On Mon, Aug 05, 2019 at 07:27:37PM -0400, Rich Felker wrote:
> I've discovered a flaw in musl libc's arch-specific math assembly code
> for i386, whereby at least the log1p function and possibly others
> return with more than one item on the x87 stack. This can lead to x87
> stack overflow in the execution of subsequent math code, causing it to
> incorrectly produce a NAN in place of the actual result. If floating
> point results are used in flow control, this can lead to runaway wrong
> code execution. For example, in Python (version 3.6.8 tested), at
> least one code path of the dtoa function becomes an infinite loop
> performing what's effectively an unbounded-length memset when entered
> under such a condition.
> 
> This bug is potentially exploitable in software which calls affected
> math functions with inputs under user control. Impact depends on how
> the application handles the ABI-violating x87 state; in Python it
> seems to be limited to producing a crash.
> 
> The bug is present in all versions after 0.9.12, up through the
> current (1.1.23) release. Only 32-bit x86 systems (aka IA32, musl's
> "i386" arch) are affected. Users of other archs, including x86_64, can
> safely ignore this issue.
> 
> Affected users are advised to apply the following patch:
> 
> https://git.musl-libc.org/cgit/musl/patch/?id=f3ed8bfe8a82af1870ddc8696ed4cc1d5aa6b441

The patch contains an error that was missed for unknown reasons,
probably failure to rebuild a file. I'm attaching an aggregate patch
that works. Alternaatively, these two commits can be applied:

https://git.musl-libc.org/cgit/musl/patch/?id=f3ed8bfe8a82af1870ddc8696ed4cc1d5aa6b441
https://git.musl-libc.org/cgit/musl/patch/?id=6818c31c9bc4bbad5357f1de14bedf781e5b349e