Re: [RFC PATCH] kernfs: release kernfs_mutex before the inode allocation

[Minchan Kim <minchan@kernel.org>]

On Wed, Nov 17, 2021 at 07:44:44AM +0100, Greg Kroah-Hartman wrote:
> On Tue, Nov 16, 2021 at 01:36:01PM -0800, Minchan Kim wrote:
> > On Tue, Nov 16, 2021 at 08:49:46PM +0100, Greg Kroah-Hartman wrote:
> > > On Tue, Nov 16, 2021 at 11:43:17AM -0800, Minchan Kim wrote:
> > > > The kernfs implementation has big lock granularity(kernfs_rwsem) so
> > > > every kernfs-based(e.g., sysfs, cgroup, dmabuf) fs are able to compete
> > > > the lock. Thus, if one of userspace goes the sleep under holding
> > > > the lock for a long time, rest of them should wait it. A example is
> > > > the holder goes direct reclaim with the lock since it needs memory
> > > > allocation. Let's fix it at common technique that release the lock
> > > > and then allocate the memory. Fortunately, kernfs looks like have
> > > > an refcount so I hope it's fine.
> > > > 
> > > > Signed-off-by: Minchan Kim <minchan@kernel.org>
> > > > ---
> > > >  fs/kernfs/dir.c             | 14 +++++++++++---
> > > >  fs/kernfs/inode.c           |  2 +-
> > > >  fs/kernfs/kernfs-internal.h |  1 +
> > > >  3 files changed, 13 insertions(+), 4 deletions(-)
> > > 
> > > What workload hits this lock to cause it to be noticable?
> > 
> > A app launching since it was dropping the frame since the
> > latency was too long.
> 
> How does running a program interact with kernfs filesystems?  Which
> one(s)?

A app launching involves dma_buf exports which creates kobject
and add it to the kernfs with down_write - kernfs_add_one.

At the same time in other CPU, a random process was accessing
sysfs and the kernfs_iop_lookup was already hoding the kernfs_rwsem
and ran under direct reclaim patch due to alloc_inode in
kerfs_get_inode.

Therefore, the app is stuck on the lock and lose frames so enduser
sees the jank.

> 
> > > There was a bunch of recent work in this area to make this much more
> > > fine-grained, and the theoritical benchmarks that people created (adding
> > > 10s of thousands of scsi disks at boot time) have gotten better.
> > > 
> > > But in that work, no one could find a real benchmark or use case that
> > > anyone could even notice this type of thing.  What do you have that
> > > shows this?
> > 
> > https://developer.android.com/studio/command-line/perfetto
> > https://perfetto.dev/docs/data-sources/cpu-scheduling
> 
> That is links to a tool, not a test we can run ourselves.
> 
> Or how about the output of that tool?
> 
> > Android has perfetto tracing system and can show where processes
> > were stuck. This case was the lock since holder was in direct reclaim
> > path.
> 
> Reclaim of what?  What is the interaction here with kernfs?  Normally
> this filesystem is not on any "fast paths" that I know of.
> 
> More specifics would be nice :)

I hope it's enough above.