net/smc/smc_rx.c

   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  */
  28 static 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  */
  53 static 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_write(&conn->local_tx_ctrl.cons, smc_curs_read(&cons, conn),
  86                        conn);
  87
  88         /* send consumer cursor update if required */
  89         /* similar to advertising new TCP rcv_wnd if required */
  90         smc_tx_consumer_update(conn, force);
  91
  92         return rc;
  93 }
  94
  95 static void smc_rx_update_cons(struct smc_sock *smc, size_t len)
  96 {
  97         struct smc_connection *conn = &smc->conn;
  98         union smc_host_cursor cons;
  99
 100         smc_curs_write(&cons, smc_curs_read(&conn->local_tx_ctrl.cons, conn),
 101                        conn);
 102         smc_rx_update_consumer(smc, cons, len);
 103 }
 104
 105 struct smc_spd_priv {
 106         struct smc_sock *smc;
 107         size_t           len;
 108 };
 109
 110 static void smc_rx_pipe_buf_release(struct pipe_inode_info *pipe,
 111                                     struct pipe_buffer *buf)
 112 {
 113         struct smc_spd_priv *priv = (struct smc_spd_priv *)buf->private;
 114         struct smc_sock *smc = priv->smc;
 115         struct smc_connection *conn;
 116         struct sock *sk = &smc->sk;
 117
 118         if (sk->sk_state == SMC_CLOSED ||
 119             sk->sk_state == SMC_PEERFINCLOSEWAIT ||
 120             sk->sk_state == SMC_APPFINCLOSEWAIT)
 121                 goto out;
 122         conn = &smc->conn;
 123         lock_sock(sk);
 124         smc_rx_update_cons(smc, priv->len);
 125         release_sock(sk);
 126         if (atomic_sub_and_test(priv->len, &conn->splice_pending))
 127                 smc_rx_wake_up(sk);
 128 out:
 129         kfree(priv);
 130         put_page(buf->page);
 131         sock_put(sk);
 132 }
 133
 134 static int smc_rx_pipe_buf_nosteal(struct pipe_inode_info *pipe,
 135                                    struct pipe_buffer *buf)
 136 {
 137         return 1;
 138 }
 139
 140 static const struct pipe_buf_operations smc_pipe_ops = {
 141         .can_merge = 0,
 142         .confirm = generic_pipe_buf_confirm,
 143         .release = smc_rx_pipe_buf_release,
 144         .steal = smc_rx_pipe_buf_nosteal,
 145         .get = generic_pipe_buf_get
 146 };
 147
 148 static void smc_rx_spd_release(struct splice_pipe_desc *spd,
 149                                unsigned int i)
 150 {
 151         put_page(spd->pages[i]);
 152 }
 153
 154 static int smc_rx_splice(struct pipe_inode_info *pipe, char *src, size_t len,
 155                          struct smc_sock *smc)
 156 {
 157         struct splice_pipe_desc spd;
 158         struct partial_page partial;
 159         struct smc_spd_priv *priv;
 160         struct page *page;
 161         int bytes;
 162
 163         page = virt_to_page(smc->conn.rmb_desc->cpu_addr);
 164         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
 165         if (!priv)
 166                 return -ENOMEM;
 167         priv->len = len;
 168         priv->smc = smc;
 169         partial.offset = src - (char *)smc->conn.rmb_desc->cpu_addr;
 170         partial.len = len;
 171         partial.private = (unsigned long)priv;
 172
 173         spd.nr_pages_max = 1;
 174         spd.nr_pages = 1;
 175         spd.pages = &page;
 176         spd.partial = &partial;
 177         spd.ops = &smc_pipe_ops;
 178         spd.spd_release = smc_rx_spd_release;
 179
 180         bytes = splice_to_pipe(pipe, &spd);
 181         if (bytes > 0) {
 182                 sock_hold(&smc->sk);
 183                 get_page(smc->conn.rmb_desc->pages);
 184                 atomic_add(bytes, &smc->conn.splice_pending);
 185         }
 186
 187         return bytes;
 188 }
 189
 190 static int smc_rx_data_available_and_no_splice_pend(struct smc_connection *conn)
 191 {
 192         return atomic_read(&conn->bytes_to_rcv) &&
 193                !atomic_read(&conn->splice_pending);
 194 }
 195
 196 /* blocks rcvbuf consumer until >=len bytes available or timeout or interrupted
 197  *   @smc    smc socket
 198  *   @timeo  pointer to max seconds to wait, pointer to value 0 for no timeout
 199  *   @fcrit  add'l criterion to evaluate as function pointer
 200  * Returns:
 201  * 1 if at least 1 byte available in rcvbuf or if socket error/shutdown.
 202  * 0 otherwise (nothing in rcvbuf nor timeout, e.g. interrupted).
 203  */
 204 int smc_rx_wait(struct smc_sock *smc, long *timeo,
 205                 int (*fcrit)(struct smc_connection *conn))
 206 {
 207         DEFINE_WAIT_FUNC(wait, woken_wake_function);
 208         struct smc_connection *conn = &smc->conn;
 209         struct sock *sk = &smc->sk;
 210         int rc;
 211
 212         if (fcrit(conn))
 213                 return 1;
 214         sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
 215         add_wait_queue(sk_sleep(sk), &wait);
 216         rc = sk_wait_event(sk, timeo,
 217                            sk->sk_err ||
 218                            sk->sk_shutdown & RCV_SHUTDOWN ||
 219                            fcrit(conn) ||
 220                            smc_cdc_rxed_any_close_or_senddone(conn),
 221                            &wait);
 222         remove_wait_queue(sk_sleep(sk), &wait);
 223         sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
 224         return rc;
 225 }
 226
 227 static int smc_rx_recv_urg(struct smc_sock *smc, struct msghdr *msg, int len,
 228                            int flags)
 229 {
 230         struct smc_connection *conn = &smc->conn;
 231         union smc_host_cursor cons;
 232         struct sock *sk = &smc->sk;
 233         int rc = 0;
 234
 235         if (sock_flag(sk, SOCK_URGINLINE) ||
 236             !(conn->urg_state == SMC_URG_VALID) ||
 237             conn->urg_state == SMC_URG_READ)
 238                 return -EINVAL;
 239
 240         if (conn->urg_state == SMC_URG_VALID) {
 241                 if (!(flags & MSG_PEEK))
 242                         smc->conn.urg_state = SMC_URG_READ;
 243                 msg->msg_flags |= MSG_OOB;
 244                 if (len > 0) {
 245                         if (!(flags & MSG_TRUNC))
 246                                 rc = memcpy_to_msg(msg, &conn->urg_rx_byte, 1);
 247                         len = 1;
 248                         smc_curs_write(&cons,
 249                                        smc_curs_read(&conn->local_tx_ctrl.cons,
 250                                                      conn),
 251                                        conn);
 252                         if (smc_curs_diff(conn->rmb_desc->len, &cons,
 253                                           &conn->urg_curs) > 1)
 254                                 conn->urg_rx_skip_pend = true;
 255                         /* Urgent Byte was already accounted for, but trigger
 256                          * skipping the urgent byte in non-inline case
 257                          */
 258                         if (!(flags & MSG_PEEK))
 259                                 smc_rx_update_consumer(smc, cons, 0);
 260                 } else {
 261                         msg->msg_flags |= MSG_TRUNC;
 262                 }
 263
 264                 return rc ? -EFAULT : len;
 265         }
 266
 267         if (sk->sk_state == SMC_CLOSED || sk->sk_shutdown & RCV_SHUTDOWN)
 268                 return 0;
 269
 270         return -EAGAIN;
 271 }
 272
 273 /* smc_rx_recvmsg - receive data from RMBE
 274  * @msg:        copy data to receive buffer
 275  * @pipe:       copy data to pipe if set - indicates splice() call
 276  *
 277  * rcvbuf consumer: main API called by socket layer.
 278  * Called under sk lock.
 279  */
 280 int smc_rx_recvmsg(struct smc_sock *smc, struct msghdr *msg,
 281                    struct pipe_inode_info *pipe, size_t len, int flags)
 282 {
 283         size_t copylen, read_done = 0, read_remaining = len;
 284         size_t chunk_len, chunk_off, chunk_len_sum;
 285         struct smc_connection *conn = &smc->conn;
 286         int (*func)(struct smc_connection *conn);
 287         union smc_host_cursor cons;
 288         int readable, chunk;
 289         char *rcvbuf_base;
 290         struct sock *sk;
 291         int splbytes;
 292         long timeo;
 293         int target;             /* Read at least these many bytes */
 294         int rc;
 295
 296         if (unlikely(flags & MSG_ERRQUEUE))
 297                 return -EINVAL; /* future work for sk.sk_family == AF_SMC */
 298
 299         sk = &smc->sk;
 300         if (sk->sk_state == SMC_LISTEN)
 301                 return -ENOTCONN;
 302         if (flags & MSG_OOB)
 303                 return smc_rx_recv_urg(smc, msg, len, flags);
 304         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
 305         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
 306
 307         /* we currently use 1 RMBE per RMB, so RMBE == RMB base addr */
 308         rcvbuf_base = conn->rmb_desc->cpu_addr;
 309
 310         do { /* while (read_remaining) */
 311                 if (read_done >= target || (pipe && read_done))
 312                         break;
 313
 314                 if (atomic_read(&conn->bytes_to_rcv))
 315                         goto copy;
 316                 else if (conn->urg_state == SMC_URG_VALID)
 317                         /* we received a single urgent Byte - skip */
 318                         smc_rx_update_cons(smc, 0);
 319
 320                 if (sk->sk_shutdown & RCV_SHUTDOWN ||
 321                     smc_cdc_rxed_any_close_or_senddone(conn) ||
 322                     conn->local_tx_ctrl.conn_state_flags.peer_conn_abort)
 323                         break;
 324
 325                 if (read_done) {
 326                         if (sk->sk_err ||
 327                             sk->sk_state == SMC_CLOSED ||
 328                             !timeo ||
 329                             signal_pending(current))
 330                                 break;
 331                 } else {
 332                         if (sk->sk_err) {
 333                                 read_done = sock_error(sk);
 334                                 break;
 335                         }
 336                         if (sk->sk_state == SMC_CLOSED) {
 337                                 if (!sock_flag(sk, SOCK_DONE)) {
 338                                         /* This occurs when user tries to read
 339                                          * from never connected socket.
 340                                          */
 341                                         read_done = -ENOTCONN;
 342                                         break;
 343                                 }
 344                                 break;
 345                         }
 346                         if (signal_pending(current)) {
 347                                 read_done = sock_intr_errno(timeo);
 348                                 break;
 349                         }
 350                         if (!timeo)
 351                                 return -EAGAIN;
 352                 }
 353
 354                 if (!smc_rx_data_available(conn)) {
 355                         smc_rx_wait(smc, &timeo, smc_rx_data_available);
 356                         continue;
 357                 }
 358
 359 copy:
 360                 /* initialize variables for 1st iteration of subsequent loop */
 361                 /* could be just 1 byte, even after waiting on data above */
 362                 readable = atomic_read(&conn->bytes_to_rcv);
 363                 splbytes = atomic_read(&conn->splice_pending);
 364                 if (!readable || (msg && splbytes)) {
 365                         if (splbytes)
 366                                 func = smc_rx_data_available_and_no_splice_pend;
 367                         else
 368                                 func = smc_rx_data_available;
 369                         smc_rx_wait(smc, &timeo, func);
 370                         continue;
 371                 }
 372
 373                 smc_curs_write(&cons,
 374                                smc_curs_read(&conn->local_tx_ctrl.cons, conn),
 375                                conn);
 376                 /* subsequent splice() calls pick up where previous left */
 377                 if (splbytes)
 378                         smc_curs_add(conn->rmb_desc->len, &cons, splbytes);
 379                 if (conn->urg_state == SMC_URG_VALID &&
 380                     sock_flag(&smc->sk, SOCK_URGINLINE) &&
 381                     readable > 1)
 382                         readable--;     /* always stop at urgent Byte */
 383                 /* not more than what user space asked for */
 384                 copylen = min_t(size_t, read_remaining, readable);
 385                 /* determine chunks where to read from rcvbuf */
 386                 /* either unwrapped case, or 1st chunk of wrapped case */
 387                 chunk_len = min_t(size_t, copylen, conn->rmb_desc->len -
 388                                   cons.count);
 389                 chunk_len_sum = chunk_len;
 390                 chunk_off = cons.count;
 391                 smc_rmb_sync_sg_for_cpu(conn);
 392                 for (chunk = 0; chunk < 2; chunk++) {
 393                         if (!(flags & MSG_TRUNC)) {
 394                                 if (msg) {
 395                                         rc = memcpy_to_msg(msg, rcvbuf_base +
 396                                                            chunk_off,
 397                                                            chunk_len);
 398                                 } else {
 399                                         rc = smc_rx_splice(pipe, rcvbuf_base +
 400                                                         chunk_off, chunk_len,
 401                                                         smc);
 402                                 }
 403                                 if (rc < 0) {
 404                                         if (!read_done)
 405                                                 read_done = -EFAULT;
 406                                         smc_rmb_sync_sg_for_device(conn);
 407                                         goto out;
 408                                 }
 409                         }
 410                         read_remaining -= chunk_len;
 411                         read_done += chunk_len;
 412
 413                         if (chunk_len_sum == copylen)
 414                                 break; /* either on 1st or 2nd iteration */
 415                         /* prepare next (== 2nd) iteration */
 416                         chunk_len = copylen - chunk_len; /* remainder */
 417                         chunk_len_sum += chunk_len;
 418                         chunk_off = 0; /* modulo offset in recv ring buffer */
 419                 }
 420                 smc_rmb_sync_sg_for_device(conn);
 421
 422                 /* update cursors */
 423                 if (!(flags & MSG_PEEK)) {
 424                         /* increased in recv tasklet smc_cdc_msg_rcv() */
 425                         smp_mb__before_atomic();
 426                         atomic_sub(copylen, &conn->bytes_to_rcv);
 427                         /* guarantee 0 <= bytes_to_rcv <= rmb_desc->len */
 428                         smp_mb__after_atomic();
 429                         if (msg && smc_rx_update_consumer(smc, cons, copylen))
 430                                 goto out;
 431                 }
 432         } while (read_remaining);
 433 out:
 434         return read_done;
 435 }
 436
 437 /* Initialize receive properties on connection establishment. NB: not __init! */
 438 void smc_rx_init(struct smc_sock *smc)
 439 {
 440         smc->sk.sk_data_ready = smc_rx_wake_up;
 441         atomic_set(&smc->conn.splice_pending, 0);
 442         smc->conn.urg_state = SMC_URG_READ;
 443 }