Anthony Liguori wrote:
> Marcelo Tosatti wrote:
>> This is Linux's behaviour for all filesystems. There is no error
>> checking on MAP_POPULATE's attempt to prefault pages.
>>   
> 
> I thought that mmap() will fail with hugetlbfs if enough large pages 
> can't be reserved?

This appears to be the case for MAP_SHARED.  For MAP_PRIVATE
the accounting is side-stepped, the result of which was the
best-effort prealloc/populate behavior we had see seen.

Marcelo and I batted this around some yesterday.  And there
doesn't appear to be any concern using MAP_SHARED for the
phys_ram map as potential access to the backing file is
still regulated by filesystem permissions.

Doing so dispenses with the syscall/useraddr hack to fault
in pages since mmap() fails as expected in the case
sufficient pages are unavailable immediately to back
MAP_SHARED pages.  We also decided to defer the map-fallback
logic.  This is essentially behavior we'd like to see
provided by kvm, ideally on a dynamic (guest fault) basis
and was technically beyond the issue we were trying to
target.

Attached is the resulting patch which also addresses the
other concerns which arose.

-john

-- 
john.cooper@third-harmonic.com