Re: [PATCH 2/2] iommu/amd: Use 'unsigned long' for domain->pt_root

[Qian Cai]

On Thu, Jun 25, 2020 at 04:52:27PM +0200, Joerg Roedel wrote:
> From: Joerg Roedel <jroedel@xxxxxxx>
> 
> Using atomic64_t can be quite expensive, so use unsigned long instead.
> This is safe because the write becomes visible atomically.
> 
> Signed-off-by: Joerg Roedel <jroedel@xxxxxxx>
> ---
>  drivers/iommu/amd/amd_iommu_types.h |  2 +-
>  drivers/iommu/amd/iommu.c           | 10 ++++++++--
>  2 files changed, 9 insertions(+), 3 deletions(-)
> 
> diff --git a/drivers/iommu/amd/amd_iommu_types.h b/drivers/iommu/amd/amd_iommu_types.h
> index 30a5d412255a..f6f102282dda 100644
> --- a/drivers/iommu/amd/amd_iommu_types.h
> +++ b/drivers/iommu/amd/amd_iommu_types.h
> @@ -468,7 +468,7 @@ struct protection_domain {
>  				       iommu core code */
>  	spinlock_t lock;	/* mostly used to lock the page table*/
>  	u16 id;			/* the domain id written to the device table */
> -	atomic64_t pt_root;	/* pgtable root and pgtable mode */
> +	unsigned long pt_root;	/* pgtable root and pgtable mode */
>  	int glx;		/* Number of levels for GCR3 table */
>  	u64 *gcr3_tbl;		/* Guest CR3 table */
>  	unsigned long flags;	/* flags to find out type of domain */
> diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
> index 5286ddcfc2f9..b0e1dc58244e 100644
> --- a/drivers/iommu/amd/iommu.c
> +++ b/drivers/iommu/amd/iommu.c
> @@ -156,7 +156,7 @@ static struct protection_domain *to_pdomain(struct iommu_domain *dom)
>  static void amd_iommu_domain_get_pgtable(struct protection_domain *domain,
>  					 struct domain_pgtable *pgtable)
>  {
> -	u64 pt_root = atomic64_read(&domain->pt_root);
> +	unsigned long pt_root = domain->pt_root;

The pt_root might be reload later in case of register pressure where the
compiler decides to not store it as a stack variable, so it needs
smp_rmb() here to match to the smp_wmb() in
amd_iommu_domain_set_pt_root() to make the load visiable to all CPUs.

Then, smp_rmb/wmb() wouldn't be able to deal with data races, so it
needs,

unsigned long pt_root = READ_ONCE(domain->pt_root);

>  
>  	pgtable->root = (u64 *)(pt_root & PAGE_MASK);
>  	pgtable->mode = pt_root & 7; /* lowest 3 bits encode pgtable mode */
> @@ -164,7 +164,13 @@ static void amd_iommu_domain_get_pgtable(struct protection_domain *domain,
>  
>  static void amd_iommu_domain_set_pt_root(struct protection_domain *domain, u64 root)
>  {
> -	atomic64_set(&domain->pt_root, root);
> +	domain->pt_root = root;

WRITE_ONCE(domain->pt_root, root);

> +
> +	/*
> +	 * The new value needs to be gobally visible in case pt_root gets
> +	 * cleared, so that the page-table can be safely freed.
> +	 */
> +	smp_wmb();
>  }
>  
>  static void amd_iommu_domain_clr_pt_root(struct protection_domain *domain)
> -- 
> 2.27.0
>