Re: [PATCH v2 2/2] drm/i915/gem: only allow WB for smem only placements

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 28/06/2021 10:38, Thomas Hellström wrote:
Hi,

On 6/28/21 11:12 AM, Matthew Auld wrote:
On 28/06/2021 08:41, Thomas Hellström wrote:

On 6/25/21 2:27 PM, Matthew Auld wrote:
We only support single mode and this should be immutable. For smem only
placements on DGFX this should be WB. On DG1 everything is snooped,
always, and so should be coherent.

I915_GEM_DOMAIN_GTT looks like it's for the aperture which is now gone
for DGFX, so hopefully can also be safely rejected.

Signed-off-by: Matthew Auld <matthew.auld@xxxxxxxxx>
Cc: Thomas Hellström <thomas.hellstrom@xxxxxxxxxxxxxxx>
Cc: Maarten Lankhorst <maarten.lankhorst@xxxxxxxxxxxxxxx>
Cc: Daniel Vetter <daniel.vetter@xxxxxxxx>
---
  drivers/gpu/drm/i915/gem/i915_gem_domain.c |  7 +++++++
  drivers/gpu/drm/i915/gem/i915_gem_mman.c   | 10 ++++++++++
  2 files changed, 17 insertions(+)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_domain.c b/drivers/gpu/drm/i915/gem/i915_gem_domain.c
index d0c91697bb22..e3459a524e64 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_domain.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_domain.c
@@ -577,6 +577,13 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data,
          goto out_unpin;
      }
+    if (IS_DGFX(to_i915(obj->base.dev)) && obj->mm.n_placements == 1 && +        i915_gem_object_placements_contain_type(obj, INTEL_MEMORY_SYSTEM) &&
+        read_domains != I915_GEM_DOMAIN_CPU) {
+        err = -EINVAL;
+        goto out_unpin;
+    }
+
      if (read_domains & I915_GEM_DOMAIN_WC)
          err = i915_gem_object_set_to_wc_domain(obj, write_domain);
      else if (read_domains & I915_GEM_DOMAIN_GTT)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_mman.c b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
index f3586b36dd53..afc9f3dc38b9 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_mman.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_mman.c
@@ -673,6 +673,7 @@ __assign_mmap_offset(struct drm_i915_gem_object *obj,
               enum i915_mmap_type mmap_type,
               u64 *offset, struct drm_file *file)
  {
+    struct drm_i915_private *i915 = to_i915(obj->base.dev);
      struct i915_mmap_offset *mmo;
      if (i915_gem_object_never_mmap(obj))
@@ -697,6 +698,15 @@ __assign_mmap_offset(struct drm_i915_gem_object *obj,           i915_gem_object_placements_contain_type(obj, INTEL_MEMORY_LOCAL))
          return -ENODEV;
+    /*
+     * For smem only placements on DGFX we need to default to WB. On DG1
+     * everything is snooped always, so should always be coherent.
+     */
+     if (IS_DGFX(i915) &&
+         mmap_type != I915_MMAP_TYPE_WB && obj->mm.n_placements == 1 && +         i915_gem_object_placements_contain_type(obj, INTEL_MEMORY_SYSTEM))
+        return -ENODEV;
+

Same thing here as in the previous patch.

Also do we need to modify i915_coherent_map_type() to also include HAS_SNOOP()?

While we're at it, that "always_coherent" argument to i915_coherent_map_type() appears scary to me and probably needs some documentation. It seems used for page-tables. Is it because we know those are always snooped?

Yeah, it's either because the caller has/will mark the pages as coherent(which translates to some special ppGTT bits), or we manually flush ourselves. In i915_coherent_map_type() we should account for DG1 somehow.

Historically I don't think we enabled snooping by default since it's considered slow compared to shared LLC. On DG1 this is a different story though.

Also the pin_map() interface is pretty much only for kernel internal objects, so I don't think we have any users which try to map userspace objects with that interface. Ok, except for vm_access it seems, but that should hopefully be a simple fix to use the correct caching mode? We can maybe add some sanity checking there if someone tries to map a userspace object?
I'm not fully sure that's sufficient, see below.

For all the other callers of pin_map() which should all be kernel internal do we still need to force WB for system memory? By design we only support a single mm.mapping there. For lmem we already use WC only.

We're only allowed to map with the same caching mode as the linear kernel mapping for discrete. Otherwise things may blow up on non-intel architectures. We can probably update 195_ttm_select_tt_caching to always use WB for system pages for kernel objects, but then we must make sure we don't try to map these WC.

Ok, do you think that should be a separate series? It looks like our internal objects don't use ttm(?). Should it? If so should we make a region for it, or can we just make create_internal use the ttm system region? It should be pretty much the same, except we don't want swapping, clearing or eviction, and ideally we would have some way of marking the pages as volatile(I think we can just keep IS_SHRINKABLE for that).

Or can we keep create_internal as is and then it's just a case of dealing with all the pin_map() callers?


/Thomas




/Thomas


      mmo = mmap_offset_attach(obj, mmap_type, file);
      if (IS_ERR(mmo))
          return PTR_ERR(mmo);



[Index of Archives]     [Linux DRI Users]     [Linux Intel Graphics]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [XFree86]     [Linux USB Devel]     [Video for Linux]     [Linux Audio Users]     [Linux Kernel]     [Linux SCSI]     [XFree86]
  Powered by Linux