net/core/datagram.c

   1 /*
   2  *      SUCS NET3:
   3  *
   4  *      Generic datagram handling routines. These are generic for all
   5  *      protocols. Possibly a generic IP version on top of these would
   6  *      make sense. Not tonight however 8-).
   7  *      This is used because UDP, RAW, PACKET, DDP, IPX, AX.25 and
   8  *      NetROM layer all have identical poll code and mostly
   9  *      identical recvmsg() code. So we share it here. The poll was
  10  *      shared before but buried in udp.c so I moved it.
  11  *
  12  *      Authors:        Alan Cox <alan@redhat.com>. (datagram_poll() from old
  13  *                                                   udp.c code)
  14  *
  15  *      Fixes:
  16  *              Alan Cox        :       NULL return from skb_peek_copy()
  17  *                                      understood
  18  *              Alan Cox        :       Rewrote skb_read_datagram to avoid the
  19  *                                      skb_peek_copy stuff.
  20  *              Alan Cox        :       Added support for SOCK_SEQPACKET.
  21  *                                      IPX can no longer use the SO_TYPE hack
  22  *                                      but AX.25 now works right, and SPX is
  23  *                                      feasible.
  24  *              Alan Cox        :       Fixed write poll of non IP protocol
  25  *                                      crash.
  26  *              Florian  La Roche:      Changed for my new skbuff handling.
  27  *              Darryl Miles    :       Fixed non-blocking SOCK_SEQPACKET.
  28  *              Linus Torvalds  :       BSD semantic fixes.
  29  *              Alan Cox        :       Datagram iovec handling
  30  *              Darryl Miles    :       Fixed non-blocking SOCK_STREAM.
  31  *              Alan Cox        :       POSIXisms
  32  *              Pete Wyckoff    :       Unconnected accept() fix.
  33  *
  34  */
  35
  36 #include <linux/module.h>
  37 #include <linux/types.h>
  38 #include <linux/kernel.h>
  39 #include <asm/uaccess.h>
  40 #include <asm/system.h>
  41 #include <linux/mm.h>
  42 #include <linux/interrupt.h>
  43 #include <linux/errno.h>
  44 #include <linux/sched.h>
  45 #include <linux/inet.h>
  46 #include <linux/tcp.h>
  47 #include <linux/netdevice.h>
  48 #include <linux/rtnetlink.h>
  49 #include <linux/poll.h>
  50 #include <linux/highmem.h>
  51
  52 #include <net/protocol.h>
  53 #include <linux/skbuff.h>
  54 #include <net/sock.h>
  55 #include <net/checksum.h>
  56
  57
  58 /*
  59  *      Is a socket 'connection oriented' ?
  60  */
  61 static inline int connection_based(struct sock *sk)
  62 {
  63         return sk->sk_type == SOCK_SEQPACKET || sk->sk_type == SOCK_STREAM;
  64 }
  65
  66 /*
  67  * Wait for a packet..
  68  */
  69 static int wait_for_packet(struct sock *sk, int *err, long *timeo_p)
  70 {
  71         int error;
  72         DEFINE_WAIT(wait);
  73
  74         prepare_to_wait_exclusive(sk->sk_sleep, &wait, TASK_INTERRUPTIBLE);
  75
  76         /* Socket errors? */
  77         error = sock_error(sk);
  78         if (error)
  79                 goto out_err;
  80
  81         if (!skb_queue_empty(&sk->sk_receive_queue))
  82                 goto out;
  83
  84         /* Socket shut down? */
  85         if (sk->sk_shutdown & RCV_SHUTDOWN)
  86                 goto out_noerr;
  87
  88         /* Sequenced packets can come disconnected.
  89          * If so we report the problem
  90          */
  91         error = -ENOTCONN;
  92         if (connection_based(sk) &&
  93             !(sk->sk_state == TCP_ESTABLISHED || sk->sk_state == TCP_LISTEN))
  94                 goto out_err;
  95
  96         /* handle signals */
  97         if (signal_pending(current))
  98                 goto interrupted;
  99
 100         error = 0;
 101         *timeo_p = schedule_timeout(*timeo_p);
 102 out:
 103         finish_wait(sk->sk_sleep, &wait);
 104         return error;
 105 interrupted:
 106         error = sock_intr_errno(*timeo_p);
 107 out_err:
 108         *err = error;
 109         goto out;
 110 out_noerr:
 111         *err = 0;
 112         error = 1;
 113         goto out;
 114 }
 115
 116 /**
 117  *      skb_recv_datagram - Receive a datagram skbuff
 118  *      @sk - socket
 119  *      @flags - MSG_ flags
 120  *      @noblock - blocking operation?
 121  *      @err - error code returned
 122  *
 123  *      Get a datagram skbuff, understands the peeking, nonblocking wakeups
 124  *      and possible races. This replaces identical code in packet, raw and
 125  *      udp, as well as the IPX AX.25 and Appletalk. It also finally fixes
 126  *      the long standing peek and read race for datagram sockets. If you
 127  *      alter this routine remember it must be re-entrant.
 128  *
 129  *      This function will lock the socket if a skb is returned, so the caller
 130  *      needs to unlock the socket in that case (usually by calling
 131  *      skb_free_datagram)
 132  *
 133  *      * It does not lock socket since today. This function is
 134  *      * free of race conditions. This measure should/can improve
 135  *      * significantly datagram socket latencies at high loads,
 136  *      * when data copying to user space takes lots of time.
 137  *      * (BTW I've just killed the last cli() in IP/IPv6/core/netlink/packet
 138  *      *  8) Great win.)
 139  *      *                                           --ANK (980729)
 140  *
 141  *      The order of the tests when we find no data waiting are specified
 142  *      quite explicitly by POSIX 1003.1g, don't change them without having
 143  *      the standard around please.
 144  */
 145 struct sk_buff *skb_recv_datagram(struct sock *sk, unsigned flags,
 146                                   int noblock, int *err)
 147 {
 148         struct sk_buff *skb;
 149         long timeo;
 150         /*
 151          * Caller is allowed not to check sk->sk_err before skb_recv_datagram()
 152          */
 153         int error = sock_error(sk);
 154
 155         if (error)
 156                 goto no_packet;
 157
 158         timeo = sock_rcvtimeo(sk, noblock);
 159
 160         do {
 161                 /* Again only user level code calls this function, so nothing
 162                  * interrupt level will suddenly eat the receive_queue.
 163                  *
 164                  * Look at current nfs client by the way...
 165                  * However, this function was corrent in any case. 8)
 166                  */
 167                 if (flags & MSG_PEEK) {
 168                         unsigned long cpu_flags;
 169
 170                         spin_lock_irqsave(&sk->sk_receive_queue.lock,
 171                                           cpu_flags);
 172                         skb = skb_peek(&sk->sk_receive_queue);
 173                         if (skb)
 174                                 atomic_inc(&skb->users);
 175                         spin_unlock_irqrestore(&sk->sk_receive_queue.lock,
 176                                                cpu_flags);
 177                 } else
 178                         skb = skb_dequeue(&sk->sk_receive_queue);
 179
 180                 if (skb)
 181                         return skb;
 182
 183                 /* User doesn't want to wait */
 184                 error = -EAGAIN;
 185                 if (!timeo)
 186                         goto no_packet;
 187
 188         } while (!wait_for_packet(sk, err, &timeo));
 189
 190         return NULL;
 191
 192 no_packet:
 193         *err = error;
 194         return NULL;
 195 }
 196
 197 void skb_free_datagram(struct sock *sk, struct sk_buff *skb)
 198 {
 199         kfree_skb(skb);
 200 }
 201
 202 /**
 203  *      skb_copy_datagram_iovec - Copy a datagram to an iovec.
 204  *      @skb - buffer to copy
 205  *      @offset - offset in the buffer to start copying from
 206  *      @iovec - io vector to copy to
 207  *      @len - amount of data to copy from buffer to iovec
 208  *
 209  *      Note: the iovec is modified during the copy.
 210  */
 211 int skb_copy_datagram_iovec(const struct sk_buff *skb, int offset,
 212                             struct iovec *to, int len)
 213 {
 214         int start = skb_headlen(skb);
 215         int i, copy = start - offset;
 216
 217         /* Copy header. */
 218         if (copy > 0) {
 219                 if (copy > len)
 220                         copy = len;
 221                 if (memcpy_toiovec(to, skb->data + offset, copy))
 222                         goto fault;
 223                 if ((len -= copy) == 0)
 224                         return 0;
 225                 offset += copy;
 226         }
 227
 228         /* Copy paged appendix. Hmm... why does this look so complicated? */
 229         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 230                 int end;
 231
 232                 BUG_TRAP(start <= offset + len);
 233
 234                 end = start + skb_shinfo(skb)->frags[i].size;
 235                 if ((copy = end - offset) > 0) {
 236                         int err;
 237                         u8  *vaddr;
 238                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 239                         struct page *page = frag->page;
 240
 241                         if (copy > len)
 242                                 copy = len;
 243                         vaddr = kmap(page);
 244                         err = memcpy_toiovec(to, vaddr + frag->page_offset +
 245                                              offset - start, copy);
 246                         kunmap(page);
 247                         if (err)
 248                                 goto fault;
 249                         if (!(len -= copy))
 250                                 return 0;
 251                         offset += copy;
 252                 }
 253                 start = end;
 254         }
 255
 256         if (skb_shinfo(skb)->frag_list) {
 257                 struct sk_buff *list = skb_shinfo(skb)->frag_list;
 258
 259                 for (; list; list = list->next) {
 260                         int end;
 261
 262                         BUG_TRAP(start <= offset + len);
 263
 264                         end = start + list->len;
 265                         if ((copy = end - offset) > 0) {
 266                                 if (copy > len)
 267                                         copy = len;
 268                                 if (skb_copy_datagram_iovec(list,
 269                                                             offset - start,
 270                                                             to, copy))
 271                                         goto fault;
 272                                 if ((len -= copy) == 0)
 273                                         return 0;
 274                                 offset += copy;
 275                         }
 276                         start = end;
 277                 }
 278         }
 279         if (!len)
 280                 return 0;
 281
 282 fault:
 283         return -EFAULT;
 284 }
 285
 286 static int skb_copy_and_csum_datagram(const struct sk_buff *skb, int offset,
 287                                       u8 __user *to, int len,
 288                                       unsigned int *csump)
 289 {
 290         int start = skb_headlen(skb);
 291         int pos = 0;
 292         int i, copy = start - offset;
 293
 294         /* Copy header. */
 295         if (copy > 0) {
 296                 int err = 0;
 297                 if (copy > len)
 298                         copy = len;
 299                 *csump = csum_and_copy_to_user(skb->data + offset, to, copy,
 300                                                *csump, &err);
 301                 if (err)
 302                         goto fault;
 303                 if ((len -= copy) == 0)
 304                         return 0;
 305                 offset += copy;
 306                 to += copy;
 307                 pos = copy;
 308         }
 309
 310         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 311                 int end;
 312
 313                 BUG_TRAP(start <= offset + len);
 314
 315                 end = start + skb_shinfo(skb)->frags[i].size;
 316                 if ((copy = end - offset) > 0) {
 317                         unsigned int csum2;
 318                         int err = 0;
 319                         u8  *vaddr;
 320                         skb_frag_t *frag = &skb_shinfo(skb)->frags[i];
 321                         struct page *page = frag->page;
 322
 323                         if (copy > len)
 324                                 copy = len;
 325                         vaddr = kmap(page);
 326                         csum2 = csum_and_copy_to_user(vaddr +
 327                                                         frag->page_offset +
 328                                                         offset - start,
 329                                                       to, copy, 0, &err);
 330                         kunmap(page);
 331                         if (err)
 332                                 goto fault;
 333                         *csump = csum_block_add(*csump, csum2, pos);
 334                         if (!(len -= copy))
 335                                 return 0;
 336                         offset += copy;
 337                         to += copy;
 338                         pos += copy;
 339                 }
 340                 start = end;
 341         }
 342
 343         if (skb_shinfo(skb)->frag_list) {
 344                 struct sk_buff *list = skb_shinfo(skb)->frag_list;
 345
 346                 for (; list; list=list->next) {
 347                         int end;
 348
 349                         BUG_TRAP(start <= offset + len);
 350
 351                         end = start + list->len;
 352                         if ((copy = end - offset) > 0) {
 353                                 unsigned int csum2 = 0;
 354                                 if (copy > len)
 355                                         copy = len;
 356                                 if (skb_copy_and_csum_datagram(list,
 357                                                                offset - start,
 358                                                                to, copy,
 359                                                                &csum2))
 360                                         goto fault;
 361                                 *csump = csum_block_add(*csump, csum2, pos);
 362                                 if ((len -= copy) == 0)
 363                                         return 0;
 364                                 offset += copy;
 365                                 to += copy;
 366                                 pos += copy;
 367                         }
 368                         start = end;
 369                 }
 370         }
 371         if (!len)
 372                 return 0;
 373
 374 fault:
 375         return -EFAULT;
 376 }
 377
 378 /**
 379  *      skb_copy_and_csum_datagram_iovec - Copy and checkum skb to user iovec.
 380  *      @skb - skbuff
 381  *      @hlen - hardware length
 382  *      @iovec - io vector
 383  *
 384  *      Caller _must_ check that skb will fit to this iovec.
 385  *
 386  *      Returns: 0       - success.
 387  *               -EINVAL - checksum failure.
 388  *               -EFAULT - fault during copy. Beware, in this case iovec
 389  *                         can be modified!
 390  */
 391 int skb_copy_and_csum_datagram_iovec(const struct sk_buff *skb,
 392                                      int hlen, struct iovec *iov)
 393 {
 394         unsigned int csum;
 395         int chunk = skb->len - hlen;
 396
 397         /* Skip filled elements.
 398          * Pretty silly, look at memcpy_toiovec, though 8)
 399          */
 400         while (!iov->iov_len)
 401                 iov++;
 402
 403         if (iov->iov_len < chunk) {
 404                 if ((unsigned short)csum_fold(skb_checksum(skb, 0, chunk + hlen,
 405                                                            skb->csum)))
 406                         goto csum_error;
 407                 if (skb_copy_datagram_iovec(skb, hlen, iov, chunk))
 408                         goto fault;
 409         } else {
 410                 csum = csum_partial(skb->data, hlen, skb->csum);
 411                 if (skb_copy_and_csum_datagram(skb, hlen, iov->iov_base,
 412                                                chunk, &csum))
 413                         goto fault;
 414                 if ((unsigned short)csum_fold(csum))
 415                         goto csum_error;
 416                 iov->iov_len -= chunk;
 417                 iov->iov_base += chunk;
 418         }
 419         return 0;
 420 csum_error:
 421         return -EINVAL;
 422 fault:
 423         return -EFAULT;
 424 }
 425
 426 /**
 427  *      datagram_poll - generic datagram poll
 428  *      @file - file struct
 429  *      @sock - socket
 430  *      @wait - poll table
 431  *
 432  *      Datagram poll: Again totally generic. This also handles
 433  *      sequenced packet sockets providing the socket receive queue
 434  *      is only ever holding data ready to receive.
 435  *
 436  *      Note: when you _don't_ use this routine for this protocol,
 437  *      and you use a different write policy from sock_writeable()
 438  *      then please supply your own write_space callback.
 439  */
 440 unsigned int datagram_poll(struct file *file, struct socket *sock,
 441                            poll_table *wait)
 442 {
 443         struct sock *sk = sock->sk;
 444         unsigned int mask;
 445
 446         poll_wait(file, sk->sk_sleep, wait);
 447         mask = 0;
 448
 449         /* exceptional events? */
 450         if (sk->sk_err || !skb_queue_empty(&sk->sk_error_queue))
 451                 mask |= POLLERR;
 452         if (sk->sk_shutdown == SHUTDOWN_MASK)
 453                 mask |= POLLHUP;
 454
 455         /* readable? */
 456         if (!skb_queue_empty(&sk->sk_receive_queue) ||
 457             (sk->sk_shutdown & RCV_SHUTDOWN))
 458                 mask |= POLLIN | POLLRDNORM;
 459
 460         /* Connection-based need to check for termination and startup */
 461         if (connection_based(sk)) {
 462                 if (sk->sk_state == TCP_CLOSE)
 463                         mask |= POLLHUP;
 464                 /* connection hasn't started yet? */
 465                 if (sk->sk_state == TCP_SYN_SENT)
 466                         return mask;
 467         }
 468
 469         /* writable? */
 470         if (sock_writeable(sk))
 471                 mask |= POLLOUT | POLLWRNORM | POLLWRBAND;
 472         else
 473                 set_bit(SOCK_ASYNC_NOSPACE, &sk->sk_socket->flags);
 474
 475         return mask;
 476 }
 477
 478 EXPORT_SYMBOL(datagram_poll);
 479 EXPORT_SYMBOL(skb_copy_and_csum_datagram_iovec);
 480 EXPORT_SYMBOL(skb_copy_datagram_iovec);
 481 EXPORT_SYMBOL(skb_free_datagram);
 482 EXPORT_SYMBOL(skb_recv_datagram);