Linux Xen balloon driver "minimum target"

Linux Xen balloon driver "minimum target"

[Dan Magenheimer]

As recently re-pointed out by George Shuklin:
http://lists.xensource.com/archives/html/xen-devel/2010-06/msg00849.html 
with brief reply commentary by Keir:
http://lists.xensource.com/archives/html/xen-devel/2010-06/msg00875.html 

the Xen ballooning code in many distros contains a
"minimum target" to allow a domain to protect itself
against silly and accidental balloon settings, it being
frightfully easy for a typo or misunderstanding (e.g.
an integer intended to represent MB but, oops, the value
should be given in KB) to result in a guest or dom0 crash.

This code originated from Novell about two years ago:
http://lists.xensource.com/archives/html/xen-devel/2008-04/msg00143.html 
and, in my experience, eliminated frequent bug reports that
"xm memset" resulted in unexpected crashes.

This minimum target is a bit over-conservative, especially
on machines with 1GB-2GB of memory.  Also, the
"logarithmic function" is applied to different values in
different kernels, so is in some cases a function of the
config file's "maxmem" and in some cases a function of "memory".

Because the variations of this code are broad (and, in upstream
Linux, non-existent), posting a patch doesn't solve the problem.
So instead, I thought I would open a discussion with my opinions
and "solution".  We'll keep this on xen-devel for now, but
perhaps should open the discussion to xen-users as well.

I am in favor of keeping the code AND inserting some form of
it in the upstream balloon driver.  There are two forms
of policy it enforces: (1) don't let an admin do something
REALLY stupid; and (2) "stupid" is dependent on the amount
of guest physical memory.  While I don't much like (2), I
don't see a way one can get (1) without (2).

As a partial answer, in my balloon driver implementation
(to support self-ballooning) I added an additional proc/sysfs
variable "min_target_kb".  This is set during init to the
result of the "minimum target" algorithm and, once set,
is used as the minimum allowed ballooning target.  BUT
this variable is writable, allowing the safety latch to
be overridden.  As a result, a system admin must shoot himself
in the foot before accidentally shooting himself in the head.

Clearly this is still not an ideal solution, but it works
for me.

Comments?

Re: Linux Xen balloon driver "minimum target"

[George Dunlap]

It sounds like this safety measure increases reliability; but at the
same time, I'm opposed to safety measures that can't be disabled or
overridden.  Seems like a reasonable solution to me.

 -G

On Fri, Jun 18, 2010 at 4:24 PM, Dan Magenheimer
<dan.magenheimer@oracle.com> wrote:
> As recently re-pointed out by George Shuklin:
> http://lists.xensource.com/archives/html/xen-devel/2010-06/msg00849.html
> with brief reply commentary by Keir:
> http://lists.xensource.com/archives/html/xen-devel/2010-06/msg00875.html
>
> the Xen ballooning code in many distros contains a
> "minimum target" to allow a domain to protect itself
> against silly and accidental balloon settings, it being
> frightfully easy for a typo or misunderstanding (e.g.
> an integer intended to represent MB but, oops, the value
> should be given in KB) to result in a guest or dom0 crash.
>
> This code originated from Novell about two years ago:
> http://lists.xensource.com/archives/html/xen-devel/2008-04/msg00143.html
> and, in my experience, eliminated frequent bug reports that
> "xm memset" resulted in unexpected crashes.
>
> This minimum target is a bit over-conservative, especially
> on machines with 1GB-2GB of memory.  Also, the
> "logarithmic function" is applied to different values in
> different kernels, so is in some cases a function of the
> config file's "maxmem" and in some cases a function of "memory".
>
> Because the variations of this code are broad (and, in upstream
> Linux, non-existent), posting a patch doesn't solve the problem.
> So instead, I thought I would open a discussion with my opinions
> and "solution".  We'll keep this on xen-devel for now, but
> perhaps should open the discussion to xen-users as well.
>
> I am in favor of keeping the code AND inserting some form of
> it in the upstream balloon driver.  There are two forms
> of policy it enforces: (1) don't let an admin do something
> REALLY stupid; and (2) "stupid" is dependent on the amount
> of guest physical memory.  While I don't much like (2), I
> don't see a way one can get (1) without (2).
>
> As a partial answer, in my balloon driver implementation
> (to support self-ballooning) I added an additional proc/sysfs
> variable "min_target_kb".  This is set during init to the
> result of the "minimum target" algorithm and, once set,
> is used as the minimum allowed ballooning target.  BUT
> this variable is writable, allowing the safety latch to
> be overridden.  As a result, a system admin must shoot himself
> in the foot before accidentally shooting himself in the head.
>
> Clearly this is still not an ideal solution, but it works
> for me.
>
> Comments?
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xensource.com
> http://lists.xensource.com/xen-devel
>