Re: [PATCH 4/6] x86/virt/tdx: Add extra memory to TDX module for the extensions

[Xu Yilun <yilun.xu@linux.intel.com>]

On Wed, Sep 02, 2026 at 11:54:25AM +0100, Kiryl Shutsemau wrote:
> On Fri, Aug 28, 2026 at 03:46:16PM +0800, Xu Yilun wrote:
> > > > > +	page = alloc_contig_pages(required_pages, GFP_KERNEL, numa_mem_id(),
> > > > > +				  &node_online_map);
> > > > 
> > > > Why contiguous? TDH.EXT.MEM.ADD takes a list of page addresses and the loop
> > > > below writes every one of them out separately.
> > > > 
> > > > alloc_pages_bulk() fits the chunking that is already here, and a short
> > > > return can be handled per chunk. alloc_contig_pages() isolates and migrates
> > > > to get its range and fails TDX init outright when it cannot find one. PAMT
> > > 
> > > Yeah, this is not the ABI requirement, but the kernel's consideration. A
> > > brief reasoning in the commit log: avoiding permanent memory fragmentation
> > > and buddy allocator efficiency loss.
> > > 
> > > Also there is some discussion:
> > > 
> > > https://lore.kernel.org/all/167d9540-2d9a-4367-bc68-b96494bc4044@intel.com/
> > > 
> > > TL;DR
> > >   - The memory will never return to the kernel.
> > >   - There is chance that this tens of megabytes will fragment tens of
> > >     gigabytes of memory forever.
> > >   - The chance of fragmentation is actually low since at boot up, but
> > >     let the buddy allocator take care of these never-returned memory
> > >     is not necessary and lowers its efficiency.
> > 
> > Hi Kiryl & David:
> > 
> > I see there is another suggestion that the whole memory adding process
> > could be a little simpler if we allocate & add pages 4k by 4k [1],
> > rather than one-time pre-allocation. The concern of this alternative is,
> > as said above, memory fragmentation.
> > 
> > [1] https://lore.kernel.org/lkml/f48b83feb2ee1d3c88b5a1627cf35b4b282d3f90.camel@intel.com/
> > 
> > And I've realized the memory fragmentation discussion is not actually
> > closed in previous thread [2]. We need more input.
> > 
> > [2] https://lore.kernel.org/all/167d9540-2d9a-4367-bc68-b96494bc4044@intel.com/
> > 
> > Let me give a brief overview of the problem:
> > 
> > Intel TDX (Trust Domain Extensions) is a feature for confidential
> > computing. A secure firmware called "TDX module" runs in an isolated
> > environment to provide services about security.
> > 
> > In Linux, the host initializes TDX module at boot up time
> > (subsys_initcall()). During the initialization, the host must donate
> > tens of mega bytes physical memory (35M ~ 110M in the forseeable
> > future) to the TDX module. These memory will *never be revoked* cause
> > the TDX Module initialization is a one way path.
> > 
> > The TDX Module doesn't require this memory be physically contiguous. But
> > the kernel side concern is if we do PAGE_SIZE allocation, it may
> > permanently fragment memory regions, stop them from allocating 2M huge
> > pages. In worst case, ~50G (110M * 512) memory regions affacted.
> > 
> > So is the physically contiguous allocation really a better choice here?
> > We appreciate inputs from mm folks. Thanks!
> 
> What matters for fragmentation is not contiguity, it is how many
> pageblocks are left partially occupied by unmovable pages that are never
> freed.
> 
> So you can ask one pageblock at a time with
> 
> 	page = alloc_pages(GFP_KERNEL | __GFP_NOWARN, order);
> 
> with fallback to lower order if you must.
> 
> It also fits the ABI: pageblock_order is 9 on x86, i.e. 512 pages, which
> is exactly TDX_HPA_LIST_MAX_NR_PAGES. One allocation is one full HPA list
> is one TDH.EXT.MEM.ADD, so the allocation loop and the chunking loop
> become the same loop.
> 
> But alloc_contig_pages() might be a good enough approximation for
> per-pageblock allocation if we do it during the boot when fragmentation
> is low.

IIUC, you mean alloc_contig_pages() also gives good de-fragmentation
that we need. But it would be slightly easier to fail cause it requires
extra contiguity that we don't need.

Multiple alloc_pages(order-9) meets our requirement exactly but the
falling back to lower order may create more fragments. And we can do
this because of the ABI definition - an HPA_LIST could happen to hold
an entire pageblock.

If I have to choose, I prefer alloc_contig_pages(). It doesn't have to
depend on HPA_LIST ABI details.

> 
> My alloc_pages_bulk() suggestion is wrong. It doesn't get any control
> over allocation placement. It will tap in the per-CPU cache first.
> 
> -- 
>   Kiryl Shutsemau / Kirill A. Shutemov