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 spin_lock(&imp->imp_lock);
282 imp->imp_server_timeout = 1;
283 spin_unlock(&imp->imp_lock);
284 imp->imp_client->cli_request_portal = OUT_PORTAL;
285 CDEBUG(D_OTHER, "%s: Set 'mds' portal and timeout\n",
287 ltd = &lod->lod_mdt_descs;
290 rc = obd_connect(env, &exp, obd, &obd->obd_uuid, data, NULL);
293 CERROR("%s: cannot connect to next dev %s (%d)\n",
294 obd->obd_name, osp, rc);
298 LASSERT(obd->obd_lu_dev);
299 LASSERT(obd->obd_lu_dev->ld_site == lod->lod_dt_dev.dd_lu_dev.ld_site);
301 ldev = obd->obd_lu_dev;
304 /* Allocate ost descriptor and fill it */
305 OBD_ALLOC_PTR(tgt_desc);
307 GOTO(out_conn, rc = -ENOMEM);
309 tgt_desc->ltd_tgt = d;
310 tgt_desc->ltd_exp = exp;
311 tgt_desc->ltd_uuid = obd->u.cli.cl_target_uuid;
312 tgt_desc->ltd_gen = gen;
313 tgt_desc->ltd_index = index;
314 tgt_desc->ltd_active = active;
317 if (index >= ltd->ltd_tgts_size) {
318 /* we have to increase the size of the lod_osts array */
321 newsize = max(ltd->ltd_tgts_size, (__u32)2);
322 while (newsize < index + 1)
323 newsize = newsize << 1;
325 /* lod_bitmap_resize() needs lod_rw_sem
326 * which we hold with th reference */
327 lod_putref(lod, ltd);
329 rc = ltd_bitmap_resize(ltd, newsize);
336 mutex_lock(<d->ltd_mutex);
337 if (cfs_bitmap_check(ltd->ltd_tgt_bitmap, index)) {
338 CERROR("%s: device %d is registered already\n", obd->obd_name,
340 GOTO(out_mutex, rc = -EEXIST);
343 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
344 OBD_ALLOC_PTR(ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK]);
345 if (ltd->ltd_tgt_idx[index / TGT_PTRS_PER_BLOCK] == NULL) {
346 CERROR("can't allocate index to add %s\n",
348 GOTO(out_mutex, rc = -ENOMEM);
352 if (!strcmp(LUSTRE_OSC_NAME, type)) {
353 /* pool and qos are not supported for MDS stack yet */
354 rc = lod_ost_pool_add(&lod->lod_pool_info, index,
357 CERROR("%s: can't set up pool, failed with %d\n",
362 rc = qos_add_tgt(lod, tgt_desc);
364 CERROR("%s: qos_add_tgt failed with %d\n",
369 /* The new OST is now a full citizen */
370 if (index >= lod->lod_desc.ld_tgt_count)
371 lod->lod_desc.ld_tgt_count = index + 1;
373 lod->lod_desc.ld_active_tgt_count++;
376 LTD_TGT(ltd, index) = tgt_desc;
377 cfs_bitmap_set(ltd->ltd_tgt_bitmap, index);
379 mutex_unlock(<d->ltd_mutex);
380 lod_putref(lod, ltd);
381 if (lod->lod_recovery_completed)
382 ldev->ld_ops->ldo_recovery_complete(env, ldev);
384 rc = lfsck_add_target(env, lod->lod_child, d, exp, index, for_ost);
386 CERROR("Fail to add LFSCK target: name = %s, type = %s, "
387 "index = %u, rc = %d\n", osp, type, index, rc);
392 lod_ost_pool_remove(&lod->lod_pool_info, index);
394 mutex_unlock(<d->ltd_mutex);
395 lod_putref(lod, ltd);
397 OBD_FREE_PTR(tgt_desc);
405 * Schedule target removal from the target table.
407 * Mark the device as dead. The device is not removed here because it may
408 * still be in use. The device will be removed in lod_putref() when the
409 * last reference is released.
411 * \param[in] env execution environment for this thread
412 * \param[in] lod LOD device the target table belongs to
413 * \param[in] ltd target table
414 * \param[in] idx index of the target
415 * \param[in] for_ost type of the target: 0 - MDT, 1 - OST
417 static void __lod_del_device(const struct lu_env *env, struct lod_device *lod,
418 struct lod_tgt_descs *ltd, unsigned idx,
421 LASSERT(LTD_TGT(ltd, idx));
423 lfsck_del_target(env, lod->lod_child, LTD_TGT(ltd, idx)->ltd_tgt,
426 if (LTD_TGT(ltd, idx)->ltd_reap == 0) {
427 LTD_TGT(ltd, idx)->ltd_reap = 1;
428 ltd->ltd_death_row++;
433 * Schedule removal of all the targets from the given target table.
435 * See more details in the description for __lod_del_device()
437 * \param[in] env execution environment for this thread
438 * \param[in] lod LOD device the target table belongs to
439 * \param[in] ltd target table
440 * \param[in] for_ost type of the target: MDT or OST
444 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
445 struct lod_tgt_descs *ltd, bool for_ost)
449 if (ltd->ltd_tgts_size <= 0)
452 mutex_lock(<d->ltd_mutex);
453 cfs_foreach_bit(ltd->ltd_tgt_bitmap, idx)
454 __lod_del_device(env, lod, ltd, idx, for_ost);
455 mutex_unlock(<d->ltd_mutex);
456 lod_putref(lod, ltd);
457 CFS_FREE_BITMAP(ltd->ltd_tgt_bitmap);
458 for (idx = 0; idx < TGT_PTRS; idx++) {
459 if (ltd->ltd_tgt_idx[idx])
460 OBD_FREE_PTR(ltd->ltd_tgt_idx[idx]);
462 ltd->ltd_tgts_size = 0;
467 * Remove device by name.
469 * Remove a device identified by \a osp from the target table. Given
470 * the device can be in use, the real deletion happens in lod_putref().
472 * \param[in] env execution environment for this thread
473 * \param[in] lod LOD device to be connected to the new OSP
474 * \param[in] ltd target table
475 * \param[in] osp name of OSP device to be removed
476 * \param[in] idx index of the target
477 * \param[in] gen generation number, not used currently
478 * \param[in] for_ost type of the target: 0 - MDT, 1 - OST
480 * \retval 0 if the device was scheduled for removal
481 * \retval -EINVAL if no device was found
483 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
484 struct lod_tgt_descs *ltd, char *osp, unsigned idx,
485 unsigned gen, bool for_ost)
487 struct obd_device *obd;
489 struct obd_uuid uuid;
492 CDEBUG(D_CONFIG, "osp:%s idx:%d gen:%d\n", osp, idx, gen);
494 obd_str2uuid(&uuid, osp);
496 obd = class_find_client_obd(&uuid, LUSTRE_OSP_NAME,
497 &lod->lod_dt_dev.dd_lu_dev.ld_obd->obd_uuid);
499 CERROR("can't find %s device\n", osp);
504 CERROR("%s: request to remove OBD %s with invalid generation %d"
505 "\n", obd->obd_name, osp, gen);
509 obd_str2uuid(&uuid, osp);
512 mutex_lock(<d->ltd_mutex);
513 /* check that the index is allocated in the bitmap */
514 if (!cfs_bitmap_check(ltd->ltd_tgt_bitmap, idx) ||
515 !LTD_TGT(ltd, idx)) {
516 CERROR("%s: device %d is not set up\n", obd->obd_name, idx);
517 GOTO(out, rc = -EINVAL);
520 /* check that the UUID matches */
521 if (!obd_uuid_equals(&uuid, <D_TGT(ltd, idx)->ltd_uuid)) {
522 CERROR("%s: LOD target UUID %s at index %d does not match %s\n",
523 obd->obd_name, obd_uuid2str(<D_TGT(ltd,idx)->ltd_uuid),
525 GOTO(out, rc = -EINVAL);
528 __lod_del_device(env, lod, ltd, idx, for_ost);
531 mutex_unlock(<d->ltd_mutex);
532 lod_putref(lod, ltd);
537 * Resize per-thread storage to hold specified size.
539 * A helper function to resize per-thread temporary storage. This storage
540 * is used to process LOV/LVM EAs and may be quite large. We do not want to
541 * allocate/release it every time, so instead we put it into the env and
542 * reallocate on demand. The memory is released when the correspondent thread
545 * \param[in] info LOD-specific storage in the environment
546 * \param[in] size new size to grow the buffer to
548 * \retval 0 on success, -ENOMEM if reallocation failed
550 int lod_ea_store_resize(struct lod_thread_info *info, size_t size)
552 __u32 round = size_roundup_power2(size);
555 lov_mds_md_size(LOV_MAX_STRIPE_COUNT, LOV_MAGIC_V3));
556 if (info->lti_ea_store) {
557 LASSERT(info->lti_ea_store_size);
558 LASSERT(info->lti_ea_store_size < round);
559 CDEBUG(D_INFO, "EA store size %d is not enough, need %d\n",
560 info->lti_ea_store_size, round);
561 OBD_FREE_LARGE(info->lti_ea_store, info->lti_ea_store_size);
562 info->lti_ea_store = NULL;
563 info->lti_ea_store_size = 0;
566 OBD_ALLOC_LARGE(info->lti_ea_store, round);
567 if (info->lti_ea_store == NULL)
569 info->lti_ea_store_size = round;
574 * Make LOV EA for striped object.
576 * Generate striping information and store it in the LOV EA of the given
577 * object. The caller must ensure nobody else is calling the function
578 * against the object concurrently. The transaction must be started.
579 * FLDB service must be running as well; it's used to map FID to the target,
580 * which is stored in LOV EA.
582 * \param[in] env execution environment for this thread
583 * \param[in] lo LOD object
584 * \param[in] th transaction handle
586 * \retval 0 if LOV EA is stored successfully
587 * \retval negative error number on failure
589 int lod_generate_and_set_lovea(const struct lu_env *env,
590 struct lod_object *lo, struct thandle *th)
592 struct lod_thread_info *info = lod_env_info(env);
593 struct dt_object *next = dt_object_child(&lo->ldo_obj);
594 const struct lu_fid *fid = lu_object_fid(&lo->ldo_obj.do_lu);
595 struct lov_mds_md_v1 *lmm;
596 struct lov_ost_data_v1 *objs;
604 magic = lo->ldo_pool != NULL ? LOV_MAGIC_V3 : LOV_MAGIC_V1;
605 lmm_size = lov_mds_md_size(lo->ldo_stripenr, magic);
606 if (info->lti_ea_store_size < lmm_size) {
607 rc = lod_ea_store_resize(info, lmm_size);
612 if (lo->ldo_pattern == 0) /* default striping */
613 lo->ldo_pattern = LOV_PATTERN_RAID0;
615 lmm = info->lti_ea_store;
617 lmm->lmm_magic = cpu_to_le32(magic);
618 lmm->lmm_pattern = cpu_to_le32(lo->ldo_pattern);
619 fid_to_lmm_oi(fid, &lmm->lmm_oi);
620 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_BAD_LMMOI))
621 lmm->lmm_oi.oi.oi_id++;
622 lmm_oi_cpu_to_le(&lmm->lmm_oi, &lmm->lmm_oi);
623 lmm->lmm_stripe_size = cpu_to_le32(lo->ldo_stripe_size);
624 lmm->lmm_stripe_count = cpu_to_le16(lo->ldo_stripenr);
625 if (lo->ldo_pattern & LOV_PATTERN_F_RELEASED)
626 lmm->lmm_stripe_count = cpu_to_le16(lo->ldo_released_stripenr);
627 lmm->lmm_layout_gen = 0;
628 if (magic == LOV_MAGIC_V1) {
629 objs = &lmm->lmm_objects[0];
631 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *) lmm;
632 size_t cplen = strlcpy(v3->lmm_pool_name, lo->ldo_pool,
633 sizeof(v3->lmm_pool_name));
634 if (cplen >= sizeof(v3->lmm_pool_name))
636 objs = &v3->lmm_objects[0];
639 for (i = 0; i < lo->ldo_stripenr; i++) {
640 struct lu_fid *fid = &info->lti_fid;
641 struct lod_device *lod;
643 int type = LU_SEQ_RANGE_OST;
645 lod = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
646 LASSERT(lo->ldo_stripe[i]);
648 *fid = *lu_object_fid(&lo->ldo_stripe[i]->do_lu);
649 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_MULTIPLE_REF)) {
650 if (cfs_fail_val == 0)
651 cfs_fail_val = fid->f_oid;
653 fid->f_oid = cfs_fail_val;
656 rc = fid_to_ostid(fid, &info->lti_ostid);
659 ostid_cpu_to_le(&info->lti_ostid, &objs[i].l_ost_oi);
660 objs[i].l_ost_gen = cpu_to_le32(0);
661 rc = lod_fld_lookup(env, lod, fid, &index, &type);
663 CERROR("%s: Can not locate "DFID": rc = %d\n",
664 lod2obd(lod)->obd_name, PFID(fid), rc);
665 lod_object_free_striping(env, lo);
668 objs[i].l_ost_idx = cpu_to_le32(index);
671 info->lti_buf.lb_buf = lmm;
672 info->lti_buf.lb_len = lmm_size;
673 rc = dt_xattr_set(env, next, &info->lti_buf, XATTR_NAME_LOV, 0,
676 lod_object_free_striping(env, lo);
684 * Fill lti_ea_store buffer in the environment with a value for the given
685 * EA. The buffer is reallocated if the value doesn't fit.
687 * \param[in,out] env execution environment for this thread
688 * .lti_ea_store buffer is filled with EA's value
689 * \param[in] lo LOD object
690 * \param[in] name name of the EA
692 * \retval 0 if EA is fetched successfully
693 * \retval negative error number on failure
695 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
698 struct lod_thread_info *info = lod_env_info(env);
699 struct dt_object *next = dt_object_child(&lo->ldo_obj);
705 if (unlikely(info->lti_ea_store == NULL)) {
706 /* just to enter in allocation block below */
710 info->lti_buf.lb_buf = info->lti_ea_store;
711 info->lti_buf.lb_len = info->lti_ea_store_size;
712 rc = dt_xattr_get(env, next, &info->lti_buf, name, BYPASS_CAPA);
715 /* if object is not striped or inaccessible */
716 if (rc == -ENODATA || rc == -ENOENT)
720 /* EA doesn't fit, reallocate new buffer */
721 rc = dt_xattr_get(env, next, &LU_BUF_NULL, name,
723 if (rc == -ENODATA || rc == -ENOENT)
729 rc = lod_ea_store_resize(info, rc);
739 * Verify the target index is present in the current configuration.
741 * \param[in] md LOD device where the target table is stored
742 * \param[in] idx target's index
744 * \retval 0 if the index is present
745 * \retval -EINVAL if not
747 static int validate_lod_and_idx(struct lod_device *md, __u32 idx)
749 if (unlikely(idx >= md->lod_ost_descs.ltd_tgts_size ||
750 !cfs_bitmap_check(md->lod_ost_bitmap, idx))) {
751 CERROR("%s: bad idx: %d of %d\n", lod2obd(md)->obd_name, idx,
752 md->lod_ost_descs.ltd_tgts_size);
756 if (unlikely(OST_TGT(md, idx) == NULL)) {
757 CERROR("%s: bad lod_tgt_desc for idx: %d\n",
758 lod2obd(md)->obd_name, idx);
762 if (unlikely(OST_TGT(md, idx)->ltd_ost == NULL)) {
763 CERROR("%s: invalid lod device, for idx: %d\n",
764 lod2obd(md)->obd_name , idx);
772 * Instantiate objects for stripes.
774 * Allocate and initialize LU-objects representing the stripes. The number
775 * of the stripes (ldo_stripenr) must be initialized already. The caller
776 * must ensure nobody else is calling the function on the object at the same
777 * time. FLDB service must be running to be able to map a FID to the targets
778 * and find appropriate device representing that target.
780 * \param[in] env execution environment for this thread
781 * \param[in,out] lo LOD object
782 * \param[in] objs an array of IDs to creates the objects from
784 * \retval 0 if the objects are instantiated successfully
785 * \retval negative error number on failure
787 int lod_initialize_objects(const struct lu_env *env, struct lod_object *lo,
788 struct lov_ost_data_v1 *objs)
790 struct lod_thread_info *info = lod_env_info(env);
791 struct lod_device *md;
792 struct lu_object *o, *n;
793 struct lu_device *nd;
794 struct dt_object **stripe;
801 md = lu2lod_dev(lo->ldo_obj.do_lu.lo_dev);
802 LASSERT(lo->ldo_stripe == NULL);
803 LASSERT(lo->ldo_stripenr > 0);
804 LASSERT(lo->ldo_stripe_size > 0);
806 stripe_len = lo->ldo_stripenr;
807 OBD_ALLOC(stripe, sizeof(stripe[0]) * stripe_len);
811 for (i = 0; i < lo->ldo_stripenr; i++) {
812 if (unlikely(lovea_slot_is_dummy(&objs[i])))
815 ostid_le_to_cpu(&objs[i].l_ost_oi, &info->lti_ostid);
816 idx = le32_to_cpu(objs[i].l_ost_idx);
817 rc = ostid_to_fid(&info->lti_fid, &info->lti_ostid, idx);
820 LASSERTF(fid_is_sane(&info->lti_fid), ""DFID" insane!\n",
821 PFID(&info->lti_fid));
822 lod_getref(&md->lod_ost_descs);
824 rc = validate_lod_and_idx(md, idx);
825 if (unlikely(rc != 0)) {
826 lod_putref(md, &md->lod_ost_descs);
830 nd = &OST_TGT(md,idx)->ltd_ost->dd_lu_dev;
831 lod_putref(md, &md->lod_ost_descs);
833 /* In the function below, .hs_keycmp resolves to
834 * u_obj_hop_keycmp() */
835 /* coverity[overrun-buffer-val] */
836 o = lu_object_find_at(env, nd, &info->lti_fid, NULL);
838 GOTO(out, rc = PTR_ERR(o));
840 n = lu_object_locate(o->lo_header, nd->ld_type);
843 stripe[i] = container_of(n, struct dt_object, do_lu);
848 for (i = 0; i < stripe_len; i++)
849 if (stripe[i] != NULL)
850 lu_object_put(env, &stripe[i]->do_lu);
852 OBD_FREE(stripe, sizeof(stripe[0]) * stripe_len);
853 lo->ldo_stripenr = 0;
855 lo->ldo_stripe = stripe;
856 lo->ldo_stripes_allocated = stripe_len;
863 * Instantiate objects for striping.
865 * Parse striping information in \a buf and instantiate the objects
866 * representing the stripes.
868 * \param[in] env execution environment for this thread
869 * \param[in] lo LOD object
870 * \param[in] buf buffer storing LOV EA to parse
872 * \retval 0 if parsing and objects creation succeed
873 * \retval negative error number on failure
875 int lod_parse_striping(const struct lu_env *env, struct lod_object *lo,
876 const struct lu_buf *buf)
878 struct lov_mds_md_v1 *lmm;
879 struct lov_ost_data_v1 *objs;
886 LASSERT(buf->lb_buf);
887 LASSERT(buf->lb_len);
889 lmm = (struct lov_mds_md_v1 *) buf->lb_buf;
890 magic = le32_to_cpu(lmm->lmm_magic);
891 pattern = le32_to_cpu(lmm->lmm_pattern);
893 if (magic != LOV_MAGIC_V1 && magic != LOV_MAGIC_V3)
894 GOTO(out, rc = -EINVAL);
895 if (lov_pattern(pattern) != LOV_PATTERN_RAID0)
896 GOTO(out, rc = -EINVAL);
898 lo->ldo_pattern = pattern;
899 lo->ldo_stripe_size = le32_to_cpu(lmm->lmm_stripe_size);
900 lo->ldo_layout_gen = le16_to_cpu(lmm->lmm_layout_gen);
901 lo->ldo_stripenr = le16_to_cpu(lmm->lmm_stripe_count);
902 /* released file stripenr fixup. */
903 if (pattern & LOV_PATTERN_F_RELEASED)
904 lo->ldo_stripenr = 0;
906 LASSERT(buf->lb_len >= lov_mds_md_size(lo->ldo_stripenr, magic));
908 if (magic == LOV_MAGIC_V3) {
909 struct lov_mds_md_v3 *v3 = (struct lov_mds_md_v3 *) lmm;
910 objs = &v3->lmm_objects[0];
911 lod_object_set_pool(lo, v3->lmm_pool_name);
913 objs = &lmm->lmm_objects[0];
916 if (lo->ldo_stripenr > 0)
917 rc = lod_initialize_objects(env, lo, objs);
924 * Initialize the object representing the stripes.
926 * Unless the stripes are initialized already, fetch LOV (for regular
927 * objects) or LMV (for directory objects) EA and call lod_parse_striping()
928 * to instantiate the objects representing the stripes.
930 * \param[in] env execution environment for this thread
931 * \param[in,out] lo LOD object
933 * \retval 0 if parsing and object creation succeed
934 * \retval negative error number on failure
936 int lod_load_striping_locked(const struct lu_env *env, struct lod_object *lo)
938 struct lod_thread_info *info = lod_env_info(env);
939 struct lu_buf *buf = &info->lti_buf;
940 struct dt_object *next = dt_object_child(&lo->ldo_obj);
944 /* already initialized? */
945 if (lo->ldo_stripe != NULL)
948 if (!dt_object_exists(next))
951 /* Do not load stripe for slaves of striped dir */
952 if (lo->ldo_dir_slave_stripe)
955 if (S_ISREG(lu_object_attr(lod2lu_obj(lo)))) {
956 rc = lod_get_lov_ea(env, lo);
960 * there is LOV EA (striping information) in this object
961 * let's parse it and create in-core objects for the stripes
963 buf->lb_buf = info->lti_ea_store;
964 buf->lb_len = info->lti_ea_store_size;
965 rc = lod_parse_striping(env, lo, buf);
966 } else if (S_ISDIR(lu_object_attr(lod2lu_obj(lo)))) {
967 rc = lod_get_lmv_ea(env, lo);
968 if (rc < (typeof(rc))sizeof(struct lmv_mds_md_v1))
969 GOTO(out, rc = rc > 0 ? -EINVAL : rc);
971 buf->lb_buf = info->lti_ea_store;
972 buf->lb_len = info->lti_ea_store_size;
973 if (rc == sizeof(struct lmv_mds_md_v1)) {
974 rc = lod_load_lmv_shards(env, lo, buf, true);
975 if (buf->lb_buf != info->lti_ea_store) {
976 OBD_FREE_LARGE(info->lti_ea_store,
977 info->lti_ea_store_size);
978 info->lti_ea_store = buf->lb_buf;
979 info->lti_ea_store_size = buf->lb_len;
987 * there is LOV EA (striping information) in this object
988 * let's parse it and create in-core objects for the stripes
990 rc = lod_parse_dir_striping(env, lo, buf);
997 * A generic function to initialize the stripe objects.
999 * A protected version of lod_load_striping_locked() - load the striping
1000 * information from storage, parse that and instantiate LU objects to
1001 * represent the stripes. The LOD object \a lo supplies a pointer to the
1002 * next sub-object in the LU stack so we can lock it. Also use \a lo to
1003 * return an array of references to the newly instantiated objects.
1005 * \param[in] env execution environment for this thread
1006 * \param[in,out] lo LOD object, where striping is stored and
1007 * which gets an array of references
1009 * \retval 0 if parsing and object creation succeed
1010 * \retval negative error number on failure
1012 int lod_load_striping(const struct lu_env *env, struct lod_object *lo)
1014 struct dt_object *next = dt_object_child(&lo->ldo_obj);
1017 /* currently this code is supposed to be called from declaration
1018 * phase only, thus the object is not expected to be locked by caller */
1019 dt_write_lock(env, next, 0);
1020 rc = lod_load_striping_locked(env, lo);
1021 dt_write_unlock(env, next);
1028 * Check the validity of all fields including the magic, stripe size,
1029 * stripe count, stripe offset and that the pool is present. Also check
1030 * that each target index points to an existing target. The additional
1031 * \a is_from_disk turns additional checks. In some cases zero fields
1032 * are allowed (like pattern=0).
1034 * \param[in] d LOD device
1035 * \param[in] buf buffer with LOV EA to verify
1036 * \param[in] is_from_disk 0 - from user, allow some fields to be 0
1037 * 1 - from disk, do not allow
1039 * \retval 0 if the striping is valid
1040 * \retval -EINVAL if striping is invalid
1042 int lod_verify_striping(struct lod_device *d, const struct lu_buf *buf,
1045 struct lov_user_md_v1 *lum;
1046 struct lov_user_md_v3 *lum3;
1047 struct pool_desc *pool = NULL;
1051 __u16 stripe_offset;
1058 LASSERT(sizeof(*lum) < sizeof(*lum3));
1060 if (buf->lb_len < sizeof(*lum)) {
1061 CDEBUG(D_IOCTL, "buf len %zu too small for lov_user_md\n",
1063 GOTO(out, rc = -EINVAL);
1066 magic = le32_to_cpu(lum->lmm_magic);
1067 if (magic != LOV_USER_MAGIC_V1 &&
1068 magic != LOV_USER_MAGIC_V3 &&
1069 magic != LOV_MAGIC_V1_DEF &&
1070 magic != LOV_MAGIC_V3_DEF) {
1071 CDEBUG(D_IOCTL, "bad userland LOV MAGIC: %#x\n", magic);
1072 GOTO(out, rc = -EINVAL);
1075 /* the user uses "0" for default stripe pattern normally. */
1076 if (!is_from_disk && lum->lmm_pattern == 0)
1077 lum->lmm_pattern = cpu_to_le32(LOV_PATTERN_RAID0);
1079 if (le32_to_cpu(lum->lmm_pattern) != LOV_PATTERN_RAID0) {
1080 CDEBUG(D_IOCTL, "bad userland stripe pattern: %#x\n",
1081 le32_to_cpu(lum->lmm_pattern));
1082 GOTO(out, rc = -EINVAL);
1085 /* 64kB is the largest common page size we see (ia64), and matches the
1087 stripe_size = le32_to_cpu(lum->lmm_stripe_size);
1088 if (stripe_size & (LOV_MIN_STRIPE_SIZE - 1)) {
1089 CDEBUG(D_IOCTL, "stripe size %u not a multiple of %u\n",
1090 stripe_size, LOV_MIN_STRIPE_SIZE);
1091 GOTO(out, rc = -EINVAL);
1094 stripe_offset = le16_to_cpu(lum->lmm_stripe_offset);
1095 if (stripe_offset != LOV_OFFSET_DEFAULT) {
1096 /* if offset is not within valid range [0, osts_size) */
1097 if (stripe_offset >= d->lod_osts_size) {
1098 CDEBUG(D_IOCTL, "stripe offset %u >= bitmap size %u\n",
1099 stripe_offset, d->lod_osts_size);
1100 GOTO(out, rc = -EINVAL);
1103 /* if lmm_stripe_offset is *not* in bitmap */
1104 if (!cfs_bitmap_check(d->lod_ost_bitmap, stripe_offset)) {
1105 CDEBUG(D_IOCTL, "stripe offset %u not in bitmap\n",
1107 GOTO(out, rc = -EINVAL);
1111 if (magic == LOV_USER_MAGIC_V1 || magic == LOV_MAGIC_V1_DEF)
1112 lum_size = offsetof(struct lov_user_md_v1,
1114 else if (magic == LOV_USER_MAGIC_V3 || magic == LOV_MAGIC_V3_DEF)
1115 lum_size = offsetof(struct lov_user_md_v3,
1118 GOTO(out, rc = -EINVAL);
1120 stripe_count = le16_to_cpu(lum->lmm_stripe_count);
1121 if (buf->lb_len != lum_size) {
1122 CDEBUG(D_IOCTL, "invalid buf len %zu for lov_user_md with "
1123 "magic %#x and stripe_count %u\n",
1124 buf->lb_len, magic, stripe_count);
1125 GOTO(out, rc = -EINVAL);
1128 if (!(magic == LOV_USER_MAGIC_V3 || magic == LOV_MAGIC_V3_DEF))
1132 if (buf->lb_len < sizeof(*lum3)) {
1133 CDEBUG(D_IOCTL, "buf len %zu too small for lov_user_md_v3\n",
1135 GOTO(out, rc = -EINVAL);
1138 /* In the function below, .hs_keycmp resolves to
1139 * pool_hashkey_keycmp() */
1140 /* coverity[overrun-buffer-val] */
1141 pool = lod_find_pool(d, lum3->lmm_pool_name);
1145 if (stripe_offset != LOV_OFFSET_DEFAULT) {
1146 rc = lod_check_index_in_pool(stripe_offset, pool);
1148 GOTO(out, rc = -EINVAL);
1151 if (is_from_disk && stripe_count > pool_tgt_count(pool)) {
1153 "stripe count %u > # OSTs %u in the pool\n",
1154 stripe_count, pool_tgt_count(pool));
1155 GOTO(out, rc = -EINVAL);
1160 lod_pool_putref(pool);
1165 void lod_fix_desc_stripe_size(__u64 *val)
1167 if (*val < LOV_MIN_STRIPE_SIZE) {
1169 LCONSOLE_INFO("Increasing default stripe size to "
1170 "minimum value %u\n",
1171 LOV_DESC_STRIPE_SIZE_DEFAULT);
1172 *val = LOV_DESC_STRIPE_SIZE_DEFAULT;
1173 } else if (*val & (LOV_MIN_STRIPE_SIZE - 1)) {
1174 *val &= ~(LOV_MIN_STRIPE_SIZE - 1);
1175 LCONSOLE_WARN("Changing default stripe size to "LPU64" (a "
1176 "multiple of %u)\n",
1177 *val, LOV_MIN_STRIPE_SIZE);
1181 void lod_fix_desc_stripe_count(__u32 *val)
1187 void lod_fix_desc_pattern(__u32 *val)
1189 /* from lov_setstripe */
1190 if ((*val != 0) && (*val != LOV_PATTERN_RAID0)) {
1191 LCONSOLE_WARN("Unknown stripe pattern: %#x\n", *val);
1196 void lod_fix_desc_qos_maxage(__u32 *val)
1198 /* fix qos_maxage */
1200 *val = LOV_DESC_QOS_MAXAGE_DEFAULT;
1204 * Used to fix insane default striping.
1206 * \param[in] desc striping description
1208 void lod_fix_desc(struct lov_desc *desc)
1210 lod_fix_desc_stripe_size(&desc->ld_default_stripe_size);
1211 lod_fix_desc_stripe_count(&desc->ld_default_stripe_count);
1212 lod_fix_desc_pattern(&desc->ld_pattern);
1213 lod_fix_desc_qos_maxage(&desc->ld_qos_maxage);
1217 * Initialize the structures used to store pools and default striping.
1219 * \param[in] lod LOD device
1220 * \param[in] lcfg configuration structure storing default striping.
1222 * \retval 0 if initialization succeeds
1223 * \retval negative error number on failure
1225 int lod_pools_init(struct lod_device *lod, struct lustre_cfg *lcfg)
1227 struct obd_device *obd;
1228 struct lov_desc *desc;
1232 obd = class_name2obd(lustre_cfg_string(lcfg, 0));
1233 LASSERT(obd != NULL);
1234 obd->obd_lu_dev = &lod->lod_dt_dev.dd_lu_dev;
1236 if (LUSTRE_CFG_BUFLEN(lcfg, 1) < 1) {
1237 CERROR("LOD setup requires a descriptor\n");
1241 desc = (struct lov_desc *)lustre_cfg_buf(lcfg, 1);
1243 if (sizeof(*desc) > LUSTRE_CFG_BUFLEN(lcfg, 1)) {
1244 CERROR("descriptor size wrong: %d > %d\n",
1245 (int)sizeof(*desc), LUSTRE_CFG_BUFLEN(lcfg, 1));
1249 if (desc->ld_magic != LOV_DESC_MAGIC) {
1250 if (desc->ld_magic == __swab32(LOV_DESC_MAGIC)) {
1251 CDEBUG(D_OTHER, "%s: Swabbing lov desc %p\n",
1252 obd->obd_name, desc);
1253 lustre_swab_lov_desc(desc);
1255 CERROR("%s: Bad lov desc magic: %#x\n",
1256 obd->obd_name, desc->ld_magic);
1263 desc->ld_active_tgt_count = 0;
1264 lod->lod_desc = *desc;
1266 lod->lod_sp_me = LUSTRE_SP_CLI;
1268 /* Set up allocation policy (QoS and RR) */
1269 INIT_LIST_HEAD(&lod->lod_qos.lq_oss_list);
1270 init_rwsem(&lod->lod_qos.lq_rw_sem);
1271 lod->lod_qos.lq_dirty = 1;
1272 lod->lod_qos.lq_rr.lqr_dirty = 1;
1273 lod->lod_qos.lq_reset = 1;
1274 /* Default priority is toward free space balance */
1275 lod->lod_qos.lq_prio_free = 232;
1276 /* Default threshold for rr (roughly 17%) */
1277 lod->lod_qos.lq_threshold_rr = 43;
1279 /* Set up OST pool environment */
1280 lod->lod_pools_hash_body = cfs_hash_create("POOLS", HASH_POOLS_CUR_BITS,
1281 HASH_POOLS_MAX_BITS,
1282 HASH_POOLS_BKT_BITS, 0,
1285 &pool_hash_operations,
1287 if (lod->lod_pools_hash_body == NULL)
1290 INIT_LIST_HEAD(&lod->lod_pool_list);
1291 lod->lod_pool_count = 0;
1292 rc = lod_ost_pool_init(&lod->lod_pool_info, 0);
1295 rc = lod_ost_pool_init(&lod->lod_qos.lq_rr.lqr_pool, 0);
1297 GOTO(out_pool_info, rc);
1302 lod_ost_pool_free(&lod->lod_pool_info);
1304 cfs_hash_putref(lod->lod_pools_hash_body);
1310 * Release the structures describing the pools.
1312 * \param[in] lod LOD device from which we release the structures
1316 int lod_pools_fini(struct lod_device *lod)
1318 struct obd_device *obd = lod2obd(lod);
1319 struct pool_desc *pool, *tmp;
1322 list_for_each_entry_safe(pool, tmp, &lod->lod_pool_list, pool_list) {
1323 /* free pool structs */
1324 CDEBUG(D_INFO, "delete pool %p\n", pool);
1325 /* In the function below, .hs_keycmp resolves to
1326 * pool_hashkey_keycmp() */
1327 /* coverity[overrun-buffer-val] */
1328 lod_pool_del(obd, pool->pool_name);
1331 cfs_hash_putref(lod->lod_pools_hash_body);
1332 lod_ost_pool_free(&(lod->lod_qos.lq_rr.lqr_pool));
1333 lod_ost_pool_free(&lod->lod_pool_info);