Re: [PATCH] mm/huge_memory: Avoid PMD-size page cache if needed

[Ryan Roberts]

On 11/07/2024 21:46, Matthew Wilcox wrote:
> On Thu, Jul 11, 2024 at 08:48:40PM +1000, Gavin Shan wrote:
>> +++ b/mm/huge_memory.c
>> @@ -136,7 +136,8 @@ unsigned long __thp_vma_allowable_orders(struct vm_area_struct *vma,
>>  
>>  		while (orders) {
>>  			addr = vma->vm_end - (PAGE_SIZE << order);
>> -			if (thp_vma_suitable_order(vma, addr, order))
>> +			if (!(vma->vm_file && order > MAX_PAGECACHE_ORDER) &&
>> +			    thp_vma_suitable_order(vma, addr, order))
>>  				break;
> 
> Why does 'orders' even contain potential orders that are larger than
> MAX_PAGECACHE_ORDER?
> 
> We do this at the top:
> 
>         orders &= vma_is_anonymous(vma) ?
>                         THP_ORDERS_ALL_ANON : THP_ORDERS_ALL_FILE;
> 
> include/linux/huge_mm.h:#define THP_ORDERS_ALL_FILE     (BIT(PMD_ORDER) | BIT(PUD_ORDER))
> 
> ... and that seems very wrong.  We support all kinds of orders for
> files, not just PMD order.  We don't support PUD order at all.
> 
> What the hell is going on here?

Just to try to justify this a little, it was my perspective when adding (anon)
mTHP that memory was either anon or file; Anything that populated vma->vm_file
was file, including shmem, DAX, etc. Before my change THP could install PMD size
mappings for anon, and PMD or PUD size mappings for file memory (but yes, PUD
was only really applicable to DAX in practice, I believe).

I agree it would be good to clean this up, but I don't think the current code is
quite as mad as you're implying, Matthew?