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 list_empty(&req->rq_replay_list)) {
1367 /* NB: don't call rq_commit_cb if it's already on
1368 * rq_replay_list, ptlrpc_free_committed() will call
1369 * it later, see LU-3618 for details */
1370 spin_unlock(&imp->imp_lock);
1371 req->rq_commit_cb(req);
1372 spin_lock(&imp->imp_lock);
1376 * Replay-enabled imports return commit-status information.
1378 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1379 imp->imp_peer_committed_transno =
1380 lustre_msg_get_last_committed(req->rq_repmsg);
1383 ptlrpc_free_committed(imp);
1385 if (!cfs_list_empty(&imp->imp_replay_list)) {
1386 struct ptlrpc_request *last;
1388 last = cfs_list_entry(imp->imp_replay_list.prev,
1389 struct ptlrpc_request,
1392 * Requests with rq_replay stay on the list even if no
1393 * commit is expected.
1395 if (last->rq_transno > imp->imp_peer_committed_transno)
1396 ptlrpc_pinger_commit_expected(imp);
1399 spin_unlock(&imp->imp_lock);
1406 * Helper function to send request \a req over the network for the first time
1407 * Also adjusts request phase.
1408 * Returns 0 on success or error code.
1410 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1412 struct obd_import *imp = req->rq_import;
1416 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1417 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1418 (!req->rq_generation_set ||
1419 req->rq_import_generation == imp->imp_generation))
1422 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1424 spin_lock(&imp->imp_lock);
1426 if (!req->rq_generation_set)
1427 req->rq_import_generation = imp->imp_generation;
1429 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1430 spin_lock(&req->rq_lock);
1431 req->rq_waiting = 1;
1432 spin_unlock(&req->rq_lock);
1434 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1435 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1436 ptlrpc_import_state_name(req->rq_send_state),
1437 ptlrpc_import_state_name(imp->imp_state));
1438 LASSERT(cfs_list_empty(&req->rq_list));
1439 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1440 cfs_atomic_inc(&req->rq_import->imp_inflight);
1441 spin_unlock(&imp->imp_lock);
1446 spin_unlock(&imp->imp_lock);
1447 req->rq_status = rc;
1448 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1452 LASSERT(cfs_list_empty(&req->rq_list));
1453 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1454 cfs_atomic_inc(&req->rq_import->imp_inflight);
1455 spin_unlock(&imp->imp_lock);
1457 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1459 rc = sptlrpc_req_refresh_ctx(req, -1);
1462 req->rq_status = rc;
1465 spin_lock(&req->rq_lock);
1466 req->rq_wait_ctx = 1;
1467 spin_unlock(&req->rq_lock);
1472 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1473 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1474 imp->imp_obd->obd_uuid.uuid,
1475 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1476 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1477 lustre_msg_get_opc(req->rq_reqmsg));
1479 rc = ptl_send_rpc(req, 0);
1481 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1482 spin_lock(&req->rq_lock);
1483 req->rq_net_err = 1;
1484 spin_unlock(&req->rq_lock);
1490 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1495 LASSERT(set->set_producer != NULL);
1497 remaining = cfs_atomic_read(&set->set_remaining);
1499 /* populate the ->set_requests list with requests until we
1500 * reach the maximum number of RPCs in flight for this set */
1501 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1502 rc = set->set_producer(set, set->set_producer_arg);
1503 if (rc == -ENOENT) {
1504 /* no more RPC to produce */
1505 set->set_producer = NULL;
1506 set->set_producer_arg = NULL;
1511 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1515 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1516 * and no more replies are expected.
1517 * (it is possible to get less replies than requests sent e.g. due to timed out
1518 * requests or requests that we had trouble to send out)
1520 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1522 cfs_list_t *tmp, *next;
1523 int force_timer_recalc = 0;
1526 if (cfs_atomic_read(&set->set_remaining) == 0)
1529 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1530 struct ptlrpc_request *req =
1531 cfs_list_entry(tmp, struct ptlrpc_request,
1533 struct obd_import *imp = req->rq_import;
1534 int unregistered = 0;
1537 if (req->rq_phase == RQ_PHASE_NEW &&
1538 ptlrpc_send_new_req(req)) {
1539 force_timer_recalc = 1;
1542 /* delayed send - skip */
1543 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1546 /* delayed resend - skip */
1547 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1548 req->rq_sent > cfs_time_current_sec())
1551 if (!(req->rq_phase == RQ_PHASE_RPC ||
1552 req->rq_phase == RQ_PHASE_BULK ||
1553 req->rq_phase == RQ_PHASE_INTERPRET ||
1554 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1555 req->rq_phase == RQ_PHASE_COMPLETE)) {
1556 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1560 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1561 LASSERT(req->rq_next_phase != req->rq_phase);
1562 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1565 * Skip processing until reply is unlinked. We
1566 * can't return to pool before that and we can't
1567 * call interpret before that. We need to make
1568 * sure that all rdma transfers finished and will
1569 * not corrupt any data.
1571 if (ptlrpc_client_recv_or_unlink(req) ||
1572 ptlrpc_client_bulk_active(req))
1576 * Turn fail_loc off to prevent it from looping
1579 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1580 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1583 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1584 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1589 * Move to next phase if reply was successfully
1592 ptlrpc_rqphase_move(req, req->rq_next_phase);
1595 if (req->rq_phase == RQ_PHASE_COMPLETE)
1598 if (req->rq_phase == RQ_PHASE_INTERPRET)
1599 GOTO(interpret, req->rq_status);
1602 * Note that this also will start async reply unlink.
1604 if (req->rq_net_err && !req->rq_timedout) {
1605 ptlrpc_expire_one_request(req, 1);
1608 * Check if we still need to wait for unlink.
1610 if (ptlrpc_client_recv_or_unlink(req) ||
1611 ptlrpc_client_bulk_active(req))
1613 /* If there is no need to resend, fail it now. */
1614 if (req->rq_no_resend) {
1615 if (req->rq_status == 0)
1616 req->rq_status = -EIO;
1617 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1618 GOTO(interpret, req->rq_status);
1625 spin_lock(&req->rq_lock);
1626 req->rq_replied = 0;
1627 spin_unlock(&req->rq_lock);
1628 if (req->rq_status == 0)
1629 req->rq_status = -EIO;
1630 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1631 GOTO(interpret, req->rq_status);
1634 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1635 * so it sets rq_intr regardless of individual rpc
1636 * timeouts. The synchronous IO waiting path sets
1637 * rq_intr irrespective of whether ptlrpcd
1638 * has seen a timeout. Our policy is to only interpret
1639 * interrupted rpcs after they have timed out, so we
1640 * need to enforce that here.
1643 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1644 req->rq_wait_ctx)) {
1645 req->rq_status = -EINTR;
1646 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1647 GOTO(interpret, req->rq_status);
1650 if (req->rq_phase == RQ_PHASE_RPC) {
1651 if (req->rq_timedout || req->rq_resend ||
1652 req->rq_waiting || req->rq_wait_ctx) {
1655 if (!ptlrpc_unregister_reply(req, 1))
1658 spin_lock(&imp->imp_lock);
1659 if (ptlrpc_import_delay_req(imp, req, &status)){
1660 /* put on delay list - only if we wait
1661 * recovery finished - before send */
1662 cfs_list_del_init(&req->rq_list);
1663 cfs_list_add_tail(&req->rq_list,
1666 spin_unlock(&imp->imp_lock);
1671 req->rq_status = status;
1672 ptlrpc_rqphase_move(req,
1673 RQ_PHASE_INTERPRET);
1674 spin_unlock(&imp->imp_lock);
1675 GOTO(interpret, req->rq_status);
1677 if (ptlrpc_no_resend(req) &&
1678 !req->rq_wait_ctx) {
1679 req->rq_status = -ENOTCONN;
1680 ptlrpc_rqphase_move(req,
1681 RQ_PHASE_INTERPRET);
1682 spin_unlock(&imp->imp_lock);
1683 GOTO(interpret, req->rq_status);
1686 cfs_list_del_init(&req->rq_list);
1687 cfs_list_add_tail(&req->rq_list,
1688 &imp->imp_sending_list);
1690 spin_unlock(&imp->imp_lock);
1692 spin_lock(&req->rq_lock);
1693 req->rq_waiting = 0;
1694 spin_unlock(&req->rq_lock);
1696 if (req->rq_timedout || req->rq_resend) {
1697 /* This is re-sending anyways,
1698 * let's mark req as resend. */
1699 spin_lock(&req->rq_lock);
1701 spin_unlock(&req->rq_lock);
1705 if (!ptlrpc_unregister_bulk(req, 1))
1708 /* ensure previous bulk fails */
1709 old_xid = req->rq_xid;
1710 req->rq_xid = ptlrpc_next_xid();
1711 CDEBUG(D_HA, "resend bulk "
1714 old_xid, req->rq_xid);
1718 * rq_wait_ctx is only touched by ptlrpcd,
1719 * so no lock is needed here.
1721 status = sptlrpc_req_refresh_ctx(req, -1);
1724 req->rq_status = status;
1725 spin_lock(&req->rq_lock);
1726 req->rq_wait_ctx = 0;
1727 spin_unlock(&req->rq_lock);
1728 force_timer_recalc = 1;
1730 spin_lock(&req->rq_lock);
1731 req->rq_wait_ctx = 1;
1732 spin_unlock(&req->rq_lock);
1737 spin_lock(&req->rq_lock);
1738 req->rq_wait_ctx = 0;
1739 spin_unlock(&req->rq_lock);
1742 rc = ptl_send_rpc(req, 0);
1744 DEBUG_REQ(D_HA, req,
1745 "send failed: rc = %d", rc);
1746 force_timer_recalc = 1;
1747 spin_lock(&req->rq_lock);
1748 req->rq_net_err = 1;
1749 spin_unlock(&req->rq_lock);
1752 /* need to reset the timeout */
1753 force_timer_recalc = 1;
1756 spin_lock(&req->rq_lock);
1758 if (ptlrpc_client_early(req)) {
1759 ptlrpc_at_recv_early_reply(req);
1760 spin_unlock(&req->rq_lock);
1764 /* Still waiting for a reply? */
1765 if (ptlrpc_client_recv(req)) {
1766 spin_unlock(&req->rq_lock);
1770 /* Did we actually receive a reply? */
1771 if (!ptlrpc_client_replied(req)) {
1772 spin_unlock(&req->rq_lock);
1776 spin_unlock(&req->rq_lock);
1778 /* unlink from net because we are going to
1779 * swab in-place of reply buffer */
1780 unregistered = ptlrpc_unregister_reply(req, 1);
1784 req->rq_status = after_reply(req);
1788 /* If there is no bulk associated with this request,
1789 * then we're done and should let the interpreter
1790 * process the reply. Similarly if the RPC returned
1791 * an error, and therefore the bulk will never arrive.
1793 if (req->rq_bulk == NULL || req->rq_status < 0) {
1794 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1795 GOTO(interpret, req->rq_status);
1798 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1801 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1802 if (ptlrpc_client_bulk_active(req))
1805 if (req->rq_bulk->bd_failure) {
1806 /* The RPC reply arrived OK, but the bulk screwed
1807 * up! Dead weird since the server told us the RPC
1808 * was good after getting the REPLY for her GET or
1809 * the ACK for her PUT. */
1810 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1811 req->rq_status = -EIO;
1814 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1817 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1819 /* This moves to "unregistering" phase we need to wait for
1821 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1822 /* start async bulk unlink too */
1823 ptlrpc_unregister_bulk(req, 1);
1827 if (!ptlrpc_unregister_bulk(req, 1))
1830 /* When calling interpret receiving already should be
1832 LASSERT(!req->rq_receiving_reply);
1834 ptlrpc_req_interpret(env, req, req->rq_status);
1836 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1838 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1839 "Completed RPC pname:cluuid:pid:xid:nid:"
1840 "opc %s:%s:%d:"LPU64":%s:%d\n",
1841 current_comm(), imp->imp_obd->obd_uuid.uuid,
1842 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1843 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1844 lustre_msg_get_opc(req->rq_reqmsg));
1846 spin_lock(&imp->imp_lock);
1847 /* Request already may be not on sending or delaying list. This
1848 * may happen in the case of marking it erroneous for the case
1849 * ptlrpc_import_delay_req(req, status) find it impossible to
1850 * allow sending this rpc and returns *status != 0. */
1851 if (!cfs_list_empty(&req->rq_list)) {
1852 cfs_list_del_init(&req->rq_list);
1853 cfs_atomic_dec(&imp->imp_inflight);
1855 spin_unlock(&imp->imp_lock);
1857 cfs_atomic_dec(&set->set_remaining);
1858 wake_up_all(&imp->imp_recovery_waitq);
1860 if (set->set_producer) {
1861 /* produce a new request if possible */
1862 if (ptlrpc_set_producer(set) > 0)
1863 force_timer_recalc = 1;
1865 /* free the request that has just been completed
1866 * in order not to pollute set->set_requests */
1867 cfs_list_del_init(&req->rq_set_chain);
1868 spin_lock(&req->rq_lock);
1870 req->rq_invalid_rqset = 0;
1871 spin_unlock(&req->rq_lock);
1873 /* record rq_status to compute the final status later */
1874 if (req->rq_status != 0)
1875 set->set_rc = req->rq_status;
1876 ptlrpc_req_finished(req);
1880 /* If we hit an error, we want to recover promptly. */
1881 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1883 EXPORT_SYMBOL(ptlrpc_check_set);
1886 * Time out request \a req. is \a async_unlink is set, that means do not wait
1887 * until LNet actually confirms network buffer unlinking.
1888 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1890 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1892 struct obd_import *imp = req->rq_import;
1896 spin_lock(&req->rq_lock);
1897 req->rq_timedout = 1;
1898 spin_unlock(&req->rq_lock);
1900 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1901 "/real "CFS_DURATION_T"]",
1902 req->rq_net_err ? "failed due to network error" :
1903 ((req->rq_real_sent == 0 ||
1904 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1905 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1906 "timed out for sent delay" : "timed out for slow reply"),
1907 req->rq_sent, req->rq_real_sent);
1909 if (imp != NULL && obd_debug_peer_on_timeout)
1910 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1912 ptlrpc_unregister_reply(req, async_unlink);
1913 ptlrpc_unregister_bulk(req, async_unlink);
1915 if (obd_dump_on_timeout)
1916 libcfs_debug_dumplog();
1919 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1923 cfs_atomic_inc(&imp->imp_timeouts);
1925 /* The DLM server doesn't want recovery run on its imports. */
1926 if (imp->imp_dlm_fake)
1929 /* If this request is for recovery or other primordial tasks,
1930 * then error it out here. */
1931 if (req->rq_ctx_init || req->rq_ctx_fini ||
1932 req->rq_send_state != LUSTRE_IMP_FULL ||
1933 imp->imp_obd->obd_no_recov) {
1934 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1935 ptlrpc_import_state_name(req->rq_send_state),
1936 ptlrpc_import_state_name(imp->imp_state));
1937 spin_lock(&req->rq_lock);
1938 req->rq_status = -ETIMEDOUT;
1940 spin_unlock(&req->rq_lock);
1944 /* if a request can't be resent we can't wait for an answer after
1946 if (ptlrpc_no_resend(req)) {
1947 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1951 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1957 * Time out all uncompleted requests in request set pointed by \a data
1958 * Callback used when waiting on sets with l_wait_event.
1961 int ptlrpc_expired_set(void *data)
1963 struct ptlrpc_request_set *set = data;
1965 time_t now = cfs_time_current_sec();
1968 LASSERT(set != NULL);
1971 * A timeout expired. See which reqs it applies to...
1973 cfs_list_for_each (tmp, &set->set_requests) {
1974 struct ptlrpc_request *req =
1975 cfs_list_entry(tmp, struct ptlrpc_request,
1978 /* don't expire request waiting for context */
1979 if (req->rq_wait_ctx)
1982 /* Request in-flight? */
1983 if (!((req->rq_phase == RQ_PHASE_RPC &&
1984 !req->rq_waiting && !req->rq_resend) ||
1985 (req->rq_phase == RQ_PHASE_BULK)))
1988 if (req->rq_timedout || /* already dealt with */
1989 req->rq_deadline > now) /* not expired */
1992 /* Deal with this guy. Do it asynchronously to not block
1993 * ptlrpcd thread. */
1994 ptlrpc_expire_one_request(req, 1);
1998 * When waiting for a whole set, we always break out of the
1999 * sleep so we can recalculate the timeout, or enable interrupts
2000 * if everyone's timed out.
2004 EXPORT_SYMBOL(ptlrpc_expired_set);
2007 * Sets rq_intr flag in \a req under spinlock.
2009 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2011 spin_lock(&req->rq_lock);
2013 spin_unlock(&req->rq_lock);
2015 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2018 * Interrupts (sets interrupted flag) all uncompleted requests in
2019 * a set \a data. Callback for l_wait_event for interruptible waits.
2021 void ptlrpc_interrupted_set(void *data)
2023 struct ptlrpc_request_set *set = data;
2026 LASSERT(set != NULL);
2027 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2029 cfs_list_for_each(tmp, &set->set_requests) {
2030 struct ptlrpc_request *req =
2031 cfs_list_entry(tmp, struct ptlrpc_request,
2034 if (req->rq_phase != RQ_PHASE_RPC &&
2035 req->rq_phase != RQ_PHASE_UNREGISTERING)
2038 ptlrpc_mark_interrupted(req);
2041 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2044 * Get the smallest timeout in the set; this does NOT set a timeout.
2046 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2049 time_t now = cfs_time_current_sec();
2051 struct ptlrpc_request *req;
2055 cfs_list_for_each(tmp, &set->set_requests) {
2056 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2059 * Request in-flight?
2061 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2062 (req->rq_phase == RQ_PHASE_BULK) ||
2063 (req->rq_phase == RQ_PHASE_NEW)))
2067 * Already timed out.
2069 if (req->rq_timedout)
2075 if (req->rq_wait_ctx)
2078 if (req->rq_phase == RQ_PHASE_NEW)
2079 deadline = req->rq_sent;
2080 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2081 deadline = req->rq_sent;
2083 deadline = req->rq_sent + req->rq_timeout;
2085 if (deadline <= now) /* actually expired already */
2086 timeout = 1; /* ASAP */
2087 else if (timeout == 0 || timeout > deadline - now)
2088 timeout = deadline - now;
2092 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2095 * Send all unset request from the set and then wait untill all
2096 * requests in the set complete (either get a reply, timeout, get an
2097 * error or otherwise be interrupted).
2098 * Returns 0 on success or error code otherwise.
2100 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2103 struct ptlrpc_request *req;
2104 struct l_wait_info lwi;
2108 if (set->set_producer)
2109 (void)ptlrpc_set_producer(set);
2111 cfs_list_for_each(tmp, &set->set_requests) {
2112 req = cfs_list_entry(tmp, struct ptlrpc_request,
2114 if (req->rq_phase == RQ_PHASE_NEW)
2115 (void)ptlrpc_send_new_req(req);
2118 if (cfs_list_empty(&set->set_requests))
2122 timeout = ptlrpc_set_next_timeout(set);
2124 /* wait until all complete, interrupted, or an in-flight
2126 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2129 if (timeout == 0 && !cfs_signal_pending())
2131 * No requests are in-flight (ether timed out
2132 * or delayed), so we can allow interrupts.
2133 * We still want to block for a limited time,
2134 * so we allow interrupts during the timeout.
2136 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2138 ptlrpc_interrupted_set, set);
2141 * At least one request is in flight, so no
2142 * interrupts are allowed. Wait until all
2143 * complete, or an in-flight req times out.
2145 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2146 ptlrpc_expired_set, set);
2148 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2150 /* LU-769 - if we ignored the signal because it was already
2151 * pending when we started, we need to handle it now or we risk
2152 * it being ignored forever */
2153 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2154 cfs_signal_pending()) {
2155 sigset_t blocked_sigs =
2156 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2158 /* In fact we only interrupt for the "fatal" signals
2159 * like SIGINT or SIGKILL. We still ignore less
2160 * important signals since ptlrpc set is not easily
2161 * reentrant from userspace again */
2162 if (cfs_signal_pending())
2163 ptlrpc_interrupted_set(set);
2164 cfs_restore_sigs(blocked_sigs);
2167 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2169 /* -EINTR => all requests have been flagged rq_intr so next
2171 * -ETIMEDOUT => someone timed out. When all reqs have
2172 * timed out, signals are enabled allowing completion with
2174 * I don't really care if we go once more round the loop in
2175 * the error cases -eeb. */
2176 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2177 cfs_list_for_each(tmp, &set->set_requests) {
2178 req = cfs_list_entry(tmp, struct ptlrpc_request,
2180 spin_lock(&req->rq_lock);
2181 req->rq_invalid_rqset = 1;
2182 spin_unlock(&req->rq_lock);
2185 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2187 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2189 rc = set->set_rc; /* rq_status of already freed requests if any */
2190 cfs_list_for_each(tmp, &set->set_requests) {
2191 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2193 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2194 if (req->rq_status != 0)
2195 rc = req->rq_status;
2198 if (set->set_interpret != NULL) {
2199 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2201 rc = interpreter (set, set->set_arg, rc);
2203 struct ptlrpc_set_cbdata *cbdata, *n;
2206 cfs_list_for_each_entry_safe(cbdata, n,
2207 &set->set_cblist, psc_item) {
2208 cfs_list_del_init(&cbdata->psc_item);
2209 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2212 OBD_FREE_PTR(cbdata);
2218 EXPORT_SYMBOL(ptlrpc_set_wait);
2221 * Helper fuction for request freeing.
2222 * Called when request count reached zero and request needs to be freed.
2223 * Removes request from all sorts of sending/replay lists it might be on,
2224 * frees network buffers if any are present.
2225 * If \a locked is set, that means caller is already holding import imp_lock
2226 * and so we no longer need to reobtain it (for certain lists manipulations)
2228 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2231 if (request == NULL) {
2236 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2237 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2238 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2239 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2240 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2241 LASSERTF(!request->rq_replay, "req %p\n", request);
2243 req_capsule_fini(&request->rq_pill);
2245 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2246 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2247 if (request->rq_import != NULL) {
2249 spin_lock(&request->rq_import->imp_lock);
2250 cfs_list_del_init(&request->rq_replay_list);
2252 spin_unlock(&request->rq_import->imp_lock);
2254 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2256 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2257 DEBUG_REQ(D_ERROR, request,
2258 "freeing request with nonzero refcount");
2262 if (request->rq_repbuf != NULL)
2263 sptlrpc_cli_free_repbuf(request);
2264 if (request->rq_export != NULL) {
2265 class_export_put(request->rq_export);
2266 request->rq_export = NULL;
2268 if (request->rq_import != NULL) {
2269 class_import_put(request->rq_import);
2270 request->rq_import = NULL;
2272 if (request->rq_bulk != NULL)
2273 ptlrpc_free_bulk_pin(request->rq_bulk);
2275 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2276 sptlrpc_cli_free_reqbuf(request);
2278 if (request->rq_cli_ctx)
2279 sptlrpc_req_put_ctx(request, !locked);
2281 if (request->rq_pool)
2282 __ptlrpc_free_req_to_pool(request);
2284 ptlrpc_request_cache_free(request);
2288 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2290 * Drop one request reference. Must be called with import imp_lock held.
2291 * When reference count drops to zero, reuqest is freed.
2293 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2295 LASSERT(spin_is_locked(&request->rq_import->imp_lock));
2296 (void)__ptlrpc_req_finished(request, 1);
2298 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2302 * Drops one reference count for request \a request.
2303 * \a locked set indicates that caller holds import imp_lock.
2304 * Frees the request whe reference count reaches zero.
2306 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2309 if (request == NULL)
2312 if (request == LP_POISON ||
2313 request->rq_reqmsg == LP_POISON) {
2314 CERROR("dereferencing freed request (bug 575)\n");
2319 DEBUG_REQ(D_INFO, request, "refcount now %u",
2320 cfs_atomic_read(&request->rq_refcount) - 1);
2322 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2323 __ptlrpc_free_req(request, locked);
2331 * Drops one reference count for a request.
2333 void ptlrpc_req_finished(struct ptlrpc_request *request)
2335 __ptlrpc_req_finished(request, 0);
2337 EXPORT_SYMBOL(ptlrpc_req_finished);
2340 * Returns xid of a \a request
2342 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2344 return request->rq_xid;
2346 EXPORT_SYMBOL(ptlrpc_req_xid);
2349 * Disengage the client's reply buffer from the network
2350 * NB does _NOT_ unregister any client-side bulk.
2351 * IDEMPOTENT, but _not_ safe against concurrent callers.
2352 * The request owner (i.e. the thread doing the I/O) must call...
2353 * Returns 0 on success or 1 if unregistering cannot be made.
2355 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2358 struct l_wait_info lwi;
2363 LASSERT(!in_interrupt());
2366 * Let's setup deadline for reply unlink.
2368 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2369 async && request->rq_reply_deadline == 0)
2370 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2373 * Nothing left to do.
2375 if (!ptlrpc_client_recv_or_unlink(request))
2378 LNetMDUnlink(request->rq_reply_md_h);
2381 * Let's check it once again.
2383 if (!ptlrpc_client_recv_or_unlink(request))
2387 * Move to "Unregistering" phase as reply was not unlinked yet.
2389 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2392 * Do not wait for unlink to finish.
2398 * We have to l_wait_event() whatever the result, to give liblustre
2399 * a chance to run reply_in_callback(), and to make sure we've
2400 * unlinked before returning a req to the pool.
2404 /* The wq argument is ignored by user-space wait_event macros */
2405 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2406 &request->rq_set->set_waitq :
2407 &request->rq_reply_waitq;
2409 /* Network access will complete in finite time but the HUGE
2410 * timeout lets us CWARN for visibility of sluggish NALs */
2411 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2412 cfs_time_seconds(1), NULL, NULL);
2413 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2416 ptlrpc_rqphase_move(request, request->rq_next_phase);
2420 LASSERT(rc == -ETIMEDOUT);
2421 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2422 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2423 request->rq_must_unlink);
2427 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2429 static void ptlrpc_free_request(struct ptlrpc_request *req)
2431 spin_lock(&req->rq_lock);
2433 spin_unlock(&req->rq_lock);
2435 if (req->rq_commit_cb != NULL)
2436 req->rq_commit_cb(req);
2437 cfs_list_del_init(&req->rq_replay_list);
2439 __ptlrpc_req_finished(req, 1);
2443 * the request is committed and dropped from the replay list of its import
2445 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2447 struct obd_import *imp = req->rq_import;
2449 spin_lock(&imp->imp_lock);
2450 if (cfs_list_empty(&req->rq_replay_list)) {
2451 spin_unlock(&imp->imp_lock);
2455 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2456 ptlrpc_free_request(req);
2458 spin_unlock(&imp->imp_lock);
2460 EXPORT_SYMBOL(ptlrpc_request_committed);
2463 * Iterates through replay_list on import and prunes
2464 * all requests have transno smaller than last_committed for the
2465 * import and don't have rq_replay set.
2466 * Since requests are sorted in transno order, stops when meetign first
2467 * transno bigger than last_committed.
2468 * caller must hold imp->imp_lock
2470 void ptlrpc_free_committed(struct obd_import *imp)
2472 struct ptlrpc_request *req, *saved;
2473 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2474 bool skip_committed_list = true;
2477 LASSERT(imp != NULL);
2478 LASSERT(spin_is_locked(&imp->imp_lock));
2481 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2482 imp->imp_generation == imp->imp_last_generation_checked) {
2483 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2484 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2487 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2488 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2489 imp->imp_generation);
2491 if (imp->imp_generation != imp->imp_last_generation_checked)
2492 skip_committed_list = false;
2494 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2495 imp->imp_last_generation_checked = imp->imp_generation;
2497 cfs_list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2499 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2500 LASSERT(req != last_req);
2503 if (req->rq_transno == 0) {
2504 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2507 if (req->rq_import_generation < imp->imp_generation) {
2508 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2512 /* not yet committed */
2513 if (req->rq_transno > imp->imp_peer_committed_transno) {
2514 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2518 if (req->rq_replay) {
2519 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2520 cfs_list_move_tail(&req->rq_replay_list,
2521 &imp->imp_committed_list);
2525 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2526 imp->imp_peer_committed_transno);
2528 ptlrpc_free_request(req);
2531 if (skip_committed_list)
2534 cfs_list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2536 LASSERT(req->rq_transno != 0);
2537 if (req->rq_import_generation < imp->imp_generation) {
2538 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2539 ptlrpc_free_request(req);
2546 void ptlrpc_cleanup_client(struct obd_import *imp)
2551 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2554 * Schedule previously sent request for resend.
2555 * For bulk requests we assign new xid (to avoid problems with
2556 * lost replies and therefore several transfers landing into same buffer
2557 * from different sending attempts).
2559 void ptlrpc_resend_req(struct ptlrpc_request *req)
2561 DEBUG_REQ(D_HA, req, "going to resend");
2562 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2563 req->rq_status = -EAGAIN;
2565 spin_lock(&req->rq_lock);
2567 req->rq_net_err = 0;
2568 req->rq_timedout = 0;
2570 __u64 old_xid = req->rq_xid;
2572 /* ensure previous bulk fails */
2573 req->rq_xid = ptlrpc_next_xid();
2574 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2575 old_xid, req->rq_xid);
2577 ptlrpc_client_wake_req(req);
2578 spin_unlock(&req->rq_lock);
2580 EXPORT_SYMBOL(ptlrpc_resend_req);
2582 /* XXX: this function and rq_status are currently unused */
2583 void ptlrpc_restart_req(struct ptlrpc_request *req)
2585 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2586 req->rq_status = -ERESTARTSYS;
2588 spin_lock(&req->rq_lock);
2589 req->rq_restart = 1;
2590 req->rq_timedout = 0;
2591 ptlrpc_client_wake_req(req);
2592 spin_unlock(&req->rq_lock);
2594 EXPORT_SYMBOL(ptlrpc_restart_req);
2597 * Grab additional reference on a request \a req
2599 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2602 cfs_atomic_inc(&req->rq_refcount);
2605 EXPORT_SYMBOL(ptlrpc_request_addref);
2608 * Add a request to import replay_list.
2609 * Must be called under imp_lock
2611 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2612 struct obd_import *imp)
2616 LASSERT(spin_is_locked(&imp->imp_lock));
2618 if (req->rq_transno == 0) {
2619 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2623 /* clear this for new requests that were resent as well
2624 as resent replayed requests. */
2625 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2627 /* don't re-add requests that have been replayed */
2628 if (!cfs_list_empty(&req->rq_replay_list))
2631 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2633 LASSERT(imp->imp_replayable);
2634 /* Balanced in ptlrpc_free_committed, usually. */
2635 ptlrpc_request_addref(req);
2636 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2637 struct ptlrpc_request *iter =
2638 cfs_list_entry(tmp, struct ptlrpc_request,
2641 /* We may have duplicate transnos if we create and then
2642 * open a file, or for closes retained if to match creating
2643 * opens, so use req->rq_xid as a secondary key.
2644 * (See bugs 684, 685, and 428.)
2645 * XXX no longer needed, but all opens need transnos!
2647 if (iter->rq_transno > req->rq_transno)
2650 if (iter->rq_transno == req->rq_transno) {
2651 LASSERT(iter->rq_xid != req->rq_xid);
2652 if (iter->rq_xid > req->rq_xid)
2656 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2660 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2662 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2665 * Send request and wait until it completes.
2666 * Returns request processing status.
2668 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2670 struct ptlrpc_request_set *set;
2674 LASSERT(req->rq_set == NULL);
2675 LASSERT(!req->rq_receiving_reply);
2677 set = ptlrpc_prep_set();
2679 CERROR("Unable to allocate ptlrpc set.");
2683 /* for distributed debugging */
2684 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2686 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2687 ptlrpc_request_addref(req);
2688 ptlrpc_set_add_req(set, req);
2689 rc = ptlrpc_set_wait(set);
2690 ptlrpc_set_destroy(set);
2694 EXPORT_SYMBOL(ptlrpc_queue_wait);
2696 struct ptlrpc_replay_async_args {
2698 int praa_old_status;
2702 * Callback used for replayed requests reply processing.
2703 * In case of succesful reply calls registeresd request replay callback.
2704 * In case of error restart replay process.
2706 static int ptlrpc_replay_interpret(const struct lu_env *env,
2707 struct ptlrpc_request *req,
2708 void * data, int rc)
2710 struct ptlrpc_replay_async_args *aa = data;
2711 struct obd_import *imp = req->rq_import;
2714 cfs_atomic_dec(&imp->imp_replay_inflight);
2716 if (!ptlrpc_client_replied(req)) {
2717 CERROR("request replay timed out, restarting recovery\n");
2718 GOTO(out, rc = -ETIMEDOUT);
2721 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2722 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2723 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2724 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2726 /** VBR: check version failure */
2727 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2728 /** replay was failed due to version mismatch */
2729 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2730 spin_lock(&imp->imp_lock);
2731 imp->imp_vbr_failed = 1;
2732 imp->imp_no_lock_replay = 1;
2733 spin_unlock(&imp->imp_lock);
2734 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2736 /** The transno had better not change over replay. */
2737 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2738 lustre_msg_get_transno(req->rq_repmsg) ||
2739 lustre_msg_get_transno(req->rq_repmsg) == 0,
2741 lustre_msg_get_transno(req->rq_reqmsg),
2742 lustre_msg_get_transno(req->rq_repmsg));
2745 spin_lock(&imp->imp_lock);
2746 /** if replays by version then gap occur on server, no trust to locks */
2747 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2748 imp->imp_no_lock_replay = 1;
2749 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2750 spin_unlock(&imp->imp_lock);
2751 LASSERT(imp->imp_last_replay_transno);
2753 /* transaction number shouldn't be bigger than the latest replayed */
2754 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2755 DEBUG_REQ(D_ERROR, req,
2756 "Reported transno "LPU64" is bigger than the "
2757 "replayed one: "LPU64, req->rq_transno,
2758 lustre_msg_get_transno(req->rq_reqmsg));
2759 GOTO(out, rc = -EINVAL);
2762 DEBUG_REQ(D_HA, req, "got rep");
2764 /* let the callback do fixups, possibly including in the request */
2765 if (req->rq_replay_cb)
2766 req->rq_replay_cb(req);
2768 if (ptlrpc_client_replied(req) &&
2769 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2770 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2771 lustre_msg_get_status(req->rq_repmsg),
2772 aa->praa_old_status);
2774 /* Put it back for re-replay. */
2775 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2779 * Errors while replay can set transno to 0, but
2780 * imp_last_replay_transno shouldn't be set to 0 anyway
2782 if (req->rq_transno == 0)
2783 CERROR("Transno is 0 during replay!\n");
2785 /* continue with recovery */
2786 rc = ptlrpc_import_recovery_state_machine(imp);
2788 req->rq_send_state = aa->praa_old_state;
2791 /* this replay failed, so restart recovery */
2792 ptlrpc_connect_import(imp);
2798 * Prepares and queues request for replay.
2799 * Adds it to ptlrpcd queue for actual sending.
2800 * Returns 0 on success.
2802 int ptlrpc_replay_req(struct ptlrpc_request *req)
2804 struct ptlrpc_replay_async_args *aa;
2807 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2809 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2810 aa = ptlrpc_req_async_args(req);
2811 memset(aa, 0, sizeof *aa);
2813 /* Prepare request to be resent with ptlrpcd */
2814 aa->praa_old_state = req->rq_send_state;
2815 req->rq_send_state = LUSTRE_IMP_REPLAY;
2816 req->rq_phase = RQ_PHASE_NEW;
2817 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2819 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2821 req->rq_interpret_reply = ptlrpc_replay_interpret;
2822 /* Readjust the timeout for current conditions */
2823 ptlrpc_at_set_req_timeout(req);
2825 /* Tell server the net_latency, so the server can calculate how long
2826 * it should wait for next replay */
2827 lustre_msg_set_service_time(req->rq_reqmsg,
2828 ptlrpc_at_get_net_latency(req));
2829 DEBUG_REQ(D_HA, req, "REPLAY");
2831 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2832 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2834 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2837 EXPORT_SYMBOL(ptlrpc_replay_req);
2840 * Aborts all in-flight request on import \a imp sending and delayed lists
2842 void ptlrpc_abort_inflight(struct obd_import *imp)
2844 cfs_list_t *tmp, *n;
2847 /* Make sure that no new requests get processed for this import.
2848 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2849 * this flag and then putting requests on sending_list or delayed_list.
2851 spin_lock(&imp->imp_lock);
2853 /* XXX locking? Maybe we should remove each request with the list
2854 * locked? Also, how do we know if the requests on the list are
2855 * being freed at this time?
2857 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2858 struct ptlrpc_request *req =
2859 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2861 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2863 spin_lock(&req->rq_lock);
2864 if (req->rq_import_generation < imp->imp_generation) {
2866 req->rq_status = -EIO;
2867 ptlrpc_client_wake_req(req);
2869 spin_unlock(&req->rq_lock);
2872 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2873 struct ptlrpc_request *req =
2874 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2876 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2878 spin_lock(&req->rq_lock);
2879 if (req->rq_import_generation < imp->imp_generation) {
2881 req->rq_status = -EIO;
2882 ptlrpc_client_wake_req(req);
2884 spin_unlock(&req->rq_lock);
2887 /* Last chance to free reqs left on the replay list, but we
2888 * will still leak reqs that haven't committed. */
2889 if (imp->imp_replayable)
2890 ptlrpc_free_committed(imp);
2892 spin_unlock(&imp->imp_lock);
2896 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2899 * Abort all uncompleted requests in request set \a set
2901 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2903 cfs_list_t *tmp, *pos;
2905 LASSERT(set != NULL);
2907 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2908 struct ptlrpc_request *req =
2909 cfs_list_entry(pos, struct ptlrpc_request,
2912 spin_lock(&req->rq_lock);
2913 if (req->rq_phase != RQ_PHASE_RPC) {
2914 spin_unlock(&req->rq_lock);
2919 req->rq_status = -EINTR;
2920 ptlrpc_client_wake_req(req);
2921 spin_unlock(&req->rq_lock);
2925 static __u64 ptlrpc_last_xid;
2926 static spinlock_t ptlrpc_last_xid_lock;
2929 * Initialize the XID for the node. This is common among all requests on
2930 * this node, and only requires the property that it is monotonically
2931 * increasing. It does not need to be sequential. Since this is also used
2932 * as the RDMA match bits, it is important that a single client NOT have
2933 * the same match bits for two different in-flight requests, hence we do
2934 * NOT want to have an XID per target or similar.
2936 * To avoid an unlikely collision between match bits after a client reboot
2937 * (which would deliver old data into the wrong RDMA buffer) initialize
2938 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2939 * If the time is clearly incorrect, we instead use a 62-bit random number.
2940 * In the worst case the random number will overflow 1M RPCs per second in
2941 * 9133 years, or permutations thereof.
2943 #define YEAR_2004 (1ULL << 30)
2944 void ptlrpc_init_xid(void)
2946 time_t now = cfs_time_current_sec();
2948 spin_lock_init(&ptlrpc_last_xid_lock);
2949 if (now < YEAR_2004) {
2950 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2951 ptlrpc_last_xid >>= 2;
2952 ptlrpc_last_xid |= (1ULL << 61);
2954 ptlrpc_last_xid = (__u64)now << 20;
2957 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
2958 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
2959 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
2963 * Increase xid and returns resulting new value to the caller.
2965 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2966 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2967 * itself uses the last bulk xid needed, so the server can determine the
2968 * the number of bulk transfers from the RPC XID and a bitmask. The starting
2969 * xid must align to a power-of-two value.
2971 * This is assumed to be true due to the initial ptlrpc_last_xid
2972 * value also being initialized to a power-of-two value. LU-1431
2974 __u64 ptlrpc_next_xid(void)
2978 spin_lock(&ptlrpc_last_xid_lock);
2979 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2980 ptlrpc_last_xid = next;
2981 spin_unlock(&ptlrpc_last_xid_lock);
2985 EXPORT_SYMBOL(ptlrpc_next_xid);
2988 * Get a glimpse at what next xid value might have been.
2989 * Returns possible next xid.
2991 __u64 ptlrpc_sample_next_xid(void)
2993 #if BITS_PER_LONG == 32
2994 /* need to avoid possible word tearing on 32-bit systems */
2997 spin_lock(&ptlrpc_last_xid_lock);
2998 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
2999 spin_unlock(&ptlrpc_last_xid_lock);
3003 /* No need to lock, since returned value is racy anyways */
3004 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3007 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3010 * Functions for operating ptlrpc workers.
3012 * A ptlrpc work is a function which will be running inside ptlrpc context.
3013 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3015 * 1. after a work is created, it can be used many times, that is:
3016 * handler = ptlrpcd_alloc_work();
3017 * ptlrpcd_queue_work();
3019 * queue it again when necessary:
3020 * ptlrpcd_queue_work();
3021 * ptlrpcd_destroy_work();
3022 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3023 * it will only be queued once in any time. Also as its name implies, it may
3024 * have delay before it really runs by ptlrpcd thread.
3026 struct ptlrpc_work_async_args {
3028 int (*cb)(const struct lu_env *, void *);
3032 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
3034 static int work_interpreter(const struct lu_env *env,
3035 struct ptlrpc_request *req, void *data, int rc)
3037 struct ptlrpc_work_async_args *arg = data;
3039 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
3040 LASSERT(arg->cb != NULL);
3042 return arg->cb(env, arg->cbdata);
3046 * Create a work for ptlrpc.
3048 void *ptlrpcd_alloc_work(struct obd_import *imp,
3049 int (*cb)(const struct lu_env *, void *), void *cbdata)
3051 struct ptlrpc_request *req = NULL;
3052 struct ptlrpc_work_async_args *args;
3058 RETURN(ERR_PTR(-EINVAL));
3060 /* copy some code from deprecated fakereq. */
3061 req = ptlrpc_request_cache_alloc(__GFP_IO);
3063 CERROR("ptlrpc: run out of memory!\n");
3064 RETURN(ERR_PTR(-ENOMEM));
3067 req->rq_send_state = LUSTRE_IMP_FULL;
3068 req->rq_type = PTL_RPC_MSG_REQUEST;
3069 req->rq_import = class_import_get(imp);
3070 req->rq_export = NULL;
3071 req->rq_interpret_reply = work_interpreter;
3072 /* don't want reply */
3073 req->rq_receiving_reply = 0;
3074 req->rq_must_unlink = 0;
3075 req->rq_no_delay = req->rq_no_resend = 1;
3077 spin_lock_init(&req->rq_lock);
3078 CFS_INIT_LIST_HEAD(&req->rq_list);
3079 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
3080 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
3081 CFS_INIT_LIST_HEAD(&req->rq_history_list);
3082 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
3083 init_waitqueue_head(&req->rq_reply_waitq);
3084 init_waitqueue_head(&req->rq_set_waitq);
3085 cfs_atomic_set(&req->rq_refcount, 1);
3087 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3088 args = ptlrpc_req_async_args(req);
3089 args->magic = PTLRPC_WORK_MAGIC;
3091 args->cbdata = cbdata;
3095 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3097 void ptlrpcd_destroy_work(void *handler)
3099 struct ptlrpc_request *req = handler;
3102 ptlrpc_req_finished(req);
3104 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3106 int ptlrpcd_queue_work(void *handler)
3108 struct ptlrpc_request *req = handler;
3111 * Check if the req is already being queued.
3113 * Here comes a trick: it lacks a way of checking if a req is being
3114 * processed reliably in ptlrpc. Here I have to use refcount of req
3115 * for this purpose. This is okay because the caller should use this
3116 * req as opaque data. - Jinshan
3118 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
3119 if (cfs_atomic_read(&req->rq_refcount) > 1)
3122 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
3123 cfs_atomic_dec(&req->rq_refcount);
3127 /* re-initialize the req */
3128 req->rq_timeout = obd_timeout;
3129 req->rq_sent = cfs_time_current_sec();
3130 req->rq_deadline = req->rq_sent + req->rq_timeout;
3131 req->rq_reply_deadline = req->rq_deadline;
3132 req->rq_phase = RQ_PHASE_INTERPRET;
3133 req->rq_next_phase = RQ_PHASE_COMPLETE;
3134 req->rq_xid = ptlrpc_next_xid();
3135 req->rq_import_generation = req->rq_import->imp_generation;
3137 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3140 EXPORT_SYMBOL(ptlrpcd_queue_work);