1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _FS_CEPH_OSD_CLIENT_H
3 #define _FS_CEPH_OSD_CLIENT_H
5 #include <linux/bitrev.h>
6 #include <linux/completion.h>
7 #include <linux/kref.h>
8 #include <linux/mempool.h>
9 #include <linux/rbtree.h>
10 #include <linux/refcount.h>
12 #include <linux/ceph/types.h>
13 #include <linux/ceph/osdmap.h>
14 #include <linux/ceph/messenger.h>
15 #include <linux/ceph/msgpool.h>
16 #include <linux/ceph/auth.h>
17 #include <linux/ceph/pagelist.h>
20 struct ceph_snap_context
;
21 struct ceph_osd_request
;
22 struct ceph_osd_client
;
25 * completion callback for async writepages
27 typedef void (*ceph_osdc_callback_t
)(struct ceph_osd_request
*);
29 #define CEPH_HOMELESS_OSD -1
31 /* a given osd we're communicating with */
34 struct ceph_osd_client
*o_osdc
;
37 struct rb_node o_node
;
38 struct ceph_connection o_con
;
39 struct rb_root o_requests
;
40 struct rb_root o_linger_requests
;
41 struct rb_root o_backoff_mappings
;
42 struct rb_root o_backoffs_by_id
;
43 struct list_head o_osd_lru
;
44 struct ceph_auth_handshake o_auth
;
45 unsigned long lru_ttl
;
46 struct list_head o_keepalive_item
;
50 #define CEPH_OSD_SLAB_OPS 2
51 #define CEPH_OSD_MAX_OPS 16
53 enum ceph_osd_data_type
{
54 CEPH_OSD_DATA_TYPE_NONE
= 0,
55 CEPH_OSD_DATA_TYPE_PAGES
,
56 CEPH_OSD_DATA_TYPE_PAGELIST
,
58 CEPH_OSD_DATA_TYPE_BIO
,
59 #endif /* CONFIG_BLOCK */
60 CEPH_OSD_DATA_TYPE_BVECS
,
63 struct ceph_osd_data
{
64 enum ceph_osd_data_type type
;
73 struct ceph_pagelist
*pagelist
;
76 struct ceph_bio_iter bio_pos
;
79 #endif /* CONFIG_BLOCK */
81 struct ceph_bvec_iter bvec_pos
;
87 struct ceph_osd_req_op
{
88 u16 op
; /* CEPH_OSD_OP_* */
89 u32 flags
; /* CEPH_OSD_OP_FLAG_* */
90 u32 indata_len
; /* request */
91 u32 outdata_len
; /* reply */
95 struct ceph_osd_data raw_data_in
;
100 struct ceph_osd_data osd_data
;
105 __u8 cmp_op
; /* CEPH_OSD_CMPXATTR_OP_* */
106 __u8 cmp_mode
; /* CEPH_OSD_CMPXATTR_MODE_* */
107 struct ceph_osd_data osd_data
;
110 const char *class_name
;
111 const char *method_name
;
112 struct ceph_osd_data request_info
;
113 struct ceph_osd_data request_data
;
114 struct ceph_osd_data response_data
;
121 __u8 op
; /* CEPH_OSD_WATCH_OP_ */
125 struct ceph_osd_data request_data
;
129 struct ceph_osd_data request_data
;
130 struct ceph_osd_data response_data
;
133 struct ceph_osd_data response_data
;
136 u64 expected_object_size
;
137 u64 expected_write_size
;
143 u32 src_fadvise_flags
;
144 struct ceph_osd_data osd_data
;
149 struct ceph_osd_request_target
{
150 struct ceph_object_id base_oid
;
151 struct ceph_object_locator base_oloc
;
152 struct ceph_object_id target_oid
;
153 struct ceph_object_locator target_oloc
;
155 struct ceph_pg pgid
; /* last raw pg we mapped to */
156 struct ceph_spg spgid
; /* last actual spg we mapped to */
159 struct ceph_osds acting
;
164 bool recovery_deletes
;
166 unsigned int flags
; /* CEPH_OSD_FLAG_* */
170 u32 last_force_resend
;
175 /* an in-flight request */
176 struct ceph_osd_request
{
177 u64 r_tid
; /* unique for this client */
178 struct rb_node r_node
;
179 struct rb_node r_mc_node
; /* map check */
180 struct work_struct r_complete_work
;
181 struct ceph_osd
*r_osd
;
183 struct ceph_osd_request_target r_t
;
184 #define r_base_oid r_t.base_oid
185 #define r_base_oloc r_t.base_oloc
186 #define r_flags r_t.flags
188 struct ceph_msg
*r_request
, *r_reply
;
189 u32 r_sent
; /* >0 if r_request is sending/sent */
191 /* request osd ops array */
192 unsigned int r_num_ops
;
196 struct ceph_osd_client
*r_osdc
;
199 struct completion r_completion
; /* private to osd_client.c */
200 ceph_osdc_callback_t r_callback
;
201 struct list_head r_unsafe_item
;
203 struct inode
*r_inode
; /* for use by callbacks */
204 void *r_priv
; /* ditto */
206 /* set by submitter */
207 u64 r_snapid
; /* for reads, CEPH_NOSNAP o/w */
208 struct ceph_snap_context
*r_snapc
; /* for writes */
209 struct timespec64 r_mtime
; /* ditto */
210 u64 r_data_offset
; /* ditto */
211 bool r_linger
; /* don't resend on failure */
214 unsigned long r_stamp
; /* jiffies, send or check time */
215 unsigned long r_start_stamp
; /* jiffies */
219 struct ceph_osd_req_op r_ops
[];
222 struct ceph_request_redirect
{
223 struct ceph_object_locator oloc
;
227 * osd request identifier
229 * caller name + incarnation# + tid to unique identify this request
231 struct ceph_osd_reqid
{
232 struct ceph_entity_name name
;
237 struct ceph_blkin_trace_info
{
240 __le64 parent_span_id
;
243 typedef void (*rados_watchcb2_t
)(void *arg
, u64 notify_id
, u64 cookie
,
244 u64 notifier_id
, void *data
, size_t data_len
);
245 typedef void (*rados_watcherrcb_t
)(void *arg
, u64 cookie
, int err
);
247 struct ceph_osd_linger_request
{
248 struct ceph_osd_client
*osdc
;
251 bool is_watch
; /* watch or notify */
253 struct ceph_osd
*osd
;
254 struct ceph_osd_request
*reg_req
;
255 struct ceph_osd_request
*ping_req
;
256 unsigned long ping_sent
;
257 unsigned long watch_valid_thru
;
258 struct list_head pending_lworks
;
260 struct ceph_osd_request_target t
;
263 struct timespec64 mtime
;
267 struct rb_node node
; /* osd */
268 struct rb_node osdc_node
; /* osdc */
269 struct rb_node mc_node
; /* map check */
270 struct list_head scan_item
;
272 struct completion reg_commit_wait
;
273 struct completion notify_finish_wait
;
274 int reg_commit_error
;
275 int notify_finish_error
;
281 rados_watchcb2_t wcb
;
282 rados_watcherrcb_t errcb
;
285 struct page
***preply_pages
;
289 struct ceph_watch_item
{
290 struct ceph_entity_name name
;
292 struct ceph_entity_addr addr
;
295 struct ceph_spg_mapping
{
297 struct ceph_spg spgid
;
299 struct rb_root backoffs
;
302 struct ceph_hobject_id
{
315 u32 hash_reverse_bits
;
318 static inline void ceph_hoid_build_hash_cache(struct ceph_hobject_id
*hoid
)
320 hoid
->hash_reverse_bits
= bitrev32(hoid
->hash
);
324 * PG-wide backoff: [begin, end)
325 * per-object backoff: begin == end
327 struct ceph_osd_backoff
{
328 struct rb_node spg_node
;
329 struct rb_node id_node
;
331 struct ceph_spg spgid
;
333 struct ceph_hobject_id
*begin
;
334 struct ceph_hobject_id
*end
;
337 #define CEPH_LINGER_ID_START 0xffff000000000000ULL
339 struct ceph_osd_client
{
340 struct ceph_client
*client
;
342 struct ceph_osdmap
*osdmap
; /* current map */
343 struct rw_semaphore lock
;
345 struct rb_root osds
; /* osds */
346 struct list_head osd_lru
; /* idle osds */
347 spinlock_t osd_lru_lock
;
349 struct ceph_osd homeless_osd
;
350 atomic64_t last_tid
; /* tid of last request */
352 struct rb_root linger_requests
; /* lingering requests */
353 struct rb_root map_checks
;
354 struct rb_root linger_map_checks
;
355 atomic_t num_requests
;
356 atomic_t num_homeless
;
358 struct delayed_work timeout_work
;
359 struct delayed_work osds_timeout_work
;
360 #ifdef CONFIG_DEBUG_FS
361 struct dentry
*debugfs_file
;
364 mempool_t
*req_mempool
;
366 struct ceph_msgpool msgpool_op
;
367 struct ceph_msgpool msgpool_op_reply
;
369 struct workqueue_struct
*notify_wq
;
370 struct workqueue_struct
*completion_wq
;
373 static inline bool ceph_osdmap_flag(struct ceph_osd_client
*osdc
, int flag
)
375 return osdc
->osdmap
->flags
& flag
;
378 extern int ceph_osdc_setup(void);
379 extern void ceph_osdc_cleanup(void);
381 extern int ceph_osdc_init(struct ceph_osd_client
*osdc
,
382 struct ceph_client
*client
);
383 extern void ceph_osdc_stop(struct ceph_osd_client
*osdc
);
385 extern void ceph_osdc_handle_reply(struct ceph_osd_client
*osdc
,
386 struct ceph_msg
*msg
);
387 extern void ceph_osdc_handle_map(struct ceph_osd_client
*osdc
,
388 struct ceph_msg
*msg
);
389 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client
*osdc
, u32 eb
);
390 void ceph_osdc_abort_requests(struct ceph_osd_client
*osdc
, int err
);
392 extern void osd_req_op_init(struct ceph_osd_request
*osd_req
,
393 unsigned int which
, u16 opcode
, u32 flags
);
395 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request
*,
397 struct page
**pages
, u64 length
,
398 u32 alignment
, bool pages_from_pool
,
401 extern void osd_req_op_extent_init(struct ceph_osd_request
*osd_req
,
402 unsigned int which
, u16 opcode
,
403 u64 offset
, u64 length
,
404 u64 truncate_size
, u32 truncate_seq
);
405 extern void osd_req_op_extent_update(struct ceph_osd_request
*osd_req
,
406 unsigned int which
, u64 length
);
407 extern void osd_req_op_extent_dup_last(struct ceph_osd_request
*osd_req
,
408 unsigned int which
, u64 offset_inc
);
410 extern struct ceph_osd_data
*osd_req_op_extent_osd_data(
411 struct ceph_osd_request
*osd_req
,
414 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request
*,
416 struct page
**pages
, u64 length
,
417 u32 alignment
, bool pages_from_pool
,
419 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request
*,
421 struct ceph_pagelist
*pagelist
);
423 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request
*osd_req
,
425 struct ceph_bio_iter
*bio_pos
,
427 #endif /* CONFIG_BLOCK */
428 void osd_req_op_extent_osd_data_bvecs(struct ceph_osd_request
*osd_req
,
430 struct bio_vec
*bvecs
, u32 num_bvecs
,
432 void osd_req_op_extent_osd_data_bvec_pos(struct ceph_osd_request
*osd_req
,
434 struct ceph_bvec_iter
*bvec_pos
);
436 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request
*,
438 struct ceph_pagelist
*pagelist
);
439 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request
*,
441 struct page
**pages
, u64 length
,
442 u32 alignment
, bool pages_from_pool
,
444 void osd_req_op_cls_request_data_bvecs(struct ceph_osd_request
*osd_req
,
446 struct bio_vec
*bvecs
, u32 num_bvecs
,
448 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request
*,
450 struct page
**pages
, u64 length
,
451 u32 alignment
, bool pages_from_pool
,
453 int osd_req_op_cls_init(struct ceph_osd_request
*osd_req
, unsigned int which
,
454 const char *class, const char *method
);
455 extern int osd_req_op_xattr_init(struct ceph_osd_request
*osd_req
, unsigned int which
,
456 u16 opcode
, const char *name
, const void *value
,
457 size_t size
, u8 cmp_op
, u8 cmp_mode
);
458 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request
*osd_req
,
460 u64 expected_object_size
,
461 u64 expected_write_size
);
463 extern struct ceph_osd_request
*ceph_osdc_alloc_request(struct ceph_osd_client
*osdc
,
464 struct ceph_snap_context
*snapc
,
465 unsigned int num_ops
,
468 int ceph_osdc_alloc_messages(struct ceph_osd_request
*req
, gfp_t gfp
);
470 extern struct ceph_osd_request
*ceph_osdc_new_request(struct ceph_osd_client
*,
471 struct ceph_file_layout
*layout
,
472 struct ceph_vino vino
,
473 u64 offset
, u64
*len
,
474 unsigned int which
, int num_ops
,
475 int opcode
, int flags
,
476 struct ceph_snap_context
*snapc
,
477 u32 truncate_seq
, u64 truncate_size
,
480 extern void ceph_osdc_get_request(struct ceph_osd_request
*req
);
481 extern void ceph_osdc_put_request(struct ceph_osd_request
*req
);
483 extern int ceph_osdc_start_request(struct ceph_osd_client
*osdc
,
484 struct ceph_osd_request
*req
,
486 extern void ceph_osdc_cancel_request(struct ceph_osd_request
*req
);
487 extern int ceph_osdc_wait_request(struct ceph_osd_client
*osdc
,
488 struct ceph_osd_request
*req
);
489 extern void ceph_osdc_sync(struct ceph_osd_client
*osdc
);
491 extern void ceph_osdc_flush_notifies(struct ceph_osd_client
*osdc
);
492 void ceph_osdc_maybe_request_map(struct ceph_osd_client
*osdc
);
494 int ceph_osdc_call(struct ceph_osd_client
*osdc
,
495 struct ceph_object_id
*oid
,
496 struct ceph_object_locator
*oloc
,
497 const char *class, const char *method
,
499 struct page
*req_page
, size_t req_len
,
500 struct page
*resp_page
, size_t *resp_len
);
502 extern int ceph_osdc_readpages(struct ceph_osd_client
*osdc
,
503 struct ceph_vino vino
,
504 struct ceph_file_layout
*layout
,
506 u32 truncate_seq
, u64 truncate_size
,
507 struct page
**pages
, int nr_pages
,
510 extern int ceph_osdc_writepages(struct ceph_osd_client
*osdc
,
511 struct ceph_vino vino
,
512 struct ceph_file_layout
*layout
,
513 struct ceph_snap_context
*sc
,
515 u32 truncate_seq
, u64 truncate_size
,
516 struct timespec64
*mtime
,
517 struct page
**pages
, int nr_pages
);
519 int ceph_osdc_copy_from(struct ceph_osd_client
*osdc
,
520 u64 src_snapid
, u64 src_version
,
521 struct ceph_object_id
*src_oid
,
522 struct ceph_object_locator
*src_oloc
,
523 u32 src_fadvise_flags
,
524 struct ceph_object_id
*dst_oid
,
525 struct ceph_object_locator
*dst_oloc
,
526 u32 dst_fadvise_flags
,
530 struct ceph_osd_linger_request
*
531 ceph_osdc_watch(struct ceph_osd_client
*osdc
,
532 struct ceph_object_id
*oid
,
533 struct ceph_object_locator
*oloc
,
534 rados_watchcb2_t wcb
,
535 rados_watcherrcb_t errcb
,
537 int ceph_osdc_unwatch(struct ceph_osd_client
*osdc
,
538 struct ceph_osd_linger_request
*lreq
);
540 int ceph_osdc_notify_ack(struct ceph_osd_client
*osdc
,
541 struct ceph_object_id
*oid
,
542 struct ceph_object_locator
*oloc
,
547 int ceph_osdc_notify(struct ceph_osd_client
*osdc
,
548 struct ceph_object_id
*oid
,
549 struct ceph_object_locator
*oloc
,
553 struct page
***preply_pages
,
555 int ceph_osdc_watch_check(struct ceph_osd_client
*osdc
,
556 struct ceph_osd_linger_request
*lreq
);
557 int ceph_osdc_list_watchers(struct ceph_osd_client
*osdc
,
558 struct ceph_object_id
*oid
,
559 struct ceph_object_locator
*oloc
,
560 struct ceph_watch_item
**watchers
,