Re: [PATCH] mm: always respect QUEUE_FLAG_STABLE_WRITES on the block device

[Jan Kara]

On Fri 05-05-23 09:07:36, Dave Chinner wrote:
> On Thu, May 04, 2023 at 05:16:48PM +0100, Matthew Wilcox wrote:
> > On Thu, May 04, 2023 at 05:55:56PM +0200, Jan Kara wrote:
> > > For bdev address_space that's easy but what Ilya also mentioned is a
> > > problem when 'stable_write' flag gets toggled on the device and in that
> > > case having to propagate the flag update to all the address_space
> > > structures is a nightmare...
> > 
> > We have a number of flags which don't take effect when modified on a
> > block device with a mounted filesystem on it.  For example, modifying
> > the readahead settings do not change existing files, only new ones.
> > Since this flag is only modifiable for debugging purposes, I think I'm
> > OK with it not affecting already-mounted filesystems.  It feels like a
> > decision that reasonable people could disagree on, though.
> 
> I think an address space flag makes sense, because then we don't
> even have to care about the special bdev sb/inode thing -
> folio->mapping will already point at the bdev mapping and so do the
> right thing.
> 
> That is, if the bdev changes stable_write state, it can toggle the
> AS_STABLE_WRITE flag on it's inode->i_mapping straight away and all
> the folios and files pointing to the bdev mapping will change
> behaviour immediately.  Everything else retains the same behaviour
> we have now - the stable_write state is persistent on the superblock
> until the filesystem mount is cycled.

Yeah, I'm fine with this behavior. I just wasn't sure whether Ilya didn't
need the sysfs change to be visible in the filesystem so that was why I
pointed that out. But apparently he doesn't need it.

								Honza
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR