On Thu, Nov 12, 2026 at 01:58:22PM +0100, Christian König wrote:
> On 11/12/26 13:20, Matthew Auld wrote:
> > Every GEM/TTM buffer object, including imported dma-bufs, allocates a
> > fake mmap offset from the per-device drm_vma_offset_manager. On multi-GPU
> > systems with large VRAM, peer-to-peer dma-buf imports across devices and
> > processes can exhaust the current 1 TiB 64-bit range and fail BO creation
> > or PRIME import with -ENOSPC. Skipping mmap offset allocation for
> > imported sg buffers is not straightforward without risking userspace
> > regressions, as existing userspace (such as Intel compute-runtime) still
> > queries mmap_offset on imported host dma-bufs even when mapping the
> > dma-buf fd directly.
> 
> I was just about to ask why the heck we grab mmap offsets for imported 
> DMA-bufs :)
> 
> > Increase DRM_FILE_PAGE_OFFSET_SIZE on 64-bit kernels so the fake offset
> > range spans up to 8 TiB (SZ_8T). This stays strictly below panthor's
> > DRM_PANTHOR_USER_MMIO_OFFSET_32BIT (1ULL << 43) carveout and within
> > 32-bit compat mmap2's pgoff limit (16 TiB with 4K pages).
> > 
> > Signed-off-by: Matthew Auld <[email protected]>
> > Cc: Thomas Hellström <[email protected]>
> > Cc: Christian König <[email protected]>
> > Cc: Matthew Brost <[email protected]>

Reviewed-by: Matthew Brost <[email protected]>

> 
> Acked-by: Christian König <[email protected]>
> 
> > ---
> >  include/drm/drm_vma_manager.h | 13 ++++++++++---
> >  1 file changed, 10 insertions(+), 3 deletions(-)
> > 
> > diff --git a/include/drm/drm_vma_manager.h b/include/drm/drm_vma_manager.h
> > index 6c2a2f21dbf0..2c3233f9dec4 100644
> > --- a/include/drm/drm_vma_manager.h
> > +++ b/include/drm/drm_vma_manager.h
> > @@ -26,16 +26,23 @@
> >  #include <drm/drm_mm.h>
> >  #include <linux/mm.h>
> >  #include <linux/rbtree.h>
> > +#include <linux/sizes.h>
> >  #include <linux/spinlock.h>
> >  #include <linux/types.h>
> >  
> > -/* We make up offsets for buffer objects so we can recognize them at
> > +/*
> > + * We make up offsets for buffer objects so we can recognize them at
> >   * mmap time. pgoff in mmap is an unsigned long, so we need to make sure
> > - * that the faked up offset will fit
> > + * that the faked up offset will fit.
> > + *
> > + * On 64-bit, cap the range at 8 TiB (SZ_8T) so offsets remain below both
> > + * panthor's DRM_PANTHOR_USER_MMIO_OFFSET_32BIT (1ULL << 43) and 32-bit
> > + * compat mmap2's pgoff ceiling (16 TiB with 4K pages).
> >   */
> >  #if BITS_PER_LONG == 64
> >  #define DRM_FILE_PAGE_OFFSET_START ((0xFFFFFFFFUL >> PAGE_SHIFT) + 1)
> > -#define DRM_FILE_PAGE_OFFSET_SIZE ((0xFFFFFFFFUL >> PAGE_SHIFT) * 256)
> > +#define DRM_FILE_PAGE_OFFSET_SIZE ((SZ_8T >> PAGE_SHIFT) - \
> > +                              DRM_FILE_PAGE_OFFSET_START)
> >  #else
> >  #define DRM_FILE_PAGE_OFFSET_START ((0xFFFFFFFUL >> PAGE_SHIFT) + 1)
> >  #define DRM_FILE_PAGE_OFFSET_SIZE ((0xFFFFFFFUL >> PAGE_SHIFT) * 16)
> 

Reply via email to