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_ff.
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 *ff = &fo->ofo_ff;
142 struct lu_buf *buf = &info->fti_buf;
145 if (fid_is_sane(&ff->ff_parent))
149 buf->lb_len = sizeof(*ff);
150 rc = dt_xattr_get(env, ofd_object_child(fo), buf, XATTR_NAME_FID);
154 if (unlikely(rc < sizeof(struct lu_fid))) {
155 fid_zero(&ff->ff_parent);
160 filter_fid_le_to_cpu(ff, ff, rc);
166 * Precreate the given number \a nr of objects in the given sequence \a oseq.
168 * This function precreates new OST objects in the given sequence.
169 * The precreation starts from \a id and creates \a nr objects sequentially.
172 * This function may create fewer objects than requested.
174 * We mark object SUID+SGID to flag it for accepting UID+GID from client on
175 * first write. Currently the permission bits on the OST are never used,
178 * Initialize a/c/m time so any client timestamp will always be newer and
179 * update the inode. The ctime = 0 case is also handled specially in
180 * osd_inode_setattr(). See LU-221, LU-1042 for details.
182 * \param[in] env execution environment
183 * \param[in] ofd OFD device
184 * \param[in] id object ID to start precreation from
185 * \param[in] oseq object sequence
186 * \param[in] nr number of objects to precreate
187 * \param[in] sync synchronous precreation flag
189 * \retval 0 if successful
190 * \retval negative value on error
192 int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd,
193 u64 id, struct ofd_seq *oseq, int nr, int sync)
195 struct ofd_thread_info *info = ofd_info(env);
196 struct ofd_object *fo = NULL;
197 struct dt_object *next;
199 struct ofd_object **batch;
200 struct lu_fid *fid = &info->fti_fid;
210 /* Don't create objects beyond the valid range for this SEQ */
211 if (unlikely(fid_seq_is_mdt0(ostid_seq(&oseq->os_oi)) &&
212 (id + nr) >= IDIF_MAX_OID)) {
213 CERROR("%s:"DOSTID" hit the IDIF_MAX_OID (1<<48)!\n",
214 ofd_name(ofd), id, ostid_seq(&oseq->os_oi));
215 RETURN(rc = -ENOSPC);
216 } else if (unlikely(!fid_seq_is_mdt0(ostid_seq(&oseq->os_oi)) &&
217 (id + nr) >= OBIF_MAX_OID)) {
218 CERROR("%s:"DOSTID" hit the OBIF_MAX_OID (1<<32)!\n",
219 ofd_name(ofd), id, ostid_seq(&oseq->os_oi));
220 RETURN(rc = -ENOSPC);
223 OBD_ALLOC(batch, nr_saved * sizeof(struct ofd_object *));
227 info->fti_attr.la_valid = LA_TYPE | LA_MODE;
228 info->fti_attr.la_mode = S_IFREG | S_ISUID | S_ISGID | 0666;
229 info->fti_dof.dof_type = dt_mode_to_dft(S_IFREG);
231 info->fti_attr.la_valid |= LA_ATIME | LA_MTIME | LA_CTIME;
232 info->fti_attr.la_atime = 0;
233 info->fti_attr.la_mtime = 0;
234 info->fti_attr.la_ctime = 0;
238 /* prepare objects */
239 *fid = *lu_object_fid(&oseq->os_lastid_obj->do_lu);
240 for (i = 0; i < nr; i++) {
241 rc = fid_set_id(fid, id + i);
250 fo = ofd_object_find(env, ofd, fid);
253 GOTO(out, rc = PTR_ERR(fo));
259 ofd_write_lock(env, fo);
262 info->fti_buf.lb_buf = &tmp;
263 info->fti_buf.lb_len = sizeof(tmp);
266 th = ofd_trans_create(env, ofd);
268 GOTO(out, rc = PTR_ERR(th));
272 rc = dt_declare_record_write(env, oseq->os_lastid_obj, &info->fti_buf,
275 GOTO(trans_stop, rc);
277 for (i = 0; i < nr; i++) {
281 if (unlikely(ofd_object_exists(fo))) {
282 /* object may exist being re-created by write replay */
283 CDEBUG(D_INODE, "object %#llx/%#llx exists: "
284 DFID"\n", ostid_seq(&oseq->os_oi), id,
285 PFID(lu_object_fid(&fo->ofo_obj.do_lu)));
289 next = ofd_object_child(fo);
290 LASSERT(next != NULL);
292 rc = dt_declare_create(env, next, &info->fti_attr, NULL,
296 GOTO(trans_stop, rc);
303 rc = dt_trans_start_local(env, ofd->ofd_osd, th);
305 GOTO(trans_stop, rc);
307 CDEBUG(D_OTHER, "%s: create new object "DFID" nr %d\n",
308 ofd_name(ofd), PFID(fid), nr);
310 /* When the LFSCK scanning the whole device to verify the LAST_ID file
311 * consistency, it will load the last_id into RAM firstly, and compare
312 * the last_id with each OST-object's ID. If the later one is larger,
313 * then it will regard the LAST_ID file crashed. But during the LFSCK
314 * scanning, the OFD may continue to create new OST-objects. Those new
315 * created OST-objects will have larger IDs than the LFSCK known ones.
316 * So from the LFSCK view, it needs to re-load the last_id from disk
317 * file, and if the latest last_id is still smaller than the object's
318 * ID, then the LAST_ID file is real crashed.
320 * To make above mechanism to work, before OFD pre-create OST-objects,
321 * it needs to update the LAST_ID file firstly, otherwise, the LFSCK
322 * may cannot get latest last_id although new OST-object created. */
323 if (!OBD_FAIL_CHECK(OBD_FAIL_LFSCK_SKIP_LASTID)) {
324 tmp = cpu_to_le64(id + nr - 1);
325 dt_write_lock(env, oseq->os_lastid_obj, 0);
326 rc = dt_record_write(env, oseq->os_lastid_obj,
327 &info->fti_buf, &info->fti_off, th);
328 dt_write_unlock(env, oseq->os_lastid_obj);
330 GOTO(trans_stop, rc);
333 for (i = 0; i < nr; i++) {
337 /* Only the new created objects need to be recorded. */
338 if (ofd->ofd_osd->dd_record_fid_accessed) {
339 struct lfsck_req_local *lrl = &ofd_info(env)->fti_lrl;
341 lfsck_pack_rfa(lrl, lu_object_fid(&fo->ofo_obj.do_lu),
342 LEL_FID_ACCESSED, LFSCK_TYPE_LAYOUT);
343 lfsck_in_notify_local(env, ofd->ofd_osd, lrl, NULL);
346 if (likely(!ofd_object_exists(fo) &&
347 !OBD_FAIL_CHECK(OBD_FAIL_LFSCK_DANGLING))) {
348 next = ofd_object_child(fo);
349 LASSERT(next != NULL);
351 rc = dt_create(env, next, &info->fti_attr, NULL,
355 GOTO(trans_stop, rc);
360 LASSERT(ofd_object_exists(fo));
362 ofd_seq_last_oid_set(oseq, id + i);
366 /* NOT all the wanted objects have been created,
367 * set the LAST_ID as the real created. */
368 if (unlikely(objects < nr)) {
372 tmp = cpu_to_le64(ofd_seq_last_oid(oseq));
373 dt_write_lock(env, oseq->os_lastid_obj, 0);
374 rc1 = dt_record_write(env, oseq->os_lastid_obj,
375 &info->fti_buf, &info->fti_off, th);
376 dt_write_unlock(env, oseq->os_lastid_obj);
378 CERROR("%s: fail to reset the LAST_ID for seq (%#llx"
379 ") from %llu to %llu\n", ofd_name(ofd),
380 ostid_seq(&oseq->os_oi), id + nr - 1,
381 ofd_seq_last_oid(oseq));
385 rc2 = ofd_trans_stop(env, ofd, th, rc);
387 CERROR("%s: failed to stop transaction: rc = %d\n",
392 for (i = 0; i < nr_saved; i++) {
395 ofd_write_unlock(env, fo);
396 ofd_object_put(env, fo);
399 OBD_FREE(batch, nr_saved * sizeof(struct ofd_object *));
401 CDEBUG((objects == 0 && rc == 0) ? D_ERROR : D_OTHER,
402 "created %d/%d objects: %d\n", objects, nr_saved, rc);
404 LASSERT(ergo(objects == 0, rc < 0));
405 RETURN(objects > 0 ? objects : rc);
409 * Fix the OFD object ownership.
411 * If the object still has SUID+SGID bits set, meaning that it was precreated
412 * by the MDT before it was assigned to any file, (see ofd_precreate_objects())
413 * then we will accept the UID/GID/PROJID if sent by the client for initializing
414 * the ownership of this object. We only allow this to happen once (so clear
415 * these bits) and later only allow setattr.
417 * \param[in] env execution environment
418 * \param[in] fo OFD object
419 * \param[in] la object attributes
420 * \param[in] is_setattr was this function called from setattr or not
422 * \retval 0 if successful
423 * \retval negative value on error
425 int ofd_attr_handle_id(const struct lu_env *env, struct ofd_object *fo,
426 struct lu_attr *la, int is_setattr)
428 struct ofd_thread_info *info = ofd_info(env);
429 struct lu_attr *ln = &info->fti_attr2;
435 if (!(la->la_valid & LA_UID) && !(la->la_valid & LA_GID) &&
436 !(la->la_valid & LA_PROJID))
439 rc = dt_attr_get(env, ofd_object_child(fo), ln);
443 LASSERT(ln->la_valid & LA_MODE);
446 * Only allow setattr to change UID/GID/PROJID, if
447 * SUID+SGID is not set which means this is not
448 * initialization of this objects.
451 if (!(ln->la_mode & S_ISUID))
452 la->la_valid &= ~(LA_UID | LA_PROJID);
453 if (!(ln->la_mode & S_ISGID))
454 la->la_valid &= ~(LA_GID | LA_PROJID);
457 /* Initialize ownership of this object, clear SUID+SGID bits*/
458 if ((la->la_valid & LA_UID) && (ln->la_mode & S_ISUID))
460 if ((la->la_valid & LA_GID) && (ln->la_mode & S_ISGID))
463 if (!(la->la_valid & LA_MODE) || !is_setattr) {
464 la->la_mode = ln->la_mode;
465 la->la_valid |= LA_MODE;
467 la->la_mode &= ~mask;
474 * Set OFD object attributes.
476 * This function sets OFD object attributes taken from incoming request.
477 * It sets not only regular attributes but also XATTR_NAME_FID extended
478 * attribute if needed. The "fid" xattr allows the object's MDT parent inode
479 * to be found and verified by LFSCK and other tools in case of inconsistency.
481 * \param[in] env execution environment
482 * \param[in] fo OFD object
483 * \param[in] la object attributes
484 * \param[in] ff filter_fid structure, contains additional attributes
486 * \retval 0 if successful
487 * \retval negative value on error
489 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
490 struct lu_attr *la, struct filter_fid *ff)
492 struct ofd_thread_info *info = ofd_info(env);
493 struct ofd_device *ofd = ofd_obj2dev(fo);
495 struct ofd_mod_data *fmd;
501 ofd_write_lock(env, fo);
502 if (!ofd_object_exists(fo))
503 GOTO(unlock, rc = -ENOENT);
505 if (la->la_valid & (LA_ATIME | LA_MTIME | LA_CTIME)) {
506 fmd = ofd_fmd_get(info->fti_exp, &fo->ofo_header.loh_fid);
507 if (fmd && fmd->fmd_mactime_xid < info->fti_xid)
508 fmd->fmd_mactime_xid = info->fti_xid;
509 ofd_fmd_put(info->fti_exp, fmd);
512 /* VBR: version recovery check */
513 rc = ofd_version_get_check(info, fo);
517 rc = ofd_attr_handle_id(env, fo, la, 1 /* is_setattr */);
522 rc = ofd_object_ff_load(env, fo);
529 th = ofd_trans_create(env, ofd);
531 GOTO(unlock, rc = PTR_ERR(th));
533 rc = dt_declare_attr_set(env, ofd_object_child(fo), la, th);
538 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
539 ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
540 else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
541 le32_add_cpu(&ff->ff_parent.f_oid, -1);
543 info->fti_buf.lb_buf = ff;
544 info->fti_buf.lb_len = sizeof(*ff);
545 rc = dt_declare_xattr_set(env, ofd_object_child(fo),
546 &info->fti_buf, XATTR_NAME_FID, 0,
552 rc = ofd_trans_start(env, ofd, la->la_valid & LA_SIZE ? fo : NULL, th);
556 rc = dt_attr_set(env, ofd_object_child(fo), la, th);
561 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
564 rc = dt_xattr_set(env, ofd_object_child(fo), &info->fti_buf,
565 XATTR_NAME_FID, 0, th);
567 filter_fid_le_to_cpu(&fo->ofo_ff, ff, sizeof(*ff));
573 rc2 = ofd_trans_stop(env, ofd, th, rc);
575 CERROR("%s: failed to stop transaction: rc = %d\n",
581 ofd_write_unlock(env, fo);
587 * Truncate/punch OFD object.
589 * This function frees all of the allocated object's space from the \a start
590 * offset to the \a end offset. For truncate() operations the \a end offset
591 * is OBD_OBJECT_EOF. The functionality to punch holes in an object via
592 * fallocate(FALLOC_FL_PUNCH_HOLE) is not yet implemented (see LU-3606).
594 * \param[in] env execution environment
595 * \param[in] fo OFD object
596 * \param[in] start start offset to punch from
597 * \param[in] end end of punch
598 * \param[in] la object attributes
599 * \param[in] ff filter_fid structure
600 * \param[in] oa obdo struct from incoming request
602 * \retval 0 if successful
603 * \retval negative value on error
605 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
606 __u64 start, __u64 end, struct lu_attr *la,
607 struct filter_fid *ff, struct obdo *oa)
609 struct ofd_thread_info *info = ofd_info(env);
610 struct ofd_device *ofd = ofd_obj2dev(fo);
611 struct ofd_mod_data *fmd;
612 struct dt_object *dob = ofd_object_child(fo);
620 /* we support truncate, not punch yet */
621 LASSERT(end == OBD_OBJECT_EOF);
623 ofd_write_lock(env, fo);
624 fmd = ofd_fmd_get(info->fti_exp, &fo->ofo_header.loh_fid);
625 if (fmd && fmd->fmd_mactime_xid < info->fti_xid)
626 fmd->fmd_mactime_xid = info->fti_xid;
627 ofd_fmd_put(info->fti_exp, fmd);
629 if (!ofd_object_exists(fo))
630 GOTO(unlock, rc = -ENOENT);
632 if (ofd->ofd_lfsck_verify_pfid && oa->o_valid & OBD_MD_FLFID) {
633 rc = ofd_verify_ff(env, fo, oa);
638 /* VBR: version recovery check */
639 rc = ofd_version_get_check(info, fo);
643 rc = ofd_attr_handle_id(env, fo, la, 0 /* !is_setattr */);
648 rc = ofd_object_ff_load(env, fo);
655 th = ofd_trans_create(env, ofd);
657 GOTO(unlock, rc = PTR_ERR(th));
659 rc = dt_declare_attr_set(env, dob, la, th);
663 rc = dt_declare_punch(env, dob, start, OBD_OBJECT_EOF, th);
668 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR1))
669 ff->ff_parent.f_oid = cpu_to_le32(1UL << 31);
670 else if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_UNMATCHED_PAIR2))
671 le32_add_cpu(&ff->ff_parent.f_oid, -1);
673 info->fti_buf.lb_buf = ff;
674 info->fti_buf.lb_len = sizeof(*ff);
675 rc = dt_declare_xattr_set(env, ofd_object_child(fo),
676 &info->fti_buf, XATTR_NAME_FID, 0,
682 rc = ofd_trans_start(env, ofd, fo, th);
686 rc = dt_punch(env, dob, start, OBD_OBJECT_EOF, th);
690 rc = dt_attr_set(env, dob, la, th);
695 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_NOPFID))
698 rc = dt_xattr_set(env, ofd_object_child(fo), &info->fti_buf,
699 XATTR_NAME_FID, 0, th);
701 filter_fid_le_to_cpu(&fo->ofo_ff, ff, sizeof(*ff));
707 rc2 = ofd_trans_stop(env, ofd, th, rc);
709 CERROR("%s: failed to stop transaction: rc = %d\n",
714 ofd_write_unlock(env, fo);
720 * Destroy OFD object.
722 * This function destroys OFD object. If object wasn't used at all (orphan)
723 * then local transaction is used, which means the transaction data is not
724 * returned back in reply.
726 * \param[in] env execution environment
727 * \param[in] fo OFD object
728 * \param[in] orphan flag to indicate that object is orphaned
730 * \retval 0 if successful
731 * \retval negative value on error
733 int ofd_destroy(const struct lu_env *env, struct ofd_object *fo,
736 struct ofd_device *ofd = ofd_obj2dev(fo);
743 ofd_write_lock(env, fo);
744 if (!ofd_object_exists(fo))
745 GOTO(unlock, rc = -ENOENT);
747 th = ofd_trans_create(env, ofd);
749 GOTO(unlock, rc = PTR_ERR(th));
751 rc = dt_declare_ref_del(env, ofd_object_child(fo), th);
755 rc = dt_declare_destroy(env, ofd_object_child(fo), th);
760 rc = dt_trans_start_local(env, ofd->ofd_osd, th);
762 rc = ofd_trans_start(env, ofd, NULL, th);
766 ofd_fmd_drop(ofd_info(env)->fti_exp, &fo->ofo_header.loh_fid);
768 dt_ref_del(env, ofd_object_child(fo), th);
769 dt_destroy(env, ofd_object_child(fo), th);
771 rc2 = ofd_trans_stop(env, ofd, th, rc);
773 CERROR("%s failed to stop transaction: %d\n",
778 ofd_write_unlock(env, fo);
783 * Get OFD object attributes.
785 * This function gets OFD object regular attributes. It is used to serve
786 * incoming request as well as for local OFD purposes.
788 * \param[in] env execution environment
789 * \param[in] fo OFD object
790 * \param[in] la object attributes
792 * \retval 0 if successful
793 * \retval negative value on error
795 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
802 if (ofd_object_exists(fo)) {
803 rc = dt_attr_get(env, ofd_object_child(fo), la);