Re: [PATCH 8/9] mm/hugetlb: convert hugetlb_wp() to take in a folio

Re: [PATCH 8/9] mm/hugetlb: convert hugetlb_wp() to take in a folio

[Gerald Schaefer]

On Thu, 19 Jan 2023 13:14:45 -0800
Sidhartha Kumar <sidhartha.kumar@oracle.com> wrote:

[...]
>  	page = pte_page(entry);
> -	if (page != pagecache_page)
> +	if (page_folio(page) != pagecache_folio)
>  		if (!trylock_page(page)) {
>  			need_wait_lock = 1;
>  			goto out_ptl;
>  		}
>  
> -	get_page(page);
> +	folio_get(pagecache_folio);
>  

We get a kernel crash on s390 in mprotect testcase from libhugetlbfs
testsuite, starting with next-20230120, bisected to this commit.

We get here with pagecache_folio == NULL, and crash in folio_get().
It doesn´t seem right to replace the get_page() with folio_get()
here, the matching put_page() at out_put_page: also wasn't changed
correspondingly. Also, pagecache_folio == NULL seems to be a valid
case here, on all architectures.

Reverting this folio_get() to get_page() fixes the crash. Not sure
though if I missed something. I think you only want to replace
pagecache_page with pagecache_folio, like in the rest of the commit,
and not page -> pagecache_folio for this get_page().

Re: [PATCH 8/9] mm/hugetlb: convert hugetlb_wp() to take in a folio

[Sidhartha Kumar]

On 1/24/23 7:23 AM, Gerald Schaefer wrote:
> On Thu, 19 Jan 2023 13:14:45 -0800
> Sidhartha Kumar <sidhartha.kumar@oracle.com> wrote:
> 
> [...]
>>   	page = pte_page(entry);
>> -	if (page != pagecache_page)
>> +	if (page_folio(page) != pagecache_folio)
>>   		if (!trylock_page(page)) {
>>   			need_wait_lock = 1;
>>   			goto out_ptl;
>>   		}
>>   
>> -	get_page(page);
>> +	folio_get(pagecache_folio);
>>   
> 
> We get a kernel crash on s390 in mprotect testcase from libhugetlbfs
> testsuite, starting with next-20230120, bisected to this commit.
> 
> We get here with pagecache_folio == NULL, and crash in folio_get().
> It doesn´t seem right to replace the get_page() with folio_get()
> here, the matching put_page() at out_put_page: also wasn't changed
> correspondingly. Also, pagecache_folio == NULL seems to be a valid
> case here, on all architectures.
> 
> Reverting this folio_get() to get_page() fixes the crash. Not sure
> though if I missed something. I think you only want to replace
> pagecache_page with pagecache_folio, like in the rest of the commit,
> and not page -> pagecache_folio for this get_page().
Ya that get_page(page) line should have stayed how it was before as 
pagecache_folio is replacing instances of pagecache_page. Thanks for 
catching this, I'll fix this change in a v2 of this patch series.

Thanks,
Sidhartha Kumar