thr3ads.net - Nouveau - [Nouveau] All-round reclocking improvements [Sep 2015]

If this information is useful, please help other people find it:
Share via:

Roy Spliet

2015-Sep-29 23:23 UTC

[Nouveau] All-round reclocking improvements

In bulletpoints:
- Add some support for G94 and G96 reclocking. Has been tested on literally
  two cards, which is hardly adequate as "full coverage". On the other
hand,
  the changes were small enough to make me confident this might work for others
  as well.
- Fix NV50 wait for VBLANK when no monitor is plugged in.
- Voltage related inprovements for GT21x.
- Slightly improve Keplers handling of DDR3 cards where the DLL should be
  disabled. Lays the foundation for a similar improvement in GDDR5 cards, but
  lacking such a card I can't test any changes here.

If you don't feel confident about patch 9 feel free to drop it. But better:
test this series and report with me on IRC #nouveau when things don't work.
Cheers,

Roy

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 1/9] bios/rammap: Identify DLLoff for >= GF100

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
---
 .../drm/nouveau/include/nvkm/subdev/bios/ramcfg.h  |  1 -
 drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c  |  2 +-
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr5.c     |  3 +-
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgk104.c  | 41 ++++++++++++++++++----
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr3.c     |  4 +--
 5 files changed, 39 insertions(+), 12 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h
b/drivers/gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h
index 3a9abd3..dd48db7 100644
--- a/drivers/gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h
+++ b/drivers/gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h
@@ -78,7 +78,6 @@ struct nvbios_ramcfg {
 			unsigned ramcfg_11_01_04:1;
 			unsigned ramcfg_11_01_08:1;
 			unsigned ramcfg_11_01_10:1;
-			unsigned ramcfg_11_01_20:1;
 			unsigned ramcfg_11_01_40:1;
 			unsigned ramcfg_11_01_80:1;
 			unsigned ramcfg_11_02_03:2;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
index f0e1fc7..3bbb1a7 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
@@ -219,7 +219,7 @@ nvbios_rammapSp(struct nvkm_bios *bios, u32 data,
 		p->ramcfg_11_01_04 = (nvbios_rd08(bios, data + 0x01) & 0x04) >>
2;
 		p->ramcfg_11_01_08 = (nvbios_rd08(bios, data + 0x01) & 0x08) >>
3;
 		p->ramcfg_11_01_10 = (nvbios_rd08(bios, data + 0x01) & 0x10) >>
4;
-		p->ramcfg_11_01_20 = (nvbios_rd08(bios, data + 0x01) & 0x20) >>
5;
+		p->ramcfg_DLLoff =   (nvbios_rd08(bios, data + 0x01) & 0x20) >>
5;
 		p->ramcfg_11_01_40 = (nvbios_rd08(bios, data + 0x01) & 0x40) >>
6;
 		p->ramcfg_11_01_80 = (nvbios_rd08(bios, data + 0x01) & 0x80) >>
7;
 		p->ramcfg_11_02_03 = (nvbios_rd08(bios, data + 0x02) & 0x03) >>
0;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr5.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr5.c
index 24f83b0..2cc074d 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr5.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr5.c
@@ -38,11 +38,12 @@ nvkm_gddr5_calc(struct nvkm_ram *ram, bool nuts)
 	int WL, CL, WR, at[2], dt, ds;
 	int rq = ram->freq < 1000000; /* XXX */
 
+	xd = !ram->next->bios.ramcfg_DLLoff;
+
 	switch (ram->next->bios.ramcfg_ver) {
 	case 0x11:
 		pd =  ram->next->bios.ramcfg_11_01_80;
 		lf =  ram->next->bios.ramcfg_11_01_40;
-		xd = !ram->next->bios.ramcfg_11_01_20;
 		vh =  ram->next->bios.ramcfg_11_02_10;
 		vr =  ram->next->bios.ramcfg_11_02_04;
 		vo =  ram->next->bios.ramcfg_11_06;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgk104.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgk104.c
index 9893556..0d20563 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgk104.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgk104.c
@@ -673,6 +673,25 @@ gk104_ram_calc_gddr5(struct gk104_ram *ram, u32 freq)
  * DDR3
 
******************************************************************************/
 
+static void
+nvkm_sddr3_dll_reset(struct gk104_ramfuc *fuc)
+{
+	ram_nuke(fuc, mr[0]);
+	ram_mask(fuc, mr[0], 0x100, 0x100);
+	ram_mask(fuc, mr[0], 0x100, 0x000);
+}
+
+static void
+nvkm_sddr3_dll_disable(struct gk104_ramfuc *fuc)
+{
+	u32 mr1_old = ram_rd32(fuc, mr[1]);
+
+	if (!(mr1_old & 0x1)) {
+		ram_mask(fuc, mr[1], 0x1, 0x1);
+		ram_nsec(fuc, 1000);
+	}
+}
+
 static int
 gk104_ram_calc_sddr3(struct gk104_ram *ram, u32 freq)
 {
@@ -702,6 +721,10 @@ gk104_ram_calc_sddr3(struct gk104_ram *ram, u32 freq)
 		ram_mask(fuc, 0x10f808, 0x04000000, 0x04000000);
 
 	ram_wr32(fuc, 0x10f314, 0x00000001); /* PRECHARGE */
+
+	if (next->bios.ramcfg_DLLoff)
+		nvkm_sddr3_dll_disable(fuc);
+
 	ram_wr32(fuc, 0x10f210, 0x00000000); /* REFRESH_AUTO = 0 */
 	ram_wr32(fuc, 0x10f310, 0x00000001); /* REFRESH */
 	ram_mask(fuc, 0x10f200, 0x80000000, 0x80000000);
@@ -879,17 +902,20 @@ gk104_ram_calc_sddr3(struct gk104_ram *ram, u32 freq)
 	ram_wr32(fuc, 0x10f210, 0x80000000); /* REFRESH_AUTO = 1 */
 	ram_nsec(fuc, 1000);
 
-	ram_nuke(fuc, mr[0]);
-	ram_mask(fuc, mr[0], 0x100, 0x100);
-	ram_mask(fuc, mr[0], 0x100, 0x000);
+	if (!next->bios.ramcfg_DLLoff) {
+		ram_mask(fuc, mr[1], 0x1, 0x0);
+		nvkm_sddr3_dll_reset(fuc);
+	}
 
-	ram_mask(fuc, mr[2], 0xfff, ram->base.mr[2]);
+	ram_mask(fuc, mr[2], 0x00000fff, ram->base.mr[2]);
+	ram_mask(fuc, mr[1], 0xffffffff, ram->base.mr[1]);
 	ram_wr32(fuc, mr[0], ram->base.mr[0]);
 	ram_nsec(fuc, 1000);
 
-	ram_nuke(fuc, mr[0]);
-	ram_mask(fuc, mr[0], 0x100, 0x100);
-	ram_mask(fuc, mr[0], 0x100, 0x000);
+	if (!next->bios.ramcfg_DLLoff) {
+		nvkm_sddr3_dll_reset(fuc);
+		ram_nsec(fuc, 1000);
+	}
 
 	if (vc == 0 && ram_have(fuc, gpio2E)) {
 		u32 temp  = ram_mask(fuc, gpio2E, 0x3000, fuc->r_func2E[0]);
@@ -1600,6 +1626,7 @@ gk104_ram_new(struct nvkm_fb *fb, struct nvkm_ram **pram)
 		break;
 	case NVKM_RAM_TYPE_DDR3:
 		ram->fuc.r_mr[0] = ramfuc_reg(0x10f300);
+		ram->fuc.r_mr[1] = ramfuc_reg(0x10f304);
 		ram->fuc.r_mr[2] = ramfuc_reg(0x10f320);
 		break;
 	default:
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr3.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr3.c
index b4edc97..2690033 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr3.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr3.c
@@ -70,6 +70,8 @@ nvkm_sddr3_calc(struct nvkm_ram *ram)
 {
 	int CWL, CL, WR, DLL = 0, ODT = 0;
 
+	DLL = !ram->next->bios.ramcfg_DLLoff;
+
 	switch (ram->next->bios.timing_ver) {
 	case 0x10:
 		if (ram->next->bios.timing_hdr < 0x17) {
@@ -79,7 +81,6 @@ nvkm_sddr3_calc(struct nvkm_ram *ram)
 		CWL = ram->next->bios.timing_10_CWL;
 		CL  = ram->next->bios.timing_10_CL;
 		WR  = ram->next->bios.timing_10_WR;
-		DLL = !ram->next->bios.ramcfg_DLLoff;
 		ODT = ram->next->bios.timing_10_ODT;
 		break;
 	case 0x20:
@@ -87,7 +88,6 @@ nvkm_sddr3_calc(struct nvkm_ram *ram)
 		CL  = (ram->next->bios.timing[1] & 0x0000001f) >> 0;
 		WR  = (ram->next->bios.timing[2] & 0x007f0000) >> 16;
 		/* XXX: Get these values from the VBIOS instead */
-		DLL = !(ram->mr[1] & 0x1);
 		ODT =   (ram->mr[1] & 0x004) >> 2 |
 			(ram->mr[1] & 0x040) >> 5 |
 		        (ram->mr[1] & 0x200) >> 7;
-- 
2.4.3

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 2/9] fb/ramgt215: Transform GPIO ramfuc method from FBVREF-specific to generic

In preparation of changing FBVDDQ, as observed on at least one GDDR3 card.
While at it, adhere to func.log[1] properly for consistency.

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
---
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c | 41 ++++++++++-------------
 drivers/gpu/drm/nouveau/nvkm/subdev/gpio/nv50.c   |  2 +-
 2 files changed, 19 insertions(+), 24 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
index 5c08ae8..0c28f38 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
@@ -34,9 +34,6 @@
 #include <subdev/clk/gt215.h>
 #include <subdev/gpio.h>
 
-/* XXX: Remove when memx gains GPIO support */
-extern int nv50_gpio_location(int line, u32 *reg, u32 *shift);
-
 struct gt215_ramfuc {
 	struct ramfuc base;
 	struct ramfuc_reg r_0x001610;
@@ -75,7 +72,7 @@ struct gt215_ramfuc {
 	struct ramfuc_reg r_0x111400;
 	struct ramfuc_reg r_0x611200;
 	struct ramfuc_reg r_mr[4];
-	struct ramfuc_reg r_gpioFBVREF;
+	struct ramfuc_reg r_gpio[4];
 };
 
 struct gt215_ltrain {
@@ -466,24 +463,27 @@ gt215_ram_lock_pll(struct gt215_ramfuc *fuc, struct
gt215_clk_info *mclk)
 }
 
 static void
-gt215_ram_fbvref(struct gt215_ramfuc *fuc, u32 val)
+gt215_ram_gpio(struct gt215_ramfuc *fuc, u8 tag, u32 val)
 {
 	struct nvkm_gpio *gpio = fuc->base.fb->subdev.device->gpio;
 	struct dcb_gpio_func func;
 	u32 reg, sh, gpio_val;
 	int ret;
 
-	if (nvkm_gpio_get(gpio, 0, 0x2e, DCB_GPIO_UNUSED) != val) {
-		ret = nvkm_gpio_find(gpio, 0, 0x2e, DCB_GPIO_UNUSED, &func);
+	if (nvkm_gpio_get(gpio, 0, tag, DCB_GPIO_UNUSED) != val) {
+		ret = nvkm_gpio_find(gpio, 0, tag, DCB_GPIO_UNUSED, &func);
 		if (ret)
 			return;
 
-		nv50_gpio_location(func.line, &reg, &sh);
-		gpio_val = ram_rd32(fuc, gpioFBVREF);
+		reg = func.line >> 3;
+		sh = (func.line & 0x7) << 2;
+		gpio_val = ram_rd32(fuc, gpio[reg]);
 		if (gpio_val & (8 << sh))
 			val = !val;
+		if (!(func.log[1] & 1))
+			val = !val;
 
-		ram_mask(fuc, gpioFBVREF, (0x3 << sh), ((val | 0x2) << sh));
+		ram_mask(fuc, gpio[reg], (0x3 << sh), ((val | 0x2) << sh));
 		ram_nsec(fuc, 20000);
 	}
 }
@@ -642,8 +642,8 @@ gt215_ram_calc(struct nvkm_ram *base, u32 freq)
 		break;
 	}
 
-	if (fuc->r_gpioFBVREF.addr && next->bios.timing_10_ODT)
-		gt215_ram_fbvref(fuc, 0);
+	if (next->bios.timing_10_ODT)
+		gt215_ram_gpio(fuc, 0x2e, 1);
 
 	/* Brace RAM for impact */
 	ram_wr32(fuc, 0x1002d4, 0x00000001);
@@ -809,8 +809,8 @@ gt215_ram_calc(struct nvkm_ram *base, u32 freq)
 	ram_mask(fuc, 0x100718, 0xffffffff, unk718);
 	ram_mask(fuc, 0x111100, 0xffffffff, r111100);
 
-	if (fuc->r_gpioFBVREF.addr && !next->bios.timing_10_ODT)
-		gt215_ram_fbvref(fuc, 1);
+	if (!next->bios.timing_10_ODT)
+		gt215_ram_gpio(fuc, 0x2e, 0);
 
 	/* Reset DLL */
 	if (!next->bios.ramcfg_DLLoff)
@@ -919,10 +919,7 @@ gt215_ram_func = {
 int
 gt215_ram_new(struct nvkm_fb *fb, struct nvkm_ram **pram)
 {
-	struct nvkm_gpio *gpio = fb->subdev.device->gpio;
-	struct dcb_gpio_func func;
 	struct gt215_ram *ram;
-	u32 reg, shift;
 	int ret, i;
 
 	if (!(ram = kzalloc(sizeof(*ram), GFP_KERNEL)))
@@ -981,12 +978,10 @@ gt215_ram_new(struct nvkm_fb *fb, struct nvkm_ram **pram)
 		ram->fuc.r_mr[2] = ramfuc_reg(0x1002e0);
 		ram->fuc.r_mr[3] = ramfuc_reg(0x1002e4);
 	}
-
-	ret = nvkm_gpio_find(gpio, 0, 0x2e, DCB_GPIO_UNUSED, &func);
-	if (ret == 0) {
-		nv50_gpio_location(func.line, &reg, &shift);
-		ram->fuc.r_gpioFBVREF = ramfuc_reg(reg);
-	}
+	ram->fuc.r_gpio[0] = ramfuc_reg(0x00e104);
+	ram->fuc.r_gpio[1] = ramfuc_reg(0x00e108);
+	ram->fuc.r_gpio[2] = ramfuc_reg(0x00e120);
+	ram->fuc.r_gpio[3] = ramfuc_reg(0x00e124);
 
 	return 0;
 }
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/gpio/nv50.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/gpio/nv50.c
index 8996649..73923fd 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/gpio/nv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/gpio/nv50.c
@@ -54,7 +54,7 @@ nv50_gpio_reset(struct nvkm_gpio *gpio, u8 match)
 	}
 }
 
-int
+static int
 nv50_gpio_location(int line, u32 *reg, u32 *shift)
 {
 	const u32 nv50_gpio_reg[4] = { 0xe104, 0xe108, 0xe280, 0xe284 };
-- 
2.4.3

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 3/9] fb/ramgt215: Change FBVDD/Q when BIOS asks for it

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
---
 .../gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h  |  1 +
 drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c      |  1 +
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c      | 18 ++++++++++++++++++
 3 files changed, 20 insertions(+)

diff --git a/drivers/gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h
b/drivers/gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h
index dd48db7..dca6c06 100644
--- a/drivers/gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h
+++ b/drivers/gpu/drm/nouveau/include/nvkm/subdev/bios/ramcfg.h
@@ -39,6 +39,7 @@ struct nvbios_ramcfg {
 	unsigned ramcfg_timing;
 	unsigned ramcfg_DLLoff;
 	unsigned ramcfg_RON;
+	unsigned ramcfg_FBVDDQ;
 	union {
 		struct {
 			unsigned ramcfg_00_03_01:1;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
index 3bbb1a7..74a4ab5 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
@@ -205,6 +205,7 @@ nvbios_rammapSp(struct nvkm_bios *bios, u32 data,
 		p->ramcfg_DLLoff   = (nvbios_rd08(bios, data + 0x02) & 0x40) >>
6;
 		p->ramcfg_10_03_0f = (nvbios_rd08(bios, data + 0x03) & 0x0f) >>
0;
 		p->ramcfg_10_04_01 = (nvbios_rd08(bios, data + 0x04) & 0x01) >>
0;
+		p->ramcfg_FBVDDQ   = (nvbios_rd08(bios, data + 0x04) & 0x08) >>
3;
 		p->ramcfg_10_05    = (nvbios_rd08(bios, data + 0x05) & 0xff) >>
0;
 		p->ramcfg_10_06    = (nvbios_rd08(bios, data + 0x06) & 0xff) >>
0;
 		p->ramcfg_10_07    = (nvbios_rd08(bios, data + 0x07) & 0xff) >>
0;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
index 0c28f38..cb75de1 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
@@ -498,6 +498,7 @@ gt215_ram_calc(struct nvkm_ram *base, u32 freq)
 	struct nvkm_device *device = subdev->device;
 	struct nvkm_bios *bios = device->bios;
 	struct gt215_clk_info mclk;
+	struct nvkm_gpio *gpio = device->gpio;
 	struct nvkm_ram_data *next;
 	u8  ver, hdr, cnt, len, strap;
 	u32 data;
@@ -656,6 +657,23 @@ gt215_ram_calc(struct nvkm_ram *base, u32 freq)
 	if (device->chipset == 0xa3 && freq <= 500000)
 		ram_mask(fuc, 0x100700, 0x00000006, 0x00000006);
 
+	/* Alter FBVDD/Q, apparently must be done with PLL disabled, thus
+	 * set it to bypass */
+	if (nvkm_gpio_get(gpio, 0, 0x18, DCB_GPIO_UNUSED) == 
+			next->bios.ramcfg_FBVDDQ) {
+		data = ram_rd32(fuc, 0x004000) & 0x9;
+
+		if (data == 0x1)
+			ram_mask(fuc, 0x004000, 0x8, 0x8);
+		if (data & 0x1)
+			ram_mask(fuc, 0x004000, 0x1, 0x0);
+
+		gt215_ram_gpio(fuc, 0x18, !next->bios.ramcfg_FBVDDQ);
+
+		if (data & 0x1)
+			ram_mask(fuc, 0x004000, 0x1, 0x1);
+	}
+
 	/* Fiddle with clocks */
 	/* There's 4 scenario's
 	 * pll->pll: first switch to a 324MHz clock, set up new PLL, switch
-- 
2.4.3

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 4/9] fb/ramgt215: Restructure r111100 calculation for DDR2

Seems to be mostly equal to DDR3 on < GT218, should improve stability for
DDR2 reclocks.

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
---
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c | 70 ++++++++++++-----------
 1 file changed, 37 insertions(+), 33 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
index cb75de1..52abcaf 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramgt215.c
@@ -771,41 +771,45 @@ gt215_ram_calc(struct nvkm_ram *base, u32 freq)
 	unk71c  = ram_rd32(fuc, 0x10071c) & ~0x00000100;
 	r111100 = ram_rd32(fuc, 0x111100) & ~0x3a800000;
 
-	if (next->bios.ramcfg_10_02_04) {
-		switch (ram->base.type) {
-		case NVKM_RAM_TYPE_DDR3:
-			if (device->chipset != 0xa8)
-				r111100 |= 0x00000004;
-			/* no break */
-		case NVKM_RAM_TYPE_DDR2:
-			r111100 |= 0x08000000;
-			break;
-		default:
-			break;
-		}
+	/* NVA8 seems to skip various bits related to ramcfg_10_02_04 */
+	if (device->chipset == 0xa8) {
+		r111100 |= 0x08000000;
+		if (!next->bios.ramcfg_10_02_04)
+ 			unk714  |= 0x00000010;
 	} else {
-		switch (ram->base.type) {
-		case NVKM_RAM_TYPE_DDR2:
-			r111100 |= 0x1a800000;
-			unk714  |= 0x00000010;
-			break;
-		case NVKM_RAM_TYPE_DDR3:
-			if (device->chipset == 0xa8) {
-				r111100 |=  0x08000000;
-			} else {
-				r111100 &= ~0x00000004;
-				r111100 |=  0x12800000;
+		if (next->bios.ramcfg_10_02_04) {
+			switch (ram->base.type) {
+			case NVKM_RAM_TYPE_DDR2:
+			case NVKM_RAM_TYPE_DDR3:
+				r111100 &= ~0x00000020;
+				if (next->bios.ramcfg_10_02_10)
+					r111100 |= 0x08000004;
+				else
+					r111100 |= 0x00000024;
+				break;
+			default:
+				break;
 			}
-			unk714  |= 0x00000010;
-			break;
-		case NVKM_RAM_TYPE_GDDR3:
-			r111100 |= 0x30000000;
-			unk714  |= 0x00000020;
-			break;
-		default:
-			break;
-		}
-	}
+		} else {
+			switch (ram->base.type) {
+			case NVKM_RAM_TYPE_DDR2:
+			case NVKM_RAM_TYPE_DDR3:
+				r111100 &= ~0x00000024;
+ 				r111100 |=  0x12800000;
+
+				if (next->bios.ramcfg_10_02_10)
+					r111100 |= 0x08000000;
+				unk714  |= 0x00000010;
+				break;
+			case NVKM_RAM_TYPE_GDDR3:
+				r111100 |= 0x30000000;
+				unk714  |= 0x00000020;
+				break;
+			default:
+				break;
+ 			}
+ 		}
+ 	}
 
 	unk714 |= (next->bios.ramcfg_10_04_01) << 8;
 
-- 
2.4.3

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 5/9] fb/ramnv50: Voltage GPIOs

Does not seem to be necessary for NVA0, hence untested by me.

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow at free.fr>
---
 drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c |  1 +
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c  | 41 +++++++++++++++++++++++
 2 files changed, 42 insertions(+)

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
index 74a4ab5..d0ae745 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/bios/rammap.c
@@ -171,6 +171,7 @@ nvbios_rammapSp_from_perf(struct nvkm_bios *bios, u32 data,
u8 size, int idx,
 	p->ramcfg_DLLoff   = (nvbios_rd08(bios, data + 0x03) & 0x04) >>
2;
 	p->ramcfg_00_03_08 = (nvbios_rd08(bios, data + 0x03) & 0x08) >>
3;
 	p->ramcfg_RON      = (nvbios_rd08(bios, data + 0x03) & 0x10) >>
3;
+	p->ramcfg_FBVDDQ   = (nvbios_rd08(bios, data + 0x03) & 0x80) >>
7;
 	p->ramcfg_00_04_02 = (nvbios_rd08(bios, data + 0x04) & 0x02) >>
1;
 	p->ramcfg_00_04_04 = (nvbios_rd08(bios, data + 0x04) & 0x04) >>
2;
 	p->ramcfg_00_04_20 = (nvbios_rd08(bios, data + 0x04) & 0x20) >>
5;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
index 9197e0e..ae6b0c4 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
@@ -33,6 +33,7 @@
 #include <subdev/bios/rammap.h>
 #include <subdev/bios/timing.h>
 #include <subdev/clk/pll.h>
+#include <subdev/gpio.h>
 
 struct nv50_ramseq {
 	struct hwsq base;
@@ -59,6 +60,7 @@ struct nv50_ramseq {
 	struct hwsq_reg r_0x611200;
 	struct hwsq_reg r_timing[9];
 	struct hwsq_reg r_mr[4];
+	struct hwsq_reg r_gpio[4];
 };
 
 struct nv50_ram {
@@ -154,6 +156,33 @@ nvkm_sddr2_dll_reset(struct nv50_ramseq *hwsq)
 	ram_nsec(hwsq, 24000);
 }
 
+static void
+nv50_ram_gpio(struct nv50_ramseq *hwsq, u8 tag, u32 val)
+{
+	struct nvkm_gpio *gpio = hwsq->base.subdev->device->gpio;
+	struct dcb_gpio_func func;
+	u32 reg, sh, gpio_val;
+	int ret;
+
+	if (nvkm_gpio_get(gpio, 0, tag, DCB_GPIO_UNUSED) != val) {
+		ret = nvkm_gpio_find(gpio, 0, tag, DCB_GPIO_UNUSED, &func);
+		if (ret)
+			return;
+
+		reg = func.line >> 3;
+		sh = (func.line & 0x7) << 2;
+		gpio_val = ram_rd32(hwsq, gpio[reg]);
+
+		if (gpio_val & (8 << sh))
+			val = !val;
+		if (!(func.log[1] & 1))
+			val = !val;
+
+		ram_mask(hwsq, gpio[reg], (0x3 << sh), ((val | 0x2) << sh));
+		ram_nsec(hwsq, 20000);
+	}
+}
+
 static int
 nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 {
@@ -250,6 +279,9 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 	ram_wait(hwsq, 0x00, 0x01); /* wait for fb disabled */
 	ram_nsec(hwsq, 2000);
 
+	if (next->bios.timing_10_ODT)
+		nv50_ram_gpio(hwsq, 0x2e, 1);
+
 	ram_wr32(hwsq, 0x1002d4, 0x00000001); /* precharge */
 	ram_wr32(hwsq, 0x1002d0, 0x00000001); /* refresh */
 	ram_wr32(hwsq, 0x1002d0, 0x00000001); /* refresh */
@@ -288,6 +320,7 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 	ram_mask(hwsq, 0x004008, 0x91ff0000, r004008);
 	if (subdev->device->chipset >= 0x96)
 		ram_wr32(hwsq, 0x100da0, r100da0);
+	nv50_ram_gpio(hwsq, 0x18, !next->bios.ramcfg_FBVDDQ);
 	ram_nsec(hwsq, 64000); /*XXX*/
 	ram_nsec(hwsq, 32000); /*XXX*/
 
@@ -364,6 +397,9 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 	}
 	ram_mask(hwsq, mr[1], 0xffffffff, ram->base.mr[1]);
 
+	if (!next->bios.timing_10_ODT)
+		nv50_ram_gpio(hwsq, 0x2e, 0);
+
 	/* Reset DLL */
 	if (!next->bios.ramcfg_DLLoff)
 		nvkm_sddr2_dll_reset(hwsq);
@@ -634,5 +670,10 @@ nv50_ram_new(struct nvkm_fb *fb, struct nvkm_ram **pram)
 		ram->hwsq.r_mr[3] = hwsq_reg(0x1002e4);
 	}
 
+	ram->hwsq.r_gpio[0] = hwsq_reg(0x00e104);
+	ram->hwsq.r_gpio[1] = hwsq_reg(0x00e108);
+	ram->hwsq.r_gpio[2] = hwsq_reg(0x00e120);
+	ram->hwsq.r_gpio[3] = hwsq_reg(0x00e124);
+
 	return 0;
 }
-- 
2.4.3

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 6/9] fb/ramnv50: Deal with cards without timing entries

Like Pierre's G94. We might want to structure Kepler similarly in a
follow-up.

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow at free.fr>
---
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr3.c   | 10 +++---
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c | 41 ++++++++++++++++++++++--
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr2.c   |  6 ++++
 3 files changed, 50 insertions(+), 7 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr3.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr3.c
index 79b523a..60ece0a 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr3.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/gddr3.c
@@ -63,7 +63,7 @@ ramgddr3_wr_lo[] = {
 	{ 5, 2 }, { 7, 4 }, { 8, 5 }, { 9, 6 }, { 10, 7 },
 	{ 11, 0 }, { 13 , 1 },
 	/* the below are mentioned in some, but not all, gddr3 docs */
-	{ 4, 1 }, { 6, 3 }, { 12, 1 },
+	{ 4, 0 }, { 6, 3 }, { 12, 1 },
 	{ -1 }
 };
 
@@ -87,15 +87,17 @@ nvkm_gddr3_calc(struct nvkm_ram *ram)
 		WR  = (ram->next->bios.timing[2] & 0x007f0000) >> 16;
 		/* XXX: Get these values from the VBIOS instead */
 		DLL = !(ram->mr[1] & 0x1);
-		ODT =  (ram->mr[1] & 0x004) >> 2 |
-		       (ram->mr[1] & 0x040) >> 5 |
-		       (ram->mr[1] & 0x200) >> 7;
 		RON = !(ram->mr[1] & 0x300) >> 8;
 		break;
 	default:
 		return -ENOSYS;
 	}
 
+	if (ram->next->bios.timing_ver == 0x20 ||
+	    ram->next->bios.ramcfg_timing == 0xff) {
+		ODT =  (ram->mr[1] & 0xc) >> 2;
+	}
+
 	hi = ram->mr[2] & 0x1;
 	CL  = ramxlat(hi ? ramgddr3_cl_hi : ramgddr3_cl_lo, CL);
 	WR  = ramxlat(ramgddr3_wr_lo, WR);
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
index ae6b0c4..1c6ae6b 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
@@ -146,6 +146,38 @@ nv50_ram_timing_calc(struct nv50_ram *ram, u32 *timing)
 	nvkm_debug(subdev, " 240: %08x\n", timing[8]);
 	return 0;
 }
+
+static int
+nv50_ram_timing_read(struct nv50_ram *ram, u32 *timing)
+{
+	unsigned int i;
+	struct nvbios_ramcfg *cfg = &ram->base.target.bios;
+	struct nvkm_subdev *subdev = &ram->base.fb->subdev;
+	struct nvkm_device *device = subdev->device;
+
+	for (i = 0; i <= 8; i++)
+		timing[i] = nvkm_rd32(device, 0x100220 + (i * 4));
+
+	/* Derive the bare minimum for the MR calculation to succeed */
+	cfg->timing_ver = 0x10;
+	T(CL) = (timing[3] & 0xff) + 1;
+
+	switch (ram->base.type) {
+	case NVKM_RAM_TYPE_DDR2:
+		T(CWL) = T(CL) - 1;
+		break;
+	case NVKM_RAM_TYPE_GDDR3:
+		T(CWL) = ((timing[2] & 0xff000000) >> 24) + 1;
+		break;
+	default:
+		return -ENOSYS;
+		break;
+	}
+
+	T(WR) = ((timing[1] >> 24) & 0xff) - 1 - T(CWL);
+
+	return 0;
+}
 #undef T
 
 static void
@@ -242,10 +274,11 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 				 strap, data, ver, hdr);
 			return -EINVAL;
 		}
+		nv50_ram_timing_calc(ram, timing);
+	} else {
+		nv50_ram_timing_read(ram, timing);
 	}
 
-	nv50_ram_timing_calc(ram, timing);
-
 	ret = ram_init(hwsq, subdev);
 	if (ret)
 		return ret;
@@ -264,8 +297,10 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 		break;
 	}
 
-	if (ret)
+	if (ret) {
+		nvkm_error(subdev, "Could not calculate MR\n");
 		return ret;
+	}
 
 	/* Always disable this bit during reclock */
 	ram_mask(hwsq, 0x100200, 0x00000800, 0x00000000);
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr2.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr2.c
index 86bf674..b9f1ffd 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr2.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/sddr2.c
@@ -76,6 +76,12 @@ nvkm_sddr2_calc(struct nvkm_ram *ram)
 		return -ENOSYS;
 	}
 
+	if (ram->next->bios.timing_ver == 0x20 ||
+	    ram->next->bios.ramcfg_timing == 0xff) {
+		ODT =  (ram->mr[1] & 0x004) >> 2 |
+		       (ram->mr[1] & 0x040) >> 5;
+	}
+
 	CL  = ramxlat(ramddr2_cl, CL);
 	WR  = ramxlat(ramddr2_wr, WR);
 	if (CL < 0 || WR < 0)
-- 
2.4.3

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 7/9] fb/ramnv50: Script changes for G94 and up

10053c is not even read on some cards, and I have no idea exactly what the
criteria are. Likely NVIDIA pre-scans the VBIOS and in their driver disables
all features that are never used. The practical effect should be the same
as this implementation though.

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow at free.fr>
---
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c | 36 ++++++++++++++++++++----
 1 file changed, 30 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
index 1c6ae6b..651b74a 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
@@ -302,6 +302,9 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 		return ret;
 	}
 
+	if (subdev->device->chipset <= 0x96 &&
!next->bios.ramcfg_00_03_02)
+		ram_mask(hwsq, 0x100710, 0x00000200, 0x00000000);
+
 	/* Always disable this bit during reclock */
 	ram_mask(hwsq, 0x100200, 0x00000800, 0x00000000);
 
@@ -353,8 +356,11 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 			next->bios.rammap_00_16_40 << 14);
 	ram_mask(hwsq, 0x00400c, 0x0000ffff, (N1 << 8) | M1);
 	ram_mask(hwsq, 0x004008, 0x91ff0000, r004008);
-	if (subdev->device->chipset >= 0x96)
+
+	/* XXX: GDDR3 only? */
+	if (subdev->device->chipset >= 0x92)
 		ram_wr32(hwsq, 0x100da0, r100da0);
+
 	nv50_ram_gpio(hwsq, 0x18, !next->bios.ramcfg_FBVDDQ);
 	ram_nsec(hwsq, 64000); /*XXX*/
 	ram_nsec(hwsq, 32000); /*XXX*/
@@ -397,19 +403,33 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 	ram_mask(hwsq, 0x100200, 0x00001000, !next->bios.ramcfg_00_04_02 <<
12);
 
 	/* XXX: A lot of this could be "chipset"/"ram type"
specific stuff */
-	unk710  = ram_rd32(hwsq, 0x100710) & ~0x00000101;
+	unk710  = ram_rd32(hwsq, 0x100710) & ~0x00000100;
 	unk714  = ram_rd32(hwsq, 0x100714) & ~0xf0000020;
 	unk718  = ram_rd32(hwsq, 0x100718) & ~0x00000100;
 	unk71c  = ram_rd32(hwsq, 0x10071c) & ~0x00000100;
+	if (subdev->device->chipset <= 0x96) {
+		unk710 &= ~0x0000006e;
+		unk714 &= ~0x00000100;
+
+		if (!next->bios.ramcfg_00_03_08)
+			unk710 |= 0x00000060;
+		if (!next->bios.ramcfg_FBVDDQ)
+			unk714 |= 0x00000100;
+		if ( next->bios.ramcfg_00_04_04)
+			unk710 |= 0x0000000e;
+	} else {
+		unk710 &= ~0x00000001;
+
+		if (!next->bios.ramcfg_00_03_08)
+			unk710 |= 0x00000001;
+	}
 
 	if ( next->bios.ramcfg_00_03_01)
 		unk71c |= 0x00000100;
 	if ( next->bios.ramcfg_00_03_02)
 		unk710 |= 0x00000100;
-	if (!next->bios.ramcfg_00_03_08) {
-		unk710 |= 0x1;
-		unk714 |= 0x20;
-	}
+	if (!next->bios.ramcfg_00_03_08)
+		unk714 |= 0x00000020;
 	if ( next->bios.ramcfg_00_04_04)
 		unk714 |= 0x70000000;
 	if ( next->bios.ramcfg_00_04_20)
@@ -420,6 +440,8 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 	ram_mask(hwsq, 0x100718, 0xffffffff, unk718);
 	ram_mask(hwsq, 0x100710, 0xffffffff, unk710);
 
+	/* XXX: G94 does not even test these regs in trace. Harmless we do it,
+ 	 * but why is it omitted? */
 	if (next->bios.rammap_00_16_20) {
 		ram_wr32(hwsq, 0x1005a0, next->bios.ramcfg_00_07 << 16 |
 					 next->bios.ramcfg_00_06 << 8 |
@@ -450,6 +472,8 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 		ram_mask(hwsq, 0x004008, 0x00004000, 0x00000000);
 	if (next->bios.ramcfg_00_03_02)
 		ram_mask(hwsq, 0x10021c, 0x00010000, 0x00010000);
+	if (subdev->device->chipset <= 0x96 &&
next->bios.ramcfg_00_03_02)
+		ram_mask(hwsq, 0x100710, 0x00000200, 0x00000200);
 
 	return 0;
 }
-- 
2.4.3

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 8/9] bus/hwsq: Implement VBLANK waiting heuristic

Avoids waiting for VBLANKS that never arrive on headless or otherwise
unconventional set-ups. Strategy taken from MEMX.

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow at free.fr>
---
 drivers/gpu/drm/nouveau/include/nvkm/subdev/bus.h |  1 +
 drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.c    | 32 +++++++++++++++++++++++
 drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.h    |  6 +++++
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c  |  3 +--
 drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramseq.h   |  1 +
 5 files changed, 41 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/include/nvkm/subdev/bus.h
b/drivers/gpu/drm/nouveau/include/nvkm/subdev/bus.h
index 6a04d9c..33a057c 100644
--- a/drivers/gpu/drm/nouveau/include/nvkm/subdev/bus.h
+++ b/drivers/gpu/drm/nouveau/include/nvkm/subdev/bus.h
@@ -14,6 +14,7 @@ int  nvkm_hwsq_fini(struct nvkm_hwsq **, bool exec);
 void nvkm_hwsq_wr32(struct nvkm_hwsq *, u32 addr, u32 data);
 void nvkm_hwsq_setf(struct nvkm_hwsq *, u8 flag, int data);
 void nvkm_hwsq_wait(struct nvkm_hwsq *, u8 flag, u8 data);
+void nvkm_hwsq_wait_vblank(struct nvkm_hwsq *);
 void nvkm_hwsq_nsec(struct nvkm_hwsq *, u32 nsec);
 
 int nv04_bus_new(struct nvkm_device *, int, struct nvkm_bus **);
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.c
index 79f1cf5..2a56689 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.c
@@ -132,6 +132,38 @@ nvkm_hwsq_wait(struct nvkm_hwsq *hwsq, u8 flag, u8 data)
 }
 
 void
+nvkm_hwsq_wait_vblank(struct nvkm_hwsq *hwsq)
+{
+	struct nvkm_subdev *subdev = hwsq->subdev;
+	struct nvkm_device *device = subdev->device;
+	u32 heads, x, y, px = 0;
+	int i, head_sync;
+
+	heads = nvkm_rd32(device, 0x610050);
+	for (i = 0; i < 2; i++) {
+		/* Heuristic: sync to head with biggest resolution */
+		if (heads & (2 << (i << 3))) {
+			x = nvkm_rd32(device, 0x610b40 + (0x540 * i));
+			y = (x & 0xffff0000) >> 16;
+			x &= 0x0000ffff;
+			if ((x * y) > px) {
+				px = (x * y);
+				head_sync = i;
+			}
+		}
+	}
+
+	if (px == 0) {
+		nvkm_debug(subdev, "WAIT VBLANK !NO ACTIVE HEAD\n");
+		return;
+	}
+
+	nvkm_debug(subdev, "WAIT VBLANK HEAD%d\n", head_sync);
+	nvkm_hwsq_wait(hwsq, head_sync ? 0x3 : 0x1, 0x0);
+	nvkm_hwsq_wait(hwsq, head_sync ? 0x3 : 0x1, 0x1);
+}
+
+void
 nvkm_hwsq_nsec(struct nvkm_hwsq *hwsq, u32 nsec)
 {
 	u8 shift = 0, usec = nsec / 1000;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.h
b/drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.h
index 8117ec5..54ec3b1 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/bus/hwsq.h
@@ -134,6 +134,12 @@ hwsq_wait(struct hwsq *ram, u8 flag, u8 data)
 }
 
 static inline void
+hwsq_wait_vblank(struct hwsq *ram)
+{
+	nvkm_hwsq_wait_vblank(ram->hwsq);
+}
+
+static inline void
 hwsq_nsec(struct hwsq *ram, u32 nsec)
 {
 	nvkm_hwsq_nsec(ram->hwsq, nsec);
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
index 651b74a..94b340f 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
@@ -308,8 +308,7 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq)
 	/* Always disable this bit during reclock */
 	ram_mask(hwsq, 0x100200, 0x00000800, 0x00000000);
 
-	ram_wait(hwsq, 0x01, 0x00); /* wait for !vblank */
-	ram_wait(hwsq, 0x01, 0x01); /* wait for vblank */
+	ram_wait_vblank(hwsq);
 	ram_wr32(hwsq, 0x611200, 0x00003300);
 	ram_wr32(hwsq, 0x002504, 0x00000001); /* block fifo */
 	ram_nsec(hwsq, 8000);
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramseq.h
b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramseq.h
index 0f1f97c..8df7306 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramseq.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramseq.h
@@ -11,5 +11,6 @@
 #define ram_mask(s,r,m,d)   hwsq_mask(&(s)->base, &(s)->r_##r,
(m), (d))
 #define ram_setf(s,f,d)     hwsq_setf(&(s)->base, (f), (d))
 #define ram_wait(s,f,d)     hwsq_wait(&(s)->base, (f), (d))
+#define ram_wait_vblank(s)  hwsq_wait_vblank(&(s)->base)
 #define ram_nsec(s,n)       hwsq_nsec(&(s)->base, (n))
 #endif
-- 
2.4.3

Roy Spliet

2015-Sep-29 23:23 UTC

head link

[Nouveau] [PATCH 9/9] clk/g84: Enable reclocking for GDDR3 G94-G200

Your milage may vary, as it's only been tested on a single G94 and one G96.

Signed-off-by: Roy Spliet <rspliet at eclipso.eu>
Tested-by: Pierre Moreau <pierre.morrow at free.fr>
---
 drivers/gpu/drm/nouveau/nvkm/subdev/clk/g84.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/g84.c
b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/g84.c
index 347da9e..f97e3ec 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/g84.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/g84.c
@@ -44,5 +44,5 @@ int
 g84_clk_new(struct nvkm_device *device, int index, struct nvkm_clk **pclk)
 {
 	return nv50_clk_new_(&g84_clk, device, index,
-			     (device->chipset == 0xa0), pclk);
+			     (device->chipset >= 0x94), pclk);
 }
-- 
2.4.3

Ben Skeggs

2015-Sep-30 03:05 UTC

head link

[Nouveau] [PATCH 7/9] fb/ramnv50: Script changes for G94 and up

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256

On 09/30/2015 09:23 AM, Roy Spliet wrote:> 10053c is not even read on some cards, and I have no idea exactly
> what the criteria are. Likely NVIDIA pre-scans the VBIOS and in
> their driver disables all features that are never used. The
> practical effect should be the same as this implementation though.I haven't fully reviewed the series properly yet, but wanted to put my
two cents worth in here.

If this works like it does on Kepler, you're going to get bitten here
if you don't do this correctly.

NVIDIA scans the full memory configurations for every "rammap" entry
(yes, even the ones that can never be referenced by the perf table)
and where all values are the same (doesn't matter if they're all on or
all off, just that they're the same) in every entry, they don't touch
that particular part of the configuration.

Depending on what DEVINIT did, this can leave something either enabled
or disabled, you don't know which, you're just supposed to just not
touch it.

I tried to ignore this on Kepler initially, but eventually got bitten
by it and had to implement it properly (gk104_ram_ctor_data() handles
this for the most part).

I'd *strongly* recommend you verify if the behaviour is the same on
these boards, and implement it if it is.

Ben.
> 
> Signed-off-by: Roy Spliet <rspliet at eclipso.eu> Tested-by: Pierre
> Moreau <pierre.morrow at free.fr> --- 
> drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c | 36
> ++++++++++++++++++++---- 1 file changed, 30 insertions(+), 6
> deletions(-)
> 
> diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c
> b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c index
> 1c6ae6b..651b74a 100644 ---
> a/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c +++
> b/drivers/gpu/drm/nouveau/nvkm/subdev/fb/ramnv50.c @@ -302,6 +302,9
> @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq) return ret; }
> 
> +	if (subdev->device->chipset <= 0x96 &&
> !next->bios.ramcfg_00_03_02) +		ram_mask(hwsq, 0x100710,
> 0x00000200, 0x00000000); + /* Always disable this bit during
> reclock */ ram_mask(hwsq, 0x100200, 0x00000800, 0x00000000);
> 
> @@ -353,8 +356,11 @@ nv50_ram_calc(struct nvkm_ram *base, u32
> freq) next->bios.rammap_00_16_40 << 14); ram_mask(hwsq, 0x00400c,
> 0x0000ffff, (N1 << 8) | M1); ram_mask(hwsq, 0x004008, 0x91ff0000,
> r004008); -	if (subdev->device->chipset >= 0x96) + +	/* XXX: GDDR3
> only? */ +	if (subdev->device->chipset >= 0x92) ram_wr32(hwsq,
> 0x100da0, r100da0); + nv50_ram_gpio(hwsq, 0x18,
> !next->bios.ramcfg_FBVDDQ); ram_nsec(hwsq, 64000); /*XXX*/ 
> ram_nsec(hwsq, 32000); /*XXX*/ @@ -397,19 +403,33 @@
> nv50_ram_calc(struct nvkm_ram *base, u32 freq) ram_mask(hwsq,
> 0x100200, 0x00001000, !next->bios.ramcfg_00_04_02 << 12);
> 
> /* XXX: A lot of this could be "chipset"/"ram type"
specific stuff
> */ -	unk710  = ram_rd32(hwsq, 0x100710) & ~0x00000101; +	unk710  >
ram_rd32(hwsq, 0x100710) & ~0x00000100; unk714  = ram_rd32(hwsq,
> 0x100714) & ~0xf0000020; unk718  = ram_rd32(hwsq, 0x100718) &
> ~0x00000100; unk71c  = ram_rd32(hwsq, 0x10071c) & ~0x00000100; +	if
> (subdev->device->chipset <= 0x96) { +		unk710 &= ~0x0000006e;
+
> unk714 &= ~0x00000100; + +		if (!next->bios.ramcfg_00_03_08) +
> unk710 |= 0x00000060; +		if (!next->bios.ramcfg_FBVDDQ) +			unk714
> |= 0x00000100; +		if ( next->bios.ramcfg_00_04_04) +			unk710 |>
0x0000000e; +	} else { +		unk710 &= ~0x00000001; + +		if
> (!next->bios.ramcfg_00_03_08) +			unk710 |= 0x00000001; +	}
> 
> if ( next->bios.ramcfg_00_03_01) unk71c |= 0x00000100; if (
> next->bios.ramcfg_00_03_02) unk710 |= 0x00000100; -	if
> (!next->bios.ramcfg_00_03_08) { -		unk710 |= 0x1; -		unk714 |> 0x20;
-	} +	if (!next->bios.ramcfg_00_03_08) +		unk714 |> 0x00000020; if (
next->bios.ramcfg_00_04_04) unk714 |= 0x70000000;
> if ( next->bios.ramcfg_00_04_20) @@ -420,6 +440,8 @@
> nv50_ram_calc(struct nvkm_ram *base, u32 freq) ram_mask(hwsq,
> 0x100718, 0xffffffff, unk718); ram_mask(hwsq, 0x100710, 0xffffffff,
> unk710);
> 
> +	/* XXX: G94 does not even test these regs in trace. Harmless we
> do it, + 	 * but why is it omitted? */ if
> (next->bios.rammap_00_16_20) { ram_wr32(hwsq, 0x1005a0,
> next->bios.ramcfg_00_07 << 16 | next->bios.ramcfg_00_06
<< 8 | @@
> -450,6 +472,8 @@ nv50_ram_calc(struct nvkm_ram *base, u32 freq) 
> ram_mask(hwsq, 0x004008, 0x00004000, 0x00000000); if
> (next->bios.ramcfg_00_03_02) ram_mask(hwsq, 0x10021c, 0x00010000,
> 0x00010000); +	if (subdev->device->chipset <= 0x96 &&
> next->bios.ramcfg_00_03_02) +		ram_mask(hwsq, 0x100710, 0x00000200,
> 0x00000200);
> 
> return 0; }
>

Apparently Analagous Threads

Search for more maybe matching threads

Nouveau - Sep 2015 - All-round reclocking improvements

[Nouveau] All-round reclocking improvements

[Nouveau] [PATCH 1/9] bios/rammap: Identify DLLoff for >= GF100

[Nouveau] [PATCH 2/9] fb/ramgt215: Transform GPIO ramfuc method from FBVREF-specific to generic

[Nouveau] [PATCH 3/9] fb/ramgt215: Change FBVDD/Q when BIOS asks for it

[Nouveau] [PATCH 4/9] fb/ramgt215: Restructure r111100 calculation for DDR2

[Nouveau] [PATCH 5/9] fb/ramnv50: Voltage GPIOs

[Nouveau] [PATCH 6/9] fb/ramnv50: Deal with cards without timing entries

[Nouveau] [PATCH 7/9] fb/ramnv50: Script changes for G94 and up

[Nouveau] [PATCH 8/9] bus/hwsq: Implement VBLANK waiting heuristic

[Nouveau] [PATCH 9/9] clk/g84: Enable reclocking for GDDR3 G94-G200

[Nouveau] [PATCH 7/9] fb/ramnv50: Script changes for G94 and up

Apparently Analagous Threads