4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Whamcloud, Inc.
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>
42 #define DEBUG_SUBSYSTEM S_OST
44 #include <linux/module.h>
45 #include <obd_cksum.h>
47 #include <lustre_net.h>
48 #include <lustre_dlm.h>
49 #include <lustre_export.h>
50 #include <lustre_debug.h>
51 #include <linux/init.h>
52 #include <lprocfs_status.h>
53 #include <libcfs/list.h>
54 #include <lustre_quota.h>
55 #include "ost_internal.h"
57 static int oss_num_threads;
58 CFS_MODULE_PARM(oss_num_threads, "i", int, 0444,
59 "number of OSS service threads to start");
61 static int ost_num_threads;
62 CFS_MODULE_PARM(ost_num_threads, "i", int, 0444,
63 "number of OST service threads to start (deprecated)");
65 static int oss_num_create_threads;
66 CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444,
67 "number of OSS create threads to start");
70 * Do not return server-side uid/gid to remote client
72 static void ost_drop_id(struct obd_export *exp, struct obdo *oa)
74 if (exp_connect_rmtclient(exp)) {
77 oa->o_valid &= ~(OBD_MD_FLUID | OBD_MD_FLGID);
82 * Validate oa from client.
83 * If the request comes from 2.0 clients, currently only RSVD seq and IDIF
85 * a. for single MDS seq = FID_SEQ_OST_MDT0,
86 * b. for CMD, seq = FID_SEQ_OST_MDT0, FID_SEQ_OST_MDT1 - FID_SEQ_OST_MAX
88 static int ost_validate_obdo(struct obd_export *exp, struct obdo *oa,
89 struct obd_ioobj *ioobj)
91 if (oa != NULL && !(oa->o_valid & OBD_MD_FLGROUP)) {
92 oa->o_seq = FID_SEQ_OST_MDT0;
94 ioobj->ioo_seq = FID_SEQ_OST_MDT0;
95 /* remove fid_seq_is_rsvd() after FID-on-OST allows SEQ > 9 */
96 } else if (oa == NULL || !(fid_seq_is_rsvd(oa->o_seq) ||
97 fid_seq_is_mdt0(oa->o_seq))) {
98 CERROR("%s: client %s sent invalid object "POSTID"\n",
99 exp->exp_obd->obd_name, obd_export_nid2str(exp),
100 oa ? oa->o_id : -1, oa ? oa->o_seq : -1);
103 obdo_from_ostid(oa, &oa->o_oi);
105 ioobj_from_obdo(ioobj, oa);
109 void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
111 struct oti_req_ack_lock *ack_lock;
117 if (req->rq_repmsg) {
118 __u64 versions[PTLRPC_NUM_VERSIONS] = { 0 };
119 lustre_msg_set_transno(req->rq_repmsg, oti->oti_transno);
120 versions[0] = oti->oti_pre_version;
121 lustre_msg_set_versions(req->rq_repmsg, versions);
123 req->rq_transno = oti->oti_transno;
125 /* XXX 4 == entries in oti_ack_locks??? */
126 for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) {
129 /* XXX not even calling target_send_reply in some cases... */
130 ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode, 0);
134 static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req,
135 struct obd_trans_info *oti)
137 struct ost_body *body, *repbody;
138 struct lustre_capa *capa = NULL;
142 /* Get the request body */
143 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
147 if (body->oa.o_id == 0)
150 rc = ost_validate_obdo(exp, &body->oa, NULL);
154 /* If there's a DLM request, cancel the locks mentioned in it*/
155 if (req_capsule_field_present(&req->rq_pill, &RMF_DLM_REQ, RCL_CLIENT)) {
156 struct ldlm_request *dlm;
158 dlm = req_capsule_client_get(&req->rq_pill, &RMF_DLM_REQ);
161 ldlm_request_cancel(req, dlm, 0);
164 /* If there's a capability, get it */
165 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
166 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
168 CERROR("Missing capability for OST DESTROY");
173 /* Prepare the reply */
174 rc = req_capsule_server_pack(&req->rq_pill);
178 /* Get the log cancellation cookie */
179 if (body->oa.o_valid & OBD_MD_FLCOOKIE)
180 oti->oti_logcookies = &body->oa.o_lcookie;
182 /* Finish the reply */
183 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
184 memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
186 /* Do the destroy and set the reply status accordingly */
187 req->rq_status = obd_destroy(req->rq_svc_thread->t_env, exp,
188 &repbody->oa, NULL, oti, NULL, capa);
193 * Helper function for getting server side [start, start+count] DLM lock
194 * if asked by client.
196 static int ost_lock_get(struct obd_export *exp, struct obdo *oa,
197 __u64 start, __u64 count, struct lustre_handle *lh,
200 struct ldlm_res_id res_id;
201 ldlm_policy_data_t policy;
202 __u64 end = start + count;
206 LASSERT(!lustre_handle_is_used(lh));
207 /* o_id and o_gr are used for localizing resource, if client miss to set
208 * them, do not trigger ASSERTION. */
209 if (unlikely((oa->o_valid & (OBD_MD_FLID | OBD_MD_FLGROUP)) !=
210 (OBD_MD_FLID | OBD_MD_FLGROUP)))
213 if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
214 !(oa->o_flags & OBD_FL_SRVLOCK))
217 osc_build_res_name(oa->o_id, oa->o_seq, &res_id);
218 CDEBUG(D_INODE, "OST-side extent lock.\n");
220 policy.l_extent.start = start & CFS_PAGE_MASK;
222 /* If ->o_blocks is EOF it means "lock till the end of the
223 * file". Otherwise, it's size of a hole being punched (in bytes) */
224 if (count == OBD_OBJECT_EOF || end < start)
225 policy.l_extent.end = OBD_OBJECT_EOF;
227 policy.l_extent.end = end | ~CFS_PAGE_MASK;
229 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
230 LDLM_EXTENT, &policy, mode, &flags,
231 ldlm_blocking_ast, ldlm_completion_ast,
232 ldlm_glimpse_ast, NULL, 0, NULL, lh));
235 /* Helper function: release lock, if any. */
236 static void ost_lock_put(struct obd_export *exp,
237 struct lustre_handle *lh, int mode)
240 if (lustre_handle_is_used(lh))
241 ldlm_lock_decref(lh, mode);
245 static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
247 struct ost_body *body, *repbody;
248 struct obd_info *oinfo;
249 struct lustre_handle lh = { 0 };
250 struct lustre_capa *capa = NULL;
254 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
258 rc = ost_validate_obdo(exp, &body->oa, NULL);
262 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
263 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
265 CERROR("Missing capability for OST GETATTR");
270 rc = req_capsule_server_pack(&req->rq_pill);
274 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
275 repbody->oa = body->oa;
277 rc = ost_lock_get(exp, &repbody->oa, 0, OBD_OBJECT_EOF, &lh, LCK_PR, 0);
281 OBD_ALLOC_PTR(oinfo);
283 GOTO(unlock, rc = -ENOMEM);
284 oinfo->oi_oa = &repbody->oa;
285 oinfo->oi_capa = capa;
287 req->rq_status = obd_getattr(req->rq_svc_thread->t_env, exp, oinfo);
291 ost_drop_id(exp, &repbody->oa);
294 ost_lock_put(exp, &lh, LCK_PR);
298 static int ost_statfs(struct ptlrpc_request *req)
300 struct obd_statfs *osfs;
304 rc = req_capsule_server_pack(&req->rq_pill);
308 osfs = req_capsule_server_get(&req->rq_pill, &RMF_OBD_STATFS);
310 req->rq_status = obd_statfs(req->rq_svc_thread->t_env, req->rq_export,
312 cfs_time_shift_64(-OBD_STATFS_CACHE_SECONDS),
314 if (req->rq_status != 0)
315 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
320 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
321 struct obd_trans_info *oti)
323 struct ost_body *body, *repbody;
327 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
331 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
335 rc = req_capsule_server_pack(&req->rq_pill);
339 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
340 repbody->oa = body->oa;
341 oti->oti_logcookies = &body->oa.o_lcookie;
343 req->rq_status = obd_create(req->rq_svc_thread->t_env, exp,
344 &repbody->oa, NULL, oti);
345 //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
349 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
350 struct obd_trans_info *oti)
352 struct ost_body *body, *repbody;
354 struct lustre_handle lh = {0,};
357 /* check that we do support OBD_CONNECT_TRUNCLOCK. */
358 CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
360 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
364 rc = ost_validate_obdo(exp, &body->oa, NULL);
368 if ((body->oa.o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
369 (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
372 rc = req_capsule_server_pack(&req->rq_pill);
376 /* standard truncate optimization: if file body is completely
377 * destroyed, don't send data back to the server. */
378 if (body->oa.o_size == 0)
379 flags |= LDLM_AST_DISCARD_DATA;
381 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
382 repbody->oa = body->oa;
384 rc = ost_lock_get(exp, &repbody->oa, repbody->oa.o_size,
385 repbody->oa.o_blocks, &lh, LCK_PW, flags);
387 struct obd_info *oinfo;
388 struct lustre_capa *capa = NULL;
390 if (repbody->oa.o_valid & OBD_MD_FLFLAGS &&
391 repbody->oa.o_flags == OBD_FL_SRVLOCK)
393 * If OBD_FL_SRVLOCK is the only bit set in
394 * ->o_flags, clear OBD_MD_FLFLAGS to avoid falling
395 * through filter_setattr() to filter_iocontrol().
397 repbody->oa.o_valid &= ~OBD_MD_FLFLAGS;
399 if (repbody->oa.o_valid & OBD_MD_FLOSSCAPA) {
400 capa = req_capsule_client_get(&req->rq_pill,
403 CERROR("Missing capability for OST PUNCH");
404 GOTO(unlock, rc = -EFAULT);
408 OBD_ALLOC_PTR(oinfo);
410 GOTO(unlock, rc = -ENOMEM);
411 oinfo->oi_oa = &repbody->oa;
412 oinfo->oi_policy.l_extent.start = oinfo->oi_oa->o_size;
413 oinfo->oi_policy.l_extent.end = oinfo->oi_oa->o_blocks;
414 oinfo->oi_capa = capa;
415 oinfo->oi_flags = OBD_FL_PUNCH;
417 req->rq_status = obd_punch(req->rq_svc_thread->t_env, exp,
421 ost_lock_put(exp, &lh, LCK_PW);
424 ost_drop_id(exp, &repbody->oa);
428 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req,
429 struct obd_trans_info *oti)
431 struct ost_body *body, *repbody;
432 struct obd_info *oinfo;
433 struct lustre_capa *capa = NULL;
437 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
441 rc = ost_validate_obdo(exp, &body->oa, NULL);
445 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
446 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
448 CERROR("Missing capability for OST SYNC");
453 rc = req_capsule_server_pack(&req->rq_pill);
457 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
458 repbody->oa = body->oa;
460 OBD_ALLOC_PTR(oinfo);
464 oinfo->oi_oa = &repbody->oa;
465 oinfo->oi_capa = capa;
466 oinfo->oi_jobid = oti->oti_jobid;
467 req->rq_status = obd_sync(req->rq_svc_thread->t_env, exp, oinfo,
468 repbody->oa.o_size, repbody->oa.o_blocks,
472 ost_drop_id(exp, &repbody->oa);
476 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
477 struct obd_trans_info *oti)
479 struct ost_body *body, *repbody;
480 struct obd_info *oinfo;
481 struct lustre_capa *capa = NULL;
485 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
489 rc = ost_validate_obdo(req->rq_export, &body->oa, NULL);
493 rc = req_capsule_server_pack(&req->rq_pill);
497 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
498 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
500 CERROR("Missing capability for OST SETATTR");
505 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
506 repbody->oa = body->oa;
508 OBD_ALLOC_PTR(oinfo);
511 oinfo->oi_oa = &repbody->oa;
512 oinfo->oi_capa = capa;
514 req->rq_status = obd_setattr(req->rq_svc_thread->t_env, exp, oinfo,
519 ost_drop_id(exp, &repbody->oa);
523 static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc,
524 cksum_type_t cksum_type)
526 struct cfs_crypto_hash_desc *hdesc;
527 unsigned int bufsize;
529 unsigned char cfs_alg = cksum_obd2cfs(cksum_type);
532 hdesc = cfs_crypto_hash_init(cfs_alg, NULL, 0);
534 CERROR("Unable to initialize checksum hash %s\n",
535 cfs_crypto_hash_name(cfs_alg));
536 return PTR_ERR(hdesc);
538 CDEBUG(D_INFO, "Checksum for algo %s\n", cfs_crypto_hash_name(cfs_alg));
539 for (i = 0; i < desc->bd_iov_count; i++) {
541 /* corrupt the data before we compute the checksum, to
542 * simulate a client->OST data error */
543 if (i == 0 && opc == OST_WRITE &&
544 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE)) {
545 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
546 int len = desc->bd_iov[i].kiov_len;
547 char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
548 memcpy(ptr, "bad3", min(4, len));
549 kunmap(desc->bd_iov[i].kiov_page);
551 cfs_crypto_hash_update_page(hdesc, desc->bd_iov[i].kiov_page,
552 desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK,
553 desc->bd_iov[i].kiov_len);
555 /* corrupt the data after we compute the checksum, to
556 * simulate an OST->client data error */
557 if (i == 0 && opc == OST_READ &&
558 OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND)) {
559 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
560 int len = desc->bd_iov[i].kiov_len;
561 char *ptr = kmap(desc->bd_iov[i].kiov_page) + off;
562 memcpy(ptr, "bad4", min(4, len));
563 kunmap(desc->bd_iov[i].kiov_page);
564 /* nobody should use corrupted page again */
565 ClearPageUptodate(desc->bd_iov[i].kiov_page);
570 err = cfs_crypto_hash_final(hdesc, (unsigned char *)&cksum, &bufsize);
572 cfs_crypto_hash_final(hdesc, NULL, NULL);
577 static int ost_brw_lock_get(int mode, struct obd_export *exp,
578 struct obd_ioobj *obj, struct niobuf_remote *nb,
579 struct lustre_handle *lh)
582 int nrbufs = obj->ioo_bufcnt;
583 struct ldlm_res_id res_id;
584 ldlm_policy_data_t policy;
588 osc_build_res_name(obj->ioo_id, obj->ioo_seq, &res_id);
589 LASSERT(mode == LCK_PR || mode == LCK_PW);
590 LASSERT(!lustre_handle_is_used(lh));
592 if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
595 for (i = 1; i < nrbufs; i ++)
596 if ((nb[0].flags & OBD_BRW_SRVLOCK) !=
597 (nb[i].flags & OBD_BRW_SRVLOCK))
600 policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
601 policy.l_extent.end = (nb[nrbufs - 1].offset +
602 nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
604 RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
605 LDLM_EXTENT, &policy, mode, &flags,
606 ldlm_blocking_ast, ldlm_completion_ast,
607 ldlm_glimpse_ast, NULL, 0, NULL, lh));
610 static void ost_brw_lock_put(int mode,
611 struct obd_ioobj *obj, struct niobuf_remote *niob,
612 struct lustre_handle *lh)
615 LASSERT(mode == LCK_PR || mode == LCK_PW);
616 LASSERT((obj->ioo_bufcnt > 0 && (niob[0].flags & OBD_BRW_SRVLOCK)) ==
617 lustre_handle_is_used(lh));
618 if (lustre_handle_is_used(lh))
619 ldlm_lock_decref(lh, mode);
623 /* Allocate thread local buffers if needed */
624 static struct ost_thread_local_cache *ost_tls_get(struct ptlrpc_request *r)
626 struct ost_thread_local_cache *tls =
627 (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
629 /* In normal mode of operation an I/O request is serviced only
630 * by ll_ost_io threads each of them has own tls buffers allocated by
632 * During recovery, an I/O request may be queued until any of the ost
633 * service threads process it. Not necessary it should be one of
634 * ll_ost_io threads. In that case we dynamically allocating tls
635 * buffers for the request service time. */
636 if (unlikely(tls == NULL)) {
637 LASSERT(r->rq_export->exp_in_recovery);
641 r->rq_svc_thread->t_data = tls;
647 /* Free thread local buffers if they were allocated only for servicing
648 * this one request */
649 static void ost_tls_put(struct ptlrpc_request *r)
651 struct ost_thread_local_cache *tls =
652 (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
654 if (unlikely(tls->temporary)) {
656 r->rq_svc_thread->t_data = NULL;
660 static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti)
662 struct ptlrpc_bulk_desc *desc = NULL;
663 struct obd_export *exp = req->rq_export;
664 struct niobuf_remote *remote_nb;
665 struct niobuf_local *local_nb;
666 struct obd_ioobj *ioo;
667 struct ost_body *body, *repbody;
668 struct lustre_capa *capa = NULL;
669 struct l_wait_info lwi;
670 struct lustre_handle lockh = { 0 };
671 int niocount, npages, nob = 0, rc, i;
673 struct ost_thread_local_cache *tls;
676 req->rq_bulk_read = 1;
678 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
679 GOTO(out, rc = -EIO);
681 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
683 /* Check if there is eviction in progress, and if so, wait for it to
685 if (unlikely(cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
686 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
687 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
688 !cfs_atomic_read(&exp->exp_obd->obd_evict_inprogress),
692 GOTO(out, rc = -ENOTCONN);
694 /* ost_body, ioobj & noibuf_remote are verified and swabbed in
695 * ost_rw_hpreq_check(). */
696 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
698 GOTO(out, rc = -EFAULT);
701 * A req_capsule_X_get_array(pill, field, ptr_to_element_count) function
702 * would be useful here and wherever we get &RMF_OBD_IOOBJ and
703 * &RMF_NIOBUF_REMOTE.
705 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
707 GOTO(out, rc = -EFAULT);
709 rc = ost_validate_obdo(exp, &body->oa, ioo);
713 niocount = ioo->ioo_bufcnt;
714 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
715 if (remote_nb == NULL)
716 GOTO(out, rc = -EFAULT);
718 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
719 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
721 CERROR("Missing capability for OST BRW READ");
722 GOTO(out, rc = -EFAULT);
726 rc = req_capsule_server_pack(&req->rq_pill);
730 tls = ost_tls_get(req);
732 GOTO(out_bulk, rc = -ENOMEM);
733 local_nb = tls->local;
735 rc = ost_brw_lock_get(LCK_PR, exp, ioo, remote_nb, &lockh);
740 * If getting the lock took more time than
741 * client was willing to wait, drop it. b=11330
743 if (cfs_time_current_sec() > req->rq_deadline ||
744 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
746 CERROR("Dropping timed-out read from %s because locking"
747 "object "LPX64" took %ld seconds (limit was %ld).\n",
748 libcfs_id2str(req->rq_peer), ioo->ioo_id,
749 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
750 req->rq_deadline - req->rq_arrival_time.tv_sec);
751 GOTO(out_lock, rc = -ETIMEDOUT);
754 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
755 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
757 npages = OST_THREAD_POOL_SIZE;
758 rc = obd_preprw(req->rq_svc_thread->t_env, OBD_BRW_READ, exp,
759 &repbody->oa, 1, ioo, remote_nb, &npages, local_nb,
764 desc = ptlrpc_prep_bulk_exp(req, npages,
765 BULK_PUT_SOURCE, OST_BULK_PORTAL);
767 GOTO(out_commitrw, rc = -ENOMEM);
770 for (i = 0; i < npages; i++) {
771 int page_rc = local_nb[i].rc;
773 if (page_rc < 0) { /* error */
779 if (page_rc != 0) { /* some data! */
780 LASSERT (local_nb[i].page != NULL);
781 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
782 local_nb[i].offset & ~CFS_PAGE_MASK,
786 if (page_rc != local_nb[i].len) { /* short read */
787 /* All subsequent pages should be 0 */
789 LASSERT(local_nb[i].rc == 0);
794 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
795 cksum_type_t cksum_type =
796 cksum_type_unpack(repbody->oa.o_valid & OBD_MD_FLFLAGS ?
797 repbody->oa.o_flags : 0);
798 repbody->oa.o_flags = cksum_type_pack(cksum_type);
799 repbody->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
800 repbody->oa.o_cksum = ost_checksum_bulk(desc, OST_READ,cksum_type);
801 CDEBUG(D_PAGE, "checksum at read origin: %x\n",
802 repbody->oa.o_cksum);
804 repbody->oa.o_valid = 0;
806 /* We're finishing using body->oa as an input variable */
808 /* Check if client was evicted while we were doing i/o before touching
811 if (likely(!CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2)))
812 rc = target_bulk_io(exp, desc, &lwi);
817 /* Must commit after prep above in all cases */
818 rc = obd_commitrw(req->rq_svc_thread->t_env, OBD_BRW_READ, exp,
819 &repbody->oa, 1, ioo, remote_nb, npages, local_nb,
823 ost_drop_id(exp, &repbody->oa);
826 ost_brw_lock_put(LCK_PR, ioo, remote_nb, &lockh);
830 if (desc && !CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))
831 ptlrpc_free_bulk(desc);
835 req->rq_status = nob;
836 ptlrpc_lprocfs_brw(req, nob);
837 target_committed_to_req(req);
839 } else if (!no_reply) {
840 /* Only reply if there was no comms problem with bulk */
841 target_committed_to_req(req);
845 /* reply out callback would free */
846 ptlrpc_req_drop_rs(req);
847 LCONSOLE_WARN("%s: Bulk IO read error with %s (at %s), "
848 "client will retry: rc %d\n",
849 exp->exp_obd->obd_name,
850 obd_uuid2str(&exp->exp_client_uuid),
851 obd_export_nid2str(exp), rc);
853 /* send a bulk after reply to simulate a network delay or reordering
855 if (unlikely(CFS_FAIL_PRECHECK(OBD_FAIL_PTLRPC_CLIENT_BULK_CB2))) {
857 struct l_wait_info lwi1;
859 CDEBUG(D_INFO, "reorder BULK\n");
860 cfs_waitq_init(&waitq);
862 lwi1 = LWI_TIMEOUT_INTR(cfs_time_seconds(3), NULL, NULL, NULL);
863 l_wait_event(waitq, 0, &lwi1);
864 rc = target_bulk_io(exp, desc, &lwi);
865 ptlrpc_free_bulk(desc);
871 static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
873 struct ptlrpc_bulk_desc *desc = NULL;
874 struct obd_export *exp = req->rq_export;
875 struct niobuf_remote *remote_nb;
876 struct niobuf_local *local_nb;
877 struct obd_ioobj *ioo;
878 struct ost_body *body, *repbody;
879 struct l_wait_info lwi;
880 struct lustre_handle lockh = {0};
881 struct lustre_capa *capa = NULL;
883 int objcount, niocount, npages;
885 obd_count client_cksum = 0, server_cksum = 0;
886 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
887 int no_reply = 0, mmap = 0;
888 __u32 o_uid = 0, o_gid = 0;
889 struct ost_thread_local_cache *tls;
892 req->rq_bulk_write = 1;
894 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
895 GOTO(out, rc = -EIO);
896 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
897 GOTO(out, rc = -EFAULT);
899 /* pause before transaction has been started */
900 OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
902 /* ost_body, ioobj & noibuf_remote are verified and swabbed in
903 * ost_rw_hpreq_check(). */
904 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
906 GOTO(out, rc = -EFAULT);
908 objcount = req_capsule_get_size(&req->rq_pill, &RMF_OBD_IOOBJ,
909 RCL_CLIENT) / sizeof(*ioo);
910 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
912 GOTO(out, rc = -EFAULT);
914 rc = ost_validate_obdo(exp, &body->oa, ioo);
918 for (niocount = i = 0; i < objcount; i++)
919 niocount += ioo[i].ioo_bufcnt;
922 * It'd be nice to have a capsule function to indicate how many elements
923 * there were in a buffer for an RMF that's declared to be an array.
924 * It's easy enough to compute the number of elements here though.
926 remote_nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
927 if (remote_nb == NULL || niocount != (req_capsule_get_size(&req->rq_pill,
928 &RMF_NIOBUF_REMOTE, RCL_CLIENT) / sizeof(*remote_nb)))
929 GOTO(out, rc = -EFAULT);
931 if ((remote_nb[0].flags & OBD_BRW_MEMALLOC) &&
932 (exp->exp_connection->c_peer.nid == exp->exp_connection->c_self))
933 cfs_memory_pressure_set();
935 if (body->oa.o_valid & OBD_MD_FLOSSCAPA) {
936 capa = req_capsule_client_get(&req->rq_pill, &RMF_CAPA1);
938 CERROR("Missing capability for OST BRW WRITE");
939 GOTO(out, rc = -EFAULT);
943 req_capsule_set_size(&req->rq_pill, &RMF_RCS, RCL_SERVER,
944 niocount * sizeof(*rcs));
945 rc = req_capsule_server_pack(&req->rq_pill);
948 CFS_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, cfs_fail_val);
949 rcs = req_capsule_server_get(&req->rq_pill, &RMF_RCS);
951 tls = ost_tls_get(req);
953 GOTO(out_bulk, rc = -ENOMEM);
954 local_nb = tls->local;
956 rc = ost_brw_lock_get(LCK_PW, exp, ioo, remote_nb, &lockh);
961 * If getting the lock took more time than
962 * client was willing to wait, drop it. b=11330
964 if (cfs_time_current_sec() > req->rq_deadline ||
965 OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
967 CERROR("Dropping timed-out write from %s because locking "
968 "object "LPX64" took %ld seconds (limit was %ld).\n",
969 libcfs_id2str(req->rq_peer), ioo->ioo_id,
970 cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
971 req->rq_deadline - req->rq_arrival_time.tv_sec);
972 GOTO(out_lock, rc = -ETIMEDOUT);
975 /* obd_preprw clobbers oa->valid, so save what we need */
976 if (body->oa.o_valid & OBD_MD_FLCKSUM) {
977 client_cksum = body->oa.o_cksum;
978 if (body->oa.o_valid & OBD_MD_FLFLAGS)
979 cksum_type = cksum_type_unpack(body->oa.o_flags);
981 if (body->oa.o_valid & OBD_MD_FLFLAGS && body->oa.o_flags & OBD_FL_MMAP)
984 /* Because we already sync grant info with client when reconnect,
985 * grant info will be cleared for resent req, then fed_grant and
986 * total_grant will not be modified in following preprw_write */
987 if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
988 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
989 body->oa.o_valid &= ~OBD_MD_FLGRANT;
992 if (exp_connect_rmtclient(exp)) {
993 o_uid = body->oa.o_uid;
994 o_gid = body->oa.o_gid;
997 repbody = req_capsule_server_get(&req->rq_pill, &RMF_OST_BODY);
998 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
1000 npages = OST_THREAD_POOL_SIZE;
1001 rc = obd_preprw(req->rq_svc_thread->t_env, OBD_BRW_WRITE, exp,
1002 &repbody->oa, objcount, ioo, remote_nb, &npages,
1003 local_nb, oti, capa);
1007 desc = ptlrpc_prep_bulk_exp(req, npages,
1008 BULK_GET_SINK, OST_BULK_PORTAL);
1010 GOTO(skip_transfer, rc = -ENOMEM);
1012 /* NB Having prepped, we must commit... */
1014 for (i = 0; i < npages; i++)
1015 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
1016 local_nb[i].offset & ~CFS_PAGE_MASK,
1019 rc = sptlrpc_svc_prep_bulk(req, desc);
1023 rc = target_bulk_io(exp, desc, &lwi);
1027 if (client_cksum != 0 && rc == 0) {
1028 static int cksum_counter;
1029 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1030 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
1031 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
1032 server_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1033 repbody->oa.o_cksum = server_cksum;
1035 if (unlikely(client_cksum != server_cksum)) {
1036 CDEBUG_LIMIT(mmap ? D_INFO : D_ERROR,
1037 "client csum %x, server csum %x\n",
1038 client_cksum, server_cksum);
1040 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
1041 CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
1042 cksum_counter, libcfs_id2str(req->rq_peer),
1047 /* Must commit after prep above in all cases */
1048 rc = obd_commitrw(req->rq_svc_thread->t_env, OBD_BRW_WRITE, exp,
1049 &repbody->oa, objcount, ioo, remote_nb, npages,
1051 if (rc == -ENOTCONN)
1052 /* quota acquire process has been given up because
1053 * either the client has been evicted or the client
1054 * has timed out the request already */
1057 if (exp_connect_rmtclient(exp)) {
1058 repbody->oa.o_uid = o_uid;
1059 repbody->oa.o_gid = o_gid;
1063 * Disable sending mtime back to the client. If the client locked the
1064 * whole object, then it has already updated the mtime on its side,
1065 * otherwise it will have to glimpse anyway (see bug 21489, comment 32)
1067 repbody->oa.o_valid &= ~(OBD_MD_FLMTIME | OBD_MD_FLATIME);
1069 if (unlikely(client_cksum != server_cksum && rc == 0 && !mmap)) {
1070 int new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1075 if (new_cksum == server_cksum)
1076 msg = "changed in transit before arrival at OST";
1077 else if (new_cksum == client_cksum)
1078 msg = "initial checksum before message complete";
1080 msg = "changed in transit AND after initial checksum";
1082 if (req->rq_peer.nid == desc->bd_sender) {
1086 router = libcfs_nid2str(desc->bd_sender);
1089 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from "
1090 "%s%s%s inode "DFID" object "
1091 LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1092 exp->exp_obd->obd_name, msg,
1093 libcfs_id2str(req->rq_peer),
1095 body->oa.o_valid & OBD_MD_FLFID ?
1096 body->oa.o_parent_seq : (__u64)0,
1097 body->oa.o_valid & OBD_MD_FLFID ?
1098 body->oa.o_parent_oid : 0,
1099 body->oa.o_valid & OBD_MD_FLFID ?
1100 body->oa.o_parent_ver : 0,
1102 body->oa.o_valid & OBD_MD_FLGROUP ?
1103 body->oa.o_seq : (__u64)0,
1105 local_nb[npages-1].offset +
1106 local_nb[npages-1].len - 1 );
1107 CERROR("client csum %x, original server csum %x, "
1108 "server csum now %x\n",
1109 client_cksum, server_cksum, new_cksum);
1115 /* set per-requested niobuf return codes */
1116 for (i = j = 0; i < niocount; i++) {
1117 int len = remote_nb[i].len;
1122 LASSERT(j < npages);
1123 if (local_nb[j].rc < 0)
1124 rcs[i] = local_nb[j].rc;
1125 len -= local_nb[j].len;
1130 LASSERT(j == npages);
1131 ptlrpc_lprocfs_brw(req, nob);
1135 ost_brw_lock_put(LCK_PW, ioo, remote_nb, &lockh);
1140 ptlrpc_free_bulk(desc);
1143 oti_to_request(oti, req);
1144 target_committed_to_req(req);
1145 rc = ptlrpc_reply(req);
1146 } else if (!no_reply) {
1147 /* Only reply if there was no comms problem with bulk */
1148 target_committed_to_req(req);
1149 req->rq_status = rc;
1152 /* reply out callback would free */
1153 ptlrpc_req_drop_rs(req);
1154 LCONSOLE_WARN("%s: Bulk IO write error with %s (at %s), "
1155 "client will retry: rc %d\n",
1156 exp->exp_obd->obd_name,
1157 obd_uuid2str(&exp->exp_client_uuid),
1158 obd_export_nid2str(exp), rc);
1160 cfs_memory_pressure_clr();
1165 * Implementation of OST_SET_INFO.
1167 * OST_SET_INFO is like ioctl(): heavily overloaded. Specifically, it takes a
1168 * "key" and a value RPC buffers as arguments, with the value's contents
1169 * interpreted according to the key.
1171 * Value types that need swabbing have swabbing done explicitly, either here or
1172 * in functions called from here. This should be corrected: all swabbing should
1173 * be done in the capsule abstraction, as that will then allow us to move
1174 * swabbing exclusively to the client without having to modify server code
1175 * outside the capsule abstraction's implementation itself. To correct this
1176 * will require minor changes to the capsule abstraction; see the comments for
1177 * req_capsule_extend() in layout.c.
1179 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
1181 struct ost_body *body = NULL, *repbody;
1182 char *key, *val = NULL;
1183 int keylen, vallen, rc = 0;
1184 int is_grant_shrink = 0;
1187 key = req_capsule_client_get(&req->rq_pill, &RMF_SETINFO_KEY);
1189 DEBUG_REQ(D_HA, req, "no set_info key");
1192 keylen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_KEY,
1195 vallen = req_capsule_get_size(&req->rq_pill, &RMF_SETINFO_VAL,
1198 if ((is_grant_shrink = KEY_IS(KEY_GRANT_SHRINK)))
1199 /* In this case the value is actually an RMF_OST_BODY, so we
1200 * transmutate the type of this PTLRPC */
1201 req_capsule_extend(&req->rq_pill, &RQF_OST_SET_GRANT_INFO);
1203 rc = req_capsule_server_pack(&req->rq_pill);
1208 if (is_grant_shrink) {
1209 body = req_capsule_client_get(&req->rq_pill,
1214 repbody = req_capsule_server_get(&req->rq_pill,
1216 memcpy(repbody, body, sizeof(*body));
1217 val = (char*)repbody;
1219 val = req_capsule_client_get(&req->rq_pill,
1224 if (KEY_IS(KEY_EVICT_BY_NID)) {
1226 obd_export_evict_by_nid(exp->exp_obd, val);
1228 } else if (KEY_IS(KEY_MDS_CONN) && ptlrpc_req_need_swab(req)) {
1229 if (vallen < sizeof(__u32))
1231 __swab32s((__u32 *)val);
1234 /* OBD will also check if KEY_IS(KEY_GRANT_SHRINK), and will cast val to
1235 * a struct ost_body * value */
1236 rc = obd_set_info_async(req->rq_svc_thread->t_env, exp, keylen,
1237 key, vallen, val, NULL);
1239 lustre_msg_set_status(req->rq_repmsg, 0);
1243 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
1246 int keylen, replylen, rc = 0;
1247 struct req_capsule *pill = &req->rq_pill;
1250 /* this common part for get_info rpc */
1251 key = req_capsule_client_get(pill, &RMF_SETINFO_KEY);
1253 DEBUG_REQ(D_HA, req, "no get_info key");
1256 keylen = req_capsule_get_size(pill, &RMF_SETINFO_KEY, RCL_CLIENT);
1258 if (KEY_IS(KEY_FIEMAP)) {
1259 struct ll_fiemap_info_key *fm_key = key;
1262 rc = ost_validate_obdo(exp, &fm_key->oa, NULL);
1267 rc = obd_get_info(req->rq_svc_thread->t_env, exp, keylen, key,
1268 &replylen, NULL, NULL);
1272 req_capsule_set_size(pill, &RMF_GENERIC_DATA,
1273 RCL_SERVER, replylen);
1275 rc = req_capsule_server_pack(pill);
1279 reply = req_capsule_server_get(pill, &RMF_GENERIC_DATA);
1283 /* call again to fill in the reply buffer */
1284 rc = obd_get_info(req->rq_svc_thread->t_env, exp, keylen, key,
1285 &replylen, reply, NULL);
1287 lustre_msg_set_status(req->rq_repmsg, 0);
1291 #ifdef HAVE_QUOTA_SUPPORT
1292 static int ost_handle_quotactl(struct ptlrpc_request *req)
1294 struct obd_quotactl *oqctl, *repoqc;
1298 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1300 GOTO(out, rc = -EPROTO);
1302 rc = req_capsule_server_pack(&req->rq_pill);
1306 repoqc = req_capsule_server_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1307 req->rq_status = obd_quotactl(req->rq_export, oqctl);
1314 static int ost_handle_quotacheck(struct ptlrpc_request *req)
1316 struct obd_quotactl *oqctl;
1320 oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1324 rc = req_capsule_server_pack(&req->rq_pill);
1328 req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1332 static int ost_handle_quota_adjust_qunit(struct ptlrpc_request *req)
1334 struct quota_adjust_qunit *oqaq, *repoqa;
1335 struct lustre_quota_ctxt *qctxt;
1339 qctxt = &req->rq_export->exp_obd->u.obt.obt_qctxt;
1340 oqaq = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
1342 GOTO(out, rc = -EPROTO);
1344 rc = req_capsule_server_pack(&req->rq_pill);
1348 repoqa = req_capsule_server_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
1349 req->rq_status = obd_quota_adjust_qunit(req->rq_export, oqaq, qctxt, NULL);
1357 static int ost_llog_handle_connect(struct obd_export *exp,
1358 struct ptlrpc_request *req)
1360 struct llogd_conn_body *body;
1364 body = req_capsule_client_get(&req->rq_pill, &RMF_LLOGD_CONN_BODY);
1365 rc = obd_llog_connect(exp, body);
1369 #define ost_init_sec_none(reply, exp) \
1371 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT | \
1372 OBD_CONNECT_RMT_CLIENT_FORCE | \
1373 OBD_CONNECT_OSS_CAPA); \
1374 cfs_spin_lock(&exp->exp_lock); \
1375 exp->exp_connect_flags = reply->ocd_connect_flags; \
1376 cfs_spin_unlock(&exp->exp_lock); \
1379 static int ost_init_sec_level(struct ptlrpc_request *req)
1381 struct obd_export *exp = req->rq_export;
1382 struct req_capsule *pill = &req->rq_pill;
1383 struct obd_device *obd = exp->exp_obd;
1384 struct filter_obd *filter = &obd->u.filter;
1385 char *client = libcfs_nid2str(req->rq_peer.nid);
1386 struct obd_connect_data *data, *reply;
1390 data = req_capsule_client_get(pill, &RMF_CONNECT_DATA);
1391 reply = req_capsule_server_get(pill, &RMF_CONNECT_DATA);
1392 if (data == NULL || reply == NULL)
1395 /* connection from MDT is always trusted */
1396 if (req->rq_auth_usr_mdt) {
1397 ost_init_sec_none(reply, exp);
1401 /* no GSS support case */
1402 if (!req->rq_auth_gss) {
1403 if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
1404 CWARN("client %s -> target %s does not user GSS, "
1405 "can not run under security level %d.\n",
1406 client, obd->obd_name, filter->fo_sec_level);
1409 ost_init_sec_none(reply, exp);
1414 /* old version case */
1415 if (unlikely(!(data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT) ||
1416 !(data->ocd_connect_flags & OBD_CONNECT_OSS_CAPA))) {
1417 if (filter->fo_sec_level > LUSTRE_SEC_NONE) {
1418 CWARN("client %s -> target %s uses old version, "
1419 "can not run under security level %d.\n",
1420 client, obd->obd_name, filter->fo_sec_level);
1423 CWARN("client %s -> target %s uses old version, "
1424 "run under security level %d.\n",
1425 client, obd->obd_name, filter->fo_sec_level);
1426 ost_init_sec_none(reply, exp);
1431 remote = data->ocd_connect_flags & OBD_CONNECT_RMT_CLIENT_FORCE;
1433 if (!req->rq_auth_remote)
1434 CDEBUG(D_SEC, "client (local realm) %s -> target %s "
1435 "asked to be remote.\n", client, obd->obd_name);
1436 } else if (req->rq_auth_remote) {
1438 CDEBUG(D_SEC, "client (remote realm) %s -> target %s is set "
1439 "as remote by default.\n", client, obd->obd_name);
1443 if (!filter->fo_fl_oss_capa) {
1444 CDEBUG(D_SEC, "client %s -> target %s is set as remote,"
1445 " but OSS capabilities are not enabled: %d.\n",
1446 client, obd->obd_name, filter->fo_fl_oss_capa);
1451 switch (filter->fo_sec_level) {
1452 case LUSTRE_SEC_NONE:
1454 ost_init_sec_none(reply, exp);
1457 CDEBUG(D_SEC, "client %s -> target %s is set as remote, "
1458 "can not run under security level %d.\n",
1459 client, obd->obd_name, filter->fo_sec_level);
1462 case LUSTRE_SEC_REMOTE:
1464 ost_init_sec_none(reply, exp);
1466 case LUSTRE_SEC_ALL:
1468 reply->ocd_connect_flags &= ~(OBD_CONNECT_RMT_CLIENT |
1469 OBD_CONNECT_RMT_CLIENT_FORCE);
1470 if (!filter->fo_fl_oss_capa)
1471 reply->ocd_connect_flags &= ~OBD_CONNECT_OSS_CAPA;
1473 cfs_spin_lock(&exp->exp_lock);
1474 exp->exp_connect_flags = reply->ocd_connect_flags;
1475 cfs_spin_unlock(&exp->exp_lock);
1487 * this should be done in filter_connect()/filter_reconnect(), but
1488 * we can't obtain information like NID, which stored in incoming
1489 * request, thus can't decide what flavor to use. so we do it here.
1491 * This hack should be removed after the OST stack be rewritten, just
1492 * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect().
1494 static int ost_connect_check_sptlrpc(struct ptlrpc_request *req)
1496 struct obd_export *exp = req->rq_export;
1497 struct filter_obd *filter = &exp->exp_obd->u.filter;
1498 struct sptlrpc_flavor flvr;
1501 if (unlikely(strcmp(exp->exp_obd->obd_type->typ_name,
1502 LUSTRE_ECHO_NAME) == 0)) {
1503 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_ANY;
1507 if (exp->exp_flvr.sf_rpc == SPTLRPC_FLVR_INVALID) {
1508 cfs_read_lock(&filter->fo_sptlrpc_lock);
1509 sptlrpc_target_choose_flavor(&filter->fo_sptlrpc_rset,
1513 cfs_read_unlock(&filter->fo_sptlrpc_lock);
1515 cfs_spin_lock(&exp->exp_lock);
1517 exp->exp_sp_peer = req->rq_sp_from;
1518 exp->exp_flvr = flvr;
1520 if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_ANY &&
1521 exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
1522 CERROR("unauthorized rpc flavor %x from %s, "
1523 "expect %x\n", req->rq_flvr.sf_rpc,
1524 libcfs_nid2str(req->rq_peer.nid),
1525 exp->exp_flvr.sf_rpc);
1529 cfs_spin_unlock(&exp->exp_lock);
1531 if (exp->exp_sp_peer != req->rq_sp_from) {
1532 CERROR("RPC source %s doesn't match %s\n",
1533 sptlrpc_part2name(req->rq_sp_from),
1534 sptlrpc_part2name(exp->exp_sp_peer));
1537 rc = sptlrpc_target_export_check(exp, req);
1544 /* Ensure that data and metadata are synced to the disk when lock is cancelled
1546 int ost_blocking_ast(struct ldlm_lock *lock,
1547 struct ldlm_lock_desc *desc,
1548 void *data, int flag)
1550 __u32 sync_lock_cancel = 0;
1551 __u32 len = sizeof(sync_lock_cancel);
1555 rc = obd_get_info(NULL, lock->l_export, sizeof(KEY_SYNC_LOCK_CANCEL),
1556 KEY_SYNC_LOCK_CANCEL, &len, &sync_lock_cancel, NULL);
1558 if (!rc && flag == LDLM_CB_CANCELING &&
1559 (lock->l_granted_mode & (LCK_PW|LCK_GROUP)) &&
1560 (sync_lock_cancel == ALWAYS_SYNC_ON_CANCEL ||
1561 (sync_lock_cancel == BLOCKING_SYNC_ON_CANCEL &&
1562 lock->l_flags & LDLM_FL_CBPENDING))) {
1563 struct obd_info *oinfo;
1567 OBD_ALLOC_PTR(oinfo);
1572 OBD_FREE_PTR(oinfo);
1575 oa->o_id = lock->l_resource->lr_name.name[0];
1576 oa->o_seq = lock->l_resource->lr_name.name[1];
1577 oa->o_valid = OBD_MD_FLID|OBD_MD_FLGROUP;
1580 rc = obd_sync(NULL, lock->l_export, oinfo,
1581 lock->l_policy_data.l_extent.start,
1582 lock->l_policy_data.l_extent.end, NULL);
1584 CERROR("Error %d syncing data on lock cancel\n", rc);
1587 OBD_FREE_PTR(oinfo);
1590 rc = ldlm_server_blocking_ast(lock, desc, data, flag);
1594 static int ost_filter_recovery_request(struct ptlrpc_request *req,
1595 struct obd_device *obd, int *process)
1597 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1598 case OST_CONNECT: /* This will never get here, but for completeness. */
1599 case OST_DISCONNECT:
1610 case OBD_LOG_CANCEL:
1612 *process = target_queue_recovery_request(req, obd);
1616 DEBUG_REQ(D_WARNING, req, "not permitted during recovery");
1622 int ost_msg_check_version(struct lustre_msg *msg)
1626 switch(lustre_msg_get_opc(msg)) {
1628 case OST_DISCONNECT:
1631 case SEC_CTX_INIT_CONT:
1633 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1635 CERROR("bad opc %u version %08x, expecting %08x\n",
1636 lustre_msg_get_opc(msg),
1637 lustre_msg_get_version(msg),
1638 LUSTRE_OBD_VERSION);
1651 #ifdef HAVE_QUOTA_SUPPORT
1652 case OST_QUOTACHECK:
1654 case OST_QUOTA_ADJUST_QUNIT:
1656 rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION);
1658 CERROR("bad opc %u version %08x, expecting %08x\n",
1659 lustre_msg_get_opc(msg),
1660 lustre_msg_get_version(msg),
1661 LUSTRE_OST_VERSION);
1666 case LDLM_BL_CALLBACK:
1667 case LDLM_CP_CALLBACK:
1668 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1670 CERROR("bad opc %u version %08x, expecting %08x\n",
1671 lustre_msg_get_opc(msg),
1672 lustre_msg_get_version(msg),
1673 LUSTRE_DLM_VERSION);
1675 case LLOG_ORIGIN_CONNECT:
1676 case OBD_LOG_CANCEL:
1677 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1679 CERROR("bad opc %u version %08x, expecting %08x\n",
1680 lustre_msg_get_opc(msg),
1681 lustre_msg_get_version(msg),
1682 LUSTRE_LOG_VERSION);
1685 CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg));
1691 struct ost_prolong_data {
1692 struct ptlrpc_request *opd_req;
1693 struct obd_export *opd_exp;
1694 struct obdo *opd_oa;
1695 struct ldlm_res_id opd_resid;
1696 struct ldlm_extent opd_extent;
1697 ldlm_mode_t opd_mode;
1698 unsigned int opd_locks;
1702 /* prolong locks for the current service time of the corresponding
1703 * portal (= OST_IO_PORTAL)
1705 static inline int prolong_timeout(struct ptlrpc_request *req)
1707 struct ptlrpc_service *svc = req->rq_rqbd->rqbd_service;
1710 return obd_timeout / 2;
1712 return max(at_est2timeout(at_get(&svc->srv_at_estimate)), ldlm_timeout);
1715 static void ost_prolong_lock_one(struct ost_prolong_data *opd,
1716 struct ldlm_lock *lock)
1718 LASSERT(lock->l_export == opd->opd_exp);
1720 if (lock->l_destroyed) /* lock already cancelled */
1723 /* XXX: never try to grab resource lock here because we're inside
1724 * exp_bl_list_lock; in ldlm_lockd.c to handle waiting list we take
1725 * res lock and then exp_bl_list_lock. */
1727 if (!(lock->l_flags & LDLM_FL_AST_SENT))
1728 /* ignore locks not being cancelled */
1732 "refreshed for req x"LPU64" ext("LPU64"->"LPU64") to %ds.\n",
1733 opd->opd_req->rq_xid, opd->opd_extent.start,
1734 opd->opd_extent.end, opd->opd_timeout);
1736 /* OK. this is a possible lock the user holds doing I/O
1737 * let's refresh eviction timer for it */
1738 ldlm_refresh_waiting_lock(lock, opd->opd_timeout);
1742 static void ost_prolong_locks(struct ost_prolong_data *data)
1744 struct obd_export *exp = data->opd_exp;
1745 struct obdo *oa = data->opd_oa;
1746 struct ldlm_lock *lock;
1749 if (oa->o_valid & OBD_MD_FLHANDLE) {
1750 /* mostly a request should be covered by only one lock, try
1752 lock = ldlm_handle2lock(&oa->o_handle);
1754 /* Fast path to check if the lock covers the whole IO
1755 * region exclusively. */
1756 if (lock->l_granted_mode == LCK_PW &&
1757 ldlm_extent_contain(&lock->l_policy_data.l_extent,
1758 &data->opd_extent)) {
1760 ost_prolong_lock_one(data, lock);
1761 LDLM_LOCK_PUT(lock);
1764 LDLM_LOCK_PUT(lock);
1769 cfs_spin_lock_bh(&exp->exp_bl_list_lock);
1770 cfs_list_for_each_entry(lock, &exp->exp_bl_list, l_exp_list) {
1771 LASSERT(lock->l_flags & LDLM_FL_AST_SENT);
1772 LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
1774 if (!ldlm_res_eq(&data->opd_resid, &lock->l_resource->lr_name))
1777 if (!ldlm_extent_overlap(&lock->l_policy_data.l_extent,
1781 ost_prolong_lock_one(data, lock);
1783 cfs_spin_unlock_bh(&exp->exp_bl_list_lock);
1789 * Returns 1 if the given PTLRPC matches the given LDLM locks, or 0 if it does
1792 static int ost_rw_hpreq_lock_match(struct ptlrpc_request *req,
1793 struct ldlm_lock *lock)
1795 struct niobuf_remote *nb;
1796 struct obd_ioobj *ioo;
1798 struct ldlm_extent ext;
1801 opc = lustre_msg_get_opc(req->rq_reqmsg);
1802 LASSERT(opc == OST_READ || opc == OST_WRITE);
1804 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1805 LASSERT(ioo != NULL);
1807 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1808 LASSERT(nb != NULL);
1810 ext.start = nb->offset;
1811 nb += ioo->ioo_bufcnt - 1;
1812 ext.end = nb->offset + nb->len - 1;
1814 LASSERT(lock->l_resource != NULL);
1815 if (!osc_res_name_eq(ioo->ioo_id, ioo->ioo_seq,
1816 &lock->l_resource->lr_name))
1820 if (opc == OST_READ)
1822 if (!(lock->l_granted_mode & mode))
1825 RETURN(ldlm_extent_overlap(&lock->l_policy_data.l_extent, &ext));
1829 * High-priority queue request check for whether the given PTLRPC request (\a
1830 * req) is blocking an LDLM lock cancel.
1832 * Returns 1 if the given given PTLRPC request (\a req) is blocking an LDLM lock
1833 * cancel, 0 if it is not, and -EFAULT if the request is malformed.
1835 * Only OST_READs, OST_WRITEs and OST_PUNCHes go on the h-p RPC queue. This
1836 * function looks only at OST_READs and OST_WRITEs.
1838 static int ost_rw_hpreq_check(struct ptlrpc_request *req)
1840 struct obd_device *obd = req->rq_export->exp_obd;
1841 struct ost_body *body;
1842 struct obd_ioobj *ioo;
1843 struct niobuf_remote *nb;
1844 struct ost_prolong_data opd = { 0 };
1849 * Use LASSERT to do sanity check because malformed RPCs should have
1850 * been filtered out in ost_hpreq_handler().
1852 opc = lustre_msg_get_opc(req->rq_reqmsg);
1853 LASSERT(opc == OST_READ || opc == OST_WRITE);
1855 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1856 LASSERT(body != NULL);
1858 ioo = req_capsule_client_get(&req->rq_pill, &RMF_OBD_IOOBJ);
1859 LASSERT(ioo != NULL);
1861 nb = req_capsule_client_get(&req->rq_pill, &RMF_NIOBUF_REMOTE);
1862 LASSERT(nb != NULL);
1863 LASSERT(!(nb->flags & OBD_BRW_SRVLOCK));
1865 osc_build_res_name(ioo->ioo_id, ioo->ioo_seq, &opd.opd_resid);
1869 if (opc == OST_READ)
1871 opd.opd_mode = mode;
1872 opd.opd_exp = req->rq_export;
1873 opd.opd_oa = &body->oa;
1874 opd.opd_extent.start = nb->offset;
1875 nb += ioo->ioo_bufcnt - 1;
1876 opd.opd_extent.end = nb->offset + nb->len - 1;
1877 opd.opd_timeout = prolong_timeout(req);
1879 DEBUG_REQ(D_RPCTRACE, req,
1880 "%s %s: refresh rw locks: " LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
1881 obd->obd_name, cfs_current()->comm,
1882 opd.opd_resid.name[0], opd.opd_resid.name[1],
1883 opd.opd_extent.start, opd.opd_extent.end);
1885 ost_prolong_locks(&opd);
1887 CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
1888 obd->obd_name, opd.opd_locks, req);
1890 RETURN(opd.opd_locks);
1893 static void ost_rw_hpreq_fini(struct ptlrpc_request *req)
1895 (void)ost_rw_hpreq_check(req);
1899 * Like ost_rw_hpreq_lock_match(), but for OST_PUNCH RPCs.
1901 static int ost_punch_hpreq_lock_match(struct ptlrpc_request *req,
1902 struct ldlm_lock *lock)
1904 struct ost_body *body;
1907 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1908 LASSERT(body != NULL);
1910 if (body->oa.o_valid & OBD_MD_FLHANDLE &&
1911 body->oa.o_handle.cookie == lock->l_handle.h_cookie)
1918 * Like ost_rw_hpreq_check(), but for OST_PUNCH RPCs.
1920 static int ost_punch_hpreq_check(struct ptlrpc_request *req)
1922 struct obd_device *obd = req->rq_export->exp_obd;
1923 struct ost_body *body;
1925 struct ost_prolong_data opd = { 0 };
1929 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
1930 LASSERT(body != NULL);
1933 LASSERT(!(oa->o_valid & OBD_MD_FLFLAGS) ||
1934 !(oa->o_flags & OBD_FL_SRVLOCK));
1937 end = start + oa->o_blocks;
1940 opd.opd_mode = LCK_PW;
1941 opd.opd_exp = req->rq_export;
1943 opd.opd_extent.start = start;
1944 opd.opd_extent.end = end;
1945 if (oa->o_blocks == OBD_OBJECT_EOF)
1946 opd.opd_extent.end = OBD_OBJECT_EOF;
1947 opd.opd_timeout = prolong_timeout(req);
1949 osc_build_res_name(oa->o_id, oa->o_seq, &opd.opd_resid);
1952 "%s: refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
1954 opd.opd_resid.name[0], opd.opd_resid.name[1],
1955 opd.opd_extent.start, opd.opd_extent.end);
1957 ost_prolong_locks(&opd);
1959 CDEBUG(D_DLMTRACE, "%s: refreshed %u locks timeout for req %p.\n",
1960 obd->obd_name, opd.opd_locks, req);
1962 RETURN(opd.opd_locks > 0);
1965 static void ost_punch_hpreq_fini(struct ptlrpc_request *req)
1967 (void)ost_punch_hpreq_check(req);
1970 struct ptlrpc_hpreq_ops ost_hpreq_rw = {
1971 .hpreq_lock_match = ost_rw_hpreq_lock_match,
1972 .hpreq_check = ost_rw_hpreq_check,
1973 .hpreq_fini = ost_rw_hpreq_fini
1976 struct ptlrpc_hpreq_ops ost_hpreq_punch = {
1977 .hpreq_lock_match = ost_punch_hpreq_lock_match,
1978 .hpreq_check = ost_punch_hpreq_check,
1979 .hpreq_fini = ost_punch_hpreq_fini
1982 /** Assign high priority operations to the request if needed. */
1983 static int ost_hpreq_handler(struct ptlrpc_request *req)
1986 if (req->rq_export) {
1987 int opc = lustre_msg_get_opc(req->rq_reqmsg);
1988 struct ost_body *body;
1990 if (opc == OST_READ || opc == OST_WRITE) {
1991 struct niobuf_remote *nb;
1992 struct obd_ioobj *ioo;
1993 int objcount, niocount;
1997 /* RPCs on the H-P queue can be inspected before
1998 * ost_handler() initializes their pills, so we
1999 * initialize that here. Capsule initialization is
2000 * idempotent, as is setting the pill's format (provided
2001 * it doesn't change).
2003 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2004 if (opc == OST_READ)
2005 req_capsule_set(&req->rq_pill,
2008 req_capsule_set(&req->rq_pill,
2009 &RQF_OST_BRW_WRITE);
2011 body = req_capsule_client_get(&req->rq_pill,
2014 CERROR("Missing/short ost_body\n");
2018 objcount = req_capsule_get_size(&req->rq_pill,
2022 if (objcount == 0) {
2023 CERROR("Missing/short ioobj\n");
2027 CERROR("too many ioobjs (%d)\n", objcount);
2031 ioo = req_capsule_client_get(&req->rq_pill,
2034 CERROR("Missing/short ioobj\n");
2038 rc = ost_validate_obdo(req->rq_export, &body->oa, ioo);
2040 CERROR("invalid object ids\n");
2044 for (niocount = i = 0; i < objcount; i++) {
2045 if (ioo[i].ioo_bufcnt == 0) {
2046 CERROR("ioo[%d] has zero bufcnt\n", i);
2049 niocount += ioo[i].ioo_bufcnt;
2051 if (niocount > PTLRPC_MAX_BRW_PAGES) {
2052 DEBUG_REQ(D_RPCTRACE, req,
2053 "bulk has too many pages (%d)",
2058 nb = req_capsule_client_get(&req->rq_pill,
2059 &RMF_NIOBUF_REMOTE);
2061 CERROR("Missing/short niobuf\n");
2065 if (niocount == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
2066 req->rq_ops = &ost_hpreq_rw;
2067 } else if (opc == OST_PUNCH) {
2068 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2069 req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
2071 body = req_capsule_client_get(&req->rq_pill,
2074 CERROR("Missing/short ost_body\n");
2078 if (!(body->oa.o_valid & OBD_MD_FLFLAGS) ||
2079 !(body->oa.o_flags & OBD_FL_SRVLOCK))
2080 req->rq_ops = &ost_hpreq_punch;
2086 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
2087 int ost_handle(struct ptlrpc_request *req)
2089 struct obd_trans_info trans_info = { 0, };
2090 struct obd_trans_info *oti = &trans_info;
2091 int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
2092 struct obd_device *obd = NULL;
2095 /* OST module is kept between remounts, but the last reference
2096 * to specific module (say, osd or ofd) kills all related keys
2097 * from the environment. so we have to refill it until the root
2098 * cause is fixed properly */
2099 lu_env_refill(req->rq_svc_thread->t_env);
2101 LASSERT(current->journal_info == NULL);
2103 /* primordial rpcs don't affect server recovery */
2104 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2106 case SEC_CTX_INIT_CONT:
2111 req_capsule_init(&req->rq_pill, req, RCL_SERVER);
2113 if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) {
2114 if (!class_connected_export(req->rq_export)) {
2115 CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
2116 lustre_msg_get_opc(req->rq_reqmsg),
2117 libcfs_id2str(req->rq_peer));
2118 req->rq_status = -ENOTCONN;
2119 GOTO(out, rc = -ENOTCONN);
2122 obd = req->rq_export->exp_obd;
2124 /* Check for aborted recovery. */
2125 if (obd->obd_recovering) {
2126 rc = ost_filter_recovery_request(req, obd,
2128 if (rc || !should_process)
2130 else if (should_process < 0) {
2131 req->rq_status = should_process;
2132 rc = ptlrpc_error(req);
2140 rc = ost_msg_check_version(req->rq_reqmsg);
2144 if (req && req->rq_reqmsg && req->rq_export &&
2145 (req->rq_export->exp_connect_flags & OBD_CONNECT_JOBSTATS))
2146 oti->oti_jobid = lustre_msg_get_jobid(req->rq_reqmsg);
2148 switch (lustre_msg_get_opc(req->rq_reqmsg)) {
2150 CDEBUG(D_INODE, "connect\n");
2151 req_capsule_set(&req->rq_pill, &RQF_OST_CONNECT);
2152 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET))
2154 rc = target_handle_connect(req);
2155 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2))
2158 rc = ost_init_sec_level(req);
2160 rc = ost_connect_check_sptlrpc(req);
2164 case OST_DISCONNECT:
2165 CDEBUG(D_INODE, "disconnect\n");
2166 req_capsule_set(&req->rq_pill, &RQF_OST_DISCONNECT);
2167 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DISCONNECT_NET))
2169 rc = target_handle_disconnect(req);
2172 CDEBUG(D_INODE, "create\n");
2173 req_capsule_set(&req->rq_pill, &RQF_OST_CREATE);
2174 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET))
2176 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2177 GOTO(out, rc = -EROFS);
2178 rc = ost_create(req->rq_export, req, oti);
2181 CDEBUG(D_INODE, "destroy\n");
2182 req_capsule_set(&req->rq_pill, &RQF_OST_DESTROY);
2183 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET))
2185 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2186 GOTO(out, rc = -EROFS);
2187 rc = ost_destroy(req->rq_export, req, oti);
2190 CDEBUG(D_INODE, "getattr\n");
2191 req_capsule_set(&req->rq_pill, &RQF_OST_GETATTR);
2192 if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET))
2194 rc = ost_getattr(req->rq_export, req);
2197 CDEBUG(D_INODE, "setattr\n");
2198 req_capsule_set(&req->rq_pill, &RQF_OST_SETATTR);
2199 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET))
2201 rc = ost_setattr(req->rq_export, req, oti);
2204 req_capsule_set(&req->rq_pill, &RQF_OST_BRW_WRITE);
2205 CDEBUG(D_INODE, "write\n");
2206 /* req->rq_request_portal would be nice, if it was set */
2207 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
2208 CERROR("%s: deny write request from %s to portal %u\n",
2209 req->rq_export->exp_obd->obd_name,
2210 obd_export_nid2str(req->rq_export),
2211 req->rq_rqbd->rqbd_service->srv_req_portal);
2212 GOTO(out, rc = -EPROTO);
2214 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
2216 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
2217 GOTO(out, rc = -ENOSPC);
2218 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2219 GOTO(out, rc = -EROFS);
2220 rc = ost_brw_write(req, oti);
2221 LASSERT(current->journal_info == NULL);
2222 /* ost_brw_write sends its own replies */
2225 req_capsule_set(&req->rq_pill, &RQF_OST_BRW_READ);
2226 CDEBUG(D_INODE, "read\n");
2227 /* req->rq_request_portal would be nice, if it was set */
2228 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
2229 CERROR("%s: deny read request from %s to portal %u\n",
2230 req->rq_export->exp_obd->obd_name,
2231 obd_export_nid2str(req->rq_export),
2232 req->rq_rqbd->rqbd_service->srv_req_portal);
2233 GOTO(out, rc = -EPROTO);
2235 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
2237 rc = ost_brw_read(req, oti);
2238 LASSERT(current->journal_info == NULL);
2239 /* ost_brw_read sends its own replies */
2242 CDEBUG(D_INODE, "punch\n");
2243 req_capsule_set(&req->rq_pill, &RQF_OST_PUNCH);
2244 if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET))
2246 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
2247 GOTO(out, rc = -EROFS);
2248 rc = ost_punch(req->rq_export, req, oti);
2251 CDEBUG(D_INODE, "statfs\n");
2252 req_capsule_set(&req->rq_pill, &RQF_OST_STATFS);
2253 if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET))
2255 rc = ost_statfs(req);
2258 CDEBUG(D_INODE, "sync\n");
2259 req_capsule_set(&req->rq_pill, &RQF_OST_SYNC);
2260 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET))
2262 rc = ost_sync(req->rq_export, req, oti);
2265 DEBUG_REQ(D_INODE, req, "set_info");
2266 req_capsule_set(&req->rq_pill, &RQF_OBD_SET_INFO);
2267 rc = ost_set_info(req->rq_export, req);
2270 DEBUG_REQ(D_INODE, req, "get_info");
2271 req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC);
2272 rc = ost_get_info(req->rq_export, req);
2274 #ifdef HAVE_QUOTA_SUPPORT
2275 case OST_QUOTACHECK:
2276 CDEBUG(D_INODE, "quotacheck\n");
2277 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK);
2278 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACHECK_NET))
2280 rc = ost_handle_quotacheck(req);
2283 CDEBUG(D_INODE, "quotactl\n");
2284 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACTL);
2285 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACTL_NET))
2287 rc = ost_handle_quotactl(req);
2289 case OST_QUOTA_ADJUST_QUNIT:
2290 CDEBUG(D_INODE, "quota_adjust_qunit\n");
2291 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTA_ADJUST_QUNIT);
2292 rc = ost_handle_quota_adjust_qunit(req);
2296 DEBUG_REQ(D_INODE, req, "ping");
2297 req_capsule_set(&req->rq_pill, &RQF_OBD_PING);
2298 rc = target_handle_ping(req);
2300 /* FIXME - just reply status */
2301 case LLOG_ORIGIN_CONNECT:
2302 DEBUG_REQ(D_INODE, req, "log connect");
2303 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_CONNECT);
2304 rc = ost_llog_handle_connect(req->rq_export, req);
2305 req->rq_status = rc;
2306 rc = req_capsule_server_pack(&req->rq_pill);
2309 RETURN(ptlrpc_reply(req));
2310 case OBD_LOG_CANCEL:
2311 CDEBUG(D_INODE, "log cancel\n");
2312 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
2313 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
2315 rc = llog_origin_handle_cancel(req);
2316 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_REP))
2318 req->rq_status = rc;
2319 rc = req_capsule_server_pack(&req->rq_pill);
2322 RETURN(ptlrpc_reply(req));
2324 CDEBUG(D_INODE, "enqueue\n");
2325 req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE);
2326 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE))
2328 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
2330 ldlm_server_glimpse_ast);
2331 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
2334 CDEBUG(D_INODE, "convert\n");
2335 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
2336 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT))
2338 rc = ldlm_handle_convert(req);
2341 CDEBUG(D_INODE, "cancel\n");
2342 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
2343 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
2345 rc = ldlm_handle_cancel(req);
2347 case LDLM_BL_CALLBACK:
2348 case LDLM_CP_CALLBACK:
2349 CDEBUG(D_INODE, "callback\n");
2350 CERROR("callbacks should not happen on OST\n");
2353 CERROR("Unexpected opcode %d\n",
2354 lustre_msg_get_opc(req->rq_reqmsg));
2355 req->rq_status = -ENOTSUPP;
2356 rc = ptlrpc_error(req);
2360 LASSERT(current->journal_info == NULL);
2363 /* If we're DISCONNECTing, the export_data is already freed */
2364 if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT)
2365 target_committed_to_req(req);
2369 oti_to_request(oti, req);
2371 target_send_reply(req, rc, fail);
2374 EXPORT_SYMBOL(ost_handle);
2376 * free per-thread pool created by ost_thread_init().
2378 static void ost_thread_done(struct ptlrpc_thread *thread)
2380 struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local
2385 LASSERT(thread != NULL);
2388 * be prepared to handle partially-initialized pools (because this is
2389 * called from ost_thread_init() for cleanup.
2391 tls = thread->t_data;
2394 thread->t_data = NULL;
2400 * initialize per-thread page pool (bug 5137).
2402 static int ost_thread_init(struct ptlrpc_thread *thread)
2404 struct ost_thread_local_cache *tls;
2408 LASSERT(thread != NULL);
2409 LASSERT(thread->t_data == NULL);
2410 LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
2415 thread->t_data = tls;
2419 #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
2421 /* Sigh - really, this is an OSS, the _server_, not the _target_ */
2422 static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
2424 struct ost_obd *ost = &obd->u.ost;
2425 struct lprocfs_static_vars lvars;
2426 int oss_min_threads;
2427 int oss_max_threads;
2428 int oss_min_create_threads;
2429 int oss_max_create_threads;
2433 rc = cfs_cleanup_group_info();
2437 lprocfs_ost_init_vars(&lvars);
2438 lprocfs_obd_setup(obd, lvars.obd_vars);
2440 cfs_mutex_init(&ost->ost_health_mutex);
2442 if (oss_num_threads) {
2443 /* If oss_num_threads is set, it is the min and the max. */
2444 if (oss_num_threads > OSS_THREADS_MAX)
2445 oss_num_threads = OSS_THREADS_MAX;
2446 if (oss_num_threads < OSS_THREADS_MIN)
2447 oss_num_threads = OSS_THREADS_MIN;
2448 oss_max_threads = oss_min_threads = oss_num_threads;
2450 /* Base min threads on memory and cpus */
2452 cfs_num_online_cpus() * CFS_NUM_CACHEPAGES >>
2453 (27 - CFS_PAGE_SHIFT);
2454 if (oss_min_threads < OSS_THREADS_MIN)
2455 oss_min_threads = OSS_THREADS_MIN;
2456 /* Insure a 4x range for dynamic threads */
2457 if (oss_min_threads > OSS_THREADS_MAX / 4)
2458 oss_min_threads = OSS_THREADS_MAX / 4;
2459 oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4 + 1);
2463 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2464 OST_MAXREPSIZE, OST_REQUEST_PORTAL,
2465 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2466 ost_handle, LUSTRE_OSS_NAME,
2467 obd->obd_proc_entry, target_print_req,
2468 oss_min_threads, oss_max_threads,
2469 "ll_ost", LCT_DT_THREAD, NULL);
2470 if (ost->ost_service == NULL) {
2471 CERROR("failed to start service\n");
2472 GOTO(out_lprocfs, rc = -ENOMEM);
2475 rc = ptlrpc_start_threads(ost->ost_service);
2477 GOTO(out_service, rc = -EINVAL);
2479 if (oss_num_create_threads) {
2480 if (oss_num_create_threads > OSS_MAX_CREATE_THREADS)
2481 oss_num_create_threads = OSS_MAX_CREATE_THREADS;
2482 if (oss_num_create_threads < OSS_MIN_CREATE_THREADS)
2483 oss_num_create_threads = OSS_MIN_CREATE_THREADS;
2484 oss_min_create_threads = oss_max_create_threads =
2485 oss_num_create_threads;
2487 oss_min_create_threads = OSS_MIN_CREATE_THREADS;
2488 oss_max_create_threads = OSS_MAX_CREATE_THREADS;
2491 ost->ost_create_service =
2492 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2493 OST_MAXREPSIZE, OST_CREATE_PORTAL,
2494 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2495 ost_handle, "ost_create",
2496 obd->obd_proc_entry, target_print_req,
2497 oss_min_create_threads, oss_max_create_threads,
2498 "ll_ost_creat", LCT_DT_THREAD, NULL);
2499 if (ost->ost_create_service == NULL) {
2500 CERROR("failed to start OST create service\n");
2501 GOTO(out_service, rc = -ENOMEM);
2504 rc = ptlrpc_start_threads(ost->ost_create_service);
2506 GOTO(out_create, rc = -EINVAL);
2508 ost->ost_io_service =
2509 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
2510 OST_MAXREPSIZE, OST_IO_PORTAL,
2511 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
2512 ost_handle, "ost_io",
2513 obd->obd_proc_entry, target_print_req,
2514 oss_min_threads, oss_max_threads,
2515 "ll_ost_io", LCT_DT_THREAD, ost_hpreq_handler);
2516 if (ost->ost_io_service == NULL) {
2517 CERROR("failed to start OST I/O service\n");
2518 GOTO(out_create, rc = -ENOMEM);
2521 ost->ost_io_service->srv_init = ost_thread_init;
2522 ost->ost_io_service->srv_done = ost_thread_done;
2523 ost->ost_io_service->srv_cpu_affinity = 1;
2524 rc = ptlrpc_start_threads(ost->ost_io_service);
2526 GOTO(out_io, rc = -EINVAL);
2528 ping_evictor_start();
2533 ptlrpc_unregister_service(ost->ost_io_service);
2534 ost->ost_io_service = NULL;
2536 ptlrpc_unregister_service(ost->ost_create_service);
2537 ost->ost_create_service = NULL;
2539 ptlrpc_unregister_service(ost->ost_service);
2540 ost->ost_service = NULL;
2542 lprocfs_obd_cleanup(obd);
2546 static int ost_cleanup(struct obd_device *obd)
2548 struct ost_obd *ost = &obd->u.ost;
2552 ping_evictor_stop();
2554 /* there is no recovery for OST OBD, all recovery is controlled by
2556 LASSERT(obd->obd_recovering == 0);
2557 cfs_mutex_lock(&ost->ost_health_mutex);
2558 ptlrpc_unregister_service(ost->ost_service);
2559 ptlrpc_unregister_service(ost->ost_create_service);
2560 ptlrpc_unregister_service(ost->ost_io_service);
2561 ost->ost_service = NULL;
2562 ost->ost_create_service = NULL;
2563 cfs_mutex_unlock(&ost->ost_health_mutex);
2565 lprocfs_obd_cleanup(obd);
2570 static int ost_health_check(const struct lu_env *env, struct obd_device *obd)
2572 struct ost_obd *ost = &obd->u.ost;
2575 cfs_mutex_lock(&ost->ost_health_mutex);
2576 rc |= ptlrpc_service_health_check(ost->ost_service);
2577 rc |= ptlrpc_service_health_check(ost->ost_create_service);
2578 rc |= ptlrpc_service_health_check(ost->ost_io_service);
2579 cfs_mutex_unlock(&ost->ost_health_mutex);
2582 * health_check to return 0 on healthy
2583 * and 1 on unhealthy.
2591 struct ost_thread_local_cache *ost_tls(struct ptlrpc_request *r)
2593 return (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
2596 /* use obd ops to offer management infrastructure */
2597 static struct obd_ops ost_obd_ops = {
2598 .o_owner = THIS_MODULE,
2599 .o_setup = ost_setup,
2600 .o_cleanup = ost_cleanup,
2601 .o_health_check = ost_health_check,
2605 static int __init ost_init(void)
2607 struct lprocfs_static_vars lvars;
2611 lprocfs_ost_init_vars(&lvars);
2612 rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
2613 LUSTRE_OSS_NAME, NULL);
2615 if (ost_num_threads != 0 && oss_num_threads == 0) {
2616 LCONSOLE_INFO("ost_num_threads module parameter is deprecated, "
2617 "use oss_num_threads instead or unset both for "
2618 "dynamic thread startup\n");
2619 oss_num_threads = ost_num_threads;
2625 static void /*__exit*/ ost_exit(void)
2627 class_unregister_type(LUSTRE_OSS_NAME);
2630 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2631 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
2632 MODULE_LICENSE("GPL");
2634 module_init(ost_init);
2635 module_exit(ost_exit);