Re: [PATCH 0/3] Transient errors in Direct I/O

[Qu Wenruo <quwenruo.btrfs@gmx.com>]

[-- Attachment #1.1: Type: text/plain, Size: 1766 bytes --]



On 2020/6/11 下午10:11, Goldwyn Rodrigues wrote:
> On 13:31 10/06, Qu Wenruo wrote:
>>
>>
>> On 2020/6/6 上午4:48, Goldwyn Rodrigues wrote:
>>> In current scenarios, for XFS, it would mean that a page invalidation
>>> would end up being a writeback error. So, if iomap returns zero, fall
>>> back to biffered I/O. XFS has never supported fallback to buffered I/O.
>>> I hope it is not "never will" ;)
>>>
>>> With mixed buffered and direct writes in btrfs, the pages may not be
>>> released the extent may be locked in the ordered extents cleanup thread,
>>
>> I'm wondering can we handle this case in a different way.
>>
>> In fact btrfs has its own special handling for invalidating pages.
>> Btrfs will first look for any ordered extents covering the page, finish
>> the ordered extent manually, then invalidate the page.
>>
>> I'm not sure why invalidate_inode_pages2_range() used in dio iomap code
>> does not use the fs specific invalidatepage(), but only do_lander_page()
>> then releasepage().
>>
>> Shouldn'y we btrfs implement the lander_page() to handle ordered extents
>> properly?
>> Or is there any special requirement?
>>
> 
> The problem is aops->launder_page() is called only if PG_Dirty is
> set. In this case it is not because we just performed a writeback.

For the dio iomap code, before btrfs_finish_ordered_io(), the pages in
ordered ranges are still dirty.
So launder_page() here can still get triggered to finish the ordered extent.

> 
> Also, it is not just ordered ordered extent writeback which may lock
> the extent, a buffered read can lock as well.
> 
That's right.
But at least that would greatly reduce the possibility for btrfs to fall
back to buffered IO, right?

Thanks,
Qu


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 488 bytes --]