Re: [PATCH] lib/sort.c: Add _nonatomic() variants with cond_resched()

[Kent Overstreet]

On Wed, Mar 26, 2025 at 11:54:15PM +0800, Kuan-Wei Chiu wrote:
> On Wed, Mar 26, 2025 at 11:26:06AM -0400, Kent Overstreet wrote:
> > Andrew - if you're ok with this patch I'd like to get it in soon as a
> > bugfix, I've been getting quite a few reports on this one.
> > 
> > I don't much care for the naming though, thoughts there?
> > 
> > -- >8 --
> > 
> > bcachefs calls sort() during recovery to sort all keys it found in the
> > journal, and this may be very large - gigabytes on large machines.
> > 
> > This has been causing "task blocked" warnings, so needs a
> > cond_resched().
> > 
> > Cc: Kuan-Wei Chiu <visitorckw@xxxxxxxxx>
> > Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
> > Signed-off-by: Kent Overstreet <kent.overstreet@xxxxxxxxx>
> > ---
> >  include/linux/sort.h | 11 +++++++++++
> >  lib/sort.c           | 46 +++++++++++++++++++++++++++++++++++++++-----
> >  2 files changed, 52 insertions(+), 5 deletions(-)
> > 
> 
> I don't have strong opinions on this, but I recall that UBIFS had a
> similar issue with list_sort(), and they addressed it by calling
> cond_resched() within the compare function. Would that approach be
> simpler and more appropriate than introducing a new API in the library
> code?

That'd be an option, but it would be heavier; sort() has nested loops so
it has a more natural place to put it.

And I'd say the nonatomic scheduling version should likely be the
default, anyways; even if other users aren't hitting the 10 second
warning, going 1 second without scheduling isn't good.

Not going to audit all the existing callers, but we should probably
provide it.