1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Copyright (c) 2002, 2003 Cluster File Systems, Inc.
6 * This file is part of the Lustre file system, http://www.lustre.org
7 * Lustre is a trademark of Cluster File Systems, Inc.
9 * You may have signed or agreed to another license before downloading
10 * this software. If so, you are bound by the terms and conditions
11 * of that agreement, and the following does not apply to you. See the
12 * LICENSE file included with this distribution for more information.
14 * If you did not agree to a different license, then this copy of Lustre
15 * is open source software; you can redistribute it and/or modify it
16 * under the terms of version 2 of the GNU General Public License as
17 * published by the Free Software Foundation.
19 * In either case, Lustre is distributed in the hope that it will be
20 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
21 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * license text for more details.
26 #define DEBUG_SUBSYSTEM S_RPC
30 #include <liblustre.h>
33 #include <obd_support.h>
34 #include <obd_class.h>
35 #include <lustre_lib.h>
36 #include <lustre_ha.h>
37 #include <lustre_import.h>
39 #include "ptlrpc_internal.h"
41 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
42 struct ptlrpc_client *cl)
44 cl->cli_request_portal = req_portal;
45 cl->cli_reply_portal = rep_portal;
49 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
51 struct ptlrpc_connection *c;
53 lnet_process_id_t peer;
56 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
58 CERROR("cannot find peer %s!\n", uuid->uuid);
62 c = ptlrpc_get_connection(peer, self, uuid);
64 memcpy(c->c_remote_uuid.uuid,
65 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
68 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
73 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
74 struct obd_uuid *uuid)
77 lnet_process_id_t peer;
80 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
82 CERROR("cannot find peer %s!\n", uuid->uuid);
91 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
93 struct ptlrpc_bulk_desc *desc;
95 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
99 spin_lock_init(&desc->bd_lock);
100 cfs_waitq_init(&desc->bd_waitq);
101 desc->bd_max_iov = npages;
102 desc->bd_iov_count = 0;
103 desc->bd_md_h = LNET_INVALID_HANDLE;
104 desc->bd_portal = portal;
105 desc->bd_type = type;
110 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
111 int npages, int type, int portal)
113 struct obd_import *imp = req->rq_import;
114 struct ptlrpc_bulk_desc *desc;
117 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
118 desc = new_bulk(npages, type, portal);
122 desc->bd_import_generation = req->rq_import_generation;
123 desc->bd_import = class_import_get(imp);
126 desc->bd_cbid.cbid_fn = client_bulk_callback;
127 desc->bd_cbid.cbid_arg = desc;
129 /* This makes req own desc, and free it when she frees herself */
135 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp (struct ptlrpc_request *req,
136 int npages, int type, int portal)
138 struct obd_export *exp = req->rq_export;
139 struct ptlrpc_bulk_desc *desc;
142 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
144 desc = new_bulk(npages, type, portal);
148 desc->bd_export = class_export_get(exp);
151 desc->bd_cbid.cbid_fn = server_bulk_callback;
152 desc->bd_cbid.cbid_arg = desc;
154 /* NB we don't assign rq_bulk here; server-side requests are
155 * re-used, and the handler frees the bulk desc explicitly. */
160 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
161 cfs_page_t *page, int pageoffset, int len)
163 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
164 LASSERT(page != NULL);
165 LASSERT(pageoffset >= 0);
167 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
171 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
174 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
178 LASSERT(desc != NULL);
179 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
180 LASSERT(!desc->bd_network_rw); /* network hands off or */
181 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
183 class_export_put(desc->bd_export);
185 class_import_put(desc->bd_import);
187 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
188 bd_iov[desc->bd_max_iov]));
192 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
194 struct list_head *l, *tmp;
195 struct ptlrpc_request *req;
200 list_for_each_safe(l, tmp, &pool->prp_req_list) {
201 req = list_entry(l, struct ptlrpc_request, rq_list);
202 list_del(&req->rq_list);
203 LASSERT (req->rq_reqmsg);
204 OBD_FREE(req->rq_reqmsg, pool->prp_rq_size);
205 OBD_FREE(req, sizeof(*req));
207 OBD_FREE(pool, sizeof(*pool));
210 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
215 while (size < pool->prp_rq_size)
218 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
219 "Trying to change pool size with nonempty pool "
220 "from %d to %d bytes\n", pool->prp_rq_size, size);
222 spin_lock(&pool->prp_lock);
223 pool->prp_rq_size = size;
224 for (i = 0; i < num_rq; i++) {
225 struct ptlrpc_request *req;
226 struct lustre_msg *msg;
228 spin_unlock(&pool->prp_lock);
229 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
232 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
234 OBD_FREE(req, sizeof(struct ptlrpc_request));
237 req->rq_reqmsg = msg;
239 spin_lock(&pool->prp_lock);
240 list_add_tail(&req->rq_list, &pool->prp_req_list);
242 spin_unlock(&pool->prp_lock);
246 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
247 void (*populate_pool)(struct ptlrpc_request_pool *, int))
249 struct ptlrpc_request_pool *pool;
251 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
255 /* Request next power of two for the allocation, because internally
256 kernel would do exactly this */
258 spin_lock_init(&pool->prp_lock);
259 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
260 pool->prp_rq_size = msgsize;
261 pool->prp_populate = populate_pool;
263 populate_pool(pool, num_rq);
265 if (list_empty(&pool->prp_req_list)) {
266 /* have not allocated a single request for the pool */
267 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
273 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
275 struct ptlrpc_request *request;
276 struct lustre_msg *reqmsg;
281 spin_lock(&pool->prp_lock);
283 /* See if we have anything in a pool, and bail out if nothing,
284 * in writeout path, where this matters, this is safe to do, because
285 * nothing is lost in this case, and when some in-flight requests
286 * complete, this code will be called again. */
287 if (unlikely(list_empty(&pool->prp_req_list))) {
288 spin_unlock(&pool->prp_lock);
292 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
294 list_del(&request->rq_list);
295 spin_unlock(&pool->prp_lock);
297 LASSERT(request->rq_reqmsg);
298 LASSERT(request->rq_pool);
300 reqmsg = request->rq_reqmsg;
301 memset(request, 0, sizeof(*request));
302 request->rq_reqmsg = reqmsg;
303 request->rq_pool = pool;
304 request->rq_reqlen = pool->prp_rq_size;
308 struct ptlrpc_request *
309 ptlrpc_prep_req_pool(struct obd_import *imp, __u32 version, int opcode,
310 int count, int *lengths, char **bufs,
311 struct ptlrpc_request_pool *pool)
313 struct ptlrpc_request *request = NULL;
317 /* The obd disconnected */
321 LASSERT(imp != LP_POISON);
322 LASSERT((unsigned long)imp->imp_client > 0x1000);
323 LASSERT(imp->imp_client != LP_POISON);
326 request = ptlrpc_prep_req_from_pool(pool);
329 OBD_ALLOC(request, sizeof(*request));
332 CERROR("request allocation out of memory\n");
336 rc = lustre_pack_request(request, imp->imp_msg_magic, count, lengths,
339 LASSERT(!request->rq_pool);
340 OBD_FREE(request, sizeof(*request));
344 lustre_msg_add_version(request->rq_reqmsg, version);
346 if (imp->imp_server_timeout)
347 request->rq_timeout = obd_timeout / 2;
349 request->rq_timeout = obd_timeout;
350 request->rq_send_state = LUSTRE_IMP_FULL;
351 request->rq_type = PTL_RPC_MSG_REQUEST;
352 request->rq_import = class_import_get(imp);
353 request->rq_export = NULL;
355 request->rq_req_cbid.cbid_fn = request_out_callback;
356 request->rq_req_cbid.cbid_arg = request;
358 request->rq_reply_cbid.cbid_fn = reply_in_callback;
359 request->rq_reply_cbid.cbid_arg = request;
361 request->rq_phase = RQ_PHASE_NEW;
363 /* XXX FIXME bug 249 */
364 request->rq_request_portal = imp->imp_client->cli_request_portal;
365 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
367 spin_lock_init(&request->rq_lock);
368 CFS_INIT_LIST_HEAD(&request->rq_list);
369 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
370 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
371 cfs_waitq_init(&request->rq_reply_waitq);
372 request->rq_xid = ptlrpc_next_xid();
373 atomic_set(&request->rq_refcount, 1);
375 lustre_msg_set_opc(request->rq_reqmsg, opcode);
376 lustre_msg_set_flags(request->rq_reqmsg, 0);
381 struct ptlrpc_request *
382 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
383 int *lengths, char **bufs)
385 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
389 struct ptlrpc_request_set *ptlrpc_prep_set(void)
391 struct ptlrpc_request_set *set;
394 OBD_ALLOC(set, sizeof *set);
397 CFS_INIT_LIST_HEAD(&set->set_requests);
398 cfs_waitq_init(&set->set_waitq);
399 set->set_remaining = 0;
400 spin_lock_init(&set->set_new_req_lock);
401 CFS_INIT_LIST_HEAD(&set->set_new_requests);
406 /* Finish with this set; opposite of prep_set. */
407 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
409 struct list_head *tmp;
410 struct list_head *next;
415 /* Requests on the set should either all be completed, or all be new */
416 expected_phase = (set->set_remaining == 0) ?
417 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
418 list_for_each (tmp, &set->set_requests) {
419 struct ptlrpc_request *req =
420 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
422 LASSERT(req->rq_phase == expected_phase);
426 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
428 list_for_each_safe(tmp, next, &set->set_requests) {
429 struct ptlrpc_request *req =
430 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
431 list_del_init(&req->rq_set_chain);
433 LASSERT(req->rq_phase == expected_phase);
435 if (req->rq_phase == RQ_PHASE_NEW) {
437 if (req->rq_interpret_reply != NULL) {
438 int (*interpreter)(struct ptlrpc_request *,
440 req->rq_interpret_reply;
442 /* higher level (i.e. LOV) failed;
443 * let the sub reqs clean up */
444 req->rq_status = -EBADR;
445 interpreter(req, &req->rq_async_args,
448 set->set_remaining--;
452 ptlrpc_req_finished (req);
455 LASSERT(set->set_remaining == 0);
457 OBD_FREE(set, sizeof(*set));
461 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
462 struct ptlrpc_request *req)
464 /* The set takes over the caller's request reference */
465 list_add_tail(&req->rq_set_chain, &set->set_requests);
467 set->set_remaining++;
469 atomic_inc(&req->rq_import->imp_inflight);
472 /* lock so many callers can add things, the context that owns the set
473 * is supposed to notice these and move them into the set proper. */
474 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
475 struct ptlrpc_request *req)
477 spin_lock(&set->set_new_req_lock);
478 /* The set takes over the caller's request reference */
479 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
481 spin_unlock(&set->set_new_req_lock);
485 * Based on the current state of the import, determine if the request
486 * can be sent, is an error, or should be delayed.
488 * Returns true if this request should be delayed. If false, and
489 * *status is set, then the request can not be sent and *status is the
490 * error code. If false and status is 0, then request can be sent.
492 * The imp->imp_lock must be held.
494 static int ptlrpc_import_delay_req(struct obd_import *imp,
495 struct ptlrpc_request *req, int *status)
500 LASSERT (status != NULL);
503 if (imp->imp_state == LUSTRE_IMP_NEW) {
504 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
507 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
508 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
510 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
511 imp->imp_state == LUSTRE_IMP_CONNECTING) {
512 /* allow CONNECT even if import is invalid */ ;
513 } else if (imp->imp_invalid) {
514 if (!imp->imp_deactive)
515 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
516 *status = -ESHUTDOWN; /* bz 12940 */
517 } else if (req->rq_import_generation != imp->imp_generation) {
518 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
520 } else if (req->rq_send_state != imp->imp_state) {
521 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
523 *status = -EWOULDBLOCK;
531 static int ptlrpc_check_reply(struct ptlrpc_request *req)
536 /* serialise with network callback */
537 spin_lock(&req->rq_lock);
542 if (req->rq_net_err && !req->rq_timedout) {
543 spin_unlock(&req->rq_lock);
544 rc = ptlrpc_expire_one_request(req);
545 spin_lock(&req->rq_lock);
559 spin_unlock(&req->rq_lock);
560 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
564 static int ptlrpc_check_status(struct ptlrpc_request *req)
569 err = lustre_msg_get_status(req->rq_repmsg);
570 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
571 struct obd_export *exp = req->rq_export;
572 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
574 LCONSOLE_ERROR_MSG(0x011, "an error ocurred while communicating"
575 " with %s The %s operation failed with %d\n",
576 exp ? obd_export_nid2str(exp) : "(no nid)",
577 ll_opcode2str(opc), err);
578 RETURN(err < 0 ? err : -EINVAL);
582 DEBUG_REQ(D_INFO, req, "status is %d", err);
583 } else if (err > 0) {
584 /* XXX: translate this error from net to host */
585 DEBUG_REQ(D_INFO, req, "status is %d", err);
591 static int after_reply(struct ptlrpc_request *req)
593 struct obd_import *imp = req->rq_import;
597 LASSERT(!req->rq_receiving_reply);
599 /* NB Until this point, the whole of the incoming message,
600 * including buflens, status etc is in the sender's byte order. */
603 /* Clear reply swab mask; this is a new reply in sender's byte order */
604 req->rq_rep_swab_mask = 0;
606 LASSERT (req->rq_nob_received <= req->rq_replen);
607 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_nob_received);
609 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
613 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
615 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
619 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
620 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
621 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
622 lustre_msg_get_type(req->rq_repmsg));
626 rc = ptlrpc_check_status(req);
628 /* Either we've been evicted, or the server has failed for
629 * some reason. Try to reconnect, and if that fails, punt to the
631 if (ptlrpc_recoverable_error(rc)) {
632 if (req->rq_send_state != LUSTRE_IMP_FULL ||
633 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
637 ptlrpc_request_handle_notconn(req);
642 /* Store transno in reqmsg for replay. */
643 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
644 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
646 if (req->rq_import->imp_replayable) {
647 spin_lock(&imp->imp_lock);
648 /* no point in adding already-committed requests to the replay
649 * list, we will just remove them immediately. b=9829 */
650 if (req->rq_transno != 0 &&
652 lustre_msg_get_last_committed(req->rq_repmsg) ||
654 ptlrpc_retain_replayable_request(req, imp);
655 else if (req->rq_commit_cb != NULL) {
656 spin_unlock(&imp->imp_lock);
657 req->rq_commit_cb(req);
658 spin_lock(&imp->imp_lock);
661 /* Replay-enabled imports return commit-status information. */
662 if (lustre_msg_get_last_committed(req->rq_repmsg))
663 imp->imp_peer_committed_transno =
664 lustre_msg_get_last_committed(req->rq_repmsg);
665 ptlrpc_free_committed(imp);
666 spin_unlock(&imp->imp_lock);
672 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
674 struct obd_import *imp;
678 LASSERT(req->rq_phase == RQ_PHASE_NEW);
679 req->rq_phase = RQ_PHASE_RPC;
681 imp = req->rq_import;
682 spin_lock(&imp->imp_lock);
684 req->rq_import_generation = imp->imp_generation;
686 if (ptlrpc_import_delay_req(imp, req, &rc)) {
687 spin_lock (&req->rq_lock);
689 spin_unlock (&req->rq_lock);
691 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
693 lustre_msg_get_status(req->rq_reqmsg) ,
694 ptlrpc_import_state_name(req->rq_send_state),
695 ptlrpc_import_state_name(imp->imp_state));
696 LASSERT(list_empty (&req->rq_list));
698 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
699 spin_unlock(&imp->imp_lock);
704 spin_unlock(&imp->imp_lock);
706 req->rq_phase = RQ_PHASE_INTERPRET;
710 /* XXX this is the same as ptlrpc_queue_wait */
711 LASSERT(list_empty(&req->rq_list));
712 list_add_tail(&req->rq_list, &imp->imp_sending_list);
713 spin_unlock(&imp->imp_lock);
715 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
716 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
717 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
718 imp->imp_obd->obd_uuid.uuid,
719 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
720 libcfs_nid2str(imp->imp_connection->c_peer.nid),
721 lustre_msg_get_opc(req->rq_reqmsg));
723 rc = ptl_send_rpc(req, 0);
725 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
732 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
733 int ptlrpc_check_set(struct ptlrpc_request_set *set)
735 struct list_head *tmp;
736 int force_timer_recalc = 0;
739 if (set->set_remaining == 0)
742 list_for_each(tmp, &set->set_requests) {
743 struct ptlrpc_request *req =
744 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
745 struct obd_import *imp = req->rq_import;
748 if (req->rq_phase == RQ_PHASE_NEW &&
749 ptlrpc_send_new_req(req)) {
750 force_timer_recalc = 1;
753 if (!(req->rq_phase == RQ_PHASE_RPC ||
754 req->rq_phase == RQ_PHASE_BULK ||
755 req->rq_phase == RQ_PHASE_INTERPRET ||
756 req->rq_phase == RQ_PHASE_COMPLETE)) {
757 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
761 if (req->rq_phase == RQ_PHASE_COMPLETE)
764 if (req->rq_phase == RQ_PHASE_INTERPRET)
765 GOTO(interpret, req->rq_status);
767 if (req->rq_net_err && !req->rq_timedout)
768 ptlrpc_expire_one_request(req);
771 ptlrpc_unregister_reply(req);
772 if (req->rq_status == 0)
773 req->rq_status = -EIO;
774 req->rq_phase = RQ_PHASE_INTERPRET;
776 spin_lock(&imp->imp_lock);
777 list_del_init(&req->rq_list);
778 spin_unlock(&imp->imp_lock);
780 GOTO(interpret, req->rq_status);
783 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
784 * will only be set after rq_timedout, but the oig waiting
785 * path sets rq_intr irrespective of whether ptlrpcd has
786 * seen a timeout. our policy is to only interpret
787 * interrupted rpcs after they have timed out */
788 if (req->rq_intr && (req->rq_timedout || req->rq_waiting)) {
789 /* NB could be on delayed list */
790 ptlrpc_unregister_reply(req);
791 req->rq_status = -EINTR;
792 req->rq_phase = RQ_PHASE_INTERPRET;
794 spin_lock(&imp->imp_lock);
795 list_del_init(&req->rq_list);
796 spin_unlock(&imp->imp_lock);
798 GOTO(interpret, req->rq_status);
801 if (req->rq_phase == RQ_PHASE_RPC) {
802 if (req->rq_timedout||req->rq_waiting||req->rq_resend) {
805 ptlrpc_unregister_reply(req);
807 spin_lock(&imp->imp_lock);
809 if (ptlrpc_import_delay_req(imp, req, &status)){
810 spin_unlock(&imp->imp_lock);
814 list_del_init(&req->rq_list);
816 req->rq_status = status;
817 req->rq_phase = RQ_PHASE_INTERPRET;
818 spin_unlock(&imp->imp_lock);
819 GOTO(interpret, req->rq_status);
821 if (req->rq_no_resend) {
822 req->rq_status = -ENOTCONN;
823 req->rq_phase = RQ_PHASE_INTERPRET;
824 spin_unlock(&imp->imp_lock);
825 GOTO(interpret, req->rq_status);
827 list_add_tail(&req->rq_list,
828 &imp->imp_sending_list);
830 spin_unlock(&imp->imp_lock);
833 if (req->rq_resend) {
834 lustre_msg_add_flags(req->rq_reqmsg,
837 __u64 old_xid = req->rq_xid;
839 ptlrpc_unregister_bulk (req);
841 /* ensure previous bulk fails */
842 req->rq_xid = ptlrpc_next_xid();
843 CDEBUG(D_HA, "resend bulk "
846 old_xid, req->rq_xid);
850 rc = ptl_send_rpc(req, 0);
852 DEBUG_REQ(D_HA, req, "send failed (%d)",
854 force_timer_recalc = 1;
857 /* need to reset the timeout */
858 force_timer_recalc = 1;
861 /* Still waiting for a reply? */
862 if (ptlrpc_client_receiving_reply(req))
865 /* Did we actually receive a reply? */
866 if (!ptlrpc_client_replied(req))
869 spin_lock(&imp->imp_lock);
870 list_del_init(&req->rq_list);
871 spin_unlock(&imp->imp_lock);
873 req->rq_status = after_reply(req);
874 if (req->rq_resend) {
875 /* Add this req to the delayed list so
876 it can be errored if the import is
877 evicted after recovery. */
878 spin_lock(&imp->imp_lock);
879 list_add_tail(&req->rq_list,
880 &imp->imp_delayed_list);
881 spin_unlock(&imp->imp_lock);
885 /* If there is no bulk associated with this request,
886 * then we're done and should let the interpreter
887 * process the reply. Similarly if the RPC returned
888 * an error, and therefore the bulk will never arrive.
890 if (req->rq_bulk == NULL || req->rq_status != 0) {
891 req->rq_phase = RQ_PHASE_INTERPRET;
892 GOTO(interpret, req->rq_status);
895 req->rq_phase = RQ_PHASE_BULK;
898 LASSERT(req->rq_phase == RQ_PHASE_BULK);
899 if (ptlrpc_bulk_active(req->rq_bulk))
902 if (!req->rq_bulk->bd_success) {
903 /* The RPC reply arrived OK, but the bulk screwed
904 * up! Dead wierd since the server told us the RPC
905 * was good after getting the REPLY for her GET or
906 * the ACK for her PUT. */
907 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
908 req->rq_status = -EIO;
909 req->rq_phase = RQ_PHASE_INTERPRET;
910 GOTO(interpret, req->rq_status);
913 req->rq_phase = RQ_PHASE_INTERPRET;
916 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
917 LASSERT(!req->rq_receiving_reply);
919 ptlrpc_unregister_reply(req);
920 if (req->rq_bulk != NULL)
921 ptlrpc_unregister_bulk (req);
923 req->rq_phase = RQ_PHASE_COMPLETE;
925 if (req->rq_interpret_reply != NULL) {
926 int (*interpreter)(struct ptlrpc_request *,void *,int) =
927 req->rq_interpret_reply;
928 req->rq_status = interpreter(req, &req->rq_async_args,
932 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
933 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
934 imp->imp_obd->obd_uuid.uuid,
935 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
936 libcfs_nid2str(imp->imp_connection->c_peer.nid),
937 lustre_msg_get_opc(req->rq_reqmsg));
939 set->set_remaining--;
941 atomic_dec(&imp->imp_inflight);
942 cfs_waitq_signal(&imp->imp_recovery_waitq);
945 /* If we hit an error, we want to recover promptly. */
946 RETURN(set->set_remaining == 0 || force_timer_recalc);
949 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
951 struct obd_import *imp = req->rq_import;
955 DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
956 req->rq_net_err ? "network error" : "timeout",
957 (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
959 if (imp != NULL && obd_debug_peer_on_timeout)
960 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
962 spin_lock(&req->rq_lock);
963 req->rq_timedout = 1;
964 spin_unlock(&req->rq_lock);
966 ptlrpc_unregister_reply (req);
968 if (obd_dump_on_timeout)
969 libcfs_debug_dumplog();
971 if (req->rq_bulk != NULL)
972 ptlrpc_unregister_bulk (req);
975 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
979 /* The DLM server doesn't want recovery run on its imports. */
980 if (imp->imp_dlm_fake)
983 /* If this request is for recovery or other primordial tasks,
984 * then error it out here. */
985 if (req->rq_send_state != LUSTRE_IMP_FULL ||
986 imp->imp_obd->obd_no_recov) {
987 spin_lock(&req->rq_lock);
988 req->rq_status = -ETIMEDOUT;
990 spin_unlock(&req->rq_lock);
994 /* if request can't be resend we can't wait answer after timeout */
995 if (req->rq_no_resend) {
996 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1000 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1005 int ptlrpc_expired_set(void *data)
1007 struct ptlrpc_request_set *set = data;
1008 struct list_head *tmp;
1009 time_t now = CURRENT_SECONDS;
1012 LASSERT(set != NULL);
1014 /* A timeout expired; see which reqs it applies to... */
1015 list_for_each (tmp, &set->set_requests) {
1016 struct ptlrpc_request *req =
1017 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1019 /* request in-flight? */
1020 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1022 (req->rq_phase == RQ_PHASE_BULK)))
1025 if (req->rq_timedout || /* already dealt with */
1026 req->rq_sent + req->rq_timeout > now) /* not expired */
1029 /* deal with this guy */
1030 ptlrpc_expire_one_request (req);
1033 /* When waiting for a whole set, we always to break out of the
1034 * sleep so we can recalculate the timeout, or enable interrupts
1035 * iff everyone's timed out.
1040 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1042 spin_lock(&req->rq_lock);
1044 spin_unlock(&req->rq_lock);
1047 void ptlrpc_interrupted_set(void *data)
1049 struct ptlrpc_request_set *set = data;
1050 struct list_head *tmp;
1052 LASSERT(set != NULL);
1053 CERROR("INTERRUPTED SET %p\n", set);
1055 list_for_each(tmp, &set->set_requests) {
1056 struct ptlrpc_request *req =
1057 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1059 if (req->rq_phase != RQ_PHASE_RPC)
1062 ptlrpc_mark_interrupted(req);
1066 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1068 struct list_head *tmp;
1069 time_t now = CURRENT_SECONDS;
1072 struct ptlrpc_request *req;
1075 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1077 list_for_each(tmp, &set->set_requests) {
1078 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1080 /* request in-flight? */
1081 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1082 (req->rq_phase == RQ_PHASE_BULK)))
1085 if (req->rq_timedout) /* already timed out */
1088 deadline = req->rq_sent + req->rq_timeout;
1089 if (deadline <= now) /* actually expired already */
1090 timeout = 1; /* ASAP */
1091 else if (timeout == 0 || timeout > deadline - now)
1092 timeout = deadline - now;
1097 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1099 struct list_head *tmp;
1100 struct ptlrpc_request *req;
1101 struct l_wait_info lwi;
1105 if (list_empty(&set->set_requests))
1108 list_for_each(tmp, &set->set_requests) {
1109 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1110 if (req->rq_phase == RQ_PHASE_NEW)
1111 (void)ptlrpc_send_new_req(req);
1115 timeout = ptlrpc_set_next_timeout(set);
1117 /* wait until all complete, interrupted, or an in-flight
1119 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1121 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1123 ptlrpc_interrupted_set, set);
1124 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1126 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1128 /* -EINTR => all requests have been flagged rq_intr so next
1130 * -ETIMEOUTD => someone timed out. When all reqs have
1131 * timed out, signals are enabled allowing completion with
1133 * I don't really care if we go once more round the loop in
1134 * the error cases -eeb. */
1135 } while (rc != 0 || set->set_remaining != 0);
1137 LASSERT(set->set_remaining == 0);
1140 list_for_each(tmp, &set->set_requests) {
1141 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1143 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1144 if (req->rq_status != 0)
1145 rc = req->rq_status;
1148 if (set->set_interpret != NULL) {
1149 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1151 rc = interpreter (set, set->set_arg, rc);
1157 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
1159 struct ptlrpc_request_pool *pool = request->rq_pool;
1161 spin_lock(&pool->prp_lock);
1162 list_add_tail(&request->rq_list, &pool->prp_req_list);
1163 spin_unlock(&pool->prp_lock);
1166 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1169 if (request == NULL) {
1174 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1175 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1176 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1177 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1179 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1180 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1181 if (request->rq_import != NULL) {
1183 spin_lock(&request->rq_import->imp_lock);
1184 list_del_init(&request->rq_replay_list);
1186 spin_unlock(&request->rq_import->imp_lock);
1188 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1190 if (atomic_read(&request->rq_refcount) != 0) {
1191 DEBUG_REQ(D_ERROR, request,
1192 "freeing request with nonzero refcount");
1196 if (request->rq_repmsg != NULL) {
1197 OBD_FREE(request->rq_repmsg, request->rq_replen);
1198 request->rq_repmsg = NULL;
1200 if (request->rq_export != NULL) {
1201 class_export_put(request->rq_export);
1202 request->rq_export = NULL;
1204 if (request->rq_import != NULL) {
1205 class_import_put(request->rq_import);
1206 request->rq_import = NULL;
1208 if (request->rq_bulk != NULL)
1209 ptlrpc_free_bulk(request->rq_bulk);
1211 if (request->rq_pool) {
1212 __ptlrpc_free_req_to_pool(request);
1214 if (request->rq_reqmsg != NULL) {
1215 OBD_FREE(request->rq_reqmsg, request->rq_reqlen);
1216 request->rq_reqmsg = NULL;
1218 OBD_FREE(request, sizeof(*request));
1223 void ptlrpc_free_req(struct ptlrpc_request *request)
1225 __ptlrpc_free_req(request, 0);
1228 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1229 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1231 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1232 (void)__ptlrpc_req_finished(request, 1);
1235 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1238 if (request == NULL)
1241 if (request == LP_POISON ||
1242 request->rq_reqmsg == LP_POISON) {
1243 CERROR("dereferencing freed request (bug 575)\n");
1248 DEBUG_REQ(D_INFO, request, "refcount now %u",
1249 atomic_read(&request->rq_refcount) - 1);
1251 if (atomic_dec_and_test(&request->rq_refcount)) {
1252 __ptlrpc_free_req(request, locked);
1259 void ptlrpc_req_finished(struct ptlrpc_request *request)
1261 __ptlrpc_req_finished(request, 0);
1264 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1266 return request->rq_xid;
1268 EXPORT_SYMBOL(ptlrpc_req_xid);
1270 /* Disengage the client's reply buffer from the network
1271 * NB does _NOT_ unregister any client-side bulk.
1272 * IDEMPOTENT, but _not_ safe against concurrent callers.
1273 * The request owner (i.e. the thread doing the I/O) must call...
1275 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1279 struct l_wait_info lwi;
1282 LASSERT(!in_interrupt ()); /* might sleep */
1284 if (!ptlrpc_client_receiving_reply(request))
1287 LNetMDUnlink (request->rq_reply_md_h);
1289 /* We have to l_wait_event() whatever the result, to give liblustre
1290 * a chance to run reply_in_callback() */
1292 if (request->rq_set != NULL)
1293 wq = &request->rq_set->set_waitq;
1295 wq = &request->rq_reply_waitq;
1298 /* Network access will complete in finite time but the HUGE
1299 * timeout lets us CWARN for visibility of sluggish NALs */
1300 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1301 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1305 LASSERT (rc == -ETIMEDOUT);
1306 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1311 /* caller must hold imp->imp_lock */
1312 void ptlrpc_free_committed(struct obd_import *imp)
1314 struct list_head *tmp, *saved;
1315 struct ptlrpc_request *req;
1316 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1319 LASSERT(imp != NULL);
1321 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1324 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1325 imp->imp_generation == imp->imp_last_generation_checked) {
1326 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1327 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1331 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1332 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1333 imp->imp_generation);
1334 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1335 imp->imp_last_generation_checked = imp->imp_generation;
1337 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1338 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1340 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1341 LASSERT(req != last_req);
1344 if (req->rq_import_generation < imp->imp_generation) {
1345 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1349 if (req->rq_replay) {
1350 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1354 /* not yet committed */
1355 if (req->rq_transno > imp->imp_peer_committed_transno) {
1356 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1360 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1361 imp->imp_peer_committed_transno);
1363 spin_lock(&req->rq_lock);
1365 spin_unlock(&req->rq_lock);
1366 if (req->rq_commit_cb != NULL)
1367 req->rq_commit_cb(req);
1368 list_del_init(&req->rq_replay_list);
1369 __ptlrpc_req_finished(req, 1);
1376 void ptlrpc_cleanup_client(struct obd_import *imp)
1383 void ptlrpc_resend_req(struct ptlrpc_request *req)
1385 DEBUG_REQ(D_HA, req, "going to resend");
1386 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1387 req->rq_status = -EAGAIN;
1389 spin_lock(&req->rq_lock);
1391 req->rq_net_err = 0;
1392 req->rq_timedout = 0;
1394 __u64 old_xid = req->rq_xid;
1396 /* ensure previous bulk fails */
1397 req->rq_xid = ptlrpc_next_xid();
1398 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1399 old_xid, req->rq_xid);
1401 ptlrpc_wake_client_req(req);
1402 spin_unlock(&req->rq_lock);
1405 /* XXX: this function and rq_status are currently unused */
1406 void ptlrpc_restart_req(struct ptlrpc_request *req)
1408 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1409 req->rq_status = -ERESTARTSYS;
1411 spin_lock(&req->rq_lock);
1412 req->rq_restart = 1;
1413 req->rq_timedout = 0;
1414 ptlrpc_wake_client_req(req);
1415 spin_unlock(&req->rq_lock);
1418 static int expired_request(void *data)
1420 struct ptlrpc_request *req = data;
1423 RETURN(ptlrpc_expire_one_request(req));
1426 static void interrupted_request(void *data)
1428 struct ptlrpc_request *req = data;
1429 DEBUG_REQ(D_HA, req, "request interrupted");
1430 spin_lock(&req->rq_lock);
1432 spin_unlock(&req->rq_lock);
1435 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1438 atomic_inc(&req->rq_refcount);
1442 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1443 struct obd_import *imp)
1445 struct list_head *tmp;
1447 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1449 /* clear this for new requests that were resent as well
1450 as resent replayed requests. */
1451 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1453 /* don't re-add requests that have been replayed */
1454 if (!list_empty(&req->rq_replay_list))
1457 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1459 LASSERT(imp->imp_replayable);
1460 /* Balanced in ptlrpc_free_committed, usually. */
1461 ptlrpc_request_addref(req);
1462 list_for_each_prev(tmp, &imp->imp_replay_list) {
1463 struct ptlrpc_request *iter =
1464 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1466 /* We may have duplicate transnos if we create and then
1467 * open a file, or for closes retained if to match creating
1468 * opens, so use req->rq_xid as a secondary key.
1469 * (See bugs 684, 685, and 428.)
1470 * XXX no longer needed, but all opens need transnos!
1472 if (iter->rq_transno > req->rq_transno)
1475 if (iter->rq_transno == req->rq_transno) {
1476 LASSERT(iter->rq_xid != req->rq_xid);
1477 if (iter->rq_xid > req->rq_xid)
1481 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1485 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1488 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1492 struct l_wait_info lwi;
1493 struct obd_import *imp = req->rq_import;
1494 cfs_duration_t timeout = 0;
1497 LASSERT(req->rq_set == NULL);
1498 LASSERT(!req->rq_receiving_reply);
1499 atomic_inc(&imp->imp_inflight);
1501 /* for distributed debugging */
1502 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1503 LASSERT(imp->imp_obd != NULL);
1504 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1505 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1506 imp->imp_obd->obd_uuid.uuid,
1507 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1508 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1509 lustre_msg_get_opc(req->rq_reqmsg));
1511 /* Mark phase here for a little debug help */
1512 req->rq_phase = RQ_PHASE_RPC;
1514 spin_lock(&imp->imp_lock);
1515 req->rq_import_generation = imp->imp_generation;
1517 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1518 list_del(&req->rq_list);
1520 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1521 spin_unlock(&imp->imp_lock);
1523 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1525 ptlrpc_import_state_name(req->rq_send_state),
1526 ptlrpc_import_state_name(imp->imp_state));
1527 lwi = LWI_INTR(interrupted_request, req);
1528 rc = l_wait_event(req->rq_reply_waitq,
1529 (req->rq_send_state == imp->imp_state ||
1530 req->rq_err || req->rq_intr),
1532 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1534 ptlrpc_import_state_name(imp->imp_state),
1535 ptlrpc_import_state_name(req->rq_send_state),
1536 req->rq_err, req->rq_intr);
1538 spin_lock(&imp->imp_lock);
1539 list_del_init(&req->rq_list);
1544 else if (req->rq_intr) {
1547 else if (req->rq_no_resend) {
1548 spin_unlock(&imp->imp_lock);
1549 GOTO(out, rc = -ETIMEDOUT);
1557 list_del_init(&req->rq_list);
1558 spin_unlock(&imp->imp_lock);
1559 req->rq_status = rc; // XXX this ok?
1563 if (req->rq_resend) {
1564 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1566 if (req->rq_bulk != NULL) {
1567 ptlrpc_unregister_bulk (req);
1569 /* bulk requests are supposed to be
1570 * idempotent, so we are free to bump the xid
1571 * here, which we need to do before
1572 * registering the bulk again (bug 6371).
1573 * print the old xid first for sanity.
1575 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1576 req->rq_xid = ptlrpc_next_xid();
1579 DEBUG_REQ(D_HA, req, "resending: ");
1582 /* XXX this is the same as ptlrpc_set_wait */
1583 LASSERT(list_empty(&req->rq_list));
1584 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1585 spin_unlock(&imp->imp_lock);
1587 rc = ptl_send_rpc(req, 0);
1589 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1592 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1593 DEBUG_REQ(D_NET, req,
1594 "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1596 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1598 l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1600 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1601 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1602 imp->imp_obd->obd_uuid.uuid,
1603 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1604 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1605 lustre_msg_get_opc(req->rq_reqmsg));
1607 spin_lock(&imp->imp_lock);
1608 list_del_init(&req->rq_list);
1609 spin_unlock(&imp->imp_lock);
1611 /* If the reply was received normally, this just grabs the spinlock
1612 * (ensuring the reply callback has returned), sees that
1613 * req->rq_receiving_reply is clear and returns. */
1614 ptlrpc_unregister_reply (req);
1617 GOTO(out, rc = -EIO);
1619 /* Resend if we need to, unless we were interrupted. */
1620 if (req->rq_resend && !req->rq_intr) {
1621 /* ...unless we were specifically told otherwise. */
1622 if (req->rq_no_resend)
1623 GOTO(out, rc = -ETIMEDOUT);
1624 spin_lock(&imp->imp_lock);
1629 /* Should only be interrupted if we timed out. */
1630 if (!req->rq_timedout)
1631 DEBUG_REQ(D_ERROR, req,
1632 "rq_intr set but rq_timedout not");
1633 GOTO(out, rc = -EINTR);
1636 if (req->rq_timedout) { /* non-recoverable timeout */
1637 GOTO(out, rc = -ETIMEDOUT);
1640 if (!req->rq_replied) {
1641 /* How can this be? -eeb */
1642 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1644 GOTO(out, rc = req->rq_status);
1647 rc = after_reply(req);
1648 /* NB may return +ve success rc */
1649 if (req->rq_resend) {
1650 spin_lock(&imp->imp_lock);
1655 if (req->rq_bulk != NULL) {
1657 /* success so far. Note that anything going wrong
1658 * with bulk now, is EXTREMELY strange, since the
1659 * server must have believed that the bulk
1660 * tranferred OK before she replied with success to
1662 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1663 brc = l_wait_event(req->rq_reply_waitq,
1664 !ptlrpc_bulk_active(req->rq_bulk),
1666 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1668 LASSERT(brc == -ETIMEDOUT);
1669 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1671 } else if (!req->rq_bulk->bd_success) {
1672 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1677 ptlrpc_unregister_bulk (req);
1680 LASSERT(!req->rq_receiving_reply);
1681 req->rq_phase = RQ_PHASE_INTERPRET;
1683 atomic_dec(&imp->imp_inflight);
1684 cfs_waitq_signal(&imp->imp_recovery_waitq);
1688 struct ptlrpc_replay_async_args {
1690 int praa_old_status;
1693 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1694 void * data, int rc)
1696 struct ptlrpc_replay_async_args *aa = data;
1697 struct obd_import *imp = req->rq_import;
1700 atomic_dec(&imp->imp_replay_inflight);
1702 if (!req->rq_replied) {
1703 CERROR("request replay timed out, restarting recovery\n");
1704 GOTO(out, rc = -ETIMEDOUT);
1707 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1708 lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN)
1709 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1711 /* The transno had better not change over replay. */
1712 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1713 lustre_msg_get_transno(req->rq_repmsg));
1715 DEBUG_REQ(D_HA, req, "got rep");
1717 /* let the callback do fixups, possibly including in the request */
1718 if (req->rq_replay_cb)
1719 req->rq_replay_cb(req);
1721 if (req->rq_replied &&
1722 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1723 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1724 lustre_msg_get_status(req->rq_repmsg),
1725 aa->praa_old_status);
1727 /* Put it back for re-replay. */
1728 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1731 spin_lock(&imp->imp_lock);
1732 imp->imp_last_replay_transno = req->rq_transno;
1733 spin_unlock(&imp->imp_lock);
1735 /* continue with recovery */
1736 rc = ptlrpc_import_recovery_state_machine(imp);
1738 req->rq_send_state = aa->praa_old_state;
1741 /* this replay failed, so restart recovery */
1742 ptlrpc_connect_import(imp, NULL);
1748 int ptlrpc_replay_req(struct ptlrpc_request *req)
1750 struct ptlrpc_replay_async_args *aa;
1753 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1755 /* Not handling automatic bulk replay yet (or ever?) */
1756 LASSERT(req->rq_bulk == NULL);
1758 DEBUG_REQ(D_HA, req, "REPLAY");
1760 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1761 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1762 memset(aa, 0, sizeof *aa);
1764 /* Prepare request to be resent with ptlrpcd */
1765 aa->praa_old_state = req->rq_send_state;
1766 req->rq_send_state = LUSTRE_IMP_REPLAY;
1767 req->rq_phase = RQ_PHASE_NEW;
1768 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
1771 req->rq_interpret_reply = ptlrpc_replay_interpret;
1772 atomic_inc(&req->rq_import->imp_replay_inflight);
1773 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1775 ptlrpcd_add_req(req);
1779 void ptlrpc_abort_inflight(struct obd_import *imp)
1781 struct list_head *tmp, *n;
1784 /* Make sure that no new requests get processed for this import.
1785 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1786 * this flag and then putting requests on sending_list or delayed_list.
1788 spin_lock(&imp->imp_lock);
1790 /* XXX locking? Maybe we should remove each request with the list
1791 * locked? Also, how do we know if the requests on the list are
1792 * being freed at this time?
1794 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1795 struct ptlrpc_request *req =
1796 list_entry(tmp, struct ptlrpc_request, rq_list);
1798 DEBUG_REQ(D_RPCTRACE, req, "inflight");
1800 spin_lock (&req->rq_lock);
1801 if (req->rq_import_generation < imp->imp_generation) {
1803 ptlrpc_wake_client_req(req);
1805 spin_unlock (&req->rq_lock);
1808 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1809 struct ptlrpc_request *req =
1810 list_entry(tmp, struct ptlrpc_request, rq_list);
1812 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
1814 spin_lock (&req->rq_lock);
1815 if (req->rq_import_generation < imp->imp_generation) {
1817 ptlrpc_wake_client_req(req);
1819 spin_unlock (&req->rq_lock);
1822 /* Last chance to free reqs left on the replay list, but we
1823 * will still leak reqs that haven't committed. */
1824 if (imp->imp_replayable)
1825 ptlrpc_free_committed(imp);
1827 spin_unlock(&imp->imp_lock);
1832 static __u64 ptlrpc_last_xid = 0;
1833 spinlock_t ptlrpc_last_xid_lock;
1835 __u64 ptlrpc_next_xid(void)
1838 spin_lock(&ptlrpc_last_xid_lock);
1839 tmp = ++ptlrpc_last_xid;
1840 spin_unlock(&ptlrpc_last_xid_lock);
1844 __u64 ptlrpc_sample_next_xid(void)
1847 spin_lock(&ptlrpc_last_xid_lock);
1848 tmp = ptlrpc_last_xid + 1;
1849 spin_unlock(&ptlrpc_last_xid_lock);
1852 EXPORT_SYMBOL(ptlrpc_sample_next_xid);