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 = { { { 0 } } };
253 struct lustre_handle lh = { 0 };
257 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
261 rc = ost_validate_obdo(exp, &body->oa, NULL);
265 rc = req_capsule_server_pack(&req->rq_pill);
269 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
270 repbody->oa = body->oa;
272 rc = ost_lock_get(exp, &body->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0);
276 oinfo.oi_oa = &repbody->oa;
277 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) {
278 oinfo.oi_capa = req_capsule_client_get(&req->rq_pill,
280 if (oinfo.oi_capa == NULL) {
281 CERROR("Missing capability for OST GETATTR");
286 req->rq_status = obd_getattr(exp, &oinfo);
287 ost_lock_put(exp, &lh, LCK_PR);
289 ost_drop_id(exp, &repbody->oa);
293 static int ost_statfs(struct ptlrpc_request *req)
295 struct obd_statfs *osfs;
299 rc = req_capsule_server_pack(&req->rq_pill);
303 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
305 req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs,
306 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
308 if (req->rq_status != 0)
309 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
314 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
315 struct obd_trans_info *oti)
317 struct ost_body *body, *repbody;
321 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
325 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
329 rc = req_capsule_server_pack(&req->rq_pill);
333 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
334 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
335 oti->oti_logcookies = &repbody->oa.o_lcookie;
337 req->rq_status = obd_create(exp, &repbody->oa, NULL, oti);
338 //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
342 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
343 struct obd_trans_info *oti)
345 struct obd_info oinfo = { { { 0 } } };
346 struct ost_body *body, *repbody;
348 struct lustre_handle lh = {0,};
351 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
352 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
354 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
358 rc = ost_validate_obdo(exp, &body->oa, NULL);
362 oinfo.oi_oa = &body->oa;
363 oinfo.oi_policy.l_extent.start = oinfo.oi_oa->o_size;
364 oinfo.oi_policy.l_extent.end = oinfo.oi_oa->o_blocks;
366 if ((oinfo.oi_oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
367 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
370 rc = req_capsule_server_pack(&req->rq_pill);
374 /* standard truncate optimization: if file body is completely
375 * destroyed, don't send data back to the server. */
376 if (oinfo.oi_oa->o_size == 0)
377 flags |= LDLM_AST_DISCARD_DATA;
379 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
380 rc = ost_lock_get(exp, oinfo.oi_oa, oinfo.oi_oa->o_size,
381 oinfo.oi_oa->o_blocks, &lh, LCK_PW, flags);
383 if (oinfo.oi_oa->o_valid & OBD_MD_FLFLAGS &&
384 oinfo.oi_oa->o_flags == OBD_FL_SRVLOCK)
386 * If OBD_FL_SRVLOCK is the only bit set in
387 * ->o_flags, clear OBD_MD_FLFLAGS to avoid falling
388 * through filter_setattr() to filter_iocontrol().
390 oinfo.oi_oa->o_valid &= ~OBD_MD_FLFLAGS;
392 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) {
393 oinfo.oi_capa = req_capsule_client_get(&req->rq_pill,
395 if (oinfo.oi_capa == NULL) {
396 CERROR("Missing capability for OST PUNCH");
400 req->rq_status = obd_punch(exp, &oinfo, oti, NULL);
401 ost_lock_put(exp, &lh, LCK_PW);
403 repbody->oa = *oinfo.oi_oa;
404 ost_drop_id(exp, &repbody->oa);
408 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
410 struct ost_body *body, *repbody;
411 struct lustre_capa *capa = NULL;
415 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
419 rc = ost_validate_obdo(exp, &body->oa, NULL);
423 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
424 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
426 CERROR("Missing capability for OST SYNC");
431 rc = req_capsule_server_pack(&req->rq_pill);
435 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
436 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
437 req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
438 repbody->oa.o_blocks, capa);
439 ost_drop_id(exp, &repbody->oa);
443 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
444 struct obd_trans_info *oti)
446 struct ost_body *body, *repbody;
448 struct obd_info oinfo = { { { 0 } } };
451 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
455 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
459 rc = req_capsule_server_pack(&req->rq_pill);
463 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
464 repbody->oa = body->oa;
466 oinfo.oi_oa = &repbody->oa;
467 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA) {
468 oinfo.oi_capa = req_capsule_client_get(&req->rq_pill,
470 if (oinfo.oi_capa == NULL) {
471 CERROR("Missing capability for OST SETATTR");
475 req->rq_status = obd_setattr(exp, &oinfo, oti);
476 ost_drop_id(exp, &repbody->oa);
480 static int ost_bulk_timeout(void *data)
483 /* We don't fail the connection here, because having the export
484 * killed makes the (vital) call to commitrw very sad.
489 static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc,
490 cksum_type_t cksum_type)
495 cksum = init_checksum(cksum_type);
496 for (i = 0; i < desc->bd_iov_count; i++) {
497 struct page *page = desc->bd_iov[i].kiov_page;
498 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
499 char *ptr = kmap(page) + off;
500 int len = desc->bd_iov[i].kiov_len;
502 /* corrupt the data before we compute the checksum, to
503 * simulate a client->OST data error */
504 if (i == 0 && opc == OST_WRITE &&
505 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE))
506 memcpy(ptr, "bad3", min(4, len));
507 cksum = compute_checksum(cksum, ptr, len, cksum_type);
508 /* corrupt the data after we compute the checksum, to
509 * simulate an OST->client data error */
510 if (i == 0 && opc == OST_READ &&
511 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
512 memcpy(ptr, "bad4", min(4, len));
513 /* nobody should use corrupted page again */
514 ClearPageUptodate(page);
522 static int ost_brw_lock_get(int mode, struct obd_export *exp,
523 struct obd_ioobj *obj, struct niobuf_remote *nb,
524 struct lustre_handle *lh)
527 int nrbufs = obj->ioo_bufcnt;
528 struct ldlm_res_id res_id;
529 ldlm_policy_data_t policy;
533 osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
534 LASSERT(mode == LCK_PR || mode == LCK_PW);
535 LASSERT(!lustre_handle_is_used(lh));
537 if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
540 for (i = 1; i < nrbufs; i ++)
541 if ((nb[0].flags & OBD_BRW_SRVLOCK) !=
542 (nb[i].flags & OBD_BRW_SRVLOCK))
545 policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
546 policy.l_extent.end = (nb[nrbufs - 1].offset +
547 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
549 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
550 LDLM_EXTENT, &policy, mode, &flags,
551 ldlm_blocking_ast, ldlm_completion_ast,
552 ldlm_glimpse_ast, NULL, 0, NULL, lh));
555 static void ost_brw_lock_put(int mode,
556 struct obd_ioobj *obj, struct niobuf_remote *niob,
557 struct lustre_handle *lh)
560 LASSERT(mode == LCK_PR || mode == LCK_PW);
561 LASSERT((obj->ioo_bufcnt > 0 && (niob[0].flags & OBD_BRW_SRVLOCK)) ==
562 lustre_handle_is_used(lh));
563 if (lustre_handle_is_used(lh))
564 ldlm_lock_decref(lh, mode);
568 struct ost_prolong_data {
569 struct obd_export *opd_exp;
570 ldlm_policy_data_t opd_policy;
572 ldlm_mode_t opd_mode;
577 static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data)
579 struct ost_prolong_data *opd = data;
581 LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
583 if (lock->l_req_mode != lock->l_granted_mode) {
584 /* scan granted locks only */
585 return LDLM_ITER_STOP;
588 if (lock->l_export != opd->opd_exp) {
589 /* prolong locks only for given client */
590 return LDLM_ITER_CONTINUE;
593 if (!(lock->l_granted_mode & opd->opd_mode)) {
594 /* we aren't interesting in all type of locks */
595 return LDLM_ITER_CONTINUE;
598 if (lock->l_policy_data.l_extent.end < opd->opd_policy.l_extent.start ||
599 lock->l_policy_data.l_extent.start > opd->opd_policy.l_extent.end) {
600 /* the request doesn't cross the lock, skip it */
601 return LDLM_ITER_CONTINUE;
604 /* Fill the obdo with the matched lock handle.
605 * XXX: it is possible in some cases the IO RPC is covered by several
606 * locks, even for the write case, so it may need to be a lock list. */
607 if (opd->opd_oa && !(opd->opd_oa->o_valid & OBD_MD_FLHANDLE)) {
608 opd->opd_oa->o_handle.cookie = lock->l_handle.h_cookie;
609 opd->opd_oa->o_valid |= OBD_MD_FLHANDLE;
612 if (!(lock->l_flags & LDLM_FL_AST_SENT)) {
613 /* ignore locks not being cancelled */
614 return LDLM_ITER_CONTINUE;
617 CDEBUG(D_DLMTRACE,"refresh lock: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
618 lock->l_resource->lr_name.name[0],
619 lock->l_resource->lr_name.name[1],
620 opd->opd_policy.l_extent.start, opd->opd_policy.l_extent.end);
621 /* OK. this is a possible lock the user holds doing I/O
622 * let's refresh eviction timer for it */
623 ldlm_refresh_waiting_lock(lock, opd->opd_timeout);
624 opd->opd_lock_match = 1;
626 return LDLM_ITER_CONTINUE;
629 static int ost_rw_prolong_locks(struct ptlrpc_request *req, struct obd_ioobj *obj,
630 struct niobuf_remote *nb, struct obdo *oa,
633 struct ldlm_res_id res_id;
634 int nrbufs = obj->ioo_bufcnt;
635 struct ost_prolong_data opd = { 0 };
638 osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
641 opd.opd_exp = req->rq_export;
642 opd.opd_policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
643 opd.opd_policy.l_extent.end = (nb[nrbufs - 1].offset +
644 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
646 /* prolong locks for the current service time of the corresponding
647 * portal (= OST_IO_PORTAL) */
648 opd.opd_timeout = AT_OFF ? obd_timeout / 2:
649 max(at_est2timeout(at_get(&req->rq_rqbd->
650 rqbd_service->srv_at_estimate)), ldlm_timeout);
652 CDEBUG(D_INFO,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
653 res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
654 opd.opd_policy.l_extent.end);
656 if (oa->o_valid & OBD_MD_FLHANDLE) {
657 struct ldlm_lock *lock;
659 lock = ldlm_handle2lock(&oa->o_handle);
661 ost_prolong_locks_iter(lock, &opd);
662 if (opd.opd_lock_match) {
667 /* Check if the lock covers the whole IO region,
668 * otherwise iterate through the resource. */
669 if (lock->l_policy_data.l_extent.end >=
670 opd.opd_policy.l_extent.end &&
671 lock->l_policy_data.l_extent.start <=
672 opd.opd_policy.l_extent.start) {
681 ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id,
682 ost_prolong_locks_iter, &opd);
683 RETURN(opd.opd_lock_match);
686 /* Allocate thread local buffers if needed */
687 static struct ost_thread_local_cache *ost_tls_get(struct ptlrpc_request *r)
689 struct ost_thread_local_cache *tls =
690 (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
692 /* In normal mode of operation an I/O request is serviced only
693 * by ll_ost_io threads each of them has own tls buffers allocated by
695 * During recovery, an I/O request may be queued until any of the ost
696 * service threads process it. Not necessary it should be one of
697 * ll_ost_io threads. In that case we dynamically allocating tls
698 * buffers for the request service time. */
699 if (unlikely(tls == NULL)) {
700 LASSERT(r->rq_export->exp_in_recovery);
704 r->rq_svc_thread->t_data = tls;
710 /* Free thread local buffers if they were allocated only for servicing
711 * this one request */
712 static void ost_tls_put(struct ptlrpc_request *r)
714 struct ost_thread_local_cache *tls =
715 (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
717 if (unlikely(tls->temporary)) {
719 r->rq_svc_thread->t_data = NULL;
723 static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti)
725 struct ptlrpc_bulk_desc *desc = NULL;
726 struct obd_export *exp = req->rq_export;
727 struct niobuf_remote *remote_nb;
728 struct niobuf_local *local_nb;
729 struct obd_ioobj *ioo;
730 struct ost_body *body, *repbody;
731 struct lustre_capa *capa = NULL;
732 struct l_wait_info lwi;
733 struct lustre_handle lockh = { 0 };
734 int niocount, npages, nob = 0, rc, i;
736 struct ost_thread_local_cache *tls;
739 req->rq_bulk_read = 1;
741 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
742 GOTO(out, rc = -EIO);
744 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
746 /* Check if there is eviction in progress, and if so, wait for it to
748 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
749 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
750 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
751 !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
755 GOTO(out, rc = -ENOTCONN);
757 /* ost_body, ioobj & noibuf_remote are verified and swabbed in
758 * ost_rw_hpreq_check(). */
759 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
761 GOTO(out, rc = -EFAULT);
764 * A req_capsule_X_get_array(pill, field, ptr_to_element_count) function
765 * would be useful here and wherever we get &RMF_OBD_IOOBJ and
766 * &RMF_NIOBUF_REMOTE.
768 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
770 GOTO(out, rc = -EFAULT);
772 rc = ost_validate_obdo(exp, &body->oa, ioo);
776 niocount = ioo->ioo_bufcnt;
777 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
778 if (remote_nb == NULL)
779 GOTO(out, rc = -EFAULT);
781 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
782 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
784 CERROR("Missing capability for OST BRW READ");
785 GOTO(out, rc = -EFAULT);
789 rc = req_capsule_server_pack(&req->rq_pill);
793 tls = ost_tls_get(req);
795 GOTO(out_bulk, rc = -ENOMEM);
796 local_nb = tls->local;
798 rc = ost_brw_lock_get(LCK_PR, exp, ioo, remote_nb, &lockh);
803 * If getting the lock took more time than
804 * client was willing to wait, drop it. b=11330
806 if (cfs_time_current_sec() > req->rq_deadline ||
807 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
809 CERROR("Dropping timed-out read from %s because locking"
810 "object "LPX64" took %ld seconds (limit was %ld).\n",
811 libcfs_id2str(req->rq_peer), ioo->ioo_id,
812 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
813 req->rq_deadline - req->rq_arrival_time.tv_sec);
814 GOTO(out_lock, rc = -ETIMEDOUT);
817 npages = OST_THREAD_POOL_SIZE;
818 rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1, ioo,
819 remote_nb, &npages, local_nb, oti, capa);
823 desc = ptlrpc_prep_bulk_exp(req, npages,
824 BULK_PUT_SOURCE, OST_BULK_PORTAL);
826 GOTO(out_lock, rc = -ENOMEM);
828 if (!lustre_handle_is_used(&lockh))
829 /* no needs to try to prolong lock if server is asked
830 * to handle locking (= OBD_BRW_SRVLOCK) */
831 ost_rw_prolong_locks(req, ioo, remote_nb, &body->oa,
835 for (i = 0; i < npages; i++) {
836 int page_rc = local_nb[i].rc;
838 if (page_rc < 0) { /* error */
844 if (page_rc != 0) { /* some data! */
845 LASSERT (local_nb[i].page != NULL);
846 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
847 local_nb[i].offset & ~CFS_PAGE_MASK,
851 if (page_rc != local_nb[i].len) { /* short read */
852 /* All subsequent pages should be 0 */
854 LASSERT(local_nb[i].rc == 0);
859 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
860 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
862 if (body->oa.o_valid & OBD_MD_FLFLAGS)
863 cksum_type = cksum_type_unpack(body->oa.o_flags);
864 body->oa.o_flags = cksum_type_pack(cksum_type);
865 body->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
866 body->oa.o_cksum = ost_checksum_bulk(desc, OST_READ, cksum_type);
867 CDEBUG(D_PAGE,"checksum at read origin: %x\n",body->oa.o_cksum);
869 body->oa.o_valid = 0;
871 /* We're finishing using body->oa as an input variable */
873 /* Check if client was evicted while we were doing i/o before touching
876 /* Check if there is eviction in progress, and if so, wait for
878 if (unlikely(cfs_atomic_read(&exp->exp_obd->
879 obd_evict_inprogress))) {
880 lwi = LWI_INTR(NULL, NULL);
881 rc = l_wait_event(exp->exp_obd->
882 obd_evict_inprogress_waitq,
883 !cfs_atomic_read(&exp->exp_obd->
884 obd_evict_inprogress),
887 /* Check if client was evicted or tried to reconnect already */
888 if (exp->exp_failed || exp->exp_abort_active_req)
891 rc = sptlrpc_svc_wrap_bulk(req, desc);
893 rc = ptlrpc_start_bulk_transfer(desc);
897 time_t start = cfs_time_current_sec();
899 long timeoutl = req->rq_deadline -
900 cfs_time_current_sec();
901 cfs_duration_t timeout = timeoutl <= 0 ?
902 CFS_TICK : cfs_time_seconds(timeoutl);
903 lwi = LWI_TIMEOUT_INTERVAL(timeout,
907 rc = l_wait_event(desc->bd_waitq,
908 !ptlrpc_server_bulk_active(desc) ||
910 exp->exp_abort_active_req,
912 LASSERT(rc == 0 || rc == -ETIMEDOUT);
913 /* Wait again if we changed deadline */
914 } while ((rc == -ETIMEDOUT) &&
915 (req->rq_deadline > cfs_time_current_sec()));
917 if (rc == -ETIMEDOUT) {
918 DEBUG_REQ(D_ERROR, req,
919 "timeout on bulk PUT after %ld%+lds",
920 req->rq_deadline - start,
921 cfs_time_current_sec() -
923 ptlrpc_abort_bulk(desc);
924 } else if (exp->exp_failed) {
925 DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT");
927 ptlrpc_abort_bulk(desc);
928 } else if (exp->exp_abort_active_req) {
929 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk PUT");
930 /* we don't reply anyway */
932 ptlrpc_abort_bulk(desc);
933 } else if (!desc->bd_success ||
934 desc->bd_nob_transferred != desc->bd_nob) {
935 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
937 "truncated" : "network error on",
938 desc->bd_nob_transferred,
940 /* XXX should this be a different errno? */
944 DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d", rc);
949 /* Must commit after prep above in all cases */
950 rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1, ioo,
951 remote_nb, npages, local_nb, oti, rc);
954 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
955 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
956 ost_drop_id(exp, &repbody->oa);
960 ost_brw_lock_put(LCK_PR, ioo, remote_nb, &lockh);
965 ptlrpc_free_bulk(desc);
969 req->rq_status = nob;
970 ptlrpc_lprocfs_brw(req, nob);
971 target_committed_to_req(req);
973 } else if (!no_reply) {
974 /* Only reply if there was no comms problem with bulk */
975 target_committed_to_req(req);
979 /* reply out callback would free */
980 ptlrpc_req_drop_rs(req);
981 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
982 "client will retry\n",
983 exp->exp_obd->obd_name,
984 exp->exp_client_uuid.uuid,
985 exp->exp_connection->c_remote_uuid.uuid,
986 libcfs_id2str(req->rq_peer));
992 static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
994 struct ptlrpc_bulk_desc *desc = NULL;
995 struct obd_export *exp = req->rq_export;
996 struct niobuf_remote *remote_nb;
997 struct niobuf_local *local_nb;
998 struct obd_ioobj *ioo;
999 struct ost_body *body, *repbody;
1000 struct l_wait_info lwi;
1001 struct lustre_handle lockh = {0};
1002 struct lustre_capa *capa = NULL;
1004 int objcount, niocount, npages;
1006 obd_count client_cksum = 0, server_cksum = 0;
1007 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
1009 __u32 o_uid = 0, o_gid = 0;
1010 struct ost_thread_local_cache *tls;
1013 req->rq_bulk_write = 1;
1015 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
1016 GOTO(out, rc = -EIO);
1017 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
1018 GOTO(out, rc = -EFAULT);
1020 /* pause before transaction has been started */
1021 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
1023 /* Check if there is eviction in progress, and if so, wait for it to
1025 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
1026 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
1027 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
1028 !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
1031 if (exp->exp_failed)
1032 GOTO(out, rc = -ENOTCONN);
1034 /* ost_body, ioobj & noibuf_remote are verified and swabbed in
1035 * ost_rw_hpreq_check(). */
1036 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1038 GOTO(out, rc = -EFAULT);
1040 if ((body->oa.o_flags & OBD_BRW_MEMALLOC) &&
1041 (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
1042 cfs_memory_pressure_set();
1044 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1045 RCL_CLIENT) / sizeof(*ioo);
1046 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1048 GOTO(out, rc = -EFAULT);
1050 rc = ost_validate_obdo(exp, &body->oa, ioo);
1054 for (niocount = i = 0; i < objcount; i++)
1055 niocount += ioo[i].ioo_bufcnt;
1058 * It'd be nice to have a capsule function to indicate how many elements
1059 * there were in a buffer for an RMF that's declared to be an array.
1060 * It's easy enough to compute the number of elements here though.
1062 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1063 if (remote_nb == NULL || niocount != (req_capsule_get_size(&req->rq_pill,
1064 &RMF_NIOBUF_REMOTE, RCL_CLIENT) / sizeof(*remote_nb)))
1065 GOTO(out, rc = -EFAULT);
1067 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
1068 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
1070 CERROR("Missing capability for OST BRW WRITE");
1071 GOTO(out, rc = -EFAULT);
1075 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
1076 niocount * sizeof(*rcs));
1077 rc = req_capsule_server_pack(&req->rq_pill);
1080 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, obd_fail_val);
1081 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
1083 tls = ost_tls_get(req);
1085 GOTO(out_bulk, rc = -ENOMEM);
1086 local_nb = tls->local;
1088 rc = ost_brw_lock_get(LCK_PW, exp, ioo, remote_nb, &lockh);
1093 * If getting the lock took more time than
1094 * client was willing to wait, drop it. b=11330
1096 if (cfs_time_current_sec() > req->rq_deadline ||
1097 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1099 CERROR("Dropping timed-out write from %s because locking "
1100 "object "LPX64" took %ld seconds (limit was %ld).\n",
1101 libcfs_id2str(req->rq_peer), ioo->ioo_id,
1102 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
1103 req->rq_deadline - req->rq_arrival_time.tv_sec);
1104 GOTO(out_lock, rc = -ETIMEDOUT);
1107 if (!lustre_handle_is_used(&lockh))
1108 /* no needs to try to prolong lock if server is asked
1109 * to handle locking (= OBD_BRW_SRVLOCK) */
1110 ost_rw_prolong_locks(req, ioo, remote_nb,&body->oa, LCK_PW);
1112 /* obd_preprw clobbers oa->valid, so save what we need */
1113 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1114 client_cksum = body->oa.o_cksum;
1115 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1116 cksum_type = cksum_type_unpack(body->oa.o_flags);
1119 /* Because we already sync grant info with client when reconnect,
1120 * grant info will be cleared for resent req, then fed_grant and
1121 * total_grant will not be modified in following preprw_write */
1122 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
1123 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
1124 body->oa.o_valid &= ~OBD_MD_FLGRANT;
1127 if (exp_connect_rmtclient(exp)) {
1128 o_uid = body->oa.o_uid;
1129 o_gid = body->oa.o_gid;
1131 npages = OST_THREAD_POOL_SIZE;
1132 rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount,
1133 ioo, remote_nb, &npages, local_nb, oti, capa);
1137 desc = ptlrpc_prep_bulk_exp(req, npages,
1138 BULK_GET_SINK, OST_BULK_PORTAL);
1140 GOTO(out_lock, rc = -ENOMEM);
1142 /* NB Having prepped, we must commit... */
1144 for (i = 0; i < npages; i++)
1145 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
1146 local_nb[i].offset & ~CFS_PAGE_MASK,
1149 rc = sptlrpc_svc_prep_bulk(req, desc);
1153 /* Check if client was evicted or tried to reconnect while we
1154 * were doing i/o before touching network */
1155 if (desc->bd_export->exp_failed ||
1156 desc->bd_export->exp_abort_active_req)
1159 rc = ptlrpc_start_bulk_transfer(desc);
1161 time_t start = cfs_time_current_sec();
1163 long timeoutl = req->rq_deadline -
1164 cfs_time_current_sec();
1165 cfs_duration_t timeout = timeoutl <= 0 ?
1166 CFS_TICK : cfs_time_seconds(timeoutl);
1167 lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
1168 ost_bulk_timeout, desc);
1169 rc = l_wait_event(desc->bd_waitq,
1170 !ptlrpc_server_bulk_active(desc) ||
1171 desc->bd_export->exp_failed ||
1172 desc->bd_export->exp_abort_active_req,
1174 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1175 /* Wait again if we changed deadline */
1176 } while ((rc == -ETIMEDOUT) &&
1177 (req->rq_deadline > cfs_time_current_sec()));
1179 if (rc == -ETIMEDOUT) {
1180 DEBUG_REQ(D_ERROR, req,
1181 "timeout on bulk GET after %ld%+lds",
1182 req->rq_deadline - start,
1183 cfs_time_current_sec() -
1185 ptlrpc_abort_bulk(desc);
1186 } else if (desc->bd_export->exp_failed) {
1187 DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET");
1189 ptlrpc_abort_bulk(desc);
1190 } else if (desc->bd_export->exp_abort_active_req) {
1191 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk GET");
1192 /* we don't reply anyway */
1194 ptlrpc_abort_bulk(desc);
1195 } else if (!desc->bd_success) {
1196 DEBUG_REQ(D_ERROR, req, "network error on bulk GET");
1197 /* XXX should this be a different errno? */
1200 rc = sptlrpc_svc_unwrap_bulk(req, desc);
1203 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc);
1207 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1208 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
1210 if (unlikely(client_cksum != 0 && rc == 0)) {
1211 static int cksum_counter;
1212 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1213 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
1214 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
1215 server_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1216 repbody->oa.o_cksum = server_cksum;
1218 if (unlikely(client_cksum != server_cksum)) {
1219 CERROR("client csum %x, server csum %x\n",
1220 client_cksum, server_cksum);
1222 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
1223 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
1224 cksum_counter, libcfs_id2str(req->rq_peer),
1229 /* Must commit after prep above in all cases */
1230 rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa, objcount, ioo,
1231 remote_nb, npages, local_nb, oti, rc);
1232 if (rc == -ENOTCONN)
1233 /* quota acquire process has been given up because
1234 * either the client has been evicted or the client
1235 * has timed out the request already */
1238 if (exp_connect_rmtclient(exp)) {
1239 repbody->oa.o_uid = o_uid;
1240 repbody->oa.o_gid = o_gid;
1244 * Disable sending mtime back to the client. If the client locked the
1245 * whole object, then it has already updated the mtime on its side,
1246 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
1248 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
1250 if (unlikely(client_cksum != server_cksum && rc == 0)) {
1251 int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1256 if (new_cksum == server_cksum)
1257 msg = "changed in transit before arrival at OST";
1258 else if (new_cksum == client_cksum)
1259 msg = "initial checksum before message complete";
1261 msg = "changed in transit AND after initial checksum";
1263 if (req->rq_peer.nid == desc->bd_sender) {
1267 router = libcfs_nid2str(desc->bd_sender);
1270 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from "
1271 "%s%s%s inode "DFID" object "
1272 LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1273 exp->exp_obd->obd_name, msg,
1274 libcfs_id2str(req->rq_peer),
1276 body->oa.o_valid & OBD_MD_FLFID ?
1277 body->oa.o_parent_seq : (__u64)0,
1278 body->oa.o_valid & OBD_MD_FLFID ?
1279 body->oa.o_parent_oid : 0,
1280 body->oa.o_valid & OBD_MD_FLFID ?
1281 body->oa.o_parent_ver : 0,
1283 body->oa.o_valid & OBD_MD_FLGROUP ?
1284 body->oa.o_seq : (__u64)0,
1286 local_nb[npages-1].offset +
1287 local_nb[npages-1].len - 1 );
1288 CERROR("client csum %x, original server csum %x, "
1289 "server csum now %x\n",
1290 client_cksum, server_cksum, new_cksum);
1296 /* set per-requested niobuf return codes */
1297 for (i = j = 0; i < niocount; i++) {
1298 int len = remote_nb[i].len;
1303 LASSERT(j < npages);
1304 if (local_nb[j].rc < 0)
1305 rcs[i] = local_nb[j].rc;
1306 len -= local_nb[j].len;
1311 LASSERT(j == npages);
1312 ptlrpc_lprocfs_brw(req, nob);
1316 ost_brw_lock_put(LCK_PW, ioo, remote_nb, &lockh);
1321 ptlrpc_free_bulk(desc);
1323 /* XXX: don't send reply if obd rdonly mode, this can cause data loss
1324 * on client, see bug 22190. Remove this when async bulk will be done.
1325 * Meanwhile, if this is umount then don't reply anything. */
1326 if (req->rq_export->exp_obd->obd_no_transno) {
1327 no_reply = req->rq_export->exp_obd->obd_stopping;
1332 oti_to_request(oti, req);
1333 target_committed_to_req(req);
1334 rc = ptlrpc_reply(req);
1335 } else if (!no_reply) {
1336 /* Only reply if there was no comms problem with bulk */
1337 target_committed_to_req(req);
1338 req->rq_status = rc;
1341 /* reply out callback would free */
1342 ptlrpc_req_drop_rs(req);
1343 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
1344 "client will retry\n",
1345 exp->exp_obd->obd_name,
1346 exp->exp_client_uuid.uuid,
1347 exp->exp_connection->c_remote_uuid.uuid,
1348 libcfs_id2str(req->rq_peer));
1350 cfs_memory_pressure_clr();
1355 * Implementation of OST_SET_INFO.
1357 * OST_SET_INFO is like ioctl(): heavily overloaded. Specifically, it takes a
1358 * "key" and a value RPC buffers as arguments, with the value's contents
1359 * interpreted according to the key.
1361 * Value types that need swabbing have swabbing done explicitly, either here or
1362 * in functions called from here. This should be corrected: all swabbing should
1363 * be done in the capsule abstraction, as that will then allow us to move
1364 * swabbing exclusively to the client without having to modify server code
1365 * outside the capsule abstraction's implementation itself. To correct this
1366 * will require minor changes to the capsule abstraction; see the comments for
1367 * req_capsule_extend() in layout.c.
1369 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
1371 struct ost_body *body = NULL, *repbody;
1372 char *key, *val = NULL;
1373 int keylen, vallen, rc = 0;
1374 int is_grant_shrink = 0;
1377 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1379 DEBUG_REQ(D_HA, req, "no set_info key");
1382 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1385 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1388 if ((is_grant_shrink = KEY_IS(KEY_GRANT_SHRINK)))
1389 /* In this case the value is actually an RMF_OST_BODY, so we
1390 * transmutate the type of this PTLRPC */
1391 req_capsule_extend(&req->rq_pill, &RQF_OST_SET_GRANT_INFO);
1393 rc = req_capsule_server_pack(&req->rq_pill);
1398 if (is_grant_shrink) {
1399 body = req_capsule_client_get(&req->rq_pill,
1404 repbody = req_capsule_server_get(&req->rq_pill,
1406 memcpy(repbody, body, sizeof(*body));
1407 val = (char*)repbody;
1409 val = req_capsule_client_get(&req->rq_pill,
1414 if (KEY_IS(KEY_EVICT_BY_NID)) {
1416 obd_export_evict_by_nid(exp->exp_obd, val);
1418 } else if (KEY_IS(KEY_MDS_CONN) && ptlrpc_req_need_swab(req)) {
1419 if (vallen < sizeof(__u32))
1421 __swab32s((__u32 *)val);
1424 /* OBD will also check if KEY_IS(KEY_GRANT_SHRINK), and will cast val to
1425 * a struct ost_body * value */
1426 rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL);
1428 lustre_msg_set_status(req->rq_repmsg, 0);
1432 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
1435 int keylen, replylen, rc = 0;
1436 struct req_capsule *pill = &req->rq_pill;
1439 /* this common part for get_info rpc */
1440 key = req_capsule_client_get(pill, &RMF_SETINFO_KEY);
1442 DEBUG_REQ(D_HA, req, "no get_info key");
1445 keylen = req_capsule_get_size(pill, &RMF_SETINFO_KEY, RCL_CLIENT);
1447 if (KEY_IS(KEY_FIEMAP)) {
1448 struct ll_fiemap_info_key *fm_key = key;
1451 rc = ost_validate_obdo(exp, &fm_key->oa, NULL);
1456 rc = obd_get_info(exp, keylen, key, &replylen, NULL, NULL);
1460 req_capsule_set_size(pill, &RMF_GENERIC_DATA,
1461 RCL_SERVER, replylen);
1463 rc = req_capsule_server_pack(pill);
1467 reply = req_capsule_server_get(pill, &RMF_GENERIC_DATA);
1471 /* call again to fill in the reply buffer */
1472 rc = obd_get_info(exp, keylen, key, &replylen, reply, NULL);
1474 lustre_msg_set_status(req->rq_repmsg, 0);
1478 #ifdef HAVE_QUOTA_SUPPORT
1479 static int ost_handle_quotactl(struct ptlrpc_request *req)
1481 struct obd_quotactl *oqctl, *repoqc;
1485 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1487 GOTO(out, rc = -EPROTO);
1489 rc = req_capsule_server_pack(&req->rq_pill);
1493 repoqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1494 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1501 static int ost_handle_quotacheck(struct ptlrpc_request *req)
1503 struct obd_quotactl *oqctl;
1507 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1511 rc = req_capsule_server_pack(&req->rq_pill);
1515 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1519 static int ost_handle_quota_adjust_qunit(struct ptlrpc_request *req)
1521 struct quota_adjust_qunit *oqaq, *repoqa;
1522 struct lustre_quota_ctxt *qctxt;
1526 qctxt = &req->rq_export->exp_obd->u.obt.obt_qctxt;
1527 oqaq = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
1529 GOTO(out, rc = -EPROTO);
1531 rc = req_capsule_server_pack(&req->rq_pill);
1535 repoqa = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
1536 req->rq_status = obd_quota_adjust_qunit(req->rq_export, oqaq, qctxt);
1544 static int ost_llog_handle_connect(struct obd_export *exp,
1545 struct ptlrpc_request *req)
1547 struct llogd_conn_body *body;
1551 body = req_capsule_client_get(&req->rq_pill, &RMF_LLOGD_CONN_BODY);
1552 rc = obd_llog_connect(exp, body);
1556 #define ost_init_sec_none(reply, exp) \
1558 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \
1559 OBD_CONNECT_RMT_CLIENT_FORCE | \
1560 OBD_CONNECT_OSS_CAPA); \
1561 cfs_spin_lock(&exp->exp_lock); \
1562 exp->exp_connect_flags = reply->ocd_connect_flags; \
1563 cfs_spin_unlock(&exp->exp_lock); \
1566 static int ost_init_sec_level(struct ptlrpc_request *req)
1568 struct obd_export *exp = req->rq_export;
1569 struct req_capsule *pill = &req->rq_pill;
1570 struct obd_device *obd = exp->exp_obd;
1571 struct filter_obd *filter = &obd->u.filter;
1572 char *client = libcfs_nid2str(req->rq_peer.nid);
1573 struct obd_connect_data *data, *reply;
1577 data = req_capsule_client_get(pill, &RMF_CONNECT_DATA);
1578 reply = req_capsule_server_get(pill, &RMF_CONNECT_DATA);
1579 if (data == NULL || reply == NULL)
1582 /* connection from MDT is always trusted */
1583 if (req->rq_auth_usr_mdt) {
1584 ost_init_sec_none(reply, exp);
1588 /* no GSS support case */
1589 if (!req->rq_auth_gss) {
1590 if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
1591 CWARN("client %s -> target %s does not user GSS, "
1592 "can not run under security level %d.\n",
1593 client, obd->obd_name, filter->fo_sec_level);
1596 ost_init_sec_none(reply, exp);
1601 /* old version case */
1602 if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) ||
1603 !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) {
1604 if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
1605 CWARN("client %s -> target %s uses old version, "
1606 "can not run under security level %d.\n",
1607 client, obd->obd_name, filter->fo_sec_level);
1610 CWARN("client %s -> target %s uses old version, "
1611 "run under security level %d.\n",
1612 client, obd->obd_name, filter->fo_sec_level);
1613 ost_init_sec_none(reply, exp);
1618 remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE;
1620 if (!req->rq_auth_remote)
1621 CDEBUG(D_SEC, "client (local realm) %s -> target %s "
1622 "asked to be remote.\n", client, obd->obd_name);
1623 } else if (req->rq_auth_remote) {
1625 CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set "
1626 "as remote by default.\n", client, obd->obd_name);
1630 if (!filter->fo_fl_oss_capa) {
1631 CDEBUG(D_SEC, "client %s -> target %s is set as remote,"
1632 " but OSS capabilities are not enabled: %d.\n",
1633 client, obd->obd_name, filter->fo_fl_oss_capa);
1638 switch (filter->fo_sec_level) {
1639 case LUSTRE_SEC_NONE:
1641 ost_init_sec_none(reply, exp);
1644 CDEBUG(D_SEC, "client %s -> target %s is set as remote, "
1645 "can not run under security level %d.\n",
1646 client, obd->obd_name, filter->fo_sec_level);
1649 case LUSTRE_SEC_REMOTE:
1651 ost_init_sec_none(reply, exp);
1653 case LUSTRE_SEC_ALL:
1655 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
1656 OBD_CONNECT_RMT_CLIENT_FORCE);
1657 if (!filter->fo_fl_oss_capa)
1658 reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA;
1660 cfs_spin_lock(&exp->exp_lock);
1661 exp->exp_connect_flags = reply->ocd_connect_flags;
1662 cfs_spin_unlock(&exp->exp_lock);
1674 * this should be done in filter_connect()/filter_reconnect(), but
1675 * we can't obtain information like NID, which stored in incoming
1676 * request, thus can't decide what flavor to use. so we do it here.
1678 * This hack should be removed after the OST stack be rewritten, just
1679 * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect().
1681 static int ost_connect_check_sptlrpc(struct ptlrpc_request *req)
1683 struct obd_export *exp = req->rq_export;
1684 struct filter_obd *filter = &exp->exp_obd->u.filter;
1685 struct sptlrpc_flavor flvr;
1688 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
1689 LUSTRE_ECHO_NAME) == 0)) {
1690 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
1694 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
1695 cfs_read_lock(&filter->fo_sptlrpc_lock);
1696 sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset,
1700 cfs_read_unlock(&filter->fo_sptlrpc_lock);
1702 cfs_spin_lock(&exp->exp_lock);
1704 exp->exp_sp_peer = req->rq_sp_from;
1705 exp->exp_flvr = flvr;
1707 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
1708 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
1709 CERROR("unauthorized rpc flavor %x from %s, "
1710 "expect %x\n", req->rq_flvr.sf_rpc,
1711 libcfs_nid2str(req->rq_peer.nid),
1712 exp->exp_flvr.sf_rpc);
1716 cfs_spin_unlock(&exp->exp_lock);
1718 if (exp->exp_sp_peer != req->rq_sp_from) {
1719 CERROR("RPC source %s doesn't match %s\n",
1720 sptlrpc_part2name(req->rq_sp_from),
1721 sptlrpc_part2name(exp->exp_sp_peer));
1724 rc = sptlrpc_target_export_check(exp, req);
1731 static int ost_filter_recovery_request(struct ptlrpc_request *req,
1732 struct obd_device *obd, int *process)
1734 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1735 case OST_CONNECT: /* This will never get here, but for completeness. */
1736 case OST_DISCONNECT:
1747 case OBD_LOG_CANCEL:
1749 *process = target_queue_recovery_request(req, obd);
1753 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1759 int ost_msg_check_version(struct lustre_msg *msg)
1763 switch(lustre_msg_get_opc(msg)) {
1765 case OST_DISCONNECT:
1768 case SEC_CTX_INIT_CONT:
1770 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1772 CERROR("bad opc %u version %08x, expecting %08x\n",
1773 lustre_msg_get_opc(msg),
1774 lustre_msg_get_version(msg),
1775 LUSTRE_OBD_VERSION);
1788 #ifdef HAVE_QUOTA_SUPPORT
1789 case OST_QUOTACHECK:
1791 case OST_QUOTA_ADJUST_QUNIT:
1793 rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION);
1795 CERROR("bad opc %u version %08x, expecting %08x\n",
1796 lustre_msg_get_opc(msg),
1797 lustre_msg_get_version(msg),
1798 LUSTRE_OST_VERSION);
1803 case LDLM_BL_CALLBACK:
1804 case LDLM_CP_CALLBACK:
1805 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1807 CERROR("bad opc %u version %08x, expecting %08x\n",
1808 lustre_msg_get_opc(msg),
1809 lustre_msg_get_version(msg),
1810 LUSTRE_DLM_VERSION);
1812 case LLOG_ORIGIN_CONNECT:
1813 case OBD_LOG_CANCEL:
1814 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1816 CERROR("bad opc %u version %08x, expecting %08x\n",
1817 lustre_msg_get_opc(msg),
1818 lustre_msg_get_version(msg),
1819 LUSTRE_LOG_VERSION);
1822 CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg));
1829 * Returns 1 if the given PTLRPC matches the given LDLM locks, or 0 if it does
1832 static int ost_rw_hpreq_lock_match(struct ptlrpc_request *req,
1833 struct ldlm_lock *lock)
1835 struct niobuf_remote *nb;
1836 struct obd_ioobj *ioo;
1837 struct ost_body *body;
1838 int objcount, niocount;
1839 int mode, opc, i, rc;
1843 opc = lustre_msg_get_opc(req->rq_reqmsg);
1844 LASSERT(opc == OST_READ || opc == OST_WRITE);
1846 /* As the request may be covered by several locks, do not look at
1847 * o_handle, look at the RPC IO region. */
1848 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1852 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1853 RCL_CLIENT) / sizeof(*ioo);
1854 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1858 rc = ost_validate_obdo(req->rq_export, &body->oa, ioo);
1862 for (niocount = i = 0; i < objcount; i++)
1863 niocount += ioo[i].ioo_bufcnt;
1865 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1867 niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE,
1868 RCL_CLIENT) / sizeof(*nb)))
1872 if (opc == OST_READ)
1875 start = nb[0].offset & CFS_PAGE_MASK;
1876 end = (nb[ioo->ioo_bufcnt - 1].offset +
1877 nb[ioo->ioo_bufcnt - 1].len - 1) | ~CFS_PAGE_MASK;
1879 LASSERT(lock->l_resource != NULL);
1880 if (!osc_res_name_eq(ioo->ioo_id, ioo->ioo_seq,
1881 &lock->l_resource->lr_name))
1884 if (!(lock->l_granted_mode & mode))
1887 if (lock->l_policy_data.l_extent.end < start ||
1888 lock->l_policy_data.l_extent.start > end)
1895 * High-priority queue request check for whether the given PTLRPC request (\a
1896 * req) is blocking an LDLM lock cancel.
1898 * Returns 1 if the given given PTLRPC request (\a req) is blocking an LDLM lock
1899 * cancel, 0 if it is not, and -EFAULT if the request is malformed.
1901 * Only OST_READs, OST_WRITEs and OST_PUNCHes go on the h-p RPC queue. This
1902 * function looks only at OST_READs and OST_WRITEs.
1904 static int ost_rw_hpreq_check(struct ptlrpc_request *req)
1906 struct niobuf_remote *nb;
1907 struct obd_ioobj *ioo;
1908 struct ost_body *body;
1909 int objcount, niocount;
1910 int mode, opc, i, rc;
1913 opc = lustre_msg_get_opc(req->rq_reqmsg);
1914 LASSERT(opc == OST_READ || opc == OST_WRITE);
1916 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1920 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1921 RCL_CLIENT) / sizeof(*ioo);
1922 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1926 rc = ost_validate_obdo(req->rq_export, &body->oa, ioo);
1930 for (niocount = i = 0; i < objcount; i++)
1931 niocount += ioo[i].ioo_bufcnt;
1932 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1934 niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE,
1935 RCL_CLIENT) / sizeof(*nb)))
1937 if (niocount != 0 && (nb[0].flags & OBD_BRW_SRVLOCK))
1941 if (opc == OST_READ)
1943 RETURN(ost_rw_prolong_locks(req, ioo, nb, &body->oa, mode));
1946 static int ost_punch_prolong_locks(struct ptlrpc_request *req, struct obdo *oa)
1948 struct ldlm_res_id res_id = { .name = { oa->o_id } };
1949 struct ost_prolong_data opd = { 0 };
1954 end = start + oa->o_blocks;
1956 opd.opd_mode = LCK_PW;
1957 opd.opd_exp = req->rq_export;
1958 opd.opd_policy.l_extent.start = start & CFS_PAGE_MASK;
1959 if (oa->o_blocks == OBD_OBJECT_EOF || end < start)
1960 opd.opd_policy.l_extent.end = OBD_OBJECT_EOF;
1962 opd.opd_policy.l_extent.end = end | ~CFS_PAGE_MASK;
1964 /* prolong locks for the current service time of the corresponding
1965 * portal (= OST_IO_PORTAL) */
1966 opd.opd_timeout = AT_OFF ? obd_timeout / 2:
1967 max(at_est2timeout(at_get(&req->rq_rqbd->
1968 rqbd_service->srv_at_estimate)), ldlm_timeout);
1970 CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
1971 res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
1972 opd.opd_policy.l_extent.end);
1975 ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id,
1976 ost_prolong_locks_iter, &opd);
1977 RETURN(opd.opd_lock_match);
1981 * Like ost_rw_hpreq_lock_match(), but for OST_PUNCH RPCs.
1983 static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req,
1984 struct ldlm_lock *lock)
1986 struct ost_body *body;
1990 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1992 RETURN(0); /* can't return -EFAULT here */
1994 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
1998 if (body->oa.o_valid & OBD_MD_FLHANDLE &&
1999 body->oa.o_handle.cookie == lock->l_handle.h_cookie)
2005 * Like ost_rw_hpreq_check(), but for OST_PUNCH RPCs.
2007 static int ost_punch_hpreq_check(struct ptlrpc_request *req)
2009 struct ost_body *body;
2012 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2016 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
2020 LASSERT(!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
2021 !(body->oa.o_flags & OBD_FL_SRVLOCK));
2023 RETURN(ost_punch_prolong_locks(req, &body->oa));
2026 struct ptlrpc_hpreq_ops ost_hpreq_rw = {
2027 .hpreq_lock_match = ost_rw_hpreq_lock_match,
2028 .hpreq_check = ost_rw_hpreq_check,
2031 struct ptlrpc_hpreq_ops ost_hpreq_punch = {
2032 .hpreq_lock_match = ost_punch_hpreq_lock_match,
2033 .hpreq_check = ost_punch_hpreq_check,
2036 /** Assign high priority operations to the request if needed. */
2037 static int ost_hpreq_handler(struct ptlrpc_request *req)
2040 if (req->rq_export) {
2041 int opc = lustre_msg_get_opc(req->rq_reqmsg);
2042 struct ost_body *body;
2044 if (opc == OST_READ || opc == OST_WRITE) {
2045 struct niobuf_remote *nb;
2046 struct obd_ioobj *ioo;
2047 int objcount, niocount;
2050 /* RPCs on the H-P queue can be inspected before
2051 * ost_handler() initializes their pills, so we
2052 * initialize that here. Capsule initialization is
2053 * idempotent, as is setting the pill's format (provided
2054 * it doesn't change).
2056 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2057 if (opc == OST_READ)
2058 req_capsule_set(&req->rq_pill,
2061 req_capsule_set(&req->rq_pill,
2062 &RQF_OST_BRW_WRITE);
2064 body = req_capsule_client_get(&req->rq_pill,
2067 CERROR("Missing/short ost_body\n");
2071 objcount = req_capsule_get_size(&req->rq_pill,
2075 if (objcount == 0) {
2076 CERROR("Missing/short ioobj\n");
2080 CERROR("too many ioobjs (%d)\n", objcount);
2084 ioo = req_capsule_client_get(&req->rq_pill,
2087 CERROR("Missing/short ioobj\n");
2091 for (niocount = i = 0; i < objcount; i++) {
2092 if (ioo[i].ioo_bufcnt == 0) {
2093 CERROR("ioo[%d] has zero bufcnt\n", i);
2096 niocount += ioo[i].ioo_bufcnt;
2098 if (niocount > PTLRPC_MAX_BRW_PAGES) {
2099 DEBUG_REQ(D_RPCTRACE, req,
2100 "bulk has too many pages (%d)",
2105 nb = req_capsule_client_get(&req->rq_pill,
2106 &RMF_NIOBUF_REMOTE);
2108 CERROR("Missing/short niobuf\n");
2112 if (niocount == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
2113 req->rq_ops = &ost_hpreq_rw;
2114 } else if (opc == OST_PUNCH) {
2115 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2116 req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
2118 body = req_capsule_client_get(&req->rq_pill,
2121 CERROR("Missing/short ost_body\n");
2125 if (!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
2126 !(body->oa.o_flags & OBD_FL_SRVLOCK))
2127 req->rq_ops = &ost_hpreq_punch;
2133 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2134 int ost_handle(struct ptlrpc_request *req)
2136 struct obd_trans_info trans_info = { 0, };
2137 struct obd_trans_info *oti = &trans_info;
2138 int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
2139 struct obd_device *obd = NULL;
2142 LASSERT(current->journal_info == NULL);
2144 /* primordial rpcs don't affect server recovery */
2145 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2147 case SEC_CTX_INIT_CONT:
2152 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2154 if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) {
2155 if (!class_connected_export(req->rq_export)) {
2156 CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
2157 lustre_msg_get_opc(req->rq_reqmsg),
2158 libcfs_id2str(req->rq_peer));
2159 req->rq_status = -ENOTCONN;
2160 GOTO(out, rc = -ENOTCONN);
2163 obd = req->rq_export->exp_obd;
2165 /* Check for aborted recovery. */
2166 if (obd->obd_recovering) {
2167 rc = ost_filter_recovery_request(req, obd,
2169 if (rc || !should_process)
2171 else if (should_process < 0) {
2172 req->rq_status = should_process;
2173 rc = ptlrpc_error(req);
2181 rc = ost_msg_check_version(req->rq_reqmsg);
2185 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2187 CDEBUG(D_INODE, "connect\n");
2188 req_capsule_set(&req->rq_pill, &RQF_OST_CONNECT);
2189 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET))
2191 rc = target_handle_connect(req);
2192 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2))
2195 rc = ost_init_sec_level(req);
2197 rc = ost_connect_check_sptlrpc(req);
2201 case OST_DISCONNECT:
2202 CDEBUG(D_INODE, "disconnect\n");
2203 req_capsule_set(&req->rq_pill, &RQF_OST_DISCONNECT);
2204 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DISCONNECT_NET))
2206 rc = target_handle_disconnect(req);
2209 CDEBUG(D_INODE, "create\n");
2210 req_capsule_set(&req->rq_pill, &RQF_OST_CREATE);
2211 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET))
2213 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2214 GOTO(out, rc = -EROFS);
2215 rc = ost_create(req->rq_export, req, oti);
2218 CDEBUG(D_INODE, "destroy\n");
2219 req_capsule_set(&req->rq_pill, &RQF_OST_DESTROY);
2220 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET))
2222 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2223 GOTO(out, rc = -EROFS);
2224 rc = ost_destroy(req->rq_export, req, oti);
2227 CDEBUG(D_INODE, "getattr\n");
2228 req_capsule_set(&req->rq_pill, &RQF_OST_GETATTR);
2229 if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET))
2231 rc = ost_getattr(req->rq_export, req);
2234 CDEBUG(D_INODE, "setattr\n");
2235 req_capsule_set(&req->rq_pill, &RQF_OST_SETATTR);
2236 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET))
2238 rc = ost_setattr(req->rq_export, req, oti);
2241 req_capsule_set(&req->rq_pill, &RQF_OST_BRW_WRITE);
2242 CDEBUG(D_INODE, "write\n");
2243 /* req->rq_request_portal would be nice, if it was set */
2244 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
2245 CERROR("%s: deny write request from %s to portal %u\n",
2246 req->rq_export->exp_obd->obd_name,
2247 obd_export_nid2str(req->rq_export),
2248 req->rq_rqbd->rqbd_service->srv_req_portal);
2249 GOTO(out, rc = -EPROTO);
2251 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
2253 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2254 GOTO(out, rc = -ENOSPC);
2255 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2256 GOTO(out, rc = -EROFS);
2257 rc = ost_brw_write(req, oti);
2258 LASSERT(current->journal_info == NULL);
2259 /* ost_brw_write sends its own replies */
2262 req_capsule_set(&req->rq_pill, &RQF_OST_BRW_READ);
2263 CDEBUG(D_INODE, "read\n");
2264 /* req->rq_request_portal would be nice, if it was set */
2265 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
2266 CERROR("%s: deny read request from %s to portal %u\n",
2267 req->rq_export->exp_obd->obd_name,
2268 obd_export_nid2str(req->rq_export),
2269 req->rq_rqbd->rqbd_service->srv_req_portal);
2270 GOTO(out, rc = -EPROTO);
2272 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
2274 rc = ost_brw_read(req, oti);
2275 LASSERT(current->journal_info == NULL);
2276 /* ost_brw_read sends its own replies */
2279 CDEBUG(D_INODE, "punch\n");
2280 req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
2281 if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET))
2283 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2284 GOTO(out, rc = -EROFS);
2285 rc = ost_punch(req->rq_export, req, oti);
2288 CDEBUG(D_INODE, "statfs\n");
2289 req_capsule_set(&req->rq_pill, &RQF_OST_STATFS);
2290 if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET))
2292 rc = ost_statfs(req);
2295 CDEBUG(D_INODE, "sync\n");
2296 req_capsule_set(&req->rq_pill, &RQF_OST_SYNC);
2297 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET))
2299 rc = ost_sync(req->rq_export, req);
2302 DEBUG_REQ(D_INODE, req, "set_info");
2303 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2304 rc = ost_set_info(req->rq_export, req);
2307 DEBUG_REQ(D_INODE, req, "get_info");
2308 req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC);
2309 rc = ost_get_info(req->rq_export, req);
2311 #ifdef HAVE_QUOTA_SUPPORT
2312 case OST_QUOTACHECK:
2313 CDEBUG(D_INODE, "quotacheck\n");
2314 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK);
2315 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACHECK_NET))
2317 rc = ost_handle_quotacheck(req);
2320 CDEBUG(D_INODE, "quotactl\n");
2321 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACTL);
2322 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACTL_NET))
2324 rc = ost_handle_quotactl(req);
2326 case OST_QUOTA_ADJUST_QUNIT:
2327 CDEBUG(D_INODE, "quota_adjust_qunit\n");
2328 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTA_ADJUST_QUNIT);
2329 rc = ost_handle_quota_adjust_qunit(req);
2333 DEBUG_REQ(D_INODE, req, "ping");
2334 req_capsule_set(&req->rq_pill, &RQF_OBD_PING);
2335 rc = target_handle_ping(req);
2337 /* FIXME - just reply status */
2338 case LLOG_ORIGIN_CONNECT:
2339 DEBUG_REQ(D_INODE, req, "log connect");
2340 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_CONNECT);
2341 rc = ost_llog_handle_connect(req->rq_export, req);
2342 req->rq_status = rc;
2343 rc = req_capsule_server_pack(&req->rq_pill);
2346 RETURN(ptlrpc_reply(req));
2347 case OBD_LOG_CANCEL:
2348 CDEBUG(D_INODE, "log cancel\n");
2349 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2350 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2352 rc = llog_origin_handle_cancel(req);
2353 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2355 req->rq_status = rc;
2356 rc = req_capsule_server_pack(&req->rq_pill);
2359 RETURN(ptlrpc_reply(req));
2361 CDEBUG(D_INODE, "enqueue\n");
2362 req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE);
2363 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE))
2365 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2366 ldlm_server_blocking_ast,
2367 ldlm_server_glimpse_ast);
2368 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
2371 CDEBUG(D_INODE, "convert\n");
2372 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2373 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT))
2375 rc = ldlm_handle_convert(req);
2378 CDEBUG(D_INODE, "cancel\n");
2379 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2380 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2382 rc = ldlm_handle_cancel(req);
2384 case LDLM_BL_CALLBACK:
2385 case LDLM_CP_CALLBACK:
2386 CDEBUG(D_INODE, "callback\n");
2387 CERROR("callbacks should not happen on OST\n");
2390 CERROR("Unexpected opcode %d\n",
2391 lustre_msg_get_opc(req->rq_reqmsg));
2392 req->rq_status = -ENOTSUPP;
2393 rc = ptlrpc_error(req);
2397 LASSERT(current->journal_info == NULL);
2400 /* If we're DISCONNECTing, the export_data is already freed */
2401 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT)
2402 target_committed_to_req(req);
2406 oti_to_request(oti, req);
2408 target_send_reply(req, rc, fail);
2411 EXPORT_SYMBOL(ost_handle);
2413 * free per-thread pool created by ost_thread_init().
2415 static void ost_thread_done(struct ptlrpc_thread *thread)
2417 struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local
2422 LASSERT(thread != NULL);
2425 * be prepared to handle partially-initialized pools (because this is
2426 * called from ost_thread_init() for cleanup.
2428 tls = thread->t_data;
2431 thread->t_data = NULL;
2437 * initialize per-thread page pool (bug 5137).
2439 static int ost_thread_init(struct ptlrpc_thread *thread)
2441 struct ost_thread_local_cache *tls;
2445 LASSERT(thread != NULL);
2446 LASSERT(thread->t_data == NULL);
2447 LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
2452 thread->t_data = tls;
2456 #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
2458 /* Sigh - really, this is an OSS, the _server_, not the _target_ */
2459 static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2461 struct ost_obd *ost = &obd->u.ost;
2462 struct lprocfs_static_vars lvars;
2463 int oss_min_threads;
2464 int oss_max_threads;
2465 int oss_min_create_threads;
2466 int oss_max_create_threads;
2470 rc = cfs_cleanup_group_info();
2474 lprocfs_ost_init_vars(&lvars);
2475 lprocfs_obd_setup(obd, lvars.obd_vars);
2477 cfs_sema_init(&ost->ost_health_sem, 1);
2479 if (oss_num_threads) {
2480 /* If oss_num_threads is set, it is the min and the max. */
2481 if (oss_num_threads > OSS_THREADS_MAX)
2482 oss_num_threads = OSS_THREADS_MAX;
2483 if (oss_num_threads < OSS_THREADS_MIN)
2484 oss_num_threads = OSS_THREADS_MIN;
2485 oss_max_threads = oss_min_threads = oss_num_threads;
2487 /* Base min threads on memory and cpus */
2489 cfs_num_possible_cpus() * CFS_NUM_CACHEPAGES >>
2490 (27 - CFS_PAGE_SHIFT);
2491 if (oss_min_threads < OSS_THREADS_MIN)
2492 oss_min_threads = OSS_THREADS_MIN;
2493 /* Insure a 4x range for dynamic threads */
2494 if (oss_min_threads > OSS_THREADS_MAX / 4)
2495 oss_min_threads = OSS_THREADS_MAX / 4;
2496 oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4 + 1);
2500 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2501 OST_MAXREPSIZE, OST_REQUEST_PORTAL,
2502 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2503 ost_handle, LUSTRE_OSS_NAME,
2504 obd->obd_proc_entry, target_print_req,
2505 oss_min_threads, oss_max_threads,
2506 "ll_ost", LCT_DT_THREAD, NULL);
2507 if (ost->ost_service == NULL) {
2508 CERROR("failed to start service\n");
2509 GOTO(out_lprocfs, rc = -ENOMEM);
2512 rc = ptlrpc_start_threads(obd, ost->ost_service);
2514 GOTO(out_service, rc = -EINVAL);
2516 if (oss_num_create_threads) {
2517 if (oss_num_create_threads > OSS_MAX_CREATE_THREADS)
2518 oss_num_create_threads = OSS_MAX_CREATE_THREADS;
2519 if (oss_num_create_threads < OSS_MIN_CREATE_THREADS)
2520 oss_num_create_threads = OSS_MIN_CREATE_THREADS;
2521 oss_min_create_threads = oss_max_create_threads =
2522 oss_num_create_threads;
2524 oss_min_create_threads = OSS_MIN_CREATE_THREADS;
2525 oss_max_create_threads = OSS_MAX_CREATE_THREADS;
2528 ost->ost_create_service =
2529 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2530 OST_MAXREPSIZE, OST_CREATE_PORTAL,
2531 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2532 ost_handle, "ost_create",
2533 obd->obd_proc_entry, target_print_req,
2534 oss_min_create_threads, oss_max_create_threads,
2535 "ll_ost_creat", LCT_DT_THREAD, NULL);
2536 if (ost->ost_create_service == NULL) {
2537 CERROR("failed to start OST create service\n");
2538 GOTO(out_service, rc = -ENOMEM);
2541 rc = ptlrpc_start_threads(obd, ost->ost_create_service);
2543 GOTO(out_create, rc = -EINVAL);
2545 ost->ost_io_service =
2546 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2547 OST_MAXREPSIZE, OST_IO_PORTAL,
2548 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2549 ost_handle, "ost_io",
2550 obd->obd_proc_entry, target_print_req,
2551 oss_min_threads, oss_max_threads,
2552 "ll_ost_io", LCT_DT_THREAD, ost_hpreq_handler);
2553 if (ost->ost_io_service == NULL) {
2554 CERROR("failed to start OST I/O service\n");
2555 GOTO(out_create, rc = -ENOMEM);
2558 ost->ost_io_service->srv_init = ost_thread_init;
2559 ost->ost_io_service->srv_done = ost_thread_done;
2560 ost->ost_io_service->srv_cpu_affinity = 1;
2561 rc = ptlrpc_start_threads(obd, ost->ost_io_service);
2563 GOTO(out_io, rc = -EINVAL);
2565 ping_evictor_start();
2570 ptlrpc_unregister_service(ost->ost_io_service);
2571 ost->ost_io_service = NULL;
2573 ptlrpc_unregister_service(ost->ost_create_service);
2574 ost->ost_create_service = NULL;
2576 ptlrpc_unregister_service(ost->ost_service);
2577 ost->ost_service = NULL;
2579 lprocfs_obd_cleanup(obd);
2583 static int ost_cleanup(struct obd_device *obd)
2585 struct ost_obd *ost = &obd->u.ost;
2589 ping_evictor_stop();
2591 cfs_spin_lock_bh(&obd->obd_processing_task_lock);
2592 if (obd->obd_recovering) {
2593 target_cancel_recovery_timer(obd);
2594 obd->obd_recovering = 0;
2596 cfs_spin_unlock_bh(&obd->obd_processing_task_lock);
2598 cfs_down(&ost->ost_health_sem);
2599 ptlrpc_unregister_service(ost->ost_service);
2600 ptlrpc_unregister_service(ost->ost_create_service);
2601 ptlrpc_unregister_service(ost->ost_io_service);
2602 ost->ost_service = NULL;
2603 ost->ost_create_service = NULL;
2604 cfs_up(&ost->ost_health_sem);
2606 lprocfs_obd_cleanup(obd);
2611 static int ost_health_check(struct obd_device *obd)
2613 struct ost_obd *ost = &obd->u.ost;
2616 cfs_down(&ost->ost_health_sem);
2617 rc |= ptlrpc_service_health_check(ost->ost_service);
2618 rc |= ptlrpc_service_health_check(ost->ost_create_service);
2619 rc |= ptlrpc_service_health_check(ost->ost_io_service);
2620 cfs_up(&ost->ost_health_sem);
2623 * health_check to return 0 on healthy
2624 * and 1 on unhealthy.
2632 struct ost_thread_local_cache *ost_tls(struct ptlrpc_request *r)
2634 return (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
2637 /* use obd ops to offer management infrastructure */
2638 static struct obd_ops ost_obd_ops = {
2639 .o_owner = THIS_MODULE,
2640 .o_setup = ost_setup,
2641 .o_cleanup = ost_cleanup,
2642 .o_health_check = ost_health_check,
2646 static int __init ost_init(void)
2648 struct lprocfs_static_vars lvars;
2652 lprocfs_ost_init_vars(&lvars);
2653 rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
2654 LUSTRE_OSS_NAME, NULL);
2656 if (ost_num_threads != 0 && oss_num_threads == 0) {
2657 LCONSOLE_INFO("ost_num_threads module parameter is deprecated, "
2658 "use oss_num_threads instead or unset both for "
2659 "dynamic thread startup\n");
2660 oss_num_threads = ost_num_threads;
2666 static void /*__exit*/ ost_exit(void)
2668 class_unregister_type(LUSTRE_OSS_NAME);
2671 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2672 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
2673 MODULE_LICENSE("GPL");
2675 module_init(ost_init);
2676 module_exit(ost_exit);