xref: /openbmc/linux/drivers/gpu/drm/vc4/vc4_plane.c (revision 519b58bb)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2015 Broadcom
4  */
5 
6 /**
7  * DOC: VC4 plane module
8  *
9  * Each DRM plane is a layer of pixels being scanned out by the HVS.
10  *
11  * At atomic modeset check time, we compute the HVS display element
12  * state that would be necessary for displaying the plane (giving us a
13  * chance to figure out if a plane configuration is invalid), then at
14  * atomic flush time the CRTC will ask us to write our element state
15  * into the region of the HVS that it has allocated for us.
16  */
17 
18 #include <drm/drm_atomic.h>
19 #include <drm/drm_atomic_helper.h>
20 #include <drm/drm_atomic_uapi.h>
21 #include <drm/drm_blend.h>
22 #include <drm/drm_drv.h>
23 #include <drm/drm_fb_dma_helper.h>
24 #include <drm/drm_fourcc.h>
25 #include <drm/drm_framebuffer.h>
26 #include <drm/drm_gem_atomic_helper.h>
27 
28 #include "uapi/drm/vc4_drm.h"
29 
30 #include "vc4_drv.h"
31 #include "vc4_regs.h"
32 
33 static const struct hvs_format {
34 	u32 drm; /* DRM_FORMAT_* */
35 	u32 hvs; /* HVS_FORMAT_* */
36 	u32 pixel_order;
37 	u32 pixel_order_hvs5;
38 	bool hvs5_only;
39 } hvs_formats[] = {
40 	{
41 		.drm = DRM_FORMAT_XRGB8888,
42 		.hvs = HVS_PIXEL_FORMAT_RGBA8888,
43 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
44 		.pixel_order_hvs5 = HVS_PIXEL_ORDER_ARGB,
45 	},
46 	{
47 		.drm = DRM_FORMAT_ARGB8888,
48 		.hvs = HVS_PIXEL_FORMAT_RGBA8888,
49 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
50 		.pixel_order_hvs5 = HVS_PIXEL_ORDER_ARGB,
51 	},
52 	{
53 		.drm = DRM_FORMAT_ABGR8888,
54 		.hvs = HVS_PIXEL_FORMAT_RGBA8888,
55 		.pixel_order = HVS_PIXEL_ORDER_ARGB,
56 		.pixel_order_hvs5 = HVS_PIXEL_ORDER_ABGR,
57 	},
58 	{
59 		.drm = DRM_FORMAT_XBGR8888,
60 		.hvs = HVS_PIXEL_FORMAT_RGBA8888,
61 		.pixel_order = HVS_PIXEL_ORDER_ARGB,
62 		.pixel_order_hvs5 = HVS_PIXEL_ORDER_ABGR,
63 	},
64 	{
65 		.drm = DRM_FORMAT_RGB565,
66 		.hvs = HVS_PIXEL_FORMAT_RGB565,
67 		.pixel_order = HVS_PIXEL_ORDER_XRGB,
68 	},
69 	{
70 		.drm = DRM_FORMAT_BGR565,
71 		.hvs = HVS_PIXEL_FORMAT_RGB565,
72 		.pixel_order = HVS_PIXEL_ORDER_XBGR,
73 	},
74 	{
75 		.drm = DRM_FORMAT_ARGB1555,
76 		.hvs = HVS_PIXEL_FORMAT_RGBA5551,
77 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
78 	},
79 	{
80 		.drm = DRM_FORMAT_XRGB1555,
81 		.hvs = HVS_PIXEL_FORMAT_RGBA5551,
82 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
83 	},
84 	{
85 		.drm = DRM_FORMAT_RGB888,
86 		.hvs = HVS_PIXEL_FORMAT_RGB888,
87 		.pixel_order = HVS_PIXEL_ORDER_XRGB,
88 	},
89 	{
90 		.drm = DRM_FORMAT_BGR888,
91 		.hvs = HVS_PIXEL_FORMAT_RGB888,
92 		.pixel_order = HVS_PIXEL_ORDER_XBGR,
93 	},
94 	{
95 		.drm = DRM_FORMAT_YUV422,
96 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_3PLANE,
97 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
98 	},
99 	{
100 		.drm = DRM_FORMAT_YVU422,
101 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_3PLANE,
102 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
103 	},
104 	{
105 		.drm = DRM_FORMAT_YUV420,
106 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_3PLANE,
107 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
108 	},
109 	{
110 		.drm = DRM_FORMAT_YVU420,
111 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_3PLANE,
112 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
113 	},
114 	{
115 		.drm = DRM_FORMAT_NV12,
116 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_2PLANE,
117 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
118 	},
119 	{
120 		.drm = DRM_FORMAT_NV21,
121 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_2PLANE,
122 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
123 	},
124 	{
125 		.drm = DRM_FORMAT_NV16,
126 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_2PLANE,
127 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
128 	},
129 	{
130 		.drm = DRM_FORMAT_NV61,
131 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_2PLANE,
132 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
133 	},
134 	{
135 		.drm = DRM_FORMAT_P030,
136 		.hvs = HVS_PIXEL_FORMAT_YCBCR_10BIT,
137 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
138 		.hvs5_only = true,
139 	},
140 };
141 
142 static const struct hvs_format *vc4_get_hvs_format(u32 drm_format)
143 {
144 	unsigned i;
145 
146 	for (i = 0; i < ARRAY_SIZE(hvs_formats); i++) {
147 		if (hvs_formats[i].drm == drm_format)
148 			return &hvs_formats[i];
149 	}
150 
151 	return NULL;
152 }
153 
154 static enum vc4_scaling_mode vc4_get_scaling_mode(u32 src, u32 dst)
155 {
156 	if (dst == src)
157 		return VC4_SCALING_NONE;
158 	if (3 * dst >= 2 * src)
159 		return VC4_SCALING_PPF;
160 	else
161 		return VC4_SCALING_TPZ;
162 }
163 
164 static bool plane_enabled(struct drm_plane_state *state)
165 {
166 	return state->fb && !WARN_ON(!state->crtc);
167 }
168 
169 static struct drm_plane_state *vc4_plane_duplicate_state(struct drm_plane *plane)
170 {
171 	struct vc4_plane_state *vc4_state;
172 
173 	if (WARN_ON(!plane->state))
174 		return NULL;
175 
176 	vc4_state = kmemdup(plane->state, sizeof(*vc4_state), GFP_KERNEL);
177 	if (!vc4_state)
178 		return NULL;
179 
180 	memset(&vc4_state->lbm, 0, sizeof(vc4_state->lbm));
181 	vc4_state->dlist_initialized = 0;
182 
183 	__drm_atomic_helper_plane_duplicate_state(plane, &vc4_state->base);
184 
185 	if (vc4_state->dlist) {
186 		vc4_state->dlist = kmemdup(vc4_state->dlist,
187 					   vc4_state->dlist_count * 4,
188 					   GFP_KERNEL);
189 		if (!vc4_state->dlist) {
190 			kfree(vc4_state);
191 			return NULL;
192 		}
193 		vc4_state->dlist_size = vc4_state->dlist_count;
194 	}
195 
196 	return &vc4_state->base;
197 }
198 
199 static void vc4_plane_destroy_state(struct drm_plane *plane,
200 				    struct drm_plane_state *state)
201 {
202 	struct vc4_dev *vc4 = to_vc4_dev(plane->dev);
203 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
204 
205 	if (drm_mm_node_allocated(&vc4_state->lbm)) {
206 		unsigned long irqflags;
207 
208 		spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
209 		drm_mm_remove_node(&vc4_state->lbm);
210 		spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
211 	}
212 
213 	kfree(vc4_state->dlist);
214 	__drm_atomic_helper_plane_destroy_state(&vc4_state->base);
215 	kfree(state);
216 }
217 
218 /* Called during init to allocate the plane's atomic state. */
219 static void vc4_plane_reset(struct drm_plane *plane)
220 {
221 	struct vc4_plane_state *vc4_state;
222 
223 	WARN_ON(plane->state);
224 
225 	vc4_state = kzalloc(sizeof(*vc4_state), GFP_KERNEL);
226 	if (!vc4_state)
227 		return;
228 
229 	__drm_atomic_helper_plane_reset(plane, &vc4_state->base);
230 }
231 
232 static void vc4_dlist_counter_increment(struct vc4_plane_state *vc4_state)
233 {
234 	if (vc4_state->dlist_count == vc4_state->dlist_size) {
235 		u32 new_size = max(4u, vc4_state->dlist_count * 2);
236 		u32 *new_dlist = kmalloc_array(new_size, 4, GFP_KERNEL);
237 
238 		if (!new_dlist)
239 			return;
240 		memcpy(new_dlist, vc4_state->dlist, vc4_state->dlist_count * 4);
241 
242 		kfree(vc4_state->dlist);
243 		vc4_state->dlist = new_dlist;
244 		vc4_state->dlist_size = new_size;
245 	}
246 
247 	vc4_state->dlist_count++;
248 }
249 
250 static void vc4_dlist_write(struct vc4_plane_state *vc4_state, u32 val)
251 {
252 	unsigned int idx = vc4_state->dlist_count;
253 
254 	vc4_dlist_counter_increment(vc4_state);
255 	vc4_state->dlist[idx] = val;
256 }
257 
258 /* Returns the scl0/scl1 field based on whether the dimensions need to
259  * be up/down/non-scaled.
260  *
261  * This is a replication of a table from the spec.
262  */
263 static u32 vc4_get_scl_field(struct drm_plane_state *state, int plane)
264 {
265 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
266 
267 	switch (vc4_state->x_scaling[plane] << 2 | vc4_state->y_scaling[plane]) {
268 	case VC4_SCALING_PPF << 2 | VC4_SCALING_PPF:
269 		return SCALER_CTL0_SCL_H_PPF_V_PPF;
270 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_PPF:
271 		return SCALER_CTL0_SCL_H_TPZ_V_PPF;
272 	case VC4_SCALING_PPF << 2 | VC4_SCALING_TPZ:
273 		return SCALER_CTL0_SCL_H_PPF_V_TPZ;
274 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_TPZ:
275 		return SCALER_CTL0_SCL_H_TPZ_V_TPZ;
276 	case VC4_SCALING_PPF << 2 | VC4_SCALING_NONE:
277 		return SCALER_CTL0_SCL_H_PPF_V_NONE;
278 	case VC4_SCALING_NONE << 2 | VC4_SCALING_PPF:
279 		return SCALER_CTL0_SCL_H_NONE_V_PPF;
280 	case VC4_SCALING_NONE << 2 | VC4_SCALING_TPZ:
281 		return SCALER_CTL0_SCL_H_NONE_V_TPZ;
282 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_NONE:
283 		return SCALER_CTL0_SCL_H_TPZ_V_NONE;
284 	default:
285 	case VC4_SCALING_NONE << 2 | VC4_SCALING_NONE:
286 		/* The unity case is independently handled by
287 		 * SCALER_CTL0_UNITY.
288 		 */
289 		return 0;
290 	}
291 }
292 
293 static int vc4_plane_margins_adj(struct drm_plane_state *pstate)
294 {
295 	struct vc4_plane_state *vc4_pstate = to_vc4_plane_state(pstate);
296 	unsigned int left, right, top, bottom, adjhdisplay, adjvdisplay;
297 	struct drm_crtc_state *crtc_state;
298 
299 	crtc_state = drm_atomic_get_new_crtc_state(pstate->state,
300 						   pstate->crtc);
301 
302 	vc4_crtc_get_margins(crtc_state, &left, &right, &top, &bottom);
303 	if (!left && !right && !top && !bottom)
304 		return 0;
305 
306 	if (left + right >= crtc_state->mode.hdisplay ||
307 	    top + bottom >= crtc_state->mode.vdisplay)
308 		return -EINVAL;
309 
310 	adjhdisplay = crtc_state->mode.hdisplay - (left + right);
311 	vc4_pstate->crtc_x = DIV_ROUND_CLOSEST(vc4_pstate->crtc_x *
312 					       adjhdisplay,
313 					       crtc_state->mode.hdisplay);
314 	vc4_pstate->crtc_x += left;
315 	if (vc4_pstate->crtc_x > crtc_state->mode.hdisplay - right)
316 		vc4_pstate->crtc_x = crtc_state->mode.hdisplay - right;
317 
318 	adjvdisplay = crtc_state->mode.vdisplay - (top + bottom);
319 	vc4_pstate->crtc_y = DIV_ROUND_CLOSEST(vc4_pstate->crtc_y *
320 					       adjvdisplay,
321 					       crtc_state->mode.vdisplay);
322 	vc4_pstate->crtc_y += top;
323 	if (vc4_pstate->crtc_y > crtc_state->mode.vdisplay - bottom)
324 		vc4_pstate->crtc_y = crtc_state->mode.vdisplay - bottom;
325 
326 	vc4_pstate->crtc_w = DIV_ROUND_CLOSEST(vc4_pstate->crtc_w *
327 					       adjhdisplay,
328 					       crtc_state->mode.hdisplay);
329 	vc4_pstate->crtc_h = DIV_ROUND_CLOSEST(vc4_pstate->crtc_h *
330 					       adjvdisplay,
331 					       crtc_state->mode.vdisplay);
332 
333 	if (!vc4_pstate->crtc_w || !vc4_pstate->crtc_h)
334 		return -EINVAL;
335 
336 	return 0;
337 }
338 
339 static int vc4_plane_setup_clipping_and_scaling(struct drm_plane_state *state)
340 {
341 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
342 	struct drm_framebuffer *fb = state->fb;
343 	struct drm_gem_dma_object *bo = drm_fb_dma_get_gem_obj(fb, 0);
344 	int num_planes = fb->format->num_planes;
345 	struct drm_crtc_state *crtc_state;
346 	u32 h_subsample = fb->format->hsub;
347 	u32 v_subsample = fb->format->vsub;
348 	int i, ret;
349 
350 	crtc_state = drm_atomic_get_existing_crtc_state(state->state,
351 							state->crtc);
352 	if (!crtc_state) {
353 		DRM_DEBUG_KMS("Invalid crtc state\n");
354 		return -EINVAL;
355 	}
356 
357 	ret = drm_atomic_helper_check_plane_state(state, crtc_state, 1,
358 						  INT_MAX, true, true);
359 	if (ret)
360 		return ret;
361 
362 	for (i = 0; i < num_planes; i++)
363 		vc4_state->offsets[i] = bo->dma_addr + fb->offsets[i];
364 
365 	/*
366 	 * We don't support subpixel source positioning for scaling,
367 	 * but fractional coordinates can be generated by clipping
368 	 * so just round for now
369 	 */
370 	vc4_state->src_x = DIV_ROUND_CLOSEST(state->src.x1, 1 << 16);
371 	vc4_state->src_y = DIV_ROUND_CLOSEST(state->src.y1, 1 << 16);
372 	vc4_state->src_w[0] = DIV_ROUND_CLOSEST(state->src.x2, 1 << 16) - vc4_state->src_x;
373 	vc4_state->src_h[0] = DIV_ROUND_CLOSEST(state->src.y2, 1 << 16) - vc4_state->src_y;
374 
375 	vc4_state->crtc_x = state->dst.x1;
376 	vc4_state->crtc_y = state->dst.y1;
377 	vc4_state->crtc_w = state->dst.x2 - state->dst.x1;
378 	vc4_state->crtc_h = state->dst.y2 - state->dst.y1;
379 
380 	ret = vc4_plane_margins_adj(state);
381 	if (ret)
382 		return ret;
383 
384 	vc4_state->x_scaling[0] = vc4_get_scaling_mode(vc4_state->src_w[0],
385 						       vc4_state->crtc_w);
386 	vc4_state->y_scaling[0] = vc4_get_scaling_mode(vc4_state->src_h[0],
387 						       vc4_state->crtc_h);
388 
389 	vc4_state->is_unity = (vc4_state->x_scaling[0] == VC4_SCALING_NONE &&
390 			       vc4_state->y_scaling[0] == VC4_SCALING_NONE);
391 
392 	if (num_planes > 1) {
393 		vc4_state->is_yuv = true;
394 
395 		vc4_state->src_w[1] = vc4_state->src_w[0] / h_subsample;
396 		vc4_state->src_h[1] = vc4_state->src_h[0] / v_subsample;
397 
398 		vc4_state->x_scaling[1] =
399 			vc4_get_scaling_mode(vc4_state->src_w[1],
400 					     vc4_state->crtc_w);
401 		vc4_state->y_scaling[1] =
402 			vc4_get_scaling_mode(vc4_state->src_h[1],
403 					     vc4_state->crtc_h);
404 
405 		/* YUV conversion requires that horizontal scaling be enabled
406 		 * on the UV plane even if vc4_get_scaling_mode() returned
407 		 * VC4_SCALING_NONE (which can happen when the down-scaling
408 		 * ratio is 0.5). Let's force it to VC4_SCALING_PPF in this
409 		 * case.
410 		 */
411 		if (vc4_state->x_scaling[1] == VC4_SCALING_NONE)
412 			vc4_state->x_scaling[1] = VC4_SCALING_PPF;
413 	} else {
414 		vc4_state->is_yuv = false;
415 		vc4_state->x_scaling[1] = VC4_SCALING_NONE;
416 		vc4_state->y_scaling[1] = VC4_SCALING_NONE;
417 	}
418 
419 	return 0;
420 }
421 
422 static void vc4_write_tpz(struct vc4_plane_state *vc4_state, u32 src, u32 dst)
423 {
424 	u32 scale, recip;
425 
426 	scale = (1 << 16) * src / dst;
427 
428 	/* The specs note that while the reciprocal would be defined
429 	 * as (1<<32)/scale, ~0 is close enough.
430 	 */
431 	recip = ~0 / scale;
432 
433 	vc4_dlist_write(vc4_state,
434 			VC4_SET_FIELD(scale, SCALER_TPZ0_SCALE) |
435 			VC4_SET_FIELD(0, SCALER_TPZ0_IPHASE));
436 	vc4_dlist_write(vc4_state,
437 			VC4_SET_FIELD(recip, SCALER_TPZ1_RECIP));
438 }
439 
440 static void vc4_write_ppf(struct vc4_plane_state *vc4_state, u32 src, u32 dst)
441 {
442 	u32 scale = (1 << 16) * src / dst;
443 
444 	vc4_dlist_write(vc4_state,
445 			SCALER_PPF_AGC |
446 			VC4_SET_FIELD(scale, SCALER_PPF_SCALE) |
447 			VC4_SET_FIELD(0, SCALER_PPF_IPHASE));
448 }
449 
450 static u32 vc4_lbm_size(struct drm_plane_state *state)
451 {
452 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
453 	struct vc4_dev *vc4 = to_vc4_dev(state->plane->dev);
454 	u32 pix_per_line;
455 	u32 lbm;
456 
457 	/* LBM is not needed when there's no vertical scaling. */
458 	if (vc4_state->y_scaling[0] == VC4_SCALING_NONE &&
459 	    vc4_state->y_scaling[1] == VC4_SCALING_NONE)
460 		return 0;
461 
462 	/*
463 	 * This can be further optimized in the RGB/YUV444 case if the PPF
464 	 * decimation factor is between 0.5 and 1.0 by using crtc_w.
465 	 *
466 	 * It's not an issue though, since in that case since src_w[0] is going
467 	 * to be greater than or equal to crtc_w.
468 	 */
469 	if (vc4_state->x_scaling[0] == VC4_SCALING_TPZ)
470 		pix_per_line = vc4_state->crtc_w;
471 	else
472 		pix_per_line = vc4_state->src_w[0];
473 
474 	if (!vc4_state->is_yuv) {
475 		if (vc4_state->y_scaling[0] == VC4_SCALING_TPZ)
476 			lbm = pix_per_line * 8;
477 		else {
478 			/* In special cases, this multiplier might be 12. */
479 			lbm = pix_per_line * 16;
480 		}
481 	} else {
482 		/* There are cases for this going down to a multiplier
483 		 * of 2, but according to the firmware source, the
484 		 * table in the docs is somewhat wrong.
485 		 */
486 		lbm = pix_per_line * 16;
487 	}
488 
489 	/* Align it to 64 or 128 (hvs5) bytes */
490 	lbm = roundup(lbm, vc4->is_vc5 ? 128 : 64);
491 
492 	/* Each "word" of the LBM memory contains 2 or 4 (hvs5) pixels */
493 	lbm /= vc4->is_vc5 ? 4 : 2;
494 
495 	return lbm;
496 }
497 
498 static void vc4_write_scaling_parameters(struct drm_plane_state *state,
499 					 int channel)
500 {
501 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
502 
503 	/* Ch0 H-PPF Word 0: Scaling Parameters */
504 	if (vc4_state->x_scaling[channel] == VC4_SCALING_PPF) {
505 		vc4_write_ppf(vc4_state,
506 			      vc4_state->src_w[channel], vc4_state->crtc_w);
507 	}
508 
509 	/* Ch0 V-PPF Words 0-1: Scaling Parameters, Context */
510 	if (vc4_state->y_scaling[channel] == VC4_SCALING_PPF) {
511 		vc4_write_ppf(vc4_state,
512 			      vc4_state->src_h[channel], vc4_state->crtc_h);
513 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
514 	}
515 
516 	/* Ch0 H-TPZ Words 0-1: Scaling Parameters, Recip */
517 	if (vc4_state->x_scaling[channel] == VC4_SCALING_TPZ) {
518 		vc4_write_tpz(vc4_state,
519 			      vc4_state->src_w[channel], vc4_state->crtc_w);
520 	}
521 
522 	/* Ch0 V-TPZ Words 0-2: Scaling Parameters, Recip, Context */
523 	if (vc4_state->y_scaling[channel] == VC4_SCALING_TPZ) {
524 		vc4_write_tpz(vc4_state,
525 			      vc4_state->src_h[channel], vc4_state->crtc_h);
526 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
527 	}
528 }
529 
530 static void vc4_plane_calc_load(struct drm_plane_state *state)
531 {
532 	unsigned int hvs_load_shift, vrefresh, i;
533 	struct drm_framebuffer *fb = state->fb;
534 	struct vc4_plane_state *vc4_state;
535 	struct drm_crtc_state *crtc_state;
536 	unsigned int vscale_factor;
537 
538 	vc4_state = to_vc4_plane_state(state);
539 	crtc_state = drm_atomic_get_existing_crtc_state(state->state,
540 							state->crtc);
541 	vrefresh = drm_mode_vrefresh(&crtc_state->adjusted_mode);
542 
543 	/* The HVS is able to process 2 pixels/cycle when scaling the source,
544 	 * 4 pixels/cycle otherwise.
545 	 * Alpha blending step seems to be pipelined and it's always operating
546 	 * at 4 pixels/cycle, so the limiting aspect here seems to be the
547 	 * scaler block.
548 	 * HVS load is expressed in clk-cycles/sec (AKA Hz).
549 	 */
550 	if (vc4_state->x_scaling[0] != VC4_SCALING_NONE ||
551 	    vc4_state->x_scaling[1] != VC4_SCALING_NONE ||
552 	    vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
553 	    vc4_state->y_scaling[1] != VC4_SCALING_NONE)
554 		hvs_load_shift = 1;
555 	else
556 		hvs_load_shift = 2;
557 
558 	vc4_state->membus_load = 0;
559 	vc4_state->hvs_load = 0;
560 	for (i = 0; i < fb->format->num_planes; i++) {
561 		/* Even if the bandwidth/plane required for a single frame is
562 		 *
563 		 * vc4_state->src_w[i] * vc4_state->src_h[i] * cpp * vrefresh
564 		 *
565 		 * when downscaling, we have to read more pixels per line in
566 		 * the time frame reserved for a single line, so the bandwidth
567 		 * demand can be punctually higher. To account for that, we
568 		 * calculate the down-scaling factor and multiply the plane
569 		 * load by this number. We're likely over-estimating the read
570 		 * demand, but that's better than under-estimating it.
571 		 */
572 		vscale_factor = DIV_ROUND_UP(vc4_state->src_h[i],
573 					     vc4_state->crtc_h);
574 		vc4_state->membus_load += vc4_state->src_w[i] *
575 					  vc4_state->src_h[i] * vscale_factor *
576 					  fb->format->cpp[i];
577 		vc4_state->hvs_load += vc4_state->crtc_h * vc4_state->crtc_w;
578 	}
579 
580 	vc4_state->hvs_load *= vrefresh;
581 	vc4_state->hvs_load >>= hvs_load_shift;
582 	vc4_state->membus_load *= vrefresh;
583 }
584 
585 static int vc4_plane_allocate_lbm(struct drm_plane_state *state)
586 {
587 	struct vc4_dev *vc4 = to_vc4_dev(state->plane->dev);
588 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
589 	unsigned long irqflags;
590 	u32 lbm_size;
591 
592 	lbm_size = vc4_lbm_size(state);
593 	if (!lbm_size)
594 		return 0;
595 
596 	if (WARN_ON(!vc4_state->lbm_offset))
597 		return -EINVAL;
598 
599 	/* Allocate the LBM memory that the HVS will use for temporary
600 	 * storage due to our scaling/format conversion.
601 	 */
602 	if (!drm_mm_node_allocated(&vc4_state->lbm)) {
603 		int ret;
604 
605 		spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
606 		ret = drm_mm_insert_node_generic(&vc4->hvs->lbm_mm,
607 						 &vc4_state->lbm,
608 						 lbm_size,
609 						 vc4->is_vc5 ? 64 : 32,
610 						 0, 0);
611 		spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
612 
613 		if (ret)
614 			return ret;
615 	} else {
616 		WARN_ON_ONCE(lbm_size != vc4_state->lbm.size);
617 	}
618 
619 	vc4_state->dlist[vc4_state->lbm_offset] = vc4_state->lbm.start;
620 
621 	return 0;
622 }
623 
624 /*
625  * The colorspace conversion matrices are held in 3 entries in the dlist.
626  * Create an array of them, with entries for each full and limited mode, and
627  * each supported colorspace.
628  */
629 static const u32 colorspace_coeffs[2][DRM_COLOR_ENCODING_MAX][3] = {
630 	{
631 		/* Limited range */
632 		{
633 			/* BT601 */
634 			SCALER_CSC0_ITR_R_601_5,
635 			SCALER_CSC1_ITR_R_601_5,
636 			SCALER_CSC2_ITR_R_601_5,
637 		}, {
638 			/* BT709 */
639 			SCALER_CSC0_ITR_R_709_3,
640 			SCALER_CSC1_ITR_R_709_3,
641 			SCALER_CSC2_ITR_R_709_3,
642 		}, {
643 			/* BT2020 */
644 			SCALER_CSC0_ITR_R_2020,
645 			SCALER_CSC1_ITR_R_2020,
646 			SCALER_CSC2_ITR_R_2020,
647 		}
648 	}, {
649 		/* Full range */
650 		{
651 			/* JFIF */
652 			SCALER_CSC0_JPEG_JFIF,
653 			SCALER_CSC1_JPEG_JFIF,
654 			SCALER_CSC2_JPEG_JFIF,
655 		}, {
656 			/* BT709 */
657 			SCALER_CSC0_ITR_R_709_3_FR,
658 			SCALER_CSC1_ITR_R_709_3_FR,
659 			SCALER_CSC2_ITR_R_709_3_FR,
660 		}, {
661 			/* BT2020 */
662 			SCALER_CSC0_ITR_R_2020_FR,
663 			SCALER_CSC1_ITR_R_2020_FR,
664 			SCALER_CSC2_ITR_R_2020_FR,
665 		}
666 	}
667 };
668 
669 static u32 vc4_hvs4_get_alpha_blend_mode(struct drm_plane_state *state)
670 {
671 	if (!state->fb->format->has_alpha)
672 		return VC4_SET_FIELD(SCALER_POS2_ALPHA_MODE_FIXED,
673 				     SCALER_POS2_ALPHA_MODE);
674 
675 	switch (state->pixel_blend_mode) {
676 	case DRM_MODE_BLEND_PIXEL_NONE:
677 		return VC4_SET_FIELD(SCALER_POS2_ALPHA_MODE_FIXED,
678 				     SCALER_POS2_ALPHA_MODE);
679 	default:
680 	case DRM_MODE_BLEND_PREMULTI:
681 		return VC4_SET_FIELD(SCALER_POS2_ALPHA_MODE_PIPELINE,
682 				     SCALER_POS2_ALPHA_MODE) |
683 			SCALER_POS2_ALPHA_PREMULT;
684 	case DRM_MODE_BLEND_COVERAGE:
685 		return VC4_SET_FIELD(SCALER_POS2_ALPHA_MODE_PIPELINE,
686 				     SCALER_POS2_ALPHA_MODE);
687 	}
688 }
689 
690 static u32 vc4_hvs5_get_alpha_blend_mode(struct drm_plane_state *state)
691 {
692 	if (!state->fb->format->has_alpha)
693 		return VC4_SET_FIELD(SCALER5_CTL2_ALPHA_MODE_FIXED,
694 				     SCALER5_CTL2_ALPHA_MODE);
695 
696 	switch (state->pixel_blend_mode) {
697 	case DRM_MODE_BLEND_PIXEL_NONE:
698 		return VC4_SET_FIELD(SCALER5_CTL2_ALPHA_MODE_FIXED,
699 				     SCALER5_CTL2_ALPHA_MODE);
700 	default:
701 	case DRM_MODE_BLEND_PREMULTI:
702 		return VC4_SET_FIELD(SCALER5_CTL2_ALPHA_MODE_PIPELINE,
703 				     SCALER5_CTL2_ALPHA_MODE) |
704 			SCALER5_CTL2_ALPHA_PREMULT;
705 	case DRM_MODE_BLEND_COVERAGE:
706 		return VC4_SET_FIELD(SCALER5_CTL2_ALPHA_MODE_PIPELINE,
707 				     SCALER5_CTL2_ALPHA_MODE);
708 	}
709 }
710 
711 /* Writes out a full display list for an active plane to the plane's
712  * private dlist state.
713  */
714 static int vc4_plane_mode_set(struct drm_plane *plane,
715 			      struct drm_plane_state *state)
716 {
717 	struct vc4_dev *vc4 = to_vc4_dev(plane->dev);
718 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
719 	struct drm_framebuffer *fb = state->fb;
720 	u32 ctl0_offset = vc4_state->dlist_count;
721 	const struct hvs_format *format = vc4_get_hvs_format(fb->format->format);
722 	u64 base_format_mod = fourcc_mod_broadcom_mod(fb->modifier);
723 	int num_planes = fb->format->num_planes;
724 	u32 h_subsample = fb->format->hsub;
725 	u32 v_subsample = fb->format->vsub;
726 	bool mix_plane_alpha;
727 	bool covers_screen;
728 	u32 scl0, scl1, pitch0;
729 	u32 tiling, src_y;
730 	u32 hvs_format = format->hvs;
731 	unsigned int rotation;
732 	int ret, i;
733 
734 	if (vc4_state->dlist_initialized)
735 		return 0;
736 
737 	ret = vc4_plane_setup_clipping_and_scaling(state);
738 	if (ret)
739 		return ret;
740 
741 	/* SCL1 is used for Cb/Cr scaling of planar formats.  For RGB
742 	 * and 4:4:4, scl1 should be set to scl0 so both channels of
743 	 * the scaler do the same thing.  For YUV, the Y plane needs
744 	 * to be put in channel 1 and Cb/Cr in channel 0, so we swap
745 	 * the scl fields here.
746 	 */
747 	if (num_planes == 1) {
748 		scl0 = vc4_get_scl_field(state, 0);
749 		scl1 = scl0;
750 	} else {
751 		scl0 = vc4_get_scl_field(state, 1);
752 		scl1 = vc4_get_scl_field(state, 0);
753 	}
754 
755 	rotation = drm_rotation_simplify(state->rotation,
756 					 DRM_MODE_ROTATE_0 |
757 					 DRM_MODE_REFLECT_X |
758 					 DRM_MODE_REFLECT_Y);
759 
760 	/* We must point to the last line when Y reflection is enabled. */
761 	src_y = vc4_state->src_y;
762 	if (rotation & DRM_MODE_REFLECT_Y)
763 		src_y += vc4_state->src_h[0] - 1;
764 
765 	switch (base_format_mod) {
766 	case DRM_FORMAT_MOD_LINEAR:
767 		tiling = SCALER_CTL0_TILING_LINEAR;
768 		pitch0 = VC4_SET_FIELD(fb->pitches[0], SCALER_SRC_PITCH);
769 
770 		/* Adjust the base pointer to the first pixel to be scanned
771 		 * out.
772 		 */
773 		for (i = 0; i < num_planes; i++) {
774 			vc4_state->offsets[i] += src_y /
775 						 (i ? v_subsample : 1) *
776 						 fb->pitches[i];
777 
778 			vc4_state->offsets[i] += vc4_state->src_x /
779 						 (i ? h_subsample : 1) *
780 						 fb->format->cpp[i];
781 		}
782 
783 		break;
784 
785 	case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED: {
786 		u32 tile_size_shift = 12; /* T tiles are 4kb */
787 		/* Whole-tile offsets, mostly for setting the pitch. */
788 		u32 tile_w_shift = fb->format->cpp[0] == 2 ? 6 : 5;
789 		u32 tile_h_shift = 5; /* 16 and 32bpp are 32 pixels high */
790 		u32 tile_w_mask = (1 << tile_w_shift) - 1;
791 		/* The height mask on 32-bit-per-pixel tiles is 63, i.e. twice
792 		 * the height (in pixels) of a 4k tile.
793 		 */
794 		u32 tile_h_mask = (2 << tile_h_shift) - 1;
795 		/* For T-tiled, the FB pitch is "how many bytes from one row to
796 		 * the next, such that
797 		 *
798 		 *	pitch * tile_h == tile_size * tiles_per_row
799 		 */
800 		u32 tiles_w = fb->pitches[0] >> (tile_size_shift - tile_h_shift);
801 		u32 tiles_l = vc4_state->src_x >> tile_w_shift;
802 		u32 tiles_r = tiles_w - tiles_l;
803 		u32 tiles_t = src_y >> tile_h_shift;
804 		/* Intra-tile offsets, which modify the base address (the
805 		 * SCALER_PITCH0_TILE_Y_OFFSET tells HVS how to walk from that
806 		 * base address).
807 		 */
808 		u32 tile_y = (src_y >> 4) & 1;
809 		u32 subtile_y = (src_y >> 2) & 3;
810 		u32 utile_y = src_y & 3;
811 		u32 x_off = vc4_state->src_x & tile_w_mask;
812 		u32 y_off = src_y & tile_h_mask;
813 
814 		/* When Y reflection is requested we must set the
815 		 * SCALER_PITCH0_TILE_LINE_DIR flag to tell HVS that all lines
816 		 * after the initial one should be fetched in descending order,
817 		 * which makes sense since we start from the last line and go
818 		 * backward.
819 		 * Don't know why we need y_off = max_y_off - y_off, but it's
820 		 * definitely required (I guess it's also related to the "going
821 		 * backward" situation).
822 		 */
823 		if (rotation & DRM_MODE_REFLECT_Y) {
824 			y_off = tile_h_mask - y_off;
825 			pitch0 = SCALER_PITCH0_TILE_LINE_DIR;
826 		} else {
827 			pitch0 = 0;
828 		}
829 
830 		tiling = SCALER_CTL0_TILING_256B_OR_T;
831 		pitch0 |= (VC4_SET_FIELD(x_off, SCALER_PITCH0_SINK_PIX) |
832 			   VC4_SET_FIELD(y_off, SCALER_PITCH0_TILE_Y_OFFSET) |
833 			   VC4_SET_FIELD(tiles_l, SCALER_PITCH0_TILE_WIDTH_L) |
834 			   VC4_SET_FIELD(tiles_r, SCALER_PITCH0_TILE_WIDTH_R));
835 		vc4_state->offsets[0] += tiles_t * (tiles_w << tile_size_shift);
836 		vc4_state->offsets[0] += subtile_y << 8;
837 		vc4_state->offsets[0] += utile_y << 4;
838 
839 		/* Rows of tiles alternate left-to-right and right-to-left. */
840 		if (tiles_t & 1) {
841 			pitch0 |= SCALER_PITCH0_TILE_INITIAL_LINE_DIR;
842 			vc4_state->offsets[0] += (tiles_w - tiles_l) <<
843 						 tile_size_shift;
844 			vc4_state->offsets[0] -= (1 + !tile_y) << 10;
845 		} else {
846 			vc4_state->offsets[0] += tiles_l << tile_size_shift;
847 			vc4_state->offsets[0] += tile_y << 10;
848 		}
849 
850 		break;
851 	}
852 
853 	case DRM_FORMAT_MOD_BROADCOM_SAND64:
854 	case DRM_FORMAT_MOD_BROADCOM_SAND128:
855 	case DRM_FORMAT_MOD_BROADCOM_SAND256: {
856 		uint32_t param = fourcc_mod_broadcom_param(fb->modifier);
857 
858 		if (param > SCALER_TILE_HEIGHT_MASK) {
859 			DRM_DEBUG_KMS("SAND height too large (%d)\n",
860 				      param);
861 			return -EINVAL;
862 		}
863 
864 		if (fb->format->format == DRM_FORMAT_P030) {
865 			hvs_format = HVS_PIXEL_FORMAT_YCBCR_10BIT;
866 			tiling = SCALER_CTL0_TILING_128B;
867 		} else {
868 			hvs_format = HVS_PIXEL_FORMAT_H264;
869 
870 			switch (base_format_mod) {
871 			case DRM_FORMAT_MOD_BROADCOM_SAND64:
872 				tiling = SCALER_CTL0_TILING_64B;
873 				break;
874 			case DRM_FORMAT_MOD_BROADCOM_SAND128:
875 				tiling = SCALER_CTL0_TILING_128B;
876 				break;
877 			case DRM_FORMAT_MOD_BROADCOM_SAND256:
878 				tiling = SCALER_CTL0_TILING_256B_OR_T;
879 				break;
880 			default:
881 				return -EINVAL;
882 			}
883 		}
884 
885 		/* Adjust the base pointer to the first pixel to be scanned
886 		 * out.
887 		 *
888 		 * For P030, y_ptr [31:4] is the 128bit word for the start pixel
889 		 * y_ptr [3:0] is the pixel (0-11) contained within that 128bit
890 		 * word that should be taken as the first pixel.
891 		 * Ditto uv_ptr [31:4] vs [3:0], however [3:0] contains the
892 		 * element within the 128bit word, eg for pixel 3 the value
893 		 * should be 6.
894 		 */
895 		for (i = 0; i < num_planes; i++) {
896 			u32 tile_w, tile, x_off, pix_per_tile;
897 
898 			if (fb->format->format == DRM_FORMAT_P030) {
899 				/*
900 				 * Spec says: bits [31:4] of the given address
901 				 * should point to the 128-bit word containing
902 				 * the desired starting pixel, and bits[3:0]
903 				 * should be between 0 and 11, indicating which
904 				 * of the 12-pixels in that 128-bit word is the
905 				 * first pixel to be used
906 				 */
907 				u32 remaining_pixels = vc4_state->src_x % 96;
908 				u32 aligned = remaining_pixels / 12;
909 				u32 last_bits = remaining_pixels % 12;
910 
911 				x_off = aligned * 16 + last_bits;
912 				tile_w = 128;
913 				pix_per_tile = 96;
914 			} else {
915 				switch (base_format_mod) {
916 				case DRM_FORMAT_MOD_BROADCOM_SAND64:
917 					tile_w = 64;
918 					break;
919 				case DRM_FORMAT_MOD_BROADCOM_SAND128:
920 					tile_w = 128;
921 					break;
922 				case DRM_FORMAT_MOD_BROADCOM_SAND256:
923 					tile_w = 256;
924 					break;
925 				default:
926 					return -EINVAL;
927 				}
928 				pix_per_tile = tile_w / fb->format->cpp[0];
929 				x_off = (vc4_state->src_x % pix_per_tile) /
930 					(i ? h_subsample : 1) *
931 					fb->format->cpp[i];
932 			}
933 
934 			tile = vc4_state->src_x / pix_per_tile;
935 
936 			vc4_state->offsets[i] += param * tile_w * tile;
937 			vc4_state->offsets[i] += src_y /
938 						 (i ? v_subsample : 1) *
939 						 tile_w;
940 			vc4_state->offsets[i] += x_off & ~(i ? 1 : 0);
941 		}
942 
943 		pitch0 = VC4_SET_FIELD(param, SCALER_TILE_HEIGHT);
944 		break;
945 	}
946 
947 	default:
948 		DRM_DEBUG_KMS("Unsupported FB tiling flag 0x%16llx",
949 			      (long long)fb->modifier);
950 		return -EINVAL;
951 	}
952 
953 	/* Don't waste cycles mixing with plane alpha if the set alpha
954 	 * is opaque or there is no per-pixel alpha information.
955 	 * In any case we use the alpha property value as the fixed alpha.
956 	 */
957 	mix_plane_alpha = state->alpha != DRM_BLEND_ALPHA_OPAQUE &&
958 			  fb->format->has_alpha;
959 
960 	if (!vc4->is_vc5) {
961 	/* Control word */
962 		vc4_dlist_write(vc4_state,
963 				SCALER_CTL0_VALID |
964 				(rotation & DRM_MODE_REFLECT_X ? SCALER_CTL0_HFLIP : 0) |
965 				(rotation & DRM_MODE_REFLECT_Y ? SCALER_CTL0_VFLIP : 0) |
966 				VC4_SET_FIELD(SCALER_CTL0_RGBA_EXPAND_ROUND, SCALER_CTL0_RGBA_EXPAND) |
967 				(format->pixel_order << SCALER_CTL0_ORDER_SHIFT) |
968 				(hvs_format << SCALER_CTL0_PIXEL_FORMAT_SHIFT) |
969 				VC4_SET_FIELD(tiling, SCALER_CTL0_TILING) |
970 				(vc4_state->is_unity ? SCALER_CTL0_UNITY : 0) |
971 				VC4_SET_FIELD(scl0, SCALER_CTL0_SCL0) |
972 				VC4_SET_FIELD(scl1, SCALER_CTL0_SCL1));
973 
974 		/* Position Word 0: Image Positions and Alpha Value */
975 		vc4_state->pos0_offset = vc4_state->dlist_count;
976 		vc4_dlist_write(vc4_state,
977 				VC4_SET_FIELD(state->alpha >> 8, SCALER_POS0_FIXED_ALPHA) |
978 				VC4_SET_FIELD(vc4_state->crtc_x, SCALER_POS0_START_X) |
979 				VC4_SET_FIELD(vc4_state->crtc_y, SCALER_POS0_START_Y));
980 
981 		/* Position Word 1: Scaled Image Dimensions. */
982 		if (!vc4_state->is_unity) {
983 			vc4_dlist_write(vc4_state,
984 					VC4_SET_FIELD(vc4_state->crtc_w,
985 						      SCALER_POS1_SCL_WIDTH) |
986 					VC4_SET_FIELD(vc4_state->crtc_h,
987 						      SCALER_POS1_SCL_HEIGHT));
988 		}
989 
990 		/* Position Word 2: Source Image Size, Alpha */
991 		vc4_state->pos2_offset = vc4_state->dlist_count;
992 		vc4_dlist_write(vc4_state,
993 				(mix_plane_alpha ? SCALER_POS2_ALPHA_MIX : 0) |
994 				vc4_hvs4_get_alpha_blend_mode(state) |
995 				VC4_SET_FIELD(vc4_state->src_w[0],
996 					      SCALER_POS2_WIDTH) |
997 				VC4_SET_FIELD(vc4_state->src_h[0],
998 					      SCALER_POS2_HEIGHT));
999 
1000 		/* Position Word 3: Context.  Written by the HVS. */
1001 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
1002 
1003 	} else {
1004 		u32 hvs_pixel_order = format->pixel_order;
1005 
1006 		if (format->pixel_order_hvs5)
1007 			hvs_pixel_order = format->pixel_order_hvs5;
1008 
1009 		/* Control word */
1010 		vc4_dlist_write(vc4_state,
1011 				SCALER_CTL0_VALID |
1012 				(hvs_pixel_order << SCALER_CTL0_ORDER_SHIFT) |
1013 				(hvs_format << SCALER_CTL0_PIXEL_FORMAT_SHIFT) |
1014 				VC4_SET_FIELD(tiling, SCALER_CTL0_TILING) |
1015 				(vc4_state->is_unity ?
1016 						SCALER5_CTL0_UNITY : 0) |
1017 				VC4_SET_FIELD(scl0, SCALER_CTL0_SCL0) |
1018 				VC4_SET_FIELD(scl1, SCALER_CTL0_SCL1) |
1019 				SCALER5_CTL0_ALPHA_EXPAND |
1020 				SCALER5_CTL0_RGB_EXPAND);
1021 
1022 		/* Position Word 0: Image Positions and Alpha Value */
1023 		vc4_state->pos0_offset = vc4_state->dlist_count;
1024 		vc4_dlist_write(vc4_state,
1025 				(rotation & DRM_MODE_REFLECT_Y ?
1026 						SCALER5_POS0_VFLIP : 0) |
1027 				VC4_SET_FIELD(vc4_state->crtc_x,
1028 					      SCALER_POS0_START_X) |
1029 				(rotation & DRM_MODE_REFLECT_X ?
1030 					      SCALER5_POS0_HFLIP : 0) |
1031 				VC4_SET_FIELD(vc4_state->crtc_y,
1032 					      SCALER5_POS0_START_Y)
1033 			       );
1034 
1035 		/* Control Word 2 */
1036 		vc4_dlist_write(vc4_state,
1037 				VC4_SET_FIELD(state->alpha >> 4,
1038 					      SCALER5_CTL2_ALPHA) |
1039 				vc4_hvs5_get_alpha_blend_mode(state) |
1040 				(mix_plane_alpha ?
1041 					SCALER5_CTL2_ALPHA_MIX : 0)
1042 			       );
1043 
1044 		/* Position Word 1: Scaled Image Dimensions. */
1045 		if (!vc4_state->is_unity) {
1046 			vc4_dlist_write(vc4_state,
1047 					VC4_SET_FIELD(vc4_state->crtc_w,
1048 						      SCALER5_POS1_SCL_WIDTH) |
1049 					VC4_SET_FIELD(vc4_state->crtc_h,
1050 						      SCALER5_POS1_SCL_HEIGHT));
1051 		}
1052 
1053 		/* Position Word 2: Source Image Size */
1054 		vc4_state->pos2_offset = vc4_state->dlist_count;
1055 		vc4_dlist_write(vc4_state,
1056 				VC4_SET_FIELD(vc4_state->src_w[0],
1057 					      SCALER5_POS2_WIDTH) |
1058 				VC4_SET_FIELD(vc4_state->src_h[0],
1059 					      SCALER5_POS2_HEIGHT));
1060 
1061 		/* Position Word 3: Context.  Written by the HVS. */
1062 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
1063 	}
1064 
1065 
1066 	/* Pointer Word 0/1/2: RGB / Y / Cb / Cr Pointers
1067 	 *
1068 	 * The pointers may be any byte address.
1069 	 */
1070 	vc4_state->ptr0_offset = vc4_state->dlist_count;
1071 	for (i = 0; i < num_planes; i++)
1072 		vc4_dlist_write(vc4_state, vc4_state->offsets[i]);
1073 
1074 	/* Pointer Context Word 0/1/2: Written by the HVS */
1075 	for (i = 0; i < num_planes; i++)
1076 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
1077 
1078 	/* Pitch word 0 */
1079 	vc4_dlist_write(vc4_state, pitch0);
1080 
1081 	/* Pitch word 1/2 */
1082 	for (i = 1; i < num_planes; i++) {
1083 		if (hvs_format != HVS_PIXEL_FORMAT_H264 &&
1084 		    hvs_format != HVS_PIXEL_FORMAT_YCBCR_10BIT) {
1085 			vc4_dlist_write(vc4_state,
1086 					VC4_SET_FIELD(fb->pitches[i],
1087 						      SCALER_SRC_PITCH));
1088 		} else {
1089 			vc4_dlist_write(vc4_state, pitch0);
1090 		}
1091 	}
1092 
1093 	/* Colorspace conversion words */
1094 	if (vc4_state->is_yuv) {
1095 		enum drm_color_encoding color_encoding = state->color_encoding;
1096 		enum drm_color_range color_range = state->color_range;
1097 		const u32 *ccm;
1098 
1099 		if (color_encoding >= DRM_COLOR_ENCODING_MAX)
1100 			color_encoding = DRM_COLOR_YCBCR_BT601;
1101 		if (color_range >= DRM_COLOR_RANGE_MAX)
1102 			color_range = DRM_COLOR_YCBCR_LIMITED_RANGE;
1103 
1104 		ccm = colorspace_coeffs[color_range][color_encoding];
1105 
1106 		vc4_dlist_write(vc4_state, ccm[0]);
1107 		vc4_dlist_write(vc4_state, ccm[1]);
1108 		vc4_dlist_write(vc4_state, ccm[2]);
1109 	}
1110 
1111 	vc4_state->lbm_offset = 0;
1112 
1113 	if (vc4_state->x_scaling[0] != VC4_SCALING_NONE ||
1114 	    vc4_state->x_scaling[1] != VC4_SCALING_NONE ||
1115 	    vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
1116 	    vc4_state->y_scaling[1] != VC4_SCALING_NONE) {
1117 		/* Reserve a slot for the LBM Base Address. The real value will
1118 		 * be set when calling vc4_plane_allocate_lbm().
1119 		 */
1120 		if (vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
1121 		    vc4_state->y_scaling[1] != VC4_SCALING_NONE) {
1122 			vc4_state->lbm_offset = vc4_state->dlist_count;
1123 			vc4_dlist_counter_increment(vc4_state);
1124 		}
1125 
1126 		if (num_planes > 1) {
1127 			/* Emit Cb/Cr as channel 0 and Y as channel
1128 			 * 1. This matches how we set up scl0/scl1
1129 			 * above.
1130 			 */
1131 			vc4_write_scaling_parameters(state, 1);
1132 		}
1133 		vc4_write_scaling_parameters(state, 0);
1134 
1135 		/* If any PPF setup was done, then all the kernel
1136 		 * pointers get uploaded.
1137 		 */
1138 		if (vc4_state->x_scaling[0] == VC4_SCALING_PPF ||
1139 		    vc4_state->y_scaling[0] == VC4_SCALING_PPF ||
1140 		    vc4_state->x_scaling[1] == VC4_SCALING_PPF ||
1141 		    vc4_state->y_scaling[1] == VC4_SCALING_PPF) {
1142 			u32 kernel = VC4_SET_FIELD(vc4->hvs->mitchell_netravali_filter.start,
1143 						   SCALER_PPF_KERNEL_OFFSET);
1144 
1145 			/* HPPF plane 0 */
1146 			vc4_dlist_write(vc4_state, kernel);
1147 			/* VPPF plane 0 */
1148 			vc4_dlist_write(vc4_state, kernel);
1149 			/* HPPF plane 1 */
1150 			vc4_dlist_write(vc4_state, kernel);
1151 			/* VPPF plane 1 */
1152 			vc4_dlist_write(vc4_state, kernel);
1153 		}
1154 	}
1155 
1156 	vc4_state->dlist[ctl0_offset] |=
1157 		VC4_SET_FIELD(vc4_state->dlist_count, SCALER_CTL0_SIZE);
1158 
1159 	/* crtc_* are already clipped coordinates. */
1160 	covers_screen = vc4_state->crtc_x == 0 && vc4_state->crtc_y == 0 &&
1161 			vc4_state->crtc_w == state->crtc->mode.hdisplay &&
1162 			vc4_state->crtc_h == state->crtc->mode.vdisplay;
1163 	/* Background fill might be necessary when the plane has per-pixel
1164 	 * alpha content or a non-opaque plane alpha and could blend from the
1165 	 * background or does not cover the entire screen.
1166 	 */
1167 	vc4_state->needs_bg_fill = fb->format->has_alpha || !covers_screen ||
1168 				   state->alpha != DRM_BLEND_ALPHA_OPAQUE;
1169 
1170 	/* Flag the dlist as initialized to avoid checking it twice in case
1171 	 * the async update check already called vc4_plane_mode_set() and
1172 	 * decided to fallback to sync update because async update was not
1173 	 * possible.
1174 	 */
1175 	vc4_state->dlist_initialized = 1;
1176 
1177 	vc4_plane_calc_load(state);
1178 
1179 	return 0;
1180 }
1181 
1182 /* If a modeset involves changing the setup of a plane, the atomic
1183  * infrastructure will call this to validate a proposed plane setup.
1184  * However, if a plane isn't getting updated, this (and the
1185  * corresponding vc4_plane_atomic_update) won't get called.  Thus, we
1186  * compute the dlist here and have all active plane dlists get updated
1187  * in the CRTC's flush.
1188  */
1189 static int vc4_plane_atomic_check(struct drm_plane *plane,
1190 				  struct drm_atomic_state *state)
1191 {
1192 	struct drm_plane_state *new_plane_state = drm_atomic_get_new_plane_state(state,
1193 										 plane);
1194 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(new_plane_state);
1195 	int ret;
1196 
1197 	vc4_state->dlist_count = 0;
1198 
1199 	if (!plane_enabled(new_plane_state))
1200 		return 0;
1201 
1202 	ret = vc4_plane_mode_set(plane, new_plane_state);
1203 	if (ret)
1204 		return ret;
1205 
1206 	return vc4_plane_allocate_lbm(new_plane_state);
1207 }
1208 
1209 static void vc4_plane_atomic_update(struct drm_plane *plane,
1210 				    struct drm_atomic_state *state)
1211 {
1212 	/* No contents here.  Since we don't know where in the CRTC's
1213 	 * dlist we should be stored, our dlist is uploaded to the
1214 	 * hardware with vc4_plane_write_dlist() at CRTC atomic_flush
1215 	 * time.
1216 	 */
1217 }
1218 
1219 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist)
1220 {
1221 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
1222 	int i;
1223 	int idx;
1224 
1225 	if (!drm_dev_enter(plane->dev, &idx))
1226 		goto out;
1227 
1228 	vc4_state->hw_dlist = dlist;
1229 
1230 	/* Can't memcpy_toio() because it needs to be 32-bit writes. */
1231 	for (i = 0; i < vc4_state->dlist_count; i++)
1232 		writel(vc4_state->dlist[i], &dlist[i]);
1233 
1234 	drm_dev_exit(idx);
1235 
1236 out:
1237 	return vc4_state->dlist_count;
1238 }
1239 
1240 u32 vc4_plane_dlist_size(const struct drm_plane_state *state)
1241 {
1242 	const struct vc4_plane_state *vc4_state =
1243 		container_of(state, typeof(*vc4_state), base);
1244 
1245 	return vc4_state->dlist_count;
1246 }
1247 
1248 /* Updates the plane to immediately (well, once the FIFO needs
1249  * refilling) scan out from at a new framebuffer.
1250  */
1251 void vc4_plane_async_set_fb(struct drm_plane *plane, struct drm_framebuffer *fb)
1252 {
1253 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
1254 	struct drm_gem_dma_object *bo = drm_fb_dma_get_gem_obj(fb, 0);
1255 	uint32_t addr;
1256 	int idx;
1257 
1258 	if (!drm_dev_enter(plane->dev, &idx))
1259 		return;
1260 
1261 	/* We're skipping the address adjustment for negative origin,
1262 	 * because this is only called on the primary plane.
1263 	 */
1264 	WARN_ON_ONCE(plane->state->crtc_x < 0 || plane->state->crtc_y < 0);
1265 	addr = bo->dma_addr + fb->offsets[0];
1266 
1267 	/* Write the new address into the hardware immediately.  The
1268 	 * scanout will start from this address as soon as the FIFO
1269 	 * needs to refill with pixels.
1270 	 */
1271 	writel(addr, &vc4_state->hw_dlist[vc4_state->ptr0_offset]);
1272 
1273 	/* Also update the CPU-side dlist copy, so that any later
1274 	 * atomic updates that don't do a new modeset on our plane
1275 	 * also use our updated address.
1276 	 */
1277 	vc4_state->dlist[vc4_state->ptr0_offset] = addr;
1278 
1279 	drm_dev_exit(idx);
1280 }
1281 
1282 static void vc4_plane_atomic_async_update(struct drm_plane *plane,
1283 					  struct drm_atomic_state *state)
1284 {
1285 	struct drm_plane_state *new_plane_state = drm_atomic_get_new_plane_state(state,
1286 										 plane);
1287 	struct vc4_plane_state *vc4_state, *new_vc4_state;
1288 	int idx;
1289 
1290 	if (!drm_dev_enter(plane->dev, &idx))
1291 		return;
1292 
1293 	swap(plane->state->fb, new_plane_state->fb);
1294 	plane->state->crtc_x = new_plane_state->crtc_x;
1295 	plane->state->crtc_y = new_plane_state->crtc_y;
1296 	plane->state->crtc_w = new_plane_state->crtc_w;
1297 	plane->state->crtc_h = new_plane_state->crtc_h;
1298 	plane->state->src_x = new_plane_state->src_x;
1299 	plane->state->src_y = new_plane_state->src_y;
1300 	plane->state->src_w = new_plane_state->src_w;
1301 	plane->state->src_h = new_plane_state->src_h;
1302 	plane->state->alpha = new_plane_state->alpha;
1303 	plane->state->pixel_blend_mode = new_plane_state->pixel_blend_mode;
1304 	plane->state->rotation = new_plane_state->rotation;
1305 	plane->state->zpos = new_plane_state->zpos;
1306 	plane->state->normalized_zpos = new_plane_state->normalized_zpos;
1307 	plane->state->color_encoding = new_plane_state->color_encoding;
1308 	plane->state->color_range = new_plane_state->color_range;
1309 	plane->state->src = new_plane_state->src;
1310 	plane->state->dst = new_plane_state->dst;
1311 	plane->state->visible = new_plane_state->visible;
1312 
1313 	new_vc4_state = to_vc4_plane_state(new_plane_state);
1314 	vc4_state = to_vc4_plane_state(plane->state);
1315 
1316 	vc4_state->crtc_x = new_vc4_state->crtc_x;
1317 	vc4_state->crtc_y = new_vc4_state->crtc_y;
1318 	vc4_state->crtc_h = new_vc4_state->crtc_h;
1319 	vc4_state->crtc_w = new_vc4_state->crtc_w;
1320 	vc4_state->src_x = new_vc4_state->src_x;
1321 	vc4_state->src_y = new_vc4_state->src_y;
1322 	memcpy(vc4_state->src_w, new_vc4_state->src_w,
1323 	       sizeof(vc4_state->src_w));
1324 	memcpy(vc4_state->src_h, new_vc4_state->src_h,
1325 	       sizeof(vc4_state->src_h));
1326 	memcpy(vc4_state->x_scaling, new_vc4_state->x_scaling,
1327 	       sizeof(vc4_state->x_scaling));
1328 	memcpy(vc4_state->y_scaling, new_vc4_state->y_scaling,
1329 	       sizeof(vc4_state->y_scaling));
1330 	vc4_state->is_unity = new_vc4_state->is_unity;
1331 	vc4_state->is_yuv = new_vc4_state->is_yuv;
1332 	memcpy(vc4_state->offsets, new_vc4_state->offsets,
1333 	       sizeof(vc4_state->offsets));
1334 	vc4_state->needs_bg_fill = new_vc4_state->needs_bg_fill;
1335 
1336 	/* Update the current vc4_state pos0, pos2 and ptr0 dlist entries. */
1337 	vc4_state->dlist[vc4_state->pos0_offset] =
1338 		new_vc4_state->dlist[vc4_state->pos0_offset];
1339 	vc4_state->dlist[vc4_state->pos2_offset] =
1340 		new_vc4_state->dlist[vc4_state->pos2_offset];
1341 	vc4_state->dlist[vc4_state->ptr0_offset] =
1342 		new_vc4_state->dlist[vc4_state->ptr0_offset];
1343 
1344 	/* Note that we can't just call vc4_plane_write_dlist()
1345 	 * because that would smash the context data that the HVS is
1346 	 * currently using.
1347 	 */
1348 	writel(vc4_state->dlist[vc4_state->pos0_offset],
1349 	       &vc4_state->hw_dlist[vc4_state->pos0_offset]);
1350 	writel(vc4_state->dlist[vc4_state->pos2_offset],
1351 	       &vc4_state->hw_dlist[vc4_state->pos2_offset]);
1352 	writel(vc4_state->dlist[vc4_state->ptr0_offset],
1353 	       &vc4_state->hw_dlist[vc4_state->ptr0_offset]);
1354 
1355 	drm_dev_exit(idx);
1356 }
1357 
1358 static int vc4_plane_atomic_async_check(struct drm_plane *plane,
1359 					struct drm_atomic_state *state)
1360 {
1361 	struct drm_plane_state *new_plane_state = drm_atomic_get_new_plane_state(state,
1362 										 plane);
1363 	struct vc4_plane_state *old_vc4_state, *new_vc4_state;
1364 	int ret;
1365 	u32 i;
1366 
1367 	ret = vc4_plane_mode_set(plane, new_plane_state);
1368 	if (ret)
1369 		return ret;
1370 
1371 	old_vc4_state = to_vc4_plane_state(plane->state);
1372 	new_vc4_state = to_vc4_plane_state(new_plane_state);
1373 
1374 	if (!new_vc4_state->hw_dlist)
1375 		return -EINVAL;
1376 
1377 	if (old_vc4_state->dlist_count != new_vc4_state->dlist_count ||
1378 	    old_vc4_state->pos0_offset != new_vc4_state->pos0_offset ||
1379 	    old_vc4_state->pos2_offset != new_vc4_state->pos2_offset ||
1380 	    old_vc4_state->ptr0_offset != new_vc4_state->ptr0_offset ||
1381 	    vc4_lbm_size(plane->state) != vc4_lbm_size(new_plane_state))
1382 		return -EINVAL;
1383 
1384 	/* Only pos0, pos2 and ptr0 DWORDS can be updated in an async update
1385 	 * if anything else has changed, fallback to a sync update.
1386 	 */
1387 	for (i = 0; i < new_vc4_state->dlist_count; i++) {
1388 		if (i == new_vc4_state->pos0_offset ||
1389 		    i == new_vc4_state->pos2_offset ||
1390 		    i == new_vc4_state->ptr0_offset ||
1391 		    (new_vc4_state->lbm_offset &&
1392 		     i == new_vc4_state->lbm_offset))
1393 			continue;
1394 
1395 		if (new_vc4_state->dlist[i] != old_vc4_state->dlist[i])
1396 			return -EINVAL;
1397 	}
1398 
1399 	return 0;
1400 }
1401 
1402 static int vc4_prepare_fb(struct drm_plane *plane,
1403 			  struct drm_plane_state *state)
1404 {
1405 	struct vc4_bo *bo;
1406 
1407 	if (!state->fb)
1408 		return 0;
1409 
1410 	bo = to_vc4_bo(&drm_fb_dma_get_gem_obj(state->fb, 0)->base);
1411 
1412 	drm_gem_plane_helper_prepare_fb(plane, state);
1413 
1414 	if (plane->state->fb == state->fb)
1415 		return 0;
1416 
1417 	return vc4_bo_inc_usecnt(bo);
1418 }
1419 
1420 static void vc4_cleanup_fb(struct drm_plane *plane,
1421 			   struct drm_plane_state *state)
1422 {
1423 	struct vc4_bo *bo;
1424 
1425 	if (plane->state->fb == state->fb || !state->fb)
1426 		return;
1427 
1428 	bo = to_vc4_bo(&drm_fb_dma_get_gem_obj(state->fb, 0)->base);
1429 	vc4_bo_dec_usecnt(bo);
1430 }
1431 
1432 static const struct drm_plane_helper_funcs vc4_plane_helper_funcs = {
1433 	.atomic_check = vc4_plane_atomic_check,
1434 	.atomic_update = vc4_plane_atomic_update,
1435 	.prepare_fb = vc4_prepare_fb,
1436 	.cleanup_fb = vc4_cleanup_fb,
1437 	.atomic_async_check = vc4_plane_atomic_async_check,
1438 	.atomic_async_update = vc4_plane_atomic_async_update,
1439 };
1440 
1441 static const struct drm_plane_helper_funcs vc5_plane_helper_funcs = {
1442 	.atomic_check = vc4_plane_atomic_check,
1443 	.atomic_update = vc4_plane_atomic_update,
1444 	.atomic_async_check = vc4_plane_atomic_async_check,
1445 	.atomic_async_update = vc4_plane_atomic_async_update,
1446 };
1447 
1448 static bool vc4_format_mod_supported(struct drm_plane *plane,
1449 				     uint32_t format,
1450 				     uint64_t modifier)
1451 {
1452 	/* Support T_TILING for RGB formats only. */
1453 	switch (format) {
1454 	case DRM_FORMAT_XRGB8888:
1455 	case DRM_FORMAT_ARGB8888:
1456 	case DRM_FORMAT_ABGR8888:
1457 	case DRM_FORMAT_XBGR8888:
1458 	case DRM_FORMAT_RGB565:
1459 	case DRM_FORMAT_BGR565:
1460 	case DRM_FORMAT_ARGB1555:
1461 	case DRM_FORMAT_XRGB1555:
1462 		switch (fourcc_mod_broadcom_mod(modifier)) {
1463 		case DRM_FORMAT_MOD_LINEAR:
1464 		case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
1465 			return true;
1466 		default:
1467 			return false;
1468 		}
1469 	case DRM_FORMAT_NV12:
1470 	case DRM_FORMAT_NV21:
1471 		switch (fourcc_mod_broadcom_mod(modifier)) {
1472 		case DRM_FORMAT_MOD_LINEAR:
1473 		case DRM_FORMAT_MOD_BROADCOM_SAND64:
1474 		case DRM_FORMAT_MOD_BROADCOM_SAND128:
1475 		case DRM_FORMAT_MOD_BROADCOM_SAND256:
1476 			return true;
1477 		default:
1478 			return false;
1479 		}
1480 	case DRM_FORMAT_P030:
1481 		switch (fourcc_mod_broadcom_mod(modifier)) {
1482 		case DRM_FORMAT_MOD_BROADCOM_SAND128:
1483 			return true;
1484 		default:
1485 			return false;
1486 		}
1487 	case DRM_FORMAT_RGBX1010102:
1488 	case DRM_FORMAT_BGRX1010102:
1489 	case DRM_FORMAT_RGBA1010102:
1490 	case DRM_FORMAT_BGRA1010102:
1491 	case DRM_FORMAT_YUV422:
1492 	case DRM_FORMAT_YVU422:
1493 	case DRM_FORMAT_YUV420:
1494 	case DRM_FORMAT_YVU420:
1495 	case DRM_FORMAT_NV16:
1496 	case DRM_FORMAT_NV61:
1497 	default:
1498 		return (modifier == DRM_FORMAT_MOD_LINEAR);
1499 	}
1500 }
1501 
1502 static const struct drm_plane_funcs vc4_plane_funcs = {
1503 	.update_plane = drm_atomic_helper_update_plane,
1504 	.disable_plane = drm_atomic_helper_disable_plane,
1505 	.reset = vc4_plane_reset,
1506 	.atomic_duplicate_state = vc4_plane_duplicate_state,
1507 	.atomic_destroy_state = vc4_plane_destroy_state,
1508 	.format_mod_supported = vc4_format_mod_supported,
1509 };
1510 
1511 struct drm_plane *vc4_plane_init(struct drm_device *dev,
1512 				 enum drm_plane_type type,
1513 				 uint32_t possible_crtcs)
1514 {
1515 	struct vc4_dev *vc4 = to_vc4_dev(dev);
1516 	struct drm_plane *plane;
1517 	struct vc4_plane *vc4_plane;
1518 	u32 formats[ARRAY_SIZE(hvs_formats)];
1519 	int num_formats = 0;
1520 	unsigned i;
1521 	static const uint64_t modifiers[] = {
1522 		DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED,
1523 		DRM_FORMAT_MOD_BROADCOM_SAND128,
1524 		DRM_FORMAT_MOD_BROADCOM_SAND64,
1525 		DRM_FORMAT_MOD_BROADCOM_SAND256,
1526 		DRM_FORMAT_MOD_LINEAR,
1527 		DRM_FORMAT_MOD_INVALID
1528 	};
1529 
1530 	for (i = 0; i < ARRAY_SIZE(hvs_formats); i++) {
1531 		if (!hvs_formats[i].hvs5_only || vc4->is_vc5) {
1532 			formats[num_formats] = hvs_formats[i].drm;
1533 			num_formats++;
1534 		}
1535 	}
1536 
1537 	vc4_plane = drmm_universal_plane_alloc(dev, struct vc4_plane, base,
1538 					       possible_crtcs,
1539 					       &vc4_plane_funcs,
1540 					       formats, num_formats,
1541 					       modifiers, type, NULL);
1542 	if (IS_ERR(vc4_plane))
1543 		return ERR_CAST(vc4_plane);
1544 	plane = &vc4_plane->base;
1545 
1546 	if (vc4->is_vc5)
1547 		drm_plane_helper_add(plane, &vc5_plane_helper_funcs);
1548 	else
1549 		drm_plane_helper_add(plane, &vc4_plane_helper_funcs);
1550 
1551 	drm_plane_create_alpha_property(plane);
1552 	drm_plane_create_blend_mode_property(plane,
1553 					     BIT(DRM_MODE_BLEND_PIXEL_NONE) |
1554 					     BIT(DRM_MODE_BLEND_PREMULTI) |
1555 					     BIT(DRM_MODE_BLEND_COVERAGE));
1556 	drm_plane_create_rotation_property(plane, DRM_MODE_ROTATE_0,
1557 					   DRM_MODE_ROTATE_0 |
1558 					   DRM_MODE_ROTATE_180 |
1559 					   DRM_MODE_REFLECT_X |
1560 					   DRM_MODE_REFLECT_Y);
1561 
1562 	drm_plane_create_color_properties(plane,
1563 					  BIT(DRM_COLOR_YCBCR_BT601) |
1564 					  BIT(DRM_COLOR_YCBCR_BT709) |
1565 					  BIT(DRM_COLOR_YCBCR_BT2020),
1566 					  BIT(DRM_COLOR_YCBCR_LIMITED_RANGE) |
1567 					  BIT(DRM_COLOR_YCBCR_FULL_RANGE),
1568 					  DRM_COLOR_YCBCR_BT709,
1569 					  DRM_COLOR_YCBCR_LIMITED_RANGE);
1570 
1571 	return plane;
1572 }
1573 
1574 int vc4_plane_create_additional_planes(struct drm_device *drm)
1575 {
1576 	struct drm_plane *cursor_plane;
1577 	struct drm_crtc *crtc;
1578 	unsigned int i;
1579 
1580 	/* Set up some arbitrary number of planes.  We're not limited
1581 	 * by a set number of physical registers, just the space in
1582 	 * the HVS (16k) and how small an plane can be (28 bytes).
1583 	 * However, each plane we set up takes up some memory, and
1584 	 * increases the cost of looping over planes, which atomic
1585 	 * modesetting does quite a bit.  As a result, we pick a
1586 	 * modest number of planes to expose, that should hopefully
1587 	 * still cover any sane usecase.
1588 	 */
1589 	for (i = 0; i < 16; i++) {
1590 		struct drm_plane *plane =
1591 			vc4_plane_init(drm, DRM_PLANE_TYPE_OVERLAY,
1592 				       GENMASK(drm->mode_config.num_crtc - 1, 0));
1593 
1594 		if (IS_ERR(plane))
1595 			continue;
1596 	}
1597 
1598 	drm_for_each_crtc(crtc, drm) {
1599 		/* Set up the legacy cursor after overlay initialization,
1600 		 * since we overlay planes on the CRTC in the order they were
1601 		 * initialized.
1602 		 */
1603 		cursor_plane = vc4_plane_init(drm, DRM_PLANE_TYPE_CURSOR,
1604 					      drm_crtc_mask(crtc));
1605 		if (!IS_ERR(cursor_plane)) {
1606 			crtc->cursor = cursor_plane;
1607 		}
1608 	}
1609 
1610 	return 0;
1611 }
1612