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_REQPORTAL |
257 OBD_CONNECT_SKIP_ORPHAN |
259 OBD_CONNECT_LVB_TYPE |
260 OBD_CONNECT_VERSION |
261 OBD_CONNECT_PINGLESS |
264 data->ocd_group = tgt_index;
265 ltd = &lod->lod_ost_descs;
267 struct obd_import *imp = obd->u.cli.cl_import;
270 data->ocd_ibits_known = MDS_INODELOCK_UPDATE;
271 data->ocd_connect_flags |= OBD_CONNECT_ACL |
273 OBD_CONNECT_MDS_MDS |
278 spin_lock(&imp->imp_lock);
279 imp->imp_server_timeout = 1;
280 spin_unlock(&imp->imp_lock);
281 imp->imp_client->cli_request_portal = OUT_PORTAL;
282 CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
284 ltd = &lod->lod_mdt_descs;
287 rc = obd_connect(env, &exp, obd, &obd->obd_uuid, data, NULL);
290 CERROR("%s: cannot connect to next dev %s (%d)\n",
291 obd->obd_name, osp, rc);
295 LASSERT(obd->obd_lu_dev);
296 LASSERT(obd->obd_lu_dev->ld_site == lod->lod_dt_dev.dd_lu_dev.ld_site);
298 ldev = obd->obd_lu_dev;
301 /* Allocate ost descriptor and fill it */
302 OBD_ALLOC_PTR(tgt_desc);
304 GOTO(out_conn, rc = -ENOMEM);
306 tgt_desc->ltd_tgt = d;
307 tgt_desc->ltd_exp = exp;
308 tgt_desc->ltd_uuid = obd->u.cli.cl_target_uuid;
309 tgt_desc->ltd_gen = gen;
310 tgt_desc->ltd_index = index;
311 tgt_desc->ltd_active = active;
314 if (index >= ltd->ltd_tgts_size) {
315 /* we have to increase the size of the lod_osts array */
318 newsize = max(ltd->ltd_tgts_size, (__u32)2);
319 while (newsize < index + 1)
320 newsize = newsize << 1;
322 /* lod_bitmap_resize() needs lod_rw_sem
323 * which we hold with th reference */
324 lod_putref(lod, ltd);
326 rc = ltd_bitmap_resize(ltd, newsize);
333 mutex_lock(<d->ltd_mutex);
334 if (cfs_bitmap_check(ltd->ltd_tgt_bitmap, index)) {
335 CERROR("%s: device %d is registered already\n", obd->obd_name,
337 GOTO(out_mutex, rc = -EEXIST);
340 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
341 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
342 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
343 CERROR("can't allocate index to add %s\n",
345 GOTO(out_mutex, rc = -ENOMEM);
350 /* pool and qos are not supported for MDS stack yet */
351 rc = lod_ost_pool_add(&lod->lod_pool_info, index,
354 CERROR("%s: can't set up pool, failed with %d\n",
359 rc = qos_add_tgt(lod, tgt_desc);
361 CERROR("%s: qos_add_tgt failed with %d\n",
366 /* The new OST is now a full citizen */
367 if (index >= lod->lod_desc.ld_tgt_count)
368 lod->lod_desc.ld_tgt_count = index + 1;
370 lod->lod_desc.ld_active_tgt_count++;
373 LTD_TGT(ltd, index) = tgt_desc;
374 cfs_bitmap_set(ltd->ltd_tgt_bitmap, index);
376 mutex_unlock(<d->ltd_mutex);
377 lod_putref(lod, ltd);
378 if (lod->lod_recovery_completed)
379 ldev->ld_ops->ldo_recovery_complete(env, ldev);
381 if (!for_ost && lod->lod_initialized) {
382 rc = lod_sub_init_llog(env, lod, tgt_desc->ltd_tgt);
384 CERROR("%s: cannot start llog on %s:rc = %d\n",
385 lod2obd(lod)->obd_name, osp, rc);
390 rc = lfsck_add_target(env, lod->lod_child, d, exp, index, for_ost);
392 CERROR("Fail to add LFSCK target: name = %s, type = %s, "
393 "index = %u, rc = %d\n", osp, type, index, rc);
394 GOTO(out_fini_llog, rc);
398 lod_sub_fini_llog(env, tgt_desc->ltd_tgt,
399 tgt_desc->ltd_recovery_thread);
401 lod_ost_pool_remove(&lod->lod_pool_info, index);
403 mutex_unlock(<d->ltd_mutex);
404 lod_putref(lod, ltd);
406 OBD_FREE_PTR(tgt_desc);
414 * Schedule target removal from the target table.
416 * Mark the device as dead. The device is not removed here because it may
417 * still be in use. The device will be removed in lod_putref() when the
418 * last reference is released.
420 * \param[in] env execution environment for this thread
421 * \param[in] lod LOD device the target table belongs to
422 * \param[in] ltd target table
423 * \param[in] idx index of the target
424 * \param[in] for_ost type of the target: 0 - MDT, 1 - OST
426 static void __lod_del_device(const struct lu_env *env, struct lod_device *lod,
427 struct lod_tgt_descs *ltd, unsigned idx,
430 LASSERT(LTD_TGT(ltd, idx));
432 lfsck_del_target(env, lod->lod_child, LTD_TGT(ltd, idx)->ltd_tgt,
435 if (!for_ost && LTD_TGT(ltd, idx)->ltd_recovery_thread != NULL) {
436 struct ptlrpc_thread *thread;
438 thread = LTD_TGT(ltd, idx)->ltd_recovery_thread;
439 OBD_FREE_PTR(thread);
442 if (LTD_TGT(ltd, idx)->ltd_reap == 0) {
443 LTD_TGT(ltd, idx)->ltd_reap = 1;
444 ltd->ltd_death_row++;
449 * Schedule removal of all the targets from the given target table.
451 * See more details in the description for __lod_del_device()
453 * \param[in] env execution environment for this thread
454 * \param[in] lod LOD device the target table belongs to
455 * \param[in] ltd target table
456 * \param[in] for_ost type of the target: MDT or OST
460 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
461 struct lod_tgt_descs *ltd, bool for_ost)
465 if (ltd->ltd_tgts_size <= 0)
468 mutex_lock(<d->ltd_mutex);
469 cfs_foreach_bit(ltd->ltd_tgt_bitmap, idx)
470 __lod_del_device(env, lod, ltd, idx, for_ost);
471 mutex_unlock(<d->ltd_mutex);
472 lod_putref(lod, ltd);
473 CFS_FREE_BITMAP(ltd->ltd_tgt_bitmap);
474 for (idx = 0; idx < TGT_PTRS; idx++) {
475 if (ltd->ltd_tgt_idx[idx])
476 OBD_FREE_PTR(ltd->ltd_tgt_idx[idx]);
478 ltd->ltd_tgts_size = 0;
483 * Remove device by name.
485 * Remove a device identified by \a osp from the target table. Given
486 * the device can be in use, the real deletion happens in lod_putref().
488 * \param[in] env execution environment for this thread
489 * \param[in] lod LOD device to be connected to the new OSP
490 * \param[in] ltd target table
491 * \param[in] osp name of OSP device to be removed
492 * \param[in] idx index of the target
493 * \param[in] gen generation number, not used currently
494 * \param[in] for_ost type of the target: 0 - MDT, 1 - OST
496 * \retval 0 if the device was scheduled for removal
497 * \retval -EINVAL if no device was found
499 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
500 struct lod_tgt_descs *ltd, char *osp, unsigned idx,
501 unsigned gen, bool for_ost)
503 struct obd_device *obd;
505 struct obd_uuid uuid;
508 CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, idx, gen);
510 obd_str2uuid(&uuid, osp);
512 obd = class_find_client_obd(&uuid, LUSTRE_OSP_NAME,
513 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
515 CERROR("can't find %s device\n", osp);
520 CERROR("%s: request to remove OBD %s with invalid generation %d"
521 "\n", obd->obd_name, osp, gen);
525 obd_str2uuid(&uuid, osp);
528 mutex_lock(<d->ltd_mutex);
529 /* check that the index is allocated in the bitmap */
530 if (!cfs_bitmap_check(ltd->ltd_tgt_bitmap, idx) ||
531 !LTD_TGT(ltd, idx)) {
532 CERROR("%s: device %d is not set up\n", obd->obd_name, idx);
533 GOTO(out, rc = -EINVAL);
536 /* check that the UUID matches */
537 if (!obd_uuid_equals(&uuid, <D_TGT(ltd, idx)->ltd_uuid)) {
538 CERROR("%s: LOD target UUID %s at index %d does not match %s\n",
539 obd->obd_name, obd_uuid2str(<D_TGT(ltd,idx)->ltd_uuid),
541 GOTO(out, rc = -EINVAL);
544 __lod_del_device(env, lod, ltd, idx, for_ost);
547 mutex_unlock(<d->ltd_mutex);
548 lod_putref(lod, ltd);
553 * Resize per-thread storage to hold specified size.
555 * A helper function to resize per-thread temporary storage. This storage
556 * is used to process LOV/LVM EAs and may be quite large. We do not want to
557 * allocate/release it every time, so instead we put it into the env and
558 * reallocate on demand. The memory is released when the correspondent thread
561 * \param[in] info LOD-specific storage in the environment
562 * \param[in] size new size to grow the buffer to
564 * \retval 0 on success, -ENOMEM if reallocation failed
566 int lod_ea_store_resize(struct lod_thread_info *info, size_t size)
568 __u32 round = size_roundup_power2(size);
571 lov_mds_md_size(LOV_MAX_STRIPE_COUNT, LOV_MAGIC_V3));
572 if (info->lti_ea_store) {
573 LASSERT(info->lti_ea_store_size);
574 LASSERT(info->lti_ea_store_size < round);
575 CDEBUG(D_INFO, "EA store size %d is not enough, need %d\n",
576 info->lti_ea_store_size, round);
577 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
578 info->lti_ea_store = NULL;
579 info->lti_ea_store_size = 0;
582 OBD_ALLOC_LARGE(info->lti_ea_store, round);
583 if (info->lti_ea_store == NULL)
585 info->lti_ea_store_size = round;
590 * Make LOV EA for striped object.
592 * Generate striping information and store it in the LOV EA of the given
593 * object. The caller must ensure nobody else is calling the function
594 * against the object concurrently. The transaction must be started.
595 * FLDB service must be running as well; it's used to map FID to the target,
596 * which is stored in LOV EA.
598 * \param[in] env execution environment for this thread
599 * \param[in] lo LOD object
600 * \param[in] th transaction handle
602 * \retval 0 if LOV EA is stored successfully
603 * \retval negative error number on failure
605 int lod_generate_and_set_lovea(const struct lu_env *env,
606 struct lod_object *lo, struct thandle *th)
608 struct lod_thread_info *info = lod_env_info(env);
609 struct dt_object *next = dt_object_child(&lo->ldo_obj);
610 const struct lu_fid *fid = lu_object_fid(&lo->ldo_obj.do_lu);
611 struct lov_mds_md_v1 *lmm;
612 struct lov_ost_data_v1 *objs;
620 magic = lo->ldo_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
621 lmm_size = lov_mds_md_size(lo->ldo_stripenr, magic);
622 if (info->lti_ea_store_size < lmm_size) {
623 rc = lod_ea_store_resize(info, lmm_size);
628 if (lo->ldo_pattern == 0) /* default striping */
629 lo->ldo_pattern = LOV_PATTERN_RAID0;
631 lmm = info->lti_ea_store;
633 lmm->lmm_magic = cpu_to_le32(magic);
634 lmm->lmm_pattern = cpu_to_le32(lo->ldo_pattern);
635 fid_to_lmm_oi(fid, &lmm->lmm_oi);
636 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
637 lmm->lmm_oi.oi.oi_id++;
638 lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
639 lmm->lmm_stripe_size = cpu_to_le32(lo->ldo_stripe_size);
640 lmm->lmm_stripe_count = cpu_to_le16(lo->ldo_stripenr);
641 if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED)
642 lmm->lmm_stripe_count = cpu_to_le16(lo->ldo_released_stripenr);
643 lmm->lmm_layout_gen = 0;
644 if (magic == LOV_MAGIC_V1) {
645 objs = &lmm->lmm_objects[0];
647 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *) lmm;
648 size_t cplen = strlcpy(v3->lmm_pool_name, lo->ldo_pool,
649 sizeof(v3->lmm_pool_name));
650 if (cplen >= sizeof(v3->lmm_pool_name))
652 objs = &v3->lmm_objects[0];
655 for (i = 0; i < lo->ldo_stripenr; i++) {
656 struct lu_fid *fid = &info->lti_fid;
657 struct lod_device *lod;
659 int type = LU_SEQ_RANGE_OST;
661 lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
662 LASSERT(lo->ldo_stripe[i]);
664 *fid = *lu_object_fid(&lo->ldo_stripe[i]->do_lu);
665 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF)) {
666 if (cfs_fail_val == 0)
667 cfs_fail_val = fid->f_oid;
669 fid->f_oid = cfs_fail_val;
672 rc = fid_to_ostid(fid, &info->lti_ostid);
675 ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
676 objs[i].l_ost_gen = cpu_to_le32(0);
677 rc = lod_fld_lookup(env, lod, fid, &index, &type);
679 CERROR("%s: Can not locate "DFID": rc = %d\n",
680 lod2obd(lod)->obd_name, PFID(fid), rc);
681 lod_object_free_striping(env, lo);
684 objs[i].l_ost_idx = cpu_to_le32(index);
687 info->lti_buf.lb_buf = lmm;
688 info->lti_buf.lb_len = lmm_size;
689 rc = lod_sub_object_xattr_set(env, next, &info->lti_buf, XATTR_NAME_LOV,
692 lod_object_free_striping(env, lo);
702 * Fill lti_ea_store buffer in the environment with a value for the given
703 * EA. The buffer is reallocated if the value doesn't fit.
705 * \param[in,out] env execution environment for this thread
706 * .lti_ea_store buffer is filled with EA's value
707 * \param[in] lo LOD object
708 * \param[in] name name of the EA
710 * \retval 0 if EA is fetched successfully
711 * \retval negative error number on failure
713 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
716 struct lod_thread_info *info = lod_env_info(env);
717 struct dt_object *next = dt_object_child(&lo->ldo_obj);
723 if (unlikely(info->lti_ea_store == NULL)) {
724 /* just to enter in allocation block below */
728 info->lti_buf.lb_buf = info->lti_ea_store;
729 info->lti_buf.lb_len = info->lti_ea_store_size;
730 rc = dt_xattr_get(env, next, &info->lti_buf, name);
733 /* if object is not striped or inaccessible */
734 if (rc == -ENODATA || rc == -ENOENT)
738 /* EA doesn't fit, reallocate new buffer */
739 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name);
740 if (rc == -ENODATA || rc == -ENOENT)
746 rc = lod_ea_store_resize(info, rc);
756 * Verify the target index is present in the current configuration.
758 * \param[in] md LOD device where the target table is stored
759 * \param[in] idx target's index
761 * \retval 0 if the index is present
762 * \retval -EINVAL if not
764 static int validate_lod_and_idx(struct lod_device *md, __u32 idx)
766 if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
767 !cfs_bitmap_check(md->lod_ost_bitmap, idx))) {
768 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
769 md->lod_ost_descs.ltd_tgts_size);
773 if (unlikely(OST_TGT(md, idx) == NULL)) {
774 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
775 lod2obd(md)->obd_name, idx);
779 if (unlikely(OST_TGT(md, idx)->ltd_ost == NULL)) {
780 CERROR("%s: invalid lod device, for idx: %d\n",
781 lod2obd(md)->obd_name , idx);
789 * Instantiate objects for stripes.
791 * Allocate and initialize LU-objects representing the stripes. The number
792 * of the stripes (ldo_stripenr) must be initialized already. The caller
793 * must ensure nobody else is calling the function on the object at the same
794 * time. FLDB service must be running to be able to map a FID to the targets
795 * and find appropriate device representing that target.
797 * \param[in] env execution environment for this thread
798 * \param[in,out] lo LOD object
799 * \param[in] objs an array of IDs to creates the objects from
801 * \retval 0 if the objects are instantiated successfully
802 * \retval negative error number on failure
804 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
805 struct lov_ost_data_v1 *objs)
807 struct lod_thread_info *info = lod_env_info(env);
808 struct lod_device *md;
809 struct lu_object *o, *n;
810 struct lu_device *nd;
811 struct dt_object **stripe;
818 md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
819 LASSERT(lo->ldo_stripe == NULL);
820 LASSERT(lo->ldo_stripenr > 0);
821 LASSERT(lo->ldo_stripe_size > 0);
823 stripe_len = lo->ldo_stripenr;
824 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
828 for (i = 0; i < lo->ldo_stripenr; i++) {
829 if (unlikely(lovea_slot_is_dummy(&objs[i])))
832 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
833 idx = le32_to_cpu(objs[i].l_ost_idx);
834 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
837 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
838 PFID(&info->lti_fid));
839 lod_getref(&md->lod_ost_descs);
841 rc = validate_lod_and_idx(md, idx);
842 if (unlikely(rc != 0)) {
843 lod_putref(md, &md->lod_ost_descs);
847 nd = &OST_TGT(md,idx)->ltd_ost->dd_lu_dev;
848 lod_putref(md, &md->lod_ost_descs);
850 /* In the function below, .hs_keycmp resolves to
851 * u_obj_hop_keycmp() */
852 /* coverity[overrun-buffer-val] */
853 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
855 GOTO(out, rc = PTR_ERR(o));
857 n = lu_object_locate(o->lo_header, nd->ld_type);
860 stripe[i] = container_of(n, struct dt_object, do_lu);
865 for (i = 0; i < stripe_len; i++)
866 if (stripe[i] != NULL)
867 lu_object_put(env, &stripe[i]->do_lu);
869 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
870 lo->ldo_stripenr = 0;
872 lo->ldo_stripe = stripe;
873 lo->ldo_stripes_allocated = stripe_len;
880 * Instantiate objects for striping.
882 * Parse striping information in \a buf and instantiate the objects
883 * representing the stripes.
885 * \param[in] env execution environment for this thread
886 * \param[in] lo LOD object
887 * \param[in] buf buffer storing LOV EA to parse
889 * \retval 0 if parsing and objects creation succeed
890 * \retval negative error number on failure
892 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
893 const struct lu_buf *buf)
895 struct lov_mds_md_v1 *lmm;
896 struct lov_ost_data_v1 *objs;
903 LASSERT(buf->lb_buf);
904 LASSERT(buf->lb_len);
906 lmm = (struct lov_mds_md_v1 *) buf->lb_buf;
907 magic = le32_to_cpu(lmm->lmm_magic);
908 pattern = le32_to_cpu(lmm->lmm_pattern);
910 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)
911 GOTO(out, rc = -EINVAL);
912 if (lov_pattern(pattern) != LOV_PATTERN_RAID0)
913 GOTO(out, rc = -EINVAL);
915 lo->ldo_pattern = pattern;
916 lo->ldo_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
917 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
918 lo->ldo_stripenr = le16_to_cpu(lmm->lmm_stripe_count);
919 /* released file stripenr fixup. */
920 if (pattern & LOV_PATTERN_F_RELEASED)
921 lo->ldo_stripenr = 0;
923 LASSERT(buf->lb_len >= lov_mds_md_size(lo->ldo_stripenr, magic));
925 if (magic == LOV_MAGIC_V3) {
926 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *) lmm;
927 objs = &v3->lmm_objects[0];
928 lod_object_set_pool(lo, v3->lmm_pool_name);
930 objs = &lmm->lmm_objects[0];
933 if (lo->ldo_stripenr > 0)
934 rc = lod_initialize_objects(env, lo, objs);
941 * Initialize the object representing the stripes.
943 * Unless the stripes are initialized already, fetch LOV (for regular
944 * objects) or LMV (for directory objects) EA and call lod_parse_striping()
945 * to instantiate the objects representing the stripes.
947 * \param[in] env execution environment for this thread
948 * \param[in,out] lo LOD object
950 * \retval 0 if parsing and object creation succeed
951 * \retval negative error number on failure
953 int lod_load_striping_locked(const struct lu_env *env, struct lod_object *lo)
955 struct lod_thread_info *info = lod_env_info(env);
956 struct lu_buf *buf = &info->lti_buf;
957 struct dt_object *next = dt_object_child(&lo->ldo_obj);
961 /* already initialized? */
962 if (lo->ldo_stripe != NULL)
965 if (!dt_object_exists(next))
968 /* Do not load stripe for slaves of striped dir */
969 if (lo->ldo_dir_slave_stripe)
972 if (S_ISREG(lu_object_attr(lod2lu_obj(lo)))) {
973 rc = lod_get_lov_ea(env, lo);
977 * there is LOV EA (striping information) in this object
978 * let's parse it and create in-core objects for the stripes
980 buf->lb_buf = info->lti_ea_store;
981 buf->lb_len = info->lti_ea_store_size;
982 rc = lod_parse_striping(env, lo, buf);
983 } else if (S_ISDIR(lu_object_attr(lod2lu_obj(lo)))) {
984 rc = lod_get_lmv_ea(env, lo);
985 if (rc < (typeof(rc))sizeof(struct lmv_mds_md_v1))
986 GOTO(out, rc = rc > 0 ? -EINVAL : rc);
988 buf->lb_buf = info->lti_ea_store;
989 buf->lb_len = info->lti_ea_store_size;
990 if (rc == sizeof(struct lmv_mds_md_v1)) {
991 rc = lod_load_lmv_shards(env, lo, buf, true);
992 if (buf->lb_buf != info->lti_ea_store) {
993 OBD_FREE_LARGE(info->lti_ea_store,
994 info->lti_ea_store_size);
995 info->lti_ea_store = buf->lb_buf;
996 info->lti_ea_store_size = buf->lb_len;
1004 * there is LOV EA (striping information) in this object
1005 * let's parse it and create in-core objects for the stripes
1007 rc = lod_parse_dir_striping(env, lo, buf);
1011 lo->ldo_striping_cached = 1;
1017 * A generic function to initialize the stripe objects.
1019 * A protected version of lod_load_striping_locked() - load the striping
1020 * information from storage, parse that and instantiate LU objects to
1021 * represent the stripes. The LOD object \a lo supplies a pointer to the
1022 * next sub-object in the LU stack so we can lock it. Also use \a lo to
1023 * return an array of references to the newly instantiated objects.
1025 * \param[in] env execution environment for this thread
1026 * \param[in,out] lo LOD object, where striping is stored and
1027 * which gets an array of references
1029 * \retval 0 if parsing and object creation succeed
1030 * \retval negative error number on failure
1032 int lod_load_striping(const struct lu_env *env, struct lod_object *lo)
1034 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1037 /* currently this code is supposed to be called from declaration
1038 * phase only, thus the object is not expected to be locked by caller */
1039 dt_write_lock(env, next, 0);
1040 rc = lod_load_striping_locked(env, lo);
1041 dt_write_unlock(env, next);
1048 * Check the validity of all fields including the magic, stripe size,
1049 * stripe count, stripe offset and that the pool is present. Also check
1050 * that each target index points to an existing target. The additional
1051 * \a is_from_disk turns additional checks. In some cases zero fields
1052 * are allowed (like pattern=0).
1054 * \param[in] d LOD device
1055 * \param[in] buf buffer with LOV EA to verify
1056 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
1057 * 1 - from disk, do not allow
1059 * \retval 0 if the striping is valid
1060 * \retval -EINVAL if striping is invalid
1062 int lod_verify_striping(struct lod_device *d, const struct lu_buf *buf,
1065 struct lov_user_md_v1 *lum;
1066 struct lov_user_md_v3 *lum3;
1067 struct pool_desc *pool = NULL;
1071 __u16 stripe_offset;
1078 LASSERT(sizeof(*lum) < sizeof(*lum3));
1080 if (buf->lb_len < sizeof(*lum)) {
1081 CDEBUG(D_IOCTL, "buf len %zu too small for lov_user_md\n",
1083 GOTO(out, rc = -EINVAL);
1086 magic = le32_to_cpu(lum->lmm_magic);
1087 if (magic != LOV_USER_MAGIC_V1 &&
1088 magic != LOV_USER_MAGIC_V3 &&
1089 magic != LOV_MAGIC_V1_DEF &&
1090 magic != LOV_MAGIC_V3_DEF) {
1091 CDEBUG(D_IOCTL, "bad userland LOV MAGIC: %#x\n", magic);
1092 GOTO(out, rc = -EINVAL);
1095 /* the user uses "0" for default stripe pattern normally. */
1096 if (!is_from_disk && lum->lmm_pattern == 0)
1097 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1099 if (le32_to_cpu(lum->lmm_pattern) != LOV_PATTERN_RAID0) {
1100 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
1101 le32_to_cpu(lum->lmm_pattern));
1102 GOTO(out, rc = -EINVAL);
1105 /* 64kB is the largest common page size we see (ia64), and matches the
1107 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1108 if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1109 CDEBUG(D_IOCTL, "stripe size %u not a multiple of %u\n",
1110 stripe_size, LOV_MIN_STRIPE_SIZE);
1111 GOTO(out, rc = -EINVAL);
1114 stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1115 if (stripe_offset != LOV_OFFSET_DEFAULT) {
1116 /* if offset is not within valid range [0, osts_size) */
1117 if (stripe_offset >= d->lod_osts_size) {
1118 CDEBUG(D_IOCTL, "stripe offset %u >= bitmap size %u\n",
1119 stripe_offset, d->lod_osts_size);
1120 GOTO(out, rc = -EINVAL);
1123 /* if lmm_stripe_offset is *not* in bitmap */
1124 if (!cfs_bitmap_check(d->lod_ost_bitmap, stripe_offset)) {
1125 CDEBUG(D_IOCTL, "stripe offset %u not in bitmap\n",
1127 GOTO(out, rc = -EINVAL);
1131 if (magic == LOV_USER_MAGIC_V1 || magic == LOV_MAGIC_V1_DEF)
1132 lum_size = offsetof(struct lov_user_md_v1,
1134 else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_MAGIC_V3_DEF)
1135 lum_size = offsetof(struct lov_user_md_v3,
1138 GOTO(out, rc = -EINVAL);
1140 stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1141 if (buf->lb_len != lum_size) {
1142 CDEBUG(D_IOCTL, "invalid buf len %zu for lov_user_md with "
1143 "magic %#x and stripe_count %u\n",
1144 buf->lb_len, magic, stripe_count);
1145 GOTO(out, rc = -EINVAL);
1148 if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_MAGIC_V3_DEF))
1152 if (buf->lb_len < sizeof(*lum3)) {
1153 CDEBUG(D_IOCTL, "buf len %zu too small for lov_user_md_v3\n",
1155 GOTO(out, rc = -EINVAL);
1158 /* In the function below, .hs_keycmp resolves to
1159 * pool_hashkey_keycmp() */
1160 /* coverity[overrun-buffer-val] */
1161 pool = lod_find_pool(d, lum3->lmm_pool_name);
1165 if (stripe_offset != LOV_OFFSET_DEFAULT) {
1166 rc = lod_check_index_in_pool(stripe_offset, pool);
1168 GOTO(out, rc = -EINVAL);
1171 if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1173 "stripe count %u > # OSTs %u in the pool\n",
1174 stripe_count, pool_tgt_count(pool));
1175 GOTO(out, rc = -EINVAL);
1180 lod_pool_putref(pool);
1185 void lod_fix_desc_stripe_size(__u64 *val)
1187 if (*val < LOV_MIN_STRIPE_SIZE) {
1189 LCONSOLE_INFO("Increasing default stripe size to "
1190 "minimum value %u\n",
1191 LOV_DESC_STRIPE_SIZE_DEFAULT);
1192 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
1193 } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
1194 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
1195 LCONSOLE_WARN("Changing default stripe size to "LPU64" (a "
1196 "multiple of %u)\n",
1197 *val, LOV_MIN_STRIPE_SIZE);
1201 void lod_fix_desc_stripe_count(__u32 *val)
1207 void lod_fix_desc_pattern(__u32 *val)
1209 /* from lov_setstripe */
1210 if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
1211 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
1216 void lod_fix_desc_qos_maxage(__u32 *val)
1218 /* fix qos_maxage */
1220 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
1224 * Used to fix insane default striping.
1226 * \param[in] desc striping description
1228 void lod_fix_desc(struct lov_desc *desc)
1230 lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
1231 lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
1232 lod_fix_desc_pattern(&desc->ld_pattern);
1233 lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
1237 * Initialize the structures used to store pools and default striping.
1239 * \param[in] lod LOD device
1240 * \param[in] lcfg configuration structure storing default striping.
1242 * \retval 0 if initialization succeeds
1243 * \retval negative error number on failure
1245 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
1247 struct obd_device *obd;
1248 struct lov_desc *desc;
1252 obd = class_name2obd(lustre_cfg_string(lcfg, 0));
1253 LASSERT(obd != NULL);
1254 obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1256 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1257 CERROR("LOD setup requires a descriptor\n");
1261 desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
1263 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1264 CERROR("descriptor size wrong: %d > %d\n",
1265 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1269 if (desc->ld_magic != LOV_DESC_MAGIC) {
1270 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
1271 CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
1272 obd->obd_name, desc);
1273 lustre_swab_lov_desc(desc);
1275 CERROR("%s: Bad lov desc magic: %#x\n",
1276 obd->obd_name, desc->ld_magic);
1283 desc->ld_active_tgt_count = 0;
1284 lod->lod_desc = *desc;
1286 lod->lod_sp_me = LUSTRE_SP_CLI;
1288 /* Set up allocation policy (QoS and RR) */
1289 INIT_LIST_HEAD(&lod->lod_qos.lq_oss_list);
1290 init_rwsem(&lod->lod_qos.lq_rw_sem);
1291 lod->lod_qos.lq_dirty = 1;
1292 lod->lod_qos.lq_rr.lqr_dirty = 1;
1293 lod->lod_qos.lq_reset = 1;
1294 /* Default priority is toward free space balance */
1295 lod->lod_qos.lq_prio_free = 232;
1296 /* Default threshold for rr (roughly 17%) */
1297 lod->lod_qos.lq_threshold_rr = 43;
1299 /* Set up OST pool environment */
1300 lod->lod_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
1301 HASH_POOLS_MAX_BITS,
1302 HASH_POOLS_BKT_BITS, 0,
1305 &pool_hash_operations,
1307 if (lod->lod_pools_hash_body == NULL)
1310 INIT_LIST_HEAD(&lod->lod_pool_list);
1311 lod->lod_pool_count = 0;
1312 rc = lod_ost_pool_init(&lod->lod_pool_info, 0);
1315 rc = lod_ost_pool_init(&lod->lod_qos.lq_rr.lqr_pool, 0);
1317 GOTO(out_pool_info, rc);
1322 lod_ost_pool_free(&lod->lod_pool_info);
1324 cfs_hash_putref(lod->lod_pools_hash_body);
1330 * Release the structures describing the pools.
1332 * \param[in] lod LOD device from which we release the structures
1336 int lod_pools_fini(struct lod_device *lod)
1338 struct obd_device *obd = lod2obd(lod);
1339 struct pool_desc *pool, *tmp;
1342 list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
1343 /* free pool structs */
1344 CDEBUG(D_INFO, "delete pool %p\n", pool);
1345 /* In the function below, .hs_keycmp resolves to
1346 * pool_hashkey_keycmp() */
1347 /* coverity[overrun-buffer-val] */
1348 lod_pool_del(obd, pool->pool_name);
1351 cfs_hash_putref(lod->lod_pools_hash_body);
1352 lod_ost_pool_free(&(lod->lod_qos.lq_rr.lqr_pool));
1353 lod_ost_pool_free(&lod->lod_pool_info);