net/smc/smc_clc.c

   1 /*
   2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   3  *
   4  *  CLC (connection layer control) handshake over initial TCP socket to
   5  *  prepare for RDMA traffic
   6  *
   7  *  Copyright IBM Corp. 2016
   8  *
   9  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  10  */
  11
  12 #include <linux/in.h>
  13 #include <net/sock.h>
  14 #include <net/tcp.h>
  15
  16 #include "smc.h"
  17 #include "smc_core.h"
  18 #include "smc_clc.h"
  19 #include "smc_ib.h"
  20
  21 /* Wait for data on the tcp-socket, analyze received data
  22  * Returns:
  23  * 0 if success and it was not a decline that we received.
  24  * SMC_CLC_DECL_REPLY if decline received for fallback w/o another decl send.
  25  * clcsock error, -EINTR, -ECONNRESET, -EPROTO otherwise.
  26  */
  27 int smc_clc_wait_msg(struct smc_sock *smc, void *buf, int buflen,
  28                      u8 expected_type)
  29 {
  30         struct sock *clc_sk = smc->clcsock->sk;
  31         struct smc_clc_msg_hdr *clcm = buf;
  32         struct msghdr msg = {NULL, 0};
  33         int reason_code = 0;
  34         struct kvec vec;
  35         int len, datlen;
  36         int krflags;
  37
  38         /* peek the first few bytes to determine length of data to receive
  39          * so we don't consume any subsequent CLC message or payload data
  40          * in the TCP byte stream
  41          */
  42         vec.iov_base = buf;
  43         vec.iov_len = buflen;
  44         krflags = MSG_PEEK | MSG_WAITALL;
  45         smc->clcsock->sk->sk_rcvtimeo = CLC_WAIT_TIME;
  46         len = kernel_recvmsg(smc->clcsock, &msg, &vec, 1,
  47                              sizeof(struct smc_clc_msg_hdr), krflags);
  48         if (signal_pending(current)) {
  49                 reason_code = -EINTR;
  50                 clc_sk->sk_err = EINTR;
  51                 smc->sk.sk_err = EINTR;
  52                 goto out;
  53         }
  54         if (clc_sk->sk_err) {
  55                 reason_code = -clc_sk->sk_err;
  56                 smc->sk.sk_err = clc_sk->sk_err;
  57                 goto out;
  58         }
  59         if (!len) { /* peer has performed orderly shutdown */
  60                 smc->sk.sk_err = ECONNRESET;
  61                 reason_code = -ECONNRESET;
  62                 goto out;
  63         }
  64         if (len < 0) {
  65                 smc->sk.sk_err = -len;
  66                 reason_code = len;
  67                 goto out;
  68         }
  69         datlen = ntohs(clcm->length);
  70         if ((len < sizeof(struct smc_clc_msg_hdr)) ||
  71             (datlen < sizeof(struct smc_clc_msg_decline)) ||
  72             (datlen > sizeof(struct smc_clc_msg_accept_confirm)) ||
  73             memcmp(clcm->eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER)) ||
  74             ((clcm->type != SMC_CLC_DECLINE) &&
  75              (clcm->type != expected_type))) {
  76                 smc->sk.sk_err = EPROTO;
  77                 reason_code = -EPROTO;
  78                 goto out;
  79         }
  80
  81         /* receive the complete CLC message */
  82         vec.iov_base = buf;
  83         vec.iov_len = buflen;
  84         memset(&msg, 0, sizeof(struct msghdr));
  85         krflags = MSG_WAITALL;
  86         smc->clcsock->sk->sk_rcvtimeo = CLC_WAIT_TIME;
  87         len = kernel_recvmsg(smc->clcsock, &msg, &vec, 1, datlen, krflags);
  88         if (len < datlen) {
  89                 smc->sk.sk_err = EPROTO;
  90                 reason_code = -EPROTO;
  91                 goto out;
  92         }
  93         if (clcm->type == SMC_CLC_DECLINE) {
  94                 reason_code = SMC_CLC_DECL_REPLY;
  95                 if (ntohl(((struct smc_clc_msg_decline *)buf)->peer_diagnosis)
  96                         == SMC_CLC_DECL_SYNCERR)
  97                         smc->conn.lgr->sync_err = true;
  98         }
  99
 100 out:
 101         return reason_code;
 102 }
 103
 104 /* send CLC DECLINE message across internal TCP socket */
 105 int smc_clc_send_decline(struct smc_sock *smc, u32 peer_diag_info,
 106                          u8 out_of_sync)
 107 {
 108         struct smc_clc_msg_decline dclc;
 109         struct msghdr msg;
 110         struct kvec vec;
 111         int len;
 112
 113         memset(&dclc, 0, sizeof(dclc));
 114         memcpy(dclc.hdr.eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER));
 115         dclc.hdr.type = SMC_CLC_DECLINE;
 116         dclc.hdr.length = htons(sizeof(struct smc_clc_msg_decline));
 117         dclc.hdr.version = SMC_CLC_V1;
 118         dclc.hdr.flag = out_of_sync ? 1 : 0;
 119         memcpy(dclc.id_for_peer, local_systemid, sizeof(local_systemid));
 120         dclc.peer_diagnosis = htonl(peer_diag_info);
 121         memcpy(dclc.trl.eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER));
 122
 123         memset(&msg, 0, sizeof(msg));
 124         vec.iov_base = &dclc;
 125         vec.iov_len = sizeof(struct smc_clc_msg_decline);
 126         len = kernel_sendmsg(smc->clcsock, &msg, &vec, 1,
 127                              sizeof(struct smc_clc_msg_decline));
 128         if (len < sizeof(struct smc_clc_msg_decline))
 129                 smc->sk.sk_err = EPROTO;
 130         if (len < 0)
 131                 smc->sk.sk_err = -len;
 132         return len;
 133 }
 134
 135 /* send CLC PROPOSAL message across internal TCP socket */
 136 int smc_clc_send_proposal(struct smc_sock *smc,
 137                           struct smc_ib_device *smcibdev,
 138                           u8 ibport)
 139 {
 140         struct smc_clc_msg_proposal pclc;
 141         int reason_code = 0;
 142         struct msghdr msg;
 143         struct kvec vec;
 144         int len, rc;
 145
 146         /* send SMC Proposal CLC message */
 147         memset(&pclc, 0, sizeof(pclc));
 148         memcpy(pclc.hdr.eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER));
 149         pclc.hdr.type = SMC_CLC_PROPOSAL;
 150         pclc.hdr.length = htons(sizeof(pclc));
 151         pclc.hdr.version = SMC_CLC_V1;          /* SMC version */
 152         memcpy(pclc.lcl.id_for_peer, local_systemid, sizeof(local_systemid));
 153         memcpy(&pclc.lcl.gid, &smcibdev->gid[ibport - 1], SMC_GID_SIZE);
 154         memcpy(&pclc.lcl.mac, &smcibdev->mac[ibport - 1],
 155                sizeof(smcibdev->mac[ibport - 1]));
 156
 157         /* determine subnet and mask from internal TCP socket */
 158         rc = smc_netinfo_by_tcpsk(smc->clcsock, &pclc.outgoing_subnet,
 159                                   &pclc.prefix_len);
 160         if (rc)
 161                 return SMC_CLC_DECL_CNFERR; /* configuration error */
 162         memcpy(pclc.trl.eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER));
 163         memset(&msg, 0, sizeof(msg));
 164         vec.iov_base = &pclc;
 165         vec.iov_len = sizeof(pclc);
 166         /* due to the few bytes needed for clc-handshake this cannot block */
 167         len = kernel_sendmsg(smc->clcsock, &msg, &vec, 1, sizeof(pclc));
 168         if (len < sizeof(pclc)) {
 169                 if (len >= 0) {
 170                         reason_code = -ENETUNREACH;
 171                         smc->sk.sk_err = -reason_code;
 172                 } else {
 173                         smc->sk.sk_err = smc->clcsock->sk->sk_err;
 174                         reason_code = -smc->sk.sk_err;
 175                 }
 176         }
 177
 178         return reason_code;
 179 }
 180
 181 /* send CLC CONFIRM message across internal TCP socket */
 182 int smc_clc_send_confirm(struct smc_sock *smc)
 183 {
 184         struct smc_connection *conn = &smc->conn;
 185         struct smc_clc_msg_accept_confirm cclc;
 186         struct smc_link *link;
 187         int reason_code = 0;
 188         struct msghdr msg;
 189         struct kvec vec;
 190         int len;
 191
 192         link = &conn->lgr->lnk[SMC_SINGLE_LINK];
 193         /* send SMC Confirm CLC msg */
 194         memset(&cclc, 0, sizeof(cclc));
 195         memcpy(cclc.hdr.eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER));
 196         cclc.hdr.type = SMC_CLC_CONFIRM;
 197         cclc.hdr.length = htons(sizeof(cclc));
 198         cclc.hdr.version = SMC_CLC_V1;          /* SMC version */
 199         memcpy(cclc.lcl.id_for_peer, local_systemid, sizeof(local_systemid));
 200         memcpy(&cclc.lcl.gid, &link->smcibdev->gid[link->ibport - 1],
 201                SMC_GID_SIZE);
 202         memcpy(&cclc.lcl.mac, &link->smcibdev->mac[link->ibport - 1],
 203                sizeof(link->smcibdev->mac));
 204
 205         /* tbd in follow-on patch: fill in rmb-related values */
 206
 207         hton24(cclc.qpn, link->roce_qp->qp_num);
 208         cclc.conn_idx = 1; /* for now: 1 RMB = 1 RMBE */
 209         cclc.rmbe_alert_token = htonl(conn->alert_token_local);
 210         cclc.qp_mtu = min(link->path_mtu, link->peer_mtu);
 211         hton24(cclc.psn, link->psn_initial);
 212
 213         memcpy(cclc.trl.eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER));
 214
 215         memset(&msg, 0, sizeof(msg));
 216         vec.iov_base = &cclc;
 217         vec.iov_len = sizeof(cclc);
 218         len = kernel_sendmsg(smc->clcsock, &msg, &vec, 1, sizeof(cclc));
 219         if (len < sizeof(cclc)) {
 220                 if (len >= 0) {
 221                         reason_code = -ENETUNREACH;
 222                         smc->sk.sk_err = -reason_code;
 223                 } else {
 224                         smc->sk.sk_err = smc->clcsock->sk->sk_err;
 225                         reason_code = -smc->sk.sk_err;
 226                 }
 227         }
 228         return reason_code;
 229 }
 230
 231 /* send CLC ACCEPT message across internal TCP socket */
 232 int smc_clc_send_accept(struct smc_sock *new_smc, int srv_first_contact)
 233 {
 234         struct smc_connection *conn = &new_smc->conn;
 235         struct smc_clc_msg_accept_confirm aclc;
 236         struct smc_link *link;
 237         struct msghdr msg;
 238         struct kvec vec;
 239         int rc = 0;
 240         int len;
 241
 242         link = &conn->lgr->lnk[SMC_SINGLE_LINK];
 243         memset(&aclc, 0, sizeof(aclc));
 244         memcpy(aclc.hdr.eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER));
 245         aclc.hdr.type = SMC_CLC_ACCEPT;
 246         aclc.hdr.length = htons(sizeof(aclc));
 247         aclc.hdr.version = SMC_CLC_V1;          /* SMC version */
 248         if (srv_first_contact)
 249                 aclc.hdr.flag = 1;
 250         memcpy(aclc.lcl.id_for_peer, local_systemid, sizeof(local_systemid));
 251         memcpy(&aclc.lcl.gid, &link->smcibdev->gid[link->ibport - 1],
 252                SMC_GID_SIZE);
 253         memcpy(&aclc.lcl.mac, link->smcibdev->mac[link->ibport - 1],
 254                sizeof(link->smcibdev->mac[link->ibport - 1]));
 255
 256         /* tbd in follow-on patch: fill in rmb-related values */
 257
 258         hton24(aclc.qpn, link->roce_qp->qp_num);
 259         aclc.conn_idx = 1;                      /* as long as 1 RMB = 1 RMBE */
 260         aclc.rmbe_alert_token = htonl(conn->alert_token_local);
 261         aclc.qp_mtu = link->path_mtu;
 262         hton24(aclc.psn, link->psn_initial);
 263         memcpy(aclc.trl.eyecatcher, SMC_EYECATCHER, sizeof(SMC_EYECATCHER));
 264
 265         memset(&msg, 0, sizeof(msg));
 266         vec.iov_base = &aclc;
 267         vec.iov_len = sizeof(aclc);
 268         len = kernel_sendmsg(new_smc->clcsock, &msg, &vec, 1, sizeof(aclc));
 269         if (len < sizeof(aclc)) {
 270                 if (len >= 0)
 271                         new_smc->sk.sk_err = EPROTO;
 272                 else
 273                         new_smc->sk.sk_err = new_smc->clcsock->sk->sk_err;
 274                 rc = sock_error(&new_smc->sk);
 275         }
 276
 277         return rc;
 278 }