1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
6 * This file is part of the Lustre file system, http://www.lustre.org
7 * Lustre is a trademark of Cluster File Systems, Inc.
9 * You may have signed or agreed to another license before downloading
10 * this software. If so, you are bound by the terms and conditions
11 * of that agreement, and the following does not apply to you. See the
12 * LICENSE file included with this distribution for more information.
14 * If you did not agree to a different license, then this copy of Lustre
15 * is open source software; you can redistribute it and/or modify it
16 * under the terms of version 2 of the GNU General Public License as
17 * published by the Free Software Foundation.
19 * In either case, Lustre is distributed in the hope that it will be
20 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
21 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * license text for more details.
26 #define DEBUG_SUBSYSTEM S_RPC
30 #include <liblustre.h>
33 #include <obd_support.h>
34 #include <obd_class.h>
35 #include <lustre_lib.h>
36 #include <lustre_ha.h>
37 #include <lustre_import.h>
39 #include "ptlrpc_internal.h"
41 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
42 struct ptlrpc_client *cl)
44 cl->cli_request_portal = req_portal;
45 cl->cli_reply_portal = rep_portal;
49 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
51 struct ptlrpc_connection *c;
53 lnet_process_id_t peer;
56 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
58 CERROR("cannot find peer %s!\n", uuid->uuid);
62 c = ptlrpc_get_connection(peer, self, uuid);
64 memcpy(c->c_remote_uuid.uuid,
65 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
68 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
73 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
74 struct obd_uuid *uuid)
77 lnet_process_id_t peer;
80 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
82 CERROR("cannot find peer %s!\n", uuid->uuid);
91 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
93 struct ptlrpc_bulk_desc *desc;
95 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
99 spin_lock_init(&desc->bd_lock);
100 cfs_waitq_init(&desc->bd_waitq);
101 desc->bd_max_iov = npages;
102 desc->bd_iov_count = 0;
103 desc->bd_md_h = LNET_INVALID_HANDLE;
104 desc->bd_portal = portal;
105 desc->bd_type = type;
110 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
111 int npages, int type, int portal)
113 struct obd_import *imp = req->rq_import;
114 struct ptlrpc_bulk_desc *desc;
117 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
118 desc = new_bulk(npages, type, portal);
122 desc->bd_import_generation = req->rq_import_generation;
123 desc->bd_import = class_import_get(imp);
126 desc->bd_cbid.cbid_fn = client_bulk_callback;
127 desc->bd_cbid.cbid_arg = desc;
129 /* This makes req own desc, and free it when she frees herself */
135 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp (struct ptlrpc_request *req,
136 int npages, int type, int portal)
138 struct obd_export *exp = req->rq_export;
139 struct ptlrpc_bulk_desc *desc;
142 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
144 desc = new_bulk(npages, type, portal);
148 desc->bd_export = class_export_get(exp);
151 desc->bd_cbid.cbid_fn = server_bulk_callback;
152 desc->bd_cbid.cbid_arg = desc;
154 /* NB we don't assign rq_bulk here; server-side requests are
155 * re-used, and the handler frees the bulk desc explicitly. */
160 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
161 cfs_page_t *page, int pageoffset, int len)
163 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
164 LASSERT(page != NULL);
165 LASSERT(pageoffset >= 0);
167 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
171 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
174 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
178 LASSERT(desc != NULL);
179 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
180 LASSERT(!desc->bd_network_rw); /* network hands off or */
181 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
183 sptlrpc_enc_pool_put_pages(desc);
186 class_export_put(desc->bd_export);
188 class_import_put(desc->bd_import);
190 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
191 bd_iov[desc->bd_max_iov]));
195 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
197 struct list_head *l, *tmp;
198 struct ptlrpc_request *req;
203 list_for_each_safe(l, tmp, &pool->prp_req_list) {
204 req = list_entry(l, struct ptlrpc_request, rq_list);
205 list_del(&req->rq_list);
206 LASSERT(req->rq_reqbuf);
207 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
208 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
209 OBD_FREE(req, sizeof(*req));
211 OBD_FREE(pool, sizeof(*pool));
214 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
219 while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
222 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
223 "Trying to change pool size with nonempty pool "
224 "from %d to %d bytes\n", pool->prp_rq_size, size);
226 spin_lock(&pool->prp_lock);
227 pool->prp_rq_size = size;
228 for (i = 0; i < num_rq; i++) {
229 struct ptlrpc_request *req;
230 struct lustre_msg *msg;
232 spin_unlock(&pool->prp_lock);
233 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
236 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
238 OBD_FREE(req, sizeof(struct ptlrpc_request));
241 req->rq_reqbuf = msg;
242 req->rq_reqbuf_len = size;
244 spin_lock(&pool->prp_lock);
245 list_add_tail(&req->rq_list, &pool->prp_req_list);
247 spin_unlock(&pool->prp_lock);
251 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
252 void (*populate_pool)(struct ptlrpc_request_pool *, int))
254 struct ptlrpc_request_pool *pool;
256 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
260 /* Request next power of two for the allocation, because internally
261 kernel would do exactly this */
263 spin_lock_init(&pool->prp_lock);
264 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
265 pool->prp_rq_size = msgsize;
266 pool->prp_populate = populate_pool;
268 populate_pool(pool, num_rq);
270 if (list_empty(&pool->prp_req_list)) {
271 /* have not allocated a single request for the pool */
272 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
278 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
280 struct ptlrpc_request *request;
281 struct lustre_msg *reqbuf;
286 spin_lock(&pool->prp_lock);
288 /* See if we have anything in a pool, and bail out if nothing,
289 * in writeout path, where this matters, this is safe to do, because
290 * nothing is lost in this case, and when some in-flight requests
291 * complete, this code will be called again. */
292 if (unlikely(list_empty(&pool->prp_req_list))) {
293 spin_unlock(&pool->prp_lock);
297 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
299 list_del(&request->rq_list);
300 spin_unlock(&pool->prp_lock);
302 LASSERT(request->rq_reqbuf);
303 LASSERT(request->rq_pool);
305 reqbuf = request->rq_reqbuf;
306 memset(request, 0, sizeof(*request));
307 request->rq_reqbuf = reqbuf;
308 request->rq_reqbuf_len = pool->prp_rq_size;
309 request->rq_pool = pool;
313 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
315 struct ptlrpc_request_pool *pool = request->rq_pool;
317 spin_lock(&pool->prp_lock);
318 LASSERT(list_empty(&request->rq_list));
319 list_add_tail(&request->rq_list, &pool->prp_req_list);
320 spin_unlock(&pool->prp_lock);
323 struct ptlrpc_request *
324 ptlrpc_prep_req_pool(struct obd_import *imp, __u32 version, int opcode,
325 int count, int *lengths, char **bufs,
326 struct ptlrpc_request_pool *pool,
327 struct ptlrpc_cli_ctx *ctx)
329 struct ptlrpc_request *request = NULL;
333 LASSERT((unsigned long)imp > 0x1000);
334 LASSERT(imp != LP_POISON);
335 LASSERT((unsigned long)imp->imp_client > 0x1000);
336 LASSERT(imp->imp_client != LP_POISON);
339 request = ptlrpc_prep_req_from_pool(pool);
342 OBD_ALLOC(request, sizeof(*request));
345 CERROR("request allocation out of memory\n");
349 request->rq_import = class_import_get(imp);
352 request->rq_cli_ctx = sptlrpc_ctx_get(ctx);
354 rc = sptlrpc_req_get_ctx(request);
359 sptlrpc_req_set_flavor(request, opcode);
361 rc = lustre_pack_request(request, imp->imp_msg_magic, count, lengths,
364 LASSERT(!request->rq_pool);
368 lustre_msg_add_version(request->rq_reqmsg, version);
370 if (imp->imp_server_timeout)
371 request->rq_timeout = obd_timeout / 2;
373 request->rq_timeout = obd_timeout;
374 request->rq_send_state = LUSTRE_IMP_FULL;
375 request->rq_type = PTL_RPC_MSG_REQUEST;
376 request->rq_export = NULL;
378 request->rq_req_cbid.cbid_fn = request_out_callback;
379 request->rq_req_cbid.cbid_arg = request;
381 request->rq_reply_cbid.cbid_fn = reply_in_callback;
382 request->rq_reply_cbid.cbid_arg = request;
384 request->rq_phase = RQ_PHASE_NEW;
386 /* XXX FIXME bug 249 */
387 request->rq_request_portal = imp->imp_client->cli_request_portal;
388 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
390 spin_lock_init(&request->rq_lock);
391 CFS_INIT_LIST_HEAD(&request->rq_list);
392 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
393 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
394 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
395 cfs_waitq_init(&request->rq_reply_waitq);
396 request->rq_xid = ptlrpc_next_xid();
397 atomic_set(&request->rq_refcount, 1);
399 lustre_msg_set_opc(request->rq_reqmsg, opcode);
400 lustre_msg_set_flags(request->rq_reqmsg, 0);
404 sptlrpc_req_put_ctx(request);
406 class_import_put(imp);
407 if (request->rq_pool)
408 __ptlrpc_free_req_to_pool(request);
410 OBD_FREE(request, sizeof(*request));
414 struct ptlrpc_request *
415 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
416 int *lengths, char **bufs)
418 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
422 struct ptlrpc_request_set *ptlrpc_prep_set(void)
424 struct ptlrpc_request_set *set;
427 OBD_ALLOC(set, sizeof *set);
430 CFS_INIT_LIST_HEAD(&set->set_requests);
431 cfs_waitq_init(&set->set_waitq);
432 set->set_remaining = 0;
433 spin_lock_init(&set->set_new_req_lock);
434 CFS_INIT_LIST_HEAD(&set->set_new_requests);
439 /* Finish with this set; opposite of prep_set. */
440 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
442 struct list_head *tmp;
443 struct list_head *next;
448 /* Requests on the set should either all be completed, or all be new */
449 expected_phase = (set->set_remaining == 0) ?
450 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
451 list_for_each (tmp, &set->set_requests) {
452 struct ptlrpc_request *req =
453 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
455 LASSERT(req->rq_phase == expected_phase);
459 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
461 list_for_each_safe(tmp, next, &set->set_requests) {
462 struct ptlrpc_request *req =
463 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
464 list_del_init(&req->rq_set_chain);
466 LASSERT(req->rq_phase == expected_phase);
468 if (req->rq_phase == RQ_PHASE_NEW) {
470 if (req->rq_interpret_reply != NULL) {
471 int (*interpreter)(struct ptlrpc_request *,
473 req->rq_interpret_reply;
475 /* higher level (i.e. LOV) failed;
476 * let the sub reqs clean up */
477 req->rq_status = -EBADR;
478 interpreter(req, &req->rq_async_args,
481 set->set_remaining--;
485 ptlrpc_req_finished (req);
488 LASSERT(set->set_remaining == 0);
490 OBD_FREE(set, sizeof(*set));
494 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
495 struct ptlrpc_request *req)
497 /* The set takes over the caller's request reference */
498 list_add_tail(&req->rq_set_chain, &set->set_requests);
500 set->set_remaining++;
502 atomic_inc(&req->rq_import->imp_inflight);
505 /* lock so many callers can add things, the context that owns the set
506 * is supposed to notice these and move them into the set proper. */
507 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
508 struct ptlrpc_request *req)
510 spin_lock(&set->set_new_req_lock);
511 /* The set takes over the caller's request reference */
512 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
514 spin_unlock(&set->set_new_req_lock);
518 * Based on the current state of the import, determine if the request
519 * can be sent, is an error, or should be delayed.
521 * Returns true if this request should be delayed. If false, and
522 * *status is set, then the request can not be sent and *status is the
523 * error code. If false and status is 0, then request can be sent.
525 * The imp->imp_lock must be held.
527 static int ptlrpc_import_delay_req(struct obd_import *imp,
528 struct ptlrpc_request *req, int *status)
533 LASSERT (status != NULL);
536 if (req->rq_ctx_init || req->rq_ctx_fini) {
537 /* always allow ctx init/fini rpc go through */
538 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
539 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
542 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
543 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
545 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
546 imp->imp_state == LUSTRE_IMP_CONNECTING) {
547 /* allow CONNECT even if import is invalid */ ;
548 } else if (imp->imp_invalid) {
549 /* If the import has been invalidated (such as by an OST
550 * failure), the request must fail with -EIO. */
551 if (!imp->imp_deactive)
552 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
554 } else if (req->rq_import_generation != imp->imp_generation) {
555 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
557 } else if (req->rq_send_state != imp->imp_state) {
558 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
560 *status = -EWOULDBLOCK;
568 static int ptlrpc_check_reply(struct ptlrpc_request *req)
573 /* serialise with network callback */
574 spin_lock(&req->rq_lock);
576 if (req->rq_replied) {
577 DEBUG_REQ(D_NET, req, "REPLIED:");
581 if (req->rq_net_err && !req->rq_timedout) {
582 spin_unlock(&req->rq_lock);
583 rc = ptlrpc_expire_one_request(req);
584 spin_lock(&req->rq_lock);
589 DEBUG_REQ(D_ERROR, req, "ABORTED:");
593 if (req->rq_resend) {
594 DEBUG_REQ(D_ERROR, req, "RESEND:");
598 if (req->rq_restart) {
599 DEBUG_REQ(D_ERROR, req, "RESTART:");
604 spin_unlock(&req->rq_lock);
605 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
609 static int ptlrpc_check_status(struct ptlrpc_request *req)
614 err = lustre_msg_get_status(req->rq_repmsg);
615 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
616 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err == %d",
618 RETURN(err < 0 ? err : -EINVAL);
622 DEBUG_REQ(D_INFO, req, "status is %d", err);
623 } else if (err > 0) {
624 /* XXX: translate this error from net to host */
625 DEBUG_REQ(D_INFO, req, "status is %d", err);
631 static int after_reply(struct ptlrpc_request *req)
633 struct obd_import *imp = req->rq_import;
637 LASSERT(!req->rq_receiving_reply);
638 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
640 /* NB Until this point, the whole of the incoming message,
641 * including buflens, status etc is in the sender's byte order. */
644 /* Clear reply swab mask; this is a new reply in sender's byte order */
645 req->rq_rep_swab_mask = 0;
647 rc = sptlrpc_cli_unwrap_reply(req);
649 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
653 /* security layer unwrap might ask resend this request */
657 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
659 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d\n", rc);
663 rc = lustre_unpack_ptlrpc_body(req->rq_repmsg);
665 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d\n", rc);
669 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
670 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
671 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)\n",
672 lustre_msg_get_type(req->rq_repmsg));
676 rc = ptlrpc_check_status(req);
678 /* Either we've been evicted, or the server has failed for
679 * some reason. Try to reconnect, and if that fails, punt to the
681 if ((rc == -ENOTCONN) || (rc == -ENODEV)) {
682 if (req->rq_send_state != LUSTRE_IMP_FULL ||
683 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
687 ptlrpc_request_handle_notconn(req);
692 /* Store transno in reqmsg for replay. */
693 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
694 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
696 if (req->rq_import->imp_replayable) {
697 spin_lock(&imp->imp_lock);
698 if (req->rq_transno != 0)
699 ptlrpc_retain_replayable_request(req, imp);
700 else if (req->rq_commit_cb != NULL) {
701 spin_unlock(&imp->imp_lock);
702 req->rq_commit_cb(req);
703 spin_lock(&imp->imp_lock);
706 /* Replay-enabled imports return commit-status information. */
707 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
708 imp->imp_peer_committed_transno =
709 lustre_msg_get_last_committed(req->rq_repmsg);
711 ptlrpc_free_committed(imp);
712 spin_unlock(&imp->imp_lock);
718 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
720 struct obd_import *imp;
724 LASSERT(req->rq_phase == RQ_PHASE_NEW);
725 req->rq_phase = RQ_PHASE_RPC;
727 imp = req->rq_import;
728 spin_lock(&imp->imp_lock);
730 req->rq_import_generation = imp->imp_generation;
732 if (ptlrpc_import_delay_req(imp, req, &rc)) {
733 spin_lock (&req->rq_lock);
735 spin_unlock (&req->rq_lock);
737 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
739 lustre_msg_get_status(req->rq_reqmsg) ,
740 ptlrpc_import_state_name(req->rq_send_state),
741 ptlrpc_import_state_name(imp->imp_state));
742 LASSERT(list_empty (&req->rq_list));
744 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
745 spin_unlock(&imp->imp_lock);
750 spin_unlock(&imp->imp_lock);
752 req->rq_phase = RQ_PHASE_INTERPRET;
756 /* XXX this is the same as ptlrpc_queue_wait */
757 LASSERT(list_empty(&req->rq_list));
758 list_add_tail(&req->rq_list, &imp->imp_sending_list);
759 spin_unlock(&imp->imp_lock);
761 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
763 rc = sptlrpc_req_refresh_ctx(req, -1);
769 /* here begins timeout counting */
770 req->rq_sent = CURRENT_SECONDS;
771 req->rq_wait_ctx = 1;
776 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
777 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
778 imp->imp_obd->obd_uuid.uuid,
779 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
780 libcfs_nid2str(imp->imp_connection->c_peer.nid),
781 lustre_msg_get_opc(req->rq_reqmsg));
783 rc = ptl_send_rpc(req, 0);
785 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
792 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
793 int ptlrpc_check_set(struct ptlrpc_request_set *set)
795 struct list_head *tmp;
796 int force_timer_recalc = 0;
799 if (set->set_remaining == 0)
802 list_for_each(tmp, &set->set_requests) {
803 struct ptlrpc_request *req =
804 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
805 struct obd_import *imp = req->rq_import;
808 if (req->rq_phase == RQ_PHASE_NEW &&
809 ptlrpc_send_new_req(req)) {
810 force_timer_recalc = 1;
813 if (!(req->rq_phase == RQ_PHASE_RPC ||
814 req->rq_phase == RQ_PHASE_BULK ||
815 req->rq_phase == RQ_PHASE_INTERPRET ||
816 req->rq_phase == RQ_PHASE_COMPLETE)) {
817 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
821 if (req->rq_phase == RQ_PHASE_COMPLETE)
824 if (req->rq_phase == RQ_PHASE_INTERPRET)
825 GOTO(interpret, req->rq_status);
827 if (req->rq_net_err && !req->rq_timedout)
828 ptlrpc_expire_one_request(req);
831 ptlrpc_unregister_reply(req);
832 if (req->rq_status == 0)
833 req->rq_status = -EIO;
834 req->rq_phase = RQ_PHASE_INTERPRET;
836 spin_lock(&imp->imp_lock);
837 list_del_init(&req->rq_list);
838 spin_unlock(&imp->imp_lock);
840 GOTO(interpret, req->rq_status);
843 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
844 * will only be set after rq_timedout, but the oig waiting
845 * path sets rq_intr irrespective of whether ptlrpcd has
846 * seen a timeout. our policy is to only interpret
847 * interrupted rpcs after they have timed out */
848 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
850 /* NB could be on delayed list */
851 ptlrpc_unregister_reply(req);
852 req->rq_status = -EINTR;
853 req->rq_phase = RQ_PHASE_INTERPRET;
855 spin_lock(&imp->imp_lock);
856 list_del_init(&req->rq_list);
857 spin_unlock(&imp->imp_lock);
859 GOTO(interpret, req->rq_status);
862 if (req->rq_phase == RQ_PHASE_RPC) {
863 if (req->rq_timedout || req->rq_resend ||
864 req->rq_waiting || req->rq_wait_ctx) {
867 /* rq_wait_ctx is only touched in ptlrpcd,
868 * no lock needed here.
870 if (req->rq_wait_ctx)
873 ptlrpc_unregister_reply(req);
875 spin_lock(&imp->imp_lock);
877 if (ptlrpc_import_delay_req(imp, req, &status)){
878 spin_unlock(&imp->imp_lock);
882 list_del_init(&req->rq_list);
884 req->rq_status = status;
885 req->rq_phase = RQ_PHASE_INTERPRET;
886 spin_unlock(&imp->imp_lock);
887 GOTO(interpret, req->rq_status);
889 if (req->rq_no_resend && !req->rq_wait_ctx) {
890 req->rq_status = -ENOTCONN;
891 req->rq_phase = RQ_PHASE_INTERPRET;
892 spin_unlock(&imp->imp_lock);
893 GOTO(interpret, req->rq_status);
895 list_add_tail(&req->rq_list,
896 &imp->imp_sending_list);
898 spin_unlock(&imp->imp_lock);
901 if (req->rq_resend) {
902 lustre_msg_add_flags(req->rq_reqmsg,
905 __u64 old_xid = req->rq_xid;
907 ptlrpc_unregister_bulk (req);
909 /* ensure previous bulk fails */
910 req->rq_xid = ptlrpc_next_xid();
911 CDEBUG(D_HA, "resend bulk "
914 old_xid, req->rq_xid);
918 status = sptlrpc_req_refresh_ctx(req, -1);
921 req->rq_status = status;
922 force_timer_recalc = 1;
924 if (!req->rq_wait_ctx) {
925 /* begins timeout counting */
926 req->rq_sent = CURRENT_SECONDS;
927 req->rq_wait_ctx = 1;
932 req->rq_wait_ctx = 0;
935 rc = ptl_send_rpc(req, 0);
937 DEBUG_REQ(D_HA, req, "send failed (%d)",
939 force_timer_recalc = 1;
942 /* need to reset the timeout */
943 force_timer_recalc = 1;
946 /* Still waiting for a reply? */
947 if (ptlrpc_client_receiving_reply(req))
950 /* Did we actually receive a reply? */
951 if (!ptlrpc_client_replied(req))
954 spin_lock(&imp->imp_lock);
955 list_del_init(&req->rq_list);
956 spin_unlock(&imp->imp_lock);
958 req->rq_status = after_reply(req);
959 if (req->rq_resend) {
960 /* Add this req to the delayed list so
961 it can be errored if the import is
962 evicted after recovery. */
963 spin_lock(&req->rq_lock);
964 list_add_tail(&req->rq_list,
965 &imp->imp_delayed_list);
966 spin_unlock(&req->rq_lock);
970 /* If there is no bulk associated with this request,
971 * then we're done and should let the interpreter
972 * process the reply. Similarly if the RPC returned
973 * an error, and therefore the bulk will never arrive.
975 if (req->rq_bulk == NULL || req->rq_status != 0) {
976 req->rq_phase = RQ_PHASE_INTERPRET;
977 GOTO(interpret, req->rq_status);
980 req->rq_phase = RQ_PHASE_BULK;
983 LASSERT(req->rq_phase == RQ_PHASE_BULK);
984 if (ptlrpc_bulk_active(req->rq_bulk))
987 if (!req->rq_bulk->bd_success) {
988 /* The RPC reply arrived OK, but the bulk screwed
989 * up! Dead wierd since the server told us the RPC
990 * was good after getting the REPLY for her GET or
991 * the ACK for her PUT. */
992 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
996 req->rq_phase = RQ_PHASE_INTERPRET;
999 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1000 LASSERT(!req->rq_receiving_reply);
1002 ptlrpc_unregister_reply(req);
1003 if (req->rq_bulk != NULL)
1004 ptlrpc_unregister_bulk (req);
1006 req->rq_phase = RQ_PHASE_COMPLETE;
1008 if (req->rq_interpret_reply != NULL) {
1009 int (*interpreter)(struct ptlrpc_request *,void *,int) =
1010 req->rq_interpret_reply;
1011 req->rq_status = interpreter(req, &req->rq_async_args,
1015 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1016 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1017 imp->imp_obd->obd_uuid.uuid,
1018 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1019 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1020 lustre_msg_get_opc(req->rq_reqmsg));
1022 set->set_remaining--;
1024 atomic_dec(&imp->imp_inflight);
1025 cfs_waitq_signal(&imp->imp_recovery_waitq);
1028 /* If we hit an error, we want to recover promptly. */
1029 RETURN(set->set_remaining == 0 || force_timer_recalc);
1032 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1034 struct obd_import *imp = req->rq_import;
1037 DEBUG_REQ(D_ERROR, req, "timeout (sent at %lu, %lus ago)",
1038 (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
1041 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1043 spin_lock(&req->rq_lock);
1044 req->rq_timedout = 1;
1045 req->rq_wait_ctx = 0;
1046 spin_unlock(&req->rq_lock);
1048 ptlrpc_unregister_reply (req);
1050 if (obd_dump_on_timeout)
1051 libcfs_debug_dumplog();
1053 if (req->rq_bulk != NULL)
1054 ptlrpc_unregister_bulk (req);
1057 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1061 /* The DLM server doesn't want recovery run on its imports. */
1062 if (imp->imp_dlm_fake)
1065 /* If this request is for recovery or other primordial tasks,
1066 * then error it out here. */
1067 if (req->rq_ctx_init || req->rq_ctx_fini ||
1068 req->rq_send_state != LUSTRE_IMP_FULL ||
1069 imp->imp_obd->obd_no_recov) {
1070 spin_lock(&req->rq_lock);
1071 req->rq_status = -ETIMEDOUT;
1073 spin_unlock(&req->rq_lock);
1077 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1082 int ptlrpc_expired_set(void *data)
1084 struct ptlrpc_request_set *set = data;
1085 struct list_head *tmp;
1086 time_t now = CURRENT_SECONDS;
1089 LASSERT(set != NULL);
1091 /* A timeout expired; see which reqs it applies to... */
1092 list_for_each (tmp, &set->set_requests) {
1093 struct ptlrpc_request *req =
1094 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1096 /* request in-flight? */
1097 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1099 (req->rq_phase == RQ_PHASE_BULK)))
1102 if (req->rq_timedout || /* already dealt with */
1103 req->rq_sent + req->rq_timeout > now) /* not expired */
1106 /* deal with this guy */
1107 ptlrpc_expire_one_request (req);
1110 /* When waiting for a whole set, we always to break out of the
1111 * sleep so we can recalculate the timeout, or enable interrupts
1112 * iff everyone's timed out.
1117 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1119 spin_lock(&req->rq_lock);
1121 spin_unlock(&req->rq_lock);
1124 void ptlrpc_interrupted_set(void *data)
1126 struct ptlrpc_request_set *set = data;
1127 struct list_head *tmp;
1129 LASSERT(set != NULL);
1130 CERROR("INTERRUPTED SET %p\n", set);
1132 list_for_each(tmp, &set->set_requests) {
1133 struct ptlrpc_request *req =
1134 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1136 if (req->rq_phase != RQ_PHASE_RPC)
1139 ptlrpc_mark_interrupted(req);
1143 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1145 struct list_head *tmp;
1146 time_t now = CURRENT_SECONDS;
1149 struct ptlrpc_request *req;
1152 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1154 list_for_each(tmp, &set->set_requests) {
1155 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1157 /* request in-flight? */
1158 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1159 (req->rq_phase == RQ_PHASE_BULK)))
1162 if (req->rq_timedout) /* already timed out */
1165 deadline = req->rq_sent + req->rq_timeout;
1166 if (deadline <= now) /* actually expired already */
1167 timeout = 1; /* ASAP */
1168 else if (timeout == 0 || timeout > deadline - now)
1169 timeout = deadline - now;
1174 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1176 struct list_head *tmp;
1177 struct ptlrpc_request *req;
1178 struct l_wait_info lwi;
1182 if (list_empty(&set->set_requests))
1185 list_for_each(tmp, &set->set_requests) {
1186 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1187 if (req->rq_phase == RQ_PHASE_NEW)
1188 (void)ptlrpc_send_new_req(req);
1192 timeout = ptlrpc_set_next_timeout(set);
1194 /* wait until all complete, interrupted, or an in-flight
1196 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
1198 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1200 ptlrpc_interrupted_set, set);
1201 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1203 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1205 /* -EINTR => all requests have been flagged rq_intr so next
1207 * -ETIMEOUTD => someone timed out. When all reqs have
1208 * timed out, signals are enabled allowing completion with
1210 * I don't really care if we go once more round the loop in
1211 * the error cases -eeb. */
1212 } while (rc != 0 || set->set_remaining != 0);
1214 LASSERT(set->set_remaining == 0);
1217 list_for_each(tmp, &set->set_requests) {
1218 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1220 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1221 if (req->rq_status != 0)
1222 rc = req->rq_status;
1225 if (set->set_interpret != NULL) {
1226 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1228 rc = interpreter (set, set->set_arg, rc);
1234 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1237 if (request == NULL) {
1242 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1243 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1244 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1245 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1246 LASSERT(request->rq_cli_ctx);
1248 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1249 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1250 if (request->rq_import != NULL) {
1252 spin_lock(&request->rq_import->imp_lock);
1253 list_del_init(&request->rq_replay_list);
1255 spin_unlock(&request->rq_import->imp_lock);
1257 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1259 if (atomic_read(&request->rq_refcount) != 0) {
1260 DEBUG_REQ(D_ERROR, request,
1261 "freeing request with nonzero refcount");
1265 if (request->rq_repbuf != NULL)
1266 sptlrpc_cli_free_repbuf(request);
1267 if (request->rq_export != NULL) {
1268 class_export_put(request->rq_export);
1269 request->rq_export = NULL;
1271 if (request->rq_import != NULL) {
1272 class_import_put(request->rq_import);
1273 request->rq_import = NULL;
1275 if (request->rq_bulk != NULL)
1276 ptlrpc_free_bulk(request->rq_bulk);
1278 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1279 sptlrpc_cli_free_reqbuf(request);
1281 sptlrpc_req_put_ctx(request);
1283 if (request->rq_pool)
1284 __ptlrpc_free_req_to_pool(request);
1286 OBD_FREE(request, sizeof(*request));
1290 void ptlrpc_free_req(struct ptlrpc_request *request)
1292 __ptlrpc_free_req(request, 0);
1295 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1296 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1298 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1299 (void)__ptlrpc_req_finished(request, 1);
1302 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1305 if (request == NULL)
1308 if (request == LP_POISON ||
1309 request->rq_reqmsg == LP_POISON) {
1310 CERROR("dereferencing freed request (bug 575)\n");
1315 DEBUG_REQ(D_INFO, request, "refcount now %u",
1316 atomic_read(&request->rq_refcount) - 1);
1318 if (atomic_dec_and_test(&request->rq_refcount)) {
1319 __ptlrpc_free_req(request, locked);
1326 void ptlrpc_req_finished(struct ptlrpc_request *request)
1328 __ptlrpc_req_finished(request, 0);
1331 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1333 return request->rq_xid;
1335 EXPORT_SYMBOL(ptlrpc_req_xid);
1337 /* Disengage the client's reply buffer from the network
1338 * NB does _NOT_ unregister any client-side bulk.
1339 * IDEMPOTENT, but _not_ safe against concurrent callers.
1340 * The request owner (i.e. the thread doing the I/O) must call...
1342 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1346 struct l_wait_info lwi;
1348 LASSERT(!in_interrupt ()); /* might sleep */
1350 if (!ptlrpc_client_receiving_reply(request))
1353 LNetMDUnlink (request->rq_reply_md_h);
1355 /* We have to l_wait_event() whatever the result, to give liblustre
1356 * a chance to run reply_in_callback() */
1358 if (request->rq_set != NULL)
1359 wq = &request->rq_set->set_waitq;
1361 wq = &request->rq_reply_waitq;
1364 /* Network access will complete in finite time but the HUGE
1365 * timeout lets us CWARN for visibility of sluggish NALs */
1366 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1367 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1371 LASSERT (rc == -ETIMEDOUT);
1372 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1376 /* caller must hold imp->imp_lock */
1377 void ptlrpc_free_committed(struct obd_import *imp)
1379 struct list_head *tmp, *saved;
1380 struct ptlrpc_request *req;
1381 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1384 LASSERT(imp != NULL);
1386 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1389 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1390 imp->imp_generation == imp->imp_last_generation_checked) {
1391 CDEBUG(D_HA, "%s: skip recheck for last_committed "LPU64"\n",
1392 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1396 CDEBUG(D_HA, "%s: committing for last_committed "LPU64" gen %d\n",
1397 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1398 imp->imp_generation);
1399 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1400 imp->imp_last_generation_checked = imp->imp_generation;
1402 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1403 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1405 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1406 LASSERT(req != last_req);
1408 /*XXX: Debug to see the corrupted request due to 11012 */
1409 CDEBUG(D_HA, "Next request: req@%p x"LPD64"/t"LPD64" o?->%s@%s:%d lens %d/%d ref %d, status %d\n",
1410 req, req->rq_xid, req->rq_transno, req->rq_import ? obd2cli_tgt(req->rq_import->imp_obd) :
1411 req->rq_export ? (char*)req->rq_export->exp_client_uuid.uuid : "<?>",
1412 req->rq_import ? (char *)req->rq_import->imp_connection->c_remote_uuid.uuid :
1413 req->rq_export ? (char *)req->rq_export->exp_connection->c_remote_uuid.uuid : "<?>",
1414 (req->rq_import && req->rq_import->imp_client) ?
1415 req->rq_import->imp_client->cli_request_portal : -1,
1416 req->rq_reqlen, req->rq_replen, atomic_read(&req->rq_refcount), req->rq_status);
1418 if (req->rq_import_generation < imp->imp_generation) {
1419 DEBUG_REQ(D_HA, req, "freeing request with old gen");
1423 if (req->rq_replay) {
1424 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1428 /* not yet committed */
1429 if (req->rq_transno > imp->imp_peer_committed_transno) {
1430 DEBUG_REQ(D_HA, req, "stopping search");
1434 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1435 imp->imp_peer_committed_transno);
1437 if (req->rq_commit_cb != NULL)
1438 req->rq_commit_cb(req);
1439 list_del_init(&req->rq_replay_list);
1440 __ptlrpc_req_finished(req, 1);
1447 void ptlrpc_cleanup_client(struct obd_import *imp)
1454 void ptlrpc_resend_req(struct ptlrpc_request *req)
1456 DEBUG_REQ(D_HA, req, "going to resend");
1457 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1458 req->rq_status = -EAGAIN;
1460 spin_lock(&req->rq_lock);
1462 req->rq_net_err = 0;
1463 req->rq_timedout = 0;
1465 __u64 old_xid = req->rq_xid;
1467 /* ensure previous bulk fails */
1468 req->rq_xid = ptlrpc_next_xid();
1469 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1470 old_xid, req->rq_xid);
1472 ptlrpc_wake_client_req(req);
1473 spin_unlock(&req->rq_lock);
1476 /* XXX: this function and rq_status are currently unused */
1477 void ptlrpc_restart_req(struct ptlrpc_request *req)
1479 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1480 req->rq_status = -ERESTARTSYS;
1482 spin_lock(&req->rq_lock);
1483 req->rq_restart = 1;
1484 req->rq_timedout = 0;
1485 ptlrpc_wake_client_req(req);
1486 spin_unlock(&req->rq_lock);
1489 static int expired_request(void *data)
1491 struct ptlrpc_request *req = data;
1494 RETURN(ptlrpc_expire_one_request(req));
1497 static void interrupted_request(void *data)
1499 struct ptlrpc_request *req = data;
1500 DEBUG_REQ(D_HA, req, "request interrupted");
1501 spin_lock(&req->rq_lock);
1503 spin_unlock(&req->rq_lock);
1506 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1509 atomic_inc(&req->rq_refcount);
1513 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1514 struct obd_import *imp)
1516 struct list_head *tmp;
1518 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1520 /* clear this for new requests that were resent as well
1521 as resent replayed requests. */
1522 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1524 /* don't re-add requests that have been replayed */
1525 if (!list_empty(&req->rq_replay_list))
1528 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1530 LASSERT(imp->imp_replayable);
1531 /* Balanced in ptlrpc_free_committed, usually. */
1532 ptlrpc_request_addref(req);
1533 list_for_each_prev(tmp, &imp->imp_replay_list) {
1534 struct ptlrpc_request *iter =
1535 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1537 /* We may have duplicate transnos if we create and then
1538 * open a file, or for closes retained if to match creating
1539 * opens, so use req->rq_xid as a secondary key.
1540 * (See bugs 684, 685, and 428.)
1541 * XXX no longer needed, but all opens need transnos!
1543 if (iter->rq_transno > req->rq_transno)
1546 if (iter->rq_transno == req->rq_transno) {
1547 LASSERT(iter->rq_xid != req->rq_xid);
1548 if (iter->rq_xid > req->rq_xid)
1552 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1556 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1559 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1563 struct l_wait_info lwi;
1564 struct obd_import *imp = req->rq_import;
1565 cfs_duration_t timeout = 0;
1568 LASSERT(req->rq_set == NULL);
1569 LASSERT(!req->rq_receiving_reply);
1570 atomic_inc(&imp->imp_inflight);
1572 /* for distributed debugging */
1573 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1574 LASSERT(imp->imp_obd != NULL);
1575 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1576 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1577 imp->imp_obd->obd_uuid.uuid,
1578 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1579 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1580 lustre_msg_get_opc(req->rq_reqmsg));
1582 /* Mark phase here for a little debug help */
1583 req->rq_phase = RQ_PHASE_RPC;
1585 spin_lock(&imp->imp_lock);
1586 req->rq_import_generation = imp->imp_generation;
1588 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1589 list_del(&req->rq_list);
1591 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1592 spin_unlock(&imp->imp_lock);
1594 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1596 ptlrpc_import_state_name(req->rq_send_state),
1597 ptlrpc_import_state_name(imp->imp_state));
1598 lwi = LWI_INTR(interrupted_request, req);
1599 rc = l_wait_event(req->rq_reply_waitq,
1600 (req->rq_send_state == imp->imp_state ||
1601 req->rq_err || req->rq_intr),
1603 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1605 ptlrpc_import_state_name(imp->imp_state),
1606 ptlrpc_import_state_name(req->rq_send_state),
1607 req->rq_err, req->rq_intr);
1609 spin_lock(&imp->imp_lock);
1610 list_del_init(&req->rq_list);
1615 else if (req->rq_intr) {
1618 else if (req->rq_no_resend) {
1619 spin_unlock(&imp->imp_lock);
1620 GOTO(out, rc = -ETIMEDOUT);
1628 list_del_init(&req->rq_list);
1629 spin_unlock(&imp->imp_lock);
1630 req->rq_status = rc; // XXX this ok?
1634 if (req->rq_resend) {
1635 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1637 if (req->rq_bulk != NULL) {
1638 ptlrpc_unregister_bulk (req);
1640 /* bulk requests are supposed to be
1641 * idempotent, so we are free to bump the xid
1642 * here, which we need to do before
1643 * registering the bulk again (bug 6371).
1644 * print the old xid first for sanity.
1646 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1647 req->rq_xid = ptlrpc_next_xid();
1650 DEBUG_REQ(D_HA, req, "resending: ");
1653 /* XXX this is the same as ptlrpc_set_wait */
1654 LASSERT(list_empty(&req->rq_list));
1655 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1656 spin_unlock(&imp->imp_lock);
1658 rc = sptlrpc_req_refresh_ctx(req, 0);
1661 /* we got fatal ctx refresh error, directly jump out
1662 * thus we can pass back the actual error code.
1664 spin_lock(&imp->imp_lock);
1665 list_del_init(&req->rq_list);
1666 spin_unlock(&imp->imp_lock);
1668 CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
1671 /* simulating we got error during send rpc */
1675 rc = ptl_send_rpc(req, 0);
1677 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1680 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1681 DEBUG_REQ(D_NET, req,
1682 "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1684 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1686 l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1687 DEBUG_REQ(D_NET, req, "-- done sleeping");
1690 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1691 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1692 imp->imp_obd->obd_uuid.uuid,
1693 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1694 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1695 lustre_msg_get_opc(req->rq_reqmsg));
1697 spin_lock(&imp->imp_lock);
1698 list_del_init(&req->rq_list);
1699 spin_unlock(&imp->imp_lock);
1701 /* If the reply was received normally, this just grabs the spinlock
1702 * (ensuring the reply callback has returned), sees that
1703 * req->rq_receiving_reply is clear and returns. */
1704 ptlrpc_unregister_reply (req);
1707 GOTO(out, rc = -EIO);
1709 /* Resend if we need to, unless we were interrupted. */
1710 if (req->rq_resend && !req->rq_intr) {
1711 /* ...unless we were specifically told otherwise. */
1712 if (req->rq_no_resend)
1713 GOTO(out, rc = -ETIMEDOUT);
1714 spin_lock(&imp->imp_lock);
1719 /* Should only be interrupted if we timed out. */
1720 if (!req->rq_timedout)
1721 DEBUG_REQ(D_ERROR, req,
1722 "rq_intr set but rq_timedout not");
1723 GOTO(out, rc = -EINTR);
1726 if (req->rq_timedout) { /* non-recoverable timeout */
1727 GOTO(out, rc = -ETIMEDOUT);
1730 if (!req->rq_replied) {
1731 /* How can this be? -eeb */
1732 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1734 GOTO(out, rc = req->rq_status);
1737 rc = after_reply(req);
1738 /* NB may return +ve success rc */
1739 if (req->rq_resend) {
1740 spin_lock(&imp->imp_lock);
1745 if (req->rq_bulk != NULL) {
1747 /* success so far. Note that anything going wrong
1748 * with bulk now, is EXTREMELY strange, since the
1749 * server must have believed that the bulk
1750 * tranferred OK before she replied with success to
1752 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1753 brc = l_wait_event(req->rq_reply_waitq,
1754 !ptlrpc_bulk_active(req->rq_bulk),
1756 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1758 LASSERT(brc == -ETIMEDOUT);
1759 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1761 } else if (!req->rq_bulk->bd_success) {
1762 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1767 ptlrpc_unregister_bulk (req);
1770 LASSERT(!req->rq_receiving_reply);
1771 req->rq_phase = RQ_PHASE_INTERPRET;
1773 atomic_dec(&imp->imp_inflight);
1774 cfs_waitq_signal(&imp->imp_recovery_waitq);
1778 struct ptlrpc_replay_async_args {
1780 int praa_old_status;
1783 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1784 void * data, int rc)
1786 struct ptlrpc_replay_async_args *aa = data;
1787 struct obd_import *imp = req->rq_import;
1790 atomic_dec(&imp->imp_replay_inflight);
1792 if (!req->rq_replied) {
1793 CERROR("request replay timed out, restarting recovery\n");
1794 GOTO(out, rc = -ETIMEDOUT);
1797 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1798 lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN)
1799 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1801 /* The transno had better not change over replay. */
1802 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1803 lustre_msg_get_transno(req->rq_repmsg));
1805 DEBUG_REQ(D_HA, req, "got rep");
1807 /* let the callback do fixups, possibly including in the request */
1808 if (req->rq_replay_cb)
1809 req->rq_replay_cb(req);
1811 if (req->rq_replied &&
1812 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1813 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1814 lustre_msg_get_status(req->rq_repmsg),
1815 aa->praa_old_status);
1817 /* Put it back for re-replay. */
1818 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1821 spin_lock(&imp->imp_lock);
1822 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
1823 imp->imp_last_replay_transno = req->rq_transno;
1824 spin_unlock(&imp->imp_lock);
1826 /* continue with recovery */
1827 rc = ptlrpc_import_recovery_state_machine(imp);
1829 req->rq_send_state = aa->praa_old_state;
1832 /* this replay failed, so restart recovery */
1833 ptlrpc_connect_import(imp, NULL);
1838 int ptlrpc_replay_req(struct ptlrpc_request *req)
1840 struct ptlrpc_replay_async_args *aa;
1843 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1845 /* Not handling automatic bulk replay yet (or ever?) */
1846 LASSERT(req->rq_bulk == NULL);
1848 DEBUG_REQ(D_HA, req, "REPLAY");
1850 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1851 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1852 memset(aa, 0, sizeof *aa);
1854 /* Prepare request to be resent with ptlrpcd */
1855 aa->praa_old_state = req->rq_send_state;
1856 req->rq_send_state = LUSTRE_IMP_REPLAY;
1857 req->rq_phase = RQ_PHASE_NEW;
1858 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
1861 req->rq_interpret_reply = ptlrpc_replay_interpret;
1862 atomic_inc(&req->rq_import->imp_replay_inflight);
1863 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1865 ptlrpcd_add_req(req);
1869 void ptlrpc_abort_inflight(struct obd_import *imp)
1871 struct list_head *tmp, *n;
1874 /* Make sure that no new requests get processed for this import.
1875 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1876 * this flag and then putting requests on sending_list or delayed_list.
1878 spin_lock(&imp->imp_lock);
1880 /* XXX locking? Maybe we should remove each request with the list
1881 * locked? Also, how do we know if the requests on the list are
1882 * being freed at this time?
1884 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1885 struct ptlrpc_request *req =
1886 list_entry(tmp, struct ptlrpc_request, rq_list);
1888 DEBUG_REQ(D_HA, req, "inflight");
1890 spin_lock (&req->rq_lock);
1891 if (req->rq_import_generation < imp->imp_generation) {
1893 ptlrpc_wake_client_req(req);
1895 spin_unlock (&req->rq_lock);
1898 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1899 struct ptlrpc_request *req =
1900 list_entry(tmp, struct ptlrpc_request, rq_list);
1902 DEBUG_REQ(D_HA, req, "aborting waiting req");
1904 spin_lock (&req->rq_lock);
1905 if (req->rq_import_generation < imp->imp_generation) {
1907 ptlrpc_wake_client_req(req);
1909 spin_unlock (&req->rq_lock);
1912 /* Last chance to free reqs left on the replay list, but we
1913 * will still leak reqs that haven't committed. */
1914 if (imp->imp_replayable)
1915 ptlrpc_free_committed(imp);
1917 spin_unlock(&imp->imp_lock);
1922 static __u64 ptlrpc_last_xid = 0;
1923 spinlock_t ptlrpc_last_xid_lock;
1925 __u64 ptlrpc_next_xid(void)
1928 spin_lock(&ptlrpc_last_xid_lock);
1929 tmp = ++ptlrpc_last_xid;
1930 spin_unlock(&ptlrpc_last_xid_lock);
1934 __u64 ptlrpc_sample_next_xid(void)
1937 spin_lock(&ptlrpc_last_xid_lock);
1938 tmp = ptlrpc_last_xid + 1;
1939 spin_unlock(&ptlrpc_last_xid_lock);
1942 EXPORT_SYMBOL(ptlrpc_sample_next_xid);