Re: [PATCH v1] mm/gup: adjust stale comment for RCU GUP-fast

[Aneesh Kumar K V <aneesh.kumar@linux.ibm.com>]

On 9/2/22 12:02 PM, David Hildenbrand wrote:
> On 01.09.22 20:35, Yang Shi wrote:
>> On Thu, Sep 1, 2022 at 11:07 AM Peter Xu <peterx@redhat.com> wrote:
>>>
>>> On Thu, Sep 01, 2022 at 10:50:48AM -0700, Yang Shi wrote:
>>>> Yeah, because THP collapse does copy the data before clearing pte. If
>>>> we want to remove pmdp_collapse_flush() by just clearing pmd, we
>>>> should clear *AND* flush pte before copying the data IIRC.
>>>
>>> Yes tlb flush is still needed.  IIUC the generic pmdp_collapse_flush() will
>>> still be working (with the pte level flushing there) but it should just
>>> start to work for all archs, so potentially we could drop the arch-specific
>>> pmdp_collapse_flush()s, mostly the ppc impl.
>>
>> I'm don't know why powperpc needs to have its specific
>> pmdp_collapse_flush() in the first place, not only the mandatory IPI
>> broadcast, but also the specific implementation of pmd tlb flush. But
>> anyway the IPI broadcast could be removed at least IMO.
>>
> 
> pmdp_collapse_flush() is overwritten on book3s only. It either translates
> to radix__pmdp_collapse_flush() or hash__pmdp_collapse_flush().
> 
> 
> radix__pmdp_collapse_flush() has a comment explaining the situation:
> 
> 
> +       /*
> +        * pmdp collapse_flush need to ensure that there are no parallel gup
> +        * walk after this call. This is needed so that we can have stable
> +        * page ref count when collapsing a page. We don't allow a collapse page
> +        * if we have gup taken on the page. We can ensure that by sending IPI
> +        * because gup walk happens with IRQ disabled.
> +        */
> 
> 
> The comment for hash__pmdp_collapse_flush() is a bit more involved:
> 
> 	/*
> 	 * Wait for all pending hash_page to finish. This is needed
> 	 * in case of subpage collapse. When we collapse normal pages
> 	 * to hugepage, we first clear the pmd, then invalidate all
> 	 * the PTE entries. The assumption here is that any low level
> 	 * page fault will see a none pmd and take the slow path that
> 	 * will wait on mmap_lock. But we could very well be in a
> 	 * hash_page with local ptep pointer value. Such a hash page
> 	 * can result in adding new HPTE entries for normal subpages.
> 	 * That means we could be modifying the page content as we
> 	 * copy them to a huge page. So wait for parallel hash_page
> 	 * to finish before invalidating HPTE entries. We can do this
> 	 * by sending an IPI to all the cpus and executing a dummy
> 	 * function there.
> 	 */
> 
> I'm not sure if that implies that the IPI is needed for some other hash-magic.
> 
> Maybe Aneesh can clarify.
> 

We still need the IPI for the hash. Another reason for architecture to override that
function is to help them use the right page size when flushing the TLB.


-aneesh