* Re: [PATCH v1 08/26] x86/virt/tdx: Add tdx_enable_ext() to enable of TDX Module Extensions
[not found] ` <167d9540-2d9a-4367-bc68-b96494bc4044@intel.com>
@ 2025-11-24 10:41 ` Xu Yilun
0 siblings, 0 replies; only message in thread
From: Xu Yilun @ 2025-11-24 10:41 UTC (permalink / raw)
To: Dave Hansen, linux-mm
Cc: linux-coco, linux-pci, chao.gao, dave.jiang, baolu.lu, yilun.xu,
zhenzhong.duan, kvm, rick.p.edgecombe, dave.hansen,
dan.j.williams, kas, x86, akpm
On Fri, Nov 21, 2025 at 07:38:03AM -0800, Dave Hansen wrote:
> On 11/21/25 07:15, Dave Hansen wrote:
> > On 11/21/25 04:54, Xu Yilun wrote:
> > ...
> >> For now, TDX Module Extensions consume quite large amount of memory
> >> (12800 pages), print this readout value on TDX Module Extentions
> >> initialization.
> > Overall, the description is looking better, thanks!
> >
> > A few more nits, though. Please don't talk about things in terms of
> > number of pages. Just give the usage in megabytes.
>
> Oh, and please at least have a discussion with the memory management
> folks about consuming this amount of memory forever. I think it's quite
> possible they will prefer it be allocated in a way other than thousands
> of plain old allocations.
>
> For example, imagine memory was fragmented and those 12800 pages came
> from 12,800 different 2M regions. Well, now you've got ~50GB of memory
> that is _permanently_ fragmented and will never be able to satisfy a 2M
> allocation.
>
> You might get an answer that it's better to do a small number of
> max-size buddy allocations than a large number of PAGE_SIZE allocations.
Loop in mm folks.
Hi mm folks, for Intel TDX (Trust Domain Extensions) feature, there is
a requirement to donate quite a number of pages (12800 x 4K = 50MB for
now) to TDX firmware (known as TDX Module) for its initialization. These
pages will never be revoked cause the TDX Module initialization is a one
way path.
The TDX Module doesn't require these pages be physically contiguous, and
the patches [1][2] in this series [3] does PAGE_SIZE allocation. But as
mentioned by Dave, the donation may _permanently_ fragment regions, stop
them from 2M huge page allocation. In worst case, 12800 x 2MB = 25GB
memory region.
So is order based buddy allocation a better choice? I believe so. And if
that fails, should we fall back to PAGE_SIZE allocation? Or PAGE_SIZE
allocation should be a hard no in this _permanent_ donation case?
[1]: https://lore.kernel.org/linux-coco/20251117022311.2443900-7-yilun.xu@linux.intel.com/
[2]: https://lore.kernel.org/linux-coco/20251117022311.2443900-9-yilun.xu@linux.intel.com/
[3]: https://lore.kernel.org/linux-coco/20251117022311.2443900-1-yilun.xu@linux.intel.com/
Thanks,
Yilun
^ permalink raw reply [flat|nested] only message in thread