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
53 /* Limit the returned fields marked valid to those that we actually might set */
54 #define OFD_VALID_FLAGS (LA_TYPE | LA_MODE | LA_SIZE | LA_BLOCKS | \
55 LA_BLKSIZE | LA_ATIME | LA_MTIME | LA_CTIME)
57 /* per-client-per-object persistent state (LRU) */
59 cfs_list_t fmd_list; /* linked to fed_mod_list */
60 struct lu_fid fmd_fid; /* FID being written to */
61 __u64 fmd_mactime_xid; /* xid highest {m,a,c}time setattr */
62 cfs_time_t fmd_expire; /* time when the fmd should expire */
63 int fmd_refcount; /* reference counter - list holds 1 */
66 #define OFD_FMD_MAX_NUM_DEFAULT 128
67 #define OFD_FMD_MAX_AGE_DEFAULT ((obd_timeout + 10) * CFS_HZ)
70 LPROC_OFD_READ_BYTES = 0,
71 LPROC_OFD_WRITE_BYTES = 1,
76 struct dt_device ofd_dt_dev;
77 struct dt_device *ofd_osd;
78 struct obd_export *ofd_osd_exp;
79 struct dt_device_param ofd_dt_conf;
80 /* DLM name-space for meta-data locks maintained by this server */
81 struct ldlm_namespace *ofd_namespace;
83 /* transaction callbacks */
84 struct dt_txn_callback ofd_txn_cb;
87 struct lu_target ofd_lut;
88 struct dt_object *ofd_last_group_file;
89 struct dt_object *ofd_health_check_file;
94 obd_id ofd_last_objids[OFD_MAX_GROUPS];
95 cfs_mutex_t ofd_create_locks[OFD_MAX_GROUPS];
96 struct dt_object *ofd_lastid_obj[OFD_MAX_GROUPS];
97 cfs_spinlock_t ofd_objid_lock;
98 unsigned long ofd_destroys_in_progress;
100 /* protect all statfs-related counters */
101 cfs_spinlock_t ofd_osfs_lock;
102 /* statfs optimization: we cache a bit */
103 struct obd_statfs ofd_osfs;
106 /* writes which might be be accounted twice in ofd_osfs.os_bavail */
107 obd_size ofd_osfs_unstable;
109 /* counters used during statfs update, protected by ofd_osfs_lock.
110 * record when some statfs refresh are in progress */
111 int ofd_statfs_inflight;
112 /* track writes completed while statfs refresh is underway.
113 * tracking is only effective when ofd_statfs_inflight > 1 */
114 obd_size ofd_osfs_inflight;
116 /* grants: all values in bytes */
117 /* grant lock to protect all grant counters */
118 cfs_spinlock_t ofd_grant_lock;
119 /* total amount of dirty data reported by clients in incoming obdo */
120 obd_size ofd_tot_dirty;
121 /* sum of filesystem space granted to clients for async writes */
122 obd_size ofd_tot_granted;
123 /* grant used by I/Os in progress (between prepare and commit) */
124 obd_size ofd_tot_pending;
125 /* free space threshold over which we stop granting space to clients
126 * ofd_grant_ratio is stored as a fixed-point fraction using
127 * OFD_GRANT_RATIO_SHIFT of the remaining free space, not in percentage
130 /* number of clients using grants */
131 int ofd_tot_granted_clients;
133 /* ofd mod data: ofd_device wide values */
134 int ofd_fmd_max_num; /* per ofd ofd_mod_data */
135 cfs_duration_t ofd_fmd_max_age; /* time to fmd expiry */
137 cfs_spinlock_t ofd_flags_lock;
138 unsigned long ofd_raid_degraded:1,
139 /* sync journal on writes */
141 /* sync on lock cancel */
142 ofd_sync_lock_cancel:2,
143 /* shall we grant space to clients not
144 * supporting OBD_CONNECT_GRANT_PARAM? */
145 ofd_grant_compat_disable:1;
147 struct lu_site ofd_site;
150 static inline struct ofd_device *ofd_dev(struct lu_device *d)
152 return container_of0(d, struct ofd_device, ofd_dt_dev.dd_lu_dev);
155 static inline struct obd_device *ofd_obd(struct ofd_device *ofd)
157 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd;
160 static inline struct ofd_device *ofd_exp(struct obd_export *exp)
162 return ofd_dev(exp->exp_obd->obd_lu_dev);
165 static inline char *ofd_name(struct ofd_device *ofd)
167 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd->obd_name;
171 struct lu_object_header ofo_header;
172 struct dt_object ofo_obj;
176 static inline struct ofd_object *ofd_obj(struct lu_object *o)
178 return container_of0(o, struct ofd_object, ofo_obj.do_lu);
181 static inline int ofd_object_exists(struct ofd_object *obj)
183 LASSERT(obj != NULL);
184 if (lu_object_is_dying(obj->ofo_obj.do_lu.lo_header))
186 return lu_object_exists(&obj->ofo_obj.do_lu);
189 static inline struct dt_object *fo2dt(struct ofd_object *obj)
191 return &obj->ofo_obj;
194 static inline struct dt_object *ofd_object_child(struct ofd_object *_obj)
196 struct lu_object *lu = &(_obj)->ofo_obj.do_lu;
198 return container_of0(lu_object_next(lu), struct dt_object, do_lu);
201 static inline struct ofd_device *ofd_obj2dev(const struct ofd_object *fo)
203 return ofd_dev(fo->ofo_obj.do_lu.lo_dev);
206 static inline struct lustre_capa *ofd_object_capa(const struct lu_env *env,
207 const struct ofd_object *obj)
209 /* TODO: see mdd_object_capa() */
213 static inline void ofd_read_lock(const struct lu_env *env,
214 struct ofd_object *fo)
216 struct dt_object *next = ofd_object_child(fo);
218 next->do_ops->do_read_lock(env, next, 0);
221 static inline void ofd_read_unlock(const struct lu_env *env,
222 struct ofd_object *fo)
224 struct dt_object *next = ofd_object_child(fo);
226 next->do_ops->do_read_unlock(env, next);
229 static inline void ofd_write_lock(const struct lu_env *env,
230 struct ofd_object *fo)
232 struct dt_object *next = ofd_object_child(fo);
234 next->do_ops->do_write_lock(env, next, 0);
237 static inline void ofd_write_unlock(const struct lu_env *env,
238 struct ofd_object *fo)
240 struct dt_object *next = ofd_object_child(fo);
242 next->do_ops->do_write_unlock(env, next);
246 * Common data shared by obdofd-level handlers. This is allocated per-thread
247 * to reduce stack consumption.
249 struct ofd_thread_info {
250 const struct lu_env *fti_env;
252 struct obd_export *fti_exp;
255 __u64 fti_pre_version;
256 __u32 fti_has_trans:1, /* has txn already */
259 struct lu_fid fti_fid;
260 struct lu_attr fti_attr;
261 struct lu_attr fti_attr2;
262 struct ldlm_res_id fti_resid;
263 struct filter_fid fti_mds_fid;
264 struct filter_fid fti_mds_fid2;
265 struct ost_id fti_ostid;
266 struct ofd_object *fti_obj;
268 char name[64]; /* for ofd_init0() */
269 struct obd_statfs osfs; /* for obdofd_statfs() */
272 /* Ops object filename */
273 struct lu_name fti_name;
274 struct dt_object_format fti_dof;
275 struct lu_buf fti_buf;
278 /* Space used by the I/O, used by grant code */
279 unsigned long fti_used;
282 extern void target_recovery_fini(struct obd_device *obd);
283 extern void target_recovery_init(struct lu_target *lut, svc_handler_t handler);
286 int ofd_update_capa_key(struct ofd_device *ofd, struct lustre_capa_key *key);
287 int ofd_auth_capa(struct obd_export *exp, struct lu_fid *fid, obd_seq seq,
288 struct lustre_capa *capa, __u64 opc);
289 void ofd_free_capa_keys(struct ofd_device *ofd);
292 extern struct lu_context_key ofd_thread_key;
295 extern struct obd_ops ofd_obd_ops;
296 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
297 struct obd_statfs *osfs, __u64 max_age,
301 obd_id ofd_last_id(struct ofd_device *ofd, obd_seq seq);
302 void ofd_last_id_set(struct ofd_device *ofd, obd_id id, obd_seq seq);
303 int ofd_last_id_write(const struct lu_env *env, struct ofd_device *ofd,
305 int ofd_group_load(const struct lu_env *env, struct ofd_device *ofd, int);
306 int ofd_fs_setup(const struct lu_env *env, struct ofd_device *ofd,
307 struct obd_device *obd);
308 void ofd_fs_cleanup(const struct lu_env *env, struct ofd_device *ofd);
311 int ofd_preprw(const struct lu_env *env,int cmd, struct obd_export *exp,
312 struct obdo *oa, int objcount, struct obd_ioobj *obj,
313 struct niobuf_remote *rnb, int *nr_local,
314 struct niobuf_local *lnb, struct obd_trans_info *oti,
315 struct lustre_capa *capa);
316 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
317 struct obdo *oa, int objcount, struct obd_ioobj *obj,
318 struct niobuf_remote *rnb, int npages,
319 struct niobuf_local *lnb, struct obd_trans_info *oti,
323 struct thandle *ofd_trans_create(const struct lu_env *env,
324 struct ofd_device *ofd);
325 int ofd_trans_start(const struct lu_env *env,
326 struct ofd_device *ofd, struct ofd_object *fo,
328 void ofd_trans_stop(const struct lu_env *env, struct ofd_device *ofd,
329 struct thandle *th, int rc);
330 int ofd_txn_stop_cb(const struct lu_env *env, struct thandle *txn,
334 void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars);
335 int lproc_ofd_attach_seqstat(struct obd_device *dev);
336 extern struct file_operations ofd_per_nid_stats_fops;
339 struct ofd_object *ofd_object_find(const struct lu_env *env,
340 struct ofd_device *ofd,
341 const struct lu_fid *fid);
342 struct ofd_object *ofd_object_find_or_create(const struct lu_env *env,
343 struct ofd_device *ofd,
344 const struct lu_fid *fid,
345 struct lu_attr *attr);
346 int ofd_object_ff_check(const struct lu_env *env, struct ofd_object *fo);
347 int ofd_precreate_object(const struct lu_env *env, struct ofd_device *ofd,
348 obd_id id, obd_seq seq);
350 void ofd_object_put(const struct lu_env *env, struct ofd_object *fo);
351 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
352 struct lu_attr *la, struct filter_fid *ff);
353 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
354 __u64 start, __u64 end, struct lu_attr *la,
355 struct filter_fid *ff);
356 int ofd_object_destroy(const struct lu_env *, struct ofd_object *, int);
357 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
359 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
360 struct lu_attr *la, int is_setattr);
363 #define OFD_GRANT_RATIO_SHIFT 8
364 static inline __u64 ofd_grant_reserved(struct ofd_device *ofd, obd_size bavail)
366 return (bavail * ofd->ofd_grant_ratio) >> OFD_GRANT_RATIO_SHIFT;
369 static inline int ofd_grant_ratio_conv(int percentage)
371 return (percentage << OFD_GRANT_RATIO_SHIFT) / 100;
374 static inline int ofd_grant_param_supp(struct obd_export *exp)
376 return !!(exp->exp_connect_flags & OBD_CONNECT_GRANT_PARAM);
379 /* Blocksize used for client not supporting OBD_CONNECT_GRANT_PARAM.
380 * That's 4KB=2^12 which is the biggest block size known to work whatever
381 * the client's page size is. */
382 #define COMPAT_BSIZE_SHIFT 12
383 static inline int ofd_grant_compat(struct obd_export *exp,
384 struct ofd_device *ofd)
386 /* Clients which don't support OBD_CONNECT_GRANT_PARAM cannot handle
387 * a block size > page size and consume CFS_PAGE_SIZE of grant when
388 * dirtying a page regardless of the block size */
389 return !!(ofd_obd(ofd)->obd_self_export != exp &&
390 ofd->ofd_blockbits > COMPAT_BSIZE_SHIFT &&
391 !ofd_grant_param_supp(exp));
394 static inline int ofd_grant_prohibit(struct obd_export *exp,
395 struct ofd_device *ofd)
397 /* When ofd_grant_compat_disable is set, we don't grant any space to
398 * clients not supporting OBD_CONNECT_GRANT_PARAM.
399 * Otherwise, space granted to such a client is inflated since it
400 * consumes CFS_PAGE_SIZE of grant space per block */
401 return !!(ofd_grant_compat(exp, ofd) && ofd->ofd_grant_compat_disable);
404 void ofd_grant_sanity_check(struct obd_device *obd, const char *func);
405 long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
407 void ofd_grant_discard(struct obd_export *exp);
408 void ofd_grant_prepare_read(const struct lu_env *env, struct obd_export *exp,
410 void ofd_grant_prepare_write(const struct lu_env *env, struct obd_export *exp,
411 struct obdo *oa, struct niobuf_remote *rnb,
413 void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp, int rc);
414 int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr);
417 int ofd_fmd_init(void);
418 void ofd_fmd_exit(void);
419 struct ofd_mod_data *ofd_fmd_find(struct obd_export *exp,
421 struct ofd_mod_data *ofd_fmd_get(struct obd_export *exp,
423 void ofd_fmd_put(struct obd_export *exp, struct ofd_mod_data *fmd);
424 void ofd_fmd_expire(struct obd_export *exp);
425 void ofd_fmd_cleanup(struct obd_export *exp);
427 void ofd_fmd_drop(struct obd_export *exp, struct lu_fid *fid);
429 #define ofd_fmd_drop(exp, fid) do {} while (0)
433 extern struct ldlm_valblock_ops ofd_lvbo;
436 int ofd_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp,
437 void *req_cookie, ldlm_mode_t mode, int flags,
440 static inline struct ofd_thread_info * ofd_info(const struct lu_env *env)
442 struct ofd_thread_info *info;
444 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
446 LASSERT(info->fti_env);
447 LASSERT(info->fti_env == env);
451 static inline struct ofd_thread_info * ofd_info_init(const struct lu_env *env,
452 struct obd_export *exp)
454 struct ofd_thread_info *info;
456 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
458 LASSERT(info->fti_exp == NULL);
459 LASSERT(info->fti_env == NULL);
460 LASSERT(info->fti_attr.la_valid == 0);
464 info->fti_pre_version = 0;
465 info->fti_transno = 0;
466 info->fti_has_trans = 0;
470 /* The same as osc_build_res_name() */
471 static inline void ofd_build_resid(const struct lu_fid *fid,
472 struct ldlm_res_id *resname)
474 if (fid_is_idif(fid)) {
475 /* get id/seq like ostid_idif_pack() does */
476 osc_build_res_name(fid_idif_id(fid_seq(fid), fid_oid(fid),
478 FID_SEQ_OST_MDT0, resname);
480 /* In the future, where OSTs have FID sequences allocated. */
481 fid_build_reg_res_name(fid, resname);
485 static inline void ofd_fid_from_resid(struct lu_fid *fid,
486 const struct ldlm_res_id *name)
488 /* if seq is FID_SEQ_OST_MDT0 then we have IDIF and resid was built
489 * using osc_build_res_name function. */
490 if (fid_seq_is_mdt0(name->name[LUSTRE_RES_ID_VER_OID_OFF])) {
493 ostid.oi_id = name->name[LUSTRE_RES_ID_SEQ_OFF];
494 ostid.oi_seq = name->name[LUSTRE_RES_ID_VER_OID_OFF];
495 fid_ostid_unpack(fid, &ostid, 0);
497 fid->f_seq = name->name[LUSTRE_RES_ID_SEQ_OFF];
498 fid->f_oid = (__u32)name->name[LUSTRE_RES_ID_VER_OID_OFF];
499 fid->f_ver = name->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32;
503 static inline void ofd_oti2info(struct ofd_thread_info *info,
504 struct obd_trans_info *oti)
506 info->fti_xid = oti->oti_xid;
507 info->fti_transno = oti->oti_transno;
508 info->fti_pre_version = oti->oti_pre_version;
511 static inline void ofd_info2oti(struct ofd_thread_info *info,
512 struct obd_trans_info *oti)
514 oti->oti_xid = info->fti_xid;
515 LASSERTF(ergo(oti->oti_transno > 0,
516 oti->oti_transno == info->fti_transno),
517 "Overwrite replay transno "LPX64" by "LPX64"\n",
518 oti->oti_transno, info->fti_transno);
519 oti->oti_transno = info->fti_transno;
520 oti->oti_pre_version = info->fti_pre_version;
523 /* sync on lock cancel is useless when we force a journal flush,
524 * and if we enable async journal commit, we should also turn on
525 * sync on lock cancel if it is not enabled already. */
526 static inline void ofd_slc_set(struct ofd_device *ofd)
528 if (ofd->ofd_syncjournal == 1)
529 ofd->ofd_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
530 else if (ofd->ofd_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
531 ofd->ofd_sync_lock_cancel = ALWAYS_SYNC_ON_CANCEL;
534 static inline void ofd_prepare_fidea(struct filter_fid *ff, struct obdo *oa)
536 if (!(oa->o_valid & OBD_MD_FLGROUP))
538 /* packing fid and converting it to LE for storing into EA.
539 * Here ->o_stripe_idx should be filled by LOV and rest of
540 * fields - by client. */
541 ff->ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
542 ff->ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
543 /* XXX: we are ignoring o_parent_ver here, since this should
544 * be the same for all objects in this fileset. */
545 ff->ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
546 ff->ff_objid = cpu_to_le64(oa->o_id);
547 ff->ff_seq = cpu_to_le64(oa->o_seq);
550 /* niobuf_remote has no rnb_ prefix in master */
551 #define rnb_offset offset
552 #define rnb_flags flags
554 /* the same for niobuf_local */
555 #define lnb_flags flags
558 #endif /* _OFD_INTERNAL_H */