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/sched.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/errno.h>
  24 #include <linux/socket.h>
  25 #include <linux/in.h>
  26 #include <linux/net.h>
  27 #include <linux/mm.h>
  28 #include <linux/udp.h>
  29 #include <linux/tcp.h>
  30 #include <linux/sunrpc/clnt.h>
  31 #include <linux/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  * How many times to try sending a request on a socket before waiting
  40  * for the socket buffer to clear.
  41  */
  42 #define XS_SENDMSG_RETRY        (10U)
  43
  44 /*
  45  * Time out for an RPC UDP socket connect.  UDP socket connects are
  46  * synchronous, but we set a timeout anyway in case of resource
  47  * exhaustion on the local host.
  48  */
  49 #define XS_UDP_CONN_TO          (5U * HZ)
  50
  51 /*
  52  * Wait duration for an RPC TCP connection to be established.  Solaris
  53  * NFS over TCP uses 60 seconds, for example, which is in line with how
  54  * long a server takes to reboot.
  55  */
  56 #define XS_TCP_CONN_TO          (60U * HZ)
  57
  58 /*
  59  * Wait duration for a reply from the RPC portmapper.
  60  */
  61 #define XS_BIND_TO              (60U * HZ)
  62
  63 /*
  64  * Delay if a UDP socket connect error occurs.  This is most likely some
  65  * kind of resource problem on the local host.
  66  */
  67 #define XS_UDP_REEST_TO         (2U * HZ)
  68
  69 /*
  70  * The reestablish timeout allows clients to delay for a bit before attempting
  71  * to reconnect to a server that just dropped our connection.
  72  *
  73  * We implement an exponential backoff when trying to reestablish a TCP
  74  * transport connection with the server.  Some servers like to drop a TCP
  75  * connection when they are overworked, so we start with a short timeout and
  76  * increase over time if the server is down or not responding.
  77  */
  78 #define XS_TCP_INIT_REEST_TO    (3U * HZ)
  79 #define XS_TCP_MAX_REEST_TO     (5U * 60 * HZ)
  80
  81 /*
  82  * TCP idle timeout; client drops the transport socket if it is idle
  83  * for this long.  Note that we also timeout UDP sockets to prevent
  84  * holding port numbers when there is no RPC traffic.
  85  */
  86 #define XS_IDLE_DISC_TO         (5U * 60 * HZ)
  87
  88 #ifdef RPC_DEBUG
  89 # undef  RPC_DEBUG_DATA
  90 # define RPCDBG_FACILITY        RPCDBG_TRANS
  91 #endif
  92
  93 #ifdef RPC_DEBUG_DATA
  94 static void xs_pktdump(char *msg, u32 *packet, unsigned int count)
  95 {
  96         u8 *buf = (u8 *) packet;
  97         int j;
  98
  99         dprintk("RPC:      %s\n", msg);
 100         for (j = 0; j < count && j < 128; j += 4) {
 101                 if (!(j & 31)) {
 102                         if (j)
 103                                 dprintk("\n");
 104                         dprintk("0x%04x ", j);
 105                 }
 106                 dprintk("%02x%02x%02x%02x ",
 107                         buf[j], buf[j+1], buf[j+2], buf[j+3]);
 108         }
 109         dprintk("\n");
 110 }
 111 #else
 112 static inline void xs_pktdump(char *msg, u32 *packet, unsigned int count)
 113 {
 114         /* NOP */
 115 }
 116 #endif
 117
 118 #define XS_SENDMSG_FLAGS        (MSG_DONTWAIT | MSG_NOSIGNAL)
 119
 120 static inline int xs_send_head(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base, unsigned int len)
 121 {
 122         struct kvec iov = {
 123                 .iov_base       = xdr->head[0].iov_base + base,
 124                 .iov_len        = len - base,
 125         };
 126         struct msghdr msg = {
 127                 .msg_name       = addr,
 128                 .msg_namelen    = addrlen,
 129                 .msg_flags      = XS_SENDMSG_FLAGS,
 130         };
 131
 132         if (xdr->len > len)
 133                 msg.msg_flags |= MSG_MORE;
 134
 135         if (likely(iov.iov_len))
 136                 return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
 137         return kernel_sendmsg(sock, &msg, NULL, 0, 0);
 138 }
 139
 140 static int xs_send_tail(struct socket *sock, struct xdr_buf *xdr, unsigned int base, unsigned int len)
 141 {
 142         struct kvec iov = {
 143                 .iov_base       = xdr->tail[0].iov_base + base,
 144                 .iov_len        = len - base,
 145         };
 146         struct msghdr msg = {
 147                 .msg_flags      = XS_SENDMSG_FLAGS,
 148         };
 149
 150         return kernel_sendmsg(sock, &msg, &iov, 1, iov.iov_len);
 151 }
 152
 153 /**
 154  * xs_sendpages - write pages directly to a socket
 155  * @sock: socket to send on
 156  * @addr: UDP only -- address of destination
 157  * @addrlen: UDP only -- length of destination address
 158  * @xdr: buffer containing this request
 159  * @base: starting position in the buffer
 160  *
 161  */
 162 static inline int xs_sendpages(struct socket *sock, struct sockaddr *addr, int addrlen, struct xdr_buf *xdr, unsigned int base)
 163 {
 164         struct page **ppage = xdr->pages;
 165         unsigned int len, pglen = xdr->page_len;
 166         int err, ret = 0;
 167         ssize_t (*sendpage)(struct socket *, struct page *, int, size_t, int);
 168
 169         if (unlikely(!sock))
 170                 return -ENOTCONN;
 171
 172         clear_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 173
 174         len = xdr->head[0].iov_len;
 175         if (base < len || (addr != NULL && base == 0)) {
 176                 err = xs_send_head(sock, addr, addrlen, xdr, base, len);
 177                 if (ret == 0)
 178                         ret = err;
 179                 else if (err > 0)
 180                         ret += err;
 181                 if (err != (len - base))
 182                         goto out;
 183                 base = 0;
 184         } else
 185                 base -= len;
 186
 187         if (unlikely(pglen == 0))
 188                 goto copy_tail;
 189         if (unlikely(base >= pglen)) {
 190                 base -= pglen;
 191                 goto copy_tail;
 192         }
 193         if (base || xdr->page_base) {
 194                 pglen -= base;
 195                 base += xdr->page_base;
 196                 ppage += base >> PAGE_CACHE_SHIFT;
 197                 base &= ~PAGE_CACHE_MASK;
 198         }
 199
 200         sendpage = sock->ops->sendpage ? : sock_no_sendpage;
 201         do {
 202                 int flags = XS_SENDMSG_FLAGS;
 203
 204                 len = PAGE_CACHE_SIZE;
 205                 if (base)
 206                         len -= base;
 207                 if (pglen < len)
 208                         len = pglen;
 209
 210                 if (pglen != len || xdr->tail[0].iov_len != 0)
 211                         flags |= MSG_MORE;
 212
 213                 /* Hmm... We might be dealing with highmem pages */
 214                 if (PageHighMem(*ppage))
 215                         sendpage = sock_no_sendpage;
 216                 err = sendpage(sock, *ppage, base, len, flags);
 217                 if (ret == 0)
 218                         ret = err;
 219                 else if (err > 0)
 220                         ret += err;
 221                 if (err != len)
 222                         goto out;
 223                 base = 0;
 224                 ppage++;
 225         } while ((pglen -= len) != 0);
 226 copy_tail:
 227         len = xdr->tail[0].iov_len;
 228         if (base < len) {
 229                 err = xs_send_tail(sock, xdr, base, len);
 230                 if (ret == 0)
 231                         ret = err;
 232                 else if (err > 0)
 233                         ret += err;
 234         }
 235 out:
 236         return ret;
 237 }
 238
 239 /**
 240  * xs_nospace - place task on wait queue if transmit was incomplete
 241  * @task: task to put to sleep
 242  *
 243  */
 244 static void xs_nospace(struct rpc_task *task)
 245 {
 246         struct rpc_rqst *req = task->tk_rqstp;
 247         struct rpc_xprt *xprt = req->rq_xprt;
 248
 249         dprintk("RPC: %4d xmit incomplete (%u left of %u)\n",
 250                         task->tk_pid, req->rq_slen - req->rq_bytes_sent,
 251                         req->rq_slen);
 252
 253         if (test_bit(SOCK_ASYNC_NOSPACE, &xprt->sock->flags)) {
 254                 /* Protect against races with write_space */
 255                 spin_lock_bh(&xprt->transport_lock);
 256
 257                 /* Don't race with disconnect */
 258                 if (!xprt_connected(xprt))
 259                         task->tk_status = -ENOTCONN;
 260                 else if (test_bit(SOCK_NOSPACE, &xprt->sock->flags))
 261                         xprt_wait_for_buffer_space(task);
 262
 263                 spin_unlock_bh(&xprt->transport_lock);
 264         } else
 265                 /* Keep holding the socket if it is blocked */
 266                 rpc_delay(task, HZ>>4);
 267 }
 268
 269 /**
 270  * xs_udp_send_request - write an RPC request to a UDP socket
 271  * @task: address of RPC task that manages the state of an RPC request
 272  *
 273  * Return values:
 274  *        0:    The request has been sent
 275  *   EAGAIN:    The socket was blocked, please call again later to
 276  *              complete the request
 277  * ENOTCONN:    Caller needs to invoke connect logic then call again
 278  *    other:    Some other error occured, the request was not sent
 279  */
 280 static int xs_udp_send_request(struct rpc_task *task)
 281 {
 282         struct rpc_rqst *req = task->tk_rqstp;
 283         struct rpc_xprt *xprt = req->rq_xprt;
 284         struct xdr_buf *xdr = &req->rq_snd_buf;
 285         int status;
 286
 287         xs_pktdump("packet data:",
 288                                 req->rq_svec->iov_base,
 289                                 req->rq_svec->iov_len);
 290
 291         req->rq_xtime = jiffies;
 292         status = xs_sendpages(xprt->sock, (struct sockaddr *) &xprt->addr,
 293                                 sizeof(xprt->addr), xdr, req->rq_bytes_sent);
 294
 295         dprintk("RPC:      xs_udp_send_request(%u) = %d\n",
 296                         xdr->len - req->rq_bytes_sent, status);
 297
 298         if (likely(status >= (int) req->rq_slen))
 299                 return 0;
 300
 301         /* Still some bytes left; set up for a retry later. */
 302         if (status > 0)
 303                 status = -EAGAIN;
 304
 305         switch (status) {
 306         case -ENETUNREACH:
 307         case -EPIPE:
 308         case -ECONNREFUSED:
 309                 /* When the server has died, an ICMP port unreachable message
 310                  * prompts ECONNREFUSED. */
 311                 break;
 312         case -EAGAIN:
 313                 xs_nospace(task);
 314                 break;
 315         default:
 316                 dprintk("RPC:      sendmsg returned unrecognized error %d\n",
 317                         -status);
 318                 break;
 319         }
 320
 321         return status;
 322 }
 323
 324 static inline void xs_encode_tcp_record_marker(struct xdr_buf *buf)
 325 {
 326         u32 reclen = buf->len - sizeof(rpc_fraghdr);
 327         rpc_fraghdr *base = buf->head[0].iov_base;
 328         *base = htonl(RPC_LAST_STREAM_FRAGMENT | reclen);
 329 }
 330
 331 /**
 332  * xs_tcp_send_request - write an RPC request to a TCP socket
 333  * @task: address of RPC task that manages the state of an RPC request
 334  *
 335  * Return values:
 336  *        0:    The request has been sent
 337  *   EAGAIN:    The socket was blocked, please call again later to
 338  *              complete the request
 339  * ENOTCONN:    Caller needs to invoke connect logic then call again
 340  *    other:    Some other error occured, the request was not sent
 341  *
 342  * XXX: In the case of soft timeouts, should we eventually give up
 343  *      if sendmsg is not able to make progress?
 344  */
 345 static int xs_tcp_send_request(struct rpc_task *task)
 346 {
 347         struct rpc_rqst *req = task->tk_rqstp;
 348         struct rpc_xprt *xprt = req->rq_xprt;
 349         struct xdr_buf *xdr = &req->rq_snd_buf;
 350         int status, retry = 0;
 351
 352         xs_encode_tcp_record_marker(&req->rq_snd_buf);
 353
 354         xs_pktdump("packet data:",
 355                                 req->rq_svec->iov_base,
 356                                 req->rq_svec->iov_len);
 357
 358         /* Continue transmitting the packet/record. We must be careful
 359          * to cope with writespace callbacks arriving _after_ we have
 360          * called sendmsg(). */
 361         while (1) {
 362                 req->rq_xtime = jiffies;
 363                 status = xs_sendpages(xprt->sock, NULL, 0, xdr,
 364                                                 req->rq_bytes_sent);
 365
 366                 dprintk("RPC:      xs_tcp_send_request(%u) = %d\n",
 367                                 xdr->len - req->rq_bytes_sent, status);
 368
 369                 if (unlikely(status < 0))
 370                         break;
 371
 372                 /* If we've sent the entire packet, immediately
 373                  * reset the count of bytes sent. */
 374                 req->rq_bytes_sent += status;
 375                 if (likely(req->rq_bytes_sent >= req->rq_slen)) {
 376                         req->rq_bytes_sent = 0;
 377                         return 0;
 378                 }
 379
 380                 status = -EAGAIN;
 381                 if (retry++ > XS_SENDMSG_RETRY)
 382                         break;
 383         }
 384
 385         switch (status) {
 386         case -EAGAIN:
 387                 xs_nospace(task);
 388                 break;
 389         case -ECONNREFUSED:
 390         case -ECONNRESET:
 391         case -ENOTCONN:
 392         case -EPIPE:
 393                 status = -ENOTCONN;
 394                 break;
 395         default:
 396                 dprintk("RPC:      sendmsg returned unrecognized error %d\n",
 397                         -status);
 398                 xprt_disconnect(xprt);
 399                 break;
 400         }
 401
 402         return status;
 403 }
 404
 405 /**
 406  * xs_close - close a socket
 407  * @xprt: transport
 408  *
 409  * This is used when all requests are complete; ie, no DRC state remains
 410  * on the server we want to save.
 411  */
 412 static void xs_close(struct rpc_xprt *xprt)
 413 {
 414         struct socket *sock = xprt->sock;
 415         struct sock *sk = xprt->inet;
 416
 417         if (!sk)
 418                 return;
 419
 420         dprintk("RPC:      xs_close xprt %p\n", xprt);
 421
 422         write_lock_bh(&sk->sk_callback_lock);
 423         xprt->inet = NULL;
 424         xprt->sock = NULL;
 425
 426         sk->sk_user_data = NULL;
 427         sk->sk_data_ready = xprt->old_data_ready;
 428         sk->sk_state_change = xprt->old_state_change;
 429         sk->sk_write_space = xprt->old_write_space;
 430         write_unlock_bh(&sk->sk_callback_lock);
 431
 432         sk->sk_no_check = 0;
 433
 434         sock_release(sock);
 435 }
 436
 437 /**
 438  * xs_destroy - prepare to shutdown a transport
 439  * @xprt: doomed transport
 440  *
 441  */
 442 static void xs_destroy(struct rpc_xprt *xprt)
 443 {
 444         dprintk("RPC:      xs_destroy xprt %p\n", xprt);
 445
 446         cancel_delayed_work(&xprt->connect_worker);
 447         flush_scheduled_work();
 448
 449         xprt_disconnect(xprt);
 450         xs_close(xprt);
 451         kfree(xprt->slot);
 452 }
 453
 454 static inline struct rpc_xprt *xprt_from_sock(struct sock *sk)
 455 {
 456         return (struct rpc_xprt *) sk->sk_user_data;
 457 }
 458
 459 /**
 460  * xs_udp_data_ready - "data ready" callback for UDP sockets
 461  * @sk: socket with data to read
 462  * @len: how much data to read
 463  *
 464  */
 465 static void xs_udp_data_ready(struct sock *sk, int len)
 466 {
 467         struct rpc_task *task;
 468         struct rpc_xprt *xprt;
 469         struct rpc_rqst *rovr;
 470         struct sk_buff *skb;
 471         int err, repsize, copied;
 472         u32 _xid, *xp;
 473
 474         read_lock(&sk->sk_callback_lock);
 475         dprintk("RPC:      xs_udp_data_ready...\n");
 476         if (!(xprt = xprt_from_sock(sk)))
 477                 goto out;
 478
 479         if ((skb = skb_recv_datagram(sk, 0, 1, &err)) == NULL)
 480                 goto out;
 481
 482         if (xprt->shutdown)
 483                 goto dropit;
 484
 485         repsize = skb->len - sizeof(struct udphdr);
 486         if (repsize < 4) {
 487                 dprintk("RPC:      impossible RPC reply size %d!\n", repsize);
 488                 goto dropit;
 489         }
 490
 491         /* Copy the XID from the skb... */
 492         xp = skb_header_pointer(skb, sizeof(struct udphdr),
 493                                 sizeof(_xid), &_xid);
 494         if (xp == NULL)
 495                 goto dropit;
 496
 497         /* Look up and lock the request corresponding to the given XID */
 498         spin_lock(&xprt->transport_lock);
 499         rovr = xprt_lookup_rqst(xprt, *xp);
 500         if (!rovr)
 501                 goto out_unlock;
 502         task = rovr->rq_task;
 503
 504         if ((copied = rovr->rq_private_buf.buflen) > repsize)
 505                 copied = repsize;
 506
 507         /* Suck it into the iovec, verify checksum if not done by hw. */
 508         if (csum_partial_copy_to_xdr(&rovr->rq_private_buf, skb))
 509                 goto out_unlock;
 510
 511         /* Something worked... */
 512         dst_confirm(skb->dst);
 513
 514         xprt_adjust_cwnd(task, copied);
 515         xprt_update_rtt(task);
 516         xprt_complete_rqst(task, copied);
 517
 518  out_unlock:
 519         spin_unlock(&xprt->transport_lock);
 520  dropit:
 521         skb_free_datagram(sk, skb);
 522  out:
 523         read_unlock(&sk->sk_callback_lock);
 524 }
 525
 526 static inline size_t xs_tcp_copy_data(skb_reader_t *desc, void *p, size_t len)
 527 {
 528         if (len > desc->count)
 529                 len = desc->count;
 530         if (skb_copy_bits(desc->skb, desc->offset, p, len)) {
 531                 dprintk("RPC:      failed to copy %zu bytes from skb. %zu bytes remain\n",
 532                                 len, desc->count);
 533                 return 0;
 534         }
 535         desc->offset += len;
 536         desc->count -= len;
 537         dprintk("RPC:      copied %zu bytes from skb. %zu bytes remain\n",
 538                         len, desc->count);
 539         return len;
 540 }
 541
 542 static inline void xs_tcp_read_fraghdr(struct rpc_xprt *xprt, skb_reader_t *desc)
 543 {
 544         size_t len, used;
 545         char *p;
 546
 547         p = ((char *) &xprt->tcp_recm) + xprt->tcp_offset;
 548         len = sizeof(xprt->tcp_recm) - xprt->tcp_offset;
 549         used = xs_tcp_copy_data(desc, p, len);
 550         xprt->tcp_offset += used;
 551         if (used != len)
 552                 return;
 553
 554         xprt->tcp_reclen = ntohl(xprt->tcp_recm);
 555         if (xprt->tcp_reclen & RPC_LAST_STREAM_FRAGMENT)
 556                 xprt->tcp_flags |= XPRT_LAST_FRAG;
 557         else
 558                 xprt->tcp_flags &= ~XPRT_LAST_FRAG;
 559         xprt->tcp_reclen &= RPC_FRAGMENT_SIZE_MASK;
 560
 561         xprt->tcp_flags &= ~XPRT_COPY_RECM;
 562         xprt->tcp_offset = 0;
 563
 564         /* Sanity check of the record length */
 565         if (unlikely(xprt->tcp_reclen < 4)) {
 566                 dprintk("RPC:      invalid TCP record fragment length\n");
 567                 xprt_disconnect(xprt);
 568                 return;
 569         }
 570         dprintk("RPC:      reading TCP record fragment of length %d\n",
 571                         xprt->tcp_reclen);
 572 }
 573
 574 static void xs_tcp_check_recm(struct rpc_xprt *xprt)
 575 {
 576         dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u, tcp_flags = %lx\n",
 577                         xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen, xprt->tcp_flags);
 578         if (xprt->tcp_offset == xprt->tcp_reclen) {
 579                 xprt->tcp_flags |= XPRT_COPY_RECM;
 580                 xprt->tcp_offset = 0;
 581                 if (xprt->tcp_flags & XPRT_LAST_FRAG) {
 582                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 583                         xprt->tcp_flags |= XPRT_COPY_XID;
 584                         xprt->tcp_copied = 0;
 585                 }
 586         }
 587 }
 588
 589 static inline void xs_tcp_read_xid(struct rpc_xprt *xprt, skb_reader_t *desc)
 590 {
 591         size_t len, used;
 592         char *p;
 593
 594         len = sizeof(xprt->tcp_xid) - xprt->tcp_offset;
 595         dprintk("RPC:      reading XID (%Zu bytes)\n", len);
 596         p = ((char *) &xprt->tcp_xid) + xprt->tcp_offset;
 597         used = xs_tcp_copy_data(desc, p, len);
 598         xprt->tcp_offset += used;
 599         if (used != len)
 600                 return;
 601         xprt->tcp_flags &= ~XPRT_COPY_XID;
 602         xprt->tcp_flags |= XPRT_COPY_DATA;
 603         xprt->tcp_copied = 4;
 604         dprintk("RPC:      reading reply for XID %08x\n",
 605                                                 ntohl(xprt->tcp_xid));
 606         xs_tcp_check_recm(xprt);
 607 }
 608
 609 static inline void xs_tcp_read_request(struct rpc_xprt *xprt, skb_reader_t *desc)
 610 {
 611         struct rpc_rqst *req;
 612         struct xdr_buf *rcvbuf;
 613         size_t len;
 614         ssize_t r;
 615
 616         /* Find and lock the request corresponding to this xid */
 617         spin_lock(&xprt->transport_lock);
 618         req = xprt_lookup_rqst(xprt, xprt->tcp_xid);
 619         if (!req) {
 620                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 621                 dprintk("RPC:      XID %08x request not found!\n",
 622                                 ntohl(xprt->tcp_xid));
 623                 spin_unlock(&xprt->transport_lock);
 624                 return;
 625         }
 626
 627         rcvbuf = &req->rq_private_buf;
 628         len = desc->count;
 629         if (len > xprt->tcp_reclen - xprt->tcp_offset) {
 630                 skb_reader_t my_desc;
 631
 632                 len = xprt->tcp_reclen - xprt->tcp_offset;
 633                 memcpy(&my_desc, desc, sizeof(my_desc));
 634                 my_desc.count = len;
 635                 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 636                                           &my_desc, xs_tcp_copy_data);
 637                 desc->count -= r;
 638                 desc->offset += r;
 639         } else
 640                 r = xdr_partial_copy_from_skb(rcvbuf, xprt->tcp_copied,
 641                                           desc, xs_tcp_copy_data);
 642
 643         if (r > 0) {
 644                 xprt->tcp_copied += r;
 645                 xprt->tcp_offset += r;
 646         }
 647         if (r != len) {
 648                 /* Error when copying to the receive buffer,
 649                  * usually because we weren't able to allocate
 650                  * additional buffer pages. All we can do now
 651                  * is turn off XPRT_COPY_DATA, so the request
 652                  * will not receive any additional updates,
 653                  * and time out.
 654                  * Any remaining data from this record will
 655                  * be discarded.
 656                  */
 657                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 658                 dprintk("RPC:      XID %08x truncated request\n",
 659                                 ntohl(xprt->tcp_xid));
 660                 dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
 661                                 xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
 662                 goto out;
 663         }
 664
 665         dprintk("RPC:      XID %08x read %Zd bytes\n",
 666                         ntohl(xprt->tcp_xid), r);
 667         dprintk("RPC:      xprt = %p, tcp_copied = %lu, tcp_offset = %u, tcp_reclen = %u\n",
 668                         xprt, xprt->tcp_copied, xprt->tcp_offset, xprt->tcp_reclen);
 669
 670         if (xprt->tcp_copied == req->rq_private_buf.buflen)
 671                 xprt->tcp_flags &= ~XPRT_COPY_DATA;
 672         else if (xprt->tcp_offset == xprt->tcp_reclen) {
 673                 if (xprt->tcp_flags & XPRT_LAST_FRAG)
 674                         xprt->tcp_flags &= ~XPRT_COPY_DATA;
 675         }
 676
 677 out:
 678         if (!(xprt->tcp_flags & XPRT_COPY_DATA))
 679                 xprt_complete_rqst(req->rq_task, xprt->tcp_copied);
 680         spin_unlock(&xprt->transport_lock);
 681         xs_tcp_check_recm(xprt);
 682 }
 683
 684 static inline void xs_tcp_read_discard(struct rpc_xprt *xprt, skb_reader_t *desc)
 685 {
 686         size_t len;
 687
 688         len = xprt->tcp_reclen - xprt->tcp_offset;
 689         if (len > desc->count)
 690                 len = desc->count;
 691         desc->count -= len;
 692         desc->offset += len;
 693         xprt->tcp_offset += len;
 694         dprintk("RPC:      discarded %Zu bytes\n", len);
 695         xs_tcp_check_recm(xprt);
 696 }
 697
 698 static int xs_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb, unsigned int offset, size_t len)
 699 {
 700         struct rpc_xprt *xprt = rd_desc->arg.data;
 701         skb_reader_t desc = {
 702                 .skb    = skb,
 703                 .offset = offset,
 704                 .count  = len,
 705                 .csum   = 0
 706         };
 707
 708         dprintk("RPC:      xs_tcp_data_recv started\n");
 709         do {
 710                 /* Read in a new fragment marker if necessary */
 711                 /* Can we ever really expect to get completely empty fragments? */
 712                 if (xprt->tcp_flags & XPRT_COPY_RECM) {
 713                         xs_tcp_read_fraghdr(xprt, &desc);
 714                         continue;
 715                 }
 716                 /* Read in the xid if necessary */
 717                 if (xprt->tcp_flags & XPRT_COPY_XID) {
 718                         xs_tcp_read_xid(xprt, &desc);
 719                         continue;
 720                 }
 721                 /* Read in the request data */
 722                 if (xprt->tcp_flags & XPRT_COPY_DATA) {
 723                         xs_tcp_read_request(xprt, &desc);
 724                         continue;
 725                 }
 726                 /* Skip over any trailing bytes on short reads */
 727                 xs_tcp_read_discard(xprt, &desc);
 728         } while (desc.count);
 729         dprintk("RPC:      xs_tcp_data_recv done\n");
 730         return len - desc.count;
 731 }
 732
 733 /**
 734  * xs_tcp_data_ready - "data ready" callback for TCP sockets
 735  * @sk: socket with data to read
 736  * @bytes: how much data to read
 737  *
 738  */
 739 static void xs_tcp_data_ready(struct sock *sk, int bytes)
 740 {
 741         struct rpc_xprt *xprt;
 742         read_descriptor_t rd_desc;
 743
 744         read_lock(&sk->sk_callback_lock);
 745         dprintk("RPC:      xs_tcp_data_ready...\n");
 746         if (!(xprt = xprt_from_sock(sk)))
 747                 goto out;
 748         if (xprt->shutdown)
 749                 goto out;
 750
 751         /* We use rd_desc to pass struct xprt to xs_tcp_data_recv */
 752         rd_desc.arg.data = xprt;
 753         rd_desc.count = 65536;
 754         tcp_read_sock(sk, &rd_desc, xs_tcp_data_recv);
 755 out:
 756         read_unlock(&sk->sk_callback_lock);
 757 }
 758
 759 /**
 760  * xs_tcp_state_change - callback to handle TCP socket state changes
 761  * @sk: socket whose state has changed
 762  *
 763  */
 764 static void xs_tcp_state_change(struct sock *sk)
 765 {
 766         struct rpc_xprt *xprt;
 767
 768         read_lock(&sk->sk_callback_lock);
 769         if (!(xprt = xprt_from_sock(sk)))
 770                 goto out;
 771         dprintk("RPC:      xs_tcp_state_change client %p...\n", xprt);
 772         dprintk("RPC:      state %x conn %d dead %d zapped %d\n",
 773                                 sk->sk_state, xprt_connected(xprt),
 774                                 sock_flag(sk, SOCK_DEAD),
 775                                 sock_flag(sk, SOCK_ZAPPED));
 776
 777         switch (sk->sk_state) {
 778         case TCP_ESTABLISHED:
 779                 spin_lock_bh(&xprt->transport_lock);
 780                 if (!xprt_test_and_set_connected(xprt)) {
 781                         /* Reset TCP record info */
 782                         xprt->tcp_offset = 0;
 783                         xprt->tcp_reclen = 0;
 784                         xprt->tcp_copied = 0;
 785                         xprt->tcp_flags = XPRT_COPY_RECM | XPRT_COPY_XID;
 786                         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
 787                         xprt_wake_pending_tasks(xprt, 0);
 788                 }
 789                 spin_unlock_bh(&xprt->transport_lock);
 790                 break;
 791         case TCP_SYN_SENT:
 792         case TCP_SYN_RECV:
 793                 break;
 794         default:
 795                 xprt_disconnect(xprt);
 796                 break;
 797         }
 798  out:
 799         read_unlock(&sk->sk_callback_lock);
 800 }
 801
 802 /**
 803  * xs_udp_write_space - callback invoked when socket buffer space
 804  *                             becomes available
 805  * @sk: socket whose state has changed
 806  *
 807  * Called when more output buffer space is available for this socket.
 808  * We try not to wake our writers until they can make "significant"
 809  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
 810  * with a bunch of small requests.
 811  */
 812 static void xs_udp_write_space(struct sock *sk)
 813 {
 814         read_lock(&sk->sk_callback_lock);
 815
 816         /* from net/core/sock.c:sock_def_write_space */
 817         if (sock_writeable(sk)) {
 818                 struct socket *sock;
 819                 struct rpc_xprt *xprt;
 820
 821                 if (unlikely(!(sock = sk->sk_socket)))
 822                         goto out;
 823                 if (unlikely(!(xprt = xprt_from_sock(sk))))
 824                         goto out;
 825                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
 826                         goto out;
 827
 828                 xprt_write_space(xprt);
 829         }
 830
 831  out:
 832         read_unlock(&sk->sk_callback_lock);
 833 }
 834
 835 /**
 836  * xs_tcp_write_space - callback invoked when socket buffer space
 837  *                             becomes available
 838  * @sk: socket whose state has changed
 839  *
 840  * Called when more output buffer space is available for this socket.
 841  * We try not to wake our writers until they can make "significant"
 842  * progress, otherwise we'll waste resources thrashing kernel_sendmsg
 843  * with a bunch of small requests.
 844  */
 845 static void xs_tcp_write_space(struct sock *sk)
 846 {
 847         read_lock(&sk->sk_callback_lock);
 848
 849         /* from net/core/stream.c:sk_stream_write_space */
 850         if (sk_stream_wspace(sk) >= sk_stream_min_wspace(sk)) {
 851                 struct socket *sock;
 852                 struct rpc_xprt *xprt;
 853
 854                 if (unlikely(!(sock = sk->sk_socket)))
 855                         goto out;
 856                 if (unlikely(!(xprt = xprt_from_sock(sk))))
 857                         goto out;
 858                 if (unlikely(!test_and_clear_bit(SOCK_NOSPACE, &sock->flags)))
 859                         goto out;
 860
 861                 xprt_write_space(xprt);
 862         }
 863
 864  out:
 865         read_unlock(&sk->sk_callback_lock);
 866 }
 867
 868 /**
 869  * xs_udp_set_buffer_size - set send and receive limits
 870  * @xprt: generic transport
 871  *
 872  * Set socket send and receive limits based on the
 873  * sndsize and rcvsize fields in the generic transport
 874  * structure.
 875  */
 876 static void xs_udp_set_buffer_size(struct rpc_xprt *xprt)
 877 {
 878         struct sock *sk = xprt->inet;
 879
 880         if (xprt->rcvsize) {
 881                 sk->sk_userlocks |= SOCK_RCVBUF_LOCK;
 882                 sk->sk_rcvbuf = xprt->rcvsize * xprt->max_reqs *  2;
 883         }
 884         if (xprt->sndsize) {
 885                 sk->sk_userlocks |= SOCK_SNDBUF_LOCK;
 886                 sk->sk_sndbuf = xprt->sndsize * xprt->max_reqs * 2;
 887                 sk->sk_write_space(sk);
 888         }
 889 }
 890
 891 /**
 892  * xs_tcp_set_buffer_size - set send and receive limits
 893  * @xprt: generic transport
 894  *
 895  * Nothing to do for TCP.
 896  */
 897 static void xs_tcp_set_buffer_size(struct rpc_xprt *xprt)
 898 {
 899         return;
 900 }
 901
 902 /**
 903  * xs_udp_timer - called when a retransmit timeout occurs on a UDP transport
 904  * @task: task that timed out
 905  *
 906  * Adjust the congestion window after a retransmit timeout has occurred.
 907  */
 908 static void xs_udp_timer(struct rpc_task *task)
 909 {
 910         xprt_adjust_cwnd(task, -ETIMEDOUT);
 911 }
 912
 913 static int xs_bindresvport(struct rpc_xprt *xprt, struct socket *sock)
 914 {
 915         struct sockaddr_in myaddr = {
 916                 .sin_family = AF_INET,
 917         };
 918         int err;
 919         unsigned short port = xprt->port;
 920
 921         do {
 922                 myaddr.sin_port = htons(port);
 923                 err = sock->ops->bind(sock, (struct sockaddr *) &myaddr,
 924                                                 sizeof(myaddr));
 925                 if (err == 0) {
 926                         xprt->port = port;
 927                         dprintk("RPC:      xs_bindresvport bound to port %u\n",
 928                                         port);
 929                         return 0;
 930                 }
 931                 if (port <= xprt_min_resvport)
 932                         port = xprt_max_resvport;
 933                 else
 934                         port--;
 935         } while (err == -EADDRINUSE && port != xprt->port);
 936
 937         dprintk("RPC:      can't bind to reserved port (%d).\n", -err);
 938         return err;
 939 }
 940
 941 /**
 942  * xs_udp_connect_worker - set up a UDP socket
 943  * @args: RPC transport to connect
 944  *
 945  * Invoked by a work queue tasklet.
 946  */
 947 static void xs_udp_connect_worker(void *args)
 948 {
 949         struct rpc_xprt *xprt = (struct rpc_xprt *) args;
 950         struct socket *sock = xprt->sock;
 951         int err, status = -EIO;
 952
 953         if (xprt->shutdown || xprt->addr.sin_port == 0)
 954                 goto out;
 955
 956         dprintk("RPC:      xs_udp_connect_worker for xprt %p\n", xprt);
 957
 958         /* Start by resetting any existing state */
 959         xs_close(xprt);
 960
 961         if ((err = sock_create_kern(PF_INET, SOCK_DGRAM, IPPROTO_UDP, &sock)) < 0) {
 962                 dprintk("RPC:      can't create UDP transport socket (%d).\n", -err);
 963                 goto out;
 964         }
 965
 966         if (xprt->resvport && xs_bindresvport(xprt, sock) < 0) {
 967                 sock_release(sock);
 968                 goto out;
 969         }
 970
 971         if (!xprt->inet) {
 972                 struct sock *sk = sock->sk;
 973
 974                 write_lock_bh(&sk->sk_callback_lock);
 975
 976                 sk->sk_user_data = xprt;
 977                 xprt->old_data_ready = sk->sk_data_ready;
 978                 xprt->old_state_change = sk->sk_state_change;
 979                 xprt->old_write_space = sk->sk_write_space;
 980                 sk->sk_data_ready = xs_udp_data_ready;
 981                 sk->sk_write_space = xs_udp_write_space;
 982                 sk->sk_no_check = UDP_CSUM_NORCV;
 983
 984                 xprt_set_connected(xprt);
 985
 986                 /* Reset to new socket */
 987                 xprt->sock = sock;
 988                 xprt->inet = sk;
 989
 990                 write_unlock_bh(&sk->sk_callback_lock);
 991         }
 992         xs_udp_set_buffer_size(xprt);
 993         status = 0;
 994 out:
 995         xprt_wake_pending_tasks(xprt, status);
 996         xprt_clear_connecting(xprt);
 997 }
 998
 999 /*
1000  * We need to preserve the port number so the reply cache on the server can
1001  * find our cached RPC replies when we get around to reconnecting.
1002  */
1003 static void xs_tcp_reuse_connection(struct rpc_xprt *xprt)
1004 {
1005         int result;
1006         struct socket *sock = xprt->sock;
1007         struct sockaddr any;
1008
1009         dprintk("RPC:      disconnecting xprt %p to reuse port\n", xprt);
1010
1011         /*
1012          * Disconnect the transport socket by doing a connect operation
1013          * with AF_UNSPEC.  This should return immediately...
1014          */
1015         memset(&any, 0, sizeof(any));
1016         any.sa_family = AF_UNSPEC;
1017         result = sock->ops->connect(sock, &any, sizeof(any), 0);
1018         if (result)
1019                 dprintk("RPC:      AF_UNSPEC connect return code %d\n",
1020                                 result);
1021 }
1022
1023 /**
1024  * xs_tcp_connect_worker - connect a TCP socket to a remote endpoint
1025  * @args: RPC transport to connect
1026  *
1027  * Invoked by a work queue tasklet.
1028  */
1029 static void xs_tcp_connect_worker(void *args)
1030 {
1031         struct rpc_xprt *xprt = (struct rpc_xprt *)args;
1032         struct socket *sock = xprt->sock;
1033         int err, status = -EIO;
1034
1035         if (xprt->shutdown || xprt->addr.sin_port == 0)
1036                 goto out;
1037
1038         dprintk("RPC:      xs_tcp_connect_worker for xprt %p\n", xprt);
1039
1040         if (!xprt->sock) {
1041                 /* start from scratch */
1042                 if ((err = sock_create_kern(PF_INET, SOCK_STREAM, IPPROTO_TCP, &sock)) < 0) {
1043                         dprintk("RPC:      can't create TCP transport socket (%d).\n", -err);
1044                         goto out;
1045                 }
1046
1047                 if (xprt->resvport && xs_bindresvport(xprt, sock) < 0) {
1048                         sock_release(sock);
1049                         goto out;
1050                 }
1051         } else
1052                 /* "close" the socket, preserving the local port */
1053                 xs_tcp_reuse_connection(xprt);
1054
1055         if (!xprt->inet) {
1056                 struct sock *sk = sock->sk;
1057
1058                 write_lock_bh(&sk->sk_callback_lock);
1059
1060                 sk->sk_user_data = xprt;
1061                 xprt->old_data_ready = sk->sk_data_ready;
1062                 xprt->old_state_change = sk->sk_state_change;
1063                 xprt->old_write_space = sk->sk_write_space;
1064                 sk->sk_data_ready = xs_tcp_data_ready;
1065                 sk->sk_state_change = xs_tcp_state_change;
1066                 sk->sk_write_space = xs_tcp_write_space;
1067
1068                 /* socket options */
1069                 sk->sk_userlocks |= SOCK_BINDPORT_LOCK;
1070                 sock_reset_flag(sk, SOCK_LINGER);
1071                 tcp_sk(sk)->linger2 = 0;
1072                 tcp_sk(sk)->nonagle |= TCP_NAGLE_OFF;
1073
1074                 xprt_clear_connected(xprt);
1075
1076                 /* Reset to new socket */
1077                 xprt->sock = sock;
1078                 xprt->inet = sk;
1079
1080                 write_unlock_bh(&sk->sk_callback_lock);
1081         }
1082
1083         /* Tell the socket layer to start connecting... */
1084         status = sock->ops->connect(sock, (struct sockaddr *) &xprt->addr,
1085                         sizeof(xprt->addr), O_NONBLOCK);
1086         dprintk("RPC: %p  connect status %d connected %d sock state %d\n",
1087                         xprt, -status, xprt_connected(xprt), sock->sk->sk_state);
1088         if (status < 0) {
1089                 switch (status) {
1090                         case -EINPROGRESS:
1091                         case -EALREADY:
1092                                 goto out_clear;
1093                         case -ECONNREFUSED:
1094                         case -ECONNRESET:
1095                                 /* retry with existing socket, after a delay */
1096                                 break;
1097                         default:
1098                                 /* get rid of existing socket, and retry */
1099                                 xs_close(xprt);
1100                                 break;
1101                 }
1102         }
1103 out:
1104         xprt_wake_pending_tasks(xprt, status);
1105 out_clear:
1106         xprt_clear_connecting(xprt);
1107 }
1108
1109 /**
1110  * xs_connect - connect a socket to a remote endpoint
1111  * @task: address of RPC task that manages state of connect request
1112  *
1113  * TCP: If the remote end dropped the connection, delay reconnecting.
1114  *
1115  * UDP socket connects are synchronous, but we use a work queue anyway
1116  * to guarantee that even unprivileged user processes can set up a
1117  * socket on a privileged port.
1118  *
1119  * If a UDP socket connect fails, the delay behavior here prevents
1120  * retry floods (hard mounts).
1121  */
1122 static void xs_connect(struct rpc_task *task)
1123 {
1124         struct rpc_xprt *xprt = task->tk_xprt;
1125
1126         if (xprt_test_and_set_connecting(xprt))
1127                 return;
1128
1129         if (xprt->sock != NULL) {
1130                 dprintk("RPC:      xs_connect delayed xprt %p for %lu seconds\n",
1131                                 xprt, xprt->reestablish_timeout / HZ);
1132                 schedule_delayed_work(&xprt->connect_worker,
1133                                         xprt->reestablish_timeout);
1134                 xprt->reestablish_timeout <<= 1;
1135                 if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO)
1136                         xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO;
1137         } else {
1138                 dprintk("RPC:      xs_connect scheduled xprt %p\n", xprt);
1139                 schedule_work(&xprt->connect_worker);
1140
1141                 /* flush_scheduled_work can sleep... */
1142                 if (!RPC_IS_ASYNC(task))
1143                         flush_scheduled_work();
1144         }
1145 }
1146
1147 static struct rpc_xprt_ops xs_udp_ops = {
1148         .set_buffer_size        = xs_udp_set_buffer_size,
1149         .reserve_xprt           = xprt_reserve_xprt_cong,
1150         .release_xprt           = xprt_release_xprt_cong,
1151         .connect                = xs_connect,
1152         .send_request           = xs_udp_send_request,
1153         .set_retrans_timeout    = xprt_set_retrans_timeout_rtt,
1154         .timer                  = xs_udp_timer,
1155         .release_request        = xprt_release_rqst_cong,
1156         .close                  = xs_close,
1157         .destroy                = xs_destroy,
1158 };
1159
1160 static struct rpc_xprt_ops xs_tcp_ops = {
1161         .set_buffer_size        = xs_tcp_set_buffer_size,
1162         .reserve_xprt           = xprt_reserve_xprt,
1163         .release_xprt           = xprt_release_xprt,
1164         .connect                = xs_connect,
1165         .send_request           = xs_tcp_send_request,
1166         .set_retrans_timeout    = xprt_set_retrans_timeout_def,
1167         .close                  = xs_close,
1168         .destroy                = xs_destroy,
1169 };
1170
1171 /**
1172  * xs_setup_udp - Set up transport to use a UDP socket
1173  * @xprt: transport to set up
1174  * @to:   timeout parameters
1175  *
1176  */
1177 int xs_setup_udp(struct rpc_xprt *xprt, struct rpc_timeout *to)
1178 {
1179         size_t slot_table_size;
1180
1181         dprintk("RPC:      setting up udp-ipv4 transport...\n");
1182
1183         xprt->max_reqs = xprt_udp_slot_table_entries;
1184         slot_table_size = xprt->max_reqs * sizeof(xprt->slot[0]);
1185         xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1186         if (xprt->slot == NULL)
1187                 return -ENOMEM;
1188         memset(xprt->slot, 0, slot_table_size);
1189
1190         xprt->prot = IPPROTO_UDP;
1191         xprt->port = xprt_max_resvport;
1192         xprt->tsh_size = 0;
1193         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1194         /* XXX: header size can vary due to auth type, IPv6, etc. */
1195         xprt->max_payload = (1U << 16) - (MAX_HEADER << 3);
1196
1197         INIT_WORK(&xprt->connect_worker, xs_udp_connect_worker, xprt);
1198         xprt->bind_timeout = XS_BIND_TO;
1199         xprt->connect_timeout = XS_UDP_CONN_TO;
1200         xprt->reestablish_timeout = XS_UDP_REEST_TO;
1201         xprt->idle_timeout = XS_IDLE_DISC_TO;
1202
1203         xprt->ops = &xs_udp_ops;
1204
1205         if (to)
1206                 xprt->timeout = *to;
1207         else
1208                 xprt_set_timeout(&xprt->timeout, 5, 5 * HZ);
1209
1210         return 0;
1211 }
1212
1213 /**
1214  * xs_setup_tcp - Set up transport to use a TCP socket
1215  * @xprt: transport to set up
1216  * @to: timeout parameters
1217  *
1218  */
1219 int xs_setup_tcp(struct rpc_xprt *xprt, struct rpc_timeout *to)
1220 {
1221         size_t slot_table_size;
1222
1223         dprintk("RPC:      setting up tcp-ipv4 transport...\n");
1224
1225         xprt->max_reqs = xprt_tcp_slot_table_entries;
1226         slot_table_size = xprt->max_reqs * sizeof(xprt->slot[0]);
1227         xprt->slot = kmalloc(slot_table_size, GFP_KERNEL);
1228         if (xprt->slot == NULL)
1229                 return -ENOMEM;
1230         memset(xprt->slot, 0, slot_table_size);
1231
1232         xprt->prot = IPPROTO_TCP;
1233         xprt->port = xprt_max_resvport;
1234         xprt->tsh_size = sizeof(rpc_fraghdr) / sizeof(u32);
1235         xprt->resvport = capable(CAP_NET_BIND_SERVICE) ? 1 : 0;
1236         xprt->max_payload = RPC_MAX_FRAGMENT_SIZE;
1237
1238         INIT_WORK(&xprt->connect_worker, xs_tcp_connect_worker, xprt);
1239         xprt->bind_timeout = XS_BIND_TO;
1240         xprt->connect_timeout = XS_TCP_CONN_TO;
1241         xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO;
1242         xprt->idle_timeout = XS_IDLE_DISC_TO;
1243
1244         xprt->ops = &xs_tcp_ops;
1245
1246         if (to)
1247                 xprt->timeout = *to;
1248         else
1249                 xprt_set_timeout(&xprt->timeout, 2, 60 * HZ);
1250
1251         return 0;
1252 }