high bit as required by posix. wouldn't want to break any standards.
idea and ok deraadt

makes it much harder.
ok bcook@ kettenis@

can copy this file (plus chacha_private.h) directly and reuse it
trivially.  Well, as long as they have a getentropy() as well..
ok beck

now using this as upstream code.  The particular problem is systems
that contain older arc4random derivations lacking arc4random_uniform().
ok tedu miod

__atexit tables and touches global variables.  From Srinavasa Nagaraju
through Android/Elliott Hughes.
ok tedu@, guenther@

ok deraadt

handlers. if this happens, restart the loop.
ok kettenis matthew millert miod

ok otto

new wording agreed by at least kettenis@ millert@ otto@

ok jsing@

ok beck@ deraadt@ jsing@ guenther@

portable code path must handle that; with brent cook

improve the random stream itself (it doesn't), but to introduce
noise in the arc4random calling pattern. Thanks to matthew@ who
pointed out bias in a previous diff, ok deraadt@ matthew@

mechanism, to aid in portability to other systems as requested.
ok matthew

artificially constrain alternative implementations. ok deraadt

Modern compiler toolchains are capable of optimizing even across
translation unit boundaries, so simply moving the memory clearing into
a separate function is not guaranteed to clear memory.
To avoid this, we take advantage of ELF weak symbol semantics, and
insert a call to an empty, weakly named function.  The semantics of
calling this function aren't determinable until load time, so the
compiler and linker need to keep the memset() call.
There are still ways a toolchain might defeat this trick (e.g.,
optimistically expecting the weak symbol to not be overloaded, and
only calling memset() if it is; promoting weak symbols to strong
symbols at link-time when emitting a static binary because they won't
be interposed; implementing load-time optimizations).  But at least
for the foreseeable future, these seem unlikely.
ok deraadt

Allow other non-zero return values in case we change our mind to
return an ssize_t byte count instead of simple success/fail.
ok deraadt, djm

MAP_INHERIT_ZERO anymore.  This restores arc4random's previous
behavior where fork children would mix in some randomness from the
parent process.
New behavior noticed by deraadt
ok deraadt, tedu

The extra argument doesn't hurt genuine atexit handlers and this fixes a
bug where we didn't provide the argument (effectively passing garbage) for
functions registered with __cxa_atexit in the main executable.
Pointed out by Dmitriy Ivanov <dimitry@google.com> and Elliott Hughes
<enh@google.com>.
ok matthew@

Now instead of calling getpid() each time a user invokes arc4random(),
we're able to rely on the kernel zero'ing out the RNG state if the
process forks.
ok deraadt, djm

discussion with matthew

of sysctl().  Mark it with XXX while we consider.

call abort().
this direction discussed at length with miod beck tedu matthew etc

ok deraadt, jmc, tedu

behavior for certain inputs.  From NetBSD.  OK tedu@

From Fritjof Bornebusch.

freed chunk is actually freeable immediately. catch more errors.
hints/ok otto

While there, sort headers.
ok tedu@

ok miod

required. try to document this fact and some of the history.
with feedback from deraadt guenther millert

while changing things, add a crypt_checkpass wrapper that handles most of
the edge cases. (not quite ready for production, though.)
ok deraadt

deterministic behavior. four selected because it's more than three, less
than five. i.e., no particular reason.

can avoid reinventing the wheel
ok guenther schwarze

ok crickets@