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 class_export_put(desc->bd_export);
185 class_import_put(desc->bd_import);
187 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
188 bd_iov[desc->bd_max_iov]));
192 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
194 struct list_head *l, *tmp;
195 struct ptlrpc_request *req;
200 list_for_each_safe(l, tmp, &pool->prp_req_list) {
201 req = list_entry(l, struct ptlrpc_request, rq_list);
202 list_del(&req->rq_list);
203 LASSERT (req->rq_reqmsg);
204 OBD_FREE(req->rq_reqmsg, pool->prp_rq_size);
205 OBD_FREE(req, sizeof(*req));
207 OBD_FREE(pool, sizeof(*pool));
210 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
215 while (size < pool->prp_rq_size)
218 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
219 "Trying to change pool size with nonempty pool "
220 "from %d to %d bytes\n", pool->prp_rq_size, size);
222 spin_lock(&pool->prp_lock);
223 pool->prp_rq_size = size;
224 for (i = 0; i < num_rq; i++) {
225 struct ptlrpc_request *req;
226 struct lustre_msg *msg;
228 spin_unlock(&pool->prp_lock);
229 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
232 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
234 OBD_FREE(req, sizeof(struct ptlrpc_request));
237 req->rq_reqmsg = msg;
239 spin_lock(&pool->prp_lock);
240 list_add_tail(&req->rq_list, &pool->prp_req_list);
242 spin_unlock(&pool->prp_lock);
246 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
247 void (*populate_pool)(struct ptlrpc_request_pool *, int))
249 struct ptlrpc_request_pool *pool;
251 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
255 /* Request next power of two for the allocation, because internally
256 kernel would do exactly this */
258 spin_lock_init(&pool->prp_lock);
259 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
260 pool->prp_rq_size = msgsize;
261 pool->prp_populate = populate_pool;
263 populate_pool(pool, num_rq);
265 if (list_empty(&pool->prp_req_list)) {
266 /* have not allocated a single request for the pool */
267 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
273 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
275 struct ptlrpc_request *request;
276 struct lustre_msg *reqmsg;
281 spin_lock(&pool->prp_lock);
283 /* See if we have anything in a pool, and bail out if nothing,
284 * in writeout path, where this matters, this is safe to do, because
285 * nothing is lost in this case, and when some in-flight requests
286 * complete, this code will be called again. */
287 if (unlikely(list_empty(&pool->prp_req_list))) {
288 spin_unlock(&pool->prp_lock);
292 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
294 list_del(&request->rq_list);
295 spin_unlock(&pool->prp_lock);
297 LASSERT(request->rq_reqmsg);
298 LASSERT(request->rq_pool);
300 reqmsg = request->rq_reqmsg;
301 memset(request, 0, sizeof(*request));
302 request->rq_reqmsg = reqmsg;
303 request->rq_pool = pool;
304 request->rq_reqlen = pool->prp_rq_size;
308 struct ptlrpc_request *
309 ptlrpc_prep_req_pool(struct obd_import *imp, __u32 version, int opcode,
310 int count, int *lengths, char **bufs,
311 struct ptlrpc_request_pool *pool)
313 struct ptlrpc_request *request = NULL;
317 /* The obd disconnected */
321 LASSERT(imp != LP_POISON);
322 LASSERT((unsigned long)imp->imp_client > 0x1000);
323 LASSERT(imp->imp_client != LP_POISON);
326 request = ptlrpc_prep_req_from_pool(pool);
329 OBD_ALLOC(request, sizeof(*request));
332 CERROR("request allocation out of memory\n");
336 rc = lustre_pack_request(request, imp->imp_msg_magic, count, lengths,
339 LASSERT(!request->rq_pool);
340 OBD_FREE(request, sizeof(*request));
344 lustre_msg_add_version(request->rq_reqmsg, version);
346 if (imp->imp_server_timeout)
347 request->rq_timeout = obd_timeout / 2;
349 request->rq_timeout = obd_timeout;
350 request->rq_send_state = LUSTRE_IMP_FULL;
351 request->rq_type = PTL_RPC_MSG_REQUEST;
352 request->rq_import = class_import_get(imp);
353 request->rq_export = NULL;
355 request->rq_req_cbid.cbid_fn = request_out_callback;
356 request->rq_req_cbid.cbid_arg = request;
358 request->rq_reply_cbid.cbid_fn = reply_in_callback;
359 request->rq_reply_cbid.cbid_arg = request;
361 request->rq_phase = RQ_PHASE_NEW;
363 /* XXX FIXME bug 249 */
364 request->rq_request_portal = imp->imp_client->cli_request_portal;
365 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
367 spin_lock_init(&request->rq_lock);
368 CFS_INIT_LIST_HEAD(&request->rq_list);
369 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
370 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
371 cfs_waitq_init(&request->rq_reply_waitq);
372 request->rq_xid = ptlrpc_next_xid();
373 atomic_set(&request->rq_refcount, 1);
375 lustre_msg_set_opc(request->rq_reqmsg, opcode);
376 lustre_msg_set_flags(request->rq_reqmsg, 0);
381 struct ptlrpc_request *
382 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
383 int *lengths, char **bufs)
385 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
389 struct ptlrpc_request_set *ptlrpc_prep_set(void)
391 struct ptlrpc_request_set *set;
394 OBD_ALLOC(set, sizeof *set);
397 CFS_INIT_LIST_HEAD(&set->set_requests);
398 cfs_waitq_init(&set->set_waitq);
399 set->set_remaining = 0;
400 spin_lock_init(&set->set_new_req_lock);
401 CFS_INIT_LIST_HEAD(&set->set_new_requests);
406 /* Finish with this set; opposite of prep_set. */
407 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
409 struct list_head *tmp;
410 struct list_head *next;
415 /* Requests on the set should either all be completed, or all be new */
416 expected_phase = (set->set_remaining == 0) ?
417 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
418 list_for_each (tmp, &set->set_requests) {
419 struct ptlrpc_request *req =
420 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
422 LASSERT(req->rq_phase == expected_phase);
426 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
428 list_for_each_safe(tmp, next, &set->set_requests) {
429 struct ptlrpc_request *req =
430 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
431 list_del_init(&req->rq_set_chain);
433 LASSERT(req->rq_phase == expected_phase);
435 if (req->rq_phase == RQ_PHASE_NEW) {
437 if (req->rq_interpret_reply != NULL) {
438 int (*interpreter)(struct ptlrpc_request *,
440 req->rq_interpret_reply;
442 /* higher level (i.e. LOV) failed;
443 * let the sub reqs clean up */
444 req->rq_status = -EBADR;
445 interpreter(req, &req->rq_async_args,
448 set->set_remaining--;
452 ptlrpc_req_finished (req);
455 LASSERT(set->set_remaining == 0);
457 OBD_FREE(set, sizeof(*set));
461 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
462 struct ptlrpc_request *req)
464 /* The set takes over the caller's request reference */
465 list_add_tail(&req->rq_set_chain, &set->set_requests);
467 set->set_remaining++;
469 atomic_inc(&req->rq_import->imp_inflight);
472 /* lock so many callers can add things, the context that owns the set
473 * is supposed to notice these and move them into the set proper. */
474 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
475 struct ptlrpc_request *req)
477 spin_lock(&set->set_new_req_lock);
478 /* The set takes over the caller's request reference */
479 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
481 spin_unlock(&set->set_new_req_lock);
485 * Based on the current state of the import, determine if the request
486 * can be sent, is an error, or should be delayed.
488 * Returns true if this request should be delayed. If false, and
489 * *status is set, then the request can not be sent and *status is the
490 * error code. If false and status is 0, then request can be sent.
492 * The imp->imp_lock must be held.
494 static int ptlrpc_import_delay_req(struct obd_import *imp,
495 struct ptlrpc_request *req, int *status)
500 LASSERT (status != NULL);
503 if (imp->imp_state == LUSTRE_IMP_NEW) {
504 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
507 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
508 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
510 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
511 imp->imp_state == LUSTRE_IMP_CONNECTING) {
512 /* allow CONNECT even if import is invalid */ ;
513 } else if (imp->imp_invalid) {
514 /* If the import has been invalidated (such as by an OST
515 * failure), the request must fail with -EIO. */
516 if (!imp->imp_deactive)
517 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
519 } else if (req->rq_import_generation != imp->imp_generation) {
520 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
522 } else if (req->rq_send_state != imp->imp_state) {
523 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
525 *status = -EWOULDBLOCK;
533 static int ptlrpc_check_reply(struct ptlrpc_request *req)
538 /* serialise with network callback */
539 spin_lock(&req->rq_lock);
544 if (req->rq_net_err && !req->rq_timedout) {
545 spin_unlock(&req->rq_lock);
546 rc = ptlrpc_expire_one_request(req);
547 spin_lock(&req->rq_lock);
561 spin_unlock(&req->rq_lock);
562 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
566 static int ptlrpc_check_status(struct ptlrpc_request *req)
571 err = lustre_msg_get_status(req->rq_repmsg);
572 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
573 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err == %d",
575 RETURN(err < 0 ? err : -EINVAL);
579 DEBUG_REQ(D_INFO, req, "status is %d", err);
580 } else if (err > 0) {
581 /* XXX: translate this error from net to host */
582 DEBUG_REQ(D_INFO, req, "status is %d", err);
588 static int after_reply(struct ptlrpc_request *req)
590 struct obd_import *imp = req->rq_import;
594 LASSERT(!req->rq_receiving_reply);
596 /* NB Until this point, the whole of the incoming message,
597 * including buflens, status etc is in the sender's byte order. */
600 /* Clear reply swab mask; this is a new reply in sender's byte order */
601 req->rq_rep_swab_mask = 0;
603 LASSERT (req->rq_nob_received <= req->rq_replen);
604 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
606 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d\n", rc);
610 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
612 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d\n", rc);
616 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
617 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
618 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)\n",
619 lustre_msg_get_type(req->rq_repmsg));
623 rc = ptlrpc_check_status(req);
625 /* Either we've been evicted, or the server has failed for
626 * some reason. Try to reconnect, and if that fails, punt to the
628 if ((rc == -ENOTCONN) || (rc == -ENODEV)) {
629 if (req->rq_send_state != LUSTRE_IMP_FULL ||
630 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
634 ptlrpc_request_handle_notconn(req);
639 /* Store transno in reqmsg for replay. */
640 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
641 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
643 if (req->rq_import->imp_replayable) {
644 spin_lock(&imp->imp_lock);
645 /* no point in adding already-committed requests to the replay
646 * list, we will just remove them immediately. b=9829 */
647 if (req->rq_transno != 0 &&
649 lustre_msg_get_last_committed(req->rq_repmsg) ||
651 ptlrpc_retain_replayable_request(req, imp);
652 else if (req->rq_commit_cb != NULL) {
653 spin_unlock(&imp->imp_lock);
654 req->rq_commit_cb(req);
655 spin_lock(&imp->imp_lock);
658 /* Replay-enabled imports return commit-status information. */
659 if (lustre_msg_get_last_committed(req->rq_repmsg))
660 imp->imp_peer_committed_transno =
661 lustre_msg_get_last_committed(req->rq_repmsg);
662 ptlrpc_free_committed(imp);
663 spin_unlock(&imp->imp_lock);
669 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
671 struct obd_import *imp;
675 LASSERT(req->rq_phase == RQ_PHASE_NEW);
676 req->rq_phase = RQ_PHASE_RPC;
678 imp = req->rq_import;
679 spin_lock(&imp->imp_lock);
681 req->rq_import_generation = imp->imp_generation;
683 if (ptlrpc_import_delay_req(imp, req, &rc)) {
684 spin_lock (&req->rq_lock);
686 spin_unlock (&req->rq_lock);
688 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
690 lustre_msg_get_status(req->rq_reqmsg) ,
691 ptlrpc_import_state_name(req->rq_send_state),
692 ptlrpc_import_state_name(imp->imp_state));
693 LASSERT(list_empty (&req->rq_list));
695 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
696 spin_unlock(&imp->imp_lock);
701 spin_unlock(&imp->imp_lock);
703 req->rq_phase = RQ_PHASE_INTERPRET;
707 /* XXX this is the same as ptlrpc_queue_wait */
708 LASSERT(list_empty(&req->rq_list));
709 list_add_tail(&req->rq_list, &imp->imp_sending_list);
710 spin_unlock(&imp->imp_lock);
712 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
713 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
714 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
715 imp->imp_obd->obd_uuid.uuid,
716 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
717 libcfs_nid2str(imp->imp_connection->c_peer.nid),
718 lustre_msg_get_opc(req->rq_reqmsg));
720 rc = ptl_send_rpc(req, 0);
722 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
729 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
730 int ptlrpc_check_set(struct ptlrpc_request_set *set)
732 struct list_head *tmp;
733 int force_timer_recalc = 0;
736 if (set->set_remaining == 0)
739 list_for_each(tmp, &set->set_requests) {
740 struct ptlrpc_request *req =
741 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
742 struct obd_import *imp = req->rq_import;
745 if (req->rq_phase == RQ_PHASE_NEW &&
746 ptlrpc_send_new_req(req)) {
747 force_timer_recalc = 1;
750 if (!(req->rq_phase == RQ_PHASE_RPC ||
751 req->rq_phase == RQ_PHASE_BULK ||
752 req->rq_phase == RQ_PHASE_INTERPRET ||
753 req->rq_phase == RQ_PHASE_COMPLETE)) {
754 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
758 if (req->rq_phase == RQ_PHASE_COMPLETE)
761 if (req->rq_phase == RQ_PHASE_INTERPRET)
762 GOTO(interpret, req->rq_status);
764 if (req->rq_net_err && !req->rq_timedout)
765 ptlrpc_expire_one_request(req);
768 ptlrpc_unregister_reply(req);
769 if (req->rq_status == 0)
770 req->rq_status = -EIO;
771 req->rq_phase = RQ_PHASE_INTERPRET;
773 spin_lock(&imp->imp_lock);
774 list_del_init(&req->rq_list);
775 spin_unlock(&imp->imp_lock);
777 GOTO(interpret, req->rq_status);
780 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
781 * will only be set after rq_timedout, but the oig waiting
782 * path sets rq_intr irrespective of whether ptlrpcd has
783 * seen a timeout. our policy is to only interpret
784 * interrupted rpcs after they have timed out */
785 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
786 /* NB could be on delayed list */
787 ptlrpc_unregister_reply(req);
788 req->rq_status = -EINTR;
789 req->rq_phase = RQ_PHASE_INTERPRET;
791 spin_lock(&imp->imp_lock);
792 list_del_init(&req->rq_list);
793 spin_unlock(&imp->imp_lock);
795 GOTO(interpret, req->rq_status);
798 if (req->rq_phase == RQ_PHASE_RPC) {
799 if (req->rq_timedout||req->rq_waiting||req->rq_resend) {
802 ptlrpc_unregister_reply(req);
804 spin_lock(&imp->imp_lock);
806 if (ptlrpc_import_delay_req(imp, req, &status)){
807 spin_unlock(&imp->imp_lock);
811 list_del_init(&req->rq_list);
813 req->rq_status = status;
814 req->rq_phase = RQ_PHASE_INTERPRET;
815 spin_unlock(&imp->imp_lock);
816 GOTO(interpret, req->rq_status);
818 if (req->rq_no_resend) {
819 req->rq_status = -ENOTCONN;
820 req->rq_phase = RQ_PHASE_INTERPRET;
821 spin_unlock(&imp->imp_lock);
822 GOTO(interpret, req->rq_status);
824 list_add_tail(&req->rq_list,
825 &imp->imp_sending_list);
827 spin_unlock(&imp->imp_lock);
830 if (req->rq_resend) {
831 lustre_msg_add_flags(req->rq_reqmsg,
834 __u64 old_xid = req->rq_xid;
836 ptlrpc_unregister_bulk (req);
838 /* ensure previous bulk fails */
839 req->rq_xid = ptlrpc_next_xid();
840 CDEBUG(D_HA, "resend bulk "
843 old_xid, req->rq_xid);
847 rc = ptl_send_rpc(req, 0);
849 DEBUG_REQ(D_HA, req, "send failed (%d)",
851 force_timer_recalc = 1;
854 /* need to reset the timeout */
855 force_timer_recalc = 1;
858 /* Still waiting for a reply? */
859 if (ptlrpc_client_receiving_reply(req))
862 /* Did we actually receive a reply? */
863 if (!ptlrpc_client_replied(req))
866 spin_lock(&imp->imp_lock);
867 list_del_init(&req->rq_list);
868 spin_unlock(&imp->imp_lock);
870 req->rq_status = after_reply(req);
871 if (req->rq_resend) {
872 /* Add this req to the delayed list so
873 it can be errored if the import is
874 evicted after recovery. */
875 spin_lock(&req->rq_lock);
876 list_add_tail(&req->rq_list,
877 &imp->imp_delayed_list);
878 spin_unlock(&req->rq_lock);
882 /* If there is no bulk associated with this request,
883 * then we're done and should let the interpreter
884 * process the reply. Similarly if the RPC returned
885 * an error, and therefore the bulk will never arrive.
887 if (req->rq_bulk == NULL || req->rq_status != 0) {
888 req->rq_phase = RQ_PHASE_INTERPRET;
889 GOTO(interpret, req->rq_status);
892 req->rq_phase = RQ_PHASE_BULK;
895 LASSERT(req->rq_phase == RQ_PHASE_BULK);
896 if (ptlrpc_bulk_active(req->rq_bulk))
899 if (!req->rq_bulk->bd_success) {
900 /* The RPC reply arrived OK, but the bulk screwed
901 * up! Dead wierd since the server told us the RPC
902 * was good after getting the REPLY for her GET or
903 * the ACK for her PUT. */
904 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
908 req->rq_phase = RQ_PHASE_INTERPRET;
911 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
912 LASSERT(!req->rq_receiving_reply);
914 ptlrpc_unregister_reply(req);
915 if (req->rq_bulk != NULL)
916 ptlrpc_unregister_bulk (req);
918 req->rq_phase = RQ_PHASE_COMPLETE;
920 if (req->rq_interpret_reply != NULL) {
921 int (*interpreter)(struct ptlrpc_request *,void *,int) =
922 req->rq_interpret_reply;
923 req->rq_status = interpreter(req, &req->rq_async_args,
927 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
928 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
929 imp->imp_obd->obd_uuid.uuid,
930 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
931 libcfs_nid2str(imp->imp_connection->c_peer.nid),
932 lustre_msg_get_opc(req->rq_reqmsg));
934 set->set_remaining--;
936 atomic_dec(&imp->imp_inflight);
937 cfs_waitq_signal(&imp->imp_recovery_waitq);
940 /* If we hit an error, we want to recover promptly. */
941 RETURN(set->set_remaining == 0 || force_timer_recalc);
944 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
946 struct obd_import *imp = req->rq_import;
949 DEBUG_REQ(D_ERROR|D_NETERROR, req, "timeout (sent at %lu, %lus ago)",
950 (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
952 if (imp != NULL && obd_debug_peer_on_timeout)
953 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
955 spin_lock(&req->rq_lock);
956 req->rq_timedout = 1;
957 spin_unlock(&req->rq_lock);
959 ptlrpc_unregister_reply (req);
961 if (obd_dump_on_timeout)
962 libcfs_debug_dumplog();
964 if (req->rq_bulk != NULL)
965 ptlrpc_unregister_bulk (req);
968 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
972 /* The DLM server doesn't want recovery run on its imports. */
973 if (imp->imp_dlm_fake)
976 /* If this request is for recovery or other primordial tasks,
977 * then error it out here. */
978 if (req->rq_send_state != LUSTRE_IMP_FULL ||
979 imp->imp_obd->obd_no_recov) {
980 spin_lock(&req->rq_lock);
981 req->rq_status = -ETIMEDOUT;
983 spin_unlock(&req->rq_lock);
987 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
992 int ptlrpc_expired_set(void *data)
994 struct ptlrpc_request_set *set = data;
995 struct list_head *tmp;
996 time_t now = CURRENT_SECONDS;
999 LASSERT(set != NULL);
1001 /* A timeout expired; see which reqs it applies to... */
1002 list_for_each (tmp, &set->set_requests) {
1003 struct ptlrpc_request *req =
1004 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1006 /* request in-flight? */
1007 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1009 (req->rq_phase == RQ_PHASE_BULK)))
1012 if (req->rq_timedout || /* already dealt with */
1013 req->rq_sent + req->rq_timeout > now) /* not expired */
1016 /* deal with this guy */
1017 ptlrpc_expire_one_request (req);
1020 /* When waiting for a whole set, we always to break out of the
1021 * sleep so we can recalculate the timeout, or enable interrupts
1022 * iff everyone's timed out.
1027 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1029 spin_lock(&req->rq_lock);
1031 spin_unlock(&req->rq_lock);
1034 void ptlrpc_interrupted_set(void *data)
1036 struct ptlrpc_request_set *set = data;
1037 struct list_head *tmp;
1039 LASSERT(set != NULL);
1040 CERROR("INTERRUPTED SET %p\n", set);
1042 list_for_each(tmp, &set->set_requests) {
1043 struct ptlrpc_request *req =
1044 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1046 if (req->rq_phase != RQ_PHASE_RPC)
1049 ptlrpc_mark_interrupted(req);
1053 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1055 struct list_head *tmp;
1056 time_t now = CURRENT_SECONDS;
1059 struct ptlrpc_request *req;
1062 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1064 list_for_each(tmp, &set->set_requests) {
1065 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1067 /* request in-flight? */
1068 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1069 (req->rq_phase == RQ_PHASE_BULK)))
1072 if (req->rq_timedout) /* already timed out */
1075 deadline = req->rq_sent + req->rq_timeout;
1076 if (deadline <= now) /* actually expired already */
1077 timeout = 1; /* ASAP */
1078 else if (timeout == 0 || timeout > deadline - now)
1079 timeout = deadline - now;
1084 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1086 struct list_head *tmp;
1087 struct ptlrpc_request *req;
1088 struct l_wait_info lwi;
1092 if (list_empty(&set->set_requests))
1095 list_for_each(tmp, &set->set_requests) {
1096 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1097 if (req->rq_phase == RQ_PHASE_NEW)
1098 (void)ptlrpc_send_new_req(req);
1102 timeout = ptlrpc_set_next_timeout(set);
1104 /* wait until all complete, interrupted, or an in-flight
1106 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
1108 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1110 ptlrpc_interrupted_set, set);
1111 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1113 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1115 /* -EINTR => all requests have been flagged rq_intr so next
1117 * -ETIMEOUTD => someone timed out. When all reqs have
1118 * timed out, signals are enabled allowing completion with
1120 * I don't really care if we go once more round the loop in
1121 * the error cases -eeb. */
1122 } while (rc != 0 || set->set_remaining != 0);
1124 LASSERT(set->set_remaining == 0);
1127 list_for_each(tmp, &set->set_requests) {
1128 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1130 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1131 if (req->rq_status != 0)
1132 rc = req->rq_status;
1135 if (set->set_interpret != NULL) {
1136 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1138 rc = interpreter (set, set->set_arg, rc);
1144 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
1146 struct ptlrpc_request_pool *pool = request->rq_pool;
1148 spin_lock(&pool->prp_lock);
1149 list_add_tail(&request->rq_list, &pool->prp_req_list);
1150 spin_unlock(&pool->prp_lock);
1153 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1156 if (request == NULL) {
1161 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1162 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1163 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1164 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1166 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1167 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1168 if (request->rq_import != NULL) {
1170 spin_lock(&request->rq_import->imp_lock);
1171 list_del_init(&request->rq_replay_list);
1173 spin_unlock(&request->rq_import->imp_lock);
1175 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1177 if (atomic_read(&request->rq_refcount) != 0) {
1178 DEBUG_REQ(D_ERROR, request,
1179 "freeing request with nonzero refcount");
1183 if (request->rq_repmsg != NULL) {
1184 OBD_FREE(request->rq_repmsg, request->rq_replen);
1185 request->rq_repmsg = NULL;
1187 if (request->rq_export != NULL) {
1188 class_export_put(request->rq_export);
1189 request->rq_export = NULL;
1191 if (request->rq_import != NULL) {
1192 class_import_put(request->rq_import);
1193 request->rq_import = NULL;
1195 if (request->rq_bulk != NULL)
1196 ptlrpc_free_bulk(request->rq_bulk);
1198 if (request->rq_pool) {
1199 __ptlrpc_free_req_to_pool(request);
1201 if (request->rq_reqmsg != NULL) {
1202 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1203 request->rq_reqmsg = NULL;
1205 OBD_FREE(request, sizeof(*request));
1210 void ptlrpc_free_req(struct ptlrpc_request *request)
1212 __ptlrpc_free_req(request, 0);
1215 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1216 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1218 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1219 (void)__ptlrpc_req_finished(request, 1);
1222 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1225 if (request == NULL)
1228 if (request == LP_POISON ||
1229 request->rq_reqmsg == LP_POISON) {
1230 CERROR("dereferencing freed request (bug 575)\n");
1235 DEBUG_REQ(D_INFO, request, "refcount now %u",
1236 atomic_read(&request->rq_refcount) - 1);
1238 if (atomic_dec_and_test(&request->rq_refcount)) {
1239 __ptlrpc_free_req(request, locked);
1246 void ptlrpc_req_finished(struct ptlrpc_request *request)
1248 __ptlrpc_req_finished(request, 0);
1251 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1253 return request->rq_xid;
1255 EXPORT_SYMBOL(ptlrpc_req_xid);
1257 /* Disengage the client's reply buffer from the network
1258 * NB does _NOT_ unregister any client-side bulk.
1259 * IDEMPOTENT, but _not_ safe against concurrent callers.
1260 * The request owner (i.e. the thread doing the I/O) must call...
1262 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1266 struct l_wait_info lwi;
1268 LASSERT(!in_interrupt ()); /* might sleep */
1270 if (!ptlrpc_client_receiving_reply(request))
1273 LNetMDUnlink (request->rq_reply_md_h);
1275 /* We have to l_wait_event() whatever the result, to give liblustre
1276 * a chance to run reply_in_callback() */
1278 if (request->rq_set != NULL)
1279 wq = &request->rq_set->set_waitq;
1281 wq = &request->rq_reply_waitq;
1284 /* Network access will complete in finite time but the HUGE
1285 * timeout lets us CWARN for visibility of sluggish NALs */
1286 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1287 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1291 LASSERT (rc == -ETIMEDOUT);
1292 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1296 /* caller must hold imp->imp_lock */
1297 void ptlrpc_free_committed(struct obd_import *imp)
1299 struct list_head *tmp, *saved;
1300 struct ptlrpc_request *req;
1301 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1304 LASSERT(imp != NULL);
1306 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1309 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1310 imp->imp_generation == imp->imp_last_generation_checked) {
1311 CDEBUG(D_HA, "%s: skip recheck for last_committed "LPU64"\n",
1312 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1316 CDEBUG(D_HA, "%s: committing for last_committed "LPU64" gen %d\n",
1317 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1318 imp->imp_generation);
1319 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1320 imp->imp_last_generation_checked = imp->imp_generation;
1322 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1323 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1325 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1326 LASSERT(req != last_req);
1329 if (req->rq_import_generation < imp->imp_generation) {
1330 DEBUG_REQ(D_HA, req, "freeing request with old gen");
1334 if (req->rq_replay) {
1335 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1339 /* not yet committed */
1340 if (req->rq_transno > imp->imp_peer_committed_transno) {
1341 DEBUG_REQ(D_HA, req, "stopping search");
1345 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1346 imp->imp_peer_committed_transno);
1348 if (req->rq_commit_cb != NULL)
1349 req->rq_commit_cb(req);
1350 list_del_init(&req->rq_replay_list);
1351 __ptlrpc_req_finished(req, 1);
1358 void ptlrpc_cleanup_client(struct obd_import *imp)
1365 void ptlrpc_resend_req(struct ptlrpc_request *req)
1367 DEBUG_REQ(D_HA, req, "going to resend");
1368 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1369 req->rq_status = -EAGAIN;
1371 spin_lock(&req->rq_lock);
1373 req->rq_net_err = 0;
1374 req->rq_timedout = 0;
1376 __u64 old_xid = req->rq_xid;
1378 /* ensure previous bulk fails */
1379 req->rq_xid = ptlrpc_next_xid();
1380 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1381 old_xid, req->rq_xid);
1383 ptlrpc_wake_client_req(req);
1384 spin_unlock(&req->rq_lock);
1387 /* XXX: this function and rq_status are currently unused */
1388 void ptlrpc_restart_req(struct ptlrpc_request *req)
1390 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1391 req->rq_status = -ERESTARTSYS;
1393 spin_lock(&req->rq_lock);
1394 req->rq_restart = 1;
1395 req->rq_timedout = 0;
1396 ptlrpc_wake_client_req(req);
1397 spin_unlock(&req->rq_lock);
1400 static int expired_request(void *data)
1402 struct ptlrpc_request *req = data;
1405 RETURN(ptlrpc_expire_one_request(req));
1408 static void interrupted_request(void *data)
1410 struct ptlrpc_request *req = data;
1411 DEBUG_REQ(D_HA, req, "request interrupted");
1412 spin_lock(&req->rq_lock);
1414 spin_unlock(&req->rq_lock);
1417 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1420 atomic_inc(&req->rq_refcount);
1424 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1425 struct obd_import *imp)
1427 struct list_head *tmp;
1429 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1431 /* clear this for new requests that were resent as well
1432 as resent replayed requests. */
1433 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1435 /* don't re-add requests that have been replayed */
1436 if (!list_empty(&req->rq_replay_list))
1439 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1441 LASSERT(imp->imp_replayable);
1442 /* Balanced in ptlrpc_free_committed, usually. */
1443 ptlrpc_request_addref(req);
1444 list_for_each_prev(tmp, &imp->imp_replay_list) {
1445 struct ptlrpc_request *iter =
1446 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1448 /* We may have duplicate transnos if we create and then
1449 * open a file, or for closes retained if to match creating
1450 * opens, so use req->rq_xid as a secondary key.
1451 * (See bugs 684, 685, and 428.)
1452 * XXX no longer needed, but all opens need transnos!
1454 if (iter->rq_transno > req->rq_transno)
1457 if (iter->rq_transno == req->rq_transno) {
1458 LASSERT(iter->rq_xid != req->rq_xid);
1459 if (iter->rq_xid > req->rq_xid)
1463 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1467 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1470 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1474 struct l_wait_info lwi;
1475 struct obd_import *imp = req->rq_import;
1476 cfs_duration_t timeout = 0;
1479 LASSERT(req->rq_set == NULL);
1480 LASSERT(!req->rq_receiving_reply);
1481 atomic_inc(&imp->imp_inflight);
1483 /* for distributed debugging */
1484 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1485 LASSERT(imp->imp_obd != NULL);
1486 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1487 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1488 imp->imp_obd->obd_uuid.uuid,
1489 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1490 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1491 lustre_msg_get_opc(req->rq_reqmsg));
1493 /* Mark phase here for a little debug help */
1494 req->rq_phase = RQ_PHASE_RPC;
1496 spin_lock(&imp->imp_lock);
1497 req->rq_import_generation = imp->imp_generation;
1499 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1500 list_del(&req->rq_list);
1502 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1503 spin_unlock(&imp->imp_lock);
1505 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1507 ptlrpc_import_state_name(req->rq_send_state),
1508 ptlrpc_import_state_name(imp->imp_state));
1509 lwi = LWI_INTR(interrupted_request, req);
1510 rc = l_wait_event(req->rq_reply_waitq,
1511 (req->rq_send_state == imp->imp_state ||
1512 req->rq_err || req->rq_intr),
1514 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1516 ptlrpc_import_state_name(imp->imp_state),
1517 ptlrpc_import_state_name(req->rq_send_state),
1518 req->rq_err, req->rq_intr);
1520 spin_lock(&imp->imp_lock);
1521 list_del_init(&req->rq_list);
1526 else if (req->rq_intr) {
1529 else if (req->rq_no_resend) {
1530 spin_unlock(&imp->imp_lock);
1531 GOTO(out, rc = -ETIMEDOUT);
1539 list_del_init(&req->rq_list);
1540 spin_unlock(&imp->imp_lock);
1541 req->rq_status = rc; // XXX this ok?
1545 if (req->rq_resend) {
1546 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1548 if (req->rq_bulk != NULL) {
1549 ptlrpc_unregister_bulk (req);
1551 /* bulk requests are supposed to be
1552 * idempotent, so we are free to bump the xid
1553 * here, which we need to do before
1554 * registering the bulk again (bug 6371).
1555 * print the old xid first for sanity.
1557 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1558 req->rq_xid = ptlrpc_next_xid();
1561 DEBUG_REQ(D_HA, req, "resending: ");
1564 /* XXX this is the same as ptlrpc_set_wait */
1565 LASSERT(list_empty(&req->rq_list));
1566 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1567 spin_unlock(&imp->imp_lock);
1569 rc = ptl_send_rpc(req, 0);
1571 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1574 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1575 DEBUG_REQ(D_NET, req,
1576 "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1578 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1580 l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1582 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1583 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1584 imp->imp_obd->obd_uuid.uuid,
1585 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1586 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1587 lustre_msg_get_opc(req->rq_reqmsg));
1589 spin_lock(&imp->imp_lock);
1590 list_del_init(&req->rq_list);
1591 spin_unlock(&imp->imp_lock);
1593 /* If the reply was received normally, this just grabs the spinlock
1594 * (ensuring the reply callback has returned), sees that
1595 * req->rq_receiving_reply is clear and returns. */
1596 ptlrpc_unregister_reply (req);
1599 GOTO(out, rc = -EIO);
1601 /* Resend if we need to, unless we were interrupted. */
1602 if (req->rq_resend && !req->rq_intr) {
1603 /* ...unless we were specifically told otherwise. */
1604 if (req->rq_no_resend)
1605 GOTO(out, rc = -ETIMEDOUT);
1606 spin_lock(&imp->imp_lock);
1611 /* Should only be interrupted if we timed out. */
1612 if (!req->rq_timedout)
1613 DEBUG_REQ(D_ERROR, req,
1614 "rq_intr set but rq_timedout not");
1615 GOTO(out, rc = -EINTR);
1618 if (req->rq_timedout) { /* non-recoverable timeout */
1619 GOTO(out, rc = -ETIMEDOUT);
1622 if (!req->rq_replied) {
1623 /* How can this be? -eeb */
1624 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1626 GOTO(out, rc = req->rq_status);
1629 rc = after_reply(req);
1630 /* NB may return +ve success rc */
1631 if (req->rq_resend) {
1632 spin_lock(&imp->imp_lock);
1637 if (req->rq_bulk != NULL) {
1639 /* success so far. Note that anything going wrong
1640 * with bulk now, is EXTREMELY strange, since the
1641 * server must have believed that the bulk
1642 * tranferred OK before she replied with success to
1644 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1645 brc = l_wait_event(req->rq_reply_waitq,
1646 !ptlrpc_bulk_active(req->rq_bulk),
1648 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1650 LASSERT(brc == -ETIMEDOUT);
1651 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1653 } else if (!req->rq_bulk->bd_success) {
1654 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1659 ptlrpc_unregister_bulk (req);
1662 LASSERT(!req->rq_receiving_reply);
1663 req->rq_phase = RQ_PHASE_INTERPRET;
1665 atomic_dec(&imp->imp_inflight);
1666 cfs_waitq_signal(&imp->imp_recovery_waitq);
1670 struct ptlrpc_replay_async_args {
1672 int praa_old_status;
1675 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1676 void * data, int rc)
1678 struct ptlrpc_replay_async_args *aa = data;
1679 struct obd_import *imp = req->rq_import;
1682 atomic_dec(&imp->imp_replay_inflight);
1684 if (!req->rq_replied) {
1685 CERROR("request replay timed out, restarting recovery\n");
1686 GOTO(out, rc = -ETIMEDOUT);
1689 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1690 lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN)
1691 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1693 /* The transno had better not change over replay. */
1694 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1695 lustre_msg_get_transno(req->rq_repmsg));
1697 DEBUG_REQ(D_HA, req, "got rep");
1699 /* let the callback do fixups, possibly including in the request */
1700 if (req->rq_replay_cb)
1701 req->rq_replay_cb(req);
1703 if (req->rq_replied &&
1704 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1705 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1706 lustre_msg_get_status(req->rq_repmsg),
1707 aa->praa_old_status);
1709 /* Put it back for re-replay. */
1710 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1713 spin_lock(&imp->imp_lock);
1714 imp->imp_last_replay_transno = req->rq_transno;
1715 spin_unlock(&imp->imp_lock);
1717 /* continue with recovery */
1718 rc = ptlrpc_import_recovery_state_machine(imp);
1720 req->rq_send_state = aa->praa_old_state;
1723 /* this replay failed, so restart recovery */
1724 ptlrpc_connect_import(imp, NULL);
1730 int ptlrpc_replay_req(struct ptlrpc_request *req)
1732 struct ptlrpc_replay_async_args *aa;
1735 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1737 /* Not handling automatic bulk replay yet (or ever?) */
1738 LASSERT(req->rq_bulk == NULL);
1740 DEBUG_REQ(D_HA, req, "REPLAY");
1742 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1743 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1744 memset(aa, 0, sizeof *aa);
1746 /* Prepare request to be resent with ptlrpcd */
1747 aa->praa_old_state = req->rq_send_state;
1748 req->rq_send_state = LUSTRE_IMP_REPLAY;
1749 req->rq_phase = RQ_PHASE_NEW;
1750 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
1753 req->rq_interpret_reply = ptlrpc_replay_interpret;
1754 atomic_inc(&req->rq_import->imp_replay_inflight);
1755 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1757 ptlrpcd_add_req(req);
1761 void ptlrpc_abort_inflight(struct obd_import *imp)
1763 struct list_head *tmp, *n;
1766 /* Make sure that no new requests get processed for this import.
1767 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1768 * this flag and then putting requests on sending_list or delayed_list.
1770 spin_lock(&imp->imp_lock);
1772 /* XXX locking? Maybe we should remove each request with the list
1773 * locked? Also, how do we know if the requests on the list are
1774 * being freed at this time?
1776 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1777 struct ptlrpc_request *req =
1778 list_entry(tmp, struct ptlrpc_request, rq_list);
1780 DEBUG_REQ(D_HA, req, "inflight");
1782 spin_lock (&req->rq_lock);
1783 if (req->rq_import_generation < imp->imp_generation) {
1785 ptlrpc_wake_client_req(req);
1787 spin_unlock (&req->rq_lock);
1790 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1791 struct ptlrpc_request *req =
1792 list_entry(tmp, struct ptlrpc_request, rq_list);
1794 DEBUG_REQ(D_HA, req, "aborting waiting req");
1796 spin_lock (&req->rq_lock);
1797 if (req->rq_import_generation < imp->imp_generation) {
1799 ptlrpc_wake_client_req(req);
1801 spin_unlock (&req->rq_lock);
1804 /* Last chance to free reqs left on the replay list, but we
1805 * will still leak reqs that haven't committed. */
1806 if (imp->imp_replayable)
1807 ptlrpc_free_committed(imp);
1809 spin_unlock(&imp->imp_lock);
1814 static __u64 ptlrpc_last_xid = 0;
1815 spinlock_t ptlrpc_last_xid_lock;
1817 __u64 ptlrpc_next_xid(void)
1820 spin_lock(&ptlrpc_last_xid_lock);
1821 tmp = ++ptlrpc_last_xid;
1822 spin_unlock(&ptlrpc_last_xid_lock);
1826 __u64 ptlrpc_sample_next_xid(void)
1829 spin_lock(&ptlrpc_last_xid_lock);
1830 tmp = ptlrpc_last_xid + 1;
1831 spin_unlock(&ptlrpc_last_xid_lock);
1834 EXPORT_SYMBOL(ptlrpc_sample_next_xid);