[PATCH] arm64: signal: Update sigcontext reservations table

[Dave Martin]

On Tue, Aug 20, 2024 at 02:37:40PM +0100, Will Deacon wrote:
> On Mon, Jul 29, 2024 at 03:41:49PM +0100, Dave Martin wrote:
> > The table tracking space usage in sigcontext.__reserved[] has got
> > a bit out of date.
> > 
> > Update it, and clarify the opt-in constraints.
> > 
> > Note, svl <= 64 would be a sufficient condition for keeping the
> > sve_context within range when in streaming SVE mode under SME, but
> > then za_context gets too big and userspace loses anyway.  To keep
> > the conditions simple, just write "svl <= 32" everywhere.
> > 
> > No functional change.
> > 
> > Signed-off-by: Dave Martin <Dave.Martin at arm.com>
> > 
> > ---
> > 
> > Note, this is a back-of-the-envelope calculation... but whatever way I
> > slice it, __reserved[] looks pretty much full (!)
> > 
> > If Mark in particular can double-check the SME impact, that would be
> > appreciated.
> > 
> > New arch features with a non-trivial amount of state that needs to be
> > saved may need to be disabled by default and require explicitly turning
> > on by a syscall unless we want to allow some ABI breakage (x86's
> > experience suggests that the world takes a long time to explode when
> > signal frames outgrow their official size, though).
> > 
> > Either way, do we need a new strategy to slow down the filling of the
> > remaining space?  There is continuing demand on it (see e.g., [1]).
> > 
> > Migration note: at least glibc since version 2.34 [2] has stopped
> > offering compile-time constant signal stack size #defines to programs
> > built with -D_GNU_SOURCE. [3]  This should mitigate ABI breaks for
> > programs that bother to size stacks correctly.
> > 
> > I haven't checked what other libcs and runtimes are doing.
> > 
> > Additional SME note: since userspace can freely switch in and out of
> > streaming SVE mode and freely enable/disable ZA and ZT0, I'm assuming
> > that none of sve_context, za_context and zt_context are mutually
> > exclusive, but please shout if I have confused myself here.
> > 
> > [1] [PATCH v4 18/29] arm64: add POE signal support
> > https://lore.kernel.org/linux-arm-kernel/20240503130147.1154804-19-joey.gouly@arm.com/
> > 
> > [2] [glibc] glibc-2.34
> > https://sourceware.org/git/?p=glibc.git;a=tag;h=9df03063320651bc629fa427eef3ac73fabb61ba
> > 
> > [3] [glibc] sysconf: Add _SC_MINSIGSTKSZ/_SC_SIGSTKSZ [BZ #20305]
> > https://sourceware.org/git/?p=glibc.git;a=commit;f=sysdeps/unix/sysv/linux/bits/sigstksz.h;h=6c57d320484988e87e446e2e60ce42816bf51d53
> > ---
> >  arch/arm64/include/uapi/asm/sigcontext.h | 11 +++++++++--
> >  1 file changed, 9 insertions(+), 2 deletions(-)
> > 
> > diff --git a/arch/arm64/include/uapi/asm/sigcontext.h b/arch/arm64/include/uapi/asm/sigcontext.h
> > index 8a45b7a411e0..2cd60fd64e9a 100644
> > --- a/arch/arm64/include/uapi/asm/sigcontext.h
> > +++ b/arch/arm64/include/uapi/asm/sigcontext.h
> > @@ -44,11 +44,18 @@ struct sigcontext {
> >   *
> >   *	0x210		fpsimd_context
> >   *	 0x10		esr_context
> > - *	0x8a0		sve_context (vl <= 64) (optional)
> > + *	0x8a0		sve_context (vl <= 64 && svl <= 32) (optional)
> > + *	 0x10		tpidr2_context (optional)
> > + *	0x410		za_context (svl <= 32) (optional)
> > + *	 0x50		zt_context (optional)
> > + *	 0x10		fpmr_context (optional)
> >   *	 0x20		extra_context (optional)
> >   *	 0x10		terminator (null _aarch64_ctx)
> >   *
> > - *	0x510		(reserved for future allocation)
> > + *	 0x90		(reserved for future allocation)
> > + *
> > + * where vl is the non-streaming SVE vector length, as set with PR_SVE_SET_VL,
> > + * and svl is the streaming SVE vector length, as set with PR_SME_SET_VL.
> 
> These are quite fiddly to check by hand, but the ones I *did* check
> appear to be correct. The discussion with Mark seems tangential to the
> actual diff, so I'm inclined to apply this if nobody objects.
> 
> Will

I'm wondering whether we should just get rid of this instead, since
it's not that maintainable: see [4].

The discussion re POR_EL0 has moved in the direction of probably
dumping this register unconditionally. [5]  Ideally we wouldn't, but
there's no way to know in advance whether POR_EL0 counts as "in use",
or to anticipate whether a signal handler might want to alter it.

Trapping POR_EL0 until something writes it or a pkey_foo() syscall is
called is an option, but this feels like overkill.

We could also come up with a way for a signal handler to hang extra
stuff off the sigframe to be restored by sigreturn even when there's no
space to allocate another record, but again, the complexity does not
feel well justified until somebody shouts about it.

I can propose something if you think it's worth discussing.

Cheers
---Dave

[4] https://lore.kernel.org/linux-arm-kernel/Zr4aJqc%2FifRXJQAd@e133380.arm.com/

[5] https://lore.kernel.org/linux-arm-kernel/ZsSgKl2JINjdpuW1@e133380.arm.com/