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