On 09/07/ , Felix Kuehling wrote:
> We need heavy-weight flushes not just for SVM. If this is broken it will
> affect ROCm either way.
Currently, TLB_FLUSH_HEAVYWEIGHT is called in 2 places,
1, kfd_ioctl_unmap_memory_from_gpu()
Under following conditions.
KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3) ||
KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2) ||
(KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 1) && dev->sdma_fw_version >= 18) ||
KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 0);
2, svm_range_unmap_from_gpus()
Unconditional.
That means TLB_FLUSH_HEAVYWEIGHT defect won't affect ROCm if
we don't use SVM for ASICs except MI series.
Regards,
Lang
> Regards,
> Felix
>
>
> On 2023-09-07 08:08, Lang Yu wrote:
> > GC 10.1.3/4 have problems with TLB_FLUSH_HEAVYWEIGHT
> > which is used by SVM in svm_range_unmap_from_gpus().
> > This causes problems on GC 10.1.3/4.
> >
> > Signed-off-by: Lang Yu <Lang.Yu@xxxxxxx>
> > ---
> > drivers/gpu/drm/amd/amdkfd/kfd_migrate.c | 22 +++++++++++++++++-----
> > 1 file changed, 17 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> > index 7d82c7da223a..dd3db3d88d59 100644
> > --- a/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> > +++ b/drivers/gpu/drm/amd/amdkfd/kfd_migrate.c
> > @@ -992,6 +992,22 @@ static const struct dev_pagemap_ops svm_migrate_pgmap_ops = {
> > /* Each VRAM page uses sizeof(struct page) on system memory */
> > #define SVM_HMM_PAGE_STRUCT_SIZE(size) ((size)/PAGE_SIZE * sizeof(struct page))
> > +static inline bool is_zone_device_needed(struct amdgpu_device *adev)
> > +{
> > + /* Page migration works on gfx9 or newer */
> > + if (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 0, 1))
> > + return false;
> > +
> > + if (adev->ip_versions[GC_HWIP][0] == IP_VERSION(10, 1, 3) ||
> > + adev->ip_versions[GC_HWIP][0] == IP_VERSION(10, 1, 4))
> > + return false;
> > +
> > + if (adev->gmc.is_app_apu)
> > + return false;
> > +
> > + return true;
> > +}
> > +
> > int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
> > {
> > struct amdgpu_kfd_dev *kfddev = &adev->kfd;
> > @@ -1000,11 +1016,7 @@ int kgd2kfd_init_zone_device(struct amdgpu_device *adev)
> > unsigned long size;
> > void *r;
> > - /* Page migration works on gfx9 or newer */
> > - if (adev->ip_versions[GC_HWIP][0] < IP_VERSION(9, 0, 1))
> > - return -EINVAL;
> > -
> > - if (adev->gmc.is_app_apu)
> > + if (!is_zone_device_needed(adev))
> > return 0;
> > pgmap = &kfddev->pgmap;
