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 <[email protected]>
  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 #include "smc_tracepoint.h"
  32
  33 #define SMC_TX_WORK_DELAY       0
  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 /* If we have pending CDC messages, do not send:
 135  * Because CQE of this CDC message will happen shortly, it gives
 136  * a chance to coalesce future sendmsg() payload in to one RDMA Write,
 137  * without need for a timer, and with no latency trade off.
 138  * Algorithm here:
 139  *  1. First message should never cork
 140  *  2. If we have pending Tx CDC messages, wait for the first CDC
 141  *     message's completion
 142  *  3. Don't cork to much data in a single RDMA Write to prevent burst
 143  *     traffic, total corked message should not exceed sendbuf/2
 144  */
 145 static bool smc_should_autocork(struct smc_sock *smc)
 146 {
 147         struct smc_connection *conn = &smc->conn;
 148         int corking_size;
 149
 150         corking_size = min_t(unsigned int, conn->sndbuf_desc->len >> 1,
 151                              sock_net(&smc->sk)->smc.sysctl_autocorking_size);
 152
 153         if (atomic_read(&conn->cdc_pend_tx_wr) == 0 ||
 154             smc_tx_prepared_sends(conn) > corking_size)
 155                 return false;
 156         return true;
 157 }
 158
 159 static bool smc_tx_should_cork(struct smc_sock *smc, struct msghdr *msg)
 160 {
 161         struct smc_connection *conn = &smc->conn;
 162
 163         if (smc_should_autocork(smc))
 164                 return true;
 165
 166         /* for a corked socket defer the RDMA writes if
 167          * sndbuf_space is still available. The applications
 168          * should known how/when to uncork it.
 169          */
 170         if ((msg->msg_flags & MSG_MORE ||
 171              smc_tx_is_corked(smc) ||
 172              msg->msg_flags & MSG_SENDPAGE_NOTLAST) &&
 173             atomic_read(&conn->sndbuf_space))
 174                 return true;
 175
 176         return false;
 177 }
 178
 179 /* sndbuf producer: main API called by socket layer.
 180  * called under sock lock.
 181  */
 182 int smc_tx_sendmsg(struct smc_sock *smc, struct msghdr *msg, size_t len)
 183 {
 184         size_t copylen, send_done = 0, send_remaining = len;
 185         size_t chunk_len, chunk_off, chunk_len_sum;
 186         struct smc_connection *conn = &smc->conn;
 187         union smc_host_cursor prep;
 188         struct sock *sk = &smc->sk;
 189         char *sndbuf_base;
 190         int tx_cnt_prep;
 191         int writespace;
 192         int rc, chunk;
 193
 194         /* This should be in poll */
 195         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
 196
 197         if (sk->sk_err || (sk->sk_shutdown & SEND_SHUTDOWN)) {
 198                 rc = -EPIPE;
 199                 goto out_err;
 200         }
 201
 202         if (sk->sk_state == SMC_INIT)
 203                 return -ENOTCONN;
 204
 205         if (len > conn->sndbuf_desc->len)
 206                 SMC_STAT_RMB_TX_SIZE_SMALL(smc, !conn->lnk);
 207
 208         if (len > conn->peer_rmbe_size)
 209                 SMC_STAT_RMB_TX_PEER_SIZE_SMALL(smc, !conn->lnk);
 210
 211         if (msg->msg_flags & MSG_OOB)
 212                 SMC_STAT_INC(smc, urg_data_cnt);
 213
 214         while (msg_data_left(msg)) {
 215                 if (smc->sk.sk_shutdown & SEND_SHUTDOWN ||
 216                     (smc->sk.sk_err == ECONNABORTED) ||
 217                     conn->killed)
 218                         return -EPIPE;
 219                 if (smc_cdc_rxed_any_close(conn))
 220                         return send_done ?: -ECONNRESET;
 221
 222                 if (msg->msg_flags & MSG_OOB)
 223                         conn->local_tx_ctrl.prod_flags.urg_data_pending = 1;
 224
 225                 if (!atomic_read(&conn->sndbuf_space) || conn->urg_tx_pend) {
 226                         if (send_done)
 227                                 return send_done;
 228                         rc = smc_tx_wait(smc, msg->msg_flags);
 229                         if (rc)
 230                                 goto out_err;
 231                         continue;
 232                 }
 233
 234                 /* initialize variables for 1st iteration of subsequent loop */
 235                 /* could be just 1 byte, even after smc_tx_wait above */
 236                 writespace = atomic_read(&conn->sndbuf_space);
 237                 /* not more than what user space asked for */
 238                 copylen = min_t(size_t, send_remaining, writespace);
 239                 /* determine start of sndbuf */
 240                 sndbuf_base = conn->sndbuf_desc->cpu_addr;
 241                 smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
 242                 tx_cnt_prep = prep.count;
 243                 /* determine chunks where to write into sndbuf */
 244                 /* either unwrapped case, or 1st chunk of wrapped case */
 245                 chunk_len = min_t(size_t, copylen, conn->sndbuf_desc->len -
 246                                   tx_cnt_prep);
 247                 chunk_len_sum = chunk_len;
 248                 chunk_off = tx_cnt_prep;
 249                 for (chunk = 0; chunk < 2; chunk++) {
 250                         rc = memcpy_from_msg(sndbuf_base + chunk_off,
 251                                              msg, chunk_len);
 252                         if (rc) {
 253                                 smc_sndbuf_sync_sg_for_device(conn);
 254                                 if (send_done)
 255                                         return send_done;
 256                                 goto out_err;
 257                         }
 258                         send_done += chunk_len;
 259                         send_remaining -= chunk_len;
 260
 261                         if (chunk_len_sum == copylen)
 262                                 break; /* either on 1st or 2nd iteration */
 263                         /* prepare next (== 2nd) iteration */
 264                         chunk_len = copylen - chunk_len; /* remainder */
 265                         chunk_len_sum += chunk_len;
 266                         chunk_off = 0; /* modulo offset in send ring buffer */
 267                 }
 268                 smc_sndbuf_sync_sg_for_device(conn);
 269                 /* update cursors */
 270                 smc_curs_add(conn->sndbuf_desc->len, &prep, copylen);
 271                 smc_curs_copy(&conn->tx_curs_prep, &prep, conn);
 272                 /* increased in send tasklet smc_cdc_tx_handler() */
 273                 smp_mb__before_atomic();
 274                 atomic_sub(copylen, &conn->sndbuf_space);
 275                 /* guarantee 0 <= sndbuf_space <= sndbuf_desc->len */
 276                 smp_mb__after_atomic();
 277                 /* since we just produced more new data into sndbuf,
 278                  * trigger sndbuf consumer: RDMA write into peer RMBE and CDC
 279                  */
 280                 if ((msg->msg_flags & MSG_OOB) && !send_remaining)
 281                         conn->urg_tx_pend = true;
 282                 /* If we need to cork, do nothing and wait for the next
 283                  * sendmsg() call or push on tx completion
 284                  */
 285                 if (!smc_tx_should_cork(smc, msg))
 286                         smc_tx_sndbuf_nonempty(conn);
 287
 288                 trace_smc_tx_sendmsg(smc, copylen);
 289         } /* while (msg_data_left(msg)) */
 290
 291         return send_done;
 292
 293 out_err:
 294         rc = sk_stream_error(sk, msg->msg_flags, rc);
 295         /* make sure we wake any epoll edge trigger waiter */
 296         if (unlikely(rc == -EAGAIN))
 297                 sk->sk_write_space(sk);
 298         return rc;
 299 }
 300
 301 int smc_tx_sendpage(struct smc_sock *smc, struct page *page, int offset,
 302                     size_t size, int flags)
 303 {
 304         struct msghdr msg = {.msg_flags = flags};
 305         char *kaddr = kmap(page);
 306         struct kvec iov;
 307         int rc;
 308
 309         iov.iov_base = kaddr + offset;
 310         iov.iov_len = size;
 311         iov_iter_kvec(&msg.msg_iter, ITER_SOURCE, &iov, 1, size);
 312         rc = smc_tx_sendmsg(smc, &msg, size);
 313         kunmap(page);
 314         return rc;
 315 }
 316
 317 /***************************** sndbuf consumer *******************************/
 318
 319 /* sndbuf consumer: actual data transfer of one target chunk with ISM write */
 320 int smcd_tx_ism_write(struct smc_connection *conn, void *data, size_t len,
 321                       u32 offset, int signal)
 322 {
 323         int rc;
 324
 325         rc = smc_ism_write(conn->lgr->smcd, conn->peer_token,
 326                            conn->peer_rmbe_idx, signal, conn->tx_off + offset,
 327                            data, len);
 328         if (rc)
 329                 conn->local_tx_ctrl.conn_state_flags.peer_conn_abort = 1;
 330         return rc;
 331 }
 332
 333 /* sndbuf consumer: actual data transfer of one target chunk with RDMA write */
 334 static int smc_tx_rdma_write(struct smc_connection *conn, int peer_rmbe_offset,
 335                              int num_sges, struct ib_rdma_wr *rdma_wr)
 336 {
 337         struct smc_link_group *lgr = conn->lgr;
 338         struct smc_link *link = conn->lnk;
 339         int rc;
 340
 341         rdma_wr->wr.wr_id = smc_wr_tx_get_next_wr_id(link);
 342         rdma_wr->wr.num_sge = num_sges;
 343         rdma_wr->remote_addr =
 344                 lgr->rtokens[conn->rtoken_idx][link->link_idx].dma_addr +
 345                 /* RMBE within RMB */
 346                 conn->tx_off +
 347                 /* offset within RMBE */
 348                 peer_rmbe_offset;
 349         rdma_wr->rkey = lgr->rtokens[conn->rtoken_idx][link->link_idx].rkey;
 350         rc = ib_post_send(link->roce_qp, &rdma_wr->wr, NULL);
 351         if (rc)
 352                 smcr_link_down_cond_sched(link);
 353         return rc;
 354 }
 355
 356 /* sndbuf consumer */
 357 static inline void smc_tx_advance_cursors(struct smc_connection *conn,
 358                                           union smc_host_cursor *prod,
 359                                           union smc_host_cursor *sent,
 360                                           size_t len)
 361 {
 362         smc_curs_add(conn->peer_rmbe_size, prod, len);
 363         /* increased in recv tasklet smc_cdc_msg_rcv() */
 364         smp_mb__before_atomic();
 365         /* data in flight reduces usable snd_wnd */
 366         atomic_sub(len, &conn->peer_rmbe_space);
 367         /* guarantee 0 <= peer_rmbe_space <= peer_rmbe_size */
 368         smp_mb__after_atomic();
 369         smc_curs_add(conn->sndbuf_desc->len, sent, len);
 370 }
 371
 372 /* SMC-R helper for smc_tx_rdma_writes() */
 373 static int smcr_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                                struct smc_rdma_wr *wr_rdma_buf)
 377 {
 378         struct smc_link *link = conn->lnk;
 379
 380         dma_addr_t dma_addr =
 381                 sg_dma_address(conn->sndbuf_desc->sgt[link->link_idx].sgl);
 382         u64 virt_addr = (uintptr_t)conn->sndbuf_desc->cpu_addr;
 383         int src_len_sum = src_len, dst_len_sum = dst_len;
 384         int sent_count = src_off;
 385         int srcchunk, dstchunk;
 386         int num_sges;
 387         int rc;
 388
 389         for (dstchunk = 0; dstchunk < 2; dstchunk++) {
 390                 struct ib_rdma_wr *wr = &wr_rdma_buf->wr_tx_rdma[dstchunk];
 391                 struct ib_sge *sge = wr->wr.sg_list;
 392                 u64 base_addr = dma_addr;
 393
 394                 if (dst_len < link->qp_attr.cap.max_inline_data) {
 395                         base_addr = virt_addr;
 396                         wr->wr.send_flags |= IB_SEND_INLINE;
 397                 } else {
 398                         wr->wr.send_flags &= ~IB_SEND_INLINE;
 399                 }
 400
 401                 num_sges = 0;
 402                 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
 403                         sge[srcchunk].addr = conn->sndbuf_desc->is_vm ?
 404                                 (virt_addr + src_off) : (base_addr + src_off);
 405                         sge[srcchunk].length = src_len;
 406                         if (conn->sndbuf_desc->is_vm)
 407                                 sge[srcchunk].lkey =
 408                                         conn->sndbuf_desc->mr[link->link_idx]->lkey;
 409                         num_sges++;
 410
 411                         src_off += src_len;
 412                         if (src_off >= conn->sndbuf_desc->len)
 413                                 src_off -= conn->sndbuf_desc->len;
 414                                                 /* modulo in send ring */
 415                         if (src_len_sum == dst_len)
 416                                 break; /* either on 1st or 2nd iteration */
 417                         /* prepare next (== 2nd) iteration */
 418                         src_len = dst_len - src_len; /* remainder */
 419                         src_len_sum += src_len;
 420                 }
 421                 rc = smc_tx_rdma_write(conn, dst_off, num_sges, wr);
 422                 if (rc)
 423                         return rc;
 424                 if (dst_len_sum == len)
 425                         break; /* either on 1st or 2nd iteration */
 426                 /* prepare next (== 2nd) iteration */
 427                 dst_off = 0; /* modulo offset in RMBE ring buffer */
 428                 dst_len = len - dst_len; /* remainder */
 429                 dst_len_sum += dst_len;
 430                 src_len = min_t(int, dst_len, conn->sndbuf_desc->len -
 431                                 sent_count);
 432                 src_len_sum = src_len;
 433         }
 434         return 0;
 435 }
 436
 437 /* SMC-D helper for smc_tx_rdma_writes() */
 438 static int smcd_tx_rdma_writes(struct smc_connection *conn, size_t len,
 439                                size_t src_off, size_t src_len,
 440                                size_t dst_off, size_t dst_len)
 441 {
 442         int src_len_sum = src_len, dst_len_sum = dst_len;
 443         int srcchunk, dstchunk;
 444         int rc;
 445
 446         for (dstchunk = 0; dstchunk < 2; dstchunk++) {
 447                 for (srcchunk = 0; srcchunk < 2; srcchunk++) {
 448                         void *data = conn->sndbuf_desc->cpu_addr + src_off;
 449
 450                         rc = smcd_tx_ism_write(conn, data, src_len, dst_off +
 451                                                sizeof(struct smcd_cdc_msg), 0);
 452                         if (rc)
 453                                 return rc;
 454                         dst_off += src_len;
 455                         src_off += src_len;
 456                         if (src_off >= conn->sndbuf_desc->len)
 457                                 src_off -= conn->sndbuf_desc->len;
 458                                                 /* modulo in send ring */
 459                         if (src_len_sum == dst_len)
 460                                 break; /* either on 1st or 2nd iteration */
 461                         /* prepare next (== 2nd) iteration */
 462                         src_len = dst_len - src_len; /* remainder */
 463                         src_len_sum += src_len;
 464                 }
 465                 if (dst_len_sum == len)
 466                         break; /* either on 1st or 2nd iteration */
 467                 /* prepare next (== 2nd) iteration */
 468                 dst_off = 0; /* modulo offset in RMBE ring buffer */
 469                 dst_len = len - dst_len; /* remainder */
 470                 dst_len_sum += dst_len;
 471                 src_len = min_t(int, dst_len, conn->sndbuf_desc->len - src_off);
 472                 src_len_sum = src_len;
 473         }
 474         return 0;
 475 }
 476
 477 /* sndbuf consumer: prepare all necessary (src&dst) chunks of data transmit;
 478  * usable snd_wnd as max transmit
 479  */
 480 static int smc_tx_rdma_writes(struct smc_connection *conn,
 481                               struct smc_rdma_wr *wr_rdma_buf)
 482 {
 483         size_t len, src_len, dst_off, dst_len; /* current chunk values */
 484         union smc_host_cursor sent, prep, prod, cons;
 485         struct smc_cdc_producer_flags *pflags;
 486         int to_send, rmbespace;
 487         int rc;
 488
 489         /* source: sndbuf */
 490         smc_curs_copy(&sent, &conn->tx_curs_sent, conn);
 491         smc_curs_copy(&prep, &conn->tx_curs_prep, conn);
 492         /* cf. wmem_alloc - (snd_max - snd_una) */
 493         to_send = smc_curs_diff(conn->sndbuf_desc->len, &sent, &prep);
 494         if (to_send <= 0)
 495                 return 0;
 496
 497         /* destination: RMBE */
 498         /* cf. snd_wnd */
 499         rmbespace = atomic_read(&conn->peer_rmbe_space);
 500         if (rmbespace <= 0) {
 501                 struct smc_sock *smc = container_of(conn, struct smc_sock,
 502                                                     conn);
 503                 SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
 504                 return 0;
 505         }
 506         smc_curs_copy(&prod, &conn->local_tx_ctrl.prod, conn);
 507         smc_curs_copy(&cons, &conn->local_rx_ctrl.cons, conn);
 508
 509         /* if usable snd_wnd closes ask peer to advertise once it opens again */
 510         pflags = &conn->local_tx_ctrl.prod_flags;
 511         pflags->write_blocked = (to_send >= rmbespace);
 512         /* cf. usable snd_wnd */
 513         len = min(to_send, rmbespace);
 514
 515         /* initialize variables for first iteration of subsequent nested loop */
 516         dst_off = prod.count;
 517         if (prod.wrap == cons.wrap) {
 518                 /* the filled destination area is unwrapped,
 519                  * hence the available free destination space is wrapped
 520                  * and we need 2 destination chunks of sum len; start with 1st
 521                  * which is limited by what's available in sndbuf
 522                  */
 523                 dst_len = min_t(size_t,
 524                                 conn->peer_rmbe_size - prod.count, len);
 525         } else {
 526                 /* the filled destination area is wrapped,
 527                  * hence the available free destination space is unwrapped
 528                  * and we need a single destination chunk of entire len
 529                  */
 530                 dst_len = len;
 531         }
 532         /* dst_len determines the maximum src_len */
 533         if (sent.count + dst_len <= conn->sndbuf_desc->len) {
 534                 /* unwrapped src case: single chunk of entire dst_len */
 535                 src_len = dst_len;
 536         } else {
 537                 /* wrapped src case: 2 chunks of sum dst_len; start with 1st: */
 538                 src_len = conn->sndbuf_desc->len - sent.count;
 539         }
 540
 541         if (conn->lgr->is_smcd)
 542                 rc = smcd_tx_rdma_writes(conn, len, sent.count, src_len,
 543                                          dst_off, dst_len);
 544         else
 545                 rc = smcr_tx_rdma_writes(conn, len, sent.count, src_len,
 546                                          dst_off, dst_len, wr_rdma_buf);
 547         if (rc)
 548                 return rc;
 549
 550         if (conn->urg_tx_pend && len == to_send)
 551                 pflags->urg_data_present = 1;
 552         smc_tx_advance_cursors(conn, &prod, &sent, len);
 553         /* update connection's cursors with advanced local cursors */
 554         smc_curs_copy(&conn->local_tx_ctrl.prod, &prod, conn);
 555                                                         /* dst: peer RMBE */
 556         smc_curs_copy(&conn->tx_curs_sent, &sent, conn);/* src: local sndbuf */
 557
 558         return 0;
 559 }
 560
 561 /* Wakeup sndbuf consumers from any context (IRQ or process)
 562  * since there is more data to transmit; usable snd_wnd as max transmit
 563  */
 564 static int smcr_tx_sndbuf_nonempty(struct smc_connection *conn)
 565 {
 566         struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
 567         struct smc_link *link = conn->lnk;
 568         struct smc_rdma_wr *wr_rdma_buf;
 569         struct smc_cdc_tx_pend *pend;
 570         struct smc_wr_buf *wr_buf;
 571         int rc;
 572
 573         if (!link || !smc_wr_tx_link_hold(link))
 574                 return -ENOLINK;
 575         rc = smc_cdc_get_free_slot(conn, link, &wr_buf, &wr_rdma_buf, &pend);
 576         if (rc < 0) {
 577                 smc_wr_tx_link_put(link);
 578                 if (rc == -EBUSY) {
 579                         struct smc_sock *smc =
 580                                 container_of(conn, struct smc_sock, conn);
 581
 582                         if (smc->sk.sk_err == ECONNABORTED)
 583                                 return sock_error(&smc->sk);
 584                         if (conn->killed)
 585                                 return -EPIPE;
 586                         rc = 0;
 587                         mod_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
 588                                          SMC_TX_WORK_DELAY);
 589                 }
 590                 return rc;
 591         }
 592
 593         spin_lock_bh(&conn->send_lock);
 594         if (link != conn->lnk) {
 595                 /* link of connection changed, tx_work will restart */
 596                 smc_wr_tx_put_slot(link,
 597                                    (struct smc_wr_tx_pend_priv *)pend);
 598                 rc = -ENOLINK;
 599                 goto out_unlock;
 600         }
 601         if (!pflags->urg_data_present) {
 602                 rc = smc_tx_rdma_writes(conn, wr_rdma_buf);
 603                 if (rc) {
 604                         smc_wr_tx_put_slot(link,
 605                                            (struct smc_wr_tx_pend_priv *)pend);
 606                         goto out_unlock;
 607                 }
 608         }
 609
 610         rc = smc_cdc_msg_send(conn, wr_buf, pend);
 611         if (!rc && pflags->urg_data_present) {
 612                 pflags->urg_data_pending = 0;
 613                 pflags->urg_data_present = 0;
 614         }
 615
 616 out_unlock:
 617         spin_unlock_bh(&conn->send_lock);
 618         smc_wr_tx_link_put(link);
 619         return rc;
 620 }
 621
 622 static int smcd_tx_sndbuf_nonempty(struct smc_connection *conn)
 623 {
 624         struct smc_cdc_producer_flags *pflags = &conn->local_tx_ctrl.prod_flags;
 625         int rc = 0;
 626
 627         spin_lock_bh(&conn->send_lock);
 628         if (!pflags->urg_data_present)
 629                 rc = smc_tx_rdma_writes(conn, NULL);
 630         if (!rc)
 631                 rc = smcd_cdc_msg_send(conn);
 632
 633         if (!rc && pflags->urg_data_present) {
 634                 pflags->urg_data_pending = 0;
 635                 pflags->urg_data_present = 0;
 636         }
 637         spin_unlock_bh(&conn->send_lock);
 638         return rc;
 639 }
 640
 641 static int __smc_tx_sndbuf_nonempty(struct smc_connection *conn)
 642 {
 643         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 644         int rc = 0;
 645
 646         /* No data in the send queue */
 647         if (unlikely(smc_tx_prepared_sends(conn) <= 0))
 648                 goto out;
 649
 650         /* Peer don't have RMBE space */
 651         if (unlikely(atomic_read(&conn->peer_rmbe_space) <= 0)) {
 652                 SMC_STAT_RMB_TX_PEER_FULL(smc, !conn->lnk);
 653                 goto out;
 654         }
 655
 656         if (conn->killed ||
 657             conn->local_rx_ctrl.conn_state_flags.peer_conn_abort) {
 658                 rc = -EPIPE;    /* connection being aborted */
 659                 goto out;
 660         }
 661         if (conn->lgr->is_smcd)
 662                 rc = smcd_tx_sndbuf_nonempty(conn);
 663         else
 664                 rc = smcr_tx_sndbuf_nonempty(conn);
 665
 666         if (!rc) {
 667                 /* trigger socket release if connection is closing */
 668                 smc_close_wake_tx_prepared(smc);
 669         }
 670
 671 out:
 672         return rc;
 673 }
 674
 675 int smc_tx_sndbuf_nonempty(struct smc_connection *conn)
 676 {
 677         int rc;
 678
 679         /* This make sure only one can send simultaneously to prevent wasting
 680          * of CPU and CDC slot.
 681          * Record whether someone has tried to push while we are pushing.
 682          */
 683         if (atomic_inc_return(&conn->tx_pushing) > 1)
 684                 return 0;
 685
 686 again:
 687         atomic_set(&conn->tx_pushing, 1);
 688         smp_wmb(); /* Make sure tx_pushing is 1 before real send */
 689         rc = __smc_tx_sndbuf_nonempty(conn);
 690
 691         /* We need to check whether someone else have added some data into
 692          * the send queue and tried to push but failed after the atomic_set()
 693          * when we are pushing.
 694          * If so, we need to push again to prevent those data hang in the send
 695          * queue.
 696          */
 697         if (unlikely(!atomic_dec_and_test(&conn->tx_pushing)))
 698                 goto again;
 699
 700         return rc;
 701 }
 702
 703 /* Wakeup sndbuf consumers from process context
 704  * since there is more data to transmit. The caller
 705  * must hold sock lock.
 706  */
 707 void smc_tx_pending(struct smc_connection *conn)
 708 {
 709         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 710         int rc;
 711
 712         if (smc->sk.sk_err)
 713                 return;
 714
 715         rc = smc_tx_sndbuf_nonempty(conn);
 716         if (!rc && conn->local_rx_ctrl.prod_flags.write_blocked &&
 717             !atomic_read(&conn->bytes_to_rcv))
 718                 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
 719 }
 720
 721 /* Wakeup sndbuf consumers from process context
 722  * since there is more data to transmit in locked
 723  * sock.
 724  */
 725 void smc_tx_work(struct work_struct *work)
 726 {
 727         struct smc_connection *conn = container_of(to_delayed_work(work),
 728                                                    struct smc_connection,
 729                                                    tx_work);
 730         struct smc_sock *smc = container_of(conn, struct smc_sock, conn);
 731
 732         lock_sock(&smc->sk);
 733         smc_tx_pending(conn);
 734         release_sock(&smc->sk);
 735 }
 736
 737 void smc_tx_consumer_update(struct smc_connection *conn, bool force)
 738 {
 739         union smc_host_cursor cfed, cons, prod;
 740         int sender_free = conn->rmb_desc->len;
 741         int to_confirm;
 742
 743         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
 744         smc_curs_copy(&cfed, &conn->rx_curs_confirmed, conn);
 745         to_confirm = smc_curs_diff(conn->rmb_desc->len, &cfed, &cons);
 746         if (to_confirm > conn->rmbe_update_limit) {
 747                 smc_curs_copy(&prod, &conn->local_rx_ctrl.prod, conn);
 748                 sender_free = conn->rmb_desc->len -
 749                               smc_curs_diff_large(conn->rmb_desc->len,
 750                                                   &cfed, &prod);
 751         }
 752
 753         if (conn->local_rx_ctrl.prod_flags.cons_curs_upd_req ||
 754             force ||
 755             ((to_confirm > conn->rmbe_update_limit) &&
 756              ((sender_free <= (conn->rmb_desc->len / 2)) ||
 757               conn->local_rx_ctrl.prod_flags.write_blocked))) {
 758                 if (conn->killed ||
 759                     conn->local_rx_ctrl.conn_state_flags.peer_conn_abort)
 760                         return;
 761                 if ((smc_cdc_get_slot_and_msg_send(conn) < 0) &&
 762                     !conn->killed) {
 763                         queue_delayed_work(conn->lgr->tx_wq, &conn->tx_work,
 764                                            SMC_TX_WORK_DELAY);
 765                         return;
 766                 }
 767         }
 768         if (conn->local_rx_ctrl.prod_flags.write_blocked &&
 769             !atomic_read(&conn->bytes_to_rcv))
 770                 conn->local_rx_ctrl.prod_flags.write_blocked = 0;
 771 }
 772
 773 /***************************** send initialize *******************************/
 774
 775 /* Initialize send properties on connection establishment. NB: not __init! */
 776 void smc_tx_init(struct smc_sock *smc)
 777 {
 778         smc->sk.sk_write_space = smc_tx_write_space;
 779 }