Re: [PATCH 2/2] mm/page_alloc: Fix race condition in unaccepted memory handling

Re: [PATCH 2/2] mm/page_alloc: Fix race condition in unaccepted memory handling

[Borislav Petkov]

On Tue, May 06, 2025 at 02:25:09PM +0300, Kirill A. Shutemov wrote:
> The page allocator tracks the number of zones that have unaccepted
> memory using static_branch_enc/dec() and uses that static branch in hot
> paths to determine if it needs to deal with unaccepted memory.
> 
> Borisal and Thomas pointed out that the tracking is racy operations on

Boris or Borislav would be nice.

> static_branch are not serialized against adding/removing unaccepted pages
> to/from the zone.

Also, that sentence needs massaging.

> The effect of this static_branch optimization is only visible on
> microbenchmark.
> 
> Instead of adding more complexity around it, remove it altogether.

Yah, good idea.

> Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> Fixes: dcdfdd40fa82 ("mm: Add support for unaccepted memory")
> Link: https://lore.kernel.org/all/20250506092445.GBaBnVXXyvnazly6iF@fat_crate.local
> Reported-by: Borislav Petkov <bp@alien8.de>

Tested-by: Borislav Petkov (AMD) <bp@alien8.de>

Thx for the quick fix.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Brendan Jackman]

On Tue May 6, 2025 at 11:25 AM UTC, Kirill A. Shutemov wrote:
> +	/* Bailout, since try_to_accept_memory_one() needs to take a lock */
> +	if (alloc_flags & ALLOC_TRYLOCK)
> +		return false;
> +

Quick lazy question: why don't we just trylock it like we do for the zone
lock?

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Kirill A. Shutemov]

On Tue, May 06, 2025 at 01:20:25PM +0000, Brendan Jackman wrote:
> On Tue May 6, 2025 at 11:25 AM UTC, Kirill A. Shutemov wrote:
> > +	/* Bailout, since try_to_accept_memory_one() needs to take a lock */
> > +	if (alloc_flags & ALLOC_TRYLOCK)
> > +		return false;
> > +
> 
> Quick lazy question: why don't we just trylock it like we do for the zone
> lock?

It is not only zone lock. There's also unaccepted_memory_lock inside
accept_memory().

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Alexei Starovoitov]

On Tue, May 6, 2025 at 4:25 AM Kirill A. Shutemov
<kirill.shutemov@linux.intel.com> wrote:
>
> try_alloc_pages() will not attempt to allocate memory if the system has
> *any* unaccepted memory. Memory is accepted as needed and can remain in
> the system indefinitely, causing the interface to always fail.
>
> Rather than immediately giving up, attempt to use already accepted
> memory on free lists.
>
> Pass 'alloc_flags' to cond_accept_memory() and do not accept new memory
> for ALLOC_TRYLOCK requests.
>
> Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> Fixes: 97769a53f117 ("mm, bpf: Introduce try_alloc_pages() for opportunistic page allocation")

Thanks for working on this, but the fixes tag is overkill.
This limitation is not causing any issues in our setups.
Improving it is certainly better, of course.
Acked-by: Alexei Starovoitov <ast@kernel.org>

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Kirill A. Shutemov]

On Tue, May 06, 2025 at 10:18:21AM -0700, Alexei Starovoitov wrote:
> On Tue, May 6, 2025 at 4:25 AM Kirill A. Shutemov
> <kirill.shutemov@linux.intel.com> wrote:
> >
> > try_alloc_pages() will not attempt to allocate memory if the system has
> > *any* unaccepted memory. Memory is accepted as needed and can remain in
> > the system indefinitely, causing the interface to always fail.
> >
> > Rather than immediately giving up, attempt to use already accepted
> > memory on free lists.
> >
> > Pass 'alloc_flags' to cond_accept_memory() and do not accept new memory
> > for ALLOC_TRYLOCK requests.
> >
> > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> > Fixes: 97769a53f117 ("mm, bpf: Introduce try_alloc_pages() for opportunistic page allocation")
> 
> Thanks for working on this, but the fixes tag is overkill.
> This limitation is not causing any issues in our setups.

Have you had chance to test it on any platform with unaccepted memory?
So far it is only Intel TDX and AMD SEV guests.

> Improving it is certainly better, of course.
> Acked-by: Alexei Starovoitov <ast@kernel.org>

Thanks!

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Alexei Starovoitov]

On Tue, May 6, 2025 at 12:00 PM Kirill A. Shutemov
<kirill.shutemov@linux.intel.com> wrote:
>
> On Tue, May 06, 2025 at 10:18:21AM -0700, Alexei Starovoitov wrote:
> > On Tue, May 6, 2025 at 4:25 AM Kirill A. Shutemov
> > <kirill.shutemov@linux.intel.com> wrote:
> > >
> > > try_alloc_pages() will not attempt to allocate memory if the system has
> > > *any* unaccepted memory. Memory is accepted as needed and can remain in
> > > the system indefinitely, causing the interface to always fail.
> > >
> > > Rather than immediately giving up, attempt to use already accepted
> > > memory on free lists.
> > >
> > > Pass 'alloc_flags' to cond_accept_memory() and do not accept new memory
> > > for ALLOC_TRYLOCK requests.
> > >
> > > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> > > Fixes: 97769a53f117 ("mm, bpf: Introduce try_alloc_pages() for opportunistic page allocation")
> >
> > Thanks for working on this, but the fixes tag is overkill.
> > This limitation is not causing any issues in our setups.
>
> Have you had chance to test it on any platform with unaccepted memory?
> So far it is only Intel TDX and AMD SEV guests.

We don't use them, and my understanding is that such
unaccepted memory will be there only during boot time.

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Andrew Morton]

On Tue,  6 May 2025 14:25:08 +0300 "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> wrote:

> try_alloc_pages() will not attempt to allocate memory if the system has
> *any* unaccepted memory. Memory is accepted as needed and can remain in
> the system indefinitely, causing the interface to always fail.
> 
> Rather than immediately giving up, attempt to use already accepted
> memory on free lists.
> 
> Pass 'alloc_flags' to cond_accept_memory() and do not accept new memory
> for ALLOC_TRYLOCK requests.

What are the userspace-visible effects, please?

Was the omission of cc:stable intentional?  I cannot locally determine
this without the above info.

If the cc:stable omission was indeed intentional then it would be better
if this series was presented as two standalone patches.

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Kirill A. Shutemov]

On Tue, May 06, 2025 at 03:05:31PM -0700, Alexei Starovoitov wrote:
> On Tue, May 6, 2025 at 12:00 PM Kirill A. Shutemov
> <kirill.shutemov@linux.intel.com> wrote:
> >
> > On Tue, May 06, 2025 at 10:18:21AM -0700, Alexei Starovoitov wrote:
> > > On Tue, May 6, 2025 at 4:25 AM Kirill A. Shutemov
> > > <kirill.shutemov@linux.intel.com> wrote:
> > > >
> > > > try_alloc_pages() will not attempt to allocate memory if the system has
> > > > *any* unaccepted memory. Memory is accepted as needed and can remain in
> > > > the system indefinitely, causing the interface to always fail.
> > > >
> > > > Rather than immediately giving up, attempt to use already accepted
> > > > memory on free lists.
> > > >
> > > > Pass 'alloc_flags' to cond_accept_memory() and do not accept new memory
> > > > for ALLOC_TRYLOCK requests.
> > > >
> > > > Signed-off-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
> > > > Fixes: 97769a53f117 ("mm, bpf: Introduce try_alloc_pages() for opportunistic page allocation")
> > >
> > > Thanks for working on this, but the fixes tag is overkill.
> > > This limitation is not causing any issues in our setups.
> >
> > Have you had chance to test it on any platform with unaccepted memory?
> > So far it is only Intel TDX and AMD SEV guests.
> 
> We don't use them, and my understanding is that such
> unaccepted memory will be there only during boot time.

That's false. Unaccepted memory can be there indefinitely after boot. It
only gets accepted on demand.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Kirill A. Shutemov]

On Tue, May 06, 2025 at 05:00:34PM -0700, Andrew Morton wrote:
> On Tue,  6 May 2025 14:25:08 +0300 "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> wrote:
> 
> > try_alloc_pages() will not attempt to allocate memory if the system has
> > *any* unaccepted memory. Memory is accepted as needed and can remain in
> > the system indefinitely, causing the interface to always fail.
> > 
> > Rather than immediately giving up, attempt to use already accepted
> > memory on free lists.
> > 
> > Pass 'alloc_flags' to cond_accept_memory() and do not accept new memory
> > for ALLOC_TRYLOCK requests.
> 
> What are the userspace-visible effects, please?

I cannot say I fully understand the implications.

The interface obviously allows failure, but the caller might expect
eventual success on retry.

So far, only BPF uses the interface. Maybe Alexei can comment on what will
happen if the function always fails.

I noticed the issue by code analysis because the second patch removes
has_unaccepted_memory().

> Was the omission of cc:stable intentional?  I cannot locally determine
> this without the above info.
> 
> If the cc:stable omission was indeed intentional then it would be better
> if this series was presented as two standalone patches.

Given that the second patch cannot be applied to current Linus' tree
without this one, it is better to add stable@.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov

Re: [PATCH 1/2] mm/page_alloc: Ensure try_alloc_pages() plays well with unaccepted memory

[Brendan Jackman]

On Tue May 6, 2025 at 1:34 PM UTC, Kirill A. Shutemov wrote:
> On Tue, May 06, 2025 at 01:20:25PM +0000, Brendan Jackman wrote:
>> On Tue May 6, 2025 at 11:25 AM UTC, Kirill A. Shutemov wrote:
>> > +	/* Bailout, since try_to_accept_memory_one() needs to take a lock */
>> > +	if (alloc_flags & ALLOC_TRYLOCK)
>> > +		return false;
>> > +
>> 
>> Quick lazy question: why don't we just trylock it like we do for the zone
>> lock?
>
> It is not only zone lock. There's also unaccepted_memory_lock inside
> accept_memory().

Right, but my lazy question was why can't we "just" trylock that too?

But anyway, that's no use because if we win the trylock we'd still have
to do __free_pages_ok().