[PATCH] amba: Fix driver_managed_dma check

[PATCH] amba: Fix driver_managed_dma check

[Robin Murphy]

Since it's not currently safe to take device_lock() in the IOMMU probe
path, that can race against really_probe() setting dev->driver before
attempting to bind. The race itself isn't so bad, since we're only
concerned with dereferencing dev->driver itself anyway, but sadly my
attempt to implement the check with minimal churn leads to a kind of
TOCTOU issue, where dev->driver becomes valid after to_amba_driver(NULL)
is already computed, and thus the check fails to work as intended.

Will and I both hit this with the platform bus, but the pattern here is
the same, so fix it for correctness too.

Reported-by: Will McVicker <willmcvicker@google.com>
Fixes: bcb81ac6ae3c ("iommu: Get DT/ACPI parsing into the proper probe path")
Signed-off-by: Robin Murphy <robin.murphy@arm.com>
---
 drivers/amba/bus.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/amba/bus.c b/drivers/amba/bus.c
index 71482d639a6d..84bc788663e6 100644
--- a/drivers/amba/bus.c
+++ b/drivers/amba/bus.c
@@ -353,7 +353,7 @@ static void amba_shutdown(struct device *dev)
 
 static int amba_dma_configure(struct device *dev)
 {
-	struct amba_driver *drv = to_amba_driver(dev->driver);
+	const struct device_driver *drv = READ_ONCE(dev->driver);
 	enum dev_dma_attr attr;
 	int ret = 0;
 
@@ -365,7 +365,7 @@ static int amba_dma_configure(struct device *dev)
 	}
 
 	/* @drv may not be valid when we're called from the IOMMU layer */
-	if (!ret && dev->driver && !drv->driver_managed_dma) {
+	if (!ret && drv && !to_amba_driver(drv)->driver_managed_dma) {
 		ret = iommu_device_use_default_domain(dev);
 		if (ret)
 			arch_teardown_dma_ops(dev);
-- 
2.39.2.101.g768bb238c484.dirty

Re: [PATCH] amba: Fix driver_managed_dma check

[William McVicker]

On 04/25/2025, Robin Murphy wrote:
> Since it's not currently safe to take device_lock() in the IOMMU probe
> path, that can race against really_probe() setting dev->driver before
> attempting to bind. The race itself isn't so bad, since we're only
> concerned with dereferencing dev->driver itself anyway, but sadly my
> attempt to implement the check with minimal churn leads to a kind of
> TOCTOU issue, where dev->driver becomes valid after to_amba_driver(NULL)
> is already computed, and thus the check fails to work as intended.
> 
> Will and I both hit this with the platform bus, but the pattern here is
> the same, so fix it for correctness too.

Thanks!

Reviewed-by: Will McVicker <willmcvicker@google.com>

> 
> Reported-by: Will McVicker <willmcvicker@google.com>
> Fixes: bcb81ac6ae3c ("iommu: Get DT/ACPI parsing into the proper probe path")
> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
> ---
>  drivers/amba/bus.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/amba/bus.c b/drivers/amba/bus.c
> index 71482d639a6d..84bc788663e6 100644
> --- a/drivers/amba/bus.c
> +++ b/drivers/amba/bus.c
> @@ -353,7 +353,7 @@ static void amba_shutdown(struct device *dev)
>  
>  static int amba_dma_configure(struct device *dev)
>  {
> -	struct amba_driver *drv = to_amba_driver(dev->driver);
> +	const struct device_driver *drv = READ_ONCE(dev->driver);
>  	enum dev_dma_attr attr;
>  	int ret = 0;
>  
> @@ -365,7 +365,7 @@ static int amba_dma_configure(struct device *dev)
>  	}
>  
>  	/* @drv may not be valid when we're called from the IOMMU layer */
> -	if (!ret && dev->driver && !drv->driver_managed_dma) {
> +	if (!ret && drv && !to_amba_driver(drv)->driver_managed_dma) {
>  		ret = iommu_device_use_default_domain(dev);
>  		if (ret)
>  			arch_teardown_dma_ops(dev);
> -- 
> 2.39.2.101.g768bb238c484.dirty
> 

Re: [PATCH] amba: Fix driver_managed_dma check

[Ketan Kishore]

On 4/25/2025 7:09 PM, Robin Murphy wrote:
> Since it's not currently safe to take device_lock() in the IOMMU probe
> path, that can race against really_probe() setting dev->driver before
> attempting to bind. The race itself isn't so bad, since we're only
> concerned with dereferencing dev->driver itself anyway, but sadly my
> attempt to implement the check with minimal churn leads to a kind of
> TOCTOU issue, where dev->driver becomes valid after to_amba_driver(NULL)
> is already computed, and thus the check fails to work as intended.
> 
> Will and I both hit this with the platform bus, but the pattern here is
> the same, so fix it for correctness too.
> 
> Reported-by: Will McVicker <willmcvicker@google.com>
> Fixes: bcb81ac6ae3c ("iommu: Get DT/ACPI parsing into the proper probe path")
> Signed-off-by: Robin Murphy <robin.murphy@arm.com>
			arch_teardown_dma_ops(dev);

Hello Robin

We faced this issue with amba_dma_configure race internally as well:
Call trace:
   amba_dma_configure
   __iommu_probe_device
   probe_iommu_group
   bus_for_each_dev
   iommu_device_register
   arm_smmu_device_probe
   platform_probe
   really_probe
   __driver_probe_device
   driver_probe_device
   __device_attach_driver
   bus_for_each_drv
   __device_attach
   device_initial_probe
   bus_probe_device
   deferred_probe_work_func
   process_scheduled_works
   worker_thread
   kthread
   ret_from_fork

We raised change for this based on already merged changes that fixed the 
same TOCTOU race for the platform bus in
commit 95deee37a123 ("platform: Fix race condition during DMA
configure at IOMMU probe time") and for fsl-mc in commit 152f33ee30ee
("bus: fsl_mc: Fix driver_managed_dma check").
Link: 
https://lore.kernel.org/all/20260825-iommu_races-v2-1-49de6de71357@oss.qualcomm.com/

We later found this change you raised which was never merged.
Can you please help to revive and take this change forward for merge.

Reviewed-by: Ketan Kishore <ketan.kishore@oss.qualcomm.com>

Regards
Ketan