Re: [RFC v3 8/8] ext4: Remove the logic to trim inode PAs

From: Jan Kara
Date: Thu Oct 06 2022 - 05:00:18 EST

On Thu 06-10-22 12:25:00, Ojaswin Mujoo wrote:
> On Thu, Sep 29, 2022 at 02:53:11PM +0200, Jan Kara wrote:
> > On Tue 27-09-22 14:46:48, Ojaswin Mujoo wrote:
> > > Earlier, inode PAs were stored in a linked list. This caused a need to
> > > periodically trim the list down inorder to avoid growing it to a very
> > > large size, as this would severly affect performance during list
> > > iteration.
> > >
> > > Recent patches changed this list to an rbtree, and since the tree scales
> > > up much better, we no longer need to have the trim functionality, hence
> > > remove it.
> > >
> > > Signed-off-by: Ojaswin Mujoo <ojaswin@xxxxxxxxxxxxx>
> > > Reviewed-by: Ritesh Harjani (IBM) <ritesh.list@xxxxxxxxx>
> >
> > I'm kind of wondering: Now there won't be performance issues with much
> > more inode PAs but probably we don't want to let them grow completely out
> > of control? E.g. I can imagine that if we'd have 1 billion of inode PAs
> > attached to an inode, things would get wonky both in terms of memory
> > consumption and also in terms of CPU time spent for the cases where we
> > still do iterate all of the PAs... Is there anything which keeps inode PAs
> > reasonably bounded?
> >
> > Honza
> >
> Hi Jan,
> Sorry for the delay in response, I was on leave for the last few days.
> So as per my understanding, after this patch, the only path where we
> would need to traverse all the PAs is the ext4_discard_preallocations()
> call where we discard all the PAs of an inode one by one (eg when
> closing the file etc). Such a discard is a colder path as we don't
> usually expect to do it as often as say allocating blocks to an inode.
> Originally, the limit was added in this patch [1] because of the time
> lost in O(N) traversal in the allocation path (ext4_mb_use_preallocated
> and ext4_mb_normalize_request). Since the rbtree addressed this
> scalability issue we had decided to remove the trim logic in this
> patchset.
> [1]

I agree the O(N) traversal is not in any performance sensitive path.

> That being said, I do agree that there should be some way to limit the
> PAs from taking up an unreasonable amount of buddy space, memory and CPU
> cycles in use cases like database files and disk files of long running
> VMs. Previously the limit was 512 PAs per inode and trim was happening
> in an LRU fashion, which is not very straightforward to implement in
> trees.
> Another approach is rather than having a hard limit, we can throttle the
> PAs based on some parameter like total active PAs in FS or FSUtil% of
> the PAs but we might need to take care of fairness so one inode is not
> holding all the PAs while others get throttled.
> Anyways, I think the trimming part would need some brainstorming to get
> right so just wondering if we could keep that as part of a separate
> patchset and remove the trimming logic for now since rbtree has
> addressed the scalability concerns in allocation path.

I agree the fact it took until 2020 for someone to notice inode PAs can
be cumulating enough for full scan to matter on block allocation means that
this is not a pressing issue. So I'm OK postponing it for now since I also
don't have a great idea how to best trim excessive preallocations.

Jan Kara <jack@xxxxxxxx>