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 on the sender.
102 * Returns pointer to the descriptor or NULL on error.
104 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned npages, unsigned max_brw,
105 unsigned type, unsigned portal)
107 struct ptlrpc_bulk_desc *desc;
110 OBD_ALLOC(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[npages]));
114 spin_lock_init(&desc->bd_lock);
115 cfs_waitq_init(&desc->bd_waitq);
116 desc->bd_max_iov = npages;
117 desc->bd_iov_count = 0;
118 desc->bd_portal = portal;
119 desc->bd_type = type;
120 desc->bd_md_count = 0;
121 LASSERT(max_brw > 0);
122 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
123 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
124 * node. Negotiated ocd_brw_size will always be <= this number. */
125 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
126 LNetInvalidateHandle(&desc->bd_mds[i]);
132 * Prepare bulk descriptor for specified outgoing request \a req that
133 * can fit \a npages * pages. \a type is bulk type. \a portal is where
134 * the bulk to be sent. Used on client-side.
135 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
138 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
139 unsigned npages, unsigned max_brw,
140 unsigned type, unsigned portal)
142 struct obd_import *imp = req->rq_import;
143 struct ptlrpc_bulk_desc *desc;
146 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
147 desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
151 desc->bd_import_generation = req->rq_import_generation;
152 desc->bd_import = class_import_get(imp);
155 desc->bd_cbid.cbid_fn = client_bulk_callback;
156 desc->bd_cbid.cbid_arg = desc;
158 /* This makes req own desc, and free it when she frees herself */
163 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
166 * Add a page \a page to the bulk descriptor \a desc.
167 * Data to transfer in the page starts at offset \a pageoffset and
168 * amount of data to transfer from the page is \a len
170 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
171 cfs_page_t *page, int pageoffset, int len, int pin)
173 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
174 LASSERT(page != NULL);
175 LASSERT(pageoffset >= 0);
177 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
184 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
186 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
189 * Uninitialize and free bulk descriptor \a desc.
190 * Works on bulk descriptors both from server and client side.
192 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin)
197 LASSERT(desc != NULL);
198 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
199 LASSERT(desc->bd_md_count == 0); /* network hands off */
200 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
202 sptlrpc_enc_pool_put_pages(desc);
205 class_export_put(desc->bd_export);
207 class_import_put(desc->bd_import);
210 for (i = 0; i < desc->bd_iov_count ; i++)
211 cfs_page_unpin(desc->bd_iov[i].kiov_page);
214 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
215 bd_iov[desc->bd_max_iov]));
218 EXPORT_SYMBOL(__ptlrpc_free_bulk);
221 * Set server timelimit for this req, i.e. how long are we willing to wait
222 * for reply before timing out this request.
224 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
230 LASSERT(req->rq_import);
233 /* non-AT settings */
235 * \a imp_server_timeout means this is reverse import and
236 * we send (currently only) ASTs to the client and cannot afford
237 * to wait too long for the reply, otherwise the other client
238 * (because of which we are sending this request) would
239 * timeout waiting for us
241 req->rq_timeout = req->rq_import->imp_server_timeout ?
242 obd_timeout / 2 : obd_timeout;
244 at = &req->rq_import->imp_at;
245 idx = import_at_get_index(req->rq_import,
246 req->rq_request_portal);
247 serv_est = at_get(&at->iat_service_estimate[idx]);
248 req->rq_timeout = at_est2timeout(serv_est);
250 /* We could get even fancier here, using history to predict increased
253 /* Let the server know what this RPC timeout is by putting it in the
255 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
257 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
259 /* Adjust max service estimate based on server value */
260 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
261 unsigned int serv_est)
267 LASSERT(req->rq_import);
268 at = &req->rq_import->imp_at;
270 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
271 /* max service estimates are tracked on the server side,
272 so just keep minimal history here */
273 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
275 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
276 "has changed from %d to %d\n",
277 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
278 oldse, at_get(&at->iat_service_estimate[idx]));
281 /* Expected network latency per remote node (secs) */
282 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
284 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
287 /* Adjust expected network latency */
288 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
289 unsigned int service_time)
291 unsigned int nl, oldnl;
293 time_t now = cfs_time_current_sec();
295 LASSERT(req->rq_import);
296 at = &req->rq_import->imp_at;
298 /* Network latency is total time less server processing time */
299 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
300 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
301 CWARN("Reported service time %u > total measured time "
302 CFS_DURATION_T"\n", service_time,
303 cfs_time_sub(now, req->rq_sent));
305 oldnl = at_measured(&at->iat_net_latency, nl);
307 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
308 "has changed from %d to %d\n",
309 req->rq_import->imp_obd->obd_name,
311 &req->rq_import->imp_connection->c_remote_uuid),
312 oldnl, at_get(&at->iat_net_latency));
315 static int unpack_reply(struct ptlrpc_request *req)
319 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
320 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
322 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
327 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
329 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
336 * Handle an early reply message, called with the rq_lock held.
337 * If anything goes wrong just ignore it - same as if it never happened
339 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
341 struct ptlrpc_request *early_req;
347 spin_unlock(&req->rq_lock);
349 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
351 spin_lock(&req->rq_lock);
355 rc = unpack_reply(early_req);
357 /* Expecting to increase the service time estimate here */
358 ptlrpc_at_adj_service(req,
359 lustre_msg_get_timeout(early_req->rq_repmsg));
360 ptlrpc_at_adj_net_latency(req,
361 lustre_msg_get_service_time(early_req->rq_repmsg));
364 sptlrpc_cli_finish_early_reply(early_req);
367 spin_lock(&req->rq_lock);
371 /* Adjust the local timeout for this req */
372 ptlrpc_at_set_req_timeout(req);
374 spin_lock(&req->rq_lock);
375 olddl = req->rq_deadline;
376 /* server assumes it now has rq_timeout from when it sent the
377 * early reply, so client should give it at least that long. */
378 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
379 ptlrpc_at_get_net_latency(req);
381 DEBUG_REQ(D_ADAPTTO, req,
382 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
383 "("CFS_DURATION_T"s)", req->rq_early_count,
384 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
385 cfs_time_sub(req->rq_deadline, olddl));
391 * Wind down request pool \a pool.
392 * Frees all requests from the pool too
394 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
397 struct ptlrpc_request *req;
399 LASSERT(pool != NULL);
401 spin_lock(&pool->prp_lock);
402 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
403 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
404 cfs_list_del(&req->rq_list);
405 LASSERT(req->rq_reqbuf);
406 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
407 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
408 OBD_FREE(req, sizeof(*req));
410 spin_unlock(&pool->prp_lock);
411 OBD_FREE(pool, sizeof(*pool));
413 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
416 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
418 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
423 while (size < pool->prp_rq_size)
426 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
427 size == pool->prp_rq_size,
428 "Trying to change pool size with nonempty pool "
429 "from %d to %d bytes\n", pool->prp_rq_size, size);
431 spin_lock(&pool->prp_lock);
432 pool->prp_rq_size = size;
433 for (i = 0; i < num_rq; i++) {
434 struct ptlrpc_request *req;
435 struct lustre_msg *msg;
437 spin_unlock(&pool->prp_lock);
438 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
441 OBD_ALLOC_LARGE(msg, size);
443 OBD_FREE(req, sizeof(struct ptlrpc_request));
446 req->rq_reqbuf = msg;
447 req->rq_reqbuf_len = size;
449 spin_lock(&pool->prp_lock);
450 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
452 spin_unlock(&pool->prp_lock);
455 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
458 * Create and initialize new request pool with given attributes:
459 * \a num_rq - initial number of requests to create for the pool
460 * \a msgsize - maximum message size possible for requests in thid pool
461 * \a populate_pool - function to be called when more requests need to be added
463 * Returns pointer to newly created pool or NULL on error.
465 struct ptlrpc_request_pool *
466 ptlrpc_init_rq_pool(int num_rq, int msgsize,
467 void (*populate_pool)(struct ptlrpc_request_pool *, int))
469 struct ptlrpc_request_pool *pool;
471 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
475 /* Request next power of two for the allocation, because internally
476 kernel would do exactly this */
478 spin_lock_init(&pool->prp_lock);
479 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
480 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
481 pool->prp_populate = populate_pool;
483 populate_pool(pool, num_rq);
485 if (cfs_list_empty(&pool->prp_req_list)) {
486 /* have not allocated a single request for the pool */
487 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
492 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
495 * Fetches one request from pool \a pool
497 static struct ptlrpc_request *
498 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
500 struct ptlrpc_request *request;
501 struct lustre_msg *reqbuf;
506 spin_lock(&pool->prp_lock);
508 /* See if we have anything in a pool, and bail out if nothing,
509 * in writeout path, where this matters, this is safe to do, because
510 * nothing is lost in this case, and when some in-flight requests
511 * complete, this code will be called again. */
512 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
513 spin_unlock(&pool->prp_lock);
517 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
519 cfs_list_del_init(&request->rq_list);
520 spin_unlock(&pool->prp_lock);
522 LASSERT(request->rq_reqbuf);
523 LASSERT(request->rq_pool);
525 reqbuf = request->rq_reqbuf;
526 memset(request, 0, sizeof(*request));
527 request->rq_reqbuf = reqbuf;
528 request->rq_reqbuf_len = pool->prp_rq_size;
529 request->rq_pool = pool;
535 * Returns freed \a request to pool.
537 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
539 struct ptlrpc_request_pool *pool = request->rq_pool;
541 spin_lock(&pool->prp_lock);
542 LASSERT(cfs_list_empty(&request->rq_list));
543 LASSERT(!request->rq_receiving_reply);
544 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
545 spin_unlock(&pool->prp_lock);
548 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
549 __u32 version, int opcode,
550 int count, __u32 *lengths, char **bufs,
551 struct ptlrpc_cli_ctx *ctx)
553 struct obd_import *imp = request->rq_import;
558 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
560 rc = sptlrpc_req_get_ctx(request);
565 sptlrpc_req_set_flavor(request, opcode);
567 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
570 LASSERT(!request->rq_pool);
574 lustre_msg_add_version(request->rq_reqmsg, version);
575 request->rq_send_state = LUSTRE_IMP_FULL;
576 request->rq_type = PTL_RPC_MSG_REQUEST;
577 request->rq_export = NULL;
579 request->rq_req_cbid.cbid_fn = request_out_callback;
580 request->rq_req_cbid.cbid_arg = request;
582 request->rq_reply_cbid.cbid_fn = reply_in_callback;
583 request->rq_reply_cbid.cbid_arg = request;
585 request->rq_reply_deadline = 0;
586 request->rq_phase = RQ_PHASE_NEW;
587 request->rq_next_phase = RQ_PHASE_UNDEFINED;
589 request->rq_request_portal = imp->imp_client->cli_request_portal;
590 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
592 ptlrpc_at_set_req_timeout(request);
594 spin_lock_init(&request->rq_lock);
595 CFS_INIT_LIST_HEAD(&request->rq_list);
596 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
597 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
598 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
599 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
600 CFS_INIT_LIST_HEAD(&request->rq_history_list);
601 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
602 cfs_waitq_init(&request->rq_reply_waitq);
603 cfs_waitq_init(&request->rq_set_waitq);
604 request->rq_xid = ptlrpc_next_xid();
605 cfs_atomic_set(&request->rq_refcount, 1);
607 lustre_msg_set_opc(request->rq_reqmsg, opcode);
611 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
613 class_import_put(imp);
617 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
618 __u32 version, int opcode, char **bufs,
619 struct ptlrpc_cli_ctx *ctx)
623 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
624 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
625 request->rq_pill.rc_area[RCL_CLIENT],
628 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
631 * Pack request buffers for network transfer, performing necessary encryption
632 * steps if necessary.
634 int ptlrpc_request_pack(struct ptlrpc_request *request,
635 __u32 version, int opcode)
638 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
642 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
643 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
644 * have to send old ptlrpc_body to keep interoprability with these
647 * Only three kinds of server->client RPCs so far:
652 * XXX This should be removed whenever we drop the interoprability with
653 * the these old clients.
655 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
656 opcode == LDLM_GL_CALLBACK)
657 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
658 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
662 EXPORT_SYMBOL(ptlrpc_request_pack);
665 * Helper function to allocate new request on import \a imp
666 * and possibly using existing request from pool \a pool if provided.
667 * Returns allocated request structure with import field filled or
671 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
672 struct ptlrpc_request_pool *pool)
674 struct ptlrpc_request *request = NULL;
677 request = ptlrpc_prep_req_from_pool(pool);
680 OBD_ALLOC_PTR(request);
683 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
684 LASSERT(imp != LP_POISON);
685 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
687 LASSERT(imp->imp_client != LP_POISON);
689 request->rq_import = class_import_get(imp);
691 CERROR("request allocation out of memory\n");
698 * Helper function for creating a request.
699 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
700 * buffer structures according to capsule template \a format.
701 * Returns allocated request structure pointer or NULL on error.
703 static struct ptlrpc_request *
704 ptlrpc_request_alloc_internal(struct obd_import *imp,
705 struct ptlrpc_request_pool * pool,
706 const struct req_format *format)
708 struct ptlrpc_request *request;
710 request = __ptlrpc_request_alloc(imp, pool);
714 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
715 req_capsule_set(&request->rq_pill, format);
720 * Allocate new request structure for import \a imp and initialize its
721 * buffer structure according to capsule template \a format.
723 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
724 const struct req_format *format)
726 return ptlrpc_request_alloc_internal(imp, NULL, format);
728 EXPORT_SYMBOL(ptlrpc_request_alloc);
731 * Allocate new request structure for import \a imp from pool \a pool and
732 * initialize its buffer structure according to capsule template \a format.
734 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
735 struct ptlrpc_request_pool * pool,
736 const struct req_format *format)
738 return ptlrpc_request_alloc_internal(imp, pool, format);
740 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
743 * For requests not from pool, free memory of the request structure.
744 * For requests obtained from a pool earlier, return request back to pool.
746 void ptlrpc_request_free(struct ptlrpc_request *request)
748 if (request->rq_pool)
749 __ptlrpc_free_req_to_pool(request);
751 OBD_FREE_PTR(request);
753 EXPORT_SYMBOL(ptlrpc_request_free);
756 * Allocate new request for operatione \a opcode and immediatelly pack it for
758 * Only used for simple requests like OBD_PING where the only important
759 * part of the request is operation itself.
760 * Returns allocated request or NULL on error.
762 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
763 const struct req_format *format,
764 __u32 version, int opcode)
766 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
770 rc = ptlrpc_request_pack(req, version, opcode);
772 ptlrpc_request_free(req);
778 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
781 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
782 * for operation \a opcode. Request would contain \a count buffers.
783 * Sizes of buffers are described in array \a lengths and buffers themselves
784 * are provided by a pointer \a bufs.
785 * Returns prepared request structure pointer or NULL on error.
787 struct ptlrpc_request *
788 ptlrpc_prep_req_pool(struct obd_import *imp,
789 __u32 version, int opcode,
790 int count, __u32 *lengths, char **bufs,
791 struct ptlrpc_request_pool *pool)
793 struct ptlrpc_request *request;
796 request = __ptlrpc_request_alloc(imp, pool);
800 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
801 lengths, bufs, NULL);
803 ptlrpc_request_free(request);
808 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
811 * Same as ptlrpc_prep_req_pool, but without pool
813 struct ptlrpc_request *
814 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
815 __u32 *lengths, char **bufs)
817 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
820 EXPORT_SYMBOL(ptlrpc_prep_req);
823 * Allocate and initialize new request set structure.
824 * Returns a pointer to the newly allocated set structure or NULL on error.
826 struct ptlrpc_request_set *ptlrpc_prep_set(void)
828 struct ptlrpc_request_set *set;
831 OBD_ALLOC(set, sizeof *set);
834 cfs_atomic_set(&set->set_refcount, 1);
835 CFS_INIT_LIST_HEAD(&set->set_requests);
836 cfs_waitq_init(&set->set_waitq);
837 cfs_atomic_set(&set->set_new_count, 0);
838 cfs_atomic_set(&set->set_remaining, 0);
839 spin_lock_init(&set->set_new_req_lock);
840 CFS_INIT_LIST_HEAD(&set->set_new_requests);
841 CFS_INIT_LIST_HEAD(&set->set_cblist);
842 set->set_max_inflight = UINT_MAX;
843 set->set_producer = NULL;
844 set->set_producer_arg = NULL;
849 EXPORT_SYMBOL(ptlrpc_prep_set);
852 * Allocate and initialize new request set structure with flow control
853 * extension. This extension allows to control the number of requests in-flight
854 * for the whole set. A callback function to generate requests must be provided
855 * and the request set will keep the number of requests sent over the wire to
857 * Returns a pointer to the newly allocated set structure or NULL on error.
859 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
863 struct ptlrpc_request_set *set;
865 set = ptlrpc_prep_set();
869 set->set_max_inflight = max;
870 set->set_producer = func;
871 set->set_producer_arg = arg;
875 EXPORT_SYMBOL(ptlrpc_prep_fcset);
878 * Wind down and free request set structure previously allocated with
880 * Ensures that all requests on the set have completed and removes
881 * all requests from the request list in a set.
882 * If any unsent request happen to be on the list, pretends that they got
883 * an error in flight and calls their completion handler.
885 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
893 /* Requests on the set should either all be completed, or all be new */
894 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
895 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
896 cfs_list_for_each (tmp, &set->set_requests) {
897 struct ptlrpc_request *req =
898 cfs_list_entry(tmp, struct ptlrpc_request,
901 LASSERT(req->rq_phase == expected_phase);
905 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
906 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
907 cfs_atomic_read(&set->set_remaining), n);
909 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
910 struct ptlrpc_request *req =
911 cfs_list_entry(tmp, struct ptlrpc_request,
913 cfs_list_del_init(&req->rq_set_chain);
915 LASSERT(req->rq_phase == expected_phase);
917 if (req->rq_phase == RQ_PHASE_NEW) {
918 ptlrpc_req_interpret(NULL, req, -EBADR);
919 cfs_atomic_dec(&set->set_remaining);
922 spin_lock(&req->rq_lock);
924 req->rq_invalid_rqset = 0;
925 spin_unlock(&req->rq_lock);
927 ptlrpc_req_finished (req);
930 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
932 ptlrpc_reqset_put(set);
935 EXPORT_SYMBOL(ptlrpc_set_destroy);
938 * Add a callback function \a fn to the set.
939 * This function would be called when all requests on this set are completed.
940 * The function will be passed \a data argument.
942 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
943 set_interpreter_func fn, void *data)
945 struct ptlrpc_set_cbdata *cbdata;
947 OBD_ALLOC_PTR(cbdata);
951 cbdata->psc_interpret = fn;
952 cbdata->psc_data = data;
953 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
957 EXPORT_SYMBOL(ptlrpc_set_add_cb);
960 * Add a new request to the general purpose request set.
961 * Assumes request reference from the caller.
963 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
964 struct ptlrpc_request *req)
966 LASSERT(cfs_list_empty(&req->rq_set_chain));
968 /* The set takes over the caller's request reference */
969 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
971 cfs_atomic_inc(&set->set_remaining);
972 req->rq_queued_time = cfs_time_current();
974 if (req->rq_reqmsg != NULL)
975 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
977 if (set->set_producer != NULL)
978 /* If the request set has a producer callback, the RPC must be
979 * sent straight away */
980 ptlrpc_send_new_req(req);
982 EXPORT_SYMBOL(ptlrpc_set_add_req);
985 * Add a request to a request with dedicated server thread
986 * and wake the thread to make any necessary processing.
987 * Currently only used for ptlrpcd.
989 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
990 struct ptlrpc_request *req)
992 struct ptlrpc_request_set *set = pc->pc_set;
995 LASSERT(req->rq_set == NULL);
996 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
998 spin_lock(&set->set_new_req_lock);
1000 * The set takes over the caller's request reference.
1003 req->rq_queued_time = cfs_time_current();
1004 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1005 count = cfs_atomic_inc_return(&set->set_new_count);
1006 spin_unlock(&set->set_new_req_lock);
1008 /* Only need to call wakeup once for the first entry. */
1010 cfs_waitq_signal(&set->set_waitq);
1012 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1013 * guarantee the async RPC can be processed ASAP, we have
1014 * no other better choice. It maybe fixed in future. */
1015 for (i = 0; i < pc->pc_npartners; i++)
1016 cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
1019 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1022 * Based on the current state of the import, determine if the request
1023 * can be sent, is an error, or should be delayed.
1025 * Returns true if this request should be delayed. If false, and
1026 * *status is set, then the request can not be sent and *status is the
1027 * error code. If false and status is 0, then request can be sent.
1029 * The imp->imp_lock must be held.
1031 static int ptlrpc_import_delay_req(struct obd_import *imp,
1032 struct ptlrpc_request *req, int *status)
1037 LASSERT (status != NULL);
1040 if (req->rq_ctx_init || req->rq_ctx_fini) {
1041 /* always allow ctx init/fini rpc go through */
1042 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1043 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1045 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1046 /* pings may safely race with umount */
1047 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1048 D_HA : D_ERROR, req, "IMP_CLOSED ");
1050 } else if (ptlrpc_send_limit_expired(req)) {
1051 /* probably doesn't need to be a D_ERROR after initial testing */
1052 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1054 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1055 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1056 /* allow CONNECT even if import is invalid */ ;
1057 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1058 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1061 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1062 if (!imp->imp_deactive)
1063 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1064 *status = -ESHUTDOWN; /* bz 12940 */
1065 } else if (req->rq_import_generation != imp->imp_generation) {
1066 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1068 } else if (req->rq_send_state != imp->imp_state) {
1069 /* invalidate in progress - any requests should be drop */
1070 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1071 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1073 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1074 *status = -EWOULDBLOCK;
1075 } else if (req->rq_allow_replay &&
1076 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1077 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1078 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1079 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1080 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1090 * Decide if the eror message regarding provided request \a req
1091 * should be printed to the console or not.
1092 * Makes it's decision on request status and other properties.
1093 * Returns 1 to print error on the system console or 0 if not.
1095 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1100 LASSERT(req->rq_reqmsg != NULL);
1101 opc = lustre_msg_get_opc(req->rq_reqmsg);
1103 /* Suppress particular reconnect errors which are to be expected. No
1104 * errors are suppressed for the initial connection on an import */
1105 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1106 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1108 /* Suppress timed out reconnect requests */
1109 if (req->rq_timedout)
1112 /* Suppress unavailable/again reconnect requests */
1113 err = lustre_msg_get_status(req->rq_repmsg);
1114 if (err == -ENODEV || err == -EAGAIN)
1122 * Check request processing status.
1123 * Returns the status.
1125 static int ptlrpc_check_status(struct ptlrpc_request *req)
1130 err = lustre_msg_get_status(req->rq_repmsg);
1131 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1132 struct obd_import *imp = req->rq_import;
1133 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1134 if (ptlrpc_console_allow(req))
1135 LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s,"
1136 " operation %s failed with %d.\n",
1137 imp->imp_obd->obd_name,
1139 imp->imp_connection->c_peer.nid),
1140 ll_opcode2str(opc), err);
1141 RETURN(err < 0 ? err : -EINVAL);
1145 DEBUG_REQ(D_INFO, req, "status is %d", err);
1146 } else if (err > 0) {
1147 /* XXX: translate this error from net to host */
1148 DEBUG_REQ(D_INFO, req, "status is %d", err);
1155 * save pre-versions of objects into request for replay.
1156 * Versions are obtained from server reply.
1159 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1161 struct lustre_msg *repmsg = req->rq_repmsg;
1162 struct lustre_msg *reqmsg = req->rq_reqmsg;
1163 __u64 *versions = lustre_msg_get_versions(repmsg);
1166 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1170 lustre_msg_set_versions(reqmsg, versions);
1171 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1172 versions[0], versions[1]);
1178 * Callback function called when client receives RPC reply for \a req.
1179 * Returns 0 on success or error code.
1180 * The return alue would be assigned to req->rq_status by the caller
1181 * as request processing status.
1182 * This function also decides if the request needs to be saved for later replay.
1184 static int after_reply(struct ptlrpc_request *req)
1186 struct obd_import *imp = req->rq_import;
1187 struct obd_device *obd = req->rq_import->imp_obd;
1189 struct timeval work_start;
1193 LASSERT(obd != NULL);
1194 /* repbuf must be unlinked */
1195 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1197 if (req->rq_reply_truncate) {
1198 if (ptlrpc_no_resend(req)) {
1199 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1200 " expected: %d, actual size: %d",
1201 req->rq_nob_received, req->rq_repbuf_len);
1205 sptlrpc_cli_free_repbuf(req);
1206 /* Pass the required reply buffer size (include
1207 * space for early reply).
1208 * NB: no need to roundup because alloc_repbuf
1209 * will roundup it */
1210 req->rq_replen = req->rq_nob_received;
1211 req->rq_nob_received = 0;
1217 * NB Until this point, the whole of the incoming message,
1218 * including buflens, status etc is in the sender's byte order.
1220 rc = sptlrpc_cli_unwrap_reply(req);
1222 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1226 /* retry indefinitely on EINPROGRESS */
1227 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1228 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1229 time_t now = cfs_time_current_sec();
1231 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1233 req->rq_nr_resend++;
1235 /* allocate new xid to avoid reply reconstruction */
1236 if (!req->rq_bulk) {
1237 /* new xid is already allocated for bulk in
1238 * ptlrpc_check_set() */
1239 req->rq_xid = ptlrpc_next_xid();
1240 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1241 "resend on EINPROGRESS");
1244 /* Readjust the timeout for current conditions */
1245 ptlrpc_at_set_req_timeout(req);
1246 /* delay resend to give a chance to the server to get ready.
1247 * The delay is increased by 1s on every resend and is capped to
1248 * the current request timeout (i.e. obd_timeout if AT is off,
1249 * or AT service time x 125% + 5s, see at_est2timeout) */
1250 if (req->rq_nr_resend > req->rq_timeout)
1251 req->rq_sent = now + req->rq_timeout;
1253 req->rq_sent = now + req->rq_nr_resend;
1257 * Security layer unwrap might ask resend this request.
1262 rc = unpack_reply(req);
1266 cfs_gettimeofday(&work_start);
1267 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1268 if (obd->obd_svc_stats != NULL) {
1269 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1271 ptlrpc_lprocfs_rpc_sent(req, timediff);
1274 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1275 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1276 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1277 lustre_msg_get_type(req->rq_repmsg));
1281 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1282 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1283 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1284 ptlrpc_at_adj_net_latency(req,
1285 lustre_msg_get_service_time(req->rq_repmsg));
1287 rc = ptlrpc_check_status(req);
1288 imp->imp_connect_error = rc;
1292 * Either we've been evicted, or the server has failed for
1293 * some reason. Try to reconnect, and if that fails, punt to
1296 if (ll_rpc_recoverable_error(rc)) {
1297 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1298 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1301 ptlrpc_request_handle_notconn(req);
1306 * Let's look if server sent slv. Do it only for RPC with
1309 ldlm_cli_update_pool(req);
1313 * Store transno in reqmsg for replay.
1315 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1316 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1317 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1320 if (imp->imp_replayable) {
1321 spin_lock(&imp->imp_lock);
1323 * No point in adding already-committed requests to the replay
1324 * list, we will just remove them immediately. b=9829
1326 if (req->rq_transno != 0 &&
1328 lustre_msg_get_last_committed(req->rq_repmsg) ||
1330 /** version recovery */
1331 ptlrpc_save_versions(req);
1332 ptlrpc_retain_replayable_request(req, imp);
1333 } else if (req->rq_commit_cb != NULL) {
1334 spin_unlock(&imp->imp_lock);
1335 req->rq_commit_cb(req);
1336 spin_lock(&imp->imp_lock);
1340 * Replay-enabled imports return commit-status information.
1342 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1343 imp->imp_peer_committed_transno =
1344 lustre_msg_get_last_committed(req->rq_repmsg);
1347 ptlrpc_free_committed(imp);
1349 if (!cfs_list_empty(&imp->imp_replay_list)) {
1350 struct ptlrpc_request *last;
1352 last = cfs_list_entry(imp->imp_replay_list.prev,
1353 struct ptlrpc_request,
1356 * Requests with rq_replay stay on the list even if no
1357 * commit is expected.
1359 if (last->rq_transno > imp->imp_peer_committed_transno)
1360 ptlrpc_pinger_commit_expected(imp);
1363 spin_unlock(&imp->imp_lock);
1370 * Helper function to send request \a req over the network for the first time
1371 * Also adjusts request phase.
1372 * Returns 0 on success or error code.
1374 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1376 struct obd_import *imp = req->rq_import;
1380 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1381 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1382 (!req->rq_generation_set ||
1383 req->rq_import_generation == imp->imp_generation))
1386 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1388 spin_lock(&imp->imp_lock);
1390 if (!req->rq_generation_set)
1391 req->rq_import_generation = imp->imp_generation;
1393 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1394 spin_lock(&req->rq_lock);
1395 req->rq_waiting = 1;
1396 spin_unlock(&req->rq_lock);
1398 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1399 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1400 ptlrpc_import_state_name(req->rq_send_state),
1401 ptlrpc_import_state_name(imp->imp_state));
1402 LASSERT(cfs_list_empty(&req->rq_list));
1403 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1404 cfs_atomic_inc(&req->rq_import->imp_inflight);
1405 spin_unlock(&imp->imp_lock);
1410 spin_unlock(&imp->imp_lock);
1411 req->rq_status = rc;
1412 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1416 LASSERT(cfs_list_empty(&req->rq_list));
1417 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1418 cfs_atomic_inc(&req->rq_import->imp_inflight);
1419 spin_unlock(&imp->imp_lock);
1421 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1423 rc = sptlrpc_req_refresh_ctx(req, -1);
1426 req->rq_status = rc;
1429 req->rq_wait_ctx = 1;
1434 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1435 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1436 imp->imp_obd->obd_uuid.uuid,
1437 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1438 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1439 lustre_msg_get_opc(req->rq_reqmsg));
1441 rc = ptl_send_rpc(req, 0);
1443 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1444 req->rq_net_err = 1;
1450 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1455 LASSERT(set->set_producer != NULL);
1457 remaining = cfs_atomic_read(&set->set_remaining);
1459 /* populate the ->set_requests list with requests until we
1460 * reach the maximum number of RPCs in flight for this set */
1461 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1462 rc = set->set_producer(set, set->set_producer_arg);
1463 if (rc == -ENOENT) {
1464 /* no more RPC to produce */
1465 set->set_producer = NULL;
1466 set->set_producer_arg = NULL;
1471 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1475 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1476 * and no more replies are expected.
1477 * (it is possible to get less replies than requests sent e.g. due to timed out
1478 * requests or requests that we had trouble to send out)
1480 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1482 cfs_list_t *tmp, *next;
1483 int force_timer_recalc = 0;
1486 if (cfs_atomic_read(&set->set_remaining) == 0)
1489 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1490 struct ptlrpc_request *req =
1491 cfs_list_entry(tmp, struct ptlrpc_request,
1493 struct obd_import *imp = req->rq_import;
1494 int unregistered = 0;
1497 if (req->rq_phase == RQ_PHASE_NEW &&
1498 ptlrpc_send_new_req(req)) {
1499 force_timer_recalc = 1;
1502 /* delayed send - skip */
1503 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1506 /* delayed resend - skip */
1507 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1508 req->rq_sent > cfs_time_current_sec())
1511 if (!(req->rq_phase == RQ_PHASE_RPC ||
1512 req->rq_phase == RQ_PHASE_BULK ||
1513 req->rq_phase == RQ_PHASE_INTERPRET ||
1514 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1515 req->rq_phase == RQ_PHASE_COMPLETE)) {
1516 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1520 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1521 LASSERT(req->rq_next_phase != req->rq_phase);
1522 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1525 * Skip processing until reply is unlinked. We
1526 * can't return to pool before that and we can't
1527 * call interpret before that. We need to make
1528 * sure that all rdma transfers finished and will
1529 * not corrupt any data.
1531 if (ptlrpc_client_recv_or_unlink(req) ||
1532 ptlrpc_client_bulk_active(req))
1536 * Turn fail_loc off to prevent it from looping
1539 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1540 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1543 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1544 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1549 * Move to next phase if reply was successfully
1552 ptlrpc_rqphase_move(req, req->rq_next_phase);
1555 if (req->rq_phase == RQ_PHASE_COMPLETE)
1558 if (req->rq_phase == RQ_PHASE_INTERPRET)
1559 GOTO(interpret, req->rq_status);
1562 * Note that this also will start async reply unlink.
1564 if (req->rq_net_err && !req->rq_timedout) {
1565 ptlrpc_expire_one_request(req, 1);
1568 * Check if we still need to wait for unlink.
1570 if (ptlrpc_client_recv_or_unlink(req) ||
1571 ptlrpc_client_bulk_active(req))
1573 /* If there is no need to resend, fail it now. */
1574 if (req->rq_no_resend) {
1575 if (req->rq_status == 0)
1576 req->rq_status = -EIO;
1577 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1578 GOTO(interpret, req->rq_status);
1585 spin_lock(&req->rq_lock);
1586 req->rq_replied = 0;
1587 spin_unlock(&req->rq_lock);
1588 if (req->rq_status == 0)
1589 req->rq_status = -EIO;
1590 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1591 GOTO(interpret, req->rq_status);
1594 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1595 * so it sets rq_intr regardless of individual rpc
1596 * timeouts. The synchronous IO waiting path sets
1597 * rq_intr irrespective of whether ptlrpcd
1598 * has seen a timeout. Our policy is to only interpret
1599 * interrupted rpcs after they have timed out, so we
1600 * need to enforce that here.
1603 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1604 req->rq_wait_ctx)) {
1605 req->rq_status = -EINTR;
1606 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1607 GOTO(interpret, req->rq_status);
1610 if (req->rq_phase == RQ_PHASE_RPC) {
1611 if (req->rq_timedout || req->rq_resend ||
1612 req->rq_waiting || req->rq_wait_ctx) {
1615 if (!ptlrpc_unregister_reply(req, 1))
1618 spin_lock(&imp->imp_lock);
1619 if (ptlrpc_import_delay_req(imp, req, &status)){
1620 /* put on delay list - only if we wait
1621 * recovery finished - before send */
1622 cfs_list_del_init(&req->rq_list);
1623 cfs_list_add_tail(&req->rq_list,
1626 spin_unlock(&imp->imp_lock);
1631 req->rq_status = status;
1632 ptlrpc_rqphase_move(req,
1633 RQ_PHASE_INTERPRET);
1634 spin_unlock(&imp->imp_lock);
1635 GOTO(interpret, req->rq_status);
1637 if (ptlrpc_no_resend(req) &&
1638 !req->rq_wait_ctx) {
1639 req->rq_status = -ENOTCONN;
1640 ptlrpc_rqphase_move(req,
1641 RQ_PHASE_INTERPRET);
1642 spin_unlock(&imp->imp_lock);
1643 GOTO(interpret, req->rq_status);
1646 cfs_list_del_init(&req->rq_list);
1647 cfs_list_add_tail(&req->rq_list,
1648 &imp->imp_sending_list);
1650 spin_unlock(&imp->imp_lock);
1652 spin_lock(&req->rq_lock);
1653 req->rq_waiting = 0;
1654 spin_unlock(&req->rq_lock);
1656 if (req->rq_timedout || req->rq_resend) {
1657 /* This is re-sending anyways,
1658 * let's mark req as resend. */
1659 spin_lock(&req->rq_lock);
1661 spin_unlock(&req->rq_lock);
1665 if (!ptlrpc_unregister_bulk(req, 1))
1668 /* ensure previous bulk fails */
1669 old_xid = req->rq_xid;
1670 req->rq_xid = ptlrpc_next_xid();
1671 CDEBUG(D_HA, "resend bulk "
1674 old_xid, req->rq_xid);
1678 * rq_wait_ctx is only touched by ptlrpcd,
1679 * so no lock is needed here.
1681 status = sptlrpc_req_refresh_ctx(req, -1);
1684 req->rq_status = status;
1685 spin_lock(&req->rq_lock);
1686 req->rq_wait_ctx = 0;
1687 spin_unlock(&req->rq_lock);
1688 force_timer_recalc = 1;
1690 spin_lock(&req->rq_lock);
1691 req->rq_wait_ctx = 1;
1692 spin_unlock(&req->rq_lock);
1697 spin_lock(&req->rq_lock);
1698 req->rq_wait_ctx = 0;
1699 spin_unlock(&req->rq_lock);
1702 rc = ptl_send_rpc(req, 0);
1704 DEBUG_REQ(D_HA, req,
1705 "send failed: rc = %d", rc);
1706 force_timer_recalc = 1;
1707 spin_lock(&req->rq_lock);
1708 req->rq_net_err = 1;
1709 spin_unlock(&req->rq_lock);
1711 /* need to reset the timeout */
1712 force_timer_recalc = 1;
1715 spin_lock(&req->rq_lock);
1717 if (ptlrpc_client_early(req)) {
1718 ptlrpc_at_recv_early_reply(req);
1719 spin_unlock(&req->rq_lock);
1723 /* Still waiting for a reply? */
1724 if (ptlrpc_client_recv(req)) {
1725 spin_unlock(&req->rq_lock);
1729 /* Did we actually receive a reply? */
1730 if (!ptlrpc_client_replied(req)) {
1731 spin_unlock(&req->rq_lock);
1735 spin_unlock(&req->rq_lock);
1737 /* unlink from net because we are going to
1738 * swab in-place of reply buffer */
1739 unregistered = ptlrpc_unregister_reply(req, 1);
1743 req->rq_status = after_reply(req);
1747 /* If there is no bulk associated with this request,
1748 * then we're done and should let the interpreter
1749 * process the reply. Similarly if the RPC returned
1750 * an error, and therefore the bulk will never arrive.
1752 if (req->rq_bulk == NULL || req->rq_status < 0) {
1753 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1754 GOTO(interpret, req->rq_status);
1757 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1760 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1761 if (ptlrpc_client_bulk_active(req))
1764 if (req->rq_bulk->bd_failure) {
1765 /* The RPC reply arrived OK, but the bulk screwed
1766 * up! Dead weird since the server told us the RPC
1767 * was good after getting the REPLY for her GET or
1768 * the ACK for her PUT. */
1769 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1770 req->rq_status = -EIO;
1773 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1776 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1778 /* This moves to "unregistering" phase we need to wait for
1780 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1781 /* start async bulk unlink too */
1782 ptlrpc_unregister_bulk(req, 1);
1786 if (!ptlrpc_unregister_bulk(req, 1))
1789 /* When calling interpret receiving already should be
1791 LASSERT(!req->rq_receiving_reply);
1793 ptlrpc_req_interpret(env, req, req->rq_status);
1795 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1797 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1798 "Completed RPC pname:cluuid:pid:xid:nid:"
1799 "opc %s:%s:%d:"LPU64":%s:%d\n",
1800 cfs_curproc_comm(), imp->imp_obd->obd_uuid.uuid,
1801 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1802 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1803 lustre_msg_get_opc(req->rq_reqmsg));
1805 spin_lock(&imp->imp_lock);
1806 /* Request already may be not on sending or delaying list. This
1807 * may happen in the case of marking it erroneous for the case
1808 * ptlrpc_import_delay_req(req, status) find it impossible to
1809 * allow sending this rpc and returns *status != 0. */
1810 if (!cfs_list_empty(&req->rq_list)) {
1811 cfs_list_del_init(&req->rq_list);
1812 cfs_atomic_dec(&imp->imp_inflight);
1814 spin_unlock(&imp->imp_lock);
1816 cfs_atomic_dec(&set->set_remaining);
1817 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1819 if (set->set_producer) {
1820 /* produce a new request if possible */
1821 if (ptlrpc_set_producer(set) > 0)
1822 force_timer_recalc = 1;
1824 /* free the request that has just been completed
1825 * in order not to pollute set->set_requests */
1826 cfs_list_del_init(&req->rq_set_chain);
1827 spin_lock(&req->rq_lock);
1829 req->rq_invalid_rqset = 0;
1830 spin_unlock(&req->rq_lock);
1832 /* record rq_status to compute the final status later */
1833 if (req->rq_status != 0)
1834 set->set_rc = req->rq_status;
1835 ptlrpc_req_finished(req);
1839 /* If we hit an error, we want to recover promptly. */
1840 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1842 EXPORT_SYMBOL(ptlrpc_check_set);
1845 * Time out request \a req. is \a async_unlink is set, that means do not wait
1846 * until LNet actually confirms network buffer unlinking.
1847 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1849 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1851 struct obd_import *imp = req->rq_import;
1855 spin_lock(&req->rq_lock);
1856 req->rq_timedout = 1;
1857 spin_unlock(&req->rq_lock);
1859 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1860 "/real "CFS_DURATION_T"]",
1861 req->rq_net_err ? "failed due to network error" :
1862 ((req->rq_real_sent == 0 ||
1863 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1864 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1865 "timed out for sent delay" : "timed out for slow reply"),
1866 req->rq_sent, req->rq_real_sent);
1868 if (imp != NULL && obd_debug_peer_on_timeout)
1869 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1871 ptlrpc_unregister_reply(req, async_unlink);
1872 ptlrpc_unregister_bulk(req, async_unlink);
1874 if (obd_dump_on_timeout)
1875 libcfs_debug_dumplog();
1878 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1882 cfs_atomic_inc(&imp->imp_timeouts);
1884 /* The DLM server doesn't want recovery run on its imports. */
1885 if (imp->imp_dlm_fake)
1888 /* If this request is for recovery or other primordial tasks,
1889 * then error it out here. */
1890 if (req->rq_ctx_init || req->rq_ctx_fini ||
1891 req->rq_send_state != LUSTRE_IMP_FULL ||
1892 imp->imp_obd->obd_no_recov) {
1893 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1894 ptlrpc_import_state_name(req->rq_send_state),
1895 ptlrpc_import_state_name(imp->imp_state));
1896 spin_lock(&req->rq_lock);
1897 req->rq_status = -ETIMEDOUT;
1899 spin_unlock(&req->rq_lock);
1903 /* if a request can't be resent we can't wait for an answer after
1905 if (ptlrpc_no_resend(req)) {
1906 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1910 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1916 * Time out all uncompleted requests in request set pointed by \a data
1917 * Callback used when waiting on sets with l_wait_event.
1920 int ptlrpc_expired_set(void *data)
1922 struct ptlrpc_request_set *set = data;
1924 time_t now = cfs_time_current_sec();
1927 LASSERT(set != NULL);
1930 * A timeout expired. See which reqs it applies to...
1932 cfs_list_for_each (tmp, &set->set_requests) {
1933 struct ptlrpc_request *req =
1934 cfs_list_entry(tmp, struct ptlrpc_request,
1937 /* don't expire request waiting for context */
1938 if (req->rq_wait_ctx)
1941 /* Request in-flight? */
1942 if (!((req->rq_phase == RQ_PHASE_RPC &&
1943 !req->rq_waiting && !req->rq_resend) ||
1944 (req->rq_phase == RQ_PHASE_BULK)))
1947 if (req->rq_timedout || /* already dealt with */
1948 req->rq_deadline > now) /* not expired */
1951 /* Deal with this guy. Do it asynchronously to not block
1952 * ptlrpcd thread. */
1953 ptlrpc_expire_one_request(req, 1);
1957 * When waiting for a whole set, we always break out of the
1958 * sleep so we can recalculate the timeout, or enable interrupts
1959 * if everyone's timed out.
1963 EXPORT_SYMBOL(ptlrpc_expired_set);
1966 * Sets rq_intr flag in \a req under spinlock.
1968 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1970 spin_lock(&req->rq_lock);
1972 spin_unlock(&req->rq_lock);
1974 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
1977 * Interrupts (sets interrupted flag) all uncompleted requests in
1978 * a set \a data. Callback for l_wait_event for interruptible waits.
1980 void ptlrpc_interrupted_set(void *data)
1982 struct ptlrpc_request_set *set = data;
1985 LASSERT(set != NULL);
1986 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
1988 cfs_list_for_each(tmp, &set->set_requests) {
1989 struct ptlrpc_request *req =
1990 cfs_list_entry(tmp, struct ptlrpc_request,
1993 if (req->rq_phase != RQ_PHASE_RPC &&
1994 req->rq_phase != RQ_PHASE_UNREGISTERING)
1997 ptlrpc_mark_interrupted(req);
2000 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2003 * Get the smallest timeout in the set; this does NOT set a timeout.
2005 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2008 time_t now = cfs_time_current_sec();
2010 struct ptlrpc_request *req;
2014 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
2016 cfs_list_for_each(tmp, &set->set_requests) {
2017 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2020 * Request in-flight?
2022 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2023 (req->rq_phase == RQ_PHASE_BULK) ||
2024 (req->rq_phase == RQ_PHASE_NEW)))
2028 * Already timed out.
2030 if (req->rq_timedout)
2036 if (req->rq_wait_ctx)
2039 if (req->rq_phase == RQ_PHASE_NEW)
2040 deadline = req->rq_sent;
2041 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2042 deadline = req->rq_sent;
2044 deadline = req->rq_sent + req->rq_timeout;
2046 if (deadline <= now) /* actually expired already */
2047 timeout = 1; /* ASAP */
2048 else if (timeout == 0 || timeout > deadline - now)
2049 timeout = deadline - now;
2053 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2056 * Send all unset request from the set and then wait untill all
2057 * requests in the set complete (either get a reply, timeout, get an
2058 * error or otherwise be interrupted).
2059 * Returns 0 on success or error code otherwise.
2061 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2064 struct ptlrpc_request *req;
2065 struct l_wait_info lwi;
2069 if (set->set_producer)
2070 (void)ptlrpc_set_producer(set);
2072 cfs_list_for_each(tmp, &set->set_requests) {
2073 req = cfs_list_entry(tmp, struct ptlrpc_request,
2075 if (req->rq_phase == RQ_PHASE_NEW)
2076 (void)ptlrpc_send_new_req(req);
2079 if (cfs_list_empty(&set->set_requests))
2083 timeout = ptlrpc_set_next_timeout(set);
2085 /* wait until all complete, interrupted, or an in-flight
2087 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2090 if (timeout == 0 && !cfs_signal_pending())
2092 * No requests are in-flight (ether timed out
2093 * or delayed), so we can allow interrupts.
2094 * We still want to block for a limited time,
2095 * so we allow interrupts during the timeout.
2097 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2099 ptlrpc_interrupted_set, set);
2102 * At least one request is in flight, so no
2103 * interrupts are allowed. Wait until all
2104 * complete, or an in-flight req times out.
2106 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2107 ptlrpc_expired_set, set);
2109 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2111 /* LU-769 - if we ignored the signal because it was already
2112 * pending when we started, we need to handle it now or we risk
2113 * it being ignored forever */
2114 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2115 cfs_signal_pending()) {
2116 cfs_sigset_t blocked_sigs =
2117 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2119 /* In fact we only interrupt for the "fatal" signals
2120 * like SIGINT or SIGKILL. We still ignore less
2121 * important signals since ptlrpc set is not easily
2122 * reentrant from userspace again */
2123 if (cfs_signal_pending())
2124 ptlrpc_interrupted_set(set);
2125 cfs_restore_sigs(blocked_sigs);
2128 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2130 /* -EINTR => all requests have been flagged rq_intr so next
2132 * -ETIMEDOUT => someone timed out. When all reqs have
2133 * timed out, signals are enabled allowing completion with
2135 * I don't really care if we go once more round the loop in
2136 * the error cases -eeb. */
2137 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2138 cfs_list_for_each(tmp, &set->set_requests) {
2139 req = cfs_list_entry(tmp, struct ptlrpc_request,
2141 spin_lock(&req->rq_lock);
2142 req->rq_invalid_rqset = 1;
2143 spin_unlock(&req->rq_lock);
2146 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2148 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2150 rc = set->set_rc; /* rq_status of already freed requests if any */
2151 cfs_list_for_each(tmp, &set->set_requests) {
2152 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2154 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2155 if (req->rq_status != 0)
2156 rc = req->rq_status;
2159 if (set->set_interpret != NULL) {
2160 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2162 rc = interpreter (set, set->set_arg, rc);
2164 struct ptlrpc_set_cbdata *cbdata, *n;
2167 cfs_list_for_each_entry_safe(cbdata, n,
2168 &set->set_cblist, psc_item) {
2169 cfs_list_del_init(&cbdata->psc_item);
2170 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2173 OBD_FREE_PTR(cbdata);
2179 EXPORT_SYMBOL(ptlrpc_set_wait);
2182 * Helper fuction for request freeing.
2183 * Called when request count reached zero and request needs to be freed.
2184 * Removes request from all sorts of sending/replay lists it might be on,
2185 * frees network buffers if any are present.
2186 * If \a locked is set, that means caller is already holding import imp_lock
2187 * and so we no longer need to reobtain it (for certain lists manipulations)
2189 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2192 if (request == NULL) {
2197 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2198 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2199 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2200 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2201 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2202 LASSERTF(!request->rq_replay, "req %p\n", request);
2204 req_capsule_fini(&request->rq_pill);
2206 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2207 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2208 if (request->rq_import != NULL) {
2210 spin_lock(&request->rq_import->imp_lock);
2211 cfs_list_del_init(&request->rq_replay_list);
2213 spin_unlock(&request->rq_import->imp_lock);
2215 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2217 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2218 DEBUG_REQ(D_ERROR, request,
2219 "freeing request with nonzero refcount");
2223 if (request->rq_repbuf != NULL)
2224 sptlrpc_cli_free_repbuf(request);
2225 if (request->rq_export != NULL) {
2226 class_export_put(request->rq_export);
2227 request->rq_export = NULL;
2229 if (request->rq_import != NULL) {
2230 class_import_put(request->rq_import);
2231 request->rq_import = NULL;
2233 if (request->rq_bulk != NULL)
2234 ptlrpc_free_bulk_pin(request->rq_bulk);
2236 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2237 sptlrpc_cli_free_reqbuf(request);
2239 if (request->rq_cli_ctx)
2240 sptlrpc_req_put_ctx(request, !locked);
2242 if (request->rq_pool)
2243 __ptlrpc_free_req_to_pool(request);
2245 OBD_FREE(request, sizeof(*request));
2249 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2251 * Drop one request reference. Must be called with import imp_lock held.
2252 * When reference count drops to zero, reuqest is freed.
2254 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2256 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2257 (void)__ptlrpc_req_finished(request, 1);
2259 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2263 * Drops one reference count for request \a request.
2264 * \a locked set indicates that caller holds import imp_lock.
2265 * Frees the request whe reference count reaches zero.
2267 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2270 if (request == NULL)
2273 if (request == LP_POISON ||
2274 request->rq_reqmsg == LP_POISON) {
2275 CERROR("dereferencing freed request (bug 575)\n");
2280 DEBUG_REQ(D_INFO, request, "refcount now %u",
2281 cfs_atomic_read(&request->rq_refcount) - 1);
2283 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2284 __ptlrpc_free_req(request, locked);
2292 * Drops one reference count for a request.
2294 void ptlrpc_req_finished(struct ptlrpc_request *request)
2296 __ptlrpc_req_finished(request, 0);
2298 EXPORT_SYMBOL(ptlrpc_req_finished);
2301 * Returns xid of a \a request
2303 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2305 return request->rq_xid;
2307 EXPORT_SYMBOL(ptlrpc_req_xid);
2310 * Disengage the client's reply buffer from the network
2311 * NB does _NOT_ unregister any client-side bulk.
2312 * IDEMPOTENT, but _not_ safe against concurrent callers.
2313 * The request owner (i.e. the thread doing the I/O) must call...
2314 * Returns 0 on success or 1 if unregistering cannot be made.
2316 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2320 struct l_wait_info lwi;
2325 LASSERT(!cfs_in_interrupt());
2328 * Let's setup deadline for reply unlink.
2330 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2331 async && request->rq_reply_deadline == 0)
2332 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2335 * Nothing left to do.
2337 if (!ptlrpc_client_recv_or_unlink(request))
2340 LNetMDUnlink(request->rq_reply_md_h);
2343 * Let's check it once again.
2345 if (!ptlrpc_client_recv_or_unlink(request))
2349 * Move to "Unregistering" phase as reply was not unlinked yet.
2351 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2354 * Do not wait for unlink to finish.
2360 * We have to l_wait_event() whatever the result, to give liblustre
2361 * a chance to run reply_in_callback(), and to make sure we've
2362 * unlinked before returning a req to the pool.
2364 if (request->rq_set != NULL)
2365 wq = &request->rq_set->set_waitq;
2367 wq = &request->rq_reply_waitq;
2370 /* Network access will complete in finite time but the HUGE
2371 * timeout lets us CWARN for visibility of sluggish NALs */
2372 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2373 cfs_time_seconds(1), NULL, NULL);
2374 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2377 ptlrpc_rqphase_move(request, request->rq_next_phase);
2381 LASSERT(rc == -ETIMEDOUT);
2382 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2383 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2384 request->rq_must_unlink);
2388 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2391 * Iterates through replay_list on import and prunes
2392 * all requests have transno smaller than last_committed for the
2393 * import and don't have rq_replay set.
2394 * Since requests are sorted in transno order, stops when meetign first
2395 * transno bigger than last_committed.
2396 * caller must hold imp->imp_lock
2398 void ptlrpc_free_committed(struct obd_import *imp)
2400 cfs_list_t *tmp, *saved;
2401 struct ptlrpc_request *req;
2402 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2405 LASSERT(imp != NULL);
2407 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2410 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2411 imp->imp_generation == imp->imp_last_generation_checked) {
2412 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2413 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2417 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2418 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2419 imp->imp_generation);
2420 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2421 imp->imp_last_generation_checked = imp->imp_generation;
2423 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2424 req = cfs_list_entry(tmp, struct ptlrpc_request,
2427 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2428 LASSERT(req != last_req);
2431 if (req->rq_transno == 0) {
2432 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2435 if (req->rq_import_generation < imp->imp_generation) {
2436 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2440 if (req->rq_replay) {
2441 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2445 /* not yet committed */
2446 if (req->rq_transno > imp->imp_peer_committed_transno) {
2447 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2451 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2452 imp->imp_peer_committed_transno);
2454 spin_lock(&req->rq_lock);
2456 spin_unlock(&req->rq_lock);
2457 if (req->rq_commit_cb != NULL)
2458 req->rq_commit_cb(req);
2459 cfs_list_del_init(&req->rq_replay_list);
2460 __ptlrpc_req_finished(req, 1);
2467 void ptlrpc_cleanup_client(struct obd_import *imp)
2473 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2476 * Schedule previously sent request for resend.
2477 * For bulk requests we assign new xid (to avoid problems with
2478 * lost replies and therefore several transfers landing into same buffer
2479 * from different sending attempts).
2481 void ptlrpc_resend_req(struct ptlrpc_request *req)
2483 DEBUG_REQ(D_HA, req, "going to resend");
2484 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2485 req->rq_status = -EAGAIN;
2487 spin_lock(&req->rq_lock);
2489 req->rq_net_err = 0;
2490 req->rq_timedout = 0;
2492 __u64 old_xid = req->rq_xid;
2494 /* ensure previous bulk fails */
2495 req->rq_xid = ptlrpc_next_xid();
2496 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2497 old_xid, req->rq_xid);
2499 ptlrpc_client_wake_req(req);
2500 spin_unlock(&req->rq_lock);
2502 EXPORT_SYMBOL(ptlrpc_resend_req);
2504 /* XXX: this function and rq_status are currently unused */
2505 void ptlrpc_restart_req(struct ptlrpc_request *req)
2507 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2508 req->rq_status = -ERESTARTSYS;
2510 spin_lock(&req->rq_lock);
2511 req->rq_restart = 1;
2512 req->rq_timedout = 0;
2513 ptlrpc_client_wake_req(req);
2514 spin_unlock(&req->rq_lock);
2516 EXPORT_SYMBOL(ptlrpc_restart_req);
2519 * Grab additional reference on a request \a req
2521 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2524 cfs_atomic_inc(&req->rq_refcount);
2527 EXPORT_SYMBOL(ptlrpc_request_addref);
2530 * Add a request to import replay_list.
2531 * Must be called under imp_lock
2533 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2534 struct obd_import *imp)
2538 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2540 if (req->rq_transno == 0) {
2541 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2545 /* clear this for new requests that were resent as well
2546 as resent replayed requests. */
2547 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2549 /* don't re-add requests that have been replayed */
2550 if (!cfs_list_empty(&req->rq_replay_list))
2553 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2555 LASSERT(imp->imp_replayable);
2556 /* Balanced in ptlrpc_free_committed, usually. */
2557 ptlrpc_request_addref(req);
2558 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2559 struct ptlrpc_request *iter =
2560 cfs_list_entry(tmp, struct ptlrpc_request,
2563 /* We may have duplicate transnos if we create and then
2564 * open a file, or for closes retained if to match creating
2565 * opens, so use req->rq_xid as a secondary key.
2566 * (See bugs 684, 685, and 428.)
2567 * XXX no longer needed, but all opens need transnos!
2569 if (iter->rq_transno > req->rq_transno)
2572 if (iter->rq_transno == req->rq_transno) {
2573 LASSERT(iter->rq_xid != req->rq_xid);
2574 if (iter->rq_xid > req->rq_xid)
2578 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2582 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2584 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2587 * Send request and wait until it completes.
2588 * Returns request processing status.
2590 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2592 struct ptlrpc_request_set *set;
2596 LASSERT(req->rq_set == NULL);
2597 LASSERT(!req->rq_receiving_reply);
2599 set = ptlrpc_prep_set();
2601 CERROR("Unable to allocate ptlrpc set.");
2605 /* for distributed debugging */
2606 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2608 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2609 ptlrpc_request_addref(req);
2610 ptlrpc_set_add_req(set, req);
2611 rc = ptlrpc_set_wait(set);
2612 ptlrpc_set_destroy(set);
2616 EXPORT_SYMBOL(ptlrpc_queue_wait);
2618 struct ptlrpc_replay_async_args {
2620 int praa_old_status;
2624 * Callback used for replayed requests reply processing.
2625 * In case of succesful reply calls registeresd request replay callback.
2626 * In case of error restart replay process.
2628 static int ptlrpc_replay_interpret(const struct lu_env *env,
2629 struct ptlrpc_request *req,
2630 void * data, int rc)
2632 struct ptlrpc_replay_async_args *aa = data;
2633 struct obd_import *imp = req->rq_import;
2636 cfs_atomic_dec(&imp->imp_replay_inflight);
2638 if (!ptlrpc_client_replied(req)) {
2639 CERROR("request replay timed out, restarting recovery\n");
2640 GOTO(out, rc = -ETIMEDOUT);
2643 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2644 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2645 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2646 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2648 /** VBR: check version failure */
2649 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2650 /** replay was failed due to version mismatch */
2651 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2652 spin_lock(&imp->imp_lock);
2653 imp->imp_vbr_failed = 1;
2654 imp->imp_no_lock_replay = 1;
2655 spin_unlock(&imp->imp_lock);
2656 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2658 /** The transno had better not change over replay. */
2659 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2660 lustre_msg_get_transno(req->rq_repmsg) ||
2661 lustre_msg_get_transno(req->rq_repmsg) == 0,
2663 lustre_msg_get_transno(req->rq_reqmsg),
2664 lustre_msg_get_transno(req->rq_repmsg));
2667 spin_lock(&imp->imp_lock);
2668 /** if replays by version then gap occur on server, no trust to locks */
2669 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2670 imp->imp_no_lock_replay = 1;
2671 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2672 spin_unlock(&imp->imp_lock);
2673 LASSERT(imp->imp_last_replay_transno);
2675 /* transaction number shouldn't be bigger than the latest replayed */
2676 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2677 DEBUG_REQ(D_ERROR, req,
2678 "Reported transno "LPU64" is bigger than the "
2679 "replayed one: "LPU64, req->rq_transno,
2680 lustre_msg_get_transno(req->rq_reqmsg));
2681 GOTO(out, rc = -EINVAL);
2684 DEBUG_REQ(D_HA, req, "got rep");
2686 /* let the callback do fixups, possibly including in the request */
2687 if (req->rq_replay_cb)
2688 req->rq_replay_cb(req);
2690 if (ptlrpc_client_replied(req) &&
2691 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2692 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2693 lustre_msg_get_status(req->rq_repmsg),
2694 aa->praa_old_status);
2696 /* Put it back for re-replay. */
2697 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2701 * Errors while replay can set transno to 0, but
2702 * imp_last_replay_transno shouldn't be set to 0 anyway
2704 if (req->rq_transno == 0)
2705 CERROR("Transno is 0 during replay!\n");
2707 /* continue with recovery */
2708 rc = ptlrpc_import_recovery_state_machine(imp);
2710 req->rq_send_state = aa->praa_old_state;
2713 /* this replay failed, so restart recovery */
2714 ptlrpc_connect_import(imp);
2720 * Prepares and queues request for replay.
2721 * Adds it to ptlrpcd queue for actual sending.
2722 * Returns 0 on success.
2724 int ptlrpc_replay_req(struct ptlrpc_request *req)
2726 struct ptlrpc_replay_async_args *aa;
2729 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2731 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2732 aa = ptlrpc_req_async_args(req);
2733 memset(aa, 0, sizeof *aa);
2735 /* Prepare request to be resent with ptlrpcd */
2736 aa->praa_old_state = req->rq_send_state;
2737 req->rq_send_state = LUSTRE_IMP_REPLAY;
2738 req->rq_phase = RQ_PHASE_NEW;
2739 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2741 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2743 req->rq_interpret_reply = ptlrpc_replay_interpret;
2744 /* Readjust the timeout for current conditions */
2745 ptlrpc_at_set_req_timeout(req);
2747 /* Tell server the net_latency, so the server can calculate how long
2748 * it should wait for next replay */
2749 lustre_msg_set_service_time(req->rq_reqmsg,
2750 ptlrpc_at_get_net_latency(req));
2751 DEBUG_REQ(D_HA, req, "REPLAY");
2753 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2754 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2756 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2759 EXPORT_SYMBOL(ptlrpc_replay_req);
2762 * Aborts all in-flight request on import \a imp sending and delayed lists
2764 void ptlrpc_abort_inflight(struct obd_import *imp)
2766 cfs_list_t *tmp, *n;
2769 /* Make sure that no new requests get processed for this import.
2770 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2771 * this flag and then putting requests on sending_list or delayed_list.
2773 spin_lock(&imp->imp_lock);
2775 /* XXX locking? Maybe we should remove each request with the list
2776 * locked? Also, how do we know if the requests on the list are
2777 * being freed at this time?
2779 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2780 struct ptlrpc_request *req =
2781 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2783 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2785 spin_lock(&req->rq_lock);
2786 if (req->rq_import_generation < imp->imp_generation) {
2788 req->rq_status = -EIO;
2789 ptlrpc_client_wake_req(req);
2791 spin_unlock(&req->rq_lock);
2794 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2795 struct ptlrpc_request *req =
2796 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2798 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2800 spin_lock(&req->rq_lock);
2801 if (req->rq_import_generation < imp->imp_generation) {
2803 req->rq_status = -EIO;
2804 ptlrpc_client_wake_req(req);
2806 spin_unlock(&req->rq_lock);
2809 /* Last chance to free reqs left on the replay list, but we
2810 * will still leak reqs that haven't committed. */
2811 if (imp->imp_replayable)
2812 ptlrpc_free_committed(imp);
2814 spin_unlock(&imp->imp_lock);
2818 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2821 * Abort all uncompleted requests in request set \a set
2823 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2825 cfs_list_t *tmp, *pos;
2827 LASSERT(set != NULL);
2829 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2830 struct ptlrpc_request *req =
2831 cfs_list_entry(pos, struct ptlrpc_request,
2834 spin_lock(&req->rq_lock);
2835 if (req->rq_phase != RQ_PHASE_RPC) {
2836 spin_unlock(&req->rq_lock);
2841 req->rq_status = -EINTR;
2842 ptlrpc_client_wake_req(req);
2843 spin_unlock(&req->rq_lock);
2847 static __u64 ptlrpc_last_xid;
2848 static spinlock_t ptlrpc_last_xid_lock;
2851 * Initialize the XID for the node. This is common among all requests on
2852 * this node, and only requires the property that it is monotonically
2853 * increasing. It does not need to be sequential. Since this is also used
2854 * as the RDMA match bits, it is important that a single client NOT have
2855 * the same match bits for two different in-flight requests, hence we do
2856 * NOT want to have an XID per target or similar.
2858 * To avoid an unlikely collision between match bits after a client reboot
2859 * (which would deliver old data into the wrong RDMA buffer) initialize
2860 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2861 * If the time is clearly incorrect, we instead use a 62-bit random number.
2862 * In the worst case the random number will overflow 1M RPCs per second in
2863 * 9133 years, or permutations thereof.
2865 #define YEAR_2004 (1ULL << 30)
2866 void ptlrpc_init_xid(void)
2868 time_t now = cfs_time_current_sec();
2870 spin_lock_init(&ptlrpc_last_xid_lock);
2871 if (now < YEAR_2004) {
2872 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2873 ptlrpc_last_xid >>= 2;
2874 ptlrpc_last_xid |= (1ULL << 61);
2876 ptlrpc_last_xid = (__u64)now << 20;
2879 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
2880 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
2881 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
2885 * Increase xid and returns resulting new value to the caller.
2887 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2888 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2889 * itself uses the last bulk xid needed, so the server can determine the
2890 * the number of bulk transfers from the RPC XID and a bitmask. The starting
2891 * xid must align to a power-of-two value.
2893 * This is assumed to be true due to the initial ptlrpc_last_xid
2894 * value also being initialized to a power-of-two value. LU-1431
2896 __u64 ptlrpc_next_xid(void)
2900 spin_lock(&ptlrpc_last_xid_lock);
2901 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2902 ptlrpc_last_xid = next;
2903 spin_unlock(&ptlrpc_last_xid_lock);
2907 EXPORT_SYMBOL(ptlrpc_next_xid);
2910 * Get a glimpse at what next xid value might have been.
2911 * Returns possible next xid.
2913 __u64 ptlrpc_sample_next_xid(void)
2915 #if BITS_PER_LONG == 32
2916 /* need to avoid possible word tearing on 32-bit systems */
2919 spin_lock(&ptlrpc_last_xid_lock);
2920 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2921 spin_unlock(&ptlrpc_last_xid_lock);
2925 /* No need to lock, since returned value is racy anyways */
2926 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2929 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2932 * Functions for operating ptlrpc workers.
2934 * A ptlrpc work is a function which will be running inside ptlrpc context.
2935 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2937 * 1. after a work is created, it can be used many times, that is:
2938 * handler = ptlrpcd_alloc_work();
2939 * ptlrpcd_queue_work();
2941 * queue it again when necessary:
2942 * ptlrpcd_queue_work();
2943 * ptlrpcd_destroy_work();
2944 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2945 * it will only be queued once in any time. Also as its name implies, it may
2946 * have delay before it really runs by ptlrpcd thread.
2948 struct ptlrpc_work_async_args {
2950 int (*cb)(const struct lu_env *, void *);
2954 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2956 static int work_interpreter(const struct lu_env *env,
2957 struct ptlrpc_request *req, void *data, int rc)
2959 struct ptlrpc_work_async_args *arg = data;
2961 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2962 LASSERT(arg->cb != NULL);
2964 return arg->cb(env, arg->cbdata);
2968 * Create a work for ptlrpc.
2970 void *ptlrpcd_alloc_work(struct obd_import *imp,
2971 int (*cb)(const struct lu_env *, void *), void *cbdata)
2973 struct ptlrpc_request *req = NULL;
2974 struct ptlrpc_work_async_args *args;
2980 RETURN(ERR_PTR(-EINVAL));
2982 /* copy some code from deprecated fakereq. */
2985 CERROR("ptlrpc: run out of memory!\n");
2986 RETURN(ERR_PTR(-ENOMEM));
2989 req->rq_send_state = LUSTRE_IMP_FULL;
2990 req->rq_type = PTL_RPC_MSG_REQUEST;
2991 req->rq_import = class_import_get(imp);
2992 req->rq_export = NULL;
2993 req->rq_interpret_reply = work_interpreter;
2994 /* don't want reply */
2995 req->rq_receiving_reply = 0;
2996 req->rq_must_unlink = 0;
2997 req->rq_no_delay = req->rq_no_resend = 1;
2999 spin_lock_init(&req->rq_lock);
3000 CFS_INIT_LIST_HEAD(&req->rq_list);
3001 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
3002 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
3003 CFS_INIT_LIST_HEAD(&req->rq_history_list);
3004 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
3005 cfs_waitq_init(&req->rq_reply_waitq);
3006 cfs_waitq_init(&req->rq_set_waitq);
3007 cfs_atomic_set(&req->rq_refcount, 1);
3009 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3010 args = ptlrpc_req_async_args(req);
3011 args->magic = PTLRPC_WORK_MAGIC;
3013 args->cbdata = cbdata;
3017 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3019 void ptlrpcd_destroy_work(void *handler)
3021 struct ptlrpc_request *req = handler;
3024 ptlrpc_req_finished(req);
3026 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3028 int ptlrpcd_queue_work(void *handler)
3030 struct ptlrpc_request *req = handler;
3033 * Check if the req is already being queued.
3035 * Here comes a trick: it lacks a way of checking if a req is being
3036 * processed reliably in ptlrpc. Here I have to use refcount of req
3037 * for this purpose. This is okay because the caller should use this
3038 * req as opaque data. - Jinshan
3040 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
3041 if (cfs_atomic_read(&req->rq_refcount) > 1)
3044 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
3045 cfs_atomic_dec(&req->rq_refcount);
3049 /* re-initialize the req */
3050 req->rq_timeout = obd_timeout;
3051 req->rq_sent = cfs_time_current_sec();
3052 req->rq_deadline = req->rq_sent + req->rq_timeout;
3053 req->rq_reply_deadline = req->rq_deadline;
3054 req->rq_phase = RQ_PHASE_INTERPRET;
3055 req->rq_next_phase = RQ_PHASE_COMPLETE;
3056 req->rq_xid = ptlrpc_next_xid();
3057 req->rq_import_generation = req->rq_import->imp_generation;
3059 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3062 EXPORT_SYMBOL(ptlrpcd_queue_work);