]>
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; | |
c647b8a8 ID |
117 | struct { |
118 | u64 expected_object_size; | |
119 | u64 expected_write_size; | |
120 | } alloc_hint; | |
79528734 AE |
121 | }; |
122 | }; | |
123 | ||
63244fa1 ID |
124 | struct ceph_osd_request_target { |
125 | struct ceph_object_id base_oid; | |
126 | struct ceph_object_locator base_oloc; | |
127 | struct ceph_object_id target_oid; | |
128 | struct ceph_object_locator target_oloc; | |
129 | ||
130 | struct ceph_pg pgid; | |
131 | u32 pg_num; | |
132 | u32 pg_num_mask; | |
133 | struct ceph_osds acting; | |
134 | struct ceph_osds up; | |
135 | int size; | |
136 | int min_size; | |
137 | bool sort_bitwise; | |
138 | ||
139 | unsigned int flags; /* CEPH_OSD_FLAG_* */ | |
140 | bool paused; | |
141 | ||
142 | int osd; | |
143 | }; | |
144 | ||
f24e9980 SW |
145 | /* an in-flight request */ |
146 | struct ceph_osd_request { | |
147 | u64 r_tid; /* unique for this client */ | |
148 | struct rb_node r_node; | |
f24e9980 | 149 | struct ceph_osd *r_osd; |
a66dd383 ID |
150 | |
151 | struct ceph_osd_request_target r_t; | |
152 | #define r_base_oid r_t.base_oid | |
153 | #define r_base_oloc r_t.base_oloc | |
154 | #define r_flags r_t.flags | |
f24e9980 SW |
155 | |
156 | struct ceph_msg *r_request, *r_reply; | |
f24e9980 | 157 | u32 r_sent; /* >0 if r_request is sending/sent */ |
1b83bef2 | 158 | |
79528734 AE |
159 | /* request osd ops array */ |
160 | unsigned int r_num_ops; | |
79528734 | 161 | |
1b83bef2 | 162 | int r_result; |
fe5da05e | 163 | bool r_got_reply; |
f24e9980 SW |
164 | |
165 | struct ceph_osd_client *r_osdc; | |
415e49a9 | 166 | struct kref r_kref; |
f24e9980 | 167 | bool r_mempool; |
fe5da05e ID |
168 | struct completion r_completion; |
169 | struct completion r_safe_completion; /* fsync waiter */ | |
26be8808 AE |
170 | ceph_osdc_callback_t r_callback; |
171 | ceph_osdc_unsafe_callback_t r_unsafe_callback; | |
f24e9980 SW |
172 | struct list_head r_unsafe_item; |
173 | ||
174 | struct inode *r_inode; /* for use by callbacks */ | |
3d14c5d2 | 175 | void *r_priv; /* ditto */ |
f24e9980 | 176 | |
bb873b53 ID |
177 | /* set by submitter */ |
178 | u64 r_snapid; /* for reads, CEPH_NOSNAP o/w */ | |
179 | struct ceph_snap_context *r_snapc; /* for writes */ | |
180 | struct timespec r_mtime; /* ditto */ | |
181 | u64 r_data_offset; /* ditto */ | |
922dab61 | 182 | bool r_linger; /* don't resend on failure */ |
f24e9980 | 183 | |
bb873b53 ID |
184 | /* internal */ |
185 | unsigned long r_stamp; /* jiffies, send or check time */ | |
186 | int r_attempts; | |
187 | struct ceph_eversion r_replay_version; /* aka reassert_version */ | |
188 | u32 r_last_force_resend; | |
3f1af42a ID |
189 | |
190 | struct ceph_osd_req_op r_ops[]; | |
f24e9980 SW |
191 | }; |
192 | ||
205ee118 ID |
193 | struct ceph_request_redirect { |
194 | struct ceph_object_locator oloc; | |
195 | }; | |
196 | ||
922dab61 ID |
197 | typedef void (*rados_watchcb2_t)(void *arg, u64 notify_id, u64 cookie, |
198 | u64 notifier_id, void *data, size_t data_len); | |
199 | typedef void (*rados_watcherrcb_t)(void *arg, u64 cookie, int err); | |
200 | ||
201 | struct ceph_osd_linger_request { | |
a40c4f10 | 202 | struct ceph_osd_client *osdc; |
922dab61 ID |
203 | u64 linger_id; |
204 | bool committed; | |
205 | ||
206 | struct ceph_osd *osd; | |
207 | struct ceph_osd_request *reg_req; | |
208 | struct ceph_osd_request *ping_req; | |
209 | unsigned long ping_sent; | |
210 | ||
211 | struct ceph_osd_request_target t; | |
212 | u32 last_force_resend; | |
213 | ||
214 | struct timespec mtime; | |
215 | ||
a40c4f10 | 216 | struct kref kref; |
922dab61 ID |
217 | struct mutex lock; |
218 | struct rb_node node; /* osd */ | |
219 | struct rb_node osdc_node; /* osdc */ | |
220 | struct list_head scan_item; | |
221 | ||
222 | struct completion reg_commit_wait; | |
223 | int reg_commit_error; | |
224 | int last_error; | |
225 | ||
226 | u32 register_gen; | |
a40c4f10 | 227 | |
922dab61 ID |
228 | rados_watchcb2_t wcb; |
229 | rados_watcherrcb_t errcb; | |
230 | void *data; | |
a40c4f10 YS |
231 | }; |
232 | ||
f24e9980 SW |
233 | struct ceph_osd_client { |
234 | struct ceph_client *client; | |
235 | ||
236 | struct ceph_osdmap *osdmap; /* current map */ | |
5aea3dcd | 237 | struct rw_semaphore lock; |
f24e9980 | 238 | |
f24e9980 | 239 | struct rb_root osds; /* osds */ |
f5a2041b | 240 | struct list_head osd_lru; /* idle osds */ |
9dd2845c | 241 | spinlock_t osd_lru_lock; |
5aea3dcd ID |
242 | struct ceph_osd homeless_osd; |
243 | atomic64_t last_tid; /* tid of last request */ | |
922dab61 ID |
244 | u64 last_linger_id; |
245 | struct rb_root linger_requests; /* lingering requests */ | |
5aea3dcd ID |
246 | atomic_t num_requests; |
247 | atomic_t num_homeless; | |
f24e9980 | 248 | struct delayed_work timeout_work; |
f5a2041b | 249 | struct delayed_work osds_timeout_work; |
039934b8 | 250 | #ifdef CONFIG_DEBUG_FS |
f24e9980 | 251 | struct dentry *debugfs_file; |
039934b8 | 252 | #endif |
f24e9980 SW |
253 | |
254 | mempool_t *req_mempool; | |
255 | ||
0d59ab81 | 256 | struct ceph_msgpool msgpool_op; |
c16e7869 | 257 | struct ceph_msgpool msgpool_op_reply; |
a40c4f10 | 258 | |
a40c4f10 | 259 | struct workqueue_struct *notify_wq; |
f24e9980 SW |
260 | }; |
261 | ||
5522ae0b AE |
262 | extern int ceph_osdc_setup(void); |
263 | extern void ceph_osdc_cleanup(void); | |
264 | ||
f24e9980 SW |
265 | extern int ceph_osdc_init(struct ceph_osd_client *osdc, |
266 | struct ceph_client *client); | |
267 | extern void ceph_osdc_stop(struct ceph_osd_client *osdc); | |
268 | ||
269 | extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc, | |
270 | struct ceph_msg *msg); | |
271 | extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc, | |
272 | struct ceph_msg *msg); | |
273 | ||
49719778 | 274 | extern void osd_req_op_init(struct ceph_osd_request *osd_req, |
144cba14 | 275 | unsigned int which, u16 opcode, u32 flags); |
49719778 AE |
276 | |
277 | extern void osd_req_op_raw_data_in_pages(struct ceph_osd_request *, | |
278 | unsigned int which, | |
279 | struct page **pages, u64 length, | |
280 | u32 alignment, bool pages_from_pool, | |
281 | bool own_pages); | |
282 | ||
c99d2d4a AE |
283 | extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req, |
284 | unsigned int which, u16 opcode, | |
33803f33 AE |
285 | u64 offset, u64 length, |
286 | u64 truncate_size, u32 truncate_seq); | |
c99d2d4a AE |
287 | extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req, |
288 | unsigned int which, u64 length); | |
2c63f49a YZ |
289 | extern void osd_req_op_extent_dup_last(struct ceph_osd_request *osd_req, |
290 | unsigned int which, u64 offset_inc); | |
a4ce40a9 AE |
291 | |
292 | extern struct ceph_osd_data *osd_req_op_extent_osd_data( | |
293 | struct ceph_osd_request *osd_req, | |
406e2c9f | 294 | unsigned int which); |
a4ce40a9 AE |
295 | |
296 | extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *, | |
406e2c9f | 297 | unsigned int which, |
a4ce40a9 AE |
298 | struct page **pages, u64 length, |
299 | u32 alignment, bool pages_from_pool, | |
300 | bool own_pages); | |
301 | extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *, | |
406e2c9f | 302 | unsigned int which, |
a4ce40a9 AE |
303 | struct ceph_pagelist *pagelist); |
304 | #ifdef CONFIG_BLOCK | |
305 | extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *, | |
406e2c9f | 306 | unsigned int which, |
a4ce40a9 AE |
307 | struct bio *bio, size_t bio_length); |
308 | #endif /* CONFIG_BLOCK */ | |
309 | ||
04017e29 AE |
310 | extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *, |
311 | unsigned int which, | |
312 | struct ceph_pagelist *pagelist); | |
6c57b554 AE |
313 | extern void osd_req_op_cls_request_data_pages(struct ceph_osd_request *, |
314 | unsigned int which, | |
315 | struct page **pages, u64 length, | |
316 | u32 alignment, bool pages_from_pool, | |
317 | bool own_pages); | |
a4ce40a9 | 318 | extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *, |
c99d2d4a | 319 | unsigned int which, |
a4ce40a9 AE |
320 | struct page **pages, u64 length, |
321 | u32 alignment, bool pages_from_pool, | |
322 | bool own_pages); | |
323 | ||
c99d2d4a AE |
324 | extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req, |
325 | unsigned int which, u16 opcode, | |
04017e29 | 326 | const char *class, const char *method); |
d74b50be YZ |
327 | extern int osd_req_op_xattr_init(struct ceph_osd_request *osd_req, unsigned int which, |
328 | u16 opcode, const char *name, const void *value, | |
329 | size_t size, u8 cmp_op, u8 cmp_mode); | |
c647b8a8 ID |
330 | extern void osd_req_op_alloc_hint_init(struct ceph_osd_request *osd_req, |
331 | unsigned int which, | |
332 | u64 expected_object_size, | |
333 | u64 expected_write_size); | |
33803f33 | 334 | |
3499e8a5 | 335 | extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, |
3499e8a5 | 336 | struct ceph_snap_context *snapc, |
acead002 | 337 | unsigned int num_ops, |
3499e8a5 | 338 | bool use_mempool, |
54a54007 | 339 | gfp_t gfp_flags); |
13d1ad16 | 340 | int ceph_osdc_alloc_messages(struct ceph_osd_request *req, gfp_t gfp); |
3499e8a5 | 341 | |
f24e9980 SW |
342 | extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *, |
343 | struct ceph_file_layout *layout, | |
344 | struct ceph_vino vino, | |
acead002 | 345 | u64 offset, u64 *len, |
715e4cd4 YZ |
346 | unsigned int which, int num_ops, |
347 | int opcode, int flags, | |
f24e9980 | 348 | struct ceph_snap_context *snapc, |
acead002 | 349 | u32 truncate_seq, u64 truncate_size, |
153e5167 | 350 | bool use_mempool); |
f24e9980 | 351 | |
9e94af20 ID |
352 | extern void ceph_osdc_get_request(struct ceph_osd_request *req); |
353 | extern void ceph_osdc_put_request(struct ceph_osd_request *req); | |
f24e9980 SW |
354 | |
355 | extern int ceph_osdc_start_request(struct ceph_osd_client *osdc, | |
356 | struct ceph_osd_request *req, | |
357 | bool nofail); | |
c9f9b93d | 358 | extern void ceph_osdc_cancel_request(struct ceph_osd_request *req); |
f24e9980 SW |
359 | extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc, |
360 | struct ceph_osd_request *req); | |
361 | extern void ceph_osdc_sync(struct ceph_osd_client *osdc); | |
362 | ||
dd935f44 JD |
363 | extern void ceph_osdc_flush_notifies(struct ceph_osd_client *osdc); |
364 | ||
f24e9980 SW |
365 | extern int ceph_osdc_readpages(struct ceph_osd_client *osdc, |
366 | struct ceph_vino vino, | |
367 | struct ceph_file_layout *layout, | |
368 | u64 off, u64 *plen, | |
369 | u32 truncate_seq, u64 truncate_size, | |
b7495fc2 SW |
370 | struct page **pages, int nr_pages, |
371 | int page_align); | |
f24e9980 SW |
372 | |
373 | extern int ceph_osdc_writepages(struct ceph_osd_client *osdc, | |
374 | struct ceph_vino vino, | |
375 | struct ceph_file_layout *layout, | |
376 | struct ceph_snap_context *sc, | |
377 | u64 off, u64 len, | |
378 | u32 truncate_seq, u64 truncate_size, | |
379 | struct timespec *mtime, | |
24808826 | 380 | struct page **pages, int nr_pages); |
f24e9980 | 381 | |
922dab61 ID |
382 | /* watch/notify */ |
383 | struct ceph_osd_linger_request * | |
384 | ceph_osdc_watch(struct ceph_osd_client *osdc, | |
385 | struct ceph_object_id *oid, | |
386 | struct ceph_object_locator *oloc, | |
387 | rados_watchcb2_t wcb, | |
388 | rados_watcherrcb_t errcb, | |
389 | void *data); | |
390 | int ceph_osdc_unwatch(struct ceph_osd_client *osdc, | |
391 | struct ceph_osd_linger_request *lreq); | |
392 | ||
393 | int ceph_osdc_notify_ack(struct ceph_osd_client *osdc, | |
394 | struct ceph_object_id *oid, | |
395 | struct ceph_object_locator *oloc, | |
396 | u64 notify_id, | |
397 | u64 cookie, | |
398 | void *payload, | |
399 | size_t payload_len); | |
f24e9980 SW |
400 | #endif |
401 |