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;
142 struct ceph_osd_request_target {
143 struct ceph_object_id base_oid;
144 struct ceph_object_locator base_oloc;
145 struct ceph_object_id target_oid;
146 struct ceph_object_locator target_oloc;
148 struct ceph_pg pgid; /* last raw pg we mapped to */
149 struct ceph_spg spgid; /* last actual spg we mapped to */
152 struct ceph_osds acting;
157 bool recovery_deletes;
159 unsigned int flags; /* CEPH_OSD_FLAG_* */
163 u32 last_force_resend;
168 /* an in-flight request */
169 struct ceph_osd_request {
170 u64 r_tid; /* unique for this client */
171 struct rb_node r_node;
172 struct rb_node r_mc_node; /* map check */
173 struct work_struct r_complete_work;
174 struct ceph_osd *r_osd;
176 struct ceph_osd_request_target r_t;
177 #define r_base_oid r_t.base_oid
178 #define r_base_oloc r_t.base_oloc
179 #define r_flags r_t.flags
181 struct ceph_msg *r_request, *r_reply;
182 u32 r_sent; /* >0 if r_request is sending/sent */
184 /* request osd ops array */
185 unsigned int r_num_ops;
189 struct ceph_osd_client *r_osdc;
192 struct completion r_completion; /* private to osd_client.c */
193 ceph_osdc_callback_t r_callback;
194 struct list_head r_unsafe_item;
196 struct inode *r_inode; /* for use by callbacks */
197 void *r_priv; /* ditto */
199 /* set by submitter */
200 u64 r_snapid; /* for reads, CEPH_NOSNAP o/w */
201 struct ceph_snap_context *r_snapc; /* for writes */
202 struct timespec64 r_mtime; /* ditto */
203 u64 r_data_offset; /* ditto */
204 bool r_linger; /* don't resend on failure */
207 unsigned long r_stamp; /* jiffies, send or check time */
208 unsigned long r_start_stamp; /* jiffies */
212 struct ceph_osd_req_op r_ops[];
215 struct ceph_request_redirect {
216 struct ceph_object_locator oloc;
220 * osd request identifier
222 * caller name + incarnation# + tid to unique identify this request
224 struct ceph_osd_reqid {
225 struct ceph_entity_name name;
230 struct ceph_blkin_trace_info {
233 __le64 parent_span_id;
236 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
237 u64 notifier_id, void *data, size_t data_len);
238 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
240 struct ceph_osd_linger_request {
241 struct ceph_osd_client *osdc;
244 bool is_watch; /* watch or notify */
246 struct ceph_osd *osd;
247 struct ceph_osd_request *reg_req;
248 struct ceph_osd_request *ping_req;
249 unsigned long ping_sent;
250 unsigned long watch_valid_thru;
251 struct list_head pending_lworks;
253 struct ceph_osd_request_target t;
256 struct timespec64 mtime;
260 struct rb_node node; /* osd */
261 struct rb_node osdc_node; /* osdc */
262 struct rb_node mc_node; /* map check */
263 struct list_head scan_item;
265 struct completion reg_commit_wait;
266 struct completion notify_finish_wait;
267 int reg_commit_error;
268 int notify_finish_error;
274 rados_watchcb2_t wcb;
275 rados_watcherrcb_t errcb;
278 struct page ***preply_pages;
282 struct ceph_watch_item {
283 struct ceph_entity_name name;
285 struct ceph_entity_addr addr;
288 struct ceph_spg_mapping {
290 struct ceph_spg spgid;
292 struct rb_root backoffs;
295 struct ceph_hobject_id {
308 u32 hash_reverse_bits;
311 static inline void ceph_hoid_build_hash_cache(struct ceph_hobject_id *hoid)
313 hoid->hash_reverse_bits = bitrev32(hoid->hash);
317 * PG-wide backoff: [begin, end)
318 * per-object backoff: begin == end
320 struct ceph_osd_backoff {
321 struct rb_node spg_node;
322 struct rb_node id_node;
324 struct ceph_spg spgid;
326 struct ceph_hobject_id *begin;
327 struct ceph_hobject_id *end;
330 #define CEPH_LINGER_ID_START 0xffff000000000000ULL
332 struct ceph_osd_client {
333 struct ceph_client *client;
335 struct ceph_osdmap *osdmap; /* current map */
336 struct rw_semaphore lock;
338 struct rb_root osds; /* osds */
339 struct list_head osd_lru; /* idle osds */
340 spinlock_t osd_lru_lock;
342 struct ceph_osd homeless_osd;
343 atomic64_t last_tid; /* tid of last request */
345 struct rb_root linger_requests; /* lingering requests */
346 struct rb_root map_checks;
347 struct rb_root linger_map_checks;
348 atomic_t num_requests;
349 atomic_t num_homeless;
350 bool abort_on_full; /* abort w/ ENOSPC when full */
352 struct delayed_work timeout_work;
353 struct delayed_work osds_timeout_work;
354 #ifdef CONFIG_DEBUG_FS
355 struct dentry *debugfs_file;
358 mempool_t *req_mempool;
360 struct ceph_msgpool msgpool_op;
361 struct ceph_msgpool msgpool_op_reply;
363 struct workqueue_struct *notify_wq;
364 struct workqueue_struct *completion_wq;
367 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
369 return osdc->osdmap->flags & flag;
372 extern int ceph_osdc_setup(void);
373 extern void ceph_osdc_cleanup(void);
375 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
376 struct ceph_client *client);
377 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
379 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
380 struct ceph_msg *msg);
381 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
382 struct ceph_msg *msg);
383 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb);
384 void ceph_osdc_abort_requests(struct ceph_osd_client *osdc, int err);
386 extern void osd_req_op_init(struct ceph_osd_request *osd_req,
387 unsigned int which, u16 opcode, u32 flags);
389 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
391 struct page **pages, u64 length,
392 u32 alignment, bool pages_from_pool,
395 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
396 unsigned int which, u16 opcode,
397 u64 offset, u64 length,
398 u64 truncate_size, u32 truncate_seq);
399 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
400 unsigned int which, u64 length);
401 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
402 unsigned int which, u64 offset_inc);
404 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
405 struct ceph_osd_request *osd_req,
408 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
410 struct page **pages, u64 length,
411 u32 alignment, bool pages_from_pool,
413 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
415 struct ceph_pagelist *pagelist);
417 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
419 struct ceph_bio_iter *bio_pos,
421 #endif /* CONFIG_BLOCK */
422 void osd_req_op_extent_osd_data_bvecs(struct ceph_osd_request *osd_req,
424 struct bio_vec *bvecs, u32 num_bvecs,
426 void osd_req_op_extent_osd_data_bvec_pos(struct ceph_osd_request *osd_req,
428 struct ceph_bvec_iter *bvec_pos);
430 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
432 struct ceph_pagelist *pagelist);
433 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
435 struct page **pages, u64 length,
436 u32 alignment, bool pages_from_pool,
438 void osd_req_op_cls_request_data_bvecs(struct ceph_osd_request *osd_req,
440 struct bio_vec *bvecs, u32 num_bvecs,
442 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
444 struct page **pages, u64 length,
445 u32 alignment, bool pages_from_pool,
447 extern int osd_req_op_cls_init(struct ceph_osd_request *osd_req,
448 unsigned int which, u16 opcode,
449 const char *class, const char *method);
450 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
451 u16 opcode, const char *name, const void *value,
452 size_t size, u8 cmp_op, u8 cmp_mode);
453 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
455 u64 expected_object_size,
456 u64 expected_write_size);
458 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
459 struct ceph_snap_context *snapc,
460 unsigned int num_ops,
463 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
465 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
466 struct ceph_file_layout *layout,
467 struct ceph_vino vino,
468 u64 offset, u64 *len,
469 unsigned int which, int num_ops,
470 int opcode, int flags,
471 struct ceph_snap_context *snapc,
472 u32 truncate_seq, u64 truncate_size,
475 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
476 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
478 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
479 struct ceph_osd_request *req,
481 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
482 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
483 struct ceph_osd_request *req);
484 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
486 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
487 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
489 int ceph_osdc_call(struct ceph_osd_client *osdc,
490 struct ceph_object_id *oid,
491 struct ceph_object_locator *oloc,
492 const char *class, const char *method,
494 struct page *req_page, size_t req_len,
495 struct page *resp_page, size_t *resp_len);
497 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
498 struct ceph_vino vino,
499 struct ceph_file_layout *layout,
501 u32 truncate_seq, u64 truncate_size,
502 struct page **pages, int nr_pages,
505 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
506 struct ceph_vino vino,
507 struct ceph_file_layout *layout,
508 struct ceph_snap_context *sc,
510 u32 truncate_seq, u64 truncate_size,
511 struct timespec64 *mtime,
512 struct page **pages, int nr_pages);
515 struct ceph_osd_linger_request *
516 ceph_osdc_watch(struct ceph_osd_client *osdc,
517 struct ceph_object_id *oid,
518 struct ceph_object_locator *oloc,
519 rados_watchcb2_t wcb,
520 rados_watcherrcb_t errcb,
522 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
523 struct ceph_osd_linger_request *lreq);
525 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
526 struct ceph_object_id *oid,
527 struct ceph_object_locator *oloc,
532 int ceph_osdc_notify(struct ceph_osd_client *osdc,
533 struct ceph_object_id *oid,
534 struct ceph_object_locator *oloc,
538 struct page ***preply_pages,
540 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
541 struct ceph_osd_linger_request *lreq);
542 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
543 struct ceph_object_id *oid,
544 struct ceph_object_locator *oloc,
545 struct ceph_watch_item **watchers,