[Xen-devel] RE: blue screen in windows balloon driver

["James Harper" <james.harper@xxxxxxxxxxxxxxxx>]

I have just pushed a change to check the
"\KernelObjects\LowMemoryCondition" event before allocating memory for
ballooning, and waiting if the event is set. This may resolve the
problems you are seeing.

What I have seen is that initially the event gets set, but then as
Windows pages some active memory out the event gets cleared again and
further ballooning down is possible. It may prevent you ballooning down
quite as low as you could before, but if it stops windows crashing then
I think it is good.

James

> -----Original Message-----
> From: MaoXiaoyun [mailto:tinnycloud@xxxxxxxxxxx]
> Sent: Monday, 28 February 2011 19:30
> To: xen devel
> Cc: James Harper
> Subject: RE: blue screen in windows balloon driver
> 
> Hi James:
> 
>      Unfortunately, We still hit the blue screen on the stress test.
>      (Start total 24 HVMS on a single 16core, 24G host,
>       each HVM owns 2G Memory, start with memory=512M,
>       and inside two eating memory  processes, each of which will each
1G
> memory)
> 
>      As I go though the code, I noticed that all memory allocation
relates to
>      "ExAllocatePoolWithTag(NonPagedPool,...)", which is from
NonePaged Pool,
>       As I know, the NonePagePool memory is the memory could not be
paged out,
>       and that is limited, and for the blue screen VMS, I also found
the free
> memory
>       is quite low, only about hundreds KB left.
> 
>      So, when memory overcommit, some of the VM will not got enough
memory,
> and if  most of its Memory is occupied by eating memory process, then
> ExAllocatePoolWithTag
> will fail, thus caused "NO_PAGES_AVALIABLE" blue screen. Is this
possible?
> 
>       Meanwhile, I will have your PVdriver tested to see if blue
exists,
> thanks.
> 
> 
> >From: tinnycloud@xxxxxxxxxxx
> >To: tinnycloud@xxxxxxxxxxx
> >Subject: FW: blue screen in windows balloon driver
> >Date: Mon, 28 Feb 2011 16:16:59 +0800
> >
> >
> >Thanks for fixing the POD. It's is better make it earlier to avoid
crash.
> >
> >The meminfo is written every 1 seconds into xenstore dir
> /local/domain/did/memory/meminfo.
> >And to avoid to many writes, only the memory changes large than 5M,
the
> thread will do the write.
> >
> >As for userspace daemon, it is our first choice, but we found it
xenstore
> daemon in dom0 comsume
> >many CPU(we test in linux only), so we decide to move it into driver.
> >
> >I've done merge my code with latest changeset 866, and do the stree
test
> later.
> >
> >many thanks.
> >
> >> Subject: RE: RE: blue screen in windows balloon driver
> >> Date: Sun, 27 Feb 2011 22:25:28 +1100
> >> From: james.harper@xxxxxxxxxxxxxxxx
> >> To: tinnycloud@xxxxxxxxxxx; xen-devel@xxxxxxxxxxxxxxxxxxx
> >> CC: george.dunlap@xxxxxxxxxxxxx
> >>
> >> > Thanks James.
> >> >
> >> > I think it is GPLPV. The driver is from
> >> http://xenbits.xen.org/ext/win-
> >> > pvdrivers.hg
> >> > But, I have done some other things
> >> >
> >> > 1) Add pod support
> >> > 2) enable a meminfo thread, periodically write VM meminfo into
> >> xenstore
> >> > We use info of Current Memory, Free memory, and Committed memory,
> >> retrived
> >> > through NativeAPI
> >> > 3) our code is based from changeset 823, attached is the diff of
my
> >> current
> >> > code with changset of 853.
> >> >
> >> > Maybe I need add my code to 853, and test again.
> >> > Thanks.
> >> >
> >>
> >> As per other post, I have just committed some patches and PoD
should now
> >> be working properly. I can start a DomU with 4GB of maxmem but only
> >> 128MB of populated memory without any problems. This now works
because I
> >> do the initial balloon down in DriverEntry, way before xenpci does
> >> anything else. Before it would blow up in DriverEntry. I think I
> >> determine the amount to initially balloon down a little differently
from
> >> you too.
> >>
> >> It takes a while to balloon down the memory though... I think
Windows
> >> tends to delay large allocations or something, because ballooning
up
> >> again is pretty much instant.
> >>
> >> How often are you writing meminfo stuff into xenstore? Could you do
that
> >> in userspace (the interface to xenstore exists and seems to work
well
> >> although it's a little tedious)? You would then be able to just run
it
> >> as a service and not need to patch GPLPV.
> >>
> >> James
> >

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel