Skip to contentSkip to conversation

iommu: handle drivers that manage iommu directly

PATCH In review Rob Clark 4 participants 9 messages Sep 6, 2019 – Sep 17, 2019

Subsystems (6)
lore
Share & tools
lore

Rob Clark

author
#0 lore raw

v3 cover iommu: handle drivers that manage iommu directly

From: Rob Clark <[email protected]>

One of the challenges we have to enable the aarch64 laptops upstream is dealing with the fact that the bootloader enables the display and takes the corresponding SMMU context-bank out of BYPASS. Unfortunately, currently, the IOMMU framework attaches a DMA (or potentially an IDENTITY) domain before the driver is probed and has a chance to intervene and shutdown scanout. Which makes things go horribly wrong.

But in this case, drm/msm is already directly managing it's IOMMUs directly, the DMA API attached iommu_domain simply gets in the way. This series adds a way that a driver can indicate to drivers/iommu that it does not wish to have an DMA managed iommu_domain attached. This way, drm/msm can shut down scanout cleanly before attaching it's own iommu_domain.

NOTE that to get things working with arm-smmu on the aarch64 laptops, you also need a patchset[1] from Bjorn Andersson to inherit SMMU config at boot, when it is already enabled.

[1] https://www.spinics.net/lists/arm-kernel/msg732246.html

NOTE that in discussion of previous revisions, RMRR came up. This is not really a replacement for RMRR (nor does RMRR really provide any more information than we already get from EFI GOP, or DT in the simplefb case). I also don't see how RMRR could help w/ SMMU handover of CB/SMR config (Bjorn's patchset[1]) without defining new tables.

This perhaps doesn't solve the more general case of bootloader enabled display for drivers that actually want to use DMA API managed IOMMU. But it does also happen to avoid a related problem with GPU, caused by the DMA domain claiming the context bank that the GPU firmware expects to use. And it avoids spurious TLB invalidation coming from the unused DMA domain. So IMHO this is a useful and necessary change.

Rob Clark (2):
iommu: add support for drivers that manage iommu explicitly drm/msm: mark devices where iommu is managed by driver

 drivers/gpu/drm/msm/adreno/adreno_device.c | 1 +
 drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c    | 1 +
 drivers/gpu/drm/msm/disp/mdp5/mdp5_kms.c   | 1 +
 drivers/gpu/drm/msm/msm_drv.c              | 1 +
 drivers/iommu/iommu.c                      | 2 +-
 drivers/iommu/of_iommu.c                   | 3 +++
 include/linux/device.h                     | 3 ++-
 7 files changed, 10 insertions(+), 2 deletions(-)
Signature · 6 lines
2.21.0

_______________________________________________
dri-devel mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/dri-devel
author
#1 lore raw

PATCH v2 1/2 iommu: add support for drivers that manage iommu explicitly

From: Rob Clark <[email protected]>

Avoid attaching any non-driver managed domain if the driver indicates that it manages the iommu directly.

---
 drivers/iommu/iommu.c    | 2 +-
 drivers/iommu/of_iommu.c | 3 +++
 include/linux/device.h   | 3 ++-
 3 files changed, 6 insertions(+), 2 deletions(-)
drivers/iommu/iommu.c +1 -1
index 0c674d80c37f..2ac5e8d48cae 100644
@@ -674,7 +674,7 @@ int iommu_group_add_device(struct iommu_group *group, struct device *dev)
 
 	mutex_lock(&group->mutex);
 	list_add_tail(&device->list, &group->devices);
-	if (group->domain)
+	if (group->domain && !(dev->driver && dev->driver->driver_manages_iommu))
 		ret = __iommu_attach_device(group->domain, dev);
 	mutex_unlock(&group->mutex);
 	if (ret)
drivers/iommu/of_iommu.c +3 -0
index 614a93aa5305..62b47e384a77 100644
@@ -221,6 +221,9 @@ const struct iommu_ops *of_iommu_configure(struct device *dev,
 	} else if (err < 0) {
 		dev_dbg(dev, "Adding to IOMMU failed: %d\n", err);
 		ops = NULL;
+	} else if (dev->driver && dev->driver->driver_manages_iommu) {
+		dev_dbg(dev, "Driver manages IOMMU\n");
+		ops = NULL;
 	}
 
 	return ops;
include/linux/device.h +2 -1
index 1aa341b2a0db..b77a11b8d9bb 100644
@@ -284,7 +284,8 @@ struct device_driver {
 	struct module		*owner;
 	const char		*mod_name;	/* used for built-in modules */
 
-	bool suppress_bind_attrs;	/* disables bind/unbind via sysfs */
+	bool suppress_bind_attrs:1;	/* disables bind/unbind via sysfs */
+	bool driver_manages_iommu:1;	/* driver manages IOMMU explicitly */
 	enum probe_type probe_type;
 
 	const struct of_device_id	*of_match_table;
Signature · 1 line
2.21.0
maintainer
#2 lore raw

Re: [PATCH v2 1/2] iommu: add support for drivers that manage iommu explicitly

drivers/iommu/iommu.c int iommu_group_add_device(struct iommu_group *group, struct device *dev) · line 677

Quoted 6 lines · Rob Clark
drivers/iommu/iommu.c +1 -1
@@ -674,7 +674,7 @@ int iommu_group_add_device(struct iommu_group *group, struct device *dev)
 
 	mutex_lock(&group->mutex);
 	list_add_tail(&device->list, &group->devices);
-	if (group->domain)
+	if (group->domain && !(dev->driver && dev->driver->driver_manages_iommu))

Hmm, this code usually runs at enumeration time when no driver is attached to the device. Actually it would be pretty dangerous when this code runs while a driver is attached to the device. How does that change make things work for you?

Regards,

Joerg

author
#3 lore raw

Re: [PATCH v2 1/2] iommu: add support for drivers that manage iommu explicitly

Quoted 4 lines · Joerg Roedel
Quoted 2 lines · Rob Clark
(quoted diff) +0 -0
@@ -674,7 +674,7 @@ int iommu_group_add_device(struct iommu_group *group, struct device *dev)
 
Quoted 10 lines

drivers/iommu/iommu.c int iommu_group_add_device(struct iommu_group *group, struct device *dev) · line 677

Quoted 4 lines
drivers/iommu/iommu.c +1 -1
      mutex_lock(&group->mutex);
      list_add_tail(&device->list, &group->devices);
-     if (group->domain)
+     if (group->domain && !(dev->driver && dev->driver->driver_manages_iommu))

Hmm, this code usually runs at enumeration time when no driver is attached to the device. Actually it would be pretty dangerous when this code runs while a driver is attached to the device. How does that change make things work for you?

I was seeing this get called via the path driver_probe_device() -> platform_dma_configure() -> of_dma_configure() -> of_iommu_configure() -> iommu_probe_device() -> ...

The only cases I was seeing where dev->driver is NULL where a few places that drivers call of_dma_configure() on their own sub-devices. But maybe there are some other paths that I did not notice?

BR,
-R

reviewer
#4 lore raw

Re: [PATCH v2 1/2] iommu: add support for drivers that manage iommu explicitly

Quoted 16 lines · Rob Clark
Quoted 4 lines · Joerg Roedel
Quoted 2 lines · Rob Clark
(quoted diff) +0 -0
@@ -674,7 +674,7 @@ int iommu_group_add_device(struct iommu_group *group, struct device *dev)
 
Quoted 10 lines

drivers/iommu/iommu.c int iommu_group_add_device(struct iommu_group *group, struct device *dev) · line 677

Quoted 4 lines
drivers/iommu/iommu.c +1 -1
       mutex_lock(&group->mutex);
       list_add_tail(&device->list, &group->devices);
-     if (group->domain)
+     if (group->domain && !(dev->driver && dev->driver->driver_manages_iommu))

Hmm, this code usually runs at enumeration time when no driver is attached to the device. Actually it would be pretty dangerous when this code runs while a driver is attached to the device. How does that change make things work for you?

Quoted 7 lines

I was seeing this get called via the path driver_probe_device() -> platform_dma_configure() -> of_dma_configure() -> of_iommu_configure() -> iommu_probe_device() -> ...

The only cases I was seeing where dev->driver is NULL where a few places that drivers call of_dma_configure() on their own sub-devices. But maybe there are some other paths that I did not notice?

For the of_iommu flow, it very much depends on your DT layout and driver probe order as to whether you catch the "proper" call from iommu_bus_notifier()/iommu_bus_init() or the late "replay" from of_iommu_configure(). I wouldn't make any assumptions of consistency.

Robin.

author
#5 lore raw

PATCH v2 2/2 drm/msm: mark devices where iommu is managed by driver

In reply to Rob Clark

From: Rob Clark <[email protected]>

---
 drivers/gpu/drm/msm/adreno/adreno_device.c | 1 +
 drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c    | 1 +
 drivers/gpu/drm/msm/disp/mdp5/mdp5_kms.c   | 1 +
 drivers/gpu/drm/msm/msm_drv.c              | 1 +
 4 files changed, 4 insertions(+)
drivers/gpu/drm/msm/adreno/adreno_device.c +1 -0
index 7f750a9510a5..19f2bd2d6cb4 100644
@@ -452,6 +452,7 @@ static struct platform_driver adreno_driver = {
 		.name = "adreno",
 		.of_match_table = dt_match,
 		.pm = &adreno_pm_ops,
+		.driver_manages_iommu = true,
 	},
 };
 
drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c +1 -0
index 5751815a26d7..dec8cc6b64dc 100644
@@ -1094,6 +1094,7 @@ static struct platform_driver dpu_driver = {
 		.name = "msm_dpu",
 		.of_match_table = dpu_dt_match,
 		.pm = &dpu_pm_ops,
+		.driver_manages_iommu = true,
 	},
 };
 
drivers/gpu/drm/msm/disp/mdp5/mdp5_kms.c +1 -0
index d93de3a569b4..eff1b000258e 100644
@@ -1134,6 +1134,7 @@ static struct platform_driver mdp5_driver = {
 		.name = "msm_mdp",
 		.of_match_table = mdp5_dt_match,
 		.pm = &mdp5_pm_ops,
+		.driver_manages_iommu = true,
 	},
 };
 
drivers/gpu/drm/msm/msm_drv.c +1 -0
index 3a4fd20a33e8..336a6d0a4cd3 100644
@@ -1388,6 +1388,7 @@ static struct platform_driver msm_platform_driver = {
 		.name   = "msm",
 		.of_match_table = dt_match,
 		.pm     = &msm_pm_ops,
+		.driver_manages_iommu = true,
 	},
 };
 
Signature · 6 lines
2.21.0

_______________________________________________
Freedreno mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/freedreno
reviewer
#6 lore raw
In reply to Rob Clark
Quoted 8 lines · Rob Clark

From: Rob Clark <[email protected]>

One of the challenges we have to enable the aarch64 laptops upstream is dealing with the fact that the bootloader enables the display and takes the corresponding SMMU context-bank out of BYPASS. Unfortunately, currently, the IOMMU framework attaches a DMA (or potentially an IDENTITY) domain before the driver is probed and has a chance to intervene and shutdown scanout. Which makes things go horribly wrong.

Nope, things already went horribly wrong in arm_smmu_device_reset() - sure, sometimes for some configurations it might *seem* like they didn't and that you can fudge the context bank state at arm's length from core code later, but the truth is that impl->cfg_probe is your last chance to guarantee that any necessary SMMU state is preserved.

The remainder of the problem involves reworking default domain allocation such that we can converge on what iommu_request_dm_for_dev() currently does but without the momentary attachment to a translation domain to cause hiccups. That's starting here:

https://lore.kernel.org/linux-iommu/cover.1566353521.git.sai.praneeth.prakhya-ral2JQCrhuEAvxtiuMwx3w@public.gmane.org/

Quoted 11 lines

But in this case, drm/msm is already directly managing it's IOMMUs directly, the DMA API attached iommu_domain simply gets in the way. This series adds a way that a driver can indicate to drivers/iommu that it does not wish to have an DMA managed iommu_domain attached. This way, drm/msm can shut down scanout cleanly before attaching it's own iommu_domain.

NOTE that to get things working with arm-smmu on the aarch64 laptops, you also need a patchset[1] from Bjorn Andersson to inherit SMMU config at boot, when it is already enabled.

Quoted 7 lines

[1] https://www.spinics.net/lists/arm-kernel/msg732246.html

NOTE that in discussion of previous revisions, RMRR came up. This is not really a replacement for RMRR (nor does RMRR really provide any more information than we already get from EFI GOP, or DT in the simplefb case). I also don't see how RMRR could help w/ SMMU handover of CB/SMR config (Bjorn's patchset[1]) without defining new tables.

The point of RMRR-like-things is that they identify not just the memory region but also the specific device accessing them, which means the IOMMU driver knows up-front which IDs etc. it must be careful not to disrupt. Obviously for SMMU that *would* be some new table (designed to encompass everything relevant) since literal RMRRs are specifically an Intel VT-d thing.

Quoted 5 lines

This perhaps doesn't solve the more general case of bootloader enabled display for drivers that actually want to use DMA API managed IOMMU. But it does also happen to avoid a related problem with GPU, caused by the DMA domain claiming the context bank that the GPU firmware expects to use.

Careful bringing that up again, or I really will rework the context bank allocator to avoid this default domain problem entirely... ;)

Robin.

Quoted 16 lines

And it avoids spurious TLB invalidation coming from the unused
DMA domain. So IMHO this is a useful and necessary change.

Rob Clark (2):
iommu: add support for drivers that manage iommu explicitly drm/msm: mark devices where iommu is managed by driver

  drivers/gpu/drm/msm/adreno/adreno_device.c | 1 +
  drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c    | 1 +
  drivers/gpu/drm/msm/disp/mdp5/mdp5_kms.c   | 1 +
  drivers/gpu/drm/msm/msm_drv.c              | 1 +
  drivers/iommu/iommu.c                      | 2 +-
  drivers/iommu/of_iommu.c                   | 3 +++
  include/linux/device.h                     | 3 ++-
  7 files changed, 10 insertions(+), 2 deletions(-)
author
#7 lore raw
Quoted 11 lines · Robin Murphy
Quoted 8 lines · Rob Clark

From: Rob Clark <[email protected]>

One of the challenges we have to enable the aarch64 laptops upstream is dealing with the fact that the bootloader enables the display and takes the corresponding SMMU context-bank out of BYPASS. Unfortunately, currently, the IOMMU framework attaches a DMA (or potentially an IDENTITY) domain before the driver is probed and has a chance to intervene and shutdown scanout. Which makes things go horribly wrong.

Quoted 5 lines

Nope, things already went horribly wrong in arm_smmu_device_reset() - sure, sometimes for some configurations it might *seem* like they didn't and that you can fudge the context bank state at arm's length from core code later, but the truth is that impl->cfg_probe is your last chance to guarantee that any necessary SMMU state is preserved.

cfg_probe is where Bjorn's patch is preserving the SMMU state. So I think that should ensure device_reset() preserves the configuration.. or at least if something is missing, that seems fixable.

Quoted 6 lines

The remainder of the problem involves reworking default domain allocation such that we can converge on what iommu_request_dm_for_dev() currently does but without the momentary attachment to a translation domain to cause hiccups. That's starting here:

https://lore.kernel.org/linux-iommu/[email protected]/

I suppose if the stream-match state and bootloader chosen context bank is preserved, then keeping it direct-mapped would avoid things starting to fault before display driver is probed. That plus some solution for GPU default domain would narrow the scope of what I need to just avoiding getting iommu dma_ops installed.

Quoted 19 lines
Quoted 11 lines

But in this case, drm/msm is already directly managing it's IOMMUs directly, the DMA API attached iommu_domain simply gets in the way. This series adds a way that a driver can indicate to drivers/iommu that it does not wish to have an DMA managed iommu_domain attached. This way, drm/msm can shut down scanout cleanly before attaching it's own iommu_domain.

NOTE that to get things working with arm-smmu on the aarch64 laptops, you also need a patchset[1] from Bjorn Andersson to inherit SMMU config at boot, when it is already enabled.

Quoted 7 lines

[1] https://www.spinics.net/lists/arm-kernel/msg732246.html

NOTE that in discussion of previous revisions, RMRR came up. This is not really a replacement for RMRR (nor does RMRR really provide any more information than we already get from EFI GOP, or DT in the simplefb case). I also don't see how RMRR could help w/ SMMU handover of CB/SMR config (Bjorn's patchset[1]) without defining new tables.

Quoted 6 lines

The point of RMRR-like-things is that they identify not just the memory region but also the specific device accessing them, which means the IOMMU driver knows up-front which IDs etc. it must be careful not to disrupt. Obviously for SMMU that *would* be some new table (designed to encompass everything relevant) since literal RMRRs are specifically an Intel VT-d thing.

Perhaps I'm not looking in the right place, but the extent of what I could find about RMRR tables was:

https://github.com/tianocore/edk2/blob/master/MdePkg/Include/IndustryStandard/DmaRemappingReportingTable.h#L122

I couldn't really see how that specifies the device. But entirely possible that I'm not seeing the whole picture.

I am a bit curious about how windows handles this, since they must have the same problem on these laptops.

Quoted 8 lines
Quoted 5 lines

This perhaps doesn't solve the more general case of bootloader enabled display for drivers that actually want to use DMA API managed IOMMU. But it does also happen to avoid a related problem with GPU, caused by the DMA domain claiming the context bank that the GPU firmware expects to use.

Careful bringing that up again, or I really will rework the context bank allocator to avoid this default domain problem entirely... ;)

That doesn't seem like a bad outcome ;-)

BR,
-R

Quoted 18 lines

Robin.

Quoted 16 lines

And it avoids spurious TLB invalidation coming from the unused
DMA domain. So IMHO this is a useful and necessary change.

Rob Clark (2):
iommu: add support for drivers that manage iommu explicitly drm/msm: mark devices where iommu is managed by driver

  drivers/gpu/drm/msm/adreno/adreno_device.c | 1 +
  drivers/gpu/drm/msm/disp/dpu1/dpu_kms.c    | 1 +
  drivers/gpu/drm/msm/disp/mdp5/mdp5_kms.c   | 1 +
  drivers/gpu/drm/msm/msm_drv.c              | 1 +
  drivers/iommu/iommu.c                      | 2 +-
  drivers/iommu/of_iommu.c                   | 3 +++
  include/linux/device.h                     | 3 ++-
  7 files changed, 10 insertions(+), 2 deletions(-)
maintainer
#8 lore raw
Quoted 15 lines · Rob Clark
Quoted 7 lines · Robin Murphy
Quoted 5 lines · Rob Clark

NOTE that in discussion of previous revisions, RMRR came up. This is not really a replacement for RMRR (nor does RMRR really provide any more information than we already get from EFI GOP, or DT in the simplefb case). I also don't see how RMRR could help w/ SMMU handover of CB/SMR config (Bjorn's patchset[1]) without defining new tables.

Quoted 6 lines

The point of RMRR-like-things is that they identify not just the memory region but also the specific device accessing them, which means the IOMMU driver knows up-front which IDs etc. it must be careful not to disrupt. Obviously for SMMU that *would* be some new table (designed to encompass everything relevant) since literal RMRRs are specifically an Intel VT-d thing.

Quoted 7 lines

Perhaps I'm not looking in the right place, but the extent of what I could find about RMRR tables was:

https://github.com/tianocore/edk2/blob/master/MdePkg/Include/IndustryStandard/DmaRemappingReportingTable.h#L122

I couldn't really see how that specifies the device. But entirely possible that I'm not seeing the whole picture.

I don't think anybody was planning to implement RMRR "as-is" for arm64, so you might be better off looking at the proposal from Thierry, although it has some issues that are still to be resolved:

http://lkml.kernel.org/r/[email protected]

Will