drm/tests: shmem: Fix intermittent DMA overflow on ppc64le/s390x
Subsystems (2)
cover drm/tests: shmem: Fix intermittent DMA overflow on ppc64le/s390x
Two drm_gem_shmem KUnit tests, drm_gem_shmem_test_get_pages_sgt() and drm_gem_shmem_test_purge(), intermittently fail on ppc64le and s390x CI with a DMA address overflow warning followed by -EIO:
DMA addr 0x00000001130b0000+65536 overflow (mask ffffffff, bus limit 0). Expected sgt is not error, but is: -5
Both tests call drm_gem_shmem_get_pages_sgt(), which internally pins backing pages and DMA-maps them via dma_map_sgtable(). The DMA mapping path is:
drm_gem_shmem_test_purge()
drm_gem_shmem_get_pages_sgt()
drm_gem_shmem_get_pages_sgt_locked() [drm_gem_shmem_helper.c]
dma_map_sgtable() [mapping.c]
__dma_map_sg_attrs()
dma_direct_map_sg() [direct.c]
dma_direct_map_phys() [kernel/dma/direct.h]
dma_capable() Checks addr against DMA mask
-> FAILS: addr > 0xFFFFFFFFKUnit devices are initialized with a 32-bit DMA mask (DMA_BIT_MASK(32)) in lib/kunit/device.c. On systems where the kernel allocates backing pages at physical addresses above 4GB, dma_capable() returns false because the address exceeds the 32-bit mask. The dma_set_mask() function updates *dev->dma_mask to the given value; setting it to DMA_BIT_MASK(64) allows any physical address to pass the check.
The failure is intermittent because pages may or may not be allocated above 4GB on any given run depend on memory pressure.
A third test in the same suite, drm_gem_shmem_test_obj_create_private, already calls dma_set_mask(drm_dev->dev, DMA_BIT_MASK(64)) before its DMA mapping. This series applies the same fix to the two remaining tests.
José Expósito (2):
drm/tests: shmem: Set DMA mask to 64-bit in
drm_gem_shmem_test_get_pages_sgt
drm/tests: shmem: Set DMA mask to 64-bit in drm_gem_shmem_test_purge
drivers/gpu/drm/tests/drm_gem_shmem_test.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
Signature · 1 line
2.54.0
PATCH 1/2 drm/tests: shmem: Set DMA mask to 64-bit in drm_gem_shmem_test_get_pages_sgt
From: José Expósito <[email protected]>
drm_gem_shmem_test_get_pages_sgt intermittently fails on ppc64le and s390x CI systems with a DMA address overflow: [1]
DMA addr 0x00000001b6110000+65536 overflow (mask ffffffff, bus limit 0)
WARNING: kernel/dma/direct.h:114 dma_direct_map_phys+0x298/0x300
drm_gem_shmem_test_get_pages_sgt: ASSERTION FAILED at
drivers/gpu/drm/tests/drm_gem_shmem_test.c:255 Expected sgt is not error, but is: -5
The call chain leading to the failure is:
drm_gem_shmem_test_get_pages_sgt()
drm_gem_shmem_get_pages_sgt()
drm_gem_shmem_get_pages_sgt_locked() [drm_gem_shmem_helper.c]
dma_map_sgtable() [mapping.c]
__dma_map_sg_attrs()
dma_direct_map_sg() [direct.c]
dma_direct_map_phys() [kernel/dma/direct.h]
dma_capable() Checks addr against DMA mask
-> FAILS: addr > 0xFFFFFFFFThe root cause is that KUnit devices are initialized with a 32-bit DMA mask (DMA_BIT_MASK(32)) in lib/kunit/device.c. On ppc64le and s390x systems with physical memory above 4GB, page allocations can land at addresses that exceed this mask. When drm_gem_shmem_get_pages_sgt() attempts to DMA-map these pages via dma_map_sgtable(), the DMA layer rejects the mapping because the physical address overflows the 32-bit mask.
The failure is intermittent because pages may or may not be allocated above 4GB on any given run depend on memory pressure.
Fix by setting a 64-bit DMA mask on the device before calling drm_gem_shmem_get_pages_sgt(), following the same pattern already used in drm_gem_shmem_test_obj_create_private().
--- drivers/gpu/drm/tests/drm_gem_shmem_test.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
drivers/gpu/drm/tests/drm_gem_shmem_test.c +5 -1
@@ -250,7 +250,11 @@ static void drm_gem_shmem_test_get_pages_sgt(struct kunit *test)struct drm_gem_shmem_object *shmem;struct sg_table *sgt;struct scatterlist *sg; -unsigned int si, ret, len =0 ; +unsigned int si, len =0 ; +int ret; + + ret =dma_set_mask (drm_dev->dev,DMA_BIT_MASK (64 )); +KUNIT_ASSERT_EQ (test, ret,0 ); shmem =drm_gem_shmem_create (drm_dev,TEST_SIZE );KUNIT_ASSERT_NOT_ERR_OR_NULL (test, shmem);
Signature · 1 line
2.54.0
PATCH 2/2 drm/tests: shmem: Set DMA mask to 64-bit in drm_gem_shmem_test_purge
In reply to José ExpósitoFrom: José Expósito <[email protected]>
drm_gem_shmem_test_purge intermittently fails on ppc64le and s390x CI systems with a DMA address overflow: [1]
DMA addr 0x0000000100307000+4096 overflow (mask ffffffff, bus limit 0)
WARNING: kernel/dma/direct.h:114 dma_direct_map_sg+0x778/0x920
drm_gem_shmem_test_purge: ASSERTION FAILED at
drivers/gpu/drm/tests/drm_gem_shmem_test.c:330 Expected sgt is not error, but is: -5
The call chain leading to the failure is:
drm_gem_shmem_test_purge()
drm_gem_shmem_get_pages_sgt()
drm_gem_shmem_get_pages_sgt_locked() [drm_gem_shmem_helper.c]
dma_map_sgtable() [mapping.c]
__dma_map_sg_attrs()
dma_direct_map_sg() [direct.c]
dma_direct_map_phys() [kernel/dma/direct.h]
dma_capable() Checks addr against DMA mask
-> FAILS: addr > 0xFFFFFFFFThe root cause is that KUnit devices are initialized with a 32-bit DMA mask (DMA_BIT_MASK(32)) in lib/kunit/device.c. On ppc64le and s390x systems with physical memory above 4GB, page allocations can land at addresses that exceed this mask. When drm_gem_shmem_get_pages_sgt() attempts to DMA-map these pages via dma_map_sgtable(), the DMA layer rejects the mapping because the physical address overflows the 32-bit mask.
The failure is intermittent because pages may or may not be allocated above 4GB on any given run depend on memory pressure.
Fix by setting a 64-bit DMA mask on the device before calling drm_gem_shmem_get_pages_sgt(), following the same pattern already used in drm_gem_shmem_test_obj_create_private().
--- drivers/gpu/drm/tests/drm_gem_shmem_test.c | 3 +++ 1 file changed, 3 insertions(+)
drivers/gpu/drm/tests/drm_gem_shmem_test.c +3 -0
@@ -325,6 +325,9 @@ static void drm_gem_shmem_test_purge(struct kunit *test)struct sg_table *sgt;int ret; + ret =dma_set_mask (drm_dev->dev,DMA_BIT_MASK (64 )); +KUNIT_ASSERT_EQ (test, ret,0 ); + shmem =drm_gem_shmem_create (drm_dev,TEST_SIZE );KUNIT_ASSERT_NOT_ERR_OR_NULL (test, shmem);
Signature · 1 line
2.54.0
Thomas Zimmermann
maintainerHi
Am 03.07.26 um 08:34 schrieb José Expósito:
Quoted 28 lines
Two
drm_gem_shmemKUnit tests,drm_gem_shmem_test_get_pages_sgt()anddrm_gem_shmem_test_purge(), intermittently fail on ppc64le and s390x CI with a DMA address overflow warning followed by-EIO:DMA addr 0x00000001130b0000+65536 overflow (mask ffffffff, bus limit 0). Expected sgt is not error, but is: -5
Both tests call
drm_gem_shmem_get_pages_sgt(), which internally pins backing pages and DMA-maps them viadma_map_sgtable(). The DMA mapping path is:drm_gem_shmem_test_purge() drm_gem_shmem_get_pages_sgt() drm_gem_shmem_get_pages_sgt_locked() [drm_gem_shmem_helper.c] dma_map_sgtable() [mapping.c] __dma_map_sg_attrs() dma_direct_map_sg() [direct.c] dma_direct_map_phys() [kernel/dma/direct.h] dma_capable() Checks addr against DMA mask -> FAILS: addr > 0xFFFFFFFFKUnit devices are initialized with a 32-bit DMA mask (
DMA_BIT_MASK(32)) inlib/kunit/device.c. On systems where the kernel allocates backing pages at physical addresses above 4GB,dma_capable()returns false because the address exceeds the 32-bit mask. Thedma_set_mask()function updates*dev->dma_maskto the given value; setting it toDMA_BIT_MASK(64)allows any physical address to pass the check.
Quoted 6 lines
The failure is intermittent because pages may or may not be allocated above 4GB on any given run depend on memory pressure.
A third test in the same suite,
drm_gem_shmem_test_obj_create_private, already callsdma_set_mask(drm_dev->dev, DMA_BIT_MASK(64))before its DMA mapping. This series applies the same fix to the two remaining tests.
Instead of doing whack-a-mole, is it possible to move the existing fix from drm_gem_shmem_test_obj_create_private() to _test_init at [1] ? The DMA mask would then be the same on all tests. [1] https://elixir.bootlin.com/linux/v7.1.2/source/drivers/gpu/drm/tests/drm_gem_shmem_test.c#L359 Best regards Thomas
Quoted 9 lines
José Expósito (2):
drm/tests: shmem: Set DMA mask to 64-bit in
drm_gem_shmem_test_get_pages_sgt
drm/tests: shmem: Set DMA mask to 64-bit indrm_gem_shmem_test_purgedrivers/gpu/drm/tests/drm_gem_shmem_test.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)
Signature · 6 lines
-- Thomas Zimmermann Graphics Driver Developer SUSE Software Solutions Germany GmbH Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)
Hi Thomas,
Thanks for your review.
Quoted 34 lines · Thomas Zimmermann
Hi
Am 03.07.26 um 08:34 schrieb José Expósito:
Quoted 31 lines
Two
drm_gem_shmemKUnit tests,drm_gem_shmem_test_get_pages_sgt()anddrm_gem_shmem_test_purge(), intermittently fail on ppc64le and s390x CI with a DMA address overflow warning followed by-EIO:DMA addr 0x00000001130b0000+65536 overflow (mask ffffffff, bus limit 0). Expected sgt is not error, but is: -5
Both tests call
drm_gem_shmem_get_pages_sgt(), which internally pins backing pages and DMA-maps them viadma_map_sgtable(). The DMA mapping path is:drm_gem_shmem_test_purge() drm_gem_shmem_get_pages_sgt() drm_gem_shmem_get_pages_sgt_locked() [drm_gem_shmem_helper.c] dma_map_sgtable() [mapping.c] __dma_map_sg_attrs() dma_direct_map_sg() [direct.c] dma_direct_map_phys() [kernel/dma/direct.h] dma_capable() Checks addr against DMA mask -> FAILS: addr > 0xFFFFFFFFKUnit devices are initialized with a 32-bit DMA mask (
DMA_BIT_MASK(32)) inlib/kunit/device.c. On systems where the kernel allocates backing pages at physical addresses above 4GB,dma_capable()returns false because the address exceeds the 32-bit mask. Thedma_set_mask()function updates*dev->dma_maskto the given value; setting it toDMA_BIT_MASK(64)allows any physical address to pass the check.The failure is intermittent because pages may or may not be allocated above 4GB on any given run depend on memory pressure.
Quoted 7 lines
Quoted 3 lines
A third test in the same suite,
drm_gem_shmem_test_obj_create_private, already callsdma_set_mask(drm_dev->dev, DMA_BIT_MASK(64))before its DMA mapping. This series applies the same fix to the two remaining tests.Instead of doing whack-a-mole, is it possible to move the existing fix from
drm_gem_shmem_test_obj_create_private()to_test_initat [1] ? The DMA mask
would then be the same on all tests. [1] https://elixir.bootlin.com/linux/v7.1.2/source/drivers/gpu/drm/tests/drm_gem_shmem_test.c#L359
Yes, I think it is possible. However, this is an intermittent failure that I couldn't reproduce locally and I tried to reduce as much as possible the risk of regression limitting the change to the 2 affected tests.
I'm afraid I don't have enough knoledge about GEM's internals to know if moving the fix to the init function could introduce unwanted side-effects in the other tests. If you think it'd be safe, I'll trust you and move it.
Jose
PS - Next week I'll be away from keyboard, so any change would need to wait a bit
Quoted 20 lines
Best regards Thomas
Quoted 9 lines
José Expósito (2):
drm/tests: shmem: Set DMA mask to 64-bit in
drm_gem_shmem_test_get_pages_sgt
drm/tests: shmem: Set DMA mask to 64-bit indrm_gem_shmem_test_purgedrivers/gpu/drm/tests/drm_gem_shmem_test.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)Signature · 6 lines
-- Thomas Zimmermann Graphics Driver Developer SUSE Software Solutions Germany GmbH Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)
Thomas Zimmermann
maintainerHi José
Am 03.07.26 um 11:00 schrieb José Expósito:
Quoted 45 lines
Hi Thomas,
Thanks for your review.
Quoted 34 lines · Thomas Zimmermann
Hi
Am 03.07.26 um 08:34 schrieb José Expósito:
Quoted 31 lines
Two
drm_gem_shmemKUnit tests,drm_gem_shmem_test_get_pages_sgt()anddrm_gem_shmem_test_purge(), intermittently fail on ppc64le and s390x CI with a DMA address overflow warning followed by-EIO:DMA addr 0x00000001130b0000+65536 overflow (mask ffffffff, bus limit 0). Expected sgt is not error, but is: -5
Both tests call
drm_gem_shmem_get_pages_sgt(), which internally pins backing pages and DMA-maps them viadma_map_sgtable(). The DMA mapping path is:drm_gem_shmem_test_purge() drm_gem_shmem_get_pages_sgt() drm_gem_shmem_get_pages_sgt_locked() [drm_gem_shmem_helper.c] dma_map_sgtable() [mapping.c] __dma_map_sg_attrs() dma_direct_map_sg() [direct.c] dma_direct_map_phys() [kernel/dma/direct.h] dma_capable() Checks addr against DMA mask -> FAILS: addr > 0xFFFFFFFFKUnit devices are initialized with a 32-bit DMA mask (
DMA_BIT_MASK(32)) inlib/kunit/device.c. On systems where the kernel allocates backing pages at physical addresses above 4GB,dma_capable()returns false because the address exceeds the 32-bit mask. Thedma_set_mask()function updates*dev->dma_maskto the given value; setting it toDMA_BIT_MASK(64)allows any physical address to pass the check.The failure is intermittent because pages may or may not be allocated above 4GB on any given run depend on memory pressure.
Quoted 6 lines
Quoted 3 lines
A third test in the same suite,
drm_gem_shmem_test_obj_create_private, already callsdma_set_mask(drm_dev->dev, DMA_BIT_MASK(64))before its DMA mapping. This series applies the same fix to the two remaining tests.Instead of doing whack-a-mole, is it possible to move the existing fix from
drm_gem_shmem_test_obj_create_private()to_test_initat [1] ? The DMA mask
would then be the same on all tests. [1] https://elixir.bootlin.com/linux/v7.1.2/source/drivers/gpu/drm/tests/drm_gem_shmem_test.c#L359
Quoted 7 lines
Yes, I think it is possible. However, this is an intermittent failure that I couldn't reproduce locally and I tried to reduce as much as possible the risk of regression limitting the change to the 2 affected tests.
I'm afraid I don't have enough knoledge about GEM's internals to know if moving the fix to the init function could introduce unwanted side-effects in the other tests. If you think it'd be safe, I'll trust you and move it.
Please do. I don't see why it should not work, and it's an easy revert it doesn't.
Best regards
Thomas
Quoted 23 lines
Jose
PS - Next week I'll be away from keyboard, so any change would need to wait a bit
Quoted 18 lines
Best regards Thomas
Quoted 8 lines
José Expósito (2):
drm/tests: shmem: Set DMA mask to 64-bit in
drm_gem_shmem_test_get_pages_sgt
drm/tests: shmem: Set DMA mask to 64-bit indrm_gem_shmem_test_purgedrivers/gpu/drm/tests/drm_gem_shmem_test.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)Signature · 6 lines
-- Thomas Zimmermann Graphics Driver Developer SUSE Software Solutions Germany GmbH Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)
Signature · 6 lines
-- Thomas Zimmermann Graphics Driver Developer SUSE Software Solutions Germany GmbH Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)
Quoted 51 lines · Thomas Zimmermann
Hi José
Am 03.07.26 um 11:00 schrieb José Expósito:Quoted 49 lines
Hi Thomas,
Thanks for your review.
Quoted 34 lines · Thomas Zimmermann
Hi
Am 03.07.26 um 08:34 schrieb José Expósito:
Quoted 31 lines
Two
drm_gem_shmemKUnit tests,drm_gem_shmem_test_get_pages_sgt()anddrm_gem_shmem_test_purge(), intermittently fail on ppc64le and s390x CI with a DMA address overflow warning followed by-EIO:DMA addr 0x00000001130b0000+65536 overflow (mask ffffffff, bus limit 0). Expected sgt is not error, but is: -5
Both tests call
drm_gem_shmem_get_pages_sgt(), which internally pins backing pages and DMA-maps them viadma_map_sgtable(). The DMA mapping path is:drm_gem_shmem_test_purge() drm_gem_shmem_get_pages_sgt() drm_gem_shmem_get_pages_sgt_locked() [drm_gem_shmem_helper.c] dma_map_sgtable() [mapping.c] __dma_map_sg_attrs() dma_direct_map_sg() [direct.c] dma_direct_map_phys() [kernel/dma/direct.h] dma_capable() Checks addr against DMA mask -> FAILS: addr > 0xFFFFFFFFKUnit devices are initialized with a 32-bit DMA mask (
DMA_BIT_MASK(32)) inlib/kunit/device.c. On systems where the kernel allocates backing pages at physical addresses above 4GB,dma_capable()returns false because the address exceeds the 32-bit mask. Thedma_set_mask()function updates*dev->dma_maskto the given value; setting it toDMA_BIT_MASK(64)allows any physical address to pass the check.The failure is intermittent because pages may or may not be allocated above 4GB on any given run depend on memory pressure.
Quoted 6 lines
Quoted 3 lines
A third test in the same suite,
drm_gem_shmem_test_obj_create_private, already callsdma_set_mask(drm_dev->dev, DMA_BIT_MASK(64))before its DMA mapping. This series applies the same fix to the two remaining tests.Instead of doing whack-a-mole, is it possible to move the existing fix from
drm_gem_shmem_test_obj_create_private()to_test_initat [1] ? The DMA mask
would then be the same on all tests. [1] https://elixir.bootlin.com/linux/v7.1.2/source/drivers/gpu/drm/tests/drm_gem_shmem_test.c#L359Yes, I think it is possible. However, this is an intermittent failure that I couldn't reproduce locally and I tried to reduce as much as possible the risk of regression limitting the change to the 2 affected tests.
Quoted 6 lines
Quoted 3 lines
I'm afraid I don't have enough knoledge about GEM's internals to know if moving the fix to the init function could introduce unwanted side-effects in the other tests. If you think it'd be safe, I'll trust you and move it.
Please do. I don't see why it should not work, and it's an easy revert it doesn't.
I just sent v2 with the suggested change. I run the test in several architectures and I didn't run into any issue:
https://lore.kernel.org/dri-devel/[email protected]/
Thanks,
Jose
Quoted 37 lines
Best regards
ThomasQuoted 23 lines
Jose
PS - Next week I'll be away from keyboard, so any change would need to wait a bit
Quoted 18 lines
Best regards Thomas
Quoted 8 lines
José Expósito (2):
drm/tests: shmem: Set DMA mask to 64-bit in
drm_gem_shmem_test_get_pages_sgt
drm/tests: shmem: Set DMA mask to 64-bit indrm_gem_shmem_test_purgedrivers/gpu/drm/tests/drm_gem_shmem_test.c | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-)Signature · 6 lines
-- Thomas Zimmermann Graphics Driver Developer SUSE Software Solutions Germany GmbH Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)Signature · 6 lines
-- Thomas Zimmermann Graphics Driver Developer SUSE Software Solutions Germany GmbH Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)