SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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/pagemap.h>
  22 #include <linux/errno.h>
  23 #include <linux/socket.h>
  24 #include <linux/in.h>
  25 #include <linux/net.h>
  26 #include <linux/mm.h>
  27 #include <linux/udp.h>
  28 #include <linux/tcp.h>
  29 #include <linux/sunrpc/clnt.h>
  30 #include <linux/sunrpc/sched.h>
  31 #include <linux/file.h>
  32
  33 #include <net/sock.h>
  34 #include <net/checksum.h>
  35 #include <net/udp.h>
  36 #include <net/tcp.h>
  37
  38 /*
  39  * xprtsock tunables
  40  */
  41 unsigned int xprt_udp_slot_table_entries = RPC_DEF_SLOT_TABLE;
  42 unsigned int xprt_tcp_slot_table_entries = RPC_DEF_SLOT_TABLE;
  43
  44 unsigned int xprt_min_resvport = RPC_DEF_MIN_RESVPORT;
  45 unsigned int xprt_max_resvport = RPC_DEF_MAX_RESVPORT;
  46
  47 /*
  48  * We can register our own files under /proc/sys/sunrpc by
  49  * calling register_sysctl_table() again.  The files in that
  50  * directory become the union of all files registered there.
  51  *
  52  * We simply need to make sure that we don't collide with
  53  * someone else's file names!
  54  */
  55
  56 #ifdef RPC_DEBUG
  57
  58 static unsigned int min_slot_table_size = RPC_MIN_SLOT_TABLE;
  59 static unsigned int max_slot_table_size = RPC_MAX_SLOT_TABLE;
  60 static unsigned int xprt_min_resvport_limit = RPC_MIN_RESVPORT;
  61 static unsigned int xprt_max_resvport_limit = RPC_MAX_RESVPORT;
  62
  63 static struct ctl_table_header *sunrpc_table_header;
  64
  65 /*
  66  * FIXME: changing the UDP slot table size should also resize the UDP
  67  *        socket buffers for existing UDP transports
  68  */
  69 static ctl_table xs_tunables_table[] = {
  70         {
  71                 .ctl_name       = CTL_SLOTTABLE_UDP,
  72                 .procname       = "udp_slot_table_entries",
  73                 .data           = &xprt_udp_slot_table_entries,
  74                 .maxlen         = sizeof(unsigned int),
  75                 .mode           = 0644,
  76                 .proc_handler   = &proc_dointvec_minmax,
  77                 .strategy       = &sysctl_intvec,
  78                 .extra1         = &min_slot_table_size,
  79                 .extra2         = &max_slot_table_size
  80         },
  81         {
  82                 .ctl_name       = CTL_SLOTTABLE_TCP,
  83                 .procname       = "tcp_slot_table_entries",
  84                 .data           = &xprt_tcp_slot_table_entries,
  85                 .maxlen         = sizeof(unsigned int),
  86                 .mode           = 0644,
  87                 .proc_handler   = &proc_dointvec_minmax,
  88                 .strategy       = &sysctl_intvec,
  89                 .extra1         = &min_slot_table_size,
  90                 .extra2         = &max_slot_table_size
  91         },
  92         {
  93                 .ctl_name       = CTL_MIN_RESVPORT,
  94                 .procname       = "min_resvport",
  95                 .data           = &xprt_min_resvport,
  96                 .maxlen         = sizeof(unsigned int),
  97                 .mode           = 0644,
  98                 .proc_handler   = &proc_dointvec_minmax,
  99                 .strategy       = &sysctl_intvec,
 100                 .extra1         = &xprt_min_resvport_limit,
 101                 .extra2         = &xprt_max_resvport_limit
 102         },
 103         {
 104                 .ctl_name       = CTL_MAX_RESVPORT,
 105                 .procname       = "max_resvport",
 106                 .data           = &xprt_max_resvport,
 107                 .maxlen         = sizeof(unsigned int),
 108                 .mode           = 0644,
 109                 .proc_handler   = &proc_dointvec_minmax,
 110                 .strategy       = &sysctl_intvec,
 111                 .extra1         = &xprt_min_resvport_limit,
 112                 .extra2         = &xprt_max_resvport_limit
 113         },
 114         {
 115                 .ctl_name = 0,
 116         },
 117 };
 118
 119 static ctl_table sunrpc_table[] = {
 120         {
 121                 .ctl_name       = CTL_SUNRPC,
 122                 .procname       = "sunrpc",
 123                 .mode           = 0555,
 124                 .child          = xs_tunables_table
 125         },
 126         {
 127                 .ctl_name = 0,
 128         },
 129 };
 130
 131 #endif
 132
 133 /*
 134  * How many times to try sending a request on a socket before waiting
 135  * for the socket buffer to clear.
 136  */
 137 #define XS_SENDMSG_RETRY        (10U)
 138
 139 /*
 140  * Time out for an RPC UDP socket connect.  UDP socket connects are
 141  * synchronous, but we set a timeout anyway in case of resource
 142  * exhaustion on the local host.
 143  */
 144 #define XS_UDP_CONN_TO          (5U * HZ)
 145
 146 /*
 147  * Wait duration for an RPC TCP connection to be established.  Solaris
 148  * NFS over TCP uses 60 seconds, for example, which is in line with how
 149  * long a server takes to reboot.
 150  */
 151 #define XS_TCP_CONN_TO          (60U * HZ)
 152
 153 /*
 154  * Wait duration for a reply from the RPC portmapper.
 155  */
 156 #define XS_BIND_TO              (60U * HZ)
 157
 158 /*
 159  * Delay if a UDP socket connect error occurs.  This is most likely some
 160  * kind of resource problem on the local host.
 161  */
 162 #define XS_UDP_REEST_TO         (2U * HZ)
 163
 164 /*
 165  * The reestablish timeout allows clients to delay for a bit before attempting
 166  * to reconnect to a server that just dropped our connection.
 167  *
 168  * We implement an exponential backoff when trying to reestablish a TCP
 169  * transport connection with the server.  Some servers like to drop a TCP
 170  * connection when they are overworked, so we start with a short timeout and
 171  * increase over time if the server is down or not responding.
 172  */
 173 #define XS_TCP_INIT_REEST_TO    (3U * HZ)
 174 #define XS_TCP_MAX_REEST_TO     (5U * 60 * HZ)
 175
 176 /*
 177  * TCP idle timeout; client drops the transport socket if it is idle
 178  * for this long.  Note that we also timeout UDP sockets to prevent
 179  * holding port numbers when there is no RPC traffic.
 180  */
 181 #define XS_IDLE_DISC_TO         (5U * 60 * HZ)
 182
 183 #ifdef RPC_DEBUG
 184 # undef  RPC_DEBUG_DATA
 185 # define RPCDBG_FACILITY        RPCDBG_TRANS
 186 #endif
 187
 188 #ifdef RPC_DEBUG_DATA
 189 static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
 190 {
 191         u8 *buf = (u8 *) packet;
 192         int j;
 193
 194         dprintk("RPC:       %s\n", msg);
 195         for (j = 0; j < count && j < 128; j += 4) {
 196                 if (!(j & 31)) {
 197                         if (j)
 198                                 dprintk("\n");
 199                         dprintk("0x%04x ", j);
 200                 }
 201                 dprintk("%02x%02x%02x%02x ",
 202                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
 203         }
 204         dprintk("\n");
 205 }
 206 #else
 207 static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
 208 {
 209         /* NOP */
 210 }
 211 #endif
 212
 213 struct sock_xprt {
 214         struct rpc_xprt         xprt;
 215
 216         /*
 217          * Network layer
 218          */
 219         struct socket *         sock;
 220         struct sock *           inet;
 221
 222         /*
 223          * State of TCP reply receive
 224          */
 225         __be32                  tcp_fraghdr,
 226                                 tcp_xid;
 227
 228         u32                     tcp_offset,
 229                                 tcp_reclen;
 230
 231         unsigned long           tcp_copied,
 232                                 tcp_flags;
 233
 234         /*
 235          * Connection of transports
 236          */
 237         struct delayed_work     connect_worker;
 238         struct sockaddr_storage addr;
 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, NIPQUAD_FMT,
 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         buf = kzalloc(8, GFP_KERNEL);
 283         if (buf) {
 284                 if (xprt->prot == IPPROTO_UDP)
 285                         snprintf(buf, 8, "udp");
 286                 else
 287                         snprintf(buf, 8, "tcp");
 288         }
 289         xprt->address_strings[RPC_DISPLAY_PROTO] = buf;
 290
 291         buf = kzalloc(48, GFP_KERNEL);
 292         if (buf) {
 293                 snprintf(buf, 48, "addr="NIPQUAD_FMT" port=%u proto=%s",
 294                         NIPQUAD(addr->sin_addr.s_addr),
 295                         ntohs(addr->sin_port),
 296                         xprt->prot == IPPROTO_UDP ? "udp" : "tcp");
 297         }
 298         xprt->address_strings[RPC_DISPLAY_ALL] = buf;
 299
 300         buf = kzalloc(10, GFP_KERNEL);
 301         if (buf) {
 302                 snprintf(buf, 10, "%02x%02x%02x%02x",
 303                                 NIPQUAD(addr->sin_addr.s_addr));
 304         }
 305         xprt->address_strings[RPC_DISPLAY_HEX_ADDR] = buf;
 306
 307         buf = kzalloc(8, GFP_KERNEL);
 308         if (buf) {
 309                 snprintf(buf, 8, "%4hx",
 310                                 ntohs(addr->sin_port));
 311         }
 312         xprt->address_strings[RPC_DISPLAY_HEX_PORT] = buf;
 313 }
 314
 315 static void xs_free_peer_addresses(struct rpc_xprt *xprt)
 316 {
 317         int i;
 318
 319         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 320                 kfree(xprt->address_strings[i]);
 321 }
 322
 323 #define XS_SENDMSG_FLAGS        (MSG_DONTWAIT | MSG_NOSIGNAL)
 324
 325 static int xs_send_kvec(struct socket *sock, struct sockaddr *addr, int addrlen, struct kvec *vec, unsigned int base, int more)
 326 {
 327         struct msghdr msg = {
 328                 .msg_name       = addr,
 329                 .msg_namelen    = addrlen,
 330                 .msg_flags      = XS_SENDMSG_FLAGS | (more ? MSG_MORE : 0),
 331         };
 332         struct kvec iov = {
 333                 .iov_base       = vec->iov_base + base,
 334                 .iov_len        = vec->iov_len - base,
 335         };
 336
 337         if (iov.iov_len != 0)
 338                 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
 339         return kernel_sendmsg(sock, &msg, NULL, 0, 0);
 340 }
 341
 342 static int xs_send_pagedata(struct socket *sock, struct xdr_buf *xdr, unsigned int base, int more)
 343 {
 344         struct page **ppage;
 345         unsigned int remainder;
 346         int err, sent = 0;
 347
 348         remainder = xdr->page_len - base;
 349         base += xdr->page_base;
 350         ppage = xdr->pages + (base >> PAGE_SHIFT);
 351         base &= ~PAGE_MASK;
 352         for(;;) {
 353                 unsigned int len = min_t(unsigned int, PAGE_SIZE - base, remainder);
 354                 int flags = XS_SENDMSG_FLAGS;
 355
 356                 remainder -= len;
 357                 if (remainder != 0 || more)
 358                         flags |= MSG_MORE;
 359                 err = sock->ops->sendpage(sock, *ppage, base, len, flags);
 360                 if (remainder == 0 || err != len)
 361                         break;
 362                 sent += err;
 363                 ppage++;
 364                 base = 0;
 365         }
 366         if (sent == 0)
 367                 return err;
 368         if (err > 0)
 369                 sent += err;
 370         return sent;
 371 }
 372
 373 /**
 374  * xs_sendpages - write pages directly to a socket
 375  * @sock: socket to send on
 376  * @addr: UDP only -- address of destination
 377  * @addrlen: UDP only -- length of destination address
 378  * @xdr: buffer containing this request
 379  * @base: starting position in the buffer
 380  *
 381  */
 382 static int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
 383 {
 384         unsigned int remainder = xdr->len - base;
 385         int err, sent = 0;
 386
 387         if (unlikely(!sock))
 388                 return -ENOTCONN;
 389
 390         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 391         if (base != 0) {
 392                 addr = NULL;
 393                 addrlen = 0;
 394         }
 395
 396         if (base < xdr->head[0].iov_len || addr != NULL) {
 397                 unsigned int len = xdr->head[0].iov_len - base;
 398                 remainder -= len;
 399                 err = xs_send_kvec(sock, addr, addrlen, &xdr->head[0], base, remainder != 0);
 400                 if (remainder == 0 || err != len)
 401                         goto out;
 402                 sent += err;
 403                 base = 0;
 404         } else
 405                 base -= xdr->head[0].iov_len;
 406
 407         if (base < xdr->page_len) {
 408                 unsigned int len = xdr->page_len - base;
 409                 remainder -= len;
 410                 err = xs_send_pagedata(sock, xdr, base, remainder != 0);
 411                 if (remainder == 0 || err != len)
 412                         goto out;
 413                 sent += err;
 414                 base = 0;
 415         } else
 416                 base -= xdr->page_len;
 417
 418         if (base >= xdr->tail[0].iov_len)
 419                 return sent;
 420         err = xs_send_kvec(sock, NULL, 0, &xdr->tail[0], base, 0);
 421 out:
 422         if (sent == 0)
 423                 return err;
 424         if (err > 0)
 425                 sent += err;
 426         return sent;
 427 }
 428
 429 /**
 430  * xs_nospace - place task on wait queue if transmit was incomplete
 431  * @task: task to put to sleep
 432  *
 433  */
 434 static void xs_nospace(struct rpc_task *task)
 435 {
 436         struct rpc_rqst *req = task->tk_rqstp;
 437         struct rpc_xprt *xprt = req->rq_xprt;
 438         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 439
 440         dprintk("RPC: %5u xmit incomplete (%u left of %u)\n",
 441                         task->tk_pid, req->rq_slen - req->rq_bytes_sent,
 442                         req->rq_slen);
 443
 444         if (test_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags)) {
 445                 /* Protect against races with write_space */
 446                 spin_lock_bh(&xprt->transport_lock);
 447
 448                 /* Don't race with disconnect */
 449                 if (!xprt_connected(xprt))
 450                         task->tk_status = -ENOTCONN;
 451                 else if (test_bit(SOCK_NOSPACE, &transport->sock->flags))
 452                         xprt_wait_for_buffer_space(task);
 453
 454                 spin_unlock_bh(&xprt->transport_lock);
 455         } else
 456                 /* Keep holding the socket if it is blocked */
 457                 rpc_delay(task, HZ>>4);
 458 }
 459
 460 /**
 461  * xs_udp_send_request - write an RPC request to a UDP socket
 462  * @task: address of RPC task that manages the state of an RPC request
 463  *
 464  * Return values:
 465  *        0:    The request has been sent
 466  *   EAGAIN:    The socket was blocked, please call again later to
 467  *              complete the request
 468  * ENOTCONN:    Caller needs to invoke connect logic then call again
 469  *    other:    Some other error occured, the request was not sent
 470  */
 471 static int xs_udp_send_request(struct rpc_task *task)
 472 {
 473         struct rpc_rqst *req = task->tk_rqstp;
 474         struct rpc_xprt *xprt = req->rq_xprt;
 475         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 476         struct xdr_buf *xdr = &req->rq_snd_buf;
 477         int status;
 478
 479         xs_pktdump("packet data:",
 480                                 req->rq_svec->iov_base,
 481                                 req->rq_svec->iov_len);
 482
 483         req->rq_xtime = jiffies;
 484         status = xs_sendpages(transport->sock,
 485                               (struct sockaddr *) &xprt->addr,
 486                               xprt->addrlen, xdr,
 487                               req->rq_bytes_sent);
 488
 489         dprintk("RPC:       xs_udp_send_request(%u) = %d\n",
 490                         xdr->len - req->rq_bytes_sent, status);
 491
 492         if (likely(status >= (int) req->rq_slen))
 493                 return 0;
 494
 495         /* Still some bytes left; set up for a retry later. */
 496         if (status > 0)
 497                 status = -EAGAIN;
 498
 499         switch (status) {
 500         case -ENETUNREACH:
 501         case -EPIPE:
 502         case -ECONNREFUSED:
 503                 /* When the server has died, an ICMP port unreachable message
 504                  * prompts ECONNREFUSED. */
 505                 break;
 506         case -EAGAIN:
 507                 xs_nospace(task);
 508                 break;
 509         default:
 510                 dprintk("RPC:       sendmsg returned unrecognized error %d\n",
 511                         -status);
 512                 break;
 513         }
 514
 515         return status;
 516 }
 517
 518 static inline void xs_encode_tcp_record_marker(struct xdr_buf *buf)
 519 {
 520         u32 reclen = buf->len - sizeof(rpc_fraghdr);
 521         rpc_fraghdr *base = buf->head[0].iov_base;
 522         *base = htonl(RPC_LAST_STREAM_FRAGMENT | reclen);
 523 }
 524
 525 /**
 526  * xs_tcp_send_request - write an RPC request to a TCP socket
 527  * @task: address of RPC task that manages the state of an RPC request
 528  *
 529  * Return values:
 530  *        0:    The request has been sent
 531  *   EAGAIN:    The socket was blocked, please call again later to
 532  *              complete the request
 533  * ENOTCONN:    Caller needs to invoke connect logic then call again
 534  *    other:    Some other error occured, the request was not sent
 535  *
 536  * XXX: In the case of soft timeouts, should we eventually give up
 537  *      if sendmsg is not able to make progress?
 538  */
 539 static int xs_tcp_send_request(struct rpc_task *task)
 540 {
 541         struct rpc_rqst *req = task->tk_rqstp;
 542         struct rpc_xprt *xprt = req->rq_xprt;
 543         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 544         struct xdr_buf *xdr = &req->rq_snd_buf;
 545         int status;
 546         unsigned int retry = 0;
 547
 548         xs_encode_tcp_record_marker(&req->rq_snd_buf);
 549
 550         xs_pktdump("packet data:",
 551                                 req->rq_svec->iov_base,
 552                                 req->rq_svec->iov_len);
 553
 554         /* Continue transmitting the packet/record. We must be careful
 555          * to cope with writespace callbacks arriving _after_ we have
 556          * called sendmsg(). */
 557         while (1) {
 558                 req->rq_xtime = jiffies;
 559                 status = xs_sendpages(transport->sock,
 560                                         NULL, 0, xdr, req->rq_bytes_sent);
 561
 562                 dprintk("RPC:       xs_tcp_send_request(%u) = %d\n",
 563                                 xdr->len - req->rq_bytes_sent, status);
 564
 565                 if (unlikely(status < 0))
 566                         break;
 567
 568                 /* If we've sent the entire packet, immediately
 569                  * reset the count of bytes sent. */
 570                 req->rq_bytes_sent += status;
 571                 task->tk_bytes_sent += status;
 572                 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
 573                         req->rq_bytes_sent = 0;
 574                         return 0;
 575                 }
 576
 577                 status = -EAGAIN;
 578                 if (retry++ > XS_SENDMSG_RETRY)
 579                         break;
 580         }
 581
 582         switch (status) {
 583         case -EAGAIN:
 584                 xs_nospace(task);
 585                 break;
 586         case -ECONNREFUSED:
 587         case -ECONNRESET:
 588         case -ENOTCONN:
 589         case -EPIPE:
 590                 status = -ENOTCONN;
 591                 break;
 592         default:
 593                 dprintk("RPC:       sendmsg returned unrecognized error %d\n",
 594                         -status);
 595                 xprt_disconnect(xprt);
 596                 break;
 597         }
 598
 599         return status;
 600 }
 601
 602 /**
 603  * xs_tcp_release_xprt - clean up after a tcp transmission
 604  * @xprt: transport
 605  * @task: rpc task
 606  *
 607  * This cleans up if an error causes us to abort the transmission of a request.
 608  * In this case, the socket may need to be reset in order to avoid confusing
 609  * the server.
 610  */
 611 static void xs_tcp_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 612 {
 613         struct rpc_rqst *req;
 614
 615         if (task != xprt->snd_task)
 616                 return;
 617         if (task == NULL)
 618                 goto out_release;
 619         req = task->tk_rqstp;
 620         if (req->rq_bytes_sent == 0)
 621                 goto out_release;
 622         if (req->rq_bytes_sent == req->rq_snd_buf.len)
 623                 goto out_release;
 624         set_bit(XPRT_CLOSE_WAIT, &task->tk_xprt->state);
 625 out_release:
 626         xprt_release_xprt(xprt, task);
 627 }
 628
 629 /**
 630  * xs_close - close a socket
 631  * @xprt: transport
 632  *
 633  * This is used when all requests are complete; ie, no DRC state remains
 634  * on the server we want to save.
 635  */
 636 static void xs_close(struct rpc_xprt *xprt)
 637 {
 638         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 639         struct socket *sock = transport->sock;
 640         struct sock *sk = transport->inet;
 641
 642         if (!sk)
 643                 goto clear_close_wait;
 644
 645         dprintk("RPC:       xs_close xprt %p\n", xprt);
 646
 647         write_lock_bh(&sk->sk_callback_lock);
 648         transport->inet = NULL;
 649         transport->sock = NULL;
 650
 651         sk->sk_user_data = NULL;
 652         sk->sk_data_ready = transport->old_data_ready;
 653         sk->sk_state_change = transport->old_state_change;
 654         sk->sk_write_space = transport->old_write_space;
 655         write_unlock_bh(&sk->sk_callback_lock);
 656
 657         sk->sk_no_check = 0;
 658
 659         sock_release(sock);
 660 clear_close_wait:
 661         smp_mb__before_clear_bit();
 662         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 663         smp_mb__after_clear_bit();
 664 }
 665
 666 /**
 667  * xs_destroy - prepare to shutdown a transport
 668  * @xprt: doomed transport
 669  *
 670  */
 671 static void xs_destroy(struct rpc_xprt *xprt)
 672 {
 673         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 674
 675         dprintk("RPC:       xs_destroy xprt %p\n", xprt);
 676
 677         cancel_rearming_delayed_work(&transport->connect_worker);
 678
 679         xprt_disconnect(xprt);
 680         xs_close(xprt);
 681         xs_free_peer_addresses(xprt);
 682         kfree(xprt->slot);
 683         kfree(xprt);
 684 }
 685
 686 static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
 687 {
 688         return (struct rpc_xprt *) sk->sk_user_data;
 689 }
 690
 691 /**
 692  * xs_udp_data_ready - "data ready" callback for UDP sockets
 693  * @sk: socket with data to read
 694  * @len: how much data to read
 695  *
 696  */
 697 static void xs_udp_data_ready(struct sock *sk, int len)
 698 {
 699         struct rpc_task *task;
 700         struct rpc_xprt *xprt;
 701         struct rpc_rqst *rovr;
 702         struct sk_buff *skb;
 703         int err, repsize, copied;
 704         u32 _xid;
 705         __be32 *xp;
 706
 707         read_lock(&sk->sk_callback_lock);
 708         dprintk("RPC:       xs_udp_data_ready...\n");
 709         if (!(xprt = xprt_from_sock(sk)))
 710                 goto out;
 711
 712         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 713                 goto out;
 714
 715         if (xprt->shutdown)
 716                 goto dropit;
 717
 718         repsize = skb->len - sizeof(struct udphdr);
 719         if (repsize < 4) {
 720                 dprintk("RPC:       impossible RPC reply size %d!\n", repsize);
 721                 goto dropit;
 722         }
 723
 724         /* Copy the XID from the skb... */
 725         xp = skb_header_pointer(skb, sizeof(struct udphdr),
 726                                 sizeof(_xid), &_xid);
 727         if (xp == NULL)
 728                 goto dropit;
 729
 730         /* Look up and lock the request corresponding to the given XID */
 731         spin_lock(&xprt->transport_lock);
 732         rovr = xprt_lookup_rqst(xprt, *xp);
 733         if (!rovr)
 734                 goto out_unlock;
 735         task = rovr->rq_task;
 736
 737         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 738                 copied = repsize;
 739
 740         /* Suck it into the iovec, verify checksum if not done by hw. */
 741         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 742                 goto out_unlock;
 743
 744         /* Something worked... */
 745         dst_confirm(skb->dst);
 746
 747         xprt_adjust_cwnd(task, copied);
 748         xprt_update_rtt(task);
 749         xprt_complete_rqst(task, copied);
 750
 751  out_unlock:
 752         spin_unlock(&xprt->transport_lock);
 753  dropit:
 754         skb_free_datagram(sk, skb);
 755  out:
 756         read_unlock(&sk->sk_callback_lock);
 757 }
 758
 759 static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
 760 {
 761         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 762         size_t len, used;
 763         char *p;
 764
 765         p = ((char *) &transport->tcp_fraghdr) + transport->tcp_offset;
 766         len = sizeof(transport->tcp_fraghdr) - transport->tcp_offset;
 767         used = xdr_skb_read_bits(desc, p, len);
 768         transport->tcp_offset += used;
 769         if (used != len)
 770                 return;
 771
 772         transport->tcp_reclen = ntohl(transport->tcp_fraghdr);
 773         if (transport->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
 774                 transport->tcp_flags |= TCP_RCV_LAST_FRAG;
 775         else
 776                 transport->tcp_flags &= ~TCP_RCV_LAST_FRAG;
 777         transport->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
 778
 779         transport->tcp_flags &= ~TCP_RCV_COPY_FRAGHDR;
 780         transport->tcp_offset = 0;
 781
 782         /* Sanity check of the record length */
 783         if (unlikely(transport->tcp_reclen < 4)) {
 784                 dprintk("RPC:       invalid TCP record fragment length\n");
 785                 xprt_disconnect(xprt);
 786                 return;
 787         }
 788         dprintk("RPC:       reading TCP record fragment of length %d\n",
 789                         transport->tcp_reclen);
 790 }
 791
 792 static void xs_tcp_check_fraghdr(struct sock_xprt *transport)
 793 {
 794         if (transport->tcp_offset == transport->tcp_reclen) {
 795                 transport->tcp_flags |= TCP_RCV_COPY_FRAGHDR;
 796                 transport->tcp_offset = 0;
 797                 if (transport->tcp_flags & TCP_RCV_LAST_FRAG) {
 798                         transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 799                         transport->tcp_flags |= TCP_RCV_COPY_XID;
 800                         transport->tcp_copied = 0;
 801                 }
 802         }
 803 }
 804
 805 static inline void xs_tcp_read_xid(struct sock_xprt *transport, struct xdr_skb_reader *desc)
 806 {
 807         size_t len, used;
 808         char *p;
 809
 810         len = sizeof(transport->tcp_xid) - transport->tcp_offset;
 811         dprintk("RPC:       reading XID (%Zu bytes)\n", len);
 812         p = ((char *) &transport->tcp_xid) + transport->tcp_offset;
 813         used = xdr_skb_read_bits(desc, p, len);
 814         transport->tcp_offset += used;
 815         if (used != len)
 816                 return;
 817         transport->tcp_flags &= ~TCP_RCV_COPY_XID;
 818         transport->tcp_flags |= TCP_RCV_COPY_DATA;
 819         transport->tcp_copied = 4;
 820         dprintk("RPC:       reading reply for XID %08x\n",
 821                         ntohl(transport->tcp_xid));
 822         xs_tcp_check_fraghdr(transport);
 823 }
 824
 825 static inline void xs_tcp_read_request(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
 826 {
 827         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 828         struct rpc_rqst *req;
 829         struct xdr_buf *rcvbuf;
 830         size_t len;
 831         ssize_t r;
 832
 833         /* Find and lock the request corresponding to this xid */
 834         spin_lock(&xprt->transport_lock);
 835         req = xprt_lookup_rqst(xprt, transport->tcp_xid);
 836         if (!req) {
 837                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 838                 dprintk("RPC:       XID %08x request not found!\n",
 839                                 ntohl(transport->tcp_xid));
 840                 spin_unlock(&xprt->transport_lock);
 841                 return;
 842         }
 843
 844         rcvbuf = &req->rq_private_buf;
 845         len = desc->count;
 846         if (len > transport->tcp_reclen - transport->tcp_offset) {
 847                 struct xdr_skb_reader my_desc;
 848
 849                 len = transport->tcp_reclen - transport->tcp_offset;
 850                 memcpy(&my_desc, desc, sizeof(my_desc));
 851                 my_desc.count = len;
 852                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
 853                                           &my_desc, xdr_skb_read_bits);
 854                 desc->count -= r;
 855                 desc->offset += r;
 856         } else
 857                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
 858                                           desc, xdr_skb_read_bits);
 859
 860         if (r > 0) {
 861                 transport->tcp_copied += r;
 862                 transport->tcp_offset += r;
 863         }
 864         if (r != len) {
 865                 /* Error when copying to the receive buffer,
 866                  * usually because we weren't able to allocate
 867                  * additional buffer pages. All we can do now
 868                  * is turn off TCP_RCV_COPY_DATA, so the request
 869                  * will not receive any additional updates,
 870                  * and time out.
 871                  * Any remaining data from this record will
 872                  * be discarded.
 873                  */
 874                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 875                 dprintk("RPC:       XID %08x truncated request\n",
 876                                 ntohl(transport->tcp_xid));
 877                 dprintk("RPC:       xprt = %p, tcp_copied = %lu, "
 878                                 "tcp_offset = %u, tcp_reclen = %u\n",
 879                                 xprt, transport->tcp_copied,
 880                                 transport->tcp_offset, transport->tcp_reclen);
 881                 goto out;
 882         }
 883
 884         dprintk("RPC:       XID %08x read %Zd bytes\n",
 885                         ntohl(transport->tcp_xid), r);
 886         dprintk("RPC:       xprt = %p, tcp_copied = %lu, tcp_offset = %u, "
 887                         "tcp_reclen = %u\n", xprt, transport->tcp_copied,
 888                         transport->tcp_offset, transport->tcp_reclen);
 889
 890         if (transport->tcp_copied == req->rq_private_buf.buflen)
 891                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 892         else if (transport->tcp_offset == transport->tcp_reclen) {
 893                 if (transport->tcp_flags & TCP_RCV_LAST_FRAG)
 894                         transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 895         }
 896
 897 out:
 898         if (!(transport->tcp_flags & TCP_RCV_COPY_DATA))
 899                 xprt_complete_rqst(req->rq_task, transport->tcp_copied);
 900         spin_unlock(&xprt->transport_lock);
 901         xs_tcp_check_fraghdr(transport);
 902 }
 903
 904 static inline void xs_tcp_read_discard(struct sock_xprt *transport, struct xdr_skb_reader *desc)
 905 {
 906         size_t len;
 907
 908         len = transport->tcp_reclen - transport->tcp_offset;
 909         if (len > desc->count)
 910                 len = desc->count;
 911         desc->count -= len;
 912         desc->offset += len;
 913         transport->tcp_offset += len;
 914         dprintk("RPC:       discarded %Zu bytes\n", len);
 915         xs_tcp_check_fraghdr(transport);
 916 }
 917
 918 static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
 919 {
 920         struct rpc_xprt *xprt = rd_desc->arg.data;
 921         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 922         struct xdr_skb_reader desc = {
 923                 .skb    = skb,
 924                 .offset = offset,
 925                 .count  = len,
 926         };
 927
 928         dprintk("RPC:       xs_tcp_data_recv started\n");
 929         do {
 930                 /* Read in a new fragment marker if necessary */
 931                 /* Can we ever really expect to get completely empty fragments? */
 932                 if (transport->tcp_flags & TCP_RCV_COPY_FRAGHDR) {
 933                         xs_tcp_read_fraghdr(xprt, &desc);
 934                         continue;
 935                 }
 936                 /* Read in the xid if necessary */
 937                 if (transport->tcp_flags & TCP_RCV_COPY_XID) {
 938                         xs_tcp_read_xid(transport, &desc);
 939                         continue;
 940                 }
 941                 /* Read in the request data */
 942                 if (transport->tcp_flags & TCP_RCV_COPY_DATA) {
 943                         xs_tcp_read_request(xprt, &desc);
 944                         continue;
 945                 }
 946                 /* Skip over any trailing bytes on short reads */
 947                 xs_tcp_read_discard(transport, &desc);
 948         } while (desc.count);
 949         dprintk("RPC:       xs_tcp_data_recv done\n");
 950         return len - desc.count;
 951 }
 952
 953 /**
 954  * xs_tcp_data_ready - "data ready" callback for TCP sockets
 955  * @sk: socket with data to read
 956  * @bytes: how much data to read
 957  *
 958  */
 959 static void xs_tcp_data_ready(struct sock *sk, int bytes)
 960 {
 961         struct rpc_xprt *xprt;
 962         read_descriptor_t rd_desc;
 963
 964         dprintk("RPC:       xs_tcp_data_ready...\n");
 965
 966         read_lock(&sk->sk_callback_lock);
 967         if (!(xprt = xprt_from_sock(sk)))
 968                 goto out;
 969         if (xprt->shutdown)
 970                 goto out;
 971
 972         /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
 973         rd_desc.arg.data = xprt;
 974         rd_desc.count = 65536;
 975         tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
 976 out:
 977         read_unlock(&sk->sk_callback_lock);
 978 }
 979
 980 /**
 981  * xs_tcp_state_change - callback to handle TCP socket state changes
 982  * @sk: socket whose state has changed
 983  *
 984  */
 985 static void xs_tcp_state_change(struct sock *sk)
 986 {
 987         struct rpc_xprt *xprt;
 988
 989         read_lock(&sk->sk_callback_lock);
 990         if (!(xprt = xprt_from_sock(sk)))
 991                 goto out;
 992         dprintk("RPC:       xs_tcp_state_change client %p...\n", xprt);
 993         dprintk("RPC:       state %x conn %d dead %d zapped %d\n",
 994                         sk->sk_state, xprt_connected(xprt),
 995                         sock_flag(sk, SOCK_DEAD),
 996                         sock_flag(sk, SOCK_ZAPPED));
 997
 998         switch (sk->sk_state) {
 999         case TCP_ESTABLISHED:
1000                 spin_lock_bh(&xprt->transport_lock);
1001                 if (!xprt_test_and_set_connected(xprt)) {
1002                         struct sock_xprt *transport = container_of(xprt,
1003                                         struct sock_xprt, xprt);
1004
1005                         /* Reset TCP record info */
1006                         transport->tcp_offset = 0;
1007                         transport->tcp_reclen = 0;
1008                         transport->tcp_copied = 0;
1009                         transport->tcp_flags =
1010                                 TCP_RCV_COPY_FRAGHDR | TCP_RCV_COPY_XID;
1011
1012                         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1013                         xprt_wake_pending_tasks(xprt, 0);
1014                 }
1015                 spin_unlock_bh(&xprt->transport_lock);
1016                 break;
1017         case TCP_SYN_SENT:
1018         case TCP_SYN_RECV:
1019                 break;
1020         case TCP_CLOSE_WAIT:
1021                 /* Try to schedule an autoclose RPC calls */
1022                 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
1023                 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
1024                         queue_work(rpciod_workqueue, &xprt->task_cleanup);
1025         default:
1026                 xprt_disconnect(xprt);
1027         }
1028  out:
1029         read_unlock(&sk->sk_callback_lock);
1030 }
1031
1032 /**
1033  * xs_udp_write_space - callback invoked when socket buffer space
1034  *                             becomes available
1035  * @sk: socket whose state has changed
1036  *
1037  * Called when more output buffer space is available for this socket.
1038  * We try not to wake our writers until they can make "significant"
1039  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1040  * with a bunch of small requests.
1041  */
1042 static void xs_udp_write_space(struct sock *sk)
1043 {
1044         read_lock(&sk->sk_callback_lock);
1045
1046         /* from net/core/sock.c:sock_def_write_space */
1047         if (sock_writeable(sk)) {
1048                 struct socket *sock;
1049                 struct rpc_xprt *xprt;
1050
1051                 if (unlikely(!(sock = sk->sk_socket)))
1052                         goto out;
1053                 if (unlikely(!(xprt = xprt_from_sock(sk))))
1054                         goto out;
1055                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
1056                         goto out;
1057
1058                 xprt_write_space(xprt);
1059         }
1060
1061  out:
1062         read_unlock(&sk->sk_callback_lock);
1063 }
1064
1065 /**
1066  * xs_tcp_write_space - callback invoked when socket buffer space
1067  *                             becomes available
1068  * @sk: socket whose state has changed
1069  *
1070  * Called when more output buffer space is available for this socket.
1071  * We try not to wake our writers until they can make "significant"
1072  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1073  * with a bunch of small requests.
1074  */
1075 static void xs_tcp_write_space(struct sock *sk)
1076 {
1077         read_lock(&sk->sk_callback_lock);
1078
1079         /* from net/core/stream.c:sk_stream_write_space */
1080         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
1081                 struct socket *sock;
1082                 struct rpc_xprt *xprt;
1083
1084                 if (unlikely(!(sock = sk->sk_socket)))
1085                         goto out;
1086                 if (unlikely(!(xprt = xprt_from_sock(sk))))
1087                         goto out;
1088                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
1089                         goto out;
1090
1091                 xprt_write_space(xprt);
1092         }
1093
1094  out:
1095         read_unlock(&sk->sk_callback_lock);
1096 }
1097
1098 static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
1099 {
1100         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1101         struct sock *sk = transport->inet;
1102
1103         if (transport->rcvsize) {
1104                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1105                 sk->sk_rcvbuf = transport->rcvsize * xprt->max_reqs * 2;
1106         }
1107         if (transport->sndsize) {
1108                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1109                 sk->sk_sndbuf = transport->sndsize * xprt->max_reqs * 2;
1110                 sk->sk_write_space(sk);
1111         }
1112 }
1113
1114 /**
1115  * xs_udp_set_buffer_size - set send and receive limits
1116  * @xprt: generic transport
1117  * @sndsize: requested size of send buffer, in bytes
1118  * @rcvsize: requested size of receive buffer, in bytes
1119  *
1120  * Set socket send and receive buffer size limits.
1121  */
1122 static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
1123 {
1124         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1125
1126         transport->sndsize = 0;
1127         if (sndsize)
1128                 transport->sndsize = sndsize + 1024;
1129         transport->rcvsize = 0;
1130         if (rcvsize)
1131                 transport->rcvsize = rcvsize + 1024;
1132
1133         xs_udp_do_set_buffer_size(xprt);
1134 }
1135
1136 /**
1137  * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
1138  * @task: task that timed out
1139  *
1140  * Adjust the congestion window after a retransmit timeout has occurred.
1141  */
1142 static void xs_udp_timer(struct rpc_task *task)
1143 {
1144         xprt_adjust_cwnd(task, -ETIMEDOUT);
1145 }
1146
1147 static unsigned short xs_get_random_port(void)
1148 {
1149         unsigned short range = xprt_max_resvport - xprt_min_resvport;
1150         unsigned short rand = (unsigned short) net_random() % range;
1151         return rand + xprt_min_resvport;
1152 }
1153
1154 /**
1155  * xs_set_port - reset the port number in the remote endpoint address
1156  * @xprt: generic transport
1157  * @port: new port number
1158  *
1159  */
1160 static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
1161 {
1162         struct sockaddr_in *sap = (struct sockaddr_in *) &xprt->addr;
1163
1164         dprintk("RPC:       setting port for xprt %p to %u\n", xprt, port);
1165
1166         sap->sin_port = htons(port);
1167 }
1168
1169 static int xs_bind(struct sock_xprt *transport, struct socket *sock)
1170 {
1171         struct sockaddr_in myaddr = {
1172                 .sin_family = AF_INET,
1173         };
1174         struct sockaddr_in *sa;
1175         int err;
1176         unsigned short port = transport->port;
1177
1178         if (!transport->xprt.resvport)
1179                 port = 0;
1180         sa = (struct sockaddr_in *)&transport->addr;
1181         myaddr.sin_addr = sa->sin_addr;
1182         do {
1183                 myaddr.sin_port = htons(port);
1184                 err = kernel_bind(sock, (struct sockaddr *) &myaddr,
1185                                                 sizeof(myaddr));
1186                 if (!transport->xprt.resvport)
1187                         break;
1188                 if (err == 0) {
1189                         transport->port = port;
1190                         break;
1191                 }
1192                 if (port <= xprt_min_resvport)
1193                         port = xprt_max_resvport;
1194                 else
1195                         port--;
1196         } while (err == -EADDRINUSE && port != transport->port);
1197         dprintk("RPC:       xs_bind "NIPQUAD_FMT":%u: %s (%d)\n",
1198                 NIPQUAD(myaddr.sin_addr), port, err ? "failed" : "ok", err);
1199         return err;
1200 }
1201
1202 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1203 static struct lock_class_key xs_key[2];
1204 static struct lock_class_key xs_slock_key[2];
1205
1206 static inline void xs_reclassify_socket(struct socket *sock)
1207 {
1208         struct sock *sk = sock->sk;
1209         BUG_ON(sk->sk_lock.owner != NULL);
1210         switch (sk->sk_family) {
1211         case AF_INET:
1212                 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFS",
1213                         &xs_slock_key[0], "sk_lock-AF_INET-NFS", &xs_key[0]);
1214                 break;
1215
1216         case AF_INET6:
1217                 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFS",
1218                         &xs_slock_key[1], "sk_lock-AF_INET6-NFS", &xs_key[1]);
1219                 break;
1220
1221         default:
1222                 BUG();
1223         }
1224 }
1225 #else
1226 static inline void xs_reclassify_socket(struct socket *sock)
1227 {
1228 }
1229 #endif
1230
1231 /**
1232  * xs_udp_connect_worker - set up a UDP socket
1233  * @work: RPC transport to connect
1234  *
1235  * Invoked by a work queue tasklet.
1236  */
1237 static void xs_udp_connect_worker(struct work_struct *work)
1238 {
1239         struct sock_xprt *transport =
1240                 container_of(work, struct sock_xprt, connect_worker.work);
1241         struct rpc_xprt *xprt = &transport->xprt;
1242         struct socket *sock = transport->sock;
1243         int err, status = -EIO;
1244
1245         if (xprt->shutdown || !xprt_bound(xprt))
1246                 goto out;
1247
1248         /* Start by resetting any existing state */
1249         xs_close(xprt);
1250
1251         if ((err = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock)) < 0) {
1252                 dprintk("RPC:       can't create UDP transport socket (%d).\n", -err);
1253                 goto out;
1254         }
1255         xs_reclassify_socket(sock);
1256
1257         if (xs_bind(transport, sock)) {
1258                 sock_release(sock);
1259                 goto out;
1260         }
1261
1262         dprintk("RPC:       worker connecting xprt %p to address: %s\n",
1263                         xprt, xprt->address_strings[RPC_DISPLAY_ALL]);
1264
1265         if (!transport->inet) {
1266                 struct sock *sk = sock->sk;
1267
1268                 write_lock_bh(&sk->sk_callback_lock);
1269
1270                 sk->sk_user_data = xprt;
1271                 transport->old_data_ready = sk->sk_data_ready;
1272                 transport->old_state_change = sk->sk_state_change;
1273                 transport->old_write_space = sk->sk_write_space;
1274                 sk->sk_data_ready = xs_udp_data_ready;
1275                 sk->sk_write_space = xs_udp_write_space;
1276                 sk->sk_no_check = UDP_CSUM_NORCV;
1277                 sk->sk_allocation = GFP_ATOMIC;
1278
1279                 xprt_set_connected(xprt);
1280
1281                 /* Reset to new socket */
1282                 transport->sock = sock;
1283                 transport->inet = sk;
1284
1285                 write_unlock_bh(&sk->sk_callback_lock);
1286         }
1287         xs_udp_do_set_buffer_size(xprt);
1288         status = 0;
1289 out:
1290         xprt_wake_pending_tasks(xprt, status);
1291         xprt_clear_connecting(xprt);
1292 }
1293
1294 /*
1295  * We need to preserve the port number so the reply cache on the server can
1296  * find our cached RPC replies when we get around to reconnecting.
1297  */
1298 static void xs_tcp_reuse_connection(struct rpc_xprt *xprt)
1299 {
1300         int result;
1301         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1302         struct sockaddr any;
1303
1304         dprintk("RPC:       disconnecting xprt %p to reuse port\n", xprt);
1305
1306         /*
1307          * Disconnect the transport socket by doing a connect operation
1308          * with AF_UNSPEC.  This should return immediately...
1309          */
1310         memset(&any, 0, sizeof(any));
1311         any.sa_family = AF_UNSPEC;
1312         result = kernel_connect(transport->sock, &any, sizeof(any), 0);
1313         if (result)
1314                 dprintk("RPC:       AF_UNSPEC connect return code %d\n",
1315                                 result);
1316 }
1317
1318 /**
1319  * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1320  * @work: RPC transport to connect
1321  *
1322  * Invoked by a work queue tasklet.
1323  */
1324 static void xs_tcp_connect_worker(struct work_struct *work)
1325 {
1326         struct sock_xprt *transport =
1327                 container_of(work, struct sock_xprt, connect_worker.work);
1328         struct rpc_xprt *xprt = &transport->xprt;
1329         struct socket *sock = transport->sock;
1330         int err, status = -EIO;
1331
1332         if (xprt->shutdown || !xprt_bound(xprt))
1333                 goto out;
1334
1335         if (!sock) {
1336                 /* start from scratch */
1337                 if ((err = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock)) < 0) {
1338                         dprintk("RPC:       can't create TCP transport "
1339                                         "socket (%d).\n", -err);
1340                         goto out;
1341                 }
1342                 xs_reclassify_socket(sock);
1343
1344                 if (xs_bind(transport, sock)) {
1345                         sock_release(sock);
1346                         goto out;
1347                 }
1348         } else
1349                 /* "close" the socket, preserving the local port */
1350                 xs_tcp_reuse_connection(xprt);
1351
1352         dprintk("RPC:       worker connecting xprt %p to address: %s\n",
1353                         xprt, xprt->address_strings[RPC_DISPLAY_ALL]);
1354
1355         if (!transport->inet) {
1356                 struct sock *sk = sock->sk;
1357
1358                 write_lock_bh(&sk->sk_callback_lock);
1359
1360                 sk->sk_user_data = xprt;
1361                 transport->old_data_ready = sk->sk_data_ready;
1362                 transport->old_state_change = sk->sk_state_change;
1363                 transport->old_write_space = sk->sk_write_space;
1364                 sk->sk_data_ready = xs_tcp_data_ready;
1365                 sk->sk_state_change = xs_tcp_state_change;
1366                 sk->sk_write_space = xs_tcp_write_space;
1367                 sk->sk_allocation = GFP_ATOMIC;
1368
1369                 /* socket options */
1370                 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
1371                 sock_reset_flag(sk, SOCK_LINGER);
1372                 tcp_sk(sk)->linger2 = 0;
1373                 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1374
1375                 xprt_clear_connected(xprt);
1376
1377                 /* Reset to new socket */
1378                 transport->sock = sock;
1379                 transport->inet = sk;
1380
1381                 write_unlock_bh(&sk->sk_callback_lock);
1382         }
1383
1384         /* Tell the socket layer to start connecting... */
1385         xprt->stat.connect_count++;
1386         xprt->stat.connect_start = jiffies;
1387         status = kernel_connect(sock, (struct sockaddr *) &xprt->addr,
1388                         xprt->addrlen, O_NONBLOCK);
1389         dprintk("RPC:       %p connect status %d connected %d sock state %d\n",
1390                         xprt, -status, xprt_connected(xprt),
1391                         sock->sk->sk_state);
1392         if (status < 0) {
1393                 switch (status) {
1394                         case -EINPROGRESS:
1395                         case -EALREADY:
1396                                 goto out_clear;
1397                         case -ECONNREFUSED:
1398                         case -ECONNRESET:
1399                                 /* retry with existing socket, after a delay */
1400                                 break;
1401                         default:
1402                                 /* get rid of existing socket, and retry */
1403                                 xs_close(xprt);
1404                                 break;
1405                 }
1406         }
1407 out:
1408         xprt_wake_pending_tasks(xprt, status);
1409 out_clear:
1410         xprt_clear_connecting(xprt);
1411 }
1412
1413 /**
1414  * xs_connect - connect a socket to a remote endpoint
1415  * @task: address of RPC task that manages state of connect request
1416  *
1417  * TCP: If the remote end dropped the connection, delay reconnecting.
1418  *
1419  * UDP socket connects are synchronous, but we use a work queue anyway
1420  * to guarantee that even unprivileged user processes can set up a
1421  * socket on a privileged port.
1422  *
1423  * If a UDP socket connect fails, the delay behavior here prevents
1424  * retry floods (hard mounts).
1425  */
1426 static void xs_connect(struct rpc_task *task)
1427 {
1428         struct rpc_xprt *xprt = task->tk_xprt;
1429         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1430
1431         if (xprt_test_and_set_connecting(xprt))
1432                 return;
1433
1434         if (transport->sock != NULL) {
1435                 dprintk("RPC:       xs_connect delayed xprt %p for %lu "
1436                                 "seconds\n",
1437                                 xprt, xprt->reestablish_timeout / HZ);
1438                 queue_delayed_work(rpciod_workqueue,
1439                                    &transport->connect_worker,
1440                                    xprt->reestablish_timeout);
1441                 xprt->reestablish_timeout <<= 1;
1442                 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
1443                         xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
1444         } else {
1445                 dprintk("RPC:       xs_connect scheduled xprt %p\n", xprt);
1446                 queue_delayed_work(rpciod_workqueue,
1447                                    &transport->connect_worker, 0);
1448         }
1449 }
1450
1451 /**
1452  * xs_udp_print_stats - display UDP socket-specifc stats
1453  * @xprt: rpc_xprt struct containing statistics
1454  * @seq: output file
1455  *
1456  */
1457 static void xs_udp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1458 {
1459         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1460
1461         seq_printf(seq, "\txprt:\tudp %u %lu %lu %lu %lu %Lu %Lu\n",
1462                         transport->port,
1463                         xprt->stat.bind_count,
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 /**
1472  * xs_tcp_print_stats - display TCP socket-specifc stats
1473  * @xprt: rpc_xprt struct containing statistics
1474  * @seq: output file
1475  *
1476  */
1477 static void xs_tcp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1478 {
1479         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1480         long idle_time = 0;
1481
1482         if (xprt_connected(xprt))
1483                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
1484
1485         seq_printf(seq, "\txprt:\ttcp %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu\n",
1486                         transport->port,
1487                         xprt->stat.bind_count,
1488                         xprt->stat.connect_count,
1489                         xprt->stat.connect_time,
1490                         idle_time,
1491                         xprt->stat.sends,
1492                         xprt->stat.recvs,
1493                         xprt->stat.bad_xids,
1494                         xprt->stat.req_u,
1495                         xprt->stat.bklog_u);
1496 }
1497
1498 static struct rpc_xprt_ops xs_udp_ops = {
1499         .set_buffer_size        = xs_udp_set_buffer_size,
1500         .reserve_xprt           = xprt_reserve_xprt_cong,
1501         .release_xprt           = xprt_release_xprt_cong,
1502         .rpcbind                = rpcb_getport_async,
1503         .set_port               = xs_set_port,
1504         .connect                = xs_connect,
1505         .buf_alloc              = rpc_malloc,
1506         .buf_free               = rpc_free,
1507         .send_request           = xs_udp_send_request,
1508         .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
1509         .timer                  = xs_udp_timer,
1510         .release_request        = xprt_release_rqst_cong,
1511         .close                  = xs_close,
1512         .destroy                = xs_destroy,
1513         .print_stats            = xs_udp_print_stats,
1514 };
1515
1516 static struct rpc_xprt_ops xs_tcp_ops = {
1517         .reserve_xprt           = xprt_reserve_xprt,
1518         .release_xprt           = xs_tcp_release_xprt,
1519         .rpcbind                = rpcb_getport_async,
1520         .set_port               = xs_set_port,
1521         .connect                = xs_connect,
1522         .buf_alloc              = rpc_malloc,
1523         .buf_free               = rpc_free,
1524         .send_request           = xs_tcp_send_request,
1525         .set_retrans_timeout    = xprt_set_retrans_timeout_def,
1526         .close                  = xs_close,
1527         .destroy                = xs_destroy,
1528         .print_stats            = xs_tcp_print_stats,
1529 };
1530
1531 static struct rpc_xprt *xs_setup_xprt(struct rpc_xprtsock_create *args, unsigned int slot_table_size)
1532 {
1533         struct rpc_xprt *xprt;
1534         struct sock_xprt *new;
1535
1536         if (args->addrlen > sizeof(xprt->addr)) {
1537                 dprintk("RPC:       xs_setup_xprt: address too large\n");
1538                 return ERR_PTR(-EBADF);
1539         }
1540
1541         new = kzalloc(sizeof(*new), GFP_KERNEL);
1542         if (new == NULL) {
1543                 dprintk("RPC:       xs_setup_xprt: couldn't allocate "
1544                                 "rpc_xprt\n");
1545                 return ERR_PTR(-ENOMEM);
1546         }
1547         xprt = &new->xprt;
1548
1549         xprt->max_reqs = slot_table_size;
1550         xprt->slot = kcalloc(xprt->max_reqs, sizeof(struct rpc_rqst), GFP_KERNEL);
1551         if (xprt->slot == NULL) {
1552                 kfree(xprt);
1553                 dprintk("RPC:       xs_setup_xprt: couldn't allocate slot "
1554                                 "table\n");
1555                 return ERR_PTR(-ENOMEM);
1556         }
1557
1558         memcpy(&xprt->addr, args->dstaddr, args->addrlen);
1559         xprt->addrlen = args->addrlen;
1560         if (args->srcaddr)
1561                 memcpy(&new->addr, args->srcaddr, args->addrlen);
1562         new->port = xs_get_random_port();
1563
1564         return xprt;
1565 }
1566
1567 /**
1568  * xs_setup_udp - Set up transport to use a UDP socket
1569  * @args: rpc transport creation arguments
1570  *
1571  */
1572 struct rpc_xprt *xs_setup_udp(struct rpc_xprtsock_create *args)
1573 {
1574         struct rpc_xprt *xprt;
1575         struct sock_xprt *transport;
1576
1577         xprt = xs_setup_xprt(args, xprt_udp_slot_table_entries);
1578         if (IS_ERR(xprt))
1579                 return xprt;
1580         transport = container_of(xprt, struct sock_xprt, xprt);
1581
1582         if (ntohs(((struct sockaddr_in *)args->dstaddr)->sin_port) != 0)
1583                 xprt_set_bound(xprt);
1584
1585         xprt->prot = IPPROTO_UDP;
1586         xprt->tsh_size = 0;
1587         /* XXX: header size can vary due to auth type, IPv6, etc. */
1588         xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1589
1590         INIT_DELAYED_WORK(&transport->connect_worker, xs_udp_connect_worker);
1591         xprt->bind_timeout = XS_BIND_TO;
1592         xprt->connect_timeout = XS_UDP_CONN_TO;
1593         xprt->reestablish_timeout = XS_UDP_REEST_TO;
1594         xprt->idle_timeout = XS_IDLE_DISC_TO;
1595
1596         xprt->ops = &xs_udp_ops;
1597
1598         if (args->timeout)
1599                 xprt->timeout = *args->timeout;
1600         else
1601                 xprt_set_timeout(&xprt->timeout, 5, 5 * HZ);
1602
1603         xs_format_peer_addresses(xprt);
1604         dprintk("RPC:       set up transport to address %s\n",
1605                         xprt->address_strings[RPC_DISPLAY_ALL]);
1606
1607         return xprt;
1608 }
1609
1610 /**
1611  * xs_setup_tcp - Set up transport to use a TCP socket
1612  * @args: rpc transport creation arguments
1613  *
1614  */
1615 struct rpc_xprt *xs_setup_tcp(struct rpc_xprtsock_create *args)
1616 {
1617         struct rpc_xprt *xprt;
1618         struct sock_xprt *transport;
1619
1620         xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries);
1621         if (IS_ERR(xprt))
1622                 return xprt;
1623         transport = container_of(xprt, struct sock_xprt, xprt);
1624
1625         if (ntohs(((struct sockaddr_in *)args->dstaddr)->sin_port) != 0)
1626                 xprt_set_bound(xprt);
1627
1628         xprt->prot = IPPROTO_TCP;
1629         xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
1630         xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
1631
1632         INIT_DELAYED_WORK(&transport->connect_worker, xs_tcp_connect_worker);
1633         xprt->bind_timeout = XS_BIND_TO;
1634         xprt->connect_timeout = XS_TCP_CONN_TO;
1635         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1636         xprt->idle_timeout = XS_IDLE_DISC_TO;
1637
1638         xprt->ops = &xs_tcp_ops;
1639
1640         if (args->timeout)
1641                 xprt->timeout = *args->timeout;
1642         else
1643                 xprt_set_timeout(&xprt->timeout, 2, 60 * HZ);
1644
1645         xs_format_peer_addresses(xprt);
1646         dprintk("RPC:       set up transport to address %s\n",
1647                         xprt->address_strings[RPC_DISPLAY_ALL]);
1648
1649         return xprt;
1650 }
1651
1652 /**
1653  * init_socket_xprt - set up xprtsock's sysctls
1654  *
1655  */
1656 int init_socket_xprt(void)
1657 {
1658 #ifdef RPC_DEBUG
1659         if (!sunrpc_table_header)
1660                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
1661 #endif
1662
1663         return 0;
1664 }
1665
1666 /**
1667  * cleanup_socket_xprt - remove xprtsock's sysctls
1668  *
1669  */
1670 void cleanup_socket_xprt(void)
1671 {
1672 #ifdef RPC_DEBUG
1673         if (sunrpc_table_header) {
1674                 unregister_sysctl_table(sunrpc_table_header);
1675                 sunrpc_table_header = NULL;
1676         }
1677 #endif
1678 }