gem: allow user-space to specify an object should be coherent
Subsystems (2)
PATCH 1/1 gem: allow user-space to specify an object should be coherent
User-space use mappable BOs notably for fences, and expects that a value update by the GPU will be immediatly visible through the user-space mapping.
ARM has a property that may prevent this from happening though: memory can be mapped multiple times only if the different mappings share the same caching properties. However all the lowmem memory is already identity-mapped into the kernel with cache enabled, so when user-space requests an uncached mapping, we actually get an "undefined caching policy" one and this has strange side-effects described on Freedesktop bug 86690.
To prevent this from happening, allow user-space to explicitly specify which objects should be coherent, and create such objects with the TTM_PL_FLAG_UNCACHED flag. This will make TTM allocate memory using the DMA API, which will fix the identify mapping and allow us to safely map the objects to user-space uncached.
---
Patches that take advantage of this in Mesa will follow up shortly. I'd to make sure the new flag is ok first before also adding it to libdrm.
drm/nouveau/include/uapi/drm/nouveau_drm.h | 1 + drm/nouveau/nouveau_gem.c | 3 +++ 2 files changed, 4 insertions(+)
drm/nouveau/include/uapi/drm/nouveau_drm.h +1 -0
@@ -39,6 +39,7 @@#define NOUVEAU_GEM_DOMAIN_VRAM (1 << 1)#define NOUVEAU_GEM_DOMAIN_GART (1 << 2)#define NOUVEAU_GEM_DOMAIN_MAPPABLE (1 << 3) +#define NOUVEAU_GEM_DOMAIN_COHERENT (1 << 4)#define NOUVEAU_GEM_TILE_COMP 0x00030000/* nv50-only */ #define NOUVEAU_GEM_TILE_LAYOUT_MASK 0x0000ff00
drm/nouveau/nouveau_gem.c +3 -0
@@ -189,6 +189,9 @@ nouveau_gem_new(struct drm_device *dev, int size, int align, uint32_t domain,if (!flags || domain &NOUVEAU_GEM_DOMAIN_CPU ) flags |=TTM_PL_FLAG_SYSTEM ; +if (domain &NOUVEAU_GEM_DOMAIN_COHERENT ) + flags |=TTM_PL_FLAG_UNCACHED ; + ret =nouveau_bo_new (dev, size, align, flags, tile_mode, tile_flags,NULL ,NULL , pnvbo);if (ret)
Signature · 6 lines
2.3.0 _______________________________________________ Nouveau mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/nouveau
A message that is not in the archive (sent off-list or not yet received): 1424922292-20688-1-git-send-email-acourbot-DDmLM1+adcrQT0dZR+AlfA@public.gmane.org
PATCH 1/1 instmem/gk20a: use roundup() macro
Use the roundup() macro to make code easier to read and fix a warning when the driver is compiled for 64 bit architectures.
---
Ben, this should probably be squashed into patch 6/6 of my "RAM device removal & IOMMU support" series, since it is not merged yet.
drm/nouveau/nvkm/subdev/instmem/gk20a.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)
drm/nouveau/nvkm/subdev/instmem/gk20a.c +2 -2
@@ -335,8 +335,8 @@ gk20a_instobj_ctor(struct nvkm_object *parent, struct nvkm_object *engine, priv->domain ? "IOMMU" : "DMA", args->size, args->align);/* Round size and align to page bounds */ - size =max ((args->size + ~PAGE_MASK ) &PAGE_MASK , (u32 )PAGE_SIZE ); - align =max ((args->align + ~PAGE_MASK ) &PAGE_MASK , (u32 )PAGE_SIZE ); + size =max (roundup (args->size,PAGE_SIZE ),PAGE_SIZE ); + align =max (roundup (args->align,PAGE_SIZE ),PAGE_SIZE );if (priv->domain) ret = gk20a_instobj_ctor_iommu(parent, engine, oclass,
Signature · 6 lines
2.3.0 _______________________________________________ Nouveau mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/nouveau
Lucas Stach
Reviewed-byAm Donnerstag, den 26.02.2015, 12:44 +0900 schrieb Alexandre Courbot:
Quoted 12 lines
User-space use mappable BOs notably for fences, and expects that a value update by the GPU will be immediatly visible through the user-space mapping.
ARM has a property that may prevent this from happening though: memory can be mapped multiple times only if the different mappings share the same caching properties. However all the lowmem memory is already identity-mapped into the kernel with cache enabled, so when user-space requests an uncached mapping, we actually get an "undefined caching policy" one and this has strange side-effects described on Freedesktop bug 86690.
Quoted 7 lines
To prevent this from happening, allow user-space to explicitly specify which objects should be coherent, and create such objects with the
TTM_PL_FLAG_UNCACHEDflag. This will make TTM allocate memory using the DMA API, which will fix the identify mapping and allow us to safely map the objects to user-space uncached.Signed-off-by Alexandre Courbot <[email protected]>
Ok, this is only needed as userspace is skipping the cpu_prep for the fence BO reads. As doing this would increase the userspace fence overhead a lot, this flag seems to be the right thing to do.
Quoted 34 lines
---
Patches that take advantage of this in Mesa will follow up shortly. I'd to make sure the new flag is ok first before also adding it to libdrm.drm/nouveau/include/uapi/drm/nouveau_drm.h | 1 + drm/nouveau/nouveau_gem.c | 3 +++ 2 files changed, 4 insertions(+)drm/nouveau/include/uapi/drm/nouveau_drm.h +1
-0@@ -39,6 +39,7 @@#define NOUVEAU_GEM_DOMAIN_VRAM (1 << 1)#define NOUVEAU_GEM_DOMAIN_GART (1 << 2)#define NOUVEAU_GEM_DOMAIN_MAPPABLE (1 << 3) +#define NOUVEAU_GEM_DOMAIN_COHERENT (1 << 4)#define NOUVEAU_GEM_TILE_COMP 0x00030000/* nv50-only */ #define NOUVEAU_GEM_TILE_LAYOUT_MASK 0x0000ff00drm/nouveau/nouveau_gem.c +3
-0@@ -189,6 +189,9 @@ nouveau_gem_new(struct drm_device *dev, int size, int align, uint32_t domain,if (!flags || domain &NOUVEAU_GEM_DOMAIN_CPU ) flags |=TTM_PL_FLAG_SYSTEM ; +if (domain &NOUVEAU_GEM_DOMAIN_COHERENT ) + flags |=TTM_PL_FLAG_UNCACHED ; + ret =nouveau_bo_new (dev, size, align, flags, tile_mode, tile_flags,NULL ,NULL , pnvbo);if (ret)