[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Xen-devel] [memory sharing] bug on get_page_and_type



Hi Tim:
 
         More thoughts on this bug.
        
         First some questions
 

         1) What PGT_writeable_page means to a page?

         2) When a page type will be changed to PGT_writeable_page?

         3) It looks like PGT_writeable_page is not sharable? Since only PGT_none can, right?

         4) Could I use get_page_type(page, PGT_writeable_page) before every is_p2m_shared() check.

              

             Since if get_page_type() success, then the page will has no chance to be shared later

             and if get_page_type() failed, it page mush has type, it is either  PGT_shared_page or other types,

             if other types, the page still has no chance to be shared.

             if PGT_shared_page, that's ok, just do usual is_p2m_shared return routine.

 

             question is, is it ok if we only get_page_type, and not to put_page_type()?

       


 
> Date: Fri, 11 Feb 2011 09:32:18 +0000
> From: Tim.Deegan@xxxxxxxxxx
> To: tinnycloud@xxxxxxxxxxx
> CC: xen-devel@xxxxxxxxxxxxxxxxxxx; George.Dunlap@xxxxxxxxxxxxx; juihaochiang@xxxxxxxxx
> Subject: Re: [Xen-devel] [memory sharing] bug on get_page_and_type
>
> At 07:04 +0000 on 11 Feb (1297407854), MaoXiaoyun wrote:
> > Thanks Tim.
> >
> > After discuss with JuiHao, How about fix in this way?
> >
> > 1) Suppose we have a function, make_page_unsharable() to substitude
> > p2m_is_shared() check, if p2mt is not shared, we increase its type count
> > to avoid it turn to shared while using it.
>
> That's a good idea. I'd rather not have the name be anything to do with
> "sharable", but we could have a function that does a p2m lookup and a
> get-page-and-type, all under the p2m lock, and use it instead of the
> lookup-check-getref idiom elsewhere.< BR>>
> Then if (as you say) the make-shareable and nominate-page actions were
> covered by the same lock (or potentially even just called the same
> function themselves) we would eliminate a lot of races.
>
> That will be too big a patch to take before 4.1.0 but I'd consider it
> immediately after the release.
>
> Tim.
>
> > 1 int make_page_unsharable(int enable)
> > 2 {
> > 3 p2m_type_t p2mt;
> > 4 unsigned long mfn;
> > 5
> > 6 p2m_lock()
> > 7 mfn = mfn_x(gfn_to_mfn(d, gmfn, &p2mt))
> > 8
> > 9 if(p2m_is_shared(p2mt)){
> > 10 p2m_unlock()
> > 11 return 1;
> > 12 }
> > 13
> > 14 get_page_type() / ***increase page type count to avoid page type turn to shared, since in
> > mem_sharing_nominate_page->page_make_sharable, only type count zero is
> > allowed to be shared */
> > 15 p2m_unlock()
> > 16
> > 17 return 0;
> > 18 }
> >
> > 2) If p2mt is not shared, we must decrease it type count after we finish using it
> > 3) To avoid competition, page_make_sharble() and p2m_change_type() in
> > mem_sharing_nominate_page() should be protected in same p2m_lock.
> >
> > comments?
> >
> >
> > > Date: Wed, 9 Feb 2011 09:57:20 +0000
> > > From: Tim.Deegan@xxxxxxxxxx
> > > To: tinnycloud@xxxxxxxxxxx
> > > CC: xen-devel@xxxxxxxxxxxxxxxxxxx; George.Dunlap@xxxxxxxxxxxxx; juihaochiang@xxxxxxxxx
> > > Subject: Re: [Xen-devel] [memory sharing] bug on get_page_and_type
> > >
> > > At 02:46 +0000 on 09 Feb (1297219562), MaoXiaoyun wrote:
> > > > I've been looking into the TOCTOU issue quite a while, but
> > > >
> > > > 1) Th ere are quite a lot judgements like "p2m_is_shared(p2mt)" or
> > > > "p2mt == p2m_ram_shared", which, for me, is hard to tell whom
> > > > are need to be protect by p2m_lock and whom are not So is
> > > > there a rule to distinguish these?
> > >
> > > Not particularly. I suspect that most of them will need to be
> > > changed, but as I said I hope we can find something nicer than
> > > scattering p2m_lock() around non-p2m code.
> > >
> > > > 2) Could we improve p2m_lock to sparse lock, which maybe better, right?
> > >
> > > Maybe, but not necessarily. Let's get it working properly first and
> > > then we can measure lock contention and see whether fancy locks are
> > > worthwhile.
> > >
> > > Tim.
> > >
> > > >
> > > > > Date: Wed, 2 Feb 2011 16:18:37 +0000
> > > > > From: Tim.Deegan@xxxxxxxxxx
> > > > > To: tinnycloud@xxxxxxxxxxx
> > > > > CC: xen-devel@xxxxxxxxxxxxxxxxxxx; George.Dunlap@xxxxxxxxxxxxx; juihaochiang@xxxxxxxxx
> > > > > Subject: Re: [Xen-devel] [memory sharing] bug on get_page_and_type
> > > > >
> > > > > At 15:43 +0000 on 02 Feb (1296661396), MaoXiaoyun wrote:
> > > > > > Hi Tim:
> > > > > >
> > > > > > Thanks for both your advice and quick reply. I will follow.
> > > > > >
> > > > > > So at last we should replace shr_lock with p2m_lock.
> > > > > > But more complicate, it seems both the
> > > > > > *check action* code and *nominate page* code need to be locked ,right?
> > > > > > If so, quite a lot of *check action* codes nee d to be locked.
> > > > >
> > > > > Yes, I think you're right about that. Unfortunately there are some very
> > > > > long TOCTOU windows in those kind of p2m lookups, which will get more
> > > > > important as the p2m gets more dynamic. I don't want to have the
> > > > > callers of p2m code touching the p2m lock directly so we may need a new
> > > > > p2m interface to address it.
> > > > >
> > > > > Tim.
> > > > >
> > >
> > > --
> > > Tim Deegan <Tim.Deegan@xxxxxxxxxx>
> > > Principal Software Engineer, Xen Platform Team
> > > Citrix Systems UK Ltd. (Company #02937203, SL9 0BG)
>
> --
> Tim Deegan <Tim.Deegan@xxxxxxxxxx>
> Principal Software Engineer, Xen Platform Team
> Citrix Systems UK Ltd. (Company #02937203, SL9 0BG)
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel

 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.