[PATCH mm-unstable] mm/khugepaged: fix collapse_pte_mapped_thp() versus uffd

[David Hildenbrand]

On 22.08.23 17:30, Jann Horn wrote:
> On Tue, Aug 22, 2023 at 5:23 PM Matthew Wilcox <willy at infradead.org> wrote:
>> On Tue, Aug 22, 2023 at 04:39:43PM +0200, Jann Horn wrote:
>>>> Perhaps something else will want that same behaviour in future (it's
>>>> tempting, but difficult to guarantee correctness); for now, it is just
>>>> userfaultfd (but by saying "_armed" rather than "_missing", I'm half-
>>>> expecting uffd to add more such exceptional modes in future).
>>>
>>> Hm, yeah, sounds okay. (I guess we'd also run into this if we ever
>>> wanted to make it possible to reliably install PTE markers with
>>> madvise() or something like that, which might be nice for allowing
>>> userspace to create guard pages without unnecessary extra VMAs...)
>>
>> I don't know what a userspace API for this would look like, but I have
>> a dream of creating guard VMAs which only live in the maple tree and
>> don't require the allocation of a struct VMA.  Use some magic reserved
>> pointer value like XA_ZERO_ENTRY to represent them ... seems more
>> robust than putting a PTE marker in the page tables?
> 
> Chrome currently uses a lot of VMAs for its heap, which I think are
> basically alternating PROT_NONE and PROT_READ|PROT_WRITE anonymous
> VMAs. Like this:
> 
> [...]
> 3a10002cf000-3a10002d0000 ---p 00000000 00:00 0
> 3a10002d0000-3a10002e6000 rw-p 00000000 00:00 0
> 3a10002e6000-3a10002e8000 ---p 00000000 00:00 0
> 3a10002e8000-3a10002f2000 rw-p 00000000 00:00 0
> 3a10002f2000-3a10002f4000 ---p 00000000 00:00 0
> 3a10002f4000-3a10002fb000 rw-p 00000000 00:00 0
> 3a10002fb000-3a10002fc000 ---p 00000000 00:00 0
> 3a10002fc000-3a1000303000 rw-p 00000000 00:00 0
> 3a1000303000-3a1000304000 ---p 00000000 00:00 0
> 3a1000304000-3a100031b000 rw-p 00000000 00:00 0
> 3a100031b000-3a100031c000 ---p 00000000 00:00 0
> 3a100031c000-3a1000326000 rw-p 00000000 00:00 0
> 3a1000326000-3a1000328000 ---p 00000000 00:00 0
> 3a1000328000-3a100033a000 rw-p 00000000 00:00 0
> 3a100033a000-3a100033c000 ---p 00000000 00:00 0
> 3a100033c000-3a100038b000 rw-p 00000000 00:00 0
> 3a100038b000-3a100038c000 ---p 00000000 00:00 0
> 3a100038c000-3a100039b000 rw-p 00000000 00:00 0
> 3a100039b000-3a100039c000 ---p 00000000 00:00 0
> 3a100039c000-3a10003af000 rw-p 00000000 00:00 0
> 3a10003af000-3a10003b0000 ---p 00000000 00:00 0
> 3a10003b0000-3a10003e8000 rw-p 00000000 00:00 0
> 3a10003e8000-3a1000401000 ---p 00000000 00:00 0
> 3a1000401000-3a1000402000 rw-p 00000000 00:00 0
> 3a1000402000-3a100040c000 ---p 00000000 00:00 0
> 3a100040c000-3a100046f000 rw-p 00000000 00:00 0
> 3a100046f000-3a1000470000 ---p 00000000 00:00 0
> 3a1000470000-3a100047a000 rw-p 00000000 00:00 0
> 3a100047a000-3a100047c000 ---p 00000000 00:00 0
> 3a100047c000-3a1000492000 rw-p 00000000 00:00 0
> 3a1000492000-3a1000494000 ---p 00000000 00:00 0
> 3a1000494000-3a10004a2000 rw-p 00000000 00:00 0
> 3a10004a2000-3a10004a4000 ---p 00000000 00:00 0
> 3a10004a4000-3a10004b6000 rw-p 00000000 00:00 0
> 3a10004b6000-3a10004b8000 ---p 00000000 00:00 0
> 3a10004b8000-3a10004ea000 rw-p 00000000 00:00 0
> 3a10004ea000-3a10004ec000 ---p 00000000 00:00 0
> 3a10004ec000-3a10005f4000 rw-p 00000000 00:00 0
> 3a10005f4000-3a1000601000 ---p 00000000 00:00 0
> 3a1000601000-3a1000602000 rw-p 00000000 00:00 0
> 3a1000602000-3a1000604000 ---p 00000000 00:00 0
> 3a1000604000-3a100062b000 rw-p 00000000 00:00 0
> 3a100062b000-3a1000801000 ---p 00000000 00:00 0
> [...]
> 
> I was thinking if you used PTE markers as guards, you could maybe turn
> all that into more or less a single VMA?

I proposed the topic "A proper API for sparse memory mappings" for the 
bi-weekly MM meeting on September 20, that would also cover exactly that 
use case. :)

-- 
Cheers,

David / dhildenb