On 2/11/22 11:00, Matthew Auld wrote:
On Fri, 11 Feb 2022 at 09:56, Thomas Hellström
<thomas.hellstrom@xxxxxxxxxxxxxxx> wrote:
On 2/11/22 10:52, Matthew Auld wrote:
On Fri, 11 Feb 2022 at 09:49, Thomas Hellström
<thomas.hellstrom@xxxxxxxxxxxxxxx> wrote:
On 2/10/22 13:13, Matthew Auld wrote:
Starting from DG2+, when dealing with LMEM, we assume that by default
all userspace allocations should be placed in the non-mappable portion
of LMEM. Note that dumb buffers are not included here, since these are
not "GPU accelerated" and likely need CPU access.
In a later patch userspace will be able to provide a hint if CPU access
to the buffer is needed.
Signed-off-by: Matthew Auld <matthew.auld@xxxxxxxxx>
Cc: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
---
drivers/gpu/drm/i915/gem/i915_gem_create.c | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_create.c b/drivers/gpu/drm/i915/gem/i915_gem_create.c
index 9402d4bf4ffc..cc9ddb943f96 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_create.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_create.c
@@ -424,6 +424,15 @@ i915_gem_create_ext_ioctl(struct drm_device *dev, void *data,
ext_data.n_placements = 1;
}
+ /*
+ * TODO: add a userspace hint to force CPU_ACCESS for the object, which
+ * can override this.
+ */
+ if (!IS_DG1(i915) && (ext_data.n_placements > 1 ||
+ ext_data.placements[0]->type !=
+ INTEL_MEMORY_SYSTEM))
+ ext_data.flags |= I915_BO_ALLOC_GPU_ONLY;
+
WRT previous review comment here, it would be easier to follow if the bo
was marked as a GPU only buffer regardless. Then for example capture and
other functions where it actually matters can choose to take action
based on, for example, whether the BAR is restricted or not?
Yeah, I completely forgot about this, sorry. Will fix now.
Actually you did reply, but I forgot to reply to that :).
Hmm, should we just drop the IS_DG1() check here(that was my first
thought), or go further and still apply even regardless of placements?
i.e it would be set on integrated
That was my first thought as well, but yes it makes sense to also drop
the placement checks and let the placement selection logic handle that
later?
One alternative approach would also be to invert the thing and have a
BO_ALLOC_CPU_REQUIRE, that is set by default on some bos and can be set
on the others using the hint, but I figure that needs to be then set
also on kernel-only buffer objects. Not sure what is simplest.
/Thomas
/Thomas
