Change order of linkage in kernel makefiles for amdkfd

AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory.

This patch changes the order in the drivers makefile, so iommu/ subsystem is linked before gpu/ subsystem. That way, if the gpu and iommu drivers are compiled inside the kernel image (not as modules), the correct order of device loading is still maintained (iommu module is loaded before gpu module).

Signed-off-by Oded Gabbay <[email protected]>

---
  drivers/Makefile | 6 ++++--
  1 file changed, 4 insertions(+), 2 deletions(-)

drivers/Makefile +4 -1

index ebee555..106200f 100644
@@ -50,7 +50,10 @@ obj-$(CONFIG_RESET_CONTROLLER)	+= reset/
  obj-y				+= tty/
  obj-y				+= char/
 
-# gpu/ comes after char for AGP vs DRM startup
+# iommu/ comes before gpu as gpu are using iommu controllers
+obj-$(CONFIG_IOMMU_SUPPORT)	+= iommu/
+
+# gpu/ comes after char for AGP vs DRM startup and after iommu
  obj-y				+= gpu/
 
  obj-$(CONFIG_CONNECTOR)		+= connector/

Hello Linus,
Dave Airlie asked me to send this patch to you for review.

This small patch-set, was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when trying use amdkfd driver on Kaveri). It replaces the previous patch-set called [PATCH 0/3] Use workqueue for device init in amdkfd
(http://lists.freedesktop.org/archives/dri-devel/2014-December/074401.html)

That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled inside the kernel (not as modules). In that case, the correct loading order, as determined by the exported symbol used by each driver, is not enforced anymore and the kernel loads them based on who was linked first. That makes radeon load first, amdkfd second and amd_iommu_v2 third.

Because the initialization of a device in amdkfd is initiated by radeon, and can only be completed if amdkfd and amd_iommu_v2 were loaded and initialized, then in the case mentioned above, this initalization fails and there is a kernel panic as some pointers are not initialized but used nontheless.

To solve this bug, this patch-set moves iommu/ before gpu/ in drivers/Makefile and also moves amdkfd/ before radeon/ in drivers/gpu/drm/Makefile.

The rationale is that in general, AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory. That's why it makes sense to initialize the iommu/ subsystem ahead of the gpu/ subsystem.

Oded

Oded Gabbay (2):
drivers: Move iommu/ before gpu/ in Makefile
drm: Put amdkfd before radeon in drm Makefile

  drivers/Makefile         | 6 ++++--
  drivers/gpu/drm/Makefile | 2 +-
  2 files changed, 5 insertions(+), 3 deletions(-)

This small patch-set, was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when trying use amdkfd driver on Kaveri). It replaces the previous patch-set called [PATCH 0/3] Use workqueue for device init in amdkfd
(http://lists.freedesktop.org/archives/dri-devel/2014-December/074401.html)

That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled inside the kernel (not as modules). In that case, the correct loading order, as determined by the exported symbol used by each driver, is not enforced anymore and the kernel loads them based on who was linked first. That makes radeon load first, amdkfd second and amd_iommu_v2 third.

Because the initialization of a device in amdkfd is initiated by radeon, and can only be completed if amdkfd and amd_iommu_v2 were loaded and initialized, then in the case mentioned above, this initalization fails and there is a kernel panic as some pointers are not initialized but used nontheless.

To solve this bug, this patch-set moves iommu/ before gpu/ in drivers/Makefile and also moves amdkfd/ before radeon/ in drivers/gpu/drm/Makefile.

The rationale is that in general, AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory. That's why it makes sense to initialize the iommu/ subsystem ahead of the gpu/ subsystem.

Quoted 24 lines · Oded Gabbay

This small patch-set, was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when trying use amdkfd driver on Kaveri). It replaces the previous patch-set called [PATCH 0/3] Use workqueue for device init in amdkfd
(http://lists.freedesktop.org/archives/dri-devel/2014-December/074401.html
)

That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled inside the kernel (not as modules). In that case, the correct loading order, as determined by the exported symbol used by each driver, is not enforced anymore and the kernel loads them based on who was linked first. That makes radeon load first, amdkfd second and amd_iommu_v2 third.

Because the initialization of a device in amdkfd is initiated by radeon, and can only be completed if amdkfd and amd_iommu_v2 were loaded and initialized, then in the case mentioned above, this initalization fails and there is a kernel panic as some pointers are not initialized but used nontheless.

To solve this bug, this patch-set moves iommu/ before gpu/ in drivers/Makefile and also moves amdkfd/ before radeon/ in drivers/gpu/drm/Makefile.

Quoted 5 lines

The rationale is that in general, AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory. That's why it makes sense to initialize the iommu/ subsystem ahead of the gpu/ subsystem.

I strongly object to this patch set. This makes assumptions about how the build system influences probe order. That's bad because seemingly unrelated changes could easily break this in the future.

We already have ways to solve this kind of dependency (driver probe deferral), and I think you should be using it to solve this particular problem rather than some linking order hack.

Coincidentally there's a separate thread currently going on that deals with IOMMUs and probe order. The solution being worked on is currently somewhat ARM-specific, so adding a couple of folks for visibility. It looks like we're going to need something more generic since this is a problem that even the "big" architectures need to solve.

Hi Thierry,

Quoted 31 lines · On Thursday 25 December 2014 14:20:59 Thierry Reding wrote:

Quoted 24 lines · Oded Gabbay

This small patch-set, was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when trying use amdkfd driver on Kaveri). It replaces the previous patch-set called [PATCH 0/3] Use workqueue for device init in amdkfd
(http://lists.freedesktop.org/archives/dri-devel/2014-December/074401.html
)

That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled inside the kernel (not as modules). In that case, the correct loading order, as determined by the exported symbol used by each driver, is not enforced anymore and the kernel loads them based on who was linked first. That makes radeon load first, amdkfd second and amd_iommu_v2 third.

Because the initialization of a device in amdkfd is initiated by radeon, and can only be completed if amdkfd and amd_iommu_v2 were loaded and initialized, then in the case mentioned above, this initalization fails and there is a kernel panic as some pointers are not initialized but used nontheless.

To solve this bug, this patch-set moves iommu/ before gpu/ in drivers/Makefile and also moves amdkfd/ before radeon/ in drivers/gpu/drm/Makefile.

Quoted 5 lines

The rationale is that in general, AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory. That's why it makes sense to initialize the iommu/ subsystem ahead of the gpu/ subsystem.

Quoted 7 lines

I strongly object to this patch set. This makes assumptions about how the build system influences probe order. That's bad because seemingly unrelated changes could easily break this in the future.

We already have ways to solve this kind of dependency (driver probe deferral), and I think you should be using it to solve this particular problem rather than some linking order hack.

While I agree with you that probe deferral is the way to go, I believe linkage ordering can still be used as an optimization to avoid deferring probe in the most common cases. I'm thus not opposed to moving iommu/ earlier in link order (provided we can properly test for side effects, as the jump is pretty large), but not as a replacement for probe deferral.

Quoted 6 lines

Coincidentally there's a separate thread currently going on that deals with IOMMUs and probe order. The solution being worked on is currently somewhat ARM-specific, so adding a couple of folks for visibility. It looks like we're going to need something more generic since this is a problem that even the "big" architectures need to solve.

Quoted 40 lines · Laurent Pinchart

Quoted 31 lines · On Thursday 25 December 2014 14:20:59 Thierry Reding wrote:

Quoted 24 lines · Oded Gabbay

This small patch-set, was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when trying use amdkfd driver on Kaveri). It replaces the previous patch-set called [PATCH 0/3] Use workqueue for device init in amdkfd
(http://lists.freedesktop.org/archives/dri-devel/2014-December/074401.ht
ml)

That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled inside the kernel (not as modules). In that case, the correct loading order, as determined by the exported symbol used by each driver, is not enforced anymore and the kernel loads them based on who was linked first. That makes radeon load first, amdkfd second and amd_iommu_v2 third.

Because the initialization of a device in amdkfd is initiated by radeon, and can only be completed if amdkfd and amd_iommu_v2 were loaded and initialized, then in the case mentioned above, this initalization fails and there is a kernel panic as some pointers are not initialized but used nontheless.

To solve this bug, this patch-set moves iommu/ before gpu/ in drivers/Makefile and also moves amdkfd/ before radeon/ in drivers/gpu/drm/Makefile.

Quoted 5 lines

The rationale is that in general, AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory. That's why it makes sense to initialize the iommu/ subsystem ahead of the gpu/ subsystem.

Quoted 7 lines

I strongly object to this patch set. This makes assumptions about how the build system influences probe order. That's bad because seemingly unrelated changes could easily break this in the future.

We already have ways to solve this kind of dependency (driver probe deferral), and I think you should be using it to solve this particular problem rather than some linking order hack.

Quoted 5 lines

While I agree with you that probe deferral is the way to go, I believe linkage ordering can still be used as an optimization to avoid deferring probe in the most common cases. I'm thus not opposed to moving iommu/ earlier in link order (provided we can properly test for side effects, as the jump is pretty large), but not as a replacement for probe deferral.

My thoughts exactly. If this was some extreme use case, than it would be justified to solve it with probe deferral. But I think that for most common cases, GPU are dependent on IOMMU and *not* vice-versa.

BTW, my first try at solving this was to use probe deferral (using workqueue), but the feedback I got from Christian and Dave was that moving iommu/ linkage before gpu/ was a much more simpler solution.

In addition, Linus said he doesn't object to this "band-aid". See:
https://lkml.org/lkml/2014/12/25/152

Oded

Quoted 5 lines

Quoted 5 lines

Coincidentally there's a separate thread currently going on that deals with IOMMUs and probe order. The solution being worked on is currently somewhat ARM-specific, so adding a couple of folks for visibility. It looks like we're going to need something more generic since this is a problem that even the "big" architectures need to solve.

Hi Oded,

Quoted 47 lines · On Sunday 28 December 2014 13:36:50 Oded Gabbay wrote:

Quoted 35 lines · Laurent Pinchart

Quoted 34 lines · On Thursday 25 December 2014 14:20:59 Thierry Reding wrote:

Quoted 24 lines · Oded Gabbay

This small patch-set, was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when trying use amdkfd driver on Kaveri). It replaces the previous patch-set called [PATCH 0/3] Use workqueue for device init in amdkfd
(http://lists.freedesktop.org/archives/dri-devel/2014-December/074401.ht
ml)

That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled inside the kernel (not as modules). In that case, the correct loading order, as determined by the exported symbol used by each driver, is not enforced anymore and the kernel loads them based on who was linked first. That makes radeon load first, amdkfd second and amd_iommu_v2 third.

Because the initialization of a device in amdkfd is initiated by radeon, and can only be completed if amdkfd and amd_iommu_v2 were loaded and initialized, then in the case mentioned above, this initalization fails and there is a kernel panic as some pointers are not initialized but used nontheless.

To solve this bug, this patch-set moves iommu/ before gpu/ in drivers/Makefile and also moves amdkfd/ before radeon/ in drivers/gpu/drm/Makefile.

Quoted 5 lines

The rationale is that in general, AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory. That's why it makes sense to initialize the iommu/ subsystem ahead of the gpu/ subsystem.

I strongly object to this patch set. This makes assumptions about how the build system influences probe order. That's bad because seemingly unrelated changes could easily break this in the future.

Quoted 8 lines

Quoted 3 lines

We already have ways to solve this kind of dependency (driver probe deferral), and I think you should be using it to solve this particular problem rather than some linking order hack.

While I agree with you that probe deferral is the way to go, I believe linkage ordering can still be used as an optimization to avoid deferring probe in the most common cases. I'm thus not opposed to moving iommu/ earlier in link order (provided we can properly test for side effects, as the jump is pretty large), but not as a replacement for probe deferral.

My thoughts exactly. If this was some extreme use case, than it would be justified to solve it with probe deferral. But I think that for most common cases, GPU are dependent on IOMMU and *not* vice-versa.

Quoted 4 lines

BTW, my first try at solving this was to use probe deferral (using workqueue), but the feedback I got from Christian and Dave was that moving iommu/ linkage before gpu/ was a much more simpler solution.

To clarify my position, I believe changing the link order can be a worthwhile optimization, but I'm uncertain about the long term viability of that change as a fix. Probe deferral has been introduced because not all probe ordering issues can be fixed through link ordering, so we should fix the problem properly.

This being said, if modifying the link order can help for now without introducing negative side effects, it would only postpone the real fix, so I'm not opposed to it.

Quoted 11 lines

In addition, Linus said he doesn't object to this "band-aid". See:
https://lkml.org/lkml/2014/12/25/152

Oded

Quoted 5 lines

Quoted 5 lines

Coincidentally there's a separate thread currently going on that deals with IOMMUs and probe order. The solution being worked on is currently somewhat ARM-specific, so adding a couple of folks for visibility. It looks like we're going to need something more generic since this is a problem that even the "big" architectures need to solve.

Am 29.12.2014 um 09:16 schrieb Laurent Pinchart:

Quoted 50 lines

Hi Oded,

Quoted 47 lines · On Sunday 28 December 2014 13:36:50 Oded Gabbay wrote:

Quoted 35 lines · Laurent Pinchart

Quoted 34 lines · On Thursday 25 December 2014 14:20:59 Thierry Reding wrote:

Quoted 24 lines · Oded Gabbay

This small patch-set, was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when trying use amdkfd driver on Kaveri). It replaces the previous patch-set called [PATCH 0/3] Use workqueue for device init in amdkfd
(http://lists.freedesktop.org/archives/dri-devel/2014-December/074401.ht
ml)

That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled inside the kernel (not as modules). In that case, the correct loading order, as determined by the exported symbol used by each driver, is not enforced anymore and the kernel loads them based on who was linked first. That makes radeon load first, amdkfd second and amd_iommu_v2 third.

Because the initialization of a device in amdkfd is initiated by radeon, and can only be completed if amdkfd and amd_iommu_v2 were loaded and initialized, then in the case mentioned above, this initalization fails and there is a kernel panic as some pointers are not initialized but used nontheless.

To solve this bug, this patch-set moves iommu/ before gpu/ in drivers/Makefile and also moves amdkfd/ before radeon/ in drivers/gpu/drm/Makefile.

Quoted 5 lines

The rationale is that in general, AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory. That's why it makes sense to initialize the iommu/ subsystem ahead of the gpu/ subsystem.

I strongly object to this patch set. This makes assumptions about how the build system influences probe order. That's bad because seemingly unrelated changes could easily break this in the future.

Quoted 8 lines

Quoted 3 lines

We already have ways to solve this kind of dependency (driver probe deferral), and I think you should be using it to solve this particular problem rather than some linking order hack.

While I agree with you that probe deferral is the way to go, I believe linkage ordering can still be used as an optimization to avoid deferring probe in the most common cases. I'm thus not opposed to moving iommu/ earlier in link order (provided we can properly test for side effects, as the jump is pretty large), but not as a replacement for probe deferral.

My thoughts exactly. If this was some extreme use case, than it would be justified to solve it with probe deferral. But I think that for most common cases, GPU are dependent on IOMMU and *not* vice-versa.

Fixing this through deferred probing sounds like the correct long term solution to me as well.

But what Thierry is referring to here is probably the approach of returning -EAGAIN from the probe method (at least that was the last status when I looked into this).

The problem with this approach is the interface design between radeon and amdkfd. amdkfd simply doesn't have a probe method which gets called when the hardware is detected and can return -EAGAIN. Instead amdkfd is called by radeon after hardware initialization when it is way to late for such a thing.

Quoted 9 lines

Quoted 3 lines

BTW, my first try at solving this was to use probe deferral (using workqueue), but the feedback I got from Christian and Dave was that moving iommu/ linkage before gpu/ was a much more simpler solution.

To clarify my position, I believe changing the link order can be a worthwhile optimization, but I'm uncertain about the long term viability of that change as a fix. Probe deferral has been introduced because not all probe ordering issues can be fixed through link ordering, so we should fix the problem properly.

Quoted 3 lines

This being said, if modifying the link order can help for now without introducing negative side effects, it would only postpone the real fix, so I'm not opposed to it.

Yeah, that sounds like the right approach to me as well.

In general I would prefer that modules compiled into the kernel load by the order of their symbol dependency just like standalone modules do.

Quoted 52 lines · Christian König

Am 29.12.2014 um 09:16 schrieb Laurent Pinchart:

Quoted 50 lines

Hi Oded,

Quoted 47 lines · On Sunday 28 December 2014 13:36:50 Oded Gabbay wrote:

Quoted 35 lines · Laurent Pinchart

Quoted 34 lines · On Thursday 25 December 2014 14:20:59 Thierry Reding wrote:

Quoted 24 lines · Oded Gabbay

This small patch-set, was created to solve the bug described at
https://bugzilla.kernel.org/show_bug.cgi?id=89661 (Kernel panic when trying use amdkfd driver on Kaveri). It replaces the previous patch-set called [PATCH 0/3] Use workqueue for device init in amdkfd
(http://lists.freedesktop.org/archives/dri-devel/2014-December/074401.ht
ml)

That bug appears only when radeon, amdkfd and amd_iommu_v2 are compiled inside the kernel (not as modules). In that case, the correct loading order, as determined by the exported symbol used by each driver, is not enforced anymore and the kernel loads them based on who was linked first. That makes radeon load first, amdkfd second and amd_iommu_v2 third.

Because the initialization of a device in amdkfd is initiated by radeon, and can only be completed if amdkfd and amd_iommu_v2 were loaded and initialized, then in the case mentioned above, this initalization fails and there is a kernel panic as some pointers are not initialized but used nontheless.

To solve this bug, this patch-set moves iommu/ before gpu/ in drivers/Makefile and also moves amdkfd/ before radeon/ in drivers/gpu/drm/Makefile.

Quoted 5 lines

The rationale is that in general, AMD GPU devices are dependent on AMD IOMMU controller functionality to allow the GPU to access a process's virtual memory address space, without the need for pinning the memory. That's why it makes sense to initialize the iommu/ subsystem ahead of the gpu/ subsystem.

I strongly object to this patch set. This makes assumptions about how the build system influences probe order. That's bad because seemingly unrelated changes could easily break this in the future.

Quoted 8 lines

Quoted 3 lines

We already have ways to solve this kind of dependency (driver probe deferral), and I think you should be using it to solve this particular problem rather than some linking order hack.

While I agree with you that probe deferral is the way to go, I believe linkage ordering can still be used as an optimization to avoid deferring probe in the most common cases. I'm thus not opposed to moving iommu/ earlier in link order (provided we can properly test for side effects, as the jump is pretty large), but not as a replacement for probe deferral.

My thoughts exactly. If this was some extreme use case, than it would be justified to solve it with probe deferral. But I think that for most common cases, GPU are dependent on IOMMU and *not* vice-versa.

Quoted 6 lines

Fixing this through deferred probing sounds like the correct long term solution to me as well.

But what Thierry is referring to here is probably the approach of returning -EAGAIN from the probe method (at least that was the last status when I looked into this).

-EPROBE_DEFER would be the one that I was referring to.

Quoted 5 lines

The problem with this approach is the interface design between radeon and amdkfd. amdkfd simply doesn't have a probe method which gets called when the hardware is detected and can return -EAGAIN. Instead amdkfd is called by radeon after hardware initialization when it is way to late for such a thing.

That sounds like a pretty brittle design. It sounds like nowhere in the code you've encoded this dependency and you rely on symbols only to ensure probe ordering.

Couldn't you simply make radeon check for availability of the IOMMU early and defer probing if it's not there yet? Or if radeon depends on the IOMMU via amdkfd, then perhaps calling into amdkfd to check for the availability of the IOMMU would be a more correct representation of the dependency.

Quoted 9 lines

Quoted 9 lines

Quoted 3 lines

BTW, my first try at solving this was to use probe deferral (using workqueue), but the feedback I got from Christian and Dave was that moving iommu/ linkage before gpu/ was a much more simpler solution.

To clarify my position, I believe changing the link order can be a worthwhile optimization, but I'm uncertain about the long term viability of that change as a fix. Probe deferral has been introduced because not all probe ordering issues can be fixed through link ordering, so we should fix the problem properly.

Quoted 5 lines

Quoted 3 lines

This being said, if modifying the link order can help for now without introducing negative side effects, it would only postpone the real fix, so I'm not opposed to it.

Yeah, that sounds like the right approach to me as well.

I don't think that's a good approach at all. It doesn't encode the dependency anywhere. All you have is some Makefile that lists a bunch of directories. What happens if anybody was to change that order for some other reason? If you're not Cc'ed on the patch and nobody else NAK's it because they are accidentally aware of the dependency that patch is going to break radeon.

Quoted 3 lines

In general I would prefer that modules compiled into the kernel load by the order of their symbol dependency just like standalone modules do.

That I generally agree with. But it can't be a replacement for properly describing the runtime dependencies between modules. There can be other reasons than link order that influence probe ordering. What if the IOMMU driver for some reason gains deferred probe support and therefore doesn't succeed on the first probe? Or only sometimes.

Properly describing the dependency is the only way to prevent any of that.

Thierry

Hi Thierry,
Generally I agree with the issues you describe in the current design. One task in our 2015 workplan is to change the whole method amdkfd is loaded, so it can independently load at any time, regardless of the order of loading between it and radeon and amd_iommu_v2. To reach that goal, I assume we will have to use some form of deferred probing.

However, for the moment, this is the best band-aid I could think of, and choosing between this band-aid or no band-aid at all, I prefer the former any day.