Re: [PATCH 1/2] exec: flush CPU TB cache in breakpoint_invalidate

[Alex Bennée]

Richard Henderson <address@hidden> writes:

> On 11/27/19 10:06 PM, Max Filippov wrote:
>> When a breakpoint is inserted at location for which there's currently no
>> virtual to physical translation no action is taken on CPU TB cache. If a
>> TB for that virtual address already exists but is not visible ATM the
>> breakpoint won't be hit next time an instruction at that address will be
>> executed.
>> 
>> Flush entire CPU TB cache in breakpoint_invalidate to force
>> re-translation of all TBs for the breakpoint address.
>> 
>> This change fixes the following scenario:
>> - linux user application is running
>> - a breakpoint is inserted from QEMU gdbstub for a user address that is
>>   not currently present in the target CPU TLB
>> - an instruction at that address is executed, but the external debugger
>>   doesn't get control.
>> 
>> Signed-off-by: Max Filippov <address@hidden>
>> ---
>> Changes RFC->v1:
>> - do tb_flush in breakpoint_invalidate unconditionally
>
> Unlike Paolo, I don't think this is a good idea.

We previously had a general tb_flush during the MTTCG implementation as
a temporary fix. It was changed back in 406bc339b0 and it would be nice
to minimise the flushing of code if we can. While most interactive users
aren't going to notice the temporary slow down it would suck for any
automated gdb scripting.

>
> If I was going to change anything here, I'd change this to not use
> cpu_get_phys_page_attrs_debug but using the caching available from the actual
> cputlb, using cc->tlb_fill() in probe mode -- something akin to 
> probe_access(),
> but not returning a host address, nor handling watchpoints nor notdirty.
>
> This would help flushing too much by distinguishing different tbs for the same
> virtual address mapping to a different physical address.
>
>
> r~


-- 
Alex Bennée