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>
11 #include <linux/ktime.h>
13 #include <linux/ceph/types.h>
14 #include <linux/ceph/osdmap.h>
15 #include <linux/ceph/messenger.h>
16 #include <linux/ceph/msgpool.h>
17 #include <linux/ceph/auth.h>
18 #include <linux/ceph/pagelist.h>
21 struct ceph_snap_context;
22 struct ceph_osd_request;
23 struct ceph_osd_client;
26 * completion callback for async writepages
28 typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *);
30 #define CEPH_HOMELESS_OSD -1
32 /* a given osd we're communicating with */
35 struct ceph_osd_client *o_osdc;
38 struct rb_node o_node;
39 struct ceph_connection o_con;
40 struct rb_root o_requests;
41 struct rb_root o_linger_requests;
42 struct rb_root o_backoff_mappings;
43 struct rb_root o_backoffs_by_id;
44 struct list_head o_osd_lru;
45 struct ceph_auth_handshake o_auth;
46 unsigned long lru_ttl;
47 struct list_head o_keepalive_item;
51 #define CEPH_OSD_SLAB_OPS 2
52 #define CEPH_OSD_MAX_OPS 16
54 enum ceph_osd_data_type {
55 CEPH_OSD_DATA_TYPE_NONE = 0,
56 CEPH_OSD_DATA_TYPE_PAGES,
57 CEPH_OSD_DATA_TYPE_PAGELIST,
59 CEPH_OSD_DATA_TYPE_BIO,
60 #endif /* CONFIG_BLOCK */
61 CEPH_OSD_DATA_TYPE_BVECS,
64 struct ceph_osd_data {
65 enum ceph_osd_data_type type;
74 struct ceph_pagelist *pagelist;
77 struct ceph_bio_iter bio_pos;
80 #endif /* CONFIG_BLOCK */
82 struct ceph_bvec_iter bvec_pos;
88 struct ceph_osd_req_op {
89 u16 op; /* CEPH_OSD_OP_* */
90 u32 flags; /* CEPH_OSD_OP_FLAG_* */
91 u32 indata_len; /* request */
92 u32 outdata_len; /* reply */
96 struct ceph_osd_data raw_data_in;
101 struct ceph_osd_data osd_data;
106 __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */
107 __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */
108 struct ceph_osd_data osd_data;
111 const char *class_name;
112 const char *method_name;
113 struct ceph_osd_data request_info;
114 struct ceph_osd_data request_data;
115 struct ceph_osd_data response_data;
122 __u8 op; /* CEPH_OSD_WATCH_OP_ */
126 struct ceph_osd_data request_data;
130 struct ceph_osd_data request_data;
131 struct ceph_osd_data response_data;
134 struct ceph_osd_data response_data;
137 u64 expected_object_size;
138 u64 expected_write_size;
139 u32 flags; /* CEPH_OSD_OP_ALLOC_HINT_FLAG_* */
145 u32 src_fadvise_flags;
146 struct ceph_osd_data osd_data;
151 struct ceph_osd_request_target {
152 struct ceph_object_id base_oid;
153 struct ceph_object_locator base_oloc;
154 struct ceph_object_id target_oid;
155 struct ceph_object_locator target_oloc;
157 struct ceph_pg pgid; /* last raw pg we mapped to */
158 struct ceph_spg spgid; /* last actual spg we mapped to */
161 struct ceph_osds acting;
166 bool recovery_deletes;
168 unsigned int flags; /* CEPH_OSD_FLAG_* */
173 u32 last_force_resend;
178 /* an in-flight request */
179 struct ceph_osd_request {
180 u64 r_tid; /* unique for this client */
181 struct rb_node r_node;
182 struct rb_node r_mc_node; /* map check */
183 struct work_struct r_complete_work;
184 struct ceph_osd *r_osd;
186 struct ceph_osd_request_target r_t;
187 #define r_base_oid r_t.base_oid
188 #define r_base_oloc r_t.base_oloc
189 #define r_flags r_t.flags
191 struct ceph_msg *r_request, *r_reply;
192 u32 r_sent; /* >0 if r_request is sending/sent */
194 /* request osd ops array */
195 unsigned int r_num_ops;
199 struct ceph_osd_client *r_osdc;
202 struct completion r_completion; /* private to osd_client.c */
203 ceph_osdc_callback_t r_callback;
205 struct inode *r_inode; /* for use by callbacks */
206 struct list_head r_private_item; /* ditto */
207 void *r_priv; /* ditto */
209 /* set by submitter */
210 u64 r_snapid; /* for reads, CEPH_NOSNAP o/w */
211 struct ceph_snap_context *r_snapc; /* for writes */
212 struct timespec64 r_mtime; /* ditto */
213 u64 r_data_offset; /* ditto */
214 bool r_linger; /* don't resend on failure */
217 unsigned long r_stamp; /* jiffies, send or check time */
218 unsigned long r_start_stamp; /* jiffies */
219 ktime_t r_start_latency; /* ktime_t */
220 ktime_t r_end_latency; /* ktime_t */
224 struct ceph_osd_req_op r_ops[];
227 struct ceph_request_redirect {
228 struct ceph_object_locator oloc;
232 * osd request identifier
234 * caller name + incarnation# + tid to unique identify this request
236 struct ceph_osd_reqid {
237 struct ceph_entity_name name;
242 struct ceph_blkin_trace_info {
245 __le64 parent_span_id;
248 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
249 u64 notifier_id, void *data, size_t data_len);
250 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
252 struct ceph_osd_linger_request {
253 struct ceph_osd_client *osdc;
256 bool is_watch; /* watch or notify */
258 struct ceph_osd *osd;
259 struct ceph_osd_request *reg_req;
260 struct ceph_osd_request *ping_req;
261 unsigned long ping_sent;
262 unsigned long watch_valid_thru;
263 struct list_head pending_lworks;
265 struct ceph_osd_request_target t;
268 struct timespec64 mtime;
272 struct rb_node node; /* osd */
273 struct rb_node osdc_node; /* osdc */
274 struct rb_node mc_node; /* map check */
275 struct list_head scan_item;
277 struct completion reg_commit_wait;
278 struct completion notify_finish_wait;
279 int reg_commit_error;
280 int notify_finish_error;
286 rados_watchcb2_t wcb;
287 rados_watcherrcb_t errcb;
290 struct page ***preply_pages;
294 struct ceph_watch_item {
295 struct ceph_entity_name name;
297 struct ceph_entity_addr addr;
300 struct ceph_spg_mapping {
302 struct ceph_spg spgid;
304 struct rb_root backoffs;
307 struct ceph_hobject_id {
320 u32 hash_reverse_bits;
323 static inline void ceph_hoid_build_hash_cache(struct ceph_hobject_id *hoid)
325 hoid->hash_reverse_bits = bitrev32(hoid->hash);
329 * PG-wide backoff: [begin, end)
330 * per-object backoff: begin == end
332 struct ceph_osd_backoff {
333 struct rb_node spg_node;
334 struct rb_node id_node;
336 struct ceph_spg spgid;
338 struct ceph_hobject_id *begin;
339 struct ceph_hobject_id *end;
342 #define CEPH_LINGER_ID_START 0xffff000000000000ULL
344 struct ceph_osd_client {
345 struct ceph_client *client;
347 struct ceph_osdmap *osdmap; /* current map */
348 struct rw_semaphore lock;
350 struct rb_root osds; /* osds */
351 struct list_head osd_lru; /* idle osds */
352 spinlock_t osd_lru_lock;
354 struct ceph_osd homeless_osd;
355 atomic64_t last_tid; /* tid of last request */
357 struct rb_root linger_requests; /* lingering requests */
358 struct rb_root map_checks;
359 struct rb_root linger_map_checks;
360 atomic_t num_requests;
361 atomic_t num_homeless;
363 struct delayed_work timeout_work;
364 struct delayed_work osds_timeout_work;
365 #ifdef CONFIG_DEBUG_FS
366 struct dentry *debugfs_file;
369 mempool_t *req_mempool;
371 struct ceph_msgpool msgpool_op;
372 struct ceph_msgpool msgpool_op_reply;
374 struct workqueue_struct *notify_wq;
375 struct workqueue_struct *completion_wq;
378 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
380 return osdc->osdmap->flags & flag;
383 extern int ceph_osdc_setup(void);
384 extern void ceph_osdc_cleanup(void);
386 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
387 struct ceph_client *client);
388 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
389 extern void ceph_osdc_reopen_osds(struct ceph_osd_client *osdc);
391 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
392 struct ceph_msg *msg);
393 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
394 struct ceph_msg *msg);
395 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb);
396 void ceph_osdc_abort_requests(struct ceph_osd_client *osdc, int err);
397 void ceph_osdc_clear_abort_err(struct ceph_osd_client *osdc);
399 #define osd_req_op_data(oreq, whch, typ, fld) \
401 struct ceph_osd_request *__oreq = (oreq); \
402 unsigned int __whch = (whch); \
403 BUG_ON(__whch >= __oreq->r_num_ops); \
404 &__oreq->r_ops[__whch].typ.fld; \
407 struct ceph_osd_req_op *osd_req_op_init(struct ceph_osd_request *osd_req,
408 unsigned int which, u16 opcode, u32 flags);
410 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
412 struct page **pages, u64 length,
413 u32 alignment, bool pages_from_pool,
416 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
417 unsigned int which, u16 opcode,
418 u64 offset, u64 length,
419 u64 truncate_size, u32 truncate_seq);
420 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
421 unsigned int which, u64 length);
422 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
423 unsigned int which, u64 offset_inc);
425 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
426 struct ceph_osd_request *osd_req,
429 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
431 struct page **pages, u64 length,
432 u32 alignment, bool pages_from_pool,
434 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
436 struct ceph_pagelist *pagelist);
438 void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *osd_req,
440 struct ceph_bio_iter *bio_pos,
442 #endif /* CONFIG_BLOCK */
443 void osd_req_op_extent_osd_data_bvecs(struct ceph_osd_request *osd_req,
445 struct bio_vec *bvecs, u32 num_bvecs,
447 void osd_req_op_extent_osd_data_bvec_pos(struct ceph_osd_request *osd_req,
449 struct ceph_bvec_iter *bvec_pos);
451 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
453 struct ceph_pagelist *pagelist);
454 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
456 struct page **pages, u64 length,
457 u32 alignment, bool pages_from_pool,
459 void osd_req_op_cls_request_data_bvecs(struct ceph_osd_request *osd_req,
461 struct bio_vec *bvecs, u32 num_bvecs,
463 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
465 struct page **pages, u64 length,
466 u32 alignment, bool pages_from_pool,
468 int osd_req_op_cls_init(struct ceph_osd_request *osd_req, unsigned int which,
469 const char *class, const char *method);
470 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
471 u16 opcode, const char *name, const void *value,
472 size_t size, u8 cmp_op, u8 cmp_mode);
473 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
475 u64 expected_object_size,
476 u64 expected_write_size,
478 extern int osd_req_op_copy_from_init(struct ceph_osd_request *req,
479 u64 src_snapid, u64 src_version,
480 struct ceph_object_id *src_oid,
481 struct ceph_object_locator *src_oloc,
482 u32 src_fadvise_flags,
483 u32 dst_fadvise_flags,
484 u32 truncate_seq, u64 truncate_size,
487 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
488 struct ceph_snap_context *snapc,
489 unsigned int num_ops,
492 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
494 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
495 struct ceph_file_layout *layout,
496 struct ceph_vino vino,
497 u64 offset, u64 *len,
498 unsigned int which, int num_ops,
499 int opcode, int flags,
500 struct ceph_snap_context *snapc,
501 u32 truncate_seq, u64 truncate_size,
504 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
505 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
507 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
508 struct ceph_osd_request *req,
510 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
511 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
512 struct ceph_osd_request *req);
513 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
515 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
516 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
518 int ceph_osdc_call(struct ceph_osd_client *osdc,
519 struct ceph_object_id *oid,
520 struct ceph_object_locator *oloc,
521 const char *class, const char *method,
523 struct page *req_page, size_t req_len,
524 struct page **resp_pages, size_t *resp_len);
527 struct ceph_osd_linger_request *
528 ceph_osdc_watch(struct ceph_osd_client *osdc,
529 struct ceph_object_id *oid,
530 struct ceph_object_locator *oloc,
531 rados_watchcb2_t wcb,
532 rados_watcherrcb_t errcb,
534 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
535 struct ceph_osd_linger_request *lreq);
537 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
538 struct ceph_object_id *oid,
539 struct ceph_object_locator *oloc,
544 int ceph_osdc_notify(struct ceph_osd_client *osdc,
545 struct ceph_object_id *oid,
546 struct ceph_object_locator *oloc,
550 struct page ***preply_pages,
552 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
553 struct ceph_osd_linger_request *lreq);
554 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
555 struct ceph_object_id *oid,
556 struct ceph_object_locator *oloc,
557 struct ceph_watch_item **watchers,