[CRIU] lxc-checkpoint problems with latest lxc and criu from git

[Tycho Andersen]

On Fri, Jul 08, 2016 at 04:56:35PM +0200, Adrian Reber wrote:
> On Fri, Jul 08, 2016 at 08:27:27AM -0600, Tycho Andersen wrote:
> > On Fri, Jul 08, 2016 at 11:07:11AM +0300, Cyrill Gorcunov wrote:
> > > On Fri, Jul 08, 2016 at 10:03:25AM +0200, Adrian Reber wrote:
> > > > Using latest lxc and criu from git I see errors during restore which I
> > > > have not seen before and which are difficult to reproduce but sometimes
> > > > I see:
> > > > 
> > > > 
> > > > Error (cgroup.c:1193): cg: Failed writing c *:* m
> > > > b *:* m
> > > > c 1:3 rwm
> > > > c 1:5 rwm
> > > > c 1:7 rwm
> > > > c 5:0 rwm
> > > > c 5:2 rwm
> > > > c 1:8 rwm
> > > > c 1:9 rwm
> > > > c 136:* rwm
> > > > c 10:229 rwm to devices//lxc/c7/devices.allow: Argument list too long
> > > > Error (cr-restore.c:1110): 5019 killed by signal 9: Killed
> > > > Error (cr-restore.c:1990): Restoring FAILED.
> > > > 
> > > > I might also have messed up my installation as I have parallel installed
> > > > criu (2.2) and lxc (2.0.3) from the package management system.
> > > > 
> > > > I just wanted to mention it in case something is really broken.
> > > 
> > > Tycho?
> > 
> > What kernel are you using? Based on my read of the kernel code we
> > should be able to write up to PAGE_SIZE bytes to devices, but maybe we
> > need to split this up into more granular writes.
> 
> It's the RHEL7 kernel. So it claims it is 3.10, but with lots of patches
> on top. So it definitely could be that it behaves differently. Where in
> the kernel code you are exactly looking at,

see security/device_cgroup.c and kernel/cgroup.c; cgroup_file_write is
declared in kernfs with an atomic_write_len of PAGE_SIZE, and the
devices cgroup doesn't declare its own atomic_write_len to override
it. Perhaps it is different on 3.10, though; I know a lot of cgroup
code has been cleaned up since then (and there is at least one other
difference in behavior Andrey pointed out to me in some other thread).

Anyway, the way we do this is a little hokey, so I'm sending some
patches shortly that should clean it up. If you don't mind applying
those and trying that would be handy.

Tycho

> 		Adrian