- according to man.openbsd.org getentropy() is in unistd.h
- according to macOS sierra's man it's in sys/random.h
- since sys/random.h is does not exist for iOS and for linux, do
not attempt to include it, rather redeclare the prototype
- make sure that `./configure`:
- uses getentropy() on macOS sierra
- does not use getentropy() if compiling for 10.11
- does not use getentropy() if compiling for ios armv7
This diff changes the logic by which configure detects getentropy() to
ensure that we don't use the system wide getentropy
- with macOS sierra if the deployment target is lower than sierra as
found by tor developers here
https://gitweb.torproject.org/tor.git/commit/?id=https://gitweb.torproject.org/tor.git/commit/?id=16fcbd21c963a9a65bf55024680c8323c8b7175d
- with iOS unconditionally because an app linking libressl compiled with
system wide getentropy has been rejected by the App store as I have
documented here
https://github.com/measurement-kit/measurement-kit/pull/994
I think something similar could also affect clock_gettime judging from
tor's patch, but this diff for now doesn't address that.
I do not have macOS < sierra, so I could only verify that configure was
not picking up system wide getentropy by compiling libressl using
export CFLAGS="-mmacosx-version-min=10.11"
As regards iOS, removing the check for getentropy and recompiling (thus
using libressl builtin getentropy()) was enough to have another iteration
of the app accepted. Otherwise testing should be possible with:
export LDFLAGS=-arch armv7 -miphoneos-version-min=7.1 -isysroot `xcrun --show-sdk-path --sdk iphoneos`
export CPPFLAGS=-arch armv7 -isysroot `xcrun --show-sdk-path --sdk iphoneos`
export CFLAGS=-arch armv7 -miphoneos-version-min=7.1 -isysroot `xcrun --show-sdk-path --sdk iphoneos`
Related ticket: https://github.com/libressl-portable/portable/issues/230
Solaris 11 recently introduced a builtin arc4random in libc which fails the tests in "make check". Found USE_BUILTIN_ARC4RANDOM, but could not get it to work. Apparently, there is a typo in the configure logic rendering USE_BUILTIN_ARC4RANDOM ineffective.
This moves the compatibility include files from include to
include/compat so we can use the awful MS C compiler
<../include/> trick to emulate the GNU #include_next extension.
This also removes a few old compat files we do not need anymore.
Rather than doing separate linker/compiler checks, just build a
non-empty program with each so that the compiler will actually try to
use the hardening features. Reduce redundancy in the macro calls by just
setting the flag that was just tested.
Also, disable hardening for mingw, since its trying to use a
libssp-0.dll file that I can't find right now. The detected hardening
flags break mingw builds currently.
Where available, enable stack smashing protection, fortify source,
no-strict-overflow, and read only relocations.
Many Linux distributions automatically enable most of these options.
They are no brainers. The difference introduced here is in asking for a
few more aggressive options. An option to disable the more aggressive
options is provided (--disable-hardening). When set, configure will fall
back to the default CFLAGS on the system - in many cases that will still
be hardened. There is no point in going further than that.
Options enabled are:
-fstack-protector-strong is a relatively new GCC-4.9 feature that is
supposed to give a better balance between performance and protection.
-all is considered too aggressive, but was used in Chromium and other
security critical systems until -strong became available. Follow their
lead and use -strong when possible. clang 6.0 supports -all but not
-strong.
_FORTIFY_SOURCE replaces certain unsafe C str* and mem* functions with
more robust equivalents when the compiler can determine the length of
the buffers involved.
-fno-strict-overflow instructs GCC to not make optimizations based on
the assumption that signed arithmetic will wrap around on overflow (e.g.
(short)0x7FFF + 1 == 0). This prevents the optimizer from doing some
unexpected things. Further improvements should trap signed overflows and
reduce the use of signed to refer to naturally unsigned quantities.
I did not set -fPIE (position independent executables). The critical
function of Open/LibreSSL is as a library, not an executable.
Tested on Ubuntu Linux 14.04.1 LTS, OS X 10.10.1 with "make check".
Signed-off-by: Jim Barlow <jim@purplerock.ca>
