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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2009, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2014, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/ofd/ofd_objects.c
34 * This file contains OSD API methods related to OBD Filter Device (OFD)
37 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
38 * Author: Mikhail Pershin <mike.pershin@intel.com>
41 #define DEBUG_SUBSYSTEM S_FILTER
43 #include <dt_object.h>
44 #include <lustre/lustre_idl.h>
45 #include <lustre_lfsck.h>
47 #include "ofd_internal.h"
50 * Get object version from disk and check it.
52 * This function checks object version from disk with
53 * ofd_thread_info::fti_pre_version filled from incoming RPC. This is part of
54 * VBR (Version-Based Recovery) and ensures that object has the same version
55 * upon replay as it has during original modification.
57 * \param[in] info execution thread OFD private data
58 * \param[in] fo OFD object
60 * \retval 0 if version matches
61 * \retval -EOVERFLOW on version mismatch
63 static int ofd_version_get_check(struct ofd_thread_info *info,
64 struct ofd_object *fo)
66 dt_obj_version_t curr_version;
68 LASSERT(ofd_object_exists(fo));
70 if (info->fti_exp == NULL)
73 curr_version = dt_version_get(info->fti_env, ofd_object_child(fo));
74 if ((__s64)curr_version == -EOPNOTSUPP)
76 /* VBR: version is checked always because costs nothing */
77 if (info->fti_pre_version != 0 &&
78 info->fti_pre_version != curr_version) {
79 CDEBUG(D_INODE, "Version mismatch %#llx != %#llx\n",
80 info->fti_pre_version, curr_version);
81 spin_lock(&info->fti_exp->exp_lock);
82 info->fti_exp->exp_vbr_failed = 1;
83 spin_unlock(&info->fti_exp->exp_lock);
86 info->fti_pre_version = curr_version;
91 * Get OFD object by FID.
93 * This function finds OFD slice of compound object with the given FID.
95 * \param[in] env execution environment
96 * \param[in] ofd OFD device
97 * \param[in] fid FID of the object
99 * \retval pointer to the found ofd_object
100 * \retval ERR_PTR(errno) in case of error
102 struct ofd_object *ofd_object_find(const struct lu_env *env,
103 struct ofd_device *ofd,
104 const struct lu_fid *fid)
106 struct ofd_object *fo;
111 o = lu_object_find(env, &ofd->ofd_dt_dev.dd_lu_dev, fid, NULL);
112 if (likely(!IS_ERR(o)))
115 fo = ERR_CAST(o); /* return error */
121 * Get FID of parent MDT object.
123 * This function reads extended attribute XATTR_NAME_FID of OFD object which
124 * contains the MDT parent object FID and saves it in ofd_object::ofo_pfid.
126 * The filter_fid::ff_parent::f_ver field currently holds
127 * the OST-object index in the parent MDT-object's layout EA,
128 * not the actual FID::f_ver of the parent. We therefore access
129 * it via the macro f_stripe_idx.
131 * \param[in] env execution environment
132 * \param[in] fo OFD object
134 * \retval 0 if successful
135 * \retval -ENODATA if there is no such xattr
136 * \retval negative value on error
138 int ofd_object_ff_load(const struct lu_env *env, struct ofd_object *fo)
140 struct ofd_thread_info *info = ofd_info(env);
141 struct filter_fid_old *ff = &info->fti_mds_fid_old;
142 struct lu_buf *buf = &info->fti_buf;
143 struct lu_fid *pfid = &fo->ofo_pfid;
146 if (fid_is_sane(pfid))
150 buf->lb_len = sizeof(*ff);
151 rc = dt_xattr_get(env, ofd_object_child(fo), buf, XATTR_NAME_FID);
155 if (rc < sizeof(struct lu_fid)) {
161 pfid->f_seq = le64_to_cpu(ff->ff_parent.f_seq);
162 pfid->f_oid = le32_to_cpu(ff->ff_parent.f_oid);
163 pfid->f_stripe_idx = le32_to_cpu(ff->ff_parent.f_stripe_idx);
169 * Precreate the given number \a nr of objects in the given sequence \a oseq.
171 * This function precreates new OST objects in the given sequence.
172 * The precreation starts from \a id and creates \a nr objects sequentially.
175 * This function may create fewer objects than requested.
177 * We mark object SUID+SGID to flag it for accepting UID+GID from client on
178 * first write. Currently the permission bits on the OST are never used,
181 * Initialize a/c/m time so any client timestamp will always be newer and
182 * update the inode. The ctime = 0 case is also handled specially in
183 * osd_inode_setattr(). See LU-221, LU-1042 for details.
185 * \param[in] env execution environment
186 * \param[in] ofd OFD device
187 * \param[in] id object ID to start precreation from
188 * \param[in] oseq object sequence
189 * \param[in] nr number of objects to precreate
190 * \param[in] sync synchronous precreation flag
192 * \retval 0 if successful
193 * \retval negative value on error
195 int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd,
196 u64 id, struct ofd_seq *oseq, int nr, int sync)
198 struct ofd_thread_info *info = ofd_info(env);
199 struct ofd_object *fo = NULL;
200 struct dt_object *next;
202 struct ofd_object **batch;
203 struct lu_fid *fid = &info->fti_fid;
213 /* Don't create objects beyond the valid range for this SEQ */
214 if (unlikely(fid_seq_is_mdt0(ostid_seq(&oseq->os_oi)) &&
215 (id + nr) >= IDIF_MAX_OID)) {
216 CERROR("%s:"DOSTID" hit the IDIF_MAX_OID (1<<48)!\n",
217 ofd_name(ofd), id, ostid_seq(&oseq->os_oi));
218 RETURN(rc = -ENOSPC);
219 } else if (unlikely(!fid_seq_is_mdt0(ostid_seq(&oseq->os_oi)) &&
220 (id + nr) >= OBIF_MAX_OID)) {
221 CERROR("%s:"DOSTID" hit the OBIF_MAX_OID (1<<32)!\n",
222 ofd_name(ofd), id, ostid_seq(&oseq->os_oi));
223 RETURN(rc = -ENOSPC);
226 OBD_ALLOC(batch, nr_saved * sizeof(struct ofd_object *));
230 info->fti_attr.la_valid = LA_TYPE | LA_MODE;
231 info->fti_attr.la_mode = S_IFREG | S_ISUID | S_ISGID | 0666;
232 info->fti_dof.dof_type = dt_mode_to_dft(S_IFREG);
234 info->fti_attr.la_valid |= LA_ATIME | LA_MTIME | LA_CTIME;
235 info->fti_attr.la_atime = 0;
236 info->fti_attr.la_mtime = 0;
237 info->fti_attr.la_ctime = 0;
241 /* prepare objects */
242 *fid = *lu_object_fid(&oseq->os_lastid_obj->do_lu);
243 for (i = 0; i < nr; i++) {
244 rc = fid_set_id(fid, id + i);
253 fo = ofd_object_find(env, ofd, fid);
256 GOTO(out, rc = PTR_ERR(fo));
262 ofd_write_lock(env, fo);
265 info->fti_buf.lb_buf = &tmp;
266 info->fti_buf.lb_len = sizeof(tmp);
269 th = ofd_trans_create(env, ofd);
271 GOTO(out, rc = PTR_ERR(th));
275 rc = dt_declare_record_write(env, oseq->os_lastid_obj, &info->fti_buf,
278 GOTO(trans_stop, rc);
280 for (i = 0; i < nr; i++) {
284 if (unlikely(ofd_object_exists(fo))) {
285 /* object may exist being re-created by write replay */
286 CDEBUG(D_INODE, "object %#llx/%#llx exists: "
287 DFID"\n", ostid_seq(&oseq->os_oi), id,
288 PFID(lu_object_fid(&fo->ofo_obj.do_lu)));
292 next = ofd_object_child(fo);
293 LASSERT(next != NULL);
295 rc = dt_declare_create(env, next, &info->fti_attr, NULL,
299 GOTO(trans_stop, rc);
306 rc = dt_trans_start_local(env, ofd->ofd_osd, th);
308 GOTO(trans_stop, rc);
310 CDEBUG(D_OTHER, "%s: create new object "DFID" nr %d\n",
311 ofd_name(ofd), PFID(fid), nr);
313 /* When the LFSCK scanning the whole device to verify the LAST_ID file
314 * consistency, it will load the last_id into RAM firstly, and compare
315 * the last_id with each OST-object's ID. If the later one is larger,
316 * then it will regard the LAST_ID file crashed. But during the LFSCK
317 * scanning, the OFD may continue to create new OST-objects. Those new
318 * created OST-objects will have larger IDs than the LFSCK known ones.
319 * So from the LFSCK view, it needs to re-load the last_id from disk
320 * file, and if the latest last_id is still smaller than the object's
321 * ID, then the LAST_ID file is real crashed.
323 * To make above mechanism to work, before OFD pre-create OST-objects,
324 * it needs to update the LAST_ID file firstly, otherwise, the LFSCK
325 * may cannot get latest last_id although new OST-object created. */
326 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_SKIP_LASTID)) {
327 tmp = cpu_to_le64(id + nr - 1);
328 dt_write_lock(env, oseq->os_lastid_obj, 0);
329 rc = dt_record_write(env, oseq->os_lastid_obj,
330 &info->fti_buf, &info->fti_off, th);
331 dt_write_unlock(env, oseq->os_lastid_obj);
333 GOTO(trans_stop, rc);
336 for (i = 0; i < nr; i++) {
340 /* Only the new created objects need to be recorded. */
341 if (ofd->ofd_osd->dd_record_fid_accessed) {
342 struct lfsck_request *lr = &ofd_info(env)->fti_lr;
344 lfsck_pack_rfa(lr, lu_object_fid(&fo->ofo_obj.do_lu),
347 lfsck_in_notify(env, ofd->ofd_osd, lr, NULL);
350 if (likely(!ofd_object_exists(fo) &&
351 !OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DANGLING))) {
352 next = ofd_object_child(fo);
353 LASSERT(next != NULL);
355 rc = dt_create(env, next, &info->fti_attr, NULL,
359 GOTO(trans_stop, rc);
364 LASSERT(ofd_object_exists(fo));
366 ofd_seq_last_oid_set(oseq, id + i);
370 /* NOT all the wanted objects have been created,
371 * set the LAST_ID as the real created. */
372 if (unlikely(objects < nr)) {
376 tmp = cpu_to_le64(ofd_seq_last_oid(oseq));
377 dt_write_lock(env, oseq->os_lastid_obj, 0);
378 rc1 = dt_record_write(env, oseq->os_lastid_obj,
379 &info->fti_buf, &info->fti_off, th);
380 dt_write_unlock(env, oseq->os_lastid_obj);
382 CERROR("%s: fail to reset the LAST_ID for seq (%#llx"
383 ") from %llu to %llu\n", ofd_name(ofd),
384 ostid_seq(&oseq->os_oi), id + nr - 1,
385 ofd_seq_last_oid(oseq));
389 rc2 = ofd_trans_stop(env, ofd, th, rc);
391 CERROR("%s: failed to stop transaction: rc = %d\n",
396 for (i = 0; i < nr_saved; i++) {
399 ofd_write_unlock(env, fo);
400 ofd_object_put(env, fo);
403 OBD_FREE(batch, nr_saved * sizeof(struct ofd_object *));
405 CDEBUG((objects == 0 && rc == 0) ? D_ERROR : D_OTHER,
406 "created %d/%d objects: %d\n", objects, nr_saved, rc);
408 LASSERT(ergo(objects == 0, rc < 0));
409 RETURN(objects > 0 ? objects : rc);
413 * Fix the OFD object ownership.
415 * If the object still has SUID+SGID bits set, meaning that it was precreated
416 * by the MDT before it was assigned to any file, (see ofd_precreate_objects())
417 * then we will accept the UID+GID if sent by the client for initializing the
418 * ownership of this object. We only allow this to happen once (so clear these
419 * bits) and later only allow setattr.
421 * \param[in] env execution environment
422 * \param[in] fo OFD object
423 * \param[in] la object attributes
424 * \param[in] is_setattr was this function called from setattr or not
426 * \retval 0 if successful
427 * \retval negative value on error
429 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
430 struct lu_attr *la, int is_setattr)
432 struct ofd_thread_info *info = ofd_info(env);
433 struct lu_attr *ln = &info->fti_attr2;
439 if (!(la->la_valid & LA_UID) && !(la->la_valid & LA_GID))
442 rc = dt_attr_get(env, ofd_object_child(fo), ln);
446 LASSERT(ln->la_valid & LA_MODE);
449 if (!(ln->la_mode & S_ISUID))
450 la->la_valid &= ~LA_UID;
451 if (!(ln->la_mode & S_ISGID))
452 la->la_valid &= ~LA_GID;
455 if ((la->la_valid & LA_UID) && (ln->la_mode & S_ISUID))
457 if ((la->la_valid & LA_GID) && (ln->la_mode & S_ISGID))
460 if (!(la->la_valid & LA_MODE) || !is_setattr) {
461 la->la_mode = ln->la_mode;
462 la->la_valid |= LA_MODE;
464 la->la_mode &= ~mask;
471 * Set OFD object attributes.
473 * This function sets OFD object attributes taken from incoming request.
474 * It sets not only regular attributes but also XATTR_NAME_FID extended
475 * attribute if needed. The "fid" xattr allows the object's MDT parent inode
476 * to be found and verified by LFSCK and other tools in case of inconsistency.
478 * \param[in] env execution environment
479 * \param[in] fo OFD object
480 * \param[in] la object attributes
481 * \param[in] ff filter_fid structure, contains additional attributes
483 * \retval 0 if successful
484 * \retval negative value on error
486 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
487 struct lu_attr *la, struct filter_fid *ff)
489 struct ofd_thread_info *info = ofd_info(env);
490 struct ofd_device *ofd = ofd_obj2dev(fo);
492 struct ofd_mod_data *fmd;
498 ofd_write_lock(env, fo);
499 if (!ofd_object_exists(fo))
500 GOTO(unlock, rc = -ENOENT);
502 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
503 fmd = ofd_fmd_get(info->fti_exp, &fo->ofo_header.loh_fid);
504 if (fmd && fmd->fmd_mactime_xid < info->fti_xid)
505 fmd->fmd_mactime_xid = info->fti_xid;
506 ofd_fmd_put(info->fti_exp, fmd);
509 /* VBR: version recovery check */
510 rc = ofd_version_get_check(info, fo);
514 rc = ofd_attr_handle_ugid(env, fo, la, 1 /* is_setattr */);
519 rc = ofd_object_ff_load(env, fo);
526 th = ofd_trans_create(env, ofd);
528 GOTO(unlock, rc = PTR_ERR(th));
530 rc = dt_declare_attr_set(env, ofd_object_child(fo), la, th);
535 info->fti_buf.lb_buf = ff;
536 info->fti_buf.lb_len = sizeof(*ff);
537 rc = dt_declare_xattr_set(env, ofd_object_child(fo),
538 &info->fti_buf, XATTR_NAME_FID, 0,
544 rc = ofd_trans_start(env, ofd, la->la_valid & LA_SIZE ? fo : NULL, th);
548 rc = dt_attr_set(env, ofd_object_child(fo), la, th);
553 rc = dt_xattr_set(env, ofd_object_child(fo), &info->fti_buf,
554 XATTR_NAME_FID, 0, th);
556 fo->ofo_pfid.f_seq = le64_to_cpu(ff->ff_parent.f_seq);
557 fo->ofo_pfid.f_oid = le32_to_cpu(ff->ff_parent.f_oid);
558 /* Currently, the filter_fid::ff_parent::f_ver is not
559 * the real parent MDT-object's FID::f_ver, instead it
560 * is the OST-object index in its parent MDT-object's
562 fo->ofo_pfid.f_stripe_idx =
563 le32_to_cpu(ff->ff_parent.f_stripe_idx);
570 rc2 = ofd_trans_stop(env, ofd, th, rc);
572 CERROR("%s: failed to stop transaction: rc = %d\n",
578 ofd_write_unlock(env, fo);
584 * Truncate/punch OFD object.
586 * This function frees all of the allocated object's space from the \a start
587 * offset to the \a end offset. For truncate() operations the \a end offset
588 * is OBD_OBJECT_EOF. The functionality to punch holes in an object via
589 * fallocate(FALLOC_FL_PUNCH_HOLE) is not yet implemented (see LU-3606).
591 * \param[in] env execution environment
592 * \param[in] fo OFD object
593 * \param[in] start start offset to punch from
594 * \param[in] end end of punch
595 * \param[in] la object attributes
596 * \param[in] ff filter_fid structure
597 * \param[in] oa obdo struct from incoming request
599 * \retval 0 if successful
600 * \retval negative value on error
602 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
603 __u64 start, __u64 end, struct lu_attr *la,
604 struct filter_fid *ff, struct obdo *oa)
606 struct ofd_thread_info *info = ofd_info(env);
607 struct ofd_device *ofd = ofd_obj2dev(fo);
608 struct ofd_mod_data *fmd;
609 struct dt_object *dob = ofd_object_child(fo);
617 /* we support truncate, not punch yet */
618 LASSERT(end == OBD_OBJECT_EOF);
620 ofd_write_lock(env, fo);
621 fmd = ofd_fmd_get(info->fti_exp, &fo->ofo_header.loh_fid);
622 if (fmd && fmd->fmd_mactime_xid < info->fti_xid)
623 fmd->fmd_mactime_xid = info->fti_xid;
624 ofd_fmd_put(info->fti_exp, fmd);
626 if (!ofd_object_exists(fo))
627 GOTO(unlock, rc = -ENOENT);
629 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
630 rc = ofd_verify_ff(env, fo, oa);
635 /* VBR: version recovery check */
636 rc = ofd_version_get_check(info, fo);
640 rc = ofd_attr_handle_ugid(env, fo, la, 0 /* !is_setattr */);
645 rc = ofd_object_ff_load(env, fo);
652 th = ofd_trans_create(env, ofd);
654 GOTO(unlock, rc = PTR_ERR(th));
656 rc = dt_declare_attr_set(env, dob, la, th);
660 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
665 info->fti_buf.lb_buf = ff;
666 info->fti_buf.lb_len = sizeof(*ff);
667 rc = dt_declare_xattr_set(env, ofd_object_child(fo),
668 &info->fti_buf, XATTR_NAME_FID, 0,
674 rc = ofd_trans_start(env, ofd, fo, th);
678 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
682 rc = dt_attr_set(env, dob, la, th);
687 rc = dt_xattr_set(env, ofd_object_child(fo), &info->fti_buf,
688 XATTR_NAME_FID, 0, th);
690 fo->ofo_pfid.f_seq = le64_to_cpu(ff->ff_parent.f_seq);
691 fo->ofo_pfid.f_oid = le32_to_cpu(ff->ff_parent.f_oid);
692 /* Currently, the filter_fid::ff_parent::f_ver is not
693 * the real parent MDT-object's FID::f_ver, instead it
694 * is the OST-object index in its parent MDT-object's
696 fo->ofo_pfid.f_stripe_idx =
697 le32_to_cpu(ff->ff_parent.f_stripe_idx);
704 rc2 = ofd_trans_stop(env, ofd, th, rc);
706 CERROR("%s: failed to stop transaction: rc = %d\n",
711 ofd_write_unlock(env, fo);
717 * Destroy OFD object.
719 * This function destroys OFD object. If object wasn't used at all (orphan)
720 * then local transaction is used, which means the transaction data is not
721 * returned back in reply.
723 * \param[in] env execution environment
724 * \param[in] fo OFD object
725 * \param[in] orphan flag to indicate that object is orphaned
727 * \retval 0 if successful
728 * \retval negative value on error
730 int ofd_object_destroy(const struct lu_env *env, struct ofd_object *fo,
733 struct ofd_device *ofd = ofd_obj2dev(fo);
740 ofd_write_lock(env, fo);
741 if (!ofd_object_exists(fo))
742 GOTO(unlock, rc = -ENOENT);
744 th = ofd_trans_create(env, ofd);
746 GOTO(unlock, rc = PTR_ERR(th));
748 rc = dt_declare_ref_del(env, ofd_object_child(fo), th);
752 rc = dt_declare_destroy(env, ofd_object_child(fo), th);
757 rc = dt_trans_start_local(env, ofd->ofd_osd, th);
759 rc = ofd_trans_start(env, ofd, NULL, th);
763 ofd_fmd_drop(ofd_info(env)->fti_exp, &fo->ofo_header.loh_fid);
765 dt_ref_del(env, ofd_object_child(fo), th);
766 dt_destroy(env, ofd_object_child(fo), th);
768 rc2 = ofd_trans_stop(env, ofd, th, rc);
770 CERROR("%s failed to stop transaction: %d\n",
775 ofd_write_unlock(env, fo);
780 * Get OFD object attributes.
782 * This function gets OFD object regular attributes. It is used to serve
783 * incoming request as well as for local OFD purposes.
785 * \param[in] env execution environment
786 * \param[in] fo OFD object
787 * \param[in] la object attributes
789 * \retval 0 if successful
790 * \retval negative value on error
792 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
799 if (ofd_object_exists(fo)) {
800 rc = dt_attr_get(env, ofd_object_child(fo), la);