Re: [RFC PATCH] exec: flush the whole TLB if a watchpoint crosses a page boundary

[Richard Henderson]

On 6/2/20 9:49 AM, Alex Bennée wrote:
> There is no particular reason why you can't have a watchpoint in TCG
> that covers a large chunk of the address space. We could be clever
> about it but these cases are pretty rare and we can assume the user
> will expect a little performance degradation.
> 
> NB: In my testing gdb will silently squash a watchpoint like:
> 
>   watch (char[0x7fffffffff]) *0x0
> 
> to a 4 byte watchpoint. Practically it will limit the maximum size
> based on max-value-size. However given enough of a tweak the sky is
> the limit.
> 
> Reported-by: Alexander Bulekov <alxndr@bu.edu>
> Signed-off-by: Alex Bennée <alex.bennee@linaro.org>
> ---
>  exec.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/exec.c b/exec.c
> index 5162f0d12f9..851ac180fe7 100644
> --- a/exec.c
> +++ b/exec.c
> @@ -1056,7 +1056,11 @@ int cpu_watchpoint_insert(CPUState *cpu, vaddr addr, 
> vaddr len,
>          QTAILQ_INSERT_TAIL(&cpu->watchpoints, wp, entry);
>      }
>  
> -    tlb_flush_page(cpu, addr);
> +    if (((addr + len) & TARGET_PAGE_MASK) != (addr & TARGET_PAGE_MASK)) {

This computation will be false when adding a watchpoint for the last len bytes
in the page, and no actual page crossing is required.  For this reason I prefer

    in_page = -(addr | TARGET_PAGE_MASK);
    if (len <= in_page)


r~

> +        tlb_flush(cpu);
> +    } else {
> +        tlb_flush_page(cpu, addr);
> +    }
>  
>      if (watchpoint)
>          *watchpoint = wp;
>