linux-kernel - Re: [PATCH v4 2/5] mm: move per-vma lock into vm_area

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <CAJuCfpHVQyVGR9sPtKLYt4a7rQUUopX0cBfk53=3RqpYK4=V8w@mail.gmail.com>
Date: Thu, 21 Nov 2024 09:05:21 -0800
From: Suren Baghdasaryan <surenb@...gle.com>
To: Shakeel Butt <shakeel.butt@...ux.dev>
Cc: akpm@...ux-foundation.org, willy@...radead.org, liam.howlett@...cle.com, 
	lorenzo.stoakes@...cle.com, mhocko@...e.com, vbabka@...e.cz, 
	hannes@...xchg.org, mjguzik@...il.com, oliver.sang@...el.com, 
	mgorman@...hsingularity.net, david@...hat.com, peterx@...hat.com, 
	oleg@...hat.com, dave@...olabs.net, paulmck@...nel.org, brauner@...nel.org, 
	dhowells@...hat.com, hdanton@...a.com, hughd@...gle.com, minchan@...gle.com, 
	jannh@...gle.com, souravpanda@...gle.com, pasha.tatashin@...een.com, 
	corbet@....net, linux-doc@...r.kernel.org, linux-mm@...ck.org, 
	linux-kernel@...r.kernel.org, kernel-team@...roid.com
Subject: Re: [PATCH v4 2/5] mm: move per-vma lock into vm_area_struct

On Wed, Nov 20, 2024 at 11:02 PM Shakeel Butt <shakeel.butt@...ux.dev> wrote:
>
> On Wed, Nov 20, 2024 at 04:33:37PM -0800, Suren Baghdasaryan wrote:
> [...]
> > > > >
> > > > > Do we just want 'struct vm_area_struct' to be cacheline aligned or do we
> > > > > want 'struct vma_lock vm_lock' to be on a separate cacheline as well?
> > > >
> > > > We want both to minimize cacheline sharing.
> > > >
> > >
> > > For later, you will need to add a pad after vm_lock as well, so any
> > > future addition will not share the cacheline with vm_lock. I would do
> > > something like below. This is a nit and can be done later.
> > >
> > > diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> > > index 7654c766cbe2..5cc4fff163a0 100644
> > > --- a/include/linux/mm_types.h
> > > +++ b/include/linux/mm_types.h
> > > @@ -751,10 +751,12 @@ struct vm_area_struct {
> > >  #endif
> > >         struct vm_userfaultfd_ctx vm_userfaultfd_ctx;
> > >  #ifdef CONFIG_PER_VMA_LOCK
> > > +       CACHELINE_PADDING(__pad1__);
> > >         /* Unstable RCU readers are allowed to read this. */
> > > -       struct vma_lock vm_lock ____cacheline_aligned_in_smp;
> > > +       struct vma_lock vm_lock;
> > > +       CACHELINE_PADDING(__pad2__);
> > >  #endif
> > > -} __randomize_layout;
> > > +} __randomize_layout ____cacheline_aligned_in_smp;
> >
> > I thought SLAB_HWCACHE_ALIGN for vm_area_cachep added in this patch
> > would have the same result, no?
> >
>
> SLAB_HWCACHE_ALIGN is more about the slub allocator allocating cache
> aligned memory. It does not say anything about the internals of the
> struct for which the kmem_cache is being created. The
> ____cacheline_aligned_in_smp tag in your patch made sure that the field
> vm_lock will be put in a new cacheline and there can be a hole between
> vm_lock and the previous field if the previous field is not ending at
> the cacheline boundary. Please note that if you add a new field after
> vm_lock (without cacheline alignment tag), it will be on the same
> cacheline as vm_lock. So, your code is achieving the vm_lock on its own
> cacheline goal but vm_lock being the only field on that cacheline is not
> being achieved.

Sorry, I should have been more clear. It's ok if some fields which are
rarely accessed in the pagefault path are placed in the same cacheling
with vm_lock. In fact I've done that to pack them better in the
previous version of this patchset here:
https://lore.kernel.org/all/20241111205506.3404479-5-surenb@google.com/
(removed for now based on the feedback). So, vm_lock being the only
field on the cacheline is not my goal. After this patchset I'm
planning to try packing the members better and save some memory.

>