Re: [RFC PATCH v5 16/45] x86/virt/tdx: Add tdx_alloc/free_control_page() helpers

[Huang, Kai]

On Tue, 2026-02-10 at 14:19 -0800, Dave Hansen wrote:
> On 2/10/26 14:15, Edgecombe, Rick P wrote:
> > I'm wasn't familiar with atomic_dec_and_lock(). I'm guess the atomic
> > part doesn't cover both decrementing *and* taking the lock?
> 
> Right. Only 1=>0 is under the lock. All other decs are outside the lock.
> 
> It doesn't do the atomic and the lock "atomically together" somehow.

Sorry I am a bit confused.  But I think the "1=>0 and lock" are atomic
together?

If so, I think we can avoid the "race" mentioned by Rick, which is handled
by "x86/virt/tdx: Optimize tdx_alloc/free_control_page() helpers".

Kirill described the race [*].  Quote it here:

---

  Consider the following scenario

	CPU0				CPU1

  tdx_pamt_put()
    atomic_dec_and_test() == true
    				    tdx_pamt_get()
				      atomic_inc_not_zero() == false
					tdx_pamt_add()
					  <takes pamt_lock>
					  // CPU0 never removed PAMTmemory
					  tdh_phymem_pamt_add() ==  
							HPA_RANGE_NOT_FREE
					  atomic_set(1);
					  <drops pamt_lock>
  <takes pamt_lock>
  // Lost the race to CPU1
  atomic_read() > 0
  <drop pamt_lock>

---

But with atomic_dec_and_lock() (assuming "1=>0 and lock" is atomic), I think
this race won't happen.  In tdx_pamt_put() on CPU0, the lock will always be
grabbed when refcount becomes 0, so PAMT pages are guaranteed to be freed. 
Therefore tdx_pamt_get() on CPU1 should never meet "HPA_RANGE_NOT_FREE".

[*]
https://lore.kernel.org/kvm/bfaswqmlsyycr3alibn6f422cjtpd6ybssjekvrrz4zdwgwfcz@pxy25ra4sln2/