In C, signed integer overflow is undefined behavior. Thus, the compiler
is allowed to assume that it will not occur. In the check for an
overflow, the developer assumes that the signed integer of type time_t
will wrap around if it overflows. However, this behavior is undefined in
the C standard. Thus, when the compiler sees this, it simplifies t +
delta < t to delta < 0. Since delta > 0 and delta < 0 can't both be
true, the entire if statement is optimized out under certain
optimization levels. Thus, the parsedate function would return
PARSEDATE_OK with an undefined value in the time, instead of return -1 =
PARSEDATE_FAIL.

Commit 0db811b6 made some existing config files pass on unexpected
values to libcurl that made it somewhat hard to track down what was
really going on.

This code detects unquoted white spaces in the parameter when parsing a
config file as that would be one symptom and it is generally a bad
syntax anyway.

It would otherwise always clobber the return code with new function
calls and it couldn't return timeout etc.

Bug: http://curl.haxx.se/bug/view.cgi?id=1310

The comment here says that SecKeychainSearch causes a deprecation
warning when used with a minimum Mac OS X SDK version of 10.7.0, which
is correct.  However, the #if guard did not match.  It was intended to
only use the code if 10.6.0 support was enabled, but it had 10.7.0
instead.  This caused a warning if the minimum was exactly 10.7.0.

curl.h should also include <sys/select.h> on OpenBSD to reliably
pull in select().  Typically, including <sys/time.h> will be enough,
but not if strict standards-compliance is requested (e.g. by defining
_XOPEN_SOURCE).

The URI that is passed in as part of the Authorization: header needs to
be cut off at '?' if CURLAUTH_DIGEST_IE is set. Previously the code only
did when calculating the MD5sum.

Bug: http://curl.haxx.se/bug/view.cgi?id=1308
Patched-by: Sergey Tatarincev

(bug introduced in 255826c4, never present in a release)

Reported-by: Dima Tisnek
Bug: http://curl.haxx.se/mail/lib-2013-12/0006.html

POP3_TYPE_ANY, or ~0, is written to pop3c->preftype in lib/pop3c.c, an
unsigned int variable.  The result of ~0 is -1, which caused a warning
due to writing a negative number to an unsigned variable.  To fix this,
make the expression ~0U so that its value is considered the unsigned
number UINT_MAX which is what SASL_AUTH_ANY does in curl_sasl.h.

... no longer provided by the configure script

... since we depend on NSS 3.14+ because of SSL_VersionRangeSet() anyway

Bug: http://curl.haxx.se/mail/lib-2013-11/0162.html

This change introduces a dependency on NSS 3.14+.

warning: declaration of 'pipe' shadows a global declaration

Reduced the separate processing of the last quantum to be performed in
the main decoding loop and renamed some variables for consistency.

Extended the basic validation in commit e17c1b25 to return a
failure when invalid base64 characters are included.

Due to the length checks introduced in commit e17c1b25 there is no
need to allow for extra space in the output buffer for a non-padded last
quantum.

1) Renamed curl_tlsinfo to curl_tlssessioninfo as discussed on the
mailing list.
2) Renamed curl_ssl_backend to curl_sslbackend so it doesn't follow our
function naming convention.
3) Updated sessioninfo.c example accordingly.

Bug: http://curl.haxx.se/bug/view.cgi?id=1297
Reported-by: Michael Osipov

The "fixed string" function wrongly bumped the "urlnum" counter which
made curl output the total number of URLs wrong when using
{one,two,three} lists in globs.

Reported-by: Michael-O
Bug: http://curl.haxx.se/bug/view.cgi?id=1305

Added a simple example to show how one can use CURLINFO_TLS_SESSION for
obtaining extensive TLS certificate information.

Systems that define SIGPIPE_VARIABLE as a noop would not compile as
restore_pipe was defined afterwards.

This fixes a NULL dereference in the case where the client asks for
CURLINFO_TLS_SESSION data after the (TLS) session has already been
destroyed (i.e. curl_easy_perform has already completed for this
handle). Instead of crashing, we now return a CURLSSLBACKEND_NONE
error.

This is an extension to the fix in 7d80ed64

. We may
call Curl_disconnect() while cleaning up the multi handle,
which could lead to openssl sending packets, which could get
a SIGPIPE.

Signed-off-by: Jeff King <peff@peff.net>

Commit 7d80ed64

 introduced some helpers to handle
sigpipe in easy.c. However, that fix was incomplete, and we
need to add more callers in other files. The first step is
making the helpers globally accessible.

Since the functions are small and should generally end up
inlined anyway, we simply define them in the header as
static functions.

Signed-off-by: Jeff King <peff@peff.net>

This fixes a rare Happy Eyeballs bug where if the first IP family runs
out of addresses before the second-family-timer fires, and the second
IP family's first connect fails immediately, no further IPs of the
second family are attempted.

When adding entries to the DNS cache with CURLOPT_RESOLVE, they are
marked 'inuse' forever to prevent them from ever being removed in normal
operations. Still, the code that pruned out-of-date DNS entries didn't
care for the 'inuse' struct field and pruned it anyway!

Reported-by: Romulo A. Ceccon
Bug: http://curl.haxx.se/bug/view.cgi?id=1303