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, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE 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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2012, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
37 #ifndef _OFD_INTERNAL_H
38 #define _OFD_INTERNAL_H
41 #include <obd_class.h>
42 #include <dt_object.h>
43 #include <md_object.h>
44 #include <lustre_fid.h>
46 #define OFD_INIT_OBJID 0
47 #define OFD_PRECREATE_BATCH_DEFAULT (OBJ_SUBDIR_COUNT * 4)
49 /* on small filesystems we should not precreate too many objects in
50 * a single transaction, otherwise we can overflow transactions */
51 #define OFD_PRECREATE_SMALL_FS (1024ULL * 1024 * 1024)
52 #define OFD_PRECREATE_BATCH_SMALL 8
54 /* Limit the returned fields marked valid to those that we actually might set */
55 #define OFD_VALID_FLAGS (LA_TYPE | LA_MODE | LA_SIZE | LA_BLOCKS | \
56 LA_BLKSIZE | LA_ATIME | LA_MTIME | LA_CTIME)
58 /* per-client-per-object persistent state (LRU) */
60 struct list_head fmd_list; /* linked to fed_mod_list */
61 struct lu_fid fmd_fid; /* FID being written to */
62 __u64 fmd_mactime_xid; /* xid highest {m,a,c}time setattr */
63 cfs_time_t fmd_expire; /* time when the fmd should expire */
64 int fmd_refcount; /* reference counter - list holds 1 */
67 #define OFD_FMD_MAX_NUM_DEFAULT 128
68 #define OFD_FMD_MAX_AGE_DEFAULT msecs_to_jiffies((obd_timeout+10)*MSEC_PER_SEC)
70 #define OFD_SOFT_SYNC_LIMIT_DEFAULT 16
74 LPROC_OFD_STATS_READ = 0,
75 LPROC_OFD_STATS_WRITE,
76 LPROC_OFD_STATS_GETATTR,
77 LPROC_OFD_STATS_SETATTR,
78 LPROC_OFD_STATS_PUNCH,
80 LPROC_OFD_STATS_DESTROY,
81 LPROC_OFD_STATS_CREATE,
82 LPROC_OFD_STATS_STATFS,
83 LPROC_OFD_STATS_GET_INFO,
84 LPROC_OFD_STATS_SET_INFO,
85 LPROC_OFD_STATS_QUOTACTL,
89 static inline void ofd_counter_incr(struct obd_export *exp, int opcode,
90 char *jobid, long amount)
92 if (exp->exp_obd && exp->exp_obd->obd_stats)
93 lprocfs_counter_add(exp->exp_obd->obd_stats, opcode, amount);
95 if (exp->exp_obd && exp->exp_obd->u.obt.obt_jobstats.ojs_hash &&
96 (exp_connect_flags(exp) & OBD_CONNECT_JOBSTATS))
97 lprocfs_job_stats_log(exp->exp_obd, jobid, opcode, amount);
99 if (exp->exp_nid_stats != NULL &&
100 exp->exp_nid_stats->nid_stats != NULL) {
101 lprocfs_counter_add(exp->exp_nid_stats->nid_stats, opcode,
107 struct list_head os_list;
109 spinlock_t os_last_oid_lock;
110 struct mutex os_create_lock;
112 struct dt_object *os_lastid_obj;
113 unsigned long os_destroys_in_progress:1;
117 struct dt_device ofd_dt_dev;
118 struct dt_device *ofd_osd;
119 struct obd_export *ofd_osd_exp;
120 struct dt_device_param ofd_dt_conf;
121 /* DLM name-space for meta-data locks maintained by this server */
122 struct ldlm_namespace *ofd_namespace;
125 struct lu_target ofd_lut;
126 struct dt_object *ofd_health_check_file;
128 int ofd_subdir_count;
129 __u64 ofd_inconsistency_self_detected;
130 __u64 ofd_inconsistency_self_repaired;
132 struct list_head ofd_seq_list;
133 rwlock_t ofd_seq_list_lock;
135 int ofd_precreate_batch;
136 spinlock_t ofd_batch_lock;
138 /* protect all statfs-related counters */
139 spinlock_t ofd_osfs_lock;
140 /* statfs optimization: we cache a bit */
141 struct obd_statfs ofd_osfs;
144 /* writes which might be be accounted twice in ofd_osfs.os_bavail */
145 u64 ofd_osfs_unstable;
147 /* counters used during statfs update, protected by ofd_osfs_lock.
148 * record when some statfs refresh are in progress */
149 int ofd_statfs_inflight;
150 /* track writes completed while statfs refresh is underway.
151 * tracking is only effective when ofd_statfs_inflight > 1 */
152 u64 ofd_osfs_inflight;
154 /* grants: all values in bytes */
155 /* grant lock to protect all grant counters */
156 spinlock_t ofd_grant_lock;
157 /* total amount of dirty data reported by clients in incoming obdo */
159 /* sum of filesystem space granted to clients for async writes */
161 /* grant used by I/Os in progress (between prepare and commit) */
163 /* free space threshold over which we stop granting space to clients
164 * ofd_grant_ratio is stored as a fixed-point fraction using
165 * OFD_GRANT_RATIO_SHIFT of the remaining free space, not in percentage
168 /* number of clients using grants */
169 int ofd_tot_granted_clients;
171 /* ofd mod data: ofd_device wide values */
172 int ofd_fmd_max_num; /* per ofd ofd_mod_data */
173 cfs_duration_t ofd_fmd_max_age; /* time to fmd expiry */
175 spinlock_t ofd_flags_lock;
176 unsigned long ofd_raid_degraded:1,
177 /* sync journal on writes */
179 /* shall we grant space to clients not
180 * supporting OBD_CONNECT_GRANT_PARAM? */
181 ofd_grant_compat_disable:1,
182 /* Protected by ofd_lastid_rwsem. */
183 ofd_lastid_rebuilding:1,
184 ofd_record_fid_accessed:1,
185 ofd_lfsck_verify_pfid:1;
186 struct seq_server_site ofd_seq_site;
187 /* the limit of SOFT_SYNC RPCs that will trigger a soft sync */
188 unsigned int ofd_soft_sync_limit;
189 /* Protect ::ofd_lastid_rebuilding */
190 struct rw_semaphore ofd_lastid_rwsem;
191 __u64 ofd_lastid_gen;
192 struct ptlrpc_thread ofd_inconsistency_thread;
193 struct list_head ofd_inconsistency_list;
194 spinlock_t ofd_inconsistency_lock;
197 static inline struct ofd_device *ofd_dev(struct lu_device *d)
199 return container_of0(d, struct ofd_device, ofd_dt_dev.dd_lu_dev);
202 static inline struct obd_device *ofd_obd(struct ofd_device *ofd)
204 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd;
207 static inline struct ofd_device *ofd_exp(struct obd_export *exp)
209 return ofd_dev(exp->exp_obd->obd_lu_dev);
212 static inline char *ofd_name(struct ofd_device *ofd)
214 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd->obd_name;
218 struct lu_object_header ofo_header;
219 struct dt_object ofo_obj;
220 struct lu_fid ofo_pfid;
221 unsigned int ofo_pfid_checking:1,
225 static inline struct ofd_object *ofd_obj(struct lu_object *o)
227 return container_of0(o, struct ofd_object, ofo_obj.do_lu);
230 static inline int ofd_object_exists(struct ofd_object *obj)
232 LASSERT(obj != NULL);
233 if (lu_object_is_dying(obj->ofo_obj.do_lu.lo_header))
235 return lu_object_exists(&obj->ofo_obj.do_lu);
238 static inline struct dt_object *fo2dt(struct ofd_object *obj)
240 return &obj->ofo_obj;
243 static inline struct dt_object *ofd_object_child(struct ofd_object *_obj)
245 struct lu_object *lu = &(_obj)->ofo_obj.do_lu;
247 return container_of0(lu_object_next(lu), struct dt_object, do_lu);
250 static inline struct ofd_device *ofd_obj2dev(const struct ofd_object *fo)
252 return ofd_dev(fo->ofo_obj.do_lu.lo_dev);
255 static inline void ofd_read_lock(const struct lu_env *env,
256 struct ofd_object *fo)
258 struct dt_object *next = ofd_object_child(fo);
260 next->do_ops->do_read_lock(env, next, 0);
263 static inline void ofd_read_unlock(const struct lu_env *env,
264 struct ofd_object *fo)
266 struct dt_object *next = ofd_object_child(fo);
268 next->do_ops->do_read_unlock(env, next);
271 static inline void ofd_write_lock(const struct lu_env *env,
272 struct ofd_object *fo)
274 struct dt_object *next = ofd_object_child(fo);
276 next->do_ops->do_write_lock(env, next, 0);
279 static inline void ofd_write_unlock(const struct lu_env *env,
280 struct ofd_object *fo)
282 struct dt_object *next = ofd_object_child(fo);
284 next->do_ops->do_write_unlock(env, next);
288 * Common data shared by obdofd-level handlers. This is allocated per-thread
289 * to reduce stack consumption.
291 struct ofd_thread_info {
292 const struct lu_env *fti_env;
294 struct obd_export *fti_exp;
296 __u64 fti_pre_version;
298 struct lu_fid fti_fid;
299 struct lu_attr fti_attr;
300 struct lu_attr fti_attr2;
301 struct ldlm_res_id fti_resid;
303 struct filter_fid fti_mds_fid;
304 struct filter_fid_old fti_mds_fid_old;
306 struct ost_id fti_ostid;
307 struct ofd_object *fti_obj;
309 char name[64]; /* for ofd_init0() */
310 struct obd_statfs osfs; /* for obdofd_statfs() */
313 /* Ops object filename */
314 struct lu_name fti_name;
315 struct dt_object_format fti_dof;
316 struct lu_buf fti_buf;
319 /* Space used by the I/O, used by grant code */
320 unsigned long fti_used;
321 struct ost_lvb fti_lvb;
322 struct lfsck_request fti_lr;
325 extern void target_recovery_fini(struct obd_device *obd);
326 extern void target_recovery_init(struct lu_target *lut, svc_handler_t handler);
329 extern struct lu_context_key ofd_thread_key;
330 int ofd_postrecov(const struct lu_env *env, struct ofd_device *ofd);
331 int ofd_fiemap_get(const struct lu_env *env, struct ofd_device *ofd,
332 struct lu_fid *fid, struct fiemap *fiemap);
335 extern struct obd_ops ofd_obd_ops;
336 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
337 struct obd_statfs *osfs, __u64 max_age,
339 int ofd_destroy_by_fid(const struct lu_env *env, struct ofd_device *ofd,
340 const struct lu_fid *fid, int orphan);
341 int ofd_statfs(const struct lu_env *env, struct obd_export *exp,
342 struct obd_statfs *osfs, __u64 max_age, __u32 flags);
343 int ofd_obd_disconnect(struct obd_export *exp);
346 u64 ofd_seq_last_oid(struct ofd_seq *oseq);
347 void ofd_seq_last_oid_set(struct ofd_seq *oseq, u64 id);
348 int ofd_seq_last_oid_write(const struct lu_env *env, struct ofd_device *ofd,
349 struct ofd_seq *oseq);
350 int ofd_seqs_init(const struct lu_env *env, struct ofd_device *ofd);
351 struct ofd_seq *ofd_seq_get(struct ofd_device *ofd, u64 seq);
352 void ofd_seq_put(const struct lu_env *env, struct ofd_seq *oseq);
354 int ofd_fs_setup(const struct lu_env *env, struct ofd_device *ofd,
355 struct obd_device *obd);
356 void ofd_fs_cleanup(const struct lu_env *env, struct ofd_device *ofd);
357 int ofd_precreate_batch(struct ofd_device *ofd, int batch);
358 struct ofd_seq *ofd_seq_load(const struct lu_env *env, struct ofd_device *ofd,
360 void ofd_seqs_fini(const struct lu_env *env, struct ofd_device *ofd);
361 void ofd_seqs_free(const struct lu_env *env, struct ofd_device *ofd);
364 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd);
365 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd);
366 int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
368 int ofd_preprw(const struct lu_env *env,int cmd, struct obd_export *exp,
369 struct obdo *oa, int objcount, struct obd_ioobj *obj,
370 struct niobuf_remote *rnb, int *nr_local,
371 struct niobuf_local *lnb);
372 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
373 struct obdo *oa, int objcount, struct obd_ioobj *obj,
374 struct niobuf_remote *rnb, int npages,
375 struct niobuf_local *lnb, int old_rc);
378 struct thandle *ofd_trans_create(const struct lu_env *env,
379 struct ofd_device *ofd);
380 int ofd_trans_start(const struct lu_env *env,
381 struct ofd_device *ofd, struct ofd_object *fo,
383 void ofd_trans_stop(const struct lu_env *env, struct ofd_device *ofd,
384 struct thandle *th, int rc);
385 int ofd_txn_stop_cb(const struct lu_env *env, struct thandle *txn,
389 #ifdef CONFIG_PROC_FS
390 extern struct lprocfs_vars lprocfs_ofd_obd_vars[];
391 void ofd_stats_counter_init(struct lprocfs_stats *stats);
393 static inline void ofd_stats_counter_init(struct lprocfs_stats *stats) {}
397 struct ofd_object *ofd_object_find(const struct lu_env *env,
398 struct ofd_device *ofd,
399 const struct lu_fid *fid);
400 int ofd_object_ff_load(const struct lu_env *env, struct ofd_object *fo);
401 int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd,
402 u64 id, struct ofd_seq *oseq, int nr, int sync);
404 void ofd_object_put(const struct lu_env *env, struct ofd_object *fo);
405 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
406 struct lu_attr *la, struct filter_fid *ff);
407 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
408 __u64 start, __u64 end, struct lu_attr *la,
409 struct filter_fid *ff, struct obdo *oa);
410 int ofd_object_destroy(const struct lu_env *, struct ofd_object *, int);
411 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
413 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
414 struct lu_attr *la, int is_setattr);
417 struct ofd_object *ofd_object_find_exists(const struct lu_env *env,
418 struct ofd_device *ofd,
419 const struct lu_fid *fid)
421 struct ofd_object *fo;
423 fo = ofd_object_find(env, ofd, fid);
424 if (!IS_ERR(fo) && !ofd_object_exists(fo)) {
425 ofd_object_put(env, fo);
426 fo = ERR_PTR(-ENOENT);
432 #define OFD_GRANT_RATIO_SHIFT 8
433 static inline u64 ofd_grant_reserved(struct ofd_device *ofd, u64 bavail)
435 return (bavail * ofd->ofd_grant_ratio) >> OFD_GRANT_RATIO_SHIFT;
438 static inline int ofd_grant_ratio_conv(int percentage)
440 return (percentage << OFD_GRANT_RATIO_SHIFT) / 100;
443 static inline int ofd_grant_param_supp(struct obd_export *exp)
445 return !!(exp_connect_flags(exp) & OBD_CONNECT_GRANT_PARAM);
448 /* Blocksize used for client not supporting OBD_CONNECT_GRANT_PARAM.
449 * That's 4KB=2^12 which is the biggest block size known to work whatever
450 * the client's page size is. */
451 #define COMPAT_BSIZE_SHIFT 12
452 static inline int ofd_grant_compat(struct obd_export *exp,
453 struct ofd_device *ofd)
455 /* Clients which don't support OBD_CONNECT_GRANT_PARAM cannot handle
456 * a block size > page size and consume PAGE_CACHE_SIZE of grant when
457 * dirtying a page regardless of the block size */
458 return !!(ofd_obd(ofd)->obd_self_export != exp &&
459 ofd->ofd_blockbits > COMPAT_BSIZE_SHIFT &&
460 !ofd_grant_param_supp(exp));
463 static inline int ofd_grant_prohibit(struct obd_export *exp,
464 struct ofd_device *ofd)
466 /* When ofd_grant_compat_disable is set, we don't grant any space to
467 * clients not supporting OBD_CONNECT_GRANT_PARAM.
468 * Otherwise, space granted to such a client is inflated since it
469 * consumes PAGE_CACHE_SIZE of grant space per block */
470 return !!(ofd_grant_compat(exp, ofd) && ofd->ofd_grant_compat_disable);
473 void ofd_grant_sanity_check(struct obd_device *obd, const char *func);
474 long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
475 u64 want, bool new_conn);
476 void ofd_grant_discard(struct obd_export *exp);
477 void ofd_grant_prepare_read(const struct lu_env *env, struct obd_export *exp,
479 void ofd_grant_prepare_write(const struct lu_env *env, struct obd_export *exp,
480 struct obdo *oa, struct niobuf_remote *rnb,
482 void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp, int rc);
483 int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr);
486 int ofd_fmd_init(void);
487 void ofd_fmd_exit(void);
488 struct ofd_mod_data *ofd_fmd_find(struct obd_export *exp,
489 const struct lu_fid *fid);
490 struct ofd_mod_data *ofd_fmd_get(struct obd_export *exp,
491 const struct lu_fid *fid);
492 void ofd_fmd_put(struct obd_export *exp, struct ofd_mod_data *fmd);
493 void ofd_fmd_expire(struct obd_export *exp);
494 void ofd_fmd_cleanup(struct obd_export *exp);
496 void ofd_fmd_drop(struct obd_export *exp, const struct lu_fid *fid);
498 #define ofd_fmd_drop(exp, fid) do {} while (0)
502 int ofd_fid_set_index(const struct lu_env *env, struct ofd_device *ofd,
504 int ofd_fid_init(const struct lu_env *env, struct ofd_device *ofd);
505 int ofd_fid_fini(const struct lu_env *env, struct ofd_device *ofd);
508 extern struct ldlm_valblock_ops ofd_lvbo;
511 int ofd_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp,
512 void *req_cookie, ldlm_mode_t mode, __u64 flags,
515 static inline struct ofd_thread_info *ofd_info(const struct lu_env *env)
517 struct ofd_thread_info *info;
519 lu_env_refill((void *)env);
520 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
525 static inline struct ofd_thread_info *ofd_info_init(const struct lu_env *env,
526 struct obd_export *exp)
528 struct ofd_thread_info *info;
530 info = ofd_info(env);
531 LASSERT(info->fti_exp == NULL);
532 LASSERT(info->fti_env == NULL);
533 LASSERT(info->fti_attr.la_valid == 0);
537 info->fti_pre_version = 0;
541 static inline struct ofd_thread_info *tsi2ofd_info(struct tgt_session_info *tsi)
543 struct ptlrpc_request *req = tgt_ses_req(tsi);
544 struct ofd_thread_info *info;
546 info = ofd_info(tsi->tsi_env);
547 LASSERT(info->fti_exp == NULL);
548 LASSERT(info->fti_env == NULL);
549 LASSERT(info->fti_attr.la_valid == 0);
551 info->fti_env = tsi->tsi_env;
552 info->fti_exp = tsi->tsi_exp;
554 info->fti_xid = req->rq_xid;
555 /** VBR: take versions from request */
556 if (req->rq_reqmsg != NULL &&
557 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
558 __u64 *pre_version = lustre_msg_get_versions(req->rq_reqmsg);
560 info->fti_pre_version = pre_version ? pre_version[0] : 0;
565 /* sync on lock cancel is useless when we force a journal flush,
566 * and if we enable async journal commit, we should also turn on
567 * sync on lock cancel if it is not enabled already. */
568 static inline void ofd_slc_set(struct ofd_device *ofd)
570 if (ofd->ofd_syncjournal == 1)
571 ofd->ofd_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
572 else if (ofd->ofd_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
573 ofd->ofd_lut.lut_sync_lock_cancel = ALWAYS_SYNC_ON_CANCEL;
576 static inline void ofd_prepare_fidea(struct filter_fid *ff,
577 const struct obdo *oa)
579 /* packing fid and converting it to LE for storing into EA.
580 * Here ->o_stripe_idx should be filled by LOV and rest of
581 * fields - by client. */
582 ff->ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
583 ff->ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
584 /* XXX: we are ignoring o_parent_ver here, since this should
585 * be the same for all objects in this fileset. */
586 ff->ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
589 #endif /* _OFD_INTERNAL_H */