xref: /linux/Documentation/gpu/rfc/i915_small_bar.rst (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
1==========================
2I915 Small BAR RFC Section
3==========================
4Starting from DG2 we will have resizable BAR support for device local-memory(i.e
5I915_MEMORY_CLASS_DEVICE), but in some cases the final BAR size might still be
6smaller than the total probed_size. In such cases, only some subset of
7I915_MEMORY_CLASS_DEVICE will be CPU accessible(for example the first 256M),
8while the remainder is only accessible via the GPU.
9
10I915_GEM_CREATE_EXT_FLAG_NEEDS_CPU_ACCESS flag
11----------------------------------------------
12New gem_create_ext flag to tell the kernel that a BO will require CPU access.
13This becomes important when placing an object in I915_MEMORY_CLASS_DEVICE, where
14underneath the device has a small BAR, meaning only some portion of it is CPU
15accessible. Without this flag the kernel will assume that CPU access is not
16required, and prioritize using the non-CPU visible portion of
17I915_MEMORY_CLASS_DEVICE.
18
19.. kernel-doc:: Documentation/gpu/rfc/i915_small_bar.h
20   :functions: __drm_i915_gem_create_ext
21
22probed_cpu_visible_size attribute
23---------------------------------
24New struct__drm_i915_memory_region attribute which returns the total size of the
25CPU accessible portion, for the particular region. This should only be
26applicable for I915_MEMORY_CLASS_DEVICE. We also report the
27unallocated_cpu_visible_size, alongside the unallocated_size.
28
29Vulkan will need this as part of creating a separate VkMemoryHeap with the
30VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT set, to represent the CPU visible portion,
31where the total size of the heap needs to be known. It also wants to be able to
32give a rough estimate of how memory can potentially be allocated.
33
34.. kernel-doc:: Documentation/gpu/rfc/i915_small_bar.h
35   :functions: __drm_i915_memory_region_info
36
37Error Capture restrictions
38--------------------------
39With error capture we have two new restrictions:
40
41    1) Error capture is best effort on small BAR systems; if the pages are not
42    CPU accessible, at the time of capture, then the kernel is free to skip
43    trying to capture them.
44
45    2) On discrete and newer integrated platforms we now reject error capture
46    on recoverable contexts. In the future the kernel may want to blit during
47    error capture, when for example something is not currently CPU accessible.
48