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 | 14 | |
224736d9 SP |
15 | /* |
16 | * Maximum object name size | |
17 | * (must be at least as big as RBD_MAX_MD_NAME_LEN -- currently 100) | |
18 | */ | |
19 | #define MAX_OBJ_NAME_SIZE 100 | |
20 | ||
f24e9980 SW |
21 | struct ceph_msg; |
22 | struct ceph_snap_context; | |
23 | struct ceph_osd_request; | |
24 | struct ceph_osd_client; | |
4e7a5dcd | 25 | struct ceph_authorizer; |
f24e9980 SW |
26 | |
27 | /* | |
28 | * completion callback for async writepages | |
29 | */ | |
30 | typedef void (*ceph_osdc_callback_t)(struct ceph_osd_request *, | |
31 | struct ceph_msg *); | |
26be8808 | 32 | typedef void (*ceph_osdc_unsafe_callback_t)(struct ceph_osd_request *, bool); |
f24e9980 SW |
33 | |
34 | /* a given osd we're communicating with */ | |
35 | struct ceph_osd { | |
36 | atomic_t o_ref; | |
37 | struct ceph_osd_client *o_osdc; | |
38 | int o_osd; | |
39 | int o_incarnation; | |
40 | struct rb_node o_node; | |
41 | struct ceph_connection o_con; | |
42 | struct list_head o_requests; | |
a40c4f10 | 43 | struct list_head o_linger_requests; |
f5a2041b | 44 | struct list_head o_osd_lru; |
6c4a1915 | 45 | struct ceph_auth_handshake o_auth; |
f5a2041b | 46 | unsigned long lru_ttl; |
422d2cb8 YS |
47 | int o_marked_for_keepalive; |
48 | struct list_head o_keepalive_item; | |
f24e9980 SW |
49 | }; |
50 | ||
1b83bef2 | 51 | |
79528734 | 52 | #define CEPH_OSD_MAX_OP 2 |
1b83bef2 | 53 | |
2ac2b7a6 | 54 | enum ceph_osd_data_type { |
ec9123c5 | 55 | CEPH_OSD_DATA_TYPE_NONE = 0, |
2ac2b7a6 | 56 | CEPH_OSD_DATA_TYPE_PAGES, |
9a5e6d09 | 57 | CEPH_OSD_DATA_TYPE_PAGELIST, |
2ac2b7a6 AE |
58 | #ifdef CONFIG_BLOCK |
59 | CEPH_OSD_DATA_TYPE_BIO, | |
60 | #endif /* CONFIG_BLOCK */ | |
61 | }; | |
62 | ||
2794a82a | 63 | struct ceph_osd_data { |
2ac2b7a6 AE |
64 | enum ceph_osd_data_type type; |
65 | union { | |
2794a82a AE |
66 | struct { |
67 | struct page **pages; | |
e0c59487 | 68 | u64 length; |
2794a82a AE |
69 | u32 alignment; |
70 | bool pages_from_pool; | |
71 | bool own_pages; | |
72 | }; | |
9a5e6d09 | 73 | struct ceph_pagelist *pagelist; |
2794a82a | 74 | #ifdef CONFIG_BLOCK |
fdce58cc AE |
75 | struct { |
76 | struct bio *bio; /* list of bios */ | |
77 | size_t bio_length; /* total in list */ | |
78 | }; | |
2794a82a AE |
79 | #endif /* CONFIG_BLOCK */ |
80 | }; | |
81 | }; | |
82 | ||
79528734 AE |
83 | struct ceph_osd_req_op { |
84 | u16 op; /* CEPH_OSD_OP_* */ | |
85 | u32 payload_len; | |
86 | union { | |
87 | struct { | |
88 | u64 offset, length; | |
89 | u64 truncate_size; | |
90 | u32 truncate_seq; | |
5476492f | 91 | struct ceph_osd_data osd_data; |
79528734 AE |
92 | } extent; |
93 | struct { | |
94 | const char *class_name; | |
95 | const char *method_name; | |
5476492f | 96 | struct ceph_osd_data request_info; |
04017e29 | 97 | struct ceph_osd_data request_data; |
5476492f | 98 | struct ceph_osd_data response_data; |
79528734 AE |
99 | __u8 class_len; |
100 | __u8 method_len; | |
101 | __u8 argc; | |
102 | } cls; | |
103 | struct { | |
104 | u64 cookie; | |
105 | u64 ver; | |
106 | u32 prot_ver; | |
107 | u32 timeout; | |
108 | __u8 flag; | |
109 | } watch; | |
110 | }; | |
111 | }; | |
112 | ||
f24e9980 SW |
113 | /* an in-flight request */ |
114 | struct ceph_osd_request { | |
115 | u64 r_tid; /* unique for this client */ | |
116 | struct rb_node r_node; | |
422d2cb8 | 117 | struct list_head r_req_lru_item; |
f24e9980 | 118 | struct list_head r_osd_item; |
a40c4f10 YS |
119 | struct list_head r_linger_item; |
120 | struct list_head r_linger_osd; | |
f24e9980 | 121 | struct ceph_osd *r_osd; |
5b191d99 | 122 | struct ceph_pg r_pgid; |
d85b7056 SW |
123 | int r_pg_osds[CEPH_PG_MAX_SIZE]; |
124 | int r_num_pg_osds; | |
f24e9980 SW |
125 | |
126 | struct ceph_msg *r_request, *r_reply; | |
f24e9980 SW |
127 | int r_flags; /* any additional flags for the osd */ |
128 | u32 r_sent; /* >0 if r_request is sending/sent */ | |
1b83bef2 | 129 | |
79528734 AE |
130 | /* request osd ops array */ |
131 | unsigned int r_num_ops; | |
132 | struct ceph_osd_req_op r_ops[CEPH_OSD_MAX_OP]; | |
133 | ||
1b83bef2 SW |
134 | /* these are updated on each send */ |
135 | __le32 *r_request_osdmap_epoch; | |
136 | __le32 *r_request_flags; | |
137 | __le64 *r_request_pool; | |
138 | void *r_request_pgid; | |
139 | __le32 *r_request_attempts; | |
140 | struct ceph_eversion *r_request_reassert_version; | |
141 | ||
142 | int r_result; | |
143 | int r_reply_op_len[CEPH_OSD_MAX_OP]; | |
144 | s32 r_reply_op_result[CEPH_OSD_MAX_OP]; | |
350b1c32 | 145 | int r_got_reply; |
a40c4f10 | 146 | int r_linger; |
0d5af164 | 147 | int r_completed; |
f24e9980 SW |
148 | |
149 | struct ceph_osd_client *r_osdc; | |
415e49a9 | 150 | struct kref r_kref; |
f24e9980 SW |
151 | bool r_mempool; |
152 | struct completion r_completion, r_safe_completion; | |
26be8808 AE |
153 | ceph_osdc_callback_t r_callback; |
154 | ceph_osdc_unsafe_callback_t r_unsafe_callback; | |
f24e9980 SW |
155 | struct ceph_eversion r_reassert_version; |
156 | struct list_head r_unsafe_item; | |
157 | ||
158 | struct inode *r_inode; /* for use by callbacks */ | |
3d14c5d2 | 159 | void *r_priv; /* ditto */ |
f24e9980 | 160 | |
224736d9 | 161 | char r_oid[MAX_OBJ_NAME_SIZE]; /* object name */ |
f24e9980 | 162 | int r_oid_len; |
2169aea6 | 163 | u64 r_snapid; |
3dd72fc0 | 164 | unsigned long r_stamp; /* send OR check time */ |
f24e9980 SW |
165 | |
166 | struct ceph_file_layout r_file_layout; | |
167 | struct ceph_snap_context *r_snapc; /* snap context for writes */ | |
f24e9980 SW |
168 | }; |
169 | ||
a40c4f10 YS |
170 | struct ceph_osd_event { |
171 | u64 cookie; | |
172 | int one_shot; | |
173 | struct ceph_osd_client *osdc; | |
174 | void (*cb)(u64, u64, u8, void *); | |
175 | void *data; | |
176 | struct rb_node node; | |
177 | struct list_head osd_node; | |
178 | struct kref kref; | |
a40c4f10 YS |
179 | }; |
180 | ||
181 | struct ceph_osd_event_work { | |
182 | struct work_struct work; | |
183 | struct ceph_osd_event *event; | |
184 | u64 ver; | |
185 | u64 notify_id; | |
186 | u8 opcode; | |
187 | }; | |
188 | ||
f24e9980 SW |
189 | struct ceph_osd_client { |
190 | struct ceph_client *client; | |
191 | ||
192 | struct ceph_osdmap *osdmap; /* current map */ | |
193 | struct rw_semaphore map_sem; | |
194 | struct completion map_waiters; | |
195 | u64 last_requested_map; | |
196 | ||
197 | struct mutex request_mutex; | |
198 | struct rb_root osds; /* osds */ | |
f5a2041b | 199 | struct list_head osd_lru; /* idle osds */ |
f24e9980 SW |
200 | u64 timeout_tid; /* tid of timeout triggering rq */ |
201 | u64 last_tid; /* tid of last request */ | |
202 | struct rb_root requests; /* pending requests */ | |
6f6c7006 SW |
203 | struct list_head req_lru; /* in-flight lru */ |
204 | struct list_head req_unsent; /* unsent/need-resend queue */ | |
205 | struct list_head req_notarget; /* map to no osd */ | |
a40c4f10 | 206 | struct list_head req_linger; /* lingering requests */ |
f24e9980 SW |
207 | int num_requests; |
208 | struct delayed_work timeout_work; | |
f5a2041b | 209 | struct delayed_work osds_timeout_work; |
039934b8 | 210 | #ifdef CONFIG_DEBUG_FS |
f24e9980 | 211 | struct dentry *debugfs_file; |
039934b8 | 212 | #endif |
f24e9980 SW |
213 | |
214 | mempool_t *req_mempool; | |
215 | ||
0d59ab81 | 216 | struct ceph_msgpool msgpool_op; |
c16e7869 | 217 | struct ceph_msgpool msgpool_op_reply; |
a40c4f10 YS |
218 | |
219 | spinlock_t event_lock; | |
220 | struct rb_root event_tree; | |
221 | u64 event_count; | |
222 | ||
223 | struct workqueue_struct *notify_wq; | |
f24e9980 SW |
224 | }; |
225 | ||
226 | extern int ceph_osdc_init(struct ceph_osd_client *osdc, | |
227 | struct ceph_client *client); | |
228 | extern void ceph_osdc_stop(struct ceph_osd_client *osdc); | |
229 | ||
230 | extern void ceph_osdc_handle_reply(struct ceph_osd_client *osdc, | |
231 | struct ceph_msg *msg); | |
232 | extern void ceph_osdc_handle_map(struct ceph_osd_client *osdc, | |
233 | struct ceph_msg *msg); | |
234 | ||
c99d2d4a AE |
235 | extern void osd_req_op_extent_init(struct ceph_osd_request *osd_req, |
236 | unsigned int which, u16 opcode, | |
33803f33 AE |
237 | u64 offset, u64 length, |
238 | u64 truncate_size, u32 truncate_seq); | |
c99d2d4a AE |
239 | extern void osd_req_op_extent_update(struct ceph_osd_request *osd_req, |
240 | unsigned int which, u64 length); | |
a4ce40a9 AE |
241 | |
242 | extern struct ceph_osd_data *osd_req_op_extent_osd_data( | |
243 | struct ceph_osd_request *osd_req, | |
244 | unsigned int which, bool write_request); | |
245 | extern struct ceph_osd_data *osd_req_op_cls_response_data( | |
246 | struct ceph_osd_request *osd_req, | |
247 | unsigned int which); | |
248 | ||
249 | extern void osd_req_op_extent_osd_data_pages(struct ceph_osd_request *, | |
250 | unsigned int which, bool write_request, | |
251 | struct page **pages, u64 length, | |
252 | u32 alignment, bool pages_from_pool, | |
253 | bool own_pages); | |
254 | extern void osd_req_op_extent_osd_data_pagelist(struct ceph_osd_request *, | |
255 | unsigned int which, bool write_request, | |
256 | struct ceph_pagelist *pagelist); | |
257 | #ifdef CONFIG_BLOCK | |
258 | extern void osd_req_op_extent_osd_data_bio(struct ceph_osd_request *, | |
259 | unsigned int which, bool write_request, | |
260 | struct bio *bio, size_t bio_length); | |
261 | #endif /* CONFIG_BLOCK */ | |
262 | ||
04017e29 AE |
263 | extern void osd_req_op_cls_request_data_pagelist(struct ceph_osd_request *, |
264 | unsigned int which, | |
265 | struct ceph_pagelist *pagelist); | |
a4ce40a9 | 266 | extern void osd_req_op_cls_response_data_pages(struct ceph_osd_request *, |
c99d2d4a | 267 | unsigned int which, |
a4ce40a9 AE |
268 | struct page **pages, u64 length, |
269 | u32 alignment, bool pages_from_pool, | |
270 | bool own_pages); | |
271 | ||
c99d2d4a AE |
272 | extern void osd_req_op_cls_init(struct ceph_osd_request *osd_req, |
273 | unsigned int which, u16 opcode, | |
04017e29 | 274 | const char *class, const char *method); |
c99d2d4a AE |
275 | extern void osd_req_op_watch_init(struct ceph_osd_request *osd_req, |
276 | unsigned int which, u16 opcode, | |
33803f33 AE |
277 | u64 cookie, u64 version, int flag); |
278 | ||
3499e8a5 | 279 | extern struct ceph_osd_request *ceph_osdc_alloc_request(struct ceph_osd_client *osdc, |
3499e8a5 | 280 | struct ceph_snap_context *snapc, |
acead002 | 281 | unsigned int num_ops, |
3499e8a5 | 282 | bool use_mempool, |
54a54007 | 283 | gfp_t gfp_flags); |
3499e8a5 | 284 | |
175face2 | 285 | extern void ceph_osdc_build_request(struct ceph_osd_request *req, u64 off, |
68b4476b | 286 | struct ceph_snap_context *snapc, |
4d6b250b | 287 | u64 snap_id, |
af77f26c | 288 | struct timespec *mtime); |
3499e8a5 | 289 | |
f24e9980 SW |
290 | extern struct ceph_osd_request *ceph_osdc_new_request(struct ceph_osd_client *, |
291 | struct ceph_file_layout *layout, | |
292 | struct ceph_vino vino, | |
acead002 | 293 | u64 offset, u64 *len, |
79528734 | 294 | int num_ops, int opcode, int flags, |
f24e9980 | 295 | struct ceph_snap_context *snapc, |
acead002 | 296 | u32 truncate_seq, u64 truncate_size, |
153e5167 | 297 | bool use_mempool); |
f24e9980 | 298 | |
a40c4f10 YS |
299 | extern void ceph_osdc_set_request_linger(struct ceph_osd_client *osdc, |
300 | struct ceph_osd_request *req); | |
301 | extern void ceph_osdc_unregister_linger_request(struct ceph_osd_client *osdc, | |
302 | struct ceph_osd_request *req); | |
303 | ||
f24e9980 SW |
304 | static inline void ceph_osdc_get_request(struct ceph_osd_request *req) |
305 | { | |
415e49a9 SW |
306 | kref_get(&req->r_kref); |
307 | } | |
308 | extern void ceph_osdc_release_request(struct kref *kref); | |
309 | static inline void ceph_osdc_put_request(struct ceph_osd_request *req) | |
310 | { | |
311 | kref_put(&req->r_kref, ceph_osdc_release_request); | |
f24e9980 | 312 | } |
f24e9980 SW |
313 | |
314 | extern int ceph_osdc_start_request(struct ceph_osd_client *osdc, | |
315 | struct ceph_osd_request *req, | |
316 | bool nofail); | |
317 | extern int ceph_osdc_wait_request(struct ceph_osd_client *osdc, | |
318 | struct ceph_osd_request *req); | |
319 | extern void ceph_osdc_sync(struct ceph_osd_client *osdc); | |
320 | ||
321 | extern int ceph_osdc_readpages(struct ceph_osd_client *osdc, | |
322 | struct ceph_vino vino, | |
323 | struct ceph_file_layout *layout, | |
324 | u64 off, u64 *plen, | |
325 | u32 truncate_seq, u64 truncate_size, | |
b7495fc2 SW |
326 | struct page **pages, int nr_pages, |
327 | int page_align); | |
f24e9980 SW |
328 | |
329 | extern int ceph_osdc_writepages(struct ceph_osd_client *osdc, | |
330 | struct ceph_vino vino, | |
331 | struct ceph_file_layout *layout, | |
332 | struct ceph_snap_context *sc, | |
333 | u64 off, u64 len, | |
334 | u32 truncate_seq, u64 truncate_size, | |
335 | struct timespec *mtime, | |
24808826 | 336 | struct page **pages, int nr_pages); |
f24e9980 | 337 | |
a40c4f10 YS |
338 | /* watch/notify events */ |
339 | extern int ceph_osdc_create_event(struct ceph_osd_client *osdc, | |
340 | void (*event_cb)(u64, u64, u8, void *), | |
3c663bbd | 341 | void *data, struct ceph_osd_event **pevent); |
a40c4f10 | 342 | extern void ceph_osdc_cancel_event(struct ceph_osd_event *event); |
a40c4f10 | 343 | extern void ceph_osdc_put_event(struct ceph_osd_event *event); |
f24e9980 SW |
344 | #endif |
345 |