Re: + compaction-add-def_blk_aops-migrate-function-for-memory-compaction.patch added to -mm tree

[Michal Hocko]

On Thu 09-03-17 13:15:07, Jan Kara wrote:
> On Thu 09-03-17 08:58:45, Vlastimil Babka wrote:
> > On 03/09/2017 12:55 AM, akpm@xxxxxxxxxxxxxxxxxxxx wrote:
> > > 
> > > The patch titled
> > >      Subject: compaction: add def_blk_aops migrate function for memory compaction
> > > has been added to the -mm tree.  Its filename is
> > >      compaction-add-def_blk_aops-migrate-function-for-memory-compaction.patch
> > > 
> > > This patch should soon appear at
> > >     http://ozlabs.org/~akpm/mmots/broken-out/compaction-add-def_blk_aops-migrate-function-for-memory-compaction.patch
> > > and later at
> > >     http://ozlabs.org/~akpm/mmotm/broken-out/compaction-add-def_blk_aops-migrate-function-for-memory-compaction.patch
> > > 
> > > Before you just go and hit "reply", please:
> > >    a) Consider who else should be cc'ed
> > >    b) Prefer to cc a suitable mailing list as well
> > >    c) Ideally: find the original patch on the mailing list and do a
> > >       reply-to-all to that, adding suitable additional cc's
> > > 
> > > *** Remember to use Documentation/SubmitChecklist when testing your code ***
> > > 
> > > The -mm tree is included into linux-next and is updated
> > > there every 3-4 working days
> > > 
> > > ------------------------------------------------------
> > > From: zhouxianrong <zhouxianrong@xxxxxxxxxx>
> > > Subject: compaction: add def_blk_aops migrate function for memory compaction
> > 
> > That's not really a mm/compaction patch, but a block layer/migration patch. I
> > don't know internals of those so well, so I added some CC's.
> 
> Thanks!
> 
> > > The reason for doing this is based on two factors.
> > > 
> > > 1. larg file read/write operations with order 0 can fragmentize
> > >    memory rapidly.
> > > 
> > > 2. when a special filesystem does not supply migratepage callback,
> > >    kernel would fallback to default function fallback_migrate_page.
> > >    but fallback_migrate_page could not migrate diry page nicely;
> > >    specially kcompactd with MIGRATE_SYNC_LIGHT could not migrate
> > >    diry pages due to this until clear_page_dirty_for_io in some
> > >    procedure. i think it is not suitable here in this scenario.
> > >    for dirty pages we should migrate it rather than skip or writeout
> > >    it in kcomapctd with MIGRATE_SYNC_LIGHT. i think this problem is
> > >    for all filesystem without migratepage not only for block device fs.
> > > 
> > > So for compaction under large file writing supply migratepage for
> > > def_blk_aops.
> > 
> > Is this really safe to do? buffer_migrate_page() has some assumptions listed in
> > its comment (and maybe more that are not listed). Do we know it's safe to use it
> > for all def_blk_aops users?
> 
> So this is definitely *not* OK for ext4 which is using def_blk_aops. Ext4
> will create temporary buffer heads that alias buffer heads attached to the
> block device page when committing a transaction (see fs/jbd2/journal.c:
> jbd2_journal_write_metadata_buffer()) and these will point to the original
> page.
> 
> Now I'm not saying ext4 / jbd2 cannot be modified so that it somehow makes
> it clear to the page migration code it should stay away from the page.
> However it demostrates that this patch is not safe as is...

Which would suggest to enahnce the documentation of buffer_migrate_page.
-- 
Michal Hocko
SUSE Labs