1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/ost/ost_handler.c
38 * Author: Peter J. Braam <braam@clusterfs.com>
39 * Author: Phil Schwan <phil@clusterfs.com>
43 # define EXPORT_SYMTAB
45 #define DEBUG_SUBSYSTEM S_OST
47 #include <linux/module.h>
48 #include <obd_cksum.h>
50 #include <lustre_net.h>
51 #include <lustre_dlm.h>
52 #include <lustre_export.h>
53 #include <lustre_debug.h>
54 #include <linux/init.h>
55 #include <lprocfs_status.h>
56 #include <libcfs/list.h>
57 #include <lustre_quota.h>
58 #include "ost_internal.h"
60 static int oss_num_threads;
61 CFS_MODULE_PARM(oss_num_threads, "i", int, 0444,
62 "number of OSS service threads to start");
64 static int ost_num_threads;
65 CFS_MODULE_PARM(ost_num_threads, "i", int, 0444,
66 "number of OST service threads to start (deprecated)");
68 static int oss_num_create_threads;
69 CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444,
70 "number of OSS create threads to start");
73 * Do not return server-side uid/gid to remote client
75 static void ost_drop_id(struct obd_export *exp, struct obdo *oa)
77 if (exp_connect_rmtclient(exp)) {
80 oa->o_valid &= ~(OBD_MD_FLUID | OBD_MD_FLGID);
85 * Validate oa from client.
86 * 1. If the request comes from 1.8 clients, it will reset o_seq with MDT0.
87 * 2. If the request comes from 2.0 clients, currently only RSVD seq and IDIF
89 * a. for single MDS seq = FID_SEQ_OST_MDT0,
90 * b. for CMD, seq = FID_SEQ_OST_MDT0, FID_SEQ_OST_MDT1 - FID_SEQ_OST_MAX
92 static int ost_validate_obdo(struct obd_export *exp, struct obdo *oa,
93 struct obd_ioobj *ioobj)
95 if (oa != NULL && (!(oa->o_valid & OBD_MD_FLGROUP) ||
96 !(exp->exp_connect_flags & OBD_CONNECT_FULL20))) {
97 oa->o_seq = FID_SEQ_OST_MDT0;
99 ioobj->ioo_seq = FID_SEQ_OST_MDT0;
100 /* remove fid_seq_is_rsvd() after FID-on-OST allows SEQ > 9 */
101 } else if (oa == NULL ||
102 !(fid_seq_is_rsvd(oa->o_seq) || fid_seq_is_idif(oa->o_seq))) {
103 CERROR("%s: client %s sent invalid object "POSTID"\n",
104 exp->exp_obd->obd_name, obd_export_nid2str(exp),
105 oa ? oa->o_id : -1, oa ? oa->o_seq : -1);
108 obdo_from_ostid(oa, &oa->o_oi);
110 ioobj_from_obdo(ioobj, oa);
114 void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
116 struct oti_req_ack_lock *ack_lock;
122 if (req->rq_repmsg) {
123 __u64 versions[PTLRPC_NUM_VERSIONS] = { 0 };
124 lustre_msg_set_transno(req->rq_repmsg, oti->oti_transno);
125 versions[0] = oti->oti_pre_version;
126 lustre_msg_set_versions(req->rq_repmsg, versions);
128 req->rq_transno = oti->oti_transno;
130 /* XXX 4 == entries in oti_ack_locks??? */
131 for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) {
134 /* XXX not even calling target_send_reply in some cases... */
135 ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode, 0);
139 static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req,
140 struct obd_trans_info *oti)
142 struct ost_body *body, *repbody;
143 struct lustre_capa *capa = NULL;
147 /* Get the request body */
148 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
152 if (body->oa.o_id == 0)
155 rc = ost_validate_obdo(exp, &body->oa, NULL);
159 /* If there's a DLM request, cancel the locks mentioned in it*/
160 if (req_capsule_field_present(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT)) {
161 struct ldlm_request *dlm;
163 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
166 ldlm_request_cancel(req, dlm, 0);
169 /* If there's a capability, get it */
170 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
171 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
173 CERROR("Missing capability for OST DESTROY");
178 /* Prepare the reply */
179 rc = req_capsule_server_pack(&req->rq_pill);
183 /* Get the log cancellation cookie */
184 if (body->oa.o_valid & OBD_MD_FLCOOKIE)
185 oti->oti_logcookies = &body->oa.o_lcookie;
187 /* Finish the reply */
188 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
189 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
191 /* Do the destroy and set the reply status accordingly */
192 req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL, capa);
197 * Helper function for getting server side [start, start+count] DLM lock
198 * if asked by client.
200 static int ost_lock_get(struct obd_export *exp, struct obdo *oa,
201 __u64 start, __u64 count, struct lustre_handle *lh,
204 struct ldlm_res_id res_id;
205 ldlm_policy_data_t policy;
206 __u64 end = start + count;
210 LASSERT(!lustre_handle_is_used(lh));
211 /* o_id and o_gr are used for localizing resource, if client miss to set
212 * them, do not trigger ASSERTION. */
213 if (unlikely((oa->o_valid & (OBD_MD_FLID | OBD_MD_FLGROUP)) !=
214 (OBD_MD_FLID | OBD_MD_FLGROUP)))
217 if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
218 !(oa->o_flags & OBD_FL_SRVLOCK))
221 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
222 CDEBUG(D_INODE, "OST-side extent lock.\n");
224 policy.l_extent.start = start & CFS_PAGE_MASK;
226 /* If ->o_blocks is EOF it means "lock till the end of the
227 * file". Otherwise, it's size of a hole being punched (in bytes) */
228 if (count == OBD_OBJECT_EOF || end < start)
229 policy.l_extent.end = OBD_OBJECT_EOF;
231 policy.l_extent.end = end | ~CFS_PAGE_MASK;
233 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
234 LDLM_EXTENT, &policy, mode, &flags,
235 ldlm_blocking_ast, ldlm_completion_ast,
236 ldlm_glimpse_ast, NULL, 0, NULL, lh));
239 /* Helper function: release lock, if any. */
240 static void ost_lock_put(struct obd_export *exp,
241 struct lustre_handle *lh, int mode)
244 if (lustre_handle_is_used(lh))
245 ldlm_lock_decref(lh, mode);
249 static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
251 struct ost_body *body, *repbody;
252 struct obd_info *oinfo;
253 struct lustre_handle lh = { 0 };
254 struct lustre_capa *capa = NULL;
258 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
262 rc = ost_validate_obdo(exp, &body->oa, NULL);
266 rc = req_capsule_server_pack(&req->rq_pill);
270 rc = ost_lock_get(exp, &body->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0);
274 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
275 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
277 CERROR("Missing capability for OST GETATTR");
278 GOTO(unlock, rc = -EFAULT);
282 OBD_ALLOC_PTR(oinfo);
284 GOTO(unlock, rc = -ENOMEM);
285 oinfo->oi_oa = &body->oa;
286 oinfo->oi_capa = capa;
288 req->rq_status = obd_getattr(exp, oinfo);
292 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
293 repbody->oa = body->oa;
294 ost_drop_id(exp, &repbody->oa);
297 ost_lock_put(exp, &lh, LCK_PR);
302 static int ost_statfs(struct ptlrpc_request *req)
304 struct obd_statfs *osfs;
308 rc = req_capsule_server_pack(&req->rq_pill);
312 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
314 req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs,
315 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
317 if (req->rq_status != 0)
318 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
323 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
324 struct obd_trans_info *oti)
326 struct ost_body *body, *repbody;
330 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
334 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
338 rc = req_capsule_server_pack(&req->rq_pill);
342 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
343 repbody->oa = body->oa;
344 oti->oti_logcookies = &body->oa.o_lcookie;
346 req->rq_status = obd_create(exp, &repbody->oa, NULL, oti);
347 //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
351 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
352 struct obd_trans_info *oti)
354 struct ost_body *body, *repbody;
356 struct lustre_handle lh = {0,};
359 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
360 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
362 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
366 rc = ost_validate_obdo(exp, &body->oa, NULL);
370 if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
371 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
374 rc = req_capsule_server_pack(&req->rq_pill);
378 /* standard truncate optimization: if file body is completely
379 * destroyed, don't send data back to the server. */
380 if (body->oa.o_size == 0)
381 flags |= LDLM_AST_DISCARD_DATA;
383 rc = ost_lock_get(exp, &body->oa, body->oa.o_size, body->oa.o_blocks,
386 struct obd_info *oinfo;
387 struct lustre_capa *capa = NULL;
389 if (body->oa.o_valid & OBD_MD_FLFLAGS &&
390 body->oa.o_flags == OBD_FL_SRVLOCK)
392 * If OBD_FL_SRVLOCK is the only bit set in
393 * ->o_flags, clear OBD_MD_FLFLAGS to avoid falling
394 * through filter_setattr() to filter_iocontrol().
396 body->oa.o_valid &= ~OBD_MD_FLFLAGS;
398 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
399 capa = req_capsule_client_get(&req->rq_pill,
402 CERROR("Missing capability for OST PUNCH");
403 GOTO(unlock, rc = -EFAULT);
407 OBD_ALLOC_PTR(oinfo);
409 GOTO(unlock, rc = -ENOMEM);
410 oinfo->oi_oa = &body->oa;
411 oinfo->oi_policy.l_extent.start = oinfo->oi_oa->o_size;
412 oinfo->oi_policy.l_extent.end = oinfo->oi_oa->o_blocks;
413 oinfo->oi_capa = capa;
415 req->rq_status = obd_punch(exp, oinfo, oti, NULL);
418 ost_lock_put(exp, &lh, LCK_PW);
421 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
422 repbody->oa = body->oa;
423 ost_drop_id(exp, &repbody->oa);
427 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
429 struct ost_body *body, *repbody;
430 struct obd_info *oinfo;
431 struct lustre_capa *capa = NULL;
435 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
439 rc = ost_validate_obdo(exp, &body->oa, NULL);
443 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
444 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
446 CERROR("Missing capability for OST SYNC");
451 rc = req_capsule_server_pack(&req->rq_pill);
455 OBD_ALLOC_PTR(oinfo);
459 oinfo->oi_oa = &body->oa;
460 oinfo->oi_capa = capa;
461 req->rq_status = obd_sync(exp, oinfo, body->oa.o_size,
462 body->oa.o_blocks, NULL);
465 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
466 repbody->oa = body->oa;
467 ost_drop_id(exp, &repbody->oa);
471 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
472 struct obd_trans_info *oti)
474 struct ost_body *body, *repbody;
475 struct obd_info *oinfo;
476 struct lustre_capa *capa = NULL;
480 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
484 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
488 rc = req_capsule_server_pack(&req->rq_pill);
492 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
493 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
495 CERROR("Missing capability for OST SETATTR");
500 OBD_ALLOC_PTR(oinfo);
503 oinfo->oi_oa = &body->oa;
504 oinfo->oi_capa = capa;
506 req->rq_status = obd_setattr(exp, oinfo, oti);
510 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
511 repbody->oa = body->oa;
512 ost_drop_id(exp, &repbody->oa);
516 static int ost_bulk_timeout(void *data)
519 /* We don't fail the connection here, because having the export
520 * killed makes the (vital) call to commitrw very sad.
525 static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc,
526 cksum_type_t cksum_type)
531 cksum = init_checksum(cksum_type);
532 for (i = 0; i < desc->bd_iov_count; i++) {
533 struct page *page = desc->bd_iov[i].kiov_page;
534 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
535 char *ptr = kmap(page) + off;
536 int len = desc->bd_iov[i].kiov_len;
538 /* corrupt the data before we compute the checksum, to
539 * simulate a client->OST data error */
540 if (i == 0 && opc == OST_WRITE &&
541 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE))
542 memcpy(ptr, "bad3", min(4, len));
543 cksum = compute_checksum(cksum, ptr, len, cksum_type);
544 /* corrupt the data after we compute the checksum, to
545 * simulate an OST->client data error */
546 if (i == 0 && opc == OST_READ &&
547 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
548 memcpy(ptr, "bad4", min(4, len));
549 /* nobody should use corrupted page again */
550 ClearPageUptodate(page);
558 static int ost_brw_lock_get(int mode, struct obd_export *exp,
559 struct obd_ioobj *obj, struct niobuf_remote *nb,
560 struct lustre_handle *lh)
563 int nrbufs = obj->ioo_bufcnt;
564 struct ldlm_res_id res_id;
565 ldlm_policy_data_t policy;
569 osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
570 LASSERT(mode == LCK_PR || mode == LCK_PW);
571 LASSERT(!lustre_handle_is_used(lh));
573 if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
576 for (i = 1; i < nrbufs; i ++)
577 if ((nb[0].flags & OBD_BRW_SRVLOCK) !=
578 (nb[i].flags & OBD_BRW_SRVLOCK))
581 policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
582 policy.l_extent.end = (nb[nrbufs - 1].offset +
583 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
585 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
586 LDLM_EXTENT, &policy, mode, &flags,
587 ldlm_blocking_ast, ldlm_completion_ast,
588 ldlm_glimpse_ast, NULL, 0, NULL, lh));
591 static void ost_brw_lock_put(int mode,
592 struct obd_ioobj *obj, struct niobuf_remote *niob,
593 struct lustre_handle *lh)
596 LASSERT(mode == LCK_PR || mode == LCK_PW);
597 LASSERT((obj->ioo_bufcnt > 0 && (niob[0].flags & OBD_BRW_SRVLOCK)) ==
598 lustre_handle_is_used(lh));
599 if (lustre_handle_is_used(lh))
600 ldlm_lock_decref(lh, mode);
604 struct ost_prolong_data {
605 struct obd_export *opd_exp;
606 ldlm_policy_data_t opd_policy;
608 ldlm_mode_t opd_mode;
613 static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data)
615 struct ost_prolong_data *opd = data;
617 LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
619 if (lock->l_req_mode != lock->l_granted_mode) {
620 /* scan granted locks only */
621 return LDLM_ITER_STOP;
624 if (lock->l_export != opd->opd_exp) {
625 /* prolong locks only for given client */
626 return LDLM_ITER_CONTINUE;
629 if (!(lock->l_granted_mode & opd->opd_mode)) {
630 /* we aren't interesting in all type of locks */
631 return LDLM_ITER_CONTINUE;
634 if (lock->l_policy_data.l_extent.end < opd->opd_policy.l_extent.start ||
635 lock->l_policy_data.l_extent.start > opd->opd_policy.l_extent.end) {
636 /* the request doesn't cross the lock, skip it */
637 return LDLM_ITER_CONTINUE;
640 /* Fill the obdo with the matched lock handle.
641 * XXX: it is possible in some cases the IO RPC is covered by several
642 * locks, even for the write case, so it may need to be a lock list. */
643 if (opd->opd_oa && !(opd->opd_oa->o_valid & OBD_MD_FLHANDLE)) {
644 opd->opd_oa->o_handle.cookie = lock->l_handle.h_cookie;
645 opd->opd_oa->o_valid |= OBD_MD_FLHANDLE;
648 if (!(lock->l_flags & LDLM_FL_AST_SENT)) {
649 /* ignore locks not being cancelled */
650 return LDLM_ITER_CONTINUE;
653 CDEBUG(D_DLMTRACE,"refresh lock: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
654 lock->l_resource->lr_name.name[0],
655 lock->l_resource->lr_name.name[1],
656 opd->opd_policy.l_extent.start, opd->opd_policy.l_extent.end);
657 /* OK. this is a possible lock the user holds doing I/O
658 * let's refresh eviction timer for it */
659 ldlm_refresh_waiting_lock(lock, opd->opd_timeout);
660 opd->opd_lock_match = 1;
662 return LDLM_ITER_CONTINUE;
665 static int ost_rw_prolong_locks(struct ptlrpc_request *req, struct obd_ioobj *obj,
666 struct niobuf_remote *nb, struct obdo *oa,
669 struct ldlm_res_id res_id;
670 int nrbufs = obj->ioo_bufcnt;
671 struct ost_prolong_data opd = { 0 };
674 osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
677 opd.opd_exp = req->rq_export;
678 opd.opd_policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
679 opd.opd_policy.l_extent.end = (nb[nrbufs - 1].offset +
680 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
682 /* prolong locks for the current service time of the corresponding
683 * portal (= OST_IO_PORTAL) */
684 opd.opd_timeout = AT_OFF ? obd_timeout / 2:
685 max(at_est2timeout(at_get(&req->rq_rqbd->
686 rqbd_service->srv_at_estimate)), ldlm_timeout);
688 CDEBUG(D_INFO,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
689 res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
690 opd.opd_policy.l_extent.end);
692 if (oa->o_valid & OBD_MD_FLHANDLE) {
693 struct ldlm_lock *lock;
695 lock = ldlm_handle2lock(&oa->o_handle);
697 ost_prolong_locks_iter(lock, &opd);
698 if (opd.opd_lock_match) {
703 /* Check if the lock covers the whole IO region,
704 * otherwise iterate through the resource. */
705 if (lock->l_policy_data.l_extent.end >=
706 opd.opd_policy.l_extent.end &&
707 lock->l_policy_data.l_extent.start <=
708 opd.opd_policy.l_extent.start) {
717 ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id,
718 ost_prolong_locks_iter, &opd);
719 RETURN(opd.opd_lock_match);
722 /* Allocate thread local buffers if needed */
723 static struct ost_thread_local_cache *ost_tls_get(struct ptlrpc_request *r)
725 struct ost_thread_local_cache *tls =
726 (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
728 /* In normal mode of operation an I/O request is serviced only
729 * by ll_ost_io threads each of them has own tls buffers allocated by
731 * During recovery, an I/O request may be queued until any of the ost
732 * service threads process it. Not necessary it should be one of
733 * ll_ost_io threads. In that case we dynamically allocating tls
734 * buffers for the request service time. */
735 if (unlikely(tls == NULL)) {
736 LASSERT(r->rq_export->exp_in_recovery);
740 r->rq_svc_thread->t_data = tls;
746 /* Free thread local buffers if they were allocated only for servicing
747 * this one request */
748 static void ost_tls_put(struct ptlrpc_request *r)
750 struct ost_thread_local_cache *tls =
751 (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
753 if (unlikely(tls->temporary)) {
755 r->rq_svc_thread->t_data = NULL;
759 static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti)
761 struct ptlrpc_bulk_desc *desc = NULL;
762 struct obd_export *exp = req->rq_export;
763 struct niobuf_remote *remote_nb;
764 struct niobuf_local *local_nb;
765 struct obd_ioobj *ioo;
766 struct ost_body *body, *repbody;
767 struct lustre_capa *capa = NULL;
768 struct l_wait_info lwi;
769 struct lustre_handle lockh = { 0 };
770 int niocount, npages, nob = 0, rc, i;
772 struct ost_thread_local_cache *tls;
775 req->rq_bulk_read = 1;
777 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
778 GOTO(out, rc = -EIO);
780 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
782 /* Check if there is eviction in progress, and if so, wait for it to
784 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
785 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
786 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
787 !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
791 GOTO(out, rc = -ENOTCONN);
793 /* ost_body, ioobj & noibuf_remote are verified and swabbed in
794 * ost_rw_hpreq_check(). */
795 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
797 GOTO(out, rc = -EFAULT);
800 * A req_capsule_X_get_array(pill, field, ptr_to_element_count) function
801 * would be useful here and wherever we get &RMF_OBD_IOOBJ and
802 * &RMF_NIOBUF_REMOTE.
804 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
806 GOTO(out, rc = -EFAULT);
808 rc = ost_validate_obdo(exp, &body->oa, ioo);
812 niocount = ioo->ioo_bufcnt;
813 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
814 if (remote_nb == NULL)
815 GOTO(out, rc = -EFAULT);
817 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
818 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
820 CERROR("Missing capability for OST BRW READ");
821 GOTO(out, rc = -EFAULT);
825 rc = req_capsule_server_pack(&req->rq_pill);
829 tls = ost_tls_get(req);
831 GOTO(out_bulk, rc = -ENOMEM);
832 local_nb = tls->local;
834 rc = ost_brw_lock_get(LCK_PR, exp, ioo, remote_nb, &lockh);
839 * If getting the lock took more time than
840 * client was willing to wait, drop it. b=11330
842 if (cfs_time_current_sec() > req->rq_deadline ||
843 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
845 CERROR("Dropping timed-out read from %s because locking"
846 "object "LPX64" took %ld seconds (limit was %ld).\n",
847 libcfs_id2str(req->rq_peer), ioo->ioo_id,
848 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
849 req->rq_deadline - req->rq_arrival_time.tv_sec);
850 GOTO(out_lock, rc = -ETIMEDOUT);
853 npages = OST_THREAD_POOL_SIZE;
854 rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1, ioo,
855 remote_nb, &npages, local_nb, oti, capa);
859 desc = ptlrpc_prep_bulk_exp(req, npages,
860 BULK_PUT_SOURCE, OST_BULK_PORTAL);
862 GOTO(out_commitrw, rc = -ENOMEM);
864 if (!lustre_handle_is_used(&lockh))
865 /* no needs to try to prolong lock if server is asked
866 * to handle locking (= OBD_BRW_SRVLOCK) */
867 ost_rw_prolong_locks(req, ioo, remote_nb, &body->oa,
871 for (i = 0; i < npages; i++) {
872 int page_rc = local_nb[i].rc;
874 if (page_rc < 0) { /* error */
880 if (page_rc != 0) { /* some data! */
881 LASSERT (local_nb[i].page != NULL);
882 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
883 local_nb[i].offset & ~CFS_PAGE_MASK,
887 if (page_rc != local_nb[i].len) { /* short read */
888 /* All subsequent pages should be 0 */
890 LASSERT(local_nb[i].rc == 0);
895 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
896 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
898 if (body->oa.o_valid & OBD_MD_FLFLAGS)
899 cksum_type = cksum_type_unpack(body->oa.o_flags);
900 body->oa.o_flags = cksum_type_pack(cksum_type);
901 body->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
902 body->oa.o_cksum = ost_checksum_bulk(desc, OST_READ, cksum_type);
903 CDEBUG(D_PAGE,"checksum at read origin: %x\n",body->oa.o_cksum);
905 body->oa.o_valid = 0;
907 /* We're finishing using body->oa as an input variable */
909 /* Check if client was evicted while we were doing i/o before touching
912 /* Check if there is eviction in progress, and if so, wait for
914 if (unlikely(cfs_atomic_read(&exp->exp_obd->
915 obd_evict_inprogress))) {
916 lwi = LWI_INTR(NULL, NULL);
917 rc = l_wait_event(exp->exp_obd->
918 obd_evict_inprogress_waitq,
919 !cfs_atomic_read(&exp->exp_obd->
920 obd_evict_inprogress),
923 /* Check if client was evicted or tried to reconnect already */
924 if (exp->exp_failed || exp->exp_abort_active_req)
927 rc = sptlrpc_svc_wrap_bulk(req, desc);
929 rc = ptlrpc_start_bulk_transfer(desc);
933 time_t start = cfs_time_current_sec();
935 long timeoutl = req->rq_deadline -
936 cfs_time_current_sec();
937 cfs_duration_t timeout = timeoutl <= 0 ?
938 CFS_TICK : cfs_time_seconds(timeoutl);
939 lwi = LWI_TIMEOUT_INTERVAL(timeout,
943 rc = l_wait_event(desc->bd_waitq,
944 !ptlrpc_server_bulk_active(desc) ||
946 exp->exp_abort_active_req,
948 LASSERT(rc == 0 || rc == -ETIMEDOUT);
949 /* Wait again if we changed deadline */
950 } while ((rc == -ETIMEDOUT) &&
951 (req->rq_deadline > cfs_time_current_sec()));
953 if (rc == -ETIMEDOUT) {
954 DEBUG_REQ(D_ERROR, req,
955 "timeout on bulk PUT after %ld%+lds",
956 req->rq_deadline - start,
957 cfs_time_current_sec() -
959 ptlrpc_abort_bulk(desc);
960 } else if (exp->exp_failed) {
961 DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT");
963 ptlrpc_abort_bulk(desc);
964 } else if (exp->exp_abort_active_req) {
965 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk PUT");
966 /* we don't reply anyway */
968 ptlrpc_abort_bulk(desc);
969 } else if (!desc->bd_success ||
970 desc->bd_nob_transferred != desc->bd_nob) {
971 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
973 "truncated" : "network error on",
974 desc->bd_nob_transferred,
976 /* XXX should this be a different errno? */
980 DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d", rc);
986 /* Must commit after prep above in all cases */
987 rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1, ioo,
988 remote_nb, npages, local_nb, oti, rc);
991 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
992 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
993 ost_drop_id(exp, &repbody->oa);
997 ost_brw_lock_put(LCK_PR, ioo, remote_nb, &lockh);
1002 ptlrpc_free_bulk(desc);
1006 req->rq_status = nob;
1007 ptlrpc_lprocfs_brw(req, nob);
1008 target_committed_to_req(req);
1010 } else if (!no_reply) {
1011 /* Only reply if there was no comms problem with bulk */
1012 target_committed_to_req(req);
1013 req->rq_status = rc;
1016 /* reply out callback would free */
1017 ptlrpc_req_drop_rs(req);
1018 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
1019 "client will retry\n",
1020 exp->exp_obd->obd_name,
1021 exp->exp_client_uuid.uuid,
1022 exp->exp_connection->c_remote_uuid.uuid,
1023 libcfs_id2str(req->rq_peer));
1029 static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
1031 struct ptlrpc_bulk_desc *desc = NULL;
1032 struct obd_export *exp = req->rq_export;
1033 struct niobuf_remote *remote_nb;
1034 struct niobuf_local *local_nb;
1035 struct obd_ioobj *ioo;
1036 struct ost_body *body, *repbody;
1037 struct l_wait_info lwi;
1038 struct lustre_handle lockh = {0};
1039 struct lustre_capa *capa = NULL;
1041 int objcount, niocount, npages;
1043 obd_count client_cksum = 0, server_cksum = 0;
1044 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
1045 int no_reply = 0, mmap = 0;
1046 __u32 o_uid = 0, o_gid = 0;
1047 struct ost_thread_local_cache *tls;
1050 req->rq_bulk_write = 1;
1052 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
1053 GOTO(out, rc = -EIO);
1054 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
1055 GOTO(out, rc = -EFAULT);
1057 /* pause before transaction has been started */
1058 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
1060 /* Check if there is eviction in progress, and if so, wait for it to
1062 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
1063 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
1064 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
1065 !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
1068 if (exp->exp_failed)
1069 GOTO(out, rc = -ENOTCONN);
1071 /* ost_body, ioobj & noibuf_remote are verified and swabbed in
1072 * ost_rw_hpreq_check(). */
1073 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1075 GOTO(out, rc = -EFAULT);
1077 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1078 RCL_CLIENT) / sizeof(*ioo);
1079 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1081 GOTO(out, rc = -EFAULT);
1083 rc = ost_validate_obdo(exp, &body->oa, ioo);
1087 for (niocount = i = 0; i < objcount; i++)
1088 niocount += ioo[i].ioo_bufcnt;
1091 * It'd be nice to have a capsule function to indicate how many elements
1092 * there were in a buffer for an RMF that's declared to be an array.
1093 * It's easy enough to compute the number of elements here though.
1095 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1096 if (remote_nb == NULL || niocount != (req_capsule_get_size(&req->rq_pill,
1097 &RMF_NIOBUF_REMOTE, RCL_CLIENT) / sizeof(*remote_nb)))
1098 GOTO(out, rc = -EFAULT);
1100 if ((remote_nb[0].flags & OBD_BRW_MEMALLOC) &&
1101 (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
1102 cfs_memory_pressure_set();
1104 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
1105 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
1107 CERROR("Missing capability for OST BRW WRITE");
1108 GOTO(out, rc = -EFAULT);
1112 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
1113 niocount * sizeof(*rcs));
1114 rc = req_capsule_server_pack(&req->rq_pill);
1117 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
1118 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
1120 tls = ost_tls_get(req);
1122 GOTO(out_bulk, rc = -ENOMEM);
1123 local_nb = tls->local;
1125 rc = ost_brw_lock_get(LCK_PW, exp, ioo, remote_nb, &lockh);
1130 * If getting the lock took more time than
1131 * client was willing to wait, drop it. b=11330
1133 if (cfs_time_current_sec() > req->rq_deadline ||
1134 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1136 CERROR("Dropping timed-out write from %s because locking "
1137 "object "LPX64" took %ld seconds (limit was %ld).\n",
1138 libcfs_id2str(req->rq_peer), ioo->ioo_id,
1139 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
1140 req->rq_deadline - req->rq_arrival_time.tv_sec);
1141 GOTO(out_lock, rc = -ETIMEDOUT);
1144 if (!lustre_handle_is_used(&lockh))
1145 /* no needs to try to prolong lock if server is asked
1146 * to handle locking (= OBD_BRW_SRVLOCK) */
1147 ost_rw_prolong_locks(req, ioo, remote_nb,&body->oa, LCK_PW);
1149 /* obd_preprw clobbers oa->valid, so save what we need */
1150 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1151 client_cksum = body->oa.o_cksum;
1152 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1153 cksum_type = cksum_type_unpack(body->oa.o_flags);
1155 if (body->oa.o_valid & OBD_MD_FLFLAGS && body->oa.o_flags & OBD_FL_MMAP)
1158 /* Because we already sync grant info with client when reconnect,
1159 * grant info will be cleared for resent req, then fed_grant and
1160 * total_grant will not be modified in following preprw_write */
1161 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
1162 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
1163 body->oa.o_valid &= ~OBD_MD_FLGRANT;
1166 if (exp_connect_rmtclient(exp)) {
1167 o_uid = body->oa.o_uid;
1168 o_gid = body->oa.o_gid;
1170 npages = OST_THREAD_POOL_SIZE;
1171 rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount,
1172 ioo, remote_nb, &npages, local_nb, oti, capa);
1176 desc = ptlrpc_prep_bulk_exp(req, npages,
1177 BULK_GET_SINK, OST_BULK_PORTAL);
1179 GOTO(skip_transfer, rc = -ENOMEM);
1181 /* NB Having prepped, we must commit... */
1183 for (i = 0; i < npages; i++)
1184 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
1185 local_nb[i].offset & ~CFS_PAGE_MASK,
1188 rc = sptlrpc_svc_prep_bulk(req, desc);
1192 /* Check if client was evicted or tried to reconnect while we
1193 * were doing i/o before touching network */
1194 if (desc->bd_export->exp_failed ||
1195 desc->bd_export->exp_abort_active_req)
1198 rc = ptlrpc_start_bulk_transfer(desc);
1200 time_t start = cfs_time_current_sec();
1202 long timeoutl = req->rq_deadline -
1203 cfs_time_current_sec();
1204 cfs_duration_t timeout = timeoutl <= 0 ?
1205 CFS_TICK : cfs_time_seconds(timeoutl);
1206 lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
1207 ost_bulk_timeout, desc);
1208 rc = l_wait_event(desc->bd_waitq,
1209 !ptlrpc_server_bulk_active(desc) ||
1210 desc->bd_export->exp_failed ||
1211 desc->bd_export->exp_abort_active_req,
1213 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1214 /* Wait again if we changed deadline */
1215 } while ((rc == -ETIMEDOUT) &&
1216 (req->rq_deadline > cfs_time_current_sec()));
1218 if (rc == -ETIMEDOUT) {
1219 DEBUG_REQ(D_ERROR, req,
1220 "timeout on bulk GET after %ld%+lds",
1221 req->rq_deadline - start,
1222 cfs_time_current_sec() -
1224 ptlrpc_abort_bulk(desc);
1225 } else if (desc->bd_export->exp_failed) {
1226 DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET");
1228 ptlrpc_abort_bulk(desc);
1229 } else if (desc->bd_export->exp_abort_active_req) {
1230 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk GET");
1231 /* we don't reply anyway */
1233 ptlrpc_abort_bulk(desc);
1234 } else if (!desc->bd_success) {
1235 DEBUG_REQ(D_ERROR, req, "network error on bulk GET");
1236 /* XXX should this be a different errno? */
1239 rc = sptlrpc_svc_unwrap_bulk(req, desc);
1242 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc);
1247 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1248 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
1250 if (unlikely(client_cksum != 0 && rc == 0)) {
1251 static int cksum_counter;
1252 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1253 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
1254 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
1255 server_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1256 repbody->oa.o_cksum = server_cksum;
1258 if (unlikely(client_cksum != server_cksum)) {
1259 CDEBUG_LIMIT(mmap ? D_INFO : D_ERROR,
1260 "client csum %x, server csum %x\n",
1261 client_cksum, server_cksum);
1263 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
1264 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
1265 cksum_counter, libcfs_id2str(req->rq_peer),
1270 /* Must commit after prep above in all cases */
1271 rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa, objcount, ioo,
1272 remote_nb, npages, local_nb, oti, rc);
1273 if (rc == -ENOTCONN)
1274 /* quota acquire process has been given up because
1275 * either the client has been evicted or the client
1276 * has timed out the request already */
1279 if (exp_connect_rmtclient(exp)) {
1280 repbody->oa.o_uid = o_uid;
1281 repbody->oa.o_gid = o_gid;
1285 * Disable sending mtime back to the client. If the client locked the
1286 * whole object, then it has already updated the mtime on its side,
1287 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
1289 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
1291 if (unlikely(client_cksum != server_cksum && rc == 0 && !mmap)) {
1292 int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1297 if (new_cksum == server_cksum)
1298 msg = "changed in transit before arrival at OST";
1299 else if (new_cksum == client_cksum)
1300 msg = "initial checksum before message complete";
1302 msg = "changed in transit AND after initial checksum";
1304 if (req->rq_peer.nid == desc->bd_sender) {
1308 router = libcfs_nid2str(desc->bd_sender);
1311 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from "
1312 "%s%s%s inode "DFID" object "
1313 LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1314 exp->exp_obd->obd_name, msg,
1315 libcfs_id2str(req->rq_peer),
1317 body->oa.o_valid & OBD_MD_FLFID ?
1318 body->oa.o_parent_seq : (__u64)0,
1319 body->oa.o_valid & OBD_MD_FLFID ?
1320 body->oa.o_parent_oid : 0,
1321 body->oa.o_valid & OBD_MD_FLFID ?
1322 body->oa.o_parent_ver : 0,
1324 body->oa.o_valid & OBD_MD_FLGROUP ?
1325 body->oa.o_seq : (__u64)0,
1327 local_nb[npages-1].offset +
1328 local_nb[npages-1].len - 1 );
1329 CERROR("client csum %x, original server csum %x, "
1330 "server csum now %x\n",
1331 client_cksum, server_cksum, new_cksum);
1337 /* set per-requested niobuf return codes */
1338 for (i = j = 0; i < niocount; i++) {
1339 int len = remote_nb[i].len;
1344 LASSERT(j < npages);
1345 if (local_nb[j].rc < 0)
1346 rcs[i] = local_nb[j].rc;
1347 len -= local_nb[j].len;
1352 LASSERT(j == npages);
1353 ptlrpc_lprocfs_brw(req, nob);
1357 ost_brw_lock_put(LCK_PW, ioo, remote_nb, &lockh);
1362 ptlrpc_free_bulk(desc);
1365 oti_to_request(oti, req);
1366 target_committed_to_req(req);
1367 rc = ptlrpc_reply(req);
1368 } else if (!no_reply) {
1369 /* Only reply if there was no comms problem with bulk */
1370 target_committed_to_req(req);
1371 req->rq_status = rc;
1374 /* reply out callback would free */
1375 ptlrpc_req_drop_rs(req);
1376 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
1377 "client will retry\n",
1378 exp->exp_obd->obd_name,
1379 exp->exp_client_uuid.uuid,
1380 exp->exp_connection->c_remote_uuid.uuid,
1381 libcfs_id2str(req->rq_peer));
1383 cfs_memory_pressure_clr();
1388 * Implementation of OST_SET_INFO.
1390 * OST_SET_INFO is like ioctl(): heavily overloaded. Specifically, it takes a
1391 * "key" and a value RPC buffers as arguments, with the value's contents
1392 * interpreted according to the key.
1394 * Value types that need swabbing have swabbing done explicitly, either here or
1395 * in functions called from here. This should be corrected: all swabbing should
1396 * be done in the capsule abstraction, as that will then allow us to move
1397 * swabbing exclusively to the client without having to modify server code
1398 * outside the capsule abstraction's implementation itself. To correct this
1399 * will require minor changes to the capsule abstraction; see the comments for
1400 * req_capsule_extend() in layout.c.
1402 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
1404 struct ost_body *body = NULL, *repbody;
1405 char *key, *val = NULL;
1406 int keylen, vallen, rc = 0;
1407 int is_grant_shrink = 0;
1410 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1412 DEBUG_REQ(D_HA, req, "no set_info key");
1415 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1418 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1421 if ((is_grant_shrink = KEY_IS(KEY_GRANT_SHRINK)))
1422 /* In this case the value is actually an RMF_OST_BODY, so we
1423 * transmutate the type of this PTLRPC */
1424 req_capsule_extend(&req->rq_pill, &RQF_OST_SET_GRANT_INFO);
1426 rc = req_capsule_server_pack(&req->rq_pill);
1431 if (is_grant_shrink) {
1432 body = req_capsule_client_get(&req->rq_pill,
1437 repbody = req_capsule_server_get(&req->rq_pill,
1439 memcpy(repbody, body, sizeof(*body));
1440 val = (char*)repbody;
1442 val = req_capsule_client_get(&req->rq_pill,
1447 if (KEY_IS(KEY_EVICT_BY_NID)) {
1449 obd_export_evict_by_nid(exp->exp_obd, val);
1451 } else if (KEY_IS(KEY_MDS_CONN) && ptlrpc_req_need_swab(req)) {
1452 if (vallen < sizeof(__u32))
1454 __swab32s((__u32 *)val);
1457 /* OBD will also check if KEY_IS(KEY_GRANT_SHRINK), and will cast val to
1458 * a struct ost_body * value */
1459 rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL);
1461 lustre_msg_set_status(req->rq_repmsg, 0);
1465 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
1468 int keylen, replylen, rc = 0;
1469 struct req_capsule *pill = &req->rq_pill;
1472 /* this common part for get_info rpc */
1473 key = req_capsule_client_get(pill, &RMF_SETINFO_KEY);
1475 DEBUG_REQ(D_HA, req, "no get_info key");
1478 keylen = req_capsule_get_size(pill, &RMF_SETINFO_KEY, RCL_CLIENT);
1480 if (KEY_IS(KEY_FIEMAP)) {
1481 struct ll_fiemap_info_key *fm_key = key;
1484 rc = ost_validate_obdo(exp, &fm_key->oa, NULL);
1489 rc = obd_get_info(exp, keylen, key, &replylen, NULL, NULL);
1493 req_capsule_set_size(pill, &RMF_GENERIC_DATA,
1494 RCL_SERVER, replylen);
1496 rc = req_capsule_server_pack(pill);
1500 reply = req_capsule_server_get(pill, &RMF_GENERIC_DATA);
1504 /* call again to fill in the reply buffer */
1505 rc = obd_get_info(exp, keylen, key, &replylen, reply, NULL);
1507 lustre_msg_set_status(req->rq_repmsg, 0);
1511 #ifdef HAVE_QUOTA_SUPPORT
1512 static int ost_handle_quotactl(struct ptlrpc_request *req)
1514 struct obd_quotactl *oqctl, *repoqc;
1518 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1520 GOTO(out, rc = -EPROTO);
1522 rc = req_capsule_server_pack(&req->rq_pill);
1526 repoqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1527 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1534 static int ost_handle_quotacheck(struct ptlrpc_request *req)
1536 struct obd_quotactl *oqctl;
1540 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1544 rc = req_capsule_server_pack(&req->rq_pill);
1548 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1552 static int ost_handle_quota_adjust_qunit(struct ptlrpc_request *req)
1554 struct quota_adjust_qunit *oqaq, *repoqa;
1555 struct lustre_quota_ctxt *qctxt;
1559 qctxt = &req->rq_export->exp_obd->u.obt.obt_qctxt;
1560 oqaq = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
1562 GOTO(out, rc = -EPROTO);
1564 rc = req_capsule_server_pack(&req->rq_pill);
1568 repoqa = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
1569 req->rq_status = obd_quota_adjust_qunit(req->rq_export, oqaq, qctxt, NULL);
1577 static int ost_llog_handle_connect(struct obd_export *exp,
1578 struct ptlrpc_request *req)
1580 struct llogd_conn_body *body;
1584 body = req_capsule_client_get(&req->rq_pill, &RMF_LLOGD_CONN_BODY);
1585 rc = obd_llog_connect(exp, body);
1589 #define ost_init_sec_none(reply, exp) \
1591 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \
1592 OBD_CONNECT_RMT_CLIENT_FORCE | \
1593 OBD_CONNECT_OSS_CAPA); \
1594 cfs_spin_lock(&exp->exp_lock); \
1595 exp->exp_connect_flags = reply->ocd_connect_flags; \
1596 cfs_spin_unlock(&exp->exp_lock); \
1599 static int ost_init_sec_level(struct ptlrpc_request *req)
1601 struct obd_export *exp = req->rq_export;
1602 struct req_capsule *pill = &req->rq_pill;
1603 struct obd_device *obd = exp->exp_obd;
1604 struct filter_obd *filter = &obd->u.filter;
1605 char *client = libcfs_nid2str(req->rq_peer.nid);
1606 struct obd_connect_data *data, *reply;
1610 data = req_capsule_client_get(pill, &RMF_CONNECT_DATA);
1611 reply = req_capsule_server_get(pill, &RMF_CONNECT_DATA);
1612 if (data == NULL || reply == NULL)
1615 /* connection from MDT is always trusted */
1616 if (req->rq_auth_usr_mdt) {
1617 ost_init_sec_none(reply, exp);
1621 /* no GSS support case */
1622 if (!req->rq_auth_gss) {
1623 if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
1624 CWARN("client %s -> target %s does not user GSS, "
1625 "can not run under security level %d.\n",
1626 client, obd->obd_name, filter->fo_sec_level);
1629 ost_init_sec_none(reply, exp);
1634 /* old version case */
1635 if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) ||
1636 !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) {
1637 if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
1638 CWARN("client %s -> target %s uses old version, "
1639 "can not run under security level %d.\n",
1640 client, obd->obd_name, filter->fo_sec_level);
1643 CWARN("client %s -> target %s uses old version, "
1644 "run under security level %d.\n",
1645 client, obd->obd_name, filter->fo_sec_level);
1646 ost_init_sec_none(reply, exp);
1651 remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE;
1653 if (!req->rq_auth_remote)
1654 CDEBUG(D_SEC, "client (local realm) %s -> target %s "
1655 "asked to be remote.\n", client, obd->obd_name);
1656 } else if (req->rq_auth_remote) {
1658 CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set "
1659 "as remote by default.\n", client, obd->obd_name);
1663 if (!filter->fo_fl_oss_capa) {
1664 CDEBUG(D_SEC, "client %s -> target %s is set as remote,"
1665 " but OSS capabilities are not enabled: %d.\n",
1666 client, obd->obd_name, filter->fo_fl_oss_capa);
1671 switch (filter->fo_sec_level) {
1672 case LUSTRE_SEC_NONE:
1674 ost_init_sec_none(reply, exp);
1677 CDEBUG(D_SEC, "client %s -> target %s is set as remote, "
1678 "can not run under security level %d.\n",
1679 client, obd->obd_name, filter->fo_sec_level);
1682 case LUSTRE_SEC_REMOTE:
1684 ost_init_sec_none(reply, exp);
1686 case LUSTRE_SEC_ALL:
1688 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
1689 OBD_CONNECT_RMT_CLIENT_FORCE);
1690 if (!filter->fo_fl_oss_capa)
1691 reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA;
1693 cfs_spin_lock(&exp->exp_lock);
1694 exp->exp_connect_flags = reply->ocd_connect_flags;
1695 cfs_spin_unlock(&exp->exp_lock);
1707 * this should be done in filter_connect()/filter_reconnect(), but
1708 * we can't obtain information like NID, which stored in incoming
1709 * request, thus can't decide what flavor to use. so we do it here.
1711 * This hack should be removed after the OST stack be rewritten, just
1712 * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect().
1714 static int ost_connect_check_sptlrpc(struct ptlrpc_request *req)
1716 struct obd_export *exp = req->rq_export;
1717 struct filter_obd *filter = &exp->exp_obd->u.filter;
1718 struct sptlrpc_flavor flvr;
1721 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
1722 LUSTRE_ECHO_NAME) == 0)) {
1723 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
1727 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
1728 cfs_read_lock(&filter->fo_sptlrpc_lock);
1729 sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset,
1733 cfs_read_unlock(&filter->fo_sptlrpc_lock);
1735 cfs_spin_lock(&exp->exp_lock);
1737 exp->exp_sp_peer = req->rq_sp_from;
1738 exp->exp_flvr = flvr;
1740 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
1741 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
1742 CERROR("unauthorized rpc flavor %x from %s, "
1743 "expect %x\n", req->rq_flvr.sf_rpc,
1744 libcfs_nid2str(req->rq_peer.nid),
1745 exp->exp_flvr.sf_rpc);
1749 cfs_spin_unlock(&exp->exp_lock);
1751 if (exp->exp_sp_peer != req->rq_sp_from) {
1752 CERROR("RPC source %s doesn't match %s\n",
1753 sptlrpc_part2name(req->rq_sp_from),
1754 sptlrpc_part2name(exp->exp_sp_peer));
1757 rc = sptlrpc_target_export_check(exp, req);
1764 /* Ensure that data and metadata are synced to the disk when lock is cancelled
1766 int ost_blocking_ast(struct ldlm_lock *lock,
1767 struct ldlm_lock_desc *desc,
1768 void *data, int flag)
1770 __u32 sync_lock_cancel = 0;
1771 __u32 len = sizeof(sync_lock_cancel);
1775 rc = obd_get_info(lock->l_export, sizeof(KEY_SYNC_LOCK_CANCEL),
1776 KEY_SYNC_LOCK_CANCEL, &len, &sync_lock_cancel, NULL);
1778 if (!rc && flag == LDLM_CB_CANCELING &&
1779 (lock->l_granted_mode & (LCK_PW|LCK_GROUP)) &&
1780 (sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1781 (sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1782 lock->l_flags & LDLM_FL_CBPENDING))) {
1783 struct obd_info *oinfo;
1787 OBD_ALLOC_PTR(oinfo);
1792 OBD_FREE_PTR(oinfo);
1795 oa->o_id = lock->l_resource->lr_name.name[0];
1796 oa->o_seq = lock->l_resource->lr_name.name[1];
1797 oa->o_valid = OBD_MD_FLID|OBD_MD_FLGROUP;
1800 rc = obd_sync(lock->l_export, oinfo,
1801 lock->l_policy_data.l_extent.start,
1802 lock->l_policy_data.l_extent.end, NULL);
1804 CERROR("Error %d syncing data on lock cancel\n", rc);
1807 OBD_FREE_PTR(oinfo);
1810 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1814 static int ost_filter_recovery_request(struct ptlrpc_request *req,
1815 struct obd_device *obd, int *process)
1817 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1818 case OST_CONNECT: /* This will never get here, but for completeness. */
1819 case OST_DISCONNECT:
1830 case OBD_LOG_CANCEL:
1832 *process = target_queue_recovery_request(req, obd);
1836 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1842 int ost_msg_check_version(struct lustre_msg *msg)
1846 switch(lustre_msg_get_opc(msg)) {
1848 case OST_DISCONNECT:
1851 case SEC_CTX_INIT_CONT:
1853 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1855 CERROR("bad opc %u version %08x, expecting %08x\n",
1856 lustre_msg_get_opc(msg),
1857 lustre_msg_get_version(msg),
1858 LUSTRE_OBD_VERSION);
1871 #ifdef HAVE_QUOTA_SUPPORT
1872 case OST_QUOTACHECK:
1874 case OST_QUOTA_ADJUST_QUNIT:
1876 rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION);
1878 CERROR("bad opc %u version %08x, expecting %08x\n",
1879 lustre_msg_get_opc(msg),
1880 lustre_msg_get_version(msg),
1881 LUSTRE_OST_VERSION);
1886 case LDLM_BL_CALLBACK:
1887 case LDLM_CP_CALLBACK:
1888 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1890 CERROR("bad opc %u version %08x, expecting %08x\n",
1891 lustre_msg_get_opc(msg),
1892 lustre_msg_get_version(msg),
1893 LUSTRE_DLM_VERSION);
1895 case LLOG_ORIGIN_CONNECT:
1896 case OBD_LOG_CANCEL:
1897 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1899 CERROR("bad opc %u version %08x, expecting %08x\n",
1900 lustre_msg_get_opc(msg),
1901 lustre_msg_get_version(msg),
1902 LUSTRE_LOG_VERSION);
1905 CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg));
1912 * Returns 1 if the given PTLRPC matches the given LDLM locks, or 0 if it does
1915 static int ost_rw_hpreq_lock_match(struct ptlrpc_request *req,
1916 struct ldlm_lock *lock)
1918 struct niobuf_remote *nb;
1919 struct obd_ioobj *ioo;
1920 struct ost_body *body;
1921 int objcount, niocount;
1922 int mode, opc, i, rc;
1926 opc = lustre_msg_get_opc(req->rq_reqmsg);
1927 LASSERT(opc == OST_READ || opc == OST_WRITE);
1929 /* As the request may be covered by several locks, do not look at
1930 * o_handle, look at the RPC IO region. */
1931 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1935 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1936 RCL_CLIENT) / sizeof(*ioo);
1937 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1941 rc = ost_validate_obdo(req->rq_export, &body->oa, ioo);
1945 for (niocount = i = 0; i < objcount; i++)
1946 niocount += ioo[i].ioo_bufcnt;
1948 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1950 niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE,
1951 RCL_CLIENT) / sizeof(*nb)))
1955 if (opc == OST_READ)
1958 start = nb[0].offset & CFS_PAGE_MASK;
1959 end = (nb[ioo->ioo_bufcnt - 1].offset +
1960 nb[ioo->ioo_bufcnt - 1].len - 1) | ~CFS_PAGE_MASK;
1962 LASSERT(lock->l_resource != NULL);
1963 if (!osc_res_name_eq(ioo->ioo_id, ioo->ioo_seq,
1964 &lock->l_resource->lr_name))
1967 if (!(lock->l_granted_mode & mode))
1970 if (lock->l_policy_data.l_extent.end < start ||
1971 lock->l_policy_data.l_extent.start > end)
1978 * High-priority queue request check for whether the given PTLRPC request (\a
1979 * req) is blocking an LDLM lock cancel.
1981 * Returns 1 if the given given PTLRPC request (\a req) is blocking an LDLM lock
1982 * cancel, 0 if it is not, and -EFAULT if the request is malformed.
1984 * Only OST_READs, OST_WRITEs and OST_PUNCHes go on the h-p RPC queue. This
1985 * function looks only at OST_READs and OST_WRITEs.
1987 static int ost_rw_hpreq_check(struct ptlrpc_request *req)
1989 struct niobuf_remote *nb;
1990 struct obd_ioobj *ioo;
1991 struct ost_body *body;
1992 int objcount, niocount;
1993 int mode, opc, i, rc;
1996 opc = lustre_msg_get_opc(req->rq_reqmsg);
1997 LASSERT(opc == OST_READ || opc == OST_WRITE);
1999 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2003 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
2004 RCL_CLIENT) / sizeof(*ioo);
2005 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
2009 rc = ost_validate_obdo(req->rq_export, &body->oa, ioo);
2013 for (niocount = i = 0; i < objcount; i++)
2014 niocount += ioo[i].ioo_bufcnt;
2015 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
2017 niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE,
2018 RCL_CLIENT) / sizeof(*nb)))
2020 if (niocount != 0 && (nb[0].flags & OBD_BRW_SRVLOCK))
2024 if (opc == OST_READ)
2026 RETURN(ost_rw_prolong_locks(req, ioo, nb, &body->oa, mode));
2029 static int ost_punch_prolong_locks(struct ptlrpc_request *req, struct obdo *oa)
2031 struct ldlm_res_id res_id = { .name = { oa->o_id } };
2032 struct ost_prolong_data opd = { 0 };
2037 end = start + oa->o_blocks;
2039 opd.opd_mode = LCK_PW;
2040 opd.opd_exp = req->rq_export;
2041 opd.opd_policy.l_extent.start = start & CFS_PAGE_MASK;
2042 if (oa->o_blocks == OBD_OBJECT_EOF || end < start)
2043 opd.opd_policy.l_extent.end = OBD_OBJECT_EOF;
2045 opd.opd_policy.l_extent.end = end | ~CFS_PAGE_MASK;
2047 /* prolong locks for the current service time of the corresponding
2048 * portal (= OST_IO_PORTAL) */
2049 opd.opd_timeout = AT_OFF ? obd_timeout / 2:
2050 max(at_est2timeout(at_get(&req->rq_rqbd->
2051 rqbd_service->srv_at_estimate)), ldlm_timeout);
2053 CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
2054 res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
2055 opd.opd_policy.l_extent.end);
2058 ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id,
2059 ost_prolong_locks_iter, &opd);
2060 RETURN(opd.opd_lock_match);
2064 * Like ost_rw_hpreq_lock_match(), but for OST_PUNCH RPCs.
2066 static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req,
2067 struct ldlm_lock *lock)
2069 struct ost_body *body;
2073 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2075 RETURN(0); /* can't return -EFAULT here */
2077 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
2081 if (body->oa.o_valid & OBD_MD_FLHANDLE &&
2082 body->oa.o_handle.cookie == lock->l_handle.h_cookie)
2088 * Like ost_rw_hpreq_check(), but for OST_PUNCH RPCs.
2090 static int ost_punch_hpreq_check(struct ptlrpc_request *req)
2092 struct ost_body *body;
2095 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2099 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
2103 LASSERT(!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
2104 !(body->oa.o_flags & OBD_FL_SRVLOCK));
2106 RETURN(ost_punch_prolong_locks(req, &body->oa));
2109 struct ptlrpc_hpreq_ops ost_hpreq_rw = {
2110 .hpreq_lock_match = ost_rw_hpreq_lock_match,
2111 .hpreq_check = ost_rw_hpreq_check,
2114 struct ptlrpc_hpreq_ops ost_hpreq_punch = {
2115 .hpreq_lock_match = ost_punch_hpreq_lock_match,
2116 .hpreq_check = ost_punch_hpreq_check,
2119 /** Assign high priority operations to the request if needed. */
2120 static int ost_hpreq_handler(struct ptlrpc_request *req)
2123 if (req->rq_export) {
2124 int opc = lustre_msg_get_opc(req->rq_reqmsg);
2125 struct ost_body *body;
2127 if (opc == OST_READ || opc == OST_WRITE) {
2128 struct niobuf_remote *nb;
2129 struct obd_ioobj *ioo;
2130 int objcount, niocount;
2133 /* RPCs on the H-P queue can be inspected before
2134 * ost_handler() initializes their pills, so we
2135 * initialize that here. Capsule initialization is
2136 * idempotent, as is setting the pill's format (provided
2137 * it doesn't change).
2139 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2140 if (opc == OST_READ)
2141 req_capsule_set(&req->rq_pill,
2144 req_capsule_set(&req->rq_pill,
2145 &RQF_OST_BRW_WRITE);
2147 body = req_capsule_client_get(&req->rq_pill,
2150 CERROR("Missing/short ost_body\n");
2154 objcount = req_capsule_get_size(&req->rq_pill,
2158 if (objcount == 0) {
2159 CERROR("Missing/short ioobj\n");
2163 CERROR("too many ioobjs (%d)\n", objcount);
2167 ioo = req_capsule_client_get(&req->rq_pill,
2170 CERROR("Missing/short ioobj\n");
2174 for (niocount = i = 0; i < objcount; i++) {
2175 if (ioo[i].ioo_bufcnt == 0) {
2176 CERROR("ioo[%d] has zero bufcnt\n", i);
2179 niocount += ioo[i].ioo_bufcnt;
2181 if (niocount > PTLRPC_MAX_BRW_PAGES) {
2182 DEBUG_REQ(D_RPCTRACE, req,
2183 "bulk has too many pages (%d)",
2188 nb = req_capsule_client_get(&req->rq_pill,
2189 &RMF_NIOBUF_REMOTE);
2191 CERROR("Missing/short niobuf\n");
2195 if (niocount == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
2196 req->rq_ops = &ost_hpreq_rw;
2197 } else if (opc == OST_PUNCH) {
2198 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2199 req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
2201 body = req_capsule_client_get(&req->rq_pill,
2204 CERROR("Missing/short ost_body\n");
2208 if (!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
2209 !(body->oa.o_flags & OBD_FL_SRVLOCK))
2210 req->rq_ops = &ost_hpreq_punch;
2216 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2217 int ost_handle(struct ptlrpc_request *req)
2219 struct obd_trans_info trans_info = { 0, };
2220 struct obd_trans_info *oti = &trans_info;
2221 int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
2222 struct obd_device *obd = NULL;
2225 LASSERT(current->journal_info == NULL);
2227 /* primordial rpcs don't affect server recovery */
2228 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2230 case SEC_CTX_INIT_CONT:
2235 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2237 if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) {
2238 if (!class_connected_export(req->rq_export)) {
2239 CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
2240 lustre_msg_get_opc(req->rq_reqmsg),
2241 libcfs_id2str(req->rq_peer));
2242 req->rq_status = -ENOTCONN;
2243 GOTO(out, rc = -ENOTCONN);
2246 obd = req->rq_export->exp_obd;
2248 /* Check for aborted recovery. */
2249 if (obd->obd_recovering) {
2250 rc = ost_filter_recovery_request(req, obd,
2252 if (rc || !should_process)
2254 else if (should_process < 0) {
2255 req->rq_status = should_process;
2256 rc = ptlrpc_error(req);
2264 rc = ost_msg_check_version(req->rq_reqmsg);
2268 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2270 CDEBUG(D_INODE, "connect\n");
2271 req_capsule_set(&req->rq_pill, &RQF_OST_CONNECT);
2272 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET))
2274 rc = target_handle_connect(req);
2275 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2))
2278 rc = ost_init_sec_level(req);
2280 rc = ost_connect_check_sptlrpc(req);
2284 case OST_DISCONNECT:
2285 CDEBUG(D_INODE, "disconnect\n");
2286 req_capsule_set(&req->rq_pill, &RQF_OST_DISCONNECT);
2287 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DISCONNECT_NET))
2289 rc = target_handle_disconnect(req);
2292 CDEBUG(D_INODE, "create\n");
2293 req_capsule_set(&req->rq_pill, &RQF_OST_CREATE);
2294 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET))
2296 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2297 GOTO(out, rc = -EROFS);
2298 rc = ost_create(req->rq_export, req, oti);
2301 CDEBUG(D_INODE, "destroy\n");
2302 req_capsule_set(&req->rq_pill, &RQF_OST_DESTROY);
2303 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET))
2305 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2306 GOTO(out, rc = -EROFS);
2307 rc = ost_destroy(req->rq_export, req, oti);
2310 CDEBUG(D_INODE, "getattr\n");
2311 req_capsule_set(&req->rq_pill, &RQF_OST_GETATTR);
2312 if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET))
2314 rc = ost_getattr(req->rq_export, req);
2317 CDEBUG(D_INODE, "setattr\n");
2318 req_capsule_set(&req->rq_pill, &RQF_OST_SETATTR);
2319 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET))
2321 rc = ost_setattr(req->rq_export, req, oti);
2324 req_capsule_set(&req->rq_pill, &RQF_OST_BRW_WRITE);
2325 CDEBUG(D_INODE, "write\n");
2326 /* req->rq_request_portal would be nice, if it was set */
2327 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
2328 CERROR("%s: deny write request from %s to portal %u\n",
2329 req->rq_export->exp_obd->obd_name,
2330 obd_export_nid2str(req->rq_export),
2331 req->rq_rqbd->rqbd_service->srv_req_portal);
2332 GOTO(out, rc = -EPROTO);
2334 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
2336 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2337 GOTO(out, rc = -ENOSPC);
2338 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2339 GOTO(out, rc = -EROFS);
2340 rc = ost_brw_write(req, oti);
2341 LASSERT(current->journal_info == NULL);
2342 /* ost_brw_write sends its own replies */
2345 req_capsule_set(&req->rq_pill, &RQF_OST_BRW_READ);
2346 CDEBUG(D_INODE, "read\n");
2347 /* req->rq_request_portal would be nice, if it was set */
2348 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
2349 CERROR("%s: deny read request from %s to portal %u\n",
2350 req->rq_export->exp_obd->obd_name,
2351 obd_export_nid2str(req->rq_export),
2352 req->rq_rqbd->rqbd_service->srv_req_portal);
2353 GOTO(out, rc = -EPROTO);
2355 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
2357 rc = ost_brw_read(req, oti);
2358 LASSERT(current->journal_info == NULL);
2359 /* ost_brw_read sends its own replies */
2362 CDEBUG(D_INODE, "punch\n");
2363 req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
2364 if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET))
2366 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2367 GOTO(out, rc = -EROFS);
2368 rc = ost_punch(req->rq_export, req, oti);
2371 CDEBUG(D_INODE, "statfs\n");
2372 req_capsule_set(&req->rq_pill, &RQF_OST_STATFS);
2373 if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET))
2375 rc = ost_statfs(req);
2378 CDEBUG(D_INODE, "sync\n");
2379 req_capsule_set(&req->rq_pill, &RQF_OST_SYNC);
2380 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET))
2382 rc = ost_sync(req->rq_export, req);
2385 DEBUG_REQ(D_INODE, req, "set_info");
2386 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2387 rc = ost_set_info(req->rq_export, req);
2390 DEBUG_REQ(D_INODE, req, "get_info");
2391 req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC);
2392 rc = ost_get_info(req->rq_export, req);
2394 #ifdef HAVE_QUOTA_SUPPORT
2395 case OST_QUOTACHECK:
2396 CDEBUG(D_INODE, "quotacheck\n");
2397 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK);
2398 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACHECK_NET))
2400 rc = ost_handle_quotacheck(req);
2403 CDEBUG(D_INODE, "quotactl\n");
2404 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACTL);
2405 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACTL_NET))
2407 rc = ost_handle_quotactl(req);
2409 case OST_QUOTA_ADJUST_QUNIT:
2410 CDEBUG(D_INODE, "quota_adjust_qunit\n");
2411 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTA_ADJUST_QUNIT);
2412 rc = ost_handle_quota_adjust_qunit(req);
2416 DEBUG_REQ(D_INODE, req, "ping");
2417 req_capsule_set(&req->rq_pill, &RQF_OBD_PING);
2418 rc = target_handle_ping(req);
2420 /* FIXME - just reply status */
2421 case LLOG_ORIGIN_CONNECT:
2422 DEBUG_REQ(D_INODE, req, "log connect");
2423 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_CONNECT);
2424 rc = ost_llog_handle_connect(req->rq_export, req);
2425 req->rq_status = rc;
2426 rc = req_capsule_server_pack(&req->rq_pill);
2429 RETURN(ptlrpc_reply(req));
2430 case OBD_LOG_CANCEL:
2431 CDEBUG(D_INODE, "log cancel\n");
2432 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2433 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2435 rc = llog_origin_handle_cancel(req);
2436 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2438 req->rq_status = rc;
2439 rc = req_capsule_server_pack(&req->rq_pill);
2442 RETURN(ptlrpc_reply(req));
2444 CDEBUG(D_INODE, "enqueue\n");
2445 req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE);
2446 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE))
2448 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2450 ldlm_server_glimpse_ast);
2451 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
2454 CDEBUG(D_INODE, "convert\n");
2455 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2456 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT))
2458 rc = ldlm_handle_convert(req);
2461 CDEBUG(D_INODE, "cancel\n");
2462 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2463 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2465 rc = ldlm_handle_cancel(req);
2467 case LDLM_BL_CALLBACK:
2468 case LDLM_CP_CALLBACK:
2469 CDEBUG(D_INODE, "callback\n");
2470 CERROR("callbacks should not happen on OST\n");
2473 CERROR("Unexpected opcode %d\n",
2474 lustre_msg_get_opc(req->rq_reqmsg));
2475 req->rq_status = -ENOTSUPP;
2476 rc = ptlrpc_error(req);
2480 LASSERT(current->journal_info == NULL);
2483 /* If we're DISCONNECTing, the export_data is already freed */
2484 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT)
2485 target_committed_to_req(req);
2489 oti_to_request(oti, req);
2491 target_send_reply(req, rc, fail);
2494 EXPORT_SYMBOL(ost_handle);
2496 * free per-thread pool created by ost_thread_init().
2498 static void ost_thread_done(struct ptlrpc_thread *thread)
2500 struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local
2505 LASSERT(thread != NULL);
2508 * be prepared to handle partially-initialized pools (because this is
2509 * called from ost_thread_init() for cleanup.
2511 tls = thread->t_data;
2514 thread->t_data = NULL;
2520 * initialize per-thread page pool (bug 5137).
2522 static int ost_thread_init(struct ptlrpc_thread *thread)
2524 struct ost_thread_local_cache *tls;
2528 LASSERT(thread != NULL);
2529 LASSERT(thread->t_data == NULL);
2530 LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
2535 thread->t_data = tls;
2539 #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
2541 /* Sigh - really, this is an OSS, the _server_, not the _target_ */
2542 static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2544 struct ost_obd *ost = &obd->u.ost;
2545 struct lprocfs_static_vars lvars;
2546 int oss_min_threads;
2547 int oss_max_threads;
2548 int oss_min_create_threads;
2549 int oss_max_create_threads;
2553 rc = cfs_cleanup_group_info();
2557 lprocfs_ost_init_vars(&lvars);
2558 lprocfs_obd_setup(obd, lvars.obd_vars);
2560 cfs_sema_init(&ost->ost_health_sem, 1);
2562 if (oss_num_threads) {
2563 /* If oss_num_threads is set, it is the min and the max. */
2564 if (oss_num_threads > OSS_THREADS_MAX)
2565 oss_num_threads = OSS_THREADS_MAX;
2566 if (oss_num_threads < OSS_THREADS_MIN)
2567 oss_num_threads = OSS_THREADS_MIN;
2568 oss_max_threads = oss_min_threads = oss_num_threads;
2570 /* Base min threads on memory and cpus */
2572 cfs_num_possible_cpus() * CFS_NUM_CACHEPAGES >>
2573 (27 - CFS_PAGE_SHIFT);
2574 if (oss_min_threads < OSS_THREADS_MIN)
2575 oss_min_threads = OSS_THREADS_MIN;
2576 /* Insure a 4x range for dynamic threads */
2577 if (oss_min_threads > OSS_THREADS_MAX / 4)
2578 oss_min_threads = OSS_THREADS_MAX / 4;
2579 oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4 + 1);
2583 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2584 OST_MAXREPSIZE, OST_REQUEST_PORTAL,
2585 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2586 ost_handle, LUSTRE_OSS_NAME,
2587 obd->obd_proc_entry, target_print_req,
2588 oss_min_threads, oss_max_threads,
2589 "ll_ost", LCT_DT_THREAD, NULL);
2590 if (ost->ost_service == NULL) {
2591 CERROR("failed to start service\n");
2592 GOTO(out_lprocfs, rc = -ENOMEM);
2595 rc = ptlrpc_start_threads(ost->ost_service);
2597 GOTO(out_service, rc = -EINVAL);
2599 if (oss_num_create_threads) {
2600 if (oss_num_create_threads > OSS_MAX_CREATE_THREADS)
2601 oss_num_create_threads = OSS_MAX_CREATE_THREADS;
2602 if (oss_num_create_threads < OSS_MIN_CREATE_THREADS)
2603 oss_num_create_threads = OSS_MIN_CREATE_THREADS;
2604 oss_min_create_threads = oss_max_create_threads =
2605 oss_num_create_threads;
2607 oss_min_create_threads = OSS_MIN_CREATE_THREADS;
2608 oss_max_create_threads = OSS_MAX_CREATE_THREADS;
2611 ost->ost_create_service =
2612 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2613 OST_MAXREPSIZE, OST_CREATE_PORTAL,
2614 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2615 ost_handle, "ost_create",
2616 obd->obd_proc_entry, target_print_req,
2617 oss_min_create_threads, oss_max_create_threads,
2618 "ll_ost_creat", LCT_DT_THREAD, NULL);
2619 if (ost->ost_create_service == NULL) {
2620 CERROR("failed to start OST create service\n");
2621 GOTO(out_service, rc = -ENOMEM);
2624 rc = ptlrpc_start_threads(ost->ost_create_service);
2626 GOTO(out_create, rc = -EINVAL);
2628 ost->ost_io_service =
2629 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2630 OST_MAXREPSIZE, OST_IO_PORTAL,
2631 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2632 ost_handle, "ost_io",
2633 obd->obd_proc_entry, target_print_req,
2634 oss_min_threads, oss_max_threads,
2635 "ll_ost_io", LCT_DT_THREAD, ost_hpreq_handler);
2636 if (ost->ost_io_service == NULL) {
2637 CERROR("failed to start OST I/O service\n");
2638 GOTO(out_create, rc = -ENOMEM);
2641 ost->ost_io_service->srv_init = ost_thread_init;
2642 ost->ost_io_service->srv_done = ost_thread_done;
2643 ost->ost_io_service->srv_cpu_affinity = 1;
2644 rc = ptlrpc_start_threads(ost->ost_io_service);
2646 GOTO(out_io, rc = -EINVAL);
2648 ping_evictor_start();
2653 ptlrpc_unregister_service(ost->ost_io_service);
2654 ost->ost_io_service = NULL;
2656 ptlrpc_unregister_service(ost->ost_create_service);
2657 ost->ost_create_service = NULL;
2659 ptlrpc_unregister_service(ost->ost_service);
2660 ost->ost_service = NULL;
2662 lprocfs_obd_cleanup(obd);
2666 static int ost_cleanup(struct obd_device *obd)
2668 struct ost_obd *ost = &obd->u.ost;
2672 ping_evictor_stop();
2674 /* there is no recovery for OST OBD, all recovery is controlled by
2676 LASSERT(obd->obd_recovering == 0);
2677 cfs_down(&ost->ost_health_sem);
2678 ptlrpc_unregister_service(ost->ost_service);
2679 ptlrpc_unregister_service(ost->ost_create_service);
2680 ptlrpc_unregister_service(ost->ost_io_service);
2681 ost->ost_service = NULL;
2682 ost->ost_create_service = NULL;
2683 cfs_up(&ost->ost_health_sem);
2685 lprocfs_obd_cleanup(obd);
2690 static int ost_health_check(struct obd_device *obd)
2692 struct ost_obd *ost = &obd->u.ost;
2695 cfs_down(&ost->ost_health_sem);
2696 rc |= ptlrpc_service_health_check(ost->ost_service);
2697 rc |= ptlrpc_service_health_check(ost->ost_create_service);
2698 rc |= ptlrpc_service_health_check(ost->ost_io_service);
2699 cfs_up(&ost->ost_health_sem);
2702 * health_check to return 0 on healthy
2703 * and 1 on unhealthy.
2711 struct ost_thread_local_cache *ost_tls(struct ptlrpc_request *r)
2713 return (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
2716 /* use obd ops to offer management infrastructure */
2717 static struct obd_ops ost_obd_ops = {
2718 .o_owner = THIS_MODULE,
2719 .o_setup = ost_setup,
2720 .o_cleanup = ost_cleanup,
2721 .o_health_check = ost_health_check,
2725 static int __init ost_init(void)
2727 struct lprocfs_static_vars lvars;
2731 lprocfs_ost_init_vars(&lvars);
2732 rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
2733 LUSTRE_OSS_NAME, NULL);
2735 if (ost_num_threads != 0 && oss_num_threads == 0) {
2736 LCONSOLE_INFO("ost_num_threads module parameter is deprecated, "
2737 "use oss_num_threads instead or unset both for "
2738 "dynamic thread startup\n");
2739 oss_num_threads = ost_num_threads;
2745 static void /*__exit*/ ost_exit(void)
2747 class_unregister_type(LUSTRE_OSS_NAME);
2750 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2751 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
2752 MODULE_LICENSE("GPL");
2754 module_init(ost_init);
2755 module_exit(ost_exit);