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"
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);
357 spin_lock(&req->rq_lock);
360 /* Adjust the local timeout for this req */
361 ptlrpc_at_set_req_timeout(req);
363 olddl = req->rq_deadline;
364 /* server assumes it now has rq_timeout from when it sent the
365 early reply, so client should give it at least that long. */
366 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
367 ptlrpc_at_get_net_latency(req);
369 DEBUG_REQ(D_ADAPTTO, req,
370 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
371 "("CFS_DURATION_T"s)", req->rq_early_count,
372 cfs_time_sub(req->rq_deadline,
373 cfs_time_current_sec()),
374 cfs_time_sub(req->rq_deadline, olddl));
381 * Wind down request pool \a pool.
382 * Frees all requests from the pool too
384 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
387 struct ptlrpc_request *req;
389 LASSERT(pool != NULL);
391 spin_lock(&pool->prp_lock);
392 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
393 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
394 cfs_list_del(&req->rq_list);
395 LASSERT(req->rq_reqbuf);
396 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
397 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
398 OBD_FREE(req, sizeof(*req));
400 spin_unlock(&pool->prp_lock);
401 OBD_FREE(pool, sizeof(*pool));
403 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
406 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
408 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
413 while (size < pool->prp_rq_size)
416 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
417 size == pool->prp_rq_size,
418 "Trying to change pool size with nonempty pool "
419 "from %d to %d bytes\n", pool->prp_rq_size, size);
421 spin_lock(&pool->prp_lock);
422 pool->prp_rq_size = size;
423 for (i = 0; i < num_rq; i++) {
424 struct ptlrpc_request *req;
425 struct lustre_msg *msg;
427 spin_unlock(&pool->prp_lock);
428 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
431 OBD_ALLOC_LARGE(msg, size);
433 OBD_FREE(req, sizeof(struct ptlrpc_request));
436 req->rq_reqbuf = msg;
437 req->rq_reqbuf_len = size;
439 spin_lock(&pool->prp_lock);
440 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
442 spin_unlock(&pool->prp_lock);
445 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
448 * Create and initialize new request pool with given attributes:
449 * \a num_rq - initial number of requests to create for the pool
450 * \a msgsize - maximum message size possible for requests in thid pool
451 * \a populate_pool - function to be called when more requests need to be added
453 * Returns pointer to newly created pool or NULL on error.
455 struct ptlrpc_request_pool *
456 ptlrpc_init_rq_pool(int num_rq, int msgsize,
457 void (*populate_pool)(struct ptlrpc_request_pool *, int))
459 struct ptlrpc_request_pool *pool;
461 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
465 /* Request next power of two for the allocation, because internally
466 kernel would do exactly this */
468 spin_lock_init(&pool->prp_lock);
469 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
470 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
471 pool->prp_populate = populate_pool;
473 populate_pool(pool, num_rq);
475 if (cfs_list_empty(&pool->prp_req_list)) {
476 /* have not allocated a single request for the pool */
477 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
482 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
485 * Fetches one request from pool \a pool
487 static struct ptlrpc_request *
488 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
490 struct ptlrpc_request *request;
491 struct lustre_msg *reqbuf;
496 spin_lock(&pool->prp_lock);
498 /* See if we have anything in a pool, and bail out if nothing,
499 * in writeout path, where this matters, this is safe to do, because
500 * nothing is lost in this case, and when some in-flight requests
501 * complete, this code will be called again. */
502 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
503 spin_unlock(&pool->prp_lock);
507 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
509 cfs_list_del_init(&request->rq_list);
510 spin_unlock(&pool->prp_lock);
512 LASSERT(request->rq_reqbuf);
513 LASSERT(request->rq_pool);
515 reqbuf = request->rq_reqbuf;
516 memset(request, 0, sizeof(*request));
517 request->rq_reqbuf = reqbuf;
518 request->rq_reqbuf_len = pool->prp_rq_size;
519 request->rq_pool = pool;
525 * Returns freed \a request to pool.
527 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
529 struct ptlrpc_request_pool *pool = request->rq_pool;
531 spin_lock(&pool->prp_lock);
532 LASSERT(cfs_list_empty(&request->rq_list));
533 LASSERT(!request->rq_receiving_reply);
534 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
535 spin_unlock(&pool->prp_lock);
538 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
539 __u32 version, int opcode,
540 int count, __u32 *lengths, char **bufs,
541 struct ptlrpc_cli_ctx *ctx)
543 struct obd_import *imp = request->rq_import;
548 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
550 rc = sptlrpc_req_get_ctx(request);
555 sptlrpc_req_set_flavor(request, opcode);
557 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
560 LASSERT(!request->rq_pool);
564 lustre_msg_add_version(request->rq_reqmsg, version);
565 request->rq_send_state = LUSTRE_IMP_FULL;
566 request->rq_type = PTL_RPC_MSG_REQUEST;
567 request->rq_export = NULL;
569 request->rq_req_cbid.cbid_fn = request_out_callback;
570 request->rq_req_cbid.cbid_arg = request;
572 request->rq_reply_cbid.cbid_fn = reply_in_callback;
573 request->rq_reply_cbid.cbid_arg = request;
575 request->rq_reply_deadline = 0;
576 request->rq_phase = RQ_PHASE_NEW;
577 request->rq_next_phase = RQ_PHASE_UNDEFINED;
579 request->rq_request_portal = imp->imp_client->cli_request_portal;
580 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
582 ptlrpc_at_set_req_timeout(request);
584 spin_lock_init(&request->rq_lock);
585 CFS_INIT_LIST_HEAD(&request->rq_list);
586 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
587 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
588 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
589 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
590 CFS_INIT_LIST_HEAD(&request->rq_history_list);
591 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
592 cfs_waitq_init(&request->rq_reply_waitq);
593 cfs_waitq_init(&request->rq_set_waitq);
594 request->rq_xid = ptlrpc_next_xid();
595 cfs_atomic_set(&request->rq_refcount, 1);
597 lustre_msg_set_opc(request->rq_reqmsg, opcode);
601 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
603 class_import_put(imp);
607 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
608 __u32 version, int opcode, char **bufs,
609 struct ptlrpc_cli_ctx *ctx)
613 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
614 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
615 request->rq_pill.rc_area[RCL_CLIENT],
618 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
621 * Pack request buffers for network transfer, performing necessary encryption
622 * steps if necessary.
624 int ptlrpc_request_pack(struct ptlrpc_request *request,
625 __u32 version, int opcode)
628 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
632 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
633 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
634 * have to send old ptlrpc_body to keep interoprability with these
637 * Only three kinds of server->client RPCs so far:
642 * XXX This should be removed whenever we drop the interoprability with
643 * the these old clients.
645 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
646 opcode == LDLM_GL_CALLBACK)
647 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
648 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
652 EXPORT_SYMBOL(ptlrpc_request_pack);
655 * Helper function to allocate new request on import \a imp
656 * and possibly using existing request from pool \a pool if provided.
657 * Returns allocated request structure with import field filled or
661 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
662 struct ptlrpc_request_pool *pool)
664 struct ptlrpc_request *request = NULL;
667 request = ptlrpc_prep_req_from_pool(pool);
670 OBD_ALLOC_PTR(request);
673 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
674 LASSERT(imp != LP_POISON);
675 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
677 LASSERT(imp->imp_client != LP_POISON);
679 request->rq_import = class_import_get(imp);
681 CERROR("request allocation out of memory\n");
688 * Helper function for creating a request.
689 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
690 * buffer structures according to capsule template \a format.
691 * Returns allocated request structure pointer or NULL on error.
693 static struct ptlrpc_request *
694 ptlrpc_request_alloc_internal(struct obd_import *imp,
695 struct ptlrpc_request_pool * pool,
696 const struct req_format *format)
698 struct ptlrpc_request *request;
700 request = __ptlrpc_request_alloc(imp, pool);
704 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
705 req_capsule_set(&request->rq_pill, format);
710 * Allocate new request structure for import \a imp and initialize its
711 * buffer structure according to capsule template \a format.
713 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
714 const struct req_format *format)
716 return ptlrpc_request_alloc_internal(imp, NULL, format);
718 EXPORT_SYMBOL(ptlrpc_request_alloc);
721 * Allocate new request structure for import \a imp from pool \a pool and
722 * initialize its buffer structure according to capsule template \a format.
724 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
725 struct ptlrpc_request_pool * pool,
726 const struct req_format *format)
728 return ptlrpc_request_alloc_internal(imp, pool, format);
730 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
733 * For requests not from pool, free memory of the request structure.
734 * For requests obtained from a pool earlier, return request back to pool.
736 void ptlrpc_request_free(struct ptlrpc_request *request)
738 if (request->rq_pool)
739 __ptlrpc_free_req_to_pool(request);
741 OBD_FREE_PTR(request);
743 EXPORT_SYMBOL(ptlrpc_request_free);
746 * Allocate new request for operatione \a opcode and immediatelly pack it for
748 * Only used for simple requests like OBD_PING where the only important
749 * part of the request is operation itself.
750 * Returns allocated request or NULL on error.
752 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
753 const struct req_format *format,
754 __u32 version, int opcode)
756 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
760 rc = ptlrpc_request_pack(req, version, opcode);
762 ptlrpc_request_free(req);
768 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
771 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
772 * for operation \a opcode. Request would contain \a count buffers.
773 * Sizes of buffers are described in array \a lengths and buffers themselves
774 * are provided by a pointer \a bufs.
775 * Returns prepared request structure pointer or NULL on error.
777 struct ptlrpc_request *
778 ptlrpc_prep_req_pool(struct obd_import *imp,
779 __u32 version, int opcode,
780 int count, __u32 *lengths, char **bufs,
781 struct ptlrpc_request_pool *pool)
783 struct ptlrpc_request *request;
786 request = __ptlrpc_request_alloc(imp, pool);
790 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
791 lengths, bufs, NULL);
793 ptlrpc_request_free(request);
798 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
801 * Same as ptlrpc_prep_req_pool, but without pool
803 struct ptlrpc_request *
804 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
805 __u32 *lengths, char **bufs)
807 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
810 EXPORT_SYMBOL(ptlrpc_prep_req);
813 * Allocate and initialize new request set structure.
814 * Returns a pointer to the newly allocated set structure or NULL on error.
816 struct ptlrpc_request_set *ptlrpc_prep_set(void)
818 struct ptlrpc_request_set *set;
821 OBD_ALLOC(set, sizeof *set);
824 cfs_atomic_set(&set->set_refcount, 1);
825 CFS_INIT_LIST_HEAD(&set->set_requests);
826 cfs_waitq_init(&set->set_waitq);
827 cfs_atomic_set(&set->set_new_count, 0);
828 cfs_atomic_set(&set->set_remaining, 0);
829 spin_lock_init(&set->set_new_req_lock);
830 CFS_INIT_LIST_HEAD(&set->set_new_requests);
831 CFS_INIT_LIST_HEAD(&set->set_cblist);
832 set->set_max_inflight = UINT_MAX;
833 set->set_producer = NULL;
834 set->set_producer_arg = NULL;
839 EXPORT_SYMBOL(ptlrpc_prep_set);
842 * Allocate and initialize new request set structure with flow control
843 * extension. This extension allows to control the number of requests in-flight
844 * for the whole set. A callback function to generate requests must be provided
845 * and the request set will keep the number of requests sent over the wire to
847 * Returns a pointer to the newly allocated set structure or NULL on error.
849 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
853 struct ptlrpc_request_set *set;
855 set = ptlrpc_prep_set();
859 set->set_max_inflight = max;
860 set->set_producer = func;
861 set->set_producer_arg = arg;
865 EXPORT_SYMBOL(ptlrpc_prep_fcset);
868 * Wind down and free request set structure previously allocated with
870 * Ensures that all requests on the set have completed and removes
871 * all requests from the request list in a set.
872 * If any unsent request happen to be on the list, pretends that they got
873 * an error in flight and calls their completion handler.
875 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
883 /* Requests on the set should either all be completed, or all be new */
884 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
885 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
886 cfs_list_for_each (tmp, &set->set_requests) {
887 struct ptlrpc_request *req =
888 cfs_list_entry(tmp, struct ptlrpc_request,
891 LASSERT(req->rq_phase == expected_phase);
895 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
896 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
897 cfs_atomic_read(&set->set_remaining), n);
899 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
900 struct ptlrpc_request *req =
901 cfs_list_entry(tmp, struct ptlrpc_request,
903 cfs_list_del_init(&req->rq_set_chain);
905 LASSERT(req->rq_phase == expected_phase);
907 if (req->rq_phase == RQ_PHASE_NEW) {
908 ptlrpc_req_interpret(NULL, req, -EBADR);
909 cfs_atomic_dec(&set->set_remaining);
912 spin_lock(&req->rq_lock);
914 req->rq_invalid_rqset = 0;
915 spin_unlock(&req->rq_lock);
917 ptlrpc_req_finished (req);
920 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
922 ptlrpc_reqset_put(set);
925 EXPORT_SYMBOL(ptlrpc_set_destroy);
928 * Add a callback function \a fn to the set.
929 * This function would be called when all requests on this set are completed.
930 * The function will be passed \a data argument.
932 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
933 set_interpreter_func fn, void *data)
935 struct ptlrpc_set_cbdata *cbdata;
937 OBD_ALLOC_PTR(cbdata);
941 cbdata->psc_interpret = fn;
942 cbdata->psc_data = data;
943 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
947 EXPORT_SYMBOL(ptlrpc_set_add_cb);
950 * Add a new request to the general purpose request set.
951 * Assumes request reference from the caller.
953 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
954 struct ptlrpc_request *req)
956 LASSERT(cfs_list_empty(&req->rq_set_chain));
958 /* The set takes over the caller's request reference */
959 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
961 cfs_atomic_inc(&set->set_remaining);
962 req->rq_queued_time = cfs_time_current();
964 if (req->rq_reqmsg != NULL)
965 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
967 if (set->set_producer != NULL)
968 /* If the request set has a producer callback, the RPC must be
969 * sent straight away */
970 ptlrpc_send_new_req(req);
972 EXPORT_SYMBOL(ptlrpc_set_add_req);
975 * Add a request to a request with dedicated server thread
976 * and wake the thread to make any necessary processing.
977 * Currently only used for ptlrpcd.
979 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
980 struct ptlrpc_request *req)
982 struct ptlrpc_request_set *set = pc->pc_set;
985 LASSERT(req->rq_set == NULL);
986 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
988 spin_lock(&set->set_new_req_lock);
990 * The set takes over the caller's request reference.
993 req->rq_queued_time = cfs_time_current();
994 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
995 count = cfs_atomic_inc_return(&set->set_new_count);
996 spin_unlock(&set->set_new_req_lock);
998 /* Only need to call wakeup once for the first entry. */
1000 cfs_waitq_signal(&set->set_waitq);
1002 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1003 * guarantee the async RPC can be processed ASAP, we have
1004 * no other better choice. It maybe fixed in future. */
1005 for (i = 0; i < pc->pc_npartners; i++)
1006 cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
1009 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1012 * Based on the current state of the import, determine if the request
1013 * can be sent, is an error, or should be delayed.
1015 * Returns true if this request should be delayed. If false, and
1016 * *status is set, then the request can not be sent and *status is the
1017 * error code. If false and status is 0, then request can be sent.
1019 * The imp->imp_lock must be held.
1021 static int ptlrpc_import_delay_req(struct obd_import *imp,
1022 struct ptlrpc_request *req, int *status)
1027 LASSERT (status != NULL);
1030 if (req->rq_ctx_init || req->rq_ctx_fini) {
1031 /* always allow ctx init/fini rpc go through */
1032 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1033 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1035 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1036 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1038 } else if (ptlrpc_send_limit_expired(req)) {
1039 /* probably doesn't need to be a D_ERROR after initial testing */
1040 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1042 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1043 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1044 /* allow CONNECT even if import is invalid */ ;
1045 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1046 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1049 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1050 if (!imp->imp_deactive)
1051 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1052 *status = -ESHUTDOWN; /* bz 12940 */
1053 } else if (req->rq_import_generation != imp->imp_generation) {
1054 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1056 } else if (req->rq_send_state != imp->imp_state) {
1057 /* invalidate in progress - any requests should be drop */
1058 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1059 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1061 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1062 *status = -EWOULDBLOCK;
1072 * Decide if the eror message regarding provided request \a req
1073 * should be printed to the console or not.
1074 * Makes it's decision on request status and other properties.
1075 * Returns 1 to print error on the system console or 0 if not.
1077 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1082 LASSERT(req->rq_reqmsg != NULL);
1083 opc = lustre_msg_get_opc(req->rq_reqmsg);
1085 /* Suppress particular reconnect errors which are to be expected. No
1086 * errors are suppressed for the initial connection on an import */
1087 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1088 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1090 /* Suppress timed out reconnect requests */
1091 if (req->rq_timedout)
1094 /* Suppress unavailable/again reconnect requests */
1095 err = lustre_msg_get_status(req->rq_repmsg);
1096 if (err == -ENODEV || err == -EAGAIN)
1104 * Check request processing status.
1105 * Returns the status.
1107 static int ptlrpc_check_status(struct ptlrpc_request *req)
1112 err = lustre_msg_get_status(req->rq_repmsg);
1113 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1114 struct obd_import *imp = req->rq_import;
1115 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1116 if (ptlrpc_console_allow(req))
1117 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1118 "communicating with %s. The %s "
1119 "operation failed with %d\n",
1121 imp->imp_connection->c_peer.nid),
1122 ll_opcode2str(opc), err);
1123 RETURN(err < 0 ? err : -EINVAL);
1127 DEBUG_REQ(D_INFO, req, "status is %d", err);
1128 } else if (err > 0) {
1129 /* XXX: translate this error from net to host */
1130 DEBUG_REQ(D_INFO, req, "status is %d", err);
1137 * save pre-versions of objects into request for replay.
1138 * Versions are obtained from server reply.
1141 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1143 struct lustre_msg *repmsg = req->rq_repmsg;
1144 struct lustre_msg *reqmsg = req->rq_reqmsg;
1145 __u64 *versions = lustre_msg_get_versions(repmsg);
1148 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1152 lustre_msg_set_versions(reqmsg, versions);
1153 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1154 versions[0], versions[1]);
1160 * Callback function called when client receives RPC reply for \a req.
1161 * Returns 0 on success or error code.
1162 * The return alue would be assigned to req->rq_status by the caller
1163 * as request processing status.
1164 * This function also decides if the request needs to be saved for later replay.
1166 static int after_reply(struct ptlrpc_request *req)
1168 struct obd_import *imp = req->rq_import;
1169 struct obd_device *obd = req->rq_import->imp_obd;
1171 struct timeval work_start;
1175 LASSERT(obd != NULL);
1176 /* repbuf must be unlinked */
1177 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1179 if (req->rq_reply_truncate) {
1180 if (ptlrpc_no_resend(req)) {
1181 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1182 " expected: %d, actual size: %d",
1183 req->rq_nob_received, req->rq_repbuf_len);
1187 sptlrpc_cli_free_repbuf(req);
1188 /* Pass the required reply buffer size (include
1189 * space for early reply).
1190 * NB: no need to roundup because alloc_repbuf
1191 * will roundup it */
1192 req->rq_replen = req->rq_nob_received;
1193 req->rq_nob_received = 0;
1199 * NB Until this point, the whole of the incoming message,
1200 * including buflens, status etc is in the sender's byte order.
1202 rc = sptlrpc_cli_unwrap_reply(req);
1204 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1208 /* retry indefinitely on EINPROGRESS */
1209 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1210 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1211 time_t now = cfs_time_current_sec();
1213 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1215 req->rq_nr_resend++;
1217 /* allocate new xid to avoid reply reconstruction */
1218 if (!req->rq_bulk) {
1219 /* new xid is already allocated for bulk in
1220 * ptlrpc_check_set() */
1221 req->rq_xid = ptlrpc_next_xid();
1222 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1223 "resend on EINPROGRESS");
1226 /* Readjust the timeout for current conditions */
1227 ptlrpc_at_set_req_timeout(req);
1228 /* delay resend to give a chance to the server to get ready.
1229 * The delay is increased by 1s on every resend and is capped to
1230 * the current request timeout (i.e. obd_timeout if AT is off,
1231 * or AT service time x 125% + 5s, see at_est2timeout) */
1232 if (req->rq_nr_resend > req->rq_timeout)
1233 req->rq_sent = now + req->rq_timeout;
1235 req->rq_sent = now + req->rq_nr_resend;
1239 * Security layer unwrap might ask resend this request.
1244 rc = unpack_reply(req);
1248 cfs_gettimeofday(&work_start);
1249 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1250 if (obd->obd_svc_stats != NULL) {
1251 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1253 ptlrpc_lprocfs_rpc_sent(req, timediff);
1256 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1257 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1258 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1259 lustre_msg_get_type(req->rq_repmsg));
1263 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1264 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1265 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1266 ptlrpc_at_adj_net_latency(req,
1267 lustre_msg_get_service_time(req->rq_repmsg));
1269 rc = ptlrpc_check_status(req);
1270 imp->imp_connect_error = rc;
1274 * Either we've been evicted, or the server has failed for
1275 * some reason. Try to reconnect, and if that fails, punt to
1278 if (ll_rpc_recoverable_error(rc)) {
1279 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1280 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1283 ptlrpc_request_handle_notconn(req);
1288 * Let's look if server sent slv. Do it only for RPC with
1291 ldlm_cli_update_pool(req);
1295 * Store transno in reqmsg for replay.
1297 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1298 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1299 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1302 if (imp->imp_replayable) {
1303 spin_lock(&imp->imp_lock);
1305 * No point in adding already-committed requests to the replay
1306 * list, we will just remove them immediately. b=9829
1308 if (req->rq_transno != 0 &&
1310 lustre_msg_get_last_committed(req->rq_repmsg) ||
1312 /** version recovery */
1313 ptlrpc_save_versions(req);
1314 ptlrpc_retain_replayable_request(req, imp);
1315 } else if (req->rq_commit_cb != NULL) {
1316 spin_unlock(&imp->imp_lock);
1317 req->rq_commit_cb(req);
1318 spin_lock(&imp->imp_lock);
1322 * Replay-enabled imports return commit-status information.
1324 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1325 imp->imp_peer_committed_transno =
1326 lustre_msg_get_last_committed(req->rq_repmsg);
1328 ptlrpc_free_committed(imp);
1330 if (req->rq_transno > imp->imp_peer_committed_transno)
1331 ptlrpc_pinger_commit_expected(imp);
1333 spin_unlock(&imp->imp_lock);
1340 * Helper function to send request \a req over the network for the first time
1341 * Also adjusts request phase.
1342 * Returns 0 on success or error code.
1344 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1346 struct obd_import *imp = req->rq_import;
1350 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1351 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1352 (!req->rq_generation_set ||
1353 req->rq_import_generation == imp->imp_generation))
1356 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1358 spin_lock(&imp->imp_lock);
1360 if (!req->rq_generation_set)
1361 req->rq_import_generation = imp->imp_generation;
1363 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1364 spin_lock(&req->rq_lock);
1365 req->rq_waiting = 1;
1366 spin_unlock(&req->rq_lock);
1368 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1369 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1370 ptlrpc_import_state_name(req->rq_send_state),
1371 ptlrpc_import_state_name(imp->imp_state));
1372 LASSERT(cfs_list_empty(&req->rq_list));
1373 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1374 cfs_atomic_inc(&req->rq_import->imp_inflight);
1375 spin_unlock(&imp->imp_lock);
1380 spin_unlock(&imp->imp_lock);
1381 req->rq_status = rc;
1382 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1386 LASSERT(cfs_list_empty(&req->rq_list));
1387 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1388 cfs_atomic_inc(&req->rq_import->imp_inflight);
1389 spin_unlock(&imp->imp_lock);
1391 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1393 rc = sptlrpc_req_refresh_ctx(req, -1);
1396 req->rq_status = rc;
1399 req->rq_wait_ctx = 1;
1404 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1405 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1406 imp->imp_obd->obd_uuid.uuid,
1407 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1408 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1409 lustre_msg_get_opc(req->rq_reqmsg));
1411 rc = ptl_send_rpc(req, 0);
1413 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1414 req->rq_net_err = 1;
1420 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1425 LASSERT(set->set_producer != NULL);
1427 remaining = cfs_atomic_read(&set->set_remaining);
1429 /* populate the ->set_requests list with requests until we
1430 * reach the maximum number of RPCs in flight for this set */
1431 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1432 rc = set->set_producer(set, set->set_producer_arg);
1433 if (rc == -ENOENT) {
1434 /* no more RPC to produce */
1435 set->set_producer = NULL;
1436 set->set_producer_arg = NULL;
1441 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1445 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1446 * and no more replies are expected.
1447 * (it is possible to get less replies than requests sent e.g. due to timed out
1448 * requests or requests that we had trouble to send out)
1450 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1452 cfs_list_t *tmp, *next;
1453 int force_timer_recalc = 0;
1456 if (cfs_atomic_read(&set->set_remaining) == 0)
1459 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1460 struct ptlrpc_request *req =
1461 cfs_list_entry(tmp, struct ptlrpc_request,
1463 struct obd_import *imp = req->rq_import;
1464 int unregistered = 0;
1467 if (req->rq_phase == RQ_PHASE_NEW &&
1468 ptlrpc_send_new_req(req)) {
1469 force_timer_recalc = 1;
1472 /* delayed send - skip */
1473 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1476 /* delayed resend - skip */
1477 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1478 req->rq_sent > cfs_time_current_sec())
1481 if (!(req->rq_phase == RQ_PHASE_RPC ||
1482 req->rq_phase == RQ_PHASE_BULK ||
1483 req->rq_phase == RQ_PHASE_INTERPRET ||
1484 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1485 req->rq_phase == RQ_PHASE_COMPLETE)) {
1486 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1490 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1491 LASSERT(req->rq_next_phase != req->rq_phase);
1492 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1495 * Skip processing until reply is unlinked. We
1496 * can't return to pool before that and we can't
1497 * call interpret before that. We need to make
1498 * sure that all rdma transfers finished and will
1499 * not corrupt any data.
1501 if (ptlrpc_client_recv_or_unlink(req) ||
1502 ptlrpc_client_bulk_active(req))
1506 * Turn fail_loc off to prevent it from looping
1509 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1510 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1513 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1514 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1519 * Move to next phase if reply was successfully
1522 ptlrpc_rqphase_move(req, req->rq_next_phase);
1525 if (req->rq_phase == RQ_PHASE_COMPLETE)
1528 if (req->rq_phase == RQ_PHASE_INTERPRET)
1529 GOTO(interpret, req->rq_status);
1532 * Note that this also will start async reply unlink.
1534 if (req->rq_net_err && !req->rq_timedout) {
1535 ptlrpc_expire_one_request(req, 1);
1538 * Check if we still need to wait for unlink.
1540 if (ptlrpc_client_recv_or_unlink(req) ||
1541 ptlrpc_client_bulk_active(req))
1543 /* If there is no need to resend, fail it now. */
1544 if (req->rq_no_resend) {
1545 if (req->rq_status == 0)
1546 req->rq_status = -EIO;
1547 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1548 GOTO(interpret, req->rq_status);
1555 spin_lock(&req->rq_lock);
1556 req->rq_replied = 0;
1557 spin_unlock(&req->rq_lock);
1558 if (req->rq_status == 0)
1559 req->rq_status = -EIO;
1560 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1561 GOTO(interpret, req->rq_status);
1564 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1565 * so it sets rq_intr regardless of individual rpc
1566 * timeouts. The synchronous IO waiting path sets
1567 * rq_intr irrespective of whether ptlrpcd
1568 * has seen a timeout. Our policy is to only interpret
1569 * interrupted rpcs after they have timed out, so we
1570 * need to enforce that here.
1573 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1574 req->rq_wait_ctx)) {
1575 req->rq_status = -EINTR;
1576 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1577 GOTO(interpret, req->rq_status);
1580 if (req->rq_phase == RQ_PHASE_RPC) {
1581 if (req->rq_timedout || req->rq_resend ||
1582 req->rq_waiting || req->rq_wait_ctx) {
1585 if (!ptlrpc_unregister_reply(req, 1))
1588 spin_lock(&imp->imp_lock);
1589 if (ptlrpc_import_delay_req(imp, req, &status)){
1590 /* put on delay list - only if we wait
1591 * recovery finished - before send */
1592 cfs_list_del_init(&req->rq_list);
1593 cfs_list_add_tail(&req->rq_list,
1596 spin_unlock(&imp->imp_lock);
1601 req->rq_status = status;
1602 ptlrpc_rqphase_move(req,
1603 RQ_PHASE_INTERPRET);
1604 spin_unlock(&imp->imp_lock);
1605 GOTO(interpret, req->rq_status);
1607 if (ptlrpc_no_resend(req) &&
1608 !req->rq_wait_ctx) {
1609 req->rq_status = -ENOTCONN;
1610 ptlrpc_rqphase_move(req,
1611 RQ_PHASE_INTERPRET);
1612 spin_unlock(&imp->imp_lock);
1613 GOTO(interpret, req->rq_status);
1616 cfs_list_del_init(&req->rq_list);
1617 cfs_list_add_tail(&req->rq_list,
1618 &imp->imp_sending_list);
1620 spin_unlock(&imp->imp_lock);
1622 spin_lock(&req->rq_lock);
1623 req->rq_waiting = 0;
1624 spin_unlock(&req->rq_lock);
1626 if (req->rq_timedout || req->rq_resend) {
1627 /* This is re-sending anyways,
1628 * let's mark req as resend. */
1629 spin_lock(&req->rq_lock);
1631 spin_unlock(&req->rq_lock);
1635 if (!ptlrpc_unregister_bulk(req, 1))
1638 /* ensure previous bulk fails */
1639 old_xid = req->rq_xid;
1640 req->rq_xid = ptlrpc_next_xid();
1641 CDEBUG(D_HA, "resend bulk "
1644 old_xid, req->rq_xid);
1648 * rq_wait_ctx is only touched by ptlrpcd,
1649 * so no lock is needed here.
1651 status = sptlrpc_req_refresh_ctx(req, -1);
1654 req->rq_status = status;
1655 spin_lock(&req->rq_lock);
1656 req->rq_wait_ctx = 0;
1657 spin_unlock(&req->rq_lock);
1658 force_timer_recalc = 1;
1660 spin_lock(&req->rq_lock);
1661 req->rq_wait_ctx = 1;
1662 spin_unlock(&req->rq_lock);
1667 spin_lock(&req->rq_lock);
1668 req->rq_wait_ctx = 0;
1669 spin_unlock(&req->rq_lock);
1672 rc = ptl_send_rpc(req, 0);
1674 DEBUG_REQ(D_HA, req,
1675 "send failed: rc = %d", rc);
1676 force_timer_recalc = 1;
1677 spin_lock(&req->rq_lock);
1678 req->rq_net_err = 1;
1679 spin_unlock(&req->rq_lock);
1681 /* need to reset the timeout */
1682 force_timer_recalc = 1;
1685 spin_lock(&req->rq_lock);
1687 if (ptlrpc_client_early(req)) {
1688 ptlrpc_at_recv_early_reply(req);
1689 spin_unlock(&req->rq_lock);
1693 /* Still waiting for a reply? */
1694 if (ptlrpc_client_recv(req)) {
1695 spin_unlock(&req->rq_lock);
1699 /* Did we actually receive a reply? */
1700 if (!ptlrpc_client_replied(req)) {
1701 spin_unlock(&req->rq_lock);
1705 spin_unlock(&req->rq_lock);
1707 /* unlink from net because we are going to
1708 * swab in-place of reply buffer */
1709 unregistered = ptlrpc_unregister_reply(req, 1);
1713 req->rq_status = after_reply(req);
1717 /* If there is no bulk associated with this request,
1718 * then we're done and should let the interpreter
1719 * process the reply. Similarly if the RPC returned
1720 * an error, and therefore the bulk will never arrive.
1722 if (req->rq_bulk == NULL || req->rq_status < 0) {
1723 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1724 GOTO(interpret, req->rq_status);
1727 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1730 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1731 if (ptlrpc_client_bulk_active(req))
1734 if (!req->rq_bulk->bd_success) {
1735 /* The RPC reply arrived OK, but the bulk screwed
1736 * up! Dead weird since the server told us the RPC
1737 * was good after getting the REPLY for her GET or
1738 * the ACK for her PUT. */
1739 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1740 req->rq_status = -EIO;
1743 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1746 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1748 /* This moves to "unregistering" phase we need to wait for
1750 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1751 /* start async bulk unlink too */
1752 ptlrpc_unregister_bulk(req, 1);
1756 if (!ptlrpc_unregister_bulk(req, 1))
1759 /* When calling interpret receiving already should be
1761 LASSERT(!req->rq_receiving_reply);
1763 ptlrpc_req_interpret(env, req, req->rq_status);
1765 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1767 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1768 "Completed RPC pname:cluuid:pid:xid:nid:"
1769 "opc %s:%s:%d:"LPU64":%s:%d\n",
1770 cfs_curproc_comm(), imp->imp_obd->obd_uuid.uuid,
1771 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1772 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1773 lustre_msg_get_opc(req->rq_reqmsg));
1775 spin_lock(&imp->imp_lock);
1776 /* Request already may be not on sending or delaying list. This
1777 * may happen in the case of marking it erroneous for the case
1778 * ptlrpc_import_delay_req(req, status) find it impossible to
1779 * allow sending this rpc and returns *status != 0. */
1780 if (!cfs_list_empty(&req->rq_list)) {
1781 cfs_list_del_init(&req->rq_list);
1782 cfs_atomic_dec(&imp->imp_inflight);
1784 spin_unlock(&imp->imp_lock);
1786 cfs_atomic_dec(&set->set_remaining);
1787 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1789 if (set->set_producer) {
1790 /* produce a new request if possible */
1791 if (ptlrpc_set_producer(set) > 0)
1792 force_timer_recalc = 1;
1794 /* free the request that has just been completed
1795 * in order not to pollute set->set_requests */
1796 cfs_list_del_init(&req->rq_set_chain);
1797 spin_lock(&req->rq_lock);
1799 req->rq_invalid_rqset = 0;
1800 spin_unlock(&req->rq_lock);
1802 /* record rq_status to compute the final status later */
1803 if (req->rq_status != 0)
1804 set->set_rc = req->rq_status;
1805 ptlrpc_req_finished(req);
1809 /* If we hit an error, we want to recover promptly. */
1810 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1812 EXPORT_SYMBOL(ptlrpc_check_set);
1815 * Time out request \a req. is \a async_unlink is set, that means do not wait
1816 * until LNet actually confirms network buffer unlinking.
1817 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1819 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1821 struct obd_import *imp = req->rq_import;
1825 spin_lock(&req->rq_lock);
1826 req->rq_timedout = 1;
1827 spin_unlock(&req->rq_lock);
1829 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1830 "/real "CFS_DURATION_T"]",
1831 req->rq_net_err ? "failed due to network error" :
1832 ((req->rq_real_sent == 0 ||
1833 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1834 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1835 "timed out for sent delay" : "timed out for slow reply"),
1836 req->rq_sent, req->rq_real_sent);
1838 if (imp != NULL && obd_debug_peer_on_timeout)
1839 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1841 ptlrpc_unregister_reply(req, async_unlink);
1842 ptlrpc_unregister_bulk(req, async_unlink);
1844 if (obd_dump_on_timeout)
1845 libcfs_debug_dumplog();
1848 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1852 cfs_atomic_inc(&imp->imp_timeouts);
1854 /* The DLM server doesn't want recovery run on its imports. */
1855 if (imp->imp_dlm_fake)
1858 /* If this request is for recovery or other primordial tasks,
1859 * then error it out here. */
1860 if (req->rq_ctx_init || req->rq_ctx_fini ||
1861 req->rq_send_state != LUSTRE_IMP_FULL ||
1862 imp->imp_obd->obd_no_recov) {
1863 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1864 ptlrpc_import_state_name(req->rq_send_state),
1865 ptlrpc_import_state_name(imp->imp_state));
1866 spin_lock(&req->rq_lock);
1867 req->rq_status = -ETIMEDOUT;
1869 spin_unlock(&req->rq_lock);
1873 /* if a request can't be resent we can't wait for an answer after
1875 if (ptlrpc_no_resend(req)) {
1876 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1880 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1886 * Time out all uncompleted requests in request set pointed by \a data
1887 * Callback used when waiting on sets with l_wait_event.
1890 int ptlrpc_expired_set(void *data)
1892 struct ptlrpc_request_set *set = data;
1894 time_t now = cfs_time_current_sec();
1897 LASSERT(set != NULL);
1900 * A timeout expired. See which reqs it applies to...
1902 cfs_list_for_each (tmp, &set->set_requests) {
1903 struct ptlrpc_request *req =
1904 cfs_list_entry(tmp, struct ptlrpc_request,
1907 /* don't expire request waiting for context */
1908 if (req->rq_wait_ctx)
1911 /* Request in-flight? */
1912 if (!((req->rq_phase == RQ_PHASE_RPC &&
1913 !req->rq_waiting && !req->rq_resend) ||
1914 (req->rq_phase == RQ_PHASE_BULK)))
1917 if (req->rq_timedout || /* already dealt with */
1918 req->rq_deadline > now) /* not expired */
1921 /* Deal with this guy. Do it asynchronously to not block
1922 * ptlrpcd thread. */
1923 ptlrpc_expire_one_request(req, 1);
1927 * When waiting for a whole set, we always break out of the
1928 * sleep so we can recalculate the timeout, or enable interrupts
1929 * if everyone's timed out.
1933 EXPORT_SYMBOL(ptlrpc_expired_set);
1936 * Sets rq_intr flag in \a req under spinlock.
1938 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1940 spin_lock(&req->rq_lock);
1942 spin_unlock(&req->rq_lock);
1944 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
1947 * Interrupts (sets interrupted flag) all uncompleted requests in
1948 * a set \a data. Callback for l_wait_event for interruptible waits.
1950 void ptlrpc_interrupted_set(void *data)
1952 struct ptlrpc_request_set *set = data;
1955 LASSERT(set != NULL);
1956 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
1958 cfs_list_for_each(tmp, &set->set_requests) {
1959 struct ptlrpc_request *req =
1960 cfs_list_entry(tmp, struct ptlrpc_request,
1963 if (req->rq_phase != RQ_PHASE_RPC &&
1964 req->rq_phase != RQ_PHASE_UNREGISTERING)
1967 ptlrpc_mark_interrupted(req);
1970 EXPORT_SYMBOL(ptlrpc_interrupted_set);
1973 * Get the smallest timeout in the set; this does NOT set a timeout.
1975 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1978 time_t now = cfs_time_current_sec();
1980 struct ptlrpc_request *req;
1984 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1986 cfs_list_for_each(tmp, &set->set_requests) {
1987 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1990 * Request in-flight?
1992 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
1993 (req->rq_phase == RQ_PHASE_BULK) ||
1994 (req->rq_phase == RQ_PHASE_NEW)))
1998 * Already timed out.
2000 if (req->rq_timedout)
2006 if (req->rq_wait_ctx)
2009 if (req->rq_phase == RQ_PHASE_NEW)
2010 deadline = req->rq_sent;
2011 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2012 deadline = req->rq_sent;
2014 deadline = req->rq_sent + req->rq_timeout;
2016 if (deadline <= now) /* actually expired already */
2017 timeout = 1; /* ASAP */
2018 else if (timeout == 0 || timeout > deadline - now)
2019 timeout = deadline - now;
2023 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2026 * Send all unset request from the set and then wait untill all
2027 * requests in the set complete (either get a reply, timeout, get an
2028 * error or otherwise be interrupted).
2029 * Returns 0 on success or error code otherwise.
2031 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2034 struct ptlrpc_request *req;
2035 struct l_wait_info lwi;
2039 if (set->set_producer)
2040 (void)ptlrpc_set_producer(set);
2042 cfs_list_for_each(tmp, &set->set_requests) {
2043 req = cfs_list_entry(tmp, struct ptlrpc_request,
2045 if (req->rq_phase == RQ_PHASE_NEW)
2046 (void)ptlrpc_send_new_req(req);
2049 if (cfs_list_empty(&set->set_requests))
2053 timeout = ptlrpc_set_next_timeout(set);
2055 /* wait until all complete, interrupted, or an in-flight
2057 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2060 if (timeout == 0 && !cfs_signal_pending())
2062 * No requests are in-flight (ether timed out
2063 * or delayed), so we can allow interrupts.
2064 * We still want to block for a limited time,
2065 * so we allow interrupts during the timeout.
2067 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2069 ptlrpc_interrupted_set, set);
2072 * At least one request is in flight, so no
2073 * interrupts are allowed. Wait until all
2074 * complete, or an in-flight req times out.
2076 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2077 ptlrpc_expired_set, set);
2079 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2081 /* LU-769 - if we ignored the signal because it was already
2082 * pending when we started, we need to handle it now or we risk
2083 * it being ignored forever */
2084 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2085 cfs_signal_pending()) {
2086 cfs_sigset_t blocked_sigs =
2087 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2089 /* In fact we only interrupt for the "fatal" signals
2090 * like SIGINT or SIGKILL. We still ignore less
2091 * important signals since ptlrpc set is not easily
2092 * reentrant from userspace again */
2093 if (cfs_signal_pending())
2094 ptlrpc_interrupted_set(set);
2095 cfs_restore_sigs(blocked_sigs);
2098 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2100 /* -EINTR => all requests have been flagged rq_intr so next
2102 * -ETIMEDOUT => someone timed out. When all reqs have
2103 * timed out, signals are enabled allowing completion with
2105 * I don't really care if we go once more round the loop in
2106 * the error cases -eeb. */
2107 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2108 cfs_list_for_each(tmp, &set->set_requests) {
2109 req = cfs_list_entry(tmp, struct ptlrpc_request,
2111 spin_lock(&req->rq_lock);
2112 req->rq_invalid_rqset = 1;
2113 spin_unlock(&req->rq_lock);
2116 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2118 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2120 rc = set->set_rc; /* rq_status of already freed requests if any */
2121 cfs_list_for_each(tmp, &set->set_requests) {
2122 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2124 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2125 if (req->rq_status != 0)
2126 rc = req->rq_status;
2129 if (set->set_interpret != NULL) {
2130 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2132 rc = interpreter (set, set->set_arg, rc);
2134 struct ptlrpc_set_cbdata *cbdata, *n;
2137 cfs_list_for_each_entry_safe(cbdata, n,
2138 &set->set_cblist, psc_item) {
2139 cfs_list_del_init(&cbdata->psc_item);
2140 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2143 OBD_FREE_PTR(cbdata);
2149 EXPORT_SYMBOL(ptlrpc_set_wait);
2152 * Helper fuction for request freeing.
2153 * Called when request count reached zero and request needs to be freed.
2154 * Removes request from all sorts of sending/replay lists it might be on,
2155 * frees network buffers if any are present.
2156 * If \a locked is set, that means caller is already holding import imp_lock
2157 * and so we no longer need to reobtain it (for certain lists manipulations)
2159 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2162 if (request == NULL) {
2167 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2168 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2169 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2170 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2171 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2172 LASSERTF(!request->rq_replay, "req %p\n", request);
2174 req_capsule_fini(&request->rq_pill);
2176 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2177 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2178 if (request->rq_import != NULL) {
2180 spin_lock(&request->rq_import->imp_lock);
2181 cfs_list_del_init(&request->rq_replay_list);
2183 spin_unlock(&request->rq_import->imp_lock);
2185 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2187 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2188 DEBUG_REQ(D_ERROR, request,
2189 "freeing request with nonzero refcount");
2193 if (request->rq_repbuf != NULL)
2194 sptlrpc_cli_free_repbuf(request);
2195 if (request->rq_export != NULL) {
2196 class_export_put(request->rq_export);
2197 request->rq_export = NULL;
2199 if (request->rq_import != NULL) {
2200 class_import_put(request->rq_import);
2201 request->rq_import = NULL;
2203 if (request->rq_bulk != NULL)
2204 ptlrpc_free_bulk_pin(request->rq_bulk);
2206 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2207 sptlrpc_cli_free_reqbuf(request);
2209 if (request->rq_cli_ctx)
2210 sptlrpc_req_put_ctx(request, !locked);
2212 if (request->rq_pool)
2213 __ptlrpc_free_req_to_pool(request);
2215 OBD_FREE(request, sizeof(*request));
2219 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2221 * Drop one request reference. Must be called with import imp_lock held.
2222 * When reference count drops to zero, reuqest is freed.
2224 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2226 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2227 (void)__ptlrpc_req_finished(request, 1);
2229 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2233 * Drops one reference count for request \a request.
2234 * \a locked set indicates that caller holds import imp_lock.
2235 * Frees the request whe reference count reaches zero.
2237 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2240 if (request == NULL)
2243 if (request == LP_POISON ||
2244 request->rq_reqmsg == LP_POISON) {
2245 CERROR("dereferencing freed request (bug 575)\n");
2250 DEBUG_REQ(D_INFO, request, "refcount now %u",
2251 cfs_atomic_read(&request->rq_refcount) - 1);
2253 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2254 __ptlrpc_free_req(request, locked);
2262 * Drops one reference count for a request.
2264 void ptlrpc_req_finished(struct ptlrpc_request *request)
2266 __ptlrpc_req_finished(request, 0);
2268 EXPORT_SYMBOL(ptlrpc_req_finished);
2271 * Returns xid of a \a request
2273 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2275 return request->rq_xid;
2277 EXPORT_SYMBOL(ptlrpc_req_xid);
2280 * Disengage the client's reply buffer from the network
2281 * NB does _NOT_ unregister any client-side bulk.
2282 * IDEMPOTENT, but _not_ safe against concurrent callers.
2283 * The request owner (i.e. the thread doing the I/O) must call...
2284 * Returns 0 on success or 1 if unregistering cannot be made.
2286 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2290 struct l_wait_info lwi;
2295 LASSERT(!cfs_in_interrupt());
2298 * Let's setup deadline for reply unlink.
2300 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2301 async && request->rq_reply_deadline == 0)
2302 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2305 * Nothing left to do.
2307 if (!ptlrpc_client_recv_or_unlink(request))
2310 LNetMDUnlink(request->rq_reply_md_h);
2313 * Let's check it once again.
2315 if (!ptlrpc_client_recv_or_unlink(request))
2319 * Move to "Unregistering" phase as reply was not unlinked yet.
2321 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2324 * Do not wait for unlink to finish.
2330 * We have to l_wait_event() whatever the result, to give liblustre
2331 * a chance to run reply_in_callback(), and to make sure we've
2332 * unlinked before returning a req to the pool.
2334 if (request->rq_set != NULL)
2335 wq = &request->rq_set->set_waitq;
2337 wq = &request->rq_reply_waitq;
2340 /* Network access will complete in finite time but the HUGE
2341 * timeout lets us CWARN for visibility of sluggish NALs */
2342 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2343 cfs_time_seconds(1), NULL, NULL);
2344 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2347 ptlrpc_rqphase_move(request, request->rq_next_phase);
2351 LASSERT(rc == -ETIMEDOUT);
2352 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2353 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2354 request->rq_must_unlink);
2358 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2361 * Iterates through replay_list on import and prunes
2362 * all requests have transno smaller than last_committed for the
2363 * import and don't have rq_replay set.
2364 * Since requests are sorted in transno order, stops when meetign first
2365 * transno bigger than last_committed.
2366 * caller must hold imp->imp_lock
2368 void ptlrpc_free_committed(struct obd_import *imp)
2370 cfs_list_t *tmp, *saved;
2371 struct ptlrpc_request *req;
2372 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2375 LASSERT(imp != NULL);
2377 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2380 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2381 imp->imp_generation == imp->imp_last_generation_checked) {
2382 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2383 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2387 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2388 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2389 imp->imp_generation);
2390 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2391 imp->imp_last_generation_checked = imp->imp_generation;
2393 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2394 req = cfs_list_entry(tmp, struct ptlrpc_request,
2397 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2398 LASSERT(req != last_req);
2401 if (req->rq_transno == 0) {
2402 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2405 if (req->rq_import_generation < imp->imp_generation) {
2406 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2410 if (req->rq_replay) {
2411 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2415 /* not yet committed */
2416 if (req->rq_transno > imp->imp_peer_committed_transno) {
2417 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2421 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2422 imp->imp_peer_committed_transno);
2424 spin_lock(&req->rq_lock);
2426 spin_unlock(&req->rq_lock);
2427 if (req->rq_commit_cb != NULL)
2428 req->rq_commit_cb(req);
2429 cfs_list_del_init(&req->rq_replay_list);
2430 __ptlrpc_req_finished(req, 1);
2437 void ptlrpc_cleanup_client(struct obd_import *imp)
2443 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2446 * Schedule previously sent request for resend.
2447 * For bulk requests we assign new xid (to avoid problems with
2448 * lost replies and therefore several transfers landing into same buffer
2449 * from different sending attempts).
2451 void ptlrpc_resend_req(struct ptlrpc_request *req)
2453 DEBUG_REQ(D_HA, req, "going to resend");
2454 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2455 req->rq_status = -EAGAIN;
2457 spin_lock(&req->rq_lock);
2459 req->rq_net_err = 0;
2460 req->rq_timedout = 0;
2462 __u64 old_xid = req->rq_xid;
2464 /* ensure previous bulk fails */
2465 req->rq_xid = ptlrpc_next_xid();
2466 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2467 old_xid, req->rq_xid);
2469 ptlrpc_client_wake_req(req);
2470 spin_unlock(&req->rq_lock);
2472 EXPORT_SYMBOL(ptlrpc_resend_req);
2474 /* XXX: this function and rq_status are currently unused */
2475 void ptlrpc_restart_req(struct ptlrpc_request *req)
2477 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2478 req->rq_status = -ERESTARTSYS;
2480 spin_lock(&req->rq_lock);
2481 req->rq_restart = 1;
2482 req->rq_timedout = 0;
2483 ptlrpc_client_wake_req(req);
2484 spin_unlock(&req->rq_lock);
2486 EXPORT_SYMBOL(ptlrpc_restart_req);
2489 * Grab additional reference on a request \a req
2491 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2494 cfs_atomic_inc(&req->rq_refcount);
2497 EXPORT_SYMBOL(ptlrpc_request_addref);
2500 * Add a request to import replay_list.
2501 * Must be called under imp_lock
2503 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2504 struct obd_import *imp)
2508 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2510 if (req->rq_transno == 0) {
2511 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2515 /* clear this for new requests that were resent as well
2516 as resent replayed requests. */
2517 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2519 /* don't re-add requests that have been replayed */
2520 if (!cfs_list_empty(&req->rq_replay_list))
2523 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2525 LASSERT(imp->imp_replayable);
2526 /* Balanced in ptlrpc_free_committed, usually. */
2527 ptlrpc_request_addref(req);
2528 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2529 struct ptlrpc_request *iter =
2530 cfs_list_entry(tmp, struct ptlrpc_request,
2533 /* We may have duplicate transnos if we create and then
2534 * open a file, or for closes retained if to match creating
2535 * opens, so use req->rq_xid as a secondary key.
2536 * (See bugs 684, 685, and 428.)
2537 * XXX no longer needed, but all opens need transnos!
2539 if (iter->rq_transno > req->rq_transno)
2542 if (iter->rq_transno == req->rq_transno) {
2543 LASSERT(iter->rq_xid != req->rq_xid);
2544 if (iter->rq_xid > req->rq_xid)
2548 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2552 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2554 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2557 * Send request and wait until it completes.
2558 * Returns request processing status.
2560 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2562 struct ptlrpc_request_set *set;
2566 LASSERT(req->rq_set == NULL);
2567 LASSERT(!req->rq_receiving_reply);
2569 set = ptlrpc_prep_set();
2571 CERROR("Unable to allocate ptlrpc set.");
2575 /* for distributed debugging */
2576 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2578 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2579 ptlrpc_request_addref(req);
2580 ptlrpc_set_add_req(set, req);
2581 rc = ptlrpc_set_wait(set);
2582 ptlrpc_set_destroy(set);
2586 EXPORT_SYMBOL(ptlrpc_queue_wait);
2588 struct ptlrpc_replay_async_args {
2590 int praa_old_status;
2594 * Callback used for replayed requests reply processing.
2595 * In case of succesful reply calls registeresd request replay callback.
2596 * In case of error restart replay process.
2598 static int ptlrpc_replay_interpret(const struct lu_env *env,
2599 struct ptlrpc_request *req,
2600 void * data, int rc)
2602 struct ptlrpc_replay_async_args *aa = data;
2603 struct obd_import *imp = req->rq_import;
2606 cfs_atomic_dec(&imp->imp_replay_inflight);
2608 if (!ptlrpc_client_replied(req)) {
2609 CERROR("request replay timed out, restarting recovery\n");
2610 GOTO(out, rc = -ETIMEDOUT);
2613 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2614 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2615 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2616 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2618 /** VBR: check version failure */
2619 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2620 /** replay was failed due to version mismatch */
2621 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2622 spin_lock(&imp->imp_lock);
2623 imp->imp_vbr_failed = 1;
2624 imp->imp_no_lock_replay = 1;
2625 spin_unlock(&imp->imp_lock);
2626 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2628 /** The transno had better not change over replay. */
2629 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2630 lustre_msg_get_transno(req->rq_repmsg) ||
2631 lustre_msg_get_transno(req->rq_repmsg) == 0,
2633 lustre_msg_get_transno(req->rq_reqmsg),
2634 lustre_msg_get_transno(req->rq_repmsg));
2637 spin_lock(&imp->imp_lock);
2638 /** if replays by version then gap occur on server, no trust to locks */
2639 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2640 imp->imp_no_lock_replay = 1;
2641 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2642 spin_unlock(&imp->imp_lock);
2643 LASSERT(imp->imp_last_replay_transno);
2645 /* transaction number shouldn't be bigger than the latest replayed */
2646 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2647 DEBUG_REQ(D_ERROR, req,
2648 "Reported transno "LPU64" is bigger than the "
2649 "replayed one: "LPU64, req->rq_transno,
2650 lustre_msg_get_transno(req->rq_reqmsg));
2651 GOTO(out, rc = -EINVAL);
2654 DEBUG_REQ(D_HA, req, "got rep");
2656 /* let the callback do fixups, possibly including in the request */
2657 if (req->rq_replay_cb)
2658 req->rq_replay_cb(req);
2660 if (ptlrpc_client_replied(req) &&
2661 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2662 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2663 lustre_msg_get_status(req->rq_repmsg),
2664 aa->praa_old_status);
2666 /* Put it back for re-replay. */
2667 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2671 * Errors while replay can set transno to 0, but
2672 * imp_last_replay_transno shouldn't be set to 0 anyway
2674 if (req->rq_transno == 0)
2675 CERROR("Transno is 0 during replay!\n");
2677 /* continue with recovery */
2678 rc = ptlrpc_import_recovery_state_machine(imp);
2680 req->rq_send_state = aa->praa_old_state;
2683 /* this replay failed, so restart recovery */
2684 ptlrpc_connect_import(imp);
2690 * Prepares and queues request for replay.
2691 * Adds it to ptlrpcd queue for actual sending.
2692 * Returns 0 on success.
2694 int ptlrpc_replay_req(struct ptlrpc_request *req)
2696 struct ptlrpc_replay_async_args *aa;
2699 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2701 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2702 aa = ptlrpc_req_async_args(req);
2703 memset(aa, 0, sizeof *aa);
2705 /* Prepare request to be resent with ptlrpcd */
2706 aa->praa_old_state = req->rq_send_state;
2707 req->rq_send_state = LUSTRE_IMP_REPLAY;
2708 req->rq_phase = RQ_PHASE_NEW;
2709 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2711 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2713 req->rq_interpret_reply = ptlrpc_replay_interpret;
2714 /* Readjust the timeout for current conditions */
2715 ptlrpc_at_set_req_timeout(req);
2717 /* Tell server the net_latency, so the server can calculate how long
2718 * it should wait for next replay */
2719 lustre_msg_set_service_time(req->rq_reqmsg,
2720 ptlrpc_at_get_net_latency(req));
2721 DEBUG_REQ(D_HA, req, "REPLAY");
2723 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2724 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2726 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2729 EXPORT_SYMBOL(ptlrpc_replay_req);
2732 * Aborts all in-flight request on import \a imp sending and delayed lists
2734 void ptlrpc_abort_inflight(struct obd_import *imp)
2736 cfs_list_t *tmp, *n;
2739 /* Make sure that no new requests get processed for this import.
2740 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2741 * this flag and then putting requests on sending_list or delayed_list.
2743 spin_lock(&imp->imp_lock);
2745 /* XXX locking? Maybe we should remove each request with the list
2746 * locked? Also, how do we know if the requests on the list are
2747 * being freed at this time?
2749 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2750 struct ptlrpc_request *req =
2751 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2753 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2755 spin_lock(&req->rq_lock);
2756 if (req->rq_import_generation < imp->imp_generation) {
2758 req->rq_status = -EIO;
2759 ptlrpc_client_wake_req(req);
2761 spin_unlock(&req->rq_lock);
2764 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2765 struct ptlrpc_request *req =
2766 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2768 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2770 spin_lock(&req->rq_lock);
2771 if (req->rq_import_generation < imp->imp_generation) {
2773 req->rq_status = -EIO;
2774 ptlrpc_client_wake_req(req);
2776 spin_unlock(&req->rq_lock);
2779 /* Last chance to free reqs left on the replay list, but we
2780 * will still leak reqs that haven't committed. */
2781 if (imp->imp_replayable)
2782 ptlrpc_free_committed(imp);
2784 spin_unlock(&imp->imp_lock);
2788 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2791 * Abort all uncompleted requests in request set \a set
2793 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2795 cfs_list_t *tmp, *pos;
2797 LASSERT(set != NULL);
2799 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2800 struct ptlrpc_request *req =
2801 cfs_list_entry(pos, struct ptlrpc_request,
2804 spin_lock(&req->rq_lock);
2805 if (req->rq_phase != RQ_PHASE_RPC) {
2806 spin_unlock(&req->rq_lock);
2811 req->rq_status = -EINTR;
2812 ptlrpc_client_wake_req(req);
2813 spin_unlock(&req->rq_lock);
2817 static __u64 ptlrpc_last_xid;
2818 static spinlock_t ptlrpc_last_xid_lock;
2821 * Initialize the XID for the node. This is common among all requests on
2822 * this node, and only requires the property that it is monotonically
2823 * increasing. It does not need to be sequential. Since this is also used
2824 * as the RDMA match bits, it is important that a single client NOT have
2825 * the same match bits for two different in-flight requests, hence we do
2826 * NOT want to have an XID per target or similar.
2828 * To avoid an unlikely collision between match bits after a client reboot
2829 * (which would deliver old data into the wrong RDMA buffer) initialize
2830 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2831 * If the time is clearly incorrect, we instead use a 62-bit random number.
2832 * In the worst case the random number will overflow 1M RPCs per second in
2833 * 9133 years, or permutations thereof.
2835 #define YEAR_2004 (1ULL << 30)
2836 void ptlrpc_init_xid(void)
2838 time_t now = cfs_time_current_sec();
2840 spin_lock_init(&ptlrpc_last_xid_lock);
2841 if (now < YEAR_2004) {
2842 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2843 ptlrpc_last_xid >>= 2;
2844 ptlrpc_last_xid |= (1ULL << 61);
2846 ptlrpc_last_xid = (__u64)now << 20;
2851 * Increase xid and returns resultng new value to the caller.
2853 __u64 ptlrpc_next_xid(void)
2856 spin_lock(&ptlrpc_last_xid_lock);
2857 tmp = ++ptlrpc_last_xid;
2858 spin_unlock(&ptlrpc_last_xid_lock);
2861 EXPORT_SYMBOL(ptlrpc_next_xid);
2864 * Get a glimpse at what next xid value might have been.
2865 * Returns possible next xid.
2867 __u64 ptlrpc_sample_next_xid(void)
2869 #if BITS_PER_LONG == 32
2870 /* need to avoid possible word tearing on 32-bit systems */
2872 spin_lock(&ptlrpc_last_xid_lock);
2873 tmp = ptlrpc_last_xid + 1;
2874 spin_unlock(&ptlrpc_last_xid_lock);
2877 /* No need to lock, since returned value is racy anyways */
2878 return ptlrpc_last_xid + 1;
2881 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2884 * Functions for operating ptlrpc workers.
2886 * A ptlrpc work is a function which will be running inside ptlrpc context.
2887 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2889 * 1. after a work is created, it can be used many times, that is:
2890 * handler = ptlrpcd_alloc_work();
2891 * ptlrpcd_queue_work();
2893 * queue it again when necessary:
2894 * ptlrpcd_queue_work();
2895 * ptlrpcd_destroy_work();
2896 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2897 * it will only be queued once in any time. Also as its name implies, it may
2898 * have delay before it really runs by ptlrpcd thread.
2900 struct ptlrpc_work_async_args {
2902 int (*cb)(const struct lu_env *, void *);
2906 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2908 static int work_interpreter(const struct lu_env *env,
2909 struct ptlrpc_request *req, void *data, int rc)
2911 struct ptlrpc_work_async_args *arg = data;
2913 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2914 LASSERT(arg->cb != NULL);
2916 return arg->cb(env, arg->cbdata);
2920 * Create a work for ptlrpc.
2922 void *ptlrpcd_alloc_work(struct obd_import *imp,
2923 int (*cb)(const struct lu_env *, void *), void *cbdata)
2925 struct ptlrpc_request *req = NULL;
2926 struct ptlrpc_work_async_args *args;
2932 RETURN(ERR_PTR(-EINVAL));
2934 /* copy some code from deprecated fakereq. */
2937 CERROR("ptlrpc: run out of memory!\n");
2938 RETURN(ERR_PTR(-ENOMEM));
2941 req->rq_send_state = LUSTRE_IMP_FULL;
2942 req->rq_type = PTL_RPC_MSG_REQUEST;
2943 req->rq_import = class_import_get(imp);
2944 req->rq_export = NULL;
2945 req->rq_interpret_reply = work_interpreter;
2946 /* don't want reply */
2947 req->rq_receiving_reply = 0;
2948 req->rq_must_unlink = 0;
2949 req->rq_no_delay = req->rq_no_resend = 1;
2951 spin_lock_init(&req->rq_lock);
2952 CFS_INIT_LIST_HEAD(&req->rq_list);
2953 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
2954 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
2955 CFS_INIT_LIST_HEAD(&req->rq_history_list);
2956 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
2957 cfs_waitq_init(&req->rq_reply_waitq);
2958 cfs_waitq_init(&req->rq_set_waitq);
2959 cfs_atomic_set(&req->rq_refcount, 1);
2961 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
2962 args = ptlrpc_req_async_args(req);
2963 args->magic = PTLRPC_WORK_MAGIC;
2965 args->cbdata = cbdata;
2969 EXPORT_SYMBOL(ptlrpcd_alloc_work);
2971 void ptlrpcd_destroy_work(void *handler)
2973 struct ptlrpc_request *req = handler;
2976 ptlrpc_req_finished(req);
2978 EXPORT_SYMBOL(ptlrpcd_destroy_work);
2980 int ptlrpcd_queue_work(void *handler)
2982 struct ptlrpc_request *req = handler;
2985 * Check if the req is already being queued.
2987 * Here comes a trick: it lacks a way of checking if a req is being
2988 * processed reliably in ptlrpc. Here I have to use refcount of req
2989 * for this purpose. This is okay because the caller should use this
2990 * req as opaque data. - Jinshan
2992 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
2993 if (cfs_atomic_read(&req->rq_refcount) > 1)
2996 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
2997 cfs_atomic_dec(&req->rq_refcount);
3001 /* re-initialize the req */
3002 req->rq_timeout = obd_timeout;
3003 req->rq_sent = cfs_time_current_sec();
3004 req->rq_deadline = req->rq_sent + req->rq_timeout;
3005 req->rq_reply_deadline = req->rq_deadline;
3006 req->rq_phase = RQ_PHASE_INTERPRET;
3007 req->rq_next_phase = RQ_PHASE_COMPLETE;
3008 req->rq_xid = ptlrpc_next_xid();
3009 req->rq_import_generation = req->rq_import->imp_generation;
3011 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3014 EXPORT_SYMBOL(ptlrpcd_queue_work);