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 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 /** Implementation of client-side PortalRPC interfaces */
39 #define DEBUG_SUBSYSTEM S_RPC
43 #include <liblustre.h>
46 #include <obd_support.h>
47 #include <obd_class.h>
48 #include <lustre_lib.h>
49 #include <lustre_ha.h>
50 #include <lustre_import.h>
51 #include <lustre_req_layout.h>
53 #include "ptlrpc_internal.h"
56 * Initialize passed in client structure \a cl.
58 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
59 struct ptlrpc_client *cl)
61 cl->cli_request_portal = req_portal;
62 cl->cli_reply_portal = rep_portal;
67 * Return PortalRPC connection for remore uud \a uuid
69 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
71 struct ptlrpc_connection *c;
73 lnet_process_id_t peer;
76 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
78 CDEBUG(D_NETERROR, "cannot find peer %s!\n", uuid->uuid);
82 c = ptlrpc_connection_get(peer, self, uuid);
84 memcpy(c->c_remote_uuid.uuid,
85 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
88 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
94 * Allocate and initialize new bulk descriptor
95 * Returns pointer to the descriptor or NULL on error.
97 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
99 struct ptlrpc_bulk_desc *desc;
101 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
105 cfs_spin_lock_init(&desc->bd_lock);
106 cfs_waitq_init(&desc->bd_waitq);
107 desc->bd_max_iov = npages;
108 desc->bd_iov_count = 0;
109 LNetInvalidateHandle(&desc->bd_md_h);
110 desc->bd_portal = portal;
111 desc->bd_type = type;
117 * Prepare bulk descriptor for specified outgoing request \a req that
118 * can fit \a npages * pages. \a type is bulk type. \a portal is where
119 * the bulk to be sent. Used on client-side.
120 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
123 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
124 int npages, int type, int portal)
126 struct obd_import *imp = req->rq_import;
127 struct ptlrpc_bulk_desc *desc;
130 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
131 desc = new_bulk(npages, type, portal);
135 desc->bd_import_generation = req->rq_import_generation;
136 desc->bd_import = class_import_get(imp);
139 desc->bd_cbid.cbid_fn = client_bulk_callback;
140 desc->bd_cbid.cbid_arg = desc;
142 /* This makes req own desc, and free it when she frees herself */
149 * Prepare bulk descriptor for specified incoming request \a req that
150 * can fit \a npages * pages. \a type is bulk type. \a portal is where
151 * the bulk to be sent. Used on server-side after request was already
153 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
156 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
157 int npages, int type, int portal)
159 struct obd_export *exp = req->rq_export;
160 struct ptlrpc_bulk_desc *desc;
163 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
165 desc = new_bulk(npages, type, portal);
169 desc->bd_export = class_export_get(exp);
172 desc->bd_cbid.cbid_fn = server_bulk_callback;
173 desc->bd_cbid.cbid_arg = desc;
175 /* NB we don't assign rq_bulk here; server-side requests are
176 * re-used, and the handler frees the bulk desc explicitly. */
182 * Add a page \a page to the bulk descriptor \a desc.
183 * Data to transfer in the page starts at offset \a pageoffset and
184 * amount of data to transfer from the page is \a len
186 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
187 cfs_page_t *page, int pageoffset, int len)
189 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
190 LASSERT(page != NULL);
191 LASSERT(pageoffset >= 0);
193 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
197 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
201 * Uninitialize and free bulk descriptor \a desc.
202 * Works on bulk descriptors both from server and client side.
204 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
208 LASSERT(desc != NULL);
209 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
210 LASSERT(!desc->bd_network_rw); /* network hands off or */
211 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
213 sptlrpc_enc_pool_put_pages(desc);
216 class_export_put(desc->bd_export);
218 class_import_put(desc->bd_import);
220 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
221 bd_iov[desc->bd_max_iov]));
226 * Set server timelimit for this req, i.e. how long are we willing to wait
227 * for reply before timing out this request.
229 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
235 LASSERT(req->rq_import);
238 /* non-AT settings */
240 * \a imp_server_timeout means this is reverse import and
241 * we send (currently only) ASTs to the client and cannot afford
242 * to wait too long for the reply, otherwise the other client
243 * (because of which we are sending this request) would
244 * timeout waiting for us
246 req->rq_timeout = req->rq_import->imp_server_timeout ?
247 obd_timeout / 2 : obd_timeout;
249 at = &req->rq_import->imp_at;
250 idx = import_at_get_index(req->rq_import,
251 req->rq_request_portal);
252 serv_est = at_get(&at->iat_service_estimate[idx]);
253 req->rq_timeout = at_est2timeout(serv_est);
255 /* We could get even fancier here, using history to predict increased
258 /* Let the server know what this RPC timeout is by putting it in the
260 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
263 /* Adjust max service estimate based on server value */
264 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
265 unsigned int serv_est)
271 LASSERT(req->rq_import);
272 at = &req->rq_import->imp_at;
274 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
275 /* max service estimates are tracked on the server side,
276 so just keep minimal history here */
277 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
279 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
280 "has changed from %d to %d\n",
281 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
282 oldse, at_get(&at->iat_service_estimate[idx]));
285 /* Expected network latency per remote node (secs) */
286 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
288 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
291 /* Adjust expected network latency */
292 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
293 unsigned int service_time)
295 unsigned int nl, oldnl;
297 time_t now = cfs_time_current_sec();
299 LASSERT(req->rq_import);
300 at = &req->rq_import->imp_at;
302 /* Network latency is total time less server processing time */
303 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
304 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
305 CWARN("Reported service time %u > total measured time "
306 CFS_DURATION_T"\n", service_time,
307 cfs_time_sub(now, req->rq_sent));
309 oldnl = at_measured(&at->iat_net_latency, nl);
311 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
312 "has changed from %d to %d\n",
313 req->rq_import->imp_obd->obd_name,
315 &req->rq_import->imp_connection->c_remote_uuid),
316 oldnl, at_get(&at->iat_net_latency));
319 static int unpack_reply(struct ptlrpc_request *req)
323 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
324 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
326 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
331 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
333 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
340 * Handle an early reply message, called with the rq_lock held.
341 * If anything goes wrong just ignore it - same as if it never happened
343 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
345 struct ptlrpc_request *early_req;
351 cfs_spin_unlock(&req->rq_lock);
353 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
355 cfs_spin_lock(&req->rq_lock);
359 rc = unpack_reply(early_req);
361 /* Expecting to increase the service time estimate here */
362 ptlrpc_at_adj_service(req,
363 lustre_msg_get_timeout(early_req->rq_repmsg));
364 ptlrpc_at_adj_net_latency(req,
365 lustre_msg_get_service_time(early_req->rq_repmsg));
368 sptlrpc_cli_finish_early_reply(early_req);
370 cfs_spin_lock(&req->rq_lock);
373 /* Adjust the local timeout for this req */
374 ptlrpc_at_set_req_timeout(req);
376 olddl = req->rq_deadline;
377 /* server assumes it now has rq_timeout from when it sent the
378 early reply, so client should give it at least that long. */
379 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
380 ptlrpc_at_get_net_latency(req);
382 DEBUG_REQ(D_ADAPTTO, req,
383 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
384 "("CFS_DURATION_T"s)", req->rq_early_count,
385 cfs_time_sub(req->rq_deadline,
386 cfs_time_current_sec()),
387 cfs_time_sub(req->rq_deadline, olddl));
394 * Wind down request pool \a pool.
395 * Frees all requests from the pool too
397 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
400 struct ptlrpc_request *req;
402 LASSERT(pool != NULL);
404 cfs_spin_lock(&pool->prp_lock);
405 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
406 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
407 cfs_list_del(&req->rq_list);
408 LASSERT(req->rq_reqbuf);
409 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
410 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
411 OBD_FREE(req, sizeof(*req));
413 cfs_spin_unlock(&pool->prp_lock);
414 OBD_FREE(pool, sizeof(*pool));
418 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
420 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
425 while (size < pool->prp_rq_size)
428 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
429 size == pool->prp_rq_size,
430 "Trying to change pool size with nonempty pool "
431 "from %d to %d bytes\n", pool->prp_rq_size, size);
433 cfs_spin_lock(&pool->prp_lock);
434 pool->prp_rq_size = size;
435 for (i = 0; i < num_rq; i++) {
436 struct ptlrpc_request *req;
437 struct lustre_msg *msg;
439 cfs_spin_unlock(&pool->prp_lock);
440 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
443 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
445 OBD_FREE(req, sizeof(struct ptlrpc_request));
448 req->rq_reqbuf = msg;
449 req->rq_reqbuf_len = size;
451 cfs_spin_lock(&pool->prp_lock);
452 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
454 cfs_spin_unlock(&pool->prp_lock);
459 * Create and initialize new request pool with given attributes:
460 * \a num_rq - initial number of requests to create for the pool
461 * \a msgsize - maximum message size possible for requests in thid pool
462 * \a populate_pool - function to be called when more requests need to be added
464 * Returns pointer to newly created pool or NULL on error.
466 struct ptlrpc_request_pool *
467 ptlrpc_init_rq_pool(int num_rq, int msgsize,
468 void (*populate_pool)(struct ptlrpc_request_pool *, int))
470 struct ptlrpc_request_pool *pool;
472 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
476 /* Request next power of two for the allocation, because internally
477 kernel would do exactly this */
479 cfs_spin_lock_init(&pool->prp_lock);
480 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
481 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
482 pool->prp_populate = populate_pool;
484 populate_pool(pool, num_rq);
486 if (cfs_list_empty(&pool->prp_req_list)) {
487 /* have not allocated a single request for the pool */
488 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
495 * Fetches one request from pool \a pool
497 static struct ptlrpc_request *
498 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
500 struct ptlrpc_request *request;
501 struct lustre_msg *reqbuf;
506 cfs_spin_lock(&pool->prp_lock);
508 /* See if we have anything in a pool, and bail out if nothing,
509 * in writeout path, where this matters, this is safe to do, because
510 * nothing is lost in this case, and when some in-flight requests
511 * complete, this code will be called again. */
512 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
513 cfs_spin_unlock(&pool->prp_lock);
517 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
519 cfs_list_del_init(&request->rq_list);
520 cfs_spin_unlock(&pool->prp_lock);
522 LASSERT(request->rq_reqbuf);
523 LASSERT(request->rq_pool);
525 reqbuf = request->rq_reqbuf;
526 memset(request, 0, sizeof(*request));
527 request->rq_reqbuf = reqbuf;
528 request->rq_reqbuf_len = pool->prp_rq_size;
529 request->rq_pool = pool;
535 * Returns freed \a request to pool.
537 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
539 struct ptlrpc_request_pool *pool = request->rq_pool;
541 cfs_spin_lock(&pool->prp_lock);
542 LASSERT(cfs_list_empty(&request->rq_list));
543 LASSERT(!request->rq_receiving_reply);
544 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
545 cfs_spin_unlock(&pool->prp_lock);
548 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
549 __u32 version, int opcode,
550 int count, __u32 *lengths, char **bufs,
551 struct ptlrpc_cli_ctx *ctx)
553 struct obd_import *imp = request->rq_import;
558 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
560 rc = sptlrpc_req_get_ctx(request);
565 sptlrpc_req_set_flavor(request, opcode);
567 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
570 LASSERT(!request->rq_pool);
574 lustre_msg_add_version(request->rq_reqmsg, version);
575 request->rq_send_state = LUSTRE_IMP_FULL;
576 request->rq_type = PTL_RPC_MSG_REQUEST;
577 request->rq_export = NULL;
579 request->rq_req_cbid.cbid_fn = request_out_callback;
580 request->rq_req_cbid.cbid_arg = request;
582 request->rq_reply_cbid.cbid_fn = reply_in_callback;
583 request->rq_reply_cbid.cbid_arg = request;
585 request->rq_reply_deadline = 0;
586 request->rq_phase = RQ_PHASE_NEW;
587 request->rq_next_phase = RQ_PHASE_UNDEFINED;
589 request->rq_request_portal = imp->imp_client->cli_request_portal;
590 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
592 ptlrpc_at_set_req_timeout(request);
594 cfs_spin_lock_init(&request->rq_lock);
595 CFS_INIT_LIST_HEAD(&request->rq_list);
596 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
597 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
598 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
599 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
600 CFS_INIT_LIST_HEAD(&request->rq_history_list);
601 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
602 cfs_waitq_init(&request->rq_reply_waitq);
603 cfs_waitq_init(&request->rq_set_waitq);
604 request->rq_xid = ptlrpc_next_xid();
605 cfs_atomic_set(&request->rq_refcount, 1);
607 lustre_msg_set_opc(request->rq_reqmsg, opcode);
611 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
613 class_import_put(imp);
617 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
618 __u32 version, int opcode, char **bufs,
619 struct ptlrpc_cli_ctx *ctx)
623 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
624 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
625 request->rq_pill.rc_area[RCL_CLIENT],
628 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
631 * Pack request buffers for network transfer, performing necessary encryption
632 * steps if necessary.
634 int ptlrpc_request_pack(struct ptlrpc_request *request,
635 __u32 version, int opcode)
637 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
641 * Helper function to allocate new request on import \a imp
642 * and possibly using existing request from pool \a pool if provided.
643 * Returns allocated request structure with import field filled or
647 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
648 struct ptlrpc_request_pool *pool)
650 struct ptlrpc_request *request = NULL;
653 request = ptlrpc_prep_req_from_pool(pool);
656 OBD_ALLOC_PTR(request);
659 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
660 LASSERT(imp != LP_POISON);
661 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
663 LASSERT(imp->imp_client != LP_POISON);
665 request->rq_import = class_import_get(imp);
667 CERROR("request allocation out of memory\n");
674 * Helper function for creating a request.
675 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
676 * buffer structures according to capsule template \a format.
677 * Returns allocated request structure pointer or NULL on error.
679 static struct ptlrpc_request *
680 ptlrpc_request_alloc_internal(struct obd_import *imp,
681 struct ptlrpc_request_pool * pool,
682 const struct req_format *format)
684 struct ptlrpc_request *request;
686 request = __ptlrpc_request_alloc(imp, pool);
690 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
691 req_capsule_set(&request->rq_pill, format);
696 * Allocate new request structure for import \a imp and initialize its
697 * buffer structure according to capsule template \a format.
699 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
700 const struct req_format *format)
702 return ptlrpc_request_alloc_internal(imp, NULL, format);
706 * Allocate new request structure for import \a imp from pool \a pool and
707 * initialize its buffer structure according to capsule template \a format.
709 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
710 struct ptlrpc_request_pool * pool,
711 const struct req_format *format)
713 return ptlrpc_request_alloc_internal(imp, pool, format);
717 * For requests not from pool, free memory of the request structure.
718 * For requests obtained from a pool earlier, return request back to pool.
720 void ptlrpc_request_free(struct ptlrpc_request *request)
722 if (request->rq_pool)
723 __ptlrpc_free_req_to_pool(request);
725 OBD_FREE_PTR(request);
729 * Allocate new request for operatione \a opcode and immediatelly pack it for
731 * Only used for simple requests like OBD_PING where the only important
732 * part of the request is operation itself.
733 * Returns allocated request or NULL on error.
735 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
736 const struct req_format *format,
737 __u32 version, int opcode)
739 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
743 rc = ptlrpc_request_pack(req, version, opcode);
745 ptlrpc_request_free(req);
753 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
754 * for operation \a opcode. Request would contain \a count buffers.
755 * Sizes of buffers are described in array \a lengths and buffers themselves
756 * are provided by a pointer \a bufs.
757 * Returns prepared request structure pointer or NULL on error.
759 struct ptlrpc_request *
760 ptlrpc_prep_req_pool(struct obd_import *imp,
761 __u32 version, int opcode,
762 int count, __u32 *lengths, char **bufs,
763 struct ptlrpc_request_pool *pool)
765 struct ptlrpc_request *request;
768 request = __ptlrpc_request_alloc(imp, pool);
772 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
773 lengths, bufs, NULL);
775 ptlrpc_request_free(request);
782 * Same as ptlrpc_prep_req_pool, but without pool
784 struct ptlrpc_request *
785 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
786 __u32 *lengths, char **bufs)
788 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
793 * Allocate "fake" request that would not be sent anywhere in the end.
794 * Only used as a hack because we have no other way of performing
795 * async actions in lustre between layers.
796 * Used on MDS to request object preallocations from more than one OST at a
799 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
800 unsigned int timeout,
801 ptlrpc_interpterer_t interpreter)
803 struct ptlrpc_request *request = NULL;
806 OBD_ALLOC(request, sizeof(*request));
808 CERROR("request allocation out of memory\n");
812 request->rq_send_state = LUSTRE_IMP_FULL;
813 request->rq_type = PTL_RPC_MSG_REQUEST;
814 request->rq_import = class_import_get(imp);
815 request->rq_export = NULL;
816 request->rq_import_generation = imp->imp_generation;
818 request->rq_timeout = timeout;
819 request->rq_sent = cfs_time_current_sec();
820 request->rq_deadline = request->rq_sent + timeout;
821 request->rq_reply_deadline = request->rq_deadline;
822 request->rq_interpret_reply = interpreter;
823 request->rq_phase = RQ_PHASE_RPC;
824 request->rq_next_phase = RQ_PHASE_INTERPRET;
825 /* don't want reply */
826 request->rq_receiving_reply = 0;
827 request->rq_must_unlink = 0;
828 request->rq_no_delay = request->rq_no_resend = 1;
829 request->rq_fake = 1;
831 cfs_spin_lock_init(&request->rq_lock);
832 CFS_INIT_LIST_HEAD(&request->rq_list);
833 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
834 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
835 CFS_INIT_LIST_HEAD(&request->rq_history_list);
836 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
837 cfs_waitq_init(&request->rq_reply_waitq);
838 cfs_waitq_init(&request->rq_set_waitq);
840 request->rq_xid = ptlrpc_next_xid();
841 cfs_atomic_set(&request->rq_refcount, 1);
847 * Indicate that processing of "fake" request is finished.
849 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
851 /* if we kill request before timeout - need adjust counter */
852 if (req->rq_phase == RQ_PHASE_RPC) {
853 struct ptlrpc_request_set *set = req->rq_set;
856 cfs_atomic_dec(&set->set_remaining);
859 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
860 cfs_list_del_init(&req->rq_list);
864 * Allocate and initialize new request set structure.
865 * Returns a pointer to the newly allocated set structure or NULL on error.
867 struct ptlrpc_request_set *ptlrpc_prep_set(void)
869 struct ptlrpc_request_set *set;
872 OBD_ALLOC(set, sizeof *set);
875 CFS_INIT_LIST_HEAD(&set->set_requests);
876 cfs_waitq_init(&set->set_waitq);
877 cfs_atomic_set(&set->set_remaining, 0);
878 cfs_spin_lock_init(&set->set_new_req_lock);
879 CFS_INIT_LIST_HEAD(&set->set_new_requests);
880 CFS_INIT_LIST_HEAD(&set->set_cblist);
886 * Wind down and free request set structure previously allocated with
888 * Ensures that all requests on the set have completed and removes
889 * all requests from the request list in a set.
890 * If any unsent request happen to be on the list, pretends that they got
891 * an error in flight and calls their completion handler.
893 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
901 /* Requests on the set should either all be completed, or all be new */
902 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
903 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
904 cfs_list_for_each (tmp, &set->set_requests) {
905 struct ptlrpc_request *req =
906 cfs_list_entry(tmp, struct ptlrpc_request,
909 LASSERT(req->rq_phase == expected_phase);
913 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
914 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
915 cfs_atomic_read(&set->set_remaining), n);
917 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
918 struct ptlrpc_request *req =
919 cfs_list_entry(tmp, struct ptlrpc_request,
921 cfs_list_del_init(&req->rq_set_chain);
923 LASSERT(req->rq_phase == expected_phase);
925 if (req->rq_phase == RQ_PHASE_NEW) {
926 ptlrpc_req_interpret(NULL, req, -EBADR);
927 cfs_atomic_dec(&set->set_remaining);
930 cfs_spin_lock(&req->rq_lock);
932 req->rq_invalid_rqset = 0;
933 cfs_spin_unlock(&req->rq_lock);
935 ptlrpc_req_finished (req);
938 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
940 OBD_FREE(set, sizeof(*set));
945 * Add a callback function \a fn to the set.
946 * This function would be called when all requests on this set are completed.
947 * The function will be passed \a data argument.
949 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
950 set_interpreter_func fn, void *data)
952 struct ptlrpc_set_cbdata *cbdata;
954 OBD_ALLOC_PTR(cbdata);
958 cbdata->psc_interpret = fn;
959 cbdata->psc_data = data;
960 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
966 * Add a new request to the general purpose request set.
967 * Assumes request reference from the caller.
969 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
970 struct ptlrpc_request *req)
972 /* The set takes over the caller's request reference */
973 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
975 cfs_atomic_inc(&set->set_remaining);
976 req->rq_queued_time = cfs_time_current(); /* Where is the best place to set this? */
980 * Add a request to a request with dedicated server thread
981 * and wake the thread to make any necessary processing.
982 * Currently only used for ptlrpcd.
983 * Returns 0 if succesful or non zero error code on error.
984 * (the only possible error for now is if the dedicated server thread
987 int ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
988 struct ptlrpc_request *req)
990 struct ptlrpc_request_set *set = pc->pc_set;
993 * Let caller know that we stopped and will not handle this request.
994 * It needs to take care itself of request.
996 if (cfs_test_bit(LIOD_STOP, &pc->pc_flags))
999 cfs_spin_lock(&set->set_new_req_lock);
1001 * The set takes over the caller's request reference.
1003 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1005 cfs_spin_unlock(&set->set_new_req_lock);
1007 cfs_waitq_signal(&set->set_waitq);
1012 * Based on the current state of the import, determine if the request
1013 * can be sent, is an error, or should be delayed.
1015 * Returns true if this request should be delayed. If false, and
1016 * *status is set, then the request can not be sent and *status is the
1017 * error code. If false and status is 0, then request can be sent.
1019 * The imp->imp_lock must be held.
1021 static int ptlrpc_import_delay_req(struct obd_import *imp,
1022 struct ptlrpc_request *req, int *status)
1027 LASSERT (status != NULL);
1030 if (req->rq_ctx_init || req->rq_ctx_fini) {
1031 /* always allow ctx init/fini rpc go through */
1032 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1033 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1036 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1037 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1039 } else if (imp->imp_obd->obd_no_recov) {
1040 *status = -ESHUTDOWN;
1041 } else if (ptlrpc_send_limit_expired(req)) {
1042 /* probably doesn't need to be a D_ERROR after initial testing */
1043 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1045 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1046 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1047 /* allow CONNECT even if import is invalid */ ;
1048 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1049 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1052 } else if (imp->imp_invalid) {
1053 if (!imp->imp_deactive)
1054 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1055 *status = -ESHUTDOWN; /* bz 12940 */
1056 } else if (req->rq_import_generation != imp->imp_generation) {
1057 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1059 } else if (req->rq_send_state != imp->imp_state) {
1060 /* invalidate in progress - any requests should be drop */
1061 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1062 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1064 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1065 *status = -EWOULDBLOCK;
1075 * Decide if the eror message regarding provided request \a req
1076 * should be printed to the console or not.
1077 * Makes it's decision on request status and other properties.
1078 * Returns 1 to print error on the system console or 0 if not.
1080 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1082 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1085 /* Suppress particular reconnect errors which are to be expected. No
1086 * errors are suppressed for the initial connection on an import */
1087 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1088 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1090 /* Suppress timed out reconnect requests */
1091 if (req->rq_timedout)
1094 /* Suppress unavailable/again reconnect requests */
1095 err = lustre_msg_get_status(req->rq_repmsg);
1096 if (err == -ENODEV || err == -EAGAIN)
1104 * Check request processing status.
1105 * Returns the status.
1107 static int ptlrpc_check_status(struct ptlrpc_request *req)
1112 err = lustre_msg_get_status(req->rq_repmsg);
1113 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1114 struct obd_import *imp = req->rq_import;
1115 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1116 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
1117 " with %s. The %s operation failed with %d\n",
1118 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1119 ll_opcode2str(opc), err);
1120 RETURN(err < 0 ? err : -EINVAL);
1124 DEBUG_REQ(D_INFO, req, "status is %d", err);
1125 } else if (err > 0) {
1126 /* XXX: translate this error from net to host */
1127 DEBUG_REQ(D_INFO, req, "status is %d", err);
1130 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1131 struct obd_import *imp = req->rq_import;
1132 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1134 if (ptlrpc_console_allow(req))
1135 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1136 "communicating with %s. The %s "
1137 "operation failed with %d\n",
1139 imp->imp_connection->c_peer.nid),
1140 ll_opcode2str(opc), err);
1142 RETURN(err < 0 ? err : -EINVAL);
1149 * save pre-versions of objects into request for replay.
1150 * Versions are obtained from server reply.
1153 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1155 struct lustre_msg *repmsg = req->rq_repmsg;
1156 struct lustre_msg *reqmsg = req->rq_reqmsg;
1157 __u64 *versions = lustre_msg_get_versions(repmsg);
1160 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1164 lustre_msg_set_versions(reqmsg, versions);
1165 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1166 versions[0], versions[1]);
1172 * Callback function called when client receives RPC reply for \a req.
1173 * Returns 0 on success or error code.
1174 * The return alue would be assigned to req->rq_status by the caller
1175 * as request processing status.
1176 * This function also decides if the request needs to be saved for later replay.
1178 static int after_reply(struct ptlrpc_request *req)
1180 struct obd_import *imp = req->rq_import;
1181 struct obd_device *obd = req->rq_import->imp_obd;
1183 struct timeval work_start;
1187 LASSERT(obd != NULL);
1188 /* repbuf must be unlinked */
1189 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1191 if (req->rq_reply_truncate) {
1192 if (ptlrpc_no_resend(req)) {
1193 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1194 " expected: %d, actual size: %d",
1195 req->rq_nob_received, req->rq_repbuf_len);
1199 sptlrpc_cli_free_repbuf(req);
1200 /* Pass the required reply buffer size (include
1201 * space for early reply).
1202 * NB: no need to roundup because alloc_repbuf
1203 * will roundup it */
1204 req->rq_replen = req->rq_nob_received;
1205 req->rq_nob_received = 0;
1211 * NB Until this point, the whole of the incoming message,
1212 * including buflens, status etc is in the sender's byte order.
1214 rc = sptlrpc_cli_unwrap_reply(req);
1216 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1221 * Security layer unwrap might ask resend this request.
1226 rc = unpack_reply(req);
1230 cfs_gettimeofday(&work_start);
1231 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1232 if (obd->obd_svc_stats != NULL) {
1233 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1235 ptlrpc_lprocfs_rpc_sent(req, timediff);
1238 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1239 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1240 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1241 lustre_msg_get_type(req->rq_repmsg));
1245 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1246 OBD_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, obd_fail_val);
1247 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1248 ptlrpc_at_adj_net_latency(req,
1249 lustre_msg_get_service_time(req->rq_repmsg));
1251 rc = ptlrpc_check_status(req);
1252 imp->imp_connect_error = rc;
1256 * Either we've been evicted, or the server has failed for
1257 * some reason. Try to reconnect, and if that fails, punt to
1260 if (ll_rpc_recoverable_error(rc)) {
1261 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1262 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1265 ptlrpc_request_handle_notconn(req);
1270 * Let's look if server sent slv. Do it only for RPC with
1273 ldlm_cli_update_pool(req);
1277 * Store transno in reqmsg for replay.
1279 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1280 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1281 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1284 if (imp->imp_replayable) {
1285 cfs_spin_lock(&imp->imp_lock);
1287 * No point in adding already-committed requests to the replay
1288 * list, we will just remove them immediately. b=9829
1290 if (req->rq_transno != 0 &&
1292 lustre_msg_get_last_committed(req->rq_repmsg) ||
1294 /** version recovery */
1295 ptlrpc_save_versions(req);
1296 ptlrpc_retain_replayable_request(req, imp);
1297 } else if (req->rq_commit_cb != NULL) {
1298 cfs_spin_unlock(&imp->imp_lock);
1299 req->rq_commit_cb(req);
1300 cfs_spin_lock(&imp->imp_lock);
1304 * Replay-enabled imports return commit-status information.
1306 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1307 imp->imp_peer_committed_transno =
1308 lustre_msg_get_last_committed(req->rq_repmsg);
1310 ptlrpc_free_committed(imp);
1311 cfs_spin_unlock(&imp->imp_lock);
1318 * Helper function to send request \a req over the network for the first time
1319 * Also adjusts request phase.
1320 * Returns 0 on success or error code.
1322 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1324 struct obd_import *imp;
1328 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1329 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()))
1332 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1334 imp = req->rq_import;
1335 cfs_spin_lock(&imp->imp_lock);
1337 req->rq_import_generation = imp->imp_generation;
1339 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1340 cfs_spin_lock(&req->rq_lock);
1341 req->rq_waiting = 1;
1342 cfs_spin_unlock(&req->rq_lock);
1344 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1345 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1346 ptlrpc_import_state_name(req->rq_send_state),
1347 ptlrpc_import_state_name(imp->imp_state));
1348 LASSERT(cfs_list_empty(&req->rq_list));
1349 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1350 cfs_atomic_inc(&req->rq_import->imp_inflight);
1351 cfs_spin_unlock(&imp->imp_lock);
1356 cfs_spin_unlock(&imp->imp_lock);
1357 req->rq_status = rc;
1358 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1362 LASSERT(cfs_list_empty(&req->rq_list));
1363 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1364 cfs_atomic_inc(&req->rq_import->imp_inflight);
1365 cfs_spin_unlock(&imp->imp_lock);
1367 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1369 rc = sptlrpc_req_refresh_ctx(req, -1);
1372 req->rq_status = rc;
1375 req->rq_wait_ctx = 1;
1380 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1381 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1382 imp->imp_obd->obd_uuid.uuid,
1383 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1384 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1385 lustre_msg_get_opc(req->rq_reqmsg));
1387 rc = ptl_send_rpc(req, 0);
1389 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1390 req->rq_net_err = 1;
1397 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1398 * and no more replies are expected.
1399 * (it is possible to get less replies than requests sent e.g. due to timed out
1400 * requests or requests that we had trouble to send out)
1402 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1405 int force_timer_recalc = 0;
1408 if (cfs_atomic_read(&set->set_remaining) == 0)
1411 cfs_list_for_each(tmp, &set->set_requests) {
1412 struct ptlrpc_request *req =
1413 cfs_list_entry(tmp, struct ptlrpc_request,
1415 struct obd_import *imp = req->rq_import;
1416 int unregistered = 0;
1419 if (req->rq_phase == RQ_PHASE_NEW &&
1420 ptlrpc_send_new_req(req)) {
1421 force_timer_recalc = 1;
1424 /* delayed send - skip */
1425 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1428 if (!(req->rq_phase == RQ_PHASE_RPC ||
1429 req->rq_phase == RQ_PHASE_BULK ||
1430 req->rq_phase == RQ_PHASE_INTERPRET ||
1431 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1432 req->rq_phase == RQ_PHASE_COMPLETE)) {
1433 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1437 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1438 LASSERT(req->rq_next_phase != req->rq_phase);
1439 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1442 * Skip processing until reply is unlinked. We
1443 * can't return to pool before that and we can't
1444 * call interpret before that. We need to make
1445 * sure that all rdma transfers finished and will
1446 * not corrupt any data.
1448 if (ptlrpc_client_recv_or_unlink(req) ||
1449 ptlrpc_client_bulk_active(req))
1453 * Turn fail_loc off to prevent it from looping
1456 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1457 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1460 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1461 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1466 * Move to next phase if reply was successfully
1469 ptlrpc_rqphase_move(req, req->rq_next_phase);
1472 if (req->rq_phase == RQ_PHASE_COMPLETE)
1475 if (req->rq_phase == RQ_PHASE_INTERPRET)
1476 GOTO(interpret, req->rq_status);
1479 * Note that this also will start async reply unlink.
1481 if (req->rq_net_err && !req->rq_timedout) {
1482 ptlrpc_expire_one_request(req, 1);
1485 * Check if we still need to wait for unlink.
1487 if (ptlrpc_client_recv_or_unlink(req) ||
1488 ptlrpc_client_bulk_active(req))
1493 cfs_spin_lock(&req->rq_lock);
1494 req->rq_replied = 0;
1495 cfs_spin_unlock(&req->rq_lock);
1496 if (req->rq_status == 0)
1497 req->rq_status = -EIO;
1498 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1499 GOTO(interpret, req->rq_status);
1502 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1503 * so it sets rq_intr regardless of individual rpc
1504 * timeouts. The synchronous IO waiting path sets
1505 * rq_intr irrespective of whether ptlrpcd
1506 * has seen a timeout. Our policy is to only interpret
1507 * interrupted rpcs after they have timed out, so we
1508 * need to enforce that here.
1511 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1512 req->rq_wait_ctx)) {
1513 req->rq_status = -EINTR;
1514 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1515 GOTO(interpret, req->rq_status);
1518 if (req->rq_phase == RQ_PHASE_RPC) {
1519 if (req->rq_timedout || req->rq_resend ||
1520 req->rq_waiting || req->rq_wait_ctx) {
1523 if (!ptlrpc_unregister_reply(req, 1))
1526 cfs_spin_lock(&imp->imp_lock);
1527 if (ptlrpc_import_delay_req(imp, req, &status)){
1528 /* put on delay list - only if we wait
1529 * recovery finished - before send */
1530 cfs_list_del_init(&req->rq_list);
1531 cfs_list_add_tail(&req->rq_list,
1534 cfs_spin_unlock(&imp->imp_lock);
1539 req->rq_status = status;
1540 ptlrpc_rqphase_move(req,
1541 RQ_PHASE_INTERPRET);
1542 cfs_spin_unlock(&imp->imp_lock);
1543 GOTO(interpret, req->rq_status);
1545 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1546 req->rq_status = -ENOTCONN;
1547 ptlrpc_rqphase_move(req,
1548 RQ_PHASE_INTERPRET);
1549 cfs_spin_unlock(&imp->imp_lock);
1550 GOTO(interpret, req->rq_status);
1553 cfs_list_del_init(&req->rq_list);
1554 cfs_list_add_tail(&req->rq_list,
1555 &imp->imp_sending_list);
1557 cfs_spin_unlock(&imp->imp_lock);
1559 cfs_spin_lock(&req->rq_lock);
1560 req->rq_waiting = 0;
1561 cfs_spin_unlock(&req->rq_lock);
1563 if (req->rq_timedout || req->rq_resend) {
1564 /* This is re-sending anyways,
1565 * let's mark req as resend. */
1566 cfs_spin_lock(&req->rq_lock);
1568 cfs_spin_unlock(&req->rq_lock);
1572 if (!ptlrpc_unregister_bulk(req, 1))
1575 /* ensure previous bulk fails */
1576 old_xid = req->rq_xid;
1577 req->rq_xid = ptlrpc_next_xid();
1578 CDEBUG(D_HA, "resend bulk "
1581 old_xid, req->rq_xid);
1585 * rq_wait_ctx is only touched by ptlrpcd,
1586 * so no lock is needed here.
1588 status = sptlrpc_req_refresh_ctx(req, -1);
1591 req->rq_status = status;
1592 cfs_spin_lock(&req->rq_lock);
1593 req->rq_wait_ctx = 0;
1594 cfs_spin_unlock(&req->rq_lock);
1595 force_timer_recalc = 1;
1597 cfs_spin_lock(&req->rq_lock);
1598 req->rq_wait_ctx = 1;
1599 cfs_spin_unlock(&req->rq_lock);
1604 cfs_spin_lock(&req->rq_lock);
1605 req->rq_wait_ctx = 0;
1606 cfs_spin_unlock(&req->rq_lock);
1609 rc = ptl_send_rpc(req, 0);
1611 DEBUG_REQ(D_HA, req, "send failed (%d)",
1613 force_timer_recalc = 1;
1614 cfs_spin_lock(&req->rq_lock);
1615 req->rq_net_err = 1;
1616 cfs_spin_unlock(&req->rq_lock);
1618 /* need to reset the timeout */
1619 force_timer_recalc = 1;
1622 cfs_spin_lock(&req->rq_lock);
1624 if (ptlrpc_client_early(req)) {
1625 ptlrpc_at_recv_early_reply(req);
1626 cfs_spin_unlock(&req->rq_lock);
1630 /* Still waiting for a reply? */
1631 if (ptlrpc_client_recv(req)) {
1632 cfs_spin_unlock(&req->rq_lock);
1636 /* Did we actually receive a reply? */
1637 if (!ptlrpc_client_replied(req)) {
1638 cfs_spin_unlock(&req->rq_lock);
1642 cfs_spin_unlock(&req->rq_lock);
1644 /* unlink from net because we are going to
1645 * swab in-place of reply buffer */
1646 unregistered = ptlrpc_unregister_reply(req, 1);
1650 req->rq_status = after_reply(req);
1654 /* If there is no bulk associated with this request,
1655 * then we're done and should let the interpreter
1656 * process the reply. Similarly if the RPC returned
1657 * an error, and therefore the bulk will never arrive.
1659 if (req->rq_bulk == NULL || req->rq_status != 0) {
1660 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1661 GOTO(interpret, req->rq_status);
1664 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1667 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1668 if (ptlrpc_client_bulk_active(req))
1671 if (!req->rq_bulk->bd_success) {
1672 /* The RPC reply arrived OK, but the bulk screwed
1673 * up! Dead weird since the server told us the RPC
1674 * was good after getting the REPLY for her GET or
1675 * the ACK for her PUT. */
1676 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1680 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1683 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1685 /* This moves to "unregistering" phase we need to wait for
1687 if (!unregistered && !ptlrpc_unregister_reply(req, 1))
1690 if (!ptlrpc_unregister_bulk(req, 1))
1693 /* When calling interpret receiving already should be
1695 LASSERT(!req->rq_receiving_reply);
1697 ptlrpc_req_interpret(env, req, req->rq_status);
1699 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1701 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1702 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1703 imp->imp_obd->obd_uuid.uuid,
1704 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1706 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1707 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1709 cfs_spin_lock(&imp->imp_lock);
1710 /* Request already may be not on sending or delaying list. This
1711 * may happen in the case of marking it erroneous for the case
1712 * ptlrpc_import_delay_req(req, status) find it impossible to
1713 * allow sending this rpc and returns *status != 0. */
1714 if (!cfs_list_empty(&req->rq_list)) {
1715 cfs_list_del_init(&req->rq_list);
1716 cfs_atomic_dec(&imp->imp_inflight);
1718 cfs_spin_unlock(&imp->imp_lock);
1720 cfs_atomic_dec(&set->set_remaining);
1721 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1724 /* If we hit an error, we want to recover promptly. */
1725 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1729 * Time out request \a req. is \a async_unlink is set, that means do not wait
1730 * until LNet actually confirms network buffer unlinking.
1731 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1733 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1735 struct obd_import *imp = req->rq_import;
1739 cfs_spin_lock(&req->rq_lock);
1740 req->rq_timedout = 1;
1741 cfs_spin_unlock(&req->rq_lock);
1743 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req,
1744 "Request x"LPU64" sent from %s to NID %s "CFS_DURATION_T"s "
1745 "ago has %s ("CFS_DURATION_T"s prior to deadline).\n",
1746 req->rq_xid, imp ? imp->imp_obd->obd_name : "<?>",
1747 imp ? libcfs_nid2str(imp->imp_connection->c_peer.nid) : "<?>",
1748 cfs_time_sub(cfs_time_current_sec(), req->rq_sent),
1749 req->rq_net_err ? "failed due to network error" : "timed out",
1750 cfs_time_sub(req->rq_deadline, req->rq_sent));
1752 if (imp != NULL && obd_debug_peer_on_timeout)
1753 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1755 ptlrpc_unregister_reply(req, async_unlink);
1756 ptlrpc_unregister_bulk(req, async_unlink);
1758 if (obd_dump_on_timeout)
1759 libcfs_debug_dumplog();
1762 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1769 cfs_atomic_inc(&imp->imp_timeouts);
1771 /* The DLM server doesn't want recovery run on its imports. */
1772 if (imp->imp_dlm_fake)
1775 /* If this request is for recovery or other primordial tasks,
1776 * then error it out here. */
1777 if (req->rq_ctx_init || req->rq_ctx_fini ||
1778 req->rq_send_state != LUSTRE_IMP_FULL ||
1779 imp->imp_obd->obd_no_recov) {
1780 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1781 ptlrpc_import_state_name(req->rq_send_state),
1782 ptlrpc_import_state_name(imp->imp_state));
1783 cfs_spin_lock(&req->rq_lock);
1784 req->rq_status = -ETIMEDOUT;
1786 cfs_spin_unlock(&req->rq_lock);
1790 /* if a request can't be resent we can't wait for an answer after
1792 if (ptlrpc_no_resend(req)) {
1793 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1797 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1803 * Time out all uncompleted requests in request set pointed by \a data
1804 * Callback used when waiting on sets with l_wait_event.
1807 int ptlrpc_expired_set(void *data)
1809 struct ptlrpc_request_set *set = data;
1811 time_t now = cfs_time_current_sec();
1814 LASSERT(set != NULL);
1817 * A timeout expired. See which reqs it applies to...
1819 cfs_list_for_each (tmp, &set->set_requests) {
1820 struct ptlrpc_request *req =
1821 cfs_list_entry(tmp, struct ptlrpc_request,
1824 /* don't expire request waiting for context */
1825 if (req->rq_wait_ctx)
1828 /* Request in-flight? */
1829 if (!((req->rq_phase == RQ_PHASE_RPC &&
1830 !req->rq_waiting && !req->rq_resend) ||
1831 (req->rq_phase == RQ_PHASE_BULK)))
1834 if (req->rq_timedout || /* already dealt with */
1835 req->rq_deadline > now) /* not expired */
1838 /* Deal with this guy. Do it asynchronously to not block
1839 * ptlrpcd thread. */
1840 ptlrpc_expire_one_request(req, 1);
1844 * When waiting for a whole set, we always break out of the
1845 * sleep so we can recalculate the timeout, or enable interrupts
1846 * if everyone's timed out.
1852 * Sets rq_intr flag in \a req under spinlock.
1854 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1856 cfs_spin_lock(&req->rq_lock);
1858 cfs_spin_unlock(&req->rq_lock);
1862 * Interrupts (sets interrupted flag) all uncompleted requests in
1863 * a set \a data. Callback for l_wait_event for interruptible waits.
1865 void ptlrpc_interrupted_set(void *data)
1867 struct ptlrpc_request_set *set = data;
1870 LASSERT(set != NULL);
1871 CERROR("INTERRUPTED SET %p\n", set);
1873 cfs_list_for_each(tmp, &set->set_requests) {
1874 struct ptlrpc_request *req =
1875 cfs_list_entry(tmp, struct ptlrpc_request,
1878 if (req->rq_phase != RQ_PHASE_RPC &&
1879 req->rq_phase != RQ_PHASE_UNREGISTERING)
1882 ptlrpc_mark_interrupted(req);
1887 * Get the smallest timeout in the set; this does NOT set a timeout.
1889 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1892 time_t now = cfs_time_current_sec();
1894 struct ptlrpc_request *req;
1898 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1900 cfs_list_for_each(tmp, &set->set_requests) {
1901 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1904 * Request in-flight?
1906 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1907 (req->rq_phase == RQ_PHASE_BULK) ||
1908 (req->rq_phase == RQ_PHASE_NEW)))
1912 * Already timed out.
1914 if (req->rq_timedout)
1920 if (req->rq_wait_ctx)
1923 if (req->rq_phase == RQ_PHASE_NEW)
1924 deadline = req->rq_sent;
1926 deadline = req->rq_sent + req->rq_timeout;
1928 if (deadline <= now) /* actually expired already */
1929 timeout = 1; /* ASAP */
1930 else if (timeout == 0 || timeout > deadline - now)
1931 timeout = deadline - now;
1937 * Send all unset request from the set and then wait untill all
1938 * requests in the set complete (either get a reply, timeout, get an
1939 * error or otherwise be interrupted).
1940 * Returns 0 on success or error code otherwise.
1942 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1945 struct ptlrpc_request *req;
1946 struct l_wait_info lwi;
1950 if (cfs_list_empty(&set->set_requests))
1953 cfs_list_for_each(tmp, &set->set_requests) {
1954 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1955 if (req->rq_phase == RQ_PHASE_NEW)
1956 (void)ptlrpc_send_new_req(req);
1960 timeout = ptlrpc_set_next_timeout(set);
1962 /* wait until all complete, interrupted, or an in-flight
1964 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1967 if (timeout == 0 && !cfs_signal_pending())
1969 * No requests are in-flight (ether timed out
1970 * or delayed), so we can allow interrupts.
1971 * We still want to block for a limited time,
1972 * so we allow interrupts during the timeout.
1974 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
1976 ptlrpc_interrupted_set, set);
1979 * At least one request is in flight, so no
1980 * interrupts are allowed. Wait until all
1981 * complete, or an in-flight req times out.
1983 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
1984 ptlrpc_expired_set, set);
1986 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
1988 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1990 /* -EINTR => all requests have been flagged rq_intr so next
1992 * -ETIMEDOUT => someone timed out. When all reqs have
1993 * timed out, signals are enabled allowing completion with
1995 * I don't really care if we go once more round the loop in
1996 * the error cases -eeb. */
1997 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
1998 cfs_list_for_each(tmp, &set->set_requests) {
1999 req = cfs_list_entry(tmp, struct ptlrpc_request,
2001 cfs_spin_lock(&req->rq_lock);
2002 req->rq_invalid_rqset = 1;
2003 cfs_spin_unlock(&req->rq_lock);
2006 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2008 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2011 cfs_list_for_each(tmp, &set->set_requests) {
2012 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2014 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2015 if (req->rq_status != 0)
2016 rc = req->rq_status;
2019 if (set->set_interpret != NULL) {
2020 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2022 rc = interpreter (set, set->set_arg, rc);
2024 struct ptlrpc_set_cbdata *cbdata, *n;
2027 cfs_list_for_each_entry_safe(cbdata, n,
2028 &set->set_cblist, psc_item) {
2029 cfs_list_del_init(&cbdata->psc_item);
2030 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2033 OBD_FREE_PTR(cbdata);
2041 * Helper fuction for request freeing.
2042 * Called when request count reached zero and request needs to be freed.
2043 * Removes request from all sorts of sending/replay lists it might be on,
2044 * frees network buffers if any are present.
2045 * If \a locked is set, that means caller is already holding import imp_lock
2046 * and so we no longer need to reobtain it (for certain lists manipulations)
2048 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2051 if (request == NULL) {
2056 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2057 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2058 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2059 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2060 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2061 LASSERTF(!request->rq_replay, "req %p\n", request);
2062 LASSERT(request->rq_cli_ctx || request->rq_fake);
2064 req_capsule_fini(&request->rq_pill);
2066 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2067 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2068 if (request->rq_import != NULL) {
2070 cfs_spin_lock(&request->rq_import->imp_lock);
2071 cfs_list_del_init(&request->rq_replay_list);
2073 cfs_spin_unlock(&request->rq_import->imp_lock);
2075 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2077 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2078 DEBUG_REQ(D_ERROR, request,
2079 "freeing request with nonzero refcount");
2083 if (request->rq_repbuf != NULL)
2084 sptlrpc_cli_free_repbuf(request);
2085 if (request->rq_export != NULL) {
2086 class_export_put(request->rq_export);
2087 request->rq_export = NULL;
2089 if (request->rq_import != NULL) {
2090 class_import_put(request->rq_import);
2091 request->rq_import = NULL;
2093 if (request->rq_bulk != NULL)
2094 ptlrpc_free_bulk(request->rq_bulk);
2096 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2097 sptlrpc_cli_free_reqbuf(request);
2099 if (request->rq_cli_ctx)
2100 sptlrpc_req_put_ctx(request, !locked);
2102 if (request->rq_pool)
2103 __ptlrpc_free_req_to_pool(request);
2105 OBD_FREE(request, sizeof(*request));
2109 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2111 * Drop one request reference. Must be called with import imp_lock held.
2112 * When reference count drops to zero, reuqest is freed.
2114 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2116 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2117 (void)__ptlrpc_req_finished(request, 1);
2122 * Drops one reference count for request \a request.
2123 * \a locked set indicates that caller holds import imp_lock.
2124 * Frees the request whe reference count reaches zero.
2126 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2129 if (request == NULL)
2132 if (request == LP_POISON ||
2133 request->rq_reqmsg == LP_POISON) {
2134 CERROR("dereferencing freed request (bug 575)\n");
2139 DEBUG_REQ(D_INFO, request, "refcount now %u",
2140 cfs_atomic_read(&request->rq_refcount) - 1);
2142 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2143 __ptlrpc_free_req(request, locked);
2151 * Drops one reference count for a request.
2153 void ptlrpc_req_finished(struct ptlrpc_request *request)
2155 __ptlrpc_req_finished(request, 0);
2159 * Returns xid of a \a request
2161 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2163 return request->rq_xid;
2165 EXPORT_SYMBOL(ptlrpc_req_xid);
2168 * Disengage the client's reply buffer from the network
2169 * NB does _NOT_ unregister any client-side bulk.
2170 * IDEMPOTENT, but _not_ safe against concurrent callers.
2171 * The request owner (i.e. the thread doing the I/O) must call...
2172 * Returns 0 on success or 1 if unregistering cannot be made.
2174 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2178 struct l_wait_info lwi;
2183 LASSERT(!cfs_in_interrupt());
2186 * Let's setup deadline for reply unlink.
2188 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2189 async && request->rq_reply_deadline == 0)
2190 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2193 * Nothing left to do.
2195 if (!ptlrpc_client_recv_or_unlink(request))
2198 LNetMDUnlink(request->rq_reply_md_h);
2201 * Let's check it once again.
2203 if (!ptlrpc_client_recv_or_unlink(request))
2207 * Move to "Unregistering" phase as reply was not unlinked yet.
2209 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2212 * Do not wait for unlink to finish.
2218 * We have to l_wait_event() whatever the result, to give liblustre
2219 * a chance to run reply_in_callback(), and to make sure we've
2220 * unlinked before returning a req to the pool.
2222 if (request->rq_set != NULL)
2223 wq = &request->rq_set->set_waitq;
2225 wq = &request->rq_reply_waitq;
2228 /* Network access will complete in finite time but the HUGE
2229 * timeout lets us CWARN for visibility of sluggish NALs */
2230 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2231 cfs_time_seconds(1), NULL, NULL);
2232 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2235 ptlrpc_rqphase_move(request, request->rq_next_phase);
2239 LASSERT(rc == -ETIMEDOUT);
2240 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2241 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2242 request->rq_must_unlink);
2248 * Iterates through replay_list on import and prunes
2249 * all requests have transno smaller than last_committed for the
2250 * import and don't have rq_replay set.
2251 * Since requests are sorted in transno order, stops when meetign first
2252 * transno bigger than last_committed.
2253 * caller must hold imp->imp_lock
2255 void ptlrpc_free_committed(struct obd_import *imp)
2257 cfs_list_t *tmp, *saved;
2258 struct ptlrpc_request *req;
2259 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2262 LASSERT(imp != NULL);
2264 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2267 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2268 imp->imp_generation == imp->imp_last_generation_checked) {
2269 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
2270 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2274 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2275 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2276 imp->imp_generation);
2277 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2278 imp->imp_last_generation_checked = imp->imp_generation;
2280 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2281 req = cfs_list_entry(tmp, struct ptlrpc_request,
2284 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2285 LASSERT(req != last_req);
2288 if (req->rq_transno == 0) {
2289 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2292 if (req->rq_import_generation < imp->imp_generation) {
2293 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2297 if (req->rq_replay) {
2298 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2302 /* not yet committed */
2303 if (req->rq_transno > imp->imp_peer_committed_transno) {
2304 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2308 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
2309 imp->imp_peer_committed_transno);
2311 cfs_spin_lock(&req->rq_lock);
2313 cfs_spin_unlock(&req->rq_lock);
2314 if (req->rq_commit_cb != NULL)
2315 req->rq_commit_cb(req);
2316 cfs_list_del_init(&req->rq_replay_list);
2317 __ptlrpc_req_finished(req, 1);
2324 void ptlrpc_cleanup_client(struct obd_import *imp)
2332 * Schedule previously sent request for resend.
2333 * For bulk requests we assign new xid (to avoid problems with
2334 * lost replies and therefore several transfers landing into same buffer
2335 * from different sending attempts).
2337 void ptlrpc_resend_req(struct ptlrpc_request *req)
2339 DEBUG_REQ(D_HA, req, "going to resend");
2340 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2341 req->rq_status = -EAGAIN;
2343 cfs_spin_lock(&req->rq_lock);
2345 req->rq_net_err = 0;
2346 req->rq_timedout = 0;
2348 __u64 old_xid = req->rq_xid;
2350 /* ensure previous bulk fails */
2351 req->rq_xid = ptlrpc_next_xid();
2352 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2353 old_xid, req->rq_xid);
2355 ptlrpc_client_wake_req(req);
2356 cfs_spin_unlock(&req->rq_lock);
2359 /* XXX: this function and rq_status are currently unused */
2360 void ptlrpc_restart_req(struct ptlrpc_request *req)
2362 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2363 req->rq_status = -ERESTARTSYS;
2365 cfs_spin_lock(&req->rq_lock);
2366 req->rq_restart = 1;
2367 req->rq_timedout = 0;
2368 ptlrpc_client_wake_req(req);
2369 cfs_spin_unlock(&req->rq_lock);
2373 * Grab additional reference on a request \a req
2375 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2378 cfs_atomic_inc(&req->rq_refcount);
2383 * Add a request to import replay_list.
2384 * Must be called under imp_lock
2386 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2387 struct obd_import *imp)
2391 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2393 if (req->rq_transno == 0) {
2394 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2398 /* clear this for new requests that were resent as well
2399 as resent replayed requests. */
2400 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2402 /* don't re-add requests that have been replayed */
2403 if (!cfs_list_empty(&req->rq_replay_list))
2406 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2408 LASSERT(imp->imp_replayable);
2409 /* Balanced in ptlrpc_free_committed, usually. */
2410 ptlrpc_request_addref(req);
2411 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2412 struct ptlrpc_request *iter =
2413 cfs_list_entry(tmp, struct ptlrpc_request,
2416 /* We may have duplicate transnos if we create and then
2417 * open a file, or for closes retained if to match creating
2418 * opens, so use req->rq_xid as a secondary key.
2419 * (See bugs 684, 685, and 428.)
2420 * XXX no longer needed, but all opens need transnos!
2422 if (iter->rq_transno > req->rq_transno)
2425 if (iter->rq_transno == req->rq_transno) {
2426 LASSERT(iter->rq_xid != req->rq_xid);
2427 if (iter->rq_xid > req->rq_xid)
2431 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2435 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2439 * Send request and wait until it completes.
2440 * Returns request processing status.
2442 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2444 struct ptlrpc_request_set *set;
2448 LASSERT(req->rq_set == NULL);
2449 LASSERT(!req->rq_receiving_reply);
2451 set = ptlrpc_prep_set();
2453 CERROR("Unable to allocate ptlrpc set.");
2457 /* for distributed debugging */
2458 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2460 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2461 ptlrpc_request_addref(req);
2462 ptlrpc_set_add_req(set, req);
2463 rc = ptlrpc_set_wait(set);
2464 ptlrpc_set_destroy(set);
2469 struct ptlrpc_replay_async_args {
2471 int praa_old_status;
2475 * Callback used for replayed requests reply processing.
2476 * In case of succesful reply calls registeresd request replay callback.
2477 * In case of error restart replay process.
2479 static int ptlrpc_replay_interpret(const struct lu_env *env,
2480 struct ptlrpc_request *req,
2481 void * data, int rc)
2483 struct ptlrpc_replay_async_args *aa = data;
2484 struct obd_import *imp = req->rq_import;
2487 cfs_atomic_dec(&imp->imp_replay_inflight);
2489 if (!ptlrpc_client_replied(req)) {
2490 CERROR("request replay timed out, restarting recovery\n");
2491 GOTO(out, rc = -ETIMEDOUT);
2494 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2495 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2496 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2497 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2499 /** VBR: check version failure */
2500 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2501 /** replay was failed due to version mismatch */
2502 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2503 cfs_spin_lock(&imp->imp_lock);
2504 imp->imp_vbr_failed = 1;
2505 imp->imp_no_lock_replay = 1;
2506 cfs_spin_unlock(&imp->imp_lock);
2508 /** The transno had better not change over replay. */
2509 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2510 lustre_msg_get_transno(req->rq_repmsg) ||
2511 lustre_msg_get_transno(req->rq_repmsg) == 0,
2513 lustre_msg_get_transno(req->rq_reqmsg),
2514 lustre_msg_get_transno(req->rq_repmsg));
2517 cfs_spin_lock(&imp->imp_lock);
2518 /** if replays by version then gap was occur on server, no trust to locks */
2519 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2520 imp->imp_no_lock_replay = 1;
2521 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2522 cfs_spin_unlock(&imp->imp_lock);
2523 LASSERT(imp->imp_last_replay_transno);
2525 DEBUG_REQ(D_HA, req, "got rep");
2527 /* let the callback do fixups, possibly including in the request */
2528 if (req->rq_replay_cb)
2529 req->rq_replay_cb(req);
2531 if (ptlrpc_client_replied(req) &&
2532 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2533 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2534 lustre_msg_get_status(req->rq_repmsg),
2535 aa->praa_old_status);
2537 /* Put it back for re-replay. */
2538 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2542 * Errors while replay can set transno to 0, but
2543 * imp_last_replay_transno shouldn't be set to 0 anyway
2545 if (req->rq_transno > 0) {
2546 cfs_spin_lock(&imp->imp_lock);
2547 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2548 imp->imp_last_replay_transno = req->rq_transno;
2549 cfs_spin_unlock(&imp->imp_lock);
2551 CERROR("Transno is 0 during replay!\n");
2552 /* continue with recovery */
2553 rc = ptlrpc_import_recovery_state_machine(imp);
2555 req->rq_send_state = aa->praa_old_state;
2558 /* this replay failed, so restart recovery */
2559 ptlrpc_connect_import(imp, NULL);
2565 * Prepares and queues request for replay.
2566 * Adds it to ptlrpcd queue for actual sending.
2567 * Returns 0 on success.
2569 int ptlrpc_replay_req(struct ptlrpc_request *req)
2571 struct ptlrpc_replay_async_args *aa;
2574 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2575 /* Not handling automatic bulk replay yet (or ever?) */
2576 LASSERT(req->rq_bulk == NULL);
2578 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2579 aa = ptlrpc_req_async_args(req);
2580 memset(aa, 0, sizeof *aa);
2582 /* Prepare request to be resent with ptlrpcd */
2583 aa->praa_old_state = req->rq_send_state;
2584 req->rq_send_state = LUSTRE_IMP_REPLAY;
2585 req->rq_phase = RQ_PHASE_NEW;
2586 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2588 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2590 req->rq_interpret_reply = ptlrpc_replay_interpret;
2591 /* Readjust the timeout for current conditions */
2592 ptlrpc_at_set_req_timeout(req);
2594 DEBUG_REQ(D_HA, req, "REPLAY");
2596 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2597 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2599 ptlrpcd_add_req(req, PSCOPE_OTHER);
2604 * Aborts all in-flight request on import \a imp sending and delayed lists
2606 void ptlrpc_abort_inflight(struct obd_import *imp)
2608 cfs_list_t *tmp, *n;
2611 /* Make sure that no new requests get processed for this import.
2612 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2613 * this flag and then putting requests on sending_list or delayed_list.
2615 cfs_spin_lock(&imp->imp_lock);
2617 /* XXX locking? Maybe we should remove each request with the list
2618 * locked? Also, how do we know if the requests on the list are
2619 * being freed at this time?
2621 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2622 struct ptlrpc_request *req =
2623 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2625 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2627 cfs_spin_lock (&req->rq_lock);
2628 if (req->rq_import_generation < imp->imp_generation) {
2630 req->rq_status = -EINTR;
2631 ptlrpc_client_wake_req(req);
2633 cfs_spin_unlock (&req->rq_lock);
2636 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2637 struct ptlrpc_request *req =
2638 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2640 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2642 cfs_spin_lock (&req->rq_lock);
2643 if (req->rq_import_generation < imp->imp_generation) {
2645 req->rq_status = -EINTR;
2646 ptlrpc_client_wake_req(req);
2648 cfs_spin_unlock (&req->rq_lock);
2651 /* Last chance to free reqs left on the replay list, but we
2652 * will still leak reqs that haven't committed. */
2653 if (imp->imp_replayable)
2654 ptlrpc_free_committed(imp);
2656 cfs_spin_unlock(&imp->imp_lock);
2662 * Abort all uncompleted requests in request set \a set
2664 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2666 cfs_list_t *tmp, *pos;
2668 LASSERT(set != NULL);
2670 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2671 struct ptlrpc_request *req =
2672 cfs_list_entry(pos, struct ptlrpc_request,
2675 cfs_spin_lock(&req->rq_lock);
2676 if (req->rq_phase != RQ_PHASE_RPC) {
2677 cfs_spin_unlock(&req->rq_lock);
2682 req->rq_status = -EINTR;
2683 ptlrpc_client_wake_req(req);
2684 cfs_spin_unlock(&req->rq_lock);
2688 static __u64 ptlrpc_last_xid;
2689 static cfs_spinlock_t ptlrpc_last_xid_lock;
2692 * Initialize the XID for the node. This is common among all requests on
2693 * this node, and only requires the property that it is monotonically
2694 * increasing. It does not need to be sequential. Since this is also used
2695 * as the RDMA match bits, it is important that a single client NOT have
2696 * the same match bits for two different in-flight requests, hence we do
2697 * NOT want to have an XID per target or similar.
2699 * To avoid an unlikely collision between match bits after a client reboot
2700 * (which would deliver old data into the wrong RDMA buffer) initialize
2701 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2702 * If the time is clearly incorrect, we instead use a 62-bit random number.
2703 * In the worst case the random number will overflow 1M RPCs per second in
2704 * 9133 years, or permutations thereof.
2706 #define YEAR_2004 (1ULL << 30)
2707 void ptlrpc_init_xid(void)
2709 time_t now = cfs_time_current_sec();
2711 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2712 if (now < YEAR_2004) {
2713 ll_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2714 ptlrpc_last_xid >>= 2;
2715 ptlrpc_last_xid |= (1ULL << 61);
2717 ptlrpc_last_xid = (__u64)now << 20;
2722 * Increase xid and returns resultng new value to the caller.
2724 __u64 ptlrpc_next_xid(void)
2727 cfs_spin_lock(&ptlrpc_last_xid_lock);
2728 tmp = ++ptlrpc_last_xid;
2729 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2734 * Get a glimpse at what next xid value might have been.
2735 * Returns possible next xid.
2737 __u64 ptlrpc_sample_next_xid(void)
2739 #if BITS_PER_LONG == 32
2740 /* need to avoid possible word tearing on 32-bit systems */
2742 cfs_spin_lock(&ptlrpc_last_xid_lock);
2743 tmp = ptlrpc_last_xid + 1;
2744 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2747 /* No need to lock, since returned value is racy anyways */
2748 return ptlrpc_last_xid + 1;
2751 EXPORT_SYMBOL(ptlrpc_sample_next_xid);