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 #include <net/tcp.h>
  23
  24 #include "smc.h"
  25 #include "smc_wr.h"
  26 #include "smc_cdc.h"
  27 #include "smc_close.h"
  28 #include "smc_ism.h"
  29 #include "smc_tx.h"
  30 #include "smc_stats.h"
  31
  32 #define SMC_TX_WORK_DELAY       0
  33 #define SMC_TX_CORK_DELAY       (HZ >> 2)       /* 250 ms */
  34
  35 /***************************** sndbuf producer *******************************/
  36
  37 /* callback implementation for sk.sk_write_space()
  38  * to wakeup sndbuf producers that blocked with smc_tx_wait().
  39  * called under sk_socket lock.
  40  */
  41 static void smc_tx_write_space(struct sock *sk)
  42 {
  43         struct socket *sock = sk->sk_socket;
  44         struct smc_sock *smc = smc_sk(sk);
  45         struct socket_wq *wq;
  46
  47         /* similar to sk_stream_write_space */
  48         if (atomic_read(&smc->conn.sndbuf_space) && sock) {
  49                 if (test_bit(SOCK_NOSPACE, &sock->flags))
  50                         SMC_STAT_RMB_TX_FULL(smc, !smc->conn.lnk);
  51                 clear_bit(SOCK_NOSPACE, &sock->flags);
  52                 rcu_read_lock();
  53                 wq = rcu_dereference(sk->sk_wq);
  54                 if (skwq_has_sleeper(wq))
  55                         wake_up_interruptible_poll(&wq->wait,
  56                                                    EPOLLOUT | EPOLLWRNORM |
  57                                                    EPOLLWRBAND);
  58                 if (wq && wq->fasync_list && !(sk->sk_shutdown & SEND_SHUTDOWN))
  59                         sock_wake_async(wq, SOCK_WAKE_SPACE, POLL_OUT);
  60                 rcu_read_unlock();
  61         }
  62 }
  63
  64 /* Wakeup sndbuf producers that blocked with smc_tx_wait().
  65  * Cf. tcp_data_snd_check()=>tcp_check_space()=>tcp_new_space().
  66  */
  67 void smc_tx_sndbuf_nonfull(struct smc_sock *smc)
  68 {
  69         if (smc->sk.sk_socket &&
  70             test_bit(SOCK_NOSPACE, &smc->sk.sk_socket->flags))
  71                 smc->sk.sk_write_space(&smc->sk);
  72 }
  73
  74 /* blocks sndbuf producer until at least one byte of free space available
  75  * or urgent Byte was consumed
  76  */
  77 static int smc_tx_wait(struct smc_sock *smc, int flags)
  78 {
  79         DEFINE_WAIT_FUNC(wait, woken_wake_function);
  80         struct smc_connection *conn = &smc->conn;
  81         struct sock *sk = &smc->sk;
  82         long timeo;
  83         int rc = 0;
  84
  85         /* similar to sk_stream_wait_memory */
  86         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
  87         add_wait_queue(sk_sleep(sk), &wait);
  88         while (1) {
  89                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
  90                 if (sk->sk_err ||
  91                     (sk->sk_shutdown & SEND_SHUTDOWN) ||
  92                     conn->killed ||
  93                     conn->local_tx_ctrl.conn_state_flags.peer_done_writing) {
  94                         rc = -EPIPE;
  95                         break;
  96                 }
  97                 if (smc_cdc_rxed_any_close(conn)) {
  98                         rc = -ECONNRESET;
  99                         break;
 100                 }
 101                 if (!timeo) {
 102                         /* ensure EPOLLOUT is subsequently generated */
 103                         set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 104                         rc = -EAGAIN;
 105                         break;
 106                 }
 107                 if (signal_pending(current)) {
 108                         rc = sock_intr_errno(timeo);
 109                         break;
 110                 }
 111                 sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 112                 if (atomic_read(&conn->sndbuf_space) && !conn->urg_tx_pend)
 113                         break; /* at least 1 byte of free & no urgent data */
 114                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 115                 sk_wait_event(sk, &timeo,
 116                               sk->sk_err ||
 117                               (sk->sk_shutdown & SEND_SHUTDOWN) ||
 118                               smc_cdc_rxed_any_close(conn) ||
 119                               (atomic_read(&conn->sndbuf_space) &&
 120                                !conn->urg_tx_pend),
 121                               &wait);
 122         }
 123         remove_wait_queue(sk_sleep(sk), &wait);
 124         return rc;
 125 }
 126
 127 static bool smc_tx_is_corked(struct smc_sock *smc)
 128 {
 129         struct tcp_sock *tp = tcp_sk(smc->clcsock->sk);
 130
 131         return (tp->nonagle & TCP_NAGLE_CORK) ? true : false;
 132 }
 133
 134 /* sndbuf producer: main API called by socket layer.
 135  * called under sock lock.
 136  */
 137 int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
 138 {
 139         size_t copylen, send_done = 0, send_remaining = len;
 140         size_t chunk_len, chunk_off, chunk_len_sum;
 141         struct smc_connection *conn = &smc->conn;
 142         union smc_host_cursor prep;
 143         struct sock *sk = &smc->sk;
 144         char *sndbuf_base;
 145         int tx_cnt_prep;
 146         int writespace;
 147         int rc, chunk;
 148
 149         /* This should be in poll */
 150         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 151
 152         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
 153                 rc = -EPIPE;
 154                 goto out_err;
 155         }
 156
 157         if (sk->sk_state == SMC_INIT)
 158                 return -ENOTCONN;
 159
 160         if (len > conn->sndbuf_desc->len)
 161                 SMC_STAT_RMB_TX_SIZE_SMALL(smc, !conn->lnk);
 162
 163         if (len > conn->peer_rmbe_size)
 164                 SMC_STAT_RMB_TX_PEER_SIZE_SMALL(smc, !conn->lnk);
 165
 166         if (msg->msg_flags & MSG_OOB)
 167                 SMC_STAT_INC(smc, urg_data_cnt);
 168
 169         while (msg_data_left(msg)) {
 170                 if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
 171                     (smc->sk.sk_err == ECONNABORTED) ||
 172                     conn->killed)
 173                         return -EPIPE;
 174                 if (smc_cdc_rxed_any_close(conn))
 175                         return send_done ?: -ECONNRESET;
 176
 177                 if (msg->msg_flags & MSG_OOB)
 178                         conn->local_tx_ctrl.prod_flags.urg_data_pending = 1;
 179
 180                 if (!atomic_read(&conn->sndbuf_space) || conn->urg_tx_pend) {
 181                         if (send_done)
 182                                 return send_done;
 183                         rc = smc_tx_wait(smc, msg->msg_flags);
 184                         if (rc)
 185                                 goto out_err;
 186                         continue;
 187                 }
 188
 189                 /* initialize variables for 1st iteration of subsequent loop */
 190                 /* could be just 1 byte, even after smc_tx_wait above */
 191                 writespace = atomic_read(&conn->sndbuf_space);
 192                 /* not more than what user space asked for */
 193                 copylen = min_t(size_t, send_remaining, writespace);
 194                 /* determine start of sndbuf */
 195                 sndbuf_base = conn->sndbuf_desc->cpu_addr;
 196                 smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
 197                 tx_cnt_prep = prep.count;
 198                 /* determine chunks where to write into sndbuf */
 199                 /* either unwrapped case, or 1st chunk of wrapped case */
 200                 chunk_len = min_t(size_t, copylen, conn->sndbuf_desc->len -
 201                                   tx_cnt_prep);
 202                 chunk_len_sum = chunk_len;
 203                 chunk_off = tx_cnt_prep;
 204                 smc_sndbuf_sync_sg_for_cpu(conn);
 205                 for (chunk = 0; chunk < 2; chunk++) {
 206                         rc = memcpy_from_msg(sndbuf_base + chunk_off,
 207                                              msg, chunk_len);
 208                         if (rc) {
 209                                 smc_sndbuf_sync_sg_for_device(conn);
 210                                 if (send_done)
 211                                         return send_done;
 212                                 goto out_err;
 213                         }
 214                         send_done += chunk_len;
 215                         send_remaining -= chunk_len;
 216
 217                         if (chunk_len_sum == copylen)
 218                                 break; /* either on 1st or 2nd iteration */
 219                         /* prepare next (== 2nd) iteration */
 220                         chunk_len = copylen - chunk_len; /* remainder */
 221                         chunk_len_sum += chunk_len;
 222                         chunk_off = 0; /* modulo offset in send ring buffer */
 223                 }
 224                 smc_sndbuf_sync_sg_for_device(conn);
 225                 /* update cursors */
 226                 smc_curs_add(conn->sndbuf_desc->len, &prep, copylen);
 227                 smc_curs_copy(&conn->tx_curs_prep, &prep, conn);
 228                 /* increased in send tasklet smc_cdc_tx_handler() */
 229                 smp_mb__before_atomic();
 230                 atomic_sub(copylen, &conn->sndbuf_space);
 231                 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
 232                 smp_mb__after_atomic();
 233                 /* since we just produced more new data into sndbuf,
 234                  * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
 235                  */
 236                 if ((msg->msg_flags & MSG_OOB) && !send_remaining)
 237                         conn->urg_tx_pend = true;
 238                 if ((msg->msg_flags & MSG_MORE || smc_tx_is_corked(smc)) &&
 239                     (atomic_read(&conn->sndbuf_space) >
 240                                                 (conn->sndbuf_desc->len >> 1)))
 241                         /* for a corked socket defer the RDMA writes if there
 242                          * is still sufficient sndbuf_space available
 243                          */
 244                         queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
 245                                            SMC_TX_CORK_DELAY);
 246                 else
 247                         smc_tx_sndbuf_nonempty(conn);
 248         } /* while (msg_data_left(msg)) */
 249
 250         return send_done;
 251
 252 out_err:
 253         rc = sk_stream_error(sk, msg->msg_flags, rc);
 254         /* make sure we wake any epoll edge trigger waiter */
 255         if (unlikely(rc == -EAGAIN))
 256                 sk->sk_write_space(sk);
 257         return rc;
 258 }
 259
 260 /***************************** sndbuf consumer *******************************/
 261
 262 /* sndbuf consumer: actual data transfer of one target chunk with ISM write */
 263 int smcd_tx_ism_write(struct smc_connection *conn, void *data, size_t len,
 264                       u32 offset, int signal)
 265 {
 266         struct smc_ism_position pos;
 267         int rc;
 268
 269         memset(&pos, 0, sizeof(pos));
 270         pos.token = conn->peer_token;
 271         pos.index = conn->peer_rmbe_idx;
 272         pos.offset = conn->tx_off + offset;
 273         pos.signal = signal;
 274         rc = smc_ism_write(conn->lgr->smcd, &pos, data, len);
 275         if (rc)
 276                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 277         return rc;
 278 }
 279
 280 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
 281 static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
 282                              int num_sges, struct ib_rdma_wr *rdma_wr)
 283 {
 284         struct smc_link_group *lgr = conn->lgr;
 285         struct smc_link *link = conn->lnk;
 286         int rc;
 287
 288         rdma_wr->wr.wr_id = smc_wr_tx_get_next_wr_id(link);
 289         rdma_wr->wr.num_sge = num_sges;
 290         rdma_wr->remote_addr =
 291                 lgr->rtokens[conn->rtoken_idx][link->link_idx].dma_addr +
 292                 /* RMBE within RMB */
 293                 conn->tx_off +
 294                 /* offset within RMBE */
 295                 peer_rmbe_offset;
 296         rdma_wr->rkey = lgr->rtokens[conn->rtoken_idx][link->link_idx].rkey;
 297         rc = ib_post_send(link->roce_qp, &rdma_wr->wr, NULL);
 298         if (rc)
 299                 smcr_link_down_cond_sched(link);
 300         return rc;
 301 }
 302
 303 /* sndbuf consumer */
 304 static inline void smc_tx_advance_cursors(struct smc_connection *conn,
 305                                           union smc_host_cursor *prod,
 306                                           union smc_host_cursor *sent,
 307                                           size_t len)
 308 {
 309         smc_curs_add(conn->peer_rmbe_size, prod, len);
 310         /* increased in recv tasklet smc_cdc_msg_rcv() */
 311         smp_mb__before_atomic();
 312         /* data in flight reduces usable snd_wnd */
 313         atomic_sub(len, &conn->peer_rmbe_space);
 314         /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
 315         smp_mb__after_atomic();
 316         smc_curs_add(conn->sndbuf_desc->len, sent, len);
 317 }
 318
 319 /* SMC-R helper for smc_tx_rdma_writes() */
 320 static int smcr_tx_rdma_writes(struct smc_connection *conn, size_t len,
 321                                size_t src_off, size_t src_len,
 322                                size_t dst_off, size_t dst_len,
 323                                struct smc_rdma_wr *wr_rdma_buf)
 324 {
 325         struct smc_link *link = conn->lnk;
 326
 327         dma_addr_t dma_addr =
 328                 sg_dma_address(conn->sndbuf_desc->sgt[link->link_idx].sgl);
 329         int src_len_sum = src_len, dst_len_sum = dst_len;
 330         int sent_count = src_off;
 331         int srcchunk, dstchunk;
 332         int num_sges;
 333         int rc;
 334
 335         for (dstchunk = 0; dstchunk < 2; dstchunk++) {
 336                 struct ib_sge *sge =
 337                         wr_rdma_buf->wr_tx_rdma[dstchunk].wr.sg_list;
 338
 339                 num_sges = 0;
 340                 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
 341                         sge[srcchunk].addr = dma_addr + src_off;
 342                         sge[srcchunk].length = src_len;
 343                         num_sges++;
 344
 345                         src_off += src_len;
 346                         if (src_off >= conn->sndbuf_desc->len)
 347                                 src_off -= conn->sndbuf_desc->len;
 348                                                 /* modulo in send ring */
 349                         if (src_len_sum == dst_len)
 350                                 break; /* either on 1st or 2nd iteration */
 351                         /* prepare next (== 2nd) iteration */
 352                         src_len = dst_len - src_len; /* remainder */
 353                         src_len_sum += src_len;
 354                 }
 355                 rc = smc_tx_rdma_write(conn, dst_off, num_sges,
 356                                        &wr_rdma_buf->wr_tx_rdma[dstchunk]);
 357                 if (rc)
 358                         return rc;
 359                 if (dst_len_sum == len)
 360                         break; /* either on 1st or 2nd iteration */
 361                 /* prepare next (== 2nd) iteration */
 362                 dst_off = 0; /* modulo offset in RMBE ring buffer */
 363                 dst_len = len - dst_len; /* remainder */
 364                 dst_len_sum += dst_len;
 365                 src_len = min_t(int, dst_len, conn->sndbuf_desc->len -
 366                                 sent_count);
 367                 src_len_sum = src_len;
 368         }
 369         return 0;
 370 }
 371
 372 /* SMC-D helper for smc_tx_rdma_writes() */
 373 static int smcd_tx_rdma_writes(struct smc_connection *conn, size_t len,
 374                                size_t src_off, size_t src_len,
 375                                size_t dst_off, size_t dst_len)
 376 {
 377         int src_len_sum = src_len, dst_len_sum = dst_len;
 378         int srcchunk, dstchunk;
 379         int rc;
 380
 381         for (dstchunk = 0; dstchunk < 2; dstchunk++) {
 382                 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
 383                         void *data = conn->sndbuf_desc->cpu_addr + src_off;
 384
 385                         rc = smcd_tx_ism_write(conn, data, src_len, dst_off +
 386                                                sizeof(struct smcd_cdc_msg), 0);
 387                         if (rc)
 388                                 return rc;
 389                         dst_off += src_len;
 390                         src_off += src_len;
 391                         if (src_off >= conn->sndbuf_desc->len)
 392                                 src_off -= conn->sndbuf_desc->len;
 393                                                 /* modulo in send ring */
 394                         if (src_len_sum == dst_len)
 395                                 break; /* either on 1st or 2nd iteration */
 396                         /* prepare next (== 2nd) iteration */
 397                         src_len = dst_len - src_len; /* remainder */
 398                         src_len_sum += src_len;
 399                 }
 400                 if (dst_len_sum == len)
 401                         break; /* either on 1st or 2nd iteration */
 402                 /* prepare next (== 2nd) iteration */
 403                 dst_off = 0; /* modulo offset in RMBE ring buffer */
 404                 dst_len = len - dst_len; /* remainder */
 405                 dst_len_sum += dst_len;
 406                 src_len = min_t(int, dst_len, conn->sndbuf_desc->len - src_off);
 407                 src_len_sum = src_len;
 408         }
 409         return 0;
 410 }
 411
 412 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
 413  * usable snd_wnd as max transmit
 414  */
 415 static int smc_tx_rdma_writes(struct smc_connection *conn,
 416                               struct smc_rdma_wr *wr_rdma_buf)
 417 {
 418         size_t len, src_len, dst_off, dst_len; /* current chunk values */
 419         union smc_host_cursor sent, prep, prod, cons;
 420         struct smc_cdc_producer_flags *pflags;
 421         int to_send, rmbespace;
 422         int rc;
 423
 424         /* source: sndbuf */
 425         smc_curs_copy(&sent, &conn->tx_curs_sent, conn);
 426         smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
 427         /* cf. wmem_alloc - (snd_max - snd_una) */
 428         to_send = smc_curs_diff(conn->sndbuf_desc->len, &sent, &prep);
 429         if (to_send <= 0)
 430                 return 0;
 431
 432         /* destination: RMBE */
 433         /* cf. snd_wnd */
 434         rmbespace = atomic_read(&conn->peer_rmbe_space);
 435         if (rmbespace <= 0) {
 436                 struct smc_sock *smc = container_of(conn, struct smc_sock,
 437                                                     conn);
 438                 SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
 439                 return 0;
 440         }
 441         smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
 442         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
 443
 444         /* if usable snd_wnd closes ask peer to advertise once it opens again */
 445         pflags = &conn->local_tx_ctrl.prod_flags;
 446         pflags->write_blocked = (to_send >= rmbespace);
 447         /* cf. usable snd_wnd */
 448         len = min(to_send, rmbespace);
 449
 450         /* initialize variables for first iteration of subsequent nested loop */
 451         dst_off = prod.count;
 452         if (prod.wrap == cons.wrap) {
 453                 /* the filled destination area is unwrapped,
 454                  * hence the available free destination space is wrapped
 455                  * and we need 2 destination chunks of sum len; start with 1st
 456                  * which is limited by what's available in sndbuf
 457                  */
 458                 dst_len = min_t(size_t,
 459                                 conn->peer_rmbe_size - prod.count, len);
 460         } else {
 461                 /* the filled destination area is wrapped,
 462                  * hence the available free destination space is unwrapped
 463                  * and we need a single destination chunk of entire len
 464                  */
 465                 dst_len = len;
 466         }
 467         /* dst_len determines the maximum src_len */
 468         if (sent.count + dst_len <= conn->sndbuf_desc->len) {
 469                 /* unwrapped src case: single chunk of entire dst_len */
 470                 src_len = dst_len;
 471         } else {
 472                 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
 473                 src_len = conn->sndbuf_desc->len - sent.count;
 474         }
 475
 476         if (conn->lgr->is_smcd)
 477                 rc = smcd_tx_rdma_writes(conn, len, sent.count, src_len,
 478                                          dst_off, dst_len);
 479         else
 480                 rc = smcr_tx_rdma_writes(conn, len, sent.count, src_len,
 481                                          dst_off, dst_len, wr_rdma_buf);
 482         if (rc)
 483                 return rc;
 484
 485         if (conn->urg_tx_pend && len == to_send)
 486                 pflags->urg_data_present = 1;
 487         smc_tx_advance_cursors(conn, &prod, &sent, len);
 488         /* update connection's cursors with advanced local cursors */
 489         smc_curs_copy(&conn->local_tx_ctrl.prod, &prod, conn);
 490                                                         /* dst: peer RMBE */
 491         smc_curs_copy(&conn->tx_curs_sent, &sent, conn);/* src: local sndbuf */
 492
 493         return 0;
 494 }
 495
 496 /* Wakeup sndbuf consumers from any context (IRQ or process)
 497  * since there is more data to transmit; usable snd_wnd as max transmit
 498  */
 499 static int _smcr_tx_sndbuf_nonempty(struct smc_connection *conn)
 500 {
 501         struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
 502         struct smc_link *link = conn->lnk;
 503         struct smc_rdma_wr *wr_rdma_buf;
 504         struct smc_cdc_tx_pend *pend;
 505         struct smc_wr_buf *wr_buf;
 506         int rc;
 507
 508         rc = smc_cdc_get_free_slot(conn, link, &wr_buf, &wr_rdma_buf, &pend);
 509         if (rc < 0) {
 510                 if (rc == -EBUSY) {
 511                         struct smc_sock *smc =
 512                                 container_of(conn, struct smc_sock, conn);
 513
 514                         if (smc->sk.sk_err == ECONNABORTED)
 515                                 return sock_error(&smc->sk);
 516                         if (conn->killed)
 517                                 return -EPIPE;
 518                         rc = 0;
 519                         mod_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
 520                                          SMC_TX_WORK_DELAY);
 521                 }
 522                 return rc;
 523         }
 524
 525         spin_lock_bh(&conn->send_lock);
 526         if (link != conn->lnk) {
 527                 /* link of connection changed, tx_work will restart */
 528                 smc_wr_tx_put_slot(link,
 529                                    (struct smc_wr_tx_pend_priv *)pend);
 530                 rc = -ENOLINK;
 531                 goto out_unlock;
 532         }
 533         if (!pflags->urg_data_present) {
 534                 rc = smc_tx_rdma_writes(conn, wr_rdma_buf);
 535                 if (rc) {
 536                         smc_wr_tx_put_slot(link,
 537                                            (struct smc_wr_tx_pend_priv *)pend);
 538                         goto out_unlock;
 539                 }
 540         }
 541
 542         rc = smc_cdc_msg_send(conn, wr_buf, pend);
 543         if (!rc && pflags->urg_data_present) {
 544                 pflags->urg_data_pending = 0;
 545                 pflags->urg_data_present = 0;
 546         }
 547
 548 out_unlock:
 549         spin_unlock_bh(&conn->send_lock);
 550         return rc;
 551 }
 552
 553 static int smcr_tx_sndbuf_nonempty(struct smc_connection *conn)
 554 {
 555         struct smc_link *link = conn->lnk;
 556         int rc = -ENOLINK;
 557
 558         if (!link)
 559                 return rc;
 560
 561         atomic_inc(&link->wr_tx_refcnt);
 562         if (smc_link_usable(link))
 563                 rc = _smcr_tx_sndbuf_nonempty(conn);
 564         if (atomic_dec_and_test(&link->wr_tx_refcnt))
 565                 wake_up_all(&link->wr_tx_wait);
 566         return rc;
 567 }
 568
 569 static int smcd_tx_sndbuf_nonempty(struct smc_connection *conn)
 570 {
 571         struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
 572         int rc = 0;
 573
 574         spin_lock_bh(&conn->send_lock);
 575         if (!pflags->urg_data_present)
 576                 rc = smc_tx_rdma_writes(conn, NULL);
 577         if (!rc)
 578                 rc = smcd_cdc_msg_send(conn);
 579
 580         if (!rc && pflags->urg_data_present) {
 581                 pflags->urg_data_pending = 0;
 582                 pflags->urg_data_present = 0;
 583         }
 584         spin_unlock_bh(&conn->send_lock);
 585         return rc;
 586 }
 587
 588 int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
 589 {
 590         int rc;
 591
 592         if (conn->killed ||
 593             conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
 594                 return -EPIPE;  /* connection being aborted */
 595         if (conn->lgr->is_smcd)
 596                 rc = smcd_tx_sndbuf_nonempty(conn);
 597         else
 598                 rc = smcr_tx_sndbuf_nonempty(conn);
 599
 600         if (!rc) {
 601                 /* trigger socket release if connection is closing */
 602                 struct smc_sock *smc = container_of(conn, struct smc_sock,
 603                                                     conn);
 604                 smc_close_wake_tx_prepared(smc);
 605         }
 606         return rc;
 607 }
 608
 609 /* Wakeup sndbuf consumers from process context
 610  * since there is more data to transmit
 611  */
 612 void smc_tx_work(struct work_struct *work)
 613 {
 614         struct smc_connection *conn = container_of(to_delayed_work(work),
 615                                                    struct smc_connection,
 616                                                    tx_work);
 617         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 618         int rc;
 619
 620         lock_sock(&smc->sk);
 621         if (smc->sk.sk_err)
 622                 goto out;
 623
 624         rc = smc_tx_sndbuf_nonempty(conn);
 625         if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
 626             !atomic_read(&conn->bytes_to_rcv))
 627                 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
 628
 629 out:
 630         release_sock(&smc->sk);
 631 }
 632
 633 void smc_tx_consumer_update(struct smc_connection *conn, bool force)
 634 {
 635         union smc_host_cursor cfed, cons, prod;
 636         int sender_free = conn->rmb_desc->len;
 637         int to_confirm;
 638
 639         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
 640         smc_curs_copy(&cfed, &conn->rx_curs_confirmed, conn);
 641         to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);
 642         if (to_confirm > conn->rmbe_update_limit) {
 643                 smc_curs_copy(&prod, &conn->local_rx_ctrl.prod, conn);
 644                 sender_free = conn->rmb_desc->len -
 645                               smc_curs_diff_large(conn->rmb_desc->len,
 646                                                   &cfed, &prod);
 647         }
 648
 649         if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
 650             force ||
 651             ((to_confirm > conn->rmbe_update_limit) &&
 652              ((sender_free <= (conn->rmb_desc->len / 2)) ||
 653               conn->local_rx_ctrl.prod_flags.write_blocked))) {
 654                 if (conn->killed ||
 655                     conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
 656                         return;
 657                 if ((smc_cdc_get_slot_and_msg_send(conn) < 0) &&
 658                     !conn->killed) {
 659                         queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
 660                                            SMC_TX_WORK_DELAY);
 661                         return;
 662                 }
 663         }
 664         if (conn->local_rx_ctrl.prod_flags.write_blocked &&
 665             !atomic_read(&conn->bytes_to_rcv))
 666                 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
 667 }
 668
 669 /***************************** send initialize *******************************/
 670
 671 /* Initialize send properties on connection establishment. NB: not __init! */
 672 void smc_tx_init(struct smc_sock *smc)
 673 {
 674         smc->sk.sk_write_space = smc_tx_write_space;
 675 }