Re: [PATCHv2 04/12] x86/virt/tdx: Add tdx_alloc/free_page() helpers

[Kiryl Shutsemau]

On Tue, Sep 16, 2025 at 12:03:26AM +0000, Edgecombe, Rick P wrote:
> On Mon, 2025-06-09 at 22:13 +0300, Kirill A. Shutemov wrote:
> > +
> > +static int tdx_pamt_add(atomic_t *pamt_refcount, unsigned long hpa,
> > +			struct list_head *pamt_pages)
> > +{
> > +	u64 err;
> > +
> > +	guard(spinlock)(&pamt_lock);
> > +
> > +	hpa = ALIGN_DOWN(hpa, PMD_SIZE);
> > +
> > +	/* Lost race to other tdx_pamt_add() */
> > +	if (atomic_read(pamt_refcount) != 0) {
> > +		atomic_inc(pamt_refcount);
> > +		return 1;
> > +	}
> > +
> > +	err = tdh_phymem_pamt_add(hpa | TDX_PS_2M, pamt_pages);
> > +
> > +	/*
> > +	 * tdx_hpa_range_not_free() is true if current task won race
> > +	 * against tdx_pamt_put().
> > +	 */
> > +	if (err && !tdx_hpa_range_not_free(err)) {
> > +		pr_err("TDH_PHYMEM_PAMT_ADD failed: %#llx\n", err);
> > +		return -EIO;
> > +	}
> > +
> > +	atomic_set(pamt_refcount, 1);
> > +
> > +	if (tdx_hpa_range_not_free(err))
> > +		return 1;
> 
> Hey Kirill,
> 
> I couldn't figure out how this tdx_hpa_range_not_free() check helps. We are
> already inside the lock also taken by any operation that might affect PAMT
> state. Can you explain more about this? Otherwise I'm going to drop it for
> inability to explain.

My git has comment for the check:

https://git.kernel.org/pub/scm/linux/kernel/git/kas/linux.git/tree/arch/x86/virt/vmx/tdx/tdx.c?h=tdx/dpamt&id=375706fe73a8499dbdddb22c13d19d7286280ad6#n2160

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 PAMT memory
					    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>

Does it make sense?

-- 
  Kiryl Shutsemau / Kirill A. Shutemov