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