Message-ID: <aLZtnqTXEpEBpb7z@dread.disaster.area>
Date: Tue, 2 Sep 2025 14:07:58 +1000
From: Dave Chinner <david@...morbit.com>
To: Josef Bacik <josef@...icpanda.com>
Cc: linux-fsdevel@...r.kernel.org, linux-btrfs@...r.kernel.org,
	kernel-team@...com, linux-ext4@...r.kernel.org,
	linux-xfs@...r.kernel.org, brauner@...nel.org,
	viro@...iv.linux.org.uk, amir73il@...il.com
Subject: Re: [PATCH v2 16/54] fs: delete the inode from the LRU list on lookup

On Thu, Aug 28, 2025 at 07:42:25AM -0400, Josef Bacik wrote:
> On Thu, Aug 28, 2025 at 07:46:56AM +1000, Dave Chinner wrote:
> > On Tue, Aug 26, 2025 at 11:39:16AM -0400, Josef Bacik wrote:
> > > When we move to holding a full reference on the inode when it is on an
> > > LRU list we need to have a mechanism to re-run the LRU add logic. The
> > > use case for this is btrfs's snapshot delete, we will lookup all the
> > > inodes and try to drop them, but if they're on the LRU we will not call
> > > ->drop_inode() because their refcount will be elevated, so we won't know
> > > that we need to drop the inode.
> > > 
> > > Fix this by simply removing the inode from it's respective LRU list when
> > > we grab a reference to it in a way that we have active users.  This will
> > > ensure that the logic to add the inode to the LRU or drop the inode will
> > > be run on the final iput from the user.
> > > 
> > > Signed-off-by: Josef Bacik <josef@...icpanda.com>
> > 
> > Have you benchmarked this for scalability?
> > 
> > The whole point of lazy LRU removal was to remove LRU lock
> > contention from the hot lookup path. I suspect that putting the LRU
> > locks back inside the lookup path is going to cause performance
> > regressions...
> > 
> > FWIW, why do we even need the inode LRU anymore?
> > 
> > We certainly don't need it anymore to keep the working set in memory
> > because that's what the dentry cache LRU does (i.e. by pinning a
> > reference to the inode whilst the dentry is active).
> > 
> > And with the introduction of the cached inode list, we don't need
> > the inode LRU to track  unreferenced dirty inodes around whilst
> > they hang out on writeback lists. The inodes on the writeback lists
> > are now referenced and tracked on the cached inode list, so they
> > don't need special hooks in the mm/ code to handle the special
> > transition from "unreferenced writeback" to "unreferenced LRU"
> > anymore, they can just be dropped from the cached inode list....
> > 
> > So rather than jumping through hoops to maintain an LRU we likely
> > don't actually need and is likely to re-introduce old scalability
> > issues, why not remove it completely?
> 
> That's next on the list, but we're already at 54 patches.  This won't be a hot
> path, we're not going to consistently find inodes on the LRU to remove.

IME, there are some workloads that hit the inode cache hard
(typically anything that has a large set of working inodes and
memory pressure is causing reclaim to run all the time).  In these
cases, we are finding inodes on the inode cache it's because we've
missed the dentry cache (due to reclaim) and so the inode we hit is
unreferenced and on the LRU. i.e. if we don't have lazy LRU removal,
when we hit the inode cache we'll also typically hit the LRU....

> My rough plans are
> 
> 1. Get this series merged.
> 2. Let it bake and see if any issues arise.
> 3. Remove the inode LRU completely.
> 4. Remove the i_hash and use an xarray for inode lookups.
>
> The inode LRU removal is going to be a big change,

When I last looked at it, it wasn't a particularly big code change
at all. It was just that mm/ depended on the LRU existing that
prevented it from being easily removable. You've addressed that
dependency by adding the cached inode list for inodes with populated
address spaces....

> and I want it to be separate
> from this work from the LRU work in case we find that we do really need the LRU.
> If that turns out to be the case then we can revisit if this is a scalability
> issue.  Thanks,

Understood, but I would prefer to do that the other way around.

i.e. rather than add complexity and potential scalability issues to
the LRU management on the way to removing the LRU at a later date,
we remove the LRU at the earliest possible opportunity.

If we have any sort of perf regression caused by the LRU removal, we
can address those cases via temporary residence on the new cached
inode list...

-Dave.
-- 
Dave Chinner
david@...morbit.com

Powered by blists - more mailing lists