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, 2013, 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 #include <lustre_capa.h>
48 #define OFD_INIT_OBJID 0
49 #define OFD_ROCOMPAT_SUPP (0)
50 #define OFD_INCOMPAT_SUPP (OBD_INCOMPAT_GROUPS | OBD_INCOMPAT_OST | \
51 OBD_INCOMPAT_COMMON_LR)
52 #define OFD_PRECREATE_BATCH_DEFAULT (OBJ_SUBDIR_COUNT * 4)
54 /* on small filesystems we should not precreate too many objects in
55 * a single transaction, otherwise we can overflow transactions */
56 #define OFD_PRECREATE_SMALL_FS (1024ULL * 1024 * 1024)
57 #define OFD_PRECREATE_BATCH_SMALL 8
59 /* Limit the returned fields marked valid to those that we actually might set */
60 #define OFD_VALID_FLAGS (LA_TYPE | LA_MODE | LA_SIZE | LA_BLOCKS | \
61 LA_BLKSIZE | LA_ATIME | LA_MTIME | LA_CTIME)
63 /* per-client-per-object persistent state (LRU) */
65 cfs_list_t fmd_list; /* linked to fed_mod_list */
66 struct lu_fid fmd_fid; /* FID being written to */
67 __u64 fmd_mactime_xid; /* xid highest {m,a,c}time setattr */
68 cfs_time_t fmd_expire; /* time when the fmd should expire */
69 int fmd_refcount; /* reference counter - list holds 1 */
72 #define OFD_FMD_MAX_NUM_DEFAULT 128
73 #define OFD_FMD_MAX_AGE_DEFAULT ((obd_timeout + 10) * HZ)
75 #define OFD_SOFT_SYNC_LIMIT_DEFAULT 16
79 LPROC_OFD_STATS_READ = 0,
80 LPROC_OFD_STATS_WRITE,
81 LPROC_OFD_STATS_GETATTR,
82 LPROC_OFD_STATS_SETATTR,
83 LPROC_OFD_STATS_PUNCH,
85 LPROC_OFD_STATS_DESTROY,
86 LPROC_OFD_STATS_CREATE,
87 LPROC_OFD_STATS_STATFS,
88 LPROC_OFD_STATS_GET_INFO,
89 LPROC_OFD_STATS_SET_INFO,
90 LPROC_OFD_STATS_QUOTACTL,
94 static inline void ofd_counter_incr(struct obd_export *exp, int opcode,
95 char *jobid, long amount)
97 if (exp->exp_obd && exp->exp_obd->obd_stats)
98 lprocfs_counter_add(exp->exp_obd->obd_stats, opcode, amount);
100 if (exp->exp_obd && exp->exp_obd->u.obt.obt_jobstats.ojs_hash &&
101 (exp_connect_flags(exp) & OBD_CONNECT_JOBSTATS))
102 lprocfs_job_stats_log(exp->exp_obd, jobid, opcode, amount);
104 if (exp->exp_nid_stats != NULL &&
105 exp->exp_nid_stats->nid_stats != NULL) {
106 lprocfs_counter_add(exp->exp_nid_stats->nid_stats, opcode,
114 spinlock_t os_last_oid_lock;
115 struct mutex os_create_lock;
117 struct dt_object *os_lastid_obj;
118 unsigned long os_destroys_in_progress:1;
122 struct dt_device ofd_dt_dev;
123 struct dt_device *ofd_osd;
124 struct obd_export *ofd_osd_exp;
125 struct dt_device_param ofd_dt_conf;
126 /* DLM name-space for meta-data locks maintained by this server */
127 struct ldlm_namespace *ofd_namespace;
130 struct lu_target ofd_lut;
131 struct dt_object *ofd_health_check_file;
133 int ofd_subdir_count;
134 __u64 ofd_inconsistency_self_detected;
135 __u64 ofd_inconsistency_self_repaired;
137 cfs_list_t ofd_seq_list;
138 rwlock_t ofd_seq_list_lock;
140 int ofd_precreate_batch;
141 spinlock_t ofd_batch_lock;
143 /* protect all statfs-related counters */
144 spinlock_t ofd_osfs_lock;
145 /* statfs optimization: we cache a bit */
146 struct obd_statfs ofd_osfs;
149 /* writes which might be be accounted twice in ofd_osfs.os_bavail */
150 obd_size ofd_osfs_unstable;
152 /* counters used during statfs update, protected by ofd_osfs_lock.
153 * record when some statfs refresh are in progress */
154 int ofd_statfs_inflight;
155 /* track writes completed while statfs refresh is underway.
156 * tracking is only effective when ofd_statfs_inflight > 1 */
157 obd_size ofd_osfs_inflight;
159 /* grants: all values in bytes */
160 /* grant lock to protect all grant counters */
161 spinlock_t ofd_grant_lock;
162 /* total amount of dirty data reported by clients in incoming obdo */
163 obd_size ofd_tot_dirty;
164 /* sum of filesystem space granted to clients for async writes */
165 obd_size ofd_tot_granted;
166 /* grant used by I/Os in progress (between prepare and commit) */
167 obd_size ofd_tot_pending;
168 /* free space threshold over which we stop granting space to clients
169 * ofd_grant_ratio is stored as a fixed-point fraction using
170 * OFD_GRANT_RATIO_SHIFT of the remaining free space, not in percentage
173 /* number of clients using grants */
174 int ofd_tot_granted_clients;
176 /* ofd mod data: ofd_device wide values */
177 int ofd_fmd_max_num; /* per ofd ofd_mod_data */
178 cfs_duration_t ofd_fmd_max_age; /* time to fmd expiry */
180 spinlock_t ofd_flags_lock;
181 unsigned long ofd_raid_degraded:1,
182 /* sync journal on writes */
184 /* shall we grant space to clients not
185 * supporting OBD_CONNECT_GRANT_PARAM? */
186 ofd_grant_compat_disable:1,
187 /* Protected by ofd_lastid_rwsem. */
188 ofd_lastid_rebuilding:1,
189 ofd_record_fid_accessed:1,
190 ofd_lfsck_verify_pfid:1;
191 struct seq_server_site ofd_seq_site;
192 /* the limit of SOFT_SYNC RPCs that will trigger a soft sync */
193 unsigned int ofd_soft_sync_limit;
194 /* Protect ::ofd_lastid_rebuilding */
195 struct rw_semaphore ofd_lastid_rwsem;
196 __u64 ofd_lastid_gen;
197 struct ptlrpc_thread ofd_inconsistency_thread;
198 struct list_head ofd_inconsistency_list;
199 spinlock_t ofd_inconsistency_lock;
202 static inline struct ofd_device *ofd_dev(struct lu_device *d)
204 return container_of0(d, struct ofd_device, ofd_dt_dev.dd_lu_dev);
207 static inline struct obd_device *ofd_obd(struct ofd_device *ofd)
209 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd;
212 static inline struct ofd_device *ofd_exp(struct obd_export *exp)
214 return ofd_dev(exp->exp_obd->obd_lu_dev);
217 static inline char *ofd_name(struct ofd_device *ofd)
219 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd->obd_name;
223 struct lu_object_header ofo_header;
224 struct dt_object ofo_obj;
225 struct lu_fid ofo_pfid;
226 unsigned int ofo_pfid_checking:1,
230 static inline struct ofd_object *ofd_obj(struct lu_object *o)
232 return container_of0(o, struct ofd_object, ofo_obj.do_lu);
235 static inline int ofd_object_exists(struct ofd_object *obj)
237 LASSERT(obj != NULL);
238 if (lu_object_is_dying(obj->ofo_obj.do_lu.lo_header))
240 return lu_object_exists(&obj->ofo_obj.do_lu);
243 static inline struct dt_object *fo2dt(struct ofd_object *obj)
245 return &obj->ofo_obj;
248 static inline struct dt_object *ofd_object_child(struct ofd_object *_obj)
250 struct lu_object *lu = &(_obj)->ofo_obj.do_lu;
252 return container_of0(lu_object_next(lu), struct dt_object, do_lu);
255 static inline struct ofd_device *ofd_obj2dev(const struct ofd_object *fo)
257 return ofd_dev(fo->ofo_obj.do_lu.lo_dev);
260 static inline struct lustre_capa *ofd_object_capa(const struct lu_env *env,
261 const struct ofd_object *obj)
263 /* TODO: see mdd_object_capa() */
267 static inline void ofd_read_lock(const struct lu_env *env,
268 struct ofd_object *fo)
270 struct dt_object *next = ofd_object_child(fo);
272 next->do_ops->do_read_lock(env, next, 0);
275 static inline void ofd_read_unlock(const struct lu_env *env,
276 struct ofd_object *fo)
278 struct dt_object *next = ofd_object_child(fo);
280 next->do_ops->do_read_unlock(env, next);
283 static inline void ofd_write_lock(const struct lu_env *env,
284 struct ofd_object *fo)
286 struct dt_object *next = ofd_object_child(fo);
288 next->do_ops->do_write_lock(env, next, 0);
291 static inline void ofd_write_unlock(const struct lu_env *env,
292 struct ofd_object *fo)
294 struct dt_object *next = ofd_object_child(fo);
296 next->do_ops->do_write_unlock(env, next);
300 * Common data shared by obdofd-level handlers. This is allocated per-thread
301 * to reduce stack consumption.
303 struct ofd_thread_info {
304 const struct lu_env *fti_env;
306 struct obd_export *fti_exp;
308 __u64 fti_pre_version;
310 struct lu_fid fti_fid;
311 struct lu_attr fti_attr;
312 struct lu_attr fti_attr2;
313 struct ldlm_res_id fti_resid;
315 struct filter_fid fti_mds_fid;
316 struct filter_fid_old fti_mds_fid_old;
318 struct ost_id fti_ostid;
319 struct ofd_object *fti_obj;
321 char name[64]; /* for ofd_init0() */
322 struct obd_statfs osfs; /* for obdofd_statfs() */
325 /* Ops object filename */
326 struct lu_name fti_name;
327 struct dt_object_format fti_dof;
328 struct lu_buf fti_buf;
331 /* Space used by the I/O, used by grant code */
332 unsigned long fti_used;
333 struct ost_lvb fti_lvb;
334 struct lfsck_request fti_lr;
337 extern void target_recovery_fini(struct obd_device *obd);
338 extern void target_recovery_init(struct lu_target *lut, svc_handler_t handler);
341 int ofd_update_capa_key(struct ofd_device *ofd, struct lustre_capa_key *key);
342 int ofd_auth_capa(struct obd_export *exp, const struct lu_fid *fid,
343 obd_seq seq, struct lustre_capa *capa, __u64 opc);
344 void ofd_free_capa_keys(struct ofd_device *ofd);
347 extern struct lu_context_key ofd_thread_key;
348 int ofd_postrecov(const struct lu_env *env, struct ofd_device *ofd);
349 int ofd_fiemap_get(const struct lu_env *env, struct ofd_device *ofd,
350 struct lu_fid *fid, struct ll_user_fiemap *fiemap);
353 extern struct obd_ops ofd_obd_ops;
354 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
355 struct obd_statfs *osfs, __u64 max_age,
357 int ofd_destroy_by_fid(const struct lu_env *env, struct ofd_device *ofd,
358 const struct lu_fid *fid, int orphan);
359 int ofd_statfs(const struct lu_env *env, struct obd_export *exp,
360 struct obd_statfs *osfs, __u64 max_age, __u32 flags);
361 int ofd_obd_disconnect(struct obd_export *exp);
364 obd_id ofd_seq_last_oid(struct ofd_seq *oseq);
365 void ofd_seq_last_oid_set(struct ofd_seq *oseq, obd_id id);
366 int ofd_seq_last_oid_write(const struct lu_env *env, struct ofd_device *ofd,
367 struct ofd_seq *oseq);
368 int ofd_seqs_init(const struct lu_env *env, struct ofd_device *ofd);
369 struct ofd_seq *ofd_seq_get(struct ofd_device *ofd, obd_seq seq);
370 void ofd_seq_put(const struct lu_env *env, struct ofd_seq *oseq);
372 int ofd_fs_setup(const struct lu_env *env, struct ofd_device *ofd,
373 struct obd_device *obd);
374 void ofd_fs_cleanup(const struct lu_env *env, struct ofd_device *ofd);
375 int ofd_precreate_batch(struct ofd_device *ofd, int batch);
376 struct ofd_seq *ofd_seq_load(const struct lu_env *env, struct ofd_device *ofd,
378 void ofd_seqs_fini(const struct lu_env *env, struct ofd_device *ofd);
379 void ofd_seqs_free(const struct lu_env *env, struct ofd_device *ofd);
382 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd);
383 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd);
384 int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
386 int ofd_preprw(const struct lu_env *env,int cmd, struct obd_export *exp,
387 struct obdo *oa, int objcount, struct obd_ioobj *obj,
388 struct niobuf_remote *rnb, int *nr_local,
389 struct niobuf_local *lnb, struct obd_trans_info *oti,
390 struct lustre_capa *capa);
391 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
392 struct obdo *oa, int objcount, struct obd_ioobj *obj,
393 struct niobuf_remote *rnb, int npages,
394 struct niobuf_local *lnb, struct obd_trans_info *oti,
398 struct thandle *ofd_trans_create(const struct lu_env *env,
399 struct ofd_device *ofd);
400 int ofd_trans_start(const struct lu_env *env,
401 struct ofd_device *ofd, struct ofd_object *fo,
403 void ofd_trans_stop(const struct lu_env *env, struct ofd_device *ofd,
404 struct thandle *th, int rc);
405 int ofd_txn_stop_cb(const struct lu_env *env, struct thandle *txn,
410 void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars);
411 void ofd_stats_counter_init(struct lprocfs_stats *stats);
413 static void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars)
415 memset(lvars, 0, sizeof(*lvars));
417 static inline void ofd_stats_counter_init(struct lprocfs_stats *stats) {}
421 struct ofd_object *ofd_object_find(const struct lu_env *env,
422 struct ofd_device *ofd,
423 const struct lu_fid *fid);
424 struct ofd_object *ofd_object_find_or_create(const struct lu_env *env,
425 struct ofd_device *ofd,
426 const struct lu_fid *fid,
427 struct lu_attr *attr);
428 int ofd_object_ff_load(const struct lu_env *env, struct ofd_object *fo);
429 int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd,
430 obd_id id, struct ofd_seq *oseq, int nr, int sync);
432 void ofd_object_put(const struct lu_env *env, struct ofd_object *fo);
433 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
434 struct lu_attr *la, struct filter_fid *ff);
435 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
436 __u64 start, __u64 end, struct lu_attr *la,
437 struct filter_fid *ff, struct obdo *oa);
438 int ofd_object_destroy(const struct lu_env *, struct ofd_object *, int);
439 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
441 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
442 struct lu_attr *la, int is_setattr);
445 struct ofd_object *ofd_object_find_exists(const struct lu_env *env,
446 struct ofd_device *ofd,
447 const struct lu_fid *fid)
449 struct ofd_object *fo;
451 fo = ofd_object_find(env, ofd, fid);
452 if (!IS_ERR(fo) && !ofd_object_exists(fo)) {
453 ofd_object_put(env, fo);
454 fo = ERR_PTR(-ENOENT);
460 #define OFD_GRANT_RATIO_SHIFT 8
461 static inline __u64 ofd_grant_reserved(struct ofd_device *ofd, obd_size bavail)
463 return (bavail * ofd->ofd_grant_ratio) >> OFD_GRANT_RATIO_SHIFT;
466 static inline int ofd_grant_ratio_conv(int percentage)
468 return (percentage << OFD_GRANT_RATIO_SHIFT) / 100;
471 static inline int ofd_grant_param_supp(struct obd_export *exp)
473 return !!(exp_connect_flags(exp) & OBD_CONNECT_GRANT_PARAM);
476 /* Blocksize used for client not supporting OBD_CONNECT_GRANT_PARAM.
477 * That's 4KB=2^12 which is the biggest block size known to work whatever
478 * the client's page size is. */
479 #define COMPAT_BSIZE_SHIFT 12
480 static inline int ofd_grant_compat(struct obd_export *exp,
481 struct ofd_device *ofd)
483 /* Clients which don't support OBD_CONNECT_GRANT_PARAM cannot handle
484 * a block size > page size and consume PAGE_CACHE_SIZE of grant when
485 * dirtying a page regardless of the block size */
486 return !!(ofd_obd(ofd)->obd_self_export != exp &&
487 ofd->ofd_blockbits > COMPAT_BSIZE_SHIFT &&
488 !ofd_grant_param_supp(exp));
491 static inline int ofd_grant_prohibit(struct obd_export *exp,
492 struct ofd_device *ofd)
494 /* When ofd_grant_compat_disable is set, we don't grant any space to
495 * clients not supporting OBD_CONNECT_GRANT_PARAM.
496 * Otherwise, space granted to such a client is inflated since it
497 * consumes PAGE_CACHE_SIZE of grant space per block */
498 return !!(ofd_grant_compat(exp, ofd) && ofd->ofd_grant_compat_disable);
501 void ofd_grant_sanity_check(struct obd_device *obd, const char *func);
502 long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
503 obd_size want, bool conservative);
504 void ofd_grant_discard(struct obd_export *exp);
505 void ofd_grant_prepare_read(const struct lu_env *env, struct obd_export *exp,
507 void ofd_grant_prepare_write(const struct lu_env *env, struct obd_export *exp,
508 struct obdo *oa, struct niobuf_remote *rnb,
510 void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp, int rc);
511 int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr);
514 int ofd_fmd_init(void);
515 void ofd_fmd_exit(void);
516 struct ofd_mod_data *ofd_fmd_find(struct obd_export *exp,
517 const struct lu_fid *fid);
518 struct ofd_mod_data *ofd_fmd_get(struct obd_export *exp,
519 const struct lu_fid *fid);
520 void ofd_fmd_put(struct obd_export *exp, struct ofd_mod_data *fmd);
521 void ofd_fmd_expire(struct obd_export *exp);
522 void ofd_fmd_cleanup(struct obd_export *exp);
524 void ofd_fmd_drop(struct obd_export *exp, const struct lu_fid *fid);
526 #define ofd_fmd_drop(exp, fid) do {} while (0)
530 int ofd_fid_set_index(const struct lu_env *env, struct ofd_device *ofd,
532 int ofd_fid_init(const struct lu_env *env, struct ofd_device *ofd);
533 int ofd_fid_fini(const struct lu_env *env, struct ofd_device *ofd);
536 extern struct ldlm_valblock_ops ofd_lvbo;
539 int ofd_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp,
540 void *req_cookie, ldlm_mode_t mode, __u64 flags,
543 static inline struct ofd_thread_info *ofd_info(const struct lu_env *env)
545 struct ofd_thread_info *info;
547 lu_env_refill((void *)env);
548 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
553 static inline struct ofd_thread_info *ofd_info_init(const struct lu_env *env,
554 struct obd_export *exp)
556 struct ofd_thread_info *info;
558 info = ofd_info(env);
559 LASSERT(info->fti_exp == NULL);
560 LASSERT(info->fti_env == NULL);
561 LASSERT(info->fti_attr.la_valid == 0);
565 info->fti_pre_version = 0;
569 static inline struct ofd_thread_info *tsi2ofd_info(struct tgt_session_info *tsi)
571 struct ptlrpc_request *req = tgt_ses_req(tsi);
572 struct ofd_thread_info *info;
574 info = ofd_info(tsi->tsi_env);
575 LASSERT(info->fti_exp == NULL);
576 LASSERT(info->fti_env == NULL);
577 LASSERT(info->fti_attr.la_valid == 0);
579 info->fti_env = tsi->tsi_env;
580 info->fti_exp = tsi->tsi_exp;
582 info->fti_xid = req->rq_xid;
583 /** VBR: take versions from request */
584 if (req->rq_reqmsg != NULL &&
585 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
586 __u64 *pre_version = lustre_msg_get_versions(req->rq_reqmsg);
588 info->fti_pre_version = pre_version ? pre_version[0] : 0;
593 static inline void ofd_oti2info(struct ofd_thread_info *info,
594 struct obd_trans_info *oti)
596 info->fti_xid = oti->oti_xid;
597 info->fti_pre_version = oti->oti_pre_version;
600 static inline void ofd_info2oti(struct ofd_thread_info *info,
601 struct obd_trans_info *oti)
603 oti->oti_xid = info->fti_xid;
604 oti->oti_pre_version = info->fti_pre_version;
607 /* sync on lock cancel is useless when we force a journal flush,
608 * and if we enable async journal commit, we should also turn on
609 * sync on lock cancel if it is not enabled already. */
610 static inline void ofd_slc_set(struct ofd_device *ofd)
612 if (ofd->ofd_syncjournal == 1)
613 ofd->ofd_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
614 else if (ofd->ofd_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
615 ofd->ofd_lut.lut_sync_lock_cancel = ALWAYS_SYNC_ON_CANCEL;
618 static inline void ofd_prepare_fidea(struct filter_fid *ff,
619 const struct obdo *oa)
621 /* packing fid and converting it to LE for storing into EA.
622 * Here ->o_stripe_idx should be filled by LOV and rest of
623 * fields - by client. */
624 ff->ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
625 ff->ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
626 /* XXX: we are ignoring o_parent_ver here, since this should
627 * be the same for all objects in this fileset. */
628 ff->ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
631 /* niobuf_remote has no rnb_ prefix in master */
632 #define rnb_offset offset
633 #define rnb_flags flags
635 /* the same for niobuf_local */
636 #define lnb_flags flags
639 #endif /* _OFD_INTERNAL_H */