[PATCH v3 00/35] Compiler-Based Capability- and Locking-Analysis

[Christoph Hellwig]

On Thu, Sep 18, 2025 at 10:45:55AM -0700, Nathan Chancellor wrote:
> On Thu, Sep 18, 2025 at 04:15:11PM +0200, Christoph Hellwig wrote:
> > On Thu, Sep 18, 2025 at 03:59:11PM +0200, Marco Elver wrote:
> > > A Clang version that supports `-Wthread-safety-pointer` and the new
> > > alias-analysis of capability pointers is required (from this version
> > > onwards):
> > > 
> > > 	https://github.com/llvm/llvm-project/commit/b4c98fcbe1504841203e610c351a3227f36c92a4 [3]
> > 
> > There's no chance to make say x86 pre-built binaries for that available?
> 
> I can use my existing kernel.org LLVM [1] build infrastructure to
> generate prebuilt x86 binaries. Just give me a bit to build and upload
> them. You may not be the only developer or maintainer who may want to
> play with this.

That did work, thanks.

I started to play around with that.  For the nvme code adding the
annotations was very simply, and I also started adding trivial
__guarded_by which instantly found issues.

For XFS it was a lot more work and I still see tons of compiler
warnings, which I'm not entirely sure how to address.  Right now I
see three major classes:

1) locks held over loop iterations like:

fs/xfs/xfs_extent_busy.c:573:26: warning: expecting spinlock 'xfs_group_hold(busyp->group)..xg_busy_extents->eb_lock' to be held at start of each loop [-Wthread-safety-analysis]
  573 |                 struct xfs_group        *xg = xfs_group_hold(busyp->group);
      |                                               ^
fs/xfs/xfs_extent_busy.c:577:3: note: spinlock acquired here
  577 |                 spin_lock(&eb->eb_lock);
      |                 ^

This is perfectly find code and needs some annotations, but I can't find
any good example.

2) Locks on returned objects, which can be NULL.  I.e., something
like crossover of __acquire_ret and __cond_acquires

3) Wrappers that take multiple locks conditionally

We have helpers that take different locks in the same object based on the
arguments like xfs_ilock() or those that take the same lock and a variable
number of objects like xfs_dqlockn based on input and sorting.  The
first are just historic and we might want to kill them, but the
sorting of objects to acquire locks in order thing is a pattern in
various places including the VFS, so we'll need some way to annotate it.