Commit | Line | Data |
---|---|---|
31b8006e SW |
1 | #ifndef __FS_CEPH_MESSENGER_H |
2 | #define __FS_CEPH_MESSENGER_H | |
3 | ||
c2e552e7 | 4 | #include <linux/kref.h> |
31b8006e SW |
5 | #include <linux/mutex.h> |
6 | #include <linux/net.h> | |
7 | #include <linux/radix-tree.h> | |
8 | #include <linux/uio.h> | |
31b8006e SW |
9 | #include <linux/workqueue.h> |
10 | ||
a1ce3928 DH |
11 | #include <linux/ceph/types.h> |
12 | #include <linux/ceph/buffer.h> | |
31b8006e SW |
13 | |
14 | struct ceph_msg; | |
15 | struct ceph_connection; | |
16 | ||
31b8006e SW |
17 | /* |
18 | * Ceph defines these callbacks for handling connection events. | |
19 | */ | |
20 | struct ceph_connection_operations { | |
21 | struct ceph_connection *(*get)(struct ceph_connection *); | |
22 | void (*put)(struct ceph_connection *); | |
23 | ||
24 | /* handle an incoming message. */ | |
25 | void (*dispatch) (struct ceph_connection *con, struct ceph_msg *m); | |
26 | ||
4e7a5dcd | 27 | /* authorize an outgoing connection */ |
a3530df3 AE |
28 | struct ceph_auth_handshake *(*get_authorizer) ( |
29 | struct ceph_connection *con, | |
8f43fb53 | 30 | int *proto, int force_new); |
4e7a5dcd | 31 | int (*verify_authorizer_reply) (struct ceph_connection *con, int len); |
9bd2e6f8 | 32 | int (*invalidate_authorizer)(struct ceph_connection *con); |
4e7a5dcd | 33 | |
31b8006e SW |
34 | /* there was some error on the socket (disconnect, whatever) */ |
35 | void (*fault) (struct ceph_connection *con); | |
36 | ||
37 | /* a remote host as terminated a message exchange session, and messages | |
38 | * we sent (or they tried to send us) may be lost. */ | |
39 | void (*peer_reset) (struct ceph_connection *con); | |
40 | ||
41 | struct ceph_msg * (*alloc_msg) (struct ceph_connection *con, | |
2450418c YS |
42 | struct ceph_msg_header *hdr, |
43 | int *skip); | |
31b8006e SW |
44 | }; |
45 | ||
31b8006e | 46 | /* use format string %s%d */ |
ca9d93a2 | 47 | #define ENTITY_NAME(n) ceph_entity_type_name((n).type), le64_to_cpu((n).num) |
31b8006e SW |
48 | |
49 | struct ceph_messenger { | |
50 | struct ceph_entity_inst inst; /* my name+address */ | |
63f2d211 | 51 | struct ceph_entity_addr my_enc_addr; |
31b8006e | 52 | |
a2a32584 | 53 | atomic_t stopping; |
31b8006e SW |
54 | bool nocrc; |
55 | ||
56 | /* | |
57 | * the global_seq counts connections i (attempt to) initiate | |
58 | * in order to disambiguate certain connect race conditions. | |
59 | */ | |
60 | u32 global_seq; | |
61 | spinlock_t global_seq_lock; | |
3d14c5d2 YS |
62 | |
63 | u32 supported_features; | |
64 | u32 required_features; | |
31b8006e SW |
65 | }; |
66 | ||
43794509 AE |
67 | enum ceph_msg_data_type { |
68 | CEPH_MSG_DATA_NONE, /* message contains no data payload */ | |
69 | CEPH_MSG_DATA_PAGES, /* data source/destination is a page array */ | |
70 | CEPH_MSG_DATA_PAGELIST, /* data source/destination is a pagelist */ | |
71 | #ifdef CONFIG_BLOCK | |
72 | CEPH_MSG_DATA_BIO, /* data source/destination is a bio list */ | |
73 | #endif /* CONFIG_BLOCK */ | |
74 | }; | |
75 | ||
76 | static __inline__ bool ceph_msg_data_type_valid(enum ceph_msg_data_type type) | |
77 | { | |
78 | switch (type) { | |
79 | case CEPH_MSG_DATA_NONE: | |
80 | case CEPH_MSG_DATA_PAGES: | |
81 | case CEPH_MSG_DATA_PAGELIST: | |
82 | #ifdef CONFIG_BLOCK | |
83 | case CEPH_MSG_DATA_BIO: | |
84 | #endif /* CONFIG_BLOCK */ | |
85 | return true; | |
86 | default: | |
87 | return false; | |
88 | } | |
89 | } | |
90 | ||
fe38a2b6 | 91 | struct ceph_msg_data_cursor { |
25aff7c5 | 92 | size_t resid; /* bytes not yet consumed */ |
fe38a2b6 | 93 | bool last_piece; /* now at last piece of data item */ |
f5db90bc | 94 | bool need_crc; /* new piece; crc update needed */ |
dd236fcb | 95 | union { |
6aaa4511 AE |
96 | #ifdef CONFIG_BLOCK |
97 | struct { /* bio */ | |
98 | struct bio *bio; /* bio from list */ | |
99 | unsigned int vector_index; /* vector from bio */ | |
100 | unsigned int vector_offset; /* bytes from vector */ | |
101 | }; | |
102 | #endif /* CONFIG_BLOCK */ | |
e766d7b5 | 103 | struct { /* pages */ |
e766d7b5 AE |
104 | unsigned int page_offset; /* offset in page */ |
105 | unsigned short page_index; /* index in array */ | |
106 | unsigned short page_count; /* pages in array */ | |
107 | }; | |
dd236fcb AE |
108 | struct { /* pagelist */ |
109 | struct page *page; /* page from list */ | |
110 | size_t offset; /* bytes from list */ | |
111 | }; | |
112 | }; | |
fe38a2b6 AE |
113 | }; |
114 | ||
43794509 AE |
115 | struct ceph_msg_data { |
116 | enum ceph_msg_data_type type; | |
117 | union { | |
118 | #ifdef CONFIG_BLOCK | |
6518be47 | 119 | struct bio *bio; |
43794509 AE |
120 | #endif /* CONFIG_BLOCK */ |
121 | struct { | |
122 | struct page **pages; /* NOT OWNER. */ | |
123 | size_t length; /* total # bytes */ | |
124 | unsigned int alignment; /* first page */ | |
125 | }; | |
126 | struct ceph_pagelist *pagelist; | |
127 | }; | |
fe38a2b6 | 128 | struct ceph_msg_data_cursor cursor; /* pagelist only */ |
43794509 | 129 | }; |
97fb1c7f | 130 | |
31b8006e SW |
131 | /* |
132 | * a single message. it contains a header (src, dest, message type, etc.), | |
133 | * footer (crc values, mainly), a "front" message body, and possibly a | |
134 | * data payload (stored in some number of pages). | |
135 | */ | |
136 | struct ceph_msg { | |
137 | struct ceph_msg_header hdr; /* header */ | |
138 | struct ceph_msg_footer footer; /* footer */ | |
139 | struct kvec front; /* unaligned blobs of message */ | |
140 | struct ceph_buffer *middle; | |
38941f80 | 141 | |
a1930804 | 142 | size_t data_length; |
6644ed7b | 143 | struct ceph_msg_data *data; /* data payload */ |
02afca6c AE |
144 | |
145 | struct ceph_connection *con; | |
146 | struct list_head list_head; /* links for connection lists */ | |
147 | ||
148 | struct kref kref; | |
31b8006e SW |
149 | bool front_is_vmalloc; |
150 | bool more_to_follow; | |
e84346b7 | 151 | bool needs_out_seq; |
31b8006e | 152 | int front_max; |
4cf9d544 | 153 | unsigned long ack_stamp; /* tx: when we were acked */ |
31b8006e SW |
154 | |
155 | struct ceph_msgpool *pool; | |
156 | }; | |
157 | ||
31b8006e SW |
158 | /* ceph connection fault delay defaults, for exponential backoff */ |
159 | #define BASE_DELAY_INTERVAL (HZ/2) | |
160 | #define MAX_DELAY_INTERVAL (5 * 60 * HZ) | |
161 | ||
31b8006e SW |
162 | /* |
163 | * A single connection with another host. | |
164 | * | |
165 | * We maintain a queue of outgoing messages, and some session state to | |
166 | * ensure that we can preserve the lossless, ordered delivery of | |
167 | * messages in the case of a TCP disconnect. | |
168 | */ | |
169 | struct ceph_connection { | |
170 | void *private; | |
31b8006e SW |
171 | |
172 | const struct ceph_connection_operations *ops; | |
173 | ||
174 | struct ceph_messenger *msgr; | |
ce2c8903 AE |
175 | |
176 | atomic_t sock_state; | |
31b8006e | 177 | struct socket *sock; |
ce2c8903 AE |
178 | struct ceph_entity_addr peer_addr; /* peer address */ |
179 | struct ceph_entity_addr peer_addr_for_me; | |
180 | ||
928443cd AE |
181 | unsigned long flags; |
182 | unsigned long state; | |
31b8006e SW |
183 | const char *error_msg; /* error message, if any */ |
184 | ||
31b8006e | 185 | struct ceph_entity_name peer_name; /* peer name */ |
ce2c8903 | 186 | |
aba558e2 | 187 | unsigned peer_features; |
31b8006e SW |
188 | u32 connect_seq; /* identify the most recent connection |
189 | attempt for this connection, client */ | |
190 | u32 peer_global_seq; /* peer's global seq for this connection */ | |
191 | ||
4e7a5dcd SW |
192 | int auth_retry; /* true if we need a newer authorizer */ |
193 | void *auth_reply_buf; /* where to put the authorizer reply */ | |
194 | int auth_reply_buf_len; | |
195 | ||
ec302645 SW |
196 | struct mutex mutex; |
197 | ||
31b8006e | 198 | /* out queue */ |
31b8006e SW |
199 | struct list_head out_queue; |
200 | struct list_head out_sent; /* sending or sent but unacked */ | |
201 | u64 out_seq; /* last message queued for send */ | |
31b8006e SW |
202 | |
203 | u64 in_seq, in_seq_acked; /* last message received, acked */ | |
204 | ||
205 | /* connection negotiation temps */ | |
206 | char in_banner[CEPH_BANNER_MAX_LEN]; | |
a16cb1f7 SW |
207 | struct ceph_msg_connect out_connect; |
208 | struct ceph_msg_connect_reply in_reply; | |
31b8006e SW |
209 | struct ceph_entity_addr actual_peer_addr; |
210 | ||
211 | /* message out temps */ | |
212 | struct ceph_msg *out_msg; /* sending message (== tail of | |
213 | out_sent) */ | |
c86a2930 | 214 | bool out_msg_done; |
31b8006e SW |
215 | |
216 | struct kvec out_kvec[8], /* sending header/footer data */ | |
217 | *out_kvec_cur; | |
218 | int out_kvec_left; /* kvec's left in out_kvec */ | |
219 | int out_skip; /* skip this many bytes */ | |
220 | int out_kvec_bytes; /* total bytes left */ | |
221 | bool out_kvec_is_msg; /* kvec refers to out_msg */ | |
222 | int out_more; /* there is more data after the kvecs */ | |
223 | __le64 out_temp_ack; /* for writing an ack */ | |
224 | ||
225 | /* message in temps */ | |
226 | struct ceph_msg_header in_hdr; | |
227 | struct ceph_msg *in_msg; | |
31b8006e SW |
228 | u32 in_front_crc, in_middle_crc, in_data_crc; /* calculated crc */ |
229 | ||
230 | char in_tag; /* protocol control byte */ | |
231 | int in_base_pos; /* bytes read */ | |
232 | __le64 in_temp_ack; /* for reading an ack */ | |
233 | ||
234 | struct delayed_work work; /* send|recv work */ | |
235 | unsigned long delay; /* current delay interval */ | |
236 | }; | |
237 | ||
238 | ||
3d14c5d2 | 239 | extern const char *ceph_pr_addr(const struct sockaddr_storage *ss); |
31b8006e SW |
240 | extern int ceph_parse_ips(const char *c, const char *end, |
241 | struct ceph_entity_addr *addr, | |
242 | int max_count, int *count); | |
243 | ||
244 | ||
245 | extern int ceph_msgr_init(void); | |
246 | extern void ceph_msgr_exit(void); | |
a922d38f | 247 | extern void ceph_msgr_flush(void); |
31b8006e | 248 | |
15d9882c AE |
249 | extern void ceph_messenger_init(struct ceph_messenger *msgr, |
250 | struct ceph_entity_addr *myaddr, | |
251 | u32 supported_features, | |
252 | u32 required_features, | |
253 | bool nocrc); | |
31b8006e | 254 | |
1bfd89f4 AE |
255 | extern void ceph_con_init(struct ceph_connection *con, void *private, |
256 | const struct ceph_connection_operations *ops, | |
b7a9e5dd | 257 | struct ceph_messenger *msgr); |
31b8006e | 258 | extern void ceph_con_open(struct ceph_connection *con, |
b7a9e5dd | 259 | __u8 entity_type, __u64 entity_num, |
31b8006e | 260 | struct ceph_entity_addr *addr); |
87b315a5 | 261 | extern bool ceph_con_opened(struct ceph_connection *con); |
31b8006e SW |
262 | extern void ceph_con_close(struct ceph_connection *con); |
263 | extern void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg); | |
6740a845 AE |
264 | |
265 | extern void ceph_msg_revoke(struct ceph_msg *msg); | |
8921d114 AE |
266 | extern void ceph_msg_revoke_incoming(struct ceph_msg *msg); |
267 | ||
31b8006e | 268 | extern void ceph_con_keepalive(struct ceph_connection *con); |
31b8006e | 269 | |
02afca6c | 270 | extern void ceph_msg_data_set_pages(struct ceph_msg *msg, struct page **pages, |
f1baeb2b | 271 | size_t length, size_t alignment); |
27fa8385 AE |
272 | extern void ceph_msg_data_set_pagelist(struct ceph_msg *msg, |
273 | struct ceph_pagelist *pagelist); | |
a1930804 AE |
274 | extern void ceph_msg_data_set_bio(struct ceph_msg *msg, struct bio *bio, |
275 | size_t length); | |
02afca6c | 276 | |
b61c2763 SW |
277 | extern struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags, |
278 | bool can_fail); | |
31b8006e SW |
279 | extern void ceph_msg_kfree(struct ceph_msg *m); |
280 | ||
31b8006e SW |
281 | |
282 | static inline struct ceph_msg *ceph_msg_get(struct ceph_msg *msg) | |
283 | { | |
c2e552e7 | 284 | kref_get(&msg->kref); |
31b8006e SW |
285 | return msg; |
286 | } | |
c2e552e7 SW |
287 | extern void ceph_msg_last_put(struct kref *kref); |
288 | static inline void ceph_msg_put(struct ceph_msg *msg) | |
289 | { | |
290 | kref_put(&msg->kref, ceph_msg_last_put); | |
291 | } | |
31b8006e | 292 | |
9ec7cab1 SW |
293 | extern void ceph_msg_dump(struct ceph_msg *msg); |
294 | ||
31b8006e | 295 | #endif |