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>
38 #include <lustre_req_layout.h>
40 #include "ptlrpc_internal.h"
42 void ptlrpc_init_client(int req_portal, int rep_portal, char *name,
43 struct ptlrpc_client *cl)
45 cl->cli_request_portal = req_portal;
46 cl->cli_reply_portal = rep_portal;
50 struct ptlrpc_connection *ptlrpc_uuid_to_connection(struct obd_uuid *uuid)
52 struct ptlrpc_connection *c;
54 lnet_process_id_t peer;
57 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
59 CERROR("cannot find peer %s!\n", uuid->uuid);
63 c = ptlrpc_get_connection(peer, self, uuid);
65 memcpy(c->c_remote_uuid.uuid,
66 uuid->uuid, sizeof(c->c_remote_uuid.uuid));
69 CDEBUG(D_INFO, "%s -> %p\n", uuid->uuid, c);
74 void ptlrpc_readdress_connection(struct ptlrpc_connection *conn,
75 struct obd_uuid *uuid)
78 lnet_process_id_t peer;
81 err = ptlrpc_uuid_to_peer(uuid, &peer, &self);
83 CERROR("cannot find peer %s!\n", uuid->uuid);
92 static inline struct ptlrpc_bulk_desc *new_bulk(int npages, int type, int portal)
94 struct ptlrpc_bulk_desc *desc;
96 OBD_ALLOC(desc, offsetof (struct ptlrpc_bulk_desc, bd_iov[npages]));
100 spin_lock_init(&desc->bd_lock);
101 cfs_waitq_init(&desc->bd_waitq);
102 desc->bd_max_iov = npages;
103 desc->bd_iov_count = 0;
104 desc->bd_md_h = LNET_INVALID_HANDLE;
105 desc->bd_portal = portal;
106 desc->bd_type = type;
111 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_imp (struct ptlrpc_request *req,
112 int npages, int type, int portal)
114 struct obd_import *imp = req->rq_import;
115 struct ptlrpc_bulk_desc *desc;
118 LASSERT(type == BULK_PUT_SINK || type == BULK_GET_SOURCE);
119 desc = new_bulk(npages, type, portal);
123 desc->bd_import_generation = req->rq_import_generation;
124 desc->bd_import = class_import_get(imp);
127 desc->bd_cbid.cbid_fn = client_bulk_callback;
128 desc->bd_cbid.cbid_arg = desc;
130 /* This makes req own desc, and free it when she frees herself */
136 struct ptlrpc_bulk_desc *ptlrpc_prep_bulk_exp(struct ptlrpc_request *req,
137 int npages, int type, int portal)
139 struct obd_export *exp = req->rq_export;
140 struct ptlrpc_bulk_desc *desc;
143 LASSERT(type == BULK_PUT_SOURCE || type == BULK_GET_SINK);
145 desc = new_bulk(npages, type, portal);
149 desc->bd_export = class_export_get(exp);
152 desc->bd_cbid.cbid_fn = server_bulk_callback;
153 desc->bd_cbid.cbid_arg = desc;
155 /* NB we don't assign rq_bulk here; server-side requests are
156 * re-used, and the handler frees the bulk desc explicitly. */
161 void ptlrpc_prep_bulk_page(struct ptlrpc_bulk_desc *desc,
162 cfs_page_t *page, int pageoffset, int len)
164 LASSERT(desc->bd_iov_count < desc->bd_max_iov);
165 LASSERT(page != NULL);
166 LASSERT(pageoffset >= 0);
168 LASSERT(pageoffset + len <= CFS_PAGE_SIZE);
172 ptlrpc_add_bulk_page(desc, page, pageoffset, len);
175 void ptlrpc_free_bulk(struct ptlrpc_bulk_desc *desc)
179 LASSERT(desc != NULL);
180 LASSERT(desc->bd_iov_count != LI_POISON); /* not freed already */
181 LASSERT(!desc->bd_network_rw); /* network hands off or */
182 LASSERT((desc->bd_export != NULL) ^ (desc->bd_import != NULL));
184 sptlrpc_enc_pool_put_pages(desc);
187 class_export_put(desc->bd_export);
189 class_import_put(desc->bd_import);
191 OBD_FREE(desc, offsetof(struct ptlrpc_bulk_desc,
192 bd_iov[desc->bd_max_iov]));
196 void ptlrpc_free_rq_pool(struct ptlrpc_request_pool *pool)
198 struct list_head *l, *tmp;
199 struct ptlrpc_request *req;
204 list_for_each_safe(l, tmp, &pool->prp_req_list) {
205 req = list_entry(l, struct ptlrpc_request, rq_list);
206 list_del(&req->rq_list);
207 LASSERT(req->rq_reqbuf);
208 LASSERT(req->rq_reqbuf_len == pool->prp_rq_size);
209 OBD_FREE(req->rq_reqbuf, pool->prp_rq_size);
210 OBD_FREE(req, sizeof(*req));
212 OBD_FREE(pool, sizeof(*pool));
215 void ptlrpc_add_rqs_to_pool(struct ptlrpc_request_pool *pool, int num_rq)
220 while (size < pool->prp_rq_size + SPTLRPC_MAX_PAYLOAD)
223 LASSERTF(list_empty(&pool->prp_req_list) || size == pool->prp_rq_size,
224 "Trying to change pool size with nonempty pool "
225 "from %d to %d bytes\n", pool->prp_rq_size, size);
227 spin_lock(&pool->prp_lock);
228 pool->prp_rq_size = size;
229 for (i = 0; i < num_rq; i++) {
230 struct ptlrpc_request *req;
231 struct lustre_msg *msg;
233 spin_unlock(&pool->prp_lock);
234 OBD_ALLOC(req, sizeof(struct ptlrpc_request));
237 OBD_ALLOC_GFP(msg, size, CFS_ALLOC_STD);
239 OBD_FREE(req, sizeof(struct ptlrpc_request));
242 req->rq_reqbuf = msg;
243 req->rq_reqbuf_len = size;
245 spin_lock(&pool->prp_lock);
246 list_add_tail(&req->rq_list, &pool->prp_req_list);
248 spin_unlock(&pool->prp_lock);
252 struct ptlrpc_request_pool *ptlrpc_init_rq_pool(int num_rq, int msgsize,
253 void (*populate_pool)(struct ptlrpc_request_pool *, int))
255 struct ptlrpc_request_pool *pool;
257 OBD_ALLOC(pool, sizeof (struct ptlrpc_request_pool));
261 /* Request next power of two for the allocation, because internally
262 kernel would do exactly this */
264 spin_lock_init(&pool->prp_lock);
265 CFS_INIT_LIST_HEAD(&pool->prp_req_list);
266 pool->prp_rq_size = msgsize;
267 pool->prp_populate = populate_pool;
269 populate_pool(pool, num_rq);
271 if (list_empty(&pool->prp_req_list)) {
272 /* have not allocated a single request for the pool */
273 OBD_FREE(pool, sizeof (struct ptlrpc_request_pool));
279 static struct ptlrpc_request *ptlrpc_prep_req_from_pool(struct ptlrpc_request_pool *pool)
281 struct ptlrpc_request *request;
282 struct lustre_msg *reqbuf;
287 spin_lock(&pool->prp_lock);
289 /* See if we have anything in a pool, and bail out if nothing,
290 * in writeout path, where this matters, this is safe to do, because
291 * nothing is lost in this case, and when some in-flight requests
292 * complete, this code will be called again. */
293 if (unlikely(list_empty(&pool->prp_req_list))) {
294 spin_unlock(&pool->prp_lock);
298 request = list_entry(pool->prp_req_list.next, struct ptlrpc_request,
300 list_del(&request->rq_list);
301 spin_unlock(&pool->prp_lock);
303 LASSERT(request->rq_reqbuf);
304 LASSERT(request->rq_pool);
306 reqbuf = request->rq_reqbuf;
307 memset(request, 0, sizeof(*request));
308 request->rq_reqbuf = reqbuf;
309 request->rq_reqbuf_len = pool->prp_rq_size;
310 request->rq_pool = pool;
314 static void __ptlrpc_free_req_to_pool(struct ptlrpc_request *request)
316 struct ptlrpc_request_pool *pool = request->rq_pool;
318 spin_lock(&pool->prp_lock);
319 LASSERT(list_empty(&request->rq_list));
320 list_add_tail(&request->rq_list, &pool->prp_req_list);
321 spin_unlock(&pool->prp_lock);
324 static int __ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
325 __u32 version, int opcode,
326 int count, int *lengths, char **bufs,
327 struct ptlrpc_cli_ctx *ctx)
329 struct obd_import *imp = request->rq_import;
334 request->rq_cli_ctx = sptlrpc_cli_ctx_get(ctx);
336 rc = sptlrpc_req_get_ctx(request);
341 sptlrpc_req_set_flavor(request, opcode);
343 rc = lustre_pack_request(request, imp->imp_msg_magic, count,
346 LASSERT(!request->rq_pool);
350 lustre_msg_add_version(request->rq_reqmsg, version);
352 if (imp->imp_server_timeout)
353 request->rq_timeout = obd_timeout / 2;
355 request->rq_timeout = obd_timeout;
356 request->rq_send_state = LUSTRE_IMP_FULL;
357 request->rq_type = PTL_RPC_MSG_REQUEST;
358 request->rq_export = NULL;
360 request->rq_req_cbid.cbid_fn = request_out_callback;
361 request->rq_req_cbid.cbid_arg = request;
363 request->rq_reply_cbid.cbid_fn = reply_in_callback;
364 request->rq_reply_cbid.cbid_arg = request;
366 request->rq_phase = RQ_PHASE_NEW;
368 /* XXX FIXME bug 249 */
369 request->rq_request_portal = imp->imp_client->cli_request_portal;
370 request->rq_reply_portal = imp->imp_client->cli_reply_portal;
372 spin_lock_init(&request->rq_lock);
373 CFS_INIT_LIST_HEAD(&request->rq_list);
374 CFS_INIT_LIST_HEAD(&request->rq_replay_list);
375 CFS_INIT_LIST_HEAD(&request->rq_mod_list);
376 CFS_INIT_LIST_HEAD(&request->rq_ctx_chain);
377 CFS_INIT_LIST_HEAD(&request->rq_set_chain);
378 cfs_waitq_init(&request->rq_reply_waitq);
379 request->rq_xid = ptlrpc_next_xid();
380 atomic_set(&request->rq_refcount, 1);
382 lustre_msg_set_opc(request->rq_reqmsg, opcode);
383 lustre_msg_set_flags(request->rq_reqmsg, 0);
387 sptlrpc_cli_ctx_put(request->rq_cli_ctx, 1);
389 class_import_put(imp);
393 int ptlrpc_request_bufs_pack(struct ptlrpc_request *request,
394 __u32 version, int opcode, char **bufs,
395 struct ptlrpc_cli_ctx *ctx)
399 count = req_capsule_filled_sizes(&request->rq_pill, RCL_CLIENT);
400 return __ptlrpc_request_bufs_pack(request, version, opcode, count,
401 request->rq_pill.rc_area[RCL_CLIENT],
404 EXPORT_SYMBOL(ptlrpc_request_bufs_pack);
406 int ptlrpc_request_pack(struct ptlrpc_request *request,
407 __u32 version, int opcode)
409 return ptlrpc_request_bufs_pack(request, version, opcode, NULL, NULL);
413 struct ptlrpc_request *__ptlrpc_request_alloc(struct obd_import *imp,
414 struct ptlrpc_request_pool *pool)
416 struct ptlrpc_request *request = NULL;
419 request = ptlrpc_prep_req_from_pool(pool);
422 OBD_ALLOC_PTR(request);
425 LASSERT((unsigned long)imp > 0x1000);
426 LASSERT(imp != LP_POISON);
427 LASSERT((unsigned long)imp->imp_client > 0x1000);
428 LASSERT(imp->imp_client != LP_POISON);
430 request->rq_import = class_import_get(imp);
432 CERROR("request allocation out of memory\n");
438 static struct ptlrpc_request *
439 ptlrpc_request_alloc_internal(struct obd_import *imp,
440 struct ptlrpc_request_pool * pool,
441 const struct req_format *format)
443 struct ptlrpc_request *request;
445 request = __ptlrpc_request_alloc(imp, pool);
449 req_capsule_init(&request->rq_pill, request, RCL_CLIENT);
450 req_capsule_set(&request->rq_pill, format);
454 struct ptlrpc_request *ptlrpc_request_alloc(struct obd_import *imp,
455 const struct req_format *format)
457 return ptlrpc_request_alloc_internal(imp, NULL, format);
460 struct ptlrpc_request *ptlrpc_request_alloc_pool(struct obd_import *imp,
461 struct ptlrpc_request_pool * pool,
462 const struct req_format *format)
464 return ptlrpc_request_alloc_internal(imp, pool, format);
467 void ptlrpc_request_free(struct ptlrpc_request *request)
469 if (request->rq_pool)
470 __ptlrpc_free_req_to_pool(request);
472 OBD_FREE_PTR(request);
475 struct ptlrpc_request *ptlrpc_request_alloc_pack(struct obd_import *imp,
476 const struct req_format *format,
477 __u32 version, int opcode)
479 struct ptlrpc_request *req = ptlrpc_request_alloc(imp, format);
483 rc = ptlrpc_request_pack(req, version, opcode);
485 ptlrpc_request_free(req);
492 struct ptlrpc_request *
493 ptlrpc_prep_req_pool(struct obd_import *imp,
494 __u32 version, int opcode,
495 int count, int *lengths, char **bufs,
496 struct ptlrpc_request_pool *pool)
498 struct ptlrpc_request *request;
501 request = __ptlrpc_request_alloc(imp, pool);
505 rc = __ptlrpc_request_bufs_pack(request, version, opcode, count,
506 lengths, bufs, NULL);
508 ptlrpc_request_free(request);
514 struct ptlrpc_request *
515 ptlrpc_prep_req(struct obd_import *imp, __u32 version, int opcode, int count,
516 int *lengths, char **bufs)
518 return ptlrpc_prep_req_pool(imp, version, opcode, count, lengths, bufs,
522 struct ptlrpc_request_set *ptlrpc_prep_set(void)
524 struct ptlrpc_request_set *set;
527 OBD_ALLOC(set, sizeof *set);
530 CFS_INIT_LIST_HEAD(&set->set_requests);
531 cfs_waitq_init(&set->set_waitq);
532 set->set_remaining = 0;
533 spin_lock_init(&set->set_new_req_lock);
534 CFS_INIT_LIST_HEAD(&set->set_new_requests);
535 CFS_INIT_LIST_HEAD(&set->set_cblist);
540 /* Finish with this set; opposite of prep_set. */
541 void ptlrpc_set_destroy(struct ptlrpc_request_set *set)
543 struct list_head *tmp;
544 struct list_head *next;
549 /* Requests on the set should either all be completed, or all be new */
550 expected_phase = (set->set_remaining == 0) ?
551 RQ_PHASE_COMPLETE : RQ_PHASE_NEW;
552 list_for_each (tmp, &set->set_requests) {
553 struct ptlrpc_request *req =
554 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
556 LASSERT(req->rq_phase == expected_phase);
560 LASSERT(set->set_remaining == 0 || set->set_remaining == n);
562 list_for_each_safe(tmp, next, &set->set_requests) {
563 struct ptlrpc_request *req =
564 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
565 list_del_init(&req->rq_set_chain);
567 LASSERT(req->rq_phase == expected_phase);
569 if (req->rq_phase == RQ_PHASE_NEW) {
571 if (req->rq_interpret_reply != NULL) {
572 int (*interpreter)(struct ptlrpc_request *,
574 req->rq_interpret_reply;
576 /* higher level (i.e. LOV) failed;
577 * let the sub reqs clean up */
578 req->rq_status = -EBADR;
579 interpreter(req, &req->rq_async_args,
582 set->set_remaining--;
586 ptlrpc_req_finished (req);
589 LASSERT(set->set_remaining == 0);
591 OBD_FREE(set, sizeof(*set));
595 int ptlrpc_set_add_cb(struct ptlrpc_request_set *set,
596 set_interpreter_func fn, void *data)
598 struct ptlrpc_set_cbdata *cbdata;
600 OBD_ALLOC_PTR(cbdata);
604 cbdata->psc_interpret = fn;
605 cbdata->psc_data = data;
606 list_add_tail(&cbdata->psc_item, &set->set_cblist);
611 void ptlrpc_set_add_req(struct ptlrpc_request_set *set,
612 struct ptlrpc_request *req)
614 /* The set takes over the caller's request reference */
615 list_add_tail(&req->rq_set_chain, &set->set_requests);
617 set->set_remaining++;
619 atomic_inc(&req->rq_import->imp_inflight);
622 /* lock so many callers can add things, the context that owns the set
623 * is supposed to notice these and move them into the set proper. */
624 void ptlrpc_set_add_new_req(struct ptlrpc_request_set *set,
625 struct ptlrpc_request *req)
627 spin_lock(&set->set_new_req_lock);
628 /* The set takes over the caller's request reference */
629 list_add_tail(&req->rq_set_chain, &set->set_new_requests);
631 spin_unlock(&set->set_new_req_lock);
635 * Based on the current state of the import, determine if the request
636 * can be sent, is an error, or should be delayed.
638 * Returns true if this request should be delayed. If false, and
639 * *status is set, then the request can not be sent and *status is the
640 * error code. If false and status is 0, then request can be sent.
642 * The imp->imp_lock must be held.
644 static int ptlrpc_import_delay_req(struct obd_import *imp,
645 struct ptlrpc_request *req, int *status)
650 LASSERT (status != NULL);
653 if (req->rq_ctx_init || req->rq_ctx_fini) {
654 /* always allow ctx init/fini rpc go through */
655 } else if (imp->imp_state == LUSTRE_IMP_NEW) {
656 DEBUG_REQ(D_ERROR, req, "Uninitialized import.");
659 } else if (imp->imp_state == LUSTRE_IMP_CLOSED) {
660 DEBUG_REQ(D_ERROR, req, "IMP_CLOSED ");
662 } else if (req->rq_send_state == LUSTRE_IMP_CONNECTING &&
663 imp->imp_state == LUSTRE_IMP_CONNECTING) {
664 /* allow CONNECT even if import is invalid */ ;
665 } else if (imp->imp_invalid) {
666 /* if it is mgc, wait for recovry. b=13464 */
667 if (imp->imp_recon_bk && !imp->imp_obd->obd_no_recov)
669 /* If the import has been invalidated (such as by an OST
670 * failure) the request must fail with -ESHUTDOWN. This
671 * indicates the requests should be discarded; an -EIO
672 * may result in a resend of the request. */
673 if (!imp->imp_deactive)
674 DEBUG_REQ(D_ERROR, req, "IMP_INVALID");
675 *status = -ESHUTDOWN; /* bz 12940 */
676 } else if (req->rq_import_generation != imp->imp_generation) {
677 DEBUG_REQ(D_ERROR, req, "req wrong generation:");
679 } else if (req->rq_send_state != imp->imp_state) {
680 if (imp->imp_obd->obd_no_recov || imp->imp_dlm_fake ||
682 *status = -EWOULDBLOCK;
690 static int ptlrpc_check_reply(struct ptlrpc_request *req)
695 /* serialise with network callback */
696 spin_lock(&req->rq_lock);
701 if (req->rq_net_err && !req->rq_timedout) {
702 spin_unlock(&req->rq_lock);
703 rc = ptlrpc_expire_one_request(req);
704 spin_lock(&req->rq_lock);
718 spin_unlock(&req->rq_lock);
719 DEBUG_REQ(D_NET, req, "rc = %d for", rc);
723 static int ptlrpc_check_status(struct ptlrpc_request *req)
728 err = lustre_msg_get_status(req->rq_repmsg);
729 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR) {
730 struct obd_import *imp = req->rq_import;
731 __u32 opc = lustre_msg_get_opc(req->rq_reqmsg);
732 LCONSOLE_ERROR_MSG(0x011,"an error occurred while communicating"
733 " with %s. The %s operation failed with %d\n",
734 libcfs_nid2str(imp->imp_connection->c_peer.nid),
735 ll_opcode2str(opc), err);
736 RETURN(err < 0 ? err : -EINVAL);
740 DEBUG_REQ(D_INFO, req, "status is %d", err);
741 } else if (err > 0) {
742 /* XXX: translate this error from net to host */
743 DEBUG_REQ(D_INFO, req, "status is %d", err);
749 static int after_reply(struct ptlrpc_request *req)
751 struct obd_import *imp = req->rq_import;
752 struct obd_device *obd = req->rq_import->imp_obd;
754 struct timeval work_start;
758 LASSERT(!req->rq_receiving_reply);
760 LASSERT(req->rq_nob_received <= req->rq_repbuf_len);
762 /* NB Until this point, the whole of the incoming message,
763 * including buflens, status etc is in the sender's byte order. */
765 /* Clear reply swab mask; this is a new reply in sender's byte order */
766 req->rq_rep_swab_mask = 0;
768 rc = sptlrpc_cli_unwrap_reply(req);
770 DEBUG_REQ(D_ERROR, req, "unwrap reply failed (%d):", rc);
774 /* security layer unwrap might ask resend this request */
778 rc = lustre_unpack_msg(req->rq_repmsg, req->rq_replen);
780 DEBUG_REQ(D_ERROR, req, "unpack_rep failed: %d", rc);
784 rc = lustre_unpack_rep_ptlrpc_body(req, MSG_PTLRPC_BODY_OFF);
786 DEBUG_REQ(D_ERROR, req, "unpack ptlrpc body failed: %d", rc);
790 do_gettimeofday(&work_start);
791 timediff = cfs_timeval_sub(&work_start, &req->rq_arrival_time, NULL);
792 if (obd->obd_svc_stats != NULL)
793 lprocfs_counter_add(obd->obd_svc_stats, PTLRPC_REQWAIT_CNTR,
796 if (lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_REPLY &&
797 lustre_msg_get_type(req->rq_repmsg) != PTL_RPC_MSG_ERR) {
798 DEBUG_REQ(D_ERROR, req, "invalid packet received (type=%u)",
799 lustre_msg_get_type(req->rq_repmsg));
803 rc = ptlrpc_check_status(req);
804 imp->imp_connect_error = rc;
807 /* Either we've been evicted, or the server has failed for
808 * some reason. Try to reconnect, and if that fails, punt to
810 if (ll_rpc_recoverable_error(rc)) {
811 if (req->rq_send_state != LUSTRE_IMP_FULL ||
812 imp->imp_obd->obd_no_recov || imp->imp_dlm_fake) {
815 ptlrpc_request_handle_notconn(req);
819 /* Let's look if server send slv. Do it only for RPC with
821 if (imp->imp_obd->obd_namespace) {
822 /* Disconnect rpc is sent when namespace is already
823 * destroyed. Let's check this and will not try update
825 ldlm_cli_update_pool(req);
829 /* Store transno in reqmsg for replay. */
830 req->rq_transno = lustre_msg_get_transno(req->rq_repmsg);
831 lustre_msg_set_transno(req->rq_reqmsg, req->rq_transno);
833 if (req->rq_import->imp_replayable) {
834 spin_lock(&imp->imp_lock);
835 /* no point in adding already-committed requests to the replay
836 * list, we will just remove them immediately. b=9829 */
837 if (req->rq_transno != 0 &&
839 lustre_msg_get_last_committed(req->rq_repmsg) ||
841 ptlrpc_retain_replayable_request(req, imp);
842 else if (req->rq_commit_cb != NULL) {
843 spin_unlock(&imp->imp_lock);
844 req->rq_commit_cb(req);
845 spin_lock(&imp->imp_lock);
848 /* Replay-enabled imports return commit-status information. */
849 if (lustre_msg_get_last_committed(req->rq_repmsg)) {
850 imp->imp_peer_committed_transno =
851 lustre_msg_get_last_committed(req->rq_repmsg);
853 ptlrpc_free_committed(imp);
854 spin_unlock(&imp->imp_lock);
860 static int ptlrpc_send_new_req(struct ptlrpc_request *req)
862 struct obd_import *imp;
866 LASSERT(req->rq_phase == RQ_PHASE_NEW);
867 if (req->rq_sent && (req->rq_sent > CURRENT_SECONDS))
870 req->rq_phase = RQ_PHASE_RPC;
872 imp = req->rq_import;
873 spin_lock(&imp->imp_lock);
875 req->rq_import_generation = imp->imp_generation;
877 if (ptlrpc_import_delay_req(imp, req, &rc)) {
878 spin_lock (&req->rq_lock);
880 spin_unlock (&req->rq_lock);
882 DEBUG_REQ(D_HA, req, "req from PID %d waiting for recovery: "
884 lustre_msg_get_status(req->rq_reqmsg) ,
885 ptlrpc_import_state_name(req->rq_send_state),
886 ptlrpc_import_state_name(imp->imp_state));
887 LASSERT(list_empty (&req->rq_list));
889 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
890 spin_unlock(&imp->imp_lock);
895 spin_unlock(&imp->imp_lock);
897 req->rq_phase = RQ_PHASE_INTERPRET;
901 /* XXX this is the same as ptlrpc_queue_wait */
902 LASSERT(list_empty(&req->rq_list));
903 list_add_tail(&req->rq_list, &imp->imp_sending_list);
904 spin_unlock(&imp->imp_lock);
906 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
908 rc = sptlrpc_req_refresh_ctx(req, -1);
914 /* here begins timeout counting */
915 req->rq_sent = CURRENT_SECONDS;
916 req->rq_wait_ctx = 1;
921 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc"
922 " %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
923 imp->imp_obd->obd_uuid.uuid,
924 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
925 libcfs_nid2str(imp->imp_connection->c_peer.nid),
926 lustre_msg_get_opc(req->rq_reqmsg));
928 rc = ptl_send_rpc(req, 0);
930 DEBUG_REQ(D_HA, req, "send failed (%d); expect timeout", rc);
937 /* this sends any unsent RPCs in @set and returns TRUE if all are sent */
938 int ptlrpc_check_set(struct ptlrpc_request_set *set)
940 struct list_head *tmp;
941 int force_timer_recalc = 0;
944 if (set->set_remaining == 0)
947 list_for_each(tmp, &set->set_requests) {
948 struct ptlrpc_request *req =
949 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
950 struct obd_import *imp = req->rq_import;
953 if (req->rq_phase == RQ_PHASE_NEW &&
954 ptlrpc_send_new_req(req)) {
955 force_timer_recalc = 1;
957 /* delayed send - skip */
958 if (req->rq_phase == RQ_PHASE_NEW && req->rq_sent)
961 if (!(req->rq_phase == RQ_PHASE_RPC ||
962 req->rq_phase == RQ_PHASE_BULK ||
963 req->rq_phase == RQ_PHASE_INTERPRET ||
964 req->rq_phase == RQ_PHASE_COMPLETE)) {
965 DEBUG_REQ(D_ERROR, req, "bad phase %x", req->rq_phase);
969 if (req->rq_phase == RQ_PHASE_COMPLETE)
972 if (req->rq_phase == RQ_PHASE_INTERPRET)
973 GOTO(interpret, req->rq_status);
975 if (req->rq_net_err && !req->rq_timedout)
976 ptlrpc_expire_one_request(req);
979 ptlrpc_unregister_reply(req);
981 if (req->rq_status == 0)
982 req->rq_status = -EIO;
983 req->rq_phase = RQ_PHASE_INTERPRET;
985 spin_lock(&imp->imp_lock);
986 list_del_init(&req->rq_list);
987 spin_unlock(&imp->imp_lock);
989 GOTO(interpret, req->rq_status);
992 /* ptlrpc_queue_wait->l_wait_event guarantees that rq_intr
993 * will only be set after rq_timedout, but the oig waiting
994 * path sets rq_intr irrespective of whether ptlrpcd has
995 * seen a timeout. our policy is to only interpret
996 * interrupted rpcs after they have timed out */
997 if (req->rq_intr && (req->rq_timedout || req->rq_waiting ||
999 /* NB could be on delayed list */
1000 ptlrpc_unregister_reply(req);
1001 req->rq_status = -EINTR;
1002 req->rq_phase = RQ_PHASE_INTERPRET;
1004 spin_lock(&imp->imp_lock);
1005 list_del_init(&req->rq_list);
1006 spin_unlock(&imp->imp_lock);
1008 GOTO(interpret, req->rq_status);
1011 if (req->rq_phase == RQ_PHASE_RPC) {
1012 if (req->rq_timedout || req->rq_resend ||
1013 req->rq_waiting || req->rq_wait_ctx) {
1016 /* rq_wait_ctx is only touched in ptlrpcd,
1017 * no lock needed here.
1019 if (req->rq_wait_ctx)
1022 ptlrpc_unregister_reply(req);
1024 spin_lock(&imp->imp_lock);
1026 if (ptlrpc_import_delay_req(imp, req, &status)){
1027 spin_unlock(&imp->imp_lock);
1031 list_del_init(&req->rq_list);
1033 req->rq_status = status;
1034 req->rq_phase = RQ_PHASE_INTERPRET;
1035 spin_unlock(&imp->imp_lock);
1036 GOTO(interpret, req->rq_status);
1038 if (req->rq_no_resend && !req->rq_wait_ctx) {
1039 req->rq_status = -ENOTCONN;
1040 req->rq_phase = RQ_PHASE_INTERPRET;
1041 spin_unlock(&imp->imp_lock);
1042 GOTO(interpret, req->rq_status);
1044 list_add_tail(&req->rq_list,
1045 &imp->imp_sending_list);
1047 spin_unlock(&imp->imp_lock);
1049 req->rq_waiting = 0;
1050 if (req->rq_resend) {
1051 lustre_msg_add_flags(req->rq_reqmsg,
1054 __u64 old_xid = req->rq_xid;
1056 ptlrpc_unregister_bulk (req);
1058 /* ensure previous bulk fails */
1059 req->rq_xid = ptlrpc_next_xid();
1060 CDEBUG(D_HA, "resend bulk "
1063 old_xid, req->rq_xid);
1067 status = sptlrpc_req_refresh_ctx(req, -1);
1070 req->rq_status = status;
1071 force_timer_recalc = 1;
1073 if (!req->rq_wait_ctx) {
1074 /* begins timeout counting */
1075 req->rq_sent = CURRENT_SECONDS;
1076 req->rq_wait_ctx = 1;
1081 req->rq_wait_ctx = 0;
1084 rc = ptl_send_rpc(req, 0);
1086 DEBUG_REQ(D_HA, req, "send failed (%d)",
1088 force_timer_recalc = 1;
1089 req->rq_net_err = 1;
1091 /* need to reset the timeout */
1092 force_timer_recalc = 1;
1095 /* Still waiting for a reply? */
1096 if (ptlrpc_client_receiving_reply(req))
1099 /* Did we actually receive a reply? */
1100 if (!ptlrpc_client_replied(req))
1103 spin_lock(&imp->imp_lock);
1104 list_del_init(&req->rq_list);
1105 spin_unlock(&imp->imp_lock);
1107 req->rq_status = after_reply(req);
1108 if (req->rq_resend) {
1109 /* Add this req to the delayed list so
1110 it can be errored if the import is
1111 evicted after recovery. */
1112 spin_lock(&imp->imp_lock);
1113 list_add_tail(&req->rq_list,
1114 &imp->imp_delayed_list);
1115 spin_unlock(&imp->imp_lock);
1119 /* If there is no bulk associated with this request,
1120 * then we're done and should let the interpreter
1121 * process the reply. Similarly if the RPC returned
1122 * an error, and therefore the bulk will never arrive.
1124 if (req->rq_bulk == NULL || req->rq_status != 0) {
1125 req->rq_phase = RQ_PHASE_INTERPRET;
1126 GOTO(interpret, req->rq_status);
1129 req->rq_phase = RQ_PHASE_BULK;
1132 LASSERT(req->rq_phase == RQ_PHASE_BULK);
1133 if (ptlrpc_bulk_active(req->rq_bulk))
1136 if (!req->rq_bulk->bd_success) {
1137 /* The RPC reply arrived OK, but the bulk screwed
1138 * up! Dead wierd since the server told us the RPC
1139 * was good after getting the REPLY for her GET or
1140 * the ACK for her PUT. */
1141 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1145 req->rq_phase = RQ_PHASE_INTERPRET;
1148 LASSERT(req->rq_phase == RQ_PHASE_INTERPRET);
1149 LASSERT(!req->rq_receiving_reply);
1151 ptlrpc_unregister_reply(req);
1152 if (req->rq_bulk != NULL)
1153 ptlrpc_unregister_bulk (req);
1155 if (req->rq_interpret_reply != NULL) {
1156 int (*interpreter)(struct ptlrpc_request *,void *,int) =
1157 req->rq_interpret_reply;
1158 req->rq_status = interpreter(req, &req->rq_async_args,
1161 req->rq_phase = RQ_PHASE_COMPLETE;
1163 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:"
1164 "opc %s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1165 imp->imp_obd->obd_uuid.uuid,
1166 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1167 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1168 lustre_msg_get_opc(req->rq_reqmsg));
1170 set->set_remaining--;
1172 atomic_dec(&imp->imp_inflight);
1173 cfs_waitq_signal(&imp->imp_recovery_waitq);
1176 /* If we hit an error, we want to recover promptly. */
1177 RETURN(set->set_remaining == 0 || force_timer_recalc);
1180 int ptlrpc_expire_one_request(struct ptlrpc_request *req)
1182 struct obd_import *imp = req->rq_import;
1186 DEBUG_REQ(D_ERROR|D_NETERROR, req, "%s (sent at %lu, %lus ago)",
1187 req->rq_net_err ? "network error" : "timeout",
1188 (long)req->rq_sent, CURRENT_SECONDS - req->rq_sent);
1190 if (imp != NULL && obd_debug_peer_on_timeout)
1191 LNetCtl(IOC_LIBCFS_DEBUG_PEER, &imp->imp_connection->c_peer);
1193 spin_lock(&req->rq_lock);
1194 req->rq_timedout = 1;
1195 req->rq_wait_ctx = 0;
1196 spin_unlock(&req->rq_lock);
1198 ptlrpc_unregister_reply (req);
1200 if (obd_dump_on_timeout)
1201 libcfs_debug_dumplog();
1203 if (req->rq_bulk != NULL)
1204 ptlrpc_unregister_bulk (req);
1207 DEBUG_REQ(D_HA, req, "NULL import: already cleaned up?");
1211 /* The DLM server doesn't want recovery run on its imports. */
1212 if (imp->imp_dlm_fake)
1215 /* If this request is for recovery or other primordial tasks,
1216 * then error it out here. */
1217 if (req->rq_ctx_init || req->rq_ctx_fini ||
1218 req->rq_send_state != LUSTRE_IMP_FULL ||
1219 imp->imp_obd->obd_no_recov) {
1220 spin_lock(&req->rq_lock);
1221 req->rq_status = -ETIMEDOUT;
1223 spin_unlock(&req->rq_lock);
1227 /* if request can't be resend we can't wait answer after timeout */
1228 if (req->rq_no_resend) {
1229 DEBUG_REQ(D_RPCTRACE, req, "TIMEOUT-NORESEND:");
1233 ptlrpc_fail_import(imp, lustre_msg_get_conn_cnt(req->rq_reqmsg));
1238 int ptlrpc_expired_set(void *data)
1240 struct ptlrpc_request_set *set = data;
1241 struct list_head *tmp;
1242 time_t now = CURRENT_SECONDS;
1245 LASSERT(set != NULL);
1247 /* A timeout expired; see which reqs it applies to... */
1248 list_for_each (tmp, &set->set_requests) {
1249 struct ptlrpc_request *req =
1250 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1252 /* request in-flight? */
1253 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting &&
1255 (req->rq_phase == RQ_PHASE_BULK)))
1258 if (req->rq_timedout || /* already dealt with */
1259 req->rq_sent + req->rq_timeout > now) /* not expired */
1262 /* deal with this guy */
1263 ptlrpc_expire_one_request (req);
1266 /* When waiting for a whole set, we always to break out of the
1267 * sleep so we can recalculate the timeout, or enable interrupts
1268 * iff everyone's timed out.
1273 void ptlrpc_mark_interrupted(struct ptlrpc_request *req)
1275 spin_lock(&req->rq_lock);
1277 spin_unlock(&req->rq_lock);
1280 void ptlrpc_interrupted_set(void *data)
1282 struct ptlrpc_request_set *set = data;
1283 struct list_head *tmp;
1285 LASSERT(set != NULL);
1286 CERROR("INTERRUPTED SET %p\n", set);
1288 list_for_each(tmp, &set->set_requests) {
1289 struct ptlrpc_request *req =
1290 list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1292 if (req->rq_phase != RQ_PHASE_RPC)
1295 ptlrpc_mark_interrupted(req);
1299 int ptlrpc_set_next_timeout(struct ptlrpc_request_set *set)
1301 struct list_head *tmp;
1302 time_t now = CURRENT_SECONDS;
1305 struct ptlrpc_request *req;
1308 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
1310 list_for_each(tmp, &set->set_requests) {
1311 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1313 /* request in-flight? */
1314 if (!((req->rq_phase == RQ_PHASE_RPC && !req->rq_waiting) ||
1315 (req->rq_phase == RQ_PHASE_BULK) ||
1316 (req->rq_phase == RQ_PHASE_NEW)))
1319 if (req->rq_timedout) /* already timed out */
1322 if (req->rq_phase == RQ_PHASE_NEW)
1323 deadline = req->rq_sent;
1325 deadline = req->rq_sent + req->rq_timeout;
1327 if (deadline <= now) /* actually expired already */
1328 timeout = 1; /* ASAP */
1329 else if (timeout == 0 || timeout > deadline - now)
1330 timeout = deadline - now;
1335 int ptlrpc_set_wait(struct ptlrpc_request_set *set)
1337 struct list_head *tmp;
1338 struct ptlrpc_request *req;
1339 struct l_wait_info lwi;
1343 if (list_empty(&set->set_requests))
1346 list_for_each(tmp, &set->set_requests) {
1347 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1348 if (req->rq_phase == RQ_PHASE_NEW)
1349 (void)ptlrpc_send_new_req(req);
1353 timeout = ptlrpc_set_next_timeout(set);
1355 /* wait until all complete, interrupted, or an in-flight
1357 CDEBUG(D_RPCTRACE, "set %p going to sleep for %d seconds\n",
1359 lwi = LWI_TIMEOUT_INTR(cfs_time_seconds(timeout ? timeout : 1),
1361 ptlrpc_interrupted_set, set);
1362 rc = l_wait_event(set->set_waitq, ptlrpc_check_set(set), &lwi);
1364 LASSERT(rc == 0 || rc == -EINTR || rc == -ETIMEDOUT);
1366 /* -EINTR => all requests have been flagged rq_intr so next
1368 * -ETIMEOUTD => someone timed out. When all reqs have
1369 * timed out, signals are enabled allowing completion with
1371 * I don't really care if we go once more round the loop in
1372 * the error cases -eeb. */
1373 } while (rc != 0 || set->set_remaining != 0);
1375 LASSERT(set->set_remaining == 0);
1378 list_for_each(tmp, &set->set_requests) {
1379 req = list_entry(tmp, struct ptlrpc_request, rq_set_chain);
1381 LASSERT(req->rq_phase == RQ_PHASE_COMPLETE);
1382 if (req->rq_status != 0)
1383 rc = req->rq_status;
1386 if (set->set_interpret != NULL) {
1387 int (*interpreter)(struct ptlrpc_request_set *set,void *,int) =
1389 rc = interpreter (set, set->set_arg, rc);
1391 struct ptlrpc_set_cbdata *cbdata, *n;
1394 list_for_each_entry_safe(cbdata, n,
1395 &set->set_cblist, psc_item) {
1396 list_del_init(&cbdata->psc_item);
1397 err = cbdata->psc_interpret(set, cbdata->psc_data, rc);
1400 OBD_FREE_PTR(cbdata);
1407 static void __ptlrpc_free_req(struct ptlrpc_request *request, int locked)
1410 if (request == NULL) {
1415 LASSERTF(!request->rq_receiving_reply, "req %p\n", request);
1416 LASSERTF(request->rq_rqbd == NULL, "req %p\n",request);/* client-side */
1417 LASSERTF(list_empty(&request->rq_list), "req %p\n", request);
1418 LASSERTF(list_empty(&request->rq_set_chain), "req %p\n", request);
1419 LASSERT(request->rq_cli_ctx);
1421 req_capsule_fini(&request->rq_pill);
1423 /* We must take it off the imp_replay_list first. Otherwise, we'll set
1424 * request->rq_reqmsg to NULL while osc_close is dereferencing it. */
1425 if (request->rq_import != NULL) {
1427 spin_lock(&request->rq_import->imp_lock);
1428 list_del_init(&request->rq_mod_list);
1429 list_del_init(&request->rq_replay_list);
1431 spin_unlock(&request->rq_import->imp_lock);
1433 LASSERTF(list_empty(&request->rq_replay_list), "req %p\n", request);
1435 if (atomic_read(&request->rq_refcount) != 0) {
1436 DEBUG_REQ(D_ERROR, request,
1437 "freeing request with nonzero refcount");
1441 if (request->rq_repbuf != NULL)
1442 sptlrpc_cli_free_repbuf(request);
1443 if (request->rq_export != NULL) {
1444 class_export_put(request->rq_export);
1445 request->rq_export = NULL;
1447 if (request->rq_import != NULL) {
1448 class_import_put(request->rq_import);
1449 request->rq_import = NULL;
1451 if (request->rq_bulk != NULL)
1452 ptlrpc_free_bulk(request->rq_bulk);
1454 if (request->rq_reqbuf != NULL || request->rq_clrbuf != NULL)
1455 sptlrpc_cli_free_reqbuf(request);
1457 sptlrpc_req_put_ctx(request, !locked);
1459 if (request->rq_pool)
1460 __ptlrpc_free_req_to_pool(request);
1462 OBD_FREE(request, sizeof(*request));
1466 void ptlrpc_free_req(struct ptlrpc_request *request)
1468 __ptlrpc_free_req(request, 0);
1471 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked);
1472 void ptlrpc_req_finished_with_imp_lock(struct ptlrpc_request *request)
1474 LASSERT_SPIN_LOCKED(&request->rq_import->imp_lock);
1475 (void)__ptlrpc_req_finished(request, 1);
1478 static int __ptlrpc_req_finished(struct ptlrpc_request *request, int locked)
1481 if (request == NULL)
1484 if (request == LP_POISON ||
1485 request->rq_reqmsg == LP_POISON) {
1486 CERROR("dereferencing freed request (bug 575)\n");
1491 DEBUG_REQ(D_INFO, request, "refcount now %u",
1492 atomic_read(&request->rq_refcount) - 1);
1494 if (atomic_dec_and_test(&request->rq_refcount)) {
1495 __ptlrpc_free_req(request, locked);
1502 void ptlrpc_req_finished(struct ptlrpc_request *request)
1504 __ptlrpc_req_finished(request, 0);
1507 __u64 ptlrpc_req_xid(struct ptlrpc_request *request)
1509 return request->rq_xid;
1511 EXPORT_SYMBOL(ptlrpc_req_xid);
1513 /* Disengage the client's reply buffer from the network
1514 * NB does _NOT_ unregister any client-side bulk.
1515 * IDEMPOTENT, but _not_ safe against concurrent callers.
1516 * The request owner (i.e. the thread doing the I/O) must call...
1518 void ptlrpc_unregister_reply (struct ptlrpc_request *request)
1522 struct l_wait_info lwi;
1524 LASSERT(!in_interrupt ()); /* might sleep */
1526 if (!ptlrpc_client_receiving_reply(request))
1529 LNetMDUnlink (request->rq_reply_md_h);
1531 /* We have to l_wait_event() whatever the result, to give liblustre
1532 * a chance to run reply_in_callback() */
1534 if (request->rq_set != NULL)
1535 wq = &request->rq_set->set_waitq;
1537 wq = &request->rq_reply_waitq;
1540 /* Network access will complete in finite time but the HUGE
1541 * timeout lets us CWARN for visibility of sluggish NALs */
1542 lwi = LWI_TIMEOUT(cfs_time_seconds(300), NULL, NULL);
1543 rc = l_wait_event (*wq, !ptlrpc_client_receiving_reply(request), &lwi);
1547 LASSERT (rc == -ETIMEDOUT);
1548 DEBUG_REQ(D_WARNING, request, "Unexpectedly long timeout");
1552 /* caller must hold imp->imp_lock */
1553 void ptlrpc_free_committed(struct obd_import *imp)
1555 struct list_head *tmp, *saved;
1556 struct ptlrpc_request *req;
1557 struct ptlrpc_request *last_req = NULL; /* temporary fire escape */
1560 LASSERT(imp != NULL);
1562 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1565 if (imp->imp_peer_committed_transno == imp->imp_last_transno_checked &&
1566 imp->imp_generation == imp->imp_last_generation_checked) {
1567 CDEBUG(D_RPCTRACE, "%s: skip recheck: last_committed "LPU64"\n",
1568 imp->imp_obd->obd_name, imp->imp_peer_committed_transno);
1572 CDEBUG(D_RPCTRACE, "%s: committing for last_committed "LPU64" gen %d\n",
1573 imp->imp_obd->obd_name, imp->imp_peer_committed_transno,
1574 imp->imp_generation);
1575 imp->imp_last_transno_checked = imp->imp_peer_committed_transno;
1576 imp->imp_last_generation_checked = imp->imp_generation;
1578 list_for_each_safe(tmp, saved, &imp->imp_replay_list) {
1579 req = list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1581 /* XXX ok to remove when 1357 resolved - rread 05/29/03 */
1582 LASSERT(req != last_req);
1585 if (req->rq_import_generation < imp->imp_generation) {
1586 DEBUG_REQ(D_RPCTRACE, req, "free request with old gen");
1590 if (req->rq_replay) {
1591 DEBUG_REQ(D_RPCTRACE, req, "keeping (FL_REPLAY)");
1595 /* not yet committed */
1596 if (req->rq_transno > imp->imp_peer_committed_transno) {
1597 DEBUG_REQ(D_RPCTRACE, req, "stopping search");
1601 DEBUG_REQ(D_RPCTRACE, req, "commit (last_committed "LPU64")",
1602 imp->imp_peer_committed_transno);
1604 spin_lock(&req->rq_lock);
1606 spin_unlock(&req->rq_lock);
1607 if (req->rq_commit_cb != NULL)
1608 req->rq_commit_cb(req);
1609 list_del_init(&req->rq_replay_list);
1610 __ptlrpc_req_finished(req, 1);
1617 void ptlrpc_cleanup_client(struct obd_import *imp)
1624 void ptlrpc_resend_req(struct ptlrpc_request *req)
1626 DEBUG_REQ(D_HA, req, "going to resend");
1627 lustre_msg_set_handle(req->rq_reqmsg, &(struct lustre_handle){ 0 });
1628 req->rq_status = -EAGAIN;
1630 spin_lock(&req->rq_lock);
1632 req->rq_net_err = 0;
1633 req->rq_timedout = 0;
1635 __u64 old_xid = req->rq_xid;
1637 /* ensure previous bulk fails */
1638 req->rq_xid = ptlrpc_next_xid();
1639 CDEBUG(D_HA, "resend bulk old x"LPU64" new x"LPU64"\n",
1640 old_xid, req->rq_xid);
1642 ptlrpc_wake_client_req(req);
1643 spin_unlock(&req->rq_lock);
1646 /* XXX: this function and rq_status are currently unused */
1647 void ptlrpc_restart_req(struct ptlrpc_request *req)
1649 DEBUG_REQ(D_HA, req, "restarting (possibly-)completed request");
1650 req->rq_status = -ERESTARTSYS;
1652 spin_lock(&req->rq_lock);
1653 req->rq_restart = 1;
1654 req->rq_timedout = 0;
1655 ptlrpc_wake_client_req(req);
1656 spin_unlock(&req->rq_lock);
1659 static int expired_request(void *data)
1661 struct ptlrpc_request *req = data;
1664 /* some failure can suspend regular timeouts */
1665 if (ptlrpc_check_suspend())
1668 RETURN(ptlrpc_expire_one_request(req));
1671 static void interrupted_request(void *data)
1673 struct ptlrpc_request *req = data;
1674 DEBUG_REQ(D_HA, req, "request interrupted");
1675 spin_lock(&req->rq_lock);
1677 spin_unlock(&req->rq_lock);
1680 struct ptlrpc_request *ptlrpc_request_addref(struct ptlrpc_request *req)
1683 atomic_inc(&req->rq_refcount);
1687 void ptlrpc_retain_replayable_request(struct ptlrpc_request *req,
1688 struct obd_import *imp)
1690 struct list_head *tmp;
1692 LASSERT_SPIN_LOCKED(&imp->imp_lock);
1694 /* clear this for new requests that were resent as well
1695 as resent replayed requests. */
1696 lustre_msg_clear_flags(req->rq_reqmsg, MSG_RESENT);
1698 /* don't re-add requests that have been replayed */
1699 if (!list_empty(&req->rq_replay_list))
1702 lustre_msg_add_flags(req->rq_reqmsg, MSG_REPLAY);
1704 LASSERT(imp->imp_replayable);
1705 /* Balanced in ptlrpc_free_committed, usually. */
1706 ptlrpc_request_addref(req);
1707 list_for_each_prev(tmp, &imp->imp_replay_list) {
1708 struct ptlrpc_request *iter =
1709 list_entry(tmp, struct ptlrpc_request, rq_replay_list);
1711 /* We may have duplicate transnos if we create and then
1712 * open a file, or for closes retained if to match creating
1713 * opens, so use req->rq_xid as a secondary key.
1714 * (See bugs 684, 685, and 428.)
1715 * XXX no longer needed, but all opens need transnos!
1717 if (iter->rq_transno > req->rq_transno)
1720 if (iter->rq_transno == req->rq_transno) {
1721 LASSERT(iter->rq_xid != req->rq_xid);
1722 if (iter->rq_xid > req->rq_xid)
1726 list_add(&req->rq_replay_list, &iter->rq_replay_list);
1730 list_add_tail(&req->rq_replay_list, &imp->imp_replay_list);
1733 int ptlrpc_queue_wait(struct ptlrpc_request *req)
1737 struct l_wait_info lwi;
1738 struct obd_import *imp = req->rq_import;
1739 cfs_duration_t timeout = 0;
1742 LASSERT(req->rq_set == NULL);
1743 LASSERT(!req->rq_receiving_reply);
1744 atomic_inc(&imp->imp_inflight);
1746 /* for distributed debugging */
1747 lustre_msg_set_status(req->rq_reqmsg, cfs_curproc_pid());
1748 LASSERT(imp->imp_obd != NULL);
1749 CDEBUG(D_RPCTRACE, "Sending RPC pname:cluuid:pid:xid:nid:opc "
1750 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1751 imp->imp_obd->obd_uuid.uuid,
1752 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1753 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1754 lustre_msg_get_opc(req->rq_reqmsg));
1756 /* Mark phase here for a little debug help */
1757 req->rq_phase = RQ_PHASE_RPC;
1759 spin_lock(&imp->imp_lock);
1761 req->rq_import_generation = imp->imp_generation;
1762 if (ptlrpc_import_delay_req(imp, req, &rc)) {
1763 list_del(&req->rq_list);
1765 list_add_tail(&req->rq_list, &imp->imp_delayed_list);
1766 spin_unlock(&imp->imp_lock);
1768 DEBUG_REQ(D_HA, req, "\"%s\" waiting for recovery: (%s != %s)",
1770 ptlrpc_import_state_name(req->rq_send_state),
1771 ptlrpc_import_state_name(imp->imp_state));
1772 lwi = LWI_INTR(interrupted_request, req);
1773 rc = l_wait_event(req->rq_reply_waitq,
1774 (req->rq_send_state == imp->imp_state ||
1775 req->rq_err || req->rq_intr),
1777 DEBUG_REQ(D_HA, req, "\"%s\" awake: (%s == %s or %d/%d == 1)",
1779 ptlrpc_import_state_name(imp->imp_state),
1780 ptlrpc_import_state_name(req->rq_send_state),
1781 req->rq_err, req->rq_intr);
1783 spin_lock(&imp->imp_lock);
1784 list_del_init(&req->rq_list);
1789 else if (req->rq_intr) {
1792 else if (req->rq_no_resend) {
1793 spin_unlock(&imp->imp_lock);
1794 GOTO(out, rc = -ETIMEDOUT);
1802 list_del_init(&req->rq_list);
1803 spin_unlock(&imp->imp_lock);
1804 req->rq_status = rc; // XXX this ok?
1808 if (req->rq_resend) {
1809 lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
1811 if (req->rq_bulk != NULL) {
1812 ptlrpc_unregister_bulk (req);
1814 /* bulk requests are supposed to be
1815 * idempotent, so we are free to bump the xid
1816 * here, which we need to do before
1817 * registering the bulk again (bug 6371).
1818 * print the old xid first for sanity.
1820 DEBUG_REQ(D_HA, req, "bumping xid for bulk: ");
1821 req->rq_xid = ptlrpc_next_xid();
1824 DEBUG_REQ(D_HA, req, "resending: ");
1827 /* XXX this is the same as ptlrpc_set_wait */
1828 LASSERT(list_empty(&req->rq_list));
1829 list_add_tail(&req->rq_list, &imp->imp_sending_list);
1830 spin_unlock(&imp->imp_lock);
1832 rc = sptlrpc_req_refresh_ctx(req, 0);
1835 /* we got fatal ctx refresh error, directly jump out
1836 * thus we can pass back the actual error code.
1838 spin_lock(&imp->imp_lock);
1839 list_del_init(&req->rq_list);
1840 spin_unlock(&imp->imp_lock);
1842 CERROR("Failed to refresh ctx of req %p: %d\n", req, rc);
1845 /* simulating we got error during send rpc */
1849 rc = ptl_send_rpc(req, 0);
1851 DEBUG_REQ(D_HA, req, "send failed (%d); recovering", rc);
1854 timeout = cfs_timeout_cap(cfs_time_seconds(req->rq_timeout));
1855 DEBUG_REQ(D_NET, req,
1856 "-- sleeping for "CFS_DURATION_T" jiffies", timeout);
1859 lwi = LWI_TIMEOUT_INTR(timeout, expired_request, interrupted_request,
1861 rc = l_wait_event(req->rq_reply_waitq, ptlrpc_check_reply(req), &lwi);
1862 if (rc == -ETIMEDOUT && ptlrpc_check_and_wait_suspend(req))
1866 CDEBUG(D_RPCTRACE, "Completed RPC pname:cluuid:pid:xid:nid:opc "
1867 "%s:%s:%d:"LPU64":%s:%d\n", cfs_curproc_comm(),
1868 imp->imp_obd->obd_uuid.uuid,
1869 lustre_msg_get_status(req->rq_reqmsg), req->rq_xid,
1870 libcfs_nid2str(imp->imp_connection->c_peer.nid),
1871 lustre_msg_get_opc(req->rq_reqmsg));
1873 spin_lock(&imp->imp_lock);
1874 list_del_init(&req->rq_list);
1875 spin_unlock(&imp->imp_lock);
1877 /* If the reply was received normally, this just grabs the spinlock
1878 * (ensuring the reply callback has returned), sees that
1879 * req->rq_receiving_reply is clear and returns. */
1880 ptlrpc_unregister_reply (req);
1883 GOTO(out, rc = -EIO);
1885 /* Resend if we need to, unless we were interrupted. */
1886 if (req->rq_resend && !req->rq_intr) {
1887 /* ...unless we were specifically told otherwise. */
1888 if (req->rq_no_resend)
1889 GOTO(out, rc = -ETIMEDOUT);
1890 spin_lock(&imp->imp_lock);
1895 /* Should only be interrupted if we timed out. */
1896 if (!req->rq_timedout)
1897 DEBUG_REQ(D_ERROR, req,
1898 "rq_intr set but rq_timedout not");
1899 GOTO(out, rc = -EINTR);
1902 if (req->rq_timedout) { /* non-recoverable timeout */
1903 GOTO(out, rc = -ETIMEDOUT);
1906 if (!req->rq_replied) {
1907 /* How can this be? -eeb */
1908 DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
1910 GOTO(out, rc = req->rq_status);
1913 rc = after_reply(req);
1914 /* NB may return +ve success rc */
1915 if (req->rq_resend) {
1916 spin_lock(&imp->imp_lock);
1921 if (req->rq_bulk != NULL) {
1923 /* success so far. Note that anything going wrong
1924 * with bulk now, is EXTREMELY strange, since the
1925 * server must have believed that the bulk
1926 * tranferred OK before she replied with success to
1928 lwi = LWI_TIMEOUT(timeout, NULL, NULL);
1929 brc = l_wait_event(req->rq_reply_waitq,
1930 !ptlrpc_bulk_active(req->rq_bulk),
1932 LASSERT(brc == 0 || brc == -ETIMEDOUT);
1934 LASSERT(brc == -ETIMEDOUT);
1935 DEBUG_REQ(D_ERROR, req, "bulk timed out");
1937 } else if (!req->rq_bulk->bd_success) {
1938 DEBUG_REQ(D_ERROR, req, "bulk transfer failed");
1943 ptlrpc_unregister_bulk (req);
1946 LASSERT(!req->rq_receiving_reply);
1947 req->rq_phase = RQ_PHASE_INTERPRET;
1949 atomic_dec(&imp->imp_inflight);
1950 cfs_waitq_signal(&imp->imp_recovery_waitq);
1954 struct ptlrpc_replay_async_args {
1956 int praa_old_status;
1959 static int ptlrpc_replay_interpret(struct ptlrpc_request *req,
1960 void * data, int rc)
1962 struct ptlrpc_replay_async_args *aa = data;
1963 struct obd_import *imp = req->rq_import;
1966 atomic_dec(&imp->imp_replay_inflight);
1968 if (!req->rq_replied) {
1969 CERROR("request replay timed out, restarting recovery\n");
1970 GOTO(out, rc = -ETIMEDOUT);
1973 if (lustre_msg_get_type(req->rq_repmsg) == PTL_RPC_MSG_ERR &&
1974 (lustre_msg_get_status(req->rq_repmsg) == -ENOTCONN ||
1975 lustre_msg_get_status(req->rq_repmsg) == -ENODEV))
1976 GOTO(out, rc = lustre_msg_get_status(req->rq_repmsg));
1978 /* The transno had better not change over replay. */
1979 LASSERT(lustre_msg_get_transno(req->rq_reqmsg) ==
1980 lustre_msg_get_transno(req->rq_repmsg));
1982 DEBUG_REQ(D_HA, req, "got rep");
1984 /* let the callback do fixups, possibly including in the request */
1985 if (req->rq_replay_cb)
1986 req->rq_replay_cb(req);
1988 if (req->rq_replied &&
1989 lustre_msg_get_status(req->rq_repmsg) != aa->praa_old_status) {
1990 DEBUG_REQ(D_ERROR, req, "status %d, old was %d",
1991 lustre_msg_get_status(req->rq_repmsg),
1992 aa->praa_old_status);
1994 /* Put it back for re-replay. */
1995 lustre_msg_set_status(req->rq_repmsg, aa->praa_old_status);
1999 * Errors while replay can set transno to 0, but
2000 * imp_last_replay_transno shouldn't be set to 0 anyway
2002 if (req->rq_transno > 0) {
2003 spin_lock(&imp->imp_lock);
2004 LASSERT(req->rq_transno <= imp->imp_last_replay_transno);
2005 imp->imp_last_replay_transno = req->rq_transno;
2006 spin_unlock(&imp->imp_lock);
2008 CERROR("Transno is 0 during replay!\n");
2009 /* continue with recovery */
2010 rc = ptlrpc_import_recovery_state_machine(imp);
2012 req->rq_send_state = aa->praa_old_state;
2015 /* this replay failed, so restart recovery */
2016 ptlrpc_connect_import(imp, NULL);
2021 int ptlrpc_replay_req(struct ptlrpc_request *req)
2023 struct ptlrpc_replay_async_args *aa;
2026 LASSERT(req->rq_import->imp_state == LUSTRE_IMP_REPLAY);
2028 /* Not handling automatic bulk replay yet (or ever?) */
2029 LASSERT(req->rq_bulk == NULL);
2031 DEBUG_REQ(D_HA, req, "REPLAY");
2033 LASSERT (sizeof (*aa) <= sizeof (req->rq_async_args));
2034 aa = (struct ptlrpc_replay_async_args *)&req->rq_async_args;
2035 memset(aa, 0, sizeof *aa);
2037 /* Prepare request to be resent with ptlrpcd */
2038 aa->praa_old_state = req->rq_send_state;
2039 req->rq_send_state = LUSTRE_IMP_REPLAY;
2040 req->rq_phase = RQ_PHASE_NEW;
2041 aa->praa_old_status = lustre_msg_get_status(req->rq_repmsg);
2044 req->rq_interpret_reply = ptlrpc_replay_interpret;
2045 atomic_inc(&req->rq_import->imp_replay_inflight);
2046 ptlrpc_request_addref(req); /* ptlrpcd needs a ref */
2048 ptlrpcd_add_req(req);
2052 void ptlrpc_abort_inflight(struct obd_import *imp)
2054 struct list_head *tmp, *n;
2057 /* Make sure that no new requests get processed for this import.
2058 * ptlrpc_{queue,set}_wait must (and does) hold imp_lock while testing
2059 * this flag and then putting requests on sending_list or delayed_list.
2061 spin_lock(&imp->imp_lock);
2063 /* XXX locking? Maybe we should remove each request with the list
2064 * locked? Also, how do we know if the requests on the list are
2065 * being freed at this time?
2067 list_for_each_safe(tmp, n, &imp->imp_sending_list) {
2068 struct ptlrpc_request *req =
2069 list_entry(tmp, struct ptlrpc_request, rq_list);
2071 DEBUG_REQ(D_RPCTRACE, req, "inflight");
2073 spin_lock (&req->rq_lock);
2074 if (req->rq_import_generation < imp->imp_generation) {
2076 ptlrpc_wake_client_req(req);
2078 spin_unlock (&req->rq_lock);
2081 list_for_each_safe(tmp, n, &imp->imp_delayed_list) {
2082 struct ptlrpc_request *req =
2083 list_entry(tmp, struct ptlrpc_request, rq_list);
2085 DEBUG_REQ(D_RPCTRACE, req, "aborting waiting req");
2087 spin_lock (&req->rq_lock);
2088 if (req->rq_import_generation < imp->imp_generation) {
2090 ptlrpc_wake_client_req(req);
2092 spin_unlock (&req->rq_lock);
2095 /* Last chance to free reqs left on the replay list, but we
2096 * will still leak reqs that haven't committed. */
2097 if (imp->imp_replayable)
2098 ptlrpc_free_committed(imp);
2100 spin_unlock(&imp->imp_lock);
2105 static __u64 ptlrpc_last_xid = 0;
2106 spinlock_t ptlrpc_last_xid_lock;
2108 __u64 ptlrpc_next_xid(void)
2111 spin_lock(&ptlrpc_last_xid_lock);
2112 tmp = ++ptlrpc_last_xid;
2113 spin_unlock(&ptlrpc_last_xid_lock);
2117 __u64 ptlrpc_sample_next_xid(void)
2120 spin_lock(&ptlrpc_last_xid_lock);
2121 tmp = ptlrpc_last_xid + 1;
2122 spin_unlock(&ptlrpc_last_xid_lock);
2125 EXPORT_SYMBOL(ptlrpc_sample_next_xid);