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