Re: [PATCH] powerpc/powernv/iommu: iommu incorrectly bypass DMA APIs

From: Shivaprasad G Bhat

Date: Fri Apr 03 2026 - 23:53:44 EST

On 4/1/26 6:35 AM, Ritesh Harjani (IBM) wrote:

++CC few people who would be interested in this fix.

Gaurav Batra <gbatra@xxxxxxxxxxxxx> writes:

In a PowerNV environment, for devices that supports DMA mask less than
64 bit but larger than 32 bits, iommu is incorrectly bypassing DMA
APIs while allocating and mapping buffers for DMA operations.

Devices are failing with ENOMEN during probe with the following messages

amdgpu 0000:01:00.0: [drm] Detected VRAM RAM=4096M, BAR=4096M
amdgpu 0000:01:00.0: [drm] RAM width 128bits GDDR5
amdgpu 0000:01:00.0: iommu: 64-bit OK but direct DMA is limited by 0
amdgpu 0000:01:00.0: dma_iommu_get_required_mask: returning bypass mask 0xfffffffffffffff
amdgpu 0000:01:00.0: 4096M of VRAM memory ready
amdgpu 0000:01:00.0: 32570M of GTT memory ready.
amdgpu 0000:01:00.0: (-12) failed to allocate kernel bo
amdgpu 0000:01:00.0: [drm] Debug VRAM access will use slowpath MM access
amdgpu 0000:01:00.0: [drm] GART: num cpu pages 4096, num gpu pages 65536
amdgpu 0000:01:00.0: [drm] PCIE GART of 256M enabled (table at 0x000000F4FFF80000).
amdgpu 0000:01:00.0: (-12) failed to allocate kernel bo
amdgpu 0000:01:00.0: (-12) create WB bo failed
amdgpu 0000:01:00.0: amdgpu_device_wb_init failed -12
amdgpu 0000:01:00.0: amdgpu_device_ip_init failed
amdgpu 0000:01:00.0: Fatal error during GPU init
amdgpu 0000:01:00.0: finishing device.
amdgpu 0000:01:00.0: probe with driver amdgpu failed with error -12
amdgpu 0000:01:00.0: ttm finalized

Fixes: 1471c517cf7d ("powerpc/iommu: bypass DMA APIs for coherent allocations for pre-mapped memory")
Reported-by: Dan Horák <dan@xxxxxxxx>
Closes: https://gitlab.freedesktop.org/drm/amd/-/work_items/5039

We could even add the lore link so that in future people can find the
discussion.
Closes: https://lore.kernel.org/linuxppc-dev/20260313142351.609bc4c3efe1184f64ca5f44@xxxxxxxx/

Tested-by: Dan Horak <dan@xxxxxxxx>
Signed-off-by: Ritesh Harjani <ritesh.list@xxxxxxxxx>

Feel free to change this to the following, I mainly only suggested the fix :)

Suggested-by: Ritesh Harjani (IBM) <ritesh.list@xxxxxxxxx>

Signed-off-by: Gaurav Batra <gbatra@xxxxxxxxxxxxx>
---

Generally this info...

I am working on testing the patch in an LPAR with AMDGPU. I will update the
results soon.

... goes below the three dashes in here ^^^
This is the same place where people also update the patch change log.

But sure, thanks for updating. As for this patch, it looks good to me.
So, feel free to add:

Reviewed-by: Ritesh Harjani (IBM) <ritesh.list@xxxxxxxxx>

arch/powerpc/kernel/dma-iommu.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/arch/powerpc/kernel/dma-iommu.c b/arch/powerpc/kernel/dma-iommu.c
index 73e10bd4d56d..8b4de508d2eb 100644
--- a/arch/powerpc/kernel/dma-iommu.c
+++ b/arch/powerpc/kernel/dma-iommu.c
@@ -67,7 +67,7 @@ bool arch_dma_unmap_sg_direct(struct device *dev, struct scatterlist *sg,
}
bool arch_dma_alloc_direct(struct device *dev)
{
- if (dev->dma_ops_bypass)
+ if (dev->dma_ops_bypass && dev->bus_dma_limit)
return true;
return false;
@@ -75,7 +75,7 @@ bool arch_dma_alloc_direct(struct device *dev)
bool arch_dma_free_direct(struct device *dev, dma_addr_t dma_handle)
{
- if (!dev->dma_ops_bypass)
+ if (!dev->dma_ops_bypass || !dev->bus_dma_limit)
return false;

While this works, looking more on why the 1471c517cf7dae needed arch_dma_[alloc|free]_direct()

functions was because dev->bus_dma_limit being not updated when the new memory pages were

getting added in add_pages() case leading to the can_map_direct() checks failing later when needed.

Looks like these functions were modeled after arch_dma_[map|unmap]_phys_direct() which were

introduced for handing a similar need.

Do we get iommu_mem_notifier() call for action MEM_GOING_ONLINE in device memory being

made online case? If yes, we can just update the bus_dma_limit for each device in the iommu

group to include the new max_pfn. This makes it sound wrong as the pages though configured

as system memory, are still device private and the memory notifiers are for the real memory

hotplug/unplug cases.

In the commit 7170130e4c below, Balibir does point this out to be a problem and prevents updating

max_pfn for device private memory. He has mentioned it to be considered for PPC too.

The dma_addressing_limited() of x86 is same as PPC can_map_direct() and we know why

arch_dma_[alloc|free]_direct() would not be needed anymore.

commit 7170130e4c72ce0caa0cb42a1627c635cc262821
Author: Balbir Singh <balbirs@xxxxxxxxxx>
Date: Tue Apr 1 11:07:52 2025 +1100

x86/mm/init: Handle the special case of device private pages in add_pages(), to not increase max_pfn and trigger dma_addressing_limited() bounce buffers

So, I think we need to bring the changes from 7170130e4c72ce0c into PPC.

With that, the original hunk setting the bypass unconditionally in 1471c517cf7dae can be

changed like below, and remove the redundant arch_dma_[alloc|free]_direct() functions.

diff --git a/arch/powerpc/kernel/dma-iommu.c b/arch/powerpc/kernel/dma-iommu.c
index aa3689d61917..120800b962cf 100644
--- a/arch/powerpc/kernel/dma-iommu.c
+++ b/arch/powerpc/kernel/dma-iommu.c
@@ -150,7 +150,7 @@ int dma_iommu_dma_supported(struct device *dev, u64 mask)
* 1:1 mapping but it is somehow limited.
* ibm,pmemory is one example.
*/
- dev->dma_ops_bypass = dev->bus_dma_limit == 0;
+ dev->dma_ops_bypass = (dev->bus_dma_limit == 0) || (min_not_zero(mask, dev->bus_dma_limit) >= dma_direct_get_required_mask(dev));
if (!dev->dma_ops_bypass)
dev_warn(dev,
"iommu: 64-bit OK but direct DMA is limited by %llx\n",

Suggested-by: Shivaprasad G Bhat <sbhat@xxxxxxxxxxxxx>

Thanks,

Shivaprasad