Re: [PATCH 0/2] target/arm: Fix unwind from dc zva and FEAT_MOPS

[Ilya Leoshkevich]

On Tue, 2024-07-02 at 16:41 -0700, Richard Henderson wrote:
> While looking into Zoltan's attempt to speed up ppc64 DCBZ
> (data cache block set to zero), I wondered what AArch64 was
> doing differently.  It turned out that Arm is the only user
> of tlb_vaddr_to_host.
> 
> None of the code sequences in use between AArch64, Power64 and S390X
> are 100% safe, with race conditions vs mmap et al, however, AArch64
> is the only one that will fail this single threaded test case.  Use
> of these new functions fixes the race condition as well, though I
> have not yet touched the other guests.
> 
> I thought about exposing accel/tcg/user-retaddr.h for direct use
> from the targets, but perhaps these wrappers are cleaner.  RFC?
> 
> 
> r~
> 
> 
> Richard Henderson (2):
>   accel/tcg: Introduce memset_ra, memmove_ra
>   target/arm: Use memset_ra, memmove_ra in helper-a64.c
> 
>  include/exec/cpu_ldst.h            | 40 ++++++++++++++++
>  accel/tcg/user-exec.c              | 22 +++++++++
>  target/arm/tcg/helper-a64.c        | 10 ++--
>  tests/tcg/multiarch/memset-fault.c | 77
> ++++++++++++++++++++++++++++++
>  4 files changed, 144 insertions(+), 5 deletions(-)
>  create mode 100644 tests/tcg/multiarch/memset-fault.c

This sounds good to me.

I haven't debugged it, but I wonder why doesn't s390x fail here.
For XC with src == dst, it does access_memset() -> do_access_memset()
-> memset() without setting the RA. And I don't think that anything
around it sets the RA either.