linux-kernel - Re: [RFC, PATCH 05/12] KVM: TDX: Add tdx_pamt

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: <cu332ci4fvpgb6akgpg3p53336qndi36px5osfv57vcq5u3din@kxu2diy723tq>
Date: Thu, 8 May 2025 16:16:50 +0300
From: "kirill.shutemov@...ux.intel.com" <kirill.shutemov@...ux.intel.com>
To: Dave Hansen <dave.hansen@...el.com>
Cc: "Huang, Kai" <kai.huang@...el.com>, 
	"pbonzini@...hat.com" <pbonzini@...hat.com>, "seanjc@...gle.com" <seanjc@...gle.com>, 
	"Edgecombe, Rick P" <rick.p.edgecombe@...el.com>, "bp@...en8.de" <bp@...en8.de>, 
	"dave.hansen@...ux.intel.com" <dave.hansen@...ux.intel.com>, "x86@...nel.org" <x86@...nel.org>, 
	"mingo@...hat.com" <mingo@...hat.com>, "Zhao, Yan Y" <yan.y.zhao@...el.com>, 
	"tglx@...utronix.de" <tglx@...utronix.de>, "kvm@...r.kernel.org" <kvm@...r.kernel.org>, 
	"linux-coco@...ts.linux.dev" <linux-coco@...ts.linux.dev>, "Yamahata, Isaku" <isaku.yamahata@...el.com>, 
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>
Subject: Re: [RFC, PATCH 05/12] KVM: TDX: Add tdx_pamt_get()/put() helpers

On Wed, May 07, 2025 at 09:31:22AM -0700, Dave Hansen wrote:
> On 5/5/25 05:44, Huang, Kai wrote:
> >> +static int tdx_pamt_add(atomic_t *pamt_refcount, unsigned long hpa,
> >> +			struct list_head *pamt_pages)
> >> +{
> >> +	u64 err;
> >> +
> >> +	hpa = ALIGN_DOWN(hpa, SZ_2M);
> >> +
> >> +	spin_lock(&pamt_lock);
> > Just curious, Can the lock be per-2M-range?
> 
> Folks, please keep it simple.
> 
> If there's lock contention on this, we'll fix the lock contention, or
> hash the physical address into a fixed number of locks.

I had this idea in mind as well.

> But having it be
> per-2M-range sounds awful. Then you have to size it, and allocate it and
> then resize it if there's ever hotplug, etc...
> 
> Kirill, could you put together some kind of torture test for this,
> please? I would imagine a workload which is sitting in a loop setting up
> and tearing down VMs on a bunch of CPUs would do it.

It has to be multiple parallel creation/teardown loops. With single TD we
won't see much concurrency. Most of PAMT allocations comes from single
VCPU.

And it makes sense to do with huge pages as it cuts number of allocated
PAMT memory allocated on TD creation by factor of 10 in my setup.

JFYI, booting a TD with huge pages consumes 1-2MB of PAMT memory. I doubt
any optimization here is justifiable.

> That ^ would be the worst possible case, I think. If you don't see lock
> contention there, you'll hopefully never see it on real systems.
> 
> I *suspect* that real systems will get bottlenecked somewhere in the
> page conversion process rather than on this lock. But it should be a
> pretty simple experiment to run.

-- 
  Kiryl Shutsemau / Kirill A. Shutemov