xref: /openbmc/linux/drivers/gpu/drm/vc4/vc4_hvs.c (revision 29c37341)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2015 Broadcom
4  */
5 
6 /**
7  * DOC: VC4 HVS module.
8  *
9  * The Hardware Video Scaler (HVS) is the piece of hardware that does
10  * translation, scaling, colorspace conversion, and compositing of
11  * pixels stored in framebuffers into a FIFO of pixels going out to
12  * the Pixel Valve (CRTC).  It operates at the system clock rate (the
13  * system audio clock gate, specifically), which is much higher than
14  * the pixel clock rate.
15  *
16  * There is a single global HVS, with multiple output FIFOs that can
17  * be consumed by the PVs.  This file just manages the resources for
18  * the HVS, while the vc4_crtc.c code actually drives HVS setup for
19  * each CRTC.
20  */
21 
22 #include <linux/component.h>
23 #include <linux/platform_device.h>
24 
25 #include <drm/drm_atomic_helper.h>
26 #include <drm/drm_vblank.h>
27 
28 #include "vc4_drv.h"
29 #include "vc4_regs.h"
30 
31 static const struct debugfs_reg32 hvs_regs[] = {
32 	VC4_REG32(SCALER_DISPCTRL),
33 	VC4_REG32(SCALER_DISPSTAT),
34 	VC4_REG32(SCALER_DISPID),
35 	VC4_REG32(SCALER_DISPECTRL),
36 	VC4_REG32(SCALER_DISPPROF),
37 	VC4_REG32(SCALER_DISPDITHER),
38 	VC4_REG32(SCALER_DISPEOLN),
39 	VC4_REG32(SCALER_DISPLIST0),
40 	VC4_REG32(SCALER_DISPLIST1),
41 	VC4_REG32(SCALER_DISPLIST2),
42 	VC4_REG32(SCALER_DISPLSTAT),
43 	VC4_REG32(SCALER_DISPLACT0),
44 	VC4_REG32(SCALER_DISPLACT1),
45 	VC4_REG32(SCALER_DISPLACT2),
46 	VC4_REG32(SCALER_DISPCTRL0),
47 	VC4_REG32(SCALER_DISPBKGND0),
48 	VC4_REG32(SCALER_DISPSTAT0),
49 	VC4_REG32(SCALER_DISPBASE0),
50 	VC4_REG32(SCALER_DISPCTRL1),
51 	VC4_REG32(SCALER_DISPBKGND1),
52 	VC4_REG32(SCALER_DISPSTAT1),
53 	VC4_REG32(SCALER_DISPBASE1),
54 	VC4_REG32(SCALER_DISPCTRL2),
55 	VC4_REG32(SCALER_DISPBKGND2),
56 	VC4_REG32(SCALER_DISPSTAT2),
57 	VC4_REG32(SCALER_DISPBASE2),
58 	VC4_REG32(SCALER_DISPALPHA2),
59 	VC4_REG32(SCALER_OLEDOFFS),
60 	VC4_REG32(SCALER_OLEDCOEF0),
61 	VC4_REG32(SCALER_OLEDCOEF1),
62 	VC4_REG32(SCALER_OLEDCOEF2),
63 };
64 
65 void vc4_hvs_dump_state(struct drm_device *dev)
66 {
67 	struct vc4_dev *vc4 = to_vc4_dev(dev);
68 	struct drm_printer p = drm_info_printer(&vc4->hvs->pdev->dev);
69 	int i;
70 
71 	drm_print_regset32(&p, &vc4->hvs->regset);
72 
73 	DRM_INFO("HVS ctx:\n");
74 	for (i = 0; i < 64; i += 4) {
75 		DRM_INFO("0x%08x (%s): 0x%08x 0x%08x 0x%08x 0x%08x\n",
76 			 i * 4, i < HVS_BOOTLOADER_DLIST_END ? "B" : "D",
77 			 readl((u32 __iomem *)vc4->hvs->dlist + i + 0),
78 			 readl((u32 __iomem *)vc4->hvs->dlist + i + 1),
79 			 readl((u32 __iomem *)vc4->hvs->dlist + i + 2),
80 			 readl((u32 __iomem *)vc4->hvs->dlist + i + 3));
81 	}
82 }
83 
84 static int vc4_hvs_debugfs_underrun(struct seq_file *m, void *data)
85 {
86 	struct drm_info_node *node = m->private;
87 	struct drm_device *dev = node->minor->dev;
88 	struct vc4_dev *vc4 = to_vc4_dev(dev);
89 	struct drm_printer p = drm_seq_file_printer(m);
90 
91 	drm_printf(&p, "%d\n", atomic_read(&vc4->underrun));
92 
93 	return 0;
94 }
95 
96 /* The filter kernel is composed of dwords each containing 3 9-bit
97  * signed integers packed next to each other.
98  */
99 #define VC4_INT_TO_COEFF(coeff) (coeff & 0x1ff)
100 #define VC4_PPF_FILTER_WORD(c0, c1, c2)				\
101 	((((c0) & 0x1ff) << 0) |				\
102 	 (((c1) & 0x1ff) << 9) |				\
103 	 (((c2) & 0x1ff) << 18))
104 
105 /* The whole filter kernel is arranged as the coefficients 0-16 going
106  * up, then a pad, then 17-31 going down and reversed within the
107  * dwords.  This means that a linear phase kernel (where it's
108  * symmetrical at the boundary between 15 and 16) has the last 5
109  * dwords matching the first 5, but reversed.
110  */
111 #define VC4_LINEAR_PHASE_KERNEL(c0, c1, c2, c3, c4, c5, c6, c7, c8,	\
112 				c9, c10, c11, c12, c13, c14, c15)	\
113 	{VC4_PPF_FILTER_WORD(c0, c1, c2),				\
114 	 VC4_PPF_FILTER_WORD(c3, c4, c5),				\
115 	 VC4_PPF_FILTER_WORD(c6, c7, c8),				\
116 	 VC4_PPF_FILTER_WORD(c9, c10, c11),				\
117 	 VC4_PPF_FILTER_WORD(c12, c13, c14),				\
118 	 VC4_PPF_FILTER_WORD(c15, c15, 0)}
119 
120 #define VC4_LINEAR_PHASE_KERNEL_DWORDS 6
121 #define VC4_KERNEL_DWORDS (VC4_LINEAR_PHASE_KERNEL_DWORDS * 2 - 1)
122 
123 /* Recommended B=1/3, C=1/3 filter choice from Mitchell/Netravali.
124  * http://www.cs.utexas.edu/~fussell/courses/cs384g/lectures/mitchell/Mitchell.pdf
125  */
126 static const u32 mitchell_netravali_1_3_1_3_kernel[] =
127 	VC4_LINEAR_PHASE_KERNEL(0, -2, -6, -8, -10, -8, -3, 2, 18,
128 				50, 82, 119, 155, 187, 213, 227);
129 
130 static int vc4_hvs_upload_linear_kernel(struct vc4_hvs *hvs,
131 					struct drm_mm_node *space,
132 					const u32 *kernel)
133 {
134 	int ret, i;
135 	u32 __iomem *dst_kernel;
136 
137 	ret = drm_mm_insert_node(&hvs->dlist_mm, space, VC4_KERNEL_DWORDS);
138 	if (ret) {
139 		DRM_ERROR("Failed to allocate space for filter kernel: %d\n",
140 			  ret);
141 		return ret;
142 	}
143 
144 	dst_kernel = hvs->dlist + space->start;
145 
146 	for (i = 0; i < VC4_KERNEL_DWORDS; i++) {
147 		if (i < VC4_LINEAR_PHASE_KERNEL_DWORDS)
148 			writel(kernel[i], &dst_kernel[i]);
149 		else {
150 			writel(kernel[VC4_KERNEL_DWORDS - i - 1],
151 			       &dst_kernel[i]);
152 		}
153 	}
154 
155 	return 0;
156 }
157 
158 static void vc4_hvs_lut_load(struct drm_crtc *crtc)
159 {
160 	struct drm_device *dev = crtc->dev;
161 	struct vc4_dev *vc4 = to_vc4_dev(dev);
162 	struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
163 	u32 i;
164 
165 	/* The LUT memory is laid out with each HVS channel in order,
166 	 * each of which takes 256 writes for R, 256 for G, then 256
167 	 * for B.
168 	 */
169 	HVS_WRITE(SCALER_GAMADDR,
170 		  SCALER_GAMADDR_AUTOINC |
171 		  (vc4_crtc->channel * 3 * crtc->gamma_size));
172 
173 	for (i = 0; i < crtc->gamma_size; i++)
174 		HVS_WRITE(SCALER_GAMDATA, vc4_crtc->lut_r[i]);
175 	for (i = 0; i < crtc->gamma_size; i++)
176 		HVS_WRITE(SCALER_GAMDATA, vc4_crtc->lut_g[i]);
177 	for (i = 0; i < crtc->gamma_size; i++)
178 		HVS_WRITE(SCALER_GAMDATA, vc4_crtc->lut_b[i]);
179 }
180 
181 static void vc4_hvs_update_gamma_lut(struct drm_crtc *crtc)
182 {
183 	struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
184 	struct drm_color_lut *lut = crtc->state->gamma_lut->data;
185 	u32 length = drm_color_lut_size(crtc->state->gamma_lut);
186 	u32 i;
187 
188 	for (i = 0; i < length; i++) {
189 		vc4_crtc->lut_r[i] = drm_color_lut_extract(lut[i].red, 8);
190 		vc4_crtc->lut_g[i] = drm_color_lut_extract(lut[i].green, 8);
191 		vc4_crtc->lut_b[i] = drm_color_lut_extract(lut[i].blue, 8);
192 	}
193 
194 	vc4_hvs_lut_load(crtc);
195 }
196 
197 int vc4_hvs_atomic_check(struct drm_crtc *crtc,
198 			 struct drm_crtc_state *state)
199 {
200 	struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(state);
201 	struct drm_device *dev = crtc->dev;
202 	struct vc4_dev *vc4 = to_vc4_dev(dev);
203 	struct drm_plane *plane;
204 	unsigned long flags;
205 	const struct drm_plane_state *plane_state;
206 	u32 dlist_count = 0;
207 	int ret;
208 
209 	/* The pixelvalve can only feed one encoder (and encoders are
210 	 * 1:1 with connectors.)
211 	 */
212 	if (hweight32(state->connector_mask) > 1)
213 		return -EINVAL;
214 
215 	drm_atomic_crtc_state_for_each_plane_state(plane, plane_state, state)
216 		dlist_count += vc4_plane_dlist_size(plane_state);
217 
218 	dlist_count++; /* Account for SCALER_CTL0_END. */
219 
220 	spin_lock_irqsave(&vc4->hvs->mm_lock, flags);
221 	ret = drm_mm_insert_node(&vc4->hvs->dlist_mm, &vc4_state->mm,
222 				 dlist_count);
223 	spin_unlock_irqrestore(&vc4->hvs->mm_lock, flags);
224 	if (ret)
225 		return ret;
226 
227 	return 0;
228 }
229 
230 static void vc4_hvs_update_dlist(struct drm_crtc *crtc)
231 {
232 	struct drm_device *dev = crtc->dev;
233 	struct vc4_dev *vc4 = to_vc4_dev(dev);
234 	struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
235 	struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc->state);
236 
237 	if (crtc->state->event) {
238 		unsigned long flags;
239 
240 		crtc->state->event->pipe = drm_crtc_index(crtc);
241 
242 		WARN_ON(drm_crtc_vblank_get(crtc) != 0);
243 
244 		spin_lock_irqsave(&dev->event_lock, flags);
245 
246 		if (!vc4_state->feed_txp || vc4_state->txp_armed) {
247 			vc4_crtc->event = crtc->state->event;
248 			crtc->state->event = NULL;
249 		}
250 
251 		HVS_WRITE(SCALER_DISPLISTX(vc4_crtc->channel),
252 			  vc4_state->mm.start);
253 
254 		spin_unlock_irqrestore(&dev->event_lock, flags);
255 	} else {
256 		HVS_WRITE(SCALER_DISPLISTX(vc4_crtc->channel),
257 			  vc4_state->mm.start);
258 	}
259 }
260 
261 void vc4_hvs_atomic_enable(struct drm_crtc *crtc,
262 			   struct drm_crtc_state *old_state)
263 {
264 	struct drm_device *dev = crtc->dev;
265 	struct vc4_dev *vc4 = to_vc4_dev(dev);
266 	struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
267 	struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc->state);
268 	struct drm_display_mode *mode = &crtc->state->adjusted_mode;
269 	bool oneshot = vc4_state->feed_txp;
270 	u32 dispctrl;
271 
272 	vc4_hvs_update_dlist(crtc);
273 
274 	/* Turn on the scaler, which will wait for vstart to start
275 	 * compositing.
276 	 * When feeding the transposer, we should operate in oneshot
277 	 * mode.
278 	 */
279 	dispctrl = SCALER_DISPCTRLX_ENABLE;
280 	dispctrl |= VC4_SET_FIELD(mode->hdisplay,
281 				  SCALER_DISPCTRLX_WIDTH) |
282 		    VC4_SET_FIELD(mode->vdisplay,
283 				  SCALER_DISPCTRLX_HEIGHT) |
284 		    (oneshot ? SCALER_DISPCTRLX_ONESHOT : 0);
285 
286 	HVS_WRITE(SCALER_DISPCTRLX(vc4_crtc->channel), dispctrl);
287 }
288 
289 void vc4_hvs_atomic_disable(struct drm_crtc *crtc,
290 			    struct drm_crtc_state *old_state)
291 {
292 	struct drm_device *dev = crtc->dev;
293 	struct vc4_dev *vc4 = to_vc4_dev(dev);
294 	struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
295 	u32 chan = vc4_crtc->channel;
296 
297 	if (HVS_READ(SCALER_DISPCTRLX(chan)) &
298 	    SCALER_DISPCTRLX_ENABLE) {
299 		HVS_WRITE(SCALER_DISPCTRLX(chan),
300 			  SCALER_DISPCTRLX_RESET);
301 
302 		/* While the docs say that reset is self-clearing, it
303 		 * seems it doesn't actually.
304 		 */
305 		HVS_WRITE(SCALER_DISPCTRLX(chan), 0);
306 	}
307 
308 	/* Once we leave, the scaler should be disabled and its fifo empty. */
309 
310 	WARN_ON_ONCE(HVS_READ(SCALER_DISPCTRLX(chan)) & SCALER_DISPCTRLX_RESET);
311 
312 	WARN_ON_ONCE(VC4_GET_FIELD(HVS_READ(SCALER_DISPSTATX(chan)),
313 				   SCALER_DISPSTATX_MODE) !=
314 		     SCALER_DISPSTATX_MODE_DISABLED);
315 
316 	WARN_ON_ONCE((HVS_READ(SCALER_DISPSTATX(chan)) &
317 		      (SCALER_DISPSTATX_FULL | SCALER_DISPSTATX_EMPTY)) !=
318 		     SCALER_DISPSTATX_EMPTY);
319 }
320 
321 void vc4_hvs_atomic_flush(struct drm_crtc *crtc,
322 			  struct drm_crtc_state *old_state)
323 {
324 	struct drm_device *dev = crtc->dev;
325 	struct vc4_dev *vc4 = to_vc4_dev(dev);
326 	struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
327 	struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc->state);
328 	struct drm_plane *plane;
329 	struct vc4_plane_state *vc4_plane_state;
330 	bool debug_dump_regs = false;
331 	bool enable_bg_fill = false;
332 	u32 __iomem *dlist_start = vc4->hvs->dlist + vc4_state->mm.start;
333 	u32 __iomem *dlist_next = dlist_start;
334 
335 	if (debug_dump_regs) {
336 		DRM_INFO("CRTC %d HVS before:\n", drm_crtc_index(crtc));
337 		vc4_hvs_dump_state(dev);
338 	}
339 
340 	/* Copy all the active planes' dlist contents to the hardware dlist. */
341 	drm_atomic_crtc_for_each_plane(plane, crtc) {
342 		/* Is this the first active plane? */
343 		if (dlist_next == dlist_start) {
344 			/* We need to enable background fill when a plane
345 			 * could be alpha blending from the background, i.e.
346 			 * where no other plane is underneath. It suffices to
347 			 * consider the first active plane here since we set
348 			 * needs_bg_fill such that either the first plane
349 			 * already needs it or all planes on top blend from
350 			 * the first or a lower plane.
351 			 */
352 			vc4_plane_state = to_vc4_plane_state(plane->state);
353 			enable_bg_fill = vc4_plane_state->needs_bg_fill;
354 		}
355 
356 		dlist_next += vc4_plane_write_dlist(plane, dlist_next);
357 	}
358 
359 	writel(SCALER_CTL0_END, dlist_next);
360 	dlist_next++;
361 
362 	WARN_ON_ONCE(dlist_next - dlist_start != vc4_state->mm.size);
363 
364 	if (enable_bg_fill)
365 		/* This sets a black background color fill, as is the case
366 		 * with other DRM drivers.
367 		 */
368 		HVS_WRITE(SCALER_DISPBKGNDX(vc4_crtc->channel),
369 			  HVS_READ(SCALER_DISPBKGNDX(vc4_crtc->channel)) |
370 			  SCALER_DISPBKGND_FILL);
371 
372 	/* Only update DISPLIST if the CRTC was already running and is not
373 	 * being disabled.
374 	 * vc4_crtc_enable() takes care of updating the dlist just after
375 	 * re-enabling VBLANK interrupts and before enabling the engine.
376 	 * If the CRTC is being disabled, there's no point in updating this
377 	 * information.
378 	 */
379 	if (crtc->state->active && old_state->active)
380 		vc4_hvs_update_dlist(crtc);
381 
382 	if (crtc->state->color_mgmt_changed) {
383 		u32 dispbkgndx = HVS_READ(SCALER_DISPBKGNDX(vc4_crtc->channel));
384 
385 		if (crtc->state->gamma_lut) {
386 			vc4_hvs_update_gamma_lut(crtc);
387 			dispbkgndx |= SCALER_DISPBKGND_GAMMA;
388 		} else {
389 			/* Unsetting DISPBKGND_GAMMA skips the gamma lut step
390 			 * in hardware, which is the same as a linear lut that
391 			 * DRM expects us to use in absence of a user lut.
392 			 */
393 			dispbkgndx &= ~SCALER_DISPBKGND_GAMMA;
394 		}
395 		HVS_WRITE(SCALER_DISPBKGNDX(vc4_crtc->channel), dispbkgndx);
396 	}
397 
398 	if (debug_dump_regs) {
399 		DRM_INFO("CRTC %d HVS after:\n", drm_crtc_index(crtc));
400 		vc4_hvs_dump_state(dev);
401 	}
402 }
403 
404 void vc4_hvs_mode_set_nofb(struct drm_crtc *crtc)
405 {
406 	struct drm_device *dev = crtc->dev;
407 	struct vc4_dev *vc4 = to_vc4_dev(dev);
408 	struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
409 	struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc->state);
410 	struct drm_display_mode *mode = &crtc->state->adjusted_mode;
411 	bool interlace = mode->flags & DRM_MODE_FLAG_INTERLACE;
412 
413 	if (vc4_crtc->data->hvs_channel == 2) {
414 		u32 dispctrl;
415 		u32 dsp3_mux;
416 
417 		/*
418 		 * SCALER_DISPCTRL_DSP3 = X, where X < 2 means 'connect DSP3 to
419 		 * FIFO X'.
420 		 * SCALER_DISPCTRL_DSP3 = 3 means 'disable DSP 3'.
421 		 *
422 		 * DSP3 is connected to FIFO2 unless the transposer is
423 		 * enabled. In this case, FIFO 2 is directly accessed by the
424 		 * TXP IP, and we need to disable the FIFO2 -> pixelvalve1
425 		 * route.
426 		 */
427 		if (vc4_state->feed_txp)
428 			dsp3_mux = VC4_SET_FIELD(3, SCALER_DISPCTRL_DSP3_MUX);
429 		else
430 			dsp3_mux = VC4_SET_FIELD(2, SCALER_DISPCTRL_DSP3_MUX);
431 
432 		dispctrl = HVS_READ(SCALER_DISPCTRL) &
433 			   ~SCALER_DISPCTRL_DSP3_MUX_MASK;
434 		HVS_WRITE(SCALER_DISPCTRL, dispctrl | dsp3_mux);
435 	}
436 
437 	HVS_WRITE(SCALER_DISPBKGNDX(vc4_crtc->channel),
438 		  SCALER_DISPBKGND_AUTOHS |
439 		  SCALER_DISPBKGND_GAMMA |
440 		  (interlace ? SCALER_DISPBKGND_INTERLACE : 0));
441 
442 	/* Reload the LUT, since the SRAMs would have been disabled if
443 	 * all CRTCs had SCALER_DISPBKGND_GAMMA unset at once.
444 	 */
445 	vc4_hvs_lut_load(crtc);
446 }
447 
448 void vc4_hvs_mask_underrun(struct drm_device *dev, int channel)
449 {
450 	struct vc4_dev *vc4 = to_vc4_dev(dev);
451 	u32 dispctrl = HVS_READ(SCALER_DISPCTRL);
452 
453 	dispctrl &= ~SCALER_DISPCTRL_DSPEISLUR(channel);
454 
455 	HVS_WRITE(SCALER_DISPCTRL, dispctrl);
456 }
457 
458 void vc4_hvs_unmask_underrun(struct drm_device *dev, int channel)
459 {
460 	struct vc4_dev *vc4 = to_vc4_dev(dev);
461 	u32 dispctrl = HVS_READ(SCALER_DISPCTRL);
462 
463 	dispctrl |= SCALER_DISPCTRL_DSPEISLUR(channel);
464 
465 	HVS_WRITE(SCALER_DISPSTAT,
466 		  SCALER_DISPSTAT_EUFLOW(channel));
467 	HVS_WRITE(SCALER_DISPCTRL, dispctrl);
468 }
469 
470 static void vc4_hvs_report_underrun(struct drm_device *dev)
471 {
472 	struct vc4_dev *vc4 = to_vc4_dev(dev);
473 
474 	atomic_inc(&vc4->underrun);
475 	DRM_DEV_ERROR(dev->dev, "HVS underrun\n");
476 }
477 
478 static irqreturn_t vc4_hvs_irq_handler(int irq, void *data)
479 {
480 	struct drm_device *dev = data;
481 	struct vc4_dev *vc4 = to_vc4_dev(dev);
482 	irqreturn_t irqret = IRQ_NONE;
483 	int channel;
484 	u32 control;
485 	u32 status;
486 
487 	status = HVS_READ(SCALER_DISPSTAT);
488 	control = HVS_READ(SCALER_DISPCTRL);
489 
490 	for (channel = 0; channel < SCALER_CHANNELS_COUNT; channel++) {
491 		/* Interrupt masking is not always honored, so check it here. */
492 		if (status & SCALER_DISPSTAT_EUFLOW(channel) &&
493 		    control & SCALER_DISPCTRL_DSPEISLUR(channel)) {
494 			vc4_hvs_mask_underrun(dev, channel);
495 			vc4_hvs_report_underrun(dev);
496 
497 			irqret = IRQ_HANDLED;
498 		}
499 	}
500 
501 	/* Clear every per-channel interrupt flag. */
502 	HVS_WRITE(SCALER_DISPSTAT, SCALER_DISPSTAT_IRQMASK(0) |
503 				   SCALER_DISPSTAT_IRQMASK(1) |
504 				   SCALER_DISPSTAT_IRQMASK(2));
505 
506 	return irqret;
507 }
508 
509 static int vc4_hvs_bind(struct device *dev, struct device *master, void *data)
510 {
511 	struct platform_device *pdev = to_platform_device(dev);
512 	struct drm_device *drm = dev_get_drvdata(master);
513 	struct vc4_dev *vc4 = drm->dev_private;
514 	struct vc4_hvs *hvs = NULL;
515 	int ret;
516 	u32 dispctrl;
517 
518 	hvs = devm_kzalloc(&pdev->dev, sizeof(*hvs), GFP_KERNEL);
519 	if (!hvs)
520 		return -ENOMEM;
521 
522 	hvs->pdev = pdev;
523 
524 	hvs->regs = vc4_ioremap_regs(pdev, 0);
525 	if (IS_ERR(hvs->regs))
526 		return PTR_ERR(hvs->regs);
527 
528 	hvs->regset.base = hvs->regs;
529 	hvs->regset.regs = hvs_regs;
530 	hvs->regset.nregs = ARRAY_SIZE(hvs_regs);
531 
532 	hvs->dlist = hvs->regs + SCALER_DLIST_START;
533 
534 	spin_lock_init(&hvs->mm_lock);
535 
536 	/* Set up the HVS display list memory manager.  We never
537 	 * overwrite the setup from the bootloader (just 128b out of
538 	 * our 16K), since we don't want to scramble the screen when
539 	 * transitioning from the firmware's boot setup to runtime.
540 	 */
541 	drm_mm_init(&hvs->dlist_mm,
542 		    HVS_BOOTLOADER_DLIST_END,
543 		    (SCALER_DLIST_SIZE >> 2) - HVS_BOOTLOADER_DLIST_END);
544 
545 	/* Set up the HVS LBM memory manager.  We could have some more
546 	 * complicated data structure that allowed reuse of LBM areas
547 	 * between planes when they don't overlap on the screen, but
548 	 * for now we just allocate globally.
549 	 */
550 	drm_mm_init(&hvs->lbm_mm, 0, 96 * 1024);
551 
552 	/* Upload filter kernels.  We only have the one for now, so we
553 	 * keep it around for the lifetime of the driver.
554 	 */
555 	ret = vc4_hvs_upload_linear_kernel(hvs,
556 					   &hvs->mitchell_netravali_filter,
557 					   mitchell_netravali_1_3_1_3_kernel);
558 	if (ret)
559 		return ret;
560 
561 	vc4->hvs = hvs;
562 
563 	dispctrl = HVS_READ(SCALER_DISPCTRL);
564 
565 	dispctrl |= SCALER_DISPCTRL_ENABLE;
566 	dispctrl |= SCALER_DISPCTRL_DISPEIRQ(0) |
567 		    SCALER_DISPCTRL_DISPEIRQ(1) |
568 		    SCALER_DISPCTRL_DISPEIRQ(2);
569 
570 	/* Set DSP3 (PV1) to use HVS channel 2, which would otherwise
571 	 * be unused.
572 	 */
573 	dispctrl &= ~SCALER_DISPCTRL_DSP3_MUX_MASK;
574 	dispctrl &= ~(SCALER_DISPCTRL_DMAEIRQ |
575 		      SCALER_DISPCTRL_SLVWREIRQ |
576 		      SCALER_DISPCTRL_SLVRDEIRQ |
577 		      SCALER_DISPCTRL_DSPEIEOF(0) |
578 		      SCALER_DISPCTRL_DSPEIEOF(1) |
579 		      SCALER_DISPCTRL_DSPEIEOF(2) |
580 		      SCALER_DISPCTRL_DSPEIEOLN(0) |
581 		      SCALER_DISPCTRL_DSPEIEOLN(1) |
582 		      SCALER_DISPCTRL_DSPEIEOLN(2) |
583 		      SCALER_DISPCTRL_DSPEISLUR(0) |
584 		      SCALER_DISPCTRL_DSPEISLUR(1) |
585 		      SCALER_DISPCTRL_DSPEISLUR(2) |
586 		      SCALER_DISPCTRL_SCLEIRQ);
587 	dispctrl |= VC4_SET_FIELD(2, SCALER_DISPCTRL_DSP3_MUX);
588 
589 	HVS_WRITE(SCALER_DISPCTRL, dispctrl);
590 
591 	ret = devm_request_irq(dev, platform_get_irq(pdev, 0),
592 			       vc4_hvs_irq_handler, 0, "vc4 hvs", drm);
593 	if (ret)
594 		return ret;
595 
596 	vc4_debugfs_add_regset32(drm, "hvs_regs", &hvs->regset);
597 	vc4_debugfs_add_file(drm, "hvs_underrun", vc4_hvs_debugfs_underrun,
598 			     NULL);
599 
600 	return 0;
601 }
602 
603 static void vc4_hvs_unbind(struct device *dev, struct device *master,
604 			   void *data)
605 {
606 	struct drm_device *drm = dev_get_drvdata(master);
607 	struct vc4_dev *vc4 = drm->dev_private;
608 
609 	if (drm_mm_node_allocated(&vc4->hvs->mitchell_netravali_filter))
610 		drm_mm_remove_node(&vc4->hvs->mitchell_netravali_filter);
611 
612 	drm_mm_takedown(&vc4->hvs->dlist_mm);
613 	drm_mm_takedown(&vc4->hvs->lbm_mm);
614 
615 	vc4->hvs = NULL;
616 }
617 
618 static const struct component_ops vc4_hvs_ops = {
619 	.bind   = vc4_hvs_bind,
620 	.unbind = vc4_hvs_unbind,
621 };
622 
623 static int vc4_hvs_dev_probe(struct platform_device *pdev)
624 {
625 	return component_add(&pdev->dev, &vc4_hvs_ops);
626 }
627 
628 static int vc4_hvs_dev_remove(struct platform_device *pdev)
629 {
630 	component_del(&pdev->dev, &vc4_hvs_ops);
631 	return 0;
632 }
633 
634 static const struct of_device_id vc4_hvs_dt_match[] = {
635 	{ .compatible = "brcm,bcm2835-hvs" },
636 	{}
637 };
638 
639 struct platform_driver vc4_hvs_driver = {
640 	.probe = vc4_hvs_dev_probe,
641 	.remove = vc4_hvs_dev_remove,
642 	.driver = {
643 		.name = "vc4_hvs",
644 		.of_match_table = vc4_hvs_dt_match,
645 	},
646 };
647