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>
35 #include <linux/lustre_sec.h>
37 #include "ptlrpc_internal.h"
39 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
40 struct ptlrpc_client *cl)
42 cl->cli_request_portal = req_portal;
43 cl->cli_reply_portal = rep_portal;
47 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
49 struct ptlrpc_connection *c;
50 struct ptlrpc_peer peer;
53 err = ptlrpc_uuid_to_peer(uuid, &peer);
55 CERROR("cannot find peer %s!\n", uuid->uuid);
59 c = ptlrpc_get_connection(&peer, uuid);
61 memcpy(c->c_remote_uuid.uuid,
62 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
65 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
70 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
71 struct obd_uuid *uuid)
73 struct ptlrpc_peer peer;
76 err = ptlrpc_uuid_to_peer(uuid, &peer);
78 CERROR("cannot find peer %s!\n", uuid->uuid);
82 memcpy(&conn->c_peer, &peer, sizeof (peer));
86 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
88 struct ptlrpc_bulk_desc *desc;
90 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
94 spin_lock_init(&desc->bd_lock);
95 init_waitqueue_head(&desc->bd_waitq);
96 desc->bd_max_iov = npages;
97 desc->bd_iov_count = 0;
98 desc->bd_md_h = PTL_INVALID_HANDLE;
99 desc->bd_portal = portal;
100 desc->bd_type = type;
105 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
106 int npages, int type, int portal)
108 struct obd_import *imp = req->rq_import;
109 struct ptlrpc_bulk_desc *desc;
111 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
112 desc = new_bulk(npages, type, portal);
116 desc->bd_import_generation = req->rq_import_generation;
117 desc->bd_import = class_import_get(imp);
120 desc->bd_cbid.cbid_fn = client_bulk_callback;
121 desc->bd_cbid.cbid_arg = desc;
123 /* This makes req own desc, and free it when she frees herself */
129 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp (struct ptlrpc_request *req,
130 int npages, int type, int portal)
132 struct obd_export *exp = req->rq_export;
133 struct ptlrpc_bulk_desc *desc;
135 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
137 desc = new_bulk(npages, type, portal);
141 desc->bd_export = class_export_get(exp);
144 desc->bd_cbid.cbid_fn = server_bulk_callback;
145 desc->bd_cbid.cbid_arg = desc;
147 /* NB we don't assign rq_bulk here; server-side requests are
148 * re-used, and the handler frees the bulk desc explicitly. */
153 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
154 struct page *page, int pageoffset, int len)
156 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
157 LASSERT(page != NULL);
158 LASSERT(pageoffset >= 0);
160 LASSERT(pageoffset + len <= PAGE_SIZE);
164 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
167 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
171 LASSERT(desc != NULL);
172 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
173 LASSERT(!desc->bd_network_rw); /* network hands off or */
174 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
176 class_export_put(desc->bd_export);
178 class_import_put(desc->bd_import);
180 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
181 bd_iov[desc->bd_max_iov]));
185 /* FIXME prep_req now should return error code other than NULL. but
186 * this is called everywhere :(
188 struct ptlrpc_request *ptlrpc_prep_req(struct obd_import *imp, __u32 version,
189 int opcode, int count, int *lengths,
192 struct ptlrpc_request *request;
196 LASSERT((unsigned long)imp > 0x1000);
198 OBD_ALLOC(request, sizeof(*request));
200 CERROR("request allocation out of memory\n");
204 request->rq_import = class_import_get(imp);
206 rc = ptlrpcs_req_get_cred(request);
208 CDEBUG(D_SEC, "failed to get credential\n");
212 /* try to refresh the cred. we do this here in order to let fewer
213 * refresh be performed in ptlrpcd context (which might block ptlrpcd).
214 * fail out only if a fatal ptlrpcs error occured.
216 ptlrpcs_req_refresh_cred(request);
217 if (request->rq_ptlrpcs_err)
218 GOTO(out_cred, rc = -EPERM);
220 rc = lustre_pack_request(request, count, lengths, bufs);
222 CERROR("cannot pack request %d\n", rc);
225 request->rq_reqmsg->version |= version;
227 if (imp->imp_server_timeout)
228 request->rq_timeout = obd_timeout / 2;
230 request->rq_timeout = obd_timeout;
232 request->rq_send_state = LUSTRE_IMP_FULL;
233 request->rq_type = PTL_RPC_MSG_REQUEST;
235 request->rq_req_cbid.cbid_fn = request_out_callback;
236 request->rq_req_cbid.cbid_arg = request;
238 request->rq_reply_cbid.cbid_fn = reply_in_callback;
239 request->rq_reply_cbid.cbid_arg = request;
241 request->rq_phase = RQ_PHASE_NEW;
243 /* XXX FIXME bug 249 */
244 request->rq_request_portal = imp->imp_client->cli_request_portal;
245 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
247 spin_lock_init(&request->rq_lock);
248 INIT_LIST_HEAD(&request->rq_list);
249 INIT_LIST_HEAD(&request->rq_replay_list);
250 INIT_LIST_HEAD(&request->rq_set_chain);
251 init_waitqueue_head(&request->rq_reply_waitq);
252 request->rq_xid = ptlrpc_next_xid();
253 atomic_set(&request->rq_refcount, 1);
255 request->rq_reqmsg->opc = opcode;
256 request->rq_reqmsg->flags = 0;
259 ptlrpcs_req_drop_cred(request);
261 class_import_put(imp);
262 OBD_FREE(request, sizeof(*request));
266 struct ptlrpc_request_set *ptlrpc_prep_set(void)
268 struct ptlrpc_request_set *set;
270 OBD_ALLOC(set, sizeof *set);
273 INIT_LIST_HEAD(&set->set_requests);
274 init_waitqueue_head(&set->set_waitq);
275 set->set_remaining = 0;
276 spin_lock_init(&set->set_new_req_lock);
277 INIT_LIST_HEAD(&set->set_new_requests);
282 /* Finish with this set; opposite of prep_set. */
283 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
285 struct list_head *tmp;
286 struct list_head *next;
291 /* Requests on the set should either all be completed, or all be new */
292 expected_phase = (set->set_remaining == 0) ?
293 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
294 list_for_each (tmp, &set->set_requests) {
295 struct ptlrpc_request *req =
296 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
298 LASSERT(req->rq_phase == expected_phase);
302 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
304 list_for_each_safe(tmp, next, &set->set_requests) {
305 struct ptlrpc_request *req =
306 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
307 list_del_init(&req->rq_set_chain);
309 LASSERT(req->rq_phase == expected_phase);
311 if (req->rq_phase == RQ_PHASE_NEW) {
313 if (req->rq_interpret_reply != NULL) {
314 int (*interpreter)(struct ptlrpc_request *,
316 req->rq_interpret_reply;
318 /* higher level (i.e. LOV) failed;
319 * let the sub reqs clean up */
320 req->rq_status = -EBADR;
321 interpreter(req, &req->rq_async_args,
324 set->set_remaining--;
328 ptlrpc_req_finished (req);
331 LASSERT(set->set_remaining == 0);
333 OBD_FREE(set, sizeof(*set));
337 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
338 struct ptlrpc_request *req)
340 /* The set takes over the caller's request reference */
341 list_add_tail(&req->rq_set_chain, &set->set_requests);
343 set->set_remaining++;
344 atomic_inc(&req->rq_import->imp_inflight);
347 /* lock so many callers can add things, the context that owns the set
348 * is supposed to notice these and move them into the set proper. */
349 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
350 struct ptlrpc_request *req)
353 spin_lock_irqsave(&set->set_new_req_lock, flags);
354 /* The set takes over the caller's request reference */
355 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
357 spin_unlock_irqrestore(&set->set_new_req_lock, flags);
361 * Based on the current state of the import, determine if the request
362 * can be sent, is an error, or should be delayed.
364 * Returns true if this request should be delayed. If false, and
365 * *status is set, then the request can not be sent and *status is the
366 * error code. If false and status is 0, then request can be sent.
368 * The imp->imp_lock must be held.
370 static int ptlrpc_import_delay_req(struct obd_import *imp,
371 struct ptlrpc_request *req, int *status)
376 LASSERT (status != NULL);
379 if (imp->imp_state == LUSTRE_IMP_NEW) {
380 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
384 else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
385 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
388 /* allow CONNECT even if import is invalid */
389 else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
390 imp->imp_state == LUSTRE_IMP_CONNECTING) {
394 * If the import has been invalidated (such as by an OST failure), the
395 * request must fail with -EIO.
397 else if (imp->imp_invalid) {
398 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
401 else if (req->rq_import_generation != imp->imp_generation) {
402 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
405 else if (req->rq_send_state != imp->imp_state) {
406 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake
408 *status = -EWOULDBLOCK;
416 static int ptlrpc_check_reply(struct ptlrpc_request *req)
422 /* serialise with network callback */
423 spin_lock_irqsave (&req->rq_lock, flags);
425 if (req->rq_replied) {
426 DEBUG_REQ(D_NET, req, "REPLIED:");
430 if (req->rq_net_err && !req->rq_timedout) {
431 spin_unlock_irqrestore (&req->rq_lock, flags);
432 rc = ptlrpc_expire_one_request(req);
433 spin_lock_irqsave (&req->rq_lock, flags);
438 DEBUG_REQ(D_ERROR, req, "ABORTED:");
442 if (req->rq_resend) {
443 DEBUG_REQ(D_ERROR, req, "RESEND:");
447 if (req->rq_restart) {
448 DEBUG_REQ(D_ERROR, req, "RESTART:");
453 spin_unlock_irqrestore (&req->rq_lock, flags);
454 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
458 static int ptlrpc_check_status(struct ptlrpc_request *req)
463 err = req->rq_repmsg->status;
464 if (req->rq_repmsg->type == PTL_RPC_MSG_ERR) {
465 DEBUG_REQ(D_ERROR, req, "type == PTL_RPC_MSG_ERR, err == %d",
467 RETURN(err < 0 ? err : -EINVAL);
471 DEBUG_REQ(D_INFO, req, "status is %d", err);
472 } else if (err > 0) {
473 /* XXX: translate this error from net to host */
474 DEBUG_REQ(D_INFO, req, "status is %d", err);
480 static int after_reply(struct ptlrpc_request *req)
483 struct obd_import *imp = req->rq_import;
487 LASSERT(!req->rq_receiving_reply);
489 /* NB Until this point, the whole of the incoming message,
490 * including buflens, status etc is in the sender's byte order. */
493 /* Clear reply swab mask; this is a new reply in sender's byte order */
494 req->rq_rep_swab_mask = 0;
496 LASSERT (req->rq_nob_received <= req->rq_repbuf_len);
497 rc = ptlrpcs_cli_unwrap_reply(req);
499 CERROR("verify reply error: %d\n", rc);
502 /* unwrap_reply may request rpc be resend */
503 if (req->rq_ptlrpcs_restart) {
508 /* unwrap_reply will set rq_replen as the actual received
511 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
513 CERROR("unpack_rep failed: %d\n", rc);
517 if (req->rq_repmsg->type != PTL_RPC_MSG_REPLY &&
518 req->rq_repmsg->type != PTL_RPC_MSG_ERR) {
519 CERROR("invalid packet type received (type=%u)\n",
520 req->rq_repmsg->type);
524 rc = ptlrpc_check_status(req);
526 /* Either we've been evicted, or the server has failed for
527 * some reason. Try to reconnect, and if that fails, punt to the
529 if (rc == -ENOTCONN) {
530 if (req->rq_send_state != LUSTRE_IMP_FULL ||
531 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
535 ptlrpc_request_handle_notconn(req);
540 /* Store transno in reqmsg for replay. */
541 req->rq_reqmsg->transno = req->rq_transno = req->rq_repmsg->transno;
543 if (req->rq_import->imp_replayable) {
544 spin_lock_irqsave(&imp->imp_lock, flags);
545 if (req->rq_transno != 0)
546 ptlrpc_retain_replayable_request(req, imp);
547 else if (req->rq_commit_cb != NULL) {
548 spin_unlock_irqrestore(&imp->imp_lock, flags);
549 req->rq_commit_cb(req);
550 spin_lock_irqsave(&imp->imp_lock, flags);
553 if (req->rq_transno > imp->imp_max_transno)
554 imp->imp_max_transno = req->rq_transno;
556 /* Replay-enabled imports return commit-status information. */
557 if (req->rq_repmsg->last_committed)
558 imp->imp_peer_committed_transno =
559 req->rq_repmsg->last_committed;
560 ptlrpc_free_committed(imp);
561 spin_unlock_irqrestore(&imp->imp_lock, flags);
567 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
569 char str[PTL_NALFMT_SIZE];
570 struct obd_import *imp;
575 LASSERT(req->rq_phase == RQ_PHASE_NEW);
576 req->rq_phase = RQ_PHASE_RPC;
578 imp = req->rq_import;
579 spin_lock_irqsave(&imp->imp_lock, flags);
581 req->rq_import_generation = imp->imp_generation;
583 if (ptlrpc_import_delay_req(imp, req, &rc)) {
584 spin_lock (&req->rq_lock);
586 spin_unlock (&req->rq_lock);
588 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
590 req->rq_reqmsg->status,
591 ptlrpc_import_state_name(req->rq_send_state),
592 ptlrpc_import_state_name(imp->imp_state));
593 LASSERT(list_empty (&req->rq_list));
595 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
596 spin_unlock_irqrestore(&imp->imp_lock, flags);
601 spin_unlock_irqrestore(&imp->imp_lock, flags);
603 req->rq_phase = RQ_PHASE_INTERPRET;
607 /* XXX this is the same as ptlrpc_queue_wait */
608 LASSERT(list_empty(&req->rq_list));
609 list_add_tail(&req->rq_list, &imp->imp_sending_list);
610 spin_unlock_irqrestore(&imp->imp_lock, flags);
612 req->rq_reqmsg->status = current->pid;
613 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc"
614 " %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
615 imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
617 imp->imp_connection->c_peer.peer_ni->pni_name,
618 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
619 req->rq_reqmsg->opc);
621 rc = ptl_send_rpc(req);
623 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
630 int ptlrpc_check_set(struct ptlrpc_request_set *set)
632 char str[PTL_NALFMT_SIZE];
634 struct list_head *tmp;
635 int force_timer_recalc = 0;
638 if (set->set_remaining == 0)
641 list_for_each(tmp, &set->set_requests) {
642 struct ptlrpc_request *req =
643 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
644 struct obd_import *imp = req->rq_import;
647 if (req->rq_phase == RQ_PHASE_NEW &&
648 ptlrpc_send_new_req(req)) {
649 force_timer_recalc = 1;
652 if (!(req->rq_phase == RQ_PHASE_RPC ||
653 req->rq_phase == RQ_PHASE_BULK ||
654 req->rq_phase == RQ_PHASE_INTERPRET ||
655 req->rq_phase == RQ_PHASE_COMPLETE)) {
656 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
660 if (req->rq_phase == RQ_PHASE_COMPLETE)
663 if (req->rq_phase == RQ_PHASE_INTERPRET)
664 GOTO(interpret, req->rq_status);
666 if (req->rq_net_err && !req->rq_timedout)
667 ptlrpc_expire_one_request(req);
669 if (req->rq_err || req->rq_ptlrpcs_err) {
670 ptlrpc_unregister_reply(req);
671 if (req->rq_status == 0)
672 req->rq_status = req->rq_err ? -EIO : -EPERM;
673 req->rq_phase = RQ_PHASE_INTERPRET;
675 spin_lock_irqsave(&imp->imp_lock, flags);
676 list_del_init(&req->rq_list);
677 spin_unlock_irqrestore(&imp->imp_lock, flags);
679 GOTO(interpret, req->rq_status);
682 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
683 * will only be set after rq_timedout, but the oig waiting
684 * path sets rq_intr irrespective of whether ptlrpcd has
685 * seen a timeout. our policy is to only interpret
686 * interrupted rpcs after they have timed out */
687 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
688 /* NB could be on delayed list */
689 ptlrpc_unregister_reply(req);
690 req->rq_status = -EINTR;
691 req->rq_phase = RQ_PHASE_INTERPRET;
693 spin_lock_irqsave(&imp->imp_lock, flags);
694 list_del_init(&req->rq_list);
695 spin_unlock_irqrestore(&imp->imp_lock, flags);
697 GOTO(interpret, req->rq_status);
700 if (req->rq_phase == RQ_PHASE_RPC) {
701 if (req->rq_timedout||req->rq_waiting||req->rq_resend) {
704 ptlrpc_unregister_reply(req);
706 spin_lock_irqsave(&imp->imp_lock, flags);
708 if (ptlrpc_import_delay_req(imp, req, &status)){
709 spin_unlock_irqrestore(&imp->imp_lock,
714 list_del_init(&req->rq_list);
716 req->rq_status = status;
717 req->rq_phase = RQ_PHASE_INTERPRET;
718 spin_unlock_irqrestore(&imp->imp_lock,
720 GOTO(interpret, req->rq_status);
722 if (req->rq_no_resend) {
723 req->rq_status = -ENOTCONN;
724 req->rq_phase = RQ_PHASE_INTERPRET;
725 spin_unlock_irqrestore(&imp->imp_lock,
727 GOTO(interpret, req->rq_status);
729 list_add_tail(&req->rq_list,
730 &imp->imp_sending_list);
732 spin_unlock_irqrestore(&imp->imp_lock, flags);
735 if (req->rq_resend) {
736 if (!req->rq_ptlrpcs_restart)
737 lustre_msg_add_flags(
741 __u64 old_xid = req->rq_xid;
743 ptlrpc_unregister_bulk (req);
745 /* ensure previous bulk fails */
746 req->rq_xid = ptlrpc_next_xid();
747 CDEBUG(D_HA, "resend bulk "
750 old_xid, req->rq_xid);
754 rc = ptl_send_rpc(req);
756 DEBUG_REQ(D_HA, req, "send failed (%d)",
758 force_timer_recalc = 1;
761 /* need to reset the timeout */
762 force_timer_recalc = 1;
765 /* Still waiting for a reply? */
766 if (ptlrpc_client_receiving_reply(req))
769 /* Did we actually receive a reply? */
770 if (!ptlrpc_client_replied(req))
773 spin_lock_irqsave(&imp->imp_lock, flags);
774 list_del_init(&req->rq_list);
775 spin_unlock_irqrestore(&imp->imp_lock, flags);
777 req->rq_status = after_reply(req);
778 if (req->rq_resend) {
779 /* Add this req to the delayed list so
780 it can be errored if the import is
781 evicted after recovery. */
782 spin_lock_irqsave (&req->rq_lock, flags);
783 list_add_tail(&req->rq_list,
784 &imp->imp_delayed_list);
785 spin_unlock_irqrestore(&req->rq_lock, flags);
789 /* If there is no bulk associated with this request,
790 * then we're done and should let the interpreter
791 * process the reply. Similarly if the RPC returned
792 * an error, and therefore the bulk will never arrive.
794 if (req->rq_bulk == NULL || req->rq_status != 0) {
795 req->rq_phase = RQ_PHASE_INTERPRET;
796 GOTO(interpret, req->rq_status);
799 req->rq_phase = RQ_PHASE_BULK;
802 LASSERT(req->rq_phase == RQ_PHASE_BULK);
803 if (ptlrpc_bulk_active(req->rq_bulk))
806 if (!req->rq_bulk->bd_success) {
807 /* The RPC reply arrived OK, but the bulk screwed
808 * up! Dead wierd since the server told us the RPC
809 * was good after getting the REPLY for her GET or
810 * the ACK for her PUT. */
811 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
815 req->rq_phase = RQ_PHASE_INTERPRET;
818 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
819 LASSERT(!req->rq_receiving_reply);
821 ptlrpc_unregister_reply(req);
822 if (req->rq_bulk != NULL)
823 ptlrpc_unregister_bulk (req);
825 req->rq_phase = RQ_PHASE_COMPLETE;
827 if (req->rq_interpret_reply != NULL) {
828 int (*interpreter)(struct ptlrpc_request *,void *,int) =
829 req->rq_interpret_reply;
830 req->rq_status = interpreter(req, &req->rq_async_args,
834 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
835 "opc %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
836 imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
838 imp->imp_connection->c_peer.peer_ni->pni_name,
839 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
840 req->rq_reqmsg->opc);
842 set->set_remaining--;
844 atomic_dec(&imp->imp_inflight);
845 wake_up(&imp->imp_recovery_waitq);
848 /* If we hit an error, we want to recover promptly. */
849 RETURN(set->set_remaining == 0 || force_timer_recalc);
852 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
855 struct obd_import *imp = req->rq_import;
859 DEBUG_REQ(D_ERROR, req, "timeout (sent at %lu, %lus ago)",
860 (long)req->rq_sent, LTIME_S(CURRENT_TIME) - req->rq_sent);
862 spin_lock_irqsave (&req->rq_lock, flags);
863 replied = req->rq_replied;
865 req->rq_timedout = 1;
866 spin_unlock_irqrestore (&req->rq_lock, flags);
871 ptlrpc_unregister_reply (req);
873 if (obd_dump_on_timeout)
874 portals_debug_dumplog();
876 if (req->rq_bulk != NULL)
877 ptlrpc_unregister_bulk (req);
880 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
884 /* The DLM server doesn't want recovery run on its imports. */
885 if (imp->imp_dlm_fake)
888 /* If this request is for recovery or other primordial tasks,
889 * then error it out here. */
890 if (req->rq_send_state != LUSTRE_IMP_FULL ||
891 imp->imp_obd->obd_no_recov) {
892 spin_lock_irqsave (&req->rq_lock, flags);
893 req->rq_status = -ETIMEDOUT;
895 spin_unlock_irqrestore (&req->rq_lock, flags);
899 ptlrpc_fail_import(imp, req->rq_import_generation);
904 int ptlrpc_expired_set(void *data)
906 struct ptlrpc_request_set *set = data;
907 struct list_head *tmp;
908 time_t now = LTIME_S(CURRENT_TIME);
911 LASSERT(set != NULL);
913 /* A timeout expired; see which reqs it applies to... */
914 list_for_each (tmp, &set->set_requests) {
915 struct ptlrpc_request *req =
916 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
918 /* request in-flight? */
919 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting
920 && !req->rq_resend) ||
921 (req->rq_phase == RQ_PHASE_BULK)))
924 if (req->rq_timedout || /* already dealt with */
925 req->rq_sent + req->rq_timeout > now) /* not expired */
928 /* deal with this guy */
929 ptlrpc_expire_one_request (req);
932 /* When waiting for a whole set, we always to break out of the
933 * sleep so we can recalculate the timeout, or enable interrupts
934 * iff everyone's timed out.
939 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
942 spin_lock_irqsave(&req->rq_lock, flags);
944 spin_unlock_irqrestore(&req->rq_lock, flags);
947 void ptlrpc_interrupted_set(void *data)
949 struct ptlrpc_request_set *set = data;
950 struct list_head *tmp;
952 LASSERT(set != NULL);
953 CERROR("INTERRUPTED SET %p\n", set);
955 list_for_each(tmp, &set->set_requests) {
956 struct ptlrpc_request *req =
957 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
959 if (req->rq_phase != RQ_PHASE_RPC)
962 ptlrpc_mark_interrupted(req);
966 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
968 struct list_head *tmp;
969 time_t now = LTIME_S(CURRENT_TIME);
972 struct ptlrpc_request *req;
975 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
977 list_for_each(tmp, &set->set_requests) {
978 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
980 /* request in-flight? */
981 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
982 (req->rq_phase == RQ_PHASE_BULK)))
985 if (req->rq_timedout) /* already timed out */
988 deadline = req->rq_sent + req->rq_timeout;
989 if (deadline <= now) /* actually expired already */
990 timeout = 1; /* ASAP */
991 else if (timeout == 0 || timeout > deadline - now)
992 timeout = deadline - now;
998 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1000 struct list_head *tmp;
1001 struct ptlrpc_request *req;
1002 struct l_wait_info lwi;
1006 LASSERT(!list_empty(&set->set_requests));
1007 list_for_each(tmp, &set->set_requests) {
1008 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1009 if (req->rq_phase == RQ_PHASE_NEW)
1010 (void)ptlrpc_send_new_req(req);
1014 timeout = ptlrpc_set_next_timeout(set);
1016 /* wait until all complete, interrupted, or an in-flight
1018 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
1020 lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
1022 ptlrpc_interrupted_set, set);
1023 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1025 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1027 /* -EINTR => all requests have been flagged rq_intr so next
1029 * -ETIMEOUTD => someone timed out. When all reqs have
1030 * timed out, signals are enabled allowing completion with
1032 * I don't really care if we go once more round the loop in
1033 * the error cases -eeb. */
1034 } while (rc != 0 || set->set_remaining != 0);
1036 LASSERT(set->set_remaining == 0);
1039 list_for_each(tmp, &set->set_requests) {
1040 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1042 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1043 if (req->rq_status != 0)
1044 rc = req->rq_status;
1047 if (set->set_interpret != NULL) {
1048 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1050 rc = interpreter (set, set->set_arg, rc);
1056 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1059 if (request == NULL) {
1064 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1065 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1066 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1067 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1068 LASSERT(request->rq_cred);
1070 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1071 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1072 if (request->rq_import != NULL) {
1073 unsigned long flags = 0;
1075 spin_lock_irqsave(&request->rq_import->imp_lock, flags);
1076 list_del_init(&request->rq_replay_list);
1078 spin_unlock_irqrestore(&request->rq_import->imp_lock,
1081 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1083 if (atomic_read(&request->rq_refcount) != 0) {
1084 DEBUG_REQ(D_ERROR, request,
1085 "freeing request with nonzero refcount");
1089 if (request->rq_repbuf != NULL)
1090 ptlrpcs_cli_free_repbuf(request);
1091 if (request->rq_reqbuf != NULL)
1092 ptlrpcs_cli_free_reqbuf(request);
1094 if (request->rq_export != NULL) {
1095 class_export_put(request->rq_export);
1096 request->rq_export = NULL;
1098 if (request->rq_import != NULL) {
1099 class_import_put(request->rq_import);
1100 request->rq_import = NULL;
1102 if (request->rq_bulk != NULL)
1103 ptlrpc_free_bulk(request->rq_bulk);
1105 ptlrpcs_req_drop_cred(request);
1106 OBD_FREE(request, sizeof(*request));
1110 void ptlrpc_free_req(struct ptlrpc_request *request)
1112 __ptlrpc_free_req(request, 0);
1115 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1116 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1118 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1119 (void)__ptlrpc_req_finished(request, 1);
1122 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1125 if (request == NULL)
1128 if (request == LP_POISON ||
1129 request->rq_reqmsg == LP_POISON) {
1130 CERROR("dereferencing freed request (bug 575)\n");
1135 DEBUG_REQ(D_INFO, request, "refcount now %u",
1136 atomic_read(&request->rq_refcount) - 1);
1138 if (atomic_dec_and_test(&request->rq_refcount)) {
1139 __ptlrpc_free_req(request, locked);
1146 void ptlrpc_req_finished(struct ptlrpc_request *request)
1148 __ptlrpc_req_finished(request, 0);
1151 /* Disengage the client's reply buffer from the network
1152 * NB does _NOT_ unregister any client-side bulk.
1153 * IDEMPOTENT, but _not_ safe against concurrent callers.
1154 * The request owner (i.e. the thread doing the I/O) must call...
1156 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1159 wait_queue_head_t *wq;
1160 struct l_wait_info lwi;
1162 LASSERT(!in_interrupt ()); /* might sleep */
1164 if (!ptlrpc_client_receiving_reply(request))
1167 PtlMDUnlink (request->rq_reply_md_h);
1169 /* We have to l_wait_event() whatever the result, to give liblustre
1170 * a chance to run reply_in_callback() */
1172 if (request->rq_set != NULL)
1173 wq = &request->rq_set->set_waitq;
1175 wq = &request->rq_reply_waitq;
1178 /* Network access will complete in finite time but the HUGE
1179 * timeout lets us CWARN for visibility of sluggish NALs */
1180 lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
1181 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1185 LASSERT (rc == -ETIMEDOUT);
1186 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1190 /* caller must hold imp->imp_lock */
1191 void ptlrpc_free_committed(struct obd_import *imp)
1193 struct list_head *tmp, *saved;
1194 struct ptlrpc_request *req;
1195 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1198 LASSERT(imp != NULL);
1200 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1202 CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
1203 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1205 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1206 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1208 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1209 LASSERT(req != last_req);
1212 if (req->rq_import_generation < imp->imp_generation) {
1213 DEBUG_REQ(D_HA, req, "freeing request with old gen");
1217 if (req->rq_replay) {
1218 #warning "REMOVE THIS CRAP ONCE 6403 SOLVED -bzzz"
1219 struct mdc_open_data {
1220 struct obd_client_handle *mod_och;
1221 struct ptlrpc_request *mod_open_req;
1222 struct ptlrpc_request *mod_close_req;
1224 struct mdc_open_data *mod = req->rq_cb_data;
1225 if (mod == NULL || mod->mod_close_req == NULL) {
1226 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1229 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY), "
1230 "closed by x"LPD64"/t"LPD64,
1231 mod->mod_close_req->rq_xid,
1232 mod->mod_close_req->rq_repmsg ?
1233 mod->mod_close_req->rq_repmsg->transno : 0);
1237 /* not yet committed */
1238 if (req->rq_transno > imp->imp_peer_committed_transno) {
1239 DEBUG_REQ(D_HA, req, "stopping search");
1243 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1244 imp->imp_peer_committed_transno);
1246 if (req->rq_commit_cb != NULL)
1247 req->rq_commit_cb(req);
1248 list_del_init(&req->rq_replay_list);
1249 __ptlrpc_req_finished(req, 1);
1256 void ptlrpc_cleanup_client(struct obd_import *imp)
1263 void ptlrpc_resend_req(struct ptlrpc_request *req)
1265 unsigned long flags;
1267 DEBUG_REQ(D_HA, req, "going to resend");
1268 req->rq_reqmsg->handle.cookie = 0;
1269 req->rq_status = -EAGAIN;
1271 spin_lock_irqsave (&req->rq_lock, flags);
1273 req->rq_net_err = 0;
1274 req->rq_timedout = 0;
1276 __u64 old_xid = req->rq_xid;
1278 /* ensure previous bulk fails */
1279 req->rq_xid = ptlrpc_next_xid();
1280 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1281 old_xid, req->rq_xid);
1283 ptlrpc_wake_client_req(req);
1284 spin_unlock_irqrestore (&req->rq_lock, flags);
1287 /* XXX: this function and rq_status are currently unused */
1288 void ptlrpc_restart_req(struct ptlrpc_request *req)
1290 unsigned long flags;
1292 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1293 req->rq_status = -ERESTARTSYS;
1295 spin_lock_irqsave (&req->rq_lock, flags);
1296 req->rq_restart = 1;
1297 req->rq_timedout = 0;
1298 ptlrpc_wake_client_req(req);
1299 spin_unlock_irqrestore (&req->rq_lock, flags);
1302 static int expired_request(void *data)
1304 struct ptlrpc_request *req = data;
1307 /* some failure can suspend regular timeouts */
1308 if (ptlrpc_check_suspend())
1311 RETURN(ptlrpc_expire_one_request(req));
1314 static void interrupted_request(void *data)
1316 unsigned long flags;
1318 struct ptlrpc_request *req = data;
1319 DEBUG_REQ(D_HA, req, "request interrupted");
1320 spin_lock_irqsave (&req->rq_lock, flags);
1322 spin_unlock_irqrestore (&req->rq_lock, flags);
1325 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1328 atomic_inc(&req->rq_refcount);
1332 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1333 struct obd_import *imp)
1335 struct list_head *tmp;
1337 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1339 /* clear this for new requests that were resent as well
1340 as resent replayed requests. */
1341 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1343 /* don't re-add requests that have been replayed */
1344 if (!list_empty(&req->rq_replay_list))
1347 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1349 LASSERT(imp->imp_replayable);
1350 /* Balanced in ptlrpc_free_committed, usually. */
1351 ptlrpc_request_addref(req);
1352 list_for_each_prev(tmp, &imp->imp_replay_list) {
1353 struct ptlrpc_request *iter =
1354 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1356 /* We may have duplicate transnos if we create and then
1357 * open a file, or for closes retained if to match creating
1358 * opens, so use req->rq_xid as a secondary key.
1359 * (See bugs 684, 685, and 428.)
1360 * XXX no longer needed, but all opens need transnos!
1362 if (iter->rq_transno > req->rq_transno)
1365 if (iter->rq_transno == req->rq_transno) {
1366 LASSERT(iter->rq_xid != req->rq_xid);
1367 if (iter->rq_xid > req->rq_xid)
1371 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1375 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1378 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1380 char str[PTL_NALFMT_SIZE];
1383 struct l_wait_info lwi;
1384 struct obd_import *imp = req->rq_import;
1385 unsigned long flags;
1389 LASSERT(req->rq_set == NULL);
1390 LASSERT(!req->rq_receiving_reply);
1391 atomic_inc(&imp->imp_inflight);
1393 if (imp->imp_connection == NULL) {
1394 CERROR("request on not connected import %s\n",
1395 imp->imp_obd->obd_name);
1399 /* for distributed debugging */
1400 req->rq_reqmsg->status = current->pid;
1401 LASSERT(imp->imp_obd != NULL);
1402 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
1403 "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1404 imp->imp_obd->obd_uuid.uuid,
1405 req->rq_reqmsg->status, req->rq_xid,
1406 imp->imp_connection->c_peer.peer_ni->pni_name,
1407 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1408 req->rq_reqmsg->opc);
1410 /* Mark phase here for a little debug help */
1411 req->rq_phase = RQ_PHASE_RPC;
1413 spin_lock_irqsave(&imp->imp_lock, flags);
1414 req->rq_import_generation = imp->imp_generation;
1416 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1417 list_del(&req->rq_list);
1419 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1420 spin_unlock_irqrestore(&imp->imp_lock, flags);
1422 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1424 ptlrpc_import_state_name(req->rq_send_state),
1425 ptlrpc_import_state_name(imp->imp_state));
1426 lwi = LWI_INTR(interrupted_request, req);
1427 rc = l_wait_event(req->rq_reply_waitq,
1428 (req->rq_send_state == imp->imp_state ||
1431 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
1433 ptlrpc_import_state_name(imp->imp_state),
1434 ptlrpc_import_state_name(req->rq_send_state),
1437 spin_lock_irqsave(&imp->imp_lock, flags);
1438 list_del_init(&req->rq_list);
1443 else if (req->rq_intr) {
1446 else if (req->rq_no_resend) {
1447 spin_unlock_irqrestore(&imp->imp_lock, flags);
1448 GOTO(out, rc = -ETIMEDOUT);
1456 list_del_init(&req->rq_list);
1457 spin_unlock_irqrestore(&imp->imp_lock, flags);
1458 req->rq_status = rc; // XXX this ok?
1462 if (req->rq_resend) {
1463 if (!req->rq_ptlrpcs_restart)
1464 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1466 if (req->rq_bulk != NULL)
1467 ptlrpc_unregister_bulk (req);
1469 DEBUG_REQ(D_HA, req, "resending: ");
1472 /* XXX this is the same as ptlrpc_set_wait */
1473 LASSERT(list_empty(&req->rq_list));
1474 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1475 spin_unlock_irqrestore(&imp->imp_lock, flags);
1477 rc = ptl_send_rpc(req);
1479 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1482 timeout = MAX(req->rq_timeout * HZ, 1);
1483 DEBUG_REQ(D_NET, req, "-- sleeping for %d jiffies", timeout);
1486 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1488 rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1489 if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1491 DEBUG_REQ(D_NET, req, "-- done sleeping");
1493 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
1494 "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1495 imp->imp_obd->obd_uuid.uuid,
1496 req->rq_reqmsg->status, req->rq_xid,
1497 imp->imp_connection->c_peer.peer_ni->pni_name,
1498 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1499 req->rq_reqmsg->opc);
1501 spin_lock_irqsave(&imp->imp_lock, flags);
1502 list_del_init(&req->rq_list);
1503 spin_unlock_irqrestore(&imp->imp_lock, flags);
1505 /* If the reply was received normally, this just grabs the spinlock
1506 * (ensuring the reply callback has returned), sees that
1507 * req->rq_receiving_reply is clear and returns. */
1508 ptlrpc_unregister_reply (req);
1511 GOTO(out, rc = -EIO);
1513 if (req->rq_ptlrpcs_err)
1514 GOTO(out, rc = -EPERM);
1516 /* Resend if we need to, unless we were interrupted. */
1517 if (req->rq_resend && !req->rq_intr) {
1518 /* ...unless we were specifically told otherwise. */
1519 if (req->rq_no_resend)
1520 GOTO(out, rc = -ETIMEDOUT);
1521 spin_lock_irqsave(&imp->imp_lock, flags);
1526 /* Should only be interrupted if we timed out. */
1527 if (!req->rq_timedout)
1528 DEBUG_REQ(D_ERROR, req,
1529 "rq_intr set but rq_timedout not");
1530 GOTO(out, rc = -EINTR);
1533 if (req->rq_timedout) { /* non-recoverable timeout */
1534 GOTO(out, rc = -ETIMEDOUT);
1537 if (!req->rq_replied) {
1538 /* How can this be? -eeb */
1539 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1541 GOTO(out, rc = req->rq_status);
1544 rc = after_reply (req);
1545 /* NB may return +ve success rc */
1546 if (req->rq_resend) {
1547 spin_lock_irqsave(&imp->imp_lock, flags);
1552 if (req->rq_bulk != NULL) {
1554 /* success so far. Note that anything going wrong
1555 * with bulk now, is EXTREMELY strange, since the
1556 * server must have believed that the bulk
1557 * tranferred OK before she replied with success to
1559 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1560 brc = l_wait_event(req->rq_reply_waitq,
1561 !ptlrpc_bulk_active(req->rq_bulk),
1563 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1565 LASSERT(brc == -ETIMEDOUT);
1566 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1568 } else if (!req->rq_bulk->bd_success) {
1569 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1574 ptlrpc_unregister_bulk (req);
1577 LASSERT(!req->rq_receiving_reply);
1578 req->rq_phase = RQ_PHASE_INTERPRET;
1580 atomic_dec(&imp->imp_inflight);
1581 wake_up(&imp->imp_recovery_waitq);
1585 struct ptlrpc_replay_async_args {
1587 int praa_old_status;
1590 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1591 void * data, int rc)
1593 struct ptlrpc_replay_async_args *aa = data;
1594 struct obd_import *imp = req->rq_import;
1595 unsigned long flags;
1597 atomic_dec(&imp->imp_replay_inflight);
1599 if (!req->rq_replied) {
1600 CERROR("request replay timed out, restarting recovery\n");
1601 GOTO(out, rc = -ETIMEDOUT);
1605 /* Clear reply swab mask; this is a new reply in sender's byte order */
1606 req->rq_rep_swab_mask = 0;
1608 LASSERT (req->rq_nob_received <= req->rq_repbuf_len);
1609 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
1611 CERROR("unpack_rep failed: %d\n", rc);
1612 GOTO(out, rc = -EPROTO);
1615 if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
1616 req->rq_repmsg->status == -ENOTCONN)
1617 GOTO(out, rc = req->rq_repmsg->status);
1619 /* The transno had better not change over replay. */
1620 LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
1622 DEBUG_REQ(D_HA, req, "got rep");
1624 /* let the callback do fixups, possibly including in the request */
1625 if (req->rq_replay_cb)
1626 req->rq_replay_cb(req);
1628 if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
1629 DEBUG_REQ(D_HA, req, "status %d, old was %d",
1630 req->rq_repmsg->status, aa->praa_old_status);
1632 /* Put it back for re-replay. */
1633 req->rq_repmsg->status = aa->praa_old_status;
1636 spin_lock_irqsave(&imp->imp_lock, flags);
1637 imp->imp_last_replay_transno = req->rq_transno;
1638 spin_unlock_irqrestore(&imp->imp_lock, flags);
1640 /* continue with recovery */
1641 rc = ptlrpc_import_recovery_state_machine(imp);
1643 req->rq_send_state = aa->praa_old_state;
1646 /* this replay failed, so restart recovery */
1647 ptlrpc_connect_import(imp, NULL);
1653 int ptlrpc_replay_req(struct ptlrpc_request *req)
1655 struct ptlrpc_replay_async_args *aa;
1658 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1660 /* Not handling automatic bulk replay yet (or ever?) */
1661 LASSERT(req->rq_bulk == NULL);
1663 DEBUG_REQ(D_HA, req, "REPLAY");
1665 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1666 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1667 memset(aa, 0, sizeof *aa);
1669 /* Prepare request to be resent with ptlrpcd */
1670 aa->praa_old_state = req->rq_send_state;
1671 req->rq_send_state = LUSTRE_IMP_REPLAY;
1672 req->rq_phase = RQ_PHASE_NEW;
1673 aa->praa_old_status = req->rq_repmsg->status;
1676 req->rq_interpret_reply = ptlrpc_replay_interpret;
1677 atomic_inc(&req->rq_import->imp_replay_inflight);
1678 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1680 ptlrpcd_add_req(req);
1684 void ptlrpc_abort_inflight(struct obd_import *imp)
1686 unsigned long flags;
1687 struct list_head *tmp, *n;
1690 /* Make sure that no new requests get processed for this import.
1691 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1692 * this flag and then putting requests on sending_list or delayed_list.
1694 spin_lock_irqsave(&imp->imp_lock, flags);
1696 /* XXX locking? Maybe we should remove each request with the list
1697 * locked? Also, how do we know if the requests on the list are
1698 * being freed at this time?
1700 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1701 struct ptlrpc_request *req =
1702 list_entry(tmp, struct ptlrpc_request, rq_list);
1704 DEBUG_REQ(D_HA, req, "inflight");
1706 spin_lock (&req->rq_lock);
1707 if (req->rq_import_generation < imp->imp_generation) {
1709 ptlrpc_wake_client_req(req);
1711 spin_unlock (&req->rq_lock);
1714 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1715 struct ptlrpc_request *req =
1716 list_entry(tmp, struct ptlrpc_request, rq_list);
1718 DEBUG_REQ(D_HA, req, "aborting waiting req");
1720 spin_lock (&req->rq_lock);
1721 if (req->rq_import_generation < imp->imp_generation) {
1723 ptlrpc_wake_client_req(req);
1725 spin_unlock (&req->rq_lock);
1728 list_for_each_safe(tmp, n, &imp->imp_rawrpc_list) {
1729 struct ptlrpc_request *req =
1730 list_entry(tmp, struct ptlrpc_request, rq_list);
1732 DEBUG_REQ(D_HA, req, "aborting raw rpc");
1734 spin_lock(&req->rq_lock);
1736 ptlrpc_wake_client_req(req);
1737 spin_unlock(&req->rq_lock);
1740 /* Last chance to free reqs left on the replay list, but we
1741 * will still leak reqs that haven't comitted. */
1742 if (imp->imp_replayable)
1743 ptlrpc_free_committed(imp);
1745 spin_unlock_irqrestore(&imp->imp_lock, flags);
1750 static __u64 ptlrpc_last_xid = 0;
1751 static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
1753 __u64 ptlrpc_next_xid(void)
1756 spin_lock(&ptlrpc_last_xid_lock);
1757 tmp = ++ptlrpc_last_xid;
1758 spin_unlock(&ptlrpc_last_xid_lock);