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;
1240 spin_lock(&req->rq_lock);
1242 spin_unlock(&req->rq_lock);
1247 * NB Until this point, the whole of the incoming message,
1248 * including buflens, status etc is in the sender's byte order.
1250 rc = sptlrpc_cli_unwrap_reply(req);
1252 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1257 * Security layer unwrap might ask resend this request.
1262 rc = unpack_reply(req);
1266 /* retry indefinitely on EINPROGRESS */
1267 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1268 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1269 time_t now = cfs_time_current_sec();
1271 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1273 req->rq_nr_resend++;
1275 /* allocate new xid to avoid reply reconstruction */
1276 if (!req->rq_bulk) {
1277 /* new xid is already allocated for bulk in
1278 * ptlrpc_check_set() */
1279 req->rq_xid = ptlrpc_next_xid();
1280 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1281 "resend on EINPROGRESS");
1284 /* Readjust the timeout for current conditions */
1285 ptlrpc_at_set_req_timeout(req);
1286 /* delay resend to give a chance to the server to get ready.
1287 * The delay is increased by 1s on every resend and is capped to
1288 * the current request timeout (i.e. obd_timeout if AT is off,
1289 * or AT service time x 125% + 5s, see at_est2timeout) */
1290 if (req->rq_nr_resend > req->rq_timeout)
1291 req->rq_sent = now + req->rq_timeout;
1293 req->rq_sent = now + req->rq_nr_resend;
1298 do_gettimeofday(&work_start);
1299 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1300 if (obd->obd_svc_stats != NULL) {
1301 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1303 ptlrpc_lprocfs_rpc_sent(req, timediff);
1306 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1307 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1308 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1309 lustre_msg_get_type(req->rq_repmsg));
1313 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1314 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1315 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1316 ptlrpc_at_adj_net_latency(req,
1317 lustre_msg_get_service_time(req->rq_repmsg));
1319 rc = ptlrpc_check_status(req);
1320 imp->imp_connect_error = rc;
1324 * Either we've been evicted, or the server has failed for
1325 * some reason. Try to reconnect, and if that fails, punt to
1328 if (ll_rpc_recoverable_error(rc)) {
1329 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1330 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1333 ptlrpc_request_handle_notconn(req);
1338 * Let's look if server sent slv. Do it only for RPC with
1341 ldlm_cli_update_pool(req);
1345 * Store transno in reqmsg for replay.
1347 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1348 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1349 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1352 if (imp->imp_replayable) {
1353 spin_lock(&imp->imp_lock);
1355 * No point in adding already-committed requests to the replay
1356 * list, we will just remove them immediately. b=9829
1358 if (req->rq_transno != 0 &&
1360 lustre_msg_get_last_committed(req->rq_repmsg) ||
1362 /** version recovery */
1363 ptlrpc_save_versions(req);
1364 ptlrpc_retain_replayable_request(req, imp);
1365 } else if (req->rq_commit_cb != NULL) {
1366 spin_unlock(&imp->imp_lock);
1367 req->rq_commit_cb(req);
1368 spin_lock(&imp->imp_lock);
1372 * Replay-enabled imports return commit-status information.
1374 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1375 imp->imp_peer_committed_transno =
1376 lustre_msg_get_last_committed(req->rq_repmsg);
1379 ptlrpc_free_committed(imp);
1381 if (!cfs_list_empty(&imp->imp_replay_list)) {
1382 struct ptlrpc_request *last;
1384 last = cfs_list_entry(imp->imp_replay_list.prev,
1385 struct ptlrpc_request,
1388 * Requests with rq_replay stay on the list even if no
1389 * commit is expected.
1391 if (last->rq_transno > imp->imp_peer_committed_transno)
1392 ptlrpc_pinger_commit_expected(imp);
1395 spin_unlock(&imp->imp_lock);
1402 * Helper function to send request \a req over the network for the first time
1403 * Also adjusts request phase.
1404 * Returns 0 on success or error code.
1406 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1408 struct obd_import *imp = req->rq_import;
1412 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1413 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1414 (!req->rq_generation_set ||
1415 req->rq_import_generation == imp->imp_generation))
1418 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1420 spin_lock(&imp->imp_lock);
1422 if (!req->rq_generation_set)
1423 req->rq_import_generation = imp->imp_generation;
1425 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1426 spin_lock(&req->rq_lock);
1427 req->rq_waiting = 1;
1428 spin_unlock(&req->rq_lock);
1430 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1431 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1432 ptlrpc_import_state_name(req->rq_send_state),
1433 ptlrpc_import_state_name(imp->imp_state));
1434 LASSERT(cfs_list_empty(&req->rq_list));
1435 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1436 cfs_atomic_inc(&req->rq_import->imp_inflight);
1437 spin_unlock(&imp->imp_lock);
1442 spin_unlock(&imp->imp_lock);
1443 req->rq_status = rc;
1444 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1448 LASSERT(cfs_list_empty(&req->rq_list));
1449 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1450 cfs_atomic_inc(&req->rq_import->imp_inflight);
1451 spin_unlock(&imp->imp_lock);
1453 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1455 rc = sptlrpc_req_refresh_ctx(req, -1);
1458 req->rq_status = rc;
1461 spin_lock(&req->rq_lock);
1462 req->rq_wait_ctx = 1;
1463 spin_unlock(&req->rq_lock);
1468 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1469 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1470 imp->imp_obd->obd_uuid.uuid,
1471 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1472 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1473 lustre_msg_get_opc(req->rq_reqmsg));
1475 rc = ptl_send_rpc(req, 0);
1477 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1478 spin_lock(&req->rq_lock);
1479 req->rq_net_err = 1;
1480 spin_unlock(&req->rq_lock);
1486 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1491 LASSERT(set->set_producer != NULL);
1493 remaining = cfs_atomic_read(&set->set_remaining);
1495 /* populate the ->set_requests list with requests until we
1496 * reach the maximum number of RPCs in flight for this set */
1497 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1498 rc = set->set_producer(set, set->set_producer_arg);
1499 if (rc == -ENOENT) {
1500 /* no more RPC to produce */
1501 set->set_producer = NULL;
1502 set->set_producer_arg = NULL;
1507 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1511 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1512 * and no more replies are expected.
1513 * (it is possible to get less replies than requests sent e.g. due to timed out
1514 * requests or requests that we had trouble to send out)
1516 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1518 cfs_list_t *tmp, *next;
1519 int force_timer_recalc = 0;
1522 if (cfs_atomic_read(&set->set_remaining) == 0)
1525 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1526 struct ptlrpc_request *req =
1527 cfs_list_entry(tmp, struct ptlrpc_request,
1529 struct obd_import *imp = req->rq_import;
1530 int unregistered = 0;
1533 if (req->rq_phase == RQ_PHASE_NEW &&
1534 ptlrpc_send_new_req(req)) {
1535 force_timer_recalc = 1;
1538 /* delayed send - skip */
1539 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1542 /* delayed resend - skip */
1543 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1544 req->rq_sent > cfs_time_current_sec())
1547 if (!(req->rq_phase == RQ_PHASE_RPC ||
1548 req->rq_phase == RQ_PHASE_BULK ||
1549 req->rq_phase == RQ_PHASE_INTERPRET ||
1550 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1551 req->rq_phase == RQ_PHASE_COMPLETE)) {
1552 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1556 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1557 LASSERT(req->rq_next_phase != req->rq_phase);
1558 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1561 * Skip processing until reply is unlinked. We
1562 * can't return to pool before that and we can't
1563 * call interpret before that. We need to make
1564 * sure that all rdma transfers finished and will
1565 * not corrupt any data.
1567 if (ptlrpc_client_recv_or_unlink(req) ||
1568 ptlrpc_client_bulk_active(req))
1572 * Turn fail_loc off to prevent it from looping
1575 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1576 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1579 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1580 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1585 * Move to next phase if reply was successfully
1588 ptlrpc_rqphase_move(req, req->rq_next_phase);
1591 if (req->rq_phase == RQ_PHASE_COMPLETE)
1594 if (req->rq_phase == RQ_PHASE_INTERPRET)
1595 GOTO(interpret, req->rq_status);
1598 * Note that this also will start async reply unlink.
1600 if (req->rq_net_err && !req->rq_timedout) {
1601 ptlrpc_expire_one_request(req, 1);
1604 * Check if we still need to wait for unlink.
1606 if (ptlrpc_client_recv_or_unlink(req) ||
1607 ptlrpc_client_bulk_active(req))
1609 /* If there is no need to resend, fail it now. */
1610 if (req->rq_no_resend) {
1611 if (req->rq_status == 0)
1612 req->rq_status = -EIO;
1613 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1614 GOTO(interpret, req->rq_status);
1621 spin_lock(&req->rq_lock);
1622 req->rq_replied = 0;
1623 spin_unlock(&req->rq_lock);
1624 if (req->rq_status == 0)
1625 req->rq_status = -EIO;
1626 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1627 GOTO(interpret, req->rq_status);
1630 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1631 * so it sets rq_intr regardless of individual rpc
1632 * timeouts. The synchronous IO waiting path sets
1633 * rq_intr irrespective of whether ptlrpcd
1634 * has seen a timeout. Our policy is to only interpret
1635 * interrupted rpcs after they have timed out, so we
1636 * need to enforce that here.
1639 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1640 req->rq_wait_ctx)) {
1641 req->rq_status = -EINTR;
1642 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1643 GOTO(interpret, req->rq_status);
1646 if (req->rq_phase == RQ_PHASE_RPC) {
1647 if (req->rq_timedout || req->rq_resend ||
1648 req->rq_waiting || req->rq_wait_ctx) {
1651 if (!ptlrpc_unregister_reply(req, 1))
1654 spin_lock(&imp->imp_lock);
1655 if (ptlrpc_import_delay_req(imp, req, &status)){
1656 /* put on delay list - only if we wait
1657 * recovery finished - before send */
1658 cfs_list_del_init(&req->rq_list);
1659 cfs_list_add_tail(&req->rq_list,
1662 spin_unlock(&imp->imp_lock);
1667 req->rq_status = status;
1668 ptlrpc_rqphase_move(req,
1669 RQ_PHASE_INTERPRET);
1670 spin_unlock(&imp->imp_lock);
1671 GOTO(interpret, req->rq_status);
1673 if (ptlrpc_no_resend(req) &&
1674 !req->rq_wait_ctx) {
1675 req->rq_status = -ENOTCONN;
1676 ptlrpc_rqphase_move(req,
1677 RQ_PHASE_INTERPRET);
1678 spin_unlock(&imp->imp_lock);
1679 GOTO(interpret, req->rq_status);
1682 cfs_list_del_init(&req->rq_list);
1683 cfs_list_add_tail(&req->rq_list,
1684 &imp->imp_sending_list);
1686 spin_unlock(&imp->imp_lock);
1688 spin_lock(&req->rq_lock);
1689 req->rq_waiting = 0;
1690 spin_unlock(&req->rq_lock);
1692 if (req->rq_timedout || req->rq_resend) {
1693 /* This is re-sending anyways,
1694 * let's mark req as resend. */
1695 spin_lock(&req->rq_lock);
1697 spin_unlock(&req->rq_lock);
1701 if (!ptlrpc_unregister_bulk(req, 1))
1704 /* ensure previous bulk fails */
1705 old_xid = req->rq_xid;
1706 req->rq_xid = ptlrpc_next_xid();
1707 CDEBUG(D_HA, "resend bulk "
1710 old_xid, req->rq_xid);
1714 * rq_wait_ctx is only touched by ptlrpcd,
1715 * so no lock is needed here.
1717 status = sptlrpc_req_refresh_ctx(req, -1);
1720 req->rq_status = status;
1721 spin_lock(&req->rq_lock);
1722 req->rq_wait_ctx = 0;
1723 spin_unlock(&req->rq_lock);
1724 force_timer_recalc = 1;
1726 spin_lock(&req->rq_lock);
1727 req->rq_wait_ctx = 1;
1728 spin_unlock(&req->rq_lock);
1733 spin_lock(&req->rq_lock);
1734 req->rq_wait_ctx = 0;
1735 spin_unlock(&req->rq_lock);
1738 rc = ptl_send_rpc(req, 0);
1740 DEBUG_REQ(D_HA, req,
1741 "send failed: rc = %d", rc);
1742 force_timer_recalc = 1;
1743 spin_lock(&req->rq_lock);
1744 req->rq_net_err = 1;
1745 spin_unlock(&req->rq_lock);
1748 /* need to reset the timeout */
1749 force_timer_recalc = 1;
1752 spin_lock(&req->rq_lock);
1754 if (ptlrpc_client_early(req)) {
1755 ptlrpc_at_recv_early_reply(req);
1756 spin_unlock(&req->rq_lock);
1760 /* Still waiting for a reply? */
1761 if (ptlrpc_client_recv(req)) {
1762 spin_unlock(&req->rq_lock);
1766 /* Did we actually receive a reply? */
1767 if (!ptlrpc_client_replied(req)) {
1768 spin_unlock(&req->rq_lock);
1772 spin_unlock(&req->rq_lock);
1774 /* unlink from net because we are going to
1775 * swab in-place of reply buffer */
1776 unregistered = ptlrpc_unregister_reply(req, 1);
1780 req->rq_status = after_reply(req);
1784 /* If there is no bulk associated with this request,
1785 * then we're done and should let the interpreter
1786 * process the reply. Similarly if the RPC returned
1787 * an error, and therefore the bulk will never arrive.
1789 if (req->rq_bulk == NULL || req->rq_status < 0) {
1790 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1791 GOTO(interpret, req->rq_status);
1794 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1797 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1798 if (ptlrpc_client_bulk_active(req))
1801 if (req->rq_bulk->bd_failure) {
1802 /* The RPC reply arrived OK, but the bulk screwed
1803 * up! Dead weird since the server told us the RPC
1804 * was good after getting the REPLY for her GET or
1805 * the ACK for her PUT. */
1806 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1807 req->rq_status = -EIO;
1810 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1813 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1815 /* This moves to "unregistering" phase we need to wait for
1817 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1818 /* start async bulk unlink too */
1819 ptlrpc_unregister_bulk(req, 1);
1823 if (!ptlrpc_unregister_bulk(req, 1))
1826 /* When calling interpret receiving already should be
1828 LASSERT(!req->rq_receiving_reply);
1830 ptlrpc_req_interpret(env, req, req->rq_status);
1832 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1834 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1835 "Completed RPC pname:cluuid:pid:xid:nid:"
1836 "opc %s:%s:%d:"LPU64":%s:%d\n",
1837 current_comm(), imp->imp_obd->obd_uuid.uuid,
1838 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1839 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1840 lustre_msg_get_opc(req->rq_reqmsg));
1842 spin_lock(&imp->imp_lock);
1843 /* Request already may be not on sending or delaying list. This
1844 * may happen in the case of marking it erroneous for the case
1845 * ptlrpc_import_delay_req(req, status) find it impossible to
1846 * allow sending this rpc and returns *status != 0. */
1847 if (!cfs_list_empty(&req->rq_list)) {
1848 cfs_list_del_init(&req->rq_list);
1849 cfs_atomic_dec(&imp->imp_inflight);
1851 spin_unlock(&imp->imp_lock);
1853 cfs_atomic_dec(&set->set_remaining);
1854 wake_up_all(&imp->imp_recovery_waitq);
1856 if (set->set_producer) {
1857 /* produce a new request if possible */
1858 if (ptlrpc_set_producer(set) > 0)
1859 force_timer_recalc = 1;
1861 /* free the request that has just been completed
1862 * in order not to pollute set->set_requests */
1863 cfs_list_del_init(&req->rq_set_chain);
1864 spin_lock(&req->rq_lock);
1866 req->rq_invalid_rqset = 0;
1867 spin_unlock(&req->rq_lock);
1869 /* record rq_status to compute the final status later */
1870 if (req->rq_status != 0)
1871 set->set_rc = req->rq_status;
1872 ptlrpc_req_finished(req);
1876 /* If we hit an error, we want to recover promptly. */
1877 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1879 EXPORT_SYMBOL(ptlrpc_check_set);
1882 * Time out request \a req. is \a async_unlink is set, that means do not wait
1883 * until LNet actually confirms network buffer unlinking.
1884 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1886 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1888 struct obd_import *imp = req->rq_import;
1892 spin_lock(&req->rq_lock);
1893 req->rq_timedout = 1;
1894 spin_unlock(&req->rq_lock);
1896 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1897 "/real "CFS_DURATION_T"]",
1898 req->rq_net_err ? "failed due to network error" :
1899 ((req->rq_real_sent == 0 ||
1900 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1901 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1902 "timed out for sent delay" : "timed out for slow reply"),
1903 req->rq_sent, req->rq_real_sent);
1905 if (imp != NULL && obd_debug_peer_on_timeout)
1906 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1908 ptlrpc_unregister_reply(req, async_unlink);
1909 ptlrpc_unregister_bulk(req, async_unlink);
1911 if (obd_dump_on_timeout)
1912 libcfs_debug_dumplog();
1915 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1919 cfs_atomic_inc(&imp->imp_timeouts);
1921 /* The DLM server doesn't want recovery run on its imports. */
1922 if (imp->imp_dlm_fake)
1925 /* If this request is for recovery or other primordial tasks,
1926 * then error it out here. */
1927 if (req->rq_ctx_init || req->rq_ctx_fini ||
1928 req->rq_send_state != LUSTRE_IMP_FULL ||
1929 imp->imp_obd->obd_no_recov) {
1930 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1931 ptlrpc_import_state_name(req->rq_send_state),
1932 ptlrpc_import_state_name(imp->imp_state));
1933 spin_lock(&req->rq_lock);
1934 req->rq_status = -ETIMEDOUT;
1936 spin_unlock(&req->rq_lock);
1940 /* if a request can't be resent we can't wait for an answer after
1942 if (ptlrpc_no_resend(req)) {
1943 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1947 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1953 * Time out all uncompleted requests in request set pointed by \a data
1954 * Callback used when waiting on sets with l_wait_event.
1957 int ptlrpc_expired_set(void *data)
1959 struct ptlrpc_request_set *set = data;
1961 time_t now = cfs_time_current_sec();
1964 LASSERT(set != NULL);
1967 * A timeout expired. See which reqs it applies to...
1969 cfs_list_for_each (tmp, &set->set_requests) {
1970 struct ptlrpc_request *req =
1971 cfs_list_entry(tmp, struct ptlrpc_request,
1974 /* don't expire request waiting for context */
1975 if (req->rq_wait_ctx)
1978 /* Request in-flight? */
1979 if (!((req->rq_phase == RQ_PHASE_RPC &&
1980 !req->rq_waiting && !req->rq_resend) ||
1981 (req->rq_phase == RQ_PHASE_BULK)))
1984 if (req->rq_timedout || /* already dealt with */
1985 req->rq_deadline > now) /* not expired */
1988 /* Deal with this guy. Do it asynchronously to not block
1989 * ptlrpcd thread. */
1990 ptlrpc_expire_one_request(req, 1);
1994 * When waiting for a whole set, we always break out of the
1995 * sleep so we can recalculate the timeout, or enable interrupts
1996 * if everyone's timed out.
2000 EXPORT_SYMBOL(ptlrpc_expired_set);
2003 * Sets rq_intr flag in \a req under spinlock.
2005 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2007 spin_lock(&req->rq_lock);
2009 spin_unlock(&req->rq_lock);
2011 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2014 * Interrupts (sets interrupted flag) all uncompleted requests in
2015 * a set \a data. Callback for l_wait_event for interruptible waits.
2017 void ptlrpc_interrupted_set(void *data)
2019 struct ptlrpc_request_set *set = data;
2022 LASSERT(set != NULL);
2023 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2025 cfs_list_for_each(tmp, &set->set_requests) {
2026 struct ptlrpc_request *req =
2027 cfs_list_entry(tmp, struct ptlrpc_request,
2030 if (req->rq_phase != RQ_PHASE_RPC &&
2031 req->rq_phase != RQ_PHASE_UNREGISTERING)
2034 ptlrpc_mark_interrupted(req);
2037 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2040 * Get the smallest timeout in the set; this does NOT set a timeout.
2042 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2045 time_t now = cfs_time_current_sec();
2047 struct ptlrpc_request *req;
2051 cfs_list_for_each(tmp, &set->set_requests) {
2052 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2055 * Request in-flight?
2057 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2058 (req->rq_phase == RQ_PHASE_BULK) ||
2059 (req->rq_phase == RQ_PHASE_NEW)))
2063 * Already timed out.
2065 if (req->rq_timedout)
2071 if (req->rq_wait_ctx)
2074 if (req->rq_phase == RQ_PHASE_NEW)
2075 deadline = req->rq_sent;
2076 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2077 deadline = req->rq_sent;
2079 deadline = req->rq_sent + req->rq_timeout;
2081 if (deadline <= now) /* actually expired already */
2082 timeout = 1; /* ASAP */
2083 else if (timeout == 0 || timeout > deadline - now)
2084 timeout = deadline - now;
2088 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2091 * Send all unset request from the set and then wait untill all
2092 * requests in the set complete (either get a reply, timeout, get an
2093 * error or otherwise be interrupted).
2094 * Returns 0 on success or error code otherwise.
2096 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2099 struct ptlrpc_request *req;
2100 struct l_wait_info lwi;
2104 if (set->set_producer)
2105 (void)ptlrpc_set_producer(set);
2107 cfs_list_for_each(tmp, &set->set_requests) {
2108 req = cfs_list_entry(tmp, struct ptlrpc_request,
2110 if (req->rq_phase == RQ_PHASE_NEW)
2111 (void)ptlrpc_send_new_req(req);
2114 if (cfs_list_empty(&set->set_requests))
2118 timeout = ptlrpc_set_next_timeout(set);
2120 /* wait until all complete, interrupted, or an in-flight
2122 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2125 if (timeout == 0 && !cfs_signal_pending())
2127 * No requests are in-flight (ether timed out
2128 * or delayed), so we can allow interrupts.
2129 * We still want to block for a limited time,
2130 * so we allow interrupts during the timeout.
2132 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2134 ptlrpc_interrupted_set, set);
2137 * At least one request is in flight, so no
2138 * interrupts are allowed. Wait until all
2139 * complete, or an in-flight req times out.
2141 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2142 ptlrpc_expired_set, set);
2144 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2146 /* LU-769 - if we ignored the signal because it was already
2147 * pending when we started, we need to handle it now or we risk
2148 * it being ignored forever */
2149 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2150 cfs_signal_pending()) {
2151 sigset_t blocked_sigs =
2152 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2154 /* In fact we only interrupt for the "fatal" signals
2155 * like SIGINT or SIGKILL. We still ignore less
2156 * important signals since ptlrpc set is not easily
2157 * reentrant from userspace again */
2158 if (cfs_signal_pending())
2159 ptlrpc_interrupted_set(set);
2160 cfs_restore_sigs(blocked_sigs);
2163 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2165 /* -EINTR => all requests have been flagged rq_intr so next
2167 * -ETIMEDOUT => someone timed out. When all reqs have
2168 * timed out, signals are enabled allowing completion with
2170 * I don't really care if we go once more round the loop in
2171 * the error cases -eeb. */
2172 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2173 cfs_list_for_each(tmp, &set->set_requests) {
2174 req = cfs_list_entry(tmp, struct ptlrpc_request,
2176 spin_lock(&req->rq_lock);
2177 req->rq_invalid_rqset = 1;
2178 spin_unlock(&req->rq_lock);
2181 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2183 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2185 rc = set->set_rc; /* rq_status of already freed requests if any */
2186 cfs_list_for_each(tmp, &set->set_requests) {
2187 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2189 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2190 if (req->rq_status != 0)
2191 rc = req->rq_status;
2194 if (set->set_interpret != NULL) {
2195 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2197 rc = interpreter (set, set->set_arg, rc);
2199 struct ptlrpc_set_cbdata *cbdata, *n;
2202 cfs_list_for_each_entry_safe(cbdata, n,
2203 &set->set_cblist, psc_item) {
2204 cfs_list_del_init(&cbdata->psc_item);
2205 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2208 OBD_FREE_PTR(cbdata);
2214 EXPORT_SYMBOL(ptlrpc_set_wait);
2217 * Helper fuction for request freeing.
2218 * Called when request count reached zero and request needs to be freed.
2219 * Removes request from all sorts of sending/replay lists it might be on,
2220 * frees network buffers if any are present.
2221 * If \a locked is set, that means caller is already holding import imp_lock
2222 * and so we no longer need to reobtain it (for certain lists manipulations)
2224 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2227 if (request == NULL) {
2232 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2233 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2234 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2235 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2236 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2237 LASSERTF(!request->rq_replay, "req %p\n", request);
2239 req_capsule_fini(&request->rq_pill);
2241 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2242 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2243 if (request->rq_import != NULL) {
2245 spin_lock(&request->rq_import->imp_lock);
2246 cfs_list_del_init(&request->rq_replay_list);
2248 spin_unlock(&request->rq_import->imp_lock);
2250 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2252 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2253 DEBUG_REQ(D_ERROR, request,
2254 "freeing request with nonzero refcount");
2258 if (request->rq_repbuf != NULL)
2259 sptlrpc_cli_free_repbuf(request);
2260 if (request->rq_export != NULL) {
2261 class_export_put(request->rq_export);
2262 request->rq_export = NULL;
2264 if (request->rq_import != NULL) {
2265 class_import_put(request->rq_import);
2266 request->rq_import = NULL;
2268 if (request->rq_bulk != NULL)
2269 ptlrpc_free_bulk_pin(request->rq_bulk);
2271 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2272 sptlrpc_cli_free_reqbuf(request);
2274 if (request->rq_cli_ctx)
2275 sptlrpc_req_put_ctx(request, !locked);
2277 if (request->rq_pool)
2278 __ptlrpc_free_req_to_pool(request);
2280 ptlrpc_request_cache_free(request);
2284 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2286 * Drop one request reference. Must be called with import imp_lock held.
2287 * When reference count drops to zero, reuqest is freed.
2289 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2291 LASSERT(spin_is_locked(&request->rq_import->imp_lock));
2292 (void)__ptlrpc_req_finished(request, 1);
2294 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2298 * Drops one reference count for request \a request.
2299 * \a locked set indicates that caller holds import imp_lock.
2300 * Frees the request whe reference count reaches zero.
2302 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2305 if (request == NULL)
2308 if (request == LP_POISON ||
2309 request->rq_reqmsg == LP_POISON) {
2310 CERROR("dereferencing freed request (bug 575)\n");
2315 DEBUG_REQ(D_INFO, request, "refcount now %u",
2316 cfs_atomic_read(&request->rq_refcount) - 1);
2318 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2319 __ptlrpc_free_req(request, locked);
2327 * Drops one reference count for a request.
2329 void ptlrpc_req_finished(struct ptlrpc_request *request)
2331 __ptlrpc_req_finished(request, 0);
2333 EXPORT_SYMBOL(ptlrpc_req_finished);
2336 * Returns xid of a \a request
2338 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2340 return request->rq_xid;
2342 EXPORT_SYMBOL(ptlrpc_req_xid);
2345 * Disengage the client's reply buffer from the network
2346 * NB does _NOT_ unregister any client-side bulk.
2347 * IDEMPOTENT, but _not_ safe against concurrent callers.
2348 * The request owner (i.e. the thread doing the I/O) must call...
2349 * Returns 0 on success or 1 if unregistering cannot be made.
2351 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2354 struct l_wait_info lwi;
2359 LASSERT(!in_interrupt());
2362 * Let's setup deadline for reply unlink.
2364 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2365 async && request->rq_reply_deadline == 0)
2366 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2369 * Nothing left to do.
2371 if (!ptlrpc_client_recv_or_unlink(request))
2374 LNetMDUnlink(request->rq_reply_md_h);
2377 * Let's check it once again.
2379 if (!ptlrpc_client_recv_or_unlink(request))
2383 * Move to "Unregistering" phase as reply was not unlinked yet.
2385 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2388 * Do not wait for unlink to finish.
2394 * We have to l_wait_event() whatever the result, to give liblustre
2395 * a chance to run reply_in_callback(), and to make sure we've
2396 * unlinked before returning a req to the pool.
2400 /* The wq argument is ignored by user-space wait_event macros */
2401 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2402 &request->rq_set->set_waitq :
2403 &request->rq_reply_waitq;
2405 /* Network access will complete in finite time but the HUGE
2406 * timeout lets us CWARN for visibility of sluggish NALs */
2407 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2408 cfs_time_seconds(1), NULL, NULL);
2409 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2412 ptlrpc_rqphase_move(request, request->rq_next_phase);
2416 LASSERT(rc == -ETIMEDOUT);
2417 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2418 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2419 request->rq_must_unlink);
2423 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2425 static void ptlrpc_free_request(struct ptlrpc_request *req)
2427 spin_lock(&req->rq_lock);
2429 spin_unlock(&req->rq_lock);
2431 if (req->rq_commit_cb != NULL)
2432 req->rq_commit_cb(req);
2433 cfs_list_del_init(&req->rq_replay_list);
2435 __ptlrpc_req_finished(req, 1);
2439 * the request is committed and dropped from the replay list of its import
2441 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2443 struct obd_import *imp = req->rq_import;
2445 spin_lock(&imp->imp_lock);
2446 if (cfs_list_empty(&req->rq_replay_list)) {
2447 spin_unlock(&imp->imp_lock);
2451 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2452 ptlrpc_free_request(req);
2454 spin_unlock(&imp->imp_lock);
2456 EXPORT_SYMBOL(ptlrpc_request_committed);
2459 * Iterates through replay_list on import and prunes
2460 * all requests have transno smaller than last_committed for the
2461 * import and don't have rq_replay set.
2462 * Since requests are sorted in transno order, stops when meetign first
2463 * transno bigger than last_committed.
2464 * caller must hold imp->imp_lock
2466 void ptlrpc_free_committed(struct obd_import *imp)
2468 struct ptlrpc_request *req, *saved;
2469 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2470 bool skip_committed_list = true;
2473 LASSERT(imp != NULL);
2474 LASSERT(spin_is_locked(&imp->imp_lock));
2477 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2478 imp->imp_generation == imp->imp_last_generation_checked) {
2479 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2480 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2484 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2485 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2486 imp->imp_generation);
2488 if (imp->imp_generation != imp->imp_last_generation_checked)
2489 skip_committed_list = false;
2491 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2492 imp->imp_last_generation_checked = imp->imp_generation;
2494 cfs_list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2496 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2497 LASSERT(req != last_req);
2500 if (req->rq_transno == 0) {
2501 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2504 if (req->rq_import_generation < imp->imp_generation) {
2505 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2509 /* not yet committed */
2510 if (req->rq_transno > imp->imp_peer_committed_transno) {
2511 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2515 if (req->rq_replay) {
2516 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2517 cfs_list_move_tail(&req->rq_replay_list,
2518 &imp->imp_committed_list);
2522 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2523 imp->imp_peer_committed_transno);
2525 ptlrpc_free_request(req);
2528 if (skip_committed_list)
2531 cfs_list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2533 LASSERT(req->rq_transno != 0);
2534 if (req->rq_import_generation < imp->imp_generation) {
2535 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2536 ptlrpc_free_request(req);
2544 void ptlrpc_cleanup_client(struct obd_import *imp)
2550 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2553 * Schedule previously sent request for resend.
2554 * For bulk requests we assign new xid (to avoid problems with
2555 * lost replies and therefore several transfers landing into same buffer
2556 * from different sending attempts).
2558 void ptlrpc_resend_req(struct ptlrpc_request *req)
2560 DEBUG_REQ(D_HA, req, "going to resend");
2561 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2562 req->rq_status = -EAGAIN;
2564 spin_lock(&req->rq_lock);
2566 req->rq_net_err = 0;
2567 req->rq_timedout = 0;
2569 __u64 old_xid = req->rq_xid;
2571 /* ensure previous bulk fails */
2572 req->rq_xid = ptlrpc_next_xid();
2573 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2574 old_xid, req->rq_xid);
2576 ptlrpc_client_wake_req(req);
2577 spin_unlock(&req->rq_lock);
2579 EXPORT_SYMBOL(ptlrpc_resend_req);
2581 /* XXX: this function and rq_status are currently unused */
2582 void ptlrpc_restart_req(struct ptlrpc_request *req)
2584 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2585 req->rq_status = -ERESTARTSYS;
2587 spin_lock(&req->rq_lock);
2588 req->rq_restart = 1;
2589 req->rq_timedout = 0;
2590 ptlrpc_client_wake_req(req);
2591 spin_unlock(&req->rq_lock);
2593 EXPORT_SYMBOL(ptlrpc_restart_req);
2596 * Grab additional reference on a request \a req
2598 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2601 cfs_atomic_inc(&req->rq_refcount);
2604 EXPORT_SYMBOL(ptlrpc_request_addref);
2607 * Add a request to import replay_list.
2608 * Must be called under imp_lock
2610 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2611 struct obd_import *imp)
2615 LASSERT(spin_is_locked(&imp->imp_lock));
2617 if (req->rq_transno == 0) {
2618 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2622 /* clear this for new requests that were resent as well
2623 as resent replayed requests. */
2624 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2626 /* don't re-add requests that have been replayed */
2627 if (!cfs_list_empty(&req->rq_replay_list))
2630 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2632 LASSERT(imp->imp_replayable);
2633 /* Balanced in ptlrpc_free_committed, usually. */
2634 ptlrpc_request_addref(req);
2635 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2636 struct ptlrpc_request *iter =
2637 cfs_list_entry(tmp, struct ptlrpc_request,
2640 /* We may have duplicate transnos if we create and then
2641 * open a file, or for closes retained if to match creating
2642 * opens, so use req->rq_xid as a secondary key.
2643 * (See bugs 684, 685, and 428.)
2644 * XXX no longer needed, but all opens need transnos!
2646 if (iter->rq_transno > req->rq_transno)
2649 if (iter->rq_transno == req->rq_transno) {
2650 LASSERT(iter->rq_xid != req->rq_xid);
2651 if (iter->rq_xid > req->rq_xid)
2655 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2659 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2661 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2664 * Send request and wait until it completes.
2665 * Returns request processing status.
2667 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2669 struct ptlrpc_request_set *set;
2673 LASSERT(req->rq_set == NULL);
2674 LASSERT(!req->rq_receiving_reply);
2676 set = ptlrpc_prep_set();
2678 CERROR("Unable to allocate ptlrpc set.");
2682 /* for distributed debugging */
2683 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2685 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2686 ptlrpc_request_addref(req);
2687 ptlrpc_set_add_req(set, req);
2688 rc = ptlrpc_set_wait(set);
2689 ptlrpc_set_destroy(set);
2693 EXPORT_SYMBOL(ptlrpc_queue_wait);
2695 struct ptlrpc_replay_async_args {
2697 int praa_old_status;
2701 * Callback used for replayed requests reply processing.
2702 * In case of succesful reply calls registeresd request replay callback.
2703 * In case of error restart replay process.
2705 static int ptlrpc_replay_interpret(const struct lu_env *env,
2706 struct ptlrpc_request *req,
2707 void * data, int rc)
2709 struct ptlrpc_replay_async_args *aa = data;
2710 struct obd_import *imp = req->rq_import;
2713 cfs_atomic_dec(&imp->imp_replay_inflight);
2715 if (!ptlrpc_client_replied(req)) {
2716 CERROR("request replay timed out, restarting recovery\n");
2717 GOTO(out, rc = -ETIMEDOUT);
2720 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2721 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2722 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2723 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2725 /** VBR: check version failure */
2726 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2727 /** replay was failed due to version mismatch */
2728 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2729 spin_lock(&imp->imp_lock);
2730 imp->imp_vbr_failed = 1;
2731 imp->imp_no_lock_replay = 1;
2732 spin_unlock(&imp->imp_lock);
2733 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2735 /** The transno had better not change over replay. */
2736 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2737 lustre_msg_get_transno(req->rq_repmsg) ||
2738 lustre_msg_get_transno(req->rq_repmsg) == 0,
2740 lustre_msg_get_transno(req->rq_reqmsg),
2741 lustre_msg_get_transno(req->rq_repmsg));
2744 spin_lock(&imp->imp_lock);
2745 /** if replays by version then gap occur on server, no trust to locks */
2746 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2747 imp->imp_no_lock_replay = 1;
2748 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2749 spin_unlock(&imp->imp_lock);
2750 LASSERT(imp->imp_last_replay_transno);
2752 /* transaction number shouldn't be bigger than the latest replayed */
2753 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2754 DEBUG_REQ(D_ERROR, req,
2755 "Reported transno "LPU64" is bigger than the "
2756 "replayed one: "LPU64, req->rq_transno,
2757 lustre_msg_get_transno(req->rq_reqmsg));
2758 GOTO(out, rc = -EINVAL);
2761 DEBUG_REQ(D_HA, req, "got rep");
2763 /* let the callback do fixups, possibly including in the request */
2764 if (req->rq_replay_cb)
2765 req->rq_replay_cb(req);
2767 if (ptlrpc_client_replied(req) &&
2768 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2769 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2770 lustre_msg_get_status(req->rq_repmsg),
2771 aa->praa_old_status);
2773 /* Put it back for re-replay. */
2774 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2778 * Errors while replay can set transno to 0, but
2779 * imp_last_replay_transno shouldn't be set to 0 anyway
2781 if (req->rq_transno == 0)
2782 CERROR("Transno is 0 during replay!\n");
2784 /* continue with recovery */
2785 rc = ptlrpc_import_recovery_state_machine(imp);
2787 req->rq_send_state = aa->praa_old_state;
2790 /* this replay failed, so restart recovery */
2791 ptlrpc_connect_import(imp);
2797 * Prepares and queues request for replay.
2798 * Adds it to ptlrpcd queue for actual sending.
2799 * Returns 0 on success.
2801 int ptlrpc_replay_req(struct ptlrpc_request *req)
2803 struct ptlrpc_replay_async_args *aa;
2806 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2808 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2809 aa = ptlrpc_req_async_args(req);
2810 memset(aa, 0, sizeof *aa);
2812 /* Prepare request to be resent with ptlrpcd */
2813 aa->praa_old_state = req->rq_send_state;
2814 req->rq_send_state = LUSTRE_IMP_REPLAY;
2815 req->rq_phase = RQ_PHASE_NEW;
2816 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2818 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2820 req->rq_interpret_reply = ptlrpc_replay_interpret;
2821 /* Readjust the timeout for current conditions */
2822 ptlrpc_at_set_req_timeout(req);
2824 /* Tell server the net_latency, so the server can calculate how long
2825 * it should wait for next replay */
2826 lustre_msg_set_service_time(req->rq_reqmsg,
2827 ptlrpc_at_get_net_latency(req));
2828 DEBUG_REQ(D_HA, req, "REPLAY");
2830 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2831 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2833 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2836 EXPORT_SYMBOL(ptlrpc_replay_req);
2839 * Aborts all in-flight request on import \a imp sending and delayed lists
2841 void ptlrpc_abort_inflight(struct obd_import *imp)
2843 cfs_list_t *tmp, *n;
2846 /* Make sure that no new requests get processed for this import.
2847 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2848 * this flag and then putting requests on sending_list or delayed_list.
2850 spin_lock(&imp->imp_lock);
2852 /* XXX locking? Maybe we should remove each request with the list
2853 * locked? Also, how do we know if the requests on the list are
2854 * being freed at this time?
2856 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2857 struct ptlrpc_request *req =
2858 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2860 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2862 spin_lock(&req->rq_lock);
2863 if (req->rq_import_generation < imp->imp_generation) {
2865 req->rq_status = -EIO;
2866 ptlrpc_client_wake_req(req);
2868 spin_unlock(&req->rq_lock);
2871 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2872 struct ptlrpc_request *req =
2873 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2875 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2877 spin_lock(&req->rq_lock);
2878 if (req->rq_import_generation < imp->imp_generation) {
2880 req->rq_status = -EIO;
2881 ptlrpc_client_wake_req(req);
2883 spin_unlock(&req->rq_lock);
2886 /* Last chance to free reqs left on the replay list, but we
2887 * will still leak reqs that haven't committed. */
2888 if (imp->imp_replayable)
2889 ptlrpc_free_committed(imp);
2891 spin_unlock(&imp->imp_lock);
2895 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2898 * Abort all uncompleted requests in request set \a set
2900 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2902 cfs_list_t *tmp, *pos;
2904 LASSERT(set != NULL);
2906 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2907 struct ptlrpc_request *req =
2908 cfs_list_entry(pos, struct ptlrpc_request,
2911 spin_lock(&req->rq_lock);
2912 if (req->rq_phase != RQ_PHASE_RPC) {
2913 spin_unlock(&req->rq_lock);
2918 req->rq_status = -EINTR;
2919 ptlrpc_client_wake_req(req);
2920 spin_unlock(&req->rq_lock);
2924 static __u64 ptlrpc_last_xid;
2925 static spinlock_t ptlrpc_last_xid_lock;
2928 * Initialize the XID for the node. This is common among all requests on
2929 * this node, and only requires the property that it is monotonically
2930 * increasing. It does not need to be sequential. Since this is also used
2931 * as the RDMA match bits, it is important that a single client NOT have
2932 * the same match bits for two different in-flight requests, hence we do
2933 * NOT want to have an XID per target or similar.
2935 * To avoid an unlikely collision between match bits after a client reboot
2936 * (which would deliver old data into the wrong RDMA buffer) initialize
2937 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2938 * If the time is clearly incorrect, we instead use a 62-bit random number.
2939 * In the worst case the random number will overflow 1M RPCs per second in
2940 * 9133 years, or permutations thereof.
2942 #define YEAR_2004 (1ULL << 30)
2943 void ptlrpc_init_xid(void)
2945 time_t now = cfs_time_current_sec();
2947 spin_lock_init(&ptlrpc_last_xid_lock);
2948 if (now < YEAR_2004) {
2949 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2950 ptlrpc_last_xid >>= 2;
2951 ptlrpc_last_xid |= (1ULL << 61);
2953 ptlrpc_last_xid = (__u64)now << 20;
2956 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
2957 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
2958 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
2962 * Increase xid and returns resulting new value to the caller.
2964 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2965 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2966 * itself uses the last bulk xid needed, so the server can determine the
2967 * the number of bulk transfers from the RPC XID and a bitmask. The starting
2968 * xid must align to a power-of-two value.
2970 * This is assumed to be true due to the initial ptlrpc_last_xid
2971 * value also being initialized to a power-of-two value. LU-1431
2973 __u64 ptlrpc_next_xid(void)
2977 spin_lock(&ptlrpc_last_xid_lock);
2978 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2979 ptlrpc_last_xid = next;
2980 spin_unlock(&ptlrpc_last_xid_lock);
2984 EXPORT_SYMBOL(ptlrpc_next_xid);
2987 * Get a glimpse at what next xid value might have been.
2988 * Returns possible next xid.
2990 __u64 ptlrpc_sample_next_xid(void)
2992 #if BITS_PER_LONG == 32
2993 /* need to avoid possible word tearing on 32-bit systems */
2996 spin_lock(&ptlrpc_last_xid_lock);
2997 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2998 spin_unlock(&ptlrpc_last_xid_lock);
3002 /* No need to lock, since returned value is racy anyways */
3003 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3006 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3009 * Functions for operating ptlrpc workers.
3011 * A ptlrpc work is a function which will be running inside ptlrpc context.
3012 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3014 * 1. after a work is created, it can be used many times, that is:
3015 * handler = ptlrpcd_alloc_work();
3016 * ptlrpcd_queue_work();
3018 * queue it again when necessary:
3019 * ptlrpcd_queue_work();
3020 * ptlrpcd_destroy_work();
3021 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3022 * it will only be queued once in any time. Also as its name implies, it may
3023 * have delay before it really runs by ptlrpcd thread.
3025 struct ptlrpc_work_async_args {
3027 int (*cb)(const struct lu_env *, void *);
3031 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
3033 static int work_interpreter(const struct lu_env *env,
3034 struct ptlrpc_request *req, void *data, int rc)
3036 struct ptlrpc_work_async_args *arg = data;
3038 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
3039 LASSERT(arg->cb != NULL);
3041 return arg->cb(env, arg->cbdata);
3045 * Create a work for ptlrpc.
3047 void *ptlrpcd_alloc_work(struct obd_import *imp,
3048 int (*cb)(const struct lu_env *, void *), void *cbdata)
3050 struct ptlrpc_request *req = NULL;
3051 struct ptlrpc_work_async_args *args;
3057 RETURN(ERR_PTR(-EINVAL));
3059 /* copy some code from deprecated fakereq. */
3060 req = ptlrpc_request_cache_alloc(__GFP_IO);
3062 CERROR("ptlrpc: run out of memory!\n");
3063 RETURN(ERR_PTR(-ENOMEM));
3066 req->rq_send_state = LUSTRE_IMP_FULL;
3067 req->rq_type = PTL_RPC_MSG_REQUEST;
3068 req->rq_import = class_import_get(imp);
3069 req->rq_export = NULL;
3070 req->rq_interpret_reply = work_interpreter;
3071 /* don't want reply */
3072 req->rq_receiving_reply = 0;
3073 req->rq_must_unlink = 0;
3074 req->rq_no_delay = req->rq_no_resend = 1;
3076 spin_lock_init(&req->rq_lock);
3077 CFS_INIT_LIST_HEAD(&req->rq_list);
3078 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
3079 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
3080 CFS_INIT_LIST_HEAD(&req->rq_history_list);
3081 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
3082 init_waitqueue_head(&req->rq_reply_waitq);
3083 init_waitqueue_head(&req->rq_set_waitq);
3084 cfs_atomic_set(&req->rq_refcount, 1);
3086 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3087 args = ptlrpc_req_async_args(req);
3088 args->magic = PTLRPC_WORK_MAGIC;
3090 args->cbdata = cbdata;
3094 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3096 void ptlrpcd_destroy_work(void *handler)
3098 struct ptlrpc_request *req = handler;
3101 ptlrpc_req_finished(req);
3103 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3105 int ptlrpcd_queue_work(void *handler)
3107 struct ptlrpc_request *req = handler;
3110 * Check if the req is already being queued.
3112 * Here comes a trick: it lacks a way of checking if a req is being
3113 * processed reliably in ptlrpc. Here I have to use refcount of req
3114 * for this purpose. This is okay because the caller should use this
3115 * req as opaque data. - Jinshan
3117 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
3118 if (cfs_atomic_read(&req->rq_refcount) > 1)
3121 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
3122 cfs_atomic_dec(&req->rq_refcount);
3126 /* re-initialize the req */
3127 req->rq_timeout = obd_timeout;
3128 req->rq_sent = cfs_time_current_sec();
3129 req->rq_deadline = req->rq_sent + req->rq_timeout;
3130 req->rq_reply_deadline = req->rq_deadline;
3131 req->rq_phase = RQ_PHASE_INTERPRET;
3132 req->rq_next_phase = RQ_PHASE_COMPLETE;
3133 req->rq_xid = ptlrpc_next_xid();
3134 req->rq_import_generation = req->rq_import->imp_generation;
3136 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3139 EXPORT_SYMBOL(ptlrpcd_queue_work);