Am 15.06.2018 um 16:27 schrieb Michel Dänzer: > From: Michel Dänzer <michel.daenzer at amd.com> > > It can be quite big, and there's no need for it to be physically > contiguous. This is less likely to fail under memory pressure (has > actually happened while running piglit). Mhm, the last time I looked a drm_mm node had around 160 bytes, which means that you need to allocate more than 50MB of VRAM to fill up a 4K page. If I'm not completely mistaken you can still nearly fit a 4K 8bit RGB texture into that. Christian. > > Cc: stable at vger.kernel.org > Signed-off-by: Michel Dänzer <michel.daenzer at amd.com> > --- > drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c | 7 ++++--- > 1 file changed, 4 insertions(+), 3 deletions(-) > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c > index 9aca653bec07..9c47e860e5e6 100644 > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c > @@ -135,7 +135,8 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, > num_nodes = DIV_ROUND_UP(mem->num_pages, pages_per_node); > } > > - nodes = kcalloc(num_nodes, sizeof(*nodes), GFP_KERNEL); > + nodes = kvmalloc_array(num_nodes, sizeof(*nodes), > + GFP_KERNEL | __GFP_ZERO); > if (!nodes) > return -ENOMEM; > > @@ -190,7 +191,7 @@ static int amdgpu_vram_mgr_new(struct ttm_mem_type_manager *man, > drm_mm_remove_node(&nodes[i]); > spin_unlock(&mgr->lock); > > - kfree(nodes); > + kvfree(nodes); > return r == -ENOSPC ? 0 : r; > } > > @@ -229,7 +230,7 @@ static void amdgpu_vram_mgr_del(struct ttm_mem_type_manager *man, > atomic64_sub(usage, &mgr->usage); > atomic64_sub(vis_usage, &mgr->vis_usage); > > - kfree(mem->mm_node); > + kvfree(mem->mm_node); > mem->mm_node = NULL; > } >
