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 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
23 * Copyright 2009 Sun Microsystems, Inc. All rights reserved
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2014, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/lod/lod_internal.h
34 * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
35 * Author: Mikhail Pershin <mike.pershin@intel.com>
38 #ifndef _LOD_INTERNAL_H
39 #define _LOD_INTERNAL_H
41 #include <libcfs/libcfs.h>
42 #include <lustre_cfg.h>
44 #include <dt_object.h>
46 #define LOV_USES_ASSIGNED_STRIPE 0
47 #define LOV_USES_DEFAULT_STRIPE 1
49 /* Special values to remove LOV EA from disk */
50 #define LOVEA_DELETE_VALUES(size, count, offset, pool) \
51 ((size) == 0 && (count) == 0 && \
52 (offset) == (typeof(offset))(-1) && (pool) == NULL)
54 #define LMVEA_DELETE_VALUES(count, offset) \
55 ((count) == 0 && (offset) == (typeof(offset))(-1))
57 #define LOV_OFFSET_DEFAULT ((__u16)-1)
60 spinlock_t lqr_alloc; /* protect allocation index */
61 __u32 lqr_start_idx; /* start index of new inode */
62 __u32 lqr_offset_idx;/* aliasing for start_idx */
63 int lqr_start_count;/* reseed counter */
64 struct ost_pool lqr_pool; /* round-robin optimized list */
65 unsigned long lqr_dirty:1; /* recalc round-robin list */
69 char pool_name[LOV_MAXPOOLNAME + 1];
70 struct ost_pool pool_obds; /* pool members */
71 atomic_t pool_refcount;
72 struct lod_qos_rr pool_rr;
73 struct hlist_node pool_hash; /* access by poolname */
74 struct list_head pool_list;
75 struct proc_dir_entry *pool_proc_entry;
76 struct obd_device *pool_lobd; /* owner */
79 #define pool_tgt_size(p) ((p)->pool_obds.op_size)
80 #define pool_tgt_count(p) ((p)->pool_obds.op_count)
81 #define pool_tgt_array(p) ((p)->pool_obds.op_array)
82 #define pool_tgt_rw_sem(p) ((p)->pool_obds.op_rw_sem)
85 struct list_head lq_oss_list;
86 struct rw_semaphore lq_rw_sem;
87 __u32 lq_active_oss_count;
88 unsigned int lq_prio_free; /* priority for free space */
89 unsigned int lq_threshold_rr;/* priority for rr */
90 struct lod_qos_rr lq_rr; /* round robin qos data */
91 bool lq_dirty:1, /* recalc qos data */
92 lq_same_space:1,/* the ost's all have approx.
93 the same space avail */
94 lq_reset:1; /* zero current penalties */
98 struct obd_uuid lqo_uuid; /* ptlrpc's c_remote_uuid */
99 struct list_head lqo_oss_list; /* link to lov_qos */
100 __u64 lqo_bavail; /* total bytes avail on OSS */
101 __u64 lqo_penalty; /* current penalty */
102 __u64 lqo_penalty_per_obj; /* penalty decrease
104 time_t lqo_used; /* last used time, seconds */
105 __u32 lqo_ost_count; /* number of osts on this oss */
109 struct lod_qos_oss *ltq_oss; /* oss info */
110 __u64 ltq_penalty; /* current penalty */
111 __u64 ltq_penalty_per_obj; /* penalty decrease
113 __u64 ltq_weight; /* net weighting */
114 time_t ltq_used; /* last used time, seconds */
115 bool ltq_usable:1; /* usable for striping */
118 struct lod_tgt_desc {
119 struct dt_device *ltd_tgt;
120 struct list_head ltd_kill;
121 struct obd_export *ltd_exp;
122 struct obd_uuid ltd_uuid;
125 struct ltd_qos ltd_qos; /* qos info per target */
126 struct obd_statfs ltd_statfs;
127 struct ptlrpc_thread *ltd_recovery_thread;
128 unsigned long ltd_active:1,/* is this target up for requests */
129 ltd_activate:1,/* should target be activated */
130 ltd_reap:1, /* should this target be deleted */
131 ltd_got_update_log:1; /* Already got update log */
134 #define TGT_PTRS 256 /* number of pointers at 1st level */
135 #define TGT_PTRS_PER_BLOCK 256 /* number of pointers at 2nd level */
137 struct lod_tgt_desc_idx {
138 struct lod_tgt_desc *ldi_tgt[TGT_PTRS_PER_BLOCK];
141 #define LTD_TGT(ltd, index) \
142 ((ltd)->ltd_tgt_idx[(index) / \
143 TGT_PTRS_PER_BLOCK]->ldi_tgt[(index) % TGT_PTRS_PER_BLOCK])
145 #define OST_TGT(lod, index) LTD_TGT(&lod->lod_ost_descs, index)
146 #define MDT_TGT(lod, index) LTD_TGT(&lod->lod_mdt_descs, index)
147 struct lod_tgt_descs {
148 /* list of known TGTs */
149 struct lod_tgt_desc_idx *ltd_tgt_idx[TGT_PTRS];
150 /* Size of the lod_tgts array, granted to be a power of 2 */
152 /* number of registered TGTs */
154 /* bitmap of TGTs available */
155 cfs_bitmap_t *ltd_tgt_bitmap;
156 /* TGTs scheduled to be deleted */
158 /* Table refcount used for delayed deletion */
160 /* mutex to serialize concurrent updates to the tgt table */
161 struct mutex ltd_mutex;
162 /* read/write semaphore used for array relocation */
163 struct rw_semaphore ltd_rw_sem;
167 struct dt_device lod_dt_dev;
168 struct obd_export *lod_child_exp;
169 struct dt_device *lod_child;
170 struct proc_dir_entry *lod_proc_entry;
171 struct lprocfs_stats *lod_stats;
172 spinlock_t lod_connects_lock;
174 unsigned int lod_recovery_completed:1,
177 lod_child_got_update_log:1;
179 /* lov settings descriptor storing static information */
180 struct lov_desc lod_desc;
182 /* use to protect ld_active_tgt_count and all ltd_active */
183 spinlock_t lod_desc_lock;
185 /* Description of OST */
186 struct lod_tgt_descs lod_ost_descs;
187 /* Description of MDT */
188 struct lod_tgt_descs lod_mdt_descs;
190 /* Recovery thread for lod_child */
191 struct ptlrpc_thread lod_child_recovery_thread;
193 /* maximum EA size underlied OSD may have */
194 unsigned int lod_osd_max_easize;
196 /*FIXME: When QOS and pool is implemented for MDT, probably these
197 * structure should be moved to lod_tgt_descs as well.
199 /* QoS info per LOD */
200 struct lod_qos lod_qos; /* qos info per lod */
203 struct ost_pool lod_pool_info; /* all OSTs in a packed array */
205 struct cfs_hash *lod_pools_hash_body; /* used for key access */
206 struct list_head lod_pool_list; /* used for sequential access */
207 struct proc_dir_entry *lod_pool_proc_entry;
209 enum lustre_sec_part lod_sp_me;
211 struct proc_dir_entry *lod_symlink;
214 #define lod_osts lod_ost_descs.ltd_tgts
215 #define lod_ost_bitmap lod_ost_descs.ltd_tgt_bitmap
216 #define lod_ostnr lod_ost_descs.ltd_tgtnr
217 #define lod_osts_size lod_ost_descs.ltd_tgts_size
218 #define ltd_ost ltd_tgt
219 #define lod_ost_desc lod_tgt_desc
221 #define lod_mdts lod_mdt_descs.ltd_tgts
222 #define lod_mdt_bitmap lod_mdt_descs.ltd_tgt_bitmap
223 #define lod_remote_mdt_count lod_mdt_descs.ltd_tgtnr
224 #define lod_mdts_size lod_mdt_descs.ltd_tgts_size
225 #define ltd_mdt ltd_tgt
226 #define lod_mdt_desc lod_tgt_desc
228 struct lod_dir_stripe_info {
229 __u32 ldsi_stripe_offset;
230 __u32 ldsi_def_stripenr;
231 __u32 ldsi_def_stripe_offset;
232 __u32 ldsi_def_hash_type;
233 __u32 ldsi_hash_type;
235 unsigned int ldsi_def_striping_set:1,
236 ldsi_def_striping_cached:1,
241 * XXX: shrink this structure, currently it's 72bytes on 32bit arch,
242 * so, slab will be allocating 128bytes
245 struct dt_object ldo_obj;
247 /* if object is striped, then the next fields describe stripes */
248 /* For striped directory, ldo_stripenr == slave stripe count */
250 __u16 ldo_layout_gen;
251 __u32 ldo_stripe_size;
253 __u16 ldo_released_stripenr;
255 struct dt_object **ldo_stripe;
256 /* to know how much memory to free, ldo_stripenr can be less */
257 /* default striping for directory represented by this object
258 * is cached in stripenr/stripe_size */
259 unsigned int ldo_stripes_allocated:16,
260 ldo_striping_cached:1,
261 ldo_def_striping_set:1,
262 ldo_def_striping_cached:1,
263 /* ldo_dir_slave_stripe indicate this is a slave stripe of
265 ldo_dir_slave_stripe:1;
266 __u32 ldo_def_stripe_size;
267 __u16 ldo_def_stripenr;
268 __u16 ldo_def_stripe_offset;
269 struct lod_dir_stripe_info *ldo_dir_stripe;
272 #define ldo_dir_stripe_offset ldo_dir_stripe->ldsi_stripe_offset
273 #define ldo_dir_def_stripenr ldo_dir_stripe->ldsi_def_stripenr
274 #define ldo_dir_hash_type ldo_dir_stripe->ldsi_hash_type
275 #define ldo_dir_def_hash_type ldo_dir_stripe->ldsi_def_hash_type
276 #define ldo_dir_striped ldo_dir_stripe->ldsi_striped
277 #define ldo_dir_def_striping_set ldo_dir_stripe->ldsi_def_striping_set
278 #define ldo_dir_def_striping_cached ldo_dir_stripe->ldsi_def_striping_cached
279 #define ldo_dir_def_stripe_offset ldo_dir_stripe->ldsi_def_stripe_offset
282 struct dt_object *lit_obj; /* object from the layer below */
283 /* stripe offset of iteration */
284 __u32 lit_stripe_index;
286 struct dt_it *lit_it; /* iterator from the layer below */
289 struct lod_thread_info {
290 /* per-thread buffer for LOV EA */
292 __u32 lti_ea_store_size;
293 /* per-thread buffer for LMV EA */
294 struct lu_buf lti_buf;
295 struct ost_id lti_ostid;
296 struct lu_fid lti_fid;
297 struct obd_statfs lti_osfs;
298 struct lu_attr lti_attr;
299 struct lod_it lti_it;
300 struct ldlm_res_id lti_res_id;
301 /* used to hold lu_dirent, sizeof(struct lu_dirent) + NAME_MAX */
302 char lti_key[sizeof(struct lu_dirent) + NAME_MAX];
303 struct dt_object_format lti_format;
304 struct lu_name lti_name;
305 struct lu_buf lti_linkea_buf;
306 struct dt_insert_rec lti_dt_rec;
307 struct llog_catid lti_cid;
308 struct llog_cookie lti_cookie;
309 struct lustre_cfg lti_lustre_cfg;
312 extern const struct lu_device_operations lod_lu_ops;
314 static inline int lu_device_is_lod(struct lu_device *d)
316 return ergo(d != NULL && d->ld_ops != NULL, d->ld_ops == &lod_lu_ops);
319 static inline struct lod_device* lu2lod_dev(struct lu_device *d)
321 LASSERT(lu_device_is_lod(d));
322 return container_of0(d, struct lod_device, lod_dt_dev.dd_lu_dev);
325 static inline struct lu_device *lod2lu_dev(struct lod_device *d)
327 return &d->lod_dt_dev.dd_lu_dev;
330 static inline struct obd_device *lod2obd(struct lod_device *d)
332 return d->lod_dt_dev.dd_lu_dev.ld_obd;
335 static inline struct lod_device *dt2lod_dev(struct dt_device *d)
337 LASSERT(lu_device_is_lod(&d->dd_lu_dev));
338 return container_of0(d, struct lod_device, lod_dt_dev);
341 static inline struct lod_object *lu2lod_obj(struct lu_object *o)
343 LASSERT(ergo(o != NULL, lu_device_is_lod(o->lo_dev)));
344 return container_of0(o, struct lod_object, ldo_obj.do_lu);
347 static inline struct lu_object *lod2lu_obj(struct lod_object *obj)
349 return &obj->ldo_obj.do_lu;
352 static inline struct lod_object *lod_obj(const struct lu_object *o)
354 LASSERT(lu_device_is_lod(o->lo_dev));
355 return container_of0(o, struct lod_object, ldo_obj.do_lu);
358 static inline struct lod_object *lod_dt_obj(const struct dt_object *d)
360 return lod_obj(&d->do_lu);
363 static inline struct dt_object* lod_object_child(struct lod_object *o)
365 return container_of0(lu_object_next(lod2lu_obj(o)),
366 struct dt_object, do_lu);
369 extern struct lu_context_key lod_thread_key;
371 static inline struct lod_thread_info *lod_env_info(const struct lu_env *env)
373 struct lod_thread_info *info;
374 info = lu_context_key_get(&env->le_ctx, &lod_thread_key);
379 static inline struct lu_name *
380 lod_name_get(const struct lu_env *env, const void *area, int len)
382 struct lu_name *lname;
384 lname = &lod_env_info(env)->lti_name;
385 lname->ln_name = area;
386 lname->ln_namelen = len;
390 #define lod_foreach_ost(__dev, index) \
391 if ((__dev)->lod_osts_size > 0) \
392 cfs_foreach_bit((__dev)->lod_ost_bitmap, (index))
394 #define lod_foreach_mdt(mdt_dev, index) \
395 cfs_foreach_bit((mdt_dev)->lod_mdt_bitmap, (index))
398 extern struct kmem_cache *lod_object_kmem;
399 int lod_fld_lookup(const struct lu_env *env, struct lod_device *lod,
400 const struct lu_fid *fid, __u32 *tgt, int *flags);
401 int lod_sub_init_llog(const struct lu_env *env, struct lod_device *lod,
402 struct dt_device *dt);
403 void lod_sub_fini_llog(const struct lu_env *env,
404 struct dt_device *dt, struct ptlrpc_thread *thread);
405 int lodname2mdt_index(char *lodname, __u32 *mdt_index);
408 void lod_getref(struct lod_tgt_descs *ltd);
409 void lod_putref(struct lod_device *lod, struct lod_tgt_descs *ltd);
410 int lod_add_device(const struct lu_env *env, struct lod_device *lod,
411 char *osp, unsigned index, unsigned gen, int mdt_index,
412 char *type, int active);
413 int lod_del_device(const struct lu_env *env, struct lod_device *lod,
414 struct lod_tgt_descs *ltd, char *osp, unsigned idx,
415 unsigned gen, bool for_ost);
416 int lod_fini_tgt(const struct lu_env *env, struct lod_device *lod,
417 struct lod_tgt_descs *ltd, bool for_ost);
418 int lod_load_striping_locked(const struct lu_env *env, struct lod_object *lo);
419 int lod_load_striping(const struct lu_env *env, struct lod_object *lo);
421 int lod_get_ea(const struct lu_env *env, struct lod_object *lo,
424 lod_get_lov_ea(const struct lu_env *env, struct lod_object *lo)
426 return lod_get_ea(env, lo, XATTR_NAME_LOV);
430 lod_get_lmv_ea(const struct lu_env *env, struct lod_object *lo)
432 return lod_get_ea(env, lo, XATTR_NAME_LMV);
436 lod_get_default_lmv_ea(const struct lu_env *env, struct lod_object *lo)
438 return lod_get_ea(env, lo, XATTR_NAME_DEFAULT_LMV);
441 void lod_fix_desc(struct lov_desc *desc);
442 void lod_fix_desc_qos_maxage(__u32 *val);
443 void lod_fix_desc_pattern(__u32 *val);
444 void lod_fix_desc_stripe_count(__u32 *val);
445 void lod_fix_desc_stripe_size(__u64 *val);
446 int lod_pools_init(struct lod_device *m, struct lustre_cfg *cfg);
447 int lod_pools_fini(struct lod_device *m);
448 int lod_parse_striping(const struct lu_env *env, struct lod_object *mo,
449 const struct lu_buf *buf);
450 int lod_parse_dir_striping(const struct lu_env *env, struct lod_object *lo,
451 const struct lu_buf *buf);
452 int lod_initialize_objects(const struct lu_env *env, struct lod_object *mo,
453 struct lov_ost_data_v1 *objs);
454 int lod_verify_striping(struct lod_device *d, const struct lu_buf *buf,
456 int lod_generate_and_set_lovea(const struct lu_env *env,
457 struct lod_object *mo, struct thandle *th);
458 int lod_ea_store_resize(struct lod_thread_info *info, size_t size);
460 int lod_ost_pool_add(struct ost_pool *op, __u32 idx, unsigned int min_count);
461 int lod_ost_pool_remove(struct ost_pool *op, __u32 idx);
462 int lod_ost_pool_extend(struct ost_pool *op, unsigned int min_count);
463 struct pool_desc *lod_find_pool(struct lod_device *lod, char *poolname);
464 void lod_pool_putref(struct pool_desc *pool);
465 int lod_ost_pool_free(struct ost_pool *op);
466 int lod_pool_del(struct obd_device *obd, char *poolname);
467 int lod_ost_pool_init(struct ost_pool *op, unsigned int count);
468 extern struct cfs_hash_ops pool_hash_operations;
469 int lod_check_index_in_pool(__u32 idx, struct pool_desc *pool);
470 int lod_pool_new(struct obd_device *obd, char *poolname);
471 int lod_pool_add(struct obd_device *obd, char *poolname, char *ostname);
472 int lod_pool_remove(struct obd_device *obd, char *poolname, char *ostname);
475 int lod_qos_prep_create(const struct lu_env *env, struct lod_object *lo,
476 struct lu_attr *attr, const struct lu_buf *buf,
478 int qos_add_tgt(struct lod_device*, struct lod_tgt_desc *);
479 int qos_del_tgt(struct lod_device *, struct lod_tgt_desc *);
480 void lod_qos_rr_init(struct lod_qos_rr *lqr);
483 int lod_procfs_init(struct lod_device *lod);
484 void lod_procfs_fini(struct lod_device *lod);
487 extern struct dt_object_operations lod_obj_ops;
488 extern struct lu_object_operations lod_lu_obj_ops;
489 int lod_load_lmv_shards(const struct lu_env *env, struct lod_object *lo,
490 struct lu_buf *buf, bool resize);
491 int lod_object_set_pool(struct lod_object *o, char *pool);
492 int lod_declare_striped_object(const struct lu_env *env, struct dt_object *dt,
493 struct lu_attr *attr,
494 const struct lu_buf *lovea, struct thandle *th);
495 int lod_striping_create(const struct lu_env *env, struct dt_object *dt,
496 struct lu_attr *attr, struct dt_object_format *dof,
498 void lod_object_free_striping(const struct lu_env *env, struct lod_object *lo);
500 /* lod_sub_object.c */
501 struct thandle *lod_sub_get_thandle(const struct lu_env *env,
503 const struct dt_object *sub_obj,
504 bool *record_update);
505 int lod_sub_object_declare_create(const struct lu_env *env,
506 struct dt_object *dt,
507 struct lu_attr *attr,
508 struct dt_allocation_hint *hint,
509 struct dt_object_format *dof,
511 int lod_sub_object_create(const struct lu_env *env, struct dt_object *dt,
512 struct lu_attr *attr,
513 struct dt_allocation_hint *hint,
514 struct dt_object_format *dof,
516 int lod_sub_object_declare_ref_add(const struct lu_env *env,
517 struct dt_object *dt,
519 int lod_sub_object_ref_add(const struct lu_env *env, struct dt_object *dt,
521 int lod_sub_object_declare_ref_del(const struct lu_env *env,
522 struct dt_object *dt,
524 int lod_sub_object_ref_del(const struct lu_env *env, struct dt_object *dt,
526 int lod_sub_object_declare_destroy(const struct lu_env *env,
527 struct dt_object *dt,
529 int lod_sub_object_destroy(const struct lu_env *env, struct dt_object *dt,
531 int lod_sub_object_declare_insert(const struct lu_env *env,
532 struct dt_object *dt,
533 const struct dt_rec *rec,
534 const struct dt_key *key,
536 int lod_sub_object_index_insert(const struct lu_env *env, struct dt_object *dt,
537 const struct dt_rec *rec,
538 const struct dt_key *key, struct thandle *th,
540 int lod_sub_object_declare_delete(const struct lu_env *env,
541 struct dt_object *dt,
542 const struct dt_key *key,
544 int lod_sub_object_delete(const struct lu_env *env, struct dt_object *dt,
545 const struct dt_key *name, struct thandle *th);
546 int lod_sub_object_declare_xattr_set(const struct lu_env *env,
547 struct dt_object *dt,
548 const struct lu_buf *buf,
549 const char *name, int fl,
551 int lod_sub_object_xattr_set(const struct lu_env *env, struct dt_object *dt,
552 const struct lu_buf *buf, const char *name, int fl,
554 int lod_sub_object_declare_attr_set(const struct lu_env *env,
555 struct dt_object *dt,
556 const struct lu_attr *attr,
558 int lod_sub_object_attr_set(const struct lu_env *env,
559 struct dt_object *dt,
560 const struct lu_attr *attr,
562 int lod_sub_object_declare_xattr_del(const struct lu_env *env,
563 struct dt_object *dt,
566 int lod_sub_object_xattr_del(const struct lu_env *env,
567 struct dt_object *dt,
570 int lod_sub_object_declare_write(const struct lu_env *env,
571 struct dt_object *dt,
572 const struct lu_buf *buf, loff_t pos,
574 ssize_t lod_sub_object_write(const struct lu_env *env, struct dt_object *dt,
575 const struct lu_buf *buf, loff_t *pos,
576 struct thandle *th, int rq);
577 int lod_sub_object_declare_punch(const struct lu_env *env,
578 struct dt_object *dt,
579 __u64 start, __u64 end,
581 int lod_sub_object_punch(const struct lu_env *env, struct dt_object *dt,
582 __u64 start, __u64 end, struct thandle *th);
584 int lod_sub_prep_llog(const struct lu_env *env, struct lod_device *lod,
585 struct dt_device *dt, int index);