RE: slow xp hibernation revisited

["James Harper" <james.harper@bendigoit.com.au>]

> 
> On 04/06/2011 05:54, "James Harper" <james.harper@bendigoit.com.au>
wrote:
> 
> >> It's the !test_bit(address_offset>>XC_PAGE_SHIFT,
> > entry->valid_mapping)
> >> that is causing the if expression to be true. From what I can see
so
> >> far, the bit representing the pfn in entry->valid_mapping is 0
because
> >> err[] returned for that pfn was -EINVAL.
> >>
> >> Maybe the test is superfluous? Is there a need to do the remap if
all
> >> the other variables in the expression are satisfied? If the remap
was
> >> already done and the page could not be mapped last time, what
reasons
> >> are there why it would succeed this time?
> >>
> >
> > FWIW, removing the test_bit makes the hibernate go faster than my
screen
> > can refresh over a slow DSL connection and in a quick 30 second test
> > doesn't appear to have any adverse effects.
> >
> > If there is a chance that a subsequent call to qemu_remap_bucket
with
> > identical parameters could successfully map a page that couldn't be
> > mapped in the previous call, are there any optimisations that could
be
> > done? Maybe only attempt to map the page being accessed rather than
all
> > pages in the bucket if the other parameters are identical?
> 
> I'm guessing this happens because of frequent guest CPU access to
non-RAM
> during hibernate? Unfortunately really the qemu checks do make sense,
I'd
> say, since the memory map of the guest can be changed dynamically ,
and we
> currently only flush the map_cache on XENMEM_decrease_reservation
> hypercalls.
> 
> One fix would be for Xen to know which regions of non-RAM are actually
> emulated device areas, and only forward those to qemu. It could then
> quick-fail on the rest.
> 
> However, the easiest fix would be to only re-try to map the one pfn
under
> test. Reloading a whole bucket takes bloody ages as they are *huge*:
256kB
> in 32-bit qemu; 4MB in 64-bit qemu. It might be easiest to do a test
re-map
> of the one page to a scratch area, then iff it succeeds, *then* call
> qemu_remap_bucket(). Then you remap the bucket only if something
really has
> changed, and you don't have to mess too much with modifying the bucket
> yourself outside of remap_bucket.
> 
> How does that sound?
> 

Sounds like a plan. Is there no way to detect the changes to the memory
map of the guest?

Looking past the test_bit call, the next statement does another test and
sets last_address_index to 0 and returns NULL. Is this just to ensure
that the next access isn't just trivially accepted?

James