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>
45 #include <lustre_capa.h>
47 #define OFD_INIT_OBJID 0
48 #define OFD_PRECREATE_BATCH_DEFAULT (OBJ_SUBDIR_COUNT * 4)
50 /* on small filesystems we should not precreate too many objects in
51 * a single transaction, otherwise we can overflow transactions */
52 #define OFD_PRECREATE_SMALL_FS (1024ULL * 1024 * 1024)
53 #define OFD_PRECREATE_BATCH_SMALL 8
55 /* Limit the returned fields marked valid to those that we actually might set */
56 #define OFD_VALID_FLAGS (LA_TYPE | LA_MODE | LA_SIZE | LA_BLOCKS | \
57 LA_BLKSIZE | LA_ATIME | LA_MTIME | LA_CTIME)
59 /* per-client-per-object persistent state (LRU) */
61 cfs_list_t fmd_list; /* linked to fed_mod_list */
62 struct lu_fid fmd_fid; /* FID being written to */
63 __u64 fmd_mactime_xid; /* xid highest {m,a,c}time setattr */
64 cfs_time_t fmd_expire; /* time when the fmd should expire */
65 int fmd_refcount; /* reference counter - list holds 1 */
68 #define OFD_FMD_MAX_NUM_DEFAULT 128
69 #define OFD_FMD_MAX_AGE_DEFAULT ((obd_timeout + 10) * HZ)
71 #define OFD_SOFT_SYNC_LIMIT_DEFAULT 16
75 LPROC_OFD_STATS_READ = 0,
76 LPROC_OFD_STATS_WRITE,
77 LPROC_OFD_STATS_GETATTR,
78 LPROC_OFD_STATS_SETATTR,
79 LPROC_OFD_STATS_PUNCH,
81 LPROC_OFD_STATS_DESTROY,
82 LPROC_OFD_STATS_CREATE,
83 LPROC_OFD_STATS_STATFS,
84 LPROC_OFD_STATS_GET_INFO,
85 LPROC_OFD_STATS_SET_INFO,
86 LPROC_OFD_STATS_QUOTACTL,
90 static inline void ofd_counter_incr(struct obd_export *exp, int opcode,
91 char *jobid, long amount)
93 if (exp->exp_obd && exp->exp_obd->obd_stats)
94 lprocfs_counter_add(exp->exp_obd->obd_stats, opcode, amount);
96 if (exp->exp_obd && exp->exp_obd->u.obt.obt_jobstats.ojs_hash &&
97 (exp_connect_flags(exp) & OBD_CONNECT_JOBSTATS))
98 lprocfs_job_stats_log(exp->exp_obd, jobid, opcode, amount);
100 if (exp->exp_nid_stats != NULL &&
101 exp->exp_nid_stats->nid_stats != NULL) {
102 lprocfs_counter_add(exp->exp_nid_stats->nid_stats, opcode,
110 spinlock_t os_last_oid_lock;
111 struct mutex os_create_lock;
113 struct dt_object *os_lastid_obj;
114 unsigned long os_destroys_in_progress:1;
118 struct dt_device ofd_dt_dev;
119 struct dt_device *ofd_osd;
120 struct obd_export *ofd_osd_exp;
121 struct dt_device_param ofd_dt_conf;
122 /* DLM name-space for meta-data locks maintained by this server */
123 struct ldlm_namespace *ofd_namespace;
126 struct lu_target ofd_lut;
127 struct dt_object *ofd_health_check_file;
129 int ofd_subdir_count;
130 __u64 ofd_inconsistency_self_detected;
131 __u64 ofd_inconsistency_self_repaired;
133 cfs_list_t ofd_seq_list;
134 rwlock_t ofd_seq_list_lock;
136 int ofd_precreate_batch;
137 spinlock_t ofd_batch_lock;
139 /* protect all statfs-related counters */
140 spinlock_t ofd_osfs_lock;
141 /* statfs optimization: we cache a bit */
142 struct obd_statfs ofd_osfs;
145 /* writes which might be be accounted twice in ofd_osfs.os_bavail */
146 obd_size ofd_osfs_unstable;
148 /* counters used during statfs update, protected by ofd_osfs_lock.
149 * record when some statfs refresh are in progress */
150 int ofd_statfs_inflight;
151 /* track writes completed while statfs refresh is underway.
152 * tracking is only effective when ofd_statfs_inflight > 1 */
153 obd_size ofd_osfs_inflight;
155 /* grants: all values in bytes */
156 /* grant lock to protect all grant counters */
157 spinlock_t ofd_grant_lock;
158 /* total amount of dirty data reported by clients in incoming obdo */
159 obd_size ofd_tot_dirty;
160 /* sum of filesystem space granted to clients for async writes */
161 obd_size ofd_tot_granted;
162 /* grant used by I/Os in progress (between prepare and commit) */
163 obd_size ofd_tot_pending;
164 /* free space threshold over which we stop granting space to clients
165 * ofd_grant_ratio is stored as a fixed-point fraction using
166 * OFD_GRANT_RATIO_SHIFT of the remaining free space, not in percentage
169 /* number of clients using grants */
170 int ofd_tot_granted_clients;
172 /* ofd mod data: ofd_device wide values */
173 int ofd_fmd_max_num; /* per ofd ofd_mod_data */
174 cfs_duration_t ofd_fmd_max_age; /* time to fmd expiry */
176 spinlock_t ofd_flags_lock;
177 unsigned long ofd_raid_degraded:1,
178 /* sync journal on writes */
180 /* shall we grant space to clients not
181 * supporting OBD_CONNECT_GRANT_PARAM? */
182 ofd_grant_compat_disable:1,
183 /* Protected by ofd_lastid_rwsem. */
184 ofd_lastid_rebuilding:1,
185 ofd_record_fid_accessed:1,
186 ofd_lfsck_verify_pfid:1;
187 struct seq_server_site ofd_seq_site;
188 /* the limit of SOFT_SYNC RPCs that will trigger a soft sync */
189 unsigned int ofd_soft_sync_limit;
190 /* Protect ::ofd_lastid_rebuilding */
191 struct rw_semaphore ofd_lastid_rwsem;
192 __u64 ofd_lastid_gen;
193 struct ptlrpc_thread ofd_inconsistency_thread;
194 struct list_head ofd_inconsistency_list;
195 spinlock_t ofd_inconsistency_lock;
198 static inline struct ofd_device *ofd_dev(struct lu_device *d)
200 return container_of0(d, struct ofd_device, ofd_dt_dev.dd_lu_dev);
203 static inline struct obd_device *ofd_obd(struct ofd_device *ofd)
205 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd;
208 static inline struct ofd_device *ofd_exp(struct obd_export *exp)
210 return ofd_dev(exp->exp_obd->obd_lu_dev);
213 static inline char *ofd_name(struct ofd_device *ofd)
215 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd->obd_name;
219 struct lu_object_header ofo_header;
220 struct dt_object ofo_obj;
221 struct lu_fid ofo_pfid;
222 unsigned int ofo_pfid_checking:1,
226 static inline struct ofd_object *ofd_obj(struct lu_object *o)
228 return container_of0(o, struct ofd_object, ofo_obj.do_lu);
231 static inline int ofd_object_exists(struct ofd_object *obj)
233 LASSERT(obj != NULL);
234 if (lu_object_is_dying(obj->ofo_obj.do_lu.lo_header))
236 return lu_object_exists(&obj->ofo_obj.do_lu);
239 static inline struct dt_object *fo2dt(struct ofd_object *obj)
241 return &obj->ofo_obj;
244 static inline struct dt_object *ofd_object_child(struct ofd_object *_obj)
246 struct lu_object *lu = &(_obj)->ofo_obj.do_lu;
248 return container_of0(lu_object_next(lu), struct dt_object, do_lu);
251 static inline struct ofd_device *ofd_obj2dev(const struct ofd_object *fo)
253 return ofd_dev(fo->ofo_obj.do_lu.lo_dev);
256 static inline struct lustre_capa *ofd_object_capa(const struct lu_env *env,
257 const struct ofd_object *obj)
259 /* TODO: see mdd_object_capa() */
263 static inline void ofd_read_lock(const struct lu_env *env,
264 struct ofd_object *fo)
266 struct dt_object *next = ofd_object_child(fo);
268 next->do_ops->do_read_lock(env, next, 0);
271 static inline void ofd_read_unlock(const struct lu_env *env,
272 struct ofd_object *fo)
274 struct dt_object *next = ofd_object_child(fo);
276 next->do_ops->do_read_unlock(env, next);
279 static inline void ofd_write_lock(const struct lu_env *env,
280 struct ofd_object *fo)
282 struct dt_object *next = ofd_object_child(fo);
284 next->do_ops->do_write_lock(env, next, 0);
287 static inline void ofd_write_unlock(const struct lu_env *env,
288 struct ofd_object *fo)
290 struct dt_object *next = ofd_object_child(fo);
292 next->do_ops->do_write_unlock(env, next);
296 * Common data shared by obdofd-level handlers. This is allocated per-thread
297 * to reduce stack consumption.
299 struct ofd_thread_info {
300 const struct lu_env *fti_env;
302 struct obd_export *fti_exp;
304 __u64 fti_pre_version;
306 struct lu_fid fti_fid;
307 struct lu_attr fti_attr;
308 struct lu_attr fti_attr2;
309 struct ldlm_res_id fti_resid;
311 struct filter_fid fti_mds_fid;
312 struct filter_fid_old fti_mds_fid_old;
314 struct ost_id fti_ostid;
315 struct ofd_object *fti_obj;
317 char name[64]; /* for ofd_init0() */
318 struct obd_statfs osfs; /* for obdofd_statfs() */
321 /* Ops object filename */
322 struct lu_name fti_name;
323 struct dt_object_format fti_dof;
324 struct lu_buf fti_buf;
327 /* Space used by the I/O, used by grant code */
328 unsigned long fti_used;
329 struct ost_lvb fti_lvb;
330 struct lfsck_request fti_lr;
333 extern void target_recovery_fini(struct obd_device *obd);
334 extern void target_recovery_init(struct lu_target *lut, svc_handler_t handler);
337 int ofd_update_capa_key(struct ofd_device *ofd, struct lustre_capa_key *key);
338 int ofd_auth_capa(struct obd_export *exp, const struct lu_fid *fid,
339 obd_seq seq, struct lustre_capa *capa, __u64 opc);
340 void ofd_free_capa_keys(struct ofd_device *ofd);
343 extern struct lu_context_key ofd_thread_key;
344 int ofd_postrecov(const struct lu_env *env, struct ofd_device *ofd);
345 int ofd_fiemap_get(const struct lu_env *env, struct ofd_device *ofd,
346 struct lu_fid *fid, struct ll_user_fiemap *fiemap);
349 extern struct obd_ops ofd_obd_ops;
350 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
351 struct obd_statfs *osfs, __u64 max_age,
353 int ofd_destroy_by_fid(const struct lu_env *env, struct ofd_device *ofd,
354 const struct lu_fid *fid, int orphan);
355 int ofd_statfs(const struct lu_env *env, struct obd_export *exp,
356 struct obd_statfs *osfs, __u64 max_age, __u32 flags);
357 int ofd_obd_disconnect(struct obd_export *exp);
360 obd_id ofd_seq_last_oid(struct ofd_seq *oseq);
361 void ofd_seq_last_oid_set(struct ofd_seq *oseq, obd_id id);
362 int ofd_seq_last_oid_write(const struct lu_env *env, struct ofd_device *ofd,
363 struct ofd_seq *oseq);
364 int ofd_seqs_init(const struct lu_env *env, struct ofd_device *ofd);
365 struct ofd_seq *ofd_seq_get(struct ofd_device *ofd, obd_seq seq);
366 void ofd_seq_put(const struct lu_env *env, struct ofd_seq *oseq);
368 int ofd_fs_setup(const struct lu_env *env, struct ofd_device *ofd,
369 struct obd_device *obd);
370 void ofd_fs_cleanup(const struct lu_env *env, struct ofd_device *ofd);
371 int ofd_precreate_batch(struct ofd_device *ofd, int batch);
372 struct ofd_seq *ofd_seq_load(const struct lu_env *env, struct ofd_device *ofd,
374 void ofd_seqs_fini(const struct lu_env *env, struct ofd_device *ofd);
375 void ofd_seqs_free(const struct lu_env *env, struct ofd_device *ofd);
378 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd);
379 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd);
380 int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
382 int ofd_preprw(const struct lu_env *env,int cmd, struct obd_export *exp,
383 struct obdo *oa, int objcount, struct obd_ioobj *obj,
384 struct niobuf_remote *rnb, int *nr_local,
385 struct niobuf_local *lnb, struct obd_trans_info *oti,
386 struct lustre_capa *capa);
387 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
388 struct obdo *oa, int objcount, struct obd_ioobj *obj,
389 struct niobuf_remote *rnb, int npages,
390 struct niobuf_local *lnb, struct obd_trans_info *oti,
394 struct thandle *ofd_trans_create(const struct lu_env *env,
395 struct ofd_device *ofd);
396 int ofd_trans_start(const struct lu_env *env,
397 struct ofd_device *ofd, struct ofd_object *fo,
399 void ofd_trans_stop(const struct lu_env *env, struct ofd_device *ofd,
400 struct thandle *th, int rc);
401 int ofd_txn_stop_cb(const struct lu_env *env, struct thandle *txn,
406 void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars);
407 void ofd_stats_counter_init(struct lprocfs_stats *stats);
409 static void lprocfs_ofd_init_vars(struct lprocfs_static_vars *lvars)
411 memset(lvars, 0, sizeof(*lvars));
413 static inline void ofd_stats_counter_init(struct lprocfs_stats *stats) {}
417 struct ofd_object *ofd_object_find(const struct lu_env *env,
418 struct ofd_device *ofd,
419 const struct lu_fid *fid);
420 int ofd_object_ff_load(const struct lu_env *env, struct ofd_object *fo);
421 int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd,
422 obd_id id, struct ofd_seq *oseq, int nr, int sync);
424 void ofd_object_put(const struct lu_env *env, struct ofd_object *fo);
425 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
426 struct lu_attr *la, struct filter_fid *ff);
427 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
428 __u64 start, __u64 end, struct lu_attr *la,
429 struct filter_fid *ff, struct obdo *oa);
430 int ofd_object_destroy(const struct lu_env *, struct ofd_object *, int);
431 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
433 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
434 struct lu_attr *la, int is_setattr);
437 struct ofd_object *ofd_object_find_exists(const struct lu_env *env,
438 struct ofd_device *ofd,
439 const struct lu_fid *fid)
441 struct ofd_object *fo;
443 fo = ofd_object_find(env, ofd, fid);
444 if (!IS_ERR(fo) && !ofd_object_exists(fo)) {
445 ofd_object_put(env, fo);
446 fo = ERR_PTR(-ENOENT);
452 #define OFD_GRANT_RATIO_SHIFT 8
453 static inline __u64 ofd_grant_reserved(struct ofd_device *ofd, obd_size bavail)
455 return (bavail * ofd->ofd_grant_ratio) >> OFD_GRANT_RATIO_SHIFT;
458 static inline int ofd_grant_ratio_conv(int percentage)
460 return (percentage << OFD_GRANT_RATIO_SHIFT) / 100;
463 static inline int ofd_grant_param_supp(struct obd_export *exp)
465 return !!(exp_connect_flags(exp) & OBD_CONNECT_GRANT_PARAM);
468 /* Blocksize used for client not supporting OBD_CONNECT_GRANT_PARAM.
469 * That's 4KB=2^12 which is the biggest block size known to work whatever
470 * the client's page size is. */
471 #define COMPAT_BSIZE_SHIFT 12
472 static inline int ofd_grant_compat(struct obd_export *exp,
473 struct ofd_device *ofd)
475 /* Clients which don't support OBD_CONNECT_GRANT_PARAM cannot handle
476 * a block size > page size and consume PAGE_CACHE_SIZE of grant when
477 * dirtying a page regardless of the block size */
478 return !!(ofd_obd(ofd)->obd_self_export != exp &&
479 ofd->ofd_blockbits > COMPAT_BSIZE_SHIFT &&
480 !ofd_grant_param_supp(exp));
483 static inline int ofd_grant_prohibit(struct obd_export *exp,
484 struct ofd_device *ofd)
486 /* When ofd_grant_compat_disable is set, we don't grant any space to
487 * clients not supporting OBD_CONNECT_GRANT_PARAM.
488 * Otherwise, space granted to such a client is inflated since it
489 * consumes PAGE_CACHE_SIZE of grant space per block */
490 return !!(ofd_grant_compat(exp, ofd) && ofd->ofd_grant_compat_disable);
493 void ofd_grant_sanity_check(struct obd_device *obd, const char *func);
494 long ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
495 obd_size want, bool conservative);
496 void ofd_grant_discard(struct obd_export *exp);
497 void ofd_grant_prepare_read(const struct lu_env *env, struct obd_export *exp,
499 void ofd_grant_prepare_write(const struct lu_env *env, struct obd_export *exp,
500 struct obdo *oa, struct niobuf_remote *rnb,
502 void ofd_grant_commit(const struct lu_env *env, struct obd_export *exp, int rc);
503 int ofd_grant_create(const struct lu_env *env, struct obd_export *exp, int *nr);
506 int ofd_fmd_init(void);
507 void ofd_fmd_exit(void);
508 struct ofd_mod_data *ofd_fmd_find(struct obd_export *exp,
509 const struct lu_fid *fid);
510 struct ofd_mod_data *ofd_fmd_get(struct obd_export *exp,
511 const struct lu_fid *fid);
512 void ofd_fmd_put(struct obd_export *exp, struct ofd_mod_data *fmd);
513 void ofd_fmd_expire(struct obd_export *exp);
514 void ofd_fmd_cleanup(struct obd_export *exp);
516 void ofd_fmd_drop(struct obd_export *exp, const struct lu_fid *fid);
518 #define ofd_fmd_drop(exp, fid) do {} while (0)
522 int ofd_fid_set_index(const struct lu_env *env, struct ofd_device *ofd,
524 int ofd_fid_init(const struct lu_env *env, struct ofd_device *ofd);
525 int ofd_fid_fini(const struct lu_env *env, struct ofd_device *ofd);
528 extern struct ldlm_valblock_ops ofd_lvbo;
531 int ofd_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp,
532 void *req_cookie, ldlm_mode_t mode, __u64 flags,
535 static inline struct ofd_thread_info *ofd_info(const struct lu_env *env)
537 struct ofd_thread_info *info;
539 lu_env_refill((void *)env);
540 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
545 static inline struct ofd_thread_info *ofd_info_init(const struct lu_env *env,
546 struct obd_export *exp)
548 struct ofd_thread_info *info;
550 info = ofd_info(env);
551 LASSERT(info->fti_exp == NULL);
552 LASSERT(info->fti_env == NULL);
553 LASSERT(info->fti_attr.la_valid == 0);
557 info->fti_pre_version = 0;
561 static inline struct ofd_thread_info *tsi2ofd_info(struct tgt_session_info *tsi)
563 struct ptlrpc_request *req = tgt_ses_req(tsi);
564 struct ofd_thread_info *info;
566 info = ofd_info(tsi->tsi_env);
567 LASSERT(info->fti_exp == NULL);
568 LASSERT(info->fti_env == NULL);
569 LASSERT(info->fti_attr.la_valid == 0);
571 info->fti_env = tsi->tsi_env;
572 info->fti_exp = tsi->tsi_exp;
574 info->fti_xid = req->rq_xid;
575 /** VBR: take versions from request */
576 if (req->rq_reqmsg != NULL &&
577 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
578 __u64 *pre_version = lustre_msg_get_versions(req->rq_reqmsg);
580 info->fti_pre_version = pre_version ? pre_version[0] : 0;
585 static inline void ofd_oti2info(struct ofd_thread_info *info,
586 struct obd_trans_info *oti)
588 info->fti_xid = oti->oti_xid;
589 info->fti_pre_version = oti->oti_pre_version;
592 static inline void ofd_info2oti(struct ofd_thread_info *info,
593 struct obd_trans_info *oti)
595 oti->oti_xid = info->fti_xid;
596 oti->oti_pre_version = info->fti_pre_version;
599 /* sync on lock cancel is useless when we force a journal flush,
600 * and if we enable async journal commit, we should also turn on
601 * sync on lock cancel if it is not enabled already. */
602 static inline void ofd_slc_set(struct ofd_device *ofd)
604 if (ofd->ofd_syncjournal == 1)
605 ofd->ofd_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
606 else if (ofd->ofd_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
607 ofd->ofd_lut.lut_sync_lock_cancel = ALWAYS_SYNC_ON_CANCEL;
610 static inline void ofd_prepare_fidea(struct filter_fid *ff,
611 const struct obdo *oa)
613 /* packing fid and converting it to LE for storing into EA.
614 * Here ->o_stripe_idx should be filled by LOV and rest of
615 * fields - by client. */
616 ff->ff_parent.f_seq = cpu_to_le64(oa->o_parent_seq);
617 ff->ff_parent.f_oid = cpu_to_le32(oa->o_parent_oid);
618 /* XXX: we are ignoring o_parent_ver here, since this should
619 * be the same for all objects in this fileset. */
620 ff->ff_parent.f_ver = cpu_to_le32(oa->o_stripe_idx);
623 /* niobuf_remote has no rnb_ prefix in master */
624 #define rnb_offset offset
625 #define rnb_flags flags
627 /* the same for niobuf_local */
628 #define lnb_flags flags
631 #endif /* _OFD_INTERNAL_H */