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
41 #include <obd_support.h>
42 #include <obd_class.h>
43 #include <lustre_lib.h>
44 #include <lustre_ha.h>
45 #include <lustre_import.h>
46 #include <lustre_req_layout.h>
48 #include "ptlrpc_internal.h"
50 static int ptlrpc_send_new_req(struct ptlrpc_request *req);
51 static int ptlrpcd_check_work(struct ptlrpc_request *req);
54 * Initialize passed in client structure \a cl.
56 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
57 struct ptlrpc_client *cl)
59 cl->cli_request_portal = req_portal;
60 cl->cli_reply_portal = rep_portal;
63 EXPORT_SYMBOL(ptlrpc_init_client);
66 * Return PortalRPC connection for remore uud \a uuid
68 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
70 struct ptlrpc_connection *c;
72 lnet_process_id_t peer;
75 /* ptlrpc_uuid_to_peer() initializes its 2nd parameter
76 * before accessing its values. */
77 /* coverity[uninit_use_in_call] */
78 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
80 CNETERR("cannot find peer %s!\n", uuid->uuid);
84 c = ptlrpc_connection_get(peer, self, uuid);
86 memcpy(c->c_remote_uuid.uuid,
87 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
90 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
94 EXPORT_SYMBOL(ptlrpc_uuid_to_connection);
97 * Allocate and initialize new bulk descriptor on the sender.
98 * Returns pointer to the descriptor or NULL on error.
100 struct ptlrpc_bulk_desc *ptlrpc_new_bulk(unsigned npages, unsigned max_brw,
101 unsigned type, unsigned portal)
103 struct ptlrpc_bulk_desc *desc;
106 OBD_ALLOC(desc, offsetof(struct ptlrpc_bulk_desc, bd_iov[npages]));
110 spin_lock_init(&desc->bd_lock);
111 init_waitqueue_head(&desc->bd_waitq);
112 desc->bd_max_iov = npages;
113 desc->bd_iov_count = 0;
114 desc->bd_portal = portal;
115 desc->bd_type = type;
116 desc->bd_md_count = 0;
117 LASSERT(max_brw > 0);
118 desc->bd_md_max_brw = min(max_brw, PTLRPC_BULK_OPS_COUNT);
119 /* PTLRPC_BULK_OPS_COUNT is the compile-time transfer limit for this
120 * node. Negotiated ocd_brw_size will always be <= this number. */
121 for (i = 0; i < PTLRPC_BULK_OPS_COUNT; i++)
122 LNetInvalidateHandle(&desc->bd_mds[i]);
128 * Prepare bulk descriptor for specified outgoing request \a req that
129 * can fit \a npages * pages. \a type is bulk type. \a portal is where
130 * the bulk to be sent. Used on client-side.
131 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
134 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
135 unsigned npages, unsigned max_brw,
136 unsigned type, unsigned portal)
138 struct obd_import *imp = req->rq_import;
139 struct ptlrpc_bulk_desc *desc;
142 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
143 desc = ptlrpc_new_bulk(npages, max_brw, type, portal);
147 desc->bd_import_generation = req->rq_import_generation;
148 desc->bd_import = class_import_get(imp);
151 desc->bd_cbid.cbid_fn = client_bulk_callback;
152 desc->bd_cbid.cbid_arg = desc;
154 /* This makes req own desc, and free it when she frees herself */
159 EXPORT_SYMBOL(ptlrpc_prep_bulk_imp);
162 * Add a page \a page to the bulk descriptor \a desc.
163 * Data to transfer in the page starts at offset \a pageoffset and
164 * amount of data to transfer from the page is \a len
166 void __ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
167 struct page *page, int pageoffset, int len, int pin)
169 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
170 LASSERT(page != NULL);
171 LASSERT(pageoffset >= 0);
173 LASSERT(pageoffset + len <= PAGE_CACHE_SIZE);
178 page_cache_get(page);
180 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
182 EXPORT_SYMBOL(__ptlrpc_prep_bulk_page);
185 * Uninitialize and free bulk descriptor \a desc.
186 * Works on bulk descriptors both from server and client side.
188 void __ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc, int unpin)
193 LASSERT(desc != NULL);
194 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
195 LASSERT(desc->bd_md_count == 0); /* network hands off */
196 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
198 sptlrpc_enc_pool_put_pages(desc);
201 class_export_put(desc->bd_export);
203 class_import_put(desc->bd_import);
206 for (i = 0; i < desc->bd_iov_count ; i++)
207 page_cache_release(desc->bd_iov[i].kiov_page);
210 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
211 bd_iov[desc->bd_max_iov]));
214 EXPORT_SYMBOL(__ptlrpc_free_bulk);
217 * Set server timelimit for this req, i.e. how long are we willing to wait
218 * for reply before timing out this request.
220 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
226 LASSERT(req->rq_import);
229 /* non-AT settings */
231 * \a imp_server_timeout means this is reverse import and
232 * we send (currently only) ASTs to the client and cannot afford
233 * to wait too long for the reply, otherwise the other client
234 * (because of which we are sending this request) would
235 * timeout waiting for us
237 req->rq_timeout = req->rq_import->imp_server_timeout ?
238 obd_timeout / 2 : obd_timeout;
240 at = &req->rq_import->imp_at;
241 idx = import_at_get_index(req->rq_import,
242 req->rq_request_portal);
243 serv_est = at_get(&at->iat_service_estimate[idx]);
244 req->rq_timeout = at_est2timeout(serv_est);
246 /* We could get even fancier here, using history to predict increased
249 /* Let the server know what this RPC timeout is by putting it in the
251 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
253 EXPORT_SYMBOL(ptlrpc_at_set_req_timeout);
255 /* Adjust max service estimate based on server value */
256 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
257 unsigned int serv_est)
263 LASSERT(req->rq_import);
264 at = &req->rq_import->imp_at;
266 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
267 /* max service estimates are tracked on the server side,
268 so just keep minimal history here */
269 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
271 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
272 "has changed from %d to %d\n",
273 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
274 oldse, at_get(&at->iat_service_estimate[idx]));
277 /* Expected network latency per remote node (secs) */
278 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
280 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
283 /* Adjust expected network latency */
284 void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
285 unsigned int service_time)
287 unsigned int nl, oldnl;
289 time_t now = cfs_time_current_sec();
291 LASSERT(req->rq_import);
293 if (service_time > now - req->rq_sent + 3) {
294 /* bz16408, however, this can also happen if early reply
295 * is lost and client RPC is expired and resent, early reply
296 * or reply of original RPC can still be fit in reply buffer
297 * of resent RPC, now client is measuring time from the
298 * resent time, but server sent back service time of original
301 CDEBUG((lustre_msg_get_flags(req->rq_reqmsg) & MSG_RESENT) ?
302 D_ADAPTTO : D_WARNING,
303 "Reported service time %u > total measured time "
304 CFS_DURATION_T"\n", service_time,
305 cfs_time_sub(now, req->rq_sent));
309 /* Network latency is total time less server processing time */
310 nl = max_t(int, now - req->rq_sent -
311 service_time, 0) + 1; /* st rounding */
312 at = &req->rq_import->imp_at;
314 oldnl = at_measured(&at->iat_net_latency, nl);
316 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
317 "has changed from %d to %d\n",
318 req->rq_import->imp_obd->obd_name,
320 &req->rq_import->imp_connection->c_remote_uuid),
321 oldnl, at_get(&at->iat_net_latency));
324 static int unpack_reply(struct ptlrpc_request *req)
328 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
329 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
331 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
336 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
338 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
345 * Handle an early reply message, called with the rq_lock held.
346 * If anything goes wrong just ignore it - same as if it never happened
348 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
349 __must_hold(&req->rq_lock)
351 struct ptlrpc_request *early_req;
357 spin_unlock(&req->rq_lock);
359 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
361 spin_lock(&req->rq_lock);
365 rc = unpack_reply(early_req);
367 /* Expecting to increase the service time estimate here */
368 ptlrpc_at_adj_service(req,
369 lustre_msg_get_timeout(early_req->rq_repmsg));
370 ptlrpc_at_adj_net_latency(req,
371 lustre_msg_get_service_time(early_req->rq_repmsg));
374 sptlrpc_cli_finish_early_reply(early_req);
377 spin_lock(&req->rq_lock);
381 /* Adjust the local timeout for this req */
382 ptlrpc_at_set_req_timeout(req);
384 spin_lock(&req->rq_lock);
385 olddl = req->rq_deadline;
386 /* server assumes it now has rq_timeout from when the request
387 * arrived, so the client should give it at least that long.
388 * since we don't know the arrival time we'll use the original
390 req->rq_deadline = req->rq_sent + req->rq_timeout +
391 ptlrpc_at_get_net_latency(req);
393 DEBUG_REQ(D_ADAPTTO, req,
394 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
395 "("CFS_DURATION_T"s)", req->rq_early_count,
396 cfs_time_sub(req->rq_deadline, cfs_time_current_sec()),
397 cfs_time_sub(req->rq_deadline, olddl));
402 struct kmem_cache *request_cache;
404 int ptlrpc_request_cache_init(void)
406 request_cache = kmem_cache_create("ptlrpc_cache",
407 sizeof(struct ptlrpc_request),
408 0, SLAB_HWCACHE_ALIGN, NULL);
409 return request_cache == NULL ? -ENOMEM : 0;
412 void ptlrpc_request_cache_fini(void)
414 kmem_cache_destroy(request_cache);
417 struct ptlrpc_request *ptlrpc_request_cache_alloc(gfp_t flags)
419 struct ptlrpc_request *req;
421 OBD_SLAB_ALLOC_PTR_GFP(req, request_cache, flags);
425 void ptlrpc_request_cache_free(struct ptlrpc_request *req)
427 OBD_SLAB_FREE_PTR(req, request_cache);
431 * Wind down request pool \a pool.
432 * Frees all requests from the pool too
434 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
436 struct list_head *l, *tmp;
437 struct ptlrpc_request *req;
439 LASSERT(pool != NULL);
441 spin_lock(&pool->prp_lock);
442 list_for_each_safe(l, tmp, &pool->prp_req_list) {
443 req = list_entry(l, struct ptlrpc_request, rq_list);
444 list_del(&req->rq_list);
445 LASSERT(req->rq_reqbuf);
446 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
447 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
448 ptlrpc_request_cache_free(req);
450 spin_unlock(&pool->prp_lock);
451 OBD_FREE(pool, sizeof(*pool));
453 EXPORT_SYMBOL(ptlrpc_free_rq_pool);
456 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
458 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
463 while (size < pool->prp_rq_size)
466 LASSERTF(list_empty(&pool->prp_req_list) ||
467 size == pool->prp_rq_size,
468 "Trying to change pool size with nonempty pool "
469 "from %d to %d bytes\n", pool->prp_rq_size, size);
471 spin_lock(&pool->prp_lock);
472 pool->prp_rq_size = size;
473 for (i = 0; i < num_rq; i++) {
474 struct ptlrpc_request *req;
475 struct lustre_msg *msg;
477 spin_unlock(&pool->prp_lock);
478 req = ptlrpc_request_cache_alloc(GFP_NOFS);
481 OBD_ALLOC_LARGE(msg, size);
483 ptlrpc_request_cache_free(req);
486 req->rq_reqbuf = msg;
487 req->rq_reqbuf_len = size;
489 spin_lock(&pool->prp_lock);
490 list_add_tail(&req->rq_list, &pool->prp_req_list);
492 spin_unlock(&pool->prp_lock);
495 EXPORT_SYMBOL(ptlrpc_add_rqs_to_pool);
498 * Create and initialize new request pool with given attributes:
499 * \a num_rq - initial number of requests to create for the pool
500 * \a msgsize - maximum message size possible for requests in thid pool
501 * \a populate_pool - function to be called when more requests need to be added
503 * Returns pointer to newly created pool or NULL on error.
505 struct ptlrpc_request_pool *
506 ptlrpc_init_rq_pool(int num_rq, int msgsize,
507 void (*populate_pool)(struct ptlrpc_request_pool *, int))
509 struct ptlrpc_request_pool *pool;
511 OBD_ALLOC(pool, sizeof(struct ptlrpc_request_pool));
515 /* Request next power of two for the allocation, because internally
516 kernel would do exactly this */
518 spin_lock_init(&pool->prp_lock);
519 INIT_LIST_HEAD(&pool->prp_req_list);
520 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
521 pool->prp_populate = populate_pool;
523 populate_pool(pool, num_rq);
525 if (list_empty(&pool->prp_req_list)) {
526 /* have not allocated a single request for the pool */
527 OBD_FREE(pool, sizeof(struct ptlrpc_request_pool));
532 EXPORT_SYMBOL(ptlrpc_init_rq_pool);
535 * Fetches one request from pool \a pool
537 static struct ptlrpc_request *
538 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
540 struct ptlrpc_request *request;
541 struct lustre_msg *reqbuf;
546 spin_lock(&pool->prp_lock);
548 /* See if we have anything in a pool, and bail out if nothing,
549 * in writeout path, where this matters, this is safe to do, because
550 * nothing is lost in this case, and when some in-flight requests
551 * complete, this code will be called again. */
552 if (unlikely(list_empty(&pool->prp_req_list))) {
553 spin_unlock(&pool->prp_lock);
557 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
559 list_del_init(&request->rq_list);
560 spin_unlock(&pool->prp_lock);
562 LASSERT(request->rq_reqbuf);
563 LASSERT(request->rq_pool);
565 reqbuf = request->rq_reqbuf;
566 memset(request, 0, sizeof(*request));
567 request->rq_reqbuf = reqbuf;
568 request->rq_reqbuf_len = pool->prp_rq_size;
569 request->rq_pool = pool;
575 * Returns freed \a request to pool.
577 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
579 struct ptlrpc_request_pool *pool = request->rq_pool;
581 spin_lock(&pool->prp_lock);
582 LASSERT(list_empty(&request->rq_list));
583 LASSERT(!request->rq_receiving_reply);
584 list_add_tail(&request->rq_list, &pool->prp_req_list);
585 spin_unlock(&pool->prp_lock);
588 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
589 __u32 version, int opcode,
590 int count, __u32 *lengths, char **bufs,
591 struct ptlrpc_cli_ctx *ctx)
593 struct obd_import *imp = request->rq_import;
598 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
600 rc = sptlrpc_req_get_ctx(request);
605 sptlrpc_req_set_flavor(request, opcode);
607 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
610 LASSERT(!request->rq_pool);
614 lustre_msg_add_version(request->rq_reqmsg, version);
615 request->rq_send_state = LUSTRE_IMP_FULL;
616 request->rq_type = PTL_RPC_MSG_REQUEST;
618 request->rq_req_cbid.cbid_fn = request_out_callback;
619 request->rq_req_cbid.cbid_arg = request;
621 request->rq_reply_cbid.cbid_fn = reply_in_callback;
622 request->rq_reply_cbid.cbid_arg = request;
624 request->rq_reply_deadline = 0;
625 request->rq_phase = RQ_PHASE_NEW;
626 request->rq_next_phase = RQ_PHASE_UNDEFINED;
628 request->rq_request_portal = imp->imp_client->cli_request_portal;
629 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
631 ptlrpc_at_set_req_timeout(request);
633 request->rq_xid = ptlrpc_next_xid();
634 lustre_msg_set_opc(request->rq_reqmsg, opcode);
638 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
640 class_import_put(imp);
644 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
645 __u32 version, int opcode, char **bufs,
646 struct ptlrpc_cli_ctx *ctx)
650 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
651 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
652 request->rq_pill.rc_area[RCL_CLIENT],
655 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
658 * Pack request buffers for network transfer, performing necessary encryption
659 * steps if necessary.
661 int ptlrpc_request_pack(struct ptlrpc_request *request,
662 __u32 version, int opcode)
665 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
669 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
670 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
671 * have to send old ptlrpc_body to keep interoprability with these
674 * Only three kinds of server->client RPCs so far:
679 * XXX This should be removed whenever we drop the interoprability with
680 * the these old clients.
682 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
683 opcode == LDLM_GL_CALLBACK)
684 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
685 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
689 EXPORT_SYMBOL(ptlrpc_request_pack);
692 * Helper function to allocate new request on import \a imp
693 * and possibly using existing request from pool \a pool if provided.
694 * Returns allocated request structure with import field filled or
698 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
699 struct ptlrpc_request_pool *pool)
701 struct ptlrpc_request *request = NULL;
704 request = ptlrpc_prep_req_from_pool(pool);
707 request = ptlrpc_request_cache_alloc(GFP_NOFS);
710 ptlrpc_cli_req_init(request);
712 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
713 LASSERT(imp != LP_POISON);
714 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p\n",
716 LASSERT(imp->imp_client != LP_POISON);
718 request->rq_import = class_import_get(imp);
720 CERROR("request allocation out of memory\n");
727 * Helper function for creating a request.
728 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
729 * buffer structures according to capsule template \a format.
730 * Returns allocated request structure pointer or NULL on error.
732 static struct ptlrpc_request *
733 ptlrpc_request_alloc_internal(struct obd_import *imp,
734 struct ptlrpc_request_pool * pool,
735 const struct req_format *format)
737 struct ptlrpc_request *request;
739 request = __ptlrpc_request_alloc(imp, pool);
743 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
744 req_capsule_set(&request->rq_pill, format);
749 * Allocate new request structure for import \a imp and initialize its
750 * buffer structure according to capsule template \a format.
752 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
753 const struct req_format *format)
755 return ptlrpc_request_alloc_internal(imp, NULL, format);
757 EXPORT_SYMBOL(ptlrpc_request_alloc);
760 * Allocate new request structure for import \a imp from pool \a pool and
761 * initialize its buffer structure according to capsule template \a format.
763 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
764 struct ptlrpc_request_pool * pool,
765 const struct req_format *format)
767 return ptlrpc_request_alloc_internal(imp, pool, format);
769 EXPORT_SYMBOL(ptlrpc_request_alloc_pool);
772 * For requests not from pool, free memory of the request structure.
773 * For requests obtained from a pool earlier, return request back to pool.
775 void ptlrpc_request_free(struct ptlrpc_request *request)
777 if (request->rq_pool)
778 __ptlrpc_free_req_to_pool(request);
780 ptlrpc_request_cache_free(request);
782 EXPORT_SYMBOL(ptlrpc_request_free);
785 * Allocate new request for operatione \a opcode and immediatelly pack it for
787 * Only used for simple requests like OBD_PING where the only important
788 * part of the request is operation itself.
789 * Returns allocated request or NULL on error.
791 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
792 const struct req_format *format,
793 __u32 version, int opcode)
795 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
799 rc = ptlrpc_request_pack(req, version, opcode);
801 ptlrpc_request_free(req);
807 EXPORT_SYMBOL(ptlrpc_request_alloc_pack);
810 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
811 * for operation \a opcode. Request would contain \a count buffers.
812 * Sizes of buffers are described in array \a lengths and buffers themselves
813 * are provided by a pointer \a bufs.
814 * Returns prepared request structure pointer or NULL on error.
816 struct ptlrpc_request *
817 ptlrpc_prep_req_pool(struct obd_import *imp,
818 __u32 version, int opcode,
819 int count, __u32 *lengths, char **bufs,
820 struct ptlrpc_request_pool *pool)
822 struct ptlrpc_request *request;
825 request = __ptlrpc_request_alloc(imp, pool);
829 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
830 lengths, bufs, NULL);
832 ptlrpc_request_free(request);
837 EXPORT_SYMBOL(ptlrpc_prep_req_pool);
840 * Same as ptlrpc_prep_req_pool, but without pool
842 struct ptlrpc_request *
843 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
844 __u32 *lengths, char **bufs)
846 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
849 EXPORT_SYMBOL(ptlrpc_prep_req);
852 * Allocate and initialize new request set structure.
853 * Returns a pointer to the newly allocated set structure or NULL on error.
855 struct ptlrpc_request_set *ptlrpc_prep_set(void)
857 struct ptlrpc_request_set *set;
860 OBD_ALLOC(set, sizeof *set);
863 atomic_set(&set->set_refcount, 1);
864 INIT_LIST_HEAD(&set->set_requests);
865 init_waitqueue_head(&set->set_waitq);
866 atomic_set(&set->set_new_count, 0);
867 atomic_set(&set->set_remaining, 0);
868 spin_lock_init(&set->set_new_req_lock);
869 INIT_LIST_HEAD(&set->set_new_requests);
870 INIT_LIST_HEAD(&set->set_cblist);
871 set->set_max_inflight = UINT_MAX;
872 set->set_producer = NULL;
873 set->set_producer_arg = NULL;
878 EXPORT_SYMBOL(ptlrpc_prep_set);
881 * Allocate and initialize new request set structure with flow control
882 * extension. This extension allows to control the number of requests in-flight
883 * for the whole set. A callback function to generate requests must be provided
884 * and the request set will keep the number of requests sent over the wire to
886 * Returns a pointer to the newly allocated set structure or NULL on error.
888 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
892 struct ptlrpc_request_set *set;
894 set = ptlrpc_prep_set();
898 set->set_max_inflight = max;
899 set->set_producer = func;
900 set->set_producer_arg = arg;
904 EXPORT_SYMBOL(ptlrpc_prep_fcset);
907 * Wind down and free request set structure previously allocated with
909 * Ensures that all requests on the set have completed and removes
910 * all requests from the request list in a set.
911 * If any unsent request happen to be on the list, pretends that they got
912 * an error in flight and calls their completion handler.
914 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
916 struct list_head *tmp;
917 struct list_head *next;
922 /* Requests on the set should either all be completed, or all be new */
923 expected_phase = (atomic_read(&set->set_remaining) == 0) ?
924 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
925 list_for_each(tmp, &set->set_requests) {
926 struct ptlrpc_request *req =
927 list_entry(tmp, struct ptlrpc_request,
930 LASSERT(req->rq_phase == expected_phase);
934 LASSERTF(atomic_read(&set->set_remaining) == 0 ||
935 atomic_read(&set->set_remaining) == n, "%d / %d\n",
936 atomic_read(&set->set_remaining), n);
938 list_for_each_safe(tmp, next, &set->set_requests) {
939 struct ptlrpc_request *req =
940 list_entry(tmp, struct ptlrpc_request,
942 list_del_init(&req->rq_set_chain);
944 LASSERT(req->rq_phase == expected_phase);
946 if (req->rq_phase == RQ_PHASE_NEW) {
947 ptlrpc_req_interpret(NULL, req, -EBADR);
948 atomic_dec(&set->set_remaining);
951 spin_lock(&req->rq_lock);
953 req->rq_invalid_rqset = 0;
954 spin_unlock(&req->rq_lock);
956 ptlrpc_req_finished (req);
959 LASSERT(atomic_read(&set->set_remaining) == 0);
961 ptlrpc_reqset_put(set);
964 EXPORT_SYMBOL(ptlrpc_set_destroy);
967 * Add a callback function \a fn to the set.
968 * This function would be called when all requests on this set are completed.
969 * The function will be passed \a data argument.
971 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
972 set_interpreter_func fn, void *data)
974 struct ptlrpc_set_cbdata *cbdata;
976 OBD_ALLOC_PTR(cbdata);
980 cbdata->psc_interpret = fn;
981 cbdata->psc_data = data;
982 list_add_tail(&cbdata->psc_item, &set->set_cblist);
986 EXPORT_SYMBOL(ptlrpc_set_add_cb);
989 * Add a new request to the general purpose request set.
990 * Assumes request reference from the caller.
992 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
993 struct ptlrpc_request *req)
995 LASSERT(list_empty(&req->rq_set_chain));
997 /* The set takes over the caller's request reference */
998 list_add_tail(&req->rq_set_chain, &set->set_requests);
1000 atomic_inc(&set->set_remaining);
1001 req->rq_queued_time = cfs_time_current();
1003 if (req->rq_reqmsg != NULL)
1004 lustre_msg_set_jobid(req->rq_reqmsg, NULL);
1006 if (set->set_producer != NULL)
1007 /* If the request set has a producer callback, the RPC must be
1008 * sent straight away */
1009 ptlrpc_send_new_req(req);
1011 EXPORT_SYMBOL(ptlrpc_set_add_req);
1014 * Add a request to a request with dedicated server thread
1015 * and wake the thread to make any necessary processing.
1016 * Currently only used for ptlrpcd.
1018 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1019 struct ptlrpc_request *req)
1021 struct ptlrpc_request_set *set = pc->pc_set;
1024 LASSERT(req->rq_set == NULL);
1025 LASSERT(test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1027 spin_lock(&set->set_new_req_lock);
1029 * The set takes over the caller's request reference.
1032 req->rq_queued_time = cfs_time_current();
1033 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1034 count = atomic_inc_return(&set->set_new_count);
1035 spin_unlock(&set->set_new_req_lock);
1037 /* Only need to call wakeup once for the first entry. */
1039 wake_up(&set->set_waitq);
1041 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1042 * guarantee the async RPC can be processed ASAP, we have
1043 * no other better choice. It maybe fixed in future. */
1044 for (i = 0; i < pc->pc_npartners; i++)
1045 wake_up(&pc->pc_partners[i]->pc_set->set_waitq);
1048 EXPORT_SYMBOL(ptlrpc_set_add_new_req);
1051 * Based on the current state of the import, determine if the request
1052 * can be sent, is an error, or should be delayed.
1054 * Returns true if this request should be delayed. If false, and
1055 * *status is set, then the request can not be sent and *status is the
1056 * error code. If false and status is 0, then request can be sent.
1058 * The imp->imp_lock must be held.
1060 static int ptlrpc_import_delay_req(struct obd_import *imp,
1061 struct ptlrpc_request *req, int *status)
1066 LASSERT (status != NULL);
1069 if (req->rq_ctx_init || req->rq_ctx_fini) {
1070 /* always allow ctx init/fini rpc go through */
1071 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1072 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1074 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1075 /* pings may safely race with umount */
1076 DEBUG_REQ(lustre_msg_get_opc(req->rq_reqmsg) == OBD_PING ?
1077 D_HA : D_ERROR, req, "IMP_CLOSED ");
1079 } else if (ptlrpc_send_limit_expired(req)) {
1080 /* probably doesn't need to be a D_ERROR after initial testing*/
1081 DEBUG_REQ(D_HA, req, "send limit expired ");
1082 *status = -ETIMEDOUT;
1083 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1084 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1085 /* allow CONNECT even if import is invalid */ ;
1086 if (atomic_read(&imp->imp_inval_count) != 0) {
1087 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1090 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1091 if (!imp->imp_deactive)
1092 DEBUG_REQ(D_NET, req, "IMP_INVALID");
1093 *status = -ESHUTDOWN; /* bz 12940 */
1094 } else if (req->rq_import_generation != imp->imp_generation) {
1095 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1097 } else if (req->rq_send_state != imp->imp_state) {
1098 /* invalidate in progress - any requests should be drop */
1099 if (atomic_read(&imp->imp_inval_count) != 0) {
1100 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1102 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1103 *status = -EWOULDBLOCK;
1104 } else if (req->rq_allow_replay &&
1105 (imp->imp_state == LUSTRE_IMP_REPLAY ||
1106 imp->imp_state == LUSTRE_IMP_REPLAY_LOCKS ||
1107 imp->imp_state == LUSTRE_IMP_REPLAY_WAIT ||
1108 imp->imp_state == LUSTRE_IMP_RECOVER)) {
1109 DEBUG_REQ(D_HA, req, "allow during recovery.\n");
1119 * Decide if the eror message regarding provided request \a req
1120 * should be printed to the console or not.
1121 * Makes it's decision on request status and other properties.
1122 * Returns 1 to print error on the system console or 0 if not.
1124 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1129 LASSERT(req->rq_reqmsg != NULL);
1130 opc = lustre_msg_get_opc(req->rq_reqmsg);
1132 /* Suppress particular reconnect errors which are to be expected. No
1133 * errors are suppressed for the initial connection on an import */
1134 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1135 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1137 /* Suppress timed out reconnect requests */
1138 if (req->rq_timedout)
1141 /* Suppress unavailable/again reconnect requests */
1142 err = lustre_msg_get_status(req->rq_repmsg);
1143 if (err == -ENODEV || err == -EAGAIN)
1151 * Check request processing status.
1152 * Returns the status.
1154 static int ptlrpc_check_status(struct ptlrpc_request *req)
1159 err = lustre_msg_get_status(req->rq_repmsg);
1160 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1161 struct obd_import *imp = req->rq_import;
1162 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1163 if (ptlrpc_console_allow(req))
1164 LCONSOLE_ERROR_MSG(0x011, "%s: Communicating with %s,"
1165 " operation %s failed with %d.\n",
1166 imp->imp_obd->obd_name,
1168 imp->imp_connection->c_peer.nid),
1169 ll_opcode2str(opc), err);
1170 RETURN(err < 0 ? err : -EINVAL);
1174 DEBUG_REQ(D_INFO, req, "status is %d", err);
1175 } else if (err > 0) {
1176 /* XXX: translate this error from net to host */
1177 DEBUG_REQ(D_INFO, req, "status is %d", err);
1184 * save pre-versions of objects into request for replay.
1185 * Versions are obtained from server reply.
1188 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1190 struct lustre_msg *repmsg = req->rq_repmsg;
1191 struct lustre_msg *reqmsg = req->rq_reqmsg;
1192 __u64 *versions = lustre_msg_get_versions(repmsg);
1195 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1199 lustre_msg_set_versions(reqmsg, versions);
1200 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1201 versions[0], versions[1]);
1207 * Callback function called when client receives RPC reply for \a req.
1208 * Returns 0 on success or error code.
1209 * The return alue would be assigned to req->rq_status by the caller
1210 * as request processing status.
1211 * This function also decides if the request needs to be saved for later replay.
1213 static int after_reply(struct ptlrpc_request *req)
1215 struct obd_import *imp = req->rq_import;
1216 struct obd_device *obd = req->rq_import->imp_obd;
1218 struct timeval work_start;
1222 LASSERT(obd != NULL);
1223 /* repbuf must be unlinked */
1224 LASSERT(!req->rq_receiving_reply && !req->rq_reply_unlink);
1226 if (req->rq_reply_truncate) {
1227 if (ptlrpc_no_resend(req)) {
1228 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1229 " expected: %d, actual size: %d",
1230 req->rq_nob_received, req->rq_repbuf_len);
1234 sptlrpc_cli_free_repbuf(req);
1235 /* Pass the required reply buffer size (include
1236 * space for early reply).
1237 * NB: no need to roundup because alloc_repbuf
1238 * will roundup it */
1239 req->rq_replen = req->rq_nob_received;
1240 req->rq_nob_received = 0;
1241 spin_lock(&req->rq_lock);
1243 spin_unlock(&req->rq_lock);
1248 * NB Until this point, the whole of the incoming message,
1249 * including buflens, status etc is in the sender's byte order.
1251 rc = sptlrpc_cli_unwrap_reply(req);
1253 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1258 * Security layer unwrap might ask resend this request.
1263 rc = unpack_reply(req);
1267 /* retry indefinitely on EINPROGRESS */
1268 if (lustre_msg_get_status(req->rq_repmsg) == -EINPROGRESS &&
1269 ptlrpc_no_resend(req) == 0 && !req->rq_no_retry_einprogress) {
1270 time_t now = cfs_time_current_sec();
1272 DEBUG_REQ(D_RPCTRACE, req, "Resending request on EINPROGRESS");
1273 spin_lock(&req->rq_lock);
1275 spin_unlock(&req->rq_lock);
1276 req->rq_nr_resend++;
1278 /* allocate new xid to avoid reply reconstruction */
1279 if (!req->rq_bulk) {
1280 /* new xid is already allocated for bulk in
1281 * ptlrpc_check_set() */
1282 req->rq_xid = ptlrpc_next_xid();
1283 DEBUG_REQ(D_RPCTRACE, req, "Allocating new xid for "
1284 "resend on EINPROGRESS");
1287 /* Readjust the timeout for current conditions */
1288 ptlrpc_at_set_req_timeout(req);
1289 /* delay resend to give a chance to the server to get ready.
1290 * The delay is increased by 1s on every resend and is capped to
1291 * the current request timeout (i.e. obd_timeout if AT is off,
1292 * or AT service time x 125% + 5s, see at_est2timeout) */
1293 if (req->rq_nr_resend > req->rq_timeout)
1294 req->rq_sent = now + req->rq_timeout;
1296 req->rq_sent = now + req->rq_nr_resend;
1301 do_gettimeofday(&work_start);
1302 timediff = cfs_timeval_sub(&work_start, &req->rq_sent_tv, NULL);
1303 if (obd->obd_svc_stats != NULL) {
1304 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1306 ptlrpc_lprocfs_rpc_sent(req, timediff);
1309 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1310 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1311 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1312 lustre_msg_get_type(req->rq_repmsg));
1316 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1317 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1318 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1319 ptlrpc_at_adj_net_latency(req,
1320 lustre_msg_get_service_time(req->rq_repmsg));
1322 rc = ptlrpc_check_status(req);
1323 imp->imp_connect_error = rc;
1327 * Either we've been evicted, or the server has failed for
1328 * some reason. Try to reconnect, and if that fails, punt to
1331 if (ll_rpc_recoverable_error(rc)) {
1332 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1333 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1336 ptlrpc_request_handle_notconn(req);
1341 * Let's look if server sent slv. Do it only for RPC with
1344 ldlm_cli_update_pool(req);
1348 * Store transno in reqmsg for replay.
1350 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1351 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1352 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1355 if (imp->imp_replayable) {
1356 spin_lock(&imp->imp_lock);
1358 * No point in adding already-committed requests to the replay
1359 * list, we will just remove them immediately. b=9829
1361 if (req->rq_transno != 0 &&
1363 lustre_msg_get_last_committed(req->rq_repmsg) ||
1365 /** version recovery */
1366 ptlrpc_save_versions(req);
1367 ptlrpc_retain_replayable_request(req, imp);
1368 } else if (req->rq_commit_cb != NULL &&
1369 list_empty(&req->rq_replay_list)) {
1370 /* NB: don't call rq_commit_cb if it's already on
1371 * rq_replay_list, ptlrpc_free_committed() will call
1372 * it later, see LU-3618 for details */
1373 spin_unlock(&imp->imp_lock);
1374 req->rq_commit_cb(req);
1375 spin_lock(&imp->imp_lock);
1379 * Replay-enabled imports return commit-status information.
1381 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1382 imp->imp_peer_committed_transno =
1383 lustre_msg_get_last_committed(req->rq_repmsg);
1386 ptlrpc_free_committed(imp);
1388 if (!list_empty(&imp->imp_replay_list)) {
1389 struct ptlrpc_request *last;
1391 last = list_entry(imp->imp_replay_list.prev,
1392 struct ptlrpc_request,
1395 * Requests with rq_replay stay on the list even if no
1396 * commit is expected.
1398 if (last->rq_transno > imp->imp_peer_committed_transno)
1399 ptlrpc_pinger_commit_expected(imp);
1402 spin_unlock(&imp->imp_lock);
1409 * Helper function to send request \a req over the network for the first time
1410 * Also adjusts request phase.
1411 * Returns 0 on success or error code.
1413 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1415 struct obd_import *imp = req->rq_import;
1419 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1420 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1421 (!req->rq_generation_set ||
1422 req->rq_import_generation == imp->imp_generation))
1425 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1427 spin_lock(&imp->imp_lock);
1429 if (!req->rq_generation_set)
1430 req->rq_import_generation = imp->imp_generation;
1432 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1433 spin_lock(&req->rq_lock);
1434 req->rq_waiting = 1;
1435 spin_unlock(&req->rq_lock);
1437 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1438 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1439 ptlrpc_import_state_name(req->rq_send_state),
1440 ptlrpc_import_state_name(imp->imp_state));
1441 LASSERT(list_empty(&req->rq_list));
1442 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1443 atomic_inc(&req->rq_import->imp_inflight);
1444 spin_unlock(&imp->imp_lock);
1449 spin_unlock(&imp->imp_lock);
1450 req->rq_status = rc;
1451 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1455 LASSERT(list_empty(&req->rq_list));
1456 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1457 atomic_inc(&req->rq_import->imp_inflight);
1458 spin_unlock(&imp->imp_lock);
1460 lustre_msg_set_status(req->rq_reqmsg, current_pid());
1462 rc = sptlrpc_req_refresh_ctx(req, -1);
1465 req->rq_status = rc;
1468 spin_lock(&req->rq_lock);
1469 req->rq_wait_ctx = 1;
1470 spin_unlock(&req->rq_lock);
1475 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1476 " %s:%s:%d:"LPU64":%s:%d\n", current_comm(),
1477 imp->imp_obd->obd_uuid.uuid,
1478 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1479 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1480 lustre_msg_get_opc(req->rq_reqmsg));
1482 rc = ptl_send_rpc(req, 0);
1484 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1485 spin_lock(&req->rq_lock);
1486 req->rq_net_err = 1;
1487 spin_unlock(&req->rq_lock);
1493 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1498 LASSERT(set->set_producer != NULL);
1500 remaining = atomic_read(&set->set_remaining);
1502 /* populate the ->set_requests list with requests until we
1503 * reach the maximum number of RPCs in flight for this set */
1504 while (atomic_read(&set->set_remaining) < set->set_max_inflight) {
1505 rc = set->set_producer(set, set->set_producer_arg);
1506 if (rc == -ENOENT) {
1507 /* no more RPC to produce */
1508 set->set_producer = NULL;
1509 set->set_producer_arg = NULL;
1514 RETURN((atomic_read(&set->set_remaining) - remaining));
1518 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1519 * and no more replies are expected.
1520 * (it is possible to get less replies than requests sent e.g. due to timed out
1521 * requests or requests that we had trouble to send out)
1523 * NOTE: This function contains a potential schedule point (cond_resched()).
1525 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1527 struct list_head *tmp, *next;
1528 struct list_head comp_reqs;
1529 int force_timer_recalc = 0;
1532 if (atomic_read(&set->set_remaining) == 0)
1535 INIT_LIST_HEAD(&comp_reqs);
1536 list_for_each_safe(tmp, next, &set->set_requests) {
1537 struct ptlrpc_request *req =
1538 list_entry(tmp, struct ptlrpc_request,
1540 struct obd_import *imp = req->rq_import;
1541 int unregistered = 0;
1544 /* This schedule point is mainly for the ptlrpcd caller of this
1545 * function. Most ptlrpc sets are not long-lived and unbounded
1546 * in length, but at the least the set used by the ptlrpcd is.
1547 * Since the processing time is unbounded, we need to insert an
1548 * explicit schedule point to make the thread well-behaved.
1552 if (req->rq_phase == RQ_PHASE_NEW &&
1553 ptlrpc_send_new_req(req)) {
1554 force_timer_recalc = 1;
1557 /* delayed send - skip */
1558 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1561 /* delayed resend - skip */
1562 if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend &&
1563 req->rq_sent > cfs_time_current_sec())
1566 if (!(req->rq_phase == RQ_PHASE_RPC ||
1567 req->rq_phase == RQ_PHASE_BULK ||
1568 req->rq_phase == RQ_PHASE_INTERPRET ||
1569 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1570 req->rq_phase == RQ_PHASE_COMPLETE)) {
1571 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1575 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1576 LASSERT(req->rq_next_phase != req->rq_phase);
1577 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1580 * Skip processing until reply is unlinked. We
1581 * can't return to pool before that and we can't
1582 * call interpret before that. We need to make
1583 * sure that all rdma transfers finished and will
1584 * not corrupt any data.
1586 if (ptlrpc_client_recv_or_unlink(req) ||
1587 ptlrpc_client_bulk_active(req))
1591 * Turn fail_loc off to prevent it from looping
1594 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1595 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1598 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1599 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1604 * Move to next phase if reply was successfully
1607 ptlrpc_rqphase_move(req, req->rq_next_phase);
1610 if (req->rq_phase == RQ_PHASE_COMPLETE) {
1611 list_move_tail(&req->rq_set_chain, &comp_reqs);
1615 if (req->rq_phase == RQ_PHASE_INTERPRET)
1616 GOTO(interpret, req->rq_status);
1619 * Note that this also will start async reply unlink.
1621 if (req->rq_net_err && !req->rq_timedout) {
1622 ptlrpc_expire_one_request(req, 1);
1625 * Check if we still need to wait for unlink.
1627 if (ptlrpc_client_recv_or_unlink(req) ||
1628 ptlrpc_client_bulk_active(req))
1630 /* If there is no need to resend, fail it now. */
1631 if (req->rq_no_resend) {
1632 if (req->rq_status == 0)
1633 req->rq_status = -EIO;
1634 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1635 GOTO(interpret, req->rq_status);
1642 spin_lock(&req->rq_lock);
1643 req->rq_replied = 0;
1644 spin_unlock(&req->rq_lock);
1645 if (req->rq_status == 0)
1646 req->rq_status = -EIO;
1647 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1648 GOTO(interpret, req->rq_status);
1651 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1652 * so it sets rq_intr regardless of individual rpc
1653 * timeouts. The synchronous IO waiting path sets
1654 * rq_intr irrespective of whether ptlrpcd
1655 * has seen a timeout. Our policy is to only interpret
1656 * interrupted rpcs after they have timed out, so we
1657 * need to enforce that here.
1660 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1661 req->rq_wait_ctx)) {
1662 req->rq_status = -EINTR;
1663 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1664 GOTO(interpret, req->rq_status);
1667 if (req->rq_phase == RQ_PHASE_RPC) {
1668 if (req->rq_timedout || req->rq_resend ||
1669 req->rq_waiting || req->rq_wait_ctx) {
1672 if (!ptlrpc_unregister_reply(req, 1)) {
1673 ptlrpc_unregister_bulk(req, 1);
1677 spin_lock(&imp->imp_lock);
1678 if (ptlrpc_import_delay_req(imp, req, &status)){
1679 /* put on delay list - only if we wait
1680 * recovery finished - before send */
1681 list_del_init(&req->rq_list);
1682 list_add_tail(&req->rq_list,
1685 spin_unlock(&imp->imp_lock);
1690 req->rq_status = status;
1691 ptlrpc_rqphase_move(req,
1692 RQ_PHASE_INTERPRET);
1693 spin_unlock(&imp->imp_lock);
1694 GOTO(interpret, req->rq_status);
1696 if (ptlrpc_no_resend(req) &&
1697 !req->rq_wait_ctx) {
1698 req->rq_status = -ENOTCONN;
1699 ptlrpc_rqphase_move(req,
1700 RQ_PHASE_INTERPRET);
1701 spin_unlock(&imp->imp_lock);
1702 GOTO(interpret, req->rq_status);
1705 list_del_init(&req->rq_list);
1706 list_add_tail(&req->rq_list,
1707 &imp->imp_sending_list);
1709 spin_unlock(&imp->imp_lock);
1711 spin_lock(&req->rq_lock);
1712 req->rq_waiting = 0;
1713 spin_unlock(&req->rq_lock);
1715 if (req->rq_timedout || req->rq_resend) {
1716 /* This is re-sending anyways,
1717 * let's mark req as resend. */
1718 spin_lock(&req->rq_lock);
1720 spin_unlock(&req->rq_lock);
1724 if (!ptlrpc_unregister_bulk(req, 1))
1727 /* ensure previous bulk fails */
1728 old_xid = req->rq_xid;
1729 req->rq_xid = ptlrpc_next_xid();
1730 CDEBUG(D_HA, "resend bulk "
1733 old_xid, req->rq_xid);
1737 * rq_wait_ctx is only touched by ptlrpcd,
1738 * so no lock is needed here.
1740 status = sptlrpc_req_refresh_ctx(req, -1);
1743 req->rq_status = status;
1744 spin_lock(&req->rq_lock);
1745 req->rq_wait_ctx = 0;
1746 spin_unlock(&req->rq_lock);
1747 force_timer_recalc = 1;
1749 spin_lock(&req->rq_lock);
1750 req->rq_wait_ctx = 1;
1751 spin_unlock(&req->rq_lock);
1756 spin_lock(&req->rq_lock);
1757 req->rq_wait_ctx = 0;
1758 spin_unlock(&req->rq_lock);
1761 rc = ptl_send_rpc(req, 0);
1763 DEBUG_REQ(D_HA, req,
1764 "send failed: rc = %d", rc);
1765 force_timer_recalc = 1;
1766 spin_lock(&req->rq_lock);
1767 req->rq_net_err = 1;
1768 spin_unlock(&req->rq_lock);
1771 /* need to reset the timeout */
1772 force_timer_recalc = 1;
1775 spin_lock(&req->rq_lock);
1777 if (ptlrpc_client_early(req)) {
1778 ptlrpc_at_recv_early_reply(req);
1779 spin_unlock(&req->rq_lock);
1783 /* Still waiting for a reply? */
1784 if (ptlrpc_client_recv(req)) {
1785 spin_unlock(&req->rq_lock);
1789 /* Did we actually receive a reply? */
1790 if (!ptlrpc_client_replied(req)) {
1791 spin_unlock(&req->rq_lock);
1795 spin_unlock(&req->rq_lock);
1797 /* unlink from net because we are going to
1798 * swab in-place of reply buffer */
1799 unregistered = ptlrpc_unregister_reply(req, 1);
1803 req->rq_status = after_reply(req);
1807 /* If there is no bulk associated with this request,
1808 * then we're done and should let the interpreter
1809 * process the reply. Similarly if the RPC returned
1810 * an error, and therefore the bulk will never arrive.
1812 if (req->rq_bulk == NULL || req->rq_status < 0) {
1813 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1814 GOTO(interpret, req->rq_status);
1817 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1820 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1821 if (ptlrpc_client_bulk_active(req))
1824 if (req->rq_bulk->bd_failure) {
1825 /* The RPC reply arrived OK, but the bulk screwed
1826 * up! Dead weird since the server told us the RPC
1827 * was good after getting the REPLY for her GET or
1828 * the ACK for her PUT. */
1829 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1830 req->rq_status = -EIO;
1833 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1836 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1838 /* This moves to "unregistering" phase we need to wait for
1840 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1841 /* start async bulk unlink too */
1842 ptlrpc_unregister_bulk(req, 1);
1846 if (!ptlrpc_unregister_bulk(req, 1))
1849 /* When calling interpret receiving already should be
1851 LASSERT(!req->rq_receiving_reply);
1853 ptlrpc_req_interpret(env, req, req->rq_status);
1855 if (ptlrpcd_check_work(req)) {
1856 atomic_dec(&set->set_remaining);
1859 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1861 CDEBUG(req->rq_reqmsg != NULL ? D_RPCTRACE : 0,
1862 "Completed RPC pname:cluuid:pid:xid:nid:"
1863 "opc %s:%s:%d:"LPU64":%s:%d\n",
1864 current_comm(), imp->imp_obd->obd_uuid.uuid,
1865 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1866 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1867 lustre_msg_get_opc(req->rq_reqmsg));
1869 spin_lock(&imp->imp_lock);
1870 /* Request already may be not on sending or delaying list. This
1871 * may happen in the case of marking it erroneous for the case
1872 * ptlrpc_import_delay_req(req, status) find it impossible to
1873 * allow sending this rpc and returns *status != 0. */
1874 if (!list_empty(&req->rq_list)) {
1875 list_del_init(&req->rq_list);
1876 atomic_dec(&imp->imp_inflight);
1878 spin_unlock(&imp->imp_lock);
1880 atomic_dec(&set->set_remaining);
1881 wake_up_all(&imp->imp_recovery_waitq);
1883 if (set->set_producer) {
1884 /* produce a new request if possible */
1885 if (ptlrpc_set_producer(set) > 0)
1886 force_timer_recalc = 1;
1888 /* free the request that has just been completed
1889 * in order not to pollute set->set_requests */
1890 list_del_init(&req->rq_set_chain);
1891 spin_lock(&req->rq_lock);
1893 req->rq_invalid_rqset = 0;
1894 spin_unlock(&req->rq_lock);
1896 /* record rq_status to compute the final status later */
1897 if (req->rq_status != 0)
1898 set->set_rc = req->rq_status;
1899 ptlrpc_req_finished(req);
1901 list_move_tail(&req->rq_set_chain, &comp_reqs);
1905 /* move completed request at the head of list so it's easier for
1906 * caller to find them */
1907 list_splice(&comp_reqs, &set->set_requests);
1909 /* If we hit an error, we want to recover promptly. */
1910 RETURN(atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1912 EXPORT_SYMBOL(ptlrpc_check_set);
1915 * Time out request \a req. is \a async_unlink is set, that means do not wait
1916 * until LNet actually confirms network buffer unlinking.
1917 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1919 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1921 struct obd_import *imp = req->rq_import;
1925 spin_lock(&req->rq_lock);
1926 req->rq_timedout = 1;
1927 spin_unlock(&req->rq_lock);
1929 DEBUG_REQ(D_WARNING, req, "Request sent has %s: [sent "CFS_DURATION_T
1930 "/real "CFS_DURATION_T"]",
1931 req->rq_net_err ? "failed due to network error" :
1932 ((req->rq_real_sent == 0 ||
1933 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1934 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1935 "timed out for sent delay" : "timed out for slow reply"),
1936 req->rq_sent, req->rq_real_sent);
1938 if (imp != NULL && obd_debug_peer_on_timeout)
1939 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1941 ptlrpc_unregister_reply(req, async_unlink);
1942 ptlrpc_unregister_bulk(req, async_unlink);
1944 if (obd_dump_on_timeout)
1945 libcfs_debug_dumplog();
1948 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1952 atomic_inc(&imp->imp_timeouts);
1954 /* The DLM server doesn't want recovery run on its imports. */
1955 if (imp->imp_dlm_fake)
1958 /* If this request is for recovery or other primordial tasks,
1959 * then error it out here. */
1960 if (req->rq_ctx_init || req->rq_ctx_fini ||
1961 req->rq_send_state != LUSTRE_IMP_FULL ||
1962 imp->imp_obd->obd_no_recov) {
1963 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1964 ptlrpc_import_state_name(req->rq_send_state),
1965 ptlrpc_import_state_name(imp->imp_state));
1966 spin_lock(&req->rq_lock);
1967 req->rq_status = -ETIMEDOUT;
1969 spin_unlock(&req->rq_lock);
1973 /* if a request can't be resent we can't wait for an answer after
1975 if (ptlrpc_no_resend(req)) {
1976 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1980 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1986 * Time out all uncompleted requests in request set pointed by \a data
1987 * Callback used when waiting on sets with l_wait_event.
1990 int ptlrpc_expired_set(void *data)
1992 struct ptlrpc_request_set *set = data;
1993 struct list_head *tmp;
1994 time_t now = cfs_time_current_sec();
1997 LASSERT(set != NULL);
2000 * A timeout expired. See which reqs it applies to...
2002 list_for_each(tmp, &set->set_requests) {
2003 struct ptlrpc_request *req =
2004 list_entry(tmp, struct ptlrpc_request,
2007 /* don't expire request waiting for context */
2008 if (req->rq_wait_ctx)
2011 /* Request in-flight? */
2012 if (!((req->rq_phase == RQ_PHASE_RPC &&
2013 !req->rq_waiting && !req->rq_resend) ||
2014 (req->rq_phase == RQ_PHASE_BULK)))
2017 if (req->rq_timedout || /* already dealt with */
2018 req->rq_deadline > now) /* not expired */
2021 /* Deal with this guy. Do it asynchronously to not block
2022 * ptlrpcd thread. */
2023 ptlrpc_expire_one_request(req, 1);
2027 * When waiting for a whole set, we always break out of the
2028 * sleep so we can recalculate the timeout, or enable interrupts
2029 * if everyone's timed out.
2033 EXPORT_SYMBOL(ptlrpc_expired_set);
2036 * Sets rq_intr flag in \a req under spinlock.
2038 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
2040 spin_lock(&req->rq_lock);
2042 spin_unlock(&req->rq_lock);
2044 EXPORT_SYMBOL(ptlrpc_mark_interrupted);
2047 * Interrupts (sets interrupted flag) all uncompleted requests in
2048 * a set \a data. Callback for l_wait_event for interruptible waits.
2050 void ptlrpc_interrupted_set(void *data)
2052 struct ptlrpc_request_set *set = data;
2053 struct list_head *tmp;
2055 LASSERT(set != NULL);
2056 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
2058 list_for_each(tmp, &set->set_requests) {
2059 struct ptlrpc_request *req =
2060 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2062 if (req->rq_phase != RQ_PHASE_RPC &&
2063 req->rq_phase != RQ_PHASE_UNREGISTERING)
2066 ptlrpc_mark_interrupted(req);
2069 EXPORT_SYMBOL(ptlrpc_interrupted_set);
2072 * Get the smallest timeout in the set; this does NOT set a timeout.
2074 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2076 struct list_head *tmp;
2077 time_t now = cfs_time_current_sec();
2079 struct ptlrpc_request *req;
2083 list_for_each(tmp, &set->set_requests) {
2084 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2087 * Request in-flight?
2089 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2090 (req->rq_phase == RQ_PHASE_BULK) ||
2091 (req->rq_phase == RQ_PHASE_NEW)))
2095 * Already timed out.
2097 if (req->rq_timedout)
2103 if (req->rq_wait_ctx)
2106 if (req->rq_phase == RQ_PHASE_NEW)
2107 deadline = req->rq_sent;
2108 else if (req->rq_phase == RQ_PHASE_RPC && req->rq_resend)
2109 deadline = req->rq_sent;
2111 deadline = req->rq_sent + req->rq_timeout;
2113 if (deadline <= now) /* actually expired already */
2114 timeout = 1; /* ASAP */
2115 else if (timeout == 0 || timeout > deadline - now)
2116 timeout = deadline - now;
2120 EXPORT_SYMBOL(ptlrpc_set_next_timeout);
2123 * Send all unset request from the set and then wait untill all
2124 * requests in the set complete (either get a reply, timeout, get an
2125 * error or otherwise be interrupted).
2126 * Returns 0 on success or error code otherwise.
2128 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2130 struct list_head *tmp;
2131 struct ptlrpc_request *req;
2132 struct l_wait_info lwi;
2136 if (set->set_producer)
2137 (void)ptlrpc_set_producer(set);
2139 list_for_each(tmp, &set->set_requests) {
2140 req = list_entry(tmp, struct ptlrpc_request,
2142 if (req->rq_phase == RQ_PHASE_NEW)
2143 (void)ptlrpc_send_new_req(req);
2146 if (list_empty(&set->set_requests))
2150 timeout = ptlrpc_set_next_timeout(set);
2152 /* wait until all complete, interrupted, or an in-flight
2154 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2157 if (timeout == 0 && !cfs_signal_pending())
2159 * No requests are in-flight (ether timed out
2160 * or delayed), so we can allow interrupts.
2161 * We still want to block for a limited time,
2162 * so we allow interrupts during the timeout.
2164 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2166 ptlrpc_interrupted_set, set);
2169 * At least one request is in flight, so no
2170 * interrupts are allowed. Wait until all
2171 * complete, or an in-flight req times out.
2173 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2174 ptlrpc_expired_set, set);
2176 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2178 /* LU-769 - if we ignored the signal because it was already
2179 * pending when we started, we need to handle it now or we risk
2180 * it being ignored forever */
2181 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2182 cfs_signal_pending()) {
2183 sigset_t blocked_sigs =
2184 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2186 /* In fact we only interrupt for the "fatal" signals
2187 * like SIGINT or SIGKILL. We still ignore less
2188 * important signals since ptlrpc set is not easily
2189 * reentrant from userspace again */
2190 if (cfs_signal_pending())
2191 ptlrpc_interrupted_set(set);
2192 cfs_restore_sigs(blocked_sigs);
2195 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2197 /* -EINTR => all requests have been flagged rq_intr so next
2199 * -ETIMEDOUT => someone timed out. When all reqs have
2200 * timed out, signals are enabled allowing completion with
2202 * I don't really care if we go once more round the loop in
2203 * the error cases -eeb. */
2204 if (rc == 0 && atomic_read(&set->set_remaining) == 0) {
2205 list_for_each(tmp, &set->set_requests) {
2206 req = list_entry(tmp, struct ptlrpc_request,
2208 spin_lock(&req->rq_lock);
2209 req->rq_invalid_rqset = 1;
2210 spin_unlock(&req->rq_lock);
2213 } while (rc != 0 || atomic_read(&set->set_remaining) != 0);
2215 LASSERT(atomic_read(&set->set_remaining) == 0);
2217 rc = set->set_rc; /* rq_status of already freed requests if any */
2218 list_for_each(tmp, &set->set_requests) {
2219 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2221 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2222 if (req->rq_status != 0)
2223 rc = req->rq_status;
2226 if (set->set_interpret != NULL) {
2227 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2229 rc = interpreter (set, set->set_arg, rc);
2231 struct ptlrpc_set_cbdata *cbdata, *n;
2234 list_for_each_entry_safe(cbdata, n,
2235 &set->set_cblist, psc_item) {
2236 list_del_init(&cbdata->psc_item);
2237 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2240 OBD_FREE_PTR(cbdata);
2246 EXPORT_SYMBOL(ptlrpc_set_wait);
2249 * Helper fuction for request freeing.
2250 * Called when request count reached zero and request needs to be freed.
2251 * Removes request from all sorts of sending/replay lists it might be on,
2252 * frees network buffers if any are present.
2253 * If \a locked is set, that means caller is already holding import imp_lock
2254 * and so we no longer need to reobtain it (for certain lists manipulations)
2256 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2260 if (request == NULL)
2263 LASSERT(!request->rq_srv_req);
2264 LASSERT(request->rq_export == NULL);
2265 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2266 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
2267 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
2268 LASSERTF(!request->rq_replay, "req %p\n", request);
2270 req_capsule_fini(&request->rq_pill);
2272 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2273 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2274 if (request->rq_import != NULL) {
2276 spin_lock(&request->rq_import->imp_lock);
2277 list_del_init(&request->rq_replay_list);
2279 spin_unlock(&request->rq_import->imp_lock);
2281 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
2283 if (atomic_read(&request->rq_refcount) != 0) {
2284 DEBUG_REQ(D_ERROR, request,
2285 "freeing request with nonzero refcount");
2289 if (request->rq_repbuf != NULL)
2290 sptlrpc_cli_free_repbuf(request);
2292 if (request->rq_import != NULL) {
2293 class_import_put(request->rq_import);
2294 request->rq_import = NULL;
2296 if (request->rq_bulk != NULL)
2297 ptlrpc_free_bulk_pin(request->rq_bulk);
2299 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2300 sptlrpc_cli_free_reqbuf(request);
2302 if (request->rq_cli_ctx)
2303 sptlrpc_req_put_ctx(request, !locked);
2305 if (request->rq_pool)
2306 __ptlrpc_free_req_to_pool(request);
2308 ptlrpc_request_cache_free(request);
2312 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2314 * Drop one request reference. Must be called with import imp_lock held.
2315 * When reference count drops to zero, reuqest is freed.
2317 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2319 assert_spin_locked(&request->rq_import->imp_lock);
2320 (void)__ptlrpc_req_finished(request, 1);
2322 EXPORT_SYMBOL(ptlrpc_req_finished_with_imp_lock);
2326 * Drops one reference count for request \a request.
2327 * \a locked set indicates that caller holds import imp_lock.
2328 * Frees the request whe reference count reaches zero.
2330 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2333 if (request == NULL)
2336 if (request == LP_POISON ||
2337 request->rq_reqmsg == LP_POISON) {
2338 CERROR("dereferencing freed request (bug 575)\n");
2343 DEBUG_REQ(D_INFO, request, "refcount now %u",
2344 atomic_read(&request->rq_refcount) - 1);
2346 if (atomic_dec_and_test(&request->rq_refcount)) {
2347 __ptlrpc_free_req(request, locked);
2355 * Drops one reference count for a request.
2357 void ptlrpc_req_finished(struct ptlrpc_request *request)
2359 __ptlrpc_req_finished(request, 0);
2361 EXPORT_SYMBOL(ptlrpc_req_finished);
2364 * Returns xid of a \a request
2366 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2368 return request->rq_xid;
2370 EXPORT_SYMBOL(ptlrpc_req_xid);
2373 * Disengage the client's reply buffer from the network
2374 * NB does _NOT_ unregister any client-side bulk.
2375 * IDEMPOTENT, but _not_ safe against concurrent callers.
2376 * The request owner (i.e. the thread doing the I/O) must call...
2377 * Returns 0 on success or 1 if unregistering cannot be made.
2379 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2382 struct l_wait_info lwi;
2387 LASSERT(!in_interrupt());
2390 * Let's setup deadline for reply unlink.
2392 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2393 async && request->rq_reply_deadline == 0)
2394 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2397 * Nothing left to do.
2399 if (!ptlrpc_client_recv_or_unlink(request))
2402 LNetMDUnlink(request->rq_reply_md_h);
2405 * Let's check it once again.
2407 if (!ptlrpc_client_recv_or_unlink(request))
2411 * Move to "Unregistering" phase as reply was not unlinked yet.
2413 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2416 * Do not wait for unlink to finish.
2422 * We have to l_wait_event() whatever the result, to give liblustre
2423 * a chance to run reply_in_callback(), and to make sure we've
2424 * unlinked before returning a req to the pool.
2427 /* The wq argument is ignored by user-space wait_event macros */
2428 wait_queue_head_t *wq = (request->rq_set != NULL) ?
2429 &request->rq_set->set_waitq :
2430 &request->rq_reply_waitq;
2431 /* Network access will complete in finite time but the HUGE
2432 * timeout lets us CWARN for visibility of sluggish NALs */
2433 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2434 cfs_time_seconds(1), NULL, NULL);
2435 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2438 ptlrpc_rqphase_move(request, request->rq_next_phase);
2442 LASSERT(rc == -ETIMEDOUT);
2443 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2444 "rvcng=%d unlnk=%d/%d", request->rq_receiving_reply,
2445 request->rq_req_unlink, request->rq_reply_unlink);
2449 EXPORT_SYMBOL(ptlrpc_unregister_reply);
2451 static void ptlrpc_free_request(struct ptlrpc_request *req)
2453 spin_lock(&req->rq_lock);
2455 spin_unlock(&req->rq_lock);
2457 if (req->rq_commit_cb != NULL)
2458 req->rq_commit_cb(req);
2459 list_del_init(&req->rq_replay_list);
2461 __ptlrpc_req_finished(req, 1);
2465 * the request is committed and dropped from the replay list of its import
2467 void ptlrpc_request_committed(struct ptlrpc_request *req, int force)
2469 struct obd_import *imp = req->rq_import;
2471 spin_lock(&imp->imp_lock);
2472 if (list_empty(&req->rq_replay_list)) {
2473 spin_unlock(&imp->imp_lock);
2477 if (force || req->rq_transno <= imp->imp_peer_committed_transno)
2478 ptlrpc_free_request(req);
2480 spin_unlock(&imp->imp_lock);
2482 EXPORT_SYMBOL(ptlrpc_request_committed);
2485 * Iterates through replay_list on import and prunes
2486 * all requests have transno smaller than last_committed for the
2487 * import and don't have rq_replay set.
2488 * Since requests are sorted in transno order, stops when meetign first
2489 * transno bigger than last_committed.
2490 * caller must hold imp->imp_lock
2492 void ptlrpc_free_committed(struct obd_import *imp)
2494 struct ptlrpc_request *req, *saved;
2495 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2496 bool skip_committed_list = true;
2499 LASSERT(imp != NULL);
2500 assert_spin_locked(&imp->imp_lock);
2502 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2503 imp->imp_generation == imp->imp_last_generation_checked) {
2504 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2505 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2508 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2509 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2510 imp->imp_generation);
2512 if (imp->imp_generation != imp->imp_last_generation_checked)
2513 skip_committed_list = false;
2515 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2516 imp->imp_last_generation_checked = imp->imp_generation;
2518 list_for_each_entry_safe(req, saved, &imp->imp_replay_list,
2520 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2521 LASSERT(req != last_req);
2524 if (req->rq_transno == 0) {
2525 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2528 if (req->rq_import_generation < imp->imp_generation) {
2529 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2533 /* not yet committed */
2534 if (req->rq_transno > imp->imp_peer_committed_transno) {
2535 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2539 if (req->rq_replay) {
2540 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2541 list_move_tail(&req->rq_replay_list,
2542 &imp->imp_committed_list);
2546 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2547 imp->imp_peer_committed_transno);
2549 ptlrpc_free_request(req);
2552 if (skip_committed_list)
2555 list_for_each_entry_safe(req, saved, &imp->imp_committed_list,
2557 LASSERT(req->rq_transno != 0);
2558 if (req->rq_import_generation < imp->imp_generation) {
2559 DEBUG_REQ(D_RPCTRACE, req, "free stale open request");
2560 ptlrpc_free_request(req);
2567 void ptlrpc_cleanup_client(struct obd_import *imp)
2572 EXPORT_SYMBOL(ptlrpc_cleanup_client);
2575 * Schedule previously sent request for resend.
2576 * For bulk requests we assign new xid (to avoid problems with
2577 * lost replies and therefore several transfers landing into same buffer
2578 * from different sending attempts).
2580 void ptlrpc_resend_req(struct ptlrpc_request *req)
2582 DEBUG_REQ(D_HA, req, "going to resend");
2583 spin_lock(&req->rq_lock);
2585 /* Request got reply but linked to the import list still.
2586 Let ptlrpc_check_set() to process it. */
2587 if (ptlrpc_client_replied(req)) {
2588 spin_unlock(&req->rq_lock);
2589 DEBUG_REQ(D_HA, req, "it has reply, so skip it");
2593 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2594 req->rq_status = -EAGAIN;
2597 req->rq_net_err = 0;
2598 req->rq_timedout = 0;
2600 __u64 old_xid = req->rq_xid;
2602 /* ensure previous bulk fails */
2603 req->rq_xid = ptlrpc_next_xid();
2604 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2605 old_xid, req->rq_xid);
2607 ptlrpc_client_wake_req(req);
2608 spin_unlock(&req->rq_lock);
2610 EXPORT_SYMBOL(ptlrpc_resend_req);
2612 /* XXX: this function and rq_status are currently unused */
2613 void ptlrpc_restart_req(struct ptlrpc_request *req)
2615 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2616 req->rq_status = -ERESTARTSYS;
2618 spin_lock(&req->rq_lock);
2619 req->rq_restart = 1;
2620 req->rq_timedout = 0;
2621 ptlrpc_client_wake_req(req);
2622 spin_unlock(&req->rq_lock);
2624 EXPORT_SYMBOL(ptlrpc_restart_req);
2627 * Grab additional reference on a request \a req
2629 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2632 atomic_inc(&req->rq_refcount);
2635 EXPORT_SYMBOL(ptlrpc_request_addref);
2638 * Add a request to import replay_list.
2639 * Must be called under imp_lock
2641 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2642 struct obd_import *imp)
2644 struct list_head *tmp;
2646 assert_spin_locked(&imp->imp_lock);
2648 if (req->rq_transno == 0) {
2649 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2653 /* clear this for new requests that were resent as well
2654 as resent replayed requests. */
2655 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2657 /* don't re-add requests that have been replayed */
2658 if (!list_empty(&req->rq_replay_list))
2661 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2663 LASSERT(imp->imp_replayable);
2664 /* Balanced in ptlrpc_free_committed, usually. */
2665 ptlrpc_request_addref(req);
2666 list_for_each_prev(tmp, &imp->imp_replay_list) {
2667 struct ptlrpc_request *iter = list_entry(tmp,
2668 struct ptlrpc_request,
2671 /* We may have duplicate transnos if we create and then
2672 * open a file, or for closes retained if to match creating
2673 * opens, so use req->rq_xid as a secondary key.
2674 * (See bugs 684, 685, and 428.)
2675 * XXX no longer needed, but all opens need transnos!
2677 if (iter->rq_transno > req->rq_transno)
2680 if (iter->rq_transno == req->rq_transno) {
2681 LASSERT(iter->rq_xid != req->rq_xid);
2682 if (iter->rq_xid > req->rq_xid)
2686 list_add(&req->rq_replay_list, &iter->rq_replay_list);
2690 list_add(&req->rq_replay_list, &imp->imp_replay_list);
2692 EXPORT_SYMBOL(ptlrpc_retain_replayable_request);
2695 * Send request and wait until it completes.
2696 * Returns request processing status.
2698 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2700 struct ptlrpc_request_set *set;
2704 LASSERT(req->rq_set == NULL);
2705 LASSERT(!req->rq_receiving_reply);
2707 set = ptlrpc_prep_set();
2709 CERROR("cannot allocate ptlrpc set: rc = %d\n", -ENOMEM);
2713 /* for distributed debugging */
2714 lustre_msg_set_status(req->rq_reqmsg, current_pid());
2716 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2717 ptlrpc_request_addref(req);
2718 ptlrpc_set_add_req(set, req);
2719 rc = ptlrpc_set_wait(set);
2720 ptlrpc_set_destroy(set);
2724 EXPORT_SYMBOL(ptlrpc_queue_wait);
2727 * Callback used for replayed requests reply processing.
2728 * In case of succesful reply calls registeresd request replay callback.
2729 * In case of error restart replay process.
2731 static int ptlrpc_replay_interpret(const struct lu_env *env,
2732 struct ptlrpc_request *req,
2733 void * data, int rc)
2735 struct ptlrpc_replay_async_args *aa = data;
2736 struct obd_import *imp = req->rq_import;
2739 atomic_dec(&imp->imp_replay_inflight);
2741 if (!ptlrpc_client_replied(req)) {
2742 CERROR("request replay timed out, restarting recovery\n");
2743 GOTO(out, rc = -ETIMEDOUT);
2746 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2747 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2748 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2749 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2751 /** VBR: check version failure */
2752 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2753 /** replay was failed due to version mismatch */
2754 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2755 spin_lock(&imp->imp_lock);
2756 imp->imp_vbr_failed = 1;
2757 imp->imp_no_lock_replay = 1;
2758 spin_unlock(&imp->imp_lock);
2759 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2761 /** The transno had better not change over replay. */
2762 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2763 lustre_msg_get_transno(req->rq_repmsg) ||
2764 lustre_msg_get_transno(req->rq_repmsg) == 0,
2766 lustre_msg_get_transno(req->rq_reqmsg),
2767 lustre_msg_get_transno(req->rq_repmsg));
2770 spin_lock(&imp->imp_lock);
2771 /** if replays by version then gap occur on server, no trust to locks */
2772 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2773 imp->imp_no_lock_replay = 1;
2774 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2775 spin_unlock(&imp->imp_lock);
2776 LASSERT(imp->imp_last_replay_transno);
2778 /* transaction number shouldn't be bigger than the latest replayed */
2779 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2780 DEBUG_REQ(D_ERROR, req,
2781 "Reported transno "LPU64" is bigger than the "
2782 "replayed one: "LPU64, req->rq_transno,
2783 lustre_msg_get_transno(req->rq_reqmsg));
2784 GOTO(out, rc = -EINVAL);
2787 DEBUG_REQ(D_HA, req, "got rep");
2789 /* let the callback do fixups, possibly including in the request */
2790 if (req->rq_replay_cb)
2791 req->rq_replay_cb(req);
2793 if (ptlrpc_client_replied(req) &&
2794 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2795 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2796 lustre_msg_get_status(req->rq_repmsg),
2797 aa->praa_old_status);
2799 /* Put it back for re-replay. */
2800 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2804 * Errors while replay can set transno to 0, but
2805 * imp_last_replay_transno shouldn't be set to 0 anyway
2807 if (req->rq_transno == 0)
2808 CERROR("Transno is 0 during replay!\n");
2810 /* continue with recovery */
2811 rc = ptlrpc_import_recovery_state_machine(imp);
2813 req->rq_send_state = aa->praa_old_state;
2816 /* this replay failed, so restart recovery */
2817 ptlrpc_connect_import(imp);
2823 * Prepares and queues request for replay.
2824 * Adds it to ptlrpcd queue for actual sending.
2825 * Returns 0 on success.
2827 int ptlrpc_replay_req(struct ptlrpc_request *req)
2829 struct ptlrpc_replay_async_args *aa;
2832 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2834 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2835 aa = ptlrpc_req_async_args(req);
2836 memset(aa, 0, sizeof *aa);
2838 /* Prepare request to be resent with ptlrpcd */
2839 aa->praa_old_state = req->rq_send_state;
2840 req->rq_send_state = LUSTRE_IMP_REPLAY;
2841 req->rq_phase = RQ_PHASE_NEW;
2842 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2844 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2846 req->rq_interpret_reply = ptlrpc_replay_interpret;
2847 /* Readjust the timeout for current conditions */
2848 ptlrpc_at_set_req_timeout(req);
2850 /* Tell server the net_latency, so the server can calculate how long
2851 * it should wait for next replay */
2852 lustre_msg_set_service_time(req->rq_reqmsg,
2853 ptlrpc_at_get_net_latency(req));
2854 DEBUG_REQ(D_HA, req, "REPLAY");
2856 atomic_inc(&req->rq_import->imp_replay_inflight);
2857 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2859 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2862 EXPORT_SYMBOL(ptlrpc_replay_req);
2865 * Aborts all in-flight request on import \a imp sending and delayed lists
2867 void ptlrpc_abort_inflight(struct obd_import *imp)
2869 struct list_head *tmp, *n;
2872 /* Make sure that no new requests get processed for this import.
2873 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2874 * this flag and then putting requests on sending_list or delayed_list.
2876 spin_lock(&imp->imp_lock);
2878 /* XXX locking? Maybe we should remove each request with the list
2879 * locked? Also, how do we know if the requests on the list are
2880 * being freed at this time?
2882 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2883 struct ptlrpc_request *req = list_entry(tmp,
2884 struct ptlrpc_request,
2887 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2889 spin_lock(&req->rq_lock);
2890 if (req->rq_import_generation < imp->imp_generation) {
2892 req->rq_status = -EIO;
2893 ptlrpc_client_wake_req(req);
2895 spin_unlock(&req->rq_lock);
2898 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2899 struct ptlrpc_request *req =
2900 list_entry(tmp, struct ptlrpc_request, rq_list);
2902 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2904 spin_lock(&req->rq_lock);
2905 if (req->rq_import_generation < imp->imp_generation) {
2907 req->rq_status = -EIO;
2908 ptlrpc_client_wake_req(req);
2910 spin_unlock(&req->rq_lock);
2913 /* Last chance to free reqs left on the replay list, but we
2914 * will still leak reqs that haven't committed. */
2915 if (imp->imp_replayable)
2916 ptlrpc_free_committed(imp);
2918 spin_unlock(&imp->imp_lock);
2922 EXPORT_SYMBOL(ptlrpc_abort_inflight);
2925 * Abort all uncompleted requests in request set \a set
2927 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2929 struct list_head *tmp, *pos;
2931 LASSERT(set != NULL);
2933 list_for_each_safe(pos, tmp, &set->set_requests) {
2934 struct ptlrpc_request *req =
2935 list_entry(pos, struct ptlrpc_request,
2938 spin_lock(&req->rq_lock);
2939 if (req->rq_phase != RQ_PHASE_RPC) {
2940 spin_unlock(&req->rq_lock);
2945 req->rq_status = -EINTR;
2946 ptlrpc_client_wake_req(req);
2947 spin_unlock(&req->rq_lock);
2951 static __u64 ptlrpc_last_xid;
2952 static spinlock_t ptlrpc_last_xid_lock;
2955 * Initialize the XID for the node. This is common among all requests on
2956 * this node, and only requires the property that it is monotonically
2957 * increasing. It does not need to be sequential. Since this is also used
2958 * as the RDMA match bits, it is important that a single client NOT have
2959 * the same match bits for two different in-flight requests, hence we do
2960 * NOT want to have an XID per target or similar.
2962 * To avoid an unlikely collision between match bits after a client reboot
2963 * (which would deliver old data into the wrong RDMA buffer) initialize
2964 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2965 * If the time is clearly incorrect, we instead use a 62-bit random number.
2966 * In the worst case the random number will overflow 1M RPCs per second in
2967 * 9133 years, or permutations thereof.
2969 #define YEAR_2004 (1ULL << 30)
2970 void ptlrpc_init_xid(void)
2972 time_t now = cfs_time_current_sec();
2974 spin_lock_init(&ptlrpc_last_xid_lock);
2975 if (now < YEAR_2004) {
2976 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2977 ptlrpc_last_xid >>= 2;
2978 ptlrpc_last_xid |= (1ULL << 61);
2980 ptlrpc_last_xid = (__u64)now << 20;
2983 /* Need to always be aligned to a power-of-two for mutli-bulk BRW */
2984 CLASSERT((PTLRPC_BULK_OPS_COUNT & (PTLRPC_BULK_OPS_COUNT - 1)) == 0);
2985 ptlrpc_last_xid &= PTLRPC_BULK_OPS_MASK;
2989 * Increase xid and returns resulting new value to the caller.
2991 * Multi-bulk BRW RPCs consume multiple XIDs for each bulk transfer, starting
2992 * at the returned xid, up to xid + PTLRPC_BULK_OPS_COUNT - 1. The BRW RPC
2993 * itself uses the last bulk xid needed, so the server can determine the
2994 * the number of bulk transfers from the RPC XID and a bitmask. The starting
2995 * xid must align to a power-of-two value.
2997 * This is assumed to be true due to the initial ptlrpc_last_xid
2998 * value also being initialized to a power-of-two value. LU-1431
3000 __u64 ptlrpc_next_xid(void)
3004 spin_lock(&ptlrpc_last_xid_lock);
3005 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3006 ptlrpc_last_xid = next;
3007 spin_unlock(&ptlrpc_last_xid_lock);
3011 EXPORT_SYMBOL(ptlrpc_next_xid);
3014 * Get a glimpse at what next xid value might have been.
3015 * Returns possible next xid.
3017 __u64 ptlrpc_sample_next_xid(void)
3019 #if BITS_PER_LONG == 32
3020 /* need to avoid possible word tearing on 32-bit systems */
3023 spin_lock(&ptlrpc_last_xid_lock);
3024 next = ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3025 spin_unlock(&ptlrpc_last_xid_lock);
3029 /* No need to lock, since returned value is racy anyways */
3030 return ptlrpc_last_xid + PTLRPC_BULK_OPS_COUNT;
3033 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
3036 * Functions for operating ptlrpc workers.
3038 * A ptlrpc work is a function which will be running inside ptlrpc context.
3039 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
3041 * 1. after a work is created, it can be used many times, that is:
3042 * handler = ptlrpcd_alloc_work();
3043 * ptlrpcd_queue_work();
3045 * queue it again when necessary:
3046 * ptlrpcd_queue_work();
3047 * ptlrpcd_destroy_work();
3048 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
3049 * it will only be queued once in any time. Also as its name implies, it may
3050 * have delay before it really runs by ptlrpcd thread.
3052 struct ptlrpc_work_async_args {
3053 int (*cb)(const struct lu_env *, void *);
3057 static void ptlrpcd_add_work_req(struct ptlrpc_request *req)
3059 /* re-initialize the req */
3060 req->rq_timeout = obd_timeout;
3061 req->rq_sent = cfs_time_current_sec();
3062 req->rq_deadline = req->rq_sent + req->rq_timeout;
3063 req->rq_reply_deadline = req->rq_deadline;
3064 req->rq_phase = RQ_PHASE_INTERPRET;
3065 req->rq_next_phase = RQ_PHASE_COMPLETE;
3066 req->rq_xid = ptlrpc_next_xid();
3067 req->rq_import_generation = req->rq_import->imp_generation;
3069 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3072 static int work_interpreter(const struct lu_env *env,
3073 struct ptlrpc_request *req, void *data, int rc)
3075 struct ptlrpc_work_async_args *arg = data;
3077 LASSERT(ptlrpcd_check_work(req));
3078 LASSERT(arg->cb != NULL);
3080 rc = arg->cb(env, arg->cbdata);
3082 list_del_init(&req->rq_set_chain);
3085 if (atomic_dec_return(&req->rq_refcount) > 1) {
3086 atomic_set(&req->rq_refcount, 2);
3087 ptlrpcd_add_work_req(req);
3092 static int worker_format;
3094 static int ptlrpcd_check_work(struct ptlrpc_request *req)
3096 return req->rq_pill.rc_fmt == (void *)&worker_format;
3100 * Create a work for ptlrpc.
3102 void *ptlrpcd_alloc_work(struct obd_import *imp,
3103 int (*cb)(const struct lu_env *, void *), void *cbdata)
3105 struct ptlrpc_request *req = NULL;
3106 struct ptlrpc_work_async_args *args;
3112 RETURN(ERR_PTR(-EINVAL));
3114 /* copy some code from deprecated fakereq. */
3115 req = ptlrpc_request_cache_alloc(GFP_NOFS);
3117 CERROR("ptlrpc: run out of memory!\n");
3118 RETURN(ERR_PTR(-ENOMEM));
3121 ptlrpc_cli_req_init(req);
3123 req->rq_send_state = LUSTRE_IMP_FULL;
3124 req->rq_type = PTL_RPC_MSG_REQUEST;
3125 req->rq_import = class_import_get(imp);
3126 req->rq_interpret_reply = work_interpreter;
3127 /* don't want reply */
3128 req->rq_receiving_reply = 0;
3129 req->rq_req_unlink = req->rq_reply_unlink = 0;
3130 req->rq_no_delay = req->rq_no_resend = 1;
3131 req->rq_pill.rc_fmt = (void *)&worker_format;
3133 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
3134 args = ptlrpc_req_async_args(req);
3136 args->cbdata = cbdata;
3140 EXPORT_SYMBOL(ptlrpcd_alloc_work);
3142 void ptlrpcd_destroy_work(void *handler)
3144 struct ptlrpc_request *req = handler;
3147 ptlrpc_req_finished(req);
3149 EXPORT_SYMBOL(ptlrpcd_destroy_work);
3151 int ptlrpcd_queue_work(void *handler)
3153 struct ptlrpc_request *req = handler;
3156 * Check if the req is already being queued.
3158 * Here comes a trick: it lacks a way of checking if a req is being
3159 * processed reliably in ptlrpc. Here I have to use refcount of req
3160 * for this purpose. This is okay because the caller should use this
3161 * req as opaque data. - Jinshan
3163 LASSERT(atomic_read(&req->rq_refcount) > 0);
3164 if (atomic_inc_return(&req->rq_refcount) == 2)
3165 ptlrpcd_add_work_req(req);
3168 EXPORT_SYMBOL(ptlrpcd_queue_work);