Am 10.11.21 um 05:31 schrieb xinhui pan:
BO might sit in a wrong lru list as there is a small period of memory
moving and lru list updating.
Lets skip eviction if we hit such mismatch.
Suggested-by: Christian König
Signed-off-by: xinhui pan
Reviewed-by: Christian König for the series.
Allow other drivers to initialize relative to current active
profile and react to platform profile changes.
Drivers wishing to utilize this should register for notification
at module load and unregister when unloading.
Notifications will come in the from a notifier call.
Signed-off-by: Mario
Various drivers provide platform profile support to let users set a hint
in their GUI whether they want to run in a high performance, low battery
life or balanced configuration.
Drivers that provide this typically work with the firmware on their system
to configure hardware. In the case of
Many OEM platform provide a platform profile knob that can be used to make
firmware tunings to the system to allow operating in a higher or lower
performance mode trading off power consumption.
Software like power-profiles-daemon to expose this knob to the UI.
As we know the user's intent to go
Reviewed-by: Alex Deucher
On Wed, Nov 10, 2021 at 12:16 AM Guchun Chen wrote:
>
> Fixes: 5b30f206dbd1("drm/amdgpu/amdgpu_vcn: convert to IP version checking")
> Signed-off-by: Flora Cui
> Signed-off-by: Guchun Chen
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c | 1 +
>
Fixes: 5b30f206dbd1("drm/amdgpu/amdgpu_vcn: convert to IP version checking")
Signed-off-by: Flora Cui
Signed-off-by: Guchun Chen
---
drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c | 1 +
drivers/gpu/drm/amd/amdgpu/amdgpu_vcn.c | 1 +
drivers/gpu/drm/amd/amdgpu/nv.c | 1 +
3
On 2021-11-09 6:04 p.m., Philip Yang wrote:
Make sure mm does not remove when prange deferred work insert mmu range
notifier, to avoid WARNING:
WARNING: CPU: 6 PID: 1787 at mm/mmu_notifier.c:932
__mmu_interval_notifier_insert+0xdd/0xf0
Workqueue: events svm_range_deferred_list_work [amdgpu]
After we move BO to a new memory region, we should put it to
the new memory manager's lru list regardless we unlock the resv or not.
Cc: sta...@vger.kernel.org
Reviewed-by: Christian König
Signed-off-by: xinhui pan
---
drivers/gpu/drm/ttm/ttm_bo.c | 2 ++
1 file changed, 2 insertions(+)
diff
BO might sit in a wrong lru list as there is a small period of memory
moving and lru list updating.
Lets skip eviction if we hit such mismatch.
Suggested-by: Christian König
Signed-off-by: xinhui pan
---
drivers/gpu/drm/ttm/ttm_bo.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
On 2021-11-09 6:04 p.m., Philip Yang wrote:
restore pages work can not find kfd process or mm struct if process is
destroyed before drain retry fault work schedule to run, this is not
failure, return 0 to avoid dump GPU vm fault kernel log.
I wonder if this could also be solved by draining page
On 2021-11-09 6:04 p.m., Philip Yang wrote:
Before restore pages takes mmap read or write lock, vma maybe removed.
Check if vma exists before creating unregistered range or verifying
range access permission, and return 0 if vma is removed to avoid restore
pages return failure to report GPU vm
On 2021-11-09 6:04 p.m., Philip Yang wrote:
If unmapping partial range, the parent prange list op is update
notifier, child range list op is unmap range, need check child range to
set drain retry fault flag.
Signed-off-by: Philip Yang
I think this could be simplified by simply setting
On 2021-11-09 2:12 p.m., Ramesh Errabolu wrote:
MMIO/DOORBELL BOs encode control data and should be pinned in GTT
domain before enabling PCIe connected peer devices in accessing it
Signed-off-by: Ramesh Errabolu
Reviewed-by: Felix Kuehling
---
On Tue, Nov 9, 2021 at 9:14 PM Guchun Chen wrote:
>
> Fixes: 67a765c6352d("drm/amdgpu: clean up set IP function")
>
> Signed-off-by: Guchun Chen
Reviewed-by: Alex Deucher
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff
Disable HDP register remapping on SRIOV and set rmmio_remap.reg_offset
to the fixed address of the VF register for hdp_v*_flush_hdp.
Signed-off-by: Felix Kuehling
---
drivers/gpu/drm/amd/amdgpu/nbio_v2_3.c | 4
drivers/gpu/drm/amd/amdgpu/nbio_v6_1.c | 4
Fixes: 67a765c6352d("drm/amdgpu: clean up set IP function")
Signed-off-by: Guchun Chen
---
drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_discovery.c
IH ring1 is used to process GPU retry fault, overflow is enabled to
drain retry fault before unmapping the range, wptr may pass rptr,
amdgpu_ih_process should check rptr equals to the latest wptr to exit,
otherwise it will continue to recover outdatad retry fault after drain
retry fault is done,
restore pages work can not find kfd process or mm struct if process is
destroyed before drain retry fault work schedule to run, this is not
failure, return 0 to avoid dump GPU vm fault kernel log.
Signed-off-by: Philip Yang
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 ++--
1 file changed, 2
Make sure mm does not remove when prange deferred work insert mmu range
notifier, to avoid WARNING:
WARNING: CPU: 6 PID: 1787 at mm/mmu_notifier.c:932
__mmu_interval_notifier_insert+0xdd/0xf0
Workqueue: events svm_range_deferred_list_work [amdgpu]
RIP:
Before restore pages takes mmap read or write lock, vma maybe removed.
Check if vma exists before creating unregistered range or verifying
range access permission, and return 0 if vma is removed to avoid restore
pages return failure to report GPU vm fault to application.
Signed-off-by: Philip
If unmapping partial range, the parent prange list op is update
notifier, child range list op is unmap range, need check child range to
set drain retry fault flag.
Signed-off-by: Philip Yang
---
drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 13 -
1 file changed, 12 insertions(+), 1
Switch to IP version checking instead of asic_type on various KFD
version checks.
Signed-off-by: Graham Sider
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +-
drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 2 +-
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 27 ++-
Converts KFD switch statements to use IP version checking instead
of asic_type.
Signed-off-by: Graham Sider
---
drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 124 +-
drivers/gpu/drm/amd/amdkfd/kfd_device.c | 8 +-
.../drm/amd/amdkfd/kfd_device_queue_manager.c | 33
Defined as GC HWIP >= IP_VERSION(9, 0, 1).
Also defines KFD_GC_VERSION to return GC HWIP version.
Signed-off-by: Graham Sider
---
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 2 +-
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 4 ++--
drivers/gpu/drm/amd/amdkfd/kfd_priv.h
[AMD Official Use Only]
Based on my experiments I am able conclude that I can avoid validating the BO
prior to pinning it. I don't have the code history that led me to validating
the BO in the first place. In any case I posted an updated patch to the
DRM-NEXT branch in addition to a standalone
MMIO/DOORBELL BOs encode control data and should be pinned in GTT
domain before enabling PCIe connected peer devices in accessing it
Signed-off-by: Ramesh Errabolu
---
.../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 70 +++
1 file changed, 70 insertions(+)
diff --git
On Tue, Nov 9, 2021 at 1:07 AM Daniel Vetter wrote:
>
> On Mon, Nov 08, 2021 at 03:39:17PM -0800, Rob Clark wrote:
> > I stumbled across this thread when I ran into the same issue, while
> > working out how to move drm/msm to use scheduler's retire +
> > timeout/recovery (and get rid of our own
In general the correct idea, but the wrong place to check that.
Calling amdgpu_ttm_bo_eviction_valuable() is only optional, but that
check must be mandatory for correct operation.
This needs to be inside ttm_bo_evict_swapout_allowable().
Christian.
Am 09.11.21 um 14:41 schrieb Pan, Xinhui:
On 04/11/21 12:46 am, Matthew Auld wrote:
> On 25/10/2021 14:00, Arunpravin wrote:
>> On contiguous allocation, we round up the size
>> to the *next* power of 2, implement a function
>> to free the unused pages after the newly allocate block.
>>
>> Signed-off-by: Arunpravin
>
> Ideally this
On 2021-11-09 00:14, Shirish S wrote:
> update developers with next level of info about unsupported
> display configuration query that led to atomic check failure.
>
> Signed-off-by: Shirish S
Reviewed-by: Harry Wentland
Harry
> ---
> .../gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 69
[AMD Official Use Only]
yes, a double check is needed.
how about change below.
As long as we detect such mismatch, it indicates another eviction is on going.
return false here is reasonable.
@@ -1335,6 +1336,8 @@ static bool amdgpu_ttm_bo_eviction_valuable(struct
ttm_buffer_object *bo,
Exactly that's the reason why we should have the double check in TTM
I've mentioned in the other mail.
Christian.
Am 09.11.21 um 14:16 schrieb Pan, Xinhui:
[AMD Official Use Only]
Actually this patch does not totally fix the mismatch of lru list with mem_type as
mem_type is changed in
Yeah, but that should never happen in the first place.
Even when the BO is on the wrong LRU TTM should check that beforehand.
In other words when we pick a BO from the LRU we should still double
check bo->resource->mem_type to make sure it is what we are searching for.
Christian.
Am
[AMD Official Use Only]
Actually this patch does not totally fix the mismatch of lru list with mem_type
as mem_type is changed in ->move() and lru list is changed after that.
During this small period, another eviction could still happed and evict this
mismatched BO from sMam(say, its lru list
[AMD Official Use Only]
Yes, a stable tag is needed. vulkan guys say 5.14 hit this issue too.
I think that amdgpu_bo_move() does support copy from sysMem to sysMem correctly.
maybe something below is needed.
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ttm.c
Mhm, I'm not sure what the rational behind that is.
Not moving the BO would make things less efficient, but should never
cause a crash.
Maybe we should add a CC: stable tag and push it to -fixes instead?
Christian.
Am 09.11.21 um 13:28 schrieb Pan, Xinhui:
[AMD Official Use Only]
I hit
[AMD Official Use Only]
I hit vulkan cts test hang with navi23.
dmesg says gmc page fault with address 0x0, 0x1000, 0x2000
And some debug log also says amdgu copy one BO from system Domain to system
Domain which is really weird.
发件人: Koenig,
Am 09.11.21 um 12:19 schrieb xinhui pan:
After we move BO to a new memory region, we should put it to
the new memory manager's lru list regardless we unlock the resv or not.
Signed-off-by: xinhui pan
Interesting find, did you trigger that somehow or did you just stumbled
over it by reading
After we move BO to a new memory region, we should put it to
the new memory manager's lru list regardless we unlock the resv or not.
Signed-off-by: xinhui pan
---
drivers/gpu/drm/ttm/ttm_bo.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c
On Mon, Nov 08, 2021 at 03:39:17PM -0800, Rob Clark wrote:
> I stumbled across this thread when I ran into the same issue, while
> working out how to move drm/msm to use scheduler's retire +
> timeout/recovery (and get rid of our own mirror list of in-flight
> jobs). We already have hw error
[AMD Official Use Only]
> -Original Message-
> From: Lazar, Lijo
> Sent: Tuesday, November 9, 2021 3:29 PM
> To: Koenig, Christian ; Borislav Petkov
> ; Paul Menzel ; Liu, Leo
>
> Cc: Deucher, Alexander ; Quan, Evan
> ; amd-gfx@lists.freedesktop.org
> Subject: Re: [PATCH] drm/amd/pm:
[AMD Official Use Only]
> -Original Message-
> From: Lazar, Lijo
> Sent: Tuesday, November 9, 2021 12:15 PM
> To: Quan, Evan ; amd-gfx@lists.freedesktop.org
> Cc: Deucher, Alexander ; Borislav Petkov
>
> Subject: Re: [PATCH] drm/amd/pm: avoid duplicate powergate/ungate
> setting
>
>
42 matches
Mail list logo