Re: [REGRESSION] x86/cpu fsgsbase breaks TLS in 32 bit rr tracees on a 64 bit system

From: Kyle Huey <me@kylehuey.com>
To: "Bae, Chang Seok" <chang.seok.bae@intel.com>
Cc: "Robert O'Callahan" <rocallahan@gmail.com>,
	Andy Lutomirski <luto@kernel.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Ingo Molnar <mingo@kernel.org>, "H . Peter Anvin" <hpa@zytor.com>,
	Andi Kleen <ak@linux.intel.com>,
	"Shankar, Ravi V" <ravi.v.shankar@intel.com>,
	LKML <linux-kernel@vger.kernel.org>,
	"Hansen, Dave" <dave.hansen@intel.com>
Subject: Re: [REGRESSION] x86/cpu fsgsbase breaks TLS in 32 bit rr tracees on a 64 bit system
Date: Fri, 21 Aug 2020 14:32:47 -0700	[thread overview]
Message-ID: <CAP045AqiT75B9o5OV+SJT8uTzo7A55Y1BxjNcaSuxtm5EoAxaQ@mail.gmail.com> (raw)
In-Reply-To: <29FD0DDA-3093-46A3-BCF4-85DEC229E30D@intel.com>

On Fri, Aug 21, 2020 at 1:08 PM Bae, Chang Seok
<chang.seok.bae@intel.com> wrote:
>
>
> > On Aug 20, 2020, at 21:41, Kyle Huey <me@kylehuey.com> wrote:
> >
> > On the x86-64 5.9-rc1 TLS is completely broken in 32 bit tracees when
> > running under rr[0]. Booting the kernel with `nofsgsbase` fixes it and
> > I bisected to https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?h=v5.8&id=673903495c85137791d5820d690229efe09c8f7b.
> >
> > STR:
> > 1. Build rr from source by
> >  a. git clone https://github.com/mozilla/rr
> >  b. mkdir rr/obj
> >  c. cd rr/obj
> >  d. cmake ..
> >  e. make -j16
> > 2. Run the simple 32 bit tracee outside of rr with `./bin/simple_32`.
> > It should print a message and exit cleanly.
> > 3. Run it under rr with `./bin/rr ./bin/simple_32`.
> >
> > It should behave the same way, but with fsgsbase enabled it will
> > segfault. The `simple_32` binary is a simple "hello world" type
> > program but it does link to pthreads, so pre-main code attempts to
> > access TLS variables.
> >
> > The interplay between 32 bit and 64 bit TLS is dark magic to me
> > unfortunately so this is all the useful information I have.
>
> As I run it and collect the ptrace logs, it starts to set FSBASE with
> some numbers, e.g. 140632147826496, and then later attempts to set GS
> with 99 through putreg(), not putreg32().
>
> With FSGSBASE, the FS/GS base is decoupled from the selector. Andy
> made putreg32() to retain the old behavior, fetching FS/GS base
> according to the index, but the putreg() does not do. So, rr probably
> relies on the old behavior as observed to setting the GS index only.
> But it was previously considered to be okay with this comment [1]:
>
>    "Our modifications to fs/gs/fs_base/gs_base are always either a)
>     setting values that the kernel set during recording to make them
>     happen during replay or b) emulating PTRACE_SET_REGS that a tracee
>     ptracer tried to set on another tracee. Either way I think the
>     effects are going to be the same as what would happen if the
>     program were run without rr."
>
> It is not straightforward to go into the rr internal to me. Robert,
> any thought?

Hmm. When we are running a 32 bit tracee in a 64 bit build of rr we
internally convert between the 32 bit and 64 bit user_regs_structs[0]
at the ptrace boundary. This does not preserve the fs/gsbase (because
there is no fs/gsbase in the 32 bit user_regs_struct, of course).

40c45904f818c1f6555294ca27afc5fda4f09e68 added magic for a 32 bit
tracer tracing a 32 bit tracee on a 64 bit kernel, but it looks like a
64 bit tracer tracing a 32 bit tracee on a 64 bit kernel *is* now
expected to preserve the fs/gsbase values (or die, in our case).

Is that correct?

- Kyle

[0] https://github.com/mozilla/rr/blob/fcd2a26680a3fc2bda5f40d732d0c72b9628357b/src/Registers.cc#L519