[RFC v3 00/12] Define and use reset domain for GPU recovery in amdgpu
This patchset is based on earlier work by Boris[1] that allowed to have an ordered workqueue at the driver level that will be used by the different schedulers to queue their timeout work. On top of that I also serialized any GPU reset we trigger from within amdgpu code to also go through the same ordered wq and in this way simplify somewhat our GPU reset code so we don't need to protect from concurrency by multiple GPU reset triggeres such as TDR on one hand and sysfs trigger or RAS trigger on the other hand. As advised by Christian and Daniel I defined a reset_domain struct such that all the entities that go through reset together will be serialized one against another. TDR triggered by multiple entities within the same domain due to the same reason will not be triggered as the first such reset will cancel all the pending resets. This is relevant only to TDR timers and not to triggered resets coming from RAS or SYSFS, those will still happen after the in flight resets finishes. v2: Add handling on SRIOV configuration, the reset notify coming from host and driver already trigger a work queue to handle the reset so drop this intermediate wq and send directly to timeout wq. (Shaoyun) v3: Lijo suggested puting 'adev->in_gpu_reset' in amdgpu_reset_domain struct. I followed his advise and also moved adev->reset_sem into same place. This in turn caused to do some follow-up refactor of the original patches where i decoupled amdgpu_reset_domain life cycle frolm XGMI hive because hive is destroyed and reconstructed for the case of reset the devices in the XGMI hive during probe for SRIOV See [2] while we need the reset sem and gpu_reset flag to always be present. This was attained by adding refcount to amdgpu_reset_domain so each device can safely point to it as long as it needs. [1] https://patchwork.kernel.org/project/dri-devel/patch/20210629073510.2764391-3-boris.brezil...@collabora.com/ [2] https://www.spinics.net/lists/amd-gfx/msg58836.html P.S Going through drm-misc-next and not amd-staging-drm-next as Boris work hasn't landed yet there. P.P.S Patches 8-12 are the refactor on top of the original V2 patchset. P.P.P.S I wasn't able yet to test the reworked code on XGMI SRIOV system because drm-misc-next fails to load there. Would appriciate if maybe jingwech can try it on his system like he tested V2. Andrey Grodzovsky (12): drm/amdgpu: Introduce reset domain drm/amdgpu: Move scheduler init to after XGMI is ready drm/amdgpu: Fix crash on modprobe drm/amdgpu: Serialize non TDR gpu recovery with TDRs drm/amd/virt: For SRIOV send GPU reset directly to TDR queue. drm/amdgpu: Drop hive->in_reset drm/amdgpu: Drop concurrent GPU reset protection for device drm/amdgpu: Rework reset domain to be refcounted. drm/amdgpu: Move reset sem into reset_domain drm/amdgpu: Move in_gpu_reset into reset_domain drm/amdgpu: Rework amdgpu_device_lock_adev Revert 'drm/amdgpu: annotate a false positive recursive locking' drivers/gpu/drm/amd/amdgpu/amdgpu.h | 15 +- drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c| 275 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 43 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c| 18 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 39 +++ drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h | 12 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 24 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.h | 3 +- drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c| 6 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 14 +- drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c | 19 +- drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c | 19 +- drivers/gpu/drm/amd/amdgpu/mxgpu_vi.c | 11 +- 16 files changed, 313 insertions(+), 199 deletions(-) -- 2.25.1
Re: [RFC v3 00/12] Define and use reset domain for GPU recovery in amdgpu
Just a gentle ping if people have more comments on this patch set ? Especially last 5 patches as first 7 are exact same as V2 and we already went over them mostly. Andrey On 2022-01-25 17:37, Andrey Grodzovsky wrote: This patchset is based on earlier work by Boris[1] that allowed to have an ordered workqueue at the driver level that will be used by the different schedulers to queue their timeout work. On top of that I also serialized any GPU reset we trigger from within amdgpu code to also go through the same ordered wq and in this way simplify somewhat our GPU reset code so we don't need to protect from concurrency by multiple GPU reset triggeres such as TDR on one hand and sysfs trigger or RAS trigger on the other hand. As advised by Christian and Daniel I defined a reset_domain struct such that all the entities that go through reset together will be serialized one against another. TDR triggered by multiple entities within the same domain due to the same reason will not be triggered as the first such reset will cancel all the pending resets. This is relevant only to TDR timers and not to triggered resets coming from RAS or SYSFS, those will still happen after the in flight resets finishes. v2: Add handling on SRIOV configuration, the reset notify coming from host and driver already trigger a work queue to handle the reset so drop this intermediate wq and send directly to timeout wq. (Shaoyun) v3: Lijo suggested puting 'adev->in_gpu_reset' in amdgpu_reset_domain struct. I followed his advise and also moved adev->reset_sem into same place. This in turn caused to do some follow-up refactor of the original patches where i decoupled amdgpu_reset_domain life cycle frolm XGMI hive because hive is destroyed and reconstructed for the case of reset the devices in the XGMI hive during probe for SRIOV See [2] while we need the reset sem and gpu_reset flag to always be present. This was attained by adding refcount to amdgpu_reset_domain so each device can safely point to it as long as it needs. [1] https://patchwork.kernel.org/project/dri-devel/patch/20210629073510.2764391-3-boris.brezil...@collabora.com/ [2] https://www.spinics.net/lists/amd-gfx/msg58836.html P.S Going through drm-misc-next and not amd-staging-drm-next as Boris work hasn't landed yet there. P.P.S Patches 8-12 are the refactor on top of the original V2 patchset. P.P.P.S I wasn't able yet to test the reworked code on XGMI SRIOV system because drm-misc-next fails to load there. Would appriciate if maybe jingwech can try it on his system like he tested V2. Andrey Grodzovsky (12): drm/amdgpu: Introduce reset domain drm/amdgpu: Move scheduler init to after XGMI is ready drm/amdgpu: Fix crash on modprobe drm/amdgpu: Serialize non TDR gpu recovery with TDRs drm/amd/virt: For SRIOV send GPU reset directly to TDR queue. drm/amdgpu: Drop hive->in_reset drm/amdgpu: Drop concurrent GPU reset protection for device drm/amdgpu: Rework reset domain to be refcounted. drm/amdgpu: Move reset sem into reset_domain drm/amdgpu: Move in_gpu_reset into reset_domain drm/amdgpu: Rework amdgpu_device_lock_adev Revert 'drm/amdgpu: annotate a false positive recursive locking' drivers/gpu/drm/amd/amdgpu/amdgpu.h | 15 +- drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c| 275 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 43 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c| 18 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 39 +++ drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h | 12 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 24 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.h | 3 +- drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c| 6 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 14 +- drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c | 19 +- drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c | 19 +- drivers/gpu/drm/amd/amdgpu/mxgpu_vi.c | 11 +- 16 files changed, 313 insertions(+), 199 deletions(-)
Re: [RFC v3 00/12] Define and use reset domain for GPU recovery in amdgpu
Just another ping, with Shyun's help I was able to do some smoke testing on XGMI SRIOV system (booting and triggering hive reset) and for now looks good. Andrey On 2022-01-28 14:36, Andrey Grodzovsky wrote: Just a gentle ping if people have more comments on this patch set ? Especially last 5 patches as first 7 are exact same as V2 and we already went over them mostly. Andrey On 2022-01-25 17:37, Andrey Grodzovsky wrote: This patchset is based on earlier work by Boris[1] that allowed to have an ordered workqueue at the driver level that will be used by the different schedulers to queue their timeout work. On top of that I also serialized any GPU reset we trigger from within amdgpu code to also go through the same ordered wq and in this way simplify somewhat our GPU reset code so we don't need to protect from concurrency by multiple GPU reset triggeres such as TDR on one hand and sysfs trigger or RAS trigger on the other hand. As advised by Christian and Daniel I defined a reset_domain struct such that all the entities that go through reset together will be serialized one against another. TDR triggered by multiple entities within the same domain due to the same reason will not be triggered as the first such reset will cancel all the pending resets. This is relevant only to TDR timers and not to triggered resets coming from RAS or SYSFS, those will still happen after the in flight resets finishes. v2: Add handling on SRIOV configuration, the reset notify coming from host and driver already trigger a work queue to handle the reset so drop this intermediate wq and send directly to timeout wq. (Shaoyun) v3: Lijo suggested puting 'adev->in_gpu_reset' in amdgpu_reset_domain struct. I followed his advise and also moved adev->reset_sem into same place. This in turn caused to do some follow-up refactor of the original patches where i decoupled amdgpu_reset_domain life cycle frolm XGMI hive because hive is destroyed and reconstructed for the case of reset the devices in the XGMI hive during probe for SRIOV See [2] while we need the reset sem and gpu_reset flag to always be present. This was attained by adding refcount to amdgpu_reset_domain so each device can safely point to it as long as it needs. [1] https://patchwork.kernel.org/project/dri-devel/patch/20210629073510.2764391-3-boris.brezil...@collabora.com/ [2] https://www.spinics.net/lists/amd-gfx/msg58836.html P.S Going through drm-misc-next and not amd-staging-drm-next as Boris work hasn't landed yet there. P.P.S Patches 8-12 are the refactor on top of the original V2 patchset. P.P.P.S I wasn't able yet to test the reworked code on XGMI SRIOV system because drm-misc-next fails to load there. Would appriciate if maybe jingwech can try it on his system like he tested V2. Andrey Grodzovsky (12): drm/amdgpu: Introduce reset domain drm/amdgpu: Move scheduler init to after XGMI is ready drm/amdgpu: Fix crash on modprobe drm/amdgpu: Serialize non TDR gpu recovery with TDRs drm/amd/virt: For SRIOV send GPU reset directly to TDR queue. drm/amdgpu: Drop hive->in_reset drm/amdgpu: Drop concurrent GPU reset protection for device drm/amdgpu: Rework reset domain to be refcounted. drm/amdgpu: Move reset sem into reset_domain drm/amdgpu: Move in_gpu_reset into reset_domain drm/amdgpu: Rework amdgpu_device_lock_adev Revert 'drm/amdgpu: annotate a false positive recursive locking' drivers/gpu/drm/amd/amdgpu/amdgpu.h | 15 +- drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 275 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 43 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c | 18 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 39 +++ drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h | 12 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 24 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.h | 3 +- drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 6 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 14 +- drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c | 19 +- drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c | 19 +- drivers/gpu/drm/amd/amdgpu/mxgpu_vi.c | 11 +- 16 files changed, 313 insertions(+), 199 deletions(-)
Re: [RFC v3 00/12] Define and use reset domain for GPU recovery in amdgpu
Hi Andrey, I have been testing your patch and it seems fine till now. Best Regards, Jingwen Chen On 2022/2/3 上午2:57, Andrey Grodzovsky wrote: > Just another ping, with Shyun's help I was able to do some smoke testing on > XGMI SRIOV system (booting and triggering hive reset) > and for now looks good. > > Andrey > > On 2022-01-28 14:36, Andrey Grodzovsky wrote: >> Just a gentle ping if people have more comments on this patch set ? >> Especially last 5 patches >> as first 7 are exact same as V2 and we already went over them mostly. >> >> Andrey >> >> On 2022-01-25 17:37, Andrey Grodzovsky wrote: >>> This patchset is based on earlier work by Boris[1] that allowed to have an >>> ordered workqueue at the driver level that will be used by the different >>> schedulers to queue their timeout work. On top of that I also serialized >>> any GPU reset we trigger from within amdgpu code to also go through the same >>> ordered wq and in this way simplify somewhat our GPU reset code so we don't >>> need >>> to protect from concurrency by multiple GPU reset triggeres such as TDR on >>> one >>> hand and sysfs trigger or RAS trigger on the other hand. >>> >>> As advised by Christian and Daniel I defined a reset_domain struct such that >>> all the entities that go through reset together will be serialized one >>> against >>> another. >>> >>> TDR triggered by multiple entities within the same domain due to the same >>> reason will not >>> be triggered as the first such reset will cancel all the pending resets. >>> This is >>> relevant only to TDR timers and not to triggered resets coming from RAS or >>> SYSFS, >>> those will still happen after the in flight resets finishes. >>> >>> v2: >>> Add handling on SRIOV configuration, the reset notify coming from host >>> and driver already trigger a work queue to handle the reset so drop this >>> intermediate wq and send directly to timeout wq. (Shaoyun) >>> >>> v3: >>> Lijo suggested puting 'adev->in_gpu_reset' in amdgpu_reset_domain struct. >>> I followed his advise and also moved adev->reset_sem into same place. This >>> in turn caused to do some follow-up refactor of the original patches >>> where i decoupled amdgpu_reset_domain life cycle frolm XGMI hive because >>> hive is destroyed and >>> reconstructed for the case of reset the devices in the XGMI hive during >>> probe for SRIOV See [2] >>> while we need the reset sem and gpu_reset flag to always be present. This >>> was attained >>> by adding refcount to amdgpu_reset_domain so each device can safely point >>> to it as long as >>> it needs. >>> >>> >>> [1] >>> https://patchwork.kernel.org/project/dri-devel/patch/20210629073510.2764391-3-boris.brezil...@collabora.com/ >>> [2] https://www.spinics.net/lists/amd-gfx/msg58836.html >>> >>> P.S Going through drm-misc-next and not amd-staging-drm-next as Boris work >>> hasn't landed yet there. >>> >>> P.P.S Patches 8-12 are the refactor on top of the original V2 patchset. >>> >>> P.P.P.S I wasn't able yet to test the reworked code on XGMI SRIOV system >>> because drm-misc-next fails to load there. >>> Would appriciate if maybe jingwech can try it on his system like he tested >>> V2. >>> >>> Andrey Grodzovsky (12): >>> drm/amdgpu: Introduce reset domain >>> drm/amdgpu: Move scheduler init to after XGMI is ready >>> drm/amdgpu: Fix crash on modprobe >>> drm/amdgpu: Serialize non TDR gpu recovery with TDRs >>> drm/amd/virt: For SRIOV send GPU reset directly to TDR queue. >>> drm/amdgpu: Drop hive->in_reset >>> drm/amdgpu: Drop concurrent GPU reset protection for device >>> drm/amdgpu: Rework reset domain to be refcounted. >>> drm/amdgpu: Move reset sem into reset_domain >>> drm/amdgpu: Move in_gpu_reset into reset_domain >>> drm/amdgpu: Rework amdgpu_device_lock_adev >>> Revert 'drm/amdgpu: annotate a false positive recursive locking' >>> >>> drivers/gpu/drm/amd/amdgpu/amdgpu.h | 15 +- >>> drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 10 +- >>> drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 275 ++ >>> drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 43 +-- >>> drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 +- >>> .../gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c | 18 +- >>> drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 39 +++ >>> drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h | 12 + >>> drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 2 + >>> drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 24 +- >>> drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.h | 3 +- >>> drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 6 +- >>> drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 14 +- >>> drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c | 19 +- >>> drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c | 19 +- >>> drivers/gpu/drm/amd/amdgpu/mxgpu_vi.c | 11 +- >>> 16 files changed, 313 insertions(+), 199 deletions(-) >>>
Re: [RFC v3 00/12] Define and use reset domain for GPU recovery in amdgpu
Thanks a lot! Andrey On 2022-02-09 01:06, JingWen Chen wrote: Hi Andrey, I have been testing your patch and it seems fine till now. Best Regards, Jingwen Chen On 2022/2/3 上午2:57, Andrey Grodzovsky wrote: Just another ping, with Shyun's help I was able to do some smoke testing on XGMI SRIOV system (booting and triggering hive reset) and for now looks good. Andrey On 2022-01-28 14:36, Andrey Grodzovsky wrote: Just a gentle ping if people have more comments on this patch set ? Especially last 5 patches as first 7 are exact same as V2 and we already went over them mostly. Andrey On 2022-01-25 17:37, Andrey Grodzovsky wrote: This patchset is based on earlier work by Boris[1] that allowed to have an ordered workqueue at the driver level that will be used by the different schedulers to queue their timeout work. On top of that I also serialized any GPU reset we trigger from within amdgpu code to also go through the same ordered wq and in this way simplify somewhat our GPU reset code so we don't need to protect from concurrency by multiple GPU reset triggeres such as TDR on one hand and sysfs trigger or RAS trigger on the other hand. As advised by Christian and Daniel I defined a reset_domain struct such that all the entities that go through reset together will be serialized one against another. TDR triggered by multiple entities within the same domain due to the same reason will not be triggered as the first such reset will cancel all the pending resets. This is relevant only to TDR timers and not to triggered resets coming from RAS or SYSFS, those will still happen after the in flight resets finishes. v2: Add handling on SRIOV configuration, the reset notify coming from host and driver already trigger a work queue to handle the reset so drop this intermediate wq and send directly to timeout wq. (Shaoyun) v3: Lijo suggested puting 'adev->in_gpu_reset' in amdgpu_reset_domain struct. I followed his advise and also moved adev->reset_sem into same place. This in turn caused to do some follow-up refactor of the original patches where i decoupled amdgpu_reset_domain life cycle frolm XGMI hive because hive is destroyed and reconstructed for the case of reset the devices in the XGMI hive during probe for SRIOV See [2] while we need the reset sem and gpu_reset flag to always be present. This was attained by adding refcount to amdgpu_reset_domain so each device can safely point to it as long as it needs. [1] https://patchwork.kernel.org/project/dri-devel/patch/20210629073510.2764391-3-boris.brezil...@collabora.com/ [2] https://www.spinics.net/lists/amd-gfx/msg58836.html P.S Going through drm-misc-next and not amd-staging-drm-next as Boris work hasn't landed yet there. P.P.S Patches 8-12 are the refactor on top of the original V2 patchset. P.P.P.S I wasn't able yet to test the reworked code on XGMI SRIOV system because drm-misc-next fails to load there. Would appriciate if maybe jingwech can try it on his system like he tested V2. Andrey Grodzovsky (12): drm/amdgpu: Introduce reset domain drm/amdgpu: Move scheduler init to after XGMI is ready drm/amdgpu: Fix crash on modprobe drm/amdgpu: Serialize non TDR gpu recovery with TDRs drm/amd/virt: For SRIOV send GPU reset directly to TDR queue. drm/amdgpu: Drop hive->in_reset drm/amdgpu: Drop concurrent GPU reset protection for device drm/amdgpu: Rework reset domain to be refcounted. drm/amdgpu: Move reset sem into reset_domain drm/amdgpu: Move in_gpu_reset into reset_domain drm/amdgpu: Rework amdgpu_device_lock_adev Revert 'drm/amdgpu: annotate a false positive recursive locking' drivers/gpu/drm/amd/amdgpu/amdgpu.h | 15 +- drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 10 +- drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 275 ++ drivers/gpu/drm/amd/amdgpu/amdgpu_fence.c | 43 +-- drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 2 +- .../gpu/drm/amd/amdgpu/amdgpu_ras_eeprom.c | 18 +- drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 39 +++ drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h | 12 + drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 2 + drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.c | 24 +- drivers/gpu/drm/amd/amdgpu/amdgpu_xgmi.h | 3 +- drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c | 6 +- drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 14 +- drivers/gpu/drm/amd/amdgpu/mxgpu_ai.c | 19 +- drivers/gpu/drm/amd/amdgpu/mxgpu_nv.c | 19 +- drivers/gpu/drm/amd/amdgpu/mxgpu_vi.c | 11 +- 16 files changed, 313 insertions(+), 199 deletions(-)