Re: [PATCH] xenpaging: libxl support

[Olaf Hering <olaf@aepfle.de>]

On Tue, Sep 27, Ian Jackson wrote:

> Olaf Hering writes ("[Xen-devel] [PATCH] xenpaging: libxl support"):
> > The patch adds four new config options:
> > xenpaging=<int> , the number of pages to page-out
> ...
> > +libxl_xenpaging_info = Struct("xenpaging_info",[
> > +    ("xenpaging",                 integer,    False, "number of pages"),
> ...
> > +        if (!xlu_cfg_get_long (config, "xenpaging", &l))
> > +            xp_info->xenpaging = l;
> 
> The config file setting should be a number of megabytes, just like
> "memory", "maxmem", etc., not a number of pages.  And at the libxl
> interface it should be a number of kilobytes like max_memkb et al.

I will change that part.

> Also as I said at the hackathon, I think that we need a more
> sophisticated approach to the interaction with ballooning.  I would
> like to see the argument to the xenpaging daemon not be a target
> number of pages to page out, but rather for it to be a total memory
> usage target.

That would mean xenpaging has to monitor tot_pages and try to reach the
configured number?

> The effect of this would be that you could tell a guest to balloon
> down, but also tell xenpaging to page it out, and balance between
> ballooning and paging is then up to the guest.

Does a guest even know its paged? If I understand it correctly, mem-set
will ask the guests balloon driver to work toward that number, but if
xenpaging is instructed to page-out the entire guest, the guest wont
notice it. 
A new mem-tot_pages command (with a better name) would limit the amount
of consumed mfns, while mem-set would limit the amount of mfns the guest
balloon driver thinks it has.

Is it that what you have in mind?

> > +static int libxl__create_xenpaging(libxl__gc *gc, char *dom_name, uint32_t domid, libxl_xenpaging_info *xp_info)
> > +{
> > + [a lot of stuff]
> 
> Can any of this be factored out and made common with the device model
> creation ?

Yes. The current device model startup uses the spawn functions, and
things like *_record_pid and libxl__wait_for_* should be part of the
spawn interface itself because the way libxl__spawn_spawn works a read
from the pipe has to happen in some way. I will prepare patches for
review.

> Also:
>   - what deletes the logfile, if anything ?

The pagefile is unlinked by xenpaging on exit.

>   - will the xenpaging daemon automatically exit if the domain dies ?

xenpaging has a watch on @releaseDomain.

Olaf