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, Intel Corporation.
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"
55 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
58 * Initialize passed in client structure \a cl.
60 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
61 struct ptlrpc_client *cl)
63 cl->cli_request_portal = req_portal;
64 cl->cli_reply_portal = rep_portal;
67 EXPORT_SYMBOL(ptlrpc_init_client);
70 * Return PortalRPC connection for remore uud \a uuid
72 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
74 struct ptlrpc_connection *c;
76 lnet_process_id_t peer;
79 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
80 * before accessing its values. */
81 /* coverity[uninit_use_in_call] */
82 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
84 CNETERR("cannot find peer %s!\n", uuid->uuid);
88 c = ptlrpc_connection_get(peer, self, uuid);
90 memcpy(c->c_remote_uuid.uuid,
91 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
94 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
98 EXPORT_SYMBOL(ptlrpc_uuid_to_connection);
101 * Allocate and initialize new bulk descriptor
102 * Returns pointer to the descriptor or NULL on error.
104 struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
106 struct ptlrpc_bulk_desc *desc;
108 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
112 spin_lock_init(&desc->bd_lock);
113 cfs_waitq_init(&desc->bd_waitq);
114 desc->bd_max_iov = npages;
115 desc->bd_iov_count = 0;
116 LNetInvalidateHandle(&desc->bd_md_h);
117 desc->bd_portal = portal;
118 desc->bd_type = type;
124 * Prepare bulk descriptor for specified outgoing request \a req that
125 * can fit \a npages * pages. \a type is bulk type. \a portal is where
126 * the bulk to be sent. Used on client-side.
127 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
130 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
131 int npages, int type, int portal)
133 struct obd_import *imp = req->rq_import;
134 struct ptlrpc_bulk_desc *desc;
137 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
138 desc = new_bulk(npages, type, portal);
142 desc->bd_import_generation = req->rq_import_generation;
143 desc->bd_import = class_import_get(imp);
146 desc->bd_cbid.cbid_fn = client_bulk_callback;
147 desc->bd_cbid.cbid_arg = desc;
149 /* This makes req own desc, and free it when she frees herself */
154 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
157 * Add a page \a page to the bulk descriptor \a desc.
158 * Data to transfer in the page starts at offset \a pageoffset and
159 * amount of data to transfer from the page is \a len
161 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
162 cfs_page_t *page, int pageoffset, int len, int pin)
164 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
165 LASSERT(page != NULL);
166 LASSERT(pageoffset >= 0);
168 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
175 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
177 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
180 * Uninitialize and free bulk descriptor \a desc.
181 * Works on bulk descriptors both from server and client side.
183 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin)
188 LASSERT(desc != NULL);
189 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
190 LASSERT(!desc->bd_network_rw); /* network hands off or */
191 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
193 sptlrpc_enc_pool_put_pages(desc);
196 class_export_put(desc->bd_export);
198 class_import_put(desc->bd_import);
201 for (i = 0; i < desc->bd_iov_count ; i++)
202 cfs_page_unpin(desc->bd_iov[i].kiov_page);
205 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
206 bd_iov[desc->bd_max_iov]));
209 EXPORT_SYMBOL(__ptlrpc_free_bulk);
212 * Set server timelimit for this req, i.e. how long are we willing to wait
213 * for reply before timing out this request.
215 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
221 LASSERT(req->rq_import);
224 /* non-AT settings */
226 * \a imp_server_timeout means this is reverse import and
227 * we send (currently only) ASTs to the client and cannot afford
228 * to wait too long for the reply, otherwise the other client
229 * (because of which we are sending this request) would
230 * timeout waiting for us
232 req->rq_timeout = req->rq_import->imp_server_timeout ?
233 obd_timeout / 2 : obd_timeout;
235 at = &req->rq_import->imp_at;
236 idx = import_at_get_index(req->rq_import,
237 req->rq_request_portal);
238 serv_est = at_get(&at->iat_service_estimate[idx]);
239 req->rq_timeout = at_est2timeout(serv_est);
241 /* We could get even fancier here, using history to predict increased
244 /* Let the server know what this RPC timeout is by putting it in the
246 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
248 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
250 /* Adjust max service estimate based on server value */
251 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
252 unsigned int serv_est)
258 LASSERT(req->rq_import);
259 at = &req->rq_import->imp_at;
261 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
262 /* max service estimates are tracked on the server side,
263 so just keep minimal history here */
264 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
266 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
267 "has changed from %d to %d\n",
268 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
269 oldse, at_get(&at->iat_service_estimate[idx]));
272 /* Expected network latency per remote node (secs) */
273 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
275 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
278 /* Adjust expected network latency */
279 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
280 unsigned int service_time)
282 unsigned int nl, oldnl;
284 time_t now = cfs_time_current_sec();
286 LASSERT(req->rq_import);
287 at = &req->rq_import->imp_at;
289 /* Network latency is total time less server processing time */
290 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
291 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
292 CWARN("Reported service time %u > total measured time "
293 CFS_DURATION_T"\n", service_time,
294 cfs_time_sub(now, req->rq_sent));
296 oldnl = at_measured(&at->iat_net_latency, nl);
298 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
299 "has changed from %d to %d\n",
300 req->rq_import->imp_obd->obd_name,
302 &req->rq_import->imp_connection->c_remote_uuid),
303 oldnl, at_get(&at->iat_net_latency));
306 static int unpack_reply(struct ptlrpc_request *req)
310 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
311 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
313 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
318 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
320 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
327 * Handle an early reply message, called with the rq_lock held.
328 * If anything goes wrong just ignore it - same as if it never happened
330 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
332 struct ptlrpc_request *early_req;
338 spin_unlock(&req->rq_lock);
340 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
342 spin_lock(&req->rq_lock);
346 rc = unpack_reply(early_req);
348 /* Expecting to increase the service time estimate here */
349 ptlrpc_at_adj_service(req,
350 lustre_msg_get_timeout(early_req->rq_repmsg));
351 ptlrpc_at_adj_net_latency(req,
352 lustre_msg_get_service_time(early_req->rq_repmsg));
355 sptlrpc_cli_finish_early_reply(early_req);
358 spin_lock(&req->rq_lock);
362 /* Adjust the local timeout for this req */
363 ptlrpc_at_set_req_timeout(req);
365 spin_lock(&req->rq_lock);
366 olddl = req->rq_deadline;
367 /* server assumes it now has rq_timeout from when it sent the
368 * early reply, so client should give it at least that long. */
369 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
370 ptlrpc_at_get_net_latency(req);
372 DEBUG_REQ(D_ADAPTTO, req,
373 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
374 "("CFS_DURATION_T"s)", req->rq_early_count,
375 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
376 cfs_time_sub(req->rq_deadline, olddl));
382 * Wind down request pool \a pool.
383 * Frees all requests from the pool too
385 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
388 struct ptlrpc_request *req;
390 LASSERT(pool != NULL);
392 spin_lock(&pool->prp_lock);
393 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
394 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
395 cfs_list_del(&req->rq_list);
396 LASSERT(req->rq_reqbuf);
397 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
398 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
399 OBD_FREE(req, sizeof(*req));
401 spin_unlock(&pool->prp_lock);
402 OBD_FREE(pool, sizeof(*pool));
404 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
407 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
409 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
414 while (size < pool->prp_rq_size)
417 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
418 size == pool->prp_rq_size,
419 "Trying to change pool size with nonempty pool "
420 "from %d to %d bytes\n", pool->prp_rq_size, size);
422 spin_lock(&pool->prp_lock);
423 pool->prp_rq_size = size;
424 for (i = 0; i < num_rq; i++) {
425 struct ptlrpc_request *req;
426 struct lustre_msg *msg;
428 spin_unlock(&pool->prp_lock);
429 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
432 OBD_ALLOC_LARGE(msg, size);
434 OBD_FREE(req, sizeof(struct ptlrpc_request));
437 req->rq_reqbuf = msg;
438 req->rq_reqbuf_len = size;
440 spin_lock(&pool->prp_lock);
441 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
443 spin_unlock(&pool->prp_lock);
446 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
449 * Create and initialize new request pool with given attributes:
450 * \a num_rq - initial number of requests to create for the pool
451 * \a msgsize - maximum message size possible for requests in thid pool
452 * \a populate_pool - function to be called when more requests need to be added
454 * Returns pointer to newly created pool or NULL on error.
456 struct ptlrpc_request_pool *
457 ptlrpc_init_rq_pool(int num_rq, int msgsize,
458 void (*populate_pool)(struct ptlrpc_request_pool *, int))
460 struct ptlrpc_request_pool *pool;
462 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
466 /* Request next power of two for the allocation, because internally
467 kernel would do exactly this */
469 spin_lock_init(&pool->prp_lock);
470 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
471 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
472 pool->prp_populate = populate_pool;
474 populate_pool(pool, num_rq);
476 if (cfs_list_empty(&pool->prp_req_list)) {
477 /* have not allocated a single request for the pool */
478 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
483 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
486 * Fetches one request from pool \a pool
488 static struct ptlrpc_request *
489 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
491 struct ptlrpc_request *request;
492 struct lustre_msg *reqbuf;
497 spin_lock(&pool->prp_lock);
499 /* See if we have anything in a pool, and bail out if nothing,
500 * in writeout path, where this matters, this is safe to do, because
501 * nothing is lost in this case, and when some in-flight requests
502 * complete, this code will be called again. */
503 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
504 spin_unlock(&pool->prp_lock);
508 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
510 cfs_list_del_init(&request->rq_list);
511 spin_unlock(&pool->prp_lock);
513 LASSERT(request->rq_reqbuf);
514 LASSERT(request->rq_pool);
516 reqbuf = request->rq_reqbuf;
517 memset(request, 0, sizeof(*request));
518 request->rq_reqbuf = reqbuf;
519 request->rq_reqbuf_len = pool->prp_rq_size;
520 request->rq_pool = pool;
526 * Returns freed \a request to pool.
528 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
530 struct ptlrpc_request_pool *pool = request->rq_pool;
532 spin_lock(&pool->prp_lock);
533 LASSERT(cfs_list_empty(&request->rq_list));
534 LASSERT(!request->rq_receiving_reply);
535 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
536 spin_unlock(&pool->prp_lock);
539 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
540 __u32 version, int opcode,
541 int count, __u32 *lengths, char **bufs,
542 struct ptlrpc_cli_ctx *ctx)
544 struct obd_import *imp = request->rq_import;
549 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
551 rc = sptlrpc_req_get_ctx(request);
556 sptlrpc_req_set_flavor(request, opcode);
558 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
561 LASSERT(!request->rq_pool);
565 lustre_msg_add_version(request->rq_reqmsg, version);
566 request->rq_send_state = LUSTRE_IMP_FULL;
567 request->rq_type = PTL_RPC_MSG_REQUEST;
568 request->rq_export = NULL;
570 request->rq_req_cbid.cbid_fn = request_out_callback;
571 request->rq_req_cbid.cbid_arg = request;
573 request->rq_reply_cbid.cbid_fn = reply_in_callback;
574 request->rq_reply_cbid.cbid_arg = request;
576 request->rq_reply_deadline = 0;
577 request->rq_phase = RQ_PHASE_NEW;
578 request->rq_next_phase = RQ_PHASE_UNDEFINED;
580 request->rq_request_portal = imp->imp_client->cli_request_portal;
581 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
583 ptlrpc_at_set_req_timeout(request);
585 spin_lock_init(&request->rq_lock);
586 CFS_INIT_LIST_HEAD(&request->rq_list);
587 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
588 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
589 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
590 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
591 CFS_INIT_LIST_HEAD(&request->rq_history_list);
592 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
593 cfs_waitq_init(&request->rq_reply_waitq);
594 cfs_waitq_init(&request->rq_set_waitq);
595 request->rq_xid = ptlrpc_next_xid();
596 cfs_atomic_set(&request->rq_refcount, 1);
598 lustre_msg_set_opc(request->rq_reqmsg, opcode);
602 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
604 class_import_put(imp);
608 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
609 __u32 version, int opcode, char **bufs,
610 struct ptlrpc_cli_ctx *ctx)
614 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
615 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
616 request->rq_pill.rc_area[RCL_CLIENT],
619 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
622 * Pack request buffers for network transfer, performing necessary encryption
623 * steps if necessary.
625 int ptlrpc_request_pack(struct ptlrpc_request *request,
626 __u32 version, int opcode)
629 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
633 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
634 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
635 * have to send old ptlrpc_body to keep interoprability with these
638 * Only three kinds of server->client RPCs so far:
643 * XXX This should be removed whenever we drop the interoprability with
644 * the these old clients.
646 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
647 opcode == LDLM_GL_CALLBACK)
648 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
649 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
653 EXPORT_SYMBOL(ptlrpc_request_pack);
656 * Helper function to allocate new request on import \a imp
657 * and possibly using existing request from pool \a pool if provided.
658 * Returns allocated request structure with import field filled or
662 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
663 struct ptlrpc_request_pool *pool)
665 struct ptlrpc_request *request = NULL;
668 request = ptlrpc_prep_req_from_pool(pool);
671 OBD_ALLOC_PTR(request);
674 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
675 LASSERT(imp != LP_POISON);
676 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
678 LASSERT(imp->imp_client != LP_POISON);
680 request->rq_import = class_import_get(imp);
682 CERROR("request allocation out of memory\n");
689 * Helper function for creating a request.
690 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
691 * buffer structures according to capsule template \a format.
692 * Returns allocated request structure pointer or NULL on error.
694 static struct ptlrpc_request *
695 ptlrpc_request_alloc_internal(struct obd_import *imp,
696 struct ptlrpc_request_pool * pool,
697 const struct req_format *format)
699 struct ptlrpc_request *request;
701 request = __ptlrpc_request_alloc(imp, pool);
705 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
706 req_capsule_set(&request->rq_pill, format);
711 * Allocate new request structure for import \a imp and initialize its
712 * buffer structure according to capsule template \a format.
714 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
715 const struct req_format *format)
717 return ptlrpc_request_alloc_internal(imp, NULL, format);
719 EXPORT_SYMBOL(ptlrpc_request_alloc);
722 * Allocate new request structure for import \a imp from pool \a pool and
723 * initialize its buffer structure according to capsule template \a format.
725 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
726 struct ptlrpc_request_pool * pool,
727 const struct req_format *format)
729 return ptlrpc_request_alloc_internal(imp, pool, format);
731 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
734 * For requests not from pool, free memory of the request structure.
735 * For requests obtained from a pool earlier, return request back to pool.
737 void ptlrpc_request_free(struct ptlrpc_request *request)
739 if (request->rq_pool)
740 __ptlrpc_free_req_to_pool(request);
742 OBD_FREE_PTR(request);
744 EXPORT_SYMBOL(ptlrpc_request_free);
747 * Allocate new request for operatione \a opcode and immediatelly pack it for
749 * Only used for simple requests like OBD_PING where the only important
750 * part of the request is operation itself.
751 * Returns allocated request or NULL on error.
753 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
754 const struct req_format *format,
755 __u32 version, int opcode)
757 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
761 rc = ptlrpc_request_pack(req, version, opcode);
763 ptlrpc_request_free(req);
769 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
772 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
773 * for operation \a opcode. Request would contain \a count buffers.
774 * Sizes of buffers are described in array \a lengths and buffers themselves
775 * are provided by a pointer \a bufs.
776 * Returns prepared request structure pointer or NULL on error.
778 struct ptlrpc_request *
779 ptlrpc_prep_req_pool(struct obd_import *imp,
780 __u32 version, int opcode,
781 int count, __u32 *lengths, char **bufs,
782 struct ptlrpc_request_pool *pool)
784 struct ptlrpc_request *request;
787 request = __ptlrpc_request_alloc(imp, pool);
791 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
792 lengths, bufs, NULL);
794 ptlrpc_request_free(request);
799 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
802 * Same as ptlrpc_prep_req_pool, but without pool
804 struct ptlrpc_request *
805 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
806 __u32 *lengths, char **bufs)
808 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
811 EXPORT_SYMBOL(ptlrpc_prep_req);
814 * Allocate and initialize new request set structure.
815 * Returns a pointer to the newly allocated set structure or NULL on error.
817 struct ptlrpc_request_set *ptlrpc_prep_set(void)
819 struct ptlrpc_request_set *set;
822 OBD_ALLOC(set, sizeof *set);
825 cfs_atomic_set(&set->set_refcount, 1);
826 CFS_INIT_LIST_HEAD(&set->set_requests);
827 cfs_waitq_init(&set->set_waitq);
828 cfs_atomic_set(&set->set_new_count, 0);
829 cfs_atomic_set(&set->set_remaining, 0);
830 spin_lock_init(&set->set_new_req_lock);
831 CFS_INIT_LIST_HEAD(&set->set_new_requests);
832 CFS_INIT_LIST_HEAD(&set->set_cblist);
833 set->set_max_inflight = UINT_MAX;
834 set->set_producer = NULL;
835 set->set_producer_arg = NULL;
840 EXPORT_SYMBOL(ptlrpc_prep_set);
843 * Allocate and initialize new request set structure with flow control
844 * extension. This extension allows to control the number of requests in-flight
845 * for the whole set. A callback function to generate requests must be provided
846 * and the request set will keep the number of requests sent over the wire to
848 * Returns a pointer to the newly allocated set structure or NULL on error.
850 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
854 struct ptlrpc_request_set *set;
856 set = ptlrpc_prep_set();
860 set->set_max_inflight = max;
861 set->set_producer = func;
862 set->set_producer_arg = arg;
866 EXPORT_SYMBOL(ptlrpc_prep_fcset);
869 * Wind down and free request set structure previously allocated with
871 * Ensures that all requests on the set have completed and removes
872 * all requests from the request list in a set.
873 * If any unsent request happen to be on the list, pretends that they got
874 * an error in flight and calls their completion handler.
876 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
884 /* Requests on the set should either all be completed, or all be new */
885 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
886 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
887 cfs_list_for_each (tmp, &set->set_requests) {
888 struct ptlrpc_request *req =
889 cfs_list_entry(tmp, struct ptlrpc_request,
892 LASSERT(req->rq_phase == expected_phase);
896 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
897 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
898 cfs_atomic_read(&set->set_remaining), n);
900 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
901 struct ptlrpc_request *req =
902 cfs_list_entry(tmp, struct ptlrpc_request,
904 cfs_list_del_init(&req->rq_set_chain);
906 LASSERT(req->rq_phase == expected_phase);
908 if (req->rq_phase == RQ_PHASE_NEW) {
909 ptlrpc_req_interpret(NULL, req, -EBADR);
910 cfs_atomic_dec(&set->set_remaining);
913 spin_lock(&req->rq_lock);
915 req->rq_invalid_rqset = 0;
916 spin_unlock(&req->rq_lock);
918 ptlrpc_req_finished (req);
921 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
923 ptlrpc_reqset_put(set);
926 EXPORT_SYMBOL(ptlrpc_set_destroy);
929 * Add a callback function \a fn to the set.
930 * This function would be called when all requests on this set are completed.
931 * The function will be passed \a data argument.
933 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
934 set_interpreter_func fn, void *data)
936 struct ptlrpc_set_cbdata *cbdata;
938 OBD_ALLOC_PTR(cbdata);
942 cbdata->psc_interpret = fn;
943 cbdata->psc_data = data;
944 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
948 EXPORT_SYMBOL(ptlrpc_set_add_cb);
951 * Add a new request to the general purpose request set.
952 * Assumes request reference from the caller.
954 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
955 struct ptlrpc_request *req)
957 LASSERT(cfs_list_empty(&req->rq_set_chain));
959 /* The set takes over the caller's request reference */
960 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
962 cfs_atomic_inc(&set->set_remaining);
963 req->rq_queued_time = cfs_time_current();
965 if (req->rq_reqmsg != NULL)
966 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
968 if (set->set_producer != NULL)
969 /* If the request set has a producer callback, the RPC must be
970 * sent straight away */
971 ptlrpc_send_new_req(req);
973 EXPORT_SYMBOL(ptlrpc_set_add_req);
976 * Add a request to a request with dedicated server thread
977 * and wake the thread to make any necessary processing.
978 * Currently only used for ptlrpcd.
980 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
981 struct ptlrpc_request *req)
983 struct ptlrpc_request_set *set = pc->pc_set;
986 LASSERT(req->rq_set == NULL);
987 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
989 spin_lock(&set->set_new_req_lock);
991 * The set takes over the caller's request reference.
994 req->rq_queued_time = cfs_time_current();
995 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
996 count = cfs_atomic_inc_return(&set->set_new_count);
997 spin_unlock(&set->set_new_req_lock);
999 /* Only need to call wakeup once for the first entry. */
1001 cfs_waitq_signal(&set->set_waitq);
1003 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1004 * guarantee the async RPC can be processed ASAP, we have
1005 * no other better choice. It maybe fixed in future. */
1006 for (i = 0; i < pc->pc_npartners; i++)
1007 cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
1010 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1013 * Based on the current state of the import, determine if the request
1014 * can be sent, is an error, or should be delayed.
1016 * Returns true if this request should be delayed. If false, and
1017 * *status is set, then the request can not be sent and *status is the
1018 * error code. If false and status is 0, then request can be sent.
1020 * The imp->imp_lock must be held.
1022 static int ptlrpc_import_delay_req(struct obd_import *imp,
1023 struct ptlrpc_request *req, int *status)
1028 LASSERT (status != NULL);
1031 if (req->rq_ctx_init || req->rq_ctx_fini) {
1032 /* always allow ctx init/fini rpc go through */
1033 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1034 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1036 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1037 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1039 } else if (ptlrpc_send_limit_expired(req)) {
1040 /* probably doesn't need to be a D_ERROR after initial testing */
1041 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1043 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1044 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1045 /* allow CONNECT even if import is invalid */ ;
1046 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1047 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1050 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1051 if (!imp->imp_deactive)
1052 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1053 *status = -ESHUTDOWN; /* bz 12940 */
1054 } else if (req->rq_import_generation != imp->imp_generation) {
1055 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1057 } else if (req->rq_send_state != imp->imp_state) {
1058 /* invalidate in progress - any requests should be drop */
1059 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1060 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1062 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1063 *status = -EWOULDBLOCK;
1073 * Decide if the eror message regarding provided request \a req
1074 * should be printed to the console or not.
1075 * Makes it's decision on request status and other properties.
1076 * Returns 1 to print error on the system console or 0 if not.
1078 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1083 LASSERT(req->rq_reqmsg != NULL);
1084 opc = lustre_msg_get_opc(req->rq_reqmsg);
1086 /* Suppress particular reconnect errors which are to be expected. No
1087 * errors are suppressed for the initial connection on an import */
1088 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1089 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1091 /* Suppress timed out reconnect requests */
1092 if (req->rq_timedout)
1095 /* Suppress unavailable/again reconnect requests */
1096 err = lustre_msg_get_status(req->rq_repmsg);
1097 if (err == -ENODEV || err == -EAGAIN)
1105 * Check request processing status.
1106 * Returns the status.
1108 static int ptlrpc_check_status(struct ptlrpc_request *req)
1113 err = lustre_msg_get_status(req->rq_repmsg);
1114 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1115 struct obd_import *imp = req->rq_import;
1116 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1117 if (ptlrpc_console_allow(req))
1118 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1119 "communicating with %s. The %s "
1120 "operation failed with %d\n",
1122 imp->imp_connection->c_peer.nid),
1123 ll_opcode2str(opc), err);
1124 RETURN(err < 0 ? err : -EINVAL);
1128 DEBUG_REQ(D_INFO, req, "status is %d", err);
1129 } else if (err > 0) {
1130 /* XXX: translate this error from net to host */
1131 DEBUG_REQ(D_INFO, req, "status is %d", err);
1138 * save pre-versions of objects into request for replay.
1139 * Versions are obtained from server reply.
1142 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1144 struct lustre_msg *repmsg = req->rq_repmsg;
1145 struct lustre_msg *reqmsg = req->rq_reqmsg;
1146 __u64 *versions = lustre_msg_get_versions(repmsg);
1149 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1153 lustre_msg_set_versions(reqmsg, versions);
1154 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1155 versions[0], versions[1]);
1161 * Callback function called when client receives RPC reply for \a req.
1162 * Returns 0 on success or error code.
1163 * The return alue would be assigned to req->rq_status by the caller
1164 * as request processing status.
1165 * This function also decides if the request needs to be saved for later replay.
1167 static int after_reply(struct ptlrpc_request *req)
1169 struct obd_import *imp = req->rq_import;
1170 struct obd_device *obd = req->rq_import->imp_obd;
1172 struct timeval work_start;
1176 LASSERT(obd != NULL);
1177 /* repbuf must be unlinked */
1178 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1180 if (req->rq_reply_truncate) {
1181 if (ptlrpc_no_resend(req)) {
1182 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1183 " expected: %d, actual size: %d",
1184 req->rq_nob_received, req->rq_repbuf_len);
1188 sptlrpc_cli_free_repbuf(req);
1189 /* Pass the required reply buffer size (include
1190 * space for early reply).
1191 * NB: no need to roundup because alloc_repbuf
1192 * will roundup it */
1193 req->rq_replen = req->rq_nob_received;
1194 req->rq_nob_received = 0;
1200 * NB Until this point, the whole of the incoming message,
1201 * including buflens, status etc is in the sender's byte order.
1203 rc = sptlrpc_cli_unwrap_reply(req);
1205 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1209 /* retry indefinitely on EINPROGRESS */
1210 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1211 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1212 time_t now = cfs_time_current_sec();
1214 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1216 req->rq_nr_resend++;
1218 /* allocate new xid to avoid reply reconstruction */
1219 if (!req->rq_bulk) {
1220 /* new xid is already allocated for bulk in
1221 * ptlrpc_check_set() */
1222 req->rq_xid = ptlrpc_next_xid();
1223 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1224 "resend on EINPROGRESS");
1227 /* Readjust the timeout for current conditions */
1228 ptlrpc_at_set_req_timeout(req);
1229 /* delay resend to give a chance to the server to get ready.
1230 * The delay is increased by 1s on every resend and is capped to
1231 * the current request timeout (i.e. obd_timeout if AT is off,
1232 * or AT service time x 125% + 5s, see at_est2timeout) */
1233 if (req->rq_nr_resend > req->rq_timeout)
1234 req->rq_sent = now + req->rq_timeout;
1236 req->rq_sent = now + req->rq_nr_resend;
1240 * Security layer unwrap might ask resend this request.
1245 rc = unpack_reply(req);
1249 cfs_gettimeofday(&work_start);
1250 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1251 if (obd->obd_svc_stats != NULL) {
1252 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1254 ptlrpc_lprocfs_rpc_sent(req, timediff);
1257 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1258 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1259 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1260 lustre_msg_get_type(req->rq_repmsg));
1264 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1265 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1266 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1267 ptlrpc_at_adj_net_latency(req,
1268 lustre_msg_get_service_time(req->rq_repmsg));
1270 rc = ptlrpc_check_status(req);
1271 imp->imp_connect_error = rc;
1275 * Either we've been evicted, or the server has failed for
1276 * some reason. Try to reconnect, and if that fails, punt to
1279 if (ll_rpc_recoverable_error(rc)) {
1280 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1281 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1284 ptlrpc_request_handle_notconn(req);
1289 * Let's look if server sent slv. Do it only for RPC with
1292 ldlm_cli_update_pool(req);
1296 * Store transno in reqmsg for replay.
1298 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1299 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1300 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1303 if (imp->imp_replayable) {
1304 spin_lock(&imp->imp_lock);
1306 * No point in adding already-committed requests to the replay
1307 * list, we will just remove them immediately. b=9829
1309 if (req->rq_transno != 0 &&
1311 lustre_msg_get_last_committed(req->rq_repmsg) ||
1313 /** version recovery */
1314 ptlrpc_save_versions(req);
1315 ptlrpc_retain_replayable_request(req, imp);
1316 } else if (req->rq_commit_cb != NULL) {
1317 spin_unlock(&imp->imp_lock);
1318 req->rq_commit_cb(req);
1319 spin_lock(&imp->imp_lock);
1323 * Replay-enabled imports return commit-status information.
1325 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1326 imp->imp_peer_committed_transno =
1327 lustre_msg_get_last_committed(req->rq_repmsg);
1329 ptlrpc_free_committed(imp);
1331 if (req->rq_transno > imp->imp_peer_committed_transno)
1332 ptlrpc_pinger_commit_expected(imp);
1334 spin_unlock(&imp->imp_lock);
1341 * Helper function to send request \a req over the network for the first time
1342 * Also adjusts request phase.
1343 * Returns 0 on success or error code.
1345 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1347 struct obd_import *imp = req->rq_import;
1351 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1352 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1353 (!req->rq_generation_set ||
1354 req->rq_import_generation == imp->imp_generation))
1357 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1359 spin_lock(&imp->imp_lock);
1361 if (!req->rq_generation_set)
1362 req->rq_import_generation = imp->imp_generation;
1364 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1365 spin_lock(&req->rq_lock);
1366 req->rq_waiting = 1;
1367 spin_unlock(&req->rq_lock);
1369 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1370 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1371 ptlrpc_import_state_name(req->rq_send_state),
1372 ptlrpc_import_state_name(imp->imp_state));
1373 LASSERT(cfs_list_empty(&req->rq_list));
1374 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1375 cfs_atomic_inc(&req->rq_import->imp_inflight);
1376 spin_unlock(&imp->imp_lock);
1381 spin_unlock(&imp->imp_lock);
1382 req->rq_status = rc;
1383 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1387 LASSERT(cfs_list_empty(&req->rq_list));
1388 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1389 cfs_atomic_inc(&req->rq_import->imp_inflight);
1390 spin_unlock(&imp->imp_lock);
1392 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1394 rc = sptlrpc_req_refresh_ctx(req, -1);
1397 req->rq_status = rc;
1400 req->rq_wait_ctx = 1;
1405 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1406 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1407 imp->imp_obd->obd_uuid.uuid,
1408 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1409 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1410 lustre_msg_get_opc(req->rq_reqmsg));
1412 rc = ptl_send_rpc(req, 0);
1414 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1415 req->rq_net_err = 1;
1421 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1426 LASSERT(set->set_producer != NULL);
1428 remaining = cfs_atomic_read(&set->set_remaining);
1430 /* populate the ->set_requests list with requests until we
1431 * reach the maximum number of RPCs in flight for this set */
1432 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1433 rc = set->set_producer(set, set->set_producer_arg);
1434 if (rc == -ENOENT) {
1435 /* no more RPC to produce */
1436 set->set_producer = NULL;
1437 set->set_producer_arg = NULL;
1442 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1446 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1447 * and no more replies are expected.
1448 * (it is possible to get less replies than requests sent e.g. due to timed out
1449 * requests or requests that we had trouble to send out)
1451 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1453 cfs_list_t *tmp, *next;
1454 int force_timer_recalc = 0;
1457 if (cfs_atomic_read(&set->set_remaining) == 0)
1460 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1461 struct ptlrpc_request *req =
1462 cfs_list_entry(tmp, struct ptlrpc_request,
1464 struct obd_import *imp = req->rq_import;
1465 int unregistered = 0;
1468 if (req->rq_phase == RQ_PHASE_NEW &&
1469 ptlrpc_send_new_req(req)) {
1470 force_timer_recalc = 1;
1473 /* delayed send - skip */
1474 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1477 /* delayed resend - skip */
1478 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1479 req->rq_sent > cfs_time_current_sec())
1482 if (!(req->rq_phase == RQ_PHASE_RPC ||
1483 req->rq_phase == RQ_PHASE_BULK ||
1484 req->rq_phase == RQ_PHASE_INTERPRET ||
1485 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1486 req->rq_phase == RQ_PHASE_COMPLETE)) {
1487 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1491 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1492 LASSERT(req->rq_next_phase != req->rq_phase);
1493 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1496 * Skip processing until reply is unlinked. We
1497 * can't return to pool before that and we can't
1498 * call interpret before that. We need to make
1499 * sure that all rdma transfers finished and will
1500 * not corrupt any data.
1502 if (ptlrpc_client_recv_or_unlink(req) ||
1503 ptlrpc_client_bulk_active(req))
1507 * Turn fail_loc off to prevent it from looping
1510 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1511 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1514 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1515 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1520 * Move to next phase if reply was successfully
1523 ptlrpc_rqphase_move(req, req->rq_next_phase);
1526 if (req->rq_phase == RQ_PHASE_COMPLETE)
1529 if (req->rq_phase == RQ_PHASE_INTERPRET)
1530 GOTO(interpret, req->rq_status);
1533 * Note that this also will start async reply unlink.
1535 if (req->rq_net_err && !req->rq_timedout) {
1536 ptlrpc_expire_one_request(req, 1);
1539 * Check if we still need to wait for unlink.
1541 if (ptlrpc_client_recv_or_unlink(req) ||
1542 ptlrpc_client_bulk_active(req))
1544 /* If there is no need to resend, fail it now. */
1545 if (req->rq_no_resend) {
1546 if (req->rq_status == 0)
1547 req->rq_status = -EIO;
1548 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1549 GOTO(interpret, req->rq_status);
1556 spin_lock(&req->rq_lock);
1557 req->rq_replied = 0;
1558 spin_unlock(&req->rq_lock);
1559 if (req->rq_status == 0)
1560 req->rq_status = -EIO;
1561 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1562 GOTO(interpret, req->rq_status);
1565 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1566 * so it sets rq_intr regardless of individual rpc
1567 * timeouts. The synchronous IO waiting path sets
1568 * rq_intr irrespective of whether ptlrpcd
1569 * has seen a timeout. Our policy is to only interpret
1570 * interrupted rpcs after they have timed out, so we
1571 * need to enforce that here.
1574 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1575 req->rq_wait_ctx)) {
1576 req->rq_status = -EINTR;
1577 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1578 GOTO(interpret, req->rq_status);
1581 if (req->rq_phase == RQ_PHASE_RPC) {
1582 if (req->rq_timedout || req->rq_resend ||
1583 req->rq_waiting || req->rq_wait_ctx) {
1586 if (!ptlrpc_unregister_reply(req, 1))
1589 spin_lock(&imp->imp_lock);
1590 if (ptlrpc_import_delay_req(imp, req, &status)){
1591 /* put on delay list - only if we wait
1592 * recovery finished - before send */
1593 cfs_list_del_init(&req->rq_list);
1594 cfs_list_add_tail(&req->rq_list,
1597 spin_unlock(&imp->imp_lock);
1602 req->rq_status = status;
1603 ptlrpc_rqphase_move(req,
1604 RQ_PHASE_INTERPRET);
1605 spin_unlock(&imp->imp_lock);
1606 GOTO(interpret, req->rq_status);
1608 if (ptlrpc_no_resend(req) &&
1609 !req->rq_wait_ctx) {
1610 req->rq_status = -ENOTCONN;
1611 ptlrpc_rqphase_move(req,
1612 RQ_PHASE_INTERPRET);
1613 spin_unlock(&imp->imp_lock);
1614 GOTO(interpret, req->rq_status);
1617 cfs_list_del_init(&req->rq_list);
1618 cfs_list_add_tail(&req->rq_list,
1619 &imp->imp_sending_list);
1621 spin_unlock(&imp->imp_lock);
1623 spin_lock(&req->rq_lock);
1624 req->rq_waiting = 0;
1625 spin_unlock(&req->rq_lock);
1627 if (req->rq_timedout || req->rq_resend) {
1628 /* This is re-sending anyways,
1629 * let's mark req as resend. */
1630 spin_lock(&req->rq_lock);
1632 spin_unlock(&req->rq_lock);
1636 if (!ptlrpc_unregister_bulk(req, 1))
1639 /* ensure previous bulk fails */
1640 old_xid = req->rq_xid;
1641 req->rq_xid = ptlrpc_next_xid();
1642 CDEBUG(D_HA, "resend bulk "
1645 old_xid, req->rq_xid);
1649 * rq_wait_ctx is only touched by ptlrpcd,
1650 * so no lock is needed here.
1652 status = sptlrpc_req_refresh_ctx(req, -1);
1655 req->rq_status = status;
1656 spin_lock(&req->rq_lock);
1657 req->rq_wait_ctx = 0;
1658 spin_unlock(&req->rq_lock);
1659 force_timer_recalc = 1;
1661 spin_lock(&req->rq_lock);
1662 req->rq_wait_ctx = 1;
1663 spin_unlock(&req->rq_lock);
1668 spin_lock(&req->rq_lock);
1669 req->rq_wait_ctx = 0;
1670 spin_unlock(&req->rq_lock);
1673 rc = ptl_send_rpc(req, 0);
1675 DEBUG_REQ(D_HA, req,
1676 "send failed: rc = %d", rc);
1677 force_timer_recalc = 1;
1678 spin_lock(&req->rq_lock);
1679 req->rq_net_err = 1;
1680 spin_unlock(&req->rq_lock);
1682 /* need to reset the timeout */
1683 force_timer_recalc = 1;
1686 spin_lock(&req->rq_lock);
1688 if (ptlrpc_client_early(req)) {
1689 ptlrpc_at_recv_early_reply(req);
1690 spin_unlock(&req->rq_lock);
1694 /* Still waiting for a reply? */
1695 if (ptlrpc_client_recv(req)) {
1696 spin_unlock(&req->rq_lock);
1700 /* Did we actually receive a reply? */
1701 if (!ptlrpc_client_replied(req)) {
1702 spin_unlock(&req->rq_lock);
1706 spin_unlock(&req->rq_lock);
1708 /* unlink from net because we are going to
1709 * swab in-place of reply buffer */
1710 unregistered = ptlrpc_unregister_reply(req, 1);
1714 req->rq_status = after_reply(req);
1718 /* If there is no bulk associated with this request,
1719 * then we're done and should let the interpreter
1720 * process the reply. Similarly if the RPC returned
1721 * an error, and therefore the bulk will never arrive.
1723 if (req->rq_bulk == NULL || req->rq_status < 0) {
1724 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1725 GOTO(interpret, req->rq_status);
1728 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1731 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1732 if (ptlrpc_client_bulk_active(req))
1735 if (!req->rq_bulk->bd_success) {
1736 /* The RPC reply arrived OK, but the bulk screwed
1737 * up! Dead weird since the server told us the RPC
1738 * was good after getting the REPLY for her GET or
1739 * the ACK for her PUT. */
1740 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1741 req->rq_status = -EIO;
1744 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1747 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1749 /* This moves to "unregistering" phase we need to wait for
1751 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1752 /* start async bulk unlink too */
1753 ptlrpc_unregister_bulk(req, 1);
1757 if (!ptlrpc_unregister_bulk(req, 1))
1760 /* When calling interpret receiving already should be
1762 LASSERT(!req->rq_receiving_reply);
1764 ptlrpc_req_interpret(env, req, req->rq_status);
1766 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1768 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1769 "Completed RPC pname:cluuid:pid:xid:nid:"
1770 "opc %s:%s:%d:"LPU64":%s:%d\n",
1771 cfs_curproc_comm(), imp->imp_obd->obd_uuid.uuid,
1772 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1773 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1774 lustre_msg_get_opc(req->rq_reqmsg));
1776 spin_lock(&imp->imp_lock);
1777 /* Request already may be not on sending or delaying list. This
1778 * may happen in the case of marking it erroneous for the case
1779 * ptlrpc_import_delay_req(req, status) find it impossible to
1780 * allow sending this rpc and returns *status != 0. */
1781 if (!cfs_list_empty(&req->rq_list)) {
1782 cfs_list_del_init(&req->rq_list);
1783 cfs_atomic_dec(&imp->imp_inflight);
1785 spin_unlock(&imp->imp_lock);
1787 cfs_atomic_dec(&set->set_remaining);
1788 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1790 if (set->set_producer) {
1791 /* produce a new request if possible */
1792 if (ptlrpc_set_producer(set) > 0)
1793 force_timer_recalc = 1;
1795 /* free the request that has just been completed
1796 * in order not to pollute set->set_requests */
1797 cfs_list_del_init(&req->rq_set_chain);
1798 spin_lock(&req->rq_lock);
1800 req->rq_invalid_rqset = 0;
1801 spin_unlock(&req->rq_lock);
1803 /* record rq_status to compute the final status later */
1804 if (req->rq_status != 0)
1805 set->set_rc = req->rq_status;
1806 ptlrpc_req_finished(req);
1810 /* If we hit an error, we want to recover promptly. */
1811 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1813 EXPORT_SYMBOL(ptlrpc_check_set);
1816 * Time out request \a req. is \a async_unlink is set, that means do not wait
1817 * until LNet actually confirms network buffer unlinking.
1818 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1820 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1822 struct obd_import *imp = req->rq_import;
1826 spin_lock(&req->rq_lock);
1827 req->rq_timedout = 1;
1828 spin_unlock(&req->rq_lock);
1830 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1831 "/real "CFS_DURATION_T"]",
1832 req->rq_net_err ? "failed due to network error" :
1833 ((req->rq_real_sent == 0 ||
1834 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1835 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1836 "timed out for sent delay" : "timed out for slow reply"),
1837 req->rq_sent, req->rq_real_sent);
1839 if (imp != NULL && obd_debug_peer_on_timeout)
1840 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1842 ptlrpc_unregister_reply(req, async_unlink);
1843 ptlrpc_unregister_bulk(req, async_unlink);
1845 if (obd_dump_on_timeout)
1846 libcfs_debug_dumplog();
1849 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1853 cfs_atomic_inc(&imp->imp_timeouts);
1855 /* The DLM server doesn't want recovery run on its imports. */
1856 if (imp->imp_dlm_fake)
1859 /* If this request is for recovery or other primordial tasks,
1860 * then error it out here. */
1861 if (req->rq_ctx_init || req->rq_ctx_fini ||
1862 req->rq_send_state != LUSTRE_IMP_FULL ||
1863 imp->imp_obd->obd_no_recov) {
1864 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1865 ptlrpc_import_state_name(req->rq_send_state),
1866 ptlrpc_import_state_name(imp->imp_state));
1867 spin_lock(&req->rq_lock);
1868 req->rq_status = -ETIMEDOUT;
1870 spin_unlock(&req->rq_lock);
1874 /* if a request can't be resent we can't wait for an answer after
1876 if (ptlrpc_no_resend(req)) {
1877 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1881 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1887 * Time out all uncompleted requests in request set pointed by \a data
1888 * Callback used when waiting on sets with l_wait_event.
1891 int ptlrpc_expired_set(void *data)
1893 struct ptlrpc_request_set *set = data;
1895 time_t now = cfs_time_current_sec();
1898 LASSERT(set != NULL);
1901 * A timeout expired. See which reqs it applies to...
1903 cfs_list_for_each (tmp, &set->set_requests) {
1904 struct ptlrpc_request *req =
1905 cfs_list_entry(tmp, struct ptlrpc_request,
1908 /* don't expire request waiting for context */
1909 if (req->rq_wait_ctx)
1912 /* Request in-flight? */
1913 if (!((req->rq_phase == RQ_PHASE_RPC &&
1914 !req->rq_waiting && !req->rq_resend) ||
1915 (req->rq_phase == RQ_PHASE_BULK)))
1918 if (req->rq_timedout || /* already dealt with */
1919 req->rq_deadline > now) /* not expired */
1922 /* Deal with this guy. Do it asynchronously to not block
1923 * ptlrpcd thread. */
1924 ptlrpc_expire_one_request(req, 1);
1928 * When waiting for a whole set, we always break out of the
1929 * sleep so we can recalculate the timeout, or enable interrupts
1930 * if everyone's timed out.
1934 EXPORT_SYMBOL(ptlrpc_expired_set);
1937 * Sets rq_intr flag in \a req under spinlock.
1939 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1941 spin_lock(&req->rq_lock);
1943 spin_unlock(&req->rq_lock);
1945 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
1948 * Interrupts (sets interrupted flag) all uncompleted requests in
1949 * a set \a data. Callback for l_wait_event for interruptible waits.
1951 void ptlrpc_interrupted_set(void *data)
1953 struct ptlrpc_request_set *set = data;
1956 LASSERT(set != NULL);
1957 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
1959 cfs_list_for_each(tmp, &set->set_requests) {
1960 struct ptlrpc_request *req =
1961 cfs_list_entry(tmp, struct ptlrpc_request,
1964 if (req->rq_phase != RQ_PHASE_RPC &&
1965 req->rq_phase != RQ_PHASE_UNREGISTERING)
1968 ptlrpc_mark_interrupted(req);
1971 EXPORT_SYMBOL(ptlrpc_interrupted_set);
1974 * Get the smallest timeout in the set; this does NOT set a timeout.
1976 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1979 time_t now = cfs_time_current_sec();
1981 struct ptlrpc_request *req;
1985 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1987 cfs_list_for_each(tmp, &set->set_requests) {
1988 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1991 * Request in-flight?
1993 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1994 (req->rq_phase == RQ_PHASE_BULK) ||
1995 (req->rq_phase == RQ_PHASE_NEW)))
1999 * Already timed out.
2001 if (req->rq_timedout)
2007 if (req->rq_wait_ctx)
2010 if (req->rq_phase == RQ_PHASE_NEW)
2011 deadline = req->rq_sent;
2012 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2013 deadline = req->rq_sent;
2015 deadline = req->rq_sent + req->rq_timeout;
2017 if (deadline <= now) /* actually expired already */
2018 timeout = 1; /* ASAP */
2019 else if (timeout == 0 || timeout > deadline - now)
2020 timeout = deadline - now;
2024 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2027 * Send all unset request from the set and then wait untill all
2028 * requests in the set complete (either get a reply, timeout, get an
2029 * error or otherwise be interrupted).
2030 * Returns 0 on success or error code otherwise.
2032 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2035 struct ptlrpc_request *req;
2036 struct l_wait_info lwi;
2040 if (set->set_producer)
2041 (void)ptlrpc_set_producer(set);
2043 cfs_list_for_each(tmp, &set->set_requests) {
2044 req = cfs_list_entry(tmp, struct ptlrpc_request,
2046 if (req->rq_phase == RQ_PHASE_NEW)
2047 (void)ptlrpc_send_new_req(req);
2050 if (cfs_list_empty(&set->set_requests))
2054 timeout = ptlrpc_set_next_timeout(set);
2056 /* wait until all complete, interrupted, or an in-flight
2058 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2061 if (timeout == 0 && !cfs_signal_pending())
2063 * No requests are in-flight (ether timed out
2064 * or delayed), so we can allow interrupts.
2065 * We still want to block for a limited time,
2066 * so we allow interrupts during the timeout.
2068 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2070 ptlrpc_interrupted_set, set);
2073 * At least one request is in flight, so no
2074 * interrupts are allowed. Wait until all
2075 * complete, or an in-flight req times out.
2077 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2078 ptlrpc_expired_set, set);
2080 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2082 /* LU-769 - if we ignored the signal because it was already
2083 * pending when we started, we need to handle it now or we risk
2084 * it being ignored forever */
2085 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2086 cfs_signal_pending()) {
2087 cfs_sigset_t blocked_sigs =
2088 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2090 /* In fact we only interrupt for the "fatal" signals
2091 * like SIGINT or SIGKILL. We still ignore less
2092 * important signals since ptlrpc set is not easily
2093 * reentrant from userspace again */
2094 if (cfs_signal_pending())
2095 ptlrpc_interrupted_set(set);
2096 cfs_restore_sigs(blocked_sigs);
2099 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2101 /* -EINTR => all requests have been flagged rq_intr so next
2103 * -ETIMEDOUT => someone timed out. When all reqs have
2104 * timed out, signals are enabled allowing completion with
2106 * I don't really care if we go once more round the loop in
2107 * the error cases -eeb. */
2108 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2109 cfs_list_for_each(tmp, &set->set_requests) {
2110 req = cfs_list_entry(tmp, struct ptlrpc_request,
2112 spin_lock(&req->rq_lock);
2113 req->rq_invalid_rqset = 1;
2114 spin_unlock(&req->rq_lock);
2117 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2119 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2121 rc = set->set_rc; /* rq_status of already freed requests if any */
2122 cfs_list_for_each(tmp, &set->set_requests) {
2123 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2125 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2126 if (req->rq_status != 0)
2127 rc = req->rq_status;
2130 if (set->set_interpret != NULL) {
2131 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2133 rc = interpreter (set, set->set_arg, rc);
2135 struct ptlrpc_set_cbdata *cbdata, *n;
2138 cfs_list_for_each_entry_safe(cbdata, n,
2139 &set->set_cblist, psc_item) {
2140 cfs_list_del_init(&cbdata->psc_item);
2141 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2144 OBD_FREE_PTR(cbdata);
2150 EXPORT_SYMBOL(ptlrpc_set_wait);
2153 * Helper fuction for request freeing.
2154 * Called when request count reached zero and request needs to be freed.
2155 * Removes request from all sorts of sending/replay lists it might be on,
2156 * frees network buffers if any are present.
2157 * If \a locked is set, that means caller is already holding import imp_lock
2158 * and so we no longer need to reobtain it (for certain lists manipulations)
2160 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2163 if (request == NULL) {
2168 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2169 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2170 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2171 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2172 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2173 LASSERTF(!request->rq_replay, "req %p\n", request);
2175 req_capsule_fini(&request->rq_pill);
2177 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2178 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2179 if (request->rq_import != NULL) {
2181 spin_lock(&request->rq_import->imp_lock);
2182 cfs_list_del_init(&request->rq_replay_list);
2184 spin_unlock(&request->rq_import->imp_lock);
2186 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2188 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2189 DEBUG_REQ(D_ERROR, request,
2190 "freeing request with nonzero refcount");
2194 if (request->rq_repbuf != NULL)
2195 sptlrpc_cli_free_repbuf(request);
2196 if (request->rq_export != NULL) {
2197 class_export_put(request->rq_export);
2198 request->rq_export = NULL;
2200 if (request->rq_import != NULL) {
2201 class_import_put(request->rq_import);
2202 request->rq_import = NULL;
2204 if (request->rq_bulk != NULL)
2205 ptlrpc_free_bulk_pin(request->rq_bulk);
2207 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2208 sptlrpc_cli_free_reqbuf(request);
2210 if (request->rq_cli_ctx)
2211 sptlrpc_req_put_ctx(request, !locked);
2213 if (request->rq_pool)
2214 __ptlrpc_free_req_to_pool(request);
2216 OBD_FREE(request, sizeof(*request));
2220 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2222 * Drop one request reference. Must be called with import imp_lock held.
2223 * When reference count drops to zero, reuqest is freed.
2225 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2227 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2228 (void)__ptlrpc_req_finished(request, 1);
2230 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2234 * Drops one reference count for request \a request.
2235 * \a locked set indicates that caller holds import imp_lock.
2236 * Frees the request whe reference count reaches zero.
2238 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2241 if (request == NULL)
2244 if (request == LP_POISON ||
2245 request->rq_reqmsg == LP_POISON) {
2246 CERROR("dereferencing freed request (bug 575)\n");
2251 DEBUG_REQ(D_INFO, request, "refcount now %u",
2252 cfs_atomic_read(&request->rq_refcount) - 1);
2254 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2255 __ptlrpc_free_req(request, locked);
2263 * Drops one reference count for a request.
2265 void ptlrpc_req_finished(struct ptlrpc_request *request)
2267 __ptlrpc_req_finished(request, 0);
2269 EXPORT_SYMBOL(ptlrpc_req_finished);
2272 * Returns xid of a \a request
2274 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2276 return request->rq_xid;
2278 EXPORT_SYMBOL(ptlrpc_req_xid);
2281 * Disengage the client's reply buffer from the network
2282 * NB does _NOT_ unregister any client-side bulk.
2283 * IDEMPOTENT, but _not_ safe against concurrent callers.
2284 * The request owner (i.e. the thread doing the I/O) must call...
2285 * Returns 0 on success or 1 if unregistering cannot be made.
2287 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2291 struct l_wait_info lwi;
2296 LASSERT(!cfs_in_interrupt());
2299 * Let's setup deadline for reply unlink.
2301 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2302 async && request->rq_reply_deadline == 0)
2303 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2306 * Nothing left to do.
2308 if (!ptlrpc_client_recv_or_unlink(request))
2311 LNetMDUnlink(request->rq_reply_md_h);
2314 * Let's check it once again.
2316 if (!ptlrpc_client_recv_or_unlink(request))
2320 * Move to "Unregistering" phase as reply was not unlinked yet.
2322 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2325 * Do not wait for unlink to finish.
2331 * We have to l_wait_event() whatever the result, to give liblustre
2332 * a chance to run reply_in_callback(), and to make sure we've
2333 * unlinked before returning a req to the pool.
2335 if (request->rq_set != NULL)
2336 wq = &request->rq_set->set_waitq;
2338 wq = &request->rq_reply_waitq;
2341 /* Network access will complete in finite time but the HUGE
2342 * timeout lets us CWARN for visibility of sluggish NALs */
2343 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2344 cfs_time_seconds(1), NULL, NULL);
2345 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2348 ptlrpc_rqphase_move(request, request->rq_next_phase);
2352 LASSERT(rc == -ETIMEDOUT);
2353 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2354 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2355 request->rq_must_unlink);
2359 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2362 * Iterates through replay_list on import and prunes
2363 * all requests have transno smaller than last_committed for the
2364 * import and don't have rq_replay set.
2365 * Since requests are sorted in transno order, stops when meetign first
2366 * transno bigger than last_committed.
2367 * caller must hold imp->imp_lock
2369 void ptlrpc_free_committed(struct obd_import *imp)
2371 cfs_list_t *tmp, *saved;
2372 struct ptlrpc_request *req;
2373 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2376 LASSERT(imp != NULL);
2378 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2381 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2382 imp->imp_generation == imp->imp_last_generation_checked) {
2383 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2384 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2388 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2389 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2390 imp->imp_generation);
2391 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2392 imp->imp_last_generation_checked = imp->imp_generation;
2394 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2395 req = cfs_list_entry(tmp, struct ptlrpc_request,
2398 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2399 LASSERT(req != last_req);
2402 if (req->rq_transno == 0) {
2403 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2406 if (req->rq_import_generation < imp->imp_generation) {
2407 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2411 if (req->rq_replay) {
2412 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2416 /* not yet committed */
2417 if (req->rq_transno > imp->imp_peer_committed_transno) {
2418 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2422 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2423 imp->imp_peer_committed_transno);
2425 spin_lock(&req->rq_lock);
2427 spin_unlock(&req->rq_lock);
2428 if (req->rq_commit_cb != NULL)
2429 req->rq_commit_cb(req);
2430 cfs_list_del_init(&req->rq_replay_list);
2431 __ptlrpc_req_finished(req, 1);
2438 void ptlrpc_cleanup_client(struct obd_import *imp)
2444 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2447 * Schedule previously sent request for resend.
2448 * For bulk requests we assign new xid (to avoid problems with
2449 * lost replies and therefore several transfers landing into same buffer
2450 * from different sending attempts).
2452 void ptlrpc_resend_req(struct ptlrpc_request *req)
2454 DEBUG_REQ(D_HA, req, "going to resend");
2455 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2456 req->rq_status = -EAGAIN;
2458 spin_lock(&req->rq_lock);
2460 req->rq_net_err = 0;
2461 req->rq_timedout = 0;
2463 __u64 old_xid = req->rq_xid;
2465 /* ensure previous bulk fails */
2466 req->rq_xid = ptlrpc_next_xid();
2467 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2468 old_xid, req->rq_xid);
2470 ptlrpc_client_wake_req(req);
2471 spin_unlock(&req->rq_lock);
2473 EXPORT_SYMBOL(ptlrpc_resend_req);
2475 /* XXX: this function and rq_status are currently unused */
2476 void ptlrpc_restart_req(struct ptlrpc_request *req)
2478 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2479 req->rq_status = -ERESTARTSYS;
2481 spin_lock(&req->rq_lock);
2482 req->rq_restart = 1;
2483 req->rq_timedout = 0;
2484 ptlrpc_client_wake_req(req);
2485 spin_unlock(&req->rq_lock);
2487 EXPORT_SYMBOL(ptlrpc_restart_req);
2490 * Grab additional reference on a request \a req
2492 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2495 cfs_atomic_inc(&req->rq_refcount);
2498 EXPORT_SYMBOL(ptlrpc_request_addref);
2501 * Add a request to import replay_list.
2502 * Must be called under imp_lock
2504 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2505 struct obd_import *imp)
2509 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2511 if (req->rq_transno == 0) {
2512 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2516 /* clear this for new requests that were resent as well
2517 as resent replayed requests. */
2518 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2520 /* don't re-add requests that have been replayed */
2521 if (!cfs_list_empty(&req->rq_replay_list))
2524 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2526 LASSERT(imp->imp_replayable);
2527 /* Balanced in ptlrpc_free_committed, usually. */
2528 ptlrpc_request_addref(req);
2529 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2530 struct ptlrpc_request *iter =
2531 cfs_list_entry(tmp, struct ptlrpc_request,
2534 /* We may have duplicate transnos if we create and then
2535 * open a file, or for closes retained if to match creating
2536 * opens, so use req->rq_xid as a secondary key.
2537 * (See bugs 684, 685, and 428.)
2538 * XXX no longer needed, but all opens need transnos!
2540 if (iter->rq_transno > req->rq_transno)
2543 if (iter->rq_transno == req->rq_transno) {
2544 LASSERT(iter->rq_xid != req->rq_xid);
2545 if (iter->rq_xid > req->rq_xid)
2549 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2553 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2555 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2558 * Send request and wait until it completes.
2559 * Returns request processing status.
2561 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2563 struct ptlrpc_request_set *set;
2567 LASSERT(req->rq_set == NULL);
2568 LASSERT(!req->rq_receiving_reply);
2570 set = ptlrpc_prep_set();
2572 CERROR("Unable to allocate ptlrpc set.");
2576 /* for distributed debugging */
2577 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2579 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2580 ptlrpc_request_addref(req);
2581 ptlrpc_set_add_req(set, req);
2582 rc = ptlrpc_set_wait(set);
2583 ptlrpc_set_destroy(set);
2587 EXPORT_SYMBOL(ptlrpc_queue_wait);
2589 struct ptlrpc_replay_async_args {
2591 int praa_old_status;
2595 * Callback used for replayed requests reply processing.
2596 * In case of succesful reply calls registeresd request replay callback.
2597 * In case of error restart replay process.
2599 static int ptlrpc_replay_interpret(const struct lu_env *env,
2600 struct ptlrpc_request *req,
2601 void * data, int rc)
2603 struct ptlrpc_replay_async_args *aa = data;
2604 struct obd_import *imp = req->rq_import;
2607 cfs_atomic_dec(&imp->imp_replay_inflight);
2609 if (!ptlrpc_client_replied(req)) {
2610 CERROR("request replay timed out, restarting recovery\n");
2611 GOTO(out, rc = -ETIMEDOUT);
2614 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2615 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2616 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2617 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2619 /** VBR: check version failure */
2620 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2621 /** replay was failed due to version mismatch */
2622 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2623 spin_lock(&imp->imp_lock);
2624 imp->imp_vbr_failed = 1;
2625 imp->imp_no_lock_replay = 1;
2626 spin_unlock(&imp->imp_lock);
2627 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2629 /** The transno had better not change over replay. */
2630 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2631 lustre_msg_get_transno(req->rq_repmsg) ||
2632 lustre_msg_get_transno(req->rq_repmsg) == 0,
2634 lustre_msg_get_transno(req->rq_reqmsg),
2635 lustre_msg_get_transno(req->rq_repmsg));
2638 spin_lock(&imp->imp_lock);
2639 /** if replays by version then gap occur on server, no trust to locks */
2640 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2641 imp->imp_no_lock_replay = 1;
2642 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2643 spin_unlock(&imp->imp_lock);
2644 LASSERT(imp->imp_last_replay_transno);
2646 /* transaction number shouldn't be bigger than the latest replayed */
2647 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2648 DEBUG_REQ(D_ERROR, req,
2649 "Reported transno "LPU64" is bigger than the "
2650 "replayed one: "LPU64, req->rq_transno,
2651 lustre_msg_get_transno(req->rq_reqmsg));
2652 GOTO(out, rc = -EINVAL);
2655 DEBUG_REQ(D_HA, req, "got rep");
2657 /* let the callback do fixups, possibly including in the request */
2658 if (req->rq_replay_cb)
2659 req->rq_replay_cb(req);
2661 if (ptlrpc_client_replied(req) &&
2662 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2663 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2664 lustre_msg_get_status(req->rq_repmsg),
2665 aa->praa_old_status);
2667 /* Put it back for re-replay. */
2668 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2672 * Errors while replay can set transno to 0, but
2673 * imp_last_replay_transno shouldn't be set to 0 anyway
2675 if (req->rq_transno == 0)
2676 CERROR("Transno is 0 during replay!\n");
2678 /* continue with recovery */
2679 rc = ptlrpc_import_recovery_state_machine(imp);
2681 req->rq_send_state = aa->praa_old_state;
2684 /* this replay failed, so restart recovery */
2685 ptlrpc_connect_import(imp);
2691 * Prepares and queues request for replay.
2692 * Adds it to ptlrpcd queue for actual sending.
2693 * Returns 0 on success.
2695 int ptlrpc_replay_req(struct ptlrpc_request *req)
2697 struct ptlrpc_replay_async_args *aa;
2700 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2702 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2703 aa = ptlrpc_req_async_args(req);
2704 memset(aa, 0, sizeof *aa);
2706 /* Prepare request to be resent with ptlrpcd */
2707 aa->praa_old_state = req->rq_send_state;
2708 req->rq_send_state = LUSTRE_IMP_REPLAY;
2709 req->rq_phase = RQ_PHASE_NEW;
2710 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2712 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2714 req->rq_interpret_reply = ptlrpc_replay_interpret;
2715 /* Readjust the timeout for current conditions */
2716 ptlrpc_at_set_req_timeout(req);
2718 /* Tell server the net_latency, so the server can calculate how long
2719 * it should wait for next replay */
2720 lustre_msg_set_service_time(req->rq_reqmsg,
2721 ptlrpc_at_get_net_latency(req));
2722 DEBUG_REQ(D_HA, req, "REPLAY");
2724 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2725 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2727 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2730 EXPORT_SYMBOL(ptlrpc_replay_req);
2733 * Aborts all in-flight request on import \a imp sending and delayed lists
2735 void ptlrpc_abort_inflight(struct obd_import *imp)
2737 cfs_list_t *tmp, *n;
2740 /* Make sure that no new requests get processed for this import.
2741 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2742 * this flag and then putting requests on sending_list or delayed_list.
2744 spin_lock(&imp->imp_lock);
2746 /* XXX locking? Maybe we should remove each request with the list
2747 * locked? Also, how do we know if the requests on the list are
2748 * being freed at this time?
2750 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2751 struct ptlrpc_request *req =
2752 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2754 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2756 spin_lock(&req->rq_lock);
2757 if (req->rq_import_generation < imp->imp_generation) {
2759 req->rq_status = -EIO;
2760 ptlrpc_client_wake_req(req);
2762 spin_unlock(&req->rq_lock);
2765 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2766 struct ptlrpc_request *req =
2767 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2769 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2771 spin_lock(&req->rq_lock);
2772 if (req->rq_import_generation < imp->imp_generation) {
2774 req->rq_status = -EIO;
2775 ptlrpc_client_wake_req(req);
2777 spin_unlock(&req->rq_lock);
2780 /* Last chance to free reqs left on the replay list, but we
2781 * will still leak reqs that haven't committed. */
2782 if (imp->imp_replayable)
2783 ptlrpc_free_committed(imp);
2785 spin_unlock(&imp->imp_lock);
2789 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2792 * Abort all uncompleted requests in request set \a set
2794 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2796 cfs_list_t *tmp, *pos;
2798 LASSERT(set != NULL);
2800 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2801 struct ptlrpc_request *req =
2802 cfs_list_entry(pos, struct ptlrpc_request,
2805 spin_lock(&req->rq_lock);
2806 if (req->rq_phase != RQ_PHASE_RPC) {
2807 spin_unlock(&req->rq_lock);
2812 req->rq_status = -EINTR;
2813 ptlrpc_client_wake_req(req);
2814 spin_unlock(&req->rq_lock);
2818 static __u64 ptlrpc_last_xid;
2819 static spinlock_t ptlrpc_last_xid_lock;
2822 * Initialize the XID for the node. This is common among all requests on
2823 * this node, and only requires the property that it is monotonically
2824 * increasing. It does not need to be sequential. Since this is also used
2825 * as the RDMA match bits, it is important that a single client NOT have
2826 * the same match bits for two different in-flight requests, hence we do
2827 * NOT want to have an XID per target or similar.
2829 * To avoid an unlikely collision between match bits after a client reboot
2830 * (which would deliver old data into the wrong RDMA buffer) initialize
2831 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2832 * If the time is clearly incorrect, we instead use a 62-bit random number.
2833 * In the worst case the random number will overflow 1M RPCs per second in
2834 * 9133 years, or permutations thereof.
2836 #define YEAR_2004 (1ULL << 30)
2837 void ptlrpc_init_xid(void)
2839 time_t now = cfs_time_current_sec();
2841 spin_lock_init(&ptlrpc_last_xid_lock);
2842 if (now < YEAR_2004) {
2843 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2844 ptlrpc_last_xid >>= 2;
2845 ptlrpc_last_xid |= (1ULL << 61);
2847 ptlrpc_last_xid = (__u64)now << 20;
2852 * Increase xid and returns resultng new value to the caller.
2854 __u64 ptlrpc_next_xid(void)
2857 spin_lock(&ptlrpc_last_xid_lock);
2858 tmp = ++ptlrpc_last_xid;
2859 spin_unlock(&ptlrpc_last_xid_lock);
2862 EXPORT_SYMBOL(ptlrpc_next_xid);
2865 * Get a glimpse at what next xid value might have been.
2866 * Returns possible next xid.
2868 __u64 ptlrpc_sample_next_xid(void)
2870 #if BITS_PER_LONG == 32
2871 /* need to avoid possible word tearing on 32-bit systems */
2873 spin_lock(&ptlrpc_last_xid_lock);
2874 tmp = ptlrpc_last_xid + 1;
2875 spin_unlock(&ptlrpc_last_xid_lock);
2878 /* No need to lock, since returned value is racy anyways */
2879 return ptlrpc_last_xid + 1;
2882 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2885 * Functions for operating ptlrpc workers.
2887 * A ptlrpc work is a function which will be running inside ptlrpc context.
2888 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2890 * 1. after a work is created, it can be used many times, that is:
2891 * handler = ptlrpcd_alloc_work();
2892 * ptlrpcd_queue_work();
2894 * queue it again when necessary:
2895 * ptlrpcd_queue_work();
2896 * ptlrpcd_destroy_work();
2897 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2898 * it will only be queued once in any time. Also as its name implies, it may
2899 * have delay before it really runs by ptlrpcd thread.
2901 struct ptlrpc_work_async_args {
2903 int (*cb)(const struct lu_env *, void *);
2907 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2909 static int work_interpreter(const struct lu_env *env,
2910 struct ptlrpc_request *req, void *data, int rc)
2912 struct ptlrpc_work_async_args *arg = data;
2914 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2915 LASSERT(arg->cb != NULL);
2917 return arg->cb(env, arg->cbdata);
2921 * Create a work for ptlrpc.
2923 void *ptlrpcd_alloc_work(struct obd_import *imp,
2924 int (*cb)(const struct lu_env *, void *), void *cbdata)
2926 struct ptlrpc_request *req = NULL;
2927 struct ptlrpc_work_async_args *args;
2933 RETURN(ERR_PTR(-EINVAL));
2935 /* copy some code from deprecated fakereq. */
2938 CERROR("ptlrpc: run out of memory!\n");
2939 RETURN(ERR_PTR(-ENOMEM));
2942 req->rq_send_state = LUSTRE_IMP_FULL;
2943 req->rq_type = PTL_RPC_MSG_REQUEST;
2944 req->rq_import = class_import_get(imp);
2945 req->rq_export = NULL;
2946 req->rq_interpret_reply = work_interpreter;
2947 /* don't want reply */
2948 req->rq_receiving_reply = 0;
2949 req->rq_must_unlink = 0;
2950 req->rq_no_delay = req->rq_no_resend = 1;
2952 spin_lock_init(&req->rq_lock);
2953 CFS_INIT_LIST_HEAD(&req->rq_list);
2954 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
2955 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
2956 CFS_INIT_LIST_HEAD(&req->rq_history_list);
2957 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
2958 cfs_waitq_init(&req->rq_reply_waitq);
2959 cfs_waitq_init(&req->rq_set_waitq);
2960 cfs_atomic_set(&req->rq_refcount, 1);
2962 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
2963 args = ptlrpc_req_async_args(req);
2964 args->magic = PTLRPC_WORK_MAGIC;
2966 args->cbdata = cbdata;
2970 EXPORT_SYMBOL(ptlrpcd_alloc_work);
2972 void ptlrpcd_destroy_work(void *handler)
2974 struct ptlrpc_request *req = handler;
2977 ptlrpc_req_finished(req);
2979 EXPORT_SYMBOL(ptlrpcd_destroy_work);
2981 int ptlrpcd_queue_work(void *handler)
2983 struct ptlrpc_request *req = handler;
2986 * Check if the req is already being queued.
2988 * Here comes a trick: it lacks a way of checking if a req is being
2989 * processed reliably in ptlrpc. Here I have to use refcount of req
2990 * for this purpose. This is okay because the caller should use this
2991 * req as opaque data. - Jinshan
2993 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
2994 if (cfs_atomic_read(&req->rq_refcount) > 1)
2997 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
2998 cfs_atomic_dec(&req->rq_refcount);
3002 /* re-initialize the req */
3003 req->rq_timeout = obd_timeout;
3004 req->rq_sent = cfs_time_current_sec();
3005 req->rq_deadline = req->rq_sent + req->rq_timeout;
3006 req->rq_reply_deadline = req->rq_deadline;
3007 req->rq_phase = RQ_PHASE_INTERPRET;
3008 req->rq_next_phase = RQ_PHASE_COMPLETE;
3009 req->rq_xid = ptlrpc_next_xid();
3010 req->rq_import_generation = req->rq_import->imp_generation;
3012 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3015 EXPORT_SYMBOL(ptlrpcd_queue_work);