Re: TTM huge page-faults WAS: Re: [RFC PATCH 1/2] x86: Don't let pgprot_modify() change the page encryption bit

From: Thomas HellstrÃm (VMware)
Date: Wed Sep 11 2019 - 11:08:48 EST

On 9/11/19 4:06 PM, Koenig, Christian wrote:
Am 11.09.19 um 12:10 schrieb Thomas HellstrÃm (VMware):
The problem seen in TTM is that we want to be able to change the
vm_page_prot from the fault handler, but it's problematic since we
have the mmap_sem typically only in read mode. Hence the fake vma
hack. From what I can tell it's reasonably well-behaved, since
pte_modify() skips the bits TTM updates, so mprotect() and mremap()
works OK. I think split_huge_pmd may run into trouble, but we don't
support it (yet) with TTM.
Ah! I actually ran into this while implementing huge page support for
TTM and never figured out why that doesn't work. Dropped CPU huge page
support because of this.
By incident, I got slightly sidetracked the other day and started
looking at this as well. Got to the point where I figured out all the
hairy alignment issues and actually got huge_fault() calls, but never
implemented the handler. I think that's definitely something worth
having. Not sure it will work for IO memory, though, (split_huge_pmd
will just skip non-page-backed memory) but if we only support
VM_SHARED (non COW) vmas there's no reason to split the huge pmds
anyway. Definitely something we should have IMO.
Well our primary use case would be IO memory, cause system memory is
only optionally allocate as huge page but we nearly always allocate VRAM
in chunks of at least 2MB because we otherwise get a huge performance

But that system memory option is on by default, right? In any case, a request for a huge_fault
would probably need to check that there is actually an underlying huge_page and otherwise fallback to ordinary faults.

Another requirement would be for IO memory allocations to be PMD_PAGE_SIZE aligned in the mappable aperture, to avoid fallbacks to ordinary faults. Probably increasing fragmentation somewhat. (Seems like pmd entries can only point to PMD_PAGE_SIZE aligned physical addresses) Would that work for you?

We could probably get away with a WRITE_ONCE() update of the
vm_page_prot before taking the page table lock since

a) We're locking out all other writers.
b) We can't race with another fault to the same vma since we hold an
address space lock ("buffer object reservation")
c) When we need to update there are no valid page table entries in the
vma, since it only happens directly after mmap(), or after an
unmap_mapping_range() with the same address space lock. When another
reader (for example split_huge_pmd()) sees a valid page table entry,
it also sees the new page protection and things are fine.
Yeah, that's exactly why I always wondered why we need this hack with
the vma copy on the stack.

But that would really be a special case. To solve this properly we'd
probably need an additional lock to protect the vm_flags and
vm_page_prot, taken after mmap_sem and i_mmap_lock.
Well we already have a special lock for this: The reservation object. So
memory barriers etc should be in place and I also think we can just
update the vm_page_prot on the fly.
I agree. This is needed for huge pages. We should make this change,
and perhaps add the justification above as a comment.
Alternatively we could introduce a new VM_* flag telling users of
vm_page_prot to just let the pages table entries be filled by faults again

An interesting idea, although we'd lose things like dirty-tracking bits.