drivers/staging/lustre/lnet/klnds/socklnd/socklnd_lib.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2011, 2012, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36
  37 #include "socklnd.h"
  38
  39 int
  40 ksocknal_lib_get_conn_addrs(ksock_conn_t *conn)
  41 {
  42         int rc = lnet_sock_getaddr(conn->ksnc_sock, 1, &conn->ksnc_ipaddr,
  43                                    &conn->ksnc_port);
  44
  45         /* Didn't need the {get,put}connsock dance to deref ksnc_sock... */
  46         LASSERT(!conn->ksnc_closing);
  47
  48         if (rc) {
  49                 CERROR("Error %d getting sock peer IP\n", rc);
  50                 return rc;
  51         }
  52
  53         rc = lnet_sock_getaddr(conn->ksnc_sock, 0, &conn->ksnc_myipaddr, NULL);
  54         if (rc) {
  55                 CERROR("Error %d getting sock local IP\n", rc);
  56                 return rc;
  57         }
  58
  59         return 0;
  60 }
  61
  62 int
  63 ksocknal_lib_zc_capable(ksock_conn_t *conn)
  64 {
  65         int caps = conn->ksnc_sock->sk->sk_route_caps;
  66
  67         if (conn->ksnc_proto == &ksocknal_protocol_v1x)
  68                 return 0;
  69
  70         /*
  71          * ZC if the socket supports scatter/gather and doesn't need software
  72          * checksums
  73          */
  74         return ((caps & NETIF_F_SG) && (caps & NETIF_F_CSUM_MASK));
  75 }
  76
  77 int
  78 ksocknal_lib_send_iov(ksock_conn_t *conn, ksock_tx_t *tx)
  79 {
  80         struct socket *sock = conn->ksnc_sock;
  81         int nob;
  82         int rc;
  83
  84         if (*ksocknal_tunables.ksnd_enable_csum && /* checksum enabled */
  85             conn->ksnc_proto == &ksocknal_protocol_v2x && /* V2.x connection  */
  86             tx->tx_nob == tx->tx_resid           && /* frist sending    */
  87             !tx->tx_msg.ksm_csum)                    /* not checksummed  */
  88                 ksocknal_lib_csum_tx(tx);
  89
  90         /*
  91          * NB we can't trust socket ops to either consume our iovs
  92          * or leave them alone.
  93          */
  94         {
  95 #if SOCKNAL_SINGLE_FRAG_TX
  96                 struct kvec scratch;
  97                 struct kvec *scratchiov = &scratch;
  98                 unsigned int niov = 1;
  99 #else
 100                 struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
 101                 unsigned int niov = tx->tx_niov;
 102 #endif
 103                 struct msghdr msg = {.msg_flags = MSG_DONTWAIT};
 104                 int i;
 105
 106                 for (nob = i = 0; i < niov; i++) {
 107                         scratchiov[i] = tx->tx_iov[i];
 108                         nob += scratchiov[i].iov_len;
 109                 }
 110
 111                 if (!list_empty(&conn->ksnc_tx_queue) ||
 112                     nob < tx->tx_resid)
 113                         msg.msg_flags |= MSG_MORE;
 114
 115                 rc = kernel_sendmsg(sock, &msg, scratchiov, niov, nob);
 116         }
 117         return rc;
 118 }
 119
 120 int
 121 ksocknal_lib_send_kiov(ksock_conn_t *conn, ksock_tx_t *tx)
 122 {
 123         struct socket *sock = conn->ksnc_sock;
 124         lnet_kiov_t *kiov = tx->tx_kiov;
 125         int rc;
 126         int nob;
 127
 128         /* Not NOOP message */
 129         LASSERT(tx->tx_lnetmsg);
 130
 131         /*
 132          * NB we can't trust socket ops to either consume our iovs
 133          * or leave them alone.
 134          */
 135         if (tx->tx_msg.ksm_zc_cookies[0]) {
 136                 /* Zero copy is enabled */
 137                 struct sock *sk = sock->sk;
 138                 struct page *page = kiov->kiov_page;
 139                 int offset = kiov->kiov_offset;
 140                 int fragsize = kiov->kiov_len;
 141                 int msgflg = MSG_DONTWAIT;
 142
 143                 CDEBUG(D_NET, "page %p + offset %x for %d\n",
 144                        page, offset, kiov->kiov_len);
 145
 146                 if (!list_empty(&conn->ksnc_tx_queue) ||
 147                     fragsize < tx->tx_resid)
 148                         msgflg |= MSG_MORE;
 149
 150                 if (sk->sk_prot->sendpage) {
 151                         rc = sk->sk_prot->sendpage(sk, page,
 152                                                    offset, fragsize, msgflg);
 153                 } else {
 154                         rc = tcp_sendpage(sk, page, offset, fragsize, msgflg);
 155                 }
 156         } else {
 157 #if SOCKNAL_SINGLE_FRAG_TX || !SOCKNAL_RISK_KMAP_DEADLOCK
 158                 struct kvec scratch;
 159                 struct kvec *scratchiov = &scratch;
 160                 unsigned int niov = 1;
 161 #else
 162 #ifdef CONFIG_HIGHMEM
 163 #warning "XXX risk of kmap deadlock on multiple frags..."
 164 #endif
 165                 struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
 166                 unsigned int niov = tx->tx_nkiov;
 167 #endif
 168                 struct msghdr msg = {.msg_flags = MSG_DONTWAIT};
 169                 int i;
 170
 171                 for (nob = i = 0; i < niov; i++) {
 172                         scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
 173                                                  kiov[i].kiov_offset;
 174                         nob += scratchiov[i].iov_len = kiov[i].kiov_len;
 175                 }
 176
 177                 if (!list_empty(&conn->ksnc_tx_queue) ||
 178                     nob < tx->tx_resid)
 179                         msg.msg_flags |= MSG_MORE;
 180
 181                 rc = kernel_sendmsg(sock, &msg, (struct kvec *)scratchiov, niov, nob);
 182
 183                 for (i = 0; i < niov; i++)
 184                         kunmap(kiov[i].kiov_page);
 185         }
 186         return rc;
 187 }
 188
 189 void
 190 ksocknal_lib_eager_ack(ksock_conn_t *conn)
 191 {
 192         int opt = 1;
 193         struct socket *sock = conn->ksnc_sock;
 194
 195         /*
 196          * Remind the socket to ACK eagerly.  If I don't, the socket might
 197          * think I'm about to send something it could piggy-back the ACK
 198          * on, introducing delay in completing zero-copy sends in my
 199          * peer.
 200          */
 201         kernel_setsockopt(sock, SOL_TCP, TCP_QUICKACK, (char *)&opt,
 202                           sizeof(opt));
 203 }
 204
 205 int
 206 ksocknal_lib_recv_iov(ksock_conn_t *conn)
 207 {
 208 #if SOCKNAL_SINGLE_FRAG_RX
 209         struct kvec scratch;
 210         struct kvec *scratchiov = &scratch;
 211         unsigned int niov = 1;
 212 #else
 213         struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
 214         unsigned int niov = conn->ksnc_rx_niov;
 215 #endif
 216         struct kvec *iov = conn->ksnc_rx_iov;
 217         struct msghdr msg = {
 218                 .msg_flags = 0
 219         };
 220         int nob;
 221         int i;
 222         int rc;
 223         int fragnob;
 224         int sum;
 225         __u32 saved_csum;
 226
 227         /*
 228          * NB we can't trust socket ops to either consume our iovs
 229          * or leave them alone.
 230          */
 231         LASSERT(niov > 0);
 232
 233         for (nob = i = 0; i < niov; i++) {
 234                 scratchiov[i] = iov[i];
 235                 nob += scratchiov[i].iov_len;
 236         }
 237         LASSERT(nob <= conn->ksnc_rx_nob_wanted);
 238
 239         rc = kernel_recvmsg(conn->ksnc_sock, &msg, scratchiov, niov, nob,
 240                             MSG_DONTWAIT);
 241
 242         saved_csum = 0;
 243         if (conn->ksnc_proto == &ksocknal_protocol_v2x) {
 244                 saved_csum = conn->ksnc_msg.ksm_csum;
 245                 conn->ksnc_msg.ksm_csum = 0;
 246         }
 247
 248         if (saved_csum) {
 249                 /* accumulate checksum */
 250                 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
 251                         LASSERT(i < niov);
 252
 253                         fragnob = iov[i].iov_len;
 254                         if (fragnob > sum)
 255                                 fragnob = sum;
 256
 257                         conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
 258                                                            iov[i].iov_base, fragnob);
 259                 }
 260                 conn->ksnc_msg.ksm_csum = saved_csum;
 261         }
 262
 263         return rc;
 264 }
 265
 266 static void
 267 ksocknal_lib_kiov_vunmap(void *addr)
 268 {
 269         if (!addr)
 270                 return;
 271
 272         vunmap(addr);
 273 }
 274
 275 static void *
 276 ksocknal_lib_kiov_vmap(lnet_kiov_t *kiov, int niov,
 277                        struct kvec *iov, struct page **pages)
 278 {
 279         void *addr;
 280         int nob;
 281         int i;
 282
 283         if (!*ksocknal_tunables.ksnd_zc_recv || !pages)
 284                 return NULL;
 285
 286         LASSERT(niov <= LNET_MAX_IOV);
 287
 288         if (niov < 2 ||
 289             niov < *ksocknal_tunables.ksnd_zc_recv_min_nfrags)
 290                 return NULL;
 291
 292         for (nob = i = 0; i < niov; i++) {
 293                 if ((kiov[i].kiov_offset && i > 0) ||
 294                     (kiov[i].kiov_offset + kiov[i].kiov_len != PAGE_SIZE && i < niov - 1))
 295                         return NULL;
 296
 297                 pages[i] = kiov[i].kiov_page;
 298                 nob += kiov[i].kiov_len;
 299         }
 300
 301         addr = vmap(pages, niov, VM_MAP, PAGE_KERNEL);
 302         if (!addr)
 303                 return NULL;
 304
 305         iov->iov_base = addr + kiov[0].kiov_offset;
 306         iov->iov_len = nob;
 307
 308         return addr;
 309 }
 310
 311 int
 312 ksocknal_lib_recv_kiov(ksock_conn_t *conn)
 313 {
 314 #if SOCKNAL_SINGLE_FRAG_RX || !SOCKNAL_RISK_KMAP_DEADLOCK
 315         struct kvec scratch;
 316         struct kvec *scratchiov = &scratch;
 317         struct page **pages = NULL;
 318         unsigned int niov = 1;
 319 #else
 320 #ifdef CONFIG_HIGHMEM
 321 #warning "XXX risk of kmap deadlock on multiple frags..."
 322 #endif
 323         struct kvec *scratchiov = conn->ksnc_scheduler->kss_scratch_iov;
 324         struct page **pages = conn->ksnc_scheduler->kss_rx_scratch_pgs;
 325         unsigned int niov = conn->ksnc_rx_nkiov;
 326 #endif
 327         lnet_kiov_t   *kiov = conn->ksnc_rx_kiov;
 328         struct msghdr msg = {
 329                 .msg_flags = 0
 330         };
 331         int nob;
 332         int i;
 333         int rc;
 334         void *base;
 335         void *addr;
 336         int sum;
 337         int fragnob;
 338         int n;
 339
 340         /*
 341          * NB we can't trust socket ops to either consume our iovs
 342          * or leave them alone.
 343          */
 344         addr = ksocknal_lib_kiov_vmap(kiov, niov, scratchiov, pages);
 345         if (addr) {
 346                 nob = scratchiov[0].iov_len;
 347                 n = 1;
 348
 349         } else {
 350                 for (nob = i = 0; i < niov; i++) {
 351                         nob += scratchiov[i].iov_len = kiov[i].kiov_len;
 352                         scratchiov[i].iov_base = kmap(kiov[i].kiov_page) +
 353                                                  kiov[i].kiov_offset;
 354                 }
 355                 n = niov;
 356         }
 357
 358         LASSERT(nob <= conn->ksnc_rx_nob_wanted);
 359
 360         rc = kernel_recvmsg(conn->ksnc_sock, &msg, (struct kvec *)scratchiov,
 361                             n, nob, MSG_DONTWAIT);
 362
 363         if (conn->ksnc_msg.ksm_csum) {
 364                 for (i = 0, sum = rc; sum > 0; i++, sum -= fragnob) {
 365                         LASSERT(i < niov);
 366
 367                         /*
 368                          * Dang! have to kmap again because I have nowhere to
 369                          * stash the mapped address.  But by doing it while the
 370                          * page is still mapped, the kernel just bumps the map
 371                          * count and returns me the address it stashed.
 372                          */
 373                         base = kmap(kiov[i].kiov_page) + kiov[i].kiov_offset;
 374                         fragnob = kiov[i].kiov_len;
 375                         if (fragnob > sum)
 376                                 fragnob = sum;
 377
 378                         conn->ksnc_rx_csum = ksocknal_csum(conn->ksnc_rx_csum,
 379                                                            base, fragnob);
 380
 381                         kunmap(kiov[i].kiov_page);
 382                 }
 383         }
 384
 385         if (addr) {
 386                 ksocknal_lib_kiov_vunmap(addr);
 387         } else {
 388                 for (i = 0; i < niov; i++)
 389                         kunmap(kiov[i].kiov_page);
 390         }
 391
 392         return rc;
 393 }
 394
 395 void
 396 ksocknal_lib_csum_tx(ksock_tx_t *tx)
 397 {
 398         int i;
 399         __u32 csum;
 400         void *base;
 401
 402         LASSERT(tx->tx_iov[0].iov_base == &tx->tx_msg);
 403         LASSERT(tx->tx_conn);
 404         LASSERT(tx->tx_conn->ksnc_proto == &ksocknal_protocol_v2x);
 405
 406         tx->tx_msg.ksm_csum = 0;
 407
 408         csum = ksocknal_csum(~0, tx->tx_iov[0].iov_base,
 409                              tx->tx_iov[0].iov_len);
 410
 411         if (tx->tx_kiov) {
 412                 for (i = 0; i < tx->tx_nkiov; i++) {
 413                         base = kmap(tx->tx_kiov[i].kiov_page) +
 414                                tx->tx_kiov[i].kiov_offset;
 415
 416                         csum = ksocknal_csum(csum, base, tx->tx_kiov[i].kiov_len);
 417
 418                         kunmap(tx->tx_kiov[i].kiov_page);
 419                 }
 420         } else {
 421                 for (i = 1; i < tx->tx_niov; i++)
 422                         csum = ksocknal_csum(csum, tx->tx_iov[i].iov_base,
 423                                              tx->tx_iov[i].iov_len);
 424         }
 425
 426         if (*ksocknal_tunables.ksnd_inject_csum_error) {
 427                 csum++;
 428                 *ksocknal_tunables.ksnd_inject_csum_error = 0;
 429         }
 430
 431         tx->tx_msg.ksm_csum = csum;
 432 }
 433
 434 int
 435 ksocknal_lib_get_conn_tunables(ksock_conn_t *conn, int *txmem, int *rxmem, int *nagle)
 436 {
 437         struct socket *sock = conn->ksnc_sock;
 438         int len;
 439         int rc;
 440
 441         rc = ksocknal_connsock_addref(conn);
 442         if (rc) {
 443                 LASSERT(conn->ksnc_closing);
 444                 *txmem = *rxmem = *nagle = 0;
 445                 return -ESHUTDOWN;
 446         }
 447
 448         rc = lnet_sock_getbuf(sock, txmem, rxmem);
 449         if (!rc) {
 450                 len = sizeof(*nagle);
 451                 rc = kernel_getsockopt(sock, SOL_TCP, TCP_NODELAY,
 452                                        (char *)nagle, &len);
 453         }
 454
 455         ksocknal_connsock_decref(conn);
 456
 457         if (!rc)
 458                 *nagle = !*nagle;
 459         else
 460                 *txmem = *rxmem = *nagle = 0;
 461
 462         return rc;
 463 }
 464
 465 int
 466 ksocknal_lib_setup_sock(struct socket *sock)
 467 {
 468         int rc;
 469         int option;
 470         int keep_idle;
 471         int keep_intvl;
 472         int keep_count;
 473         int do_keepalive;
 474         struct linger linger;
 475
 476         sock->sk->sk_allocation = GFP_NOFS;
 477
 478         /*
 479          * Ensure this socket aborts active sends immediately when we close
 480          * it.
 481          */
 482         linger.l_onoff = 0;
 483         linger.l_linger = 0;
 484
 485         rc = kernel_setsockopt(sock, SOL_SOCKET, SO_LINGER, (char *)&linger,
 486                                sizeof(linger));
 487         if (rc) {
 488                 CERROR("Can't set SO_LINGER: %d\n", rc);
 489                 return rc;
 490         }
 491
 492         option = -1;
 493         rc = kernel_setsockopt(sock, SOL_TCP, TCP_LINGER2, (char *)&option,
 494                                sizeof(option));
 495         if (rc) {
 496                 CERROR("Can't set SO_LINGER2: %d\n", rc);
 497                 return rc;
 498         }
 499
 500         if (!*ksocknal_tunables.ksnd_nagle) {
 501                 option = 1;
 502
 503                 rc = kernel_setsockopt(sock, SOL_TCP, TCP_NODELAY,
 504                                        (char *)&option, sizeof(option));
 505                 if (rc) {
 506                         CERROR("Can't disable nagle: %d\n", rc);
 507                         return rc;
 508                 }
 509         }
 510
 511         rc = lnet_sock_setbuf(sock, *ksocknal_tunables.ksnd_tx_buffer_size,
 512                               *ksocknal_tunables.ksnd_rx_buffer_size);
 513         if (rc) {
 514                 CERROR("Can't set buffer tx %d, rx %d buffers: %d\n",
 515                        *ksocknal_tunables.ksnd_tx_buffer_size,
 516                        *ksocknal_tunables.ksnd_rx_buffer_size, rc);
 517                 return rc;
 518         }
 519
 520 /* TCP_BACKOFF_* sockopt tunables unsupported in stock kernels */
 521
 522         /* snapshot tunables */
 523         keep_idle  = *ksocknal_tunables.ksnd_keepalive_idle;
 524         keep_count = *ksocknal_tunables.ksnd_keepalive_count;
 525         keep_intvl = *ksocknal_tunables.ksnd_keepalive_intvl;
 526
 527         do_keepalive = (keep_idle > 0 && keep_count > 0 && keep_intvl > 0);
 528
 529         option = (do_keepalive ? 1 : 0);
 530         rc = kernel_setsockopt(sock, SOL_SOCKET, SO_KEEPALIVE, (char *)&option,
 531                                sizeof(option));
 532         if (rc) {
 533                 CERROR("Can't set SO_KEEPALIVE: %d\n", rc);
 534                 return rc;
 535         }
 536
 537         if (!do_keepalive)
 538                 return 0;
 539
 540         rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPIDLE, (char *)&keep_idle,
 541                                sizeof(keep_idle));
 542         if (rc) {
 543                 CERROR("Can't set TCP_KEEPIDLE: %d\n", rc);
 544                 return rc;
 545         }
 546
 547         rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPINTVL,
 548                                (char *)&keep_intvl, sizeof(keep_intvl));
 549         if (rc) {
 550                 CERROR("Can't set TCP_KEEPINTVL: %d\n", rc);
 551                 return rc;
 552         }
 553
 554         rc = kernel_setsockopt(sock, SOL_TCP, TCP_KEEPCNT, (char *)&keep_count,
 555                                sizeof(keep_count));
 556         if (rc) {
 557                 CERROR("Can't set TCP_KEEPCNT: %d\n", rc);
 558                 return rc;
 559         }
 560
 561         return 0;
 562 }
 563
 564 void
 565 ksocknal_lib_push_conn(ksock_conn_t *conn)
 566 {
 567         struct sock *sk;
 568         struct tcp_sock *tp;
 569         int nonagle;
 570         int val = 1;
 571         int rc;
 572
 573         rc = ksocknal_connsock_addref(conn);
 574         if (rc)                     /* being shut down */
 575                 return;
 576
 577         sk = conn->ksnc_sock->sk;
 578         tp = tcp_sk(sk);
 579
 580         lock_sock(sk);
 581         nonagle = tp->nonagle;
 582         tp->nonagle = 1;
 583         release_sock(sk);
 584
 585         rc = kernel_setsockopt(conn->ksnc_sock, SOL_TCP, TCP_NODELAY,
 586                                (char *)&val, sizeof(val));
 587         LASSERT(!rc);
 588
 589         lock_sock(sk);
 590         tp->nonagle = nonagle;
 591         release_sock(sk);
 592
 593         ksocknal_connsock_decref(conn);
 594 }
 595
 596 /*
 597  * socket call back in Linux
 598  */
 599 static void
 600 ksocknal_data_ready(struct sock *sk)
 601 {
 602         ksock_conn_t *conn;
 603
 604         /* interleave correctly with closing sockets... */
 605         LASSERT(!in_irq());
 606         read_lock(&ksocknal_data.ksnd_global_lock);
 607
 608         conn = sk->sk_user_data;
 609         if (!conn) {         /* raced with ksocknal_terminate_conn */
 610                 LASSERT(sk->sk_data_ready != &ksocknal_data_ready);
 611                 sk->sk_data_ready(sk);
 612         } else {
 613                 ksocknal_read_callback(conn);
 614         }
 615
 616         read_unlock(&ksocknal_data.ksnd_global_lock);
 617 }
 618
 619 static void
 620 ksocknal_write_space(struct sock *sk)
 621 {
 622         ksock_conn_t *conn;
 623         int wspace;
 624         int min_wpace;
 625
 626         /* interleave correctly with closing sockets... */
 627         LASSERT(!in_irq());
 628         read_lock(&ksocknal_data.ksnd_global_lock);
 629
 630         conn = sk->sk_user_data;
 631         wspace = sk_stream_wspace(sk);
 632         min_wpace = sk_stream_min_wspace(sk);
 633
 634         CDEBUG(D_NET, "sk %p wspace %d low water %d conn %p%s%s%s\n",
 635                sk, wspace, min_wpace, conn,
 636                !conn ? "" : (conn->ksnc_tx_ready ?
 637                                       " ready" : " blocked"),
 638                !conn ? "" : (conn->ksnc_tx_scheduled ?
 639                                       " scheduled" : " idle"),
 640                !conn ? "" : (list_empty(&conn->ksnc_tx_queue) ?
 641                                       " empty" : " queued"));
 642
 643         if (!conn) {         /* raced with ksocknal_terminate_conn */
 644                 LASSERT(sk->sk_write_space != &ksocknal_write_space);
 645                 sk->sk_write_space(sk);
 646
 647                 read_unlock(&ksocknal_data.ksnd_global_lock);
 648                 return;
 649         }
 650
 651         if (wspace >= min_wpace) {            /* got enough space */
 652                 ksocknal_write_callback(conn);
 653
 654                 /*
 655                  * Clear SOCK_NOSPACE _after_ ksocknal_write_callback so the
 656                  * ENOMEM check in ksocknal_transmit is race-free (think about
 657                  * it).
 658                  */
 659                 clear_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
 660         }
 661
 662         read_unlock(&ksocknal_data.ksnd_global_lock);
 663 }
 664
 665 void
 666 ksocknal_lib_save_callback(struct socket *sock, ksock_conn_t *conn)
 667 {
 668         conn->ksnc_saved_data_ready = sock->sk->sk_data_ready;
 669         conn->ksnc_saved_write_space = sock->sk->sk_write_space;
 670 }
 671
 672 void
 673 ksocknal_lib_set_callback(struct socket *sock,  ksock_conn_t *conn)
 674 {
 675         sock->sk->sk_user_data = conn;
 676         sock->sk->sk_data_ready = ksocknal_data_ready;
 677         sock->sk->sk_write_space = ksocknal_write_space;
 678         return;
 679 }
 680
 681 void
 682 ksocknal_lib_reset_callback(struct socket *sock, ksock_conn_t *conn)
 683 {
 684         /*
 685          * Remove conn's network callbacks.
 686          * NB I _have_ to restore the callback, rather than storing a noop,
 687          * since the socket could survive past this module being unloaded!!
 688          */
 689         sock->sk->sk_data_ready = conn->ksnc_saved_data_ready;
 690         sock->sk->sk_write_space = conn->ksnc_saved_write_space;
 691
 692         /*
 693          * A callback could be in progress already; they hold a read lock
 694          * on ksnd_global_lock (to serialise with me) and NOOP if
 695          * sk_user_data is NULL.
 696          */
 697         sock->sk->sk_user_data = NULL;
 698
 699         return ;
 700 }
 701
 702 int
 703 ksocknal_lib_memory_pressure(ksock_conn_t *conn)
 704 {
 705         int rc = 0;
 706         ksock_sched_t *sched;
 707
 708         sched = conn->ksnc_scheduler;
 709         spin_lock_bh(&sched->kss_lock);
 710
 711         if (!test_bit(SOCK_NOSPACE, &conn->ksnc_sock->flags) &&
 712             !conn->ksnc_tx_ready) {
 713                 /*
 714                  * SOCK_NOSPACE is set when the socket fills
 715                  * and cleared in the write_space callback
 716                  * (which also sets ksnc_tx_ready).  If
 717                  * SOCK_NOSPACE and ksnc_tx_ready are BOTH
 718                  * zero, I didn't fill the socket and
 719                  * write_space won't reschedule me, so I
 720                  * return -ENOMEM to get my caller to retry
 721                  * after a timeout
 722                  */
 723                 rc = -ENOMEM;
 724         }
 725
 726         spin_unlock_bh(&sched->kss_lock);
 727
 728         return rc;
 729 }