Re: [PATCH 2/4] mm: move per-vma lock into vm_area_struct
From: Suren Baghdasaryan
Date: Tue Nov 12 2024 - 12:01:26 EST
On Tue, Nov 12, 2024 at 8:08 AM Suren Baghdasaryan <surenb@xxxxxxxxxx> wrote:
>
> On Tue, Nov 12, 2024 at 7:57 AM Vlastimil Babka <vbabka@xxxxxxx> wrote:
> >
> > On 11/11/24 21:55, Suren Baghdasaryan wrote:
> > > @@ -511,7 +476,6 @@ void __vm_area_free(struct vm_area_struct *vma)
> > > {
> > > vma_numab_state_free(vma);
> > > free_anon_vma_name(vma);
> > > - vma_lock_free(vma);
> > > kmem_cache_free(vm_area_cachep, vma);
> > > }
> >
> > Have you investigated if this allows to perform vma_numab_state_free() and
> > free_anon_vma_name() immediately, and only kfree_rcu() the vma itself,
> > instead of performing all this in a call_rcu() callback?
>
> Yes, it should be fine to free them immediately. lock_vma_under_rcu()
> does not use neither vma->numab_state, nor vma->anon_name.
>
> >
> > Of course if we succeed converting vma's to SLAB_TYPESAFE_RCU this immediate
> > freeing of numab state and anon_vma_name would be implied, but maybe it's an
> > useful intermediate step on its own.
>
> I'm thinking maybe I should post SLAB_TYPESAFE_RCU conversion before
> anything else. It's simple and quite uncontroversial. I will probably
> do that today.
Uh, I forgot that I can't post SLAB_TYPESAFE_RCU until I eliminate
this vma_lock_free() call inside __vm_area_free(). So, I have to
bundle moving vm_lock into vm_area_struct with SLAB_TYPESAFE_RCU.