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
 301 static void xs_free_peer_addresses(struct rpc_xprt *xprt)
 302 {
 303         int i;
 304
 305         for (i = 0; i < RPC_DISPLAY_MAX; i++)
 306                 kfree(xprt->address_strings[i]);
 307 }
 308
 309 #define XS_SENDMSG_FLAGS        (MSG_DONTWAIT | MSG_NOSIGNAL)
 310
 311 static int xs_send_kvec(struct socket *sock, struct sockaddr *addr, int addrlen, struct kvec *vec, unsigned int base, int more)
 312 {
 313         struct msghdr msg = {
 314                 .msg_name       = addr,
 315                 .msg_namelen    = addrlen,
 316                 .msg_flags      = XS_SENDMSG_FLAGS | (more ? MSG_MORE : 0),
 317         };
 318         struct kvec iov = {
 319                 .iov_base       = vec->iov_base + base,
 320                 .iov_len        = vec->iov_len - base,
 321         };
 322
 323         if (iov.iov_len != 0)
 324                 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
 325         return kernel_sendmsg(sock, &msg, NULL, 0, 0);
 326 }
 327
 328 static int xs_send_pagedata(struct socket *sock, struct xdr_buf *xdr, unsigned int base, int more)
 329 {
 330         struct page **ppage;
 331         unsigned int remainder;
 332         int err, sent = 0;
 333
 334         remainder = xdr->page_len - base;
 335         base += xdr->page_base;
 336         ppage = xdr->pages + (base >> PAGE_SHIFT);
 337         base &= ~PAGE_MASK;
 338         for(;;) {
 339                 unsigned int len = min_t(unsigned int, PAGE_SIZE - base, remainder);
 340                 int flags = XS_SENDMSG_FLAGS;
 341
 342                 remainder -= len;
 343                 if (remainder != 0 || more)
 344                         flags |= MSG_MORE;
 345                 err = sock->ops->sendpage(sock, *ppage, base, len, flags);
 346                 if (remainder == 0 || err != len)
 347                         break;
 348                 sent += err;
 349                 ppage++;
 350                 base = 0;
 351         }
 352         if (sent == 0)
 353                 return err;
 354         if (err > 0)
 355                 sent += err;
 356         return sent;
 357 }
 358
 359 /**
 360  * xs_sendpages - write pages directly to a socket
 361  * @sock: socket to send on
 362  * @addr: UDP only -- address of destination
 363  * @addrlen: UDP only -- length of destination address
 364  * @xdr: buffer containing this request
 365  * @base: starting position in the buffer
 366  *
 367  */
 368 static int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
 369 {
 370         unsigned int remainder = xdr->len - base;
 371         int err, sent = 0;
 372
 373         if (unlikely(!sock))
 374                 return -ENOTCONN;
 375
 376         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 377         if (base != 0) {
 378                 addr = NULL;
 379                 addrlen = 0;
 380         }
 381
 382         if (base < xdr->head[0].iov_len || addr != NULL) {
 383                 unsigned int len = xdr->head[0].iov_len - base;
 384                 remainder -= len;
 385                 err = xs_send_kvec(sock, addr, addrlen, &xdr->head[0], base, remainder != 0);
 386                 if (remainder == 0 || err != len)
 387                         goto out;
 388                 sent += err;
 389                 base = 0;
 390         } else
 391                 base -= xdr->head[0].iov_len;
 392
 393         if (base < xdr->page_len) {
 394                 unsigned int len = xdr->page_len - base;
 395                 remainder -= len;
 396                 err = xs_send_pagedata(sock, xdr, base, remainder != 0);
 397                 if (remainder == 0 || err != len)
 398                         goto out;
 399                 sent += err;
 400                 base = 0;
 401         } else
 402                 base -= xdr->page_len;
 403
 404         if (base >= xdr->tail[0].iov_len)
 405                 return sent;
 406         err = xs_send_kvec(sock, NULL, 0, &xdr->tail[0], base, 0);
 407 out:
 408         if (sent == 0)
 409                 return err;
 410         if (err > 0)
 411                 sent += err;
 412         return sent;
 413 }
 414
 415 /**
 416  * xs_nospace - place task on wait queue if transmit was incomplete
 417  * @task: task to put to sleep
 418  *
 419  */
 420 static void xs_nospace(struct rpc_task *task)
 421 {
 422         struct rpc_rqst *req = task->tk_rqstp;
 423         struct rpc_xprt *xprt = req->rq_xprt;
 424         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 425
 426         dprintk("RPC: %5u xmit incomplete (%u left of %u)\n",
 427                         task->tk_pid, req->rq_slen - req->rq_bytes_sent,
 428                         req->rq_slen);
 429
 430         if (test_bit(SOCK_ASYNC_NOSPACE, &transport->sock->flags)) {
 431                 /* Protect against races with write_space */
 432                 spin_lock_bh(&xprt->transport_lock);
 433
 434                 /* Don't race with disconnect */
 435                 if (!xprt_connected(xprt))
 436                         task->tk_status = -ENOTCONN;
 437                 else if (test_bit(SOCK_NOSPACE, &transport->sock->flags))
 438                         xprt_wait_for_buffer_space(task);
 439
 440                 spin_unlock_bh(&xprt->transport_lock);
 441         } else
 442                 /* Keep holding the socket if it is blocked */
 443                 rpc_delay(task, HZ>>4);
 444 }
 445
 446 /**
 447  * xs_udp_send_request - write an RPC request to a UDP socket
 448  * @task: address of RPC task that manages the state of an RPC request
 449  *
 450  * Return values:
 451  *        0:    The request has been sent
 452  *   EAGAIN:    The socket was blocked, please call again later to
 453  *              complete the request
 454  * ENOTCONN:    Caller needs to invoke connect logic then call again
 455  *    other:    Some other error occured, the request was not sent
 456  */
 457 static int xs_udp_send_request(struct rpc_task *task)
 458 {
 459         struct rpc_rqst *req = task->tk_rqstp;
 460         struct rpc_xprt *xprt = req->rq_xprt;
 461         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 462         struct xdr_buf *xdr = &req->rq_snd_buf;
 463         int status;
 464
 465         xs_pktdump("packet data:",
 466                                 req->rq_svec->iov_base,
 467                                 req->rq_svec->iov_len);
 468
 469         req->rq_xtime = jiffies;
 470         status = xs_sendpages(transport->sock,
 471                               (struct sockaddr *) &xprt->addr,
 472                               xprt->addrlen, xdr,
 473                               req->rq_bytes_sent);
 474
 475         dprintk("RPC:       xs_udp_send_request(%u) = %d\n",
 476                         xdr->len - req->rq_bytes_sent, status);
 477
 478         if (likely(status >= (int) req->rq_slen))
 479                 return 0;
 480
 481         /* Still some bytes left; set up for a retry later. */
 482         if (status > 0)
 483                 status = -EAGAIN;
 484
 485         switch (status) {
 486         case -ENETUNREACH:
 487         case -EPIPE:
 488         case -ECONNREFUSED:
 489                 /* When the server has died, an ICMP port unreachable message
 490                  * prompts ECONNREFUSED. */
 491                 break;
 492         case -EAGAIN:
 493                 xs_nospace(task);
 494                 break;
 495         default:
 496                 dprintk("RPC:       sendmsg returned unrecognized error %d\n",
 497                         -status);
 498                 break;
 499         }
 500
 501         return status;
 502 }
 503
 504 static inline void xs_encode_tcp_record_marker(struct xdr_buf *buf)
 505 {
 506         u32 reclen = buf->len - sizeof(rpc_fraghdr);
 507         rpc_fraghdr *base = buf->head[0].iov_base;
 508         *base = htonl(RPC_LAST_STREAM_FRAGMENT | reclen);
 509 }
 510
 511 /**
 512  * xs_tcp_send_request - write an RPC request to a TCP socket
 513  * @task: address of RPC task that manages the state of an RPC request
 514  *
 515  * Return values:
 516  *        0:    The request has been sent
 517  *   EAGAIN:    The socket was blocked, please call again later to
 518  *              complete the request
 519  * ENOTCONN:    Caller needs to invoke connect logic then call again
 520  *    other:    Some other error occured, the request was not sent
 521  *
 522  * XXX: In the case of soft timeouts, should we eventually give up
 523  *      if sendmsg is not able to make progress?
 524  */
 525 static int xs_tcp_send_request(struct rpc_task *task)
 526 {
 527         struct rpc_rqst *req = task->tk_rqstp;
 528         struct rpc_xprt *xprt = req->rq_xprt;
 529         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 530         struct xdr_buf *xdr = &req->rq_snd_buf;
 531         int status;
 532         unsigned int retry = 0;
 533
 534         xs_encode_tcp_record_marker(&req->rq_snd_buf);
 535
 536         xs_pktdump("packet data:",
 537                                 req->rq_svec->iov_base,
 538                                 req->rq_svec->iov_len);
 539
 540         /* Continue transmitting the packet/record. We must be careful
 541          * to cope with writespace callbacks arriving _after_ we have
 542          * called sendmsg(). */
 543         while (1) {
 544                 req->rq_xtime = jiffies;
 545                 status = xs_sendpages(transport->sock,
 546                                         NULL, 0, xdr, req->rq_bytes_sent);
 547
 548                 dprintk("RPC:       xs_tcp_send_request(%u) = %d\n",
 549                                 xdr->len - req->rq_bytes_sent, status);
 550
 551                 if (unlikely(status < 0))
 552                         break;
 553
 554                 /* If we've sent the entire packet, immediately
 555                  * reset the count of bytes sent. */
 556                 req->rq_bytes_sent += status;
 557                 task->tk_bytes_sent += status;
 558                 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
 559                         req->rq_bytes_sent = 0;
 560                         return 0;
 561                 }
 562
 563                 status = -EAGAIN;
 564                 if (retry++ > XS_SENDMSG_RETRY)
 565                         break;
 566         }
 567
 568         switch (status) {
 569         case -EAGAIN:
 570                 xs_nospace(task);
 571                 break;
 572         case -ECONNREFUSED:
 573         case -ECONNRESET:
 574         case -ENOTCONN:
 575         case -EPIPE:
 576                 status = -ENOTCONN;
 577                 break;
 578         default:
 579                 dprintk("RPC:       sendmsg returned unrecognized error %d\n",
 580                         -status);
 581                 xprt_disconnect(xprt);
 582                 break;
 583         }
 584
 585         return status;
 586 }
 587
 588 /**
 589  * xs_tcp_release_xprt - clean up after a tcp transmission
 590  * @xprt: transport
 591  * @task: rpc task
 592  *
 593  * This cleans up if an error causes us to abort the transmission of a request.
 594  * In this case, the socket may need to be reset in order to avoid confusing
 595  * the server.
 596  */
 597 static void xs_tcp_release_xprt(struct rpc_xprt *xprt, struct rpc_task *task)
 598 {
 599         struct rpc_rqst *req;
 600
 601         if (task != xprt->snd_task)
 602                 return;
 603         if (task == NULL)
 604                 goto out_release;
 605         req = task->tk_rqstp;
 606         if (req->rq_bytes_sent == 0)
 607                 goto out_release;
 608         if (req->rq_bytes_sent == req->rq_snd_buf.len)
 609                 goto out_release;
 610         set_bit(XPRT_CLOSE_WAIT, &task->tk_xprt->state);
 611 out_release:
 612         xprt_release_xprt(xprt, task);
 613 }
 614
 615 /**
 616  * xs_close - close a socket
 617  * @xprt: transport
 618  *
 619  * This is used when all requests are complete; ie, no DRC state remains
 620  * on the server we want to save.
 621  */
 622 static void xs_close(struct rpc_xprt *xprt)
 623 {
 624         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 625         struct socket *sock = transport->sock;
 626         struct sock *sk = transport->inet;
 627
 628         if (!sk)
 629                 goto clear_close_wait;
 630
 631         dprintk("RPC:       xs_close xprt %p\n", xprt);
 632
 633         write_lock_bh(&sk->sk_callback_lock);
 634         transport->inet = NULL;
 635         transport->sock = NULL;
 636
 637         sk->sk_user_data = NULL;
 638         sk->sk_data_ready = transport->old_data_ready;
 639         sk->sk_state_change = transport->old_state_change;
 640         sk->sk_write_space = transport->old_write_space;
 641         write_unlock_bh(&sk->sk_callback_lock);
 642
 643         sk->sk_no_check = 0;
 644
 645         sock_release(sock);
 646 clear_close_wait:
 647         smp_mb__before_clear_bit();
 648         clear_bit(XPRT_CLOSE_WAIT, &xprt->state);
 649         smp_mb__after_clear_bit();
 650 }
 651
 652 /**
 653  * xs_destroy - prepare to shutdown a transport
 654  * @xprt: doomed transport
 655  *
 656  */
 657 static void xs_destroy(struct rpc_xprt *xprt)
 658 {
 659         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 660
 661         dprintk("RPC:       xs_destroy xprt %p\n", xprt);
 662
 663         cancel_rearming_delayed_work(&transport->connect_worker);
 664
 665         xprt_disconnect(xprt);
 666         xs_close(xprt);
 667         xs_free_peer_addresses(xprt);
 668         kfree(xprt->slot);
 669         kfree(xprt);
 670 }
 671
 672 static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
 673 {
 674         return (struct rpc_xprt *) sk->sk_user_data;
 675 }
 676
 677 /**
 678  * xs_udp_data_ready - "data ready" callback for UDP sockets
 679  * @sk: socket with data to read
 680  * @len: how much data to read
 681  *
 682  */
 683 static void xs_udp_data_ready(struct sock *sk, int len)
 684 {
 685         struct rpc_task *task;
 686         struct rpc_xprt *xprt;
 687         struct rpc_rqst *rovr;
 688         struct sk_buff *skb;
 689         int err, repsize, copied;
 690         u32 _xid;
 691         __be32 *xp;
 692
 693         read_lock(&sk->sk_callback_lock);
 694         dprintk("RPC:       xs_udp_data_ready...\n");
 695         if (!(xprt = xprt_from_sock(sk)))
 696                 goto out;
 697
 698         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 699                 goto out;
 700
 701         if (xprt->shutdown)
 702                 goto dropit;
 703
 704         repsize = skb->len - sizeof(struct udphdr);
 705         if (repsize < 4) {
 706                 dprintk("RPC:       impossible RPC reply size %d!\n", repsize);
 707                 goto dropit;
 708         }
 709
 710         /* Copy the XID from the skb... */
 711         xp = skb_header_pointer(skb, sizeof(struct udphdr),
 712                                 sizeof(_xid), &_xid);
 713         if (xp == NULL)
 714                 goto dropit;
 715
 716         /* Look up and lock the request corresponding to the given XID */
 717         spin_lock(&xprt->transport_lock);
 718         rovr = xprt_lookup_rqst(xprt, *xp);
 719         if (!rovr)
 720                 goto out_unlock;
 721         task = rovr->rq_task;
 722
 723         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 724                 copied = repsize;
 725
 726         /* Suck it into the iovec, verify checksum if not done by hw. */
 727         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 728                 goto out_unlock;
 729
 730         /* Something worked... */
 731         dst_confirm(skb->dst);
 732
 733         xprt_adjust_cwnd(task, copied);
 734         xprt_update_rtt(task);
 735         xprt_complete_rqst(task, copied);
 736
 737  out_unlock:
 738         spin_unlock(&xprt->transport_lock);
 739  dropit:
 740         skb_free_datagram(sk, skb);
 741  out:
 742         read_unlock(&sk->sk_callback_lock);
 743 }
 744
 745 static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
 746 {
 747         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 748         size_t len, used;
 749         char *p;
 750
 751         p = ((char *) &transport->tcp_fraghdr) + transport->tcp_offset;
 752         len = sizeof(transport->tcp_fraghdr) - transport->tcp_offset;
 753         used = xdr_skb_read_bits(desc, p, len);
 754         transport->tcp_offset += used;
 755         if (used != len)
 756                 return;
 757
 758         transport->tcp_reclen = ntohl(transport->tcp_fraghdr);
 759         if (transport->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
 760                 transport->tcp_flags |= TCP_RCV_LAST_FRAG;
 761         else
 762                 transport->tcp_flags &= ~TCP_RCV_LAST_FRAG;
 763         transport->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
 764
 765         transport->tcp_flags &= ~TCP_RCV_COPY_FRAGHDR;
 766         transport->tcp_offset = 0;
 767
 768         /* Sanity check of the record length */
 769         if (unlikely(transport->tcp_reclen < 4)) {
 770                 dprintk("RPC:       invalid TCP record fragment length\n");
 771                 xprt_disconnect(xprt);
 772                 return;
 773         }
 774         dprintk("RPC:       reading TCP record fragment of length %d\n",
 775                         transport->tcp_reclen);
 776 }
 777
 778 static void xs_tcp_check_fraghdr(struct sock_xprt *transport)
 779 {
 780         if (transport->tcp_offset == transport->tcp_reclen) {
 781                 transport->tcp_flags |= TCP_RCV_COPY_FRAGHDR;
 782                 transport->tcp_offset = 0;
 783                 if (transport->tcp_flags & TCP_RCV_LAST_FRAG) {
 784                         transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 785                         transport->tcp_flags |= TCP_RCV_COPY_XID;
 786                         transport->tcp_copied = 0;
 787                 }
 788         }
 789 }
 790
 791 static inline void xs_tcp_read_xid(struct sock_xprt *transport, struct xdr_skb_reader *desc)
 792 {
 793         size_t len, used;
 794         char *p;
 795
 796         len = sizeof(transport->tcp_xid) - transport->tcp_offset;
 797         dprintk("RPC:       reading XID (%Zu bytes)\n", len);
 798         p = ((char *) &transport->tcp_xid) + transport->tcp_offset;
 799         used = xdr_skb_read_bits(desc, p, len);
 800         transport->tcp_offset += used;
 801         if (used != len)
 802                 return;
 803         transport->tcp_flags &= ~TCP_RCV_COPY_XID;
 804         transport->tcp_flags |= TCP_RCV_COPY_DATA;
 805         transport->tcp_copied = 4;
 806         dprintk("RPC:       reading reply for XID %08x\n",
 807                         ntohl(transport->tcp_xid));
 808         xs_tcp_check_fraghdr(transport);
 809 }
 810
 811 static inline void xs_tcp_read_request(struct rpc_xprt *xprt, struct xdr_skb_reader *desc)
 812 {
 813         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 814         struct rpc_rqst *req;
 815         struct xdr_buf *rcvbuf;
 816         size_t len;
 817         ssize_t r;
 818
 819         /* Find and lock the request corresponding to this xid */
 820         spin_lock(&xprt->transport_lock);
 821         req = xprt_lookup_rqst(xprt, transport->tcp_xid);
 822         if (!req) {
 823                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 824                 dprintk("RPC:       XID %08x request not found!\n",
 825                                 ntohl(transport->tcp_xid));
 826                 spin_unlock(&xprt->transport_lock);
 827                 return;
 828         }
 829
 830         rcvbuf = &req->rq_private_buf;
 831         len = desc->count;
 832         if (len > transport->tcp_reclen - transport->tcp_offset) {
 833                 struct xdr_skb_reader my_desc;
 834
 835                 len = transport->tcp_reclen - transport->tcp_offset;
 836                 memcpy(&my_desc, desc, sizeof(my_desc));
 837                 my_desc.count = len;
 838                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
 839                                           &my_desc, xdr_skb_read_bits);
 840                 desc->count -= r;
 841                 desc->offset += r;
 842         } else
 843                 r = xdr_partial_copy_from_skb(rcvbuf, transport->tcp_copied,
 844                                           desc, xdr_skb_read_bits);
 845
 846         if (r > 0) {
 847                 transport->tcp_copied += r;
 848                 transport->tcp_offset += r;
 849         }
 850         if (r != len) {
 851                 /* Error when copying to the receive buffer,
 852                  * usually because we weren't able to allocate
 853                  * additional buffer pages. All we can do now
 854                  * is turn off TCP_RCV_COPY_DATA, so the request
 855                  * will not receive any additional updates,
 856                  * and time out.
 857                  * Any remaining data from this record will
 858                  * be discarded.
 859                  */
 860                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 861                 dprintk("RPC:       XID %08x truncated request\n",
 862                                 ntohl(transport->tcp_xid));
 863                 dprintk("RPC:       xprt = %p, tcp_copied = %lu, "
 864                                 "tcp_offset = %u, tcp_reclen = %u\n",
 865                                 xprt, transport->tcp_copied,
 866                                 transport->tcp_offset, transport->tcp_reclen);
 867                 goto out;
 868         }
 869
 870         dprintk("RPC:       XID %08x read %Zd bytes\n",
 871                         ntohl(transport->tcp_xid), r);
 872         dprintk("RPC:       xprt = %p, tcp_copied = %lu, tcp_offset = %u, "
 873                         "tcp_reclen = %u\n", xprt, transport->tcp_copied,
 874                         transport->tcp_offset, transport->tcp_reclen);
 875
 876         if (transport->tcp_copied == req->rq_private_buf.buflen)
 877                 transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 878         else if (transport->tcp_offset == transport->tcp_reclen) {
 879                 if (transport->tcp_flags & TCP_RCV_LAST_FRAG)
 880                         transport->tcp_flags &= ~TCP_RCV_COPY_DATA;
 881         }
 882
 883 out:
 884         if (!(transport->tcp_flags & TCP_RCV_COPY_DATA))
 885                 xprt_complete_rqst(req->rq_task, transport->tcp_copied);
 886         spin_unlock(&xprt->transport_lock);
 887         xs_tcp_check_fraghdr(transport);
 888 }
 889
 890 static inline void xs_tcp_read_discard(struct sock_xprt *transport, struct xdr_skb_reader *desc)
 891 {
 892         size_t len;
 893
 894         len = transport->tcp_reclen - transport->tcp_offset;
 895         if (len > desc->count)
 896                 len = desc->count;
 897         desc->count -= len;
 898         desc->offset += len;
 899         transport->tcp_offset += len;
 900         dprintk("RPC:       discarded %Zu bytes\n", len);
 901         xs_tcp_check_fraghdr(transport);
 902 }
 903
 904 static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
 905 {
 906         struct rpc_xprt *xprt = rd_desc->arg.data;
 907         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
 908         struct xdr_skb_reader desc = {
 909                 .skb    = skb,
 910                 .offset = offset,
 911                 .count  = len,
 912         };
 913
 914         dprintk("RPC:       xs_tcp_data_recv started\n");
 915         do {
 916                 /* Read in a new fragment marker if necessary */
 917                 /* Can we ever really expect to get completely empty fragments? */
 918                 if (transport->tcp_flags & TCP_RCV_COPY_FRAGHDR) {
 919                         xs_tcp_read_fraghdr(xprt, &desc);
 920                         continue;
 921                 }
 922                 /* Read in the xid if necessary */
 923                 if (transport->tcp_flags & TCP_RCV_COPY_XID) {
 924                         xs_tcp_read_xid(transport, &desc);
 925                         continue;
 926                 }
 927                 /* Read in the request data */
 928                 if (transport->tcp_flags & TCP_RCV_COPY_DATA) {
 929                         xs_tcp_read_request(xprt, &desc);
 930                         continue;
 931                 }
 932                 /* Skip over any trailing bytes on short reads */
 933                 xs_tcp_read_discard(transport, &desc);
 934         } while (desc.count);
 935         dprintk("RPC:       xs_tcp_data_recv done\n");
 936         return len - desc.count;
 937 }
 938
 939 /**
 940  * xs_tcp_data_ready - "data ready" callback for TCP sockets
 941  * @sk: socket with data to read
 942  * @bytes: how much data to read
 943  *
 944  */
 945 static void xs_tcp_data_ready(struct sock *sk, int bytes)
 946 {
 947         struct rpc_xprt *xprt;
 948         read_descriptor_t rd_desc;
 949
 950         dprintk("RPC:       xs_tcp_data_ready...\n");
 951
 952         read_lock(&sk->sk_callback_lock);
 953         if (!(xprt = xprt_from_sock(sk)))
 954                 goto out;
 955         if (xprt->shutdown)
 956                 goto out;
 957
 958         /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
 959         rd_desc.arg.data = xprt;
 960         rd_desc.count = 65536;
 961         tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
 962 out:
 963         read_unlock(&sk->sk_callback_lock);
 964 }
 965
 966 /**
 967  * xs_tcp_state_change - callback to handle TCP socket state changes
 968  * @sk: socket whose state has changed
 969  *
 970  */
 971 static void xs_tcp_state_change(struct sock *sk)
 972 {
 973         struct rpc_xprt *xprt;
 974
 975         read_lock(&sk->sk_callback_lock);
 976         if (!(xprt = xprt_from_sock(sk)))
 977                 goto out;
 978         dprintk("RPC:       xs_tcp_state_change client %p...\n", xprt);
 979         dprintk("RPC:       state %x conn %d dead %d zapped %d\n",
 980                         sk->sk_state, xprt_connected(xprt),
 981                         sock_flag(sk, SOCK_DEAD),
 982                         sock_flag(sk, SOCK_ZAPPED));
 983
 984         switch (sk->sk_state) {
 985         case TCP_ESTABLISHED:
 986                 spin_lock_bh(&xprt->transport_lock);
 987                 if (!xprt_test_and_set_connected(xprt)) {
 988                         struct sock_xprt *transport = container_of(xprt,
 989                                         struct sock_xprt, xprt);
 990
 991                         /* Reset TCP record info */
 992                         transport->tcp_offset = 0;
 993                         transport->tcp_reclen = 0;
 994                         transport->tcp_copied = 0;
 995                         transport->tcp_flags =
 996                                 TCP_RCV_COPY_FRAGHDR | TCP_RCV_COPY_XID;
 997
 998                         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
 999                         xprt_wake_pending_tasks(xprt, 0);
1000                 }
1001                 spin_unlock_bh(&xprt->transport_lock);
1002                 break;
1003         case TCP_SYN_SENT:
1004         case TCP_SYN_RECV:
1005                 break;
1006         case TCP_CLOSE_WAIT:
1007                 /* Try to schedule an autoclose RPC calls */
1008                 set_bit(XPRT_CLOSE_WAIT, &xprt->state);
1009                 if (test_and_set_bit(XPRT_LOCKED, &xprt->state) == 0)
1010                         queue_work(rpciod_workqueue, &xprt->task_cleanup);
1011         default:
1012                 xprt_disconnect(xprt);
1013         }
1014  out:
1015         read_unlock(&sk->sk_callback_lock);
1016 }
1017
1018 /**
1019  * xs_udp_write_space - callback invoked when socket buffer space
1020  *                             becomes available
1021  * @sk: socket whose state has changed
1022  *
1023  * Called when more output buffer space is available for this socket.
1024  * We try not to wake our writers until they can make "significant"
1025  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1026  * with a bunch of small requests.
1027  */
1028 static void xs_udp_write_space(struct sock *sk)
1029 {
1030         read_lock(&sk->sk_callback_lock);
1031
1032         /* from net/core/sock.c:sock_def_write_space */
1033         if (sock_writeable(sk)) {
1034                 struct socket *sock;
1035                 struct rpc_xprt *xprt;
1036
1037                 if (unlikely(!(sock = sk->sk_socket)))
1038                         goto out;
1039                 if (unlikely(!(xprt = xprt_from_sock(sk))))
1040                         goto out;
1041                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
1042                         goto out;
1043
1044                 xprt_write_space(xprt);
1045         }
1046
1047  out:
1048         read_unlock(&sk->sk_callback_lock);
1049 }
1050
1051 /**
1052  * xs_tcp_write_space - callback invoked when socket buffer space
1053  *                             becomes available
1054  * @sk: socket whose state has changed
1055  *
1056  * Called when more output buffer space is available for this socket.
1057  * We try not to wake our writers until they can make "significant"
1058  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
1059  * with a bunch of small requests.
1060  */
1061 static void xs_tcp_write_space(struct sock *sk)
1062 {
1063         read_lock(&sk->sk_callback_lock);
1064
1065         /* from net/core/stream.c:sk_stream_write_space */
1066         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
1067                 struct socket *sock;
1068                 struct rpc_xprt *xprt;
1069
1070                 if (unlikely(!(sock = sk->sk_socket)))
1071                         goto out;
1072                 if (unlikely(!(xprt = xprt_from_sock(sk))))
1073                         goto out;
1074                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
1075                         goto out;
1076
1077                 xprt_write_space(xprt);
1078         }
1079
1080  out:
1081         read_unlock(&sk->sk_callback_lock);
1082 }
1083
1084 static void xs_udp_do_set_buffer_size(struct rpc_xprt *xprt)
1085 {
1086         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1087         struct sock *sk = transport->inet;
1088
1089         if (transport->rcvsize) {
1090                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
1091                 sk->sk_rcvbuf = transport->rcvsize * xprt->max_reqs * 2;
1092         }
1093         if (transport->sndsize) {
1094                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
1095                 sk->sk_sndbuf = transport->sndsize * xprt->max_reqs * 2;
1096                 sk->sk_write_space(sk);
1097         }
1098 }
1099
1100 /**
1101  * xs_udp_set_buffer_size - set send and receive limits
1102  * @xprt: generic transport
1103  * @sndsize: requested size of send buffer, in bytes
1104  * @rcvsize: requested size of receive buffer, in bytes
1105  *
1106  * Set socket send and receive buffer size limits.
1107  */
1108 static void xs_udp_set_buffer_size(struct rpc_xprt *xprt, size_t sndsize, size_t rcvsize)
1109 {
1110         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1111
1112         transport->sndsize = 0;
1113         if (sndsize)
1114                 transport->sndsize = sndsize + 1024;
1115         transport->rcvsize = 0;
1116         if (rcvsize)
1117                 transport->rcvsize = rcvsize + 1024;
1118
1119         xs_udp_do_set_buffer_size(xprt);
1120 }
1121
1122 /**
1123  * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
1124  * @task: task that timed out
1125  *
1126  * Adjust the congestion window after a retransmit timeout has occurred.
1127  */
1128 static void xs_udp_timer(struct rpc_task *task)
1129 {
1130         xprt_adjust_cwnd(task, -ETIMEDOUT);
1131 }
1132
1133 static unsigned short xs_get_random_port(void)
1134 {
1135         unsigned short range = xprt_max_resvport - xprt_min_resvport;
1136         unsigned short rand = (unsigned short) net_random() % range;
1137         return rand + xprt_min_resvport;
1138 }
1139
1140 /**
1141  * xs_set_port - reset the port number in the remote endpoint address
1142  * @xprt: generic transport
1143  * @port: new port number
1144  *
1145  */
1146 static void xs_set_port(struct rpc_xprt *xprt, unsigned short port)
1147 {
1148         struct sockaddr_in *sap = (struct sockaddr_in *) &xprt->addr;
1149
1150         dprintk("RPC:       setting port for xprt %p to %u\n", xprt, port);
1151
1152         sap->sin_port = htons(port);
1153 }
1154
1155 static int xs_bind(struct sock_xprt *transport, struct socket *sock)
1156 {
1157         struct sockaddr_in myaddr = {
1158                 .sin_family = AF_INET,
1159         };
1160         struct sockaddr_in *sa;
1161         int err;
1162         unsigned short port = transport->port;
1163
1164         if (!transport->xprt.resvport)
1165                 port = 0;
1166         sa = (struct sockaddr_in *)&transport->addr;
1167         myaddr.sin_addr = sa->sin_addr;
1168         do {
1169                 myaddr.sin_port = htons(port);
1170                 err = kernel_bind(sock, (struct sockaddr *) &myaddr,
1171                                                 sizeof(myaddr));
1172                 if (!transport->xprt.resvport)
1173                         break;
1174                 if (err == 0) {
1175                         transport->port = port;
1176                         break;
1177                 }
1178                 if (port <= xprt_min_resvport)
1179                         port = xprt_max_resvport;
1180                 else
1181                         port--;
1182         } while (err == -EADDRINUSE && port != transport->port);
1183         dprintk("RPC:       xs_bind "NIPQUAD_FMT":%u: %s (%d)\n",
1184                 NIPQUAD(myaddr.sin_addr), port, err ? "failed" : "ok", err);
1185         return err;
1186 }
1187
1188 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1189 static struct lock_class_key xs_key[2];
1190 static struct lock_class_key xs_slock_key[2];
1191
1192 static inline void xs_reclassify_socket(struct socket *sock)
1193 {
1194         struct sock *sk = sock->sk;
1195         BUG_ON(sk->sk_lock.owner != NULL);
1196         switch (sk->sk_family) {
1197         case AF_INET:
1198                 sock_lock_init_class_and_name(sk, "slock-AF_INET-NFS",
1199                         &xs_slock_key[0], "sk_lock-AF_INET-NFS", &xs_key[0]);
1200                 break;
1201
1202         case AF_INET6:
1203                 sock_lock_init_class_and_name(sk, "slock-AF_INET6-NFS",
1204                         &xs_slock_key[1], "sk_lock-AF_INET6-NFS", &xs_key[1]);
1205                 break;
1206
1207         default:
1208                 BUG();
1209         }
1210 }
1211 #else
1212 static inline void xs_reclassify_socket(struct socket *sock)
1213 {
1214 }
1215 #endif
1216
1217 /**
1218  * xs_udp_connect_worker - set up a UDP socket
1219  * @work: RPC transport to connect
1220  *
1221  * Invoked by a work queue tasklet.
1222  */
1223 static void xs_udp_connect_worker(struct work_struct *work)
1224 {
1225         struct sock_xprt *transport =
1226                 container_of(work, struct sock_xprt, connect_worker.work);
1227         struct rpc_xprt *xprt = &transport->xprt;
1228         struct socket *sock = transport->sock;
1229         int err, status = -EIO;
1230
1231         if (xprt->shutdown || !xprt_bound(xprt))
1232                 goto out;
1233
1234         /* Start by resetting any existing state */
1235         xs_close(xprt);
1236
1237         if ((err = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock)) < 0) {
1238                 dprintk("RPC:       can't create UDP transport socket (%d).\n", -err);
1239                 goto out;
1240         }
1241         xs_reclassify_socket(sock);
1242
1243         if (xs_bind(transport, sock)) {
1244                 sock_release(sock);
1245                 goto out;
1246         }
1247
1248         dprintk("RPC:       worker connecting xprt %p to address: %s\n",
1249                         xprt, xprt->address_strings[RPC_DISPLAY_ALL]);
1250
1251         if (!transport->inet) {
1252                 struct sock *sk = sock->sk;
1253
1254                 write_lock_bh(&sk->sk_callback_lock);
1255
1256                 sk->sk_user_data = xprt;
1257                 transport->old_data_ready = sk->sk_data_ready;
1258                 transport->old_state_change = sk->sk_state_change;
1259                 transport->old_write_space = sk->sk_write_space;
1260                 sk->sk_data_ready = xs_udp_data_ready;
1261                 sk->sk_write_space = xs_udp_write_space;
1262                 sk->sk_no_check = UDP_CSUM_NORCV;
1263                 sk->sk_allocation = GFP_ATOMIC;
1264
1265                 xprt_set_connected(xprt);
1266
1267                 /* Reset to new socket */
1268                 transport->sock = sock;
1269                 transport->inet = sk;
1270
1271                 write_unlock_bh(&sk->sk_callback_lock);
1272         }
1273         xs_udp_do_set_buffer_size(xprt);
1274         status = 0;
1275 out:
1276         xprt_wake_pending_tasks(xprt, status);
1277         xprt_clear_connecting(xprt);
1278 }
1279
1280 /*
1281  * We need to preserve the port number so the reply cache on the server can
1282  * find our cached RPC replies when we get around to reconnecting.
1283  */
1284 static void xs_tcp_reuse_connection(struct rpc_xprt *xprt)
1285 {
1286         int result;
1287         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1288         struct sockaddr any;
1289
1290         dprintk("RPC:       disconnecting xprt %p to reuse port\n", xprt);
1291
1292         /*
1293          * Disconnect the transport socket by doing a connect operation
1294          * with AF_UNSPEC.  This should return immediately...
1295          */
1296         memset(&any, 0, sizeof(any));
1297         any.sa_family = AF_UNSPEC;
1298         result = kernel_connect(transport->sock, &any, sizeof(any), 0);
1299         if (result)
1300                 dprintk("RPC:       AF_UNSPEC connect return code %d\n",
1301                                 result);
1302 }
1303
1304 /**
1305  * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1306  * @work: RPC transport to connect
1307  *
1308  * Invoked by a work queue tasklet.
1309  */
1310 static void xs_tcp_connect_worker(struct work_struct *work)
1311 {
1312         struct sock_xprt *transport =
1313                 container_of(work, struct sock_xprt, connect_worker.work);
1314         struct rpc_xprt *xprt = &transport->xprt;
1315         struct socket *sock = transport->sock;
1316         int err, status = -EIO;
1317
1318         if (xprt->shutdown || !xprt_bound(xprt))
1319                 goto out;
1320
1321         if (!sock) {
1322                 /* start from scratch */
1323                 if ((err = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock)) < 0) {
1324                         dprintk("RPC:       can't create TCP transport "
1325                                         "socket (%d).\n", -err);
1326                         goto out;
1327                 }
1328                 xs_reclassify_socket(sock);
1329
1330                 if (xs_bind(transport, sock)) {
1331                         sock_release(sock);
1332                         goto out;
1333                 }
1334         } else
1335                 /* "close" the socket, preserving the local port */
1336                 xs_tcp_reuse_connection(xprt);
1337
1338         dprintk("RPC:       worker connecting xprt %p to address: %s\n",
1339                         xprt, xprt->address_strings[RPC_DISPLAY_ALL]);
1340
1341         if (!transport->inet) {
1342                 struct sock *sk = sock->sk;
1343
1344                 write_lock_bh(&sk->sk_callback_lock);
1345
1346                 sk->sk_user_data = xprt;
1347                 transport->old_data_ready = sk->sk_data_ready;
1348                 transport->old_state_change = sk->sk_state_change;
1349                 transport->old_write_space = sk->sk_write_space;
1350                 sk->sk_data_ready = xs_tcp_data_ready;
1351                 sk->sk_state_change = xs_tcp_state_change;
1352                 sk->sk_write_space = xs_tcp_write_space;
1353                 sk->sk_allocation = GFP_ATOMIC;
1354
1355                 /* socket options */
1356                 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
1357                 sock_reset_flag(sk, SOCK_LINGER);
1358                 tcp_sk(sk)->linger2 = 0;
1359                 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1360
1361                 xprt_clear_connected(xprt);
1362
1363                 /* Reset to new socket */
1364                 transport->sock = sock;
1365                 transport->inet = sk;
1366
1367                 write_unlock_bh(&sk->sk_callback_lock);
1368         }
1369
1370         /* Tell the socket layer to start connecting... */
1371         xprt->stat.connect_count++;
1372         xprt->stat.connect_start = jiffies;
1373         status = kernel_connect(sock, (struct sockaddr *) &xprt->addr,
1374                         xprt->addrlen, O_NONBLOCK);
1375         dprintk("RPC:       %p connect status %d connected %d sock state %d\n",
1376                         xprt, -status, xprt_connected(xprt),
1377                         sock->sk->sk_state);
1378         if (status < 0) {
1379                 switch (status) {
1380                         case -EINPROGRESS:
1381                         case -EALREADY:
1382                                 goto out_clear;
1383                         case -ECONNREFUSED:
1384                         case -ECONNRESET:
1385                                 /* retry with existing socket, after a delay */
1386                                 break;
1387                         default:
1388                                 /* get rid of existing socket, and retry */
1389                                 xs_close(xprt);
1390                                 break;
1391                 }
1392         }
1393 out:
1394         xprt_wake_pending_tasks(xprt, status);
1395 out_clear:
1396         xprt_clear_connecting(xprt);
1397 }
1398
1399 /**
1400  * xs_connect - connect a socket to a remote endpoint
1401  * @task: address of RPC task that manages state of connect request
1402  *
1403  * TCP: If the remote end dropped the connection, delay reconnecting.
1404  *
1405  * UDP socket connects are synchronous, but we use a work queue anyway
1406  * to guarantee that even unprivileged user processes can set up a
1407  * socket on a privileged port.
1408  *
1409  * If a UDP socket connect fails, the delay behavior here prevents
1410  * retry floods (hard mounts).
1411  */
1412 static void xs_connect(struct rpc_task *task)
1413 {
1414         struct rpc_xprt *xprt = task->tk_xprt;
1415         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1416
1417         if (xprt_test_and_set_connecting(xprt))
1418                 return;
1419
1420         if (transport->sock != NULL) {
1421                 dprintk("RPC:       xs_connect delayed xprt %p for %lu "
1422                                 "seconds\n",
1423                                 xprt, xprt->reestablish_timeout / HZ);
1424                 queue_delayed_work(rpciod_workqueue,
1425                                    &transport->connect_worker,
1426                                    xprt->reestablish_timeout);
1427                 xprt->reestablish_timeout <<= 1;
1428                 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
1429                         xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
1430         } else {
1431                 dprintk("RPC:       xs_connect scheduled xprt %p\n", xprt);
1432                 queue_delayed_work(rpciod_workqueue,
1433                                    &transport->connect_worker, 0);
1434         }
1435 }
1436
1437 /**
1438  * xs_udp_print_stats - display UDP socket-specifc stats
1439  * @xprt: rpc_xprt struct containing statistics
1440  * @seq: output file
1441  *
1442  */
1443 static void xs_udp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1444 {
1445         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1446
1447         seq_printf(seq, "\txprt:\tudp %u %lu %lu %lu %lu %Lu %Lu\n",
1448                         transport->port,
1449                         xprt->stat.bind_count,
1450                         xprt->stat.sends,
1451                         xprt->stat.recvs,
1452                         xprt->stat.bad_xids,
1453                         xprt->stat.req_u,
1454                         xprt->stat.bklog_u);
1455 }
1456
1457 /**
1458  * xs_tcp_print_stats - display TCP socket-specifc stats
1459  * @xprt: rpc_xprt struct containing statistics
1460  * @seq: output file
1461  *
1462  */
1463 static void xs_tcp_print_stats(struct rpc_xprt *xprt, struct seq_file *seq)
1464 {
1465         struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt);
1466         long idle_time = 0;
1467
1468         if (xprt_connected(xprt))
1469                 idle_time = (long)(jiffies - xprt->last_used) / HZ;
1470
1471         seq_printf(seq, "\txprt:\ttcp %u %lu %lu %lu %ld %lu %lu %lu %Lu %Lu\n",
1472                         transport->port,
1473                         xprt->stat.bind_count,
1474                         xprt->stat.connect_count,
1475                         xprt->stat.connect_time,
1476                         idle_time,
1477                         xprt->stat.sends,
1478                         xprt->stat.recvs,
1479                         xprt->stat.bad_xids,
1480                         xprt->stat.req_u,
1481                         xprt->stat.bklog_u);
1482 }
1483
1484 static struct rpc_xprt_ops xs_udp_ops = {
1485         .set_buffer_size        = xs_udp_set_buffer_size,
1486         .reserve_xprt           = xprt_reserve_xprt_cong,
1487         .release_xprt           = xprt_release_xprt_cong,
1488         .rpcbind                = rpcb_getport_async,
1489         .set_port               = xs_set_port,
1490         .connect                = xs_connect,
1491         .buf_alloc              = rpc_malloc,
1492         .buf_free               = rpc_free,
1493         .send_request           = xs_udp_send_request,
1494         .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
1495         .timer                  = xs_udp_timer,
1496         .release_request        = xprt_release_rqst_cong,
1497         .close                  = xs_close,
1498         .destroy                = xs_destroy,
1499         .print_stats            = xs_udp_print_stats,
1500 };
1501
1502 static struct rpc_xprt_ops xs_tcp_ops = {
1503         .reserve_xprt           = xprt_reserve_xprt,
1504         .release_xprt           = xs_tcp_release_xprt,
1505         .rpcbind                = rpcb_getport_async,
1506         .set_port               = xs_set_port,
1507         .connect                = xs_connect,
1508         .buf_alloc              = rpc_malloc,
1509         .buf_free               = rpc_free,
1510         .send_request           = xs_tcp_send_request,
1511         .set_retrans_timeout    = xprt_set_retrans_timeout_def,
1512         .close                  = xs_close,
1513         .destroy                = xs_destroy,
1514         .print_stats            = xs_tcp_print_stats,
1515 };
1516
1517 static struct rpc_xprt *xs_setup_xprt(struct rpc_xprtsock_create *args, unsigned int slot_table_size)
1518 {
1519         struct rpc_xprt *xprt;
1520         struct sock_xprt *new;
1521
1522         if (args->addrlen > sizeof(xprt->addr)) {
1523                 dprintk("RPC:       xs_setup_xprt: address too large\n");
1524                 return ERR_PTR(-EBADF);
1525         }
1526
1527         new = kzalloc(sizeof(*new), GFP_KERNEL);
1528         if (new == NULL) {
1529                 dprintk("RPC:       xs_setup_xprt: couldn't allocate "
1530                                 "rpc_xprt\n");
1531                 return ERR_PTR(-ENOMEM);
1532         }
1533         xprt = &new->xprt;
1534
1535         xprt->max_reqs = slot_table_size;
1536         xprt->slot = kcalloc(xprt->max_reqs, sizeof(struct rpc_rqst), GFP_KERNEL);
1537         if (xprt->slot == NULL) {
1538                 kfree(xprt);
1539                 dprintk("RPC:       xs_setup_xprt: couldn't allocate slot "
1540                                 "table\n");
1541                 return ERR_PTR(-ENOMEM);
1542         }
1543
1544         memcpy(&xprt->addr, args->dstaddr, args->addrlen);
1545         xprt->addrlen = args->addrlen;
1546         if (args->srcaddr)
1547                 memcpy(&new->addr, args->srcaddr, args->addrlen);
1548         new->port = xs_get_random_port();
1549
1550         return xprt;
1551 }
1552
1553 /**
1554  * xs_setup_udp - Set up transport to use a UDP socket
1555  * @args: rpc transport creation arguments
1556  *
1557  */
1558 struct rpc_xprt *xs_setup_udp(struct rpc_xprtsock_create *args)
1559 {
1560         struct rpc_xprt *xprt;
1561         struct sock_xprt *transport;
1562
1563         xprt = xs_setup_xprt(args, xprt_udp_slot_table_entries);
1564         if (IS_ERR(xprt))
1565                 return xprt;
1566         transport = container_of(xprt, struct sock_xprt, xprt);
1567
1568         if (ntohs(((struct sockaddr_in *)args->dstaddr)->sin_port) != 0)
1569                 xprt_set_bound(xprt);
1570
1571         xprt->prot = IPPROTO_UDP;
1572         xprt->tsh_size = 0;
1573         /* XXX: header size can vary due to auth type, IPv6, etc. */
1574         xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1575
1576         INIT_DELAYED_WORK(&transport->connect_worker, xs_udp_connect_worker);
1577         xprt->bind_timeout = XS_BIND_TO;
1578         xprt->connect_timeout = XS_UDP_CONN_TO;
1579         xprt->reestablish_timeout = XS_UDP_REEST_TO;
1580         xprt->idle_timeout = XS_IDLE_DISC_TO;
1581
1582         xprt->ops = &xs_udp_ops;
1583
1584         if (args->timeout)
1585                 xprt->timeout = *args->timeout;
1586         else
1587                 xprt_set_timeout(&xprt->timeout, 5, 5 * HZ);
1588
1589         xs_format_peer_addresses(xprt);
1590         dprintk("RPC:       set up transport to address %s\n",
1591                         xprt->address_strings[RPC_DISPLAY_ALL]);
1592
1593         return xprt;
1594 }
1595
1596 /**
1597  * xs_setup_tcp - Set up transport to use a TCP socket
1598  * @args: rpc transport creation arguments
1599  *
1600  */
1601 struct rpc_xprt *xs_setup_tcp(struct rpc_xprtsock_create *args)
1602 {
1603         struct rpc_xprt *xprt;
1604         struct sock_xprt *transport;
1605
1606         xprt = xs_setup_xprt(args, xprt_tcp_slot_table_entries);
1607         if (IS_ERR(xprt))
1608                 return xprt;
1609         transport = container_of(xprt, struct sock_xprt, xprt);
1610
1611         if (ntohs(((struct sockaddr_in *)args->dstaddr)->sin_port) != 0)
1612                 xprt_set_bound(xprt);
1613
1614         xprt->prot = IPPROTO_TCP;
1615         xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
1616         xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
1617
1618         INIT_DELAYED_WORK(&transport->connect_worker, xs_tcp_connect_worker);
1619         xprt->bind_timeout = XS_BIND_TO;
1620         xprt->connect_timeout = XS_TCP_CONN_TO;
1621         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1622         xprt->idle_timeout = XS_IDLE_DISC_TO;
1623
1624         xprt->ops = &xs_tcp_ops;
1625
1626         if (args->timeout)
1627                 xprt->timeout = *args->timeout;
1628         else
1629                 xprt_set_timeout(&xprt->timeout, 2, 60 * HZ);
1630
1631         xs_format_peer_addresses(xprt);
1632         dprintk("RPC:       set up transport to address %s\n",
1633                         xprt->address_strings[RPC_DISPLAY_ALL]);
1634
1635         return xprt;
1636 }
1637
1638 /**
1639  * init_socket_xprt - set up xprtsock's sysctls
1640  *
1641  */
1642 int init_socket_xprt(void)
1643 {
1644 #ifdef RPC_DEBUG
1645         if (!sunrpc_table_header)
1646                 sunrpc_table_header = register_sysctl_table(sunrpc_table);
1647 #endif
1648
1649         return 0;
1650 }
1651
1652 /**
1653  * cleanup_socket_xprt - remove xprtsock's sysctls
1654  *
1655  */
1656 void cleanup_socket_xprt(void)
1657 {
1658 #ifdef RPC_DEBUG
1659         if (sunrpc_table_header) {
1660                 unregister_sysctl_table(sunrpc_table_header);
1661                 sunrpc_table_header = NULL;
1662         }
1663 #endif
1664 }