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_commitrw, 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);
950 /* Must commit after prep above in all cases */
951 rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1, ioo,
952 remote_nb, npages, local_nb, oti, rc);
955 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
956 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
957 ost_drop_id(exp, &repbody->oa);
961 ost_brw_lock_put(LCK_PR, ioo, remote_nb, &lockh);
966 ptlrpc_free_bulk(desc);
970 req->rq_status = nob;
971 ptlrpc_lprocfs_brw(req, nob);
972 target_committed_to_req(req);
974 } else if (!no_reply) {
975 /* Only reply if there was no comms problem with bulk */
976 target_committed_to_req(req);
980 /* reply out callback would free */
981 ptlrpc_req_drop_rs(req);
982 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
983 "client will retry\n",
984 exp->exp_obd->obd_name,
985 exp->exp_client_uuid.uuid,
986 exp->exp_connection->c_remote_uuid.uuid,
987 libcfs_id2str(req->rq_peer));
993 static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
995 struct ptlrpc_bulk_desc *desc = NULL;
996 struct obd_export *exp = req->rq_export;
997 struct niobuf_remote *remote_nb;
998 struct niobuf_local *local_nb;
999 struct obd_ioobj *ioo;
1000 struct ost_body *body, *repbody;
1001 struct l_wait_info lwi;
1002 struct lustre_handle lockh = {0};
1003 struct lustre_capa *capa = NULL;
1005 int objcount, niocount, npages;
1007 obd_count client_cksum = 0, server_cksum = 0;
1008 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
1009 int no_reply = 0, mmap = 0;
1010 __u32 o_uid = 0, o_gid = 0;
1011 struct ost_thread_local_cache *tls;
1014 req->rq_bulk_write = 1;
1016 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
1017 GOTO(out, rc = -EIO);
1018 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
1019 GOTO(out, rc = -EFAULT);
1021 /* pause before transaction has been started */
1022 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
1024 /* Check if there is eviction in progress, and if so, wait for it to
1026 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
1027 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
1028 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
1029 !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
1032 if (exp->exp_failed)
1033 GOTO(out, rc = -ENOTCONN);
1035 /* ost_body, ioobj & noibuf_remote are verified and swabbed in
1036 * ost_rw_hpreq_check(). */
1037 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1039 GOTO(out, rc = -EFAULT);
1041 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1042 RCL_CLIENT) / sizeof(*ioo);
1043 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1045 GOTO(out, rc = -EFAULT);
1047 rc = ost_validate_obdo(exp, &body->oa, ioo);
1051 for (niocount = i = 0; i < objcount; i++)
1052 niocount += ioo[i].ioo_bufcnt;
1055 * It'd be nice to have a capsule function to indicate how many elements
1056 * there were in a buffer for an RMF that's declared to be an array.
1057 * It's easy enough to compute the number of elements here though.
1059 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1060 if (remote_nb == NULL || niocount != (req_capsule_get_size(&req->rq_pill,
1061 &RMF_NIOBUF_REMOTE, RCL_CLIENT) / sizeof(*remote_nb)))
1062 GOTO(out, rc = -EFAULT);
1064 if ((remote_nb[0].flags & OBD_BRW_MEMALLOC) &&
1065 (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
1066 cfs_memory_pressure_set();
1068 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
1069 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
1071 CERROR("Missing capability for OST BRW WRITE");
1072 GOTO(out, rc = -EFAULT);
1076 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
1077 niocount * sizeof(*rcs));
1078 rc = req_capsule_server_pack(&req->rq_pill);
1081 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, obd_fail_val);
1082 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
1084 tls = ost_tls_get(req);
1086 GOTO(out_bulk, rc = -ENOMEM);
1087 local_nb = tls->local;
1089 rc = ost_brw_lock_get(LCK_PW, exp, ioo, remote_nb, &lockh);
1094 * If getting the lock took more time than
1095 * client was willing to wait, drop it. b=11330
1097 if (cfs_time_current_sec() > req->rq_deadline ||
1098 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1100 CERROR("Dropping timed-out write from %s because locking "
1101 "object "LPX64" took %ld seconds (limit was %ld).\n",
1102 libcfs_id2str(req->rq_peer), ioo->ioo_id,
1103 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
1104 req->rq_deadline - req->rq_arrival_time.tv_sec);
1105 GOTO(out_lock, rc = -ETIMEDOUT);
1108 if (!lustre_handle_is_used(&lockh))
1109 /* no needs to try to prolong lock if server is asked
1110 * to handle locking (= OBD_BRW_SRVLOCK) */
1111 ost_rw_prolong_locks(req, ioo, remote_nb,&body->oa, LCK_PW);
1113 /* obd_preprw clobbers oa->valid, so save what we need */
1114 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1115 client_cksum = body->oa.o_cksum;
1116 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1117 cksum_type = cksum_type_unpack(body->oa.o_flags);
1119 if (body->oa.o_valid & OBD_MD_FLFLAGS && body->oa.o_flags & OBD_FL_MMAP)
1122 /* Because we already sync grant info with client when reconnect,
1123 * grant info will be cleared for resent req, then fed_grant and
1124 * total_grant will not be modified in following preprw_write */
1125 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
1126 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
1127 body->oa.o_valid &= ~OBD_MD_FLGRANT;
1130 if (exp_connect_rmtclient(exp)) {
1131 o_uid = body->oa.o_uid;
1132 o_gid = body->oa.o_gid;
1134 npages = OST_THREAD_POOL_SIZE;
1135 rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount,
1136 ioo, remote_nb, &npages, local_nb, oti, capa);
1140 desc = ptlrpc_prep_bulk_exp(req, npages,
1141 BULK_GET_SINK, OST_BULK_PORTAL);
1143 GOTO(skip_transfer, rc = -ENOMEM);
1145 /* NB Having prepped, we must commit... */
1147 for (i = 0; i < npages; i++)
1148 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
1149 local_nb[i].offset & ~CFS_PAGE_MASK,
1152 rc = sptlrpc_svc_prep_bulk(req, desc);
1156 /* Check if client was evicted or tried to reconnect while we
1157 * were doing i/o before touching network */
1158 if (desc->bd_export->exp_failed ||
1159 desc->bd_export->exp_abort_active_req)
1162 rc = ptlrpc_start_bulk_transfer(desc);
1164 time_t start = cfs_time_current_sec();
1166 long timeoutl = req->rq_deadline -
1167 cfs_time_current_sec();
1168 cfs_duration_t timeout = timeoutl <= 0 ?
1169 CFS_TICK : cfs_time_seconds(timeoutl);
1170 lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
1171 ost_bulk_timeout, desc);
1172 rc = l_wait_event(desc->bd_waitq,
1173 !ptlrpc_server_bulk_active(desc) ||
1174 desc->bd_export->exp_failed ||
1175 desc->bd_export->exp_abort_active_req,
1177 LASSERT(rc == 0 || rc == -ETIMEDOUT);
1178 /* Wait again if we changed deadline */
1179 } while ((rc == -ETIMEDOUT) &&
1180 (req->rq_deadline > cfs_time_current_sec()));
1182 if (rc == -ETIMEDOUT) {
1183 DEBUG_REQ(D_ERROR, req,
1184 "timeout on bulk GET after %ld%+lds",
1185 req->rq_deadline - start,
1186 cfs_time_current_sec() -
1188 ptlrpc_abort_bulk(desc);
1189 } else if (desc->bd_export->exp_failed) {
1190 DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET");
1192 ptlrpc_abort_bulk(desc);
1193 } else if (desc->bd_export->exp_abort_active_req) {
1194 DEBUG_REQ(D_ERROR, req, "Reconnect on bulk GET");
1195 /* we don't reply anyway */
1197 ptlrpc_abort_bulk(desc);
1198 } else if (!desc->bd_success) {
1199 DEBUG_REQ(D_ERROR, req, "network error on bulk GET");
1200 /* XXX should this be a different errno? */
1203 rc = sptlrpc_svc_unwrap_bulk(req, desc);
1206 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc);
1211 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
1212 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
1214 if (unlikely(client_cksum != 0 && rc == 0)) {
1215 static int cksum_counter;
1216 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1217 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
1218 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
1219 server_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1220 repbody->oa.o_cksum = server_cksum;
1222 if (unlikely(client_cksum != server_cksum)) {
1223 CDEBUG_LIMIT(mmap ? D_INFO : D_ERROR,
1224 "client csum %x, server csum %x\n",
1225 client_cksum, server_cksum);
1227 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
1228 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
1229 cksum_counter, libcfs_id2str(req->rq_peer),
1234 /* Must commit after prep above in all cases */
1235 rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa, objcount, ioo,
1236 remote_nb, npages, local_nb, oti, rc);
1237 if (rc == -ENOTCONN)
1238 /* quota acquire process has been given up because
1239 * either the client has been evicted or the client
1240 * has timed out the request already */
1243 if (exp_connect_rmtclient(exp)) {
1244 repbody->oa.o_uid = o_uid;
1245 repbody->oa.o_gid = o_gid;
1249 * Disable sending mtime back to the client. If the client locked the
1250 * whole object, then it has already updated the mtime on its side,
1251 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
1253 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
1255 if (unlikely(client_cksum != server_cksum && rc == 0 && !mmap)) {
1256 int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1261 if (new_cksum == server_cksum)
1262 msg = "changed in transit before arrival at OST";
1263 else if (new_cksum == client_cksum)
1264 msg = "initial checksum before message complete";
1266 msg = "changed in transit AND after initial checksum";
1268 if (req->rq_peer.nid == desc->bd_sender) {
1272 router = libcfs_nid2str(desc->bd_sender);
1275 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from "
1276 "%s%s%s inode "DFID" object "
1277 LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1278 exp->exp_obd->obd_name, msg,
1279 libcfs_id2str(req->rq_peer),
1281 body->oa.o_valid & OBD_MD_FLFID ?
1282 body->oa.o_parent_seq : (__u64)0,
1283 body->oa.o_valid & OBD_MD_FLFID ?
1284 body->oa.o_parent_oid : 0,
1285 body->oa.o_valid & OBD_MD_FLFID ?
1286 body->oa.o_parent_ver : 0,
1288 body->oa.o_valid & OBD_MD_FLGROUP ?
1289 body->oa.o_seq : (__u64)0,
1291 local_nb[npages-1].offset +
1292 local_nb[npages-1].len - 1 );
1293 CERROR("client csum %x, original server csum %x, "
1294 "server csum now %x\n",
1295 client_cksum, server_cksum, new_cksum);
1301 /* set per-requested niobuf return codes */
1302 for (i = j = 0; i < niocount; i++) {
1303 int len = remote_nb[i].len;
1308 LASSERT(j < npages);
1309 if (local_nb[j].rc < 0)
1310 rcs[i] = local_nb[j].rc;
1311 len -= local_nb[j].len;
1316 LASSERT(j == npages);
1317 ptlrpc_lprocfs_brw(req, nob);
1321 ost_brw_lock_put(LCK_PW, ioo, remote_nb, &lockh);
1326 ptlrpc_free_bulk(desc);
1329 oti_to_request(oti, req);
1330 target_committed_to_req(req);
1331 rc = ptlrpc_reply(req);
1332 } else if (!no_reply) {
1333 /* Only reply if there was no comms problem with bulk */
1334 target_committed_to_req(req);
1335 req->rq_status = rc;
1338 /* reply out callback would free */
1339 ptlrpc_req_drop_rs(req);
1340 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
1341 "client will retry\n",
1342 exp->exp_obd->obd_name,
1343 exp->exp_client_uuid.uuid,
1344 exp->exp_connection->c_remote_uuid.uuid,
1345 libcfs_id2str(req->rq_peer));
1347 cfs_memory_pressure_clr();
1352 * Implementation of OST_SET_INFO.
1354 * OST_SET_INFO is like ioctl(): heavily overloaded. Specifically, it takes a
1355 * "key" and a value RPC buffers as arguments, with the value's contents
1356 * interpreted according to the key.
1358 * Value types that need swabbing have swabbing done explicitly, either here or
1359 * in functions called from here. This should be corrected: all swabbing should
1360 * be done in the capsule abstraction, as that will then allow us to move
1361 * swabbing exclusively to the client without having to modify server code
1362 * outside the capsule abstraction's implementation itself. To correct this
1363 * will require minor changes to the capsule abstraction; see the comments for
1364 * req_capsule_extend() in layout.c.
1366 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
1368 struct ost_body *body = NULL, *repbody;
1369 char *key, *val = NULL;
1370 int keylen, vallen, rc = 0;
1371 int is_grant_shrink = 0;
1374 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1376 DEBUG_REQ(D_HA, req, "no set_info key");
1379 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1382 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1385 if ((is_grant_shrink = KEY_IS(KEY_GRANT_SHRINK)))
1386 /* In this case the value is actually an RMF_OST_BODY, so we
1387 * transmutate the type of this PTLRPC */
1388 req_capsule_extend(&req->rq_pill, &RQF_OST_SET_GRANT_INFO);
1390 rc = req_capsule_server_pack(&req->rq_pill);
1395 if (is_grant_shrink) {
1396 body = req_capsule_client_get(&req->rq_pill,
1401 repbody = req_capsule_server_get(&req->rq_pill,
1403 memcpy(repbody, body, sizeof(*body));
1404 val = (char*)repbody;
1406 val = req_capsule_client_get(&req->rq_pill,
1411 if (KEY_IS(KEY_EVICT_BY_NID)) {
1413 obd_export_evict_by_nid(exp->exp_obd, val);
1415 } else if (KEY_IS(KEY_MDS_CONN) && ptlrpc_req_need_swab(req)) {
1416 if (vallen < sizeof(__u32))
1418 __swab32s((__u32 *)val);
1421 /* OBD will also check if KEY_IS(KEY_GRANT_SHRINK), and will cast val to
1422 * a struct ost_body * value */
1423 rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL);
1425 lustre_msg_set_status(req->rq_repmsg, 0);
1429 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
1432 int keylen, replylen, rc = 0;
1433 struct req_capsule *pill = &req->rq_pill;
1436 /* this common part for get_info rpc */
1437 key = req_capsule_client_get(pill, &RMF_SETINFO_KEY);
1439 DEBUG_REQ(D_HA, req, "no get_info key");
1442 keylen = req_capsule_get_size(pill, &RMF_SETINFO_KEY, RCL_CLIENT);
1444 if (KEY_IS(KEY_FIEMAP)) {
1445 struct ll_fiemap_info_key *fm_key = key;
1448 rc = ost_validate_obdo(exp, &fm_key->oa, NULL);
1453 rc = obd_get_info(exp, keylen, key, &replylen, NULL, NULL);
1457 req_capsule_set_size(pill, &RMF_GENERIC_DATA,
1458 RCL_SERVER, replylen);
1460 rc = req_capsule_server_pack(pill);
1464 reply = req_capsule_server_get(pill, &RMF_GENERIC_DATA);
1468 /* call again to fill in the reply buffer */
1469 rc = obd_get_info(exp, keylen, key, &replylen, reply, NULL);
1471 lustre_msg_set_status(req->rq_repmsg, 0);
1475 #ifdef HAVE_QUOTA_SUPPORT
1476 static int ost_handle_quotactl(struct ptlrpc_request *req)
1478 struct obd_quotactl *oqctl, *repoqc;
1482 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1484 GOTO(out, rc = -EPROTO);
1486 rc = req_capsule_server_pack(&req->rq_pill);
1490 repoqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1491 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1498 static int ost_handle_quotacheck(struct ptlrpc_request *req)
1500 struct obd_quotactl *oqctl;
1504 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1508 rc = req_capsule_server_pack(&req->rq_pill);
1512 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1516 static int ost_handle_quota_adjust_qunit(struct ptlrpc_request *req)
1518 struct quota_adjust_qunit *oqaq, *repoqa;
1519 struct lustre_quota_ctxt *qctxt;
1523 qctxt = &req->rq_export->exp_obd->u.obt.obt_qctxt;
1524 oqaq = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
1526 GOTO(out, rc = -EPROTO);
1528 rc = req_capsule_server_pack(&req->rq_pill);
1532 repoqa = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
1533 req->rq_status = obd_quota_adjust_qunit(req->rq_export, oqaq, qctxt, NULL);
1541 static int ost_llog_handle_connect(struct obd_export *exp,
1542 struct ptlrpc_request *req)
1544 struct llogd_conn_body *body;
1548 body = req_capsule_client_get(&req->rq_pill, &RMF_LLOGD_CONN_BODY);
1549 rc = obd_llog_connect(exp, body);
1553 #define ost_init_sec_none(reply, exp) \
1555 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \
1556 OBD_CONNECT_RMT_CLIENT_FORCE | \
1557 OBD_CONNECT_OSS_CAPA); \
1558 cfs_spin_lock(&exp->exp_lock); \
1559 exp->exp_connect_flags = reply->ocd_connect_flags; \
1560 cfs_spin_unlock(&exp->exp_lock); \
1563 static int ost_init_sec_level(struct ptlrpc_request *req)
1565 struct obd_export *exp = req->rq_export;
1566 struct req_capsule *pill = &req->rq_pill;
1567 struct obd_device *obd = exp->exp_obd;
1568 struct filter_obd *filter = &obd->u.filter;
1569 char *client = libcfs_nid2str(req->rq_peer.nid);
1570 struct obd_connect_data *data, *reply;
1574 data = req_capsule_client_get(pill, &RMF_CONNECT_DATA);
1575 reply = req_capsule_server_get(pill, &RMF_CONNECT_DATA);
1576 if (data == NULL || reply == NULL)
1579 /* connection from MDT is always trusted */
1580 if (req->rq_auth_usr_mdt) {
1581 ost_init_sec_none(reply, exp);
1585 /* no GSS support case */
1586 if (!req->rq_auth_gss) {
1587 if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
1588 CWARN("client %s -> target %s does not user GSS, "
1589 "can not run under security level %d.\n",
1590 client, obd->obd_name, filter->fo_sec_level);
1593 ost_init_sec_none(reply, exp);
1598 /* old version case */
1599 if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) ||
1600 !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) {
1601 if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
1602 CWARN("client %s -> target %s uses old version, "
1603 "can not run under security level %d.\n",
1604 client, obd->obd_name, filter->fo_sec_level);
1607 CWARN("client %s -> target %s uses old version, "
1608 "run under security level %d.\n",
1609 client, obd->obd_name, filter->fo_sec_level);
1610 ost_init_sec_none(reply, exp);
1615 remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE;
1617 if (!req->rq_auth_remote)
1618 CDEBUG(D_SEC, "client (local realm) %s -> target %s "
1619 "asked to be remote.\n", client, obd->obd_name);
1620 } else if (req->rq_auth_remote) {
1622 CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set "
1623 "as remote by default.\n", client, obd->obd_name);
1627 if (!filter->fo_fl_oss_capa) {
1628 CDEBUG(D_SEC, "client %s -> target %s is set as remote,"
1629 " but OSS capabilities are not enabled: %d.\n",
1630 client, obd->obd_name, filter->fo_fl_oss_capa);
1635 switch (filter->fo_sec_level) {
1636 case LUSTRE_SEC_NONE:
1638 ost_init_sec_none(reply, exp);
1641 CDEBUG(D_SEC, "client %s -> target %s is set as remote, "
1642 "can not run under security level %d.\n",
1643 client, obd->obd_name, filter->fo_sec_level);
1646 case LUSTRE_SEC_REMOTE:
1648 ost_init_sec_none(reply, exp);
1650 case LUSTRE_SEC_ALL:
1652 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
1653 OBD_CONNECT_RMT_CLIENT_FORCE);
1654 if (!filter->fo_fl_oss_capa)
1655 reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA;
1657 cfs_spin_lock(&exp->exp_lock);
1658 exp->exp_connect_flags = reply->ocd_connect_flags;
1659 cfs_spin_unlock(&exp->exp_lock);
1671 * this should be done in filter_connect()/filter_reconnect(), but
1672 * we can't obtain information like NID, which stored in incoming
1673 * request, thus can't decide what flavor to use. so we do it here.
1675 * This hack should be removed after the OST stack be rewritten, just
1676 * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect().
1678 static int ost_connect_check_sptlrpc(struct ptlrpc_request *req)
1680 struct obd_export *exp = req->rq_export;
1681 struct filter_obd *filter = &exp->exp_obd->u.filter;
1682 struct sptlrpc_flavor flvr;
1685 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
1686 LUSTRE_ECHO_NAME) == 0)) {
1687 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
1691 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
1692 cfs_read_lock(&filter->fo_sptlrpc_lock);
1693 sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset,
1697 cfs_read_unlock(&filter->fo_sptlrpc_lock);
1699 cfs_spin_lock(&exp->exp_lock);
1701 exp->exp_sp_peer = req->rq_sp_from;
1702 exp->exp_flvr = flvr;
1704 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
1705 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
1706 CERROR("unauthorized rpc flavor %x from %s, "
1707 "expect %x\n", req->rq_flvr.sf_rpc,
1708 libcfs_nid2str(req->rq_peer.nid),
1709 exp->exp_flvr.sf_rpc);
1713 cfs_spin_unlock(&exp->exp_lock);
1715 if (exp->exp_sp_peer != req->rq_sp_from) {
1716 CERROR("RPC source %s doesn't match %s\n",
1717 sptlrpc_part2name(req->rq_sp_from),
1718 sptlrpc_part2name(exp->exp_sp_peer));
1721 rc = sptlrpc_target_export_check(exp, req);
1728 /* Ensure that data and metadata are synced to the disk when lock is cancelled
1730 int ost_blocking_ast(struct ldlm_lock *lock,
1731 struct ldlm_lock_desc *desc,
1732 void *data, int flag)
1734 __u32 sync_lock_cancel = 0;
1735 __u32 len = sizeof(sync_lock_cancel);
1739 rc = obd_get_info(lock->l_export, sizeof(KEY_SYNC_LOCK_CANCEL),
1740 KEY_SYNC_LOCK_CANCEL, &len, &sync_lock_cancel, NULL);
1742 if (!rc && flag == LDLM_CB_CANCELING &&
1743 (lock->l_granted_mode & (LCK_PW|LCK_GROUP)) &&
1744 (sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1745 (sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1746 lock->l_flags & LDLM_FL_CBPENDING))) {
1751 oa->o_id = lock->l_resource->lr_name.name[0];
1752 oa->o_seq = lock->l_resource->lr_name.name[1];
1753 oa->o_valid = OBD_MD_FLID|OBD_MD_FLGROUP;
1755 rc = obd_sync(lock->l_export, oa, NULL,
1756 lock->l_policy_data.l_extent.start,
1757 lock->l_policy_data.l_extent.end, NULL);
1759 CERROR("Error %d syncing data on lock cancel\n", rc);
1764 return ldlm_server_blocking_ast(lock, desc, data, flag);
1767 static int ost_filter_recovery_request(struct ptlrpc_request *req,
1768 struct obd_device *obd, int *process)
1770 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1771 case OST_CONNECT: /* This will never get here, but for completeness. */
1772 case OST_DISCONNECT:
1783 case OBD_LOG_CANCEL:
1785 *process = target_queue_recovery_request(req, obd);
1789 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1795 int ost_msg_check_version(struct lustre_msg *msg)
1799 switch(lustre_msg_get_opc(msg)) {
1801 case OST_DISCONNECT:
1804 case SEC_CTX_INIT_CONT:
1806 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1808 CERROR("bad opc %u version %08x, expecting %08x\n",
1809 lustre_msg_get_opc(msg),
1810 lustre_msg_get_version(msg),
1811 LUSTRE_OBD_VERSION);
1824 #ifdef HAVE_QUOTA_SUPPORT
1825 case OST_QUOTACHECK:
1827 case OST_QUOTA_ADJUST_QUNIT:
1829 rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION);
1831 CERROR("bad opc %u version %08x, expecting %08x\n",
1832 lustre_msg_get_opc(msg),
1833 lustre_msg_get_version(msg),
1834 LUSTRE_OST_VERSION);
1839 case LDLM_BL_CALLBACK:
1840 case LDLM_CP_CALLBACK:
1841 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1843 CERROR("bad opc %u version %08x, expecting %08x\n",
1844 lustre_msg_get_opc(msg),
1845 lustre_msg_get_version(msg),
1846 LUSTRE_DLM_VERSION);
1848 case LLOG_ORIGIN_CONNECT:
1849 case OBD_LOG_CANCEL:
1850 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1852 CERROR("bad opc %u version %08x, expecting %08x\n",
1853 lustre_msg_get_opc(msg),
1854 lustre_msg_get_version(msg),
1855 LUSTRE_LOG_VERSION);
1858 CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg));
1865 * Returns 1 if the given PTLRPC matches the given LDLM locks, or 0 if it does
1868 static int ost_rw_hpreq_lock_match(struct ptlrpc_request *req,
1869 struct ldlm_lock *lock)
1871 struct niobuf_remote *nb;
1872 struct obd_ioobj *ioo;
1873 struct ost_body *body;
1874 int objcount, niocount;
1875 int mode, opc, i, rc;
1879 opc = lustre_msg_get_opc(req->rq_reqmsg);
1880 LASSERT(opc == OST_READ || opc == OST_WRITE);
1882 /* As the request may be covered by several locks, do not look at
1883 * o_handle, look at the RPC IO region. */
1884 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1888 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1889 RCL_CLIENT) / sizeof(*ioo);
1890 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1894 rc = ost_validate_obdo(req->rq_export, &body->oa, ioo);
1898 for (niocount = i = 0; i < objcount; i++)
1899 niocount += ioo[i].ioo_bufcnt;
1901 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1903 niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE,
1904 RCL_CLIENT) / sizeof(*nb)))
1908 if (opc == OST_READ)
1911 start = nb[0].offset & CFS_PAGE_MASK;
1912 end = (nb[ioo->ioo_bufcnt - 1].offset +
1913 nb[ioo->ioo_bufcnt - 1].len - 1) | ~CFS_PAGE_MASK;
1915 LASSERT(lock->l_resource != NULL);
1916 if (!osc_res_name_eq(ioo->ioo_id, ioo->ioo_seq,
1917 &lock->l_resource->lr_name))
1920 if (!(lock->l_granted_mode & mode))
1923 if (lock->l_policy_data.l_extent.end < start ||
1924 lock->l_policy_data.l_extent.start > end)
1931 * High-priority queue request check for whether the given PTLRPC request (\a
1932 * req) is blocking an LDLM lock cancel.
1934 * Returns 1 if the given given PTLRPC request (\a req) is blocking an LDLM lock
1935 * cancel, 0 if it is not, and -EFAULT if the request is malformed.
1937 * Only OST_READs, OST_WRITEs and OST_PUNCHes go on the h-p RPC queue. This
1938 * function looks only at OST_READs and OST_WRITEs.
1940 static int ost_rw_hpreq_check(struct ptlrpc_request *req)
1942 struct niobuf_remote *nb;
1943 struct obd_ioobj *ioo;
1944 struct ost_body *body;
1945 int objcount, niocount;
1946 int mode, opc, i, rc;
1949 opc = lustre_msg_get_opc(req->rq_reqmsg);
1950 LASSERT(opc == OST_READ || opc == OST_WRITE);
1952 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1956 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
1957 RCL_CLIENT) / sizeof(*ioo);
1958 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1962 rc = ost_validate_obdo(req->rq_export, &body->oa, ioo);
1966 for (niocount = i = 0; i < objcount; i++)
1967 niocount += ioo[i].ioo_bufcnt;
1968 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1970 niocount != (req_capsule_get_size(&req->rq_pill, &RMF_NIOBUF_REMOTE,
1971 RCL_CLIENT) / sizeof(*nb)))
1973 if (niocount != 0 && (nb[0].flags & OBD_BRW_SRVLOCK))
1977 if (opc == OST_READ)
1979 RETURN(ost_rw_prolong_locks(req, ioo, nb, &body->oa, mode));
1982 static int ost_punch_prolong_locks(struct ptlrpc_request *req, struct obdo *oa)
1984 struct ldlm_res_id res_id = { .name = { oa->o_id } };
1985 struct ost_prolong_data opd = { 0 };
1990 end = start + oa->o_blocks;
1992 opd.opd_mode = LCK_PW;
1993 opd.opd_exp = req->rq_export;
1994 opd.opd_policy.l_extent.start = start & CFS_PAGE_MASK;
1995 if (oa->o_blocks == OBD_OBJECT_EOF || end < start)
1996 opd.opd_policy.l_extent.end = OBD_OBJECT_EOF;
1998 opd.opd_policy.l_extent.end = end | ~CFS_PAGE_MASK;
2000 /* prolong locks for the current service time of the corresponding
2001 * portal (= OST_IO_PORTAL) */
2002 opd.opd_timeout = AT_OFF ? obd_timeout / 2:
2003 max(at_est2timeout(at_get(&req->rq_rqbd->
2004 rqbd_service->srv_at_estimate)), ldlm_timeout);
2006 CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
2007 res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
2008 opd.opd_policy.l_extent.end);
2011 ldlm_resource_iterate(req->rq_export->exp_obd->obd_namespace, &res_id,
2012 ost_prolong_locks_iter, &opd);
2013 RETURN(opd.opd_lock_match);
2017 * Like ost_rw_hpreq_lock_match(), but for OST_PUNCH RPCs.
2019 static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req,
2020 struct ldlm_lock *lock)
2022 struct ost_body *body;
2026 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2028 RETURN(0); /* can't return -EFAULT here */
2030 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
2034 if (body->oa.o_valid & OBD_MD_FLHANDLE &&
2035 body->oa.o_handle.cookie == lock->l_handle.h_cookie)
2041 * Like ost_rw_hpreq_check(), but for OST_PUNCH RPCs.
2043 static int ost_punch_hpreq_check(struct ptlrpc_request *req)
2045 struct ost_body *body;
2048 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
2052 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
2056 LASSERT(!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
2057 !(body->oa.o_flags & OBD_FL_SRVLOCK));
2059 RETURN(ost_punch_prolong_locks(req, &body->oa));
2062 struct ptlrpc_hpreq_ops ost_hpreq_rw = {
2063 .hpreq_lock_match = ost_rw_hpreq_lock_match,
2064 .hpreq_check = ost_rw_hpreq_check,
2067 struct ptlrpc_hpreq_ops ost_hpreq_punch = {
2068 .hpreq_lock_match = ost_punch_hpreq_lock_match,
2069 .hpreq_check = ost_punch_hpreq_check,
2072 /** Assign high priority operations to the request if needed. */
2073 static int ost_hpreq_handler(struct ptlrpc_request *req)
2076 if (req->rq_export) {
2077 int opc = lustre_msg_get_opc(req->rq_reqmsg);
2078 struct ost_body *body;
2080 if (opc == OST_READ || opc == OST_WRITE) {
2081 struct niobuf_remote *nb;
2082 struct obd_ioobj *ioo;
2083 int objcount, niocount;
2086 /* RPCs on the H-P queue can be inspected before
2087 * ost_handler() initializes their pills, so we
2088 * initialize that here. Capsule initialization is
2089 * idempotent, as is setting the pill's format (provided
2090 * it doesn't change).
2092 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2093 if (opc == OST_READ)
2094 req_capsule_set(&req->rq_pill,
2097 req_capsule_set(&req->rq_pill,
2098 &RQF_OST_BRW_WRITE);
2100 body = req_capsule_client_get(&req->rq_pill,
2103 CERROR("Missing/short ost_body\n");
2107 objcount = req_capsule_get_size(&req->rq_pill,
2111 if (objcount == 0) {
2112 CERROR("Missing/short ioobj\n");
2116 CERROR("too many ioobjs (%d)\n", objcount);
2120 ioo = req_capsule_client_get(&req->rq_pill,
2123 CERROR("Missing/short ioobj\n");
2127 for (niocount = i = 0; i < objcount; i++) {
2128 if (ioo[i].ioo_bufcnt == 0) {
2129 CERROR("ioo[%d] has zero bufcnt\n", i);
2132 niocount += ioo[i].ioo_bufcnt;
2134 if (niocount > PTLRPC_MAX_BRW_PAGES) {
2135 DEBUG_REQ(D_RPCTRACE, req,
2136 "bulk has too many pages (%d)",
2141 nb = req_capsule_client_get(&req->rq_pill,
2142 &RMF_NIOBUF_REMOTE);
2144 CERROR("Missing/short niobuf\n");
2148 if (niocount == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
2149 req->rq_ops = &ost_hpreq_rw;
2150 } else if (opc == OST_PUNCH) {
2151 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2152 req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
2154 body = req_capsule_client_get(&req->rq_pill,
2157 CERROR("Missing/short ost_body\n");
2161 if (!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
2162 !(body->oa.o_flags & OBD_FL_SRVLOCK))
2163 req->rq_ops = &ost_hpreq_punch;
2169 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2170 int ost_handle(struct ptlrpc_request *req)
2172 struct obd_trans_info trans_info = { 0, };
2173 struct obd_trans_info *oti = &trans_info;
2174 int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
2175 struct obd_device *obd = NULL;
2178 LASSERT(current->journal_info == NULL);
2180 /* primordial rpcs don't affect server recovery */
2181 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2183 case SEC_CTX_INIT_CONT:
2188 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2190 if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) {
2191 if (!class_connected_export(req->rq_export)) {
2192 CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
2193 lustre_msg_get_opc(req->rq_reqmsg),
2194 libcfs_id2str(req->rq_peer));
2195 req->rq_status = -ENOTCONN;
2196 GOTO(out, rc = -ENOTCONN);
2199 obd = req->rq_export->exp_obd;
2201 /* Check for aborted recovery. */
2202 if (obd->obd_recovering) {
2203 rc = ost_filter_recovery_request(req, obd,
2205 if (rc || !should_process)
2207 else if (should_process < 0) {
2208 req->rq_status = should_process;
2209 rc = ptlrpc_error(req);
2217 rc = ost_msg_check_version(req->rq_reqmsg);
2221 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2223 CDEBUG(D_INODE, "connect\n");
2224 req_capsule_set(&req->rq_pill, &RQF_OST_CONNECT);
2225 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET))
2227 rc = target_handle_connect(req);
2228 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2))
2231 rc = ost_init_sec_level(req);
2233 rc = ost_connect_check_sptlrpc(req);
2237 case OST_DISCONNECT:
2238 CDEBUG(D_INODE, "disconnect\n");
2239 req_capsule_set(&req->rq_pill, &RQF_OST_DISCONNECT);
2240 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DISCONNECT_NET))
2242 rc = target_handle_disconnect(req);
2245 CDEBUG(D_INODE, "create\n");
2246 req_capsule_set(&req->rq_pill, &RQF_OST_CREATE);
2247 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET))
2249 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2250 GOTO(out, rc = -EROFS);
2251 rc = ost_create(req->rq_export, req, oti);
2254 CDEBUG(D_INODE, "destroy\n");
2255 req_capsule_set(&req->rq_pill, &RQF_OST_DESTROY);
2256 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET))
2258 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2259 GOTO(out, rc = -EROFS);
2260 rc = ost_destroy(req->rq_export, req, oti);
2263 CDEBUG(D_INODE, "getattr\n");
2264 req_capsule_set(&req->rq_pill, &RQF_OST_GETATTR);
2265 if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET))
2267 rc = ost_getattr(req->rq_export, req);
2270 CDEBUG(D_INODE, "setattr\n");
2271 req_capsule_set(&req->rq_pill, &RQF_OST_SETATTR);
2272 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET))
2274 rc = ost_setattr(req->rq_export, req, oti);
2277 req_capsule_set(&req->rq_pill, &RQF_OST_BRW_WRITE);
2278 CDEBUG(D_INODE, "write\n");
2279 /* req->rq_request_portal would be nice, if it was set */
2280 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
2281 CERROR("%s: deny write request from %s to portal %u\n",
2282 req->rq_export->exp_obd->obd_name,
2283 obd_export_nid2str(req->rq_export),
2284 req->rq_rqbd->rqbd_service->srv_req_portal);
2285 GOTO(out, rc = -EPROTO);
2287 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
2289 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2290 GOTO(out, rc = -ENOSPC);
2291 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2292 GOTO(out, rc = -EROFS);
2293 rc = ost_brw_write(req, oti);
2294 LASSERT(current->journal_info == NULL);
2295 /* ost_brw_write sends its own replies */
2298 req_capsule_set(&req->rq_pill, &RQF_OST_BRW_READ);
2299 CDEBUG(D_INODE, "read\n");
2300 /* req->rq_request_portal would be nice, if it was set */
2301 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
2302 CERROR("%s: deny read request from %s to portal %u\n",
2303 req->rq_export->exp_obd->obd_name,
2304 obd_export_nid2str(req->rq_export),
2305 req->rq_rqbd->rqbd_service->srv_req_portal);
2306 GOTO(out, rc = -EPROTO);
2308 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
2310 rc = ost_brw_read(req, oti);
2311 LASSERT(current->journal_info == NULL);
2312 /* ost_brw_read sends its own replies */
2315 CDEBUG(D_INODE, "punch\n");
2316 req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
2317 if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET))
2319 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2320 GOTO(out, rc = -EROFS);
2321 rc = ost_punch(req->rq_export, req, oti);
2324 CDEBUG(D_INODE, "statfs\n");
2325 req_capsule_set(&req->rq_pill, &RQF_OST_STATFS);
2326 if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET))
2328 rc = ost_statfs(req);
2331 CDEBUG(D_INODE, "sync\n");
2332 req_capsule_set(&req->rq_pill, &RQF_OST_SYNC);
2333 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET))
2335 rc = ost_sync(req->rq_export, req);
2338 DEBUG_REQ(D_INODE, req, "set_info");
2339 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2340 rc = ost_set_info(req->rq_export, req);
2343 DEBUG_REQ(D_INODE, req, "get_info");
2344 req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC);
2345 rc = ost_get_info(req->rq_export, req);
2347 #ifdef HAVE_QUOTA_SUPPORT
2348 case OST_QUOTACHECK:
2349 CDEBUG(D_INODE, "quotacheck\n");
2350 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK);
2351 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACHECK_NET))
2353 rc = ost_handle_quotacheck(req);
2356 CDEBUG(D_INODE, "quotactl\n");
2357 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACTL);
2358 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACTL_NET))
2360 rc = ost_handle_quotactl(req);
2362 case OST_QUOTA_ADJUST_QUNIT:
2363 CDEBUG(D_INODE, "quota_adjust_qunit\n");
2364 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTA_ADJUST_QUNIT);
2365 rc = ost_handle_quota_adjust_qunit(req);
2369 DEBUG_REQ(D_INODE, req, "ping");
2370 req_capsule_set(&req->rq_pill, &RQF_OBD_PING);
2371 rc = target_handle_ping(req);
2373 /* FIXME - just reply status */
2374 case LLOG_ORIGIN_CONNECT:
2375 DEBUG_REQ(D_INODE, req, "log connect");
2376 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_CONNECT);
2377 rc = ost_llog_handle_connect(req->rq_export, req);
2378 req->rq_status = rc;
2379 rc = req_capsule_server_pack(&req->rq_pill);
2382 RETURN(ptlrpc_reply(req));
2383 case OBD_LOG_CANCEL:
2384 CDEBUG(D_INODE, "log cancel\n");
2385 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2386 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2388 rc = llog_origin_handle_cancel(req);
2389 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2391 req->rq_status = rc;
2392 rc = req_capsule_server_pack(&req->rq_pill);
2395 RETURN(ptlrpc_reply(req));
2397 CDEBUG(D_INODE, "enqueue\n");
2398 req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE);
2399 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE))
2401 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2403 ldlm_server_glimpse_ast);
2404 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
2407 CDEBUG(D_INODE, "convert\n");
2408 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2409 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT))
2411 rc = ldlm_handle_convert(req);
2414 CDEBUG(D_INODE, "cancel\n");
2415 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2416 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2418 rc = ldlm_handle_cancel(req);
2420 case LDLM_BL_CALLBACK:
2421 case LDLM_CP_CALLBACK:
2422 CDEBUG(D_INODE, "callback\n");
2423 CERROR("callbacks should not happen on OST\n");
2426 CERROR("Unexpected opcode %d\n",
2427 lustre_msg_get_opc(req->rq_reqmsg));
2428 req->rq_status = -ENOTSUPP;
2429 rc = ptlrpc_error(req);
2433 LASSERT(current->journal_info == NULL);
2436 /* If we're DISCONNECTing, the export_data is already freed */
2437 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT)
2438 target_committed_to_req(req);
2442 oti_to_request(oti, req);
2444 target_send_reply(req, rc, fail);
2447 EXPORT_SYMBOL(ost_handle);
2449 * free per-thread pool created by ost_thread_init().
2451 static void ost_thread_done(struct ptlrpc_thread *thread)
2453 struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local
2458 LASSERT(thread != NULL);
2461 * be prepared to handle partially-initialized pools (because this is
2462 * called from ost_thread_init() for cleanup.
2464 tls = thread->t_data;
2467 thread->t_data = NULL;
2473 * initialize per-thread page pool (bug 5137).
2475 static int ost_thread_init(struct ptlrpc_thread *thread)
2477 struct ost_thread_local_cache *tls;
2481 LASSERT(thread != NULL);
2482 LASSERT(thread->t_data == NULL);
2483 LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
2488 thread->t_data = tls;
2492 #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
2494 /* Sigh - really, this is an OSS, the _server_, not the _target_ */
2495 static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2497 struct ost_obd *ost = &obd->u.ost;
2498 struct lprocfs_static_vars lvars;
2499 int oss_min_threads;
2500 int oss_max_threads;
2501 int oss_min_create_threads;
2502 int oss_max_create_threads;
2506 rc = cfs_cleanup_group_info();
2510 lprocfs_ost_init_vars(&lvars);
2511 lprocfs_obd_setup(obd, lvars.obd_vars);
2513 cfs_sema_init(&ost->ost_health_sem, 1);
2515 if (oss_num_threads) {
2516 /* If oss_num_threads is set, it is the min and the max. */
2517 if (oss_num_threads > OSS_THREADS_MAX)
2518 oss_num_threads = OSS_THREADS_MAX;
2519 if (oss_num_threads < OSS_THREADS_MIN)
2520 oss_num_threads = OSS_THREADS_MIN;
2521 oss_max_threads = oss_min_threads = oss_num_threads;
2523 /* Base min threads on memory and cpus */
2525 cfs_num_possible_cpus() * CFS_NUM_CACHEPAGES >>
2526 (27 - CFS_PAGE_SHIFT);
2527 if (oss_min_threads < OSS_THREADS_MIN)
2528 oss_min_threads = OSS_THREADS_MIN;
2529 /* Insure a 4x range for dynamic threads */
2530 if (oss_min_threads > OSS_THREADS_MAX / 4)
2531 oss_min_threads = OSS_THREADS_MAX / 4;
2532 oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4 + 1);
2536 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2537 OST_MAXREPSIZE, OST_REQUEST_PORTAL,
2538 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2539 ost_handle, LUSTRE_OSS_NAME,
2540 obd->obd_proc_entry, target_print_req,
2541 oss_min_threads, oss_max_threads,
2542 "ll_ost", LCT_DT_THREAD, NULL);
2543 if (ost->ost_service == NULL) {
2544 CERROR("failed to start service\n");
2545 GOTO(out_lprocfs, rc = -ENOMEM);
2548 rc = ptlrpc_start_threads(ost->ost_service);
2550 GOTO(out_service, rc = -EINVAL);
2552 if (oss_num_create_threads) {
2553 if (oss_num_create_threads > OSS_MAX_CREATE_THREADS)
2554 oss_num_create_threads = OSS_MAX_CREATE_THREADS;
2555 if (oss_num_create_threads < OSS_MIN_CREATE_THREADS)
2556 oss_num_create_threads = OSS_MIN_CREATE_THREADS;
2557 oss_min_create_threads = oss_max_create_threads =
2558 oss_num_create_threads;
2560 oss_min_create_threads = OSS_MIN_CREATE_THREADS;
2561 oss_max_create_threads = OSS_MAX_CREATE_THREADS;
2564 ost->ost_create_service =
2565 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2566 OST_MAXREPSIZE, OST_CREATE_PORTAL,
2567 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2568 ost_handle, "ost_create",
2569 obd->obd_proc_entry, target_print_req,
2570 oss_min_create_threads, oss_max_create_threads,
2571 "ll_ost_creat", LCT_DT_THREAD, NULL);
2572 if (ost->ost_create_service == NULL) {
2573 CERROR("failed to start OST create service\n");
2574 GOTO(out_service, rc = -ENOMEM);
2577 rc = ptlrpc_start_threads(ost->ost_create_service);
2579 GOTO(out_create, rc = -EINVAL);
2581 ost->ost_io_service =
2582 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2583 OST_MAXREPSIZE, OST_IO_PORTAL,
2584 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2585 ost_handle, "ost_io",
2586 obd->obd_proc_entry, target_print_req,
2587 oss_min_threads, oss_max_threads,
2588 "ll_ost_io", LCT_DT_THREAD, ost_hpreq_handler);
2589 if (ost->ost_io_service == NULL) {
2590 CERROR("failed to start OST I/O service\n");
2591 GOTO(out_create, rc = -ENOMEM);
2594 ost->ost_io_service->srv_init = ost_thread_init;
2595 ost->ost_io_service->srv_done = ost_thread_done;
2596 ost->ost_io_service->srv_cpu_affinity = 1;
2597 rc = ptlrpc_start_threads(ost->ost_io_service);
2599 GOTO(out_io, rc = -EINVAL);
2601 ping_evictor_start();
2606 ptlrpc_unregister_service(ost->ost_io_service);
2607 ost->ost_io_service = NULL;
2609 ptlrpc_unregister_service(ost->ost_create_service);
2610 ost->ost_create_service = NULL;
2612 ptlrpc_unregister_service(ost->ost_service);
2613 ost->ost_service = NULL;
2615 lprocfs_obd_cleanup(obd);
2619 static int ost_cleanup(struct obd_device *obd)
2621 struct ost_obd *ost = &obd->u.ost;
2625 ping_evictor_stop();
2627 /* there is no recovery for OST OBD, all recovery is controlled by
2629 LASSERT(obd->obd_recovering == 0);
2630 cfs_down(&ost->ost_health_sem);
2631 ptlrpc_unregister_service(ost->ost_service);
2632 ptlrpc_unregister_service(ost->ost_create_service);
2633 ptlrpc_unregister_service(ost->ost_io_service);
2634 ost->ost_service = NULL;
2635 ost->ost_create_service = NULL;
2636 cfs_up(&ost->ost_health_sem);
2638 lprocfs_obd_cleanup(obd);
2643 static int ost_health_check(struct obd_device *obd)
2645 struct ost_obd *ost = &obd->u.ost;
2648 cfs_down(&ost->ost_health_sem);
2649 rc |= ptlrpc_service_health_check(ost->ost_service);
2650 rc |= ptlrpc_service_health_check(ost->ost_create_service);
2651 rc |= ptlrpc_service_health_check(ost->ost_io_service);
2652 cfs_up(&ost->ost_health_sem);
2655 * health_check to return 0 on healthy
2656 * and 1 on unhealthy.
2664 struct ost_thread_local_cache *ost_tls(struct ptlrpc_request *r)
2666 return (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
2669 /* use obd ops to offer management infrastructure */
2670 static struct obd_ops ost_obd_ops = {
2671 .o_owner = THIS_MODULE,
2672 .o_setup = ost_setup,
2673 .o_cleanup = ost_cleanup,
2674 .o_health_check = ost_health_check,
2678 static int __init ost_init(void)
2680 struct lprocfs_static_vars lvars;
2684 lprocfs_ost_init_vars(&lvars);
2685 rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
2686 LUSTRE_OSS_NAME, NULL);
2688 if (ost_num_threads != 0 && oss_num_threads == 0) {
2689 LCONSOLE_INFO("ost_num_threads module parameter is deprecated, "
2690 "use oss_num_threads instead or unset both for "
2691 "dynamic thread startup\n");
2692 oss_num_threads = ost_num_threads;
2698 static void /*__exit*/ ost_exit(void)
2700 class_unregister_type(LUSTRE_OSS_NAME);
2703 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2704 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
2705 MODULE_LICENSE("GPL");
2707 module_init(ost_init);
2708 module_exit(ost_exit);