Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   1/*
   2 * Copyright © 2012 Intel Corporation
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice (including the next
  12 * paragraph) shall be included in all copies or substantial portions of the
  13 * Software.
  14 *
  15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21 * IN THE SOFTWARE.
  22 *
  23 * Authors:
  24 *    Eugeni Dodonov <eugeni.dodonov@intel.com>
  25 *
  26 */
  27
  28#include <linux/cpufreq.h>
  29#include <drm/drm_plane_helper.h>
  30#include "i915_drv.h"
  31#include "intel_drv.h"
  32#include "../../../platform/x86/intel_ips.h"
  33#include <linux/module.h>
  34#include <drm/drm_atomic_helper.h>
  35
  36/**
  37 * DOC: RC6
  38 *
  39 * RC6 is a special power stage which allows the GPU to enter an very
  40 * low-voltage mode when idle, using down to 0V while at this stage.  This
  41 * stage is entered automatically when the GPU is idle when RC6 support is
  42 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
  43 *
  44 * There are different RC6 modes available in Intel GPU, which differentiate
  45 * among each other with the latency required to enter and leave RC6 and
  46 * voltage consumed by the GPU in different states.
  47 *
  48 * The combination of the following flags define which states GPU is allowed
  49 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
  50 * RC6pp is deepest RC6. Their support by hardware varies according to the
  51 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
  52 * which brings the most power savings; deeper states save more power, but
  53 * require higher latency to switch to and wake up.
  54 */
  55
  56static void gen9_init_clock_gating(struct drm_i915_private *dev_priv)
  57{
  58	if (HAS_LLC(dev_priv)) {
  59		/*
  60		 * WaCompressedResourceDisplayNewHashMode:skl,kbl
  61		 * Display WA #0390: skl,kbl
  62		 *
  63		 * Must match Sampler, Pixel Back End, and Media. See
  64		 * WaCompressedResourceSamplerPbeMediaNewHashMode.
  65		 */
  66		I915_WRITE(CHICKEN_PAR1_1,
  67			   I915_READ(CHICKEN_PAR1_1) |
  68			   SKL_DE_COMPRESSED_HASH_MODE);
  69	}
  70
  71	/* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl,cfl */
  72	I915_WRITE(CHICKEN_PAR1_1,
  73		   I915_READ(CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
  74
  75	/* WaEnableChickenDCPR:skl,bxt,kbl,glk,cfl */
  76	I915_WRITE(GEN8_CHICKEN_DCPR_1,
  77		   I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
  78
  79	/* WaFbcTurnOffFbcWatermark:skl,bxt,kbl,cfl */
  80	/* WaFbcWakeMemOn:skl,bxt,kbl,glk,cfl */
  81	I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
  82		   DISP_FBC_WM_DIS |
  83		   DISP_FBC_MEMORY_WAKE);
  84
  85	/* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl,cfl */
  86	I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
  87		   ILK_DPFC_DISABLE_DUMMY0);
  88
  89	if (IS_SKYLAKE(dev_priv)) {
  90		/* WaDisableDopClockGating */
  91		I915_WRITE(GEN7_MISCCPCTL, I915_READ(GEN7_MISCCPCTL)
  92			   & ~GEN7_DOP_CLOCK_GATE_ENABLE);
  93	}
  94}
  95
  96static void bxt_init_clock_gating(struct drm_i915_private *dev_priv)
  97{
  98	gen9_init_clock_gating(dev_priv);
  99
 100	/* WaDisableSDEUnitClockGating:bxt */
 101	I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
 102		   GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
 103
 104	/*
 105	 * FIXME:
 106	 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
 107	 */
 108	I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
 109		   GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
 110
 111	/*
 112	 * Wa: Backlight PWM may stop in the asserted state, causing backlight
 113	 * to stay fully on.
 114	 */
 115	I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
 116		   PWM1_GATING_DIS | PWM2_GATING_DIS);
 117}
 118
 119static void glk_init_clock_gating(struct drm_i915_private *dev_priv)
 120{
 121	gen9_init_clock_gating(dev_priv);
 122
 123	/*
 124	 * WaDisablePWMClockGating:glk
 125	 * Backlight PWM may stop in the asserted state, causing backlight
 126	 * to stay fully on.
 127	 */
 128	I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
 129		   PWM1_GATING_DIS | PWM2_GATING_DIS);
 130
 131	/* WaDDIIOTimeout:glk */
 132	if (IS_GLK_REVID(dev_priv, 0, GLK_REVID_A1)) {
 133		u32 val = I915_READ(CHICKEN_MISC_2);
 134		val &= ~(GLK_CL0_PWR_DOWN |
 135			 GLK_CL1_PWR_DOWN |
 136			 GLK_CL2_PWR_DOWN);
 137		I915_WRITE(CHICKEN_MISC_2, val);
 138	}
 139
 140}
 141
 142static void i915_pineview_get_mem_freq(struct drm_i915_private *dev_priv)
 143{
 144	u32 tmp;
 145
 146	tmp = I915_READ(CLKCFG);
 147
 148	switch (tmp & CLKCFG_FSB_MASK) {
 149	case CLKCFG_FSB_533:
 150		dev_priv->fsb_freq = 533; /* 133*4 */
 151		break;
 152	case CLKCFG_FSB_800:
 153		dev_priv->fsb_freq = 800; /* 200*4 */
 154		break;
 155	case CLKCFG_FSB_667:
 156		dev_priv->fsb_freq =  667; /* 167*4 */
 157		break;
 158	case CLKCFG_FSB_400:
 159		dev_priv->fsb_freq = 400; /* 100*4 */
 160		break;
 161	}
 162
 163	switch (tmp & CLKCFG_MEM_MASK) {
 164	case CLKCFG_MEM_533:
 165		dev_priv->mem_freq = 533;
 166		break;
 167	case CLKCFG_MEM_667:
 168		dev_priv->mem_freq = 667;
 169		break;
 170	case CLKCFG_MEM_800:
 171		dev_priv->mem_freq = 800;
 172		break;
 173	}
 174
 175	/* detect pineview DDR3 setting */
 176	tmp = I915_READ(CSHRDDR3CTL);
 177	dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 178}
 179
 180static void i915_ironlake_get_mem_freq(struct drm_i915_private *dev_priv)
 181{
 182	u16 ddrpll, csipll;
 183
 184	ddrpll = I915_READ16(DDRMPLL1);
 185	csipll = I915_READ16(CSIPLL0);
 186
 187	switch (ddrpll & 0xff) {
 188	case 0xc:
 189		dev_priv->mem_freq = 800;
 190		break;
 191	case 0x10:
 192		dev_priv->mem_freq = 1066;
 193		break;
 194	case 0x14:
 195		dev_priv->mem_freq = 1333;
 196		break;
 197	case 0x18:
 198		dev_priv->mem_freq = 1600;
 199		break;
 200	default:
 201		DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 202				 ddrpll & 0xff);
 203		dev_priv->mem_freq = 0;
 204		break;
 205	}
 206
 207	dev_priv->ips.r_t = dev_priv->mem_freq;
 208
 209	switch (csipll & 0x3ff) {
 210	case 0x00c:
 211		dev_priv->fsb_freq = 3200;
 212		break;
 213	case 0x00e:
 214		dev_priv->fsb_freq = 3733;
 215		break;
 216	case 0x010:
 217		dev_priv->fsb_freq = 4266;
 218		break;
 219	case 0x012:
 220		dev_priv->fsb_freq = 4800;
 221		break;
 222	case 0x014:
 223		dev_priv->fsb_freq = 5333;
 224		break;
 225	case 0x016:
 226		dev_priv->fsb_freq = 5866;
 227		break;
 228	case 0x018:
 229		dev_priv->fsb_freq = 6400;
 230		break;
 231	default:
 232		DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 233				 csipll & 0x3ff);
 234		dev_priv->fsb_freq = 0;
 235		break;
 236	}
 237
 238	if (dev_priv->fsb_freq == 3200) {
 239		dev_priv->ips.c_m = 0;
 240	} else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 241		dev_priv->ips.c_m = 1;
 242	} else {
 243		dev_priv->ips.c_m = 2;
 244	}
 245}
 246
 247static const struct cxsr_latency cxsr_latency_table[] = {
 248	{1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 249	{1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 250	{1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 251	{1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 252	{1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 253
 254	{1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 255	{1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 256	{1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 257	{1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 258	{1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 259
 260	{1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 261	{1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 262	{1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 263	{1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 264	{1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 265
 266	{0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 267	{0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 268	{0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 269	{0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 270	{0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 271
 272	{0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 273	{0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 274	{0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 275	{0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 276	{0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 277
 278	{0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 279	{0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 280	{0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 281	{0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 282	{0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 283};
 284
 285static const struct cxsr_latency *intel_get_cxsr_latency(bool is_desktop,
 286							 bool is_ddr3,
 287							 int fsb,
 288							 int mem)
 289{
 290	const struct cxsr_latency *latency;
 291	int i;
 292
 293	if (fsb == 0 || mem == 0)
 294		return NULL;
 295
 296	for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 297		latency = &cxsr_latency_table[i];
 298		if (is_desktop == latency->is_desktop &&
 299		    is_ddr3 == latency->is_ddr3 &&
 300		    fsb == latency->fsb_freq && mem == latency->mem_freq)
 301			return latency;
 302	}
 303
 304	DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 305
 306	return NULL;
 307}
 308
 309static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
 310{
 311	u32 val;
 312
 313	mutex_lock(&dev_priv->pcu_lock);
 314
 315	val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
 316	if (enable)
 317		val &= ~FORCE_DDR_HIGH_FREQ;
 318	else
 319		val |= FORCE_DDR_HIGH_FREQ;
 320	val &= ~FORCE_DDR_LOW_FREQ;
 321	val |= FORCE_DDR_FREQ_REQ_ACK;
 322	vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
 323
 324	if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
 325		      FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
 326		DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
 327
 328	mutex_unlock(&dev_priv->pcu_lock);
 329}
 330
 331static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
 332{
 333	u32 val;
 334
 335	mutex_lock(&dev_priv->pcu_lock);
 336
 337	val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
 338	if (enable)
 339		val |= DSP_MAXFIFO_PM5_ENABLE;
 340	else
 341		val &= ~DSP_MAXFIFO_PM5_ENABLE;
 342	vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
 343
 344	mutex_unlock(&dev_priv->pcu_lock);
 345}
 346
 347#define FW_WM(value, plane) \
 348	(((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
 349
 350static bool _intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 351{
 352	bool was_enabled;
 353	u32 val;
 354
 355	if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
 356		was_enabled = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
 357		I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
 358		POSTING_READ(FW_BLC_SELF_VLV);
 359	} else if (IS_G4X(dev_priv) || IS_I965GM(dev_priv)) {
 360		was_enabled = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
 361		I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
 362		POSTING_READ(FW_BLC_SELF);
 363	} else if (IS_PINEVIEW(dev_priv)) {
 364		val = I915_READ(DSPFW3);
 365		was_enabled = val & PINEVIEW_SELF_REFRESH_EN;
 366		if (enable)
 367			val |= PINEVIEW_SELF_REFRESH_EN;
 368		else
 369			val &= ~PINEVIEW_SELF_REFRESH_EN;
 370		I915_WRITE(DSPFW3, val);
 371		POSTING_READ(DSPFW3);
 372	} else if (IS_I945G(dev_priv) || IS_I945GM(dev_priv)) {
 373		was_enabled = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
 374		val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
 375			       _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
 376		I915_WRITE(FW_BLC_SELF, val);
 377		POSTING_READ(FW_BLC_SELF);
 378	} else if (IS_I915GM(dev_priv)) {
 379		/*
 380		 * FIXME can't find a bit like this for 915G, and
 381		 * and yet it does have the related watermark in
 382		 * FW_BLC_SELF. What's going on?
 383		 */
 384		was_enabled = I915_READ(INSTPM) & INSTPM_SELF_EN;
 385		val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
 386			       _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
 387		I915_WRITE(INSTPM, val);
 388		POSTING_READ(INSTPM);
 389	} else {
 390		return false;
 391	}
 392
 393	trace_intel_memory_cxsr(dev_priv, was_enabled, enable);
 394
 395	DRM_DEBUG_KMS("memory self-refresh is %s (was %s)\n",
 396		      enableddisabled(enable),
 397		      enableddisabled(was_enabled));
 398
 399	return was_enabled;
 400}
 401
 402/**
 403 * intel_set_memory_cxsr - Configure CxSR state
 404 * @dev_priv: i915 device
 405 * @enable: Allow vs. disallow CxSR
 406 *
 407 * Allow or disallow the system to enter a special CxSR
 408 * (C-state self refresh) state. What typically happens in CxSR mode
 409 * is that several display FIFOs may get combined into a single larger
 410 * FIFO for a particular plane (so called max FIFO mode) to allow the
 411 * system to defer memory fetches longer, and the memory will enter
 412 * self refresh.
 413 *
 414 * Note that enabling CxSR does not guarantee that the system enter
 415 * this special mode, nor does it guarantee that the system stays
 416 * in that mode once entered. So this just allows/disallows the system
 417 * to autonomously utilize the CxSR mode. Other factors such as core
 418 * C-states will affect when/if the system actually enters/exits the
 419 * CxSR mode.
 420 *
 421 * Note that on VLV/CHV this actually only controls the max FIFO mode,
 422 * and the system is free to enter/exit memory self refresh at any time
 423 * even when the use of CxSR has been disallowed.
 424 *
 425 * While the system is actually in the CxSR/max FIFO mode, some plane
 426 * control registers will not get latched on vblank. Thus in order to
 427 * guarantee the system will respond to changes in the plane registers
 428 * we must always disallow CxSR prior to making changes to those registers.
 429 * Unfortunately the system will re-evaluate the CxSR conditions at
 430 * frame start which happens after vblank start (which is when the plane
 431 * registers would get latched), so we can't proceed with the plane update
 432 * during the same frame where we disallowed CxSR.
 433 *
 434 * Certain platforms also have a deeper HPLL SR mode. Fortunately the
 435 * HPLL SR mode depends on CxSR itself, so we don't have to hand hold
 436 * the hardware w.r.t. HPLL SR when writing to plane registers.
 437 * Disallowing just CxSR is sufficient.
 438 */
 439bool intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 440{
 441	bool ret;
 442
 443	mutex_lock(&dev_priv->wm.wm_mutex);
 444	ret = _intel_set_memory_cxsr(dev_priv, enable);
 445	if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
 446		dev_priv->wm.vlv.cxsr = enable;
 447	else if (IS_G4X(dev_priv))
 448		dev_priv->wm.g4x.cxsr = enable;
 449	mutex_unlock(&dev_priv->wm.wm_mutex);
 450
 451	return ret;
 452}
 453
 454/*
 455 * Latency for FIFO fetches is dependent on several factors:
 456 *   - memory configuration (speed, channels)
 457 *   - chipset
 458 *   - current MCH state
 459 * It can be fairly high in some situations, so here we assume a fairly
 460 * pessimal value.  It's a tradeoff between extra memory fetches (if we
 461 * set this value too high, the FIFO will fetch frequently to stay full)
 462 * and power consumption (set it too low to save power and we might see
 463 * FIFO underruns and display "flicker").
 464 *
 465 * A value of 5us seems to be a good balance; safe for very low end
 466 * platforms but not overly aggressive on lower latency configs.
 467 */
 468static const int pessimal_latency_ns = 5000;
 469
 470#define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
 471	((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
 472
 473static void vlv_get_fifo_size(struct intel_crtc_state *crtc_state)
 474{
 475	struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
 476	struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
 477	struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
 478	enum pipe pipe = crtc->pipe;
 479	int sprite0_start, sprite1_start;
 480
 481	switch (pipe) {
 482		uint32_t dsparb, dsparb2, dsparb3;
 483	case PIPE_A:
 484		dsparb = I915_READ(DSPARB);
 485		dsparb2 = I915_READ(DSPARB2);
 486		sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
 487		sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
 488		break;
 489	case PIPE_B:
 490		dsparb = I915_READ(DSPARB);
 491		dsparb2 = I915_READ(DSPARB2);
 492		sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
 493		sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
 494		break;
 495	case PIPE_C:
 496		dsparb2 = I915_READ(DSPARB2);
 497		dsparb3 = I915_READ(DSPARB3);
 498		sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
 499		sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
 500		break;
 501	default:
 502		MISSING_CASE(pipe);
 503		return;
 504	}
 505
 506	fifo_state->plane[PLANE_PRIMARY] = sprite0_start;
 507	fifo_state->plane[PLANE_SPRITE0] = sprite1_start - sprite0_start;
 508	fifo_state->plane[PLANE_SPRITE1] = 511 - sprite1_start;
 509	fifo_state->plane[PLANE_CURSOR] = 63;
 510}
 511
 512static int i9xx_get_fifo_size(struct drm_i915_private *dev_priv,
 513			      enum i9xx_plane_id i9xx_plane)
 514{
 515	uint32_t dsparb = I915_READ(DSPARB);
 516	int size;
 517
 518	size = dsparb & 0x7f;
 519	if (i9xx_plane == PLANE_B)
 520		size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 521
 522	DRM_DEBUG_KMS("FIFO size - (0x%08x) %c: %d\n",
 523		      dsparb, plane_name(i9xx_plane), size);
 524
 525	return size;
 526}
 527
 528static int i830_get_fifo_size(struct drm_i915_private *dev_priv,
 529			      enum i9xx_plane_id i9xx_plane)
 530{
 531	uint32_t dsparb = I915_READ(DSPARB);
 532	int size;
 533
 534	size = dsparb & 0x1ff;
 535	if (i9xx_plane == PLANE_B)
 536		size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 537	size >>= 1; /* Convert to cachelines */
 538
 539	DRM_DEBUG_KMS("FIFO size - (0x%08x) %c: %d\n",
 540		      dsparb, plane_name(i9xx_plane), size);
 541
 542	return size;
 543}
 544
 545static int i845_get_fifo_size(struct drm_i915_private *dev_priv,
 546			      enum i9xx_plane_id i9xx_plane)
 547{
 548	uint32_t dsparb = I915_READ(DSPARB);
 549	int size;
 550
 551	size = dsparb & 0x7f;
 552	size >>= 2; /* Convert to cachelines */
 553
 554	DRM_DEBUG_KMS("FIFO size - (0x%08x) %c: %d\n",
 555		      dsparb, plane_name(i9xx_plane), size);
 556
 557	return size;
 558}
 559
 560/* Pineview has different values for various configs */
 561static const struct intel_watermark_params pineview_display_wm = {
 562	.fifo_size = PINEVIEW_DISPLAY_FIFO,
 563	.max_wm = PINEVIEW_MAX_WM,
 564	.default_wm = PINEVIEW_DFT_WM,
 565	.guard_size = PINEVIEW_GUARD_WM,
 566	.cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 567};
 568static const struct intel_watermark_params pineview_display_hplloff_wm = {
 569	.fifo_size = PINEVIEW_DISPLAY_FIFO,
 570	.max_wm = PINEVIEW_MAX_WM,
 571	.default_wm = PINEVIEW_DFT_HPLLOFF_WM,
 572	.guard_size = PINEVIEW_GUARD_WM,
 573	.cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 574};
 575static const struct intel_watermark_params pineview_cursor_wm = {
 576	.fifo_size = PINEVIEW_CURSOR_FIFO,
 577	.max_wm = PINEVIEW_CURSOR_MAX_WM,
 578	.default_wm = PINEVIEW_CURSOR_DFT_WM,
 579	.guard_size = PINEVIEW_CURSOR_GUARD_WM,
 580	.cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 581};
 582static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 583	.fifo_size = PINEVIEW_CURSOR_FIFO,
 584	.max_wm = PINEVIEW_CURSOR_MAX_WM,
 585	.default_wm = PINEVIEW_CURSOR_DFT_WM,
 586	.guard_size = PINEVIEW_CURSOR_GUARD_WM,
 587	.cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 588};
 589static const struct intel_watermark_params i965_cursor_wm_info = {
 590	.fifo_size = I965_CURSOR_FIFO,
 591	.max_wm = I965_CURSOR_MAX_WM,
 592	.default_wm = I965_CURSOR_DFT_WM,
 593	.guard_size = 2,
 594	.cacheline_size = I915_FIFO_LINE_SIZE,
 595};
 596static const struct intel_watermark_params i945_wm_info = {
 597	.fifo_size = I945_FIFO_SIZE,
 598	.max_wm = I915_MAX_WM,
 599	.default_wm = 1,
 600	.guard_size = 2,
 601	.cacheline_size = I915_FIFO_LINE_SIZE,
 602};
 603static const struct intel_watermark_params i915_wm_info = {
 604	.fifo_size = I915_FIFO_SIZE,
 605	.max_wm = I915_MAX_WM,
 606	.default_wm = 1,
 607	.guard_size = 2,
 608	.cacheline_size = I915_FIFO_LINE_SIZE,
 609};
 610static const struct intel_watermark_params i830_a_wm_info = {
 611	.fifo_size = I855GM_FIFO_SIZE,
 612	.max_wm = I915_MAX_WM,
 613	.default_wm = 1,
 614	.guard_size = 2,
 615	.cacheline_size = I830_FIFO_LINE_SIZE,
 616};
 617static const struct intel_watermark_params i830_bc_wm_info = {
 618	.fifo_size = I855GM_FIFO_SIZE,
 619	.max_wm = I915_MAX_WM/2,
 620	.default_wm = 1,
 621	.guard_size = 2,
 622	.cacheline_size = I830_FIFO_LINE_SIZE,
 623};
 624static const struct intel_watermark_params i845_wm_info = {
 625	.fifo_size = I830_FIFO_SIZE,
 626	.max_wm = I915_MAX_WM,
 627	.default_wm = 1,
 628	.guard_size = 2,
 629	.cacheline_size = I830_FIFO_LINE_SIZE,
 630};
 631
 632/**
 633 * intel_wm_method1 - Method 1 / "small buffer" watermark formula
 634 * @pixel_rate: Pipe pixel rate in kHz
 635 * @cpp: Plane bytes per pixel
 636 * @latency: Memory wakeup latency in 0.1us units
 637 *
 638 * Compute the watermark using the method 1 or "small buffer"
 639 * formula. The caller may additonally add extra cachelines
 640 * to account for TLB misses and clock crossings.
 641 *
 642 * This method is concerned with the short term drain rate
 643 * of the FIFO, ie. it does not account for blanking periods
 644 * which would effectively reduce the average drain rate across
 645 * a longer period. The name "small" refers to the fact the
 646 * FIFO is relatively small compared to the amount of data
 647 * fetched.
 648 *
 649 * The FIFO level vs. time graph might look something like:
 650 *
 651 *   |\   |\
 652 *   | \  | \
 653 * __---__---__ (- plane active, _ blanking)
 654 * -> time
 655 *
 656 * or perhaps like this:
 657 *
 658 *   |\|\  |\|\
 659 * __----__----__ (- plane active, _ blanking)
 660 * -> time
 661 *
 662 * Returns:
 663 * The watermark in bytes
 664 */
 665static unsigned int intel_wm_method1(unsigned int pixel_rate,
 666				     unsigned int cpp,
 667				     unsigned int latency)
 668{
 669	uint64_t ret;
 670
 671	ret = (uint64_t) pixel_rate * cpp * latency;
 672	ret = DIV_ROUND_UP_ULL(ret, 10000);
 673
 674	return ret;
 675}
 676
 677/**
 678 * intel_wm_method2 - Method 2 / "large buffer" watermark formula
 679 * @pixel_rate: Pipe pixel rate in kHz
 680 * @htotal: Pipe horizontal total
 681 * @width: Plane width in pixels
 682 * @cpp: Plane bytes per pixel
 683 * @latency: Memory wakeup latency in 0.1us units
 684 *
 685 * Compute the watermark using the method 2 or "large buffer"
 686 * formula. The caller may additonally add extra cachelines
 687 * to account for TLB misses and clock crossings.
 688 *
 689 * This method is concerned with the long term drain rate
 690 * of the FIFO, ie. it does account for blanking periods
 691 * which effectively reduce the average drain rate across
 692 * a longer period. The name "large" refers to the fact the
 693 * FIFO is relatively large compared to the amount of data
 694 * fetched.
 695 *
 696 * The FIFO level vs. time graph might look something like:
 697 *
 698 *    |\___       |\___
 699 *    |    \___   |    \___
 700 *    |        \  |        \
 701 * __ --__--__--__--__--__--__ (- plane active, _ blanking)
 702 * -> time
 703 *
 704 * Returns:
 705 * The watermark in bytes
 706 */
 707static unsigned int intel_wm_method2(unsigned int pixel_rate,
 708				     unsigned int htotal,
 709				     unsigned int width,
 710				     unsigned int cpp,
 711				     unsigned int latency)
 712{
 713	unsigned int ret;
 714
 715	/*
 716	 * FIXME remove once all users are computing
 717	 * watermarks in the correct place.
 718	 */
 719	if (WARN_ON_ONCE(htotal == 0))
 720		htotal = 1;
 721
 722	ret = (latency * pixel_rate) / (htotal * 10000);
 723	ret = (ret + 1) * width * cpp;
 724
 725	return ret;
 726}
 727
 728/**
 729 * intel_calculate_wm - calculate watermark level
 730 * @pixel_rate: pixel clock
 731 * @wm: chip FIFO params
 732 * @fifo_size: size of the FIFO buffer
 733 * @cpp: bytes per pixel
 734 * @latency_ns: memory latency for the platform
 735 *
 736 * Calculate the watermark level (the level at which the display plane will
 737 * start fetching from memory again).  Each chip has a different display
 738 * FIFO size and allocation, so the caller needs to figure that out and pass
 739 * in the correct intel_watermark_params structure.
 740 *
 741 * As the pixel clock runs, the FIFO will be drained at a rate that depends
 742 * on the pixel size.  When it reaches the watermark level, it'll start
 743 * fetching FIFO line sized based chunks from memory until the FIFO fills
 744 * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 745 * will occur, and a display engine hang could result.
 746 */
 747static unsigned int intel_calculate_wm(int pixel_rate,
 748				       const struct intel_watermark_params *wm,
 749				       int fifo_size, int cpp,
 750				       unsigned int latency_ns)
 751{
 752	int entries, wm_size;
 753
 754	/*
 755	 * Note: we need to make sure we don't overflow for various clock &
 756	 * latency values.
 757	 * clocks go from a few thousand to several hundred thousand.
 758	 * latency is usually a few thousand
 759	 */
 760	entries = intel_wm_method1(pixel_rate, cpp,
 761				   latency_ns / 100);
 762	entries = DIV_ROUND_UP(entries, wm->cacheline_size) +
 763		wm->guard_size;
 764	DRM_DEBUG_KMS("FIFO entries required for mode: %d\n", entries);
 765
 766	wm_size = fifo_size - entries;
 767	DRM_DEBUG_KMS("FIFO watermark level: %d\n", wm_size);
 768
 769	/* Don't promote wm_size to unsigned... */
 770	if (wm_size > wm->max_wm)
 771		wm_size = wm->max_wm;
 772	if (wm_size <= 0)
 773		wm_size = wm->default_wm;
 774
 775	/*
 776	 * Bspec seems to indicate that the value shouldn't be lower than
 777	 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
 778	 * Lets go for 8 which is the burst size since certain platforms
 779	 * already use a hardcoded 8 (which is what the spec says should be
 780	 * done).
 781	 */
 782	if (wm_size <= 8)
 783		wm_size = 8;
 784
 785	return wm_size;
 786}
 787
 788static bool is_disabling(int old, int new, int threshold)
 789{
 790	return old >= threshold && new < threshold;
 791}
 792
 793static bool is_enabling(int old, int new, int threshold)
 794{
 795	return old < threshold && new >= threshold;
 796}
 797
 798static int intel_wm_num_levels(struct drm_i915_private *dev_priv)
 799{
 800	return dev_priv->wm.max_level + 1;
 801}
 802
 803static bool intel_wm_plane_visible(const struct intel_crtc_state *crtc_state,
 804				   const struct intel_plane_state *plane_state)
 805{
 806	struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
 807
 808	/* FIXME check the 'enable' instead */
 809	if (!crtc_state->base.active)
 810		return false;
 811
 812	/*
 813	 * Treat cursor with fb as always visible since cursor updates
 814	 * can happen faster than the vrefresh rate, and the current
 815	 * watermark code doesn't handle that correctly. Cursor updates
 816	 * which set/clear the fb or change the cursor size are going
 817	 * to get throttled by intel_legacy_cursor_update() to work
 818	 * around this problem with the watermark code.
 819	 */
 820	if (plane->id == PLANE_CURSOR)
 821		return plane_state->base.fb != NULL;
 822	else
 823		return plane_state->base.visible;
 824}
 825
 826static struct intel_crtc *single_enabled_crtc(struct drm_i915_private *dev_priv)
 827{
 828	struct intel_crtc *crtc, *enabled = NULL;
 829
 830	for_each_intel_crtc(&dev_priv->drm, crtc) {
 831		if (intel_crtc_active(crtc)) {
 832			if (enabled)
 833				return NULL;
 834			enabled = crtc;
 835		}
 836	}
 837
 838	return enabled;
 839}
 840
 841static void pineview_update_wm(struct intel_crtc *unused_crtc)
 842{
 843	struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
 844	struct intel_crtc *crtc;
 845	const struct cxsr_latency *latency;
 846	u32 reg;
 847	unsigned int wm;
 848
 849	latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
 850					 dev_priv->is_ddr3,
 851					 dev_priv->fsb_freq,
 852					 dev_priv->mem_freq);
 853	if (!latency) {
 854		DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 855		intel_set_memory_cxsr(dev_priv, false);
 856		return;
 857	}
 858
 859	crtc = single_enabled_crtc(dev_priv);
 860	if (crtc) {
 861		const struct drm_display_mode *adjusted_mode =
 862			&crtc->config->base.adjusted_mode;
 863		const struct drm_framebuffer *fb =
 864			crtc->base.primary->state->fb;
 865		int cpp = fb->format->cpp[0];
 866		int clock = adjusted_mode->crtc_clock;
 867
 868		/* Display SR */
 869		wm = intel_calculate_wm(clock, &pineview_display_wm,
 870					pineview_display_wm.fifo_size,
 871					cpp, latency->display_sr);
 872		reg = I915_READ(DSPFW1);
 873		reg &= ~DSPFW_SR_MASK;
 874		reg |= FW_WM(wm, SR);
 875		I915_WRITE(DSPFW1, reg);
 876		DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
 877
 878		/* cursor SR */
 879		wm = intel_calculate_wm(clock, &pineview_cursor_wm,
 880					pineview_display_wm.fifo_size,
 881					4, latency->cursor_sr);
 882		reg = I915_READ(DSPFW3);
 883		reg &= ~DSPFW_CURSOR_SR_MASK;
 884		reg |= FW_WM(wm, CURSOR_SR);
 885		I915_WRITE(DSPFW3, reg);
 886
 887		/* Display HPLL off SR */
 888		wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
 889					pineview_display_hplloff_wm.fifo_size,
 890					cpp, latency->display_hpll_disable);
 891		reg = I915_READ(DSPFW3);
 892		reg &= ~DSPFW_HPLL_SR_MASK;
 893		reg |= FW_WM(wm, HPLL_SR);
 894		I915_WRITE(DSPFW3, reg);
 895
 896		/* cursor HPLL off SR */
 897		wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
 898					pineview_display_hplloff_wm.fifo_size,
 899					4, latency->cursor_hpll_disable);
 900		reg = I915_READ(DSPFW3);
 901		reg &= ~DSPFW_HPLL_CURSOR_MASK;
 902		reg |= FW_WM(wm, HPLL_CURSOR);
 903		I915_WRITE(DSPFW3, reg);
 904		DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
 905
 906		intel_set_memory_cxsr(dev_priv, true);
 907	} else {
 908		intel_set_memory_cxsr(dev_priv, false);
 909	}
 910}
 911
 912/*
 913 * Documentation says:
 914 * "If the line size is small, the TLB fetches can get in the way of the
 915 *  data fetches, causing some lag in the pixel data return which is not
 916 *  accounted for in the above formulas. The following adjustment only
 917 *  needs to be applied if eight whole lines fit in the buffer at once.
 918 *  The WM is adjusted upwards by the difference between the FIFO size
 919 *  and the size of 8 whole lines. This adjustment is always performed
 920 *  in the actual pixel depth regardless of whether FBC is enabled or not."
 921 */
 922static unsigned int g4x_tlb_miss_wa(int fifo_size, int width, int cpp)
 923{
 924	int tlb_miss = fifo_size * 64 - width * cpp * 8;
 925
 926	return max(0, tlb_miss);
 927}
 928
 929static void g4x_write_wm_values(struct drm_i915_private *dev_priv,
 930				const struct g4x_wm_values *wm)
 931{
 932	enum pipe pipe;
 933
 934	for_each_pipe(dev_priv, pipe)
 935		trace_g4x_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
 936
 937	I915_WRITE(DSPFW1,
 938		   FW_WM(wm->sr.plane, SR) |
 939		   FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
 940		   FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
 941		   FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
 942	I915_WRITE(DSPFW2,
 943		   (wm->fbc_en ? DSPFW_FBC_SR_EN : 0) |
 944		   FW_WM(wm->sr.fbc, FBC_SR) |
 945		   FW_WM(wm->hpll.fbc, FBC_HPLL_SR) |
 946		   FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEB) |
 947		   FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
 948		   FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
 949	I915_WRITE(DSPFW3,
 950		   (wm->hpll_en ? DSPFW_HPLL_SR_EN : 0) |
 951		   FW_WM(wm->sr.cursor, CURSOR_SR) |
 952		   FW_WM(wm->hpll.cursor, HPLL_CURSOR) |
 953		   FW_WM(wm->hpll.plane, HPLL_SR));
 954
 955	POSTING_READ(DSPFW1);
 956}
 957
 958#define FW_WM_VLV(value, plane) \
 959	(((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
 960
 961static void vlv_write_wm_values(struct drm_i915_private *dev_priv,
 962				const struct vlv_wm_values *wm)
 963{
 964	enum pipe pipe;
 965
 966	for_each_pipe(dev_priv, pipe) {
 967		trace_vlv_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
 968
 969		I915_WRITE(VLV_DDL(pipe),
 970			   (wm->ddl[pipe].plane[PLANE_CURSOR] << DDL_CURSOR_SHIFT) |
 971			   (wm->ddl[pipe].plane[PLANE_SPRITE1] << DDL_SPRITE_SHIFT(1)) |
 972			   (wm->ddl[pipe].plane[PLANE_SPRITE0] << DDL_SPRITE_SHIFT(0)) |
 973			   (wm->ddl[pipe].plane[PLANE_PRIMARY] << DDL_PLANE_SHIFT));
 974	}
 975
 976	/*
 977	 * Zero the (unused) WM1 watermarks, and also clear all the
 978	 * high order bits so that there are no out of bounds values
 979	 * present in the registers during the reprogramming.
 980	 */
 981	I915_WRITE(DSPHOWM, 0);
 982	I915_WRITE(DSPHOWM1, 0);
 983	I915_WRITE(DSPFW4, 0);
 984	I915_WRITE(DSPFW5, 0);
 985	I915_WRITE(DSPFW6, 0);
 986
 987	I915_WRITE(DSPFW1,
 988		   FW_WM(wm->sr.plane, SR) |
 989		   FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
 990		   FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
 991		   FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
 992	I915_WRITE(DSPFW2,
 993		   FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE1], SPRITEB) |
 994		   FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
 995		   FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
 996	I915_WRITE(DSPFW3,
 997		   FW_WM(wm->sr.cursor, CURSOR_SR));
 998
 999	if (IS_CHERRYVIEW(dev_priv)) {
1000		I915_WRITE(DSPFW7_CHV,
1001			   FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
1002			   FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
1003		I915_WRITE(DSPFW8_CHV,
1004			   FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE1], SPRITEF) |
1005			   FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE0], SPRITEE));
1006		I915_WRITE(DSPFW9_CHV,
1007			   FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_PRIMARY], PLANEC) |
1008			   FW_WM(wm->pipe[PIPE_C].plane[PLANE_CURSOR], CURSORC));
1009		I915_WRITE(DSPHOWM,
1010			   FW_WM(wm->sr.plane >> 9, SR_HI) |
1011			   FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE1] >> 8, SPRITEF_HI) |
1012			   FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE0] >> 8, SPRITEE_HI) |
1013			   FW_WM(wm->pipe[PIPE_C].plane[PLANE_PRIMARY] >> 8, PLANEC_HI) |
1014			   FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
1015			   FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
1016			   FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
1017			   FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
1018			   FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
1019			   FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
1020	} else {
1021		I915_WRITE(DSPFW7,
1022			   FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
1023			   FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
1024		I915_WRITE(DSPHOWM,
1025			   FW_WM(wm->sr.plane >> 9, SR_HI) |
1026			   FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
1027			   FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
1028			   FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
1029			   FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
1030			   FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
1031			   FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
1032	}
1033
1034	POSTING_READ(DSPFW1);
1035}
1036
1037#undef FW_WM_VLV
1038
1039static void g4x_setup_wm_latency(struct drm_i915_private *dev_priv)
1040{
1041	/* all latencies in usec */
1042	dev_priv->wm.pri_latency[G4X_WM_LEVEL_NORMAL] = 5;
1043	dev_priv->wm.pri_latency[G4X_WM_LEVEL_SR] = 12;
1044	dev_priv->wm.pri_latency[G4X_WM_LEVEL_HPLL] = 35;
1045
1046	dev_priv->wm.max_level = G4X_WM_LEVEL_HPLL;
1047}
1048
1049static int g4x_plane_fifo_size(enum plane_id plane_id, int level)
1050{
1051	/*
1052	 * DSPCNTR[13] supposedly controls whether the
1053	 * primary plane can use the FIFO space otherwise
1054	 * reserved for the sprite plane. It's not 100% clear
1055	 * what the actual FIFO size is, but it looks like we
1056	 * can happily set both primary and sprite watermarks
1057	 * up to 127 cachelines. So that would seem to mean
1058	 * that either DSPCNTR[13] doesn't do anything, or that
1059	 * the total FIFO is >= 256 cachelines in size. Either
1060	 * way, we don't seem to have to worry about this
1061	 * repartitioning as the maximum watermark value the
1062	 * register can hold for each plane is lower than the
1063	 * minimum FIFO size.
1064	 */
1065	switch (plane_id) {
1066	case PLANE_CURSOR:
1067		return 63;
1068	case PLANE_PRIMARY:
1069		return level == G4X_WM_LEVEL_NORMAL ? 127 : 511;
1070	case PLANE_SPRITE0:
1071		return level == G4X_WM_LEVEL_NORMAL ? 127 : 0;
1072	default:
1073		MISSING_CASE(plane_id);
1074		return 0;
1075	}
1076}
1077
1078static int g4x_fbc_fifo_size(int level)
1079{
1080	switch (level) {
1081	case G4X_WM_LEVEL_SR:
1082		return 7;
1083	case G4X_WM_LEVEL_HPLL:
1084		return 15;
1085	default:
1086		MISSING_CASE(level);
1087		return 0;
1088	}
1089}
1090
1091static uint16_t g4x_compute_wm(const struct intel_crtc_state *crtc_state,
1092			       const struct intel_plane_state *plane_state,
1093			       int level)
1094{
1095	struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1096	struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1097	const struct drm_display_mode *adjusted_mode =
1098		&crtc_state->base.adjusted_mode;
1099	unsigned int latency = dev_priv->wm.pri_latency[level] * 10;
1100	unsigned int clock, htotal, cpp, width, wm;
1101
1102	if (latency == 0)
1103		return USHRT_MAX;
1104
1105	if (!intel_wm_plane_visible(crtc_state, plane_state))
1106		return 0;
1107
1108	/*
1109	 * Not 100% sure which way ELK should go here as the
1110	 * spec only says CL/CTG should assume 32bpp and BW
1111	 * doesn't need to. But as these things followed the
1112	 * mobile vs. desktop lines on gen3 as well, let's
1113	 * assume ELK doesn't need this.
1114	 *
1115	 * The spec also fails to list such a restriction for
1116	 * the HPLL watermark, which seems a little strange.
1117	 * Let's use 32bpp for the HPLL watermark as well.
1118	 */
1119	if (IS_GM45(dev_priv) && plane->id == PLANE_PRIMARY &&
1120	    level != G4X_WM_LEVEL_NORMAL)
1121		cpp = 4;
1122	else
1123		cpp = plane_state->base.fb->format->cpp[0];
1124
1125	clock = adjusted_mode->crtc_clock;
1126	htotal = adjusted_mode->crtc_htotal;
1127
1128	if (plane->id == PLANE_CURSOR)
1129		width = plane_state->base.crtc_w;
1130	else
1131		width = drm_rect_width(&plane_state->base.dst);
1132
1133	if (plane->id == PLANE_CURSOR) {
1134		wm = intel_wm_method2(clock, htotal, width, cpp, latency);
1135	} else if (plane->id == PLANE_PRIMARY &&
1136		   level == G4X_WM_LEVEL_NORMAL) {
1137		wm = intel_wm_method1(clock, cpp, latency);
1138	} else {
1139		unsigned int small, large;
1140
1141		small = intel_wm_method1(clock, cpp, latency);
1142		large = intel_wm_method2(clock, htotal, width, cpp, latency);
1143
1144		wm = min(small, large);
1145	}
1146
1147	wm += g4x_tlb_miss_wa(g4x_plane_fifo_size(plane->id, level),
1148			      width, cpp);
1149
1150	wm = DIV_ROUND_UP(wm, 64) + 2;
1151
1152	return min_t(unsigned int, wm, USHRT_MAX);
1153}
1154
1155static bool g4x_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1156				 int level, enum plane_id plane_id, u16 value)
1157{
1158	struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1159	bool dirty = false;
1160
1161	for (; level < intel_wm_num_levels(dev_priv); level++) {
1162		struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1163
1164		dirty |= raw->plane[plane_id] != value;
1165		raw->plane[plane_id] = value;
1166	}
1167
1168	return dirty;
1169}
1170
1171static bool g4x_raw_fbc_wm_set(struct intel_crtc_state *crtc_state,
1172			       int level, u16 value)
1173{
1174	struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1175	bool dirty = false;
1176
1177	/* NORMAL level doesn't have an FBC watermark */
1178	level = max(level, G4X_WM_LEVEL_SR);
1179
1180	for (; level < intel_wm_num_levels(dev_priv); level++) {
1181		struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1182
1183		dirty |= raw->fbc != value;
1184		raw->fbc = value;
1185	}
1186
1187	return dirty;
1188}
1189
1190static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
1191				   const struct intel_plane_state *pstate,
1192				   uint32_t pri_val);
1193
1194static bool g4x_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1195				     const struct intel_plane_state *plane_state)
1196{
1197	struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1198	int num_levels = intel_wm_num_levels(to_i915(plane->base.dev));
1199	enum plane_id plane_id = plane->id;
1200	bool dirty = false;
1201	int level;
1202
1203	if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1204		dirty |= g4x_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1205		if (plane_id == PLANE_PRIMARY)
1206			dirty |= g4x_raw_fbc_wm_set(crtc_state, 0, 0);
1207		goto out;
1208	}
1209
1210	for (level = 0; level < num_levels; level++) {
1211		struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1212		int wm, max_wm;
1213
1214		wm = g4x_compute_wm(crtc_state, plane_state, level);
1215		max_wm = g4x_plane_fifo_size(plane_id, level);
1216
1217		if (wm > max_wm)
1218			break;
1219
1220		dirty |= raw->plane[plane_id] != wm;
1221		raw->plane[plane_id] = wm;
1222
1223		if (plane_id != PLANE_PRIMARY ||
1224		    level == G4X_WM_LEVEL_NORMAL)
1225			continue;
1226
1227		wm = ilk_compute_fbc_wm(crtc_state, plane_state,
1228					raw->plane[plane_id]);
1229		max_wm = g4x_fbc_fifo_size(level);
1230
1231		/*
1232		 * FBC wm is not mandatory as we
1233		 * can always just disable its use.
1234		 */
1235		if (wm > max_wm)
1236			wm = USHRT_MAX;
1237
1238		dirty |= raw->fbc != wm;
1239		raw->fbc = wm;
1240	}
1241
1242	/* mark watermarks as invalid */
1243	dirty |= g4x_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1244
1245	if (plane_id == PLANE_PRIMARY)
1246		dirty |= g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
1247
1248 out:
1249	if (dirty) {
1250		DRM_DEBUG_KMS("%s watermarks: normal=%d, SR=%d, HPLL=%d\n",
1251			      plane->base.name,
1252			      crtc_state->wm.g4x.raw[G4X_WM_LEVEL_NORMAL].plane[plane_id],
1253			      crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].plane[plane_id],
1254			      crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].plane[plane_id]);
1255
1256		if (plane_id == PLANE_PRIMARY)
1257			DRM_DEBUG_KMS("FBC watermarks: SR=%d, HPLL=%d\n",
1258				      crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].fbc,
1259				      crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].fbc);
1260	}
1261
1262	return dirty;
1263}
1264
1265static bool g4x_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1266				      enum plane_id plane_id, int level)
1267{
1268	const struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1269
1270	return raw->plane[plane_id] <= g4x_plane_fifo_size(plane_id, level);
1271}
1272
1273static bool g4x_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state,
1274				     int level)
1275{
1276	struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1277
1278	if (level > dev_priv->wm.max_level)
1279		return false;
1280
1281	return g4x_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1282		g4x_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1283		g4x_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1284}
1285
1286/* mark all levels starting from 'level' as invalid */
1287static void g4x_invalidate_wms(struct intel_crtc *crtc,
1288			       struct g4x_wm_state *wm_state, int level)
1289{
1290	if (level <= G4X_WM_LEVEL_NORMAL) {
1291		enum plane_id plane_id;
1292
1293		for_each_plane_id_on_crtc(crtc, plane_id)
1294			wm_state->wm.plane[plane_id] = USHRT_MAX;
1295	}
1296
1297	if (level <= G4X_WM_LEVEL_SR) {
1298		wm_state->cxsr = false;
1299		wm_state->sr.cursor = USHRT_MAX;
1300		wm_state->sr.plane = USHRT_MAX;
1301		wm_state->sr.fbc = USHRT_MAX;
1302	}
1303
1304	if (level <= G4X_WM_LEVEL_HPLL) {
1305		wm_state->hpll_en = false;
1306		wm_state->hpll.cursor = USHRT_MAX;
1307		wm_state->hpll.plane = USHRT_MAX;
1308		wm_state->hpll.fbc = USHRT_MAX;
1309	}
1310}
1311
1312static int g4x_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1313{
1314	struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1315	struct intel_atomic_state *state =
1316		to_intel_atomic_state(crtc_state->base.state);
1317	struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
1318	int num_active_planes = hweight32(crtc_state->active_planes &
1319					  ~BIT(PLANE_CURSOR));
1320	const struct g4x_pipe_wm *raw;
1321	const struct intel_plane_state *old_plane_state;
1322	const struct intel_plane_state *new_plane_state;
1323	struct intel_plane *plane;
1324	enum plane_id plane_id;
1325	int i, level;
1326	unsigned int dirty = 0;
1327
1328	for_each_oldnew_intel_plane_in_state(state, plane,
1329					     old_plane_state,
1330					     new_plane_state, i) {
1331		if (new_plane_state->base.crtc != &crtc->base &&
1332		    old_plane_state->base.crtc != &crtc->base)
1333			continue;
1334
1335		if (g4x_raw_plane_wm_compute(crtc_state, new_plane_state))
1336			dirty |= BIT(plane->id);
1337	}
1338
1339	if (!dirty)
1340		return 0;
1341
1342	level = G4X_WM_LEVEL_NORMAL;
1343	if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1344		goto out;
1345
1346	raw = &crtc_state->wm.g4x.raw[level];
1347	for_each_plane_id_on_crtc(crtc, plane_id)
1348		wm_state->wm.plane[plane_id] = raw->plane[plane_id];
1349
1350	level = G4X_WM_LEVEL_SR;
1351
1352	if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1353		goto out;
1354
1355	raw = &crtc_state->wm.g4x.raw[level];
1356	wm_state->sr.plane = raw->plane[PLANE_PRIMARY];
1357	wm_state->sr.cursor = raw->plane[PLANE_CURSOR];
1358	wm_state->sr.fbc = raw->fbc;
1359
1360	wm_state->cxsr = num_active_planes == BIT(PLANE_PRIMARY);
1361
1362	level = G4X_WM_LEVEL_HPLL;
1363
1364	if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1365		goto out;
1366
1367	raw = &crtc_state->wm.g4x.raw[level];
1368	wm_state->hpll.plane = raw->plane[PLANE_PRIMARY];
1369	wm_state->hpll.cursor = raw->plane[PLANE_CURSOR];
1370	wm_state->hpll.fbc = raw->fbc;
1371
1372	wm_state->hpll_en = wm_state->cxsr;
1373
1374	level++;
1375
1376 out:
1377	if (level == G4X_WM_LEVEL_NORMAL)
1378		return -EINVAL;
1379
1380	/* invalidate the higher levels */
1381	g4x_invalidate_wms(crtc, wm_state, level);
1382
1383	/*
1384	 * Determine if the FBC watermark(s) can be used. IF
1385	 * this isn't the case we prefer to disable the FBC
1386	 ( watermark(s) rather than disable the SR/HPLL
1387	 * level(s) entirely.
1388	 */
1389	wm_state->fbc_en = level > G4X_WM_LEVEL_NORMAL;
1390
1391	if (level >= G4X_WM_LEVEL_SR &&
1392	    wm_state->sr.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_SR))
1393		wm_state->fbc_en = false;
1394	else if (level >= G4X_WM_LEVEL_HPLL &&
1395		 wm_state->hpll.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_HPLL))
1396		wm_state->fbc_en = false;
1397
1398	return 0;
1399}
1400
1401static int g4x_compute_intermediate_wm(struct drm_device *dev,
1402				       struct intel_crtc *crtc,
1403				       struct intel_crtc_state *new_crtc_state)
1404{
1405	struct g4x_wm_state *intermediate = &new_crtc_state->wm.g4x.intermediate;
1406	const struct g4x_wm_state *optimal = &new_crtc_state->wm.g4x.optimal;
1407	struct intel_atomic_state *intel_state =
1408		to_intel_atomic_state(new_crtc_state->base.state);
1409	const struct intel_crtc_state *old_crtc_state =
1410		intel_atomic_get_old_crtc_state(intel_state, crtc);
1411	const struct g4x_wm_state *active = &old_crtc_state->wm.g4x.optimal;
1412	enum plane_id plane_id;
1413
1414	if (!new_crtc_state->base.active || drm_atomic_crtc_needs_modeset(&new_crtc_state->base)) {
1415		*intermediate = *optimal;
1416
1417		intermediate->cxsr = false;
1418		intermediate->hpll_en = false;
1419		goto out;
1420	}
1421
1422	intermediate->cxsr = optimal->cxsr && active->cxsr &&
1423		!new_crtc_state->disable_cxsr;
1424	intermediate->hpll_en = optimal->hpll_en && active->hpll_en &&
1425		!new_crtc_state->disable_cxsr;
1426	intermediate->fbc_en = optimal->fbc_en && active->fbc_en;
1427
1428	for_each_plane_id_on_crtc(crtc, plane_id) {
1429		intermediate->wm.plane[plane_id] =
1430			max(optimal->wm.plane[plane_id],
1431			    active->wm.plane[plane_id]);
1432
1433		WARN_ON(intermediate->wm.plane[plane_id] >
1434			g4x_plane_fifo_size(plane_id, G4X_WM_LEVEL_NORMAL));
1435	}
1436
1437	intermediate->sr.plane = max(optimal->sr.plane,
1438				     active->sr.plane);
1439	intermediate->sr.cursor = max(optimal->sr.cursor,
1440				      active->sr.cursor);
1441	intermediate->sr.fbc = max(optimal->sr.fbc,
1442				   active->sr.fbc);
1443
1444	intermediate->hpll.plane = max(optimal->hpll.plane,
1445				       active->hpll.plane);
1446	intermediate->hpll.cursor = max(optimal->hpll.cursor,
1447					active->hpll.cursor);
1448	intermediate->hpll.fbc = max(optimal->hpll.fbc,
1449				     active->hpll.fbc);
1450
1451	WARN_ON((intermediate->sr.plane >
1452		 g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_SR) ||
1453		 intermediate->sr.cursor >
1454		 g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_SR)) &&
1455		intermediate->cxsr);
1456	WARN_ON((intermediate->sr.plane >
1457		 g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_HPLL) ||
1458		 intermediate->sr.cursor >
1459		 g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_HPLL)) &&
1460		intermediate->hpll_en);
1461
1462	WARN_ON(intermediate->sr.fbc > g4x_fbc_fifo_size(1) &&
1463		intermediate->fbc_en && intermediate->cxsr);
1464	WARN_ON(intermediate->hpll.fbc > g4x_fbc_fifo_size(2) &&
1465		intermediate->fbc_en && intermediate->hpll_en);
1466
1467out:
1468	/*
1469	 * If our intermediate WM are identical to the final WM, then we can
1470	 * omit the post-vblank programming; only update if it's different.
1471	 */
1472	if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
1473		new_crtc_state->wm.need_postvbl_update = true;
1474
1475	return 0;
1476}
1477
1478static void g4x_merge_wm(struct drm_i915_private *dev_priv,
1479			 struct g4x_wm_values *wm)
1480{
1481	struct intel_crtc *crtc;
1482	int num_active_crtcs = 0;
1483
1484	wm->cxsr = true;
1485	wm->hpll_en = true;
1486	wm->fbc_en = true;
1487
1488	for_each_intel_crtc(&dev_priv->drm, crtc) {
1489		const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1490
1491		if (!crtc->active)
1492			continue;
1493
1494		if (!wm_state->cxsr)
1495			wm->cxsr = false;
1496		if (!wm_state->hpll_en)
1497			wm->hpll_en = false;
1498		if (!wm_state->fbc_en)
1499			wm->fbc_en = false;
1500
1501		num_active_crtcs++;
1502	}
1503
1504	if (num_active_crtcs != 1) {
1505		wm->cxsr = false;
1506		wm->hpll_en = false;
1507		wm->fbc_en = false;
1508	}
1509
1510	for_each_intel_crtc(&dev_priv->drm, crtc) {
1511		const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1512		enum pipe pipe = crtc->pipe;
1513
1514		wm->pipe[pipe] = wm_state->wm;
1515		if (crtc->active && wm->cxsr)
1516			wm->sr = wm_state->sr;
1517		if (crtc->active && wm->hpll_en)
1518			wm->hpll = wm_state->hpll;
1519	}
1520}
1521
1522static void g4x_program_watermarks(struct drm_i915_private *dev_priv)
1523{
1524	struct g4x_wm_values *old_wm = &dev_priv->wm.g4x;
1525	struct g4x_wm_values new_wm = {};
1526
1527	g4x_merge_wm(dev_priv, &new_wm);
1528
1529	if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
1530		return;
1531
1532	if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
1533		_intel_set_memory_cxsr(dev_priv, false);
1534
1535	g4x_write_wm_values(dev_priv, &new_wm);
1536
1537	if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
1538		_intel_set_memory_cxsr(dev_priv, true);
1539
1540	*old_wm = new_wm;
1541}
1542
1543static void g4x_initial_watermarks(struct intel_atomic_state *state,
1544				   struct intel_crtc_state *crtc_state)
1545{
1546	struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1547	struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1548
1549	mutex_lock(&dev_priv->wm.wm_mutex);
1550	crtc->wm.active.g4x = crtc_state->wm.g4x.intermediate;
1551	g4x_program_watermarks(dev_priv);
1552	mutex_unlock(&dev_priv->wm.wm_mutex);
1553}
1554
1555static void g4x_optimize_watermarks(struct intel_atomic_state *state,
1556				    struct intel_crtc_state *crtc_state)
1557{
1558	struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1559	struct intel_crtc *intel_crtc = to_intel_crtc(crtc_state->base.crtc);
1560
1561	if (!crtc_state->wm.need_postvbl_update)
1562		return;
1563
1564	mutex_lock(&dev_priv->wm.wm_mutex);
1565	intel_crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
1566	g4x_program_watermarks(dev_priv);
1567	mutex_unlock(&dev_priv->wm.wm_mutex);
1568}
1569
1570/* latency must be in 0.1us units. */
1571static unsigned int vlv_wm_method2(unsigned int pixel_rate,
1572				   unsigned int htotal,
1573				   unsigned int width,
1574				   unsigned int cpp,
1575				   unsigned int latency)
1576{
1577	unsigned int ret;
1578
1579	ret = intel_wm_method2(pixel_rate, htotal,
1580			       width, cpp, latency);
1581	ret = DIV_ROUND_UP(ret, 64);
1582
1583	return ret;
1584}
1585
1586static void vlv_setup_wm_latency(struct drm_i915_private *dev_priv)
1587{
1588	/* all latencies in usec */
1589	dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
1590
1591	dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
1592
1593	if (IS_CHERRYVIEW(dev_priv)) {
1594		dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
1595		dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
1596
1597		dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
1598	}
1599}
1600
1601static uint16_t vlv_compute_wm_level(const struct intel_crtc_state *crtc_state,
1602				     const struct intel_plane_state *plane_state,
1603				     int level)
1604{
1605	struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1606	struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1607	const struct drm_display_mode *adjusted_mode =
1608		&crtc_state->base.adjusted_mode;
1609	unsigned int clock, htotal, cpp, width, wm;
1610
1611	if (dev_priv->wm.pri_latency[level] == 0)
1612		return USHRT_MAX;
1613
1614	if (!intel_wm_plane_visible(crtc_state, plane_state))
1615		return 0;
1616
1617	cpp = plane_state->base.fb->format->cpp[0];
1618	clock = adjusted_mode->crtc_clock;
1619	htotal = adjusted_mode->crtc_htotal;
1620	width = crtc_state->pipe_src_w;
1621
1622	if (plane->id == PLANE_CURSOR) {
1623		/*
1624		 * FIXME the formula gives values that are
1625		 * too big for the cursor FIFO, and hence we
1626		 * would never be able to use cursors. For
1627		 * now just hardcode the watermark.
1628		 */
1629		wm = 63;
1630	} else {
1631		wm = vlv_wm_method2(clock, htotal, width, cpp,
1632				    dev_priv->wm.pri_latency[level] * 10);
1633	}
1634
1635	return min_t(unsigned int, wm, USHRT_MAX);
1636}
1637
1638static bool vlv_need_sprite0_fifo_workaround(unsigned int active_planes)
1639{
1640	return (active_planes & (BIT(PLANE_SPRITE0) |
1641				 BIT(PLANE_SPRITE1))) == BIT(PLANE_SPRITE1);
1642}
1643
1644static int vlv_compute_fifo(struct intel_crtc_state *crtc_state)
1645{
1646	struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1647	const struct g4x_pipe_wm *raw =
1648		&crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2];
1649	struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
1650	unsigned int active_planes = crtc_state->active_planes & ~BIT(PLANE_CURSOR);
1651	int num_active_planes = hweight32(active_planes);
1652	const int fifo_size = 511;
1653	int fifo_extra, fifo_left = fifo_size;
1654	int sprite0_fifo_extra = 0;
1655	unsigned int total_rate;
1656	enum plane_id plane_id;
1657
1658	/*
1659	 * When enabling sprite0 after sprite1 has already been enabled
1660	 * we tend to get an underrun unless sprite0 already has some
1661	 * FIFO space allcoated. Hence we always allocate at least one
1662	 * cacheline for sprite0 whenever sprite1 is enabled.
1663	 *
1664	 * All other plane enable sequences appear immune to this problem.
1665	 */
1666	if (vlv_need_sprite0_fifo_workaround(active_planes))
1667		sprite0_fifo_extra = 1;
1668
1669	total_rate = raw->plane[PLANE_PRIMARY] +
1670		raw->plane[PLANE_SPRITE0] +
1671		raw->plane[PLANE_SPRITE1] +
1672		sprite0_fifo_extra;
1673
1674	if (total_rate > fifo_size)
1675		return -EINVAL;
1676
1677	if (total_rate == 0)
1678		total_rate = 1;
1679
1680	for_each_plane_id_on_crtc(crtc, plane_id) {
1681		unsigned int rate;
1682
1683		if ((active_planes & BIT(plane_id)) == 0) {
1684			fifo_state->plane[plane_id] = 0;
1685			continue;
1686		}
1687
1688		rate = raw->plane[plane_id];
1689		fifo_state->plane[plane_id] = fifo_size * rate / total_rate;
1690		fifo_left -= fifo_state->plane[plane_id];
1691	}
1692
1693	fifo_state->plane[PLANE_SPRITE0] += sprite0_fifo_extra;
1694	fifo_left -= sprite0_fifo_extra;
1695
1696	fifo_state->plane[PLANE_CURSOR] = 63;
1697
1698	fifo_extra = DIV_ROUND_UP(fifo_left, num_active_planes ?: 1);
1699
1700	/* spread the remainder evenly */
1701	for_each_plane_id_on_crtc(crtc, plane_id) {
1702		int plane_extra;
1703
1704		if (fifo_left == 0)
1705			break;
1706
1707		if ((active_planes & BIT(plane_id)) == 0)
1708			continue;
1709
1710		plane_extra = min(fifo_extra, fifo_left);
1711		fifo_state->plane[plane_id] += plane_extra;
1712		fifo_left -= plane_extra;
1713	}
1714
1715	WARN_ON(active_planes != 0 && fifo_left != 0);
1716
1717	/* give it all to the first plane if none are active */
1718	if (active_planes == 0) {
1719		WARN_ON(fifo_left != fifo_size);
1720		fifo_state->plane[PLANE_PRIMARY] = fifo_left;
1721	}
1722
1723	return 0;
1724}
1725
1726/* mark all levels starting from 'level' as invalid */
1727static void vlv_invalidate_wms(struct intel_crtc *crtc,
1728			       struct vlv_wm_state *wm_state, int level)
1729{
1730	struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1731
1732	for (; level < intel_wm_num_levels(dev_priv); level++) {
1733		enum plane_id plane_id;
1734
1735		for_each_plane_id_on_crtc(crtc, plane_id)
1736			wm_state->wm[level].plane[plane_id] = USHRT_MAX;
1737
1738		wm_state->sr[level].cursor = USHRT_MAX;
1739		wm_state->sr[level].plane = USHRT_MAX;
1740	}
1741}
1742
1743static u16 vlv_invert_wm_value(u16 wm, u16 fifo_size)
1744{
1745	if (wm > fifo_size)
1746		return USHRT_MAX;
1747	else
1748		return fifo_size - wm;
1749}
1750
1751/*
1752 * Starting from 'level' set all higher
1753 * levels to 'value' in the "raw" watermarks.
1754 */
1755static bool vlv_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1756				 int level, enum plane_id plane_id, u16 value)
1757{
1758	struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1759	int num_levels = intel_wm_num_levels(dev_priv);
1760	bool dirty = false;
1761
1762	for (; level < num_levels; level++) {
1763		struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1764
1765		dirty |= raw->plane[plane_id] != value;
1766		raw->plane[plane_id] = value;
1767	}
1768
1769	return dirty;
1770}
1771
1772static bool vlv_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1773				     const struct intel_plane_state *plane_state)
1774{
1775	struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1776	enum plane_id plane_id = plane->id;
1777	int num_levels = intel_wm_num_levels(to_i915(plane->base.dev));
1778	int level;
1779	bool dirty = false;
1780
1781	if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1782		dirty |= vlv_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1783		goto out;
1784	}
1785
1786	for (level = 0; level < num_levels; level++) {
1787		struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1788		int wm = vlv_compute_wm_level(crtc_state, plane_state, level);
1789		int max_wm = plane_id == PLANE_CURSOR ? 63 : 511;
1790
1791		if (wm > max_wm)
1792			break;
1793
1794		dirty |= raw->plane[plane_id] != wm;
1795		raw->plane[plane_id] = wm;
1796	}
1797
1798	/* mark all higher levels as invalid */
1799	dirty |= vlv_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1800
1801out:
1802	if (dirty)
1803		DRM_DEBUG_KMS("%s watermarks: PM2=%d, PM5=%d, DDR DVFS=%d\n",
1804			      plane->base.name,
1805			      crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2].plane[plane_id],
1806			      crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM5].plane[plane_id],
1807			      crtc_state->wm.vlv.raw[VLV_WM_LEVEL_DDR_DVFS].plane[plane_id]);
1808
1809	return dirty;
1810}
1811
1812static bool vlv_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1813				      enum plane_id plane_id, int level)
1814{
1815	const struct g4x_pipe_wm *raw =
1816		&crtc_state->wm.vlv.raw[level];
1817	const struct vlv_fifo_state *fifo_state =
1818		&crtc_state->wm.vlv.fifo_state;
1819
1820	return raw->plane[plane_id] <= fifo_state->plane[plane_id];
1821}
1822
1823static bool vlv_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state, int level)
1824{
1825	return vlv_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1826		vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1827		vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE1, level) &&
1828		vlv_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1829}
1830
1831static int vlv_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1832{
1833	struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1834	struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1835	struct intel_atomic_state *state =
1836		to_intel_atomic_state(crtc_state->base.state);
1837	struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
1838	const struct vlv_fifo_state *fifo_state =
1839		&crtc_state->wm.vlv.fifo_state;
1840	int num_active_planes = hweight32(crtc_state->active_planes &
1841					  ~BIT(PLANE_CURSOR));
1842	bool needs_modeset = drm_atomic_crtc_needs_modeset(&crtc_state->base);
1843	const struct intel_plane_state *old_plane_state;
1844	const struct intel_plane_state *new_plane_state;
1845	struct intel_plane *plane;
1846	enum plane_id plane_id;
1847	int level, ret, i;
1848	unsigned int dirty = 0;
1849
1850	for_each_oldnew_intel_plane_in_state(state, plane,
1851					     old_plane_state,
1852					     new_plane_state, i) {
1853		if (new_plane_state->base.crtc != &crtc->base &&
1854		    old_plane_state->base.crtc != &crtc->base)
1855			continue;
1856
1857		if (vlv_raw_plane_wm_compute(crtc_state, new_plane_state))
1858			dirty |= BIT(plane->id);
1859	}
1860
1861	/*
1862	 * DSPARB registers may have been reset due to the
1863	 * power well being turned off. Make sure we restore
1864	 * them to a consistent state even if no primary/sprite
1865	 * planes are initially active.
1866	 */
1867	if (needs_modeset)
1868		crtc_state->fifo_changed = true;
1869
1870	if (!dirty)
1871		return 0;
1872
1873	/* cursor changes don't warrant a FIFO recompute */
1874	if (dirty & ~BIT(PLANE_CURSOR)) {
1875		const struct intel_crtc_state *old_crtc_state =
1876			intel_atomic_get_old_crtc_state(state, crtc);
1877		const struct vlv_fifo_state *old_fifo_state =
1878			&old_crtc_state->wm.vlv.fifo_state;
1879
1880		ret = vlv_compute_fifo(crtc_state);
1881		if (ret)
1882			return ret;
1883
1884		if (needs_modeset ||
1885		    memcmp(old_fifo_state, fifo_state,
1886			   sizeof(*fifo_state)) != 0)
1887			crtc_state->fifo_changed = true;
1888	}
1889
1890	/* initially allow all levels */
1891	wm_state->num_levels = intel_wm_num_levels(dev_priv);
1892	/*
1893	 * Note that enabling cxsr with no primary/sprite planes
1894	 * enabled can wedge the pipe. Hence we only allow cxsr
1895	 * with exactly one enabled primary/sprite plane.
1896	 */
1897	wm_state->cxsr = crtc->pipe != PIPE_C && num_active_planes == 1;
1898
1899	for (level = 0; level < wm_state->num_levels; level++) {
1900		const struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1901		const int sr_fifo_size = INTEL_INFO(dev_priv)->num_pipes * 512 - 1;
1902
1903		if (!vlv_raw_crtc_wm_is_valid(crtc_state, level))
1904			break;
1905
1906		for_each_plane_id_on_crtc(crtc, plane_id) {
1907			wm_state->wm[level].plane[plane_id] =
1908				vlv_invert_wm_value(raw->plane[plane_id],
1909						    fifo_state->plane[plane_id]);
1910		}
1911
1912		wm_state->sr[level].plane =
1913			vlv_invert_wm_value(max3(raw->plane[PLANE_PRIMARY],
1914						 raw->plane[PLANE_SPRITE0],
1915						 raw->plane[PLANE_SPRITE1]),
1916					    sr_fifo_size);
1917
1918		wm_state->sr[level].cursor =
1919			vlv_invert_wm_value(raw->plane[PLANE_CURSOR],
1920					    63);
1921	}
1922
1923	if (level == 0)
1924		return -EINVAL;
1925
1926	/* limit to only levels we can actually handle */
1927	wm_state->num_levels = level;
1928
1929	/* invalidate the higher levels */
1930	vlv_invalidate_wms(crtc, wm_state, level);
1931
1932	return 0;
1933}
1934
1935#define VLV_FIFO(plane, value) \
1936	(((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1937
1938static void vlv_atomic_update_fifo(struct intel_atomic_state *state,
1939				   struct intel_crtc_state *crtc_state)
1940{
1941	struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1942	struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1943	const struct vlv_fifo_state *fifo_state =
1944		&crtc_state->wm.vlv.fifo_state;
1945	int sprite0_start, sprite1_start, fifo_size;
1946
1947	if (!crtc_state->fifo_changed)
1948		return;
1949
1950	sprite0_start = fifo_state->plane[PLANE_PRIMARY];
1951	sprite1_start = fifo_state->plane[PLANE_SPRITE0] + sprite0_start;
1952	fifo_size = fifo_state->plane[PLANE_SPRITE1] + sprite1_start;
1953
1954	WARN_ON(fifo_state->plane[PLANE_CURSOR] != 63);
1955	WARN_ON(fifo_size != 511);
1956
1957	trace_vlv_fifo_size(crtc, sprite0_start, sprite1_start, fifo_size);
1958
1959	/*
1960	 * uncore.lock serves a double purpose here. It allows us to
1961	 * use the less expensive I915_{READ,WRITE}_FW() functions, and
1962	 * it protects the DSPARB registers from getting clobbered by
1963	 * parallel updates from multiple pipes.
1964	 *
1965	 * intel_pipe_update_start() has already disabled interrupts
1966	 * for us, so a plain spin_lock() is sufficient here.
1967	 */
1968	spin_lock(&dev_priv->uncore.lock);
1969
1970	switch (crtc->pipe) {
1971		uint32_t dsparb, dsparb2, dsparb3;
1972	case PIPE_A:
1973		dsparb = I915_READ_FW(DSPARB);
1974		dsparb2 = I915_READ_FW(DSPARB2);
1975
1976		dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1977			    VLV_FIFO(SPRITEB, 0xff));
1978		dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1979			   VLV_FIFO(SPRITEB, sprite1_start));
1980
1981		dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1982			     VLV_FIFO(SPRITEB_HI, 0x1));
1983		dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1984			   VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1985
1986		I915_WRITE_FW(DSPARB, dsparb);
1987		I915_WRITE_FW(DSPARB2, dsparb2);
1988		break;
1989	case PIPE_B:
1990		dsparb = I915_READ_FW(DSPARB);
1991		dsparb2 = I915_READ_FW(DSPARB2);
1992
1993		dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1994			    VLV_FIFO(SPRITED, 0xff));
1995		dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1996			   VLV_FIFO(SPRITED, sprite1_start));
1997
1998		dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1999			     VLV_FIFO(SPRITED_HI, 0xff));
2000		dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
2001			   VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
2002
2003		I915_WRITE_FW(DSPARB, dsparb);
2004		I915_WRITE_FW(DSPARB2, dsparb2);
2005		break;
2006	case PIPE_C:
2007		dsparb3 = I915_READ_FW(DSPARB3);
2008		dsparb2 = I915_READ_FW(DSPARB2);
2009
2010		dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
2011			     VLV_FIFO(SPRITEF, 0xff));
2012		dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
2013			    VLV_FIFO(SPRITEF, sprite1_start));
2014
2015		dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
2016			     VLV_FIFO(SPRITEF_HI, 0xff));
2017		dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
2018			   VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
2019
2020		I915_WRITE_FW(DSPARB3, dsparb3);
2021		I915_WRITE_FW(DSPARB2, dsparb2);
2022		break;
2023	default:
2024		break;
2025	}
2026
2027	POSTING_READ_FW(DSPARB);
2028
2029	spin_unlock(&dev_priv->uncore.lock);
2030}
2031
2032#undef VLV_FIFO
2033
2034static int vlv_compute_intermediate_wm(struct drm_device *dev,
2035				       struct intel_crtc *crtc,
2036				       struct intel_crtc_state *new_crtc_state)
2037{
2038	struct vlv_wm_state *intermediate = &new_crtc_state->wm.vlv.intermediate;
2039	const struct vlv_wm_state *optimal = &new_crtc_state->wm.vlv.optimal;
2040	struct intel_atomic_state *intel_state =
2041		to_intel_atomic_state(new_crtc_state->base.state);
2042	const struct intel_crtc_state *old_crtc_state =
2043		intel_atomic_get_old_crtc_state(intel_state, crtc);
2044	const struct vlv_wm_state *active = &old_crtc_state->wm.vlv.optimal;
2045	int level;
2046
2047	if (!new_crtc_state->base.active || drm_atomic_crtc_needs_modeset(&new_crtc_state->base)) {
2048		*intermediate = *optimal;
2049
2050		intermediate->cxsr = false;
2051		goto out;
2052	}
2053
2054	intermediate->num_levels = min(optimal->num_levels, active->num_levels);
2055	intermediate->cxsr = optimal->cxsr && active->cxsr &&
2056		!new_crtc_state->disable_cxsr;
2057
2058	for (level = 0; level < intermediate->num_levels; level++) {
2059		enum plane_id plane_id;
2060
2061		for_each_plane_id_on_crtc(crtc, plane_id) {
2062			intermediate->wm[level].plane[plane_id] =
2063				min(optimal->wm[level].plane[plane_id],
2064				    active->wm[level].plane[plane_id]);
2065		}
2066
2067		intermediate->sr[level].plane = min(optimal->sr[level].plane,
2068						    active->sr[level].plane);
2069		intermediate->sr[level].cursor = min(optimal->sr[level].cursor,
2070						     active->sr[level].cursor);
2071	}
2072
2073	vlv_invalidate_wms(crtc, intermediate, level);
2074
2075out:
2076	/*
2077	 * If our intermediate WM are identical to the final WM, then we can
2078	 * omit the post-vblank programming; only update if it's different.
2079	 */
2080	if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
2081		new_crtc_state->wm.need_postvbl_update = true;
2082
2083	return 0;
2084}
2085
2086static void vlv_merge_wm(struct drm_i915_private *dev_priv,
2087			 struct vlv_wm_values *wm)
2088{
2089	struct intel_crtc *crtc;
2090	int num_active_crtcs = 0;
2091
2092	wm->level = dev_priv->wm.max_level;
2093	wm->cxsr = true;
2094
2095	for_each_intel_crtc(&dev_priv->drm, crtc) {
2096		const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2097
2098		if (!crtc->active)
2099			continue;
2100
2101		if (!wm_state->cxsr)
2102			wm->cxsr = false;
2103
2104		num_active_crtcs++;
2105		wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
2106	}
2107
2108	if (num_active_crtcs != 1)
2109		wm->cxsr = false;
2110
2111	if (num_active_crtcs > 1)
2112		wm->level = VLV_WM_LEVEL_PM2;
2113
2114	for_each_intel_crtc(&dev_priv->drm, crtc) {
2115		const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2116		enum pipe pipe = crtc->pipe;
2117
2118		wm->pipe[pipe] = wm_state->wm[wm->level];
2119		if (crtc->active && wm->cxsr)
2120			wm->sr = wm_state->sr[wm->level];
2121
2122		wm->ddl[pipe].plane[PLANE_PRIMARY] = DDL_PRECISION_HIGH | 2;
2123		wm->ddl[pipe].plane[PLANE_SPRITE0] = DDL_PRECISION_HIGH | 2;
2124		wm->ddl[pipe].plane[PLANE_SPRITE1] = DDL_PRECISION_HIGH | 2;
2125		wm->ddl[pipe].plane[PLANE_CURSOR] = DDL_PRECISION_HIGH | 2;
2126	}
2127}
2128
2129static void vlv_program_watermarks(struct drm_i915_private *dev_priv)
2130{
2131	struct vlv_wm_values *old_wm = &dev_priv->wm.vlv;
2132	struct vlv_wm_values new_wm = {};
2133
2134	vlv_merge_wm(dev_priv, &new_wm);
2135
2136	if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
2137		return;
2138
2139	if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2140		chv_set_memory_dvfs(dev_priv, false);
2141
2142	if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2143		chv_set_memory_pm5(dev_priv, false);
2144
2145	if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
2146		_intel_set_memory_cxsr(dev_priv, false);
2147
2148	vlv_write_wm_values(dev_priv, &new_wm);
2149
2150	if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
2151		_intel_set_memory_cxsr(dev_priv, true);
2152
2153	if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2154		chv_set_memory_pm5(dev_priv, true);
2155
2156	if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2157		chv_set_memory_dvfs(dev_priv, true);
2158
2159	*old_wm = new_wm;
2160}
2161
2162static void vlv_initial_watermarks(struct intel_atomic_state *state,
2163				   struct intel_crtc_state *crtc_state)
2164{
2165	struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
2166	struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
2167
2168	mutex_lock(&dev_priv->wm.wm_mutex);
2169	crtc->wm.active.vlv = crtc_state->wm.vlv.intermediate;
2170	vlv_program_watermarks(dev_priv);
2171	mutex_unlock(&dev_priv->wm.wm_mutex);
2172}
2173
2174static void vlv_optimize_watermarks(struct intel_atomic_state *state,
2175				    struct intel_crtc_state *crtc_state)
2176{
2177	struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
2178	struct intel_crtc *intel_crtc = to_intel_crtc(crtc_state->base.crtc);
2179
2180	if (!crtc_state->wm.need_postvbl_update)
2181		return;
2182
2183	mutex_lock(&dev_priv->wm.wm_mutex);
2184	intel_crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
2185	vlv_program_watermarks(dev_priv);
2186	mutex_unlock(&dev_priv->wm.wm_mutex);
2187}
2188
2189static void i965_update_wm(struct intel_crtc *unused_crtc)
2190{
2191	struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2192	struct intel_crtc *crtc;
2193	int srwm = 1;
2194	int cursor_sr = 16;
2195	bool cxsr_enabled;
2196
2197	/* Calc sr entries for one plane configs */
2198	crtc = single_enabled_crtc(dev_priv);
2199	if (crtc) {
2200		/* self-refresh has much higher latency */
2201		static const int sr_latency_ns = 12000;
2202		const struct drm_display_mode *adjusted_mode =
2203			&crtc->config->base.adjusted_mode;
2204		const struct drm_framebuffer *fb =
2205			crtc->base.primary->state->fb;
2206		int clock = adjusted_mode->crtc_clock;
2207		int htotal = adjusted_mode->crtc_htotal;
2208		int hdisplay = crtc->config->pipe_src_w;
2209		int cpp = fb->format->cpp[0];
2210		int entries;
2211
2212		entries = intel_wm_method2(clock, htotal,
2213					   hdisplay, cpp, sr_latency_ns / 100);
2214		entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
2215		srwm = I965_FIFO_SIZE - entries;
2216		if (srwm < 0)
2217			srwm = 1;
2218		srwm &= 0x1ff;
2219		DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
2220			      entries, srwm);
2221
2222		entries = intel_wm_method2(clock, htotal,
2223					   crtc->base.cursor->state->crtc_w, 4,
2224					   sr_latency_ns / 100);
2225		entries = DIV_ROUND_UP(entries,
2226				       i965_cursor_wm_info.cacheline_size) +
2227			i965_cursor_wm_info.guard_size;
2228
2229		cursor_sr = i965_cursor_wm_info.fifo_size - entries;
2230		if (cursor_sr > i965_cursor_wm_info.max_wm)
2231			cursor_sr = i965_cursor_wm_info.max_wm;
2232
2233		DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
2234			      "cursor %d\n", srwm, cursor_sr);
2235
2236		cxsr_enabled = true;
2237	} else {
2238		cxsr_enabled = false;
2239		/* Turn off self refresh if both pipes are enabled */
2240		intel_set_memory_cxsr(dev_priv, false);
2241	}
2242
2243	DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
2244		      srwm);
2245
2246	/* 965 has limitations... */
2247	I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
2248		   FW_WM(8, CURSORB) |
2249		   FW_WM(8, PLANEB) |
2250		   FW_WM(8, PLANEA));
2251	I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
2252		   FW_WM(8, PLANEC_OLD));
2253	/* update cursor SR watermark */
2254	I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
2255
2256	if (cxsr_enabled)
2257		intel_set_memory_cxsr(dev_priv, true);
2258}
2259
2260#undef FW_WM
2261
2262static void i9xx_update_wm(struct intel_crtc *unused_crtc)
2263{
2264	struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2265	const struct intel_watermark_params *wm_info;
2266	uint32_t fwater_lo;
2267	uint32_t fwater_hi;
2268	int cwm, srwm = 1;
2269	int fifo_size;
2270	int planea_wm, planeb_wm;
2271	struct intel_crtc *crtc, *enabled = NULL;
2272
2273	if (IS_I945GM(dev_priv))
2274		wm_info = &i945_wm_info;
2275	else if (!IS_GEN2(dev_priv))
2276		wm_info = &i915_wm_info;
2277	else
2278		wm_info = &i830_a_wm_info;
2279
2280	fifo_size = dev_priv->display.get_fifo_size(dev_priv, PLANE_A);
2281	crtc = intel_get_crtc_for_plane(dev_priv, PLANE_A);
2282	if (intel_crtc_active(crtc)) {
2283		const struct drm_display_mode *adjusted_mode =
2284			&crtc->config->base.adjusted_mode;
2285		const struct drm_framebuffer *fb =
2286			crtc->base.primary->state->fb;
2287		int cpp;
2288
2289		if (IS_GEN2(dev_priv))
2290			cpp = 4;
2291		else
2292			cpp = fb->format->cpp[0];
2293
2294		planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2295					       wm_info, fifo_size, cpp,
2296					       pessimal_latency_ns);
2297		enabled = crtc;
2298	} else {
2299		planea_wm = fifo_size - wm_info->guard_size;
2300		if (planea_wm > (long)wm_info->max_wm)
2301			planea_wm = wm_info->max_wm;
2302	}
2303
2304	if (IS_GEN2(dev_priv))
2305		wm_info = &i830_bc_wm_info;
2306
2307	fifo_size = dev_priv->display.get_fifo_size(dev_priv, PLANE_B);
2308	crtc = intel_get_crtc_for_plane(dev_priv, PLANE_B);
2309	if (intel_crtc_active(crtc)) {
2310		const struct drm_display_mode *adjusted_mode =
2311			&crtc->config->base.adjusted_mode;
2312		const struct drm_framebuffer *fb =
2313			crtc->base.primary->state->fb;
2314		int cpp;
2315
2316		if (IS_GEN2(dev_priv))
2317			cpp = 4;
2318		else
2319			cpp = fb->format->cpp[0];
2320
2321		planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2322					       wm_info, fifo_size, cpp,
2323					       pessimal_latency_ns);
2324		if (enabled == NULL)
2325			enabled = crtc;
2326		else
2327			enabled = NULL;
2328	} else {
2329		planeb_wm = fifo_size - wm_info->guard_size;
2330		if (planeb_wm > (long)wm_info->max_wm)
2331			planeb_wm = wm_info->max_wm;
2332	}
2333
2334	DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
2335
2336	if (IS_I915GM(dev_priv) && enabled) {
2337		struct drm_i915_gem_object *obj;
2338
2339		obj = intel_fb_obj(enabled->base.primary->state->fb);
2340
2341		/* self-refresh seems busted with untiled */
2342		if (!i915_gem_object_is_tiled(obj))
2343			enabled = NULL;
2344	}
2345
2346	/*
2347	 * Overlay gets an aggressive default since video jitter is bad.
2348	 */
2349	cwm = 2;
2350
2351	/* Play safe and disable self-refresh before adjusting watermarks. */
2352	intel_set_memory_cxsr(dev_priv, false);
2353
2354	/* Calc sr entries for one plane configs */
2355	if (HAS_FW_BLC(dev_priv) && enabled) {
2356		/* self-refresh has much higher latency */
2357		static const int sr_latency_ns = 6000;
2358		const struct drm_display_mode *adjusted_mode =
2359			&enabled->config->base.adjusted_mode;
2360		const struct drm_framebuffer *fb =
2361			enabled->base.primary->state->fb;
2362		int clock = adjusted_mode->crtc_clock;
2363		int htotal = adjusted_mode->crtc_htotal;
2364		int hdisplay = enabled->config->pipe_src_w;
2365		int cpp;
2366		int entries;
2367
2368		if (IS_I915GM(dev_priv) || IS_I945GM(dev_priv))
2369			cpp = 4;
2370		else
2371			cpp = fb->format->cpp[0];
2372
2373		entries = intel_wm_method2(clock, htotal, hdisplay, cpp,
2374					   sr_latency_ns / 100);
2375		entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
2376		DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
2377		srwm = wm_info->fifo_size - entries;
2378		if (srwm < 0)
2379			srwm = 1;
2380
2381		if (IS_I945G(dev_priv) || IS_I945GM(dev_priv))
2382			I915_WRITE(FW_BLC_SELF,
2383				   FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
2384		else
2385			I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
2386	}
2387
2388	DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
2389		      planea_wm, planeb_wm, cwm, srwm);
2390
2391	fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
2392	fwater_hi = (cwm & 0x1f);
2393
2394	/* Set request length to 8 cachelines per fetch */
2395	fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
2396	fwater_hi = fwater_hi | (1 << 8);
2397
2398	I915_WRITE(FW_BLC, fwater_lo);
2399	I915_WRITE(FW_BLC2, fwater_hi);
2400
2401	if (enabled)
2402		intel_set_memory_cxsr(dev_priv, true);
2403}
2404
2405static void i845_update_wm(struct intel_crtc *unused_crtc)
2406{
2407	struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2408	struct intel_crtc *crtc;
2409	const struct drm_display_mode *adjusted_mode;
2410	uint32_t fwater_lo;
2411	int planea_wm;
2412
2413	crtc = single_enabled_crtc(dev_priv);
2414	if (crtc == NULL)
2415		return;
2416
2417	adjusted_mode = &crtc->config->base.adjusted_mode;
2418	planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2419				       &i845_wm_info,
2420				       dev_priv->display.get_fifo_size(dev_priv, PLANE_A),
2421				       4, pessimal_latency_ns);
2422	fwater_lo = I915_READ(FW_BLC) & ~0xfff;
2423	fwater_lo |= (3<<8) | planea_wm;
2424
2425	DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
2426
2427	I915_WRITE(FW_BLC, fwater_lo);
2428}
2429
2430/* latency must be in 0.1us units. */
2431static unsigned int ilk_wm_method1(unsigned int pixel_rate,
2432				   unsigned int cpp,
2433				   unsigned int latency)
2434{
2435	unsigned int ret;
2436
2437	ret = intel_wm_method1(pixel_rate, cpp, latency);
2438	ret = DIV_ROUND_UP(ret, 64) + 2;
2439
2440	return ret;
2441}
2442
2443/* latency must be in 0.1us units. */
2444static unsigned int ilk_wm_method2(unsigned int pixel_rate,
2445				   unsigned int htotal,
2446				   unsigned int width,
2447				   unsigned int cpp,
2448				   unsigned int latency)
2449{
2450	unsigned int ret;
2451
2452	ret = intel_wm_method2(pixel_rate, htotal,
2453			       width, cpp, latency);
2454	ret = DIV_ROUND_UP(ret, 64) + 2;
2455
2456	return ret;
2457}
2458
2459static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
2460			   uint8_t cpp)
2461{
2462	/*
2463	 * Neither of these should be possible since this function shouldn't be
2464	 * called if the CRTC is off or the plane is invisible.  But let's be
2465	 * extra paranoid to avoid a potential divide-by-zero if we screw up
2466	 * elsewhere in the driver.
2467	 */
2468	if (WARN_ON(!cpp))
2469		return 0;
2470	if (WARN_ON(!horiz_pixels))
2471		return 0;
2472
2473	return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
2474}
2475
2476struct ilk_wm_maximums {
2477	uint16_t pri;
2478	uint16_t spr;
2479	uint16_t cur;
2480	uint16_t fbc;
2481};
2482
2483/*
2484 * For both WM_PIPE and WM_LP.
2485 * mem_value must be in 0.1us units.
2486 */
2487static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
2488				   const struct intel_plane_state *pstate,
2489				   uint32_t mem_value,
2490				   bool is_lp)
2491{
2492	uint32_t method1, method2;
2493	int cpp;
2494
2495	if (!intel_wm_plane_visible(cstate, pstate))
2496		return 0;
2497
2498	cpp = pstate->base.fb->format->cpp[0];
2499
2500	method1 = ilk_wm_method1(cstate->pixel_rate, cpp, mem_value);
2501
2502	if (!is_lp)
2503		return method1;
2504
2505	method2 = ilk_wm_method2(cstate->pixel_rate,
2506				 cstate->base.adjusted_mode.crtc_htotal,
2507				 drm_rect_width(&pstate->base.dst),
2508				 cpp, mem_value);
2509
2510	return min(method1, method2);
2511}
2512
2513/*
2514 * For both WM_PIPE and WM_LP.
2515 * mem_value must be in 0.1us units.
2516 */
2517static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
2518				   const struct intel_plane_state *pstate,
2519				   uint32_t mem_value)
2520{
2521	uint32_t method1, method2;
2522	int cpp;
2523
2524	if (!intel_wm_plane_visible(cstate, pstate))
2525		return 0;
2526
2527	cpp = pstate->base.fb->format->cpp[0];
2528
2529	method1 = ilk_wm_method1(cstate->pixel_rate, cpp, mem_value);
2530	method2 = ilk_wm_method2(cstate->pixel_rate,
2531				 cstate->base.adjusted_mode.crtc_htotal,
2532				 drm_rect_width(&pstate->base.dst),
2533				 cpp, mem_value);
2534	return min(method1, method2);
2535}
2536
2537/*
2538 * For both WM_PIPE and WM_LP.
2539 * mem_value must be in 0.1us units.
2540 */
2541static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
2542				   const struct intel_plane_state *pstate,
2543				   uint32_t mem_value)
2544{
2545	int cpp;
2546
2547	if (!intel_wm_plane_visible(cstate, pstate))
2548		return 0;
2549
2550	cpp = pstate->base.fb->format->cpp[0];
2551
2552	return ilk_wm_method2(cstate->pixel_rate,
2553			      cstate->base.adjusted_mode.crtc_htotal,
2554			      pstate->base.crtc_w, cpp, mem_value);
2555}
2556
2557/* Only for WM_LP. */
2558static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
2559				   const struct intel_plane_state *pstate,
2560				   uint32_t pri_val)
2561{
2562	int cpp;
2563
2564	if (!intel_wm_plane_visible(cstate, pstate))
2565		return 0;
2566
2567	cpp = pstate->base.fb->format->cpp[0];
2568
2569	return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->base.dst), cpp);
2570}
2571
2572static unsigned int
2573ilk_display_fifo_size(const struct drm_i915_private *dev_priv)
2574{
2575	if (INTEL_GEN(dev_priv) >= 8)
2576		return 3072;
2577	else if (INTEL_GEN(dev_priv) >= 7)
2578		return 768;
2579	else
2580		return 512;
2581}
2582
2583static unsigned int
2584ilk_plane_wm_reg_max(const struct drm_i915_private *dev_priv,
2585		     int level, bool is_sprite)
2586{
2587	if (INTEL_GEN(dev_priv) >= 8)
2588		/* BDW primary/sprite plane watermarks */
2589		return level == 0 ? 255 : 2047;
2590	else if (INTEL_GEN(dev_priv) >= 7)
2591		/* IVB/HSW primary/sprite plane watermarks */
2592		return level == 0 ? 127 : 1023;
2593	else if (!is_sprite)
2594		/* ILK/SNB primary plane watermarks */
2595		return level == 0 ? 127 : 511;
2596	else
2597		/* ILK/SNB sprite plane watermarks */
2598		return level == 0 ? 63 : 255;
2599}
2600
2601static unsigned int
2602ilk_cursor_wm_reg_max(const struct drm_i915_private *dev_priv, int level)
2603{
2604	if (INTEL_GEN(dev_priv) >= 7)
2605		return level == 0 ? 63 : 255;
2606	else
2607		return level == 0 ? 31 : 63;
2608}
2609
2610static unsigned int ilk_fbc_wm_reg_max(const struct drm_i915_private *dev_priv)
2611{
2612	if (INTEL_GEN(dev_priv) >= 8)
2613		return 31;
2614	else
2615		return 15;
2616}
2617
2618/* Calculate the maximum primary/sprite plane watermark */
2619static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
2620				     int level,
2621				     const struct intel_wm_config *config,
2622				     enum intel_ddb_partitioning ddb_partitioning,
2623				     bool is_sprite)
2624{
2625	struct drm_i915_private *dev_priv = to_i915(dev);
2626	unsigned int fifo_size = ilk_display_fifo_size(dev_priv);
2627
2628	/* if sprites aren't enabled, sprites get nothing */
2629	if (is_sprite && !config->sprites_enabled)
2630		return 0;
2631
2632	/* HSW allows LP1+ watermarks even with multiple pipes */
2633	if (level == 0 || config->num_pipes_active > 1) {
2634		fifo_size /= INTEL_INFO(dev_priv)->num_pipes;
2635
2636		/*
2637		 * For some reason the non self refresh
2638		 * FIFO size is only half of the self
2639		 * refresh FIFO size on ILK/SNB.
2640		 */
2641		if (INTEL_GEN(dev_priv) <= 6)
2642			fifo_size /= 2;
2643	}
2644
2645	if (config->sprites_enabled) {
2646		/* level 0 is always calculated with 1:1 split */
2647		if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
2648			if (is_sprite)
2649				fifo_size *= 5;
2650			fifo_size /= 6;
2651		} else {
2652			fifo_size /= 2;
2653		}
2654	}
2655
2656	/* clamp to max that the registers can hold */
2657	return min(fifo_size, ilk_plane_wm_reg_max(dev_priv, level, is_sprite));
2658}
2659
2660/* Calculate the maximum cursor plane watermark */
2661static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
2662				      int level,
2663				      const struct intel_wm_config *config)
2664{
2665	/* HSW LP1+ watermarks w/ multiple pipes */
2666	if (level > 0 && config->num_pipes_active > 1)
2667		return 64;
2668
2669	/* otherwise just report max that registers can hold */
2670	return ilk_cursor_wm_reg_max(to_i915(dev), level);
2671}
2672
2673static void ilk_compute_wm_maximums(const struct drm_device *dev,
2674				    int level,
2675				    const struct intel_wm_config *config,
2676				    enum intel_ddb_partitioning ddb_partitioning,
2677				    struct ilk_wm_maximums *max)
2678{
2679	max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
2680	max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
2681	max->cur = ilk_cursor_wm_max(dev, level, config);
2682	max->fbc = ilk_fbc_wm_reg_max(to_i915(dev));
2683}
2684
2685static void ilk_compute_wm_reg_maximums(const struct drm_i915_private *dev_priv,
2686					int level,
2687					struct ilk_wm_maximums *max)
2688{
2689	max->pri = ilk_plane_wm_reg_max(dev_priv, level, false);
2690	max->spr = ilk_plane_wm_reg_max(dev_priv, level, true);
2691	max->cur = ilk_cursor_wm_reg_max(dev_priv, level);
2692	max->fbc = ilk_fbc_wm_reg_max(dev_priv);
2693}
2694
2695static bool ilk_validate_wm_level(int level,
2696				  const struct ilk_wm_maximums *max,
2697				  struct intel_wm_level *result)
2698{
2699	bool ret;
2700
2701	/* already determined to be invalid? */
2702	if (!result->enable)
2703		return false;
2704
2705	result->enable = result->pri_val <= max->pri &&
2706			 result->spr_val <= max->spr &&
2707			 result->cur_val <= max->cur;
2708
2709	ret = result->enable;
2710
2711	/*
2712	 * HACK until we can pre-compute everything,
2713	 * and thus fail gracefully if LP0 watermarks
2714	 * are exceeded...
2715	 */
2716	if (level == 0 && !result->enable) {
2717		if (result->pri_val > max->pri)
2718			DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2719				      level, result->pri_val, max->pri);
2720		if (result->spr_val > max->spr)
2721			DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2722				      level, result->spr_val, max->spr);
2723		if (result->cur_val > max->cur)
2724			DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2725				      level, result->cur_val, max->cur);
2726
2727		result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2728		result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2729		result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2730		result->enable = true;
2731	}
2732
2733	return ret;
2734}
2735
2736static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2737				 const struct intel_crtc *intel_crtc,
2738				 int level,
2739				 struct intel_crtc_state *cstate,
2740				 const struct intel_plane_state *pristate,
2741				 const struct intel_plane_state *sprstate,
2742				 const struct intel_plane_state *curstate,
2743				 struct intel_wm_level *result)
2744{
2745	uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2746	uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2747	uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2748
2749	/* WM1+ latency values stored in 0.5us units */
2750	if (level > 0) {
2751		pri_latency *= 5;
2752		spr_latency *= 5;
2753		cur_latency *= 5;
2754	}
2755
2756	if (pristate) {
2757		result->pri_val = ilk_compute_pri_wm(cstate, pristate,
2758						     pri_latency, level);
2759		result->fbc_val = ilk_compute_fbc_wm(cstate, pristate, result->pri_val);
2760	}
2761
2762	if (sprstate)
2763		result->spr_val = ilk_compute_spr_wm(cstate, sprstate, spr_latency);
2764
2765	if (curstate)
2766		result->cur_val = ilk_compute_cur_wm(cstate, curstate, cur_latency);
2767
2768	result->enable = true;
2769}
2770
2771static uint32_t
2772hsw_compute_linetime_wm(const struct intel_crtc_state *cstate)
2773{
2774	const struct intel_atomic_state *intel_state =
2775		to_intel_atomic_state(cstate->base.state);
2776	const struct drm_display_mode *adjusted_mode =
2777		&cstate->base.adjusted_mode;
2778	u32 linetime, ips_linetime;
2779
2780	if (!cstate->base.active)
2781		return 0;
2782	if (WARN_ON(adjusted_mode->crtc_clock == 0))
2783		return 0;
2784	if (WARN_ON(intel_state->cdclk.logical.cdclk == 0))
2785		return 0;
2786
2787	/* The WM are computed with base on how long it takes to fill a single
2788	 * row at the given clock rate, multiplied by 8.
2789	 * */
2790	linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2791				     adjusted_mode->crtc_clock);
2792	ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2793					 intel_state->cdclk.logical.cdclk);
2794
2795	return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2796	       PIPE_WM_LINETIME_TIME(linetime);
2797}
2798
2799static void intel_read_wm_latency(struct drm_i915_private *dev_priv,
2800				  uint16_t wm[8])
2801{
2802	if (INTEL_GEN(dev_priv) >= 9) {
2803		uint32_t val;
2804		int ret, i;
2805		int level, max_level = ilk_wm_max_level(dev_priv);
2806
2807		/* read the first set of memory latencies[0:3] */
2808		val = 0; /* data0 to be programmed to 0 for first set */
2809		mutex_lock(&dev_priv->pcu_lock);
2810		ret = sandybridge_pcode_read(dev_priv,
2811					     GEN9_PCODE_READ_MEM_LATENCY,
2812					     &val);
2813		mutex_unlock(&dev_priv->pcu_lock);
2814
2815		if (ret) {
2816			DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2817			return;
2818		}
2819
2820		wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2821		wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2822				GEN9_MEM_LATENCY_LEVEL_MASK;
2823		wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2824				GEN9_MEM_LATENCY_LEVEL_MASK;
2825		wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2826				GEN9_MEM_LATENCY_LEVEL_MASK;
2827
2828		/* read the second set of memory latencies[4:7] */
2829		val = 1; /* data0 to be programmed to 1 for second set */
2830		mutex_lock(&dev_priv->pcu_lock);
2831		ret = sandybridge_pcode_read(dev_priv,
2832					     GEN9_PCODE_READ_MEM_LATENCY,
2833					     &val);
2834		mutex_unlock(&dev_priv->pcu_lock);
2835		if (ret) {
2836			DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2837			return;
2838		}
2839
2840		wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2841		wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2842				GEN9_MEM_LATENCY_LEVEL_MASK;
2843		wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2844				GEN9_MEM_LATENCY_LEVEL_MASK;
2845		wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2846				GEN9_MEM_LATENCY_LEVEL_MASK;
2847
2848		/*
2849		 * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2850		 * need to be disabled. We make sure to sanitize the values out
2851		 * of the punit to satisfy this requirement.
2852		 */
2853		for (level = 1; level <= max_level; level++) {
2854			if (wm[level] == 0) {
2855				for (i = level + 1; i <= max_level; i++)
2856					wm[i] = 0;
2857				break;
2858			}
2859		}
2860
2861		/*
2862		 * WaWmMemoryReadLatency:skl+,glk
2863		 *
2864		 * punit doesn't take into account the read latency so we need
2865		 * to add 2us to the various latency levels we retrieve from the
2866		 * punit when level 0 response data us 0us.
2867		 */
2868		if (wm[0] == 0) {
2869			wm[0] += 2;
2870			for (level = 1; level <= max_level; level++) {
2871				if (wm[level] == 0)
2872					break;
2873				wm[level] += 2;
2874			}
2875		}
2876
2877	} else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2878		uint64_t sskpd = I915_READ64(MCH_SSKPD);
2879
2880		wm[0] = (sskpd >> 56) & 0xFF;
2881		if (wm[0] == 0)
2882			wm[0] = sskpd & 0xF;
2883		wm[1] = (sskpd >> 4) & 0xFF;
2884		wm[2] = (sskpd >> 12) & 0xFF;
2885		wm[3] = (sskpd >> 20) & 0x1FF;
2886		wm[4] = (sskpd >> 32) & 0x1FF;
2887	} else if (INTEL_GEN(dev_priv) >= 6) {
2888		uint32_t sskpd = I915_READ(MCH_SSKPD);
2889
2890		wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2891		wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2892		wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2893		wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2894	} else if (INTEL_GEN(dev_priv) >= 5) {
2895		uint32_t mltr = I915_READ(MLTR_ILK);
2896
2897		/* ILK primary LP0 latency is 700 ns */
2898		wm[0] = 7;
2899		wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2900		wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2901	} else {
2902		MISSING_CASE(INTEL_DEVID(dev_priv));
2903	}
2904}
2905
2906static void intel_fixup_spr_wm_latency(struct drm_i915_private *dev_priv,
2907				       uint16_t wm[5])
2908{
2909	/* ILK sprite LP0 latency is 1300 ns */
2910	if (IS_GEN5(dev_priv))
2911		wm[0] = 13;
2912}
2913
2914static void intel_fixup_cur_wm_latency(struct drm_i915_private *dev_priv,
2915				       uint16_t wm[5])
2916{
2917	/* ILK cursor LP0 latency is 1300 ns */
2918	if (IS_GEN5(dev_priv))
2919		wm[0] = 13;
2920}
2921
2922int ilk_wm_max_level(const struct drm_i915_private *dev_priv)
2923{
2924	/* how many WM levels are we expecting */
2925	if (INTEL_GEN(dev_priv) >= 9)
2926		return 7;
2927	else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2928		return 4;
2929	else if (INTEL_GEN(dev_priv) >= 6)
2930		return 3;
2931	else
2932		return 2;
2933}
2934
2935static void intel_print_wm_latency(struct drm_i915_private *dev_priv,
2936				   const char *name,
2937				   const uint16_t wm[8])
2938{
2939	int level, max_level = ilk_wm_max_level(dev_priv);
2940
2941	for (level = 0; level <= max_level; level++) {
2942		unsigned int latency = wm[level];
2943
2944		if (latency == 0) {
2945			DRM_ERROR("%s WM%d latency not provided\n",
2946				  name, level);
2947			continue;
2948		}
2949
2950		/*
2951		 * - latencies are in us on gen9.
2952		 * - before then, WM1+ latency values are in 0.5us units
2953		 */
2954		if (INTEL_GEN(dev_priv) >= 9)
2955			latency *= 10;
2956		else if (level > 0)
2957			latency *= 5;
2958
2959		DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2960			      name, level, wm[level],
2961			      latency / 10, latency % 10);
2962	}
2963}
2964
2965static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2966				    uint16_t wm[5], uint16_t min)
2967{
2968	int level, max_level = ilk_wm_max_level(dev_priv);
2969
2970	if (wm[0] >= min)
2971		return false;
2972
2973	wm[0] = max(wm[0], min);
2974	for (level = 1; level <= max_level; level++)
2975		wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2976
2977	return true;
2978}
2979
2980static void snb_wm_latency_quirk(struct drm_i915_private *dev_priv)
2981{
2982	bool changed;
2983
2984	/*
2985	 * The BIOS provided WM memory latency values are often
2986	 * inadequate for high resolution displays. Adjust them.
2987	 */
2988	changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2989		ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2990		ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2991
2992	if (!changed)
2993		return;
2994
2995	DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2996	intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2997	intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2998	intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2999}
3000
3001static void ilk_setup_wm_latency(struct drm_i915_private *dev_priv)
3002{
3003	intel_read_wm_latency(dev_priv, dev_priv->wm.pri_latency);
3004
3005	memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
3006	       sizeof(dev_priv->wm.pri_latency));
3007	memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
3008	       sizeof(dev_priv->wm.pri_latency));
3009
3010	intel_fixup_spr_wm_latency(dev_priv, dev_priv->wm.spr_latency);
3011	intel_fixup_cur_wm_latency(dev_priv, dev_priv->wm.cur_latency);
3012
3013	intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
3014	intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
3015	intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
3016
3017	if (IS_GEN6(dev_priv))
3018		snb_wm_latency_quirk(dev_priv);
3019}
3020
3021static void skl_setup_wm_latency(struct drm_i915_private *dev_priv)
3022{
3023	intel_read_wm_latency(dev_priv, dev_priv->wm.skl_latency);
3024	intel_print_wm_latency(dev_priv, "Gen9 Plane", dev_priv->wm.skl_latency);
3025}
3026
3027static bool ilk_validate_pipe_wm(struct drm_device *dev,
3028				 struct intel_pipe_wm *pipe_wm)
3029{
3030	/* LP0 watermark maximums depend on this pipe alone */
3031	const struct intel_wm_config config = {
3032		.num_pipes_active = 1,
3033		.sprites_enabled = pipe_wm->sprites_enabled,
3034		.sprites_scaled = pipe_wm->sprites_scaled,
3035	};
3036	struct ilk_wm_maximums max;
3037
3038	/* LP0 watermarks always use 1/2 DDB partitioning */
3039	ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
3040
3041	/* At least LP0 must be valid */
3042	if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
3043		DRM_DEBUG_KMS("LP0 watermark invalid\n");
3044		return false;
3045	}
3046
3047	return true;
3048}
3049
3050/* Compute new watermarks for the pipe */
3051static int ilk_compute_pipe_wm(struct intel_crtc_state *cstate)
3052{
3053	struct drm_atomic_state *state = cstate->base.state;
3054	struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
3055	struct intel_pipe_wm *pipe_wm;
3056	struct drm_device *dev = state->dev;
3057	const struct drm_i915_private *dev_priv = to_i915(dev);
3058	struct drm_plane *plane;
3059	const struct drm_plane_state *plane_state;
3060	const struct intel_plane_state *pristate = NULL;
3061	const struct intel_plane_state *sprstate = NULL;
3062	const struct intel_plane_state *curstate = NULL;
3063	int level, max_level = ilk_wm_max_level(dev_priv), usable_level;
3064	struct ilk_wm_maximums max;
3065
3066	pipe_wm = &cstate->wm.ilk.optimal;
3067
3068	drm_atomic_crtc_state_for_each_plane_state(plane, plane_state, &cstate->base) {
3069		const struct intel_plane_state *ps = to_intel_plane_state(plane_state);
3070
3071		if (plane->type == DRM_PLANE_TYPE_PRIMARY)
3072			pristate = ps;
3073		else if (plane->type == DRM_PLANE_TYPE_OVERLAY)
3074			sprstate = ps;
3075		else if (plane->type == DRM_PLANE_TYPE_CURSOR)
3076			curstate = ps;
3077	}
3078
3079	pipe_wm->pipe_enabled = cstate->base.active;
3080	if (sprstate) {
3081		pipe_wm->sprites_enabled = sprstate->base.visible;
3082		pipe_wm->sprites_scaled = sprstate->base.visible &&
3083			(drm_rect_width(&sprstate->base.dst) != drm_rect_width(&sprstate->base.src) >> 16 ||
3084			 drm_rect_height(&sprstate->base.dst) != drm_rect_height(&sprstate->base.src) >> 16);
3085	}
3086
3087	usable_level = max_level;
3088
3089	/* ILK/SNB: LP2+ watermarks only w/o sprites */
3090	if (INTEL_GEN(dev_priv) <= 6 && pipe_wm->sprites_enabled)
3091		usable_level = 1;
3092
3093	/* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
3094	if (pipe_wm->sprites_scaled)
3095		usable_level = 0;
3096
3097	memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
3098	ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate,
3099			     pristate, sprstate, curstate, &pipe_wm->wm[0]);
3100
3101	if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
3102		pipe_wm->linetime = hsw_compute_linetime_wm(cstate);
3103
3104	if (!ilk_validate_pipe_wm(dev, pipe_wm))
3105		return -EINVAL;
3106
3107	ilk_compute_wm_reg_maximums(dev_priv, 1, &max);
3108
3109	for (level = 1; level <= usable_level; level++) {
3110		struct intel_wm_level *wm = &pipe_wm->wm[level];
3111
3112		ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate,
3113				     pristate, sprstate, curstate, wm);
3114
3115		/*
3116		 * Disable any watermark level that exceeds the
3117		 * register maximums since such watermarks are
3118		 * always invalid.
3119		 */
3120		if (!ilk_validate_wm_level(level, &max, wm)) {
3121			memset(wm, 0, sizeof(*wm));
3122			break;
3123		}
3124	}
3125
3126	return 0;
3127}
3128
3129/*
3130 * Build a set of 'intermediate' watermark values that satisfy both the old
3131 * state and the new state.  These can be programmed to the hardware
3132 * immediately.
3133 */
3134static int ilk_compute_intermediate_wm(struct drm_device *dev,
3135				       struct intel_crtc *intel_crtc,
3136				       struct intel_crtc_state *newstate)
3137{
3138	struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
3139	struct intel_atomic_state *intel_state =
3140		to_intel_atomic_state(newstate->base.state);
3141	const struct intel_crtc_state *oldstate =
3142		intel_atomic_get_old_crtc_state(intel_state, intel_crtc);
3143	const struct intel_pipe_wm *b = &oldstate->wm.ilk.optimal;
3144	int level, max_level = ilk_wm_max_level(to_i915(dev));
3145
3146	/*
3147	 * Start with the final, target watermarks, then combine with the
3148	 * currently active watermarks to get values that are safe both before
3149	 * and after the vblank.
3150	 */
3151	*a = newstate->wm.ilk.optimal;
3152	if (!newstate->base.active || drm_atomic_crtc_needs_modeset(&newstate->base))
3153		return 0;
3154
3155	a->pipe_enabled |= b->pipe_enabled;
3156	a->sprites_enabled |= b->sprites_enabled;
3157	a->sprites_scaled |= b->sprites_scaled;
3158
3159	for (level = 0; level <= max_level; level++) {
3160		struct intel_wm_level *a_wm = &a->wm[level];
3161		const struct intel_wm_level *b_wm = &b->wm[level];
3162
3163		a_wm->enable &= b_wm->enable;
3164		a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
3165		a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
3166		a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
3167		a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
3168	}
3169
3170	/*
3171	 * We need to make sure that these merged watermark values are
3172	 * actually a valid configuration themselves.  If they're not,
3173	 * there's no safe way to transition from the old state to
3174	 * the new state, so we need to fail the atomic transaction.
3175	 */
3176	if (!ilk_validate_pipe_wm(dev, a))
3177		return -EINVAL;
3178
3179	/*
3180	 * If our intermediate WM are identical to the final WM, then we can
3181	 * omit the post-vblank programming; only update if it's different.
3182	 */
3183	if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) != 0)
3184		newstate->wm.need_postvbl_update = true;
3185
3186	return 0;
3187}
3188
3189/*
3190 * Merge the watermarks from all active pipes for a specific level.
3191 */
3192static void ilk_merge_wm_level(struct drm_device *dev,
3193			       int level,
3194			       struct intel_wm_level *ret_wm)
3195{
3196	const struct intel_crtc *intel_crtc;
3197
3198	ret_wm->enable = true;
3199
3200	for_each_intel_crtc(dev, intel_crtc) {
3201		const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
3202		const struct intel_wm_level *wm = &active->wm[level];
3203
3204		if (!active->pipe_enabled)
3205			continue;
3206
3207		/*
3208		 * The watermark values may have been used in the past,
3209		 * so we must maintain them in the registers for some
3210		 * time even if the level is now disabled.
3211		 */
3212		if (!wm->enable)
3213			ret_wm->enable = false;
3214
3215		ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
3216		ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
3217		ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
3218		ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
3219	}
3220}
3221
3222/*
3223 * Merge all low power watermarks for all active pipes.
3224 */
3225static void ilk_wm_merge(struct drm_device *dev,
3226			 const struct intel_wm_config *config,
3227			 const struct ilk_wm_maximums *max,
3228			 struct intel_pipe_wm *merged)
3229{
3230	struct drm_i915_private *dev_priv = to_i915(dev);
3231	int level, max_level = ilk_wm_max_level(dev_priv);
3232	int last_enabled_level = max_level;
3233
3234	/* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
3235	if ((INTEL_GEN(dev_priv) <= 6 || IS_IVYBRIDGE(dev_priv)) &&
3236	    config->num_pipes_active > 1)
3237		last_enabled_level = 0;
3238
3239	/* ILK: FBC WM must be disabled always */
3240	merged->fbc_wm_enabled = INTEL_GEN(dev_priv) >= 6;
3241
3242	/* merge each WM1+ level */
3243	for (level = 1; level <= max_level; level++) {
3244		struct intel_wm_level *wm = &merged->wm[level];
3245
3246		ilk_merge_wm_level(dev, level, wm);
3247
3248		if (level > last_enabled_level)
3249			wm->enable = false;
3250		else if (!ilk_validate_wm_level(level, max, wm))
3251			/* make sure all following levels get disabled */
3252			last_enabled_level = level - 1;
3253
3254		/*
3255		 * The spec says it is preferred to disable
3256		 * FBC WMs instead of disabling a WM level.
3257		 */
3258		if (wm->fbc_val > max->fbc) {
3259			if (wm->enable)
3260				merged->fbc_wm_enabled = false;
3261			wm->fbc_val = 0;
3262		}
3263	}
3264
3265	/* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
3266	/*
3267	 * FIXME this is racy. FBC might get enabled later.
3268	 * What we should check here is whether FBC can be
3269	 * enabled sometime later.
3270	 */
3271	if (IS_GEN5(dev_priv) && !merged->fbc_wm_enabled &&
3272	    intel_fbc_is_active(dev_priv)) {
3273		for (level = 2; level <= max_level; level++) {
3274			struct intel_wm_level *wm = &merged->wm[level];
3275
3276			wm->enable = false;
3277		}
3278	}
3279}
3280
3281static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
3282{
3283	/* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
3284	return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
3285}
3286
3287/* The value we need to program into the WM_LPx latency field */
3288static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
3289{
3290	struct drm_i915_private *dev_priv = to_i915(dev);
3291
3292	if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
3293		return 2 * level;
3294	else
3295		return dev_priv->wm.pri_latency[level];
3296}
3297
3298static void ilk_compute_wm_results(struct drm_device *dev,
3299				   const struct intel_pipe_wm *merged,
3300				   enum intel_ddb_partitioning partitioning,
3301				   struct ilk_wm_values *results)
3302{
3303	struct drm_i915_private *dev_priv = to_i915(dev);
3304	struct intel_crtc *intel_crtc;
3305	int level, wm_lp;
3306
3307	results->enable_fbc_wm = merged->fbc_wm_enabled;
3308	results->partitioning = partitioning;
3309
3310	/* LP1+ register values */
3311	for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3312		const struct intel_wm_level *r;
3313
3314		level = ilk_wm_lp_to_level(wm_lp, merged);
3315
3316		r = &merged->wm[level];
3317
3318		/*
3319		 * Maintain the watermark values even if the level is
3320		 * disabled. Doing otherwise could cause underruns.
3321		 */
3322		results->wm_lp[wm_lp - 1] =
3323			(ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
3324			(r->pri_val << WM1_LP_SR_SHIFT) |
3325			r->cur_val;
3326
3327		if (r->enable)
3328			results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
3329
3330		if (INTEL_GEN(dev_priv) >= 8)
3331			results->wm_lp[wm_lp - 1] |=
3332				r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
3333		else
3334			results->wm_lp[wm_lp - 1] |=
3335				r->fbc_val << WM1_LP_FBC_SHIFT;
3336
3337		/*
3338		 * Always set WM1S_LP_EN when spr_val != 0, even if the
3339		 * level is disabled. Doing otherwise could cause underruns.
3340		 */
3341		if (INTEL_GEN(dev_priv) <= 6 && r->spr_val) {
3342			WARN_ON(wm_lp != 1);
3343			results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
3344		} else
3345			results->wm_lp_spr[wm_lp - 1] = r->spr_val;
3346	}
3347
3348	/* LP0 register values */
3349	for_each_intel_crtc(dev, intel_crtc) {
3350		enum pipe pipe = intel_crtc->pipe;
3351		const struct intel_wm_level *r =
3352			&intel_crtc->wm.active.ilk.wm[0];
3353
3354		if (WARN_ON(!r->enable))
3355			continue;
3356
3357		results->wm_linetime[pipe] = intel_crtc->wm.active.ilk.linetime;
3358
3359		results->wm_pipe[pipe] =
3360			(r->pri_val << WM0_PIPE_PLANE_SHIFT) |
3361			(r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
3362			r->cur_val;
3363	}
3364}
3365
3366/* Find the result with the highest level enabled. Check for enable_fbc_wm in
3367 * case both are at the same level. Prefer r1 in case they're the same. */
3368static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
3369						  struct intel_pipe_wm *r1,
3370						  struct intel_pipe_wm *r2)
3371{
3372	int level, max_level = ilk_wm_max_level(to_i915(dev));
3373	int level1 = 0, level2 = 0;
3374
3375	for (level = 1; level <= max_level; level++) {
3376		if (r1->wm[level].enable)
3377			level1 = level;
3378		if (r2->wm[level].enable)
3379			level2 = level;
3380	}
3381
3382	if (level1 == level2) {
3383		if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
3384			return r2;
3385		else
3386			return r1;
3387	} else if (level1 > level2) {
3388		return r1;
3389	} else {
3390		return r2;
3391	}
3392}
3393
3394/* dirty bits used to track which watermarks need changes */
3395#define WM_DIRTY_PIPE(pipe) (1 << (pipe))
3396#define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
3397#define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
3398#define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
3399#define WM_DIRTY_FBC (1 << 24)
3400#define WM_DIRTY_DDB (1 << 25)
3401
3402static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
3403					 const struct ilk_wm_values *old,
3404					 const struct ilk_wm_values *new)
3405{
3406	unsigned int dirty = 0;
3407	enum pipe pipe;
3408	int wm_lp;
3409
3410	for_each_pipe(dev_priv, pipe) {
3411		if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
3412			dirty |= WM_DIRTY_LINETIME(pipe);
3413			/* Must disable LP1+ watermarks too */
3414			dirty |= WM_DIRTY_LP_ALL;
3415		}
3416
3417		if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
3418			dirty |= WM_DIRTY_PIPE(pipe);
3419			/* Must disable LP1+ watermarks too */
3420			dirty |= WM_DIRTY_LP_ALL;
3421		}
3422	}
3423
3424	if (old->enable_fbc_wm != new->enable_fbc_wm) {
3425		dirty |= WM_DIRTY_FBC;
3426		/* Must disable LP1+ watermarks too */
3427		dirty |= WM_DIRTY_LP_ALL;
3428	}
3429
3430	if (old->partitioning != new->partitioning) {
3431		dirty |= WM_DIRTY_DDB;
3432		/* Must disable LP1+ watermarks too */
3433		dirty |= WM_DIRTY_LP_ALL;
3434	}
3435
3436	/* LP1+ watermarks already deemed dirty, no need to continue */
3437	if (dirty & WM_DIRTY_LP_ALL)
3438		return dirty;
3439
3440	/* Find the lowest numbered LP1+ watermark in need of an update... */
3441	for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3442		if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
3443		    old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
3444			break;
3445	}
3446
3447	/* ...and mark it and all higher numbered LP1+ watermarks as dirty */
3448	for (; wm_lp <= 3; wm_lp++)
3449		dirty |= WM_DIRTY_LP(wm_lp);
3450
3451	return dirty;
3452}
3453
3454static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
3455			       unsigned int dirty)
3456{
3457	struct ilk_wm_values *previous = &dev_priv->wm.hw;
3458	bool changed = false;
3459
3460	if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
3461		previous->wm_lp[2] &= ~WM1_LP_SR_EN;
3462		I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
3463		changed = true;
3464	}
3465	if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
3466		previous->wm_lp[1] &= ~WM1_LP_SR_EN;
3467		I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
3468		changed = true;
3469	}
3470	if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
3471		previous->wm_lp[0] &= ~WM1_LP_SR_EN;
3472		I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
3473		changed = true;
3474	}
3475
3476	/*
3477	 * Don't touch WM1S_LP_EN here.
3478	 * Doing so could cause underruns.
3479	 */
3480
3481	return changed;
3482}
3483
3484/*
3485 * The spec says we shouldn't write when we don't need, because every write
3486 * causes WMs to be re-evaluated, expending some power.
3487 */
3488static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
3489				struct ilk_wm_values *results)
3490{
3491	struct ilk_wm_values *previous = &dev_priv->wm.hw;
3492	unsigned int dirty;
3493	uint32_t val;
3494
3495	dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
3496	if (!dirty)
3497		return;
3498
3499	_ilk_disable_lp_wm(dev_priv, dirty);
3500
3501	if (dirty & WM_DIRTY_PIPE(PIPE_A))
3502		I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
3503	if (dirty & WM_DIRTY_PIPE(PIPE_B))
3504		I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
3505	if (dirty & WM_DIRTY_PIPE(PIPE_C))
3506		I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
3507
3508	if (dirty & WM_DIRTY_LINETIME(PIPE_A))
3509		I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
3510	if (dirty & WM_DIRTY_LINETIME(PIPE_B))
3511		I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
3512	if (dirty & WM_DIRTY_LINETIME(PIPE_C))
3513		I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
3514
3515	if (dirty & WM_DIRTY_DDB) {
3516		if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
3517			val = I915_READ(WM_MISC);
3518			if (results->partitioning == INTEL_DDB_PART_1_2)
3519				val &= ~WM_MISC_DATA_PARTITION_5_6;
3520			else
3521				val |= WM_MISC_DATA_PARTITION_5_6;
3522			I915_WRITE(WM_MISC, val);
3523		} else {
3524			val = I915_READ(DISP_ARB_CTL2);
3525			if (results->partitioning == INTEL_DDB_PART_1_2)
3526				val &= ~DISP_DATA_PARTITION_5_6;
3527			else
3528				val |= DISP_DATA_PARTITION_5_6;
3529			I915_WRITE(DISP_ARB_CTL2, val);
3530		}
3531	}
3532
3533	if (dirty & WM_DIRTY_FBC) {
3534		val = I915_READ(DISP_ARB_CTL);
3535		if (results->enable_fbc_wm)
3536			val &= ~DISP_FBC_WM_DIS;
3537		else
3538			val |= DISP_FBC_WM_DIS;
3539		I915_WRITE(DISP_ARB_CTL, val);
3540	}
3541
3542	if (dirty & WM_DIRTY_LP(1) &&
3543	    previous->wm_lp_spr[0] != results->wm_lp_spr[0])
3544		I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
3545
3546	if (INTEL_GEN(dev_priv) >= 7) {
3547		if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
3548			I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
3549		if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
3550			I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
3551	}
3552
3553	if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
3554		I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
3555	if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
3556		I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
3557	if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
3558		I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
3559
3560	dev_priv->wm.hw = *results;
3561}
3562
3563bool ilk_disable_lp_wm(struct drm_device *dev)
3564{
3565	struct drm_i915_private *dev_priv = to_i915(dev);
3566
3567	return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
3568}
3569
3570/*
3571 * FIXME: We still don't have the proper code detect if we need to apply the WA,
3572 * so assume we'll always need it in order to avoid underruns.
3573 */
3574static bool skl_needs_memory_bw_wa(struct intel_atomic_state *state)
3575{
3576	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
3577
3578	if (IS_GEN9_BC(dev_priv) || IS_BROXTON(dev_priv))
3579		return true;
3580
3581	return false;
3582}
3583
3584static bool
3585intel_has_sagv(struct drm_i915_private *dev_priv)
3586{
3587	if (IS_KABYLAKE(dev_priv) || IS_COFFEELAKE(dev_priv) ||
3588	    IS_CANNONLAKE(dev_priv))
3589		return true;
3590
3591	if (IS_SKYLAKE(dev_priv) &&
3592	    dev_priv->sagv_status != I915_SAGV_NOT_CONTROLLED)
3593		return true;
3594
3595	return false;
3596}
3597
3598/*
3599 * SAGV dynamically adjusts the system agent voltage and clock frequencies
3600 * depending on power and performance requirements. The display engine access
3601 * to system memory is blocked during the adjustment time. Because of the
3602 * blocking time, having this enabled can cause full system hangs and/or pipe
3603 * underruns if we don't meet all of the following requirements:
3604 *
3605 *  - <= 1 pipe enabled
3606 *  - All planes can enable watermarks for latencies >= SAGV engine block time
3607 *  - We're not using an interlaced display configuration
3608 */
3609int
3610intel_enable_sagv(struct drm_i915_private *dev_priv)
3611{
3612	int ret;
3613
3614	if (!intel_has_sagv(dev_priv))
3615		return 0;
3616
3617	if (dev_priv->sagv_status == I915_SAGV_ENABLED)
3618		return 0;
3619
3620	DRM_DEBUG_KMS("Enabling the SAGV\n");
3621	mutex_lock(&dev_priv->pcu_lock);
3622
3623	ret = sandybridge_pcode_write(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3624				      GEN9_SAGV_ENABLE);
3625
3626	/* We don't need to wait for the SAGV when enabling */
3627	mutex_unlock(&dev_priv->pcu_lock);
3628
3629	/*
3630	 * Some skl systems, pre-release machines in particular,
3631	 * don't actually have an SAGV.
3632	 */
3633	if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3634		DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
3635		dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3636		return 0;
3637	} else if (ret < 0) {
3638		DRM_ERROR("Failed to enable the SAGV\n");
3639		return ret;
3640	}
3641
3642	dev_priv->sagv_status = I915_SAGV_ENABLED;
3643	return 0;
3644}
3645
3646int
3647intel_disable_sagv(struct drm_i915_private *dev_priv)
3648{
3649	int ret;
3650
3651	if (!intel_has_sagv(dev_priv))
3652		return 0;
3653
3654	if (dev_priv->sagv_status == I915_SAGV_DISABLED)
3655		return 0;
3656
3657	DRM_DEBUG_KMS("Disabling the SAGV\n");
3658	mutex_lock(&dev_priv->pcu_lock);
3659
3660	/* bspec says to keep retrying for at least 1 ms */
3661	ret = skl_pcode_request(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3662				GEN9_SAGV_DISABLE,
3663				GEN9_SAGV_IS_DISABLED, GEN9_SAGV_IS_DISABLED,
3664				1);
3665	mutex_unlock(&dev_priv->pcu_lock);
3666
3667	/*
3668	 * Some skl systems, pre-release machines in particular,
3669	 * don't actually have an SAGV.
3670	 */
3671	if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3672		DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
3673		dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3674		return 0;
3675	} else if (ret < 0) {
3676		DRM_ERROR("Failed to disable the SAGV (%d)\n", ret);
3677		return ret;
3678	}
3679
3680	dev_priv->sagv_status = I915_SAGV_DISABLED;
3681	return 0;
3682}
3683
3684bool intel_can_enable_sagv(struct drm_atomic_state *state)
3685{
3686	struct drm_device *dev = state->dev;
3687	struct drm_i915_private *dev_priv = to_i915(dev);
3688	struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3689	struct intel_crtc *crtc;
3690	struct intel_plane *plane;
3691	struct intel_crtc_state *cstate;
3692	enum pipe pipe;
3693	int level, latency;
3694	int sagv_block_time_us;
3695
3696	if (!intel_has_sagv(dev_priv))
3697		return false;
3698
3699	if (IS_GEN9(dev_priv))
3700		sagv_block_time_us = 30;
3701	else if (IS_GEN10(dev_priv))
3702		sagv_block_time_us = 20;
3703	else
3704		sagv_block_time_us = 10;
3705
3706	/*
3707	 * SKL+ workaround: bspec recommends we disable the SAGV when we have
3708	 * more then one pipe enabled
3709	 *
3710	 * If there are no active CRTCs, no additional checks need be performed
3711	 */
3712	if (hweight32(intel_state->active_crtcs) == 0)
3713		return true;
3714	else if (hweight32(intel_state->active_crtcs) > 1)
3715		return false;
3716
3717	/* Since we're now guaranteed to only have one active CRTC... */
3718	pipe = ffs(intel_state->active_crtcs) - 1;
3719	crtc = intel_get_crtc_for_pipe(dev_priv, pipe);
3720	cstate = to_intel_crtc_state(crtc->base.state);
3721
3722	if (crtc->base.state->adjusted_mode.flags & DRM_MODE_FLAG_INTERLACE)
3723		return false;
3724
3725	for_each_intel_plane_on_crtc(dev, crtc, plane) {
3726		struct skl_plane_wm *wm =
3727			&cstate->wm.skl.optimal.planes[plane->id];
3728
3729		/* Skip this plane if it's not enabled */
3730		if (!wm->wm[0].plane_en)
3731			continue;
3732
3733		/* Find the highest enabled wm level for this plane */
3734		for (level = ilk_wm_max_level(dev_priv);
3735		     !wm->wm[level].plane_en; --level)
3736		     { }
3737
3738		latency = dev_priv->wm.skl_latency[level];
3739
3740		if (skl_needs_memory_bw_wa(intel_state) &&
3741		    plane->base.state->fb->modifier ==
3742		    I915_FORMAT_MOD_X_TILED)
3743			latency += 15;
3744
3745		/*
3746		 * If any of the planes on this pipe don't enable wm levels that
3747		 * incur memory latencies higher than sagv_block_time_us we
3748		 * can't enable the SAGV.
3749		 */
3750		if (latency < sagv_block_time_us)
3751			return false;
3752	}
3753
3754	return true;
3755}
3756
3757static void
3758skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
3759				   const struct intel_crtc_state *cstate,
3760				   struct skl_ddb_entry *alloc, /* out */
3761				   int *num_active /* out */)
3762{
3763	struct drm_atomic_state *state = cstate->base.state;
3764	struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3765	struct drm_i915_private *dev_priv = to_i915(dev);
3766	struct drm_crtc *for_crtc = cstate->base.crtc;
3767	unsigned int pipe_size, ddb_size;
3768	int nth_active_pipe;
3769
3770	if (WARN_ON(!state) || !cstate->base.active) {
3771		alloc->start = 0;
3772		alloc->end = 0;
3773		*num_active = hweight32(dev_priv->active_crtcs);
3774		return;
3775	}
3776
3777	if (intel_state->active_pipe_changes)
3778		*num_active = hweight32(intel_state->active_crtcs);
3779	else
3780		*num_active = hweight32(dev_priv->active_crtcs);
3781
3782	ddb_size = INTEL_INFO(dev_priv)->ddb_size;
3783	WARN_ON(ddb_size == 0);
3784
3785	if (INTEL_GEN(dev_priv) < 11)
3786		ddb_size -= 4; /* 4 blocks for bypass path allocation */
3787
3788	/*
3789	 * If the state doesn't change the active CRTC's, then there's
3790	 * no need to recalculate; the existing pipe allocation limits
3791	 * should remain unchanged.  Note that we're safe from racing
3792	 * commits since any racing commit that changes the active CRTC
3793	 * list would need to grab _all_ crtc locks, including the one
3794	 * we currently hold.
3795	 */
3796	if (!intel_state->active_pipe_changes) {
3797		/*
3798		 * alloc may be cleared by clear_intel_crtc_state,
3799		 * copy from old state to be sure
3800		 */
3801		*alloc = to_intel_crtc_state(for_crtc->state)->wm.skl.ddb;
3802		return;
3803	}
3804
3805	nth_active_pipe = hweight32(intel_state->active_crtcs &
3806				    (drm_crtc_mask(for_crtc) - 1));
3807	pipe_size = ddb_size / hweight32(intel_state->active_crtcs);
3808	alloc->start = nth_active_pipe * ddb_size / *num_active;
3809	alloc->end = alloc->start + pipe_size;
3810}
3811
3812static unsigned int skl_cursor_allocation(int num_active)
3813{
3814	if (num_active == 1)
3815		return 32;
3816
3817	return 8;
3818}
3819
3820static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
3821{
3822	entry->start = reg & 0x3ff;
3823	entry->end = (reg >> 16) & 0x3ff;
3824	if (entry->end)
3825		entry->end += 1;
3826}
3827
3828void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
3829			  struct skl_ddb_allocation *ddb /* out */)
3830{
3831	struct intel_crtc *crtc;
3832
3833	memset(ddb, 0, sizeof(*ddb));
3834
3835	for_each_intel_crtc(&dev_priv->drm, crtc) {
3836		enum intel_display_power_domain power_domain;
3837		enum plane_id plane_id;
3838		enum pipe pipe = crtc->pipe;
3839
3840		power_domain = POWER_DOMAIN_PIPE(pipe);
3841		if (!intel_display_power_get_if_enabled(dev_priv, power_domain))
3842			continue;
3843
3844		for_each_plane_id_on_crtc(crtc, plane_id) {
3845			u32 val;
3846
3847			if (plane_id != PLANE_CURSOR)
3848				val = I915_READ(PLANE_BUF_CFG(pipe, plane_id));
3849			else
3850				val = I915_READ(CUR_BUF_CFG(pipe));
3851
3852			skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane_id], val);
3853		}
3854
3855		intel_display_power_put(dev_priv, power_domain);
3856	}
3857}
3858
3859/*
3860 * Determines the downscale amount of a plane for the purposes of watermark calculations.
3861 * The bspec defines downscale amount as:
3862 *
3863 * """
3864 * Horizontal down scale amount = maximum[1, Horizontal source size /
3865 *                                           Horizontal destination size]
3866 * Vertical down scale amount = maximum[1, Vertical source size /
3867 *                                         Vertical destination size]
3868 * Total down scale amount = Horizontal down scale amount *
3869 *                           Vertical down scale amount
3870 * """
3871 *
3872 * Return value is provided in 16.16 fixed point form to retain fractional part.
3873 * Caller should take care of dividing & rounding off the value.
3874 */
3875static uint_fixed_16_16_t
3876skl_plane_downscale_amount(const struct intel_crtc_state *cstate,
3877			   const struct intel_plane_state *pstate)
3878{
3879	struct intel_plane *plane = to_intel_plane(pstate->base.plane);
3880	uint32_t src_w, src_h, dst_w, dst_h;
3881	uint_fixed_16_16_t fp_w_ratio, fp_h_ratio;
3882	uint_fixed_16_16_t downscale_h, downscale_w;
3883
3884	if (WARN_ON(!intel_wm_plane_visible(cstate, pstate)))
3885		return u32_to_fixed16(0);
3886
3887	/* n.b., src is 16.16 fixed point, dst is whole integer */
3888	if (plane->id == PLANE_CURSOR) {
3889		/*
3890		 * Cursors only support 0/180 degree rotation,
3891		 * hence no need to account for rotation here.
3892		 */
3893		src_w = pstate->base.src_w >> 16;
3894		src_h = pstate->base.src_h >> 16;
3895		dst_w = pstate->base.crtc_w;
3896		dst_h = pstate->base.crtc_h;
3897	} else {
3898		/*
3899		 * Src coordinates are already rotated by 270 degrees for
3900		 * the 90/270 degree plane rotation cases (to match the
3901		 * GTT mapping), hence no need to account for rotation here.
3902		 */
3903		src_w = drm_rect_width(&pstate->base.src) >> 16;
3904		src_h = drm_rect_height(&pstate->base.src) >> 16;
3905		dst_w = drm_rect_width(&pstate->base.dst);
3906		dst_h = drm_rect_height(&pstate->base.dst);
3907	}
3908
3909	fp_w_ratio = div_fixed16(src_w, dst_w);
3910	fp_h_ratio = div_fixed16(src_h, dst_h);
3911	downscale_w = max_fixed16(fp_w_ratio, u32_to_fixed16(1));
3912	downscale_h = max_fixed16(fp_h_ratio, u32_to_fixed16(1));
3913
3914	return mul_fixed16(downscale_w, downscale_h);
3915}
3916
3917static uint_fixed_16_16_t
3918skl_pipe_downscale_amount(const struct intel_crtc_state *crtc_state)
3919{
3920	uint_fixed_16_16_t pipe_downscale = u32_to_fixed16(1);
3921
3922	if (!crtc_state->base.enable)
3923		return pipe_downscale;
3924
3925	if (crtc_state->pch_pfit.enabled) {
3926		uint32_t src_w, src_h, dst_w, dst_h;
3927		uint32_t pfit_size = crtc_state->pch_pfit.size;
3928		uint_fixed_16_16_t fp_w_ratio, fp_h_ratio;
3929		uint_fixed_16_16_t downscale_h, downscale_w;
3930
3931		src_w = crtc_state->pipe_src_w;
3932		src_h = crtc_state->pipe_src_h;
3933		dst_w = pfit_size >> 16;
3934		dst_h = pfit_size & 0xffff;
3935
3936		if (!dst_w || !dst_h)
3937			return pipe_downscale;
3938
3939		fp_w_ratio = div_fixed16(src_w, dst_w);
3940		fp_h_ratio = div_fixed16(src_h, dst_h);
3941		downscale_w = max_fixed16(fp_w_ratio, u32_to_fixed16(1));
3942		downscale_h = max_fixed16(fp_h_ratio, u32_to_fixed16(1));
3943
3944		pipe_downscale = mul_fixed16(downscale_w, downscale_h);
3945	}
3946
3947	return pipe_downscale;
3948}
3949
3950int skl_check_pipe_max_pixel_rate(struct intel_crtc *intel_crtc,
3951				  struct intel_crtc_state *cstate)
3952{
3953	struct drm_i915_private *dev_priv = to_i915(intel_crtc->base.dev);
3954	struct drm_crtc_state *crtc_state = &cstate->base;
3955	struct drm_atomic_state *state = crtc_state->state;
3956	struct drm_plane *plane;
3957	const struct drm_plane_state *pstate;
3958	struct intel_plane_state *intel_pstate;
3959	int crtc_clock, dotclk;
3960	uint32_t pipe_max_pixel_rate;
3961	uint_fixed_16_16_t pipe_downscale;
3962	uint_fixed_16_16_t max_downscale = u32_to_fixed16(1);
3963
3964	if (!cstate->base.enable)
3965		return 0;
3966
3967	drm_atomic_crtc_state_for_each_plane_state(plane, pstate, crtc_state) {
3968		uint_fixed_16_16_t plane_downscale;
3969		uint_fixed_16_16_t fp_9_div_8 = div_fixed16(9, 8);
3970		int bpp;
3971
3972		if (!intel_wm_plane_visible(cstate,
3973					    to_intel_plane_state(pstate)))
3974			continue;
3975
3976		if (WARN_ON(!pstate->fb))
3977			return -EINVAL;
3978
3979		intel_pstate = to_intel_plane_state(pstate);
3980		plane_downscale = skl_plane_downscale_amount(cstate,
3981							     intel_pstate);
3982		bpp = pstate->fb->format->cpp[0] * 8;
3983		if (bpp == 64)
3984			plane_downscale = mul_fixed16(plane_downscale,
3985						      fp_9_div_8);
3986
3987		max_downscale = max_fixed16(plane_downscale, max_downscale);
3988	}
3989	pipe_downscale = skl_pipe_downscale_amount(cstate);
3990
3991	pipe_downscale = mul_fixed16(pipe_downscale, max_downscale);
3992
3993	crtc_clock = crtc_state->adjusted_mode.crtc_clock;
3994	dotclk = to_intel_atomic_state(state)->cdclk.logical.cdclk;
3995
3996	if (IS_GEMINILAKE(dev_priv) || INTEL_GEN(dev_priv) >= 10)
3997		dotclk *= 2;
3998
3999	pipe_max_pixel_rate = div_round_up_u32_fixed16(dotclk, pipe_downscale);
4000
4001	if (pipe_max_pixel_rate < crtc_clock) {
4002		DRM_DEBUG_KMS("Max supported pixel clock with scaling exceeded\n");
4003		return -EINVAL;
4004	}
4005
4006	return 0;
4007}
4008
4009static unsigned int
4010skl_plane_relative_data_rate(const struct intel_crtc_state *cstate,
4011			     const struct drm_plane_state *pstate,
4012			     int y)
4013{
4014	struct intel_plane *plane = to_intel_plane(pstate->plane);
4015	struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
4016	uint32_t data_rate;
4017	uint32_t width = 0, height = 0;
4018	struct drm_framebuffer *fb;
4019	u32 format;
4020	uint_fixed_16_16_t down_scale_amount;
4021
4022	if (!intel_pstate->base.visible)
4023		return 0;
4024
4025	fb = pstate->fb;
4026	format = fb->format->format;
4027
4028	if (plane->id == PLANE_CURSOR)
4029		return 0;
4030	if (y && format != DRM_FORMAT_NV12)
4031		return 0;
4032
4033	/*
4034	 * Src coordinates are already rotated by 270 degrees for
4035	 * the 90/270 degree plane rotation cases (to match the
4036	 * GTT mapping), hence no need to account for rotation here.
4037	 */
4038	width = drm_rect_width(&intel_pstate->base.src) >> 16;
4039	height = drm_rect_height(&intel_pstate->base.src) >> 16;
4040
4041	/* for planar format */
4042	if (format == DRM_FORMAT_NV12) {
4043		if (y)  /* y-plane data rate */
4044			data_rate = width * height *
4045				fb->format->cpp[0];
4046		else    /* uv-plane data rate */
4047			data_rate = (width / 2) * (height / 2) *
4048				fb->format->cpp[1];
4049	} else {
4050		/* for packed formats */
4051		data_rate = width * height * fb->format->cpp[0];
4052	}
4053
4054	down_scale_amount = skl_plane_downscale_amount(cstate, intel_pstate);
4055
4056	return mul_round_up_u32_fixed16(data_rate, down_scale_amount);
4057}
4058
4059/*
4060 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
4061 * a 8192x4096@32bpp framebuffer:
4062 *   3 * 4096 * 8192  * 4 < 2^32
4063 */
4064static unsigned int
4065skl_get_total_relative_data_rate(struct intel_crtc_state *intel_cstate,
4066				 unsigned *plane_data_rate,
4067				 unsigned *plane_y_data_rate)
4068{
4069	struct drm_crtc_state *cstate = &intel_cstate->base;
4070	struct drm_atomic_state *state = cstate->state;
4071	struct drm_plane *plane;
4072	const struct drm_plane_state *pstate;
4073	unsigned int total_data_rate = 0;
4074
4075	if (WARN_ON(!state))
4076		return 0;
4077
4078	/* Calculate and cache data rate for each plane */
4079	drm_atomic_crtc_state_for_each_plane_state(plane, pstate, cstate) {
4080		enum plane_id plane_id = to_intel_plane(plane)->id;
4081		unsigned int rate;
4082
4083		/* packed/uv */
4084		rate = skl_plane_relative_data_rate(intel_cstate,
4085						    pstate, 0);
4086		plane_data_rate[plane_id] = rate;
4087
4088		total_data_rate += rate;
4089
4090		/* y-plane */
4091		rate = skl_plane_relative_data_rate(intel_cstate,
4092						    pstate, 1);
4093		plane_y_data_rate[plane_id] = rate;
4094
4095		total_data_rate += rate;
4096	}
4097
4098	return total_data_rate;
4099}
4100
4101static uint16_t
4102skl_ddb_min_alloc(const struct drm_plane_state *pstate,
4103		  const int y)
4104{
4105	struct drm_framebuffer *fb = pstate->fb;
4106	struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
4107	uint32_t src_w, src_h;
4108	uint32_t min_scanlines = 8;
4109	uint8_t plane_bpp;
4110
4111	if (WARN_ON(!fb))
4112		return 0;
4113
4114	/* For packed formats, no y-plane, return 0 */
4115	if (y && fb->format->format != DRM_FORMAT_NV12)
4116		return 0;
4117
4118	/* For Non Y-tile return 8-blocks */
4119	if (fb->modifier != I915_FORMAT_MOD_Y_TILED &&
4120	    fb->modifier != I915_FORMAT_MOD_Yf_TILED &&
4121	    fb->modifier != I915_FORMAT_MOD_Y_TILED_CCS &&
4122	    fb->modifier != I915_FORMAT_MOD_Yf_TILED_CCS)
4123		return 8;
4124
4125	/*
4126	 * Src coordinates are already rotated by 270 degrees for
4127	 * the 90/270 degree plane rotation cases (to match the
4128	 * GTT mapping), hence no need to account for rotation here.
4129	 */
4130	src_w = drm_rect_width(&intel_pstate->base.src) >> 16;
4131	src_h = drm_rect_height(&intel_pstate->base.src) >> 16;
4132
4133	/* Halve UV plane width and height for NV12 */
4134	if (fb->format->format == DRM_FORMAT_NV12 && !y) {
4135		src_w /= 2;
4136		src_h /= 2;
4137	}
4138
4139	if (fb->format->format == DRM_FORMAT_NV12 && !y)
4140		plane_bpp = fb->format->cpp[1];
4141	else
4142		plane_bpp = fb->format->cpp[0];
4143
4144	if (drm_rotation_90_or_270(pstate->rotation)) {
4145		switch (plane_bpp) {
4146		case 1:
4147			min_scanlines = 32;
4148			break;
4149		case 2:
4150			min_scanlines = 16;
4151			break;
4152		case 4:
4153			min_scanlines = 8;
4154			break;
4155		case 8:
4156			min_scanlines = 4;
4157			break;
4158		default:
4159			WARN(1, "Unsupported pixel depth %u for rotation",
4160			     plane_bpp);
4161			min_scanlines = 32;
4162		}
4163	}
4164
4165	return DIV_ROUND_UP((4 * src_w * plane_bpp), 512) * min_scanlines/4 + 3;
4166}
4167
4168static void
4169skl_ddb_calc_min(const struct intel_crtc_state *cstate, int num_active,
4170		 uint16_t *minimum, uint16_t *y_minimum)
4171{
4172	const struct drm_plane_state *pstate;
4173	struct drm_plane *plane;
4174
4175	drm_atomic_crtc_state_for_each_plane_state(plane, pstate, &cstate->base) {
4176		enum plane_id plane_id = to_intel_plane(plane)->id;
4177
4178		if (plane_id == PLANE_CURSOR)
4179			continue;
4180
4181		if (!pstate->visible)
4182			continue;
4183
4184		minimum[plane_id] = skl_ddb_min_alloc(pstate, 0);
4185		y_minimum[plane_id] = skl_ddb_min_alloc(pstate, 1);
4186	}
4187
4188	minimum[PLANE_CURSOR] = skl_cursor_allocation(num_active);
4189}
4190
4191static int
4192skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
4193		      struct skl_ddb_allocation *ddb /* out */)
4194{
4195	struct drm_atomic_state *state = cstate->base.state;
4196	struct drm_crtc *crtc = cstate->base.crtc;
4197	struct drm_device *dev = crtc->dev;
4198	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4199	enum pipe pipe = intel_crtc->pipe;
4200	struct skl_ddb_entry *alloc = &cstate->wm.skl.ddb;
4201	uint16_t alloc_size, start;
4202	uint16_t minimum[I915_MAX_PLANES] = {};
4203	uint16_t y_minimum[I915_MAX_PLANES] = {};
4204	unsigned int total_data_rate;
4205	enum plane_id plane_id;
4206	int num_active;
4207	unsigned plane_data_rate[I915_MAX_PLANES] = {};
4208	unsigned plane_y_data_rate[I915_MAX_PLANES] = {};
4209	uint16_t total_min_blocks = 0;
4210
4211	/* Clear the partitioning for disabled planes. */
4212	memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
4213	memset(ddb->y_plane[pipe], 0, sizeof(ddb->y_plane[pipe]));
4214
4215	if (WARN_ON(!state))
4216		return 0;
4217
4218	if (!cstate->base.active) {
4219		alloc->start = alloc->end = 0;
4220		return 0;
4221	}
4222
4223	skl_ddb_get_pipe_allocation_limits(dev, cstate, alloc, &num_active);
4224	alloc_size = skl_ddb_entry_size(alloc);
4225	if (alloc_size == 0)
4226		return 0;
4227
4228	skl_ddb_calc_min(cstate, num_active, minimum, y_minimum);
4229
4230	/*
4231	 * 1. Allocate the mininum required blocks for each active plane
4232	 * and allocate the cursor, it doesn't require extra allocation
4233	 * proportional to the data rate.
4234	 */
4235
4236	for_each_plane_id_on_crtc(intel_crtc, plane_id) {
4237		total_min_blocks += minimum[plane_id];
4238		total_min_blocks += y_minimum[plane_id];
4239	}
4240
4241	if (total_min_blocks > alloc_size) {
4242		DRM_DEBUG_KMS("Requested display configuration exceeds system DDB limitations");
4243		DRM_DEBUG_KMS("minimum required %d/%d\n", total_min_blocks,
4244							alloc_size);
4245		return -EINVAL;
4246	}
4247
4248	alloc_size -= total_min_blocks;
4249	ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - minimum[PLANE_CURSOR];
4250	ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
4251
4252	/*
4253	 * 2. Distribute the remaining space in proportion to the amount of
4254	 * data each plane needs to fetch from memory.
4255	 *
4256	 * FIXME: we may not allocate every single block here.
4257	 */
4258	total_data_rate = skl_get_total_relative_data_rate(cstate,
4259							   plane_data_rate,
4260							   plane_y_data_rate);
4261	if (total_data_rate == 0)
4262		return 0;
4263
4264	start = alloc->start;
4265	for_each_plane_id_on_crtc(intel_crtc, plane_id) {
4266		unsigned int data_rate, y_data_rate;
4267		uint16_t plane_blocks, y_plane_blocks = 0;
4268
4269		if (plane_id == PLANE_CURSOR)
4270			continue;
4271
4272		data_rate = plane_data_rate[plane_id];
4273
4274		/*
4275		 * allocation for (packed formats) or (uv-plane part of planar format):
4276		 * promote the expression to 64 bits to avoid overflowing, the
4277		 * result is < available as data_rate / total_data_rate < 1
4278		 */
4279		plane_blocks = minimum[plane_id];
4280		plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
4281					total_data_rate);
4282
4283		/* Leave disabled planes at (0,0) */
4284		if (data_rate) {
4285			ddb->plane[pipe][plane_id].start = start;
4286			ddb->plane[pipe][plane_id].end = start + plane_blocks;
4287		}
4288
4289		start += plane_blocks;
4290
4291		/*
4292		 * allocation for y_plane part of planar format:
4293		 */
4294		y_data_rate = plane_y_data_rate[plane_id];
4295
4296		y_plane_blocks = y_minimum[plane_id];
4297		y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
4298					total_data_rate);
4299
4300		if (y_data_rate) {
4301			ddb->y_plane[pipe][plane_id].start = start;
4302			ddb->y_plane[pipe][plane_id].end = start + y_plane_blocks;
4303		}
4304
4305		start += y_plane_blocks;
4306	}
4307
4308	return 0;
4309}
4310
4311/*
4312 * The max latency should be 257 (max the punit can code is 255 and we add 2us
4313 * for the read latency) and cpp should always be <= 8, so that
4314 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
4315 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
4316*/
4317static uint_fixed_16_16_t
4318skl_wm_method1(const struct drm_i915_private *dev_priv, uint32_t pixel_rate,
4319	       uint8_t cpp, uint32_t latency, uint32_t dbuf_block_size)
4320{
4321	uint32_t wm_intermediate_val;
4322	uint_fixed_16_16_t ret;
4323
4324	if (latency == 0)
4325		return FP_16_16_MAX;
4326
4327	wm_intermediate_val = latency * pixel_rate * cpp;
4328	ret = div_fixed16(wm_intermediate_val, 1000 * dbuf_block_size);
4329
4330	if (INTEL_GEN(dev_priv) >= 10)
4331		ret = add_fixed16_u32(ret, 1);
4332
4333	return ret;
4334}
4335
4336static uint_fixed_16_16_t skl_wm_method2(uint32_t pixel_rate,
4337			uint32_t pipe_htotal,
4338			uint32_t latency,
4339			uint_fixed_16_16_t plane_blocks_per_line)
4340{
4341	uint32_t wm_intermediate_val;
4342	uint_fixed_16_16_t ret;
4343
4344	if (latency == 0)
4345		return FP_16_16_MAX;
4346
4347	wm_intermediate_val = latency * pixel_rate;
4348	wm_intermediate_val = DIV_ROUND_UP(wm_intermediate_val,
4349					   pipe_htotal * 1000);
4350	ret = mul_u32_fixed16(wm_intermediate_val, plane_blocks_per_line);
4351	return ret;
4352}
4353
4354static uint_fixed_16_16_t
4355intel_get_linetime_us(struct intel_crtc_state *cstate)
4356{
4357	uint32_t pixel_rate;
4358	uint32_t crtc_htotal;
4359	uint_fixed_16_16_t linetime_us;
4360
4361	if (!cstate->base.active)
4362		return u32_to_fixed16(0);
4363
4364	pixel_rate = cstate->pixel_rate;
4365
4366	if (WARN_ON(pixel_rate == 0))
4367		return u32_to_fixed16(0);
4368
4369	crtc_htotal = cstate->base.adjusted_mode.crtc_htotal;
4370	linetime_us = div_fixed16(crtc_htotal * 1000, pixel_rate);
4371
4372	return linetime_us;
4373}
4374
4375static uint32_t
4376skl_adjusted_plane_pixel_rate(const struct intel_crtc_state *cstate,
4377			      const struct intel_plane_state *pstate)
4378{
4379	uint64_t adjusted_pixel_rate;
4380	uint_fixed_16_16_t downscale_amount;
4381
4382	/* Shouldn't reach here on disabled planes... */
4383	if (WARN_ON(!intel_wm_plane_visible(cstate, pstate)))
4384		return 0;
4385
4386	/*
4387	 * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
4388	 * with additional adjustments for plane-specific scaling.
4389	 */
4390	adjusted_pixel_rate = cstate->pixel_rate;
4391	downscale_amount = skl_plane_downscale_amount(cstate, pstate);
4392
4393	return mul_round_up_u32_fixed16(adjusted_pixel_rate,
4394					    downscale_amount);
4395}
4396
4397static int
4398skl_compute_plane_wm_params(const struct drm_i915_private *dev_priv,
4399			    struct intel_crtc_state *cstate,
4400			    const struct intel_plane_state *intel_pstate,
4401			    struct skl_wm_params *wp)
4402{
4403	struct intel_plane *plane = to_intel_plane(intel_pstate->base.plane);
4404	const struct drm_plane_state *pstate = &intel_pstate->base;
4405	const struct drm_framebuffer *fb = pstate->fb;
4406	uint32_t interm_pbpl;
4407	struct intel_atomic_state *state =
4408		to_intel_atomic_state(cstate->base.state);
4409	bool apply_memory_bw_wa = skl_needs_memory_bw_wa(state);
4410
4411	if (!intel_wm_plane_visible(cstate, intel_pstate))
4412		return 0;
4413
4414	wp->y_tiled = fb->modifier == I915_FORMAT_MOD_Y_TILED ||
4415		      fb->modifier == I915_FORMAT_MOD_Yf_TILED ||
4416		      fb->modifier == I915_FORMAT_MOD_Y_TILED_CCS ||
4417		      fb->modifier == I915_FORMAT_MOD_Yf_TILED_CCS;
4418	wp->x_tiled = fb->modifier == I915_FORMAT_MOD_X_TILED;
4419	wp->rc_surface = fb->modifier == I915_FORMAT_MOD_Y_TILED_CCS ||
4420			 fb->modifier == I915_FORMAT_MOD_Yf_TILED_CCS;
4421
4422	if (plane->id == PLANE_CURSOR) {
4423		wp->width = intel_pstate->base.crtc_w;
4424	} else {
4425		/*
4426		 * Src coordinates are already rotated by 270 degrees for
4427		 * the 90/270 degree plane rotation cases (to match the
4428		 * GTT mapping), hence no need to account for rotation here.
4429		 */
4430		wp->width = drm_rect_width(&intel_pstate->base.src) >> 16;
4431	}
4432
4433	wp->cpp = (fb->format->format == DRM_FORMAT_NV12) ? fb->format->cpp[1] :
4434							    fb->format->cpp[0];
4435	wp->plane_pixel_rate = skl_adjusted_plane_pixel_rate(cstate,
4436							     intel_pstate);
4437
4438	if (INTEL_GEN(dev_priv) >= 11 &&
4439	    fb->modifier == I915_FORMAT_MOD_Yf_TILED && wp->cpp == 8)
4440		wp->dbuf_block_size = 256;
4441	else
4442		wp->dbuf_block_size = 512;
4443
4444	if (drm_rotation_90_or_270(pstate->rotation)) {
4445
4446		switch (wp->cpp) {
4447		case 1:
4448			wp->y_min_scanlines = 16;
4449			break;
4450		case 2:
4451			wp->y_min_scanlines = 8;
4452			break;
4453		case 4:
4454			wp->y_min_scanlines = 4;
4455			break;
4456		default:
4457			MISSING_CASE(wp->cpp);
4458			return -EINVAL;
4459		}
4460	} else {
4461		wp->y_min_scanlines = 4;
4462	}
4463
4464	if (apply_memory_bw_wa)
4465		wp->y_min_scanlines *= 2;
4466
4467	wp->plane_bytes_per_line = wp->width * wp->cpp;
4468	if (wp->y_tiled) {
4469		interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line *
4470					   wp->y_min_scanlines,
4471					   wp->dbuf_block_size);
4472
4473		if (INTEL_GEN(dev_priv) >= 10)
4474			interm_pbpl++;
4475
4476		wp->plane_blocks_per_line = div_fixed16(interm_pbpl,
4477							wp->y_min_scanlines);
4478	} else if (wp->x_tiled && IS_GEN9(dev_priv)) {
4479		interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line,
4480					   wp->dbuf_block_size);
4481		wp->plane_blocks_per_line = u32_to_fixed16(interm_pbpl);
4482	} else {
4483		interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line,
4484					   wp->dbuf_block_size) + 1;
4485		wp->plane_blocks_per_line = u32_to_fixed16(interm_pbpl);
4486	}
4487
4488	wp->y_tile_minimum = mul_u32_fixed16(wp->y_min_scanlines,
4489					     wp->plane_blocks_per_line);
4490	wp->linetime_us = fixed16_to_u32_round_up(
4491					intel_get_linetime_us(cstate));
4492
4493	return 0;
4494}
4495
4496static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
4497				struct intel_crtc_state *cstate,
4498				const struct intel_plane_state *intel_pstate,
4499				uint16_t ddb_allocation,
4500				int level,
4501				const struct skl_wm_params *wp,
4502				uint16_t *out_blocks, /* out */
4503				uint8_t *out_lines, /* out */
4504				bool *enabled /* out */)
4505{
4506	const struct drm_plane_state *pstate = &intel_pstate->base;
4507	uint32_t latency = dev_priv->wm.skl_latency[level];
4508	uint_fixed_16_16_t method1, method2;
4509	uint_fixed_16_16_t selected_result;
4510	uint32_t res_blocks, res_lines;
4511	struct intel_atomic_state *state =
4512		to_intel_atomic_state(cstate->base.state);
4513	bool apply_memory_bw_wa = skl_needs_memory_bw_wa(state);
4514	uint32_t min_disp_buf_needed;
4515
4516	if (latency == 0 ||
4517	    !intel_wm_plane_visible(cstate, intel_pstate)) {
4518		*enabled = false;
4519		return 0;
4520	}
4521
4522	/* Display WA #1141: kbl,cfl */
4523	if ((IS_KABYLAKE(dev_priv) || IS_COFFEELAKE(dev_priv) ||
4524	    IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_B0)) &&
4525	    dev_priv->ipc_enabled)
4526		latency += 4;
4527
4528	if (apply_memory_bw_wa && wp->x_tiled)
4529		latency += 15;
4530
4531	method1 = skl_wm_method1(dev_priv, wp->plane_pixel_rate,
4532				 wp->cpp, latency, wp->dbuf_block_size);
4533	method2 = skl_wm_method2(wp->plane_pixel_rate,
4534				 cstate->base.adjusted_mode.crtc_htotal,
4535				 latency,
4536				 wp->plane_blocks_per_line);
4537
4538	if (wp->y_tiled) {
4539		selected_result = max_fixed16(method2, wp->y_tile_minimum);
4540	} else {
4541		if ((wp->cpp * cstate->base.adjusted_mode.crtc_htotal /
4542		     wp->dbuf_block_size < 1) &&
4543		     (wp->plane_bytes_per_line / wp->dbuf_block_size < 1))
4544			selected_result = method2;
4545		else if (ddb_allocation >=
4546			 fixed16_to_u32_round_up(wp->plane_blocks_per_line))
4547			selected_result = min_fixed16(method1, method2);
4548		else if (latency >= wp->linetime_us)
4549			selected_result = min_fixed16(method1, method2);
4550		else
4551			selected_result = method1;
4552	}
4553
4554	res_blocks = fixed16_to_u32_round_up(selected_result) + 1;
4555	res_lines = div_round_up_fixed16(selected_result,
4556					 wp->plane_blocks_per_line);
4557
4558	/* Display WA #1125: skl,bxt,kbl,glk */
4559	if (level == 0 && wp->rc_surface)
4560		res_blocks += fixed16_to_u32_round_up(wp->y_tile_minimum);
4561
4562	/* Display WA #1126: skl,bxt,kbl,glk */
4563	if (level >= 1 && level <= 7) {
4564		if (wp->y_tiled) {
4565			res_blocks += fixed16_to_u32_round_up(
4566							wp->y_tile_minimum);
4567			res_lines += wp->y_min_scanlines;
4568		} else {
4569			res_blocks++;
4570		}
4571	}
4572
4573	if (INTEL_GEN(dev_priv) >= 11) {
4574		if (wp->y_tiled) {
4575			uint32_t extra_lines;
4576			uint_fixed_16_16_t fp_min_disp_buf_needed;
4577
4578			if (res_lines % wp->y_min_scanlines == 0)
4579				extra_lines = wp->y_min_scanlines;
4580			else
4581				extra_lines = wp->y_min_scanlines * 2 -
4582					      res_lines % wp->y_min_scanlines;
4583
4584			fp_min_disp_buf_needed = mul_u32_fixed16(res_lines +
4585						extra_lines,
4586						wp->plane_blocks_per_line);
4587			min_disp_buf_needed = fixed16_to_u32_round_up(
4588						fp_min_disp_buf_needed);
4589		} else {
4590			min_disp_buf_needed = DIV_ROUND_UP(res_blocks * 11, 10);
4591		}
4592	} else {
4593		min_disp_buf_needed = res_blocks;
4594	}
4595
4596	if ((level > 0 && res_lines > 31) ||
4597	    res_blocks >= ddb_allocation ||
4598	    min_disp_buf_needed >= ddb_allocation) {
4599		*enabled = false;
4600
4601		/*
4602		 * If there are no valid level 0 watermarks, then we can't
4603		 * support this display configuration.
4604		 */
4605		if (level) {
4606			return 0;
4607		} else {
4608			struct drm_plane *plane = pstate->plane;
4609
4610			DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
4611			DRM_DEBUG_KMS("[PLANE:%d:%s] blocks required = %u/%u, lines required = %u/31\n",
4612				      plane->base.id, plane->name,
4613				      res_blocks, ddb_allocation, res_lines);
4614			return -EINVAL;
4615		}
4616	}
4617
4618	/* The number of lines are ignored for the level 0 watermark. */
4619	*out_lines = level ? res_lines : 0;
4620	*out_blocks = res_blocks;
4621	*enabled = true;
4622
4623	return 0;
4624}
4625
4626static int
4627skl_compute_wm_levels(const struct drm_i915_private *dev_priv,
4628		      struct skl_ddb_allocation *ddb,
4629		      struct intel_crtc_state *cstate,
4630		      const struct intel_plane_state *intel_pstate,
4631		      const struct skl_wm_params *wm_params,
4632		      struct skl_plane_wm *wm)
4633{
4634	struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4635	struct drm_plane *plane = intel_pstate->base.plane;
4636	struct intel_plane *intel_plane = to_intel_plane(plane);
4637	uint16_t ddb_blocks;
4638	enum pipe pipe = intel_crtc->pipe;
4639	int level, max_level = ilk_wm_max_level(dev_priv);
4640	int ret;
4641
4642	if (WARN_ON(!intel_pstate->base.fb))
4643		return -EINVAL;
4644
4645	ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][intel_plane->id]);
4646
4647	for (level = 0; level <= max_level; level++) {
4648		struct skl_wm_level *result = &wm->wm[level];
4649
4650		ret = skl_compute_plane_wm(dev_priv,
4651					   cstate,
4652					   intel_pstate,
4653					   ddb_blocks,
4654					   level,
4655					   wm_params,
4656					   &result->plane_res_b,
4657					   &result->plane_res_l,
4658					   &result->plane_en);
4659		if (ret)
4660			return ret;
4661	}
4662
4663	return 0;
4664}
4665
4666static uint32_t
4667skl_compute_linetime_wm(struct intel_crtc_state *cstate)
4668{
4669	struct drm_atomic_state *state = cstate->base.state;
4670	struct drm_i915_private *dev_priv = to_i915(state->dev);
4671	uint_fixed_16_16_t linetime_us;
4672	uint32_t linetime_wm;
4673
4674	linetime_us = intel_get_linetime_us(cstate);
4675
4676	if (is_fixed16_zero(linetime_us))
4677		return 0;
4678
4679	linetime_wm = fixed16_to_u32_round_up(mul_u32_fixed16(8, linetime_us));
4680
4681	/* Display WA #1135: bxt:ALL GLK:ALL */
4682	if ((IS_BROXTON(dev_priv) || IS_GEMINILAKE(dev_priv)) &&
4683	    dev_priv->ipc_enabled)
4684		linetime_wm /= 2;
4685
4686	return linetime_wm;
4687}
4688
4689static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
4690				      struct skl_wm_params *wp,
4691				      struct skl_wm_level *wm_l0,
4692				      uint16_t ddb_allocation,
4693				      struct skl_wm_level *trans_wm /* out */)
4694{
4695	struct drm_device *dev = cstate->base.crtc->dev;
4696	const struct drm_i915_private *dev_priv = to_i915(dev);
4697	uint16_t trans_min, trans_y_tile_min;
4698	const uint16_t trans_amount = 10; /* This is configurable amount */
4699	uint16_t trans_offset_b, res_blocks;
4700
4701	if (!cstate->base.active)
4702		goto exit;
4703
4704	/* Transition WM are not recommended by HW team for GEN9 */
4705	if (INTEL_GEN(dev_priv) <= 9)
4706		goto exit;
4707
4708	/* Transition WM don't make any sense if ipc is disabled */
4709	if (!dev_priv->ipc_enabled)
4710		goto exit;
4711
4712	trans_min = 0;
4713	if (INTEL_GEN(dev_priv) >= 10)
4714		trans_min = 4;
4715
4716	trans_offset_b = trans_min + trans_amount;
4717
4718	if (wp->y_tiled) {
4719		trans_y_tile_min = (uint16_t) mul_round_up_u32_fixed16(2,
4720							wp->y_tile_minimum);
4721		res_blocks = max(wm_l0->plane_res_b, trans_y_tile_min) +
4722				trans_offset_b;
4723	} else {
4724		res_blocks = wm_l0->plane_res_b + trans_offset_b;
4725
4726		/* WA BUG:1938466 add one block for non y-tile planes */
4727		if (IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_A0))
4728			res_blocks += 1;
4729
4730	}
4731
4732	res_blocks += 1;
4733
4734	if (res_blocks < ddb_allocation) {
4735		trans_wm->plane_res_b = res_blocks;
4736		trans_wm->plane_en = true;
4737		return;
4738	}
4739
4740exit:
4741	trans_wm->plane_en = false;
4742}
4743
4744static int skl_build_pipe_wm(struct intel_crtc_state *cstate,
4745			     struct skl_ddb_allocation *ddb,
4746			     struct skl_pipe_wm *pipe_wm)
4747{
4748	struct drm_device *dev = cstate->base.crtc->dev;
4749	struct drm_crtc_state *crtc_state = &cstate->base;
4750	const struct drm_i915_private *dev_priv = to_i915(dev);
4751	struct drm_plane *plane;
4752	const struct drm_plane_state *pstate;
4753	struct skl_plane_wm *wm;
4754	int ret;
4755
4756	/*
4757	 * We'll only calculate watermarks for planes that are actually
4758	 * enabled, so make sure all other planes are set as disabled.
4759	 */
4760	memset(pipe_wm->planes, 0, sizeof(pipe_wm->planes));
4761
4762	drm_atomic_crtc_state_for_each_plane_state(plane, pstate, crtc_state) {
4763		const struct intel_plane_state *intel_pstate =
4764						to_intel_plane_state(pstate);
4765		enum plane_id plane_id = to_intel_plane(plane)->id;
4766		struct skl_wm_params wm_params;
4767		enum pipe pipe = to_intel_crtc(cstate->base.crtc)->pipe;
4768		uint16_t ddb_blocks;
4769
4770		wm = &pipe_wm->planes[plane_id];
4771		ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][plane_id]);
4772		memset(&wm_params, 0, sizeof(struct skl_wm_params));
4773
4774		ret = skl_compute_plane_wm_params(dev_priv, cstate,
4775						  intel_pstate, &wm_params);
4776		if (ret)
4777			return ret;
4778
4779		ret = skl_compute_wm_levels(dev_priv, ddb, cstate,
4780					    intel_pstate, &wm_params, wm);
4781		if (ret)
4782			return ret;
4783		skl_compute_transition_wm(cstate, &wm_params, &wm->wm[0],
4784					  ddb_blocks, &wm->trans_wm);
4785	}
4786	pipe_wm->linetime = skl_compute_linetime_wm(cstate);
4787
4788	return 0;
4789}
4790
4791static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
4792				i915_reg_t reg,
4793				const struct skl_ddb_entry *entry)
4794{
4795	if (entry->end)
4796		I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
4797	else
4798		I915_WRITE(reg, 0);
4799}
4800
4801static void skl_write_wm_level(struct drm_i915_private *dev_priv,
4802			       i915_reg_t reg,
4803			       const struct skl_wm_level *level)
4804{
4805	uint32_t val = 0;
4806
4807	if (level->plane_en) {
4808		val |= PLANE_WM_EN;
4809		val |= level->plane_res_b;
4810		val |= level->plane_res_l << PLANE_WM_LINES_SHIFT;
4811	}
4812
4813	I915_WRITE(reg, val);
4814}
4815
4816static void skl_write_plane_wm(struct intel_crtc *intel_crtc,
4817			       const struct skl_plane_wm *wm,
4818			       const struct skl_ddb_allocation *ddb,
4819			       enum plane_id plane_id)
4820{
4821	struct drm_crtc *crtc = &intel_crtc->base;
4822	struct drm_device *dev = crtc->dev;
4823	struct drm_i915_private *dev_priv = to_i915(dev);
4824	int level, max_level = ilk_wm_max_level(dev_priv);
4825	enum pipe pipe = intel_crtc->pipe;
4826
4827	for (level = 0; level <= max_level; level++) {
4828		skl_write_wm_level(dev_priv, PLANE_WM(pipe, plane_id, level),
4829				   &wm->wm[level]);
4830	}
4831	skl_write_wm_level(dev_priv, PLANE_WM_TRANS(pipe, plane_id),
4832			   &wm->trans_wm);
4833
4834	skl_ddb_entry_write(dev_priv, PLANE_BUF_CFG(pipe, plane_id),
4835			    &ddb->plane[pipe][plane_id]);
4836	if (INTEL_GEN(dev_priv) < 11)
4837		skl_ddb_entry_write(dev_priv,
4838				    PLANE_NV12_BUF_CFG(pipe, plane_id),
4839				    &ddb->y_plane[pipe][plane_id]);
4840}
4841
4842static void skl_write_cursor_wm(struct intel_crtc *intel_crtc,
4843				const struct skl_plane_wm *wm,
4844				const struct skl_ddb_allocation *ddb)
4845{
4846	struct drm_crtc *crtc = &intel_crtc->base;
4847	struct drm_device *dev = crtc->dev;
4848	struct drm_i915_private *dev_priv = to_i915(dev);
4849	int level, max_level = ilk_wm_max_level(dev_priv);
4850	enum pipe pipe = intel_crtc->pipe;
4851
4852	for (level = 0; level <= max_level; level++) {
4853		skl_write_wm_level(dev_priv, CUR_WM(pipe, level),
4854				   &wm->wm[level]);
4855	}
4856	skl_write_wm_level(dev_priv, CUR_WM_TRANS(pipe), &wm->trans_wm);
4857
4858	skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
4859			    &ddb->plane[pipe][PLANE_CURSOR]);
4860}
4861
4862bool skl_wm_level_equals(const struct skl_wm_level *l1,
4863			 const struct skl_wm_level *l2)
4864{
4865	if (l1->plane_en != l2->plane_en)
4866		return false;
4867
4868	/* If both planes aren't enabled, the rest shouldn't matter */
4869	if (!l1->plane_en)
4870		return true;
4871
4872	return (l1->plane_res_l == l2->plane_res_l &&
4873		l1->plane_res_b == l2->plane_res_b);
4874}
4875
4876static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry *a,
4877					   const struct skl_ddb_entry *b)
4878{
4879	return a->start < b->end && b->start < a->end;
4880}
4881
4882bool skl_ddb_allocation_overlaps(struct drm_i915_private *dev_priv,
4883				 const struct skl_ddb_entry **entries,
4884				 const struct skl_ddb_entry *ddb,
4885				 int ignore)
4886{
4887	enum pipe pipe;
4888
4889	for_each_pipe(dev_priv, pipe) {
4890		if (pipe != ignore && entries[pipe] &&
4891		    skl_ddb_entries_overlap(ddb, entries[pipe]))
4892			return true;
4893	}
4894
4895	return false;
4896}
4897
4898static int skl_update_pipe_wm(struct drm_crtc_state *cstate,
4899			      const struct skl_pipe_wm *old_pipe_wm,
4900			      struct skl_pipe_wm *pipe_wm, /* out */
4901			      struct skl_ddb_allocation *ddb, /* out */
4902			      bool *changed /* out */)
4903{
4904	struct intel_crtc_state *intel_cstate = to_intel_crtc_state(cstate);
4905	int ret;
4906
4907	ret = skl_build_pipe_wm(intel_cstate, ddb, pipe_wm);
4908	if (ret)
4909		return ret;
4910
4911	if (!memcmp(old_pipe_wm, pipe_wm, sizeof(*pipe_wm)))
4912		*changed = false;
4913	else
4914		*changed = true;
4915
4916	return 0;
4917}
4918
4919static uint32_t
4920pipes_modified(struct drm_atomic_state *state)
4921{
4922	struct drm_crtc *crtc;
4923	struct drm_crtc_state *cstate;
4924	uint32_t i, ret = 0;
4925
4926	for_each_new_crtc_in_state(state, crtc, cstate, i)
4927		ret |= drm_crtc_mask(crtc);
4928
4929	return ret;
4930}
4931
4932static int
4933skl_ddb_add_affected_planes(struct intel_crtc_state *cstate)
4934{
4935	struct drm_atomic_state *state = cstate->base.state;
4936	struct drm_device *dev = state->dev;
4937	struct drm_crtc *crtc = cstate->base.crtc;
4938	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4939	struct drm_i915_private *dev_priv = to_i915(dev);
4940	struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4941	struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
4942	struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
4943	struct drm_plane_state *plane_state;
4944	struct drm_plane *plane;
4945	enum pipe pipe = intel_crtc->pipe;
4946
4947	WARN_ON(!drm_atomic_get_existing_crtc_state(state, crtc));
4948
4949	drm_for_each_plane_mask(plane, dev, cstate->base.plane_mask) {
4950		enum plane_id plane_id = to_intel_plane(plane)->id;
4951
4952		if (skl_ddb_entry_equal(&cur_ddb->plane[pipe][plane_id],
4953					&new_ddb->plane[pipe][plane_id]) &&
4954		    skl_ddb_entry_equal(&cur_ddb->y_plane[pipe][plane_id],
4955					&new_ddb->y_plane[pipe][plane_id]))
4956			continue;
4957
4958		plane_state = drm_atomic_get_plane_state(state, plane);
4959		if (IS_ERR(plane_state))
4960			return PTR_ERR(plane_state);
4961	}
4962
4963	return 0;
4964}
4965
4966static int
4967skl_compute_ddb(struct drm_atomic_state *state)
4968{
4969	struct drm_device *dev = state->dev;
4970	struct drm_i915_private *dev_priv = to_i915(dev);
4971	struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4972	struct intel_crtc *intel_crtc;
4973	struct skl_ddb_allocation *ddb = &intel_state->wm_results.ddb;
4974	uint32_t realloc_pipes = pipes_modified(state);
4975	int ret;
4976
4977	/*
4978	 * If this is our first atomic update following hardware readout,
4979	 * we can't trust the DDB that the BIOS programmed for us.  Let's
4980	 * pretend that all pipes switched active status so that we'll
4981	 * ensure a full DDB recompute.
4982	 */
4983	if (dev_priv->wm.distrust_bios_wm) {
4984		ret = drm_modeset_lock(&dev->mode_config.connection_mutex,
4985				       state->acquire_ctx);
4986		if (ret)
4987			return ret;
4988
4989		intel_state->active_pipe_changes = ~0;
4990
4991		/*
4992		 * We usually only initialize intel_state->active_crtcs if we
4993		 * we're doing a modeset; make sure this field is always
4994		 * initialized during the sanitization process that happens
4995		 * on the first commit too.
4996		 */
4997		if (!intel_state->modeset)
4998			intel_state->active_crtcs = dev_priv->active_crtcs;
4999	}
5000
5001	/*
5002	 * If the modeset changes which CRTC's are active, we need to
5003	 * recompute the DDB allocation for *all* active pipes, even
5004	 * those that weren't otherwise being modified in any way by this
5005	 * atomic commit.  Due to the shrinking of the per-pipe allocations
5006	 * when new active CRTC's are added, it's possible for a pipe that
5007	 * we were already using and aren't changing at all here to suddenly
5008	 * become invalid if its DDB needs exceeds its new allocation.
5009	 *
5010	 * Note that if we wind up doing a full DDB recompute, we can't let
5011	 * any other display updates race with this transaction, so we need
5012	 * to grab the lock on *all* CRTC's.
5013	 */
5014	if (intel_state->active_pipe_changes) {
5015		realloc_pipes = ~0;
5016		intel_state->wm_results.dirty_pipes = ~0;
5017	}
5018
5019	/*
5020	 * We're not recomputing for the pipes not included in the commit, so
5021	 * make sure we start with the current state.
5022	 */
5023	memcpy(ddb, &dev_priv->wm.skl_hw.ddb, sizeof(*ddb));
5024
5025	for_each_intel_crtc_mask(dev, intel_crtc, realloc_pipes) {
5026		struct intel_crtc_state *cstate;
5027
5028		cstate = intel_atomic_get_crtc_state(state, intel_crtc);
5029		if (IS_ERR(cstate))
5030			return PTR_ERR(cstate);
5031
5032		ret = skl_allocate_pipe_ddb(cstate, ddb);
5033		if (ret)
5034			return ret;
5035
5036		ret = skl_ddb_add_affected_planes(cstate);
5037		if (ret)
5038			return ret;
5039	}
5040
5041	return 0;
5042}
5043
5044static void
5045skl_copy_wm_for_pipe(struct skl_wm_values *dst,
5046		     struct skl_wm_values *src,
5047		     enum pipe pipe)
5048{
5049	memcpy(dst->ddb.y_plane[pipe], src->ddb.y_plane[pipe],
5050	       sizeof(dst->ddb.y_plane[pipe]));
5051	memcpy(dst->ddb.plane[pipe], src->ddb.plane[pipe],
5052	       sizeof(dst->ddb.plane[pipe]));
5053}
5054
5055static void
5056skl_print_wm_changes(const struct drm_atomic_state *state)
5057{
5058	const struct drm_device *dev = state->dev;
5059	const struct drm_i915_private *dev_priv = to_i915(dev);
5060	const struct intel_atomic_state *intel_state =
5061		to_intel_atomic_state(state);
5062	const struct drm_crtc *crtc;
5063	const struct drm_crtc_state *cstate;
5064	const struct intel_plane *intel_plane;
5065	const struct skl_ddb_allocation *old_ddb = &dev_priv->wm.skl_hw.ddb;
5066	const struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
5067	int i;
5068
5069	for_each_new_crtc_in_state(state, crtc, cstate, i) {
5070		const struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
5071		enum pipe pipe = intel_crtc->pipe;
5072
5073		for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
5074			enum plane_id plane_id = intel_plane->id;
5075			const struct skl_ddb_entry *old, *new;
5076
5077			old = &old_ddb->plane[pipe][plane_id];
5078			new = &new_ddb->plane[pipe][plane_id];
5079
5080			if (skl_ddb_entry_equal(old, new))
5081				continue;
5082
5083			DRM_DEBUG_ATOMIC("[PLANE:%d:%s] ddb (%d - %d) -> (%d - %d)\n",
5084					 intel_plane->base.base.id,
5085					 intel_plane->base.name,
5086					 old->start, old->end,
5087					 new->start, new->end);
5088		}
5089	}
5090}
5091
5092static int
5093skl_compute_wm(struct drm_atomic_state *state)
5094{
5095	struct drm_crtc *crtc;
5096	struct drm_crtc_state *cstate;
5097	struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
5098	struct skl_wm_values *results = &intel_state->wm_results;
5099	struct drm_device *dev = state->dev;
5100	struct skl_pipe_wm *pipe_wm;
5101	bool changed = false;
5102	int ret, i;
5103
5104	/*
5105	 * When we distrust bios wm we always need to recompute to set the
5106	 * expected DDB allocations for each CRTC.
5107	 */
5108	if (to_i915(dev)->wm.distrust_bios_wm)
5109		changed = true;
5110
5111	/*
5112	 * If this transaction isn't actually touching any CRTC's, don't
5113	 * bother with watermark calculation.  Note that if we pass this
5114	 * test, we're guaranteed to hold at least one CRTC state mutex,
5115	 * which means we can safely use values like dev_priv->active_crtcs
5116	 * since any racing commits that want to update them would need to
5117	 * hold _all_ CRTC state mutexes.
5118	 */
5119	for_each_new_crtc_in_state(state, crtc, cstate, i)
5120		changed = true;
5121
5122	if (!changed)
5123		return 0;
5124
5125	/* Clear all dirty flags */
5126	results->dirty_pipes = 0;
5127
5128	ret = skl_compute_ddb(state);
5129	if (ret)
5130		return ret;
5131
5132	/*
5133	 * Calculate WM's for all pipes that are part of this transaction.
5134	 * Note that the DDB allocation above may have added more CRTC's that
5135	 * weren't otherwise being modified (and set bits in dirty_pipes) if
5136	 * pipe allocations had to change.
5137	 *
5138	 * FIXME:  Now that we're doing this in the atomic check phase, we
5139	 * should allow skl_update_pipe_wm() to return failure in cases where
5140	 * no suitable watermark values can be found.
5141	 */
5142	for_each_new_crtc_in_state(state, crtc, cstate, i) {
5143		struct intel_crtc_state *intel_cstate =
5144			to_intel_crtc_state(cstate);
5145		const struct skl_pipe_wm *old_pipe_wm =
5146			&to_intel_crtc_state(crtc->state)->wm.skl.optimal;
5147
5148		pipe_wm = &intel_cstate->wm.skl.optimal;
5149		ret = skl_update_pipe_wm(cstate, old_pipe_wm, pipe_wm,
5150					 &results->ddb, &changed);
5151		if (ret)
5152			return ret;
5153
5154		if (changed)
5155			results->dirty_pipes |= drm_crtc_mask(crtc);
5156
5157		if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
5158			/* This pipe's WM's did not change */
5159			continue;
5160
5161		intel_cstate->update_wm_pre = true;
5162	}
5163
5164	skl_print_wm_changes(state);
5165
5166	return 0;
5167}
5168
5169static void skl_atomic_update_crtc_wm(struct intel_atomic_state *state,
5170				      struct intel_crtc_state *cstate)
5171{
5172	struct intel_crtc *crtc = to_intel_crtc(cstate->base.crtc);
5173	struct drm_i915_private *dev_priv = to_i915(state->base.dev);
5174	struct skl_pipe_wm *pipe_wm = &cstate->wm.skl.optimal;
5175	const struct skl_ddb_allocation *ddb = &state->wm_results.ddb;
5176	enum pipe pipe = crtc->pipe;
5177	enum plane_id plane_id;
5178
5179	if (!(state->wm_results.dirty_pipes & drm_crtc_mask(&crtc->base)))
5180		return;
5181
5182	I915_WRITE(PIPE_WM_LINETIME(pipe), pipe_wm->linetime);
5183
5184	for_each_plane_id_on_crtc(crtc, plane_id) {
5185		if (plane_id != PLANE_CURSOR)
5186			skl_write_plane_wm(crtc, &pipe_wm->planes[plane_id],
5187					   ddb, plane_id);
5188		else
5189			skl_write_cursor_wm(crtc, &pipe_wm->planes[plane_id],
5190					    ddb);
5191	}
5192}
5193
5194static void skl_initial_wm(struct intel_atomic_state *state,
5195			   struct intel_crtc_state *cstate)
5196{
5197	struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
5198	struct drm_device *dev = intel_crtc->base.dev;
5199	struct drm_i915_private *dev_priv = to_i915(dev);
5200	struct skl_wm_values *results = &state->wm_results;
5201	struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw;
5202	enum pipe pipe = intel_crtc->pipe;
5203
5204	if ((results->dirty_pipes & drm_crtc_mask(&intel_crtc->base)) == 0)
5205		return;
5206
5207	mutex_lock(&dev_priv->wm.wm_mutex);
5208
5209	if (cstate->base.active_changed)
5210		skl_atomic_update_crtc_wm(state, cstate);
5211
5212	skl_copy_wm_for_pipe(hw_vals, results, pipe);
5213
5214	mutex_unlock(&dev_priv->wm.wm_mutex);
5215}
5216
5217static void ilk_compute_wm_config(struct drm_device *dev,
5218				  struct intel_wm_config *config)
5219{
5220	struct intel_crtc *crtc;
5221
5222	/* Compute the currently _active_ config */
5223	for_each_intel_crtc(dev, crtc) {
5224		const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
5225
5226		if (!wm->pipe_enabled)
5227			continue;
5228
5229		config->sprites_enabled |= wm->sprites_enabled;
5230		config->sprites_scaled |= wm->sprites_scaled;
5231		config->num_pipes_active++;
5232	}
5233}
5234
5235static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
5236{
5237	struct drm_device *dev = &dev_priv->drm;
5238	struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
5239	struct ilk_wm_maximums max;
5240	struct intel_wm_config config = {};
5241	struct ilk_wm_values results = {};
5242	enum intel_ddb_partitioning partitioning;
5243
5244	ilk_compute_wm_config(dev, &config);
5245
5246	ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
5247	ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
5248
5249	/* 5/6 split only in single pipe config on IVB+ */
5250	if (INTEL_GEN(dev_priv) >= 7 &&
5251	    config.num_pipes_active == 1 && config.sprites_enabled) {
5252		ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
5253		ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
5254
5255		best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
5256	} else {
5257		best_lp_wm = &lp_wm_1_2;
5258	}
5259
5260	partitioning = (best_lp_wm == &lp_wm_1_2) ?
5261		       INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
5262
5263	ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
5264
5265	ilk_write_wm_values(dev_priv, &results);
5266}
5267
5268static void ilk_initial_watermarks(struct intel_atomic_state *state,
5269				   struct intel_crtc_state *cstate)
5270{
5271	struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
5272	struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
5273
5274	mutex_lock(&dev_priv->wm.wm_mutex);
5275	intel_crtc->wm.active.ilk = cstate->wm.ilk.intermediate;
5276	ilk_program_watermarks(dev_priv);
5277	mutex_unlock(&dev_priv->wm.wm_mutex);
5278}
5279
5280static void ilk_optimize_watermarks(struct intel_atomic_state *state,
5281				    struct intel_crtc_state *cstate)
5282{
5283	struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
5284	struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
5285
5286	mutex_lock(&dev_priv->wm.wm_mutex);
5287	if (cstate->wm.need_postvbl_update) {
5288		intel_crtc->wm.active.ilk = cstate->wm.ilk.optimal;
5289		ilk_program_watermarks(dev_priv);
5290	}
5291	mutex_unlock(&dev_priv->wm.wm_mutex);
5292}
5293
5294static inline void skl_wm_level_from_reg_val(uint32_t val,
5295					     struct skl_wm_level *level)
5296{
5297	level->plane_en = val & PLANE_WM_EN;
5298	level->plane_res_b = val & PLANE_WM_BLOCKS_MASK;
5299	level->plane_res_l = (val >> PLANE_WM_LINES_SHIFT) &
5300		PLANE_WM_LINES_MASK;
5301}
5302
5303void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc,
5304			      struct skl_pipe_wm *out)
5305{
5306	struct drm_i915_private *dev_priv = to_i915(crtc->dev);
5307	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
5308	enum pipe pipe = intel_crtc->pipe;
5309	int level, max_level;
5310	enum plane_id plane_id;
5311	uint32_t val;
5312
5313	max_level = ilk_wm_max_level(dev_priv);
5314
5315	for_each_plane_id_on_crtc(intel_crtc, plane_id) {
5316		struct skl_plane_wm *wm = &out->planes[plane_id];
5317
5318		for (level = 0; level <= max_level; level++) {
5319			if (plane_id != PLANE_CURSOR)
5320				val = I915_READ(PLANE_WM(pipe, plane_id, level));
5321			else
5322				val = I915_READ(CUR_WM(pipe, level));
5323
5324			skl_wm_level_from_reg_val(val, &wm->wm[level]);
5325		}
5326
5327		if (plane_id != PLANE_CURSOR)
5328			val = I915_READ(PLANE_WM_TRANS(pipe, plane_id));
5329		else
5330			val = I915_READ(CUR_WM_TRANS(pipe));
5331
5332		skl_wm_level_from_reg_val(val, &wm->trans_wm);
5333	}
5334
5335	if (!intel_crtc->active)
5336		return;
5337
5338	out->linetime = I915_READ(PIPE_WM_LINETIME(pipe));
5339}
5340
5341void skl_wm_get_hw_state(struct drm_device *dev)
5342{
5343	struct drm_i915_private *dev_priv = to_i915(dev);
5344	struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
5345	struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
5346	struct drm_crtc *crtc;
5347	struct intel_crtc *intel_crtc;
5348	struct intel_crtc_state *cstate;
5349
5350	skl_ddb_get_hw_state(dev_priv, ddb);
5351	list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
5352		intel_crtc = to_intel_crtc(crtc);
5353		cstate = to_intel_crtc_state(crtc->state);
5354
5355		skl_pipe_wm_get_hw_state(crtc, &cstate->wm.skl.optimal);
5356
5357		if (intel_crtc->active)
5358			hw->dirty_pipes |= drm_crtc_mask(crtc);
5359	}
5360
5361	if (dev_priv->active_crtcs) {
5362		/* Fully recompute DDB on first atomic commit */
5363		dev_priv->wm.distrust_bios_wm = true;
5364	} else {
5365		/* Easy/common case; just sanitize DDB now if everything off */
5366		memset(ddb, 0, sizeof(*ddb));
5367	}
5368}
5369
5370static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
5371{
5372	struct drm_device *dev = crtc->dev;
5373	struct drm_i915_private *dev_priv = to_i915(dev);
5374	struct ilk_wm_values *hw = &dev_priv->wm.hw;
5375	struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
5376	struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
5377	struct intel_pipe_wm *active = &cstate->wm.ilk.optimal;
5378	enum pipe pipe = intel_crtc->pipe;
5379	static const i915_reg_t wm0_pipe_reg[] = {
5380		[PIPE_A] = WM0_PIPEA_ILK,
5381		[PIPE_B] = WM0_PIPEB_ILK,
5382		[PIPE_C] = WM0_PIPEC_IVB,
5383	};
5384
5385	hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
5386	if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
5387		hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
5388
5389	memset(active, 0, sizeof(*active));
5390
5391	active->pipe_enabled = intel_crtc->active;
5392
5393	if (active->pipe_enabled) {
5394		u32 tmp = hw->wm_pipe[pipe];
5395
5396		/*
5397		 * For active pipes LP0 watermark is marked as
5398		 * enabled, and LP1+ watermaks as disabled since
5399		 * we can't really reverse compute them in case
5400		 * multiple pipes are active.
5401		 */
5402		active->wm[0].enable = true;
5403		active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
5404		active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
5405		active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
5406		active->linetime = hw->wm_linetime[pipe];
5407	} else {
5408		int level, max_level = ilk_wm_max_level(dev_priv);
5409
5410		/*
5411		 * For inactive pipes, all watermark levels
5412		 * should be marked as enabled but zeroed,
5413		 * which is what we'd compute them to.
5414		 */
5415		for (level = 0; level <= max_level; level++)
5416			active->wm[level].enable = true;
5417	}
5418
5419	intel_crtc->wm.active.ilk = *active;
5420}
5421
5422#define _FW_WM(value, plane) \
5423	(((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
5424#define _FW_WM_VLV(value, plane) \
5425	(((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
5426
5427static void g4x_read_wm_values(struct drm_i915_private *dev_priv,
5428			       struct g4x_wm_values *wm)
5429{
5430	uint32_t tmp;
5431
5432	tmp = I915_READ(DSPFW1);
5433	wm->sr.plane = _FW_WM(tmp, SR);
5434	wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
5435	wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEB);
5436	wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEA);
5437
5438	tmp = I915_READ(DSPFW2);
5439	wm->fbc_en = tmp & DSPFW_FBC_SR_EN;
5440	wm->sr.fbc = _FW_WM(tmp, FBC_SR);
5441	wm->hpll.fbc = _FW_WM(tmp, FBC_HPLL_SR);
5442	wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEB);
5443	wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
5444	wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEA);
5445
5446	tmp = I915_READ(DSPFW3);
5447	wm->hpll_en = tmp & DSPFW_HPLL_SR_EN;
5448	wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
5449	wm->hpll.cursor = _FW_WM(tmp, HPLL_CURSOR);
5450	wm->hpll.plane = _FW_WM(tmp, HPLL_SR);
5451}
5452
5453static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
5454			       struct vlv_wm_values *wm)
5455{
5456	enum pipe pipe;
5457	uint32_t tmp;
5458
5459	for_each_pipe(dev_priv, pipe) {
5460		tmp = I915_READ(VLV_DDL(pipe));
5461
5462		wm->ddl[pipe].plane[PLANE_PRIMARY] =
5463			(tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5464		wm->ddl[pipe].plane[PLANE_CURSOR] =
5465			(tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5466		wm->ddl[pipe].plane[PLANE_SPRITE0] =
5467			(tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5468		wm->ddl[pipe].plane[PLANE_SPRITE1] =
5469			(tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5470	}
5471
5472	tmp = I915_READ(DSPFW1);
5473	wm->sr.plane = _FW_WM(tmp, SR);
5474	wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
5475	wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEB);
5476	wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEA);
5477
5478	tmp = I915_READ(DSPFW2);
5479	wm->pipe[PIPE_A].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEB);
5480	wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
5481	wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEA);
5482
5483	tmp = I915_READ(DSPFW3);
5484	wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
5485
5486	if (IS_CHERRYVIEW(dev_priv)) {
5487		tmp = I915_READ(DSPFW7_CHV);
5488		wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
5489		wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
5490
5491		tmp = I915_READ(DSPFW8_CHV);
5492		wm->pipe[PIPE_C].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEF);
5493		wm->pipe[PIPE_C].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEE);
5494
5495		tmp = I915_READ(DSPFW9_CHV);
5496		wm->pipe[PIPE_C].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEC);
5497		wm->pipe[PIPE_C].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORC);
5498
5499		tmp = I915_READ(DSPHOWM);
5500		wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
5501		wm->pipe[PIPE_C].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
5502		wm->pipe[PIPE_C].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
5503		wm->pipe[PIPE_C].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEC_HI) << 8;
5504		wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
5505		wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
5506		wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
5507		wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
5508		wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
5509		wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
5510	} else {
5511		tmp = I915_READ(DSPFW7);
5512		wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
5513		wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
5514
5515		tmp = I915_READ(DSPHOWM);
5516		wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
5517		wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
5518		wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
5519		wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
5520		wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
5521		wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
5522		wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
5523	}
5524}
5525
5526#undef _FW_WM
5527#undef _FW_WM_VLV
5528
5529void g4x_wm_get_hw_state(struct drm_device *dev)
5530{
5531	struct drm_i915_private *dev_priv = to_i915(dev);
5532	struct g4x_wm_values *wm = &dev_priv->wm.g4x;
5533	struct intel_crtc *crtc;
5534
5535	g4x_read_wm_values(dev_priv, wm);
5536
5537	wm->cxsr = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
5538
5539	for_each_intel_crtc(dev, crtc) {
5540		struct intel_crtc_state *crtc_state =
5541			to_intel_crtc_state(crtc->base.state);
5542		struct g4x_wm_state *active = &crtc->wm.active.g4x;
5543		struct g4x_pipe_wm *raw;
5544		enum pipe pipe = crtc->pipe;
5545		enum plane_id plane_id;
5546		int level, max_level;
5547
5548		active->cxsr = wm->cxsr;
5549		active->hpll_en = wm->hpll_en;
5550		active->fbc_en = wm->fbc_en;
5551
5552		active->sr = wm->sr;
5553		active->hpll = wm->hpll;
5554
5555		for_each_plane_id_on_crtc(crtc, plane_id) {
5556			active->wm.plane[plane_id] =
5557				wm->pipe[pipe].plane[plane_id];
5558		}
5559
5560		if (wm->cxsr && wm->hpll_en)
5561			max_level = G4X_WM_LEVEL_HPLL;
5562		else if (wm->cxsr)
5563			max_level = G4X_WM_LEVEL_SR;
5564		else
5565			max_level = G4X_WM_LEVEL_NORMAL;
5566
5567		level = G4X_WM_LEVEL_NORMAL;
5568		raw = &crtc_state->wm.g4x.raw[level];
5569		for_each_plane_id_on_crtc(crtc, plane_id)
5570			raw->plane[plane_id] = active->wm.plane[plane_id];
5571
5572		if (++level > max_level)
5573			goto out;
5574
5575		raw = &crtc_state->wm.g4x.raw[level];
5576		raw->plane[PLANE_PRIMARY] = active->sr.plane;
5577		raw->plane[PLANE_CURSOR] = active->sr.cursor;
5578		raw->plane[PLANE_SPRITE0] = 0;
5579		raw->fbc = active->sr.fbc;
5580
5581		if (++level > max_level)
5582			goto out;
5583
5584		raw = &crtc_state->wm.g4x.raw[level];
5585		raw->plane[PLANE_PRIMARY] = active->hpll.plane;
5586		raw->plane[PLANE_CURSOR] = active->hpll.cursor;
5587		raw->plane[PLANE_SPRITE0] = 0;
5588		raw->fbc = active->hpll.fbc;
5589
5590	out:
5591		for_each_plane_id_on_crtc(crtc, plane_id)
5592			g4x_raw_plane_wm_set(crtc_state, level,
5593					     plane_id, USHRT_MAX);
5594		g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
5595
5596		crtc_state->wm.g4x.optimal = *active;
5597		crtc_state->wm.g4x.intermediate = *active;
5598
5599		DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite=%d\n",
5600			      pipe_name(pipe),
5601			      wm->pipe[pipe].plane[PLANE_PRIMARY],
5602			      wm->pipe[pipe].plane[PLANE_CURSOR],
5603			      wm->pipe[pipe].plane[PLANE_SPRITE0]);
5604	}
5605
5606	DRM_DEBUG_KMS("Initial SR watermarks: plane=%d, cursor=%d fbc=%d\n",
5607		      wm->sr.plane, wm->sr.cursor, wm->sr.fbc);
5608	DRM_DEBUG_KMS("Initial HPLL watermarks: plane=%d, SR cursor=%d fbc=%d\n",
5609		      wm->hpll.plane, wm->hpll.cursor, wm->hpll.fbc);
5610	DRM_DEBUG_KMS("Initial SR=%s HPLL=%s FBC=%s\n",
5611		      yesno(wm->cxsr), yesno(wm->hpll_en), yesno(wm->fbc_en));
5612}
5613
5614void g4x_wm_sanitize(struct drm_i915_private *dev_priv)
5615{
5616	struct intel_plane *plane;
5617	struct intel_crtc *crtc;
5618
5619	mutex_lock(&dev_priv->wm.wm_mutex);
5620
5621	for_each_intel_plane(&dev_priv->drm, plane) {
5622		struct intel_crtc *crtc =
5623			intel_get_crtc_for_pipe(dev_priv, plane->pipe);
5624		struct intel_crtc_state *crtc_state =
5625			to_intel_crtc_state(crtc->base.state);
5626		struct intel_plane_state *plane_state =
5627			to_intel_plane_state(plane->base.state);
5628		struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
5629		enum plane_id plane_id = plane->id;
5630		int level;
5631
5632		if (plane_state->base.visible)
5633			continue;
5634
5635		for (level = 0; level < 3; level++) {
5636			struct g4x_pipe_wm *raw =
5637				&crtc_state->wm.g4x.raw[level];
5638
5639			raw->plane[plane_id] = 0;
5640			wm_state->wm.plane[plane_id] = 0;
5641		}
5642
5643		if (plane_id == PLANE_PRIMARY) {
5644			for (level = 0; level < 3; level++) {
5645				struct g4x_pipe_wm *raw =
5646					&crtc_state->wm.g4x.raw[level];
5647				raw->fbc = 0;
5648			}
5649
5650			wm_state->sr.fbc = 0;
5651			wm_state->hpll.fbc = 0;
5652			wm_state->fbc_en = false;
5653		}
5654	}
5655
5656	for_each_intel_crtc(&dev_priv->drm, crtc) {
5657		struct intel_crtc_state *crtc_state =
5658			to_intel_crtc_state(crtc->base.state);
5659
5660		crtc_state->wm.g4x.intermediate =
5661			crtc_state->wm.g4x.optimal;
5662		crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
5663	}
5664
5665	g4x_program_watermarks(dev_priv);
5666
5667	mutex_unlock(&dev_priv->wm.wm_mutex);
5668}
5669
5670void vlv_wm_get_hw_state(struct drm_device *dev)
5671{
5672	struct drm_i915_private *dev_priv = to_i915(dev);
5673	struct vlv_wm_values *wm = &dev_priv->wm.vlv;
5674	struct intel_crtc *crtc;
5675	u32 val;
5676
5677	vlv_read_wm_values(dev_priv, wm);
5678
5679	wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
5680	wm->level = VLV_WM_LEVEL_PM2;
5681
5682	if (IS_CHERRYVIEW(dev_priv)) {
5683		mutex_lock(&dev_priv->pcu_lock);
5684
5685		val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
5686		if (val & DSP_MAXFIFO_PM5_ENABLE)
5687			wm->level = VLV_WM_LEVEL_PM5;
5688
5689		/*
5690		 * If DDR DVFS is disabled in the BIOS, Punit
5691		 * will never ack the request. So if that happens
5692		 * assume we don't have to enable/disable DDR DVFS
5693		 * dynamically. To test that just set the REQ_ACK
5694		 * bit to poke the Punit, but don't change the
5695		 * HIGH/LOW bits so that we don't actually change
5696		 * the current state.
5697		 */
5698		val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
5699		val |= FORCE_DDR_FREQ_REQ_ACK;
5700		vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
5701
5702		if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
5703			      FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
5704			DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
5705				      "assuming DDR DVFS is disabled\n");
5706			dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
5707		} else {
5708			val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
5709			if ((val & FORCE_DDR_HIGH_FREQ) == 0)
5710				wm->level = VLV_WM_LEVEL_DDR_DVFS;
5711		}
5712
5713		mutex_unlock(&dev_priv->pcu_lock);
5714	}
5715
5716	for_each_intel_crtc(dev, crtc) {
5717		struct intel_crtc_state *crtc_state =
5718			to_intel_crtc_state(crtc->base.state);
5719		struct vlv_wm_state *active = &crtc->wm.active.vlv;
5720		const struct vlv_fifo_state *fifo_state =
5721			&crtc_state->wm.vlv.fifo_state;
5722		enum pipe pipe = crtc->pipe;
5723		enum plane_id plane_id;
5724		int level;
5725
5726		vlv_get_fifo_size(crtc_state);
5727
5728		active->num_levels = wm->level + 1;
5729		active->cxsr = wm->cxsr;
5730
5731		for (level = 0; level < active->num_levels; level++) {
5732			struct g4x_pipe_wm *raw =
5733				&crtc_state->wm.vlv.raw[level];
5734
5735			active->sr[level].plane = wm->sr.plane;
5736			active->sr[level].cursor = wm->sr.cursor;
5737
5738			for_each_plane_id_on_crtc(crtc, plane_id) {
5739				active->wm[level].plane[plane_id] =
5740					wm->pipe[pipe].plane[plane_id];
5741
5742				raw->plane[plane_id] =
5743					vlv_invert_wm_value(active->wm[level].plane[plane_id],
5744							    fifo_state->plane[plane_id]);
5745			}
5746		}
5747
5748		for_each_plane_id_on_crtc(crtc, plane_id)
5749			vlv_raw_plane_wm_set(crtc_state, level,
5750					     plane_id, USHRT_MAX);
5751		vlv_invalidate_wms(crtc, active, level);
5752
5753		crtc_state->wm.vlv.optimal = *active;
5754		crtc_state->wm.vlv.intermediate = *active;
5755
5756		DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
5757			      pipe_name(pipe),
5758			      wm->pipe[pipe].plane[PLANE_PRIMARY],
5759			      wm->pipe[pipe].plane[PLANE_CURSOR],
5760			      wm->pipe[pipe].plane[PLANE_SPRITE0],
5761			      wm->pipe[pipe].plane[PLANE_SPRITE1]);
5762	}
5763
5764	DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
5765		      wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
5766}
5767
5768void vlv_wm_sanitize(struct drm_i915_private *dev_priv)
5769{
5770	struct intel_plane *plane;
5771	struct intel_crtc *crtc;
5772
5773	mutex_lock(&dev_priv->wm.wm_mutex);
5774
5775	for_each_intel_plane(&dev_priv->drm, plane) {
5776		struct intel_crtc *crtc =
5777			intel_get_crtc_for_pipe(dev_priv, plane->pipe);
5778		struct intel_crtc_state *crtc_state =
5779			to_intel_crtc_state(crtc->base.state);
5780		struct intel_plane_state *plane_state =
5781			to_intel_plane_state(plane->base.state);
5782		struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
5783		const struct vlv_fifo_state *fifo_state =
5784			&crtc_state->wm.vlv.fifo_state;
5785		enum plane_id plane_id = plane->id;
5786		int level;
5787
5788		if (plane_state->base.visible)
5789			continue;
5790
5791		for (level = 0; level < wm_state->num_levels; level++) {
5792			struct g4x_pipe_wm *raw =
5793				&crtc_state->wm.vlv.raw[level];
5794
5795			raw->plane[plane_id] = 0;
5796
5797			wm_state->wm[level].plane[plane_id] =
5798				vlv_invert_wm_value(raw->plane[plane_id],
5799						    fifo_state->plane[plane_id]);
5800		}
5801	}
5802
5803	for_each_intel_crtc(&dev_priv->drm, crtc) {
5804		struct intel_crtc_state *crtc_state =
5805			to_intel_crtc_state(crtc->base.state);
5806
5807		crtc_state->wm.vlv.intermediate =
5808			crtc_state->wm.vlv.optimal;
5809		crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
5810	}
5811
5812	vlv_program_watermarks(dev_priv);
5813
5814	mutex_unlock(&dev_priv->wm.wm_mutex);
5815}
5816
5817/*
5818 * FIXME should probably kill this and improve
5819 * the real watermark readout/sanitation instead
5820 */
5821static void ilk_init_lp_watermarks(struct drm_i915_private *dev_priv)
5822{
5823	I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
5824	I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
5825	I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
5826
5827	/*
5828	 * Don't touch WM1S_LP_EN here.
5829	 * Doing so could cause underruns.
5830	 */
5831}
5832
5833void ilk_wm_get_hw_state(struct drm_device *dev)
5834{
5835	struct drm_i915_private *dev_priv = to_i915(dev);
5836	struct ilk_wm_values *hw = &dev_priv->wm.hw;
5837	struct drm_crtc *crtc;
5838
5839	ilk_init_lp_watermarks(dev_priv);
5840
5841	for_each_crtc(dev, crtc)
5842		ilk_pipe_wm_get_hw_state(crtc);
5843
5844	hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
5845	hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
5846	hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
5847
5848	hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
5849	if (INTEL_GEN(dev_priv) >= 7) {
5850		hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
5851		hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
5852	}
5853
5854	if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
5855		hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
5856			INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
5857	else if (IS_IVYBRIDGE(dev_priv))
5858		hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
5859			INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
5860
5861	hw->enable_fbc_wm =
5862		!(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
5863}
5864
5865/**
5866 * intel_update_watermarks - update FIFO watermark values based on current modes
5867 * @crtc: the #intel_crtc on which to compute the WM
5868 *
5869 * Calculate watermark values for the various WM regs based on current mode
5870 * and plane configuration.
5871 *
5872 * There are several cases to deal with here:
5873 *   - normal (i.e. non-self-refresh)
5874 *   - self-refresh (SR) mode
5875 *   - lines are large relative to FIFO size (buffer can hold up to 2)
5876 *   - lines are small relative to FIFO size (buffer can hold more than 2
5877 *     lines), so need to account for TLB latency
5878 *
5879 *   The normal calculation is:
5880 *     watermark = dotclock * bytes per pixel * latency
5881 *   where latency is platform & configuration dependent (we assume pessimal
5882 *   values here).
5883 *
5884 *   The SR calculation is:
5885 *     watermark = (trunc(latency/line time)+1) * surface width *
5886 *       bytes per pixel
5887 *   where
5888 *     line time = htotal / dotclock
5889 *     surface width = hdisplay for normal plane and 64 for cursor
5890 *   and latency is assumed to be high, as above.
5891 *
5892 * The final value programmed to the register should always be rounded up,
5893 * and include an extra 2 entries to account for clock crossings.
5894 *
5895 * We don't use the sprite, so we can ignore that.  And on Crestline we have
5896 * to set the non-SR watermarks to 8.
5897 */
5898void intel_update_watermarks(struct intel_crtc *crtc)
5899{
5900	struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5901
5902	if (dev_priv->display.update_wm)
5903		dev_priv->display.update_wm(crtc);
5904}
5905
5906void intel_enable_ipc(struct drm_i915_private *dev_priv)
5907{
5908	u32 val;
5909
5910	/* Display WA #0477 WaDisableIPC: skl */
5911	if (IS_SKYLAKE(dev_priv)) {
5912		dev_priv->ipc_enabled = false;
5913		return;
5914	}
5915
5916	val = I915_READ(DISP_ARB_CTL2);
5917
5918	if (dev_priv->ipc_enabled)
5919		val |= DISP_IPC_ENABLE;
5920	else
5921		val &= ~DISP_IPC_ENABLE;
5922
5923	I915_WRITE(DISP_ARB_CTL2, val);
5924}
5925
5926void intel_init_ipc(struct drm_i915_private *dev_priv)
5927{
5928	dev_priv->ipc_enabled = false;
5929	if (!HAS_IPC(dev_priv))
5930		return;
5931
5932	dev_priv->ipc_enabled = true;
5933	intel_enable_ipc(dev_priv);
5934}
5935
5936/*
5937 * Lock protecting IPS related data structures
5938 */
5939DEFINE_SPINLOCK(mchdev_lock);
5940
5941/* Global for IPS driver to get at the current i915 device. Protected by
5942 * mchdev_lock. */
5943static struct drm_i915_private *i915_mch_dev;
5944
5945bool ironlake_set_drps(struct drm_i915_private *dev_priv, u8 val)
5946{
5947	u16 rgvswctl;
5948
5949	lockdep_assert_held(&mchdev_lock);
5950
5951	rgvswctl = I915_READ16(MEMSWCTL);
5952	if (rgvswctl & MEMCTL_CMD_STS) {
5953		DRM_DEBUG("gpu busy, RCS change rejected\n");
5954		return false; /* still busy with another command */
5955	}
5956
5957	rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
5958		(val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
5959	I915_WRITE16(MEMSWCTL, rgvswctl);
5960	POSTING_READ16(MEMSWCTL);
5961
5962	rgvswctl |= MEMCTL_CMD_STS;
5963	I915_WRITE16(MEMSWCTL, rgvswctl);
5964
5965	return true;
5966}
5967
5968static void ironlake_enable_drps(struct drm_i915_private *dev_priv)
5969{
5970	u32 rgvmodectl;
5971	u8 fmax, fmin, fstart, vstart;
5972
5973	spin_lock_irq(&mchdev_lock);
5974
5975	rgvmodectl = I915_READ(MEMMODECTL);
5976
5977	/* Enable temp reporting */
5978	I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
5979	I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
5980
5981	/* 100ms RC evaluation intervals */
5982	I915_WRITE(RCUPEI, 100000);
5983	I915_WRITE(RCDNEI, 100000);
5984
5985	/* Set max/min thresholds to 90ms and 80ms respectively */
5986	I915_WRITE(RCBMAXAVG, 90000);
5987	I915_WRITE(RCBMINAVG, 80000);
5988
5989	I915_WRITE(MEMIHYST, 1);
5990
5991	/* Set up min, max, and cur for interrupt handling */
5992	fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
5993	fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
5994	fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
5995		MEMMODE_FSTART_SHIFT;
5996
5997	vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
5998		PXVFREQ_PX_SHIFT;
5999
6000	dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
6001	dev_priv->ips.fstart = fstart;
6002
6003	dev_priv->ips.max_delay = fstart;
6004	dev_priv->ips.min_delay = fmin;
6005	dev_priv->ips.cur_delay = fstart;
6006
6007	DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
6008			 fmax, fmin, fstart);
6009
6010	I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
6011
6012	/*
6013	 * Interrupts will be enabled in ironlake_irq_postinstall
6014	 */
6015
6016	I915_WRITE(VIDSTART, vstart);
6017	POSTING_READ(VIDSTART);
6018
6019	rgvmodectl |= MEMMODE_SWMODE_EN;
6020	I915_WRITE(MEMMODECTL, rgvmodectl);
6021
6022	if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
6023		DRM_ERROR("stuck trying to change perf mode\n");
6024	mdelay(1);
6025
6026	ironlake_set_drps(dev_priv, fstart);
6027
6028	dev_priv->ips.last_count1 = I915_READ(DMIEC) +
6029		I915_READ(DDREC) + I915_READ(CSIEC);
6030	dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
6031	dev_priv->ips.last_count2 = I915_READ(GFXEC);
6032	dev_priv->ips.last_time2 = ktime_get_raw_ns();
6033
6034	spin_unlock_irq(&mchdev_lock);
6035}
6036
6037static void ironlake_disable_drps(struct drm_i915_private *dev_priv)
6038{
6039	u16 rgvswctl;
6040
6041	spin_lock_irq(&mchdev_lock);
6042
6043	rgvswctl = I915_READ16(MEMSWCTL);
6044
6045	/* Ack interrupts, disable EFC interrupt */
6046	I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
6047	I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
6048	I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
6049	I915_WRITE(DEIIR, DE_PCU_EVENT);
6050	I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
6051
6052	/* Go back to the starting frequency */
6053	ironlake_set_drps(dev_priv, dev_priv->ips.fstart);
6054	mdelay(1);
6055	rgvswctl |= MEMCTL_CMD_STS;
6056	I915_WRITE(MEMSWCTL, rgvswctl);
6057	mdelay(1);
6058
6059	spin_unlock_irq(&mchdev_lock);
6060}
6061
6062/* There's a funny hw issue where the hw returns all 0 when reading from
6063 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
6064 * ourselves, instead of doing a rmw cycle (which might result in us clearing
6065 * all limits and the gpu stuck at whatever frequency it is at atm).
6066 */
6067static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
6068{
6069	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6070	u32 limits;
6071
6072	/* Only set the down limit when we've reached the lowest level to avoid
6073	 * getting more interrupts, otherwise leave this clear. This prevents a
6074	 * race in the hw when coming out of rc6: There's a tiny window where
6075	 * the hw runs at the minimal clock before selecting the desired
6076	 * frequency, if the down threshold expires in that window we will not
6077	 * receive a down interrupt. */
6078	if (INTEL_GEN(dev_priv) >= 9) {
6079		limits = (rps->max_freq_softlimit) << 23;
6080		if (val <= rps->min_freq_softlimit)
6081			limits |= (rps->min_freq_softlimit) << 14;
6082	} else {
6083		limits = rps->max_freq_softlimit << 24;
6084		if (val <= rps->min_freq_softlimit)
6085			limits |= rps->min_freq_softlimit << 16;
6086	}
6087
6088	return limits;
6089}
6090
6091static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
6092{
6093	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6094	int new_power;
6095	u32 threshold_up = 0, threshold_down = 0; /* in % */
6096	u32 ei_up = 0, ei_down = 0;
6097
6098	new_power = rps->power;
6099	switch (rps->power) {
6100	case LOW_POWER:
6101		if (val > rps->efficient_freq + 1 &&
6102		    val > rps->cur_freq)
6103			new_power = BETWEEN;
6104		break;
6105
6106	case BETWEEN:
6107		if (val <= rps->efficient_freq &&
6108		    val < rps->cur_freq)
6109			new_power = LOW_POWER;
6110		else if (val >= rps->rp0_freq &&
6111			 val > rps->cur_freq)
6112			new_power = HIGH_POWER;
6113		break;
6114
6115	case HIGH_POWER:
6116		if (val < (rps->rp1_freq + rps->rp0_freq) >> 1 &&
6117		    val < rps->cur_freq)
6118			new_power = BETWEEN;
6119		break;
6120	}
6121	/* Max/min bins are special */
6122	if (val <= rps->min_freq_softlimit)
6123		new_power = LOW_POWER;
6124	if (val >= rps->max_freq_softlimit)
6125		new_power = HIGH_POWER;
6126	if (new_power == rps->power)
6127		return;
6128
6129	/* Note the units here are not exactly 1us, but 1280ns. */
6130	switch (new_power) {
6131	case LOW_POWER:
6132		/* Upclock if more than 95% busy over 16ms */
6133		ei_up = 16000;
6134		threshold_up = 95;
6135
6136		/* Downclock if less than 85% busy over 32ms */
6137		ei_down = 32000;
6138		threshold_down = 85;
6139		break;
6140
6141	case BETWEEN:
6142		/* Upclock if more than 90% busy over 13ms */
6143		ei_up = 13000;
6144		threshold_up = 90;
6145
6146		/* Downclock if less than 75% busy over 32ms */
6147		ei_down = 32000;
6148		threshold_down = 75;
6149		break;
6150
6151	case HIGH_POWER:
6152		/* Upclock if more than 85% busy over 10ms */
6153		ei_up = 10000;
6154		threshold_up = 85;
6155
6156		/* Downclock if less than 60% busy over 32ms */
6157		ei_down = 32000;
6158		threshold_down = 60;
6159		break;
6160	}
6161
6162	/* When byt can survive without system hang with dynamic
6163	 * sw freq adjustments, this restriction can be lifted.
6164	 */
6165	if (IS_VALLEYVIEW(dev_priv))
6166		goto skip_hw_write;
6167
6168	I915_WRITE(GEN6_RP_UP_EI,
6169		   GT_INTERVAL_FROM_US(dev_priv, ei_up));
6170	I915_WRITE(GEN6_RP_UP_THRESHOLD,
6171		   GT_INTERVAL_FROM_US(dev_priv,
6172				       ei_up * threshold_up / 100));
6173
6174	I915_WRITE(GEN6_RP_DOWN_EI,
6175		   GT_INTERVAL_FROM_US(dev_priv, ei_down));
6176	I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
6177		   GT_INTERVAL_FROM_US(dev_priv,
6178				       ei_down * threshold_down / 100));
6179
6180	I915_WRITE(GEN6_RP_CONTROL,
6181		   GEN6_RP_MEDIA_TURBO |
6182		   GEN6_RP_MEDIA_HW_NORMAL_MODE |
6183		   GEN6_RP_MEDIA_IS_GFX |
6184		   GEN6_RP_ENABLE |
6185		   GEN6_RP_UP_BUSY_AVG |
6186		   GEN6_RP_DOWN_IDLE_AVG);
6187
6188skip_hw_write:
6189	rps->power = new_power;
6190	rps->up_threshold = threshold_up;
6191	rps->down_threshold = threshold_down;
6192	rps->last_adj = 0;
6193}
6194
6195static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
6196{
6197	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6198	u32 mask = 0;
6199
6200	/* We use UP_EI_EXPIRED interupts for both up/down in manual mode */
6201	if (val > rps->min_freq_softlimit)
6202		mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
6203	if (val < rps->max_freq_softlimit)
6204		mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
6205
6206	mask &= dev_priv->pm_rps_events;
6207
6208	return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
6209}
6210
6211/* gen6_set_rps is called to update the frequency request, but should also be
6212 * called when the range (min_delay and max_delay) is modified so that we can
6213 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
6214static int gen6_set_rps(struct drm_i915_private *dev_priv, u8 val)
6215{
6216	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6217
6218	/* min/max delay may still have been modified so be sure to
6219	 * write the limits value.
6220	 */
6221	if (val != rps->cur_freq) {
6222		gen6_set_rps_thresholds(dev_priv, val);
6223
6224		if (INTEL_GEN(dev_priv) >= 9)
6225			I915_WRITE(GEN6_RPNSWREQ,
6226				   GEN9_FREQUENCY(val));
6227		else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
6228			I915_WRITE(GEN6_RPNSWREQ,
6229				   HSW_FREQUENCY(val));
6230		else
6231			I915_WRITE(GEN6_RPNSWREQ,
6232				   GEN6_FREQUENCY(val) |
6233				   GEN6_OFFSET(0) |
6234				   GEN6_AGGRESSIVE_TURBO);
6235	}
6236
6237	/* Make sure we continue to get interrupts
6238	 * until we hit the minimum or maximum frequencies.
6239	 */
6240	I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
6241	I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
6242
6243	rps->cur_freq = val;
6244	trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
6245
6246	return 0;
6247}
6248
6249static int valleyview_set_rps(struct drm_i915_private *dev_priv, u8 val)
6250{
6251	int err;
6252
6253	if (WARN_ONCE(IS_CHERRYVIEW(dev_priv) && (val & 1),
6254		      "Odd GPU freq value\n"))
6255		val &= ~1;
6256
6257	I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
6258
6259	if (val != dev_priv->gt_pm.rps.cur_freq) {
6260		err = vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
6261		if (err)
6262			return err;
6263
6264		gen6_set_rps_thresholds(dev_priv, val);
6265	}
6266
6267	dev_priv->gt_pm.rps.cur_freq = val;
6268	trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
6269
6270	return 0;
6271}
6272
6273/* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
6274 *
6275 * * If Gfx is Idle, then
6276 * 1. Forcewake Media well.
6277 * 2. Request idle freq.
6278 * 3. Release Forcewake of Media well.
6279*/
6280static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
6281{
6282	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6283	u32 val = rps->idle_freq;
6284	int err;
6285
6286	if (rps->cur_freq <= val)
6287		return;
6288
6289	/* The punit delays the write of the frequency and voltage until it
6290	 * determines the GPU is awake. During normal usage we don't want to
6291	 * waste power changing the frequency if the GPU is sleeping (rc6).
6292	 * However, the GPU and driver is now idle and we do not want to delay
6293	 * switching to minimum voltage (reducing power whilst idle) as we do
6294	 * not expect to be woken in the near future and so must flush the
6295	 * change by waking the device.
6296	 *
6297	 * We choose to take the media powerwell (either would do to trick the
6298	 * punit into committing the voltage change) as that takes a lot less
6299	 * power than the render powerwell.
6300	 */
6301	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
6302	err = valleyview_set_rps(dev_priv, val);
6303	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
6304
6305	if (err)
6306		DRM_ERROR("Failed to set RPS for idle\n");
6307}
6308
6309void gen6_rps_busy(struct drm_i915_private *dev_priv)
6310{
6311	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6312
6313	mutex_lock(&dev_priv->pcu_lock);
6314	if (rps->enabled) {
6315		u8 freq;
6316
6317		if (dev_priv->pm_rps_events & GEN6_PM_RP_UP_EI_EXPIRED)
6318			gen6_rps_reset_ei(dev_priv);
6319		I915_WRITE(GEN6_PMINTRMSK,
6320			   gen6_rps_pm_mask(dev_priv, rps->cur_freq));
6321
6322		gen6_enable_rps_interrupts(dev_priv);
6323
6324		/* Use the user's desired frequency as a guide, but for better
6325		 * performance, jump directly to RPe as our starting frequency.
6326		 */
6327		freq = max(rps->cur_freq,
6328			   rps->efficient_freq);
6329
6330		if (intel_set_rps(dev_priv,
6331				  clamp(freq,
6332					rps->min_freq_softlimit,
6333					rps->max_freq_softlimit)))
6334			DRM_DEBUG_DRIVER("Failed to set idle frequency\n");
6335	}
6336	mutex_unlock(&dev_priv->pcu_lock);
6337}
6338
6339void gen6_rps_idle(struct drm_i915_private *dev_priv)
6340{
6341	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6342
6343	/* Flush our bottom-half so that it does not race with us
6344	 * setting the idle frequency and so that it is bounded by
6345	 * our rpm wakeref. And then disable the interrupts to stop any
6346	 * futher RPS reclocking whilst we are asleep.
6347	 */
6348	gen6_disable_rps_interrupts(dev_priv);
6349
6350	mutex_lock(&dev_priv->pcu_lock);
6351	if (rps->enabled) {
6352		if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
6353			vlv_set_rps_idle(dev_priv);
6354		else
6355			gen6_set_rps(dev_priv, rps->idle_freq);
6356		rps->last_adj = 0;
6357		I915_WRITE(GEN6_PMINTRMSK,
6358			   gen6_sanitize_rps_pm_mask(dev_priv, ~0));
6359	}
6360	mutex_unlock(&dev_priv->pcu_lock);
6361}
6362
6363void gen6_rps_boost(struct i915_request *rq,
6364		    struct intel_rps_client *rps_client)
6365{
6366	struct intel_rps *rps = &rq->i915->gt_pm.rps;
6367	unsigned long flags;
6368	bool boost;
6369
6370	/* This is intentionally racy! We peek at the state here, then
6371	 * validate inside the RPS worker.
6372	 */
6373	if (!rps->enabled)
6374		return;
6375
6376	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags))
6377		return;
6378
6379	/* Serializes with i915_request_retire() */
6380	boost = false;
6381	spin_lock_irqsave(&rq->lock, flags);
6382	if (!rq->waitboost && !dma_fence_is_signaled_locked(&rq->fence)) {
6383		boost = !atomic_fetch_inc(&rps->num_waiters);
6384		rq->waitboost = true;
6385	}
6386	spin_unlock_irqrestore(&rq->lock, flags);
6387	if (!boost)
6388		return;
6389
6390	if (READ_ONCE(rps->cur_freq) < rps->boost_freq)
6391		schedule_work(&rps->work);
6392
6393	atomic_inc(rps_client ? &rps_client->boosts : &rps->boosts);
6394}
6395
6396int intel_set_rps(struct drm_i915_private *dev_priv, u8 val)
6397{
6398	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6399	int err;
6400
6401	lockdep_assert_held(&dev_priv->pcu_lock);
6402	GEM_BUG_ON(val > rps->max_freq);
6403	GEM_BUG_ON(val < rps->min_freq);
6404
6405	if (!rps->enabled) {
6406		rps->cur_freq = val;
6407		return 0;
6408	}
6409
6410	if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
6411		err = valleyview_set_rps(dev_priv, val);
6412	else
6413		err = gen6_set_rps(dev_priv, val);
6414
6415	return err;
6416}
6417
6418static void gen9_disable_rc6(struct drm_i915_private *dev_priv)
6419{
6420	I915_WRITE(GEN6_RC_CONTROL, 0);
6421	I915_WRITE(GEN9_PG_ENABLE, 0);
6422}
6423
6424static void gen9_disable_rps(struct drm_i915_private *dev_priv)
6425{
6426	I915_WRITE(GEN6_RP_CONTROL, 0);
6427}
6428
6429static void gen6_disable_rc6(struct drm_i915_private *dev_priv)
6430{
6431	I915_WRITE(GEN6_RC_CONTROL, 0);
6432}
6433
6434static void gen6_disable_rps(struct drm_i915_private *dev_priv)
6435{
6436	I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
6437	I915_WRITE(GEN6_RP_CONTROL, 0);
6438}
6439
6440static void cherryview_disable_rc6(struct drm_i915_private *dev_priv)
6441{
6442	I915_WRITE(GEN6_RC_CONTROL, 0);
6443}
6444
6445static void cherryview_disable_rps(struct drm_i915_private *dev_priv)
6446{
6447	I915_WRITE(GEN6_RP_CONTROL, 0);
6448}
6449
6450static void valleyview_disable_rc6(struct drm_i915_private *dev_priv)
6451{
6452	/* We're doing forcewake before Disabling RC6,
6453	 * This what the BIOS expects when going into suspend */
6454	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6455
6456	I915_WRITE(GEN6_RC_CONTROL, 0);
6457
6458	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6459}
6460
6461static void valleyview_disable_rps(struct drm_i915_private *dev_priv)
6462{
6463	I915_WRITE(GEN6_RP_CONTROL, 0);
6464}
6465
6466static bool bxt_check_bios_rc6_setup(struct drm_i915_private *dev_priv)
6467{
6468	bool enable_rc6 = true;
6469	unsigned long rc6_ctx_base;
6470	u32 rc_ctl;
6471	int rc_sw_target;
6472
6473	rc_ctl = I915_READ(GEN6_RC_CONTROL);
6474	rc_sw_target = (I915_READ(GEN6_RC_STATE) & RC_SW_TARGET_STATE_MASK) >>
6475		       RC_SW_TARGET_STATE_SHIFT;
6476	DRM_DEBUG_DRIVER("BIOS enabled RC states: "
6477			 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
6478			 onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
6479			 onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
6480			 rc_sw_target);
6481
6482	if (!(I915_READ(RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
6483		DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
6484		enable_rc6 = false;
6485	}
6486
6487	/*
6488	 * The exact context size is not known for BXT, so assume a page size
6489	 * for this check.
6490	 */
6491	rc6_ctx_base = I915_READ(RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
6492	if (!((rc6_ctx_base >= dev_priv->dsm_reserved.start) &&
6493	      (rc6_ctx_base + PAGE_SIZE < dev_priv->dsm_reserved.end))) {
6494		DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
6495		enable_rc6 = false;
6496	}
6497
6498	if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1) &&
6499	      ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1) &&
6500	      ((I915_READ(PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1) &&
6501	      ((I915_READ(PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1))) {
6502		DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
6503		enable_rc6 = false;
6504	}
6505
6506	if (!I915_READ(GEN8_PUSHBUS_CONTROL) ||
6507	    !I915_READ(GEN8_PUSHBUS_ENABLE) ||
6508	    !I915_READ(GEN8_PUSHBUS_SHIFT)) {
6509		DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
6510		enable_rc6 = false;
6511	}
6512
6513	if (!I915_READ(GEN6_GFXPAUSE)) {
6514		DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
6515		enable_rc6 = false;
6516	}
6517
6518	if (!I915_READ(GEN8_MISC_CTRL0)) {
6519		DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
6520		enable_rc6 = false;
6521	}
6522
6523	return enable_rc6;
6524}
6525
6526static bool sanitize_rc6(struct drm_i915_private *i915)
6527{
6528	struct intel_device_info *info = mkwrite_device_info(i915);
6529
6530	/* Powersaving is controlled by the host when inside a VM */
6531	if (intel_vgpu_active(i915))
6532		info->has_rc6 = 0;
6533
6534	if (info->has_rc6 &&
6535	    IS_GEN9_LP(i915) && !bxt_check_bios_rc6_setup(i915)) {
6536		DRM_INFO("RC6 disabled by BIOS\n");
6537		info->has_rc6 = 0;
6538	}
6539
6540	/*
6541	 * We assume that we do not have any deep rc6 levels if we don't have
6542	 * have the previous rc6 level supported, i.e. we use HAS_RC6()
6543	 * as the initial coarse check for rc6 in general, moving on to
6544	 * progressively finer/deeper levels.
6545	 */
6546	if (!info->has_rc6 && info->has_rc6p)
6547		info->has_rc6p = 0;
6548
6549	return info->has_rc6;
6550}
6551
6552static void gen6_init_rps_frequencies(struct drm_i915_private *dev_priv)
6553{
6554	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6555
6556	/* All of these values are in units of 50MHz */
6557
6558	/* static values from HW: RP0 > RP1 > RPn (min_freq) */
6559	if (IS_GEN9_LP(dev_priv)) {
6560		u32 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
6561		rps->rp0_freq = (rp_state_cap >> 16) & 0xff;
6562		rps->rp1_freq = (rp_state_cap >>  8) & 0xff;
6563		rps->min_freq = (rp_state_cap >>  0) & 0xff;
6564	} else {
6565		u32 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
6566		rps->rp0_freq = (rp_state_cap >>  0) & 0xff;
6567		rps->rp1_freq = (rp_state_cap >>  8) & 0xff;
6568		rps->min_freq = (rp_state_cap >> 16) & 0xff;
6569	}
6570	/* hw_max = RP0 until we check for overclocking */
6571	rps->max_freq = rps->rp0_freq;
6572
6573	rps->efficient_freq = rps->rp1_freq;
6574	if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv) ||
6575	    IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
6576		u32 ddcc_status = 0;
6577
6578		if (sandybridge_pcode_read(dev_priv,
6579					   HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
6580					   &ddcc_status) == 0)
6581			rps->efficient_freq =
6582				clamp_t(u8,
6583					((ddcc_status >> 8) & 0xff),
6584					rps->min_freq,
6585					rps->max_freq);
6586	}
6587
6588	if (IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
6589		/* Store the frequency values in 16.66 MHZ units, which is
6590		 * the natural hardware unit for SKL
6591		 */
6592		rps->rp0_freq *= GEN9_FREQ_SCALER;
6593		rps->rp1_freq *= GEN9_FREQ_SCALER;
6594		rps->min_freq *= GEN9_FREQ_SCALER;
6595		rps->max_freq *= GEN9_FREQ_SCALER;
6596		rps->efficient_freq *= GEN9_FREQ_SCALER;
6597	}
6598}
6599
6600static void reset_rps(struct drm_i915_private *dev_priv,
6601		      int (*set)(struct drm_i915_private *, u8))
6602{
6603	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6604	u8 freq = rps->cur_freq;
6605
6606	/* force a reset */
6607	rps->power = -1;
6608	rps->cur_freq = -1;
6609
6610	if (set(dev_priv, freq))
6611		DRM_ERROR("Failed to reset RPS to initial values\n");
6612}
6613
6614/* See the Gen9_GT_PM_Programming_Guide doc for the below */
6615static void gen9_enable_rps(struct drm_i915_private *dev_priv)
6616{
6617	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6618
6619	/* Program defaults and thresholds for RPS */
6620	if (IS_GEN9(dev_priv))
6621		I915_WRITE(GEN6_RC_VIDEO_FREQ,
6622			GEN9_FREQUENCY(dev_priv->gt_pm.rps.rp1_freq));
6623
6624	/* 1 second timeout*/
6625	I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
6626		GT_INTERVAL_FROM_US(dev_priv, 1000000));
6627
6628	I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
6629
6630	/* Leaning on the below call to gen6_set_rps to program/setup the
6631	 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
6632	 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
6633	reset_rps(dev_priv, gen6_set_rps);
6634
6635	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6636}
6637
6638static void gen9_enable_rc6(struct drm_i915_private *dev_priv)
6639{
6640	struct intel_engine_cs *engine;
6641	enum intel_engine_id id;
6642	u32 rc6_mode;
6643
6644	/* 1a: Software RC state - RC0 */
6645	I915_WRITE(GEN6_RC_STATE, 0);
6646
6647	/* 1b: Get forcewake during program sequence. Although the driver
6648	 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6649	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6650
6651	/* 2a: Disable RC states. */
6652	I915_WRITE(GEN6_RC_CONTROL, 0);
6653
6654	/* 2b: Program RC6 thresholds.*/
6655	if (INTEL_GEN(dev_priv) >= 10) {
6656		I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16 | 85);
6657		I915_WRITE(GEN10_MEDIA_WAKE_RATE_LIMIT, 150);
6658	} else if (IS_SKYLAKE(dev_priv)) {
6659		/*
6660		 * WaRsDoubleRc6WrlWithCoarsePowerGating:skl Doubling WRL only
6661		 * when CPG is enabled
6662		 */
6663		I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
6664	} else {
6665		I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
6666	}
6667
6668	I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
6669	I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
6670	for_each_engine(engine, dev_priv, id)
6671		I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6672
6673	if (HAS_GUC(dev_priv))
6674		I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
6675
6676	I915_WRITE(GEN6_RC_SLEEP, 0);
6677
6678	/*
6679	 * 2c: Program Coarse Power Gating Policies.
6680	 *
6681	 * Bspec's guidance is to use 25us (really 25 * 1280ns) here. What we
6682	 * use instead is a more conservative estimate for the maximum time
6683	 * it takes us to service a CS interrupt and submit a new ELSP - that
6684	 * is the time which the GPU is idle waiting for the CPU to select the
6685	 * next request to execute. If the idle hysteresis is less than that
6686	 * interrupt service latency, the hardware will automatically gate
6687	 * the power well and we will then incur the wake up cost on top of
6688	 * the service latency. A similar guide from intel_pstate is that we
6689	 * do not want the enable hysteresis to less than the wakeup latency.
6690	 *
6691	 * igt/gem_exec_nop/sequential provides a rough estimate for the
6692	 * service latency, and puts it around 10us for Broadwell (and other
6693	 * big core) and around 40us for Broxton (and other low power cores).
6694	 * [Note that for legacy ringbuffer submission, this is less than 1us!]
6695	 * However, the wakeup latency on Broxton is closer to 100us. To be
6696	 * conservative, we have to factor in a context switch on top (due
6697	 * to ksoftirqd).
6698	 */
6699	I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 250);
6700	I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 250);
6701
6702	/* 3a: Enable RC6 */
6703	I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
6704
6705	/* WaRsUseTimeoutMode:cnl (pre-prod) */
6706	if (IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_C0))
6707		rc6_mode = GEN7_RC_CTL_TO_MODE;
6708	else
6709		rc6_mode = GEN6_RC_CTL_EI_MODE(1);
6710
6711	I915_WRITE(GEN6_RC_CONTROL,
6712		   GEN6_RC_CTL_HW_ENABLE |
6713		   GEN6_RC_CTL_RC6_ENABLE |
6714		   rc6_mode);
6715
6716	/*
6717	 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
6718	 * WaRsDisableCoarsePowerGating:skl,cnl - Render/Media PG need to be disabled with RC6.
6719	 */
6720	if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv))
6721		I915_WRITE(GEN9_PG_ENABLE, 0);
6722	else
6723		I915_WRITE(GEN9_PG_ENABLE,
6724			   GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE);
6725
6726	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6727}
6728
6729static void gen8_enable_rc6(struct drm_i915_private *dev_priv)
6730{
6731	struct intel_engine_cs *engine;
6732	enum intel_engine_id id;
6733
6734	/* 1a: Software RC state - RC0 */
6735	I915_WRITE(GEN6_RC_STATE, 0);
6736
6737	/* 1b: Get forcewake during program sequence. Although the driver
6738	 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6739	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6740
6741	/* 2a: Disable RC states. */
6742	I915_WRITE(GEN6_RC_CONTROL, 0);
6743
6744	/* 2b: Program RC6 thresholds.*/
6745	I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
6746	I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
6747	I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
6748	for_each_engine(engine, dev_priv, id)
6749		I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6750	I915_WRITE(GEN6_RC_SLEEP, 0);
6751	I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
6752
6753	/* 3: Enable RC6 */
6754
6755	I915_WRITE(GEN6_RC_CONTROL,
6756		   GEN6_RC_CTL_HW_ENABLE |
6757		   GEN7_RC_CTL_TO_MODE |
6758		   GEN6_RC_CTL_RC6_ENABLE);
6759
6760	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6761}
6762
6763static void gen8_enable_rps(struct drm_i915_private *dev_priv)
6764{
6765	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6766
6767	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6768
6769	/* 1 Program defaults and thresholds for RPS*/
6770	I915_WRITE(GEN6_RPNSWREQ,
6771		   HSW_FREQUENCY(rps->rp1_freq));
6772	I915_WRITE(GEN6_RC_VIDEO_FREQ,
6773		   HSW_FREQUENCY(rps->rp1_freq));
6774	/* NB: Docs say 1s, and 1000000 - which aren't equivalent */
6775	I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
6776
6777	/* Docs recommend 900MHz, and 300 MHz respectively */
6778	I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
6779		   rps->max_freq_softlimit << 24 |
6780		   rps->min_freq_softlimit << 16);
6781
6782	I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
6783	I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
6784	I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
6785	I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
6786
6787	I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6788
6789	/* 2: Enable RPS */
6790	I915_WRITE(GEN6_RP_CONTROL,
6791		   GEN6_RP_MEDIA_TURBO |
6792		   GEN6_RP_MEDIA_HW_NORMAL_MODE |
6793		   GEN6_RP_MEDIA_IS_GFX |
6794		   GEN6_RP_ENABLE |
6795		   GEN6_RP_UP_BUSY_AVG |
6796		   GEN6_RP_DOWN_IDLE_AVG);
6797
6798	reset_rps(dev_priv, gen6_set_rps);
6799
6800	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6801}
6802
6803static void gen6_enable_rc6(struct drm_i915_private *dev_priv)
6804{
6805	struct intel_engine_cs *engine;
6806	enum intel_engine_id id;
6807	u32 rc6vids, rc6_mask;
6808	u32 gtfifodbg;
6809	int ret;
6810
6811	I915_WRITE(GEN6_RC_STATE, 0);
6812
6813	/* Clear the DBG now so we don't confuse earlier errors */
6814	gtfifodbg = I915_READ(GTFIFODBG);
6815	if (gtfifodbg) {
6816		DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
6817		I915_WRITE(GTFIFODBG, gtfifodbg);
6818	}
6819
6820	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6821
6822	/* disable the counters and set deterministic thresholds */
6823	I915_WRITE(GEN6_RC_CONTROL, 0);
6824
6825	I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
6826	I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
6827	I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
6828	I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
6829	I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
6830
6831	for_each_engine(engine, dev_priv, id)
6832		I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6833
6834	I915_WRITE(GEN6_RC_SLEEP, 0);
6835	I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
6836	if (IS_IVYBRIDGE(dev_priv))
6837		I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
6838	else
6839		I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
6840	I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
6841	I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
6842
6843	/* We don't use those on Haswell */
6844	rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
6845	if (HAS_RC6p(dev_priv))
6846		rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
6847	if (HAS_RC6pp(dev_priv))
6848		rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
6849	I915_WRITE(GEN6_RC_CONTROL,
6850		   rc6_mask |
6851		   GEN6_RC_CTL_EI_MODE(1) |
6852		   GEN6_RC_CTL_HW_ENABLE);
6853
6854	rc6vids = 0;
6855	ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
6856	if (IS_GEN6(dev_priv) && ret) {
6857		DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
6858	} else if (IS_GEN6(dev_priv) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
6859		DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
6860			  GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
6861		rc6vids &= 0xffff00;
6862		rc6vids |= GEN6_ENCODE_RC6_VID(450);
6863		ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
6864		if (ret)
6865			DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
6866	}
6867
6868	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6869}
6870
6871static void gen6_enable_rps(struct drm_i915_private *dev_priv)
6872{
6873	/* Here begins a magic sequence of register writes to enable
6874	 * auto-downclocking.
6875	 *
6876	 * Perhaps there might be some value in exposing these to
6877	 * userspace...
6878	 */
6879	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6880
6881	/* Power down if completely idle for over 50ms */
6882	I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
6883	I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6884
6885	reset_rps(dev_priv, gen6_set_rps);
6886
6887	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6888}
6889
6890static void gen6_update_ring_freq(struct drm_i915_private *dev_priv)
6891{
6892	struct intel_rps *rps = &dev_priv->gt_pm.rps;
6893	int min_freq = 15;
6894	unsigned int gpu_freq;
6895	unsigned int max_ia_freq, min_ring_freq;
6896	unsigned int max_gpu_freq, min_gpu_freq;
6897	int scaling_factor = 180;
6898	struct cpufreq_policy *policy;
6899
6900	WARN_ON(!mutex_is_locked(&dev_priv->pcu_lock));
6901
6902	policy = cpufreq_cpu_get(0);
6903	if (policy) {
6904		max_ia_freq = policy->cpuinfo.max_freq;
6905		cpufreq_cpu_put(policy);
6906	} else {
6907		/*
6908		 * Default to measured freq if none found, PCU will ensure we
6909		 * don't go over
6910		 */
6911		max_ia_freq = tsc_khz;
6912	}
6913
6914	/* Convert from kHz to MHz */
6915	max_ia_freq /= 1000;
6916
6917	min_ring_freq = I915_READ(DCLK) & 0xf;
6918	/* convert DDR frequency from units of 266.6MHz to bandwidth */
6919	min_ring_freq = mult_frac(min_ring_freq, 8, 3);
6920
6921	if (IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
6922		/* Convert GT frequency to 50 HZ units */
6923		min_gpu_freq = rps->min_freq / GEN9_FREQ_SCALER;
6924		max_gpu_freq = rps->max_freq / GEN9_FREQ_SCALER;
6925	} else {
6926		min_gpu_freq = rps->min_freq;
6927		max_gpu_freq = rps->max_freq;
6928	}
6929
6930	/*
6931	 * For each potential GPU frequency, load a ring frequency we'd like
6932	 * to use for memory access.  We do this by specifying the IA frequency
6933	 * the PCU should use as a reference to determine the ring frequency.
6934	 */
6935	for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
6936		int diff = max_gpu_freq - gpu_freq;
6937		unsigned int ia_freq = 0, ring_freq = 0;
6938
6939		if (IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
6940			/*
6941			 * ring_freq = 2 * GT. ring_freq is in 100MHz units
6942			 * No floor required for ring frequency on SKL.
6943			 */
6944			ring_freq = gpu_freq;
6945		} else if (INTEL_GEN(dev_priv) >= 8) {
6946			/* max(2 * GT, DDR). NB: GT is 50MHz units */
6947			ring_freq = max(min_ring_freq, gpu_freq);
6948		} else if (IS_HASWELL(dev_priv)) {
6949			ring_freq = mult_frac(gpu_freq, 5, 4);
6950			ring_freq = max(min_ring_freq, ring_freq);
6951			/* leave ia_freq as the default, chosen by cpufreq */
6952		} else {
6953			/* On older processors, there is no separate ring
6954			 * clock domain, so in order to boost the bandwidth
6955			 * of the ring, we need to upclock the CPU (ia_freq).
6956			 *
6957			 * For GPU frequencies less than 750MHz,
6958			 * just use the lowest ring freq.
6959			 */
6960			if (gpu_freq < min_freq)
6961				ia_freq = 800;
6962			else
6963				ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
6964			ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
6965		}
6966
6967		sandybridge_pcode_write(dev_priv,
6968					GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
6969					ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
6970					ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
6971					gpu_freq);
6972	}
6973}
6974
6975static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
6976{
6977	u32 val, rp0;
6978
6979	val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
6980
6981	switch (INTEL_INFO(dev_priv)->sseu.eu_total) {
6982	case 8:
6983		/* (2 * 4) config */
6984		rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
6985		break;
6986	case 12:
6987		/* (2 * 6) config */
6988		rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
6989		break;
6990	case 16:
6991		/* (2 * 8) config */
6992	default:
6993		/* Setting (2 * 8) Min RP0 for any other combination */
6994		rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
6995		break;
6996	}
6997
6998	rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
6999
7000	return rp0;
7001}
7002
7003static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
7004{
7005	u32 val, rpe;
7006
7007	val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
7008	rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
7009
7010	return rpe;
7011}
7012
7013static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
7014{
7015	u32 val, rp1;
7016
7017	val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
7018	rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
7019
7020	return rp1;
7021}
7022
7023static u32 cherryview_rps_min_freq(struct drm_i915_private *dev_priv)
7024{
7025	u32 val, rpn;
7026
7027	val = vlv_punit_read(dev_priv, FB_GFX_FMIN_AT_VMIN_FUSE);
7028	rpn = ((val >> FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT) &
7029		       FB_GFX_FREQ_FUSE_MASK);
7030
7031	return rpn;
7032}
7033
7034static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
7035{
7036	u32 val, rp1;
7037
7038	val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
7039
7040	rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
7041
7042	return rp1;
7043}
7044
7045static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
7046{
7047	u32 val, rp0;
7048
7049	val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
7050
7051	rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
7052	/* Clamp to max */
7053	rp0 = min_t(u32, rp0, 0xea);
7054
7055	return rp0;
7056}
7057
7058static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
7059{
7060	u32 val, rpe;
7061
7062	val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
7063	rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
7064	val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
7065	rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
7066
7067	return rpe;
7068}
7069
7070static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
7071{
7072	u32 val;
7073
7074	val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
7075	/*
7076	 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
7077	 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
7078	 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
7079	 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
7080	 * to make sure it matches what Punit accepts.
7081	 */
7082	return max_t(u32, val, 0xc0);
7083}
7084
7085/* Check that the pctx buffer wasn't move under us. */
7086static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
7087{
7088	unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
7089
7090	WARN_ON(pctx_addr != dev_priv->dsm.start +
7091			     dev_priv->vlv_pctx->stolen->start);
7092}
7093
7094
7095/* Check that the pcbr address is not empty. */
7096static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
7097{
7098	unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
7099
7100	WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
7101}
7102
7103static void cherryview_setup_pctx(struct drm_i915_private *dev_priv)
7104{
7105	resource_size_t pctx_paddr, paddr;
7106	resource_size_t pctx_size = 32*1024;
7107	u32 pcbr;
7108
7109	pcbr = I915_READ(VLV_PCBR);
7110	if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
7111		DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
7112		paddr = dev_priv->dsm.end + 1 - pctx_size;
7113		GEM_BUG_ON(paddr > U32_MAX);
7114
7115		pctx_paddr = (paddr & (~4095));
7116		I915_WRITE(VLV_PCBR, pctx_paddr);
7117	}
7118
7119	DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
7120}
7121
7122static void valleyview_setup_pctx(struct drm_i915_private *dev_priv)
7123{
7124	struct drm_i915_gem_object *pctx;
7125	resource_size_t pctx_paddr;
7126	resource_size_t pctx_size = 24*1024;
7127	u32 pcbr;
7128
7129	pcbr = I915_READ(VLV_PCBR);
7130	if (pcbr) {
7131		/* BIOS set it up already, grab the pre-alloc'd space */
7132		resource_size_t pcbr_offset;
7133
7134		pcbr_offset = (pcbr & (~4095)) - dev_priv->dsm.start;
7135		pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv,
7136								      pcbr_offset,
7137								      I915_GTT_OFFSET_NONE,
7138								      pctx_size);
7139		goto out;
7140	}
7141
7142	DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
7143
7144	/*
7145	 * From the Gunit register HAS:
7146	 * The Gfx driver is expected to program this register and ensure
7147	 * proper allocation within Gfx stolen memory.  For example, this
7148	 * register should be programmed such than the PCBR range does not
7149	 * overlap with other ranges, such as the frame buffer, protected
7150	 * memory, or any other relevant ranges.
7151	 */
7152	pctx = i915_gem_object_create_stolen(dev_priv, pctx_size);
7153	if (!pctx) {
7154		DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
7155		goto out;
7156	}
7157
7158	GEM_BUG_ON(range_overflows_t(u64,
7159				     dev_priv->dsm.start,
7160				     pctx->stolen->start,
7161				     U32_MAX));
7162	pctx_paddr = dev_priv->dsm.start + pctx->stolen->start;
7163	I915_WRITE(VLV_PCBR, pctx_paddr);
7164
7165out:
7166	DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
7167	dev_priv->vlv_pctx = pctx;
7168}
7169
7170static void valleyview_cleanup_pctx(struct drm_i915_private *dev_priv)
7171{
7172	if (WARN_ON(!dev_priv->vlv_pctx))
7173		return;
7174
7175	i915_gem_object_put(dev_priv->vlv_pctx);
7176	dev_priv->vlv_pctx = NULL;
7177}
7178
7179static void vlv_init_gpll_ref_freq(struct drm_i915_private *dev_priv)
7180{
7181	dev_priv->gt_pm.rps.gpll_ref_freq =
7182		vlv_get_cck_clock(dev_priv, "GPLL ref",
7183				  CCK_GPLL_CLOCK_CONTROL,
7184				  dev_priv->czclk_freq);
7185
7186	DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
7187			 dev_priv->gt_pm.rps.gpll_ref_freq);
7188}
7189
7190static void valleyview_init_gt_powersave(struct drm_i915_private *dev_priv)
7191{
7192	struct intel_rps *rps = &dev_priv->gt_pm.rps;
7193	u32 val;
7194
7195	valleyview_setup_pctx(dev_priv);
7196
7197	vlv_init_gpll_ref_freq(dev_priv);
7198
7199	val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
7200	switch ((val >> 6) & 3) {
7201	case 0:
7202	case 1:
7203		dev_priv->mem_freq = 800;
7204		break;
7205	case 2:
7206		dev_priv->mem_freq = 1066;
7207		break;
7208	case 3:
7209		dev_priv->mem_freq = 1333;
7210		break;
7211	}
7212	DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
7213
7214	rps->max_freq = valleyview_rps_max_freq(dev_priv);
7215	rps->rp0_freq = rps->max_freq;
7216	DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
7217			 intel_gpu_freq(dev_priv, rps->max_freq),
7218			 rps->max_freq);
7219
7220	rps->efficient_freq = valleyview_rps_rpe_freq(dev_priv);
7221	DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
7222			 intel_gpu_freq(dev_priv, rps->efficient_freq),
7223			 rps->efficient_freq);
7224
7225	rps->rp1_freq = valleyview_rps_guar_freq(dev_priv);
7226	DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
7227			 intel_gpu_freq(dev_priv, rps->rp1_freq),
7228			 rps->rp1_freq);
7229
7230	rps->min_freq = valleyview_rps_min_freq(dev_priv);
7231	DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
7232			 intel_gpu_freq(dev_priv, rps->min_freq),
7233			 rps->min_freq);
7234}
7235
7236static void cherryview_init_gt_powersave(struct drm_i915_private *dev_priv)
7237{
7238	struct intel_rps *rps = &dev_priv->gt_pm.rps;
7239	u32 val;
7240
7241	cherryview_setup_pctx(dev_priv);
7242
7243	vlv_init_gpll_ref_freq(dev_priv);
7244
7245	mutex_lock(&dev_priv->sb_lock);
7246	val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
7247	mutex_unlock(&dev_priv->sb_lock);
7248
7249	switch ((val >> 2) & 0x7) {
7250	case 3:
7251		dev_priv->mem_freq = 2000;
7252		break;
7253	default:
7254		dev_priv->mem_freq = 1600;
7255		break;
7256	}
7257	DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
7258
7259	rps->max_freq = cherryview_rps_max_freq(dev_priv);
7260	rps->rp0_freq = rps->max_freq;
7261	DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
7262			 intel_gpu_freq(dev_priv, rps->max_freq),
7263			 rps->max_freq);
7264
7265	rps->efficient_freq = cherryview_rps_rpe_freq(dev_priv);
7266	DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
7267			 intel_gpu_freq(dev_priv, rps->efficient_freq),
7268			 rps->efficient_freq);
7269
7270	rps->rp1_freq = cherryview_rps_guar_freq(dev_priv);
7271	DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
7272			 intel_gpu_freq(dev_priv, rps->rp1_freq),
7273			 rps->rp1_freq);
7274
7275	rps->min_freq = cherryview_rps_min_freq(dev_priv);
7276	DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
7277			 intel_gpu_freq(dev_priv, rps->min_freq),
7278			 rps->min_freq);
7279
7280	WARN_ONCE((rps->max_freq | rps->efficient_freq | rps->rp1_freq |
7281		   rps->min_freq) & 1,
7282		  "Odd GPU freq values\n");
7283}
7284
7285static void valleyview_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
7286{
7287	valleyview_cleanup_pctx(dev_priv);
7288}
7289
7290static void cherryview_enable_rc6(struct drm_i915_private *dev_priv)
7291{
7292	struct intel_engine_cs *engine;
7293	enum intel_engine_id id;
7294	u32 gtfifodbg, rc6_mode, pcbr;
7295
7296	gtfifodbg = I915_READ(GTFIFODBG) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV |
7297					     GT_FIFO_FREE_ENTRIES_CHV);
7298	if (gtfifodbg) {
7299		DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
7300				 gtfifodbg);
7301		I915_WRITE(GTFIFODBG, gtfifodbg);
7302	}
7303
7304	cherryview_check_pctx(dev_priv);
7305
7306	/* 1a & 1b: Get forcewake during program sequence. Although the driver
7307	 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
7308	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
7309
7310	/*  Disable RC states. */
7311	I915_WRITE(GEN6_RC_CONTROL, 0);
7312
7313	/* 2a: Program RC6 thresholds.*/
7314	I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
7315	I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
7316	I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
7317
7318	for_each_engine(engine, dev_priv, id)
7319		I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
7320	I915_WRITE(GEN6_RC_SLEEP, 0);
7321
7322	/* TO threshold set to 500 us ( 0x186 * 1.28 us) */
7323	I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
7324
7325	/* Allows RC6 residency counter to work */
7326	I915_WRITE(VLV_COUNTER_CONTROL,
7327		   _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
7328				      VLV_MEDIA_RC6_COUNT_EN |
7329				      VLV_RENDER_RC6_COUNT_EN));
7330
7331	/* For now we assume BIOS is allocating and populating the PCBR  */
7332	pcbr = I915_READ(VLV_PCBR);
7333
7334	/* 3: Enable RC6 */
7335	rc6_mode = 0;
7336	if (pcbr >> VLV_PCBR_ADDR_SHIFT)
7337		rc6_mode = GEN7_RC_CTL_TO_MODE;
7338	I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
7339
7340	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
7341}
7342
7343static void cherryview_enable_rps(struct drm_i915_private *dev_priv)
7344{
7345	u32 val;
7346
7347	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
7348
7349	/* 1: Program defaults and thresholds for RPS*/
7350	I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
7351	I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
7352	I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
7353	I915_WRITE(GEN6_RP_UP_EI, 66000);
7354	I915_WRITE(GEN6_RP_DOWN_EI, 350000);
7355
7356	I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
7357
7358	/* 2: Enable RPS */
7359	I915_WRITE(GEN6_RP_CONTROL,
7360		   GEN6_RP_MEDIA_HW_NORMAL_MODE |
7361		   GEN6_RP_MEDIA_IS_GFX |
7362		   GEN6_RP_ENABLE |
7363		   GEN6_RP_UP_BUSY_AVG |
7364		   GEN6_RP_DOWN_IDLE_AVG);
7365
7366	/* Setting Fixed Bias */
7367	val = VLV_OVERRIDE_EN |
7368		  VLV_SOC_TDP_EN |
7369		  CHV_BIAS_CPU_50_SOC_50;
7370	vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
7371
7372	val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
7373
7374	/* RPS code assumes GPLL is used */
7375	WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
7376
7377	DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
7378	DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
7379
7380	reset_rps(dev_priv, valleyview_set_rps);
7381
7382	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
7383}
7384
7385static void valleyview_enable_rc6(struct drm_i915_private *dev_priv)
7386{
7387	struct intel_engine_cs *engine;
7388	enum intel_engine_id id;
7389	u32 gtfifodbg;
7390
7391	valleyview_check_pctx(dev_priv);
7392
7393	gtfifodbg = I915_READ(GTFIFODBG);
7394	if (gtfifodbg) {
7395		DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
7396				 gtfifodbg);
7397		I915_WRITE(GTFIFODBG, gtfifodbg);
7398	}
7399
7400	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
7401
7402	/*  Disable RC states. */
7403	I915_WRITE(GEN6_RC_CONTROL, 0);
7404
7405	I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
7406	I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
7407	I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
7408
7409	for_each_engine(engine, dev_priv, id)
7410		I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
7411
7412	I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
7413
7414	/* Allows RC6 residency counter to work */
7415	I915_WRITE(VLV_COUNTER_CONTROL,
7416		   _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
7417				      VLV_MEDIA_RC0_COUNT_EN |
7418				      VLV_RENDER_RC0_COUNT_EN |
7419				      VLV_MEDIA_RC6_COUNT_EN |
7420				      VLV_RENDER_RC6_COUNT_EN));
7421
7422	I915_WRITE(GEN6_RC_CONTROL,
7423		   GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL);
7424
7425	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
7426}
7427
7428static void valleyview_enable_rps(struct drm_i915_private *dev_priv)
7429{
7430	u32 val;
7431
7432	intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
7433
7434	I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
7435	I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
7436	I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
7437	I915_WRITE(GEN6_RP_UP_EI, 66000);
7438	I915_WRITE(GEN6_RP_DOWN_EI, 350000);
7439
7440	I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
7441
7442	I915_WRITE(GEN6_RP_CONTROL,
7443		   GEN6_RP_MEDIA_TURBO |
7444		   GEN6_RP_MEDIA_HW_NORMAL_MODE |
7445		   GEN6_RP_MEDIA_IS_GFX |
7446		   GEN6_RP_ENABLE |
7447		   GEN6_RP_UP_BUSY_AVG |
7448		   GEN6_RP_DOWN_IDLE_CONT);
7449
7450	/* Setting Fixed Bias */
7451	val = VLV_OVERRIDE_EN |
7452		  VLV_SOC_TDP_EN |
7453		  VLV_BIAS_CPU_125_SOC_875;
7454	vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
7455
7456	val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
7457
7458	/* RPS code assumes GPLL is used */
7459	WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
7460
7461	DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
7462	DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
7463
7464	reset_rps(dev_priv, valleyview_set_rps);
7465
7466	intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
7467}
7468
7469static unsigned long intel_pxfreq(u32 vidfreq)
7470{
7471	unsigned long freq;
7472	int div = (vidfreq & 0x3f0000) >> 16;
7473	int post = (vidfreq & 0x3000) >> 12;
7474	int pre = (vidfreq & 0x7);
7475
7476	if (!pre)
7477		return 0;
7478
7479	freq = ((div * 133333) / ((1<<post) * pre));
7480
7481	return freq;
7482}
7483
7484static const struct cparams {
7485	u16 i;
7486	u16 t;
7487	u16 m;
7488	u16 c;
7489} cparams[] = {
7490	{ 1, 1333, 301, 28664 },
7491	{ 1, 1066, 294, 24460 },
7492	{ 1, 800, 294, 25192 },
7493	{ 0, 1333, 276, 27605 },
7494	{ 0, 1066, 276, 27605 },
7495	{ 0, 800, 231, 23784 },
7496};
7497
7498static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
7499{
7500	u64 total_count, diff, ret;
7501	u32 count1, count2, count3, m = 0, c = 0;
7502	unsigned long now = jiffies_to_msecs(jiffies), diff1;
7503	int i;
7504
7505	lockdep_assert_held(&mchdev_lock);
7506
7507	diff1 = now - dev_priv->ips.last_time1;
7508
7509	/* Prevent division-by-zero if we are asking too fast.
7510	 * Also, we don't get interesting results if we are polling
7511	 * faster than once in 10ms, so just return the saved value
7512	 * in such cases.
7513	 */
7514	if (diff1 <= 10)
7515		return dev_priv->ips.chipset_power;
7516
7517	count1 = I915_READ(DMIEC);
7518	count2 = I915_READ(DDREC);
7519	count3 = I915_READ(CSIEC);
7520
7521	total_count = count1 + count2 + count3;
7522
7523	/* FIXME: handle per-counter overflow */
7524	if (total_count < dev_priv->ips.last_count1) {
7525		diff = ~0UL - dev_priv->ips.last_count1;
7526		diff += total_count;
7527	} else {
7528		diff = total_count - dev_priv->ips.last_count1;
7529	}
7530
7531	for (i = 0; i < ARRAY_SIZE(cparams); i++) {
7532		if (cparams[i].i == dev_priv->ips.c_m &&
7533		    cparams[i].t == dev_priv->ips.r_t) {
7534			m = cparams[i].m;
7535			c = cparams[i].c;
7536			break;
7537		}
7538	}
7539
7540	diff = div_u64(diff, diff1);
7541	ret = ((m * diff) + c);
7542	ret = div_u64(ret, 10);
7543
7544	dev_priv->ips.last_count1 = total_count;
7545	dev_priv->ips.last_time1 = now;
7546
7547	dev_priv->ips.chipset_power = ret;
7548
7549	return ret;
7550}
7551
7552unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
7553{
7554	unsigned long val;
7555
7556	if (!IS_GEN5(dev_priv))
7557		return 0;
7558
7559	spin_lock_irq(&mchdev_lock);
7560
7561	val = __i915_chipset_val(dev_priv);
7562
7563	spin_unlock_irq(&mchdev_lock);
7564
7565	return val;
7566}
7567
7568unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
7569{
7570	unsigned long m, x, b;
7571	u32 tsfs;
7572
7573	tsfs = I915_READ(TSFS);
7574
7575	m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
7576	x = I915_READ8(TR1);
7577
7578	b = tsfs & TSFS_INTR_MASK;
7579
7580	return ((m * x) / 127) - b;
7581}
7582
7583static int _pxvid_to_vd(u8 pxvid)
7584{
7585	if (pxvid == 0)
7586		return 0;
7587
7588	if (pxvid >= 8 && pxvid < 31)
7589		pxvid = 31;
7590
7591	return (pxvid + 2) * 125;
7592}
7593
7594static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
7595{
7596	const int vd = _pxvid_to_vd(pxvid);
7597	const int vm = vd - 1125;
7598
7599	if (INTEL_INFO(dev_priv)->is_mobile)
7600		return vm > 0 ? vm : 0;
7601
7602	return vd;
7603}
7604
7605static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
7606{
7607	u64 now, diff, diffms;
7608	u32 count;
7609
7610	lockdep_assert_held(&mchdev_lock);
7611
7612	now = ktime_get_raw_ns();
7613	diffms = now - dev_priv->ips.last_time2;
7614	do_div(diffms, NSEC_PER_MSEC);
7615
7616	/* Don't divide by 0 */
7617	if (!diffms)
7618		return;
7619
7620	count = I915_READ(GFXEC);
7621
7622	if (count < dev_priv->ips.last_count2) {
7623		diff = ~0UL - dev_priv->ips.last_count2;
7624		diff += count;
7625	} else {
7626		diff = count - dev_priv->ips.last_count2;
7627	}
7628
7629	dev_priv->ips.last_count2 = count;
7630	dev_priv->ips.last_time2 = now;
7631
7632	/* More magic constants... */
7633	diff = diff * 1181;
7634	diff = div_u64(diff, diffms * 10);
7635	dev_priv->ips.gfx_power = diff;
7636}
7637
7638void i915_update_gfx_val(struct drm_i915_private *dev_priv)
7639{
7640	if (!IS_GEN5(dev_priv))
7641		return;
7642
7643	spin_lock_irq(&mchdev_lock);
7644
7645	__i915_update_gfx_val(dev_priv);
7646
7647	spin_unlock_irq(&mchdev_lock);
7648}
7649
7650static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
7651{
7652	unsigned long t, corr, state1, corr2, state2;
7653	u32 pxvid, ext_v;
7654
7655	lockdep_assert_held(&mchdev_lock);
7656
7657	pxvid = I915_READ(PXVFREQ(dev_priv->gt_pm.rps.cur_freq));
7658	pxvid = (pxvid >> 24) & 0x7f;
7659	ext_v = pvid_to_extvid(dev_priv, pxvid);
7660
7661	state1 = ext_v;
7662
7663	t = i915_mch_val(dev_priv);
7664
7665	/* Revel in the empirically derived constants */
7666
7667	/* Correction factor in 1/100000 units */
7668	if (t > 80)
7669		corr = ((t * 2349) + 135940);
7670	else if (t >= 50)
7671		corr = ((t * 964) + 29317);
7672	else /* < 50 */
7673		corr = ((t * 301) + 1004);
7674
7675	corr = corr * ((150142 * state1) / 10000 - 78642);
7676	corr /= 100000;
7677	corr2 = (corr * dev_priv->ips.corr);
7678
7679	state2 = (corr2 * state1) / 10000;
7680	state2 /= 100; /* convert to mW */
7681
7682	__i915_update_gfx_val(dev_priv);
7683
7684	return dev_priv->ips.gfx_power + state2;
7685}
7686
7687unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
7688{
7689	unsigned long val;
7690
7691	if (!IS_GEN5(dev_priv))
7692		return 0;
7693
7694	spin_lock_irq(&mchdev_lock);
7695
7696	val = __i915_gfx_val(dev_priv);
7697
7698	spin_unlock_irq(&mchdev_lock);
7699
7700	return val;
7701}
7702
7703/**
7704 * i915_read_mch_val - return value for IPS use
7705 *
7706 * Calculate and return a value for the IPS driver to use when deciding whether
7707 * we have thermal and power headroom to increase CPU or GPU power budget.
7708 */
7709unsigned long i915_read_mch_val(void)
7710{
7711	struct drm_i915_private *dev_priv;
7712	unsigned long chipset_val, graphics_val, ret = 0;
7713
7714	spin_lock_irq(&mchdev_lock);
7715	if (!i915_mch_dev)
7716		goto out_unlock;
7717	dev_priv = i915_mch_dev;
7718
7719	chipset_val = __i915_chipset_val(dev_priv);
7720	graphics_val = __i915_gfx_val(dev_priv);
7721
7722	ret = chipset_val + graphics_val;
7723
7724out_unlock:
7725	spin_unlock_irq(&mchdev_lock);
7726
7727	return ret;
7728}
7729EXPORT_SYMBOL_GPL(i915_read_mch_val);
7730
7731/**
7732 * i915_gpu_raise - raise GPU frequency limit
7733 *
7734 * Raise the limit; IPS indicates we have thermal headroom.
7735 */
7736bool i915_gpu_raise(void)
7737{
7738	struct drm_i915_private *dev_priv;
7739	bool ret = true;
7740
7741	spin_lock_irq(&mchdev_lock);
7742	if (!i915_mch_dev) {
7743		ret = false;
7744		goto out_unlock;
7745	}
7746	dev_priv = i915_mch_dev;
7747
7748	if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
7749		dev_priv->ips.max_delay--;
7750
7751out_unlock:
7752	spin_unlock_irq(&mchdev_lock);
7753
7754	return ret;
7755}
7756EXPORT_SYMBOL_GPL(i915_gpu_raise);
7757
7758/**
7759 * i915_gpu_lower - lower GPU frequency limit
7760 *
7761 * IPS indicates we're close to a thermal limit, so throttle back the GPU
7762 * frequency maximum.
7763 */
7764bool i915_gpu_lower(void)
7765{
7766	struct drm_i915_private *dev_priv;
7767	bool ret = true;
7768
7769	spin_lock_irq(&mchdev_lock);
7770	if (!i915_mch_dev) {
7771		ret = false;
7772		goto out_unlock;
7773	}
7774	dev_priv = i915_mch_dev;
7775
7776	if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
7777		dev_priv->ips.max_delay++;
7778
7779out_unlock:
7780	spin_unlock_irq(&mchdev_lock);
7781
7782	return ret;
7783}
7784EXPORT_SYMBOL_GPL(i915_gpu_lower);
7785
7786/**
7787 * i915_gpu_busy - indicate GPU business to IPS
7788 *
7789 * Tell the IPS driver whether or not the GPU is busy.
7790 */
7791bool i915_gpu_busy(void)
7792{
7793	bool ret = false;
7794
7795	spin_lock_irq(&mchdev_lock);
7796	if (i915_mch_dev)
7797		ret = i915_mch_dev->gt.awake;
7798	spin_unlock_irq(&mchdev_lock);
7799
7800	return ret;
7801}
7802EXPORT_SYMBOL_GPL(i915_gpu_busy);
7803
7804/**
7805 * i915_gpu_turbo_disable - disable graphics turbo
7806 *
7807 * Disable graphics turbo by resetting the max frequency and setting the
7808 * current frequency to the default.
7809 */
7810bool i915_gpu_turbo_disable(void)
7811{
7812	struct drm_i915_private *dev_priv;
7813	bool ret = true;
7814
7815	spin_lock_irq(&mchdev_lock);
7816	if (!i915_mch_dev) {
7817		ret = false;
7818		goto out_unlock;
7819	}
7820	dev_priv = i915_mch_dev;
7821
7822	dev_priv->ips.max_delay = dev_priv->ips.fstart;
7823
7824	if (!ironlake_set_drps(dev_priv, dev_priv->ips.fstart))
7825		ret = false;
7826
7827out_unlock:
7828	spin_unlock_irq(&mchdev_lock);
7829
7830	return ret;
7831}
7832EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
7833
7834/**
7835 * Tells the intel_ips driver that the i915 driver is now loaded, if
7836 * IPS got loaded first.
7837 *
7838 * This awkward dance is so that neither module has to depend on the
7839 * other in order for IPS to do the appropriate communication of
7840 * GPU turbo limits to i915.
7841 */
7842static void
7843ips_ping_for_i915_load(void)
7844{
7845	void (*link)(void);
7846
7847	link = symbol_get(ips_link_to_i915_driver);
7848	if (link) {
7849		link();
7850		symbol_put(ips_link_to_i915_driver);
7851	}
7852}
7853
7854void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
7855{
7856	/* We only register the i915 ips part with intel-ips once everything is
7857	 * set up, to avoid intel-ips sneaking in and reading bogus values. */
7858	spin_lock_irq(&mchdev_lock);
7859	i915_mch_dev = dev_priv;
7860	spin_unlock_irq(&mchdev_lock);
7861
7862	ips_ping_for_i915_load();
7863}
7864
7865void intel_gpu_ips_teardown(void)
7866{
7867	spin_lock_irq(&mchdev_lock);
7868	i915_mch_dev = NULL;
7869	spin_unlock_irq(&mchdev_lock);
7870}
7871
7872static void intel_init_emon(struct drm_i915_private *dev_priv)
7873{
7874	u32 lcfuse;
7875	u8 pxw[16];
7876	int i;
7877
7878	/* Disable to program */
7879	I915_WRITE(ECR, 0);
7880	POSTING_READ(ECR);
7881
7882	/* Program energy weights for various events */
7883	I915_WRITE(SDEW, 0x15040d00);
7884	I915_WRITE(CSIEW0, 0x007f0000);
7885	I915_WRITE(CSIEW1, 0x1e220004);
7886	I915_WRITE(CSIEW2, 0x04000004);
7887
7888	for (i = 0; i < 5; i++)
7889		I915_WRITE(PEW(i), 0);
7890	for (i = 0; i < 3; i++)
7891		I915_WRITE(DEW(i), 0);
7892
7893	/* Program P-state weights to account for frequency power adjustment */
7894	for (i = 0; i < 16; i++) {
7895		u32 pxvidfreq = I915_READ(PXVFREQ(i));
7896		unsigned long freq = intel_pxfreq(pxvidfreq);
7897		unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
7898			PXVFREQ_PX_SHIFT;
7899		unsigned long val;
7900
7901		val = vid * vid;
7902		val *= (freq / 1000);
7903		val *= 255;
7904		val /= (127*127*900);
7905		if (val > 0xff)
7906			DRM_ERROR("bad pxval: %ld\n", val);
7907		pxw[i] = val;
7908	}
7909	/* Render standby states get 0 weight */
7910	pxw[14] = 0;
7911	pxw[15] = 0;
7912
7913	for (i = 0; i < 4; i++) {
7914		u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
7915			(pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
7916		I915_WRITE(PXW(i), val);
7917	}
7918
7919	/* Adjust magic regs to magic values (more experimental results) */
7920	I915_WRITE(OGW0, 0);
7921	I915_WRITE(OGW1, 0);
7922	I915_WRITE(EG0, 0x00007f00);
7923	I915_WRITE(EG1, 0x0000000e);
7924	I915_WRITE(EG2, 0x000e0000);
7925	I915_WRITE(EG3, 0x68000300);
7926	I915_WRITE(EG4, 0x42000000);
7927	I915_WRITE(EG5, 0x00140031);
7928	I915_WRITE(EG6, 0);
7929	I915_WRITE(EG7, 0);
7930
7931	for (i = 0; i < 8; i++)
7932		I915_WRITE(PXWL(i), 0);
7933
7934	/* Enable PMON + select events */
7935	I915_WRITE(ECR, 0x80000019);
7936
7937	lcfuse = I915_READ(LCFUSE02);
7938
7939	dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
7940}
7941
7942void intel_init_gt_powersave(struct drm_i915_private *dev_priv)
7943{
7944	struct intel_rps *rps = &dev_priv->gt_pm.rps;
7945
7946	/*
7947	 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
7948	 * requirement.
7949	 */
7950	if (!sanitize_rc6(dev_priv)) {
7951		DRM_INFO("RC6 disabled, disabling runtime PM support\n");
7952		intel_runtime_pm_get(dev_priv);
7953	}
7954
7955	mutex_lock(&dev_priv->pcu_lock);
7956
7957	/* Initialize RPS limits (for userspace) */
7958	if (IS_CHERRYVIEW(dev_priv))
7959		cherryview_init_gt_powersave(dev_priv);
7960	else if (IS_VALLEYVIEW(dev_priv))
7961		valleyview_init_gt_powersave(dev_priv);
7962	else if (INTEL_GEN(dev_priv) >= 6)
7963		gen6_init_rps_frequencies(dev_priv);
7964
7965	/* Derive initial user preferences/limits from the hardware limits */
7966	rps->idle_freq = rps->min_freq;
7967	rps->cur_freq = rps->idle_freq;
7968
7969	rps->max_freq_softlimit = rps->max_freq;
7970	rps->min_freq_softlimit = rps->min_freq;
7971
7972	if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
7973		rps->min_freq_softlimit =
7974			max_t(int,
7975			      rps->efficient_freq,
7976			      intel_freq_opcode(dev_priv, 450));
7977
7978	/* After setting max-softlimit, find the overclock max freq */
7979	if (IS_GEN6(dev_priv) ||
7980	    IS_IVYBRIDGE(dev_priv) || IS_HASWELL(dev_priv)) {
7981		u32 params = 0;
7982
7983		sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &params);
7984		if (params & BIT(31)) { /* OC supported */
7985			DRM_DEBUG_DRIVER("Overclocking supported, max: %dMHz, overclock: %dMHz\n",
7986					 (rps->max_freq & 0xff) * 50,
7987					 (params & 0xff) * 50);
7988			rps->max_freq = params & 0xff;
7989		}
7990	}
7991
7992	/* Finally allow us to boost to max by default */
7993	rps->boost_freq = rps->max_freq;
7994
7995	mutex_unlock(&dev_priv->pcu_lock);
7996}
7997
7998void intel_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
7999{
8000	if (IS_VALLEYVIEW(dev_priv))
8001		valleyview_cleanup_gt_powersave(dev_priv);
8002
8003	if (!HAS_RC6(dev_priv))
8004		intel_runtime_pm_put(dev_priv);
8005}
8006
8007/**
8008 * intel_suspend_gt_powersave - suspend PM work and helper threads
8009 * @dev_priv: i915 device
8010 *
8011 * We don't want to disable RC6 or other features here, we just want
8012 * to make sure any work we've queued has finished and won't bother
8013 * us while we're suspended.
8014 */
8015void intel_suspend_gt_powersave(struct drm_i915_private *dev_priv)
8016{
8017	if (INTEL_GEN(dev_priv) < 6)
8018		return;
8019
8020	/* gen6_rps_idle() will be called later to disable interrupts */
8021}
8022
8023void intel_sanitize_gt_powersave(struct drm_i915_private *dev_priv)
8024{
8025	dev_priv->gt_pm.rps.enabled = true; /* force RPS disabling */
8026	dev_priv->gt_pm.rc6.enabled = true; /* force RC6 disabling */
8027	intel_disable_gt_powersave(dev_priv);
8028
8029	if (INTEL_GEN(dev_priv) < 11)
8030		gen6_reset_rps_interrupts(dev_priv);
8031	else
8032		WARN_ON_ONCE(1);
8033}
8034
8035static inline void intel_disable_llc_pstate(struct drm_i915_private *i915)
8036{
8037	lockdep_assert_held(&i915->pcu_lock);
8038
8039	if (!i915->gt_pm.llc_pstate.enabled)
8040		return;
8041
8042	/* Currently there is no HW configuration to be done to disable. */
8043
8044	i915->gt_pm.llc_pstate.enabled = false;
8045}
8046
8047static void intel_disable_rc6(struct drm_i915_private *dev_priv)
8048{
8049	lockdep_assert_held(&dev_priv->pcu_lock);
8050
8051	if (!dev_priv->gt_pm.rc6.enabled)
8052		return;
8053
8054	if (INTEL_GEN(dev_priv) >= 9)
8055		gen9_disable_rc6(dev_priv);
8056	else if (IS_CHERRYVIEW(dev_priv))
8057		cherryview_disable_rc6(dev_priv);
8058	else if (IS_VALLEYVIEW(dev_priv))
8059		valleyview_disable_rc6(dev_priv);
8060	else if (INTEL_GEN(dev_priv) >= 6)
8061		gen6_disable_rc6(dev_priv);
8062
8063	dev_priv->gt_pm.rc6.enabled = false;
8064}
8065
8066static void intel_disable_rps(struct drm_i915_private *dev_priv)
8067{
8068	lockdep_assert_held(&dev_priv->pcu_lock);
8069
8070	if (!dev_priv->gt_pm.rps.enabled)
8071		return;
8072
8073	if (INTEL_GEN(dev_priv) >= 9)
8074		gen9_disable_rps(dev_priv);
8075	else if (IS_CHERRYVIEW(dev_priv))
8076		cherryview_disable_rps(dev_priv);
8077	else if (IS_VALLEYVIEW(dev_priv))
8078		valleyview_disable_rps(dev_priv);
8079	else if (INTEL_GEN(dev_priv) >= 6)
8080		gen6_disable_rps(dev_priv);
8081	else if (IS_IRONLAKE_M(dev_priv))
8082		ironlake_disable_drps(dev_priv);
8083
8084	dev_priv->gt_pm.rps.enabled = false;
8085}
8086
8087void intel_disable_gt_powersave(struct drm_i915_private *dev_priv)
8088{
8089	mutex_lock(&dev_priv->pcu_lock);
8090
8091	intel_disable_rc6(dev_priv);
8092	intel_disable_rps(dev_priv);
8093	if (HAS_LLC(dev_priv))
8094		intel_disable_llc_pstate(dev_priv);
8095
8096	mutex_unlock(&dev_priv->pcu_lock);
8097}
8098
8099static inline void intel_enable_llc_pstate(struct drm_i915_private *i915)
8100{
8101	lockdep_assert_held(&i915->pcu_lock);
8102
8103	if (i915->gt_pm.llc_pstate.enabled)
8104		return;
8105
8106	gen6_update_ring_freq(i915);
8107
8108	i915->gt_pm.llc_pstate.enabled = true;
8109}
8110
8111static void intel_enable_rc6(struct drm_i915_private *dev_priv)
8112{
8113	lockdep_assert_held(&dev_priv->pcu_lock);
8114
8115	if (dev_priv->gt_pm.rc6.enabled)
8116		return;
8117
8118	if (IS_CHERRYVIEW(dev_priv))
8119		cherryview_enable_rc6(dev_priv);
8120	else if (IS_VALLEYVIEW(dev_priv))
8121		valleyview_enable_rc6(dev_priv);
8122	else if (INTEL_GEN(dev_priv) >= 9)
8123		gen9_enable_rc6(dev_priv);
8124	else if (IS_BROADWELL(dev_priv))
8125		gen8_enable_rc6(dev_priv);
8126	else if (INTEL_GEN(dev_priv) >= 6)
8127		gen6_enable_rc6(dev_priv);
8128
8129	dev_priv->gt_pm.rc6.enabled = true;
8130}
8131
8132static void intel_enable_rps(struct drm_i915_private *dev_priv)
8133{
8134	struct intel_rps *rps = &dev_priv->gt_pm.rps;
8135
8136	lockdep_assert_held(&dev_priv->pcu_lock);
8137
8138	if (rps->enabled)
8139		return;
8140
8141	if (IS_CHERRYVIEW(dev_priv)) {
8142		cherryview_enable_rps(dev_priv);
8143	} else if (IS_VALLEYVIEW(dev_priv)) {
8144		valleyview_enable_rps(dev_priv);
8145	} else if (WARN_ON_ONCE(INTEL_GEN(dev_priv) >= 11)) {
8146		/* TODO */
8147	} else if (INTEL_GEN(dev_priv) >= 9) {
8148		gen9_enable_rps(dev_priv);
8149	} else if (IS_BROADWELL(dev_priv)) {
8150		gen8_enable_rps(dev_priv);
8151	} else if (INTEL_GEN(dev_priv) >= 6) {
8152		gen6_enable_rps(dev_priv);
8153	} else if (IS_IRONLAKE_M(dev_priv)) {
8154		ironlake_enable_drps(dev_priv);
8155		intel_init_emon(dev_priv);
8156	}
8157
8158	WARN_ON(rps->max_freq < rps->min_freq);
8159	WARN_ON(rps->idle_freq > rps->max_freq);
8160
8161	WARN_ON(rps->efficient_freq < rps->min_freq);
8162	WARN_ON(rps->efficient_freq > rps->max_freq);
8163
8164	rps->enabled = true;
8165}
8166
8167void intel_enable_gt_powersave(struct drm_i915_private *dev_priv)
8168{
8169	/* Powersaving is controlled by the host when inside a VM */
8170	if (intel_vgpu_active(dev_priv))
8171		return;
8172
8173	mutex_lock(&dev_priv->pcu_lock);
8174
8175	if (HAS_RC6(dev_priv))
8176		intel_enable_rc6(dev_priv);
8177	intel_enable_rps(dev_priv);
8178	if (HAS_LLC(dev_priv))
8179		intel_enable_llc_pstate(dev_priv);
8180
8181	mutex_unlock(&dev_priv->pcu_lock);
8182}
8183
8184static void ibx_init_clock_gating(struct drm_i915_private *dev_priv)
8185{
8186	/*
8187	 * On Ibex Peak and Cougar Point, we need to disable clock
8188	 * gating for the panel power sequencer or it will fail to
8189	 * start up when no ports are active.
8190	 */
8191	I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
8192}
8193
8194static void g4x_disable_trickle_feed(struct drm_i915_private *dev_priv)
8195{
8196	enum pipe pipe;
8197
8198	for_each_pipe(dev_priv, pipe) {
8199		I915_WRITE(DSPCNTR(pipe),
8200			   I915_READ(DSPCNTR(pipe)) |
8201			   DISPPLANE_TRICKLE_FEED_DISABLE);
8202
8203		I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
8204		POSTING_READ(DSPSURF(pipe));
8205	}
8206}
8207
8208static void ilk_init_clock_gating(struct drm_i915_private *dev_priv)
8209{
8210	uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
8211
8212	/*
8213	 * Required for FBC
8214	 * WaFbcDisableDpfcClockGating:ilk
8215	 */
8216	dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
8217		   ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
8218		   ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
8219
8220	I915_WRITE(PCH_3DCGDIS0,
8221		   MARIUNIT_CLOCK_GATE_DISABLE |
8222		   SVSMUNIT_CLOCK_GATE_DISABLE);
8223	I915_WRITE(PCH_3DCGDIS1,
8224		   VFMUNIT_CLOCK_GATE_DISABLE);
8225
8226	/*
8227	 * According to the spec the following bits should be set in
8228	 * order to enable memory self-refresh
8229	 * The bit 22/21 of 0x42004
8230	 * The bit 5 of 0x42020
8231	 * The bit 15 of 0x45000
8232	 */
8233	I915_WRITE(ILK_DISPLAY_CHICKEN2,
8234		   (I915_READ(ILK_DISPLAY_CHICKEN2) |
8235		    ILK_DPARB_GATE | ILK_VSDPFD_FULL));
8236	dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
8237	I915_WRITE(DISP_ARB_CTL,
8238		   (I915_READ(DISP_ARB_CTL) |
8239		    DISP_FBC_WM_DIS));
8240
8241	/*
8242	 * Based on the document from hardware guys the following bits
8243	 * should be set unconditionally in order to enable FBC.
8244	 * The bit 22 of 0x42000
8245	 * The bit 22 of 0x42004
8246	 * The bit 7,8,9 of 0x42020.
8247	 */
8248	if (IS_IRONLAKE_M(dev_priv)) {
8249		/* WaFbcAsynchFlipDisableFbcQueue:ilk */
8250		I915_WRITE(ILK_DISPLAY_CHICKEN1,
8251			   I915_READ(ILK_DISPLAY_CHICKEN1) |
8252			   ILK_FBCQ_DIS);
8253		I915_WRITE(ILK_DISPLAY_CHICKEN2,
8254			   I915_READ(ILK_DISPLAY_CHICKEN2) |
8255			   ILK_DPARB_GATE);
8256	}
8257
8258	I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
8259
8260	I915_WRITE(ILK_DISPLAY_CHICKEN2,
8261		   I915_READ(ILK_DISPLAY_CHICKEN2) |
8262		   ILK_ELPIN_409_SELECT);
8263	I915_WRITE(_3D_CHICKEN2,
8264		   _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
8265		   _3D_CHICKEN2_WM_READ_PIPELINED);
8266
8267	/* WaDisableRenderCachePipelinedFlush:ilk */
8268	I915_WRITE(CACHE_MODE_0,
8269		   _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
8270
8271	/* WaDisable_RenderCache_OperationalFlush:ilk */
8272	I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8273
8274	g4x_disable_trickle_feed(dev_priv);
8275
8276	ibx_init_clock_gating(dev_priv);
8277}
8278
8279static void cpt_init_clock_gating(struct drm_i915_private *dev_priv)
8280{
8281	int pipe;
8282	uint32_t val;
8283
8284	/*
8285	 * On Ibex Peak and Cougar Point, we need to disable clock
8286	 * gating for the panel power sequencer or it will fail to
8287	 * start up when no ports are active.
8288	 */
8289	I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
8290		   PCH_DPLUNIT_CLOCK_GATE_DISABLE |
8291		   PCH_CPUNIT_CLOCK_GATE_DISABLE);
8292	I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
8293		   DPLS_EDP_PPS_FIX_DIS);
8294	/* The below fixes the weird display corruption, a few pixels shifted
8295	 * downward, on (only) LVDS of some HP laptops with IVY.
8296	 */
8297	for_each_pipe(dev_priv, pipe) {
8298		val = I915_READ(TRANS_CHICKEN2(pipe));
8299		val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
8300		val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
8301		if (dev_priv->vbt.fdi_rx_polarity_inverted)
8302			val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
8303		val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
8304		val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
8305		val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
8306		I915_WRITE(TRANS_CHICKEN2(pipe), val);
8307	}
8308	/* WADP0ClockGatingDisable */
8309	for_each_pipe(dev_priv, pipe) {
8310		I915_WRITE(TRANS_CHICKEN1(pipe),
8311			   TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
8312	}
8313}
8314
8315static void gen6_check_mch_setup(struct drm_i915_private *dev_priv)
8316{
8317	uint32_t tmp;
8318
8319	tmp = I915_READ(MCH_SSKPD);
8320	if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
8321		DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
8322			      tmp);
8323}
8324
8325static void gen6_init_clock_gating(struct drm_i915_private *dev_priv)
8326{
8327	uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
8328
8329	I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
8330
8331	I915_WRITE(ILK_DISPLAY_CHICKEN2,
8332		   I915_READ(ILK_DISPLAY_CHICKEN2) |
8333		   ILK_ELPIN_409_SELECT);
8334
8335	/* WaDisableHiZPlanesWhenMSAAEnabled:snb */
8336	I915_WRITE(_3D_CHICKEN,
8337		   _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
8338
8339	/* WaDisable_RenderCache_OperationalFlush:snb */
8340	I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8341
8342	/*
8343	 * BSpec recoomends 8x4 when MSAA is used,
8344	 * however in practice 16x4 seems fastest.
8345	 *
8346	 * Note that PS/WM thread counts depend on the WIZ hashing
8347	 * disable bit, which we don't touch here, but it's good
8348	 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8349	 */
8350	I915_WRITE(GEN6_GT_MODE,
8351		   _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
8352
8353	I915_WRITE(CACHE_MODE_0,
8354		   _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
8355
8356	I915_WRITE(GEN6_UCGCTL1,
8357		   I915_READ(GEN6_UCGCTL1) |
8358		   GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
8359		   GEN6_CSUNIT_CLOCK_GATE_DISABLE);
8360
8361	/* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
8362	 * gating disable must be set.  Failure to set it results in
8363	 * flickering pixels due to Z write ordering failures after
8364	 * some amount of runtime in the Mesa "fire" demo, and Unigine
8365	 * Sanctuary and Tropics, and apparently anything else with
8366	 * alpha test or pixel discard.
8367	 *
8368	 * According to the spec, bit 11 (RCCUNIT) must also be set,
8369	 * but we didn't debug actual testcases to find it out.
8370	 *
8371	 * WaDisableRCCUnitClockGating:snb
8372	 * WaDisableRCPBUnitClockGating:snb
8373	 */
8374	I915_WRITE(GEN6_UCGCTL2,
8375		   GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
8376		   GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
8377
8378	/* WaStripsFansDisableFastClipPerformanceFix:snb */
8379	I915_WRITE(_3D_CHICKEN3,
8380		   _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
8381
8382	/*
8383	 * Bspec says:
8384	 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
8385	 * 3DSTATE_SF number of SF output attributes is more than 16."
8386	 */
8387	I915_WRITE(_3D_CHICKEN3,
8388		   _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
8389
8390	/*
8391	 * According to the spec the following bits should be
8392	 * set in order to enable memory self-refresh and fbc:
8393	 * The bit21 and bit22 of 0x42000
8394	 * The bit21 and bit22 of 0x42004
8395	 * The bit5 and bit7 of 0x42020
8396	 * The bit14 of 0x70180
8397	 * The bit14 of 0x71180
8398	 *
8399	 * WaFbcAsynchFlipDisableFbcQueue:snb
8400	 */
8401	I915_WRITE(ILK_DISPLAY_CHICKEN1,
8402		   I915_READ(ILK_DISPLAY_CHICKEN1) |
8403		   ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
8404	I915_WRITE(ILK_DISPLAY_CHICKEN2,
8405		   I915_READ(ILK_DISPLAY_CHICKEN2) |
8406		   ILK_DPARB_GATE | ILK_VSDPFD_FULL);
8407	I915_WRITE(ILK_DSPCLK_GATE_D,
8408		   I915_READ(ILK_DSPCLK_GATE_D) |
8409		   ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
8410		   ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
8411
8412	g4x_disable_trickle_feed(dev_priv);
8413
8414	cpt_init_clock_gating(dev_priv);
8415
8416	gen6_check_mch_setup(dev_priv);
8417}
8418
8419static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
8420{
8421	uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
8422
8423	/*
8424	 * WaVSThreadDispatchOverride:ivb,vlv
8425	 *
8426	 * This actually overrides the dispatch
8427	 * mode for all thread types.
8428	 */
8429	reg &= ~GEN7_FF_SCHED_MASK;
8430	reg |= GEN7_FF_TS_SCHED_HW;
8431	reg |= GEN7_FF_VS_SCHED_HW;
8432	reg |= GEN7_FF_DS_SCHED_HW;
8433
8434	I915_WRITE(GEN7_FF_THREAD_MODE, reg);
8435}
8436
8437static void lpt_init_clock_gating(struct drm_i915_private *dev_priv)
8438{
8439	/*
8440	 * TODO: this bit should only be enabled when really needed, then
8441	 * disabled when not needed anymore in order to save power.
8442	 */
8443	if (HAS_PCH_LPT_LP(dev_priv))
8444		I915_WRITE(SOUTH_DSPCLK_GATE_D,
8445			   I915_READ(SOUTH_DSPCLK_GATE_D) |
8446			   PCH_LP_PARTITION_LEVEL_DISABLE);
8447
8448	/* WADPOClockGatingDisable:hsw */
8449	I915_WRITE(TRANS_CHICKEN1(PIPE_A),
8450		   I915_READ(TRANS_CHICKEN1(PIPE_A)) |
8451		   TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
8452}
8453
8454static void lpt_suspend_hw(struct drm_i915_private *dev_priv)
8455{
8456	if (HAS_PCH_LPT_LP(dev_priv)) {
8457		uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
8458
8459		val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
8460		I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
8461	}
8462}
8463
8464static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
8465				   int general_prio_credits,
8466				   int high_prio_credits)
8467{
8468	u32 misccpctl;
8469	u32 val;
8470
8471	/* WaTempDisableDOPClkGating:bdw */
8472	misccpctl = I915_READ(GEN7_MISCCPCTL);
8473	I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
8474
8475	val = I915_READ(GEN8_L3SQCREG1);
8476	val &= ~L3_PRIO_CREDITS_MASK;
8477	val |= L3_GENERAL_PRIO_CREDITS(general_prio_credits);
8478	val |= L3_HIGH_PRIO_CREDITS(high_prio_credits);
8479	I915_WRITE(GEN8_L3SQCREG1, val);
8480
8481	/*
8482	 * Wait at least 100 clocks before re-enabling clock gating.
8483	 * See the definition of L3SQCREG1 in BSpec.
8484	 */
8485	POSTING_READ(GEN8_L3SQCREG1);
8486	udelay(1);
8487	I915_WRITE(GEN7_MISCCPCTL, misccpctl);
8488}
8489
8490static void cnp_init_clock_gating(struct drm_i915_private *dev_priv)
8491{
8492	if (!HAS_PCH_CNP(dev_priv))
8493		return;
8494
8495	/* Display WA #1181 WaSouthDisplayDisablePWMCGEGating: cnp */
8496	I915_WRITE(SOUTH_DSPCLK_GATE_D, I915_READ(SOUTH_DSPCLK_GATE_D) |
8497		   CNP_PWM_CGE_GATING_DISABLE);
8498}
8499
8500static void cnl_init_clock_gating(struct drm_i915_private *dev_priv)
8501{
8502	u32 val;
8503	cnp_init_clock_gating(dev_priv);
8504
8505	/* This is not an Wa. Enable for better image quality */
8506	I915_WRITE(_3D_CHICKEN3,
8507		   _MASKED_BIT_ENABLE(_3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE));
8508
8509	/* WaEnableChickenDCPR:cnl */
8510	I915_WRITE(GEN8_CHICKEN_DCPR_1,
8511		   I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
8512
8513	/* WaFbcWakeMemOn:cnl */
8514	I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
8515		   DISP_FBC_MEMORY_WAKE);
8516
8517	val = I915_READ(SLICE_UNIT_LEVEL_CLKGATE);
8518	/* ReadHitWriteOnlyDisable:cnl */
8519	val |= RCCUNIT_CLKGATE_DIS;
8520	/* WaSarbUnitClockGatingDisable:cnl (pre-prod) */
8521	if (IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_B0))
8522		val |= SARBUNIT_CLKGATE_DIS;
8523	I915_WRITE(SLICE_UNIT_LEVEL_CLKGATE, val);
8524
8525	/* Wa_2201832410:cnl */
8526	val = I915_READ(SUBSLICE_UNIT_LEVEL_CLKGATE);
8527	val |= GWUNIT_CLKGATE_DIS;
8528	I915_WRITE(SUBSLICE_UNIT_LEVEL_CLKGATE, val);
8529
8530	/* WaDisableVFclkgate:cnl */
8531	/* WaVFUnitClockGatingDisable:cnl */
8532	val = I915_READ(UNSLICE_UNIT_LEVEL_CLKGATE);
8533	val |= VFUNIT_CLKGATE_DIS;
8534	I915_WRITE(UNSLICE_UNIT_LEVEL_CLKGATE, val);
8535}
8536
8537static void cfl_init_clock_gating(struct drm_i915_private *dev_priv)
8538{
8539	cnp_init_clock_gating(dev_priv);
8540	gen9_init_clock_gating(dev_priv);
8541
8542	/* WaFbcNukeOnHostModify:cfl */
8543	I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
8544		   ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
8545}
8546
8547static void kbl_init_clock_gating(struct drm_i915_private *dev_priv)
8548{
8549	gen9_init_clock_gating(dev_priv);
8550
8551	/* WaDisableSDEUnitClockGating:kbl */
8552	if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
8553		I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
8554			   GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
8555
8556	/* WaDisableGamClockGating:kbl */
8557	if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
8558		I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
8559			   GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
8560
8561	/* WaFbcNukeOnHostModify:kbl */
8562	I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
8563		   ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
8564}
8565
8566static void skl_init_clock_gating(struct drm_i915_private *dev_priv)
8567{
8568	gen9_init_clock_gating(dev_priv);
8569
8570	/* WAC6entrylatency:skl */
8571	I915_WRITE(FBC_LLC_READ_CTRL, I915_READ(FBC_LLC_READ_CTRL) |
8572		   FBC_LLC_FULLY_OPEN);
8573
8574	/* WaFbcNukeOnHostModify:skl */
8575	I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
8576		   ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
8577}
8578
8579static void bdw_init_clock_gating(struct drm_i915_private *dev_priv)
8580{
8581	/* The GTT cache must be disabled if the system is using 2M pages. */
8582	bool can_use_gtt_cache = !HAS_PAGE_SIZES(dev_priv,
8583						 I915_GTT_PAGE_SIZE_2M);
8584	enum pipe pipe;
8585
8586	/* WaSwitchSolVfFArbitrationPriority:bdw */
8587	I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
8588
8589	/* WaPsrDPAMaskVBlankInSRD:bdw */
8590	I915_WRITE(CHICKEN_PAR1_1,
8591		   I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
8592
8593	/* WaPsrDPRSUnmaskVBlankInSRD:bdw */
8594	for_each_pipe(dev_priv, pipe) {
8595		I915_WRITE(CHICKEN_PIPESL_1(pipe),
8596			   I915_READ(CHICKEN_PIPESL_1(pipe)) |
8597			   BDW_DPRS_MASK_VBLANK_SRD);
8598	}
8599
8600	/* WaVSRefCountFullforceMissDisable:bdw */
8601	/* WaDSRefCountFullforceMissDisable:bdw */
8602	I915_WRITE(GEN7_FF_THREAD_MODE,
8603		   I915_READ(GEN7_FF_THREAD_MODE) &
8604		   ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
8605
8606	I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
8607		   _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
8608
8609	/* WaDisableSDEUnitClockGating:bdw */
8610	I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
8611		   GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
8612
8613	/* WaProgramL3SqcReg1Default:bdw */
8614	gen8_set_l3sqc_credits(dev_priv, 30, 2);
8615
8616	/* WaGttCachingOffByDefault:bdw */
8617	I915_WRITE(HSW_GTT_CACHE_EN, can_use_gtt_cache ? GTT_CACHE_EN_ALL : 0);
8618
8619	/* WaKVMNotificationOnConfigChange:bdw */
8620	I915_WRITE(CHICKEN_PAR2_1, I915_READ(CHICKEN_PAR2_1)
8621		   | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
8622
8623	lpt_init_clock_gating(dev_priv);
8624
8625	/* WaDisableDopClockGating:bdw
8626	 *
8627	 * Also see the CHICKEN2 write in bdw_init_workarounds() to disable DOP
8628	 * clock gating.
8629	 */
8630	I915_WRITE(GEN6_UCGCTL1,
8631		   I915_READ(GEN6_UCGCTL1) | GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE);
8632}
8633
8634static void hsw_init_clock_gating(struct drm_i915_private *dev_priv)
8635{
8636	/* L3 caching of data atomics doesn't work -- disable it. */
8637	I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
8638	I915_WRITE(HSW_ROW_CHICKEN3,
8639		   _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
8640
8641	/* This is required by WaCatErrorRejectionIssue:hsw */
8642	I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
8643			I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
8644			GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
8645
8646	/* WaVSRefCountFullforceMissDisable:hsw */
8647	I915_WRITE(GEN7_FF_THREAD_MODE,
8648		   I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
8649
8650	/* WaDisable_RenderCache_OperationalFlush:hsw */
8651	I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8652
8653	/* enable HiZ Raw Stall Optimization */
8654	I915_WRITE(CACHE_MODE_0_GEN7,
8655		   _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
8656
8657	/* WaDisable4x2SubspanOptimization:hsw */
8658	I915_WRITE(CACHE_MODE_1,
8659		   _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
8660
8661	/*
8662	 * BSpec recommends 8x4 when MSAA is used,
8663	 * however in practice 16x4 seems fastest.
8664	 *
8665	 * Note that PS/WM thread counts depend on the WIZ hashing
8666	 * disable bit, which we don't touch here, but it's good
8667	 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8668	 */
8669	I915_WRITE(GEN7_GT_MODE,
8670		   _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
8671
8672	/* WaSampleCChickenBitEnable:hsw */
8673	I915_WRITE(HALF_SLICE_CHICKEN3,
8674		   _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
8675
8676	/* WaSwitchSolVfFArbitrationPriority:hsw */
8677	I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
8678
8679	lpt_init_clock_gating(dev_priv);
8680}
8681
8682static void ivb_init_clock_gating(struct drm_i915_private *dev_priv)
8683{
8684	uint32_t snpcr;
8685
8686	I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
8687
8688	/* WaDisableEarlyCull:ivb */
8689	I915_WRITE(_3D_CHICKEN3,
8690		   _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
8691
8692	/* WaDisableBackToBackFlipFix:ivb */
8693	I915_WRITE(IVB_CHICKEN3,
8694		   CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
8695		   CHICKEN3_DGMG_DONE_FIX_DISABLE);
8696
8697	/* WaDisablePSDDualDispatchEnable:ivb */
8698	if (IS_IVB_GT1(dev_priv))
8699		I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
8700			   _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
8701
8702	/* WaDisable_RenderCache_OperationalFlush:ivb */
8703	I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8704
8705	/* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
8706	I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
8707		   GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
8708
8709	/* WaApplyL3ControlAndL3ChickenMode:ivb */
8710	I915_WRITE(GEN7_L3CNTLREG1,
8711			GEN7_WA_FOR_GEN7_L3_CONTROL);
8712	I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
8713		   GEN7_WA_L3_CHICKEN_MODE);
8714	if (IS_IVB_GT1(dev_priv))
8715		I915_WRITE(GEN7_ROW_CHICKEN2,
8716			   _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
8717	else {
8718		/* must write both registers */
8719		I915_WRITE(GEN7_ROW_CHICKEN2,
8720			   _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
8721		I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
8722			   _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
8723	}
8724
8725	/* WaForceL3Serialization:ivb */
8726	I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
8727		   ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
8728
8729	/*
8730	 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
8731	 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
8732	 */
8733	I915_WRITE(GEN6_UCGCTL2,
8734		   GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
8735
8736	/* This is required by WaCatErrorRejectionIssue:ivb */
8737	I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
8738			I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
8739			GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
8740
8741	g4x_disable_trickle_feed(dev_priv);
8742
8743	gen7_setup_fixed_func_scheduler(dev_priv);
8744
8745	if (0) { /* causes HiZ corruption on ivb:gt1 */
8746		/* enable HiZ Raw Stall Optimization */
8747		I915_WRITE(CACHE_MODE_0_GEN7,
8748			   _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
8749	}
8750
8751	/* WaDisable4x2SubspanOptimization:ivb */
8752	I915_WRITE(CACHE_MODE_1,
8753		   _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
8754
8755	/*
8756	 * BSpec recommends 8x4 when MSAA is used,
8757	 * however in practice 16x4 seems fastest.
8758	 *
8759	 * Note that PS/WM thread counts depend on the WIZ hashing
8760	 * disable bit, which we don't touch here, but it's good
8761	 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8762	 */
8763	I915_WRITE(GEN7_GT_MODE,
8764		   _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
8765
8766	snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
8767	snpcr &= ~GEN6_MBC_SNPCR_MASK;
8768	snpcr |= GEN6_MBC_SNPCR_MED;
8769	I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
8770
8771	if (!HAS_PCH_NOP(dev_priv))
8772		cpt_init_clock_gating(dev_priv);
8773
8774	gen6_check_mch_setup(dev_priv);
8775}
8776
8777static void vlv_init_clock_gating(struct drm_i915_private *dev_priv)
8778{
8779	/* WaDisableEarlyCull:vlv */
8780	I915_WRITE(_3D_CHICKEN3,
8781		   _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
8782
8783	/* WaDisableBackToBackFlipFix:vlv */
8784	I915_WRITE(IVB_CHICKEN3,
8785		   CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
8786		   CHICKEN3_DGMG_DONE_FIX_DISABLE);
8787
8788	/* WaPsdDispatchEnable:vlv */
8789	/* WaDisablePSDDualDispatchEnable:vlv */
8790	I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
8791		   _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
8792				      GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
8793
8794	/* WaDisable_RenderCache_OperationalFlush:vlv */
8795	I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8796
8797	/* WaForceL3Serialization:vlv */
8798	I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
8799		   ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
8800
8801	/* WaDisableDopClockGating:vlv */
8802	I915_WRITE(GEN7_ROW_CHICKEN2,
8803		   _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
8804
8805	/* This is required by WaCatErrorRejectionIssue:vlv */
8806	I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
8807		   I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
8808		   GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
8809
8810	gen7_setup_fixed_func_scheduler(dev_priv);
8811
8812	/*
8813	 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
8814	 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
8815	 */
8816	I915_WRITE(GEN6_UCGCTL2,
8817		   GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
8818
8819	/* WaDisableL3Bank2xClockGate:vlv
8820	 * Disabling L3 clock gating- MMIO 940c[25] = 1
8821	 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
8822	I915_WRITE(GEN7_UCGCTL4,
8823		   I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
8824
8825	/*
8826	 * BSpec says this must be set, even though
8827	 * WaDisable4x2SubspanOptimization isn't listed for VLV.
8828	 */
8829	I915_WRITE(CACHE_MODE_1,
8830		   _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
8831
8832	/*
8833	 * BSpec recommends 8x4 when MSAA is used,
8834	 * however in practice 16x4 seems fastest.
8835	 *
8836	 * Note that PS/WM thread counts depend on the WIZ hashing
8837	 * disable bit, which we don't touch here, but it's good
8838	 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8839	 */
8840	I915_WRITE(GEN7_GT_MODE,
8841		   _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
8842
8843	/*
8844	 * WaIncreaseL3CreditsForVLVB0:vlv
8845	 * This is the hardware default actually.
8846	 */
8847	I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
8848
8849	/*
8850	 * WaDisableVLVClockGating_VBIIssue:vlv
8851	 * Disable clock gating on th GCFG unit to prevent a delay
8852	 * in the reporting of vblank events.
8853	 */
8854	I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
8855}
8856
8857static void chv_init_clock_gating(struct drm_i915_private *dev_priv)
8858{
8859	/* WaVSRefCountFullforceMissDisable:chv */
8860	/* WaDSRefCountFullforceMissDisable:chv */
8861	I915_WRITE(GEN7_FF_THREAD_MODE,
8862		   I915_READ(GEN7_FF_THREAD_MODE) &
8863		   ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
8864
8865	/* WaDisableSemaphoreAndSyncFlipWait:chv */
8866	I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
8867		   _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
8868
8869	/* WaDisableCSUnitClockGating:chv */
8870	I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
8871		   GEN6_CSUNIT_CLOCK_GATE_DISABLE);
8872
8873	/* WaDisableSDEUnitClockGating:chv */
8874	I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
8875		   GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
8876
8877	/*
8878	 * WaProgramL3SqcReg1Default:chv
8879	 * See gfxspecs/Related Documents/Performance Guide/
8880	 * LSQC Setting Recommendations.
8881	 */
8882	gen8_set_l3sqc_credits(dev_priv, 38, 2);
8883
8884	/*
8885	 * GTT cache may not work with big pages, so if those
8886	 * are ever enabled GTT cache may need to be disabled.
8887	 */
8888	I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
8889}
8890
8891static void g4x_init_clock_gating(struct drm_i915_private *dev_priv)
8892{
8893	uint32_t dspclk_gate;
8894
8895	I915_WRITE(RENCLK_GATE_D1, 0);
8896	I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
8897		   GS_UNIT_CLOCK_GATE_DISABLE |
8898		   CL_UNIT_CLOCK_GATE_DISABLE);
8899	I915_WRITE(RAMCLK_GATE_D, 0);
8900	dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
8901		OVRUNIT_CLOCK_GATE_DISABLE |
8902		OVCUNIT_CLOCK_GATE_DISABLE;
8903	if (IS_GM45(dev_priv))
8904		dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
8905	I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
8906
8907	/* WaDisableRenderCachePipelinedFlush */
8908	I915_WRITE(CACHE_MODE_0,
8909		   _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
8910
8911	/* WaDisable_RenderCache_OperationalFlush:g4x */
8912	I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8913
8914	g4x_disable_trickle_feed(dev_priv);
8915}
8916
8917static void i965gm_init_clock_gating(struct drm_i915_private *dev_priv)
8918{
8919	I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
8920	I915_WRITE(RENCLK_GATE_D2, 0);
8921	I915_WRITE(DSPCLK_GATE_D, 0);
8922	I915_WRITE(RAMCLK_GATE_D, 0);
8923	I915_WRITE16(DEUC, 0);
8924	I915_WRITE(MI_ARB_STATE,
8925		   _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
8926
8927	/* WaDisable_RenderCache_OperationalFlush:gen4 */
8928	I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8929}
8930
8931static void i965g_init_clock_gating(struct drm_i915_private *dev_priv)
8932{
8933	I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
8934		   I965_RCC_CLOCK_GATE_DISABLE |
8935		   I965_RCPB_CLOCK_GATE_DISABLE |
8936		   I965_ISC_CLOCK_GATE_DISABLE |
8937		   I965_FBC_CLOCK_GATE_DISABLE);
8938	I915_WRITE(RENCLK_GATE_D2, 0);
8939	I915_WRITE(MI_ARB_STATE,
8940		   _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
8941
8942	/* WaDisable_RenderCache_OperationalFlush:gen4 */
8943	I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8944}
8945
8946static void gen3_init_clock_gating(struct drm_i915_private *dev_priv)
8947{
8948	u32 dstate = I915_READ(D_STATE);
8949
8950	dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
8951		DSTATE_DOT_CLOCK_GATING;
8952	I915_WRITE(D_STATE, dstate);
8953
8954	if (IS_PINEVIEW(dev_priv))
8955		I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
8956
8957	/* IIR "flip pending" means done if this bit is set */
8958	I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
8959
8960	/* interrupts should cause a wake up from C3 */
8961	I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
8962
8963	/* On GEN3 we really need to make sure the ARB C3 LP bit is set */
8964	I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
8965
8966	I915_WRITE(MI_ARB_STATE,
8967		   _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
8968}
8969
8970static void i85x_init_clock_gating(struct drm_i915_private *dev_priv)
8971{
8972	I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
8973
8974	/* interrupts should cause a wake up from C3 */
8975	I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
8976		   _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
8977
8978	I915_WRITE(MEM_MODE,
8979		   _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
8980}
8981
8982static void i830_init_clock_gating(struct drm_i915_private *dev_priv)
8983{
8984	I915_WRITE(MEM_MODE,
8985		   _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
8986		   _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
8987}
8988
8989void intel_init_clock_gating(struct drm_i915_private *dev_priv)
8990{
8991	dev_priv->display.init_clock_gating(dev_priv);
8992}
8993
8994void intel_suspend_hw(struct drm_i915_private *dev_priv)
8995{
8996	if (HAS_PCH_LPT(dev_priv))
8997		lpt_suspend_hw(dev_priv);
8998}
8999
9000static void nop_init_clock_gating(struct drm_i915_private *dev_priv)
9001{
9002	DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
9003}
9004
9005/**
9006 * intel_init_clock_gating_hooks - setup the clock gating hooks
9007 * @dev_priv: device private
9008 *
9009 * Setup the hooks that configure which clocks of a given platform can be
9010 * gated and also apply various GT and display specific workarounds for these
9011 * platforms. Note that some GT specific workarounds are applied separately
9012 * when GPU contexts or batchbuffers start their execution.
9013 */
9014void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
9015{
9016	if (IS_CANNONLAKE(dev_priv))
9017		dev_priv->display.init_clock_gating = cnl_init_clock_gating;
9018	else if (IS_COFFEELAKE(dev_priv))
9019		dev_priv->display.init_clock_gating = cfl_init_clock_gating;
9020	else if (IS_SKYLAKE(dev_priv))
9021		dev_priv->display.init_clock_gating = skl_init_clock_gating;
9022	else if (IS_KABYLAKE(dev_priv))
9023		dev_priv->display.init_clock_gating = kbl_init_clock_gating;
9024	else if (IS_BROXTON(dev_priv))
9025		dev_priv->display.init_clock_gating = bxt_init_clock_gating;
9026	else if (IS_GEMINILAKE(dev_priv))
9027		dev_priv->display.init_clock_gating = glk_init_clock_gating;
9028	else if (IS_BROADWELL(dev_priv))
9029		dev_priv->display.init_clock_gating = bdw_init_clock_gating;
9030	else if (IS_CHERRYVIEW(dev_priv))
9031		dev_priv->display.init_clock_gating = chv_init_clock_gating;
9032	else if (IS_HASWELL(dev_priv))
9033		dev_priv->display.init_clock_gating = hsw_init_clock_gating;
9034	else if (IS_IVYBRIDGE(dev_priv))
9035		dev_priv->display.init_clock_gating = ivb_init_clock_gating;
9036	else if (IS_VALLEYVIEW(dev_priv))
9037		dev_priv->display.init_clock_gating = vlv_init_clock_gating;
9038	else if (IS_GEN6(dev_priv))
9039		dev_priv->display.init_clock_gating = gen6_init_clock_gating;
9040	else if (IS_GEN5(dev_priv))
9041		dev_priv->display.init_clock_gating = ilk_init_clock_gating;
9042	else if (IS_G4X(dev_priv))
9043		dev_priv->display.init_clock_gating = g4x_init_clock_gating;
9044	else if (IS_I965GM(dev_priv))
9045		dev_priv->display.init_clock_gating = i965gm_init_clock_gating;
9046	else if (IS_I965G(dev_priv))
9047		dev_priv->display.init_clock_gating = i965g_init_clock_gating;
9048	else if (IS_GEN3(dev_priv))
9049		dev_priv->display.init_clock_gating = gen3_init_clock_gating;
9050	else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
9051		dev_priv->display.init_clock_gating = i85x_init_clock_gating;
9052	else if (IS_GEN2(dev_priv))
9053		dev_priv->display.init_clock_gating = i830_init_clock_gating;
9054	else {
9055		MISSING_CASE(INTEL_DEVID(dev_priv));
9056		dev_priv->display.init_clock_gating = nop_init_clock_gating;
9057	}
9058}
9059
9060/* Set up chip specific power management-related functions */
9061void intel_init_pm(struct drm_i915_private *dev_priv)
9062{
9063	intel_fbc_init(dev_priv);
9064
9065	/* For cxsr */
9066	if (IS_PINEVIEW(dev_priv))
9067		i915_pineview_get_mem_freq(dev_priv);
9068	else if (IS_GEN5(dev_priv))
9069		i915_ironlake_get_mem_freq(dev_priv);
9070
9071	/* For FIFO watermark updates */
9072	if (INTEL_GEN(dev_priv) >= 9) {
9073		skl_setup_wm_latency(dev_priv);
9074		dev_priv->display.initial_watermarks = skl_initial_wm;
9075		dev_priv->display.atomic_update_watermarks = skl_atomic_update_crtc_wm;
9076		dev_priv->display.compute_global_watermarks = skl_compute_wm;
9077	} else if (HAS_PCH_SPLIT(dev_priv)) {
9078		ilk_setup_wm_latency(dev_priv);
9079
9080		if ((IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[1] &&
9081		     dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
9082		    (!IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[0] &&
9083		     dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
9084			dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
9085			dev_priv->display.compute_intermediate_wm =
9086				ilk_compute_intermediate_wm;
9087			dev_priv->display.initial_watermarks =
9088				ilk_initial_watermarks;
9089			dev_priv->display.optimize_watermarks =
9090				ilk_optimize_watermarks;
9091		} else {
9092			DRM_DEBUG_KMS("Failed to read display plane latency. "
9093				      "Disable CxSR\n");
9094		}
9095	} else if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
9096		vlv_setup_wm_latency(dev_priv);
9097		dev_priv->display.compute_pipe_wm = vlv_compute_pipe_wm;
9098		dev_priv->display.compute_intermediate_wm = vlv_compute_intermediate_wm;
9099		dev_priv->display.initial_watermarks = vlv_initial_watermarks;
9100		dev_priv->display.optimize_watermarks = vlv_optimize_watermarks;
9101		dev_priv->display.atomic_update_watermarks = vlv_atomic_update_fifo;
9102	} else if (IS_G4X(dev_priv)) {
9103		g4x_setup_wm_latency(dev_priv);
9104		dev_priv->display.compute_pipe_wm = g4x_compute_pipe_wm;
9105		dev_priv->display.compute_intermediate_wm = g4x_compute_intermediate_wm;
9106		dev_priv->display.initial_watermarks = g4x_initial_watermarks;
9107		dev_priv->display.optimize_watermarks = g4x_optimize_watermarks;
9108	} else if (IS_PINEVIEW(dev_priv)) {
9109		if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
9110					    dev_priv->is_ddr3,
9111					    dev_priv->fsb_freq,
9112					    dev_priv->mem_freq)) {
9113			DRM_INFO("failed to find known CxSR latency "
9114				 "(found ddr%s fsb freq %d, mem freq %d), "
9115				 "disabling CxSR\n",
9116				 (dev_priv->is_ddr3 == 1) ? "3" : "2",
9117				 dev_priv->fsb_freq, dev_priv->mem_freq);
9118			/* Disable CxSR and never update its watermark again */
9119			intel_set_memory_cxsr(dev_priv, false);
9120			dev_priv->display.update_wm = NULL;
9121		} else
9122			dev_priv->display.update_wm = pineview_update_wm;
9123	} else if (IS_GEN4(dev_priv)) {
9124		dev_priv->display.update_wm = i965_update_wm;
9125	} else if (IS_GEN3(dev_priv)) {
9126		dev_priv->display.update_wm = i9xx_update_wm;
9127		dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
9128	} else if (IS_GEN2(dev_priv)) {
9129		if (INTEL_INFO(dev_priv)->num_pipes == 1) {
9130			dev_priv->display.update_wm = i845_update_wm;
9131			dev_priv->display.get_fifo_size = i845_get_fifo_size;
9132		} else {
9133			dev_priv->display.update_wm = i9xx_update_wm;
9134			dev_priv->display.get_fifo_size = i830_get_fifo_size;
9135		}
9136	} else {
9137		DRM_ERROR("unexpected fall-through in intel_init_pm\n");
9138	}
9139}
9140
9141static inline int gen6_check_mailbox_status(struct drm_i915_private *dev_priv)
9142{
9143	uint32_t flags =
9144		I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
9145
9146	switch (flags) {
9147	case GEN6_PCODE_SUCCESS:
9148		return 0;
9149	case GEN6_PCODE_UNIMPLEMENTED_CMD:
9150		return -ENODEV;
9151	case GEN6_PCODE_ILLEGAL_CMD:
9152		return -ENXIO;
9153	case GEN6_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
9154	case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
9155		return -EOVERFLOW;
9156	case GEN6_PCODE_TIMEOUT:
9157		return -ETIMEDOUT;
9158	default:
9159		MISSING_CASE(flags);
9160		return 0;
9161	}
9162}
9163
9164static inline int gen7_check_mailbox_status(struct drm_i915_private *dev_priv)
9165{
9166	uint32_t flags =
9167		I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
9168
9169	switch (flags) {
9170	case GEN6_PCODE_SUCCESS:
9171		return 0;
9172	case GEN6_PCODE_ILLEGAL_CMD:
9173		return -ENXIO;
9174	case GEN7_PCODE_TIMEOUT:
9175		return -ETIMEDOUT;
9176	case GEN7_PCODE_ILLEGAL_DATA:
9177		return -EINVAL;
9178	case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
9179		return -EOVERFLOW;
9180	default:
9181		MISSING_CASE(flags);
9182		return 0;
9183	}
9184}
9185
9186int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
9187{
9188	int status;
9189
9190	WARN_ON(!mutex_is_locked(&dev_priv->pcu_lock));
9191
9192	/* GEN6_PCODE_* are outside of the forcewake domain, we can
9193	 * use te fw I915_READ variants to reduce the amount of work
9194	 * required when reading/writing.
9195	 */
9196
9197	if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
9198		DRM_DEBUG_DRIVER("warning: pcode (read from mbox %x) mailbox access failed for %ps\n",
9199				 mbox, __builtin_return_address(0));
9200		return -EAGAIN;
9201	}
9202
9203	I915_WRITE_FW(GEN6_PCODE_DATA, *val);
9204	I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
9205	I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
9206
9207	if (__intel_wait_for_register_fw(dev_priv,
9208					 GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
9209					 500, 0, NULL)) {
9210		DRM_ERROR("timeout waiting for pcode read (from mbox %x) to finish for %ps\n",
9211			  mbox, __builtin_return_address(0));
9212		return -ETIMEDOUT;
9213	}
9214
9215	*val = I915_READ_FW(GEN6_PCODE_DATA);
9216	I915_WRITE_FW(GEN6_PCODE_DATA, 0);
9217
9218	if (INTEL_GEN(dev_priv) > 6)
9219		status = gen7_check_mailbox_status(dev_priv);
9220	else
9221		status = gen6_check_mailbox_status(dev_priv);
9222
9223	if (status) {
9224		DRM_DEBUG_DRIVER("warning: pcode (read from mbox %x) mailbox access failed for %ps: %d\n",
9225				 mbox, __builtin_return_address(0), status);
9226		return status;
9227	}
9228
9229	return 0;
9230}
9231
9232int sandybridge_pcode_write_timeout(struct drm_i915_private *dev_priv,
9233				    u32 mbox, u32 val,
9234				    int fast_timeout_us, int slow_timeout_ms)
9235{
9236	int status;
9237
9238	WARN_ON(!mutex_is_locked(&dev_priv->pcu_lock));
9239
9240	/* GEN6_PCODE_* are outside of the forcewake domain, we can
9241	 * use te fw I915_READ variants to reduce the amount of work
9242	 * required when reading/writing.
9243	 */
9244
9245	if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
9246		DRM_DEBUG_DRIVER("warning: pcode (write of 0x%08x to mbox %x) mailbox access failed for %ps\n",
9247				 val, mbox, __builtin_return_address(0));
9248		return -EAGAIN;
9249	}
9250
9251	I915_WRITE_FW(GEN6_PCODE_DATA, val);
9252	I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
9253	I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
9254
9255	if (__intel_wait_for_register_fw(dev_priv,
9256					 GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
9257					 fast_timeout_us, slow_timeout_ms,
9258					 NULL)) {
9259		DRM_ERROR("timeout waiting for pcode write of 0x%08x to mbox %x to finish for %ps\n",
9260			  val, mbox, __builtin_return_address(0));
9261		return -ETIMEDOUT;
9262	}
9263
9264	I915_WRITE_FW(GEN6_PCODE_DATA, 0);
9265
9266	if (INTEL_GEN(dev_priv) > 6)
9267		status = gen7_check_mailbox_status(dev_priv);
9268	else
9269		status = gen6_check_mailbox_status(dev_priv);
9270
9271	if (status) {
9272		DRM_DEBUG_DRIVER("warning: pcode (write of 0x%08x to mbox %x) mailbox access failed for %ps: %d\n",
9273				 val, mbox, __builtin_return_address(0), status);
9274		return status;
9275	}
9276
9277	return 0;
9278}
9279
9280static bool skl_pcode_try_request(struct drm_i915_private *dev_priv, u32 mbox,
9281				  u32 request, u32 reply_mask, u32 reply,
9282				  u32 *status)
9283{
9284	u32 val = request;
9285
9286	*status = sandybridge_pcode_read(dev_priv, mbox, &val);
9287
9288	return *status || ((val & reply_mask) == reply);
9289}
9290
9291/**
9292 * skl_pcode_request - send PCODE request until acknowledgment
9293 * @dev_priv: device private
9294 * @mbox: PCODE mailbox ID the request is targeted for
9295 * @request: request ID
9296 * @reply_mask: mask used to check for request acknowledgment
9297 * @reply: value used to check for request acknowledgment
9298 * @timeout_base_ms: timeout for polling with preemption enabled
9299 *
9300 * Keep resending the @request to @mbox until PCODE acknowledges it, PCODE
9301 * reports an error or an overall timeout of @timeout_base_ms+50 ms expires.
9302 * The request is acknowledged once the PCODE reply dword equals @reply after
9303 * applying @reply_mask. Polling is first attempted with preemption enabled
9304 * for @timeout_base_ms and if this times out for another 50 ms with
9305 * preemption disabled.
9306 *
9307 * Returns 0 on success, %-ETIMEDOUT in case of a timeout, <0 in case of some
9308 * other error as reported by PCODE.
9309 */
9310int skl_pcode_request(struct drm_i915_private *dev_priv, u32 mbox, u32 request,
9311		      u32 reply_mask, u32 reply, int timeout_base_ms)
9312{
9313	u32 status;
9314	int ret;
9315
9316	WARN_ON(!mutex_is_locked(&dev_priv->pcu_lock));
9317
9318#define COND skl_pcode_try_request(dev_priv, mbox, request, reply_mask, reply, \
9319				   &status)
9320
9321	/*
9322	 * Prime the PCODE by doing a request first. Normally it guarantees
9323	 * that a subsequent request, at most @timeout_base_ms later, succeeds.
9324	 * _wait_for() doesn't guarantee when its passed condition is evaluated
9325	 * first, so send the first request explicitly.
9326	 */
9327	if (COND) {
9328		ret = 0;
9329		goto out;
9330	}
9331	ret = _wait_for(COND, timeout_base_ms * 1000, 10, 10);
9332	if (!ret)
9333		goto out;
9334
9335	/*
9336	 * The above can time out if the number of requests was low (2 in the
9337	 * worst case) _and_ PCODE was busy for some reason even after a
9338	 * (queued) request and @timeout_base_ms delay. As a workaround retry
9339	 * the poll with preemption disabled to maximize the number of
9340	 * requests. Increase the timeout from @timeout_base_ms to 50ms to
9341	 * account for interrupts that could reduce the number of these
9342	 * requests, and for any quirks of the PCODE firmware that delays
9343	 * the request completion.
9344	 */
9345	DRM_DEBUG_KMS("PCODE timeout, retrying with preemption disabled\n");
9346	WARN_ON_ONCE(timeout_base_ms > 3);
9347	preempt_disable();
9348	ret = wait_for_atomic(COND, 50);
9349	preempt_enable();
9350
9351out:
9352	return ret ? ret : status;
9353#undef COND
9354}
9355
9356static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
9357{
9358	struct intel_rps *rps = &dev_priv->gt_pm.rps;
9359
9360	/*
9361	 * N = val - 0xb7
9362	 * Slow = Fast = GPLL ref * N
9363	 */
9364	return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * (val - 0xb7), 1000);
9365}
9366
9367static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
9368{
9369	struct intel_rps *rps = &dev_priv->gt_pm.rps;
9370
9371	return DIV_ROUND_CLOSEST(1000 * val, rps->gpll_ref_freq) + 0xb7;
9372}
9373
9374static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
9375{
9376	struct intel_rps *rps = &dev_priv->gt_pm.rps;
9377
9378	/*
9379	 * N = val / 2
9380	 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
9381	 */
9382	return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * val, 2 * 2 * 1000);
9383}
9384
9385static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
9386{
9387	struct intel_rps *rps = &dev_priv->gt_pm.rps;
9388
9389	/* CHV needs even values */
9390	return DIV_ROUND_CLOSEST(2 * 1000 * val, rps->gpll_ref_freq) * 2;
9391}
9392
9393int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
9394{
9395	if (INTEL_GEN(dev_priv) >= 9)
9396		return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
9397					 GEN9_FREQ_SCALER);
9398	else if (IS_CHERRYVIEW(dev_priv))
9399		return chv_gpu_freq(dev_priv, val);
9400	else if (IS_VALLEYVIEW(dev_priv))
9401		return byt_gpu_freq(dev_priv, val);
9402	else
9403		return val * GT_FREQUENCY_MULTIPLIER;
9404}
9405
9406int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
9407{
9408	if (INTEL_GEN(dev_priv) >= 9)
9409		return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
9410					 GT_FREQUENCY_MULTIPLIER);
9411	else if (IS_CHERRYVIEW(dev_priv))
9412		return chv_freq_opcode(dev_priv, val);
9413	else if (IS_VALLEYVIEW(dev_priv))
9414		return byt_freq_opcode(dev_priv, val);
9415	else
9416		return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
9417}
9418
9419void intel_pm_setup(struct drm_i915_private *dev_priv)
9420{
9421	mutex_init(&dev_priv->pcu_lock);
9422
9423	atomic_set(&dev_priv->gt_pm.rps.num_waiters, 0);
9424
9425	dev_priv->runtime_pm.suspended = false;
9426	atomic_set(&dev_priv->runtime_pm.wakeref_count, 0);
9427}
9428
9429static u64 vlv_residency_raw(struct drm_i915_private *dev_priv,
9430			     const i915_reg_t reg)
9431{
9432	u32 lower, upper, tmp;
9433	int loop = 2;
9434
9435	/*
9436	 * The register accessed do not need forcewake. We borrow
9437	 * uncore lock to prevent concurrent access to range reg.
9438	 */
9439	lockdep_assert_held(&dev_priv->uncore.lock);
9440
9441	/*
9442	 * vlv and chv residency counters are 40 bits in width.
9443	 * With a control bit, we can choose between upper or lower
9444	 * 32bit window into this counter.
9445	 *
9446	 * Although we always use the counter in high-range mode elsewhere,
9447	 * userspace may attempt to read the value before rc6 is initialised,
9448	 * before we have set the default VLV_COUNTER_CONTROL value. So always
9449	 * set the high bit to be safe.
9450	 */
9451	I915_WRITE_FW(VLV_COUNTER_CONTROL,
9452		      _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
9453	upper = I915_READ_FW(reg);
9454	do {
9455		tmp = upper;
9456
9457		I915_WRITE_FW(VLV_COUNTER_CONTROL,
9458			      _MASKED_BIT_DISABLE(VLV_COUNT_RANGE_HIGH));
9459		lower = I915_READ_FW(reg);
9460
9461		I915_WRITE_FW(VLV_COUNTER_CONTROL,
9462			      _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
9463		upper = I915_READ_FW(reg);
9464	} while (upper != tmp && --loop);
9465
9466	/*
9467	 * Everywhere else we always use VLV_COUNTER_CONTROL with the
9468	 * VLV_COUNT_RANGE_HIGH bit set - so it is safe to leave it set
9469	 * now.
9470	 */
9471
9472	return lower | (u64)upper << 8;
9473}
9474
9475u64 intel_rc6_residency_ns(struct drm_i915_private *dev_priv,
9476			   const i915_reg_t reg)
9477{
9478	u64 time_hw, prev_hw, overflow_hw;
9479	unsigned int fw_domains;
9480	unsigned long flags;
9481	unsigned int i;
9482	u32 mul, div;
9483
9484	if (!HAS_RC6(dev_priv))
9485		return 0;
9486
9487	/*
9488	 * Store previous hw counter values for counter wrap-around handling.
9489	 *
9490	 * There are only four interesting registers and they live next to each
9491	 * other so we can use the relative address, compared to the smallest
9492	 * one as the index into driver storage.
9493	 */
9494	i = (i915_mmio_reg_offset(reg) -
9495	     i915_mmio_reg_offset(GEN6_GT_GFX_RC6_LOCKED)) / sizeof(u32);
9496	if (WARN_ON_ONCE(i >= ARRAY_SIZE(dev_priv->gt_pm.rc6.cur_residency)))
9497		return 0;
9498
9499	fw_domains = intel_uncore_forcewake_for_reg(dev_priv, reg, FW_REG_READ);
9500
9501	spin_lock_irqsave(&dev_priv->uncore.lock, flags);
9502	intel_uncore_forcewake_get__locked(dev_priv, fw_domains);
9503
9504	/* On VLV and CHV, residency time is in CZ units rather than 1.28us */
9505	if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
9506		mul = 1000000;
9507		div = dev_priv->czclk_freq;
9508		overflow_hw = BIT_ULL(40);
9509		time_hw = vlv_residency_raw(dev_priv, reg);
9510	} else {
9511		/* 833.33ns units on Gen9LP, 1.28us elsewhere. */
9512		if (IS_GEN9_LP(dev_priv)) {
9513			mul = 10000;
9514			div = 12;
9515		} else {
9516			mul = 1280;
9517			div = 1;
9518		}
9519
9520		overflow_hw = BIT_ULL(32);
9521		time_hw = I915_READ_FW(reg);
9522	}
9523
9524	/*
9525	 * Counter wrap handling.
9526	 *
9527	 * But relying on a sufficient frequency of queries otherwise counters
9528	 * can still wrap.
9529	 */
9530	prev_hw = dev_priv->gt_pm.rc6.prev_hw_residency[i];
9531	dev_priv->gt_pm.rc6.prev_hw_residency[i] = time_hw;
9532
9533	/* RC6 delta from last sample. */
9534	if (time_hw >= prev_hw)
9535		time_hw -= prev_hw;
9536	else
9537		time_hw += overflow_hw - prev_hw;
9538
9539	/* Add delta to RC6 extended raw driver copy. */
9540	time_hw += dev_priv->gt_pm.rc6.cur_residency[i];
9541	dev_priv->gt_pm.rc6.cur_residency[i] = time_hw;
9542
9543	intel_uncore_forcewake_put__locked(dev_priv, fw_domains);
9544	spin_unlock_irqrestore(&dev_priv->uncore.lock, flags);
9545
9546	return mul_u64_u32_div(time_hw, mul, div);
9547}
9548
9549u32 intel_get_cagf(struct drm_i915_private *dev_priv, u32 rpstat)
9550{
9551	u32 cagf;
9552
9553	if (INTEL_GEN(dev_priv) >= 9)
9554		cagf = (rpstat & GEN9_CAGF_MASK) >> GEN9_CAGF_SHIFT;
9555	else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
9556		cagf = (rpstat & HSW_CAGF_MASK) >> HSW_CAGF_SHIFT;
9557	else
9558		cagf = (rpstat & GEN6_CAGF_MASK) >> GEN6_CAGF_SHIFT;
9559
9560	return  cagf;
9561}