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.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
39 /** Implementation of client-side PortalRPC interfaces */
41 #define DEBUG_SUBSYSTEM S_RPC
45 #include <liblustre.h>
48 #include <obd_support.h>
49 #include <obd_class.h>
50 #include <lustre_lib.h>
51 #include <lustre_ha.h>
52 #include <lustre_import.h>
53 #include <lustre_req_layout.h>
55 #include "ptlrpc_internal.h"
58 * Initialize passed in client structure \a cl.
60 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
61 struct ptlrpc_client *cl)
63 cl->cli_request_portal = req_portal;
64 cl->cli_reply_portal = rep_portal;
69 * Return PortalRPC connection for remore uud \a uuid
71 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
73 struct ptlrpc_connection *c;
75 lnet_process_id_t peer;
78 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
80 CNETERR("cannot find peer %s!\n", uuid->uuid);
84 c = ptlrpc_connection_get(peer, self, uuid);
86 memcpy(c->c_remote_uuid.uuid,
87 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
90 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
96 * Allocate and initialize new bulk descriptor
97 * Returns pointer to the descriptor or NULL on error.
99 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
101 struct ptlrpc_bulk_desc *desc;
103 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
107 cfs_spin_lock_init(&desc->bd_lock);
108 cfs_waitq_init(&desc->bd_waitq);
109 desc->bd_max_iov = npages;
110 desc->bd_iov_count = 0;
111 LNetInvalidateHandle(&desc->bd_md_h);
112 desc->bd_portal = portal;
113 desc->bd_type = type;
119 * Prepare bulk descriptor for specified outgoing request \a req that
120 * can fit \a npages * pages. \a type is bulk type. \a portal is where
121 * the bulk to be sent. Used on client-side.
122 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
125 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
126 int npages, int type, int portal)
128 struct obd_import *imp = req->rq_import;
129 struct ptlrpc_bulk_desc *desc;
132 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
133 desc = new_bulk(npages, type, portal);
137 desc->bd_import_generation = req->rq_import_generation;
138 desc->bd_import = class_import_get(imp);
141 desc->bd_cbid.cbid_fn = client_bulk_callback;
142 desc->bd_cbid.cbid_arg = desc;
144 /* This makes req own desc, and free it when she frees herself */
151 * Prepare bulk descriptor for specified incoming request \a req that
152 * can fit \a npages * pages. \a type is bulk type. \a portal is where
153 * the bulk to be sent. Used on server-side after request was already
155 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
158 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
159 int npages, int type, int portal)
161 struct obd_export *exp = req->rq_export;
162 struct ptlrpc_bulk_desc *desc;
165 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
167 desc = new_bulk(npages, type, portal);
171 desc->bd_export = class_export_get(exp);
174 desc->bd_cbid.cbid_fn = server_bulk_callback;
175 desc->bd_cbid.cbid_arg = desc;
177 /* NB we don't assign rq_bulk here; server-side requests are
178 * re-used, and the handler frees the bulk desc explicitly. */
184 * Add a page \a page to the bulk descriptor \a desc.
185 * Data to transfer in the page starts at offset \a pageoffset and
186 * amount of data to transfer from the page is \a len
188 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
189 cfs_page_t *page, int pageoffset, int len)
191 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
192 LASSERT(page != NULL);
193 LASSERT(pageoffset >= 0);
195 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
200 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
204 * Uninitialize and free bulk descriptor \a desc.
205 * Works on bulk descriptors both from server and client side.
207 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
212 LASSERT(desc != NULL);
213 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
214 LASSERT(!desc->bd_network_rw); /* network hands off or */
215 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
217 sptlrpc_enc_pool_put_pages(desc);
220 class_export_put(desc->bd_export);
222 class_import_put(desc->bd_import);
224 for (i = 0; i < desc->bd_iov_count ; i++)
225 cfs_page_unpin(desc->bd_iov[i].kiov_page);
227 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
228 bd_iov[desc->bd_max_iov]));
233 * Set server timelimit for this req, i.e. how long are we willing to wait
234 * for reply before timing out this request.
236 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
242 LASSERT(req->rq_import);
245 /* non-AT settings */
247 * \a imp_server_timeout means this is reverse import and
248 * we send (currently only) ASTs to the client and cannot afford
249 * to wait too long for the reply, otherwise the other client
250 * (because of which we are sending this request) would
251 * timeout waiting for us
253 req->rq_timeout = req->rq_import->imp_server_timeout ?
254 obd_timeout / 2 : obd_timeout;
256 at = &req->rq_import->imp_at;
257 idx = import_at_get_index(req->rq_import,
258 req->rq_request_portal);
259 serv_est = at_get(&at->iat_service_estimate[idx]);
260 req->rq_timeout = at_est2timeout(serv_est);
262 /* We could get even fancier here, using history to predict increased
265 /* Let the server know what this RPC timeout is by putting it in the
267 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
270 /* Adjust max service estimate based on server value */
271 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
272 unsigned int serv_est)
278 LASSERT(req->rq_import);
279 at = &req->rq_import->imp_at;
281 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
282 /* max service estimates are tracked on the server side,
283 so just keep minimal history here */
284 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
286 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
287 "has changed from %d to %d\n",
288 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
289 oldse, at_get(&at->iat_service_estimate[idx]));
292 /* Expected network latency per remote node (secs) */
293 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
295 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
298 /* Adjust expected network latency */
299 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
300 unsigned int service_time)
302 unsigned int nl, oldnl;
304 time_t now = cfs_time_current_sec();
306 LASSERT(req->rq_import);
307 at = &req->rq_import->imp_at;
309 /* Network latency is total time less server processing time */
310 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
311 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
312 CWARN("Reported service time %u > total measured time "
313 CFS_DURATION_T"\n", service_time,
314 cfs_time_sub(now, req->rq_sent));
316 oldnl = at_measured(&at->iat_net_latency, nl);
318 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
319 "has changed from %d to %d\n",
320 req->rq_import->imp_obd->obd_name,
322 &req->rq_import->imp_connection->c_remote_uuid),
323 oldnl, at_get(&at->iat_net_latency));
326 static int unpack_reply(struct ptlrpc_request *req)
330 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
331 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
333 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
338 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
340 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
347 * Handle an early reply message, called with the rq_lock held.
348 * If anything goes wrong just ignore it - same as if it never happened
350 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
352 struct ptlrpc_request *early_req;
358 cfs_spin_unlock(&req->rq_lock);
360 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
362 cfs_spin_lock(&req->rq_lock);
366 rc = unpack_reply(early_req);
368 /* Expecting to increase the service time estimate here */
369 ptlrpc_at_adj_service(req,
370 lustre_msg_get_timeout(early_req->rq_repmsg));
371 ptlrpc_at_adj_net_latency(req,
372 lustre_msg_get_service_time(early_req->rq_repmsg));
375 sptlrpc_cli_finish_early_reply(early_req);
377 cfs_spin_lock(&req->rq_lock);
380 /* Adjust the local timeout for this req */
381 ptlrpc_at_set_req_timeout(req);
383 olddl = req->rq_deadline;
384 /* server assumes it now has rq_timeout from when it sent the
385 early reply, so client should give it at least that long. */
386 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
387 ptlrpc_at_get_net_latency(req);
389 DEBUG_REQ(D_ADAPTTO, req,
390 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
391 "("CFS_DURATION_T"s)", req->rq_early_count,
392 cfs_time_sub(req->rq_deadline,
393 cfs_time_current_sec()),
394 cfs_time_sub(req->rq_deadline, olddl));
401 * Wind down request pool \a pool.
402 * Frees all requests from the pool too
404 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
407 struct ptlrpc_request *req;
409 LASSERT(pool != NULL);
411 cfs_spin_lock(&pool->prp_lock);
412 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
413 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
414 cfs_list_del(&req->rq_list);
415 LASSERT(req->rq_reqbuf);
416 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
417 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
418 OBD_FREE(req, sizeof(*req));
420 cfs_spin_unlock(&pool->prp_lock);
421 OBD_FREE(pool, sizeof(*pool));
425 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
427 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
432 while (size < pool->prp_rq_size)
435 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
436 size == pool->prp_rq_size,
437 "Trying to change pool size with nonempty pool "
438 "from %d to %d bytes\n", pool->prp_rq_size, size);
440 cfs_spin_lock(&pool->prp_lock);
441 pool->prp_rq_size = size;
442 for (i = 0; i < num_rq; i++) {
443 struct ptlrpc_request *req;
444 struct lustre_msg *msg;
446 cfs_spin_unlock(&pool->prp_lock);
447 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
450 OBD_ALLOC_LARGE(msg, size);
452 OBD_FREE(req, sizeof(struct ptlrpc_request));
455 req->rq_reqbuf = msg;
456 req->rq_reqbuf_len = size;
458 cfs_spin_lock(&pool->prp_lock);
459 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
461 cfs_spin_unlock(&pool->prp_lock);
466 * Create and initialize new request pool with given attributes:
467 * \a num_rq - initial number of requests to create for the pool
468 * \a msgsize - maximum message size possible for requests in thid pool
469 * \a populate_pool - function to be called when more requests need to be added
471 * Returns pointer to newly created pool or NULL on error.
473 struct ptlrpc_request_pool *
474 ptlrpc_init_rq_pool(int num_rq, int msgsize,
475 void (*populate_pool)(struct ptlrpc_request_pool *, int))
477 struct ptlrpc_request_pool *pool;
479 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
483 /* Request next power of two for the allocation, because internally
484 kernel would do exactly this */
486 cfs_spin_lock_init(&pool->prp_lock);
487 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
488 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
489 pool->prp_populate = populate_pool;
491 populate_pool(pool, num_rq);
493 if (cfs_list_empty(&pool->prp_req_list)) {
494 /* have not allocated a single request for the pool */
495 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
502 * Fetches one request from pool \a pool
504 static struct ptlrpc_request *
505 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
507 struct ptlrpc_request *request;
508 struct lustre_msg *reqbuf;
513 cfs_spin_lock(&pool->prp_lock);
515 /* See if we have anything in a pool, and bail out if nothing,
516 * in writeout path, where this matters, this is safe to do, because
517 * nothing is lost in this case, and when some in-flight requests
518 * complete, this code will be called again. */
519 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
520 cfs_spin_unlock(&pool->prp_lock);
524 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
526 cfs_list_del_init(&request->rq_list);
527 cfs_spin_unlock(&pool->prp_lock);
529 LASSERT(request->rq_reqbuf);
530 LASSERT(request->rq_pool);
532 reqbuf = request->rq_reqbuf;
533 memset(request, 0, sizeof(*request));
534 request->rq_reqbuf = reqbuf;
535 request->rq_reqbuf_len = pool->prp_rq_size;
536 request->rq_pool = pool;
542 * Returns freed \a request to pool.
544 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
546 struct ptlrpc_request_pool *pool = request->rq_pool;
548 cfs_spin_lock(&pool->prp_lock);
549 LASSERT(cfs_list_empty(&request->rq_list));
550 LASSERT(!request->rq_receiving_reply);
551 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
552 cfs_spin_unlock(&pool->prp_lock);
555 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
556 __u32 version, int opcode,
557 int count, __u32 *lengths, char **bufs,
558 struct ptlrpc_cli_ctx *ctx)
560 struct obd_import *imp = request->rq_import;
565 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
567 rc = sptlrpc_req_get_ctx(request);
572 sptlrpc_req_set_flavor(request, opcode);
574 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
577 LASSERT(!request->rq_pool);
581 lustre_msg_add_version(request->rq_reqmsg, version);
582 request->rq_send_state = LUSTRE_IMP_FULL;
583 request->rq_type = PTL_RPC_MSG_REQUEST;
584 request->rq_export = NULL;
586 request->rq_req_cbid.cbid_fn = request_out_callback;
587 request->rq_req_cbid.cbid_arg = request;
589 request->rq_reply_cbid.cbid_fn = reply_in_callback;
590 request->rq_reply_cbid.cbid_arg = request;
592 request->rq_reply_deadline = 0;
593 request->rq_phase = RQ_PHASE_NEW;
594 request->rq_next_phase = RQ_PHASE_UNDEFINED;
596 request->rq_request_portal = imp->imp_client->cli_request_portal;
597 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
599 ptlrpc_at_set_req_timeout(request);
601 cfs_spin_lock_init(&request->rq_lock);
602 CFS_INIT_LIST_HEAD(&request->rq_list);
603 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
604 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
605 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
606 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
607 CFS_INIT_LIST_HEAD(&request->rq_history_list);
608 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
609 cfs_waitq_init(&request->rq_reply_waitq);
610 cfs_waitq_init(&request->rq_set_waitq);
611 request->rq_xid = ptlrpc_next_xid();
612 cfs_atomic_set(&request->rq_refcount, 1);
614 lustre_msg_set_opc(request->rq_reqmsg, opcode);
618 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
620 class_import_put(imp);
624 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
625 __u32 version, int opcode, char **bufs,
626 struct ptlrpc_cli_ctx *ctx)
630 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
631 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
632 request->rq_pill.rc_area[RCL_CLIENT],
635 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
638 * Pack request buffers for network transfer, performing necessary encryption
639 * steps if necessary.
641 int ptlrpc_request_pack(struct ptlrpc_request *request,
642 __u32 version, int opcode)
644 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
648 * Helper function to allocate new request on import \a imp
649 * and possibly using existing request from pool \a pool if provided.
650 * Returns allocated request structure with import field filled or
654 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
655 struct ptlrpc_request_pool *pool)
657 struct ptlrpc_request *request = NULL;
660 request = ptlrpc_prep_req_from_pool(pool);
663 OBD_ALLOC_PTR(request);
666 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
667 LASSERT(imp != LP_POISON);
668 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
670 LASSERT(imp->imp_client != LP_POISON);
672 request->rq_import = class_import_get(imp);
674 CERROR("request allocation out of memory\n");
681 * Helper function for creating a request.
682 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
683 * buffer structures according to capsule template \a format.
684 * Returns allocated request structure pointer or NULL on error.
686 static struct ptlrpc_request *
687 ptlrpc_request_alloc_internal(struct obd_import *imp,
688 struct ptlrpc_request_pool * pool,
689 const struct req_format *format)
691 struct ptlrpc_request *request;
693 request = __ptlrpc_request_alloc(imp, pool);
697 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
698 req_capsule_set(&request->rq_pill, format);
703 * Allocate new request structure for import \a imp and initialize its
704 * buffer structure according to capsule template \a format.
706 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
707 const struct req_format *format)
709 return ptlrpc_request_alloc_internal(imp, NULL, format);
713 * Allocate new request structure for import \a imp from pool \a pool and
714 * initialize its buffer structure according to capsule template \a format.
716 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
717 struct ptlrpc_request_pool * pool,
718 const struct req_format *format)
720 return ptlrpc_request_alloc_internal(imp, pool, format);
724 * For requests not from pool, free memory of the request structure.
725 * For requests obtained from a pool earlier, return request back to pool.
727 void ptlrpc_request_free(struct ptlrpc_request *request)
729 if (request->rq_pool)
730 __ptlrpc_free_req_to_pool(request);
732 OBD_FREE_PTR(request);
736 * Allocate new request for operatione \a opcode and immediatelly pack it for
738 * Only used for simple requests like OBD_PING where the only important
739 * part of the request is operation itself.
740 * Returns allocated request or NULL on error.
742 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
743 const struct req_format *format,
744 __u32 version, int opcode)
746 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
750 rc = ptlrpc_request_pack(req, version, opcode);
752 ptlrpc_request_free(req);
760 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
761 * for operation \a opcode. Request would contain \a count buffers.
762 * Sizes of buffers are described in array \a lengths and buffers themselves
763 * are provided by a pointer \a bufs.
764 * Returns prepared request structure pointer or NULL on error.
766 struct ptlrpc_request *
767 ptlrpc_prep_req_pool(struct obd_import *imp,
768 __u32 version, int opcode,
769 int count, __u32 *lengths, char **bufs,
770 struct ptlrpc_request_pool *pool)
772 struct ptlrpc_request *request;
775 request = __ptlrpc_request_alloc(imp, pool);
779 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
780 lengths, bufs, NULL);
782 ptlrpc_request_free(request);
789 * Same as ptlrpc_prep_req_pool, but without pool
791 struct ptlrpc_request *
792 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
793 __u32 *lengths, char **bufs)
795 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
800 * Allocate "fake" request that would not be sent anywhere in the end.
801 * Only used as a hack because we have no other way of performing
802 * async actions in lustre between layers.
803 * Used on MDS to request object preallocations from more than one OST at a
806 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
807 unsigned int timeout,
808 ptlrpc_interpterer_t interpreter)
810 struct ptlrpc_request *request = NULL;
813 OBD_ALLOC(request, sizeof(*request));
815 CERROR("request allocation out of memory\n");
819 request->rq_send_state = LUSTRE_IMP_FULL;
820 request->rq_type = PTL_RPC_MSG_REQUEST;
821 request->rq_import = class_import_get(imp);
822 request->rq_export = NULL;
823 request->rq_import_generation = imp->imp_generation;
825 request->rq_timeout = timeout;
826 request->rq_sent = cfs_time_current_sec();
827 request->rq_deadline = request->rq_sent + timeout;
828 request->rq_reply_deadline = request->rq_deadline;
829 request->rq_interpret_reply = interpreter;
830 request->rq_phase = RQ_PHASE_RPC;
831 request->rq_next_phase = RQ_PHASE_INTERPRET;
832 /* don't want reply */
833 request->rq_receiving_reply = 0;
834 request->rq_must_unlink = 0;
835 request->rq_no_delay = request->rq_no_resend = 1;
836 request->rq_fake = 1;
838 cfs_spin_lock_init(&request->rq_lock);
839 CFS_INIT_LIST_HEAD(&request->rq_list);
840 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
841 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
842 CFS_INIT_LIST_HEAD(&request->rq_history_list);
843 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
844 cfs_waitq_init(&request->rq_reply_waitq);
845 cfs_waitq_init(&request->rq_set_waitq);
847 request->rq_xid = ptlrpc_next_xid();
848 cfs_atomic_set(&request->rq_refcount, 1);
854 * Indicate that processing of "fake" request is finished.
856 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
858 struct ptlrpc_request_set *set = req->rq_set;
861 /* hold ref on the request to prevent others (ptlrpcd) to free it */
862 ptlrpc_request_addref(req);
863 cfs_list_del_init(&req->rq_list);
865 /* if we kill request before timeout - need adjust counter */
866 if (req->rq_phase == RQ_PHASE_RPC && set != NULL &&
867 cfs_atomic_dec_and_test(&set->set_remaining))
870 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
872 /* Only need to call wakeup once when to be empty. */
874 cfs_waitq_signal(&set->set_waitq);
875 ptlrpc_req_finished(req);
879 * Allocate and initialize new request set structure.
880 * Returns a pointer to the newly allocated set structure or NULL on error.
882 struct ptlrpc_request_set *ptlrpc_prep_set(void)
884 struct ptlrpc_request_set *set;
887 OBD_ALLOC(set, sizeof *set);
890 cfs_atomic_set(&set->set_refcount, 1);
891 CFS_INIT_LIST_HEAD(&set->set_requests);
892 cfs_waitq_init(&set->set_waitq);
893 cfs_atomic_set(&set->set_new_count, 0);
894 cfs_atomic_set(&set->set_remaining, 0);
895 cfs_spin_lock_init(&set->set_new_req_lock);
896 CFS_INIT_LIST_HEAD(&set->set_new_requests);
897 CFS_INIT_LIST_HEAD(&set->set_cblist);
903 * Wind down and free request set structure previously allocated with
905 * Ensures that all requests on the set have completed and removes
906 * all requests from the request list in a set.
907 * If any unsent request happen to be on the list, pretends that they got
908 * an error in flight and calls their completion handler.
910 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
918 /* Requests on the set should either all be completed, or all be new */
919 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
920 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
921 cfs_list_for_each (tmp, &set->set_requests) {
922 struct ptlrpc_request *req =
923 cfs_list_entry(tmp, struct ptlrpc_request,
926 LASSERT(req->rq_phase == expected_phase);
930 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
931 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
932 cfs_atomic_read(&set->set_remaining), n);
934 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
935 struct ptlrpc_request *req =
936 cfs_list_entry(tmp, struct ptlrpc_request,
938 cfs_list_del_init(&req->rq_set_chain);
940 LASSERT(req->rq_phase == expected_phase);
942 if (req->rq_phase == RQ_PHASE_NEW) {
943 ptlrpc_req_interpret(NULL, req, -EBADR);
944 cfs_atomic_dec(&set->set_remaining);
947 cfs_spin_lock(&req->rq_lock);
949 req->rq_invalid_rqset = 0;
950 cfs_spin_unlock(&req->rq_lock);
952 ptlrpc_req_finished (req);
955 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
957 ptlrpc_reqset_put(set);
962 * Add a callback function \a fn to the set.
963 * This function would be called when all requests on this set are completed.
964 * The function will be passed \a data argument.
966 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
967 set_interpreter_func fn, void *data)
969 struct ptlrpc_set_cbdata *cbdata;
971 OBD_ALLOC_PTR(cbdata);
975 cbdata->psc_interpret = fn;
976 cbdata->psc_data = data;
977 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
983 * Add a new request to the general purpose request set.
984 * Assumes request reference from the caller.
986 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
987 struct ptlrpc_request *req)
989 LASSERT(cfs_list_empty(&req->rq_set_chain));
991 /* The set takes over the caller's request reference */
992 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
994 cfs_atomic_inc(&set->set_remaining);
995 req->rq_queued_time = cfs_time_current();
999 * Add a request to a request with dedicated server thread
1000 * and wake the thread to make any necessary processing.
1001 * Currently only used for ptlrpcd.
1003 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1004 struct ptlrpc_request *req)
1006 struct ptlrpc_request_set *set = pc->pc_set;
1009 LASSERT(req->rq_set == NULL);
1010 LASSERT(cfs_test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1012 cfs_spin_lock(&set->set_new_req_lock);
1014 * The set takes over the caller's request reference.
1017 req->rq_queued_time = cfs_time_current();
1018 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1019 count = cfs_atomic_inc_return(&set->set_new_count);
1020 cfs_spin_unlock(&set->set_new_req_lock);
1022 /* Only need to call wakeup once for the first entry. */
1024 cfs_waitq_signal(&set->set_waitq);
1026 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1027 * guarantee the async RPC can be processed ASAP, we have
1028 * no other better choice. It maybe fixed in future. */
1029 for (i = 0; i < pc->pc_npartners; i++)
1030 cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
1035 * Based on the current state of the import, determine if the request
1036 * can be sent, is an error, or should be delayed.
1038 * Returns true if this request should be delayed. If false, and
1039 * *status is set, then the request can not be sent and *status is the
1040 * error code. If false and status is 0, then request can be sent.
1042 * The imp->imp_lock must be held.
1044 static int ptlrpc_import_delay_req(struct obd_import *imp,
1045 struct ptlrpc_request *req, int *status)
1050 LASSERT (status != NULL);
1053 if (req->rq_ctx_init || req->rq_ctx_fini) {
1054 /* always allow ctx init/fini rpc go through */
1055 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1056 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1059 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1060 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1062 } else if (ptlrpc_send_limit_expired(req)) {
1063 /* probably doesn't need to be a D_ERROR after initial testing */
1064 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1066 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1067 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1068 /* allow CONNECT even if import is invalid */ ;
1069 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1070 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1073 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1074 if (!imp->imp_deactive)
1075 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1076 *status = -ESHUTDOWN; /* bz 12940 */
1077 } else if (req->rq_import_generation != imp->imp_generation) {
1078 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1080 } else if (req->rq_send_state != imp->imp_state) {
1081 /* invalidate in progress - any requests should be drop */
1082 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1083 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1085 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1086 *status = -EWOULDBLOCK;
1096 * Decide if the eror message regarding provided request \a req
1097 * should be printed to the console or not.
1098 * Makes it's decision on request status and other properties.
1099 * Returns 1 to print error on the system console or 0 if not.
1101 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1106 /* Fake requests include no rq_reqmsg */
1110 LASSERT(req->rq_reqmsg != NULL);
1111 opc = lustre_msg_get_opc(req->rq_reqmsg);
1113 /* Suppress particular reconnect errors which are to be expected. No
1114 * errors are suppressed for the initial connection on an import */
1115 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1116 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1118 /* Suppress timed out reconnect requests */
1119 if (req->rq_timedout)
1122 /* Suppress unavailable/again reconnect requests */
1123 err = lustre_msg_get_status(req->rq_repmsg);
1124 if (err == -ENODEV || err == -EAGAIN)
1132 * Check request processing status.
1133 * Returns the status.
1135 static int ptlrpc_check_status(struct ptlrpc_request *req)
1140 err = lustre_msg_get_status(req->rq_repmsg);
1141 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1142 struct obd_import *imp = req->rq_import;
1143 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1144 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
1145 " with %s. The %s operation failed with %d\n",
1146 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1147 ll_opcode2str(opc), err);
1148 RETURN(err < 0 ? err : -EINVAL);
1152 DEBUG_REQ(D_INFO, req, "status is %d", err);
1153 } else if (err > 0) {
1154 /* XXX: translate this error from net to host */
1155 DEBUG_REQ(D_INFO, req, "status is %d", err);
1158 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1159 struct obd_import *imp = req->rq_import;
1160 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1162 if (ptlrpc_console_allow(req))
1163 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1164 "communicating with %s. The %s "
1165 "operation failed with %d\n",
1167 imp->imp_connection->c_peer.nid),
1168 ll_opcode2str(opc), err);
1170 RETURN(err < 0 ? err : -EINVAL);
1177 * save pre-versions of objects into request for replay.
1178 * Versions are obtained from server reply.
1181 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1183 struct lustre_msg *repmsg = req->rq_repmsg;
1184 struct lustre_msg *reqmsg = req->rq_reqmsg;
1185 __u64 *versions = lustre_msg_get_versions(repmsg);
1188 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1192 lustre_msg_set_versions(reqmsg, versions);
1193 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1194 versions[0], versions[1]);
1200 * Callback function called when client receives RPC reply for \a req.
1201 * Returns 0 on success or error code.
1202 * The return alue would be assigned to req->rq_status by the caller
1203 * as request processing status.
1204 * This function also decides if the request needs to be saved for later replay.
1206 static int after_reply(struct ptlrpc_request *req)
1208 struct obd_import *imp = req->rq_import;
1209 struct obd_device *obd = req->rq_import->imp_obd;
1211 struct timeval work_start;
1215 LASSERT(obd != NULL);
1216 /* repbuf must be unlinked */
1217 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1219 if (req->rq_reply_truncate) {
1220 if (ptlrpc_no_resend(req)) {
1221 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1222 " expected: %d, actual size: %d",
1223 req->rq_nob_received, req->rq_repbuf_len);
1227 sptlrpc_cli_free_repbuf(req);
1228 /* Pass the required reply buffer size (include
1229 * space for early reply).
1230 * NB: no need to roundup because alloc_repbuf
1231 * will roundup it */
1232 req->rq_replen = req->rq_nob_received;
1233 req->rq_nob_received = 0;
1239 * NB Until this point, the whole of the incoming message,
1240 * including buflens, status etc is in the sender's byte order.
1242 rc = sptlrpc_cli_unwrap_reply(req);
1244 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1249 * Security layer unwrap might ask resend this request.
1254 rc = unpack_reply(req);
1258 cfs_gettimeofday(&work_start);
1259 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1260 if (obd->obd_svc_stats != NULL) {
1261 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1263 ptlrpc_lprocfs_rpc_sent(req, timediff);
1266 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1267 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1268 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1269 lustre_msg_get_type(req->rq_repmsg));
1273 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1274 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1275 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1276 ptlrpc_at_adj_net_latency(req,
1277 lustre_msg_get_service_time(req->rq_repmsg));
1279 rc = ptlrpc_check_status(req);
1280 imp->imp_connect_error = rc;
1284 * Either we've been evicted, or the server has failed for
1285 * some reason. Try to reconnect, and if that fails, punt to
1288 if (ll_rpc_recoverable_error(rc)) {
1289 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1290 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1293 ptlrpc_request_handle_notconn(req);
1298 * Let's look if server sent slv. Do it only for RPC with
1301 ldlm_cli_update_pool(req);
1305 * Store transno in reqmsg for replay.
1307 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1308 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1309 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1312 if (imp->imp_replayable) {
1313 cfs_spin_lock(&imp->imp_lock);
1315 * No point in adding already-committed requests to the replay
1316 * list, we will just remove them immediately. b=9829
1318 if (req->rq_transno != 0 &&
1320 lustre_msg_get_last_committed(req->rq_repmsg) ||
1322 /** version recovery */
1323 ptlrpc_save_versions(req);
1324 ptlrpc_retain_replayable_request(req, imp);
1325 } else if (req->rq_commit_cb != NULL) {
1326 cfs_spin_unlock(&imp->imp_lock);
1327 req->rq_commit_cb(req);
1328 cfs_spin_lock(&imp->imp_lock);
1332 * Replay-enabled imports return commit-status information.
1334 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1335 imp->imp_peer_committed_transno =
1336 lustre_msg_get_last_committed(req->rq_repmsg);
1338 ptlrpc_free_committed(imp);
1340 if (req->rq_transno > imp->imp_peer_committed_transno)
1341 ptlrpc_pinger_commit_expected(imp);
1343 cfs_spin_unlock(&imp->imp_lock);
1350 * Helper function to send request \a req over the network for the first time
1351 * Also adjusts request phase.
1352 * Returns 0 on success or error code.
1354 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1356 struct obd_import *imp;
1360 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1361 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1364 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1366 imp = req->rq_import;
1367 cfs_spin_lock(&imp->imp_lock);
1369 req->rq_import_generation = imp->imp_generation;
1371 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1372 cfs_spin_lock(&req->rq_lock);
1373 req->rq_waiting = 1;
1374 cfs_spin_unlock(&req->rq_lock);
1376 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1377 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1378 ptlrpc_import_state_name(req->rq_send_state),
1379 ptlrpc_import_state_name(imp->imp_state));
1380 LASSERT(cfs_list_empty(&req->rq_list));
1381 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1382 cfs_atomic_inc(&req->rq_import->imp_inflight);
1383 cfs_spin_unlock(&imp->imp_lock);
1388 cfs_spin_unlock(&imp->imp_lock);
1389 req->rq_status = rc;
1390 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1394 LASSERT(cfs_list_empty(&req->rq_list));
1395 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1396 cfs_atomic_inc(&req->rq_import->imp_inflight);
1397 cfs_spin_unlock(&imp->imp_lock);
1399 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1401 rc = sptlrpc_req_refresh_ctx(req, -1);
1404 req->rq_status = rc;
1407 req->rq_wait_ctx = 1;
1412 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1413 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1414 imp->imp_obd->obd_uuid.uuid,
1415 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1416 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1417 lustre_msg_get_opc(req->rq_reqmsg));
1419 rc = ptl_send_rpc(req, 0);
1421 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1422 req->rq_net_err = 1;
1429 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1430 * and no more replies are expected.
1431 * (it is possible to get less replies than requests sent e.g. due to timed out
1432 * requests or requests that we had trouble to send out)
1434 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1437 int force_timer_recalc = 0;
1440 if (cfs_atomic_read(&set->set_remaining) == 0)
1443 cfs_list_for_each(tmp, &set->set_requests) {
1444 struct ptlrpc_request *req =
1445 cfs_list_entry(tmp, struct ptlrpc_request,
1447 struct obd_import *imp = req->rq_import;
1448 int unregistered = 0;
1451 if (req->rq_phase == RQ_PHASE_NEW &&
1452 ptlrpc_send_new_req(req)) {
1453 force_timer_recalc = 1;
1456 /* delayed send - skip */
1457 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1460 if (!(req->rq_phase == RQ_PHASE_RPC ||
1461 req->rq_phase == RQ_PHASE_BULK ||
1462 req->rq_phase == RQ_PHASE_INTERPRET ||
1463 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1464 req->rq_phase == RQ_PHASE_COMPLETE)) {
1465 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1469 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1470 LASSERT(req->rq_next_phase != req->rq_phase);
1471 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1474 * Skip processing until reply is unlinked. We
1475 * can't return to pool before that and we can't
1476 * call interpret before that. We need to make
1477 * sure that all rdma transfers finished and will
1478 * not corrupt any data.
1480 if (ptlrpc_client_recv_or_unlink(req) ||
1481 ptlrpc_client_bulk_active(req))
1485 * Turn fail_loc off to prevent it from looping
1488 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1489 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1492 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1493 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1498 * Move to next phase if reply was successfully
1501 ptlrpc_rqphase_move(req, req->rq_next_phase);
1504 if (req->rq_phase == RQ_PHASE_COMPLETE)
1507 if (req->rq_phase == RQ_PHASE_INTERPRET)
1508 GOTO(interpret, req->rq_status);
1511 * Note that this also will start async reply unlink.
1513 if (req->rq_net_err && !req->rq_timedout) {
1514 ptlrpc_expire_one_request(req, 1);
1517 * Check if we still need to wait for unlink.
1519 if (ptlrpc_client_recv_or_unlink(req) ||
1520 ptlrpc_client_bulk_active(req))
1522 /* If there is no need to resend, fail it now. */
1523 if (req->rq_no_resend) {
1524 if (req->rq_status == 0)
1525 req->rq_status = -EIO;
1526 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1527 GOTO(interpret, req->rq_status);
1534 cfs_spin_lock(&req->rq_lock);
1535 req->rq_replied = 0;
1536 cfs_spin_unlock(&req->rq_lock);
1537 if (req->rq_status == 0)
1538 req->rq_status = -EIO;
1539 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1540 GOTO(interpret, req->rq_status);
1543 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1544 * so it sets rq_intr regardless of individual rpc
1545 * timeouts. The synchronous IO waiting path sets
1546 * rq_intr irrespective of whether ptlrpcd
1547 * has seen a timeout. Our policy is to only interpret
1548 * interrupted rpcs after they have timed out, so we
1549 * need to enforce that here.
1552 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1553 req->rq_wait_ctx)) {
1554 req->rq_status = -EINTR;
1555 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1556 GOTO(interpret, req->rq_status);
1559 if (req->rq_phase == RQ_PHASE_RPC) {
1560 if (req->rq_timedout || req->rq_resend ||
1561 req->rq_waiting || req->rq_wait_ctx) {
1564 if (!ptlrpc_unregister_reply(req, 1))
1567 cfs_spin_lock(&imp->imp_lock);
1568 if (ptlrpc_import_delay_req(imp, req, &status)){
1569 /* put on delay list - only if we wait
1570 * recovery finished - before send */
1571 cfs_list_del_init(&req->rq_list);
1572 cfs_list_add_tail(&req->rq_list,
1575 cfs_spin_unlock(&imp->imp_lock);
1580 req->rq_status = status;
1581 ptlrpc_rqphase_move(req,
1582 RQ_PHASE_INTERPRET);
1583 cfs_spin_unlock(&imp->imp_lock);
1584 GOTO(interpret, req->rq_status);
1586 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1587 req->rq_status = -ENOTCONN;
1588 ptlrpc_rqphase_move(req,
1589 RQ_PHASE_INTERPRET);
1590 cfs_spin_unlock(&imp->imp_lock);
1591 GOTO(interpret, req->rq_status);
1594 cfs_list_del_init(&req->rq_list);
1595 cfs_list_add_tail(&req->rq_list,
1596 &imp->imp_sending_list);
1598 cfs_spin_unlock(&imp->imp_lock);
1600 cfs_spin_lock(&req->rq_lock);
1601 req->rq_waiting = 0;
1602 cfs_spin_unlock(&req->rq_lock);
1604 if (req->rq_timedout || req->rq_resend) {
1605 /* This is re-sending anyways,
1606 * let's mark req as resend. */
1607 cfs_spin_lock(&req->rq_lock);
1609 cfs_spin_unlock(&req->rq_lock);
1613 if (!ptlrpc_unregister_bulk(req, 1))
1616 /* ensure previous bulk fails */
1617 old_xid = req->rq_xid;
1618 req->rq_xid = ptlrpc_next_xid();
1619 CDEBUG(D_HA, "resend bulk "
1622 old_xid, req->rq_xid);
1626 * rq_wait_ctx is only touched by ptlrpcd,
1627 * so no lock is needed here.
1629 status = sptlrpc_req_refresh_ctx(req, -1);
1632 req->rq_status = status;
1633 cfs_spin_lock(&req->rq_lock);
1634 req->rq_wait_ctx = 0;
1635 cfs_spin_unlock(&req->rq_lock);
1636 force_timer_recalc = 1;
1638 cfs_spin_lock(&req->rq_lock);
1639 req->rq_wait_ctx = 1;
1640 cfs_spin_unlock(&req->rq_lock);
1645 cfs_spin_lock(&req->rq_lock);
1646 req->rq_wait_ctx = 0;
1647 cfs_spin_unlock(&req->rq_lock);
1650 rc = ptl_send_rpc(req, 0);
1652 DEBUG_REQ(D_HA, req, "send failed (%d)",
1654 force_timer_recalc = 1;
1655 cfs_spin_lock(&req->rq_lock);
1656 req->rq_net_err = 1;
1657 cfs_spin_unlock(&req->rq_lock);
1659 /* need to reset the timeout */
1660 force_timer_recalc = 1;
1663 cfs_spin_lock(&req->rq_lock);
1665 if (ptlrpc_client_early(req)) {
1666 ptlrpc_at_recv_early_reply(req);
1667 cfs_spin_unlock(&req->rq_lock);
1671 /* Still waiting for a reply? */
1672 if (ptlrpc_client_recv(req)) {
1673 cfs_spin_unlock(&req->rq_lock);
1677 /* Did we actually receive a reply? */
1678 if (!ptlrpc_client_replied(req)) {
1679 cfs_spin_unlock(&req->rq_lock);
1683 cfs_spin_unlock(&req->rq_lock);
1685 /* unlink from net because we are going to
1686 * swab in-place of reply buffer */
1687 unregistered = ptlrpc_unregister_reply(req, 1);
1691 req->rq_status = after_reply(req);
1695 /* If there is no bulk associated with this request,
1696 * then we're done and should let the interpreter
1697 * process the reply. Similarly if the RPC returned
1698 * an error, and therefore the bulk will never arrive.
1700 if (req->rq_bulk == NULL || req->rq_status < 0) {
1701 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1702 GOTO(interpret, req->rq_status);
1705 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1708 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1709 if (ptlrpc_client_bulk_active(req))
1712 if (!req->rq_bulk->bd_success) {
1713 /* The RPC reply arrived OK, but the bulk screwed
1714 * up! Dead weird since the server told us the RPC
1715 * was good after getting the REPLY for her GET or
1716 * the ACK for her PUT. */
1717 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1718 req->rq_status = -EIO;
1721 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1724 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1726 /* This moves to "unregistering" phase we need to wait for
1728 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1729 /* start async bulk unlink too */
1730 ptlrpc_unregister_bulk(req, 1);
1734 if (!ptlrpc_unregister_bulk(req, 1))
1737 /* When calling interpret receiving already should be
1739 LASSERT(!req->rq_receiving_reply);
1741 ptlrpc_req_interpret(env, req, req->rq_status);
1743 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1745 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1746 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1747 imp->imp_obd->obd_uuid.uuid,
1748 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1750 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1751 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1753 cfs_spin_lock(&imp->imp_lock);
1754 /* Request already may be not on sending or delaying list. This
1755 * may happen in the case of marking it erroneous for the case
1756 * ptlrpc_import_delay_req(req, status) find it impossible to
1757 * allow sending this rpc and returns *status != 0. */
1758 if (!cfs_list_empty(&req->rq_list)) {
1759 cfs_list_del_init(&req->rq_list);
1760 cfs_atomic_dec(&imp->imp_inflight);
1762 cfs_spin_unlock(&imp->imp_lock);
1764 cfs_atomic_dec(&set->set_remaining);
1765 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1768 /* If we hit an error, we want to recover promptly. */
1769 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1773 * Time out request \a req. is \a async_unlink is set, that means do not wait
1774 * until LNet actually confirms network buffer unlinking.
1775 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1777 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1779 struct obd_import *imp = req->rq_import;
1783 cfs_spin_lock(&req->rq_lock);
1784 req->rq_timedout = 1;
1785 cfs_spin_unlock(&req->rq_lock);
1787 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req, "Request "
1788 " sent has %s: [sent "CFS_DURATION_T"/"
1789 "real "CFS_DURATION_T"]",
1790 req->rq_net_err ? "failed due to network error" :
1791 ((req->rq_real_sent == 0 ||
1792 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1793 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1794 "timed out for sent delay" : "timed out for slow reply"),
1795 req->rq_sent, req->rq_real_sent);
1797 if (imp != NULL && obd_debug_peer_on_timeout)
1798 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1800 ptlrpc_unregister_reply(req, async_unlink);
1801 ptlrpc_unregister_bulk(req, async_unlink);
1803 if (obd_dump_on_timeout)
1804 libcfs_debug_dumplog();
1807 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1814 cfs_atomic_inc(&imp->imp_timeouts);
1816 /* The DLM server doesn't want recovery run on its imports. */
1817 if (imp->imp_dlm_fake)
1820 /* If this request is for recovery or other primordial tasks,
1821 * then error it out here. */
1822 if (req->rq_ctx_init || req->rq_ctx_fini ||
1823 req->rq_send_state != LUSTRE_IMP_FULL ||
1824 imp->imp_obd->obd_no_recov) {
1825 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1826 ptlrpc_import_state_name(req->rq_send_state),
1827 ptlrpc_import_state_name(imp->imp_state));
1828 cfs_spin_lock(&req->rq_lock);
1829 req->rq_status = -ETIMEDOUT;
1831 cfs_spin_unlock(&req->rq_lock);
1835 /* if a request can't be resent we can't wait for an answer after
1837 if (ptlrpc_no_resend(req)) {
1838 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1842 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1848 * Time out all uncompleted requests in request set pointed by \a data
1849 * Callback used when waiting on sets with l_wait_event.
1852 int ptlrpc_expired_set(void *data)
1854 struct ptlrpc_request_set *set = data;
1856 time_t now = cfs_time_current_sec();
1859 LASSERT(set != NULL);
1862 * A timeout expired. See which reqs it applies to...
1864 cfs_list_for_each (tmp, &set->set_requests) {
1865 struct ptlrpc_request *req =
1866 cfs_list_entry(tmp, struct ptlrpc_request,
1869 /* don't expire request waiting for context */
1870 if (req->rq_wait_ctx)
1873 /* Request in-flight? */
1874 if (!((req->rq_phase == RQ_PHASE_RPC &&
1875 !req->rq_waiting && !req->rq_resend) ||
1876 (req->rq_phase == RQ_PHASE_BULK)))
1879 if (req->rq_timedout || /* already dealt with */
1880 req->rq_deadline > now) /* not expired */
1883 /* Deal with this guy. Do it asynchronously to not block
1884 * ptlrpcd thread. */
1885 ptlrpc_expire_one_request(req, 1);
1889 * When waiting for a whole set, we always break out of the
1890 * sleep so we can recalculate the timeout, or enable interrupts
1891 * if everyone's timed out.
1897 * Sets rq_intr flag in \a req under spinlock.
1899 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1901 cfs_spin_lock(&req->rq_lock);
1903 cfs_spin_unlock(&req->rq_lock);
1907 * Interrupts (sets interrupted flag) all uncompleted requests in
1908 * a set \a data. Callback for l_wait_event for interruptible waits.
1910 void ptlrpc_interrupted_set(void *data)
1912 struct ptlrpc_request_set *set = data;
1915 LASSERT(set != NULL);
1916 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
1918 cfs_list_for_each(tmp, &set->set_requests) {
1919 struct ptlrpc_request *req =
1920 cfs_list_entry(tmp, struct ptlrpc_request,
1923 if (req->rq_phase != RQ_PHASE_RPC &&
1924 req->rq_phase != RQ_PHASE_UNREGISTERING)
1927 ptlrpc_mark_interrupted(req);
1932 * Get the smallest timeout in the set; this does NOT set a timeout.
1934 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1937 time_t now = cfs_time_current_sec();
1939 struct ptlrpc_request *req;
1943 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1945 cfs_list_for_each(tmp, &set->set_requests) {
1946 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1949 * Request in-flight?
1951 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1952 (req->rq_phase == RQ_PHASE_BULK) ||
1953 (req->rq_phase == RQ_PHASE_NEW)))
1957 * Already timed out.
1959 if (req->rq_timedout)
1965 if (req->rq_wait_ctx)
1968 if (req->rq_phase == RQ_PHASE_NEW)
1969 deadline = req->rq_sent;
1971 deadline = req->rq_sent + req->rq_timeout;
1973 if (deadline <= now) /* actually expired already */
1974 timeout = 1; /* ASAP */
1975 else if (timeout == 0 || timeout > deadline - now)
1976 timeout = deadline - now;
1982 * Send all unset request from the set and then wait untill all
1983 * requests in the set complete (either get a reply, timeout, get an
1984 * error or otherwise be interrupted).
1985 * Returns 0 on success or error code otherwise.
1987 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1990 struct ptlrpc_request *req;
1991 struct l_wait_info lwi;
1995 if (cfs_list_empty(&set->set_requests))
1998 cfs_list_for_each(tmp, &set->set_requests) {
1999 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2000 if (req->rq_phase == RQ_PHASE_NEW)
2001 (void)ptlrpc_send_new_req(req);
2005 timeout = ptlrpc_set_next_timeout(set);
2007 /* wait until all complete, interrupted, or an in-flight
2009 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2012 if (timeout == 0 && !cfs_signal_pending())
2014 * No requests are in-flight (ether timed out
2015 * or delayed), so we can allow interrupts.
2016 * We still want to block for a limited time,
2017 * so we allow interrupts during the timeout.
2019 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2021 ptlrpc_interrupted_set, set);
2024 * At least one request is in flight, so no
2025 * interrupts are allowed. Wait until all
2026 * complete, or an in-flight req times out.
2028 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2029 ptlrpc_expired_set, set);
2031 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2033 /* LU-769 - if we ignored the signal because it was already
2034 * pending when we started, we need to handle it now or we risk
2035 * it being ignored forever */
2036 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2037 cfs_signal_pending()) {
2038 cfs_sigset_t blocked_sigs =
2039 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2041 /* In fact we only interrupt for the "fatal" signals
2042 * like SIGINT or SIGKILL. We still ignore less
2043 * important signals since ptlrpc set is not easily
2044 * reentrant from userspace again */
2045 if (cfs_signal_pending())
2046 ptlrpc_interrupted_set(set);
2047 cfs_block_sigs(blocked_sigs);
2050 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2052 /* -EINTR => all requests have been flagged rq_intr so next
2054 * -ETIMEDOUT => someone timed out. When all reqs have
2055 * timed out, signals are enabled allowing completion with
2057 * I don't really care if we go once more round the loop in
2058 * the error cases -eeb. */
2059 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2060 cfs_list_for_each(tmp, &set->set_requests) {
2061 req = cfs_list_entry(tmp, struct ptlrpc_request,
2063 cfs_spin_lock(&req->rq_lock);
2064 req->rq_invalid_rqset = 1;
2065 cfs_spin_unlock(&req->rq_lock);
2068 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2070 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2073 cfs_list_for_each(tmp, &set->set_requests) {
2074 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2076 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2077 if (req->rq_status != 0)
2078 rc = req->rq_status;
2081 if (set->set_interpret != NULL) {
2082 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2084 rc = interpreter (set, set->set_arg, rc);
2086 struct ptlrpc_set_cbdata *cbdata, *n;
2089 cfs_list_for_each_entry_safe(cbdata, n,
2090 &set->set_cblist, psc_item) {
2091 cfs_list_del_init(&cbdata->psc_item);
2092 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2095 OBD_FREE_PTR(cbdata);
2103 * Helper fuction for request freeing.
2104 * Called when request count reached zero and request needs to be freed.
2105 * Removes request from all sorts of sending/replay lists it might be on,
2106 * frees network buffers if any are present.
2107 * If \a locked is set, that means caller is already holding import imp_lock
2108 * and so we no longer need to reobtain it (for certain lists manipulations)
2110 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2113 if (request == NULL) {
2118 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2119 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2120 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2121 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2122 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2123 LASSERTF(!request->rq_replay, "req %p\n", request);
2125 req_capsule_fini(&request->rq_pill);
2127 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2128 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2129 if (request->rq_import != NULL) {
2131 cfs_spin_lock(&request->rq_import->imp_lock);
2132 cfs_list_del_init(&request->rq_replay_list);
2134 cfs_spin_unlock(&request->rq_import->imp_lock);
2136 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2138 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2139 DEBUG_REQ(D_ERROR, request,
2140 "freeing request with nonzero refcount");
2144 if (request->rq_repbuf != NULL)
2145 sptlrpc_cli_free_repbuf(request);
2146 if (request->rq_export != NULL) {
2147 class_export_put(request->rq_export);
2148 request->rq_export = NULL;
2150 if (request->rq_import != NULL) {
2151 class_import_put(request->rq_import);
2152 request->rq_import = NULL;
2154 if (request->rq_bulk != NULL)
2155 ptlrpc_free_bulk(request->rq_bulk);
2157 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2158 sptlrpc_cli_free_reqbuf(request);
2160 if (request->rq_cli_ctx)
2161 sptlrpc_req_put_ctx(request, !locked);
2163 if (request->rq_pool)
2164 __ptlrpc_free_req_to_pool(request);
2166 OBD_FREE(request, sizeof(*request));
2170 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2172 * Drop one request reference. Must be called with import imp_lock held.
2173 * When reference count drops to zero, reuqest is freed.
2175 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2177 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2178 (void)__ptlrpc_req_finished(request, 1);
2183 * Drops one reference count for request \a request.
2184 * \a locked set indicates that caller holds import imp_lock.
2185 * Frees the request whe reference count reaches zero.
2187 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2190 if (request == NULL)
2193 if (request == LP_POISON ||
2194 request->rq_reqmsg == LP_POISON) {
2195 CERROR("dereferencing freed request (bug 575)\n");
2200 DEBUG_REQ(D_INFO, request, "refcount now %u",
2201 cfs_atomic_read(&request->rq_refcount) - 1);
2203 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2204 __ptlrpc_free_req(request, locked);
2212 * Drops one reference count for a request.
2214 void ptlrpc_req_finished(struct ptlrpc_request *request)
2216 __ptlrpc_req_finished(request, 0);
2220 * Returns xid of a \a request
2222 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2224 return request->rq_xid;
2226 EXPORT_SYMBOL(ptlrpc_req_xid);
2229 * Disengage the client's reply buffer from the network
2230 * NB does _NOT_ unregister any client-side bulk.
2231 * IDEMPOTENT, but _not_ safe against concurrent callers.
2232 * The request owner (i.e. the thread doing the I/O) must call...
2233 * Returns 0 on success or 1 if unregistering cannot be made.
2235 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2239 struct l_wait_info lwi;
2244 LASSERT(!cfs_in_interrupt());
2247 * Let's setup deadline for reply unlink.
2249 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2250 async && request->rq_reply_deadline == 0)
2251 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2254 * Nothing left to do.
2256 if (!ptlrpc_client_recv_or_unlink(request))
2259 LNetMDUnlink(request->rq_reply_md_h);
2262 * Let's check it once again.
2264 if (!ptlrpc_client_recv_or_unlink(request))
2268 * Move to "Unregistering" phase as reply was not unlinked yet.
2270 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2273 * Do not wait for unlink to finish.
2279 * We have to l_wait_event() whatever the result, to give liblustre
2280 * a chance to run reply_in_callback(), and to make sure we've
2281 * unlinked before returning a req to the pool.
2283 if (request->rq_set != NULL)
2284 wq = &request->rq_set->set_waitq;
2286 wq = &request->rq_reply_waitq;
2289 /* Network access will complete in finite time but the HUGE
2290 * timeout lets us CWARN for visibility of sluggish NALs */
2291 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2292 cfs_time_seconds(1), NULL, NULL);
2293 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2296 ptlrpc_rqphase_move(request, request->rq_next_phase);
2300 LASSERT(rc == -ETIMEDOUT);
2301 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2302 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2303 request->rq_must_unlink);
2309 * Iterates through replay_list on import and prunes
2310 * all requests have transno smaller than last_committed for the
2311 * import and don't have rq_replay set.
2312 * Since requests are sorted in transno order, stops when meetign first
2313 * transno bigger than last_committed.
2314 * caller must hold imp->imp_lock
2316 void ptlrpc_free_committed(struct obd_import *imp)
2318 cfs_list_t *tmp, *saved;
2319 struct ptlrpc_request *req;
2320 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2323 LASSERT(imp != NULL);
2325 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2328 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2329 imp->imp_generation == imp->imp_last_generation_checked) {
2330 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2331 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2335 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2336 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2337 imp->imp_generation);
2338 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2339 imp->imp_last_generation_checked = imp->imp_generation;
2341 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2342 req = cfs_list_entry(tmp, struct ptlrpc_request,
2345 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2346 LASSERT(req != last_req);
2349 if (req->rq_transno == 0) {
2350 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2353 if (req->rq_import_generation < imp->imp_generation) {
2354 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2358 if (req->rq_replay) {
2359 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2363 /* not yet committed */
2364 if (req->rq_transno > imp->imp_peer_committed_transno) {
2365 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2369 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2370 imp->imp_peer_committed_transno);
2372 cfs_spin_lock(&req->rq_lock);
2374 cfs_spin_unlock(&req->rq_lock);
2375 if (req->rq_commit_cb != NULL)
2376 req->rq_commit_cb(req);
2377 cfs_list_del_init(&req->rq_replay_list);
2378 __ptlrpc_req_finished(req, 1);
2385 void ptlrpc_cleanup_client(struct obd_import *imp)
2393 * Schedule previously sent request for resend.
2394 * For bulk requests we assign new xid (to avoid problems with
2395 * lost replies and therefore several transfers landing into same buffer
2396 * from different sending attempts).
2398 void ptlrpc_resend_req(struct ptlrpc_request *req)
2400 DEBUG_REQ(D_HA, req, "going to resend");
2401 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2402 req->rq_status = -EAGAIN;
2404 cfs_spin_lock(&req->rq_lock);
2406 req->rq_net_err = 0;
2407 req->rq_timedout = 0;
2409 __u64 old_xid = req->rq_xid;
2411 /* ensure previous bulk fails */
2412 req->rq_xid = ptlrpc_next_xid();
2413 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2414 old_xid, req->rq_xid);
2416 ptlrpc_client_wake_req(req);
2417 cfs_spin_unlock(&req->rq_lock);
2420 /* XXX: this function and rq_status are currently unused */
2421 void ptlrpc_restart_req(struct ptlrpc_request *req)
2423 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2424 req->rq_status = -ERESTARTSYS;
2426 cfs_spin_lock(&req->rq_lock);
2427 req->rq_restart = 1;
2428 req->rq_timedout = 0;
2429 ptlrpc_client_wake_req(req);
2430 cfs_spin_unlock(&req->rq_lock);
2434 * Grab additional reference on a request \a req
2436 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2439 cfs_atomic_inc(&req->rq_refcount);
2444 * Add a request to import replay_list.
2445 * Must be called under imp_lock
2447 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2448 struct obd_import *imp)
2452 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2454 if (req->rq_transno == 0) {
2455 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2459 /* clear this for new requests that were resent as well
2460 as resent replayed requests. */
2461 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2463 /* don't re-add requests that have been replayed */
2464 if (!cfs_list_empty(&req->rq_replay_list))
2467 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2469 LASSERT(imp->imp_replayable);
2470 /* Balanced in ptlrpc_free_committed, usually. */
2471 ptlrpc_request_addref(req);
2472 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2473 struct ptlrpc_request *iter =
2474 cfs_list_entry(tmp, struct ptlrpc_request,
2477 /* We may have duplicate transnos if we create and then
2478 * open a file, or for closes retained if to match creating
2479 * opens, so use req->rq_xid as a secondary key.
2480 * (See bugs 684, 685, and 428.)
2481 * XXX no longer needed, but all opens need transnos!
2483 if (iter->rq_transno > req->rq_transno)
2486 if (iter->rq_transno == req->rq_transno) {
2487 LASSERT(iter->rq_xid != req->rq_xid);
2488 if (iter->rq_xid > req->rq_xid)
2492 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2496 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2500 * Send request and wait until it completes.
2501 * Returns request processing status.
2503 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2505 struct ptlrpc_request_set *set;
2509 LASSERT(req->rq_set == NULL);
2510 LASSERT(!req->rq_receiving_reply);
2512 set = ptlrpc_prep_set();
2514 CERROR("Unable to allocate ptlrpc set.");
2518 /* for distributed debugging */
2519 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2521 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2522 ptlrpc_request_addref(req);
2523 ptlrpc_set_add_req(set, req);
2524 rc = ptlrpc_set_wait(set);
2525 ptlrpc_set_destroy(set);
2530 struct ptlrpc_replay_async_args {
2532 int praa_old_status;
2536 * Callback used for replayed requests reply processing.
2537 * In case of succesful reply calls registeresd request replay callback.
2538 * In case of error restart replay process.
2540 static int ptlrpc_replay_interpret(const struct lu_env *env,
2541 struct ptlrpc_request *req,
2542 void * data, int rc)
2544 struct ptlrpc_replay_async_args *aa = data;
2545 struct obd_import *imp = req->rq_import;
2548 cfs_atomic_dec(&imp->imp_replay_inflight);
2550 if (!ptlrpc_client_replied(req)) {
2551 CERROR("request replay timed out, restarting recovery\n");
2552 GOTO(out, rc = -ETIMEDOUT);
2555 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2556 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2557 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2558 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2560 /** VBR: check version failure */
2561 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2562 /** replay was failed due to version mismatch */
2563 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2564 cfs_spin_lock(&imp->imp_lock);
2565 imp->imp_vbr_failed = 1;
2566 imp->imp_no_lock_replay = 1;
2567 cfs_spin_unlock(&imp->imp_lock);
2568 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2570 /** The transno had better not change over replay. */
2571 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2572 lustre_msg_get_transno(req->rq_repmsg) ||
2573 lustre_msg_get_transno(req->rq_repmsg) == 0,
2575 lustre_msg_get_transno(req->rq_reqmsg),
2576 lustre_msg_get_transno(req->rq_repmsg));
2579 cfs_spin_lock(&imp->imp_lock);
2580 /** if replays by version then gap was occur on server, no trust to locks */
2581 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2582 imp->imp_no_lock_replay = 1;
2583 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2584 cfs_spin_unlock(&imp->imp_lock);
2585 LASSERT(imp->imp_last_replay_transno);
2587 /* transaction number shouldn't be bigger than the latest replayed */
2588 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2589 DEBUG_REQ(D_ERROR, req,
2590 "Reported transno "LPU64" is bigger than the "
2591 "replayed one: "LPU64, req->rq_transno,
2592 lustre_msg_get_transno(req->rq_reqmsg));
2593 GOTO(out, rc = -EINVAL);
2596 DEBUG_REQ(D_HA, req, "got rep");
2598 /* let the callback do fixups, possibly including in the request */
2599 if (req->rq_replay_cb)
2600 req->rq_replay_cb(req);
2602 if (ptlrpc_client_replied(req) &&
2603 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2604 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2605 lustre_msg_get_status(req->rq_repmsg),
2606 aa->praa_old_status);
2608 /* Put it back for re-replay. */
2609 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2613 * Errors while replay can set transno to 0, but
2614 * imp_last_replay_transno shouldn't be set to 0 anyway
2616 if (req->rq_transno == 0)
2617 CERROR("Transno is 0 during replay!\n");
2619 /* continue with recovery */
2620 rc = ptlrpc_import_recovery_state_machine(imp);
2622 req->rq_send_state = aa->praa_old_state;
2625 /* this replay failed, so restart recovery */
2626 ptlrpc_connect_import(imp);
2632 * Prepares and queues request for replay.
2633 * Adds it to ptlrpcd queue for actual sending.
2634 * Returns 0 on success.
2636 int ptlrpc_replay_req(struct ptlrpc_request *req)
2638 struct ptlrpc_replay_async_args *aa;
2641 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2643 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2644 aa = ptlrpc_req_async_args(req);
2645 memset(aa, 0, sizeof *aa);
2647 /* Prepare request to be resent with ptlrpcd */
2648 aa->praa_old_state = req->rq_send_state;
2649 req->rq_send_state = LUSTRE_IMP_REPLAY;
2650 req->rq_phase = RQ_PHASE_NEW;
2651 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2653 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2655 req->rq_interpret_reply = ptlrpc_replay_interpret;
2656 /* Readjust the timeout for current conditions */
2657 ptlrpc_at_set_req_timeout(req);
2659 /* Tell server the net_latency, so the server can calculate how long
2660 * it should wait for next replay */
2661 lustre_msg_set_service_time(req->rq_reqmsg,
2662 ptlrpc_at_get_net_latency(req));
2663 DEBUG_REQ(D_HA, req, "REPLAY");
2665 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2666 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2668 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2673 * Aborts all in-flight request on import \a imp sending and delayed lists
2675 void ptlrpc_abort_inflight(struct obd_import *imp)
2677 cfs_list_t *tmp, *n;
2680 /* Make sure that no new requests get processed for this import.
2681 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2682 * this flag and then putting requests on sending_list or delayed_list.
2684 cfs_spin_lock(&imp->imp_lock);
2686 /* XXX locking? Maybe we should remove each request with the list
2687 * locked? Also, how do we know if the requests on the list are
2688 * being freed at this time?
2690 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2691 struct ptlrpc_request *req =
2692 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2694 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2696 cfs_spin_lock (&req->rq_lock);
2697 if (req->rq_import_generation < imp->imp_generation) {
2699 req->rq_status = -EINTR;
2700 ptlrpc_client_wake_req(req);
2702 cfs_spin_unlock (&req->rq_lock);
2705 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2706 struct ptlrpc_request *req =
2707 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2709 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2711 cfs_spin_lock (&req->rq_lock);
2712 if (req->rq_import_generation < imp->imp_generation) {
2714 req->rq_status = -EINTR;
2715 ptlrpc_client_wake_req(req);
2717 cfs_spin_unlock (&req->rq_lock);
2720 /* Last chance to free reqs left on the replay list, but we
2721 * will still leak reqs that haven't committed. */
2722 if (imp->imp_replayable)
2723 ptlrpc_free_committed(imp);
2725 cfs_spin_unlock(&imp->imp_lock);
2731 * Abort all uncompleted requests in request set \a set
2733 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2735 cfs_list_t *tmp, *pos;
2737 LASSERT(set != NULL);
2739 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2740 struct ptlrpc_request *req =
2741 cfs_list_entry(pos, struct ptlrpc_request,
2744 cfs_spin_lock(&req->rq_lock);
2745 if (req->rq_phase != RQ_PHASE_RPC) {
2746 cfs_spin_unlock(&req->rq_lock);
2751 req->rq_status = -EINTR;
2752 ptlrpc_client_wake_req(req);
2753 cfs_spin_unlock(&req->rq_lock);
2757 static __u64 ptlrpc_last_xid;
2758 static cfs_spinlock_t ptlrpc_last_xid_lock;
2761 * Initialize the XID for the node. This is common among all requests on
2762 * this node, and only requires the property that it is monotonically
2763 * increasing. It does not need to be sequential. Since this is also used
2764 * as the RDMA match bits, it is important that a single client NOT have
2765 * the same match bits for two different in-flight requests, hence we do
2766 * NOT want to have an XID per target or similar.
2768 * To avoid an unlikely collision between match bits after a client reboot
2769 * (which would deliver old data into the wrong RDMA buffer) initialize
2770 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2771 * If the time is clearly incorrect, we instead use a 62-bit random number.
2772 * In the worst case the random number will overflow 1M RPCs per second in
2773 * 9133 years, or permutations thereof.
2775 #define YEAR_2004 (1ULL << 30)
2776 void ptlrpc_init_xid(void)
2778 time_t now = cfs_time_current_sec();
2780 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2781 if (now < YEAR_2004) {
2782 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2783 ptlrpc_last_xid >>= 2;
2784 ptlrpc_last_xid |= (1ULL << 61);
2786 ptlrpc_last_xid = (__u64)now << 20;
2791 * Increase xid and returns resultng new value to the caller.
2793 __u64 ptlrpc_next_xid(void)
2796 cfs_spin_lock(&ptlrpc_last_xid_lock);
2797 tmp = ++ptlrpc_last_xid;
2798 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2803 * Get a glimpse at what next xid value might have been.
2804 * Returns possible next xid.
2806 __u64 ptlrpc_sample_next_xid(void)
2808 #if BITS_PER_LONG == 32
2809 /* need to avoid possible word tearing on 32-bit systems */
2811 cfs_spin_lock(&ptlrpc_last_xid_lock);
2812 tmp = ptlrpc_last_xid + 1;
2813 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2816 /* No need to lock, since returned value is racy anyways */
2817 return ptlrpc_last_xid + 1;
2820 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2823 * Functions for operating ptlrpc workers.
2825 * A ptlrpc work is a function which will be running inside ptlrpc context.
2826 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2828 * 1. after a work is created, it can be used many times, that is:
2829 * handler = ptlrpcd_alloc_work();
2830 * ptlrpcd_queue_work();
2832 * queue it again when necessary:
2833 * ptlrpcd_queue_work();
2834 * ptlrpcd_destroy_work();
2835 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2836 * it will only be queued once in any time. Also as its name implies, it may
2837 * have delay before it really runs by ptlrpcd thread.
2839 struct ptlrpc_work_async_args {
2841 int (*cb)(const struct lu_env *, void *);
2845 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2847 static int work_interpreter(const struct lu_env *env,
2848 struct ptlrpc_request *req, void *data, int rc)
2850 struct ptlrpc_work_async_args *arg = data;
2852 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2853 LASSERT(arg->cb != NULL);
2855 return arg->cb(env, arg->cbdata);
2859 * Create a work for ptlrpc.
2861 void *ptlrpcd_alloc_work(struct obd_import *imp,
2862 int (*cb)(const struct lu_env *, void *), void *cbdata)
2864 struct ptlrpc_request *req = NULL;
2865 struct ptlrpc_work_async_args *args;
2871 RETURN(ERR_PTR(-EINVAL));
2873 /* copy some code from deprecated fakereq. */
2876 CERROR("ptlrpc: run out of memory!\n");
2877 RETURN(ERR_PTR(-ENOMEM));
2880 req->rq_send_state = LUSTRE_IMP_FULL;
2881 req->rq_type = PTL_RPC_MSG_REQUEST;
2882 req->rq_import = class_import_get(imp);
2883 req->rq_export = NULL;
2884 req->rq_interpret_reply = work_interpreter;
2885 /* don't want reply */
2886 req->rq_receiving_reply = 0;
2887 req->rq_must_unlink = 0;
2888 req->rq_no_delay = req->rq_no_resend = 1;
2890 cfs_spin_lock_init(&req->rq_lock);
2891 CFS_INIT_LIST_HEAD(&req->rq_list);
2892 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
2893 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
2894 CFS_INIT_LIST_HEAD(&req->rq_history_list);
2895 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
2896 cfs_waitq_init(&req->rq_reply_waitq);
2897 cfs_waitq_init(&req->rq_set_waitq);
2898 cfs_atomic_set(&req->rq_refcount, 1);
2900 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
2901 args = ptlrpc_req_async_args(req);
2902 args->magic = PTLRPC_WORK_MAGIC;
2904 args->cbdata = cbdata;
2908 EXPORT_SYMBOL(ptlrpcd_alloc_work);
2910 void ptlrpcd_destroy_work(void *handler)
2912 struct ptlrpc_request *req = handler;
2915 ptlrpc_req_finished(req);
2917 EXPORT_SYMBOL(ptlrpcd_destroy_work);
2919 int ptlrpcd_queue_work(void *handler)
2921 struct ptlrpc_request *req = handler;
2924 * Check if the req is already being queued.
2926 * Here comes a trick: it lacks a way of checking if a req is being
2927 * processed reliably in ptlrpc. Here I have to use refcount of req
2928 * for this purpose. This is okay because the caller should use this
2929 * req as opaque data. - Jinshan
2931 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
2932 if (cfs_atomic_read(&req->rq_refcount) > 1)
2935 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
2936 cfs_atomic_dec(&req->rq_refcount);
2940 /* re-initialize the req */
2941 req->rq_timeout = obd_timeout;
2942 req->rq_sent = cfs_time_current_sec();
2943 req->rq_deadline = req->rq_sent + req->rq_timeout;
2944 req->rq_reply_deadline = req->rq_deadline;
2945 req->rq_phase = RQ_PHASE_INTERPRET;
2946 req->rq_next_phase = RQ_PHASE_COMPLETE;
2947 req->rq_xid = ptlrpc_next_xid();
2948 req->rq_import_generation = req->rq_import->imp_generation;
2950 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2953 EXPORT_SYMBOL(ptlrpcd_queue_work);