LXR linux/net/smc/smc

   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * Shared Memory Communications over RDMA (SMC-R) and RoCE
   4 *
   5 * Manage RMBE
   6 * copy new RMBE data into user space
   7 *
   8 * Copyright IBM Corp. 2016
   9 *
  10 * Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  11 */
  12
  13#include <linux/net.h>
  14#include <linux/rcupdate.h>
  15#include <linux/sched/signal.h>
  16
  17#include <net/sock.h>
  18
  19#include "smc.h"
  20#include "smc_core.h"
  21#include "smc_cdc.h"
  22#include "smc_tx.h" /* smc_tx_consumer_update() */
  23#include "smc_rx.h"
  24
  25/* callback implementation to wakeup consumers blocked with smc_rx_wait().
  26 * indirectly called by smc_cdc_msg_recv_action().
  27 */
  28static void smc_rx_wake_up(struct sock *sk)
  29{
  30        struct socket_wq *wq;
  31
  32        /* derived from sock_def_readable() */
  33        /* called already in smc_listen_work() */
  34        rcu_read_lock();
  35        wq = rcu_dereference(sk->sk_wq);
  36        if (skwq_has_sleeper(wq))
  37                wake_up_interruptible_sync_poll(&wq->wait, EPOLLIN | EPOLLPRI |
  38                                                EPOLLRDNORM | EPOLLRDBAND);
  39        sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_IN);
  40        if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
  41            (sk->sk_state == SMC_CLOSED))
  42                sk_wake_async(sk, SOCK_WAKE_WAITD, POLL_HUP);
  43        rcu_read_unlock();
  44}
  45
  46/* Update consumer cursor
  47 *   @conn   connection to update
  48 *   @cons   consumer cursor
  49 *   @len    number of Bytes consumed
  50 *   Returns:
  51 *   1 if we should end our receive, 0 otherwise
  52 */
  53static int smc_rx_update_consumer(struct smc_sock *smc,
  54                                  union smc_host_cursor cons, size_t len)
  55{
  56        struct smc_connection *conn = &smc->conn;
  57        struct sock *sk = &smc->sk;
  58        bool force = false;
  59        int diff, rc = 0;
  60
  61        smc_curs_add(conn->rmb_desc->len, &cons, len);
  62
  63        /* did we process urgent data? */
  64        if (conn->urg_state == SMC_URG_VALID || conn->urg_rx_skip_pend) {
  65                diff = smc_curs_comp(conn->rmb_desc->len, &cons,
  66                                     &conn->urg_curs);
  67                if (sock_flag(sk, SOCK_URGINLINE)) {
  68                        if (diff == 0) {
  69                                force = true;
  70                                rc = 1;
  71                                conn->urg_state = SMC_URG_READ;
  72                        }
  73                } else {
  74                        if (diff == 1) {
  75                                /* skip urgent byte */
  76                                force = true;
  77                                smc_curs_add(conn->rmb_desc->len, &cons, 1);
  78                                conn->urg_rx_skip_pend = false;
  79                        } else if (diff < -1)
  80                                /* we read past urgent byte */
  81                                conn->urg_state = SMC_URG_READ;
  82                }
  83        }
  84
  85        smc_curs_copy(&conn->local_tx_ctrl.cons, &cons, conn);
  86
  87        /* send consumer cursor update if required */
  88        /* similar to advertising new TCP rcv_wnd if required */
  89        smc_tx_consumer_update(conn, force);
  90
  91        return rc;
  92}
  93
  94static void smc_rx_update_cons(struct smc_sock *smc, size_t len)
  95{
  96        struct smc_connection *conn = &smc->conn;
  97        union smc_host_cursor cons;
  98
  99        smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
 100        smc_rx_update_consumer(smc, cons, len);
 101}
 102
 103struct smc_spd_priv {
 104        struct smc_sock *smc;
 105        size_t           len;
 106};
 107
 108static void smc_rx_pipe_buf_release(struct pipe_inode_info *pipe,
 109                                    struct pipe_buffer *buf)
 110{
 111        struct smc_spd_priv *priv = (struct smc_spd_priv *)buf->private;
 112        struct smc_sock *smc = priv->smc;
 113        struct smc_connection *conn;
 114        struct sock *sk = &smc->sk;
 115
 116        if (sk->sk_state == SMC_CLOSED ||
 117            sk->sk_state == SMC_PEERFINCLOSEWAIT ||
 118            sk->sk_state == SMC_APPFINCLOSEWAIT)
 119                goto out;
 120        conn = &smc->conn;
 121        lock_sock(sk);
 122        smc_rx_update_cons(smc, priv->len);
 123        release_sock(sk);
 124        if (atomic_sub_and_test(priv->len, &conn->splice_pending))
 125                smc_rx_wake_up(sk);
 126out:
 127        kfree(priv);
 128        put_page(buf->page);
 129        sock_put(sk);
 130}
 131
 132static int smc_rx_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 133                                   struct pipe_buffer *buf)
 134{
 135        return 1;
 136}
 137
 138static const struct pipe_buf_operations smc_pipe_ops = {
 139        .confirm = generic_pipe_buf_confirm,
 140        .release = smc_rx_pipe_buf_release,
 141        .steal = smc_rx_pipe_buf_nosteal,
 142        .get = generic_pipe_buf_get
 143};
 144
 145static void smc_rx_spd_release(struct splice_pipe_desc *spd,
 146                               unsigned int i)
 147{
 148        put_page(spd->pages[i]);
 149}
 150
 151static int smc_rx_splice(struct pipe_inode_info *pipe, char *src, size_t len,
 152                         struct smc_sock *smc)
 153{
 154        struct splice_pipe_desc spd;
 155        struct partial_page partial;
 156        struct smc_spd_priv *priv;
 157        int bytes;
 158
 159        priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 160        if (!priv)
 161                return -ENOMEM;
 162        priv->len = len;
 163        priv->smc = smc;
 164        partial.offset = src - (char *)smc->conn.rmb_desc->cpu_addr;
 165        partial.len = len;
 166        partial.private = (unsigned long)priv;
 167
 168        spd.nr_pages_max = 1;
 169        spd.nr_pages = 1;
 170        spd.pages = &smc->conn.rmb_desc->pages;
 171        spd.partial = &partial;
 172        spd.ops = &smc_pipe_ops;
 173        spd.spd_release = smc_rx_spd_release;
 174
 175        bytes = splice_to_pipe(pipe, &spd);
 176        if (bytes > 0) {
 177                sock_hold(&smc->sk);
 178                get_page(smc->conn.rmb_desc->pages);
 179                atomic_add(bytes, &smc->conn.splice_pending);
 180        }
 181
 182        return bytes;
 183}
 184
 185static int smc_rx_data_available_and_no_splice_pend(struct smc_connection *conn)
 186{
 187        return atomic_read(&conn->bytes_to_rcv) &&
 188               !atomic_read(&conn->splice_pending);
 189}
 190
 191/* blocks rcvbuf consumer until >=len bytes available or timeout or interrupted
 192 *   @smc    smc socket
 193 *   @timeo  pointer to max seconds to wait, pointer to value 0 for no timeout
 194 *   @fcrit  add'l criterion to evaluate as function pointer
 195 * Returns:
 196 * 1 if at least 1 byte available in rcvbuf or if socket error/shutdown.
 197 * 0 otherwise (nothing in rcvbuf nor timeout, e.g. interrupted).
 198 */
 199int smc_rx_wait(struct smc_sock *smc, long *timeo,
 200                int (*fcrit)(struct smc_connection *conn))
 201{
 202        DEFINE_WAIT_FUNC(wait, woken_wake_function);
 203        struct smc_connection *conn = &smc->conn;
 204        struct sock *sk = &smc->sk;
 205        int rc;
 206
 207        if (fcrit(conn))
 208                return 1;
 209        sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
 210        add_wait_queue(sk_sleep(sk), &wait);
 211        rc = sk_wait_event(sk, timeo,
 212                           sk->sk_err ||
 213                           sk->sk_shutdown & RCV_SHUTDOWN ||
 214                           fcrit(conn) ||
 215                           smc_cdc_rxed_any_close_or_senddone(conn),
 216                           &wait);
 217        remove_wait_queue(sk_sleep(sk), &wait);
 218        sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
 219        return rc;
 220}
 221
 222static int smc_rx_recv_urg(struct smc_sock *smc, struct msghdr *msg, int len,
 223                           int flags)
 224{
 225        struct smc_connection *conn = &smc->conn;
 226        union smc_host_cursor cons;
 227        struct sock *sk = &smc->sk;
 228        int rc = 0;
 229
 230        if (sock_flag(sk, SOCK_URGINLINE) ||
 231            !(conn->urg_state == SMC_URG_VALID) ||
 232            conn->urg_state == SMC_URG_READ)
 233                return -EINVAL;
 234
 235        if (conn->urg_state == SMC_URG_VALID) {
 236                if (!(flags & MSG_PEEK))
 237                        smc->conn.urg_state = SMC_URG_READ;
 238                msg->msg_flags |= MSG_OOB;
 239                if (len > 0) {
 240                        if (!(flags & MSG_TRUNC))
 241                                rc = memcpy_to_msg(msg, &conn->urg_rx_byte, 1);
 242                        len = 1;
 243                        smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
 244                        if (smc_curs_diff(conn->rmb_desc->len, &cons,
 245                                          &conn->urg_curs) > 1)
 246                                conn->urg_rx_skip_pend = true;
 247                        /* Urgent Byte was already accounted for, but trigger
 248                         * skipping the urgent byte in non-inline case
 249                         */
 250                        if (!(flags & MSG_PEEK))
 251                                smc_rx_update_consumer(smc, cons, 0);
 252                } else {
 253                        msg->msg_flags |= MSG_TRUNC;
 254                }
 255
 256                return rc ? -EFAULT : len;
 257        }
 258
 259        if (sk->sk_state == SMC_CLOSED || sk->sk_shutdown & RCV_SHUTDOWN)
 260                return 0;
 261
 262        return -EAGAIN;
 263}
 264
 265/* smc_rx_recvmsg - receive data from RMBE
 266 * @msg:        copy data to receive buffer
 267 * @pipe:       copy data to pipe if set - indicates splice() call
 268 *
 269 * rcvbuf consumer: main API called by socket layer.
 270 * Called under sk lock.
 271 */
 272int smc_rx_recvmsg(struct smc_sock *smc, struct msghdr *msg,
 273                   struct pipe_inode_info *pipe, size_t len, int flags)
 274{
 275        size_t copylen, read_done = 0, read_remaining = len;
 276        size_t chunk_len, chunk_off, chunk_len_sum;
 277        struct smc_connection *conn = &smc->conn;
 278        int (*func)(struct smc_connection *conn);
 279        union smc_host_cursor cons;
 280        int readable, chunk;
 281        char *rcvbuf_base;
 282        struct sock *sk;
 283        int splbytes;
 284        long timeo;
 285        int target;             /* Read at least these many bytes */
 286        int rc;
 287
 288        if (unlikely(flags & MSG_ERRQUEUE))
 289                return -EINVAL; /* future work for sk.sk_family == AF_SMC */
 290
 291        sk = &smc->sk;
 292        if (sk->sk_state == SMC_LISTEN)
 293                return -ENOTCONN;
 294        if (flags & MSG_OOB)
 295                return smc_rx_recv_urg(smc, msg, len, flags);
 296        timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
 297        target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
 298
 299        /* we currently use 1 RMBE per RMB, so RMBE == RMB base addr */
 300        rcvbuf_base = conn->rx_off + conn->rmb_desc->cpu_addr;
 301
 302        do { /* while (read_remaining) */
 303                if (read_done >= target || (pipe && read_done))
 304                        break;
 305
 306                if (atomic_read(&conn->bytes_to_rcv))
 307                        goto copy;
 308                else if (conn->urg_state == SMC_URG_VALID)
 309                        /* we received a single urgent Byte - skip */
 310                        smc_rx_update_cons(smc, 0);
 311
 312                if (sk->sk_shutdown & RCV_SHUTDOWN ||
 313                    smc_cdc_rxed_any_close_or_senddone(conn) ||
 314                    conn->local_tx_ctrl.conn_state_flags.peer_conn_abort)
 315                        break;
 316
 317                if (read_done) {
 318                        if (sk->sk_err ||
 319                            sk->sk_state == SMC_CLOSED ||
 320                            !timeo ||
 321                            signal_pending(current))
 322                                break;
 323                } else {
 324                        if (sk->sk_err) {
 325                                read_done = sock_error(sk);
 326                                break;
 327                        }
 328                        if (sk->sk_state == SMC_CLOSED) {
 329                                if (!sock_flag(sk, SOCK_DONE)) {
 330                                        /* This occurs when user tries to read
 331                                         * from never connected socket.
 332                                         */
 333                                        read_done = -ENOTCONN;
 334                                        break;
 335                                }
 336                                break;
 337                        }
 338                        if (signal_pending(current)) {
 339                                read_done = sock_intr_errno(timeo);
 340                                break;
 341                        }
 342                        if (!timeo)
 343                                return -EAGAIN;
 344                }
 345
 346                if (!smc_rx_data_available(conn)) {
 347                        smc_rx_wait(smc, &timeo, smc_rx_data_available);
 348                        continue;
 349                }
 350
 351copy:
 352                /* initialize variables for 1st iteration of subsequent loop */
 353                /* could be just 1 byte, even after waiting on data above */
 354                readable = atomic_read(&conn->bytes_to_rcv);
 355                splbytes = atomic_read(&conn->splice_pending);
 356                if (!readable || (msg && splbytes)) {
 357                        if (splbytes)
 358                                func = smc_rx_data_available_and_no_splice_pend;
 359                        else
 360                                func = smc_rx_data_available;
 361                        smc_rx_wait(smc, &timeo, func);
 362                        continue;
 363                }
 364
 365                smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
 366                /* subsequent splice() calls pick up where previous left */
 367                if (splbytes)
 368                        smc_curs_add(conn->rmb_desc->len, &cons, splbytes);
 369                if (conn->urg_state == SMC_URG_VALID &&
 370                    sock_flag(&smc->sk, SOCK_URGINLINE) &&
 371                    readable > 1)
 372                        readable--;     /* always stop at urgent Byte */
 373                /* not more than what user space asked for */
 374                copylen = min_t(size_t, read_remaining, readable);
 375                /* determine chunks where to read from rcvbuf */
 376                /* either unwrapped case, or 1st chunk of wrapped case */
 377                chunk_len = min_t(size_t, copylen, conn->rmb_desc->len -
 378                                  cons.count);
 379                chunk_len_sum = chunk_len;
 380                chunk_off = cons.count;
 381                smc_rmb_sync_sg_for_cpu(conn);
 382                for (chunk = 0; chunk < 2; chunk++) {
 383                        if (!(flags & MSG_TRUNC)) {
 384                                if (msg) {
 385                                        rc = memcpy_to_msg(msg, rcvbuf_base +
 386                                                           chunk_off,
 387                                                           chunk_len);
 388                                } else {
 389                                        rc = smc_rx_splice(pipe, rcvbuf_base +
 390                                                        chunk_off, chunk_len,
 391                                                        smc);
 392                                }
 393                                if (rc < 0) {
 394                                        if (!read_done)
 395                                                read_done = -EFAULT;
 396                                        smc_rmb_sync_sg_for_device(conn);
 397                                        goto out;
 398                                }
 399                        }
 400                        read_remaining -= chunk_len;
 401                        read_done += chunk_len;
 402
 403                        if (chunk_len_sum == copylen)
 404                                break; /* either on 1st or 2nd iteration */
 405                        /* prepare next (== 2nd) iteration */
 406                        chunk_len = copylen - chunk_len; /* remainder */
 407                        chunk_len_sum += chunk_len;
 408                        chunk_off = 0; /* modulo offset in recv ring buffer */
 409                }
 410                smc_rmb_sync_sg_for_device(conn);
 411
 412                /* update cursors */
 413                if (!(flags & MSG_PEEK)) {
 414                        /* increased in recv tasklet smc_cdc_msg_rcv() */
 415                        smp_mb__before_atomic();
 416                        atomic_sub(copylen, &conn->bytes_to_rcv);
 417                        /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
 418                        smp_mb__after_atomic();
 419                        if (msg && smc_rx_update_consumer(smc, cons, copylen))
 420                                goto out;
 421                }
 422        } while (read_remaining);
 423out:
 424        return read_done;
 425}
 426
 427/* Initialize receive properties on connection establishment. NB: not __init! */
 428void smc_rx_init(struct smc_sock *smc)
 429{
 430        smc->sk.sk_data_ready = smc_rx_wake_up;
 431        atomic_set(&smc->conn.splice_pending, 0);
 432        smc->conn.urg_state = SMC_URG_READ;
 433}
 434