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_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_ctx_chain);
397 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
398 cfs_waitq_init(&request->rq_reply_waitq);
399 request->rq_xid = ptlrpc_next_xid();
400 atomic_set(&request->rq_refcount, 1);
402 lustre_msg_set_opc(request->rq_reqmsg, opcode);
403 lustre_msg_set_flags(request->rq_reqmsg, 0);
407 sptlrpc_req_put_ctx(request);
409 class_import_put(imp);
410 if (request->rq_pool)
411 __ptlrpc_free_req_to_pool(request);
413 OBD_FREE(request, sizeof(*request));
417 struct ptlrpc_request *
418 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
419 int *lengths, char **bufs)
421 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
425 struct ptlrpc_request_set *ptlrpc_prep_set(void)
427 struct ptlrpc_request_set *set;
430 OBD_ALLOC(set, sizeof *set);
433 CFS_INIT_LIST_HEAD(&set->set_requests);
434 cfs_waitq_init(&set->set_waitq);
435 set->set_remaining = 0;
436 spin_lock_init(&set->set_new_req_lock);
437 CFS_INIT_LIST_HEAD(&set->set_new_requests);
442 /* Finish with this set; opposite of prep_set. */
443 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
445 struct list_head *tmp;
446 struct list_head *next;
451 /* Requests on the set should either all be completed, or all be new */
452 expected_phase = (set->set_remaining == 0) ?
453 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
454 list_for_each (tmp, &set->set_requests) {
455 struct ptlrpc_request *req =
456 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
458 LASSERT(req->rq_phase == expected_phase);
462 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
464 list_for_each_safe(tmp, next, &set->set_requests) {
465 struct ptlrpc_request *req =
466 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
467 list_del_init(&req->rq_set_chain);
469 LASSERT(req->rq_phase == expected_phase);
471 if (req->rq_phase == RQ_PHASE_NEW) {
473 if (req->rq_interpret_reply != NULL) {
474 int (*interpreter)(struct ptlrpc_request *,
476 req->rq_interpret_reply;
478 /* higher level (i.e. LOV) failed;
479 * let the sub reqs clean up */
480 req->rq_status = -EBADR;
481 interpreter(req, &req->rq_async_args,
484 set->set_remaining--;
488 ptlrpc_req_finished (req);
491 LASSERT(set->set_remaining == 0);
493 OBD_FREE(set, sizeof(*set));
497 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
498 struct ptlrpc_request *req)
500 /* The set takes over the caller's request reference */
501 list_add_tail(&req->rq_set_chain, &set->set_requests);
503 set->set_remaining++;
505 atomic_inc(&req->rq_import->imp_inflight);
508 /* lock so many callers can add things, the context that owns the set
509 * is supposed to notice these and move them into the set proper. */
510 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
511 struct ptlrpc_request *req)
513 spin_lock(&set->set_new_req_lock);
514 /* The set takes over the caller's request reference */
515 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
517 spin_unlock(&set->set_new_req_lock);
521 * Based on the current state of the import, determine if the request
522 * can be sent, is an error, or should be delayed.
524 * Returns true if this request should be delayed. If false, and
525 * *status is set, then the request can not be sent and *status is the
526 * error code. If false and status is 0, then request can be sent.
528 * The imp->imp_lock must be held.
530 static int ptlrpc_import_delay_req(struct obd_import *imp,
531 struct ptlrpc_request *req, int *status)
536 LASSERT (status != NULL);
539 if (req->rq_ctx_init || req->rq_ctx_fini) {
540 /* always allow ctx init/fini rpc go through */
541 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
542 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
545 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
546 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
548 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
549 imp->imp_state == LUSTRE_IMP_CONNECTING) {
550 /* allow CONNECT even if import is invalid */ ;
551 } else if (imp->imp_invalid) {
552 /* If the import has been invalidated (such as by an OST
553 * failure), the request must fail with -EIO. */
554 if (!imp->imp_deactive)
555 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
557 } else if (req->rq_import_generation != imp->imp_generation) {
558 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
560 } else if (req->rq_send_state != imp->imp_state) {
561 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
563 *status = -EWOULDBLOCK;
571 static int ptlrpc_check_reply(struct ptlrpc_request *req)
576 /* serialise with network callback */
577 spin_lock(&req->rq_lock);
582 if (req->rq_net_err && !req->rq_timedout) {
583 spin_unlock(&req->rq_lock);
584 rc = ptlrpc_expire_one_request(req);
585 spin_lock(&req->rq_lock);
599 spin_unlock(&req->rq_lock);
600 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
604 static int ptlrpc_check_status(struct ptlrpc_request *req)
609 err = lustre_msg_get_status(req->rq_repmsg);
610 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
611 struct obd_import *imp = req->rq_import;
612 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
613 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
614 " with %s. The %s operation failed with %d\n",
615 libcfs_nid2str(imp->imp_connection->c_peer.nid),
616 ll_opcode2str(opc), err);
617 RETURN(err < 0 ? err : -EINVAL);
621 DEBUG_REQ(D_INFO, req, "status is %d", err);
622 } else if (err > 0) {
623 /* XXX: translate this error from net to host */
624 DEBUG_REQ(D_INFO, req, "status is %d", err);
630 static int after_reply(struct ptlrpc_request *req)
632 struct obd_import *imp = req->rq_import;
636 LASSERT(!req->rq_receiving_reply);
637 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
639 /* NB Until this point, the whole of the incoming message,
640 * 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;
678 /* Either we've been evicted, or the server has failed for
679 * some reason. Try to reconnect, and if that fails, punt to the
681 if ((rc == -ENOTCONN) || (rc == -ENODEV)) {
682 if (req->rq_send_state != LUSTRE_IMP_FULL ||
683 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
687 ptlrpc_request_handle_notconn(req);
692 /* Store transno in reqmsg for replay. */
693 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
694 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
696 if (req->rq_import->imp_replayable) {
697 spin_lock(&imp->imp_lock);
698 /* no point in adding already-committed requests to the replay
699 * list, we will just remove them immediately. b=9829 */
700 if (req->rq_transno != 0 &&
702 lustre_msg_get_last_committed(req->rq_repmsg) ||
704 ptlrpc_retain_replayable_request(req, imp);
705 else if (req->rq_commit_cb != NULL) {
706 spin_unlock(&imp->imp_lock);
707 req->rq_commit_cb(req);
708 spin_lock(&imp->imp_lock);
711 /* Replay-enabled imports return commit-status information. */
712 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
713 imp->imp_peer_committed_transno =
714 lustre_msg_get_last_committed(req->rq_repmsg);
716 ptlrpc_free_committed(imp);
717 spin_unlock(&imp->imp_lock);
723 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
725 struct obd_import *imp;
729 LASSERT(req->rq_phase == RQ_PHASE_NEW);
730 req->rq_phase = RQ_PHASE_RPC;
732 imp = req->rq_import;
733 spin_lock(&imp->imp_lock);
735 req->rq_import_generation = imp->imp_generation;
737 if (ptlrpc_import_delay_req(imp, req, &rc)) {
738 spin_lock (&req->rq_lock);
740 spin_unlock (&req->rq_lock);
742 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
744 lustre_msg_get_status(req->rq_reqmsg) ,
745 ptlrpc_import_state_name(req->rq_send_state),
746 ptlrpc_import_state_name(imp->imp_state));
747 LASSERT(list_empty (&req->rq_list));
749 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
750 spin_unlock(&imp->imp_lock);
755 spin_unlock(&imp->imp_lock);
757 req->rq_phase = RQ_PHASE_INTERPRET;
761 /* XXX this is the same as ptlrpc_queue_wait */
762 LASSERT(list_empty(&req->rq_list));
763 list_add_tail(&req->rq_list, &imp->imp_sending_list);
764 spin_unlock(&imp->imp_lock);
766 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
768 rc = sptlrpc_req_refresh_ctx(req, -1);
774 /* here begins timeout counting */
775 req->rq_sent = CURRENT_SECONDS;
776 req->rq_wait_ctx = 1;
781 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
782 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
783 imp->imp_obd->obd_uuid.uuid,
784 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
785 libcfs_nid2str(imp->imp_connection->c_peer.nid),
786 lustre_msg_get_opc(req->rq_reqmsg));
788 rc = ptl_send_rpc(req, 0);
790 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
797 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
798 int ptlrpc_check_set(struct ptlrpc_request_set *set)
800 struct list_head *tmp;
801 int force_timer_recalc = 0;
804 if (set->set_remaining == 0)
807 list_for_each(tmp, &set->set_requests) {
808 struct ptlrpc_request *req =
809 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
810 struct obd_import *imp = req->rq_import;
813 if (req->rq_phase == RQ_PHASE_NEW &&
814 ptlrpc_send_new_req(req)) {
815 force_timer_recalc = 1;
818 if (!(req->rq_phase == RQ_PHASE_RPC ||
819 req->rq_phase == RQ_PHASE_BULK ||
820 req->rq_phase == RQ_PHASE_INTERPRET ||
821 req->rq_phase == RQ_PHASE_COMPLETE)) {
822 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
826 if (req->rq_phase == RQ_PHASE_COMPLETE)
829 if (req->rq_phase == RQ_PHASE_INTERPRET)
830 GOTO(interpret, req->rq_status);
832 if (req->rq_net_err && !req->rq_timedout)
833 ptlrpc_expire_one_request(req);
836 ptlrpc_unregister_reply(req);
838 if (req->rq_status == 0)
839 req->rq_status = -EIO;
840 req->rq_phase = RQ_PHASE_INTERPRET;
842 spin_lock(&imp->imp_lock);
843 list_del_init(&req->rq_list);
844 spin_unlock(&imp->imp_lock);
846 GOTO(interpret, req->rq_status);
849 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
850 * will only be set after rq_timedout, but the oig waiting
851 * path sets rq_intr irrespective of whether ptlrpcd has
852 * seen a timeout. our policy is to only interpret
853 * interrupted rpcs after they have timed out */
854 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
856 /* NB could be on delayed list */
857 ptlrpc_unregister_reply(req);
858 req->rq_status = -EINTR;
859 req->rq_phase = RQ_PHASE_INTERPRET;
861 spin_lock(&imp->imp_lock);
862 list_del_init(&req->rq_list);
863 spin_unlock(&imp->imp_lock);
865 GOTO(interpret, req->rq_status);
868 if (req->rq_phase == RQ_PHASE_RPC) {
869 if (req->rq_timedout || req->rq_resend ||
870 req->rq_waiting || req->rq_wait_ctx) {
873 /* rq_wait_ctx is only touched in ptlrpcd,
874 * no lock needed here.
876 if (req->rq_wait_ctx)
879 ptlrpc_unregister_reply(req);
881 spin_lock(&imp->imp_lock);
883 if (ptlrpc_import_delay_req(imp, req, &status)){
884 spin_unlock(&imp->imp_lock);
888 list_del_init(&req->rq_list);
890 req->rq_status = status;
891 req->rq_phase = RQ_PHASE_INTERPRET;
892 spin_unlock(&imp->imp_lock);
893 GOTO(interpret, req->rq_status);
895 if (req->rq_no_resend && !req->rq_wait_ctx) {
896 req->rq_status = -ENOTCONN;
897 req->rq_phase = RQ_PHASE_INTERPRET;
898 spin_unlock(&imp->imp_lock);
899 GOTO(interpret, req->rq_status);
901 list_add_tail(&req->rq_list,
902 &imp->imp_sending_list);
904 spin_unlock(&imp->imp_lock);
907 if (req->rq_resend) {
908 lustre_msg_add_flags(req->rq_reqmsg,
911 __u64 old_xid = req->rq_xid;
913 ptlrpc_unregister_bulk (req);
915 /* ensure previous bulk fails */
916 req->rq_xid = ptlrpc_next_xid();
917 CDEBUG(D_HA, "resend bulk "
920 old_xid, req->rq_xid);
924 status = sptlrpc_req_refresh_ctx(req, -1);
927 req->rq_status = status;
928 force_timer_recalc = 1;
930 if (!req->rq_wait_ctx) {
931 /* begins timeout counting */
932 req->rq_sent = CURRENT_SECONDS;
933 req->rq_wait_ctx = 1;
938 req->rq_wait_ctx = 0;
941 rc = ptl_send_rpc(req, 0);
943 DEBUG_REQ(D_HA, req, "send failed (%d)",
945 force_timer_recalc = 1;
948 /* need to reset the timeout */
949 force_timer_recalc = 1;
952 /* Still waiting for a reply? */
953 if (ptlrpc_client_receiving_reply(req))
956 /* Did we actually receive a reply? */
957 if (!ptlrpc_client_replied(req))
960 spin_lock(&imp->imp_lock);
961 list_del_init(&req->rq_list);
962 spin_unlock(&imp->imp_lock);
964 req->rq_status = after_reply(req);
965 if (req->rq_resend) {
966 /* Add this req to the delayed list so
967 it can be errored if the import is
968 evicted after recovery. */
969 spin_lock(&imp->imp_lock);
970 list_add_tail(&req->rq_list,
971 &imp->imp_delayed_list);
972 spin_unlock(&imp->imp_lock);
976 /* If there is no bulk associated with this request,
977 * then we're done and should let the interpreter
978 * process the reply. Similarly if the RPC returned
979 * an error, and therefore the bulk will never arrive.
981 if (req->rq_bulk == NULL || req->rq_status != 0) {
982 req->rq_phase = RQ_PHASE_INTERPRET;
983 GOTO(interpret, req->rq_status);
986 req->rq_phase = RQ_PHASE_BULK;
989 LASSERT(req->rq_phase == RQ_PHASE_BULK);
990 if (ptlrpc_bulk_active(req->rq_bulk))
993 if (!req->rq_bulk->bd_success) {
994 /* The RPC reply arrived OK, but the bulk screwed
995 * up! Dead wierd since the server told us the RPC
996 * was good after getting the REPLY for her GET or
997 * the ACK for her PUT. */
998 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1002 req->rq_phase = RQ_PHASE_INTERPRET;
1005 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1006 LASSERT(!req->rq_receiving_reply);
1008 ptlrpc_unregister_reply(req);
1009 if (req->rq_bulk != NULL)
1010 ptlrpc_unregister_bulk (req);
1012 req->rq_phase = RQ_PHASE_COMPLETE;
1014 if (req->rq_interpret_reply != NULL) {
1015 int (*interpreter)(struct ptlrpc_request *,void *,int) =
1016 req->rq_interpret_reply;
1017 req->rq_status = interpreter(req, &req->rq_async_args,
1021 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1022 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1023 imp->imp_obd->obd_uuid.uuid,
1024 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1025 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1026 lustre_msg_get_opc(req->rq_reqmsg));
1028 set->set_remaining--;
1030 atomic_dec(&imp->imp_inflight);
1031 cfs_waitq_signal(&imp->imp_recovery_waitq);
1034 /* If we hit an error, we want to recover promptly. */
1035 RETURN(set->set_remaining == 0 || force_timer_recalc);
1038 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1040 struct obd_import *imp = req->rq_import;
1044 DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
1045 req->rq_net_err ? "network error" : "timeout",
1046 (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
1048 if (imp != NULL && obd_debug_peer_on_timeout)
1049 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1051 spin_lock(&req->rq_lock);
1052 req->rq_timedout = 1;
1053 req->rq_wait_ctx = 0;
1054 spin_unlock(&req->rq_lock);
1056 ptlrpc_unregister_reply (req);
1058 if (obd_dump_on_timeout)
1059 libcfs_debug_dumplog();
1061 if (req->rq_bulk != NULL)
1062 ptlrpc_unregister_bulk (req);
1065 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1069 /* The DLM server doesn't want recovery run on its imports. */
1070 if (imp->imp_dlm_fake)
1073 /* If this request is for recovery or other primordial tasks,
1074 * then error it out here. */
1075 if (req->rq_ctx_init || req->rq_ctx_fini ||
1076 req->rq_send_state != LUSTRE_IMP_FULL ||
1077 imp->imp_obd->obd_no_recov) {
1078 spin_lock(&req->rq_lock);
1079 req->rq_status = -ETIMEDOUT;
1081 spin_unlock(&req->rq_lock);
1085 /* if request can't be resend we can't wait answer after timeout */
1086 if (req->rq_no_resend) {
1087 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1091 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1096 int ptlrpc_expired_set(void *data)
1098 struct ptlrpc_request_set *set = data;
1099 struct list_head *tmp;
1100 time_t now = CURRENT_SECONDS;
1103 LASSERT(set != NULL);
1105 /* A timeout expired; see which reqs it applies to... */
1106 list_for_each (tmp, &set->set_requests) {
1107 struct ptlrpc_request *req =
1108 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1110 /* request in-flight? */
1111 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1113 (req->rq_phase == RQ_PHASE_BULK)))
1116 if (req->rq_timedout || /* already dealt with */
1117 req->rq_sent + req->rq_timeout > now) /* not expired */
1120 /* deal with this guy */
1121 ptlrpc_expire_one_request (req);
1124 /* When waiting for a whole set, we always to break out of the
1125 * sleep so we can recalculate the timeout, or enable interrupts
1126 * iff everyone's timed out.
1131 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1133 spin_lock(&req->rq_lock);
1135 spin_unlock(&req->rq_lock);
1138 void ptlrpc_interrupted_set(void *data)
1140 struct ptlrpc_request_set *set = data;
1141 struct list_head *tmp;
1143 LASSERT(set != NULL);
1144 CERROR("INTERRUPTED SET %p\n", set);
1146 list_for_each(tmp, &set->set_requests) {
1147 struct ptlrpc_request *req =
1148 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1150 if (req->rq_phase != RQ_PHASE_RPC)
1153 ptlrpc_mark_interrupted(req);
1157 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1159 struct list_head *tmp;
1160 time_t now = CURRENT_SECONDS;
1163 struct ptlrpc_request *req;
1166 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1168 list_for_each(tmp, &set->set_requests) {
1169 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1171 /* request in-flight? */
1172 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1173 (req->rq_phase == RQ_PHASE_BULK)))
1176 if (req->rq_timedout) /* already timed out */
1179 deadline = req->rq_sent + req->rq_timeout;
1180 if (deadline <= now) /* actually expired already */
1181 timeout = 1; /* ASAP */
1182 else if (timeout == 0 || timeout > deadline - now)
1183 timeout = deadline - now;
1188 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1190 struct list_head *tmp;
1191 struct ptlrpc_request *req;
1192 struct l_wait_info lwi;
1196 if (list_empty(&set->set_requests))
1199 list_for_each(tmp, &set->set_requests) {
1200 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1201 if (req->rq_phase == RQ_PHASE_NEW)
1202 (void)ptlrpc_send_new_req(req);
1206 timeout = ptlrpc_set_next_timeout(set);
1208 /* wait until all complete, interrupted, or an in-flight
1210 CDEBUG(D_HA, "set %p going to sleep for %d seconds\n",
1212 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1214 ptlrpc_interrupted_set, set);
1215 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1217 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1219 /* -EINTR => all requests have been flagged rq_intr so next
1221 * -ETIMEOUTD => someone timed out. When all reqs have
1222 * timed out, signals are enabled allowing completion with
1224 * I don't really care if we go once more round the loop in
1225 * the error cases -eeb. */
1226 } while (rc != 0 || set->set_remaining != 0);
1228 LASSERT(set->set_remaining == 0);
1231 list_for_each(tmp, &set->set_requests) {
1232 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1234 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1235 if (req->rq_status != 0)
1236 rc = req->rq_status;
1239 if (set->set_interpret != NULL) {
1240 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1242 rc = interpreter (set, set->set_arg, rc);
1248 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1251 if (request == NULL) {
1256 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1257 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1258 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1259 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1260 LASSERT(request->rq_cli_ctx);
1262 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1263 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1264 if (request->rq_import != NULL) {
1266 spin_lock(&request->rq_import->imp_lock);
1267 list_del_init(&request->rq_replay_list);
1269 spin_unlock(&request->rq_import->imp_lock);
1271 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1273 if (atomic_read(&request->rq_refcount) != 0) {
1274 DEBUG_REQ(D_ERROR, request,
1275 "freeing request with nonzero refcount");
1279 if (request->rq_repbuf != NULL)
1280 sptlrpc_cli_free_repbuf(request);
1281 if (request->rq_export != NULL) {
1282 class_export_put(request->rq_export);
1283 request->rq_export = NULL;
1285 if (request->rq_import != NULL) {
1286 class_import_put(request->rq_import);
1287 request->rq_import = NULL;
1289 if (request->rq_bulk != NULL)
1290 ptlrpc_free_bulk(request->rq_bulk);
1292 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1293 sptlrpc_cli_free_reqbuf(request);
1295 sptlrpc_req_put_ctx(request);
1297 if (request->rq_pool)
1298 __ptlrpc_free_req_to_pool(request);
1300 OBD_FREE(request, sizeof(*request));
1304 void ptlrpc_free_req(struct ptlrpc_request *request)
1306 __ptlrpc_free_req(request, 0);
1309 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1310 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1312 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1313 (void)__ptlrpc_req_finished(request, 1);
1316 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1319 if (request == NULL)
1322 if (request == LP_POISON ||
1323 request->rq_reqmsg == LP_POISON) {
1324 CERROR("dereferencing freed request (bug 575)\n");
1329 DEBUG_REQ(D_INFO, request, "refcount now %u",
1330 atomic_read(&request->rq_refcount) - 1);
1332 if (atomic_dec_and_test(&request->rq_refcount)) {
1333 __ptlrpc_free_req(request, locked);
1340 void ptlrpc_req_finished(struct ptlrpc_request *request)
1342 __ptlrpc_req_finished(request, 0);
1345 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1347 return request->rq_xid;
1349 EXPORT_SYMBOL(ptlrpc_req_xid);
1351 /* Disengage the client's reply buffer from the network
1352 * NB does _NOT_ unregister any client-side bulk.
1353 * IDEMPOTENT, but _not_ safe against concurrent callers.
1354 * The request owner (i.e. the thread doing the I/O) must call...
1356 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1360 struct l_wait_info lwi;
1362 LASSERT(!in_interrupt ()); /* might sleep */
1364 if (!ptlrpc_client_receiving_reply(request))
1367 LNetMDUnlink (request->rq_reply_md_h);
1369 /* We have to l_wait_event() whatever the result, to give liblustre
1370 * a chance to run reply_in_callback() */
1372 if (request->rq_set != NULL)
1373 wq = &request->rq_set->set_waitq;
1375 wq = &request->rq_reply_waitq;
1378 /* Network access will complete in finite time but the HUGE
1379 * timeout lets us CWARN for visibility of sluggish NALs */
1380 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1381 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1385 LASSERT (rc == -ETIMEDOUT);
1386 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1390 /* caller must hold imp->imp_lock */
1391 void ptlrpc_free_committed(struct obd_import *imp)
1393 struct list_head *tmp, *saved;
1394 struct ptlrpc_request *req;
1395 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1398 LASSERT(imp != NULL);
1400 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1403 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1404 imp->imp_generation == imp->imp_last_generation_checked) {
1405 CDEBUG(D_HA, "%s: skip recheck for last_committed "LPU64"\n",
1406 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1410 CDEBUG(D_HA, "%s: committing for last_committed "LPU64" gen %d\n",
1411 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1412 imp->imp_generation);
1413 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1414 imp->imp_last_generation_checked = imp->imp_generation;
1416 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1417 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1419 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1420 LASSERT(req != last_req);
1423 if (req->rq_import_generation < imp->imp_generation) {
1424 DEBUG_REQ(D_HA, req, "freeing request with old gen");
1428 if (req->rq_replay) {
1429 DEBUG_REQ(D_HA, req, "keeping (FL_REPLAY)");
1433 /* not yet committed */
1434 if (req->rq_transno > imp->imp_peer_committed_transno) {
1435 DEBUG_REQ(D_HA, req, "stopping search");
1439 DEBUG_REQ(D_HA, req, "committing (last_committed "LPU64")",
1440 imp->imp_peer_committed_transno);
1442 spin_lock(&req->rq_lock);
1444 spin_unlock(&req->rq_lock);
1445 if (req->rq_commit_cb != NULL)
1446 req->rq_commit_cb(req);
1447 list_del_init(&req->rq_replay_list);
1448 __ptlrpc_req_finished(req, 1);
1455 void ptlrpc_cleanup_client(struct obd_import *imp)
1462 void ptlrpc_resend_req(struct ptlrpc_request *req)
1464 DEBUG_REQ(D_HA, req, "going to resend");
1465 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1466 req->rq_status = -EAGAIN;
1468 spin_lock(&req->rq_lock);
1470 req->rq_net_err = 0;
1471 req->rq_timedout = 0;
1473 __u64 old_xid = req->rq_xid;
1475 /* ensure previous bulk fails */
1476 req->rq_xid = ptlrpc_next_xid();
1477 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1478 old_xid, req->rq_xid);
1480 ptlrpc_wake_client_req(req);
1481 spin_unlock(&req->rq_lock);
1484 /* XXX: this function and rq_status are currently unused */
1485 void ptlrpc_restart_req(struct ptlrpc_request *req)
1487 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1488 req->rq_status = -ERESTARTSYS;
1490 spin_lock(&req->rq_lock);
1491 req->rq_restart = 1;
1492 req->rq_timedout = 0;
1493 ptlrpc_wake_client_req(req);
1494 spin_unlock(&req->rq_lock);
1497 static int expired_request(void *data)
1499 struct ptlrpc_request *req = data;
1502 /* some failure can suspend regular timeouts */
1503 if (ptlrpc_check_suspend())
1506 RETURN(ptlrpc_expire_one_request(req));
1509 static void interrupted_request(void *data)
1511 struct ptlrpc_request *req = data;
1512 DEBUG_REQ(D_HA, req, "request interrupted");
1513 spin_lock(&req->rq_lock);
1515 spin_unlock(&req->rq_lock);
1518 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1521 atomic_inc(&req->rq_refcount);
1525 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1526 struct obd_import *imp)
1528 struct list_head *tmp;
1530 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1532 /* clear this for new requests that were resent as well
1533 as resent replayed requests. */
1534 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1536 /* don't re-add requests that have been replayed */
1537 if (!list_empty(&req->rq_replay_list))
1540 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1542 LASSERT(imp->imp_replayable);
1543 /* Balanced in ptlrpc_free_committed, usually. */
1544 ptlrpc_request_addref(req);
1545 list_for_each_prev(tmp, &imp->imp_replay_list) {
1546 struct ptlrpc_request *iter =
1547 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1549 /* We may have duplicate transnos if we create and then
1550 * open a file, or for closes retained if to match creating
1551 * opens, so use req->rq_xid as a secondary key.
1552 * (See bugs 684, 685, and 428.)
1553 * XXX no longer needed, but all opens need transnos!
1555 if (iter->rq_transno > req->rq_transno)
1558 if (iter->rq_transno == req->rq_transno) {
1559 LASSERT(iter->rq_xid != req->rq_xid);
1560 if (iter->rq_xid > req->rq_xid)
1564 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1568 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1571 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1575 struct l_wait_info lwi;
1576 struct obd_import *imp = req->rq_import;
1577 cfs_duration_t timeout = 0;
1580 LASSERT(req->rq_set == NULL);
1581 LASSERT(!req->rq_receiving_reply);
1582 atomic_inc(&imp->imp_inflight);
1584 /* for distributed debugging */
1585 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1586 LASSERT(imp->imp_obd != NULL);
1587 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1588 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1589 imp->imp_obd->obd_uuid.uuid,
1590 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1591 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1592 lustre_msg_get_opc(req->rq_reqmsg));
1594 /* Mark phase here for a little debug help */
1595 req->rq_phase = RQ_PHASE_RPC;
1597 spin_lock(&imp->imp_lock);
1598 req->rq_import_generation = imp->imp_generation;
1600 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1601 list_del(&req->rq_list);
1603 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1604 spin_unlock(&imp->imp_lock);
1606 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1608 ptlrpc_import_state_name(req->rq_send_state),
1609 ptlrpc_import_state_name(imp->imp_state));
1610 lwi = LWI_INTR(interrupted_request, req);
1611 rc = l_wait_event(req->rq_reply_waitq,
1612 (req->rq_send_state == imp->imp_state ||
1613 req->rq_err || req->rq_intr),
1615 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1617 ptlrpc_import_state_name(imp->imp_state),
1618 ptlrpc_import_state_name(req->rq_send_state),
1619 req->rq_err, req->rq_intr);
1621 spin_lock(&imp->imp_lock);
1622 list_del_init(&req->rq_list);
1627 else if (req->rq_intr) {
1630 else if (req->rq_no_resend) {
1631 spin_unlock(&imp->imp_lock);
1632 GOTO(out, rc = -ETIMEDOUT);
1640 list_del_init(&req->rq_list);
1641 spin_unlock(&imp->imp_lock);
1642 req->rq_status = rc; // XXX this ok?
1646 if (req->rq_resend) {
1647 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1649 if (req->rq_bulk != NULL) {
1650 ptlrpc_unregister_bulk (req);
1652 /* bulk requests are supposed to be
1653 * idempotent, so we are free to bump the xid
1654 * here, which we need to do before
1655 * registering the bulk again (bug 6371).
1656 * print the old xid first for sanity.
1658 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1659 req->rq_xid = ptlrpc_next_xid();
1662 DEBUG_REQ(D_HA, req, "resending: ");
1665 /* XXX this is the same as ptlrpc_set_wait */
1666 LASSERT(list_empty(&req->rq_list));
1667 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1668 spin_unlock(&imp->imp_lock);
1670 rc = sptlrpc_req_refresh_ctx(req, 0);
1673 /* we got fatal ctx refresh error, directly jump out
1674 * thus we can pass back the actual error code.
1676 spin_lock(&imp->imp_lock);
1677 list_del_init(&req->rq_list);
1678 spin_unlock(&imp->imp_lock);
1680 CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
1683 /* simulating we got error during send rpc */
1687 rc = ptl_send_rpc(req, 0);
1689 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1692 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1693 DEBUG_REQ(D_NET, req,
1694 "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1697 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1699 rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1700 if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1704 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1705 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1706 imp->imp_obd->obd_uuid.uuid,
1707 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1708 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1709 lustre_msg_get_opc(req->rq_reqmsg));
1711 spin_lock(&imp->imp_lock);
1712 list_del_init(&req->rq_list);
1713 spin_unlock(&imp->imp_lock);
1715 /* If the reply was received normally, this just grabs the spinlock
1716 * (ensuring the reply callback has returned), sees that
1717 * req->rq_receiving_reply is clear and returns. */
1718 ptlrpc_unregister_reply (req);
1721 GOTO(out, rc = -EIO);
1723 /* Resend if we need to, unless we were interrupted. */
1724 if (req->rq_resend && !req->rq_intr) {
1725 /* ...unless we were specifically told otherwise. */
1726 if (req->rq_no_resend)
1727 GOTO(out, rc = -ETIMEDOUT);
1728 spin_lock(&imp->imp_lock);
1733 /* Should only be interrupted if we timed out. */
1734 if (!req->rq_timedout)
1735 DEBUG_REQ(D_ERROR, req,
1736 "rq_intr set but rq_timedout not");
1737 GOTO(out, rc = -EINTR);
1740 if (req->rq_timedout) { /* non-recoverable timeout */
1741 GOTO(out, rc = -ETIMEDOUT);
1744 if (!req->rq_replied) {
1745 /* How can this be? -eeb */
1746 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1748 GOTO(out, rc = req->rq_status);
1751 rc = after_reply(req);
1752 /* NB may return +ve success rc */
1753 if (req->rq_resend) {
1754 spin_lock(&imp->imp_lock);
1759 if (req->rq_bulk != NULL) {
1761 /* success so far. Note that anything going wrong
1762 * with bulk now, is EXTREMELY strange, since the
1763 * server must have believed that the bulk
1764 * tranferred OK before she replied with success to
1766 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1767 brc = l_wait_event(req->rq_reply_waitq,
1768 !ptlrpc_bulk_active(req->rq_bulk),
1770 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1772 LASSERT(brc == -ETIMEDOUT);
1773 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1775 } else if (!req->rq_bulk->bd_success) {
1776 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1781 ptlrpc_unregister_bulk (req);
1784 LASSERT(!req->rq_receiving_reply);
1785 req->rq_phase = RQ_PHASE_INTERPRET;
1787 atomic_dec(&imp->imp_inflight);
1788 cfs_waitq_signal(&imp->imp_recovery_waitq);
1792 struct ptlrpc_replay_async_args {
1794 int praa_old_status;
1797 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1798 void * data, int rc)
1800 struct ptlrpc_replay_async_args *aa = data;
1801 struct obd_import *imp = req->rq_import;
1804 atomic_dec(&imp->imp_replay_inflight);
1806 if (!req->rq_replied) {
1807 CERROR("request replay timed out, restarting recovery\n");
1808 GOTO(out, rc = -ETIMEDOUT);
1811 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1812 lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN)
1813 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1815 /* The transno had better not change over replay. */
1816 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1817 lustre_msg_get_transno(req->rq_repmsg));
1819 DEBUG_REQ(D_HA, req, "got rep");
1821 /* let the callback do fixups, possibly including in the request */
1822 if (req->rq_replay_cb)
1823 req->rq_replay_cb(req);
1825 if (req->rq_replied &&
1826 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1827 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1828 lustre_msg_get_status(req->rq_repmsg),
1829 aa->praa_old_status);
1831 /* Put it back for re-replay. */
1832 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1836 * Errors while replay can set transno to 0, but
1837 * imp_last_replay_transno shouldn't be set to 0 anyway
1839 if (req->rq_transno > 0) {
1840 spin_lock(&imp->imp_lock);
1841 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
1842 imp->imp_last_replay_transno = req->rq_transno;
1843 spin_unlock(&imp->imp_lock);
1845 CERROR("Transno is 0 during replay!\n");
1846 /* continue with recovery */
1847 rc = ptlrpc_import_recovery_state_machine(imp);
1849 req->rq_send_state = aa->praa_old_state;
1852 /* this replay failed, so restart recovery */
1853 ptlrpc_connect_import(imp, NULL);
1858 int ptlrpc_replay_req(struct ptlrpc_request *req)
1860 struct ptlrpc_replay_async_args *aa;
1863 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
1865 /* Not handling automatic bulk replay yet (or ever?) */
1866 LASSERT(req->rq_bulk == NULL);
1868 DEBUG_REQ(D_HA, req, "REPLAY");
1870 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
1871 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
1872 memset(aa, 0, sizeof *aa);
1874 /* Prepare request to be resent with ptlrpcd */
1875 aa->praa_old_state = req->rq_send_state;
1876 req->rq_send_state = LUSTRE_IMP_REPLAY;
1877 req->rq_phase = RQ_PHASE_NEW;
1878 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
1881 req->rq_interpret_reply = ptlrpc_replay_interpret;
1882 atomic_inc(&req->rq_import->imp_replay_inflight);
1883 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
1885 ptlrpcd_add_req(req);
1889 void ptlrpc_abort_inflight(struct obd_import *imp)
1891 struct list_head *tmp, *n;
1894 /* Make sure that no new requests get processed for this import.
1895 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
1896 * this flag and then putting requests on sending_list or delayed_list.
1898 spin_lock(&imp->imp_lock);
1900 /* XXX locking? Maybe we should remove each request with the list
1901 * locked? Also, how do we know if the requests on the list are
1902 * being freed at this time?
1904 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
1905 struct ptlrpc_request *req =
1906 list_entry(tmp, struct ptlrpc_request, rq_list);
1908 DEBUG_REQ(D_HA, req, "inflight");
1910 spin_lock (&req->rq_lock);
1911 if (req->rq_import_generation < imp->imp_generation) {
1913 ptlrpc_wake_client_req(req);
1915 spin_unlock (&req->rq_lock);
1918 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
1919 struct ptlrpc_request *req =
1920 list_entry(tmp, struct ptlrpc_request, rq_list);
1922 DEBUG_REQ(D_HA, req, "aborting waiting req");
1924 spin_lock (&req->rq_lock);
1925 if (req->rq_import_generation < imp->imp_generation) {
1927 ptlrpc_wake_client_req(req);
1929 spin_unlock (&req->rq_lock);
1932 /* Last chance to free reqs left on the replay list, but we
1933 * will still leak reqs that haven't committed. */
1934 if (imp->imp_replayable)
1935 ptlrpc_free_committed(imp);
1937 spin_unlock(&imp->imp_lock);
1942 static __u64 ptlrpc_last_xid = 0;
1943 spinlock_t ptlrpc_last_xid_lock;
1945 __u64 ptlrpc_next_xid(void)
1948 spin_lock(&ptlrpc_last_xid_lock);
1949 tmp = ++ptlrpc_last_xid;
1950 spin_unlock(&ptlrpc_last_xid_lock);
1954 __u64 ptlrpc_sample_next_xid(void)
1957 spin_lock(&ptlrpc_last_xid_lock);
1958 tmp = ptlrpc_last_xid + 1;
1959 spin_unlock(&ptlrpc_last_xid_lock);
1962 EXPORT_SYMBOL(ptlrpc_sample_next_xid);