Maarten Lankhorst
2015-Mar-14 19:30 UTC
[Nouveau] [PATCH ddx] Add support for VRAM-less devices to the ddx
With this patch the DDX almost works with GK20A, the missing piece is adding
COHERENT mappings to the right places. ;-)
If you specify NOUVEAU_BO_APER the kernel will truncate valid_domains to the
domains specified at creation time.
This means that as long as we only specify the correct domain in
nouveau_allocate_surface the effect is still the same.
Signed-off-by: Maarten Lankhorst <dev at mblankhorst.nl>
---
The only thing still missing is adding NOUVEAU_BO_COHERENT to
nouveau_allocate_surface, the scratch buffer, and nouveau_exa_scratch.
After that xorg stays up long enough to crash in
nouveau_exa_share_pixmap_backing, I'm still investigating that. :-)
diff --git a/src/nouveau_dri2.c b/src/nouveau_dri2.c
index e3445b2..69fcd8c 100644
--- a/src/nouveau_dri2.c
+++ b/src/nouveau_dri2.c
@@ -684,7 +684,7 @@ nouveau_dri2_finish_swap(DrawablePtr draw, unsigned int
frame,
/* Reference the back buffer to sync it to vblank */
nouveau_pushbuf_refn(push, &(struct nouveau_pushbuf_refn) {
src_bo,
- NOUVEAU_BO_VRAM | NOUVEAU_BO_RD
+ NOUVEAU_BO_APER | NOUVEAU_BO_RD
}, 1);
if (pNv->Architecture >= NV_FERMI)
@@ -725,7 +725,7 @@ nouveau_dri2_finish_swap(DrawablePtr draw, unsigned int
frame,
* on occluded drawables. */
nouveau_pushbuf_refn(push, &(struct nouveau_pushbuf_refn) {
dst_bo,
- NOUVEAU_BO_VRAM | NOUVEAU_BO_RD
+ NOUVEAU_BO_APER | NOUVEAU_BO_RD
}, 1);
REGION_TRANSLATE(0, ®, -draw->x, -draw->y);
diff --git a/src/nouveau_exa.c b/src/nouveau_exa.c
index def66ac..3a93d02 100644
--- a/src/nouveau_exa.c
+++ b/src/nouveau_exa.c
@@ -263,7 +263,7 @@ nouveau_exa_download_from_screen(PixmapPtr pspix, int x, int
y, int w, int h,
goto memcpy;
if (!NVAccelM2MF(pNv, w, lines, cpp, 0, tmp_offset,
- nouveau_pixmap_bo(pspix), NOUVEAU_BO_VRAM,
+ nouveau_pixmap_bo(pspix), NOUVEAU_BO_APER,
src_pitch, pspix->drawable.height, x, y,
tmp, NOUVEAU_BO_GART, tmp_pitch,
lines, 0, 0))
@@ -361,7 +361,7 @@ nouveau_exa_upload_to_screen(PixmapPtr pdpix, int x, int y,
int w, int h,
if (!NVAccelM2MF(pNv, w, lines, cpp, tmp_offset, 0, tmp,
NOUVEAU_BO_GART, tmp_pitch, lines, 0, 0,
- nouveau_pixmap_bo(pdpix), NOUVEAU_BO_VRAM,
+ nouveau_pixmap_bo(pdpix), NOUVEAU_BO_APER,
dst_pitch, pdpix->drawable.height, x, y))
goto memcpy;
diff --git a/src/nv_accel_common.c b/src/nv_accel_common.c
index 4484c1c..e1a8a12 100644
--- a/src/nv_accel_common.c
+++ b/src/nv_accel_common.c
@@ -37,7 +37,7 @@ nouveau_allocate_surface(ScrnInfoPtr scrn, int width, int
height, int bpp,
Bool tiled = (usage_hint & NOUVEAU_CREATE_PIXMAP_TILED);
Bool shared = FALSE;
union nouveau_bo_config cfg = {};
- int flags = NOUVEAU_BO_MAP | (bpp >= 8 ? NOUVEAU_BO_VRAM : 0);
+ int flags = NOUVEAU_BO_MAP;
int cpp = bpp / 8, ret;
#ifdef NOUVEAU_PIXMAP_SHARING
@@ -46,7 +46,7 @@ nouveau_allocate_surface(ScrnInfoPtr scrn, int width, int
height, int bpp,
flags = NOUVEAU_BO_MAP;
if (bpp >= 8)
- flags |= shared ? NOUVEAU_BO_GART : NOUVEAU_BO_VRAM;
+ flags |= shared ? NOUVEAU_BO_GART : pNv->vram_domain;
if (pNv->Architecture >= NV_TESLA) {
if (scanout) {
@@ -677,7 +677,7 @@ NVAccelCommonInit(ScrnInfoPtr pScrn)
pNv->pushbuf->user_priv = pNv->bufctx;
/* Scratch buffer */
- ret = nouveau_bo_new(pNv->dev, NOUVEAU_BO_VRAM | NOUVEAU_BO_MAP,
+ ret = nouveau_bo_new(pNv->dev, pNv->vram_domain | NOUVEAU_BO_MAP,
128 * 1024, 128 * 1024, NULL, &pNv->scratch);
if (!ret)
ret = nouveau_bo_map(pNv->scratch, 0, pNv->client);
diff --git a/src/nv_driver.c b/src/nv_driver.c
index 8e2ae03..b0e9d11 100644
--- a/src/nv_driver.c
+++ b/src/nv_driver.c
@@ -954,7 +954,10 @@ NVPreInit(ScrnInfoPtr pScrn, int flags)
* The first thing we should figure out is the depth, bpp, etc.
*/
- if (dev->vram_size <= 16 * 1024 * 1024)
+ pNv->vram_domain = NOUVEAU_BO_VRAM;
+ if (!dev->vram_size)
+ pNv->vram_domain = NOUVEAU_BO_GART;
+ else if (dev->vram_size <= 16 * 1024 * 1024)
defaultDepth = 16;
if (!xf86SetDepthBpp(pScrn, defaultDepth, 0, 0, Support32bppFb)) {
NVPreInitFail("\n");
diff --git a/src/nv_type.h b/src/nv_type.h
index e6ab192..c5a2684 100644
--- a/src/nv_type.h
+++ b/src/nv_type.h
@@ -122,6 +122,7 @@ typedef struct _NVRec {
struct nouveau_bo *scratch;
Bool ce_enabled;
+ uint32_t vram_domain;
struct nouveau_object *ce_channel;
struct nouveau_pushbuf *ce_pushbuf;
struct nouveau_object *NvCopy;
diff --git a/src/nvc0_accel.c b/src/nvc0_accel.c
index 848ca87..7188230 100644
--- a/src/nvc0_accel.c
+++ b/src/nvc0_accel.c
@@ -242,7 +242,7 @@ NVAccelInit3D_NVC0(ScrnInfoPtr pScrn)
if (nouveau_pushbuf_space(push, 512, 0, 0) ||
nouveau_pushbuf_refn (push, &(struct nouveau_pushbuf_refn) {
- pNv->scratch, NOUVEAU_BO_VRAM |
+ pNv->scratch, NOUVEAU_BO_APER |
NOUVEAU_BO_WR }, 1))
return FALSE;
diff --git a/src/nvc0_exa.c b/src/nvc0_exa.c
index 1f33353..596361e 100644
--- a/src/nvc0_exa.c
+++ b/src/nvc0_exa.c
@@ -96,11 +96,10 @@ NVC0EXAAcquireSurface2D(PixmapPtr ppix, int is_src, uint32_t
fmt)
{
NVC0EXA_LOCALS(ppix);
struct nouveau_bo *bo = nouveau_pixmap_bo(ppix);
- struct nouveau_pixmap *nvpix = nouveau_pixmap(ppix);
int mthd = is_src ? NV50_2D_SRC_FORMAT : NV50_2D_DST_FORMAT;
uint32_t bo_flags;
- bo_flags = nvpix->shared ? NOUVEAU_BO_GART : NOUVEAU_BO_VRAM;
+ bo_flags = NOUVEAU_BO_APER;
bo_flags |= is_src ? NOUVEAU_BO_RD : NOUVEAU_BO_WR;
if (!nv50_style_tiled_pixmap(ppix)) {
@@ -579,7 +578,7 @@ NVC0EXAPictTexture(NVPtr pNv, PixmapPtr ppix, PicturePtr
ppict, unsigned unit)
if (!nv50_style_tiled_pixmap(ppix))
NOUVEAU_FALLBACK("pixmap is scanout buffer\n");
- PUSH_REFN (push, bo, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
+ PUSH_REFN (push, bo, NOUVEAU_BO_APER | NOUVEAU_BO_RD);
PUSH_DATAu(push, pNv->scratch, TIC_OFFSET + (unit * 32), 8);
switch (ppict->format) {
case PICT_a8r8g8b8:
@@ -884,14 +883,14 @@ NVC0EXAPrepareComposite(int op,
PUSH_DATA (push, 0);
PUSH_RESET(push);
- PUSH_REFN (push, pNv->scratch, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR);
+ PUSH_REFN (push, pNv->scratch, NOUVEAU_BO_APER | NOUVEAU_BO_RDWR);
if (pspict->pDrawable)
PUSH_REFN (push, nouveau_pixmap_bo(pspix),
- NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
- PUSH_REFN (push, dst, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
+ NOUVEAU_BO_APER | NOUVEAU_BO_RD);
+ PUSH_REFN (push, dst, NOUVEAU_BO_APER | NOUVEAU_BO_WR);
if (pmpict && pmpict->pDrawable)
PUSH_REFN (push, nouveau_pixmap_bo(pmpix),
- NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
+ NOUVEAU_BO_APER | NOUVEAU_BO_RD);
nouveau_pushbuf_bufctx(push, pNv->bufctx);
if (nouveau_pushbuf_validate(push)) {
diff --git a/src/nvc0_xv.c b/src/nvc0_xv.c
index d1d8f18..31e2309 100644
--- a/src/nvc0_xv.c
+++ b/src/nvc0_xv.c
@@ -69,9 +69,9 @@ nvc0_xv_image_put(ScrnInfoPtr pScrn,
NVPtr pNv = NVPTR(pScrn);
struct nouveau_bo *dst = nouveau_pixmap_bo(ppix);
struct nouveau_pushbuf_refn refs[] = {
- { pNv->scratch, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR },
- { src, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD },
- { dst, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR },
+ { pNv->scratch, NOUVEAU_BO_APER | NOUVEAU_BO_RDWR },
+ { src, NOUVEAU_BO_APER | NOUVEAU_BO_RD },
+ { dst, NOUVEAU_BO_APER | NOUVEAU_BO_WR },
};
struct nouveau_pushbuf *push = pNv->pushbuf;
uint32_t mode = 0xd0005000 | (src->config.nvc0.tile_mode << 18);
Maarten Lankhorst
2015-Mar-14 21:43 UTC
[Nouveau] [PATCH v2 ddx] Add support for VRAM-less devices to the ddx
With this patch the DDX almost works with GK20A, the missing piece is adding
COHERENT mappings to the right places.
If you specify NOUVEAU_BO_APER the kernel will truncate valid_domains to the
domains specified at creation time.
This means that as long as we only specify the correct domain in
nouveau_allocate_surface the effect is still the same.
Signed-off-by: Maarten Lankhorst <dev at mblankhorst.nl>
---
Changes since v1:
Fix nouveau_exa_create_pixmap to not trigger the <= 32 MB code for
vram_size == 0.
This makes tegra work correctly with set_shared_pixmap.
diff --git a/src/nouveau_dri2.c b/src/nouveau_dri2.c
index e3445b2..69fcd8c 100644
--- a/src/nouveau_dri2.c
+++ b/src/nouveau_dri2.c
@@ -684,7 +684,7 @@ nouveau_dri2_finish_swap(DrawablePtr draw, unsigned int
frame,
/* Reference the back buffer to sync it to vblank */
nouveau_pushbuf_refn(push, &(struct nouveau_pushbuf_refn) {
src_bo,
- NOUVEAU_BO_VRAM | NOUVEAU_BO_RD
+ NOUVEAU_BO_APER | NOUVEAU_BO_RD
}, 1);
if (pNv->Architecture >= NV_FERMI)
@@ -725,7 +725,7 @@ nouveau_dri2_finish_swap(DrawablePtr draw, unsigned int
frame,
* on occluded drawables. */
nouveau_pushbuf_refn(push, &(struct nouveau_pushbuf_refn) {
dst_bo,
- NOUVEAU_BO_VRAM | NOUVEAU_BO_RD
+ NOUVEAU_BO_APER | NOUVEAU_BO_RD
}, 1);
REGION_TRANSLATE(0, ®, -draw->x, -draw->y);
diff --git a/src/nouveau_exa.c b/src/nouveau_exa.c
index def66ac..dbba664 100644
--- a/src/nouveau_exa.c
+++ b/src/nouveau_exa.c
@@ -121,7 +121,7 @@ nouveau_exa_create_pixmap(ScreenPtr pScreen, int width, int
height, int depth,
if (!width || !height)
return calloc(1, sizeof(*nvpix));
- if (!pNv->exa_force_cp && pNv->dev->vram_size <= 32 * 1024
* 1024)
+ if (!pNv->exa_force_cp && pNv->dev->vram_size &&
pNv->dev->vram_size <= 32 * 1024 * 1024)
return NULL;
nvpix = calloc(1, sizeof(*nvpix));
@@ -263,7 +263,7 @@ nouveau_exa_download_from_screen(PixmapPtr pspix, int x, int
y, int w, int h,
goto memcpy;
if (!NVAccelM2MF(pNv, w, lines, cpp, 0, tmp_offset,
- nouveau_pixmap_bo(pspix), NOUVEAU_BO_VRAM,
+ nouveau_pixmap_bo(pspix), NOUVEAU_BO_APER,
src_pitch, pspix->drawable.height, x, y,
tmp, NOUVEAU_BO_GART, tmp_pitch,
lines, 0, 0))
@@ -361,7 +361,7 @@ nouveau_exa_upload_to_screen(PixmapPtr pdpix, int x, int y,
int w, int h,
if (!NVAccelM2MF(pNv, w, lines, cpp, tmp_offset, 0, tmp,
NOUVEAU_BO_GART, tmp_pitch, lines, 0, 0,
- nouveau_pixmap_bo(pdpix), NOUVEAU_BO_VRAM,
+ nouveau_pixmap_bo(pdpix), NOUVEAU_BO_APER,
dst_pitch, pdpix->drawable.height, x, y))
goto memcpy;
diff --git a/src/nv_accel_common.c b/src/nv_accel_common.c
index 4484c1c..e1a8a12 100644
--- a/src/nv_accel_common.c
+++ b/src/nv_accel_common.c
@@ -37,7 +37,7 @@ nouveau_allocate_surface(ScrnInfoPtr scrn, int width, int
height, int bpp,
Bool tiled = (usage_hint & NOUVEAU_CREATE_PIXMAP_TILED);
Bool shared = FALSE;
union nouveau_bo_config cfg = {};
- int flags = NOUVEAU_BO_MAP | (bpp >= 8 ? NOUVEAU_BO_VRAM : 0);
+ int flags = NOUVEAU_BO_MAP;
int cpp = bpp / 8, ret;
#ifdef NOUVEAU_PIXMAP_SHARING
@@ -46,7 +46,7 @@ nouveau_allocate_surface(ScrnInfoPtr scrn, int width, int
height, int bpp,
flags = NOUVEAU_BO_MAP;
if (bpp >= 8)
- flags |= shared ? NOUVEAU_BO_GART : NOUVEAU_BO_VRAM;
+ flags |= shared ? NOUVEAU_BO_GART : pNv->vram_domain;
if (pNv->Architecture >= NV_TESLA) {
if (scanout) {
@@ -677,7 +677,7 @@ NVAccelCommonInit(ScrnInfoPtr pScrn)
pNv->pushbuf->user_priv = pNv->bufctx;
/* Scratch buffer */
- ret = nouveau_bo_new(pNv->dev, NOUVEAU_BO_VRAM | NOUVEAU_BO_MAP,
+ ret = nouveau_bo_new(pNv->dev, pNv->vram_domain | NOUVEAU_BO_MAP,
128 * 1024, 128 * 1024, NULL, &pNv->scratch);
if (!ret)
ret = nouveau_bo_map(pNv->scratch, 0, pNv->client);
diff --git a/src/nv_driver.c b/src/nv_driver.c
index 8e2ae03..e1aacae 100644
--- a/src/nv_driver.c
+++ b/src/nv_driver.c
@@ -954,7 +954,10 @@ NVPreInit(ScrnInfoPtr pScrn, int flags)
* The first thing we should figure out is the depth, bpp, etc.
*/
- if (dev->vram_size <= 16 * 1024 * 1024)
+ pNv->vram_domain = NOUVEAU_BO_VRAM;
+ if (!dev->vram_size)
+ pNv->vram_domain = NOUVEAU_BO_GART;
+ else if (dev->vram_size <= 16 * 1024 * 1024)
defaultDepth = 16;
if (!xf86SetDepthBpp(pScrn, defaultDepth, 0, 0, Support32bppFb)) {
NVPreInitFail("\n");
diff --git a/src/nv_type.h b/src/nv_type.h
index e6ab192..c5a2684 100644
--- a/src/nv_type.h
+++ b/src/nv_type.h
@@ -122,6 +122,7 @@ typedef struct _NVRec {
struct nouveau_bo *scratch;
Bool ce_enabled;
+ uint32_t vram_domain;
struct nouveau_object *ce_channel;
struct nouveau_pushbuf *ce_pushbuf;
struct nouveau_object *NvCopy;
diff --git a/src/nvc0_accel.c b/src/nvc0_accel.c
index 848ca87..7188230 100644
--- a/src/nvc0_accel.c
+++ b/src/nvc0_accel.c
@@ -242,7 +242,7 @@ NVAccelInit3D_NVC0(ScrnInfoPtr pScrn)
if (nouveau_pushbuf_space(push, 512, 0, 0) ||
nouveau_pushbuf_refn (push, &(struct nouveau_pushbuf_refn) {
- pNv->scratch, NOUVEAU_BO_VRAM |
+ pNv->scratch, NOUVEAU_BO_APER |
NOUVEAU_BO_WR }, 1))
return FALSE;
diff --git a/src/nvc0_exa.c b/src/nvc0_exa.c
index 1f33353..596361e 100644
--- a/src/nvc0_exa.c
+++ b/src/nvc0_exa.c
@@ -96,11 +96,10 @@ NVC0EXAAcquireSurface2D(PixmapPtr ppix, int is_src, uint32_t
fmt)
{
NVC0EXA_LOCALS(ppix);
struct nouveau_bo *bo = nouveau_pixmap_bo(ppix);
- struct nouveau_pixmap *nvpix = nouveau_pixmap(ppix);
int mthd = is_src ? NV50_2D_SRC_FORMAT : NV50_2D_DST_FORMAT;
uint32_t bo_flags;
- bo_flags = nvpix->shared ? NOUVEAU_BO_GART : NOUVEAU_BO_VRAM;
+ bo_flags = NOUVEAU_BO_APER;
bo_flags |= is_src ? NOUVEAU_BO_RD : NOUVEAU_BO_WR;
if (!nv50_style_tiled_pixmap(ppix)) {
@@ -579,7 +578,7 @@ NVC0EXAPictTexture(NVPtr pNv, PixmapPtr ppix, PicturePtr
ppict, unsigned unit)
if (!nv50_style_tiled_pixmap(ppix))
NOUVEAU_FALLBACK("pixmap is scanout buffer\n");
- PUSH_REFN (push, bo, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
+ PUSH_REFN (push, bo, NOUVEAU_BO_APER | NOUVEAU_BO_RD);
PUSH_DATAu(push, pNv->scratch, TIC_OFFSET + (unit * 32), 8);
switch (ppict->format) {
case PICT_a8r8g8b8:
@@ -884,14 +883,14 @@ NVC0EXAPrepareComposite(int op,
PUSH_DATA (push, 0);
PUSH_RESET(push);
- PUSH_REFN (push, pNv->scratch, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR);
+ PUSH_REFN (push, pNv->scratch, NOUVEAU_BO_APER | NOUVEAU_BO_RDWR);
if (pspict->pDrawable)
PUSH_REFN (push, nouveau_pixmap_bo(pspix),
- NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
- PUSH_REFN (push, dst, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR);
+ NOUVEAU_BO_APER | NOUVEAU_BO_RD);
+ PUSH_REFN (push, dst, NOUVEAU_BO_APER | NOUVEAU_BO_WR);
if (pmpict && pmpict->pDrawable)
PUSH_REFN (push, nouveau_pixmap_bo(pmpix),
- NOUVEAU_BO_VRAM | NOUVEAU_BO_RD);
+ NOUVEAU_BO_APER | NOUVEAU_BO_RD);
nouveau_pushbuf_bufctx(push, pNv->bufctx);
if (nouveau_pushbuf_validate(push)) {
diff --git a/src/nvc0_xv.c b/src/nvc0_xv.c
index d1d8f18..31e2309 100644
--- a/src/nvc0_xv.c
+++ b/src/nvc0_xv.c
@@ -69,9 +69,9 @@ nvc0_xv_image_put(ScrnInfoPtr pScrn,
NVPtr pNv = NVPTR(pScrn);
struct nouveau_bo *dst = nouveau_pixmap_bo(ppix);
struct nouveau_pushbuf_refn refs[] = {
- { pNv->scratch, NOUVEAU_BO_VRAM | NOUVEAU_BO_RDWR },
- { src, NOUVEAU_BO_VRAM | NOUVEAU_BO_RD },
- { dst, NOUVEAU_BO_VRAM | NOUVEAU_BO_WR },
+ { pNv->scratch, NOUVEAU_BO_APER | NOUVEAU_BO_RDWR },
+ { src, NOUVEAU_BO_APER | NOUVEAU_BO_RD },
+ { dst, NOUVEAU_BO_APER | NOUVEAU_BO_WR },
};
struct nouveau_pushbuf *push = pNv->pushbuf;
uint32_t mode = 0xd0005000 | (src->config.nvc0.tile_mode << 18);