net/smc/smc_cdc.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  * Connection Data Control (CDC)
   6  * handles flow control
   7  *
   8  * Copyright IBM Corp. 2016
   9  *
  10  * Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  11  */
  12
  13 #include <linux/spinlock.h>
  14
  15 #include "smc.h"
  16 #include "smc_wr.h"
  17 #include "smc_cdc.h"
  18 #include "smc_tx.h"
  19 #include "smc_rx.h"
  20 #include "smc_close.h"
  21
  22 /********************************** send *************************************/
  23
  24 /* handler for send/transmission completion of a CDC msg */
  25 static void smc_cdc_tx_handler(struct smc_wr_tx_pend_priv *pnd_snd,
  26                                struct smc_link *link,
  27                                enum ib_wc_status wc_status)
  28 {
  29         struct smc_cdc_tx_pend *cdcpend = (struct smc_cdc_tx_pend *)pnd_snd;
  30         struct smc_connection *conn = cdcpend->conn;
  31         struct smc_sock *smc;
  32         int diff;
  33
  34         if (!conn)
  35                 /* already dismissed */
  36                 return;
  37
  38         smc = container_of(conn, struct smc_sock, conn);
  39         bh_lock_sock(&smc->sk);
  40         if (!wc_status) {
  41                 diff = smc_curs_diff(cdcpend->conn->sndbuf_desc->len,
  42                                      &cdcpend->conn->tx_curs_fin,
  43                                      &cdcpend->cursor);
  44                 /* sndbuf_space is decreased in smc_sendmsg */
  45                 smp_mb__before_atomic();
  46                 atomic_add(diff, &cdcpend->conn->sndbuf_space);
  47                 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
  48                 smp_mb__after_atomic();
  49                 smc_curs_copy(&conn->tx_curs_fin, &cdcpend->cursor, conn);
  50                 smc_curs_copy(&conn->local_tx_ctrl_fin, &cdcpend->p_cursor,
  51                               conn);
  52                 conn->tx_cdc_seq_fin = cdcpend->ctrl_seq;
  53         }
  54         smc_tx_sndbuf_nonfull(smc);
  55         bh_unlock_sock(&smc->sk);
  56 }
  57
  58 int smc_cdc_get_free_slot(struct smc_connection *conn,
  59                           struct smc_wr_buf **wr_buf,
  60                           struct smc_rdma_wr **wr_rdma_buf,
  61                           struct smc_cdc_tx_pend **pend)
  62 {
  63         struct smc_link *link = conn->lnk;
  64         int rc;
  65
  66         rc = smc_wr_tx_get_free_slot(link, smc_cdc_tx_handler, wr_buf,
  67                                      wr_rdma_buf,
  68                                      (struct smc_wr_tx_pend_priv **)pend);
  69         if (conn->killed)
  70                 /* abnormal termination */
  71                 rc = -EPIPE;
  72         return rc;
  73 }
  74
  75 static inline void smc_cdc_add_pending_send(struct smc_connection *conn,
  76                                             struct smc_cdc_tx_pend *pend)
  77 {
  78         BUILD_BUG_ON_MSG(
  79                 sizeof(struct smc_cdc_msg) > SMC_WR_BUF_SIZE,
  80                 "must increase SMC_WR_BUF_SIZE to at least sizeof(struct smc_cdc_msg)");
  81         BUILD_BUG_ON_MSG(
  82                 offsetofend(struct smc_cdc_msg, reserved) > SMC_WR_TX_SIZE,
  83                 "must adapt SMC_WR_TX_SIZE to sizeof(struct smc_cdc_msg); if not all smc_wr upper layer protocols use the same message size any more, must start to set link->wr_tx_sges[i].length on each individual smc_wr_tx_send()");
  84         BUILD_BUG_ON_MSG(
  85                 sizeof(struct smc_cdc_tx_pend) > SMC_WR_TX_PEND_PRIV_SIZE,
  86                 "must increase SMC_WR_TX_PEND_PRIV_SIZE to at least sizeof(struct smc_cdc_tx_pend)");
  87         pend->conn = conn;
  88         pend->cursor = conn->tx_curs_sent;
  89         pend->p_cursor = conn->local_tx_ctrl.prod;
  90         pend->ctrl_seq = conn->tx_cdc_seq;
  91 }
  92
  93 int smc_cdc_msg_send(struct smc_connection *conn,
  94                      struct smc_wr_buf *wr_buf,
  95                      struct smc_cdc_tx_pend *pend)
  96 {
  97         struct smc_link *link = conn->lnk;
  98         union smc_host_cursor cfed;
  99         int rc;
 100
 101         smc_cdc_add_pending_send(conn, pend);
 102
 103         conn->tx_cdc_seq++;
 104         conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
 105         smc_host_msg_to_cdc((struct smc_cdc_msg *)wr_buf, conn, &cfed);
 106         rc = smc_wr_tx_send(link, (struct smc_wr_tx_pend_priv *)pend);
 107         if (!rc) {
 108                 smc_curs_copy(&conn->rx_curs_confirmed, &cfed, conn);
 109                 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
 110         } else {
 111                 conn->tx_cdc_seq--;
 112                 conn->local_tx_ctrl.seqno = conn->tx_cdc_seq;
 113         }
 114
 115         return rc;
 116 }
 117
 118 static int smcr_cdc_get_slot_and_msg_send(struct smc_connection *conn)
 119 {
 120         struct smc_cdc_tx_pend *pend;
 121         struct smc_wr_buf *wr_buf;
 122         int rc;
 123
 124         rc = smc_cdc_get_free_slot(conn, &wr_buf, NULL, &pend);
 125         if (rc)
 126                 return rc;
 127
 128         spin_lock_bh(&conn->send_lock);
 129         rc = smc_cdc_msg_send(conn, wr_buf, pend);
 130         spin_unlock_bh(&conn->send_lock);
 131         return rc;
 132 }
 133
 134 int smc_cdc_get_slot_and_msg_send(struct smc_connection *conn)
 135 {
 136         int rc;
 137
 138         if (!conn->lgr || (conn->lgr->is_smcd && conn->lgr->peer_shutdown))
 139                 return -EPIPE;
 140
 141         if (conn->lgr->is_smcd) {
 142                 spin_lock_bh(&conn->send_lock);
 143                 rc = smcd_cdc_msg_send(conn);
 144                 spin_unlock_bh(&conn->send_lock);
 145         } else {
 146                 rc = smcr_cdc_get_slot_and_msg_send(conn);
 147         }
 148
 149         return rc;
 150 }
 151
 152 static bool smc_cdc_tx_filter(struct smc_wr_tx_pend_priv *tx_pend,
 153                               unsigned long data)
 154 {
 155         struct smc_connection *conn = (struct smc_connection *)data;
 156         struct smc_cdc_tx_pend *cdc_pend =
 157                 (struct smc_cdc_tx_pend *)tx_pend;
 158
 159         return cdc_pend->conn == conn;
 160 }
 161
 162 static void smc_cdc_tx_dismisser(struct smc_wr_tx_pend_priv *tx_pend)
 163 {
 164         struct smc_cdc_tx_pend *cdc_pend =
 165                 (struct smc_cdc_tx_pend *)tx_pend;
 166
 167         cdc_pend->conn = NULL;
 168 }
 169
 170 void smc_cdc_tx_dismiss_slots(struct smc_connection *conn)
 171 {
 172         struct smc_link *link = conn->lnk;
 173
 174         smc_wr_tx_dismiss_slots(link, SMC_CDC_MSG_TYPE,
 175                                 smc_cdc_tx_filter, smc_cdc_tx_dismisser,
 176                                 (unsigned long)conn);
 177 }
 178
 179 /* Send a SMC-D CDC header.
 180  * This increments the free space available in our send buffer.
 181  * Also update the confirmed receive buffer with what was sent to the peer.
 182  */
 183 int smcd_cdc_msg_send(struct smc_connection *conn)
 184 {
 185         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 186         union smc_host_cursor curs;
 187         struct smcd_cdc_msg cdc;
 188         int rc, diff;
 189
 190         memset(&cdc, 0, sizeof(cdc));
 191         cdc.common.type = SMC_CDC_MSG_TYPE;
 192         curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.prod.acurs);
 193         cdc.prod.wrap = curs.wrap;
 194         cdc.prod.count = curs.count;
 195         curs.acurs.counter = atomic64_read(&conn->local_tx_ctrl.cons.acurs);
 196         cdc.cons.wrap = curs.wrap;
 197         cdc.cons.count = curs.count;
 198         cdc.cons.prod_flags = conn->local_tx_ctrl.prod_flags;
 199         cdc.cons.conn_state_flags = conn->local_tx_ctrl.conn_state_flags;
 200         rc = smcd_tx_ism_write(conn, &cdc, sizeof(cdc), 0, 1);
 201         if (rc)
 202                 return rc;
 203         smc_curs_copy(&conn->rx_curs_confirmed, &curs, conn);
 204         conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
 205         /* Calculate transmitted data and increment free send buffer space */
 206         diff = smc_curs_diff(conn->sndbuf_desc->len, &conn->tx_curs_fin,
 207                              &conn->tx_curs_sent);
 208         /* increased by confirmed number of bytes */
 209         smp_mb__before_atomic();
 210         atomic_add(diff, &conn->sndbuf_space);
 211         /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
 212         smp_mb__after_atomic();
 213         smc_curs_copy(&conn->tx_curs_fin, &conn->tx_curs_sent, conn);
 214
 215         smc_tx_sndbuf_nonfull(smc);
 216         return rc;
 217 }
 218
 219 /********************************* receive ***********************************/
 220
 221 static inline bool smc_cdc_before(u16 seq1, u16 seq2)
 222 {
 223         return (s16)(seq1 - seq2) < 0;
 224 }
 225
 226 static void smc_cdc_handle_urg_data_arrival(struct smc_sock *smc,
 227                                             int *diff_prod)
 228 {
 229         struct smc_connection *conn = &smc->conn;
 230         char *base;
 231
 232         /* new data included urgent business */
 233         smc_curs_copy(&conn->urg_curs, &conn->local_rx_ctrl.prod, conn);
 234         conn->urg_state = SMC_URG_VALID;
 235         if (!sock_flag(&smc->sk, SOCK_URGINLINE))
 236                 /* we'll skip the urgent byte, so don't account for it */
 237                 (*diff_prod)--;
 238         base = (char *)conn->rmb_desc->cpu_addr + conn->rx_off;
 239         if (conn->urg_curs.count)
 240                 conn->urg_rx_byte = *(base + conn->urg_curs.count - 1);
 241         else
 242                 conn->urg_rx_byte = *(base + conn->rmb_desc->len - 1);
 243         sk_send_sigurg(&smc->sk);
 244 }
 245
 246 static void smc_cdc_msg_recv_action(struct smc_sock *smc,
 247                                     struct smc_cdc_msg *cdc)
 248 {
 249         union smc_host_cursor cons_old, prod_old;
 250         struct smc_connection *conn = &smc->conn;
 251         int diff_cons, diff_prod;
 252
 253         smc_curs_copy(&prod_old, &conn->local_rx_ctrl.prod, conn);
 254         smc_curs_copy(&cons_old, &conn->local_rx_ctrl.cons, conn);
 255         smc_cdc_msg_to_host(&conn->local_rx_ctrl, cdc, conn);
 256
 257         diff_cons = smc_curs_diff(conn->peer_rmbe_size, &cons_old,
 258                                   &conn->local_rx_ctrl.cons);
 259         if (diff_cons) {
 260                 /* peer_rmbe_space is decreased during data transfer with RDMA
 261                  * write
 262                  */
 263                 smp_mb__before_atomic();
 264                 atomic_add(diff_cons, &conn->peer_rmbe_space);
 265                 /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
 266                 smp_mb__after_atomic();
 267         }
 268
 269         diff_prod = smc_curs_diff(conn->rmb_desc->len, &prod_old,
 270                                   &conn->local_rx_ctrl.prod);
 271         if (diff_prod) {
 272                 if (conn->local_rx_ctrl.prod_flags.urg_data_present)
 273                         smc_cdc_handle_urg_data_arrival(smc, &diff_prod);
 274                 /* bytes_to_rcv is decreased in smc_recvmsg */
 275                 smp_mb__before_atomic();
 276                 atomic_add(diff_prod, &conn->bytes_to_rcv);
 277                 /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
 278                 smp_mb__after_atomic();
 279                 smc->sk.sk_data_ready(&smc->sk);
 280         } else {
 281                 if (conn->local_rx_ctrl.prod_flags.write_blocked)
 282                         smc->sk.sk_data_ready(&smc->sk);
 283                 if (conn->local_rx_ctrl.prod_flags.urg_data_pending)
 284                         conn->urg_state = SMC_URG_NOTYET;
 285         }
 286
 287         /* trigger sndbuf consumer: RDMA write into peer RMBE and CDC */
 288         if ((diff_cons && smc_tx_prepared_sends(conn)) ||
 289             conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
 290             conn->local_rx_ctrl.prod_flags.urg_data_pending)
 291                 smc_tx_sndbuf_nonempty(conn);
 292
 293         if (diff_cons && conn->urg_tx_pend &&
 294             atomic_read(&conn->peer_rmbe_space) == conn->peer_rmbe_size) {
 295                 /* urg data confirmed by peer, indicate we're ready for more */
 296                 conn->urg_tx_pend = false;
 297                 smc->sk.sk_write_space(&smc->sk);
 298         }
 299
 300         if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
 301                 smc->sk.sk_err = ECONNRESET;
 302                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 303         }
 304         if (smc_cdc_rxed_any_close_or_senddone(conn)) {
 305                 smc->sk.sk_shutdown |= RCV_SHUTDOWN;
 306                 if (smc->clcsock && smc->clcsock->sk)
 307                         smc->clcsock->sk->sk_shutdown |= RCV_SHUTDOWN;
 308                 sock_set_flag(&smc->sk, SOCK_DONE);
 309                 sock_hold(&smc->sk); /* sock_put in close_work */
 310                 if (!schedule_work(&conn->close_work))
 311                         sock_put(&smc->sk);
 312         }
 313 }
 314
 315 /* called under tasklet context */
 316 static void smc_cdc_msg_recv(struct smc_sock *smc, struct smc_cdc_msg *cdc)
 317 {
 318         sock_hold(&smc->sk);
 319         bh_lock_sock(&smc->sk);
 320         smc_cdc_msg_recv_action(smc, cdc);
 321         bh_unlock_sock(&smc->sk);
 322         sock_put(&smc->sk); /* no free sk in softirq-context */
 323 }
 324
 325 /* Schedule a tasklet for this connection. Triggered from the ISM device IRQ
 326  * handler to indicate update in the DMBE.
 327  *
 328  * Context:
 329  * - tasklet context
 330  */
 331 static void smcd_cdc_rx_tsklet(unsigned long data)
 332 {
 333         struct smc_connection *conn = (struct smc_connection *)data;
 334         struct smcd_cdc_msg *data_cdc;
 335         struct smcd_cdc_msg cdc;
 336         struct smc_sock *smc;
 337
 338         if (!conn || conn->killed)
 339                 return;
 340
 341         data_cdc = (struct smcd_cdc_msg *)conn->rmb_desc->cpu_addr;
 342         smcd_curs_copy(&cdc.prod, &data_cdc->prod, conn);
 343         smcd_curs_copy(&cdc.cons, &data_cdc->cons, conn);
 344         smc = container_of(conn, struct smc_sock, conn);
 345         smc_cdc_msg_recv(smc, (struct smc_cdc_msg *)&cdc);
 346 }
 347
 348 /* Initialize receive tasklet. Called from ISM device IRQ handler to start
 349  * receiver side.
 350  */
 351 void smcd_cdc_rx_init(struct smc_connection *conn)
 352 {
 353         tasklet_init(&conn->rx_tsklet, smcd_cdc_rx_tsklet, (unsigned long)conn);
 354 }
 355
 356 /***************************** init, exit, misc ******************************/
 357
 358 static void smc_cdc_rx_handler(struct ib_wc *wc, void *buf)
 359 {
 360         struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
 361         struct smc_cdc_msg *cdc = buf;
 362         struct smc_connection *conn;
 363         struct smc_link_group *lgr;
 364         struct smc_sock *smc;
 365
 366         if (wc->byte_len < offsetof(struct smc_cdc_msg, reserved))
 367                 return; /* short message */
 368         if (cdc->len != SMC_WR_TX_SIZE)
 369                 return; /* invalid message */
 370
 371         /* lookup connection */
 372         lgr = smc_get_lgr(link);
 373         read_lock_bh(&lgr->conns_lock);
 374         conn = smc_lgr_find_conn(ntohl(cdc->token), lgr);
 375         read_unlock_bh(&lgr->conns_lock);
 376         if (!conn)
 377                 return;
 378         smc = container_of(conn, struct smc_sock, conn);
 379
 380         if (!cdc->prod_flags.failover_validation) {
 381                 if (smc_cdc_before(ntohs(cdc->seqno),
 382                                    conn->local_rx_ctrl.seqno))
 383                         /* received seqno is old */
 384                         return;
 385         }
 386         smc_cdc_msg_recv(smc, cdc);
 387 }
 388
 389 static struct smc_wr_rx_handler smc_cdc_rx_handlers[] = {
 390         {
 391                 .handler        = smc_cdc_rx_handler,
 392                 .type           = SMC_CDC_MSG_TYPE
 393         },
 394         {
 395                 .handler        = NULL,
 396         }
 397 };
 398
 399 int __init smc_cdc_init(void)
 400 {
 401         struct smc_wr_rx_handler *handler;
 402         int rc = 0;
 403
 404         for (handler = smc_cdc_rx_handlers; handler->handler; handler++) {
 405                 INIT_HLIST_NODE(&handler->list);
 406                 rc = smc_wr_rx_register_handler(handler);
 407                 if (rc)
 408                         break;
 409         }
 410         return rc;
 411 }