1 #ifndef _FS_CEPH_OSD_CLIENT_H
2 #define _FS_CEPH_OSD_CLIENT_H
4 #include <linux/bitrev.h>
5 #include <linux/completion.h>
6 #include <linux/kref.h>
7 #include <linux/mempool.h>
8 #include <linux/rbtree.h>
9 #include <linux/refcount.h>
11 #include <linux/ceph/types.h>
12 #include <linux/ceph/osdmap.h>
13 #include <linux/ceph/messenger.h>
14 #include <linux/ceph/msgpool.h>
15 #include <linux/ceph/auth.h>
16 #include <linux/ceph/pagelist.h>
19 struct ceph_snap_context;
20 struct ceph_osd_request;
21 struct ceph_osd_client;
24 * completion callback for async writepages
26 typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *);
28 #define CEPH_HOMELESS_OSD -1
30 /* a given osd we're communicating with */
33 struct ceph_osd_client *o_osdc;
36 struct rb_node o_node;
37 struct ceph_connection o_con;
38 struct rb_root o_requests;
39 struct rb_root o_linger_requests;
40 struct rb_root o_backoff_mappings;
41 struct rb_root o_backoffs_by_id;
42 struct list_head o_osd_lru;
43 struct ceph_auth_handshake o_auth;
44 unsigned long lru_ttl;
45 struct list_head o_keepalive_item;
49 #define CEPH_OSD_SLAB_OPS 2
50 #define CEPH_OSD_MAX_OPS 16
52 enum ceph_osd_data_type {
53 CEPH_OSD_DATA_TYPE_NONE = 0,
54 CEPH_OSD_DATA_TYPE_PAGES,
55 CEPH_OSD_DATA_TYPE_PAGELIST,
57 CEPH_OSD_DATA_TYPE_BIO,
58 #endif /* CONFIG_BLOCK */
61 struct ceph_osd_data {
62 enum ceph_osd_data_type type;
71 struct ceph_pagelist *pagelist;
74 struct bio *bio; /* list of bios */
75 size_t bio_length; /* total in list */
77 #endif /* CONFIG_BLOCK */
81 struct ceph_osd_req_op {
82 u16 op; /* CEPH_OSD_OP_* */
83 u32 flags; /* CEPH_OSD_OP_FLAG_* */
84 u32 indata_len; /* request */
85 u32 outdata_len; /* reply */
89 struct ceph_osd_data raw_data_in;
94 struct ceph_osd_data osd_data;
99 __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */
100 __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */
101 struct ceph_osd_data osd_data;
104 const char *class_name;
105 const char *method_name;
106 struct ceph_osd_data request_info;
107 struct ceph_osd_data request_data;
108 struct ceph_osd_data response_data;
115 __u8 op; /* CEPH_OSD_WATCH_OP_ */
119 struct ceph_osd_data request_data;
123 struct ceph_osd_data request_data;
124 struct ceph_osd_data response_data;
127 struct ceph_osd_data response_data;
130 u64 expected_object_size;
131 u64 expected_write_size;
136 struct ceph_osd_request_target {
137 struct ceph_object_id base_oid;
138 struct ceph_object_locator base_oloc;
139 struct ceph_object_id target_oid;
140 struct ceph_object_locator target_oloc;
142 struct ceph_pg pgid; /* last raw pg we mapped to */
143 struct ceph_spg spgid; /* last actual spg we mapped to */
146 struct ceph_osds acting;
151 bool recovery_deletes;
153 unsigned int flags; /* CEPH_OSD_FLAG_* */
157 u32 last_force_resend;
162 /* an in-flight request */
163 struct ceph_osd_request {
164 u64 r_tid; /* unique for this client */
165 struct rb_node r_node;
166 struct rb_node r_mc_node; /* map check */
167 struct ceph_osd *r_osd;
169 struct ceph_osd_request_target r_t;
170 #define r_base_oid r_t.base_oid
171 #define r_base_oloc r_t.base_oloc
172 #define r_flags r_t.flags
174 struct ceph_msg *r_request, *r_reply;
175 u32 r_sent; /* >0 if r_request is sending/sent */
177 /* request osd ops array */
178 unsigned int r_num_ops;
182 struct ceph_osd_client *r_osdc;
185 struct completion r_completion; /* private to osd_client.c */
186 ceph_osdc_callback_t r_callback;
187 struct list_head r_unsafe_item;
189 struct inode *r_inode; /* for use by callbacks */
190 void *r_priv; /* ditto */
192 /* set by submitter */
193 u64 r_snapid; /* for reads, CEPH_NOSNAP o/w */
194 struct ceph_snap_context *r_snapc; /* for writes */
195 struct timespec r_mtime; /* ditto */
196 u64 r_data_offset; /* ditto */
197 bool r_linger; /* don't resend on failure */
198 bool r_abort_on_full; /* return ENOSPC when full */
201 unsigned long r_stamp; /* jiffies, send or check time */
202 unsigned long r_start_stamp; /* jiffies */
206 struct ceph_osd_req_op r_ops[];
209 struct ceph_request_redirect {
210 struct ceph_object_locator oloc;
214 * osd request identifier
216 * caller name + incarnation# + tid to unique identify this request
218 struct ceph_osd_reqid {
219 struct ceph_entity_name name;
224 struct ceph_blkin_trace_info {
227 __le64 parent_span_id;
230 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
231 u64 notifier_id, void *data, size_t data_len);
232 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
234 struct ceph_osd_linger_request {
235 struct ceph_osd_client *osdc;
238 bool is_watch; /* watch or notify */
240 struct ceph_osd *osd;
241 struct ceph_osd_request *reg_req;
242 struct ceph_osd_request *ping_req;
243 unsigned long ping_sent;
244 unsigned long watch_valid_thru;
245 struct list_head pending_lworks;
247 struct ceph_osd_request_target t;
250 struct timespec mtime;
254 struct rb_node node; /* osd */
255 struct rb_node osdc_node; /* osdc */
256 struct rb_node mc_node; /* map check */
257 struct list_head scan_item;
259 struct completion reg_commit_wait;
260 struct completion notify_finish_wait;
261 int reg_commit_error;
262 int notify_finish_error;
268 rados_watchcb2_t wcb;
269 rados_watcherrcb_t errcb;
272 struct page ***preply_pages;
276 struct ceph_watch_item {
277 struct ceph_entity_name name;
279 struct ceph_entity_addr addr;
282 struct ceph_spg_mapping {
284 struct ceph_spg spgid;
286 struct rb_root backoffs;
289 struct ceph_hobject_id {
302 u32 hash_reverse_bits;
305 static inline void ceph_hoid_build_hash_cache(struct ceph_hobject_id *hoid)
307 hoid->hash_reverse_bits = bitrev32(hoid->hash);
311 * PG-wide backoff: [begin, end)
312 * per-object backoff: begin == end
314 struct ceph_osd_backoff {
315 struct rb_node spg_node;
316 struct rb_node id_node;
318 struct ceph_spg spgid;
320 struct ceph_hobject_id *begin;
321 struct ceph_hobject_id *end;
324 #define CEPH_LINGER_ID_START 0xffff000000000000ULL
326 struct ceph_osd_client {
327 struct ceph_client *client;
329 struct ceph_osdmap *osdmap; /* current map */
330 struct rw_semaphore lock;
332 struct rb_root osds; /* osds */
333 struct list_head osd_lru; /* idle osds */
334 spinlock_t osd_lru_lock;
336 struct ceph_osd homeless_osd;
337 atomic64_t last_tid; /* tid of last request */
339 struct rb_root linger_requests; /* lingering requests */
340 struct rb_root map_checks;
341 struct rb_root linger_map_checks;
342 atomic_t num_requests;
343 atomic_t num_homeless;
344 struct delayed_work timeout_work;
345 struct delayed_work osds_timeout_work;
346 #ifdef CONFIG_DEBUG_FS
347 struct dentry *debugfs_file;
350 mempool_t *req_mempool;
352 struct ceph_msgpool msgpool_op;
353 struct ceph_msgpool msgpool_op_reply;
355 struct workqueue_struct *notify_wq;
358 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
360 return osdc->osdmap->flags & flag;
363 extern int ceph_osdc_setup(void);
364 extern void ceph_osdc_cleanup(void);
366 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
367 struct ceph_client *client);
368 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
370 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
371 struct ceph_msg *msg);
372 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
373 struct ceph_msg *msg);
374 void ceph_osdc_update_epoch_barrier(struct ceph_osd_client *osdc, u32 eb);
376 extern void osd_req_op_init(struct ceph_osd_request *osd_req,
377 unsigned int which, u16 opcode, u32 flags);
379 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
381 struct page **pages, u64 length,
382 u32 alignment, bool pages_from_pool,
385 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
386 unsigned int which, u16 opcode,
387 u64 offset, u64 length,
388 u64 truncate_size, u32 truncate_seq);
389 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
390 unsigned int which, u64 length);
391 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
392 unsigned int which, u64 offset_inc);
394 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
395 struct ceph_osd_request *osd_req,
398 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
400 struct page **pages, u64 length,
401 u32 alignment, bool pages_from_pool,
403 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
405 struct ceph_pagelist *pagelist);
407 extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *,
409 struct bio *bio, size_t bio_length);
410 #endif /* CONFIG_BLOCK */
412 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
414 struct ceph_pagelist *pagelist);
415 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
417 struct page **pages, u64 length,
418 u32 alignment, bool pages_from_pool,
420 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
422 struct page **pages, u64 length,
423 u32 alignment, bool pages_from_pool,
425 extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req,
426 unsigned int which, u16 opcode,
427 const char *class, const char *method);
428 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
429 u16 opcode, const char *name, const void *value,
430 size_t size, u8 cmp_op, u8 cmp_mode);
431 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
433 u64 expected_object_size,
434 u64 expected_write_size);
436 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
437 struct ceph_snap_context *snapc,
438 unsigned int num_ops,
441 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
443 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
444 struct ceph_file_layout *layout,
445 struct ceph_vino vino,
446 u64 offset, u64 *len,
447 unsigned int which, int num_ops,
448 int opcode, int flags,
449 struct ceph_snap_context *snapc,
450 u32 truncate_seq, u64 truncate_size,
453 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
454 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
456 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
457 struct ceph_osd_request *req,
459 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
460 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
461 struct ceph_osd_request *req);
462 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
464 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
465 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
467 int ceph_osdc_call(struct ceph_osd_client *osdc,
468 struct ceph_object_id *oid,
469 struct ceph_object_locator *oloc,
470 const char *class, const char *method,
472 struct page *req_page, size_t req_len,
473 struct page *resp_page, size_t *resp_len);
475 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
476 struct ceph_vino vino,
477 struct ceph_file_layout *layout,
479 u32 truncate_seq, u64 truncate_size,
480 struct page **pages, int nr_pages,
483 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
484 struct ceph_vino vino,
485 struct ceph_file_layout *layout,
486 struct ceph_snap_context *sc,
488 u32 truncate_seq, u64 truncate_size,
489 struct timespec *mtime,
490 struct page **pages, int nr_pages);
493 struct ceph_osd_linger_request *
494 ceph_osdc_watch(struct ceph_osd_client *osdc,
495 struct ceph_object_id *oid,
496 struct ceph_object_locator *oloc,
497 rados_watchcb2_t wcb,
498 rados_watcherrcb_t errcb,
500 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
501 struct ceph_osd_linger_request *lreq);
503 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
504 struct ceph_object_id *oid,
505 struct ceph_object_locator *oloc,
510 int ceph_osdc_notify(struct ceph_osd_client *osdc,
511 struct ceph_object_id *oid,
512 struct ceph_object_locator *oloc,
516 struct page ***preply_pages,
518 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
519 struct ceph_osd_linger_request *lreq);
520 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
521 struct ceph_object_id *oid,
522 struct ceph_object_locator *oloc,
523 struct ceph_watch_item **watchers,