blob: eb31e108a64df95dd63a090561cd3fb31fed4ead [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
Sage Weil31b80062009-10-06 11:31:13 -070052/* use format string %s%d */
Sage Weilca9d93a2010-05-12 14:48:20 -070053#define ENTITY_NAME(n) ceph_entity_type_name((n).type), le64_to_cpu((n).num)
Sage Weil31b80062009-10-06 11:31:13 -070054
55struct ceph_messenger {
56 struct ceph_entity_inst inst; /* my name+address */
Sage Weil63f2d212009-11-03 15:17:56 -080057 struct ceph_entity_addr my_enc_addr;
Sage Weil31b80062009-10-06 11:31:13 -070058 struct page *zero_page; /* used in certain error cases */
59
60 bool nocrc;
61
62 /*
63 * the global_seq counts connections i (attempt to) initiate
64 * in order to disambiguate certain connect race conditions.
65 */
66 u32 global_seq;
67 spinlock_t global_seq_lock;
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -070068
69 u32 supported_features;
70 u32 required_features;
Sage Weil31b80062009-10-06 11:31:13 -070071};
72
73/*
74 * a single message. it contains a header (src, dest, message type, etc.),
75 * footer (crc values, mainly), a "front" message body, and possibly a
76 * data payload (stored in some number of pages).
77 */
78struct ceph_msg {
79 struct ceph_msg_header hdr; /* header */
80 struct ceph_msg_footer footer; /* footer */
81 struct kvec front; /* unaligned blobs of message */
82 struct ceph_buffer *middle;
83 struct page **pages; /* data payload. NOT OWNER. */
84 unsigned nr_pages; /* size of page array */
Sage Weilc5c6b192010-11-09 12:40:00 -080085 unsigned page_alignment; /* io offset in first page */
Sage Weil58bb3b32009-12-23 12:12:31 -080086 struct ceph_pagelist *pagelist; /* instead of pages */
Sage Weil31b80062009-10-06 11:31:13 -070087 struct list_head list_head;
Sage Weilc2e552e2009-12-07 15:55:05 -080088 struct kref kref;
Yehuda Sadeh68b44762010-04-06 15:01:27 -070089 struct bio *bio; /* instead of pages/pagelist */
90 struct bio *bio_iter; /* bio iterator */
91 int bio_seg; /* current bio segment */
92 struct ceph_pagelist *trail; /* the trailing part of the data */
Sage Weil31b80062009-10-06 11:31:13 -070093 bool front_is_vmalloc;
94 bool more_to_follow;
Sage Weile84346b2010-05-11 21:20:38 -070095 bool needs_out_seq;
Sage Weil31b80062009-10-06 11:31:13 -070096 int front_max;
97
98 struct ceph_msgpool *pool;
99};
100
101struct ceph_msg_pos {
102 int page, page_pos; /* which page; offset in page */
103 int data_pos; /* offset in data payload */
104 int did_page_crc; /* true if we've calculated crc for current page */
105};
106
107/* ceph connection fault delay defaults, for exponential backoff */
108#define BASE_DELAY_INTERVAL (HZ/2)
109#define MAX_DELAY_INTERVAL (5 * 60 * HZ)
110
111/*
112 * ceph_connection state bit flags
Sage Weil31b80062009-10-06 11:31:13 -0700113 */
114#define LOSSYTX 0 /* we can close channel or drop messages on errors */
Sage Weileed0ef22009-11-10 14:34:36 -0800115#define CONNECTING 1
116#define NEGOTIATING 2
Sage Weil31b80062009-10-06 11:31:13 -0700117#define KEEPALIVE_PENDING 3
118#define WRITE_PENDING 4 /* we have data ready to send */
Sage Weil31b80062009-10-06 11:31:13 -0700119#define STANDBY 8 /* no outgoing messages, socket closed. we keep
120 * the ceph_connection around to maintain shared
121 * state with the peer. */
122#define CLOSED 10 /* we've closed the connection */
123#define SOCK_CLOSED 11 /* socket state changed to closed */
Sage Weil31b80062009-10-06 11:31:13 -0700124#define OPENING 13 /* open connection w/ (possibly new) peer */
125#define DEAD 14 /* dead, about to kfree */
Sage Weil60bf8bf2011-03-04 12:24:28 -0800126#define BACKOFF 15
Sage Weil31b80062009-10-06 11:31:13 -0700127
128/*
129 * A single connection with another host.
130 *
131 * We maintain a queue of outgoing messages, and some session state to
132 * ensure that we can preserve the lossless, ordered delivery of
133 * messages in the case of a TCP disconnect.
134 */
135struct ceph_connection {
136 void *private;
137 atomic_t nref;
138
139 const struct ceph_connection_operations *ops;
140
141 struct ceph_messenger *msgr;
142 struct socket *sock;
143 unsigned long state; /* connection state (see flags above) */
144 const char *error_msg; /* error message, if any */
145
146 struct ceph_entity_addr peer_addr; /* peer address */
147 struct ceph_entity_name peer_name; /* peer name */
148 struct ceph_entity_addr peer_addr_for_me;
Sage Weilaba558e2010-05-12 15:23:30 -0700149 unsigned peer_features;
Sage Weil31b80062009-10-06 11:31:13 -0700150 u32 connect_seq; /* identify the most recent connection
151 attempt for this connection, client */
152 u32 peer_global_seq; /* peer's global seq for this connection */
153
Sage Weil4e7a5dc2009-11-18 16:19:57 -0800154 int auth_retry; /* true if we need a newer authorizer */
155 void *auth_reply_buf; /* where to put the authorizer reply */
156 int auth_reply_buf_len;
157
Sage Weilec302642009-12-22 10:43:42 -0800158 struct mutex mutex;
159
Sage Weil31b80062009-10-06 11:31:13 -0700160 /* out queue */
Sage Weil31b80062009-10-06 11:31:13 -0700161 struct list_head out_queue;
162 struct list_head out_sent; /* sending or sent but unacked */
163 u64 out_seq; /* last message queued for send */
Sage Weil31b80062009-10-06 11:31:13 -0700164 bool out_keepalive_pending;
165
166 u64 in_seq, in_seq_acked; /* last message received, acked */
167
168 /* connection negotiation temps */
169 char in_banner[CEPH_BANNER_MAX_LEN];
170 union {
171 struct { /* outgoing connection */
172 struct ceph_msg_connect out_connect;
173 struct ceph_msg_connect_reply in_reply;
174 };
175 struct { /* incoming */
176 struct ceph_msg_connect in_connect;
177 struct ceph_msg_connect_reply out_reply;
178 };
179 };
180 struct ceph_entity_addr actual_peer_addr;
181
182 /* message out temps */
183 struct ceph_msg *out_msg; /* sending message (== tail of
184 out_sent) */
Sage Weilc86a2932009-12-14 14:04:30 -0800185 bool out_msg_done;
Sage Weil31b80062009-10-06 11:31:13 -0700186 struct ceph_msg_pos out_msg_pos;
187
188 struct kvec out_kvec[8], /* sending header/footer data */
189 *out_kvec_cur;
190 int out_kvec_left; /* kvec's left in out_kvec */
191 int out_skip; /* skip this many bytes */
192 int out_kvec_bytes; /* total bytes left */
193 bool out_kvec_is_msg; /* kvec refers to out_msg */
194 int out_more; /* there is more data after the kvecs */
195 __le64 out_temp_ack; /* for writing an ack */
196
197 /* message in temps */
198 struct ceph_msg_header in_hdr;
199 struct ceph_msg *in_msg;
200 struct ceph_msg_pos in_msg_pos;
201 u32 in_front_crc, in_middle_crc, in_data_crc; /* calculated crc */
202
203 char in_tag; /* protocol control byte */
204 int in_base_pos; /* bytes read */
205 __le64 in_temp_ack; /* for reading an ack */
206
207 struct delayed_work work; /* send|recv work */
208 unsigned long delay; /* current delay interval */
209};
210
211
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700212extern const char *ceph_pr_addr(const struct sockaddr_storage *ss);
Sage Weil31b80062009-10-06 11:31:13 -0700213extern int ceph_parse_ips(const char *c, const char *end,
214 struct ceph_entity_addr *addr,
215 int max_count, int *count);
216
217
218extern int ceph_msgr_init(void);
219extern void ceph_msgr_exit(void);
Sage Weila922d382010-05-29 09:41:23 -0700220extern void ceph_msgr_flush(void);
Sage Weil31b80062009-10-06 11:31:13 -0700221
222extern struct ceph_messenger *ceph_messenger_create(
Yehuda Sadeh3d14c5d2010-04-06 15:14:15 -0700223 struct ceph_entity_addr *myaddr,
224 u32 features, u32 required);
Sage Weil31b80062009-10-06 11:31:13 -0700225extern void ceph_messenger_destroy(struct ceph_messenger *);
226
227extern void ceph_con_init(struct ceph_messenger *msgr,
228 struct ceph_connection *con);
Sage Weil31b80062009-10-06 11:31:13 -0700229extern void ceph_con_open(struct ceph_connection *con,
230 struct ceph_entity_addr *addr);
Sage Weil87b315a2010-03-22 14:51:18 -0700231extern bool ceph_con_opened(struct ceph_connection *con);
Sage Weil31b80062009-10-06 11:31:13 -0700232extern void ceph_con_close(struct ceph_connection *con);
233extern void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg);
234extern void ceph_con_revoke(struct ceph_connection *con, struct ceph_msg *msg);
Yehuda Sadeh0d59ab82010-01-13 17:03:23 -0800235extern void ceph_con_revoke_message(struct ceph_connection *con,
236 struct ceph_msg *msg);
Sage Weil31b80062009-10-06 11:31:13 -0700237extern void ceph_con_keepalive(struct ceph_connection *con);
238extern struct ceph_connection *ceph_con_get(struct ceph_connection *con);
239extern void ceph_con_put(struct ceph_connection *con);
240
Yehuda Sadeh34d23762010-04-06 14:33:58 -0700241extern struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags);
Sage Weil31b80062009-10-06 11:31:13 -0700242extern void ceph_msg_kfree(struct ceph_msg *m);
243
Sage Weil31b80062009-10-06 11:31:13 -0700244
245static inline struct ceph_msg *ceph_msg_get(struct ceph_msg *msg)
246{
Sage Weilc2e552e2009-12-07 15:55:05 -0800247 kref_get(&msg->kref);
Sage Weil31b80062009-10-06 11:31:13 -0700248 return msg;
249}
Sage Weilc2e552e2009-12-07 15:55:05 -0800250extern void ceph_msg_last_put(struct kref *kref);
251static inline void ceph_msg_put(struct ceph_msg *msg)
252{
253 kref_put(&msg->kref, ceph_msg_last_put);
254}
Sage Weil31b80062009-10-06 11:31:13 -0700255
Sage Weil9ec7cab2009-12-14 15:13:47 -0800256extern void ceph_msg_dump(struct ceph_msg *msg);
257
Sage Weil31b80062009-10-06 11:31:13 -0700258#endif