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 dt_device_param ofd_dt_conf;
79 /* DLM name-space for meta-data locks maintained by this server */
80 struct ldlm_namespace *ofd_namespace;
82 /* transaction callbacks */
83 struct dt_txn_callback ofd_txn_cb;
86 struct lu_target ofd_lut;
87 struct dt_object *ofd_last_group_file;
88 struct dt_object *ofd_health_check_file;
93 obd_id ofd_last_objids[OFD_MAX_GROUPS];
94 cfs_mutex_t ofd_create_locks[OFD_MAX_GROUPS];
95 struct dt_object *ofd_lastid_obj[OFD_MAX_GROUPS];
96 cfs_spinlock_t ofd_objid_lock;
97 unsigned long ofd_destroys_in_progress;
99 /* protect all statfs-related counters */
100 cfs_spinlock_t ofd_osfs_lock;
101 /* statfs optimization: we cache a bit */
102 struct obd_statfs ofd_osfs;
105 /* writes which might be be accounted twice in ofd_osfs.os_bavail */
106 obd_size ofd_osfs_unstable;
108 /* counters used during statfs update, protected by ofd_osfs_lock.
109 * record when some statfs refresh are in progress */
110 int ofd_statfs_inflight;
111 /* track writes completed while statfs refresh is underway.
112 * tracking is only effective when ofd_statfs_inflight > 1 */
113 obd_size ofd_osfs_inflight;
115 /* grants: all values in bytes */
116 /* grant lock to protect all grant counters */
117 cfs_spinlock_t ofd_grant_lock;
118 /* total amount of dirty data reported by clients in incoming obdo */
119 obd_size ofd_tot_dirty;
120 /* sum of filesystem space granted to clients for async writes */
121 obd_size ofd_tot_granted;
122 /* grant used by I/Os in progress (between prepare and commit) */
123 obd_size ofd_tot_pending;
124 /* free space threshold over which we stop granting space to clients
125 * ofd_grant_ratio is stored as a fixed-point fraction using
126 * OFD_GRANT_RATIO_SHIFT of the remaining free space, not in percentage
129 /* number of clients using grants */
130 int ofd_tot_granted_clients;
132 /* ofd mod data: ofd_device wide values */
133 int ofd_fmd_max_num; /* per ofd ofd_mod_data */
134 cfs_duration_t ofd_fmd_max_age; /* time to fmd expiry */
136 cfs_spinlock_t ofd_flags_lock;
137 unsigned long ofd_raid_degraded:1,
138 /* sync journal on writes */
140 /* sync on lock cancel */
141 ofd_sync_lock_cancel:2,
142 /* shall we grant space to clients not
143 * supporting OBD_CONNECT_GRANT_PARAM? */
144 ofd_grant_compat_disable:1;
146 struct lu_site ofd_site;
149 static inline struct ofd_device *ofd_dev(struct lu_device *d)
151 return container_of0(d, struct ofd_device, ofd_dt_dev.dd_lu_dev);
154 static inline struct obd_device *ofd_obd(struct ofd_device *ofd)
156 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd;
159 static inline struct ofd_device *ofd_exp(struct obd_export *exp)
161 return ofd_dev(exp->exp_obd->obd_lu_dev);
164 static inline char *ofd_name(struct ofd_device *ofd)
166 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd->obd_name;
170 struct lu_object_header ofo_header;
171 struct dt_object ofo_obj;
175 static inline struct ofd_object *ofd_obj(struct lu_object *o)
177 return container_of0(o, struct ofd_object, ofo_obj.do_lu);
180 static inline int ofd_object_exists(struct ofd_object *obj)
182 LASSERT(obj != NULL);
183 if (lu_object_is_dying(obj->ofo_obj.do_lu.lo_header))
185 return lu_object_exists(&obj->ofo_obj.do_lu);
188 static inline struct dt_object *fo2dt(struct ofd_object *obj)
190 return &obj->ofo_obj;
193 static inline struct dt_object *ofd_object_child(struct ofd_object *_obj)
195 struct lu_object *lu = &(_obj)->ofo_obj.do_lu;
197 return container_of0(lu_object_next(lu), struct dt_object, do_lu);
200 static inline struct ofd_device *ofd_obj2dev(const struct ofd_object *fo)
202 return ofd_dev(fo->ofo_obj.do_lu.lo_dev);
205 static inline struct lustre_capa *ofd_object_capa(const struct lu_env *env,
206 const struct ofd_object *obj)
208 /* TODO: see mdd_object_capa() */
212 static inline void ofd_read_lock(const struct lu_env *env,
213 struct ofd_object *fo)
215 struct dt_object *next = ofd_object_child(fo);
217 next->do_ops->do_read_lock(env, next, 0);
220 static inline void ofd_read_unlock(const struct lu_env *env,
221 struct ofd_object *fo)
223 struct dt_object *next = ofd_object_child(fo);
225 next->do_ops->do_read_unlock(env, next);
228 static inline void ofd_write_lock(const struct lu_env *env,
229 struct ofd_object *fo)
231 struct dt_object *next = ofd_object_child(fo);
233 next->do_ops->do_write_lock(env, next, 0);
236 static inline void ofd_write_unlock(const struct lu_env *env,
237 struct ofd_object *fo)
239 struct dt_object *next = ofd_object_child(fo);
241 next->do_ops->do_write_unlock(env, next);
245 * Common data shared by obdofd-level handlers. This is allocated per-thread
246 * to reduce stack consumption.
248 struct ofd_thread_info {
249 const struct lu_env *fti_env;
251 struct obd_export *fti_exp;
254 __u64 fti_pre_version;
255 __u32 fti_has_trans:1, /* has txn already */
258 struct lu_fid fti_fid;
259 struct lu_attr fti_attr;
260 struct lu_attr fti_attr2;
261 struct ldlm_res_id fti_resid;
262 struct filter_fid fti_mds_fid;
263 struct filter_fid fti_mds_fid2;
264 struct ost_id fti_ostid;
265 struct ofd_object *fti_obj;
267 char name[64]; /* for ofd_init0() */
268 struct obd_statfs osfs; /* for obdofd_statfs() */
271 /* Ops object filename */
272 struct lu_name fti_name;
273 struct dt_object_format fti_dof;
274 struct lu_buf fti_buf;
277 /* Space used by the I/O, used by grant code */
278 unsigned long fti_used;
281 extern void target_recovery_fini(struct obd_device *obd);
282 extern void target_recovery_init(struct lu_target *lut, svc_handler_t handler);
285 int ofd_update_capa_key(struct ofd_device *ofd, struct lustre_capa_key *key);
286 int ofd_auth_capa(struct obd_export *exp, struct lu_fid *fid, obd_seq seq,
287 struct lustre_capa *capa, __u64 opc);
288 void ofd_free_capa_keys(struct ofd_device *ofd);
291 extern struct lu_context_key ofd_thread_key;
294 extern struct obd_ops ofd_obd_ops;
295 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
296 struct obd_statfs *osfs, __u64 max_age,
300 obd_id ofd_last_id(struct ofd_device *ofd, obd_seq seq);
301 void ofd_last_id_set(struct ofd_device *ofd, obd_id id, obd_seq seq);
302 int ofd_last_id_write(const struct lu_env *env, struct ofd_device *ofd,
304 int ofd_group_load(const struct lu_env *env, struct ofd_device *ofd, int);
305 int ofd_fs_setup(const struct lu_env *env, struct ofd_device *ofd,
306 struct obd_device *obd);
307 void ofd_fs_cleanup(const struct lu_env *env, struct ofd_device *ofd);
310 int ofd_preprw(const struct lu_env *env,int cmd, struct obd_export *exp,
311 struct obdo *oa, int objcount, struct obd_ioobj *obj,
312 struct niobuf_remote *rnb, int *nr_local,
313 struct niobuf_local *lnb, struct obd_trans_info *oti,
314 struct lustre_capa *capa);
315 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
316 struct obdo *oa, int objcount, struct obd_ioobj *obj,
317 struct niobuf_remote *rnb, int npages,
318 struct niobuf_local *lnb, struct obd_trans_info *oti,
322 struct thandle *ofd_trans_create(const struct lu_env *env,
323 struct ofd_device *ofd);
324 int ofd_trans_start(const struct lu_env *env,
325 struct ofd_device *ofd, struct ofd_object *fo,
327 void ofd_trans_stop(const struct lu_env *env, struct ofd_device *ofd,
328 struct thandle *th, int rc);
329 int ofd_txn_stop_cb(const struct lu_env *env, struct thandle *txn,
333 void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars);
334 int lproc_ofd_attach_seqstat(struct obd_device *dev);
335 extern struct file_operations ofd_per_nid_stats_fops;
338 struct ofd_object *ofd_object_find(const struct lu_env *env,
339 struct ofd_device *ofd,
340 const struct lu_fid *fid);
341 struct ofd_object *ofd_object_find_or_create(const struct lu_env *env,
342 struct ofd_device *ofd,
343 const struct lu_fid *fid,
344 struct lu_attr *attr);
345 int ofd_object_ff_check(const struct lu_env *env, struct ofd_object *fo);
346 int ofd_precreate_object(const struct lu_env *env, struct ofd_device *ofd,
347 obd_id id, obd_seq seq);
349 void ofd_object_put(const struct lu_env *env, struct ofd_object *fo);
350 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
351 struct lu_attr *la, struct filter_fid *ff);
352 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
353 __u64 start, __u64 end, struct lu_attr *la,
354 struct filter_fid *ff);
355 int ofd_object_destroy(const struct lu_env *, struct ofd_object *, int);
356 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
358 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
359 struct lu_attr *la, int is_setattr);
362 #define OFD_GRANT_RATIO_SHIFT 8
363 static inline __u64 ofd_grant_reserved(struct ofd_device *ofd, obd_size bavail)
365 return (bavail * ofd->ofd_grant_ratio) >> OFD_GRANT_RATIO_SHIFT;
368 static inline int ofd_grant_ratio_conv(int percentage)
370 return (percentage << OFD_GRANT_RATIO_SHIFT) / 100;
373 static inline int ofd_grant_param_supp(struct obd_export *exp)
375 return !!(exp->exp_connect_flags & OBD_CONNECT_GRANT_PARAM);
378 /* Blocksize used for client not supporting OBD_CONNECT_GRANT_PARAM.
379 * That's 4KB=2^12 which is the biggest block size known to work whatever
380 * the client's page size is. */
381 #define COMPAT_BSIZE_SHIFT 12
382 static inline int ofd_grant_compat(struct obd_export *exp,
383 struct ofd_device *ofd)
385 /* Clients which don't support OBD_CONNECT_GRANT_PARAM cannot handle
386 * a block size > page size and consume CFS_PAGE_SIZE of grant when
387 * dirtying a page regardless of the block size */
388 return !!(ofd_obd(ofd)->obd_self_export != exp &&
389 ofd->ofd_blockbits > COMPAT_BSIZE_SHIFT &&
390 !ofd_grant_param_supp(exp));
393 static inline int ofd_grant_prohibit(struct obd_export *exp,
394 struct ofd_device *ofd)
396 /* When ofd_grant_compat_disable is set, we don't grant any space to
397 * clients not supporting OBD_CONNECT_GRANT_PARAM.
398 * Otherwise, space granted to such a client is inflated since it
399 * consumes CFS_PAGE_SIZE of grant space per block */
400 return !!(ofd_grant_compat(exp, ofd) && ofd->ofd_grant_compat_disable);
403 void ofd_grant_sanity_check(struct obd_device *obd, const char *func);
404 long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
406 void ofd_grant_discard(struct obd_export *exp);
407 void ofd_grant_prepare_read(const struct lu_env *env, struct obd_export *exp,
409 void ofd_grant_prepare_write(const struct lu_env *env, struct obd_export *exp,
410 struct obdo *oa, struct niobuf_remote *rnb,
412 void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp, int rc);
413 int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr);
416 int ofd_fmd_init(void);
417 void ofd_fmd_exit(void);
418 struct ofd_mod_data *ofd_fmd_find(struct obd_export *exp,
420 struct ofd_mod_data *ofd_fmd_get(struct obd_export *exp,
422 void ofd_fmd_put(struct obd_export *exp, struct ofd_mod_data *fmd);
423 void ofd_fmd_expire(struct obd_export *exp);
424 void ofd_fmd_cleanup(struct obd_export *exp);
426 void ofd_fmd_drop(struct obd_export *exp, struct lu_fid *fid);
428 #define ofd_fmd_drop(exp, fid) do {} while (0)
432 extern struct ldlm_valblock_ops ofd_lvbo;
435 int ofd_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp,
436 void *req_cookie, ldlm_mode_t mode, int flags,
439 static inline struct ofd_thread_info * ofd_info(const struct lu_env *env)
441 struct ofd_thread_info *info;
443 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
445 LASSERT(info->fti_env);
446 LASSERT(info->fti_env == env);
450 static inline struct ofd_thread_info * ofd_info_init(const struct lu_env *env,
451 struct obd_export *exp)
453 struct ofd_thread_info *info;
455 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
457 LASSERT(info->fti_exp == NULL);
458 LASSERT(info->fti_env == NULL);
459 LASSERT(info->fti_attr.la_valid == 0);
463 info->fti_pre_version = 0;
464 info->fti_transno = 0;
465 info->fti_has_trans = 0;
469 /* The same as osc_build_res_name() */
470 static inline void ofd_build_resid(const struct lu_fid *fid,
471 struct ldlm_res_id *resname)
473 if (fid_is_idif(fid)) {
474 /* get id/seq like ostid_idif_pack() does */
475 osc_build_res_name(fid_idif_id(fid_seq(fid), fid_oid(fid),
477 FID_SEQ_OST_MDT0, resname);
479 /* In the future, where OSTs have FID sequences allocated. */
480 fid_build_reg_res_name(fid, resname);
484 static inline void ofd_fid_from_resid(struct lu_fid *fid,
485 const struct ldlm_res_id *name)
487 /* if seq is FID_SEQ_OST_MDT0 then we have IDIF and resid was built
488 * using osc_build_res_name function. */
489 if (fid_seq_is_mdt0(name->name[LUSTRE_RES_ID_VER_OID_OFF])) {
492 ostid.oi_id = name->name[LUSTRE_RES_ID_SEQ_OFF];
493 ostid.oi_seq = name->name[LUSTRE_RES_ID_VER_OID_OFF];
494 fid_ostid_unpack(fid, &ostid, 0);
496 fid->f_seq = name->name[LUSTRE_RES_ID_SEQ_OFF];
497 fid->f_oid = (__u32)name->name[LUSTRE_RES_ID_VER_OID_OFF];
498 fid->f_ver = name->name[LUSTRE_RES_ID_VER_OID_OFF] >> 32;
502 static inline void ofd_oti2info(struct ofd_thread_info *info,
503 struct obd_trans_info *oti)
505 info->fti_xid = oti->oti_xid;
506 info->fti_transno = oti->oti_transno;
507 info->fti_pre_version = oti->oti_pre_version;
510 static inline void ofd_info2oti(struct ofd_thread_info *info,
511 struct obd_trans_info *oti)
513 oti->oti_xid = info->fti_xid;
514 LASSERTF(ergo(oti->oti_transno > 0,
515 oti->oti_transno == info->fti_transno),
516 "Overwrite replay transno "LPX64" by "LPX64"\n",
517 oti->oti_transno, info->fti_transno);
518 oti->oti_transno = info->fti_transno;
519 oti->oti_pre_version = info->fti_pre_version;
522 /* sync on lock cancel is useless when we force a journal flush,
523 * and if we enable async journal commit, we should also turn on
524 * sync on lock cancel if it is not enabled already. */
525 static inline void ofd_slc_set(struct ofd_device *ofd)
527 if (ofd->ofd_syncjournal == 1)
528 ofd->ofd_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
529 else if (ofd->ofd_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
530 ofd->ofd_sync_lock_cancel = ALWAYS_SYNC_ON_CANCEL;
533 static inline void ofd_prepare_fidea(struct filter_fid *ff, struct obdo *oa)
535 if (!(oa->o_valid & OBD_MD_FLGROUP))
537 /* packing fid and converting it to LE for storing into EA.
538 * Here ->o_stripe_idx should be filled by LOV and rest of
539 * fields - by client. */
540 ff->ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
541 ff->ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
542 /* XXX: we are ignoring o_parent_ver here, since this should
543 * be the same for all objects in this fileset. */
544 ff->ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
545 ff->ff_objid = cpu_to_le64(oa->o_id);
546 ff->ff_seq = cpu_to_le64(oa->o_seq);
549 /* niobuf_remote has no rnb_ prefix in master */
550 #define rnb_offset offset
551 #define rnb_flags flags
553 /* the same for niobuf_local */
554 #define lnb_flags flags
557 #endif /* _OFD_INTERNAL_H */