net/smc/smc_tx.c

   1 /*
   2  * Shared Memory Communications over RDMA (SMC-R) and RoCE
   3  *
   4  * Manage send buffer.
   5  * Producer:
   6  * Copy user space data into send buffer, if send buffer space available.
   7  * Consumer:
   8  * Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available.
   9  *
  10  * Copyright IBM Corp. 2016
  11  *
  12  * Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  13  */
  14
  15 #include <linux/net.h>
  16 #include <linux/rcupdate.h>
  17 #include <linux/workqueue.h>
  18 #include <net/sock.h>
  19
  20 #include "smc.h"
  21 #include "smc_wr.h"
  22 #include "smc_cdc.h"
  23 #include "smc_tx.h"
  24
  25 /***************************** sndbuf producer *******************************/
  26
  27 /* callback implementation for sk.sk_write_space()
  28  * to wakeup sndbuf producers that blocked with smc_tx_wait_memory().
  29  * called under sk_socket lock.
  30  */
  31 static void smc_tx_write_space(struct sock *sk)
  32 {
  33         struct socket *sock = sk->sk_socket;
  34         struct smc_sock *smc = smc_sk(sk);
  35         struct socket_wq *wq;
  36
  37         /* similar to sk_stream_write_space */
  38         if (atomic_read(&smc->conn.sndbuf_space) && sock) {
  39                 clear_bit(SOCK_NOSPACE, &sock->flags);
  40                 rcu_read_lock();
  41                 wq = rcu_dereference(sk->sk_wq);
  42                 if (skwq_has_sleeper(wq))
  43                         wake_up_interruptible_poll(&wq->wait,
  44                                                    POLLOUT | POLLWRNORM |
  45                                                    POLLWRBAND);
  46                 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))
  47                         sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT);
  48                 rcu_read_unlock();
  49         }
  50 }
  51
  52 /* Wakeup sndbuf producers that blocked with smc_tx_wait_memory().
  53  * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
  54  */
  55 void smc_tx_sndbuf_nonfull(struct smc_sock *smc)
  56 {
  57         if (smc->sk.sk_socket &&
  58             test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))
  59                 smc->sk.sk_write_space(&smc->sk);
  60 }
  61
  62 /* blocks sndbuf producer until at least one byte of free space available */
  63 static int smc_tx_wait_memory(struct smc_sock *smc, int flags)
  64 {
  65         DEFINE_WAIT_FUNC(wait, woken_wake_function);
  66         struct smc_connection *conn = &smc->conn;
  67         struct sock *sk = &smc->sk;
  68         bool noblock;
  69         long timeo;
  70         int rc = 0;
  71
  72         /* similar to sk_stream_wait_memory */
  73         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
  74         noblock = timeo ? false : true;
  75         add_wait_queue(sk_sleep(sk), &wait);
  76         while (1) {
  77                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
  78                 if (sk->sk_err ||
  79                     (sk->sk_shutdown & SEND_SHUTDOWN) ||
  80                     conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {
  81                         rc = -EPIPE;
  82                         break;
  83                 }
  84                 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
  85                         rc = -ECONNRESET;
  86                         break;
  87                 }
  88                 if (!timeo) {
  89                         if (noblock)
  90                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
  91                         rc = -EAGAIN;
  92                         break;
  93                 }
  94                 if (signal_pending(current)) {
  95                         rc = sock_intr_errno(timeo);
  96                         break;
  97                 }
  98                 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
  99                 if (atomic_read(&conn->sndbuf_space))
 100                         break; /* at least 1 byte of free space available */
 101                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 102                 sk->sk_write_pending++;
 103                 sk_wait_event(sk, &timeo,
 104                               sk->sk_err ||
 105                               (sk->sk_shutdown & SEND_SHUTDOWN) ||
 106                               smc_cdc_rxed_any_close_or_senddone(conn) ||
 107                               atomic_read(&conn->sndbuf_space),
 108                               &wait);
 109                 sk->sk_write_pending--;
 110         }
 111         remove_wait_queue(sk_sleep(sk), &wait);
 112         return rc;
 113 }
 114
 115 /* sndbuf producer: main API called by socket layer.
 116  * called under sock lock.
 117  */
 118 int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
 119 {
 120         size_t copylen, send_done = 0, send_remaining = len;
 121         size_t chunk_len, chunk_off, chunk_len_sum;
 122         struct smc_connection *conn = &smc->conn;
 123         union smc_host_cursor prep;
 124         struct sock *sk = &smc->sk;
 125         char *sndbuf_base;
 126         int tx_cnt_prep;
 127         int writespace;
 128         int rc, chunk;
 129
 130         /* This should be in poll */
 131         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 132
 133         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
 134                 rc = -EPIPE;
 135                 goto out_err;
 136         }
 137
 138         while (msg_data_left(msg)) {
 139                 if (sk->sk_state == SMC_INIT)
 140                         return -ENOTCONN;
 141                 if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
 142                     (smc->sk.sk_err == ECONNABORTED) ||
 143                     conn->local_tx_ctrl.conn_state_flags.peer_conn_abort)
 144                         return -EPIPE;
 145                 if (smc_cdc_rxed_any_close(conn))
 146                         return send_done ?: -ECONNRESET;
 147
 148                 if (!atomic_read(&conn->sndbuf_space)) {
 149                         rc = smc_tx_wait_memory(smc, msg->msg_flags);
 150                         if (rc) {
 151                                 if (send_done)
 152                                         return send_done;
 153                                 goto out_err;
 154                         }
 155                         continue;
 156                 }
 157
 158                 /* initialize variables for 1st iteration of subsequent loop */
 159                 /* could be just 1 byte, even after smc_tx_wait_memory above */
 160                 writespace = atomic_read(&conn->sndbuf_space);
 161                 /* not more than what user space asked for */
 162                 copylen = min_t(size_t, send_remaining, writespace);
 163                 /* determine start of sndbuf */
 164                 sndbuf_base = conn->sndbuf_desc->cpu_addr;
 165                 smc_curs_write(&prep,
 166                                smc_curs_read(&conn->tx_curs_prep, conn),
 167                                conn);
 168                 tx_cnt_prep = prep.count;
 169                 /* determine chunks where to write into sndbuf */
 170                 /* either unwrapped case, or 1st chunk of wrapped case */
 171                 chunk_len = min_t(size_t,
 172                                   copylen, conn->sndbuf_size - tx_cnt_prep);
 173                 chunk_len_sum = chunk_len;
 174                 chunk_off = tx_cnt_prep;
 175                 for (chunk = 0; chunk < 2; chunk++) {
 176                         rc = memcpy_from_msg(sndbuf_base + chunk_off,
 177                                              msg, chunk_len);
 178                         if (rc) {
 179                                 if (send_done)
 180                                         return send_done;
 181                                 goto out_err;
 182                         }
 183                         send_done += chunk_len;
 184                         send_remaining -= chunk_len;
 185
 186                         if (chunk_len_sum == copylen)
 187                                 break; /* either on 1st or 2nd iteration */
 188                         /* prepare next (== 2nd) iteration */
 189                         chunk_len = copylen - chunk_len; /* remainder */
 190                         chunk_len_sum += chunk_len;
 191                         chunk_off = 0; /* modulo offset in send ring buffer */
 192                 }
 193                 /* update cursors */
 194                 smc_curs_add(conn->sndbuf_size, &prep, copylen);
 195                 smc_curs_write(&conn->tx_curs_prep,
 196                                smc_curs_read(&prep, conn),
 197                                conn);
 198                 /* increased in send tasklet smc_cdc_tx_handler() */
 199                 smp_mb__before_atomic();
 200                 atomic_sub(copylen, &conn->sndbuf_space);
 201                 /* guarantee 0 <= sndbuf_space <= sndbuf_size */
 202                 smp_mb__after_atomic();
 203                 /* since we just produced more new data into sndbuf,
 204                  * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
 205                  */
 206                 smc_tx_sndbuf_nonempty(conn);
 207         } /* while (msg_data_left(msg)) */
 208
 209         return send_done;
 210
 211 out_err:
 212         rc = sk_stream_error(sk, msg->msg_flags, rc);
 213         /* make sure we wake any epoll edge trigger waiter */
 214         if (unlikely(rc == -EAGAIN))
 215                 sk->sk_write_space(sk);
 216         return rc;
 217 }
 218
 219 /***************************** sndbuf consumer *******************************/
 220
 221 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
 222 static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
 223                              int num_sges, struct ib_sge sges[])
 224 {
 225         struct smc_link_group *lgr = conn->lgr;
 226         struct ib_send_wr *failed_wr = NULL;
 227         struct ib_rdma_wr rdma_wr;
 228         struct smc_link *link;
 229         int rc;
 230
 231         memset(&rdma_wr, 0, sizeof(rdma_wr));
 232         link = &lgr->lnk[SMC_SINGLE_LINK];
 233         rdma_wr.wr.wr_id = smc_wr_tx_get_next_wr_id(link);
 234         rdma_wr.wr.sg_list = sges;
 235         rdma_wr.wr.num_sge = num_sges;
 236         rdma_wr.wr.opcode = IB_WR_RDMA_WRITE;
 237         rdma_wr.remote_addr =
 238                 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr +
 239                 /* RMBE within RMB */
 240                 ((conn->peer_conn_idx - 1) * conn->peer_rmbe_size) +
 241                 /* offset within RMBE */
 242                 peer_rmbe_offset;
 243         rdma_wr.rkey = lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey;
 244         rc = ib_post_send(link->roce_qp, &rdma_wr.wr, &failed_wr);
 245         if (rc)
 246                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 247         return rc;
 248 }
 249
 250 /* sndbuf consumer */
 251 static inline void smc_tx_advance_cursors(struct smc_connection *conn,
 252                                           union smc_host_cursor *prod,
 253                                           union smc_host_cursor *sent,
 254                                           size_t len)
 255 {
 256         smc_curs_add(conn->peer_rmbe_size, prod, len);
 257         /* increased in recv tasklet smc_cdc_msg_rcv() */
 258         smp_mb__before_atomic();
 259         /* data in flight reduces usable snd_wnd */
 260         atomic_sub(len, &conn->peer_rmbe_space);
 261         /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
 262         smp_mb__after_atomic();
 263         smc_curs_add(conn->sndbuf_size, sent, len);
 264 }
 265
 266 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
 267  * usable snd_wnd as max transmit
 268  */
 269 static int smc_tx_rdma_writes(struct smc_connection *conn)
 270 {
 271         size_t src_off, src_len, dst_off, dst_len; /* current chunk values */
 272         size_t len, dst_len_sum, src_len_sum, dstchunk, srcchunk;
 273         union smc_host_cursor sent, prep, prod, cons;
 274         struct ib_sge sges[SMC_IB_MAX_SEND_SGE];
 275         struct smc_link_group *lgr = conn->lgr;
 276         int to_send, rmbespace;
 277         struct smc_link *link;
 278         int num_sges;
 279         int rc;
 280
 281         /* source: sndbuf */
 282         smc_curs_write(&sent, smc_curs_read(&conn->tx_curs_sent, conn), conn);
 283         smc_curs_write(&prep, smc_curs_read(&conn->tx_curs_prep, conn), conn);
 284         /* cf. wmem_alloc - (snd_max - snd_una) */
 285         to_send = smc_curs_diff(conn->sndbuf_size, &sent, &prep);
 286         if (to_send <= 0)
 287                 return 0;
 288
 289         /* destination: RMBE */
 290         /* cf. snd_wnd */
 291         rmbespace = atomic_read(&conn->peer_rmbe_space);
 292         if (rmbespace <= 0)
 293                 return 0;
 294         smc_curs_write(&prod,
 295                        smc_curs_read(&conn->local_tx_ctrl.prod, conn),
 296                        conn);
 297         smc_curs_write(&cons,
 298                        smc_curs_read(&conn->local_rx_ctrl.cons, conn),
 299                        conn);
 300
 301         /* if usable snd_wnd closes ask peer to advertise once it opens again */
 302         conn->local_tx_ctrl.prod_flags.write_blocked = (to_send >= rmbespace);
 303         /* cf. usable snd_wnd */
 304         len = min(to_send, rmbespace);
 305
 306         /* initialize variables for first iteration of subsequent nested loop */
 307         link = &lgr->lnk[SMC_SINGLE_LINK];
 308         dst_off = prod.count;
 309         if (prod.wrap == cons.wrap) {
 310                 /* the filled destination area is unwrapped,
 311                  * hence the available free destination space is wrapped
 312                  * and we need 2 destination chunks of sum len; start with 1st
 313                  * which is limited by what's available in sndbuf
 314                  */
 315                 dst_len = min_t(size_t,
 316                                 conn->peer_rmbe_size - prod.count, len);
 317         } else {
 318                 /* the filled destination area is wrapped,
 319                  * hence the available free destination space is unwrapped
 320                  * and we need a single destination chunk of entire len
 321                  */
 322                 dst_len = len;
 323         }
 324         dst_len_sum = dst_len;
 325         src_off = sent.count;
 326         /* dst_len determines the maximum src_len */
 327         if (sent.count + dst_len <= conn->sndbuf_size) {
 328                 /* unwrapped src case: single chunk of entire dst_len */
 329                 src_len = dst_len;
 330         } else {
 331                 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
 332                 src_len = conn->sndbuf_size - sent.count;
 333         }
 334         src_len_sum = src_len;
 335         for (dstchunk = 0; dstchunk < 2; dstchunk++) {
 336                 num_sges = 0;
 337                 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
 338                         sges[srcchunk].addr =
 339                                 conn->sndbuf_desc->dma_addr[SMC_SINGLE_LINK] +
 340                                 src_off;
 341                         sges[srcchunk].length = src_len;
 342                         sges[srcchunk].lkey = link->roce_pd->local_dma_lkey;
 343                         num_sges++;
 344                         src_off += src_len;
 345                         if (src_off >= conn->sndbuf_size)
 346                                 src_off -= conn->sndbuf_size;
 347                                                 /* modulo in send ring */
 348                         if (src_len_sum == dst_len)
 349                                 break; /* either on 1st or 2nd iteration */
 350                         /* prepare next (== 2nd) iteration */
 351                         src_len = dst_len - src_len; /* remainder */
 352                         src_len_sum += src_len;
 353                 }
 354                 rc = smc_tx_rdma_write(conn, dst_off, num_sges, sges);
 355                 if (rc)
 356                         return rc;
 357                 if (dst_len_sum == len)
 358                         break; /* either on 1st or 2nd iteration */
 359                 /* prepare next (== 2nd) iteration */
 360                 dst_off = 0; /* modulo offset in RMBE ring buffer */
 361                 dst_len = len - dst_len; /* remainder */
 362                 dst_len_sum += dst_len;
 363                 src_len = min_t(int,
 364                                 dst_len, conn->sndbuf_size - sent.count);
 365                 src_len_sum = src_len;
 366         }
 367
 368         smc_tx_advance_cursors(conn, &prod, &sent, len);
 369         /* update connection's cursors with advanced local cursors */
 370         smc_curs_write(&conn->local_tx_ctrl.prod,
 371                        smc_curs_read(&prod, conn),
 372                        conn);
 373                                                         /* dst: peer RMBE */
 374         smc_curs_write(&conn->tx_curs_sent,
 375                        smc_curs_read(&sent, conn),
 376                        conn);
 377                                                         /* src: local sndbuf */
 378
 379         return 0;
 380 }
 381
 382 /* Wakeup sndbuf consumers from any context (IRQ or process)
 383  * since there is more data to transmit; usable snd_wnd as max transmit
 384  */
 385 int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
 386 {
 387         struct smc_cdc_tx_pend *pend;
 388         struct smc_wr_buf *wr_buf;
 389         int rc;
 390
 391         spin_lock_bh(&conn->send_lock);
 392         rc = smc_cdc_get_free_slot(&conn->lgr->lnk[SMC_SINGLE_LINK], &wr_buf,
 393                                    &pend);
 394         if (rc < 0) {
 395                 if (rc == -EBUSY) {
 396                         struct smc_sock *smc =
 397                                 container_of(conn, struct smc_sock, conn);
 398
 399                         if (smc->sk.sk_err == ECONNABORTED) {
 400                                 rc = sock_error(&smc->sk);
 401                                 goto out_unlock;
 402                         }
 403                         rc = 0;
 404                         schedule_work(&conn->tx_work);
 405                 }
 406                 goto out_unlock;
 407         }
 408
 409         rc = smc_tx_rdma_writes(conn);
 410         if (rc) {
 411                 smc_wr_tx_put_slot(&conn->lgr->lnk[SMC_SINGLE_LINK],
 412                                    (struct smc_wr_tx_pend_priv *)pend);
 413                 goto out_unlock;
 414         }
 415
 416         rc = smc_cdc_msg_send(conn, wr_buf, pend);
 417
 418 out_unlock:
 419         spin_unlock_bh(&conn->send_lock);
 420         return rc;
 421 }
 422
 423 /* Wakeup sndbuf consumers from process context
 424  * since there is more data to transmit
 425  */
 426 static void smc_tx_work(struct work_struct *work)
 427 {
 428         struct smc_connection *conn = container_of(work,
 429                                                    struct smc_connection,
 430                                                    tx_work);
 431         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 432
 433         lock_sock(&smc->sk);
 434         smc_tx_sndbuf_nonempty(conn);
 435         release_sock(&smc->sk);
 436 }
 437
 438 void smc_tx_consumer_update(struct smc_connection *conn)
 439 {
 440         union smc_host_cursor cfed, cons;
 441         struct smc_cdc_tx_pend *pend;
 442         struct smc_wr_buf *wr_buf;
 443         int to_confirm, rc;
 444
 445         smc_curs_write(&cons,
 446                        smc_curs_read(&conn->local_tx_ctrl.cons, conn),
 447                        conn);
 448         smc_curs_write(&cfed,
 449                        smc_curs_read(&conn->rx_curs_confirmed, conn),
 450                        conn);
 451         to_confirm = smc_curs_diff(conn->rmbe_size, &cfed, &cons);
 452
 453         if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
 454             ((to_confirm > conn->rmbe_update_limit) &&
 455              ((to_confirm > (conn->rmbe_size / 2)) ||
 456               conn->local_rx_ctrl.prod_flags.write_blocked))) {
 457                 rc = smc_cdc_get_free_slot(&conn->lgr->lnk[SMC_SINGLE_LINK],
 458                                            &wr_buf, &pend);
 459                 if (!rc)
 460                         rc = smc_cdc_msg_send(conn, wr_buf, pend);
 461                 if (rc < 0) {
 462                         schedule_work(&conn->tx_work);
 463                         return;
 464                 }
 465                 smc_curs_write(&conn->rx_curs_confirmed,
 466                                smc_curs_read(&conn->local_tx_ctrl.cons, conn),
 467                                conn);
 468                 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
 469         }
 470         if (conn->local_rx_ctrl.prod_flags.write_blocked &&
 471             !atomic_read(&conn->bytes_to_rcv))
 472                 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
 473 }
 474
 475 /***************************** send initialize *******************************/
 476
 477 /* Initialize send properties on connection establishment. NB: not __init! */
 478 void smc_tx_init(struct smc_sock *smc)
 479 {
 480         smc->sk.sk_write_space = smc_tx_write_space;
 481         INIT_WORK(&smc->conn.tx_work, smc_tx_work);
 482         spin_lock_init(&smc->conn.send_lock);
 483 }