On Mon, Nov 13, 2017 at 10:17:30PM +0100, Thomas Gleixner wrote: > On Mon, 13 Nov 2017, Kirill A. Shutemov wrote: > > On Mon, Nov 13, 2017 at 05:57:03PM +0100, Thomas Gleixner wrote: > > > On Mon, 13 Nov 2017, Kirill A. Shutemov wrote: > > > > > > > On Mon, Nov 13, 2017 at 04:43:26PM +0100, Thomas Gleixner wrote: > > > > > On Tue, 7 Nov 2017, Kirill A. Shutemov wrote: > > > > > > > > > > > In case of 5-level paging, we don't put any mapping above 47-bit, > > > > > > unless > > > > > > userspace explicitly asked for it. > > > > > > > > > > > > Userspace can ask for allocation from full address space by > > > > > > specifying > > > > > > hint address above 47-bit. > > > > > > > > > > > > Nicholas noticed that current implementation violates this > > > > > > interface: > > > > > > we can get vma partly in high addresses if we ask for a mapping at > > > > > > very > > > > > > end of 47-bit address space. > > > > > > > > > > > > Let's make sure that, when consider hint address for non-MAP_FIXED > > > > > > mapping, start and end of resulting vma are on the same side of > > > > > > 47-bit > > > > > > border. > > > > > > > > > > What happens for mappings with MAP_FIXED which cross the border? > > > > > > > > It will succeed with 5-level paging. > > > > > > And why is this allowed? > > > > > > > It should be safe as with 4-level paging such request would fail and > > > > it's > > > > reasonable to expect that userspace is not relying on the failure to > > > > function properly. > > > > > > Huch? > > > > > > The first rule when looking at user space is that is broken or > > > hostile. Reasonable and user space are mutually exclusive. > > > > That's basically the same assumption we made to implement current > > interface of allocation memory above 47-bits. > > > > The premise is that nobody in right mind would try mmap(addr, MAP_FIXED) > > where addr >= (1UL << 47) as it will always fail. So we can allow this to > > succeed on 5-level paging machine as a way to allocate from larger address > > space. > > > > By the same logic we can allow allocation for cases where addr is below > > (1UL << 47), but addr+size is above the limit. > > Makes some sense, but it would be nice to have this documented exactly in > arch_get_unmapped_area_topdown(), i.e. the function where you are adding > the border check to. Otherwise 3 month from now somebody will look at that > and ask exactly the same question again.
Okay, I'll update the patch. -- Kirill A. Shutemov
