Re: [Xen-devel] [PATCH] libxl: fix "xl mem-set" regression from 0c029c4da2

[Ian Campbell <ian.campbell@xxxxxxxxxx>]

On Wed, 2015-05-13 at 08:18 +0100, Jan Beulich wrote:
> >>> On 22.04.15 at 19:55, <Ian.Campbell@xxxxxxxxxx> wrote:
> > On Wed, 2015-04-22 at 17:33 +0100, Jan Beulich wrote:
> >> >>> On 22.04.15 at 17:36, <ian.campbell@xxxxxxxxxx> wrote:
> >> > On Wed, 2015-04-22 at 15:41 +0100, Jan Beulich wrote:
> >> >> >>> On 22.04.15 at 16:01, <ian.campbell@xxxxxxxxxx> wrote:
> >> >> > On Wed, 2015-04-22 at 13:02 +0100, Jan Beulich wrote:
> >> >> >> Said commit ("libxl_set_memory_target: retain the same maxmem offset 
> >> >> >> on
> >> >> >> top of the current target") caused a regression for "xl mem-set"
> >> >> >> against Dom0: While prior to creation of the first domain this works,
> >> >> >> the first domain creation involving ballooning breaks. Due to 
> >> >> >> "enforce"
> >> >> >> not being set in the domain creation case, and due to Dom0's initial
> >> >> >> ->max_pages (in the hypervisor) being UINT_MAX, the calculation of
> >> >> >> "memorykb" in the first "xl mem-set" adusting the target upwards
> >> >> >> subsequent to domain creation and termination may cause an overflow,
> >> >> >> resulting in Dom0's maximum getting to a very small value. This small
> >> >> >> maximum will the make the subsequent setting of the PoD target fail.
> >> >> >> 
> >> >> >> Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx>
> >> >> >> ---
> >> >> >> Note that this only fixes the immediate problem - there appear to be
> >> >> >> further issues lurking here:
> >> >> >> - libxl_set_memory_target()'s *_memkb variables all being 32-bit,
> >> >> >> - libxl_domain_setmaxmem()'s max_memkb parameter being 32-bit,
> >> >> > 
> >> >> > I think that increasing the width of these variables wouldn't break 
> >> >> > the
> >> >> > API guarantee which we make, at least not in a practical way, since 
> >> >> > any
> >> >> > existing 32-bit arguments passed will just get promoted.
> >> >> 
> >> >> No, not even on 64-bit. On 32-bit, two arguments slots are needed
> >> >> for what so far requires only one. On 64-bit (at least x86-64), the
> >> >> calling code isn't required to zero-extend a value calculated in a
> >> >> register (e.g. a result of earlier calculations which had more than
> >> >> 32 significant bits could be passed unchanged to the called function);
> >> >> it just so happens that 32-bit arithmetic on registers would always
> >> >> implicitly zero the upper halves (and iirc that's the same on ARM64).
> >> > 
> >> > You seem to be talking about ABI? As I tried to note in my response for
> >> > libxl we only make guarantees about the API (P not B in the middle).
> >> 
> >> Oh, okay. That would mean the libxl side is fine. The change to the
> >> libxc interface might however still be a problem for qemu (Stefano
> >> was telling me something to the effect of a compatibility layer in
> >> upstream qemu to deal with such version differences, but of course
> >> the implication - without versioned libxc interfaces - would still be
> >> that existing qemu binaries on top of updated libxc would no
> >> longer work).
> > 
> > Correct, but we would surely bump the SONAME for a change such as this,
> > so it wouldn't matter in practice.
> 
> So what's the situation with this patch? Can it go in? Is someone
> working on a better fix for the described problem?

Stefano, Are you?

Regardless it seems to me that this patch is correct in its own right,
having maxmem_kb be a 64-bit type is the correct type for it to have.

Ian.


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel