From: Thierry Reding <redacted>
An unfortunate interaction between the 32-bit ARM DMA/IOMMU mapping code
and Tegra SMMU driver changes to support IOMMU groups introduced a boot-
time regression on Tegra124. This was caught very late because none of
the standard configurations that are tested on Tegra enable the ARM DMA/
IOMMU mapping code since it is not needed.
The reason for the failure is that the GPU found on Tegra uses a special
bit in physical addresses to determine whether or not a buffer is mapped
through the SMMU. In order to achieve this, the Nouveau driver needs to
explicitly understand which buffers are mapped through the SMMU and
which aren't. Hiding usage of the SMMU behind the DMA API is bound to
fail because the knowledge doesn't exist. Furthermore, the GPU has its
own IOMMU and in most cases doesn't need buffers to be physically or
virtually contiguous. One notable exception is for compressible buffers
which need to be mapped with large pages, which in turn require all the
small pages in a large page to be contiguous. This can be achieved with
an SMMU mapping, though it isn't currently supported in Nouveau. Since
Translating through the SMMU is unnecessary and can have a negative
impact on performance for the common case, so we want to avoid it when
possible.
This series of patches adds a 32-bit ARM specific API that allows a
driver to detach the device from the DMA/IOMMU mapping so that it can
provide its own implementation for dealing with the SMMU. The second
patch makes use of that new API in the Nouveau driver to fix the
regression.
Thierry
Thierry Reding (2):
ARM: dma-mapping: Implement arm_dma_iommu_detach_device()
drm/nouveau: tegra: Detach from ARM DMA/IOMMU mapping
arch/arm/include/asm/dma-mapping.h | 3 +++
arch/arm/mm/dma-mapping-nommu.c | 4 ++++
arch/arm/mm/dma-mapping.c | 16 ++++++++++++++++
.../gpu/drm/nouveau/nvkm/engine/device/tegra.c | 5 +++++
4 files changed, 28 insertions(+)
--
2.17.0
From: Thierry Reding <redacted>
Implement this function to enable drivers from detaching from any IOMMU
domains that architecture code might have attached them to so that they
can take exclusive control of the IOMMU via the IOMMU API.
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- make API 32-bit ARM specific
- avoid extra local variable
Changes in v2:
- fix compilation
arch/arm/include/asm/dma-mapping.h | 3 +++
arch/arm/mm/dma-mapping-nommu.c | 4 ++++
arch/arm/mm/dma-mapping.c | 16 ++++++++++++++++
3 files changed, 23 insertions(+)
@@ -103,6 +103,9 @@ extern void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,#define arch_teardown_dma_ops arch_teardown_dma_opsexternvoidarch_teardown_dma_ops(structdevice*dev);+#define arm_dma_iommu_detach_device arm_dma_iommu_detach_device+externvoidarm_dma_iommu_detach_device(structdevice*dev);+/* do not use this function in a driver */staticinlineboolis_device_dma_coherent(structdevice*dev){
From: Thierry Reding <redacted>
Implement this function to enable drivers from detaching from any IOMMU
domains that architecture code might have attached them to so that they
can take exclusive control of the IOMMU via the IOMMU API.
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- make API 32-bit ARM specific
- avoid extra local variable
Changes in v2:
- fix compilation
arch/arm/include/asm/dma-mapping.h | 3 +++
arch/arm/mm/dma-mapping-nommu.c | 4 ++++
arch/arm/mm/dma-mapping.c | 16 ++++++++++++++++
3 files changed, 23 insertions(+)
@@ -103,6 +103,9 @@ extern void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,#define arch_teardown_dma_ops arch_teardown_dma_opsexternvoidarch_teardown_dma_ops(structdevice*dev);+#define arm_dma_iommu_detach_device arm_dma_iommu_detach_device+externvoidarm_dma_iommu_detach_device(structdevice*dev);+/* do not use this function in a driver */staticinlineboolis_device_dma_coherent(structdevice*dev){
Potentially freeing the mapping before you try to operate on it is never
the best idea. Plus arm_iommu_detach_device() already releases a
reference appropriately anyway, so it's a double-free.
I really don't see why we need an extra function that essentially just
duplicates arm_iommu_detach_device(). The only real difference here is
that here you reset the DMA ops more appropriately, but I think the
existing function should be fixed to do that anyway, since
set_dma_ops(dev, NULL) now just behaves as an unconditional fallback to
the noncoherent arm_dma_ops, which clearly isn't always right.
Robin.
On Wed, May 30, 2018 at 10:59:30AM +0100, Robin Murphy wrote:
On 30/05/18 09:03, Thierry Reding wrote:
quoted
From: Thierry Reding <redacted>
Implement this function to enable drivers from detaching from any IOMMU
domains that architecture code might have attached them to so that they
can take exclusive control of the IOMMU via the IOMMU API.
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- make API 32-bit ARM specific
- avoid extra local variable
Changes in v2:
- fix compilation
arch/arm/include/asm/dma-mapping.h | 3 +++
arch/arm/mm/dma-mapping-nommu.c | 4 ++++
arch/arm/mm/dma-mapping.c | 16 ++++++++++++++++
3 files changed, 23 insertions(+)
@@ -103,6 +103,9 @@ extern void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,#define arch_teardown_dma_ops arch_teardown_dma_opsexternvoidarch_teardown_dma_ops(structdevice*dev);+#define arm_dma_iommu_detach_device arm_dma_iommu_detach_device+externvoidarm_dma_iommu_detach_device(structdevice*dev);+/* do not use this function in a driver */staticinlineboolis_device_dma_coherent(structdevice*dev){
Potentially freeing the mapping before you try to operate on it is never the
best idea. Plus arm_iommu_detach_device() already releases a reference
appropriately anyway, so it's a double-free.
But the reference released by arm_iommu_detach_device() is to balance
out the reference acquired by arm_iommu_attach_device(), isn't it? In
the above, the arm_iommu_release_mapping() is supposed to drop the
final reference which was obtained by arm_iommu_create_mapping(). The
mapping shouldn't go away irrespective of the order in which these
will be called.
I really don't see why we need an extra function that essentially just
duplicates arm_iommu_detach_device(). The only real difference here is that
here you reset the DMA ops more appropriately, but I think the existing
function should be fixed to do that anyway, since set_dma_ops(dev, NULL) now
just behaves as an unconditional fallback to the noncoherent arm_dma_ops,
which clearly isn't always right.
The idea behind making this an extra function is that we can call it
unconditionally and it will do the right things. Granted, that already
doesn't quite work as elegantly anymore as I had hoped since this is
now an ARM specific function, so we need an #ifdef guard anyway.
I don't care very strongly either way, so if you and Christoph can both
agree that we just want arm_iommu_detach_device() to call the proper
variant of set_dma_ops(), that's fine with me, too.
Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20180530/38fbec80/attachment-0001.sig>
On Wed, May 30, 2018 at 02:54:46PM +0200, Thierry Reding wrote:
On Wed, May 30, 2018 at 10:59:30AM +0100, Robin Murphy wrote:
quoted
On 30/05/18 09:03, Thierry Reding wrote:
quoted
From: Thierry Reding <redacted>
Implement this function to enable drivers from detaching from any IOMMU
domains that architecture code might have attached them to so that they
can take exclusive control of the IOMMU via the IOMMU API.
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- make API 32-bit ARM specific
- avoid extra local variable
Changes in v2:
- fix compilation
arch/arm/include/asm/dma-mapping.h | 3 +++
arch/arm/mm/dma-mapping-nommu.c | 4 ++++
arch/arm/mm/dma-mapping.c | 16 ++++++++++++++++
3 files changed, 23 insertions(+)
@@ -103,6 +103,9 @@ extern void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,#define arch_teardown_dma_ops arch_teardown_dma_opsexternvoidarch_teardown_dma_ops(structdevice*dev);+#define arm_dma_iommu_detach_device arm_dma_iommu_detach_device+externvoidarm_dma_iommu_detach_device(structdevice*dev);+/* do not use this function in a driver */staticinlineboolis_device_dma_coherent(structdevice*dev){
Potentially freeing the mapping before you try to operate on it is never the
best idea. Plus arm_iommu_detach_device() already releases a reference
appropriately anyway, so it's a double-free.
But the reference released by arm_iommu_detach_device() is to balance
out the reference acquired by arm_iommu_attach_device(), isn't it? In
the above, the arm_iommu_release_mapping() is supposed to drop the
final reference which was obtained by arm_iommu_create_mapping(). The
mapping shouldn't go away irrespective of the order in which these
will be called.
Going over the DMA/IOMMU code I just remembered that I drew inspiration
from arm_teardown_iommu_dma_ops() for the initial proposal which also
calls both arm_iommu_detach_device() and arm_iommu_release_mapping().
That said, one other possibility to implement this would be to export
the 32-bit and 64-bit ARM implementations of arch_teardown_dma_ops()
and use that instead. linux/dma-mapping.h implements a stub for
architectures that don't provide one, so it should work without any
#ifdef guards.
That combined with the set_dma_ops() fix in arm_iommu_detach_device()
should fix this pretty nicely.
Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20180530/64c953e2/attachment.sig>
On Wed, May 30, 2018 at 02:54:46PM +0200, Thierry Reding wrote:
quoted
On Wed, May 30, 2018 at 10:59:30AM +0100, Robin Murphy wrote:
quoted
On 30/05/18 09:03, Thierry Reding wrote:
quoted
From: Thierry Reding <redacted>
Implement this function to enable drivers from detaching from any IOMMU
domains that architecture code might have attached them to so that they
can take exclusive control of the IOMMU via the IOMMU API.
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- make API 32-bit ARM specific
- avoid extra local variable
Changes in v2:
- fix compilation
arch/arm/include/asm/dma-mapping.h | 3 +++
arch/arm/mm/dma-mapping-nommu.c | 4 ++++
arch/arm/mm/dma-mapping.c | 16 ++++++++++++++++
3 files changed, 23 insertions(+)
@@ -103,6 +103,9 @@ extern void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,#define arch_teardown_dma_ops arch_teardown_dma_opsexternvoidarch_teardown_dma_ops(structdevice*dev);+#define arm_dma_iommu_detach_device arm_dma_iommu_detach_device+externvoidarm_dma_iommu_detach_device(structdevice*dev);+/* do not use this function in a driver */staticinlineboolis_device_dma_coherent(structdevice*dev){
Potentially freeing the mapping before you try to operate on it is never the
best idea. Plus arm_iommu_detach_device() already releases a reference
appropriately anyway, so it's a double-free.
But the reference released by arm_iommu_detach_device() is to balance
out the reference acquired by arm_iommu_attach_device(), isn't it? In
the above, the arm_iommu_release_mapping() is supposed to drop the
final reference which was obtained by arm_iommu_create_mapping(). The
mapping shouldn't go away irrespective of the order in which these
will be called.
Going over the DMA/IOMMU code I just remembered that I drew inspiration
from arm_teardown_iommu_dma_ops() for the initial proposal which also
calls both arm_iommu_detach_device() and arm_iommu_release_mapping().
That said, one other possibility to implement this would be to export
the 32-bit and 64-bit ARM implementations of arch_teardown_dma_ops()
and use that instead. linux/dma-mapping.h implements a stub for
architectures that don't provide one, so it should work without any
#ifdef guards.
That combined with the set_dma_ops() fix in arm_iommu_detach_device()
should fix this pretty nicely.
OK, having a second look at the ARM code I see I had indeed overlooked
that extra reference held until arm_teardown_iommu_dma_ops() - mea culpa
- but frankly that looks wrong anyway, as it basically defeats the point
of refcounting the mapping at all. AFAICS arm_setup_iommu_dma_ops()
should just be made to behave 'normally' by unconditionally dropping the
initial reference after calling __arm_iommu_attach_device(), then we
don't need all these odd and confusing release calls dotted around at all.
Robin.
On Wed, May 30, 2018 at 02:42:50PM +0100, Robin Murphy wrote:
On 30/05/18 14:12, Thierry Reding wrote:
quoted
On Wed, May 30, 2018 at 02:54:46PM +0200, Thierry Reding wrote:
quoted
On Wed, May 30, 2018 at 10:59:30AM +0100, Robin Murphy wrote:
quoted
On 30/05/18 09:03, Thierry Reding wrote:
quoted
From: Thierry Reding <redacted>
Implement this function to enable drivers from detaching from any IOMMU
domains that architecture code might have attached them to so that they
can take exclusive control of the IOMMU via the IOMMU API.
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- make API 32-bit ARM specific
- avoid extra local variable
Changes in v2:
- fix compilation
arch/arm/include/asm/dma-mapping.h | 3 +++
arch/arm/mm/dma-mapping-nommu.c | 4 ++++
arch/arm/mm/dma-mapping.c | 16 ++++++++++++++++
3 files changed, 23 insertions(+)
@@ -103,6 +103,9 @@ extern void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,#define arch_teardown_dma_ops arch_teardown_dma_opsexternvoidarch_teardown_dma_ops(structdevice*dev);+#define arm_dma_iommu_detach_device arm_dma_iommu_detach_device+externvoidarm_dma_iommu_detach_device(structdevice*dev);+/* do not use this function in a driver */staticinlineboolis_device_dma_coherent(structdevice*dev){
Potentially freeing the mapping before you try to operate on it is never the
best idea. Plus arm_iommu_detach_device() already releases a reference
appropriately anyway, so it's a double-free.
But the reference released by arm_iommu_detach_device() is to balance
out the reference acquired by arm_iommu_attach_device(), isn't it? In
the above, the arm_iommu_release_mapping() is supposed to drop the
final reference which was obtained by arm_iommu_create_mapping(). The
mapping shouldn't go away irrespective of the order in which these
will be called.
Going over the DMA/IOMMU code I just remembered that I drew inspiration
from arm_teardown_iommu_dma_ops() for the initial proposal which also
calls both arm_iommu_detach_device() and arm_iommu_release_mapping().
That said, one other possibility to implement this would be to export
the 32-bit and 64-bit ARM implementations of arch_teardown_dma_ops()
and use that instead. linux/dma-mapping.h implements a stub for
architectures that don't provide one, so it should work without any
#ifdef guards.
That combined with the set_dma_ops() fix in arm_iommu_detach_device()
should fix this pretty nicely.
OK, having a second look at the ARM code I see I had indeed overlooked that
extra reference held until arm_teardown_iommu_dma_ops() - mea culpa - but
frankly that looks wrong anyway, as it basically defeats the point of
refcounting the mapping at all. AFAICS arm_setup_iommu_dma_ops() should just
be made to behave 'normally' by unconditionally dropping the initial
reference after calling __arm_iommu_attach_device(), then we don't need all
these odd and confusing release calls dotted around at all.
From: Thierry Reding <redacted>
Depending on the kernel configuration, early ARM architecture setup code
may have attached the GPU to a DMA/IOMMU mapping that transparently uses
the IOMMU to back the DMA API. Tegra requires special handling for IOMMU
backed buffers (a special bit in the GPU's MMU page tables indicates the
memory path to take: via the SMMU or directly to the memory controller).
Transparently backing DMA memory with an IOMMU prevents Nouveau from
properly handling such memory accesses and causes memory access faults.
As a side-note: buffers other than those allocated in instance memory
don't need to be physically contiguous from the GPU's perspective since
the GPU can map them into contiguous buffers using its own MMU. Mapping
these buffers through the IOMMU is unnecessary and will even lead to
performance degradation because of the additional translation. One
exception to this are compressible buffers which need large pages. In
order to enable these large pages, multiple small pages will have to be
combined into one large (I/O virtually contiguous) mapping via the
IOMMU. However, that is a topic outside the scope of this fix and isn't
currently supported. An implementation will want to explicitly create
these large pages in the Nouveau driver, so detaching from a DMA/IOMMU
mapping would still be required.
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- clarify the use of IOMMU mapping for compressible buffers
- squash multiple patches into this
drivers/gpu/drm/nouveau/nvkm/engine/device/tegra.c | 5 +++++
1 file changed, 5 insertions(+)
@@ -105,6 +105,11 @@ nvkm_device_tegra_probe_iommu(struct nvkm_device_tegra *tdev)unsignedlongpgsize_bitmap;intret;+#if IS_ENABLED(CONFIG_ARM)+/* make sure we can use the IOMMU exclusively */+arm_dma_iommu_detach_device(dev);+#endif+if(!tdev->func->iommu_bit)return;
From: Thierry Reding <redacted>
Depending on the kernel configuration, early ARM architecture setup code
may have attached the GPU to a DMA/IOMMU mapping that transparently uses
the IOMMU to back the DMA API. Tegra requires special handling for IOMMU
backed buffers (a special bit in the GPU's MMU page tables indicates the
memory path to take: via the SMMU or directly to the memory controller).
Transparently backing DMA memory with an IOMMU prevents Nouveau from
properly handling such memory accesses and causes memory access faults.
As a side-note: buffers other than those allocated in instance memory
don't need to be physically contiguous from the GPU's perspective since
the GPU can map them into contiguous buffers using its own MMU. Mapping
these buffers through the IOMMU is unnecessary and will even lead to
performance degradation because of the additional translation. One
exception to this are compressible buffers which need large pages. In
order to enable these large pages, multiple small pages will have to be
combined into one large (I/O virtually contiguous) mapping via the
IOMMU. However, that is a topic outside the scope of this fix and isn't
currently supported. An implementation will want to explicitly create
these large pages in the Nouveau driver, so detaching from a DMA/IOMMU
mapping would still be required.
I wonder if it might make sense to have a hook in iommu_attach_device()
to notify the arch DMA API code when moving devices between unmanaged
and DMA ops domains? That seems like it might be the most low-impact way
to address the overall problem long-term.
quoted hunk
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- clarify the use of IOMMU mapping for compressible buffers
- squash multiple patches into this
drivers/gpu/drm/nouveau/nvkm/engine/device/tegra.c | 5 +++++
1 file changed, 5 insertions(+)
Wouldn't CONFIG_ARM_DMA_USE_IOMMU be even more appropriate?
+ /* make sure we can use the IOMMU exclusively */
+ arm_dma_iommu_detach_device(dev);
As before, I would just use the existing infrastructure the same way the
Exynos DRM driver currently does in __exynos_iommu_attach() (albeit
without then reattaching to another DMA ops mapping).
Robin.
On Wed, May 30, 2018 at 11:30:25AM +0100, Robin Murphy wrote:
On 30/05/18 09:03, Thierry Reding wrote:
quoted
From: Thierry Reding <redacted>
Depending on the kernel configuration, early ARM architecture setup code
may have attached the GPU to a DMA/IOMMU mapping that transparently uses
the IOMMU to back the DMA API. Tegra requires special handling for IOMMU
backed buffers (a special bit in the GPU's MMU page tables indicates the
memory path to take: via the SMMU or directly to the memory controller).
Transparently backing DMA memory with an IOMMU prevents Nouveau from
properly handling such memory accesses and causes memory access faults.
As a side-note: buffers other than those allocated in instance memory
don't need to be physically contiguous from the GPU's perspective since
the GPU can map them into contiguous buffers using its own MMU. Mapping
these buffers through the IOMMU is unnecessary and will even lead to
performance degradation because of the additional translation. One
exception to this are compressible buffers which need large pages. In
order to enable these large pages, multiple small pages will have to be
combined into one large (I/O virtually contiguous) mapping via the
IOMMU. However, that is a topic outside the scope of this fix and isn't
currently supported. An implementation will want to explicitly create
these large pages in the Nouveau driver, so detaching from a DMA/IOMMU
mapping would still be required.
I wonder if it might make sense to have a hook in iommu_attach_device() to
notify the arch DMA API code when moving devices between unmanaged and DMA
ops domains? That seems like it might be the most low-impact way to address
the overall problem long-term.
quoted
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- clarify the use of IOMMU mapping for compressible buffers
- squash multiple patches into this
drivers/gpu/drm/nouveau/nvkm/engine/device/tegra.c | 5 +++++
1 file changed, 5 insertions(+)
Wouldn't CONFIG_ARM_DMA_USE_IOMMU be even more appropriate?
Not necessarily. arm_dma_iommu_detach_device() is always defined on ARM,
only with CONFIG_ARM_DMA_USE_IOMMU=n it will be empty. So this check is
a guard to make sure we don't call the function when it isn't available,
but it may still not do anything.
quoted
+ /* make sure we can use the IOMMU exclusively */
+ arm_dma_iommu_detach_device(dev);
As before, I would just use the existing infrastructure the same way the
Exynos DRM driver currently does in __exynos_iommu_attach() (albeit without
then reattaching to another DMA ops mapping).
That's pretty much what I initially did and which was shot down by
Christoph. As I said earlier, at this point I don't really care what
color the shed will be. Can you and Christoph come to an agreement
on what it should be?
Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20180530/b6e5bbcc/attachment.sig>
On Wed, May 30, 2018 at 11:30:25AM +0100, Robin Murphy wrote:
quoted
On 30/05/18 09:03, Thierry Reding wrote:
quoted
From: Thierry Reding <redacted>
Depending on the kernel configuration, early ARM architecture setup code
may have attached the GPU to a DMA/IOMMU mapping that transparently uses
the IOMMU to back the DMA API. Tegra requires special handling for IOMMU
backed buffers (a special bit in the GPU's MMU page tables indicates the
memory path to take: via the SMMU or directly to the memory controller).
Transparently backing DMA memory with an IOMMU prevents Nouveau from
properly handling such memory accesses and causes memory access faults.
As a side-note: buffers other than those allocated in instance memory
don't need to be physically contiguous from the GPU's perspective since
the GPU can map them into contiguous buffers using its own MMU. Mapping
these buffers through the IOMMU is unnecessary and will even lead to
performance degradation because of the additional translation. One
exception to this are compressible buffers which need large pages. In
order to enable these large pages, multiple small pages will have to be
combined into one large (I/O virtually contiguous) mapping via the
IOMMU. However, that is a topic outside the scope of this fix and isn't
currently supported. An implementation will want to explicitly create
these large pages in the Nouveau driver, so detaching from a DMA/IOMMU
mapping would still be required.
I wonder if it might make sense to have a hook in iommu_attach_device() to
notify the arch DMA API code when moving devices between unmanaged and DMA
ops domains? That seems like it might be the most low-impact way to address
the overall problem long-term.
quoted
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- clarify the use of IOMMU mapping for compressible buffers
- squash multiple patches into this
drivers/gpu/drm/nouveau/nvkm/engine/device/tegra.c | 5 +++++
1 file changed, 5 insertions(+)
Wouldn't CONFIG_ARM_DMA_USE_IOMMU be even more appropriate?
Not necessarily. arm_dma_iommu_detach_device() is always defined on ARM,
only with CONFIG_ARM_DMA_USE_IOMMU=n it will be empty. So this check is
a guard to make sure we don't call the function when it isn't available,
but it may still not do anything.
Calling a function under condition A, which only does anything under
condition B, when B depends on A, is identical in behaviour to only
calling the function under condition B, except needlessly harder to follow.
quoted
quoted
+ /* make sure we can use the IOMMU exclusively */
+ arm_dma_iommu_detach_device(dev);
As before, I would just use the existing infrastructure the same way the
Exynos DRM driver currently does in __exynos_iommu_attach() (albeit without
then reattaching to another DMA ops mapping).
That's pretty much what I initially did and which was shot down by
Christoph. As I said earlier, at this point I don't really care what
color the shed will be. Can you and Christoph come to an agreement
on what it should be?
What I was getting at is that arm_iommu_detach_device() already *is* the
exact function Christoph was asking for, it just needs a minor fix
instead of adding explicit set_dma_ops() fiddling at its callsites which
only obfuscates the fact that it's supposed to be responsible for
resetting the device's DMA ops already.
Robin.
On Wed, May 30, 2018 at 02:30:51PM +0100, Robin Murphy wrote:
On 30/05/18 14:00, Thierry Reding wrote:
quoted
On Wed, May 30, 2018 at 11:30:25AM +0100, Robin Murphy wrote:
quoted
On 30/05/18 09:03, Thierry Reding wrote:
quoted
From: Thierry Reding <redacted>
Depending on the kernel configuration, early ARM architecture setup code
may have attached the GPU to a DMA/IOMMU mapping that transparently uses
the IOMMU to back the DMA API. Tegra requires special handling for IOMMU
backed buffers (a special bit in the GPU's MMU page tables indicates the
memory path to take: via the SMMU or directly to the memory controller).
Transparently backing DMA memory with an IOMMU prevents Nouveau from
properly handling such memory accesses and causes memory access faults.
As a side-note: buffers other than those allocated in instance memory
don't need to be physically contiguous from the GPU's perspective since
the GPU can map them into contiguous buffers using its own MMU. Mapping
these buffers through the IOMMU is unnecessary and will even lead to
performance degradation because of the additional translation. One
exception to this are compressible buffers which need large pages. In
order to enable these large pages, multiple small pages will have to be
combined into one large (I/O virtually contiguous) mapping via the
IOMMU. However, that is a topic outside the scope of this fix and isn't
currently supported. An implementation will want to explicitly create
these large pages in the Nouveau driver, so detaching from a DMA/IOMMU
mapping would still be required.
I wonder if it might make sense to have a hook in iommu_attach_device() to
notify the arch DMA API code when moving devices between unmanaged and DMA
ops domains? That seems like it might be the most low-impact way to address
the overall problem long-term.
quoted
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- clarify the use of IOMMU mapping for compressible buffers
- squash multiple patches into this
drivers/gpu/drm/nouveau/nvkm/engine/device/tegra.c | 5 +++++
1 file changed, 5 insertions(+)
Wouldn't CONFIG_ARM_DMA_USE_IOMMU be even more appropriate?
Not necessarily. arm_dma_iommu_detach_device() is always defined on ARM,
only with CONFIG_ARM_DMA_USE_IOMMU=n it will be empty. So this check is
a guard to make sure we don't call the function when it isn't available,
but it may still not do anything.
Calling a function under condition A, which only does anything under
condition B, when B depends on A, is identical in behaviour to only calling
the function under condition B, except needlessly harder to follow.
quoted
quoted
quoted
+ /* make sure we can use the IOMMU exclusively */
+ arm_dma_iommu_detach_device(dev);
As before, I would just use the existing infrastructure the same way the
Exynos DRM driver currently does in __exynos_iommu_attach() (albeit without
then reattaching to another DMA ops mapping).
That's pretty much what I initially did and which was shot down by
Christoph. As I said earlier, at this point I don't really care what
color the shed will be. Can you and Christoph come to an agreement
on what it should be?
What I was getting at is that arm_iommu_detach_device() already *is* the
exact function Christoph was asking for, it just needs a minor fix instead
of adding explicit set_dma_ops() fiddling at its callsites which only
obfuscates the fact that it's supposed to be responsible for resetting the
device's DMA ops already.
It still has the downside of callers having to explicitly check for the
existence of a mapping, otherwise they'll cause a warning to be printed
to the kernel log.
That's not all that bad, though. I'll prepare version 4 with those
changes.
Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20180530/02b467fa/attachment.sig>
On Wed, May 30, 2018 at 02:30:51PM +0100, Robin Murphy wrote:
quoted
On 30/05/18 14:00, Thierry Reding wrote:
quoted
On Wed, May 30, 2018 at 11:30:25AM +0100, Robin Murphy wrote:
quoted
On 30/05/18 09:03, Thierry Reding wrote:
quoted
From: Thierry Reding <redacted>
Depending on the kernel configuration, early ARM architecture setup code
may have attached the GPU to a DMA/IOMMU mapping that transparently uses
the IOMMU to back the DMA API. Tegra requires special handling for IOMMU
backed buffers (a special bit in the GPU's MMU page tables indicates the
memory path to take: via the SMMU or directly to the memory controller).
Transparently backing DMA memory with an IOMMU prevents Nouveau from
properly handling such memory accesses and causes memory access faults.
As a side-note: buffers other than those allocated in instance memory
don't need to be physically contiguous from the GPU's perspective since
the GPU can map them into contiguous buffers using its own MMU. Mapping
these buffers through the IOMMU is unnecessary and will even lead to
performance degradation because of the additional translation. One
exception to this are compressible buffers which need large pages. In
order to enable these large pages, multiple small pages will have to be
combined into one large (I/O virtually contiguous) mapping via the
IOMMU. However, that is a topic outside the scope of this fix and isn't
currently supported. An implementation will want to explicitly create
these large pages in the Nouveau driver, so detaching from a DMA/IOMMU
mapping would still be required.
I wonder if it might make sense to have a hook in iommu_attach_device() to
notify the arch DMA API code when moving devices between unmanaged and DMA
ops domains? That seems like it might be the most low-impact way to address
the overall problem long-term.
quoted
Signed-off-by: Thierry Reding <redacted>
---
Changes in v3:
- clarify the use of IOMMU mapping for compressible buffers
- squash multiple patches into this
drivers/gpu/drm/nouveau/nvkm/engine/device/tegra.c | 5 +++++
1 file changed, 5 insertions(+)
Wouldn't CONFIG_ARM_DMA_USE_IOMMU be even more appropriate?
Not necessarily. arm_dma_iommu_detach_device() is always defined on ARM,
only with CONFIG_ARM_DMA_USE_IOMMU=n it will be empty. So this check is
a guard to make sure we don't call the function when it isn't available,
but it may still not do anything.
Calling a function under condition A, which only does anything under
condition B, when B depends on A, is identical in behaviour to only calling
the function under condition B, except needlessly harder to follow.
quoted
quoted
quoted
+ /* make sure we can use the IOMMU exclusively */
+ arm_dma_iommu_detach_device(dev);
As before, I would just use the existing infrastructure the same way the
Exynos DRM driver currently does in __exynos_iommu_attach() (albeit without
then reattaching to another DMA ops mapping).
That's pretty much what I initially did and which was shot down by
Christoph. As I said earlier, at this point I don't really care what
color the shed will be. Can you and Christoph come to an agreement
on what it should be?
What I was getting at is that arm_iommu_detach_device() already *is* the
exact function Christoph was asking for, it just needs a minor fix instead
of adding explicit set_dma_ops() fiddling at its callsites which only
obfuscates the fact that it's supposed to be responsible for resetting the
device's DMA ops already.
It still has the downside of callers having to explicitly check for the
existence of a mapping, otherwise they'll cause a warning to be printed
to the kernel log.
Or we could look at the way it's actually used, and reconsider whether
the warning is really appropriate. That's always an option ;)
Robin.
That's not all that bad, though. I'll prepare version 4 with those
changes.
Thierry