Re: [PATCH 6.15 000/588] 6.15.4-rc2 review

From: Barry K. Nathan
Date: Thu Jun 26 2025 - 05:05:37 EST


On 6/25/25 05:15, Barry K. Nathan wrote:
On 6/25/25 02:08, Holger Hoffstätte wrote:
On 2025-06-25 10:25, Greg Kroah-Hartman wrote:
On Wed, Jun 25, 2025 at 10:00:56AM +0200, Holger Hoffstätte wrote:
(cc: Christian Brauner>

Since 6.15.4-rc1 I noticed that some KDE apps (kded6, kate (the text editor))
started going into a tailspin with 100% per-process CPU.

The symptom is 100% reproducible: open a new file with kate, save empty file,
make changes, save, watch CPU go 100%. perf top shows copy_to_user running wild.

First I tried to reproduce on 6.15.3 - no problem, everything works fine.

After checking the list of patches for 6.15.4 I reverted the anon_inode series
(all 3 for the first attempt) and the problem is gone.

Will try to reduce further & can gladly try additional fixes, but for now
I'd say these patches are not yet suitable for stable.

Does this same issue also happen for you on 6.16-rc3?

Curiously it does *not* happen on 6.16-rc3, so that's good.
I edited/saved several files and everything works as it should.

In 6.15.4-rc the problem occurs (as suspected) with:
anon_inode-use-a-proper-mode-internally.patch aka cfd86ef7e8e7 upstream.

thanks
Holger

For what it's worth, I can confirm this reproduces easily and consistently on Debian trixie's KDE (6.3.5), with either Wayland or X11. It reproduces with kernel 6.15.4-rc2, and with 6.15.3+anon_inode-use-a- proper-mode-internally.patch, but not with vanilla 6.15.3 or with 6.16-rc3.

By the way, my test VM has both GNOME and KDE installed. If I boot one of the affected kernels and log into a GNOME session, I don't get any GNOME processes chewing up CPU the way that some of the KDE processes do. However, if I then start kate within the GNOME session and follow the steps to reproduce (create a new file, save it immediately, type a few characters, save again), kate still starts using 100% CPU.

After some testing and bisecting, I found that "anon_inode: use a proper mode internally" needs to be followed up with "fs: add S_ANON_INODE" (upstream commit 19bbfe7b5fcc) in order to avoid this regression.
--
-Barry K. Nathan <barryn@xxxxxxxxx>