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 (lod_comp->llc_pattern == 0) /* default striping */
757 lod_comp->llc_pattern = LOV_PATTERN_RAID0;
759 lmm->lmm_magic = cpu_to_le32(magic);
760 lmm->lmm_pattern = cpu_to_le32(lod_comp->llc_pattern);
761 fid_to_lmm_oi(fid, &lmm->lmm_oi);
762 if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
763 lmm->lmm_oi.oi.oi_id++;
764 lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
766 lmm->lmm_stripe_size = cpu_to_le32(lod_comp->llc_stripe_size);
767 lmm->lmm_stripe_count = cpu_to_le16(lod_comp->llc_stripe_count);
769 * for dir and uninstantiated component, lmm_layout_gen stores
770 * default stripe offset.
772 lmm->lmm_layout_gen =
773 (is_dir || !lod_comp_inited(lod_comp)) ?
774 cpu_to_le16(lod_comp->llc_stripe_offset) :
775 cpu_to_le16(lod_comp->llc_layout_gen);
777 if (magic == LOV_MAGIC_V1) {
778 objs = &lmm->lmm_objects[0];
780 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
781 size_t cplen = strscpy(v3->lmm_pool_name,
783 sizeof(v3->lmm_pool_name));
786 objs = &v3->lmm_objects[0];
788 lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
789 stripe_count = lod_comp_entry_stripe_count(lo, comp_idx, is_dir);
790 if (stripe_count == 0 && !is_dir &&
791 !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
792 !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
793 /* Try again if all active targets are disconnected.
794 * It is possible when MDS does failover. */
795 if (!lod->lod_ost_active_count &&
801 if (!is_dir && lo->ldo_is_composite)
802 lod_comp_shrink_stripe_count(lod_comp, &stripe_count);
804 if (is_dir || lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
807 /* generate ost_idx of this component stripe */
808 for (i = 0; i < stripe_count; i++) {
809 struct dt_object *object;
810 __u32 ost_idx = (__u32)-1UL;
811 int type = LU_SEQ_RANGE_OST;
813 if (lod_comp->llc_stripe && lod_comp->llc_stripe[i]) {
814 object = lod_comp->llc_stripe[i];
815 /* instantiated component */
816 info->lti_fid = *lu_object_fid(&object->do_lu);
818 if (CFS_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF) &&
820 if (cfs_fail_val == 0)
821 cfs_fail_val = info->lti_fid.f_oid;
823 info->lti_fid.f_oid = cfs_fail_val;
826 rc = fid_to_ostid(&info->lti_fid, &info->lti_ostid);
829 ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
830 objs[i].l_ost_gen = cpu_to_le32(0);
831 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_FLD_LOOKUP))
834 rc = lod_fld_lookup(env, lod, &info->lti_fid,
837 CERROR("%s: Can not locate "DFID": rc = %d\n",
838 lod2obd(lod)->obd_name,
839 PFID(&info->lti_fid), rc);
842 } else if (lod_comp->llc_ostlist.op_array &&
843 lod_comp->llc_ostlist.op_count) {
844 /* user specified ost list */
845 ost_idx = lod_comp->llc_ostlist.op_array[i];
848 * with un-instantiated or with no specified ost list
849 * component, its l_ost_idx does not matter.
851 objs[i].l_ost_idx = cpu_to_le32(ost_idx);
853 /* simulation of broken LOVEA */
854 if (CFS_FAIL_CHECK(OBD_FAIL_LOV_INVALID_OSTIDX) &&
855 comp_idx == 0 && i == 0 && lo->ldo_mirror_count > 1) {
856 objs[i].l_ost_idx = cpu_to_le32(0xffffffff);
861 if (lmm_size != NULL)
862 *lmm_size = lov_mds_md_size(stripe_count, magic);
867 * Generate on-disk lov_hsm_md structure based on the information in
868 * the lod_object->ldo_comp_entries.
870 static int lod_gen_component_ea_foreign(const struct lu_env *env,
871 struct lod_object *lo,
872 struct lod_layout_component *lod_comp,
873 void *lmm, int *lmm_size)
875 struct lov_foreign_md *lfm = (struct lov_foreign_md *)lmm;
879 lfm->lfm_magic = cpu_to_le32(LOV_MAGIC_FOREIGN);
880 lfm->lfm_length = cpu_to_le32(lod_comp->llc_length);
881 lfm->lfm_type = cpu_to_le32(lod_comp->llc_type);
882 lfm->lfm_flags = cpu_to_le32(lod_comp->llc_foreign_flags);
884 if (lov_hsm_type_supported(lod_comp->llc_type)) {
885 if (lod_comp->llc_length != sizeof(struct lov_hsm_base))
888 lov_foreign_hsm_to_le(lfm, &lod_comp->llc_hsm);
892 *lmm_size = lov_foreign_md_size(lod_comp->llc_length);
898 * Generate on-disk lov_mds_md structure based on the information in
899 * the lod_object->ldo_comp_entries.
901 * \param[in] env execution environment for this thread
902 * \param[in] lo LOD object
903 * \param[in] lmm buffer to cotain the on-disk lov_mds_md
904 * \param[in|out] lmm_size buffer size/lmm size
905 * \param[in] is_dir generate lov ea for dir or file? For dir case,
906 * the stripe info is from the default stripe
907 * template, which is collected in lod_ah_init(),
908 * either from parent object or root object; for
909 * file case, it's from the @lo object
911 * \retval 0 if on disk structure is created successfully
912 * \retval negative error number on failure
914 int lod_generate_lovea(const struct lu_env *env, struct lod_object *lo,
915 struct lov_mds_md *lmm, int *lmm_size, bool is_dir)
917 struct lov_comp_md_entry_v1 *lcme;
918 struct lov_comp_md_v1 *lcm;
919 struct lod_layout_component *comp_entries;
920 __u16 comp_cnt, mirror_cnt;
921 bool is_composite, is_foreign = false;
922 int i, rc = 0, offset;
926 comp_cnt = lo->ldo_def_striping->lds_def_comp_cnt;
927 mirror_cnt = lo->ldo_def_striping->lds_def_mirror_cnt;
928 comp_entries = lo->ldo_def_striping->lds_def_comp_entries;
930 lo->ldo_def_striping->lds_def_striping_is_composite;
932 comp_cnt = lo->ldo_comp_cnt;
933 mirror_cnt = lo->ldo_mirror_count;
934 comp_entries = lo->ldo_comp_entries;
935 is_composite = lo->ldo_is_composite;
936 is_foreign = lo->ldo_is_foreign;
939 LASSERT(lmm_size != NULL);
942 struct lov_foreign_md *lfm;
944 lfm = (struct lov_foreign_md *)lmm;
945 memcpy(lfm, lo->ldo_foreign_lov, lo->ldo_foreign_lov_size);
946 /* need to store little-endian */
947 if (cpu_to_le32(LOV_MAGIC_FOREIGN) != LOV_MAGIC_FOREIGN) {
948 __swab32s(&lfm->lfm_magic);
949 __swab32s(&lfm->lfm_length);
950 __swab32s(&lfm->lfm_type);
951 __swab32s(&lfm->lfm_flags);
953 *lmm_size = lo->ldo_foreign_lov_size;
957 LASSERT(comp_cnt != 0 && comp_entries != NULL);
960 rc = lod_gen_component_ea(env, lo, 0, lmm, lmm_size, is_dir);
964 lcm = (struct lov_comp_md_v1 *)lmm;
965 memset(lcm, 0, sizeof(*lcm));
967 lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
968 lcm->lcm_entry_count = cpu_to_le16(comp_cnt);
969 lcm->lcm_mirror_count = cpu_to_le16(mirror_cnt - 1);
971 lcm->lcm_flags = cpu_to_le16(lo->ldo_flr_state);
973 lcm->lcm_flags = LCM_FL_NONE;
975 offset = sizeof(*lcm) + sizeof(*lcme) * comp_cnt;
976 LASSERT(offset % sizeof(__u64) == 0);
978 for (i = 0; i < comp_cnt; i++) {
979 struct lod_layout_component *lod_comp;
980 struct lov_mds_md *sub_md;
983 lod_comp = &comp_entries[i];
984 lcme = &lcm->lcm_entries[i];
986 LASSERT(ergo(!is_dir, lod_comp->llc_id != LCME_ID_INVAL));
987 lcme->lcme_id = cpu_to_le32(lod_comp->llc_id);
989 /* component could be un-inistantiated */
990 lcme->lcme_flags = cpu_to_le32(lod_comp->llc_flags);
991 if (lod_comp->llc_flags & LCME_FL_NOSYNC)
992 lcme->lcme_timestamp =
993 cpu_to_le64(lod_comp->llc_timestamp);
994 if (lod_comp->llc_flags & LCME_FL_EXTENSION && !is_dir)
995 lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_SEL);
997 lcme->lcme_extent.e_start =
998 cpu_to_le64(lod_comp->llc_extent.e_start);
999 lcme->lcme_extent.e_end =
1000 cpu_to_le64(lod_comp->llc_extent.e_end);
1001 lcme->lcme_offset = cpu_to_le32(offset);
1003 sub_md = (struct lov_mds_md *)((char *)lcm + offset);
1004 if (lod_comp->llc_magic == LOV_MAGIC_FOREIGN) {
1005 if (!lov_hsm_type_supported(lod_comp->llc_type)) {
1006 CDEBUG(D_LAYOUT, "Unknown HSM type: %u\n",
1007 lod_comp->llc_type);
1008 GOTO(out, rc = -EINVAL);
1010 rc = lod_gen_component_ea_foreign(env, lo, lod_comp,
1013 rc = lod_gen_component_ea(env, lo, i, sub_md,
1018 lcme->lcme_size = cpu_to_le32(size);
1020 LASSERTF((offset <= *lmm_size) && (offset % sizeof(__u64) == 0),
1021 "offset:%d lmm_size:%d\n", offset, *lmm_size);
1023 lcm->lcm_size = cpu_to_le32(offset);
1024 lcm->lcm_layout_gen = cpu_to_le32(is_dir ? 0 : lo->ldo_layout_gen);
1026 lustre_print_user_md(D_LAYOUT, (struct lov_user_md *)lmm,
1037 * Fill lti_ea_store buffer in the environment with a value for the given
1038 * EA. The buffer is reallocated if the value doesn't fit.
1040 * \param[in,out] env execution environment for this thread
1041 * .lti_ea_store buffer is filled with EA's value
1042 * \param[in] lo LOD object
1043 * \param[in] name name of the EA
1045 * \retval > 0 if EA is fetched successfully
1046 * \retval 0 if EA is empty
1047 * \retval negative error number on failure
1049 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
1052 struct lod_thread_info *info = lod_env_info(env);
1053 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1059 if (unlikely(info->lti_ea_store == NULL)) {
1060 /* just to enter in allocation block below */
1064 info->lti_buf.lb_buf = info->lti_ea_store;
1065 info->lti_buf.lb_len = info->lti_ea_store_size;
1066 rc = dt_xattr_get(env, next, &info->lti_buf, name);
1069 /* if object is not striped or inaccessible */
1070 if (rc == -ENODATA || rc == -ENOENT)
1073 if (rc == -ERANGE) {
1074 /* EA doesn't fit, reallocate new buffer */
1075 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name);
1076 if (rc == -ENODATA || rc == -ENOENT)
1082 if (rc <= info->lti_ea_store_size) {
1083 /* sometimes LOVEA can shrink in parallel */
1084 LASSERT(count++ < 10);
1087 rc = lod_ea_store_resize(info, rc);
1097 * Verify the target index is present in the current configuration.
1099 * \param[in] md LOD device where the target table is stored
1100 * \param[in] idx target's index
1102 * \retval 0 if the index is present
1103 * \retval -EINVAL if not
1105 int validate_lod_and_idx(struct lod_device *md, __u32 idx)
1107 if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
1108 !test_bit(idx, md->lod_ost_bitmap))) {
1109 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
1110 md->lod_ost_descs.ltd_tgts_size);
1114 if (unlikely(OST_TGT(md, idx) == NULL)) {
1115 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
1116 lod2obd(md)->obd_name, idx);
1120 if (unlikely(OST_TGT(md, idx)->ltd_tgt == NULL)) {
1121 CERROR("%s: invalid lod device, for idx: %d\n",
1122 lod2obd(md)->obd_name , idx);
1130 * Instantiate objects for stripes.
1132 * Allocate and initialize LU-objects representing the stripes. The number
1133 * of the stripes (llc_stripe_count) must be initialized already. The caller
1134 * must ensure nobody else is calling the function on the object at the same
1135 * time. FLDB service must be running to be able to map a FID to the targets
1136 * and find appropriate device representing that target.
1138 * \param[in] env execution environment for this thread
1139 * \param[in,out] lo LOD object
1140 * \param[in] objs an array of IDs to creates the objects from
1141 * \param[in] comp_idx index of ldo_comp_entries
1143 * \retval 0 if the objects are instantiated successfully
1144 * \retval negative error number on failure
1146 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
1147 struct lov_ost_data_v1 *objs, int comp_idx)
1149 struct lod_layout_component *lod_comp;
1150 struct lod_thread_info *info = lod_env_info(env);
1151 struct lod_device *md;
1152 struct lu_object *o, *n;
1153 struct lu_device *nd;
1154 struct dt_object **stripe = NULL;
1155 __u32 *ost_indices = NULL;
1161 LASSERT(lo != NULL);
1162 md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1164 LASSERT(lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL);
1165 lod_comp = &lo->ldo_comp_entries[comp_idx];
1167 LASSERT(lod_comp->llc_stripe == NULL);
1168 LASSERT(lod_comp->llc_stripe_count > 0);
1169 LASSERT(lod_comp->llc_stripe_size > 0);
1171 stripe_len = lod_comp->llc_stripe_count;
1172 OBD_ALLOC_PTR_ARRAY(stripe, stripe_len);
1175 OBD_ALLOC_PTR_ARRAY(ost_indices, stripe_len);
1177 GOTO(out, rc = -ENOMEM);
1179 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1180 if (unlikely(lovea_slot_is_dummy(&objs[i])))
1183 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
1184 idx = le32_to_cpu(objs[i].l_ost_idx);
1185 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
1188 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
1189 PFID(&info->lti_fid));
1190 lod_getref(&md->lod_ost_descs);
1192 rc = validate_lod_and_idx(md, idx);
1193 if (unlikely(rc != 0)) {
1194 lod_putref(md, &md->lod_ost_descs);
1198 nd = &OST_TGT(md, idx)->ltd_tgt->dd_lu_dev;
1199 lod_putref(md, &md->lod_ost_descs);
1201 /* In the function below, .hs_keycmp resolves to
1202 * u_obj_hop_keycmp() */
1203 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
1205 GOTO(out, rc = PTR_ERR(o));
1207 n = lu_object_locate(o->lo_header, nd->ld_type);
1210 stripe[i] = container_of(n, struct dt_object, do_lu);
1211 ost_indices[i] = idx;
1216 for (i = 0; i < stripe_len; i++)
1217 if (stripe[i] != NULL)
1218 dt_object_put(env, stripe[i]);
1220 OBD_FREE_PTR_ARRAY(stripe, stripe_len);
1221 lod_comp->llc_stripe_count = 0;
1223 OBD_FREE_PTR_ARRAY(ost_indices, stripe_len);
1225 lod_comp->llc_stripe = stripe;
1226 lod_comp->llc_ost_indices = ost_indices;
1227 lod_comp->llc_stripes_allocated = stripe_len;
1233 int lod_init_comp_foreign(struct lod_layout_component *lod_comp, void *lmm)
1235 struct lov_foreign_md *lfm;
1237 lfm = (struct lov_foreign_md *)lmm;
1238 lod_comp->llc_length = le32_to_cpu(lfm->lfm_length);
1239 lod_comp->llc_type = le32_to_cpu(lfm->lfm_type);
1241 if (!lov_hsm_type_supported(lod_comp->llc_type)) {
1243 "Unsupport HSM type: %u length: %u flags: %08X\n",
1244 lod_comp->llc_type, lod_comp->llc_length,
1245 le32_to_cpu(lfm->lfm_flags));
1250 * Currently it only stores the file FID as the field @lhm_archive_uuid
1251 * which is used to be the identifier within HSM backend for the archive
1253 * Thus the length of foreign layout value (HSM is a kind of foreign
1254 * layout type) is: sizeof(lhm_archive_id) + sizeof(lhm_archive_ver) +
1256 * It should fix to support other kinds of identifier for different HSM
1257 * solutions such as S3.
1259 if (lod_comp->llc_length != sizeof(struct lov_hsm_base)) {
1260 CDEBUG(D_LAYOUT, "Invalid HSM len: %u, should be %zu\n",
1261 lod_comp->llc_length, sizeof(struct lov_hsm_base));
1265 lod_comp->llc_foreign_flags = le32_to_cpu(lfm->lfm_flags);
1266 lov_foreign_hsm_to_cpu(&lod_comp->llc_hsm, lfm);
1271 * Instantiate objects for striping.
1273 * Parse striping information in \a buf and instantiate the objects
1274 * representing the stripes.
1276 * \param[in] env execution environment for this thread
1277 * \param[in] lo LOD object
1278 * \param[in] buf buffer storing LOV EA to parse
1279 * \param[in] lvf verify flags when parsing the layout
1281 * \retval 0 if parsing and objects creation succeed
1282 * \retval negative error number on failure
1284 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
1285 const struct lu_buf *buf, enum layout_verify_flags lvf)
1287 struct lod_device *d = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1288 struct lov_mds_md_v1 *lmm;
1289 struct lov_comp_md_v1 *comp_v1 = NULL;
1290 struct lov_foreign_md *foreign = NULL;
1291 struct lov_ost_data_v1 *objs;
1292 __u32 magic, pattern;
1293 __u16 mirror_cnt = 0;
1297 __u16 mirror_id = MIRROR_ID_NEG;
1299 int stale_mirrors = 0;
1303 LASSERT(buf->lb_buf);
1304 LASSERT(buf->lb_len);
1305 LASSERT(mutex_is_locked(&lo->ldo_layout_mutex));
1307 lmm = (struct lov_mds_md_v1 *)buf->lb_buf;
1308 magic = le32_to_cpu(lmm->lmm_magic);
1310 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1311 magic != LOV_MAGIC_COMP_V1 && magic != LOV_MAGIC_FOREIGN &&
1312 magic != LOV_MAGIC_SEL)
1313 GOTO(out, rc = -EINVAL);
1315 lod_striping_free_nolock(env, lo);
1317 if (magic == LOV_MAGIC_COMP_V1 || magic == LOV_MAGIC_SEL) {
1318 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1319 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1321 GOTO(out, rc = -EINVAL);
1323 lo->ldo_layout_gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1324 lo->ldo_is_composite = 1;
1325 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1327 lo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1330 lo->ldo_flr_state = LCM_FL_NONE;
1331 } else if (magic == LOV_MAGIC_FOREIGN) {
1334 foreign = (struct lov_foreign_md *)buf->lb_buf;
1335 length = offsetof(typeof(*foreign), lfm_value);
1336 if (buf->lb_len < length ||
1337 buf->lb_len < (length + le32_to_cpu(foreign->lfm_length))) {
1339 "buf len %zu too small for lov_foreign_md\n",
1341 GOTO(out, rc = -EINVAL);
1344 /* just cache foreign LOV EA raw */
1345 rc = lod_alloc_foreign_lov(lo, length);
1348 memcpy(lo->ldo_foreign_lov, buf->lb_buf, length);
1352 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1353 lo->ldo_is_composite = 0;
1356 rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
1360 for (i = 0; i < comp_cnt; i++) {
1361 struct lod_layout_component *lod_comp;
1362 struct lu_extent *ext;
1365 lod_comp = &lo->ldo_comp_entries[i];
1366 if (lo->ldo_is_composite) {
1367 offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1368 lmm = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1370 ext = &comp_v1->lcm_entries[i].lcme_extent;
1371 lod_comp->llc_extent.e_start =
1372 le64_to_cpu(ext->e_start);
1373 if (lod_comp->llc_extent.e_start &
1374 (LOV_MIN_STRIPE_SIZE - 1)) {
1376 "extent start %llu is not a multiple of min size %u\n",
1377 lod_comp->llc_extent.e_start,
1378 LOV_MIN_STRIPE_SIZE);
1379 GOTO(out, rc = -EINVAL);
1382 lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1383 if (lod_comp->llc_extent.e_end != LUSTRE_EOF &&
1384 lod_comp->llc_extent.e_end &
1385 (LOV_MIN_STRIPE_SIZE - 1)) {
1387 "extent end %llu is not a multiple of min size %u\n",
1388 lod_comp->llc_extent.e_end,
1389 LOV_MIN_STRIPE_SIZE);
1390 GOTO(out, rc = -EINVAL);
1393 lod_comp->llc_flags =
1394 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1396 if (lod_comp->llc_flags & LCME_FL_NOSYNC)
1397 lod_comp->llc_timestamp = le64_to_cpu(
1398 comp_v1->lcm_entries[i].lcme_timestamp);
1400 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1401 if (lod_comp->llc_id == LCME_ID_INVAL)
1402 GOTO(out, rc = -EINVAL);
1404 if (lvf & LVF_ALL_STALE) {
1405 if (mirror_id_of(lod_comp->llc_id) ==
1407 /* remaining comps in the mirror */
1408 stale |= lod_comp->llc_flags &
1412 * new mirror, check last mirror's
1419 mirror_id_of(lod_comp->llc_id);
1421 /* the first comp of the new mirror */
1422 stale = lod_comp->llc_flags &
1427 if ((lod_comp->llc_flags & LCME_FL_EXTENSION) &&
1428 comp_v1->lcm_magic != cpu_to_le32(LOV_MAGIC_SEL)) {
1429 CWARN("%s: EXTENSION flags=%x set on component[%u]=%x of non-SEL file "DFID" with magic=%#08x\n",
1430 lod2obd(d)->obd_name,
1431 lod_comp->llc_flags, lod_comp->llc_id, i,
1432 PFID(lod_object_fid(lo)),
1433 le32_to_cpu(comp_v1->lcm_magic));
1436 lod_comp->llc_magic = le32_to_cpu(lmm->lmm_magic);
1437 if (lod_comp->llc_magic == LOV_MAGIC_FOREIGN) {
1438 rc = lod_init_comp_foreign(lod_comp, lmm);
1444 lod_comp->llc_magic = le32_to_cpu(lmm->lmm_magic);
1445 lod_comp_set_init(lod_comp);
1448 pattern = le32_to_cpu(lmm->lmm_pattern);
1449 if (!lov_pattern_supported(lov_pattern(pattern)))
1450 GOTO(out, rc = -EINVAL);
1452 if (pattern & LOV_PATTERN_MDT) {
1453 if (lod_comp->llc_extent.e_start != 0) {
1454 CERROR("%s: DOM entry must be the first stripe "
1455 "in a mirror\n", lod2obd(d)->obd_name);
1456 GOTO(out, rc = -EINVAL);
1459 dom_size = lod_comp->llc_extent.e_end;
1460 } else if (dom_size != lod_comp->llc_extent.e_end) {
1461 CERROR("%s: DOM entries with different sizes "
1462 "%#llx/%#llx\n", lod2obd(d)->obd_name,
1463 dom_size, lod_comp->llc_extent.e_end);
1464 GOTO(out, rc = -EINVAL);
1468 lod_comp->llc_pattern = pattern;
1469 lod_comp->llc_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
1470 lod_comp->llc_stripe_count = le16_to_cpu(lmm->lmm_stripe_count);
1471 lod_comp->llc_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1473 if (lmm->lmm_magic == cpu_to_le32(LOV_MAGIC_V3)) {
1474 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
1476 lod_set_pool(&lod_comp->llc_pool, v3->lmm_pool_name);
1477 objs = &v3->lmm_objects[0];
1479 lod_set_pool(&lod_comp->llc_pool, NULL);
1480 objs = &lmm->lmm_objects[0];
1484 * If uninstantiated template component has valid l_ost_idx,
1485 * then user has specified ost list for this component.
1487 if (!lod_comp_inited(lod_comp)) {
1490 if (objs[0].l_ost_idx != (__u32)-1UL) {
1493 stripe_count = lod_comp_entry_stripe_count(
1495 if (stripe_count == 0 &&
1496 !(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1497 !(lod_comp->llc_pattern & LOV_PATTERN_MDT))
1498 GOTO(out, rc = -E2BIG);
1500 * load the user specified ost list, when this
1501 * component is instantiated later, it will be
1502 * used in lod_alloc_ost_list().
1504 lod_comp->llc_ostlist.op_count = stripe_count;
1505 lod_comp->llc_ostlist.op_size =
1506 stripe_count * sizeof(__u32);
1507 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1508 lod_comp->llc_ostlist.op_size);
1509 if (!lod_comp->llc_ostlist.op_array)
1510 GOTO(out, rc = -ENOMEM);
1512 for (j = 0; j < stripe_count; j++)
1513 lod_comp->llc_ostlist.op_array[j] =
1514 le32_to_cpu(objs[j].l_ost_idx);
1517 * this component OST objects starts from the
1518 * first ost_idx, lod_alloc_ost_list() will
1521 lod_comp->llc_stripe_offset = objs[0].l_ost_idx;
1524 * for uninstantiated component,
1525 * lmm_layout_gen stores default stripe offset.
1527 lod_comp->llc_stripe_offset =
1528 lmm->lmm_layout_gen;
1532 /* skip un-instantiated component object initialization */
1533 if (!lod_comp_inited(lod_comp))
1536 if (!(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1537 !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1538 rc = lod_initialize_objects(env, lo, objs, i);
1544 if (lo->ldo_is_composite && (lvf & LVF_ALL_STALE)) {
1545 /* check the last mirror stale-ness */
1549 if (mirror_cnt == stale_mirrors) {
1551 CERROR("%s: can not set all stale mirrors for "
1553 lod2obd(d)->obd_name, PFID(lod_object_fid(lo)),
1559 rc = lod_fill_mirrors(lo);
1565 lod_striping_free_nolock(env, lo);
1570 * Check whether the striping (LOVEA for regular file, LMVEA for directory)
1571 * is already cached.
1573 * \param[in] lo LOD object
1575 * \retval True if the striping is cached, otherwise
1578 static bool lod_striping_loaded(struct lod_object *lo)
1580 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1581 lo->ldo_comp_cached)
1584 if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1585 if (lo->ldo_dir_stripe_loaded)
1588 /* Never load LMV stripe for slaves of striped dir */
1589 if (lo->ldo_dir_slave_stripe)
1597 * A generic function to initialize the stripe objects.
1599 * A protected version of lod_striping_load_locked() - load the striping
1600 * information from storage, parse that and instantiate LU objects to
1601 * represent the stripes. The LOD object \a lo supplies a pointer to the
1602 * next sub-object in the LU stack so we can lock it. Also use \a lo to
1603 * return an array of references to the newly instantiated objects.
1605 * \param[in] env execution environment for this thread
1606 * \param[in,out] lo LOD object, where striping is stored and
1607 * which gets an array of references
1609 * \retval 0 if parsing and object creation succeed
1610 * \retval negative error number on failure
1612 int lod_striping_load(const struct lu_env *env, struct lod_object *lo)
1614 struct lod_thread_info *info = lod_env_info(env);
1615 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1616 struct lu_buf *buf = &info->lti_buf;
1621 if (!dt_object_exists(next))
1624 if (lod_striping_loaded(lo))
1627 mutex_lock(&lo->ldo_layout_mutex);
1628 if (lod_striping_loaded(lo))
1629 GOTO(unlock, rc = 0);
1631 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr)) {
1632 rc = lod_get_lov_ea(env, lo);
1637 * there is LOV EA (striping information) in this object
1638 * let's parse it and create in-core objects for the stripes
1640 buf->lb_buf = info->lti_ea_store;
1641 buf->lb_len = info->lti_ea_store_size;
1642 rc = lod_parse_striping(env, lo, buf, 0);
1644 lo->ldo_comp_cached = 1;
1645 } else if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1646 rc = lod_get_lmv_ea(env, lo);
1648 if (rc > (int)sizeof(struct lmv_foreign_md)) {
1649 struct lmv_foreign_md *lfm = info->lti_ea_store;
1651 if (le32_to_cpu(lfm->lfm_magic) == LMV_MAGIC_FOREIGN) {
1652 lo->ldo_foreign_lmv = info->lti_ea_store;
1653 lo->ldo_foreign_lmv_size =
1654 info->lti_ea_store_size;
1655 info->lti_ea_store = NULL;
1656 info->lti_ea_store_size = 0;
1658 lo->ldo_dir_stripe_loaded = 1;
1659 lo->ldo_is_foreign = 1;
1660 GOTO(unlock, rc = 0);
1664 if (rc < (int)sizeof(struct lmv_mds_md_v1)) {
1665 /* Let's set stripe_loaded to avoid further
1666 * stripe loading especially for non-stripe directory,
1667 * which can hurt performance. (See LU-9840)
1670 lo->ldo_dir_stripe_loaded = 1;
1671 GOTO(unlock, rc = rc > 0 ? -EINVAL : rc);
1673 buf->lb_buf = info->lti_ea_store;
1674 buf->lb_len = info->lti_ea_store_size;
1675 if (rc == sizeof(struct lmv_mds_md_v1)) {
1676 rc = lod_load_lmv_shards(env, lo, buf, true);
1677 if (buf->lb_buf != info->lti_ea_store) {
1678 OBD_FREE_LARGE(info->lti_ea_store,
1679 info->lti_ea_store_size);
1680 info->lti_ea_store = buf->lb_buf;
1681 info->lti_ea_store_size = buf->lb_len;
1689 * there is LMV EA (striping information) in this object
1690 * let's parse it and create in-core objects for the stripes
1692 rc = lod_parse_dir_striping(env, lo, buf);
1694 lo->ldo_dir_stripe_loaded = 1;
1698 mutex_unlock(&lo->ldo_layout_mutex);
1703 int lod_striping_reload(const struct lu_env *env, struct lod_object *lo,
1704 const struct lu_buf *buf, enum layout_verify_flags lvf)
1710 mutex_lock(&lo->ldo_layout_mutex);
1711 rc = lod_parse_striping(env, lo, buf, lvf);
1712 mutex_unlock(&lo->ldo_layout_mutex);
1718 * Verify lov_user_md_v1/v3 striping.
1720 * Check the validity of all fields including the magic, stripe size,
1721 * stripe count, stripe offset and that the pool is present. Also check
1722 * that each target index points to an existing target. The additional
1723 * \a is_from_disk turns additional checks. In some cases zero fields
1724 * are allowed (like pattern=0).
1726 * \param[in] d LOD device
1727 * \param[in] buf buffer with LOV EA to verify
1728 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
1729 * 1 - from disk, do not allow
1731 * \retval 0 if the striping is valid
1732 * \retval -EINVAL if striping is invalid
1734 static int lod_verify_v1v3(struct lod_device *d, const struct lu_buf *buf,
1737 struct lov_user_md_v1 *lum;
1738 struct lov_user_md_v3 *lum3;
1739 struct lod_pool_desc *pool = NULL;
1743 __u16 stripe_offset;
1750 if (buf->lb_len < sizeof(*lum)) {
1751 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1753 GOTO(out, rc = -EINVAL);
1756 magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1757 if (magic != LOV_USER_MAGIC_V1 &&
1758 magic != LOV_USER_MAGIC_V3 &&
1759 magic != LOV_USER_MAGIC_SPECIFIC) {
1760 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1761 le32_to_cpu(lum->lmm_magic));
1762 GOTO(out, rc = -EINVAL);
1765 /* the user uses "0" for default stripe pattern normally. */
1766 if (!is_from_disk && lum->lmm_pattern == LOV_PATTERN_NONE)
1767 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1769 if (!lov_pattern_supported(le32_to_cpu(lum->lmm_pattern))) {
1770 CDEBUG(D_LAYOUT, "bad userland stripe pattern: %#x\n",
1771 le32_to_cpu(lum->lmm_pattern));
1772 GOTO(out, rc = -EINVAL);
1775 /* a released lum comes from creating orphan on hsm release,
1776 * doesn't make sense to verify it. */
1777 if (le32_to_cpu(lum->lmm_pattern) & LOV_PATTERN_F_RELEASED)
1780 /* 64kB is the largest common page size we see (ia64), and matches the
1782 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1783 if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1784 CDEBUG(D_LAYOUT, "stripe size %u not a multiple of %u\n",
1785 stripe_size, LOV_MIN_STRIPE_SIZE);
1786 GOTO(out, rc = -EINVAL);
1789 stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1790 if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT &&
1791 !(lov_pattern(le32_to_cpu(lum->lmm_pattern)) & LOV_PATTERN_MDT)) {
1792 /* if offset is not within valid range [0, osts_size) */
1793 if (stripe_offset >= d->lod_ost_descs.ltd_tgts_size) {
1794 CDEBUG(D_LAYOUT, "stripe offset %u >= bitmap size %u\n",
1795 stripe_offset, d->lod_ost_descs.ltd_tgts_size);
1796 GOTO(out, rc = -EINVAL);
1799 /* if lmm_stripe_offset is *not* in bitmap */
1800 if (!test_bit(stripe_offset, d->lod_ost_bitmap)) {
1801 CDEBUG(D_LAYOUT, "stripe offset %u not in bitmap\n",
1803 GOTO(out, rc = -EINVAL);
1807 if (magic == LOV_USER_MAGIC_V1)
1808 lum_size = offsetof(struct lov_user_md_v1,
1810 else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC)
1811 lum_size = offsetof(struct lov_user_md_v3,
1814 GOTO(out, rc = -EINVAL);
1816 stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1817 if (buf->lb_len < lum_size) {
1818 CDEBUG(D_LAYOUT, "invalid buf len %zu/%zu for lov_user_md with "
1819 "magic %#x and stripe_count %u\n",
1820 buf->lb_len, lum_size, magic, stripe_count);
1821 GOTO(out, rc = -EINVAL);
1824 if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC))
1828 /* In the function below, .hs_keycmp resolves to
1829 * pool_hashkey_keycmp() */
1830 pool = lod_find_pool(d, lum3->lmm_pool_name);
1834 if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT) {
1835 rc = lod_check_index_in_pool(stripe_offset, pool);
1837 GOTO(out, rc = -EINVAL);
1840 if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1841 CDEBUG(D_LAYOUT, "stripe count %u > # OSTs %u in the pool\n",
1842 stripe_count, pool_tgt_count(pool));
1843 GOTO(out, rc = -EINVAL);
1848 lod_pool_putref(pool);
1854 struct lov_comp_md_entry_v1 *comp_entry_v1(struct lov_comp_md_v1 *comp, int i)
1856 LASSERTF((le32_to_cpu(comp->lcm_magic) & ~LOV_MAGIC_DEFINED) ==
1857 LOV_USER_MAGIC_COMP_V1 ||
1858 (le32_to_cpu(comp->lcm_magic) & ~LOV_MAGIC_DEFINED) ==
1859 LOV_USER_MAGIC_SEL, "Wrong magic %x\n",
1860 le32_to_cpu(comp->lcm_magic));
1861 LASSERTF(i >= 0 && i < le16_to_cpu(comp->lcm_entry_count),
1862 "bad index %d, max = %d\n",
1863 i, le16_to_cpu(comp->lcm_entry_count));
1865 return &comp->lcm_entries[i];
1868 #define for_each_comp_entry_v1(comp, entry) \
1869 for (entry = comp_entry_v1(comp, 0); \
1870 entry <= comp_entry_v1(comp, \
1871 le16_to_cpu(comp->lcm_entry_count) - 1); \
1874 static int lod_erase_dom_stripe(struct lov_comp_md_v1 *comp_v1,
1875 struct lov_comp_md_entry_v1 *dom_ent)
1877 struct lov_comp_md_entry_v1 *ent;
1879 __u32 dom_off, dom_size, comp_size, off;
1881 unsigned int size, shift;
1883 entries = le16_to_cpu(comp_v1->lcm_entry_count) - 1;
1884 LASSERT(entries > 0);
1885 comp_v1->lcm_entry_count = cpu_to_le16(entries);
1887 comp_size = le32_to_cpu(comp_v1->lcm_size);
1888 dom_off = le32_to_cpu(dom_ent->lcme_offset);
1889 dom_size = le32_to_cpu(dom_ent->lcme_size);
1891 /* all entries offsets are shifted by entry size at least */
1892 shift = sizeof(*dom_ent);
1893 for_each_comp_entry_v1(comp_v1, ent) {
1894 off = le32_to_cpu(ent->lcme_offset);
1895 if (off == dom_off) {
1896 /* Entry deletion creates two holes in layout data:
1897 * - hole in entries array
1898 * - hole in layout data at dom_off with dom_size
1900 * First memmove is one entry shift from next entry
1901 * start with size up to dom_off in blob
1904 src = (void *)(ent + 1);
1905 size = (unsigned long)((void *)comp_v1 + dom_off - src);
1906 memmove(dst, src, size);
1907 /* take 'off' from just moved entry */
1908 off = le32_to_cpu(ent->lcme_offset);
1909 /* second memmove is blob tail after 'off' up to
1912 dst = (void *)comp_v1 + dom_off - sizeof(*ent);
1913 src = (void *)comp_v1 + off;
1914 size = (unsigned long)(comp_size - off);
1915 memmove(dst, src, size);
1916 /* all entries offsets after DoM entry are shifted by
1917 * dom_size additionally
1921 ent->lcme_offset = cpu_to_le32(off - shift);
1923 comp_v1->lcm_size = cpu_to_le32(comp_size - shift);
1925 /* notify a caller to re-check entry */
1929 void lod_dom_stripesize_recalc(struct lod_device *d)
1931 __u64 threshold_mb = d->lod_dom_threshold_free_mb;
1932 __u32 max_size = d->lod_dom_stripesize_max_kb;
1933 __u32 def_size = d->lod_dom_stripesize_cur_kb;
1935 /* use maximum allowed value if free space is above threshold */
1936 if (d->lod_lsfs_free_mb >= threshold_mb) {
1937 def_size = max_size;
1938 } else if (!d->lod_lsfs_free_mb || max_size <= LOD_DOM_MIN_SIZE_KB) {
1941 /* recalc threshold like it would be with def_size as max */
1942 threshold_mb = mult_frac(threshold_mb, def_size, max_size);
1943 if (d->lod_lsfs_free_mb < threshold_mb)
1944 def_size = rounddown(def_size / 2, LOD_DOM_MIN_SIZE_KB);
1945 else if (d->lod_lsfs_free_mb > threshold_mb * 2)
1946 def_size = max_t(unsigned int, def_size * 2,
1947 LOD_DOM_MIN_SIZE_KB);
1950 if (d->lod_dom_stripesize_cur_kb != def_size) {
1951 CDEBUG(D_LAYOUT, "Change default DOM stripe size %d->%d\n",
1952 d->lod_dom_stripesize_cur_kb, def_size);
1953 d->lod_dom_stripesize_cur_kb = def_size;
1957 static __u32 lod_dom_stripesize_limit(const struct lu_env *env,
1958 struct lod_device *d)
1962 /* set bfree as fraction of total space */
1963 if (CFS_FAIL_CHECK(OBD_FAIL_MDS_STATFS_SPOOF)) {
1964 spin_lock(&d->lod_lsfs_lock);
1965 d->lod_lsfs_free_mb = mult_frac(d->lod_lsfs_total_mb,
1966 min_t(int, cfs_fail_val, 100), 100);
1967 GOTO(recalc, rc = 0);
1970 if (d->lod_lsfs_age < ktime_get_seconds() - LOD_DOM_SFS_MAX_AGE) {
1971 struct obd_statfs sfs;
1973 spin_lock(&d->lod_lsfs_lock);
1974 if (d->lod_lsfs_age > ktime_get_seconds() - LOD_DOM_SFS_MAX_AGE)
1975 GOTO(unlock, rc = 0);
1977 d->lod_lsfs_age = ktime_get_seconds();
1978 spin_unlock(&d->lod_lsfs_lock);
1979 rc = dt_statfs(env, d->lod_child, &sfs);
1982 "%s: failed to get OSD statfs: rc = %d\n",
1983 lod2obd(d)->obd_name, rc);
1986 /* udpate local OSD cached statfs data */
1987 spin_lock(&d->lod_lsfs_lock);
1988 d->lod_lsfs_total_mb = (sfs.os_blocks * sfs.os_bsize) >> 20;
1989 d->lod_lsfs_free_mb = (sfs.os_bfree * sfs.os_bsize) >> 20;
1991 lod_dom_stripesize_recalc(d);
1993 spin_unlock(&d->lod_lsfs_lock);
1996 return d->lod_dom_stripesize_cur_kb << 10;
1999 static int lod_dom_stripesize_choose(const struct lu_env *env,
2000 struct lod_device *d,
2001 struct lov_comp_md_v1 *comp_v1,
2002 struct lov_comp_md_entry_v1 *dom_ent,
2005 struct lov_comp_md_entry_v1 *ent;
2006 struct lu_extent *dom_ext, *ext;
2007 struct lov_user_md_v1 *lum;
2008 __u32 max_stripe_size;
2011 bool dom_next_entry = false;
2013 dom_ext = &dom_ent->lcme_extent;
2014 dom_mid = mirror_id_of(le32_to_cpu(dom_ent->lcme_id));
2015 max_stripe_size = lod_dom_stripesize_limit(env, d);
2017 /* Check stripe size againts current per-MDT limit */
2018 if (stripe_size <= max_stripe_size)
2021 lum = (void *)comp_v1 + le32_to_cpu(dom_ent->lcme_offset);
2022 CDEBUG(D_LAYOUT, "overwrite DoM component size %u with MDT limit %u\n",
2023 stripe_size, max_stripe_size);
2024 lum->lmm_stripe_size = cpu_to_le32(max_stripe_size);
2026 /* In common case the DoM stripe is first entry in a mirror and
2027 * can be deleted only if it is not single entry in layout or
2028 * mirror, otherwise error should be returned.
2030 for_each_comp_entry_v1(comp_v1, ent) {
2034 mid = mirror_id_of(le32_to_cpu(ent->lcme_id));
2038 ext = &ent->lcme_extent;
2039 if (ext->e_start != dom_ext->e_end)
2042 /* Found next component after the DoM one with the same
2043 * mirror_id and adjust its start with DoM component end.
2045 * NOTE: we are considering here that there can be only one
2046 * DoM component in a file, all replicas are located on OSTs
2047 * always and don't need adjustment since use own layouts.
2049 ext->e_start = cpu_to_le64(max_stripe_size);
2050 dom_next_entry = true;
2054 if (max_stripe_size == 0) {
2055 /* DoM component size is zero due to server setting, remove
2056 * it from the layout but only if next component exists in
2057 * the same mirror. That must be checked prior calling the
2058 * lod_erase_dom_stripe().
2060 if (!dom_next_entry)
2063 rc = lod_erase_dom_stripe(comp_v1, dom_ent);
2065 /* Update DoM extent end finally */
2066 dom_ext->e_end = cpu_to_le64(max_stripe_size);
2073 * Verify LOV striping.
2075 * \param[in] d LOD device
2076 * \param[in] buf buffer with LOV EA to verify
2077 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
2078 * 1 - from disk, do not allow
2079 * \param[in] start extent start for composite layout
2081 * \retval 0 if the striping is valid
2082 * \retval -EINVAL if striping is invalid
2084 int lod_verify_striping(const struct lu_env *env, struct lod_device *d,
2085 struct lod_object *lo, const struct lu_buf *buf,
2088 struct lov_user_md_v1 *lum;
2089 struct lov_comp_md_v1 *comp_v1;
2090 struct lov_comp_md_entry_v1 *ent;
2091 struct lu_extent *ext;
2094 __u32 stripe_size = 0;
2095 __u16 prev_mid = -1, mirror_id = -1;
2101 if (buf->lb_len < sizeof(lum->lmm_magic)) {
2102 CDEBUG(D_LAYOUT, "invalid buf len %zu\n", buf->lb_len);
2108 magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
2109 /* treat foreign LOV EA/object case first
2110 * XXX is it expected to try setting again a foreign?
2111 * XXX should we care about different current vs new layouts ?
2113 if (unlikely(magic == LOV_USER_MAGIC_FOREIGN)) {
2114 struct lov_foreign_md *lfm = buf->lb_buf;
2116 if (buf->lb_len < offsetof(typeof(*lfm), lfm_value)) {
2118 "buf len %zu < min lov_foreign_md size (%zu)\n",
2119 buf->lb_len, offsetof(typeof(*lfm),
2124 if (lov_foreign_size_le(lfm) > buf->lb_len) {
2126 "buf len %zu < this lov_foreign_md size (%zu)\n",
2127 buf->lb_len, lov_foreign_size_le(lfm));
2130 /* Don't do anything with foreign layouts */
2134 /* normal LOV/layout cases */
2136 if (buf->lb_len < sizeof(*lum)) {
2137 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
2143 case LOV_USER_MAGIC_FOREIGN:
2145 case LOV_USER_MAGIC_V1:
2146 case LOV_USER_MAGIC_V3:
2147 case LOV_USER_MAGIC_SPECIFIC:
2148 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) &
2150 /* DoM must use composite layout */
2151 CDEBUG(D_LAYOUT, "DoM without composite layout\n");
2154 RETURN(lod_verify_v1v3(d, buf, is_from_disk));
2155 case LOV_USER_MAGIC_COMP_V1:
2156 case LOV_USER_MAGIC_SEL:
2159 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
2160 le32_to_cpu(lum->lmm_magic));
2164 /* magic == LOV_USER_MAGIC_COMP_V1 */
2165 comp_v1 = buf->lb_buf;
2166 if (buf->lb_len < le32_to_cpu(comp_v1->lcm_size)) {
2167 CDEBUG(D_LAYOUT, "buf len %zu is less than %u\n",
2168 buf->lb_len, le32_to_cpu(comp_v1->lcm_size));
2174 if (le16_to_cpu(comp_v1->lcm_entry_count) == 0) {
2175 CDEBUG(D_LAYOUT, "entry count is zero\n");
2179 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
2180 lo->ldo_comp_cnt > 0) {
2181 /* could be called from lustre.lov.add */
2182 __u32 cnt = lo->ldo_comp_cnt;
2184 ext = &lo->ldo_comp_entries[cnt - 1].llc_extent;
2185 prev_end = ext->e_end;
2190 for_each_comp_entry_v1(comp_v1, ent) {
2191 ext = &ent->lcme_extent;
2193 if (le64_to_cpu(ext->e_start) > le64_to_cpu(ext->e_end) ||
2194 le64_to_cpu(ext->e_start) & (LOV_MIN_STRIPE_SIZE - 1) ||
2195 ((__s64)le64_to_cpu(ext->e_start) < 0 &&
2196 le64_to_cpu(ext->e_start) != LUSTRE_EOF) ||
2197 (le64_to_cpu(ext->e_end) != LUSTRE_EOF &&
2198 le64_to_cpu(ext->e_end) & (LOV_MIN_STRIPE_SIZE - 1)) ||
2199 ((__s64)le64_to_cpu(ext->e_end) < 0 &&
2200 le64_to_cpu(ext->e_end) != LUSTRE_EOF)) {
2201 CDEBUG(D_LAYOUT, "invalid extent "DEXT"\n",
2202 le64_to_cpu(ext->e_start),
2203 le64_to_cpu(ext->e_end));
2208 /* lcme_id contains valid value */
2209 if (le32_to_cpu(ent->lcme_id) == 0 ||
2210 le32_to_cpu(ent->lcme_id) > LCME_ID_MAX) {
2211 CDEBUG(D_LAYOUT, "invalid id %u\n",
2212 le32_to_cpu(ent->lcme_id));
2216 if (le16_to_cpu(comp_v1->lcm_mirror_count) > 0) {
2217 mirror_id = mirror_id_of(
2218 le32_to_cpu(ent->lcme_id));
2220 /* first component must start with 0 */
2221 if (mirror_id != prev_mid &&
2222 le64_to_cpu(ext->e_start) != 0) {
2224 "invalid start:%llu, expect:0\n",
2225 le64_to_cpu(ext->e_start));
2229 prev_mid = mirror_id;
2233 if (le64_to_cpu(ext->e_start) == 0) {
2238 /* the next must be adjacent with the previous one */
2239 if (le64_to_cpu(ext->e_start) != prev_end) {
2241 "invalid start actual:%llu, expect:%llu\n",
2242 le64_to_cpu(ext->e_start), prev_end);
2246 tmp.lb_buf = (char *)comp_v1 + le32_to_cpu(ent->lcme_offset);
2247 tmp.lb_len = le32_to_cpu(ent->lcme_size);
2250 if (le32_to_cpu(lum->lmm_magic) == LOV_MAGIC_FOREIGN) {
2251 struct lov_foreign_md *lfm;
2252 struct lov_hsm_md *lhm;
2257 * Currently when the foreign layout is used as a basic
2258 * layout component, it only supports HSM foreign types:
2259 * LU_FOREIGN_TYPE_{POSIX, S3, PCCRW, PCCRO}.
2261 lfm = (struct lov_foreign_md *)lum;
2262 ftype = le32_to_cpu(lfm->lfm_type);
2263 if (!lov_hsm_type_supported(ftype)) {
2265 "Foreign type %#x is not HSM\n", ftype);
2269 /* Current HSM component must cover [0, EOF]. */
2270 if (le64_to_cpu(ext->e_start) > 0) {
2271 CDEBUG(D_LAYOUT, "Invalid HSM component with %llu extent start\n",
2272 le64_to_cpu(ext->e_start));
2275 if (le64_to_cpu(ext->e_end) != LUSTRE_EOF) {
2276 CDEBUG(D_LAYOUT, "Invalid HSM component with %llu extent end\n",
2277 le64_to_cpu(ext->e_end));
2281 lhm = (struct lov_hsm_md *)lfm;
2282 hsmsize = lov_foreign_size_le(lhm);
2284 if (le32_to_cpu(lhm->lhm_length) !=
2285 sizeof(struct lov_hsm_base)) {
2287 "Invalid HSM component size %u != %u\n",
2288 le32_to_cpu(ent->lcme_size), hsmsize);
2292 if (le32_to_cpu(ent->lcme_size) < hsmsize) {
2294 "Invalid HSM component size %u != %u\n",
2295 le32_to_cpu(ent->lcme_size), hsmsize);
2298 if (le32_to_cpu(lhm->lhm_flags) & ~HSM_FLAGS_MASK ||
2299 !(le32_to_cpu(lhm->lhm_flags) & HSM_FLAGS_MASK)) {
2301 "Invalid HSM component flags %#x\n",
2302 le32_to_cpu(lhm->lhm_flags));
2308 /* Check DoM entry is always the first one */
2309 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) &
2311 /* DoM component must be the first in a mirror */
2312 if (le64_to_cpu(ext->e_start) > 0) {
2313 CDEBUG(D_LAYOUT, "invalid DoM component "
2314 "with %llu extent start\n",
2315 le64_to_cpu(ext->e_start));
2318 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
2319 /* There is just one stripe on MDT and it must
2320 * cover whole component size. */
2321 if (stripe_size != le64_to_cpu(ext->e_end)) {
2322 CDEBUG(D_LAYOUT, "invalid DoM layout "
2323 "stripe size %u != %llu "
2324 "(component size)\n",
2325 stripe_size, prev_end);
2328 /* Check and adjust stripe size by per-MDT limit */
2329 rc = lod_dom_stripesize_choose(env, d, comp_v1, ent,
2331 /* DoM entry was removed, re-check layout from start */
2332 if (rc == -ERESTART)
2337 if (le16_to_cpu(lum->lmm_stripe_count) == 1)
2338 lum->lmm_stripe_count = 0;
2339 /* Any stripe count is forbidden on DoM component */
2340 if (lum->lmm_stripe_count > 0) {
2342 "invalid DoM layout stripe count %u, must be 0\n",
2343 le16_to_cpu(lum->lmm_stripe_count));
2347 /* Any pool is forbidden on DoM component */
2348 if (lum->lmm_magic == LOV_USER_MAGIC_V3) {
2349 struct lov_user_md_v3 *v3 = (void *)lum;
2351 if (v3->lmm_pool_name[0] != '\0') {
2353 "DoM component cannot have pool assigned\n");
2359 prev_end = le64_to_cpu(ext->e_end);
2361 rc = lod_verify_v1v3(d, &tmp, is_from_disk);
2365 if (prev_end == LUSTRE_EOF || ext->e_start == prev_end)
2368 /* extent end must be aligned with the stripe_size */
2369 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
2370 if (stripe_size && prev_end % stripe_size) {
2371 CDEBUG(D_LAYOUT, "stripe size isn't aligned, "
2372 "stripe_sz: %u, [%llu, %llu)\n",
2373 stripe_size, ext->e_start, prev_end);
2378 /* make sure that the mirror_count is telling the truth */
2379 if (mirror_count != le16_to_cpu(comp_v1->lcm_mirror_count) + 1)
2386 * set the default stripe size, if unset.
2388 * \param[in,out] val number of bytes per OST stripe
2390 * The minimum stripe size is 64KB to ensure that a single stripe is an
2391 * even multiple of a client PAGE_SIZE (IA64, PPC, etc). Otherwise, it
2392 * is difficult to split dirty pages across OSCs during writes.
2394 void lod_fix_desc_stripe_size(__u64 *val)
2396 if (*val < LOV_MIN_STRIPE_SIZE) {
2398 LCONSOLE_INFO("Increasing default stripe size to "
2399 "minimum value %u\n",
2400 LOV_DESC_STRIPE_SIZE_DEFAULT);
2401 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
2402 } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
2403 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
2404 LCONSOLE_WARN("Changing default stripe size to %llu (a "
2405 "multiple of %u)\n",
2406 *val, LOV_MIN_STRIPE_SIZE);
2411 * set the filesystem default number of stripes, if unset.
2413 * \param[in,out] val number of stripes
2415 * A value of "0" means "use the system-wide default stripe count", which
2416 * has either been inherited by now, or falls back to 1 stripe per file.
2417 * A value of "-1" (0xffffffff) means "stripe over all available OSTs",
2418 * and is a valid value, so is left unchanged here.
2420 void lod_fix_desc_stripe_count(__u32 *val)
2427 * set the filesystem default layout pattern
2429 * \param[in,out] val LOV_PATTERN_* layout
2431 * A value of "0" means "use the system-wide default layout type", which
2432 * has either been inherited by now, or falls back to plain RAID0 striping.
2434 void lod_fix_desc_pattern(__u32 *val)
2436 /* from lov_setstripe */
2437 if ((*val != 0) && !lov_pattern_supported_normal_comp(*val)) {
2438 LCONSOLE_WARN("lod: Unknown stripe pattern: %#x\n", *val);
2443 void lod_fix_lmv_desc_pattern(__u32 *val)
2445 if ((*val) && !lmv_is_known_hash_type(*val)) {
2446 LCONSOLE_WARN("lod: Unknown md stripe pattern: %#x\n", *val);
2451 void lod_fix_desc_qos_maxage(__u32 *val)
2453 /* fix qos_maxage */
2455 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
2459 * Used to fix insane default striping.
2461 * \param[in] desc striping description
2463 void lod_fix_desc(struct lov_desc *desc)
2465 lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
2466 lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
2467 lod_fix_desc_pattern(&desc->ld_pattern);
2468 lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
2471 static void lod_fix_lmv_desc(struct lmv_desc *desc)
2473 desc->ld_active_tgt_count = 0;
2474 lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
2475 lod_fix_lmv_desc_pattern(&desc->ld_pattern);
2476 lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
2480 * Initialize the structures used to store pools and default striping.
2482 * \param[in] lod LOD device
2483 * \param[in] lcfg configuration structure storing default striping.
2485 * \retval 0 if initialization succeeds
2486 * \retval negative error number on failure
2488 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
2490 struct obd_device *obd;
2491 struct lov_desc *desc;
2495 obd = class_name2obd(lustre_cfg_string(lcfg, 0));
2496 LASSERT(obd != NULL);
2497 obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
2499 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
2500 CERROR("LOD setup requires a descriptor\n");
2504 desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
2506 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
2507 CERROR("descriptor size wrong: %d > %d\n",
2508 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
2512 if (desc->ld_magic != LOV_DESC_MAGIC) {
2513 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
2514 CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
2515 obd->obd_name, desc);
2516 lustre_swab_lov_desc(desc);
2518 CERROR("%s: Bad lov desc magic: %#x\n",
2519 obd->obd_name, desc->ld_magic);
2526 desc->ld_active_tgt_count = 0;
2527 lod->lod_ost_descs.ltd_lov_desc = *desc;
2529 /* NB: config doesn't contain lmv_desc, alter it via sysfs. */
2530 lod_fix_lmv_desc(&lod->lod_mdt_descs.ltd_lmv_desc);
2532 lod->lod_sp_me = LUSTRE_SP_CLI;
2534 /* Set up OST pool environment */
2535 lod->lod_pool_count = 0;
2536 rc = lod_pool_hash_init(&lod->lod_pools_hash_body);
2540 INIT_LIST_HEAD(&lod->lod_pool_list);
2541 lod->lod_pool_count = 0;
2542 rc = lu_tgt_pool_init(&lod->lod_mdt_descs.ltd_tgt_pool, 0);
2546 rc = lu_tgt_pool_init(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool, 0);
2548 GOTO(out_mdt_pool, rc);
2550 rc = lu_tgt_pool_init(&lod->lod_ost_descs.ltd_tgt_pool, 0);
2552 GOTO(out_mdt_rr_pool, rc);
2554 rc = lu_tgt_pool_init(&lod->lod_ost_descs.ltd_qos.lq_rr.lqr_pool, 0);
2556 GOTO(out_ost_pool, rc);
2561 lu_tgt_pool_free(&lod->lod_ost_descs.ltd_tgt_pool);
2563 lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool);
2565 lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_tgt_pool);
2567 lod_pool_hash_destroy(&lod->lod_pools_hash_body);
2573 * Release the structures describing the pools.
2575 * \param[in] lod LOD device from which we release the structures
2579 int lod_pools_fini(struct lod_device *lod)
2581 struct obd_device *obd = lod2obd(lod);
2582 struct lod_pool_desc *pool, *tmp;
2585 list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
2586 /* free pool structs */
2587 CDEBUG(D_INFO, "delete pool %p\n", pool);
2588 /* In the function below, .hs_keycmp resolves to
2589 * pool_hashkey_keycmp() */
2590 lod_pool_del(obd, pool->pool_name);
2593 lod_pool_hash_destroy(&lod->lod_pools_hash_body);
2594 lu_tgt_pool_free(&lod->lod_ost_descs.ltd_qos.lq_rr.lqr_pool);
2595 lu_tgt_pool_free(&lod->lod_ost_descs.ltd_tgt_pool);
2596 lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_qos.lq_rr.lqr_pool);
2597 lu_tgt_pool_free(&lod->lod_mdt_descs.ltd_tgt_pool);