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,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License version 2 for more details. A copy is
14 * included in the COPYING 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 2009 Sun Microsystems, Inc. All rights reserved
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2017, Intel Corporation.
29 * lustre/lod/lod_lov.c
31 * A set of helpers to maintain Logical Object Volume (LOV)
32 * Extended Attribute (EA) and known OST targets
34 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
37 #define DEBUG_SUBSYSTEM S_MDS
39 #include <obd_class.h>
40 #include <lustre_lfsck.h>
41 #include <lustre_lmv.h>
42 #include <lustre_swab.h>
44 #include "lod_internal.h"
47 * Increase reference count on the target table.
49 * Increase reference count on the target table usage to prevent racing with
50 * addition/deletion. Any function that expects the table to remain
51 * stationary must take a ref.
53 * \param[in] ltd target table (lod_ost_descs or lod_mdt_descs)
55 void lod_getref(struct lod_tgt_descs *ltd)
57 down_read(<d->ltd_rw_sem);
58 mutex_lock(<d->ltd_mutex);
60 mutex_unlock(<d->ltd_mutex);
64 * Decrease reference count on the target table.
66 * Companion of lod_getref() to release a reference on the target table.
67 * If this is the last reference and the OST entry was scheduled for deletion,
68 * the descriptor is removed from the table.
70 * \param[in] lod LOD device from which we release a reference
71 * \param[in] ltd target table (lod_ost_descs or lod_mdt_descs)
73 void lod_putref(struct lod_device *lod, struct lod_tgt_descs *ltd)
75 mutex_lock(<d->ltd_mutex);
77 if (ltd->ltd_refcount == 0 && ltd->ltd_death_row) {
78 struct lod_tgt_desc *tgt_desc, *tmp;
81 CDEBUG(D_CONFIG, "destroying %d ltd desc\n",
84 ltd_foreach_tgt_safe(ltd, tgt_desc, tmp) {
86 if (!tgt_desc->ltd_reap)
89 list_add(&tgt_desc->ltd_kill, &kill);
90 lu_tgt_pool_remove(<d->ltd_tgt_pool,
92 ltd_del_tgt(ltd, tgt_desc);
95 mutex_unlock(<d->ltd_mutex);
96 up_read(<d->ltd_rw_sem);
98 list_for_each_entry_safe(tgt_desc, tmp, &kill, ltd_kill) {
101 list_del(&tgt_desc->ltd_kill);
102 rc = obd_disconnect(tgt_desc->ltd_exp);
104 CERROR("%s: failed to disconnect %s: rc = %d\n",
105 lod2obd(lod)->obd_name,
106 obd_uuid2str(&tgt_desc->ltd_uuid), rc);
107 OBD_FREE_PTR(tgt_desc);
110 mutex_unlock(<d->ltd_mutex);
111 up_read(<d->ltd_rw_sem);
116 * Connect LOD to a new OSP and add it to the target table.
118 * Connect to the OSP device passed, initialize all the internal
119 * structures related to the device and add it to the target table.
121 * \param[in] env execution environment for this thread
122 * \param[in] lod LOD device to be connected to the new OSP
123 * \param[in] osp name of OSP device name to be added
124 * \param[in] index index of the new target
125 * \param[in] gen target's generation number
126 * \param[in] tgt_index OSP's group
127 * \param[in] type type of device (mdc or osc)
128 * \param[in] active state of OSP: 0 - inactive, 1 - active
130 * \retval 0 if added successfully
131 * \retval negative error number on failure
133 int lod_add_device(const struct lu_env *env, struct lod_device *lod,
134 char *osp, unsigned index, unsigned gen, int tgt_index,
135 char *type, int active)
137 struct obd_connect_data *data = NULL;
138 struct obd_export *exp = NULL;
139 struct obd_device *obd;
140 struct lu_device *lu_dev;
141 struct dt_device *dt_dev;
143 struct lod_tgt_desc *tgt_desc;
144 struct lod_tgt_descs *ltd;
145 struct lustre_cfg *lcfg;
146 struct obd_uuid obd_uuid;
148 bool connected = false;
151 CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, index, gen);
154 CERROR("request to add OBD %s with invalid generation: %d\n",
159 obd_str2uuid(&obd_uuid, osp);
161 obd = class_find_client_obd(&obd_uuid, LUSTRE_OSP_NAME,
162 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
164 CERROR("can't find %s device\n", osp);
168 LASSERT(obd->obd_lu_dev != NULL);
169 LASSERT(obd->obd_lu_dev->ld_site == lod->lod_dt_dev.dd_lu_dev.ld_site);
171 lu_dev = obd->obd_lu_dev;
172 dt_dev = lu2dt_dev(lu_dev);
176 GOTO(out_cleanup, rc = -ENOMEM);
178 data->ocd_connect_flags = OBD_CONNECT_INDEX | OBD_CONNECT_VERSION |
180 data->ocd_version = LUSTRE_VERSION_CODE;
181 data->ocd_index = index;
183 if (strcmp(LUSTRE_OSC_NAME, type) == 0) {
185 data->ocd_connect_flags |= OBD_CONNECT_AT |
188 #ifdef HAVE_LRU_RESIZE_SUPPORT
189 OBD_CONNECT_LRU_RESIZE |
192 OBD_CONNECT_REQPORTAL |
193 OBD_CONNECT_SKIP_ORPHAN |
195 OBD_CONNECT_LVB_TYPE |
196 OBD_CONNECT_VERSION |
197 OBD_CONNECT_PINGLESS |
199 OBD_CONNECT_BULK_MBITS;
200 data->ocd_connect_flags2 = OBD_CONNECT2_REPLAY_CREATE;
202 data->ocd_group = tgt_index;
203 ltd = &lod->lod_ost_descs;
205 struct obd_import *imp = obd->u.cli.cl_import;
208 data->ocd_ibits_known = MDS_INODELOCK_UPDATE;
209 data->ocd_connect_flags |= OBD_CONNECT_ACL |
211 OBD_CONNECT_MDS_MDS |
212 OBD_CONNECT_MULTIMODRPCS |
217 OBD_CONNECT_BULK_MBITS;
218 spin_lock(&imp->imp_lock);
219 imp->imp_server_timeout = 1;
220 spin_unlock(&imp->imp_lock);
221 imp->imp_client->cli_request_portal = OUT_PORTAL;
222 CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
224 ltd = &lod->lod_mdt_descs;
227 rc = obd_connect(env, &exp, obd, &obd->obd_uuid, data, NULL);
230 CERROR("%s: cannot connect to next dev %s (%d)\n",
231 obd->obd_name, osp, rc);
232 GOTO(out_cleanup, rc);
236 /* Allocate ost descriptor and fill it */
237 OBD_ALLOC_PTR(tgt_desc);
239 GOTO(out_cleanup, rc = -ENOMEM);
241 tgt_desc->ltd_tgt = dt_dev;
242 tgt_desc->ltd_exp = exp;
243 tgt_desc->ltd_uuid = obd->u.cli.cl_target_uuid;
244 tgt_desc->ltd_gen = gen;
245 tgt_desc->ltd_index = index;
246 tgt_desc->ltd_active = active;
248 down_write(<d->ltd_rw_sem);
249 mutex_lock(<d->ltd_mutex);
250 rc = ltd_add_tgt(ltd, tgt_desc);
254 rc = lu_qos_add_tgt(<d->ltd_qos, tgt_desc);
256 GOTO(out_del_tgt, rc);
258 rc = lu_tgt_pool_add(<d->ltd_tgt_pool, index,
259 ltd->ltd_lov_desc.ld_tgt_count);
261 CERROR("%s: can't set up pool, failed with %d\n",
263 GOTO(out_del_tgt, rc);
266 mutex_unlock(<d->ltd_mutex);
267 up_write(<d->ltd_rw_sem);
269 if (lod->lod_recovery_completed)
270 lu_dev->ld_ops->ldo_recovery_complete(env, lu_dev);
272 if (!for_ost && lod->lod_initialized) {
273 rc = lod_sub_init_llog(env, lod, tgt_desc->ltd_tgt);
275 CERROR("%s: cannot start llog on %s:rc = %d\n",
276 lod2obd(lod)->obd_name, osp, rc);
281 rc = lfsck_add_target(env, lod->lod_child, dt_dev, exp, index, for_ost);
283 CERROR("Fail to add LFSCK target: name = %s, type = %s, "
284 "index = %u, rc = %d\n", osp, type, index, rc);
285 GOTO(out_fini_llog, rc);
289 lod_sub_fini_llog(env, tgt_desc->ltd_tgt,
290 &tgt_desc->ltd_recovery_task);
292 down_write(<d->ltd_rw_sem);
293 mutex_lock(<d->ltd_mutex);
294 lu_tgt_pool_remove(<d->ltd_tgt_pool, index);
296 ltd_del_tgt(ltd, tgt_desc);
298 mutex_unlock(<d->ltd_mutex);
299 up_write(<d->ltd_rw_sem);
300 OBD_FREE_PTR(tgt_desc);
302 /* XXX OSP needs us to send down LCFG_CLEANUP because it uses
303 * objects from the MDT stack. See LU-7184. */
304 lcfg = &lod_env_info(env)->lti_lustre_cfg;
305 memset(lcfg, 0, sizeof(*lcfg));
306 lcfg->lcfg_version = LUSTRE_CFG_VERSION;
307 lcfg->lcfg_command = LCFG_CLEANUP;
308 lu_dev->ld_ops->ldo_process_config(env, lu_dev, lcfg);
317 * Schedule target removal from the target table.
319 * Mark the device as dead. The device is not removed here because it may
320 * still be in use. The device will be removed in lod_putref() when the
321 * last reference is released.
323 * \param[in] env execution environment for this thread
324 * \param[in] lod LOD device the target table belongs to
325 * \param[in] ltd target table
326 * \param[in] tgt target
328 static void __lod_del_device(const struct lu_env *env, struct lod_device *lod,
329 struct lod_tgt_descs *ltd, struct lu_tgt_desc *tgt)
331 lfsck_del_target(env, lod->lod_child, tgt->ltd_tgt, tgt->ltd_index,
334 if (!tgt->ltd_reap) {
336 ltd->ltd_death_row++;
341 * Schedule removal of all the targets from the given target table.
343 * See more details in the description for __lod_del_device()
345 * \param[in] env execution environment for this thread
346 * \param[in] lod LOD device the target table belongs to
347 * \param[in] ltd target table
351 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
352 struct lod_tgt_descs *ltd)
354 struct lu_tgt_desc *tgt;
356 if (ltd->ltd_tgts_size <= 0)
360 mutex_lock(<d->ltd_mutex);
361 ltd_foreach_tgt(ltd, tgt)
362 __lod_del_device(env, lod, ltd, tgt);
363 mutex_unlock(<d->ltd_mutex);
364 lod_putref(lod, ltd);
366 lu_tgt_descs_fini(ltd);
372 * Remove device by name.
374 * Remove a device identified by \a osp from the target table. Given
375 * the device can be in use, the real deletion happens in lod_putref().
377 * \param[in] env execution environment for this thread
378 * \param[in] lod LOD device to be connected to the new OSP
379 * \param[in] ltd target table
380 * \param[in] osp name of OSP device to be removed
381 * \param[in] idx index of the target
382 * \param[in] gen generation number, not used currently
384 * \retval 0 if the device was scheduled for removal
385 * \retval -EINVAL if no device was found
387 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
388 struct lod_tgt_descs *ltd, char *osp, unsigned int idx,
391 struct obd_device *obd;
392 struct lu_tgt_desc *tgt;
393 struct obd_uuid uuid;
398 CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, idx, gen);
400 obd_str2uuid(&uuid, osp);
402 obd = class_find_client_obd(&uuid, LUSTRE_OSP_NAME,
403 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
405 CERROR("can't find %s device\n", osp);
410 CERROR("%s: request to remove OBD %s with invalid generation %d"
411 "\n", obd->obd_name, osp, gen);
415 obd_str2uuid(&uuid, osp);
418 mutex_lock(<d->ltd_mutex);
419 tgt = LTD_TGT(ltd, idx);
420 /* check that the index is allocated in the bitmap */
421 if (!test_bit(idx, ltd->ltd_tgt_bitmap) || !tgt) {
422 CERROR("%s: device %d is not set up\n", obd->obd_name, idx);
423 GOTO(out, rc = -EINVAL);
426 /* check that the UUID matches */
427 if (!obd_uuid_equals(&uuid, &tgt->ltd_uuid)) {
428 CERROR("%s: LOD target UUID %s at index %d does not match %s\n",
429 obd->obd_name, obd_uuid2str(&tgt->ltd_uuid), idx, osp);
430 GOTO(out, rc = -EINVAL);
433 __lod_del_device(env, lod, ltd, tgt);
436 mutex_unlock(<d->ltd_mutex);
437 lod_putref(lod, ltd);
442 * Resize per-thread storage to hold specified size.
444 * A helper function to resize per-thread temporary storage. This storage
445 * is used to process LOV/LVM EAs and may be quite large. We do not want to
446 * allocate/release it every time, so instead we put it into the env and
447 * reallocate on demand. The memory is released when the correspondent thread
450 * \param[in] info LOD-specific storage in the environment
451 * \param[in] size new size to grow the buffer to
453 * \retval 0 on success, -ENOMEM if reallocation failed
455 int lod_ea_store_resize(struct lod_thread_info *info, size_t size)
457 __u32 round = size_roundup_power2(size);
459 if (info->lti_ea_store) {
460 LASSERT(info->lti_ea_store_size);
461 CDEBUG(D_INFO, "EA store size %d is not enough, need %d\n",
462 info->lti_ea_store_size, round);
463 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
464 info->lti_ea_store = NULL;
465 info->lti_ea_store_size = 0;
468 OBD_ALLOC_LARGE(info->lti_ea_store, round);
469 if (info->lti_ea_store == NULL)
471 info->lti_ea_store_size = round;
476 static void lod_free_comp_buffer(struct lod_layout_component *entries,
477 __u16 count, __u32 bufsize)
479 struct lod_layout_component *entry;
482 for (i = 0; i < count; i++) {
484 if (entry->llc_magic == LOV_MAGIC_FOREIGN)
486 if (entry->llc_pool != NULL)
487 lod_set_pool(&entry->llc_pool, NULL);
488 if (entry->llc_ostlist.op_array) {
489 OBD_FREE(entry->llc_ostlist.op_array,
490 entry->llc_ostlist.op_size);
491 entry->llc_ostlist.op_array = NULL;
492 entry->llc_ostlist.op_size = 0;
494 LASSERT(entry->llc_stripe == NULL);
495 LASSERT(entry->llc_stripes_allocated == 0);
499 OBD_FREE_LARGE(entries, bufsize);
502 void lod_free_def_comp_entries(struct lod_default_striping *lds)
504 lod_free_comp_buffer(lds->lds_def_comp_entries,
505 lds->lds_def_comp_size_cnt,
507 sizeof(*lds->lds_def_comp_entries) *
508 lds->lds_def_comp_size_cnt));
509 lds->lds_def_comp_entries = NULL;
510 lds->lds_def_comp_cnt = 0;
511 lds->lds_def_striping_is_composite = 0;
512 lds->lds_def_comp_size_cnt = 0;
516 * Resize per-thread storage to hold default striping component entries
518 * A helper function to resize per-thread temporary storage. This storage
519 * is used to hold default LOV/LVM EAs and may be quite large. We do not want
520 * to allocate/release it every time, so instead we put it into the env and
521 * reallocate it on demand. The memory is released when the correspondent
522 * thread is finished.
524 * \param[in,out] lds default striping
525 * \param[in] count new component count to grow the buffer to
527 * \retval 0 on success, -ENOMEM if reallocation failed
529 int lod_def_striping_comp_resize(struct lod_default_striping *lds, __u16 count)
531 struct lod_layout_component *entries;
532 __u32 new = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
534 __u32 old = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
535 lds->lds_def_comp_size_cnt);
540 OBD_ALLOC_LARGE(entries, new);
544 if (lds->lds_def_comp_entries != NULL) {
545 CDEBUG(D_INFO, "default striping component size %d is not "
546 "enough, need %d\n", old, new);
547 lod_free_def_comp_entries(lds);
550 lds->lds_def_comp_entries = entries;
551 lds->lds_def_comp_size_cnt = count;
556 void lod_free_comp_entries(struct lod_object *lo)
558 if (lo->ldo_mirrors) {
559 OBD_FREE_PTR_ARRAY(lo->ldo_mirrors, lo->ldo_mirror_count);
560 lo->ldo_mirrors = NULL;
561 lo->ldo_mirror_count = 0;
563 lod_free_comp_buffer(lo->ldo_comp_entries,
565 sizeof(*lo->ldo_comp_entries) * lo->ldo_comp_cnt);
566 lo->ldo_comp_entries = NULL;
567 lo->ldo_comp_cnt = 0;
568 lo->ldo_is_composite = 0;
571 int lod_alloc_comp_entries(struct lod_object *lo,
572 int mirror_count, int comp_count)
574 LASSERT(comp_count != 0);
575 LASSERT(lo->ldo_comp_cnt == 0 && lo->ldo_comp_entries == NULL);
577 if (mirror_count > 0) {
578 OBD_ALLOC_PTR_ARRAY(lo->ldo_mirrors, mirror_count);
579 if (!lo->ldo_mirrors)
582 lo->ldo_mirror_count = mirror_count;
585 OBD_ALLOC_LARGE(lo->ldo_comp_entries,
586 sizeof(*lo->ldo_comp_entries) * comp_count);
587 if (lo->ldo_comp_entries == NULL) {
588 OBD_FREE_PTR_ARRAY(lo->ldo_mirrors, mirror_count);
589 lo->ldo_mirrors = NULL;
590 lo->ldo_mirror_count = 0;
594 lo->ldo_comp_cnt = comp_count;
595 lo->ldo_is_foreign = 0;
599 int lod_fill_mirrors(struct lod_object *lo)
601 struct lod_device *lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
602 struct lod_layout_component *lod_comp;
603 bool found_preferred = false;
605 __u16 mirror_id = 0xffff;
609 LASSERT(equi(!lo->ldo_is_composite, lo->ldo_mirror_count == 0));
611 if (!lo->ldo_is_composite)
614 lod_comp = &lo->ldo_comp_entries[0];
616 for (i = 0; i < lo->ldo_comp_cnt; i++, lod_comp++) {
617 bool stale = lod_comp->llc_flags & LCME_FL_STALE;
618 bool preferred = lod_comp->llc_flags & LCME_FL_PREF_WR;
619 bool mirror_hsm = lod_is_hsm(lod_comp);
620 bool init = (lod_comp->llc_stripe != NULL) &&
621 !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
622 !(lod_comp->llc_pattern & LOV_PATTERN_MDT);
626 /* calculate component preference over all used OSTs */
627 for (j = 0; init && j < lod_comp->llc_stripes_allocated; j++) {
628 __u32 idx = lod_comp->llc_ost_indices[j];
629 struct lod_tgt_desc *ltd;
631 if (lod_comp->llc_stripe[j] == NULL)
634 if (unlikely(idx >= lod->lod_ost_descs.ltd_tgts_size)) {
635 CERROR("%s: "DFID" OST idx %u > max %u\n",
636 lod2obd(lod)->obd_name,
637 PFID(lu_object_fid(&lo->ldo_obj.do_lu)),
638 idx, lod->lod_ost_descs.ltd_tgts_size);
641 ltd = OST_TGT(lod, idx);
642 if (unlikely(!ltd)) {
643 CERROR("%s: "DFID" OST idx %u is NULL\n",
644 lod2obd(lod)->obd_name,
645 PFID(lu_object_fid(&lo->ldo_obj.do_lu)),
650 if (ltd->ltd_statfs.os_state & OS_STATFS_NONROT)
654 if (mirror_id_of(lod_comp->llc_id) == mirror_id) {
655 /* Currently HSM mirror does not support PFL. */
656 if (lo->ldo_mirrors[mirror_idx].lme_hsm)
658 lo->ldo_mirrors[mirror_idx].lme_stale |= stale;
659 lo->ldo_mirrors[mirror_idx].lme_prefer |= preferred;
660 lo->ldo_mirrors[mirror_idx].lme_preference += pref;
661 lo->ldo_mirrors[mirror_idx].lme_end = i;
665 if (mirror_idx >= 0 && preferred &&
666 !lo->ldo_mirrors[mirror_idx].lme_stale)
667 found_preferred = true;
671 if (mirror_idx >= lo->ldo_mirror_count)
674 if (mirror_hsm && (lod_comp->llc_extent.e_start != 0 ||
675 lod_comp->llc_extent.e_end != LUSTRE_EOF))
678 mirror_id = mirror_id_of(lod_comp->llc_id);
680 lo->ldo_mirrors[mirror_idx].lme_id = mirror_id;
681 lo->ldo_mirrors[mirror_idx].lme_stale = stale;
682 lo->ldo_mirrors[mirror_idx].lme_prefer = preferred;
683 lo->ldo_mirrors[mirror_idx].lme_hsm = mirror_hsm;
684 lo->ldo_mirrors[mirror_idx].lme_preference = pref;
685 lo->ldo_mirrors[mirror_idx].lme_start = i;
686 lo->ldo_mirrors[mirror_idx].lme_end = i;
688 if (mirror_idx != lo->ldo_mirror_count - 1)
691 if (!found_preferred && mirror_idx > 0) {
695 * if no explicited preferred found, then find a mirror
696 * with higher number of non-rotational OSTs
699 for (i = 0; i <= mirror_idx; i++) {
700 if (lo->ldo_mirrors[i].lme_stale)
702 if (lo->ldo_mirrors[i].lme_preference > pref) {
703 pref = lo->ldo_mirrors[i].lme_preference;
709 lo->ldo_mirrors[best].lme_prefer = 1;
716 * Generate on-disk lov_mds_md structure for each layout component based on
717 * the information in lod_object->ldo_comp_entries[i].
719 * \param[in] env execution environment for this thread
720 * \param[in] lo LOD object
721 * \param[in] comp_idx index of ldo_comp_entries
722 * \param[in] lmm buffer to cotain the on-disk lov_mds_md
723 * \param[in|out] lmm_size buffer size/lmm size
724 * \param[in] is_dir generate lov ea for dir or file? For dir case,
725 * the stripe info is from the default stripe
726 * template, which is collected in lod_ah_init(),
727 * either from parent object or root object; for
728 * file case, it's from the @lo object
730 * \retval 0 if on disk structure is created successfully
731 * \retval negative error number on failure
733 static int lod_gen_component_ea(const struct lu_env *env,
734 struct lod_object *lo, int comp_idx,
735 struct lov_mds_md *lmm, int *lmm_size,
738 struct lod_thread_info *info = lod_env_info(env);
739 const struct lu_fid *fid = lu_object_fid(&lo->ldo_obj.do_lu);
740 struct lod_device *lod;
741 struct lov_ost_data_v1 *objs;
742 struct lod_layout_component *lod_comp;
751 &lo->ldo_def_striping->lds_def_comp_entries[comp_idx];
753 lod_comp = &lo->ldo_comp_entries[comp_idx];
755 magic = lod_comp->llc_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
756 if (is_dir && lod_comp->llc_ostlist.op_count)
757 magic = LOV_MAGIC_SPECIFIC;
759 if (lod_comp->llc_pattern == 0) /* default striping */
760 lod_comp->llc_pattern = LOV_PATTERN_RAID0;
762 lmm->lmm_magic = cpu_to_le32(magic);
763 lmm->lmm_pattern = cpu_to_le32(lod_comp->llc_pattern);
764 fid_to_lmm_oi(fid, &lmm->lmm_oi);
765 if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
766 lmm->lmm_oi.oi.oi_id++;
767 lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
769 lmm->lmm_stripe_size = cpu_to_le32(lod_comp->llc_stripe_size);
770 lmm->lmm_stripe_count = cpu_to_le16(lod_comp->llc_stripe_count);
772 * for dir and uninstantiated component, lmm_layout_gen stores
773 * default stripe offset.
775 lmm->lmm_layout_gen =
776 (is_dir || !lod_comp_inited(lod_comp)) ?
777 cpu_to_le16(lod_comp->llc_stripe_offset) :
778 cpu_to_le16(lod_comp->llc_layout_gen);
780 if (magic == LOV_MAGIC_V1) {
781 objs = &lmm->lmm_objects[0];
783 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
784 ssize_t cplen = strscpy(v3->lmm_pool_name,
785 lod_comp->llc_pool ? : "\0",
786 sizeof(v3->lmm_pool_name));
790 objs = &v3->lmm_objects[0];
792 lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
793 stripe_count = lod_comp_entry_stripe_count(lo, comp_idx, is_dir);
794 if (stripe_count == 0 && !is_dir &&
795 !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
796 !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
797 /* Try again if all active targets are disconnected.
798 * It is possible when MDS does failover. */
799 if (!lod->lod_ost_active_count &&
805 if (!is_dir && lo->ldo_is_composite)
806 lod_comp_shrink_stripe_count(lod_comp, &stripe_count);
808 if ((is_dir && magic != LOV_MAGIC_SPECIFIC) ||
809 lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
812 /* generate ost_idx of this component stripe */
813 for (i = 0; i < stripe_count; i++) {
814 struct dt_object *object;
815 __u32 ost_idx = (__u32)-1UL;
816 int type = LU_SEQ_RANGE_OST;
818 if (lod_comp->llc_stripe && lod_comp->llc_stripe[i]) {
819 object = lod_comp->llc_stripe[i];
820 /* instantiated component */
821 info->lti_fid = *lu_object_fid(&object->do_lu);
823 if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF) &&
825 if (cfs_fail_val == 0)
826 cfs_fail_val = info->lti_fid.f_oid;
828 info->lti_fid.f_oid = cfs_fail_val;
831 rc = fid_to_ostid(&info->lti_fid, &info->lti_ostid);
834 ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
835 objs[i].l_ost_gen = cpu_to_le32(0);
836 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_FLD_LOOKUP))
839 rc = lod_fld_lookup(env, lod, &info->lti_fid,
842 CERROR("%s: Can not locate "DFID": rc = %d\n",
843 lod2obd(lod)->obd_name,
844 PFID(&info->lti_fid), rc);
847 } else if (lod_comp->llc_ostlist.op_array &&
848 lod_comp->llc_ostlist.op_count) {
849 /* user specified ost list */
850 ost_idx = lod_comp->llc_ostlist.op_array[i];
853 * with un-instantiated or with no specified ost list
854 * component, its l_ost_idx does not matter.
856 objs[i].l_ost_idx = cpu_to_le32(ost_idx);
858 /* simulation of broken LOVEA */
859 if (CFS_FAIL_CHECK(OBD_FAIL_LOV_INVALID_OSTIDX) &&
860 comp_idx == 0 && i == 0 && lo->ldo_mirror_count > 1) {
861 objs[i].l_ost_idx = cpu_to_le32(0xffffffff);
866 if (lmm_size != NULL)
867 *lmm_size = lov_mds_md_size(stripe_count, magic);
872 * Generate on-disk lov_hsm_md structure based on the information in
873 * the lod_object->ldo_comp_entries.
875 static int lod_gen_component_ea_foreign(const struct lu_env *env,
876 struct lod_object *lo,
877 struct lod_layout_component *lod_comp,
878 void *lmm, int *lmm_size)
880 struct lov_foreign_md *lfm = (struct lov_foreign_md *)lmm;
884 lfm->lfm_magic = cpu_to_le32(LOV_MAGIC_FOREIGN);
885 lfm->lfm_length = cpu_to_le32(lod_comp->llc_length);
886 lfm->lfm_type = cpu_to_le32(lod_comp->llc_type);
887 lfm->lfm_flags = cpu_to_le32(lod_comp->llc_foreign_flags);
889 if (lov_hsm_type_supported(lod_comp->llc_type)) {
890 if (lod_comp->llc_length != sizeof(struct lov_hsm_base))
893 lov_foreign_hsm_to_le(lfm, &lod_comp->llc_hsm);
897 *lmm_size = lov_foreign_md_size(lod_comp->llc_length);
903 * Generate on-disk lov_mds_md structure based on the information in
904 * the lod_object->ldo_comp_entries.
906 * \param[in] env execution environment for this thread
907 * \param[in] lo LOD object
908 * \param[in] lmm buffer to cotain the on-disk lov_mds_md
909 * \param[in|out] lmm_size buffer size/lmm size
910 * \param[in] is_dir generate lov ea for dir or file? For dir case,
911 * the stripe info is from the default stripe
912 * template, which is collected in lod_ah_init(),
913 * either from parent object or root object; for
914 * file case, it's from the @lo object
916 * \retval 0 if on disk structure is created successfully
917 * \retval negative error number on failure
919 int lod_generate_lovea(const struct lu_env *env, struct lod_object *lo,
920 struct lov_mds_md *lmm, int *lmm_size, bool is_dir)
922 struct lov_comp_md_entry_v1 *lcme;
923 struct lov_comp_md_v1 *lcm;
924 struct lod_layout_component *comp_entries;
925 __u16 comp_cnt, mirror_cnt;
926 bool is_composite, is_foreign = false;
927 int i, rc = 0, offset;
931 comp_cnt = lo->ldo_def_striping->lds_def_comp_cnt;
932 mirror_cnt = lo->ldo_def_striping->lds_def_mirror_cnt;
933 comp_entries = lo->ldo_def_striping->lds_def_comp_entries;
935 lo->ldo_def_striping->lds_def_striping_is_composite;
937 comp_cnt = lo->ldo_comp_cnt;
938 mirror_cnt = lo->ldo_mirror_count;
939 comp_entries = lo->ldo_comp_entries;
940 is_composite = lo->ldo_is_composite;
941 is_foreign = lo->ldo_is_foreign;
944 LASSERT(lmm_size != NULL);
947 struct lov_foreign_md *lfm;
949 lfm = (struct lov_foreign_md *)lmm;
950 memcpy(lfm, lo->ldo_foreign_lov, lo->ldo_foreign_lov_size);
951 /* need to store little-endian */
952 if (cpu_to_le32(LOV_MAGIC_FOREIGN) != LOV_MAGIC_FOREIGN) {
953 __swab32s(&lfm->lfm_magic);
954 __swab32s(&lfm->lfm_length);
955 __swab32s(&lfm->lfm_type);
956 __swab32s(&lfm->lfm_flags);
958 *lmm_size = lo->ldo_foreign_lov_size;
962 LASSERT(comp_cnt != 0 && comp_entries != NULL);
965 rc = lod_gen_component_ea(env, lo, 0, lmm, lmm_size, is_dir);
969 lcm = (struct lov_comp_md_v1 *)lmm;
970 memset(lcm, 0, sizeof(*lcm));
972 lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
973 lcm->lcm_entry_count = cpu_to_le16(comp_cnt);
974 lcm->lcm_mirror_count = cpu_to_le16(mirror_cnt - 1);
976 lcm->lcm_flags = cpu_to_le16(lo->ldo_flr_state);
978 lcm->lcm_flags = LCM_FL_NONE;
980 offset = sizeof(*lcm) + sizeof(*lcme) * comp_cnt;
981 LASSERT(offset % sizeof(__u64) == 0);
983 for (i = 0; i < comp_cnt; i++) {
984 struct lod_layout_component *lod_comp;
985 struct lov_mds_md *sub_md;
988 lod_comp = &comp_entries[i];
989 lcme = &lcm->lcm_entries[i];
991 LASSERT(ergo(!is_dir, lod_comp->llc_id != LCME_ID_INVAL));
992 lcme->lcme_id = cpu_to_le32(lod_comp->llc_id);
994 /* component could be un-inistantiated */
995 lcme->lcme_flags = cpu_to_le32(lod_comp->llc_flags);
996 if (lod_comp->llc_flags & LCME_FL_NOSYNC)
997 lcme->lcme_timestamp =
998 cpu_to_le64(lod_comp->llc_timestamp);
999 if (lod_comp->llc_flags & LCME_FL_EXTENSION && !is_dir)
1000 lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_SEL);
1002 lcme->lcme_extent.e_start =
1003 cpu_to_le64(lod_comp->llc_extent.e_start);
1004 lcme->lcme_extent.e_end =
1005 cpu_to_le64(lod_comp->llc_extent.e_end);
1006 lcme->lcme_offset = cpu_to_le32(offset);
1008 sub_md = (struct lov_mds_md *)((char *)lcm + offset);
1009 if (lod_comp->llc_magic == LOV_MAGIC_FOREIGN) {
1010 if (!lov_hsm_type_supported(lod_comp->llc_type)) {
1011 CDEBUG(D_LAYOUT, "Unknown HSM type: %u\n",
1012 lod_comp->llc_type);
1013 GOTO(out, rc = -EINVAL);
1015 rc = lod_gen_component_ea_foreign(env, lo, lod_comp,
1018 rc = lod_gen_component_ea(env, lo, i, sub_md,
1023 lcme->lcme_size = cpu_to_le32(size);
1025 LASSERTF((offset <= *lmm_size) && (offset % sizeof(__u64) == 0),
1026 "offset:%d lmm_size:%d\n", offset, *lmm_size);
1028 lcm->lcm_size = cpu_to_le32(offset);
1029 lcm->lcm_layout_gen = cpu_to_le32(is_dir ? 0 : lo->ldo_layout_gen);
1031 lustre_print_user_md(D_LAYOUT, (struct lov_user_md *)lmm,
1042 * Fill lti_ea_store buffer in the environment with a value for the given
1043 * EA. The buffer is reallocated if the value doesn't fit.
1045 * \param[in,out] env execution environment for this thread
1046 * .lti_ea_store buffer is filled with EA's value
1047 * \param[in] lo LOD object
1048 * \param[in] name name of the EA
1050 * \retval > 0 if EA is fetched successfully
1051 * \retval 0 if EA is empty
1052 * \retval negative error number on failure
1054 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
1057 struct lod_thread_info *info = lod_env_info(env);
1058 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1064 if (unlikely(info->lti_ea_store == NULL)) {
1065 /* just to enter in allocation block below */
1069 info->lti_buf.lb_buf = info->lti_ea_store;
1070 info->lti_buf.lb_len = info->lti_ea_store_size;
1071 rc = dt_xattr_get(env, next, &info->lti_buf, name);
1074 /* if object is not striped or inaccessible */
1075 if (rc == -ENODATA || rc == -ENOENT)
1078 if (rc == -ERANGE) {
1079 /* EA doesn't fit, reallocate new buffer */
1080 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name);
1081 if (rc == -ENODATA || rc == -ENOENT)
1087 if (rc <= info->lti_ea_store_size) {
1088 /* sometimes LOVEA can shrink in parallel */
1089 LASSERT(count++ < 10);
1092 rc = lod_ea_store_resize(info, rc);
1102 * Verify the target index is present in the current configuration.
1104 * \param[in] md LOD device where the target table is stored
1105 * \param[in] idx target's index
1107 * \retval 0 if the index is present
1108 * \retval -EINVAL if not
1110 int validate_lod_and_idx(struct lod_device *md, __u32 idx)
1112 if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
1113 !test_bit(idx, md->lod_ost_bitmap))) {
1114 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
1115 md->lod_ost_descs.ltd_tgts_size);
1119 if (unlikely(OST_TGT(md, idx) == NULL)) {
1120 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
1121 lod2obd(md)->obd_name, idx);
1125 if (unlikely(OST_TGT(md, idx)->ltd_tgt == NULL)) {
1126 CERROR("%s: invalid lod device, for idx: %d\n",
1127 lod2obd(md)->obd_name , idx);
1135 * Instantiate objects for stripes.
1137 * Allocate and initialize LU-objects representing the stripes. The number
1138 * of the stripes (llc_stripe_count) must be initialized already. The caller
1139 * must ensure nobody else is calling the function on the object at the same
1140 * time. FLDB service must be running to be able to map a FID to the targets
1141 * and find appropriate device representing that target.
1143 * \param[in] env execution environment for this thread
1144 * \param[in,out] lo LOD object
1145 * \param[in] objs an array of IDs to creates the objects from
1146 * \param[in] comp_idx index of ldo_comp_entries
1148 * \retval 0 if the objects are instantiated successfully
1149 * \retval negative error number on failure
1151 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
1152 struct lov_ost_data_v1 *objs, int comp_idx)
1154 struct lod_layout_component *lod_comp;
1155 struct lod_thread_info *info = lod_env_info(env);
1156 struct lod_device *md;
1157 struct lu_object *o, *n;
1158 struct lu_device *nd;
1159 struct dt_object **stripe = NULL;
1160 __u32 *ost_indices = NULL;
1166 LASSERT(lo != NULL);
1167 md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1169 LASSERT(lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL);
1170 lod_comp = &lo->ldo_comp_entries[comp_idx];
1172 LASSERT(lod_comp->llc_stripe == NULL);
1173 LASSERT(lod_comp->llc_stripe_count > 0);
1174 LASSERT(lod_comp->llc_stripe_size > 0);
1176 stripe_len = lod_comp->llc_stripe_count;
1177 OBD_ALLOC_PTR_ARRAY(stripe, stripe_len);
1180 OBD_ALLOC_PTR_ARRAY(ost_indices, stripe_len);
1182 GOTO(out, rc = -ENOMEM);
1184 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1185 if (unlikely(lovea_slot_is_dummy(&objs[i])))
1188 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
1189 idx = le32_to_cpu(objs[i].l_ost_idx);
1190 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
1193 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
1194 PFID(&info->lti_fid));
1195 lod_getref(&md->lod_ost_descs);
1197 rc = validate_lod_and_idx(md, idx);
1198 if (unlikely(rc != 0)) {
1199 lod_putref(md, &md->lod_ost_descs);
1203 nd = &OST_TGT(md, idx)->ltd_tgt->dd_lu_dev;
1204 lod_putref(md, &md->lod_ost_descs);
1206 /* In the function below, .hs_keycmp resolves to
1207 * u_obj_hop_keycmp() */
1208 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
1210 GOTO(out, rc = PTR_ERR(o));
1212 n = lu_object_locate(o->lo_header, nd->ld_type);
1215 stripe[i] = container_of(n, struct dt_object, do_lu);
1216 ost_indices[i] = idx;
1221 for (i = 0; i < stripe_len; i++)
1222 if (stripe[i] != NULL)
1223 dt_object_put(env, stripe[i]);
1225 OBD_FREE_PTR_ARRAY(stripe, stripe_len);
1226 lod_comp->llc_stripe_count = 0;
1228 OBD_FREE_PTR_ARRAY(ost_indices, stripe_len);
1230 lod_comp->llc_stripe = stripe;
1231 lod_comp->llc_ost_indices = ost_indices;
1232 lod_comp->llc_stripes_allocated = stripe_len;
1238 int lod_init_comp_foreign(struct lod_layout_component *lod_comp, void *lmm)
1240 struct lov_foreign_md *lfm;
1242 lfm = (struct lov_foreign_md *)lmm;
1243 lod_comp->llc_length = le32_to_cpu(lfm->lfm_length);
1244 lod_comp->llc_type = le32_to_cpu(lfm->lfm_type);
1246 if (!lov_hsm_type_supported(lod_comp->llc_type)) {
1248 "Unsupport HSM type: %u length: %u flags: %08X\n",
1249 lod_comp->llc_type, lod_comp->llc_length,
1250 le32_to_cpu(lfm->lfm_flags));
1255 * Currently it only stores the file FID as the field @lhm_archive_uuid
1256 * which is used to be the identifier within HSM backend for the archive
1258 * Thus the length of foreign layout value (HSM is a kind of foreign
1259 * layout type) is: sizeof(lhm_archive_id) + sizeof(lhm_archive_ver) +
1261 * It should fix to support other kinds of identifier for different HSM
1262 * solutions such as S3.
1264 if (lod_comp->llc_length != sizeof(struct lov_hsm_base)) {
1265 CDEBUG(D_LAYOUT, "Invalid HSM len: %u, should be %zu\n",
1266 lod_comp->llc_length, sizeof(struct lov_hsm_base));
1270 lod_comp->llc_foreign_flags = le32_to_cpu(lfm->lfm_flags);
1271 lov_foreign_hsm_to_cpu(&lod_comp->llc_hsm, lfm);
1276 * Instantiate objects for striping.
1278 * Parse striping information in \a buf and instantiate the objects
1279 * representing the stripes.
1281 * \param[in] env execution environment for this thread
1282 * \param[in] lo LOD object
1283 * \param[in] buf buffer storing LOV EA to parse
1284 * \param[in] lvf verify flags when parsing the layout
1286 * \retval 0 if parsing and objects creation succeed
1287 * \retval negative error number on failure
1289 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
1290 const struct lu_buf *buf, enum layout_verify_flags lvf)
1292 struct lod_device *d = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1293 struct lov_mds_md_v1 *lmm;
1294 struct lov_comp_md_v1 *comp_v1 = NULL;
1295 struct lov_foreign_md *foreign = NULL;
1296 struct lov_ost_data_v1 *objs;
1297 __u32 magic, pattern;
1298 __u16 mirror_cnt = 0;
1302 __u16 mirror_id = MIRROR_ID_NEG;
1304 int stale_mirrors = 0;
1308 LASSERT(buf->lb_buf);
1309 LASSERT(buf->lb_len);
1310 LASSERT(mutex_is_locked(&lo->ldo_layout_mutex));
1312 lmm = (struct lov_mds_md_v1 *)buf->lb_buf;
1313 magic = le32_to_cpu(lmm->lmm_magic);
1315 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1316 magic != LOV_MAGIC_COMP_V1 && magic != LOV_MAGIC_FOREIGN &&
1317 magic != LOV_MAGIC_SEL)
1318 GOTO(out, rc = -EINVAL);
1320 lod_striping_free_nolock(env, lo);
1322 if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
1323 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1324 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1326 GOTO(out, rc = -EINVAL);
1328 lo->ldo_layout_gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1329 lo->ldo_is_composite = 1;
1330 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1332 lo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1335 lo->ldo_flr_state = LCM_FL_NONE;
1336 } else if (magic == LOV_MAGIC_FOREIGN) {
1339 foreign = (struct lov_foreign_md *)buf->lb_buf;
1340 length = offsetof(typeof(*foreign), lfm_value);
1341 if (buf->lb_len < length ||
1342 buf->lb_len < (length + le32_to_cpu(foreign->lfm_length))) {
1344 "buf len %zu too small for lov_foreign_md\n",
1346 GOTO(out, rc = -EINVAL);
1349 /* just cache foreign LOV EA raw */
1350 rc = lod_alloc_foreign_lov(lo, length);
1353 memcpy(lo->ldo_foreign_lov, buf->lb_buf, length);
1357 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1358 lo->ldo_is_composite = 0;
1361 rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
1365 for (i = 0; i < comp_cnt; i++) {
1366 struct lod_layout_component *lod_comp;
1367 struct lu_extent *ext;
1370 lod_comp = &lo->ldo_comp_entries[i];
1371 if (lo->ldo_is_composite) {
1372 offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1373 lmm = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1375 ext = &comp_v1->lcm_entries[i].lcme_extent;
1376 lod_comp->llc_extent.e_start =
1377 le64_to_cpu(ext->e_start);
1378 if (lod_comp->llc_extent.e_start &
1379 (LOV_MIN_STRIPE_SIZE - 1)) {
1381 "extent start %llu is not a multiple of min size %u\n",
1382 lod_comp->llc_extent.e_start,
1383 LOV_MIN_STRIPE_SIZE);
1384 GOTO(out, rc = -EINVAL);
1387 lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1388 if (lod_comp->llc_extent.e_end != LUSTRE_EOF &&
1389 lod_comp->llc_extent.e_end &
1390 (LOV_MIN_STRIPE_SIZE - 1)) {
1392 "extent end %llu is not a multiple of min size %u\n",
1393 lod_comp->llc_extent.e_end,
1394 LOV_MIN_STRIPE_SIZE);
1395 GOTO(out, rc = -EINVAL);
1398 lod_comp->llc_flags =
1399 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1401 if (lod_comp->llc_flags & LCME_FL_NOSYNC)
1402 lod_comp->llc_timestamp = le64_to_cpu(
1403 comp_v1->lcm_entries[i].lcme_timestamp);
1405 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1406 if (lod_comp->llc_id == LCME_ID_INVAL)
1407 GOTO(out, rc = -EINVAL);
1409 if (lvf & LVF_ALL_STALE) {
1410 if (mirror_id_of(lod_comp->llc_id) ==
1412 /* remaining comps in the mirror */
1413 stale |= lod_comp->llc_flags &
1417 * new mirror, check last mirror's
1424 mirror_id_of(lod_comp->llc_id);
1426 /* the first comp of the new mirror */
1427 stale = lod_comp->llc_flags &
1432 if ((lod_comp->llc_flags & LCME_FL_EXTENSION) &&
1433 comp_v1->lcm_magic != cpu_to_le32(LOV_MAGIC_SEL)) {
1434 CWARN("%s: EXTENSION flags=%x set on component[%u]=%x of non-SEL file "DFID" with magic=%#08x\n",
1435 lod2obd(d)->obd_name,
1436 lod_comp->llc_flags, lod_comp->llc_id, i,
1437 PFID(lod_object_fid(lo)),
1438 le32_to_cpu(comp_v1->lcm_magic));
1441 lod_comp->llc_magic = le32_to_cpu(lmm->lmm_magic);
1442 if (lod_comp->llc_magic == LOV_MAGIC_FOREIGN) {
1443 rc = lod_init_comp_foreign(lod_comp, lmm);
1449 lod_comp->llc_magic = le32_to_cpu(lmm->lmm_magic);
1450 lod_comp_set_init(lod_comp);
1453 pattern = le32_to_cpu(lmm->lmm_pattern);
1454 if (!lov_pattern_supported(lov_pattern(pattern)))
1455 GOTO(out, rc = -EINVAL);
1457 if (pattern & LOV_PATTERN_MDT) {
1458 if (lod_comp->llc_extent.e_start != 0) {
1459 CERROR("%s: DOM entry must be the first stripe "
1460 "in a mirror\n", lod2obd(d)->obd_name);
1461 GOTO(out, rc = -EINVAL);
1464 dom_size = lod_comp->llc_extent.e_end;
1465 } else if (dom_size != lod_comp->llc_extent.e_end) {
1466 CERROR("%s: DOM entries with different sizes "
1467 "%#llx/%#llx\n", lod2obd(d)->obd_name,
1468 dom_size, lod_comp->llc_extent.e_end);
1469 GOTO(out, rc = -EINVAL);
1473 lod_comp->llc_pattern = pattern;
1474 lod_comp->llc_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
1475 lod_comp->llc_stripe_count = le16_to_cpu(lmm->lmm_stripe_count);
1476 lod_comp->llc_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1478 if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)) {
1479 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
1481 lod_set_pool(&lod_comp->llc_pool, v3->lmm_pool_name);
1482 objs = &v3->lmm_objects[0];
1484 lod_set_pool(&lod_comp->llc_pool, NULL);
1485 objs = &lmm->lmm_objects[0];
1489 * If uninstantiated template component has valid l_ost_idx,
1490 * then user has specified ost list for this component.
1492 if (!lod_comp_inited(lod_comp)) {
1495 if (objs[0].l_ost_idx != (__u32)-1UL) {
1498 stripe_count = lod_comp_entry_stripe_count(
1500 if (stripe_count == 0 &&
1501 !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1502 !(lod_comp->llc_pattern & LOV_PATTERN_MDT))
1503 GOTO(out, rc = -E2BIG);
1505 * load the user specified ost list, when this
1506 * component is instantiated later, it will be
1507 * used in lod_alloc_ost_list().
1509 lod_comp->llc_ostlist.op_count = stripe_count;
1510 lod_comp->llc_ostlist.op_size =
1511 stripe_count * sizeof(__u32);
1512 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1513 lod_comp->llc_ostlist.op_size);
1514 if (!lod_comp->llc_ostlist.op_array)
1515 GOTO(out, rc = -ENOMEM);
1517 for (j = 0; j < stripe_count; j++)
1518 lod_comp->llc_ostlist.op_array[j] =
1519 le32_to_cpu(objs[j].l_ost_idx);
1522 * this component OST objects starts from the
1523 * first ost_idx, lod_alloc_ost_list() will
1526 lod_comp->llc_stripe_offset = objs[0].l_ost_idx;
1529 * for uninstantiated component,
1530 * lmm_layout_gen stores default stripe offset.
1532 lod_comp->llc_stripe_offset =
1533 lmm->lmm_layout_gen;
1537 /* skip un-instantiated component object initialization */
1538 if (!lod_comp_inited(lod_comp))
1541 if (!(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1542 !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1543 rc = lod_initialize_objects(env, lo, objs, i);
1549 if (lo->ldo_is_composite && (lvf & LVF_ALL_STALE)) {
1550 /* check the last mirror stale-ness */
1554 if (mirror_cnt == stale_mirrors) {
1556 CERROR("%s: can not set all stale mirrors for "
1558 lod2obd(d)->obd_name, PFID(lod_object_fid(lo)),
1564 rc = lod_fill_mirrors(lo);
1570 lod_striping_free_nolock(env, lo);
1575 * Check whether the striping (LOVEA for regular file, LMVEA for directory)
1576 * is already cached.
1578 * \param[in] lo LOD object
1580 * \retval True if the striping is cached, otherwise
1583 static bool lod_striping_loaded(struct lod_object *lo)
1585 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1586 lo->ldo_comp_cached)
1589 if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1590 if (lo->ldo_dir_stripe_loaded)
1593 /* Never load LMV stripe for slaves of striped dir */
1594 if (lo->ldo_dir_slave_stripe)
1602 * A generic function to initialize the stripe objects.
1604 * A protected version of lod_striping_load_locked() - load the striping
1605 * information from storage, parse that and instantiate LU objects to
1606 * represent the stripes. The LOD object \a lo supplies a pointer to the
1607 * next sub-object in the LU stack so we can lock it. Also use \a lo to
1608 * return an array of references to the newly instantiated objects.
1610 * \param[in] env execution environment for this thread
1611 * \param[in,out] lo LOD object, where striping is stored and
1612 * which gets an array of references
1614 * \retval 0 if parsing and object creation succeed
1615 * \retval negative error number on failure
1617 int lod_striping_load(const struct lu_env *env, struct lod_object *lo)
1619 struct lod_thread_info *info = lod_env_info(env);
1620 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1621 struct lu_buf *buf = &info->lti_buf;
1626 if (!dt_object_exists(next))
1629 if (lod_striping_loaded(lo))
1632 mutex_lock(&lo->ldo_layout_mutex);
1633 if (lod_striping_loaded(lo))
1634 GOTO(unlock, rc = 0);
1636 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr)) {
1637 rc = lod_get_lov_ea(env, lo);
1642 * there is LOV EA (striping information) in this object
1643 * let's parse it and create in-core objects for the stripes
1645 buf->lb_buf = info->lti_ea_store;
1646 buf->lb_len = info->lti_ea_store_size;
1647 rc = lod_parse_striping(env, lo, buf, 0);
1649 lo->ldo_comp_cached = 1;
1650 } else if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1651 rc = lod_get_lmv_ea(env, lo);
1653 if (rc > (int)sizeof(struct lmv_foreign_md)) {
1654 struct lmv_foreign_md *lfm = info->lti_ea_store;
1656 if (le32_to_cpu(lfm->lfm_magic) == LMV_MAGIC_FOREIGN) {
1657 lo->ldo_foreign_lmv = info->lti_ea_store;
1658 lo->ldo_foreign_lmv_size =
1659 info->lti_ea_store_size;
1660 info->lti_ea_store = NULL;
1661 info->lti_ea_store_size = 0;
1663 lo->ldo_dir_stripe_loaded = 1;
1664 lo->ldo_is_foreign = 1;
1665 GOTO(unlock, rc = 0);
1669 if (rc < (int)sizeof(struct lmv_mds_md_v1)) {
1670 /* Let's set stripe_loaded to avoid further
1671 * stripe loading especially for non-stripe directory,
1672 * which can hurt performance. (See LU-9840)
1675 lo->ldo_dir_stripe_loaded = 1;
1676 GOTO(unlock, rc = rc > 0 ? -EINVAL : rc);
1678 buf->lb_buf = info->lti_ea_store;
1679 buf->lb_len = info->lti_ea_store_size;
1680 if (rc == sizeof(struct lmv_mds_md_v1)) {
1681 rc = lod_load_lmv_shards(env, lo, buf, true);
1682 if (buf->lb_buf != info->lti_ea_store) {
1683 OBD_FREE_LARGE(info->lti_ea_store,
1684 info->lti_ea_store_size);
1685 info->lti_ea_store = buf->lb_buf;
1686 info->lti_ea_store_size = buf->lb_len;
1694 * there is LMV EA (striping information) in this object
1695 * let's parse it and create in-core objects for the stripes
1697 rc = lod_parse_dir_striping(env, lo, buf);
1699 lo->ldo_dir_stripe_loaded = 1;
1703 mutex_unlock(&lo->ldo_layout_mutex);
1708 int lod_striping_reload(const struct lu_env *env, struct lod_object *lo,
1709 const struct lu_buf *buf, enum layout_verify_flags lvf)
1715 mutex_lock(&lo->ldo_layout_mutex);
1716 rc = lod_parse_striping(env, lo, buf, lvf);
1717 mutex_unlock(&lo->ldo_layout_mutex);
1723 * Verify lov_user_md_v1/v3 striping.
1725 * Check the validity of all fields including the magic, stripe size,
1726 * stripe count, stripe offset and that the pool is present. Also check
1727 * that each target index points to an existing target. The additional
1728 * \a is_from_disk turns additional checks. In some cases zero fields
1729 * are allowed (like pattern=0).
1731 * \param[in] d LOD device
1732 * \param[in] buf buffer with LOV EA to verify
1733 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
1734 * 1 - from disk, do not allow
1736 * \retval 0 if the striping is valid
1737 * \retval -EINVAL if striping is invalid
1739 static int lod_verify_v1v3(struct lod_device *d, const struct lu_buf *buf,
1742 struct lov_user_md_v1 *lum;
1743 struct lov_user_md_v3 *lum3;
1744 struct lod_pool_desc *pool = NULL;
1748 __u16 stripe_offset;
1755 if (buf->lb_len < sizeof(*lum)) {
1756 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1758 GOTO(out, rc = -EINVAL);
1761 magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1762 if (magic != LOV_USER_MAGIC_V1 &&
1763 magic != LOV_USER_MAGIC_V3 &&
1764 magic != LOV_USER_MAGIC_SPECIFIC) {
1765 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1766 le32_to_cpu(lum->lmm_magic));
1767 GOTO(out, rc = -EINVAL);
1770 /* the user uses "0" for default stripe pattern normally. */
1771 if (!is_from_disk && lum->lmm_pattern == LOV_PATTERN_NONE)
1772 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1774 if (!lov_pattern_supported(le32_to_cpu(lum->lmm_pattern))) {
1775 CDEBUG(D_LAYOUT, "bad userland stripe pattern: %#x\n",
1776 le32_to_cpu(lum->lmm_pattern));
1777 GOTO(out, rc = -EINVAL);
1780 /* a released lum comes from creating orphan on hsm release,
1781 * doesn't make sense to verify it. */
1782 if (le32_to_cpu(lum->lmm_pattern) & LOV_PATTERN_F_RELEASED)
1785 /* 64kB is the largest common page size we see (ia64), and matches the
1787 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1788 if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1789 CDEBUG(D_LAYOUT, "stripe size %u not a multiple of %u\n",
1790 stripe_size, LOV_MIN_STRIPE_SIZE);
1791 GOTO(out, rc = -EINVAL);
1794 stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1795 if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT &&
1796 !(lov_pattern(le32_to_cpu(lum->lmm_pattern)) & LOV_PATTERN_MDT)) {
1797 /* if offset is not within valid range [0, osts_size) */
1798 if (stripe_offset >= d->lod_ost_descs.ltd_tgts_size) {
1799 CDEBUG(D_LAYOUT, "stripe offset %u >= bitmap size %u\n",
1800 stripe_offset, d->lod_ost_descs.ltd_tgts_size);
1801 GOTO(out, rc = -EINVAL);
1804 /* if lmm_stripe_offset is *not* in bitmap */
1805 if (!test_bit(stripe_offset, d->lod_ost_bitmap)) {
1806 CDEBUG(D_LAYOUT, "stripe offset %u not in bitmap\n",
1808 GOTO(out, rc = -EINVAL);
1812 if (magic == LOV_USER_MAGIC_V1)
1813 lum_size = offsetof(struct lov_user_md_v1,
1815 else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC)
1816 lum_size = offsetof(struct lov_user_md_v3,
1819 GOTO(out, rc = -EINVAL);
1821 stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1822 if (buf->lb_len < lum_size) {
1823 CDEBUG(D_LAYOUT, "invalid buf len %zu/%zu for lov_user_md with "
1824 "magic %#x and stripe_count %u\n",
1825 buf->lb_len, lum_size, magic, stripe_count);
1826 GOTO(out, rc = -EINVAL);
1829 if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC))
1833 /* In the function below, .hs_keycmp resolves to
1834 * pool_hashkey_keycmp() */
1835 pool = lod_find_pool(d, lum3->lmm_pool_name);
1839 if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT) {
1840 rc = lod_check_index_in_pool(stripe_offset, pool);
1842 GOTO(out, rc = -EINVAL);
1845 if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1846 CDEBUG(D_LAYOUT, "stripe count %u > # OSTs %u in the pool\n",
1847 stripe_count, pool_tgt_count(pool));
1848 GOTO(out, rc = -EINVAL);
1853 lod_pool_putref(pool);
1859 struct lov_comp_md_entry_v1 *comp_entry_v1(struct lov_comp_md_v1 *comp, int i)
1861 LASSERTF((le32_to_cpu(comp->lcm_magic) & ~LOV_MAGIC_DEFINED) ==
1862 LOV_USER_MAGIC_COMP_V1 ||
1863 (le32_to_cpu(comp->lcm_magic) & ~LOV_MAGIC_DEFINED) ==
1864 LOV_USER_MAGIC_SEL, "Wrong magic %x\n",
1865 le32_to_cpu(comp->lcm_magic));
1866 LASSERTF(i >= 0 && i < le16_to_cpu(comp->lcm_entry_count),
1867 "bad index %d, max = %d\n",
1868 i, le16_to_cpu(comp->lcm_entry_count));
1870 return &comp->lcm_entries[i];
1873 #define for_each_comp_entry_v1(comp, entry) \
1874 for (entry = comp_entry_v1(comp, 0); \
1875 entry <= comp_entry_v1(comp, \
1876 le16_to_cpu(comp->lcm_entry_count) - 1); \
1879 static int lod_erase_dom_stripe(struct lov_comp_md_v1 *comp_v1,
1880 struct lov_comp_md_entry_v1 *dom_ent)
1882 struct lov_comp_md_entry_v1 *ent;
1884 __u32 dom_off, dom_size, comp_size, off;
1886 unsigned int size, shift;
1888 entries = le16_to_cpu(comp_v1->lcm_entry_count) - 1;
1889 LASSERT(entries > 0);
1890 comp_v1->lcm_entry_count = cpu_to_le16(entries);
1892 comp_size = le32_to_cpu(comp_v1->lcm_size);
1893 dom_off = le32_to_cpu(dom_ent->lcme_offset);
1894 dom_size = le32_to_cpu(dom_ent->lcme_size);
1896 /* all entries offsets are shifted by entry size at least */
1897 shift = sizeof(*dom_ent);
1898 for_each_comp_entry_v1(comp_v1, ent) {
1899 off = le32_to_cpu(ent->lcme_offset);
1900 if (off == dom_off) {
1901 /* Entry deletion creates two holes in layout data:
1902 * - hole in entries array
1903 * - hole in layout data at dom_off with dom_size
1905 * First memmove is one entry shift from next entry
1906 * start with size up to dom_off in blob
1909 src = (void *)(ent + 1);
1910 size = (unsigned long)((void *)comp_v1 + dom_off - src);
1911 memmove(dst, src, size);
1912 /* take 'off' from just moved entry */
1913 off = le32_to_cpu(ent->lcme_offset);
1914 /* second memmove is blob tail after 'off' up to
1917 dst = (void *)comp_v1 + dom_off - sizeof(*ent);
1918 src = (void *)comp_v1 + off;
1919 size = (unsigned long)(comp_size - off);
1920 memmove(dst, src, size);
1921 /* all entries offsets after DoM entry are shifted by
1922 * dom_size additionally
1926 ent->lcme_offset = cpu_to_le32(off - shift);
1928 comp_v1->lcm_size = cpu_to_le32(comp_size - shift);
1930 /* notify a caller to re-check entry */
1934 void lod_dom_stripesize_recalc(struct lod_device *d)
1936 __u64 threshold_mb = d->lod_dom_threshold_free_mb;
1937 __u32 max_size = d->lod_dom_stripesize_max_kb;
1938 __u32 def_size = d->lod_dom_stripesize_cur_kb;
1940 /* use maximum allowed value if free space is above threshold */
1941 if (d->lod_lsfs_free_mb >= threshold_mb) {
1942 def_size = max_size;
1943 } else if (!d->lod_lsfs_free_mb || max_size <= LOD_DOM_MIN_SIZE_KB) {
1946 /* recalc threshold like it would be with def_size as max */
1947 threshold_mb = mult_frac(threshold_mb, def_size, max_size);
1948 if (d->lod_lsfs_free_mb < threshold_mb)
1949 def_size = rounddown(def_size / 2, LOD_DOM_MIN_SIZE_KB);
1950 else if (d->lod_lsfs_free_mb > threshold_mb * 2)
1951 def_size = max_t(unsigned int, def_size * 2,
1952 LOD_DOM_MIN_SIZE_KB);
1955 if (d->lod_dom_stripesize_cur_kb != def_size) {
1956 CDEBUG(D_LAYOUT, "Change default DOM stripe size %d->%d\n",
1957 d->lod_dom_stripesize_cur_kb, def_size);
1958 d->lod_dom_stripesize_cur_kb = def_size;
1962 static __u32 lod_dom_stripesize_limit(const struct lu_env *env,
1963 struct lod_device *d)
1967 /* set bfree as fraction of total space */
1968 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_STATFS_SPOOF)) {
1969 spin_lock(&d->lod_lsfs_lock);
1970 d->lod_lsfs_free_mb = mult_frac(d->lod_lsfs_total_mb,
1971 min_t(int, cfs_fail_val, 100), 100);
1972 GOTO(recalc, rc = 0);
1975 if (d->lod_lsfs_age < ktime_get_seconds() - LOD_DOM_SFS_MAX_AGE) {
1976 struct obd_statfs sfs;
1978 spin_lock(&d->lod_lsfs_lock);
1979 if (d->lod_lsfs_age > ktime_get_seconds() - LOD_DOM_SFS_MAX_AGE)
1980 GOTO(unlock, rc = 0);
1982 d->lod_lsfs_age = ktime_get_seconds();
1983 spin_unlock(&d->lod_lsfs_lock);
1984 rc = dt_statfs(env, d->lod_child, &sfs);
1987 "%s: failed to get OSD statfs: rc = %d\n",
1988 lod2obd(d)->obd_name, rc);
1991 /* udpate local OSD cached statfs data */
1992 spin_lock(&d->lod_lsfs_lock);
1993 d->lod_lsfs_total_mb = (sfs.os_blocks * sfs.os_bsize) >> 20;
1994 d->lod_lsfs_free_mb = (sfs.os_bfree * sfs.os_bsize) >> 20;
1996 lod_dom_stripesize_recalc(d);
1998 spin_unlock(&d->lod_lsfs_lock);
2001 return d->lod_dom_stripesize_cur_kb << 10;
2004 static int lod_dom_stripesize_choose(const struct lu_env *env,
2005 struct lod_device *d,
2006 struct lov_comp_md_v1 *comp_v1,
2007 struct lov_comp_md_entry_v1 *dom_ent,
2010 struct lov_comp_md_entry_v1 *ent;
2011 struct lu_extent *dom_ext, *ext;
2012 struct lov_user_md_v1 *lum;
2013 __u32 max_stripe_size;
2016 bool dom_next_entry = false;
2018 dom_ext = &dom_ent->lcme_extent;
2019 dom_mid = mirror_id_of(le32_to_cpu(dom_ent->lcme_id));
2020 max_stripe_size = lod_dom_stripesize_limit(env, d);
2022 /* Check stripe size againts current per-MDT limit */
2023 if (stripe_size <= max_stripe_size)
2026 lum = (void *)comp_v1 + le32_to_cpu(dom_ent->lcme_offset);
2027 CDEBUG(D_LAYOUT, "overwrite DoM component size %u with MDT limit %u\n",
2028 stripe_size, max_stripe_size);
2029 lum->lmm_stripe_size = cpu_to_le32(max_stripe_size);
2031 /* In common case the DoM stripe is first entry in a mirror and
2032 * can be deleted only if it is not single entry in layout or
2033 * mirror, otherwise error should be returned.
2035 for_each_comp_entry_v1(comp_v1, ent) {
2039 mid = mirror_id_of(le32_to_cpu(ent->lcme_id));
2043 ext = &ent->lcme_extent;
2044 if (ext->e_start != dom_ext->e_end)
2047 /* Found next component after the DoM one with the same
2048 * mirror_id and adjust its start with DoM component end.
2050 * NOTE: we are considering here that there can be only one
2051 * DoM component in a file, all replicas are located on OSTs
2052 * always and don't need adjustment since use own layouts.
2054 ext->e_start = cpu_to_le64(max_stripe_size);
2055 dom_next_entry = true;
2059 if (max_stripe_size == 0) {
2060 /* DoM component size is zero due to server setting, remove
2061 * it from the layout but only if next component exists in
2062 * the same mirror. That must be checked prior calling the
2063 * lod_erase_dom_stripe().
2065 if (!dom_next_entry)
2068 rc = lod_erase_dom_stripe(comp_v1, dom_ent);
2070 /* Update DoM extent end finally */
2071 dom_ext->e_end = cpu_to_le64(max_stripe_size);
2078 * Verify LOV striping.
2080 * \param[in] d LOD device
2081 * \param[in] buf buffer with LOV EA to verify
2082 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
2083 * 1 - from disk, do not allow
2084 * \param[in] start extent start for composite layout
2086 * \retval 0 if the striping is valid
2087 * \retval -EINVAL if striping is invalid
2089 int lod_verify_striping(const struct lu_env *env, struct lod_device *d,
2090 struct lod_object *lo, const struct lu_buf *buf,
2093 struct lov_user_md_v1 *lum;
2094 struct lov_comp_md_v1 *comp_v1;
2095 struct lov_comp_md_entry_v1 *ent;
2096 struct lu_extent *ext;
2099 __u32 stripe_size = 0;
2100 __u16 prev_mid = -1, mirror_id = -1;
2106 if (buf->lb_len < sizeof(lum->lmm_magic)) {
2107 CDEBUG(D_LAYOUT, "invalid buf len %zu\n", buf->lb_len);
2113 magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
2114 /* treat foreign LOV EA/object case first
2115 * XXX is it expected to try setting again a foreign?
2116 * XXX should we care about different current vs new layouts ?
2118 if (unlikely(magic == LOV_USER_MAGIC_FOREIGN)) {
2119 struct lov_foreign_md *lfm = buf->lb_buf;
2121 if (buf->lb_len < offsetof(typeof(*lfm), lfm_value)) {
2123 "buf len %zu < min lov_foreign_md size (%zu)\n",
2124 buf->lb_len, offsetof(typeof(*lfm),
2129 if (lov_foreign_size_le(lfm) > buf->lb_len) {
2131 "buf len %zu < this lov_foreign_md size (%zu)\n",
2132 buf->lb_len, lov_foreign_size_le(lfm));
2135 /* Don't do anything with foreign layouts */
2139 /* normal LOV/layout cases */
2141 if (buf->lb_len < sizeof(*lum)) {
2142 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
2148 case LOV_USER_MAGIC_FOREIGN:
2150 case LOV_USER_MAGIC_V1:
2151 case LOV_USER_MAGIC_V3:
2152 case LOV_USER_MAGIC_SPECIFIC:
2153 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) &
2155 /* DoM must use composite layout */
2156 CDEBUG(D_LAYOUT, "DoM without composite layout\n");
2159 RETURN(lod_verify_v1v3(d, buf, is_from_disk));
2160 case LOV_USER_MAGIC_COMP_V1:
2161 case LOV_USER_MAGIC_SEL:
2164 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
2165 le32_to_cpu(lum->lmm_magic));
2169 /* magic == LOV_USER_MAGIC_COMP_V1 */
2170 comp_v1 = buf->lb_buf;
2171 if (buf->lb_len < le32_to_cpu(comp_v1->lcm_size)) {
2172 CDEBUG(D_LAYOUT, "buf len %zu is less than %u\n",
2173 buf->lb_len, le32_to_cpu(comp_v1->lcm_size));
2179 if (le16_to_cpu(comp_v1->lcm_entry_count) == 0) {
2180 CDEBUG(D_LAYOUT, "entry count is zero\n");
2184 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
2185 lo->ldo_comp_cnt > 0) {
2186 /* could be called from lustre.lov.add */
2187 __u32 cnt = lo->ldo_comp_cnt;
2189 ext = &lo->ldo_comp_entries[cnt - 1].llc_extent;
2190 prev_end = ext->e_end;
2195 for_each_comp_entry_v1(comp_v1, ent) {
2196 ext = &ent->lcme_extent;
2198 if (le64_to_cpu(ext->e_start) > le64_to_cpu(ext->e_end) ||
2199 le64_to_cpu(ext->e_start) & (LOV_MIN_STRIPE_SIZE - 1) ||
2200 ((__s64)le64_to_cpu(ext->e_start) < 0 &&
2201 le64_to_cpu(ext->e_start) != LUSTRE_EOF) ||
2202 (le64_to_cpu(ext->e_end) != LUSTRE_EOF &&
2203 le64_to_cpu(ext->e_end) & (LOV_MIN_STRIPE_SIZE - 1)) ||
2204 ((__s64)le64_to_cpu(ext->e_end) < 0 &&
2205 le64_to_cpu(ext->e_end) != LUSTRE_EOF)) {
2206 CDEBUG(D_LAYOUT, "invalid extent "DEXT"\n",
2207 le64_to_cpu(ext->e_start),
2208 le64_to_cpu(ext->e_end));
2213 /* lcme_id contains valid value */
2214 if (le32_to_cpu(ent->lcme_id) == 0 ||
2215 le32_to_cpu(ent->lcme_id) > LCME_ID_MAX) {
2216 CDEBUG(D_LAYOUT, "invalid id %u\n",
2217 le32_to_cpu(ent->lcme_id));
2221 if (le16_to_cpu(comp_v1->lcm_mirror_count) > 0) {
2222 mirror_id = mirror_id_of(
2223 le32_to_cpu(ent->lcme_id));
2225 /* first component must start with 0 */
2226 if (mirror_id != prev_mid &&
2227 le64_to_cpu(ext->e_start) != 0) {
2229 "invalid start:%llu, expect:0\n",
2230 le64_to_cpu(ext->e_start));
2234 prev_mid = mirror_id;
2238 if (le64_to_cpu(ext->e_start) == 0) {
2243 /* the next must be adjacent with the previous one */
2244 if (le64_to_cpu(ext->e_start) != prev_end) {
2246 "invalid start actual:%llu, expect:%llu\n",
2247 le64_to_cpu(ext->e_start), prev_end);
2251 tmp.lb_buf = (char *)comp_v1 + le32_to_cpu(ent->lcme_offset);
2252 tmp.lb_len = le32_to_cpu(ent->lcme_size);
2255 if (le32_to_cpu(lum->lmm_magic) == LOV_MAGIC_FOREIGN) {
2256 struct lov_foreign_md *lfm;
2257 struct lov_hsm_md *lhm;
2262 * Currently when the foreign layout is used as a basic
2263 * layout component, it only supports HSM foreign types:
2264 * LU_FOREIGN_TYPE_{POSIX, S3, PCCRW, PCCRO}.
2266 lfm = (struct lov_foreign_md *)lum;
2267 ftype = le32_to_cpu(lfm->lfm_type);
2268 if (!lov_hsm_type_supported(ftype)) {
2270 "Foreign type %#x is not HSM\n", ftype);
2274 /* Current HSM component must cover [0, EOF]. */
2275 if (le64_to_cpu(ext->e_start) > 0) {
2276 CDEBUG(D_LAYOUT, "Invalid HSM component with %llu extent start\n",
2277 le64_to_cpu(ext->e_start));
2280 if (le64_to_cpu(ext->e_end) != LUSTRE_EOF) {
2281 CDEBUG(D_LAYOUT, "Invalid HSM component with %llu extent end\n",
2282 le64_to_cpu(ext->e_end));
2286 lhm = (struct lov_hsm_md *)lfm;
2287 hsmsize = lov_foreign_size_le(lhm);
2289 if (le32_to_cpu(lhm->lhm_length) !=
2290 sizeof(struct lov_hsm_base)) {
2292 "Invalid HSM component size %u != %u\n",
2293 le32_to_cpu(ent->lcme_size), hsmsize);
2297 if (le32_to_cpu(ent->lcme_size) < hsmsize) {
2299 "Invalid HSM component size %u != %u\n",
2300 le32_to_cpu(ent->lcme_size), hsmsize);
2303 if (le32_to_cpu(lhm->lhm_flags) & ~HSM_FLAGS_MASK ||
2304 !(le32_to_cpu(lhm->lhm_flags) & HSM_FLAGS_MASK)) {
2306 "Invalid HSM component flags %#x\n",
2307 le32_to_cpu(lhm->lhm_flags));
2313 /* Check DoM entry is always the first one */
2314 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) &
2316 /* DoM component must be the first in a mirror */
2317 if (le64_to_cpu(ext->e_start) > 0) {
2318 CDEBUG(D_LAYOUT, "invalid DoM component "
2319 "with %llu extent start\n",
2320 le64_to_cpu(ext->e_start));
2323 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
2324 /* There is just one stripe on MDT and it must
2325 * cover whole component size. */
2326 if (stripe_size != le64_to_cpu(ext->e_end)) {
2327 CDEBUG(D_LAYOUT, "invalid DoM layout "
2328 "stripe size %u != %llu "
2329 "(component size)\n",
2330 stripe_size, prev_end);
2333 /* Check and adjust stripe size by per-MDT limit */
2334 rc = lod_dom_stripesize_choose(env, d, comp_v1, ent,
2336 /* DoM entry was removed, re-check layout from start */
2337 if (rc == -ERESTART)
2342 if (le16_to_cpu(lum->lmm_stripe_count) == 1)
2343 lum->lmm_stripe_count = 0;
2344 /* Any stripe count is forbidden on DoM component */
2345 if (lum->lmm_stripe_count > 0) {
2347 "invalid DoM layout stripe count %u, must be 0\n",
2348 le16_to_cpu(lum->lmm_stripe_count));
2352 /* Any pool is forbidden on DoM component */
2353 if (lum->lmm_magic == LOV_USER_MAGIC_V3) {
2354 struct lov_user_md_v3 *v3 = (void *)lum;
2356 if (v3->lmm_pool_name[0] != '\0') {
2358 "DoM component cannot have pool assigned\n");
2364 prev_end = le64_to_cpu(ext->e_end);
2366 rc = lod_verify_v1v3(d, &tmp, is_from_disk);
2370 if (prev_end == LUSTRE_EOF || ext->e_start == prev_end)
2373 /* extent end must be aligned with the stripe_size */
2374 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
2375 if (stripe_size && prev_end % stripe_size) {
2376 CDEBUG(D_LAYOUT, "stripe size isn't aligned, "
2377 "stripe_sz: %u, [%llu, %llu)\n",
2378 stripe_size, ext->e_start, prev_end);
2383 /* make sure that the mirror_count is telling the truth */
2384 if (mirror_count != le16_to_cpu(comp_v1->lcm_mirror_count) + 1)
2391 * set the default stripe size, if unset.
2393 * \param[in,out] val number of bytes per OST stripe
2395 * The minimum stripe size is 64KB to ensure that a single stripe is an
2396 * even multiple of a client PAGE_SIZE (IA64, PPC, etc). Otherwise, it
2397 * is difficult to split dirty pages across OSCs during writes.
2399 void lod_fix_desc_stripe_size(__u64 *val)
2401 if (*val < LOV_MIN_STRIPE_SIZE) {
2403 LCONSOLE_INFO("Increasing default stripe size to "
2404 "minimum value %u\n",
2405 LOV_DESC_STRIPE_SIZE_DEFAULT);
2406 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
2407 } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
2408 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
2409 LCONSOLE_WARN("Changing default stripe size to %llu (a "
2410 "multiple of %u)\n",
2411 *val, LOV_MIN_STRIPE_SIZE);
2416 * set the filesystem default number of stripes, if unset.
2418 * \param[in,out] val number of stripes
2420 * A value of "0" means "use the system-wide default stripe count", which
2421 * has either been inherited by now, or falls back to 1 stripe per file.
2422 * A value of "-1" (0xffffffff) means "stripe over all available OSTs",
2423 * and is a valid value, so is left unchanged here.
2425 void lod_fix_desc_stripe_count(__u32 *val)
2432 * set the filesystem default layout pattern
2434 * \param[in,out] val LOV_PATTERN_* layout
2436 * A value of "0" means "use the system-wide default layout type", which
2437 * has either been inherited by now, or falls back to plain RAID0 striping.
2439 void lod_fix_desc_pattern(__u32 *val)
2441 /* from lov_setstripe */
2442 if ((*val != 0) && !lov_pattern_supported_normal_comp(*val)) {
2443 LCONSOLE_WARN("lod: Unknown stripe pattern: %#x\n", *val);
2448 void lod_fix_lmv_desc_pattern(__u32 *val)
2450 if ((*val) && !lmv_is_known_hash_type(*val)) {
2451 LCONSOLE_WARN("lod: Unknown md stripe pattern: %#x\n", *val);
2456 void lod_fix_desc_qos_maxage(__u32 *val)
2458 /* fix qos_maxage */
2460 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
2464 * Used to fix insane default striping.
2466 * \param[in] desc striping description
2468 void lod_fix_desc(struct lov_desc *desc)
2470 lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
2471 lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
2472 lod_fix_desc_pattern(&desc->ld_pattern);
2473 lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
2476 static void lod_fix_lmv_desc(struct lmv_desc *desc)
2478 desc->ld_active_tgt_count = 0;
2479 lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
2480 lod_fix_lmv_desc_pattern(&desc->ld_pattern);
2481 lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
2485 * Initialize the structures used to store pools and default striping.
2487 * \param[in] lod LOD device
2488 * \param[in] lcfg configuration structure storing default striping.
2490 * \retval 0 if initialization succeeds
2491 * \retval negative error number on failure
2493 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
2495 struct obd_device *obd;
2496 struct lov_desc *desc;
2500 obd = class_name2obd(lustre_cfg_string(lcfg, 0));
2501 LASSERT(obd != NULL);
2502 obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
2504 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
2505 CERROR("LOD setup requires a descriptor\n");
2509 desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
2511 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
2512 CERROR("descriptor size wrong: %d > %d\n",
2513 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
2517 if (desc->ld_magic != LOV_DESC_MAGIC) {
2518 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
2519 CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
2520 obd->obd_name, desc);
2521 lustre_swab_lov_desc(desc);
2523 CERROR("%s: Bad lov desc magic: %#x\n",
2524 obd->obd_name, desc->ld_magic);
2531 desc->ld_active_tgt_count = 0;
2532 lod->lod_ost_descs.ltd_lov_desc = *desc;
2534 /* NB: config doesn't contain lmv_desc, alter it via sysfs. */
2535 lod_fix_lmv_desc(&lod->lod_mdt_descs.ltd_lmv_desc);
2537 lod->lod_sp_me = LUSTRE_SP_CLI;
2539 /* Set up OST pool environment */
2540 lod->lod_pool_count = 0;
2541 rc = lod_pool_hash_init(&lod->lod_pools_hash_body);
2545 INIT_LIST_HEAD(&lod->lod_pool_list);
2546 lod->lod_pool_count = 0;
2547 rc = lu_tgt_pool_init(&lod->lod_mdt_descs.ltd_tgt_pool, 0);
2551 rc = lu_tgt_pool_init(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool, 0);
2553 GOTO(out_mdt_pool, rc);
2555 rc = lu_tgt_pool_init(&lod->lod_ost_descs.ltd_tgt_pool, 0);
2557 GOTO(out_mdt_rr_pool, rc);
2559 rc = lu_tgt_pool_init(&lod->lod_ost_descs.ltd_qos.lq_rr.lqr_pool, 0);
2561 GOTO(out_ost_pool, rc);
2566 lu_tgt_pool_free(&lod->lod_ost_descs.ltd_tgt_pool);
2568 lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool);
2570 lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_tgt_pool);
2572 lod_pool_hash_destroy(&lod->lod_pools_hash_body);
2578 * Release the structures describing the pools.
2580 * \param[in] lod LOD device from which we release the structures
2584 int lod_pools_fini(struct lod_device *lod)
2586 struct obd_device *obd = lod2obd(lod);
2587 struct lod_pool_desc *pool, *tmp;
2590 list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
2591 /* free pool structs */
2592 CDEBUG(D_INFO, "delete pool %p\n", pool);
2593 /* In the function below, .hs_keycmp resolves to
2594 * pool_hashkey_keycmp() */
2595 lod_pool_del(obd, pool->pool_name);
2598 lod_pool_hash_destroy(&lod->lod_pools_hash_body);
2599 lu_tgt_pool_free(&lod->lod_ost_descs.ltd_qos.lq_rr.lqr_pool);
2600 lu_tgt_pool_free(&lod->lod_ost_descs.ltd_tgt_pool);
2601 lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool);
2602 lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_tgt_pool);