It's not strictly clear if the API contract allows us to call strstr()
on a string that isn't zero terminated even when we know it will find
the substring, and clang's ASAN check dislikes us for it.
Also added a check of the return code in case it fails, even if I can't
think of a situation how that can trigger.
Detected by OSS-Fuzz
Closes#2513
Bug: https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=7760
Curl_cert_hostcheck operates with the host character set, therefore the
ASCII subjectAltName string retrieved with OpenSSL must be converted to
the host encoding before comparison.
Closes#2493
This triggered an assert if called more than once in debug mode (and a
memory leak if not debug build). With the right sequence of HTTP/2
headers incoming it can happen.
Detected by OSS-Fuzz
Closes#2507
Bug: https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=7764
- Move verify_certificate functionality in schannel.c into a new
file called schannel_verify.c. Additionally, some structure defintions
from schannel.c have been moved to schannel.h to allow them to be
used in schannel_verify.c.
- Make verify_certificate functionality for Schannel available on
all versions of Windows instead of just Windows CE. verify_certificate
will be invoked on Windows CE or when the user specifies
CURLOPT_CAINFO and CURLOPT_SSL_VERIFYPEER.
- In verify_certificate, create a custom certificate chain engine that
exclusively trusts the certificate store backed by the CURLOPT_CAINFO
file.
- doc updates of --cacert/CAINFO support for schannel
- Use CERT_NAME_SEARCH_ALL_NAMES_FLAG when invoking CertGetNameString
when available. This implements a TODO in schannel.c to improve
handling of multiple SANs in a certificate. In particular, all SANs
will now be searched instead of just the first name.
- Update tool_operate.c to not search for the curl-ca-bundle.crt file
when using Schannel to maintain backward compatibility. Previously,
any curl-ca-bundle.crt file found in that search would have been
ignored by Schannel. But, with CAINFO support, the file found by
that search would have been used as the certificate store and
could cause issues for any users that have curl-ca-bundle.crt in
the search path.
- Update url.c to not set the build time CURL_CA_BUNDLE if the selected
SSL backend is Schannel. We allow setting CA location for schannel
only when explicitly specified by the user via CURLOPT_CAINFO /
--cacert.
- Add new test cases 3000 and 3001. These test cases check that the first
and last SAN, respectively, matches the connection hostname. New test
certificates have been added for these cases. For 3000, the certificate
prefix is Server-localhost-firstSAN and for 3001, the certificate
prefix is Server-localhost-secondSAN.
- Remove TODO 15.2 (Add support for custom server certificate
validation), this commit addresses it.
Closes https://github.com/curl/curl/pull/1325
- Fix warning 'integer from pointer without a cast' on 3rd arg in
CertOpenStore. The arg type HCRYPTPROV may be a pointer or integer
type of the same size.
Follow-up to e35b025.
Caught by Marc's CI builds.
Users can now specify a client certificate in system certificates store
explicitly using expression like `--cert "CurrentUser\MY\<thumbprint>"`
Closes#2376
If you pass empty user/pass asking curl to use Windows Credential
Storage (as stated in the docs) and it has valid credentials for the
domain, e.g.
curl -v -u : --ntlm example.com
currently authentication fails.
This change fixes it by providing proper SPN string to the SSPI API
calls.
Fixes https://github.com/curl/curl/issues/1622
Closes https://github.com/curl/curl/pull/1660
The ifdefs have become quite long. Also, the condition for the
definition of CURLOPT_SERVICE_NAME and for setting it from
CURLOPT_SERVICE_NAME have diverged. We will soon also need the two
options for NTLM, at least when using SSPI, for
https://github.com/curl/curl/pull/1660.
Just make the definitions unconditional to make that easier.
Closes https://github.com/curl/curl/pull/2479
When a zeroed out allocation is required, use calloc() rather than
malloc() followed by an explicit memset(). The result will be the
same, but using calloc() everywhere increases consistency in the
codebase and avoids the risk of subtle bugs when code is injected
between malloc and memset by accident.
Closes https://github.com/curl/curl/pull/2497
In debug mode, MingGW-w64's GCC 7.3 issues null-dereference warnings
when dereferencing pointers after DEBUGASSERT-ing that they are not
NULL.
Fix this by removing the DEBUGASSERTs.
Suggested-by: Daniel Stenberg
Ref: https://github.com/curl/curl/pull/2463
unit1309 and vtls/gtls: error: arithmetic on a null pointer treated as a
cast from integer to pointer is a GNU extension
Reported-by: Rikard Falkeborn
Fixes#2466Closes#2468
In the situation of a client connecting to an FTP server using an IPv6
tunnel proxy, the connection info will indicate that the connection is
IPv6. However, because the server behing the proxy is IPv4, it is
permissable to attempt PSV mode. In the case of the FTP server being
IPv4 only, EPSV will always fail, and with the current logic curl will
be unable to connect to the server, as the IPv6 fwdproxy causes curl to
think that EPSV is impossible.
Closes#2432
curl 7.57.0 and up interpret this according to Appendix E.3.2 of RFC
8089 but then returns an error saying this is unimplemented. This is
actually a regression in behavior on both Windows and Unix.
Before curl 7.57.0 this URL was treated as a path of "//foo/bar" and
then passed to the relevant OS API. This means that the behavior of this
case is actually OS dependent.
The Unix path resolution rules say that the OS must handle swallowing
the extra "/" and so this path is the same as "/foo/bar"
The Windows path resolution rules say that this is a UNC path and
automatically handles the SMB access for the program. So curl on Windows
was already doing Appendix E.3.2 without any special code in curl.
Regression
Closes#2438
This reverts commit dc85437736.
libcurl (with the OpenSSL backend) performs server certificate verification
even if verifypeer == 0 and the verification result is available using
CURLINFO_SSL_VERIFYRESULT. The commit that is being reverted caused the
CURLINFO_SSL_VERIFYRESULT to not have useful information for the
verifypeer == 0 use case (it would always have
X509_V_ERR_UNABLE_TO_GET_ISSUER_CERT_LOCALLY).
Closes#2451
This fixes a segfault occurring when a name of the (invalid) form "domain..tld"
is processed.
test46 updated to cover this case.
Follow-up to commit c990ead.
Ref: https://github.com/curl/curl/pull/2440
In order to make curl_multi_timeout() return suitable "sleep" times even
when there's no socket to wait for while the name is being resolved in a
helper thread.
It will increases the timeouts as time passes.
Closes#2419
If a connection has received a GOAWAY frame while not being used, the
function now reads frames off the connection before trying to reuse it
to avoid reusing connections the server has told us not to use.
Reported-by: Alex Baines
Fixes#1967Closes#2402
Currently CMake cannot detect Brotli support. This adds detection of the
libraries and associated header files. It also adds this to the
generated config.
Closes#2392
This patch adds CURLOPT_DNS_SHUFFLE_ADDRESSES to explicitly request
shuffling of IP addresses returned for a hostname when there is more
than one. This is useful when the application knows that a round robin
approach is appropriate and is willing to accept the consequences of
potentially discarding some preference order returned by the system's
implementation.
Closes#1694
When a transfer is requested to get done and it is put in the pending
queue when limited by number of connections, total or per-host, libcurl
would previously very aggressively retry *ALL* pending transfers to get
them transferring. That was very time consuming.
By reducing the aggressiveness in how pending are being retried, we
waste MUCH less time on putting transfers back into pending again.
Some test cases got a factor 30(!) speed improvement with this change.
Reported-by: Cyril B
Fixes#2369Closes#2383
Especially unpausing a transfer might have to move the socket back to the
"currently used sockets" hash to get monitored. Otherwise it would never get
any more data and get stuck. Easily triggered with pausing using the
multi_socket API.
Reported-by: Philip Prindeville
Bug: https://curl.haxx.se/mail/lib-2018-03/0048.htmlFixes#2393Closes#2391
Due to very frequent updates of the rate limit "window", it could
attempt to rate limit within the same milliseconds and that then made
the calculations wrong, leading to it not behaving correctly on very
fast transfers.
This new logic updates the rate limit "window" to be no shorter than the
last three seconds and only updating the timestamps for this when
switching between the states TOOFAST/PERFORM.
Reported-by: 刘佩东
Fixes#2386Closes#2388
Refuse to operate when given path components featuring byte values lower
than 32.
Previously, inserting a %00 sequence early in the directory part when
using the 'singlecwd' ftp method could make curl write a zero byte
outside of the allocated buffer.
Test case 340 verifies.
CVE-2018-1000120
Reported-by: Duy Phan Thanh
Bug: https://curl.haxx.se/docs/adv_2018-9cd6.html
gss_seal/gss_unseal have been deprecated in favor of
gss_wrap/gss_unwrap with GSS-API v2 from January 1997 [1]. The first
version of "The Kerberos Version 5 GSS-API Mechanism" [2] from June
1996 already says "GSS_Wrap() (formerly GSS_Seal())" and
"GSS_Unwrap() (formerly GSS_Unseal())".
Use the nondeprecated functions to avoid deprecation warnings.
[1] https://tools.ietf.org/html/rfc2078
[2] https://tools.ietf.org/html/rfc1964
Closes https://github.com/curl/curl/pull/2356
On MinGW and Cygwin, GCC and clang have been complaining about macro
redefinitions since 4272a0b0fc. Fix this
by undefining the macros before redefining them as suggested in
https://github.com/curl/curl/pull/2269.
Suggested-by: Daniel Stenberg
When targeting x64, MinGW-w64 complains about conversions between
32-bit long and 64-bit pointers. Fix this by reusing the
GNUTLS_POINTER_TO_SOCKET_CAST / GNUTLS_SOCKET_TO_POINTER_CAST logic
from gtls.c, moving it to warnless.h as CURLX_POINTER_TO_INTEGER_CAST /
CURLX_INTEGER_TO_POINTER_CAST.
Closes https://github.com/curl/curl/pull/2341
- Add new option CURLOPT_RESOLVER_START_FUNCTION to set a callback that
will be called every time before a new resolve request is started
(ie before a host is resolved) with a pointer to backend-specific
resolver data. Currently this is only useful for ares.
- Add new option CURLOPT_RESOLVER_START_DATA to set a user pointer to
pass to the resolver start callback.
Closes https://github.com/curl/curl/pull/2311
- In keeping with the naming of our other connect timeout options rename
CURLOPT_HAPPY_EYEBALLS_TIMEOUT to CURLOPT_HAPPY_EYEBALLS_TIMEOUT_MS.
This change adds the _MS suffix since the option expects milliseconds.
This is more intuitive for our users since other connect timeout options
that expect milliseconds use _MS such as CURLOPT_TIMEOUT_MS,
CURLOPT_CONNECTTIMEOUT_MS, CURLOPT_ACCEPTTIMEOUT_MS.
The tool option already uses an -ms suffix, --happy-eyeballs-timeout-ms.
Follow-up to 2427d94 which added the lib and tool option yesterday.
Ref: https://github.com/curl/curl/pull/2260
- Add new option CURLOPT_HAPPY_EYEBALLS_TIMEOUT to set libcurl's happy
eyeball timeout value.
- Add new optval macro CURL_HET_DEFAULT to represent the default happy
eyeballs timeout value (currently 200 ms).
- Add new tool option --happy-eyeballs-timeout-ms to expose
CURLOPT_HAPPY_EYEBALLS_TIMEOUT. The -ms suffix is used because the
other -timeout options in the tool expect seconds not milliseconds.
Closes https://github.com/curl/curl/pull/2260
This enables users to preresolve but still take advantage of happy
eyeballs and trying multiple addresses if some are not connecting.
Ref: https://github.com/curl/curl/pull/2260
Previously, it would only check for max length if the existing alloc
buffer was to small to fit it, which often would make the header still
get used.
Reported-by: Guido Berhoerster
Bug: https://curl.haxx.se/mail/lib-2018-02/0056.htmlCloses#2315
The list of state names (used in debug builds) was out of sync in
relation to the list of states (used in all builds).
I now added an assert to make sure the sizes of the two lists match, to
aid in detecting this mistake better in the future.
Regression since c92d2e14cf, shipped in 7.58.0.
Reported-by: Somnath Kundu
Fixes#2312Closes#2313
RFC 5321 4.1.1.4 specifies the CRLF terminating the DATA command
should be taken into account when chasing the <CRLF>.<CRLF> end marker.
Thus a leading dot character in data is also subject to escaping.
Tests 911 and test server are adapted to this situation.
New tests 951 and 952 check proper handling of initial dot in data.
Closes#2304
Some servers return a "content-encoding" header with a non-standard
"none" value.
Add "none" as an alias to "identity" as a work-around, to avoid
unrecognised content encoding type errors.
Signed-off-by: Mohammad AlSaleh <CE.Mohammad.AlSaleh@gmail.com>
Closes https://github.com/curl/curl/pull/2298
Windows 10.0.17061 SDK introduces support for Unix Domain Sockets.
Added the necessary include file to curl_addrinfo.c.
Note: The SDK (which is considered beta) has to be installed, VS 2017
project file has to be re-targeted for Windows 10.0.17061 and #define
enabled in config-win32.h.
When peer verification is disabled, calling
SSL_CTX_load_verify_locations is not necessary. Only call it when
verification is enabled to save resources and increase performance.
Closes#2290
Reduce code duplication by making Curl_mime_contenttype available and
used by the formdata function. This also makes the formdata function
recognize a set of more file extensions by default.
PR #2280 brought this to my attention.
Closes#2282
Whenever an expected pattern syntax rule cannot be matched, the
character starting the rule loses its special meaning and the parsing
is resumed:
- backslash at the end of pattern string matches itself.
- Error in [:keyword:] results in set containing :\[dekorwy.
Unit test 1307 updated for this new situation.
Closes#2273
... since the libc provided one are locale dependent in a way we don't
want. Also, the "native" isalnum() (for example) works differently on
different platforms which caused test 1307 failures on macos only.
Closes#2269
Make curl_getdate() handle dates before 1970 as well (returning negative
values).
Make test 517 test dates for 64 bit time_t.
This fixes bug (3) mentioned in #2238Closes#2250
... unless CURLOPT_UNRESTRICTED_AUTH is set to allow them. This matches how
curl already handles Authorization headers created internally.
Note: this changes behavior slightly, for the sake of reducing mistakes.
Added test 317 and 318 to verify.
Reported-by: Craig de Stigter
Bug: https://curl.haxx.se/docs/adv_2018-b3bf.html
In case an identity didn't match[0], the state machine would fail in
state SSH_AUTH_AGENT instead of progressing to the next identity in
ssh-agent. As a result, ssh-agent authentication only worked if the
identity required happened to be the first added to ssh-agent.
This was introduced as part of commit c4eb10e2f0, which
stated that the "else" statement was required to prevent getting stuck
in state SSH_AUTH_AGENT. Given the state machine's logic and libssh2's
interface I couldn't see how this could happen or reproduce it and I
also couldn't find a more detailed description of the problem which
would explain a test case to reproduce the problem this was supposed to
fix.
[0] libssh2_agent_userauth returning LIBSSH2_ERROR_AUTHENTICATION_FAILED
Closes#2248
Follow-up to 84fcaa2e7. libressl does not have the API even if it says it is
late OpenSSL version...
Fixes#2246Closes#2247
Reported-by: jungle-boogie on github
1. don't use "ULL" suffix since unsupported in older MSVC
2. use curl_off_t instead of custom long long ifdefs
3. make get_posix_time() not do unaligned data access
Fixes#2211Closes#2240
Reported-by: Chester Liu
A mime tree attached to an easy handle using CURLOPT_MIMEPOST is
strongly bound to the handle: there is a pointer to the easy handle in
each item of the mime tree and following the parent pointer list
of mime items ends in a dummy part stored within the handle.
Because of this binding, a mime tree cannot be shared between different
easy handles, thus it needs to be cloned upon easy handle duplication.
There is no way for the caller to get the duplicated mime tree
handle: it is then set to be automatically destroyed upon freeing the
new easy handle.
New test 654 checks proper mime structure duplication/release.
Add a warning note in curl_mime_data_cb() documentation about sharing
user data between duplicated handles.
Closes#2235
Prior to this change the stored byte count of each trailer was
miscalculated and 1 less than required. It appears any trailer
after the first that was passed to Curl_client_write would be truncated
or corrupted as well as the size. Potentially the size of some
subsequent trailer could be erroneously extracted from the contents of
that trailer, and since that size is used by client write an
out-of-bounds read could occur and cause a crash or be otherwise
processed by client write.
The bug appears to have been born in 0761a51 (precedes 7.49.0).
Closes https://github.com/curl/curl/pull/2231
Decoding loop implementation did not concern the case when all
received data is consumed by Brotli decoder and the size of decoded
data internally hold by Brotli decoder is greater than CURL_MAX_WRITE_SIZE.
For content with unencoded length greater than CURL_MAX_WRITE_SIZE this
can result in the loss of data at the end of content.
Closes#2194
Move curl_mime_initpart() and curl_mime_cleanpart() calls to lower-level
functions dealing with UserDefined structure contents.
This avoids memory leakages on curl-generated part mime headers.
New test 2073 checks this using the cli tool --next option: it
triggers a valgrind error if bug is present.
Bug: https://curl.haxx.se/mail/lib-2017-12/0060.html
Reported-by: Martin Galvan
- When zlib version is < 1.2.0.4, process gzip trailer before considering
extra data as an error.
- Inflate with Z_BLOCK instead of Z_SYNC_FLUSH to maximize correct data
and minimize corrupt data output.
- Do not try to restart deflate decompression in raw mode if output has
started or if the leading data is not available anymore.
- New test 232 checks inflating raw-deflated content.
Closes#2068
scan-build would warn on a potential access of an uninitialized
buffer. I deem it a false positive and had to add this somewhat ugly
work-around to silence it.
Fixed undefined symbol of getenv() which does not exist when compiling
for Windows 10 App (CURL_WINDOWS_APP). Replaced getenv() with
curl_getenv() which is aware of getenv() absence when CURL_WINDOWS_APP
is defined.
Closes#2171
Prune the DNS cache immediately after the dns entry is unlocked in
multi_done. Timed out entries will then get discarded in a more orderly
fashion.
Test506 is updated
Reported-by: Oleg Pudeyev
Fixes#2169Closes#2170
Prior to this change SSLKEYLOGFILE used line buffering on WIN32 just
like it does for other platforms. However, the Windows CRT does not
actually support line buffering (_IOLBF) and will use full buffering
(_IOFBF) instead. We can't use full buffering because multiple processes
may be writing to the file and that could lead to corruption, and since
full buffering is the only buffering available this commit disables
buffering for Windows SSLKEYLOGFILE entirely (_IONBF).
Ref: https://github.com/curl/curl/pull/1346#issuecomment-350530901
These are OS/2-specific things added to the code in the year 2000. They
were always ugly. If there's any user left, they still don't need it
done this way.
Closes#2166
- Allow proxy_ssl to be checked for pending data even when connssl does
not yet have an SSL handle.
This change is for posterity. Currently there doesn't seem to be a code
path that will cause a pending data check when proxyssl could have
pending data and the connssl handle doesn't yet exist [1].
[1]: Recall that an https proxy connection starts out in connssl but if
the destination is also https then the proxy SSL backend data is moved
from connssl to proxyssl, which means connssl handle is temporarily
empty until an SSL handle for the destination can be created.
Ref: https://github.com/curl/curl/commit/f4a6238#commitcomment-24396542
Closes https://github.com/curl/curl/pull/1916
Connections that are used for HTTP/1.1 Pipelining or HTTP/2 multiplexing
only get additional transfers added to them if the existing connection
is held by the same multi or easy handle. libcurl does not support doing
HTTP/2 streams in different threads using a shared connection.
Closes#2152
If the lock is released before the dealings with the bundle is over, it may
have changed by another thread in the mean time.
Fixes#2132Fixes#2151Closes#2139
For pop3/imap/smtp, added test 891 to somewhat verify the pop3
case.
For this, I enhanced the pingpong test server to be able to send back
responses with LF-only instead of always using CRLF.
Closes#2150
Figured out while reviewing code in the libssh backend. The pointer was
checked for NULL after having been dereferenced, so we know it would
always equal true or it would've crashed.
Pointed-out-by: Nikos Mavrogiannopoulos
Bug #2143Closes#2148
The previous code was incorrectly following the libssh2 error detection
for libssh2_sftp_statvfs, which is not correct for libssh's sftp_statvfs.
Fixes#2142
Signed-off-by: Nikos Mavrogiannopoulos <nmav@gnutls.org>
The SFTP back-end supports asynchronous reading only, limited
to 32-bit file length. Writing is synchronous with no other
limitations.
This also brings keyboard-interactive authentication.
Signed-off-by: Nikos Mavrogiannopoulos <nmav@gnutls.org>
That also updates tests to expect the right error code
libssh2 back-end returns CURLE_SSH error if the remote file
is not found. Expect instead CURLE_REMOTE_FILE_NOT_FOUND
which is sent by the libssh backend.
Signed-off-by: Nikos Mavrogiannopoulos <nmav@redhat.com>
libssh is an alternative library to libssh2.
https://www.libssh.org/
That patch set also introduces support for ECDSA
ed25519 keys, as well as gssapi authentication.
Signed-off-by: Nikos Mavrogiannopoulos <nmav@redhat.com>
Absent any 'symbol map' or script to limit what gets exported, static
linking of libraries previously resulted in a libcurl with curl's and
those other symbols being (re-)exported.
This did not happen if 'versioned symbols' were enabled (which is not
the default) because then a version script is employed.
This limits exports to everything starting in 'curl_*'., which is
what "libcurl.vers" exports.
This avoids strange side-effects such as with mixing methods
from system libraries and those erroneously offered by libcurl.
Closes#2127
Originally, my idea was to allocate the two structures (or more
precisely, the connectdata structure and the four SSL backend-specific
strucutres required for ssl[0..1] and proxy_ssl[0..1]) in one go, so
that they all could be free()d together.
However, getting the alignment right is tricky. Too tricky.
So let's just bite the bullet and allocate the SSL backend-specific
data separately.
As a consequence, we now have to be very careful to release the memory
allocated for the SSL backend-specific data whenever we release any
connectdata.
Signed-off-by: Johannes Schindelin <johannes.schindelin@gmx.de>
Closes#2119
commit d3ab7c5a21 broke the boringssl build since it doesn't have
RSA_flags(), so we disable that code block for boringssl builds.
Reported-by: W. Mark Kubacki
Fixes#2117
This bit is no longer used. It is not clear what it meant for users to
"init the TLS" in a world with different TLS backends and since the
introduction of multissl, libcurl didn't properly work if inited without
this bit set.
Not a single user responded to the call for users of it:
https://curl.haxx.se/mail/lib-2017-11/0072.html
Reported-by: Evgeny Grin
Assisted-by: Jay Satiro
Fixes#2089Fixes#2083Closes#2107
- Align the array of ssl_backend_data on a max 32 byte boundary.
8 is likely to be ok but I went with 32 for posterity should one of
the ssl_backend_data structs change to contain a larger sized variable
in the future.
Prior to this change (since dev 70f1db3, release 7.56) the connectdata
structure was undersized by 4 bytes in 32-bit builds with ssl enabled
because long long * was mistakenly used for alignment instead of
long long, with the intention being an 8 byte boundary. Also long long
may not be an available type.
The undersized connectdata could lead to oob read/write past the end in
what was expected to be the last 4 bytes of the connection's secondary
socket https proxy ssl_backend_data struct (the secondary socket in a
connection is used by ftp, others?).
Closes https://github.com/curl/curl/issues/2093
CVE-2017-8818
Bug: https://curl.haxx.se/docs/adv_2017-af0a.html
With this check present, scan-build warns that we might dereference this
point in other places where it isn't first checked for NULL. Thus, if it
*can* be NULL we have a problem on a few places. However, this pointer
should not be possible to be NULL here so I remove the check and thus
also three different scan-build warnings.
Closes#2111
* LOTS of comment updates
* explicit error for SMB shares (e.g. "file:////share/path/file")
* more strict handling of authority (i.e. "//localhost/")
* now accepts dodgy old "C:|" drive letters
* more precise handling of drive letters in and out of Windows
(especially recognising both "file:c:/" and "file:/c:/")
Closes#2110
The new API added in Linux 4.11 only requires setting a socket option
before connecting, without the whole sento() machinery.
Notably, this makes it possible to use TFO with SSL connections on Linux
as well, without the need to mess around with OpenSSL (or whatever other
SSL library) internals.
Closes#2056
Host names like "127.0.0.1 moo" would otherwise be accepted by some
getaddrinfo() implementations.
Updated test 1034 and 1035 accordingly.
Fixes#2073Closes#2092
... so that IPv6 addresses can be passed like they can for connect-to
and how they're used in URLs.
Added test 1324 to verify
Reported-by: Alex Malinovich
Fixes#2087Closes#2091
There is a conflict on symbol 'free_func' between openssl/crypto.h and
zlib.h on AIX. This is an attempt to resolve it.
Bug: https://curl.haxx.se/mail/lib-2017-11/0032.html
Reported-By: Michael Felt
