5 files changed, 1188 insertions, 234 deletions

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gf100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gf100.c
index 78c449b41..89d554311 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gf100.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gf100.c
@@ -99,7 +99,7 @@ read_div(struct gf100_clk *clk, int doff, u32 dsrc, u32 dctl)
 {
 	struct nvkm_device *device = clk->base.subdev.device;
 	u32 ssrc = nvkm_rd32(device, dsrc + (doff * 4));
-	u32 sctl = nvkm_rd32(device, dctl + (doff * 4));
+	u32 sclk, sctl, sdiv = 2;
 
 	switch (ssrc & 0x00000003) {
 	case 0:
@@ -109,13 +109,21 @@ read_div(struct gf100_clk *clk, int doff, u32 dsrc, u32 dctl)
 	case 2:
 		return 100000;
 	case 3:
-		if (sctl & 0x80000000) {
-			u32 sclk = read_vco(clk, dsrc + (doff * 4));
-			u32 sdiv = (sctl & 0x0000003f) + 2;
-			return (sclk * 2) / sdiv;
+		sclk = read_vco(clk, dsrc + (doff * 4));
+
+		/* Memclk has doff of 0 despite its alt. location */
+		if (doff <= 2) {
+			sctl = nvkm_rd32(device, dctl + (doff * 4));
+
+			if (sctl & 0x80000000) {
+				if (ssrc & 0x100)
+					sctl >>= 8;
+
+				sdiv = (sctl & 0x3f) + 2;
+			}
 		}
 
-		return read_vco(clk, dsrc + (doff * 4));
+		return (sclk * 2) / sdiv;
 	default:
 		return 0;
 	}
@@ -366,11 +374,17 @@ gf100_clk_prog_2(struct gf100_clk *clk, int idx)
 		if (info->coef) {
 			nvkm_wr32(device, addr + 0x04, info->coef);
 			nvkm_mask(device, addr + 0x00, 0x00000001, 0x00000001);
+
+			/* Test PLL lock */
+			nvkm_mask(device, addr + 0x00, 0x00000010, 0x00000000);
 			nvkm_msec(device, 2000,
 				if (nvkm_rd32(device, addr + 0x00) & 0x00020000)
 					break;
 			);
-			nvkm_mask(device, addr + 0x00, 0x00020004, 0x00000004);
+			nvkm_mask(device, addr + 0x00, 0x00000010, 0x00000010);
+
+			/* Enable sync mode */
+			nvkm_mask(device, addr + 0x00, 0x00000004, 0x00000004);
 		}
 	}
 }
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk104.c b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk104.c
index 975c401bc..06bc0d2d6 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk104.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk104.c
@@ -393,11 +393,17 @@ gk104_clk_prog_2(struct gk104_clk *clk, int idx)
 	if (info->coef) {
 		nvkm_wr32(device, addr + 0x04, info->coef);
 		nvkm_mask(device, addr + 0x00, 0x00000001, 0x00000001);
+
+		/* Test PLL lock */
+		nvkm_mask(device, addr + 0x00, 0x00000010, 0x00000000);
 		nvkm_msec(device, 2000,
 			if (nvkm_rd32(device, addr + 0x00) & 0x00020000)
 				break;
 		);
-		nvkm_mask(device, addr + 0x00, 0x00020004, 0x00000004);
+		nvkm_mask(device, addr + 0x00, 0x00000010, 0x00000010);
+
+		/* Enable sync mode */
+		nvkm_mask(device, addr + 0x00, 0x00000004, 0x00000004);
 	}
 }
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.c b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.c
index 5f0ee24e3..218893e3e 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.c
@@ -28,69 +28,6 @@
 #include <core/tegra.h>
 #include <subdev/timer.h>
 
-#define KHZ (1000)
-#define MHZ (KHZ * 1000)
-
-#define MASK(w)	((1 << w) - 1)
-
-#define GPCPLL_CFG		(SYS_GPCPLL_CFG_BASE + 0)
-#define GPCPLL_CFG_ENABLE	BIT(0)
-#define GPCPLL_CFG_IDDQ		BIT(1)
-#define GPCPLL_CFG_LOCK_DET_OFF	BIT(4)
-#define GPCPLL_CFG_LOCK		BIT(17)
-
-#define GPCPLL_COEFF		(SYS_GPCPLL_CFG_BASE + 4)
-#define GPCPLL_COEFF_M_SHIFT	0
-#define GPCPLL_COEFF_M_WIDTH	8
-#define GPCPLL_COEFF_N_SHIFT	8
-#define GPCPLL_COEFF_N_WIDTH	8
-#define GPCPLL_COEFF_P_SHIFT	16
-#define GPCPLL_COEFF_P_WIDTH	6
-
-#define GPCPLL_CFG2			(SYS_GPCPLL_CFG_BASE + 0xc)
-#define GPCPLL_CFG2_SETUP2_SHIFT	16
-#define GPCPLL_CFG2_PLL_STEPA_SHIFT	24
-
-#define GPCPLL_CFG3			(SYS_GPCPLL_CFG_BASE + 0x18)
-#define GPCPLL_CFG3_PLL_STEPB_SHIFT	16
-
-#define GPC_BCASE_GPCPLL_CFG_BASE		0x00132800
-#define GPCPLL_NDIV_SLOWDOWN			(SYS_GPCPLL_CFG_BASE + 0x1c)
-#define GPCPLL_NDIV_SLOWDOWN_NDIV_LO_SHIFT	0
-#define GPCPLL_NDIV_SLOWDOWN_NDIV_MID_SHIFT	8
-#define GPCPLL_NDIV_SLOWDOWN_STEP_SIZE_LO2MID_SHIFT	16
-#define GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT	22
-#define GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT	31
-
-#define SEL_VCO				(SYS_GPCPLL_CFG_BASE + 0x100)
-#define SEL_VCO_GPC2CLK_OUT_SHIFT	0
-
-#define GPC2CLK_OUT			(SYS_GPCPLL_CFG_BASE + 0x250)
-#define GPC2CLK_OUT_SDIV14_INDIV4_WIDTH	1
-#define GPC2CLK_OUT_SDIV14_INDIV4_SHIFT	31
-#define GPC2CLK_OUT_SDIV14_INDIV4_MODE	1
-#define GPC2CLK_OUT_VCODIV_WIDTH	6
-#define GPC2CLK_OUT_VCODIV_SHIFT	8
-#define GPC2CLK_OUT_VCODIV1		0
-#define GPC2CLK_OUT_VCODIV_MASK		(MASK(GPC2CLK_OUT_VCODIV_WIDTH) << \
-					GPC2CLK_OUT_VCODIV_SHIFT)
-#define GPC2CLK_OUT_BYPDIV_WIDTH	6
-#define GPC2CLK_OUT_BYPDIV_SHIFT	0
-#define GPC2CLK_OUT_BYPDIV31		0x3c
-#define GPC2CLK_OUT_INIT_MASK	((MASK(GPC2CLK_OUT_SDIV14_INDIV4_WIDTH) << \
-		GPC2CLK_OUT_SDIV14_INDIV4_SHIFT)\
-		| (MASK(GPC2CLK_OUT_VCODIV_WIDTH) << GPC2CLK_OUT_VCODIV_SHIFT)\
-		| (MASK(GPC2CLK_OUT_BYPDIV_WIDTH) << GPC2CLK_OUT_BYPDIV_SHIFT))
-#define GPC2CLK_OUT_INIT_VAL	((GPC2CLK_OUT_SDIV14_INDIV4_MODE << \
-		GPC2CLK_OUT_SDIV14_INDIV4_SHIFT) \
-		| (GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT) \
-		| (GPC2CLK_OUT_BYPDIV31 << GPC2CLK_OUT_BYPDIV_SHIFT))
-
-#define GPC_BCAST_NDIV_SLOWDOWN_DEBUG	(GPC_BCASE_GPCPLL_CFG_BASE + 0xa0)
-#define GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_SHIFT	24
-#define GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK \
-	    (0x1 << GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_SHIFT)
-
 static const u8 _pl_to_div[] = {
 /* PL:   0, 1, 2, 3, 4, 5, 6,  7,  8,  9, 10, 11, 12, 13, 14 */
 /* p: */ 1, 2, 3, 4, 5, 6, 8, 10, 12, 16, 12, 16, 20, 24, 32,
@@ -124,7 +61,7 @@ static const struct gk20a_clk_pllg_params gk20a_pllg_params = {
 	.min_pl = 1, .max_pl = 32,
 };
 
-static void
+void
 gk20a_pllg_read_mnp(struct gk20a_clk *clk, struct gk20a_pll *pll)
 {
 	struct nvkm_device *device = clk->base.subdev.device;
@@ -136,20 +73,33 @@ gk20a_pllg_read_mnp(struct gk20a_clk *clk, struct gk20a_pll *pll)
 	pll->pl = (val >> GPCPLL_COEFF_P_SHIFT) & MASK(GPCPLL_COEFF_P_WIDTH);
 }
 
-static u32
-gk20a_pllg_calc_rate(struct gk20a_clk *clk)
+void
+gk20a_pllg_write_mnp(struct gk20a_clk *clk, const struct gk20a_pll *pll)
+{
+	struct nvkm_device *device = clk->base.subdev.device;
+	u32 val;
+
+	val = (pll->m & MASK(GPCPLL_COEFF_M_WIDTH)) << GPCPLL_COEFF_M_SHIFT;
+	val |= (pll->n & MASK(GPCPLL_COEFF_N_WIDTH)) << GPCPLL_COEFF_N_SHIFT;
+	val |= (pll->pl & MASK(GPCPLL_COEFF_P_WIDTH)) << GPCPLL_COEFF_P_SHIFT;
+	nvkm_wr32(device, GPCPLL_COEFF, val);
+}
+
+u32
+gk20a_pllg_calc_rate(struct gk20a_clk *clk, struct gk20a_pll *pll)
 {
 	u32 rate;
 	u32 divider;
 
-	rate = clk->parent_rate * clk->pll.n;
-	divider = clk->pll.m * clk->pl_to_div(clk->pll.pl);
+	rate = clk->parent_rate * pll->n;
+	divider = pll->m * clk->pl_to_div(pll->pl);
 
 	return rate / divider / 2;
 }
 
-static int
-gk20a_pllg_calc_mnp(struct gk20a_clk *clk, unsigned long rate)
+int
+gk20a_pllg_calc_mnp(struct gk20a_clk *clk, unsigned long rate,
+		    struct gk20a_pll *pll)
 {
 	struct nvkm_subdev *subdev = &clk->base.subdev;
 	u32 target_clk_f, ref_clk_f, target_freq;
@@ -163,16 +113,13 @@ gk20a_pllg_calc_mnp(struct gk20a_clk *clk, unsigned long rate)
 	target_clk_f = rate * 2 / KHZ;
 	ref_clk_f = clk->parent_rate / KHZ;
 
-	max_vco_f = clk->params->max_vco;
+	target_vco_f = target_clk_f + target_clk_f / 50;
+	max_vco_f = max(clk->params->max_vco, target_vco_f);
 	min_vco_f = clk->params->min_vco;
 	best_m = clk->params->max_m;
 	best_n = clk->params->min_n;
 	best_pl = clk->params->min_pl;
 
-	target_vco_f = target_clk_f + target_clk_f / 50;
-	if (max_vco_f < target_vco_f)
-		max_vco_f = target_vco_f;
-
 	/* min_pl <= high_pl <= max_pl */
 	high_pl = (max_vco_f + target_vco_f - 1) / target_vco_f;
 	high_pl = min(high_pl, clk->params->max_pl);
@@ -195,9 +142,7 @@ gk20a_pllg_calc_mnp(struct gk20a_clk *clk, unsigned long rate)
 		target_vco_f = target_clk_f * clk->pl_to_div(pl);
 
 		for (m = clk->params->min_m; m <= clk->params->max_m; m++) {
-			u32 u_f, vco_f;
-
-			u_f = ref_clk_f / m;
+			u32 u_f = ref_clk_f / m;
 
 			if (u_f < clk->params->min_u)
 				break;
@@ -211,6 +156,8 @@ gk20a_pllg_calc_mnp(struct gk20a_clk *clk, unsigned long rate)
 				break;
 
 			for (; n <= n2; n++) {
+				u32 vco_f;
+
 				if (n < clk->params->min_n)
 					continue;
 				if (n > clk->params->max_n)
@@ -247,16 +194,16 @@ found_match:
 			   "no best match for target @ %dMHz on gpc_pll",
 			   target_clk_f / KHZ);
 
-	clk->pll.m = best_m;
-	clk->pll.n = best_n;
-	clk->pll.pl = best_pl;
+	pll->m = best_m;
+	pll->n = best_n;
+	pll->pl = best_pl;
 
-	target_freq = gk20a_pllg_calc_rate(clk);
+	target_freq = gk20a_pllg_calc_rate(clk, pll);
 
 	nvkm_debug(subdev,
-		   "actual target freq %d MHz, M %d, N %d, PL %d(div%d)\n",
-		   target_freq / MHZ, clk->pll.m, clk->pll.n, clk->pll.pl,
-		   clk->pl_to_div(clk->pll.pl));
+		   "actual target freq %d KHz, M %d, N %d, PL %d(div%d)\n",
+		   target_freq / KHZ, pll->m, pll->n, pll->pl,
+		   clk->pl_to_div(pll->pl));
 	return 0;
 }
 
@@ -265,45 +212,36 @@ gk20a_pllg_slide(struct gk20a_clk *clk, u32 n)
 {
 	struct nvkm_subdev *subdev = &clk->base.subdev;
 	struct nvkm_device *device = subdev->device;
-	u32 val;
-	int ramp_timeout;
+	struct gk20a_pll pll;
+	int ret = 0;
 
 	/* get old coefficients */
-	val = nvkm_rd32(device, GPCPLL_COEFF);
+	gk20a_pllg_read_mnp(clk, &pll);
 	/* do nothing if NDIV is the same */
-	if (n == ((val >> GPCPLL_COEFF_N_SHIFT) & MASK(GPCPLL_COEFF_N_WIDTH)))
+	if (n == pll.n)
 		return 0;
 
-	/* setup */
-	nvkm_mask(device, GPCPLL_CFG2, 0xff << GPCPLL_CFG2_PLL_STEPA_SHIFT,
-		0x2b << GPCPLL_CFG2_PLL_STEPA_SHIFT);
-	nvkm_mask(device, GPCPLL_CFG3, 0xff << GPCPLL_CFG3_PLL_STEPB_SHIFT,
-		0xb << GPCPLL_CFG3_PLL_STEPB_SHIFT);
-
 	/* pll slowdown mode */
 	nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN,
 		BIT(GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT),
 		BIT(GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT));
 
 	/* new ndiv ready for ramp */
-	val = nvkm_rd32(device, GPCPLL_COEFF);
-	val &= ~(MASK(GPCPLL_COEFF_N_WIDTH) << GPCPLL_COEFF_N_SHIFT);
-	val |= (n & MASK(GPCPLL_COEFF_N_WIDTH)) << GPCPLL_COEFF_N_SHIFT;
+	pll.n = n;
 	udelay(1);
-	nvkm_wr32(device, GPCPLL_COEFF, val);
+	gk20a_pllg_write_mnp(clk, &pll);
 
 	/* dynamic ramp to new ndiv */
-	val = nvkm_rd32(device, GPCPLL_NDIV_SLOWDOWN);
-	val |= 0x1 << GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT;
 	udelay(1);
-	nvkm_wr32(device, GPCPLL_NDIV_SLOWDOWN, val);
+	nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN,
+		  BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT),
+		  BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT));
 
-	for (ramp_timeout = 500; ramp_timeout > 0; ramp_timeout--) {
-		udelay(1);
-		val = nvkm_rd32(device, GPC_BCAST_NDIV_SLOWDOWN_DEBUG);
-		if (val & GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK)
-			break;
-	}
+	/* wait for ramping to complete */
+	if (nvkm_wait_usec(device, 500, GPC_BCAST_NDIV_SLOWDOWN_DEBUG,
+		GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK,
+		GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK) < 0)
+		ret = -ETIMEDOUT;
 
 	/* exit slowdown mode */
 	nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN,
@@ -311,21 +249,35 @@ gk20a_pllg_slide(struct gk20a_clk *clk, u32 n)
 		BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT), 0);
 	nvkm_rd32(device, GPCPLL_NDIV_SLOWDOWN);
 
-	if (ramp_timeout <= 0) {
-		nvkm_error(subdev, "gpcpll dynamic ramp timeout\n");
-		return -ETIMEDOUT;
-	}
-
-	return 0;
+	return ret;
 }
 
-static void
+static int
 gk20a_pllg_enable(struct gk20a_clk *clk)
 {
 	struct nvkm_device *device = clk->base.subdev.device;
+	u32 val;
 
 	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_ENABLE, GPCPLL_CFG_ENABLE);
 	nvkm_rd32(device, GPCPLL_CFG);
+
+	/* enable lock detection */
+	val = nvkm_rd32(device, GPCPLL_CFG);
+	if (val & GPCPLL_CFG_LOCK_DET_OFF) {
+		val &= ~GPCPLL_CFG_LOCK_DET_OFF;
+		nvkm_wr32(device, GPCPLL_CFG, val);
+	}
+
+	/* wait for lock */
+	if (nvkm_wait_usec(device, 300, GPCPLL_CFG, GPCPLL_CFG_LOCK,
+			   GPCPLL_CFG_LOCK) < 0)
+		return -ETIMEDOUT;
+
+	/* switch to VCO mode */
+	nvkm_mask(device, SEL_VCO, BIT(SEL_VCO_GPC2CLK_OUT_SHIFT),
+		BIT(SEL_VCO_GPC2CLK_OUT_SHIFT));
+
+	return 0;
 }
 
 static void
@@ -333,117 +285,81 @@ gk20a_pllg_disable(struct gk20a_clk *clk)
 {
 	struct nvkm_device *device = clk->base.subdev.device;
 
+	/* put PLL in bypass before disabling it */
+	nvkm_mask(device, SEL_VCO, BIT(SEL_VCO_GPC2CLK_OUT_SHIFT), 0);
+
 	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_ENABLE, 0);
 	nvkm_rd32(device, GPCPLL_CFG);
 }
 
 static int
-_gk20a_pllg_program_mnp(struct gk20a_clk *clk, bool allow_slide)
+gk20a_pllg_program_mnp(struct gk20a_clk *clk, const struct gk20a_pll *pll)
 {
 	struct nvkm_subdev *subdev = &clk->base.subdev;
 	struct nvkm_device *device = subdev->device;
-	u32 val, cfg;
-	struct gk20a_pll old_pll;
-	u32 n_lo;
-
-	/* get old coefficients */
-	gk20a_pllg_read_mnp(clk, &old_pll);
-
-	/* do NDIV slide if there is no change in M and PL */
-	cfg = nvkm_rd32(device, GPCPLL_CFG);
-	if (allow_slide && clk->pll.m == old_pll.m &&
-	    clk->pll.pl == old_pll.pl && (cfg & GPCPLL_CFG_ENABLE)) {
-		return gk20a_pllg_slide(clk, clk->pll.n);
-	}
-
-	/* slide down to NDIV_LO */
-	if (allow_slide && (cfg & GPCPLL_CFG_ENABLE)) {
-		int ret;
-
-		n_lo = DIV_ROUND_UP(old_pll.m * clk->params->min_vco,
-				    clk->parent_rate / KHZ);
-		ret = gk20a_pllg_slide(clk, n_lo);
+	struct gk20a_pll cur_pll;
+	int ret;
 
-		if (ret)
-			return ret;
-	}
+	gk20a_pllg_read_mnp(clk, &cur_pll);
 
-	/* split FO-to-bypass jump in halfs by setting out divider 1:2 */
+	/* split VCO-to-bypass jump in half by setting out divider 1:2 */
 	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK,
-		0x2 << GPC2CLK_OUT_VCODIV_SHIFT);
-
-	/* put PLL in bypass before programming it */
-	val = nvkm_rd32(device, SEL_VCO);
-	val &= ~(BIT(SEL_VCO_GPC2CLK_OUT_SHIFT));
+		  GPC2CLK_OUT_VCODIV2 << GPC2CLK_OUT_VCODIV_SHIFT);
+	/* Intentional 2nd write to assure linear divider operation */
+	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK,
+		  GPC2CLK_OUT_VCODIV2 << GPC2CLK_OUT_VCODIV_SHIFT);
+	nvkm_rd32(device, GPC2CLK_OUT);
 	udelay(2);
-	nvkm_wr32(device, SEL_VCO, val);
-
-	/* get out from IDDQ */
-	val = nvkm_rd32(device, GPCPLL_CFG);
-	if (val & GPCPLL_CFG_IDDQ) {
-		val &= ~GPCPLL_CFG_IDDQ;
-		nvkm_wr32(device, GPCPLL_CFG, val);
-		nvkm_rd32(device, GPCPLL_CFG);
-		udelay(2);
-	}
 
 	gk20a_pllg_disable(clk);
 
-	nvkm_debug(subdev, "%s: m=%d n=%d pl=%d\n", __func__,
-		   clk->pll.m, clk->pll.n, clk->pll.pl);
-
-	n_lo = DIV_ROUND_UP(clk->pll.m * clk->params->min_vco,
-			    clk->parent_rate / KHZ);
-	val = clk->pll.m << GPCPLL_COEFF_M_SHIFT;
-	val |= (allow_slide ? n_lo : clk->pll.n) << GPCPLL_COEFF_N_SHIFT;
-	val |= clk->pll.pl << GPCPLL_COEFF_P_SHIFT;
-	nvkm_wr32(device, GPCPLL_COEFF, val);
+	gk20a_pllg_write_mnp(clk, pll);
 
-	gk20a_pllg_enable(clk);
-
-	val = nvkm_rd32(device, GPCPLL_CFG);
-	if (val & GPCPLL_CFG_LOCK_DET_OFF) {
-		val &= ~GPCPLL_CFG_LOCK_DET_OFF;
-		nvkm_wr32(device, GPCPLL_CFG, val);
-	}
-
-	if (nvkm_usec(device, 300,
-		if (nvkm_rd32(device, GPCPLL_CFG) & GPCPLL_CFG_LOCK)
-			break;
-	) < 0)
-		return -ETIMEDOUT;
-
-	/* switch to VCO mode */
-	nvkm_mask(device, SEL_VCO, BIT(SEL_VCO_GPC2CLK_OUT_SHIFT),
-		  BIT(SEL_VCO_GPC2CLK_OUT_SHIFT));
+	ret = gk20a_pllg_enable(clk);
+	if (ret)
+		return ret;
 
 	/* restore out divider 1:1 */
-	val = nvkm_rd32(device, GPC2CLK_OUT);
-	if ((val & GPC2CLK_OUT_VCODIV_MASK) !=
-	    (GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT)) {
-		val &= ~GPC2CLK_OUT_VCODIV_MASK;
-		val |= GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT;
-		udelay(2);
-		nvkm_wr32(device, GPC2CLK_OUT, val);
-		/* Intentional 2nd write to assure linear divider operation */
-		nvkm_wr32(device, GPC2CLK_OUT, val);
-		nvkm_rd32(device, GPC2CLK_OUT);
-	}
+	udelay(2);
+	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK,
+		  GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT);
+	/* Intentional 2nd write to assure linear divider operation */
+	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK,
+		  GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT);
+	nvkm_rd32(device, GPC2CLK_OUT);
 
-	/* slide up to new NDIV */
-	return allow_slide ? gk20a_pllg_slide(clk, clk->pll.n) : 0;
+	return 0;
 }
 
 static int
-gk20a_pllg_program_mnp(struct gk20a_clk *clk)
+gk20a_pllg_program_mnp_slide(struct gk20a_clk *clk, const struct gk20a_pll *pll)
 {
-	int err;
+	struct gk20a_pll cur_pll;
+	int ret;
 
-	err = _gk20a_pllg_program_mnp(clk, true);
-	if (err)
-		err = _gk20a_pllg_program_mnp(clk, false);
+	if (gk20a_pllg_is_enabled(clk)) {
+		gk20a_pllg_read_mnp(clk, &cur_pll);
+
+		/* just do NDIV slide if there is no change to M and PL */
+		if (pll->m == cur_pll.m && pll->pl == cur_pll.pl)
+			return gk20a_pllg_slide(clk, pll->n);
+
+		/* slide down to current NDIV_LO */
+		cur_pll.n = gk20a_pllg_n_lo(clk, &cur_pll);
+		ret = gk20a_pllg_slide(clk, cur_pll.n);
+		if (ret)
+			return ret;
+	}
+
+	/* program MNP with the new clock parameters and new NDIV_LO */
+	cur_pll = *pll;
+	cur_pll.n = gk20a_pllg_n_lo(clk, &cur_pll);
+	ret = gk20a_pllg_program_mnp(clk, &cur_pll);
+	if (ret)
+		return ret;
 
-	return err;
+	/* slide up to new NDIV */
+	return gk20a_pllg_slide(clk, pll->n);
 }
 
 static struct nvkm_pstate
@@ -546,13 +462,14 @@ gk20a_clk_read(struct nvkm_clk *base, enum nv_clk_src src)
 	struct gk20a_clk *clk = gk20a_clk(base);
 	struct nvkm_subdev *subdev = &clk->base.subdev;
 	struct nvkm_device *device = subdev->device;
+	struct gk20a_pll pll;
 
 	switch (src) {
 	case nv_clk_src_crystal:
 		return device->crystal;
 	case nv_clk_src_gpc:
-		gk20a_pllg_read_mnp(clk, &clk->pll);
-		return gk20a_pllg_calc_rate(clk) / GK20A_CLK_GPC_MDIV;
+		gk20a_pllg_read_mnp(clk, &pll);
+		return gk20a_pllg_calc_rate(clk, &pll) / GK20A_CLK_GPC_MDIV;
 	default:
 		nvkm_error(subdev, "invalid clock source %d\n", src);
 		return -EINVAL;
@@ -565,15 +482,20 @@ gk20a_clk_calc(struct nvkm_clk *base, struct nvkm_cstate *cstate)
 	struct gk20a_clk *clk = gk20a_clk(base);
 
 	return gk20a_pllg_calc_mnp(clk, cstate->domain[nv_clk_src_gpc] *
-					 GK20A_CLK_GPC_MDIV);
+					 GK20A_CLK_GPC_MDIV, &clk->pll);
 }
 
 int
 gk20a_clk_prog(struct nvkm_clk *base)
 {
 	struct gk20a_clk *clk = gk20a_clk(base);
+	int ret;
+
+	ret = gk20a_pllg_program_mnp_slide(clk, &clk->pll);
+	if (ret)
+		ret = gk20a_pllg_program_mnp(clk, &clk->pll);
 
-	return gk20a_pllg_program_mnp(clk);
+	return ret;
 }
 
 void
@@ -581,29 +503,62 @@ gk20a_clk_tidy(struct nvkm_clk *base)
 {
 }
 
+int
+gk20a_clk_setup_slide(struct gk20a_clk *clk)
+{
+	struct nvkm_subdev *subdev = &clk->base.subdev;
+	struct nvkm_device *device = subdev->device;
+	u32 step_a, step_b;
+
+	switch (clk->parent_rate) {
+	case 12000000:
+	case 12800000:
+	case 13000000:
+		step_a = 0x2b;
+		step_b = 0x0b;
+		break;
+	case 19200000:
+		step_a = 0x12;
+		step_b = 0x08;
+		break;
+	case 38400000:
+		step_a = 0x04;
+		step_b = 0x05;
+		break;
+	default:
+		nvkm_error(subdev, "invalid parent clock rate %u KHz",
+			   clk->parent_rate / KHZ);
+		return -EINVAL;
+	}
+
+	nvkm_mask(device, GPCPLL_CFG2, 0xff << GPCPLL_CFG2_PLL_STEPA_SHIFT,
+		step_a << GPCPLL_CFG2_PLL_STEPA_SHIFT);
+	nvkm_mask(device, GPCPLL_CFG3, 0xff << GPCPLL_CFG3_PLL_STEPB_SHIFT,
+		step_b << GPCPLL_CFG3_PLL_STEPB_SHIFT);
+
+	return 0;
+}
+
 void
 gk20a_clk_fini(struct nvkm_clk *base)
 {
 	struct nvkm_device *device = base->subdev.device;
 	struct gk20a_clk *clk = gk20a_clk(base);
-	u32 val;
 
 	/* slide to VCO min */
-	val = nvkm_rd32(device, GPCPLL_CFG);
-	if (val & GPCPLL_CFG_ENABLE) {
+	if (gk20a_pllg_is_enabled(clk)) {
 		struct gk20a_pll pll;
 		u32 n_lo;
 
 		gk20a_pllg_read_mnp(clk, &pll);
-		n_lo = DIV_ROUND_UP(pll.m * clk->params->min_vco,
-				    clk->parent_rate / KHZ);
+		n_lo = gk20a_pllg_n_lo(clk, &pll);
 		gk20a_pllg_slide(clk, n_lo);
 	}
 
-	/* put PLL in bypass before disabling it */
-	nvkm_mask(device, SEL_VCO, BIT(SEL_VCO_GPC2CLK_OUT_SHIFT), 0);
-
 	gk20a_pllg_disable(clk);
+
+	/* set IDDQ */
+	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_IDDQ, 1);
 }
 
 static int
@@ -614,9 +569,18 @@ gk20a_clk_init(struct nvkm_clk *base)
 	struct nvkm_device *device = subdev->device;
 	int ret;
 
+	/* get out from IDDQ */
+	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_IDDQ, 0);
+	nvkm_rd32(device, GPCPLL_CFG);
+	udelay(5);
+
 	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_INIT_MASK,
 		  GPC2CLK_OUT_INIT_VAL);
 
+	ret = gk20a_clk_setup_slide(clk);
+	if (ret)
+		return ret;
+
 	/* Start with lowest frequency */
 	base->func->calc(base, &base->func->pstates[0].base);
 	ret = base->func->prog(&clk->base);
@@ -646,7 +610,7 @@ gk20a_clk = {
 };
 
 int
-_gk20a_clk_ctor(struct nvkm_device *device, int index,
+gk20a_clk_ctor(struct nvkm_device *device, int index,
 		const struct nvkm_clk_func *func,
 		const struct gk20a_clk_pllg_params *params,
 		struct gk20a_clk *clk)
@@ -685,7 +649,7 @@ gk20a_clk_new(struct nvkm_device *device, int index, struct nvkm_clk **pclk)
 		return -ENOMEM;
 	*pclk = &clk->base;
 
-	ret = _gk20a_clk_ctor(device, index, &gk20a_clk, &gk20a_pllg_params,
+	ret = gk20a_clk_ctor(device, index, &gk20a_clk, &gk20a_pllg_params,
 			      clk);
 
 	clk->pl_to_div = pl_to_div;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.h b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.h
index 13c467401..0d1450972 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gk20a.h
@@ -24,9 +24,79 @@
 #ifndef __NVKM_CLK_GK20A_H__
 #define __NVKM_CLK_GK20A_H__
 
+#define KHZ (1000)
+#define MHZ (KHZ * 1000)
+
+#define MASK(w)	((1 << (w)) - 1)
+
 #define GK20A_CLK_GPC_MDIV 1000
 
 #define SYS_GPCPLL_CFG_BASE	0x00137000
+#define GPCPLL_CFG		(SYS_GPCPLL_CFG_BASE + 0)
+#define GPCPLL_CFG_ENABLE	BIT(0)
+#define GPCPLL_CFG_IDDQ		BIT(1)
+#define GPCPLL_CFG_LOCK_DET_OFF	BIT(4)
+#define GPCPLL_CFG_LOCK		BIT(17)
+
+#define GPCPLL_CFG2		(SYS_GPCPLL_CFG_BASE + 0xc)
+#define GPCPLL_CFG2_SETUP2_SHIFT	16
+#define GPCPLL_CFG2_PLL_STEPA_SHIFT	24
+
+#define GPCPLL_CFG3			(SYS_GPCPLL_CFG_BASE + 0x18)
+#define GPCPLL_CFG3_VCO_CTRL_SHIFT		0
+#define GPCPLL_CFG3_VCO_CTRL_WIDTH		9
+#define GPCPLL_CFG3_VCO_CTRL_MASK		\
+	(MASK(GPCPLL_CFG3_VCO_CTRL_WIDTH) << GPCPLL_CFG3_VCO_CTRL_SHIFT)
+#define GPCPLL_CFG3_PLL_STEPB_SHIFT		16
+#define GPCPLL_CFG3_PLL_STEPB_WIDTH		8
+
+#define GPCPLL_COEFF		(SYS_GPCPLL_CFG_BASE + 4)
+#define GPCPLL_COEFF_M_SHIFT	0
+#define GPCPLL_COEFF_M_WIDTH	8
+#define GPCPLL_COEFF_N_SHIFT	8
+#define GPCPLL_COEFF_N_WIDTH	8
+#define GPCPLL_COEFF_N_MASK	\
+	(MASK(GPCPLL_COEFF_N_WIDTH) << GPCPLL_COEFF_N_SHIFT)
+#define GPCPLL_COEFF_P_SHIFT	16
+#define GPCPLL_COEFF_P_WIDTH	6
+
+#define GPCPLL_NDIV_SLOWDOWN			(SYS_GPCPLL_CFG_BASE + 0x1c)
+#define GPCPLL_NDIV_SLOWDOWN_NDIV_LO_SHIFT	0
+#define GPCPLL_NDIV_SLOWDOWN_NDIV_MID_SHIFT	8
+#define GPCPLL_NDIV_SLOWDOWN_STEP_SIZE_LO2MID_SHIFT	16
+#define GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT	22
+#define GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT	31
+
+#define GPC_BCAST_GPCPLL_CFG_BASE		0x00132800
+#define GPC_BCAST_NDIV_SLOWDOWN_DEBUG	(GPC_BCAST_GPCPLL_CFG_BASE + 0xa0)
+#define GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_SHIFT	24
+#define GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK \
+	(0x1 << GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_SHIFT)
+
+#define SEL_VCO				(SYS_GPCPLL_CFG_BASE + 0x100)
+#define SEL_VCO_GPC2CLK_OUT_SHIFT	0
+
+#define GPC2CLK_OUT			(SYS_GPCPLL_CFG_BASE + 0x250)
+#define GPC2CLK_OUT_SDIV14_INDIV4_WIDTH	1
+#define GPC2CLK_OUT_SDIV14_INDIV4_SHIFT	31
+#define GPC2CLK_OUT_SDIV14_INDIV4_MODE	1
+#define GPC2CLK_OUT_VCODIV_WIDTH	6
+#define GPC2CLK_OUT_VCODIV_SHIFT	8
+#define GPC2CLK_OUT_VCODIV1		0
+#define GPC2CLK_OUT_VCODIV2		2
+#define GPC2CLK_OUT_VCODIV_MASK		(MASK(GPC2CLK_OUT_VCODIV_WIDTH) << \
+					GPC2CLK_OUT_VCODIV_SHIFT)
+#define GPC2CLK_OUT_BYPDIV_WIDTH	6
+#define GPC2CLK_OUT_BYPDIV_SHIFT	0
+#define GPC2CLK_OUT_BYPDIV31		0x3c
+#define GPC2CLK_OUT_INIT_MASK	((MASK(GPC2CLK_OUT_SDIV14_INDIV4_WIDTH) << \
+		GPC2CLK_OUT_SDIV14_INDIV4_SHIFT)\
+		| (MASK(GPC2CLK_OUT_VCODIV_WIDTH) << GPC2CLK_OUT_VCODIV_SHIFT)\
+		| (MASK(GPC2CLK_OUT_BYPDIV_WIDTH) << GPC2CLK_OUT_BYPDIV_SHIFT))
+#define GPC2CLK_OUT_INIT_VAL	((GPC2CLK_OUT_SDIV14_INDIV4_MODE << \
+		GPC2CLK_OUT_SDIV14_INDIV4_SHIFT) \
+		| (GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT) \
+		| (GPC2CLK_OUT_BYPDIV31 << GPC2CLK_OUT_BYPDIV_SHIFT))
 
 /* All frequencies in Khz */
 struct gk20a_clk_pllg_params {
@@ -54,7 +124,29 @@ struct gk20a_clk {
 };
 #define gk20a_clk(p) container_of((p), struct gk20a_clk, base)
 
-int _gk20a_clk_ctor(struct nvkm_device *, int, const struct nvkm_clk_func *,
+u32 gk20a_pllg_calc_rate(struct gk20a_clk *, struct gk20a_pll *);
+int gk20a_pllg_calc_mnp(struct gk20a_clk *, unsigned long, struct gk20a_pll *);
+void gk20a_pllg_read_mnp(struct gk20a_clk *, struct gk20a_pll *);
+void gk20a_pllg_write_mnp(struct gk20a_clk *, const struct gk20a_pll *);
+
+static inline bool
+gk20a_pllg_is_enabled(struct gk20a_clk *clk)
+{
+	struct nvkm_device *device = clk->base.subdev.device;
+	u32 val;
+
+	val = nvkm_rd32(device, GPCPLL_CFG);
+	return val & GPCPLL_CFG_ENABLE;
+}
+
+static inline u32
+gk20a_pllg_n_lo(struct gk20a_clk *clk, struct gk20a_pll *pll)
+{
+	return DIV_ROUND_UP(pll->m * clk->params->min_vco,
+			    clk->parent_rate / KHZ);
+}
+
+int gk20a_clk_ctor(struct nvkm_device *, int, const struct nvkm_clk_func *,
 		    const struct gk20a_clk_pllg_params *, struct gk20a_clk *);
 void gk20a_clk_fini(struct nvkm_clk *);
 int gk20a_clk_read(struct nvkm_clk *, enum nv_clk_src);
@@ -62,4 +154,6 @@ int gk20a_clk_calc(struct nvkm_clk *, struct nvkm_cstate *);
 int gk20a_clk_prog(struct nvkm_clk *);
 void gk20a_clk_tidy(struct nvkm_clk *);
 
+int gk20a_clk_setup_slide(struct gk20a_clk *);
+
 #endif
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gm20b.c b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gm20b.c
index 71b2bbb61..b284e949f 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gm20b.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/clk/gm20b.c
@@ -21,20 +21,123 @@
  */
 
 #include <subdev/clk.h>
+#include <subdev/volt.h>
+#include <subdev/timer.h>
 #include <core/device.h>
+#include <core/tegra.h>
 
 #include "priv.h"
 #include "gk20a.h"
 
-#define KHZ (1000)
-#define MHZ (KHZ * 1000)
-
-#define MASK(w)	((1 << w) - 1)
+#define GPCPLL_CFG_SYNC_MODE	BIT(2)
 
 #define BYPASSCTRL_SYS	(SYS_GPCPLL_CFG_BASE + 0x340)
 #define BYPASSCTRL_SYS_GPCPLL_SHIFT	0
 #define BYPASSCTRL_SYS_GPCPLL_WIDTH	1
 
+#define GPCPLL_CFG2_SDM_DIN_SHIFT	0
+#define GPCPLL_CFG2_SDM_DIN_WIDTH	8
+#define GPCPLL_CFG2_SDM_DIN_MASK	\
+	(MASK(GPCPLL_CFG2_SDM_DIN_WIDTH) << GPCPLL_CFG2_SDM_DIN_SHIFT)
+#define GPCPLL_CFG2_SDM_DIN_NEW_SHIFT	8
+#define GPCPLL_CFG2_SDM_DIN_NEW_WIDTH	15
+#define GPCPLL_CFG2_SDM_DIN_NEW_MASK	\
+	(MASK(GPCPLL_CFG2_SDM_DIN_NEW_WIDTH) << GPCPLL_CFG2_SDM_DIN_NEW_SHIFT)
+#define GPCPLL_CFG2_SETUP2_SHIFT	16
+#define GPCPLL_CFG2_PLL_STEPA_SHIFT	24
+
+#define GPCPLL_DVFS0	(SYS_GPCPLL_CFG_BASE + 0x10)
+#define GPCPLL_DVFS0_DFS_COEFF_SHIFT	0
+#define GPCPLL_DVFS0_DFS_COEFF_WIDTH	7
+#define GPCPLL_DVFS0_DFS_COEFF_MASK	\
+	(MASK(GPCPLL_DVFS0_DFS_COEFF_WIDTH) << GPCPLL_DVFS0_DFS_COEFF_SHIFT)
+#define GPCPLL_DVFS0_DFS_DET_MAX_SHIFT	8
+#define GPCPLL_DVFS0_DFS_DET_MAX_WIDTH	7
+#define GPCPLL_DVFS0_DFS_DET_MAX_MASK	\
+	(MASK(GPCPLL_DVFS0_DFS_DET_MAX_WIDTH) << GPCPLL_DVFS0_DFS_DET_MAX_SHIFT)
+
+#define GPCPLL_DVFS1		(SYS_GPCPLL_CFG_BASE + 0x14)
+#define GPCPLL_DVFS1_DFS_EXT_DET_SHIFT		0
+#define GPCPLL_DVFS1_DFS_EXT_DET_WIDTH		7
+#define GPCPLL_DVFS1_DFS_EXT_STRB_SHIFT		7
+#define GPCPLL_DVFS1_DFS_EXT_STRB_WIDTH		1
+#define GPCPLL_DVFS1_DFS_EXT_CAL_SHIFT		8
+#define GPCPLL_DVFS1_DFS_EXT_CAL_WIDTH		7
+#define GPCPLL_DVFS1_DFS_EXT_SEL_SHIFT		15
+#define GPCPLL_DVFS1_DFS_EXT_SEL_WIDTH		1
+#define GPCPLL_DVFS1_DFS_CTRL_SHIFT		16
+#define GPCPLL_DVFS1_DFS_CTRL_WIDTH		12
+#define GPCPLL_DVFS1_EN_SDM_SHIFT		28
+#define GPCPLL_DVFS1_EN_SDM_WIDTH		1
+#define GPCPLL_DVFS1_EN_SDM_BIT			BIT(28)
+#define GPCPLL_DVFS1_EN_DFS_SHIFT		29
+#define GPCPLL_DVFS1_EN_DFS_WIDTH		1
+#define GPCPLL_DVFS1_EN_DFS_BIT			BIT(29)
+#define GPCPLL_DVFS1_EN_DFS_CAL_SHIFT		30
+#define GPCPLL_DVFS1_EN_DFS_CAL_WIDTH		1
+#define GPCPLL_DVFS1_EN_DFS_CAL_BIT		BIT(30)
+#define GPCPLL_DVFS1_DFS_CAL_DONE_SHIFT		31
+#define GPCPLL_DVFS1_DFS_CAL_DONE_WIDTH		1
+#define GPCPLL_DVFS1_DFS_CAL_DONE_BIT		BIT(31)
+
+#define GPC_BCAST_GPCPLL_DVFS2	(GPC_BCAST_GPCPLL_CFG_BASE + 0x20)
+#define GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT	BIT(16)
+
+#define GPCPLL_CFG3_PLL_DFS_TESTOUT_SHIFT	24
+#define GPCPLL_CFG3_PLL_DFS_TESTOUT_WIDTH	7
+
+#define DFS_DET_RANGE	6	/* -2^6 ... 2^6-1 */
+#define SDM_DIN_RANGE	12	/* -2^12 ... 2^12-1 */
+
+struct gm20b_clk_dvfs_params {
+	s32 coeff_slope;
+	s32 coeff_offs;
+	u32 vco_ctrl;
+};
+
+static const struct gm20b_clk_dvfs_params gm20b_dvfs_params = {
+	.coeff_slope = -165230,
+	.coeff_offs = 214007,
+	.vco_ctrl = 0x7 << 3,
+};
+
+/*
+ * base.n is now the *integer* part of the N factor.
+ * sdm_din contains n's decimal part.
+ */
+struct gm20b_pll {
+	struct gk20a_pll base;
+	u32 sdm_din;
+};
+
+struct gm20b_clk_dvfs {
+	u32 dfs_coeff;
+	s32 dfs_det_max;
+	s32 dfs_ext_cal;
+};
+
+struct gm20b_clk {
+	/* currently applied parameters */
+	struct gk20a_clk base;
+	struct gm20b_clk_dvfs dvfs;
+	u32 uv;
+
+	/* new parameters to apply */
+	struct gk20a_pll new_pll;
+	struct gm20b_clk_dvfs new_dvfs;
+	u32 new_uv;
+
+	const struct gm20b_clk_dvfs_params *dvfs_params;
+
+	/* fused parameters */
+	s32 uvdet_slope;
+	s32 uvdet_offs;
+
+	/* safe frequency we can use at minimum voltage */
+	u32 safe_fmax_vmin;
+};
+#define gm20b_clk(p) container_of((gk20a_clk(p)), struct gm20b_clk, base)
+
 static u32 pl_to_div(u32 pl)
 {
 	return pl;
@@ -53,6 +156,484 @@ static const struct gk20a_clk_pllg_params gm20b_pllg_params = {
 	.min_pl = 1, .max_pl = 31,
 };
 
+static void
+gm20b_pllg_read_mnp(struct gm20b_clk *clk, struct gm20b_pll *pll)
+{
+	struct nvkm_subdev *subdev = &clk->base.base.subdev;
+	struct nvkm_device *device = subdev->device;
+	u32 val;
+
+	gk20a_pllg_read_mnp(&clk->base, &pll->base);
+	val = nvkm_rd32(device, GPCPLL_CFG2);
+	pll->sdm_din = (val >> GPCPLL_CFG2_SDM_DIN_SHIFT) &
+		       MASK(GPCPLL_CFG2_SDM_DIN_WIDTH);
+}
+
+static void
+gm20b_pllg_write_mnp(struct gm20b_clk *clk, const struct gm20b_pll *pll)
+{
+	struct nvkm_device *device = clk->base.base.subdev.device;
+
+	nvkm_mask(device, GPCPLL_CFG2, GPCPLL_CFG2_SDM_DIN_MASK,
+		  pll->sdm_din << GPCPLL_CFG2_SDM_DIN_SHIFT);
+	gk20a_pllg_write_mnp(&clk->base, &pll->base);
+}
+
+/*
+ * Determine DFS_COEFF for the requested voltage. Always select external
+ * calibration override equal to the voltage, and set maximum detection
+ * limit "0" (to make sure that PLL output remains under F/V curve when
+ * voltage increases).
+ */
+static void
+gm20b_dvfs_calc_det_coeff(struct gm20b_clk *clk, s32 uv,
+			  struct gm20b_clk_dvfs *dvfs)
+{
+	struct nvkm_subdev *subdev = &clk->base.base.subdev;
+	const struct gm20b_clk_dvfs_params *p = clk->dvfs_params;
+	u32 coeff;
+	/* Work with mv as uv would likely trigger an overflow */
+	s32 mv = DIV_ROUND_CLOSEST(uv, 1000);
+
+	/* coeff = slope * voltage + offset */
+	coeff = DIV_ROUND_CLOSEST(mv * p->coeff_slope, 1000) + p->coeff_offs;
+	coeff = DIV_ROUND_CLOSEST(coeff, 1000);
+	dvfs->dfs_coeff = min_t(u32, coeff, MASK(GPCPLL_DVFS0_DFS_COEFF_WIDTH));
+
+	dvfs->dfs_ext_cal = DIV_ROUND_CLOSEST(uv - clk->uvdet_offs,
+					     clk->uvdet_slope);
+	/* should never happen */
+	if (abs(dvfs->dfs_ext_cal) >= BIT(DFS_DET_RANGE))
+		nvkm_error(subdev, "dfs_ext_cal overflow!\n");
+
+	dvfs->dfs_det_max = 0;
+
+	nvkm_debug(subdev, "%s uv: %d coeff: %x, ext_cal: %d, det_max: %d\n",
+		   __func__, uv, dvfs->dfs_coeff, dvfs->dfs_ext_cal,
+		   dvfs->dfs_det_max);
+}
+
+/*
+ * Solve equation for integer and fractional part of the effective NDIV:
+ *
+ * n_eff = n_int + 1/2 + (SDM_DIN / 2^(SDM_DIN_RANGE + 1)) +
+ *         (DVFS_COEFF * DVFS_DET_DELTA) / 2^DFS_DET_RANGE
+ *
+ * The SDM_DIN LSB is finally shifted out, since it is not accessible by sw.
+ */
+static void
+gm20b_dvfs_calc_ndiv(struct gm20b_clk *clk, u32 n_eff, u32 *n_int, u32 *sdm_din)
+{
+	struct nvkm_subdev *subdev = &clk->base.base.subdev;
+	const struct gk20a_clk_pllg_params *p = clk->base.params;
+	u32 n;
+	s32 det_delta;
+	u32 rem, rem_range;
+
+	/* calculate current ext_cal and subtract previous one */
+	det_delta = DIV_ROUND_CLOSEST(((s32)clk->uv) - clk->uvdet_offs,
+				      clk->uvdet_slope);
+	det_delta -= clk->dvfs.dfs_ext_cal;
+	det_delta = min(det_delta, clk->dvfs.dfs_det_max);
+	det_delta *= clk->dvfs.dfs_coeff;
+
+	/* integer part of n */
+	n = (n_eff << DFS_DET_RANGE) - det_delta;
+	/* should never happen! */
+	if (n <= 0) {
+		nvkm_error(subdev, "ndiv <= 0 - setting to 1...\n");
+		n = 1 << DFS_DET_RANGE;
+	}
+	if (n >> DFS_DET_RANGE > p->max_n) {
+		nvkm_error(subdev, "ndiv > max_n - setting to max_n...\n");
+		n = p->max_n << DFS_DET_RANGE;
+	}
+	*n_int = n >> DFS_DET_RANGE;
+
+	/* fractional part of n */
+	rem = ((u32)n) & MASK(DFS_DET_RANGE);
+	rem_range = SDM_DIN_RANGE + 1 - DFS_DET_RANGE;
+	/* subtract 2^SDM_DIN_RANGE to account for the 1/2 of the equation */
+	rem = (rem << rem_range) - BIT(SDM_DIN_RANGE);
+	/* lose 8 LSB and clip - sdm_din only keeps the most significant byte */
+	*sdm_din = (rem >> BITS_PER_BYTE) & MASK(GPCPLL_CFG2_SDM_DIN_WIDTH);
+
+	nvkm_debug(subdev, "%s n_eff: %d, n_int: %d, sdm_din: %d\n", __func__,
+		   n_eff, *n_int, *sdm_din);
+}
+
+static int
+gm20b_pllg_slide(struct gm20b_clk *clk, u32 n)
+{
+	struct nvkm_subdev *subdev = &clk->base.base.subdev;
+	struct nvkm_device *device = subdev->device;
+	struct gm20b_pll pll;
+	u32 n_int, sdm_din;
+	int ret = 0;
+
+	/* calculate the new n_int/sdm_din for this n/uv */
+	gm20b_dvfs_calc_ndiv(clk, n, &n_int, &sdm_din);
+
+	/* get old coefficients */
+	gm20b_pllg_read_mnp(clk, &pll);
+	/* do nothing if NDIV is the same */
+	if (n_int == pll.base.n && sdm_din == pll.sdm_din)
+		return 0;
+
+	/* pll slowdown mode */
+	nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN,
+		BIT(GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT),
+		BIT(GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT));
+
+	/* new ndiv ready for ramp */
+	/* in DVFS mode SDM is updated via "new" field */
+	nvkm_mask(device, GPCPLL_CFG2, GPCPLL_CFG2_SDM_DIN_NEW_MASK,
+		  sdm_din << GPCPLL_CFG2_SDM_DIN_NEW_SHIFT);
+	pll.base.n = n_int;
+	udelay(1);
+	gk20a_pllg_write_mnp(&clk->base, &pll.base);
+
+	/* dynamic ramp to new ndiv */
+	udelay(1);
+	nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN,
+		  BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT),
+		  BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT));
+
+	/* wait for ramping to complete */
+	if (nvkm_wait_usec(device, 500, GPC_BCAST_NDIV_SLOWDOWN_DEBUG,
+		GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK,
+		GPC_BCAST_NDIV_SLOWDOWN_DEBUG_PLL_DYNRAMP_DONE_SYNCED_MASK) < 0)
+		ret = -ETIMEDOUT;
+
+	/* in DVFS mode complete SDM update */
+	nvkm_mask(device, GPCPLL_CFG2, GPCPLL_CFG2_SDM_DIN_MASK,
+		  sdm_din << GPCPLL_CFG2_SDM_DIN_SHIFT);
+
+	/* exit slowdown mode */
+	nvkm_mask(device, GPCPLL_NDIV_SLOWDOWN,
+		BIT(GPCPLL_NDIV_SLOWDOWN_SLOWDOWN_USING_PLL_SHIFT) |
+		BIT(GPCPLL_NDIV_SLOWDOWN_EN_DYNRAMP_SHIFT), 0);
+	nvkm_rd32(device, GPCPLL_NDIV_SLOWDOWN);
+
+	return ret;
+}
+
+static int
+gm20b_pllg_enable(struct gm20b_clk *clk)
+{
+	struct nvkm_device *device = clk->base.base.subdev.device;
+
+	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_ENABLE, GPCPLL_CFG_ENABLE);
+	nvkm_rd32(device, GPCPLL_CFG);
+
+	/* In DVFS mode lock cannot be used - so just delay */
+	udelay(40);
+
+	/* set SYNC_MODE for glitchless switch out of bypass */
+	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_SYNC_MODE,
+		       GPCPLL_CFG_SYNC_MODE);
+	nvkm_rd32(device, GPCPLL_CFG);
+
+	/* switch to VCO mode */
+	nvkm_mask(device, SEL_VCO, BIT(SEL_VCO_GPC2CLK_OUT_SHIFT),
+		  BIT(SEL_VCO_GPC2CLK_OUT_SHIFT));
+
+	return 0;
+}
+
+static void
+gm20b_pllg_disable(struct gm20b_clk *clk)
+{
+	struct nvkm_device *device = clk->base.base.subdev.device;
+
+	/* put PLL in bypass before disabling it */
+	nvkm_mask(device, SEL_VCO, BIT(SEL_VCO_GPC2CLK_OUT_SHIFT), 0);
+
+	/* clear SYNC_MODE before disabling PLL */
+	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_SYNC_MODE, 0);
+
+	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_ENABLE, 0);
+	nvkm_rd32(device, GPCPLL_CFG);
+}
+
+static int
+gm20b_pllg_program_mnp(struct gm20b_clk *clk, const struct gk20a_pll *pll)
+{
+	struct nvkm_subdev *subdev = &clk->base.base.subdev;
+	struct nvkm_device *device = subdev->device;
+	struct gm20b_pll cur_pll;
+	u32 n_int, sdm_din;
+	/* if we only change pdiv, we can do a glitchless transition */
+	bool pdiv_only;
+	int ret;
+
+	gm20b_dvfs_calc_ndiv(clk, pll->n, &n_int, &sdm_din);
+	gm20b_pllg_read_mnp(clk, &cur_pll);
+	pdiv_only = cur_pll.base.n == n_int && cur_pll.sdm_din == sdm_din &&
+		    cur_pll.base.m == pll->m;
+
+	/* need full sequence if clock not enabled yet */
+	if (!gk20a_pllg_is_enabled(&clk->base))
+		pdiv_only = false;
+
+	/* split VCO-to-bypass jump in half by setting out divider 1:2 */
+	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK,
+		  GPC2CLK_OUT_VCODIV2 << GPC2CLK_OUT_VCODIV_SHIFT);
+	/* Intentional 2nd write to assure linear divider operation */
+	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK,
+		  GPC2CLK_OUT_VCODIV2 << GPC2CLK_OUT_VCODIV_SHIFT);
+	nvkm_rd32(device, GPC2CLK_OUT);
+	udelay(2);
+
+	if (pdiv_only) {
+		u32 old = cur_pll.base.pl;
+		u32 new = pll->pl;
+
+		/*
+		 * we can do a glitchless transition only if the old and new PL
+		 * parameters share at least one bit set to 1. If this is not
+		 * the case, calculate and program an interim PL that will allow
+		 * us to respect that rule.
+		 */
+		if ((old & new) == 0) {
+			cur_pll.base.pl = min(old | BIT(ffs(new) - 1),
+					      new | BIT(ffs(old) - 1));
+			gk20a_pllg_write_mnp(&clk->base, &cur_pll.base);
+		}
+
+		cur_pll.base.pl = new;
+		gk20a_pllg_write_mnp(&clk->base, &cur_pll.base);
+	} else {
+		/* disable before programming if more than pdiv changes */
+		gm20b_pllg_disable(clk);
+
+		cur_pll.base = *pll;
+		cur_pll.base.n = n_int;
+		cur_pll.sdm_din = sdm_din;
+		gm20b_pllg_write_mnp(clk, &cur_pll);
+
+		ret = gm20b_pllg_enable(clk);
+		if (ret)
+			return ret;
+	}
+
+	/* restore out divider 1:1 */
+	udelay(2);
+	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK,
+		  GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT);
+	/* Intentional 2nd write to assure linear divider operation */
+	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_VCODIV_MASK,
+		  GPC2CLK_OUT_VCODIV1 << GPC2CLK_OUT_VCODIV_SHIFT);
+	nvkm_rd32(device, GPC2CLK_OUT);
+
+	return 0;
+}
+
+static int
+gm20b_pllg_program_mnp_slide(struct gm20b_clk *clk, const struct gk20a_pll *pll)
+{
+	struct gk20a_pll cur_pll;
+	int ret;
+
+	if (gk20a_pllg_is_enabled(&clk->base)) {
+		gk20a_pllg_read_mnp(&clk->base, &cur_pll);
+
+		/* just do NDIV slide if there is no change to M and PL */
+		if (pll->m == cur_pll.m && pll->pl == cur_pll.pl)
+			return gm20b_pllg_slide(clk, pll->n);
+
+		/* slide down to current NDIV_LO */
+		cur_pll.n = gk20a_pllg_n_lo(&clk->base, &cur_pll);
+		ret = gm20b_pllg_slide(clk, cur_pll.n);
+		if (ret)
+			return ret;
+	}
+
+	/* program MNP with the new clock parameters and new NDIV_LO */
+	cur_pll = *pll;
+	cur_pll.n = gk20a_pllg_n_lo(&clk->base, &cur_pll);
+	ret = gm20b_pllg_program_mnp(clk, &cur_pll);
+	if (ret)
+		return ret;
+
+	/* slide up to new NDIV */
+	return gm20b_pllg_slide(clk, pll->n);
+}
+
+static int
+gm20b_clk_calc(struct nvkm_clk *base, struct nvkm_cstate *cstate)
+{
+	struct gm20b_clk *clk = gm20b_clk(base);
+	struct nvkm_subdev *subdev = &base->subdev;
+	struct nvkm_volt *volt = base->subdev.device->volt;
+	int ret;
+
+	ret = gk20a_pllg_calc_mnp(&clk->base, cstate->domain[nv_clk_src_gpc] *
+					     GK20A_CLK_GPC_MDIV, &clk->new_pll);
+	if (ret)
+		return ret;
+
+	clk->new_uv = volt->vid[cstate->voltage].uv;
+	gm20b_dvfs_calc_det_coeff(clk, clk->new_uv, &clk->new_dvfs);
+
+	nvkm_debug(subdev, "%s uv: %d uv\n", __func__, clk->new_uv);
+
+	return 0;
+}
+
+/*
+ * Compute PLL parameters that are always safe for the current voltage
+ */
+static void
+gm20b_dvfs_calc_safe_pll(struct gm20b_clk *clk, struct gk20a_pll *pll)
+{
+	u32 rate = gk20a_pllg_calc_rate(&clk->base, pll) / KHZ;
+	u32 parent_rate = clk->base.parent_rate / KHZ;
+	u32 nmin, nsafe;
+
+	/* remove a safe margin of 10% */
+	if (rate > clk->safe_fmax_vmin)
+		rate = rate * (100 - 10) / 100;
+
+	/* gpc2clk */
+	rate *= 2;
+
+	nmin = DIV_ROUND_UP(pll->m * clk->base.params->min_vco, parent_rate);
+	nsafe = pll->m * rate / (clk->base.parent_rate);
+
+	if (nsafe < nmin) {
+		pll->pl = DIV_ROUND_UP(nmin * parent_rate, pll->m * rate);
+		nsafe = nmin;
+	}
+
+	pll->n = nsafe;
+}
+
+static void
+gm20b_dvfs_program_coeff(struct gm20b_clk *clk, u32 coeff)
+{
+	struct nvkm_device *device = clk->base.base.subdev.device;
+
+	/* strobe to read external DFS coefficient */
+	nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2,
+		  GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT,
+		  GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT);
+
+	nvkm_mask(device, GPCPLL_DVFS0, GPCPLL_DVFS0_DFS_COEFF_MASK,
+		  coeff << GPCPLL_DVFS0_DFS_COEFF_SHIFT);
+
+	udelay(1);
+	nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2,
+		  GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT, 0);
+}
+
+static void
+gm20b_dvfs_program_ext_cal(struct gm20b_clk *clk, u32 dfs_det_cal)
+{
+	struct nvkm_device *device = clk->base.base.subdev.device;
+	u32 val;
+
+	nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2, MASK(DFS_DET_RANGE + 1),
+		  dfs_det_cal);
+	udelay(1);
+
+	val = nvkm_rd32(device, GPCPLL_DVFS1);
+	if (!(val & BIT(25))) {
+		/* Use external value to overwrite calibration value */
+		val |= BIT(25) | BIT(16);
+		nvkm_wr32(device, GPCPLL_DVFS1, val);
+	}
+}
+
+static void
+gm20b_dvfs_program_dfs_detection(struct gm20b_clk *clk,
+				 struct gm20b_clk_dvfs *dvfs)
+{
+	struct nvkm_device *device = clk->base.base.subdev.device;
+
+	/* strobe to read external DFS coefficient */
+	nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2,
+		  GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT,
+		  GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT);
+
+	nvkm_mask(device, GPCPLL_DVFS0,
+		  GPCPLL_DVFS0_DFS_COEFF_MASK | GPCPLL_DVFS0_DFS_DET_MAX_MASK,
+		  dvfs->dfs_coeff << GPCPLL_DVFS0_DFS_COEFF_SHIFT |
+		  dvfs->dfs_det_max << GPCPLL_DVFS0_DFS_DET_MAX_SHIFT);
+
+	udelay(1);
+	nvkm_mask(device, GPC_BCAST_GPCPLL_DVFS2,
+		  GPC_BCAST_GPCPLL_DVFS2_DFS_EXT_STROBE_BIT, 0);
+
+	gm20b_dvfs_program_ext_cal(clk, dvfs->dfs_ext_cal);
+}
+
+static int
+gm20b_clk_prog(struct nvkm_clk *base)
+{
+	struct gm20b_clk *clk = gm20b_clk(base);
+	u32 cur_freq;
+	int ret;
+
+	/* No change in DVFS settings? */
+	if (clk->uv == clk->new_uv)
+		goto prog;
+
+	/*
+	 * Interim step for changing DVFS detection settings: low enough
+	 * frequency to be safe at at DVFS coeff = 0.
+	 *
+	 * 1. If voltage is increasing:
+	 * - safe frequency target matches the lowest - old - frequency
+	 * - DVFS settings are still old
+	 * - Voltage already increased to new level by volt, but maximum
+	 *   detection limit assures PLL output remains under F/V curve
+	 *
+	 * 2. If voltage is decreasing:
+	 * - safe frequency target matches the lowest - new - frequency
+	 * - DVFS settings are still old
+	 * - Voltage is also old, it will be lowered by volt afterwards
+	 *
+	 * Interim step can be skipped if old frequency is below safe minimum,
+	 * i.e., it is low enough to be safe at any voltage in operating range
+	 * with zero DVFS coefficient.
+	 */
+	cur_freq = nvkm_clk_read(&clk->base.base, nv_clk_src_gpc);
+	if (cur_freq > clk->safe_fmax_vmin) {
+		struct gk20a_pll pll_safe;
+
+		if (clk->uv < clk->new_uv)
+			/* voltage will raise: safe frequency is current one */
+			pll_safe = clk->base.pll;
+		else
+			/* voltage will drop: safe frequency is new one */
+			pll_safe = clk->new_pll;
+
+		gm20b_dvfs_calc_safe_pll(clk, &pll_safe);
+		ret = gm20b_pllg_program_mnp_slide(clk, &pll_safe);
+		if (ret)
+			return ret;
+	}
+
+	/*
+	 * DVFS detection settings transition:
+	 * - Set DVFS coefficient zero
+	 * - Set calibration level to new voltage
+	 * - Set DVFS coefficient to match new voltage
+	 */
+	gm20b_dvfs_program_coeff(clk, 0);
+	gm20b_dvfs_program_ext_cal(clk, clk->new_dvfs.dfs_ext_cal);
+	gm20b_dvfs_program_coeff(clk, clk->new_dvfs.dfs_coeff);
+	gm20b_dvfs_program_dfs_detection(clk, &clk->new_dvfs);
+
+prog:
+	clk->uv = clk->new_uv;
+	clk->dvfs = clk->new_dvfs;
+	clk->base.pll = clk->new_pll;
+
+	return gm20b_pllg_program_mnp_slide(clk, &clk->base.pll);
+}
+
 static struct nvkm_pstate
 gm20b_pstates[] = {
 	{
@@ -133,9 +714,99 @@ gm20b_pstates[] = {
 			.voltage = 12,
 		},
 	},
-
 };
 
+static void
+gm20b_clk_fini(struct nvkm_clk *base)
+{
+	struct nvkm_device *device = base->subdev.device;
+	struct gm20b_clk *clk = gm20b_clk(base);
+
+	/* slide to VCO min */
+	if (gk20a_pllg_is_enabled(&clk->base)) {
+		struct gk20a_pll pll;
+		u32 n_lo;
+
+		gk20a_pllg_read_mnp(&clk->base, &pll);
+		n_lo = gk20a_pllg_n_lo(&clk->base, &pll);
+		gm20b_pllg_slide(clk, n_lo);
+	}
+
+	gm20b_pllg_disable(clk);
+
+	/* set IDDQ */
+	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_IDDQ, 1);
+}
+
+static int
+gm20b_clk_init_dvfs(struct gm20b_clk *clk)
+{
+	struct nvkm_subdev *subdev = &clk->base.base.subdev;
+	struct nvkm_device *device = subdev->device;
+	bool fused = clk->uvdet_offs && clk->uvdet_slope;
+	static const s32 ADC_SLOPE_UV = 10000; /* default ADC detection slope */
+	u32 data;
+	int ret;
+
+	/* Enable NA DVFS */
+	nvkm_mask(device, GPCPLL_DVFS1, GPCPLL_DVFS1_EN_DFS_BIT,
+		  GPCPLL_DVFS1_EN_DFS_BIT);
+
+	/* Set VCO_CTRL */
+	if (clk->dvfs_params->vco_ctrl)
+		nvkm_mask(device, GPCPLL_CFG3, GPCPLL_CFG3_VCO_CTRL_MASK,
+		      clk->dvfs_params->vco_ctrl << GPCPLL_CFG3_VCO_CTRL_SHIFT);
+
+	if (fused) {
+		/* Start internal calibration, but ignore results */
+		nvkm_mask(device, GPCPLL_DVFS1, GPCPLL_DVFS1_EN_DFS_CAL_BIT,
+			  GPCPLL_DVFS1_EN_DFS_CAL_BIT);
+
+		/* got uvdev parameters from fuse, skip calibration */
+		goto calibrated;
+	}
+
+	/*
+	 * If calibration parameters are not fused, start internal calibration,
+	 * wait for completion, and use results along with default slope to
+	 * calculate ADC offset during boot.
+	 */
+	nvkm_mask(device, GPCPLL_DVFS1, GPCPLL_DVFS1_EN_DFS_CAL_BIT,
+			  GPCPLL_DVFS1_EN_DFS_CAL_BIT);
+
+	/* Wait for internal calibration done (spec < 2us). */
+	ret = nvkm_wait_usec(device, 10, GPCPLL_DVFS1,
+			     GPCPLL_DVFS1_DFS_CAL_DONE_BIT,
+			     GPCPLL_DVFS1_DFS_CAL_DONE_BIT);
+	if (ret < 0) {
+		nvkm_error(subdev, "GPCPLL calibration timeout\n");
+		return -ETIMEDOUT;
+	}
+
+	data = nvkm_rd32(device, GPCPLL_CFG3) >>
+			 GPCPLL_CFG3_PLL_DFS_TESTOUT_SHIFT;
+	data &= MASK(GPCPLL_CFG3_PLL_DFS_TESTOUT_WIDTH);
+
+	clk->uvdet_slope = ADC_SLOPE_UV;
+	clk->uvdet_offs = ((s32)clk->uv) - data * ADC_SLOPE_UV;
+
+	nvkm_debug(subdev, "calibrated DVFS parameters: offs %d, slope %d\n",
+		   clk->uvdet_offs, clk->uvdet_slope);
+
+calibrated:
+	/* Compute and apply initial DVFS parameters */
+	gm20b_dvfs_calc_det_coeff(clk, clk->uv, &clk->dvfs);
+	gm20b_dvfs_program_coeff(clk, 0);
+	gm20b_dvfs_program_ext_cal(clk, clk->dvfs.dfs_ext_cal);
+	gm20b_dvfs_program_coeff(clk, clk->dvfs.dfs_coeff);
+	gm20b_dvfs_program_dfs_detection(clk, &clk->new_dvfs);
+
+	return 0;
+}
+
+/* Forward declaration to detect speedo >=1 in gm20b_clk_init() */
+static const struct nvkm_clk_func gm20b_clk;
+
 static int
 gm20b_clk_init(struct nvkm_clk *base)
 {
@@ -143,15 +814,56 @@ gm20b_clk_init(struct nvkm_clk *base)
 	struct nvkm_subdev *subdev = &clk->base.subdev;
 	struct nvkm_device *device = subdev->device;
 	int ret;
+	u32 data;
+
+	/* get out from IDDQ */
+	nvkm_mask(device, GPCPLL_CFG, GPCPLL_CFG_IDDQ, 0);
+	nvkm_rd32(device, GPCPLL_CFG);
+	udelay(5);
+
+	nvkm_mask(device, GPC2CLK_OUT, GPC2CLK_OUT_INIT_MASK,
+		  GPC2CLK_OUT_INIT_VAL);
 
 	/* Set the global bypass control to VCO */
 	nvkm_mask(device, BYPASSCTRL_SYS,
 	       MASK(BYPASSCTRL_SYS_GPCPLL_WIDTH) << BYPASSCTRL_SYS_GPCPLL_SHIFT,
 	       0);
 
+	ret = gk20a_clk_setup_slide(clk);
+	if (ret)
+		return ret;
+
+	/* If not fused, set RAM SVOP PDP data 0x2, and enable fuse override */
+	data = nvkm_rd32(device, 0x021944);
+	if (!(data & 0x3)) {
+		data |= 0x2;
+		nvkm_wr32(device, 0x021944, data);
+
+		data = nvkm_rd32(device, 0x021948);
+		data |=  0x1;
+		nvkm_wr32(device, 0x021948, data);
+	}
+
+	/* Disable idle slow down  */
+	nvkm_mask(device, 0x20160, 0x003f0000, 0x0);
+
+	/* speedo >= 1? */
+	if (clk->base.func == &gm20b_clk) {
+		struct gm20b_clk *_clk = gm20b_clk(base);
+		struct nvkm_volt *volt = device->volt;
+
+		/* Get current voltage */
+		_clk->uv = nvkm_volt_get(volt);
+
+		/* Initialize DVFS */
+		ret = gm20b_clk_init_dvfs(_clk);
+		if (ret)
+			return ret;
+	}
+
 	/* Start with lowest frequency */
 	base->func->calc(base, &base->func->pstates[0].base);
-	ret = base->func->prog(&clk->base);
+	ret = base->func->prog(base);
 	if (ret) {
 		nvkm_error(subdev, "cannot initialize clock\n");
 		return ret;
@@ -169,6 +881,7 @@ gm20b_clk_speedo0 = {
 	.prog = gk20a_clk_prog,
 	.tidy = gk20a_clk_tidy,
 	.pstates = gm20b_pstates,
+	/* Speedo 0 only supports 12 voltages */
 	.nr_pstates = ARRAY_SIZE(gm20b_pstates) - 1,
 	.domains = {
 		{ nv_clk_src_crystal, 0xff },
@@ -177,8 +890,26 @@ gm20b_clk_speedo0 = {
 	},
 };
 
-int
-gm20b_clk_new(struct nvkm_device *device, int index, struct nvkm_clk **pclk)
+static const struct nvkm_clk_func
+gm20b_clk = {
+	.init = gm20b_clk_init,
+	.fini = gm20b_clk_fini,
+	.read = gk20a_clk_read,
+	.calc = gm20b_clk_calc,
+	.prog = gm20b_clk_prog,
+	.tidy = gk20a_clk_tidy,
+	.pstates = gm20b_pstates,
+	.nr_pstates = ARRAY_SIZE(gm20b_pstates),
+	.domains = {
+		{ nv_clk_src_crystal, 0xff },
+		{ nv_clk_src_gpc, 0xff, 0, "core", GK20A_CLK_GPC_MDIV },
+		{ nv_clk_src_max },
+	},
+};
+
+static int
+gm20b_clk_new_speedo0(struct nvkm_device *device, int index,
+		      struct nvkm_clk **pclk)
 {
 	struct gk20a_clk *clk;
 	int ret;
@@ -188,11 +919,156 @@ gm20b_clk_new(struct nvkm_device *device, int index, struct nvkm_clk **pclk)
 		return -ENOMEM;
 	*pclk = &clk->base;
 
-	ret = _gk20a_clk_ctor(device, index, &gm20b_clk_speedo0,
-			      &gm20b_pllg_params, clk);
+	ret = gk20a_clk_ctor(device, index, &gm20b_clk_speedo0,
+			     &gm20b_pllg_params, clk);
 
 	clk->pl_to_div = pl_to_div;
 	clk->div_to_pl = div_to_pl;
 
 	return ret;
 }
+
+/* FUSE register */
+#define FUSE_RESERVED_CALIB0	0x204
+#define FUSE_RESERVED_CALIB0_INTERCEPT_FRAC_SHIFT	0
+#define FUSE_RESERVED_CALIB0_INTERCEPT_FRAC_WIDTH	4
+#define FUSE_RESERVED_CALIB0_INTERCEPT_INT_SHIFT	4
+#define FUSE_RESERVED_CALIB0_INTERCEPT_INT_WIDTH	10
+#define FUSE_RESERVED_CALIB0_SLOPE_FRAC_SHIFT		14
+#define FUSE_RESERVED_CALIB0_SLOPE_FRAC_WIDTH		10
+#define FUSE_RESERVED_CALIB0_SLOPE_INT_SHIFT		24
+#define FUSE_RESERVED_CALIB0_SLOPE_INT_WIDTH		6
+#define FUSE_RESERVED_CALIB0_FUSE_REV_SHIFT		30
+#define FUSE_RESERVED_CALIB0_FUSE_REV_WIDTH		2
+
+static int
+gm20b_clk_init_fused_params(struct gm20b_clk *clk)
+{
+	struct nvkm_subdev *subdev = &clk->base.base.subdev;
+	u32 val = 0;
+	u32 rev = 0;
+
+#if IS_ENABLED(CONFIG_ARCH_TEGRA)
+	tegra_fuse_readl(FUSE_RESERVED_CALIB0, &val);
+	rev = (val >> FUSE_RESERVED_CALIB0_FUSE_REV_SHIFT) &
+	      MASK(FUSE_RESERVED_CALIB0_FUSE_REV_WIDTH);
+#endif
+
+	/* No fused parameters, we will calibrate later */
+	if (rev == 0)
+		return -EINVAL;
+
+	/* Integer part in mV + fractional part in uV */
+	clk->uvdet_slope = ((val >> FUSE_RESERVED_CALIB0_SLOPE_INT_SHIFT) &
+			MASK(FUSE_RESERVED_CALIB0_SLOPE_INT_WIDTH)) * 1000 +
+			((val >> FUSE_RESERVED_CALIB0_SLOPE_FRAC_SHIFT) &
+			MASK(FUSE_RESERVED_CALIB0_SLOPE_FRAC_WIDTH));
+
+	/* Integer part in mV + fractional part in 100uV */
+	clk->uvdet_offs = ((val >> FUSE_RESERVED_CALIB0_INTERCEPT_INT_SHIFT) &
+			MASK(FUSE_RESERVED_CALIB0_INTERCEPT_INT_WIDTH)) * 1000 +
+			((val >> FUSE_RESERVED_CALIB0_INTERCEPT_FRAC_SHIFT) &
+			 MASK(FUSE_RESERVED_CALIB0_INTERCEPT_FRAC_WIDTH)) * 100;
+
+	nvkm_debug(subdev, "fused calibration data: slope %d, offs %d\n",
+		   clk->uvdet_slope, clk->uvdet_offs);
+	return 0;
+}
+
+static int
+gm20b_clk_init_safe_fmax(struct gm20b_clk *clk)
+{
+	struct nvkm_subdev *subdev = &clk->base.base.subdev;
+	struct nvkm_volt *volt = subdev->device->volt;
+	struct nvkm_pstate *pstates = clk->base.base.func->pstates;
+	int nr_pstates = clk->base.base.func->nr_pstates;
+	int vmin, id = 0;
+	u32 fmax = 0;
+	int i;
+
+	/* find lowest voltage we can use */
+	vmin = volt->vid[0].uv;
+	for (i = 1; i < volt->vid_nr; i++) {
+		if (volt->vid[i].uv <= vmin) {
+			vmin = volt->vid[i].uv;
+			id = volt->vid[i].vid;
+		}
+	}
+
+	/* find max frequency at this voltage */
+	for (i = 0; i < nr_pstates; i++)
+		if (pstates[i].base.voltage == id)
+			fmax = max(fmax,
+				   pstates[i].base.domain[nv_clk_src_gpc]);
+
+	if (!fmax) {
+		nvkm_error(subdev, "failed to evaluate safe fmax\n");
+		return -EINVAL;
+	}
+
+	/* we are safe at 90% of the max frequency */
+	clk->safe_fmax_vmin = fmax * (100 - 10) / 100;
+	nvkm_debug(subdev, "safe fmax @ vmin = %u Khz\n", clk->safe_fmax_vmin);
+
+	return 0;
+}
+
+int
+gm20b_clk_new(struct nvkm_device *device, int index, struct nvkm_clk **pclk)
+{
+	struct nvkm_device_tegra *tdev = device->func->tegra(device);
+	struct gm20b_clk *clk;
+	struct nvkm_subdev *subdev;
+	struct gk20a_clk_pllg_params *clk_params;
+	int ret;
+
+	/* Speedo 0 GPUs cannot use noise-aware PLL */
+	if (tdev->gpu_speedo_id == 0)
+		return gm20b_clk_new_speedo0(device, index, pclk);
+
+	/* Speedo >= 1, use NAPLL */
+	clk = kzalloc(sizeof(*clk) + sizeof(*clk_params), GFP_KERNEL);
+	if (!clk)
+		return -ENOMEM;
+	*pclk = &clk->base.base;
+	subdev = &clk->base.base.subdev;
+
+	/* duplicate the clock parameters since we will patch them below */
+	clk_params = (void *) (clk + 1);
+	*clk_params = gm20b_pllg_params;
+	ret = gk20a_clk_ctor(device, index, &gm20b_clk, clk_params,
+			     &clk->base);
+	if (ret)
+		return ret;
+
+	/*
+	 * NAPLL can only work with max_u, clamp the m range so
+	 * gk20a_pllg_calc_mnp always uses it
+	 */
+	clk_params->max_m = clk_params->min_m = DIV_ROUND_UP(clk_params->max_u,
+						(clk->base.parent_rate / KHZ));
+	if (clk_params->max_m == 0) {
+		nvkm_warn(subdev, "cannot use NAPLL, using legacy clock...\n");
+		kfree(clk);
+		return gm20b_clk_new_speedo0(device, index, pclk);
+	}
+
+	clk->base.pl_to_div = pl_to_div;
+	clk->base.div_to_pl = div_to_pl;
+
+	clk->dvfs_params = &gm20b_dvfs_params;
+
+	ret = gm20b_clk_init_fused_params(clk);
+	/*
+	 * we will calibrate during init - should never happen on
+	 * prod parts
+	 */
+	if (ret)
+		nvkm_warn(subdev, "no fused calibration parameters\n");
+
+	ret = gm20b_clk_init_safe_fmax(clk);
+	if (ret)
+		return ret;
+
+	return 0;
+}