Re: [PATCH v2 0/9] Free user PTE page table pages

[David Hildenbrand <david@redhat.com>]

On 01.09.21 18:07, Jason Gunthorpe wrote:
> On Wed, Sep 01, 2021 at 02:32:08PM +0200, David Hildenbrand wrote:
>   
>> b) pmd_trans_unstable_or_pte_try_get() and friends are really ugly.
> 
> I suspect the good API here is really more like:

That was my exactly my first idea and I tried to rework the code for 
roughly 2 days and failed.

Especially in pagefault logic, we temporarily unmap/unlock to map/lock 
again later and don't want the page table to just vanish.

I think I met similar cases when allocating a page table and not wanting 
it to vanish and not wanting to map/lock it. But I don't recall all the 
corner cases: it didn't work for me.

> 
>    ptep = pte_try_map(pmdp, &pmd_value)
>    if (!ptep) {
>       // pmd_value is guarenteed to not be a PTE table pointer.
>       if (pmd_XXX(pmd_value))
>    }
> 
> Ie the core code will do whatever stuff, including the THP data race
> avoidance, to either return the next level page table or the value of
> a pmd that is not a enxt level page table. Callers are much clearer in
> this way.
> 
> Eg this is a fairly representative sample user:
> 
> static int smaps_pte_range(pmd_t *pmd, unsigned long addr, unsigned long end,
> 			   struct mm_walk *walk)
> {
> 	if (pmd_trans_unstable(pmd))
> 		goto out;
> 	pte = pte_offset_map_lock(vma->vm_mm, pmd, addr, &ptl);
> 
> And it is obviously pretty easy to integrate any refcount into
> pte_try_map and pte_unmap as in my other email.

It didn't work when I tried.

-- 
Thanks,

David / dhildenb