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 Lustre, http://www.lustre.org.
8 * Lustre is free software; you can redistribute it and/or
9 * modify it under the terms of version 2 of the GNU General Public
10 * License as published by the Free Software Foundation.
12 * Lustre is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 * GNU General Public License for more details.
17 * You should have received a copy of the GNU General Public License
18 * along with Lustre; if not, write to the Free Software
19 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #define DEBUG_SUBSYSTEM S_RPC
27 #include <liblustre.h>
30 #include <linux/obd_support.h>
31 #include <linux/obd_class.h>
32 #include <linux/lustre_lib.h>
33 #include <linux/lustre_ha.h>
34 #include <linux/lustre_import.h>
36 #include "ptlrpc_internal.h"
38 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
39 struct ptlrpc_client *cl)
41 cl->cli_request_portal = req_portal;
42 cl->cli_reply_portal = rep_portal;
46 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
48 struct ptlrpc_connection *c;
49 struct ptlrpc_peer peer;
52 err = ptlrpc_uuid_to_peer(uuid, &peer);
54 CERROR("cannot find peer %s!\n", uuid->uuid);
58 c = ptlrpc_get_connection(&peer, uuid);
60 memcpy(c->c_remote_uuid.uuid,
61 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
64 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
69 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
70 struct obd_uuid *uuid)
72 struct ptlrpc_peer peer;
75 err = ptlrpc_uuid_to_peer(uuid, &peer);
77 CERROR("cannot find peer %s!\n", uuid->uuid);
81 memcpy(&conn->c_peer, &peer, sizeof (peer));
85 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
87 struct ptlrpc_bulk_desc *desc;
89 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
93 spin_lock_init(&desc->bd_lock);
94 init_waitqueue_head(&desc->bd_waitq);
95 desc->bd_max_iov = npages;
96 desc->bd_iov_count = 0;
97 desc->bd_md_h = PTL_INVALID_HANDLE;
98 desc->bd_portal = portal;
104 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
105 int npages, int type, int portal)
107 struct obd_import *imp = req->rq_import;
108 struct ptlrpc_bulk_desc *desc;
110 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
111 desc = new_bulk(npages, type, portal);
115 desc->bd_import_generation = req->rq_import_generation;
116 desc->bd_import = class_import_get(imp);
119 desc->bd_cbid.cbid_fn = client_bulk_callback;
120 desc->bd_cbid.cbid_arg = desc;
122 /* This makes req own desc, and free it when she frees herself */
128 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp (struct ptlrpc_request *req,
129 int npages, int type, int portal)
131 struct obd_export *exp = req->rq_export;
132 struct ptlrpc_bulk_desc *desc;
134 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
136 desc = new_bulk(npages, type, portal);
140 desc->bd_export = class_export_get(exp);
143 desc->bd_cbid.cbid_fn = server_bulk_callback;
144 desc->bd_cbid.cbid_arg = desc;
146 /* NB we don't assign rq_bulk here; server-side requests are
147 * re-used, and the handler frees the bulk desc explicitly. */
152 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
153 struct page *page, int pageoffset, int len)
155 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
156 LASSERT(page != NULL);
157 LASSERT(pageoffset >= 0);
159 LASSERT(pageoffset + len <= PAGE_SIZE);
163 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
166 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
170 LASSERT(desc != NULL);
171 LASSERT(desc->bd_iov_count != 0x5a5a5a5a); /* not freed already */
172 LASSERT(!desc->bd_network_rw); /* network hands off or */
173 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
175 class_export_put(desc->bd_export);
177 class_import_put(desc->bd_import);
179 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
180 bd_iov[desc->bd_max_iov]));
184 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, int opcode,
185 int count, int *lengths, char **bufs)
187 struct ptlrpc_request *request;
191 LASSERT((unsigned long)imp > 0x1000);
193 OBD_ALLOC(request, sizeof(*request));
195 CERROR("request allocation out of memory\n");
199 rc = lustre_pack_request(request, count, lengths, bufs);
201 CERROR("cannot pack request %d\n", rc);
202 OBD_FREE(request, sizeof(*request));
206 if (imp->imp_server_timeout)
207 request->rq_timeout = obd_timeout / 2;
209 request->rq_timeout = obd_timeout;
210 request->rq_send_state = LUSTRE_IMP_FULL;
211 request->rq_type = PTL_RPC_MSG_REQUEST;
212 request->rq_import = class_import_get(imp);
214 request->rq_req_cbid.cbid_fn = request_out_callback;
215 request->rq_req_cbid.cbid_arg = request;
217 request->rq_reply_cbid.cbid_fn = reply_in_callback;
218 request->rq_reply_cbid.cbid_arg = request;
220 request->rq_phase = RQ_PHASE_NEW;
222 /* XXX FIXME bug 249 */
223 request->rq_request_portal = imp->imp_client->cli_request_portal;
224 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
226 spin_lock_init(&request->rq_lock);
227 INIT_LIST_HEAD(&request->rq_list);
228 INIT_LIST_HEAD(&request->rq_replay_list);
229 INIT_LIST_HEAD(&request->rq_set_chain);
230 init_waitqueue_head(&request->rq_reply_waitq);
231 request->rq_xid = ptlrpc_next_xid();
232 atomic_set(&request->rq_refcount, 1);
234 request->rq_reqmsg->opc = opcode;
235 request->rq_reqmsg->flags = 0;
240 struct ptlrpc_request_set *ptlrpc_prep_set(void)
242 struct ptlrpc_request_set *set;
244 OBD_ALLOC(set, sizeof *set);
247 INIT_LIST_HEAD(&set->set_requests);
248 init_waitqueue_head(&set->set_waitq);
249 set->set_remaining = 0;
250 spin_lock_init(&set->set_new_req_lock);
251 INIT_LIST_HEAD(&set->set_new_requests);
256 /* Finish with this set; opposite of prep_set. */
257 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
259 struct list_head *tmp;
260 struct list_head *next;
265 /* Requests on the set should either all be completed, or all be new */
266 expected_phase = (set->set_remaining == 0) ?
267 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
268 list_for_each (tmp, &set->set_requests) {
269 struct ptlrpc_request *req =
270 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
272 LASSERT(req->rq_phase == expected_phase);
276 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
278 list_for_each_safe(tmp, next, &set->set_requests) {
279 struct ptlrpc_request *req =
280 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
281 list_del_init(&req->rq_set_chain);
283 LASSERT(req->rq_phase == expected_phase);
285 if (req->rq_phase == RQ_PHASE_NEW) {
287 if (req->rq_interpret_reply != NULL) {
288 int (*interpreter)(struct ptlrpc_request *,
290 req->rq_interpret_reply;
292 /* higher level (i.e. LOV) failed;
293 * let the sub reqs clean up */
294 req->rq_status = -EBADR;
295 interpreter(req, &req->rq_async_args,
298 set->set_remaining--;
302 ptlrpc_req_finished (req);
305 LASSERT(set->set_remaining == 0);
307 OBD_FREE(set, sizeof(*set));
311 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
312 struct ptlrpc_request *req)
314 /* The set takes over the caller's request reference */
315 list_add_tail(&req->rq_set_chain, &set->set_requests);
317 set->set_remaining++;
318 atomic_inc(&req->rq_import->imp_inflight);
321 /* lock so many callers can add things, the context that owns the set
322 * is supposed to notice these and move them into the set proper. */
323 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
324 struct ptlrpc_request *req)
327 spin_lock_irqsave(&set->set_new_req_lock, flags);
328 /* The set takes over the caller's request reference */
329 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
331 spin_unlock_irqrestore(&set->set_new_req_lock, flags);
335 * Based on the current state of the import, determine if the request
336 * can be sent, is an error, or should be delayed.
338 * Returns true if this request should be delayed. If false, and
339 * *status is set, then the request can not be sent and *status is the
340 * error code. If false and status is 0, then request can be sent.
342 * The imp->imp_lock must be held.
344 static int ptlrpc_import_delay_req(struct obd_import *imp,
345 struct ptlrpc_request *req, int *status)
350 LASSERT (status != NULL);
353 if (imp->imp_state == LUSTRE_IMP_NEW) {
354 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
358 else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
359 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
362 /* allow CONNECT even if import is invalid */
363 else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
364 imp->imp_state == LUSTRE_IMP_CONNECTING) {
368 * If the import has been invalidated (such as by an OST failure), the
369 * request must fail with -EIO.
371 else if (imp->imp_invalid) {
372 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
375 else if (req->rq_import_generation != imp->imp_generation) {
376 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
379 else if (req->rq_send_state != imp->imp_state) {
380 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake
382 *status = -EWOULDBLOCK;
390 static int ptlrpc_check_reply(struct ptlrpc_request *req)
396 /* serialise with network callback */
397 spin_lock_irqsave (&req->rq_lock, flags);
399 if (req->rq_replied) {
400 DEBUG_REQ(D_NET, req, "REPLIED:");
404 if (req->rq_net_err && !req->rq_timedout) {
405 spin_unlock_irqrestore (&req->rq_lock, flags);
406 rc = ptlrpc_expire_one_request(req);
407 spin_lock_irqsave (&req->rq_lock, flags);
412 DEBUG_REQ(D_ERROR, req, "ABORTED:");
416 if (req->rq_resend) {
417 DEBUG_REQ(D_ERROR, req, "RESEND:");
421 if (req->rq_restart) {
422 DEBUG_REQ(D_ERROR, req, "RESTART:");
427 spin_unlock_irqrestore (&req->rq_lock, flags);
428 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
432 static int ptlrpc_check_status(struct ptlrpc_request *req)
437 err = req->rq_repmsg->status;
438 if (req->rq_repmsg->type == PTL_RPC_MSG_ERR) {
439 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err == %d",
441 RETURN(err < 0 ? err : -EINVAL);
445 DEBUG_REQ(D_INFO, req, "status is %d", err);
446 } else if (err > 0) {
447 /* XXX: translate this error from net to host */
448 DEBUG_REQ(D_INFO, req, "status is %d", err);
454 static int after_reply(struct ptlrpc_request *req)
457 struct obd_import *imp = req->rq_import;
461 LASSERT(!req->rq_receiving_reply);
463 /* NB Until this point, the whole of the incoming message,
464 * including buflens, status etc is in the sender's byte order. */
467 /* Clear reply swab mask; this is a new reply in sender's byte order */
468 req->rq_rep_swab_mask = 0;
470 LASSERT (req->rq_nob_received <= req->rq_replen);
471 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
473 CERROR("unpack_rep failed: %d\n", rc);
477 if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
478 req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
479 CERROR("invalid packet type received (type=%u)\n",
480 req->rq_repmsg->type);
484 /* Store transno in reqmsg for replay. */
485 req->rq_reqmsg->transno = req->rq_transno = req->rq_repmsg->transno;
487 rc = ptlrpc_check_status(req);
489 /* Either we've been evicted, or the server has failed for
490 * some reason. Try to reconnect, and if that fails, punt to the
492 if (rc == -ENOTCONN) {
493 if (req->rq_send_state != LUSTRE_IMP_FULL ||
494 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
498 ptlrpc_request_handle_notconn(req);
503 if (req->rq_import->imp_replayable) {
504 spin_lock_irqsave(&imp->imp_lock, flags);
505 if (req->rq_replay || req->rq_transno != 0)
506 ptlrpc_retain_replayable_request(req, imp);
507 else if (req->rq_commit_cb != NULL) {
508 spin_unlock_irqrestore(&imp->imp_lock, flags);
509 req->rq_commit_cb(req);
510 spin_lock_irqsave(&imp->imp_lock, flags);
513 if (req->rq_transno > imp->imp_max_transno)
514 imp->imp_max_transno = req->rq_transno;
516 /* Replay-enabled imports return commit-status information. */
517 if (req->rq_repmsg->last_committed)
518 imp->imp_peer_committed_transno =
519 req->rq_repmsg->last_committed;
520 ptlrpc_free_committed(imp);
521 spin_unlock_irqrestore(&imp->imp_lock, flags);
527 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
529 char str[PTL_NALFMT_SIZE];
530 struct obd_import *imp;
535 LASSERT(req->rq_phase == RQ_PHASE_NEW);
536 req->rq_phase = RQ_PHASE_RPC;
538 imp = req->rq_import;
539 spin_lock_irqsave(&imp->imp_lock, flags);
541 req->rq_import_generation = imp->imp_generation;
543 if (ptlrpc_import_delay_req(imp, req, &rc)) {
544 spin_lock (&req->rq_lock);
546 spin_unlock (&req->rq_lock);
548 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
550 req->rq_reqmsg->status,
551 ptlrpc_import_state_name(req->rq_send_state),
552 ptlrpc_import_state_name(imp->imp_state));
553 LASSERT(list_empty (&req->rq_list));
555 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
556 spin_unlock_irqrestore(&imp->imp_lock, flags);
561 spin_unlock_irqrestore(&imp->imp_lock, flags);
563 req->rq_phase = RQ_PHASE_INTERPRET;
567 /* XXX this is the same as ptlrpc_queue_wait */
568 LASSERT(list_empty(&req->rq_list));
569 list_add_tail(&req->rq_list, &imp->imp_sending_list);
570 spin_unlock_irqrestore(&imp->imp_lock, flags);
572 req->rq_reqmsg->status = current->pid;
573 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc"
574 " %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
575 imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
577 imp->imp_connection->c_peer.peer_ni->pni_name,
578 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
579 req->rq_reqmsg->opc);
581 rc = ptl_send_rpc(req);
583 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
590 int ptlrpc_check_set(struct ptlrpc_request_set *set)
592 char str[PTL_NALFMT_SIZE];
594 struct list_head *tmp;
595 int force_timer_recalc = 0;
598 if (set->set_remaining == 0)
601 list_for_each(tmp, &set->set_requests) {
602 struct ptlrpc_request *req =
603 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
604 struct obd_import *imp = req->rq_import;
607 if (req->rq_phase == RQ_PHASE_NEW &&
608 ptlrpc_send_new_req(req)) {
609 force_timer_recalc = 1;
612 if (!(req->rq_phase == RQ_PHASE_RPC ||
613 req->rq_phase == RQ_PHASE_BULK ||
614 req->rq_phase == RQ_PHASE_INTERPRET ||
615 req->rq_phase == RQ_PHASE_COMPLETE)) {
616 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
620 if (req->rq_phase == RQ_PHASE_COMPLETE)
623 if (req->rq_phase == RQ_PHASE_INTERPRET)
624 GOTO(interpret, req->rq_status);
626 if (req->rq_net_err && !req->rq_timedout)
627 ptlrpc_expire_one_request(req);
630 ptlrpc_unregister_reply(req);
631 if (req->rq_status == 0)
632 req->rq_status = -EIO;
633 req->rq_phase = RQ_PHASE_INTERPRET;
635 spin_lock_irqsave(&imp->imp_lock, flags);
636 list_del_init(&req->rq_list);
637 spin_unlock_irqrestore(&imp->imp_lock, flags);
639 GOTO(interpret, req->rq_status);
642 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
643 * will only be set after rq_timedout, but the oig waiting
644 * path sets rq_intr irrespective of whether ptlrpcd has
645 * seen a timeout. our policy is to only interpret
646 * interrupted rpcs after they have timed out */
647 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
648 /* NB could be on delayed list */
649 ptlrpc_unregister_reply(req);
650 req->rq_status = -EINTR;
651 req->rq_phase = RQ_PHASE_INTERPRET;
653 spin_lock_irqsave(&imp->imp_lock, flags);
654 list_del_init(&req->rq_list);
655 spin_unlock_irqrestore(&imp->imp_lock, flags);
657 GOTO(interpret, req->rq_status);
660 if (req->rq_phase == RQ_PHASE_RPC) {
661 if (req->rq_waiting || req->rq_resend) {
664 spin_lock_irqsave(&imp->imp_lock, flags);
666 if (ptlrpc_import_delay_req(imp, req, &status)){
667 spin_unlock_irqrestore(&imp->imp_lock,
672 list_del_init(&req->rq_list);
674 req->rq_status = status;
675 req->rq_phase = RQ_PHASE_INTERPRET;
676 spin_unlock_irqrestore(&imp->imp_lock,
678 GOTO(interpret, req->rq_status);
680 if (req->rq_no_resend) {
681 req->rq_status = -ENOTCONN;
682 req->rq_phase = RQ_PHASE_INTERPRET;
683 spin_unlock_irqrestore(&imp->imp_lock,
685 GOTO(interpret, req->rq_status);
687 list_add_tail(&req->rq_list,
688 &imp->imp_sending_list);
690 spin_unlock_irqrestore(&imp->imp_lock, flags);
693 if (req->rq_resend) {
694 lustre_msg_add_flags(req->rq_reqmsg,
696 ptlrpc_unregister_reply(req);
698 __u64 old_xid = req->rq_xid;
700 ptlrpc_unregister_bulk (req);
702 /* ensure previous bulk fails */
703 req->rq_xid = ptlrpc_next_xid();
704 CDEBUG(D_HA, "resend bulk "
707 old_xid, req->rq_xid);
711 rc = ptl_send_rpc(req);
713 DEBUG_REQ(D_HA, req, "send failed (%d)",
715 force_timer_recalc = 1;
718 /* need to reset the timeout */
719 force_timer_recalc = 1;
722 /* Still waiting for a reply? */
723 if (ptlrpc_client_receiving_reply(req))
726 /* Did we actually receive a reply? */
727 if (!ptlrpc_client_replied(req))
730 spin_lock_irqsave(&imp->imp_lock, flags);
731 list_del_init(&req->rq_list);
732 spin_unlock_irqrestore(&imp->imp_lock, flags);
734 req->rq_status = after_reply(req);
735 if (req->rq_resend) {
736 /* Add this req to the delayed list so
737 it can be errored if the import is
738 evicted after recovery. */
739 spin_lock_irqsave (&req->rq_lock, flags);
740 list_add_tail(&req->rq_list,
741 &imp->imp_delayed_list);
742 spin_unlock_irqrestore(&req->rq_lock, flags);
746 /* If there is no bulk associated with this request,
747 * then we're done and should let the interpreter
748 * process the reply. Similarly if the RPC returned
749 * an error, and therefore the bulk will never arrive.
751 if (req->rq_bulk == NULL || req->rq_status != 0) {
752 req->rq_phase = RQ_PHASE_INTERPRET;
753 GOTO(interpret, req->rq_status);
756 req->rq_phase = RQ_PHASE_BULK;
759 LASSERT(req->rq_phase == RQ_PHASE_BULK);
760 if (ptlrpc_bulk_active(req->rq_bulk))
763 if (!req->rq_bulk->bd_success) {
764 /* The RPC reply arrived OK, but the bulk screwed
765 * up! Dead wierd since the server told us the RPC
766 * was good after getting the REPLY for her GET or
767 * the ACK for her PUT. */
768 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
772 req->rq_phase = RQ_PHASE_INTERPRET;
775 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
776 LASSERT(!req->rq_receiving_reply);
778 ptlrpc_unregister_reply(req);
779 if (req->rq_bulk != NULL)
780 ptlrpc_unregister_bulk (req);
782 req->rq_phase = RQ_PHASE_COMPLETE;
784 if (req->rq_interpret_reply != NULL) {
785 int (*interpreter)(struct ptlrpc_request *,void *,int) =
786 req->rq_interpret_reply;
787 req->rq_status = interpreter(req, &req->rq_async_args,
791 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
792 "opc %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
793 imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
795 imp->imp_connection->c_peer.peer_ni->pni_name,
796 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
797 req->rq_reqmsg->opc);
799 set->set_remaining--;
801 atomic_dec(&imp->imp_inflight);
802 wake_up(&imp->imp_recovery_waitq);
805 /* If we hit an error, we want to recover promptly. */
806 RETURN(set->set_remaining == 0 || force_timer_recalc);
809 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
812 struct obd_import *imp = req->rq_import;
815 DEBUG_REQ(D_ERROR, req, "timeout");
817 spin_lock_irqsave (&req->rq_lock, flags);
818 req->rq_timedout = 1;
819 spin_unlock_irqrestore (&req->rq_lock, flags);
821 ptlrpc_unregister_reply (req);
823 if (req->rq_bulk != NULL)
824 ptlrpc_unregister_bulk (req);
827 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
831 /* The DLM server doesn't want recovery run on its imports. */
832 if (imp->imp_dlm_fake)
835 /* If this request is for recovery or other primordial tasks,
836 * then error it out here. */
837 if (req->rq_send_state != LUSTRE_IMP_FULL ||
838 imp->imp_obd->obd_no_recov) {
839 spin_lock_irqsave (&req->rq_lock, flags);
840 req->rq_status = -ETIMEDOUT;
842 spin_unlock_irqrestore (&req->rq_lock, flags);
846 ptlrpc_fail_import(imp, req->rq_import_generation);
851 int ptlrpc_expired_set(void *data)
853 struct ptlrpc_request_set *set = data;
854 struct list_head *tmp;
855 time_t now = LTIME_S (CURRENT_TIME);
858 LASSERT(set != NULL);
860 /* A timeout expired; see which reqs it applies to... */
861 list_for_each (tmp, &set->set_requests) {
862 struct ptlrpc_request *req =
863 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
865 /* request in-flight? */
866 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting
867 && !req->rq_resend) ||
868 (req->rq_phase == RQ_PHASE_BULK)))
871 if (req->rq_timedout || /* already dealt with */
872 req->rq_sent + req->rq_timeout > now) /* not expired */
875 /* deal with this guy */
876 ptlrpc_expire_one_request (req);
879 /* When waiting for a whole set, we always to break out of the
880 * sleep so we can recalculate the timeout, or enable interrupts
881 * iff everyone's timed out.
886 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
889 spin_lock_irqsave(&req->rq_lock, flags);
891 spin_unlock_irqrestore(&req->rq_lock, flags);
894 void ptlrpc_interrupted_set(void *data)
896 struct ptlrpc_request_set *set = data;
897 struct list_head *tmp;
899 LASSERT(set != NULL);
900 CERROR("INTERRUPTED SET %p\n", set);
902 list_for_each(tmp, &set->set_requests) {
903 struct ptlrpc_request *req =
904 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
906 if (req->rq_phase != RQ_PHASE_RPC)
909 ptlrpc_mark_interrupted(req);
913 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
915 struct list_head *tmp;
916 time_t now = LTIME_S(CURRENT_TIME);
919 struct ptlrpc_request *req;
922 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
924 list_for_each(tmp, &set->set_requests) {
925 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
927 /* request in-flight? */
928 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
929 (req->rq_phase == RQ_PHASE_BULK)))
932 if (req->rq_timedout) /* already timed out */
935 deadline = req->rq_sent + req->rq_timeout;
936 if (deadline <= now) /* actually expired already */
937 timeout = 1; /* ASAP */
938 else if (timeout == 0 || timeout > deadline - now)
939 timeout = deadline - now;
945 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
947 struct list_head *tmp;
948 struct ptlrpc_request *req;
949 struct l_wait_info lwi;
953 LASSERT(!list_empty(&set->set_requests));
954 list_for_each(tmp, &set->set_requests) {
955 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
956 if (req->rq_phase == RQ_PHASE_NEW)
957 (void)ptlrpc_send_new_req(req);
961 timeout = ptlrpc_set_next_timeout(set);
963 /* wait until all complete, interrupted, or an in-flight
965 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
967 lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
969 ptlrpc_interrupted_set, set);
970 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
972 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
974 /* -EINTR => all requests have been flagged rq_intr so next
976 * -ETIMEOUTD => someone timed out. When all reqs have
977 * timed out, signals are enabled allowing completion with
979 * I don't really care if we go once more round the loop in
980 * the error cases -eeb. */
983 LASSERT(set->set_remaining == 0);
986 list_for_each(tmp, &set->set_requests) {
987 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
989 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
990 if (req->rq_status != 0)
994 if (set->set_interpret != NULL) {
995 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
997 rc = interpreter (set, &set->set_args, rc);
1003 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1006 if (request == NULL) {
1011 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1012 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1013 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1014 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1016 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1017 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1018 if (request->rq_import != NULL) {
1019 unsigned long flags = 0;
1021 spin_lock_irqsave(&request->rq_import->imp_lock, flags);
1022 list_del_init(&request->rq_replay_list);
1024 spin_unlock_irqrestore(&request->rq_import->imp_lock,
1027 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1029 if (atomic_read(&request->rq_refcount) != 0) {
1030 DEBUG_REQ(D_ERROR, request,
1031 "freeing request with nonzero refcount");
1035 if (request->rq_repmsg != NULL) {
1036 OBD_FREE(request->rq_repmsg, request->rq_replen);
1037 request->rq_repmsg = NULL;
1039 if (request->rq_reqmsg != NULL) {
1040 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1041 request->rq_reqmsg = NULL;
1043 if (request->rq_export != NULL) {
1044 class_export_put(request->rq_export);
1045 request->rq_export = NULL;
1047 if (request->rq_import != NULL) {
1048 class_import_put(request->rq_import);
1049 request->rq_import = NULL;
1051 if (request->rq_bulk != NULL)
1052 ptlrpc_free_bulk(request->rq_bulk);
1054 OBD_FREE(request, sizeof(*request));
1058 void ptlrpc_free_req(struct ptlrpc_request *request)
1060 __ptlrpc_free_req(request, 0);
1063 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1064 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1066 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1067 (void)__ptlrpc_req_finished(request, 1);
1070 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1073 if (request == NULL)
1076 if (request == LP_POISON ||
1077 request->rq_reqmsg == LP_POISON) {
1078 CERROR("dereferencing freed request (bug 575)\n");
1083 DEBUG_REQ(D_INFO, request, "refcount now %u",
1084 atomic_read(&request->rq_refcount) - 1);
1086 if (atomic_dec_and_test(&request->rq_refcount)) {
1087 __ptlrpc_free_req(request, locked);
1094 void ptlrpc_req_finished(struct ptlrpc_request *request)
1096 __ptlrpc_req_finished(request, 0);
1099 /* Disengage the client's reply buffer from the network
1100 * NB does _NOT_ unregister any client-side bulk.
1101 * IDEMPOTENT, but _not_ safe against concurrent callers.
1102 * The request owner (i.e. the thread doing the I/O) must call...
1104 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1107 wait_queue_head_t *wq;
1108 struct l_wait_info lwi;
1110 LASSERT(!in_interrupt ()); /* might sleep */
1112 if (!ptlrpc_client_receiving_reply(request))
1115 PtlMDUnlink (request->rq_reply_md_h);
1117 /* We have to l_wait_event() whatever the result, to give liblustre
1118 * a chance to run reply_in_callback() */
1120 if (request->rq_set == NULL)
1121 wq = &request->rq_set->set_waitq;
1123 wq = &request->rq_reply_waitq;
1126 /* Network access will complete in finite time but the HUGE
1127 * timeout lets us CWARN for visibility of sluggish NALs */
1128 lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
1129 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1133 LASSERT (rc == -ETIMEDOUT);
1134 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1138 /* caller must hold imp->imp_lock */
1139 void ptlrpc_free_committed(struct obd_import *imp)
1141 struct list_head *tmp, *saved;
1142 struct ptlrpc_request *req;
1143 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1146 LASSERT(imp != NULL);
1148 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1150 CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
1151 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1153 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1154 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1156 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1157 LASSERT(req != last_req);
1160 if (req->rq_import_generation < imp->imp_generation) {
1161 DEBUG_REQ(D_HA, req, "freeing request with old gen");
1165 if (req->rq_replay) {
1166 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1170 /* not yet committed */
1171 if (req->rq_transno > imp->imp_peer_committed_transno) {
1172 DEBUG_REQ(D_HA, req, "stopping search");
1176 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1177 imp->imp_peer_committed_transno);
1179 if (req->rq_commit_cb != NULL)
1180 req->rq_commit_cb(req);
1181 list_del_init(&req->rq_replay_list);
1182 __ptlrpc_req_finished(req, 1);
1189 void ptlrpc_cleanup_client(struct obd_import *imp)
1196 void ptlrpc_resend_req(struct ptlrpc_request *req)
1198 unsigned long flags;
1200 DEBUG_REQ(D_HA, req, "going to resend");
1201 req->rq_reqmsg->handle.cookie = 0;
1202 req->rq_status = -EAGAIN;
1204 spin_lock_irqsave (&req->rq_lock, flags);
1206 req->rq_net_err = 0;
1207 req->rq_timedout = 0;
1209 __u64 old_xid = req->rq_xid;
1211 /* ensure previous bulk fails */
1212 req->rq_xid = ptlrpc_next_xid();
1213 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1214 old_xid, req->rq_xid);
1216 ptlrpc_wake_client_req(req);
1217 spin_unlock_irqrestore (&req->rq_lock, flags);
1220 /* XXX: this function and rq_status are currently unused */
1221 void ptlrpc_restart_req(struct ptlrpc_request *req)
1223 unsigned long flags;
1225 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1226 req->rq_status = -ERESTARTSYS;
1228 spin_lock_irqsave (&req->rq_lock, flags);
1229 req->rq_restart = 1;
1230 req->rq_timedout = 0;
1231 ptlrpc_wake_client_req(req);
1232 spin_unlock_irqrestore (&req->rq_lock, flags);
1235 static int expired_request(void *data)
1237 struct ptlrpc_request *req = data;
1240 RETURN(ptlrpc_expire_one_request(req));
1243 static void interrupted_request(void *data)
1245 unsigned long flags;
1247 struct ptlrpc_request *req = data;
1248 DEBUG_REQ(D_HA, req, "request interrupted");
1249 spin_lock_irqsave (&req->rq_lock, flags);
1251 spin_unlock_irqrestore (&req->rq_lock, flags);
1254 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1257 atomic_inc(&req->rq_refcount);
1261 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1262 struct obd_import *imp)
1264 struct list_head *tmp;
1266 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1268 /* don't re-add requests that have been replayed */
1269 if (!list_empty(&req->rq_replay_list))
1272 lustre_msg_add_flags(req->rq_reqmsg,
1275 LASSERT(imp->imp_replayable);
1276 /* Balanced in ptlrpc_free_committed, usually. */
1277 ptlrpc_request_addref(req);
1278 list_for_each_prev(tmp, &imp->imp_replay_list) {
1279 struct ptlrpc_request *iter =
1280 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1282 /* We may have duplicate transnos if we create and then
1283 * open a file, or for closes retained if to match creating
1284 * opens, so use req->rq_xid as a secondary key.
1285 * (See bugs 684, 685, and 428.)
1286 * XXX no longer needed, but all opens need transnos!
1288 if (iter->rq_transno > req->rq_transno)
1291 if (iter->rq_transno == req->rq_transno) {
1292 LASSERT(iter->rq_xid != req->rq_xid);
1293 if (iter->rq_xid > req->rq_xid)
1297 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1301 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1304 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1306 char str[PTL_NALFMT_SIZE];
1309 struct l_wait_info lwi;
1310 struct obd_import *imp = req->rq_import;
1311 unsigned long flags;
1315 LASSERT(req->rq_set == NULL);
1316 LASSERT(!req->rq_receiving_reply);
1317 atomic_inc(&imp->imp_inflight);
1319 /* for distributed debugging */
1320 req->rq_reqmsg->status = current->pid;
1321 LASSERT(imp->imp_obd != NULL);
1322 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
1323 "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1324 imp->imp_obd->obd_uuid.uuid,
1325 req->rq_reqmsg->status, req->rq_xid,
1326 imp->imp_connection->c_peer.peer_ni->pni_name,
1327 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1328 req->rq_reqmsg->opc);
1330 /* Mark phase here for a little debug help */
1331 req->rq_phase = RQ_PHASE_RPC;
1333 spin_lock_irqsave(&imp->imp_lock, flags);
1334 req->rq_import_generation = imp->imp_generation;
1336 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1337 list_del(&req->rq_list);
1339 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1340 spin_unlock_irqrestore(&imp->imp_lock, flags);
1342 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1344 ptlrpc_import_state_name(req->rq_send_state),
1345 ptlrpc_import_state_name(imp->imp_state));
1346 lwi = LWI_INTR(interrupted_request, req);
1347 rc = l_wait_event(req->rq_reply_waitq,
1348 (req->rq_send_state == imp->imp_state ||
1351 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
1353 ptlrpc_import_state_name(imp->imp_state),
1354 ptlrpc_import_state_name(req->rq_send_state),
1357 spin_lock_irqsave(&imp->imp_lock, flags);
1358 list_del_init(&req->rq_list);
1363 else if (req->rq_intr) {
1366 else if (req->rq_no_resend) {
1367 spin_unlock_irqrestore(&imp->imp_lock, flags);
1368 GOTO(out, rc = -ETIMEDOUT);
1376 list_del_init(&req->rq_list);
1377 spin_unlock_irqrestore(&imp->imp_lock, flags);
1378 req->rq_status = rc; // XXX this ok?
1382 if (req->rq_resend) {
1383 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1385 if (req->rq_bulk != NULL)
1386 ptlrpc_unregister_bulk (req);
1388 DEBUG_REQ(D_HA, req, "resending: ");
1391 /* XXX this is the same as ptlrpc_set_wait */
1392 LASSERT(list_empty(&req->rq_list));
1393 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1394 spin_unlock_irqrestore(&imp->imp_lock, flags);
1396 rc = ptl_send_rpc(req);
1398 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1401 timeout = MAX(req->rq_timeout * HZ, 1);
1402 DEBUG_REQ(D_NET, req, "-- sleeping");
1404 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1406 l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1407 DEBUG_REQ(D_NET, req, "-- done sleeping");
1409 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
1410 "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1411 imp->imp_obd->obd_uuid.uuid,
1412 req->rq_reqmsg->status, req->rq_xid,
1413 imp->imp_connection->c_peer.peer_ni->pni_name,
1414 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1415 req->rq_reqmsg->opc);
1417 spin_lock_irqsave(&imp->imp_lock, flags);
1418 list_del_init(&req->rq_list);
1419 spin_unlock_irqrestore(&imp->imp_lock, flags);
1421 /* If the reply was received normally, this just grabs the spinlock
1422 * (ensuring the reply callback has returned), sees that
1423 * req->rq_receiving_reply is clear and returns. */
1424 ptlrpc_unregister_reply (req);
1427 GOTO(out, rc = -EIO);
1429 /* Resend if we need to, unless we were interrupted. */
1430 if (req->rq_resend && !req->rq_intr) {
1431 /* ...unless we were specifically told otherwise. */
1432 if (req->rq_no_resend)
1433 GOTO(out, rc = -ETIMEDOUT);
1434 spin_lock_irqsave(&imp->imp_lock, flags);
1439 /* Should only be interrupted if we timed out. */
1440 if (!req->rq_timedout)
1441 DEBUG_REQ(D_ERROR, req,
1442 "rq_intr set but rq_timedout not");
1443 GOTO(out, rc = -EINTR);
1446 if (req->rq_timedout) { /* non-recoverable timeout */
1447 GOTO(out, rc = -ETIMEDOUT);
1450 if (!req->rq_replied) {
1451 /* How can this be? -eeb */
1452 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1454 GOTO(out, rc = req->rq_status);
1457 rc = after_reply (req);
1458 /* NB may return +ve success rc */
1459 if (req->rq_resend) {
1460 spin_lock_irqsave(&imp->imp_lock, flags);
1465 if (req->rq_bulk != NULL) {
1467 /* success so far. Note that anything going wrong
1468 * with bulk now, is EXTREMELY strange, since the
1469 * server must have believed that the bulk
1470 * tranferred OK before she replied with success to
1472 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1473 brc = l_wait_event(req->rq_reply_waitq,
1474 !ptlrpc_bulk_active(req->rq_bulk),
1476 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1478 LASSERT(brc == -ETIMEDOUT);
1479 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1481 } else if (!req->rq_bulk->bd_success) {
1482 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1487 ptlrpc_unregister_bulk (req);
1490 LASSERT(!req->rq_receiving_reply);
1491 req->rq_phase = RQ_PHASE_INTERPRET;
1493 atomic_dec(&imp->imp_inflight);
1494 wake_up(&imp->imp_recovery_waitq);
1498 struct ptlrpc_replay_async_args {
1500 int praa_old_status;
1503 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1504 void * data, int rc)
1506 struct ptlrpc_replay_async_args *aa = data;
1507 struct obd_import *imp = req->rq_import;
1508 unsigned long flags;
1510 atomic_dec(&imp->imp_replay_inflight);
1512 if (!req->rq_replied) {
1513 CERROR("request replay timed out, restarting recovery\n");
1514 GOTO(out, rc = -ETIMEDOUT);
1518 /* Clear reply swab mask; this is a new reply in sender's byte order */
1519 req->rq_rep_swab_mask = 0;
1521 LASSERT (req->rq_nob_received <= req->rq_replen);
1522 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
1524 CERROR("unpack_rep failed: %d\n", rc);
1525 GOTO(out, rc = -EPROTO);
1528 if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
1529 req->rq_repmsg->status == -ENOTCONN)
1530 GOTO(out, rc = req->rq_repmsg->status);
1532 /* The transno had better not change over replay. */
1533 LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
1535 DEBUG_REQ(D_HA, req, "got rep");
1537 /* let the callback do fixups, possibly including in the request */
1538 if (req->rq_replay_cb)
1539 req->rq_replay_cb(req);
1541 if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
1542 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1543 req->rq_repmsg->status, aa->praa_old_status);
1545 /* Put it back for re-replay. */
1546 req->rq_repmsg->status = aa->praa_old_status;
1549 spin_lock_irqsave(&imp->imp_lock, flags);
1550 imp->imp_last_replay_transno = req->rq_transno;
1551 spin_unlock_irqrestore(&imp->imp_lock, flags);
1553 /* continue with recovery */
1554 rc = ptlrpc_import_recovery_state_machine(imp);
1556 req->rq_send_state = aa->praa_old_state;
1559 /* this replay failed, so restart recovery */
1560 ptlrpc_connect_import(imp, NULL);
1566 int ptlrpc_replay_req(struct ptlrpc_request *req)
1568 struct ptlrpc_replay_async_args *aa;
1571 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1573 /* Not handling automatic bulk replay yet (or ever?) */
1574 LASSERT(req->rq_bulk == NULL);
1576 DEBUG_REQ(D_HA, req, "REPLAY");
1578 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1579 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1580 memset(aa, 0, sizeof *aa);
1582 /* Prepare request to be resent with ptlrpcd */
1583 aa->praa_old_state = req->rq_send_state;
1584 req->rq_send_state = LUSTRE_IMP_REPLAY;
1585 req->rq_phase = RQ_PHASE_NEW;
1587 * Q: "How can a req get on the replay list if it wasn't replied?"
1588 * A: "If we failed during the replay of this request, it will still
1589 * be on the list, but rq_replied will have been reset to 0."
1591 if (req->rq_replied) {
1592 aa->praa_old_status = req->rq_repmsg->status;
1594 req->rq_replied = 0;
1597 req->rq_interpret_reply = ptlrpc_replay_interpret;
1598 atomic_inc(&req->rq_import->imp_replay_inflight);
1599 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1601 ptlrpcd_add_req(req);
1605 void ptlrpc_abort_inflight(struct obd_import *imp)
1607 unsigned long flags;
1608 struct list_head *tmp, *n;
1611 /* Make sure that no new requests get processed for this import.
1612 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1613 * this flag and then putting requests on sending_list or delayed_list.
1615 spin_lock_irqsave(&imp->imp_lock, flags);
1617 /* XXX locking? Maybe we should remove each request with the list
1618 * locked? Also, how do we know if the requests on the list are
1619 * being freed at this time?
1621 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1622 struct ptlrpc_request *req =
1623 list_entry(tmp, struct ptlrpc_request, rq_list);
1625 DEBUG_REQ(D_HA, req, "inflight");
1627 spin_lock (&req->rq_lock);
1628 if (req->rq_import_generation < imp->imp_generation) {
1630 ptlrpc_wake_client_req(req);
1632 spin_unlock (&req->rq_lock);
1635 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1636 struct ptlrpc_request *req =
1637 list_entry(tmp, struct ptlrpc_request, rq_list);
1639 DEBUG_REQ(D_HA, req, "aborting waiting req");
1641 spin_lock (&req->rq_lock);
1642 if (req->rq_import_generation < imp->imp_generation) {
1644 ptlrpc_wake_client_req(req);
1646 spin_unlock (&req->rq_lock);
1649 /* Last chance to free reqs left on the replay list, but we
1650 * will still leak reqs that haven't comitted. */
1651 if (imp->imp_replayable)
1652 ptlrpc_free_committed(imp);
1654 spin_unlock_irqrestore(&imp->imp_lock, flags);
1659 static __u64 ptlrpc_last_xid = 0;
1660 static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
1662 __u64 ptlrpc_next_xid(void)
1665 spin_lock(&ptlrpc_last_xid_lock);
1666 tmp = ++ptlrpc_last_xid;
1667 spin_unlock(&ptlrpc_last_xid_lock);