xref: /linux/drivers/gpu/drm/i915/display/i9xx_wm.c (revision 815e260a18a3af4dab59025ee99a7156c0e8b5e0)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2023 Intel Corporation
4  */
5 
6 #include <linux/iopoll.h>
7 
8 #include <drm/drm_print.h>
9 
10 #include "soc/intel_dram.h"
11 
12 #include "i915_drv.h"
13 #include "i915_reg.h"
14 #include "i9xx_wm.h"
15 #include "i9xx_wm_regs.h"
16 #include "intel_atomic.h"
17 #include "intel_bo.h"
18 #include "intel_de.h"
19 #include "intel_display.h"
20 #include "intel_display_regs.h"
21 #include "intel_display_trace.h"
22 #include "intel_fb.h"
23 #include "intel_mchbar_regs.h"
24 #include "intel_wm.h"
25 #include "skl_watermark.h"
26 #include "vlv_sideband.h"
27 
28 struct intel_watermark_params {
29 	u16 fifo_size;
30 	u16 max_wm;
31 	u8 default_wm;
32 	u8 guard_size;
33 	u8 cacheline_size;
34 };
35 
36 /* used in computing the new watermarks state */
37 struct intel_wm_config {
38 	unsigned int num_pipes_active;
39 	bool sprites_enabled;
40 	bool sprites_scaled;
41 };
42 
43 struct cxsr_latency {
44 	bool is_desktop : 1;
45 	bool is_ddr3 : 1;
46 	u16 fsb_freq;
47 	u16 mem_freq;
48 	u16 display_sr;
49 	u16 display_hpll_disable;
50 	u16 cursor_sr;
51 	u16 cursor_hpll_disable;
52 };
53 
54 static const struct cxsr_latency cxsr_latency_table[] = {
55 	{1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
56 	{1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
57 	{1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
58 	{1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
59 	{1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
60 
61 	{1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
62 	{1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
63 	{1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
64 	{1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
65 	{1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
66 
67 	{1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
68 	{1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
69 	{1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
70 	{1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
71 	{1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
72 
73 	{0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
74 	{0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
75 	{0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
76 	{0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
77 	{0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
78 
79 	{0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
80 	{0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
81 	{0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
82 	{0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
83 	{0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
84 
85 	{0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
86 	{0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
87 	{0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
88 	{0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
89 	{0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
90 };
91 
92 static const struct cxsr_latency *pnv_get_cxsr_latency(struct intel_display *display)
93 {
94 	const struct dram_info *dram_info = intel_dram_info(display->drm);
95 	bool is_ddr3 = dram_info->type == INTEL_DRAM_DDR3;
96 	int i;
97 
98 	for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
99 		const struct cxsr_latency *latency = &cxsr_latency_table[i];
100 		bool is_desktop = !display->platform.mobile;
101 
102 		if (is_desktop == latency->is_desktop &&
103 		    is_ddr3 == latency->is_ddr3 &&
104 		    DIV_ROUND_CLOSEST(dram_info->fsb_freq, 1000) == latency->fsb_freq &&
105 		    DIV_ROUND_CLOSEST(dram_info->mem_freq, 1000) == latency->mem_freq)
106 			return latency;
107 	}
108 
109 	drm_dbg_kms(display->drm,
110 		    "Could not find CxSR latency for %s, FSB %u kHz, MEM %u kHz\n",
111 		    intel_dram_type_str(dram_info->type),
112 		    dram_info->fsb_freq, dram_info->mem_freq);
113 
114 	return NULL;
115 }
116 
117 static void chv_set_memory_dvfs(struct intel_display *display, bool enable)
118 {
119 	u32 val;
120 	int ret;
121 
122 	vlv_punit_get(display->drm);
123 
124 	val = vlv_punit_read(display->drm, PUNIT_REG_DDR_SETUP2);
125 	if (enable)
126 		val &= ~FORCE_DDR_HIGH_FREQ;
127 	else
128 		val |= FORCE_DDR_HIGH_FREQ;
129 	val &= ~FORCE_DDR_LOW_FREQ;
130 	val |= FORCE_DDR_FREQ_REQ_ACK;
131 	vlv_punit_write(display->drm, PUNIT_REG_DDR_SETUP2, val);
132 
133 	ret = poll_timeout_us(val = vlv_punit_read(display->drm, PUNIT_REG_DDR_SETUP2),
134 			      (val & FORCE_DDR_FREQ_REQ_ACK) == 0,
135 			      500, 3000, false);
136 	if (ret)
137 		drm_err(display->drm,
138 			"timed out waiting for Punit DDR DVFS request\n");
139 
140 	vlv_punit_put(display->drm);
141 }
142 
143 static void chv_set_memory_pm5(struct intel_display *display, bool enable)
144 {
145 	u32 val;
146 
147 	vlv_punit_get(display->drm);
148 
149 	val = vlv_punit_read(display->drm, PUNIT_REG_DSPSSPM);
150 	if (enable)
151 		val |= DSP_MAXFIFO_PM5_ENABLE;
152 	else
153 		val &= ~DSP_MAXFIFO_PM5_ENABLE;
154 	vlv_punit_write(display->drm, PUNIT_REG_DSPSSPM, val);
155 
156 	vlv_punit_put(display->drm);
157 }
158 
159 #define FW_WM(value, plane) \
160 	(((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
161 
162 static bool _intel_set_memory_cxsr(struct intel_display *display, bool enable)
163 {
164 	bool was_enabled;
165 	u32 val;
166 
167 	if (display->platform.valleyview || display->platform.cherryview) {
168 		was_enabled = intel_de_read(display, FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
169 		intel_de_write(display, FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
170 		intel_de_posting_read(display, FW_BLC_SELF_VLV);
171 	} else if (display->platform.g4x || display->platform.i965gm) {
172 		was_enabled = intel_de_read(display, FW_BLC_SELF) & FW_BLC_SELF_EN;
173 		intel_de_write(display, FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
174 		intel_de_posting_read(display, FW_BLC_SELF);
175 	} else if (display->platform.pineview) {
176 		val = intel_de_read(display, DSPFW3(display));
177 		was_enabled = val & PINEVIEW_SELF_REFRESH_EN;
178 		if (enable)
179 			val |= PINEVIEW_SELF_REFRESH_EN;
180 		else
181 			val &= ~PINEVIEW_SELF_REFRESH_EN;
182 		intel_de_write(display, DSPFW3(display), val);
183 		intel_de_posting_read(display, DSPFW3(display));
184 	} else if (display->platform.i945g || display->platform.i945gm) {
185 		was_enabled = intel_de_read(display, FW_BLC_SELF) & FW_BLC_SELF_EN;
186 		val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
187 			       _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
188 		intel_de_write(display, FW_BLC_SELF, val);
189 		intel_de_posting_read(display, FW_BLC_SELF);
190 	} else if (display->platform.i915gm) {
191 		/*
192 		 * FIXME can't find a bit like this for 915G, and
193 		 * yet it does have the related watermark in
194 		 * FW_BLC_SELF. What's going on?
195 		 */
196 		was_enabled = intel_de_read(display, INSTPM) & INSTPM_SELF_EN;
197 		val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
198 			       _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
199 		intel_de_write(display, INSTPM, val);
200 		intel_de_posting_read(display, INSTPM);
201 	} else {
202 		return false;
203 	}
204 
205 	trace_intel_memory_cxsr(display, was_enabled, enable);
206 
207 	drm_dbg_kms(display->drm, "memory self-refresh is %s (was %s)\n",
208 		    str_enabled_disabled(enable),
209 		    str_enabled_disabled(was_enabled));
210 
211 	return was_enabled;
212 }
213 
214 /**
215  * intel_set_memory_cxsr - Configure CxSR state
216  * @display: display device
217  * @enable: Allow vs. disallow CxSR
218  *
219  * Allow or disallow the system to enter a special CxSR
220  * (C-state self refresh) state. What typically happens in CxSR mode
221  * is that several display FIFOs may get combined into a single larger
222  * FIFO for a particular plane (so called max FIFO mode) to allow the
223  * system to defer memory fetches longer, and the memory will enter
224  * self refresh.
225  *
226  * Note that enabling CxSR does not guarantee that the system enter
227  * this special mode, nor does it guarantee that the system stays
228  * in that mode once entered. So this just allows/disallows the system
229  * to autonomously utilize the CxSR mode. Other factors such as core
230  * C-states will affect when/if the system actually enters/exits the
231  * CxSR mode.
232  *
233  * Note that on VLV/CHV this actually only controls the max FIFO mode,
234  * and the system is free to enter/exit memory self refresh at any time
235  * even when the use of CxSR has been disallowed.
236  *
237  * While the system is actually in the CxSR/max FIFO mode, some plane
238  * control registers will not get latched on vblank. Thus in order to
239  * guarantee the system will respond to changes in the plane registers
240  * we must always disallow CxSR prior to making changes to those registers.
241  * Unfortunately the system will re-evaluate the CxSR conditions at
242  * frame start which happens after vblank start (which is when the plane
243  * registers would get latched), so we can't proceed with the plane update
244  * during the same frame where we disallowed CxSR.
245  *
246  * Certain platforms also have a deeper HPLL SR mode. Fortunately the
247  * HPLL SR mode depends on CxSR itself, so we don't have to hand hold
248  * the hardware w.r.t. HPLL SR when writing to plane registers.
249  * Disallowing just CxSR is sufficient.
250  */
251 bool intel_set_memory_cxsr(struct intel_display *display, bool enable)
252 {
253 	bool ret;
254 
255 	mutex_lock(&display->wm.wm_mutex);
256 	ret = _intel_set_memory_cxsr(display, enable);
257 	if (display->platform.valleyview || display->platform.cherryview)
258 		display->wm.vlv.cxsr = enable;
259 	else if (display->platform.g4x)
260 		display->wm.g4x.cxsr = enable;
261 	mutex_unlock(&display->wm.wm_mutex);
262 
263 	return ret;
264 }
265 
266 /*
267  * Latency for FIFO fetches is dependent on several factors:
268  *   - memory configuration (speed, channels)
269  *   - chipset
270  *   - current MCH state
271  * It can be fairly high in some situations, so here we assume a fairly
272  * pessimal value.  It's a tradeoff between extra memory fetches (if we
273  * set this value too high, the FIFO will fetch frequently to stay full)
274  * and power consumption (set it too low to save power and we might see
275  * FIFO underruns and display "flicker").
276  *
277  * A value of 5us seems to be a good balance; safe for very low end
278  * platforms but not overly aggressive on lower latency configs.
279  */
280 static const int pessimal_latency_ns = 5000;
281 
282 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
283 	((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
284 
285 static void vlv_get_fifo_size(struct intel_crtc_state *crtc_state)
286 {
287 	struct intel_display *display = to_intel_display(crtc_state);
288 	struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
289 	struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
290 	enum pipe pipe = crtc->pipe;
291 	int sprite0_start, sprite1_start;
292 	u32 dsparb, dsparb2, dsparb3;
293 
294 	switch (pipe) {
295 	case PIPE_A:
296 		dsparb = intel_de_read(display, DSPARB(display));
297 		dsparb2 = intel_de_read(display, DSPARB2);
298 		sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
299 		sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
300 		break;
301 	case PIPE_B:
302 		dsparb = intel_de_read(display, DSPARB(display));
303 		dsparb2 = intel_de_read(display, DSPARB2);
304 		sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
305 		sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
306 		break;
307 	case PIPE_C:
308 		dsparb2 = intel_de_read(display, DSPARB2);
309 		dsparb3 = intel_de_read(display, DSPARB3);
310 		sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
311 		sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
312 		break;
313 	default:
314 		MISSING_CASE(pipe);
315 		return;
316 	}
317 
318 	fifo_state->plane[PLANE_PRIMARY] = sprite0_start;
319 	fifo_state->plane[PLANE_SPRITE0] = sprite1_start - sprite0_start;
320 	fifo_state->plane[PLANE_SPRITE1] = 511 - sprite1_start;
321 	fifo_state->plane[PLANE_CURSOR] = 63;
322 }
323 
324 static int i9xx_get_fifo_size(struct intel_display *display,
325 			      enum i9xx_plane_id i9xx_plane)
326 {
327 	u32 dsparb = intel_de_read(display, DSPARB(display));
328 	int size;
329 
330 	size = dsparb & 0x7f;
331 	if (i9xx_plane == PLANE_B)
332 		size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
333 
334 	drm_dbg_kms(display->drm, "FIFO size - (0x%08x) %c: %d\n",
335 		    dsparb, plane_name(i9xx_plane), size);
336 
337 	return size;
338 }
339 
340 static int i830_get_fifo_size(struct intel_display *display,
341 			      enum i9xx_plane_id i9xx_plane)
342 {
343 	u32 dsparb = intel_de_read(display, DSPARB(display));
344 	int size;
345 
346 	size = dsparb & 0x1ff;
347 	if (i9xx_plane == PLANE_B)
348 		size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
349 	size >>= 1; /* Convert to cachelines */
350 
351 	drm_dbg_kms(display->drm, "FIFO size - (0x%08x) %c: %d\n",
352 		    dsparb, plane_name(i9xx_plane), size);
353 
354 	return size;
355 }
356 
357 static int i845_get_fifo_size(struct intel_display *display,
358 			      enum i9xx_plane_id i9xx_plane)
359 {
360 	u32 dsparb = intel_de_read(display, DSPARB(display));
361 	int size;
362 
363 	size = dsparb & 0x7f;
364 	size >>= 2; /* Convert to cachelines */
365 
366 	drm_dbg_kms(display->drm, "FIFO size - (0x%08x) %c: %d\n",
367 		    dsparb, plane_name(i9xx_plane), size);
368 
369 	return size;
370 }
371 
372 /* Pineview has different values for various configs */
373 static const struct intel_watermark_params pnv_display_wm = {
374 	.fifo_size = PINEVIEW_DISPLAY_FIFO,
375 	.max_wm = PINEVIEW_MAX_WM,
376 	.default_wm = PINEVIEW_DFT_WM,
377 	.guard_size = PINEVIEW_GUARD_WM,
378 	.cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
379 };
380 
381 static const struct intel_watermark_params pnv_display_hplloff_wm = {
382 	.fifo_size = PINEVIEW_DISPLAY_FIFO,
383 	.max_wm = PINEVIEW_MAX_WM,
384 	.default_wm = PINEVIEW_DFT_HPLLOFF_WM,
385 	.guard_size = PINEVIEW_GUARD_WM,
386 	.cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
387 };
388 
389 static const struct intel_watermark_params pnv_cursor_wm = {
390 	.fifo_size = PINEVIEW_CURSOR_FIFO,
391 	.max_wm = PINEVIEW_CURSOR_MAX_WM,
392 	.default_wm = PINEVIEW_CURSOR_DFT_WM,
393 	.guard_size = PINEVIEW_CURSOR_GUARD_WM,
394 	.cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
395 };
396 
397 static const struct intel_watermark_params pnv_cursor_hplloff_wm = {
398 	.fifo_size = PINEVIEW_CURSOR_FIFO,
399 	.max_wm = PINEVIEW_CURSOR_MAX_WM,
400 	.default_wm = PINEVIEW_CURSOR_DFT_WM,
401 	.guard_size = PINEVIEW_CURSOR_GUARD_WM,
402 	.cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
403 };
404 
405 static const struct intel_watermark_params i965_cursor_wm_info = {
406 	.fifo_size = I965_CURSOR_FIFO,
407 	.max_wm = I965_CURSOR_MAX_WM,
408 	.default_wm = I965_CURSOR_DFT_WM,
409 	.guard_size = 2,
410 	.cacheline_size = I915_FIFO_LINE_SIZE,
411 };
412 
413 static const struct intel_watermark_params i945_wm_info = {
414 	.fifo_size = I945_FIFO_SIZE,
415 	.max_wm = I915_MAX_WM,
416 	.default_wm = 1,
417 	.guard_size = 2,
418 	.cacheline_size = I915_FIFO_LINE_SIZE,
419 };
420 
421 static const struct intel_watermark_params i915_wm_info = {
422 	.fifo_size = I915_FIFO_SIZE,
423 	.max_wm = I915_MAX_WM,
424 	.default_wm = 1,
425 	.guard_size = 2,
426 	.cacheline_size = I915_FIFO_LINE_SIZE,
427 };
428 
429 static const struct intel_watermark_params i830_a_wm_info = {
430 	.fifo_size = I855GM_FIFO_SIZE,
431 	.max_wm = I915_MAX_WM,
432 	.default_wm = 1,
433 	.guard_size = 2,
434 	.cacheline_size = I830_FIFO_LINE_SIZE,
435 };
436 
437 static const struct intel_watermark_params i830_bc_wm_info = {
438 	.fifo_size = I855GM_FIFO_SIZE,
439 	.max_wm = I915_MAX_WM / 2,
440 	.default_wm = 1,
441 	.guard_size = 2,
442 	.cacheline_size = I830_FIFO_LINE_SIZE,
443 };
444 
445 static const struct intel_watermark_params i845_wm_info = {
446 	.fifo_size = I830_FIFO_SIZE,
447 	.max_wm = I915_MAX_WM,
448 	.default_wm = 1,
449 	.guard_size = 2,
450 	.cacheline_size = I830_FIFO_LINE_SIZE,
451 };
452 
453 /**
454  * intel_wm_method1 - Method 1 / "small buffer" watermark formula
455  * @pixel_rate: Pipe pixel rate in kHz
456  * @cpp: Plane bytes per pixel
457  * @latency: Memory wakeup latency in 0.1us units
458  *
459  * Compute the watermark using the method 1 or "small buffer"
460  * formula. The caller may additionally add extra cachelines
461  * to account for TLB misses and clock crossings.
462  *
463  * This method is concerned with the short term drain rate
464  * of the FIFO, ie. it does not account for blanking periods
465  * which would effectively reduce the average drain rate across
466  * a longer period. The name "small" refers to the fact the
467  * FIFO is relatively small compared to the amount of data
468  * fetched.
469  *
470  * The FIFO level vs. time graph might look something like:
471  *
472  *   |\   |\
473  *   | \  | \
474  * __---__---__ (- plane active, _ blanking)
475  * -> time
476  *
477  * or perhaps like this:
478  *
479  *   |\|\  |\|\
480  * __----__----__ (- plane active, _ blanking)
481  * -> time
482  *
483  * Returns:
484  * The watermark in bytes
485  */
486 static unsigned int intel_wm_method1(unsigned int pixel_rate,
487 				     unsigned int cpp,
488 				     unsigned int latency)
489 {
490 	u64 ret;
491 
492 	ret = mul_u32_u32(pixel_rate, cpp * latency);
493 	ret = DIV_ROUND_UP_ULL(ret, 10000);
494 
495 	return ret;
496 }
497 
498 /**
499  * intel_wm_method2 - Method 2 / "large buffer" watermark formula
500  * @pixel_rate: Pipe pixel rate in kHz
501  * @htotal: Pipe horizontal total
502  * @width: Plane width in pixels
503  * @cpp: Plane bytes per pixel
504  * @latency: Memory wakeup latency in 0.1us units
505  *
506  * Compute the watermark using the method 2 or "large buffer"
507  * formula. The caller may additionally add extra cachelines
508  * to account for TLB misses and clock crossings.
509  *
510  * This method is concerned with the long term drain rate
511  * of the FIFO, ie. it does account for blanking periods
512  * which effectively reduce the average drain rate across
513  * a longer period. The name "large" refers to the fact the
514  * FIFO is relatively large compared to the amount of data
515  * fetched.
516  *
517  * The FIFO level vs. time graph might look something like:
518  *
519  *    |\___       |\___
520  *    |    \___   |    \___
521  *    |        \  |        \
522  * __ --__--__--__--__--__--__ (- plane active, _ blanking)
523  * -> time
524  *
525  * Returns:
526  * The watermark in bytes
527  */
528 static unsigned int intel_wm_method2(unsigned int pixel_rate,
529 				     unsigned int htotal,
530 				     unsigned int width,
531 				     unsigned int cpp,
532 				     unsigned int latency)
533 {
534 	unsigned int ret;
535 
536 	/*
537 	 * FIXME remove once all users are computing
538 	 * watermarks in the correct place.
539 	 */
540 	if (WARN_ON_ONCE(htotal == 0))
541 		htotal = 1;
542 
543 	ret = (latency * pixel_rate) / (htotal * 10000);
544 	ret = (ret + 1) * width * cpp;
545 
546 	return ret;
547 }
548 
549 /**
550  * intel_calculate_wm - calculate watermark level
551  * @display: display device
552  * @pixel_rate: pixel clock
553  * @wm: chip FIFO params
554  * @fifo_size: size of the FIFO buffer
555  * @cpp: bytes per pixel
556  * @latency_ns: memory latency for the platform
557  *
558  * Calculate the watermark level (the level at which the display plane will
559  * start fetching from memory again).  Each chip has a different display
560  * FIFO size and allocation, so the caller needs to figure that out and pass
561  * in the correct intel_watermark_params structure.
562  *
563  * As the pixel clock runs, the FIFO will be drained at a rate that depends
564  * on the pixel size.  When it reaches the watermark level, it'll start
565  * fetching FIFO line sized based chunks from memory until the FIFO fills
566  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
567  * will occur, and a display engine hang could result.
568  */
569 static unsigned int intel_calculate_wm(struct intel_display *display,
570 				       int pixel_rate,
571 				       const struct intel_watermark_params *wm,
572 				       int fifo_size, int cpp,
573 				       unsigned int latency_ns)
574 {
575 	int entries, wm_size;
576 
577 	/*
578 	 * Note: we need to make sure we don't overflow for various clock &
579 	 * latency values.
580 	 * clocks go from a few thousand to several hundred thousand.
581 	 * latency is usually a few thousand
582 	 */
583 	entries = intel_wm_method1(pixel_rate, cpp,
584 				   latency_ns / 100);
585 	entries = DIV_ROUND_UP(entries, wm->cacheline_size) +
586 		wm->guard_size;
587 	drm_dbg_kms(display->drm, "FIFO entries required for mode: %d\n", entries);
588 
589 	wm_size = fifo_size - entries;
590 	drm_dbg_kms(display->drm, "FIFO watermark level: %d\n", wm_size);
591 
592 	/* Don't promote wm_size to unsigned... */
593 	if (wm_size > wm->max_wm)
594 		wm_size = wm->max_wm;
595 	if (wm_size <= 0)
596 		wm_size = wm->default_wm;
597 
598 	/*
599 	 * Bspec seems to indicate that the value shouldn't be lower than
600 	 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
601 	 * Lets go for 8 which is the burst size since certain platforms
602 	 * already use a hardcoded 8 (which is what the spec says should be
603 	 * done).
604 	 */
605 	if (wm_size <= 8)
606 		wm_size = 8;
607 
608 	return wm_size;
609 }
610 
611 static bool is_disabling(int old, int new, int threshold)
612 {
613 	return old >= threshold && new < threshold;
614 }
615 
616 static bool is_enabling(int old, int new, int threshold)
617 {
618 	return old < threshold && new >= threshold;
619 }
620 
621 static bool intel_crtc_active(struct intel_crtc *crtc)
622 {
623 	/* Be paranoid as we can arrive here with only partial
624 	 * state retrieved from the hardware during setup.
625 	 *
626 	 * We can ditch the adjusted_mode.crtc_clock check as soon
627 	 * as Haswell has gained clock readout/fastboot support.
628 	 *
629 	 * We can ditch the crtc->primary->state->fb check as soon as we can
630 	 * properly reconstruct framebuffers.
631 	 *
632 	 * FIXME: The intel_crtc->active here should be switched to
633 	 * crtc->state->active once we have proper CRTC states wired up
634 	 * for atomic.
635 	 */
636 	return crtc->active && crtc->base.primary->state->fb &&
637 		crtc->config->hw.adjusted_mode.crtc_clock;
638 }
639 
640 static struct intel_crtc *single_enabled_crtc(struct intel_display *display)
641 {
642 	struct intel_crtc *crtc, *enabled = NULL;
643 
644 	for_each_intel_crtc(display->drm, crtc) {
645 		if (intel_crtc_active(crtc)) {
646 			if (enabled)
647 				return NULL;
648 			enabled = crtc;
649 		}
650 	}
651 
652 	return enabled;
653 }
654 
655 static void pnv_update_wm(struct intel_display *display)
656 {
657 	struct intel_crtc *crtc;
658 	const struct cxsr_latency *latency;
659 	u32 reg;
660 	unsigned int wm;
661 
662 	latency = pnv_get_cxsr_latency(display);
663 	if (!latency) {
664 		drm_dbg_kms(display->drm, "Unknown FSB/MEM, disabling CxSR\n");
665 		intel_set_memory_cxsr(display, false);
666 		return;
667 	}
668 
669 	crtc = single_enabled_crtc(display);
670 	if (crtc) {
671 		const struct drm_framebuffer *fb =
672 			crtc->base.primary->state->fb;
673 		int pixel_rate = crtc->config->pixel_rate;
674 		int cpp = fb->format->cpp[0];
675 
676 		/* Display SR */
677 		wm = intel_calculate_wm(display, pixel_rate,
678 					&pnv_display_wm,
679 					pnv_display_wm.fifo_size,
680 					cpp, latency->display_sr);
681 		reg = intel_de_read(display, DSPFW1(display));
682 		reg &= ~DSPFW_SR_MASK;
683 		reg |= FW_WM(wm, SR);
684 		intel_de_write(display, DSPFW1(display), reg);
685 		drm_dbg_kms(display->drm, "DSPFW1 register is %x\n", reg);
686 
687 		/* cursor SR */
688 		wm = intel_calculate_wm(display, pixel_rate,
689 					&pnv_cursor_wm,
690 					pnv_display_wm.fifo_size,
691 					4, latency->cursor_sr);
692 		intel_de_rmw(display, DSPFW3(display),
693 			     DSPFW_CURSOR_SR_MASK, FW_WM(wm, CURSOR_SR));
694 
695 		/* Display HPLL off SR */
696 		wm = intel_calculate_wm(display, pixel_rate,
697 					&pnv_display_hplloff_wm,
698 					pnv_display_hplloff_wm.fifo_size,
699 					cpp, latency->display_hpll_disable);
700 		intel_de_rmw(display, DSPFW3(display),
701 			     DSPFW_HPLL_SR_MASK, FW_WM(wm, HPLL_SR));
702 
703 		/* cursor HPLL off SR */
704 		wm = intel_calculate_wm(display, pixel_rate,
705 					&pnv_cursor_hplloff_wm,
706 					pnv_display_hplloff_wm.fifo_size,
707 					4, latency->cursor_hpll_disable);
708 		reg = intel_de_read(display, DSPFW3(display));
709 		reg &= ~DSPFW_HPLL_CURSOR_MASK;
710 		reg |= FW_WM(wm, HPLL_CURSOR);
711 		intel_de_write(display, DSPFW3(display), reg);
712 		drm_dbg_kms(display->drm, "DSPFW3 register is %x\n", reg);
713 
714 		intel_set_memory_cxsr(display, true);
715 	} else {
716 		intel_set_memory_cxsr(display, false);
717 	}
718 }
719 
720 static bool i9xx_wm_need_update(const struct intel_plane_state *old_plane_state,
721 				const struct intel_plane_state *new_plane_state)
722 {
723 	/* Update watermarks on tiling or size changes. */
724 	if (old_plane_state->uapi.visible != new_plane_state->uapi.visible)
725 		return true;
726 
727 	if (!old_plane_state->hw.fb || !new_plane_state->hw.fb)
728 		return false;
729 
730 	if (old_plane_state->hw.fb->modifier != new_plane_state->hw.fb->modifier ||
731 	    old_plane_state->hw.rotation != new_plane_state->hw.rotation ||
732 	    drm_rect_width(&old_plane_state->uapi.src) != drm_rect_width(&new_plane_state->uapi.src) ||
733 	    drm_rect_height(&old_plane_state->uapi.src) != drm_rect_height(&new_plane_state->uapi.src) ||
734 	    drm_rect_width(&old_plane_state->uapi.dst) != drm_rect_width(&new_plane_state->uapi.dst) ||
735 	    drm_rect_height(&old_plane_state->uapi.dst) != drm_rect_height(&new_plane_state->uapi.dst))
736 		return true;
737 
738 	return false;
739 }
740 
741 static void i9xx_wm_compute(struct intel_crtc_state *new_crtc_state,
742 			    const struct intel_plane_state *old_plane_state,
743 			    const struct intel_plane_state *new_plane_state)
744 {
745 	bool turn_off, turn_on, visible, was_visible, mode_changed;
746 
747 	mode_changed = intel_crtc_needs_modeset(new_crtc_state);
748 	was_visible = old_plane_state->uapi.visible;
749 	visible = new_plane_state->uapi.visible;
750 
751 	if (!was_visible && !visible)
752 		return;
753 
754 	turn_off = was_visible && (!visible || mode_changed);
755 	turn_on = visible && (!was_visible || mode_changed);
756 
757 	/* FIXME nuke when all wm code is atomic */
758 	if (turn_on) {
759 		new_crtc_state->update_wm_pre = true;
760 	} else if (turn_off) {
761 		new_crtc_state->update_wm_post = true;
762 	} else if (i9xx_wm_need_update(old_plane_state, new_plane_state)) {
763 		/* FIXME bollocks */
764 		new_crtc_state->update_wm_pre = true;
765 		new_crtc_state->update_wm_post = true;
766 	}
767 }
768 
769 static int i9xx_compute_watermarks(struct intel_atomic_state *state,
770 				   struct intel_crtc *crtc)
771 {
772 	struct intel_crtc_state *new_crtc_state =
773 		intel_atomic_get_new_crtc_state(state, crtc);
774 	const struct intel_plane_state *old_plane_state;
775 	const struct intel_plane_state *new_plane_state;
776 	struct intel_plane *plane;
777 	int i;
778 
779 	for_each_oldnew_intel_plane_in_state(state, plane, old_plane_state,
780 					     new_plane_state, i) {
781 		if (plane->pipe != crtc->pipe)
782 			continue;
783 
784 		i9xx_wm_compute(new_crtc_state, old_plane_state, new_plane_state);
785 	}
786 
787 	return 0;
788 }
789 
790 /*
791  * Documentation says:
792  * "If the line size is small, the TLB fetches can get in the way of the
793  *  data fetches, causing some lag in the pixel data return which is not
794  *  accounted for in the above formulas. The following adjustment only
795  *  needs to be applied if eight whole lines fit in the buffer at once.
796  *  The WM is adjusted upwards by the difference between the FIFO size
797  *  and the size of 8 whole lines. This adjustment is always performed
798  *  in the actual pixel depth regardless of whether FBC is enabled or not."
799  */
800 static unsigned int g4x_tlb_miss_wa(int fifo_size, int width, int cpp)
801 {
802 	int tlb_miss = fifo_size * 64 - width * cpp * 8;
803 
804 	return max(0, tlb_miss);
805 }
806 
807 static void g4x_write_wm_values(struct intel_display *display,
808 				const struct g4x_wm_values *wm)
809 {
810 	enum pipe pipe;
811 
812 	for_each_pipe(display, pipe)
813 		trace_g4x_wm(intel_crtc_for_pipe(display, pipe), wm);
814 
815 	intel_de_write(display, DSPFW1(display),
816 		       FW_WM(wm->sr.plane, SR) |
817 		       FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
818 		       FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
819 		       FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
820 	intel_de_write(display, DSPFW2(display),
821 		       (wm->fbc_en ? DSPFW_FBC_SR_EN : 0) |
822 		       FW_WM(wm->sr.fbc, FBC_SR) |
823 		       FW_WM(wm->hpll.fbc, FBC_HPLL_SR) |
824 		       FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEB) |
825 		       FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
826 		       FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
827 	intel_de_write(display, DSPFW3(display),
828 		       (wm->hpll_en ? DSPFW_HPLL_SR_EN : 0) |
829 		       FW_WM(wm->sr.cursor, CURSOR_SR) |
830 		       FW_WM(wm->hpll.cursor, HPLL_CURSOR) |
831 		       FW_WM(wm->hpll.plane, HPLL_SR));
832 
833 	intel_de_posting_read(display, DSPFW1(display));
834 }
835 
836 #define FW_WM_VLV(value, plane) \
837 	(((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
838 
839 static void vlv_write_wm_values(struct intel_display *display,
840 				const struct vlv_wm_values *wm)
841 {
842 	enum pipe pipe;
843 
844 	for_each_pipe(display, pipe) {
845 		trace_vlv_wm(intel_crtc_for_pipe(display, pipe), wm);
846 
847 		intel_de_write(display, VLV_DDL(pipe),
848 			       (wm->ddl[pipe].plane[PLANE_CURSOR] << DDL_CURSOR_SHIFT) |
849 			       (wm->ddl[pipe].plane[PLANE_SPRITE1] << DDL_SPRITE_SHIFT(1)) |
850 			       (wm->ddl[pipe].plane[PLANE_SPRITE0] << DDL_SPRITE_SHIFT(0)) |
851 			       (wm->ddl[pipe].plane[PLANE_PRIMARY] << DDL_PLANE_SHIFT));
852 	}
853 
854 	/*
855 	 * Zero the (unused) WM1 watermarks, and also clear all the
856 	 * high order bits so that there are no out of bounds values
857 	 * present in the registers during the reprogramming.
858 	 */
859 	intel_de_write(display, DSPHOWM, 0);
860 	intel_de_write(display, DSPHOWM1, 0);
861 	intel_de_write(display, DSPFW4, 0);
862 	intel_de_write(display, DSPFW5, 0);
863 	intel_de_write(display, DSPFW6, 0);
864 
865 	intel_de_write(display, DSPFW1(display),
866 		       FW_WM(wm->sr.plane, SR) |
867 		       FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
868 		       FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
869 		       FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
870 	intel_de_write(display, DSPFW2(display),
871 		       FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE1], SPRITEB) |
872 		       FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
873 		       FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
874 	intel_de_write(display, DSPFW3(display),
875 		       FW_WM(wm->sr.cursor, CURSOR_SR));
876 
877 	if (display->platform.cherryview) {
878 		intel_de_write(display, DSPFW7_CHV,
879 			       FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
880 			       FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
881 		intel_de_write(display, DSPFW8_CHV,
882 			       FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE1], SPRITEF) |
883 			       FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE0], SPRITEE));
884 		intel_de_write(display, DSPFW9_CHV,
885 			       FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_PRIMARY], PLANEC) |
886 			       FW_WM(wm->pipe[PIPE_C].plane[PLANE_CURSOR], CURSORC));
887 		intel_de_write(display, DSPHOWM,
888 			       FW_WM(wm->sr.plane >> 9, SR_HI) |
889 			       FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE1] >> 8, SPRITEF_HI) |
890 			       FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE0] >> 8, SPRITEE_HI) |
891 			       FW_WM(wm->pipe[PIPE_C].plane[PLANE_PRIMARY] >> 8, PLANEC_HI) |
892 			       FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
893 			       FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
894 			       FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
895 			       FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
896 			       FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
897 			       FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
898 	} else {
899 		intel_de_write(display, DSPFW7,
900 			       FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
901 			       FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
902 		intel_de_write(display, DSPHOWM,
903 			       FW_WM(wm->sr.plane >> 9, SR_HI) |
904 			       FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
905 			       FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
906 			       FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
907 			       FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
908 			       FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
909 			       FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
910 	}
911 
912 	intel_de_posting_read(display, DSPFW1(display));
913 }
914 
915 #undef FW_WM_VLV
916 
917 static void g4x_setup_wm_latency(struct intel_display *display)
918 {
919 	/* all latencies in usec */
920 	display->wm.pri_latency[G4X_WM_LEVEL_NORMAL] = 5;
921 	display->wm.pri_latency[G4X_WM_LEVEL_SR] = 12;
922 	display->wm.pri_latency[G4X_WM_LEVEL_HPLL] = 35;
923 
924 	display->wm.num_levels = G4X_WM_LEVEL_HPLL + 1;
925 }
926 
927 static int g4x_plane_fifo_size(enum plane_id plane_id, int level)
928 {
929 	/*
930 	 * DSPCNTR[13] supposedly controls whether the
931 	 * primary plane can use the FIFO space otherwise
932 	 * reserved for the sprite plane. It's not 100% clear
933 	 * what the actual FIFO size is, but it looks like we
934 	 * can happily set both primary and sprite watermarks
935 	 * up to 127 cachelines. So that would seem to mean
936 	 * that either DSPCNTR[13] doesn't do anything, or that
937 	 * the total FIFO is >= 256 cachelines in size. Either
938 	 * way, we don't seem to have to worry about this
939 	 * repartitioning as the maximum watermark value the
940 	 * register can hold for each plane is lower than the
941 	 * minimum FIFO size.
942 	 */
943 	switch (plane_id) {
944 	case PLANE_CURSOR:
945 		return 63;
946 	case PLANE_PRIMARY:
947 		return level == G4X_WM_LEVEL_NORMAL ? 127 : 511;
948 	case PLANE_SPRITE0:
949 		return level == G4X_WM_LEVEL_NORMAL ? 127 : 0;
950 	default:
951 		MISSING_CASE(plane_id);
952 		return 0;
953 	}
954 }
955 
956 static int g4x_fbc_fifo_size(int level)
957 {
958 	switch (level) {
959 	case G4X_WM_LEVEL_SR:
960 		return 7;
961 	case G4X_WM_LEVEL_HPLL:
962 		return 15;
963 	default:
964 		MISSING_CASE(level);
965 		return 0;
966 	}
967 }
968 
969 static u16 g4x_compute_wm(const struct intel_crtc_state *crtc_state,
970 			  const struct intel_plane_state *plane_state,
971 			  int level)
972 {
973 	struct intel_display *display = to_intel_display(plane_state);
974 	struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
975 	const struct drm_display_mode *pipe_mode =
976 		&crtc_state->hw.pipe_mode;
977 	unsigned int latency = display->wm.pri_latency[level] * 10;
978 	unsigned int pixel_rate, htotal, cpp, width, wm;
979 
980 	if (latency == 0)
981 		return USHRT_MAX;
982 
983 	if (!intel_wm_plane_visible(crtc_state, plane_state))
984 		return 0;
985 
986 	cpp = plane_state->hw.fb->format->cpp[0];
987 
988 	/*
989 	 * WaUse32BppForSRWM:ctg,elk
990 	 *
991 	 * The spec fails to list this restriction for the
992 	 * HPLL watermark, which seems a little strange.
993 	 * Let's use 32bpp for the HPLL watermark as well.
994 	 */
995 	if (plane->id == PLANE_PRIMARY &&
996 	    level != G4X_WM_LEVEL_NORMAL)
997 		cpp = max(cpp, 4u);
998 
999 	pixel_rate = crtc_state->pixel_rate;
1000 	htotal = pipe_mode->crtc_htotal;
1001 	width = drm_rect_width(&plane_state->uapi.src) >> 16;
1002 
1003 	if (plane->id == PLANE_CURSOR) {
1004 		wm = intel_wm_method2(pixel_rate, htotal, width, cpp, latency);
1005 	} else if (plane->id == PLANE_PRIMARY &&
1006 		   level == G4X_WM_LEVEL_NORMAL) {
1007 		wm = intel_wm_method1(pixel_rate, cpp, latency);
1008 	} else {
1009 		unsigned int small, large;
1010 
1011 		small = intel_wm_method1(pixel_rate, cpp, latency);
1012 		large = intel_wm_method2(pixel_rate, htotal, width, cpp, latency);
1013 
1014 		wm = min(small, large);
1015 	}
1016 
1017 	wm += g4x_tlb_miss_wa(g4x_plane_fifo_size(plane->id, level),
1018 			      width, cpp);
1019 
1020 	wm = DIV_ROUND_UP(wm, 64) + 2;
1021 
1022 	return min_t(unsigned int, wm, USHRT_MAX);
1023 }
1024 
1025 static bool g4x_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1026 				 int level, enum plane_id plane_id, u16 value)
1027 {
1028 	struct intel_display *display = to_intel_display(crtc_state);
1029 	bool dirty = false;
1030 
1031 	for (; level < display->wm.num_levels; level++) {
1032 		struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1033 
1034 		dirty |= raw->plane[plane_id] != value;
1035 		raw->plane[plane_id] = value;
1036 	}
1037 
1038 	return dirty;
1039 }
1040 
1041 static bool g4x_raw_fbc_wm_set(struct intel_crtc_state *crtc_state,
1042 			       int level, u16 value)
1043 {
1044 	struct intel_display *display = to_intel_display(crtc_state);
1045 	bool dirty = false;
1046 
1047 	/* NORMAL level doesn't have an FBC watermark */
1048 	level = max(level, G4X_WM_LEVEL_SR);
1049 
1050 	for (; level < display->wm.num_levels; level++) {
1051 		struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1052 
1053 		dirty |= raw->fbc != value;
1054 		raw->fbc = value;
1055 	}
1056 
1057 	return dirty;
1058 }
1059 
1060 static u32 ilk_compute_fbc_wm(const struct intel_crtc_state *crtc_state,
1061 			      const struct intel_plane_state *plane_state,
1062 			      u32 pri_val);
1063 
1064 static bool g4x_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1065 				     const struct intel_plane_state *plane_state)
1066 {
1067 	struct intel_display *display = to_intel_display(crtc_state);
1068 	struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1069 	enum plane_id plane_id = plane->id;
1070 	bool dirty = false;
1071 	int level;
1072 
1073 	if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1074 		dirty |= g4x_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1075 		if (plane_id == PLANE_PRIMARY)
1076 			dirty |= g4x_raw_fbc_wm_set(crtc_state, 0, 0);
1077 		goto out;
1078 	}
1079 
1080 	for (level = 0; level < display->wm.num_levels; level++) {
1081 		struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1082 		int wm, max_wm;
1083 
1084 		wm = g4x_compute_wm(crtc_state, plane_state, level);
1085 		max_wm = g4x_plane_fifo_size(plane_id, level);
1086 
1087 		if (wm > max_wm)
1088 			break;
1089 
1090 		dirty |= raw->plane[plane_id] != wm;
1091 		raw->plane[plane_id] = wm;
1092 
1093 		if (plane_id != PLANE_PRIMARY ||
1094 		    level == G4X_WM_LEVEL_NORMAL)
1095 			continue;
1096 
1097 		wm = ilk_compute_fbc_wm(crtc_state, plane_state,
1098 					raw->plane[plane_id]);
1099 		max_wm = g4x_fbc_fifo_size(level);
1100 
1101 		/*
1102 		 * FBC wm is not mandatory as we
1103 		 * can always just disable its use.
1104 		 */
1105 		if (wm > max_wm)
1106 			wm = USHRT_MAX;
1107 
1108 		dirty |= raw->fbc != wm;
1109 		raw->fbc = wm;
1110 	}
1111 
1112 	/* mark watermarks as invalid */
1113 	dirty |= g4x_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1114 
1115 	if (plane_id == PLANE_PRIMARY)
1116 		dirty |= g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
1117 
1118  out:
1119 	if (dirty) {
1120 		drm_dbg_kms(display->drm,
1121 			    "%s watermarks: normal=%d, SR=%d, HPLL=%d\n",
1122 			    plane->base.name,
1123 			    crtc_state->wm.g4x.raw[G4X_WM_LEVEL_NORMAL].plane[plane_id],
1124 			    crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].plane[plane_id],
1125 			    crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].plane[plane_id]);
1126 
1127 		if (plane_id == PLANE_PRIMARY)
1128 			drm_dbg_kms(display->drm,
1129 				    "FBC watermarks: SR=%d, HPLL=%d\n",
1130 				    crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].fbc,
1131 				    crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].fbc);
1132 	}
1133 
1134 	return dirty;
1135 }
1136 
1137 static bool g4x_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1138 				      enum plane_id plane_id, int level)
1139 {
1140 	const struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1141 
1142 	return raw->plane[plane_id] <= g4x_plane_fifo_size(plane_id, level);
1143 }
1144 
1145 static bool g4x_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state,
1146 				     int level)
1147 {
1148 	struct intel_display *display = to_intel_display(crtc_state);
1149 
1150 	if (level >= display->wm.num_levels)
1151 		return false;
1152 
1153 	return g4x_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1154 		g4x_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1155 		g4x_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1156 }
1157 
1158 /* mark all levels starting from 'level' as invalid */
1159 static void g4x_invalidate_wms(struct intel_crtc *crtc,
1160 			       struct g4x_wm_state *wm_state, int level)
1161 {
1162 	if (level <= G4X_WM_LEVEL_NORMAL) {
1163 		enum plane_id plane_id;
1164 
1165 		for_each_plane_id_on_crtc(crtc, plane_id)
1166 			wm_state->wm.plane[plane_id] = USHRT_MAX;
1167 	}
1168 
1169 	if (level <= G4X_WM_LEVEL_SR) {
1170 		wm_state->cxsr = false;
1171 		wm_state->sr.cursor = USHRT_MAX;
1172 		wm_state->sr.plane = USHRT_MAX;
1173 		wm_state->sr.fbc = USHRT_MAX;
1174 	}
1175 
1176 	if (level <= G4X_WM_LEVEL_HPLL) {
1177 		wm_state->hpll_en = false;
1178 		wm_state->hpll.cursor = USHRT_MAX;
1179 		wm_state->hpll.plane = USHRT_MAX;
1180 		wm_state->hpll.fbc = USHRT_MAX;
1181 	}
1182 }
1183 
1184 static bool g4x_compute_fbc_en(const struct g4x_wm_state *wm_state,
1185 			       int level)
1186 {
1187 	if (level < G4X_WM_LEVEL_SR)
1188 		return false;
1189 
1190 	if (level >= G4X_WM_LEVEL_SR &&
1191 	    wm_state->sr.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_SR))
1192 		return false;
1193 
1194 	if (level >= G4X_WM_LEVEL_HPLL &&
1195 	    wm_state->hpll.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_HPLL))
1196 		return false;
1197 
1198 	return true;
1199 }
1200 
1201 static int _g4x_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1202 {
1203 	struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1204 	struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
1205 	u8 active_planes = crtc_state->active_planes & ~BIT(PLANE_CURSOR);
1206 	const struct g4x_pipe_wm *raw;
1207 	enum plane_id plane_id;
1208 	int level;
1209 
1210 	level = G4X_WM_LEVEL_NORMAL;
1211 	if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1212 		goto out;
1213 
1214 	raw = &crtc_state->wm.g4x.raw[level];
1215 	for_each_plane_id_on_crtc(crtc, plane_id)
1216 		wm_state->wm.plane[plane_id] = raw->plane[plane_id];
1217 
1218 	level = G4X_WM_LEVEL_SR;
1219 	if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1220 		goto out;
1221 
1222 	raw = &crtc_state->wm.g4x.raw[level];
1223 	wm_state->sr.plane = raw->plane[PLANE_PRIMARY];
1224 	wm_state->sr.cursor = raw->plane[PLANE_CURSOR];
1225 	wm_state->sr.fbc = raw->fbc;
1226 
1227 	wm_state->cxsr = active_planes == BIT(PLANE_PRIMARY);
1228 
1229 	level = G4X_WM_LEVEL_HPLL;
1230 	if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1231 		goto out;
1232 
1233 	raw = &crtc_state->wm.g4x.raw[level];
1234 	wm_state->hpll.plane = raw->plane[PLANE_PRIMARY];
1235 	wm_state->hpll.cursor = raw->plane[PLANE_CURSOR];
1236 	wm_state->hpll.fbc = raw->fbc;
1237 
1238 	wm_state->hpll_en = wm_state->cxsr;
1239 
1240 	level++;
1241 
1242  out:
1243 	if (level == G4X_WM_LEVEL_NORMAL)
1244 		return -EINVAL;
1245 
1246 	/* invalidate the higher levels */
1247 	g4x_invalidate_wms(crtc, wm_state, level);
1248 
1249 	/*
1250 	 * Determine if the FBC watermark(s) can be used. IF
1251 	 * this isn't the case we prefer to disable the FBC
1252 	 * watermark(s) rather than disable the SR/HPLL
1253 	 * level(s) entirely. 'level-1' is the highest valid
1254 	 * level here.
1255 	 */
1256 	wm_state->fbc_en = g4x_compute_fbc_en(wm_state, level - 1);
1257 
1258 	return 0;
1259 }
1260 
1261 static int g4x_compute_pipe_wm(struct intel_atomic_state *state,
1262 			       struct intel_crtc *crtc)
1263 {
1264 	struct intel_crtc_state *crtc_state =
1265 		intel_atomic_get_new_crtc_state(state, crtc);
1266 	const struct intel_plane_state *old_plane_state;
1267 	const struct intel_plane_state *new_plane_state;
1268 	struct intel_plane *plane;
1269 	unsigned int dirty = 0;
1270 	int i;
1271 
1272 	for_each_oldnew_intel_plane_in_state(state, plane,
1273 					     old_plane_state,
1274 					     new_plane_state, i) {
1275 		if (new_plane_state->hw.crtc != &crtc->base &&
1276 		    old_plane_state->hw.crtc != &crtc->base)
1277 			continue;
1278 
1279 		if (g4x_raw_plane_wm_compute(crtc_state, new_plane_state))
1280 			dirty |= BIT(plane->id);
1281 	}
1282 
1283 	if (!dirty)
1284 		return 0;
1285 
1286 	return _g4x_compute_pipe_wm(crtc_state);
1287 }
1288 
1289 static int g4x_compute_intermediate_wm(struct intel_atomic_state *state,
1290 				       struct intel_crtc *crtc)
1291 {
1292 	struct intel_display *display = to_intel_display(state);
1293 	struct intel_crtc_state *new_crtc_state =
1294 		intel_atomic_get_new_crtc_state(state, crtc);
1295 	const struct intel_crtc_state *old_crtc_state =
1296 		intel_atomic_get_old_crtc_state(state, crtc);
1297 	struct g4x_wm_state *intermediate = &new_crtc_state->wm.g4x.intermediate;
1298 	const struct g4x_wm_state *optimal = &new_crtc_state->wm.g4x.optimal;
1299 	const struct g4x_wm_state *active = &old_crtc_state->wm.g4x.optimal;
1300 	enum plane_id plane_id;
1301 
1302 	if (!new_crtc_state->hw.active ||
1303 	    intel_crtc_needs_modeset(new_crtc_state)) {
1304 		*intermediate = *optimal;
1305 
1306 		intermediate->cxsr = false;
1307 		intermediate->hpll_en = false;
1308 		goto out;
1309 	}
1310 
1311 	intermediate->cxsr = optimal->cxsr && active->cxsr &&
1312 		!new_crtc_state->disable_cxsr;
1313 	intermediate->hpll_en = optimal->hpll_en && active->hpll_en &&
1314 		!new_crtc_state->disable_cxsr;
1315 	intermediate->fbc_en = optimal->fbc_en && active->fbc_en;
1316 
1317 	for_each_plane_id_on_crtc(crtc, plane_id) {
1318 		intermediate->wm.plane[plane_id] =
1319 			max(optimal->wm.plane[plane_id],
1320 			    active->wm.plane[plane_id]);
1321 
1322 		drm_WARN_ON(display->drm, intermediate->wm.plane[plane_id] >
1323 			    g4x_plane_fifo_size(plane_id, G4X_WM_LEVEL_NORMAL));
1324 	}
1325 
1326 	intermediate->sr.plane = max(optimal->sr.plane,
1327 				     active->sr.plane);
1328 	intermediate->sr.cursor = max(optimal->sr.cursor,
1329 				      active->sr.cursor);
1330 	intermediate->sr.fbc = max(optimal->sr.fbc,
1331 				   active->sr.fbc);
1332 
1333 	intermediate->hpll.plane = max(optimal->hpll.plane,
1334 				       active->hpll.plane);
1335 	intermediate->hpll.cursor = max(optimal->hpll.cursor,
1336 					active->hpll.cursor);
1337 	intermediate->hpll.fbc = max(optimal->hpll.fbc,
1338 				     active->hpll.fbc);
1339 
1340 	drm_WARN_ON(display->drm,
1341 		    (intermediate->sr.plane >
1342 		     g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_SR) ||
1343 		     intermediate->sr.cursor >
1344 		     g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_SR)) &&
1345 		    intermediate->cxsr);
1346 	drm_WARN_ON(display->drm,
1347 		    (intermediate->sr.plane >
1348 		     g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_HPLL) ||
1349 		     intermediate->sr.cursor >
1350 		     g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_HPLL)) &&
1351 		    intermediate->hpll_en);
1352 
1353 	drm_WARN_ON(display->drm,
1354 		    intermediate->sr.fbc > g4x_fbc_fifo_size(1) &&
1355 		    intermediate->fbc_en && intermediate->cxsr);
1356 	drm_WARN_ON(display->drm,
1357 		    intermediate->hpll.fbc > g4x_fbc_fifo_size(2) &&
1358 		    intermediate->fbc_en && intermediate->hpll_en);
1359 
1360 out:
1361 	/*
1362 	 * If our intermediate WM are identical to the final WM, then we can
1363 	 * omit the post-vblank programming; only update if it's different.
1364 	 */
1365 	if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
1366 		new_crtc_state->wm.need_postvbl_update = true;
1367 
1368 	return 0;
1369 }
1370 
1371 static int g4x_compute_watermarks(struct intel_atomic_state *state,
1372 				  struct intel_crtc *crtc)
1373 {
1374 	int ret;
1375 
1376 	ret = g4x_compute_pipe_wm(state, crtc);
1377 	if (ret)
1378 		return ret;
1379 
1380 	ret = g4x_compute_intermediate_wm(state, crtc);
1381 	if (ret)
1382 		return ret;
1383 
1384 	return 0;
1385 }
1386 
1387 static void g4x_merge_wm(struct intel_display *display,
1388 			 struct g4x_wm_values *wm)
1389 {
1390 	struct intel_crtc *crtc;
1391 	int num_active_pipes = 0;
1392 
1393 	wm->cxsr = true;
1394 	wm->hpll_en = true;
1395 	wm->fbc_en = true;
1396 
1397 	for_each_intel_crtc(display->drm, crtc) {
1398 		const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1399 
1400 		if (!crtc->active)
1401 			continue;
1402 
1403 		if (!wm_state->cxsr)
1404 			wm->cxsr = false;
1405 		if (!wm_state->hpll_en)
1406 			wm->hpll_en = false;
1407 		if (!wm_state->fbc_en)
1408 			wm->fbc_en = false;
1409 
1410 		num_active_pipes++;
1411 	}
1412 
1413 	if (num_active_pipes != 1) {
1414 		wm->cxsr = false;
1415 		wm->hpll_en = false;
1416 		wm->fbc_en = false;
1417 	}
1418 
1419 	for_each_intel_crtc(display->drm, crtc) {
1420 		const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1421 		enum pipe pipe = crtc->pipe;
1422 
1423 		wm->pipe[pipe] = wm_state->wm;
1424 		if (crtc->active && wm->cxsr)
1425 			wm->sr = wm_state->sr;
1426 		if (crtc->active && wm->hpll_en)
1427 			wm->hpll = wm_state->hpll;
1428 	}
1429 }
1430 
1431 static void g4x_program_watermarks(struct intel_display *display)
1432 {
1433 	struct g4x_wm_values *old_wm = &display->wm.g4x;
1434 	struct g4x_wm_values new_wm = {};
1435 
1436 	g4x_merge_wm(display, &new_wm);
1437 
1438 	if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
1439 		return;
1440 
1441 	if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
1442 		_intel_set_memory_cxsr(display, false);
1443 
1444 	g4x_write_wm_values(display, &new_wm);
1445 
1446 	if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
1447 		_intel_set_memory_cxsr(display, true);
1448 
1449 	*old_wm = new_wm;
1450 }
1451 
1452 static void g4x_initial_watermarks(struct intel_atomic_state *state,
1453 				   struct intel_crtc *crtc)
1454 {
1455 	struct intel_display *display = to_intel_display(crtc);
1456 	const struct intel_crtc_state *crtc_state =
1457 		intel_atomic_get_new_crtc_state(state, crtc);
1458 
1459 	mutex_lock(&display->wm.wm_mutex);
1460 	crtc->wm.active.g4x = crtc_state->wm.g4x.intermediate;
1461 	g4x_program_watermarks(display);
1462 	mutex_unlock(&display->wm.wm_mutex);
1463 }
1464 
1465 static void g4x_optimize_watermarks(struct intel_atomic_state *state,
1466 				    struct intel_crtc *crtc)
1467 {
1468 	struct intel_display *display = to_intel_display(crtc);
1469 	const struct intel_crtc_state *crtc_state =
1470 		intel_atomic_get_new_crtc_state(state, crtc);
1471 
1472 	if (!crtc_state->wm.need_postvbl_update)
1473 		return;
1474 
1475 	mutex_lock(&display->wm.wm_mutex);
1476 	crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
1477 	g4x_program_watermarks(display);
1478 	mutex_unlock(&display->wm.wm_mutex);
1479 }
1480 
1481 /* latency must be in 0.1us units. */
1482 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
1483 				   unsigned int htotal,
1484 				   unsigned int width,
1485 				   unsigned int cpp,
1486 				   unsigned int latency)
1487 {
1488 	unsigned int ret;
1489 
1490 	ret = intel_wm_method2(pixel_rate, htotal,
1491 			       width, cpp, latency);
1492 	ret = DIV_ROUND_UP(ret, 64);
1493 
1494 	return ret;
1495 }
1496 
1497 static void vlv_setup_wm_latency(struct intel_display *display)
1498 {
1499 	/* all latencies in usec */
1500 	display->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
1501 
1502 	display->wm.num_levels = VLV_WM_LEVEL_PM2 + 1;
1503 
1504 	if (display->platform.cherryview) {
1505 		display->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
1506 		display->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
1507 
1508 		display->wm.num_levels = VLV_WM_LEVEL_DDR_DVFS + 1;
1509 	}
1510 }
1511 
1512 static u16 vlv_compute_wm_level(const struct intel_crtc_state *crtc_state,
1513 				const struct intel_plane_state *plane_state,
1514 				int level)
1515 {
1516 	struct intel_display *display = to_intel_display(plane_state);
1517 	struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1518 	const struct drm_display_mode *pipe_mode =
1519 		&crtc_state->hw.pipe_mode;
1520 	unsigned int pixel_rate, htotal, cpp, width, wm;
1521 
1522 	if (display->wm.pri_latency[level] == 0)
1523 		return USHRT_MAX;
1524 
1525 	if (!intel_wm_plane_visible(crtc_state, plane_state))
1526 		return 0;
1527 
1528 	cpp = plane_state->hw.fb->format->cpp[0];
1529 	pixel_rate = crtc_state->pixel_rate;
1530 	htotal = pipe_mode->crtc_htotal;
1531 	width = drm_rect_width(&plane_state->uapi.src) >> 16;
1532 
1533 	if (plane->id == PLANE_CURSOR) {
1534 		/*
1535 		 * FIXME the formula gives values that are
1536 		 * too big for the cursor FIFO, and hence we
1537 		 * would never be able to use cursors. For
1538 		 * now just hardcode the watermark.
1539 		 */
1540 		wm = 63;
1541 	} else {
1542 		wm = vlv_wm_method2(pixel_rate, htotal, width, cpp,
1543 				    display->wm.pri_latency[level] * 10);
1544 	}
1545 
1546 	return min_t(unsigned int, wm, USHRT_MAX);
1547 }
1548 
1549 static bool vlv_need_sprite0_fifo_workaround(unsigned int active_planes)
1550 {
1551 	return (active_planes & (BIT(PLANE_SPRITE0) |
1552 				 BIT(PLANE_SPRITE1))) == BIT(PLANE_SPRITE1);
1553 }
1554 
1555 static int vlv_compute_fifo(struct intel_crtc_state *crtc_state)
1556 {
1557 	struct intel_display *display = to_intel_display(crtc_state);
1558 	struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1559 	const struct g4x_pipe_wm *raw =
1560 		&crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2];
1561 	struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
1562 	u8 active_planes = crtc_state->active_planes & ~BIT(PLANE_CURSOR);
1563 	int num_active_planes = hweight8(active_planes);
1564 	const int fifo_size = 511;
1565 	int fifo_extra, fifo_left = fifo_size;
1566 	int sprite0_fifo_extra = 0;
1567 	unsigned int total_rate;
1568 	enum plane_id plane_id;
1569 
1570 	/*
1571 	 * When enabling sprite0 after sprite1 has already been enabled
1572 	 * we tend to get an underrun unless sprite0 already has some
1573 	 * FIFO space allocated. Hence we always allocate at least one
1574 	 * cacheline for sprite0 whenever sprite1 is enabled.
1575 	 *
1576 	 * All other plane enable sequences appear immune to this problem.
1577 	 */
1578 	if (vlv_need_sprite0_fifo_workaround(active_planes))
1579 		sprite0_fifo_extra = 1;
1580 
1581 	total_rate = raw->plane[PLANE_PRIMARY] +
1582 		raw->plane[PLANE_SPRITE0] +
1583 		raw->plane[PLANE_SPRITE1] +
1584 		sprite0_fifo_extra;
1585 
1586 	if (total_rate > fifo_size)
1587 		return -EINVAL;
1588 
1589 	if (total_rate == 0)
1590 		total_rate = 1;
1591 
1592 	for_each_plane_id_on_crtc(crtc, plane_id) {
1593 		unsigned int rate;
1594 
1595 		if ((active_planes & BIT(plane_id)) == 0) {
1596 			fifo_state->plane[plane_id] = 0;
1597 			continue;
1598 		}
1599 
1600 		rate = raw->plane[plane_id];
1601 		fifo_state->plane[plane_id] = fifo_size * rate / total_rate;
1602 		fifo_left -= fifo_state->plane[plane_id];
1603 	}
1604 
1605 	fifo_state->plane[PLANE_SPRITE0] += sprite0_fifo_extra;
1606 	fifo_left -= sprite0_fifo_extra;
1607 
1608 	fifo_state->plane[PLANE_CURSOR] = 63;
1609 
1610 	fifo_extra = DIV_ROUND_UP(fifo_left, num_active_planes ?: 1);
1611 
1612 	/* spread the remainder evenly */
1613 	for_each_plane_id_on_crtc(crtc, plane_id) {
1614 		int plane_extra;
1615 
1616 		if (fifo_left == 0)
1617 			break;
1618 
1619 		if ((active_planes & BIT(plane_id)) == 0)
1620 			continue;
1621 
1622 		plane_extra = min(fifo_extra, fifo_left);
1623 		fifo_state->plane[plane_id] += plane_extra;
1624 		fifo_left -= plane_extra;
1625 	}
1626 
1627 	drm_WARN_ON(display->drm, active_planes != 0 && fifo_left != 0);
1628 
1629 	/* give it all to the first plane if none are active */
1630 	if (active_planes == 0) {
1631 		drm_WARN_ON(display->drm, fifo_left != fifo_size);
1632 		fifo_state->plane[PLANE_PRIMARY] = fifo_left;
1633 	}
1634 
1635 	return 0;
1636 }
1637 
1638 /* mark all levels starting from 'level' as invalid */
1639 static void vlv_invalidate_wms(struct intel_crtc *crtc,
1640 			       struct vlv_wm_state *wm_state, int level)
1641 {
1642 	struct intel_display *display = to_intel_display(crtc);
1643 
1644 	for (; level < display->wm.num_levels; level++) {
1645 		enum plane_id plane_id;
1646 
1647 		for_each_plane_id_on_crtc(crtc, plane_id)
1648 			wm_state->wm[level].plane[plane_id] = USHRT_MAX;
1649 
1650 		wm_state->sr[level].cursor = USHRT_MAX;
1651 		wm_state->sr[level].plane = USHRT_MAX;
1652 	}
1653 }
1654 
1655 static u16 vlv_invert_wm_value(u16 wm, u16 fifo_size)
1656 {
1657 	if (wm > fifo_size)
1658 		return USHRT_MAX;
1659 	else
1660 		return fifo_size - wm;
1661 }
1662 
1663 /*
1664  * Starting from 'level' set all higher
1665  * levels to 'value' in the "raw" watermarks.
1666  */
1667 static bool vlv_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1668 				 int level, enum plane_id plane_id, u16 value)
1669 {
1670 	struct intel_display *display = to_intel_display(crtc_state);
1671 	bool dirty = false;
1672 
1673 	for (; level < display->wm.num_levels; level++) {
1674 		struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1675 
1676 		dirty |= raw->plane[plane_id] != value;
1677 		raw->plane[plane_id] = value;
1678 	}
1679 
1680 	return dirty;
1681 }
1682 
1683 static bool vlv_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1684 				     const struct intel_plane_state *plane_state)
1685 {
1686 	struct intel_display *display = to_intel_display(crtc_state);
1687 	struct intel_plane *plane = to_intel_plane(plane_state->uapi.plane);
1688 	enum plane_id plane_id = plane->id;
1689 	int level;
1690 	bool dirty = false;
1691 
1692 	if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1693 		dirty |= vlv_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1694 		goto out;
1695 	}
1696 
1697 	for (level = 0; level < display->wm.num_levels; level++) {
1698 		struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1699 		int wm = vlv_compute_wm_level(crtc_state, plane_state, level);
1700 		int max_wm = plane_id == PLANE_CURSOR ? 63 : 511;
1701 
1702 		if (wm > max_wm)
1703 			break;
1704 
1705 		dirty |= raw->plane[plane_id] != wm;
1706 		raw->plane[plane_id] = wm;
1707 	}
1708 
1709 	/* mark all higher levels as invalid */
1710 	dirty |= vlv_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1711 
1712 out:
1713 	if (dirty)
1714 		drm_dbg_kms(display->drm,
1715 			    "%s watermarks: PM2=%d, PM5=%d, DDR DVFS=%d\n",
1716 			    plane->base.name,
1717 			    crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2].plane[plane_id],
1718 			    crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM5].plane[plane_id],
1719 			    crtc_state->wm.vlv.raw[VLV_WM_LEVEL_DDR_DVFS].plane[plane_id]);
1720 
1721 	return dirty;
1722 }
1723 
1724 static bool vlv_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1725 				      enum plane_id plane_id, int level)
1726 {
1727 	const struct g4x_pipe_wm *raw =
1728 		&crtc_state->wm.vlv.raw[level];
1729 	const struct vlv_fifo_state *fifo_state =
1730 		&crtc_state->wm.vlv.fifo_state;
1731 
1732 	return raw->plane[plane_id] <= fifo_state->plane[plane_id];
1733 }
1734 
1735 static bool vlv_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state, int level)
1736 {
1737 	return vlv_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1738 		vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1739 		vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE1, level) &&
1740 		vlv_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1741 }
1742 
1743 static int _vlv_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1744 {
1745 	struct intel_display *display = to_intel_display(crtc_state);
1746 	struct intel_crtc *crtc = to_intel_crtc(crtc_state->uapi.crtc);
1747 	struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
1748 	const struct vlv_fifo_state *fifo_state =
1749 		&crtc_state->wm.vlv.fifo_state;
1750 	u8 active_planes = crtc_state->active_planes & ~BIT(PLANE_CURSOR);
1751 	int num_active_planes = hweight8(active_planes);
1752 	enum plane_id plane_id;
1753 	int level;
1754 
1755 	/* initially allow all levels */
1756 	wm_state->num_levels = display->wm.num_levels;
1757 	/*
1758 	 * Note that enabling cxsr with no primary/sprite planes
1759 	 * enabled can wedge the pipe. Hence we only allow cxsr
1760 	 * with exactly one enabled primary/sprite plane.
1761 	 */
1762 	wm_state->cxsr = crtc->pipe != PIPE_C && num_active_planes == 1;
1763 
1764 	for (level = 0; level < wm_state->num_levels; level++) {
1765 		const struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1766 		const int sr_fifo_size = INTEL_NUM_PIPES(display) * 512 - 1;
1767 
1768 		if (!vlv_raw_crtc_wm_is_valid(crtc_state, level))
1769 			break;
1770 
1771 		for_each_plane_id_on_crtc(crtc, plane_id) {
1772 			wm_state->wm[level].plane[plane_id] =
1773 				vlv_invert_wm_value(raw->plane[plane_id],
1774 						    fifo_state->plane[plane_id]);
1775 		}
1776 
1777 		wm_state->sr[level].plane =
1778 			vlv_invert_wm_value(max3(raw->plane[PLANE_PRIMARY],
1779 						 raw->plane[PLANE_SPRITE0],
1780 						 raw->plane[PLANE_SPRITE1]),
1781 					    sr_fifo_size);
1782 
1783 		wm_state->sr[level].cursor =
1784 			vlv_invert_wm_value(raw->plane[PLANE_CURSOR],
1785 					    63);
1786 	}
1787 
1788 	if (level == 0)
1789 		return -EINVAL;
1790 
1791 	/* limit to only levels we can actually handle */
1792 	wm_state->num_levels = level;
1793 
1794 	/* invalidate the higher levels */
1795 	vlv_invalidate_wms(crtc, wm_state, level);
1796 
1797 	return 0;
1798 }
1799 
1800 static int vlv_compute_pipe_wm(struct intel_atomic_state *state,
1801 			       struct intel_crtc *crtc)
1802 {
1803 	struct intel_crtc_state *crtc_state =
1804 		intel_atomic_get_new_crtc_state(state, crtc);
1805 	const struct intel_plane_state *old_plane_state;
1806 	const struct intel_plane_state *new_plane_state;
1807 	struct intel_plane *plane;
1808 	unsigned int dirty = 0;
1809 	int i;
1810 
1811 	for_each_oldnew_intel_plane_in_state(state, plane,
1812 					     old_plane_state,
1813 					     new_plane_state, i) {
1814 		if (new_plane_state->hw.crtc != &crtc->base &&
1815 		    old_plane_state->hw.crtc != &crtc->base)
1816 			continue;
1817 
1818 		if (vlv_raw_plane_wm_compute(crtc_state, new_plane_state))
1819 			dirty |= BIT(plane->id);
1820 	}
1821 
1822 	/*
1823 	 * DSPARB registers may have been reset due to the
1824 	 * power well being turned off. Make sure we restore
1825 	 * them to a consistent state even if no primary/sprite
1826 	 * planes are initially active. We also force a FIFO
1827 	 * recomputation so that we are sure to sanitize the
1828 	 * FIFO setting we took over from the BIOS even if there
1829 	 * are no active planes on the crtc.
1830 	 */
1831 	if (intel_crtc_needs_modeset(crtc_state))
1832 		dirty = ~0;
1833 
1834 	if (!dirty)
1835 		return 0;
1836 
1837 	/* cursor changes don't warrant a FIFO recompute */
1838 	if (dirty & ~BIT(PLANE_CURSOR)) {
1839 		const struct intel_crtc_state *old_crtc_state =
1840 			intel_atomic_get_old_crtc_state(state, crtc);
1841 		const struct vlv_fifo_state *old_fifo_state =
1842 			&old_crtc_state->wm.vlv.fifo_state;
1843 		const struct vlv_fifo_state *new_fifo_state =
1844 			&crtc_state->wm.vlv.fifo_state;
1845 		int ret;
1846 
1847 		ret = vlv_compute_fifo(crtc_state);
1848 		if (ret)
1849 			return ret;
1850 
1851 		if (intel_crtc_needs_modeset(crtc_state) ||
1852 		    memcmp(old_fifo_state, new_fifo_state,
1853 			   sizeof(*new_fifo_state)) != 0)
1854 			crtc_state->fifo_changed = true;
1855 	}
1856 
1857 	return _vlv_compute_pipe_wm(crtc_state);
1858 }
1859 
1860 #define VLV_FIFO(plane, value) \
1861 	(((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1862 
1863 static void vlv_atomic_update_fifo(struct intel_atomic_state *state,
1864 				   struct intel_crtc *crtc)
1865 {
1866 	struct intel_display *display = to_intel_display(crtc);
1867 	struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1868 	struct intel_uncore *uncore = &dev_priv->uncore;
1869 	const struct intel_crtc_state *crtc_state =
1870 		intel_atomic_get_new_crtc_state(state, crtc);
1871 	const struct vlv_fifo_state *fifo_state =
1872 		&crtc_state->wm.vlv.fifo_state;
1873 	int sprite0_start, sprite1_start, fifo_size;
1874 	u32 dsparb, dsparb2, dsparb3;
1875 
1876 	if (!crtc_state->fifo_changed)
1877 		return;
1878 
1879 	sprite0_start = fifo_state->plane[PLANE_PRIMARY];
1880 	sprite1_start = fifo_state->plane[PLANE_SPRITE0] + sprite0_start;
1881 	fifo_size = fifo_state->plane[PLANE_SPRITE1] + sprite1_start;
1882 
1883 	drm_WARN_ON(display->drm, fifo_state->plane[PLANE_CURSOR] != 63);
1884 	drm_WARN_ON(display->drm, fifo_size != 511);
1885 
1886 	trace_vlv_fifo_size(crtc, sprite0_start, sprite1_start, fifo_size);
1887 
1888 	/*
1889 	 * uncore.lock serves a double purpose here. It allows us to
1890 	 * use the less expensive I915_{READ,WRITE}_FW() functions, and
1891 	 * it protects the DSPARB registers from getting clobbered by
1892 	 * parallel updates from multiple pipes.
1893 	 *
1894 	 * intel_pipe_update_start() has already disabled interrupts
1895 	 * for us, so a plain spin_lock() is sufficient here.
1896 	 */
1897 	spin_lock(&uncore->lock);
1898 
1899 	switch (crtc->pipe) {
1900 	case PIPE_A:
1901 		dsparb = intel_de_read_fw(display, DSPARB(display));
1902 		dsparb2 = intel_de_read_fw(display, DSPARB2);
1903 
1904 		dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1905 			    VLV_FIFO(SPRITEB, 0xff));
1906 		dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1907 			   VLV_FIFO(SPRITEB, sprite1_start));
1908 
1909 		dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1910 			     VLV_FIFO(SPRITEB_HI, 0x1));
1911 		dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1912 			   VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1913 
1914 		intel_de_write_fw(display, DSPARB(display), dsparb);
1915 		intel_de_write_fw(display, DSPARB2, dsparb2);
1916 		break;
1917 	case PIPE_B:
1918 		dsparb = intel_de_read_fw(display, DSPARB(display));
1919 		dsparb2 = intel_de_read_fw(display, DSPARB2);
1920 
1921 		dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1922 			    VLV_FIFO(SPRITED, 0xff));
1923 		dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1924 			   VLV_FIFO(SPRITED, sprite1_start));
1925 
1926 		dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1927 			     VLV_FIFO(SPRITED_HI, 0xff));
1928 		dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1929 			   VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1930 
1931 		intel_de_write_fw(display, DSPARB(display), dsparb);
1932 		intel_de_write_fw(display, DSPARB2, dsparb2);
1933 		break;
1934 	case PIPE_C:
1935 		dsparb3 = intel_de_read_fw(display, DSPARB3);
1936 		dsparb2 = intel_de_read_fw(display, DSPARB2);
1937 
1938 		dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
1939 			     VLV_FIFO(SPRITEF, 0xff));
1940 		dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
1941 			    VLV_FIFO(SPRITEF, sprite1_start));
1942 
1943 		dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
1944 			     VLV_FIFO(SPRITEF_HI, 0xff));
1945 		dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
1946 			   VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
1947 
1948 		intel_de_write_fw(display, DSPARB3, dsparb3);
1949 		intel_de_write_fw(display, DSPARB2, dsparb2);
1950 		break;
1951 	default:
1952 		break;
1953 	}
1954 
1955 	intel_de_read_fw(display, DSPARB(display));
1956 
1957 	spin_unlock(&uncore->lock);
1958 }
1959 
1960 #undef VLV_FIFO
1961 
1962 static int vlv_compute_intermediate_wm(struct intel_atomic_state *state,
1963 				       struct intel_crtc *crtc)
1964 {
1965 	struct intel_crtc_state *new_crtc_state =
1966 		intel_atomic_get_new_crtc_state(state, crtc);
1967 	const struct intel_crtc_state *old_crtc_state =
1968 		intel_atomic_get_old_crtc_state(state, crtc);
1969 	struct vlv_wm_state *intermediate = &new_crtc_state->wm.vlv.intermediate;
1970 	const struct vlv_wm_state *optimal = &new_crtc_state->wm.vlv.optimal;
1971 	const struct vlv_wm_state *active = &old_crtc_state->wm.vlv.optimal;
1972 	int level;
1973 
1974 	if (!new_crtc_state->hw.active ||
1975 	    intel_crtc_needs_modeset(new_crtc_state)) {
1976 		*intermediate = *optimal;
1977 
1978 		intermediate->cxsr = false;
1979 		goto out;
1980 	}
1981 
1982 	intermediate->num_levels = min(optimal->num_levels, active->num_levels);
1983 	intermediate->cxsr = optimal->cxsr && active->cxsr &&
1984 		!new_crtc_state->disable_cxsr;
1985 
1986 	for (level = 0; level < intermediate->num_levels; level++) {
1987 		enum plane_id plane_id;
1988 
1989 		for_each_plane_id_on_crtc(crtc, plane_id) {
1990 			intermediate->wm[level].plane[plane_id] =
1991 				min(optimal->wm[level].plane[plane_id],
1992 				    active->wm[level].plane[plane_id]);
1993 		}
1994 
1995 		intermediate->sr[level].plane = min(optimal->sr[level].plane,
1996 						    active->sr[level].plane);
1997 		intermediate->sr[level].cursor = min(optimal->sr[level].cursor,
1998 						     active->sr[level].cursor);
1999 	}
2000 
2001 	vlv_invalidate_wms(crtc, intermediate, level);
2002 
2003 out:
2004 	/*
2005 	 * If our intermediate WM are identical to the final WM, then we can
2006 	 * omit the post-vblank programming; only update if it's different.
2007 	 */
2008 	if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
2009 		new_crtc_state->wm.need_postvbl_update = true;
2010 
2011 	return 0;
2012 }
2013 
2014 static int vlv_compute_watermarks(struct intel_atomic_state *state,
2015 				  struct intel_crtc *crtc)
2016 {
2017 	int ret;
2018 
2019 	ret = vlv_compute_pipe_wm(state, crtc);
2020 	if (ret)
2021 		return ret;
2022 
2023 	ret = vlv_compute_intermediate_wm(state, crtc);
2024 	if (ret)
2025 		return ret;
2026 
2027 	return 0;
2028 }
2029 
2030 static void vlv_merge_wm(struct intel_display *display,
2031 			 struct vlv_wm_values *wm)
2032 {
2033 	struct intel_crtc *crtc;
2034 	int num_active_pipes = 0;
2035 
2036 	wm->level = display->wm.num_levels - 1;
2037 	wm->cxsr = true;
2038 
2039 	for_each_intel_crtc(display->drm, crtc) {
2040 		const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2041 
2042 		if (!crtc->active)
2043 			continue;
2044 
2045 		if (!wm_state->cxsr)
2046 			wm->cxsr = false;
2047 
2048 		num_active_pipes++;
2049 		wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
2050 	}
2051 
2052 	if (num_active_pipes != 1)
2053 		wm->cxsr = false;
2054 
2055 	if (num_active_pipes > 1)
2056 		wm->level = VLV_WM_LEVEL_PM2;
2057 
2058 	for_each_intel_crtc(display->drm, crtc) {
2059 		const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2060 		enum pipe pipe = crtc->pipe;
2061 
2062 		wm->pipe[pipe] = wm_state->wm[wm->level];
2063 		if (crtc->active && wm->cxsr)
2064 			wm->sr = wm_state->sr[wm->level];
2065 
2066 		wm->ddl[pipe].plane[PLANE_PRIMARY] = DDL_PRECISION_HIGH | 2;
2067 		wm->ddl[pipe].plane[PLANE_SPRITE0] = DDL_PRECISION_HIGH | 2;
2068 		wm->ddl[pipe].plane[PLANE_SPRITE1] = DDL_PRECISION_HIGH | 2;
2069 		wm->ddl[pipe].plane[PLANE_CURSOR] = DDL_PRECISION_HIGH | 2;
2070 	}
2071 }
2072 
2073 static void vlv_program_watermarks(struct intel_display *display)
2074 {
2075 	struct vlv_wm_values *old_wm = &display->wm.vlv;
2076 	struct vlv_wm_values new_wm = {};
2077 
2078 	vlv_merge_wm(display, &new_wm);
2079 
2080 	if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
2081 		return;
2082 
2083 	if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2084 		chv_set_memory_dvfs(display, false);
2085 
2086 	if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2087 		chv_set_memory_pm5(display, false);
2088 
2089 	if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
2090 		_intel_set_memory_cxsr(display, false);
2091 
2092 	vlv_write_wm_values(display, &new_wm);
2093 
2094 	if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
2095 		_intel_set_memory_cxsr(display, true);
2096 
2097 	if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2098 		chv_set_memory_pm5(display, true);
2099 
2100 	if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2101 		chv_set_memory_dvfs(display, true);
2102 
2103 	*old_wm = new_wm;
2104 }
2105 
2106 static void vlv_initial_watermarks(struct intel_atomic_state *state,
2107 				   struct intel_crtc *crtc)
2108 {
2109 	struct intel_display *display = to_intel_display(crtc);
2110 	const struct intel_crtc_state *crtc_state =
2111 		intel_atomic_get_new_crtc_state(state, crtc);
2112 
2113 	mutex_lock(&display->wm.wm_mutex);
2114 	crtc->wm.active.vlv = crtc_state->wm.vlv.intermediate;
2115 	vlv_program_watermarks(display);
2116 	mutex_unlock(&display->wm.wm_mutex);
2117 }
2118 
2119 static void vlv_optimize_watermarks(struct intel_atomic_state *state,
2120 				    struct intel_crtc *crtc)
2121 {
2122 	struct intel_display *display = to_intel_display(crtc);
2123 	const struct intel_crtc_state *crtc_state =
2124 		intel_atomic_get_new_crtc_state(state, crtc);
2125 
2126 	if (!crtc_state->wm.need_postvbl_update)
2127 		return;
2128 
2129 	mutex_lock(&display->wm.wm_mutex);
2130 	crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
2131 	vlv_program_watermarks(display);
2132 	mutex_unlock(&display->wm.wm_mutex);
2133 }
2134 
2135 static void i965_update_wm(struct intel_display *display)
2136 {
2137 	struct intel_crtc *crtc;
2138 	int srwm = 1;
2139 	int cursor_sr = 16;
2140 	bool cxsr_enabled;
2141 
2142 	/* Calc sr entries for one plane configs */
2143 	crtc = single_enabled_crtc(display);
2144 	if (crtc) {
2145 		/* self-refresh has much higher latency */
2146 		static const int sr_latency_ns = 12000;
2147 		const struct drm_display_mode *pipe_mode =
2148 			&crtc->config->hw.pipe_mode;
2149 		const struct drm_framebuffer *fb =
2150 			crtc->base.primary->state->fb;
2151 		int pixel_rate = crtc->config->pixel_rate;
2152 		int htotal = pipe_mode->crtc_htotal;
2153 		int width = drm_rect_width(&crtc->base.primary->state->src) >> 16;
2154 		int cpp = fb->format->cpp[0];
2155 		int entries;
2156 
2157 		entries = intel_wm_method2(pixel_rate, htotal,
2158 					   width, cpp, sr_latency_ns / 100);
2159 		entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
2160 		srwm = I965_FIFO_SIZE - entries;
2161 		if (srwm < 0)
2162 			srwm = 1;
2163 		srwm &= 0x1ff;
2164 		drm_dbg_kms(display->drm,
2165 			    "self-refresh entries: %d, wm: %d\n",
2166 			    entries, srwm);
2167 
2168 		entries = intel_wm_method2(pixel_rate, htotal,
2169 					   crtc->base.cursor->state->crtc_w, 4,
2170 					   sr_latency_ns / 100);
2171 		entries = DIV_ROUND_UP(entries,
2172 				       i965_cursor_wm_info.cacheline_size) +
2173 			i965_cursor_wm_info.guard_size;
2174 
2175 		cursor_sr = i965_cursor_wm_info.fifo_size - entries;
2176 		if (cursor_sr > i965_cursor_wm_info.max_wm)
2177 			cursor_sr = i965_cursor_wm_info.max_wm;
2178 
2179 		drm_dbg_kms(display->drm,
2180 			    "self-refresh watermark: display plane %d "
2181 			    "cursor %d\n", srwm, cursor_sr);
2182 
2183 		cxsr_enabled = true;
2184 	} else {
2185 		cxsr_enabled = false;
2186 		/* Turn off self refresh if both pipes are enabled */
2187 		intel_set_memory_cxsr(display, false);
2188 	}
2189 
2190 	drm_dbg_kms(display->drm,
2191 		    "Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
2192 		    srwm);
2193 
2194 	/* 965 has limitations... */
2195 	intel_de_write(display, DSPFW1(display),
2196 		       FW_WM(srwm, SR) |
2197 		       FW_WM(8, CURSORB) |
2198 		       FW_WM(8, PLANEB) |
2199 		       FW_WM(8, PLANEA));
2200 	intel_de_write(display, DSPFW2(display),
2201 		       FW_WM(8, CURSORA) |
2202 		       FW_WM(8, PLANEC_OLD));
2203 	/* update cursor SR watermark */
2204 	intel_de_write(display, DSPFW3(display),
2205 		       FW_WM(cursor_sr, CURSOR_SR));
2206 
2207 	if (cxsr_enabled)
2208 		intel_set_memory_cxsr(display, true);
2209 }
2210 
2211 #undef FW_WM
2212 
2213 static struct intel_crtc *intel_crtc_for_plane(struct intel_display *display,
2214 					       enum i9xx_plane_id i9xx_plane)
2215 {
2216 	struct intel_plane *plane;
2217 
2218 	for_each_intel_plane(display->drm, plane) {
2219 		if (plane->id == PLANE_PRIMARY &&
2220 		    plane->i9xx_plane == i9xx_plane)
2221 			return intel_crtc_for_pipe(display, plane->pipe);
2222 	}
2223 
2224 	return NULL;
2225 }
2226 
2227 static void i9xx_update_wm(struct intel_display *display)
2228 {
2229 	const struct intel_watermark_params *wm_info;
2230 	u32 fwater_lo;
2231 	u32 fwater_hi;
2232 	int cwm, srwm = 1;
2233 	int fifo_size;
2234 	int planea_wm, planeb_wm;
2235 	struct intel_crtc *crtc;
2236 
2237 	if (display->platform.i945gm)
2238 		wm_info = &i945_wm_info;
2239 	else if (DISPLAY_VER(display) != 2)
2240 		wm_info = &i915_wm_info;
2241 	else
2242 		wm_info = &i830_a_wm_info;
2243 
2244 	if (DISPLAY_VER(display) == 2)
2245 		fifo_size = i830_get_fifo_size(display, PLANE_A);
2246 	else
2247 		fifo_size = i9xx_get_fifo_size(display, PLANE_A);
2248 	crtc = intel_crtc_for_plane(display, PLANE_A);
2249 	if (intel_crtc_active(crtc)) {
2250 		const struct drm_framebuffer *fb =
2251 			crtc->base.primary->state->fb;
2252 		int cpp;
2253 
2254 		if (DISPLAY_VER(display) == 2)
2255 			cpp = 4;
2256 		else
2257 			cpp = fb->format->cpp[0];
2258 
2259 		planea_wm = intel_calculate_wm(display, crtc->config->pixel_rate,
2260 					       wm_info, fifo_size, cpp,
2261 					       pessimal_latency_ns);
2262 	} else {
2263 		planea_wm = fifo_size - wm_info->guard_size;
2264 		if (planea_wm > (long)wm_info->max_wm)
2265 			planea_wm = wm_info->max_wm;
2266 	}
2267 
2268 	if (DISPLAY_VER(display) == 2)
2269 		wm_info = &i830_bc_wm_info;
2270 
2271 	if (DISPLAY_VER(display) == 2)
2272 		fifo_size = i830_get_fifo_size(display, PLANE_B);
2273 	else
2274 		fifo_size = i9xx_get_fifo_size(display, PLANE_B);
2275 	crtc = intel_crtc_for_plane(display, PLANE_B);
2276 	if (intel_crtc_active(crtc)) {
2277 		const struct drm_framebuffer *fb =
2278 			crtc->base.primary->state->fb;
2279 		int cpp;
2280 
2281 		if (DISPLAY_VER(display) == 2)
2282 			cpp = 4;
2283 		else
2284 			cpp = fb->format->cpp[0];
2285 
2286 		planeb_wm = intel_calculate_wm(display, crtc->config->pixel_rate,
2287 					       wm_info, fifo_size, cpp,
2288 					       pessimal_latency_ns);
2289 	} else {
2290 		planeb_wm = fifo_size - wm_info->guard_size;
2291 		if (planeb_wm > (long)wm_info->max_wm)
2292 			planeb_wm = wm_info->max_wm;
2293 	}
2294 
2295 	drm_dbg_kms(display->drm,
2296 		    "FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
2297 
2298 	crtc = single_enabled_crtc(display);
2299 	if (display->platform.i915gm && crtc) {
2300 		const struct drm_framebuffer *fb =
2301 			crtc->base.primary->state->fb;
2302 
2303 		/* self-refresh seems busted with untiled */
2304 		if (fb->modifier == DRM_FORMAT_MOD_LINEAR)
2305 			crtc = NULL;
2306 	}
2307 
2308 	/*
2309 	 * Overlay gets an aggressive default since video jitter is bad.
2310 	 */
2311 	cwm = 2;
2312 
2313 	/* Play safe and disable self-refresh before adjusting watermarks. */
2314 	intel_set_memory_cxsr(display, false);
2315 
2316 	/* Calc sr entries for one plane configs */
2317 	if (HAS_FW_BLC(display) && crtc) {
2318 		/* self-refresh has much higher latency */
2319 		static const int sr_latency_ns = 6000;
2320 		const struct drm_display_mode *pipe_mode =
2321 			&crtc->config->hw.pipe_mode;
2322 		const struct drm_framebuffer *fb =
2323 			crtc->base.primary->state->fb;
2324 		int pixel_rate = crtc->config->pixel_rate;
2325 		int htotal = pipe_mode->crtc_htotal;
2326 		int width = drm_rect_width(&crtc->base.primary->state->src) >> 16;
2327 		int cpp;
2328 		int entries;
2329 
2330 		if (display->platform.i915gm || display->platform.i945gm)
2331 			cpp = 4;
2332 		else
2333 			cpp = fb->format->cpp[0];
2334 
2335 		entries = intel_wm_method2(pixel_rate, htotal, width, cpp,
2336 					   sr_latency_ns / 100);
2337 		entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
2338 		drm_dbg_kms(display->drm,
2339 			    "self-refresh entries: %d\n", entries);
2340 		srwm = wm_info->fifo_size - entries;
2341 		if (srwm < 0)
2342 			srwm = 1;
2343 
2344 		if (display->platform.i945g || display->platform.i945gm)
2345 			intel_de_write(display, FW_BLC_SELF,
2346 				       FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
2347 		else
2348 			intel_de_write(display, FW_BLC_SELF, srwm & 0x3f);
2349 	}
2350 
2351 	drm_dbg_kms(display->drm,
2352 		    "Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
2353 		     planea_wm, planeb_wm, cwm, srwm);
2354 
2355 	fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
2356 	fwater_hi = (cwm & 0x1f);
2357 
2358 	/* Set request length to 8 cachelines per fetch */
2359 	fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
2360 	fwater_hi = fwater_hi | (1 << 8);
2361 
2362 	intel_de_write(display, FW_BLC, fwater_lo);
2363 	intel_de_write(display, FW_BLC2, fwater_hi);
2364 
2365 	if (crtc)
2366 		intel_set_memory_cxsr(display, true);
2367 }
2368 
2369 static void i845_update_wm(struct intel_display *display)
2370 {
2371 	struct intel_crtc *crtc;
2372 	u32 fwater_lo;
2373 	int planea_wm;
2374 
2375 	crtc = single_enabled_crtc(display);
2376 	if (crtc == NULL)
2377 		return;
2378 
2379 	planea_wm = intel_calculate_wm(display, crtc->config->pixel_rate,
2380 				       &i845_wm_info,
2381 				       i845_get_fifo_size(display, PLANE_A),
2382 				       4, pessimal_latency_ns);
2383 	fwater_lo = intel_de_read(display, FW_BLC) & ~0xfff;
2384 	fwater_lo |= (3<<8) | planea_wm;
2385 
2386 	drm_dbg_kms(display->drm,
2387 		    "Setting FIFO watermarks - A: %d\n", planea_wm);
2388 
2389 	intel_de_write(display, FW_BLC, fwater_lo);
2390 }
2391 
2392 /* latency must be in 0.1us units. */
2393 static unsigned int ilk_wm_method1(unsigned int pixel_rate,
2394 				   unsigned int cpp,
2395 				   unsigned int latency)
2396 {
2397 	unsigned int ret;
2398 
2399 	ret = intel_wm_method1(pixel_rate, cpp, latency);
2400 	ret = DIV_ROUND_UP(ret, 64) + 2;
2401 
2402 	return ret;
2403 }
2404 
2405 /* latency must be in 0.1us units. */
2406 static unsigned int ilk_wm_method2(unsigned int pixel_rate,
2407 				   unsigned int htotal,
2408 				   unsigned int width,
2409 				   unsigned int cpp,
2410 				   unsigned int latency)
2411 {
2412 	unsigned int ret;
2413 
2414 	ret = intel_wm_method2(pixel_rate, htotal,
2415 			       width, cpp, latency);
2416 	ret = DIV_ROUND_UP(ret, 64) + 2;
2417 
2418 	return ret;
2419 }
2420 
2421 static u32 ilk_wm_fbc(u32 pri_val, u32 horiz_pixels, u8 cpp)
2422 {
2423 	/*
2424 	 * Neither of these should be possible since this function shouldn't be
2425 	 * called if the CRTC is off or the plane is invisible.  But let's be
2426 	 * extra paranoid to avoid a potential divide-by-zero if we screw up
2427 	 * elsewhere in the driver.
2428 	 */
2429 	if (WARN_ON(!cpp))
2430 		return 0;
2431 	if (WARN_ON(!horiz_pixels))
2432 		return 0;
2433 
2434 	return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
2435 }
2436 
2437 struct ilk_wm_maximums {
2438 	u16 pri;
2439 	u16 spr;
2440 	u16 cur;
2441 	u16 fbc;
2442 };
2443 
2444 /*
2445  * For both WM_PIPE and WM_LP.
2446  * mem_value must be in 0.1us units.
2447  */
2448 static u32 ilk_compute_pri_wm(const struct intel_crtc_state *crtc_state,
2449 			      const struct intel_plane_state *plane_state,
2450 			      u32 mem_value, bool is_lp)
2451 {
2452 	u32 method1, method2;
2453 	int cpp;
2454 
2455 	if (mem_value == 0)
2456 		return U32_MAX;
2457 
2458 	if (!intel_wm_plane_visible(crtc_state, plane_state))
2459 		return 0;
2460 
2461 	cpp = plane_state->hw.fb->format->cpp[0];
2462 
2463 	method1 = ilk_wm_method1(crtc_state->pixel_rate, cpp, mem_value);
2464 
2465 	if (!is_lp)
2466 		return method1;
2467 
2468 	method2 = ilk_wm_method2(crtc_state->pixel_rate,
2469 				 crtc_state->hw.pipe_mode.crtc_htotal,
2470 				 drm_rect_width(&plane_state->uapi.src) >> 16,
2471 				 cpp, mem_value);
2472 
2473 	return min(method1, method2);
2474 }
2475 
2476 /*
2477  * For both WM_PIPE and WM_LP.
2478  * mem_value must be in 0.1us units.
2479  */
2480 static u32 ilk_compute_spr_wm(const struct intel_crtc_state *crtc_state,
2481 			      const struct intel_plane_state *plane_state,
2482 			      u32 mem_value)
2483 {
2484 	u32 method1, method2;
2485 	int cpp;
2486 
2487 	if (mem_value == 0)
2488 		return U32_MAX;
2489 
2490 	if (!intel_wm_plane_visible(crtc_state, plane_state))
2491 		return 0;
2492 
2493 	cpp = plane_state->hw.fb->format->cpp[0];
2494 
2495 	method1 = ilk_wm_method1(crtc_state->pixel_rate, cpp, mem_value);
2496 	method2 = ilk_wm_method2(crtc_state->pixel_rate,
2497 				 crtc_state->hw.pipe_mode.crtc_htotal,
2498 				 drm_rect_width(&plane_state->uapi.src) >> 16,
2499 				 cpp, mem_value);
2500 	return min(method1, method2);
2501 }
2502 
2503 /*
2504  * For both WM_PIPE and WM_LP.
2505  * mem_value must be in 0.1us units.
2506  */
2507 static u32 ilk_compute_cur_wm(const struct intel_crtc_state *crtc_state,
2508 			      const struct intel_plane_state *plane_state,
2509 			      u32 mem_value)
2510 {
2511 	int cpp;
2512 
2513 	if (mem_value == 0)
2514 		return U32_MAX;
2515 
2516 	if (!intel_wm_plane_visible(crtc_state, plane_state))
2517 		return 0;
2518 
2519 	cpp = plane_state->hw.fb->format->cpp[0];
2520 
2521 	return ilk_wm_method2(crtc_state->pixel_rate,
2522 			      crtc_state->hw.pipe_mode.crtc_htotal,
2523 			      drm_rect_width(&plane_state->uapi.src) >> 16,
2524 			      cpp, mem_value);
2525 }
2526 
2527 /* Only for WM_LP. */
2528 static u32 ilk_compute_fbc_wm(const struct intel_crtc_state *crtc_state,
2529 			      const struct intel_plane_state *plane_state,
2530 			      u32 pri_val)
2531 {
2532 	int cpp;
2533 
2534 	if (!intel_wm_plane_visible(crtc_state, plane_state))
2535 		return 0;
2536 
2537 	cpp = plane_state->hw.fb->format->cpp[0];
2538 
2539 	return ilk_wm_fbc(pri_val, drm_rect_width(&plane_state->uapi.src) >> 16,
2540 			  cpp);
2541 }
2542 
2543 static unsigned int
2544 ilk_display_fifo_size(struct intel_display *display)
2545 {
2546 	if (DISPLAY_VER(display) >= 8)
2547 		return 3072;
2548 	else if (DISPLAY_VER(display) >= 7)
2549 		return 768;
2550 	else
2551 		return 512;
2552 }
2553 
2554 static unsigned int
2555 ilk_plane_wm_reg_max(struct intel_display *display,
2556 		     int level, bool is_sprite)
2557 {
2558 	if (DISPLAY_VER(display) >= 8)
2559 		/* BDW primary/sprite plane watermarks */
2560 		return level == 0 ? 255 : 2047;
2561 	else if (DISPLAY_VER(display) >= 7)
2562 		/* IVB/HSW primary/sprite plane watermarks */
2563 		return level == 0 ? 127 : 1023;
2564 	else if (!is_sprite)
2565 		/* ILK/SNB primary plane watermarks */
2566 		return level == 0 ? 127 : 511;
2567 	else
2568 		/* ILK/SNB sprite plane watermarks */
2569 		return level == 0 ? 63 : 255;
2570 }
2571 
2572 static unsigned int
2573 ilk_cursor_wm_reg_max(struct intel_display *display, int level)
2574 {
2575 	if (DISPLAY_VER(display) >= 7)
2576 		return level == 0 ? 63 : 255;
2577 	else
2578 		return level == 0 ? 31 : 63;
2579 }
2580 
2581 static unsigned int ilk_fbc_wm_reg_max(struct intel_display *display)
2582 {
2583 	if (DISPLAY_VER(display) >= 8)
2584 		return 31;
2585 	else
2586 		return 15;
2587 }
2588 
2589 /* Calculate the maximum primary/sprite plane watermark */
2590 static unsigned int ilk_plane_wm_max(struct intel_display *display,
2591 				     int level,
2592 				     const struct intel_wm_config *config,
2593 				     enum intel_ddb_partitioning ddb_partitioning,
2594 				     bool is_sprite)
2595 {
2596 	unsigned int fifo_size = ilk_display_fifo_size(display);
2597 
2598 	/* if sprites aren't enabled, sprites get nothing */
2599 	if (is_sprite && !config->sprites_enabled)
2600 		return 0;
2601 
2602 	/* HSW allows LP1+ watermarks even with multiple pipes */
2603 	if (level == 0 || config->num_pipes_active > 1) {
2604 		fifo_size /= INTEL_NUM_PIPES(display);
2605 
2606 		/*
2607 		 * For some reason the non self refresh
2608 		 * FIFO size is only half of the self
2609 		 * refresh FIFO size on ILK/SNB.
2610 		 */
2611 		if (DISPLAY_VER(display) < 7)
2612 			fifo_size /= 2;
2613 	}
2614 
2615 	if (config->sprites_enabled) {
2616 		/* level 0 is always calculated with 1:1 split */
2617 		if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
2618 			if (is_sprite)
2619 				fifo_size *= 5;
2620 			fifo_size /= 6;
2621 		} else {
2622 			fifo_size /= 2;
2623 		}
2624 	}
2625 
2626 	/* clamp to max that the registers can hold */
2627 	return min(fifo_size, ilk_plane_wm_reg_max(display, level, is_sprite));
2628 }
2629 
2630 /* Calculate the maximum cursor plane watermark */
2631 static unsigned int ilk_cursor_wm_max(struct intel_display *display,
2632 				      int level,
2633 				      const struct intel_wm_config *config)
2634 {
2635 	/* HSW LP1+ watermarks w/ multiple pipes */
2636 	if (level > 0 && config->num_pipes_active > 1)
2637 		return 64;
2638 
2639 	/* otherwise just report max that registers can hold */
2640 	return ilk_cursor_wm_reg_max(display, level);
2641 }
2642 
2643 static void ilk_compute_wm_maximums(struct intel_display *display,
2644 				    int level,
2645 				    const struct intel_wm_config *config,
2646 				    enum intel_ddb_partitioning ddb_partitioning,
2647 				    struct ilk_wm_maximums *max)
2648 {
2649 	max->pri = ilk_plane_wm_max(display, level, config, ddb_partitioning, false);
2650 	max->spr = ilk_plane_wm_max(display, level, config, ddb_partitioning, true);
2651 	max->cur = ilk_cursor_wm_max(display, level, config);
2652 	max->fbc = ilk_fbc_wm_reg_max(display);
2653 }
2654 
2655 static void ilk_compute_wm_reg_maximums(struct intel_display *display,
2656 					int level,
2657 					struct ilk_wm_maximums *max)
2658 {
2659 	max->pri = ilk_plane_wm_reg_max(display, level, false);
2660 	max->spr = ilk_plane_wm_reg_max(display, level, true);
2661 	max->cur = ilk_cursor_wm_reg_max(display, level);
2662 	max->fbc = ilk_fbc_wm_reg_max(display);
2663 }
2664 
2665 static bool ilk_validate_wm_level(struct intel_display *display,
2666 				  int level,
2667 				  const struct ilk_wm_maximums *max,
2668 				  struct intel_wm_level *result)
2669 {
2670 	bool ret;
2671 
2672 	/* already determined to be invalid? */
2673 	if (!result->enable)
2674 		return false;
2675 
2676 	result->enable = result->pri_val <= max->pri &&
2677 			 result->spr_val <= max->spr &&
2678 			 result->cur_val <= max->cur;
2679 
2680 	ret = result->enable;
2681 
2682 	/*
2683 	 * HACK until we can pre-compute everything,
2684 	 * and thus fail gracefully if LP0 watermarks
2685 	 * are exceeded...
2686 	 */
2687 	if (level == 0 && !result->enable) {
2688 		if (result->pri_val > max->pri)
2689 			drm_dbg_kms(display->drm,
2690 				    "Primary WM%d too large %u (max %u)\n",
2691 				    level, result->pri_val, max->pri);
2692 		if (result->spr_val > max->spr)
2693 			drm_dbg_kms(display->drm,
2694 				    "Sprite WM%d too large %u (max %u)\n",
2695 				    level, result->spr_val, max->spr);
2696 		if (result->cur_val > max->cur)
2697 			drm_dbg_kms(display->drm,
2698 				    "Cursor WM%d too large %u (max %u)\n",
2699 				    level, result->cur_val, max->cur);
2700 
2701 		result->pri_val = min_t(u32, result->pri_val, max->pri);
2702 		result->spr_val = min_t(u32, result->spr_val, max->spr);
2703 		result->cur_val = min_t(u32, result->cur_val, max->cur);
2704 		result->enable = true;
2705 	}
2706 
2707 	return ret;
2708 }
2709 
2710 static void ilk_compute_wm_level(struct intel_display *display,
2711 				 const struct intel_crtc *crtc,
2712 				 int level,
2713 				 struct intel_crtc_state *crtc_state,
2714 				 const struct intel_plane_state *pristate,
2715 				 const struct intel_plane_state *sprstate,
2716 				 const struct intel_plane_state *curstate,
2717 				 struct intel_wm_level *result)
2718 {
2719 	u16 pri_latency = display->wm.pri_latency[level];
2720 	u16 spr_latency = display->wm.spr_latency[level];
2721 	u16 cur_latency = display->wm.cur_latency[level];
2722 
2723 	/* WM1+ latency values stored in 0.5us units */
2724 	if (level > 0) {
2725 		pri_latency *= 5;
2726 		spr_latency *= 5;
2727 		cur_latency *= 5;
2728 	}
2729 
2730 	if (pristate) {
2731 		result->pri_val = ilk_compute_pri_wm(crtc_state, pristate,
2732 						     pri_latency, level);
2733 		result->fbc_val = ilk_compute_fbc_wm(crtc_state, pristate, result->pri_val);
2734 	}
2735 
2736 	if (sprstate)
2737 		result->spr_val = ilk_compute_spr_wm(crtc_state, sprstate, spr_latency);
2738 
2739 	if (curstate)
2740 		result->cur_val = ilk_compute_cur_wm(crtc_state, curstate, cur_latency);
2741 
2742 	result->enable = true;
2743 }
2744 
2745 static void hsw_read_wm_latency(struct intel_display *display, u16 wm[])
2746 {
2747 	struct drm_i915_private *i915 = to_i915(display->drm);
2748 	u64 sskpd;
2749 
2750 	display->wm.num_levels = 5;
2751 
2752 	sskpd = intel_uncore_read64(&i915->uncore, MCH_SSKPD);
2753 
2754 	wm[0] = REG_FIELD_GET64(SSKPD_NEW_WM0_MASK_HSW, sskpd);
2755 	if (wm[0] == 0)
2756 		wm[0] = REG_FIELD_GET64(SSKPD_OLD_WM0_MASK_HSW, sskpd);
2757 	wm[1] = REG_FIELD_GET64(SSKPD_WM1_MASK_HSW, sskpd);
2758 	wm[2] = REG_FIELD_GET64(SSKPD_WM2_MASK_HSW, sskpd);
2759 	wm[3] = REG_FIELD_GET64(SSKPD_WM3_MASK_HSW, sskpd);
2760 	wm[4] = REG_FIELD_GET64(SSKPD_WM4_MASK_HSW, sskpd);
2761 }
2762 
2763 static void snb_read_wm_latency(struct intel_display *display, u16 wm[])
2764 {
2765 	struct drm_i915_private *i915 = to_i915(display->drm);
2766 	u32 sskpd;
2767 
2768 	display->wm.num_levels = 4;
2769 
2770 	sskpd = intel_uncore_read(&i915->uncore, MCH_SSKPD);
2771 
2772 	wm[0] = REG_FIELD_GET(SSKPD_WM0_MASK_SNB, sskpd);
2773 	wm[1] = REG_FIELD_GET(SSKPD_WM1_MASK_SNB, sskpd);
2774 	wm[2] = REG_FIELD_GET(SSKPD_WM2_MASK_SNB, sskpd);
2775 	wm[3] = REG_FIELD_GET(SSKPD_WM3_MASK_SNB, sskpd);
2776 }
2777 
2778 static void ilk_read_wm_latency(struct intel_display *display, u16 wm[])
2779 {
2780 	struct drm_i915_private *i915 = to_i915(display->drm);
2781 	u32 mltr;
2782 
2783 	display->wm.num_levels = 3;
2784 
2785 	mltr = intel_uncore_read(&i915->uncore, MLTR_ILK);
2786 
2787 	/* ILK primary LP0 latency is 700 ns */
2788 	wm[0] = 7;
2789 	wm[1] = REG_FIELD_GET(MLTR_WM1_MASK, mltr);
2790 	wm[2] = REG_FIELD_GET(MLTR_WM2_MASK, mltr);
2791 }
2792 
2793 static void intel_fixup_spr_wm_latency(struct intel_display *display, u16 wm[5])
2794 {
2795 	/* ILK sprite LP0 latency is 1300 ns */
2796 	if (DISPLAY_VER(display) == 5)
2797 		wm[0] = 13;
2798 }
2799 
2800 static void intel_fixup_cur_wm_latency(struct intel_display *display, u16 wm[5])
2801 {
2802 	/* ILK cursor LP0 latency is 1300 ns */
2803 	if (DISPLAY_VER(display) == 5)
2804 		wm[0] = 13;
2805 }
2806 
2807 static bool ilk_increase_wm_latency(struct intel_display *display, u16 wm[5], u16 min)
2808 {
2809 	int level;
2810 
2811 	if (wm[0] >= min)
2812 		return false;
2813 
2814 	wm[0] = max(wm[0], min);
2815 	for (level = 1; level < display->wm.num_levels; level++)
2816 		wm[level] = max_t(u16, wm[level], DIV_ROUND_UP(min, 5));
2817 
2818 	return true;
2819 }
2820 
2821 static void snb_wm_latency_quirk(struct intel_display *display)
2822 {
2823 	bool changed;
2824 
2825 	/*
2826 	 * The BIOS provided WM memory latency values are often
2827 	 * inadequate for high resolution displays. Adjust them.
2828 	 */
2829 	changed = ilk_increase_wm_latency(display, display->wm.pri_latency, 12);
2830 	changed |= ilk_increase_wm_latency(display, display->wm.spr_latency, 12);
2831 	changed |= ilk_increase_wm_latency(display, display->wm.cur_latency, 12);
2832 
2833 	if (!changed)
2834 		return;
2835 
2836 	drm_dbg_kms(display->drm,
2837 		    "WM latency values increased to avoid potential underruns\n");
2838 	intel_print_wm_latency(display, "Primary", display->wm.pri_latency);
2839 	intel_print_wm_latency(display, "Sprite", display->wm.spr_latency);
2840 	intel_print_wm_latency(display, "Cursor", display->wm.cur_latency);
2841 }
2842 
2843 static void snb_wm_lp3_irq_quirk(struct intel_display *display)
2844 {
2845 	/*
2846 	 * On some SNB machines (Thinkpad X220 Tablet at least)
2847 	 * LP3 usage can cause vblank interrupts to be lost.
2848 	 * The DEIIR bit will go high but it looks like the CPU
2849 	 * never gets interrupted.
2850 	 *
2851 	 * It's not clear whether other interrupt source could
2852 	 * be affected or if this is somehow limited to vblank
2853 	 * interrupts only. To play it safe we disable LP3
2854 	 * watermarks entirely.
2855 	 */
2856 	if (display->wm.pri_latency[3] == 0 &&
2857 	    display->wm.spr_latency[3] == 0 &&
2858 	    display->wm.cur_latency[3] == 0)
2859 		return;
2860 
2861 	display->wm.pri_latency[3] = 0;
2862 	display->wm.spr_latency[3] = 0;
2863 	display->wm.cur_latency[3] = 0;
2864 
2865 	drm_dbg_kms(display->drm,
2866 		    "LP3 watermarks disabled due to potential for lost interrupts\n");
2867 	intel_print_wm_latency(display, "Primary", display->wm.pri_latency);
2868 	intel_print_wm_latency(display, "Sprite", display->wm.spr_latency);
2869 	intel_print_wm_latency(display, "Cursor", display->wm.cur_latency);
2870 }
2871 
2872 static void ilk_setup_wm_latency(struct intel_display *display)
2873 {
2874 	if (display->platform.broadwell || display->platform.haswell)
2875 		hsw_read_wm_latency(display, display->wm.pri_latency);
2876 	else if (DISPLAY_VER(display) >= 6)
2877 		snb_read_wm_latency(display, display->wm.pri_latency);
2878 	else
2879 		ilk_read_wm_latency(display, display->wm.pri_latency);
2880 
2881 	memcpy(display->wm.spr_latency, display->wm.pri_latency,
2882 	       sizeof(display->wm.pri_latency));
2883 	memcpy(display->wm.cur_latency, display->wm.pri_latency,
2884 	       sizeof(display->wm.pri_latency));
2885 
2886 	intel_fixup_spr_wm_latency(display, display->wm.spr_latency);
2887 	intel_fixup_cur_wm_latency(display, display->wm.cur_latency);
2888 
2889 	intel_print_wm_latency(display, "Primary", display->wm.pri_latency);
2890 	intel_print_wm_latency(display, "Sprite", display->wm.spr_latency);
2891 	intel_print_wm_latency(display, "Cursor", display->wm.cur_latency);
2892 
2893 	if (DISPLAY_VER(display) == 6) {
2894 		snb_wm_latency_quirk(display);
2895 		snb_wm_lp3_irq_quirk(display);
2896 	}
2897 }
2898 
2899 static bool ilk_validate_pipe_wm(struct intel_display *display,
2900 				 struct intel_pipe_wm *pipe_wm)
2901 {
2902 	/* LP0 watermark maximums depend on this pipe alone */
2903 	const struct intel_wm_config config = {
2904 		.num_pipes_active = 1,
2905 		.sprites_enabled = pipe_wm->sprites_enabled,
2906 		.sprites_scaled = pipe_wm->sprites_scaled,
2907 	};
2908 	struct ilk_wm_maximums max;
2909 
2910 	/* LP0 watermarks always use 1/2 DDB partitioning */
2911 	ilk_compute_wm_maximums(display, 0, &config, INTEL_DDB_PART_1_2, &max);
2912 
2913 	/* At least LP0 must be valid */
2914 	if (!ilk_validate_wm_level(display, 0, &max, &pipe_wm->wm[0])) {
2915 		drm_dbg_kms(display->drm, "LP0 watermark invalid\n");
2916 		return false;
2917 	}
2918 
2919 	return true;
2920 }
2921 
2922 /* Compute new watermarks for the pipe */
2923 static int ilk_compute_pipe_wm(struct intel_atomic_state *state,
2924 			       struct intel_crtc *crtc)
2925 {
2926 	struct intel_display *display = to_intel_display(state);
2927 	struct intel_crtc_state *crtc_state =
2928 		intel_atomic_get_new_crtc_state(state, crtc);
2929 	struct intel_pipe_wm *pipe_wm;
2930 	struct intel_plane *plane;
2931 	const struct intel_plane_state *plane_state;
2932 	const struct intel_plane_state *pristate = NULL;
2933 	const struct intel_plane_state *sprstate = NULL;
2934 	const struct intel_plane_state *curstate = NULL;
2935 	struct ilk_wm_maximums max;
2936 	int level, usable_level;
2937 
2938 	pipe_wm = &crtc_state->wm.ilk.optimal;
2939 
2940 	intel_atomic_crtc_state_for_each_plane_state(plane, plane_state, crtc_state) {
2941 		if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
2942 			pristate = plane_state;
2943 		else if (plane->base.type == DRM_PLANE_TYPE_OVERLAY)
2944 			sprstate = plane_state;
2945 		else if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
2946 			curstate = plane_state;
2947 	}
2948 
2949 	pipe_wm->pipe_enabled = crtc_state->hw.active;
2950 	pipe_wm->sprites_enabled = crtc_state->active_planes & BIT(PLANE_SPRITE0);
2951 	pipe_wm->sprites_scaled = crtc_state->scaled_planes & BIT(PLANE_SPRITE0);
2952 
2953 	usable_level = display->wm.num_levels - 1;
2954 
2955 	/* ILK/SNB: LP2+ watermarks only w/o sprites */
2956 	if (DISPLAY_VER(display) < 7 && pipe_wm->sprites_enabled)
2957 		usable_level = 1;
2958 
2959 	/* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2960 	if (pipe_wm->sprites_scaled)
2961 		usable_level = 0;
2962 
2963 	memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
2964 	ilk_compute_wm_level(display, crtc, 0, crtc_state,
2965 			     pristate, sprstate, curstate, &pipe_wm->wm[0]);
2966 
2967 	if (!ilk_validate_pipe_wm(display, pipe_wm))
2968 		return -EINVAL;
2969 
2970 	ilk_compute_wm_reg_maximums(display, 1, &max);
2971 
2972 	for (level = 1; level <= usable_level; level++) {
2973 		struct intel_wm_level *wm = &pipe_wm->wm[level];
2974 
2975 		ilk_compute_wm_level(display, crtc, level, crtc_state,
2976 				     pristate, sprstate, curstate, wm);
2977 
2978 		/*
2979 		 * Disable any watermark level that exceeds the
2980 		 * register maximums since such watermarks are
2981 		 * always invalid.
2982 		 */
2983 		if (!ilk_validate_wm_level(display, level, &max, wm)) {
2984 			memset(wm, 0, sizeof(*wm));
2985 			break;
2986 		}
2987 	}
2988 
2989 	return 0;
2990 }
2991 
2992 /*
2993  * Build a set of 'intermediate' watermark values that satisfy both the old
2994  * state and the new state.  These can be programmed to the hardware
2995  * immediately.
2996  */
2997 static int ilk_compute_intermediate_wm(struct intel_atomic_state *state,
2998 				       struct intel_crtc *crtc)
2999 {
3000 	struct intel_display *display = to_intel_display(crtc);
3001 	struct intel_crtc_state *new_crtc_state =
3002 		intel_atomic_get_new_crtc_state(state, crtc);
3003 	const struct intel_crtc_state *old_crtc_state =
3004 		intel_atomic_get_old_crtc_state(state, crtc);
3005 	struct intel_pipe_wm *intermediate = &new_crtc_state->wm.ilk.intermediate;
3006 	const struct intel_pipe_wm *optimal = &new_crtc_state->wm.ilk.optimal;
3007 	const struct intel_pipe_wm *active = &old_crtc_state->wm.ilk.optimal;
3008 	int level;
3009 
3010 	/*
3011 	 * Start with the final, target watermarks, then combine with the
3012 	 * currently active watermarks to get values that are safe both before
3013 	 * and after the vblank.
3014 	 */
3015 	*intermediate = *optimal;
3016 	if (!new_crtc_state->hw.active ||
3017 	    intel_crtc_needs_modeset(new_crtc_state) ||
3018 	    state->skip_intermediate_wm)
3019 		return 0;
3020 
3021 	intermediate->pipe_enabled |= active->pipe_enabled;
3022 	intermediate->sprites_enabled |= active->sprites_enabled;
3023 	intermediate->sprites_scaled |= active->sprites_scaled;
3024 
3025 	for (level = 0; level < display->wm.num_levels; level++) {
3026 		struct intel_wm_level *intermediate_wm = &intermediate->wm[level];
3027 		const struct intel_wm_level *active_wm = &active->wm[level];
3028 
3029 		intermediate_wm->enable &= active_wm->enable;
3030 		intermediate_wm->pri_val = max(intermediate_wm->pri_val,
3031 					       active_wm->pri_val);
3032 		intermediate_wm->spr_val = max(intermediate_wm->spr_val,
3033 					       active_wm->spr_val);
3034 		intermediate_wm->cur_val = max(intermediate_wm->cur_val,
3035 					       active_wm->cur_val);
3036 		intermediate_wm->fbc_val = max(intermediate_wm->fbc_val,
3037 					       active_wm->fbc_val);
3038 	}
3039 
3040 	/*
3041 	 * We need to make sure that these merged watermark values are
3042 	 * actually a valid configuration themselves.  If they're not,
3043 	 * there's no safe way to transition from the old state to
3044 	 * the new state, so we need to fail the atomic transaction.
3045 	 */
3046 	if (!ilk_validate_pipe_wm(display, intermediate))
3047 		return -EINVAL;
3048 
3049 	/*
3050 	 * If our intermediate WM are identical to the final WM, then we can
3051 	 * omit the post-vblank programming; only update if it's different.
3052 	 */
3053 	if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
3054 		new_crtc_state->wm.need_postvbl_update = true;
3055 
3056 	return 0;
3057 }
3058 
3059 static int ilk_compute_watermarks(struct intel_atomic_state *state,
3060 				  struct intel_crtc *crtc)
3061 {
3062 	int ret;
3063 
3064 	ret = ilk_compute_pipe_wm(state, crtc);
3065 	if (ret)
3066 		return ret;
3067 
3068 	ret = ilk_compute_intermediate_wm(state, crtc);
3069 	if (ret)
3070 		return ret;
3071 
3072 	return 0;
3073 }
3074 
3075 /*
3076  * Merge the watermarks from all active pipes for a specific level.
3077  */
3078 static void ilk_merge_wm_level(struct intel_display *display,
3079 			       int level,
3080 			       struct intel_wm_level *ret_wm)
3081 {
3082 	const struct intel_crtc *crtc;
3083 
3084 	ret_wm->enable = true;
3085 
3086 	for_each_intel_crtc(display->drm, crtc) {
3087 		const struct intel_pipe_wm *active = &crtc->wm.active.ilk;
3088 		const struct intel_wm_level *wm = &active->wm[level];
3089 
3090 		if (!active->pipe_enabled)
3091 			continue;
3092 
3093 		/*
3094 		 * The watermark values may have been used in the past,
3095 		 * so we must maintain them in the registers for some
3096 		 * time even if the level is now disabled.
3097 		 */
3098 		if (!wm->enable)
3099 			ret_wm->enable = false;
3100 
3101 		ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
3102 		ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
3103 		ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
3104 		ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
3105 	}
3106 }
3107 
3108 /*
3109  * Merge all low power watermarks for all active pipes.
3110  */
3111 static void ilk_wm_merge(struct intel_display *display,
3112 			 const struct intel_wm_config *config,
3113 			 const struct ilk_wm_maximums *max,
3114 			 struct intel_pipe_wm *merged)
3115 {
3116 	int level, num_levels = display->wm.num_levels;
3117 	int last_enabled_level = num_levels - 1;
3118 
3119 	/* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
3120 	if ((DISPLAY_VER(display) < 7 || display->platform.ivybridge) &&
3121 	    config->num_pipes_active > 1)
3122 		last_enabled_level = 0;
3123 
3124 	/* ILK: FBC WM must be disabled always */
3125 	merged->fbc_wm_enabled = DISPLAY_VER(display) >= 6;
3126 
3127 	/* merge each WM1+ level */
3128 	for (level = 1; level < num_levels; level++) {
3129 		struct intel_wm_level *wm = &merged->wm[level];
3130 
3131 		ilk_merge_wm_level(display, level, wm);
3132 
3133 		if (level > last_enabled_level)
3134 			wm->enable = false;
3135 		else if (!ilk_validate_wm_level(display, level, max, wm))
3136 			/* make sure all following levels get disabled */
3137 			last_enabled_level = level - 1;
3138 
3139 		/*
3140 		 * The spec says it is preferred to disable
3141 		 * FBC WMs instead of disabling a WM level.
3142 		 */
3143 		if (wm->fbc_val > max->fbc) {
3144 			if (wm->enable)
3145 				merged->fbc_wm_enabled = false;
3146 			wm->fbc_val = 0;
3147 		}
3148 	}
3149 
3150 	/* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
3151 	if (DISPLAY_VER(display) == 5 && HAS_FBC(display) &&
3152 	    display->params.enable_fbc && !merged->fbc_wm_enabled) {
3153 		for (level = 2; level < num_levels; level++) {
3154 			struct intel_wm_level *wm = &merged->wm[level];
3155 
3156 			wm->enable = false;
3157 		}
3158 	}
3159 }
3160 
3161 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
3162 {
3163 	/* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
3164 	return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
3165 }
3166 
3167 /* The value we need to program into the WM_LPx latency field */
3168 static unsigned int ilk_wm_lp_latency(struct intel_display *display,
3169 				      int level)
3170 {
3171 	if (display->platform.haswell || display->platform.broadwell)
3172 		return 2 * level;
3173 	else
3174 		return display->wm.pri_latency[level];
3175 }
3176 
3177 static void ilk_compute_wm_results(struct intel_display *display,
3178 				   const struct intel_pipe_wm *merged,
3179 				   enum intel_ddb_partitioning partitioning,
3180 				   struct ilk_wm_values *results)
3181 {
3182 	struct intel_crtc *crtc;
3183 	int level, wm_lp;
3184 
3185 	results->enable_fbc_wm = merged->fbc_wm_enabled;
3186 	results->partitioning = partitioning;
3187 
3188 	/* LP1+ register values */
3189 	for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3190 		const struct intel_wm_level *r;
3191 
3192 		level = ilk_wm_lp_to_level(wm_lp, merged);
3193 
3194 		r = &merged->wm[level];
3195 
3196 		/*
3197 		 * Maintain the watermark values even if the level is
3198 		 * disabled. Doing otherwise could cause underruns.
3199 		 */
3200 		results->wm_lp[wm_lp - 1] =
3201 			WM_LP_LATENCY(ilk_wm_lp_latency(display, level)) |
3202 			WM_LP_PRIMARY(r->pri_val) |
3203 			WM_LP_CURSOR(r->cur_val);
3204 
3205 		if (r->enable)
3206 			results->wm_lp[wm_lp - 1] |= WM_LP_ENABLE;
3207 
3208 		if (DISPLAY_VER(display) >= 8)
3209 			results->wm_lp[wm_lp - 1] |= WM_LP_FBC_BDW(r->fbc_val);
3210 		else
3211 			results->wm_lp[wm_lp - 1] |= WM_LP_FBC_ILK(r->fbc_val);
3212 
3213 		results->wm_lp_spr[wm_lp - 1] = WM_LP_SPRITE(r->spr_val);
3214 
3215 		/*
3216 		 * Always set WM_LP_SPRITE_EN when spr_val != 0, even if the
3217 		 * level is disabled. Doing otherwise could cause underruns.
3218 		 */
3219 		if (DISPLAY_VER(display) < 7 && r->spr_val) {
3220 			drm_WARN_ON(display->drm, wm_lp != 1);
3221 			results->wm_lp_spr[wm_lp - 1] |= WM_LP_SPRITE_ENABLE;
3222 		}
3223 	}
3224 
3225 	/* LP0 register values */
3226 	for_each_intel_crtc(display->drm, crtc) {
3227 		enum pipe pipe = crtc->pipe;
3228 		const struct intel_pipe_wm *pipe_wm = &crtc->wm.active.ilk;
3229 		const struct intel_wm_level *r = &pipe_wm->wm[0];
3230 
3231 		if (drm_WARN_ON(display->drm, !r->enable))
3232 			continue;
3233 
3234 		results->wm_pipe[pipe] =
3235 			WM0_PIPE_PRIMARY(r->pri_val) |
3236 			WM0_PIPE_SPRITE(r->spr_val) |
3237 			WM0_PIPE_CURSOR(r->cur_val);
3238 	}
3239 }
3240 
3241 /*
3242  * Find the result with the highest level enabled. Check for enable_fbc_wm in
3243  * case both are at the same level. Prefer r1 in case they're the same.
3244  */
3245 static struct intel_pipe_wm *
3246 ilk_find_best_result(struct intel_display *display,
3247 		     struct intel_pipe_wm *r1,
3248 		     struct intel_pipe_wm *r2)
3249 {
3250 	int level, level1 = 0, level2 = 0;
3251 
3252 	for (level = 1; level < display->wm.num_levels; level++) {
3253 		if (r1->wm[level].enable)
3254 			level1 = level;
3255 		if (r2->wm[level].enable)
3256 			level2 = level;
3257 	}
3258 
3259 	if (level1 == level2) {
3260 		if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
3261 			return r2;
3262 		else
3263 			return r1;
3264 	} else if (level1 > level2) {
3265 		return r1;
3266 	} else {
3267 		return r2;
3268 	}
3269 }
3270 
3271 /* dirty bits used to track which watermarks need changes */
3272 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
3273 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
3274 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
3275 #define WM_DIRTY_FBC (1 << 24)
3276 #define WM_DIRTY_DDB (1 << 25)
3277 
3278 static unsigned int ilk_compute_wm_dirty(struct intel_display *display,
3279 					 const struct ilk_wm_values *old,
3280 					 const struct ilk_wm_values *new)
3281 {
3282 	unsigned int dirty = 0;
3283 	enum pipe pipe;
3284 	int wm_lp;
3285 
3286 	for_each_pipe(display, pipe) {
3287 		if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
3288 			dirty |= WM_DIRTY_PIPE(pipe);
3289 			/* Must disable LP1+ watermarks too */
3290 			dirty |= WM_DIRTY_LP_ALL;
3291 		}
3292 	}
3293 
3294 	if (old->enable_fbc_wm != new->enable_fbc_wm) {
3295 		dirty |= WM_DIRTY_FBC;
3296 		/* Must disable LP1+ watermarks too */
3297 		dirty |= WM_DIRTY_LP_ALL;
3298 	}
3299 
3300 	if (old->partitioning != new->partitioning) {
3301 		dirty |= WM_DIRTY_DDB;
3302 		/* Must disable LP1+ watermarks too */
3303 		dirty |= WM_DIRTY_LP_ALL;
3304 	}
3305 
3306 	/* LP1+ watermarks already deemed dirty, no need to continue */
3307 	if (dirty & WM_DIRTY_LP_ALL)
3308 		return dirty;
3309 
3310 	/* Find the lowest numbered LP1+ watermark in need of an update... */
3311 	for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3312 		if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
3313 		    old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
3314 			break;
3315 	}
3316 
3317 	/* ...and mark it and all higher numbered LP1+ watermarks as dirty */
3318 	for (; wm_lp <= 3; wm_lp++)
3319 		dirty |= WM_DIRTY_LP(wm_lp);
3320 
3321 	return dirty;
3322 }
3323 
3324 static bool _ilk_disable_lp_wm(struct intel_display *display,
3325 			       unsigned int dirty)
3326 {
3327 	struct ilk_wm_values *previous = &display->wm.hw;
3328 	bool changed = false;
3329 
3330 	if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM_LP_ENABLE) {
3331 		previous->wm_lp[2] &= ~WM_LP_ENABLE;
3332 		intel_de_write(display, WM3_LP_ILK, previous->wm_lp[2]);
3333 		changed = true;
3334 	}
3335 	if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM_LP_ENABLE) {
3336 		previous->wm_lp[1] &= ~WM_LP_ENABLE;
3337 		intel_de_write(display, WM2_LP_ILK, previous->wm_lp[1]);
3338 		changed = true;
3339 	}
3340 	if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM_LP_ENABLE) {
3341 		previous->wm_lp[0] &= ~WM_LP_ENABLE;
3342 		intel_de_write(display, WM1_LP_ILK, previous->wm_lp[0]);
3343 		changed = true;
3344 	}
3345 
3346 	/*
3347 	 * Don't touch WM_LP_SPRITE_ENABLE here.
3348 	 * Doing so could cause underruns.
3349 	 */
3350 
3351 	return changed;
3352 }
3353 
3354 /*
3355  * The spec says we shouldn't write when we don't need, because every write
3356  * causes WMs to be re-evaluated, expending some power.
3357  */
3358 static void ilk_write_wm_values(struct intel_display *display,
3359 				struct ilk_wm_values *results)
3360 {
3361 	struct ilk_wm_values *previous = &display->wm.hw;
3362 	unsigned int dirty;
3363 
3364 	dirty = ilk_compute_wm_dirty(display, previous, results);
3365 	if (!dirty)
3366 		return;
3367 
3368 	_ilk_disable_lp_wm(display, dirty);
3369 
3370 	if (dirty & WM_DIRTY_PIPE(PIPE_A))
3371 		intel_de_write(display, WM0_PIPE_ILK(PIPE_A), results->wm_pipe[0]);
3372 	if (dirty & WM_DIRTY_PIPE(PIPE_B))
3373 		intel_de_write(display, WM0_PIPE_ILK(PIPE_B), results->wm_pipe[1]);
3374 	if (dirty & WM_DIRTY_PIPE(PIPE_C))
3375 		intel_de_write(display, WM0_PIPE_ILK(PIPE_C), results->wm_pipe[2]);
3376 
3377 	if (dirty & WM_DIRTY_DDB) {
3378 		if (display->platform.haswell || display->platform.broadwell)
3379 			intel_de_rmw(display, WM_MISC, WM_MISC_DATA_PARTITION_5_6,
3380 				     results->partitioning == INTEL_DDB_PART_1_2 ? 0 :
3381 				     WM_MISC_DATA_PARTITION_5_6);
3382 		else
3383 			intel_de_rmw(display, DISP_ARB_CTL2, DISP_DATA_PARTITION_5_6,
3384 				     results->partitioning == INTEL_DDB_PART_1_2 ? 0 :
3385 				     DISP_DATA_PARTITION_5_6);
3386 	}
3387 
3388 	if (dirty & WM_DIRTY_FBC)
3389 		intel_de_rmw(display, DISP_ARB_CTL, DISP_FBC_WM_DIS,
3390 			     results->enable_fbc_wm ? 0 : DISP_FBC_WM_DIS);
3391 
3392 	if (dirty & WM_DIRTY_LP(1) &&
3393 	    previous->wm_lp_spr[0] != results->wm_lp_spr[0])
3394 		intel_de_write(display, WM1S_LP_ILK, results->wm_lp_spr[0]);
3395 
3396 	if (DISPLAY_VER(display) >= 7) {
3397 		if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
3398 			intel_de_write(display, WM2S_LP_IVB, results->wm_lp_spr[1]);
3399 		if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
3400 			intel_de_write(display, WM3S_LP_IVB, results->wm_lp_spr[2]);
3401 	}
3402 
3403 	if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
3404 		intel_de_write(display, WM1_LP_ILK, results->wm_lp[0]);
3405 	if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
3406 		intel_de_write(display, WM2_LP_ILK, results->wm_lp[1]);
3407 	if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
3408 		intel_de_write(display, WM3_LP_ILK, results->wm_lp[2]);
3409 
3410 	display->wm.hw = *results;
3411 }
3412 
3413 bool ilk_disable_cxsr(struct intel_display *display)
3414 {
3415 	return _ilk_disable_lp_wm(display, WM_DIRTY_LP_ALL);
3416 }
3417 
3418 static void ilk_compute_wm_config(struct intel_display *display,
3419 				  struct intel_wm_config *config)
3420 {
3421 	struct intel_crtc *crtc;
3422 
3423 	/* Compute the currently _active_ config */
3424 	for_each_intel_crtc(display->drm, crtc) {
3425 		const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
3426 
3427 		if (!wm->pipe_enabled)
3428 			continue;
3429 
3430 		config->sprites_enabled |= wm->sprites_enabled;
3431 		config->sprites_scaled |= wm->sprites_scaled;
3432 		config->num_pipes_active++;
3433 	}
3434 }
3435 
3436 static void ilk_program_watermarks(struct intel_display *display)
3437 {
3438 	struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
3439 	struct ilk_wm_maximums max;
3440 	struct intel_wm_config config = {};
3441 	struct ilk_wm_values results = {};
3442 	enum intel_ddb_partitioning partitioning;
3443 
3444 	ilk_compute_wm_config(display, &config);
3445 
3446 	ilk_compute_wm_maximums(display, 1, &config, INTEL_DDB_PART_1_2, &max);
3447 	ilk_wm_merge(display, &config, &max, &lp_wm_1_2);
3448 
3449 	/* 5/6 split only in single pipe config on IVB+ */
3450 	if (DISPLAY_VER(display) >= 7 &&
3451 	    config.num_pipes_active == 1 && config.sprites_enabled) {
3452 		ilk_compute_wm_maximums(display, 1, &config, INTEL_DDB_PART_5_6, &max);
3453 		ilk_wm_merge(display, &config, &max, &lp_wm_5_6);
3454 
3455 		best_lp_wm = ilk_find_best_result(display, &lp_wm_1_2, &lp_wm_5_6);
3456 	} else {
3457 		best_lp_wm = &lp_wm_1_2;
3458 	}
3459 
3460 	partitioning = (best_lp_wm == &lp_wm_1_2) ?
3461 		       INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
3462 
3463 	ilk_compute_wm_results(display, best_lp_wm, partitioning, &results);
3464 
3465 	ilk_write_wm_values(display, &results);
3466 }
3467 
3468 static void ilk_initial_watermarks(struct intel_atomic_state *state,
3469 				   struct intel_crtc *crtc)
3470 {
3471 	struct intel_display *display = to_intel_display(crtc);
3472 	const struct intel_crtc_state *crtc_state =
3473 		intel_atomic_get_new_crtc_state(state, crtc);
3474 
3475 	mutex_lock(&display->wm.wm_mutex);
3476 	crtc->wm.active.ilk = crtc_state->wm.ilk.intermediate;
3477 	ilk_program_watermarks(display);
3478 	mutex_unlock(&display->wm.wm_mutex);
3479 }
3480 
3481 static void ilk_optimize_watermarks(struct intel_atomic_state *state,
3482 				    struct intel_crtc *crtc)
3483 {
3484 	struct intel_display *display = to_intel_display(crtc);
3485 	const struct intel_crtc_state *crtc_state =
3486 		intel_atomic_get_new_crtc_state(state, crtc);
3487 
3488 	if (!crtc_state->wm.need_postvbl_update)
3489 		return;
3490 
3491 	mutex_lock(&display->wm.wm_mutex);
3492 	crtc->wm.active.ilk = crtc_state->wm.ilk.optimal;
3493 	ilk_program_watermarks(display);
3494 	mutex_unlock(&display->wm.wm_mutex);
3495 }
3496 
3497 static void ilk_pipe_wm_get_hw_state(struct intel_crtc *crtc)
3498 {
3499 	struct intel_display *display = to_intel_display(crtc);
3500 	struct ilk_wm_values *hw = &display->wm.hw;
3501 	struct intel_crtc_state *crtc_state = to_intel_crtc_state(crtc->base.state);
3502 	struct intel_pipe_wm *active = &crtc_state->wm.ilk.optimal;
3503 	enum pipe pipe = crtc->pipe;
3504 
3505 	hw->wm_pipe[pipe] = intel_de_read(display, WM0_PIPE_ILK(pipe));
3506 
3507 	memset(active, 0, sizeof(*active));
3508 
3509 	active->pipe_enabled = crtc->active;
3510 
3511 	if (active->pipe_enabled) {
3512 		u32 tmp = hw->wm_pipe[pipe];
3513 
3514 		/*
3515 		 * For active pipes LP0 watermark is marked as
3516 		 * enabled, and LP1+ watermaks as disabled since
3517 		 * we can't really reverse compute them in case
3518 		 * multiple pipes are active.
3519 		 */
3520 		active->wm[0].enable = true;
3521 		active->wm[0].pri_val = REG_FIELD_GET(WM0_PIPE_PRIMARY_MASK, tmp);
3522 		active->wm[0].spr_val = REG_FIELD_GET(WM0_PIPE_SPRITE_MASK, tmp);
3523 		active->wm[0].cur_val = REG_FIELD_GET(WM0_PIPE_CURSOR_MASK, tmp);
3524 	} else {
3525 		int level;
3526 
3527 		/*
3528 		 * For inactive pipes, all watermark levels
3529 		 * should be marked as enabled but zeroed,
3530 		 * which is what we'd compute them to.
3531 		 */
3532 		for (level = 0; level < display->wm.num_levels; level++)
3533 			active->wm[level].enable = true;
3534 	}
3535 
3536 	crtc->wm.active.ilk = *active;
3537 }
3538 
3539 static int ilk_sanitize_watermarks_add_affected(struct drm_atomic_state *state)
3540 {
3541 	struct drm_plane *plane;
3542 	struct intel_crtc *crtc;
3543 
3544 	for_each_intel_crtc(state->dev, crtc) {
3545 		struct intel_crtc_state *crtc_state;
3546 
3547 		crtc_state = intel_atomic_get_crtc_state(state, crtc);
3548 		if (IS_ERR(crtc_state))
3549 			return PTR_ERR(crtc_state);
3550 
3551 		if (crtc_state->hw.active) {
3552 			/*
3553 			 * Preserve the inherited flag to avoid
3554 			 * taking the full modeset path.
3555 			 */
3556 			crtc_state->inherited = true;
3557 		}
3558 	}
3559 
3560 	drm_for_each_plane(plane, state->dev) {
3561 		struct drm_plane_state *plane_state;
3562 
3563 		plane_state = drm_atomic_get_plane_state(state, plane);
3564 		if (IS_ERR(plane_state))
3565 			return PTR_ERR(plane_state);
3566 	}
3567 
3568 	return 0;
3569 }
3570 
3571 /*
3572  * Calculate what we think the watermarks should be for the state we've read
3573  * out of the hardware and then immediately program those watermarks so that
3574  * we ensure the hardware settings match our internal state.
3575  *
3576  * We can calculate what we think WM's should be by creating a duplicate of the
3577  * current state (which was constructed during hardware readout) and running it
3578  * through the atomic check code to calculate new watermark values in the
3579  * state object.
3580  */
3581 void ilk_wm_sanitize(struct intel_display *display)
3582 {
3583 	struct drm_atomic_state *state;
3584 	struct intel_atomic_state *intel_state;
3585 	struct intel_crtc *crtc;
3586 	struct intel_crtc_state *crtc_state;
3587 	struct drm_modeset_acquire_ctx ctx;
3588 	int ret;
3589 	int i;
3590 
3591 	/* Only supported on platforms that use atomic watermark design */
3592 	if (!display->funcs.wm->optimize_watermarks)
3593 		return;
3594 
3595 	if (drm_WARN_ON(display->drm, DISPLAY_VER(display) >= 9))
3596 		return;
3597 
3598 	state = drm_atomic_state_alloc(display->drm);
3599 	if (drm_WARN_ON(display->drm, !state))
3600 		return;
3601 
3602 	intel_state = to_intel_atomic_state(state);
3603 
3604 	drm_modeset_acquire_init(&ctx, 0);
3605 
3606 	state->acquire_ctx = &ctx;
3607 	to_intel_atomic_state(state)->internal = true;
3608 
3609 retry:
3610 	/*
3611 	 * Hardware readout is the only time we don't want to calculate
3612 	 * intermediate watermarks (since we don't trust the current
3613 	 * watermarks).
3614 	 */
3615 	if (!HAS_GMCH(display))
3616 		intel_state->skip_intermediate_wm = true;
3617 
3618 	ret = ilk_sanitize_watermarks_add_affected(state);
3619 	if (ret)
3620 		goto fail;
3621 
3622 	ret = intel_atomic_check(display->drm, state);
3623 	if (ret)
3624 		goto fail;
3625 
3626 	/* Write calculated watermark values back */
3627 	for_each_new_intel_crtc_in_state(intel_state, crtc, crtc_state, i) {
3628 		crtc_state->wm.need_postvbl_update = true;
3629 		intel_optimize_watermarks(intel_state, crtc);
3630 
3631 		to_intel_crtc_state(crtc->base.state)->wm = crtc_state->wm;
3632 	}
3633 
3634 fail:
3635 	if (ret == -EDEADLK) {
3636 		drm_atomic_state_clear(state);
3637 		drm_modeset_backoff(&ctx);
3638 		goto retry;
3639 	}
3640 
3641 	/*
3642 	 * If we fail here, it means that the hardware appears to be
3643 	 * programmed in a way that shouldn't be possible, given our
3644 	 * understanding of watermark requirements.  This might mean a
3645 	 * mistake in the hardware readout code or a mistake in the
3646 	 * watermark calculations for a given platform.  Raise a WARN
3647 	 * so that this is noticeable.
3648 	 *
3649 	 * If this actually happens, we'll have to just leave the
3650 	 * BIOS-programmed watermarks untouched and hope for the best.
3651 	 */
3652 	drm_WARN(display->drm, ret,
3653 		 "Could not determine valid watermarks for inherited state\n");
3654 
3655 	drm_atomic_state_put(state);
3656 
3657 	drm_modeset_drop_locks(&ctx);
3658 	drm_modeset_acquire_fini(&ctx);
3659 }
3660 
3661 #define _FW_WM(value, plane) \
3662 	(((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
3663 #define _FW_WM_VLV(value, plane) \
3664 	(((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
3665 
3666 static void g4x_read_wm_values(struct intel_display *display,
3667 			       struct g4x_wm_values *wm)
3668 {
3669 	u32 tmp;
3670 
3671 	tmp = intel_de_read(display, DSPFW1(display));
3672 	wm->sr.plane = _FW_WM(tmp, SR);
3673 	wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
3674 	wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEB);
3675 	wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEA);
3676 
3677 	tmp = intel_de_read(display, DSPFW2(display));
3678 	wm->fbc_en = tmp & DSPFW_FBC_SR_EN;
3679 	wm->sr.fbc = _FW_WM(tmp, FBC_SR);
3680 	wm->hpll.fbc = _FW_WM(tmp, FBC_HPLL_SR);
3681 	wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEB);
3682 	wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
3683 	wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEA);
3684 
3685 	tmp = intel_de_read(display, DSPFW3(display));
3686 	wm->hpll_en = tmp & DSPFW_HPLL_SR_EN;
3687 	wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
3688 	wm->hpll.cursor = _FW_WM(tmp, HPLL_CURSOR);
3689 	wm->hpll.plane = _FW_WM(tmp, HPLL_SR);
3690 }
3691 
3692 static void vlv_read_wm_values(struct intel_display *display,
3693 			       struct vlv_wm_values *wm)
3694 {
3695 	enum pipe pipe;
3696 	u32 tmp;
3697 
3698 	for_each_pipe(display, pipe) {
3699 		tmp = intel_de_read(display, VLV_DDL(pipe));
3700 
3701 		wm->ddl[pipe].plane[PLANE_PRIMARY] =
3702 			(tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3703 		wm->ddl[pipe].plane[PLANE_CURSOR] =
3704 			(tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3705 		wm->ddl[pipe].plane[PLANE_SPRITE0] =
3706 			(tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3707 		wm->ddl[pipe].plane[PLANE_SPRITE1] =
3708 			(tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3709 	}
3710 
3711 	tmp = intel_de_read(display, DSPFW1(display));
3712 	wm->sr.plane = _FW_WM(tmp, SR);
3713 	wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
3714 	wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEB);
3715 	wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEA);
3716 
3717 	tmp = intel_de_read(display, DSPFW2(display));
3718 	wm->pipe[PIPE_A].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEB);
3719 	wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
3720 	wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEA);
3721 
3722 	tmp = intel_de_read(display, DSPFW3(display));
3723 	wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
3724 
3725 	if (display->platform.cherryview) {
3726 		tmp = intel_de_read(display, DSPFW7_CHV);
3727 		wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
3728 		wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
3729 
3730 		tmp = intel_de_read(display, DSPFW8_CHV);
3731 		wm->pipe[PIPE_C].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEF);
3732 		wm->pipe[PIPE_C].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEE);
3733 
3734 		tmp = intel_de_read(display, DSPFW9_CHV);
3735 		wm->pipe[PIPE_C].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEC);
3736 		wm->pipe[PIPE_C].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORC);
3737 
3738 		tmp = intel_de_read(display, DSPHOWM);
3739 		wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
3740 		wm->pipe[PIPE_C].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
3741 		wm->pipe[PIPE_C].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
3742 		wm->pipe[PIPE_C].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEC_HI) << 8;
3743 		wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
3744 		wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
3745 		wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
3746 		wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
3747 		wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
3748 		wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
3749 	} else {
3750 		tmp = intel_de_read(display, DSPFW7);
3751 		wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
3752 		wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
3753 
3754 		tmp = intel_de_read(display, DSPHOWM);
3755 		wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
3756 		wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
3757 		wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
3758 		wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
3759 		wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
3760 		wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
3761 		wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
3762 	}
3763 }
3764 
3765 #undef _FW_WM
3766 #undef _FW_WM_VLV
3767 
3768 static void g4x_wm_get_hw_state(struct intel_display *display)
3769 {
3770 	struct g4x_wm_values *wm = &display->wm.g4x;
3771 	struct intel_crtc *crtc;
3772 
3773 	g4x_read_wm_values(display, wm);
3774 
3775 	wm->cxsr = intel_de_read(display, FW_BLC_SELF) & FW_BLC_SELF_EN;
3776 
3777 	for_each_intel_crtc(display->drm, crtc) {
3778 		struct intel_crtc_state *crtc_state =
3779 			to_intel_crtc_state(crtc->base.state);
3780 		struct g4x_wm_state *active = &crtc->wm.active.g4x;
3781 		struct g4x_pipe_wm *raw;
3782 		enum pipe pipe = crtc->pipe;
3783 		enum plane_id plane_id;
3784 		int level, max_level;
3785 
3786 		active->cxsr = wm->cxsr;
3787 		active->hpll_en = wm->hpll_en;
3788 		active->fbc_en = wm->fbc_en;
3789 
3790 		active->sr = wm->sr;
3791 		active->hpll = wm->hpll;
3792 
3793 		for_each_plane_id_on_crtc(crtc, plane_id) {
3794 			active->wm.plane[plane_id] =
3795 				wm->pipe[pipe].plane[plane_id];
3796 		}
3797 
3798 		if (wm->cxsr && wm->hpll_en)
3799 			max_level = G4X_WM_LEVEL_HPLL;
3800 		else if (wm->cxsr)
3801 			max_level = G4X_WM_LEVEL_SR;
3802 		else
3803 			max_level = G4X_WM_LEVEL_NORMAL;
3804 
3805 		level = G4X_WM_LEVEL_NORMAL;
3806 		raw = &crtc_state->wm.g4x.raw[level];
3807 		for_each_plane_id_on_crtc(crtc, plane_id)
3808 			raw->plane[plane_id] = active->wm.plane[plane_id];
3809 
3810 		level = G4X_WM_LEVEL_SR;
3811 		if (level > max_level)
3812 			goto out;
3813 
3814 		raw = &crtc_state->wm.g4x.raw[level];
3815 		raw->plane[PLANE_PRIMARY] = active->sr.plane;
3816 		raw->plane[PLANE_CURSOR] = active->sr.cursor;
3817 		raw->plane[PLANE_SPRITE0] = 0;
3818 		raw->fbc = active->sr.fbc;
3819 
3820 		level = G4X_WM_LEVEL_HPLL;
3821 		if (level > max_level)
3822 			goto out;
3823 
3824 		raw = &crtc_state->wm.g4x.raw[level];
3825 		raw->plane[PLANE_PRIMARY] = active->hpll.plane;
3826 		raw->plane[PLANE_CURSOR] = active->hpll.cursor;
3827 		raw->plane[PLANE_SPRITE0] = 0;
3828 		raw->fbc = active->hpll.fbc;
3829 
3830 		level++;
3831 	out:
3832 		for_each_plane_id_on_crtc(crtc, plane_id)
3833 			g4x_raw_plane_wm_set(crtc_state, level,
3834 					     plane_id, USHRT_MAX);
3835 		g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
3836 
3837 		g4x_invalidate_wms(crtc, active, level);
3838 
3839 		crtc_state->wm.g4x.optimal = *active;
3840 		crtc_state->wm.g4x.intermediate = *active;
3841 
3842 		drm_dbg_kms(display->drm,
3843 			    "Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite=%d\n",
3844 			    pipe_name(pipe),
3845 			    wm->pipe[pipe].plane[PLANE_PRIMARY],
3846 			    wm->pipe[pipe].plane[PLANE_CURSOR],
3847 			    wm->pipe[pipe].plane[PLANE_SPRITE0]);
3848 	}
3849 
3850 	drm_dbg_kms(display->drm,
3851 		    "Initial SR watermarks: plane=%d, cursor=%d fbc=%d\n",
3852 		    wm->sr.plane, wm->sr.cursor, wm->sr.fbc);
3853 	drm_dbg_kms(display->drm,
3854 		    "Initial HPLL watermarks: plane=%d, SR cursor=%d fbc=%d\n",
3855 		    wm->hpll.plane, wm->hpll.cursor, wm->hpll.fbc);
3856 	drm_dbg_kms(display->drm, "Initial SR=%s HPLL=%s FBC=%s\n",
3857 		    str_yes_no(wm->cxsr), str_yes_no(wm->hpll_en),
3858 		    str_yes_no(wm->fbc_en));
3859 }
3860 
3861 static void g4x_wm_sanitize(struct intel_display *display)
3862 {
3863 	struct intel_plane *plane;
3864 	struct intel_crtc *crtc;
3865 
3866 	mutex_lock(&display->wm.wm_mutex);
3867 
3868 	for_each_intel_plane(display->drm, plane) {
3869 		struct intel_crtc *crtc =
3870 			intel_crtc_for_pipe(display, plane->pipe);
3871 		struct intel_crtc_state *crtc_state =
3872 			to_intel_crtc_state(crtc->base.state);
3873 		struct intel_plane_state *plane_state =
3874 			to_intel_plane_state(plane->base.state);
3875 		enum plane_id plane_id = plane->id;
3876 		int level;
3877 
3878 		if (plane_state->uapi.visible)
3879 			continue;
3880 
3881 		for (level = 0; level < display->wm.num_levels; level++) {
3882 			struct g4x_pipe_wm *raw =
3883 				&crtc_state->wm.g4x.raw[level];
3884 
3885 			raw->plane[plane_id] = 0;
3886 
3887 			if (plane_id == PLANE_PRIMARY)
3888 				raw->fbc = 0;
3889 		}
3890 	}
3891 
3892 	for_each_intel_crtc(display->drm, crtc) {
3893 		struct intel_crtc_state *crtc_state =
3894 			to_intel_crtc_state(crtc->base.state);
3895 		int ret;
3896 
3897 		ret = _g4x_compute_pipe_wm(crtc_state);
3898 		drm_WARN_ON(display->drm, ret);
3899 
3900 		crtc_state->wm.g4x.intermediate =
3901 			crtc_state->wm.g4x.optimal;
3902 		crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
3903 	}
3904 
3905 	g4x_program_watermarks(display);
3906 
3907 	mutex_unlock(&display->wm.wm_mutex);
3908 }
3909 
3910 static void vlv_wm_get_hw_state(struct intel_display *display)
3911 {
3912 	struct vlv_wm_values *wm = &display->wm.vlv;
3913 	struct intel_crtc *crtc;
3914 	u32 val;
3915 	int ret;
3916 
3917 	vlv_read_wm_values(display, wm);
3918 
3919 	wm->cxsr = intel_de_read(display, FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
3920 	wm->level = VLV_WM_LEVEL_PM2;
3921 
3922 	if (display->platform.cherryview) {
3923 		vlv_punit_get(display->drm);
3924 
3925 		val = vlv_punit_read(display->drm, PUNIT_REG_DSPSSPM);
3926 		if (val & DSP_MAXFIFO_PM5_ENABLE)
3927 			wm->level = VLV_WM_LEVEL_PM5;
3928 
3929 		/*
3930 		 * If DDR DVFS is disabled in the BIOS, Punit
3931 		 * will never ack the request. So if that happens
3932 		 * assume we don't have to enable/disable DDR DVFS
3933 		 * dynamically. To test that just set the REQ_ACK
3934 		 * bit to poke the Punit, but don't change the
3935 		 * HIGH/LOW bits so that we don't actually change
3936 		 * the current state.
3937 		 */
3938 		val = vlv_punit_read(display->drm, PUNIT_REG_DDR_SETUP2);
3939 		val |= FORCE_DDR_FREQ_REQ_ACK;
3940 		vlv_punit_write(display->drm, PUNIT_REG_DDR_SETUP2, val);
3941 
3942 		ret = poll_timeout_us(val = vlv_punit_read(display->drm, PUNIT_REG_DDR_SETUP2),
3943 				      (val & FORCE_DDR_FREQ_REQ_ACK) == 0,
3944 				      500, 3000, false);
3945 		if (ret) {
3946 			drm_dbg_kms(display->drm,
3947 				    "Punit not acking DDR DVFS request, "
3948 				    "assuming DDR DVFS is disabled\n");
3949 			display->wm.num_levels = VLV_WM_LEVEL_PM5 + 1;
3950 		} else {
3951 			val = vlv_punit_read(display->drm, PUNIT_REG_DDR_SETUP2);
3952 			if ((val & FORCE_DDR_HIGH_FREQ) == 0)
3953 				wm->level = VLV_WM_LEVEL_DDR_DVFS;
3954 		}
3955 
3956 		vlv_punit_put(display->drm);
3957 	}
3958 
3959 	for_each_intel_crtc(display->drm, crtc) {
3960 		struct intel_crtc_state *crtc_state =
3961 			to_intel_crtc_state(crtc->base.state);
3962 		struct vlv_wm_state *active = &crtc->wm.active.vlv;
3963 		const struct vlv_fifo_state *fifo_state =
3964 			&crtc_state->wm.vlv.fifo_state;
3965 		enum pipe pipe = crtc->pipe;
3966 		enum plane_id plane_id;
3967 		int level;
3968 
3969 		vlv_get_fifo_size(crtc_state);
3970 
3971 		active->num_levels = wm->level + 1;
3972 		active->cxsr = wm->cxsr;
3973 
3974 		for (level = 0; level < active->num_levels; level++) {
3975 			struct g4x_pipe_wm *raw =
3976 				&crtc_state->wm.vlv.raw[level];
3977 
3978 			active->sr[level].plane = wm->sr.plane;
3979 			active->sr[level].cursor = wm->sr.cursor;
3980 
3981 			for_each_plane_id_on_crtc(crtc, plane_id) {
3982 				active->wm[level].plane[plane_id] =
3983 					wm->pipe[pipe].plane[plane_id];
3984 
3985 				raw->plane[plane_id] =
3986 					vlv_invert_wm_value(active->wm[level].plane[plane_id],
3987 							    fifo_state->plane[plane_id]);
3988 			}
3989 		}
3990 
3991 		for_each_plane_id_on_crtc(crtc, plane_id)
3992 			vlv_raw_plane_wm_set(crtc_state, level,
3993 					     plane_id, USHRT_MAX);
3994 		vlv_invalidate_wms(crtc, active, level);
3995 
3996 		crtc_state->wm.vlv.optimal = *active;
3997 		crtc_state->wm.vlv.intermediate = *active;
3998 
3999 		drm_dbg_kms(display->drm,
4000 			    "Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4001 			    pipe_name(pipe),
4002 			    wm->pipe[pipe].plane[PLANE_PRIMARY],
4003 			    wm->pipe[pipe].plane[PLANE_CURSOR],
4004 			    wm->pipe[pipe].plane[PLANE_SPRITE0],
4005 			    wm->pipe[pipe].plane[PLANE_SPRITE1]);
4006 	}
4007 
4008 	drm_dbg_kms(display->drm,
4009 		    "Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4010 		    wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
4011 }
4012 
4013 static void vlv_wm_sanitize(struct intel_display *display)
4014 {
4015 	struct intel_plane *plane;
4016 	struct intel_crtc *crtc;
4017 
4018 	mutex_lock(&display->wm.wm_mutex);
4019 
4020 	for_each_intel_plane(display->drm, plane) {
4021 		struct intel_crtc *crtc =
4022 			intel_crtc_for_pipe(display, plane->pipe);
4023 		struct intel_crtc_state *crtc_state =
4024 			to_intel_crtc_state(crtc->base.state);
4025 		struct intel_plane_state *plane_state =
4026 			to_intel_plane_state(plane->base.state);
4027 		enum plane_id plane_id = plane->id;
4028 		int level;
4029 
4030 		if (plane_state->uapi.visible)
4031 			continue;
4032 
4033 		for (level = 0; level < display->wm.num_levels; level++) {
4034 			struct g4x_pipe_wm *raw =
4035 				&crtc_state->wm.vlv.raw[level];
4036 
4037 			raw->plane[plane_id] = 0;
4038 		}
4039 	}
4040 
4041 	for_each_intel_crtc(display->drm, crtc) {
4042 		struct intel_crtc_state *crtc_state =
4043 			to_intel_crtc_state(crtc->base.state);
4044 		int ret;
4045 
4046 		ret = _vlv_compute_pipe_wm(crtc_state);
4047 		drm_WARN_ON(display->drm, ret);
4048 
4049 		crtc_state->wm.vlv.intermediate =
4050 			crtc_state->wm.vlv.optimal;
4051 		crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
4052 	}
4053 
4054 	vlv_program_watermarks(display);
4055 
4056 	mutex_unlock(&display->wm.wm_mutex);
4057 }
4058 
4059 /*
4060  * FIXME should probably kill this and improve
4061  * the real watermark readout/sanitation instead
4062  */
4063 static void ilk_init_lp_watermarks(struct intel_display *display)
4064 {
4065 	intel_de_rmw(display, WM3_LP_ILK, WM_LP_ENABLE, 0);
4066 	intel_de_rmw(display, WM2_LP_ILK, WM_LP_ENABLE, 0);
4067 	intel_de_rmw(display, WM1_LP_ILK, WM_LP_ENABLE, 0);
4068 
4069 	/*
4070 	 * Don't touch WM_LP_SPRITE_ENABLE here.
4071 	 * Doing so could cause underruns.
4072 	 */
4073 }
4074 
4075 static void ilk_wm_get_hw_state(struct intel_display *display)
4076 {
4077 	struct ilk_wm_values *hw = &display->wm.hw;
4078 	struct intel_crtc *crtc;
4079 
4080 	ilk_init_lp_watermarks(display);
4081 
4082 	for_each_intel_crtc(display->drm, crtc)
4083 		ilk_pipe_wm_get_hw_state(crtc);
4084 
4085 	hw->wm_lp[0] = intel_de_read(display, WM1_LP_ILK);
4086 	hw->wm_lp[1] = intel_de_read(display, WM2_LP_ILK);
4087 	hw->wm_lp[2] = intel_de_read(display, WM3_LP_ILK);
4088 
4089 	hw->wm_lp_spr[0] = intel_de_read(display, WM1S_LP_ILK);
4090 	if (DISPLAY_VER(display) >= 7) {
4091 		hw->wm_lp_spr[1] = intel_de_read(display, WM2S_LP_IVB);
4092 		hw->wm_lp_spr[2] = intel_de_read(display, WM3S_LP_IVB);
4093 	}
4094 
4095 	if (display->platform.haswell || display->platform.broadwell)
4096 		hw->partitioning = (intel_de_read(display, WM_MISC) &
4097 				    WM_MISC_DATA_PARTITION_5_6) ?
4098 			INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4099 	else if (display->platform.ivybridge)
4100 		hw->partitioning = (intel_de_read(display, DISP_ARB_CTL2) &
4101 				    DISP_DATA_PARTITION_5_6) ?
4102 			INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4103 
4104 	hw->enable_fbc_wm =
4105 		!(intel_de_read(display, DISP_ARB_CTL) & DISP_FBC_WM_DIS);
4106 }
4107 
4108 static const struct intel_wm_funcs ilk_wm_funcs = {
4109 	.compute_watermarks = ilk_compute_watermarks,
4110 	.initial_watermarks = ilk_initial_watermarks,
4111 	.optimize_watermarks = ilk_optimize_watermarks,
4112 	.get_hw_state = ilk_wm_get_hw_state,
4113 };
4114 
4115 static const struct intel_wm_funcs vlv_wm_funcs = {
4116 	.compute_watermarks = vlv_compute_watermarks,
4117 	.initial_watermarks = vlv_initial_watermarks,
4118 	.optimize_watermarks = vlv_optimize_watermarks,
4119 	.atomic_update_watermarks = vlv_atomic_update_fifo,
4120 	.get_hw_state = vlv_wm_get_hw_state,
4121 	.sanitize = vlv_wm_sanitize,
4122 };
4123 
4124 static const struct intel_wm_funcs g4x_wm_funcs = {
4125 	.compute_watermarks = g4x_compute_watermarks,
4126 	.initial_watermarks = g4x_initial_watermarks,
4127 	.optimize_watermarks = g4x_optimize_watermarks,
4128 	.get_hw_state = g4x_wm_get_hw_state,
4129 	.sanitize = g4x_wm_sanitize,
4130 };
4131 
4132 static const struct intel_wm_funcs pnv_wm_funcs = {
4133 	.compute_watermarks = i9xx_compute_watermarks,
4134 	.update_wm = pnv_update_wm,
4135 };
4136 
4137 static const struct intel_wm_funcs i965_wm_funcs = {
4138 	.compute_watermarks = i9xx_compute_watermarks,
4139 	.update_wm = i965_update_wm,
4140 };
4141 
4142 static const struct intel_wm_funcs i9xx_wm_funcs = {
4143 	.compute_watermarks = i9xx_compute_watermarks,
4144 	.update_wm = i9xx_update_wm,
4145 };
4146 
4147 static const struct intel_wm_funcs i845_wm_funcs = {
4148 	.compute_watermarks = i9xx_compute_watermarks,
4149 	.update_wm = i845_update_wm,
4150 };
4151 
4152 static const struct intel_wm_funcs nop_funcs = {
4153 };
4154 
4155 void i9xx_wm_init(struct intel_display *display)
4156 {
4157 	/* For FIFO watermark updates */
4158 	if (HAS_PCH_SPLIT(display)) {
4159 		ilk_setup_wm_latency(display);
4160 		display->funcs.wm = &ilk_wm_funcs;
4161 	} else if (display->platform.valleyview || display->platform.cherryview) {
4162 		vlv_setup_wm_latency(display);
4163 		display->funcs.wm = &vlv_wm_funcs;
4164 	} else if (display->platform.g4x) {
4165 		g4x_setup_wm_latency(display);
4166 		display->funcs.wm = &g4x_wm_funcs;
4167 	} else if (display->platform.pineview) {
4168 		if (!pnv_get_cxsr_latency(display)) {
4169 			drm_info(display->drm, "Unknown FSB/MEM, disabling CxSR\n");
4170 			/* Disable CxSR and never update its watermark again */
4171 			intel_set_memory_cxsr(display, false);
4172 			display->funcs.wm = &nop_funcs;
4173 		} else {
4174 			display->funcs.wm = &pnv_wm_funcs;
4175 		}
4176 	} else if (DISPLAY_VER(display) == 4) {
4177 		display->funcs.wm = &i965_wm_funcs;
4178 	} else if (DISPLAY_VER(display) == 3) {
4179 		display->funcs.wm = &i9xx_wm_funcs;
4180 	} else if (DISPLAY_VER(display) == 2) {
4181 		if (INTEL_NUM_PIPES(display) == 1)
4182 			display->funcs.wm = &i845_wm_funcs;
4183 		else
4184 			display->funcs.wm = &i9xx_wm_funcs;
4185 	} else {
4186 		drm_err(display->drm,
4187 			"unexpected fall-through in %s\n", __func__);
4188 		display->funcs.wm = &nop_funcs;
4189 	}
4190 }
4191