Re: [Xen-ia64-devel] [PATCH]: vcpu_match_tr_entry & vcpu_ptc_ga

[Tristan Gingold <Tristan.Gingold@xxxxxxxx>]

Le Vendredi 31 Mars 2006 04:40, Isaku Yamahata a écrit :
> On Thu, Mar 30, 2006 at 04:13:31PM +0100, Tristan Gingold wrote:
> > after re-reading code, I think the only possible race is here:
> >
> >       if (/* is_data && */ vcpu_match_tr_entry(trp,address,rid)) {
> >               if (vcpu->domain==dom0 && !in_tpa) *pteval =
> > trp->page_flags;
> >
> > Do you agree ?
>
> I think so.
>
> > Maybe which should force the p bit on pteval ?
>
> I guess such a trick doesn't solve the problem.
> The essential problem is that xen's software tlb cache(vcpu->arch.dtlb)
> and physical cpu tlb cache aren't updated atomically.
>
>
>         CPU0                            CPU1
>
>         a domain issues ptc.ga          a domain issues a access
>
>                                         ia64_do_page_fault()
>                                           vcpu_translate()
>                                           vcpu->arch.dtlb cache hit.
>
>         vcpu_ptc_ga()
>         ia64_global_tlb_purge()
>                                         vcpu->rach.dtlb is purged somehow
>
>                                           vcpu_translate() returns
>                                           IA64_NO_FAULT
>
>                                           translate_domain_pte()
>                                           vcpu_itc_no_srlz()
>                                             tlb insert
>                                             re-cache vcpu->arch.dtlb
>                                             vhpt update
>
>                                          tlb cache entry which should
>                                          be purged remains.
>
>
>                                         domain's access succeeds without
>                                         domain seeing fault.
>
>
>                                         synchronization
>
>
>                                         a domain again issues access and
>                                         succeeds without seeing faults.
>                                         However this access should result
>                                         in fault.
Correct.

Thank you for pointing this issue.

Tristan.



_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel