drm/amdkfd: only allow heavy-weight TLB flush on some ASICs for SVM too
Subsystems (3)
PATCH 1/1 drm/amdkfd: only allow heavy-weight TLB flush on some ASICs for SVM too
The idea is from commit a50fe7078035 ("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commit f61c40c0757a ("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.
--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -8
@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8 -0
@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +bool kfd_is_locked (void );/* Compute profile */
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3 -1
@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); }return r;
Signature · 1 line
2.25.1
Quoted 37 lines · Lang Yu
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid); · line 1338
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8
-0void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +
It is a cosmetic change for function kfd_flush_tlb_after_unmap, and not related to the topic. You can separate that into another patch.
Regards,
Eric
Quoted 18 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0
-0bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); }return r;
Quoted 41 lines · Eric Huang
Quoted 38 lines · Lang Yu
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +2
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid); · line 1338
Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +6
-0+return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +It is a cosmetic change for function
kfd_flush_tlb_after_unmap, and not related to the topic. You can separate that into another patch.
Okay. Thanks.
Regards,
Lang
Quoted 19 lines
Regards,
EricQuoted 16 lines
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +0
-0bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); }return r;
Felix Kuehling
maintainerAm 2022-04-14 um 04:19 schrieb Lang Yu:
Quoted 55 lines
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0
-0
drivers/gpu/drm/amd/amdkfd/kfd_svm.c svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, · line 1234
Quoted 7 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT );
Then you probably need to add another flush_tlb call in svm_range_map_to_gpus.
Regards,
Felix
Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0
-0}return r;
Quoted 55 lines · Felix Kuehling
Am 2022-04-14 um 04:19 schrieb Lang Yu:
Quoted 53 lines
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0
-0@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ;
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, · line 1234
Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1} -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT );Then you probably need to add another
flush_tlbcall insvm_range_map_to_gpus.
There is a TLB_FLUSH_LEGACY call in svm_range_map_to_gpus same with kfd_ioctl_map_memory_to_gpu. Do we still need to add another one?
Regards,
Lang
Quoted 6 lines
Regards,
FelixQuoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0
-0}return r;
Am 2022-04-14 um 22:47 schrieb Lang Yu:
Quoted 62 lines · Felix Kuehling
Quoted 53 lines
Am 2022-04-14 um 04:19 schrieb Lang Yu:
Quoted 52 lines
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0
-0@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start, · line 1234
Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT );Then you probably need to add another
flush_tlbcall insvm_range_map_to_gpus.
Quoted 2 lines
There is a
TLB_FLUSH_LEGACYcall insvm_range_map_to_gpussame withkfd_ioctl_map_memory_to_gpu. Do we still need to add another one?
Right, I missed that one. I think that should cover it and the patch looks good to me.
Quoted 10 lines
Regards,
LangQuoted 6 lines
Regards,
FelixQuoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_svm.c +0
-0}return r;
Dear Lang,
Thank you for the patch.
Am 14.04.22 um 10:19 schrieb Lang Yu:
Quoted 4 lines
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.
Please list the ASICs, you know of having problems, and even add how to reproduce this.
Kind regards,
Paul
Quoted 60 lines
Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); }return r;
Quoted 5 lines · Paul Menzel
Dear Lang,
Thank you for the patch.
Quoted 8 lines
Am 14.04.22 um 10:19 schrieb Lang Yu:
Quoted 4 lines
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.Please list the ASICs, you know of having problems, and even add how to reproduce this.
Actually, this is ported from previous commits. You can find more details from the commits I mentioned. At the moment the ASICs except Aldebaran and Arcturus probably have the problem. And running a SVM application could reproduce the issue.
Thanks,
Lang
Quoted 61 lines
Kind regards,
Paul
Quoted 55 lines
Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); }return r;
Dear Lang,
Am 15.04.22 um 05:20 schrieb Lang Yu:
Quoted 8 lines · Paul Menzel
Quoted 7 lines
Am 14.04.22 um 10:19 schrieb Lang Yu:
Quoted 4 lines
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.
Quoted 6 lines
Quoted 2 lines
Please list the ASICs, you know of having problems, and even add how to reproduce this.
Actually, this is ported from previous commits. You can find more details from the commits I mentioned. At the moment the ASICs except Aldebaran and Arcturus probably have the problem.
I think, it’s always good to make it as easy as possible for reviewers and, later, people reading a commit, and include the necessary information directly in the commit message. It’d be great if you amended the commit message.
Quoted 1 line
And running a SVM application could reproduce the issue.
Thanks. How will it fail though?
(Also, a small implementation note would be nice to have. Maybe: Move the helper function into the header kfd_priv.h, and use in svm_range_unmap_from_gpus().)
Kind regards,
Paul
Quoted 55 lines
Quoted 55 lines
Quoted 55 lines
Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); }return r;
Quoted 15 lines · Paul Menzel
Dear Lang,
Am 15.04.22 um 05:20 schrieb Lang Yu:
Quoted 11 lines · Paul Menzel
Quoted 9 lines
Am 14.04.22 um 10:19 schrieb Lang Yu:
Quoted 4 lines
The idea is from commit
a50fe7078035("drm/amdkfd: Only apply heavy-weight TLB flush on Aldebaran") and commitf61c40c0757a("drm/amdkfd: enable heavy-weight TLB flush on Arcturus"). Otherwise, we will run into problems on some ASICs when running SVM applications.Please list the ASICs, you know of having problems, and even add how to reproduce this.
Quoted 8 lines
Quoted 3 lines
Actually, this is ported from previous commits. You can find more details from the commits I mentioned. At the moment the ASICs except Aldebaran and Arcturus probably have the problem.
I think, it’s always good to make it as easy as possible for reviewers and, later, people reading a commit, and include the necessary information directly in the commit message. It’d be great if you amended the commit message.
Yes, I agree with you. Will amended the commit message.
Quoted 3 lines
Quoted 1 line
And running a SVM application could reproduce the issue.
Thanks. How will it fail though?
Will describe more details in commit message.
Quoted 3 lines
(Also, a small implementation note would be nice to have. Maybe: Move the helper function into the header
kfd_priv.h, and use insvm_range_unmap_from_gpus().)
Will separate this change into another patch suggested by Eric.
Thanks,
Lang
Quoted 60 lines
Kind regards,
Paul
Quoted 55 lines
Quoted 55 lines
Quoted 55 lines
Signed-off-by Lang Yu <[email protected]>--- drivers/gpu/drm/amd/amdkfd/kfd_chardev.c | 8 -------- drivers/gpu/drm/amd/amdkfd/kfd_priv.h | 8 ++++++++ drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 4 +++- 3 files changed, 11 insertions(+), 9 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0
-8@@ -1128,14 +1128,6 @@ static int kfd_ioctl_free_memory_of_gpu(struct file *filep,return ret; } -static bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) -{ -return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || - (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && - dev->adev->sdma.instance[0 ].fw_version >=18 ) || -KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); -} -static int kfd_ioctl_map_memory_to_gpu (struct file *filep,struct kfd_process *p,void *data) {drivers/gpu/drm/amd/amdkfd/kfd_priv.h +8
-0@@ -1328,6 +1328,14 @@ void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid);void kfd_flush_tlb (struct kfd_process_device *pdd,enum TLB_FLUSH_TYPE type); +static inline bool kfd_flush_tlb_after_unmap (struct kfd_dev *dev) +{ +return KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,2 ) || + (KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,1 ) && + dev->adev->sdma.instance[0 ].fw_version >=18 ) || +KFD_GC_VERSION (dev) ==IP_VERSION (9 ,4 ,0 ); +} +bool kfd_is_locked (void );/* Compute profile */ drivers/gpu/drm/amd/amdkfd/kfd_svm.c +3
-1@@ -1229,7 +1229,9 @@ svm_range_unmap_from_gpus(struct svm_range *prange, unsigned long start,if (r)break ; } -kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); + +if (kfd_flush_tlb_after_unmap (pdd->dev)) +kfd_flush_tlb (pdd,TLB_FLUSH_HEAVYWEIGHT ); }return r;