Re: [PATCH] drm/ttm: Do not put non-struct page memory into PUD/PMDs

["Thomas Hellström" <thomas.hellstrom@linux.intel.com>]

On 10/19/21 20:52, Jason Gunthorpe wrote:
> On Tue, Oct 19, 2021 at 08:49:29PM +0200, Thomas Hellström wrote:
>> Hi,
>>
>> On 10/19/21 20:21, Jason Gunthorpe wrote:
>>> PUD and PMD entries do not have a special bit.
>>>
>>> get_user_pages_fast() considers any page that passed pmd_huge() as
>>> usable:
>>>
>>> 		if (unlikely(pmd_trans_huge(pmd) || pmd_huge(pmd) ||
>>> 			     pmd_devmap(pmd))) {
>>>
>>> And vmf_insert_pfn_pmd_prot() unconditionally sets
>>>
>>> 	entry = pmd_mkhuge(pfn_t_pmd(pfn, prot));
>>>
>>> eg on x86 the page will be _PAGE_PRESENT | PAGE_PSE.
>>>
>>> As such gup_huge_pmd() will try to deref a struct page:
>>>
>>> 	head = try_grab_compound_head(pmd_page(orig), refs, flags);
>>>
>>> and thus crash.
>>>
>>> Prevent this by never using IO memory with vmf_insert_pfn_pud/pmd_prot().
>> Actually I think if fast gup will break even page backed memory since the
>> backing drivers don't assume anybody else takes a refcount / pincount.
>> Normal pages have PTE_SPECIAL and VM_PFNMAP to block that.
> Erk, yes, that is even worse.
>
>> (Side note I was recommended to introduce a PTE_HUGESPECIAL bit for
>> this and basically had  a patch ready but got scared off when trying
>> to handle 64-bit PTE atomic updates on x86-32)
> Right, a PMD_SPECIAL bit is needed to make this work.
Yes, PMD_SPECIAL it was :)
>
>> It might be that we (Intel) try to resurrect this code using
>> PTE_HUGESPECIAL in the near future for i915, but until that, I think
>> it's the safest option to disable it completely.
> Okay, do you want a patch to just delete this function?

That'd be great.

Thanks,

Thomas


>
> Jason