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, 2013, 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 init_waitqueue_head(&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 struct page *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 <= PAGE_CACHE_SIZE);
182 page_cache_get(page);
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 page_cache_release(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));
390 struct kmem_cache *request_cache;
392 int ptlrpc_request_cache_init()
394 request_cache = kmem_cache_create("ptlrpc_cache",
395 sizeof(struct ptlrpc_request),
396 0, SLAB_HWCACHE_ALIGN, NULL);
397 return request_cache == NULL ? -ENOMEM : 0;
400 void ptlrpc_request_cache_fini()
402 kmem_cache_destroy(request_cache);
405 struct ptlrpc_request *ptlrpc_request_cache_alloc(int flags)
407 struct ptlrpc_request *req;
409 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
413 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
415 OBD_SLAB_FREE_PTR(req, request_cache);
419 * Wind down request pool \a pool.
420 * Frees all requests from the pool too
422 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
425 struct ptlrpc_request *req;
427 LASSERT(pool != NULL);
429 spin_lock(&pool->prp_lock);
430 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
431 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
432 cfs_list_del(&req->rq_list);
433 LASSERT(req->rq_reqbuf);
434 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
435 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
436 ptlrpc_request_cache_free(req);
438 spin_unlock(&pool->prp_lock);
439 OBD_FREE(pool, sizeof(*pool));
441 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
444 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
446 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
451 while (size < pool->prp_rq_size)
454 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
455 size == pool->prp_rq_size,
456 "Trying to change pool size with nonempty pool "
457 "from %d to %d bytes\n", pool->prp_rq_size, size);
459 spin_lock(&pool->prp_lock);
460 pool->prp_rq_size = size;
461 for (i = 0; i < num_rq; i++) {
462 struct ptlrpc_request *req;
463 struct lustre_msg *msg;
465 spin_unlock(&pool->prp_lock);
466 req = ptlrpc_request_cache_alloc(__GFP_IO);
469 OBD_ALLOC_LARGE(msg, size);
471 ptlrpc_request_cache_free(req);
474 req->rq_reqbuf = msg;
475 req->rq_reqbuf_len = size;
477 spin_lock(&pool->prp_lock);
478 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
480 spin_unlock(&pool->prp_lock);
483 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
486 * Create and initialize new request pool with given attributes:
487 * \a num_rq - initial number of requests to create for the pool
488 * \a msgsize - maximum message size possible for requests in thid pool
489 * \a populate_pool - function to be called when more requests need to be added
491 * Returns pointer to newly created pool or NULL on error.
493 struct ptlrpc_request_pool *
494 ptlrpc_init_rq_pool(int num_rq, int msgsize,
495 void (*populate_pool)(struct ptlrpc_request_pool *, int))
497 struct ptlrpc_request_pool *pool;
499 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
503 /* Request next power of two for the allocation, because internally
504 kernel would do exactly this */
506 spin_lock_init(&pool->prp_lock);
507 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
508 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
509 pool->prp_populate = populate_pool;
511 populate_pool(pool, num_rq);
513 if (cfs_list_empty(&pool->prp_req_list)) {
514 /* have not allocated a single request for the pool */
515 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
520 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
523 * Fetches one request from pool \a pool
525 static struct ptlrpc_request *
526 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
528 struct ptlrpc_request *request;
529 struct lustre_msg *reqbuf;
534 spin_lock(&pool->prp_lock);
536 /* See if we have anything in a pool, and bail out if nothing,
537 * in writeout path, where this matters, this is safe to do, because
538 * nothing is lost in this case, and when some in-flight requests
539 * complete, this code will be called again. */
540 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
541 spin_unlock(&pool->prp_lock);
545 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
547 cfs_list_del_init(&request->rq_list);
548 spin_unlock(&pool->prp_lock);
550 LASSERT(request->rq_reqbuf);
551 LASSERT(request->rq_pool);
553 reqbuf = request->rq_reqbuf;
554 memset(request, 0, sizeof(*request));
555 request->rq_reqbuf = reqbuf;
556 request->rq_reqbuf_len = pool->prp_rq_size;
557 request->rq_pool = pool;
563 * Returns freed \a request to pool.
565 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
567 struct ptlrpc_request_pool *pool = request->rq_pool;
569 spin_lock(&pool->prp_lock);
570 LASSERT(cfs_list_empty(&request->rq_list));
571 LASSERT(!request->rq_receiving_reply);
572 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
573 spin_unlock(&pool->prp_lock);
576 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
577 __u32 version, int opcode,
578 int count, __u32 *lengths, char **bufs,
579 struct ptlrpc_cli_ctx *ctx)
581 struct obd_import *imp = request->rq_import;
586 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
588 rc = sptlrpc_req_get_ctx(request);
593 sptlrpc_req_set_flavor(request, opcode);
595 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
598 LASSERT(!request->rq_pool);
602 lustre_msg_add_version(request->rq_reqmsg, version);
603 request->rq_send_state = LUSTRE_IMP_FULL;
604 request->rq_type = PTL_RPC_MSG_REQUEST;
605 request->rq_export = NULL;
607 request->rq_req_cbid.cbid_fn = request_out_callback;
608 request->rq_req_cbid.cbid_arg = request;
610 request->rq_reply_cbid.cbid_fn = reply_in_callback;
611 request->rq_reply_cbid.cbid_arg = request;
613 request->rq_reply_deadline = 0;
614 request->rq_phase = RQ_PHASE_NEW;
615 request->rq_next_phase = RQ_PHASE_UNDEFINED;
617 request->rq_request_portal = imp->imp_client->cli_request_portal;
618 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
620 ptlrpc_at_set_req_timeout(request);
622 spin_lock_init(&request->rq_lock);
623 CFS_INIT_LIST_HEAD(&request->rq_list);
624 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
625 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
626 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
627 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
628 CFS_INIT_LIST_HEAD(&request->rq_history_list);
629 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
630 init_waitqueue_head(&request->rq_reply_waitq);
631 init_waitqueue_head(&request->rq_set_waitq);
632 request->rq_xid = ptlrpc_next_xid();
633 cfs_atomic_set(&request->rq_refcount, 1);
635 lustre_msg_set_opc(request->rq_reqmsg, opcode);
639 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
641 class_import_put(imp);
645 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
646 __u32 version, int opcode, char **bufs,
647 struct ptlrpc_cli_ctx *ctx)
651 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
652 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
653 request->rq_pill.rc_area[RCL_CLIENT],
656 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
659 * Pack request buffers for network transfer, performing necessary encryption
660 * steps if necessary.
662 int ptlrpc_request_pack(struct ptlrpc_request *request,
663 __u32 version, int opcode)
666 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
670 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
671 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
672 * have to send old ptlrpc_body to keep interoprability with these
675 * Only three kinds of server->client RPCs so far:
680 * XXX This should be removed whenever we drop the interoprability with
681 * the these old clients.
683 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
684 opcode == LDLM_GL_CALLBACK)
685 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
686 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
690 EXPORT_SYMBOL(ptlrpc_request_pack);
693 * Helper function to allocate new request on import \a imp
694 * and possibly using existing request from pool \a pool if provided.
695 * Returns allocated request structure with import field filled or
699 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
700 struct ptlrpc_request_pool *pool)
702 struct ptlrpc_request *request = NULL;
705 request = ptlrpc_prep_req_from_pool(pool);
708 request = ptlrpc_request_cache_alloc(__GFP_IO);
711 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
712 LASSERT(imp != LP_POISON);
713 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
715 LASSERT(imp->imp_client != LP_POISON);
717 request->rq_import = class_import_get(imp);
719 CERROR("request allocation out of memory\n");
726 * Helper function for creating a request.
727 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
728 * buffer structures according to capsule template \a format.
729 * Returns allocated request structure pointer or NULL on error.
731 static struct ptlrpc_request *
732 ptlrpc_request_alloc_internal(struct obd_import *imp,
733 struct ptlrpc_request_pool * pool,
734 const struct req_format *format)
736 struct ptlrpc_request *request;
738 request = __ptlrpc_request_alloc(imp, pool);
742 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
743 req_capsule_set(&request->rq_pill, format);
748 * Allocate new request structure for import \a imp and initialize its
749 * buffer structure according to capsule template \a format.
751 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
752 const struct req_format *format)
754 return ptlrpc_request_alloc_internal(imp, NULL, format);
756 EXPORT_SYMBOL(ptlrpc_request_alloc);
759 * Allocate new request structure for import \a imp from pool \a pool and
760 * initialize its buffer structure according to capsule template \a format.
762 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
763 struct ptlrpc_request_pool * pool,
764 const struct req_format *format)
766 return ptlrpc_request_alloc_internal(imp, pool, format);
768 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
771 * For requests not from pool, free memory of the request structure.
772 * For requests obtained from a pool earlier, return request back to pool.
774 void ptlrpc_request_free(struct ptlrpc_request *request)
776 if (request->rq_pool)
777 __ptlrpc_free_req_to_pool(request);
779 ptlrpc_request_cache_free(request);
781 EXPORT_SYMBOL(ptlrpc_request_free);
784 * Allocate new request for operatione \a opcode and immediatelly pack it for
786 * Only used for simple requests like OBD_PING where the only important
787 * part of the request is operation itself.
788 * Returns allocated request or NULL on error.
790 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
791 const struct req_format *format,
792 __u32 version, int opcode)
794 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
798 rc = ptlrpc_request_pack(req, version, opcode);
800 ptlrpc_request_free(req);
806 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
809 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
810 * for operation \a opcode. Request would contain \a count buffers.
811 * Sizes of buffers are described in array \a lengths and buffers themselves
812 * are provided by a pointer \a bufs.
813 * Returns prepared request structure pointer or NULL on error.
815 struct ptlrpc_request *
816 ptlrpc_prep_req_pool(struct obd_import *imp,
817 __u32 version, int opcode,
818 int count, __u32 *lengths, char **bufs,
819 struct ptlrpc_request_pool *pool)
821 struct ptlrpc_request *request;
824 request = __ptlrpc_request_alloc(imp, pool);
828 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
829 lengths, bufs, NULL);
831 ptlrpc_request_free(request);
836 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
839 * Same as ptlrpc_prep_req_pool, but without pool
841 struct ptlrpc_request *
842 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
843 __u32 *lengths, char **bufs)
845 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
848 EXPORT_SYMBOL(ptlrpc_prep_req);
851 * Allocate and initialize new request set structure.
852 * Returns a pointer to the newly allocated set structure or NULL on error.
854 struct ptlrpc_request_set *ptlrpc_prep_set(void)
856 struct ptlrpc_request_set *set;
859 OBD_ALLOC(set, sizeof *set);
862 cfs_atomic_set(&set->set_refcount, 1);
863 CFS_INIT_LIST_HEAD(&set->set_requests);
864 init_waitqueue_head(&set->set_waitq);
865 cfs_atomic_set(&set->set_new_count, 0);
866 cfs_atomic_set(&set->set_remaining, 0);
867 spin_lock_init(&set->set_new_req_lock);
868 CFS_INIT_LIST_HEAD(&set->set_new_requests);
869 CFS_INIT_LIST_HEAD(&set->set_cblist);
870 set->set_max_inflight = UINT_MAX;
871 set->set_producer = NULL;
872 set->set_producer_arg = NULL;
877 EXPORT_SYMBOL(ptlrpc_prep_set);
880 * Allocate and initialize new request set structure with flow control
881 * extension. This extension allows to control the number of requests in-flight
882 * for the whole set. A callback function to generate requests must be provided
883 * and the request set will keep the number of requests sent over the wire to
885 * Returns a pointer to the newly allocated set structure or NULL on error.
887 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
891 struct ptlrpc_request_set *set;
893 set = ptlrpc_prep_set();
897 set->set_max_inflight = max;
898 set->set_producer = func;
899 set->set_producer_arg = arg;
903 EXPORT_SYMBOL(ptlrpc_prep_fcset);
906 * Wind down and free request set structure previously allocated with
908 * Ensures that all requests on the set have completed and removes
909 * all requests from the request list in a set.
910 * If any unsent request happen to be on the list, pretends that they got
911 * an error in flight and calls their completion handler.
913 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
921 /* Requests on the set should either all be completed, or all be new */
922 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
923 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
924 cfs_list_for_each (tmp, &set->set_requests) {
925 struct ptlrpc_request *req =
926 cfs_list_entry(tmp, struct ptlrpc_request,
929 LASSERT(req->rq_phase == expected_phase);
933 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
934 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
935 cfs_atomic_read(&set->set_remaining), n);
937 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
938 struct ptlrpc_request *req =
939 cfs_list_entry(tmp, struct ptlrpc_request,
941 cfs_list_del_init(&req->rq_set_chain);
943 LASSERT(req->rq_phase == expected_phase);
945 if (req->rq_phase == RQ_PHASE_NEW) {
946 ptlrpc_req_interpret(NULL, req, -EBADR);
947 cfs_atomic_dec(&set->set_remaining);
950 spin_lock(&req->rq_lock);
952 req->rq_invalid_rqset = 0;
953 spin_unlock(&req->rq_lock);
955 ptlrpc_req_finished (req);
958 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
960 ptlrpc_reqset_put(set);
963 EXPORT_SYMBOL(ptlrpc_set_destroy);
966 * Add a callback function \a fn to the set.
967 * This function would be called when all requests on this set are completed.
968 * The function will be passed \a data argument.
970 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
971 set_interpreter_func fn, void *data)
973 struct ptlrpc_set_cbdata *cbdata;
975 OBD_ALLOC_PTR(cbdata);
979 cbdata->psc_interpret = fn;
980 cbdata->psc_data = data;
981 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
985 EXPORT_SYMBOL(ptlrpc_set_add_cb);
988 * Add a new request to the general purpose request set.
989 * Assumes request reference from the caller.
991 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
992 struct ptlrpc_request *req)
994 LASSERT(cfs_list_empty(&req->rq_set_chain));
996 /* The set takes over the caller's request reference */
997 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
999 cfs_atomic_inc(&set->set_remaining);
1000 req->rq_queued_time = cfs_time_current();
1002 if (req->rq_reqmsg != NULL)
1003 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1005 if (set->set_producer != NULL)
1006 /* If the request set has a producer callback, the RPC must be
1007 * sent straight away */
1008 ptlrpc_send_new_req(req);
1010 EXPORT_SYMBOL(ptlrpc_set_add_req);
1013 * Add a request to a request with dedicated server thread
1014 * and wake the thread to make any necessary processing.
1015 * Currently only used for ptlrpcd.
1017 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1018 struct ptlrpc_request *req)
1020 struct ptlrpc_request_set *set = pc->pc_set;
1023 LASSERT(req->rq_set == NULL);
1024 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1026 spin_lock(&set->set_new_req_lock);
1028 * The set takes over the caller's request reference.
1031 req->rq_queued_time = cfs_time_current();
1032 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1033 count = cfs_atomic_inc_return(&set->set_new_count);
1034 spin_unlock(&set->set_new_req_lock);
1036 /* Only need to call wakeup once for the first entry. */
1038 wake_up(&set->set_waitq);
1040 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1041 * guarantee the async RPC can be processed ASAP, we have
1042 * no other better choice. It maybe fixed in future. */
1043 for (i = 0; i < pc->pc_npartners; i++)
1044 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1047 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1050 * Based on the current state of the import, determine if the request
1051 * can be sent, is an error, or should be delayed.
1053 * Returns true if this request should be delayed. If false, and
1054 * *status is set, then the request can not be sent and *status is the
1055 * error code. If false and status is 0, then request can be sent.
1057 * The imp->imp_lock must be held.
1059 static int ptlrpc_import_delay_req(struct obd_import *imp,
1060 struct ptlrpc_request *req, int *status)
1065 LASSERT (status != NULL);
1068 if (req->rq_ctx_init || req->rq_ctx_fini) {
1069 /* always allow ctx init/fini rpc go through */
1070 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1071 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1073 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1074 /* pings may safely race with umount */
1075 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1076 D_HA : D_ERROR, req, "IMP_CLOSED ");
1078 } else if (ptlrpc_send_limit_expired(req)) {
1079 /* probably doesn't need to be a D_ERROR after initial testing */
1080 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1082 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1083 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1084 /* allow CONNECT even if import is invalid */ ;
1085 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1086 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1089 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1090 if (!imp->imp_deactive)
1091 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1092 *status = -ESHUTDOWN; /* bz 12940 */
1093 } else if (req->rq_import_generation != imp->imp_generation) {
1094 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1096 } else if (req->rq_send_state != imp->imp_state) {
1097 /* invalidate in progress - any requests should be drop */
1098 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1099 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1101 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1102 *status = -EWOULDBLOCK;
1103 } else if (req->rq_allow_replay &&
1104 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1105 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1106 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1107 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1108 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1118 * Decide if the eror message regarding provided request \a req
1119 * should be printed to the console or not.
1120 * Makes it's decision on request status and other properties.
1121 * Returns 1 to print error on the system console or 0 if not.
1123 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1128 LASSERT(req->rq_reqmsg != NULL);
1129 opc = lustre_msg_get_opc(req->rq_reqmsg);
1131 /* Suppress particular reconnect errors which are to be expected. No
1132 * errors are suppressed for the initial connection on an import */
1133 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1134 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1136 /* Suppress timed out reconnect requests */
1137 if (req->rq_timedout)
1140 /* Suppress unavailable/again reconnect requests */
1141 err = lustre_msg_get_status(req->rq_repmsg);
1142 if (err == -ENODEV || err == -EAGAIN)
1150 * Check request processing status.
1151 * Returns the status.
1153 static int ptlrpc_check_status(struct ptlrpc_request *req)
1158 err = lustre_msg_get_status(req->rq_repmsg);
1159 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1160 struct obd_import *imp = req->rq_import;
1161 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1162 if (ptlrpc_console_allow(req))
1163 LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s,"
1164 " operation %s failed with %d.\n",
1165 imp->imp_obd->obd_name,
1167 imp->imp_connection->c_peer.nid),
1168 ll_opcode2str(opc), err);
1169 RETURN(err < 0 ? err : -EINVAL);
1173 DEBUG_REQ(D_INFO, req, "status is %d", err);
1174 } else if (err > 0) {
1175 /* XXX: translate this error from net to host */
1176 DEBUG_REQ(D_INFO, req, "status is %d", err);
1183 * save pre-versions of objects into request for replay.
1184 * Versions are obtained from server reply.
1187 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1189 struct lustre_msg *repmsg = req->rq_repmsg;
1190 struct lustre_msg *reqmsg = req->rq_reqmsg;
1191 __u64 *versions = lustre_msg_get_versions(repmsg);
1194 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1198 lustre_msg_set_versions(reqmsg, versions);
1199 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1200 versions[0], versions[1]);
1206 * Callback function called when client receives RPC reply for \a req.
1207 * Returns 0 on success or error code.
1208 * The return alue would be assigned to req->rq_status by the caller
1209 * as request processing status.
1210 * This function also decides if the request needs to be saved for later replay.
1212 static int after_reply(struct ptlrpc_request *req)
1214 struct obd_import *imp = req->rq_import;
1215 struct obd_device *obd = req->rq_import->imp_obd;
1217 struct timeval work_start;
1221 LASSERT(obd != NULL);
1222 /* repbuf must be unlinked */
1223 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1225 if (req->rq_reply_truncate) {
1226 if (ptlrpc_no_resend(req)) {
1227 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1228 " expected: %d, actual size: %d",
1229 req->rq_nob_received, req->rq_repbuf_len);
1233 sptlrpc_cli_free_repbuf(req);
1234 /* Pass the required reply buffer size (include
1235 * space for early reply).
1236 * NB: no need to roundup because alloc_repbuf
1237 * will roundup it */
1238 req->rq_replen = req->rq_nob_received;
1239 req->rq_nob_received = 0;
1245 * NB Until this point, the whole of the incoming message,
1246 * including buflens, status etc is in the sender's byte order.
1248 rc = sptlrpc_cli_unwrap_reply(req);
1250 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1255 * Security layer unwrap might ask resend this request.
1260 rc = unpack_reply(req);
1264 /* retry indefinitely on EINPROGRESS */
1265 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1266 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1267 time_t now = cfs_time_current_sec();
1269 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1271 req->rq_nr_resend++;
1273 /* allocate new xid to avoid reply reconstruction */
1274 if (!req->rq_bulk) {
1275 /* new xid is already allocated for bulk in
1276 * ptlrpc_check_set() */
1277 req->rq_xid = ptlrpc_next_xid();
1278 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1279 "resend on EINPROGRESS");
1282 /* Readjust the timeout for current conditions */
1283 ptlrpc_at_set_req_timeout(req);
1284 /* delay resend to give a chance to the server to get ready.
1285 * The delay is increased by 1s on every resend and is capped to
1286 * the current request timeout (i.e. obd_timeout if AT is off,
1287 * or AT service time x 125% + 5s, see at_est2timeout) */
1288 if (req->rq_nr_resend > req->rq_timeout)
1289 req->rq_sent = now + req->rq_timeout;
1291 req->rq_sent = now + req->rq_nr_resend;
1296 do_gettimeofday(&work_start);
1297 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1298 if (obd->obd_svc_stats != NULL) {
1299 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1301 ptlrpc_lprocfs_rpc_sent(req, timediff);
1304 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1305 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1306 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1307 lustre_msg_get_type(req->rq_repmsg));
1311 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1312 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1313 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1314 ptlrpc_at_adj_net_latency(req,
1315 lustre_msg_get_service_time(req->rq_repmsg));
1317 rc = ptlrpc_check_status(req);
1318 imp->imp_connect_error = rc;
1322 * Either we've been evicted, or the server has failed for
1323 * some reason. Try to reconnect, and if that fails, punt to
1326 if (ll_rpc_recoverable_error(rc)) {
1327 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1328 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1331 ptlrpc_request_handle_notconn(req);
1336 * Let's look if server sent slv. Do it only for RPC with
1339 ldlm_cli_update_pool(req);
1343 * Store transno in reqmsg for replay.
1345 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1346 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1347 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1350 if (imp->imp_replayable) {
1351 spin_lock(&imp->imp_lock);
1353 * No point in adding already-committed requests to the replay
1354 * list, we will just remove them immediately. b=9829
1356 if (req->rq_transno != 0 &&
1358 lustre_msg_get_last_committed(req->rq_repmsg) ||
1360 /** version recovery */
1361 ptlrpc_save_versions(req);
1362 ptlrpc_retain_replayable_request(req, imp);
1363 } else if (req->rq_commit_cb != NULL) {
1364 spin_unlock(&imp->imp_lock);
1365 req->rq_commit_cb(req);
1366 spin_lock(&imp->imp_lock);
1370 * Replay-enabled imports return commit-status information.
1372 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1373 imp->imp_peer_committed_transno =
1374 lustre_msg_get_last_committed(req->rq_repmsg);
1377 ptlrpc_free_committed(imp);
1379 if (!cfs_list_empty(&imp->imp_replay_list)) {
1380 struct ptlrpc_request *last;
1382 last = cfs_list_entry(imp->imp_replay_list.prev,
1383 struct ptlrpc_request,
1386 * Requests with rq_replay stay on the list even if no
1387 * commit is expected.
1389 if (last->rq_transno > imp->imp_peer_committed_transno)
1390 ptlrpc_pinger_commit_expected(imp);
1393 spin_unlock(&imp->imp_lock);
1400 * Helper function to send request \a req over the network for the first time
1401 * Also adjusts request phase.
1402 * Returns 0 on success or error code.
1404 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1406 struct obd_import *imp = req->rq_import;
1410 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1411 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1412 (!req->rq_generation_set ||
1413 req->rq_import_generation == imp->imp_generation))
1416 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1418 spin_lock(&imp->imp_lock);
1420 if (!req->rq_generation_set)
1421 req->rq_import_generation = imp->imp_generation;
1423 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1424 spin_lock(&req->rq_lock);
1425 req->rq_waiting = 1;
1426 spin_unlock(&req->rq_lock);
1428 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1429 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1430 ptlrpc_import_state_name(req->rq_send_state),
1431 ptlrpc_import_state_name(imp->imp_state));
1432 LASSERT(cfs_list_empty(&req->rq_list));
1433 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1434 cfs_atomic_inc(&req->rq_import->imp_inflight);
1435 spin_unlock(&imp->imp_lock);
1440 spin_unlock(&imp->imp_lock);
1441 req->rq_status = rc;
1442 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1446 LASSERT(cfs_list_empty(&req->rq_list));
1447 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1448 cfs_atomic_inc(&req->rq_import->imp_inflight);
1449 spin_unlock(&imp->imp_lock);
1451 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1453 rc = sptlrpc_req_refresh_ctx(req, -1);
1456 req->rq_status = rc;
1459 req->rq_wait_ctx = 1;
1464 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1465 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1466 imp->imp_obd->obd_uuid.uuid,
1467 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1468 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1469 lustre_msg_get_opc(req->rq_reqmsg));
1471 rc = ptl_send_rpc(req, 0);
1473 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1474 req->rq_net_err = 1;
1480 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1485 LASSERT(set->set_producer != NULL);
1487 remaining = cfs_atomic_read(&set->set_remaining);
1489 /* populate the ->set_requests list with requests until we
1490 * reach the maximum number of RPCs in flight for this set */
1491 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1492 rc = set->set_producer(set, set->set_producer_arg);
1493 if (rc == -ENOENT) {
1494 /* no more RPC to produce */
1495 set->set_producer = NULL;
1496 set->set_producer_arg = NULL;
1501 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1505 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1506 * and no more replies are expected.
1507 * (it is possible to get less replies than requests sent e.g. due to timed out
1508 * requests or requests that we had trouble to send out)
1510 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1512 cfs_list_t *tmp, *next;
1513 int force_timer_recalc = 0;
1516 if (cfs_atomic_read(&set->set_remaining) == 0)
1519 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1520 struct ptlrpc_request *req =
1521 cfs_list_entry(tmp, struct ptlrpc_request,
1523 struct obd_import *imp = req->rq_import;
1524 int unregistered = 0;
1527 if (req->rq_phase == RQ_PHASE_NEW &&
1528 ptlrpc_send_new_req(req)) {
1529 force_timer_recalc = 1;
1532 /* delayed send - skip */
1533 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1536 /* delayed resend - skip */
1537 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1538 req->rq_sent > cfs_time_current_sec())
1541 if (!(req->rq_phase == RQ_PHASE_RPC ||
1542 req->rq_phase == RQ_PHASE_BULK ||
1543 req->rq_phase == RQ_PHASE_INTERPRET ||
1544 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1545 req->rq_phase == RQ_PHASE_COMPLETE)) {
1546 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1550 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1551 LASSERT(req->rq_next_phase != req->rq_phase);
1552 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1555 * Skip processing until reply is unlinked. We
1556 * can't return to pool before that and we can't
1557 * call interpret before that. We need to make
1558 * sure that all rdma transfers finished and will
1559 * not corrupt any data.
1561 if (ptlrpc_client_recv_or_unlink(req) ||
1562 ptlrpc_client_bulk_active(req))
1566 * Turn fail_loc off to prevent it from looping
1569 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1570 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1573 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1574 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1579 * Move to next phase if reply was successfully
1582 ptlrpc_rqphase_move(req, req->rq_next_phase);
1585 if (req->rq_phase == RQ_PHASE_COMPLETE)
1588 if (req->rq_phase == RQ_PHASE_INTERPRET)
1589 GOTO(interpret, req->rq_status);
1592 * Note that this also will start async reply unlink.
1594 if (req->rq_net_err && !req->rq_timedout) {
1595 ptlrpc_expire_one_request(req, 1);
1598 * Check if we still need to wait for unlink.
1600 if (ptlrpc_client_recv_or_unlink(req) ||
1601 ptlrpc_client_bulk_active(req))
1603 /* If there is no need to resend, fail it now. */
1604 if (req->rq_no_resend) {
1605 if (req->rq_status == 0)
1606 req->rq_status = -EIO;
1607 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1608 GOTO(interpret, req->rq_status);
1615 spin_lock(&req->rq_lock);
1616 req->rq_replied = 0;
1617 spin_unlock(&req->rq_lock);
1618 if (req->rq_status == 0)
1619 req->rq_status = -EIO;
1620 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1621 GOTO(interpret, req->rq_status);
1624 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1625 * so it sets rq_intr regardless of individual rpc
1626 * timeouts. The synchronous IO waiting path sets
1627 * rq_intr irrespective of whether ptlrpcd
1628 * has seen a timeout. Our policy is to only interpret
1629 * interrupted rpcs after they have timed out, so we
1630 * need to enforce that here.
1633 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1634 req->rq_wait_ctx)) {
1635 req->rq_status = -EINTR;
1636 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1637 GOTO(interpret, req->rq_status);
1640 if (req->rq_phase == RQ_PHASE_RPC) {
1641 if (req->rq_timedout || req->rq_resend ||
1642 req->rq_waiting || req->rq_wait_ctx) {
1645 if (!ptlrpc_unregister_reply(req, 1))
1648 spin_lock(&imp->imp_lock);
1649 if (ptlrpc_import_delay_req(imp, req, &status)){
1650 /* put on delay list - only if we wait
1651 * recovery finished - before send */
1652 cfs_list_del_init(&req->rq_list);
1653 cfs_list_add_tail(&req->rq_list,
1656 spin_unlock(&imp->imp_lock);
1661 req->rq_status = status;
1662 ptlrpc_rqphase_move(req,
1663 RQ_PHASE_INTERPRET);
1664 spin_unlock(&imp->imp_lock);
1665 GOTO(interpret, req->rq_status);
1667 if (ptlrpc_no_resend(req) &&
1668 !req->rq_wait_ctx) {
1669 req->rq_status = -ENOTCONN;
1670 ptlrpc_rqphase_move(req,
1671 RQ_PHASE_INTERPRET);
1672 spin_unlock(&imp->imp_lock);
1673 GOTO(interpret, req->rq_status);
1676 cfs_list_del_init(&req->rq_list);
1677 cfs_list_add_tail(&req->rq_list,
1678 &imp->imp_sending_list);
1680 spin_unlock(&imp->imp_lock);
1682 spin_lock(&req->rq_lock);
1683 req->rq_waiting = 0;
1684 spin_unlock(&req->rq_lock);
1686 if (req->rq_timedout || req->rq_resend) {
1687 /* This is re-sending anyways,
1688 * let's mark req as resend. */
1689 spin_lock(&req->rq_lock);
1691 spin_unlock(&req->rq_lock);
1695 if (!ptlrpc_unregister_bulk(req, 1))
1698 /* ensure previous bulk fails */
1699 old_xid = req->rq_xid;
1700 req->rq_xid = ptlrpc_next_xid();
1701 CDEBUG(D_HA, "resend bulk "
1704 old_xid, req->rq_xid);
1708 * rq_wait_ctx is only touched by ptlrpcd,
1709 * so no lock is needed here.
1711 status = sptlrpc_req_refresh_ctx(req, -1);
1714 req->rq_status = status;
1715 spin_lock(&req->rq_lock);
1716 req->rq_wait_ctx = 0;
1717 spin_unlock(&req->rq_lock);
1718 force_timer_recalc = 1;
1720 spin_lock(&req->rq_lock);
1721 req->rq_wait_ctx = 1;
1722 spin_unlock(&req->rq_lock);
1727 spin_lock(&req->rq_lock);
1728 req->rq_wait_ctx = 0;
1729 spin_unlock(&req->rq_lock);
1732 rc = ptl_send_rpc(req, 0);
1734 DEBUG_REQ(D_HA, req,
1735 "send failed: rc = %d", rc);
1736 force_timer_recalc = 1;
1737 spin_lock(&req->rq_lock);
1738 req->rq_net_err = 1;
1739 spin_unlock(&req->rq_lock);
1742 /* need to reset the timeout */
1743 force_timer_recalc = 1;
1746 spin_lock(&req->rq_lock);
1748 if (ptlrpc_client_early(req)) {
1749 ptlrpc_at_recv_early_reply(req);
1750 spin_unlock(&req->rq_lock);
1754 /* Still waiting for a reply? */
1755 if (ptlrpc_client_recv(req)) {
1756 spin_unlock(&req->rq_lock);
1760 /* Did we actually receive a reply? */
1761 if (!ptlrpc_client_replied(req)) {
1762 spin_unlock(&req->rq_lock);
1766 spin_unlock(&req->rq_lock);
1768 /* unlink from net because we are going to
1769 * swab in-place of reply buffer */
1770 unregistered = ptlrpc_unregister_reply(req, 1);
1774 req->rq_status = after_reply(req);
1778 /* If there is no bulk associated with this request,
1779 * then we're done and should let the interpreter
1780 * process the reply. Similarly if the RPC returned
1781 * an error, and therefore the bulk will never arrive.
1783 if (req->rq_bulk == NULL || req->rq_status < 0) {
1784 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1785 GOTO(interpret, req->rq_status);
1788 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1791 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1792 if (ptlrpc_client_bulk_active(req))
1795 if (req->rq_bulk->bd_failure) {
1796 /* The RPC reply arrived OK, but the bulk screwed
1797 * up! Dead weird since the server told us the RPC
1798 * was good after getting the REPLY for her GET or
1799 * the ACK for her PUT. */
1800 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1801 req->rq_status = -EIO;
1804 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1807 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1809 /* This moves to "unregistering" phase we need to wait for
1811 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1812 /* start async bulk unlink too */
1813 ptlrpc_unregister_bulk(req, 1);
1817 if (!ptlrpc_unregister_bulk(req, 1))
1820 /* When calling interpret receiving already should be
1822 LASSERT(!req->rq_receiving_reply);
1824 ptlrpc_req_interpret(env, req, req->rq_status);
1826 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1828 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1829 "Completed RPC pname:cluuid:pid:xid:nid:"
1830 "opc %s:%s:%d:"LPU64":%s:%d\n",
1831 current_comm(), imp->imp_obd->obd_uuid.uuid,
1832 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1833 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1834 lustre_msg_get_opc(req->rq_reqmsg));
1836 spin_lock(&imp->imp_lock);
1837 /* Request already may be not on sending or delaying list. This
1838 * may happen in the case of marking it erroneous for the case
1839 * ptlrpc_import_delay_req(req, status) find it impossible to
1840 * allow sending this rpc and returns *status != 0. */
1841 if (!cfs_list_empty(&req->rq_list)) {
1842 cfs_list_del_init(&req->rq_list);
1843 cfs_atomic_dec(&imp->imp_inflight);
1845 spin_unlock(&imp->imp_lock);
1847 cfs_atomic_dec(&set->set_remaining);
1848 wake_up_all(&imp->imp_recovery_waitq);
1850 if (set->set_producer) {
1851 /* produce a new request if possible */
1852 if (ptlrpc_set_producer(set) > 0)
1853 force_timer_recalc = 1;
1855 /* free the request that has just been completed
1856 * in order not to pollute set->set_requests */
1857 cfs_list_del_init(&req->rq_set_chain);
1858 spin_lock(&req->rq_lock);
1860 req->rq_invalid_rqset = 0;
1861 spin_unlock(&req->rq_lock);
1863 /* record rq_status to compute the final status later */
1864 if (req->rq_status != 0)
1865 set->set_rc = req->rq_status;
1866 ptlrpc_req_finished(req);
1870 /* If we hit an error, we want to recover promptly. */
1871 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1873 EXPORT_SYMBOL(ptlrpc_check_set);
1876 * Time out request \a req. is \a async_unlink is set, that means do not wait
1877 * until LNet actually confirms network buffer unlinking.
1878 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1880 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1882 struct obd_import *imp = req->rq_import;
1886 spin_lock(&req->rq_lock);
1887 req->rq_timedout = 1;
1888 spin_unlock(&req->rq_lock);
1890 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1891 "/real "CFS_DURATION_T"]",
1892 req->rq_net_err ? "failed due to network error" :
1893 ((req->rq_real_sent == 0 ||
1894 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1895 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1896 "timed out for sent delay" : "timed out for slow reply"),
1897 req->rq_sent, req->rq_real_sent);
1899 if (imp != NULL && obd_debug_peer_on_timeout)
1900 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1902 ptlrpc_unregister_reply(req, async_unlink);
1903 ptlrpc_unregister_bulk(req, async_unlink);
1905 if (obd_dump_on_timeout)
1906 libcfs_debug_dumplog();
1909 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1913 cfs_atomic_inc(&imp->imp_timeouts);
1915 /* The DLM server doesn't want recovery run on its imports. */
1916 if (imp->imp_dlm_fake)
1919 /* If this request is for recovery or other primordial tasks,
1920 * then error it out here. */
1921 if (req->rq_ctx_init || req->rq_ctx_fini ||
1922 req->rq_send_state != LUSTRE_IMP_FULL ||
1923 imp->imp_obd->obd_no_recov) {
1924 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1925 ptlrpc_import_state_name(req->rq_send_state),
1926 ptlrpc_import_state_name(imp->imp_state));
1927 spin_lock(&req->rq_lock);
1928 req->rq_status = -ETIMEDOUT;
1930 spin_unlock(&req->rq_lock);
1934 /* if a request can't be resent we can't wait for an answer after
1936 if (ptlrpc_no_resend(req)) {
1937 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1941 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1947 * Time out all uncompleted requests in request set pointed by \a data
1948 * Callback used when waiting on sets with l_wait_event.
1951 int ptlrpc_expired_set(void *data)
1953 struct ptlrpc_request_set *set = data;
1955 time_t now = cfs_time_current_sec();
1958 LASSERT(set != NULL);
1961 * A timeout expired. See which reqs it applies to...
1963 cfs_list_for_each (tmp, &set->set_requests) {
1964 struct ptlrpc_request *req =
1965 cfs_list_entry(tmp, struct ptlrpc_request,
1968 /* don't expire request waiting for context */
1969 if (req->rq_wait_ctx)
1972 /* Request in-flight? */
1973 if (!((req->rq_phase == RQ_PHASE_RPC &&
1974 !req->rq_waiting && !req->rq_resend) ||
1975 (req->rq_phase == RQ_PHASE_BULK)))
1978 if (req->rq_timedout || /* already dealt with */
1979 req->rq_deadline > now) /* not expired */
1982 /* Deal with this guy. Do it asynchronously to not block
1983 * ptlrpcd thread. */
1984 ptlrpc_expire_one_request(req, 1);
1988 * When waiting for a whole set, we always break out of the
1989 * sleep so we can recalculate the timeout, or enable interrupts
1990 * if everyone's timed out.
1994 EXPORT_SYMBOL(ptlrpc_expired_set);
1997 * Sets rq_intr flag in \a req under spinlock.
1999 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2001 spin_lock(&req->rq_lock);
2003 spin_unlock(&req->rq_lock);
2005 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2008 * Interrupts (sets interrupted flag) all uncompleted requests in
2009 * a set \a data. Callback for l_wait_event for interruptible waits.
2011 void ptlrpc_interrupted_set(void *data)
2013 struct ptlrpc_request_set *set = data;
2016 LASSERT(set != NULL);
2017 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2019 cfs_list_for_each(tmp, &set->set_requests) {
2020 struct ptlrpc_request *req =
2021 cfs_list_entry(tmp, struct ptlrpc_request,
2024 if (req->rq_phase != RQ_PHASE_RPC &&
2025 req->rq_phase != RQ_PHASE_UNREGISTERING)
2028 ptlrpc_mark_interrupted(req);
2031 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2034 * Get the smallest timeout in the set; this does NOT set a timeout.
2036 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2039 time_t now = cfs_time_current_sec();
2041 struct ptlrpc_request *req;
2045 cfs_list_for_each(tmp, &set->set_requests) {
2046 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2049 * Request in-flight?
2051 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2052 (req->rq_phase == RQ_PHASE_BULK) ||
2053 (req->rq_phase == RQ_PHASE_NEW)))
2057 * Already timed out.
2059 if (req->rq_timedout)
2065 if (req->rq_wait_ctx)
2068 if (req->rq_phase == RQ_PHASE_NEW)
2069 deadline = req->rq_sent;
2070 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2071 deadline = req->rq_sent;
2073 deadline = req->rq_sent + req->rq_timeout;
2075 if (deadline <= now) /* actually expired already */
2076 timeout = 1; /* ASAP */
2077 else if (timeout == 0 || timeout > deadline - now)
2078 timeout = deadline - now;
2082 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2085 * Send all unset request from the set and then wait untill all
2086 * requests in the set complete (either get a reply, timeout, get an
2087 * error or otherwise be interrupted).
2088 * Returns 0 on success or error code otherwise.
2090 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2093 struct ptlrpc_request *req;
2094 struct l_wait_info lwi;
2098 if (set->set_producer)
2099 (void)ptlrpc_set_producer(set);
2101 cfs_list_for_each(tmp, &set->set_requests) {
2102 req = cfs_list_entry(tmp, struct ptlrpc_request,
2104 if (req->rq_phase == RQ_PHASE_NEW)
2105 (void)ptlrpc_send_new_req(req);
2108 if (cfs_list_empty(&set->set_requests))
2112 timeout = ptlrpc_set_next_timeout(set);
2114 /* wait until all complete, interrupted, or an in-flight
2116 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2119 if (timeout == 0 && !cfs_signal_pending())
2121 * No requests are in-flight (ether timed out
2122 * or delayed), so we can allow interrupts.
2123 * We still want to block for a limited time,
2124 * so we allow interrupts during the timeout.
2126 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2128 ptlrpc_interrupted_set, set);
2131 * At least one request is in flight, so no
2132 * interrupts are allowed. Wait until all
2133 * complete, or an in-flight req times out.
2135 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2136 ptlrpc_expired_set, set);
2138 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2140 /* LU-769 - if we ignored the signal because it was already
2141 * pending when we started, we need to handle it now or we risk
2142 * it being ignored forever */
2143 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2144 cfs_signal_pending()) {
2145 cfs_sigset_t blocked_sigs =
2146 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2148 /* In fact we only interrupt for the "fatal" signals
2149 * like SIGINT or SIGKILL. We still ignore less
2150 * important signals since ptlrpc set is not easily
2151 * reentrant from userspace again */
2152 if (cfs_signal_pending())
2153 ptlrpc_interrupted_set(set);
2154 cfs_restore_sigs(blocked_sigs);
2157 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2159 /* -EINTR => all requests have been flagged rq_intr so next
2161 * -ETIMEDOUT => someone timed out. When all reqs have
2162 * timed out, signals are enabled allowing completion with
2164 * I don't really care if we go once more round the loop in
2165 * the error cases -eeb. */
2166 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2167 cfs_list_for_each(tmp, &set->set_requests) {
2168 req = cfs_list_entry(tmp, struct ptlrpc_request,
2170 spin_lock(&req->rq_lock);
2171 req->rq_invalid_rqset = 1;
2172 spin_unlock(&req->rq_lock);
2175 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2177 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2179 rc = set->set_rc; /* rq_status of already freed requests if any */
2180 cfs_list_for_each(tmp, &set->set_requests) {
2181 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2183 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2184 if (req->rq_status != 0)
2185 rc = req->rq_status;
2188 if (set->set_interpret != NULL) {
2189 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2191 rc = interpreter (set, set->set_arg, rc);
2193 struct ptlrpc_set_cbdata *cbdata, *n;
2196 cfs_list_for_each_entry_safe(cbdata, n,
2197 &set->set_cblist, psc_item) {
2198 cfs_list_del_init(&cbdata->psc_item);
2199 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2202 OBD_FREE_PTR(cbdata);
2208 EXPORT_SYMBOL(ptlrpc_set_wait);
2211 * Helper fuction for request freeing.
2212 * Called when request count reached zero and request needs to be freed.
2213 * Removes request from all sorts of sending/replay lists it might be on,
2214 * frees network buffers if any are present.
2215 * If \a locked is set, that means caller is already holding import imp_lock
2216 * and so we no longer need to reobtain it (for certain lists manipulations)
2218 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2221 if (request == NULL) {
2226 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2227 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2228 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2229 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2230 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2231 LASSERTF(!request->rq_replay, "req %p\n", request);
2233 req_capsule_fini(&request->rq_pill);
2235 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2236 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2237 if (request->rq_import != NULL) {
2239 spin_lock(&request->rq_import->imp_lock);
2240 cfs_list_del_init(&request->rq_replay_list);
2242 spin_unlock(&request->rq_import->imp_lock);
2244 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2246 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2247 DEBUG_REQ(D_ERROR, request,
2248 "freeing request with nonzero refcount");
2252 if (request->rq_repbuf != NULL)
2253 sptlrpc_cli_free_repbuf(request);
2254 if (request->rq_export != NULL) {
2255 class_export_put(request->rq_export);
2256 request->rq_export = NULL;
2258 if (request->rq_import != NULL) {
2259 class_import_put(request->rq_import);
2260 request->rq_import = NULL;
2262 if (request->rq_bulk != NULL)
2263 ptlrpc_free_bulk_pin(request->rq_bulk);
2265 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2266 sptlrpc_cli_free_reqbuf(request);
2268 if (request->rq_cli_ctx)
2269 sptlrpc_req_put_ctx(request, !locked);
2271 if (request->rq_pool)
2272 __ptlrpc_free_req_to_pool(request);
2274 ptlrpc_request_cache_free(request);
2278 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2280 * Drop one request reference. Must be called with import imp_lock held.
2281 * When reference count drops to zero, reuqest is freed.
2283 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2285 LASSERT(spin_is_locked(&request->rq_import->imp_lock));
2286 (void)__ptlrpc_req_finished(request, 1);
2288 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2292 * Drops one reference count for request \a request.
2293 * \a locked set indicates that caller holds import imp_lock.
2294 * Frees the request whe reference count reaches zero.
2296 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2299 if (request == NULL)
2302 if (request == LP_POISON ||
2303 request->rq_reqmsg == LP_POISON) {
2304 CERROR("dereferencing freed request (bug 575)\n");
2309 DEBUG_REQ(D_INFO, request, "refcount now %u",
2310 cfs_atomic_read(&request->rq_refcount) - 1);
2312 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2313 __ptlrpc_free_req(request, locked);
2321 * Drops one reference count for a request.
2323 void ptlrpc_req_finished(struct ptlrpc_request *request)
2325 __ptlrpc_req_finished(request, 0);
2327 EXPORT_SYMBOL(ptlrpc_req_finished);
2330 * Returns xid of a \a request
2332 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2334 return request->rq_xid;
2336 EXPORT_SYMBOL(ptlrpc_req_xid);
2339 * Disengage the client's reply buffer from the network
2340 * NB does _NOT_ unregister any client-side bulk.
2341 * IDEMPOTENT, but _not_ safe against concurrent callers.
2342 * The request owner (i.e. the thread doing the I/O) must call...
2343 * Returns 0 on success or 1 if unregistering cannot be made.
2345 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2348 struct l_wait_info lwi;
2353 LASSERT(!cfs_in_interrupt());
2356 * Let's setup deadline for reply unlink.
2358 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2359 async && request->rq_reply_deadline == 0)
2360 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2363 * Nothing left to do.
2365 if (!ptlrpc_client_recv_or_unlink(request))
2368 LNetMDUnlink(request->rq_reply_md_h);
2371 * Let's check it once again.
2373 if (!ptlrpc_client_recv_or_unlink(request))
2377 * Move to "Unregistering" phase as reply was not unlinked yet.
2379 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2382 * Do not wait for unlink to finish.
2388 * We have to l_wait_event() whatever the result, to give liblustre
2389 * a chance to run reply_in_callback(), and to make sure we've
2390 * unlinked before returning a req to the pool.
2394 /* The wq argument is ignored by user-space wait_event macros */
2395 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2396 &request->rq_set->set_waitq :
2397 &request->rq_reply_waitq;
2399 /* Network access will complete in finite time but the HUGE
2400 * timeout lets us CWARN for visibility of sluggish NALs */
2401 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2402 cfs_time_seconds(1), NULL, NULL);
2403 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2406 ptlrpc_rqphase_move(request, request->rq_next_phase);
2410 LASSERT(rc == -ETIMEDOUT);
2411 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2412 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2413 request->rq_must_unlink);
2417 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2420 * Iterates through replay_list on import and prunes
2421 * all requests have transno smaller than last_committed for the
2422 * import and don't have rq_replay set.
2423 * Since requests are sorted in transno order, stops when meetign first
2424 * transno bigger than last_committed.
2425 * caller must hold imp->imp_lock
2427 void ptlrpc_free_committed(struct obd_import *imp)
2429 cfs_list_t *tmp, *saved;
2430 struct ptlrpc_request *req;
2431 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2434 LASSERT(imp != NULL);
2435 LASSERT(spin_is_locked(&imp->imp_lock));
2438 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2439 imp->imp_generation == imp->imp_last_generation_checked) {
2440 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2441 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2445 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2446 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2447 imp->imp_generation);
2448 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2449 imp->imp_last_generation_checked = imp->imp_generation;
2451 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2452 req = cfs_list_entry(tmp, struct ptlrpc_request,
2455 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2456 LASSERT(req != last_req);
2459 if (req->rq_transno == 0) {
2460 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2463 if (req->rq_import_generation < imp->imp_generation) {
2464 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2468 if (req->rq_replay) {
2469 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2473 /* not yet committed */
2474 if (req->rq_transno > imp->imp_peer_committed_transno) {
2475 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2479 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2480 imp->imp_peer_committed_transno);
2482 spin_lock(&req->rq_lock);
2484 spin_unlock(&req->rq_lock);
2485 if (req->rq_commit_cb != NULL)
2486 req->rq_commit_cb(req);
2487 cfs_list_del_init(&req->rq_replay_list);
2488 __ptlrpc_req_finished(req, 1);
2495 void ptlrpc_cleanup_client(struct obd_import *imp)
2501 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2504 * Schedule previously sent request for resend.
2505 * For bulk requests we assign new xid (to avoid problems with
2506 * lost replies and therefore several transfers landing into same buffer
2507 * from different sending attempts).
2509 void ptlrpc_resend_req(struct ptlrpc_request *req)
2511 DEBUG_REQ(D_HA, req, "going to resend");
2512 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2513 req->rq_status = -EAGAIN;
2515 spin_lock(&req->rq_lock);
2517 req->rq_net_err = 0;
2518 req->rq_timedout = 0;
2520 __u64 old_xid = req->rq_xid;
2522 /* ensure previous bulk fails */
2523 req->rq_xid = ptlrpc_next_xid();
2524 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2525 old_xid, req->rq_xid);
2527 ptlrpc_client_wake_req(req);
2528 spin_unlock(&req->rq_lock);
2530 EXPORT_SYMBOL(ptlrpc_resend_req);
2532 /* XXX: this function and rq_status are currently unused */
2533 void ptlrpc_restart_req(struct ptlrpc_request *req)
2535 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2536 req->rq_status = -ERESTARTSYS;
2538 spin_lock(&req->rq_lock);
2539 req->rq_restart = 1;
2540 req->rq_timedout = 0;
2541 ptlrpc_client_wake_req(req);
2542 spin_unlock(&req->rq_lock);
2544 EXPORT_SYMBOL(ptlrpc_restart_req);
2547 * Grab additional reference on a request \a req
2549 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2552 cfs_atomic_inc(&req->rq_refcount);
2555 EXPORT_SYMBOL(ptlrpc_request_addref);
2558 * Add a request to import replay_list.
2559 * Must be called under imp_lock
2561 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2562 struct obd_import *imp)
2566 LASSERT(spin_is_locked(&imp->imp_lock));
2568 if (req->rq_transno == 0) {
2569 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2573 /* clear this for new requests that were resent as well
2574 as resent replayed requests. */
2575 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2577 /* don't re-add requests that have been replayed */
2578 if (!cfs_list_empty(&req->rq_replay_list))
2581 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2583 LASSERT(imp->imp_replayable);
2584 /* Balanced in ptlrpc_free_committed, usually. */
2585 ptlrpc_request_addref(req);
2586 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2587 struct ptlrpc_request *iter =
2588 cfs_list_entry(tmp, struct ptlrpc_request,
2591 /* We may have duplicate transnos if we create and then
2592 * open a file, or for closes retained if to match creating
2593 * opens, so use req->rq_xid as a secondary key.
2594 * (See bugs 684, 685, and 428.)
2595 * XXX no longer needed, but all opens need transnos!
2597 if (iter->rq_transno > req->rq_transno)
2600 if (iter->rq_transno == req->rq_transno) {
2601 LASSERT(iter->rq_xid != req->rq_xid);
2602 if (iter->rq_xid > req->rq_xid)
2606 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2610 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2612 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2615 * Send request and wait until it completes.
2616 * Returns request processing status.
2618 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2620 struct ptlrpc_request_set *set;
2624 LASSERT(req->rq_set == NULL);
2625 LASSERT(!req->rq_receiving_reply);
2627 set = ptlrpc_prep_set();
2629 CERROR("Unable to allocate ptlrpc set.");
2633 /* for distributed debugging */
2634 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2636 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2637 ptlrpc_request_addref(req);
2638 ptlrpc_set_add_req(set, req);
2639 rc = ptlrpc_set_wait(set);
2640 ptlrpc_set_destroy(set);
2644 EXPORT_SYMBOL(ptlrpc_queue_wait);
2646 struct ptlrpc_replay_async_args {
2648 int praa_old_status;
2652 * Callback used for replayed requests reply processing.
2653 * In case of succesful reply calls registeresd request replay callback.
2654 * In case of error restart replay process.
2656 static int ptlrpc_replay_interpret(const struct lu_env *env,
2657 struct ptlrpc_request *req,
2658 void * data, int rc)
2660 struct ptlrpc_replay_async_args *aa = data;
2661 struct obd_import *imp = req->rq_import;
2664 cfs_atomic_dec(&imp->imp_replay_inflight);
2666 if (!ptlrpc_client_replied(req)) {
2667 CERROR("request replay timed out, restarting recovery\n");
2668 GOTO(out, rc = -ETIMEDOUT);
2671 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2672 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2673 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2674 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2676 /** VBR: check version failure */
2677 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2678 /** replay was failed due to version mismatch */
2679 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2680 spin_lock(&imp->imp_lock);
2681 imp->imp_vbr_failed = 1;
2682 imp->imp_no_lock_replay = 1;
2683 spin_unlock(&imp->imp_lock);
2684 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2686 /** The transno had better not change over replay. */
2687 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2688 lustre_msg_get_transno(req->rq_repmsg) ||
2689 lustre_msg_get_transno(req->rq_repmsg) == 0,
2691 lustre_msg_get_transno(req->rq_reqmsg),
2692 lustre_msg_get_transno(req->rq_repmsg));
2695 spin_lock(&imp->imp_lock);
2696 /** if replays by version then gap occur on server, no trust to locks */
2697 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2698 imp->imp_no_lock_replay = 1;
2699 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2700 spin_unlock(&imp->imp_lock);
2701 LASSERT(imp->imp_last_replay_transno);
2703 /* transaction number shouldn't be bigger than the latest replayed */
2704 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2705 DEBUG_REQ(D_ERROR, req,
2706 "Reported transno "LPU64" is bigger than the "
2707 "replayed one: "LPU64, req->rq_transno,
2708 lustre_msg_get_transno(req->rq_reqmsg));
2709 GOTO(out, rc = -EINVAL);
2712 DEBUG_REQ(D_HA, req, "got rep");
2714 /* let the callback do fixups, possibly including in the request */
2715 if (req->rq_replay_cb)
2716 req->rq_replay_cb(req);
2718 if (ptlrpc_client_replied(req) &&
2719 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2720 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2721 lustre_msg_get_status(req->rq_repmsg),
2722 aa->praa_old_status);
2724 /* Put it back for re-replay. */
2725 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2729 * Errors while replay can set transno to 0, but
2730 * imp_last_replay_transno shouldn't be set to 0 anyway
2732 if (req->rq_transno == 0)
2733 CERROR("Transno is 0 during replay!\n");
2735 /* continue with recovery */
2736 rc = ptlrpc_import_recovery_state_machine(imp);
2738 req->rq_send_state = aa->praa_old_state;
2741 /* this replay failed, so restart recovery */
2742 ptlrpc_connect_import(imp);
2748 * Prepares and queues request for replay.
2749 * Adds it to ptlrpcd queue for actual sending.
2750 * Returns 0 on success.
2752 int ptlrpc_replay_req(struct ptlrpc_request *req)
2754 struct ptlrpc_replay_async_args *aa;
2757 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2759 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2760 aa = ptlrpc_req_async_args(req);
2761 memset(aa, 0, sizeof *aa);
2763 /* Prepare request to be resent with ptlrpcd */
2764 aa->praa_old_state = req->rq_send_state;
2765 req->rq_send_state = LUSTRE_IMP_REPLAY;
2766 req->rq_phase = RQ_PHASE_NEW;
2767 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2769 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2771 req->rq_interpret_reply = ptlrpc_replay_interpret;
2772 /* Readjust the timeout for current conditions */
2773 ptlrpc_at_set_req_timeout(req);
2775 /* Tell server the net_latency, so the server can calculate how long
2776 * it should wait for next replay */
2777 lustre_msg_set_service_time(req->rq_reqmsg,
2778 ptlrpc_at_get_net_latency(req));
2779 DEBUG_REQ(D_HA, req, "REPLAY");
2781 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2782 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2784 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2787 EXPORT_SYMBOL(ptlrpc_replay_req);
2790 * Aborts all in-flight request on import \a imp sending and delayed lists
2792 void ptlrpc_abort_inflight(struct obd_import *imp)
2794 cfs_list_t *tmp, *n;
2797 /* Make sure that no new requests get processed for this import.
2798 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2799 * this flag and then putting requests on sending_list or delayed_list.
2801 spin_lock(&imp->imp_lock);
2803 /* XXX locking? Maybe we should remove each request with the list
2804 * locked? Also, how do we know if the requests on the list are
2805 * being freed at this time?
2807 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2808 struct ptlrpc_request *req =
2809 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2811 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2813 spin_lock(&req->rq_lock);
2814 if (req->rq_import_generation < imp->imp_generation) {
2816 req->rq_status = -EIO;
2817 ptlrpc_client_wake_req(req);
2819 spin_unlock(&req->rq_lock);
2822 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2823 struct ptlrpc_request *req =
2824 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2826 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2828 spin_lock(&req->rq_lock);
2829 if (req->rq_import_generation < imp->imp_generation) {
2831 req->rq_status = -EIO;
2832 ptlrpc_client_wake_req(req);
2834 spin_unlock(&req->rq_lock);
2837 /* Last chance to free reqs left on the replay list, but we
2838 * will still leak reqs that haven't committed. */
2839 if (imp->imp_replayable)
2840 ptlrpc_free_committed(imp);
2842 spin_unlock(&imp->imp_lock);
2846 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2849 * Abort all uncompleted requests in request set \a set
2851 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2853 cfs_list_t *tmp, *pos;
2855 LASSERT(set != NULL);
2857 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2858 struct ptlrpc_request *req =
2859 cfs_list_entry(pos, struct ptlrpc_request,
2862 spin_lock(&req->rq_lock);
2863 if (req->rq_phase != RQ_PHASE_RPC) {
2864 spin_unlock(&req->rq_lock);
2869 req->rq_status = -EINTR;
2870 ptlrpc_client_wake_req(req);
2871 spin_unlock(&req->rq_lock);
2875 static __u64 ptlrpc_last_xid;
2876 static spinlock_t ptlrpc_last_xid_lock;
2879 * Initialize the XID for the node. This is common among all requests on
2880 * this node, and only requires the property that it is monotonically
2881 * increasing. It does not need to be sequential. Since this is also used
2882 * as the RDMA match bits, it is important that a single client NOT have
2883 * the same match bits for two different in-flight requests, hence we do
2884 * NOT want to have an XID per target or similar.
2886 * To avoid an unlikely collision between match bits after a client reboot
2887 * (which would deliver old data into the wrong RDMA buffer) initialize
2888 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2889 * If the time is clearly incorrect, we instead use a 62-bit random number.
2890 * In the worst case the random number will overflow 1M RPCs per second in
2891 * 9133 years, or permutations thereof.
2893 #define YEAR_2004 (1ULL << 30)
2894 void ptlrpc_init_xid(void)
2896 time_t now = cfs_time_current_sec();
2898 spin_lock_init(&ptlrpc_last_xid_lock);
2899 if (now < YEAR_2004) {
2900 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2901 ptlrpc_last_xid >>= 2;
2902 ptlrpc_last_xid |= (1ULL << 61);
2904 ptlrpc_last_xid = (__u64)now << 20;
2907 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
2908 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
2909 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
2913 * Increase xid and returns resulting new value to the caller.
2915 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2916 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2917 * itself uses the last bulk xid needed, so the server can determine the
2918 * the number of bulk transfers from the RPC XID and a bitmask. The starting
2919 * xid must align to a power-of-two value.
2921 * This is assumed to be true due to the initial ptlrpc_last_xid
2922 * value also being initialized to a power-of-two value. LU-1431
2924 __u64 ptlrpc_next_xid(void)
2928 spin_lock(&ptlrpc_last_xid_lock);
2929 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2930 ptlrpc_last_xid = next;
2931 spin_unlock(&ptlrpc_last_xid_lock);
2935 EXPORT_SYMBOL(ptlrpc_next_xid);
2938 * Get a glimpse at what next xid value might have been.
2939 * Returns possible next xid.
2941 __u64 ptlrpc_sample_next_xid(void)
2943 #if BITS_PER_LONG == 32
2944 /* need to avoid possible word tearing on 32-bit systems */
2947 spin_lock(&ptlrpc_last_xid_lock);
2948 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2949 spin_unlock(&ptlrpc_last_xid_lock);
2953 /* No need to lock, since returned value is racy anyways */
2954 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2957 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2960 * Functions for operating ptlrpc workers.
2962 * A ptlrpc work is a function which will be running inside ptlrpc context.
2963 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2965 * 1. after a work is created, it can be used many times, that is:
2966 * handler = ptlrpcd_alloc_work();
2967 * ptlrpcd_queue_work();
2969 * queue it again when necessary:
2970 * ptlrpcd_queue_work();
2971 * ptlrpcd_destroy_work();
2972 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2973 * it will only be queued once in any time. Also as its name implies, it may
2974 * have delay before it really runs by ptlrpcd thread.
2976 struct ptlrpc_work_async_args {
2978 int (*cb)(const struct lu_env *, void *);
2982 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2984 static int work_interpreter(const struct lu_env *env,
2985 struct ptlrpc_request *req, void *data, int rc)
2987 struct ptlrpc_work_async_args *arg = data;
2989 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2990 LASSERT(arg->cb != NULL);
2992 return arg->cb(env, arg->cbdata);
2996 * Create a work for ptlrpc.
2998 void *ptlrpcd_alloc_work(struct obd_import *imp,
2999 int (*cb)(const struct lu_env *, void *), void *cbdata)
3001 struct ptlrpc_request *req = NULL;
3002 struct ptlrpc_work_async_args *args;
3008 RETURN(ERR_PTR(-EINVAL));
3010 /* copy some code from deprecated fakereq. */
3011 req = ptlrpc_request_cache_alloc(__GFP_IO);
3013 CERROR("ptlrpc: run out of memory!\n");
3014 RETURN(ERR_PTR(-ENOMEM));
3017 req->rq_send_state = LUSTRE_IMP_FULL;
3018 req->rq_type = PTL_RPC_MSG_REQUEST;
3019 req->rq_import = class_import_get(imp);
3020 req->rq_export = NULL;
3021 req->rq_interpret_reply = work_interpreter;
3022 /* don't want reply */
3023 req->rq_receiving_reply = 0;
3024 req->rq_must_unlink = 0;
3025 req->rq_no_delay = req->rq_no_resend = 1;
3027 spin_lock_init(&req->rq_lock);
3028 CFS_INIT_LIST_HEAD(&req->rq_list);
3029 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
3030 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
3031 CFS_INIT_LIST_HEAD(&req->rq_history_list);
3032 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
3033 init_waitqueue_head(&req->rq_reply_waitq);
3034 init_waitqueue_head(&req->rq_set_waitq);
3035 cfs_atomic_set(&req->rq_refcount, 1);
3037 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3038 args = ptlrpc_req_async_args(req);
3039 args->magic = PTLRPC_WORK_MAGIC;
3041 args->cbdata = cbdata;
3045 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3047 void ptlrpcd_destroy_work(void *handler)
3049 struct ptlrpc_request *req = handler;
3052 ptlrpc_req_finished(req);
3054 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3056 int ptlrpcd_queue_work(void *handler)
3058 struct ptlrpc_request *req = handler;
3061 * Check if the req is already being queued.
3063 * Here comes a trick: it lacks a way of checking if a req is being
3064 * processed reliably in ptlrpc. Here I have to use refcount of req
3065 * for this purpose. This is okay because the caller should use this
3066 * req as opaque data. - Jinshan
3068 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
3069 if (cfs_atomic_read(&req->rq_refcount) > 1)
3072 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
3073 cfs_atomic_dec(&req->rq_refcount);
3077 /* re-initialize the req */
3078 req->rq_timeout = obd_timeout;
3079 req->rq_sent = cfs_time_current_sec();
3080 req->rq_deadline = req->rq_sent + req->rq_timeout;
3081 req->rq_reply_deadline = req->rq_deadline;
3082 req->rq_phase = RQ_PHASE_INTERPRET;
3083 req->rq_next_phase = RQ_PHASE_COMPLETE;
3084 req->rq_xid = ptlrpc_next_xid();
3085 req->rq_import_generation = req->rq_import->imp_generation;
3087 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3090 EXPORT_SYMBOL(ptlrpcd_queue_work);