net/smc/smc_tx.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  * Manage send buffer.
   6  * Producer:
   7  * Copy user space data into send buffer, if send buffer space available.
   8  * Consumer:
   9  * Trigger RDMA write into RMBE of peer and send CDC, if RMBE space available.
  10  *
  11  * Copyright IBM Corp. 2016
  12  *
  13  * Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  14  */
  15
  16 #include <linux/net.h>
  17 #include <linux/rcupdate.h>
  18 #include <linux/workqueue.h>
  19 #include <linux/sched/signal.h>
  20
  21 #include <net/sock.h>
  22
  23 #include "smc.h"
  24 #include "smc_wr.h"
  25 #include "smc_cdc.h"
  26 #include "smc_tx.h"
  27
  28 #define SMC_TX_WORK_DELAY       HZ
  29
  30 /***************************** sndbuf producer *******************************/
  31
  32 /* callback implementation for sk.sk_write_space()
  33  * to wakeup sndbuf producers that blocked with smc_tx_wait_memory().
  34  * called under sk_socket lock.
  35  */
  36 static void smc_tx_write_space(struct sock *sk)
  37 {
  38         struct socket *sock = sk->sk_socket;
  39         struct smc_sock *smc = smc_sk(sk);
  40         struct socket_wq *wq;
  41
  42         /* similar to sk_stream_write_space */
  43         if (atomic_read(&smc->conn.sndbuf_space) && sock) {
  44                 clear_bit(SOCK_NOSPACE, &sock->flags);
  45                 rcu_read_lock();
  46                 wq = rcu_dereference(sk->sk_wq);
  47                 if (skwq_has_sleeper(wq))
  48                         wake_up_interruptible_poll(&wq->wait,
  49                                                    POLLOUT | POLLWRNORM |
  50                                                    POLLWRBAND);
  51                 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))
  52                         sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT);
  53                 rcu_read_unlock();
  54         }
  55 }
  56
  57 /* Wakeup sndbuf producers that blocked with smc_tx_wait_memory().
  58  * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
  59  */
  60 void smc_tx_sndbuf_nonfull(struct smc_sock *smc)
  61 {
  62         if (smc->sk.sk_socket &&
  63             test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))
  64                 smc->sk.sk_write_space(&smc->sk);
  65 }
  66
  67 /* blocks sndbuf producer until at least one byte of free space available */
  68 static int smc_tx_wait_memory(struct smc_sock *smc, int flags)
  69 {
  70         DEFINE_WAIT_FUNC(wait, woken_wake_function);
  71         struct smc_connection *conn = &smc->conn;
  72         struct sock *sk = &smc->sk;
  73         long timeo;
  74         int rc = 0;
  75
  76         /* similar to sk_stream_wait_memory */
  77         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
  78         add_wait_queue(sk_sleep(sk), &wait);
  79         while (1) {
  80                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
  81                 if (sk->sk_err ||
  82                     (sk->sk_shutdown & SEND_SHUTDOWN) ||
  83                     conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {
  84                         rc = -EPIPE;
  85                         break;
  86                 }
  87                 if (conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
  88                         rc = -ECONNRESET;
  89                         break;
  90                 }
  91                 if (!timeo) {
  92                         /* ensure EPOLLOUT is subsequently generated */
  93                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
  94                         rc = -EAGAIN;
  95                         break;
  96                 }
  97                 if (signal_pending(current)) {
  98                         rc = sock_intr_errno(timeo);
  99                         break;
 100                 }
 101                 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 102                 if (atomic_read(&conn->sndbuf_space))
 103                         break; /* at least 1 byte of free space available */
 104                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 105                 sk->sk_write_pending++;
 106                 sk_wait_event(sk, &timeo,
 107                               sk->sk_err ||
 108                               (sk->sk_shutdown & SEND_SHUTDOWN) ||
 109                               smc_cdc_rxed_any_close_or_senddone(conn) ||
 110                               atomic_read(&conn->sndbuf_space),
 111                               &wait);
 112                 sk->sk_write_pending--;
 113         }
 114         remove_wait_queue(sk_sleep(sk), &wait);
 115         return rc;
 116 }
 117
 118 /* sndbuf producer: main API called by socket layer.
 119  * called under sock lock.
 120  */
 121 int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
 122 {
 123         size_t copylen, send_done = 0, send_remaining = len;
 124         size_t chunk_len, chunk_off, chunk_len_sum;
 125         struct smc_connection *conn = &smc->conn;
 126         union smc_host_cursor prep;
 127         struct sock *sk = &smc->sk;
 128         char *sndbuf_base;
 129         int tx_cnt_prep;
 130         int writespace;
 131         int rc, chunk;
 132
 133         /* This should be in poll */
 134         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 135
 136         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
 137                 rc = -EPIPE;
 138                 goto out_err;
 139         }
 140
 141         while (msg_data_left(msg)) {
 142                 if (sk->sk_state == SMC_INIT)
 143                         return -ENOTCONN;
 144                 if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
 145                     (smc->sk.sk_err == ECONNABORTED) ||
 146                     conn->local_tx_ctrl.conn_state_flags.peer_conn_abort)
 147                         return -EPIPE;
 148                 if (smc_cdc_rxed_any_close(conn))
 149                         return send_done ?: -ECONNRESET;
 150
 151                 if (!atomic_read(&conn->sndbuf_space)) {
 152                         if (send_done)
 153                                 return send_done;
 154                         rc = smc_tx_wait_memory(smc, msg->msg_flags);
 155                         if (rc)
 156                                 goto out_err;
 157                         continue;
 158                 }
 159
 160                 /* initialize variables for 1st iteration of subsequent loop */
 161                 /* could be just 1 byte, even after smc_tx_wait_memory above */
 162                 writespace = atomic_read(&conn->sndbuf_space);
 163                 /* not more than what user space asked for */
 164                 copylen = min_t(size_t, send_remaining, writespace);
 165                 /* determine start of sndbuf */
 166                 sndbuf_base = conn->sndbuf_desc->cpu_addr;
 167                 smc_curs_write(&prep,
 168                                smc_curs_read(&conn->tx_curs_prep, conn),
 169                                conn);
 170                 tx_cnt_prep = prep.count;
 171                 /* determine chunks where to write into sndbuf */
 172                 /* either unwrapped case, or 1st chunk of wrapped case */
 173                 chunk_len = min_t(size_t,
 174                                   copylen, conn->sndbuf_size - tx_cnt_prep);
 175                 chunk_len_sum = chunk_len;
 176                 chunk_off = tx_cnt_prep;
 177                 smc_sndbuf_sync_sg_for_cpu(conn);
 178                 for (chunk = 0; chunk < 2; chunk++) {
 179                         rc = memcpy_from_msg(sndbuf_base + chunk_off,
 180                                              msg, chunk_len);
 181                         if (rc) {
 182                                 smc_sndbuf_sync_sg_for_device(conn);
 183                                 if (send_done)
 184                                         return send_done;
 185                                 goto out_err;
 186                         }
 187                         send_done += chunk_len;
 188                         send_remaining -= chunk_len;
 189
 190                         if (chunk_len_sum == copylen)
 191                                 break; /* either on 1st or 2nd iteration */
 192                         /* prepare next (== 2nd) iteration */
 193                         chunk_len = copylen - chunk_len; /* remainder */
 194                         chunk_len_sum += chunk_len;
 195                         chunk_off = 0; /* modulo offset in send ring buffer */
 196                 }
 197                 smc_sndbuf_sync_sg_for_device(conn);
 198                 /* update cursors */
 199                 smc_curs_add(conn->sndbuf_size, &prep, copylen);
 200                 smc_curs_write(&conn->tx_curs_prep,
 201                                smc_curs_read(&prep, conn),
 202                                conn);
 203                 /* increased in send tasklet smc_cdc_tx_handler() */
 204                 smp_mb__before_atomic();
 205                 atomic_sub(copylen, &conn->sndbuf_space);
 206                 /* guarantee 0 <= sndbuf_space <= sndbuf_size */
 207                 smp_mb__after_atomic();
 208                 /* since we just produced more new data into sndbuf,
 209                  * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
 210                  */
 211                 smc_tx_sndbuf_nonempty(conn);
 212         } /* while (msg_data_left(msg)) */
 213
 214         return send_done;
 215
 216 out_err:
 217         rc = sk_stream_error(sk, msg->msg_flags, rc);
 218         /* make sure we wake any epoll edge trigger waiter */
 219         if (unlikely(rc == -EAGAIN))
 220                 sk->sk_write_space(sk);
 221         return rc;
 222 }
 223
 224 /***************************** sndbuf consumer *******************************/
 225
 226 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
 227 static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
 228                              int num_sges, struct ib_sge sges[])
 229 {
 230         struct smc_link_group *lgr = conn->lgr;
 231         struct ib_send_wr *failed_wr = NULL;
 232         struct ib_rdma_wr rdma_wr;
 233         struct smc_link *link;
 234         int rc;
 235
 236         memset(&rdma_wr, 0, sizeof(rdma_wr));
 237         link = &lgr->lnk[SMC_SINGLE_LINK];
 238         rdma_wr.wr.wr_id = smc_wr_tx_get_next_wr_id(link);
 239         rdma_wr.wr.sg_list = sges;
 240         rdma_wr.wr.num_sge = num_sges;
 241         rdma_wr.wr.opcode = IB_WR_RDMA_WRITE;
 242         rdma_wr.remote_addr =
 243                 lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].dma_addr +
 244                 /* RMBE within RMB */
 245                 ((conn->peer_conn_idx - 1) * conn->peer_rmbe_size) +
 246                 /* offset within RMBE */
 247                 peer_rmbe_offset;
 248         rdma_wr.rkey = lgr->rtokens[conn->rtoken_idx][SMC_SINGLE_LINK].rkey;
 249         rc = ib_post_send(link->roce_qp, &rdma_wr.wr, &failed_wr);
 250         if (rc)
 251                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 252         return rc;
 253 }
 254
 255 /* sndbuf consumer */
 256 static inline void smc_tx_advance_cursors(struct smc_connection *conn,
 257                                           union smc_host_cursor *prod,
 258                                           union smc_host_cursor *sent,
 259                                           size_t len)
 260 {
 261         smc_curs_add(conn->peer_rmbe_size, prod, len);
 262         /* increased in recv tasklet smc_cdc_msg_rcv() */
 263         smp_mb__before_atomic();
 264         /* data in flight reduces usable snd_wnd */
 265         atomic_sub(len, &conn->peer_rmbe_space);
 266         /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
 267         smp_mb__after_atomic();
 268         smc_curs_add(conn->sndbuf_size, sent, len);
 269 }
 270
 271 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
 272  * usable snd_wnd as max transmit
 273  */
 274 static int smc_tx_rdma_writes(struct smc_connection *conn)
 275 {
 276         size_t src_off, src_len, dst_off, dst_len; /* current chunk values */
 277         size_t len, dst_len_sum, src_len_sum, dstchunk, srcchunk;
 278         union smc_host_cursor sent, prep, prod, cons;
 279         struct ib_sge sges[SMC_IB_MAX_SEND_SGE];
 280         struct smc_link_group *lgr = conn->lgr;
 281         int to_send, rmbespace;
 282         struct smc_link *link;
 283         dma_addr_t dma_addr;
 284         int num_sges;
 285         int rc;
 286
 287         /* source: sndbuf */
 288         smc_curs_write(&sent, smc_curs_read(&conn->tx_curs_sent, conn), conn);
 289         smc_curs_write(&prep, smc_curs_read(&conn->tx_curs_prep, conn), conn);
 290         /* cf. wmem_alloc - (snd_max - snd_una) */
 291         to_send = smc_curs_diff(conn->sndbuf_size, &sent, &prep);
 292         if (to_send <= 0)
 293                 return 0;
 294
 295         /* destination: RMBE */
 296         /* cf. snd_wnd */
 297         rmbespace = atomic_read(&conn->peer_rmbe_space);
 298         if (rmbespace <= 0)
 299                 return 0;
 300         smc_curs_write(&prod,
 301                        smc_curs_read(&conn->local_tx_ctrl.prod, conn),
 302                        conn);
 303         smc_curs_write(&cons,
 304                        smc_curs_read(&conn->local_rx_ctrl.cons, conn),
 305                        conn);
 306
 307         /* if usable snd_wnd closes ask peer to advertise once it opens again */
 308         conn->local_tx_ctrl.prod_flags.write_blocked = (to_send >= rmbespace);
 309         /* cf. usable snd_wnd */
 310         len = min(to_send, rmbespace);
 311
 312         /* initialize variables for first iteration of subsequent nested loop */
 313         link = &lgr->lnk[SMC_SINGLE_LINK];
 314         dst_off = prod.count;
 315         if (prod.wrap == cons.wrap) {
 316                 /* the filled destination area is unwrapped,
 317                  * hence the available free destination space is wrapped
 318                  * and we need 2 destination chunks of sum len; start with 1st
 319                  * which is limited by what's available in sndbuf
 320                  */
 321                 dst_len = min_t(size_t,
 322                                 conn->peer_rmbe_size - prod.count, len);
 323         } else {
 324                 /* the filled destination area is wrapped,
 325                  * hence the available free destination space is unwrapped
 326                  * and we need a single destination chunk of entire len
 327                  */
 328                 dst_len = len;
 329         }
 330         dst_len_sum = dst_len;
 331         src_off = sent.count;
 332         /* dst_len determines the maximum src_len */
 333         if (sent.count + dst_len <= conn->sndbuf_size) {
 334                 /* unwrapped src case: single chunk of entire dst_len */
 335                 src_len = dst_len;
 336         } else {
 337                 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
 338                 src_len = conn->sndbuf_size - sent.count;
 339         }
 340         src_len_sum = src_len;
 341         dma_addr = sg_dma_address(conn->sndbuf_desc->sgt[SMC_SINGLE_LINK].sgl);
 342         for (dstchunk = 0; dstchunk < 2; dstchunk++) {
 343                 num_sges = 0;
 344                 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
 345                         sges[srcchunk].addr = dma_addr + src_off;
 346                         sges[srcchunk].length = src_len;
 347                         sges[srcchunk].lkey = link->roce_pd->local_dma_lkey;
 348                         num_sges++;
 349                         src_off += src_len;
 350                         if (src_off >= conn->sndbuf_size)
 351                                 src_off -= conn->sndbuf_size;
 352                                                 /* modulo in send ring */
 353                         if (src_len_sum == dst_len)
 354                                 break; /* either on 1st or 2nd iteration */
 355                         /* prepare next (== 2nd) iteration */
 356                         src_len = dst_len - src_len; /* remainder */
 357                         src_len_sum += src_len;
 358                 }
 359                 rc = smc_tx_rdma_write(conn, dst_off, num_sges, sges);
 360                 if (rc)
 361                         return rc;
 362                 if (dst_len_sum == len)
 363                         break; /* either on 1st or 2nd iteration */
 364                 /* prepare next (== 2nd) iteration */
 365                 dst_off = 0; /* modulo offset in RMBE ring buffer */
 366                 dst_len = len - dst_len; /* remainder */
 367                 dst_len_sum += dst_len;
 368                 src_len = min_t(int,
 369                                 dst_len, conn->sndbuf_size - sent.count);
 370                 src_len_sum = src_len;
 371         }
 372
 373         smc_tx_advance_cursors(conn, &prod, &sent, len);
 374         /* update connection's cursors with advanced local cursors */
 375         smc_curs_write(&conn->local_tx_ctrl.prod,
 376                        smc_curs_read(&prod, conn),
 377                        conn);
 378                                                         /* dst: peer RMBE */
 379         smc_curs_write(&conn->tx_curs_sent,
 380                        smc_curs_read(&sent, conn),
 381                        conn);
 382                                                         /* src: local sndbuf */
 383
 384         return 0;
 385 }
 386
 387 /* Wakeup sndbuf consumers from any context (IRQ or process)
 388  * since there is more data to transmit; usable snd_wnd as max transmit
 389  */
 390 int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
 391 {
 392         struct smc_cdc_tx_pend *pend;
 393         struct smc_wr_buf *wr_buf;
 394         int rc;
 395
 396         spin_lock_bh(&conn->send_lock);
 397         rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend);
 398         if (rc < 0) {
 399                 if (rc == -EBUSY) {
 400                         struct smc_sock *smc =
 401                                 container_of(conn, struct smc_sock, conn);
 402
 403                         if (smc->sk.sk_err == ECONNABORTED) {
 404                                 rc = sock_error(&smc->sk);
 405                                 goto out_unlock;
 406                         }
 407                         rc = 0;
 408                         schedule_delayed_work(&conn->tx_work,
 409                                               SMC_TX_WORK_DELAY);
 410                 }
 411                 goto out_unlock;
 412         }
 413
 414         rc = smc_tx_rdma_writes(conn);
 415         if (rc) {
 416                 smc_wr_tx_put_slot(&conn->lgr->lnk[SMC_SINGLE_LINK],
 417                                    (struct smc_wr_tx_pend_priv *)pend);
 418                 goto out_unlock;
 419         }
 420
 421         rc = smc_cdc_msg_send(conn, wr_buf, pend);
 422
 423 out_unlock:
 424         spin_unlock_bh(&conn->send_lock);
 425         return rc;
 426 }
 427
 428 /* Wakeup sndbuf consumers from process context
 429  * since there is more data to transmit
 430  */
 431 static void smc_tx_work(struct work_struct *work)
 432 {
 433         struct smc_connection *conn = container_of(to_delayed_work(work),
 434                                                    struct smc_connection,
 435                                                    tx_work);
 436         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 437         int rc;
 438
 439         lock_sock(&smc->sk);
 440         rc = smc_tx_sndbuf_nonempty(conn);
 441         if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
 442             !atomic_read(&conn->bytes_to_rcv))
 443                 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
 444         release_sock(&smc->sk);
 445 }
 446
 447 void smc_tx_consumer_update(struct smc_connection *conn)
 448 {
 449         union smc_host_cursor cfed, cons;
 450         struct smc_cdc_tx_pend *pend;
 451         struct smc_wr_buf *wr_buf;
 452         int to_confirm, rc;
 453
 454         smc_curs_write(&cons,
 455                        smc_curs_read(&conn->local_tx_ctrl.cons, conn),
 456                        conn);
 457         smc_curs_write(&cfed,
 458                        smc_curs_read(&conn->rx_curs_confirmed, conn),
 459                        conn);
 460         to_confirm = smc_curs_diff(conn->rmbe_size, &cfed, &cons);
 461
 462         if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
 463             ((to_confirm > conn->rmbe_update_limit) &&
 464              ((to_confirm > (conn->rmbe_size / 2)) ||
 465               conn->local_rx_ctrl.prod_flags.write_blocked))) {
 466                 rc = smc_cdc_get_free_slot(conn, &wr_buf, &pend);
 467                 if (!rc)
 468                         rc = smc_cdc_msg_send(conn, wr_buf, pend);
 469                 if (rc < 0) {
 470                         schedule_delayed_work(&conn->tx_work,
 471                                               SMC_TX_WORK_DELAY);
 472                         return;
 473                 }
 474                 smc_curs_write(&conn->rx_curs_confirmed,
 475                                smc_curs_read(&conn->local_tx_ctrl.cons, conn),
 476                                conn);
 477                 conn->local_rx_ctrl.prod_flags.cons_curs_upd_req = 0;
 478         }
 479         if (conn->local_rx_ctrl.prod_flags.write_blocked &&
 480             !atomic_read(&conn->bytes_to_rcv))
 481                 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
 482 }
 483
 484 /***************************** send initialize *******************************/
 485
 486 /* Initialize send properties on connection establishment. NB: not __init! */
 487 void smc_tx_init(struct smc_sock *smc)
 488 {
 489         smc->sk.sk_write_space = smc_tx_write_space;
 490         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 491         spin_lock_init(&smc->conn.send_lock);
 492 }