Re: [RFC 3/3] mm: make swapin readahead to improve thp collapse rate

[Rik van Riel <riel@redhat.com>]

On 06/16/2015 05:15 PM, Andrew Morton wrote:
> On Sun, 14 Jun 2015 18:04:43 +0300 Ebru Akagunduz <ebru.akagunduz@gmail.com> wrote:
> 
>> This patch makes swapin readahead to improve thp collapse rate.
>> When khugepaged scanned pages, there can be a few of the pages
>> in swap area.
>>
>> With the patch THP can collapse 4kB pages into a THP when
>> there are up to max_ptes_swap swap ptes in a 2MB range.
>>
>> The patch was tested with a test program that allocates
>> 800MB of memory, writes to it, and then sleeps. I force
>> the system to swap out all. Afterwards, the test program
>> touches the area by writing, it skips a page in each
>> 20 pages of the area.
>>
>> Without the patch, system did not swap in readahead.
>> THP rate was %47 of the program of the memory, it
>> did not change over time.
>>
>> With this patch, after 10 minutes of waiting khugepaged had
>> collapsed %99 of the program's memory.
>>
>> ...
>>
>> +/*
>> + * Bring missing pages in from swap, to complete THP collapse.
>> + * Only done if khugepaged_scan_pmd believes it is worthwhile.
>> + *
>> + * Called and returns without pte mapped or spinlocks held,
>> + * but with mmap_sem held to protect against vma changes.
>> + */
>> +
>> +static void __collapse_huge_page_swapin(struct mm_struct *mm,
>> +					struct vm_area_struct *vma,
>> +					unsigned long address, pmd_t *pmd,
>> +					pte_t *pte)
>> +{
>> +	unsigned long _address;
>> +	pte_t pteval = *pte;
>> +	int swap_pte = 0;
>> +
>> +	pte = pte_offset_map(pmd, address);
>> +	for (_address = address; _address < address + HPAGE_PMD_NR*PAGE_SIZE;
>> +	     pte++, _address += PAGE_SIZE) {
>> +		pteval = *pte;
>> +		if (is_swap_pte(pteval)) {
>> +			swap_pte++;
>> +			do_swap_page(mm, vma, _address, pte, pmd, 0x0, pteval);
>> +			/* pte is unmapped now, we need to map it */
>> +			pte = pte_offset_map(pmd, _address);
>> +		}
>> +	}
>> +	pte--;
>> +	pte_unmap(pte);
>> +	trace_mm_collapse_huge_page_swapin(mm, vma->vm_start, swap_pte);
>> +}
> 
> This is doing a series of synchronous reads.  That will be sloooow on
> spinning disks.
>
> This function should be significantly faster if it first gets all the
> necessary I/O underway.  I don't think we have a function which exactly
> does this.  Perhaps generalise swapin_readahead() or open-code
> something like

Looking at do_swap_page() and __lock_page_or_retry(), I guess
there already is a way to do the above.

Passing a "flags" of FAULT_FLAG_ALLOW_RETRY|FAULT_FLAG_RETRY_NOWAIT
to do_swap_page() should result in do_swap_page() returning with
the pte unmapped and the mmap_sem still held if the page was not
immediately available to map into the pte (trylock_page succeeds).

Ebru, can you try passing the above as the flags argument to
do_swap_page(), and see what happens?

-- 
All rights reversed