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, 2014, 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>
43 #include "lod_internal.h"
46 * Increase reference count on the target table.
48 * Increase reference count on the target table usage to prevent racing with
49 * addition/deletion. Any function that expects the table to remain
50 * stationary must take a ref.
52 * \param[in] ltd target table (lod_ost_descs or lod_mdt_descs)
54 void lod_getref(struct lod_tgt_descs *ltd)
56 down_read(<d->ltd_rw_sem);
57 mutex_lock(<d->ltd_mutex);
59 mutex_unlock(<d->ltd_mutex);
63 * Decrease reference count on the target table.
65 * Companion of lod_getref() to release a reference on the target table.
66 * If this is the last reference and the OST entry was scheduled for deletion,
67 * the descriptor is removed from the table.
69 * \param[in] lod LOD device from which we release a reference
70 * \param[in] ltd target table (lod_ost_descs or lod_mdt_descs)
72 void lod_putref(struct lod_device *lod, struct lod_tgt_descs *ltd)
74 mutex_lock(<d->ltd_mutex);
76 if (ltd->ltd_refcount == 0 && ltd->ltd_death_row) {
77 struct lod_tgt_desc *tgt_desc, *tmp;
78 struct list_head kill;
81 CDEBUG(D_CONFIG, "destroying %d ltd desc\n",
84 INIT_LIST_HEAD(&kill);
86 cfs_foreach_bit(ltd->ltd_tgt_bitmap, idx) {
87 tgt_desc = LTD_TGT(ltd, idx);
90 if (!tgt_desc->ltd_reap)
93 list_add(&tgt_desc->ltd_kill, &kill);
94 LTD_TGT(ltd, idx) = NULL;
95 /*FIXME: only support ost pool for now */
96 if (ltd == &lod->lod_ost_descs) {
97 lod_ost_pool_remove(&lod->lod_pool_info, idx);
98 if (tgt_desc->ltd_active)
99 lod->lod_desc.ld_active_tgt_count--;
102 cfs_bitmap_clear(ltd->ltd_tgt_bitmap, idx);
103 ltd->ltd_death_row--;
105 mutex_unlock(<d->ltd_mutex);
106 up_read(<d->ltd_rw_sem);
108 list_for_each_entry_safe(tgt_desc, tmp, &kill, ltd_kill) {
110 list_del(&tgt_desc->ltd_kill);
111 if (ltd == &lod->lod_ost_descs) {
112 /* remove from QoS structures */
113 rc = qos_del_tgt(lod, tgt_desc);
115 CERROR("%s: qos_del_tgt(%s) failed:"
117 lod2obd(lod)->obd_name,
118 obd_uuid2str(&tgt_desc->ltd_uuid),
121 rc = obd_disconnect(tgt_desc->ltd_exp);
123 CERROR("%s: failed to disconnect %s: rc = %d\n",
124 lod2obd(lod)->obd_name,
125 obd_uuid2str(&tgt_desc->ltd_uuid), rc);
126 OBD_FREE_PTR(tgt_desc);
129 mutex_unlock(<d->ltd_mutex);
130 up_read(<d->ltd_rw_sem);
135 * Expand size of target table.
137 * When the target table is full, we have to extend the table. To do so,
138 * we allocate new memory with some reserve, move data from the old table
139 * to the new one and release memory consumed by the old table.
140 * Notice we take ltd_rw_sem exclusively to ensure atomic switch.
142 * \param[in] ltd target table
143 * \param[in] newsize new size of the table
145 * \retval 0 on success
146 * \retval -ENOMEM if reallocation failed
148 static int ltd_bitmap_resize(struct lod_tgt_descs *ltd, __u32 newsize)
150 cfs_bitmap_t *new_bitmap, *old_bitmap = NULL;
154 /* grab write reference on the lod. Relocating the array requires
155 * exclusive access */
157 down_write(<d->ltd_rw_sem);
158 if (newsize <= ltd->ltd_tgts_size)
159 /* someone else has already resize the array */
162 /* allocate new bitmap */
163 new_bitmap = CFS_ALLOCATE_BITMAP(newsize);
165 GOTO(out, rc = -ENOMEM);
167 if (ltd->ltd_tgts_size > 0) {
168 /* the bitmap already exists, we need
169 * to copy data from old one */
170 cfs_bitmap_copy(new_bitmap, ltd->ltd_tgt_bitmap);
171 old_bitmap = ltd->ltd_tgt_bitmap;
174 ltd->ltd_tgts_size = newsize;
175 ltd->ltd_tgt_bitmap = new_bitmap;
178 CFS_FREE_BITMAP(old_bitmap);
180 CDEBUG(D_CONFIG, "tgt size: %d\n", ltd->ltd_tgts_size);
184 up_write(<d->ltd_rw_sem);
189 * Connect LOD to a new OSP and add it to the target table.
191 * Connect to the OSP device passed, initialize all the internal
192 * structures related to the device and add it to the target table.
194 * \param[in] env execution environment for this thread
195 * \param[in] lod LOD device to be connected to the new OSP
196 * \param[in] osp name of OSP device name to be added
197 * \param[in] index index of the new target
198 * \param[in] gen target's generation number
199 * \param[in] tgt_index OSP's group
200 * \param[in] type type of device (mdc or osc)
201 * \param[in] active state of OSP: 0 - inactive, 1 - active
203 * \retval 0 if added successfully
204 * \retval negative error number on failure
206 int lod_add_device(const struct lu_env *env, struct lod_device *lod,
207 char *osp, unsigned index, unsigned gen, int tgt_index,
208 char *type, int active)
210 struct obd_connect_data *data = NULL;
211 struct obd_export *exp = NULL;
212 struct obd_device *obd;
213 struct lu_device *ldev;
216 struct lod_tgt_desc *tgt_desc;
217 struct lod_tgt_descs *ltd;
218 struct obd_uuid obd_uuid;
222 CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, index, gen);
225 CERROR("request to add OBD %s with invalid generation: %d\n",
230 obd_str2uuid(&obd_uuid, osp);
232 obd = class_find_client_obd(&obd_uuid, LUSTRE_OSP_NAME,
233 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
235 CERROR("can't find %s device\n", osp);
243 data->ocd_connect_flags = OBD_CONNECT_INDEX | OBD_CONNECT_VERSION;
244 data->ocd_version = LUSTRE_VERSION_CODE;
245 data->ocd_index = index;
247 if (strcmp(LUSTRE_OSC_NAME, type) == 0) {
249 data->ocd_connect_flags |= OBD_CONNECT_AT |
252 #ifdef HAVE_LRU_RESIZE_SUPPORT
253 OBD_CONNECT_LRU_RESIZE |
256 OBD_CONNECT_OSS_CAPA |
257 OBD_CONNECT_REQPORTAL |
258 OBD_CONNECT_SKIP_ORPHAN |
260 OBD_CONNECT_LVB_TYPE |
261 OBD_CONNECT_VERSION |
262 OBD_CONNECT_PINGLESS |
265 data->ocd_group = tgt_index;
266 ltd = &lod->lod_ost_descs;
268 struct obd_import *imp = obd->u.cli.cl_import;
271 data->ocd_ibits_known = MDS_INODELOCK_UPDATE;
272 data->ocd_connect_flags |= OBD_CONNECT_ACL |
273 OBD_CONNECT_MDS_CAPA |
274 OBD_CONNECT_OSS_CAPA |
276 OBD_CONNECT_MDS_MDS |
281 /* XXX set MDS-MDS flags, remove this when running this
283 data->ocd_connect_flags |= OBD_CONNECT_MDS_MDS;
284 spin_lock(&imp->imp_lock);
285 imp->imp_server_timeout = 1;
286 spin_unlock(&imp->imp_lock);
287 imp->imp_client->cli_request_portal = OUT_PORTAL;
288 CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
290 ltd = &lod->lod_mdt_descs;
293 rc = obd_connect(env, &exp, obd, &obd->obd_uuid, data, NULL);
296 CERROR("%s: cannot connect to next dev %s (%d)\n",
297 obd->obd_name, osp, rc);
301 LASSERT(obd->obd_lu_dev);
302 LASSERT(obd->obd_lu_dev->ld_site == lod->lod_dt_dev.dd_lu_dev.ld_site);
304 ldev = obd->obd_lu_dev;
307 /* Allocate ost descriptor and fill it */
308 OBD_ALLOC_PTR(tgt_desc);
310 GOTO(out_conn, rc = -ENOMEM);
312 tgt_desc->ltd_tgt = d;
313 tgt_desc->ltd_exp = exp;
314 tgt_desc->ltd_uuid = obd->u.cli.cl_target_uuid;
315 tgt_desc->ltd_gen = gen;
316 tgt_desc->ltd_index = index;
317 tgt_desc->ltd_active = active;
320 if (index >= ltd->ltd_tgts_size) {
321 /* we have to increase the size of the lod_osts array */
324 newsize = max(ltd->ltd_tgts_size, (__u32)2);
325 while (newsize < index + 1)
326 newsize = newsize << 1;
328 /* lod_bitmap_resize() needs lod_rw_sem
329 * which we hold with th reference */
330 lod_putref(lod, ltd);
332 rc = ltd_bitmap_resize(ltd, newsize);
339 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);
355 if (!strcmp(LUSTRE_OSC_NAME, type)) {
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);
384 if (lod->lod_recovery_completed)
385 ldev->ld_ops->ldo_recovery_complete(env, ldev);
387 rc = lfsck_add_target(env, lod->lod_child, d, exp, index, for_ost);
389 CERROR("Fail to add LFSCK target: name = %s, type = %s, "
390 "index = %u, rc = %d\n", osp, type, index, rc);
395 lod_ost_pool_remove(&lod->lod_pool_info, index);
397 mutex_unlock(<d->ltd_mutex);
398 lod_putref(lod, ltd);
400 OBD_FREE_PTR(tgt_desc);
408 * Schedule target removal from the target table.
410 * Mark the device as dead. The device is not removed here because it may
411 * still be in use. The device will be removed in lod_putref() when the
412 * last reference is released.
414 * \param[in] env execution environment for this thread
415 * \param[in] lod LOD device the target table belongs to
416 * \param[in] ltd target table
417 * \param[in] idx index of the target
418 * \param[in] for_ost type of the target: 0 - MDT, 1 - OST
420 static void __lod_del_device(const struct lu_env *env, struct lod_device *lod,
421 struct lod_tgt_descs *ltd, unsigned idx,
424 LASSERT(LTD_TGT(ltd, idx));
426 lfsck_del_target(env, lod->lod_child, LTD_TGT(ltd, idx)->ltd_tgt,
429 if (LTD_TGT(ltd, idx)->ltd_reap == 0) {
430 LTD_TGT(ltd, idx)->ltd_reap = 1;
431 ltd->ltd_death_row++;
436 * Schedule removal of all the targets from the given target table.
438 * See more details in the description for __lod_del_device()
440 * \param[in] env execution environment for this thread
441 * \param[in] lod LOD device the target table belongs to
442 * \param[in] ltd target table
443 * \param[in] for_ost type of the target: MDT or OST
447 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
448 struct lod_tgt_descs *ltd, bool for_ost)
452 if (ltd->ltd_tgts_size <= 0)
455 mutex_lock(<d->ltd_mutex);
456 cfs_foreach_bit(ltd->ltd_tgt_bitmap, idx)
457 __lod_del_device(env, lod, ltd, idx, for_ost);
458 mutex_unlock(<d->ltd_mutex);
459 lod_putref(lod, ltd);
460 CFS_FREE_BITMAP(ltd->ltd_tgt_bitmap);
461 for (idx = 0; idx < TGT_PTRS; idx++) {
462 if (ltd->ltd_tgt_idx[idx])
463 OBD_FREE_PTR(ltd->ltd_tgt_idx[idx]);
465 ltd->ltd_tgts_size = 0;
470 * Remove device by name.
472 * Remove a device identified by \a osp from the target table. Given
473 * the device can be in use, the real deletion happens in lod_putref().
475 * \param[in] env execution environment for this thread
476 * \param[in] lod LOD device to be connected to the new OSP
477 * \param[in] ltd target table
478 * \param[in] osp name of OSP device to be removed
479 * \param[in] idx index of the target
480 * \param[in] gen generation number, not used currently
481 * \param[in] for_ost type of the target: 0 - MDT, 1 - OST
483 * \retval 0 if the device was scheduled for removal
484 * \retval -EINVAL if no device was found
486 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
487 struct lod_tgt_descs *ltd, char *osp, unsigned idx,
488 unsigned gen, bool for_ost)
490 struct obd_device *obd;
492 struct obd_uuid uuid;
495 CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, idx, gen);
497 obd_str2uuid(&uuid, osp);
499 obd = class_find_client_obd(&uuid, LUSTRE_OSP_NAME,
500 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
502 CERROR("can't find %s device\n", osp);
507 CERROR("%s: request to remove OBD %s with invalid generation %d"
508 "\n", obd->obd_name, osp, gen);
512 obd_str2uuid(&uuid, osp);
515 mutex_lock(<d->ltd_mutex);
516 /* check that the index is allocated in the bitmap */
517 if (!cfs_bitmap_check(ltd->ltd_tgt_bitmap, idx) ||
518 !LTD_TGT(ltd, idx)) {
519 CERROR("%s: device %d is not set up\n", obd->obd_name, idx);
520 GOTO(out, rc = -EINVAL);
523 /* check that the UUID matches */
524 if (!obd_uuid_equals(&uuid, <D_TGT(ltd, idx)->ltd_uuid)) {
525 CERROR("%s: LOD target UUID %s at index %d does not match %s\n",
526 obd->obd_name, obd_uuid2str(<D_TGT(ltd,idx)->ltd_uuid),
528 GOTO(out, rc = -EINVAL);
531 __lod_del_device(env, lod, ltd, idx, for_ost);
534 mutex_unlock(<d->ltd_mutex);
535 lod_putref(lod, ltd);
540 * Resize per-thread storage to hold specified size.
542 * A helper function to resize per-thread temporary storage. This storage
543 * is used to process LOV/LVM EAs and may be quite large. We do not want to
544 * allocate/release it every time, so instead we put it into the env and
545 * reallocate on demand. The memory is released when the correspondent thread
548 * \param[in] info LOD-specific storage in the environment
549 * \param[in] size new size to grow the buffer to
551 * \retval 0 on success, -ENOMEM if reallocation failed
553 int lod_ea_store_resize(struct lod_thread_info *info, size_t size)
555 __u32 round = size_roundup_power2(size);
558 lov_mds_md_size(LOV_MAX_STRIPE_COUNT, LOV_MAGIC_V3));
559 if (info->lti_ea_store) {
560 LASSERT(info->lti_ea_store_size);
561 LASSERT(info->lti_ea_store_size < round);
562 CDEBUG(D_INFO, "EA store size %d is not enough, need %d\n",
563 info->lti_ea_store_size, round);
564 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
565 info->lti_ea_store = NULL;
566 info->lti_ea_store_size = 0;
569 OBD_ALLOC_LARGE(info->lti_ea_store, round);
570 if (info->lti_ea_store == NULL)
572 info->lti_ea_store_size = round;
577 * Make LOV EA for striped object.
579 * Generate striping information and store it in the LOV EA of the given
580 * object. The caller must ensure nobody else is calling the function
581 * against the object concurrently. The transaction must be started.
582 * FLDB service must be running as well; it's used to map FID to the target,
583 * which is stored in LOV EA.
585 * \param[in] env execution environment for this thread
586 * \param[in] lo LOD object
587 * \param[in] th transaction handle
589 * \retval 0 if LOV EA is stored successfully
590 * \retval negative error number on failure
592 int lod_generate_and_set_lovea(const struct lu_env *env,
593 struct lod_object *lo, struct thandle *th)
595 struct lod_thread_info *info = lod_env_info(env);
596 struct dt_object *next = dt_object_child(&lo->ldo_obj);
597 const struct lu_fid *fid = lu_object_fid(&lo->ldo_obj.do_lu);
598 struct lov_mds_md_v1 *lmm;
599 struct lov_ost_data_v1 *objs;
607 magic = lo->ldo_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
608 lmm_size = lov_mds_md_size(lo->ldo_stripenr, magic);
609 if (info->lti_ea_store_size < lmm_size) {
610 rc = lod_ea_store_resize(info, lmm_size);
615 if (lo->ldo_pattern == 0) /* default striping */
616 lo->ldo_pattern = LOV_PATTERN_RAID0;
618 lmm = info->lti_ea_store;
620 lmm->lmm_magic = cpu_to_le32(magic);
621 lmm->lmm_pattern = cpu_to_le32(lo->ldo_pattern);
622 fid_to_lmm_oi(fid, &lmm->lmm_oi);
623 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
624 lmm->lmm_oi.oi.oi_id++;
625 lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
626 lmm->lmm_stripe_size = cpu_to_le32(lo->ldo_stripe_size);
627 lmm->lmm_stripe_count = cpu_to_le16(lo->ldo_stripenr);
628 if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED)
629 lmm->lmm_stripe_count = cpu_to_le16(lo->ldo_released_stripenr);
630 lmm->lmm_layout_gen = 0;
631 if (magic == LOV_MAGIC_V1) {
632 objs = &lmm->lmm_objects[0];
634 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *) lmm;
635 size_t cplen = strlcpy(v3->lmm_pool_name, lo->ldo_pool,
636 sizeof(v3->lmm_pool_name));
637 if (cplen >= sizeof(v3->lmm_pool_name))
639 objs = &v3->lmm_objects[0];
642 for (i = 0; i < lo->ldo_stripenr; i++) {
643 struct lu_fid *fid = &info->lti_fid;
644 struct lod_device *lod;
646 int type = LU_SEQ_RANGE_OST;
648 lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
649 LASSERT(lo->ldo_stripe[i]);
651 *fid = *lu_object_fid(&lo->ldo_stripe[i]->do_lu);
652 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF)) {
653 if (cfs_fail_val == 0)
654 cfs_fail_val = fid->f_oid;
656 fid->f_oid = cfs_fail_val;
659 rc = fid_to_ostid(fid, &info->lti_ostid);
662 ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
663 objs[i].l_ost_gen = cpu_to_le32(0);
664 rc = lod_fld_lookup(env, lod, fid, &index, &type);
666 CERROR("%s: Can not locate "DFID": rc = %d\n",
667 lod2obd(lod)->obd_name, PFID(fid), rc);
668 lod_object_free_striping(env, lo);
671 objs[i].l_ost_idx = cpu_to_le32(index);
674 info->lti_buf.lb_buf = lmm;
675 info->lti_buf.lb_len = lmm_size;
676 rc = dt_xattr_set(env, next, &info->lti_buf, XATTR_NAME_LOV, 0,
679 lod_object_free_striping(env, lo);
687 * Fill lti_ea_store buffer in the environment with a value for the given
688 * EA. The buffer is reallocated if the value doesn't fit.
690 * \param[in,out] env execution environment for this thread
691 * .lti_ea_store buffer is filled with EA's value
692 * \param[in] lo LOD object
693 * \param[in] name name of the EA
695 * \retval 0 if EA is fetched successfully
696 * \retval negative error number on failure
698 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
701 struct lod_thread_info *info = lod_env_info(env);
702 struct dt_object *next = dt_object_child(&lo->ldo_obj);
708 if (unlikely(info->lti_ea_store == NULL)) {
709 /* just to enter in allocation block below */
713 info->lti_buf.lb_buf = info->lti_ea_store;
714 info->lti_buf.lb_len = info->lti_ea_store_size;
715 rc = dt_xattr_get(env, next, &info->lti_buf, name, BYPASS_CAPA);
718 /* if object is not striped or inaccessible */
719 if (rc == -ENODATA || rc == -ENOENT)
723 /* EA doesn't fit, reallocate new buffer */
724 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name,
726 if (rc == -ENODATA || rc == -ENOENT)
732 rc = lod_ea_store_resize(info, rc);
742 * Store default striping.
744 * Store default striping for the files in the given directory. The data
745 * are stored in the LOD-object representing the directory (ldo_def_* fields).
746 * If default striping matches virtual fs-wide default striping, then we
747 * store nothing. This mean that the files in the directory will be created
748 * with filesystem-wide default striping. The transaction must be started.
750 * \param[in] env execution environment for this thread
751 * \param[in] dt dt object representing directory in LOD layer
752 * \param[in] th transaction handle
754 * \retval 0 if stored successfully or no need to store
755 * \retval negative error number on failure
757 int lod_store_def_striping(const struct lu_env *env, struct dt_object *dt,
760 struct lod_thread_info *info = lod_env_info(env);
761 struct lod_object *lo = lod_dt_obj(dt);
762 struct dt_object *next = dt_object_child(dt);
763 struct lov_user_md_v3 *v3;
767 if (S_ISDIR(dt->do_lu.lo_header->loh_attr))
770 * store striping defaults into new directory
771 * used to implement defaults inheritance
774 /* probably nothing to inherite */
775 if (lo->ldo_striping_cached == 0)
778 if (LOVEA_DELETE_VALUES(lo->ldo_def_stripe_size, lo->ldo_def_stripenr,
779 lo->ldo_def_stripe_offset))
782 v3 = info->lti_ea_store;
783 if (info->lti_ea_store_size < sizeof(*v3)) {
784 rc = lod_ea_store_resize(info, sizeof(*v3));
787 v3 = info->lti_ea_store;
789 memset(v3, 0, sizeof(*v3));
790 v3->lmm_magic = cpu_to_le32(LOV_USER_MAGIC_V3);
791 v3->lmm_stripe_count = cpu_to_le16(lo->ldo_def_stripenr);
792 v3->lmm_stripe_offset = cpu_to_le16(lo->ldo_def_stripe_offset);
793 v3->lmm_stripe_size = cpu_to_le32(lo->ldo_def_stripe_size);
794 if (lo->ldo_pool != NULL) {
795 strlcpy(v3->lmm_pool_name, lo->ldo_pool,
796 sizeof(v3->lmm_pool_name));
797 v3->lmm_pool_name[sizeof(v3->lmm_pool_name) - 1] = '\0';
799 info->lti_buf.lb_buf = v3;
800 info->lti_buf.lb_len = sizeof(*v3);
801 rc = dt_xattr_set(env, next, &info->lti_buf, XATTR_NAME_LOV, 0, th,
808 * Verify the target index is present in the current configuration.
810 * \param[in] md LOD device where the target table is stored
811 * \param[in] idx target's index
813 * \retval 0 if the index is present
814 * \retval -EINVAL if not
816 static int validate_lod_and_idx(struct lod_device *md, __u32 idx)
818 if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
819 !cfs_bitmap_check(md->lod_ost_bitmap, idx))) {
820 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
821 md->lod_ost_descs.ltd_tgts_size);
825 if (unlikely(OST_TGT(md, idx) == NULL)) {
826 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
827 lod2obd(md)->obd_name, idx);
831 if (unlikely(OST_TGT(md, idx)->ltd_ost == NULL)) {
832 CERROR("%s: invalid lod device, for idx: %d\n",
833 lod2obd(md)->obd_name , idx);
841 * Instantiate objects for stripes.
843 * Allocate and initialize LU-objects representing the stripes. The number
844 * of the stripes (ldo_stripenr) must be initialized already. The caller
845 * must ensure nobody else is calling the function on the object at the same
846 * time. FLDB service must be running to be able to map a FID to the targets
847 * and find appropriate device representing that target.
849 * \param[in] env execution environment for this thread
850 * \param[in,out] lo LOD object
851 * \param[in] objs an array of IDs to creates the objects from
853 * \retval 0 if the objects are instantiated successfully
854 * \retval negative error number on failure
856 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
857 struct lov_ost_data_v1 *objs)
859 struct lod_thread_info *info = lod_env_info(env);
860 struct lod_device *md;
861 struct lu_object *o, *n;
862 struct lu_device *nd;
863 struct dt_object **stripe;
870 md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
871 LASSERT(lo->ldo_stripe == NULL);
872 LASSERT(lo->ldo_stripenr > 0);
873 LASSERT(lo->ldo_stripe_size > 0);
875 stripe_len = lo->ldo_stripenr;
876 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
880 for (i = 0; i < lo->ldo_stripenr; i++) {
881 if (unlikely(lovea_slot_is_dummy(&objs[i])))
884 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
885 idx = le32_to_cpu(objs[i].l_ost_idx);
886 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
889 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
890 PFID(&info->lti_fid));
891 lod_getref(&md->lod_ost_descs);
893 rc = validate_lod_and_idx(md, idx);
894 if (unlikely(rc != 0)) {
895 lod_putref(md, &md->lod_ost_descs);
899 nd = &OST_TGT(md,idx)->ltd_ost->dd_lu_dev;
900 lod_putref(md, &md->lod_ost_descs);
902 /* In the function below, .hs_keycmp resolves to
903 * u_obj_hop_keycmp() */
904 /* coverity[overrun-buffer-val] */
905 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
907 GOTO(out, rc = PTR_ERR(o));
909 n = lu_object_locate(o->lo_header, nd->ld_type);
912 stripe[i] = container_of(n, struct dt_object, do_lu);
917 for (i = 0; i < stripe_len; i++)
918 if (stripe[i] != NULL)
919 lu_object_put(env, &stripe[i]->do_lu);
921 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
922 lo->ldo_stripenr = 0;
924 lo->ldo_stripe = stripe;
925 lo->ldo_stripes_allocated = stripe_len;
932 * Instantiate objects for striping.
934 * Parse striping information in \a buf and instantiate the objects
935 * representing the stripes.
937 * \param[in] env execution environment for this thread
938 * \param[in] lo LOD object
939 * \param[in] buf buffer storing LOV EA to parse
941 * \retval 0 if parsing and objects creation succeed
942 * \retval negative error number on failure
944 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
945 const struct lu_buf *buf)
947 struct lov_mds_md_v1 *lmm;
948 struct lov_ost_data_v1 *objs;
955 LASSERT(buf->lb_buf);
956 LASSERT(buf->lb_len);
958 lmm = (struct lov_mds_md_v1 *) buf->lb_buf;
959 magic = le32_to_cpu(lmm->lmm_magic);
960 pattern = le32_to_cpu(lmm->lmm_pattern);
962 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)
963 GOTO(out, rc = -EINVAL);
964 if (lov_pattern(pattern) != LOV_PATTERN_RAID0)
965 GOTO(out, rc = -EINVAL);
967 lo->ldo_pattern = pattern;
968 lo->ldo_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
969 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
970 lo->ldo_stripenr = le16_to_cpu(lmm->lmm_stripe_count);
971 /* released file stripenr fixup. */
972 if (pattern & LOV_PATTERN_F_RELEASED)
973 lo->ldo_stripenr = 0;
975 LASSERT(buf->lb_len >= lov_mds_md_size(lo->ldo_stripenr, magic));
977 if (magic == LOV_MAGIC_V3) {
978 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *) lmm;
979 objs = &v3->lmm_objects[0];
980 lod_object_set_pool(lo, v3->lmm_pool_name);
982 objs = &lmm->lmm_objects[0];
985 if (lo->ldo_stripenr > 0)
986 rc = lod_initialize_objects(env, lo, objs);
993 * Initialize the object representing the stripes.
995 * Unless the stripes are initialized already, fetch LOV (for regular
996 * objects) or LMV (for directory objects) EA and call lod_parse_striping()
997 * to instantiate the objects representing the stripes.
999 * \param[in] env execution environment for this thread
1000 * \param[in,out] lo LOD object
1002 * \retval 0 if parsing and object creation succeed
1003 * \retval negative error number on failure
1005 int lod_load_striping_locked(const struct lu_env *env, struct lod_object *lo)
1007 struct lod_thread_info *info = lod_env_info(env);
1008 struct lu_buf *buf = &info->lti_buf;
1009 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1013 /* already initialized? */
1014 if (lo->ldo_stripe != NULL)
1017 if (!dt_object_exists(next))
1020 /* Do not load stripe for slaves of striped dir */
1021 if (lo->ldo_dir_slave_stripe)
1024 if (S_ISREG(lu_object_attr(lod2lu_obj(lo)))) {
1025 rc = lod_get_lov_ea(env, lo);
1029 * there is LOV EA (striping information) in this object
1030 * let's parse it and create in-core objects for the stripes
1032 buf->lb_buf = info->lti_ea_store;
1033 buf->lb_len = info->lti_ea_store_size;
1034 rc = lod_parse_striping(env, lo, buf);
1035 } else if (S_ISDIR(lu_object_attr(lod2lu_obj(lo)))) {
1036 rc = lod_get_lmv_ea(env, lo);
1037 if (rc < (typeof(rc))sizeof(struct lmv_mds_md_v1))
1038 GOTO(out, rc = rc > 0 ? -EINVAL : rc);
1040 buf->lb_buf = info->lti_ea_store;
1041 buf->lb_len = info->lti_ea_store_size;
1042 if (rc == sizeof(struct lmv_mds_md_v1)) {
1043 rc = lod_load_lmv_shards(env, lo, buf, true);
1044 if (buf->lb_buf != info->lti_ea_store) {
1045 OBD_FREE_LARGE(info->lti_ea_store,
1046 info->lti_ea_store_size);
1047 info->lti_ea_store = buf->lb_buf;
1048 info->lti_ea_store_size = buf->lb_len;
1056 * there is LOV EA (striping information) in this object
1057 * let's parse it and create in-core objects for the stripes
1059 rc = lod_parse_dir_striping(env, lo, buf);
1066 * A generic function to initialize the stripe objects.
1068 * A protected version of lod_load_striping_locked() - load the striping
1069 * information from storage, parse that and instantiate LU objects to
1070 * represent the stripes. The LOD object \a lo supplies a pointer to the
1071 * next sub-object in the LU stack so we can lock it. Also use \a lo to
1072 * return an array of references to the newly instantiated objects.
1074 * \param[in] env execution environment for this thread
1075 * \param[in,out] lo LOD object, where striping is stored and
1076 * which gets an array of references
1078 * \retval 0 if parsing and object creation succeed
1079 * \retval negative error number on failure
1081 int lod_load_striping(const struct lu_env *env, struct lod_object *lo)
1083 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1086 /* currently this code is supposed to be called from declaration
1087 * phase only, thus the object is not expected to be locked by caller */
1088 dt_write_lock(env, next, 0);
1089 rc = lod_load_striping_locked(env, lo);
1090 dt_write_unlock(env, next);
1097 * Check the validity of all fields including the magic, stripe size,
1098 * stripe count, stripe offset and that the pool is present. Also check
1099 * that each target index points to an existing target. The additional
1100 * \a is_from_disk turns additional checks. In some cases zero fields
1101 * are allowed (like pattern=0).
1103 * \param[in] d LOD device
1104 * \param[in] buf buffer with LOV EA to verify
1105 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
1106 * 1 - from disk, do not allow
1108 * \retval 0 if the striping is valid
1109 * \retval -EINVAL if striping is invalid
1111 int lod_verify_striping(struct lod_device *d, const struct lu_buf *buf,
1114 struct lov_user_md_v1 *lum;
1115 struct lov_user_md_v3 *lum3;
1116 struct pool_desc *pool = NULL;
1120 __u16 stripe_offset;
1127 LASSERT(sizeof(*lum) < sizeof(*lum3));
1129 if (buf->lb_len < sizeof(*lum)) {
1130 CDEBUG(D_IOCTL, "buf len %zu too small for lov_user_md\n",
1132 GOTO(out, rc = -EINVAL);
1135 magic = le32_to_cpu(lum->lmm_magic);
1136 if (magic != LOV_USER_MAGIC_V1 &&
1137 magic != LOV_USER_MAGIC_V3 &&
1138 magic != LOV_MAGIC_V1_DEF &&
1139 magic != LOV_MAGIC_V3_DEF) {
1140 CDEBUG(D_IOCTL, "bad userland LOV MAGIC: %#x\n", magic);
1141 GOTO(out, rc = -EINVAL);
1144 /* the user uses "0" for default stripe pattern normally. */
1145 if (!is_from_disk && lum->lmm_pattern == 0)
1146 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1148 if (le32_to_cpu(lum->lmm_pattern) != LOV_PATTERN_RAID0) {
1149 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
1150 le32_to_cpu(lum->lmm_pattern));
1151 GOTO(out, rc = -EINVAL);
1154 /* 64kB is the largest common page size we see (ia64), and matches the
1156 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1157 if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1158 CDEBUG(D_IOCTL, "stripe size %u not a multiple of %u\n",
1159 stripe_size, LOV_MIN_STRIPE_SIZE);
1160 GOTO(out, rc = -EINVAL);
1163 stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1164 if (stripe_offset != LOV_OFFSET_DEFAULT) {
1165 /* if offset is not within valid range [0, osts_size) */
1166 if (stripe_offset >= d->lod_osts_size) {
1167 CDEBUG(D_IOCTL, "stripe offset %u >= bitmap size %u\n",
1168 stripe_offset, d->lod_osts_size);
1169 GOTO(out, rc = -EINVAL);
1172 /* if lmm_stripe_offset is *not* in bitmap */
1173 if (!cfs_bitmap_check(d->lod_ost_bitmap, stripe_offset)) {
1174 CDEBUG(D_IOCTL, "stripe offset %u not in bitmap\n",
1176 GOTO(out, rc = -EINVAL);
1180 if (magic == LOV_USER_MAGIC_V1 || magic == LOV_MAGIC_V1_DEF)
1181 lum_size = offsetof(struct lov_user_md_v1,
1183 else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_MAGIC_V3_DEF)
1184 lum_size = offsetof(struct lov_user_md_v3,
1187 GOTO(out, rc = -EINVAL);
1189 stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1190 if (buf->lb_len != lum_size) {
1191 CDEBUG(D_IOCTL, "invalid buf len %zu for lov_user_md with "
1192 "magic %#x and stripe_count %u\n",
1193 buf->lb_len, magic, stripe_count);
1194 GOTO(out, rc = -EINVAL);
1197 if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_MAGIC_V3_DEF))
1201 if (buf->lb_len < sizeof(*lum3)) {
1202 CDEBUG(D_IOCTL, "buf len %zu too small for lov_user_md_v3\n",
1204 GOTO(out, rc = -EINVAL);
1207 /* In the function below, .hs_keycmp resolves to
1208 * pool_hashkey_keycmp() */
1209 /* coverity[overrun-buffer-val] */
1210 pool = lod_find_pool(d, lum3->lmm_pool_name);
1214 if (stripe_offset != LOV_OFFSET_DEFAULT) {
1215 rc = lod_check_index_in_pool(stripe_offset, pool);
1217 GOTO(out, rc = -EINVAL);
1220 if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1222 "stripe count %u > # OSTs %u in the pool\n",
1223 stripe_count, pool_tgt_count(pool));
1224 GOTO(out, rc = -EINVAL);
1229 lod_pool_putref(pool);
1234 void lod_fix_desc_stripe_size(__u64 *val)
1236 if (*val < LOV_MIN_STRIPE_SIZE) {
1238 LCONSOLE_INFO("Increasing default stripe size to "
1239 "minimum value %u\n",
1240 LOV_DESC_STRIPE_SIZE_DEFAULT);
1241 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
1242 } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
1243 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
1244 LCONSOLE_WARN("Changing default stripe size to "LPU64" (a "
1245 "multiple of %u)\n",
1246 *val, LOV_MIN_STRIPE_SIZE);
1250 void lod_fix_desc_stripe_count(__u32 *val)
1256 void lod_fix_desc_pattern(__u32 *val)
1258 /* from lov_setstripe */
1259 if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
1260 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
1265 void lod_fix_desc_qos_maxage(__u32 *val)
1267 /* fix qos_maxage */
1269 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
1273 * Used to fix insane default striping.
1275 * \param[in] desc striping description
1277 void lod_fix_desc(struct lov_desc *desc)
1279 lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
1280 lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
1281 lod_fix_desc_pattern(&desc->ld_pattern);
1282 lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
1286 * Initialize the structures used to store pools and default striping.
1288 * \param[in] lod LOD device
1289 * \param[in] lcfg configuration structure storing default striping.
1291 * \retval 0 if initialization succeeds
1292 * \retval negative error number on failure
1294 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
1296 struct obd_device *obd;
1297 struct lov_desc *desc;
1301 obd = class_name2obd(lustre_cfg_string(lcfg, 0));
1302 LASSERT(obd != NULL);
1303 obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1305 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1306 CERROR("LOD setup requires a descriptor\n");
1310 desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
1312 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1313 CERROR("descriptor size wrong: %d > %d\n",
1314 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1318 if (desc->ld_magic != LOV_DESC_MAGIC) {
1319 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
1320 CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
1321 obd->obd_name, desc);
1322 lustre_swab_lov_desc(desc);
1324 CERROR("%s: Bad lov desc magic: %#x\n",
1325 obd->obd_name, desc->ld_magic);
1332 desc->ld_active_tgt_count = 0;
1333 lod->lod_desc = *desc;
1335 lod->lod_sp_me = LUSTRE_SP_CLI;
1337 /* Set up allocation policy (QoS and RR) */
1338 INIT_LIST_HEAD(&lod->lod_qos.lq_oss_list);
1339 init_rwsem(&lod->lod_qos.lq_rw_sem);
1340 lod->lod_qos.lq_dirty = 1;
1341 lod->lod_qos.lq_rr.lqr_dirty = 1;
1342 lod->lod_qos.lq_reset = 1;
1343 /* Default priority is toward free space balance */
1344 lod->lod_qos.lq_prio_free = 232;
1345 /* Default threshold for rr (roughly 17%) */
1346 lod->lod_qos.lq_threshold_rr = 43;
1348 /* Set up OST pool environment */
1349 lod->lod_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
1350 HASH_POOLS_MAX_BITS,
1351 HASH_POOLS_BKT_BITS, 0,
1354 &pool_hash_operations,
1356 if (lod->lod_pools_hash_body == NULL)
1359 INIT_LIST_HEAD(&lod->lod_pool_list);
1360 lod->lod_pool_count = 0;
1361 rc = lod_ost_pool_init(&lod->lod_pool_info, 0);
1364 rc = lod_ost_pool_init(&lod->lod_qos.lq_rr.lqr_pool, 0);
1366 GOTO(out_pool_info, rc);
1371 lod_ost_pool_free(&lod->lod_pool_info);
1373 cfs_hash_putref(lod->lod_pools_hash_body);
1379 * Release the structures describing the pools.
1381 * \param[in] lod LOD device from which we release the structures
1385 int lod_pools_fini(struct lod_device *lod)
1387 struct obd_device *obd = lod2obd(lod);
1388 struct pool_desc *pool, *tmp;
1391 list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
1392 /* free pool structs */
1393 CDEBUG(D_INFO, "delete pool %p\n", pool);
1394 /* In the function below, .hs_keycmp resolves to
1395 * pool_hashkey_keycmp() */
1396 /* coverity[overrun-buffer-val] */
1397 lod_pool_del(obd, pool->pool_name);
1400 cfs_hash_putref(lod->lod_pools_hash_body);
1401 lod_ost_pool_free(&(lod->lod_qos.lq_rr.lqr_pool));
1402 lod_ost_pool_free(&lod->lod_pool_info);