[svn] Try to reuse connections that return error codes.

[wget] / src / http.c

/* HTTP support.
   Copyright (C) 1995, 1996, 1997, 1998, 2000, 2001, 2002
   Free Software Foundation, Inc.

This file is part of GNU Wget.

GNU Wget is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 2 of the License, or
 (at your option) any later version.

GNU Wget is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
GNU General Public License for more details.

You should have received a copy of the GNU General Public License
along with Wget; if not, write to the Free Software
Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.

In addition, as a special exception, the Free Software Foundation
gives permission to link the code of its release of Wget with the
OpenSSL project's "OpenSSL" library (or with modified versions of it
that use the same license as the "OpenSSL" library), and distribute
the linked executables.  You must obey the GNU General Public License
in all respects for all of the code used other than "OpenSSL".  If you
modify this file, you may extend this exception to your version of the
file, but you are not obligated to do so.  If you do not wish to do
so, delete this exception statement from your version.  */

#include <config.h>

#include <stdio.h>
#include <stdlib.h>
#include <sys/types.h>
#ifdef HAVE_STRING_H
# include <string.h>
#else
# include <strings.h>
#endif
#ifdef HAVE_UNISTD_H
# include <unistd.h>
#endif
#include <assert.h>
#include <errno.h>
#if TIME_WITH_SYS_TIME
# include <sys/time.h>
# include <time.h>
#else
# if HAVE_SYS_TIME_H
#  include <sys/time.h>
# else
#  include <time.h>
# endif
#endif
#ifndef errno
extern int errno;
#endif

#include "wget.h"
#include "utils.h"
#include "url.h"
#include "host.h"
#include "retr.h"
#include "connect.h"
#include "netrc.h"
#ifdef HAVE_SSL
# include "gen_sslfunc.h"
#endif /* HAVE_SSL */
#include "cookies.h"
#ifdef USE_DIGEST
# include "gen-md5.h"
#endif
#include "convert.h"

extern char *version_string;
extern LARGE_INT total_downloaded_bytes;

#ifndef MIN
# define MIN(x, y) ((x) > (y) ? (y) : (x))
#endif

\f
static int cookies_loaded_p;
struct cookie_jar *wget_cookie_jar;

#define TEXTHTML_S "text/html"
#define TEXTXHTML_S "application/xhtml+xml"

/* Some status code validation macros: */
#define H_20X(x)        (((x) >= 200) && ((x) < 300))
#define H_PARTIAL(x)    ((x) == HTTP_STATUS_PARTIAL_CONTENTS)
#define H_REDIRECTED(x) ((x) == HTTP_STATUS_MOVED_PERMANENTLY   \
                         || (x) == HTTP_STATUS_MOVED_TEMPORARILY \
                         || (x) == HTTP_STATUS_TEMPORARY_REDIRECT)

/* HTTP/1.0 status codes from RFC1945, provided for reference.  */
/* Successful 2xx.  */
#define HTTP_STATUS_OK                  200
#define HTTP_STATUS_CREATED             201
#define HTTP_STATUS_ACCEPTED            202
#define HTTP_STATUS_NO_CONTENT          204
#define HTTP_STATUS_PARTIAL_CONTENTS    206

/* Redirection 3xx.  */
#define HTTP_STATUS_MULTIPLE_CHOICES    300
#define HTTP_STATUS_MOVED_PERMANENTLY   301
#define HTTP_STATUS_MOVED_TEMPORARILY   302
#define HTTP_STATUS_NOT_MODIFIED        304
#define HTTP_STATUS_TEMPORARY_REDIRECT  307

/* Client error 4xx.  */
#define HTTP_STATUS_BAD_REQUEST         400
#define HTTP_STATUS_UNAUTHORIZED        401
#define HTTP_STATUS_FORBIDDEN           403
#define HTTP_STATUS_NOT_FOUND           404

/* Server errors 5xx.  */
#define HTTP_STATUS_INTERNAL            500
#define HTTP_STATUS_NOT_IMPLEMENTED     501
#define HTTP_STATUS_BAD_GATEWAY         502
#define HTTP_STATUS_UNAVAILABLE         503
\f
enum rp {
  rel_none, rel_name, rel_value, rel_both
};

struct request {
  const char *method;
  char *arg;

  struct request_header {
    char *name, *value;
    enum rp release_policy;
  } *headers;
  int hcount, hcapacity;
};

/* Create a new, empty request.  At least request_set_method must be
   called before the request can be used.  */

static struct request *
request_new ()
{
  struct request *req = xnew0 (struct request);
  req->hcapacity = 8;
  req->headers = xnew_array (struct request_header, req->hcapacity);
  return req;
}

/* Set the request's method and its arguments.  METH should be a
   literal string (or it should outlive the request) because it will
   not be freed.  ARG will be freed by request_free.  */

static void
request_set_method (struct request *req, const char *meth, char *arg)
{
  req->method = meth;
  req->arg = arg;
}

/* Return the method string passed with the last call to
   request_set_method.  */

static const char *
request_method (const struct request *req)
{
  return req->method;
}

/* Free one header according to the release policy specified with
   request_set_header.  */

static void
release_header (struct request_header *hdr)
{
  switch (hdr->release_policy)
    {
    case rel_none:
      break;
    case rel_name:
      xfree (hdr->name);
      break;
    case rel_value:
      xfree (hdr->value);
      break;
    case rel_both:
      xfree (hdr->name);
      xfree (hdr->value);
      break;
    }
}

/* Set the request named NAME to VALUE.  Specifically, this means that
   a "NAME: VALUE\r\n" header line will be used in the request.  If a
   header with the same name previously existed in the request, its
   value will be replaced by this one.

   RELEASE_POLICY determines whether NAME and VALUE should be released
   (freed) with request_free.  Allowed values are:

    - rel_none     - don't free NAME or VALUE
    - rel_name     - free NAME when done
    - rel_value    - free VALUE when done
    - rel_both     - free both NAME and VALUE when done

   Setting release policy is useful when arguments come from different
   sources.  For example:

     // Don't free literal strings!
     request_set_header (req, "Pragma", "no-cache", rel_none);

     // Don't free a global variable, we'll need it later.
     request_set_header (req, "Referer", opt.referer, rel_none);

     // Value freshly allocated, free it when done.
     request_set_header (req, "Range", aprintf ("bytes=%ld-", hs->restval),
                         rel_value);
   */

static void
request_set_header (struct request *req, char *name, char *value,
                    enum rp release_policy)
{
  struct request_header *hdr;
  int i;
  if (!value)
    return;
  for (i = 0; i < req->hcount; i++)
    {
      hdr = &req->headers[i];
      if (0 == strcasecmp (name, hdr->name))
        {
          /* Replace existing header. */
          release_header (hdr);
          hdr->name = name;
          hdr->value = value;
          hdr->release_policy = release_policy;
          return;
        }
    }

  /* Install new header. */

  if (req->hcount >= req->hcount)
    {
      req->hcapacity <<= 1;
      req->headers = xrealloc (req->headers,
                               req->hcapacity * sizeof (struct request_header));
    }
  hdr = &req->headers[req->hcount++];
  hdr->name = name;
  hdr->value = value;
  hdr->release_policy = release_policy;
}

/* Like request_set_header, but sets the whole header line, as
   provided by the user using the `--header' option.  For example,
   request_set_user_header (req, "Foo: bar") works just like
   request_set_header (req, "Foo", "bar").  */

static void
request_set_user_header (struct request *req, const char *header)
{
  char *name;
  const char *p = strchr (header, ':');
  if (!p)
    return;
  BOUNDED_TO_ALLOCA (header, p, name);
  ++p;
  while (ISSPACE (*p))
    ++p;
  request_set_header (req, xstrdup (name), (char *) p, rel_name);
}

#define APPEND(p, str) do {                     \
  int A_len = strlen (str);                     \
  memcpy (p, str, A_len);                       \
  p += A_len;                                   \
} while (0)

/* Construct the request and write it to FD using fd_write.  */

static int
request_send (const struct request *req, int fd)
{
  char *request_string, *p;
  int i, size, write_error;

  /* Count the request size. */
  size = 0;

  /* METHOD " " ARG " " "HTTP/1.0" "\r\n" */
  size += strlen (req->method) + 1 + strlen (req->arg) + 1 + 8 + 2;

  for (i = 0; i < req->hcount; i++)
    {
      struct request_header *hdr = &req->headers[i];
      /* NAME ": " VALUE "\r\n" */
      size += strlen (hdr->name) + 2 + strlen (hdr->value) + 2;
    }

  /* "\r\n\0" */
  size += 3;

  p = request_string = alloca_array (char, size);

  /* Generate the request. */

  APPEND (p, req->method); *p++ = ' ';
  APPEND (p, req->arg);    *p++ = ' ';
  memcpy (p, "HTTP/1.0\r\n", 10); p += 10;

  for (i = 0; i < req->hcount; i++)
    {
      struct request_header *hdr = &req->headers[i];
      APPEND (p, hdr->name);
      *p++ = ':', *p++ = ' ';
      APPEND (p, hdr->value);
      *p++ = '\r', *p++ = '\n';
    }

  *p++ = '\r', *p++ = '\n', *p++ = '\0';
  assert (p - request_string == size);

#undef APPEND

  DEBUGP (("\n---request begin---\n%s---request end---\n", request_string));

  /* Send the request to the server. */

  write_error = fd_write (fd, request_string, size - 1, -1);
  if (write_error < 0)
    logprintf (LOG_VERBOSE, _("Failed writing HTTP request: %s.\n"),
               strerror (errno));
  return write_error;
}

/* Release the resources used by REQ. */

static void
request_free (struct request *req)
{
  int i;
  xfree_null (req->arg);
  for (i = 0; i < req->hcount; i++)
    release_header (&req->headers[i]);
  xfree_null (req->headers);
  xfree (req);
}

/* Send the contents of FILE_NAME to SOCK/SSL.  Make sure that exactly
   PROMISED_SIZE bytes are sent over the wire -- if the file is
   longer, read only that much; if the file is shorter, report an error.  */

static int
post_file (int sock, const char *file_name, long promised_size)
{
  static char chunk[8192];
  long written = 0;
  int write_error;
  FILE *fp;

  DEBUGP (("[writing POST file %s ... ", file_name));

  fp = fopen (file_name, "rb");
  if (!fp)
    return -1;
  while (!feof (fp) && written < promised_size)
    {
      int towrite;
      int length = fread (chunk, 1, sizeof (chunk), fp);
      if (length == 0)
        break;
      towrite = MIN (promised_size - written, length);
      write_error = fd_write (sock, chunk, towrite, -1);
      if (write_error < 0)
        {
          fclose (fp);
          return -1;
        }
      written += towrite;
    }
  fclose (fp);

  /* If we've written less than was promised, report a (probably
     nonsensical) error rather than break the promise.  */
  if (written < promised_size)
    {
      errno = EINVAL;
      return -1;
    }

  assert (written == promised_size);
  DEBUGP (("done]\n"));
  return 0;
}
\f
static const char *
head_terminator (const char *hunk, int oldlen, int peeklen)
{
  const char *start, *end;

  /* If at first peek, verify whether HUNK starts with "HTTP".  If
     not, this is a HTTP/0.9 request and we must bail out without
     reading anything.  */
  if (oldlen == 0 && 0 != memcmp (hunk, "HTTP", MIN (peeklen, 4)))
    return hunk;

  if (oldlen < 4)
    start = hunk;
  else
    start = hunk + oldlen - 4;
  end = hunk + oldlen + peeklen;

  for (; start < end - 1; start++)
    if (*start == '\n')
      {
        if (start < end - 2
            && start[1] == '\r'
            && start[2] == '\n')
          return start + 3;
        if (start[1] == '\n')
          return start + 2;
      }
  return NULL;
}

/* Read the HTTP request head from FD and return it.  The error
   conditions are the same as with fd_read_hunk.

   To support HTTP/0.9 responses, this function tries to make sure
   that the data begins with "HTTP".  If this is not the case, no data
   is read and an empty request is returned, so that the remaining
   data can be treated as body.  */

static char *
fd_read_http_head (int fd)
{
  return fd_read_hunk (fd, head_terminator, 512);
}

struct response {
  /* The response data. */
  const char *data;

  /* The array of pointers that indicate where each header starts.
     For example, given this HTTP response:

       HTTP/1.0 200 Ok
       Description: some
        text
       Etag: x

     The headers are located like this:

     "HTTP/1.0 200 Ok\r\nDescription: some\r\n text\r\nEtag: x\r\n\r\n"
     ^                   ^                             ^          ^
     headers[0]          headers[1]                    headers[2] headers[3]

     I.e. headers[0] points to the beginning of the request,
     headers[1] points to the end of the first header and the
     beginning of the second one, etc.  */

  const char **headers;
};

/* Create a new response object from the text of the HTTP response,
   available in HEAD.  That text is automatically split into
   constituent header lines for fast retrieval using
   response_header_*.  */

static struct response *
response_new (const char *head)
{
  const char *hdr;
  int count, size;

  struct response *resp = xnew0 (struct response);
  resp->data = head;

  if (*head == '\0')
    {
      /* Empty head means that we're dealing with a headerless
         (HTTP/0.9) response.  In that case, don't set HEADERS at
         all.  */
      return resp;
    }

  /* Split HEAD into header lines, so that response_header_* functions
     don't need to do this over and over again.  */

  size = count = 0;
  hdr = head;
  while (1)
    {
      DO_REALLOC (resp->headers, size, count + 1, const char *);
      resp->headers[count++] = hdr;

      /* Break upon encountering an empty line. */
      if (!hdr[0] || (hdr[0] == '\r' && hdr[1] == '\n') || hdr[0] == '\n')
        break;

      /* Find the end of HDR, including continuations. */
      do
        {
          const char *end = strchr (hdr, '\n');
          if (end)
            hdr = end + 1;
          else
            hdr += strlen (hdr);
        }
      while (*hdr == ' ' || *hdr == '\t');
    }
  DO_REALLOC (resp->headers, size, count + 1, const char *);
  resp->headers[count++] = NULL;

  return resp;
}

/* Locate the header named NAME in the request data.  If found, set
   *BEGPTR to its starting, and *ENDPTR to its ending position, and
   return 1.  Otherwise return 0.

   This function is used as a building block for response_header_copy
   and response_header_strdup.  */

static int
response_header_bounds (const struct response *resp, const char *name,
                        const char **begptr, const char **endptr)
{
  int i;
  const char **headers = resp->headers;
  int name_len;

  if (!headers || !headers[1])
    return 0;

  name_len = strlen (name);

  for (i = 1; headers[i + 1]; i++)
    {
      const char *b = headers[i];
      const char *e = headers[i + 1];
      if (e - b > name_len
          && b[name_len] == ':'
          && 0 == strncasecmp (b, name, name_len))
        {
          b += name_len + 1;
          while (b < e && ISSPACE (*b))
            ++b;
          while (b < e && ISSPACE (e[-1]))
            --e;
          *begptr = b;
          *endptr = e;
          return 1;
        }
    }
  return 0;
}

/* Copy the response header named NAME to buffer BUF, no longer than
   BUFSIZE (BUFSIZE includes the terminating 0).  If the header
   exists, 1 is returned, otherwise 0.  If there should be no limit on
   the size of the header, use response_header_strdup instead.

   If BUFSIZE is 0, no data is copied, but the boolean indication of
   whether the header is present is still returned.  */

static int
response_header_copy (const struct response *resp, const char *name,
                      char *buf, int bufsize)
{
  const char *b, *e;
  if (!response_header_bounds (resp, name, &b, &e))
    return 0;
  if (bufsize)
    {
      int len = MIN (e - b, bufsize);
      strncpy (buf, b, len);
      buf[len] = '\0';
    }
  return 1;
}

/* Return the value of header named NAME in RESP, allocated with
   malloc.  If such a header does not exist in RESP, return NULL.  */

static char *
response_header_strdup (const struct response *resp, const char *name)
{
  const char *b, *e;
  if (!response_header_bounds (resp, name, &b, &e))
    return NULL;
  return strdupdelim (b, e);
}

/* Parse the HTTP status line, which is of format:

   HTTP-Version SP Status-Code SP Reason-Phrase

   The function returns the status-code, or -1 if the status line
   appears malformed.  The pointer to "reason-phrase" message is
   returned in *MESSAGE.  */

static int
response_status (const struct response *resp, char **message)
{
  int status;
  const char *p, *end;

  if (!resp->headers)
    {
      /* For a HTTP/0.9 response, assume status 200. */
      if (message)
        *message = xstrdup (_("No headers, assuming HTTP/0.9"));
      return 200;
    }

  p = resp->headers[0];
  end = resp->headers[1];

  if (!end)
    return -1;

  /* "HTTP" */
  if (end - p < 4 || 0 != strncmp (p, "HTTP", 4))
    return -1;
  p += 4;

  /* Match the HTTP version.  This is optional because Gnutella
     servers have been reported to not specify HTTP version.  */
  if (p < end && *p == '/')
    {
      ++p;
      while (p < end && ISDIGIT (*p))
        ++p;
      if (p < end && *p == '.')
        ++p; 
      while (p < end && ISDIGIT (*p))
        ++p;
    }

  while (p < end && ISSPACE (*p))
    ++p;
  if (end - p < 3 || !ISDIGIT (p[0]) || !ISDIGIT (p[1]) || !ISDIGIT (p[2]))
    return -1;

  status = 100 * (p[0] - '0') + 10 * (p[1] - '0') + (p[2] - '0');
  p += 3;

  if (message)
    {
      while (p < end && ISSPACE (*p))
        ++p;
      while (p < end && ISSPACE (end[-1]))
        --end;
      *message = strdupdelim (p, end);
    }

  return status;
}

/* Release the resources used by RESP.  */

static void
response_free (struct response *resp)
{
  xfree_null (resp->headers);
  xfree (resp);
}

/* Print [b, e) to the log, omitting the trailing CRLF.  */

static void
print_server_response_1 (const char *prefix, const char *b, const char *e)
{
  char *ln;
  if (b < e && e[-1] == '\n')
    --e;
  if (b < e && e[-1] == '\r')
    --e;
  BOUNDED_TO_ALLOCA (b, e, ln);
  logprintf (LOG_VERBOSE, "%s%s\n", prefix, ln);
}

/* Print the server response, line by line, omitting the trailing CR
   characters, prefixed with PREFIX.  */

static void
print_server_response (const struct response *resp, const char *prefix)
{
  int i;
  if (!resp->headers)
    return;
  for (i = 0; resp->headers[i + 1]; i++)
    print_server_response_1 (prefix, resp->headers[i], resp->headers[i + 1]);
}

/* Parse the `Content-Range' header and extract the information it
   contains.  Returns 1 if successful, -1 otherwise.  */
static int
parse_content_range (const char *hdr, long *first_byte_ptr,
                     long *last_byte_ptr, long *entity_length_ptr)
{
  long num;

  /* Ancient versions of Netscape proxy server, presumably predating
     rfc2068, sent out `Content-Range' without the "bytes"
     specifier.  */
  if (!strncasecmp (hdr, "bytes", 5))
    {
      hdr += 5;
      /* "JavaWebServer/1.1.1" sends "bytes: x-y/z", contrary to the
         HTTP spec. */
      if (*hdr == ':')
        ++hdr;
      while (ISSPACE (*hdr))
        ++hdr;
      if (!*hdr)
        return 0;
    }
  if (!ISDIGIT (*hdr))
    return 0;
  for (num = 0; ISDIGIT (*hdr); hdr++)
    num = 10 * num + (*hdr - '0');
  if (*hdr != '-' || !ISDIGIT (*(hdr + 1)))
    return 0;
  *first_byte_ptr = num;
  ++hdr;
  for (num = 0; ISDIGIT (*hdr); hdr++)
    num = 10 * num + (*hdr - '0');
  if (*hdr != '/' || !ISDIGIT (*(hdr + 1)))
    return 0;
  *last_byte_ptr = num;
  ++hdr;
  for (num = 0; ISDIGIT (*hdr); hdr++)
    num = 10 * num + (*hdr - '0');
  *entity_length_ptr = num;
  return 1;
}

/* Read the body of the request, but don't store it anywhere.  This is
   useful when reading error responses that are not logged anywhere,
   but which need to be read so the same connection can be reused.  */

static void
skip_body (int fd, long contlen)
{
  int oldverbose;
  long dummy;

  /* Skipping the body doesn't make sense if the content length is
     unknown because, in that case, persistent connections cannot be
     used.  (#### This is not the case with HTTP/1.1 where they can
     still be used with the magic of the "chunked" transfer!)  */
  if (contlen == -1)
    return;

  oldverbose = opt.verbose;
  opt.verbose = 0;
  fd_read_body (fd, NULL, &dummy, 0, contlen, 1, NULL);
  opt.verbose = oldverbose;
}
\f
/* Persistent connections.  Currently, we cache the most recently used
   connection as persistent, provided that the HTTP server agrees to
   make it such.  The persistence data is stored in the variables
   below.  Ideally, it should be possible to cache an arbitrary fixed
   number of these connections.  */

/* Whether a persistent connection is active. */
static int pconn_active;

static struct {
  /* The socket of the connection.  */
  int socket;

  /* Host and port of the currently active persistent connection. */
  char *host;
  int port;

  /* Whether a ssl handshake has occoured on this connection.  */
  int ssl;
} pconn;

/* Mark the persistent connection as invalid and free the resources it
   uses.  This is used by the CLOSE_* macros after they forcefully
   close a registered persistent connection.  */

static void
invalidate_persistent (void)
{
  DEBUGP (("Disabling further reuse of socket %d.\n", pconn.socket));
  pconn_active = 0;
  fd_close (pconn.socket);
  xfree (pconn.host);
  xzero (pconn);
}

/* Register FD, which should be a TCP/IP connection to HOST:PORT, as
   persistent.  This will enable someone to use the same connection
   later.  In the context of HTTP, this must be called only AFTER the
   response has been received and the server has promised that the
   connection will remain alive.

   If a previous connection was persistent, it is closed. */

static void
register_persistent (const char *host, int port, int fd, int ssl)
{
  if (pconn_active)
    {
      if (pconn.socket == fd)
        {
          /* The connection FD is already registered. */
          return;
        }
      else
        {
          /* The old persistent connection is still active; close it
             first.  This situation arises whenever a persistent
             connection exists, but we then connect to a different
             host, and try to register a persistent connection to that
             one.  */
          invalidate_persistent ();
        }
    }

  pconn_active = 1;
  pconn.socket = fd;
  pconn.host = xstrdup (host);
  pconn.port = port;
  pconn.ssl = ssl;

  DEBUGP (("Registered socket %d for persistent reuse.\n", fd));
}

/* Return non-zero if a persistent connection is available for
   connecting to HOST:PORT.  */

static int
persistent_available_p (const char *host, int port, int ssl,
                        int *host_lookup_failed)
{
  /* First, check whether a persistent connection is active at all.  */
  if (!pconn_active)
    return 0;

  /* If we want SSL and the last connection wasn't or vice versa,
     don't use it.  Checking for host and port is not enough because
     HTTP and HTTPS can apparently coexist on the same port.  */
  if (ssl != pconn.ssl)
    return 0;

  /* If we're not connecting to the same port, we're not interested. */
  if (port != pconn.port)
    return 0;

  /* If the host is the same, we're in business.  If not, there is
     still hope -- read below.  */
  if (0 != strcasecmp (host, pconn.host))
    {
      /* If pconn.socket is already talking to HOST, we needn't
         reconnect.  This happens often when both sites are virtual
         hosts distinguished only by name and served by the same
         network interface, and hence the same web server (possibly
         set up by the ISP and serving many different web sites).
         This admittedly non-standard optimization does not contradict
         HTTP and works well with popular server software.  */

      int found;
      ip_address ip;
      struct address_list *al;

      if (ssl)
        /* Don't try to talk to two different SSL sites over the same
           secure connection!  (Besides, it's not clear if name-based
           virtual hosting is even possible with SSL.)  */
        return 0;

      /* If pconn.socket's peer is one of the IP addresses HOST
         resolves to, pconn.socket is for all intents and purposes
         already talking to HOST.  */

      if (!socket_ip_address (pconn.socket, &ip, ENDPOINT_PEER))
        {
          /* Can't get the peer's address -- something must be very
             wrong with the connection.  */
          invalidate_persistent ();
          return 0;
        }
      al = lookup_host (host, 0);
      if (!al)
        {
          *host_lookup_failed = 1;
          return 0;
        }

      found = address_list_contains (al, &ip);
      address_list_release (al);

      if (!found)
        return 0;

      /* The persistent connection's peer address was found among the
         addresses HOST resolved to; therefore, pconn.sock is in fact
         already talking to HOST -- no need to reconnect.  */
    }

  /* Finally, check whether the connection is still open.  This is
     important because most server implement a liberal (short) timeout
     on persistent connections.  Wget can of course always reconnect
     if the connection doesn't work out, but it's nicer to know in
     advance.  This test is a logical followup of the first test, but
     is "expensive" and therefore placed at the end of the list.  */

  if (!test_socket_open (pconn.socket))
    {
      /* Oops, the socket is no longer open.  Now that we know that,
         let's invalidate the persistent connection before returning
         0.  */
      invalidate_persistent ();
      return 0;
    }

  return 1;
}

/* The idea behind these two CLOSE macros is to distinguish between
   two cases: one when the job we've been doing is finished, and we
   want to close the connection and leave, and two when something is
   seriously wrong and we're closing the connection as part of
   cleanup.

   In case of keep_alive, CLOSE_FINISH should leave the connection
   open, while CLOSE_INVALIDATE should still close it.

   Note that the semantics of the flag `keep_alive' is "this
   connection *will* be reused (the server has promised not to close
   the connection once we're done)", while the semantics of
   `pc_active_p && (fd) == pc_last_fd' is "we're *now* using an
   active, registered connection".  */

#define CLOSE_FINISH(fd) do {                   \
  if (!keep_alive)                              \
    {                                           \
      if (pconn_active && (fd) == pconn.socket) \
        invalidate_persistent ();               \
      else                                      \
        {                                       \
          fd_close (fd);                        \
          fd = -1;                              \
        }                                       \
    }                                           \
} while (0)

#define CLOSE_INVALIDATE(fd) do {               \
  if (pconn_active && (fd) == pconn.socket)     \
    invalidate_persistent ();                   \
  else                                          \
    fd_close (fd);                              \
  fd = -1;                                      \
} while (0)
\f
struct http_stat
{
  long len;                     /* received length */
  long contlen;                 /* expected length */
  long restval;                 /* the restart value */
  int res;                      /* the result of last read */
  char *newloc;                 /* new location (redirection) */
  char *remote_time;            /* remote time-stamp string */
  char *error;                  /* textual HTTP error */
  int statcode;                 /* status code */
  double dltime;                /* time of the download in msecs */
  int no_truncate;              /* whether truncating the file is
                                   forbidden. */
  const char *referer;          /* value of the referer header. */
  char **local_file;            /* local file. */
};

static void
free_hstat (struct http_stat *hs)
{
  xfree_null (hs->newloc);
  xfree_null (hs->remote_time);
  xfree_null (hs->error);

  /* Guard against being called twice. */
  hs->newloc = NULL;
  hs->remote_time = NULL;
  hs->error = NULL;
}

static char *create_authorization_line PARAMS ((const char *, const char *,
                                                const char *, const char *,
                                                const char *));
static char *basic_authentication_encode PARAMS ((const char *, const char *));
static int known_authentication_scheme_p PARAMS ((const char *));

time_t http_atotm PARAMS ((const char *));

#define BEGINS_WITH(line, string_constant)                              \
  (!strncasecmp (line, string_constant, sizeof (string_constant) - 1)   \
   && (ISSPACE (line[sizeof (string_constant) - 1])                     \
       || !line[sizeof (string_constant) - 1]))

/* Retrieve a document through HTTP protocol.  It recognizes status
   code, and correctly handles redirections.  It closes the network
   socket.  If it receives an error from the functions below it, it
   will print it if there is enough information to do so (almost
   always), returning the error to the caller (i.e. http_loop).

   Various HTTP parameters are stored to hs.

   If PROXY is non-NULL, the connection will be made to the proxy
   server, and u->url will be requested.  */
static uerr_t
gethttp (struct url *u, struct http_stat *hs, int *dt, struct url *proxy)
{
  struct request *req;

  char *type;
  char *user, *passwd;
  char *proxyauth;
  int statcode;
  int write_error;
  long contlen, contrange;
  struct url *conn;
  FILE *fp;

  int sock = -1;

  /* Whether authorization has been already tried. */
  int auth_tried_already = 0;

  /* Whether our connection to the remote host is through SSL.  */
  int using_ssl = 0;

  char *head;
  struct response *resp;
  char hdrval[256];
  char *message;

  /* Whether this connection will be kept alive after the HTTP request
     is done. */
  int keep_alive;

  /* Whether keep-alive should be inhibited. */
  int inhibit_keep_alive = !opt.http_keep_alive;

  /* Headers sent when using POST. */
  long post_data_size = 0;

  int host_lookup_failed = 0;

#ifdef HAVE_SSL
  if (u->scheme == SCHEME_HTTPS)
    {
      /* Initialize the SSL context.  After this has once been done,
         it becomes a no-op.  */
      switch (ssl_init ())
        {
        case SSLERRCTXCREATE:
          /* this is fatal */
          logprintf (LOG_NOTQUIET, _("Failed to set up an SSL context\n"));
          return SSLERRCTXCREATE;
        case SSLERRCERTFILE:
          /* try without certfile */
          logprintf (LOG_NOTQUIET,
                     _("Failed to load certificates from %s\n"),
                     opt.sslcertfile);
          logprintf (LOG_NOTQUIET,
                     _("Trying without the specified certificate\n"));
          break;
        case SSLERRCERTKEY:
          logprintf (LOG_NOTQUIET,
                     _("Failed to get certificate key from %s\n"),
                     opt.sslcertkey);
          logprintf (LOG_NOTQUIET,
                     _("Trying without the specified certificate\n"));
          break;
        default:
          break;
        }
    }
#endif /* HAVE_SSL */

  if (!(*dt & HEAD_ONLY))
    /* If we're doing a GET on the URL, as opposed to just a HEAD, we need to
       know the local filename so we can save to it. */
    assert (*hs->local_file != NULL);

  auth_tried_already = 0;

  /* Initialize certain elements of struct http_stat.  */
  hs->len = 0L;
  hs->contlen = -1;
  hs->res = -1;
  hs->newloc = NULL;
  hs->remote_time = NULL;
  hs->error = NULL;

  conn = u;

  proxyauth = NULL;
  if (proxy)
    {
      char *proxy_user, *proxy_passwd;
      /* For normal username and password, URL components override
         command-line/wgetrc parameters.  With proxy
         authentication, it's the reverse, because proxy URLs are
         normally the "permanent" ones, so command-line args
         should take precedence.  */
      if (opt.proxy_user && opt.proxy_passwd)
        {
          proxy_user = opt.proxy_user;
          proxy_passwd = opt.proxy_passwd;
        }
      else
        {
          proxy_user = proxy->user;
          proxy_passwd = proxy->passwd;
        }
      /* #### This does not appear right.  Can't the proxy request,
         say, `Digest' authentication?  */
      if (proxy_user && proxy_passwd)
        proxyauth = basic_authentication_encode (proxy_user, proxy_passwd);

      /* If we're using a proxy, we will be connecting to the proxy
         server.  */
      conn = proxy;
    }

  /* Prepare the request to send. */

  req = request_new ();
  {
    const char *meth = "GET";
    if (*dt & HEAD_ONLY)
      meth = "HEAD";
    else if (opt.post_file_name || opt.post_data)
      meth = "POST";
    /* Use the full path, i.e. one that includes the leading slash and
       the query string.  E.g. if u->path is "foo/bar" and u->query is
       "param=value", full_path will be "/foo/bar?param=value".  */
    request_set_method (req, meth,
                        proxy ? xstrdup (u->url) : url_full_path (u));
  }

  request_set_header (req, "Referer", (char *) hs->referer, rel_none);
  if (*dt & SEND_NOCACHE)
    request_set_header (req, "Pragma", "no-cache", rel_none);
  if (hs->restval)
    request_set_header (req, "Range",
                        aprintf ("bytes=%ld-", hs->restval), rel_value);
  if (opt.useragent)
    request_set_header (req, "User-Agent", opt.useragent, rel_none);
  else
    request_set_header (req, "User-Agent",
                        aprintf ("Wget/%s", version_string), rel_value);
  request_set_header (req, "Accept", "*/*", rel_none);

  /* Find the username and password for authentication. */
  user = u->user;
  passwd = u->passwd;
  search_netrc (u->host, (const char **)&user, (const char **)&passwd, 0);
  user = user ? user : opt.http_user;
  passwd = passwd ? passwd : opt.http_passwd;

  if (user && passwd)
    {
      /* We have the username and the password, but haven't tried
         any authorization yet.  Let's see if the "Basic" method
         works.  If not, we'll come back here and construct a
         proper authorization method with the right challenges.

         If we didn't employ this kind of logic, every URL that
         requires authorization would have to be processed twice,
         which is very suboptimal and generates a bunch of false
         "unauthorized" errors in the server log.

         #### But this logic also has a serious problem when used
         with stronger authentications: we *first* transmit the
         username and the password in clear text, and *then* attempt a
         stronger authentication scheme.  That cannot be right!  We
         are only fortunate that almost everyone still uses the
         `Basic' scheme anyway.

         There should be an option to prevent this from happening, for
         those who use strong authentication schemes and value their
         passwords.  */
      request_set_header (req, "Authorization",
                          basic_authentication_encode (user, passwd),
                          rel_value);
    }

  {
    /* Whether we need to print the host header with braces around
       host, e.g. "Host: [3ffe:8100:200:2::2]:1234" instead of the
       usual "Host: symbolic-name:1234". */
    int squares = strchr (u->host, ':') != NULL;
    if (u->port == scheme_default_port (u->scheme))
      request_set_header (req, "Host",
                          aprintf (squares ? "[%s]" : "%s", u->host),
                          rel_value);
    else
      request_set_header (req, "Host",
                          aprintf (squares ? "[%s]:%d" : "%s:%d",
                                   u->host, u->port),
                          rel_value);
  }

  if (!inhibit_keep_alive)
    request_set_header (req, "Connection", "Keep-Alive", rel_none);

  if (opt.cookies)
    request_set_header (req, "Cookie",
                        cookie_header (wget_cookie_jar,
                                       u->host, u->port, u->path,
#ifdef HAVE_SSL
                                       u->scheme == SCHEME_HTTPS
#else
                                       0
#endif
                                       ),
                        rel_value);

  if (opt.post_data || opt.post_file_name)
    {
      request_set_header (req, "Content-Type",
                          "application/x-www-form-urlencoded", rel_none);
      if (opt.post_data)
        post_data_size = strlen (opt.post_data);
      else
        {
          post_data_size = file_size (opt.post_file_name);
          if (post_data_size == -1)
            {
              logprintf (LOG_NOTQUIET, "POST data file missing: %s\n",
                         opt.post_file_name);
              post_data_size = 0;
            }
        }
      request_set_header (req, "Content-Length",
                          aprintf ("Content-Length: %ld", post_data_size),
                          rel_value);
    }

  /* Add the user headers. */
  if (opt.user_headers)
    {
      int i;
      for (i = 0; opt.user_headers[i]; i++)
        request_set_user_header (req, opt.user_headers[i]);
    }

 retry_with_auth:
  /* We need to come back here when the initial attempt to retrieve
     without authorization header fails.  (Expected to happen at least
     for the Digest authorization scheme.)  */

  keep_alive = 0;

  /* Establish the connection.  */

  if (!inhibit_keep_alive)
    {
      /* Look for a persistent connection to target host, unless a
         proxy is used.  The exception is when SSL is in use, in which
         case the proxy is nothing but a passthrough to the target
         host, registered as a connection to the latter.  */
      struct url *relevant = conn;
#ifdef HAVE_SSL
      if (u->scheme == SCHEME_HTTPS)
        relevant = u;
#endif

      if (persistent_available_p (relevant->host, relevant->port,
#ifdef HAVE_SSL
                                  relevant->scheme == SCHEME_HTTPS,
#else
                                  0,
#endif
                                  &host_lookup_failed))
        {
          sock = pconn.socket;
          using_ssl = pconn.ssl;
          logprintf (LOG_VERBOSE, _("Reusing existing connection to %s:%d.\n"),
                     pconn.host, pconn.port);
          DEBUGP (("Reusing fd %d.\n", sock));
        }
    }

  if (sock < 0)
    {
      /* In its current implementation, persistent_available_p will
         look up conn->host in some cases.  If that lookup failed, we
         don't need to bother with connect_to_host.  */
      if (host_lookup_failed)
        return HOSTERR;

      sock = connect_to_host (conn->host, conn->port);
      if (sock == E_HOST)
        return HOSTERR;
      else if (sock < 0)
        return (retryable_socket_connect_error (errno)
                ? CONERROR : CONIMPOSSIBLE);

#ifdef HAVE_SSL
      if (proxy && u->scheme == SCHEME_HTTPS)
        {
          /* When requesting SSL URLs through proxies, use the
             CONNECT method to request passthrough.  */
          struct request *connreq = request_new ();
          request_set_method (connreq, "CONNECT",
                              aprintf ("%s:%d", u->host, u->port));
          if (proxyauth)
            {
              request_set_header (connreq, "Proxy-Authorization",
                                  proxyauth, rel_value);
              /* Now that PROXYAUTH is part of the CONNECT request,
                 zero it out so we don't send proxy authorization with
                 the regular request below.  */
              proxyauth = NULL;
            }

          write_error = request_send (connreq, sock);
          request_free (connreq);
          if (write_error < 0)
            {
              logprintf (LOG_VERBOSE, _("Failed writing to proxy: %s.\n"),
                         strerror (errno));
              CLOSE_INVALIDATE (sock);
              return WRITEFAILED;
            }

          head = fd_read_http_head (sock);
          if (!head)
            {
              logprintf (LOG_VERBOSE, _("Failed reading proxy response: %s\n"),
                         strerror (errno));
              CLOSE_INVALIDATE (sock);
              return HERR;
            }
          message = NULL;
          if (!*head)
            {
              xfree (head);
              goto failed_tunnel;
            }
          DEBUGP (("proxy responded with: [%s]\n", head));

          resp = response_new (head);
          statcode = response_status (resp, &message);
          response_free (resp);
          if (statcode != 200)
            {
            failed_tunnel:
              logprintf (LOG_NOTQUIET, _("Proxy tunneling failed: %s"),
                         message ? message : "?");
              xfree_null (message);
              return CONSSLERR;
            }
          xfree (message);

          /* SOCK is now *really* connected to u->host, so update CONN
             to reflect this.  That way register_persistent will
             register SOCK as being connected to u->host:u->port.  */
          conn = u;
        }

      if (conn->scheme == SCHEME_HTTPS)
        {
          if (!ssl_connect (sock))
            {
              fd_close (sock);
              return CONSSLERR;
            }
          using_ssl = 1;
        }
#endif /* HAVE_SSL */
    }

  /* Send the request to server.  */
  write_error = request_send (req, sock);

  if (write_error >= 0)
    {
      if (opt.post_data)
        {
          DEBUGP (("[POST data: %s]\n", opt.post_data));
          write_error = fd_write (sock, opt.post_data, post_data_size, -1);
        }
      else if (opt.post_file_name && post_data_size != 0)
        write_error = post_file (sock, opt.post_file_name, post_data_size);
    }

  if (write_error < 0)
    {
      logprintf (LOG_VERBOSE, _("Failed writing HTTP request: %s.\n"),
                 strerror (errno));
      CLOSE_INVALIDATE (sock);
      request_free (req);
      return WRITEFAILED;
    }
  logprintf (LOG_VERBOSE, _("%s request sent, awaiting response... "),
             proxy ? "Proxy" : "HTTP");
  contlen = contrange = -1;
  type = NULL;
  statcode = -1;
  *dt &= ~RETROKF;

  head = fd_read_http_head (sock);
  if (!head)
    {
      if (errno == 0)
        {
          logputs (LOG_NOTQUIET, _("No data received.\n"));
          CLOSE_INVALIDATE (sock);
          request_free (req);
          return HEOF;
        }
      else
        {
          logprintf (LOG_NOTQUIET, _("Read error (%s) in headers.\n"),
                     strerror (errno));
          CLOSE_INVALIDATE (sock);
          request_free (req);
          return HERR;
        }
    }
  DEBUGP (("\n---response begin---\n%s---response end---\n", head));

  resp = response_new (head);

  /* Check for status line.  */
  message = NULL;
  statcode = response_status (resp, &message);
  if (!opt.server_response)
    logprintf (LOG_VERBOSE, "%2d %s\n", statcode, message ? message : "");
  else
    {
      logprintf (LOG_VERBOSE, "\n");
      print_server_response (resp, "  ");
    }

  if (response_header_copy (resp, "Content-Length", hdrval, sizeof (hdrval)))
    contlen = strtol (hdrval, NULL, 10);

  /* Check for keep-alive related responses. */
  if (!inhibit_keep_alive && contlen != -1)
    {
      if (response_header_copy (resp, "Keep-Alive", NULL, 0))
        keep_alive = 1;
      else if (response_header_copy (resp, "Connection", hdrval,
                                     sizeof (hdrval)))
        {
          if (0 == strcasecmp (hdrval, "Keep-Alive"))
            keep_alive = 1;
        }
    }
  if (keep_alive)
    /* The server has promised that it will not close the connection
       when we're done.  This means that we can register it.  */
    register_persistent (conn->host, conn->port, sock, using_ssl);

  if (statcode == HTTP_STATUS_UNAUTHORIZED)
    {
      /* Authorization is required.  */
      skip_body (sock, contlen);
      CLOSE_FINISH (sock);
      if (auth_tried_already || !(user && passwd))
        {
          /* If we have tried it already, then there is not point
             retrying it.  */
          logputs (LOG_NOTQUIET, _("Authorization failed.\n"));
        }
      else
        {
          char *www_authenticate = response_header_strdup (resp,
                                                           "WWW-Authenticate");
          /* If the authentication scheme is unknown or if it's the
             "Basic" authentication (which we try by default), there's
             no sense in retrying.  */
          if (!www_authenticate
              || !known_authentication_scheme_p (www_authenticate)
              || BEGINS_WITH (www_authenticate, "Basic"))
            {
              xfree_null (www_authenticate);
              logputs (LOG_NOTQUIET, _("Unknown authentication scheme.\n"));
            }
          else
            {
              char *pth;
              auth_tried_already = 1;
              pth = url_full_path (u);
              request_set_header (req, "Authorization",
                                  create_authorization_line (www_authenticate,
                                                             user, passwd,
                                                             request_method (req),
                                                             pth),
                                  rel_value);
              xfree (pth);
              xfree (www_authenticate);
              goto retry_with_auth;
            }
        }
      request_free (req);
      return AUTHFAILED;
    }
  request_free (req);

  hs->statcode = statcode;
  if (statcode == -1)
    hs->error = xstrdup (_("Malformed status line"));
  else if (!*message)
    hs->error = xstrdup (_("(no description)"));
  else
    hs->error = xstrdup (message);

  type = response_header_strdup (resp, "Content-Type");
  if (type)
    {
      char *tmp = strchr (type, ';');
      if (tmp)
        {
          while (tmp > type && ISSPACE (tmp[-1]))
            --tmp;
          *tmp = '\0';
        }
    }
  hs->newloc = response_header_strdup (resp, "Location");
  hs->remote_time = response_header_strdup (resp, "Last-Modified");
  {
    char *set_cookie = response_header_strdup (resp, "Set-Cookie");
    if (set_cookie)
      {
        /* The jar should have been created by now. */
        assert (wget_cookie_jar != NULL);
        cookie_handle_set_cookie (wget_cookie_jar, u->host, u->port, u->path,
                                  set_cookie);
        xfree (set_cookie);
      }
  }
  if (response_header_copy (resp, "Content-Range", hdrval, sizeof (hdrval)))
    {
      long first_byte_pos, last_byte_pos, entity_length;
      if (parse_content_range (hdrval, &first_byte_pos, &last_byte_pos,
                               &entity_length))
        contrange = first_byte_pos;
    }
  response_free (resp);

  /* 20x responses are counted among successful by default.  */
  if (H_20X (statcode))
    *dt |= RETROKF;

  /* Return if redirected.  */
  if (H_REDIRECTED (statcode) || statcode == HTTP_STATUS_MULTIPLE_CHOICES)
    {
      /* RFC2068 says that in case of the 300 (multiple choices)
         response, the server can output a preferred URL through
         `Location' header; otherwise, the request should be treated
         like GET.  So, if the location is set, it will be a
         redirection; otherwise, just proceed normally.  */
      if (statcode == HTTP_STATUS_MULTIPLE_CHOICES && !hs->newloc)
        *dt |= RETROKF;
      else
        {
          logprintf (LOG_VERBOSE,
                     _("Location: %s%s\n"),
                     hs->newloc ? hs->newloc : _("unspecified"),
                     hs->newloc ? _(" [following]") : "");
          if (keep_alive)
            skip_body (sock, contlen);
          CLOSE_FINISH (sock);
          xfree_null (type);
          return NEWLOCATION;
        }
    }

  /* If content-type is not given, assume text/html.  This is because
     of the multitude of broken CGI's that "forget" to generate the
     content-type.  */
  if (!type ||
        0 == strncasecmp (type, TEXTHTML_S, strlen (TEXTHTML_S)) ||
        0 == strncasecmp (type, TEXTXHTML_S, strlen (TEXTXHTML_S)))
    *dt |= TEXTHTML;
  else
    *dt &= ~TEXTHTML;

  if (opt.html_extension && (*dt & TEXTHTML))
    /* -E / --html-extension / html_extension = on was specified, and this is a
       text/html file.  If some case-insensitive variation on ".htm[l]" isn't
       already the file's suffix, tack on ".html". */
    {
      char*  last_period_in_local_filename = strrchr(*hs->local_file, '.');

      if (last_period_in_local_filename == NULL
          || !(0 == strcasecmp (last_period_in_local_filename, ".htm")
               || 0 == strcasecmp (last_period_in_local_filename, ".html")))
        {
          size_t  local_filename_len = strlen(*hs->local_file);
          
          *hs->local_file = xrealloc(*hs->local_file,
                                     local_filename_len + sizeof(".html"));
          strcpy(*hs->local_file + local_filename_len, ".html");

          *dt |= ADDED_HTML_EXTENSION;
        }
    }

  if (contrange == -1)
    {
      /* We did not get a content-range header.  This means that the
         server did not honor our `Range' request.  Normally, this
         means we should reset hs->restval and continue normally.  */

      /* However, if `-c' is used, we need to be a bit more careful:

         1. If `-c' is specified and the file already existed when
         Wget was started, it would be a bad idea for us to start
         downloading it from scratch, effectively truncating it.  I
         believe this cannot happen unless `-c' was specified.

         2. If `-c' is used on a file that is already fully
         downloaded, we're requesting bytes after the end of file,
         which can result in server not honoring `Range'.  If this is
         the case, `Content-Length' will be equal to the length of the
         file.  */
      if (opt.always_rest)
        {
          /* Check for condition #2. */
          if (hs->restval > 0               /* restart was requested. */
              && contlen != -1              /* we got content-length. */
              && hs->restval >= contlen     /* file fully downloaded
                                               or has shrunk.  */
              )
            {
              logputs (LOG_VERBOSE, _("\
\n    The file is already fully retrieved; nothing to do.\n\n"));
              /* In case the caller inspects. */
              hs->len = contlen;
              hs->res = 0;
              /* Mark as successfully retrieved. */
              *dt |= RETROKF;
              xfree_null (type);
              CLOSE_INVALIDATE (sock);  /* would be CLOSE_FINISH, but there
                                           might be more bytes in the body. */
              return RETRUNNEEDED;
            }

          /* Check for condition #1. */
          if (hs->no_truncate)
            {
              logprintf (LOG_NOTQUIET,
                         _("\
\n\
Continued download failed on this file, which conflicts with `-c'.\n\
Refusing to truncate existing file `%s'.\n\n"), *hs->local_file);
              xfree_null (type);
              CLOSE_INVALIDATE (sock); /* see above */
              return CONTNOTSUPPORTED;
            }

          /* Fallthrough */
        }

      hs->restval = 0;
    }
  else if (contrange != hs->restval ||
           (H_PARTIAL (statcode) && contrange == -1))
    {
      /* This means the whole request was somehow misunderstood by the
         server.  Bail out.  */
      xfree_null (type);
      CLOSE_INVALIDATE (sock);
      return RANGEERR;
    }

  if (hs->restval)
    {
      if (contlen != -1)
        contlen += contrange;
      else
        contrange = -1;        /* If conent-length was not sent,
                                  content-range will be ignored.  */
    }
  hs->contlen = contlen;

  if (opt.verbose)
    {
      if (*dt & RETROKF)
        {
          /* No need to print this output if the body won't be
             downloaded at all, or if the original server response is
             printed.  */
          logputs (LOG_VERBOSE, _("Length: "));
          if (contlen != -1)
            {
              logputs (LOG_VERBOSE, legible (contlen));
              if (contrange != -1)
                logprintf (LOG_VERBOSE, _(" (%s to go)"),
                           legible (contlen - contrange));
            }
          else
            logputs (LOG_VERBOSE,
                     opt.ignore_length ? _("ignored") : _("unspecified"));
          if (type)
            logprintf (LOG_VERBOSE, " [%s]\n", type);
          else
            logputs (LOG_VERBOSE, "\n");
        }
    }
  xfree_null (type);
  type = NULL;                  /* We don't need it any more.  */

  /* Return if we have no intention of further downloading.  */
  if (!(*dt & RETROKF) || (*dt & HEAD_ONLY))
    {
      /* In case the caller cares to look...  */
      hs->len = 0L;
      hs->res = 0;
      xfree_null (type);
      /* Pre-1.10 Wget used CLOSE_INVALIDATE here.  Now we trust the
         servers not to send body in response to a HEAD request.  If
         you encounter such a server (more likely a broken CGI), use
         `--no-http-keep-alive'.  */
      CLOSE_FINISH (sock);
      return RETRFINISHED;
    }

  /* Open the local file.  */
  if (!opt.dfp)
    {
      mkalldirs (*hs->local_file);
      if (opt.backups)
        rotate_backups (*hs->local_file);
      fp = fopen (*hs->local_file, hs->restval ? "ab" : "wb");
      if (!fp)
        {
          logprintf (LOG_NOTQUIET, "%s: %s\n", *hs->local_file, strerror (errno));
          CLOSE_INVALIDATE (sock); /* would be CLOSE_FINISH, but there
                                      might be more bytes in the body. */
          return FOPENERR;
        }
    }
  else                          /* opt.dfp */
    {
      extern int global_download_count;
      fp = opt.dfp;
      /* To ensure that repeated "from scratch" downloads work for -O
         files, we rewind the file pointer, unless restval is
         non-zero.  (This works only when -O is used on regular files,
         but it's still a valuable feature.)

         However, this loses when more than one URL is specified on
         the command line the second rewinds eradicates the contents
         of the first download.  Thus we disable the above trick for
         all the downloads except the very first one.

         #### A possible solution to this would be to remember the
         file position in the output document and to seek to that
         position, instead of rewinding.

         We don't truncate stdout, since that breaks
         "wget -O - [...] >> foo".
      */
      if (!hs->restval && global_download_count == 0 && opt.dfp != stdout)
        {
          /* This will silently fail for streams that don't correspond
             to regular files, but that's OK.  */
          rewind (fp);
          /* ftruncate is needed because opt.dfp is opened in append
             mode if opt.always_rest is set.  */
          ftruncate (fileno (fp), 0);
          clearerr (fp);
        }
    }

  /* #### This confuses the code that checks for file size.  There
     should be some overhead information.  */
  if (opt.save_headers)
    fwrite (head, 1, strlen (head), fp);

  /* Get the contents of the document.  */
  hs->res = fd_read_body (sock, fp, &hs->len, hs->restval,
                          (contlen != -1 ? contlen : 0),
                          keep_alive, &hs->dltime);

  if (hs->res >= 0)
    CLOSE_FINISH (sock);
  else
    CLOSE_INVALIDATE (sock);

  {
    /* Close or flush the file.  We have to be careful to check for
       error here.  Checking the result of fwrite() is not enough --
       errors could go unnoticed!  */
    int flush_res;
    if (!opt.dfp)
      flush_res = fclose (fp);
    else
      flush_res = fflush (fp);
    if (flush_res == EOF)
      hs->res = -2;
  }
  if (hs->res == -2)
    return FWRITEERR;
  return RETRFINISHED;
}

/* The genuine HTTP loop!  This is the part where the retrieval is
   retried, and retried, and retried, and...  */
uerr_t
http_loop (struct url *u, char **newloc, char **local_file, const char *referer,
           int *dt, struct url *proxy)
{
  int count;
  int use_ts, got_head = 0;     /* time-stamping info */
  char *filename_plus_orig_suffix;
  char *local_filename = NULL;
  char *tms, *locf, *tmrate;
  uerr_t err;
  time_t tml = -1, tmr = -1;    /* local and remote time-stamps */
  long local_size = 0;          /* the size of the local file */
  size_t filename_len;
  struct http_stat hstat;       /* HTTP status */
  struct stat st;
  char *dummy = NULL;

  /* This used to be done in main(), but it's a better idea to do it
     here so that we don't go through the hoops if we're just using
     FTP or whatever. */
  if (opt.cookies)
    {
      if (!wget_cookie_jar)
        wget_cookie_jar = cookie_jar_new ();
      if (opt.cookies_input && !cookies_loaded_p)
        {
          cookie_jar_load (wget_cookie_jar, opt.cookies_input);
          cookies_loaded_p = 1;
        }
    }

  *newloc = NULL;

  /* Warn on (likely bogus) wildcard usage in HTTP.  Don't use
     has_wildcards_p because it would also warn on `?', and we know that
     shows up in CGI paths a *lot*.  */
  if (strchr (u->url, '*'))
    logputs (LOG_VERBOSE, _("Warning: wildcards not supported in HTTP.\n"));

  /* Determine the local filename.  */
  if (local_file && *local_file)
    hstat.local_file = local_file;
  else if (local_file)
    {
      *local_file = url_file_name (u);
      hstat.local_file = local_file;
    }
  else
    {
      dummy = url_file_name (u);
      hstat.local_file = &dummy;
    }

  if (!opt.output_document)
    locf = *hstat.local_file;
  else
    locf = opt.output_document;

  hstat.referer = referer;

  filename_len = strlen (*hstat.local_file);
  filename_plus_orig_suffix = alloca (filename_len + sizeof (".orig"));

  if (opt.noclobber && file_exists_p (*hstat.local_file))
    {
      /* If opt.noclobber is turned on and file already exists, do not
         retrieve the file */
      logprintf (LOG_VERBOSE, _("\
File `%s' already there, will not retrieve.\n"), *hstat.local_file);
      /* If the file is there, we suppose it's retrieved OK.  */
      *dt |= RETROKF;

      /* #### Bogusness alert.  */
      /* If its suffix is "html" or "htm" or similar, assume text/html.  */
      if (has_html_suffix_p (*hstat.local_file))
        *dt |= TEXTHTML;

      xfree_null (dummy);
      return RETROK;
    }

  use_ts = 0;
  if (opt.timestamping)
    {
      int local_dot_orig_file_exists = 0;

      if (opt.backup_converted)
        /* If -K is specified, we'll act on the assumption that it was specified
           last time these files were downloaded as well, and instead of just
           comparing local file X against server file X, we'll compare local
           file X.orig (if extant, else X) against server file X.  If -K
           _wasn't_ specified last time, or the server contains files called
           *.orig, -N will be back to not operating correctly with -k. */
        {
          /* Would a single s[n]printf() call be faster?  --dan

             Definitely not.  sprintf() is horribly slow.  It's a
             different question whether the difference between the two
             affects a program.  Usually I'd say "no", but at one
             point I profiled Wget, and found that a measurable and
             non-negligible amount of time was lost calling sprintf()
             in url.c.  Replacing sprintf with inline calls to
             strcpy() and long_to_string() made a difference.
             --hniksic */
          memcpy (filename_plus_orig_suffix, *hstat.local_file, filename_len);
          memcpy (filename_plus_orig_suffix + filename_len,
                  ".orig", sizeof (".orig"));

          /* Try to stat() the .orig file. */
          if (stat (filename_plus_orig_suffix, &st) == 0)
            {
              local_dot_orig_file_exists = 1;
              local_filename = filename_plus_orig_suffix;
            }
        }      

      if (!local_dot_orig_file_exists)
        /* Couldn't stat() <file>.orig, so try to stat() <file>. */
        if (stat (*hstat.local_file, &st) == 0)
          local_filename = *hstat.local_file;

      if (local_filename != NULL)
        /* There was a local file, so we'll check later to see if the version
           the server has is the same version we already have, allowing us to
           skip a download. */
        {
          use_ts = 1;
          tml = st.st_mtime;
#ifdef WINDOWS
          /* Modification time granularity is 2 seconds for Windows, so
             increase local time by 1 second for later comparison. */
          tml++;
#endif
          local_size = st.st_size;
          got_head = 0;
        }
    }
  /* Reset the counter.  */
  count = 0;
  *dt = 0 | ACCEPTRANGES;
  /* THE loop */
  do
    {
      /* Increment the pass counter.  */
      ++count;
      sleep_between_retrievals (count);
      /* Get the current time string.  */
      tms = time_str (NULL);
      /* Print fetch message, if opt.verbose.  */
      if (opt.verbose)
        {
          char *hurl = url_string (u, 1);
          char tmp[15];
          strcpy (tmp, "        ");
          if (count > 1)
            sprintf (tmp, _("(try:%2d)"), count);
          logprintf (LOG_VERBOSE, "--%s--  %s\n  %s => `%s'\n",
                     tms, hurl, tmp, locf);
#ifdef WINDOWS
          ws_changetitle (hurl, 1);
#endif
          xfree (hurl);
        }

      /* Default document type is empty.  However, if spider mode is
         on or time-stamping is employed, HEAD_ONLY commands is
         encoded within *dt.  */
      if (opt.spider || (use_ts && !got_head))
        *dt |= HEAD_ONLY;
      else
        *dt &= ~HEAD_ONLY;
      /* Assume no restarting.  */
      hstat.restval = 0L;
      /* Decide whether or not to restart.  */
      if (((count > 1 && (*dt & ACCEPTRANGES)) || opt.always_rest)
          /* #### this calls access() and then stat(); could be optimized. */
          && file_exists_p (locf))
        if (stat (locf, &st) == 0 && S_ISREG (st.st_mode))
          hstat.restval = st.st_size;

      /* In `-c' is used and the file is existing and non-empty,
         refuse to truncate it if the server doesn't support continued
         downloads.  */
      hstat.no_truncate = 0;
      if (opt.always_rest && hstat.restval)
        hstat.no_truncate = 1;

      /* Decide whether to send the no-cache directive.  We send it in
         two cases:
           a) we're using a proxy, and we're past our first retrieval.
              Some proxies are notorious for caching incomplete data, so
              we require a fresh get.
           b) caching is explicitly inhibited. */
      if ((proxy && count > 1)  /* a */
          || !opt.allow_cache   /* b */
          )
        *dt |= SEND_NOCACHE;
      else
        *dt &= ~SEND_NOCACHE;

      /* Try fetching the document, or at least its head.  */
      err = gethttp (u, &hstat, dt, proxy);

      /* It's unfortunate that wget determines the local filename before finding
         out the Content-Type of the file.  Barring a major restructuring of the
         code, we need to re-set locf here, since gethttp() may have xrealloc()d
         *hstat.local_file to tack on ".html". */
      if (!opt.output_document)
        locf = *hstat.local_file;
      else
        locf = opt.output_document;

      /* Time?  */
      tms = time_str (NULL);
      /* Get the new location (with or without the redirection).  */
      if (hstat.newloc)
        *newloc = xstrdup (hstat.newloc);
      switch (err)
        {
        case HERR: case HEOF: case CONSOCKERR: case CONCLOSED:
        case CONERROR: case READERR: case WRITEFAILED:
        case RANGEERR:
          /* Non-fatal errors continue executing the loop, which will
             bring them to "while" statement at the end, to judge
             whether the number of tries was exceeded.  */
          free_hstat (&hstat);
          printwhat (count, opt.ntry);
          continue;
          break;
        case HOSTERR: case CONIMPOSSIBLE: case PROXERR: case AUTHFAILED: 
        case SSLERRCTXCREATE: case CONTNOTSUPPORTED:
          /* Fatal errors just return from the function.  */
          free_hstat (&hstat);
          xfree_null (dummy);
          return err;
          break;
        case FWRITEERR: case FOPENERR:
          /* Another fatal error.  */
          logputs (LOG_VERBOSE, "\n");
          logprintf (LOG_NOTQUIET, _("Cannot write to `%s' (%s).\n"),
                     *hstat.local_file, strerror (errno));
          free_hstat (&hstat);
          xfree_null (dummy);
          return err;
          break;
        case CONSSLERR:
          /* Another fatal error.  */
          logputs (LOG_VERBOSE, "\n");
          logprintf (LOG_NOTQUIET, _("Unable to establish SSL connection.\n"));
          free_hstat (&hstat);
          xfree_null (dummy);
          return err;
          break;
        case NEWLOCATION:
          /* Return the new location to the caller.  */
          if (!hstat.newloc)
            {
              logprintf (LOG_NOTQUIET,
                         _("ERROR: Redirection (%d) without location.\n"),
                         hstat.statcode);
              free_hstat (&hstat);
              xfree_null (dummy);
              return WRONGCODE;
            }
          free_hstat (&hstat);
          xfree_null (dummy);
          return NEWLOCATION;
          break;
        case RETRUNNEEDED:
          /* The file was already fully retrieved. */
          free_hstat (&hstat);
          xfree_null (dummy);
          return RETROK;
          break;
        case RETRFINISHED:
          /* Deal with you later.  */
          break;
        default:
          /* All possibilities should have been exhausted.  */
          abort ();
        }
      if (!(*dt & RETROKF))
        {
          if (!opt.verbose)
            {
              /* #### Ugly ugly ugly! */
              char *hurl = url_string (u, 1);
              logprintf (LOG_NONVERBOSE, "%s:\n", hurl);
              xfree (hurl);
            }
          logprintf (LOG_NOTQUIET, _("%s ERROR %d: %s.\n"),
                     tms, hstat.statcode, hstat.error);
          logputs (LOG_VERBOSE, "\n");
          free_hstat (&hstat);
          xfree_null (dummy);
          return WRONGCODE;
        }

      /* Did we get the time-stamp?  */
      if (!got_head)
        {
          if (opt.timestamping && !hstat.remote_time)
            {
              logputs (LOG_NOTQUIET, _("\
Last-modified header missing -- time-stamps turned off.\n"));
            }
          else if (hstat.remote_time)
            {
              /* Convert the date-string into struct tm.  */
              tmr = http_atotm (hstat.remote_time);
              if (tmr == (time_t) (-1))
                logputs (LOG_VERBOSE, _("\
Last-modified header invalid -- time-stamp ignored.\n"));
            }
        }

      /* The time-stamping section.  */
      if (use_ts)
        {
          got_head = 1;
          *dt &= ~HEAD_ONLY;
          use_ts = 0;           /* no more time-stamping */
          count = 0;            /* the retrieve count for HEAD is
                                   reset */
          if (hstat.remote_time && tmr != (time_t) (-1))
            {
              /* Now time-stamping can be used validly.  Time-stamping
                 means that if the sizes of the local and remote file
                 match, and local file is newer than the remote file,
                 it will not be retrieved.  Otherwise, the normal
                 download procedure is resumed.  */
              if (tml >= tmr &&
                  (hstat.contlen == -1 || local_size == hstat.contlen))
                {
                  logprintf (LOG_VERBOSE, _("\
Server file no newer than local file `%s' -- not retrieving.\n\n"),
                             local_filename);
                  free_hstat (&hstat);
                  xfree_null (dummy);
                  return RETROK;
                }
              else if (tml >= tmr)
                logprintf (LOG_VERBOSE, _("\
The sizes do not match (local %ld) -- retrieving.\n"), local_size);
              else
                logputs (LOG_VERBOSE,
                         _("Remote file is newer, retrieving.\n"));
            }
          free_hstat (&hstat);
          continue;
        }
      if ((tmr != (time_t) (-1))
          && !opt.spider
          && ((hstat.len == hstat.contlen) ||
              ((hstat.res == 0) &&
               ((hstat.contlen == -1) ||
                (hstat.len >= hstat.contlen && !opt.kill_longer)))))
        {
          /* #### This code repeats in http.c and ftp.c.  Move it to a
             function!  */
          const char *fl = NULL;
          if (opt.output_document)
            {
              if (opt.od_known_regular)
                fl = opt.output_document;
            }
          else
            fl = *hstat.local_file;
          if (fl)
            touch (fl, tmr);
        }
      /* End of time-stamping section.  */

      if (opt.spider)
        {
          logprintf (LOG_NOTQUIET, "%d %s\n\n", hstat.statcode, hstat.error);
          xfree_null (dummy);
          return RETROK;
        }

      tmrate = retr_rate (hstat.len - hstat.restval, hstat.dltime, 0);

      if (hstat.len == hstat.contlen)
        {
          if (*dt & RETROKF)
            {
              logprintf (LOG_VERBOSE,
                         _("%s (%s) - `%s' saved [%ld/%ld]\n\n"),
                         tms, tmrate, locf, hstat.len, hstat.contlen);
              logprintf (LOG_NONVERBOSE,
                         "%s URL:%s [%ld/%ld] -> \"%s\" [%d]\n",
                         tms, u->url, hstat.len, hstat.contlen, locf, count);
            }
          ++opt.numurls;
          total_downloaded_bytes += hstat.len;

          /* Remember that we downloaded the file for later ".orig" code. */
          if (*dt & ADDED_HTML_EXTENSION)
            downloaded_file(FILE_DOWNLOADED_AND_HTML_EXTENSION_ADDED, locf);
          else
            downloaded_file(FILE_DOWNLOADED_NORMALLY, locf);

          free_hstat (&hstat);
          xfree_null (dummy);
          return RETROK;
        }
      else if (hstat.res == 0) /* No read error */
        {
          if (hstat.contlen == -1)  /* We don't know how much we were supposed
                                       to get, so assume we succeeded. */ 
            {
              if (*dt & RETROKF)
                {
                  logprintf (LOG_VERBOSE,
                             _("%s (%s) - `%s' saved [%ld]\n\n"),
                             tms, tmrate, locf, hstat.len);
                  logprintf (LOG_NONVERBOSE,
                             "%s URL:%s [%ld] -> \"%s\" [%d]\n",
                             tms, u->url, hstat.len, locf, count);
                }
              ++opt.numurls;
              total_downloaded_bytes += hstat.len;

              /* Remember that we downloaded the file for later ".orig" code. */
              if (*dt & ADDED_HTML_EXTENSION)
                downloaded_file(FILE_DOWNLOADED_AND_HTML_EXTENSION_ADDED, locf);
              else
                downloaded_file(FILE_DOWNLOADED_NORMALLY, locf);
              
              free_hstat (&hstat);
              xfree_null (dummy);
              return RETROK;
            }
          else if (hstat.len < hstat.contlen) /* meaning we lost the
                                                 connection too soon */
            {
              logprintf (LOG_VERBOSE,
                         _("%s (%s) - Connection closed at byte %ld. "),
                         tms, tmrate, hstat.len);
              printwhat (count, opt.ntry);
              free_hstat (&hstat);
              continue;
            }
          else if (!opt.kill_longer) /* meaning we got more than expected */
            {
              logprintf (LOG_VERBOSE,
                         _("%s (%s) - `%s' saved [%ld/%ld])\n\n"),
                         tms, tmrate, locf, hstat.len, hstat.contlen);
              logprintf (LOG_NONVERBOSE,
                         "%s URL:%s [%ld/%ld] -> \"%s\" [%d]\n",
                         tms, u->url, hstat.len, hstat.contlen, locf, count);
              ++opt.numurls;
              total_downloaded_bytes += hstat.len;

              /* Remember that we downloaded the file for later ".orig" code. */
              if (*dt & ADDED_HTML_EXTENSION)
                downloaded_file(FILE_DOWNLOADED_AND_HTML_EXTENSION_ADDED, locf);
              else
                downloaded_file(FILE_DOWNLOADED_NORMALLY, locf);
              
              free_hstat (&hstat);
              xfree_null (dummy);
              return RETROK;
            }
          else                  /* the same, but not accepted */
            {
              logprintf (LOG_VERBOSE,
                         _("%s (%s) - Connection closed at byte %ld/%ld. "),
                         tms, tmrate, hstat.len, hstat.contlen);
              printwhat (count, opt.ntry);
              free_hstat (&hstat);
              continue;
            }
        }
      else                      /* now hstat.res can only be -1 */
        {
          if (hstat.contlen == -1)
            {
              logprintf (LOG_VERBOSE,
                         _("%s (%s) - Read error at byte %ld (%s)."),
                         tms, tmrate, hstat.len, strerror (errno));
              printwhat (count, opt.ntry);
              free_hstat (&hstat);
              continue;
            }
          else                  /* hstat.res == -1 and contlen is given */
            {
              logprintf (LOG_VERBOSE,
                         _("%s (%s) - Read error at byte %ld/%ld (%s). "),
                         tms, tmrate, hstat.len, hstat.contlen,
                         strerror (errno));
              printwhat (count, opt.ntry);
              free_hstat (&hstat);
              continue;
            }
        }
      /* not reached */
      break;
    }
  while (!opt.ntry || (count < opt.ntry));
  return TRYLIMEXC;
}
\f
/* Converts struct tm to time_t, assuming the data in tm is UTC rather
   than local timezone.

   mktime is similar but assumes struct tm, also known as the
   "broken-down" form of time, is in local time zone.  mktime_from_utc
   uses mktime to make the conversion understanding that an offset
   will be introduced by the local time assumption.

   mktime_from_utc then measures the introduced offset by applying
   gmtime to the initial result and applying mktime to the resulting
   "broken-down" form.  The difference between the two mktime results
   is the measured offset which is then subtracted from the initial
   mktime result to yield a calendar time which is the value returned.

   tm_isdst in struct tm is set to 0 to force mktime to introduce a
   consistent offset (the non DST offset) since tm and tm+o might be
   on opposite sides of a DST change.

   Some implementations of mktime return -1 for the nonexistent
   localtime hour at the beginning of DST.  In this event, use
   mktime(tm - 1hr) + 3600.

   Schematically
     mktime(tm)   --> t+o
     gmtime(t+o)  --> tm+o
     mktime(tm+o) --> t+2o
     t+o - (t+2o - t+o) = t

   Note that glibc contains a function of the same purpose named
   `timegm' (reverse of gmtime).  But obviously, it is not universally
   available, and unfortunately it is not straightforwardly
   extractable for use here.  Perhaps configure should detect timegm
   and use it where available.

   Contributed by Roger Beeman <beeman@cisco.com>, with the help of
   Mark Baushke <mdb@cisco.com> and the rest of the Gurus at CISCO.
   Further improved by Roger with assistance from Edward J. Sabol
   based on input by Jamie Zawinski.  */

static time_t
mktime_from_utc (struct tm *t)
{
  time_t tl, tb;
  struct tm *tg;

  tl = mktime (t);
  if (tl == -1)
    {
      t->tm_hour--;
      tl = mktime (t);
      if (tl == -1)
        return -1; /* can't deal with output from strptime */
      tl += 3600;
    }
  tg = gmtime (&tl);
  tg->tm_isdst = 0;
  tb = mktime (tg);
  if (tb == -1)
    {
      tg->tm_hour--;
      tb = mktime (tg);
      if (tb == -1)
        return -1; /* can't deal with output from gmtime */
      tb += 3600;
    }
  return (tl - (tb - tl));
}

/* Check whether the result of strptime() indicates success.
   strptime() returns the pointer to how far it got to in the string.
   The processing has been successful if the string is at `GMT' or
   `+X', or at the end of the string.

   In extended regexp parlance, the function returns 1 if P matches
   "^ *(GMT|[+-][0-9]|$)", 0 otherwise.  P being NULL (which strptime
   can return) is considered a failure and 0 is returned.  */
static int
check_end (const char *p)
{
  if (!p)
    return 0;
  while (ISSPACE (*p))
    ++p;
  if (!*p
      || (p[0] == 'G' && p[1] == 'M' && p[2] == 'T')
      || ((p[0] == '+' || p[0] == '-') && ISDIGIT (p[1])))
    return 1;
  else
    return 0;
}

/* Convert the textual specification of time in TIME_STRING to the
   number of seconds since the Epoch.

   TIME_STRING can be in any of the three formats RFC2068 allows the
   HTTP servers to emit -- RFC1123-date, RFC850-date or asctime-date.
   Timezones are ignored, and should be GMT.

   Return the computed time_t representation, or -1 if the conversion
   fails.

   This function uses strptime with various string formats for parsing
   TIME_STRING.  This results in a parser that is not as lenient in
   interpreting TIME_STRING as I would like it to be.  Being based on
   strptime, it always allows shortened months, one-digit days, etc.,
   but due to the multitude of formats in which time can be
   represented, an ideal HTTP time parser would be even more
   forgiving.  It should completely ignore things like week days and
   concentrate only on the various forms of representing years,
   months, days, hours, minutes, and seconds.  For example, it would
   be nice if it accepted ISO 8601 out of the box.

   I've investigated free and PD code for this purpose, but none was
   usable.  getdate was big and unwieldy, and had potential copyright
   issues, or so I was informed.  Dr. Marcus Hennecke's atotm(),
   distributed with phttpd, is excellent, but we cannot use it because
   it is not assigned to the FSF.  So I stuck it with strptime.  */

time_t
http_atotm (const char *time_string)
{
  /* NOTE: Solaris strptime man page claims that %n and %t match white
     space, but that's not universally available.  Instead, we simply
     use ` ' to mean "skip all WS", which works under all strptime
     implementations I've tested.  */

  static const char *time_formats[] = {
    "%a, %d %b %Y %T",          /* RFC1123: Thu, 29 Jan 1998 22:12:57 */
    "%A, %d-%b-%y %T",          /* RFC850:  Thursday, 29-Jan-98 22:12:57 */
    "%a, %d-%b-%Y %T",          /* pseudo-RFC850:  Thu, 29-Jan-1998 22:12:57
                                   (google.com uses this for their cookies.) */
    "%a %b %d %T %Y"            /* asctime: Thu Jan 29 22:12:57 1998 */
  };

  int i;
  struct tm t;

  /* According to Roger Beeman, we need to initialize tm_isdst, since
     strptime won't do it.  */
  t.tm_isdst = 0;

  /* Note that under foreign locales Solaris strptime() fails to
     recognize English dates, which renders this function useless.  We
     solve this by being careful not to affect LC_TIME when
     initializing locale.

     Another solution would be to temporarily set locale to C, invoke
     strptime(), and restore it back.  This is slow and dirty,
     however, and locale support other than LC_MESSAGES can mess other
     things, so I rather chose to stick with just setting LC_MESSAGES.

     GNU strptime does not have this problem because it recognizes
     both international and local dates.  */

  for (i = 0; i < countof (time_formats); i++)
    if (check_end (strptime (time_string, time_formats[i], &t)))
      return mktime_from_utc (&t);

  /* All formats have failed.  */
  return -1;
}
\f
/* Authorization support: We support two authorization schemes:

   * `Basic' scheme, consisting of base64-ing USER:PASSWORD string;

   * `Digest' scheme, added by Junio Hamano <junio@twinsun.com>,
   consisting of answering to the server's challenge with the proper
   MD5 digests.  */

/* How many bytes it will take to store LEN bytes in base64.  */
#define BASE64_LENGTH(len) (4 * (((len) + 2) / 3))

/* Encode the string S of length LENGTH to base64 format and place it
   to STORE.  STORE will be 0-terminated, and must point to a writable
   buffer of at least 1+BASE64_LENGTH(length) bytes.  */
static void
base64_encode (const char *s, char *store, int length)
{
  /* Conversion table.  */
  static char tbl[64] = {
    'A','B','C','D','E','F','G','H',
    'I','J','K','L','M','N','O','P',
    'Q','R','S','T','U','V','W','X',
    'Y','Z','a','b','c','d','e','f',
    'g','h','i','j','k','l','m','n',
    'o','p','q','r','s','t','u','v',
    'w','x','y','z','0','1','2','3',
    '4','5','6','7','8','9','+','/'
  };
  int i;
  unsigned char *p = (unsigned char *)store;

  /* Transform the 3x8 bits to 4x6 bits, as required by base64.  */
  for (i = 0; i < length; i += 3)
    {
      *p++ = tbl[s[0] >> 2];
      *p++ = tbl[((s[0] & 3) << 4) + (s[1] >> 4)];
      *p++ = tbl[((s[1] & 0xf) << 2) + (s[2] >> 6)];
      *p++ = tbl[s[2] & 0x3f];
      s += 3;
    }
  /* Pad the result if necessary...  */
  if (i == length + 1)
    *(p - 1) = '=';
  else if (i == length + 2)
    *(p - 1) = *(p - 2) = '=';
  /* ...and zero-terminate it.  */
  *p = '\0';
}

/* Create the authentication header contents for the `Basic' scheme.
   This is done by encoding the string `USER:PASS' in base64 and
   prepending `HEADER: Basic ' to it.  */
static char *
basic_authentication_encode (const char *user, const char *passwd)
{
  char *t1, *t2, *res;
  int len1 = strlen (user) + 1 + strlen (passwd);
  int len2 = BASE64_LENGTH (len1);

  t1 = (char *)alloca (len1 + 1);
  sprintf (t1, "%s:%s", user, passwd);

  t2 = (char *)alloca (len2 + 1);
  base64_encode (t1, t2, len1);

  res = (char *)xmalloc (6 + len2 + 1);
  sprintf (res, "Basic %s", t2);

  return res;
}

#define SKIP_WS(x) do {                         \
  while (ISSPACE (*(x)))                        \
    ++(x);                                      \
} while (0)

#ifdef USE_DIGEST
/* Parse HTTP `WWW-Authenticate:' header.  AU points to the beginning
   of a field in such a header.  If the field is the one specified by
   ATTR_NAME ("realm", "opaque", and "nonce" are used by the current
   digest authorization code), extract its value in the (char*)
   variable pointed by RET.  Returns negative on a malformed header,
   or number of bytes that have been parsed by this call.  */
static int
extract_header_attr (const char *au, const char *attr_name, char **ret)
{
  const char *cp, *ep;

  ep = cp = au;

  if (strncmp (cp, attr_name, strlen (attr_name)) == 0)
    {
      cp += strlen (attr_name);
      if (!*cp)
        return -1;
      SKIP_WS (cp);
      if (*cp != '=')
        return -1;
      if (!*++cp)
        return -1;
      SKIP_WS (cp);
      if (*cp != '\"')
        return -1;
      if (!*++cp)
        return -1;
      for (ep = cp; *ep && *ep != '\"'; ep++)
        ;
      if (!*ep)
        return -1;
      xfree_null (*ret);
      *ret = strdupdelim (cp, ep);
      return ep - au + 1;
    }
  else
    return 0;
}

/* Dump the hexadecimal representation of HASH to BUF.  HASH should be
   an array of 16 bytes containing the hash keys, and BUF should be a
   buffer of 33 writable characters (32 for hex digits plus one for
   zero termination).  */
static void
dump_hash (unsigned char *buf, const unsigned char *hash)
{
  int i;

  for (i = 0; i < MD5_HASHLEN; i++, hash++)
    {
      *buf++ = XNUM_TO_digit (*hash >> 4);
      *buf++ = XNUM_TO_digit (*hash & 0xf);
    }
  *buf = '\0';
}

/* Take the line apart to find the challenge, and compose a digest
   authorization header.  See RFC2069 section 2.1.2.  */
static char *
digest_authentication_encode (const char *au, const char *user,
                              const char *passwd, const char *method,
                              const char *path)
{
  static char *realm, *opaque, *nonce;
  static struct {
    const char *name;
    char **variable;
  } options[] = {
    { "realm", &realm },
    { "opaque", &opaque },
    { "nonce", &nonce }
  };
  char *res;

  realm = opaque = nonce = NULL;

  au += 6;                      /* skip over `Digest' */
  while (*au)
    {
      int i;

      SKIP_WS (au);
      for (i = 0; i < countof (options); i++)
        {
          int skip = extract_header_attr (au, options[i].name,
                                          options[i].variable);
          if (skip < 0)
            {
              xfree_null (realm);
              xfree_null (opaque);
              xfree_null (nonce);
              return NULL;
            }
          else if (skip)
            {
              au += skip;
              break;
            }
        }
      if (i == countof (options))
        {
          while (*au && *au != '=')
            au++;
          if (*au && *++au)
            {
              SKIP_WS (au);
              if (*au == '\"')
                {
                  au++;
                  while (*au && *au != '\"')
                    au++;
                  if (*au)
                    au++;
                }
            }
        }
      while (*au && *au != ',')
        au++;
      if (*au)
        au++;
    }
  if (!realm || !nonce || !user || !passwd || !path || !method)
    {
      xfree_null (realm);
      xfree_null (opaque);
      xfree_null (nonce);
      return NULL;
    }

  /* Calculate the digest value.  */
  {
    ALLOCA_MD5_CONTEXT (ctx);
    unsigned char hash[MD5_HASHLEN];
    unsigned char a1buf[MD5_HASHLEN * 2 + 1], a2buf[MD5_HASHLEN * 2 + 1];
    unsigned char response_digest[MD5_HASHLEN * 2 + 1];

    /* A1BUF = H(user ":" realm ":" password) */
    gen_md5_init (ctx);
    gen_md5_update ((unsigned char *)user, strlen (user), ctx);
    gen_md5_update ((unsigned char *)":", 1, ctx);
    gen_md5_update ((unsigned char *)realm, strlen (realm), ctx);
    gen_md5_update ((unsigned char *)":", 1, ctx);
    gen_md5_update ((unsigned char *)passwd, strlen (passwd), ctx);
    gen_md5_finish (ctx, hash);
    dump_hash (a1buf, hash);

    /* A2BUF = H(method ":" path) */
    gen_md5_init (ctx);
    gen_md5_update ((unsigned char *)method, strlen (method), ctx);
    gen_md5_update ((unsigned char *)":", 1, ctx);
    gen_md5_update ((unsigned char *)path, strlen (path), ctx);
    gen_md5_finish (ctx, hash);
    dump_hash (a2buf, hash);

    /* RESPONSE_DIGEST = H(A1BUF ":" nonce ":" A2BUF) */
    gen_md5_init (ctx);
    gen_md5_update (a1buf, MD5_HASHLEN * 2, ctx);
    gen_md5_update ((unsigned char *)":", 1, ctx);
    gen_md5_update ((unsigned char *)nonce, strlen (nonce), ctx);
    gen_md5_update ((unsigned char *)":", 1, ctx);
    gen_md5_update (a2buf, MD5_HASHLEN * 2, ctx);
    gen_md5_finish (ctx, hash);
    dump_hash (response_digest, hash);

    res = (char*) xmalloc (strlen (user)
                           + strlen (user)
                           + strlen (realm)
                           + strlen (nonce)
                           + strlen (path)
                           + 2 * MD5_HASHLEN /*strlen (response_digest)*/
                           + (opaque ? strlen (opaque) : 0)
                           + 128);
    sprintf (res, "Digest \
username=\"%s\", realm=\"%s\", nonce=\"%s\", uri=\"%s\", response=\"%s\"",
             user, realm, nonce, path, response_digest);
    if (opaque)
      {
        char *p = res + strlen (res);
        strcat (p, ", opaque=\"");
        strcat (p, opaque);
        strcat (p, "\"");
      }
  }
  return res;
}
#endif /* USE_DIGEST */


#define BEGINS_WITH(line, string_constant)                              \
  (!strncasecmp (line, string_constant, sizeof (string_constant) - 1)   \
   && (ISSPACE (line[sizeof (string_constant) - 1])                     \
       || !line[sizeof (string_constant) - 1]))

static int
known_authentication_scheme_p (const char *au)
{
  return BEGINS_WITH (au, "Basic")
    || BEGINS_WITH (au, "Digest")
    || BEGINS_WITH (au, "NTLM");
}

#undef BEGINS_WITH

/* Create the HTTP authorization request header.  When the
   `WWW-Authenticate' response header is seen, according to the
   authorization scheme specified in that header (`Basic' and `Digest'
   are supported by the current implementation), produce an
   appropriate HTTP authorization request header.  */
static char *
create_authorization_line (const char *au, const char *user,
                           const char *passwd, const char *method,
                           const char *path)
{
  if (0 == strncasecmp (au, "Basic", 5))
    return basic_authentication_encode (user, passwd);
#ifdef USE_DIGEST
  if (0 == strncasecmp (au, "Digest", 6))
    return digest_authentication_encode (au, user, passwd, method, path);
#endif /* USE_DIGEST */
  return NULL;
}
\f
void
http_cleanup (void)
{
}