[Intel-gfx] Regression on linux-next (next-20231107)

[Krister Johansen]

Hi Chaitanya,

On Thu, Nov 09, 2023 at 05:00:09PM +0000, Borah, Chaitanya Kumar wrote:
> Hello Krister,
>  
> Hope you are doing well. I am Chaitanya from the linux graphics team in Intel.
>  
> This mail is regarding a regression we are seeing in our CI runs[1] for some machines (dg2 and adl-p) on linux-next  repository.
> 
> Since the version next-20231107 [2], we are seeing the following error
> ```````````````````````````````````````````````````````````````````````````````
> <4>[   32.015910] stack segment: 0000 [#1] PREEMPT SMP NOPTI
> <4>[   32.021048] CPU: 15 PID: 766 Comm: fusermount Not tainted 6.6.0-next-20231107-next-20231107-g5cd631a52568+ #1
> <4>[   32.031135] Hardware name: Intel Corporation Raptor Lake Client Platform/RPL-S ADP-S DDR5 UDIMM CRB, BIOS RPLSFWI1.R00.4221.A00.2305271351 05/27/2023
> <4>[   32.044657] RIP: 0010:fuse_evict_inode+0x61/0x150 [fuse]
> `````````````````````````````````````````````````````````````````````````````````
> 
> Details log can be found in [3].
> 
> After bisecting the tree, the following patch [4] seems to be the first "bad" commit
> 
>  `````````````````````````````````````````````````````````````````````````````````````````````````````````
> 513dfacefd712bcbfab64e1a9c9c3e0d51c2dca5 is the first bad commit
> commit 513dfacefd712bcbfab64e1a9c9c3e0d51c2dca5
> Author: Krister Johansen kjlx at templeofstupid.com
> Date:   Fri Nov 3 10:39:47 2023 -0700
> 
>     fuse: share lookup state between submount and its parent
> 
>     Fuse submounts do not perform a lookup for the nodeid that they inherit
>     from their parent.  Instead, the code decrements the nlookup on the
>     submount's fuse_inode when it is instantiated, and no forget is
>     performed when a submount root is evicted.
> 
>     Trouble arises when the submount's parent is evicted despite the
>     submount itself being in use.  In this author's case, the submount was
>     in a container and deatched from the initial mount namespace via a
>     MNT_DEATCH operation.  When memory pressure triggered the shrinker, the
>     inode from the parent was evicted, which triggered enough forgets to
>     render the submount's nodeid invalid.
> 
>     Since submounts should still function, even if their parent goes away,
>     solve this problem by sharing refcounted state between the parent and
>     its submount.  When all of the references on this shared state reach
>     zero, it's safe to forget the final lookup of the fuse nodeid.
> 
>  `````````````````````````````````````````````````````````````````````````````````````````````````````````
>  
> We also verified that if we revert the patch the issue is not seen.
> 
> Could you please check why the patch causes this regression and provide a fix if necessary?

Apologies for the inconvenience.  I've reproduced the problem, tested a
fix, and am in the process of preparing patches to send to Miklos.  I'll
cc the people on this e-mail in that thread.

> [3] http://gfx-ci.igk.intel.com/tree/linux-next/next-20231109/bat-dg2-14/boot0.txt

This link didn't resolve in DNS when I tried to access it.  I needed to
use intel-gfx-ci.01.org as the hostname instead.

Thanks,

-K