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;
544 if (req->rq_import->imp_replayable) {
545 spin_lock_irqsave(&imp->imp_lock, flags);
546 if (req->rq_transno != 0)
547 ptlrpc_retain_replayable_request(req, imp);
548 else if (req->rq_commit_cb != NULL) {
549 spin_unlock_irqrestore(&imp->imp_lock, flags);
550 req->rq_commit_cb(req);
551 spin_lock_irqsave(&imp->imp_lock, flags);
554 if (req->rq_transno > imp->imp_max_transno)
555 imp->imp_max_transno = req->rq_transno;
557 /* Replay-enabled imports return commit-status information. */
558 if (req->rq_repmsg->last_committed)
559 imp->imp_peer_committed_transno =
560 req->rq_repmsg->last_committed;
561 ptlrpc_free_committed(imp);
562 spin_unlock_irqrestore(&imp->imp_lock, flags);
568 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
570 char str[PTL_NALFMT_SIZE];
571 struct obd_import *imp;
576 LASSERT(req->rq_phase == RQ_PHASE_NEW);
577 req->rq_phase = RQ_PHASE_RPC;
579 imp = req->rq_import;
580 spin_lock_irqsave(&imp->imp_lock, flags);
582 req->rq_import_generation = imp->imp_generation;
584 if (ptlrpc_import_delay_req(imp, req, &rc)) {
585 spin_lock (&req->rq_lock);
587 spin_unlock (&req->rq_lock);
589 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
591 req->rq_reqmsg->status,
592 ptlrpc_import_state_name(req->rq_send_state),
593 ptlrpc_import_state_name(imp->imp_state));
594 LASSERT(list_empty (&req->rq_list));
596 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
597 spin_unlock_irqrestore(&imp->imp_lock, flags);
602 spin_unlock_irqrestore(&imp->imp_lock, flags);
604 req->rq_phase = RQ_PHASE_INTERPRET;
608 /* XXX this is the same as ptlrpc_queue_wait */
609 LASSERT(list_empty(&req->rq_list));
610 list_add_tail(&req->rq_list, &imp->imp_sending_list);
611 spin_unlock_irqrestore(&imp->imp_lock, flags);
613 req->rq_reqmsg->status = current->pid;
614 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc"
615 " %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
616 imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
618 imp->imp_connection->c_peer.peer_ni->pni_name,
619 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
620 req->rq_reqmsg->opc);
622 rc = ptl_send_rpc(req);
624 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
631 int ptlrpc_check_set(struct ptlrpc_request_set *set)
633 char str[PTL_NALFMT_SIZE];
635 struct list_head *tmp;
636 int force_timer_recalc = 0;
639 if (set->set_remaining == 0)
642 list_for_each(tmp, &set->set_requests) {
643 struct ptlrpc_request *req =
644 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
645 struct obd_import *imp = req->rq_import;
648 if (req->rq_phase == RQ_PHASE_NEW &&
649 ptlrpc_send_new_req(req)) {
650 force_timer_recalc = 1;
653 if (!(req->rq_phase == RQ_PHASE_RPC ||
654 req->rq_phase == RQ_PHASE_BULK ||
655 req->rq_phase == RQ_PHASE_INTERPRET ||
656 req->rq_phase == RQ_PHASE_COMPLETE)) {
657 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
661 if (req->rq_phase == RQ_PHASE_COMPLETE)
664 if (req->rq_phase == RQ_PHASE_INTERPRET)
665 GOTO(interpret, req->rq_status);
667 if (req->rq_net_err && !req->rq_timedout)
668 ptlrpc_expire_one_request(req);
670 if (req->rq_err || req->rq_ptlrpcs_err) {
671 ptlrpc_unregister_reply(req);
672 if (req->rq_status == 0)
673 req->rq_status = req->rq_err ? -EIO : -EPERM;
674 req->rq_phase = RQ_PHASE_INTERPRET;
676 spin_lock_irqsave(&imp->imp_lock, flags);
677 list_del_init(&req->rq_list);
678 spin_unlock_irqrestore(&imp->imp_lock, flags);
680 GOTO(interpret, req->rq_status);
683 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
684 * will only be set after rq_timedout, but the oig waiting
685 * path sets rq_intr irrespective of whether ptlrpcd has
686 * seen a timeout. our policy is to only interpret
687 * interrupted rpcs after they have timed out */
688 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
689 /* NB could be on delayed list */
690 ptlrpc_unregister_reply(req);
691 req->rq_status = -EINTR;
692 req->rq_phase = RQ_PHASE_INTERPRET;
694 spin_lock_irqsave(&imp->imp_lock, flags);
695 list_del_init(&req->rq_list);
696 spin_unlock_irqrestore(&imp->imp_lock, flags);
698 GOTO(interpret, req->rq_status);
701 if (req->rq_phase == RQ_PHASE_RPC) {
702 if (req->rq_timedout||req->rq_waiting||req->rq_resend) {
705 ptlrpc_unregister_reply(req);
707 spin_lock_irqsave(&imp->imp_lock, flags);
709 if (ptlrpc_import_delay_req(imp, req, &status)){
710 spin_unlock_irqrestore(&imp->imp_lock,
715 list_del_init(&req->rq_list);
717 req->rq_status = status;
718 req->rq_phase = RQ_PHASE_INTERPRET;
719 spin_unlock_irqrestore(&imp->imp_lock,
721 GOTO(interpret, req->rq_status);
723 if (req->rq_no_resend) {
724 req->rq_status = -ENOTCONN;
725 req->rq_phase = RQ_PHASE_INTERPRET;
726 spin_unlock_irqrestore(&imp->imp_lock,
728 GOTO(interpret, req->rq_status);
730 list_add_tail(&req->rq_list,
731 &imp->imp_sending_list);
733 spin_unlock_irqrestore(&imp->imp_lock, flags);
736 if (req->rq_resend) {
737 if (!req->rq_ptlrpcs_restart)
738 lustre_msg_add_flags(
742 __u64 old_xid = req->rq_xid;
744 ptlrpc_unregister_bulk (req);
746 /* ensure previous bulk fails */
747 req->rq_xid = ptlrpc_next_xid();
748 CDEBUG(D_HA, "resend bulk "
751 old_xid, req->rq_xid);
755 rc = ptl_send_rpc(req);
757 DEBUG_REQ(D_HA, req, "send failed (%d)",
759 force_timer_recalc = 1;
762 /* need to reset the timeout */
763 force_timer_recalc = 1;
766 /* Still waiting for a reply? */
767 if (ptlrpc_client_receiving_reply(req))
770 /* Did we actually receive a reply? */
771 if (!ptlrpc_client_replied(req))
774 spin_lock_irqsave(&imp->imp_lock, flags);
775 list_del_init(&req->rq_list);
776 spin_unlock_irqrestore(&imp->imp_lock, flags);
778 req->rq_status = after_reply(req);
779 if (req->rq_resend) {
780 /* Add this req to the delayed list so
781 it can be errored if the import is
782 evicted after recovery. */
783 spin_lock_irqsave (&req->rq_lock, flags);
784 list_add_tail(&req->rq_list,
785 &imp->imp_delayed_list);
786 spin_unlock_irqrestore(&req->rq_lock, flags);
790 /* If there is no bulk associated with this request,
791 * then we're done and should let the interpreter
792 * process the reply. Similarly if the RPC returned
793 * an error, and therefore the bulk will never arrive.
795 if (req->rq_bulk == NULL || req->rq_status != 0) {
796 req->rq_phase = RQ_PHASE_INTERPRET;
797 GOTO(interpret, req->rq_status);
800 req->rq_phase = RQ_PHASE_BULK;
803 LASSERT(req->rq_phase == RQ_PHASE_BULK);
804 if (ptlrpc_bulk_active(req->rq_bulk))
807 if (!req->rq_bulk->bd_success) {
808 /* The RPC reply arrived OK, but the bulk screwed
809 * up! Dead wierd since the server told us the RPC
810 * was good after getting the REPLY for her GET or
811 * the ACK for her PUT. */
812 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
816 req->rq_phase = RQ_PHASE_INTERPRET;
819 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
820 LASSERT(!req->rq_receiving_reply);
822 ptlrpc_unregister_reply(req);
823 if (req->rq_bulk != NULL)
824 ptlrpc_unregister_bulk (req);
826 req->rq_phase = RQ_PHASE_COMPLETE;
828 if (req->rq_interpret_reply != NULL) {
829 int (*interpreter)(struct ptlrpc_request *,void *,int) =
830 req->rq_interpret_reply;
831 req->rq_status = interpreter(req, &req->rq_async_args,
835 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:"
836 "opc %s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
837 imp->imp_obd->obd_uuid.uuid, req->rq_reqmsg->status,
839 imp->imp_connection->c_peer.peer_ni->pni_name,
840 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
841 req->rq_reqmsg->opc);
843 set->set_remaining--;
845 atomic_dec(&imp->imp_inflight);
846 wake_up(&imp->imp_recovery_waitq);
849 /* If we hit an error, we want to recover promptly. */
850 RETURN(set->set_remaining == 0 || force_timer_recalc);
853 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
856 struct obd_import *imp = req->rq_import;
860 DEBUG_REQ(D_ERROR, req, "timeout (sent at %lu, %lus ago)",
861 (long)req->rq_sent, LTIME_S(CURRENT_TIME) - req->rq_sent);
863 spin_lock_irqsave (&req->rq_lock, flags);
864 replied = req->rq_replied;
866 req->rq_timedout = 1;
867 spin_unlock_irqrestore (&req->rq_lock, flags);
872 ptlrpc_unregister_reply (req);
874 if (obd_dump_on_timeout)
875 portals_debug_dumplog();
877 if (req->rq_bulk != NULL)
878 ptlrpc_unregister_bulk (req);
881 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
885 /* The DLM server doesn't want recovery run on its imports. */
886 if (imp->imp_dlm_fake)
889 /* If this request is for recovery or other primordial tasks,
890 * then error it out here. */
891 if (req->rq_send_state != LUSTRE_IMP_FULL ||
892 imp->imp_obd->obd_no_recov) {
893 spin_lock_irqsave (&req->rq_lock, flags);
894 req->rq_status = -ETIMEDOUT;
896 spin_unlock_irqrestore (&req->rq_lock, flags);
900 ptlrpc_fail_import(imp, req->rq_import_generation);
905 int ptlrpc_expired_set(void *data)
907 struct ptlrpc_request_set *set = data;
908 struct list_head *tmp;
909 time_t now = LTIME_S(CURRENT_TIME);
912 LASSERT(set != NULL);
914 /* A timeout expired; see which reqs it applies to... */
915 list_for_each (tmp, &set->set_requests) {
916 struct ptlrpc_request *req =
917 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
919 /* request in-flight? */
920 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting
921 && !req->rq_resend) ||
922 (req->rq_phase == RQ_PHASE_BULK)))
925 if (req->rq_timedout || /* already dealt with */
926 req->rq_sent + req->rq_timeout > now) /* not expired */
929 /* deal with this guy */
930 ptlrpc_expire_one_request (req);
933 /* When waiting for a whole set, we always to break out of the
934 * sleep so we can recalculate the timeout, or enable interrupts
935 * iff everyone's timed out.
940 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
943 spin_lock_irqsave(&req->rq_lock, flags);
945 spin_unlock_irqrestore(&req->rq_lock, flags);
948 void ptlrpc_interrupted_set(void *data)
950 struct ptlrpc_request_set *set = data;
951 struct list_head *tmp;
953 LASSERT(set != NULL);
954 CERROR("INTERRUPTED SET %p\n", set);
956 list_for_each(tmp, &set->set_requests) {
957 struct ptlrpc_request *req =
958 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
960 if (req->rq_phase != RQ_PHASE_RPC)
963 ptlrpc_mark_interrupted(req);
967 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
969 struct list_head *tmp;
970 time_t now = LTIME_S(CURRENT_TIME);
973 struct ptlrpc_request *req;
976 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
978 list_for_each(tmp, &set->set_requests) {
979 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
981 /* request in-flight? */
982 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
983 (req->rq_phase == RQ_PHASE_BULK)))
986 if (req->rq_timedout) /* already timed out */
989 deadline = req->rq_sent + req->rq_timeout;
990 if (deadline <= now) /* actually expired already */
991 timeout = 1; /* ASAP */
992 else if (timeout == 0 || timeout > deadline - now)
993 timeout = deadline - now;
999 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1001 struct list_head *tmp;
1002 struct ptlrpc_request *req;
1003 struct l_wait_info lwi;
1007 LASSERT(!list_empty(&set->set_requests));
1008 list_for_each(tmp, &set->set_requests) {
1009 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1010 if (req->rq_phase == RQ_PHASE_NEW)
1011 (void)ptlrpc_send_new_req(req);
1015 timeout = ptlrpc_set_next_timeout(set);
1017 /* wait until all complete, interrupted, or an in-flight
1019 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
1021 lwi = LWI_TIMEOUT_INTR((timeout ? timeout : 1) * HZ,
1023 ptlrpc_interrupted_set, set);
1024 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1026 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1028 /* -EINTR => all requests have been flagged rq_intr so next
1030 * -ETIMEOUTD => someone timed out. When all reqs have
1031 * timed out, signals are enabled allowing completion with
1033 * I don't really care if we go once more round the loop in
1034 * the error cases -eeb. */
1035 } while (rc != 0 || set->set_remaining != 0);
1037 LASSERT(set->set_remaining == 0);
1040 list_for_each(tmp, &set->set_requests) {
1041 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1043 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1044 if (req->rq_status != 0)
1045 rc = req->rq_status;
1048 if (set->set_interpret != NULL) {
1049 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1051 rc = interpreter (set, set->set_arg, rc);
1057 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1060 if (request == NULL) {
1065 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1066 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1067 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1068 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1069 LASSERT(request->rq_cred);
1071 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1072 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1073 if (request->rq_import != NULL) {
1074 unsigned long flags = 0;
1076 spin_lock_irqsave(&request->rq_import->imp_lock, flags);
1077 list_del_init(&request->rq_replay_list);
1079 spin_unlock_irqrestore(&request->rq_import->imp_lock,
1082 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1084 if (atomic_read(&request->rq_refcount) != 0) {
1085 DEBUG_REQ(D_ERROR, request,
1086 "freeing request with nonzero refcount");
1090 if (request->rq_repbuf != NULL)
1091 ptlrpcs_cli_free_repbuf(request);
1092 if (request->rq_reqbuf != NULL)
1093 ptlrpcs_cli_free_reqbuf(request);
1095 if (request->rq_export != NULL) {
1096 class_export_put(request->rq_export);
1097 request->rq_export = NULL;
1099 if (request->rq_import != NULL) {
1100 class_import_put(request->rq_import);
1101 request->rq_import = NULL;
1103 if (request->rq_bulk != NULL)
1104 ptlrpc_free_bulk(request->rq_bulk);
1106 ptlrpcs_req_drop_cred(request);
1107 OBD_FREE(request, sizeof(*request));
1111 void ptlrpc_free_req(struct ptlrpc_request *request)
1113 __ptlrpc_free_req(request, 0);
1116 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1117 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1119 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1120 (void)__ptlrpc_req_finished(request, 1);
1123 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1126 if (request == NULL)
1129 if (request == LP_POISON ||
1130 request->rq_reqmsg == LP_POISON) {
1131 CERROR("dereferencing freed request (bug 575)\n");
1136 DEBUG_REQ(D_INFO, request, "refcount now %u",
1137 atomic_read(&request->rq_refcount) - 1);
1139 if (atomic_dec_and_test(&request->rq_refcount)) {
1140 __ptlrpc_free_req(request, locked);
1147 void ptlrpc_req_finished(struct ptlrpc_request *request)
1149 __ptlrpc_req_finished(request, 0);
1152 /* Disengage the client's reply buffer from the network
1153 * NB does _NOT_ unregister any client-side bulk.
1154 * IDEMPOTENT, but _not_ safe against concurrent callers.
1155 * The request owner (i.e. the thread doing the I/O) must call...
1157 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1160 wait_queue_head_t *wq;
1161 struct l_wait_info lwi;
1163 LASSERT(!in_interrupt ()); /* might sleep */
1165 if (!ptlrpc_client_receiving_reply(request))
1168 PtlMDUnlink (request->rq_reply_md_h);
1170 /* We have to l_wait_event() whatever the result, to give liblustre
1171 * a chance to run reply_in_callback() */
1173 if (request->rq_set != NULL)
1174 wq = &request->rq_set->set_waitq;
1176 wq = &request->rq_reply_waitq;
1179 /* Network access will complete in finite time but the HUGE
1180 * timeout lets us CWARN for visibility of sluggish NALs */
1181 lwi = LWI_TIMEOUT(300 * HZ, NULL, NULL);
1182 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1186 LASSERT (rc == -ETIMEDOUT);
1187 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1191 /* caller must hold imp->imp_lock */
1192 void ptlrpc_free_committed(struct obd_import *imp)
1194 struct list_head *tmp, *saved;
1195 struct ptlrpc_request *req;
1196 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1199 LASSERT(imp != NULL);
1201 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1203 CDEBUG(D_HA, "%s: committing for last_committed "LPU64"\n",
1204 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1206 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1207 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1209 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1210 LASSERT(req != last_req);
1213 if (req->rq_import_generation < imp->imp_generation) {
1214 DEBUG_REQ(D_HA, req, "freeing request with old gen");
1218 if (req->rq_replay) {
1219 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1223 /* not yet committed */
1224 if (req->rq_transno > imp->imp_peer_committed_transno) {
1225 DEBUG_REQ(D_HA, req, "stopping search");
1229 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1230 imp->imp_peer_committed_transno);
1232 if (req->rq_commit_cb != NULL)
1233 req->rq_commit_cb(req);
1234 list_del_init(&req->rq_replay_list);
1235 __ptlrpc_req_finished(req, 1);
1242 void ptlrpc_cleanup_client(struct obd_import *imp)
1249 void ptlrpc_resend_req(struct ptlrpc_request *req)
1251 unsigned long flags;
1253 DEBUG_REQ(D_HA, req, "going to resend");
1254 req->rq_reqmsg->handle.cookie = 0;
1255 req->rq_status = -EAGAIN;
1257 spin_lock_irqsave (&req->rq_lock, flags);
1259 req->rq_net_err = 0;
1260 req->rq_timedout = 0;
1262 __u64 old_xid = req->rq_xid;
1264 /* ensure previous bulk fails */
1265 req->rq_xid = ptlrpc_next_xid();
1266 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1267 old_xid, req->rq_xid);
1269 ptlrpc_wake_client_req(req);
1270 spin_unlock_irqrestore (&req->rq_lock, flags);
1273 /* XXX: this function and rq_status are currently unused */
1274 void ptlrpc_restart_req(struct ptlrpc_request *req)
1276 unsigned long flags;
1278 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1279 req->rq_status = -ERESTARTSYS;
1281 spin_lock_irqsave (&req->rq_lock, flags);
1282 req->rq_restart = 1;
1283 req->rq_timedout = 0;
1284 ptlrpc_wake_client_req(req);
1285 spin_unlock_irqrestore (&req->rq_lock, flags);
1288 static int expired_request(void *data)
1290 struct ptlrpc_request *req = data;
1293 /* some failure can suspend regular timeouts */
1294 if (ptlrpc_check_suspend())
1297 RETURN(ptlrpc_expire_one_request(req));
1300 static void interrupted_request(void *data)
1302 unsigned long flags;
1304 struct ptlrpc_request *req = data;
1305 DEBUG_REQ(D_HA, req, "request interrupted");
1306 spin_lock_irqsave (&req->rq_lock, flags);
1308 spin_unlock_irqrestore (&req->rq_lock, flags);
1311 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1314 atomic_inc(&req->rq_refcount);
1318 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1319 struct obd_import *imp)
1321 struct list_head *tmp;
1323 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1325 /* clear this for new requests that were resent as well
1326 as resent replayed requests. */
1327 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1329 /* don't re-add requests that have been replayed */
1330 if (!list_empty(&req->rq_replay_list))
1333 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1335 LASSERT(imp->imp_replayable);
1336 /* Balanced in ptlrpc_free_committed, usually. */
1337 ptlrpc_request_addref(req);
1338 list_for_each_prev(tmp, &imp->imp_replay_list) {
1339 struct ptlrpc_request *iter =
1340 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1342 /* We may have duplicate transnos if we create and then
1343 * open a file, or for closes retained if to match creating
1344 * opens, so use req->rq_xid as a secondary key.
1345 * (See bugs 684, 685, and 428.)
1346 * XXX no longer needed, but all opens need transnos!
1348 if (iter->rq_transno > req->rq_transno)
1351 if (iter->rq_transno == req->rq_transno) {
1352 LASSERT(iter->rq_xid != req->rq_xid);
1353 if (iter->rq_xid > req->rq_xid)
1357 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1361 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1364 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1366 char str[PTL_NALFMT_SIZE];
1369 struct l_wait_info lwi;
1370 struct obd_import *imp = req->rq_import;
1371 unsigned long flags;
1375 LASSERT(req->rq_set == NULL);
1376 LASSERT(!req->rq_receiving_reply);
1377 atomic_inc(&imp->imp_inflight);
1379 if (imp->imp_connection == NULL) {
1380 CERROR("request on not connected import %s\n",
1381 imp->imp_obd->obd_name);
1385 /* for distributed debugging */
1386 req->rq_reqmsg->status = current->pid;
1387 LASSERT(imp->imp_obd != NULL);
1388 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:ni:nid:opc "
1389 "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1390 imp->imp_obd->obd_uuid.uuid,
1391 req->rq_reqmsg->status, req->rq_xid,
1392 imp->imp_connection->c_peer.peer_ni->pni_name,
1393 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1394 req->rq_reqmsg->opc);
1396 /* Mark phase here for a little debug help */
1397 req->rq_phase = RQ_PHASE_RPC;
1399 spin_lock_irqsave(&imp->imp_lock, flags);
1400 req->rq_import_generation = imp->imp_generation;
1402 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1403 list_del(&req->rq_list);
1405 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1406 spin_unlock_irqrestore(&imp->imp_lock, flags);
1408 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1410 ptlrpc_import_state_name(req->rq_send_state),
1411 ptlrpc_import_state_name(imp->imp_state));
1412 lwi = LWI_INTR(interrupted_request, req);
1413 rc = l_wait_event(req->rq_reply_waitq,
1414 (req->rq_send_state == imp->imp_state ||
1417 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d == 1)",
1419 ptlrpc_import_state_name(imp->imp_state),
1420 ptlrpc_import_state_name(req->rq_send_state),
1423 spin_lock_irqsave(&imp->imp_lock, flags);
1424 list_del_init(&req->rq_list);
1429 else if (req->rq_intr) {
1432 else if (req->rq_no_resend) {
1433 spin_unlock_irqrestore(&imp->imp_lock, flags);
1434 GOTO(out, rc = -ETIMEDOUT);
1442 list_del_init(&req->rq_list);
1443 spin_unlock_irqrestore(&imp->imp_lock, flags);
1444 req->rq_status = rc; // XXX this ok?
1448 if (req->rq_resend) {
1449 if (!req->rq_ptlrpcs_restart)
1450 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1452 if (req->rq_bulk != NULL)
1453 ptlrpc_unregister_bulk (req);
1455 DEBUG_REQ(D_HA, req, "resending: ");
1458 /* XXX this is the same as ptlrpc_set_wait */
1459 LASSERT(list_empty(&req->rq_list));
1460 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1461 spin_unlock_irqrestore(&imp->imp_lock, flags);
1463 rc = ptl_send_rpc(req);
1465 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1468 timeout = MAX(req->rq_timeout * HZ, 1);
1469 DEBUG_REQ(D_NET, req, "-- sleeping for %d jiffies", timeout);
1472 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1474 rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1475 if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1477 DEBUG_REQ(D_NET, req, "-- done sleeping");
1479 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:ni:nid:opc "
1480 "%s:%s:%d:"LPU64":%s:%s:%d\n", current->comm,
1481 imp->imp_obd->obd_uuid.uuid,
1482 req->rq_reqmsg->status, req->rq_xid,
1483 imp->imp_connection->c_peer.peer_ni->pni_name,
1484 ptlrpc_peernid2str(&imp->imp_connection->c_peer, str),
1485 req->rq_reqmsg->opc);
1487 spin_lock_irqsave(&imp->imp_lock, flags);
1488 list_del_init(&req->rq_list);
1489 spin_unlock_irqrestore(&imp->imp_lock, flags);
1491 /* If the reply was received normally, this just grabs the spinlock
1492 * (ensuring the reply callback has returned), sees that
1493 * req->rq_receiving_reply is clear and returns. */
1494 ptlrpc_unregister_reply (req);
1497 GOTO(out, rc = -EIO);
1499 if (req->rq_ptlrpcs_err)
1500 GOTO(out, rc = -EPERM);
1502 /* Resend if we need to, unless we were interrupted. */
1503 if (req->rq_resend && !req->rq_intr) {
1504 /* ...unless we were specifically told otherwise. */
1505 if (req->rq_no_resend)
1506 GOTO(out, rc = -ETIMEDOUT);
1507 spin_lock_irqsave(&imp->imp_lock, flags);
1512 /* Should only be interrupted if we timed out. */
1513 if (!req->rq_timedout)
1514 DEBUG_REQ(D_ERROR, req,
1515 "rq_intr set but rq_timedout not");
1516 GOTO(out, rc = -EINTR);
1519 if (req->rq_timedout) { /* non-recoverable timeout */
1520 GOTO(out, rc = -ETIMEDOUT);
1523 if (!req->rq_replied) {
1524 /* How can this be? -eeb */
1525 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1527 GOTO(out, rc = req->rq_status);
1530 rc = after_reply (req);
1531 /* NB may return +ve success rc */
1532 if (req->rq_resend) {
1533 spin_lock_irqsave(&imp->imp_lock, flags);
1538 if (req->rq_bulk != NULL) {
1540 /* success so far. Note that anything going wrong
1541 * with bulk now, is EXTREMELY strange, since the
1542 * server must have believed that the bulk
1543 * tranferred OK before she replied with success to
1545 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1546 brc = l_wait_event(req->rq_reply_waitq,
1547 !ptlrpc_bulk_active(req->rq_bulk),
1549 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1551 LASSERT(brc == -ETIMEDOUT);
1552 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1554 } else if (!req->rq_bulk->bd_success) {
1555 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1560 ptlrpc_unregister_bulk (req);
1563 LASSERT(!req->rq_receiving_reply);
1564 req->rq_phase = RQ_PHASE_INTERPRET;
1566 atomic_dec(&imp->imp_inflight);
1567 wake_up(&imp->imp_recovery_waitq);
1571 struct ptlrpc_replay_async_args {
1573 int praa_old_status;
1576 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1577 void * data, int rc)
1579 struct ptlrpc_replay_async_args *aa = data;
1580 struct obd_import *imp = req->rq_import;
1581 unsigned long flags;
1583 atomic_dec(&imp->imp_replay_inflight);
1585 if (!req->rq_replied) {
1586 CERROR("request replay timed out, restarting recovery\n");
1587 GOTO(out, rc = -ETIMEDOUT);
1591 /* Clear reply swab mask; this is a new reply in sender's byte order */
1592 req->rq_rep_swab_mask = 0;
1594 LASSERT (req->rq_nob_received <= req->rq_repbuf_len);
1595 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
1597 CERROR("unpack_rep failed: %d\n", rc);
1598 GOTO(out, rc = -EPROTO);
1601 if (req->rq_repmsg->type == PTL_RPC_MSG_ERR &&
1602 req->rq_repmsg->status == -ENOTCONN)
1603 GOTO(out, rc = req->rq_repmsg->status);
1605 /* The transno had better not change over replay. */
1606 LASSERT(req->rq_reqmsg->transno == req->rq_repmsg->transno);
1608 DEBUG_REQ(D_HA, req, "got rep");
1610 /* let the callback do fixups, possibly including in the request */
1611 if (req->rq_replay_cb)
1612 req->rq_replay_cb(req);
1614 if (req->rq_replied && req->rq_repmsg->status != aa->praa_old_status) {
1615 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1616 req->rq_repmsg->status, aa->praa_old_status);
1618 /* Put it back for re-replay. */
1619 req->rq_repmsg->status = aa->praa_old_status;
1622 spin_lock_irqsave(&imp->imp_lock, flags);
1623 imp->imp_last_replay_transno = req->rq_transno;
1624 spin_unlock_irqrestore(&imp->imp_lock, flags);
1626 /* continue with recovery */
1627 rc = ptlrpc_import_recovery_state_machine(imp);
1629 req->rq_send_state = aa->praa_old_state;
1632 /* this replay failed, so restart recovery */
1633 ptlrpc_connect_import(imp, NULL);
1639 int ptlrpc_replay_req(struct ptlrpc_request *req)
1641 struct ptlrpc_replay_async_args *aa;
1644 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1646 /* Not handling automatic bulk replay yet (or ever?) */
1647 LASSERT(req->rq_bulk == NULL);
1649 DEBUG_REQ(D_HA, req, "REPLAY");
1651 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1652 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1653 memset(aa, 0, sizeof *aa);
1655 /* Prepare request to be resent with ptlrpcd */
1656 aa->praa_old_state = req->rq_send_state;
1657 req->rq_send_state = LUSTRE_IMP_REPLAY;
1658 req->rq_phase = RQ_PHASE_NEW;
1659 aa->praa_old_status = req->rq_repmsg->status;
1662 req->rq_interpret_reply = ptlrpc_replay_interpret;
1663 atomic_inc(&req->rq_import->imp_replay_inflight);
1664 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1666 ptlrpcd_add_req(req);
1670 void ptlrpc_abort_inflight(struct obd_import *imp)
1672 unsigned long flags;
1673 struct list_head *tmp, *n;
1676 /* Make sure that no new requests get processed for this import.
1677 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1678 * this flag and then putting requests on sending_list or delayed_list.
1680 spin_lock_irqsave(&imp->imp_lock, flags);
1682 /* XXX locking? Maybe we should remove each request with the list
1683 * locked? Also, how do we know if the requests on the list are
1684 * being freed at this time?
1686 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1687 struct ptlrpc_request *req =
1688 list_entry(tmp, struct ptlrpc_request, rq_list);
1690 DEBUG_REQ(D_HA, req, "inflight");
1692 spin_lock (&req->rq_lock);
1693 if (req->rq_import_generation < imp->imp_generation) {
1695 ptlrpc_wake_client_req(req);
1697 spin_unlock (&req->rq_lock);
1700 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1701 struct ptlrpc_request *req =
1702 list_entry(tmp, struct ptlrpc_request, rq_list);
1704 DEBUG_REQ(D_HA, req, "aborting waiting req");
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_rawrpc_list) {
1715 struct ptlrpc_request *req =
1716 list_entry(tmp, struct ptlrpc_request, rq_list);
1718 DEBUG_REQ(D_HA, req, "aborting raw rpc");
1720 spin_lock(&req->rq_lock);
1722 ptlrpc_wake_client_req(req);
1723 spin_unlock(&req->rq_lock);
1726 /* Last chance to free reqs left on the replay list, but we
1727 * will still leak reqs that haven't comitted. */
1728 if (imp->imp_replayable)
1729 ptlrpc_free_committed(imp);
1731 spin_unlock_irqrestore(&imp->imp_lock, flags);
1736 static __u64 ptlrpc_last_xid = 0;
1737 static spinlock_t ptlrpc_last_xid_lock = SPIN_LOCK_UNLOCKED;
1739 __u64 ptlrpc_next_xid(void)
1742 spin_lock(&ptlrpc_last_xid_lock);
1743 tmp = ++ptlrpc_last_xid;
1744 spin_unlock(&ptlrpc_last_xid_lock);