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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2012, 2016, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
33 #ifndef _OFD_INTERNAL_H
34 #define _OFD_INTERNAL_H
37 #include <obd_class.h>
38 #include <dt_object.h>
39 #include <md_object.h>
40 #include <lustre_fid.h>
42 #define OFD_INIT_OBJID 0
43 #define OFD_PRECREATE_BATCH_DEFAULT (OBJ_SUBDIR_COUNT * 4)
45 /* on small filesystems we should not precreate too many objects in
46 * a single transaction, otherwise we can overflow transactions */
47 #define OFD_PRECREATE_SMALL_FS (1024ULL * 1024 * 1024)
48 #define OFD_PRECREATE_BATCH_SMALL 8
50 /* Limit the returned fields marked valid to those that we actually might set */
51 #define OFD_VALID_FLAGS (LA_TYPE | LA_MODE | LA_SIZE | LA_BLOCKS | \
52 LA_BLKSIZE | LA_ATIME | LA_MTIME | LA_CTIME)
54 /* per-client-per-object persistent state (LRU) */
56 struct list_head fmd_list; /* linked to fed_mod_list */
57 struct lu_fid fmd_fid; /* FID being written to */
58 __u64 fmd_mactime_xid; /* xid highest {m,a,c}time setattr */
59 cfs_time_t fmd_expire; /* time when the fmd should expire */
60 int fmd_refcount; /* reference counter - list holds 1 */
63 #define OFD_FMD_MAX_NUM_DEFAULT 128
64 #define OFD_FMD_MAX_AGE_DEFAULT msecs_to_jiffies((obd_timeout+10)*MSEC_PER_SEC)
66 #define OFD_SOFT_SYNC_LIMIT_DEFAULT 16
70 LPROC_OFD_STATS_READ = 0,
71 LPROC_OFD_STATS_WRITE,
72 LPROC_OFD_STATS_GETATTR,
73 LPROC_OFD_STATS_SETATTR,
74 LPROC_OFD_STATS_PUNCH,
76 LPROC_OFD_STATS_DESTROY,
77 LPROC_OFD_STATS_CREATE,
78 LPROC_OFD_STATS_STATFS,
79 LPROC_OFD_STATS_GET_INFO,
80 LPROC_OFD_STATS_SET_INFO,
81 LPROC_OFD_STATS_QUOTACTL,
85 static inline void ofd_counter_incr(struct obd_export *exp, int opcode,
86 char *jobid, long amount)
88 if (exp->exp_obd && exp->exp_obd->obd_stats)
89 lprocfs_counter_add(exp->exp_obd->obd_stats, opcode, amount);
91 if (exp->exp_obd && exp->exp_obd->u.obt.obt_jobstats.ojs_hash &&
92 (exp_connect_flags(exp) & OBD_CONNECT_JOBSTATS))
93 lprocfs_job_stats_log(exp->exp_obd, jobid, opcode, amount);
95 if (exp->exp_nid_stats != NULL &&
96 exp->exp_nid_stats->nid_stats != NULL) {
97 lprocfs_counter_add(exp->exp_nid_stats->nid_stats, opcode,
103 struct list_head os_list;
105 spinlock_t os_last_oid_lock;
106 struct mutex os_create_lock;
108 struct dt_object *os_lastid_obj;
109 unsigned long os_destroys_in_progress:1;
113 struct dt_device ofd_dt_dev;
114 struct dt_device *ofd_osd;
115 struct obd_export *ofd_osd_exp;
116 struct dt_device_param ofd_dt_conf;
117 /* DLM name-space for meta-data locks maintained by this server */
118 struct ldlm_namespace *ofd_namespace;
121 struct lu_target ofd_lut;
122 struct dt_object *ofd_health_check_file;
123 struct local_oid_storage *ofd_los;
125 __u64 ofd_inconsistency_self_detected;
126 __u64 ofd_inconsistency_self_repaired;
128 struct list_head ofd_seq_list;
129 rwlock_t ofd_seq_list_lock;
131 int ofd_precreate_batch;
132 spinlock_t ofd_batch_lock;
134 /* protect all statfs-related counters */
135 spinlock_t ofd_osfs_lock;
136 /* statfs optimization: we cache a bit */
137 struct obd_statfs ofd_osfs;
140 /* counters used during statfs update, protected by ofd_osfs_lock.
141 * record when some statfs refresh are in progress */
142 int ofd_statfs_inflight;
144 /* writes between prep & commit which might be accounted twice in
145 * ofd_osfs.os_bavail */
146 u64 ofd_osfs_unstable;
148 /* track writes completed while statfs refresh is underway.
149 * tracking is only effective when ofd_statfs_inflight > 1 */
150 u64 ofd_osfs_inflight;
152 /* grants: all values in bytes */
153 /* grant lock to protect all grant counters */
154 spinlock_t ofd_grant_lock;
155 /* total amount of dirty data reported by clients in incoming obdo */
157 /* sum of filesystem space granted to clients for async writes */
159 /* grant used by I/Os in progress (between prepare and commit) */
161 /* number of clients using grants */
162 int ofd_tot_granted_clients;
164 /* preferred BRW size, decided by storage type and capability */
166 /* checksum types supported on this node */
167 enum cksum_types ofd_cksum_types_supported;
169 /* ofd mod data: ofd_device wide values */
170 int ofd_fmd_max_num; /* per ofd ofd_mod_data */
171 cfs_duration_t ofd_fmd_max_age; /* time to fmd expiry */
173 spinlock_t ofd_flags_lock;
174 unsigned long ofd_raid_degraded:1,
175 /* sync journal on writes */
177 /* shall we grant space to clients not
178 * supporting OBD_CONNECT_GRANT_PARAM? */
179 ofd_grant_compat_disable:1,
180 /* Protected by ofd_lastid_rwsem. */
181 ofd_lastid_rebuilding:1,
182 ofd_record_fid_accessed:1,
183 ofd_lfsck_verify_pfid:1,
185 struct seq_server_site ofd_seq_site;
186 /* the limit of SOFT_SYNC RPCs that will trigger a soft sync */
187 unsigned int ofd_soft_sync_limit;
188 /* Protect ::ofd_lastid_rebuilding */
189 struct rw_semaphore ofd_lastid_rwsem;
190 __u64 ofd_lastid_gen;
191 struct ptlrpc_thread ofd_inconsistency_thread;
192 struct list_head ofd_inconsistency_list;
193 spinlock_t ofd_inconsistency_lock;
196 static inline struct ofd_device *ofd_dev(struct lu_device *d)
198 return container_of0(d, struct ofd_device, ofd_dt_dev.dd_lu_dev);
201 static inline struct obd_device *ofd_obd(struct ofd_device *ofd)
203 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd;
206 static inline struct ofd_device *ofd_exp(struct obd_export *exp)
208 return ofd_dev(exp->exp_obd->obd_lu_dev);
211 static inline char *ofd_name(struct ofd_device *ofd)
213 return ofd->ofd_dt_dev.dd_lu_dev.ld_obd->obd_name;
217 struct lu_object_header ofo_header;
218 struct dt_object ofo_obj;
219 struct lu_fid ofo_pfid;
220 unsigned int ofo_pfid_checking:1,
224 static inline struct ofd_object *ofd_obj(struct lu_object *o)
226 return container_of0(o, struct ofd_object, ofo_obj.do_lu);
229 static inline int ofd_object_exists(struct ofd_object *obj)
231 LASSERT(obj != NULL);
232 if (lu_object_is_dying(obj->ofo_obj.do_lu.lo_header))
234 return lu_object_exists(&obj->ofo_obj.do_lu);
237 static inline struct dt_object *fo2dt(struct ofd_object *obj)
239 return &obj->ofo_obj;
242 static inline struct dt_object *ofd_object_child(struct ofd_object *_obj)
244 struct lu_object *lu = &(_obj)->ofo_obj.do_lu;
246 return container_of0(lu_object_next(lu), struct dt_object, do_lu);
249 static inline struct ofd_device *ofd_obj2dev(const struct ofd_object *fo)
251 return ofd_dev(fo->ofo_obj.do_lu.lo_dev);
254 static inline void ofd_read_lock(const struct lu_env *env,
255 struct ofd_object *fo)
257 struct dt_object *next = ofd_object_child(fo);
259 next->do_ops->do_read_lock(env, next, 0);
262 static inline void ofd_read_unlock(const struct lu_env *env,
263 struct ofd_object *fo)
265 struct dt_object *next = ofd_object_child(fo);
267 next->do_ops->do_read_unlock(env, next);
270 static inline void ofd_write_lock(const struct lu_env *env,
271 struct ofd_object *fo)
273 struct dt_object *next = ofd_object_child(fo);
275 next->do_ops->do_write_lock(env, next, 0);
278 static inline void ofd_write_unlock(const struct lu_env *env,
279 struct ofd_object *fo)
281 struct dt_object *next = ofd_object_child(fo);
283 next->do_ops->do_write_unlock(env, next);
287 * Common data shared by obdofd-level handlers. This is allocated per-thread
288 * to reduce stack consumption.
290 struct ofd_thread_info {
291 const struct lu_env *fti_env;
293 struct obd_export *fti_exp;
295 __u64 fti_pre_version;
297 struct lu_fid fti_fid;
298 struct lu_attr fti_attr;
299 struct lu_attr fti_attr2;
300 struct ldlm_res_id fti_resid;
302 struct filter_fid fti_mds_fid;
303 struct filter_fid_old fti_mds_fid_old;
305 struct ost_id fti_ostid;
306 struct ofd_object *fti_obj;
308 char name[64]; /* for ofd_init0() */
309 struct obd_statfs osfs; /* for obdofd_statfs() */
312 /* Ops object filename */
313 struct lu_name fti_name;
314 struct dt_object_format fti_dof;
315 struct lu_buf fti_buf;
318 /* Space used by the I/O, used by grant code */
319 unsigned long fti_used;
320 struct ost_lvb fti_lvb;
322 struct lfsck_request fti_lr;
323 struct obd_connect_data fti_ocd;
327 extern void target_recovery_fini(struct obd_device *obd);
328 extern void target_recovery_init(struct lu_target *lut, svc_handler_t handler);
331 extern struct lu_context_key ofd_thread_key;
332 int ofd_postrecov(const struct lu_env *env, struct ofd_device *ofd);
333 int ofd_fiemap_get(const struct lu_env *env, struct ofd_device *ofd,
334 struct lu_fid *fid, struct fiemap *fiemap);
337 extern struct obd_ops ofd_obd_ops;
338 int ofd_statfs_internal(const struct lu_env *env, struct ofd_device *ofd,
339 struct obd_statfs *osfs, __u64 max_age,
341 int ofd_destroy_by_fid(const struct lu_env *env, struct ofd_device *ofd,
342 const struct lu_fid *fid, int orphan);
343 int ofd_statfs(const struct lu_env *env, struct obd_export *exp,
344 struct obd_statfs *osfs, __u64 max_age, __u32 flags);
345 int ofd_obd_disconnect(struct obd_export *exp);
348 u64 ofd_seq_last_oid(struct ofd_seq *oseq);
349 void ofd_seq_last_oid_set(struct ofd_seq *oseq, u64 id);
350 int ofd_seq_last_oid_write(const struct lu_env *env, struct ofd_device *ofd,
351 struct ofd_seq *oseq);
352 int ofd_seqs_init(const struct lu_env *env, struct ofd_device *ofd);
353 struct ofd_seq *ofd_seq_get(struct ofd_device *ofd, u64 seq);
354 void ofd_seq_put(const struct lu_env *env, struct ofd_seq *oseq);
356 int ofd_fs_setup(const struct lu_env *env, struct ofd_device *ofd,
357 struct obd_device *obd);
358 void ofd_fs_cleanup(const struct lu_env *env, struct ofd_device *ofd);
359 int ofd_precreate_batch(struct ofd_device *ofd, int batch);
360 struct ofd_seq *ofd_seq_load(const struct lu_env *env, struct ofd_device *ofd,
362 void ofd_seqs_fini(const struct lu_env *env, struct ofd_device *ofd);
363 void ofd_seqs_free(const struct lu_env *env, struct ofd_device *ofd);
366 int ofd_start_inconsistency_verification_thread(struct ofd_device *ofd);
367 int ofd_stop_inconsistency_verification_thread(struct ofd_device *ofd);
368 int ofd_verify_ff(const struct lu_env *env, struct ofd_object *fo,
370 int ofd_preprw(const struct lu_env *env,int cmd, struct obd_export *exp,
371 struct obdo *oa, int objcount, struct obd_ioobj *obj,
372 struct niobuf_remote *rnb, int *nr_local,
373 struct niobuf_local *lnb);
374 int ofd_commitrw(const struct lu_env *env, int cmd, struct obd_export *exp,
375 struct obdo *oa, int objcount, struct obd_ioobj *obj,
376 struct niobuf_remote *rnb, int npages,
377 struct niobuf_local *lnb, int old_rc);
380 struct thandle *ofd_trans_create(const struct lu_env *env,
381 struct ofd_device *ofd);
382 int ofd_trans_start(const struct lu_env *env,
383 struct ofd_device *ofd, struct ofd_object *fo,
385 int ofd_trans_stop(const struct lu_env *env, struct ofd_device *ofd,
386 struct thandle *th, int rc);
387 int ofd_txn_stop_cb(const struct lu_env *env, struct thandle *txn,
391 #ifdef CONFIG_PROC_FS
392 extern struct lprocfs_vars lprocfs_ofd_obd_vars[];
393 void ofd_stats_counter_init(struct lprocfs_stats *stats);
395 static inline void ofd_stats_counter_init(struct lprocfs_stats *stats) {}
399 struct ofd_object *ofd_object_find(const struct lu_env *env,
400 struct ofd_device *ofd,
401 const struct lu_fid *fid);
402 int ofd_object_ff_load(const struct lu_env *env, struct ofd_object *fo);
403 int ofd_precreate_objects(const struct lu_env *env, struct ofd_device *ofd,
404 u64 id, struct ofd_seq *oseq, int nr, int sync);
406 static inline void ofd_object_put(const struct lu_env *env,
407 struct ofd_object *fo)
409 dt_object_put(env, &fo->ofo_obj);
411 int ofd_attr_set(const struct lu_env *env, struct ofd_object *fo,
412 struct lu_attr *la, struct filter_fid *ff);
413 int ofd_object_punch(const struct lu_env *env, struct ofd_object *fo,
414 __u64 start, __u64 end, struct lu_attr *la,
415 struct filter_fid *ff, struct obdo *oa);
416 int ofd_object_destroy(const struct lu_env *, struct ofd_object *, int);
417 int ofd_attr_get(const struct lu_env *env, struct ofd_object *fo,
419 int ofd_attr_handle_ugid(const struct lu_env *env, struct ofd_object *fo,
420 struct lu_attr *la, int is_setattr);
423 struct ofd_object *ofd_object_find_exists(const struct lu_env *env,
424 struct ofd_device *ofd,
425 const struct lu_fid *fid)
427 struct ofd_object *fo;
429 fo = ofd_object_find(env, ofd, fid);
430 if (!IS_ERR(fo) && !ofd_object_exists(fo)) {
431 ofd_object_put(env, fo);
432 fo = ERR_PTR(-ENOENT);
438 static inline int ofd_grant_param_supp(struct obd_export *exp)
440 return !!(exp_connect_flags(exp) & OBD_CONNECT_GRANT_PARAM);
443 /* Blocksize used for client not supporting OBD_CONNECT_GRANT_PARAM.
444 * That's 4KB=2^12 which is the biggest block size known to work whatever
445 * the client's page size is. */
446 #define COMPAT_BSIZE_SHIFT 12
448 static inline int ofd_grant_prohibit(struct obd_export *exp,
449 struct ofd_device *ofd)
451 /* When ofd_grant_compat_disable is set, we don't grant any space to
452 * clients not supporting OBD_CONNECT_GRANT_PARAM.
453 * Otherwise, space granted to such a client is inflated since it
454 * consumes PAGE_SIZE of grant space per block */
455 return !!(ofd_obd(ofd)->obd_self_export != exp &&
456 !ofd_grant_param_supp(exp) && ofd->ofd_grant_compat_disable);
459 void ofd_grant_sanity_check(struct obd_device *obd, const char *func);
460 void ofd_grant_connect(const struct lu_env *env, struct obd_export *exp,
461 struct obd_connect_data *data, bool new_conn);
462 void ofd_grant_discard(struct obd_export *exp);
463 void ofd_grant_prepare_read(const struct lu_env *env, struct obd_export *exp,
465 void ofd_grant_prepare_write(const struct lu_env *env, struct obd_export *exp,
466 struct obdo *oa, struct niobuf_remote *rnb,
468 void ofd_grant_commit(struct obd_export *exp, unsigned long grant_used, int rc);
469 int ofd_grant_commit_cb_add(struct thandle *th, struct obd_export *exp,
470 unsigned long grant);
471 long ofd_grant_create(const struct lu_env *env, struct obd_export *exp,
475 int ofd_fmd_init(void);
476 void ofd_fmd_exit(void);
477 struct ofd_mod_data *ofd_fmd_find(struct obd_export *exp,
478 const struct lu_fid *fid);
479 struct ofd_mod_data *ofd_fmd_get(struct obd_export *exp,
480 const struct lu_fid *fid);
481 void ofd_fmd_put(struct obd_export *exp, struct ofd_mod_data *fmd);
482 void ofd_fmd_expire(struct obd_export *exp);
483 void ofd_fmd_cleanup(struct obd_export *exp);
485 void ofd_fmd_drop(struct obd_export *exp, const struct lu_fid *fid);
487 #define ofd_fmd_drop(exp, fid) do {} while (0)
491 int ofd_fid_set_index(const struct lu_env *env, struct ofd_device *ofd,
493 int ofd_fid_init(const struct lu_env *env, struct ofd_device *ofd);
494 int ofd_fid_fini(const struct lu_env *env, struct ofd_device *ofd);
497 extern struct ldlm_valblock_ops ofd_lvbo;
500 int ofd_intent_policy(struct ldlm_namespace *ns, struct ldlm_lock **lockp,
501 void *req_cookie, enum ldlm_mode mode, __u64 flags,
504 static inline struct ofd_thread_info *ofd_info(const struct lu_env *env)
506 struct ofd_thread_info *info;
508 lu_env_refill((void *)env);
509 info = lu_context_key_get(&env->le_ctx, &ofd_thread_key);
514 static inline struct ofd_thread_info *ofd_info_init(const struct lu_env *env,
515 struct obd_export *exp)
517 struct ofd_thread_info *info;
519 info = ofd_info(env);
520 LASSERT(info->fti_exp == NULL);
521 LASSERT(info->fti_env == NULL);
522 LASSERT(info->fti_attr.la_valid == 0);
526 info->fti_pre_version = 0;
530 static inline struct ofd_thread_info *tsi2ofd_info(struct tgt_session_info *tsi)
532 struct ptlrpc_request *req = tgt_ses_req(tsi);
533 struct ofd_thread_info *info;
535 info = ofd_info(tsi->tsi_env);
536 LASSERT(info->fti_exp == NULL);
537 LASSERT(info->fti_env == NULL);
538 LASSERT(info->fti_attr.la_valid == 0);
540 info->fti_env = tsi->tsi_env;
541 info->fti_exp = tsi->tsi_exp;
543 info->fti_xid = req->rq_xid;
544 /** VBR: take versions from request */
545 if (req->rq_reqmsg != NULL &&
546 lustre_msg_get_flags(req->rq_reqmsg) & MSG_REPLAY) {
547 __u64 *pre_version = lustre_msg_get_versions(req->rq_reqmsg);
549 info->fti_pre_version = pre_version ? pre_version[0] : 0;
554 /* sync on lock cancel is useless when we force a journal flush,
555 * and if we enable async journal commit, we should also turn on
556 * sync on lock cancel if it is not enabled already. */
557 static inline void ofd_slc_set(struct ofd_device *ofd)
559 if (ofd->ofd_syncjournal == 1)
560 ofd->ofd_lut.lut_sync_lock_cancel = NEVER_SYNC_ON_CANCEL;
561 else if (ofd->ofd_lut.lut_sync_lock_cancel == NEVER_SYNC_ON_CANCEL)
562 ofd->ofd_lut.lut_sync_lock_cancel = ALWAYS_SYNC_ON_CANCEL;
565 static inline void ofd_prepare_fidea(struct filter_fid *ff,
566 const struct obdo *oa)
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);
578 static inline int ofd_validate_seq(struct obd_export *exp, __u64 seq)
580 struct filter_export_data *fed = &exp->exp_filter_data;
582 if (unlikely(seq == FID_SEQ_OST_MDT0 && fed->fed_group != 0)) {
583 /* IDIF request only operates on MDT0 group */
584 CERROR("%s: Invalid sequence %#llx for group %u\n",
585 exp->exp_obd->obd_name, seq, fed->fed_group);
591 #endif /* _OFD_INTERNAL_H */