On Fri, Oct 18, 2024 at 09:27:48AM +0200, Thomas Hellström wrote:
> On Thu, 2024-10-17 at 16:39 -0700, Matthew Brost wrote:
> > Non-contiguous VRAM cannot be mapped in Xe nor can non-visible VRAM
> > easily be accessed. Add ttm_bo_access, which is similar to
> > ttm_bo_vm_access, to access such memory.
> >
> > Visible VRAM access is only supported at the momement but a follow up
> > can add GPU access to non-visible VRAM.
> >
> > Suggested-by: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
> > Signed-off-by: Matthew Brost <matthew.brost@xxxxxxxxx>
> > ---
> > drivers/gpu/drm/ttm/ttm_bo_vm.c | 20 +++++++++-----
> > drivers/gpu/drm/xe/xe_bo.c | 48
> > +++++++++++++++++++++++++++++++++
> > include/drm/ttm/ttm_bo.h | 2 ++
> > 3 files changed, 64 insertions(+), 6 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c
> > b/drivers/gpu/drm/ttm/ttm_bo_vm.c
> > index 2c699ed1963a..b53cc064da44 100644
> > --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c
> > +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c
> > @@ -405,13 +405,9 @@ static int ttm_bo_vm_access_kmap(struct
> > ttm_buffer_object *bo,
> > return len;
> > }
> >
> > -int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
> > - void *buf, int len, int write)
> > +int ttm_bo_access(struct ttm_buffer_object *bo, unsigned long
> > offset,
> > + void *buf, int len, int write)
> > {
> > - struct ttm_buffer_object *bo = vma->vm_private_data;
> > - unsigned long offset = (addr) - vma->vm_start +
> > - ((vma->vm_pgoff - drm_vma_node_start(&bo-
> > >base.vma_node))
> > - << PAGE_SHIFT);
> > int ret;
> >
> > if (len < 1 || (offset + len) > bo->base.size)
> > @@ -439,6 +435,18 @@ int ttm_bo_vm_access(struct vm_area_struct *vma,
> > unsigned long addr,
> >
> > return ret;
> > }
> > +EXPORT_SYMBOL(ttm_bo_access);
> > +
> > +int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr,
> > + void *buf, int len, int write)
> > +{
> > + struct ttm_buffer_object *bo = vma->vm_private_data;
> > + unsigned long offset = (addr) - vma->vm_start +
> > + ((vma->vm_pgoff - drm_vma_node_start(&bo-
> > >base.vma_node))
> > + << PAGE_SHIFT);
> > +
> > + return ttm_bo_access(bo, offset, buf, len, write);
> > +}
> > EXPORT_SYMBOL(ttm_bo_vm_access);
> >
> > static const struct vm_operations_struct ttm_bo_vm_ops = {
> > diff --git a/drivers/gpu/drm/xe/xe_bo.c b/drivers/gpu/drm/xe/xe_bo.c
> > index 5b232f2951b1..267f3b03a6d0 100644
> > --- a/drivers/gpu/drm/xe/xe_bo.c
> > +++ b/drivers/gpu/drm/xe/xe_bo.c
> > @@ -1111,6 +1111,53 @@ static void xe_ttm_bo_swap_notify(struct
> > ttm_buffer_object *ttm_bo)
> > }
> > }
> >
> > +static int xe_ttm_access_memory(struct ttm_buffer_object *ttm_bo,
> > + unsigned long offset, void *buf, int
> > len,
> > + int write)
> > +{
> > + struct xe_bo *bo = ttm_to_xe_bo(ttm_bo);
> > + struct xe_device *xe = ttm_to_xe_device(ttm_bo->bdev);
> > + struct iosys_map vmap;
> > + struct xe_res_cursor cursor;
> > + struct xe_mem_region *vram;
> > + void __iomem *virtual;
> > + int bytes_left = len;
> > +
> > + xe_bo_assert_held(bo);
> > +
> > + if (!mem_type_is_vram(ttm_bo->resource->mem_type))
> > + return -EIO;
> > +
> > + /* FIXME: Use GPU for non-visible VRAM */
> > + if (!(bo->flags & XE_BO_FLAG_NEEDS_CPU_ACCESS))
> > + return -EINVAL;
>
> Is it possible to check whether the bo is actually already in cpu
> visible space even if the flag is missing? Also is there an error code
> convention for the access() callback? Should we return -EIO here as
> well?
>
We should be able to setup of cursor, walk it, and it if all in visible
space proceed with the memcpy. Let me look at that. Eventually we will
need GPU access based on the eudebug feedback but making this better for
now is probably a good idea.
Yea, EIO seems like a better return code too.
> > +
> > + vram = res_to_mem_region(ttm_bo->resource);
> > + xe_res_first(ttm_bo->resource, offset & ~PAGE_MASK, 0,
> > &cursor);
> > +
> > + do {
> > + int wcount = PAGE_SIZE - (offset & PAGE_MASK) >
> > bytes_left ?
> > + bytes_left : PAGE_SIZE - (offset &
> > PAGE_MASK);
> > +
> > + virtual = (u8 __force *)vram->mapping +
> > cursor.start;
>
> Is it possible to use (u8 __iomem *)?
>
Yes, this is copy / paste from somewhere else in the driver.
> > +
> > + iosys_map_set_vaddr_iomem(&vmap, (void __iomem
> > *)virtual);
> And no typecast here (virtual should be __iomem already)
>
Will fix.
> > + if (write)
> > + xe_map_memcpy_to(xe, &vmap, offset &
> > PAGE_MASK, buf,
> > + wcount);
> > + else
> > + xe_map_memcpy_from(xe, buf, &vmap, offset &
> > PAGE_MASK,
> > + wcount);
> > +
> > + offset += wcount;
> > + buf += wcount;
> > + bytes_left -= wcount;
> > + xe_res_next(&cursor, PAGE_SIZE);
> > + } while (bytes_left);
> > +
> > + return len;
> > +}
> > +
> Otherwise LTGM.
> /Thomas
>
Thanks. Will split the TTM part and Xe parts in individual patches, add
kernel doc, and fixup comments in next rev.
Matt
>
> > const struct ttm_device_funcs xe_ttm_funcs = {
> > .ttm_tt_create = xe_ttm_tt_create,
> > .ttm_tt_populate = xe_ttm_tt_populate,
> > @@ -1120,6 +1167,7 @@ const struct ttm_device_funcs xe_ttm_funcs = {
> > .move = xe_bo_move,
> > .io_mem_reserve = xe_ttm_io_mem_reserve,
> > .io_mem_pfn = xe_ttm_io_mem_pfn,
> > + .access_memory = xe_ttm_access_memory,
> > .release_notify = xe_ttm_bo_release_notify,
> > .eviction_valuable = ttm_bo_eviction_valuable,
> > .delete_mem_notify = xe_ttm_bo_delete_mem_notify,
> > diff --git a/include/drm/ttm/ttm_bo.h b/include/drm/ttm/ttm_bo.h
> > index 5804408815be..8ea11cd8df39 100644
> > --- a/include/drm/ttm/ttm_bo.h
> > +++ b/include/drm/ttm/ttm_bo.h
> > @@ -421,6 +421,8 @@ void ttm_bo_unpin(struct ttm_buffer_object *bo);
> > int ttm_bo_evict_first(struct ttm_device *bdev,
> > struct ttm_resource_manager *man,
> > struct ttm_operation_ctx *ctx);
> > +int ttm_bo_access(struct ttm_buffer_object *bo, unsigned long
> > offset,
> > + void *buf, int len, int write);
> > vm_fault_t ttm_bo_vm_reserve(struct ttm_buffer_object *bo,
> > struct vm_fault *vmf);
> > vm_fault_t ttm_bo_vm_fault_reserved(struct vm_fault *vmf,
>
