.
2178 void ptlrpc_req_finished(struct ptlrpc_request *request)
2180 __ptlrpc_req_finished(request, 0);
2184 * Returns xid of a \a request
2186 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2188 return request->rq_xid;
2190 EXPORT_SYMBOL(ptlrpc_req_xid);
2193 * Disengage the client's reply buffer from the network
2194 * NB does _NOT_ unregister any client-side bulk.
2195 * IDEMPOTENT, but _not_ safe against concurrent callers.
2196 * The request owner (i.e. the thread doing the I/O) must call...
2197 * Returns 0 on success or 1 if unregistering cannot be made.
2199 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2203 struct l_wait_info lwi;
2208 LASSERT(!cfs_in_interrupt());
2211 * Let's setup deadline for reply unlink.
2213 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2214 async && request->rq_reply_deadline == 0)
2215 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2218 * Nothing left to do.
2220 if (!ptlrpc_client_recv_or_unlink(request))
2223 LNetMDUnlink(request->rq_reply_md_h);
2226 * Let's check it once again.
2228 if (!ptlrpc_client_recv_or_unlink(request))
2232 * Move to "Unregistering" phase as reply was not unlinked yet.
2234 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2237 * Do not wait for unlink to finish.
2243 * We have to l_wait_event() whatever the result, to give liblustre
2244 * a chance to run reply_in_callback(), and to make sure we've
2245 * unlinked before returning a req to the pool.
2247 if (request->rq_set != NULL)
2248 wq = &request->rq_set->set_waitq;
2250 wq = &request->rq_reply_waitq;
2253 /* Network access will complete in finite time but the HUGE
2254 * timeout lets us CWARN for visibility of sluggish NALs */
2255 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2256 cfs_time_seconds(1), NULL, NULL);
2257 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2260 ptlrpc_rqphase_move(request, request->rq_next_phase);
2264 LASSERT(rc == -ETIMEDOUT);
2265 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2266 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2267 request->rq_must_unlink);
2273 * Iterates through replay_list on import and prunes
2274 * all requests have transno smaller than last_committed for the
2275 * import and don't have rq_replay set.
2276 * Since requests are sorted in transno order, stops when meetign first
2277 * transno bigger than last_committed.
2278 * caller must hold imp->imp_lock
2280 void ptlrpc_free_committed(struct obd_import *imp)
2282 cfs_list_t *tmp, *saved;
2283 struct ptlrpc_request *req;
2284 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2287 LASSERT(imp != NULL);
2289 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2292 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2293 imp->imp_generation == imp->imp_last_generation_checked) {
2294 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2295 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2299 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2300 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2301 imp->imp_generation);
2302 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2303 imp->imp_last_generation_checked = imp->imp_generation;
2305 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2306 req = cfs_list_entry(tmp, struct ptlrpc_request,
2309 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2310 LASSERT(req != last_req);
2313 if (req->rq_transno == 0) {
2314 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2317 if (req->rq_import_generation < imp->imp_generation) {
2318 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2322 if (req->rq_replay) {
2323 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2327 /* not yet committed */
2328 if (req->rq_transno > imp->imp_peer_committed_transno) {
2329 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2333 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2334 imp->imp_peer_committed_transno);
2336 cfs_spin_lock(&req->rq_lock);
2338 cfs_spin_unlock(&req->rq_lock);
2339 if (req->rq_commit_cb != NULL)
2340 req->rq_commit_cb(req);
2341 cfs_list_del_init(&req->rq_replay_list);
2342 __ptlrpc_req_finished(req, 1);
2349 void ptlrpc_cleanup_client(struct obd_import *imp)
2357 * Schedule previously sent request for resend.
2358 * For bulk requests we assign new xid (to avoid problems with
2359 * lost replies and therefore several transfers landing into same buffer
2360 * from different sending attempts).
2362 void ptlrpc_resend_req(struct ptlrpc_request *req)
2364 DEBUG_REQ(D_HA, req, "going to resend");
2365 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2366 req->rq_status = -EAGAIN;
2368 cfs_spin_lock(&req->rq_lock);
2370 req->rq_net_err = 0;
2371 req->rq_timedout = 0;
2373 __u64 old_xid = req->rq_xid;
2375 /* ensure previous bulk fails */
2376 req->rq_xid = ptlrpc_next_xid();
2377 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2378 old_xid, req->rq_xid);
2380 ptlrpc_client_wake_req(req);
2381 cfs_spin_unlock(&req->rq_lock);
2384 /* XXX: this function and rq_status are currently unused */
2385 void ptlrpc_restart_req(struct ptlrpc_request *req)
2387 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2388 req->rq_status = -ERESTARTSYS;
2390 cfs_spin_lock(&req->rq_lock);
2391 req->rq_restart = 1;
2392 req->rq_timedout = 0;
2393 ptlrpc_client_wake_req(req);
2394 cfs_spin_unlock(&req->rq_lock);
2398 * Grab additional reference on a request \a req
2400 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2403 cfs_atomic_inc(&req->rq_refcount);
2408 * Add a request to import replay_list.
2409 * Must be called under imp_lock
2411 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2412 struct obd_import *imp)
2416 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2418 if (req->rq_transno == 0) {
2419 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2423 /* clear this for new requests that were resent as well
2424 as resent replayed requests. */
2425 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2427 /* don't re-add requests that have been replayed */
2428 if (!cfs_list_empty(&req->rq_replay_list))
2431 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2433 LASSERT(imp->imp_replayable);
2434 /* Balanced in ptlrpc_free_committed, usually. */
2435 ptlrpc_request_addref(req);
2436 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2437 struct ptlrpc_request *iter =
2438 cfs_list_entry(tmp, struct ptlrpc_request,
2441 /* We may have duplicate transnos if we create and then
2442 * open a file, or for closes retained if to match creating
2443 * opens, so use req->rq_xid as a secondary key.
2444 * (See bugs 684, 685, and 428.)
2445 * XXX no longer needed, but all opens need transnos!
2447 if (iter->rq_transno > req->rq_transno)
2450 if (iter->rq_transno == req->rq_transno) {
2451 LASSERT(iter->rq_xid != req->rq_xid);
2452 if (iter->rq_xid > req->rq_xid)
2456 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2460 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2464 * Send request and wait until it completes.
2465 * Returns request processing status.
2467 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2469 struct ptlrpc_request_set *set;
2473 LASSERT(req->rq_set == NULL);
2474 LASSERT(!req->rq_receiving_reply);
2476 set = ptlrpc_prep_set();
2478 CERROR("Unable to allocate ptlrpc set.");
2482 /* for distributed debugging */
2483 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2485 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2486 ptlrpc_request_addref(req);
2487 ptlrpc_set_add_req(set, req);
2488 rc = ptlrpc_set_wait(set);
2489 ptlrpc_set_destroy(set);
2494 struct ptlrpc_replay_async_args {
2496 int praa_old_status;
2500 * Callback used for replayed requests reply processing.
2501 * In case of succesful reply calls registeresd request replay callback.
2502 * In case of error restart replay process.
2504 static int ptlrpc_replay_interpret(const struct lu_env *env,
2505 struct ptlrpc_request *req,
2506 void * data, int rc)
2508 struct ptlrpc_replay_async_args *aa = data;
2509 struct obd_import *imp = req->rq_import;
2512 cfs_atomic_dec(&imp->imp_replay_inflight);
2514 if (!ptlrpc_client_replied(req)) {
2515 CERROR("request replay timed out, restarting recovery\n");
2516 GOTO(out, rc = -ETIMEDOUT);
2519 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2520 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2521 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2522 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2524 /** VBR: check version failure */
2525 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2526 /** replay was failed due to version mismatch */
2527 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2528 cfs_spin_lock(&imp->imp_lock);
2529 imp->imp_vbr_failed = 1;
2530 imp->imp_no_lock_replay = 1;
2531 cfs_spin_unlock(&imp->imp_lock);
2532 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2534 /** The transno had better not change over replay. */
2535 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2536 lustre_msg_get_transno(req->rq_repmsg) ||
2537 lustre_msg_get_transno(req->rq_repmsg) == 0,
2539 lustre_msg_get_transno(req->rq_reqmsg),
2540 lustre_msg_get_transno(req->rq_repmsg));
2543 cfs_spin_lock(&imp->imp_lock);
2544 /** if replays by version then gap was occur on server, no trust to locks */
2545 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2546 imp->imp_no_lock_replay = 1;
2547 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2548 cfs_spin_unlock(&imp->imp_lock);
2549 LASSERT(imp->imp_last_replay_transno);
2551 /* transaction number shouldn't be bigger than the latest replayed */
2552 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2553 DEBUG_REQ(D_ERROR, req,
2554 "Reported transno "LPU64" is bigger than the "
2555 "replayed one: "LPU64, req->rq_transno,
2556 lustre_msg_get_transno(req->rq_reqmsg));
2557 GOTO(out, rc = -EINVAL);
2560 DEBUG_REQ(D_HA, req, "got rep");
2562 /* let the callback do fixups, possibly including in the request */
2563 if (req->rq_replay_cb)
2564 req->rq_replay_cb(req);
2566 if (ptlrpc_client_replied(req) &&
2567 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2568 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2569 lustre_msg_get_status(req->rq_repmsg),
2570 aa->praa_old_status);
2572 /* Put it back for re-replay. */
2573 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2577 * Errors while replay can set transno to 0, but
2578 * imp_last_replay_transno shouldn't be set to 0 anyway
2580 if (req->rq_transno == 0)
2581 CERROR("Transno is 0 during replay!\n");
2583 /* continue with recovery */
2584 rc = ptlrpc_import_recovery_state_machine(imp);
2586 req->rq_send_state = aa->praa_old_state;
2589 /* this replay failed, so restart recovery */
2590 ptlrpc_connect_import(imp, NULL);
2596 * Prepares and queues request for replay.
2597 * Adds it to ptlrpcd queue for actual sending.
2598 * Returns 0 on success.
2600 int ptlrpc_replay_req(struct ptlrpc_request *req)
2602 struct ptlrpc_replay_async_args *aa;
2605 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2607 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2608 aa = ptlrpc_req_async_args(req);
2609 memset(aa, 0, sizeof *aa);
2611 /* Prepare request to be resent with ptlrpcd */
2612 aa->praa_old_state = req->rq_send_state;
2613 req->rq_send_state = LUSTRE_IMP_REPLAY;
2614 req->rq_phase = RQ_PHASE_NEW;
2615 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2617 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2619 req->rq_interpret_reply = ptlrpc_replay_interpret;
2620 /* Readjust the timeout for current conditions */
2621 ptlrpc_at_set_req_timeout(req);
2623 DEBUG_REQ(D_HA, req, "REPLAY");
2625 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2626 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2628 ptlrpcd_add_req(req, PSCOPE_OTHER);
2633 * Aborts all in-flight request on import \a imp sending and delayed lists
2635 void ptlrpc_abort_inflight(struct obd_import *imp)
2637 cfs_list_t *tmp, *n;
2640 /* Make sure that no new requests get processed for this import.
2641 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2642 * this flag and then putting requests on sending_list or delayed_list.
2644 cfs_spin_lock(&imp->imp_lock);
2646 /* XXX locking? Maybe we should remove each request with the list
2647 * locked? Also, how do we know if the requests on the list are
2648 * being freed at this time?
2650 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2651 struct ptlrpc_request *req =
2652 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2654 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2656 cfs_spin_lock (&req->rq_lock);
2657 if (req->rq_import_generation < imp->imp_generation) {
2659 req->rq_status = -EINTR;
2660 ptlrpc_client_wake_req(req);
2662 cfs_spin_unlock (&req->rq_lock);
2665 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2666 struct ptlrpc_request *req =
2667 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2669 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2671 cfs_spin_lock (&req->rq_lock);
2672 if (req->rq_import_generation < imp->imp_generation) {
2674 req->rq_status = -EINTR;
2675 ptlrpc_client_wake_req(req);
2677 cfs_spin_unlock (&req->rq_lock);
2680 /* Last chance to free reqs left on the replay list, but we
2681 * will still leak reqs that haven't committed. */
2682 if (imp->imp_replayable)
2683 ptlrpc_free_committed(imp);
2685 cfs_spin_unlock(&imp->imp_lock);
2691 * Abort all uncompleted requests in request set \a set
2693 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2695 cfs_list_t *tmp, *pos;
2697 LASSERT(set != NULL);
2699 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2700 struct ptlrpc_request *req =
2701 cfs_list_entry(pos, struct ptlrpc_request,
2704 cfs_spin_lock(&req->rq_lock);
2705 if (req->rq_phase != RQ_PHASE_RPC) {
2706 cfs_spin_unlock(&req->rq_lock);
2711 req->rq_status = -EINTR;
2712 ptlrpc_client_wake_req(req);
2713 cfs_spin_unlock(&req->rq_lock);
2717 static __u64 ptlrpc_last_xid;
2718 static cfs_spinlock_t ptlrpc_last_xid_lock;
2721 * Initialize the XID for the node. This is common among all requests on
2722 * this node, and only requires the property that it is monotonically
2723 * increasing. It does not need to be sequential. Since this is also used
2724 * as the RDMA match bits, it is important that a single client NOT have
2725 * the same match bits for two different in-flight requests, hence we do
2726 * NOT want to have an XID per target or similar.
2728 * To avoid an unlikely collision between match bits after a client reboot
2729 * (which would deliver old data into the wrong RDMA buffer) initialize
2730 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2731 * If the time is clearly incorrect, we instead use a 62-bit random number.
2732 * In the worst case the random number will overflow 1M RPCs per second in
2733 * 9133 years, or permutations thereof.
2735 #define YEAR_2004 (1ULL << 30)
2736 void ptlrpc_init_xid(void)
2738 time_t now = cfs_time_current_sec();
2740 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2741 if (now < YEAR_2004) {
2742 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2743 ptlrpc_last_xid >>= 2;
2744 ptlrpc_last_xid |= (1ULL << 61);
2746 ptlrpc_last_xid = (__u64)now << 20;
2751 * Increase xid and returns resultng new value to the caller.
2753 __u64 ptlrpc_next_xid(void)
2756 cfs_spin_lock(&ptlrpc_last_xid_lock);
2757 tmp = ++ptlrpc_last_xid;
2758 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2763 * Get a glimpse at what next xid value might have been.
2764 * Returns possible next xid.
2766 __u64 ptlrpc_sample_next_xid(void)
2768 #if BITS_PER_LONG == 32
2769 /* need to avoid possible word tearing on 32-bit systems */
2771 cfs_spin_lock(&ptlrpc_last_xid_lock);
2772 tmp = ptlrpc_last_xid + 1;
2773 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2776 /* No need to lock, since returned value is racy anyways */
2777 return ptlrpc_last_xid + 1;
2780 EXPORT_SYMBOL(ptlrpc_sample_next_xid);