1 #ifndef _FS_CEPH_OSD_CLIENT_H
2 #define _FS_CEPH_OSD_CLIENT_H
4 #include <linux/completion.h>
5 #include <linux/kref.h>
6 #include <linux/mempool.h>
7 #include <linux/rbtree.h>
9 #include <linux/ceph/types.h>
10 #include <linux/ceph/osdmap.h>
11 #include <linux/ceph/messenger.h>
12 #include <linux/ceph/msgpool.h>
13 #include <linux/ceph/auth.h>
14 #include <linux/ceph/pagelist.h>
17 struct ceph_snap_context;
18 struct ceph_osd_request;
19 struct ceph_osd_client;
22 * completion callback for async writepages
24 typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *);
25 typedef void (*ceph_osdc_unsafe_callback_t)(struct ceph_osd_request *, bool);
27 #define CEPH_HOMELESS_OSD -1
29 /* a given osd we're communicating with */
32 struct ceph_osd_client *o_osdc;
35 struct rb_node o_node;
36 struct ceph_connection o_con;
37 struct rb_root o_requests;
38 struct rb_root o_linger_requests;
39 struct list_head o_osd_lru;
40 struct ceph_auth_handshake o_auth;
41 unsigned long lru_ttl;
42 struct list_head o_keepalive_item;
46 #define CEPH_OSD_SLAB_OPS 2
47 #define CEPH_OSD_MAX_OPS 16
49 enum ceph_osd_data_type {
50 CEPH_OSD_DATA_TYPE_NONE = 0,
51 CEPH_OSD_DATA_TYPE_PAGES,
52 CEPH_OSD_DATA_TYPE_PAGELIST,
54 CEPH_OSD_DATA_TYPE_BIO,
55 #endif /* CONFIG_BLOCK */
58 struct ceph_osd_data {
59 enum ceph_osd_data_type type;
68 struct ceph_pagelist *pagelist;
71 struct bio *bio; /* list of bios */
72 size_t bio_length; /* total in list */
74 #endif /* CONFIG_BLOCK */
78 struct ceph_osd_req_op {
79 u16 op; /* CEPH_OSD_OP_* */
80 u32 flags; /* CEPH_OSD_OP_FLAG_* */
81 u32 indata_len; /* request */
82 u32 outdata_len; /* reply */
86 struct ceph_osd_data raw_data_in;
91 struct ceph_osd_data osd_data;
96 __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */
97 __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */
98 struct ceph_osd_data osd_data;
101 const char *class_name;
102 const char *method_name;
103 struct ceph_osd_data request_info;
104 struct ceph_osd_data request_data;
105 struct ceph_osd_data response_data;
112 __u8 op; /* CEPH_OSD_WATCH_OP_ */
116 struct ceph_osd_data request_data;
120 struct ceph_osd_data request_data;
121 struct ceph_osd_data response_data;
124 struct ceph_osd_data response_data;
127 u64 expected_object_size;
128 u64 expected_write_size;
133 struct ceph_osd_request_target {
134 struct ceph_object_id base_oid;
135 struct ceph_object_locator base_oloc;
136 struct ceph_object_id target_oid;
137 struct ceph_object_locator target_oloc;
142 struct ceph_osds acting;
148 unsigned int flags; /* CEPH_OSD_FLAG_* */
154 /* an in-flight request */
155 struct ceph_osd_request {
156 u64 r_tid; /* unique for this client */
157 struct rb_node r_node;
158 struct rb_node r_mc_node; /* map check */
159 struct ceph_osd *r_osd;
161 struct ceph_osd_request_target r_t;
162 #define r_base_oid r_t.base_oid
163 #define r_base_oloc r_t.base_oloc
164 #define r_flags r_t.flags
166 struct ceph_msg *r_request, *r_reply;
167 u32 r_sent; /* >0 if r_request is sending/sent */
169 /* request osd ops array */
170 unsigned int r_num_ops;
175 struct ceph_osd_client *r_osdc;
178 struct completion r_completion;
179 struct completion r_safe_completion; /* fsync waiter */
180 ceph_osdc_callback_t r_callback;
181 ceph_osdc_unsafe_callback_t r_unsafe_callback;
182 struct list_head r_unsafe_item;
184 struct inode *r_inode; /* for use by callbacks */
185 void *r_priv; /* ditto */
187 /* set by submitter */
188 u64 r_snapid; /* for reads, CEPH_NOSNAP o/w */
189 struct ceph_snap_context *r_snapc; /* for writes */
190 struct timespec r_mtime; /* ditto */
191 u64 r_data_offset; /* ditto */
192 bool r_linger; /* don't resend on failure */
195 unsigned long r_stamp; /* jiffies, send or check time */
197 struct ceph_eversion r_replay_version; /* aka reassert_version */
198 u32 r_last_force_resend;
201 struct ceph_osd_req_op r_ops[];
204 struct ceph_request_redirect {
205 struct ceph_object_locator oloc;
208 typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie,
209 u64 notifier_id, void *data, size_t data_len);
210 typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err);
212 struct ceph_osd_linger_request {
213 struct ceph_osd_client *osdc;
216 bool is_watch; /* watch or notify */
218 struct ceph_osd *osd;
219 struct ceph_osd_request *reg_req;
220 struct ceph_osd_request *ping_req;
221 unsigned long ping_sent;
222 unsigned long watch_valid_thru;
223 struct list_head pending_lworks;
225 struct ceph_osd_request_target t;
226 u32 last_force_resend;
229 struct timespec mtime;
233 struct rb_node node; /* osd */
234 struct rb_node osdc_node; /* osdc */
235 struct rb_node mc_node; /* map check */
236 struct list_head scan_item;
238 struct completion reg_commit_wait;
239 struct completion notify_finish_wait;
240 int reg_commit_error;
241 int notify_finish_error;
247 rados_watchcb2_t wcb;
248 rados_watcherrcb_t errcb;
251 struct page ***preply_pages;
255 struct ceph_watch_item {
256 struct ceph_entity_name name;
258 struct ceph_entity_addr addr;
261 struct ceph_osd_client {
262 struct ceph_client *client;
264 struct ceph_osdmap *osdmap; /* current map */
265 struct rw_semaphore lock;
267 struct rb_root osds; /* osds */
268 struct list_head osd_lru; /* idle osds */
269 spinlock_t osd_lru_lock;
270 struct ceph_osd homeless_osd;
271 atomic64_t last_tid; /* tid of last request */
273 struct rb_root linger_requests; /* lingering requests */
274 struct rb_root map_checks;
275 struct rb_root linger_map_checks;
276 atomic_t num_requests;
277 atomic_t num_homeless;
278 struct delayed_work timeout_work;
279 struct delayed_work osds_timeout_work;
280 #ifdef CONFIG_DEBUG_FS
281 struct dentry *debugfs_file;
284 mempool_t *req_mempool;
286 struct ceph_msgpool msgpool_op;
287 struct ceph_msgpool msgpool_op_reply;
289 struct workqueue_struct *notify_wq;
292 static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag)
294 return osdc->osdmap->flags & flag;
297 extern int ceph_osdc_setup(void);
298 extern void ceph_osdc_cleanup(void);
300 extern int ceph_osdc_init(struct ceph_osd_client *osdc,
301 struct ceph_client *client);
302 extern void ceph_osdc_stop(struct ceph_osd_client *osdc);
304 extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc,
305 struct ceph_msg *msg);
306 extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc,
307 struct ceph_msg *msg);
309 extern void osd_req_op_init(struct ceph_osd_request *osd_req,
310 unsigned int which, u16 opcode, u32 flags);
312 extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *,
314 struct page **pages, u64 length,
315 u32 alignment, bool pages_from_pool,
318 extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req,
319 unsigned int which, u16 opcode,
320 u64 offset, u64 length,
321 u64 truncate_size, u32 truncate_seq);
322 extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req,
323 unsigned int which, u64 length);
324 extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req,
325 unsigned int which, u64 offset_inc);
327 extern struct ceph_osd_data *osd_req_op_extent_osd_data(
328 struct ceph_osd_request *osd_req,
331 extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *,
333 struct page **pages, u64 length,
334 u32 alignment, bool pages_from_pool,
336 extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *,
338 struct ceph_pagelist *pagelist);
340 extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *,
342 struct bio *bio, size_t bio_length);
343 #endif /* CONFIG_BLOCK */
345 extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *,
347 struct ceph_pagelist *pagelist);
348 extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *,
350 struct page **pages, u64 length,
351 u32 alignment, bool pages_from_pool,
353 extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *,
355 struct page **pages, u64 length,
356 u32 alignment, bool pages_from_pool,
358 extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req,
359 unsigned int which, u16 opcode,
360 const char *class, const char *method);
361 extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which,
362 u16 opcode, const char *name, const void *value,
363 size_t size, u8 cmp_op, u8 cmp_mode);
364 extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req,
366 u64 expected_object_size,
367 u64 expected_write_size);
369 extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc,
370 struct ceph_snap_context *snapc,
371 unsigned int num_ops,
374 int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp);
376 extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *,
377 struct ceph_file_layout *layout,
378 struct ceph_vino vino,
379 u64 offset, u64 *len,
380 unsigned int which, int num_ops,
381 int opcode, int flags,
382 struct ceph_snap_context *snapc,
383 u32 truncate_seq, u64 truncate_size,
386 extern void ceph_osdc_get_request(struct ceph_osd_request *req);
387 extern void ceph_osdc_put_request(struct ceph_osd_request *req);
389 extern int ceph_osdc_start_request(struct ceph_osd_client *osdc,
390 struct ceph_osd_request *req,
392 extern void ceph_osdc_cancel_request(struct ceph_osd_request *req);
393 extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc,
394 struct ceph_osd_request *req);
395 extern void ceph_osdc_sync(struct ceph_osd_client *osdc);
397 extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc);
398 void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc);
400 int ceph_osdc_call(struct ceph_osd_client *osdc,
401 struct ceph_object_id *oid,
402 struct ceph_object_locator *oloc,
403 const char *class, const char *method,
405 struct page *req_page, size_t req_len,
406 struct page *resp_page, size_t *resp_len);
408 extern int ceph_osdc_readpages(struct ceph_osd_client *osdc,
409 struct ceph_vino vino,
410 struct ceph_file_layout *layout,
412 u32 truncate_seq, u64 truncate_size,
413 struct page **pages, int nr_pages,
416 extern int ceph_osdc_writepages(struct ceph_osd_client *osdc,
417 struct ceph_vino vino,
418 struct ceph_file_layout *layout,
419 struct ceph_snap_context *sc,
421 u32 truncate_seq, u64 truncate_size,
422 struct timespec *mtime,
423 struct page **pages, int nr_pages);
426 struct ceph_osd_linger_request *
427 ceph_osdc_watch(struct ceph_osd_client *osdc,
428 struct ceph_object_id *oid,
429 struct ceph_object_locator *oloc,
430 rados_watchcb2_t wcb,
431 rados_watcherrcb_t errcb,
433 int ceph_osdc_unwatch(struct ceph_osd_client *osdc,
434 struct ceph_osd_linger_request *lreq);
436 int ceph_osdc_notify_ack(struct ceph_osd_client *osdc,
437 struct ceph_object_id *oid,
438 struct ceph_object_locator *oloc,
443 int ceph_osdc_notify(struct ceph_osd_client *osdc,
444 struct ceph_object_id *oid,
445 struct ceph_object_locator *oloc,
449 struct page ***preply_pages,
451 int ceph_osdc_watch_check(struct ceph_osd_client *osdc,
452 struct ceph_osd_linger_request *lreq);
453 int ceph_osdc_list_watchers(struct ceph_osd_client *osdc,
454 struct ceph_object_id *oid,
455 struct ceph_object_locator *oloc,
456 struct ceph_watch_item **watchers,