Hello, On 6/30/26 16:10, Matthew Auld wrote: > On 30/06/2026 14:55, Maarten Lankhorst wrote: >> On systems with media GT, extra latency is added when accessing stolen >> memory when the GT is in MC6. Since we additionally aren't counting how >> much memory is used for stolen and we could in theory fill up the >> entire stolen area with DPT's, avoid using stolen and only use the >> default memory region. >> >> Additionally, this may also result in random system hangs under load. >> >> Link: https://gitlab.freedesktop.org/drm/xe/kernel/-/work_items/7513 >> Signed-off-by: Maarten Lankhorst <[email protected]> > > Fixes: 775d0adc01a5 ("drm/xe/fbdev: Limit the usage of stolen for LNL+") > Cc: <[email protected]> # v6.12+ > > ? > >> --- >> drivers/gpu/drm/xe/display/xe_fb_pin.c | 33 +++++++------------------- >> 1 file changed, 8 insertions(+), 25 deletions(-) >> >> diff --git a/drivers/gpu/drm/xe/display/xe_fb_pin.c >> b/drivers/gpu/drm/xe/display/xe_fb_pin.c >> index 5f4a0cd8deca0..73469ea5f333d 100644 >> --- a/drivers/gpu/drm/xe/display/xe_fb_pin.c >> +++ b/drivers/gpu/drm/xe/display/xe_fb_pin.c >> @@ -164,31 +164,14 @@ static int __xe_pin_fb_vma_dpt(struct drm_gem_object >> *obj, >> dpt_size = ALIGN(intel_rotation_info_size(&view->rotated) * 8, >> XE_PAGE_SIZE); >> - if (IS_DGFX(xe)) >> - dpt = xe_bo_create_pin_map_at_novm(xe, tile0, >> - dpt_size, ~0ull, >> - ttm_bo_type_kernel, >> - XE_BO_FLAG_VRAM0 | >> - XE_BO_FLAG_GGTT | >> - XE_BO_FLAG_PAGETABLE, >> - pin_params->alignment, false); >> - else >> - dpt = xe_bo_create_pin_map_at_novm(xe, tile0, >> - dpt_size, ~0ull, >> - ttm_bo_type_kernel, >> - XE_BO_FLAG_STOLEN | >> - XE_BO_FLAG_GGTT | >> - XE_BO_FLAG_PAGETABLE, >> - pin_params->alignment, false); >> - if (IS_ERR(dpt)) >> - dpt = xe_bo_create_pin_map_at_novm(xe, tile0, >> - dpt_size, ~0ull, >> - ttm_bo_type_kernel, >> - XE_BO_FLAG_SYSTEM | >> - XE_BO_FLAG_GGTT | >> - XE_BO_FLAG_PAGETABLE | >> - XE_BO_FLAG_FORCE_WC, >> - pin_params->alignment, false); > > Technically this changes the behaviour for dgpu, where it looks like it will > fall back to system memory? Surprised that would even work though, and not > get a bunch of pipe errors? I would assume VRAM is a requirement.
I just tried it out to answer the question, turns out this fails spectacularly! It gives a pipe fault error on every vblank on my bmg. It probably never worked, but as long as we never fall back it never was an issue. The new behavior is correct, I'll update the wording of the commit message for this. > Either keeping the same behaviour, or updating the commit to mention this > change, or splitting that out (maybe needs a separate fixes?), > Reviewed-by: Matthew Auld <[email protected]> > Can I extend that to patch 2 and 3 too? >> + dpt = xe_bo_create_pin_map_at_novm(xe, tile0, >> + dpt_size, ~0ull, >> + ttm_bo_type_kernel, >> + XE_BO_FLAG_VRAM_IF_DGFX(tile0) | >> + XE_BO_FLAG_GGTT | >> + XE_BO_FLAG_PAGETABLE | >> + XE_BO_FLAG_FORCE_WC, >> + pin_params->alignment, false); >> if (IS_ERR(dpt)) >> return PTR_ERR(dpt); >> >
