blob: a343dae73cdcf4f6f458480e4b910e5b9889093e [file] [log] [blame]
Sage Weil31b80062009-10-06 11:31:13 -07001#ifndef __FS_CEPH_MESSENGER_H
2#define __FS_CEPH_MESSENGER_H
3
Sage Weilc2e552e2009-12-07 15:55:05 -08004#include <linux/kref.h>
Sage Weil31b80062009-10-06 11:31:13 -07005#include <linux/mutex.h>
6#include <linux/net.h>
7#include <linux/radix-tree.h>
8#include <linux/uio.h>
9#include <linux/version.h>
10#include <linux/workqueue.h>
11
12#include "types.h"
13#include "buffer.h"
14
15struct ceph_msg;
16struct ceph_connection;
17
18extern struct workqueue_struct *ceph_msgr_wq; /* receive work queue */
19
20/*
21 * Ceph defines these callbacks for handling connection events.
22 */
23struct ceph_connection_operations {
24 struct ceph_connection *(*get)(struct ceph_connection *);
25 void (*put)(struct ceph_connection *);
26
27 /* handle an incoming message. */
28 void (*dispatch) (struct ceph_connection *con, struct ceph_msg *m);
29
Sage Weil4e7a5dc2009-11-18 16:19:57 -080030 /* authorize an outgoing connection */
31 int (*get_authorizer) (struct ceph_connection *con,
32 void **buf, int *len, int *proto,
33 void **reply_buf, int *reply_len, int force_new);
34 int (*verify_authorizer_reply) (struct ceph_connection *con, int len);
Sage Weil9bd2e6f2010-02-02 16:21:06 -080035 int (*invalidate_authorizer)(struct ceph_connection *con);
Sage Weil4e7a5dc2009-11-18 16:19:57 -080036
Sage Weil31b80062009-10-06 11:31:13 -070037 /* protocol version mismatch */
38 void (*bad_proto) (struct ceph_connection *con);
39
40 /* there was some error on the socket (disconnect, whatever) */
41 void (*fault) (struct ceph_connection *con);
42
43 /* a remote host as terminated a message exchange session, and messages
44 * we sent (or they tried to send us) may be lost. */
45 void (*peer_reset) (struct ceph_connection *con);
46
47 struct ceph_msg * (*alloc_msg) (struct ceph_connection *con,
Yehuda Sadeh24504182010-01-08 13:58:34 -080048 struct ceph_msg_header *hdr,
49 int *skip);
Sage Weil31b80062009-10-06 11:31:13 -070050};
51
52extern const char *ceph_name_type_str(int t);
53
54/* use format string %s%d */
55#define ENTITY_NAME(n) ceph_name_type_str((n).type), le64_to_cpu((n).num)
56
57struct ceph_messenger {
58 struct ceph_entity_inst inst; /* my name+address */
Sage Weil63f2d212009-11-03 15:17:56 -080059 struct ceph_entity_addr my_enc_addr;
Sage Weil31b80062009-10-06 11:31:13 -070060 struct page *zero_page; /* used in certain error cases */
61
62 bool nocrc;
63
64 /*
65 * the global_seq counts connections i (attempt to) initiate
66 * in order to disambiguate certain connect race conditions.
67 */
68 u32 global_seq;
69 spinlock_t global_seq_lock;
70};
71
72/*
73 * a single message. it contains a header (src, dest, message type, etc.),
74 * footer (crc values, mainly), a "front" message body, and possibly a
75 * data payload (stored in some number of pages).
76 */
77struct ceph_msg {
78 struct ceph_msg_header hdr; /* header */
79 struct ceph_msg_footer footer; /* footer */
80 struct kvec front; /* unaligned blobs of message */
81 struct ceph_buffer *middle;
82 struct page **pages; /* data payload. NOT OWNER. */
83 unsigned nr_pages; /* size of page array */
Sage Weil58bb3b32009-12-23 12:12:31 -080084 struct ceph_pagelist *pagelist; /* instead of pages */
Sage Weil31b80062009-10-06 11:31:13 -070085 struct list_head list_head;
Sage Weilc2e552e2009-12-07 15:55:05 -080086 struct kref kref;
Sage Weil31b80062009-10-06 11:31:13 -070087 bool front_is_vmalloc;
88 bool more_to_follow;
89 int front_max;
90
91 struct ceph_msgpool *pool;
92};
93
94struct ceph_msg_pos {
95 int page, page_pos; /* which page; offset in page */
96 int data_pos; /* offset in data payload */
97 int did_page_crc; /* true if we've calculated crc for current page */
98};
99
100/* ceph connection fault delay defaults, for exponential backoff */
101#define BASE_DELAY_INTERVAL (HZ/2)
102#define MAX_DELAY_INTERVAL (5 * 60 * HZ)
103
104/*
105 * ceph_connection state bit flags
106 *
107 * QUEUED and BUSY are used together to ensure that only a single
108 * thread is currently opening, reading or writing data to the socket.
109 */
110#define LOSSYTX 0 /* we can close channel or drop messages on errors */
Sage Weileed0ef22009-11-10 14:34:36 -0800111#define CONNECTING 1
112#define NEGOTIATING 2
Sage Weil31b80062009-10-06 11:31:13 -0700113#define KEEPALIVE_PENDING 3
114#define WRITE_PENDING 4 /* we have data ready to send */
115#define QUEUED 5 /* there is work queued on this connection */
116#define BUSY 6 /* work is being done */
117#define STANDBY 8 /* no outgoing messages, socket closed. we keep
118 * the ceph_connection around to maintain shared
119 * state with the peer. */
120#define CLOSED 10 /* we've closed the connection */
121#define SOCK_CLOSED 11 /* socket state changed to closed */
Sage Weil31b80062009-10-06 11:31:13 -0700122#define OPENING 13 /* open connection w/ (possibly new) peer */
123#define DEAD 14 /* dead, about to kfree */
124
125/*
126 * A single connection with another host.
127 *
128 * We maintain a queue of outgoing messages, and some session state to
129 * ensure that we can preserve the lossless, ordered delivery of
130 * messages in the case of a TCP disconnect.
131 */
132struct ceph_connection {
133 void *private;
134 atomic_t nref;
135
136 const struct ceph_connection_operations *ops;
137
138 struct ceph_messenger *msgr;
139 struct socket *sock;
140 unsigned long state; /* connection state (see flags above) */
141 const char *error_msg; /* error message, if any */
142
143 struct ceph_entity_addr peer_addr; /* peer address */
144 struct ceph_entity_name peer_name; /* peer name */
145 struct ceph_entity_addr peer_addr_for_me;
146 u32 connect_seq; /* identify the most recent connection
147 attempt for this connection, client */
148 u32 peer_global_seq; /* peer's global seq for this connection */
149
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800150 int auth_retry; /* true if we need a newer authorizer */
151 void *auth_reply_buf; /* where to put the authorizer reply */
152 int auth_reply_buf_len;
153
Sage Weilec302642009-12-22 10:43:42 -0800154 struct mutex mutex;
155
Sage Weil31b80062009-10-06 11:31:13 -0700156 /* out queue */
Sage Weil31b80062009-10-06 11:31:13 -0700157 struct list_head out_queue;
158 struct list_head out_sent; /* sending or sent but unacked */
159 u64 out_seq; /* last message queued for send */
160 u64 out_seq_sent; /* last message sent */
161 bool out_keepalive_pending;
162
163 u64 in_seq, in_seq_acked; /* last message received, acked */
164
165 /* connection negotiation temps */
166 char in_banner[CEPH_BANNER_MAX_LEN];
167 union {
168 struct { /* outgoing connection */
169 struct ceph_msg_connect out_connect;
170 struct ceph_msg_connect_reply in_reply;
171 };
172 struct { /* incoming */
173 struct ceph_msg_connect in_connect;
174 struct ceph_msg_connect_reply out_reply;
175 };
176 };
177 struct ceph_entity_addr actual_peer_addr;
178
179 /* message out temps */
180 struct ceph_msg *out_msg; /* sending message (== tail of
181 out_sent) */
Sage Weilc86a2932009-12-14 14:04:30 -0800182 bool out_msg_done;
Sage Weil31b80062009-10-06 11:31:13 -0700183 struct ceph_msg_pos out_msg_pos;
184
185 struct kvec out_kvec[8], /* sending header/footer data */
186 *out_kvec_cur;
187 int out_kvec_left; /* kvec's left in out_kvec */
188 int out_skip; /* skip this many bytes */
189 int out_kvec_bytes; /* total bytes left */
190 bool out_kvec_is_msg; /* kvec refers to out_msg */
191 int out_more; /* there is more data after the kvecs */
192 __le64 out_temp_ack; /* for writing an ack */
193
194 /* message in temps */
195 struct ceph_msg_header in_hdr;
196 struct ceph_msg *in_msg;
197 struct ceph_msg_pos in_msg_pos;
198 u32 in_front_crc, in_middle_crc, in_data_crc; /* calculated crc */
199
200 char in_tag; /* protocol control byte */
201 int in_base_pos; /* bytes read */
202 __le64 in_temp_ack; /* for reading an ack */
203
204 struct delayed_work work; /* send|recv work */
205 unsigned long delay; /* current delay interval */
206};
207
208
209extern const char *pr_addr(const struct sockaddr_storage *ss);
210extern int ceph_parse_ips(const char *c, const char *end,
211 struct ceph_entity_addr *addr,
212 int max_count, int *count);
213
214
215extern int ceph_msgr_init(void);
216extern void ceph_msgr_exit(void);
217
218extern struct ceph_messenger *ceph_messenger_create(
219 struct ceph_entity_addr *myaddr);
220extern void ceph_messenger_destroy(struct ceph_messenger *);
221
222extern void ceph_con_init(struct ceph_messenger *msgr,
223 struct ceph_connection *con);
Sage Weil31b80062009-10-06 11:31:13 -0700224extern void ceph_con_open(struct ceph_connection *con,
225 struct ceph_entity_addr *addr);
Sage Weil87b315a2010-03-22 14:51:18 -0700226extern bool ceph_con_opened(struct ceph_connection *con);
Sage Weil31b80062009-10-06 11:31:13 -0700227extern void ceph_con_close(struct ceph_connection *con);
228extern void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg);
229extern void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg);
Yehuda Sadeh0d59ab82010-01-13 17:03:23 -0800230extern void ceph_con_revoke_message(struct ceph_connection *con,
231 struct ceph_msg *msg);
Sage Weil31b80062009-10-06 11:31:13 -0700232extern void ceph_con_keepalive(struct ceph_connection *con);
233extern struct ceph_connection *ceph_con_get(struct ceph_connection *con);
234extern void ceph_con_put(struct ceph_connection *con);
235
236extern struct ceph_msg *ceph_msg_new(int type, int front_len,
237 int page_len, int page_off,
238 struct page **pages);
239extern void ceph_msg_kfree(struct ceph_msg *m);
240
Sage Weil31b80062009-10-06 11:31:13 -0700241
242static inline struct ceph_msg *ceph_msg_get(struct ceph_msg *msg)
243{
Sage Weilc2e552e2009-12-07 15:55:05 -0800244 kref_get(&msg->kref);
Sage Weil31b80062009-10-06 11:31:13 -0700245 return msg;
246}
Sage Weilc2e552e2009-12-07 15:55:05 -0800247extern void ceph_msg_last_put(struct kref *kref);
248static inline void ceph_msg_put(struct ceph_msg *msg)
249{
250 kref_put(&msg->kref, ceph_msg_last_put);
251}
Sage Weil31b80062009-10-06 11:31:13 -0700252
Sage Weil9ec7cab2009-12-14 15:13:47 -0800253extern void ceph_msg_dump(struct ceph_msg *msg);
254
Sage Weil31b80062009-10-06 11:31:13 -0700255#endif