include/linux/ceph/messenger.h

   1 #ifndef __FS_CEPH_MESSENGER_H
   2 #define __FS_CEPH_MESSENGER_H
   3
   4 #include <linux/kref.h>
   5 #include <linux/mutex.h>
   6 #include <linux/net.h>
   7 #include <linux/radix-tree.h>
   8 #include <linux/uio.h>
   9 #include <linux/workqueue.h>
  10
  11 #include <linux/ceph/types.h>
  12 #include <linux/ceph/buffer.h>
  13
  14 struct ceph_msg;
  15 struct ceph_connection;
  16
  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
  27         /* authorize an outgoing connection */
  28         struct ceph_auth_handshake *(*get_authorizer) (
  29                                 struct ceph_connection *con,
  30                                int *proto, int force_new);
  31         int (*verify_authorizer_reply) (struct ceph_connection *con, int len);
  32         int (*invalidate_authorizer)(struct ceph_connection *con);
  33
  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,
  42                                         struct ceph_msg_header *hdr,
  43                                         int *skip);
  44 };
  45
  46 /* use format string %s%d */
  47 #define ENTITY_NAME(n) ceph_entity_type_name((n).type), le64_to_cpu((n).num)
  48
  49 struct ceph_messenger {
  50         struct ceph_entity_inst inst;    /* my name+address */
  51         struct ceph_entity_addr my_enc_addr;
  52
  53         atomic_t stopping;
  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;
  62
  63         u32 supported_features;
  64         u32 required_features;
  65 };
  66
  67 #define ceph_msg_has_pages(m)           ((m)->p.type == CEPH_MSG_DATA_PAGES)
  68 #define ceph_msg_has_pagelist(m)        ((m)->l.type == CEPH_MSG_DATA_PAGELIST)
  69 #ifdef CONFIG_BLOCK
  70 #define ceph_msg_has_bio(m)             ((m)->b.type == CEPH_MSG_DATA_BIO)
  71 #endif /* CONFIG_BLOCK */
  72
  73 enum ceph_msg_data_type {
  74         CEPH_MSG_DATA_NONE,     /* message contains no data payload */
  75         CEPH_MSG_DATA_PAGES,    /* data source/destination is a page array */
  76         CEPH_MSG_DATA_PAGELIST, /* data source/destination is a pagelist */
  77 #ifdef CONFIG_BLOCK
  78         CEPH_MSG_DATA_BIO,      /* data source/destination is a bio list */
  79 #endif /* CONFIG_BLOCK */
  80 };
  81
  82 static __inline__ bool ceph_msg_data_type_valid(enum ceph_msg_data_type type)
  83 {
  84         switch (type) {
  85         case CEPH_MSG_DATA_NONE:
  86         case CEPH_MSG_DATA_PAGES:
  87         case CEPH_MSG_DATA_PAGELIST:
  88 #ifdef CONFIG_BLOCK
  89         case CEPH_MSG_DATA_BIO:
  90 #endif /* CONFIG_BLOCK */
  91                 return true;
  92         default:
  93                 return false;
  94         }
  95 }
  96
  97 struct ceph_msg_data_cursor {
  98         size_t          resid;          /* bytes not yet consumed */
  99         bool            last_piece;     /* now at last piece of data item */
 100         union {
 101 #ifdef CONFIG_BLOCK
 102                 struct {                                /* bio */
 103                         struct bio      *bio;           /* bio from list */
 104                         unsigned int    vector_index;   /* vector from bio */
 105                         unsigned int    vector_offset;  /* bytes from vector */
 106                 };
 107 #endif /* CONFIG_BLOCK */
 108                 struct {                                /* pages */
 109                         unsigned int    page_offset;    /* offset in page */
 110                         unsigned short  page_index;     /* index in array */
 111                         unsigned short  page_count;     /* pages in array */
 112                 };
 113                 struct {                                /* pagelist */
 114                         struct page     *page;          /* page from list */
 115                         size_t          offset;         /* bytes from list */
 116                 };
 117         };
 118 };
 119
 120 struct ceph_msg_data {
 121         enum ceph_msg_data_type         type;
 122         union {
 123 #ifdef CONFIG_BLOCK
 124                 struct bio              *bio;
 125 #endif /* CONFIG_BLOCK */
 126                 struct {
 127                         struct page     **pages;        /* NOT OWNER. */
 128                         size_t          length;         /* total # bytes */
 129                         unsigned int    alignment;      /* first page */
 130                 };
 131                 struct ceph_pagelist    *pagelist;
 132         };
 133         struct ceph_msg_data_cursor     cursor;         /* pagelist only */
 134 };
 135
 136 /*
 137  * a single message.  it contains a header (src, dest, message type, etc.),
 138  * footer (crc values, mainly), a "front" message body, and possibly a
 139  * data payload (stored in some number of pages).
 140  */
 141 struct ceph_msg {
 142         struct ceph_msg_header hdr;     /* header */
 143         struct ceph_msg_footer footer;  /* footer */
 144         struct kvec front;              /* unaligned blobs of message */
 145         struct ceph_buffer *middle;
 146
 147         /* data payload */
 148         struct ceph_msg_data    p;      /* pages */
 149         struct ceph_msg_data    l;      /* pagelist */
 150 #ifdef CONFIG_BLOCK
 151         struct ceph_msg_data    b;      /* bio */
 152 #endif /* CONFIG_BLOCK */
 153
 154         struct ceph_connection *con;
 155         struct list_head list_head;     /* links for connection lists */
 156
 157         struct kref kref;
 158         bool front_is_vmalloc;
 159         bool more_to_follow;
 160         bool needs_out_seq;
 161         int front_max;
 162         unsigned long ack_stamp;        /* tx: when we were acked */
 163
 164         struct ceph_msgpool *pool;
 165 };
 166
 167 struct ceph_msg_pos {
 168         int page, page_pos;  /* which page; offset in page */
 169         int data_pos;        /* offset in data payload */
 170         bool did_page_crc;   /* true if we've calculated crc for current page */
 171 };
 172
 173 /* ceph connection fault delay defaults, for exponential backoff */
 174 #define BASE_DELAY_INTERVAL     (HZ/2)
 175 #define MAX_DELAY_INTERVAL      (5 * 60 * HZ)
 176
 177 /*
 178  * A single connection with another host.
 179  *
 180  * We maintain a queue of outgoing messages, and some session state to
 181  * ensure that we can preserve the lossless, ordered delivery of
 182  * messages in the case of a TCP disconnect.
 183  */
 184 struct ceph_connection {
 185         void *private;
 186
 187         const struct ceph_connection_operations *ops;
 188
 189         struct ceph_messenger *msgr;
 190
 191         atomic_t sock_state;
 192         struct socket *sock;
 193         struct ceph_entity_addr peer_addr; /* peer address */
 194         struct ceph_entity_addr peer_addr_for_me;
 195
 196         unsigned long flags;
 197         unsigned long state;
 198         const char *error_msg;  /* error message, if any */
 199
 200         struct ceph_entity_name peer_name; /* peer name */
 201
 202         unsigned peer_features;
 203         u32 connect_seq;      /* identify the most recent connection
 204                                  attempt for this connection, client */
 205         u32 peer_global_seq;  /* peer's global seq for this connection */
 206
 207         int auth_retry;       /* true if we need a newer authorizer */
 208         void *auth_reply_buf;   /* where to put the authorizer reply */
 209         int auth_reply_buf_len;
 210
 211         struct mutex mutex;
 212
 213         /* out queue */
 214         struct list_head out_queue;
 215         struct list_head out_sent;   /* sending or sent but unacked */
 216         u64 out_seq;                 /* last message queued for send */
 217
 218         u64 in_seq, in_seq_acked;  /* last message received, acked */
 219
 220         /* connection negotiation temps */
 221         char in_banner[CEPH_BANNER_MAX_LEN];
 222         struct ceph_msg_connect out_connect;
 223         struct ceph_msg_connect_reply in_reply;
 224         struct ceph_entity_addr actual_peer_addr;
 225
 226         /* message out temps */
 227         struct ceph_msg *out_msg;        /* sending message (== tail of
 228                                             out_sent) */
 229         bool out_msg_done;
 230         struct ceph_msg_pos out_msg_pos;
 231
 232         struct kvec out_kvec[8],         /* sending header/footer data */
 233                 *out_kvec_cur;
 234         int out_kvec_left;   /* kvec's left in out_kvec */
 235         int out_skip;        /* skip this many bytes */
 236         int out_kvec_bytes;  /* total bytes left */
 237         bool out_kvec_is_msg; /* kvec refers to out_msg */
 238         int out_more;        /* there is more data after the kvecs */
 239         __le64 out_temp_ack; /* for writing an ack */
 240
 241         /* message in temps */
 242         struct ceph_msg_header in_hdr;
 243         struct ceph_msg *in_msg;
 244         struct ceph_msg_pos in_msg_pos;
 245         u32 in_front_crc, in_middle_crc, in_data_crc;  /* calculated crc */
 246
 247         char in_tag;         /* protocol control byte */
 248         int in_base_pos;     /* bytes read */
 249         __le64 in_temp_ack;  /* for reading an ack */
 250
 251         struct delayed_work work;           /* send|recv work */
 252         unsigned long       delay;          /* current delay interval */
 253 };
 254
 255
 256 extern const char *ceph_pr_addr(const struct sockaddr_storage *ss);
 257 extern int ceph_parse_ips(const char *c, const char *end,
 258                           struct ceph_entity_addr *addr,
 259                           int max_count, int *count);
 260
 261
 262 extern int ceph_msgr_init(void);
 263 extern void ceph_msgr_exit(void);
 264 extern void ceph_msgr_flush(void);
 265
 266 extern void ceph_messenger_init(struct ceph_messenger *msgr,
 267                         struct ceph_entity_addr *myaddr,
 268                         u32 supported_features,
 269                         u32 required_features,
 270                         bool nocrc);
 271
 272 extern void ceph_con_init(struct ceph_connection *con, void *private,
 273                         const struct ceph_connection_operations *ops,
 274                         struct ceph_messenger *msgr);
 275 extern void ceph_con_open(struct ceph_connection *con,
 276                           __u8 entity_type, __u64 entity_num,
 277                           struct ceph_entity_addr *addr);
 278 extern bool ceph_con_opened(struct ceph_connection *con);
 279 extern void ceph_con_close(struct ceph_connection *con);
 280 extern void ceph_con_send(struct ceph_connection *con, struct ceph_msg *msg);
 281
 282 extern void ceph_msg_revoke(struct ceph_msg *msg);
 283 extern void ceph_msg_revoke_incoming(struct ceph_msg *msg);
 284
 285 extern void ceph_con_keepalive(struct ceph_connection *con);
 286
 287 extern void ceph_msg_data_set_pages(struct ceph_msg *msg, struct page **pages,
 288                                 size_t length, size_t alignment);
 289 extern void ceph_msg_data_set_pagelist(struct ceph_msg *msg,
 290                                 struct ceph_pagelist *pagelist);
 291 extern void ceph_msg_data_set_bio(struct ceph_msg *msg, struct bio *bio);
 292
 293 extern struct ceph_msg *ceph_msg_new(int type, int front_len, gfp_t flags,
 294                                      bool can_fail);
 295 extern void ceph_msg_kfree(struct ceph_msg *m);
 296
 297
 298 static inline struct ceph_msg *ceph_msg_get(struct ceph_msg *msg)
 299 {
 300         kref_get(&msg->kref);
 301         return msg;
 302 }
 303 extern void ceph_msg_last_put(struct kref *kref);
 304 static inline void ceph_msg_put(struct ceph_msg *msg)
 305 {
 306         kref_put(&msg->kref, ceph_msg_last_put);
 307 }
 308
 309 extern void ceph_msg_dump(struct ceph_msg *msg);
 310
 311 #endif