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, 2016, 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;
79 struct list_head kill;
82 CDEBUG(D_CONFIG, "destroying %d ltd desc\n",
85 INIT_LIST_HEAD(&kill);
87 cfs_foreach_bit(ltd->ltd_tgt_bitmap, idx) {
88 tgt_desc = LTD_TGT(ltd, idx);
91 if (!tgt_desc->ltd_reap)
94 list_add(&tgt_desc->ltd_kill, &kill);
95 LTD_TGT(ltd, idx) = NULL;
96 /*FIXME: only support ost pool for now */
97 if (ltd == &lod->lod_ost_descs) {
98 lod_ost_pool_remove(&lod->lod_pool_info, idx);
99 if (tgt_desc->ltd_active)
100 lod->lod_desc.ld_active_tgt_count--;
103 cfs_bitmap_clear(ltd->ltd_tgt_bitmap, idx);
104 ltd->ltd_death_row--;
106 mutex_unlock(<d->ltd_mutex);
107 up_read(<d->ltd_rw_sem);
109 list_for_each_entry_safe(tgt_desc, tmp, &kill, ltd_kill) {
111 list_del(&tgt_desc->ltd_kill);
112 if (ltd == &lod->lod_ost_descs) {
113 /* remove from QoS structures */
114 rc = qos_del_tgt(lod, tgt_desc);
116 CERROR("%s: qos_del_tgt(%s) failed:"
118 lod2obd(lod)->obd_name,
119 obd_uuid2str(&tgt_desc->ltd_uuid),
122 rc = obd_disconnect(tgt_desc->ltd_exp);
124 CERROR("%s: failed to disconnect %s: rc = %d\n",
125 lod2obd(lod)->obd_name,
126 obd_uuid2str(&tgt_desc->ltd_uuid), rc);
127 OBD_FREE_PTR(tgt_desc);
130 mutex_unlock(<d->ltd_mutex);
131 up_read(<d->ltd_rw_sem);
136 * Expand size of target table.
138 * When the target table is full, we have to extend the table. To do so,
139 * we allocate new memory with some reserve, move data from the old table
140 * to the new one and release memory consumed by the old table.
141 * Notice we take ltd_rw_sem exclusively to ensure atomic switch.
143 * \param[in] ltd target table
144 * \param[in] newsize new size of the table
146 * \retval 0 on success
147 * \retval -ENOMEM if reallocation failed
149 static int ltd_bitmap_resize(struct lod_tgt_descs *ltd, __u32 newsize)
151 struct cfs_bitmap *new_bitmap, *old_bitmap = NULL;
155 /* grab write reference on the lod. Relocating the array requires
156 * exclusive access */
158 down_write(<d->ltd_rw_sem);
159 if (newsize <= ltd->ltd_tgts_size)
160 /* someone else has already resize the array */
163 /* allocate new bitmap */
164 new_bitmap = CFS_ALLOCATE_BITMAP(newsize);
166 GOTO(out, rc = -ENOMEM);
168 if (ltd->ltd_tgts_size > 0) {
169 /* the bitmap already exists, we need
170 * to copy data from old one */
171 cfs_bitmap_copy(new_bitmap, ltd->ltd_tgt_bitmap);
172 old_bitmap = ltd->ltd_tgt_bitmap;
175 ltd->ltd_tgts_size = newsize;
176 ltd->ltd_tgt_bitmap = new_bitmap;
179 CFS_FREE_BITMAP(old_bitmap);
181 CDEBUG(D_CONFIG, "tgt size: %d\n", ltd->ltd_tgts_size);
185 up_write(<d->ltd_rw_sem);
190 * Connect LOD to a new OSP and add it to the target table.
192 * Connect to the OSP device passed, initialize all the internal
193 * structures related to the device and add it to the target table.
195 * \param[in] env execution environment for this thread
196 * \param[in] lod LOD device to be connected to the new OSP
197 * \param[in] osp name of OSP device name to be added
198 * \param[in] index index of the new target
199 * \param[in] gen target's generation number
200 * \param[in] tgt_index OSP's group
201 * \param[in] type type of device (mdc or osc)
202 * \param[in] active state of OSP: 0 - inactive, 1 - active
204 * \retval 0 if added successfully
205 * \retval negative error number on failure
207 int lod_add_device(const struct lu_env *env, struct lod_device *lod,
208 char *osp, unsigned index, unsigned gen, int tgt_index,
209 char *type, int active)
211 struct obd_connect_data *data = NULL;
212 struct obd_export *exp = NULL;
213 struct obd_device *obd;
214 struct lu_device *lu_dev;
215 struct dt_device *dt_dev;
217 struct lod_tgt_desc *tgt_desc;
218 struct lod_tgt_descs *ltd;
219 struct lustre_cfg *lcfg;
220 struct obd_uuid obd_uuid;
225 CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, index, gen);
228 CERROR("request to add OBD %s with invalid generation: %d\n",
233 obd_str2uuid(&obd_uuid, osp);
235 obd = class_find_client_obd(&obd_uuid, LUSTRE_OSP_NAME,
236 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
238 CERROR("can't find %s device\n", osp);
242 LASSERT(obd->obd_lu_dev != NULL);
243 LASSERT(obd->obd_lu_dev->ld_site == lod->lod_dt_dev.dd_lu_dev.ld_site);
245 lu_dev = obd->obd_lu_dev;
246 dt_dev = lu2dt_dev(lu_dev);
250 GOTO(out_cleanup, rc = -ENOMEM);
252 data->ocd_connect_flags = OBD_CONNECT_INDEX | OBD_CONNECT_VERSION;
253 data->ocd_version = LUSTRE_VERSION_CODE;
254 data->ocd_index = index;
256 if (strcmp(LUSTRE_OSC_NAME, type) == 0) {
258 data->ocd_connect_flags |= OBD_CONNECT_AT |
261 #ifdef HAVE_LRU_RESIZE_SUPPORT
262 OBD_CONNECT_LRU_RESIZE |
265 OBD_CONNECT_REQPORTAL |
266 OBD_CONNECT_SKIP_ORPHAN |
268 OBD_CONNECT_LVB_TYPE |
269 OBD_CONNECT_VERSION |
270 OBD_CONNECT_PINGLESS |
272 OBD_CONNECT_BULK_MBITS;
274 data->ocd_group = tgt_index;
275 ltd = &lod->lod_ost_descs;
277 struct obd_import *imp = obd->u.cli.cl_import;
280 data->ocd_ibits_known = MDS_INODELOCK_UPDATE;
281 data->ocd_connect_flags |= OBD_CONNECT_ACL |
283 OBD_CONNECT_MDS_MDS |
288 OBD_CONNECT_BULK_MBITS;
289 spin_lock(&imp->imp_lock);
290 imp->imp_server_timeout = 1;
291 spin_unlock(&imp->imp_lock);
292 imp->imp_client->cli_request_portal = OUT_PORTAL;
293 CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
295 ltd = &lod->lod_mdt_descs;
298 rc = obd_connect(env, &exp, obd, &obd->obd_uuid, data, NULL);
301 CERROR("%s: cannot connect to next dev %s (%d)\n",
302 obd->obd_name, osp, rc);
303 GOTO(out_cleanup, rc);
306 /* Allocate ost descriptor and fill it */
307 OBD_ALLOC_PTR(tgt_desc);
309 GOTO(out_conn, rc = -ENOMEM);
311 tgt_desc->ltd_tgt = dt_dev;
312 tgt_desc->ltd_exp = exp;
313 tgt_desc->ltd_uuid = obd->u.cli.cl_target_uuid;
314 tgt_desc->ltd_gen = gen;
315 tgt_desc->ltd_index = index;
316 tgt_desc->ltd_active = active;
319 if (index >= ltd->ltd_tgts_size) {
320 /* we have to increase the size of the lod_osts array */
323 newsize = max(ltd->ltd_tgts_size, (__u32)2);
324 while (newsize < index + 1)
325 newsize = newsize << 1;
327 /* lod_bitmap_resize() needs lod_rw_sem
328 * which we hold with th reference */
329 lod_putref(lod, ltd);
331 rc = ltd_bitmap_resize(ltd, newsize);
338 mutex_lock(<d->ltd_mutex);
340 if (cfs_bitmap_check(ltd->ltd_tgt_bitmap, index)) {
341 CERROR("%s: device %d is registered already\n", obd->obd_name,
343 GOTO(out_mutex, rc = -EEXIST);
346 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
347 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
348 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
349 CERROR("can't allocate index to add %s\n",
351 GOTO(out_mutex, rc = -ENOMEM);
356 /* pool and qos are not supported for MDS stack yet */
357 rc = lod_ost_pool_add(&lod->lod_pool_info, index,
360 CERROR("%s: can't set up pool, failed with %d\n",
365 rc = qos_add_tgt(lod, tgt_desc);
367 CERROR("%s: qos_add_tgt failed with %d\n",
372 /* The new OST is now a full citizen */
373 if (index >= lod->lod_desc.ld_tgt_count)
374 lod->lod_desc.ld_tgt_count = index + 1;
376 lod->lod_desc.ld_active_tgt_count++;
379 LTD_TGT(ltd, index) = tgt_desc;
380 cfs_bitmap_set(ltd->ltd_tgt_bitmap, index);
382 mutex_unlock(<d->ltd_mutex);
383 lod_putref(lod, ltd);
385 if (lod->lod_recovery_completed)
386 lu_dev->ld_ops->ldo_recovery_complete(env, lu_dev);
388 if (!for_ost && lod->lod_initialized) {
389 rc = lod_sub_init_llog(env, lod, tgt_desc->ltd_tgt);
391 CERROR("%s: cannot start llog on %s:rc = %d\n",
392 lod2obd(lod)->obd_name, osp, rc);
397 rc = lfsck_add_target(env, lod->lod_child, dt_dev, exp, index, for_ost);
399 CERROR("Fail to add LFSCK target: name = %s, type = %s, "
400 "index = %u, rc = %d\n", osp, type, index, rc);
401 GOTO(out_fini_llog, rc);
405 lod_sub_fini_llog(env, tgt_desc->ltd_tgt,
406 tgt_desc->ltd_recovery_thread);
409 mutex_lock(<d->ltd_mutex);
411 if (!for_ost && LTD_TGT(ltd, index)->ltd_recovery_thread != NULL) {
412 struct ptlrpc_thread *thread;
414 thread = LTD_TGT(ltd, index)->ltd_recovery_thread;
415 OBD_FREE_PTR(thread);
418 cfs_bitmap_clear(ltd->ltd_tgt_bitmap, index);
419 LTD_TGT(ltd, index) = NULL;
421 lod_ost_pool_remove(&lod->lod_pool_info, index);
424 mutex_unlock(<d->ltd_mutex);
425 lod_putref(lod, ltd);
428 OBD_FREE_PTR(tgt_desc);
432 /* XXX OSP needs us to send down LCFG_CLEANUP because it uses
433 * objects from the MDT stack. See LU-7184. */
434 lcfg = &lod_env_info(env)->lti_lustre_cfg;
435 memset(lcfg, 0, sizeof(*lcfg));
436 lcfg->lcfg_version = LUSTRE_CFG_VERSION;
437 lcfg->lcfg_command = LCFG_CLEANUP;
438 lu_dev->ld_ops->ldo_process_config(env, lu_dev, lcfg);
444 * Schedule target removal from the target table.
446 * Mark the device as dead. The device is not removed here because it may
447 * still be in use. The device will be removed in lod_putref() when the
448 * last reference is released.
450 * \param[in] env execution environment for this thread
451 * \param[in] lod LOD device the target table belongs to
452 * \param[in] ltd target table
453 * \param[in] idx index of the target
454 * \param[in] for_ost type of the target: 0 - MDT, 1 - OST
456 static void __lod_del_device(const struct lu_env *env, struct lod_device *lod,
457 struct lod_tgt_descs *ltd, unsigned idx,
460 LASSERT(LTD_TGT(ltd, idx));
462 lfsck_del_target(env, lod->lod_child, LTD_TGT(ltd, idx)->ltd_tgt,
465 if (!for_ost && LTD_TGT(ltd, idx)->ltd_recovery_thread != NULL) {
466 struct ptlrpc_thread *thread;
468 thread = LTD_TGT(ltd, idx)->ltd_recovery_thread;
469 OBD_FREE_PTR(thread);
472 if (LTD_TGT(ltd, idx)->ltd_reap == 0) {
473 LTD_TGT(ltd, idx)->ltd_reap = 1;
474 ltd->ltd_death_row++;
479 * Schedule removal of all the targets from the given target table.
481 * See more details in the description for __lod_del_device()
483 * \param[in] env execution environment for this thread
484 * \param[in] lod LOD device the target table belongs to
485 * \param[in] ltd target table
486 * \param[in] for_ost type of the target: MDT or OST
490 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
491 struct lod_tgt_descs *ltd, bool for_ost)
495 if (ltd->ltd_tgts_size <= 0)
498 mutex_lock(<d->ltd_mutex);
499 cfs_foreach_bit(ltd->ltd_tgt_bitmap, idx)
500 __lod_del_device(env, lod, ltd, idx, for_ost);
501 mutex_unlock(<d->ltd_mutex);
502 lod_putref(lod, ltd);
503 CFS_FREE_BITMAP(ltd->ltd_tgt_bitmap);
504 for (idx = 0; idx < TGT_PTRS; idx++) {
505 if (ltd->ltd_tgt_idx[idx])
506 OBD_FREE_PTR(ltd->ltd_tgt_idx[idx]);
508 ltd->ltd_tgts_size = 0;
513 * Remove device by name.
515 * Remove a device identified by \a osp from the target table. Given
516 * the device can be in use, the real deletion happens in lod_putref().
518 * \param[in] env execution environment for this thread
519 * \param[in] lod LOD device to be connected to the new OSP
520 * \param[in] ltd target table
521 * \param[in] osp name of OSP device to be removed
522 * \param[in] idx index of the target
523 * \param[in] gen generation number, not used currently
524 * \param[in] for_ost type of the target: 0 - MDT, 1 - OST
526 * \retval 0 if the device was scheduled for removal
527 * \retval -EINVAL if no device was found
529 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
530 struct lod_tgt_descs *ltd, char *osp, unsigned idx,
531 unsigned gen, bool for_ost)
533 struct obd_device *obd;
535 struct obd_uuid uuid;
538 CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, idx, gen);
540 obd_str2uuid(&uuid, osp);
542 obd = class_find_client_obd(&uuid, LUSTRE_OSP_NAME,
543 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
545 CERROR("can't find %s device\n", osp);
550 CERROR("%s: request to remove OBD %s with invalid generation %d"
551 "\n", obd->obd_name, osp, gen);
555 obd_str2uuid(&uuid, osp);
558 mutex_lock(<d->ltd_mutex);
559 /* check that the index is allocated in the bitmap */
560 if (!cfs_bitmap_check(ltd->ltd_tgt_bitmap, idx) ||
561 !LTD_TGT(ltd, idx)) {
562 CERROR("%s: device %d is not set up\n", obd->obd_name, idx);
563 GOTO(out, rc = -EINVAL);
566 /* check that the UUID matches */
567 if (!obd_uuid_equals(&uuid, <D_TGT(ltd, idx)->ltd_uuid)) {
568 CERROR("%s: LOD target UUID %s at index %d does not match %s\n",
569 obd->obd_name, obd_uuid2str(<D_TGT(ltd,idx)->ltd_uuid),
571 GOTO(out, rc = -EINVAL);
574 __lod_del_device(env, lod, ltd, idx, for_ost);
577 mutex_unlock(<d->ltd_mutex);
578 lod_putref(lod, ltd);
583 * Resize per-thread storage to hold specified size.
585 * A helper function to resize per-thread temporary storage. This storage
586 * is used to process LOV/LVM EAs and may be quite large. We do not want to
587 * allocate/release it every time, so instead we put it into the env and
588 * reallocate on demand. The memory is released when the correspondent thread
591 * \param[in] info LOD-specific storage in the environment
592 * \param[in] size new size to grow the buffer to
594 * \retval 0 on success, -ENOMEM if reallocation failed
596 int lod_ea_store_resize(struct lod_thread_info *info, size_t size)
598 __u32 round = size_roundup_power2(size);
601 lov_mds_md_size(LOV_MAX_STRIPE_COUNT, LOV_MAGIC_V3));
602 if (info->lti_ea_store) {
603 LASSERT(info->lti_ea_store_size);
604 LASSERT(info->lti_ea_store_size < round);
605 CDEBUG(D_INFO, "EA store size %d is not enough, need %d\n",
606 info->lti_ea_store_size, round);
607 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
608 info->lti_ea_store = NULL;
609 info->lti_ea_store_size = 0;
612 OBD_ALLOC_LARGE(info->lti_ea_store, round);
613 if (info->lti_ea_store == NULL)
615 info->lti_ea_store_size = round;
620 static void lod_free_comp_buffer(struct lod_layout_component *entries,
621 __u16 count, __u32 bufsize)
623 struct lod_layout_component *entry;
626 for (i = 0; i < count; i++) {
628 if (entry->llc_pool != NULL)
629 lod_set_pool(&entry->llc_pool, NULL);
630 if (entry->llc_ostlist.op_array)
631 OBD_FREE(entry->llc_ostlist.op_array,
632 entry->llc_ostlist.op_size);
633 LASSERT(entry->llc_stripe == NULL);
634 LASSERT(entry->llc_stripes_allocated == 0);
638 OBD_FREE_LARGE(entries, bufsize);
641 void lod_free_def_comp_entries(struct lod_default_striping *lds)
643 lod_free_comp_buffer(lds->lds_def_comp_entries,
644 lds->lds_def_comp_size_cnt,
646 sizeof(*lds->lds_def_comp_entries) *
647 lds->lds_def_comp_size_cnt));
648 lds->lds_def_comp_entries = NULL;
649 lds->lds_def_comp_cnt = 0;
650 lds->lds_def_striping_is_composite = 0;
651 lds->lds_def_comp_size_cnt = 0;
655 * Resize per-thread storage to hold default striping component entries
657 * A helper function to resize per-thread temporary storage. This storage
658 * is used to hold default LOV/LVM EAs and may be quite large. We do not want
659 * to allocate/release it every time, so instead we put it into the env and
660 * reallocate it on demand. The memory is released when the correspondent
661 * thread is finished.
663 * \param[in,out] lds default striping
664 * \param[in] count new component count to grow the buffer to
666 * \retval 0 on success, -ENOMEM if reallocation failed
668 int lod_def_striping_comp_resize(struct lod_default_striping *lds, __u16 count)
670 struct lod_layout_component *entries;
671 __u32 new = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
673 __u32 old = size_roundup_power2(sizeof(*lds->lds_def_comp_entries) *
674 lds->lds_def_comp_size_cnt);
679 OBD_ALLOC_LARGE(entries, new);
683 if (lds->lds_def_comp_entries != NULL) {
684 CDEBUG(D_INFO, "default striping component size %d is not "
685 "enough, need %d\n", old, new);
686 lod_free_def_comp_entries(lds);
689 lds->lds_def_comp_entries = entries;
690 lds->lds_def_comp_size_cnt = count;
695 void lod_free_comp_entries(struct lod_object *lo)
697 if (lo->ldo_mirrors) {
698 OBD_FREE(lo->ldo_mirrors,
699 sizeof(*lo->ldo_mirrors) * lo->ldo_mirror_count);
700 lo->ldo_mirrors = NULL;
701 lo->ldo_mirror_count = 0;
703 lod_free_comp_buffer(lo->ldo_comp_entries,
705 sizeof(*lo->ldo_comp_entries) * lo->ldo_comp_cnt);
706 lo->ldo_comp_entries = NULL;
707 lo->ldo_comp_cnt = 0;
708 lo->ldo_is_composite = 0;
711 int lod_alloc_comp_entries(struct lod_object *lo,
712 int mirror_count, int comp_count)
714 LASSERT(comp_count != 0);
715 LASSERT(lo->ldo_comp_cnt == 0 && lo->ldo_comp_entries == NULL);
717 if (mirror_count > 0) {
718 OBD_ALLOC(lo->ldo_mirrors,
719 sizeof(*lo->ldo_mirrors) * mirror_count);
720 if (!lo->ldo_mirrors)
723 lo->ldo_mirror_count = mirror_count;
726 OBD_ALLOC_LARGE(lo->ldo_comp_entries,
727 sizeof(*lo->ldo_comp_entries) * comp_count);
728 if (lo->ldo_comp_entries == NULL) {
729 OBD_FREE(lo->ldo_mirrors,
730 sizeof(*lo->ldo_mirrors) * mirror_count);
731 lo->ldo_mirror_count = 0;
735 lo->ldo_comp_cnt = comp_count;
739 int lod_fill_mirrors(struct lod_object *lo)
741 struct lod_layout_component *lod_comp;
743 __u16 mirror_id = 0xffff;
747 LASSERT(equi(!lo->ldo_is_composite, lo->ldo_mirror_count == 0));
749 if (!lo->ldo_is_composite)
752 lod_comp = &lo->ldo_comp_entries[0];
753 for (i = 0; i < lo->ldo_comp_cnt; i++, lod_comp++) {
754 int stale = !!(lod_comp->llc_flags & LCME_FL_STALE);
756 if (mirror_id_of(lod_comp->llc_id) == mirror_id) {
757 lo->ldo_mirrors[mirror_idx].lme_stale |= stale;
758 lo->ldo_mirrors[mirror_idx].lme_end = i;
764 if (mirror_idx >= lo->ldo_mirror_count)
767 mirror_id = mirror_id_of(lod_comp->llc_id);
769 lo->ldo_mirrors[mirror_idx].lme_id = mirror_id;
770 lo->ldo_mirrors[mirror_idx].lme_stale = stale;
771 lo->ldo_mirrors[mirror_idx].lme_start = i;
772 lo->ldo_mirrors[mirror_idx].lme_end = i;
774 if (mirror_idx != lo->ldo_mirror_count - 1)
781 * Generate on-disk lov_mds_md structure for each layout component based on
782 * the information in lod_object->ldo_comp_entries[i].
784 * \param[in] env execution environment for this thread
785 * \param[in] lo LOD object
786 * \param[in] comp_idx index of ldo_comp_entries
787 * \param[in] lmm buffer to cotain the on-disk lov_mds_md
788 * \param[in|out] lmm_size buffer size/lmm size
789 * \param[in] is_dir generate lov ea for dir or file? For dir case,
790 * the stripe info is from the default stripe
791 * template, which is collected in lod_ah_init(),
792 * either from parent object or root object; for
793 * file case, it's from the @lo object
795 * \retval 0 if on disk structure is created successfully
796 * \retval negative error number on failure
798 static int lod_gen_component_ea(const struct lu_env *env,
799 struct lod_object *lo, int comp_idx,
800 struct lov_mds_md *lmm, int *lmm_size,
803 struct lod_thread_info *info = lod_env_info(env);
804 const struct lu_fid *fid = lu_object_fid(&lo->ldo_obj.do_lu);
805 struct lod_device *lod;
806 struct lov_ost_data_v1 *objs;
807 struct lod_layout_component *lod_comp;
816 &lo->ldo_def_striping->lds_def_comp_entries[comp_idx];
818 lod_comp = &lo->ldo_comp_entries[comp_idx];
820 magic = lod_comp->llc_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
821 if (lod_comp->llc_pattern == 0) /* default striping */
822 lod_comp->llc_pattern = LOV_PATTERN_RAID0;
824 lmm->lmm_magic = cpu_to_le32(magic);
825 lmm->lmm_pattern = cpu_to_le32(lod_comp->llc_pattern);
826 fid_to_lmm_oi(fid, &lmm->lmm_oi);
827 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
828 lmm->lmm_oi.oi.oi_id++;
829 lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
831 lmm->lmm_stripe_size = cpu_to_le32(lod_comp->llc_stripe_size);
832 lmm->lmm_stripe_count = cpu_to_le16(lod_comp->llc_stripe_count);
834 * for dir and uninstantiated component, lmm_layout_gen stores
835 * default stripe offset.
837 lmm->lmm_layout_gen =
838 (is_dir || !lod_comp_inited(lod_comp)) ?
839 cpu_to_le16(lod_comp->llc_stripe_offset) :
840 cpu_to_le16(lod_comp->llc_layout_gen);
842 if (magic == LOV_MAGIC_V1) {
843 objs = &lmm->lmm_objects[0];
845 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
846 size_t cplen = strlcpy(v3->lmm_pool_name,
848 sizeof(v3->lmm_pool_name));
849 if (cplen >= sizeof(v3->lmm_pool_name))
851 objs = &v3->lmm_objects[0];
853 stripe_count = lod_comp_entry_stripe_count(lo, lod_comp, is_dir);
854 if (!is_dir && lo->ldo_is_composite)
855 lod_comp_shrink_stripe_count(lod_comp, &stripe_count);
857 if (is_dir || lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED)
860 /* generate ost_idx of this component stripe */
861 lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
862 for (i = 0; i < stripe_count; i++) {
863 struct dt_object *object;
864 __u32 ost_idx = (__u32)-1UL;
865 int type = LU_SEQ_RANGE_OST;
867 if (lod_comp->llc_stripe && lod_comp->llc_stripe[i]) {
868 object = lod_comp->llc_stripe[i];
869 /* instantiated component */
870 info->lti_fid = *lu_object_fid(&object->do_lu);
872 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF) &&
874 if (cfs_fail_val == 0)
875 cfs_fail_val = info->lti_fid.f_oid;
877 info->lti_fid.f_oid = cfs_fail_val;
880 rc = fid_to_ostid(&info->lti_fid, &info->lti_ostid);
883 ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
884 objs[i].l_ost_gen = cpu_to_le32(0);
885 if (OBD_FAIL_CHECK(OBD_FAIL_MDS_FLD_LOOKUP))
888 rc = lod_fld_lookup(env, lod, &info->lti_fid,
891 CERROR("%s: Can not locate "DFID": rc = %d\n",
892 lod2obd(lod)->obd_name,
893 PFID(&info->lti_fid), rc);
896 } else if (lod_comp->llc_ostlist.op_array) {
897 /* user specified ost list */
898 ost_idx = lod_comp->llc_ostlist.op_array[i];
901 * with un-instantiated or with no specified ost list
902 * component, its l_ost_idx does not matter.
904 objs[i].l_ost_idx = cpu_to_le32(ost_idx);
907 if (lmm_size != NULL)
908 *lmm_size = lov_mds_md_size(stripe_count, magic);
913 * Generate on-disk lov_mds_md structure based on the information in
914 * the lod_object->ldo_comp_entries.
916 * \param[in] env execution environment for this thread
917 * \param[in] lo LOD object
918 * \param[in] lmm buffer to cotain the on-disk lov_mds_md
919 * \param[in|out] lmm_size buffer size/lmm size
920 * \param[in] is_dir generate lov ea for dir or file? For dir case,
921 * the stripe info is from the default stripe
922 * template, which is collected in lod_ah_init(),
923 * either from parent object or root object; for
924 * file case, it's from the @lo object
926 * \retval 0 if on disk structure is created successfully
927 * \retval negative error number on failure
929 int lod_generate_lovea(const struct lu_env *env, struct lod_object *lo,
930 struct lov_mds_md *lmm, int *lmm_size, bool is_dir)
932 struct lov_comp_md_entry_v1 *lcme;
933 struct lov_comp_md_v1 *lcm;
934 struct lod_layout_component *comp_entries;
935 __u16 comp_cnt, mirror_cnt;
937 int i, rc = 0, offset;
941 comp_cnt = lo->ldo_def_striping->lds_def_comp_cnt;
942 mirror_cnt = lo->ldo_def_striping->lds_def_mirror_cnt;
943 comp_entries = lo->ldo_def_striping->lds_def_comp_entries;
945 lo->ldo_def_striping->lds_def_striping_is_composite;
947 comp_cnt = lo->ldo_comp_cnt;
948 mirror_cnt = lo->ldo_mirror_count;
949 comp_entries = lo->ldo_comp_entries;
950 is_composite = lo->ldo_is_composite;
953 LASSERT(lmm_size != NULL);
954 LASSERT(comp_cnt != 0 && comp_entries != NULL);
957 rc = lod_gen_component_ea(env, lo, 0, lmm, lmm_size, is_dir);
961 lcm = (struct lov_comp_md_v1 *)lmm;
962 lcm->lcm_magic = cpu_to_le32(LOV_MAGIC_COMP_V1);
963 lcm->lcm_entry_count = cpu_to_le16(comp_cnt);
964 lcm->lcm_mirror_count = cpu_to_le16(mirror_cnt - 1);
965 lcm->lcm_flags = cpu_to_le16(lo->ldo_flr_state);
967 offset = sizeof(*lcm) + sizeof(*lcme) * comp_cnt;
968 LASSERT(offset % sizeof(__u64) == 0);
970 for (i = 0; i < comp_cnt; i++) {
971 struct lod_layout_component *lod_comp;
972 struct lov_mds_md *sub_md;
975 lod_comp = &comp_entries[i];
976 lcme = &lcm->lcm_entries[i];
978 LASSERT(ergo(!is_dir, lod_comp->llc_id != LCME_ID_INVAL));
979 lcme->lcme_id = cpu_to_le32(lod_comp->llc_id);
981 /* component could be un-inistantiated */
982 lcme->lcme_flags = cpu_to_le32(lod_comp->llc_flags);
983 lcme->lcme_extent.e_start =
984 cpu_to_le64(lod_comp->llc_extent.e_start);
985 lcme->lcme_extent.e_end =
986 cpu_to_le64(lod_comp->llc_extent.e_end);
987 lcme->lcme_offset = cpu_to_le32(offset);
989 sub_md = (struct lov_mds_md *)((char *)lcm + offset);
990 rc = lod_gen_component_ea(env, lo, i, sub_md, &size, is_dir);
993 lcme->lcme_size = cpu_to_le32(size);
995 LASSERTF((offset <= *lmm_size) && (offset % sizeof(__u64) == 0),
996 "offset:%d lmm_size:%d\n", offset, *lmm_size);
998 lcm->lcm_size = cpu_to_le32(offset);
999 lcm->lcm_layout_gen = cpu_to_le32(is_dir ? 0 : lo->ldo_layout_gen);
1001 lustre_print_user_md(D_LAYOUT, (struct lov_user_md *)lmm,
1012 * Fill lti_ea_store buffer in the environment with a value for the given
1013 * EA. The buffer is reallocated if the value doesn't fit.
1015 * \param[in,out] env execution environment for this thread
1016 * .lti_ea_store buffer is filled with EA's value
1017 * \param[in] lo LOD object
1018 * \param[in] name name of the EA
1020 * \retval > 0 if EA is fetched successfully
1021 * \retval 0 if EA is empty
1022 * \retval negative error number on failure
1024 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
1027 struct lod_thread_info *info = lod_env_info(env);
1028 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1034 if (unlikely(info->lti_ea_store == NULL)) {
1035 /* just to enter in allocation block below */
1039 info->lti_buf.lb_buf = info->lti_ea_store;
1040 info->lti_buf.lb_len = info->lti_ea_store_size;
1041 rc = dt_xattr_get(env, next, &info->lti_buf, name);
1044 /* if object is not striped or inaccessible */
1045 if (rc == -ENODATA || rc == -ENOENT)
1048 if (rc == -ERANGE) {
1049 /* EA doesn't fit, reallocate new buffer */
1050 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name);
1051 if (rc == -ENODATA || rc == -ENOENT)
1057 rc = lod_ea_store_resize(info, rc);
1067 * Verify the target index is present in the current configuration.
1069 * \param[in] md LOD device where the target table is stored
1070 * \param[in] idx target's index
1072 * \retval 0 if the index is present
1073 * \retval -EINVAL if not
1075 static int validate_lod_and_idx(struct lod_device *md, __u32 idx)
1077 if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
1078 !cfs_bitmap_check(md->lod_ost_bitmap, idx))) {
1079 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
1080 md->lod_ost_descs.ltd_tgts_size);
1084 if (unlikely(OST_TGT(md, idx) == NULL)) {
1085 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
1086 lod2obd(md)->obd_name, idx);
1090 if (unlikely(OST_TGT(md, idx)->ltd_ost == NULL)) {
1091 CERROR("%s: invalid lod device, for idx: %d\n",
1092 lod2obd(md)->obd_name , idx);
1100 * Instantiate objects for stripes.
1102 * Allocate and initialize LU-objects representing the stripes. The number
1103 * of the stripes (ldo_stripe_count) must be initialized already. The caller
1104 * must ensure nobody else is calling the function on the object at the same
1105 * time. FLDB service must be running to be able to map a FID to the targets
1106 * and find appropriate device representing that target.
1108 * \param[in] env execution environment for this thread
1109 * \param[in,out] lo LOD object
1110 * \param[in] objs an array of IDs to creates the objects from
1111 * \param[in] comp_idx index of ldo_comp_entries
1113 * \retval 0 if the objects are instantiated successfully
1114 * \retval negative error number on failure
1116 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
1117 struct lov_ost_data_v1 *objs, int comp_idx)
1119 struct lod_layout_component *lod_comp;
1120 struct lod_thread_info *info = lod_env_info(env);
1121 struct lod_device *md;
1122 struct lu_object *o, *n;
1123 struct lu_device *nd;
1124 struct dt_object **stripe;
1130 LASSERT(lo != NULL);
1131 md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
1133 LASSERT(lo->ldo_comp_cnt != 0 && lo->ldo_comp_entries != NULL);
1134 lod_comp = &lo->ldo_comp_entries[comp_idx];
1136 LASSERT(lod_comp->llc_stripe == NULL);
1137 LASSERT(lod_comp->llc_stripe_count > 0);
1138 LASSERT(lod_comp->llc_stripe_size > 0);
1140 stripe_len = lod_comp->llc_stripe_count;
1141 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
1145 for (i = 0; i < lod_comp->llc_stripe_count; i++) {
1146 if (unlikely(lovea_slot_is_dummy(&objs[i])))
1149 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
1150 idx = le32_to_cpu(objs[i].l_ost_idx);
1151 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
1154 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
1155 PFID(&info->lti_fid));
1156 lod_getref(&md->lod_ost_descs);
1158 rc = validate_lod_and_idx(md, idx);
1159 if (unlikely(rc != 0)) {
1160 lod_putref(md, &md->lod_ost_descs);
1164 nd = &OST_TGT(md,idx)->ltd_ost->dd_lu_dev;
1165 lod_putref(md, &md->lod_ost_descs);
1167 /* In the function below, .hs_keycmp resolves to
1168 * u_obj_hop_keycmp() */
1169 /* coverity[overrun-buffer-val] */
1170 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
1172 GOTO(out, rc = PTR_ERR(o));
1174 n = lu_object_locate(o->lo_header, nd->ld_type);
1177 stripe[i] = container_of(n, struct dt_object, do_lu);
1182 for (i = 0; i < stripe_len; i++)
1183 if (stripe[i] != NULL)
1184 dt_object_put(env, stripe[i]);
1186 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
1187 lod_comp->llc_stripe_count = 0;
1189 lod_comp->llc_stripe = stripe;
1190 lod_comp->llc_stripes_allocated = stripe_len;
1197 * Instantiate objects for striping.
1199 * Parse striping information in \a buf and instantiate the objects
1200 * representing the stripes.
1202 * \param[in] env execution environment for this thread
1203 * \param[in] lo LOD object
1204 * \param[in] buf buffer storing LOV EA to parse
1206 * \retval 0 if parsing and objects creation succeed
1207 * \retval negative error number on failure
1209 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
1210 const struct lu_buf *buf)
1212 struct lov_mds_md_v1 *lmm;
1213 struct lov_comp_md_v1 *comp_v1 = NULL;
1214 struct lov_ost_data_v1 *objs;
1215 __u32 magic, pattern;
1218 __u16 mirror_cnt = 0;
1222 LASSERT(buf->lb_buf);
1223 LASSERT(buf->lb_len);
1225 lmm = (struct lov_mds_md_v1 *)buf->lb_buf;
1226 magic = le32_to_cpu(lmm->lmm_magic);
1228 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3 &&
1229 magic != LOV_MAGIC_COMP_V1)
1230 GOTO(out, rc = -EINVAL);
1232 lod_free_comp_entries(lo);
1234 if (magic == LOV_MAGIC_COMP_V1) {
1235 comp_v1 = (struct lov_comp_md_v1 *)lmm;
1236 comp_cnt = le16_to_cpu(comp_v1->lcm_entry_count);
1238 GOTO(out, rc = -EINVAL);
1239 lo->ldo_layout_gen = le32_to_cpu(comp_v1->lcm_layout_gen);
1240 lo->ldo_is_composite = 1;
1241 lo->ldo_flr_state = le16_to_cpu(comp_v1->lcm_flags) &
1243 mirror_cnt = le16_to_cpu(comp_v1->lcm_mirror_count) + 1;
1246 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1247 lo->ldo_is_composite = 0;
1250 rc = lod_alloc_comp_entries(lo, mirror_cnt, comp_cnt);
1254 for (i = 0; i < comp_cnt; i++) {
1255 struct lod_layout_component *lod_comp;
1256 struct lu_extent *ext;
1259 lod_comp = &lo->ldo_comp_entries[i];
1260 if (lo->ldo_is_composite) {
1261 offs = le32_to_cpu(comp_v1->lcm_entries[i].lcme_offset);
1262 lmm = (struct lov_mds_md_v1 *)((char *)comp_v1 + offs);
1263 magic = le32_to_cpu(lmm->lmm_magic);
1265 ext = &comp_v1->lcm_entries[i].lcme_extent;
1266 lod_comp->llc_extent.e_start =
1267 le64_to_cpu(ext->e_start);
1268 lod_comp->llc_extent.e_end = le64_to_cpu(ext->e_end);
1269 lod_comp->llc_flags =
1270 le32_to_cpu(comp_v1->lcm_entries[i].lcme_flags);
1272 le32_to_cpu(comp_v1->lcm_entries[i].lcme_id);
1273 if (lod_comp->llc_id == LCME_ID_INVAL)
1274 GOTO(out, rc = -EINVAL);
1276 lod_comp_set_init(lod_comp);
1279 pattern = le32_to_cpu(lmm->lmm_pattern);
1280 if (lov_pattern(pattern) != LOV_PATTERN_RAID0 &&
1281 lov_pattern(pattern) != LOV_PATTERN_MDT)
1282 GOTO(out, rc = -EINVAL);
1284 lod_comp->llc_pattern = pattern;
1285 lod_comp->llc_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
1286 lod_comp->llc_stripe_count = le16_to_cpu(lmm->lmm_stripe_count);
1287 lod_comp->llc_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
1289 if (magic == LOV_MAGIC_V3) {
1290 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *)lmm;
1291 lod_set_pool(&lod_comp->llc_pool, v3->lmm_pool_name);
1292 objs = &v3->lmm_objects[0];
1294 lod_set_pool(&lod_comp->llc_pool, NULL);
1295 objs = &lmm->lmm_objects[0];
1299 * If uninstantiated template component has valid l_ost_idx,
1300 * then user has specified ost list for this component.
1302 if (!lod_comp_inited(lod_comp)) {
1303 if (objs[0].l_ost_idx != (__u32)-1UL) {
1305 * load the user specified ost list, when this
1306 * component is instantiated later, it will be
1307 * used in lod_alloc_ost_list().
1309 lod_comp->llc_ostlist.op_count =
1310 lod_comp->llc_stripe_count;
1311 lod_comp->llc_ostlist.op_size =
1312 lod_comp->llc_stripe_count *
1314 OBD_ALLOC(lod_comp->llc_ostlist.op_array,
1315 lod_comp->llc_ostlist.op_size);
1316 if (!lod_comp->llc_ostlist.op_array)
1317 GOTO(out, rc = -ENOMEM);
1319 for (j = 0; j < lod_comp->llc_stripe_count; j++)
1320 lod_comp->llc_ostlist.op_array[j] =
1321 le32_to_cpu(objs[j].l_ost_idx);
1324 * this component OST objects starts from the
1325 * first ost_idx, lod_alloc_ost_list() will
1328 lod_comp->llc_stripe_offset = objs[0].l_ost_idx;
1331 * for uninstantiated component,
1332 * lmm_layout_gen stores default stripe offset.
1334 lod_comp->llc_stripe_offset =
1335 lmm->lmm_layout_gen;
1339 /* skip un-instantiated component object initialization */
1340 if (!lod_comp_inited(lod_comp))
1343 if (!(lod_comp->llc_pattern & LOV_PATTERN_F_RELEASED) &&
1344 !(lod_comp->llc_pattern & LOV_PATTERN_MDT)) {
1345 rc = lod_initialize_objects(env, lo, objs, i);
1351 rc = lod_fill_mirrors(lo);
1357 lod_object_free_striping(env, lo);
1362 * Check whether the striping (LOVEA for regular file, LMVEA for directory)
1363 * is already cached.
1365 * \param[in] lo LOD object
1367 * \retval True if the striping is cached, otherwise
1370 static bool lod_striping_loaded(struct lod_object *lo)
1372 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1373 lo->ldo_comp_cached)
1376 if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1377 if (lo->ldo_stripe != NULL || lo->ldo_dir_stripe_loaded)
1380 /* Never load LMV stripe for slaves of striped dir */
1381 if (lo->ldo_dir_slave_stripe)
1389 * Initialize the object representing the stripes.
1391 * Unless the stripes are initialized already, fetch LOV (for regular
1392 * objects) or LMV (for directory objects) EA and call lod_parse_striping()
1393 * to instantiate the objects representing the stripes. Caller should
1394 * hold the dt_write_lock(next).
1396 * \param[in] env execution environment for this thread
1397 * \param[in,out] lo LOD object
1399 * \retval 0 if parsing and object creation succeed
1400 * \retval negative error number on failure
1402 int lod_load_striping_locked(const struct lu_env *env, struct lod_object *lo)
1404 struct lod_thread_info *info = lod_env_info(env);
1405 struct lu_buf *buf = &info->lti_buf;
1406 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1410 if (!dt_object_exists(next))
1413 if (lod_striping_loaded(lo))
1416 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr)) {
1417 rc = lod_get_lov_ea(env, lo);
1421 * there is LOV EA (striping information) in this object
1422 * let's parse it and create in-core objects for the stripes
1424 buf->lb_buf = info->lti_ea_store;
1425 buf->lb_len = info->lti_ea_store_size;
1426 rc = lod_parse_striping(env, lo, buf);
1428 lo->ldo_comp_cached = 1;
1429 } else if (S_ISDIR(lod2lu_obj(lo)->lo_header->loh_attr)) {
1430 rc = lod_get_lmv_ea(env, lo);
1431 if (rc < (typeof(rc))sizeof(struct lmv_mds_md_v1)) {
1432 /* Let's set stripe_loaded to avoid further
1433 * stripe loading especially for non-stripe directory,
1434 * which can hurt performance. (See LU-9840)
1437 lo->ldo_dir_stripe_loaded = 1;
1438 GOTO(out, rc = rc > 0 ? -EINVAL : rc);
1440 buf->lb_buf = info->lti_ea_store;
1441 buf->lb_len = info->lti_ea_store_size;
1442 if (rc == sizeof(struct lmv_mds_md_v1)) {
1443 rc = lod_load_lmv_shards(env, lo, buf, true);
1444 if (buf->lb_buf != info->lti_ea_store) {
1445 OBD_FREE_LARGE(info->lti_ea_store,
1446 info->lti_ea_store_size);
1447 info->lti_ea_store = buf->lb_buf;
1448 info->lti_ea_store_size = buf->lb_len;
1456 * there is LMV EA (striping information) in this object
1457 * let's parse it and create in-core objects for the stripes
1459 rc = lod_parse_dir_striping(env, lo, buf);
1461 lo->ldo_dir_stripe_loaded = 1;
1468 * A generic function to initialize the stripe objects.
1470 * A protected version of lod_load_striping_locked() - load the striping
1471 * information from storage, parse that and instantiate LU objects to
1472 * represent the stripes. The LOD object \a lo supplies a pointer to the
1473 * next sub-object in the LU stack so we can lock it. Also use \a lo to
1474 * return an array of references to the newly instantiated objects.
1476 * \param[in] env execution environment for this thread
1477 * \param[in,out] lo LOD object, where striping is stored and
1478 * which gets an array of references
1480 * \retval 0 if parsing and object creation succeed
1481 * \retval negative error number on failure
1483 int lod_load_striping(const struct lu_env *env, struct lod_object *lo)
1485 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1488 if (!dt_object_exists(next))
1491 /* Check without locking first */
1492 if (lod_striping_loaded(lo))
1495 /* currently this code is supposed to be called from declaration
1496 * phase only, thus the object is not expected to be locked by caller */
1497 dt_write_lock(env, next, 0);
1498 rc = lod_load_striping_locked(env, lo);
1499 dt_write_unlock(env, next);
1504 * Verify lov_user_md_v1/v3 striping.
1506 * Check the validity of all fields including the magic, stripe size,
1507 * stripe count, stripe offset and that the pool is present. Also check
1508 * that each target index points to an existing target. The additional
1509 * \a is_from_disk turns additional checks. In some cases zero fields
1510 * are allowed (like pattern=0).
1512 * \param[in] d LOD device
1513 * \param[in] buf buffer with LOV EA to verify
1514 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
1515 * 1 - from disk, do not allow
1517 * \retval 0 if the striping is valid
1518 * \retval -EINVAL if striping is invalid
1520 static int lod_verify_v1v3(struct lod_device *d, const struct lu_buf *buf,
1523 struct lov_user_md_v1 *lum;
1524 struct lov_user_md_v3 *lum3;
1525 struct pool_desc *pool = NULL;
1529 __u16 stripe_offset;
1536 if (buf->lb_len < sizeof(*lum)) {
1537 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1539 GOTO(out, rc = -EINVAL);
1542 magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1543 if (magic != LOV_USER_MAGIC_V1 &&
1544 magic != LOV_USER_MAGIC_V3 &&
1545 magic != LOV_USER_MAGIC_SPECIFIC) {
1546 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1547 le32_to_cpu(lum->lmm_magic));
1548 GOTO(out, rc = -EINVAL);
1551 /* the user uses "0" for default stripe pattern normally. */
1552 if (!is_from_disk && lum->lmm_pattern == LOV_PATTERN_NONE)
1553 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1555 if (!lov_pattern_supported(le32_to_cpu(lum->lmm_pattern))) {
1556 CDEBUG(D_LAYOUT, "bad userland stripe pattern: %#x\n",
1557 le32_to_cpu(lum->lmm_pattern));
1558 GOTO(out, rc = -EINVAL);
1561 /* a released lum comes from creating orphan on hsm release,
1562 * doesn't make sense to verify it. */
1563 if (le32_to_cpu(lum->lmm_pattern) & LOV_PATTERN_F_RELEASED)
1566 /* 64kB is the largest common page size we see (ia64), and matches the
1568 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1569 if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1570 CDEBUG(D_LAYOUT, "stripe size %u not a multiple of %u\n",
1571 stripe_size, LOV_MIN_STRIPE_SIZE);
1572 GOTO(out, rc = -EINVAL);
1575 stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1576 if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT &&
1577 lov_pattern(le32_to_cpu(lum->lmm_pattern)) != LOV_PATTERN_MDT) {
1578 /* if offset is not within valid range [0, osts_size) */
1579 if (stripe_offset >= d->lod_osts_size) {
1580 CDEBUG(D_LAYOUT, "stripe offset %u >= bitmap size %u\n",
1581 stripe_offset, d->lod_osts_size);
1582 GOTO(out, rc = -EINVAL);
1585 /* if lmm_stripe_offset is *not* in bitmap */
1586 if (!cfs_bitmap_check(d->lod_ost_bitmap, stripe_offset)) {
1587 CDEBUG(D_LAYOUT, "stripe offset %u not in bitmap\n",
1589 GOTO(out, rc = -EINVAL);
1593 if (magic == LOV_USER_MAGIC_V1)
1594 lum_size = offsetof(struct lov_user_md_v1,
1596 else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC)
1597 lum_size = offsetof(struct lov_user_md_v3,
1600 GOTO(out, rc = -EINVAL);
1602 stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1603 if (buf->lb_len < lum_size) {
1604 CDEBUG(D_LAYOUT, "invalid buf len %zu/%zu for lov_user_md with "
1605 "magic %#x and stripe_count %u\n",
1606 buf->lb_len, lum_size, magic, stripe_count);
1607 GOTO(out, rc = -EINVAL);
1610 if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_USER_MAGIC_SPECIFIC))
1614 /* In the function below, .hs_keycmp resolves to
1615 * pool_hashkey_keycmp() */
1616 /* coverity[overrun-buffer-val] */
1617 pool = lod_find_pool(d, lum3->lmm_pool_name);
1621 if (!is_from_disk && stripe_offset != LOV_OFFSET_DEFAULT) {
1622 rc = lod_check_index_in_pool(stripe_offset, pool);
1624 GOTO(out, rc = -EINVAL);
1627 if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1628 CDEBUG(D_LAYOUT, "stripe count %u > # OSTs %u in the pool\n",
1629 stripe_count, pool_tgt_count(pool));
1630 GOTO(out, rc = -EINVAL);
1635 lod_pool_putref(pool);
1641 * Verify LOV striping.
1643 * \param[in] d LOD device
1644 * \param[in] buf buffer with LOV EA to verify
1645 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
1646 * 1 - from disk, do not allow
1647 * \param[in] start extent start for composite layout
1649 * \retval 0 if the striping is valid
1650 * \retval -EINVAL if striping is invalid
1652 int lod_verify_striping(struct lod_device *d, struct lod_object *lo,
1653 const struct lu_buf *buf, bool is_from_disk)
1655 struct lov_user_md_v1 *lum;
1656 struct lov_comp_md_v1 *comp_v1;
1663 if (buf->lb_len < sizeof(*lum)) {
1664 CDEBUG(D_LAYOUT, "buf len %zu too small for lov_user_md\n",
1669 magic = le32_to_cpu(lum->lmm_magic) & ~LOV_MAGIC_DEFINED;
1670 if (magic != LOV_USER_MAGIC_V1 &&
1671 magic != LOV_USER_MAGIC_V3 &&
1672 magic != LOV_USER_MAGIC_SPECIFIC &&
1673 magic != LOV_USER_MAGIC_COMP_V1) {
1674 CDEBUG(D_LAYOUT, "bad userland LOV MAGIC: %#x\n",
1675 le32_to_cpu(lum->lmm_magic));
1679 if (magic == LOV_USER_MAGIC_COMP_V1) {
1680 struct lov_comp_md_entry_v1 *ent;
1681 struct lu_extent *ext;
1682 struct lov_desc *desc = &d->lod_desc;
1685 __u32 stripe_size = 0;
1687 comp_v1 = buf->lb_buf;
1688 if (buf->lb_len < le32_to_cpu(comp_v1->lcm_size)) {
1689 CDEBUG(D_LAYOUT, "buf len %zu is less than %u\n",
1690 buf->lb_len, le32_to_cpu(comp_v1->lcm_size));
1694 if (le16_to_cpu(comp_v1->lcm_entry_count) == 0) {
1695 CDEBUG(D_LAYOUT, "entry count is zero\n");
1699 if (S_ISREG(lod2lu_obj(lo)->lo_header->loh_attr) &&
1700 lo->ldo_comp_cnt > 0) {
1701 __u32 cnt = lo->ldo_comp_cnt;
1703 ext = &lo->ldo_comp_entries[cnt - 1].llc_extent;
1704 prev_end = ext->e_end;
1707 for (i = 0; i < le16_to_cpu(comp_v1->lcm_entry_count); i++) {
1708 ent = &comp_v1->lcm_entries[i];
1709 ext = &ent->lcme_extent;
1712 (le32_to_cpu(ent->lcme_id) == 0 ||
1713 le32_to_cpu(ent->lcme_id) > LCME_ID_MAX)) {
1714 CDEBUG(D_LAYOUT, "invalid id %u\n",
1715 le32_to_cpu(ent->lcme_id));
1719 if (le64_to_cpu(ext->e_start) >=
1720 le64_to_cpu(ext->e_end)) {
1721 CDEBUG(D_LAYOUT, "invalid extent "
1723 le64_to_cpu(ext->e_start),
1724 le64_to_cpu(ext->e_end));
1728 /* first component must start with 0, and the next
1729 * must be adjacent with the previous one */
1730 if (le64_to_cpu(ext->e_start) != prev_end) {
1731 CDEBUG(D_LAYOUT, "invalid start "
1732 "actual:%llu, expect:%llu\n",
1733 le64_to_cpu(ext->e_start), prev_end);
1737 prev_end = le64_to_cpu(ext->e_end);
1739 tmp.lb_buf = (char *)comp_v1 +
1740 le32_to_cpu(ent->lcme_offset);
1741 tmp.lb_len = le32_to_cpu(ent->lcme_size);
1743 /* Checks for DoM entry in composite layout. */
1745 if (lov_pattern(le32_to_cpu(lum->lmm_pattern)) ==
1747 /* DoM component can be only the first entry */
1749 CDEBUG(D_LAYOUT, "invalid DoM layout "
1750 "entry found at %i index\n", i);
1753 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1754 /* There is just one stripe on MDT and it must
1755 * cover whole component size. */
1756 if (stripe_size != prev_end) {
1757 CDEBUG(D_LAYOUT, "invalid DoM layout "
1758 "stripe size %u != %llu "
1759 "(component size)\n",
1760 stripe_size, prev_end);
1763 /* Check stripe size againts per-MDT limit */
1764 if (stripe_size > d->lod_dom_max_stripesize) {
1765 CDEBUG(D_LAYOUT, "DoM component size "
1766 "%u is bigger than MDT limit "
1767 "%u, check dom_max_stripesize"
1770 d->lod_dom_max_stripesize);
1774 rc = lod_verify_v1v3(d, &tmp, is_from_disk);
1780 /* extent end must be aligned with the stripe_size */
1781 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1782 if (stripe_size == 0)
1783 stripe_size = desc->ld_default_stripe_size;
1784 if (stripe_size == 0 ||
1785 (prev_end != LUSTRE_EOF &&
1786 (prev_end & (stripe_size - 1)))) {
1787 CDEBUG(D_LAYOUT, "stripe size isn't aligned. "
1788 " stripe_sz: %u, [%llu, %llu)\n",
1789 stripe_size, ext->e_start, prev_end);
1794 rc = lod_verify_v1v3(d, buf, is_from_disk);
1801 * set the default stripe size, if unset.
1803 * \param[in,out] val number of bytes per OST stripe
1805 * The minimum stripe size is 64KB to ensure that a single stripe is an
1806 * even multiple of a client PAGE_SIZE (IA64, PPC, etc). Otherwise, it
1807 * is difficult to split dirty pages across OSCs during writes.
1809 void lod_fix_desc_stripe_size(__u64 *val)
1811 if (*val < LOV_MIN_STRIPE_SIZE) {
1813 LCONSOLE_INFO("Increasing default stripe size to "
1814 "minimum value %u\n",
1815 LOV_DESC_STRIPE_SIZE_DEFAULT);
1816 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
1817 } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
1818 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
1819 LCONSOLE_WARN("Changing default stripe size to %llu (a "
1820 "multiple of %u)\n",
1821 *val, LOV_MIN_STRIPE_SIZE);
1826 * set the filesystem default number of stripes, if unset.
1828 * \param[in,out] val number of stripes
1830 * A value of "0" means "use the system-wide default stripe count", which
1831 * has either been inherited by now, or falls back to 1 stripe per file.
1832 * A value of "-1" (0xffffffff) means "stripe over all available OSTs",
1833 * and is a valid value, so is left unchanged here.
1835 void lod_fix_desc_stripe_count(__u32 *val)
1842 * set the filesystem default layout pattern
1844 * \param[in,out] val LOV_PATTERN_* layout
1846 * A value of "0" means "use the system-wide default layout type", which
1847 * has either been inherited by now, or falls back to plain RAID0 striping.
1849 void lod_fix_desc_pattern(__u32 *val)
1851 /* from lov_setstripe */
1852 if ((*val != 0) && (*val != LOV_PATTERN_RAID0) &&
1853 (*val != LOV_PATTERN_MDT)) {
1854 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
1859 void lod_fix_desc_qos_maxage(__u32 *val)
1861 /* fix qos_maxage */
1863 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
1867 * Used to fix insane default striping.
1869 * \param[in] desc striping description
1871 void lod_fix_desc(struct lov_desc *desc)
1873 lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
1874 lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
1875 lod_fix_desc_pattern(&desc->ld_pattern);
1876 lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
1880 * Initialize the structures used to store pools and default striping.
1882 * \param[in] lod LOD device
1883 * \param[in] lcfg configuration structure storing default striping.
1885 * \retval 0 if initialization succeeds
1886 * \retval negative error number on failure
1888 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
1890 struct obd_device *obd;
1891 struct lov_desc *desc;
1895 obd = class_name2obd(lustre_cfg_string(lcfg, 0));
1896 LASSERT(obd != NULL);
1897 obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1899 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1900 CERROR("LOD setup requires a descriptor\n");
1904 desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
1906 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1907 CERROR("descriptor size wrong: %d > %d\n",
1908 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1912 if (desc->ld_magic != LOV_DESC_MAGIC) {
1913 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
1914 CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
1915 obd->obd_name, desc);
1916 lustre_swab_lov_desc(desc);
1918 CERROR("%s: Bad lov desc magic: %#x\n",
1919 obd->obd_name, desc->ld_magic);
1926 desc->ld_active_tgt_count = 0;
1927 lod->lod_desc = *desc;
1929 lod->lod_sp_me = LUSTRE_SP_CLI;
1931 /* Set up allocation policy (QoS and RR) */
1932 INIT_LIST_HEAD(&lod->lod_qos.lq_oss_list);
1933 init_rwsem(&lod->lod_qos.lq_rw_sem);
1934 lod->lod_qos.lq_dirty = 1;
1935 lod->lod_qos.lq_rr.lqr_dirty = 1;
1936 lod->lod_qos.lq_reset = 1;
1937 /* Default priority is toward free space balance */
1938 lod->lod_qos.lq_prio_free = 232;
1939 /* Default threshold for rr (roughly 17%) */
1940 lod->lod_qos.lq_threshold_rr = 43;
1942 /* Set up OST pool environment */
1943 lod->lod_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
1944 HASH_POOLS_MAX_BITS,
1945 HASH_POOLS_BKT_BITS, 0,
1948 &pool_hash_operations,
1950 if (lod->lod_pools_hash_body == NULL)
1953 INIT_LIST_HEAD(&lod->lod_pool_list);
1954 lod->lod_pool_count = 0;
1955 rc = lod_ost_pool_init(&lod->lod_pool_info, 0);
1958 lod_qos_rr_init(&lod->lod_qos.lq_rr);
1959 rc = lod_ost_pool_init(&lod->lod_qos.lq_rr.lqr_pool, 0);
1961 GOTO(out_pool_info, rc);
1966 lod_ost_pool_free(&lod->lod_pool_info);
1968 cfs_hash_putref(lod->lod_pools_hash_body);
1974 * Release the structures describing the pools.
1976 * \param[in] lod LOD device from which we release the structures
1980 int lod_pools_fini(struct lod_device *lod)
1982 struct obd_device *obd = lod2obd(lod);
1983 struct pool_desc *pool, *tmp;
1986 list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
1987 /* free pool structs */
1988 CDEBUG(D_INFO, "delete pool %p\n", pool);
1989 /* In the function below, .hs_keycmp resolves to
1990 * pool_hashkey_keycmp() */
1991 /* coverity[overrun-buffer-val] */
1992 lod_pool_del(obd, pool->pool_name);
1995 cfs_hash_putref(lod->lod_pools_hash_body);
1996 lod_ost_pool_free(&(lod->lod_qos.lq_rr.lqr_pool));
1997 lod_ost_pool_free(&lod->lod_pool_info);