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