[PATCH] drm/amdgpu: Use device specific BO size & stride check.
Bas Nieuwenhuizen
bas at basnieuwenhuizen.nl
Tue May 4 09:43:34 UTC 2021
The builtin size check isn't really the right thing for AMD
modifiers due to a couple of reasons:
1) In the format structs we don't do set any of the tilesize / blocks
etc. to avoid having format arrays per modifier/GPU
2) The pitch on the main plane is pixel_pitch * bytes_per_pixel even
for tiled ...
3) The pitch for the DCC planes is really the pixel pitch of the main
surface that would be covered by it ...
Note that we only handle GFX9+ case but we do this after converting
the implicit modifier to an explicit modifier, so on GFX9+ all
framebuffers should be checked here.
There is a TODO about DCC alignment, but it isn't worse than before
and I'd need to dig a bunch into the specifics. Getting this out in
a reasonable timeframe to make sure it gets the appropriate testing
seemed more important.
Finally as I've found that debugging addfb2 failures is a pita I was
generous adding explicit error messages to every failure case.
Fixes: f258907fdd83 ("drm/amdgpu: Verify bo size can fit framebuffer size on init.")
Signed-off-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
---
For drm-fixes, replaces the 'Revert "drm/amdgpu: Verify bo size can fit framebuffer size on init."'
revert on the ML, intended for -fixes.
drivers/gpu/drm/amd/amdgpu/amdgpu_display.c | 181 +++++++++++++++++++-
1 file changed, 175 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c
index 2e622c1675d7..80d81774fda8 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_display.c
@@ -837,6 +837,174 @@ static int convert_tiling_flags_to_modifier(struct amdgpu_framebuffer *afb)
return 0;
}
+static void get_block_dimensions(unsigned int block_log2, unsigned int cpp,
+ unsigned int *width, unsigned int *height)
+{
+ unsigned int cpp_log2 = ilog2(cpp);
+ unsigned int pixel_log2 = block_log2 - cpp_log2;
+ unsigned int width_log2 = (pixel_log2 + 1) / 2;
+ unsigned int height_log2 = pixel_log2 - width_log2;
+
+ *width = 1 << width_log2;
+ *height = 1 << height_log2;
+}
+
+static unsigned int get_dcc_block_size(uint64_t modifier, bool rb_aligned,
+ bool pipe_aligned)
+{
+ unsigned int ver = AMD_FMT_MOD_GET(TILE_VERSION, modifier);
+
+ switch (ver) {
+ case AMD_FMT_MOD_TILE_VER_GFX9: {
+ /*
+ * TODO: for pipe aligned we may need to check the alignment of the
+ * total size of the surface, which may need to be bigger than the
+ * natural alignment due to some HW workarounds
+ */
+ return max(10 + (rb_aligned ? (int)AMD_FMT_MOD_GET(RB, modifier) : 0), 12);
+ }
+ case AMD_FMT_MOD_TILE_VER_GFX10:
+ case AMD_FMT_MOD_TILE_VER_GFX10_RBPLUS: {
+ int pipes_log2 = AMD_FMT_MOD_GET(PIPE_XOR_BITS, modifier);
+
+ if (ver == AMD_FMT_MOD_TILE_VER_GFX10_RBPLUS && pipes_log2 > 1 &&
+ AMD_FMT_MOD_GET(PACKERS, modifier) == pipes_log2)
+ ++pipes_log2;
+
+ return max(8 + (pipe_aligned ? pipes_log2 : 0), 12);
+ }
+ default:
+ return 0;
+ }
+}
+
+static int amdgpu_display_verify_plane(struct amdgpu_framebuffer *rfb, int plane,
+ const struct drm_format_info *format,
+ unsigned int block_width, unsigned int block_height,
+ unsigned int block_size_log2)
+{
+ unsigned int width = rfb->base.width /
+ ((plane && plane < format->num_planes) ? format->hsub : 1);
+ unsigned int height = rfb->base.height /
+ ((plane && plane < format->num_planes) ? format->vsub : 1);
+ unsigned int cpp = plane < format->num_planes ? format->cpp[plane] : 1;
+ unsigned int block_pitch = block_width * cpp;
+ unsigned int min_pitch = ALIGN(width * cpp, block_pitch);
+ unsigned int block_size = 1 << block_size_log2;
+ uint64_t size;
+
+ if (rfb->base.pitches[plane] % block_pitch) {
+ drm_dbg_kms(rfb->base.dev,
+ "pitch %d for plane %d is not a multiple of block pitch %d\n",
+ rfb->base.pitches[plane], plane, block_pitch);
+ return -EINVAL;
+ }
+ if (rfb->base.pitches[plane] < min_pitch) {
+ drm_dbg_kms(rfb->base.dev,
+ "pitch %d for plane %d is less than minimum pitch %d\n",
+ rfb->base.pitches[plane], plane, min_pitch);
+ return -EINVAL;
+ }
+
+ /* Force at least natural alignment. */
+ if (rfb->base.offsets[plane] % block_size) {
+ drm_dbg_kms(rfb->base.dev,
+ "offset 0x%x for plane %d is not a multiple of block pitch 0x%x\n",
+ rfb->base.offsets[plane], plane, block_size);
+ return -EINVAL;
+ }
+
+ size = rfb->base.offsets[plane] +
+ (uint64_t)rfb->base.pitches[plane] / block_pitch *
+ block_size * DIV_ROUND_UP(height, block_height);
+
+ if (rfb->base.obj[0]->size < size) {
+ drm_dbg_kms(rfb->base.dev,
+ "BO size 0x%zx is less than 0x%llx required for plane %d\n",
+ rfb->base.obj[0]->size, size, plane);
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
+
+static int amdgpu_display_verify_sizes(struct amdgpu_framebuffer *rfb)
+{
+ const struct drm_format_info *format_info = drm_format_info(rfb->base.format->format);
+ uint64_t modifier = rfb->base.modifier;
+ int ret;
+ unsigned int i, block_width, block_height, block_size_log2;
+
+ if (!rfb->base.dev->mode_config.allow_fb_modifiers)
+ return 0;
+
+ for (i = 0; i < format_info->num_planes; ++i) {
+ if (modifier == DRM_FORMAT_MOD_LINEAR) {
+ block_width = 256 / format_info->cpp[i];
+ block_height = 1;
+ block_size_log2 = 8;
+ } else {
+ int swizzle = AMD_FMT_MOD_GET(TILE, modifier);
+
+ switch ((swizzle & ~3) + 1) {
+ case DC_SW_256B_S:
+ block_size_log2 = 8;
+ break;
+ case DC_SW_4KB_S:
+ case DC_SW_4KB_S_X:
+ block_size_log2 = 12;
+ break;
+ case DC_SW_64KB_S:
+ case DC_SW_64KB_S_T:
+ case DC_SW_64KB_S_X:
+ block_size_log2 = 16;
+ break;
+ default:
+ drm_dbg_kms(rfb->base.dev,
+ "Swizzle mode with unknown block size: %d\n", swizzle);
+ return -EINVAL;
+ }
+
+ get_block_dimensions(block_size_log2, format_info->cpp[i],
+ &block_width, &block_height);
+ }
+
+ ret = amdgpu_display_verify_plane(rfb, i, format_info,
+ block_width, block_height, block_size_log2);
+ if (ret)
+ return ret;
+ }
+
+ if (AMD_FMT_MOD_GET(DCC, modifier)) {
+ if (AMD_FMT_MOD_GET(DCC_RETILE, modifier)) {
+ block_size_log2 = get_dcc_block_size(modifier, false, false);
+ get_block_dimensions(block_size_log2 + 8, format_info->cpp[0],
+ &block_width, &block_height);
+ ret = amdgpu_display_verify_plane(rfb, i, format_info,
+ block_width, block_height,
+ block_size_log2);
+ if (ret)
+ return ret;
+
+ ++i;
+ block_size_log2 = get_dcc_block_size(modifier, true, true);
+ } else {
+ bool pipe_aligned = AMD_FMT_MOD_GET(DCC_PIPE_ALIGN, modifier);
+
+ block_size_log2 = get_dcc_block_size(modifier, true, pipe_aligned);
+ }
+ get_block_dimensions(block_size_log2 + 8, format_info->cpp[0],
+ &block_width, &block_height);
+ ret = amdgpu_display_verify_plane(rfb, i, format_info,
+ block_width, block_height, block_size_log2);
+ if (ret)
+ return ret;
+ }
+
+ return 0;
+}
+
static int amdgpu_display_get_fb_info(const struct amdgpu_framebuffer *amdgpu_fb,
uint64_t *tiling_flags, bool *tmz_surface)
{
@@ -902,10 +1070,8 @@ int amdgpu_display_gem_fb_verify_and_init(
int ret;
rfb->base.obj[0] = obj;
-
- /* Verify that bo size can fit the fb size. */
- ret = drm_gem_fb_init_with_funcs(dev, &rfb->base, file_priv, mode_cmd,
- &amdgpu_fb_funcs);
+ drm_helper_mode_fill_fb_struct(dev, &rfb->base, mode_cmd);
+ ret = drm_framebuffer_init(dev, &rfb->base, &amdgpu_fb_funcs);
if (ret)
goto err;
/* Verify that the modifier is supported. */
@@ -967,9 +1133,12 @@ int amdgpu_display_framebuffer_init(struct drm_device *dev,
}
}
- for (i = 1; i < rfb->base.format->num_planes; ++i) {
+ ret = amdgpu_display_verify_sizes(rfb);
+ if (ret)
+ return ret;
+
+ for (i = 0; i < rfb->base.format->num_planes; ++i) {
drm_gem_object_get(rfb->base.obj[0]);
- drm_gem_object_put(rfb->base.obj[i]);
rfb->base.obj[i] = rfb->base.obj[0];
}
--
2.31.1
More information about the amd-gfx
mailing list