On 12.10.2021 14:43, Matthew Wilcox wrote:
> On Tue, Oct 12, 2021 at 02:02:37PM +0200, Marek Szyprowski wrote:
>> Decrease nr_thps counter in file's mapping to ensure that the page cache
>> won't be dropped excessively on file write access if page has been
>> already splitted.
>>
>> Signed-off-by: Marek Szyprowski <m.szyprowski@xxxxxxxxxxx>
>> Fixes: 09d91cda0e82 ("mm,thp: avoid writes to file with THP in pagecache")
>> Fixes: 06d3eff62d9d ("mm/thp: fix node page state in split_huge_page_to_list()")
>> ---
>> I've analyzed the code a few times but either I missed something or the
>> nr_thps counter is not decremented during the THP split on non-shmem file
>> pages.
> This looks OK to me, but have you tested it? If so, what workload did
> you use? The way you wrote this changelog makes it sound like you only
> read the code and there have been rather too many bugs introduced recently
> that way :-(
Well, indeed I've found it while reading the code. However I've just
tried a test scenario, where one runs a big binary, kernel remaps it
with THPs, then one forces THP split with
/sys/kernel/debug/split_huge_pages. During any further open of that
binary with O_RDWR or O_WRITEONLY kernel drops page cache for it,
because of non-zero thps counter.
Best regards
--
Marek Szyprowski, PhD
Samsung R&D Institute Poland
