4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 /** Implementation of client-side PortalRPC interfaces */
39 #define DEBUG_SUBSYSTEM S_RPC
43 #include <liblustre.h>
46 #include <obd_support.h>
47 #include <obd_class.h>
48 #include <lustre_lib.h>
49 #include <lustre_ha.h>
50 #include <lustre_import.h>
51 #include <lustre_req_layout.h>
53 #include "ptlrpc_internal.h"
56 * Initialize passed in client structure \a cl.
58 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
59 struct ptlrpc_client *cl)
61 cl->cli_request_portal = req_portal;
62 cl->cli_reply_portal = rep_portal;
67 * Return PortalRPC connection for remore uud \a uuid
69 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
71 struct ptlrpc_connection *c;
73 lnet_process_id_t peer;
76 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
78 CNETERR("cannot find peer %s!\n", uuid->uuid);
82 c = ptlrpc_connection_get(peer, self, uuid);
84 memcpy(c->c_remote_uuid.uuid,
85 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
88 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
94 * Allocate and initialize new bulk descriptor
95 * Returns pointer to the descriptor or NULL on error.
97 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 * Add a page \a page to the bulk descriptor \a desc.
150 * Data to transfer in the page starts at offset \a pageoffset and
151 * amount of data to transfer from the page is \a len
153 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
154 cfs_page_t *page, int pageoffset, int len)
156 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
157 LASSERT(page != NULL);
158 LASSERT(pageoffset >= 0);
160 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
165 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
169 * Uninitialize and free bulk descriptor \a desc.
170 * Works on bulk descriptors both from server and client side.
172 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
177 LASSERT(desc != NULL);
178 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
179 LASSERT(!desc->bd_network_rw); /* network hands off or */
180 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
182 sptlrpc_enc_pool_put_pages(desc);
185 class_export_put(desc->bd_export);
187 class_import_put(desc->bd_import);
189 for (i = 0; i < desc->bd_iov_count ; i++)
190 cfs_page_unpin(desc->bd_iov[i].kiov_page);
192 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
193 bd_iov[desc->bd_max_iov]));
198 * Set server timelimit for this req, i.e. how long are we willing to wait
199 * for reply before timing out this request.
201 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
207 LASSERT(req->rq_import);
210 /* non-AT settings */
212 * \a imp_server_timeout means this is reverse import and
213 * we send (currently only) ASTs to the client and cannot afford
214 * to wait too long for the reply, otherwise the other client
215 * (because of which we are sending this request) would
216 * timeout waiting for us
218 req->rq_timeout = req->rq_import->imp_server_timeout ?
219 obd_timeout / 2 : obd_timeout;
221 at = &req->rq_import->imp_at;
222 idx = import_at_get_index(req->rq_import,
223 req->rq_request_portal);
224 serv_est = at_get(&at->iat_service_estimate[idx]);
225 req->rq_timeout = at_est2timeout(serv_est);
227 /* We could get even fancier here, using history to predict increased
230 /* Let the server know what this RPC timeout is by putting it in the
232 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
235 /* Adjust max service estimate based on server value */
236 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
237 unsigned int serv_est)
243 LASSERT(req->rq_import);
244 at = &req->rq_import->imp_at;
246 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
247 /* max service estimates are tracked on the server side,
248 so just keep minimal history here */
249 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
251 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
252 "has changed from %d to %d\n",
253 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
254 oldse, at_get(&at->iat_service_estimate[idx]));
257 /* Expected network latency per remote node (secs) */
258 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
260 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
263 /* Adjust expected network latency */
264 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
265 unsigned int service_time)
267 unsigned int nl, oldnl;
269 time_t now = cfs_time_current_sec();
271 LASSERT(req->rq_import);
272 at = &req->rq_import->imp_at;
274 /* Network latency is total time less server processing time */
275 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
276 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
277 CWARN("Reported service time %u > total measured time "
278 CFS_DURATION_T"\n", service_time,
279 cfs_time_sub(now, req->rq_sent));
281 oldnl = at_measured(&at->iat_net_latency, nl);
283 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
284 "has changed from %d to %d\n",
285 req->rq_import->imp_obd->obd_name,
287 &req->rq_import->imp_connection->c_remote_uuid),
288 oldnl, at_get(&at->iat_net_latency));
291 static int unpack_reply(struct ptlrpc_request *req)
295 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
296 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
298 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
303 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
305 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
312 * Handle an early reply message, called with the rq_lock held.
313 * If anything goes wrong just ignore it - same as if it never happened
315 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
317 struct ptlrpc_request *early_req;
323 cfs_spin_unlock(&req->rq_lock);
325 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
327 cfs_spin_lock(&req->rq_lock);
331 rc = unpack_reply(early_req);
333 /* Expecting to increase the service time estimate here */
334 ptlrpc_at_adj_service(req,
335 lustre_msg_get_timeout(early_req->rq_repmsg));
336 ptlrpc_at_adj_net_latency(req,
337 lustre_msg_get_service_time(early_req->rq_repmsg));
340 sptlrpc_cli_finish_early_reply(early_req);
342 cfs_spin_lock(&req->rq_lock);
345 /* Adjust the local timeout for this req */
346 ptlrpc_at_set_req_timeout(req);
348 olddl = req->rq_deadline;
349 /* server assumes it now has rq_timeout from when it sent the
350 early reply, so client should give it at least that long. */
351 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
352 ptlrpc_at_get_net_latency(req);
354 DEBUG_REQ(D_ADAPTTO, req,
355 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
356 "("CFS_DURATION_T"s)", req->rq_early_count,
357 cfs_time_sub(req->rq_deadline,
358 cfs_time_current_sec()),
359 cfs_time_sub(req->rq_deadline, olddl));
366 * Wind down request pool \a pool.
367 * Frees all requests from the pool too
369 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
372 struct ptlrpc_request *req;
374 LASSERT(pool != NULL);
376 cfs_spin_lock(&pool->prp_lock);
377 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
378 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
379 cfs_list_del(&req->rq_list);
380 LASSERT(req->rq_reqbuf);
381 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
382 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
383 OBD_FREE(req, sizeof(*req));
385 cfs_spin_unlock(&pool->prp_lock);
386 OBD_FREE(pool, sizeof(*pool));
390 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
392 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
397 while (size < pool->prp_rq_size)
400 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
401 size == pool->prp_rq_size,
402 "Trying to change pool size with nonempty pool "
403 "from %d to %d bytes\n", pool->prp_rq_size, size);
405 cfs_spin_lock(&pool->prp_lock);
406 pool->prp_rq_size = size;
407 for (i = 0; i < num_rq; i++) {
408 struct ptlrpc_request *req;
409 struct lustre_msg *msg;
411 cfs_spin_unlock(&pool->prp_lock);
412 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
415 OBD_ALLOC_LARGE(msg, size);
417 OBD_FREE(req, sizeof(struct ptlrpc_request));
420 req->rq_reqbuf = msg;
421 req->rq_reqbuf_len = size;
423 cfs_spin_lock(&pool->prp_lock);
424 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
426 cfs_spin_unlock(&pool->prp_lock);
431 * Create and initialize new request pool with given attributes:
432 * \a num_rq - initial number of requests to create for the pool
433 * \a msgsize - maximum message size possible for requests in thid pool
434 * \a populate_pool - function to be called when more requests need to be added
436 * Returns pointer to newly created pool or NULL on error.
438 struct ptlrpc_request_pool *
439 ptlrpc_init_rq_pool(int num_rq, int msgsize,
440 void (*populate_pool)(struct ptlrpc_request_pool *, int))
442 struct ptlrpc_request_pool *pool;
444 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
448 /* Request next power of two for the allocation, because internally
449 kernel would do exactly this */
451 cfs_spin_lock_init(&pool->prp_lock);
452 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
453 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
454 pool->prp_populate = populate_pool;
456 populate_pool(pool, num_rq);
458 if (cfs_list_empty(&pool->prp_req_list)) {
459 /* have not allocated a single request for the pool */
460 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
467 * Fetches one request from pool \a pool
469 static struct ptlrpc_request *
470 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
472 struct ptlrpc_request *request;
473 struct lustre_msg *reqbuf;
478 cfs_spin_lock(&pool->prp_lock);
480 /* See if we have anything in a pool, and bail out if nothing,
481 * in writeout path, where this matters, this is safe to do, because
482 * nothing is lost in this case, and when some in-flight requests
483 * complete, this code will be called again. */
484 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
485 cfs_spin_unlock(&pool->prp_lock);
489 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
491 cfs_list_del_init(&request->rq_list);
492 cfs_spin_unlock(&pool->prp_lock);
494 LASSERT(request->rq_reqbuf);
495 LASSERT(request->rq_pool);
497 reqbuf = request->rq_reqbuf;
498 memset(request, 0, sizeof(*request));
499 request->rq_reqbuf = reqbuf;
500 request->rq_reqbuf_len = pool->prp_rq_size;
501 request->rq_pool = pool;
507 * Returns freed \a request to pool.
509 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
511 struct ptlrpc_request_pool *pool = request->rq_pool;
513 cfs_spin_lock(&pool->prp_lock);
514 LASSERT(cfs_list_empty(&request->rq_list));
515 LASSERT(!request->rq_receiving_reply);
516 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
517 cfs_spin_unlock(&pool->prp_lock);
520 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
521 __u32 version, int opcode,
522 int count, __u32 *lengths, char **bufs,
523 struct ptlrpc_cli_ctx *ctx)
525 struct obd_import *imp = request->rq_import;
530 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
532 rc = sptlrpc_req_get_ctx(request);
537 sptlrpc_req_set_flavor(request, opcode);
539 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
542 LASSERT(!request->rq_pool);
546 lustre_msg_add_version(request->rq_reqmsg, version);
547 request->rq_send_state = LUSTRE_IMP_FULL;
548 request->rq_type = PTL_RPC_MSG_REQUEST;
549 request->rq_export = NULL;
551 request->rq_req_cbid.cbid_fn = request_out_callback;
552 request->rq_req_cbid.cbid_arg = request;
554 request->rq_reply_cbid.cbid_fn = reply_in_callback;
555 request->rq_reply_cbid.cbid_arg = request;
557 request->rq_reply_deadline = 0;
558 request->rq_phase = RQ_PHASE_NEW;
559 request->rq_next_phase = RQ_PHASE_UNDEFINED;
561 request->rq_request_portal = imp->imp_client->cli_request_portal;
562 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
564 ptlrpc_at_set_req_timeout(request);
566 cfs_spin_lock_init(&request->rq_lock);
567 CFS_INIT_LIST_HEAD(&request->rq_list);
568 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
569 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
570 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
571 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
572 CFS_INIT_LIST_HEAD(&request->rq_history_list);
573 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
574 cfs_waitq_init(&request->rq_reply_waitq);
575 cfs_waitq_init(&request->rq_set_waitq);
576 request->rq_xid = ptlrpc_next_xid();
577 cfs_atomic_set(&request->rq_refcount, 1);
579 lustre_msg_set_opc(request->rq_reqmsg, opcode);
583 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
585 class_import_put(imp);
589 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
590 __u32 version, int opcode, char **bufs,
591 struct ptlrpc_cli_ctx *ctx)
595 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
596 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
597 request->rq_pill.rc_area[RCL_CLIENT],
600 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
603 * Pack request buffers for network transfer, performing necessary encryption
604 * steps if necessary.
606 int ptlrpc_request_pack(struct ptlrpc_request *request,
607 __u32 version, int opcode)
610 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
614 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
615 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
616 * have to send old ptlrpc_body to keep interoprability with these
619 * Only three kinds of server->client RPCs so far:
624 * XXX This should be removed whenever we drop the interoprability with
625 * the these old clients.
627 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
628 opcode == LDLM_GL_CALLBACK)
629 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
630 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
636 * Helper function to allocate new request on import \a imp
637 * and possibly using existing request from pool \a pool if provided.
638 * Returns allocated request structure with import field filled or
642 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
643 struct ptlrpc_request_pool *pool)
645 struct ptlrpc_request *request = NULL;
648 request = ptlrpc_prep_req_from_pool(pool);
651 OBD_ALLOC_PTR(request);
654 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
655 LASSERT(imp != LP_POISON);
656 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
658 LASSERT(imp->imp_client != LP_POISON);
660 request->rq_import = class_import_get(imp);
662 CERROR("request allocation out of memory\n");
669 * Helper function for creating a request.
670 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
671 * buffer structures according to capsule template \a format.
672 * Returns allocated request structure pointer or NULL on error.
674 static struct ptlrpc_request *
675 ptlrpc_request_alloc_internal(struct obd_import *imp,
676 struct ptlrpc_request_pool * pool,
677 const struct req_format *format)
679 struct ptlrpc_request *request;
681 request = __ptlrpc_request_alloc(imp, pool);
685 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
686 req_capsule_set(&request->rq_pill, format);
691 * Allocate new request structure for import \a imp and initialize its
692 * buffer structure according to capsule template \a format.
694 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
695 const struct req_format *format)
697 return ptlrpc_request_alloc_internal(imp, NULL, format);
701 * Allocate new request structure for import \a imp from pool \a pool and
702 * initialize its buffer structure according to capsule template \a format.
704 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
705 struct ptlrpc_request_pool * pool,
706 const struct req_format *format)
708 return ptlrpc_request_alloc_internal(imp, pool, format);
712 * For requests not from pool, free memory of the request structure.
713 * For requests obtained from a pool earlier, return request back to pool.
715 void ptlrpc_request_free(struct ptlrpc_request *request)
717 if (request->rq_pool)
718 __ptlrpc_free_req_to_pool(request);
720 OBD_FREE_PTR(request);
724 * Allocate new request for operatione \a opcode and immediatelly pack it for
726 * Only used for simple requests like OBD_PING where the only important
727 * part of the request is operation itself.
728 * Returns allocated request or NULL on error.
730 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
731 const struct req_format *format,
732 __u32 version, int opcode)
734 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
738 rc = ptlrpc_request_pack(req, version, opcode);
740 ptlrpc_request_free(req);
748 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
749 * for operation \a opcode. Request would contain \a count buffers.
750 * Sizes of buffers are described in array \a lengths and buffers themselves
751 * are provided by a pointer \a bufs.
752 * Returns prepared request structure pointer or NULL on error.
754 struct ptlrpc_request *
755 ptlrpc_prep_req_pool(struct obd_import *imp,
756 __u32 version, int opcode,
757 int count, __u32 *lengths, char **bufs,
758 struct ptlrpc_request_pool *pool)
760 struct ptlrpc_request *request;
763 request = __ptlrpc_request_alloc(imp, pool);
767 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
768 lengths, bufs, NULL);
770 ptlrpc_request_free(request);
777 * Same as ptlrpc_prep_req_pool, but without pool
779 struct ptlrpc_request *
780 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
781 __u32 *lengths, char **bufs)
783 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
788 * Allocate "fake" request that would not be sent anywhere in the end.
789 * Only used as a hack because we have no other way of performing
790 * async actions in lustre between layers.
791 * Used on MDS to request object preallocations from more than one OST at a
794 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
795 unsigned int timeout,
796 ptlrpc_interpterer_t interpreter)
798 struct ptlrpc_request *request = NULL;
801 OBD_ALLOC(request, sizeof(*request));
803 CERROR("request allocation out of memory\n");
807 request->rq_send_state = LUSTRE_IMP_FULL;
808 request->rq_type = PTL_RPC_MSG_REQUEST;
809 request->rq_import = class_import_get(imp);
810 request->rq_export = NULL;
811 request->rq_import_generation = imp->imp_generation;
813 request->rq_timeout = timeout;
814 request->rq_sent = cfs_time_current_sec();
815 request->rq_deadline = request->rq_sent + timeout;
816 request->rq_reply_deadline = request->rq_deadline;
817 request->rq_interpret_reply = interpreter;
818 request->rq_phase = RQ_PHASE_RPC;
819 request->rq_next_phase = RQ_PHASE_INTERPRET;
820 /* don't want reply */
821 request->rq_receiving_reply = 0;
822 request->rq_must_unlink = 0;
823 request->rq_no_delay = request->rq_no_resend = 1;
824 request->rq_fake = 1;
826 cfs_spin_lock_init(&request->rq_lock);
827 CFS_INIT_LIST_HEAD(&request->rq_list);
828 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
829 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
830 CFS_INIT_LIST_HEAD(&request->rq_history_list);
831 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
832 cfs_waitq_init(&request->rq_reply_waitq);
833 cfs_waitq_init(&request->rq_set_waitq);
835 request->rq_xid = ptlrpc_next_xid();
836 cfs_atomic_set(&request->rq_refcount, 1);
842 * Indicate that processing of "fake" request is finished.
844 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
846 struct ptlrpc_request_set *set = req->rq_set;
849 /* hold ref on the request to prevent others (ptlrpcd) to free it */
850 ptlrpc_request_addref(req);
851 cfs_list_del_init(&req->rq_list);
853 /* if we kill request before timeout - need adjust counter */
854 if (req->rq_phase == RQ_PHASE_RPC && set != NULL &&
855 cfs_atomic_dec_and_test(&set->set_remaining))
858 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
860 /* Only need to call wakeup once when to be empty. */
862 cfs_waitq_signal(&set->set_waitq);
863 ptlrpc_req_finished(req);
867 * Allocate and initialize new request set structure.
868 * Returns a pointer to the newly allocated set structure or NULL on error.
870 struct ptlrpc_request_set *ptlrpc_prep_set(void)
872 struct ptlrpc_request_set *set;
875 OBD_ALLOC(set, sizeof *set);
878 cfs_atomic_set(&set->set_refcount, 1);
879 CFS_INIT_LIST_HEAD(&set->set_requests);
880 cfs_waitq_init(&set->set_waitq);
881 cfs_atomic_set(&set->set_new_count, 0);
882 cfs_atomic_set(&set->set_remaining, 0);
883 cfs_spin_lock_init(&set->set_new_req_lock);
884 CFS_INIT_LIST_HEAD(&set->set_new_requests);
885 CFS_INIT_LIST_HEAD(&set->set_cblist);
891 * Wind down and free request set structure previously allocated with
893 * Ensures that all requests on the set have completed and removes
894 * all requests from the request list in a set.
895 * If any unsent request happen to be on the list, pretends that they got
896 * an error in flight and calls their completion handler.
898 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
906 /* Requests on the set should either all be completed, or all be new */
907 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
908 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
909 cfs_list_for_each (tmp, &set->set_requests) {
910 struct ptlrpc_request *req =
911 cfs_list_entry(tmp, struct ptlrpc_request,
914 LASSERT(req->rq_phase == expected_phase);
918 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
919 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
920 cfs_atomic_read(&set->set_remaining), n);
922 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
923 struct ptlrpc_request *req =
924 cfs_list_entry(tmp, struct ptlrpc_request,
926 cfs_list_del_init(&req->rq_set_chain);
928 LASSERT(req->rq_phase == expected_phase);
930 if (req->rq_phase == RQ_PHASE_NEW) {
931 ptlrpc_req_interpret(NULL, req, -EBADR);
932 cfs_atomic_dec(&set->set_remaining);
935 cfs_spin_lock(&req->rq_lock);
937 req->rq_invalid_rqset = 0;
938 cfs_spin_unlock(&req->rq_lock);
940 ptlrpc_req_finished (req);
943 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
945 ptlrpc_reqset_put(set);
950 * Add a callback function \a fn to the set.
951 * This function would be called when all requests on this set are completed.
952 * The function will be passed \a data argument.
954 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
955 set_interpreter_func fn, void *data)
957 struct ptlrpc_set_cbdata *cbdata;
959 OBD_ALLOC_PTR(cbdata);
963 cbdata->psc_interpret = fn;
964 cbdata->psc_data = data;
965 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
971 * Add a new request to the general purpose request set.
972 * Assumes request reference from the caller.
974 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
975 struct ptlrpc_request *req)
977 char jobid[JOBSTATS_JOBID_SIZE];
978 LASSERT(cfs_list_empty(&req->rq_set_chain));
980 /* The set takes over the caller's request reference */
981 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
983 cfs_atomic_inc(&set->set_remaining);
984 req->rq_queued_time = cfs_time_current();
986 if (req->rq_reqmsg) {
987 lustre_get_jobid(jobid);
988 lustre_msg_set_jobid(req->rq_reqmsg, jobid);
993 * Add a request to a request with dedicated server thread
994 * and wake the thread to make any necessary processing.
995 * Currently only used for ptlrpcd.
997 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
998 struct ptlrpc_request *req)
1000 struct ptlrpc_request_set *set = pc->pc_set;
1003 LASSERT(req->rq_set == NULL);
1004 LASSERT(cfs_test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1006 cfs_spin_lock(&set->set_new_req_lock);
1008 * The set takes over the caller's request reference.
1011 req->rq_queued_time = cfs_time_current();
1012 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1013 count = cfs_atomic_inc_return(&set->set_new_count);
1014 cfs_spin_unlock(&set->set_new_req_lock);
1016 /* Only need to call wakeup once for the first entry. */
1018 cfs_waitq_signal(&set->set_waitq);
1020 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1021 * guarantee the async RPC can be processed ASAP, we have
1022 * no other better choice. It maybe fixed in future. */
1023 for (i = 0; i < pc->pc_npartners; i++)
1024 cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
1029 * Based on the current state of the import, determine if the request
1030 * can be sent, is an error, or should be delayed.
1032 * Returns true if this request should be delayed. If false, and
1033 * *status is set, then the request can not be sent and *status is the
1034 * error code. If false and status is 0, then request can be sent.
1036 * The imp->imp_lock must be held.
1038 static int ptlrpc_import_delay_req(struct obd_import *imp,
1039 struct ptlrpc_request *req, int *status)
1044 LASSERT (status != NULL);
1047 if (req->rq_ctx_init || req->rq_ctx_fini) {
1048 /* always allow ctx init/fini rpc go through */
1049 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1050 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1053 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1054 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1056 } else if (ptlrpc_send_limit_expired(req)) {
1057 /* probably doesn't need to be a D_ERROR after initial testing */
1058 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1060 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1061 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1062 /* allow CONNECT even if import is invalid */ ;
1063 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1064 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1067 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1068 if (!imp->imp_deactive)
1069 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1070 *status = -ESHUTDOWN; /* bz 12940 */
1071 } else if (req->rq_import_generation != imp->imp_generation) {
1072 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1074 } else if (req->rq_send_state != imp->imp_state) {
1075 /* invalidate in progress - any requests should be drop */
1076 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1077 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1079 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1080 *status = -EWOULDBLOCK;
1090 * Decide if the eror message regarding provided request \a req
1091 * should be printed to the console or not.
1092 * Makes it's decision on request status and other properties.
1093 * Returns 1 to print error on the system console or 0 if not.
1095 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1100 /* Fake requests include no rq_reqmsg */
1104 LASSERT(req->rq_reqmsg != NULL);
1105 opc = lustre_msg_get_opc(req->rq_reqmsg);
1107 /* Suppress particular reconnect errors which are to be expected. No
1108 * errors are suppressed for the initial connection on an import */
1109 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1110 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1112 /* Suppress timed out reconnect requests */
1113 if (req->rq_timedout)
1116 /* Suppress unavailable/again reconnect requests */
1117 err = lustre_msg_get_status(req->rq_repmsg);
1118 if (err == -ENODEV || err == -EAGAIN)
1126 * Check request processing status.
1127 * Returns the status.
1129 static int ptlrpc_check_status(struct ptlrpc_request *req)
1134 err = lustre_msg_get_status(req->rq_repmsg);
1135 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1136 struct obd_import *imp = req->rq_import;
1137 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1138 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
1139 " with %s. The %s operation failed with %d\n",
1140 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1141 ll_opcode2str(opc), err);
1142 RETURN(err < 0 ? err : -EINVAL);
1146 DEBUG_REQ(D_INFO, req, "status is %d", err);
1147 } else if (err > 0) {
1148 /* XXX: translate this error from net to host */
1149 DEBUG_REQ(D_INFO, req, "status is %d", err);
1152 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1153 struct obd_import *imp = req->rq_import;
1154 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1156 if (ptlrpc_console_allow(req))
1157 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1158 "communicating with %s. The %s "
1159 "operation failed with %d\n",
1161 imp->imp_connection->c_peer.nid),
1162 ll_opcode2str(opc), err);
1164 RETURN(err < 0 ? err : -EINVAL);
1171 * save pre-versions of objects into request for replay.
1172 * Versions are obtained from server reply.
1175 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1177 struct lustre_msg *repmsg = req->rq_repmsg;
1178 struct lustre_msg *reqmsg = req->rq_reqmsg;
1179 __u64 *versions = lustre_msg_get_versions(repmsg);
1182 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1186 lustre_msg_set_versions(reqmsg, versions);
1187 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1188 versions[0], versions[1]);
1194 * Callback function called when client receives RPC reply for \a req.
1195 * Returns 0 on success or error code.
1196 * The return alue would be assigned to req->rq_status by the caller
1197 * as request processing status.
1198 * This function also decides if the request needs to be saved for later replay.
1200 static int after_reply(struct ptlrpc_request *req)
1202 struct obd_import *imp = req->rq_import;
1203 struct obd_device *obd = req->rq_import->imp_obd;
1205 struct timeval work_start;
1209 LASSERT(obd != NULL);
1210 /* repbuf must be unlinked */
1211 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1213 if (req->rq_reply_truncate) {
1214 if (ptlrpc_no_resend(req)) {
1215 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1216 " expected: %d, actual size: %d",
1217 req->rq_nob_received, req->rq_repbuf_len);
1221 sptlrpc_cli_free_repbuf(req);
1222 /* Pass the required reply buffer size (include
1223 * space for early reply).
1224 * NB: no need to roundup because alloc_repbuf
1225 * will roundup it */
1226 req->rq_replen = req->rq_nob_received;
1227 req->rq_nob_received = 0;
1233 * NB Until this point, the whole of the incoming message,
1234 * including buflens, status etc is in the sender's byte order.
1236 rc = sptlrpc_cli_unwrap_reply(req);
1238 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1243 * Security layer unwrap might ask resend this request.
1248 rc = unpack_reply(req);
1252 cfs_gettimeofday(&work_start);
1253 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1254 if (obd->obd_svc_stats != NULL) {
1255 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1257 ptlrpc_lprocfs_rpc_sent(req, timediff);
1260 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1261 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1262 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1263 lustre_msg_get_type(req->rq_repmsg));
1267 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1268 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1269 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1270 ptlrpc_at_adj_net_latency(req,
1271 lustre_msg_get_service_time(req->rq_repmsg));
1273 rc = ptlrpc_check_status(req);
1274 imp->imp_connect_error = rc;
1278 * Either we've been evicted, or the server has failed for
1279 * some reason. Try to reconnect, and if that fails, punt to
1282 if (ll_rpc_recoverable_error(rc)) {
1283 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1284 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1287 ptlrpc_request_handle_notconn(req);
1292 * Let's look if server sent slv. Do it only for RPC with
1295 ldlm_cli_update_pool(req);
1299 * Store transno in reqmsg for replay.
1301 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1302 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1303 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1306 if (imp->imp_replayable) {
1307 cfs_spin_lock(&imp->imp_lock);
1309 * No point in adding already-committed requests to the replay
1310 * list, we will just remove them immediately. b=9829
1312 if (req->rq_transno != 0 &&
1314 lustre_msg_get_last_committed(req->rq_repmsg) ||
1316 /** version recovery */
1317 ptlrpc_save_versions(req);
1318 ptlrpc_retain_replayable_request(req, imp);
1319 } else if (req->rq_commit_cb != NULL) {
1320 cfs_spin_unlock(&imp->imp_lock);
1321 req->rq_commit_cb(req);
1322 cfs_spin_lock(&imp->imp_lock);
1326 * Replay-enabled imports return commit-status information.
1328 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1329 imp->imp_peer_committed_transno =
1330 lustre_msg_get_last_committed(req->rq_repmsg);
1332 ptlrpc_free_committed(imp);
1334 if (req->rq_transno > imp->imp_peer_committed_transno)
1335 ptlrpc_pinger_commit_expected(imp);
1337 cfs_spin_unlock(&imp->imp_lock);
1344 * Helper function to send request \a req over the network for the first time
1345 * Also adjusts request phase.
1346 * Returns 0 on success or error code.
1348 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1350 struct obd_import *imp = req->rq_import;
1354 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1355 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1356 (!req->rq_generation_set ||
1357 req->rq_import_generation == imp->imp_generation))
1360 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1362 cfs_spin_lock(&imp->imp_lock);
1364 if (!req->rq_generation_set)
1365 req->rq_import_generation = imp->imp_generation;
1367 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1368 cfs_spin_lock(&req->rq_lock);
1369 req->rq_waiting = 1;
1370 cfs_spin_unlock(&req->rq_lock);
1372 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1373 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1374 ptlrpc_import_state_name(req->rq_send_state),
1375 ptlrpc_import_state_name(imp->imp_state));
1376 LASSERT(cfs_list_empty(&req->rq_list));
1377 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1378 cfs_atomic_inc(&req->rq_import->imp_inflight);
1379 cfs_spin_unlock(&imp->imp_lock);
1384 cfs_spin_unlock(&imp->imp_lock);
1385 req->rq_status = rc;
1386 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1390 LASSERT(cfs_list_empty(&req->rq_list));
1391 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1392 cfs_atomic_inc(&req->rq_import->imp_inflight);
1393 cfs_spin_unlock(&imp->imp_lock);
1395 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1397 rc = sptlrpc_req_refresh_ctx(req, -1);
1400 req->rq_status = rc;
1403 req->rq_wait_ctx = 1;
1408 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1409 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1410 imp->imp_obd->obd_uuid.uuid,
1411 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1412 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1413 lustre_msg_get_opc(req->rq_reqmsg));
1415 rc = ptl_send_rpc(req, 0);
1417 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1418 req->rq_net_err = 1;
1425 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1426 * and no more replies are expected.
1427 * (it is possible to get less replies than requests sent e.g. due to timed out
1428 * requests or requests that we had trouble to send out)
1430 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1433 int force_timer_recalc = 0;
1436 if (cfs_atomic_read(&set->set_remaining) == 0)
1439 cfs_list_for_each(tmp, &set->set_requests) {
1440 struct ptlrpc_request *req =
1441 cfs_list_entry(tmp, struct ptlrpc_request,
1443 struct obd_import *imp = req->rq_import;
1444 int unregistered = 0;
1447 if (req->rq_phase == RQ_PHASE_NEW &&
1448 ptlrpc_send_new_req(req)) {
1449 force_timer_recalc = 1;
1452 /* delayed send - skip */
1453 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1456 if (!(req->rq_phase == RQ_PHASE_RPC ||
1457 req->rq_phase == RQ_PHASE_BULK ||
1458 req->rq_phase == RQ_PHASE_INTERPRET ||
1459 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1460 req->rq_phase == RQ_PHASE_COMPLETE)) {
1461 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1465 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1466 LASSERT(req->rq_next_phase != req->rq_phase);
1467 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1470 * Skip processing until reply is unlinked. We
1471 * can't return to pool before that and we can't
1472 * call interpret before that. We need to make
1473 * sure that all rdma transfers finished and will
1474 * not corrupt any data.
1476 if (ptlrpc_client_recv_or_unlink(req) ||
1477 ptlrpc_client_bulk_active(req))
1481 * Turn fail_loc off to prevent it from looping
1484 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1485 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1488 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1489 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1494 * Move to next phase if reply was successfully
1497 ptlrpc_rqphase_move(req, req->rq_next_phase);
1500 if (req->rq_phase == RQ_PHASE_COMPLETE)
1503 if (req->rq_phase == RQ_PHASE_INTERPRET)
1504 GOTO(interpret, req->rq_status);
1507 * Note that this also will start async reply unlink.
1509 if (req->rq_net_err && !req->rq_timedout) {
1510 ptlrpc_expire_one_request(req, 1);
1513 * Check if we still need to wait for unlink.
1515 if (ptlrpc_client_recv_or_unlink(req) ||
1516 ptlrpc_client_bulk_active(req))
1518 /* If there is no need to resend, fail it now. */
1519 if (req->rq_no_resend) {
1520 if (req->rq_status == 0)
1521 req->rq_status = -EIO;
1522 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1523 GOTO(interpret, req->rq_status);
1530 cfs_spin_lock(&req->rq_lock);
1531 req->rq_replied = 0;
1532 cfs_spin_unlock(&req->rq_lock);
1533 if (req->rq_status == 0)
1534 req->rq_status = -EIO;
1535 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1536 GOTO(interpret, req->rq_status);
1539 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1540 * so it sets rq_intr regardless of individual rpc
1541 * timeouts. The synchronous IO waiting path sets
1542 * rq_intr irrespective of whether ptlrpcd
1543 * has seen a timeout. Our policy is to only interpret
1544 * interrupted rpcs after they have timed out, so we
1545 * need to enforce that here.
1548 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1549 req->rq_wait_ctx)) {
1550 req->rq_status = -EINTR;
1551 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1552 GOTO(interpret, req->rq_status);
1555 if (req->rq_phase == RQ_PHASE_RPC) {
1556 if (req->rq_timedout || req->rq_resend ||
1557 req->rq_waiting || req->rq_wait_ctx) {
1560 if (!ptlrpc_unregister_reply(req, 1))
1563 cfs_spin_lock(&imp->imp_lock);
1564 if (ptlrpc_import_delay_req(imp, req, &status)){
1565 /* put on delay list - only if we wait
1566 * recovery finished - before send */
1567 cfs_list_del_init(&req->rq_list);
1568 cfs_list_add_tail(&req->rq_list,
1571 cfs_spin_unlock(&imp->imp_lock);
1576 req->rq_status = status;
1577 ptlrpc_rqphase_move(req,
1578 RQ_PHASE_INTERPRET);
1579 cfs_spin_unlock(&imp->imp_lock);
1580 GOTO(interpret, req->rq_status);
1582 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1583 req->rq_status = -ENOTCONN;
1584 ptlrpc_rqphase_move(req,
1585 RQ_PHASE_INTERPRET);
1586 cfs_spin_unlock(&imp->imp_lock);
1587 GOTO(interpret, req->rq_status);
1590 cfs_list_del_init(&req->rq_list);
1591 cfs_list_add_tail(&req->rq_list,
1592 &imp->imp_sending_list);
1594 cfs_spin_unlock(&imp->imp_lock);
1596 cfs_spin_lock(&req->rq_lock);
1597 req->rq_waiting = 0;
1598 cfs_spin_unlock(&req->rq_lock);
1600 if (req->rq_timedout || req->rq_resend) {
1601 /* This is re-sending anyways,
1602 * let's mark req as resend. */
1603 cfs_spin_lock(&req->rq_lock);
1605 cfs_spin_unlock(&req->rq_lock);
1609 if (!ptlrpc_unregister_bulk(req, 1))
1612 /* ensure previous bulk fails */
1613 old_xid = req->rq_xid;
1614 req->rq_xid = ptlrpc_next_xid();
1615 CDEBUG(D_HA, "resend bulk "
1618 old_xid, req->rq_xid);
1622 * rq_wait_ctx is only touched by ptlrpcd,
1623 * so no lock is needed here.
1625 status = sptlrpc_req_refresh_ctx(req, -1);
1628 req->rq_status = status;
1629 cfs_spin_lock(&req->rq_lock);
1630 req->rq_wait_ctx = 0;
1631 cfs_spin_unlock(&req->rq_lock);
1632 force_timer_recalc = 1;
1634 cfs_spin_lock(&req->rq_lock);
1635 req->rq_wait_ctx = 1;
1636 cfs_spin_unlock(&req->rq_lock);
1641 cfs_spin_lock(&req->rq_lock);
1642 req->rq_wait_ctx = 0;
1643 cfs_spin_unlock(&req->rq_lock);
1646 rc = ptl_send_rpc(req, 0);
1648 DEBUG_REQ(D_HA, req, "send failed (%d)",
1650 force_timer_recalc = 1;
1651 cfs_spin_lock(&req->rq_lock);
1652 req->rq_net_err = 1;
1653 cfs_spin_unlock(&req->rq_lock);
1655 /* need to reset the timeout */
1656 force_timer_recalc = 1;
1659 cfs_spin_lock(&req->rq_lock);
1661 if (ptlrpc_client_early(req)) {
1662 ptlrpc_at_recv_early_reply(req);
1663 cfs_spin_unlock(&req->rq_lock);
1667 /* Still waiting for a reply? */
1668 if (ptlrpc_client_recv(req)) {
1669 cfs_spin_unlock(&req->rq_lock);
1673 /* Did we actually receive a reply? */
1674 if (!ptlrpc_client_replied(req)) {
1675 cfs_spin_unlock(&req->rq_lock);
1679 cfs_spin_unlock(&req->rq_lock);
1681 /* unlink from net because we are going to
1682 * swab in-place of reply buffer */
1683 unregistered = ptlrpc_unregister_reply(req, 1);
1687 req->rq_status = after_reply(req);
1691 /* If there is no bulk associated with this request,
1692 * then we're done and should let the interpreter
1693 * process the reply. Similarly if the RPC returned
1694 * an error, and therefore the bulk will never arrive.
1696 if (req->rq_bulk == NULL || req->rq_status < 0) {
1697 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1698 GOTO(interpret, req->rq_status);
1701 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1704 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1705 if (ptlrpc_client_bulk_active(req))
1708 if (!req->rq_bulk->bd_success) {
1709 /* The RPC reply arrived OK, but the bulk screwed
1710 * up! Dead weird since the server told us the RPC
1711 * was good after getting the REPLY for her GET or
1712 * the ACK for her PUT. */
1713 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1714 req->rq_status = -EIO;
1717 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1720 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1722 /* This moves to "unregistering" phase we need to wait for
1724 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1725 /* start async bulk unlink too */
1726 ptlrpc_unregister_bulk(req, 1);
1730 if (!ptlrpc_unregister_bulk(req, 1))
1733 /* When calling interpret receiving already should be
1735 LASSERT(!req->rq_receiving_reply);
1737 ptlrpc_req_interpret(env, req, req->rq_status);
1739 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1741 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1742 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1743 imp->imp_obd->obd_uuid.uuid,
1744 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1746 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1747 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1749 cfs_spin_lock(&imp->imp_lock);
1750 /* Request already may be not on sending or delaying list. This
1751 * may happen in the case of marking it erroneous for the case
1752 * ptlrpc_import_delay_req(req, status) find it impossible to
1753 * allow sending this rpc and returns *status != 0. */
1754 if (!cfs_list_empty(&req->rq_list)) {
1755 cfs_list_del_init(&req->rq_list);
1756 cfs_atomic_dec(&imp->imp_inflight);
1758 cfs_spin_unlock(&imp->imp_lock);
1760 cfs_atomic_dec(&set->set_remaining);
1761 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1764 /* If we hit an error, we want to recover promptly. */
1765 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1769 * Time out request \a req. is \a async_unlink is set, that means do not wait
1770 * until LNet actually confirms network buffer unlinking.
1771 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1773 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1775 struct obd_import *imp = req->rq_import;
1779 cfs_spin_lock(&req->rq_lock);
1780 req->rq_timedout = 1;
1781 cfs_spin_unlock(&req->rq_lock);
1783 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req, "Request "
1784 " sent has %s: [sent "CFS_DURATION_T"/"
1785 "real "CFS_DURATION_T"]",
1786 req->rq_net_err ? "failed due to network error" :
1787 ((req->rq_real_sent == 0 ||
1788 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1789 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1790 "timed out for sent delay" : "timed out for slow reply"),
1791 req->rq_sent, req->rq_real_sent);
1793 if (imp != NULL && obd_debug_peer_on_timeout)
1794 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1796 ptlrpc_unregister_reply(req, async_unlink);
1797 ptlrpc_unregister_bulk(req, async_unlink);
1799 if (obd_dump_on_timeout)
1800 libcfs_debug_dumplog();
1803 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1810 cfs_atomic_inc(&imp->imp_timeouts);
1812 /* The DLM server doesn't want recovery run on its imports. */
1813 if (imp->imp_dlm_fake)
1816 /* If this request is for recovery or other primordial tasks,
1817 * then error it out here. */
1818 if (req->rq_ctx_init || req->rq_ctx_fini ||
1819 req->rq_send_state != LUSTRE_IMP_FULL ||
1820 imp->imp_obd->obd_no_recov) {
1821 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1822 ptlrpc_import_state_name(req->rq_send_state),
1823 ptlrpc_import_state_name(imp->imp_state));
1824 cfs_spin_lock(&req->rq_lock);
1825 req->rq_status = -ETIMEDOUT;
1827 cfs_spin_unlock(&req->rq_lock);
1831 /* if a request can't be resent we can't wait for an answer after
1833 if (ptlrpc_no_resend(req)) {
1834 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1838 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1844 * Time out all uncompleted requests in request set pointed by \a data
1845 * Callback used when waiting on sets with l_wait_event.
1848 int ptlrpc_expired_set(void *data)
1850 struct ptlrpc_request_set *set = data;
1852 time_t now = cfs_time_current_sec();
1855 LASSERT(set != NULL);
1858 * A timeout expired. See which reqs it applies to...
1860 cfs_list_for_each (tmp, &set->set_requests) {
1861 struct ptlrpc_request *req =
1862 cfs_list_entry(tmp, struct ptlrpc_request,
1865 /* don't expire request waiting for context */
1866 if (req->rq_wait_ctx)
1869 /* Request in-flight? */
1870 if (!((req->rq_phase == RQ_PHASE_RPC &&
1871 !req->rq_waiting && !req->rq_resend) ||
1872 (req->rq_phase == RQ_PHASE_BULK)))
1875 if (req->rq_timedout || /* already dealt with */
1876 req->rq_deadline > now) /* not expired */
1879 /* Deal with this guy. Do it asynchronously to not block
1880 * ptlrpcd thread. */
1881 ptlrpc_expire_one_request(req, 1);
1885 * When waiting for a whole set, we always break out of the
1886 * sleep so we can recalculate the timeout, or enable interrupts
1887 * if everyone's timed out.
1893 * Sets rq_intr flag in \a req under spinlock.
1895 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1897 cfs_spin_lock(&req->rq_lock);
1899 cfs_spin_unlock(&req->rq_lock);
1903 * Interrupts (sets interrupted flag) all uncompleted requests in
1904 * a set \a data. Callback for l_wait_event for interruptible waits.
1906 void ptlrpc_interrupted_set(void *data)
1908 struct ptlrpc_request_set *set = data;
1911 LASSERT(set != NULL);
1912 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
1914 cfs_list_for_each(tmp, &set->set_requests) {
1915 struct ptlrpc_request *req =
1916 cfs_list_entry(tmp, struct ptlrpc_request,
1919 if (req->rq_phase != RQ_PHASE_RPC &&
1920 req->rq_phase != RQ_PHASE_UNREGISTERING)
1923 ptlrpc_mark_interrupted(req);
1928 * Get the smallest timeout in the set; this does NOT set a timeout.
1930 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1933 time_t now = cfs_time_current_sec();
1935 struct ptlrpc_request *req;
1939 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1941 cfs_list_for_each(tmp, &set->set_requests) {
1942 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1945 * Request in-flight?
1947 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1948 (req->rq_phase == RQ_PHASE_BULK) ||
1949 (req->rq_phase == RQ_PHASE_NEW)))
1953 * Already timed out.
1955 if (req->rq_timedout)
1961 if (req->rq_wait_ctx)
1964 if (req->rq_phase == RQ_PHASE_NEW)
1965 deadline = req->rq_sent;
1967 deadline = req->rq_sent + req->rq_timeout;
1969 if (deadline <= now) /* actually expired already */
1970 timeout = 1; /* ASAP */
1971 else if (timeout == 0 || timeout > deadline - now)
1972 timeout = deadline - now;
1978 * Send all unset request from the set and then wait untill all
1979 * requests in the set complete (either get a reply, timeout, get an
1980 * error or otherwise be interrupted).
1981 * Returns 0 on success or error code otherwise.
1983 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1986 struct ptlrpc_request *req;
1987 struct l_wait_info lwi;
1991 if (cfs_list_empty(&set->set_requests))
1994 cfs_list_for_each(tmp, &set->set_requests) {
1995 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1996 if (req->rq_phase == RQ_PHASE_NEW)
1997 (void)ptlrpc_send_new_req(req);
2001 timeout = ptlrpc_set_next_timeout(set);
2003 /* wait until all complete, interrupted, or an in-flight
2005 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2008 if (timeout == 0 && !cfs_signal_pending())
2010 * No requests are in-flight (ether timed out
2011 * or delayed), so we can allow interrupts.
2012 * We still want to block for a limited time,
2013 * so we allow interrupts during the timeout.
2015 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2017 ptlrpc_interrupted_set, set);
2020 * At least one request is in flight, so no
2021 * interrupts are allowed. Wait until all
2022 * complete, or an in-flight req times out.
2024 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2025 ptlrpc_expired_set, set);
2027 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2029 /* LU-769 - if we ignored the signal because it was already
2030 * pending when we started, we need to handle it now or we risk
2031 * it being ignored forever */
2032 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2033 cfs_signal_pending()) {
2034 cfs_sigset_t blocked_sigs =
2035 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2037 /* In fact we only interrupt for the "fatal" signals
2038 * like SIGINT or SIGKILL. We still ignore less
2039 * important signals since ptlrpc set is not easily
2040 * reentrant from userspace again */
2041 if (cfs_signal_pending())
2042 ptlrpc_interrupted_set(set);
2043 cfs_block_sigs(blocked_sigs);
2046 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2048 /* -EINTR => all requests have been flagged rq_intr so next
2050 * -ETIMEDOUT => someone timed out. When all reqs have
2051 * timed out, signals are enabled allowing completion with
2053 * I don't really care if we go once more round the loop in
2054 * the error cases -eeb. */
2055 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2056 cfs_list_for_each(tmp, &set->set_requests) {
2057 req = cfs_list_entry(tmp, struct ptlrpc_request,
2059 cfs_spin_lock(&req->rq_lock);
2060 req->rq_invalid_rqset = 1;
2061 cfs_spin_unlock(&req->rq_lock);
2064 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2066 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2069 cfs_list_for_each(tmp, &set->set_requests) {
2070 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2072 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2073 if (req->rq_status != 0)
2074 rc = req->rq_status;
2077 if (set->set_interpret != NULL) {
2078 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2080 rc = interpreter (set, set->set_arg, rc);
2082 struct ptlrpc_set_cbdata *cbdata, *n;
2085 cfs_list_for_each_entry_safe(cbdata, n,
2086 &set->set_cblist, psc_item) {
2087 cfs_list_del_init(&cbdata->psc_item);
2088 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2091 OBD_FREE_PTR(cbdata);
2099 * Helper fuction for request freeing.
2100 * Called when request count reached zero and request needs to be freed.
2101 * Removes request from all sorts of sending/replay lists it might be on,
2102 * frees network buffers if any are present.
2103 * If \a locked is set, that means caller is already holding import imp_lock
2104 * and so we no longer need to reobtain it (for certain lists manipulations)
2106 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2109 if (request == NULL) {
2114 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2115 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2116 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2117 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2118 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2119 LASSERTF(!request->rq_replay, "req %p\n", request);
2121 req_capsule_fini(&request->rq_pill);
2123 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2124 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2125 if (request->rq_import != NULL) {
2127 cfs_spin_lock(&request->rq_import->imp_lock);
2128 cfs_list_del_init(&request->rq_replay_list);
2130 cfs_spin_unlock(&request->rq_import->imp_lock);
2132 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2134 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2135 DEBUG_REQ(D_ERROR, request,
2136 "freeing request with nonzero refcount");
2140 if (request->rq_repbuf != NULL)
2141 sptlrpc_cli_free_repbuf(request);
2142 if (request->rq_export != NULL) {
2143 class_export_put(request->rq_export);
2144 request->rq_export = NULL;
2146 if (request->rq_import != NULL) {
2147 class_import_put(request->rq_import);
2148 request->rq_import = NULL;
2150 if (request->rq_bulk != NULL)
2151 ptlrpc_free_bulk(request->rq_bulk);
2153 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2154 sptlrpc_cli_free_reqbuf(request);
2156 if (request->rq_cli_ctx)
2157 sptlrpc_req_put_ctx(request, !locked);
2159 if (request->rq_pool)
2160 __ptlrpc_free_req_to_pool(request);
2162 OBD_FREE(request, sizeof(*request));
2166 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2168 * Drop one request reference. Must be called with import imp_lock held.
2169 * When reference count drops to zero, reuqest is freed.
2171 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2173 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2174 (void)__ptlrpc_req_finished(request, 1);
2179 * Drops one reference count for request \a request.
2180 * \a locked set indicates that caller holds import imp_lock.
2181 * Frees the request whe reference count reaches zero.
2183 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2186 if (request == NULL)
2189 if (request == LP_POISON ||
2190 request->rq_reqmsg == LP_POISON) {
2191 CERROR("dereferencing freed request (bug 575)\n");
2196 DEBUG_REQ(D_INFO, request, "refcount now %u",
2197 cfs_atomic_read(&request->rq_refcount) - 1);
2199 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2200 __ptlrpc_free_req(request, locked);
2208 * Drops one reference count for a request.
2210 void ptlrpc_req_finished(struct ptlrpc_request *request)
2212 __ptlrpc_req_finished(request, 0);
2216 * Returns xid of a \a request
2218 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2220 return request->rq_xid;
2222 EXPORT_SYMBOL(ptlrpc_req_xid);
2225 * Disengage the client's reply buffer from the network
2226 * NB does _NOT_ unregister any client-side bulk.
2227 * IDEMPOTENT, but _not_ safe against concurrent callers.
2228 * The request owner (i.e. the thread doing the I/O) must call...
2229 * Returns 0 on success or 1 if unregistering cannot be made.
2231 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2235 struct l_wait_info lwi;
2240 LASSERT(!cfs_in_interrupt());
2243 * Let's setup deadline for reply unlink.
2245 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2246 async && request->rq_reply_deadline == 0)
2247 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2250 * Nothing left to do.
2252 if (!ptlrpc_client_recv_or_unlink(request))
2255 LNetMDUnlink(request->rq_reply_md_h);
2258 * Let's check it once again.
2260 if (!ptlrpc_client_recv_or_unlink(request))
2264 * Move to "Unregistering" phase as reply was not unlinked yet.
2266 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2269 * Do not wait for unlink to finish.
2275 * We have to l_wait_event() whatever the result, to give liblustre
2276 * a chance to run reply_in_callback(), and to make sure we've
2277 * unlinked before returning a req to the pool.
2279 if (request->rq_set != NULL)
2280 wq = &request->rq_set->set_waitq;
2282 wq = &request->rq_reply_waitq;
2285 /* Network access will complete in finite time but the HUGE
2286 * timeout lets us CWARN for visibility of sluggish NALs */
2287 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2288 cfs_time_seconds(1), NULL, NULL);
2289 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2292 ptlrpc_rqphase_move(request, request->rq_next_phase);
2296 LASSERT(rc == -ETIMEDOUT);
2297 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2298 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2299 request->rq_must_unlink);
2305 * Iterates through replay_list on import and prunes
2306 * all requests have transno smaller than last_committed for the
2307 * import and don't have rq_replay set.
2308 * Since requests are sorted in transno order, stops when meetign first
2309 * transno bigger than last_committed.
2310 * caller must hold imp->imp_lock
2312 void ptlrpc_free_committed(struct obd_import *imp)
2314 cfs_list_t *tmp, *saved;
2315 struct ptlrpc_request *req;
2316 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2319 LASSERT(imp != NULL);
2321 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2324 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2325 imp->imp_generation == imp->imp_last_generation_checked) {
2326 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2327 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2331 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2332 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2333 imp->imp_generation);
2334 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2335 imp->imp_last_generation_checked = imp->imp_generation;
2337 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2338 req = cfs_list_entry(tmp, struct ptlrpc_request,
2341 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2342 LASSERT(req != last_req);
2345 if (req->rq_transno == 0) {
2346 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2349 if (req->rq_import_generation < imp->imp_generation) {
2350 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2354 if (req->rq_replay) {
2355 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2359 /* not yet committed */
2360 if (req->rq_transno > imp->imp_peer_committed_transno) {
2361 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2365 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2366 imp->imp_peer_committed_transno);
2368 cfs_spin_lock(&req->rq_lock);
2370 cfs_spin_unlock(&req->rq_lock);
2371 if (req->rq_commit_cb != NULL)
2372 req->rq_commit_cb(req);
2373 cfs_list_del_init(&req->rq_replay_list);
2374 __ptlrpc_req_finished(req, 1);
2381 void ptlrpc_cleanup_client(struct obd_import *imp)
2389 * Schedule previously sent request for resend.
2390 * For bulk requests we assign new xid (to avoid problems with
2391 * lost replies and therefore several transfers landing into same buffer
2392 * from different sending attempts).
2394 void ptlrpc_resend_req(struct ptlrpc_request *req)
2396 DEBUG_REQ(D_HA, req, "going to resend");
2397 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2398 req->rq_status = -EAGAIN;
2400 cfs_spin_lock(&req->rq_lock);
2402 req->rq_net_err = 0;
2403 req->rq_timedout = 0;
2405 __u64 old_xid = req->rq_xid;
2407 /* ensure previous bulk fails */
2408 req->rq_xid = ptlrpc_next_xid();
2409 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2410 old_xid, req->rq_xid);
2412 ptlrpc_client_wake_req(req);
2413 cfs_spin_unlock(&req->rq_lock);
2416 /* XXX: this function and rq_status are currently unused */
2417 void ptlrpc_restart_req(struct ptlrpc_request *req)
2419 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2420 req->rq_status = -ERESTARTSYS;
2422 cfs_spin_lock(&req->rq_lock);
2423 req->rq_restart = 1;
2424 req->rq_timedout = 0;
2425 ptlrpc_client_wake_req(req);
2426 cfs_spin_unlock(&req->rq_lock);
2430 * Grab additional reference on a request \a req
2432 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2435 cfs_atomic_inc(&req->rq_refcount);
2440 * Add a request to import replay_list.
2441 * Must be called under imp_lock
2443 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2444 struct obd_import *imp)
2448 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2450 if (req->rq_transno == 0) {
2451 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2455 /* clear this for new requests that were resent as well
2456 as resent replayed requests. */
2457 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2459 /* don't re-add requests that have been replayed */
2460 if (!cfs_list_empty(&req->rq_replay_list))
2463 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2465 LASSERT(imp->imp_replayable);
2466 /* Balanced in ptlrpc_free_committed, usually. */
2467 ptlrpc_request_addref(req);
2468 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2469 struct ptlrpc_request *iter =
2470 cfs_list_entry(tmp, struct ptlrpc_request,
2473 /* We may have duplicate transnos if we create and then
2474 * open a file, or for closes retained if to match creating
2475 * opens, so use req->rq_xid as a secondary key.
2476 * (See bugs 684, 685, and 428.)
2477 * XXX no longer needed, but all opens need transnos!
2479 if (iter->rq_transno > req->rq_transno)
2482 if (iter->rq_transno == req->rq_transno) {
2483 LASSERT(iter->rq_xid != req->rq_xid);
2484 if (iter->rq_xid > req->rq_xid)
2488 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2492 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2496 * Send request and wait until it completes.
2497 * Returns request processing status.
2499 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2501 struct ptlrpc_request_set *set;
2505 LASSERT(req->rq_set == NULL);
2506 LASSERT(!req->rq_receiving_reply);
2508 set = ptlrpc_prep_set();
2510 CERROR("Unable to allocate ptlrpc set.");
2514 /* for distributed debugging */
2515 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2517 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2518 ptlrpc_request_addref(req);
2519 ptlrpc_set_add_req(set, req);
2520 rc = ptlrpc_set_wait(set);
2521 ptlrpc_set_destroy(set);
2526 struct ptlrpc_replay_async_args {
2528 int praa_old_status;
2532 * Callback used for replayed requests reply processing.
2533 * In case of succesful reply calls registeresd request replay callback.
2534 * In case of error restart replay process.
2536 static int ptlrpc_replay_interpret(const struct lu_env *env,
2537 struct ptlrpc_request *req,
2538 void * data, int rc)
2540 struct ptlrpc_replay_async_args *aa = data;
2541 struct obd_import *imp = req->rq_import;
2544 cfs_atomic_dec(&imp->imp_replay_inflight);
2546 if (!ptlrpc_client_replied(req)) {
2547 CERROR("request replay timed out, restarting recovery\n");
2548 GOTO(out, rc = -ETIMEDOUT);
2551 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2552 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2553 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2554 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2556 /** VBR: check version failure */
2557 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2558 /** replay was failed due to version mismatch */
2559 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2560 cfs_spin_lock(&imp->imp_lock);
2561 imp->imp_vbr_failed = 1;
2562 imp->imp_no_lock_replay = 1;
2563 cfs_spin_unlock(&imp->imp_lock);
2564 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2566 /** The transno had better not change over replay. */
2567 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2568 lustre_msg_get_transno(req->rq_repmsg) ||
2569 lustre_msg_get_transno(req->rq_repmsg) == 0,
2571 lustre_msg_get_transno(req->rq_reqmsg),
2572 lustre_msg_get_transno(req->rq_repmsg));
2575 cfs_spin_lock(&imp->imp_lock);
2576 /** if replays by version then gap was occur on server, no trust to locks */
2577 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2578 imp->imp_no_lock_replay = 1;
2579 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2580 cfs_spin_unlock(&imp->imp_lock);
2581 LASSERT(imp->imp_last_replay_transno);
2583 /* transaction number shouldn't be bigger than the latest replayed */
2584 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2585 DEBUG_REQ(D_ERROR, req,
2586 "Reported transno "LPU64" is bigger than the "
2587 "replayed one: "LPU64, req->rq_transno,
2588 lustre_msg_get_transno(req->rq_reqmsg));
2589 GOTO(out, rc = -EINVAL);
2592 DEBUG_REQ(D_HA, req, "got rep");
2594 /* let the callback do fixups, possibly including in the request */
2595 if (req->rq_replay_cb)
2596 req->rq_replay_cb(req);
2598 if (ptlrpc_client_replied(req) &&
2599 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2600 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2601 lustre_msg_get_status(req->rq_repmsg),
2602 aa->praa_old_status);
2604 /* Put it back for re-replay. */
2605 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2609 * Errors while replay can set transno to 0, but
2610 * imp_last_replay_transno shouldn't be set to 0 anyway
2612 if (req->rq_transno == 0)
2613 CERROR("Transno is 0 during replay!\n");
2615 /* continue with recovery */
2616 rc = ptlrpc_import_recovery_state_machine(imp);
2618 req->rq_send_state = aa->praa_old_state;
2621 /* this replay failed, so restart recovery */
2622 ptlrpc_connect_import(imp);
2628 * Prepares and queues request for replay.
2629 * Adds it to ptlrpcd queue for actual sending.
2630 * Returns 0 on success.
2632 int ptlrpc_replay_req(struct ptlrpc_request *req)
2634 struct ptlrpc_replay_async_args *aa;
2637 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2639 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2640 aa = ptlrpc_req_async_args(req);
2641 memset(aa, 0, sizeof *aa);
2643 /* Prepare request to be resent with ptlrpcd */
2644 aa->praa_old_state = req->rq_send_state;
2645 req->rq_send_state = LUSTRE_IMP_REPLAY;
2646 req->rq_phase = RQ_PHASE_NEW;
2647 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2649 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2651 req->rq_interpret_reply = ptlrpc_replay_interpret;
2652 /* Readjust the timeout for current conditions */
2653 ptlrpc_at_set_req_timeout(req);
2655 /* Tell server the net_latency, so the server can calculate how long
2656 * it should wait for next replay */
2657 lustre_msg_set_service_time(req->rq_reqmsg,
2658 ptlrpc_at_get_net_latency(req));
2659 DEBUG_REQ(D_HA, req, "REPLAY");
2661 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2662 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2664 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2669 * Aborts all in-flight request on import \a imp sending and delayed lists
2671 void ptlrpc_abort_inflight(struct obd_import *imp)
2673 cfs_list_t *tmp, *n;
2676 /* Make sure that no new requests get processed for this import.
2677 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2678 * this flag and then putting requests on sending_list or delayed_list.
2680 cfs_spin_lock(&imp->imp_lock);
2682 /* XXX locking? Maybe we should remove each request with the list
2683 * locked? Also, how do we know if the requests on the list are
2684 * being freed at this time?
2686 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2687 struct ptlrpc_request *req =
2688 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2690 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2692 cfs_spin_lock (&req->rq_lock);
2693 if (req->rq_import_generation < imp->imp_generation) {
2695 req->rq_status = -EINTR;
2696 ptlrpc_client_wake_req(req);
2698 cfs_spin_unlock (&req->rq_lock);
2701 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2702 struct ptlrpc_request *req =
2703 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2705 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2707 cfs_spin_lock (&req->rq_lock);
2708 if (req->rq_import_generation < imp->imp_generation) {
2710 req->rq_status = -EINTR;
2711 ptlrpc_client_wake_req(req);
2713 cfs_spin_unlock (&req->rq_lock);
2716 /* Last chance to free reqs left on the replay list, but we
2717 * will still leak reqs that haven't committed. */
2718 if (imp->imp_replayable)
2719 ptlrpc_free_committed(imp);
2721 cfs_spin_unlock(&imp->imp_lock);
2727 * Abort all uncompleted requests in request set \a set
2729 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2731 cfs_list_t *tmp, *pos;
2733 LASSERT(set != NULL);
2735 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2736 struct ptlrpc_request *req =
2737 cfs_list_entry(pos, struct ptlrpc_request,
2740 cfs_spin_lock(&req->rq_lock);
2741 if (req->rq_phase != RQ_PHASE_RPC) {
2742 cfs_spin_unlock(&req->rq_lock);
2747 req->rq_status = -EINTR;
2748 ptlrpc_client_wake_req(req);
2749 cfs_spin_unlock(&req->rq_lock);
2753 static __u64 ptlrpc_last_xid;
2754 static cfs_spinlock_t ptlrpc_last_xid_lock;
2757 * Initialize the XID for the node. This is common among all requests on
2758 * this node, and only requires the property that it is monotonically
2759 * increasing. It does not need to be sequential. Since this is also used
2760 * as the RDMA match bits, it is important that a single client NOT have
2761 * the same match bits for two different in-flight requests, hence we do
2762 * NOT want to have an XID per target or similar.
2764 * To avoid an unlikely collision between match bits after a client reboot
2765 * (which would deliver old data into the wrong RDMA buffer) initialize
2766 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2767 * If the time is clearly incorrect, we instead use a 62-bit random number.
2768 * In the worst case the random number will overflow 1M RPCs per second in
2769 * 9133 years, or permutations thereof.
2771 #define YEAR_2004 (1ULL << 30)
2772 void ptlrpc_init_xid(void)
2774 time_t now = cfs_time_current_sec();
2776 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2777 if (now < YEAR_2004) {
2778 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2779 ptlrpc_last_xid >>= 2;
2780 ptlrpc_last_xid |= (1ULL << 61);
2782 ptlrpc_last_xid = (__u64)now << 20;
2787 * Increase xid and returns resultng new value to the caller.
2789 __u64 ptlrpc_next_xid(void)
2792 cfs_spin_lock(&ptlrpc_last_xid_lock);
2793 tmp = ++ptlrpc_last_xid;
2794 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2799 * Get a glimpse at what next xid value might have been.
2800 * Returns possible next xid.
2802 __u64 ptlrpc_sample_next_xid(void)
2804 #if BITS_PER_LONG == 32
2805 /* need to avoid possible word tearing on 32-bit systems */
2807 cfs_spin_lock(&ptlrpc_last_xid_lock);
2808 tmp = ptlrpc_last_xid + 1;
2809 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2812 /* No need to lock, since returned value is racy anyways */
2813 return ptlrpc_last_xid + 1;
2816 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2819 * Functions for operating ptlrpc workers.
2821 * A ptlrpc work is a function which will be running inside ptlrpc context.
2822 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2824 * 1. after a work is created, it can be used many times, that is:
2825 * handler = ptlrpcd_alloc_work();
2826 * ptlrpcd_queue_work();
2828 * queue it again when necessary:
2829 * ptlrpcd_queue_work();
2830 * ptlrpcd_destroy_work();
2831 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2832 * it will only be queued once in any time. Also as its name implies, it may
2833 * have delay before it really runs by ptlrpcd thread.
2835 struct ptlrpc_work_async_args {
2837 int (*cb)(const struct lu_env *, void *);
2841 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2843 static int work_interpreter(const struct lu_env *env,
2844 struct ptlrpc_request *req, void *data, int rc)
2846 struct ptlrpc_work_async_args *arg = data;
2848 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2849 LASSERT(arg->cb != NULL);
2851 return arg->cb(env, arg->cbdata);
2855 * Create a work for ptlrpc.
2857 void *ptlrpcd_alloc_work(struct obd_import *imp,
2858 int (*cb)(const struct lu_env *, void *), void *cbdata)
2860 struct ptlrpc_request *req = NULL;
2861 struct ptlrpc_work_async_args *args;
2867 RETURN(ERR_PTR(-EINVAL));
2869 /* copy some code from deprecated fakereq. */
2872 CERROR("ptlrpc: run out of memory!\n");
2873 RETURN(ERR_PTR(-ENOMEM));
2876 req->rq_send_state = LUSTRE_IMP_FULL;
2877 req->rq_type = PTL_RPC_MSG_REQUEST;
2878 req->rq_import = class_import_get(imp);
2879 req->rq_export = NULL;
2880 req->rq_interpret_reply = work_interpreter;
2881 /* don't want reply */
2882 req->rq_receiving_reply = 0;
2883 req->rq_must_unlink = 0;
2884 req->rq_no_delay = req->rq_no_resend = 1;
2886 cfs_spin_lock_init(&req->rq_lock);
2887 CFS_INIT_LIST_HEAD(&req->rq_list);
2888 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
2889 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
2890 CFS_INIT_LIST_HEAD(&req->rq_history_list);
2891 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
2892 cfs_waitq_init(&req->rq_reply_waitq);
2893 cfs_waitq_init(&req->rq_set_waitq);
2894 cfs_atomic_set(&req->rq_refcount, 1);
2896 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
2897 args = ptlrpc_req_async_args(req);
2898 args->magic = PTLRPC_WORK_MAGIC;
2900 args->cbdata = cbdata;
2904 EXPORT_SYMBOL(ptlrpcd_alloc_work);
2906 void ptlrpcd_destroy_work(void *handler)
2908 struct ptlrpc_request *req = handler;
2911 ptlrpc_req_finished(req);
2913 EXPORT_SYMBOL(ptlrpcd_destroy_work);
2915 int ptlrpcd_queue_work(void *handler)
2917 struct ptlrpc_request *req = handler;
2920 * Check if the req is already being queued.
2922 * Here comes a trick: it lacks a way of checking if a req is being
2923 * processed reliably in ptlrpc. Here I have to use refcount of req
2924 * for this purpose. This is okay because the caller should use this
2925 * req as opaque data. - Jinshan
2927 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
2928 if (cfs_atomic_read(&req->rq_refcount) > 1)
2931 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
2932 cfs_atomic_dec(&req->rq_refcount);
2936 /* re-initialize the req */
2937 req->rq_timeout = obd_timeout;
2938 req->rq_sent = cfs_time_current_sec();
2939 req->rq_deadline = req->rq_sent + req->rq_timeout;
2940 req->rq_reply_deadline = req->rq_deadline;
2941 req->rq_phase = RQ_PHASE_INTERPRET;
2942 req->rq_next_phase = RQ_PHASE_COMPLETE;
2943 req->rq_xid = ptlrpc_next_xid();
2944 req->rq_import_generation = req->rq_import->imp_generation;
2946 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
2949 EXPORT_SYMBOL(ptlrpcd_queue_work);