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) 2011, 2012, Whamcloud, Inc.
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 /* 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 cfs_list_t 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 ((obd_timeout + 10) * CFS_HZ)
71 LPROC_OFD_READ_BYTES = 0,
72 LPROC_OFD_WRITE_BYTES = 1,
78 LPROC_OFD_STATS_READ = 0,
79 LPROC_OFD_STATS_WRITE = 1,
80 LPROC_OFD_STATS_SETATTR = 2,
81 LPROC_OFD_STATS_PUNCH = 3,
82 LPROC_OFD_STATS_SYNC = 4,
86 static inline void ofd_counter_incr(struct obd_export *exp, int opcode,
87 char *jobid, long amount)
89 if (exp->exp_obd && exp->exp_obd->u.obt.obt_jobstats.ojs_hash &&
90 (exp->exp_connect_flags & OBD_CONNECT_JOBSTATS))
91 lprocfs_job_stats_log(exp->exp_obd, jobid, opcode, amount);
95 struct dt_device ofd_dt_dev;
96 struct dt_device *ofd_osd;
97 struct obd_export *ofd_osd_exp;
98 struct dt_device_param ofd_dt_conf;
99 /* DLM name-space for meta-data locks maintained by this server */
100 struct ldlm_namespace *ofd_namespace;
102 /* transaction callbacks */
103 struct dt_txn_callback ofd_txn_cb;
106 struct lu_target ofd_lut;
107 struct dt_object *ofd_last_group_file;
108 struct dt_object *ofd_health_check_file;
110 int ofd_subdir_count;
113 obd_id ofd_last_objids[OFD_MAX_GROUPS];
114 struct mutex ofd_create_locks[OFD_MAX_GROUPS];
115 struct dt_object *ofd_lastid_obj[OFD_MAX_GROUPS];
116 spinlock_t ofd_objid_lock;
117 unsigned long ofd_destroys_in_progress;
118 int ofd_precreate_batch;
120 /* protect all statfs-related counters */
121 spinlock_t ofd_osfs_lock;
122 /* statfs optimization: we cache a bit */
123 struct obd_statfs ofd_osfs;
126 /* writes which might be be accounted twice in ofd_osfs.os_bavail */
127 obd_size ofd_osfs_unstable;
129 /* counters used during statfs update, protected by ofd_osfs_lock.
130 * record when some statfs refresh are in progress */
131 int ofd_statfs_inflight;
132 /* track writes completed while statfs refresh is underway.
133 * tracking is only effective when ofd_statfs_inflight > 1 */
134 obd_size ofd_osfs_inflight;
136 /* grants: all values in bytes */
137 /* grant lock to protect all grant counters */
138 spinlock_t ofd_grant_lock;
139 /* total amount of dirty data reported by clients in incoming obdo */
140 obd_size ofd_tot_dirty;
141 /* sum of filesystem space granted to clients for async writes */
142 obd_size ofd_tot_granted;
143 /* grant used by I/Os in progress (between prepare and commit) */
144 obd_size ofd_tot_pending;
145 /* free space threshold over which we stop granting space to clients
146 * ofd_grant_ratio is stored as a fixed-point fraction using
147 * OFD_GRANT_RATIO_SHIFT of the remaining free space, not in percentage
150 /* number of clients using grants */
151 int ofd_tot_granted_clients;
153 /* ofd mod data: ofd_device wide values */
154 int ofd_fmd_max_num; /* per ofd ofd_mod_data */
155 cfs_duration_t ofd_fmd_max_age; /* time to fmd expiry */
157 spinlock_t ofd_flags_lock;
158 unsigned long ofd_raid_degraded:1,
159 /* sync journal on writes */
161 /* sync on lock cancel */
162 ofd_sync_lock_cancel:2,
163 /* shall we grant space to clients not
164 * supporting OBD_CONNECT_GRANT_PARAM? */
165 ofd_grant_compat_disable:1;
168 static inline struct ofd_device *ofd_dev(struct lu_device *d)
170 return container_of0(d, struct ofd_device, ofd_dt_dev.dd_lu_dev);
173 static inline struct obd_device *ofd_obd(struct ofd_device *ofd)
175 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd;
178 static inline struct ofd_device *ofd_exp(struct obd_export *exp)
180 return ofd_dev(exp->exp_obd->obd_lu_dev);
183 static inline char *ofd_name(struct ofd_device *ofd)
185 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd->obd_name;
189 struct lu_object_header ofo_header;
190 struct dt_object ofo_obj;
194 static inline struct ofd_object *ofd_obj(struct lu_object *o)
196 return container_of0(o, struct ofd_object, ofo_obj.do_lu);
199 static inline int ofd_object_exists(struct ofd_object *obj)
201 LASSERT(obj != NULL);
202 if (lu_object_is_dying(obj->ofo_obj.do_lu.lo_header))
204 return lu_object_exists(&obj->ofo_obj.do_lu);
207 static inline struct dt_object *fo2dt(struct ofd_object *obj)
209 return &obj->ofo_obj;
212 static inline struct dt_object *ofd_object_child(struct ofd_object *_obj)
214 struct lu_object *lu = &(_obj)->ofo_obj.do_lu;
216 return container_of0(lu_object_next(lu), struct dt_object, do_lu);
219 static inline struct ofd_device *ofd_obj2dev(const struct ofd_object *fo)
221 return ofd_dev(fo->ofo_obj.do_lu.lo_dev);
224 static inline struct lustre_capa *ofd_object_capa(const struct lu_env *env,
225 const struct ofd_object *obj)
227 /* TODO: see mdd_object_capa() */
231 static inline void ofd_read_lock(const struct lu_env *env,
232 struct ofd_object *fo)
234 struct dt_object *next = ofd_object_child(fo);
236 next->do_ops->do_read_lock(env, next, 0);
239 static inline void ofd_read_unlock(const struct lu_env *env,
240 struct ofd_object *fo)
242 struct dt_object *next = ofd_object_child(fo);
244 next->do_ops->do_read_unlock(env, next);
247 static inline void ofd_write_lock(const struct lu_env *env,
248 struct ofd_object *fo)
250 struct dt_object *next = ofd_object_child(fo);
252 next->do_ops->do_write_lock(env, next, 0);
255 static inline void ofd_write_unlock(const struct lu_env *env,
256 struct ofd_object *fo)
258 struct dt_object *next = ofd_object_child(fo);
260 next->do_ops->do_write_unlock(env, next);
264 * Common data shared by obdofd-level handlers. This is allocated per-thread
265 * to reduce stack consumption.
267 struct ofd_thread_info {
268 const struct lu_env *fti_env;
270 struct obd_export *fti_exp;
273 __u64 fti_pre_version;
274 __u32 fti_has_trans:1, /* has txn already */
277 struct lu_fid fti_fid;
278 struct lu_attr fti_attr;
279 struct lu_attr fti_attr2;
280 struct ldlm_res_id fti_resid;
281 struct filter_fid fti_mds_fid;
282 struct filter_fid fti_mds_fid2;
283 struct ost_id fti_ostid;
284 struct ofd_object *fti_obj;
286 char name[64]; /* for ofd_init0() */
287 struct obd_statfs osfs; /* for obdofd_statfs() */
290 /* Ops object filename */
291 struct lu_name fti_name;
292 struct dt_object_format fti_dof;
293 struct lu_buf fti_buf;
296 /* Space used by the I/O, used by grant code */
297 unsigned long fti_used;
300 extern void target_recovery_fini(struct obd_device *obd);
301 extern void target_recovery_init(struct lu_target *lut, svc_handler_t handler);
304 int ofd_update_capa_key(struct ofd_device *ofd, struct lustre_capa_key *key);
305 int ofd_auth_capa(struct obd_export *exp, struct lu_fid *fid, obd_seq seq,
306 struct lustre_capa *capa, __u64 opc);
307 void ofd_free_capa_keys(struct ofd_device *ofd);
310 extern struct lu_context_key ofd_thread_key;
311 int ofd_postrecov(const struct lu_env *env, struct ofd_device *ofd);
314 extern struct obd_ops ofd_obd_ops;
315 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
316 struct obd_statfs *osfs, __u64 max_age,
320 obd_id ofd_last_id(struct ofd_device *ofd, obd_seq seq);
321 void ofd_last_id_set(struct ofd_device *ofd, obd_id id, obd_seq seq);
322 int ofd_last_id_write(const struct lu_env *env, struct ofd_device *ofd,
324 int ofd_group_load(const struct lu_env *env, struct ofd_device *ofd, int);
325 int ofd_fs_setup(const struct lu_env *env, struct ofd_device *ofd,
326 struct obd_device *obd);
327 void ofd_fs_cleanup(const struct lu_env *env, struct ofd_device *ofd);
328 int ofd_precreate_batch(struct ofd_device *ofd, int batch);
331 int ofd_preprw(const struct lu_env *env,int cmd, struct obd_export *exp,
332 struct obdo *oa, int objcount, struct obd_ioobj *obj,
333 struct niobuf_remote *rnb, int *nr_local,
334 struct niobuf_local *lnb, struct obd_trans_info *oti,
335 struct lustre_capa *capa);
336 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
337 struct obdo *oa, int objcount, struct obd_ioobj *obj,
338 struct niobuf_remote *rnb, int npages,
339 struct niobuf_local *lnb, struct obd_trans_info *oti,
343 struct thandle *ofd_trans_create(const struct lu_env *env,
344 struct ofd_device *ofd);
345 int ofd_trans_start(const struct lu_env *env,
346 struct ofd_device *ofd, struct ofd_object *fo,
348 void ofd_trans_stop(const struct lu_env *env, struct ofd_device *ofd,
349 struct thandle *th, int rc);
350 int ofd_txn_stop_cb(const struct lu_env *env, struct thandle *txn,
355 void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars);
356 int lproc_ofd_attach_seqstat(struct obd_device *dev);
357 extern struct file_operations ofd_per_nid_stats_fops;
358 void ofd_stats_counter_init(struct lprocfs_stats *stats);
360 static void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars)
362 memset(lvars, 0, sizeof(*lvars));
364 static inline int lproc_ofd_attach_seqstat(struct obd_device *dev) {}
365 static inline void ofd_stats_counter_init(struct lprocfs_stats *stats) {}
369 struct ofd_object *ofd_object_find(const struct lu_env *env,
370 struct ofd_device *ofd,
371 const struct lu_fid *fid);
372 struct ofd_object *ofd_object_find_or_create(const struct lu_env *env,
373 struct ofd_device *ofd,
374 const struct lu_fid *fid,
375 struct lu_attr *attr);
376 int ofd_object_ff_check(const struct lu_env *env, struct ofd_object *fo);
377 int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd,
378 obd_id id, obd_seq group, int nr);
380 void ofd_object_put(const struct lu_env *env, struct ofd_object *fo);
381 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
382 struct lu_attr *la, struct filter_fid *ff);
383 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
384 __u64 start, __u64 end, struct lu_attr *la,
385 struct filter_fid *ff);
386 int ofd_object_destroy(const struct lu_env *, struct ofd_object *, int);
387 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
389 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
390 struct lu_attr *la, int is_setattr);
393 #define OFD_GRANT_RATIO_SHIFT 8
394 static inline __u64 ofd_grant_reserved(struct ofd_device *ofd, obd_size bavail)
396 return (bavail * ofd->ofd_grant_ratio) >> OFD_GRANT_RATIO_SHIFT;
399 static inline int ofd_grant_ratio_conv(int percentage)
401 return (percentage << OFD_GRANT_RATIO_SHIFT) / 100;
404 static inline int ofd_grant_param_supp(struct obd_export *exp)
406 return !!(exp->exp_connect_flags & OBD_CONNECT_GRANT_PARAM);
409 /* Blocksize used for client not supporting OBD_CONNECT_GRANT_PARAM.
410 * That's 4KB=2^12 which is the biggest block size known to work whatever
411 * the client's page size is. */
412 #define COMPAT_BSIZE_SHIFT 12
413 static inline int ofd_grant_compat(struct obd_export *exp,
414 struct ofd_device *ofd)
416 /* Clients which don't support OBD_CONNECT_GRANT_PARAM cannot handle
417 * a block size > page size and consume CFS_PAGE_SIZE of grant when
418 * dirtying a page regardless of the block size */
419 return !!(ofd_obd(ofd)->obd_self_export != exp &&
420 ofd->ofd_blockbits > COMPAT_BSIZE_SHIFT &&
421 !ofd_grant_param_supp(exp));
424 static inline int ofd_grant_prohibit(struct obd_export *exp,
425 struct ofd_device *ofd)
427 /* When ofd_grant_compat_disable is set, we don't grant any space to
428 * clients not supporting OBD_CONNECT_GRANT_PARAM.
429 * Otherwise, space granted to such a client is inflated since it
430 * consumes CFS_PAGE_SIZE of grant space per block */
431 return !!(ofd_grant_compat(exp, ofd) && ofd->ofd_grant_compat_disable);
434 void ofd_grant_sanity_check(struct obd_device *obd, const char *func);
435 long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
437 void ofd_grant_discard(struct obd_export *exp);
438 void ofd_grant_prepare_read(const struct lu_env *env, struct obd_export *exp,
440 void ofd_grant_prepare_write(const struct lu_env *env, struct obd_export *exp,
441 struct obdo *oa, struct niobuf_remote *rnb,
443 void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp, int rc);
444 int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr);
447 int ofd_fmd_init(void);
448 void ofd_fmd_exit(void);
449 struct ofd_mod_data *ofd_fmd_find(struct obd_export *exp,
451 struct ofd_mod_data *ofd_fmd_get(struct obd_export *exp,
453 void ofd_fmd_put(struct obd_export *exp, struct ofd_mod_data *fmd);
454 void ofd_fmd_expire(struct obd_export *exp);
455 void ofd_fmd_cleanup(struct obd_export *exp);
457 void ofd_fmd_drop(struct obd_export *exp, struct lu_fid *fid);
459 #define ofd_fmd_drop(exp, fid) do {} while (0)
463 extern struct ldlm_valblock_ops ofd_lvbo;
466 int ofd_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp,
467 void *req_cookie, ldlm_mode_t mode, __u64 flags,
470 static inline struct ofd_thread_info * ofd_info(const struct lu_env *env)
472 struct ofd_thread_info *info;
474 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
476 LASSERT(info->fti_env);
477 LASSERT(info->fti_env == env);
481 static inline struct ofd_thread_info * ofd_info_init(const struct lu_env *env,
482 struct obd_export *exp)
484 struct ofd_thread_info *info;
486 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
488 LASSERT(info->fti_exp == NULL);
489 LASSERT(info->fti_env == NULL);
490 LASSERT(info->fti_attr.la_valid == 0);
494 info->fti_pre_version = 0;
495 info->fti_transno = 0;
496 info->fti_has_trans = 0;
500 /* The same as osc_build_res_name() */
501 static inline void ofd_build_resid(const struct lu_fid *fid,
502 struct ldlm_res_id *resname)
504 if (fid_is_idif(fid)) {
505 /* get id/seq like ostid_idif_pack() does */
506 osc_build_res_name(fid_idif_id(fid_seq(fid), fid_oid(fid),
508 FID_SEQ_OST_MDT0, resname);
510 /* In the future, where OSTs have FID sequences allocated. */
511 fid_build_reg_res_name(fid, resname);
515 static inline void ofd_fid_from_resid(struct lu_fid *fid,
516 const struct ldlm_res_id *name)
518 /* if seq is FID_SEQ_OST_MDT0 then we have IDIF and resid was built
519 * using osc_build_res_name function. */
520 if (fid_seq_is_mdt0(name->name[LUSTRE_RES_ID_VER_OID_OFF])) {
523 ostid.oi_id = name->name[LUSTRE_RES_ID_SEQ_OFF];
524 ostid.oi_seq = name->name[LUSTRE_RES_ID_VER_OID_OFF];
525 fid_ostid_unpack(fid, &ostid, 0);
527 fid->f_seq = name->name[LUSTRE_RES_ID_SEQ_OFF];
528 fid->f_oid = (__u32)name->name[LUSTRE_RES_ID_VER_OID_OFF];
529 fid->f_ver = name->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32;
533 static inline void ofd_oti2info(struct ofd_thread_info *info,
534 struct obd_trans_info *oti)
536 info->fti_xid = oti->oti_xid;
537 info->fti_transno = oti->oti_transno;
538 info->fti_pre_version = oti->oti_pre_version;
541 static inline void ofd_info2oti(struct ofd_thread_info *info,
542 struct obd_trans_info *oti)
544 oti->oti_xid = info->fti_xid;
545 LASSERTF(ergo(oti->oti_transno > 0,
546 oti->oti_transno == info->fti_transno),
547 "Overwrite replay transno "LPX64" by "LPX64"\n",
548 oti->oti_transno, info->fti_transno);
549 oti->oti_transno = info->fti_transno;
550 oti->oti_pre_version = info->fti_pre_version;
553 /* sync on lock cancel is useless when we force a journal flush,
554 * and if we enable async journal commit, we should also turn on
555 * sync on lock cancel if it is not enabled already. */
556 static inline void ofd_slc_set(struct ofd_device *ofd)
558 if (ofd->ofd_syncjournal == 1)
559 ofd->ofd_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
560 else if (ofd->ofd_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
561 ofd->ofd_sync_lock_cancel = ALWAYS_SYNC_ON_CANCEL;
564 static inline void ofd_prepare_fidea(struct filter_fid *ff, struct obdo *oa)
566 if (!(oa->o_valid & OBD_MD_FLGROUP))
568 /* packing fid and converting it to LE for storing into EA.
569 * Here ->o_stripe_idx should be filled by LOV and rest of
570 * fields - by client. */
571 ff->ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
572 ff->ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
573 /* XXX: we are ignoring o_parent_ver here, since this should
574 * be the same for all objects in this fileset. */
575 ff->ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
576 ff->ff_objid = cpu_to_le64(oa->o_id);
577 ff->ff_seq = cpu_to_le64(oa->o_seq);
580 /* niobuf_remote has no rnb_ prefix in master */
581 #define rnb_offset offset
582 #define rnb_flags flags
584 /* the same for niobuf_local */
585 #define lnb_flags flags
588 #endif /* _OFD_INTERNAL_H */