linux-kernel - Re: [PATCHv2 00/12] TDX: Enable Dynamic 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: <aKXqUf9QBpLOeB3Q@google.com>
Date: Wed, 20 Aug 2025 08:31:29 -0700
From: Sean Christopherson <seanjc@...gle.com>
To: Rick P Edgecombe <rick.p.edgecombe@...el.com>
Cc: "kas@...nel.org" <kas@...nel.org>, Chao Gao <chao.gao@...el.com>, 
	Dave Hansen <dave.hansen@...el.com>, "x86@...nel.org" <x86@...nel.org>, "bp@...en8.de" <bp@...en8.de>, 
	Kai Huang <kai.huang@...el.com>, 
	"linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>, "tglx@...utronix.de" <tglx@...utronix.de>, 
	Yan Y Zhao <yan.y.zhao@...el.com>, "mingo@...hat.com" <mingo@...hat.com>, 
	"pbonzini@...hat.com" <pbonzini@...hat.com>, Isaku Yamahata <isaku.yamahata@...el.com>, 
	"dave.hansen@...ux.intel.com" <dave.hansen@...ux.intel.com>, 
	"linux-coco@...ts.linux.dev" <linux-coco@...ts.linux.dev>, "kvm@...r.kernel.org" <kvm@...r.kernel.org>
Subject: Re: [PATCHv2 00/12] TDX: Enable Dynamic PAMT

On Fri, Aug 15, 2025, Rick P Edgecombe wrote:
> On Thu, 2025-08-14 at 11:55 +0100, Kiryl Shutsemau wrote:
> > > > > (similar pattern on the unmapping)
> > > > > 
> > > > > So it will only be valid contention if two threads try to fault in the >
> > > > > > *same* 2MB
> > > > > DPAMT region *and* lose that race around 1-3, but invalid contention if
> > > > > > > threads try
> > > > > to execute 2-4 at the same time for any different 2MB regions.
> > > > > 
> > > > > Let me go verify.
> 
> It lost the race only once over a couple runs. So it seems mostly invalid
> contention.
> 
> > > 
> > > Note that in absence of the global lock here, concurrent PAMT.ADD would
> > > also trigger some cache bouncing during pamt_walk() on taking shared
> > > lock on 1G PAMT entry and exclusive lock on 2M entries in the same
> > > cache (4 PAMT_2M entries per cache line). This is hidden by the global
> > > lock.
> > > 
> > > You would not recover full contention time by removing the global lock.
> 
> Hmm, yea. Another consideration is that performance sensitive users will
> probably be using huge pages, in which case 4k PAMT will be mostly skipped.
> 
> But man, the number and complexity of the locks is getting a bit high across the
> whole stack. I don't have any easy ideas.

FWIW, I'm not concerned about bouncing cachelines, I'm concerned about the cost
of the SEAMCALLs.  The latency due to bouncing a cache line due to "false"
contention is probably in the noise compared to waiting thousands of cycles for
other SEAMCALLs to complete.

That's also my concern with tying PAMT management to S-EPT population.  E.g. if
a use case triggers a decent amount S-EPT churn, then dynamic PAMT support will
exacerbate the S-EPT overhead.

But IIUC, that's a limitation of the TDX-Module design, i.e. there's no way to
hand it a pool of PAMT pages to manage.  And I suppose if a use case is churning
S-EPT, then it's probably going to be sad no matter what.  So, as long as the KVM
side of things isn't completely awful, I can live with on-demand PAMT management.

As for the global lock, I don't really care what we go with for initial support,
just so long as there's clear line of sight to an elegant solution _if_ we need
shard the lock.