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.

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 Still, the code that pruned out-of-date DNS entries didn't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

Removed unnecessary SMTP_STOP state changes on failure.

Removed hard return on failure in smtp_state_data_resp().

Added new API for returning a SSL backend type and pointer, in order to
allow access to the TLS internals, that may then be used to obtain X509
certificate information for example.

Commands such as NOOP, RSET and HELP would be sent with a space at the
end of the command, for example: "NOOP ".

Otherwise a NOOP operation would be performed which a) only returns a
single line response and not a multiline response where -I needs to be
used, and b) provides an inconsistent user experience compared to that
of the POP3 and IMAP protocols.

The buffer allocated by smtp_parse_custom_request() was not freed.

singleipconnect() could return the file descriptor of an open socket
even though the function returned a CURLE_COULDNT_CONNECT error code
from commit ed1662c3 and 02fbc26d.

This could cause tests 19, 704 and 1233 to fail on FreeBSD, AIX and
Solaris.

Incorrectly processed multiline server greeting responses as "wanted"
continue responses in smtp_endofresp(), from commit f16c0de4,
which in turn broke the SMTP server detection in the test suite,
because the EHLO command would not be sent.

warning: unused variable 'smtpc'

Added a loop to smtp_statemach_act() in which Curl_pp_readresp() is
called until the cache is drained. Without this multiple responses
received in a single packet could result in a hang or delay.

Similar to the processing of untagged CAPABILITY responses in IMAP moved
the processing of multiline EHLO responses to smtp_state_ehlo_resp() and
introduced an internal response code of one to differentiate a multiline
continuation from the end of command. This also allows for the separate
processing of multiline responses from commands such as VRFY and EXPN.

singleipconnect() did not return the open socket descriptor on some
errors, thereby sometimes causing a socket leak. This patch ensures
the socket is always returned.

Even though this is only a formality (since not many people build on
Mavericks while targeting Leopard), since we still support Leopard
at the earliest, we might as well be pedantic.

It turns out that some of the constants necessary to make this feature
work are missing from Snow Leopard's Security framework even though
they are defined in the headers.

Bug: http://curl.haxx.se/mail/lib-2013-11/0076.html
Reported by: myriachan

Rather than set the authentication options as part of the login details
specified in the URL, or via the older CURLOPT_USERPWD option, added a
new libcurl option to allow the login options to be set separately.

This patch fixes and issue introduced in commit 7d7df831, if the
tunnel state was TUNNEL_CONNECT, waitconnect_getsock() would return a
bitmask indicating a readable socket but never stored the socket in the
return array.

The temporary sockets used for Happy Eyeballs were not closed properly,
if curl exited prematurely, which this patch fixes.