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 sptlrpc_enc_pool_put_pages(desc);
186 class_export_put(desc->bd_export);
188 class_import_put(desc->bd_import);
190 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
191 bd_iov[desc->bd_max_iov]));
195 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
197 struct list_head *l, *tmp;
198 struct ptlrpc_request *req;
203 list_for_each_safe(l, tmp, &pool->prp_req_list) {
204 req = list_entry(l, struct ptlrpc_request, rq_list);
205 list_del(&req->rq_list);
206 LASSERT(req->rq_reqbuf);
207 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
208 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
209 OBD_FREE(req, sizeof(*req));
211 OBD_FREE(pool, sizeof(*pool));
214 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
219 while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
222 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
223 "Trying to change pool size with nonempty pool "
224 "from %d to %d bytes\n", pool->prp_rq_size, size);
226 spin_lock(&pool->prp_lock);
227 pool->prp_rq_size = size;
228 for (i = 0; i < num_rq; i++) {
229 struct ptlrpc_request *req;
230 struct lustre_msg *msg;
232 spin_unlock(&pool->prp_lock);
233 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
236 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
238 OBD_FREE(req, sizeof(struct ptlrpc_request));
241 req->rq_reqbuf = msg;
242 req->rq_reqbuf_len = size;
244 spin_lock(&pool->prp_lock);
245 list_add_tail(&req->rq_list, &pool->prp_req_list);
247 spin_unlock(&pool->prp_lock);
251 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
252 void (*populate_pool)(struct ptlrpc_request_pool *, int))
254 struct ptlrpc_request_pool *pool;
256 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
260 /* Request next power of two for the allocation, because internally
261 kernel would do exactly this */
263 spin_lock_init(&pool->prp_lock);
264 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
265 pool->prp_rq_size = msgsize;
266 pool->prp_populate = populate_pool;
268 populate_pool(pool, num_rq);
270 if (list_empty(&pool->prp_req_list)) {
271 /* have not allocated a single request for the pool */
272 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
278 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
280 struct ptlrpc_request *request;
281 struct lustre_msg *reqbuf;
286 spin_lock(&pool->prp_lock);
288 /* See if we have anything in a pool, and bail out if nothing,
289 * in writeout path, where this matters, this is safe to do, because
290 * nothing is lost in this case, and when some in-flight requests
291 * complete, this code will be called again. */
292 if (unlikely(list_empty(&pool->prp_req_list))) {
293 spin_unlock(&pool->prp_lock);
297 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
299 list_del(&request->rq_list);
300 spin_unlock(&pool->prp_lock);
302 LASSERT(request->rq_reqbuf);
303 LASSERT(request->rq_pool);
305 reqbuf = request->rq_reqbuf;
306 memset(request, 0, sizeof(*request));
307 request->rq_reqbuf = reqbuf;
308 request->rq_reqbuf_len = pool->prp_rq_size;
309 request->rq_pool = pool;
313 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
315 struct ptlrpc_request_pool *pool = request->rq_pool;
317 spin_lock(&pool->prp_lock);
318 LASSERT(list_empty(&request->rq_list));
319 list_add_tail(&request->rq_list, &pool->prp_req_list);
320 spin_unlock(&pool->prp_lock);
323 struct ptlrpc_request *
324 ptlrpc_prep_req_pool(struct obd_import *imp, __u32 version, int opcode,
325 int count, int *lengths, char **bufs,
326 struct ptlrpc_request_pool *pool,
327 struct ptlrpc_cli_ctx *ctx)
329 struct ptlrpc_request *request = NULL;
333 /* The obd disconnected */
337 LASSERT(imp != LP_POISON);
338 LASSERT((unsigned long)imp->imp_client > 0x1000);
339 LASSERT(imp->imp_client != LP_POISON);
342 request = ptlrpc_prep_req_from_pool(pool);
345 OBD_ALLOC(request, sizeof(*request));
348 CERROR("request allocation out of memory\n");
352 request->rq_import = class_import_get(imp);
355 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
357 rc = sptlrpc_req_get_ctx(request);
362 sptlrpc_req_set_flavor(request, opcode);
364 rc = lustre_pack_request(request, imp->imp_msg_magic, count, lengths,
367 LASSERT(!request->rq_pool);
371 lustre_msg_add_version(request->rq_reqmsg, version);
373 if (imp->imp_server_timeout)
374 request->rq_timeout = obd_timeout / 2;
376 request->rq_timeout = obd_timeout;
377 request->rq_send_state = LUSTRE_IMP_FULL;
378 request->rq_type = PTL_RPC_MSG_REQUEST;
379 request->rq_export = NULL;
381 request->rq_req_cbid.cbid_fn = request_out_callback;
382 request->rq_req_cbid.cbid_arg = request;
384 request->rq_reply_cbid.cbid_fn = reply_in_callback;
385 request->rq_reply_cbid.cbid_arg = request;
387 request->rq_phase = RQ_PHASE_NEW;
389 /* XXX FIXME bug 249 */
390 request->rq_request_portal = imp->imp_client->cli_request_portal;
391 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
393 spin_lock_init(&request->rq_lock);
394 CFS_INIT_LIST_HEAD(&request->rq_list);
395 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
396 CFS_INIT_LIST_HEAD(&request->rq_mod_list);
397 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
398 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
399 cfs_waitq_init(&request->rq_reply_waitq);
400 request->rq_xid = ptlrpc_next_xid();
401 atomic_set(&request->rq_refcount, 1);
403 lustre_msg_set_opc(request->rq_reqmsg, opcode);
404 lustre_msg_set_flags(request->rq_reqmsg, 0);
408 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
410 class_import_put(imp);
411 if (request->rq_pool)
412 __ptlrpc_free_req_to_pool(request);
414 OBD_FREE(request, sizeof(*request));
418 struct ptlrpc_request *
419 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
420 int *lengths, char **bufs)
422 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
426 struct ptlrpc_request_set *ptlrpc_prep_set(void)
428 struct ptlrpc_request_set *set;
431 OBD_ALLOC(set, sizeof *set);
434 CFS_INIT_LIST_HEAD(&set->set_requests);
435 cfs_waitq_init(&set->set_waitq);
436 set->set_remaining = 0;
437 spin_lock_init(&set->set_new_req_lock);
438 CFS_INIT_LIST_HEAD(&set->set_new_requests);
443 /* Finish with this set; opposite of prep_set. */
444 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
446 struct list_head *tmp;
447 struct list_head *next;
452 /* Requests on the set should either all be completed, or all be new */
453 expected_phase = (set->set_remaining == 0) ?
454 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
455 list_for_each (tmp, &set->set_requests) {
456 struct ptlrpc_request *req =
457 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
459 LASSERT(req->rq_phase == expected_phase);
463 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
465 list_for_each_safe(tmp, next, &set->set_requests) {
466 struct ptlrpc_request *req =
467 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
468 list_del_init(&req->rq_set_chain);
470 LASSERT(req->rq_phase == expected_phase);
472 if (req->rq_phase == RQ_PHASE_NEW) {
474 if (req->rq_interpret_reply != NULL) {
475 int (*interpreter)(struct ptlrpc_request *,
477 req->rq_interpret_reply;
479 /* higher level (i.e. LOV) failed;
480 * let the sub reqs clean up */
481 req->rq_status = -EBADR;
482 interpreter(req, &req->rq_async_args,
485 set->set_remaining--;
489 ptlrpc_req_finished (req);
492 LASSERT(set->set_remaining == 0);
494 OBD_FREE(set, sizeof(*set));
498 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
499 struct ptlrpc_request *req)
501 /* The set takes over the caller's request reference */
502 list_add_tail(&req->rq_set_chain, &set->set_requests);
504 set->set_remaining++;
506 atomic_inc(&req->rq_import->imp_inflight);
509 /* lock so many callers can add things, the context that owns the set
510 * is supposed to notice these and move them into the set proper. */
511 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
512 struct ptlrpc_request *req)
514 spin_lock(&set->set_new_req_lock);
515 /* The set takes over the caller's request reference */
516 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
518 spin_unlock(&set->set_new_req_lock);
522 * Based on the current state of the import, determine if the request
523 * can be sent, is an error, or should be delayed.
525 * Returns true if this request should be delayed. If false, and
526 * *status is set, then the request can not be sent and *status is the
527 * error code. If false and status is 0, then request can be sent.
529 * The imp->imp_lock must be held.
531 static int ptlrpc_import_delay_req(struct obd_import *imp,
532 struct ptlrpc_request *req, int *status)
537 LASSERT (status != NULL);
540 if (req->rq_ctx_init || req->rq_ctx_fini) {
541 /* always allow ctx init/fini rpc go through */
542 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
543 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
546 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
547 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
549 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
550 imp->imp_state == LUSTRE_IMP_CONNECTING) {
551 /* allow CONNECT even if import is invalid */ ;
552 } else if (imp->imp_invalid) {
553 /* If the import has been invalidated (such as by an OST
554 * failure), the request must fail with -EIO. */
555 if (!imp->imp_deactive)
556 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
558 } else if (req->rq_import_generation != imp->imp_generation) {
559 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
561 } else if (req->rq_send_state != imp->imp_state) {
562 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
564 *status = -EWOULDBLOCK;
572 static int ptlrpc_check_reply(struct ptlrpc_request *req)
577 /* serialise with network callback */
578 spin_lock(&req->rq_lock);
583 if (req->rq_net_err && !req->rq_timedout) {
584 spin_unlock(&req->rq_lock);
585 rc = ptlrpc_expire_one_request(req);
586 spin_lock(&req->rq_lock);
600 spin_unlock(&req->rq_lock);
601 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
605 static int ptlrpc_check_status(struct ptlrpc_request *req)
610 err = lustre_msg_get_status(req->rq_repmsg);
611 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
612 struct obd_import *imp = req->rq_import;
613 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
614 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
615 " with %s. The %s operation failed with %d\n",
616 libcfs_nid2str(imp->imp_connection->c_peer.nid),
617 ll_opcode2str(opc), err);
618 RETURN(err < 0 ? err : -EINVAL);
622 DEBUG_REQ(D_INFO, req, "status is %d", err);
623 } else if (err > 0) {
624 /* XXX: translate this error from net to host */
625 DEBUG_REQ(D_INFO, req, "status is %d", err);
631 static int after_reply(struct ptlrpc_request *req)
633 struct obd_import *imp = req->rq_import;
637 LASSERT(!req->rq_receiving_reply);
638 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
640 /* NB Until this point, the whole of the incoming message,
641 * including buflens, status etc is in the sender's byte order. */
643 /* Clear reply swab mask; this is a new reply in sender's byte order */
644 req->rq_rep_swab_mask = 0;
646 rc = sptlrpc_cli_unwrap_reply(req);
648 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
652 /* security layer unwrap might ask resend this request */
656 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
658 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
662 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
664 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
668 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
669 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
670 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
671 lustre_msg_get_type(req->rq_repmsg));
675 rc = ptlrpc_check_status(req);
676 imp->imp_connect_error = rc;
679 /* Either we've been evicted, or the server has failed for
680 * some reason. Try to reconnect, and if that fails, punt to
682 if (ll_rpc_recoverable_error(rc)) {
683 if (req->rq_send_state != LUSTRE_IMP_FULL ||
684 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
687 ptlrpc_request_handle_notconn(req);
691 /* Let's look if server send slv. Do it only for RPC with
693 if (imp->imp_obd->obd_namespace) {
694 /* Disconnect rpc is sent when namespace is already
695 * destroyed. Let's check this and will not try update
697 ldlm_cli_update_pool(req);
701 /* Store transno in reqmsg for replay. */
702 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
703 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
705 if (req->rq_import->imp_replayable) {
706 spin_lock(&imp->imp_lock);
707 /* no point in adding already-committed requests to the replay
708 * list, we will just remove them immediately. b=9829 */
709 if (req->rq_transno != 0 &&
711 lustre_msg_get_last_committed(req->rq_repmsg) ||
713 ptlrpc_retain_replayable_request(req, imp);
714 else if (req->rq_commit_cb != NULL) {
715 spin_unlock(&imp->imp_lock);
716 req->rq_commit_cb(req);
717 spin_lock(&imp->imp_lock);
720 /* Replay-enabled imports return commit-status information. */
721 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
722 imp->imp_peer_committed_transno =
723 lustre_msg_get_last_committed(req->rq_repmsg);
725 ptlrpc_free_committed(imp);
726 spin_unlock(&imp->imp_lock);
732 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
734 struct obd_import *imp;
738 LASSERT(req->rq_phase == RQ_PHASE_NEW);
739 if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
742 req->rq_phase = RQ_PHASE_RPC;
744 imp = req->rq_import;
745 spin_lock(&imp->imp_lock);
747 req->rq_import_generation = imp->imp_generation;
749 if (ptlrpc_import_delay_req(imp, req, &rc)) {
750 spin_lock (&req->rq_lock);
752 spin_unlock (&req->rq_lock);
754 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
756 lustre_msg_get_status(req->rq_reqmsg) ,
757 ptlrpc_import_state_name(req->rq_send_state),
758 ptlrpc_import_state_name(imp->imp_state));
759 LASSERT(list_empty (&req->rq_list));
761 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
762 spin_unlock(&imp->imp_lock);
767 spin_unlock(&imp->imp_lock);
769 req->rq_phase = RQ_PHASE_INTERPRET;
773 /* XXX this is the same as ptlrpc_queue_wait */
774 LASSERT(list_empty(&req->rq_list));
775 list_add_tail(&req->rq_list, &imp->imp_sending_list);
776 spin_unlock(&imp->imp_lock);
778 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
780 rc = sptlrpc_req_refresh_ctx(req, -1);
786 /* here begins timeout counting */
787 req->rq_sent = CURRENT_SECONDS;
788 req->rq_wait_ctx = 1;
793 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
794 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
795 imp->imp_obd->obd_uuid.uuid,
796 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
797 libcfs_nid2str(imp->imp_connection->c_peer.nid),
798 lustre_msg_get_opc(req->rq_reqmsg));
800 rc = ptl_send_rpc(req, 0);
802 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
809 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
810 int ptlrpc_check_set(struct ptlrpc_request_set *set)
812 struct list_head *tmp;
813 int force_timer_recalc = 0;
816 if (set->set_remaining == 0)
819 list_for_each(tmp, &set->set_requests) {
820 struct ptlrpc_request *req =
821 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
822 struct obd_import *imp = req->rq_import;
825 if (req->rq_phase == RQ_PHASE_NEW &&
826 ptlrpc_send_new_req(req)) {
827 force_timer_recalc = 1;
829 /* delayed send - skip */
830 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
833 if (!(req->rq_phase == RQ_PHASE_RPC ||
834 req->rq_phase == RQ_PHASE_BULK ||
835 req->rq_phase == RQ_PHASE_INTERPRET ||
836 req->rq_phase == RQ_PHASE_COMPLETE)) {
837 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
841 if (req->rq_phase == RQ_PHASE_COMPLETE)
844 if (req->rq_phase == RQ_PHASE_INTERPRET)
845 GOTO(interpret, req->rq_status);
847 if (req->rq_net_err && !req->rq_timedout)
848 ptlrpc_expire_one_request(req);
851 ptlrpc_unregister_reply(req);
853 if (req->rq_status == 0)
854 req->rq_status = -EIO;
855 req->rq_phase = RQ_PHASE_INTERPRET;
857 spin_lock(&imp->imp_lock);
858 list_del_init(&req->rq_list);
859 spin_unlock(&imp->imp_lock);
861 GOTO(interpret, req->rq_status);
864 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
865 * will only be set after rq_timedout, but the oig waiting
866 * path sets rq_intr irrespective of whether ptlrpcd has
867 * seen a timeout. our policy is to only interpret
868 * interrupted rpcs after they have timed out */
869 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
871 /* NB could be on delayed list */
872 ptlrpc_unregister_reply(req);
873 req->rq_status = -EINTR;
874 req->rq_phase = RQ_PHASE_INTERPRET;
876 spin_lock(&imp->imp_lock);
877 list_del_init(&req->rq_list);
878 spin_unlock(&imp->imp_lock);
880 GOTO(interpret, req->rq_status);
883 if (req->rq_phase == RQ_PHASE_RPC) {
884 if (req->rq_timedout || req->rq_resend ||
885 req->rq_waiting || req->rq_wait_ctx) {
888 /* rq_wait_ctx is only touched in ptlrpcd,
889 * no lock needed here.
891 if (req->rq_wait_ctx)
894 ptlrpc_unregister_reply(req);
896 spin_lock(&imp->imp_lock);
898 if (ptlrpc_import_delay_req(imp, req, &status)){
899 spin_unlock(&imp->imp_lock);
903 list_del_init(&req->rq_list);
905 req->rq_status = status;
906 req->rq_phase = RQ_PHASE_INTERPRET;
907 spin_unlock(&imp->imp_lock);
908 GOTO(interpret, req->rq_status);
910 if (req->rq_no_resend && !req->rq_wait_ctx) {
911 req->rq_status = -ENOTCONN;
912 req->rq_phase = RQ_PHASE_INTERPRET;
913 spin_unlock(&imp->imp_lock);
914 GOTO(interpret, req->rq_status);
916 list_add_tail(&req->rq_list,
917 &imp->imp_sending_list);
919 spin_unlock(&imp->imp_lock);
922 if (req->rq_resend) {
923 lustre_msg_add_flags(req->rq_reqmsg,
926 __u64 old_xid = req->rq_xid;
928 ptlrpc_unregister_bulk (req);
930 /* ensure previous bulk fails */
931 req->rq_xid = ptlrpc_next_xid();
932 CDEBUG(D_HA, "resend bulk "
935 old_xid, req->rq_xid);
939 status = sptlrpc_req_refresh_ctx(req, -1);
942 req->rq_status = status;
943 force_timer_recalc = 1;
945 if (!req->rq_wait_ctx) {
946 /* begins timeout counting */
947 req->rq_sent = CURRENT_SECONDS;
948 req->rq_wait_ctx = 1;
953 req->rq_wait_ctx = 0;
956 rc = ptl_send_rpc(req, 0);
958 DEBUG_REQ(D_HA, req, "send failed (%d)",
960 force_timer_recalc = 1;
963 /* need to reset the timeout */
964 force_timer_recalc = 1;
967 /* Still waiting for a reply? */
968 if (ptlrpc_client_receiving_reply(req))
971 /* Did we actually receive a reply? */
972 if (!ptlrpc_client_replied(req))
975 spin_lock(&imp->imp_lock);
976 list_del_init(&req->rq_list);
977 spin_unlock(&imp->imp_lock);
979 req->rq_status = after_reply(req);
980 if (req->rq_resend) {
981 /* Add this req to the delayed list so
982 it can be errored if the import is
983 evicted after recovery. */
984 spin_lock(&imp->imp_lock);
985 list_add_tail(&req->rq_list,
986 &imp->imp_delayed_list);
987 spin_unlock(&imp->imp_lock);
991 /* If there is no bulk associated with this request,
992 * then we're done and should let the interpreter
993 * process the reply. Similarly if the RPC returned
994 * an error, and therefore the bulk will never arrive.
996 if (req->rq_bulk == NULL || req->rq_status != 0) {
997 req->rq_phase = RQ_PHASE_INTERPRET;
998 GOTO(interpret, req->rq_status);
1001 req->rq_phase = RQ_PHASE_BULK;
1004 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1005 if (ptlrpc_bulk_active(req->rq_bulk))
1008 if (!req->rq_bulk->bd_success) {
1009 /* The RPC reply arrived OK, but the bulk screwed
1010 * up! Dead wierd since the server told us the RPC
1011 * was good after getting the REPLY for her GET or
1012 * the ACK for her PUT. */
1013 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1017 req->rq_phase = RQ_PHASE_INTERPRET;
1020 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1021 LASSERT(!req->rq_receiving_reply);
1023 ptlrpc_unregister_reply(req);
1024 if (req->rq_bulk != NULL)
1025 ptlrpc_unregister_bulk (req);
1027 req->rq_phase = RQ_PHASE_COMPLETE;
1029 if (req->rq_interpret_reply != NULL) {
1030 int (*interpreter)(struct ptlrpc_request *,void *,int) =
1031 req->rq_interpret_reply;
1032 req->rq_status = interpreter(req, &req->rq_async_args,
1036 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1037 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1038 imp->imp_obd->obd_uuid.uuid,
1039 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1040 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1041 lustre_msg_get_opc(req->rq_reqmsg));
1043 set->set_remaining--;
1045 atomic_dec(&imp->imp_inflight);
1046 cfs_waitq_signal(&imp->imp_recovery_waitq);
1049 /* If we hit an error, we want to recover promptly. */
1050 RETURN(set->set_remaining == 0 || force_timer_recalc);
1053 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1055 struct obd_import *imp = req->rq_import;
1059 DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
1060 req->rq_net_err ? "network error" : "timeout",
1061 (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
1063 if (imp != NULL && obd_debug_peer_on_timeout)
1064 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1066 spin_lock(&req->rq_lock);
1067 req->rq_timedout = 1;
1068 req->rq_wait_ctx = 0;
1069 spin_unlock(&req->rq_lock);
1071 ptlrpc_unregister_reply (req);
1073 if (obd_dump_on_timeout)
1074 libcfs_debug_dumplog();
1076 if (req->rq_bulk != NULL)
1077 ptlrpc_unregister_bulk (req);
1080 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1084 /* The DLM server doesn't want recovery run on its imports. */
1085 if (imp->imp_dlm_fake)
1088 /* If this request is for recovery or other primordial tasks,
1089 * then error it out here. */
1090 if (req->rq_ctx_init || req->rq_ctx_fini ||
1091 req->rq_send_state != LUSTRE_IMP_FULL ||
1092 imp->imp_obd->obd_no_recov) {
1093 spin_lock(&req->rq_lock);
1094 req->rq_status = -ETIMEDOUT;
1096 spin_unlock(&req->rq_lock);
1100 /* if request can't be resend we can't wait answer after timeout */
1101 if (req->rq_no_resend) {
1102 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1106 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1111 int ptlrpc_expired_set(void *data)
1113 struct ptlrpc_request_set *set = data;
1114 struct list_head *tmp;
1115 time_t now = CURRENT_SECONDS;
1118 LASSERT(set != NULL);
1120 /* A timeout expired; see which reqs it applies to... */
1121 list_for_each (tmp, &set->set_requests) {
1122 struct ptlrpc_request *req =
1123 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1125 /* request in-flight? */
1126 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1128 (req->rq_phase == RQ_PHASE_BULK)))
1131 if (req->rq_timedout || /* already dealt with */
1132 req->rq_sent + req->rq_timeout > now) /* not expired */
1135 /* deal with this guy */
1136 ptlrpc_expire_one_request (req);
1139 /* When waiting for a whole set, we always to break out of the
1140 * sleep so we can recalculate the timeout, or enable interrupts
1141 * iff everyone's timed out.
1146 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1148 spin_lock(&req->rq_lock);
1150 spin_unlock(&req->rq_lock);
1153 void ptlrpc_interrupted_set(void *data)
1155 struct ptlrpc_request_set *set = data;
1156 struct list_head *tmp;
1158 LASSERT(set != NULL);
1159 CERROR("INTERRUPTED SET %p\n", set);
1161 list_for_each(tmp, &set->set_requests) {
1162 struct ptlrpc_request *req =
1163 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1165 if (req->rq_phase != RQ_PHASE_RPC)
1168 ptlrpc_mark_interrupted(req);
1172 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1174 struct list_head *tmp;
1175 time_t now = CURRENT_SECONDS;
1178 struct ptlrpc_request *req;
1181 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1183 list_for_each(tmp, &set->set_requests) {
1184 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1186 /* request in-flight? */
1187 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1188 (req->rq_phase == RQ_PHASE_BULK) ||
1189 (req->rq_phase == RQ_PHASE_NEW)))
1192 if (req->rq_timedout) /* already timed out */
1195 if (req->rq_phase == RQ_PHASE_NEW)
1196 deadline = req->rq_sent;
1198 deadline = req->rq_sent + req->rq_timeout;
1200 if (deadline <= now) /* actually expired already */
1201 timeout = 1; /* ASAP */
1202 else if (timeout == 0 || timeout > deadline - now)
1203 timeout = deadline - now;
1208 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1210 struct list_head *tmp;
1211 struct ptlrpc_request *req;
1212 struct l_wait_info lwi;
1216 if (list_empty(&set->set_requests))
1219 list_for_each(tmp, &set->set_requests) {
1220 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1221 if (req->rq_phase == RQ_PHASE_NEW)
1222 (void)ptlrpc_send_new_req(req);
1226 timeout = ptlrpc_set_next_timeout(set);
1228 /* wait until all complete, interrupted, or an in-flight
1230 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1232 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1234 ptlrpc_interrupted_set, set);
1235 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1237 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1239 /* -EINTR => all requests have been flagged rq_intr so next
1241 * -ETIMEOUTD => someone timed out. When all reqs have
1242 * timed out, signals are enabled allowing completion with
1244 * I don't really care if we go once more round the loop in
1245 * the error cases -eeb. */
1246 } while (rc != 0 || set->set_remaining != 0);
1248 LASSERT(set->set_remaining == 0);
1251 list_for_each(tmp, &set->set_requests) {
1252 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1254 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1255 if (req->rq_status != 0)
1256 rc = req->rq_status;
1259 if (set->set_interpret != NULL) {
1260 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1262 rc = interpreter (set, set->set_arg, rc);
1268 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1271 if (request == NULL) {
1276 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1277 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1278 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1279 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1280 LASSERT(request->rq_cli_ctx);
1282 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1283 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1284 if (request->rq_import != NULL) {
1286 spin_lock(&request->rq_import->imp_lock);
1287 list_del_init(&request->rq_mod_list);
1288 list_del_init(&request->rq_replay_list);
1290 spin_unlock(&request->rq_import->imp_lock);
1292 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1294 if (atomic_read(&request->rq_refcount) != 0) {
1295 DEBUG_REQ(D_ERROR, request,
1296 "freeing request with nonzero refcount");
1300 if (request->rq_repbuf != NULL)
1301 sptlrpc_cli_free_repbuf(request);
1302 if (request->rq_export != NULL) {
1303 class_export_put(request->rq_export);
1304 request->rq_export = NULL;
1306 if (request->rq_import != NULL) {
1307 class_import_put(request->rq_import);
1308 request->rq_import = NULL;
1310 if (request->rq_bulk != NULL)
1311 ptlrpc_free_bulk(request->rq_bulk);
1313 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1314 sptlrpc_cli_free_reqbuf(request);
1316 sptlrpc_req_put_ctx(request, !locked);
1318 if (request->rq_pool)
1319 __ptlrpc_free_req_to_pool(request);
1321 OBD_FREE(request, sizeof(*request));
1325 void ptlrpc_free_req(struct ptlrpc_request *request)
1327 __ptlrpc_free_req(request, 0);
1330 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1331 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1333 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1334 (void)__ptlrpc_req_finished(request, 1);
1337 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1340 if (request == NULL)
1343 if (request == LP_POISON ||
1344 request->rq_reqmsg == LP_POISON) {
1345 CERROR("dereferencing freed request (bug 575)\n");
1350 DEBUG_REQ(D_INFO, request, "refcount now %u",
1351 atomic_read(&request->rq_refcount) - 1);
1353 if (atomic_dec_and_test(&request->rq_refcount)) {
1354 __ptlrpc_free_req(request, locked);
1361 void ptlrpc_req_finished(struct ptlrpc_request *request)
1363 __ptlrpc_req_finished(request, 0);
1366 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1368 return request->rq_xid;
1370 EXPORT_SYMBOL(ptlrpc_req_xid);
1372 /* Disengage the client's reply buffer from the network
1373 * NB does _NOT_ unregister any client-side bulk.
1374 * IDEMPOTENT, but _not_ safe against concurrent callers.
1375 * The request owner (i.e. the thread doing the I/O) must call...
1377 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1381 struct l_wait_info lwi;
1383 LASSERT(!in_interrupt ()); /* might sleep */
1385 if (!ptlrpc_client_receiving_reply(request))
1388 LNetMDUnlink (request->rq_reply_md_h);
1390 /* We have to l_wait_event() whatever the result, to give liblustre
1391 * a chance to run reply_in_callback() */
1393 if (request->rq_set != NULL)
1394 wq = &request->rq_set->set_waitq;
1396 wq = &request->rq_reply_waitq;
1399 /* Network access will complete in finite time but the HUGE
1400 * timeout lets us CWARN for visibility of sluggish NALs */
1401 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1402 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1406 LASSERT (rc == -ETIMEDOUT);
1407 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1411 /* caller must hold imp->imp_lock */
1412 void ptlrpc_free_committed(struct obd_import *imp)
1414 struct list_head *tmp, *saved;
1415 struct ptlrpc_request *req;
1416 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1419 LASSERT(imp != NULL);
1421 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1424 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1425 imp->imp_generation == imp->imp_last_generation_checked) {
1426 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1427 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1431 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1432 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1433 imp->imp_generation);
1434 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1435 imp->imp_last_generation_checked = imp->imp_generation;
1437 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1438 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1440 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1441 LASSERT(req != last_req);
1444 if (req->rq_import_generation < imp->imp_generation) {
1445 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1449 if (req->rq_replay) {
1450 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1454 /* not yet committed */
1455 if (req->rq_transno > imp->imp_peer_committed_transno) {
1456 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1460 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1461 imp->imp_peer_committed_transno);
1463 spin_lock(&req->rq_lock);
1465 spin_unlock(&req->rq_lock);
1466 if (req->rq_commit_cb != NULL)
1467 req->rq_commit_cb(req);
1468 list_del_init(&req->rq_replay_list);
1469 __ptlrpc_req_finished(req, 1);
1476 void ptlrpc_cleanup_client(struct obd_import *imp)
1483 void ptlrpc_resend_req(struct ptlrpc_request *req)
1485 DEBUG_REQ(D_HA, req, "going to resend");
1486 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1487 req->rq_status = -EAGAIN;
1489 spin_lock(&req->rq_lock);
1491 req->rq_net_err = 0;
1492 req->rq_timedout = 0;
1494 __u64 old_xid = req->rq_xid;
1496 /* ensure previous bulk fails */
1497 req->rq_xid = ptlrpc_next_xid();
1498 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1499 old_xid, req->rq_xid);
1501 ptlrpc_wake_client_req(req);
1502 spin_unlock(&req->rq_lock);
1505 /* XXX: this function and rq_status are currently unused */
1506 void ptlrpc_restart_req(struct ptlrpc_request *req)
1508 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1509 req->rq_status = -ERESTARTSYS;
1511 spin_lock(&req->rq_lock);
1512 req->rq_restart = 1;
1513 req->rq_timedout = 0;
1514 ptlrpc_wake_client_req(req);
1515 spin_unlock(&req->rq_lock);
1518 static int expired_request(void *data)
1520 struct ptlrpc_request *req = data;
1523 /* some failure can suspend regular timeouts */
1524 if (ptlrpc_check_suspend())
1527 RETURN(ptlrpc_expire_one_request(req));
1530 static void interrupted_request(void *data)
1532 struct ptlrpc_request *req = data;
1533 DEBUG_REQ(D_HA, req, "request interrupted");
1534 spin_lock(&req->rq_lock);
1536 spin_unlock(&req->rq_lock);
1539 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1542 atomic_inc(&req->rq_refcount);
1546 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1547 struct obd_import *imp)
1549 struct list_head *tmp;
1551 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1553 /* clear this for new requests that were resent as well
1554 as resent replayed requests. */
1555 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1557 /* don't re-add requests that have been replayed */
1558 if (!list_empty(&req->rq_replay_list))
1561 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1563 LASSERT(imp->imp_replayable);
1564 /* Balanced in ptlrpc_free_committed, usually. */
1565 ptlrpc_request_addref(req);
1566 list_for_each_prev(tmp, &imp->imp_replay_list) {
1567 struct ptlrpc_request *iter =
1568 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1570 /* We may have duplicate transnos if we create and then
1571 * open a file, or for closes retained if to match creating
1572 * opens, so use req->rq_xid as a secondary key.
1573 * (See bugs 684, 685, and 428.)
1574 * XXX no longer needed, but all opens need transnos!
1576 if (iter->rq_transno > req->rq_transno)
1579 if (iter->rq_transno == req->rq_transno) {
1580 LASSERT(iter->rq_xid != req->rq_xid);
1581 if (iter->rq_xid > req->rq_xid)
1585 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1589 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1592 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1596 struct l_wait_info lwi;
1597 struct obd_import *imp = req->rq_import;
1598 cfs_duration_t timeout = 0;
1601 LASSERT(req->rq_set == NULL);
1602 LASSERT(!req->rq_receiving_reply);
1603 atomic_inc(&imp->imp_inflight);
1605 /* for distributed debugging */
1606 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1607 LASSERT(imp->imp_obd != NULL);
1608 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1609 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1610 imp->imp_obd->obd_uuid.uuid,
1611 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1612 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1613 lustre_msg_get_opc(req->rq_reqmsg));
1615 /* Mark phase here for a little debug help */
1616 req->rq_phase = RQ_PHASE_RPC;
1618 spin_lock(&imp->imp_lock);
1619 req->rq_import_generation = imp->imp_generation;
1621 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1622 list_del(&req->rq_list);
1624 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1625 spin_unlock(&imp->imp_lock);
1627 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1629 ptlrpc_import_state_name(req->rq_send_state),
1630 ptlrpc_import_state_name(imp->imp_state));
1631 lwi = LWI_INTR(interrupted_request, req);
1632 rc = l_wait_event(req->rq_reply_waitq,
1633 (req->rq_send_state == imp->imp_state ||
1634 req->rq_err || req->rq_intr),
1636 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1638 ptlrpc_import_state_name(imp->imp_state),
1639 ptlrpc_import_state_name(req->rq_send_state),
1640 req->rq_err, req->rq_intr);
1642 spin_lock(&imp->imp_lock);
1643 list_del_init(&req->rq_list);
1648 else if (req->rq_intr) {
1651 else if (req->rq_no_resend) {
1652 spin_unlock(&imp->imp_lock);
1653 GOTO(out, rc = -ETIMEDOUT);
1661 list_del_init(&req->rq_list);
1662 spin_unlock(&imp->imp_lock);
1663 req->rq_status = rc; // XXX this ok?
1667 if (req->rq_resend) {
1668 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1670 if (req->rq_bulk != NULL) {
1671 ptlrpc_unregister_bulk (req);
1673 /* bulk requests are supposed to be
1674 * idempotent, so we are free to bump the xid
1675 * here, which we need to do before
1676 * registering the bulk again (bug 6371).
1677 * print the old xid first for sanity.
1679 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1680 req->rq_xid = ptlrpc_next_xid();
1683 DEBUG_REQ(D_HA, req, "resending: ");
1686 /* XXX this is the same as ptlrpc_set_wait */
1687 LASSERT(list_empty(&req->rq_list));
1688 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1689 spin_unlock(&imp->imp_lock);
1691 rc = sptlrpc_req_refresh_ctx(req, 0);
1694 /* we got fatal ctx refresh error, directly jump out
1695 * thus we can pass back the actual error code.
1697 spin_lock(&imp->imp_lock);
1698 list_del_init(&req->rq_list);
1699 spin_unlock(&imp->imp_lock);
1701 CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
1704 /* simulating we got error during send rpc */
1708 rc = ptl_send_rpc(req, 0);
1710 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1713 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1714 DEBUG_REQ(D_NET, req,
1715 "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1718 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1720 rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1721 if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1725 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1726 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1727 imp->imp_obd->obd_uuid.uuid,
1728 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1729 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1730 lustre_msg_get_opc(req->rq_reqmsg));
1732 spin_lock(&imp->imp_lock);
1733 list_del_init(&req->rq_list);
1734 spin_unlock(&imp->imp_lock);
1736 /* If the reply was received normally, this just grabs the spinlock
1737 * (ensuring the reply callback has returned), sees that
1738 * req->rq_receiving_reply is clear and returns. */
1739 ptlrpc_unregister_reply (req);
1742 GOTO(out, rc = -EIO);
1744 /* Resend if we need to, unless we were interrupted. */
1745 if (req->rq_resend && !req->rq_intr) {
1746 /* ...unless we were specifically told otherwise. */
1747 if (req->rq_no_resend)
1748 GOTO(out, rc = -ETIMEDOUT);
1749 spin_lock(&imp->imp_lock);
1754 /* Should only be interrupted if we timed out. */
1755 if (!req->rq_timedout)
1756 DEBUG_REQ(D_ERROR, req,
1757 "rq_intr set but rq_timedout not");
1758 GOTO(out, rc = -EINTR);
1761 if (req->rq_timedout) { /* non-recoverable timeout */
1762 GOTO(out, rc = -ETIMEDOUT);
1765 if (!req->rq_replied) {
1766 /* How can this be? -eeb */
1767 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1769 GOTO(out, rc = req->rq_status);
1772 rc = after_reply(req);
1773 /* NB may return +ve success rc */
1774 if (req->rq_resend) {
1775 spin_lock(&imp->imp_lock);
1780 if (req->rq_bulk != NULL) {
1782 /* success so far. Note that anything going wrong
1783 * with bulk now, is EXTREMELY strange, since the
1784 * server must have believed that the bulk
1785 * tranferred OK before she replied with success to
1787 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1788 brc = l_wait_event(req->rq_reply_waitq,
1789 !ptlrpc_bulk_active(req->rq_bulk),
1791 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1793 LASSERT(brc == -ETIMEDOUT);
1794 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1796 } else if (!req->rq_bulk->bd_success) {
1797 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1802 ptlrpc_unregister_bulk (req);
1805 LASSERT(!req->rq_receiving_reply);
1806 req->rq_phase = RQ_PHASE_INTERPRET;
1808 atomic_dec(&imp->imp_inflight);
1809 cfs_waitq_signal(&imp->imp_recovery_waitq);
1813 struct ptlrpc_replay_async_args {
1815 int praa_old_status;
1818 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1819 void * data, int rc)
1821 struct ptlrpc_replay_async_args *aa = data;
1822 struct obd_import *imp = req->rq_import;
1825 atomic_dec(&imp->imp_replay_inflight);
1827 if (!req->rq_replied) {
1828 CERROR("request replay timed out, restarting recovery\n");
1829 GOTO(out, rc = -ETIMEDOUT);
1832 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1833 lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN)
1834 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1836 /* The transno had better not change over replay. */
1837 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1838 lustre_msg_get_transno(req->rq_repmsg));
1840 DEBUG_REQ(D_HA, req, "got rep");
1842 /* let the callback do fixups, possibly including in the request */
1843 if (req->rq_replay_cb)
1844 req->rq_replay_cb(req);
1846 if (req->rq_replied &&
1847 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1848 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1849 lustre_msg_get_status(req->rq_repmsg),
1850 aa->praa_old_status);
1852 /* Put it back for re-replay. */
1853 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1857 * Errors while replay can set transno to 0, but
1858 * imp_last_replay_transno shouldn't be set to 0 anyway
1860 if (req->rq_transno > 0) {
1861 spin_lock(&imp->imp_lock);
1862 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
1863 imp->imp_last_replay_transno = req->rq_transno;
1864 spin_unlock(&imp->imp_lock);
1866 CERROR("Transno is 0 during replay!\n");
1867 /* continue with recovery */
1868 rc = ptlrpc_import_recovery_state_machine(imp);
1870 req->rq_send_state = aa->praa_old_state;
1873 /* this replay failed, so restart recovery */
1874 ptlrpc_connect_import(imp, NULL);
1879 int ptlrpc_replay_req(struct ptlrpc_request *req)
1881 struct ptlrpc_replay_async_args *aa;
1884 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1886 /* Not handling automatic bulk replay yet (or ever?) */
1887 LASSERT(req->rq_bulk == NULL);
1889 DEBUG_REQ(D_HA, req, "REPLAY");
1891 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1892 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1893 memset(aa, 0, sizeof *aa);
1895 /* Prepare request to be resent with ptlrpcd */
1896 aa->praa_old_state = req->rq_send_state;
1897 req->rq_send_state = LUSTRE_IMP_REPLAY;
1898 req->rq_phase = RQ_PHASE_NEW;
1899 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
1902 req->rq_interpret_reply = ptlrpc_replay_interpret;
1903 atomic_inc(&req->rq_import->imp_replay_inflight);
1904 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1906 ptlrpcd_add_req(req);
1910 void ptlrpc_abort_inflight(struct obd_import *imp)
1912 struct list_head *tmp, *n;
1915 /* Make sure that no new requests get processed for this import.
1916 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1917 * this flag and then putting requests on sending_list or delayed_list.
1919 spin_lock(&imp->imp_lock);
1921 /* XXX locking? Maybe we should remove each request with the list
1922 * locked? Also, how do we know if the requests on the list are
1923 * being freed at this time?
1925 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1926 struct ptlrpc_request *req =
1927 list_entry(tmp, struct ptlrpc_request, rq_list);
1929 DEBUG_REQ(D_RPCTRACE, req, "inflight");
1931 spin_lock (&req->rq_lock);
1932 if (req->rq_import_generation < imp->imp_generation) {
1934 ptlrpc_wake_client_req(req);
1936 spin_unlock (&req->rq_lock);
1939 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1940 struct ptlrpc_request *req =
1941 list_entry(tmp, struct ptlrpc_request, rq_list);
1943 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
1945 spin_lock (&req->rq_lock);
1946 if (req->rq_import_generation < imp->imp_generation) {
1948 ptlrpc_wake_client_req(req);
1950 spin_unlock (&req->rq_lock);
1953 /* Last chance to free reqs left on the replay list, but we
1954 * will still leak reqs that haven't committed. */
1955 if (imp->imp_replayable)
1956 ptlrpc_free_committed(imp);
1958 spin_unlock(&imp->imp_lock);
1963 static __u64 ptlrpc_last_xid = 0;
1964 spinlock_t ptlrpc_last_xid_lock;
1966 __u64 ptlrpc_next_xid(void)
1969 spin_lock(&ptlrpc_last_xid_lock);
1970 tmp = ++ptlrpc_last_xid;
1971 spin_unlock(&ptlrpc_last_xid_lock);
1975 __u64 ptlrpc_sample_next_xid(void)
1978 spin_lock(&ptlrpc_last_xid_lock);
1979 tmp = ptlrpc_last_xid + 1;
1980 spin_unlock(&ptlrpc_last_xid_lock);
1983 EXPORT_SYMBOL(ptlrpc_sample_next_xid);