Re: [PATCH -mm -v4 08/21] mm, THP, swap: Support to read a huge swap cluster for swapin a THP

[Huang\, Ying]

Matthew Wilcox <willy@xxxxxxxxxxxxx> writes:

> On Fri, Jun 22, 2018 at 11:51:38AM +0800, Huang, Ying wrote:
>> +++ b/mm/swap_state.c
>> @@ -426,33 +447,37 @@ struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
>>  		/*
>>  		 * call radix_tree_preload() while we can wait.
>>  		 */
>> -		err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
>> +		err = radix_tree_maybe_preload_order(gfp_mask & GFP_KERNEL,
>> +						     compound_order(new_page));
>>  		if (err)
>>  			break;
>
> There's no more preloading in the XArray world, so this can just be dropped.

Sure.

>>  		/*
>>  		 * Swap entry may have been freed since our caller observed it.
>>  		 */
>> +		err = swapcache_prepare(hentry, huge_cluster);
>> +		if (err) {
>>  			radix_tree_preload_end();
>> -			break;
>> +			if (err == -EEXIST) {
>> +				/*
>> +				 * We might race against get_swap_page() and
>> +				 * stumble across a SWAP_HAS_CACHE swap_map
>> +				 * entry whose page has not been brought into
>> +				 * the swapcache yet.
>> +				 */
>> +				cond_resched();
>> +				continue;
>> +			} else if (err == -ENOTDIR) {
>> +				/* huge swap cluster is split under us */
>> +				continue;
>> +			} else		/* swp entry is obsolete ? */
>> +				break;
>
> I'm not entirely happy about -ENOTDIR being overloaded to mean this.
> Maybe we can return a new enum rather than an errno?

Can we use -ESTALE instead?  The "huge swap cluster is split under us"
means the swap entry is kind of "staled".

> Also, I'm not sure that a true/false parameter is the right approach for
> "is this a huge page".  I think we'll have usecases for swap entries which
> are both larger and smaller than PMD_SIZE.

OK.  I can change the interface to number of swap entries style to make
it more flexible.

> I was hoping to encode the swap entry size into the entry; we only need one
> extra bit to do that (no matter the size of the entry).  I detailed the
> encoding scheme here:
>
> https://plus.google.com/117536210417097546339/posts/hvctn17WUZu
>
> (let me know if that doesn't work for you; I'm not very experienced with
> this G+ thing)

The encoding method looks good.  To use it, we need to

- Encode swap entry and size into swap_entry_size
- Call function with swap_entry_size
- Decode swap_entry_size to swap entry and size

It appears that there is no real benefit?

Best Regards,
Huang, Ying