Message-ID: <20200220145241.GI3704@minyard.net>
Date:   Thu, 20 Feb 2020 08:52:41 -0600
From:   Corey Minyard <minyard@....org>
To:     Daniel Thompson <daniel.thompson@...aro.org>
Cc:     Will Deacon <will@...nel.org>,
        Catalin Marinas <catalin.marinas@....com>,
        linux-arm-kernel@...ts.infradead.org, linux-kernel@...r.kernel.org,
        Corey Minyard <cminyard@...sta.com>
Subject: Re: [PATCH v2] arm64:kgdb: Fix kernel single-stepping

On Thu, Feb 20, 2020 at 02:06:50PM +0000, Daniel Thompson wrote:
> On Wed, Feb 19, 2020 at 09:24:03AM -0600, minyard@....org wrote:
> > From: Corey Minyard <cminyard@...sta.com>
> > 
> > I was working on a single-step bug on kgdb on an ARM64 system, and I saw
> > this scenario:
> > 
> > * A single step is setup to return to el1
> > * The ERET return to el1
> > * An interrupt is pending and runs before the instruction
> > * As soon as PSTATE.D (the debug disable bit) is cleared, the single
> >     step happens in that location, not where it should have.
> > 
> > This appears to be due to PSTATE.SS not being cleared when the exception
> > happens.  Per section D.2.12.5 of the ARMv8 reference manual, that
> > appears to be incorrect, it says "As part of exception entry, the PE
> > does all of the following: ...  Sets PSTATE.SS to 0."
> > 
> > However, I appear to not be the first person who has noticed this.  In
> > the el0-only portion of the kernel_entry macro in entry.S, I found the
> > following comment: "Ensure MDSCR_EL1.SS is clear, since we can unmask
> > debug exceptions when scheduling."  Exactly the same scenario, except
> > coming from a userland single step, not a kernel one.
> > 
> > As I was studying this, though, I realized that the following scenario
> > had an issue:
> > 
> > * Kernel enables MDSCR.SS, MDSCR.KDE, MDSCR.MDE (unnecessary), and
> >   PSTATE.SS to enable a single step in el1, for kgdb or kprobes,
> >   on the current CPU's MDSCR register and the process' PSTATE.SS
> >   register.
> > * Kernel returns from the exception with ERET.
> > * An interrupt or page fault happens on the instruction, causing the
> >   instruction to not be run, but the exception handler runs.
> > * The exception causes the task to migrate to a new core.
> > * The return from the exception runs on a different processor now,
> >   where the MDSCR values are not set up for a single step.
> > * The single step fails to happen.
> > 
> > This is bad for kgdb, of course, but it seems really bad for kprobes if
> > this happens.
> > 
> > To fix both these problems, rework the handling of single steps to clear
> > things out upon entry to the kernel from el1, and then to set up single
> > step when returning to el1, and not do the setup in debug-monitors.c.
> > This means that single stepping does not use
> > enable/disable_debug_monitors(); it is no longer necessary to track
> > those flags for single stepping.  This is much like single stepping is
> > handled for el0.  A new flag is added in pt_regs to enable single
> > stepping from el1.  Unfortunately, the old value of PSTATE.SS cannot be
> > used for this because of the hardware bug mentioned earlier.
> > 
> > As part of this, there is an interaction between single stepping and the
> > other users of debug monitors with the MDSCR.KDE bit.  That bit has to
> > be set for both hardware breakpoints at el1 and single stepping at el1.
> > A new variable was created to store the cpu-wide value of MDSCR.KDE; the
> > single stepping code makes sure not to clear that bit on kernel entry if
> > it's set in the per-cpu variable.
> > 
> > After fixing this and doing some more testing, I ran into another issue:
> > 
> > * Kernel enables the pt_regs single step
> > * Kernel returns from the exception with ERET.
> > * An interrupt or page fault happens on the instruction, causing the
> >   instruction to not be run, but the exception handler runs.
> > * The exception handling hits a breakpoint and stops.
> > * The user continues from the breakpoint, so the kernel is no longer
> >   expecting a single step.
> > * On the return from the first exception, the single step flag in
> >   pt_regs is still set, so a single step trap happens.
> > * The kernel keels over from an unexpected single step.
> > 
> > There's no easy way to find the pt_regs that has the single step flag
> > set.  So a thread info flag was added so that the single step could be
> > disabled in this case.  Both that flag and the flag in pt_regs must be
> > set to enable a single step.
> > 
> > Signed-off-by: Corey Minyard <cminyard@...sta.com>
> 
> I've pointed the kgdbtest suite at this patch (and run one of the
> historically unstable test cases an extra 100 times just in case).
> 
> kgdbtest hasn't got great coverage, runs the code in qemu and some
> of the strongest tests are still marked XFAIL on arm64 (for reasons
> unrelated to stepping).
> 
> So the best I can say based on the above is that the test suite does not
> observe any regression (but equally no improvement). Nevertheless FWIW:

Thanks for testing this.  This is not a surprise, you would either have
to have a broken processor like the one I'm using, or you would have to
have a migration occur on the instruction being single-stepped, which
would be extremely unlikely.

Since I've already gained some experience here, I'll try to look at
fixing things here for ARM64.

-corey

> 
> 
> Tested-by: Daniel Thompson <daniel.thompson@...aro.org>
> 
> 
> Daniel.

Powered by blists - more mailing lists