net/smc/smc.h

   1 /*
   2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   3  *
   4  *  Definitions for the SMC module (socket related)
   5  *
   6  *  Copyright IBM Corp. 2016
   7  *
   8  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
   9  */
  10 #ifndef __SMC_H
  11 #define __SMC_H
  12
  13 #include <linux/socket.h>
  14 #include <linux/types.h>
  15 #include <linux/compiler.h> /* __aligned */
  16 #include <net/sock.h>
  17
  18 #include "smc_ib.h"
  19
  20 #define SMCPROTO_SMC            0       /* SMC protocol */
  21
  22 #define SMC_MAX_PORTS           2       /* Max # of ports */
  23
  24 extern struct proto smc_proto;
  25
  26 #ifdef ATOMIC64_INIT
  27 #define KERNEL_HAS_ATOMIC64
  28 #endif
  29
  30 enum smc_state {                /* possible states of an SMC socket */
  31         SMC_ACTIVE      = 1,
  32         SMC_INIT        = 2,
  33         SMC_CLOSED      = 7,
  34         SMC_LISTEN      = 10,
  35         /* normal close */
  36         SMC_PEERCLOSEWAIT1      = 20,
  37         SMC_PEERCLOSEWAIT2      = 21,
  38         SMC_APPFINCLOSEWAIT     = 24,
  39         SMC_APPCLOSEWAIT1       = 22,
  40         SMC_APPCLOSEWAIT2       = 23,
  41         SMC_PEERFINCLOSEWAIT    = 25,
  42         /* abnormal close */
  43         SMC_PEERABORTWAIT       = 26,
  44         SMC_PROCESSABORT        = 27,
  45 };
  46
  47 struct smc_link_group;
  48
  49 struct smc_wr_rx_hdr {  /* common prefix part of LLC and CDC to demultiplex */
  50         u8                      type;
  51 } __aligned(1);
  52
  53 struct smc_cdc_conn_state_flags {
  54 #if defined(__BIG_ENDIAN_BITFIELD)
  55         u8      peer_done_writing : 1;  /* Sending done indicator */
  56         u8      peer_conn_closed : 1;   /* Peer connection closed indicator */
  57         u8      peer_conn_abort : 1;    /* Abnormal close indicator */
  58         u8      reserved : 5;
  59 #elif defined(__LITTLE_ENDIAN_BITFIELD)
  60         u8      reserved : 5;
  61         u8      peer_conn_abort : 1;
  62         u8      peer_conn_closed : 1;
  63         u8      peer_done_writing : 1;
  64 #endif
  65 };
  66
  67 struct smc_cdc_producer_flags {
  68 #if defined(__BIG_ENDIAN_BITFIELD)
  69         u8      write_blocked : 1;      /* Writing Blocked, no rx buf space */
  70         u8      urg_data_pending : 1;   /* Urgent Data Pending */
  71         u8      urg_data_present : 1;   /* Urgent Data Present */
  72         u8      cons_curs_upd_req : 1;  /* cursor update requested */
  73         u8      failover_validation : 1;/* message replay due to failover */
  74         u8      reserved : 3;
  75 #elif defined(__LITTLE_ENDIAN_BITFIELD)
  76         u8      reserved : 3;
  77         u8      failover_validation : 1;
  78         u8      cons_curs_upd_req : 1;
  79         u8      urg_data_present : 1;
  80         u8      urg_data_pending : 1;
  81         u8      write_blocked : 1;
  82 #endif
  83 };
  84
  85 /* in host byte order */
  86 union smc_host_cursor { /* SMC cursor - an offset in an RMBE */
  87         struct {
  88                 u16     reserved;
  89                 u16     wrap;           /* window wrap sequence number */
  90                 u32     count;          /* cursor (= offset) part */
  91         };
  92 #ifdef KERNEL_HAS_ATOMIC64
  93         atomic64_t              acurs;  /* for atomic processing */
  94 #else
  95         u64                     acurs;  /* for atomic processing */
  96 #endif
  97 } __aligned(8);
  98
  99 /* in host byte order, except for flag bitfields in network byte order */
 100 struct smc_host_cdc_msg {               /* Connection Data Control message */
 101         struct smc_wr_rx_hdr            common; /* .type = 0xFE */
 102         u8                              len;    /* length = 44 */
 103         u16                             seqno;  /* connection seq # */
 104         u32                             token;  /* alert_token */
 105         union smc_host_cursor           prod;           /* producer cursor */
 106         union smc_host_cursor           cons;           /* consumer cursor,
 107                                                          * piggy backed "ack"
 108                                                          */
 109         struct smc_cdc_producer_flags   prod_flags;     /* conn. tx/rx status */
 110         struct smc_cdc_conn_state_flags conn_state_flags; /* peer conn. status*/
 111         u8                              reserved[18];
 112 } __aligned(8);
 113
 114 struct smc_connection {
 115         struct rb_node          alert_node;
 116         struct smc_link_group   *lgr;           /* link group of connection */
 117         u32                     alert_token_local; /* unique conn. id */
 118         u8                      peer_conn_idx;  /* from tcp handshake */
 119         int                     peer_rmbe_size; /* size of peer rx buffer */
 120         atomic_t                peer_rmbe_space;/* remaining free bytes in peer
 121                                                  * rmbe
 122                                                  */
 123         int                     rtoken_idx;     /* idx to peer RMB rkey/addr */
 124
 125         struct smc_buf_desc     *sndbuf_desc;   /* send buffer descriptor */
 126         int                     sndbuf_size;    /* sndbuf size <== sock wmem */
 127         struct smc_buf_desc     *rmb_desc;      /* RMBE descriptor */
 128         int                     rmbe_size;      /* RMBE size <== sock rmem */
 129         int                     rmbe_size_short;/* compressed notation */
 130         int                     rmbe_update_limit;
 131                                                 /* lower limit for consumer
 132                                                  * cursor update
 133                                                  */
 134
 135         struct smc_host_cdc_msg local_tx_ctrl;  /* host byte order staging
 136                                                  * buffer for CDC msg send
 137                                                  * .prod cf. TCP snd_nxt
 138                                                  * .cons cf. TCP sends ack
 139                                                  */
 140         union smc_host_cursor   tx_curs_prep;   /* tx - prepared data
 141                                                  * snd_max..wmem_alloc
 142                                                  */
 143         union smc_host_cursor   tx_curs_sent;   /* tx - sent data
 144                                                  * snd_nxt ?
 145                                                  */
 146         union smc_host_cursor   tx_curs_fin;    /* tx - confirmed by peer
 147                                                  * snd-wnd-begin ?
 148                                                  */
 149         atomic_t                sndbuf_space;   /* remaining space in sndbuf */
 150         u16                     tx_cdc_seq;     /* sequence # for CDC send */
 151         spinlock_t              send_lock;      /* protect wr_sends */
 152         struct work_struct      tx_work;        /* retry of smc_cdc_msg_send */
 153
 154         struct smc_host_cdc_msg local_rx_ctrl;  /* filled during event_handl.
 155                                                  * .prod cf. TCP rcv_nxt
 156                                                  * .cons cf. TCP snd_una
 157                                                  */
 158         union smc_host_cursor   rx_curs_confirmed; /* confirmed to peer
 159                                                     * source of snd_una ?
 160                                                     */
 161         atomic_t                bytes_to_rcv;   /* arrived data,
 162                                                  * not yet received
 163                                                  */
 164 #ifndef KERNEL_HAS_ATOMIC64
 165         spinlock_t              acurs_lock;     /* protect cursors */
 166 #endif
 167 };
 168
 169 struct smc_sock {                               /* smc sock container */
 170         struct sock             sk;
 171         struct socket           *clcsock;       /* internal tcp socket */
 172         struct smc_connection   conn;           /* smc connection */
 173         struct sockaddr         *addr;          /* inet connect address */
 174         struct smc_sock         *listen_smc;    /* listen parent */
 175         struct work_struct      tcp_listen_work;/* handle tcp socket accepts */
 176         struct work_struct      smc_listen_work;/* prepare new accept socket */
 177         struct list_head        accept_q;       /* sockets to be accepted */
 178         spinlock_t              accept_q_lock;  /* protects accept_q */
 179         struct delayed_work     sock_put_work;  /* final socket freeing */
 180         bool                    use_fallback;   /* fallback to tcp */
 181         u8                      wait_close_tx_prepared : 1;
 182                                                 /* shutdown wr or close
 183                                                  * started, waiting for unsent
 184                                                  * data to be sent
 185                                                  */
 186 };
 187
 188 static inline struct smc_sock *smc_sk(const struct sock *sk)
 189 {
 190         return (struct smc_sock *)sk;
 191 }
 192
 193 #define SMC_SYSTEMID_LEN                8
 194
 195 extern u8       local_systemid[SMC_SYSTEMID_LEN]; /* unique system identifier */
 196
 197 /* convert an u32 value into network byte order, store it into a 3 byte field */
 198 static inline void hton24(u8 *net, u32 host)
 199 {
 200         __be32 t;
 201
 202         t = cpu_to_be32(host);
 203         memcpy(net, ((u8 *)&t) + 1, 3);
 204 }
 205
 206 /* convert a received 3 byte field into host byte order*/
 207 static inline u32 ntoh24(u8 *net)
 208 {
 209         __be32 t = 0;
 210
 211         memcpy(((u8 *)&t) + 1, net, 3);
 212         return be32_to_cpu(t);
 213 }
 214
 215 #define SMC_BUF_MIN_SIZE 16384          /* minimum size of an RMB */
 216
 217 #define SMC_RMBE_SIZES  16      /* number of distinct sizes for an RMBE */
 218 /* theoretically, the RFC states that largest size would be 512K,
 219  * i.e. compressed 5 and thus 6 sizes (0..5), despite
 220  * struct smc_clc_msg_accept_confirm.rmbe_size being a 4 bit value (0..15)
 221  */
 222
 223 /* convert the RMB size into the compressed notation - minimum 16K.
 224  * In contrast to plain ilog2, this rounds towards the next power of 2,
 225  * so the socket application gets at least its desired sndbuf / rcvbuf size.
 226  */
 227 static inline u8 smc_compress_bufsize(int size)
 228 {
 229         u8 compressed;
 230
 231         if (size <= SMC_BUF_MIN_SIZE)
 232                 return 0;
 233
 234         size = (size - 1) >> 14;
 235         compressed = ilog2(size) + 1;
 236         if (compressed >= SMC_RMBE_SIZES)
 237                 compressed = SMC_RMBE_SIZES - 1;
 238         return compressed;
 239 }
 240
 241 /* convert the RMB size from compressed notation into integer */
 242 static inline int smc_uncompress_bufsize(u8 compressed)
 243 {
 244         u32 size;
 245
 246         size = 0x00000001 << (((int)compressed) + 14);
 247         return (int)size;
 248 }
 249
 250 #ifdef CONFIG_XFRM
 251 static inline bool using_ipsec(struct smc_sock *smc)
 252 {
 253         return (smc->clcsock->sk->sk_policy[0] ||
 254                 smc->clcsock->sk->sk_policy[1]) ? 1 : 0;
 255 }
 256 #else
 257 static inline bool using_ipsec(struct smc_sock *smc)
 258 {
 259         return 0;
 260 }
 261 #endif
 262
 263 struct smc_clc_msg_local;
 264
 265 int smc_netinfo_by_tcpsk(struct socket *clcsock, __be32 *subnet,
 266                          u8 *prefix_len);
 267 void smc_conn_free(struct smc_connection *conn);
 268 int smc_conn_create(struct smc_sock *smc, __be32 peer_in_addr,
 269                     struct smc_ib_device *smcibdev, u8 ibport,
 270                     struct smc_clc_msg_local *lcl, int srv_first_contact);
 271 struct sock *smc_accept_dequeue(struct sock *parent, struct socket *new_sock);
 272 void smc_close_non_accepted(struct sock *sk);
 273
 274 #endif  /* __SMC_H */