Commit | Line | Data |
---|---|---|
f24e9980 SW |
1 | #ifndef _FS_CEPH_OSD_CLIENT_H |
2 | #define _FS_CEPH_OSD_CLIENT_H | |
3 | ||
4 | #include <linux/completion.h> | |
415e49a9 | 5 | #include <linux/kref.h> |
f24e9980 SW |
6 | #include <linux/mempool.h> |
7 | #include <linux/rbtree.h> | |
8 | ||
6c4a1915 AE |
9 | #include <linux/ceph/types.h> |
10 | #include <linux/ceph/osdmap.h> | |
11 | #include <linux/ceph/messenger.h> | |
b2aa5d0b | 12 | #include <linux/ceph/msgpool.h> |
6c4a1915 | 13 | #include <linux/ceph/auth.h> |
c885837f | 14 | #include <linux/ceph/pagelist.h> |
f24e9980 SW |
15 | |
16 | struct ceph_msg; | |
17 | struct ceph_snap_context; | |
18 | struct ceph_osd_request; | |
19 | struct ceph_osd_client; | |
20 | ||
21 | /* | |
22 | * completion callback for async writepages | |
23 | */ | |
85e084fe | 24 | typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *); |
26be8808 | 25 | typedef void (*ceph_osdc_unsafe_callback_t)(struct ceph_osd_request *, bool); |
f24e9980 | 26 | |
63244fa1 ID |
27 | #define CEPH_HOMELESS_OSD -1 |
28 | ||
f24e9980 SW |
29 | /* a given osd we're communicating with */ |
30 | struct ceph_osd { | |
31 | atomic_t o_ref; | |
32 | struct ceph_osd_client *o_osdc; | |
33 | int o_osd; | |
34 | int o_incarnation; | |
35 | struct rb_node o_node; | |
36 | struct ceph_connection o_con; | |
5aea3dcd | 37 | struct rb_root o_requests; |
922dab61 | 38 | struct rb_root o_linger_requests; |
f5a2041b | 39 | struct list_head o_osd_lru; |
6c4a1915 | 40 | struct ceph_auth_handshake o_auth; |
f5a2041b | 41 | unsigned long lru_ttl; |
422d2cb8 | 42 | struct list_head o_keepalive_item; |
5aea3dcd | 43 | struct mutex lock; |
f24e9980 SW |
44 | }; |
45 | ||
3f1af42a ID |
46 | #define CEPH_OSD_SLAB_OPS 2 |
47 | #define CEPH_OSD_MAX_OPS 16 | |
1b83bef2 | 48 | |
2ac2b7a6 | 49 | enum ceph_osd_data_type { |
ec9123c5 | 50 | CEPH_OSD_DATA_TYPE_NONE = 0, |
2ac2b7a6 | 51 | CEPH_OSD_DATA_TYPE_PAGES, |
9a5e6d09 | 52 | CEPH_OSD_DATA_TYPE_PAGELIST, |
2ac2b7a6 AE |
53 | #ifdef CONFIG_BLOCK |
54 | CEPH_OSD_DATA_TYPE_BIO, | |
55 | #endif /* CONFIG_BLOCK */ | |
56 | }; | |
57 | ||
2794a82a | 58 | struct ceph_osd_data { |
2ac2b7a6 AE |
59 | enum ceph_osd_data_type type; |
60 | union { | |
2794a82a AE |
61 | struct { |
62 | struct page **pages; | |
e0c59487 | 63 | u64 length; |
2794a82a AE |
64 | u32 alignment; |
65 | bool pages_from_pool; | |
66 | bool own_pages; | |
67 | }; | |
9a5e6d09 | 68 | struct ceph_pagelist *pagelist; |
2794a82a | 69 | #ifdef CONFIG_BLOCK |
fdce58cc AE |
70 | struct { |
71 | struct bio *bio; /* list of bios */ | |
72 | size_t bio_length; /* total in list */ | |
73 | }; | |
2794a82a AE |
74 | #endif /* CONFIG_BLOCK */ |
75 | }; | |
76 | }; | |
77 | ||
79528734 AE |
78 | struct ceph_osd_req_op { |
79 | u16 op; /* CEPH_OSD_OP_* */ | |
7b25bf5f | 80 | u32 flags; /* CEPH_OSD_OP_FLAG_* */ |
de2aa102 | 81 | u32 indata_len; /* request */ |
7665d85b YZ |
82 | u32 outdata_len; /* reply */ |
83 | s32 rval; | |
84 | ||
79528734 | 85 | union { |
49719778 | 86 | struct ceph_osd_data raw_data_in; |
79528734 AE |
87 | struct { |
88 | u64 offset, length; | |
89 | u64 truncate_size; | |
90 | u32 truncate_seq; | |
5476492f | 91 | struct ceph_osd_data osd_data; |
79528734 | 92 | } extent; |
d74b50be | 93 | struct { |
d7d5a007 ID |
94 | u32 name_len; |
95 | u32 value_len; | |
d74b50be YZ |
96 | __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */ |
97 | __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */ | |
98 | struct ceph_osd_data osd_data; | |
99 | } xattr; | |
79528734 AE |
100 | struct { |
101 | const char *class_name; | |
102 | const char *method_name; | |
5476492f | 103 | struct ceph_osd_data request_info; |
04017e29 | 104 | struct ceph_osd_data request_data; |
5476492f | 105 | struct ceph_osd_data response_data; |
79528734 AE |
106 | __u8 class_len; |
107 | __u8 method_len; | |
bb873b53 | 108 | u32 indata_len; |
79528734 AE |
109 | } cls; |
110 | struct { | |
111 | u64 cookie; | |
922dab61 ID |
112 | __u8 op; /* CEPH_OSD_WATCH_OP_ */ |
113 | u32 gen; | |
79528734 | 114 | } watch; |
922dab61 ID |
115 | struct { |
116 | struct ceph_osd_data request_data; | |
117 | } notify_ack; | |
19079203 ID |
118 | struct { |
119 | u64 cookie; | |
120 | struct ceph_osd_data request_data; | |
121 | struct ceph_osd_data response_data; | |
122 | } notify; | |
a4ed38d7 DF |
123 | struct { |
124 | struct ceph_osd_data response_data; | |
125 | } list_watchers; | |
c647b8a8 ID |
126 | struct { |
127 | u64 expected_object_size; | |
128 | u64 expected_write_size; | |
129 | } alloc_hint; | |
79528734 AE |
130 | }; |
131 | }; | |
132 | ||
63244fa1 ID |
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; | |
138 | ||
139 | struct ceph_pg pgid; | |
140 | u32 pg_num; | |
141 | u32 pg_num_mask; | |
142 | struct ceph_osds acting; | |
143 | struct ceph_osds up; | |
144 | int size; | |
145 | int min_size; | |
146 | bool sort_bitwise; | |
147 | ||
148 | unsigned int flags; /* CEPH_OSD_FLAG_* */ | |
149 | bool paused; | |
150 | ||
151 | int osd; | |
152 | }; | |
153 | ||
f24e9980 SW |
154 | /* an in-flight request */ |
155 | struct ceph_osd_request { | |
156 | u64 r_tid; /* unique for this client */ | |
157 | struct rb_node r_node; | |
4609245e | 158 | struct rb_node r_mc_node; /* map check */ |
f24e9980 | 159 | struct ceph_osd *r_osd; |
a66dd383 ID |
160 | |
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 | |
f24e9980 SW |
165 | |
166 | struct ceph_msg *r_request, *r_reply; | |
f24e9980 | 167 | u32 r_sent; /* >0 if r_request is sending/sent */ |
1b83bef2 | 168 | |
79528734 AE |
169 | /* request osd ops array */ |
170 | unsigned int r_num_ops; | |
79528734 | 171 | |
1b83bef2 | 172 | int r_result; |
fe5da05e | 173 | bool r_got_reply; |
f24e9980 SW |
174 | |
175 | struct ceph_osd_client *r_osdc; | |
415e49a9 | 176 | struct kref r_kref; |
f24e9980 | 177 | bool r_mempool; |
fe5da05e ID |
178 | struct completion r_completion; |
179 | struct completion r_safe_completion; /* fsync waiter */ | |
26be8808 AE |
180 | ceph_osdc_callback_t r_callback; |
181 | ceph_osdc_unsafe_callback_t r_unsafe_callback; | |
f24e9980 SW |
182 | struct list_head r_unsafe_item; |
183 | ||
184 | struct inode *r_inode; /* for use by callbacks */ | |
3d14c5d2 | 185 | void *r_priv; /* ditto */ |
f24e9980 | 186 | |
bb873b53 ID |
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 */ | |
922dab61 | 192 | bool r_linger; /* don't resend on failure */ |
f24e9980 | 193 | |
bb873b53 ID |
194 | /* internal */ |
195 | unsigned long r_stamp; /* jiffies, send or check time */ | |
196 | int r_attempts; | |
197 | struct ceph_eversion r_replay_version; /* aka reassert_version */ | |
198 | u32 r_last_force_resend; | |
4609245e | 199 | u32 r_map_dne_bound; |
3f1af42a ID |
200 | |
201 | struct ceph_osd_req_op r_ops[]; | |
f24e9980 SW |
202 | }; |
203 | ||
205ee118 ID |
204 | struct ceph_request_redirect { |
205 | struct ceph_object_locator oloc; | |
206 | }; | |
207 | ||
922dab61 ID |
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); | |
211 | ||
212 | struct ceph_osd_linger_request { | |
a40c4f10 | 213 | struct ceph_osd_client *osdc; |
922dab61 ID |
214 | u64 linger_id; |
215 | bool committed; | |
19079203 | 216 | bool is_watch; /* watch or notify */ |
922dab61 ID |
217 | |
218 | struct ceph_osd *osd; | |
219 | struct ceph_osd_request *reg_req; | |
220 | struct ceph_osd_request *ping_req; | |
221 | unsigned long ping_sent; | |
b07d3c4b ID |
222 | unsigned long watch_valid_thru; |
223 | struct list_head pending_lworks; | |
922dab61 ID |
224 | |
225 | struct ceph_osd_request_target t; | |
226 | u32 last_force_resend; | |
4609245e | 227 | u32 map_dne_bound; |
922dab61 ID |
228 | |
229 | struct timespec mtime; | |
230 | ||
a40c4f10 | 231 | struct kref kref; |
922dab61 ID |
232 | struct mutex lock; |
233 | struct rb_node node; /* osd */ | |
234 | struct rb_node osdc_node; /* osdc */ | |
4609245e | 235 | struct rb_node mc_node; /* map check */ |
922dab61 ID |
236 | struct list_head scan_item; |
237 | ||
238 | struct completion reg_commit_wait; | |
19079203 | 239 | struct completion notify_finish_wait; |
922dab61 | 240 | int reg_commit_error; |
19079203 | 241 | int notify_finish_error; |
922dab61 ID |
242 | int last_error; |
243 | ||
244 | u32 register_gen; | |
19079203 | 245 | u64 notify_id; |
a40c4f10 | 246 | |
922dab61 ID |
247 | rados_watchcb2_t wcb; |
248 | rados_watcherrcb_t errcb; | |
249 | void *data; | |
19079203 ID |
250 | |
251 | struct page ***preply_pages; | |
252 | size_t *preply_len; | |
a40c4f10 YS |
253 | }; |
254 | ||
a4ed38d7 DF |
255 | struct ceph_watch_item { |
256 | struct ceph_entity_name name; | |
257 | u64 cookie; | |
258 | struct ceph_entity_addr addr; | |
259 | }; | |
260 | ||
f24e9980 SW |
261 | struct ceph_osd_client { |
262 | struct ceph_client *client; | |
263 | ||
264 | struct ceph_osdmap *osdmap; /* current map */ | |
5aea3dcd | 265 | struct rw_semaphore lock; |
f24e9980 | 266 | |
f24e9980 | 267 | struct rb_root osds; /* osds */ |
f5a2041b | 268 | struct list_head osd_lru; /* idle osds */ |
9dd2845c | 269 | spinlock_t osd_lru_lock; |
5aea3dcd ID |
270 | struct ceph_osd homeless_osd; |
271 | atomic64_t last_tid; /* tid of last request */ | |
922dab61 ID |
272 | u64 last_linger_id; |
273 | struct rb_root linger_requests; /* lingering requests */ | |
4609245e ID |
274 | struct rb_root map_checks; |
275 | struct rb_root linger_map_checks; | |
5aea3dcd ID |
276 | atomic_t num_requests; |
277 | atomic_t num_homeless; | |
f24e9980 | 278 | struct delayed_work timeout_work; |
f5a2041b | 279 | struct delayed_work osds_timeout_work; |
039934b8 | 280 | #ifdef CONFIG_DEBUG_FS |
f24e9980 | 281 | struct dentry *debugfs_file; |
039934b8 | 282 | #endif |
f24e9980 SW |
283 | |
284 | mempool_t *req_mempool; | |
285 | ||
0d59ab81 | 286 | struct ceph_msgpool msgpool_op; |
c16e7869 | 287 | struct ceph_msgpool msgpool_op_reply; |
a40c4f10 | 288 | |
a40c4f10 | 289 | struct workqueue_struct *notify_wq; |
f24e9980 SW |
290 | }; |
291 | ||
b7ec35b3 ID |
292 | static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag) |
293 | { | |
294 | return osdc->osdmap->flags & flag; | |
295 | } | |
296 | ||
5522ae0b AE |
297 | extern int ceph_osdc_setup(void); |
298 | extern void ceph_osdc_cleanup(void); | |
299 | ||
f24e9980 SW |
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); | |
303 | ||
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); | |
308 | ||
49719778 | 309 | extern void osd_req_op_init(struct ceph_osd_request *osd_req, |
144cba14 | 310 | unsigned int which, u16 opcode, u32 flags); |
49719778 AE |
311 | |
312 | extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *, | |
313 | unsigned int which, | |
314 | struct page **pages, u64 length, | |
315 | u32 alignment, bool pages_from_pool, | |
316 | bool own_pages); | |
317 | ||
c99d2d4a AE |
318 | extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req, |
319 | unsigned int which, u16 opcode, | |
33803f33 AE |
320 | u64 offset, u64 length, |
321 | u64 truncate_size, u32 truncate_seq); | |
c99d2d4a AE |
322 | extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req, |
323 | unsigned int which, u64 length); | |
2c63f49a YZ |
324 | extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, |
325 | unsigned int which, u64 offset_inc); | |
a4ce40a9 AE |
326 | |
327 | extern struct ceph_osd_data *osd_req_op_extent_osd_data( | |
328 | struct ceph_osd_request *osd_req, | |
406e2c9f | 329 | unsigned int which); |
a4ce40a9 AE |
330 | |
331 | extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *, | |
406e2c9f | 332 | unsigned int which, |
a4ce40a9 AE |
333 | struct page **pages, u64 length, |
334 | u32 alignment, bool pages_from_pool, | |
335 | bool own_pages); | |
336 | extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *, | |
406e2c9f | 337 | unsigned int which, |
a4ce40a9 AE |
338 | struct ceph_pagelist *pagelist); |
339 | #ifdef CONFIG_BLOCK | |
340 | extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *, | |
406e2c9f | 341 | unsigned int which, |
a4ce40a9 AE |
342 | struct bio *bio, size_t bio_length); |
343 | #endif /* CONFIG_BLOCK */ | |
344 | ||
04017e29 AE |
345 | extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *, |
346 | unsigned int which, | |
347 | struct ceph_pagelist *pagelist); | |
6c57b554 AE |
348 | extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *, |
349 | unsigned int which, | |
350 | struct page **pages, u64 length, | |
351 | u32 alignment, bool pages_from_pool, | |
352 | bool own_pages); | |
a4ce40a9 | 353 | extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *, |
c99d2d4a | 354 | unsigned int which, |
a4ce40a9 AE |
355 | struct page **pages, u64 length, |
356 | u32 alignment, bool pages_from_pool, | |
357 | bool own_pages); | |
c99d2d4a AE |
358 | extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req, |
359 | unsigned int which, u16 opcode, | |
04017e29 | 360 | const char *class, const char *method); |
d74b50be YZ |
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); | |
c647b8a8 ID |
364 | extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, |
365 | unsigned int which, | |
366 | u64 expected_object_size, | |
367 | u64 expected_write_size); | |
33803f33 | 368 | |
3499e8a5 | 369 | extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, |
3499e8a5 | 370 | struct ceph_snap_context *snapc, |
acead002 | 371 | unsigned int num_ops, |
3499e8a5 | 372 | bool use_mempool, |
54a54007 | 373 | gfp_t gfp_flags); |
13d1ad16 | 374 | int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp); |
3499e8a5 | 375 | |
f24e9980 SW |
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, | |
acead002 | 379 | u64 offset, u64 *len, |
715e4cd4 YZ |
380 | unsigned int which, int num_ops, |
381 | int opcode, int flags, | |
f24e9980 | 382 | struct ceph_snap_context *snapc, |
acead002 | 383 | u32 truncate_seq, u64 truncate_size, |
153e5167 | 384 | bool use_mempool); |
f24e9980 | 385 | |
9e94af20 ID |
386 | extern void ceph_osdc_get_request(struct ceph_osd_request *req); |
387 | extern void ceph_osdc_put_request(struct ceph_osd_request *req); | |
f24e9980 SW |
388 | |
389 | extern int ceph_osdc_start_request(struct ceph_osd_client *osdc, | |
390 | struct ceph_osd_request *req, | |
391 | bool nofail); | |
c9f9b93d | 392 | extern void ceph_osdc_cancel_request(struct ceph_osd_request *req); |
f24e9980 SW |
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); | |
396 | ||
dd935f44 | 397 | extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc); |
7cca78c9 | 398 | void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc); |
dd935f44 | 399 | |
428a7158 DF |
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, | |
404 | unsigned int flags, | |
405 | struct page *req_page, size_t req_len, | |
406 | struct page *resp_page, size_t *resp_len); | |
407 | ||
f24e9980 SW |
408 | extern int ceph_osdc_readpages(struct ceph_osd_client *osdc, |
409 | struct ceph_vino vino, | |
410 | struct ceph_file_layout *layout, | |
411 | u64 off, u64 *plen, | |
412 | u32 truncate_seq, u64 truncate_size, | |
b7495fc2 SW |
413 | struct page **pages, int nr_pages, |
414 | int page_align); | |
f24e9980 SW |
415 | |
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, | |
420 | u64 off, u64 len, | |
421 | u32 truncate_seq, u64 truncate_size, | |
422 | struct timespec *mtime, | |
24808826 | 423 | struct page **pages, int nr_pages); |
f24e9980 | 424 | |
922dab61 ID |
425 | /* watch/notify */ |
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, | |
432 | void *data); | |
433 | int ceph_osdc_unwatch(struct ceph_osd_client *osdc, | |
434 | struct ceph_osd_linger_request *lreq); | |
435 | ||
436 | int ceph_osdc_notify_ack(struct ceph_osd_client *osdc, | |
437 | struct ceph_object_id *oid, | |
438 | struct ceph_object_locator *oloc, | |
439 | u64 notify_id, | |
440 | u64 cookie, | |
441 | void *payload, | |
442 | size_t payload_len); | |
19079203 ID |
443 | int ceph_osdc_notify(struct ceph_osd_client *osdc, |
444 | struct ceph_object_id *oid, | |
445 | struct ceph_object_locator *oloc, | |
446 | void *payload, | |
447 | size_t payload_len, | |
448 | u32 timeout, | |
449 | struct page ***preply_pages, | |
450 | size_t *preply_len); | |
b07d3c4b ID |
451 | int ceph_osdc_watch_check(struct ceph_osd_client *osdc, |
452 | struct ceph_osd_linger_request *lreq); | |
a4ed38d7 DF |
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, | |
457 | u32 *num_watchers); | |
f24e9980 SW |
458 | #endif |
459 |