Re: save & restore failed when tmem enabled in Xen 4.1 & Xen 4.3

[Yang Hongyang <yanghy@cn.fujitsu.com>]

On 06/08/2015 02:30 PM, yunfang tai wrote:
> Hi Andrew,
>      Thank you for your reply!
>      I do not know much about migration V2. Was it integrated to Xen? If
> integrated, from which version?

It's intended to be integrated to Xen4.6. However, the libxc part has already
been merged into upstream, but the libxl part still work in progress.

>      Thank you!!
>
> Best Regards,
> Yunfang
>
> 2015-06-06 3:00 GMT+08:00 Andrew Cooper <andrew.cooper3@citrix.com
> <mailto:andrew.cooper3@citrix.com>>:
>
>     On 05/06/15 19:45, Konrad Rzeszutek Wilk wrote:
>      > On Thu, Jun 04, 2015 at 10:27:06PM +0800, yunfang tai wrote:
>      >> Hi all,
>      > Hey!
>      >>     Recently, I am testing the TMEM support on Xen. I discovered that when
>      >> enabled TMEM in ubuntu 14.10 as guest on Xen 4.1 & Xen 4.3, "xm save" & "xm
>      >> restore“ failed after there are more than 1000 pages put in persistent pool
>      >> of TMEM in Xen. My operations are list as follows:
>      > Is it exactly 1000 or just about? I presume it does not matter how much but
>      > that you discovered it by having 1000 of them?
>      >
>      >> In ubuntu guest (8 cores , 8GB):
>      >> sudo modprobe tmem
>      >> (than wait for the selfballoon to finish)
>      >> dd if=/dev/zero of=/tmp/test.img bs=10M count=1000
>      >> dd if=/tmp/test.img of=/dev/null bs=10M
>      >> dd if=/tmp/test.img of=/dev/null bs=10M
>      >> .........
>      >> (until more than 1000 pages put in persistent pool)
>      >> In Domain 0:
>      >> (add tmem in grub.cfg)
>      >> xm save ubuntu test.save
>      >> xm restore ubuntu test.save
>      >>
>      >> When TMEM is not enabled, save & restore success after these operations.
>      >> But if TMEM is enabled, save & restore fail.
>      > Are there any errors from the logs? Anything?
>      >> Does anyone test about save & restore when enabled TMEM in Xen?? Is there
>      >> anything I do wrong?
>      > Well lets see what broke. But I think Andrew discovered that the
>      > migration protocol when it came to 'tmem' was not up to snuff. CC-ing him
>      > just to confirm.
>      >
>      > (Andrew, for the persistent part of this - it conceptually should
>      > get all of the tmem memory that pushed to the hypervisor back in the
>      > image. When you were looking at migrationv2 did you just skim through
>      > that or mostly ignored it?)
>
>     Took a look at the code, attempted to figure out what was going on, then
>     decided to ignore it for the time being.
>
>     As a baseline, there is no error checking of hypercalls or their
>     returned data putting the data into the stream.
>
>     Migration v2 currently has no TMEM support, and I would suggest
>     re-implementing it from scratch over attempting to port what currently
>     exists for legacy.
>
>     ~Andrew
>
>
>
>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@lists.xen.org
> http://lists.xen.org/xen-devel
>

-- 
Thanks,
Yang.