Re: [PATCH 0/3] xenoprof: XSA-313 follow-up

[Julien Grall <julien@xxxxxxx>]

Hi Paul,

On 20/04/2020 16:01, Paul Durrant wrote:
> > -----Original Message-----
> > From: Julien Grall <julien@xxxxxxx>
> > Sent: 20 April 2020 15:15
> > To: paul@xxxxxxx; 'Jan Beulich' <jbeulich@xxxxxxxx>; 
> > xen-devel@xxxxxxxxxxxxxxxxxxxx
> > Cc: 'Andrew Cooper' <andrew.cooper3@xxxxxxxxxx>; 'George Dunlap' 
> > <george.dunlap@xxxxxxxxxx>; 'Ian
> > Jackson' <ian.jackson@xxxxxxxxxxxxx>; 'Stefano Stabellini' 
> > <sstabellini@xxxxxxxxxx>; 'Wei Liu'
> > <wl@xxxxxxx>
> > Subject: Re: [PATCH 0/3] xenoprof: XSA-313 follow-up
> >
> > Hi Paul,
> >
> > On 15/04/2020 09:50, Paul Durrant wrote:
> > > > -----Original Message-----
> > > > From: Jan Beulich <jbeulich@xxxxxxxx>
> > > > Sent: 15 April 2020 09:45
> > > > To: xen-devel@xxxxxxxxxxxxxxxxxxxx
> > > > Cc: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>; George Dunlap 
> > > > <george.dunlap@xxxxxxxxxx>; Ian
> > Jackson
> > > > <ian.jackson@xxxxxxxxxxxxx>; Julien Grall <julien@xxxxxxx>; Stefano Stabellini
> > > > <sstabellini@xxxxxxxxxx>; Wei Liu <wl@xxxxxxx>; Paul Durrant <paul@xxxxxxx>
> > > > Subject: [PATCH 0/3] xenoprof: XSA-313 follow-up
> > > >
> > > > Patch 1 was considered to become part of the XSA, but it was then
> > > > decided against. The other two are a little bit of cleanup, albeit
> > > > there's certainly far more room for tidying. Yet then again Paul,
> > > > in his mail from Mar 13, was asking whether we shouldn't drop
> > > > xenoprof altogether, at which point cleaning up the code would be
> > > > wasted effort.
> > > That's still my opinion. This is a large chunk of (only passively maintained) 
> > > code which I think is
> > of very limited value (since it relates to an old tool, and it only works for 
> > PV domains).
> >
> > While there are no active user we are aware of, this is an example on
> > how to implement a profiler backend with Xen. So I would agree with
> > Andrew here.
> >
> > IIRC, the reason behind your request is it makes difficult for your
> > xenheap work. Am I correct? If so, do you have a thread explaining the
> > issues?
> After shared info and grant table, it is the only other occurrence of a xenheap 
> page shared with a non-system domain. Also, it cannot be trivially replaced 
> with an 'extra' domheap page because its assignment changes. Thus a whole bunch 
> of cleanup work that I was hoping to do (largely in domain_relinquish_resources 
> and free_domheap_pages) is either ruled out, or would have to special-case this 
> type of page.
My knowledge of xenoprof is very limited, so I might be wrong.

From my understanding, a buffer can only be shared between two domains:
   - When in passive mode, the buffer will be shared with the primary 
profiler (always the hwdom per the check in xenoprof_op_init()).
   - When in active mode, the buffer will be shared with the domain 
requesting to be profiled.
Would it be possible to consider to have a separate buffer for the 
passive mode and active mode?
> Also, I am unconvinced that PV guests are sufficiently common these days (apart 
> from dom0) that profiling them is of any real use.
Even an HVM guest can't profile itself, I think it would be useful to 
have dom0 to profile an HVM guest. Are you suggesting this doesn't work?
Cheers,

--
Julien Grall