svn commit: r267858 - in head/sys/dev: virtio/balloon xen/balloon

[Attilio Rao]

On Wed, Jun 25, 2014 at 3:29 PM, Roger Pau Monné <royger at freebsd.org> wrote:
> On 25/06/14 13:58, Attilio Rao wrote:
>> On Wed, Jun 25, 2014 at 11:51 AM, Roger Pau Monné <royger at freebsd.org> wrote:
>>> Author: royger
>>> Date: Wed Jun 25 09:51:08 2014
>>> New Revision: 267858
>>> URL: http://svnweb.freebsd.org/changeset/base/267858
>>>
>>> Log:
>>>   xen/virtio: fix balloon drivers to not mark pages as WIRED
>>>
>>>   Prevent the Xen and VirtIO balloon drivers from marking pages as
>>>   wired. This prevents them from increasing the system wired page count,
>>>   which can lead to mlock failing because of hitting the limit in
>>>   vm.max_wired.
>>
>> This change is conceptually wrong.
>> The pages balloon is allocating are unmanaged and they should be wired
>> by definition. Alan and I are considering enforcing this (mandatory
>> wired pages for unmanaged pages allocation) directly in the KPI.
>> This in practice just seem an artifact to deal with scarce  wired
>> memory limit. I suggest that for the XEN case this limit gets bumped
>> rather relying on similar type of hacks.
>
> IMHO, marking them as wired seems wrong too, those pages are not wired,
> they are simply not there any more. This was discussed in:

I'm not entirely sure what do you mean with "not there anymore", so
I'm just guessing and I assume that you mean "pages are not collected
in any vm object and then they are not referenced in any pagequeue".
By extension this also matches what "unmanaged page" means.

If the page is unmanaged it means that the pagedaemon won't see it, so
they won't be swapped out anyway. Wiring them it will enforce more
sanity checking on the page. The max_wired concern may be real,
however, please see below.

> http://lists.freebsd.org/pipermail/freebsd-virtualization/2014-June/002643.html
>
> If there's consensus I will revert the change, but I would say that
> increasing vm.max_wired for VMs is also a gross hack.

Why? If VM needs more wired memory I assume that we can tune up the
default value of max_wired?

I think that however your case makes an interesting point: if we want
to make unmanaged pages as inherently wired, we likely need a little
bit higher max_wired value. When I completed a patch for this, pho@
couldn't reproduce any similar issue even with stress-testing (and
also, the places to allocate unmanaged pages and not requesting
VM_ALLOC_WIRED were very little, almost 0, with the exception of
vm_page_alloc_contig() calls) but I think it is a valid proposition.

However I would still like to have more control on kernel-specific
wired memory for processes. I'm for example thinking to ARC vs. buffer
cache, where I expect the wired memory consumption to be much bigger
for the former case.

Attilio


-- 
Peace can only be achieved by understanding - A. Einstein