[Intel-gfx] [PATCH 0/1] Ensure zero alignment on gens < 4

[Tvrtko Ursulin]

On 22/11/2021 19:13, Zbigniew Kempczyński wrote:
> In short - we want to enforce alignment == 0 for gen4+ GEM object
> settings.
> 
> Before we merge this we need to inspect all UMD we expect can use
> this. My investigation was narrowed to UMD code:
> 
> 1. IGT
> 2. Mesa
> 3. Media-Driver
> 4. NEO
> 5. libdrm
> 6. xf86-intel-video
> 
> I would like to ask subsystem developers / maintainers to confirm
> my analysis.
> 
> 1. IGT:
>     We've already removed / fixed most of the code where alignment != 0.
>     What left was few multi-card subtests I'm not able to rewrite due
>     to lack of such hw (nv + intel on the board).
> 
> 2. Mesa:
>     gallium/drivers/iris/iris_batch.c,iris_bufmgr.c - it uses softpinning
>     only with alignment handled by allocator, so drm_i915_gem_exec_object2
>     alignment field == 0.
> 
>     drivers/dri/i965/brw_batch.c,brw_screen.c - it uses relocations but
>     it is supported by allocator, there're no direct alignment settings
>     to value != 0.
> 
>     vulcan/anv_batch_chain.c: drm_i915_gem_exec_object2 objects are
>     initialized within anv_execbuf_add_bo() and .alignment field
>     is set to 0 there. There's no other place where I've found vulcan
>     driver touches it both for softpinning / relocations.
> 
> 3. Media-Driver:
>     It contains modified libdrm code and three functions which do
>     allocations, all of them uses mos_gem_bo_alloc_internal():
>     - mos_gem_bo_alloc() - internally uses alignment == 0, that's ok
>     - mos_gem_bo_alloc_tiled() - same as mos_gem_bo_alloc()
>     - mos_gem_bo_alloc_for_render() - this one passes alignment from
>       the caller and it may be != 0. But I haven't found practical
>       usage of this function externally (using mos_bo_alloc_for_render()
>       wrapper).
>     There's another userptr allocation function: mos_bo_alloc_userptr()
>     but it doesn't use alignment.
> 
> 4. NEO:
>     Uses softpinning only with alignment == 0:
>     source/os_interface/linux/drm_buffer_object.cpp:
>     void BufferObject::fillExecObject() has execObject.alignment = 0;
> 
> 5. libdrm:
>     Corresponding functions to Media-Driver:
>     drm_intel_bo_alloc(), drm_intel_bo_alloc_for_render(),
>     drm_intel_bo_alloc_userptr() and drm_intel_bo_alloc_tiled().
>     Alignment field is used in drm_intel_bo_alloc_for_render()
>     so couple not rewritten IGTs may encounter issue here (alignment
>     passed in IGTs which still uses libdrm == 4096).
> 
> 6. xf86-intel-video:
>     src/sna/kgem.c: _kgem_submit() - alignment is set to 0 so this
>     shouldn't be a problem.

You also need to figure out not only what codebase currently uses this, 
but what maybe has an older version in the field which used to, right? 
Otherwise kernel upgrade can break someones old userspace which is not 
allowed. Just raising this for consideration if it isn't already on your 
radar.

Regards,

Tvrtko

> 
> 
> Cc: Petri Latvala <petri.latvala at intel.com>
> Cc: Jason Ekstrand <jason at jlekstrand.net>
> Cc: Dmitry Rogozhkin <dmitry.v.rogozhkin at intel.com>
> Cc: Michal Mrozek <michal.mrozek at intel.com>
> Cc: José Roberto de Souza <jose.souza at intel.com>
> Cc: Chris Wilson <chris at chris-wilson.co.uk>
> Cc: Daniel Vetter <daniel.vetter at ffwll.ch>
> 
> Zbigniew Kempczyński (1):
>    i915/gem/i915_gem_execbuffer: Disallow passing alignment
> 
>   drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c | 11 +++++++++++
>   1 file changed, 11 insertions(+)
>