Re: [PATCH] mm/filemap: invalidating pages is still necessary when io with IOCB_NOWAIT

[Jens Axboe]

On 5/23/24 3:11 PM, Matthew Wilcox wrote:
> On Thu, May 23, 2024 at 01:08:02PM -0700, Andrew Morton wrote:
>> On Mon, 13 May 2024 21:23:39 +0800 Liu Wei <liuwei09@xxxxxxxx> wrote:
>>
>>> After commit (6be96d3ad3 fs: return if direct I/O will trigger writeback),
>>> when we issuing AIO with direct I/O and IOCB_NOWAIT on a block device, the
>>> process context will not be blocked.
>>>
>>> However, if the device already has page cache in memory, EAGAIN will be
>>> returned. And even when trying to reissue the AIO with direct I/O and
>>> IOCB_NOWAIT again, we consistently receive EAGAIN.
>>>
>>> Maybe a better way to deal with it: filemap_fdatawrite_range dirty pages
>>> with WB_SYNC_NONE flag, and invalidate_mapping_pages unmapped pages at
>>> the same time.
>>
>> Can't userspace do this?  If EAGAIN, sync the fd and retry the IO?
> 
> I don't think that it can, because the pages will still be there, even
> if now clean?  I think the idea was to punt to a worker thread which
> could sleep and retry without NOWAIT.  But let's see what someone
> involved in this patch has to say about the intent.

Right, the idea is that if you get -EAGAIN, a non-blocking attempt
wasn't possible. You'd need to retry from somewhere where you CAN block.
Any issuer very much can do that, whether it's in-kernel or not.

It'd be somewhat fragile to make assumptions on what can cause the
-EAGAIN and try to rectify them, and then try again with IOCB_NOWAIT.

-- 
Jens Axboe