drm/amdgpu: block ring buffer access during GPU recovery
Subsystems (2)
PATCH 1/1 drm/amdgpu: block ring buffer access during GPU recovery
When GPU is in reset, its status isn't stable and ring buffer also need be reset when resuming. Therefore driver should protect GPU recovery thread from ring buffer accessed by other threads. Otherwise GPU will randomly hang during recovery.
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +15 -4
@@ -319,8 +319,13 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, {uint32_t ret; -if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_rreg (adev, reg); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { + ret =amdgpu_kiq_rreg (adev, reg); +up_read (&adev->reset_sem); +return ret; + }if ((reg *4 ) < adev->rmmio_size) ret =readl (((void __iomem *)adev->rmmio) + (reg *4 ));
@@ -332,6 +337,7 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, ret =readl (((void __iomem *)adev->rmmio) + (mmMM_DATA *4 ));spin_unlock_irqrestore (&adev->mmio_idx_lock, flags); } +trace_amdgpu_mm_rreg (adev->pdev->device, reg, ret);return ret; }
@@ -407,8 +413,13 @@ void static inline amdgpu_mm_wreg_mmio(struct amdgpu_device *adev, uint32_t reg,void amdgpu_mm_wreg (struct amdgpu_device *adev,uint32_t reg,uint32_t v,uint32_t acc_flags) { -if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_wreg (adev, reg, v); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { +amdgpu_kiq_wreg (adev, reg, v); +up_read (&adev->reset_sem); +return ; + }amdgpu_mm_wreg_mmio (adev, reg, v, acc_flags); }
drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +3 -1
@@ -287,7 +287,7 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];const unsigned eng =17 ;
@@ -297,6 +297,8 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); + +up_read (&adev->reset_sem);return ; }
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +7 -3
@@ -501,12 +501,13 @@ static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); +up_read (&adev->reset_sem);return ; }
@@ -599,7 +600,8 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (amdgpu_in_reset (adev))return -EIO ; -if (ring->sched.ready) { +if (ring->sched.ready && +down_read_trylock (&adev->reset_sem)) {/* Vega20+XGMI caches PTEs in TC and TLB. Add a * heavy-weight TLB flush (type 2), which flushes * both. Due to a race condition with concurrent
@@ -626,6 +628,7 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (r) {amdgpu_ring_undo (ring);spin_unlock (&adev->gfx.kiq.ring_lock); +up_read (&adev->reset_sem);return -ETIME ; }
@@ -634,9 +637,10 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, r =amdgpu_fence_wait_polling (ring, seq, adev->usec_timeout);if (r <1 ) {dev_err (adev->dev,"wait for kiq fence error: %ld.\n" , r); +up_read (&adev->reset_sem);return -ETIME ; } - +up_read (&adev->reset_sem);return 0 ; }
Signature · 6 lines
2.17.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 14 lines · Dennis Li
When GPU is in reset, its status isn't stable and ring buffer also need be reset when resuming. Therefore driver should protect GPU recovery thread from ring buffer accessed by other threads. Otherwise GPU will randomly hang during recovery.
Signed-off-by Dennis Li <[email protected]>drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -319,8 +319,13 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, {uint32_t ret;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, · line 328
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +7
-2-if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_rreg (adev, reg); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { + ret =amdgpu_kiq_rreg (adev, reg); +up_read (&adev->reset_sem); +return ret; + }
Isn't adev->reset_sem non-recursive ? How this works when you try to access registers from within GPU reset thread while adev->reset_sem is already write locked from amdgpu_device_lock_adev earlier in the same thread ?
Andrey
Quoted 88 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +8
-2if ((reg *4 ) < adev->rmmio_size) ret =readl (((void __iomem *)adev->rmmio) + (reg *4 ));@@ -332,6 +337,7 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, ret =readl (((void __iomem *)adev->rmmio) + (mmMM_DATA *4 ));spin_unlock_irqrestore (&adev->mmio_idx_lock, flags); } +trace_amdgpu_mm_rreg (adev->pdev->device, reg, ret);return ret; }@@ -407,8 +413,13 @@ void static inline amdgpu_mm_wreg_mmio(struct amdgpu_device *adev, uint32_t reg,void amdgpu_mm_wreg (struct amdgpu_device *adev,uint32_t reg,uint32_t v,uint32_t acc_flags) { -if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_wreg (adev, reg, v); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { +amdgpu_kiq_wreg (adev, reg, v); +up_read (&adev->reset_sem); +return ; + }amdgpu_mm_wreg_mmio (adev, reg, v, acc_flags); }drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +3
-1@@ -287,7 +287,7 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];const unsigned eng =17 ;@@ -297,6 +297,8 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); + +up_read (&adev->reset_sem);return ; }drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +5
-2@@ -501,12 +501,13 @@ static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); +up_read (&adev->reset_sem);return ; }@@ -599,7 +600,8 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (amdgpu_in_reset (adev))return -EIO ; -if (ring->sched.ready) { +if (ring->sched.ready && +down_read_trylock (&adev->reset_sem)) {/* Vega20+XGMI caches PTEs in TC and TLB. Add a * heavy-weight TLB flush (type 2), which flushes * both. Due to a race condition with concurrent @@ -626,6 +628,7 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (r) {amdgpu_ring_undo (ring);spin_unlock (&adev->gfx.kiq.ring_lock); +up_read (&adev->reset_sem);return -ETIME ; }@@ -634,9 +637,10 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, r =amdgpu_fence_wait_polling (ring, seq, adev->usec_timeout);
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, · line 646
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +2
-1if (r <1 ) {dev_err (adev->dev,"wait for kiq fence error: %ld.\n" , r); +up_read (&adev->reset_sem);return -ETIME ; } - +up_read (&adev->reset_sem);return 0 ; }
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
[AMD Official Use Only - Internal Distribution Only]
Hi, Andrey,
RE- Isn't adev->reset_sem non-recursive ? How this works when you try to access registers from within GPU reset thread while adev->reset_sem is already write locked from amdgpu_device_lock_adev earlier in the same thread ?
Deli: down_read_trylock will fail in this case, return false immediately and will not lock adev->reset_sem. In GPU reset thread, we should use MMIO instead of KIQ to access registers.
Best Regards
Dennis Li
-----Original Message-----
From: Grodzovsky, Andrey <[email protected]>
Sent: Tuesday, September 1, 2020 9:40 AM
To: Li, Dennis <[email protected]>; [email protected]; Deucher, Alexander <[email protected]>; Kuehling, Felix <[email protected]>; Zhang, Hawking <[email protected]>; Koenig, Christian <[email protected]>
Subject: Re: [PATCH] drm/amdgpu: block ring buffer access during GPU recovery
Quoted 15 lines · Dennis Li
When GPU is in reset, its status isn't stable and ring buffer also need be reset when resuming. Therefore driver should protect GPU recovery thread from ring buffer accessed by other threads. Otherwise GPU will randomly hang during recovery.
Signed-off-by Dennis Li <[email protected]>drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -319,8 +319,13 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, {uint32_t ret;
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, · line 328
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +7
-2-if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_rreg (adev, reg); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { + ret =amdgpu_kiq_rreg (adev, reg); +up_read (&adev->reset_sem); +return ret; + }
Isn't adev->reset_sem non-recursive ? How this works when you try to access registers from within GPU reset thread while adev->reset_sem is already write locked from amdgpu_device_lock_adev earlier in the same thread ?
Andrey
Quoted 91 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-2if ((reg *4 ) < adev->rmmio_size) ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); @@ -332,6 +337,7 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));spin_unlock_irqrestore (&adev->mmio_idx_lock, flags); } +trace_amdgpu_mm_rreg (adev->pdev->device, reg, ret);return ret; }@@ -407,8 +413,13 @@ void static inline amdgpu_mm_wreg_mmio(struct amdgpu_device *adev, uint32_t reg,void amdgpu_mm_wreg (struct amdgpu_device *adev,uint32_t reg,uint32_t v,uint32_t acc_flags) { -if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_wreg (adev, reg, v); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { +amdgpu_kiq_wreg (adev, reg, v); +up_read (&adev->reset_sem); +return ; + }amdgpu_mm_wreg_mmio (adev, reg, v, acc_flags); }drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +3
-1@@ -287,7 +287,7 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];const unsigned eng =17 ;@@ -297,6 +297,8 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,amdgpu_virt_kiq_reg_write_reg_wait (adev ,req ,ack ,inv_req , 1 << vmid); + +up_read (&adev->reset_sem);return ; }drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +4
-2@@ -501,12 +501,13 @@ static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); +up_read (&adev->reset_sem);return ; }@@ -599,7 +600,8 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (amdgpu_in_reset (adev))return -EIO ; -if (ring->sched.ready) { +if (ring->sched.ready && +down_read_trylock (&adev->reset_sem)) {/* Vega20+XGMI caches PTEs in TC and TLB. Add a * heavy-weight TLB flush (type 2), which flushes * both. Due to a race condition with concurrent @@ -626,6 +628,7 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct
amdgpu_device*adev,
if (r) {
amdgpu_ring_undo(ring);spin_unlock(&adev->gfx.kiq.ring_lock);drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +1
-0+up_read (&adev->reset_sem);return -ETIME ; }@@ -634,9 +637,10 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, r =amdgpu_fence_wait_polling (ring, seq, adev->usec_timeout);
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, · line 646
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +2
-1if (r <1 ) {dev_err (adev->dev,"wait for kiq fence error: %ld.\n" , r); +up_read (&adev->reset_sem);return -ETIME ; } - +up_read (&adev->reset_sem);return 0 ; }
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Yeah, correct.
What we maybe should do is to add a WARN_ON() which tests if the current thread is the one which has locked the semaphore to catch this case.
Regards,
Christian.
Am 01.09.20 um 04:45 schrieb Li, Dennis:
Quoted 138 lines
[AMD Official Use Only - Internal Distribution Only]
Hi, Andrey,
RE- Isn't
adev->reset_semnon-recursive ? How this works when you try to access registers from within GPU reset thread whileadev->reset_semis already write locked fromamdgpu_device_lock_adevearlier in the same thread ?Deli:
down_read_trylockwill fail in this case, return false immediately and will not lockadev->reset_sem. In GPU reset thread, we should use MMIO instead of KIQ to access registers.Best Regards
Dennis Li
-----Original Message-----
From: Grodzovsky, Andrey <[email protected]>
Sent: Tuesday, September 1, 2020 9:40 AM
To: Li, Dennis <[email protected]>; [email protected]; Deucher, Alexander <[email protected]>; Kuehling, Felix <[email protected]>; Zhang, Hawking <[email protected]>; Koenig, Christian <[email protected]>
Subject: Re: [PATCH] drm/amdgpu: block ring buffer access during GPU recoveryQuoted 15 lines · Dennis Li
When GPU is in reset, its status isn't stable and ring buffer also need be reset when resuming. Therefore driver should protect GPU recovery thread from ring buffer accessed by other threads. Otherwise GPU will randomly hang during recovery.
Signed-off-by Dennis Li <[email protected]>drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -319,8 +319,13 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, {uint32_t ret;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, · line 328
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +7
-2-if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_rreg (adev, reg); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { + ret =amdgpu_kiq_rreg (adev, reg); +up_read (&adev->reset_sem); +return ret; + }Isn't
adev->reset_semnon-recursive ? How this works when you try to access registers from within GPU reset thread whileadev->reset_semis already write locked fromamdgpu_device_lock_adevearlier in the same thread ?Andrey
Quoted 89 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-2if ((reg *4 ) < adev->rmmio_size) ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); @@ -332,6 +337,7 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));spin_unlock_irqrestore (&adev->mmio_idx_lock, flags); } +trace_amdgpu_mm_rreg (adev->pdev->device, reg, ret);return ret; }@@ -407,8 +413,13 @@ void static inline amdgpu_mm_wreg_mmio(struct amdgpu_device *adev, uint32_t reg,void amdgpu_mm_wreg (struct amdgpu_device *adev,uint32_t reg,uint32_t v,uint32_t acc_flags) { -if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_wreg (adev, reg, v); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { +amdgpu_kiq_wreg (adev, reg, v); +up_read (&adev->reset_sem); +return ; + }amdgpu_mm_wreg_mmio (adev, reg, v, acc_flags); }drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +3
-1@@ -287,7 +287,7 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];const unsigned eng =17 ;@@ -297,6 +297,8 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,amdgpu_virt_kiq_reg_write_reg_wait (adev ,req ,ack ,inv_req , 1 << vmid); + +up_read (&adev->reset_sem);return ; }drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +4
-2@@ -501,12 +501,13 @@ static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); +up_read (&adev->reset_sem);return ; }@@ -599,7 +600,8 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (amdgpu_in_reset (adev))return -EIO ; -if (ring->sched.ready) { +if (ring->sched.ready && +down_read_trylock (&adev->reset_sem)) {/* Vega20+XGMI caches PTEs in TC and TLB. Add a * heavy-weight TLB flush (type 2), which flushes * both. Due to a race condition with concurrent @@ -626,6 +628,7 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct
amdgpu_device*adev,if (r) { amdgpu_ring_undo(ring); spin_unlock(&adev->gfx.kiq.ring_lock);drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +1
-0+up_read (&adev->reset_sem);return -ETIME ; }
Quoted 16 lines
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, · line 646
Quoted 12 lines
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +2
-1@@ -634,9 +637,10 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, r =amdgpu_fence_wait_polling (ring, seq, adev->usec_timeout);if (r <1 ) {dev_err (adev->dev,"wait for kiq fence error: %ld.\n" , r); +up_read (&adev->reset_sem);return -ETIME ; } - +up_read (&adev->reset_sem);return 0 ; }_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Andrey Grodzovsky
Acked-byNow i get it, I missed the 'else' part.
Andrey
Quoted 138 lines · Dennis
[AMD Official Use Only - Internal Distribution Only]
Hi, Andrey,
RE- Isn't
adev->reset_semnon-recursive ? How this works when you try to access registers from within GPU reset thread whileadev->reset_semis already write locked fromamdgpu_device_lock_adevearlier in the same thread ?Deli:
down_read_trylockwill fail in this case, return false immediately and will not lockadev->reset_sem. In GPU reset thread, we should use MMIO instead of KIQ to access registers.Best Regards
Dennis Li
-----Original Message-----
From: Grodzovsky, Andrey <[email protected]>
Sent: Tuesday, September 1, 2020 9:40 AM
To: Li, Dennis <[email protected]>; [email protected]; Deucher, Alexander <[email protected]>; Kuehling, Felix <[email protected]>; Zhang, Hawking <[email protected]>; Koenig, Christian <[email protected]>
Subject: Re: [PATCH] drm/amdgpu: block ring buffer access during GPU recoveryQuoted 15 lines · Dennis Li
When GPU is in reset, its status isn't stable and ring buffer also need be reset when resuming. Therefore driver should protect GPU recovery thread from ring buffer accessed by other threads. Otherwise GPU will randomly hang during recovery.
Signed-off-by Dennis Li <[email protected]>drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0@@ -319,8 +319,13 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, {uint32_t ret;drivers/gpu/drm/amd/amdgpu/amdgpu_device.c uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, · line 328
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +7
-2-if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_rreg (adev, reg); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { + ret =amdgpu_kiq_rreg (adev, reg); +up_read (&adev->reset_sem); +return ret; + }Isn't
adev->reset_semnon-recursive ? How this works when you try to access registers from within GPU reset thread whileadev->reset_semis already write locked fromamdgpu_device_lock_adevearlier in the same thread ?Andrey
Quoted 89 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +9
-2if ((reg *4 ) < adev->rmmio_size) ret = readl(((void __iomem *)adev->rmmio) + (reg * 4)); @@ -332,6 +337,7 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, ret = readl(((void __iomem *)adev->rmmio) + (mmMM_DATA * 4));spin_unlock_irqrestore (&adev->mmio_idx_lock, flags); } +trace_amdgpu_mm_rreg (adev->pdev->device, reg, ret);return ret; }@@ -407,8 +413,13 @@ void static inline amdgpu_mm_wreg_mmio(struct amdgpu_device *adev, uint32_t reg,void amdgpu_mm_wreg (struct amdgpu_device *adev,uint32_t reg,uint32_t v,uint32_t acc_flags) { -if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_wreg (adev, reg, v); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { +amdgpu_kiq_wreg (adev, reg, v); +up_read (&adev->reset_sem); +return ; + }amdgpu_mm_wreg_mmio (adev, reg, v, acc_flags); }drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +3
-1@@ -287,7 +287,7 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];const unsigned eng =17 ;@@ -297,6 +297,8 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,amdgpu_virt_kiq_reg_write_reg_wait (adev ,req ,ack ,inv_req , 1 << vmid); + +up_read (&adev->reset_sem);return ; }drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +4
-2@@ -501,12 +501,13 @@ static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); +up_read (&adev->reset_sem);return ; }@@ -599,7 +600,8 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (amdgpu_in_reset (adev))return -EIO ; -if (ring->sched.ready) { +if (ring->sched.ready && +down_read_trylock (&adev->reset_sem)) {/* Vega20+XGMI caches PTEs in TC and TLB. Add a * heavy-weight TLB flush (type 2), which flushes * both. Due to a race condition with concurrent @@ -626,6 +628,7 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct
amdgpu_device*adev,if (r) { amdgpu_ring_undo(ring); spin_unlock(&adev->gfx.kiq.ring_lock);drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +1
-0+up_read (&adev->reset_sem);return -ETIME ; }
Quoted 12 lines
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, · line 646
Quoted 12 lines
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +2
-1@@ -634,9 +637,10 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, r =amdgpu_fence_wait_polling (ring, seq, adev->usec_timeout);if (r <1 ) {dev_err (adev->dev,"wait for kiq fence error: %ld.\n" , r); +up_read (&adev->reset_sem);return -ETIME ; } - +up_read (&adev->reset_sem);return 0 ; }
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Am 01.09.20 um 03:17 schrieb Dennis Li:
Quoted 4 lines
When GPU is in reset, its status isn't stable and ring buffer also need be reset when resuming. Therefore driver should protect GPU recovery thread from ring buffer accessed by other threads. Otherwise GPU will randomly hang during recovery.
One style comment inline, apart from that looks good to me.
Quoted 8 lines
Signed-off-by Dennis Li <[email protected]>drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0
-0
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, · line 324
Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +3
-2@@ -319,8 +319,13 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, {uint32_t ret; -if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_rreg (adev, reg); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) {
This is not correctly indented, the following lines of an if clause should be under the ( of the first line. Same for other cases as well.
Regards,
Christian.
Quoted 92 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +12
-2+ ret =amdgpu_kiq_rreg (adev, reg); +up_read (&adev->reset_sem); +return ret; + }if ((reg *4 ) < adev->rmmio_size) ret =readl (((void __iomem *)adev->rmmio) + (reg *4 ));@@ -332,6 +337,7 @@ uint32_t amdgpu_mm_rreg(struct amdgpu_device *adev, uint32_t reg, ret =readl (((void __iomem *)adev->rmmio) + (mmMM_DATA *4 ));spin_unlock_irqrestore (&adev->mmio_idx_lock, flags); } +trace_amdgpu_mm_rreg (adev->pdev->device, reg, ret);return ret; }@@ -407,8 +413,13 @@ void static inline amdgpu_mm_wreg_mmio(struct amdgpu_device *adev, uint32_t reg,void amdgpu_mm_wreg (struct amdgpu_device *adev,uint32_t reg,uint32_t v,uint32_t acc_flags) { -if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) &&amdgpu_sriov_runtime (adev)) -return amdgpu_kiq_wreg (adev, reg, v); +if (!(acc_flags &AMDGPU_REGS_NO_KIQ ) && +amdgpu_sriov_runtime (adev) && +down_read_trylock (&adev->reset_sem)) { +amdgpu_kiq_wreg (adev, reg, v); +up_read (&adev->reset_sem); +return ; + }amdgpu_mm_wreg_mmio (adev, reg, v, acc_flags); }drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c +3
-1@@ -287,7 +287,7 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {struct amdgpu_vmhub *hub = &adev->vmhub[vmhub];const unsigned eng =17 ;@@ -297,6 +297,8 @@ static void gmc_v10_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); + +up_read (&adev->reset_sem);return ; }drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +5
-2@@ -501,12 +501,13 @@ static void gmc_v9_0_flush_gpu_tlb(struct amdgpu_device *adev, uint32_t vmid,*/ if (adev->gfx.kiq.ring.sched.ready && (amdgpu_sriov_runtime (adev) || !amdgpu_sriov_vf (adev)) && - !amdgpu_in_reset (adev)) { +down_read_trylock (&adev->reset_sem)) {uint32_t req = hub->vm_inv_eng0_req + hub->eng_distance * eng;uint32_t ack = hub->vm_inv_eng0_ack + hub->eng_distance * eng;amdgpu_virt_kiq_reg_write_reg_wait (adev, req, ack, inv_req,1 << vmid); +up_read (&adev->reset_sem);return ; }@@ -599,7 +600,8 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (amdgpu_in_reset (adev))return -EIO ; -if (ring->sched.ready) { +if (ring->sched.ready && +down_read_trylock (&adev->reset_sem)) {/* Vega20+XGMI caches PTEs in TC and TLB. Add a * heavy-weight TLB flush (type 2), which flushes * both. Due to a race condition with concurrent @@ -626,6 +628,7 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev,if (r) {amdgpu_ring_undo (ring);spin_unlock (&adev->gfx.kiq.ring_lock); +up_read (&adev->reset_sem);return -ETIME ; }@@ -634,9 +637,10 @@ static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, r =amdgpu_fence_wait_polling (ring, seq, adev->usec_timeout);
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c static int gmc_v9_0_flush_gpu_tlb_pasid(struct amdgpu_device *adev, · line 646
Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c +2
-1if (r <1 ) {dev_err (adev->dev,"wait for kiq fence error: %ld.\n" , r); +up_read (&adev->reset_sem);return -ETIME ; } - +up_read (&adev->reset_sem);return 0 ; }
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx