net/sunrpc/xprtsock.c

   1 /*
   2  * linux/net/sunrpc/xprtsock.c
   3  *
   4  * Client-side transport implementation for sockets.
   5  *
   6  * TCP callback races fixes (C) 1998 Red Hat Software <[email protected]>
   7  * TCP send fixes (C) 1998 Red Hat Software <[email protected]>
   8  * TCP NFS related read + write fixes
   9  *  (C) 1999 Dave Airlie, University of Limerick, Ireland <[email protected]>
  10  *
  11  * Rewrite of larges part of the code in order to stabilize TCP stuff.
  12  * Fix behaviour when socket buffer is full.
  13  *  (C) 1999 Trond Myklebust <[email protected]>
  14  *
  15  * IP socket transport implementation, (C) 2005 Chuck Lever <[email protected]>
  16  */
  17
  18 #include <linux/types.h>
  19 #include <linux/slab.h>
  20 #include <linux/capability.h>
  21 #include <linux/sched.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/errno.h>
  24 #include <linux/socket.h>
  25 #include <linux/in.h>
  26 #include <linux/net.h>
  27 #include <linux/mm.h>
  28 #include <linux/udp.h>
  29 #include <linux/tcp.h>
  30 #include <linux/sunrpc/clnt.h>
  31 #include <linux/sunrpc/sched.h>
  32 #include <linux/file.h>
  33
  34 #include <net/sock.h>
  35 #include <net/checksum.h>
  36 #include <net/udp.h>
  37 #include <net/tcp.h>
  38
  39 /*
  40  * xprtsock tunables
  41  */
  42 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
  43 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
  44
  45 unsigned int xprt_min_resvport = RPC_DEF_MIN_RESVPORT;
  46 unsigned int xprt_max_resvport = RPC_DEF_MAX_RESVPORT;
  47
  48 /*
  49  * How many times to try sending a request on a socket before waiting
  50  * for the socket buffer to clear.
  51  */
  52 #define XS_SENDMSG_RETRY        (10U)
  53
  54 /*
  55  * Time out for an RPC UDP socket connect.  UDP socket connects are
  56  * synchronous, but we set a timeout anyway in case of resource
  57  * exhaustion on the local host.
  58  */
  59 #define XS_UDP_CONN_TO          (5U * HZ)
  60
  61 /*
  62  * Wait duration for an RPC TCP connection to be established.  Solaris
  63  * NFS over TCP uses 60 seconds, for example, which is in line with how
  64  * long a server takes to reboot.
  65  */
  66 #define XS_TCP_CONN_TO          (60U * HZ)
  67
  68 /*
  69  * Wait duration for a reply from the RPC portmapper.
  70  */
  71 #define XS_BIND_TO              (60U * HZ)
  72
  73 /*
  74  * Delay if a UDP socket connect error occurs.  This is most likely some
  75  * kind of resource problem on the local host.
  76  */
  77 #define XS_UDP_REEST_TO         (2U * HZ)
  78
  79 /*
  80  * The reestablish timeout allows clients to delay for a bit before attempting
  81  * to reconnect to a server that just dropped our connection.
  82  *
  83  * We implement an exponential backoff when trying to reestablish a TCP
  84  * transport connection with the server.  Some servers like to drop a TCP
  85  * connection when they are overworked, so we start with a short timeout and
  86  * increase over time if the server is down or not responding.
  87  */
  88 #define XS_TCP_INIT_REEST_TO    (3U * HZ)
  89 #define XS_TCP_MAX_REEST_TO     (5U * 60 * HZ)
  90
  91 /*
  92  * TCP idle timeout; client drops the transport socket if it is idle
  93  * for this long.  Note that we also timeout UDP sockets to prevent
  94  * holding port numbers when there is no RPC traffic.
  95  */
  96 #define XS_IDLE_DISC_TO         (5U * 60 * HZ)
  97
  98 #ifdef RPC_DEBUG
  99 # undef  RPC_DEBUG_DATA
 100 # define RPCDBG_FACILITY        RPCDBG_TRANS
 101 #endif
 102
 103 #ifdef RPC_DEBUG_DATA
 104 static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
 105 {
 106         u8 *buf = (u8 *) packet;
 107         int j;
 108
 109         dprintk("RPC:      %s\n", msg);
 110         for (j = 0; j < count && j < 128; j += 4) {
 111                 if (!(j & 31)) {
 112                         if (j)
 113                                 dprintk("\n");
 114                         dprintk("0x%04x ", j);
 115                 }
 116                 dprintk("%02x%02x%02x%02x ",
 117                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
 118         }
 119         dprintk("\n");
 120 }
 121 #else
 122 static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
 123 {
 124         /* NOP */
 125 }
 126 #endif
 127
 128 struct sock_xprt {
 129         struct rpc_xprt         xprt;
 130
 131         /*
 132          * Network layer
 133          */
 134         struct socket *         sock;
 135         struct sock *           inet;
 136
 137         /*
 138          * State of TCP reply receive
 139          */
 140         __be32                  tcp_fraghdr,
 141                                 tcp_xid;
 142
 143         u32                     tcp_offset,
 144                                 tcp_reclen;
 145
 146         unsigned long           tcp_copied,
 147                                 tcp_flags;
 148 };
 149
 150 static void xs_format_peer_addresses(struct rpc_xprt *xprt)
 151 {
 152         struct sockaddr_in *addr = (struct sockaddr_in *) &xprt->addr;
 153         char *buf;
 154
 155         buf = kzalloc(20, GFP_KERNEL);
 156         if (buf) {
 157                 snprintf(buf, 20, "%u.%u.%u.%u",
 158                                 NIPQUAD(addr->sin_addr.s_addr));
 159         }
 160         xprt->address_strings[RPC_DISPLAY_ADDR] = buf;
 161
 162         buf = kzalloc(8, GFP_KERNEL);
 163         if (buf) {
 164                 snprintf(buf, 8, "%u",
 165                                 ntohs(addr->sin_port));
 166         }
 167         xprt->address_strings[RPC_DISPLAY_PORT] = buf;
 168
 169         if (xprt->prot == IPPROTO_UDP)
 170                 xprt->address_strings[RPC_DISPLAY_PROTO] = "udp";
 171         else
 172                 xprt->address_strings[RPC_DISPLAY_PROTO] = "tcp";
 173
 174         buf = kzalloc(48, GFP_KERNEL);
 175         if (buf) {
 176                 snprintf(buf, 48, "addr=%u.%u.%u.%u port=%u proto=%s",
 177                         NIPQUAD(addr->sin_addr.s_addr),
 178                         ntohs(addr->sin_port),
 179                         xprt->prot == IPPROTO_UDP ? "udp" : "tcp");
 180         }
 181         xprt->address_strings[RPC_DISPLAY_ALL] = buf;
 182 }
 183
 184 static void xs_free_peer_addresses(struct rpc_xprt *xprt)
 185 {
 186         kfree(xprt->address_strings[RPC_DISPLAY_ADDR]);
 187         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
 188         kfree(xprt->address_strings[RPC_DISPLAY_ALL]);
 189 }
 190
 191 #define XS_SENDMSG_FLAGS        (MSG_DONTWAIT | MSG_NOSIGNAL)
 192
 193 static int xs_send_kvec(struct socket *sock, struct sockaddr *addr, int addrlen, struct kvec *vec, unsigned int base, int more)
 194 {
 195         struct msghdr msg = {
 196                 .msg_name       = addr,
 197                 .msg_namelen    = addrlen,
 198                 .msg_flags      = XS_SENDMSG_FLAGS | (more ? MSG_MORE : 0),
 199         };
 200         struct kvec iov = {
 201                 .iov_base       = vec->iov_base + base,
 202                 .iov_len        = vec->iov_len - base,
 203         };
 204
 205         if (iov.iov_len != 0)
 206                 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
 207         return kernel_sendmsg(sock, &msg, NULL, 0, 0);
 208 }
 209
 210 static int xs_send_pagedata(struct socket *sock, struct xdr_buf *xdr, unsigned int base, int more)
 211 {
 212         struct page **ppage;
 213         unsigned int remainder;
 214         int err, sent = 0;
 215
 216         remainder = xdr->page_len - base;
 217         base += xdr->page_base;
 218         ppage = xdr->pages + (base >> PAGE_SHIFT);
 219         base &= ~PAGE_MASK;
 220         for(;;) {
 221                 unsigned int len = min_t(unsigned int, PAGE_SIZE - base, remainder);
 222                 int flags = XS_SENDMSG_FLAGS;
 223
 224                 remainder -= len;
 225                 if (remainder != 0 || more)
 226                         flags |= MSG_MORE;
 227                 err = sock->ops->sendpage(sock, *ppage, base, len, flags);
 228                 if (remainder == 0 || err != len)
 229                         break;
 230                 sent += err;
 231                 ppage++;
 232                 base = 0;
 233         }
 234         if (sent == 0)
 235                 return err;
 236         if (err > 0)
 237                 sent += err;
 238         return sent;
 239 }
 240
 241 /**
 242  * xs_sendpages - write pages directly to a socket
 243  * @sock: socket to send on
 244  * @addr: UDP only -- address of destination
 245  * @addrlen: UDP only -- length of destination address
 246  * @xdr: buffer containing this request
 247  * @base: starting position in the buffer
 248  *
 249  */
 250 static int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
 251 {
 252         unsigned int remainder = xdr->len - base;
 253         int err, sent = 0;
 254
 255         if (unlikely(!sock))
 256                 return -ENOTCONN;
 257
 258         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 259         if (base != 0) {
 260                 addr = NULL;
 261                 addrlen = 0;
 262         }
 263
 264         if (base < xdr->head[0].iov_len || addr != NULL) {
 265                 unsigned int len = xdr->head[0].iov_len - base;
 266                 remainder -= len;
 267                 err = xs_send_kvec(sock, addr, addrlen, &xdr->head[0], base, remainder != 0);
 268                 if (remainder == 0 || err != len)
 269                         goto out;
 270                 sent += err;
 271                 base = 0;
 272         } else
 273                 base -= xdr->head[0].iov_len;
 274
 275         if (base < xdr->page_len) {
 276                 unsigned int len = xdr->page_len - base;
 277                 remainder -= len;
 278                 err = xs_send_pagedata(sock, xdr, base, remainder != 0);
 279                 if (remainder == 0 || err != len)
 280                         goto out;
 281                 sent += err;
 282                 base = 0;
 283         } else
 284                 base -= xdr->page_len;
 285
 286         if (base >= xdr->tail[0].iov_len)
 287                 return sent;
 288         err = xs_send_kvec(sock, NULL, 0, &xdr->tail[0], base, 0);
 289 out:
 290         if (sent == 0)
 291                 return err;
 292         if (err > 0)
 293                 sent += err;
 294         return sent;
 295 }
 296
 297 /**
 298  * xs_nospace - place task on wait queue if transmit was incomplete
 299  * @task: task to put to sleep
 300  *
 301  */
 302 static void xs_nospace(struct rpc_task *task)
 303 {
 304         struct rpc_rqst *req = task->tk_rqstp;
 305         struct rpc_xprt *xprt = req->rq_xprt;
 306         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 307
 308         dprintk("RPC: %4d xmit incomplete (%u left of %u)\n",
 309                         task->tk_pid, req->rq_slen - req->rq_bytes_sent,
 310                         req->rq_slen);
 311
 312         if (test_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags)) {
 313                 /* Protect against races with write_space */
 314                 spin_lock_bh(&xprt->transport_lock);
 315
 316                 /* Don't race with disconnect */
 317                 if (!xprt_connected(xprt))
 318                         task->tk_status = -ENOTCONN;
 319                 else if (test_bit(SOCK_NOSPACE, &transport->sock->flags))
 320                         xprt_wait_for_buffer_space(task);
 321
 322                 spin_unlock_bh(&xprt->transport_lock);
 323         } else
 324                 /* Keep holding the socket if it is blocked */
 325                 rpc_delay(task, HZ>>4);
 326 }
 327
 328 /**
 329  * xs_udp_send_request - write an RPC request to a UDP socket
 330  * @task: address of RPC task that manages the state of an RPC request
 331  *
 332  * Return values:
 333  *        0:    The request has been sent
 334  *   EAGAIN:    The socket was blocked, please call again later to
 335  *              complete the request
 336  * ENOTCONN:    Caller needs to invoke connect logic then call again
 337  *    other:    Some other error occured, the request was not sent
 338  */
 339 static int xs_udp_send_request(struct rpc_task *task)
 340 {
 341         struct rpc_rqst *req = task->tk_rqstp;
 342         struct rpc_xprt *xprt = req->rq_xprt;
 343         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 344         struct xdr_buf *xdr = &req->rq_snd_buf;
 345         int status;
 346
 347         xs_pktdump("packet data:",
 348                                 req->rq_svec->iov_base,
 349                                 req->rq_svec->iov_len);
 350
 351         req->rq_xtime = jiffies;
 352         status = xs_sendpages(transport->sock,
 353                               (struct sockaddr *) &xprt->addr,
 354                               xprt->addrlen, xdr,
 355                               req->rq_bytes_sent);
 356
 357         dprintk("RPC:      xs_udp_send_request(%u) = %d\n",
 358                         xdr->len - req->rq_bytes_sent, status);
 359
 360         if (likely(status >= (int) req->rq_slen))
 361                 return 0;
 362
 363         /* Still some bytes left; set up for a retry later. */
 364         if (status > 0)
 365                 status = -EAGAIN;
 366
 367         switch (status) {
 368         case -ENETUNREACH:
 369         case -EPIPE:
 370         case -ECONNREFUSED:
 371                 /* When the server has died, an ICMP port unreachable message
 372                  * prompts ECONNREFUSED. */
 373                 break;
 374         case -EAGAIN:
 375                 xs_nospace(task);
 376                 break;
 377         default:
 378                 dprintk("RPC:      sendmsg returned unrecognized error %d\n",
 379                         -status);
 380                 break;
 381         }
 382
 383         return status;
 384 }
 385
 386 static inline void xs_encode_tcp_record_marker(struct xdr_buf *buf)
 387 {
 388         u32 reclen = buf->len - sizeof(rpc_fraghdr);
 389         rpc_fraghdr *base = buf->head[0].iov_base;
 390         *base = htonl(RPC_LAST_STREAM_FRAGMENT | reclen);
 391 }
 392
 393 /**
 394  * xs_tcp_send_request - write an RPC request to a TCP socket
 395  * @task: address of RPC task that manages the state of an RPC request
 396  *
 397  * Return values:
 398  *        0:    The request has been sent
 399  *   EAGAIN:    The socket was blocked, please call again later to
 400  *              complete the request
 401  * ENOTCONN:    Caller needs to invoke connect logic then call again
 402  *    other:    Some other error occured, the request was not sent
 403  *
 404  * XXX: In the case of soft timeouts, should we eventually give up
 405  *      if sendmsg is not able to make progress?
 406  */
 407 static int xs_tcp_send_request(struct rpc_task *task)
 408 {
 409         struct rpc_rqst *req = task->tk_rqstp;
 410         struct rpc_xprt *xprt = req->rq_xprt;
 411         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 412         struct xdr_buf *xdr = &req->rq_snd_buf;
 413         int status, retry = 0;
 414
 415         xs_encode_tcp_record_marker(&req->rq_snd_buf);
 416
 417         xs_pktdump("packet data:",
 418                                 req->rq_svec->iov_base,
 419                                 req->rq_svec->iov_len);
 420
 421         /* Continue transmitting the packet/record. We must be careful
 422          * to cope with writespace callbacks arriving _after_ we have
 423          * called sendmsg(). */
 424         while (1) {
 425                 req->rq_xtime = jiffies;
 426                 status = xs_sendpages(transport->sock,
 427                                         NULL, 0, xdr, req->rq_bytes_sent);
 428
 429                 dprintk("RPC:      xs_tcp_send_request(%u) = %d\n",
 430                                 xdr->len - req->rq_bytes_sent, status);
 431
 432                 if (unlikely(status < 0))
 433                         break;
 434
 435                 /* If we've sent the entire packet, immediately
 436                  * reset the count of bytes sent. */
 437                 req->rq_bytes_sent += status;
 438                 task->tk_bytes_sent += status;
 439                 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
 440                         req->rq_bytes_sent = 0;
 441                         return 0;
 442                 }
 443
 444                 status = -EAGAIN;
 445                 if (retry++ > XS_SENDMSG_RETRY)
 446                         break;
 447         }
 448
 449         switch (status) {
 450         case -EAGAIN:
 451                 xs_nospace(task);
 452                 break;
 453         case -ECONNREFUSED:
 454         case -ECONNRESET:
 455         case -ENOTCONN:
 456         case -EPIPE:
 457                 status = -ENOTCONN;
 458                 break;
 459         default:
 460                 dprintk("RPC:      sendmsg returned unrecognized error %d\n",
 461                         -status);
 462                 xprt_disconnect(xprt);
 463                 break;
 464         }
 465
 466         return status;
 467 }
 468
 469 /**
 470  * xs_tcp_release_xprt - clean up after a tcp transmission
 471  * @xprt: transport
 472  * @task: rpc task
 473  *
 474  * This cleans up if an error causes us to abort the transmission of a request.
 475  * In this case, the socket may need to be reset in order to avoid confusing
 476  * the server.
 477  */
 478 static void xs_tcp_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 479 {
 480         struct rpc_rqst *req;
 481
 482         if (task != xprt->snd_task)
 483                 return;
 484         if (task == NULL)
 485                 goto out_release;
 486         req = task->tk_rqstp;
 487         if (req->rq_bytes_sent == 0)
 488                 goto out_release;
 489         if (req->rq_bytes_sent == req->rq_snd_buf.len)
 490                 goto out_release;
 491         set_bit(XPRT_CLOSE_WAIT, &task->tk_xprt->state);
 492 out_release:
 493         xprt_release_xprt(xprt, task);
 494 }
 495
 496 /**
 497  * xs_close - close a socket
 498  * @xprt: transport
 499  *
 500  * This is used when all requests are complete; ie, no DRC state remains
 501  * on the server we want to save.
 502  */
 503 static void xs_close(struct rpc_xprt *xprt)
 504 {
 505         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 506         struct socket *sock = transport->sock;
 507         struct sock *sk = transport->inet;
 508
 509         if (!sk)
 510                 goto clear_close_wait;
 511
 512         dprintk("RPC:      xs_close xprt %p\n", xprt);
 513
 514         write_lock_bh(&sk->sk_callback_lock);
 515         transport->inet = NULL;
 516         transport->sock = NULL;
 517
 518         sk->sk_user_data = NULL;
 519         sk->sk_data_ready = xprt->old_data_ready;
 520         sk->sk_state_change = xprt->old_state_change;
 521         sk->sk_write_space = xprt->old_write_space;
 522         write_unlock_bh(&sk->sk_callback_lock);
 523
 524         sk->sk_no_check = 0;
 525
 526         sock_release(sock);
 527 clear_close_wait:
 528         smp_mb__before_clear_bit();
 529         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 530         smp_mb__after_clear_bit();
 531 }
 532
 533 /**
 534  * xs_destroy - prepare to shutdown a transport
 535  * @xprt: doomed transport
 536  *
 537  */
 538 static void xs_destroy(struct rpc_xprt *xprt)
 539 {
 540         dprintk("RPC:      xs_destroy xprt %p\n", xprt);
 541
 542         cancel_delayed_work(&xprt->connect_worker);
 543         flush_scheduled_work();
 544
 545         xprt_disconnect(xprt);
 546         xs_close(xprt);
 547         xs_free_peer_addresses(xprt);
 548         kfree(xprt->slot);
 549         kfree(xprt);
 550 }
 551
 552 static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
 553 {
 554         return (struct rpc_xprt *) sk->sk_user_data;
 555 }
 556
 557 /**
 558  * xs_udp_data_ready - "data ready" callback for UDP sockets
 559  * @sk: socket with data to read
 560  * @len: how much data to read
 561  *
 562  */
 563 static void xs_udp_data_ready(struct sock *sk, int len)
 564 {
 565         struct rpc_task *task;
 566         struct rpc_xprt *xprt;
 567         struct rpc_rqst *rovr;
 568         struct sk_buff *skb;
 569         int err, repsize, copied;
 570         u32 _xid;
 571         __be32 *xp;
 572
 573         read_lock(&sk->sk_callback_lock);
 574         dprintk("RPC:      xs_udp_data_ready...\n");
 575         if (!(xprt = xprt_from_sock(sk)))
 576                 goto out;
 577
 578         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 579                 goto out;
 580
 581         if (xprt->shutdown)
 582                 goto dropit;
 583
 584         repsize = skb->len - sizeof(struct udphdr);
 585         if (repsize < 4) {
 586                 dprintk("RPC:      impossible RPC reply size %d!\n", repsize);
 587                 goto dropit;
 588         }
 589
 590         /* Copy the XID from the skb... */
 591         xp = skb_header_pointer(skb, sizeof(struct udphdr),
 592                                 sizeof(_xid), &_xid);
 593         if (xp == NULL)
 594                 goto dropit;
 595
 596         /* Look up and lock the request corresponding to the given XID */
 597         spin_lock(&xprt->transport_lock);
 598         rovr = xprt_lookup_rqst(xprt, *xp);
 599         if (!rovr)
 600                 goto out_unlock;
 601         task = rovr->rq_task;
 602
 603         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 604                 copied = repsize;
 605
 606         /* Suck it into the iovec, verify checksum if not done by hw. */
 607         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 608                 goto out_unlock;
 609
 610         /* Something worked... */
 611         dst_confirm(skb->dst);
 612
 613         xprt_adjust_cwnd(task, copied);
 614         xprt_update_rtt(task);
 615         xprt_complete_rqst(task, copied);
 616
 617  out_unlock:
 618         spin_unlock(&xprt->transport_lock);
 619  dropit:
 620         skb_free_datagram(sk, skb);
 621  out:
 622         read_unlock(&sk->sk_callback_lock);
 623 }
 624
 625 static inline size_t xs_tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
 626 {
 627         if (len > desc->count)
 628                 len = desc->count;
 629         if (skb_copy_bits(desc->skb, desc->offset, p, len)) {
 630                 dprintk("RPC:      failed to copy %zu bytes from skb. %zu bytes remain\n",
 631                                 len, desc->count);
 632                 return 0;
 633         }
 634         desc->offset += len;
 635         desc->count -= len;
 636         dprintk("RPC:      copied %zu bytes from skb. %zu bytes remain\n",
 637                         len, desc->count);
 638         return len;
 639 }
 640
 641 static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
 642 {
 643         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 644         size_t len, used;
 645         char *p;
 646
 647         p = ((char *) &transport->tcp_fraghdr) + transport->tcp_offset;
 648         len = sizeof(transport->tcp_fraghdr) - transport->tcp_offset;
 649         used = xs_tcp_copy_data(desc, p, len);
 650         transport->tcp_offset += used;
 651         if (used != len)
 652                 return;
 653
 654         transport->tcp_reclen = ntohl(transport->tcp_fraghdr);
 655         if (transport->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
 656                 transport->tcp_flags |= XPRT_LAST_FRAG;
 657         else
 658                 transport->tcp_flags &= ~XPRT_LAST_FRAG;
 659         transport->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
 660
 661         transport->tcp_flags &= ~XPRT_COPY_RECM;
 662         transport->tcp_offset = 0;
 663
 664         /* Sanity check of the record length */
 665         if (unlikely(transport->tcp_reclen < 4)) {
 666                 dprintk("RPC:      invalid TCP record fragment length\n");
 667                 xprt_disconnect(xprt);
 668                 return;
 669         }
 670         dprintk("RPC:      reading TCP record fragment of length %d\n",
 671                         transport->tcp_reclen);
 672 }
 673
 674 static void xs_tcp_check_fraghdr(struct sock_xprt *transport)
 675 {
 676         if (transport->tcp_offset == transport->tcp_reclen) {
 677                 transport->tcp_flags |= XPRT_COPY_RECM;
 678                 transport->tcp_offset = 0;
 679                 if (transport->tcp_flags & XPRT_LAST_FRAG) {
 680                         transport->tcp_flags &= ~XPRT_COPY_DATA;
 681                         transport->tcp_flags |= XPRT_COPY_XID;
 682                         transport->tcp_copied = 0;
 683                 }
 684         }
 685 }
 686
 687 static inline void xs_tcp_read_xid(struct sock_xprt *transport, skb_reader_t *desc)
 688 {
 689         size_t len, used;
 690         char *p;
 691
 692         len = sizeof(transport->tcp_xid) - transport->tcp_offset;
 693         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
 694         p = ((char *) &transport->tcp_xid) + transport->tcp_offset;
 695         used = xs_tcp_copy_data(desc, p, len);
 696         transport->tcp_offset += used;
 697         if (used != len)
 698                 return;
 699         transport->tcp_flags &= ~XPRT_COPY_XID;
 700         transport->tcp_flags |= XPRT_COPY_DATA;
 701         transport->tcp_copied = 4;
 702         dprintk("RPC:      reading reply for XID %08x\n",
 703                         ntohl(transport->tcp_xid));
 704         xs_tcp_check_fraghdr(transport);
 705 }
 706
 707 static inline void xs_tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
 708 {
 709         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 710         struct rpc_rqst *req;
 711         struct xdr_buf *rcvbuf;
 712         size_t len;
 713         ssize_t r;
 714
 715         /* Find and lock the request corresponding to this xid */
 716         spin_lock(&xprt->transport_lock);
 717         req = xprt_lookup_rqst(xprt, transport->tcp_xid);
 718         if (!req) {
 719                 transport->tcp_flags &= ~XPRT_COPY_DATA;
 720                 dprintk("RPC:      XID %08x request not found!\n",
 721                                 ntohl(transport->tcp_xid));
 722                 spin_unlock(&xprt->transport_lock);
 723                 return;
 724         }
 725
 726         rcvbuf = &req->rq_private_buf;
 727         len = desc->count;
 728         if (len > transport->tcp_reclen - transport->tcp_offset) {
 729                 skb_reader_t my_desc;
 730
 731                 len = transport->tcp_reclen - transport->tcp_offset;
 732                 memcpy(&my_desc, desc, sizeof(my_desc));
 733                 my_desc.count = len;
 734                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
 735                                           &my_desc, xs_tcp_copy_data);
 736                 desc->count -= r;
 737                 desc->offset += r;
 738         } else
 739                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
 740                                           desc, xs_tcp_copy_data);
 741
 742         if (r > 0) {
 743                 transport->tcp_copied += r;
 744                 transport->tcp_offset += r;
 745         }
 746         if (r != len) {
 747                 /* Error when copying to the receive buffer,
 748                  * usually because we weren't able to allocate
 749                  * additional buffer pages. All we can do now
 750                  * is turn off XPRT_COPY_DATA, so the request
 751                  * will not receive any additional updates,
 752                  * and time out.
 753                  * Any remaining data from this record will
 754                  * be discarded.
 755                  */
 756                 transport->tcp_flags &= ~XPRT_COPY_DATA;
 757                 dprintk("RPC:      XID %08x truncated request\n",
 758                                 ntohl(transport->tcp_xid));
 759                 dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
 760                                 xprt, transport->tcp_copied, transport->tcp_offset,
 761                                         transport->tcp_reclen);
 762                 goto out;
 763         }
 764
 765         dprintk("RPC:      XID %08x read %Zd bytes\n",
 766                         ntohl(transport->tcp_xid), r);
 767         dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
 768                         xprt, transport->tcp_copied, transport->tcp_offset,
 769                                 transport->tcp_reclen);
 770
 771         if (transport->tcp_copied == req->rq_private_buf.buflen)
 772                 transport->tcp_flags &= ~XPRT_COPY_DATA;
 773         else if (transport->tcp_offset == transport->tcp_reclen) {
 774                 if (transport->tcp_flags & XPRT_LAST_FRAG)
 775                         transport->tcp_flags &= ~XPRT_COPY_DATA;
 776         }
 777
 778 out:
 779         if (!(transport->tcp_flags & XPRT_COPY_DATA))
 780                 xprt_complete_rqst(req->rq_task, transport->tcp_copied);
 781         spin_unlock(&xprt->transport_lock);
 782         xs_tcp_check_fraghdr(transport);
 783 }
 784
 785 static inline void xs_tcp_read_discard(struct sock_xprt *transport, skb_reader_t *desc)
 786 {
 787         size_t len;
 788
 789         len = transport->tcp_reclen - transport->tcp_offset;
 790         if (len > desc->count)
 791                 len = desc->count;
 792         desc->count -= len;
 793         desc->offset += len;
 794         transport->tcp_offset += len;
 795         dprintk("RPC:      discarded %Zu bytes\n", len);
 796         xs_tcp_check_fraghdr(transport);
 797 }
 798
 799 static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
 800 {
 801         struct rpc_xprt *xprt = rd_desc->arg.data;
 802         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 803         skb_reader_t desc = {
 804                 .skb    = skb,
 805                 .offset = offset,
 806                 .count  = len,
 807         };
 808
 809         dprintk("RPC:      xs_tcp_data_recv started\n");
 810         do {
 811                 /* Read in a new fragment marker if necessary */
 812                 /* Can we ever really expect to get completely empty fragments? */
 813                 if (transport->tcp_flags & XPRT_COPY_RECM) {
 814                         xs_tcp_read_fraghdr(xprt, &desc);
 815                         continue;
 816                 }
 817                 /* Read in the xid if necessary */
 818                 if (transport->tcp_flags & XPRT_COPY_XID) {
 819                         xs_tcp_read_xid(transport, &desc);
 820                         continue;
 821                 }
 822                 /* Read in the request data */
 823                 if (transport->tcp_flags & XPRT_COPY_DATA) {
 824                         xs_tcp_read_request(xprt, &desc);
 825                         continue;
 826                 }
 827                 /* Skip over any trailing bytes on short reads */
 828                 xs_tcp_read_discard(transport, &desc);
 829         } while (desc.count);
 830         dprintk("RPC:      xs_tcp_data_recv done\n");
 831         return len - desc.count;
 832 }
 833
 834 /**
 835  * xs_tcp_data_ready - "data ready" callback for TCP sockets
 836  * @sk: socket with data to read
 837  * @bytes: how much data to read
 838  *
 839  */
 840 static void xs_tcp_data_ready(struct sock *sk, int bytes)
 841 {
 842         struct rpc_xprt *xprt;
 843         read_descriptor_t rd_desc;
 844
 845         read_lock(&sk->sk_callback_lock);
 846         dprintk("RPC:      xs_tcp_data_ready...\n");
 847         if (!(xprt = xprt_from_sock(sk)))
 848                 goto out;
 849         if (xprt->shutdown)
 850                 goto out;
 851
 852         /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
 853         rd_desc.arg.data = xprt;
 854         rd_desc.count = 65536;
 855         tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
 856 out:
 857         read_unlock(&sk->sk_callback_lock);
 858 }
 859
 860 /**
 861  * xs_tcp_state_change - callback to handle TCP socket state changes
 862  * @sk: socket whose state has changed
 863  *
 864  */
 865 static void xs_tcp_state_change(struct sock *sk)
 866 {
 867         struct rpc_xprt *xprt;
 868
 869         read_lock(&sk->sk_callback_lock);
 870         if (!(xprt = xprt_from_sock(sk)))
 871                 goto out;
 872         dprintk("RPC:      xs_tcp_state_change client %p...\n", xprt);
 873         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
 874                                 sk->sk_state, xprt_connected(xprt),
 875                                 sock_flag(sk, SOCK_DEAD),
 876                                 sock_flag(sk, SOCK_ZAPPED));
 877
 878         switch (sk->sk_state) {
 879         case TCP_ESTABLISHED:
 880                 spin_lock_bh(&xprt->transport_lock);
 881                 if (!xprt_test_and_set_connected(xprt)) {
 882                         struct sock_xprt *transport = container_of(xprt,
 883                                         struct sock_xprt, xprt);
 884
 885                         /* Reset TCP record info */
 886                         transport->tcp_offset = 0;
 887                         transport->tcp_reclen = 0;
 888                         transport->tcp_copied = 0;
 889                         transport->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
 890
 891                         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
 892                         xprt_wake_pending_tasks(xprt, 0);
 893                 }
 894                 spin_unlock_bh(&xprt->transport_lock);
 895                 break;
 896         case TCP_SYN_SENT:
 897         case TCP_SYN_RECV:
 898                 break;
 899         case TCP_CLOSE_WAIT:
 900                 /* Try to schedule an autoclose RPC calls */
 901                 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
 902                 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
 903                         schedule_work(&xprt->task_cleanup);
 904         default:
 905                 xprt_disconnect(xprt);
 906         }
 907  out:
 908         read_unlock(&sk->sk_callback_lock);
 909 }
 910
 911 /**
 912  * xs_udp_write_space - callback invoked when socket buffer space
 913  *                             becomes available
 914  * @sk: socket whose state has changed
 915  *
 916  * Called when more output buffer space is available for this socket.
 917  * We try not to wake our writers until they can make "significant"
 918  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
 919  * with a bunch of small requests.
 920  */
 921 static void xs_udp_write_space(struct sock *sk)
 922 {
 923         read_lock(&sk->sk_callback_lock);
 924
 925         /* from net/core/sock.c:sock_def_write_space */
 926         if (sock_writeable(sk)) {
 927                 struct socket *sock;
 928                 struct rpc_xprt *xprt;
 929
 930                 if (unlikely(!(sock = sk->sk_socket)))
 931                         goto out;
 932                 if (unlikely(!(xprt = xprt_from_sock(sk))))
 933                         goto out;
 934                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
 935                         goto out;
 936
 937                 xprt_write_space(xprt);
 938         }
 939
 940  out:
 941         read_unlock(&sk->sk_callback_lock);
 942 }
 943
 944 /**
 945  * xs_tcp_write_space - callback invoked when socket buffer space
 946  *                             becomes available
 947  * @sk: socket whose state has changed
 948  *
 949  * Called when more output buffer space is available for this socket.
 950  * We try not to wake our writers until they can make "significant"
 951  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
 952  * with a bunch of small requests.
 953  */
 954 static void xs_tcp_write_space(struct sock *sk)
 955 {
 956         read_lock(&sk->sk_callback_lock);
 957
 958         /* from net/core/stream.c:sk_stream_write_space */
 959         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 960                 struct socket *sock;
 961                 struct rpc_xprt *xprt;
 962
 963                 if (unlikely(!(sock = sk->sk_socket)))
 964                         goto out;
 965                 if (unlikely(!(xprt = xprt_from_sock(sk))))
 966                         goto out;
 967                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
 968                         goto out;
 969
 970                 xprt_write_space(xprt);
 971         }
 972
 973  out:
 974         read_unlock(&sk->sk_callback_lock);
 975 }
 976
 977 static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
 978 {
 979         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 980         struct sock *sk = transport->inet;
 981
 982         if (xprt->rcvsize) {
 983                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
 984                 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
 985         }
 986         if (xprt->sndsize) {
 987                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 988                 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
 989                 sk->sk_write_space(sk);
 990         }
 991 }
 992
 993 /**
 994  * xs_udp_set_buffer_size - set send and receive limits
 995  * @xprt: generic transport
 996  * @sndsize: requested size of send buffer, in bytes
 997  * @rcvsize: requested size of receive buffer, in bytes
 998  *
 999  * Set socket send and receive buffer size limits.
1000  */
1001 static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
1002 {
1003         xprt->sndsize = 0;
1004         if (sndsize)
1005                 xprt->sndsize = sndsize + 1024;
1006         xprt->rcvsize = 0;
1007         if (rcvsize)
1008                 xprt->rcvsize = rcvsize + 1024;
1009
1010         xs_udp_do_set_buffer_size(xprt);
1011 }
1012
1013 /**
1014  * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
1015  * @task: task that timed out
1016  *
1017  * Adjust the congestion window after a retransmit timeout has occurred.
1018  */
1019 static void xs_udp_timer(struct rpc_task *task)
1020 {
1021         xprt_adjust_cwnd(task, -ETIMEDOUT);
1022 }
1023
1024 static unsigned short xs_get_random_port(void)
1025 {
1026         unsigned short range = xprt_max_resvport - xprt_min_resvport;
1027         unsigned short rand = (unsigned short) net_random() % range;
1028         return rand + xprt_min_resvport;
1029 }
1030
1031 /**
1032  * xs_print_peer_address - format an IPv4 address for printing
1033  * @xprt: generic transport
1034  * @format: flags field indicating which parts of the address to render
1035  */
1036 static char *xs_print_peer_address(struct rpc_xprt *xprt, enum rpc_display_format_t format)
1037 {
1038         if (xprt->address_strings[format] != NULL)
1039                 return xprt->address_strings[format];
1040         else
1041                 return "unprintable";
1042 }
1043
1044 /**
1045  * xs_set_port - reset the port number in the remote endpoint address
1046  * @xprt: generic transport
1047  * @port: new port number
1048  *
1049  */
1050 static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
1051 {
1052         struct sockaddr_in *sap = (struct sockaddr_in *) &xprt->addr;
1053
1054         dprintk("RPC:      setting port for xprt %p to %u\n", xprt, port);
1055
1056         sap->sin_port = htons(port);
1057 }
1058
1059 static int xs_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
1060 {
1061         struct sockaddr_in myaddr = {
1062                 .sin_family = AF_INET,
1063         };
1064         int err;
1065         unsigned short port = xprt->port;
1066
1067         do {
1068                 myaddr.sin_port = htons(port);
1069                 err = kernel_bind(sock, (struct sockaddr *) &myaddr,
1070                                                 sizeof(myaddr));
1071                 if (err == 0) {
1072                         xprt->port = port;
1073                         dprintk("RPC:      xs_bindresvport bound to port %u\n",
1074                                         port);
1075                         return 0;
1076                 }
1077                 if (port <= xprt_min_resvport)
1078                         port = xprt_max_resvport;
1079                 else
1080                         port--;
1081         } while (err == -EADDRINUSE && port != xprt->port);
1082
1083         dprintk("RPC:      can't bind to reserved port (%d).\n", -err);
1084         return err;
1085 }
1086
1087 /**
1088  * xs_udp_connect_worker - set up a UDP socket
1089  * @args: RPC transport to connect
1090  *
1091  * Invoked by a work queue tasklet.
1092  */
1093 static void xs_udp_connect_worker(void *args)
1094 {
1095         struct rpc_xprt *xprt = (struct rpc_xprt *) args;
1096         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1097         struct socket *sock = transport->sock;
1098         int err, status = -EIO;
1099
1100         if (xprt->shutdown || !xprt_bound(xprt))
1101                 goto out;
1102
1103         /* Start by resetting any existing state */
1104         xs_close(xprt);
1105
1106         if ((err = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock)) < 0) {
1107                 dprintk("RPC:      can't create UDP transport socket (%d).\n", -err);
1108                 goto out;
1109         }
1110
1111         if (xprt->resvport && xs_bindresvport(xprt, sock) < 0) {
1112                 sock_release(sock);
1113                 goto out;
1114         }
1115
1116         dprintk("RPC:      worker connecting xprt %p to address: %s\n",
1117                         xprt, xs_print_peer_address(xprt, RPC_DISPLAY_ALL));
1118
1119         if (!transport->inet) {
1120                 struct sock *sk = sock->sk;
1121
1122                 write_lock_bh(&sk->sk_callback_lock);
1123
1124                 sk->sk_user_data = xprt;
1125                 xprt->old_data_ready = sk->sk_data_ready;
1126                 xprt->old_state_change = sk->sk_state_change;
1127                 xprt->old_write_space = sk->sk_write_space;
1128                 sk->sk_data_ready = xs_udp_data_ready;
1129                 sk->sk_write_space = xs_udp_write_space;
1130                 sk->sk_no_check = UDP_CSUM_NORCV;
1131                 sk->sk_allocation = GFP_ATOMIC;
1132
1133                 xprt_set_connected(xprt);
1134
1135                 /* Reset to new socket */
1136                 transport->sock = sock;
1137                 transport->inet = sk;
1138
1139                 write_unlock_bh(&sk->sk_callback_lock);
1140         }
1141         xs_udp_do_set_buffer_size(xprt);
1142         status = 0;
1143 out:
1144         xprt_wake_pending_tasks(xprt, status);
1145         xprt_clear_connecting(xprt);
1146 }
1147
1148 /*
1149  * We need to preserve the port number so the reply cache on the server can
1150  * find our cached RPC replies when we get around to reconnecting.
1151  */
1152 static void xs_tcp_reuse_connection(struct rpc_xprt *xprt)
1153 {
1154         int result;
1155         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1156         struct sockaddr any;
1157
1158         dprintk("RPC:      disconnecting xprt %p to reuse port\n", xprt);
1159
1160         /*
1161          * Disconnect the transport socket by doing a connect operation
1162          * with AF_UNSPEC.  This should return immediately...
1163          */
1164         memset(&any, 0, sizeof(any));
1165         any.sa_family = AF_UNSPEC;
1166         result = kernel_connect(transport->sock, &any, sizeof(any), 0);
1167         if (result)
1168                 dprintk("RPC:      AF_UNSPEC connect return code %d\n",
1169                                 result);
1170 }
1171
1172 /**
1173  * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1174  * @args: RPC transport to connect
1175  *
1176  * Invoked by a work queue tasklet.
1177  */
1178 static void xs_tcp_connect_worker(void *args)
1179 {
1180         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
1181         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1182         struct socket *sock = transport->sock;
1183         int err, status = -EIO;
1184
1185         if (xprt->shutdown || !xprt_bound(xprt))
1186                 goto out;
1187
1188         if (!sock) {
1189                 /* start from scratch */
1190                 if ((err = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock)) < 0) {
1191                         dprintk("RPC:      can't create TCP transport socket (%d).\n", -err);
1192                         goto out;
1193                 }
1194
1195                 if (xprt->resvport && xs_bindresvport(xprt, sock) < 0) {
1196                         sock_release(sock);
1197                         goto out;
1198                 }
1199         } else
1200                 /* "close" the socket, preserving the local port */
1201                 xs_tcp_reuse_connection(xprt);
1202
1203         dprintk("RPC:      worker connecting xprt %p to address: %s\n",
1204                         xprt, xs_print_peer_address(xprt, RPC_DISPLAY_ALL));
1205
1206         if (!transport->inet) {
1207                 struct sock *sk = sock->sk;
1208
1209                 write_lock_bh(&sk->sk_callback_lock);
1210
1211                 sk->sk_user_data = xprt;
1212                 xprt->old_data_ready = sk->sk_data_ready;
1213                 xprt->old_state_change = sk->sk_state_change;
1214                 xprt->old_write_space = sk->sk_write_space;
1215                 sk->sk_data_ready = xs_tcp_data_ready;
1216                 sk->sk_state_change = xs_tcp_state_change;
1217                 sk->sk_write_space = xs_tcp_write_space;
1218                 sk->sk_allocation = GFP_ATOMIC;
1219
1220                 /* socket options */
1221                 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
1222                 sock_reset_flag(sk, SOCK_LINGER);
1223                 tcp_sk(sk)->linger2 = 0;
1224                 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1225
1226                 xprt_clear_connected(xprt);
1227
1228                 /* Reset to new socket */
1229                 transport->sock = sock;
1230                 transport->inet = sk;
1231
1232                 write_unlock_bh(&sk->sk_callback_lock);
1233         }
1234
1235         /* Tell the socket layer to start connecting... */
1236         xprt->stat.connect_count++;
1237         xprt->stat.connect_start = jiffies;
1238         status = kernel_connect(sock, (struct sockaddr *) &xprt->addr,
1239                         xprt->addrlen, O_NONBLOCK);
1240         dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
1241                         xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
1242         if (status < 0) {
1243                 switch (status) {
1244                         case -EINPROGRESS:
1245                         case -EALREADY:
1246                                 goto out_clear;
1247                         case -ECONNREFUSED:
1248                         case -ECONNRESET:
1249                                 /* retry with existing socket, after a delay */
1250                                 break;
1251                         default:
1252                                 /* get rid of existing socket, and retry */
1253                                 xs_close(xprt);
1254                                 break;
1255                 }
1256         }
1257 out:
1258         xprt_wake_pending_tasks(xprt, status);
1259 out_clear:
1260         xprt_clear_connecting(xprt);
1261 }
1262
1263 /**
1264  * xs_connect - connect a socket to a remote endpoint
1265  * @task: address of RPC task that manages state of connect request
1266  *
1267  * TCP: If the remote end dropped the connection, delay reconnecting.
1268  *
1269  * UDP socket connects are synchronous, but we use a work queue anyway
1270  * to guarantee that even unprivileged user processes can set up a
1271  * socket on a privileged port.
1272  *
1273  * If a UDP socket connect fails, the delay behavior here prevents
1274  * retry floods (hard mounts).
1275  */
1276 static void xs_connect(struct rpc_task *task)
1277 {
1278         struct rpc_xprt *xprt = task->tk_xprt;
1279         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1280
1281         if (xprt_test_and_set_connecting(xprt))
1282                 return;
1283
1284         if (transport->sock != NULL) {
1285                 dprintk("RPC:      xs_connect delayed xprt %p for %lu seconds\n",
1286                                 xprt, xprt->reestablish_timeout / HZ);
1287                 schedule_delayed_work(&xprt->connect_worker,
1288                                         xprt->reestablish_timeout);
1289                 xprt->reestablish_timeout <<= 1;
1290                 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
1291                         xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
1292         } else {
1293                 dprintk("RPC:      xs_connect scheduled xprt %p\n", xprt);
1294                 schedule_work(&xprt->connect_worker);
1295
1296                 /* flush_scheduled_work can sleep... */
1297                 if (!RPC_IS_ASYNC(task))
1298                         flush_scheduled_work();
1299         }
1300 }
1301
1302 /**
1303  * xs_udp_print_stats - display UDP socket-specifc stats
1304  * @xprt: rpc_xprt struct containing statistics
1305  * @seq: output file
1306  *
1307  */
1308 static void xs_udp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1309 {
1310         seq_printf(seq, "\txprt:\tudp %u %lu %lu %lu %lu %Lu %Lu\n",
1311                         xprt->port,
1312                         xprt->stat.bind_count,
1313                         xprt->stat.sends,
1314                         xprt->stat.recvs,
1315                         xprt->stat.bad_xids,
1316                         xprt->stat.req_u,
1317                         xprt->stat.bklog_u);
1318 }
1319
1320 /**
1321  * xs_tcp_print_stats - display TCP socket-specifc stats
1322  * @xprt: rpc_xprt struct containing statistics
1323  * @seq: output file
1324  *
1325  */
1326 static void xs_tcp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1327 {
1328         long idle_time = 0;
1329
1330         if (xprt_connected(xprt))
1331                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
1332
1333         seq_printf(seq, "\txprt:\ttcp %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu\n",
1334                         xprt->port,
1335                         xprt->stat.bind_count,
1336                         xprt->stat.connect_count,
1337                         xprt->stat.connect_time,
1338                         idle_time,
1339                         xprt->stat.sends,
1340                         xprt->stat.recvs,
1341                         xprt->stat.bad_xids,
1342                         xprt->stat.req_u,
1343                         xprt->stat.bklog_u);
1344 }
1345
1346 static struct rpc_xprt_ops xs_udp_ops = {
1347         .set_buffer_size        = xs_udp_set_buffer_size,
1348         .print_addr             = xs_print_peer_address,
1349         .reserve_xprt           = xprt_reserve_xprt_cong,
1350         .release_xprt           = xprt_release_xprt_cong,
1351         .rpcbind                = rpc_getport,
1352         .set_port               = xs_set_port,
1353         .connect                = xs_connect,
1354         .buf_alloc              = rpc_malloc,
1355         .buf_free               = rpc_free,
1356         .send_request           = xs_udp_send_request,
1357         .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
1358         .timer                  = xs_udp_timer,
1359         .release_request        = xprt_release_rqst_cong,
1360         .close                  = xs_close,
1361         .destroy                = xs_destroy,
1362         .print_stats            = xs_udp_print_stats,
1363 };
1364
1365 static struct rpc_xprt_ops xs_tcp_ops = {
1366         .print_addr             = xs_print_peer_address,
1367         .reserve_xprt           = xprt_reserve_xprt,
1368         .release_xprt           = xs_tcp_release_xprt,
1369         .rpcbind                = rpc_getport,
1370         .set_port               = xs_set_port,
1371         .connect                = xs_connect,
1372         .buf_alloc              = rpc_malloc,
1373         .buf_free               = rpc_free,
1374         .send_request           = xs_tcp_send_request,
1375         .set_retrans_timeout    = xprt_set_retrans_timeout_def,
1376         .close                  = xs_close,
1377         .destroy                = xs_destroy,
1378         .print_stats            = xs_tcp_print_stats,
1379 };
1380
1381 static struct rpc_xprt *xs_setup_xprt(struct sockaddr *addr, size_t addrlen, unsigned int slot_table_size)
1382 {
1383         struct rpc_xprt *xprt;
1384         struct sock_xprt *new;
1385
1386         if (addrlen > sizeof(xprt->addr)) {
1387                 dprintk("RPC:      xs_setup_xprt: address too large\n");
1388                 return ERR_PTR(-EBADF);
1389         }
1390
1391         new = kzalloc(sizeof(*new), GFP_KERNEL);
1392         if (new == NULL) {
1393                 dprintk("RPC:      xs_setup_xprt: couldn't allocate rpc_xprt\n");
1394                 return ERR_PTR(-ENOMEM);
1395         }
1396         xprt = &new->xprt;
1397
1398         xprt->max_reqs = slot_table_size;
1399         xprt->slot = kcalloc(xprt->max_reqs, sizeof(struct rpc_rqst), GFP_KERNEL);
1400         if (xprt->slot == NULL) {
1401                 kfree(xprt);
1402                 dprintk("RPC:      xs_setup_xprt: couldn't allocate slot table\n");
1403                 return ERR_PTR(-ENOMEM);
1404         }
1405
1406         memcpy(&xprt->addr, addr, addrlen);
1407         xprt->addrlen = addrlen;
1408         xprt->port = xs_get_random_port();
1409
1410         return xprt;
1411 }
1412
1413 /**
1414  * xs_setup_udp - Set up transport to use a UDP socket
1415  * @addr: address of remote server
1416  * @addrlen: length of address in bytes
1417  * @to:   timeout parameters
1418  *
1419  */
1420 struct rpc_xprt *xs_setup_udp(struct sockaddr *addr, size_t addrlen, struct rpc_timeout *to)
1421 {
1422         struct rpc_xprt *xprt;
1423
1424         xprt = xs_setup_xprt(addr, addrlen, xprt_udp_slot_table_entries);
1425         if (IS_ERR(xprt))
1426                 return xprt;
1427
1428         if (ntohs(((struct sockaddr_in *)addr)->sin_port) != 0)
1429                 xprt_set_bound(xprt);
1430
1431         xprt->prot = IPPROTO_UDP;
1432         xprt->tsh_size = 0;
1433         /* XXX: header size can vary due to auth type, IPv6, etc. */
1434         xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1435
1436         INIT_WORK(&xprt->connect_worker, xs_udp_connect_worker, xprt);
1437         xprt->bind_timeout = XS_BIND_TO;
1438         xprt->connect_timeout = XS_UDP_CONN_TO;
1439         xprt->reestablish_timeout = XS_UDP_REEST_TO;
1440         xprt->idle_timeout = XS_IDLE_DISC_TO;
1441
1442         xprt->ops = &xs_udp_ops;
1443
1444         if (to)
1445                 xprt->timeout = *to;
1446         else
1447                 xprt_set_timeout(&xprt->timeout, 5, 5 * HZ);
1448
1449         xs_format_peer_addresses(xprt);
1450         dprintk("RPC:      set up transport to address %s\n",
1451                         xs_print_peer_address(xprt, RPC_DISPLAY_ALL));
1452
1453         return xprt;
1454 }
1455
1456 /**
1457  * xs_setup_tcp - Set up transport to use a TCP socket
1458  * @addr: address of remote server
1459  * @addrlen: length of address in bytes
1460  * @to: timeout parameters
1461  *
1462  */
1463 struct rpc_xprt *xs_setup_tcp(struct sockaddr *addr, size_t addrlen, struct rpc_timeout *to)
1464 {
1465         struct rpc_xprt *xprt;
1466
1467         xprt = xs_setup_xprt(addr, addrlen, xprt_tcp_slot_table_entries);
1468         if (IS_ERR(xprt))
1469                 return xprt;
1470
1471         if (ntohs(((struct sockaddr_in *)addr)->sin_port) != 0)
1472                 xprt_set_bound(xprt);
1473
1474         xprt->prot = IPPROTO_TCP;
1475         xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
1476         xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
1477
1478         INIT_WORK(&xprt->connect_worker, xs_tcp_connect_worker, xprt);
1479         xprt->bind_timeout = XS_BIND_TO;
1480         xprt->connect_timeout = XS_TCP_CONN_TO;
1481         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1482         xprt->idle_timeout = XS_IDLE_DISC_TO;
1483
1484         xprt->ops = &xs_tcp_ops;
1485
1486         if (to)
1487                 xprt->timeout = *to;
1488         else
1489                 xprt_set_timeout(&xprt->timeout, 2, 60 * HZ);
1490
1491         xs_format_peer_addresses(xprt);
1492         dprintk("RPC:      set up transport to address %s\n",
1493                         xs_print_peer_address(xprt, RPC_DISPLAY_ALL));
1494
1495         return xprt;
1496 }