Re: [PATCH 1/2] mm/filemap: don't initiate writeback if mapping has no dirty pages

[Konstantin Khlebnikov <khlebnikov@yandex-team.ru>]

On 30.07.2019 17:14, Jan Kara wrote:
> On Tue 23-07-19 11:16:51, Konstantin Khlebnikov wrote:
>> On 23.07.2019 3:52, Andrew Morton wrote:
>>>
>>> (cc linux-fsdevel and Jan)
> 
> Thanks for CC Andrew.
> 
>>> On Mon, 22 Jul 2019 12:36:08 +0300 Konstantin Khlebnikov <khlebnikov@yandex-team.ru> wrote:
>>>
>>>> Functions like filemap_write_and_wait_range() should do nothing if inode
>>>> has no dirty pages or pages currently under writeback. But they anyway
>>>> construct struct writeback_control and this does some atomic operations
>>>> if CONFIG_CGROUP_WRITEBACK=y - on fast path it locks inode->i_lock and
>>>> updates state of writeback ownership, on slow path might be more work.
>>>> Current this path is safely avoided only when inode mapping has no pages.
>>>>
>>>> For example generic_file_read_iter() calls filemap_write_and_wait_range()
>>>> at each O_DIRECT read - pretty hot path.
> 
> Yes, but in common case mapping_needs_writeback() is false for files you do
> direct IO to (exactly the case with no pages in the mapping). So you
> shouldn't see the overhead at all. So which case you really care about?
> 
>>>> This patch skips starting new writeback if mapping has no dirty tags set.
>>>> If writeback is already in progress filemap_write_and_wait_range() will
>>>> wait for it.
>>>>
>>>> ...
>>>>
>>>> --- a/mm/filemap.c
>>>> +++ b/mm/filemap.c
>>>> @@ -408,7 +408,8 @@ int __filemap_fdatawrite_range(struct address_space *mapping, loff_t start,
>>>>    		.range_end = end,
>>>>    	};
>>>> -	if (!mapping_cap_writeback_dirty(mapping))
>>>> +	if (!mapping_cap_writeback_dirty(mapping) ||
>>>> +	    !mapping_tagged(mapping, PAGECACHE_TAG_DIRTY))
>>>>    		return 0;
>>>>    	wbc_attach_fdatawrite_inode(&wbc, mapping->host);
>>>
>>> How does this play with tagged_writepages?  We assume that no tagging
>>> has been performed by any __filemap_fdatawrite_range() caller?
>>>
>>
>> Checking also PAGECACHE_TAG_TOWRITE is cheap but seems redundant.
>>
>> To-write tags are supposed to be a subset of dirty tags:
>> to-write is set only when dirty is set and cleared after starting writeback.
>>
>> Special case set_page_writeback_keepwrite() which does not clear to-write
>> should be for dirty page thus dirty tag is not going to be cleared either.
>> Ext4 calls it after redirty_page_for_writepage()
>> XFS even without clear_page_dirty_for_io()
>>
>> Anyway to-write tag without dirty tag or at clear page is confusing.
> 
> Yeah, TOWRITE tag is intended to be internal to writepages logic so your
> patch is fine in that regard. Overall the patch looks good to me so I'm
> just wondering a bit about the motivation...

In our case file mixes cached pages and O_DIRECT read. Kind of database
were index header is memory mapped while the rest data read via O_DIRECT.
I suppose for sharing index between multiple instances.

On this path we also hit this bug:
https://lore.kernel.org/lkml/156355839560.2063.5265687291430814589.stgit@buzz/
so that's why I've started looking into this code.