1 files changed, 881 insertions, 11 deletions

diff --git a/drm/nouveau/nvkm/subdev/clk/gm20b.c b/drm/nouveau/nvkm/subdev/clk/gm20b.c
index 68749bbbe..3089e4bc5 100644
--- a/drm/nouveau/nvkm/subdev/clk/gm20b.c
+++ b/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)
 {
@@ -143,19 +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;
 
-	ret = gk20a_clk_setup_slide(clk);
-	if (ret)
-		return 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);
 
 	/* 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;
@@ -173,6 +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 },
@@ -181,8 +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;
@@ -200,3 +927,146 @@ gm20b_clk_new(struct nvkm_device *device, int index, struct nvkm_clk **pclk)
 
 	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;
+	u32 rev;
+
+	tegra_fuse_readl(FUSE_RESERVED_CALIB0, &val);
+	rev = (val >> FUSE_RESERVED_CALIB0_FUSE_REV_SHIFT) &
+	      MASK(FUSE_RESERVED_CALIB0_FUSE_REV_WIDTH);
+
+	/* 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;
+}