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> | |
12 | #include <linux/ceph/auth.h> | |
c885837f | 13 | #include <linux/ceph/pagelist.h> |
f24e9980 SW |
14 | |
15 | struct ceph_msg; | |
16 | struct ceph_snap_context; | |
17 | struct ceph_osd_request; | |
18 | struct ceph_osd_client; | |
19 | ||
20 | /* | |
21 | * completion callback for async writepages | |
22 | */ | |
85e084fe | 23 | typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *); |
26be8808 | 24 | typedef void (*ceph_osdc_unsafe_callback_t)(struct ceph_osd_request *, bool); |
f24e9980 | 25 | |
63244fa1 ID |
26 | #define CEPH_HOMELESS_OSD -1 |
27 | ||
f24e9980 SW |
28 | /* a given osd we're communicating with */ |
29 | struct ceph_osd { | |
30 | atomic_t o_ref; | |
31 | struct ceph_osd_client *o_osdc; | |
32 | int o_osd; | |
33 | int o_incarnation; | |
34 | struct rb_node o_node; | |
35 | struct ceph_connection o_con; | |
5aea3dcd | 36 | struct rb_root o_requests; |
922dab61 | 37 | struct rb_root o_linger_requests; |
f5a2041b | 38 | struct list_head o_osd_lru; |
6c4a1915 | 39 | struct ceph_auth_handshake o_auth; |
f5a2041b | 40 | unsigned long lru_ttl; |
422d2cb8 | 41 | struct list_head o_keepalive_item; |
5aea3dcd | 42 | struct mutex lock; |
f24e9980 SW |
43 | }; |
44 | ||
3f1af42a ID |
45 | #define CEPH_OSD_SLAB_OPS 2 |
46 | #define CEPH_OSD_MAX_OPS 16 | |
1b83bef2 | 47 | |
2ac2b7a6 | 48 | enum ceph_osd_data_type { |
ec9123c5 | 49 | CEPH_OSD_DATA_TYPE_NONE = 0, |
2ac2b7a6 | 50 | CEPH_OSD_DATA_TYPE_PAGES, |
9a5e6d09 | 51 | CEPH_OSD_DATA_TYPE_PAGELIST, |
2ac2b7a6 AE |
52 | #ifdef CONFIG_BLOCK |
53 | CEPH_OSD_DATA_TYPE_BIO, | |
54 | #endif /* CONFIG_BLOCK */ | |
55 | }; | |
56 | ||
2794a82a | 57 | struct ceph_osd_data { |
2ac2b7a6 AE |
58 | enum ceph_osd_data_type type; |
59 | union { | |
2794a82a AE |
60 | struct { |
61 | struct page **pages; | |
e0c59487 | 62 | u64 length; |
2794a82a AE |
63 | u32 alignment; |
64 | bool pages_from_pool; | |
65 | bool own_pages; | |
66 | }; | |
9a5e6d09 | 67 | struct ceph_pagelist *pagelist; |
2794a82a | 68 | #ifdef CONFIG_BLOCK |
fdce58cc AE |
69 | struct { |
70 | struct bio *bio; /* list of bios */ | |
71 | size_t bio_length; /* total in list */ | |
72 | }; | |
2794a82a AE |
73 | #endif /* CONFIG_BLOCK */ |
74 | }; | |
75 | }; | |
76 | ||
79528734 AE |
77 | struct ceph_osd_req_op { |
78 | u16 op; /* CEPH_OSD_OP_* */ | |
7b25bf5f | 79 | u32 flags; /* CEPH_OSD_OP_FLAG_* */ |
de2aa102 | 80 | u32 indata_len; /* request */ |
7665d85b YZ |
81 | u32 outdata_len; /* reply */ |
82 | s32 rval; | |
83 | ||
79528734 | 84 | union { |
49719778 | 85 | struct ceph_osd_data raw_data_in; |
79528734 AE |
86 | struct { |
87 | u64 offset, length; | |
88 | u64 truncate_size; | |
89 | u32 truncate_seq; | |
5476492f | 90 | struct ceph_osd_data osd_data; |
79528734 | 91 | } extent; |
d74b50be | 92 | struct { |
d7d5a007 ID |
93 | u32 name_len; |
94 | u32 value_len; | |
d74b50be YZ |
95 | __u8 cmp_op; /* CEPH_OSD_CMPXATTR_OP_* */ |
96 | __u8 cmp_mode; /* CEPH_OSD_CMPXATTR_MODE_* */ | |
97 | struct ceph_osd_data osd_data; | |
98 | } xattr; | |
79528734 AE |
99 | struct { |
100 | const char *class_name; | |
101 | const char *method_name; | |
5476492f | 102 | struct ceph_osd_data request_info; |
04017e29 | 103 | struct ceph_osd_data request_data; |
5476492f | 104 | struct ceph_osd_data response_data; |
79528734 AE |
105 | __u8 class_len; |
106 | __u8 method_len; | |
bb873b53 | 107 | u32 indata_len; |
79528734 AE |
108 | } cls; |
109 | struct { | |
110 | u64 cookie; | |
922dab61 ID |
111 | __u8 op; /* CEPH_OSD_WATCH_OP_ */ |
112 | u32 gen; | |
79528734 | 113 | } watch; |
922dab61 ID |
114 | struct { |
115 | struct ceph_osd_data request_data; | |
116 | } notify_ack; | |
19079203 ID |
117 | struct { |
118 | u64 cookie; | |
119 | struct ceph_osd_data request_data; | |
120 | struct ceph_osd_data response_data; | |
121 | } notify; | |
c647b8a8 ID |
122 | struct { |
123 | u64 expected_object_size; | |
124 | u64 expected_write_size; | |
125 | } alloc_hint; | |
79528734 AE |
126 | }; |
127 | }; | |
128 | ||
63244fa1 ID |
129 | struct ceph_osd_request_target { |
130 | struct ceph_object_id base_oid; | |
131 | struct ceph_object_locator base_oloc; | |
132 | struct ceph_object_id target_oid; | |
133 | struct ceph_object_locator target_oloc; | |
134 | ||
135 | struct ceph_pg pgid; | |
136 | u32 pg_num; | |
137 | u32 pg_num_mask; | |
138 | struct ceph_osds acting; | |
139 | struct ceph_osds up; | |
140 | int size; | |
141 | int min_size; | |
142 | bool sort_bitwise; | |
143 | ||
144 | unsigned int flags; /* CEPH_OSD_FLAG_* */ | |
145 | bool paused; | |
146 | ||
147 | int osd; | |
148 | }; | |
149 | ||
f24e9980 SW |
150 | /* an in-flight request */ |
151 | struct ceph_osd_request { | |
152 | u64 r_tid; /* unique for this client */ | |
153 | struct rb_node r_node; | |
4609245e | 154 | struct rb_node r_mc_node; /* map check */ |
f24e9980 | 155 | struct ceph_osd *r_osd; |
a66dd383 ID |
156 | |
157 | struct ceph_osd_request_target r_t; | |
158 | #define r_base_oid r_t.base_oid | |
159 | #define r_base_oloc r_t.base_oloc | |
160 | #define r_flags r_t.flags | |
f24e9980 SW |
161 | |
162 | struct ceph_msg *r_request, *r_reply; | |
f24e9980 | 163 | u32 r_sent; /* >0 if r_request is sending/sent */ |
1b83bef2 | 164 | |
79528734 AE |
165 | /* request osd ops array */ |
166 | unsigned int r_num_ops; | |
79528734 | 167 | |
1b83bef2 | 168 | int r_result; |
fe5da05e | 169 | bool r_got_reply; |
f24e9980 SW |
170 | |
171 | struct ceph_osd_client *r_osdc; | |
415e49a9 | 172 | struct kref r_kref; |
f24e9980 | 173 | bool r_mempool; |
fe5da05e ID |
174 | struct completion r_completion; |
175 | struct completion r_safe_completion; /* fsync waiter */ | |
26be8808 AE |
176 | ceph_osdc_callback_t r_callback; |
177 | ceph_osdc_unsafe_callback_t r_unsafe_callback; | |
f24e9980 SW |
178 | struct list_head r_unsafe_item; |
179 | ||
180 | struct inode *r_inode; /* for use by callbacks */ | |
3d14c5d2 | 181 | void *r_priv; /* ditto */ |
f24e9980 | 182 | |
bb873b53 ID |
183 | /* set by submitter */ |
184 | u64 r_snapid; /* for reads, CEPH_NOSNAP o/w */ | |
185 | struct ceph_snap_context *r_snapc; /* for writes */ | |
186 | struct timespec r_mtime; /* ditto */ | |
187 | u64 r_data_offset; /* ditto */ | |
922dab61 | 188 | bool r_linger; /* don't resend on failure */ |
f24e9980 | 189 | |
bb873b53 ID |
190 | /* internal */ |
191 | unsigned long r_stamp; /* jiffies, send or check time */ | |
192 | int r_attempts; | |
193 | struct ceph_eversion r_replay_version; /* aka reassert_version */ | |
194 | u32 r_last_force_resend; | |
4609245e | 195 | u32 r_map_dne_bound; |
3f1af42a ID |
196 | |
197 | struct ceph_osd_req_op r_ops[]; | |
f24e9980 SW |
198 | }; |
199 | ||
205ee118 ID |
200 | struct ceph_request_redirect { |
201 | struct ceph_object_locator oloc; | |
202 | }; | |
203 | ||
922dab61 ID |
204 | typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie, |
205 | u64 notifier_id, void *data, size_t data_len); | |
206 | typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err); | |
207 | ||
208 | struct ceph_osd_linger_request { | |
a40c4f10 | 209 | struct ceph_osd_client *osdc; |
922dab61 ID |
210 | u64 linger_id; |
211 | bool committed; | |
19079203 | 212 | bool is_watch; /* watch or notify */ |
922dab61 ID |
213 | |
214 | struct ceph_osd *osd; | |
215 | struct ceph_osd_request *reg_req; | |
216 | struct ceph_osd_request *ping_req; | |
217 | unsigned long ping_sent; | |
b07d3c4b ID |
218 | unsigned long watch_valid_thru; |
219 | struct list_head pending_lworks; | |
922dab61 ID |
220 | |
221 | struct ceph_osd_request_target t; | |
222 | u32 last_force_resend; | |
4609245e | 223 | u32 map_dne_bound; |
922dab61 ID |
224 | |
225 | struct timespec mtime; | |
226 | ||
a40c4f10 | 227 | struct kref kref; |
922dab61 ID |
228 | struct mutex lock; |
229 | struct rb_node node; /* osd */ | |
230 | struct rb_node osdc_node; /* osdc */ | |
4609245e | 231 | struct rb_node mc_node; /* map check */ |
922dab61 ID |
232 | struct list_head scan_item; |
233 | ||
234 | struct completion reg_commit_wait; | |
19079203 | 235 | struct completion notify_finish_wait; |
922dab61 | 236 | int reg_commit_error; |
19079203 | 237 | int notify_finish_error; |
922dab61 ID |
238 | int last_error; |
239 | ||
240 | u32 register_gen; | |
19079203 | 241 | u64 notify_id; |
a40c4f10 | 242 | |
922dab61 ID |
243 | rados_watchcb2_t wcb; |
244 | rados_watcherrcb_t errcb; | |
245 | void *data; | |
19079203 ID |
246 | |
247 | struct page ***preply_pages; | |
248 | size_t *preply_len; | |
a40c4f10 YS |
249 | }; |
250 | ||
f24e9980 SW |
251 | struct ceph_osd_client { |
252 | struct ceph_client *client; | |
253 | ||
254 | struct ceph_osdmap *osdmap; /* current map */ | |
5aea3dcd | 255 | struct rw_semaphore lock; |
f24e9980 | 256 | |
f24e9980 | 257 | struct rb_root osds; /* osds */ |
f5a2041b | 258 | struct list_head osd_lru; /* idle osds */ |
9dd2845c | 259 | spinlock_t osd_lru_lock; |
5aea3dcd ID |
260 | struct ceph_osd homeless_osd; |
261 | atomic64_t last_tid; /* tid of last request */ | |
922dab61 ID |
262 | u64 last_linger_id; |
263 | struct rb_root linger_requests; /* lingering requests */ | |
4609245e ID |
264 | struct rb_root map_checks; |
265 | struct rb_root linger_map_checks; | |
5aea3dcd ID |
266 | atomic_t num_requests; |
267 | atomic_t num_homeless; | |
f24e9980 | 268 | struct delayed_work timeout_work; |
f5a2041b | 269 | struct delayed_work osds_timeout_work; |
039934b8 | 270 | #ifdef CONFIG_DEBUG_FS |
f24e9980 | 271 | struct dentry *debugfs_file; |
039934b8 | 272 | #endif |
f24e9980 SW |
273 | |
274 | mempool_t *req_mempool; | |
275 | ||
0d59ab81 | 276 | struct ceph_msgpool msgpool_op; |
c16e7869 | 277 | struct ceph_msgpool msgpool_op_reply; |
a40c4f10 | 278 | |
a40c4f10 | 279 | struct workqueue_struct *notify_wq; |
f24e9980 SW |
280 | }; |
281 | ||
b7ec35b3 ID |
282 | static inline bool ceph_osdmap_flag(struct ceph_osd_client *osdc, int flag) |
283 | { | |
284 | return osdc->osdmap->flags & flag; | |
285 | } | |
286 | ||
5522ae0b AE |
287 | extern int ceph_osdc_setup(void); |
288 | extern void ceph_osdc_cleanup(void); | |
289 | ||
f24e9980 SW |
290 | extern int ceph_osdc_init(struct ceph_osd_client *osdc, |
291 | struct ceph_client *client); | |
292 | extern void ceph_osdc_stop(struct ceph_osd_client *osdc); | |
293 | ||
294 | extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc, | |
295 | struct ceph_msg *msg); | |
296 | extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc, | |
297 | struct ceph_msg *msg); | |
298 | ||
49719778 | 299 | extern void osd_req_op_init(struct ceph_osd_request *osd_req, |
144cba14 | 300 | unsigned int which, u16 opcode, u32 flags); |
49719778 AE |
301 | |
302 | extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *, | |
303 | unsigned int which, | |
304 | struct page **pages, u64 length, | |
305 | u32 alignment, bool pages_from_pool, | |
306 | bool own_pages); | |
307 | ||
c99d2d4a AE |
308 | extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req, |
309 | unsigned int which, u16 opcode, | |
33803f33 AE |
310 | u64 offset, u64 length, |
311 | u64 truncate_size, u32 truncate_seq); | |
c99d2d4a AE |
312 | extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req, |
313 | unsigned int which, u64 length); | |
2c63f49a YZ |
314 | extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, |
315 | unsigned int which, u64 offset_inc); | |
a4ce40a9 AE |
316 | |
317 | extern struct ceph_osd_data *osd_req_op_extent_osd_data( | |
318 | struct ceph_osd_request *osd_req, | |
406e2c9f | 319 | unsigned int which); |
a4ce40a9 AE |
320 | |
321 | extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *, | |
406e2c9f | 322 | unsigned int which, |
a4ce40a9 AE |
323 | struct page **pages, u64 length, |
324 | u32 alignment, bool pages_from_pool, | |
325 | bool own_pages); | |
326 | extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *, | |
406e2c9f | 327 | unsigned int which, |
a4ce40a9 AE |
328 | struct ceph_pagelist *pagelist); |
329 | #ifdef CONFIG_BLOCK | |
330 | extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *, | |
406e2c9f | 331 | unsigned int which, |
a4ce40a9 AE |
332 | struct bio *bio, size_t bio_length); |
333 | #endif /* CONFIG_BLOCK */ | |
334 | ||
04017e29 AE |
335 | extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *, |
336 | unsigned int which, | |
337 | struct ceph_pagelist *pagelist); | |
6c57b554 AE |
338 | extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *, |
339 | unsigned int which, | |
340 | struct page **pages, u64 length, | |
341 | u32 alignment, bool pages_from_pool, | |
342 | bool own_pages); | |
a4ce40a9 | 343 | extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *, |
c99d2d4a | 344 | unsigned int which, |
a4ce40a9 AE |
345 | struct page **pages, u64 length, |
346 | u32 alignment, bool pages_from_pool, | |
347 | bool own_pages); | |
348 | ||
c99d2d4a AE |
349 | extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req, |
350 | unsigned int which, u16 opcode, | |
04017e29 | 351 | const char *class, const char *method); |
d74b50be YZ |
352 | extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, |
353 | u16 opcode, const char *name, const void *value, | |
354 | size_t size, u8 cmp_op, u8 cmp_mode); | |
c647b8a8 ID |
355 | extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, |
356 | unsigned int which, | |
357 | u64 expected_object_size, | |
358 | u64 expected_write_size); | |
33803f33 | 359 | |
3499e8a5 | 360 | extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, |
3499e8a5 | 361 | struct ceph_snap_context *snapc, |
acead002 | 362 | unsigned int num_ops, |
3499e8a5 | 363 | bool use_mempool, |
54a54007 | 364 | gfp_t gfp_flags); |
13d1ad16 | 365 | int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp); |
3499e8a5 | 366 | |
f24e9980 SW |
367 | extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *, |
368 | struct ceph_file_layout *layout, | |
369 | struct ceph_vino vino, | |
acead002 | 370 | u64 offset, u64 *len, |
715e4cd4 YZ |
371 | unsigned int which, int num_ops, |
372 | int opcode, int flags, | |
f24e9980 | 373 | struct ceph_snap_context *snapc, |
acead002 | 374 | u32 truncate_seq, u64 truncate_size, |
153e5167 | 375 | bool use_mempool); |
f24e9980 | 376 | |
9e94af20 ID |
377 | extern void ceph_osdc_get_request(struct ceph_osd_request *req); |
378 | extern void ceph_osdc_put_request(struct ceph_osd_request *req); | |
f24e9980 SW |
379 | |
380 | extern int ceph_osdc_start_request(struct ceph_osd_client *osdc, | |
381 | struct ceph_osd_request *req, | |
382 | bool nofail); | |
c9f9b93d | 383 | extern void ceph_osdc_cancel_request(struct ceph_osd_request *req); |
f24e9980 SW |
384 | extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc, |
385 | struct ceph_osd_request *req); | |
386 | extern void ceph_osdc_sync(struct ceph_osd_client *osdc); | |
387 | ||
dd935f44 | 388 | extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc); |
7cca78c9 | 389 | void ceph_osdc_maybe_request_map(struct ceph_osd_client *osdc); |
dd935f44 | 390 | |
f24e9980 SW |
391 | extern int ceph_osdc_readpages(struct ceph_osd_client *osdc, |
392 | struct ceph_vino vino, | |
393 | struct ceph_file_layout *layout, | |
394 | u64 off, u64 *plen, | |
395 | u32 truncate_seq, u64 truncate_size, | |
b7495fc2 SW |
396 | struct page **pages, int nr_pages, |
397 | int page_align); | |
f24e9980 SW |
398 | |
399 | extern int ceph_osdc_writepages(struct ceph_osd_client *osdc, | |
400 | struct ceph_vino vino, | |
401 | struct ceph_file_layout *layout, | |
402 | struct ceph_snap_context *sc, | |
403 | u64 off, u64 len, | |
404 | u32 truncate_seq, u64 truncate_size, | |
405 | struct timespec *mtime, | |
24808826 | 406 | struct page **pages, int nr_pages); |
f24e9980 | 407 | |
922dab61 ID |
408 | /* watch/notify */ |
409 | struct ceph_osd_linger_request * | |
410 | ceph_osdc_watch(struct ceph_osd_client *osdc, | |
411 | struct ceph_object_id *oid, | |
412 | struct ceph_object_locator *oloc, | |
413 | rados_watchcb2_t wcb, | |
414 | rados_watcherrcb_t errcb, | |
415 | void *data); | |
416 | int ceph_osdc_unwatch(struct ceph_osd_client *osdc, | |
417 | struct ceph_osd_linger_request *lreq); | |
418 | ||
419 | int ceph_osdc_notify_ack(struct ceph_osd_client *osdc, | |
420 | struct ceph_object_id *oid, | |
421 | struct ceph_object_locator *oloc, | |
422 | u64 notify_id, | |
423 | u64 cookie, | |
424 | void *payload, | |
425 | size_t payload_len); | |
19079203 ID |
426 | int ceph_osdc_notify(struct ceph_osd_client *osdc, |
427 | struct ceph_object_id *oid, | |
428 | struct ceph_object_locator *oloc, | |
429 | void *payload, | |
430 | size_t payload_len, | |
431 | u32 timeout, | |
432 | struct page ***preply_pages, | |
433 | size_t *preply_len); | |
b07d3c4b ID |
434 | int ceph_osdc_watch_check(struct ceph_osd_client *osdc, |
435 | struct ceph_osd_linger_request *lreq); | |
f24e9980 SW |
436 | #endif |
437 |