4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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;
69 * Return PortalRPC connection for remore uud \a uuid
71 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
73 struct ptlrpc_connection *c;
75 lnet_process_id_t peer;
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);
96 * Allocate and initialize new bulk descriptor
97 * Returns pointer to the descriptor or NULL on error.
99 struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
101 struct ptlrpc_bulk_desc *desc;
103 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
107 cfs_spin_lock_init(&desc->bd_lock);
108 cfs_waitq_init(&desc->bd_waitq);
109 desc->bd_max_iov = npages;
110 desc->bd_iov_count = 0;
111 LNetInvalidateHandle(&desc->bd_md_h);
112 desc->bd_portal = portal;
113 desc->bd_type = type;
119 * Prepare bulk descriptor for specified outgoing request \a req that
120 * can fit \a npages * pages. \a type is bulk type. \a portal is where
121 * the bulk to be sent. Used on client-side.
122 * Returns pointer to newly allocatrd initialized bulk descriptor or NULL on
125 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp(struct ptlrpc_request *req,
126 int npages, int type, int portal)
128 struct obd_import *imp = req->rq_import;
129 struct ptlrpc_bulk_desc *desc;
132 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
133 desc = new_bulk(npages, type, portal);
137 desc->bd_import_generation = req->rq_import_generation;
138 desc->bd_import = class_import_get(imp);
141 desc->bd_cbid.cbid_fn = client_bulk_callback;
142 desc->bd_cbid.cbid_arg = desc;
144 /* This makes req own desc, and free it when she frees herself */
151 * Add a page \a page to the bulk descriptor \a desc.
152 * Data to transfer in the page starts at offset \a pageoffset and
153 * amount of data to transfer from the page is \a len
155 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
156 cfs_page_t *page, int pageoffset, int len)
158 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
159 LASSERT(page != NULL);
160 LASSERT(pageoffset >= 0);
162 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
167 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
171 * Uninitialize and free bulk descriptor \a desc.
172 * Works on bulk descriptors both from server and client side.
174 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
179 LASSERT(desc != NULL);
180 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
181 LASSERT(!desc->bd_network_rw); /* network hands off or */
182 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
184 sptlrpc_enc_pool_put_pages(desc);
187 class_export_put(desc->bd_export);
189 class_import_put(desc->bd_import);
191 for (i = 0; i < desc->bd_iov_count ; i++)
192 cfs_page_unpin(desc->bd_iov[i].kiov_page);
194 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
195 bd_iov[desc->bd_max_iov]));
200 * Set server timelimit for this req, i.e. how long are we willing to wait
201 * for reply before timing out this request.
203 void ptlrpc_at_set_req_timeout(struct ptlrpc_request *req)
209 LASSERT(req->rq_import);
212 /* non-AT settings */
214 * \a imp_server_timeout means this is reverse import and
215 * we send (currently only) ASTs to the client and cannot afford
216 * to wait too long for the reply, otherwise the other client
217 * (because of which we are sending this request) would
218 * timeout waiting for us
220 req->rq_timeout = req->rq_import->imp_server_timeout ?
221 obd_timeout / 2 : obd_timeout;
223 at = &req->rq_import->imp_at;
224 idx = import_at_get_index(req->rq_import,
225 req->rq_request_portal);
226 serv_est = at_get(&at->iat_service_estimate[idx]);
227 req->rq_timeout = at_est2timeout(serv_est);
229 /* We could get even fancier here, using history to predict increased
232 /* Let the server know what this RPC timeout is by putting it in the
234 lustre_msg_set_timeout(req->rq_reqmsg, req->rq_timeout);
237 /* Adjust max service estimate based on server value */
238 static void ptlrpc_at_adj_service(struct ptlrpc_request *req,
239 unsigned int serv_est)
245 LASSERT(req->rq_import);
246 at = &req->rq_import->imp_at;
248 idx = import_at_get_index(req->rq_import, req->rq_request_portal);
249 /* max service estimates are tracked on the server side,
250 so just keep minimal history here */
251 oldse = at_measured(&at->iat_service_estimate[idx], serv_est);
253 CDEBUG(D_ADAPTTO, "The RPC service estimate for %s ptl %d "
254 "has changed from %d to %d\n",
255 req->rq_import->imp_obd->obd_name,req->rq_request_portal,
256 oldse, at_get(&at->iat_service_estimate[idx]));
259 /* Expected network latency per remote node (secs) */
260 int ptlrpc_at_get_net_latency(struct ptlrpc_request *req)
262 return AT_OFF ? 0 : at_get(&req->rq_import->imp_at.iat_net_latency);
265 /* Adjust expected network latency */
266 static void ptlrpc_at_adj_net_latency(struct ptlrpc_request *req,
267 unsigned int service_time)
269 unsigned int nl, oldnl;
271 time_t now = cfs_time_current_sec();
273 LASSERT(req->rq_import);
274 at = &req->rq_import->imp_at;
276 /* Network latency is total time less server processing time */
277 nl = max_t(int, now - req->rq_sent - service_time, 0) +1/*st rounding*/;
278 if (service_time > now - req->rq_sent + 3 /* bz16408 */)
279 CWARN("Reported service time %u > total measured time "
280 CFS_DURATION_T"\n", service_time,
281 cfs_time_sub(now, req->rq_sent));
283 oldnl = at_measured(&at->iat_net_latency, nl);
285 CDEBUG(D_ADAPTTO, "The network latency for %s (nid %s) "
286 "has changed from %d to %d\n",
287 req->rq_import->imp_obd->obd_name,
289 &req->rq_import->imp_connection->c_remote_uuid),
290 oldnl, at_get(&at->iat_net_latency));
293 static int unpack_reply(struct ptlrpc_request *req)
297 if (SPTLRPC_FLVR_POLICY(req->rq_flvr.sf_rpc) != SPTLRPC_POLICY_NULL) {
298 rc = ptlrpc_unpack_rep_msg(req, req->rq_replen);
300 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
305 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
307 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
314 * Handle an early reply message, called with the rq_lock held.
315 * If anything goes wrong just ignore it - same as if it never happened
317 static int ptlrpc_at_recv_early_reply(struct ptlrpc_request *req)
319 struct ptlrpc_request *early_req;
325 cfs_spin_unlock(&req->rq_lock);
327 rc = sptlrpc_cli_unwrap_early_reply(req, &early_req);
329 cfs_spin_lock(&req->rq_lock);
333 rc = unpack_reply(early_req);
335 /* Expecting to increase the service time estimate here */
336 ptlrpc_at_adj_service(req,
337 lustre_msg_get_timeout(early_req->rq_repmsg));
338 ptlrpc_at_adj_net_latency(req,
339 lustre_msg_get_service_time(early_req->rq_repmsg));
342 sptlrpc_cli_finish_early_reply(early_req);
344 cfs_spin_lock(&req->rq_lock);
347 /* Adjust the local timeout for this req */
348 ptlrpc_at_set_req_timeout(req);
350 olddl = req->rq_deadline;
351 /* server assumes it now has rq_timeout from when it sent the
352 early reply, so client should give it at least that long. */
353 req->rq_deadline = cfs_time_current_sec() + req->rq_timeout +
354 ptlrpc_at_get_net_latency(req);
356 DEBUG_REQ(D_ADAPTTO, req,
357 "Early reply #%d, new deadline in "CFS_DURATION_T"s "
358 "("CFS_DURATION_T"s)", req->rq_early_count,
359 cfs_time_sub(req->rq_deadline,
360 cfs_time_current_sec()),
361 cfs_time_sub(req->rq_deadline, olddl));
368 * Wind down request pool \a pool.
369 * Frees all requests from the pool too
371 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
374 struct ptlrpc_request *req;
376 LASSERT(pool != NULL);
378 cfs_spin_lock(&pool->prp_lock);
379 cfs_list_for_each_safe(l, tmp, &pool->prp_req_list) {
380 req = cfs_list_entry(l, struct ptlrpc_request, rq_list);
381 cfs_list_del(&req->rq_list);
382 LASSERT(req->rq_reqbuf);
383 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
384 OBD_FREE_LARGE(req->rq_reqbuf, pool->prp_rq_size);
385 OBD_FREE(req, sizeof(*req));
387 cfs_spin_unlock(&pool->prp_lock);
388 OBD_FREE(pool, sizeof(*pool));
392 * Allocates, initializes and adds \a num_rq requests to the pool \a pool
394 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
399 while (size < pool->prp_rq_size)
402 LASSERTF(cfs_list_empty(&pool->prp_req_list) ||
403 size == pool->prp_rq_size,
404 "Trying to change pool size with nonempty pool "
405 "from %d to %d bytes\n", pool->prp_rq_size, size);
407 cfs_spin_lock(&pool->prp_lock);
408 pool->prp_rq_size = size;
409 for (i = 0; i < num_rq; i++) {
410 struct ptlrpc_request *req;
411 struct lustre_msg *msg;
413 cfs_spin_unlock(&pool->prp_lock);
414 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
417 OBD_ALLOC_LARGE(msg, size);
419 OBD_FREE(req, sizeof(struct ptlrpc_request));
422 req->rq_reqbuf = msg;
423 req->rq_reqbuf_len = size;
425 cfs_spin_lock(&pool->prp_lock);
426 cfs_list_add_tail(&req->rq_list, &pool->prp_req_list);
428 cfs_spin_unlock(&pool->prp_lock);
433 * Create and initialize new request pool with given attributes:
434 * \a num_rq - initial number of requests to create for the pool
435 * \a msgsize - maximum message size possible for requests in thid pool
436 * \a populate_pool - function to be called when more requests need to be added
438 * Returns pointer to newly created pool or NULL on error.
440 struct ptlrpc_request_pool *
441 ptlrpc_init_rq_pool(int num_rq, int msgsize,
442 void (*populate_pool)(struct ptlrpc_request_pool *, int))
444 struct ptlrpc_request_pool *pool;
446 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
450 /* Request next power of two for the allocation, because internally
451 kernel would do exactly this */
453 cfs_spin_lock_init(&pool->prp_lock);
454 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
455 pool->prp_rq_size = msgsize + SPTLRPC_MAX_PAYLOAD;
456 pool->prp_populate = populate_pool;
458 populate_pool(pool, num_rq);
460 if (cfs_list_empty(&pool->prp_req_list)) {
461 /* have not allocated a single request for the pool */
462 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
469 * Fetches one request from pool \a pool
471 static struct ptlrpc_request *
472 ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
474 struct ptlrpc_request *request;
475 struct lustre_msg *reqbuf;
480 cfs_spin_lock(&pool->prp_lock);
482 /* See if we have anything in a pool, and bail out if nothing,
483 * in writeout path, where this matters, this is safe to do, because
484 * nothing is lost in this case, and when some in-flight requests
485 * complete, this code will be called again. */
486 if (unlikely(cfs_list_empty(&pool->prp_req_list))) {
487 cfs_spin_unlock(&pool->prp_lock);
491 request = cfs_list_entry(pool->prp_req_list.next, struct ptlrpc_request,
493 cfs_list_del_init(&request->rq_list);
494 cfs_spin_unlock(&pool->prp_lock);
496 LASSERT(request->rq_reqbuf);
497 LASSERT(request->rq_pool);
499 reqbuf = request->rq_reqbuf;
500 memset(request, 0, sizeof(*request));
501 request->rq_reqbuf = reqbuf;
502 request->rq_reqbuf_len = pool->prp_rq_size;
503 request->rq_pool = pool;
509 * Returns freed \a request to pool.
511 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
513 struct ptlrpc_request_pool *pool = request->rq_pool;
515 cfs_spin_lock(&pool->prp_lock);
516 LASSERT(cfs_list_empty(&request->rq_list));
517 LASSERT(!request->rq_receiving_reply);
518 cfs_list_add_tail(&request->rq_list, &pool->prp_req_list);
519 cfs_spin_unlock(&pool->prp_lock);
522 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
523 __u32 version, int opcode,
524 int count, __u32 *lengths, char **bufs,
525 struct ptlrpc_cli_ctx *ctx)
527 struct obd_import *imp = request->rq_import;
532 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
534 rc = sptlrpc_req_get_ctx(request);
539 sptlrpc_req_set_flavor(request, opcode);
541 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
544 LASSERT(!request->rq_pool);
548 lustre_msg_add_version(request->rq_reqmsg, version);
549 request->rq_send_state = LUSTRE_IMP_FULL;
550 request->rq_type = PTL_RPC_MSG_REQUEST;
551 request->rq_export = NULL;
553 request->rq_req_cbid.cbid_fn = request_out_callback;
554 request->rq_req_cbid.cbid_arg = request;
556 request->rq_reply_cbid.cbid_fn = reply_in_callback;
557 request->rq_reply_cbid.cbid_arg = request;
559 request->rq_reply_deadline = 0;
560 request->rq_phase = RQ_PHASE_NEW;
561 request->rq_next_phase = RQ_PHASE_UNDEFINED;
563 request->rq_request_portal = imp->imp_client->cli_request_portal;
564 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
566 ptlrpc_at_set_req_timeout(request);
568 cfs_spin_lock_init(&request->rq_lock);
569 CFS_INIT_LIST_HEAD(&request->rq_list);
570 CFS_INIT_LIST_HEAD(&request->rq_timed_list);
571 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
572 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
573 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
574 CFS_INIT_LIST_HEAD(&request->rq_history_list);
575 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
576 cfs_waitq_init(&request->rq_reply_waitq);
577 cfs_waitq_init(&request->rq_set_waitq);
578 request->rq_xid = ptlrpc_next_xid();
579 cfs_atomic_set(&request->rq_refcount, 1);
581 lustre_msg_set_opc(request->rq_reqmsg, opcode);
585 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
587 class_import_put(imp);
591 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
592 __u32 version, int opcode, char **bufs,
593 struct ptlrpc_cli_ctx *ctx)
597 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
598 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
599 request->rq_pill.rc_area[RCL_CLIENT],
602 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
605 * Pack request buffers for network transfer, performing necessary encryption
606 * steps if necessary.
608 int ptlrpc_request_pack(struct ptlrpc_request *request,
609 __u32 version, int opcode)
612 rc = ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
616 /* For some old 1.8 clients (< 1.8.7), they will LASSERT the size of
617 * ptlrpc_body sent from server equal to local ptlrpc_body size, so we
618 * have to send old ptlrpc_body to keep interoprability with these
621 * Only three kinds of server->client RPCs so far:
626 * XXX This should be removed whenever we drop the interoprability with
627 * the these old clients.
629 if (opcode == LDLM_BL_CALLBACK || opcode == LDLM_CP_CALLBACK ||
630 opcode == LDLM_GL_CALLBACK)
631 req_capsule_shrink(&request->rq_pill, &RMF_PTLRPC_BODY,
632 sizeof(struct ptlrpc_body_v2), RCL_CLIENT);
638 * Helper function to allocate new request on import \a imp
639 * and possibly using existing request from pool \a pool if provided.
640 * Returns allocated request structure with import field filled or
644 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
645 struct ptlrpc_request_pool *pool)
647 struct ptlrpc_request *request = NULL;
650 request = ptlrpc_prep_req_from_pool(pool);
653 OBD_ALLOC_PTR(request);
656 LASSERTF((unsigned long)imp > 0x1000, "%p", imp);
657 LASSERT(imp != LP_POISON);
658 LASSERTF((unsigned long)imp->imp_client > 0x1000, "%p",
660 LASSERT(imp->imp_client != LP_POISON);
662 request->rq_import = class_import_get(imp);
664 CERROR("request allocation out of memory\n");
671 * Helper function for creating a request.
672 * Calls __ptlrpc_request_alloc to allocate new request sturcture and inits
673 * buffer structures according to capsule template \a format.
674 * Returns allocated request structure pointer or NULL on error.
676 static struct ptlrpc_request *
677 ptlrpc_request_alloc_internal(struct obd_import *imp,
678 struct ptlrpc_request_pool * pool,
679 const struct req_format *format)
681 struct ptlrpc_request *request;
683 request = __ptlrpc_request_alloc(imp, pool);
687 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
688 req_capsule_set(&request->rq_pill, format);
693 * Allocate new request structure for import \a imp and initialize its
694 * buffer structure according to capsule template \a format.
696 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
697 const struct req_format *format)
699 return ptlrpc_request_alloc_internal(imp, NULL, format);
703 * Allocate new request structure for import \a imp from pool \a pool and
704 * initialize its buffer structure according to capsule template \a format.
706 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
707 struct ptlrpc_request_pool * pool,
708 const struct req_format *format)
710 return ptlrpc_request_alloc_internal(imp, pool, format);
714 * For requests not from pool, free memory of the request structure.
715 * For requests obtained from a pool earlier, return request back to pool.
717 void ptlrpc_request_free(struct ptlrpc_request *request)
719 if (request->rq_pool)
720 __ptlrpc_free_req_to_pool(request);
722 OBD_FREE_PTR(request);
726 * Allocate new request for operatione \a opcode and immediatelly pack it for
728 * Only used for simple requests like OBD_PING where the only important
729 * part of the request is operation itself.
730 * Returns allocated request or NULL on error.
732 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
733 const struct req_format *format,
734 __u32 version, int opcode)
736 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
740 rc = ptlrpc_request_pack(req, version, opcode);
742 ptlrpc_request_free(req);
750 * Prepare request (fetched from pool \a poolif not NULL) on import \a imp
751 * for operation \a opcode. Request would contain \a count buffers.
752 * Sizes of buffers are described in array \a lengths and buffers themselves
753 * are provided by a pointer \a bufs.
754 * Returns prepared request structure pointer or NULL on error.
756 struct ptlrpc_request *
757 ptlrpc_prep_req_pool(struct obd_import *imp,
758 __u32 version, int opcode,
759 int count, __u32 *lengths, char **bufs,
760 struct ptlrpc_request_pool *pool)
762 struct ptlrpc_request *request;
765 request = __ptlrpc_request_alloc(imp, pool);
769 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
770 lengths, bufs, NULL);
772 ptlrpc_request_free(request);
779 * Same as ptlrpc_prep_req_pool, but without pool
781 struct ptlrpc_request *
782 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
783 __u32 *lengths, char **bufs)
785 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
790 * Allocate "fake" request that would not be sent anywhere in the end.
791 * Only used as a hack because we have no other way of performing
792 * async actions in lustre between layers.
793 * Used on MDS to request object preallocations from more than one OST at a
796 struct ptlrpc_request *ptlrpc_prep_fakereq(struct obd_import *imp,
797 unsigned int timeout,
798 ptlrpc_interpterer_t interpreter)
800 struct ptlrpc_request *request = NULL;
803 OBD_ALLOC(request, sizeof(*request));
805 CERROR("request allocation out of memory\n");
809 request->rq_send_state = LUSTRE_IMP_FULL;
810 request->rq_type = PTL_RPC_MSG_REQUEST;
811 request->rq_import = class_import_get(imp);
812 request->rq_export = NULL;
813 request->rq_import_generation = imp->imp_generation;
815 request->rq_timeout = timeout;
816 request->rq_sent = cfs_time_current_sec();
817 request->rq_deadline = request->rq_sent + timeout;
818 request->rq_reply_deadline = request->rq_deadline;
819 request->rq_interpret_reply = interpreter;
820 request->rq_phase = RQ_PHASE_RPC;
821 request->rq_next_phase = RQ_PHASE_INTERPRET;
822 /* don't want reply */
823 request->rq_receiving_reply = 0;
824 request->rq_must_unlink = 0;
825 request->rq_no_delay = request->rq_no_resend = 1;
826 request->rq_fake = 1;
828 cfs_spin_lock_init(&request->rq_lock);
829 CFS_INIT_LIST_HEAD(&request->rq_list);
830 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
831 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
832 CFS_INIT_LIST_HEAD(&request->rq_history_list);
833 CFS_INIT_LIST_HEAD(&request->rq_exp_list);
834 cfs_waitq_init(&request->rq_reply_waitq);
835 cfs_waitq_init(&request->rq_set_waitq);
837 request->rq_xid = ptlrpc_next_xid();
838 cfs_atomic_set(&request->rq_refcount, 1);
844 * Indicate that processing of "fake" request is finished.
846 void ptlrpc_fakereq_finished(struct ptlrpc_request *req)
848 struct ptlrpc_request_set *set = req->rq_set;
851 /* hold ref on the request to prevent others (ptlrpcd) to free it */
852 ptlrpc_request_addref(req);
853 cfs_list_del_init(&req->rq_list);
855 /* if we kill request before timeout - need adjust counter */
856 if (req->rq_phase == RQ_PHASE_RPC && set != NULL &&
857 cfs_atomic_dec_and_test(&set->set_remaining))
860 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
862 /* Only need to call wakeup once when to be empty. */
864 cfs_waitq_signal(&set->set_waitq);
865 ptlrpc_req_finished(req);
869 * Allocate and initialize new request set structure.
870 * Returns a pointer to the newly allocated set structure or NULL on error.
872 struct ptlrpc_request_set *ptlrpc_prep_set(void)
874 struct ptlrpc_request_set *set;
877 OBD_ALLOC(set, sizeof *set);
880 cfs_atomic_set(&set->set_refcount, 1);
881 CFS_INIT_LIST_HEAD(&set->set_requests);
882 cfs_waitq_init(&set->set_waitq);
883 cfs_atomic_set(&set->set_new_count, 0);
884 cfs_atomic_set(&set->set_remaining, 0);
885 cfs_spin_lock_init(&set->set_new_req_lock);
886 CFS_INIT_LIST_HEAD(&set->set_new_requests);
887 CFS_INIT_LIST_HEAD(&set->set_cblist);
888 set->set_max_inflight = UINT_MAX;
889 set->set_producer = NULL;
890 set->set_producer_arg = NULL;
897 * Allocate and initialize new request set structure with flow control
898 * extension. This extension allows to control the number of requests in-flight
899 * for the whole set. A callback function to generate requests must be provided
900 * and the request set will keep the number of requests sent over the wire to
902 * Returns a pointer to the newly allocated set structure or NULL on error.
904 struct ptlrpc_request_set *ptlrpc_prep_fcset(int max, set_producer_func func,
908 struct ptlrpc_request_set *set;
910 set = ptlrpc_prep_set();
914 set->set_max_inflight = max;
915 set->set_producer = func;
916 set->set_producer_arg = arg;
920 EXPORT_SYMBOL(ptlrpc_prep_fcset);
923 * Wind down and free request set structure previously allocated with
925 * Ensures that all requests on the set have completed and removes
926 * all requests from the request list in a set.
927 * If any unsent request happen to be on the list, pretends that they got
928 * an error in flight and calls their completion handler.
930 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
938 /* Requests on the set should either all be completed, or all be new */
939 expected_phase = (cfs_atomic_read(&set->set_remaining) == 0) ?
940 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
941 cfs_list_for_each (tmp, &set->set_requests) {
942 struct ptlrpc_request *req =
943 cfs_list_entry(tmp, struct ptlrpc_request,
946 LASSERT(req->rq_phase == expected_phase);
950 LASSERTF(cfs_atomic_read(&set->set_remaining) == 0 ||
951 cfs_atomic_read(&set->set_remaining) == n, "%d / %d\n",
952 cfs_atomic_read(&set->set_remaining), n);
954 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
955 struct ptlrpc_request *req =
956 cfs_list_entry(tmp, struct ptlrpc_request,
958 cfs_list_del_init(&req->rq_set_chain);
960 LASSERT(req->rq_phase == expected_phase);
962 if (req->rq_phase == RQ_PHASE_NEW) {
963 ptlrpc_req_interpret(NULL, req, -EBADR);
964 cfs_atomic_dec(&set->set_remaining);
967 cfs_spin_lock(&req->rq_lock);
969 req->rq_invalid_rqset = 0;
970 cfs_spin_unlock(&req->rq_lock);
972 ptlrpc_req_finished (req);
975 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
977 ptlrpc_reqset_put(set);
982 * Add a callback function \a fn to the set.
983 * This function would be called when all requests on this set are completed.
984 * The function will be passed \a data argument.
986 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
987 set_interpreter_func fn, void *data)
989 struct ptlrpc_set_cbdata *cbdata;
991 OBD_ALLOC_PTR(cbdata);
995 cbdata->psc_interpret = fn;
996 cbdata->psc_data = data;
997 cfs_list_add_tail(&cbdata->psc_item, &set->set_cblist);
1003 * Add a new request to the general purpose request set.
1004 * Assumes request reference from the caller.
1006 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
1007 struct ptlrpc_request *req)
1009 char jobid[JOBSTATS_JOBID_SIZE];
1010 LASSERT(cfs_list_empty(&req->rq_set_chain));
1012 /* The set takes over the caller's request reference */
1013 cfs_list_add_tail(&req->rq_set_chain, &set->set_requests);
1015 cfs_atomic_inc(&set->set_remaining);
1016 req->rq_queued_time = cfs_time_current();
1018 if (req->rq_reqmsg) {
1019 lustre_get_jobid(jobid);
1020 lustre_msg_set_jobid(req->rq_reqmsg, jobid);
1023 if (set->set_producer != NULL)
1024 /* If the request set has a producer callback, the RPC must be
1025 * sent straight away */
1026 ptlrpc_send_new_req(req);
1030 * Add a request to a request with dedicated server thread
1031 * and wake the thread to make any necessary processing.
1032 * Currently only used for ptlrpcd.
1034 void ptlrpc_set_add_new_req(struct ptlrpcd_ctl *pc,
1035 struct ptlrpc_request *req)
1037 struct ptlrpc_request_set *set = pc->pc_set;
1040 LASSERT(req->rq_set == NULL);
1041 LASSERT(cfs_test_bit(LIOD_STOP, &pc->pc_flags) == 0);
1043 cfs_spin_lock(&set->set_new_req_lock);
1045 * The set takes over the caller's request reference.
1048 req->rq_queued_time = cfs_time_current();
1049 cfs_list_add_tail(&req->rq_set_chain, &set->set_new_requests);
1050 count = cfs_atomic_inc_return(&set->set_new_count);
1051 cfs_spin_unlock(&set->set_new_req_lock);
1053 /* Only need to call wakeup once for the first entry. */
1055 cfs_waitq_signal(&set->set_waitq);
1057 /* XXX: It maybe unnecessary to wakeup all the partners. But to
1058 * guarantee the async RPC can be processed ASAP, we have
1059 * no other better choice. It maybe fixed in future. */
1060 for (i = 0; i < pc->pc_npartners; i++)
1061 cfs_waitq_signal(&pc->pc_partners[i]->pc_set->set_waitq);
1066 * Based on the current state of the import, determine if the request
1067 * can be sent, is an error, or should be delayed.
1069 * Returns true if this request should be delayed. If false, and
1070 * *status is set, then the request can not be sent and *status is the
1071 * error code. If false and status is 0, then request can be sent.
1073 * The imp->imp_lock must be held.
1075 static int ptlrpc_import_delay_req(struct obd_import *imp,
1076 struct ptlrpc_request *req, int *status)
1081 LASSERT (status != NULL);
1084 if (req->rq_ctx_init || req->rq_ctx_fini) {
1085 /* always allow ctx init/fini rpc go through */
1086 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
1087 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
1090 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
1091 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
1093 } else if (ptlrpc_send_limit_expired(req)) {
1094 /* probably doesn't need to be a D_ERROR after initial testing */
1095 DEBUG_REQ(D_ERROR, req, "send limit expired ");
1097 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
1098 imp->imp_state == LUSTRE_IMP_CONNECTING) {
1099 /* allow CONNECT even if import is invalid */ ;
1100 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1101 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1104 } else if (imp->imp_invalid || imp->imp_obd->obd_no_recov) {
1105 if (!imp->imp_deactive)
1106 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
1107 *status = -ESHUTDOWN; /* bz 12940 */
1108 } else if (req->rq_import_generation != imp->imp_generation) {
1109 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
1111 } else if (req->rq_send_state != imp->imp_state) {
1112 /* invalidate in progress - any requests should be drop */
1113 if (cfs_atomic_read(&imp->imp_inval_count) != 0) {
1114 DEBUG_REQ(D_ERROR, req, "invalidate in flight");
1116 } else if (imp->imp_dlm_fake || req->rq_no_delay) {
1117 *status = -EWOULDBLOCK;
1127 * Decide if the eror message regarding provided request \a req
1128 * should be printed to the console or not.
1129 * Makes it's decision on request status and other properties.
1130 * Returns 1 to print error on the system console or 0 if not.
1132 static int ptlrpc_console_allow(struct ptlrpc_request *req)
1137 /* Fake requests include no rq_reqmsg */
1141 LASSERT(req->rq_reqmsg != NULL);
1142 opc = lustre_msg_get_opc(req->rq_reqmsg);
1144 /* Suppress particular reconnect errors which are to be expected. No
1145 * errors are suppressed for the initial connection on an import */
1146 if ((lustre_handle_is_used(&req->rq_import->imp_remote_handle)) &&
1147 (opc == OST_CONNECT || opc == MDS_CONNECT || opc == MGS_CONNECT)) {
1149 /* Suppress timed out reconnect requests */
1150 if (req->rq_timedout)
1153 /* Suppress unavailable/again reconnect requests */
1154 err = lustre_msg_get_status(req->rq_repmsg);
1155 if (err == -ENODEV || err == -EAGAIN)
1163 * Check request processing status.
1164 * Returns the status.
1166 static int ptlrpc_check_status(struct ptlrpc_request *req)
1171 err = lustre_msg_get_status(req->rq_repmsg);
1172 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1173 struct obd_import *imp = req->rq_import;
1174 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1175 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
1176 " with %s. The %s operation failed with %d\n",
1177 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1178 ll_opcode2str(opc), err);
1179 RETURN(err < 0 ? err : -EINVAL);
1183 DEBUG_REQ(D_INFO, req, "status is %d", err);
1184 } else if (err > 0) {
1185 /* XXX: translate this error from net to host */
1186 DEBUG_REQ(D_INFO, req, "status is %d", err);
1189 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
1190 struct obd_import *imp = req->rq_import;
1191 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
1193 if (ptlrpc_console_allow(req))
1194 LCONSOLE_ERROR_MSG(0x011,"an error occurred while "
1195 "communicating with %s. The %s "
1196 "operation failed with %d\n",
1198 imp->imp_connection->c_peer.nid),
1199 ll_opcode2str(opc), err);
1201 RETURN(err < 0 ? err : -EINVAL);
1208 * save pre-versions of objects into request for replay.
1209 * Versions are obtained from server reply.
1212 static void ptlrpc_save_versions(struct ptlrpc_request *req)
1214 struct lustre_msg *repmsg = req->rq_repmsg;
1215 struct lustre_msg *reqmsg = req->rq_reqmsg;
1216 __u64 *versions = lustre_msg_get_versions(repmsg);
1219 if (lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)
1223 lustre_msg_set_versions(reqmsg, versions);
1224 CDEBUG(D_INFO, "Client save versions ["LPX64"/"LPX64"]\n",
1225 versions[0], versions[1]);
1231 * Callback function called when client receives RPC reply for \a req.
1232 * Returns 0 on success or error code.
1233 * The return alue would be assigned to req->rq_status by the caller
1234 * as request processing status.
1235 * This function also decides if the request needs to be saved for later replay.
1237 static int after_reply(struct ptlrpc_request *req)
1239 struct obd_import *imp = req->rq_import;
1240 struct obd_device *obd = req->rq_import->imp_obd;
1242 struct timeval work_start;
1246 LASSERT(obd != NULL);
1247 /* repbuf must be unlinked */
1248 LASSERT(!req->rq_receiving_reply && !req->rq_must_unlink);
1250 if (req->rq_reply_truncate) {
1251 if (ptlrpc_no_resend(req)) {
1252 DEBUG_REQ(D_ERROR, req, "reply buffer overflow,"
1253 " expected: %d, actual size: %d",
1254 req->rq_nob_received, req->rq_repbuf_len);
1258 sptlrpc_cli_free_repbuf(req);
1259 /* Pass the required reply buffer size (include
1260 * space for early reply).
1261 * NB: no need to roundup because alloc_repbuf
1262 * will roundup it */
1263 req->rq_replen = req->rq_nob_received;
1264 req->rq_nob_received = 0;
1270 * NB Until this point, the whole of the incoming message,
1271 * including buflens, status etc is in the sender's byte order.
1273 rc = sptlrpc_cli_unwrap_reply(req);
1275 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
1280 * Security layer unwrap might ask resend this request.
1285 rc = unpack_reply(req);
1289 cfs_gettimeofday(&work_start);
1290 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
1291 if (obd->obd_svc_stats != NULL) {
1292 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
1294 ptlrpc_lprocfs_rpc_sent(req, timediff);
1297 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
1298 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
1299 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
1300 lustre_msg_get_type(req->rq_repmsg));
1304 if (lustre_msg_get_opc(req->rq_reqmsg) != OBD_PING)
1305 CFS_FAIL_TIMEOUT(OBD_FAIL_PTLRPC_PAUSE_REP, cfs_fail_val);
1306 ptlrpc_at_adj_service(req, lustre_msg_get_timeout(req->rq_repmsg));
1307 ptlrpc_at_adj_net_latency(req,
1308 lustre_msg_get_service_time(req->rq_repmsg));
1310 rc = ptlrpc_check_status(req);
1311 imp->imp_connect_error = rc;
1315 * Either we've been evicted, or the server has failed for
1316 * some reason. Try to reconnect, and if that fails, punt to
1319 if (ll_rpc_recoverable_error(rc)) {
1320 if (req->rq_send_state != LUSTRE_IMP_FULL ||
1321 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
1324 ptlrpc_request_handle_notconn(req);
1329 * Let's look if server sent slv. Do it only for RPC with
1332 ldlm_cli_update_pool(req);
1336 * Store transno in reqmsg for replay.
1338 if (!(lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY)) {
1339 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
1340 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
1343 if (imp->imp_replayable) {
1344 cfs_spin_lock(&imp->imp_lock);
1346 * No point in adding already-committed requests to the replay
1347 * list, we will just remove them immediately. b=9829
1349 if (req->rq_transno != 0 &&
1351 lustre_msg_get_last_committed(req->rq_repmsg) ||
1353 /** version recovery */
1354 ptlrpc_save_versions(req);
1355 ptlrpc_retain_replayable_request(req, imp);
1356 } else if (req->rq_commit_cb != NULL) {
1357 cfs_spin_unlock(&imp->imp_lock);
1358 req->rq_commit_cb(req);
1359 cfs_spin_lock(&imp->imp_lock);
1363 * Replay-enabled imports return commit-status information.
1365 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
1366 imp->imp_peer_committed_transno =
1367 lustre_msg_get_last_committed(req->rq_repmsg);
1369 ptlrpc_free_committed(imp);
1371 if (req->rq_transno > imp->imp_peer_committed_transno)
1372 ptlrpc_pinger_commit_expected(imp);
1374 cfs_spin_unlock(&imp->imp_lock);
1381 * Helper function to send request \a req over the network for the first time
1382 * Also adjusts request phase.
1383 * Returns 0 on success or error code.
1385 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
1387 struct obd_import *imp = req->rq_import;
1391 LASSERT(req->rq_phase == RQ_PHASE_NEW);
1392 if (req->rq_sent && (req->rq_sent > cfs_time_current_sec()) &&
1393 (!req->rq_generation_set ||
1394 req->rq_import_generation == imp->imp_generation))
1397 ptlrpc_rqphase_move(req, RQ_PHASE_RPC);
1399 cfs_spin_lock(&imp->imp_lock);
1401 if (!req->rq_generation_set)
1402 req->rq_import_generation = imp->imp_generation;
1404 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1405 cfs_spin_lock(&req->rq_lock);
1406 req->rq_waiting = 1;
1407 cfs_spin_unlock(&req->rq_lock);
1409 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
1410 "(%s != %s)", lustre_msg_get_status(req->rq_reqmsg),
1411 ptlrpc_import_state_name(req->rq_send_state),
1412 ptlrpc_import_state_name(imp->imp_state));
1413 LASSERT(cfs_list_empty(&req->rq_list));
1414 cfs_list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1415 cfs_atomic_inc(&req->rq_import->imp_inflight);
1416 cfs_spin_unlock(&imp->imp_lock);
1421 cfs_spin_unlock(&imp->imp_lock);
1422 req->rq_status = rc;
1423 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1427 LASSERT(cfs_list_empty(&req->rq_list));
1428 cfs_list_add_tail(&req->rq_list, &imp->imp_sending_list);
1429 cfs_atomic_inc(&req->rq_import->imp_inflight);
1430 cfs_spin_unlock(&imp->imp_lock);
1432 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1434 rc = sptlrpc_req_refresh_ctx(req, -1);
1437 req->rq_status = rc;
1440 req->rq_wait_ctx = 1;
1445 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
1446 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1447 imp->imp_obd->obd_uuid.uuid,
1448 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1449 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1450 lustre_msg_get_opc(req->rq_reqmsg));
1452 rc = ptl_send_rpc(req, 0);
1454 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
1455 req->rq_net_err = 1;
1461 static inline int ptlrpc_set_producer(struct ptlrpc_request_set *set)
1466 LASSERT(set->set_producer != NULL);
1468 remaining = cfs_atomic_read(&set->set_remaining);
1470 /* populate the ->set_requests list with requests until we
1471 * reach the maximum number of RPCs in flight for this set */
1472 while (cfs_atomic_read(&set->set_remaining) < set->set_max_inflight) {
1473 rc = set->set_producer(set, set->set_producer_arg);
1474 if (rc == -ENOENT) {
1475 /* no more RPC to produce */
1476 set->set_producer = NULL;
1477 set->set_producer_arg = NULL;
1482 RETURN((cfs_atomic_read(&set->set_remaining) - remaining));
1486 * this sends any unsent RPCs in \a set and returns 1 if all are sent
1487 * and no more replies are expected.
1488 * (it is possible to get less replies than requests sent e.g. due to timed out
1489 * requests or requests that we had trouble to send out)
1491 int ptlrpc_check_set(const struct lu_env *env, struct ptlrpc_request_set *set)
1493 cfs_list_t *tmp, *next;
1494 int force_timer_recalc = 0;
1497 if (cfs_atomic_read(&set->set_remaining) == 0)
1500 cfs_list_for_each_safe(tmp, next, &set->set_requests) {
1501 struct ptlrpc_request *req =
1502 cfs_list_entry(tmp, struct ptlrpc_request,
1504 struct obd_import *imp = req->rq_import;
1505 int unregistered = 0;
1508 if (req->rq_phase == RQ_PHASE_NEW &&
1509 ptlrpc_send_new_req(req)) {
1510 force_timer_recalc = 1;
1513 /* delayed send - skip */
1514 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
1517 if (!(req->rq_phase == RQ_PHASE_RPC ||
1518 req->rq_phase == RQ_PHASE_BULK ||
1519 req->rq_phase == RQ_PHASE_INTERPRET ||
1520 req->rq_phase == RQ_PHASE_UNREGISTERING ||
1521 req->rq_phase == RQ_PHASE_COMPLETE)) {
1522 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
1526 if (req->rq_phase == RQ_PHASE_UNREGISTERING) {
1527 LASSERT(req->rq_next_phase != req->rq_phase);
1528 LASSERT(req->rq_next_phase != RQ_PHASE_UNDEFINED);
1531 * Skip processing until reply is unlinked. We
1532 * can't return to pool before that and we can't
1533 * call interpret before that. We need to make
1534 * sure that all rdma transfers finished and will
1535 * not corrupt any data.
1537 if (ptlrpc_client_recv_or_unlink(req) ||
1538 ptlrpc_client_bulk_active(req))
1542 * Turn fail_loc off to prevent it from looping
1545 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
1546 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK,
1549 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK)) {
1550 OBD_FAIL_CHECK_ORSET(OBD_FAIL_PTLRPC_LONG_BULK_UNLINK,
1555 * Move to next phase if reply was successfully
1558 ptlrpc_rqphase_move(req, req->rq_next_phase);
1561 if (req->rq_phase == RQ_PHASE_COMPLETE)
1564 if (req->rq_phase == RQ_PHASE_INTERPRET)
1565 GOTO(interpret, req->rq_status);
1568 * Note that this also will start async reply unlink.
1570 if (req->rq_net_err && !req->rq_timedout) {
1571 ptlrpc_expire_one_request(req, 1);
1574 * Check if we still need to wait for unlink.
1576 if (ptlrpc_client_recv_or_unlink(req) ||
1577 ptlrpc_client_bulk_active(req))
1579 /* If there is no need to resend, fail it now. */
1580 if (req->rq_no_resend) {
1581 if (req->rq_status == 0)
1582 req->rq_status = -EIO;
1583 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1584 GOTO(interpret, req->rq_status);
1591 cfs_spin_lock(&req->rq_lock);
1592 req->rq_replied = 0;
1593 cfs_spin_unlock(&req->rq_lock);
1594 if (req->rq_status == 0)
1595 req->rq_status = -EIO;
1596 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1597 GOTO(interpret, req->rq_status);
1600 /* ptlrpc_set_wait->l_wait_event sets lwi_allow_intr
1601 * so it sets rq_intr regardless of individual rpc
1602 * timeouts. The synchronous IO waiting path sets
1603 * rq_intr irrespective of whether ptlrpcd
1604 * has seen a timeout. Our policy is to only interpret
1605 * interrupted rpcs after they have timed out, so we
1606 * need to enforce that here.
1609 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
1610 req->rq_wait_ctx)) {
1611 req->rq_status = -EINTR;
1612 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1613 GOTO(interpret, req->rq_status);
1616 if (req->rq_phase == RQ_PHASE_RPC) {
1617 if (req->rq_timedout || req->rq_resend ||
1618 req->rq_waiting || req->rq_wait_ctx) {
1621 if (!ptlrpc_unregister_reply(req, 1))
1624 cfs_spin_lock(&imp->imp_lock);
1625 if (ptlrpc_import_delay_req(imp, req, &status)){
1626 /* put on delay list - only if we wait
1627 * recovery finished - before send */
1628 cfs_list_del_init(&req->rq_list);
1629 cfs_list_add_tail(&req->rq_list,
1632 cfs_spin_unlock(&imp->imp_lock);
1637 req->rq_status = status;
1638 ptlrpc_rqphase_move(req,
1639 RQ_PHASE_INTERPRET);
1640 cfs_spin_unlock(&imp->imp_lock);
1641 GOTO(interpret, req->rq_status);
1643 if (ptlrpc_no_resend(req) && !req->rq_wait_ctx) {
1644 req->rq_status = -ENOTCONN;
1645 ptlrpc_rqphase_move(req,
1646 RQ_PHASE_INTERPRET);
1647 cfs_spin_unlock(&imp->imp_lock);
1648 GOTO(interpret, req->rq_status);
1651 cfs_list_del_init(&req->rq_list);
1652 cfs_list_add_tail(&req->rq_list,
1653 &imp->imp_sending_list);
1655 cfs_spin_unlock(&imp->imp_lock);
1657 cfs_spin_lock(&req->rq_lock);
1658 req->rq_waiting = 0;
1659 cfs_spin_unlock(&req->rq_lock);
1661 if (req->rq_timedout || req->rq_resend) {
1662 /* This is re-sending anyways,
1663 * let's mark req as resend. */
1664 cfs_spin_lock(&req->rq_lock);
1666 cfs_spin_unlock(&req->rq_lock);
1670 if (!ptlrpc_unregister_bulk(req, 1))
1673 /* ensure previous bulk fails */
1674 old_xid = req->rq_xid;
1675 req->rq_xid = ptlrpc_next_xid();
1676 CDEBUG(D_HA, "resend bulk "
1679 old_xid, req->rq_xid);
1683 * rq_wait_ctx is only touched by ptlrpcd,
1684 * so no lock is needed here.
1686 status = sptlrpc_req_refresh_ctx(req, -1);
1689 req->rq_status = status;
1690 cfs_spin_lock(&req->rq_lock);
1691 req->rq_wait_ctx = 0;
1692 cfs_spin_unlock(&req->rq_lock);
1693 force_timer_recalc = 1;
1695 cfs_spin_lock(&req->rq_lock);
1696 req->rq_wait_ctx = 1;
1697 cfs_spin_unlock(&req->rq_lock);
1702 cfs_spin_lock(&req->rq_lock);
1703 req->rq_wait_ctx = 0;
1704 cfs_spin_unlock(&req->rq_lock);
1707 rc = ptl_send_rpc(req, 0);
1709 DEBUG_REQ(D_HA, req, "send failed (%d)",
1711 force_timer_recalc = 1;
1712 cfs_spin_lock(&req->rq_lock);
1713 req->rq_net_err = 1;
1714 cfs_spin_unlock(&req->rq_lock);
1716 /* need to reset the timeout */
1717 force_timer_recalc = 1;
1720 cfs_spin_lock(&req->rq_lock);
1722 if (ptlrpc_client_early(req)) {
1723 ptlrpc_at_recv_early_reply(req);
1724 cfs_spin_unlock(&req->rq_lock);
1728 /* Still waiting for a reply? */
1729 if (ptlrpc_client_recv(req)) {
1730 cfs_spin_unlock(&req->rq_lock);
1734 /* Did we actually receive a reply? */
1735 if (!ptlrpc_client_replied(req)) {
1736 cfs_spin_unlock(&req->rq_lock);
1740 cfs_spin_unlock(&req->rq_lock);
1742 /* unlink from net because we are going to
1743 * swab in-place of reply buffer */
1744 unregistered = ptlrpc_unregister_reply(req, 1);
1748 req->rq_status = after_reply(req);
1752 /* If there is no bulk associated with this request,
1753 * then we're done and should let the interpreter
1754 * process the reply. Similarly if the RPC returned
1755 * an error, and therefore the bulk will never arrive.
1757 if (req->rq_bulk == NULL || req->rq_status < 0) {
1758 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1759 GOTO(interpret, req->rq_status);
1762 ptlrpc_rqphase_move(req, RQ_PHASE_BULK);
1765 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1766 if (ptlrpc_client_bulk_active(req))
1769 if (!req->rq_bulk->bd_success) {
1770 /* The RPC reply arrived OK, but the bulk screwed
1771 * up! Dead weird since the server told us the RPC
1772 * was good after getting the REPLY for her GET or
1773 * the ACK for her PUT. */
1774 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1775 req->rq_status = -EIO;
1778 ptlrpc_rqphase_move(req, RQ_PHASE_INTERPRET);
1781 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1783 /* This moves to "unregistering" phase we need to wait for
1785 if (!unregistered && !ptlrpc_unregister_reply(req, 1)) {
1786 /* start async bulk unlink too */
1787 ptlrpc_unregister_bulk(req, 1);
1791 if (!ptlrpc_unregister_bulk(req, 1))
1794 /* When calling interpret receiving already should be
1796 LASSERT(!req->rq_receiving_reply);
1798 ptlrpc_req_interpret(env, req, req->rq_status);
1800 ptlrpc_rqphase_move(req, RQ_PHASE_COMPLETE);
1802 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1803 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1804 imp->imp_obd->obd_uuid.uuid,
1805 req->rq_reqmsg ? lustre_msg_get_status(req->rq_reqmsg):-1,
1807 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1808 req->rq_reqmsg ? lustre_msg_get_opc(req->rq_reqmsg) : -1);
1810 cfs_spin_lock(&imp->imp_lock);
1811 /* Request already may be not on sending or delaying list. This
1812 * may happen in the case of marking it erroneous for the case
1813 * ptlrpc_import_delay_req(req, status) find it impossible to
1814 * allow sending this rpc and returns *status != 0. */
1815 if (!cfs_list_empty(&req->rq_list)) {
1816 cfs_list_del_init(&req->rq_list);
1817 cfs_atomic_dec(&imp->imp_inflight);
1819 cfs_spin_unlock(&imp->imp_lock);
1821 cfs_atomic_dec(&set->set_remaining);
1822 cfs_waitq_broadcast(&imp->imp_recovery_waitq);
1824 if (set->set_producer) {
1825 /* produce a new request if possible */
1826 if (ptlrpc_set_producer(set) > 0)
1827 force_timer_recalc = 1;
1829 /* free the request that has just been completed
1830 * in order not to pollute set->set_requests */
1831 cfs_list_del_init(&req->rq_set_chain);
1832 cfs_spin_lock(&req->rq_lock);
1834 req->rq_invalid_rqset = 0;
1835 cfs_spin_unlock(&req->rq_lock);
1837 /* record rq_status to compute the final status later */
1838 if (req->rq_status != 0)
1839 set->set_rc = req->rq_status;
1840 ptlrpc_req_finished(req);
1844 /* If we hit an error, we want to recover promptly. */
1845 RETURN(cfs_atomic_read(&set->set_remaining) == 0 || force_timer_recalc);
1849 * Time out request \a req. is \a async_unlink is set, that means do not wait
1850 * until LNet actually confirms network buffer unlinking.
1851 * Return 1 if we should give up further retrying attempts or 0 otherwise.
1853 int ptlrpc_expire_one_request(struct ptlrpc_request *req, int async_unlink)
1855 struct obd_import *imp = req->rq_import;
1859 cfs_spin_lock(&req->rq_lock);
1860 req->rq_timedout = 1;
1861 cfs_spin_unlock(&req->rq_lock);
1863 DEBUG_REQ(req->rq_fake ? D_INFO : D_WARNING, req, "Request "
1864 " sent has %s: [sent "CFS_DURATION_T"/"
1865 "real "CFS_DURATION_T"]",
1866 req->rq_net_err ? "failed due to network error" :
1867 ((req->rq_real_sent == 0 ||
1868 cfs_time_before(req->rq_real_sent, req->rq_sent) ||
1869 cfs_time_aftereq(req->rq_real_sent, req->rq_deadline)) ?
1870 "timed out for sent delay" : "timed out for slow reply"),
1871 req->rq_sent, req->rq_real_sent);
1873 if (imp != NULL && obd_debug_peer_on_timeout)
1874 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1876 ptlrpc_unregister_reply(req, async_unlink);
1877 ptlrpc_unregister_bulk(req, async_unlink);
1879 if (obd_dump_on_timeout)
1880 libcfs_debug_dumplog();
1883 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1890 cfs_atomic_inc(&imp->imp_timeouts);
1892 /* The DLM server doesn't want recovery run on its imports. */
1893 if (imp->imp_dlm_fake)
1896 /* If this request is for recovery or other primordial tasks,
1897 * then error it out here. */
1898 if (req->rq_ctx_init || req->rq_ctx_fini ||
1899 req->rq_send_state != LUSTRE_IMP_FULL ||
1900 imp->imp_obd->obd_no_recov) {
1901 DEBUG_REQ(D_RPCTRACE, req, "err -110, sent_state=%s (now=%s)",
1902 ptlrpc_import_state_name(req->rq_send_state),
1903 ptlrpc_import_state_name(imp->imp_state));
1904 cfs_spin_lock(&req->rq_lock);
1905 req->rq_status = -ETIMEDOUT;
1907 cfs_spin_unlock(&req->rq_lock);
1911 /* if a request can't be resent we can't wait for an answer after
1913 if (ptlrpc_no_resend(req)) {
1914 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1918 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1924 * Time out all uncompleted requests in request set pointed by \a data
1925 * Callback used when waiting on sets with l_wait_event.
1928 int ptlrpc_expired_set(void *data)
1930 struct ptlrpc_request_set *set = data;
1932 time_t now = cfs_time_current_sec();
1935 LASSERT(set != NULL);
1938 * A timeout expired. See which reqs it applies to...
1940 cfs_list_for_each (tmp, &set->set_requests) {
1941 struct ptlrpc_request *req =
1942 cfs_list_entry(tmp, struct ptlrpc_request,
1945 /* don't expire request waiting for context */
1946 if (req->rq_wait_ctx)
1949 /* Request in-flight? */
1950 if (!((req->rq_phase == RQ_PHASE_RPC &&
1951 !req->rq_waiting && !req->rq_resend) ||
1952 (req->rq_phase == RQ_PHASE_BULK)))
1955 if (req->rq_timedout || /* already dealt with */
1956 req->rq_deadline > now) /* not expired */
1959 /* Deal with this guy. Do it asynchronously to not block
1960 * ptlrpcd thread. */
1961 ptlrpc_expire_one_request(req, 1);
1965 * When waiting for a whole set, we always break out of the
1966 * sleep so we can recalculate the timeout, or enable interrupts
1967 * if everyone's timed out.
1973 * Sets rq_intr flag in \a req under spinlock.
1975 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1977 cfs_spin_lock(&req->rq_lock);
1979 cfs_spin_unlock(&req->rq_lock);
1983 * Interrupts (sets interrupted flag) all uncompleted requests in
1984 * a set \a data. Callback for l_wait_event for interruptible waits.
1986 void ptlrpc_interrupted_set(void *data)
1988 struct ptlrpc_request_set *set = data;
1991 LASSERT(set != NULL);
1992 CDEBUG(D_RPCTRACE, "INTERRUPTED SET %p\n", set);
1994 cfs_list_for_each(tmp, &set->set_requests) {
1995 struct ptlrpc_request *req =
1996 cfs_list_entry(tmp, struct ptlrpc_request,
1999 if (req->rq_phase != RQ_PHASE_RPC &&
2000 req->rq_phase != RQ_PHASE_UNREGISTERING)
2003 ptlrpc_mark_interrupted(req);
2008 * Get the smallest timeout in the set; this does NOT set a timeout.
2010 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
2013 time_t now = cfs_time_current_sec();
2015 struct ptlrpc_request *req;
2019 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
2021 cfs_list_for_each(tmp, &set->set_requests) {
2022 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2025 * Request in-flight?
2027 if (!(((req->rq_phase == RQ_PHASE_RPC) && !req->rq_waiting) ||
2028 (req->rq_phase == RQ_PHASE_BULK) ||
2029 (req->rq_phase == RQ_PHASE_NEW)))
2033 * Already timed out.
2035 if (req->rq_timedout)
2041 if (req->rq_wait_ctx)
2044 if (req->rq_phase == RQ_PHASE_NEW)
2045 deadline = req->rq_sent;
2047 deadline = req->rq_sent + req->rq_timeout;
2049 if (deadline <= now) /* actually expired already */
2050 timeout = 1; /* ASAP */
2051 else if (timeout == 0 || timeout > deadline - now)
2052 timeout = deadline - now;
2058 * Send all unset request from the set and then wait untill all
2059 * requests in the set complete (either get a reply, timeout, get an
2060 * error or otherwise be interrupted).
2061 * Returns 0 on success or error code otherwise.
2063 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
2066 struct ptlrpc_request *req;
2067 struct l_wait_info lwi;
2071 if (set->set_producer)
2072 (void)ptlrpc_set_producer(set);
2074 cfs_list_for_each(tmp, &set->set_requests) {
2075 req = cfs_list_entry(tmp, struct ptlrpc_request,
2077 if (req->rq_phase == RQ_PHASE_NEW)
2078 (void)ptlrpc_send_new_req(req);
2081 if (cfs_list_empty(&set->set_requests))
2085 timeout = ptlrpc_set_next_timeout(set);
2087 /* wait until all complete, interrupted, or an in-flight
2089 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
2092 if (timeout == 0 && !cfs_signal_pending())
2094 * No requests are in-flight (ether timed out
2095 * or delayed), so we can allow interrupts.
2096 * We still want to block for a limited time,
2097 * so we allow interrupts during the timeout.
2099 lwi = LWI_TIMEOUT_INTR_ALL(cfs_time_seconds(1),
2101 ptlrpc_interrupted_set, set);
2104 * At least one request is in flight, so no
2105 * interrupts are allowed. Wait until all
2106 * complete, or an in-flight req times out.
2108 lwi = LWI_TIMEOUT(cfs_time_seconds(timeout? timeout : 1),
2109 ptlrpc_expired_set, set);
2111 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(NULL, set), &lwi);
2113 /* LU-769 - if we ignored the signal because it was already
2114 * pending when we started, we need to handle it now or we risk
2115 * it being ignored forever */
2116 if (rc == -ETIMEDOUT && !lwi.lwi_allow_intr &&
2117 cfs_signal_pending()) {
2118 cfs_sigset_t blocked_sigs =
2119 cfs_block_sigsinv(LUSTRE_FATAL_SIGS);
2121 /* In fact we only interrupt for the "fatal" signals
2122 * like SIGINT or SIGKILL. We still ignore less
2123 * important signals since ptlrpc set is not easily
2124 * reentrant from userspace again */
2125 if (cfs_signal_pending())
2126 ptlrpc_interrupted_set(set);
2127 cfs_restore_sigs(blocked_sigs);
2130 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
2132 /* -EINTR => all requests have been flagged rq_intr so next
2134 * -ETIMEDOUT => someone timed out. When all reqs have
2135 * timed out, signals are enabled allowing completion with
2137 * I don't really care if we go once more round the loop in
2138 * the error cases -eeb. */
2139 if (rc == 0 && cfs_atomic_read(&set->set_remaining) == 0) {
2140 cfs_list_for_each(tmp, &set->set_requests) {
2141 req = cfs_list_entry(tmp, struct ptlrpc_request,
2143 cfs_spin_lock(&req->rq_lock);
2144 req->rq_invalid_rqset = 1;
2145 cfs_spin_unlock(&req->rq_lock);
2148 } while (rc != 0 || cfs_atomic_read(&set->set_remaining) != 0);
2150 LASSERT(cfs_atomic_read(&set->set_remaining) == 0);
2152 rc = set->set_rc; /* rq_status of already freed requests if any */
2153 cfs_list_for_each(tmp, &set->set_requests) {
2154 req = cfs_list_entry(tmp, struct ptlrpc_request, rq_set_chain);
2156 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
2157 if (req->rq_status != 0)
2158 rc = req->rq_status;
2161 if (set->set_interpret != NULL) {
2162 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
2164 rc = interpreter (set, set->set_arg, rc);
2166 struct ptlrpc_set_cbdata *cbdata, *n;
2169 cfs_list_for_each_entry_safe(cbdata, n,
2170 &set->set_cblist, psc_item) {
2171 cfs_list_del_init(&cbdata->psc_item);
2172 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
2175 OBD_FREE_PTR(cbdata);
2183 * Helper fuction for request freeing.
2184 * Called when request count reached zero and request needs to be freed.
2185 * Removes request from all sorts of sending/replay lists it might be on,
2186 * frees network buffers if any are present.
2187 * If \a locked is set, that means caller is already holding import imp_lock
2188 * and so we no longer need to reobtain it (for certain lists manipulations)
2190 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
2193 if (request == NULL) {
2198 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
2199 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
2200 LASSERTF(cfs_list_empty(&request->rq_list), "req %p\n", request);
2201 LASSERTF(cfs_list_empty(&request->rq_set_chain), "req %p\n", request);
2202 LASSERTF(cfs_list_empty(&request->rq_exp_list), "req %p\n", request);
2203 LASSERTF(!request->rq_replay, "req %p\n", request);
2205 req_capsule_fini(&request->rq_pill);
2207 /* We must take it off the imp_replay_list first. Otherwise, we'll set
2208 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
2209 if (request->rq_import != NULL) {
2211 cfs_spin_lock(&request->rq_import->imp_lock);
2212 cfs_list_del_init(&request->rq_replay_list);
2214 cfs_spin_unlock(&request->rq_import->imp_lock);
2216 LASSERTF(cfs_list_empty(&request->rq_replay_list), "req %p\n", request);
2218 if (cfs_atomic_read(&request->rq_refcount) != 0) {
2219 DEBUG_REQ(D_ERROR, request,
2220 "freeing request with nonzero refcount");
2224 if (request->rq_repbuf != NULL)
2225 sptlrpc_cli_free_repbuf(request);
2226 if (request->rq_export != NULL) {
2227 class_export_put(request->rq_export);
2228 request->rq_export = NULL;
2230 if (request->rq_import != NULL) {
2231 class_import_put(request->rq_import);
2232 request->rq_import = NULL;
2234 if (request->rq_bulk != NULL)
2235 ptlrpc_free_bulk(request->rq_bulk);
2237 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
2238 sptlrpc_cli_free_reqbuf(request);
2240 if (request->rq_cli_ctx)
2241 sptlrpc_req_put_ctx(request, !locked);
2243 if (request->rq_pool)
2244 __ptlrpc_free_req_to_pool(request);
2246 OBD_FREE(request, sizeof(*request));
2250 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
2252 * Drop one request reference. Must be called with import imp_lock held.
2253 * When reference count drops to zero, reuqest is freed.
2255 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
2257 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
2258 (void)__ptlrpc_req_finished(request, 1);
2263 * Drops one reference count for request \a request.
2264 * \a locked set indicates that caller holds import imp_lock.
2265 * Frees the request whe reference count reaches zero.
2267 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
2270 if (request == NULL)
2273 if (request == LP_POISON ||
2274 request->rq_reqmsg == LP_POISON) {
2275 CERROR("dereferencing freed request (bug 575)\n");
2280 DEBUG_REQ(D_INFO, request, "refcount now %u",
2281 cfs_atomic_read(&request->rq_refcount) - 1);
2283 if (cfs_atomic_dec_and_test(&request->rq_refcount)) {
2284 __ptlrpc_free_req(request, locked);
2292 * Drops one reference count for a request.
2294 void ptlrpc_req_finished(struct ptlrpc_request *request)
2296 __ptlrpc_req_finished(request, 0);
2300 * Returns xid of a \a request
2302 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
2304 return request->rq_xid;
2306 EXPORT_SYMBOL(ptlrpc_req_xid);
2309 * Disengage the client's reply buffer from the network
2310 * NB does _NOT_ unregister any client-side bulk.
2311 * IDEMPOTENT, but _not_ safe against concurrent callers.
2312 * The request owner (i.e. the thread doing the I/O) must call...
2313 * Returns 0 on success or 1 if unregistering cannot be made.
2315 int ptlrpc_unregister_reply(struct ptlrpc_request *request, int async)
2319 struct l_wait_info lwi;
2324 LASSERT(!cfs_in_interrupt());
2327 * Let's setup deadline for reply unlink.
2329 if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK) &&
2330 async && request->rq_reply_deadline == 0)
2331 request->rq_reply_deadline = cfs_time_current_sec()+LONG_UNLINK;
2334 * Nothing left to do.
2336 if (!ptlrpc_client_recv_or_unlink(request))
2339 LNetMDUnlink(request->rq_reply_md_h);
2342 * Let's check it once again.
2344 if (!ptlrpc_client_recv_or_unlink(request))
2348 * Move to "Unregistering" phase as reply was not unlinked yet.
2350 ptlrpc_rqphase_move(request, RQ_PHASE_UNREGISTERING);
2353 * Do not wait for unlink to finish.
2359 * We have to l_wait_event() whatever the result, to give liblustre
2360 * a chance to run reply_in_callback(), and to make sure we've
2361 * unlinked before returning a req to the pool.
2363 if (request->rq_set != NULL)
2364 wq = &request->rq_set->set_waitq;
2366 wq = &request->rq_reply_waitq;
2369 /* Network access will complete in finite time but the HUGE
2370 * timeout lets us CWARN for visibility of sluggish NALs */
2371 lwi = LWI_TIMEOUT_INTERVAL(cfs_time_seconds(LONG_UNLINK),
2372 cfs_time_seconds(1), NULL, NULL);
2373 rc = l_wait_event(*wq, !ptlrpc_client_recv_or_unlink(request),
2376 ptlrpc_rqphase_move(request, request->rq_next_phase);
2380 LASSERT(rc == -ETIMEDOUT);
2381 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout "
2382 "rvcng=%d unlnk=%d", request->rq_receiving_reply,
2383 request->rq_must_unlink);
2389 * Iterates through replay_list on import and prunes
2390 * all requests have transno smaller than last_committed for the
2391 * import and don't have rq_replay set.
2392 * Since requests are sorted in transno order, stops when meetign first
2393 * transno bigger than last_committed.
2394 * caller must hold imp->imp_lock
2396 void ptlrpc_free_committed(struct obd_import *imp)
2398 cfs_list_t *tmp, *saved;
2399 struct ptlrpc_request *req;
2400 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
2403 LASSERT(imp != NULL);
2405 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2408 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
2409 imp->imp_generation == imp->imp_last_generation_checked) {
2410 CDEBUG(D_INFO, "%s: skip recheck: last_committed "LPU64"\n",
2411 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
2415 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
2416 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
2417 imp->imp_generation);
2418 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
2419 imp->imp_last_generation_checked = imp->imp_generation;
2421 cfs_list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
2422 req = cfs_list_entry(tmp, struct ptlrpc_request,
2425 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
2426 LASSERT(req != last_req);
2429 if (req->rq_transno == 0) {
2430 DEBUG_REQ(D_EMERG, req, "zero transno during replay");
2433 if (req->rq_import_generation < imp->imp_generation) {
2434 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
2438 if (req->rq_replay) {
2439 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
2443 /* not yet committed */
2444 if (req->rq_transno > imp->imp_peer_committed_transno) {
2445 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
2449 DEBUG_REQ(D_INFO, req, "commit (last_committed "LPU64")",
2450 imp->imp_peer_committed_transno);
2452 cfs_spin_lock(&req->rq_lock);
2454 cfs_spin_unlock(&req->rq_lock);
2455 if (req->rq_commit_cb != NULL)
2456 req->rq_commit_cb(req);
2457 cfs_list_del_init(&req->rq_replay_list);
2458 __ptlrpc_req_finished(req, 1);
2465 void ptlrpc_cleanup_client(struct obd_import *imp)
2473 * Schedule previously sent request for resend.
2474 * For bulk requests we assign new xid (to avoid problems with
2475 * lost replies and therefore several transfers landing into same buffer
2476 * from different sending attempts).
2478 void ptlrpc_resend_req(struct ptlrpc_request *req)
2480 DEBUG_REQ(D_HA, req, "going to resend");
2481 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
2482 req->rq_status = -EAGAIN;
2484 cfs_spin_lock(&req->rq_lock);
2486 req->rq_net_err = 0;
2487 req->rq_timedout = 0;
2489 __u64 old_xid = req->rq_xid;
2491 /* ensure previous bulk fails */
2492 req->rq_xid = ptlrpc_next_xid();
2493 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
2494 old_xid, req->rq_xid);
2496 ptlrpc_client_wake_req(req);
2497 cfs_spin_unlock(&req->rq_lock);
2500 /* XXX: this function and rq_status are currently unused */
2501 void ptlrpc_restart_req(struct ptlrpc_request *req)
2503 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
2504 req->rq_status = -ERESTARTSYS;
2506 cfs_spin_lock(&req->rq_lock);
2507 req->rq_restart = 1;
2508 req->rq_timedout = 0;
2509 ptlrpc_client_wake_req(req);
2510 cfs_spin_unlock(&req->rq_lock);
2514 * Grab additional reference on a request \a req
2516 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
2519 cfs_atomic_inc(&req->rq_refcount);
2524 * Add a request to import replay_list.
2525 * Must be called under imp_lock
2527 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
2528 struct obd_import *imp)
2532 LASSERT_SPIN_LOCKED(&imp->imp_lock);
2534 if (req->rq_transno == 0) {
2535 DEBUG_REQ(D_EMERG, req, "saving request with zero transno");
2539 /* clear this for new requests that were resent as well
2540 as resent replayed requests. */
2541 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
2543 /* don't re-add requests that have been replayed */
2544 if (!cfs_list_empty(&req->rq_replay_list))
2547 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
2549 LASSERT(imp->imp_replayable);
2550 /* Balanced in ptlrpc_free_committed, usually. */
2551 ptlrpc_request_addref(req);
2552 cfs_list_for_each_prev(tmp, &imp->imp_replay_list) {
2553 struct ptlrpc_request *iter =
2554 cfs_list_entry(tmp, struct ptlrpc_request,
2557 /* We may have duplicate transnos if we create and then
2558 * open a file, or for closes retained if to match creating
2559 * opens, so use req->rq_xid as a secondary key.
2560 * (See bugs 684, 685, and 428.)
2561 * XXX no longer needed, but all opens need transnos!
2563 if (iter->rq_transno > req->rq_transno)
2566 if (iter->rq_transno == req->rq_transno) {
2567 LASSERT(iter->rq_xid != req->rq_xid);
2568 if (iter->rq_xid > req->rq_xid)
2572 cfs_list_add(&req->rq_replay_list, &iter->rq_replay_list);
2576 cfs_list_add(&req->rq_replay_list, &imp->imp_replay_list);
2580 * Send request and wait until it completes.
2581 * Returns request processing status.
2583 int ptlrpc_queue_wait(struct ptlrpc_request *req)
2585 struct ptlrpc_request_set *set;
2589 LASSERT(req->rq_set == NULL);
2590 LASSERT(!req->rq_receiving_reply);
2592 set = ptlrpc_prep_set();
2594 CERROR("Unable to allocate ptlrpc set.");
2598 /* for distributed debugging */
2599 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
2601 /* add a ref for the set (see comment in ptlrpc_set_add_req) */
2602 ptlrpc_request_addref(req);
2603 ptlrpc_set_add_req(set, req);
2604 rc = ptlrpc_set_wait(set);
2605 ptlrpc_set_destroy(set);
2610 struct ptlrpc_replay_async_args {
2612 int praa_old_status;
2616 * Callback used for replayed requests reply processing.
2617 * In case of succesful reply calls registeresd request replay callback.
2618 * In case of error restart replay process.
2620 static int ptlrpc_replay_interpret(const struct lu_env *env,
2621 struct ptlrpc_request *req,
2622 void * data, int rc)
2624 struct ptlrpc_replay_async_args *aa = data;
2625 struct obd_import *imp = req->rq_import;
2628 cfs_atomic_dec(&imp->imp_replay_inflight);
2630 if (!ptlrpc_client_replied(req)) {
2631 CERROR("request replay timed out, restarting recovery\n");
2632 GOTO(out, rc = -ETIMEDOUT);
2635 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
2636 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
2637 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
2638 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
2640 /** VBR: check version failure */
2641 if (lustre_msg_get_status(req->rq_repmsg) == -EOVERFLOW) {
2642 /** replay was failed due to version mismatch */
2643 DEBUG_REQ(D_WARNING, req, "Version mismatch during replay\n");
2644 cfs_spin_lock(&imp->imp_lock);
2645 imp->imp_vbr_failed = 1;
2646 imp->imp_no_lock_replay = 1;
2647 cfs_spin_unlock(&imp->imp_lock);
2648 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2650 /** The transno had better not change over replay. */
2651 LASSERTF(lustre_msg_get_transno(req->rq_reqmsg) ==
2652 lustre_msg_get_transno(req->rq_repmsg) ||
2653 lustre_msg_get_transno(req->rq_repmsg) == 0,
2655 lustre_msg_get_transno(req->rq_reqmsg),
2656 lustre_msg_get_transno(req->rq_repmsg));
2659 cfs_spin_lock(&imp->imp_lock);
2660 /** if replays by version then gap was occur on server, no trust to locks */
2661 if (lustre_msg_get_flags(req->rq_repmsg) & MSG_VERSION_REPLAY)
2662 imp->imp_no_lock_replay = 1;
2663 imp->imp_last_replay_transno = lustre_msg_get_transno(req->rq_reqmsg);
2664 cfs_spin_unlock(&imp->imp_lock);
2665 LASSERT(imp->imp_last_replay_transno);
2667 /* transaction number shouldn't be bigger than the latest replayed */
2668 if (req->rq_transno > lustre_msg_get_transno(req->rq_reqmsg)) {
2669 DEBUG_REQ(D_ERROR, req,
2670 "Reported transno "LPU64" is bigger than the "
2671 "replayed one: "LPU64, req->rq_transno,
2672 lustre_msg_get_transno(req->rq_reqmsg));
2673 GOTO(out, rc = -EINVAL);
2676 DEBUG_REQ(D_HA, req, "got rep");
2678 /* let the callback do fixups, possibly including in the request */
2679 if (req->rq_replay_cb)
2680 req->rq_replay_cb(req);
2682 if (ptlrpc_client_replied(req) &&
2683 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
2684 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
2685 lustre_msg_get_status(req->rq_repmsg),
2686 aa->praa_old_status);
2688 /* Put it back for re-replay. */
2689 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
2693 * Errors while replay can set transno to 0, but
2694 * imp_last_replay_transno shouldn't be set to 0 anyway
2696 if (req->rq_transno == 0)
2697 CERROR("Transno is 0 during replay!\n");
2699 /* continue with recovery */
2700 rc = ptlrpc_import_recovery_state_machine(imp);
2702 req->rq_send_state = aa->praa_old_state;
2705 /* this replay failed, so restart recovery */
2706 ptlrpc_connect_import(imp);
2712 * Prepares and queues request for replay.
2713 * Adds it to ptlrpcd queue for actual sending.
2714 * Returns 0 on success.
2716 int ptlrpc_replay_req(struct ptlrpc_request *req)
2718 struct ptlrpc_replay_async_args *aa;
2721 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2723 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2724 aa = ptlrpc_req_async_args(req);
2725 memset(aa, 0, sizeof *aa);
2727 /* Prepare request to be resent with ptlrpcd */
2728 aa->praa_old_state = req->rq_send_state;
2729 req->rq_send_state = LUSTRE_IMP_REPLAY;
2730 req->rq_phase = RQ_PHASE_NEW;
2731 req->rq_next_phase = RQ_PHASE_UNDEFINED;
2733 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2735 req->rq_interpret_reply = ptlrpc_replay_interpret;
2736 /* Readjust the timeout for current conditions */
2737 ptlrpc_at_set_req_timeout(req);
2739 /* Tell server the net_latency, so the server can calculate how long
2740 * it should wait for next replay */
2741 lustre_msg_set_service_time(req->rq_reqmsg,
2742 ptlrpc_at_get_net_latency(req));
2743 DEBUG_REQ(D_HA, req, "REPLAY");
2745 cfs_atomic_inc(&req->rq_import->imp_replay_inflight);
2746 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2748 ptlrpcd_add_req(req, PDL_POLICY_LOCAL, -1);
2753 * Aborts all in-flight request on import \a imp sending and delayed lists
2755 void ptlrpc_abort_inflight(struct obd_import *imp)
2757 cfs_list_t *tmp, *n;
2760 /* Make sure that no new requests get processed for this import.
2761 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2762 * this flag and then putting requests on sending_list or delayed_list.
2764 cfs_spin_lock(&imp->imp_lock);
2766 /* XXX locking? Maybe we should remove each request with the list
2767 * locked? Also, how do we know if the requests on the list are
2768 * being freed at this time?
2770 cfs_list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2771 struct ptlrpc_request *req =
2772 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2774 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2776 cfs_spin_lock (&req->rq_lock);
2777 if (req->rq_import_generation < imp->imp_generation) {
2779 req->rq_status = -EINTR;
2780 ptlrpc_client_wake_req(req);
2782 cfs_spin_unlock (&req->rq_lock);
2785 cfs_list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2786 struct ptlrpc_request *req =
2787 cfs_list_entry(tmp, struct ptlrpc_request, rq_list);
2789 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2791 cfs_spin_lock (&req->rq_lock);
2792 if (req->rq_import_generation < imp->imp_generation) {
2794 req->rq_status = -EINTR;
2795 ptlrpc_client_wake_req(req);
2797 cfs_spin_unlock (&req->rq_lock);
2800 /* Last chance to free reqs left on the replay list, but we
2801 * will still leak reqs that haven't committed. */
2802 if (imp->imp_replayable)
2803 ptlrpc_free_committed(imp);
2805 cfs_spin_unlock(&imp->imp_lock);
2811 * Abort all uncompleted requests in request set \a set
2813 void ptlrpc_abort_set(struct ptlrpc_request_set *set)
2815 cfs_list_t *tmp, *pos;
2817 LASSERT(set != NULL);
2819 cfs_list_for_each_safe(pos, tmp, &set->set_requests) {
2820 struct ptlrpc_request *req =
2821 cfs_list_entry(pos, struct ptlrpc_request,
2824 cfs_spin_lock(&req->rq_lock);
2825 if (req->rq_phase != RQ_PHASE_RPC) {
2826 cfs_spin_unlock(&req->rq_lock);
2831 req->rq_status = -EINTR;
2832 ptlrpc_client_wake_req(req);
2833 cfs_spin_unlock(&req->rq_lock);
2837 static __u64 ptlrpc_last_xid;
2838 static cfs_spinlock_t ptlrpc_last_xid_lock;
2841 * Initialize the XID for the node. This is common among all requests on
2842 * this node, and only requires the property that it is monotonically
2843 * increasing. It does not need to be sequential. Since this is also used
2844 * as the RDMA match bits, it is important that a single client NOT have
2845 * the same match bits for two different in-flight requests, hence we do
2846 * NOT want to have an XID per target or similar.
2848 * To avoid an unlikely collision between match bits after a client reboot
2849 * (which would deliver old data into the wrong RDMA buffer) initialize
2850 * the XID based on the current time, assuming a maximum RPC rate of 1M RPC/s.
2851 * If the time is clearly incorrect, we instead use a 62-bit random number.
2852 * In the worst case the random number will overflow 1M RPCs per second in
2853 * 9133 years, or permutations thereof.
2855 #define YEAR_2004 (1ULL << 30)
2856 void ptlrpc_init_xid(void)
2858 time_t now = cfs_time_current_sec();
2860 cfs_spin_lock_init(&ptlrpc_last_xid_lock);
2861 if (now < YEAR_2004) {
2862 cfs_get_random_bytes(&ptlrpc_last_xid, sizeof(ptlrpc_last_xid));
2863 ptlrpc_last_xid >>= 2;
2864 ptlrpc_last_xid |= (1ULL << 61);
2866 ptlrpc_last_xid = (__u64)now << 20;
2871 * Increase xid and returns resultng new value to the caller.
2873 __u64 ptlrpc_next_xid(void)
2876 cfs_spin_lock(&ptlrpc_last_xid_lock);
2877 tmp = ++ptlrpc_last_xid;
2878 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2883 * Get a glimpse at what next xid value might have been.
2884 * Returns possible next xid.
2886 __u64 ptlrpc_sample_next_xid(void)
2888 #if BITS_PER_LONG == 32
2889 /* need to avoid possible word tearing on 32-bit systems */
2891 cfs_spin_lock(&ptlrpc_last_xid_lock);
2892 tmp = ptlrpc_last_xid + 1;
2893 cfs_spin_unlock(&ptlrpc_last_xid_lock);
2896 /* No need to lock, since returned value is racy anyways */
2897 return ptlrpc_last_xid + 1;
2900 EXPORT_SYMBOL(ptlrpc_sample_next_xid);
2903 * Functions for operating ptlrpc workers.
2905 * A ptlrpc work is a function which will be running inside ptlrpc context.
2906 * The callback shouldn't sleep otherwise it will block that ptlrpcd thread.
2908 * 1. after a work is created, it can be used many times, that is:
2909 * handler = ptlrpcd_alloc_work();
2910 * ptlrpcd_queue_work();
2912 * queue it again when necessary:
2913 * ptlrpcd_queue_work();
2914 * ptlrpcd_destroy_work();
2915 * 2. ptlrpcd_queue_work() can be called by multiple processes meanwhile, but
2916 * it will only be queued once in any time. Also as its name implies, it may
2917 * have delay before it really runs by ptlrpcd thread.
2919 struct ptlrpc_work_async_args {
2921 int (*cb)(const struct lu_env *, void *);
2925 #define PTLRPC_WORK_MAGIC 0x6655436b676f4f44ULL /* magic code */
2927 static int work_interpreter(const struct lu_env *env,
2928 struct ptlrpc_request *req, void *data, int rc)
2930 struct ptlrpc_work_async_args *arg = data;
2932 LASSERT(arg->magic == PTLRPC_WORK_MAGIC);
2933 LASSERT(arg->cb != NULL);
2935 return arg->cb(env, arg->cbdata);
2939 * Create a work for ptlrpc.
2941 void *ptlrpcd_alloc_work(struct obd_import *imp,
2942 int (*cb)(const struct lu_env *, void *), void *cbdata)
2944 struct ptlrpc_request *req = NULL;
2945 struct ptlrpc_work_async_args *args;
2951 RETURN(ERR_PTR(-EINVAL));
2953 /* copy some code from deprecated fakereq. */
2956 CERROR("ptlrpc: run out of memory!\n");
2957 RETURN(ERR_PTR(-ENOMEM));
2960 req->rq_send_state = LUSTRE_IMP_FULL;
2961 req->rq_type = PTL_RPC_MSG_REQUEST;
2962 req->rq_import = class_import_get(imp);
2963 req->rq_export = NULL;
2964 req->rq_interpret_reply = work_interpreter;
2965 /* don't want reply */
2966 req->rq_receiving_reply = 0;
2967 req->rq_must_unlink = 0;
2968 req->rq_no_delay = req->rq_no_resend = 1;
2970 cfs_spin_lock_init(&req->rq_lock);
2971 CFS_INIT_LIST_HEAD(&req->rq_list);
2972 CFS_INIT_LIST_HEAD(&req->rq_replay_list);
2973 CFS_INIT_LIST_HEAD(&req->rq_set_chain);
2974 CFS_INIT_LIST_HEAD(&req->rq_history_list);
2975 CFS_INIT_LIST_HEAD(&req->rq_exp_list);
2976 cfs_waitq_init(&req->rq_reply_waitq);
2977 cfs_waitq_init(&req->rq_set_waitq);
2978 cfs_atomic_set(&req->rq_refcount, 1);
2980 CLASSERT (sizeof(*args) <= sizeof(req->rq_async_args));
2981 args = ptlrpc_req_async_args(req);
2982 args->magic = PTLRPC_WORK_MAGIC;
2984 args->cbdata = cbdata;
2988 EXPORT_SYMBOL(ptlrpcd_alloc_work);
2990 void ptlrpcd_destroy_work(void *handler)
2992 struct ptlrpc_request *req = handler;
2995 ptlrpc_req_finished(req);
2997 EXPORT_SYMBOL(ptlrpcd_destroy_work);
2999 int ptlrpcd_queue_work(void *handler)
3001 struct ptlrpc_request *req = handler;
3004 * Check if the req is already being queued.
3006 * Here comes a trick: it lacks a way of checking if a req is being
3007 * processed reliably in ptlrpc. Here I have to use refcount of req
3008 * for this purpose. This is okay because the caller should use this
3009 * req as opaque data. - Jinshan
3011 LASSERT(cfs_atomic_read(&req->rq_refcount) > 0);
3012 if (cfs_atomic_read(&req->rq_refcount) > 1)
3015 if (cfs_atomic_inc_return(&req->rq_refcount) > 2) { /* race */
3016 cfs_atomic_dec(&req->rq_refcount);
3020 /* re-initialize the req */
3021 req->rq_timeout = obd_timeout;
3022 req->rq_sent = cfs_time_current_sec();
3023 req->rq_deadline = req->rq_sent + req->rq_timeout;
3024 req->rq_reply_deadline = req->rq_deadline;
3025 req->rq_phase = RQ_PHASE_INTERPRET;
3026 req->rq_next_phase = RQ_PHASE_COMPLETE;
3027 req->rq_xid = ptlrpc_next_xid();
3028 req->rq_import_generation = req->rq_import->imp_generation;
3030 ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
3033 EXPORT_SYMBOL(ptlrpcd_queue_work);