Re: [Qemu-devel] [RFC 0/6] memory: make dirty_memory[] accesses atomic

[Stefan Hajnoczi]

On Thu, Nov 27, 2014 at 01:21:54PM +0000, Peter Maydell wrote:
> On 27 November 2014 at 12:29, Stefan Hajnoczi <address@hidden> wrote:
> > 1. Convert all cpu_physical_memory_*_dirty() callers to use the API 
> > atomically.
> >    There are TCG callers who things along the lines of:
> >
> >      if (!cpu_physical_memory_get_dirty(addr)) {
> >          cpu_physical_memory_set_dirty(addr);  /* not atomic! */
> >      }
> 
> Which bit of code is this? Note that for the TCG DIRTY_MEMORY_CODE
> flag you have bigger problems than just whether the bitmap updates
> are atomic, because the sequence is:
>  write to memory
>  if (!dirty) {
>      flush generated code tbs;
>      set dirty;
>  }
> 
> and what you care about is that the existence of cached translations
> for this area of memory should be in sync with the state of the dirty
> bit, so the whole operation of "flush affected translations and set
> the dirty bit" needs to be thread-safe, I think.

This is an example of what I mean.

I'm not going to work on making TCG thread-safe in this series, and
there is no dangerous race condition in this code if we leave it as is.

But I'm not 100% sure of all cases, so I'll audit them again.

Stefan

pgpwDYlw2Vex6.pgp
Description: PGP signature