Re: [RFC 2/6] mm/migrate_pages: split unmap_and_move() to _unmap() and _move()

[Yang Shi <shy828301@gmail.com>]

On Tue, Sep 27, 2022 at 7:57 PM John Hubbard <jhubbard@nvidia.com> wrote:
>
> On 9/27/22 19:14, Yang Shi wrote:
> > IIRC, the writeback may not call generic_writepages. On my ext4
> > filesystem, the writeback call stack looks like:
> >
> > @[
> >     ext4_writepages+1
> >     do_writepages+191
> >     __writeback_single_inode+65
> >     writeback_sb_inodes+477
> >     __writeback_inodes_wb+76
> >     wb_writeback+457
> >     wb_workfn+680
> >     process_one_work+485
> >     worker_thread+80
> >     kthread+231
> >     ret_from_fork+34
> > ]: 2
> >
>
> Sure, that's fine for ext4, in that particular case, but
>
> a) not all filesystems have .writepages hooked up. That's why
> do_writepages() checks for .writepages(), and falls back to
> .writepage():
>
>         if (mapping->a_ops->writepages)
>                 ret = mapping->a_ops->writepages(mapping, wbc);
>         else
>                 ret = generic_writepages(mapping, wbc);
>
> , and
>
> b) there are also a lot of places and ways to invoke writebacks. There
> are periodic writebacks, and there are data integrity (WB_SYNC_ALL)
> writebacks, and other places where a page needs to be cleaned--so, a lot
> of call sites. Some of which will land on a .writepage() sometimes, even
> now.
>
> For just one example, I see migrate.c's writeout() function directly
> calling writepage():
>
>         rc = mapping->a_ops->writepage(&folio->page, &wbc);

Thanks, John. You are right. I think "deprecated" is inaccurate,
.writepage is actually no longer used in memory reclaim context, but
it is still used for other contexts.

Anyway I think what we tried to figure out in the first place is
whether it is possible that filesystem writeback have dead lock with
the new batch migration or not. I think the conclusion didn't change.

>
>
> thanks,
>
> --
> John Hubbard
> NVIDIA
>