Re: [PATCH v2 2/2] xfs: kick extra large ioends to completion workqueue

[Matthew Wilcox <willy@infradead.org>]

On Thu, May 06, 2021 at 12:34:06PM -0700, Darrick J. Wong wrote:
> On Tue, Oct 06, 2020 at 03:07:20PM +0100, Matthew Wilcox wrote:
> > On Mon, Oct 05, 2020 at 08:55:37PM -0700, Darrick J. Wong wrote:
> > > On Mon, Oct 05, 2020 at 11:21:02AM -0400, Brian Foster wrote:
> > > > We've had reports of soft lockup warnings in the iomap ioend
> > > > completion path due to very large bios and/or bio chains. Divert any
> > > > ioends with 256k or more pages to process to the workqueue so
> > > > completion occurs in non-atomic context and can reschedule to avoid
> > > > soft lockup warnings.
> > > 
> > > Hmmmm... is there any way we can just make end_page_writeback faster?
> > 
> > There are ways to make it faster.  I don't know if they're a "just"
> > solution ...
> > 
> > 1. We can use THPs.  That will reduce the number of pages being operated
> > on.  I hear somebody might have a patch set for that.  Incidentally,
> > this patch set will clash with the THP patchset, so one of us is going to
> > have to rebase on the other's work.  Not a complaint, just acknowledging
> > that some coordination will be needed for the 5.11 merge window.
> 
> How far off is this, anyway?  I assume it's in line behind the folio
> series?

Right.  The folio series found all kinds of fun places where the
accounting was wrong (eg accounting for an N-page I/O as a single page),
so the THP work is all renamed folio now.  The folio patchset I posted
yesterday [1] is _most_ of what is necessary from an XFS point of view.
There's probably another three dozen mm patches to actually enable
multi-page folios after that, and a lot more patches to optimise the
mm/vfs for multi-page folios, but your side of things is almost all
published and reviewable.

[1] https://lore.kernel.org/linux-fsdevel/20210505150628.111735-1-willy@infradead.org/

> > 2. We could create end_writeback_pages(struct pagevec *pvec) which
> > calls a new test_clear_writeback_pages(pvec).  That could amortise
> > taking the memcg lock and finding the lruvec and taking the mapping
> > lock -- assuming these pages are sufficiently virtually contiguous.
> > It can definitely amortise all the statistics updates.
> 
> /me kinda wonders if THPs arent the better solution for people who want
> to run large ios.

Yes, definitely.  It does rather depend on their usage patterns,
but if they're working on a file-contiguous chunk of memory, this
can help a lot.

> > 3. We can make wake_up_page(page, PG_writeback); more efficient.  If
> > you can produce this situation on demand, I had a patch for that which
> > languished due to lack of interest.
> 
> I can (well, someone can) so I'll talk to you internally about their
> seeekret reproducer.

Fantastic.  If it's something that needs to get backported to a
stable-ABI kernel ... this isn't going to be a viable solution.