Re: f2fs async buffered write patch

[Yangtao Li]

+cc willy@xxxxxxxxxxxxx


On 2023/6/20 4:43, Jens Axboe wrote:

> Hi,
>
> I came across this patch in a news posting:
>
> https://git.kernel.org/pub/scm/linux/kernel/git/jaegeuk/f2fs.git/commit/?h=dev&id=d618126911829523e35a61f4a5a4ad159b1b2c8d
>
> which has me a bit worried. As far as I can tell, all that patch does is
> set FMODE_BUF_WASYNC, and then just hope that the lower layers handle
> the rest?
>
> What happens if iocb->ki_flags & IOCB_NOWAIT is true, and now we do:
>
> generic_perform_write(iocb, from)
> 	...
> 	->write_begin() <- does this block?

Most file systems have not been changed to the iomap method,

and the generic_perform_write method is still used. So it seems

 reasonable to make IOCB_NOWAIT be handled correctly by the

 generic_perform_write function. This means that we need to modify

 several places mentioned by Jens in the generic_perform_write function,

 and need to pass AOP_FLAG_xxx or iocb into write_begin.


I noticed that Matthew Wilcox removed the flags parameter in write_begin

in a previous commit, maybe we can add it back?


Thx

> 	...
> 	->write_end() <- or this one?
> 	...
> 	balance_dirty_pages_ratelimited() <- this one surely does...
>
> If you look just one level down the latter to
> balance_dirty_pages_ratelimited_flags(), you'll even see the 'flags'
> argument documented there.
>
> This looks pretty haphazard and cannot possibly work as-is, so please
> get this reverted until f2fs is converted to iomap, or IOCB_NOWAIT is
> handled by generic_perform_write() and below.