Re: [PATCH RFC 2/2] fs: fully sync all fses even for an emergency sync

[Christian Brauner <brauner@kernel.org>]

On Tue, Nov 04, 2025 at 09:39:14AM +0100, Jan Kara wrote:
> On Tue 04-11-25 09:28:27, Jan Kara wrote:
> > On Tue 04-11-25 07:25:06, Qu Wenruo wrote:
> > > 
> > > 
> > > 在 2025/11/3 22:26, Christoph Hellwig 写道:
> > > > The emergency sync being non-blocking goes back to day 1.  I think the
> > > > idea behind it is to not lock up a already messed up system by
> > > > blocking forever, even if it is in workqueue.  Changing this feels
> > > > a bit risky to me.
> > > 
> > > Considering everything is already done in task context (baked by the global
> > > per-cpu workqueue), it at least won't block anything else.
> > > 
> > > And I'd say if the fs is already screwed up and hanging, the
> > > sync_inodes_one_sb() call are more likely to hang than the final sync_fs()
> > > call.
> > 
> > Well, but notice that sync_inodes_one_sb() is always called with wait == 0
> > from do_sync_work() exactly to skip inodes already marked as under
> > writeback, locked pages or pages under writeback as waiting for these has
> > high chances of locking up. Suddently calling sync_fs_one_sb() with wait ==
> > 1 can change things. That being said for ext4 the chances of locking up
> > ext4_sync_fs() with wait == 1 after sync_fs_one_sb() managed to do
> > non-trivial work are fairly minimal so I don't have strong objections
> > myself.
> 
> Ah, ok, now I've checked the code and read patch 1 in this thread. Indeed
> sync_inodes_one_sb() ignores the wait parameter and waits for everything.
> Given we've been running like this for over 10 years and nobody complained
> I agree calling sync_fs_one_sb() with wait == 1 is worth trying if it makes
> life better for btrfs.

Agreed. But emergency_sync() is really just a best-effort thing and I'm
rather weary accumulating complexity for it otherwise.