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) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Author: Alex Tomas <bzzz@whamcloud.com>
41 #define DEBUG_SUBSYSTEM S_FILTER
43 #include "ofd_internal.h"
45 static int ofd_preprw_read(const struct lu_env *env, struct obd_export *exp,
46 struct ofd_device *ofd, struct lu_fid *fid,
47 struct lu_attr *la, int niocount,
48 struct niobuf_remote *rnb, int *nr_local,
49 struct niobuf_local *lnb,
50 struct obd_trans_info *oti)
52 struct ofd_object *fo;
53 int i, j, rc, tot_bytes = 0;
58 fo = ofd_object_find(env, ofd, fid);
63 ofd_read_lock(env, fo);
64 if (!ofd_object_exists(fo))
65 GOTO(unlock, rc = -ENOENT);
67 /* parse remote buffers to local buffers and prepare the latter */
69 for (i = 0, j = 0; i < niocount; i++) {
70 rc = dt_bufs_get(env, ofd_object_child(fo), rnb + i,
71 lnb + j, 0, ofd_object_capa(env, fo));
74 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
75 /* correct index for local buffers to continue with */
78 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
79 tot_bytes += rnb[i].rnb_len;
82 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
83 rc = dt_attr_get(env, ofd_object_child(fo), la,
84 ofd_object_capa(env, fo));
88 rc = dt_read_prep(env, ofd_object_child(fo), lnb, *nr_local);
91 lprocfs_counter_add(ofd_obd(ofd)->obd_stats,
92 LPROC_OFD_READ_BYTES, tot_bytes);
93 ofd_counter_incr(exp, LPROC_OFD_STATS_READ,
94 oti->oti_jobid, tot_bytes);
98 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
100 ofd_read_unlock(env, fo);
101 ofd_object_put(env, fo);
105 static int ofd_preprw_write(const struct lu_env *env, struct obd_export *exp,
106 struct ofd_device *ofd, struct lu_fid *fid,
107 struct lu_attr *la, struct obdo *oa,
108 int objcount, struct obd_ioobj *obj,
109 struct niobuf_remote *rnb, int *nr_local,
110 struct niobuf_local *lnb,
111 struct obd_trans_info *oti)
113 struct ofd_object *fo;
114 int i, j, k, rc = 0, tot_bytes = 0;
117 LASSERT(env != NULL);
118 LASSERT(objcount == 1);
120 if (unlikely(exp->exp_obd->obd_recovering)) {
121 struct ofd_thread_info *info = ofd_info(env);
123 /* copied from ofd_precreate_object */
124 /* XXX this should be consolidated to use the same code
125 * instead of a copy, due to the ongoing risk of bugs. */
126 memset(&info->fti_attr, 0, sizeof(info->fti_attr));
127 info->fti_attr.la_valid = LA_TYPE | LA_MODE;
128 info->fti_attr.la_mode = S_IFREG | S_ISUID | S_ISGID | 0666;
129 info->fti_attr.la_valid |= LA_ATIME | LA_MTIME | LA_CTIME;
130 /* Initialize a/c/m time so any client timestamp will always
131 * be newer and update the inode. ctime = 0 is also handled
132 * specially in osd_inode_setattr(). See LU-221, LU-1042 */
133 info->fti_attr.la_atime = 0;
134 info->fti_attr.la_mtime = 0;
135 info->fti_attr.la_ctime = 0;
137 fo = ofd_object_find_or_create(env, ofd, fid, &info->fti_attr);
139 fo = ofd_object_find(env, ofd, fid);
143 GOTO(out, rc = PTR_ERR(fo));
146 ofd_read_lock(env, fo);
147 if (!ofd_object_exists(fo)) {
148 CERROR("%s: BRW to missing obj "DOSTID"\n",
149 exp->exp_obd->obd_name, POSTID(&obj->ioo_oid));
150 ofd_read_unlock(env, fo);
151 ofd_object_put(env, fo);
152 GOTO(out, rc = -ENOENT);
155 /* Always sync if syncjournal parameter is set */
156 oti->oti_sync_write = ofd->ofd_syncjournal;
158 /* Process incoming grant info, set OBD_BRW_GRANTED flag and grant some
159 * space back if possible */
160 ofd_grant_prepare_write(env, exp, oa, rnb, obj->ioo_bufcnt);
162 /* parse remote buffers to local buffers and prepare the latter */
164 for (i = 0, j = 0; i < obj->ioo_bufcnt; i++) {
165 rc = dt_bufs_get(env, ofd_object_child(fo),
167 ofd_object_capa(env, fo));
168 if (unlikely(rc < 0))
170 LASSERT(rc <= PTLRPC_MAX_BRW_PAGES);
171 /* correct index for local buffers to continue with */
172 for (k = 0; k < rc; k++) {
173 lnb[j+k].lnb_flags = rnb[i].rnb_flags;
174 if (!(rnb[i].rnb_flags & OBD_BRW_GRANTED))
175 lnb[j+k].lnb_rc = -ENOSPC;
176 if (!(rnb[i].rnb_flags & OBD_BRW_ASYNC))
177 oti->oti_sync_write = 1;
178 /* remote client can't break through quota */
179 if (exp_connect_rmtclient(exp))
180 lnb[j+k].lnb_flags &= ~OBD_BRW_NOQUOTA;
184 LASSERT(j <= PTLRPC_MAX_BRW_PAGES);
185 tot_bytes += rnb[i].rnb_len;
187 LASSERT(*nr_local > 0 && *nr_local <= PTLRPC_MAX_BRW_PAGES);
189 rc = dt_write_prep(env, ofd_object_child(fo), lnb, *nr_local);
190 if (unlikely(rc != 0))
193 lprocfs_counter_add(ofd_obd(ofd)->obd_stats,
194 LPROC_OFD_WRITE_BYTES, tot_bytes);
195 ofd_counter_incr(exp, LPROC_OFD_STATS_WRITE,
196 oti->oti_jobid, tot_bytes);
199 dt_bufs_put(env, ofd_object_child(fo), lnb, *nr_local);
200 ofd_read_unlock(env, fo);
201 /* ofd_grant_prepare_write() was called, so we must commit */
202 ofd_grant_commit(env, exp, rc);
204 /* let's still process incoming grant information packed in the oa,
205 * but without enforcing grant since we won't proceed with the write.
206 * Just like a read request actually. */
207 ofd_grant_prepare_read(env, exp, oa);
211 int ofd_preprw(const struct lu_env* env, int cmd, struct obd_export *exp,
212 struct obdo *oa, int objcount, struct obd_ioobj *obj,
213 struct niobuf_remote *rnb, int *nr_local,
214 struct niobuf_local *lnb, struct obd_trans_info *oti,
215 struct lustre_capa *capa)
217 struct ofd_device *ofd = ofd_exp(exp);
218 struct ofd_thread_info *info;
221 rc = lu_env_refill((struct lu_env *)env);
223 info = ofd_info_init(env, exp);
227 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOENT)) {
228 struct ofd_seq *oseq;
229 oseq = ofd_seq_load(env, ofd, ostid_seq(&oa->o_oi));
231 CERROR("%s: Can not find seq for "DOSTID
232 ": rc = %ld\n", ofd_name(ofd), POSTID(&oa->o_oi),
237 if (oseq->os_destroys_in_progress == 0) {
238 /* don't fail lookups for orphan recovery, it causes
239 * later LBUGs when objects still exist during
241 ofd_seq_put(env, oseq);
244 ofd_seq_put(env, oseq);
247 LASSERT(objcount == 1);
248 LASSERT(obj->ioo_bufcnt > 0);
250 rc = ostid_to_fid(&info->fti_fid, &oa->o_oi, 0);
251 if (unlikely(rc != 0))
254 if (cmd == OBD_BRW_WRITE) {
255 rc = ofd_auth_capa(exp, &info->fti_fid, ostid_seq(&oa->o_oi),
256 capa, CAPA_OPC_OSS_WRITE);
258 la_from_obdo(&info->fti_attr, oa, OBD_MD_FLGETATTR);
259 rc = ofd_preprw_write(env, exp, ofd, &info->fti_fid,
260 &info->fti_attr, oa, objcount,
261 obj, rnb, nr_local, lnb, oti);
263 } else if (cmd == OBD_BRW_READ) {
264 rc = ofd_auth_capa(exp, &info->fti_fid, ostid_seq(&oa->o_oi),
265 capa, CAPA_OPC_OSS_READ);
267 ofd_grant_prepare_read(env, exp, oa);
268 rc = ofd_preprw_read(env, exp, ofd, &info->fti_fid,
269 &info->fti_attr, obj->ioo_bufcnt,
270 rnb, nr_local, lnb, oti);
271 obdo_from_la(oa, &info->fti_attr, LA_ATIME);
274 CERROR("%s: wrong cmd %d received!\n",
275 exp->exp_obd->obd_name, cmd);
282 ofd_commitrw_read(const struct lu_env *env, struct ofd_device *ofd,
283 struct lu_fid *fid, int objcount, int niocount,
284 struct niobuf_local *lnb)
286 struct ofd_object *fo;
290 LASSERT(niocount > 0);
292 fo = ofd_object_find(env, ofd, fid);
296 LASSERT(ofd_object_exists(fo));
297 dt_bufs_put(env, ofd_object_child(fo), lnb, niocount);
299 ofd_read_unlock(env, fo);
300 ofd_object_put(env, fo);
301 /* second put is pair to object_get in ofd_preprw_read */
302 ofd_object_put(env, fo);
308 ofd_write_attr_set(const struct lu_env *env, struct ofd_device *ofd,
309 struct ofd_object *ofd_obj, struct lu_attr *la,
310 struct filter_fid *ff)
312 struct ofd_thread_info *info = ofd_info(env);
313 __u64 valid = la->la_valid;
316 struct dt_object *dt_obj;
323 dt_obj = ofd_object_child(ofd_obj);
324 LASSERT(dt_obj != NULL);
326 la->la_valid &= LA_UID | LA_GID;
328 rc = ofd_attr_handle_ugid(env, ofd_obj, la, 0 /* !is_setattr */);
333 rc = ofd_object_ff_check(env, ofd_obj);
340 if (!la->la_valid && !ff_needed)
341 /* no attributes to set */
344 th = ofd_trans_create(env, ofd);
346 GOTO(out, rc = PTR_ERR(th));
349 rc = dt_declare_attr_set(env, dt_obj, la, th);
355 info->fti_buf.lb_buf = ff;
356 info->fti_buf.lb_len = sizeof(*ff);
357 rc = dt_declare_xattr_set(env, dt_obj, &info->fti_buf,
358 XATTR_NAME_FID, 0, th);
363 /* We don't need a transno for this operation which will be re-executed
364 * anyway when the OST_WRITE (with a transno assigned) is replayed */
365 rc = dt_trans_start_local(env, ofd->ofd_osd , th);
371 rc = dt_attr_set(env, dt_obj, la, th,
372 ofd_object_capa(env, ofd_obj));
377 /* set filter fid EA */
379 rc = dt_xattr_set(env, dt_obj, &info->fti_buf, XATTR_NAME_FID,
387 dt_trans_stop(env, ofd->ofd_osd, th);
389 la->la_valid = valid;
394 ofd_commitrw_write(const struct lu_env *env, struct ofd_device *ofd,
395 struct lu_fid *fid, struct lu_attr *la,
396 struct filter_fid *ff, int objcount,
397 int niocount, struct niobuf_local *lnb,
398 struct obd_trans_info *oti, int old_rc)
400 struct ofd_thread_info *info = ofd_info(env);
401 struct ofd_object *fo;
409 LASSERT(objcount == 1);
411 fo = ofd_object_find(env, ofd, fid);
413 LASSERT(ofd_object_exists(fo));
415 o = ofd_object_child(fo);
419 GOTO(out, rc = old_rc);
422 * The first write to each object must set some attributes. It is
423 * important to set the uid/gid before calling
424 * dt_declare_write_commit() since quota enforcement is now handled in
427 rc = ofd_write_attr_set(env, ofd, fo, la, ff);
431 la->la_valid &= LA_ATIME | LA_MTIME | LA_CTIME;
434 th = ofd_trans_create(env, ofd);
436 GOTO(out, rc = PTR_ERR(th));
438 th->th_sync |= oti->oti_sync_write;
440 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DQACQ_NET))
441 GOTO(out_stop, rc = -EINPROGRESS);
443 rc = dt_declare_write_commit(env, o, lnb, niocount, th);
448 /* update [mac]time if needed */
449 rc = dt_declare_attr_set(env, o, la, th);
454 rc = ofd_trans_start(env, ofd, fo, th);
458 rc = dt_write_commit(env, o, lnb, niocount, th);
463 rc = dt_attr_set(env, o, la, th, ofd_object_capa(env, fo));
468 /* get attr to return */
469 rc = dt_attr_get(env, o, la, ofd_object_capa(env, fo));
472 /* Force commit to make the just-deleted blocks
473 * reusable. LU-456 */
477 ofd_trans_stop(env, ofd, th, rc);
478 if (rc == -ENOSPC && retries++ < 3) {
479 CDEBUG(D_INODE, "retry after force commit, retries:%d\n",
485 dt_bufs_put(env, o, lnb, niocount);
486 ofd_read_unlock(env, fo);
487 ofd_object_put(env, fo);
488 /* second put is pair to object_get in ofd_preprw_write */
489 ofd_object_put(env, fo);
490 ofd_grant_commit(env, info->fti_exp, old_rc);
494 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
495 struct obdo *oa, int objcount, struct obd_ioobj *obj,
496 struct niobuf_remote *rnb, int npages,
497 struct niobuf_local *lnb, struct obd_trans_info *oti,
500 struct ofd_thread_info *info;
501 struct ofd_mod_data *fmd;
503 struct ofd_device *ofd = ofd_exp(exp);
504 struct filter_fid *ff = NULL;
507 info = ofd_info(env);
508 ofd_oti2info(info, oti);
512 rc = ostid_to_fid(&info->fti_fid, &oa->o_oi, 0);
513 if (unlikely(rc != 0))
515 if (cmd == OBD_BRW_WRITE) {
516 /* Don't update timestamps if this write is older than a
517 * setattr which modifies the timestamps. b=10150 */
519 /* XXX when we start having persistent reservations this needs
520 * to be changed to ofd_fmd_get() to create the fmd if it
521 * doesn't already exist so we can store the reservation handle
523 valid = OBD_MD_FLUID | OBD_MD_FLGID;
524 fmd = ofd_fmd_find(exp, &info->fti_fid);
525 if (!fmd || fmd->fmd_mactime_xid < info->fti_xid)
526 valid |= OBD_MD_FLATIME | OBD_MD_FLMTIME |
528 ofd_fmd_put(exp, fmd);
529 la_from_obdo(&info->fti_attr, oa, valid);
531 if (oa->o_valid & OBD_MD_FLFID) {
532 ff = &info->fti_mds_fid;
533 ofd_prepare_fidea(ff, oa);
536 rc = ofd_commitrw_write(env, ofd, &info->fti_fid,
537 &info->fti_attr, ff, objcount, npages,
540 obdo_from_la(oa, &info->fti_attr,
541 OFD_VALID_FLAGS | LA_GID | LA_UID);
543 obdo_from_la(oa, &info->fti_attr, LA_GID | LA_UID);
545 /* don't report overquota flag if we failed before reaching
547 if (old_rc == 0 && (rc == 0 || rc == -EDQUOT)) {
548 /* return the overquota flags to client */
549 if (lnb[0].lnb_flags & OBD_BRW_OVER_USRQUOTA) {
550 if (oa->o_valid & OBD_MD_FLFLAGS)
551 oa->o_flags |= OBD_FL_NO_USRQUOTA;
553 oa->o_flags = OBD_FL_NO_USRQUOTA;
556 if (lnb[0].lnb_flags & OBD_BRW_OVER_GRPQUOTA) {
557 if (oa->o_valid & OBD_MD_FLFLAGS)
558 oa->o_flags |= OBD_FL_NO_GRPQUOTA;
560 oa->o_flags = OBD_FL_NO_GRPQUOTA;
563 oa->o_valid |= OBD_MD_FLFLAGS;
564 oa->o_valid |= OBD_MD_FLUSRQUOTA | OBD_MD_FLGRPQUOTA;
566 } else if (cmd == OBD_BRW_READ) {
567 struct ldlm_namespace *ns = ofd->ofd_namespace;
569 /* If oa != NULL then ofd_preprw_read updated the inode
570 * atime and we should update the lvb so that other glimpses
571 * will also get the updated value. bug 5972 */
572 if (oa && ns && ns->ns_lvbo && ns->ns_lvbo->lvbo_update) {
573 struct ldlm_resource *rs = NULL;
575 ost_fid_build_resid(&info->fti_fid, &info->fti_resid);
576 rs = ldlm_resource_get(ns, NULL, &info->fti_resid,
579 ns->ns_lvbo->lvbo_update(rs, NULL, 1);
580 ldlm_resource_putref(rs);
583 rc = ofd_commitrw_read(env, ofd, &info->fti_fid, objcount,
592 ofd_info2oti(info, oti);