On Tue, Dec 7, 2021 at 4:59 PM Felix Kuehling <felix.kuehling@xxxxxxx> wrote:
>
> On 2021-12-07 9:59 a.m., Philip Yang wrote:
> > If host and amdgpu IOMMU is not enabled or IOMMU is pass through mode,
> > dma_map_page return address is equal to page physical address, use this
> > to set adev->iommu_isolation flag which will be used to optimize memory
> > usage for multi GPU mappings.
> >
> > Signed-off-by: Philip Yang <Philip.Yang@xxxxxxx>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu.h | 2 ++
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 27 ++++++++++++++++++++++
> > 2 files changed, 29 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> > index c5cfe2926ca1..fbbe8c7b5d0c 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
> > @@ -1097,6 +1097,8 @@ struct amdgpu_device {
> >
> > struct amdgpu_reset_control *reset_cntl;
> > uint32_t ip_versions[MAX_HWIP][HWIP_MAX_INSTANCE];
> > +
> > + bool iommu_isolation;
> > };
> >
> > static inline struct amdgpu_device *drm_to_adev(struct drm_device *ddev)
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > index 3c5afa45173c..6d0f3c477670 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > @@ -3364,6 +3364,31 @@ static int amdgpu_device_get_job_timeout_settings(struct amdgpu_device *adev)
> > return ret;
> > }
> >
> > +/**
> > + * amdgpu_device_check_iommu_isolation - check if IOMMU isolation is enabled
> > + *
> > + * @adev: amdgpu_device pointer
> > + *
> > + * device is in IOMMU isolation mode if dma_map_page return address is not equal
> > + * to page physical address.
> > + */
> > +static void amdgpu_device_check_iommu_isolation(struct amdgpu_device *adev)
> > +{
> > + struct page *page;
> > + dma_addr_t addr;
> > +
> > + page = alloc_page(GFP_KERNEL);
> > + if (!page)
> > + return;
> > + addr = dma_map_page(adev->dev, page, 0, PAGE_SIZE, DMA_BIDIRECTIONAL);
> > + if (dma_mapping_error(adev->dev, addr))
> > + goto out_free_page;
> > + adev->iommu_isolation = (addr != page_to_phys(page));
> > + dma_unmap_page(adev->dev, addr, PAGE_SIZE, DMA_BIDIRECTIONAL);
>
> This is a bit of a hack. Unfortunately it seems there isn't a much
> better way to do this. I guess you could copy the implementation of
> dma_map_direct in kernel/dma/mapping.c, but that's more brittle.
>
> I think this hack only tells you whether system memory is direct-mapped.
> The answer may be different for peer VRAM (which isn't supported
> upstream yet, but it's coming). I think this can happen when the IOMMU
> is in pass-through mode by default but still used to DMA map physical
> addresses that are outside the dma mask of the GPU. So a more future
> proof way would be to store a direct-mapped flag for each GPU-GPU and
> GPU-System pair somehow. For the GPU->GPU direct mapping flag you'd need
> to try to DMA-map a page from one GPU's VRAM to the other device.
> Anyway, that can be done in a later change.
>
> For now I'd just change the name of the flag from iommu_isolation to
> direct_map_ram or ram_is_direct_mapped or similar to be more specific
> about what it means.
This thread might be useful:
https://lists.freedesktop.org/archives/dri-devel/2021-November/330557.html
https://lists.freedesktop.org/archives/dri-devel/2021-November/330583.html
Alex
>
> Regards,
> Felix
>
>
> > +out_free_page:
> > + __free_page(page);
> > +}
> > +
> > static const struct attribute *amdgpu_dev_attributes[] = {
> > &dev_attr_product_name.attr,
> > &dev_attr_product_number.attr,
> > @@ -3767,6 +3792,8 @@ int amdgpu_device_init(struct amdgpu_device *adev,
> > queue_delayed_work(system_wq, &mgpu_info.delayed_reset_work,
> > msecs_to_jiffies(AMDGPU_RESUME_MS));
> >
> > + amdgpu_device_check_iommu_isolation(adev);
> > +
> > return 0;
> >
> > release_ras_con:
