libceph: make message data be a pointer
[GitHub/mt8127/android_kernel_alcatel_ttab.git] / include / linux / ceph / messenger.h
CommitLineData
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
14struct ceph_msg;
15struct ceph_connection;
16
31b8006e
SW
17/*
18 * Ceph defines these callbacks for handling connection events.
19 */
20struct 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
49struct 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
67enum 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
76static __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 91struct 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
115struct 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 */
136struct 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
6644ed7b 142 struct ceph_msg_data *data; /* data payload */
02afca6c
AE
143
144 struct ceph_connection *con;
145 struct list_head list_head; /* links for connection lists */
146
147 struct kref kref;
31b8006e
SW
148 bool front_is_vmalloc;
149 bool more_to_follow;
e84346b7 150 bool needs_out_seq;
31b8006e 151 int front_max;
4cf9d544 152 unsigned long ack_stamp; /* tx: when we were acked */
31b8006e
SW
153
154 struct ceph_msgpool *pool;
155};
156
31b8006e
SW
157/* ceph connection fault delay defaults, for exponential backoff */
158#define BASE_DELAY_INTERVAL (HZ/2)
159#define MAX_DELAY_INTERVAL (5 * 60 * HZ)
160
31b8006e
SW
161/*
162 * A single connection with another host.
163 *
164 * We maintain a queue of outgoing messages, and some session state to
165 * ensure that we can preserve the lossless, ordered delivery of
166 * messages in the case of a TCP disconnect.
167 */
168struct ceph_connection {
169 void *private;
31b8006e
SW
170
171 const struct ceph_connection_operations *ops;
172
173 struct ceph_messenger *msgr;
ce2c8903
AE
174
175 atomic_t sock_state;
31b8006e 176 struct socket *sock;
ce2c8903
AE
177 struct ceph_entity_addr peer_addr; /* peer address */
178 struct ceph_entity_addr peer_addr_for_me;
179
928443cd
AE
180 unsigned long flags;
181 unsigned long state;
31b8006e
SW
182 const char *error_msg; /* error message, if any */
183
31b8006e 184 struct ceph_entity_name peer_name; /* peer name */
ce2c8903 185
aba558e2 186 unsigned peer_features;
31b8006e
SW
187 u32 connect_seq; /* identify the most recent connection
188 attempt for this connection, client */
189 u32 peer_global_seq; /* peer's global seq for this connection */
190
4e7a5dcd
SW
191 int auth_retry; /* true if we need a newer authorizer */
192 void *auth_reply_buf; /* where to put the authorizer reply */
193 int auth_reply_buf_len;
194
ec302645
SW
195 struct mutex mutex;
196
31b8006e 197 /* out queue */
31b8006e
SW
198 struct list_head out_queue;
199 struct list_head out_sent; /* sending or sent but unacked */
200 u64 out_seq; /* last message queued for send */
31b8006e
SW
201
202 u64 in_seq, in_seq_acked; /* last message received, acked */
203
204 /* connection negotiation temps */
205 char in_banner[CEPH_BANNER_MAX_LEN];
a16cb1f7
SW
206 struct ceph_msg_connect out_connect;
207 struct ceph_msg_connect_reply in_reply;
31b8006e
SW
208 struct ceph_entity_addr actual_peer_addr;
209
210 /* message out temps */
211 struct ceph_msg *out_msg; /* sending message (== tail of
212 out_sent) */
c86a2930 213 bool out_msg_done;
31b8006e
SW
214
215 struct kvec out_kvec[8], /* sending header/footer data */
216 *out_kvec_cur;
217 int out_kvec_left; /* kvec's left in out_kvec */
218 int out_skip; /* skip this many bytes */
219 int out_kvec_bytes; /* total bytes left */
220 bool out_kvec_is_msg; /* kvec refers to out_msg */
221 int out_more; /* there is more data after the kvecs */
222 __le64 out_temp_ack; /* for writing an ack */
223
224 /* message in temps */
225 struct ceph_msg_header in_hdr;
226 struct ceph_msg *in_msg;
31b8006e
SW
227 u32 in_front_crc, in_middle_crc, in_data_crc; /* calculated crc */
228
229 char in_tag; /* protocol control byte */
230 int in_base_pos; /* bytes read */
231 __le64 in_temp_ack; /* for reading an ack */
232
233 struct delayed_work work; /* send|recv work */
234 unsigned long delay; /* current delay interval */
235};
236
237
3d14c5d2 238extern const char *ceph_pr_addr(const struct sockaddr_storage *ss);
31b8006e
SW
239extern int ceph_parse_ips(const char *c, const char *end,
240 struct ceph_entity_addr *addr,
241 int max_count, int *count);
242
243
244extern int ceph_msgr_init(void);
245extern void ceph_msgr_exit(void);
a922d38f 246extern void ceph_msgr_flush(void);
31b8006e 247
15d9882c
AE
248extern void ceph_messenger_init(struct ceph_messenger *msgr,
249 struct ceph_entity_addr *myaddr,
250 u32 supported_features,
251 u32 required_features,
252 bool nocrc);
31b8006e 253
1bfd89f4
AE
254extern void ceph_con_init(struct ceph_connection *con, void *private,
255 const struct ceph_connection_operations *ops,
b7a9e5dd 256 struct ceph_messenger *msgr);
31b8006e 257extern void ceph_con_open(struct ceph_connection *con,
b7a9e5dd 258 __u8 entity_type, __u64 entity_num,
31b8006e 259 struct ceph_entity_addr *addr);
87b315a5 260extern bool ceph_con_opened(struct ceph_connection *con);
31b8006e
SW
261extern void ceph_con_close(struct ceph_connection *con);
262extern void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg);
6740a845
AE
263
264extern void ceph_msg_revoke(struct ceph_msg *msg);
8921d114
AE
265extern void ceph_msg_revoke_incoming(struct ceph_msg *msg);
266
31b8006e 267extern void ceph_con_keepalive(struct ceph_connection *con);
31b8006e 268
02afca6c 269extern void ceph_msg_data_set_pages(struct ceph_msg *msg, struct page **pages,
f1baeb2b 270 size_t length, size_t alignment);
27fa8385
AE
271extern void ceph_msg_data_set_pagelist(struct ceph_msg *msg,
272 struct ceph_pagelist *pagelist);
273extern void ceph_msg_data_set_bio(struct ceph_msg *msg, struct bio *bio);
02afca6c 274
b61c2763
SW
275extern struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags,
276 bool can_fail);
31b8006e
SW
277extern void ceph_msg_kfree(struct ceph_msg *m);
278
31b8006e
SW
279
280static inline struct ceph_msg *ceph_msg_get(struct ceph_msg *msg)
281{
c2e552e7 282 kref_get(&msg->kref);
31b8006e
SW
283 return msg;
284}
c2e552e7
SW
285extern void ceph_msg_last_put(struct kref *kref);
286static inline void ceph_msg_put(struct ceph_msg *msg)
287{
288 kref_put(&msg->kref, ceph_msg_last_put);
289}
31b8006e 290
9ec7cab1
SW
291extern void ceph_msg_dump(struct ceph_msg *msg);
292
31b8006e 293#endif