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 <alan@redhat.com>
   7  * TCP send fixes (C) 1998 Red Hat Software <alan@redhat.com>
   8  * TCP NFS related read + write fixes
   9  *  (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  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 <trond.myklebust@fys.uio.no>
  14  *
  15  * IP socket transport implementation, (C) 2005 Chuck Lever <cel@netapp.com>
  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  * We can register our own files under /proc/sys/sunrpc by
  50  * calling register_sysctl_table() again.  The files in that
  51  * directory become the union of all files registered there.
  52  *
  53  * We simply need to make sure that we don't collide with
  54  * someone else's file names!
  55  */
  56
  57 #ifdef RPC_DEBUG
  58
  59 static unsigned int min_slot_table_size = RPC_MIN_SLOT_TABLE;
  60 static unsigned int max_slot_table_size = RPC_MAX_SLOT_TABLE;
  61 static unsigned int xprt_min_resvport_limit = RPC_MIN_RESVPORT;
  62 static unsigned int xprt_max_resvport_limit = RPC_MAX_RESVPORT;
  63
  64 static struct ctl_table_header *sunrpc_table_header;
  65
  66 /*
  67  * FIXME: changing the UDP slot table size should also resize the UDP
  68  *        socket buffers for existing UDP transports
  69  */
  70 static ctl_table xs_tunables_table[] = {
  71         {
  72                 .ctl_name       = CTL_SLOTTABLE_UDP,
  73                 .procname       = "udp_slot_table_entries",
  74                 .data           = &xprt_udp_slot_table_entries,
  75                 .maxlen         = sizeof(unsigned int),
  76                 .mode           = 0644,
  77                 .proc_handler   = &proc_dointvec_minmax,
  78                 .strategy       = &sysctl_intvec,
  79                 .extra1         = &min_slot_table_size,
  80                 .extra2         = &max_slot_table_size
  81         },
  82         {
  83                 .ctl_name       = CTL_SLOTTABLE_TCP,
  84                 .procname       = "tcp_slot_table_entries",
  85                 .data           = &xprt_tcp_slot_table_entries,
  86                 .maxlen         = sizeof(unsigned int),
  87                 .mode           = 0644,
  88                 .proc_handler   = &proc_dointvec_minmax,
  89                 .strategy       = &sysctl_intvec,
  90                 .extra1         = &min_slot_table_size,
  91                 .extra2         = &max_slot_table_size
  92         },
  93         {
  94                 .ctl_name       = CTL_MIN_RESVPORT,
  95                 .procname       = "min_resvport",
  96                 .data           = &xprt_min_resvport,
  97                 .maxlen         = sizeof(unsigned int),
  98                 .mode           = 0644,
  99                 .proc_handler   = &proc_dointvec_minmax,
 100                 .strategy       = &sysctl_intvec,
 101                 .extra1         = &xprt_min_resvport_limit,
 102                 .extra2         = &xprt_max_resvport_limit
 103         },
 104         {
 105                 .ctl_name       = CTL_MAX_RESVPORT,
 106                 .procname       = "max_resvport",
 107                 .data           = &xprt_max_resvport,
 108                 .maxlen         = sizeof(unsigned int),
 109                 .mode           = 0644,
 110                 .proc_handler   = &proc_dointvec_minmax,
 111                 .strategy       = &sysctl_intvec,
 112                 .extra1         = &xprt_min_resvport_limit,
 113                 .extra2         = &xprt_max_resvport_limit
 114         },
 115         {
 116                 .ctl_name = 0,
 117         },
 118 };
 119
 120 static ctl_table sunrpc_table[] = {
 121         {
 122                 .ctl_name       = CTL_SUNRPC,
 123                 .procname       = "sunrpc",
 124                 .mode           = 0555,
 125                 .child          = xs_tunables_table
 126         },
 127         {
 128                 .ctl_name = 0,
 129         },
 130 };
 131
 132 #endif
 133
 134 /*
 135  * How many times to try sending a request on a socket before waiting
 136  * for the socket buffer to clear.
 137  */
 138 #define XS_SENDMSG_RETRY        (10U)
 139
 140 /*
 141  * Time out for an RPC UDP socket connect.  UDP socket connects are
 142  * synchronous, but we set a timeout anyway in case of resource
 143  * exhaustion on the local host.
 144  */
 145 #define XS_UDP_CONN_TO          (5U * HZ)
 146
 147 /*
 148  * Wait duration for an RPC TCP connection to be established.  Solaris
 149  * NFS over TCP uses 60 seconds, for example, which is in line with how
 150  * long a server takes to reboot.
 151  */
 152 #define XS_TCP_CONN_TO          (60U * HZ)
 153
 154 /*
 155  * Wait duration for a reply from the RPC portmapper.
 156  */
 157 #define XS_BIND_TO              (60U * HZ)
 158
 159 /*
 160  * Delay if a UDP socket connect error occurs.  This is most likely some
 161  * kind of resource problem on the local host.
 162  */
 163 #define XS_UDP_REEST_TO         (2U * HZ)
 164
 165 /*
 166  * The reestablish timeout allows clients to delay for a bit before attempting
 167  * to reconnect to a server that just dropped our connection.
 168  *
 169  * We implement an exponential backoff when trying to reestablish a TCP
 170  * transport connection with the server.  Some servers like to drop a TCP
 171  * connection when they are overworked, so we start with a short timeout and
 172  * increase over time if the server is down or not responding.
 173  */
 174 #define XS_TCP_INIT_REEST_TO    (3U * HZ)
 175 #define XS_TCP_MAX_REEST_TO     (5U * 60 * HZ)
 176
 177 /*
 178  * TCP idle timeout; client drops the transport socket if it is idle
 179  * for this long.  Note that we also timeout UDP sockets to prevent
 180  * holding port numbers when there is no RPC traffic.
 181  */
 182 #define XS_IDLE_DISC_TO         (5U * 60 * HZ)
 183
 184 #ifdef RPC_DEBUG
 185 # undef  RPC_DEBUG_DATA
 186 # define RPCDBG_FACILITY        RPCDBG_TRANS
 187 #endif
 188
 189 #ifdef RPC_DEBUG_DATA
 190 static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
 191 {
 192         u8 *buf = (u8 *) packet;
 193         int j;
 194
 195         dprintk("RPC:      %s\n", msg);
 196         for (j = 0; j < count && j < 128; j += 4) {
 197                 if (!(j & 31)) {
 198                         if (j)
 199                                 dprintk("\n");
 200                         dprintk("0x%04x ", j);
 201                 }
 202                 dprintk("%02x%02x%02x%02x ",
 203                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
 204         }
 205         dprintk("\n");
 206 }
 207 #else
 208 static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
 209 {
 210         /* NOP */
 211 }
 212 #endif
 213
 214 struct sock_xprt {
 215         struct rpc_xprt         xprt;
 216
 217         /*
 218          * Network layer
 219          */
 220         struct socket *         sock;
 221         struct sock *           inet;
 222
 223         /*
 224          * State of TCP reply receive
 225          */
 226         __be32                  tcp_fraghdr,
 227                                 tcp_xid;
 228
 229         u32                     tcp_offset,
 230                                 tcp_reclen;
 231
 232         unsigned long           tcp_copied,
 233                                 tcp_flags;
 234
 235         /*
 236          * Connection of transports
 237          */
 238         struct delayed_work     connect_worker;
 239         unsigned short          port;
 240
 241         /*
 242          * UDP socket buffer size parameters
 243          */
 244         size_t                  rcvsize,
 245                                 sndsize;
 246
 247         /*
 248          * Saved socket callback addresses
 249          */
 250         void                    (*old_data_ready)(struct sock *, int);
 251         void                    (*old_state_change)(struct sock *);
 252         void                    (*old_write_space)(struct sock *);
 253 };
 254
 255 /*
 256  * TCP receive state flags
 257  */
 258 #define TCP_RCV_LAST_FRAG       (1UL << 0)
 259 #define TCP_RCV_COPY_FRAGHDR    (1UL << 1)
 260 #define TCP_RCV_COPY_XID        (1UL << 2)
 261 #define TCP_RCV_COPY_DATA       (1UL << 3)
 262
 263 static void xs_format_peer_addresses(struct rpc_xprt *xprt)
 264 {
 265         struct sockaddr_in *addr = (struct sockaddr_in *) &xprt->addr;
 266         char *buf;
 267
 268         buf = kzalloc(20, GFP_KERNEL);
 269         if (buf) {
 270                 snprintf(buf, 20, "%u.%u.%u.%u",
 271                                 NIPQUAD(addr->sin_addr.s_addr));
 272         }
 273         xprt->address_strings[RPC_DISPLAY_ADDR] = buf;
 274
 275         buf = kzalloc(8, GFP_KERNEL);
 276         if (buf) {
 277                 snprintf(buf, 8, "%u",
 278                                 ntohs(addr->sin_port));
 279         }
 280         xprt->address_strings[RPC_DISPLAY_PORT] = buf;
 281
 282         if (xprt->prot == IPPROTO_UDP)
 283                 xprt->address_strings[RPC_DISPLAY_PROTO] = "udp";
 284         else
 285                 xprt->address_strings[RPC_DISPLAY_PROTO] = "tcp";
 286
 287         buf = kzalloc(48, GFP_KERNEL);
 288         if (buf) {
 289                 snprintf(buf, 48, "addr=%u.%u.%u.%u port=%u proto=%s",
 290                         NIPQUAD(addr->sin_addr.s_addr),
 291                         ntohs(addr->sin_port),
 292                         xprt->prot == IPPROTO_UDP ? "udp" : "tcp");
 293         }
 294         xprt->address_strings[RPC_DISPLAY_ALL] = buf;
 295 }
 296
 297 static void xs_free_peer_addresses(struct rpc_xprt *xprt)
 298 {
 299         kfree(xprt->address_strings[RPC_DISPLAY_ADDR]);
 300         kfree(xprt->address_strings[RPC_DISPLAY_PORT]);
 301         kfree(xprt->address_strings[RPC_DISPLAY_ALL]);
 302 }
 303
 304 #define XS_SENDMSG_FLAGS        (MSG_DONTWAIT | MSG_NOSIGNAL)
 305
 306 static int xs_send_kvec(struct socket *sock, struct sockaddr *addr, int addrlen, struct kvec *vec, unsigned int base, int more)
 307 {
 308         struct msghdr msg = {
 309                 .msg_name       = addr,
 310                 .msg_namelen    = addrlen,
 311                 .msg_flags      = XS_SENDMSG_FLAGS | (more ? MSG_MORE : 0),
 312         };
 313         struct kvec iov = {
 314                 .iov_base       = vec->iov_base + base,
 315                 .iov_len        = vec->iov_len - base,
 316         };
 317
 318         if (iov.iov_len != 0)
 319                 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
 320         return kernel_sendmsg(sock, &msg, NULL, 0, 0);
 321 }
 322
 323 static int xs_send_pagedata(struct socket *sock, struct xdr_buf *xdr, unsigned int base, int more)
 324 {
 325         struct page **ppage;
 326         unsigned int remainder;
 327         int err, sent = 0;
 328
 329         remainder = xdr->page_len - base;
 330         base += xdr->page_base;
 331         ppage = xdr->pages + (base >> PAGE_SHIFT);
 332         base &= ~PAGE_MASK;
 333         for(;;) {
 334                 unsigned int len = min_t(unsigned int, PAGE_SIZE - base, remainder);
 335                 int flags = XS_SENDMSG_FLAGS;
 336
 337                 remainder -= len;
 338                 if (remainder != 0 || more)
 339                         flags |= MSG_MORE;
 340                 err = sock->ops->sendpage(sock, *ppage, base, len, flags);
 341                 if (remainder == 0 || err != len)
 342                         break;
 343                 sent += err;
 344                 ppage++;
 345                 base = 0;
 346         }
 347         if (sent == 0)
 348                 return err;
 349         if (err > 0)
 350                 sent += err;
 351         return sent;
 352 }
 353
 354 /**
 355  * xs_sendpages - write pages directly to a socket
 356  * @sock: socket to send on
 357  * @addr: UDP only -- address of destination
 358  * @addrlen: UDP only -- length of destination address
 359  * @xdr: buffer containing this request
 360  * @base: starting position in the buffer
 361  *
 362  */
 363 static int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
 364 {
 365         unsigned int remainder = xdr->len - base;
 366         int err, sent = 0;
 367
 368         if (unlikely(!sock))
 369                 return -ENOTCONN;
 370
 371         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 372         if (base != 0) {
 373                 addr = NULL;
 374                 addrlen = 0;
 375         }
 376
 377         if (base < xdr->head[0].iov_len || addr != NULL) {
 378                 unsigned int len = xdr->head[0].iov_len - base;
 379                 remainder -= len;
 380                 err = xs_send_kvec(sock, addr, addrlen, &xdr->head[0], base, remainder != 0);
 381                 if (remainder == 0 || err != len)
 382                         goto out;
 383                 sent += err;
 384                 base = 0;
 385         } else
 386                 base -= xdr->head[0].iov_len;
 387
 388         if (base < xdr->page_len) {
 389                 unsigned int len = xdr->page_len - base;
 390                 remainder -= len;
 391                 err = xs_send_pagedata(sock, xdr, base, remainder != 0);
 392                 if (remainder == 0 || err != len)
 393                         goto out;
 394                 sent += err;
 395                 base = 0;
 396         } else
 397                 base -= xdr->page_len;
 398
 399         if (base >= xdr->tail[0].iov_len)
 400                 return sent;
 401         err = xs_send_kvec(sock, NULL, 0, &xdr->tail[0], base, 0);
 402 out:
 403         if (sent == 0)
 404                 return err;
 405         if (err > 0)
 406                 sent += err;
 407         return sent;
 408 }
 409
 410 /**
 411  * xs_nospace - place task on wait queue if transmit was incomplete
 412  * @task: task to put to sleep
 413  *
 414  */
 415 static void xs_nospace(struct rpc_task *task)
 416 {
 417         struct rpc_rqst *req = task->tk_rqstp;
 418         struct rpc_xprt *xprt = req->rq_xprt;
 419         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 420
 421         dprintk("RPC: %4d xmit incomplete (%u left of %u)\n",
 422                         task->tk_pid, req->rq_slen - req->rq_bytes_sent,
 423                         req->rq_slen);
 424
 425         if (test_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags)) {
 426                 /* Protect against races with write_space */
 427                 spin_lock_bh(&xprt->transport_lock);
 428
 429                 /* Don't race with disconnect */
 430                 if (!xprt_connected(xprt))
 431                         task->tk_status = -ENOTCONN;
 432                 else if (test_bit(SOCK_NOSPACE, &transport->sock->flags))
 433                         xprt_wait_for_buffer_space(task);
 434
 435                 spin_unlock_bh(&xprt->transport_lock);
 436         } else
 437                 /* Keep holding the socket if it is blocked */
 438                 rpc_delay(task, HZ>>4);
 439 }
 440
 441 /**
 442  * xs_udp_send_request - write an RPC request to a UDP socket
 443  * @task: address of RPC task that manages the state of an RPC request
 444  *
 445  * Return values:
 446  *        0:    The request has been sent
 447  *   EAGAIN:    The socket was blocked, please call again later to
 448  *              complete the request
 449  * ENOTCONN:    Caller needs to invoke connect logic then call again
 450  *    other:    Some other error occured, the request was not sent
 451  */
 452 static int xs_udp_send_request(struct rpc_task *task)
 453 {
 454         struct rpc_rqst *req = task->tk_rqstp;
 455         struct rpc_xprt *xprt = req->rq_xprt;
 456         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 457         struct xdr_buf *xdr = &req->rq_snd_buf;
 458         int status;
 459
 460         xs_pktdump("packet data:",
 461                                 req->rq_svec->iov_base,
 462                                 req->rq_svec->iov_len);
 463
 464         req->rq_xtime = jiffies;
 465         status = xs_sendpages(transport->sock,
 466                               (struct sockaddr *) &xprt->addr,
 467                               xprt->addrlen, xdr,
 468                               req->rq_bytes_sent);
 469
 470         dprintk("RPC:      xs_udp_send_request(%u) = %d\n",
 471                         xdr->len - req->rq_bytes_sent, status);
 472
 473         if (likely(status >= (int) req->rq_slen))
 474                 return 0;
 475
 476         /* Still some bytes left; set up for a retry later. */
 477         if (status > 0)
 478                 status = -EAGAIN;
 479
 480         switch (status) {
 481         case -ENETUNREACH:
 482         case -EPIPE:
 483         case -ECONNREFUSED:
 484                 /* When the server has died, an ICMP port unreachable message
 485                  * prompts ECONNREFUSED. */
 486                 break;
 487         case -EAGAIN:
 488                 xs_nospace(task);
 489                 break;
 490         default:
 491                 dprintk("RPC:      sendmsg returned unrecognized error %d\n",
 492                         -status);
 493                 break;
 494         }
 495
 496         return status;
 497 }
 498
 499 static inline void xs_encode_tcp_record_marker(struct xdr_buf *buf)
 500 {
 501         u32 reclen = buf->len - sizeof(rpc_fraghdr);
 502         rpc_fraghdr *base = buf->head[0].iov_base;
 503         *base = htonl(RPC_LAST_STREAM_FRAGMENT | reclen);
 504 }
 505
 506 /**
 507  * xs_tcp_send_request - write an RPC request to a TCP socket
 508  * @task: address of RPC task that manages the state of an RPC request
 509  *
 510  * Return values:
 511  *        0:    The request has been sent
 512  *   EAGAIN:    The socket was blocked, please call again later to
 513  *              complete the request
 514  * ENOTCONN:    Caller needs to invoke connect logic then call again
 515  *    other:    Some other error occured, the request was not sent
 516  *
 517  * XXX: In the case of soft timeouts, should we eventually give up
 518  *      if sendmsg is not able to make progress?
 519  */
 520 static int xs_tcp_send_request(struct rpc_task *task)
 521 {
 522         struct rpc_rqst *req = task->tk_rqstp;
 523         struct rpc_xprt *xprt = req->rq_xprt;
 524         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 525         struct xdr_buf *xdr = &req->rq_snd_buf;
 526         int status, retry = 0;
 527
 528         xs_encode_tcp_record_marker(&req->rq_snd_buf);
 529
 530         xs_pktdump("packet data:",
 531                                 req->rq_svec->iov_base,
 532                                 req->rq_svec->iov_len);
 533
 534         /* Continue transmitting the packet/record. We must be careful
 535          * to cope with writespace callbacks arriving _after_ we have
 536          * called sendmsg(). */
 537         while (1) {
 538                 req->rq_xtime = jiffies;
 539                 status = xs_sendpages(transport->sock,
 540                                         NULL, 0, xdr, req->rq_bytes_sent);
 541
 542                 dprintk("RPC:      xs_tcp_send_request(%u) = %d\n",
 543                                 xdr->len - req->rq_bytes_sent, status);
 544
 545                 if (unlikely(status < 0))
 546                         break;
 547
 548                 /* If we've sent the entire packet, immediately
 549                  * reset the count of bytes sent. */
 550                 req->rq_bytes_sent += status;
 551                 task->tk_bytes_sent += status;
 552                 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
 553                         req->rq_bytes_sent = 0;
 554                         return 0;
 555                 }
 556
 557                 status = -EAGAIN;
 558                 if (retry++ > XS_SENDMSG_RETRY)
 559                         break;
 560         }
 561
 562         switch (status) {
 563         case -EAGAIN:
 564                 xs_nospace(task);
 565                 break;
 566         case -ECONNREFUSED:
 567         case -ECONNRESET:
 568         case -ENOTCONN:
 569         case -EPIPE:
 570                 status = -ENOTCONN;
 571                 break;
 572         default:
 573                 dprintk("RPC:      sendmsg returned unrecognized error %d\n",
 574                         -status);
 575                 xprt_disconnect(xprt);
 576                 break;
 577         }
 578
 579         return status;
 580 }
 581
 582 /**
 583  * xs_tcp_release_xprt - clean up after a tcp transmission
 584  * @xprt: transport
 585  * @task: rpc task
 586  *
 587  * This cleans up if an error causes us to abort the transmission of a request.
 588  * In this case, the socket may need to be reset in order to avoid confusing
 589  * the server.
 590  */
 591 static void xs_tcp_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 592 {
 593         struct rpc_rqst *req;
 594
 595         if (task != xprt->snd_task)
 596                 return;
 597         if (task == NULL)
 598                 goto out_release;
 599         req = task->tk_rqstp;
 600         if (req->rq_bytes_sent == 0)
 601                 goto out_release;
 602         if (req->rq_bytes_sent == req->rq_snd_buf.len)
 603                 goto out_release;
 604         set_bit(XPRT_CLOSE_WAIT, &task->tk_xprt->state);
 605 out_release:
 606         xprt_release_xprt(xprt, task);
 607 }
 608
 609 /**
 610  * xs_close - close a socket
 611  * @xprt: transport
 612  *
 613  * This is used when all requests are complete; ie, no DRC state remains
 614  * on the server we want to save.
 615  */
 616 static void xs_close(struct rpc_xprt *xprt)
 617 {
 618         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 619         struct socket *sock = transport->sock;
 620         struct sock *sk = transport->inet;
 621
 622         if (!sk)
 623                 goto clear_close_wait;
 624
 625         dprintk("RPC:      xs_close xprt %p\n", xprt);
 626
 627         write_lock_bh(&sk->sk_callback_lock);
 628         transport->inet = NULL;
 629         transport->sock = NULL;
 630
 631         sk->sk_user_data = NULL;
 632         sk->sk_data_ready = transport->old_data_ready;
 633         sk->sk_state_change = transport->old_state_change;
 634         sk->sk_write_space = transport->old_write_space;
 635         write_unlock_bh(&sk->sk_callback_lock);
 636
 637         sk->sk_no_check = 0;
 638
 639         sock_release(sock);
 640 clear_close_wait:
 641         smp_mb__before_clear_bit();
 642         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 643         smp_mb__after_clear_bit();
 644 }
 645
 646 /**
 647  * xs_destroy - prepare to shutdown a transport
 648  * @xprt: doomed transport
 649  *
 650  */
 651 static void xs_destroy(struct rpc_xprt *xprt)
 652 {
 653         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 654
 655         dprintk("RPC:      xs_destroy xprt %p\n", xprt);
 656
 657         cancel_delayed_work(&transport->connect_worker);
 658         flush_scheduled_work();
 659
 660         xprt_disconnect(xprt);
 661         xs_close(xprt);
 662         xs_free_peer_addresses(xprt);
 663         kfree(xprt->slot);
 664         kfree(xprt);
 665 }
 666
 667 static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
 668 {
 669         return (struct rpc_xprt *) sk->sk_user_data;
 670 }
 671
 672 /**
 673  * xs_udp_data_ready - "data ready" callback for UDP sockets
 674  * @sk: socket with data to read
 675  * @len: how much data to read
 676  *
 677  */
 678 static void xs_udp_data_ready(struct sock *sk, int len)
 679 {
 680         struct rpc_task *task;
 681         struct rpc_xprt *xprt;
 682         struct rpc_rqst *rovr;
 683         struct sk_buff *skb;
 684         int err, repsize, copied;
 685         u32 _xid;
 686         __be32 *xp;
 687
 688         read_lock(&sk->sk_callback_lock);
 689         dprintk("RPC:      xs_udp_data_ready...\n");
 690         if (!(xprt = xprt_from_sock(sk)))
 691                 goto out;
 692
 693         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 694                 goto out;
 695
 696         if (xprt->shutdown)
 697                 goto dropit;
 698
 699         repsize = skb->len - sizeof(struct udphdr);
 700         if (repsize < 4) {
 701                 dprintk("RPC:      impossible RPC reply size %d!\n", repsize);
 702                 goto dropit;
 703         }
 704
 705         /* Copy the XID from the skb... */
 706         xp = skb_header_pointer(skb, sizeof(struct udphdr),
 707                                 sizeof(_xid), &_xid);
 708         if (xp == NULL)
 709                 goto dropit;
 710
 711         /* Look up and lock the request corresponding to the given XID */
 712         spin_lock(&xprt->transport_lock);
 713         rovr = xprt_lookup_rqst(xprt, *xp);
 714         if (!rovr)
 715                 goto out_unlock;
 716         task = rovr->rq_task;
 717
 718         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 719                 copied = repsize;
 720
 721         /* Suck it into the iovec, verify checksum if not done by hw. */
 722         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 723                 goto out_unlock;
 724
 725         /* Something worked... */
 726         dst_confirm(skb->dst);
 727
 728         xprt_adjust_cwnd(task, copied);
 729         xprt_update_rtt(task);
 730         xprt_complete_rqst(task, copied);
 731
 732  out_unlock:
 733         spin_unlock(&xprt->transport_lock);
 734  dropit:
 735         skb_free_datagram(sk, skb);
 736  out:
 737         read_unlock(&sk->sk_callback_lock);
 738 }
 739
 740 static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
 741 {
 742         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 743         size_t len, used;
 744         char *p;
 745
 746         p = ((char *) &transport->tcp_fraghdr) + transport->tcp_offset;
 747         len = sizeof(transport->tcp_fraghdr) - transport->tcp_offset;
 748         used = xdr_skb_read_bits(desc, p, len);
 749         transport->tcp_offset += used;
 750         if (used != len)
 751                 return;
 752
 753         transport->tcp_reclen = ntohl(transport->tcp_fraghdr);
 754         if (transport->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
 755                 transport->tcp_flags |= TCP_RCV_LAST_FRAG;
 756         else
 757                 transport->tcp_flags &= ~TCP_RCV_LAST_FRAG;
 758         transport->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
 759
 760         transport->tcp_flags &= ~TCP_RCV_COPY_FRAGHDR;
 761         transport->tcp_offset = 0;
 762
 763         /* Sanity check of the record length */
 764         if (unlikely(transport->tcp_reclen < 4)) {
 765                 dprintk("RPC:      invalid TCP record fragment length\n");
 766                 xprt_disconnect(xprt);
 767                 return;
 768         }
 769         dprintk("RPC:      reading TCP record fragment of length %d\n",
 770                         transport->tcp_reclen);
 771 }
 772
 773 static void xs_tcp_check_fraghdr(struct sock_xprt *transport)
 774 {
 775         if (transport->tcp_offset == transport->tcp_reclen) {
 776                 transport->tcp_flags |= TCP_RCV_COPY_FRAGHDR;
 777                 transport->tcp_offset = 0;
 778                 if (transport->tcp_flags & TCP_RCV_LAST_FRAG) {
 779                         transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 780                         transport->tcp_flags |= TCP_RCV_COPY_XID;
 781                         transport->tcp_copied = 0;
 782                 }
 783         }
 784 }
 785
 786 static inline void xs_tcp_read_xid(struct sock_xprt *transport, struct xdr_skb_reader *desc)
 787 {
 788         size_t len, used;
 789         char *p;
 790
 791         len = sizeof(transport->tcp_xid) - transport->tcp_offset;
 792         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
 793         p = ((char *) &transport->tcp_xid) + transport->tcp_offset;
 794         used = xdr_skb_read_bits(desc, p, len);
 795         transport->tcp_offset += used;
 796         if (used != len)
 797                 return;
 798         transport->tcp_flags &= ~TCP_RCV_COPY_XID;
 799         transport->tcp_flags |= TCP_RCV_COPY_DATA;
 800         transport->tcp_copied = 4;
 801         dprintk("RPC:      reading reply for XID %08x\n",
 802                         ntohl(transport->tcp_xid));
 803         xs_tcp_check_fraghdr(transport);
 804 }
 805
 806 static inline void xs_tcp_read_request(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
 807 {
 808         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 809         struct rpc_rqst *req;
 810         struct xdr_buf *rcvbuf;
 811         size_t len;
 812         ssize_t r;
 813
 814         /* Find and lock the request corresponding to this xid */
 815         spin_lock(&xprt->transport_lock);
 816         req = xprt_lookup_rqst(xprt, transport->tcp_xid);
 817         if (!req) {
 818                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 819                 dprintk("RPC:      XID %08x request not found!\n",
 820                                 ntohl(transport->tcp_xid));
 821                 spin_unlock(&xprt->transport_lock);
 822                 return;
 823         }
 824
 825         rcvbuf = &req->rq_private_buf;
 826         len = desc->count;
 827         if (len > transport->tcp_reclen - transport->tcp_offset) {
 828                 struct xdr_skb_reader my_desc;
 829
 830                 len = transport->tcp_reclen - transport->tcp_offset;
 831                 memcpy(&my_desc, desc, sizeof(my_desc));
 832                 my_desc.count = len;
 833                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
 834                                           &my_desc, xdr_skb_read_bits);
 835                 desc->count -= r;
 836                 desc->offset += r;
 837         } else
 838                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
 839                                           desc, xdr_skb_read_bits);
 840
 841         if (r > 0) {
 842                 transport->tcp_copied += r;
 843                 transport->tcp_offset += r;
 844         }
 845         if (r != len) {
 846                 /* Error when copying to the receive buffer,
 847                  * usually because we weren't able to allocate
 848                  * additional buffer pages. All we can do now
 849                  * is turn off TCP_RCV_COPY_DATA, so the request
 850                  * will not receive any additional updates,
 851                  * and time out.
 852                  * Any remaining data from this record will
 853                  * be discarded.
 854                  */
 855                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 856                 dprintk("RPC:      XID %08x truncated request\n",
 857                                 ntohl(transport->tcp_xid));
 858                 dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
 859                                 xprt, transport->tcp_copied, transport->tcp_offset,
 860                                         transport->tcp_reclen);
 861                 goto out;
 862         }
 863
 864         dprintk("RPC:      XID %08x read %Zd bytes\n",
 865                         ntohl(transport->tcp_xid), r);
 866         dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
 867                         xprt, transport->tcp_copied, transport->tcp_offset,
 868                                 transport->tcp_reclen);
 869
 870         if (transport->tcp_copied == req->rq_private_buf.buflen)
 871                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 872         else if (transport->tcp_offset == transport->tcp_reclen) {
 873                 if (transport->tcp_flags & TCP_RCV_LAST_FRAG)
 874                         transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 875         }
 876
 877 out:
 878         if (!(transport->tcp_flags & TCP_RCV_COPY_DATA))
 879                 xprt_complete_rqst(req->rq_task, transport->tcp_copied);
 880         spin_unlock(&xprt->transport_lock);
 881         xs_tcp_check_fraghdr(transport);
 882 }
 883
 884 static inline void xs_tcp_read_discard(struct sock_xprt *transport, struct xdr_skb_reader *desc)
 885 {
 886         size_t len;
 887
 888         len = transport->tcp_reclen - transport->tcp_offset;
 889         if (len > desc->count)
 890                 len = desc->count;
 891         desc->count -= len;
 892         desc->offset += len;
 893         transport->tcp_offset += len;
 894         dprintk("RPC:      discarded %Zu bytes\n", len);
 895         xs_tcp_check_fraghdr(transport);
 896 }
 897
 898 static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
 899 {
 900         struct rpc_xprt *xprt = rd_desc->arg.data;
 901         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 902         struct xdr_skb_reader desc = {
 903                 .skb    = skb,
 904                 .offset = offset,
 905                 .count  = len,
 906         };
 907
 908         dprintk("RPC:      xs_tcp_data_recv started\n");
 909         do {
 910                 /* Read in a new fragment marker if necessary */
 911                 /* Can we ever really expect to get completely empty fragments? */
 912                 if (transport->tcp_flags & TCP_RCV_COPY_FRAGHDR) {
 913                         xs_tcp_read_fraghdr(xprt, &desc);
 914                         continue;
 915                 }
 916                 /* Read in the xid if necessary */
 917                 if (transport->tcp_flags & TCP_RCV_COPY_XID) {
 918                         xs_tcp_read_xid(transport, &desc);
 919                         continue;
 920                 }
 921                 /* Read in the request data */
 922                 if (transport->tcp_flags & TCP_RCV_COPY_DATA) {
 923                         xs_tcp_read_request(xprt, &desc);
 924                         continue;
 925                 }
 926                 /* Skip over any trailing bytes on short reads */
 927                 xs_tcp_read_discard(transport, &desc);
 928         } while (desc.count);
 929         dprintk("RPC:      xs_tcp_data_recv done\n");
 930         return len - desc.count;
 931 }
 932
 933 /**
 934  * xs_tcp_data_ready - "data ready" callback for TCP sockets
 935  * @sk: socket with data to read
 936  * @bytes: how much data to read
 937  *
 938  */
 939 static void xs_tcp_data_ready(struct sock *sk, int bytes)
 940 {
 941         struct rpc_xprt *xprt;
 942         read_descriptor_t rd_desc;
 943
 944         read_lock(&sk->sk_callback_lock);
 945         dprintk("RPC:      xs_tcp_data_ready...\n");
 946         if (!(xprt = xprt_from_sock(sk)))
 947                 goto out;
 948         if (xprt->shutdown)
 949                 goto out;
 950
 951         /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
 952         rd_desc.arg.data = xprt;
 953         rd_desc.count = 65536;
 954         tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
 955 out:
 956         read_unlock(&sk->sk_callback_lock);
 957 }
 958
 959 /**
 960  * xs_tcp_state_change - callback to handle TCP socket state changes
 961  * @sk: socket whose state has changed
 962  *
 963  */
 964 static void xs_tcp_state_change(struct sock *sk)
 965 {
 966         struct rpc_xprt *xprt;
 967
 968         read_lock(&sk->sk_callback_lock);
 969         if (!(xprt = xprt_from_sock(sk)))
 970                 goto out;
 971         dprintk("RPC:      xs_tcp_state_change client %p...\n", xprt);
 972         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
 973                                 sk->sk_state, xprt_connected(xprt),
 974                                 sock_flag(sk, SOCK_DEAD),
 975                                 sock_flag(sk, SOCK_ZAPPED));
 976
 977         switch (sk->sk_state) {
 978         case TCP_ESTABLISHED:
 979                 spin_lock_bh(&xprt->transport_lock);
 980                 if (!xprt_test_and_set_connected(xprt)) {
 981                         struct sock_xprt *transport = container_of(xprt,
 982                                         struct sock_xprt, xprt);
 983
 984                         /* Reset TCP record info */
 985                         transport->tcp_offset = 0;
 986                         transport->tcp_reclen = 0;
 987                         transport->tcp_copied = 0;
 988                         transport->tcp_flags =
 989                                 TCP_RCV_COPY_FRAGHDR | TCP_RCV_COPY_XID;
 990
 991                         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
 992                         xprt_wake_pending_tasks(xprt, 0);
 993                 }
 994                 spin_unlock_bh(&xprt->transport_lock);
 995                 break;
 996         case TCP_SYN_SENT:
 997         case TCP_SYN_RECV:
 998                 break;
 999         case TCP_CLOSE_WAIT:
1000                 /* Try to schedule an autoclose RPC calls */
1001                 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
1002                 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
1003                         schedule_work(&xprt->task_cleanup);
1004         default:
1005                 xprt_disconnect(xprt);
1006         }
1007  out:
1008         read_unlock(&sk->sk_callback_lock);
1009 }
1010
1011 /**
1012  * xs_udp_write_space - callback invoked when socket buffer space
1013  *                             becomes available
1014  * @sk: socket whose state has changed
1015  *
1016  * Called when more output buffer space is available for this socket.
1017  * We try not to wake our writers until they can make "significant"
1018  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1019  * with a bunch of small requests.
1020  */
1021 static void xs_udp_write_space(struct sock *sk)
1022 {
1023         read_lock(&sk->sk_callback_lock);
1024
1025         /* from net/core/sock.c:sock_def_write_space */
1026         if (sock_writeable(sk)) {
1027                 struct socket *sock;
1028                 struct rpc_xprt *xprt;
1029
1030                 if (unlikely(!(sock = sk->sk_socket)))
1031                         goto out;
1032                 if (unlikely(!(xprt = xprt_from_sock(sk))))
1033                         goto out;
1034                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
1035                         goto out;
1036
1037                 xprt_write_space(xprt);
1038         }
1039
1040  out:
1041         read_unlock(&sk->sk_callback_lock);
1042 }
1043
1044 /**
1045  * xs_tcp_write_space - callback invoked when socket buffer space
1046  *                             becomes available
1047  * @sk: socket whose state has changed
1048  *
1049  * Called when more output buffer space is available for this socket.
1050  * We try not to wake our writers until they can make "significant"
1051  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1052  * with a bunch of small requests.
1053  */
1054 static void xs_tcp_write_space(struct sock *sk)
1055 {
1056         read_lock(&sk->sk_callback_lock);
1057
1058         /* from net/core/stream.c:sk_stream_write_space */
1059         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
1060                 struct socket *sock;
1061                 struct rpc_xprt *xprt;
1062
1063                 if (unlikely(!(sock = sk->sk_socket)))
1064                         goto out;
1065                 if (unlikely(!(xprt = xprt_from_sock(sk))))
1066                         goto out;
1067                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
1068                         goto out;
1069
1070                 xprt_write_space(xprt);
1071         }
1072
1073  out:
1074         read_unlock(&sk->sk_callback_lock);
1075 }
1076
1077 static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
1078 {
1079         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1080         struct sock *sk = transport->inet;
1081
1082         if (transport->rcvsize) {
1083                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1084                 sk->sk_rcvbuf = transport->rcvsize * xprt->max_reqs * 2;
1085         }
1086         if (transport->sndsize) {
1087                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1088                 sk->sk_sndbuf = transport->sndsize * xprt->max_reqs * 2;
1089                 sk->sk_write_space(sk);
1090         }
1091 }
1092
1093 /**
1094  * xs_udp_set_buffer_size - set send and receive limits
1095  * @xprt: generic transport
1096  * @sndsize: requested size of send buffer, in bytes
1097  * @rcvsize: requested size of receive buffer, in bytes
1098  *
1099  * Set socket send and receive buffer size limits.
1100  */
1101 static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
1102 {
1103         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1104
1105         transport->sndsize = 0;
1106         if (sndsize)
1107                 transport->sndsize = sndsize + 1024;
1108         transport->rcvsize = 0;
1109         if (rcvsize)
1110                 transport->rcvsize = rcvsize + 1024;
1111
1112         xs_udp_do_set_buffer_size(xprt);
1113 }
1114
1115 /**
1116  * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
1117  * @task: task that timed out
1118  *
1119  * Adjust the congestion window after a retransmit timeout has occurred.
1120  */
1121 static void xs_udp_timer(struct rpc_task *task)
1122 {
1123         xprt_adjust_cwnd(task, -ETIMEDOUT);
1124 }
1125
1126 static unsigned short xs_get_random_port(void)
1127 {
1128         unsigned short range = xprt_max_resvport - xprt_min_resvport;
1129         unsigned short rand = (unsigned short) net_random() % range;
1130         return rand + xprt_min_resvport;
1131 }
1132
1133 /**
1134  * xs_set_port - reset the port number in the remote endpoint address
1135  * @xprt: generic transport
1136  * @port: new port number
1137  *
1138  */
1139 static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
1140 {
1141         struct sockaddr_in *sap = (struct sockaddr_in *) &xprt->addr;
1142
1143         dprintk("RPC:      setting port for xprt %p to %u\n", xprt, port);
1144
1145         sap->sin_port = htons(port);
1146 }
1147
1148 static int xs_bindresvport(struct sock_xprt *transport, struct socket *sock)
1149 {
1150         struct sockaddr_in myaddr = {
1151                 .sin_family = AF_INET,
1152         };
1153         int err;
1154         unsigned short port = transport->port;
1155
1156         do {
1157                 myaddr.sin_port = htons(port);
1158                 err = kernel_bind(sock, (struct sockaddr *) &myaddr,
1159                                                 sizeof(myaddr));
1160                 if (err == 0) {
1161                         transport->port = port;
1162                         dprintk("RPC:      xs_bindresvport bound to port %u\n",
1163                                         port);
1164                         return 0;
1165                 }
1166                 if (port <= xprt_min_resvport)
1167                         port = xprt_max_resvport;
1168                 else
1169                         port--;
1170         } while (err == -EADDRINUSE && port != transport->port);
1171
1172         dprintk("RPC:      can't bind to reserved port (%d).\n", -err);
1173         return err;
1174 }
1175
1176 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1177 static struct lock_class_key xs_key[2];
1178 static struct lock_class_key xs_slock_key[2];
1179
1180 static inline void xs_reclassify_socket(struct socket *sock)
1181 {
1182         struct sock *sk = sock->sk;
1183         BUG_ON(sk->sk_lock.owner != NULL);
1184         switch (sk->sk_family) {
1185         case AF_INET:
1186                 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFS",
1187                         &xs_slock_key[0], "sk_lock-AF_INET-NFS", &xs_key[0]);
1188                 break;
1189
1190         case AF_INET6:
1191                 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFS",
1192                         &xs_slock_key[1], "sk_lock-AF_INET6-NFS", &xs_key[1]);
1193                 break;
1194
1195         default:
1196                 BUG();
1197         }
1198 }
1199 #else
1200 static inline void xs_reclassify_socket(struct socket *sock)
1201 {
1202 }
1203 #endif
1204
1205 /**
1206  * xs_udp_connect_worker - set up a UDP socket
1207  * @work: RPC transport to connect
1208  *
1209  * Invoked by a work queue tasklet.
1210  */
1211 static void xs_udp_connect_worker(struct work_struct *work)
1212 {
1213         struct sock_xprt *transport =
1214                 container_of(work, struct sock_xprt, connect_worker.work);
1215         struct rpc_xprt *xprt = &transport->xprt;
1216         struct socket *sock = transport->sock;
1217         int err, status = -EIO;
1218
1219         if (xprt->shutdown || !xprt_bound(xprt))
1220                 goto out;
1221
1222         /* Start by resetting any existing state */
1223         xs_close(xprt);
1224
1225         if ((err = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock)) < 0) {
1226                 dprintk("RPC:      can't create UDP transport socket (%d).\n", -err);
1227                 goto out;
1228         }
1229         xs_reclassify_socket(sock);
1230
1231         if (xprt->resvport && xs_bindresvport(transport, sock) < 0) {
1232                 sock_release(sock);
1233                 goto out;
1234         }
1235
1236         dprintk("RPC:      worker connecting xprt %p to address: %s\n",
1237                         xprt, xprt->address_strings[RPC_DISPLAY_ALL]);
1238
1239         if (!transport->inet) {
1240                 struct sock *sk = sock->sk;
1241
1242                 write_lock_bh(&sk->sk_callback_lock);
1243
1244                 sk->sk_user_data = xprt;
1245                 transport->old_data_ready = sk->sk_data_ready;
1246                 transport->old_state_change = sk->sk_state_change;
1247                 transport->old_write_space = sk->sk_write_space;
1248                 sk->sk_data_ready = xs_udp_data_ready;
1249                 sk->sk_write_space = xs_udp_write_space;
1250                 sk->sk_no_check = UDP_CSUM_NORCV;
1251                 sk->sk_allocation = GFP_ATOMIC;
1252
1253                 xprt_set_connected(xprt);
1254
1255                 /* Reset to new socket */
1256                 transport->sock = sock;
1257                 transport->inet = sk;
1258
1259                 write_unlock_bh(&sk->sk_callback_lock);
1260         }
1261         xs_udp_do_set_buffer_size(xprt);
1262         status = 0;
1263 out:
1264         xprt_wake_pending_tasks(xprt, status);
1265         xprt_clear_connecting(xprt);
1266 }
1267
1268 /*
1269  * We need to preserve the port number so the reply cache on the server can
1270  * find our cached RPC replies when we get around to reconnecting.
1271  */
1272 static void xs_tcp_reuse_connection(struct rpc_xprt *xprt)
1273 {
1274         int result;
1275         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1276         struct sockaddr any;
1277
1278         dprintk("RPC:      disconnecting xprt %p to reuse port\n", xprt);
1279
1280         /*
1281          * Disconnect the transport socket by doing a connect operation
1282          * with AF_UNSPEC.  This should return immediately...
1283          */
1284         memset(&any, 0, sizeof(any));
1285         any.sa_family = AF_UNSPEC;
1286         result = kernel_connect(transport->sock, &any, sizeof(any), 0);
1287         if (result)
1288                 dprintk("RPC:      AF_UNSPEC connect return code %d\n",
1289                                 result);
1290 }
1291
1292 /**
1293  * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1294  * @work: RPC transport to connect
1295  *
1296  * Invoked by a work queue tasklet.
1297  */
1298 static void xs_tcp_connect_worker(struct work_struct *work)
1299 {
1300         struct sock_xprt *transport =
1301                 container_of(work, struct sock_xprt, connect_worker.work);
1302         struct rpc_xprt *xprt = &transport->xprt;
1303         struct socket *sock = transport->sock;
1304         int err, status = -EIO;
1305
1306         if (xprt->shutdown || !xprt_bound(xprt))
1307                 goto out;
1308
1309         if (!sock) {
1310                 /* start from scratch */
1311                 if ((err = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock)) < 0) {
1312                         dprintk("RPC:      can't create TCP transport socket (%d).\n", -err);
1313                         goto out;
1314                 }
1315                 xs_reclassify_socket(sock);
1316
1317                 if (xprt->resvport && xs_bindresvport(transport, sock) < 0) {
1318                         sock_release(sock);
1319                         goto out;
1320                 }
1321         } else
1322                 /* "close" the socket, preserving the local port */
1323                 xs_tcp_reuse_connection(xprt);
1324
1325         dprintk("RPC:      worker connecting xprt %p to address: %s\n",
1326                         xprt, xprt->address_strings[RPC_DISPLAY_ALL]);
1327
1328         if (!transport->inet) {
1329                 struct sock *sk = sock->sk;
1330
1331                 write_lock_bh(&sk->sk_callback_lock);
1332
1333                 sk->sk_user_data = xprt;
1334                 transport->old_data_ready = sk->sk_data_ready;
1335                 transport->old_state_change = sk->sk_state_change;
1336                 transport->old_write_space = sk->sk_write_space;
1337                 sk->sk_data_ready = xs_tcp_data_ready;
1338                 sk->sk_state_change = xs_tcp_state_change;
1339                 sk->sk_write_space = xs_tcp_write_space;
1340                 sk->sk_allocation = GFP_ATOMIC;
1341
1342                 /* socket options */
1343                 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
1344                 sock_reset_flag(sk, SOCK_LINGER);
1345                 tcp_sk(sk)->linger2 = 0;
1346                 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1347
1348                 xprt_clear_connected(xprt);
1349
1350                 /* Reset to new socket */
1351                 transport->sock = sock;
1352                 transport->inet = sk;
1353
1354                 write_unlock_bh(&sk->sk_callback_lock);
1355         }
1356
1357         /* Tell the socket layer to start connecting... */
1358         xprt->stat.connect_count++;
1359         xprt->stat.connect_start = jiffies;
1360         status = kernel_connect(sock, (struct sockaddr *) &xprt->addr,
1361                         xprt->addrlen, O_NONBLOCK);
1362         dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
1363                         xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
1364         if (status < 0) {
1365                 switch (status) {
1366                         case -EINPROGRESS:
1367                         case -EALREADY:
1368                                 goto out_clear;
1369                         case -ECONNREFUSED:
1370                         case -ECONNRESET:
1371                                 /* retry with existing socket, after a delay */
1372                                 break;
1373                         default:
1374                                 /* get rid of existing socket, and retry */
1375                                 xs_close(xprt);
1376                                 break;
1377                 }
1378         }
1379 out:
1380         xprt_wake_pending_tasks(xprt, status);
1381 out_clear:
1382         xprt_clear_connecting(xprt);
1383 }
1384
1385 /**
1386  * xs_connect - connect a socket to a remote endpoint
1387  * @task: address of RPC task that manages state of connect request
1388  *
1389  * TCP: If the remote end dropped the connection, delay reconnecting.
1390  *
1391  * UDP socket connects are synchronous, but we use a work queue anyway
1392  * to guarantee that even unprivileged user processes can set up a
1393  * socket on a privileged port.
1394  *
1395  * If a UDP socket connect fails, the delay behavior here prevents
1396  * retry floods (hard mounts).
1397  */
1398 static void xs_connect(struct rpc_task *task)
1399 {
1400         struct rpc_xprt *xprt = task->tk_xprt;
1401         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1402
1403         if (xprt_test_and_set_connecting(xprt))
1404                 return;
1405
1406         if (transport->sock != NULL) {
1407                 dprintk("RPC:      xs_connect delayed xprt %p for %lu seconds\n",
1408                                 xprt, xprt->reestablish_timeout / HZ);
1409                 schedule_delayed_work(&transport->connect_worker,
1410                                         xprt->reestablish_timeout);
1411                 xprt->reestablish_timeout <<= 1;
1412                 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
1413                         xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
1414         } else {
1415                 dprintk("RPC:      xs_connect scheduled xprt %p\n", xprt);
1416                 schedule_delayed_work(&transport->connect_worker, 0);
1417
1418                 /* flush_scheduled_work can sleep... */
1419                 if (!RPC_IS_ASYNC(task))
1420                         flush_scheduled_work();
1421         }
1422 }
1423
1424 /**
1425  * xs_udp_print_stats - display UDP socket-specifc stats
1426  * @xprt: rpc_xprt struct containing statistics
1427  * @seq: output file
1428  *
1429  */
1430 static void xs_udp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1431 {
1432         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1433
1434         seq_printf(seq, "\txprt:\tudp %u %lu %lu %lu %lu %Lu %Lu\n",
1435                         transport->port,
1436                         xprt->stat.bind_count,
1437                         xprt->stat.sends,
1438                         xprt->stat.recvs,
1439                         xprt->stat.bad_xids,
1440                         xprt->stat.req_u,
1441                         xprt->stat.bklog_u);
1442 }
1443
1444 /**
1445  * xs_tcp_print_stats - display TCP socket-specifc stats
1446  * @xprt: rpc_xprt struct containing statistics
1447  * @seq: output file
1448  *
1449  */
1450 static void xs_tcp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1451 {
1452         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1453         long idle_time = 0;
1454
1455         if (xprt_connected(xprt))
1456                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
1457
1458         seq_printf(seq, "\txprt:\ttcp %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu\n",
1459                         transport->port,
1460                         xprt->stat.bind_count,
1461                         xprt->stat.connect_count,
1462                         xprt->stat.connect_time,
1463                         idle_time,
1464                         xprt->stat.sends,
1465                         xprt->stat.recvs,
1466                         xprt->stat.bad_xids,
1467                         xprt->stat.req_u,
1468                         xprt->stat.bklog_u);
1469 }
1470
1471 static struct rpc_xprt_ops xs_udp_ops = {
1472         .set_buffer_size        = xs_udp_set_buffer_size,
1473         .reserve_xprt           = xprt_reserve_xprt_cong,
1474         .release_xprt           = xprt_release_xprt_cong,
1475         .rpcbind                = rpc_getport,
1476         .set_port               = xs_set_port,
1477         .connect                = xs_connect,
1478         .buf_alloc              = rpc_malloc,
1479         .buf_free               = rpc_free,
1480         .send_request           = xs_udp_send_request,
1481         .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
1482         .timer                  = xs_udp_timer,
1483         .release_request        = xprt_release_rqst_cong,
1484         .close                  = xs_close,
1485         .destroy                = xs_destroy,
1486         .print_stats            = xs_udp_print_stats,
1487 };
1488
1489 static struct rpc_xprt_ops xs_tcp_ops = {
1490         .reserve_xprt           = xprt_reserve_xprt,
1491         .release_xprt           = xs_tcp_release_xprt,
1492         .rpcbind                = rpc_getport,
1493         .set_port               = xs_set_port,
1494         .connect                = xs_connect,
1495         .buf_alloc              = rpc_malloc,
1496         .buf_free               = rpc_free,
1497         .send_request           = xs_tcp_send_request,
1498         .set_retrans_timeout    = xprt_set_retrans_timeout_def,
1499         .close                  = xs_close,
1500         .destroy                = xs_destroy,
1501         .print_stats            = xs_tcp_print_stats,
1502 };
1503
1504 static struct rpc_xprt *xs_setup_xprt(struct sockaddr *addr, size_t addrlen, unsigned int slot_table_size)
1505 {
1506         struct rpc_xprt *xprt;
1507         struct sock_xprt *new;
1508
1509         if (addrlen > sizeof(xprt->addr)) {
1510                 dprintk("RPC:      xs_setup_xprt: address too large\n");
1511                 return ERR_PTR(-EBADF);
1512         }
1513
1514         new = kzalloc(sizeof(*new), GFP_KERNEL);
1515         if (new == NULL) {
1516                 dprintk("RPC:      xs_setup_xprt: couldn't allocate rpc_xprt\n");
1517                 return ERR_PTR(-ENOMEM);
1518         }
1519         xprt = &new->xprt;
1520
1521         xprt->max_reqs = slot_table_size;
1522         xprt->slot = kcalloc(xprt->max_reqs, sizeof(struct rpc_rqst), GFP_KERNEL);
1523         if (xprt->slot == NULL) {
1524                 kfree(xprt);
1525                 dprintk("RPC:      xs_setup_xprt: couldn't allocate slot table\n");
1526                 return ERR_PTR(-ENOMEM);
1527         }
1528
1529         memcpy(&xprt->addr, addr, addrlen);
1530         xprt->addrlen = addrlen;
1531         new->port = xs_get_random_port();
1532
1533         return xprt;
1534 }
1535
1536 /**
1537  * xs_setup_udp - Set up transport to use a UDP socket
1538  * @addr: address of remote server
1539  * @addrlen: length of address in bytes
1540  * @to:   timeout parameters
1541  *
1542  */
1543 struct rpc_xprt *xs_setup_udp(struct sockaddr *addr, size_t addrlen, struct rpc_timeout *to)
1544 {
1545         struct rpc_xprt *xprt;
1546         struct sock_xprt *transport;
1547
1548         xprt = xs_setup_xprt(addr, addrlen, xprt_udp_slot_table_entries);
1549         if (IS_ERR(xprt))
1550                 return xprt;
1551         transport = container_of(xprt, struct sock_xprt, xprt);
1552
1553         if (ntohs(((struct sockaddr_in *)addr)->sin_port) != 0)
1554                 xprt_set_bound(xprt);
1555
1556         xprt->prot = IPPROTO_UDP;
1557         xprt->tsh_size = 0;
1558         /* XXX: header size can vary due to auth type, IPv6, etc. */
1559         xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1560
1561         INIT_DELAYED_WORK(&transport->connect_worker, xs_udp_connect_worker);
1562         xprt->bind_timeout = XS_BIND_TO;
1563         xprt->connect_timeout = XS_UDP_CONN_TO;
1564         xprt->reestablish_timeout = XS_UDP_REEST_TO;
1565         xprt->idle_timeout = XS_IDLE_DISC_TO;
1566
1567         xprt->ops = &xs_udp_ops;
1568
1569         if (to)
1570                 xprt->timeout = *to;
1571         else
1572                 xprt_set_timeout(&xprt->timeout, 5, 5 * HZ);
1573
1574         xs_format_peer_addresses(xprt);
1575         dprintk("RPC:      set up transport to address %s\n",
1576                         xprt->address_strings[RPC_DISPLAY_ALL]);
1577
1578         return xprt;
1579 }
1580
1581 /**
1582  * xs_setup_tcp - Set up transport to use a TCP socket
1583  * @addr: address of remote server
1584  * @addrlen: length of address in bytes
1585  * @to: timeout parameters
1586  *
1587  */
1588 struct rpc_xprt *xs_setup_tcp(struct sockaddr *addr, size_t addrlen, struct rpc_timeout *to)
1589 {
1590         struct rpc_xprt *xprt;
1591         struct sock_xprt *transport;
1592
1593         xprt = xs_setup_xprt(addr, addrlen, xprt_tcp_slot_table_entries);
1594         if (IS_ERR(xprt))
1595                 return xprt;
1596         transport = container_of(xprt, struct sock_xprt, xprt);
1597
1598         if (ntohs(((struct sockaddr_in *)addr)->sin_port) != 0)
1599                 xprt_set_bound(xprt);
1600
1601         xprt->prot = IPPROTO_TCP;
1602         xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
1603         xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
1604
1605         INIT_DELAYED_WORK(&transport->connect_worker, xs_tcp_connect_worker);
1606         xprt->bind_timeout = XS_BIND_TO;
1607         xprt->connect_timeout = XS_TCP_CONN_TO;
1608         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1609         xprt->idle_timeout = XS_IDLE_DISC_TO;
1610
1611         xprt->ops = &xs_tcp_ops;
1612
1613         if (to)
1614                 xprt->timeout = *to;
1615         else
1616                 xprt_set_timeout(&xprt->timeout, 2, 60 * HZ);
1617
1618         xs_format_peer_addresses(xprt);
1619         dprintk("RPC:      set up transport to address %s\n",
1620                         xprt->address_strings[RPC_DISPLAY_ALL]);
1621
1622         return xprt;
1623 }
1624
1625 /**
1626  * init_socket_xprt - set up xprtsock's sysctls
1627  *
1628  */
1629 int init_socket_xprt(void)
1630 {
1631 #ifdef RPC_DEBUG
1632         if (!sunrpc_table_header) {
1633                 sunrpc_table_header = register_sysctl_table(sunrpc_table, 1);
1634 #ifdef CONFIG_PROC_FS
1635                 if (sunrpc_table[0].de)
1636                         sunrpc_table[0].de->owner = THIS_MODULE;
1637 #endif
1638         }
1639 #endif
1640
1641         return 0;
1642 }
1643
1644 /**
1645  * cleanup_socket_xprt - remove xprtsock's sysctls
1646  *
1647  */
1648 void cleanup_socket_xprt(void)
1649 {
1650 #ifdef RPC_DEBUG
1651         if (sunrpc_table_header) {
1652                 unregister_sysctl_table(sunrpc_table_header);
1653                 sunrpc_table_header = NULL;
1654         }
1655 #endif
1656 }