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