linux-kernel - Re: [PATCH] mm: tag kernel stack pages

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: Windows password security audit tool. GUI, reports in PDF.

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

Message-ID: <0ba60468-fc6d-4f07-a9ea-e16b8bcd5575@redhat.com>
Date: Thu, 4 Sep 2025 12:31:34 +0200
From: David Hildenbrand <david@...hat.com>
To: Matthew Wilcox <willy@...radead.org>
Cc: "Vishal Moola (Oracle)" <vishal.moola@...il.com>, linux-mm@...ck.org,
 linux-kernel@...r.kernel.org, Andrew Morton <akpm@...ux-foundation.org>
Subject: Re: [PATCH] mm: tag kernel stack pages

On 02.09.25 21:52, Matthew Wilcox wrote:
> On Thu, Aug 21, 2025 at 02:44:31PM +0200, David Hildenbrand wrote:
>> On 20.08.25 22:20, Vishal Moola (Oracle) wrote:
>>> Currently, we have no way to distinguish a kernel stack page from an
>>> unidentified page. Being able to track this information can be
>>> beneficial for optimizing kernel memory usage (i.e. analyzing
>>> fragmentation, location etc.). Knowing a page is being used for a kernel
>>> stack gives us more insight about pages that are certainly immovable and
>>> important to kernel functionality.
>>
>> It's a very niche use case. Anything that's not clearly a folio or a special
>> movable_ops page is certainly immovable. So we can identify pretty reliable
>> what's movable and what's not.
>>
>> Happy to learn how you would want to use that knowledge to reduce
>> fragmentation. :)
>>
>> So this reads a bit hand-wavy.
> 
> I have a theory that we should always be attempting to do aligned
> allocations if we can, falling back to individual allocations if
> we can't.  This is an attempt to gather some data to inform us whether
> that theory is true, and to help us measure whether any effort we
> take to improve that situation is effective.
> 
> Eyeballing the output of tools/testing/page-types certainly lends
> some credence to this.  On x86-64 with its 16KiB stacks and 4KiB
> page size, we often see four consecutive pages allocated as type
> KernelStack, and as you'd expect only about 25% of the time are they
> aligned to a 16KiB boundary.  That is, at least 75% of the time they
> prevent _two_ order-2 pages from being available.

I assume, ideally, you'd also know whether all these stack pages belong 
to the same thread, not various ones, right? ("context" can matter as well)

> 
> As you say, they're not movable.  I'm not sure if it makes sense to
> go to the effort of making them movable; it'd require interacting
> with the scheduler (to prevent the task we're relocating from
> being scheduled), and I don't think the realtime people would be
> terribly keen on that idea.  So that isn't one of the ideas we
> have on the table for improving matters.

Yeah, while possible I am also not sure if we always want that.

> 
> Ideas we have been batting around:
> 
>   - Have kernel stacks try to do an order-N allocation and vmap()
>     the result, fall back to current implementation
>   - Have vmalloc try to do an order-N allocation, fall back down the
>     orders on failure to allocate
>   - Change the alloc_bulk implementation to do the order-N allocation
>     and fall back
> 
> I'm sure other possibilities also exist.
> 
>> staring at [1], we allocate from vmalloc, so I would assume that these will
>> be vmalloc-typed pages in the future and we cannot change the type later.
>>
>> [1] https://kernelnewbies.org/MatthewWilcox/Memdescs
> 
> I see the vmalloc subtype as being a "we don't know any better" type.

I guess this could get nasty once we would have metadata assigned to the 
vmalloc allocations (struct vmdesc).

> We could allocate another subtype of type 0 to mean "kernel stacks"
> and have it be implicit that kernel stacks are allocated from vmalloc.

Yes, that would work.

> This would probably require that we have a vmalloc interface that lets us
> specify a subtype, which I think is probably something we'd want anyway.

vmalloc subtypes don't sound like a bad idea.

> 
> I think it's fine to say "This doesn't add enough value to merge it
> upstream".  I will note one minor advantage which is that typing these
> pages as PGTY_kstack today prevents them from being inadvertently mapped
> to userspace (whether by malicious code or innocent bug).

Yes, as raised elsewhere, if we can do this consistently today (stack -> 
PGTY_kstack), fine with me.

-- 
Cheers

David / dhildenb