Re: [PATCH v2 08/20] KVM: s390: KVM page table management functions: allocation

[Heiko Carstens <hca@linux.ibm.com>]

On Wed, Sep 17, 2025 at 01:25:56PM +0200, Claudio Imbrenda wrote:
> On Wed, 17 Sep 2025 09:27:33 +0200
> Heiko Carstens <hca@linux.ibm.com> wrote:
> 
> > On Tue, Sep 16, 2025 at 07:36:44PM +0200, Heiko Carstens wrote:
> > > On Tue, Sep 16, 2025 at 07:06:06PM +0200, Christian Borntraeger wrote:  
> > > > 
> > > > Am 16.09.25 um 19:05 schrieb Claudio Imbrenda:
> > > >   
> > > > > > > I think GFP_ATOMIC actually gives more guarantees?  
> > > > > > 
> > > > > > In real life GFP_ATOMIC can fail, GFP_KERNEL does not.All gfp allocation failures
> > > > > > are usually the atomic ones.  
> > > > > 
> > > > > interesting... then I guess I need GFP_KERNEL | GFP_ATOMIC ?  
> > > > 
> > > > No. ATOMIC always means: can fail. 
> 
> my issue is that GFP_KERNEL can sleep, and this allocation is sometimes
> called from atomic contexts (e.g. while holding spinlocks)
> 
> the right way to do this would be with mempools, to allocate memory
> (and potentially sleep) when we are not in atomic context, and use it
> whenever needed. this is on my to-do list for the future, but right now
> I'd like to avoid having to refactor a ton of code.

I doubt this is accetable even for an intermediate solution. As soon
as the host is under memory pressure and starts doing I/O to free up
memory, you will end up in -ENOMEM situations for simple guest page
allocations.

What happens with a guest in such a situation? Is this gracefully
handled without that the guest is terminated?

> > Another correction: it should be GFP_KERNEL_ACCOUNT instead, like it
> > used to be before in gmap_alloc_crst() since those page tables should
> > be accounted to the process which is allocating them.
> 
> gfp_types.h says:
> 
>  * %GFP_KERNEL_ACCOUNT is the same as GFP_KERNEL, except the allocation is
>  * accounted to kmemcg.
> 
> I thought that without GFP_KERNEL_ACCOUNT, it would be accounted toward
> the process?
> The documentation is a little confusing here

Documentation/core-api/memory-allocation.rst:

  * Untrusted allocations triggered from userspace should be a subject
    of kmem accounting and must have ``__GFP_ACCOUNT`` bit set. There
    is the handy ``GFP_KERNEL_ACCOUNT`` shortcut for ``GFP_KERNEL``
    allocations that should be accounted.

This is done for all page table allocations, except if mm == init_mm
(and unfortunately also not for the s390 specific crste alloc
functions - I'll send a patch for that).