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, 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 <lustre_fid.h>
45 #include <lustre_capa.h>
47 #define OFD_INIT_OBJID 0
48 #define OFD_ROCOMPAT_SUPP (0)
49 #define OFD_INCOMPAT_SUPP (OBD_INCOMPAT_GROUPS | OBD_INCOMPAT_OST | \
50 OBD_INCOMPAT_COMMON_LR)
51 #define OFD_MAX_GROUPS 256
52 #define OFD_PRECREATE_BATCH_DEFAULT (FILTER_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) * CFS_HZ)
76 LPROC_OFD_READ_BYTES = 0,
77 LPROC_OFD_WRITE_BYTES = 1,
83 LPROC_OFD_STATS_READ = 0,
84 LPROC_OFD_STATS_WRITE = 1,
85 LPROC_OFD_STATS_SETATTR = 2,
86 LPROC_OFD_STATS_PUNCH = 3,
87 LPROC_OFD_STATS_SYNC = 4,
91 static inline void ofd_counter_incr(struct obd_export *exp, int opcode,
92 char *jobid, long amount)
94 if (exp->exp_obd && exp->exp_obd->u.obt.obt_jobstats.ojs_hash &&
95 (exp->exp_connect_flags & OBD_CONNECT_JOBSTATS))
96 lprocfs_job_stats_log(exp->exp_obd, jobid, opcode, amount);
100 struct dt_device ofd_dt_dev;
101 struct dt_device *ofd_osd;
102 struct obd_export *ofd_osd_exp;
103 struct dt_device_param ofd_dt_conf;
104 /* DLM name-space for meta-data locks maintained by this server */
105 struct ldlm_namespace *ofd_namespace;
107 /* transaction callbacks */
108 struct dt_txn_callback ofd_txn_cb;
111 struct lu_target ofd_lut;
112 struct dt_object *ofd_last_group_file;
113 struct dt_object *ofd_health_check_file;
115 int ofd_subdir_count;
118 obd_id ofd_last_objids[OFD_MAX_GROUPS];
119 struct mutex ofd_create_locks[OFD_MAX_GROUPS];
120 struct dt_object *ofd_lastid_obj[OFD_MAX_GROUPS];
121 spinlock_t ofd_objid_lock;
122 unsigned long ofd_destroys_in_progress;
123 int ofd_precreate_batch;
125 /* protect all statfs-related counters */
126 spinlock_t ofd_osfs_lock;
127 /* statfs optimization: we cache a bit */
128 struct obd_statfs ofd_osfs;
131 /* writes which might be be accounted twice in ofd_osfs.os_bavail */
132 obd_size ofd_osfs_unstable;
134 /* counters used during statfs update, protected by ofd_osfs_lock.
135 * record when some statfs refresh are in progress */
136 int ofd_statfs_inflight;
137 /* track writes completed while statfs refresh is underway.
138 * tracking is only effective when ofd_statfs_inflight > 1 */
139 obd_size ofd_osfs_inflight;
141 /* grants: all values in bytes */
142 /* grant lock to protect all grant counters */
143 spinlock_t ofd_grant_lock;
144 /* total amount of dirty data reported by clients in incoming obdo */
145 obd_size ofd_tot_dirty;
146 /* sum of filesystem space granted to clients for async writes */
147 obd_size ofd_tot_granted;
148 /* grant used by I/Os in progress (between prepare and commit) */
149 obd_size ofd_tot_pending;
150 /* free space threshold over which we stop granting space to clients
151 * ofd_grant_ratio is stored as a fixed-point fraction using
152 * OFD_GRANT_RATIO_SHIFT of the remaining free space, not in percentage
155 /* number of clients using grants */
156 int ofd_tot_granted_clients;
158 /* ofd mod data: ofd_device wide values */
159 int ofd_fmd_max_num; /* per ofd ofd_mod_data */
160 cfs_duration_t ofd_fmd_max_age; /* time to fmd expiry */
162 spinlock_t ofd_flags_lock;
163 unsigned long ofd_raid_degraded:1,
164 /* sync journal on writes */
166 /* sync on lock cancel */
167 ofd_sync_lock_cancel:2,
168 /* shall we grant space to clients not
169 * supporting OBD_CONNECT_GRANT_PARAM? */
170 ofd_grant_compat_disable:1;
173 static inline struct ofd_device *ofd_dev(struct lu_device *d)
175 return container_of0(d, struct ofd_device, ofd_dt_dev.dd_lu_dev);
178 static inline struct obd_device *ofd_obd(struct ofd_device *ofd)
180 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd;
183 static inline struct ofd_device *ofd_exp(struct obd_export *exp)
185 return ofd_dev(exp->exp_obd->obd_lu_dev);
188 static inline char *ofd_name(struct ofd_device *ofd)
190 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd->obd_name;
194 struct lu_object_header ofo_header;
195 struct dt_object ofo_obj;
199 static inline struct ofd_object *ofd_obj(struct lu_object *o)
201 return container_of0(o, struct ofd_object, ofo_obj.do_lu);
204 static inline int ofd_object_exists(struct ofd_object *obj)
206 LASSERT(obj != NULL);
207 if (lu_object_is_dying(obj->ofo_obj.do_lu.lo_header))
209 return lu_object_exists(&obj->ofo_obj.do_lu);
212 static inline struct dt_object *fo2dt(struct ofd_object *obj)
214 return &obj->ofo_obj;
217 static inline struct dt_object *ofd_object_child(struct ofd_object *_obj)
219 struct lu_object *lu = &(_obj)->ofo_obj.do_lu;
221 return container_of0(lu_object_next(lu), struct dt_object, do_lu);
224 static inline struct ofd_device *ofd_obj2dev(const struct ofd_object *fo)
226 return ofd_dev(fo->ofo_obj.do_lu.lo_dev);
229 static inline struct lustre_capa *ofd_object_capa(const struct lu_env *env,
230 const struct ofd_object *obj)
232 /* TODO: see mdd_object_capa() */
236 static inline void ofd_read_lock(const struct lu_env *env,
237 struct ofd_object *fo)
239 struct dt_object *next = ofd_object_child(fo);
241 next->do_ops->do_read_lock(env, next, 0);
244 static inline void ofd_read_unlock(const struct lu_env *env,
245 struct ofd_object *fo)
247 struct dt_object *next = ofd_object_child(fo);
249 next->do_ops->do_read_unlock(env, next);
252 static inline void ofd_write_lock(const struct lu_env *env,
253 struct ofd_object *fo)
255 struct dt_object *next = ofd_object_child(fo);
257 next->do_ops->do_write_lock(env, next, 0);
260 static inline void ofd_write_unlock(const struct lu_env *env,
261 struct ofd_object *fo)
263 struct dt_object *next = ofd_object_child(fo);
265 next->do_ops->do_write_unlock(env, next);
269 * Common data shared by obdofd-level handlers. This is allocated per-thread
270 * to reduce stack consumption.
272 struct ofd_thread_info {
273 const struct lu_env *fti_env;
275 struct obd_export *fti_exp;
278 __u64 fti_pre_version;
279 __u32 fti_has_trans:1, /* has txn already */
282 struct lu_fid fti_fid;
283 struct lu_attr fti_attr;
284 struct lu_attr fti_attr2;
285 struct ldlm_res_id fti_resid;
286 struct filter_fid fti_mds_fid;
287 struct filter_fid fti_mds_fid2;
288 struct ost_id fti_ostid;
289 struct ofd_object *fti_obj;
291 char name[64]; /* for ofd_init0() */
292 struct obd_statfs osfs; /* for obdofd_statfs() */
295 /* Ops object filename */
296 struct lu_name fti_name;
297 struct dt_object_format fti_dof;
298 struct lu_buf fti_buf;
301 /* Space used by the I/O, used by grant code */
302 unsigned long fti_used;
303 struct ost_lvb fti_lvb;
306 extern void target_recovery_fini(struct obd_device *obd);
307 extern void target_recovery_init(struct lu_target *lut, svc_handler_t handler);
310 int ofd_update_capa_key(struct ofd_device *ofd, struct lustre_capa_key *key);
311 int ofd_auth_capa(struct obd_export *exp, struct lu_fid *fid, obd_seq seq,
312 struct lustre_capa *capa, __u64 opc);
313 void ofd_free_capa_keys(struct ofd_device *ofd);
316 extern struct lu_context_key ofd_thread_key;
317 int ofd_postrecov(const struct lu_env *env, struct ofd_device *ofd);
320 extern struct obd_ops ofd_obd_ops;
321 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
322 struct obd_statfs *osfs, __u64 max_age,
326 obd_id ofd_last_id(struct ofd_device *ofd, obd_seq seq);
327 void ofd_last_id_set(struct ofd_device *ofd, obd_id id, obd_seq seq);
328 int ofd_last_id_write(const struct lu_env *env, struct ofd_device *ofd,
330 int ofd_group_load(const struct lu_env *env, struct ofd_device *ofd, int);
331 int ofd_fs_setup(const struct lu_env *env, struct ofd_device *ofd,
332 struct obd_device *obd);
333 void ofd_fs_cleanup(const struct lu_env *env, struct ofd_device *ofd);
334 int ofd_precreate_batch(struct ofd_device *ofd, int batch);
337 int ofd_preprw(const struct lu_env *env,int cmd, struct obd_export *exp,
338 struct obdo *oa, int objcount, struct obd_ioobj *obj,
339 struct niobuf_remote *rnb, int *nr_local,
340 struct niobuf_local *lnb, struct obd_trans_info *oti,
341 struct lustre_capa *capa);
342 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
343 struct obdo *oa, int objcount, struct obd_ioobj *obj,
344 struct niobuf_remote *rnb, int npages,
345 struct niobuf_local *lnb, struct obd_trans_info *oti,
349 struct thandle *ofd_trans_create(const struct lu_env *env,
350 struct ofd_device *ofd);
351 int ofd_trans_start(const struct lu_env *env,
352 struct ofd_device *ofd, struct ofd_object *fo,
354 void ofd_trans_stop(const struct lu_env *env, struct ofd_device *ofd,
355 struct thandle *th, int rc);
356 int ofd_txn_stop_cb(const struct lu_env *env, struct thandle *txn,
361 void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars);
362 int lproc_ofd_attach_seqstat(struct obd_device *dev);
363 extern struct file_operations ofd_per_nid_stats_fops;
364 void ofd_stats_counter_init(struct lprocfs_stats *stats);
366 static void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars)
368 memset(lvars, 0, sizeof(*lvars));
370 static inline int lproc_ofd_attach_seqstat(struct obd_device *dev) {}
371 static inline void ofd_stats_counter_init(struct lprocfs_stats *stats) {}
375 struct ofd_object *ofd_object_find(const struct lu_env *env,
376 struct ofd_device *ofd,
377 const struct lu_fid *fid);
378 struct ofd_object *ofd_object_find_or_create(const struct lu_env *env,
379 struct ofd_device *ofd,
380 const struct lu_fid *fid,
381 struct lu_attr *attr);
382 int ofd_object_ff_check(const struct lu_env *env, struct ofd_object *fo);
383 int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd,
384 obd_id id, obd_seq group, int nr);
386 void ofd_object_put(const struct lu_env *env, struct ofd_object *fo);
387 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
388 struct lu_attr *la, struct filter_fid *ff);
389 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
390 __u64 start, __u64 end, struct lu_attr *la,
391 struct filter_fid *ff);
392 int ofd_object_destroy(const struct lu_env *, struct ofd_object *, int);
393 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
395 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
396 struct lu_attr *la, int is_setattr);
399 #define OFD_GRANT_RATIO_SHIFT 8
400 static inline __u64 ofd_grant_reserved(struct ofd_device *ofd, obd_size bavail)
402 return (bavail * ofd->ofd_grant_ratio) >> OFD_GRANT_RATIO_SHIFT;
405 static inline int ofd_grant_ratio_conv(int percentage)
407 return (percentage << OFD_GRANT_RATIO_SHIFT) / 100;
410 static inline int ofd_grant_param_supp(struct obd_export *exp)
412 return !!(exp->exp_connect_flags & OBD_CONNECT_GRANT_PARAM);
415 /* Blocksize used for client not supporting OBD_CONNECT_GRANT_PARAM.
416 * That's 4KB=2^12 which is the biggest block size known to work whatever
417 * the client's page size is. */
418 #define COMPAT_BSIZE_SHIFT 12
419 static inline int ofd_grant_compat(struct obd_export *exp,
420 struct ofd_device *ofd)
422 /* Clients which don't support OBD_CONNECT_GRANT_PARAM cannot handle
423 * a block size > page size and consume CFS_PAGE_SIZE of grant when
424 * dirtying a page regardless of the block size */
425 return !!(ofd_obd(ofd)->obd_self_export != exp &&
426 ofd->ofd_blockbits > COMPAT_BSIZE_SHIFT &&
427 !ofd_grant_param_supp(exp));
430 static inline int ofd_grant_prohibit(struct obd_export *exp,
431 struct ofd_device *ofd)
433 /* When ofd_grant_compat_disable is set, we don't grant any space to
434 * clients not supporting OBD_CONNECT_GRANT_PARAM.
435 * Otherwise, space granted to such a client is inflated since it
436 * consumes CFS_PAGE_SIZE of grant space per block */
437 return !!(ofd_grant_compat(exp, ofd) && ofd->ofd_grant_compat_disable);
440 void ofd_grant_sanity_check(struct obd_device *obd, const char *func);
441 long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
443 void ofd_grant_discard(struct obd_export *exp);
444 void ofd_grant_prepare_read(const struct lu_env *env, struct obd_export *exp,
446 void ofd_grant_prepare_write(const struct lu_env *env, struct obd_export *exp,
447 struct obdo *oa, struct niobuf_remote *rnb,
449 void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp, int rc);
450 int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr);
453 int ofd_fmd_init(void);
454 void ofd_fmd_exit(void);
455 struct ofd_mod_data *ofd_fmd_find(struct obd_export *exp,
457 struct ofd_mod_data *ofd_fmd_get(struct obd_export *exp,
459 void ofd_fmd_put(struct obd_export *exp, struct ofd_mod_data *fmd);
460 void ofd_fmd_expire(struct obd_export *exp);
461 void ofd_fmd_cleanup(struct obd_export *exp);
463 void ofd_fmd_drop(struct obd_export *exp, struct lu_fid *fid);
465 #define ofd_fmd_drop(exp, fid) do {} while (0)
469 extern struct ldlm_valblock_ops ofd_lvbo;
472 int ofd_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp,
473 void *req_cookie, ldlm_mode_t mode, __u64 flags,
476 static inline struct ofd_thread_info * ofd_info(const struct lu_env *env)
478 struct ofd_thread_info *info;
480 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
482 LASSERT(info->fti_env);
483 LASSERT(info->fti_env == env);
487 static inline struct ofd_thread_info * ofd_info_init(const struct lu_env *env,
488 struct obd_export *exp)
490 struct ofd_thread_info *info;
492 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
494 LASSERT(info->fti_exp == NULL);
495 LASSERT(info->fti_env == NULL);
496 LASSERT(info->fti_attr.la_valid == 0);
500 info->fti_pre_version = 0;
501 info->fti_transno = 0;
502 info->fti_has_trans = 0;
506 /* The same as osc_build_res_name() */
507 static inline void ofd_build_resid(const struct lu_fid *fid,
508 struct ldlm_res_id *resname)
510 if (fid_is_idif(fid)) {
511 /* get id/seq like ostid_idif_pack() does */
512 osc_build_res_name(fid_idif_id(fid_seq(fid), fid_oid(fid),
514 FID_SEQ_OST_MDT0, resname);
516 /* In the future, where OSTs have FID sequences allocated. */
517 fid_build_reg_res_name(fid, resname);
521 static inline void ofd_fid_from_resid(struct lu_fid *fid,
522 const struct ldlm_res_id *name)
524 /* if seq is FID_SEQ_OST_MDT0 then we have IDIF and resid was built
525 * using osc_build_res_name function. */
526 if (fid_seq_is_mdt0(name->name[LUSTRE_RES_ID_VER_OID_OFF])) {
529 ostid.oi_id = name->name[LUSTRE_RES_ID_SEQ_OFF];
530 ostid.oi_seq = name->name[LUSTRE_RES_ID_VER_OID_OFF];
531 fid_ostid_unpack(fid, &ostid, 0);
533 fid->f_seq = name->name[LUSTRE_RES_ID_SEQ_OFF];
534 fid->f_oid = (__u32)name->name[LUSTRE_RES_ID_VER_OID_OFF];
535 fid->f_ver = name->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32;
539 static inline void ofd_oti2info(struct ofd_thread_info *info,
540 struct obd_trans_info *oti)
542 info->fti_xid = oti->oti_xid;
543 info->fti_transno = oti->oti_transno;
544 info->fti_pre_version = oti->oti_pre_version;
547 static inline void ofd_info2oti(struct ofd_thread_info *info,
548 struct obd_trans_info *oti)
550 oti->oti_xid = info->fti_xid;
551 LASSERTF(ergo(oti->oti_transno > 0,
552 oti->oti_transno == info->fti_transno),
553 "Overwrite replay transno "LPX64" by "LPX64"\n",
554 oti->oti_transno, info->fti_transno);
555 oti->oti_transno = info->fti_transno;
556 oti->oti_pre_version = info->fti_pre_version;
559 /* sync on lock cancel is useless when we force a journal flush,
560 * and if we enable async journal commit, we should also turn on
561 * sync on lock cancel if it is not enabled already. */
562 static inline void ofd_slc_set(struct ofd_device *ofd)
564 if (ofd->ofd_syncjournal == 1)
565 ofd->ofd_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
566 else if (ofd->ofd_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
567 ofd->ofd_sync_lock_cancel = ALWAYS_SYNC_ON_CANCEL;
570 static inline void ofd_prepare_fidea(struct filter_fid *ff, struct obdo *oa)
572 if (!(oa->o_valid & OBD_MD_FLGROUP))
574 /* packing fid and converting it to LE for storing into EA.
575 * Here ->o_stripe_idx should be filled by LOV and rest of
576 * fields - by client. */
577 ff->ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
578 ff->ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
579 /* XXX: we are ignoring o_parent_ver here, since this should
580 * be the same for all objects in this fileset. */
581 ff->ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
582 ff->ff_objid = cpu_to_le64(oa->o_id);
583 ff->ff_seq = cpu_to_le64(oa->o_seq);
586 /* niobuf_remote has no rnb_ prefix in master */
587 #define rnb_offset offset
588 #define rnb_flags flags
590 /* the same for niobuf_local */
591 #define lnb_flags flags
594 #endif /* _OFD_INTERNAL_H */