| Message ID | 20240221125324.718192-1-mripard@kernel.org (mailing list archive) |
|---|---|
| State | Accepted |
| Delegated to: | Brendan Higgins |
| Headers | show |
| Series | kunit: Setup DMA masks on the kunit device | expand |
On Wed, Feb 21, 2024 at 01:53:24PM +0100, Maxime Ripard wrote: > Commit d393acce7b3f ("drm/tests: Switch to kunit devices") switched the > DRM device creation helpers from an ad-hoc implementation to the new > kunit device creation helpers introduced in commit d03c720e03bd ("kunit: > Add APIs for managing devices"). > > However, while the DRM helpers were using a platform_device, the kunit > helpers are using a dedicated bus and device type. > > That situation creates small differences in the initialisation, and one > of them is that the kunit devices do not have the DMA masks setup. In > turn, this means that we can't do any kind of DMA buffer allocation > anymore, which creates a regression on some (downstream for now) tests. > This is wrong. The mainline kernel crashes for me without this patch. With v6.8-rc6: [ 6.401720] ok 3 Above the allocation limit [ 6.407679] ------------[ cut here ]------------ [ 6.407927] WARNING: CPU: 1 PID: 1592 at kernel/dma/mapping.c:503 dma_alloc_attrs+0x36/0x60 ... [ 6.413161] # ttm_pool_alloc_basic: ASSERTION FAILED at drivers/gpu/drm/ttm/tests/ttm_pool_test.c:162 [ 6.413161] Expected err == 0, but [ 6.413161] err == -12 (0xfffffffffffffff4) [ 6.416872] not ok 4 One page, with coherent DMA mappings enabled [ 6.420623] ------------[ cut here ]------------ [ 6.420900] list_add corruption. prev->next should be next (ffffffff86d3e930), but was 6b6b6b6b6b6b6b6b. (prev=ffff8f3143621e20). and downhill from there. Granted, this is due to bad cleanup in the ttm unit tests after a failure, but the offending code causing that failure is already in the mainline kernel. > Let's set up a default DMA mask that should work on any platform to fix > it. > > Fixes: d03c720e03bd ("kunit: Add APIs for managing devices") > Signed-off-by: Maxime Ripard <mripard@kernel.org> Tested-by: Guenter Roeck <linux@roeck-us.net> Guenter
On Wed, 21 Feb 2024 at 20:53, Maxime Ripard <mripard@kernel.org> wrote: > > Commit d393acce7b3f ("drm/tests: Switch to kunit devices") switched the > DRM device creation helpers from an ad-hoc implementation to the new > kunit device creation helpers introduced in commit d03c720e03bd ("kunit: > Add APIs for managing devices"). > > However, while the DRM helpers were using a platform_device, the kunit > helpers are using a dedicated bus and device type. > > That situation creates small differences in the initialisation, and one > of them is that the kunit devices do not have the DMA masks setup. In > turn, this means that we can't do any kind of DMA buffer allocation > anymore, which creates a regression on some (downstream for now) tests. > > Let's set up a default DMA mask that should work on any platform to fix > it. > > Fixes: d03c720e03bd ("kunit: Add APIs for managing devices") > Signed-off-by: Maxime Ripard <mripard@kernel.org> > --- Thanks for this. As Guenter notes, this fixes the ttm tests, which are otherwise completely broken: ./tools/testing/kunit/kunit.py run --arch x86_64 --kunitconfig drivers/gpu/drm/ttm/tests/ While I suspect there's probably a discussion to have around what things KUnit devices should set up (and how much flexibility users should have there), I think this is definitely a fix worth having in the meantime. Reviewed-by: David Gow <davidgow@google.com> Cheers, -- David > lib/kunit/device.c | 4 ++++ > 1 file changed, 4 insertions(+) > > diff --git a/lib/kunit/device.c b/lib/kunit/device.c > index 644a38a1f5b1..9ea399049749 100644 > --- a/lib/kunit/device.c > +++ b/lib/kunit/device.c > @@ -10,6 +10,7 @@ > */ > > #include <linux/device.h> > +#include <linux/dma-mapping.h> > > #include <kunit/test.h> > #include <kunit/device.h> > @@ -133,6 +134,9 @@ static struct kunit_device *kunit_device_register_internal(struct kunit *test, > return ERR_PTR(err); > } > > + kunit_dev->dev.dma_mask = &kunit_dev->dev.coherent_dma_mask; > + kunit_dev->dev.coherent_dma_mask = DMA_BIT_MASK(32); > + > kunit_add_action(test, device_unregister_wrapper, &kunit_dev->dev); > > return kunit_dev; > -- > 2.43.2 >
diff --git a/lib/kunit/device.c b/lib/kunit/device.c index 644a38a1f5b1..9ea399049749 100644 --- a/lib/kunit/device.c +++ b/lib/kunit/device.c @@ -10,6 +10,7 @@ */ #include <linux/device.h> +#include <linux/dma-mapping.h> #include <kunit/test.h> #include <kunit/device.h> @@ -133,6 +134,9 @@ static struct kunit_device *kunit_device_register_internal(struct kunit *test, return ERR_PTR(err); } + kunit_dev->dev.dma_mask = &kunit_dev->dev.coherent_dma_mask; + kunit_dev->dev.coherent_dma_mask = DMA_BIT_MASK(32); + kunit_add_action(test, device_unregister_wrapper, &kunit_dev->dev); return kunit_dev;
Commit d393acce7b3f ("drm/tests: Switch to kunit devices") switched the DRM device creation helpers from an ad-hoc implementation to the new kunit device creation helpers introduced in commit d03c720e03bd ("kunit: Add APIs for managing devices"). However, while the DRM helpers were using a platform_device, the kunit helpers are using a dedicated bus and device type. That situation creates small differences in the initialisation, and one of them is that the kunit devices do not have the DMA masks setup. In turn, this means that we can't do any kind of DMA buffer allocation anymore, which creates a regression on some (downstream for now) tests. Let's set up a default DMA mask that should work on any platform to fix it. Fixes: d03c720e03bd ("kunit: Add APIs for managing devices") Signed-off-by: Maxime Ripard <mripard@kernel.org> --- lib/kunit/device.c | 4 ++++ 1 file changed, 4 insertions(+)