Re: [PATCH] arm64/gcs: Don't call gcs_free() during flush_gcs()

[Will Deacon]

On Thu, Jun 12, 2025 at 03:51:19PM +0100, Mark Brown wrote:
> On Thu, Jun 12, 2025 at 03:47:44PM +0100, Catalin Marinas wrote:
> > On Thu, Jun 12, 2025 at 12:40:42PM +0100, Mark Brown wrote:
> > > On Wed, Jun 11, 2025 at 06:34:15PM +0100, Catalin Marinas wrote:
> 
> > > > Another caller of gcs_free() is deactivate_mm(). It's not clear to me
> > > > when we need to free the shadow stack on this path. On the exit_mm()
> > > > path for example we have mmput() -> exit_mmap() that takes care of
> > > > unmapping everything. Similarly on the exec_mmap() path.
> 
> > > We need that one to clean up the GCS for threads that had it allocated
> > > for compatibility, you can see the leak that results without it easily
> > > with the glibc testsuite (or anything else that does threads, the glibc
> > > tests just spot it).  Most of the checking for arch_release_task_struct()
> > > is verifying that deactivate_mm() is guaranteed to be called eveywhere
> > > it's relevant, I need to page that back in.
> 
> > Makes sense. I think we should only keep gcs_free() in one place,
> > ideally deactivate_mm() as that's more related to mm rather than the
> > task_struct.
> 
> Yes, me too - I just need to double check.

Having looking a little at the code, I think that
arch_release_task_struct() might be better than deactivate_mm(). The
latter takes an 'mm' parameter which we ignore but I think happens to
be 'current->mm'and so things work. Given that, and that we don't do any
GCS management on the activate_mm() path, freeing the GCS in the
task-centric functions makes more sense to me.

Will