[PATCH v1 0/1] Don't split flush for amd_iommu_domain_flush_all()

[PATCH v1 0/1] Don't split flush for amd_iommu_domain_flush_all()

[Weinan Liu]

This patch is a respin of Josef Bacik's original work[1] to fix the
performance issues and soft lockups in the AMD IOMMU driver when running
within a VM. I am taking over the respin as Josef currently has limited
time for this

[1]:https://lore.kernel.org/linux-iommu/ad8652c5e9f8aeee05e2103f4987589cdd4a3fd0.1772659768.git.josef@toxicpanda.com

Weinan Liu (1):
  iommu/amd: Don't split flush for amd_iommu_domain_flush_all()

 drivers/iommu/amd/iommu.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

-- 
2.54.0.rc0.605.g598a273b03-goog

[PATCH v1 1/1] iommu/amd: Don't split flush for amd_iommu_domain_flush_all()

[Weinan Liu]

We have observed multiple full invalidations occurring during device
detach when we are done using the vfio-device.

blocked_domain_attach_device()
  -> detach_device()
    -> amd_iommu_domain_flush_all()
      -> amd_iommu_domain_flush_pages(..., CMD_INV_IOMMU_ALL_PAGES_ADDRESS)

      	while (size != 0) {

          -> __domain_flush_pages( flush_size /* power of 2 flush_size */)
            -> domain_flush_pages_v1()
              -> build_inv_iommu_pages()
                -> build_inv_address()

         }

build_inv_address() will trigger a full invalidation  if the chunk
size > (1 << 51). Consequently, the guest will issue multiple full
invalidations for a single call to  amd_iommu_domain_flush_all()

Without this patch, we will see 10 time instead of 1 time full
invalidations for every amd_iommu_domain_flush_all().

Fixes: a270be1b3fdf ("iommu/amd: Use only natural aligned flushes in a VM")

Suggested-by: Josef Bacik <josef@toxicpanda.com>
Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Weinan Liu <wnliu@google.com>
---
 drivers/iommu/amd/iommu.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
index 760d5f4623b5..bcec8721d228 100644
--- a/drivers/iommu/amd/iommu.c
+++ b/drivers/iommu/amd/iommu.c
@@ -1769,7 +1769,8 @@ void amd_iommu_domain_flush_pages(struct protection_domain *domain,
 {
 	lockdep_assert_held(&domain->lock);
 
-	if (likely(!amd_iommu_np_cache)) {
+	if (likely(!amd_iommu_np_cache) ||
+		size == CMD_INV_IOMMU_ALL_PAGES_ADDRESS) {
 		__domain_flush_pages(domain, address, size);
 
 		/* Wait until IOMMU TLB and all device IOTLB flushes are complete */
-- 
2.54.0.rc0.605.g598a273b03-goog

Re: [PATCH v1 1/1] iommu/amd: Don't split flush for amd_iommu_domain_flush_all()

[Jason Gunthorpe]

On Tue, Apr 14, 2026 at 09:06:26PM +0000, Weinan Liu wrote:
> diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
> index 760d5f4623b5..bcec8721d228 100644
> --- a/drivers/iommu/amd/iommu.c
> +++ b/drivers/iommu/amd/iommu.c
> @@ -1769,7 +1769,8 @@ void amd_iommu_domain_flush_pages(struct protection_domain *domain,
>  {
>  	lockdep_assert_held(&domain->lock);
>  
> -	if (likely(!amd_iommu_np_cache)) {
> +	if (likely(!amd_iommu_np_cache) ||
> +		size == CMD_INV_IOMMU_ALL_PAGES_ADDRESS) {

This is better as size >= (1ULL<<52)

Which is the break point where it switches over to full invalidation
anyhow, there is no sense then in doing the special np flow..

Otherwise

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

Jason

Re: [PATCH v1 1/1] iommu/amd: Don't split flush for amd_iommu_domain_flush_all()

[Samiullah Khawaja]

On Tue, Apr 14, 2026 at 09:06:26PM +0000, Weinan Liu wrote:
>We have observed multiple full invalidations occurring during device
>detach when we are done using the vfio-device.
>
>blocked_domain_attach_device()
>  -> detach_device()
>    -> amd_iommu_domain_flush_all()
>      -> amd_iommu_domain_flush_pages(..., CMD_INV_IOMMU_ALL_PAGES_ADDRESS)
>
>      	while (size != 0) {
>
>          -> __domain_flush_pages( flush_size /* power of 2 flush_size */)
>            -> domain_flush_pages_v1()
>              -> build_inv_iommu_pages()
>                -> build_inv_address()
>
>         }
>
>build_inv_address() will trigger a full invalidation  if the chunk
>size > (1 << 51). Consequently, the guest will issue multiple full
>invalidations for a single call to  amd_iommu_domain_flush_all()
>
>Without this patch, we will see 10 time instead of 1 time full
>invalidations for every amd_iommu_domain_flush_all().
>
>Fixes: a270be1b3fdf ("iommu/amd: Use only natural aligned flushes in a VM")
>
>Suggested-by: Josef Bacik <josef@toxicpanda.com>
>Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
>Signed-off-by: Weinan Liu <wnliu@google.com>
>---
> drivers/iommu/amd/iommu.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
>diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
>index 760d5f4623b5..bcec8721d228 100644
>--- a/drivers/iommu/amd/iommu.c
>+++ b/drivers/iommu/amd/iommu.c
>@@ -1769,7 +1769,8 @@ void amd_iommu_domain_flush_pages(struct protection_domain *domain,
> {
> 	lockdep_assert_held(&domain->lock);
>
>-	if (likely(!amd_iommu_np_cache)) {
>+	if (likely(!amd_iommu_np_cache) ||
>+		size == CMD_INV_IOMMU_ALL_PAGES_ADDRESS) {
> 		__domain_flush_pages(domain, address, size);
>
> 		/* Wait until IOMMU TLB and all device IOTLB flushes are complete */
>-- 
>2.54.0.rc0.605.g598a273b03-goog
>
>

Reviewed-by: Samiullah Khawaja <skhawaja@google.com>

[PATCH v1 1/1] iommu/amd: Don't split flush for amd_iommu_domain_flush_all()

[Weinan Liu]

We have observed multiple full invalidations occurring during device
detach when we are done using the vfio-device.

blocked_domain_attach_device()
  -> detach_device()
    -> amd_iommu_domain_flush_all()
      -> amd_iommu_domain_flush_pages(..., CMD_INV_IOMMU_ALL_PAGES_ADDRESS)

      	while (size != 0) {

          -> __domain_flush_pages( flush_size /* power of 2 flush_size */)
            -> domain_flush_pages_v1()
              -> build_inv_iommu_pages()
                -> build_inv_address()

         }

build_inv_address() will trigger a full invalidation  if the chunk
size > (1 << 51). Consequently, the guest will issue multiple full
invalidations for a single call to  amd_iommu_domain_flush_all()

Without this patch, we will see 10 time instead of 1 time full
invalidations for every amd_iommu_domain_flush_all().

Fixes: a270be1b3fdf ("iommu/amd: Use only natural aligned flushes in a VM")

Suggested-by: Josef Bacik <josef@toxicpanda.com>
Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
Signed-off-by: Weinan Liu <wnliu@google.com>
---
 drivers/iommu/amd/iommu.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/drivers/iommu/amd/iommu.c b/drivers/iommu/amd/iommu.c
index 760d5f4623b5..59650a44ab50 100644
--- a/drivers/iommu/amd/iommu.c
+++ b/drivers/iommu/amd/iommu.c
@@ -1769,7 +1769,8 @@ void amd_iommu_domain_flush_pages(struct protection_domain *domain,
 {
 	lockdep_assert_held(&domain->lock);
 
-	if (likely(!amd_iommu_np_cache)) {
+	if (likely(!amd_iommu_np_cache) ||
+		size >= (1ULL<<52)) {
 		__domain_flush_pages(domain, address, size);
 
 		/* Wait until IOMMU TLB and all device IOTLB flushes are complete */
-- 
2.54.0.rc0.605.g598a273b03-goog

Re: [PATCH v1 1/1] iommu/amd: Don't split flush for amd_iommu_domain_flush_all()

[Jason Gunthorpe]

On Wed, Apr 15, 2026 at 12:30:46AM +0000, Weinan Liu wrote:
> We have observed multiple full invalidations occurring during device
> detach when we are done using the vfio-device.
> 
> blocked_domain_attach_device()
>   -> detach_device()
>     -> amd_iommu_domain_flush_all()
>       -> amd_iommu_domain_flush_pages(..., CMD_INV_IOMMU_ALL_PAGES_ADDRESS)
> 
>       	while (size != 0) {
> 
>           -> __domain_flush_pages( flush_size /* power of 2 flush_size */)
>             -> domain_flush_pages_v1()
>               -> build_inv_iommu_pages()
>                 -> build_inv_address()
> 
>          }
> 
> build_inv_address() will trigger a full invalidation  if the chunk
> size > (1 << 51). Consequently, the guest will issue multiple full
> invalidations for a single call to  amd_iommu_domain_flush_all()
> 
> Without this patch, we will see 10 time instead of 1 time full
> invalidations for every amd_iommu_domain_flush_all().
> 
> Fixes: a270be1b3fdf ("iommu/amd: Use only natural aligned flushes in a VM")
> 
> Suggested-by: Josef Bacik <josef@toxicpanda.com>
> Suggested-by: Jason Gunthorpe <jgg@nvidia.com>
> Signed-off-by: Weinan Liu <wnliu@google.com>
> ---
>  drivers/iommu/amd/iommu.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)

Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>

Jason