xref: /linux/drivers/gpu/drm/vc4/vc4_plane.c (revision ca853314e78b0a65c20b6a889a23c31f918d4aa2)
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_fb_cma_helper.h>
22 #include <drm/drm_fourcc.h>
23 #include <drm/drm_gem_framebuffer_helper.h>
24 #include <drm/drm_plane_helper.h>
25 
26 #include "uapi/drm/vc4_drm.h"
27 
28 #include "vc4_drv.h"
29 #include "vc4_regs.h"
30 
31 static const struct hvs_format {
32 	u32 drm; /* DRM_FORMAT_* */
33 	u32 hvs; /* HVS_FORMAT_* */
34 	u32 pixel_order;
35 	u32 pixel_order_hvs5;
36 } hvs_formats[] = {
37 	{
38 		.drm = DRM_FORMAT_XRGB8888,
39 		.hvs = HVS_PIXEL_FORMAT_RGBA8888,
40 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
41 		.pixel_order_hvs5 = HVS_PIXEL_ORDER_ARGB,
42 	},
43 	{
44 		.drm = DRM_FORMAT_ARGB8888,
45 		.hvs = HVS_PIXEL_FORMAT_RGBA8888,
46 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
47 		.pixel_order_hvs5 = HVS_PIXEL_ORDER_ARGB,
48 	},
49 	{
50 		.drm = DRM_FORMAT_ABGR8888,
51 		.hvs = HVS_PIXEL_FORMAT_RGBA8888,
52 		.pixel_order = HVS_PIXEL_ORDER_ARGB,
53 		.pixel_order_hvs5 = HVS_PIXEL_ORDER_ABGR,
54 	},
55 	{
56 		.drm = DRM_FORMAT_XBGR8888,
57 		.hvs = HVS_PIXEL_FORMAT_RGBA8888,
58 		.pixel_order = HVS_PIXEL_ORDER_ARGB,
59 		.pixel_order_hvs5 = HVS_PIXEL_ORDER_ABGR,
60 	},
61 	{
62 		.drm = DRM_FORMAT_RGB565,
63 		.hvs = HVS_PIXEL_FORMAT_RGB565,
64 		.pixel_order = HVS_PIXEL_ORDER_XRGB,
65 	},
66 	{
67 		.drm = DRM_FORMAT_BGR565,
68 		.hvs = HVS_PIXEL_FORMAT_RGB565,
69 		.pixel_order = HVS_PIXEL_ORDER_XBGR,
70 	},
71 	{
72 		.drm = DRM_FORMAT_ARGB1555,
73 		.hvs = HVS_PIXEL_FORMAT_RGBA5551,
74 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
75 	},
76 	{
77 		.drm = DRM_FORMAT_XRGB1555,
78 		.hvs = HVS_PIXEL_FORMAT_RGBA5551,
79 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
80 	},
81 	{
82 		.drm = DRM_FORMAT_RGB888,
83 		.hvs = HVS_PIXEL_FORMAT_RGB888,
84 		.pixel_order = HVS_PIXEL_ORDER_XRGB,
85 	},
86 	{
87 		.drm = DRM_FORMAT_BGR888,
88 		.hvs = HVS_PIXEL_FORMAT_RGB888,
89 		.pixel_order = HVS_PIXEL_ORDER_XBGR,
90 	},
91 	{
92 		.drm = DRM_FORMAT_YUV422,
93 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_3PLANE,
94 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
95 	},
96 	{
97 		.drm = DRM_FORMAT_YVU422,
98 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_3PLANE,
99 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
100 	},
101 	{
102 		.drm = DRM_FORMAT_YUV420,
103 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_3PLANE,
104 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
105 	},
106 	{
107 		.drm = DRM_FORMAT_YVU420,
108 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_3PLANE,
109 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
110 	},
111 	{
112 		.drm = DRM_FORMAT_NV12,
113 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_2PLANE,
114 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
115 	},
116 	{
117 		.drm = DRM_FORMAT_NV21,
118 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_2PLANE,
119 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
120 	},
121 	{
122 		.drm = DRM_FORMAT_NV16,
123 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_2PLANE,
124 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
125 	},
126 	{
127 		.drm = DRM_FORMAT_NV61,
128 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_2PLANE,
129 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
130 	},
131 };
132 
133 static const struct hvs_format *vc4_get_hvs_format(u32 drm_format)
134 {
135 	unsigned i;
136 
137 	for (i = 0; i < ARRAY_SIZE(hvs_formats); i++) {
138 		if (hvs_formats[i].drm == drm_format)
139 			return &hvs_formats[i];
140 	}
141 
142 	return NULL;
143 }
144 
145 static enum vc4_scaling_mode vc4_get_scaling_mode(u32 src, u32 dst)
146 {
147 	if (dst == src)
148 		return VC4_SCALING_NONE;
149 	if (3 * dst >= 2 * src)
150 		return VC4_SCALING_PPF;
151 	else
152 		return VC4_SCALING_TPZ;
153 }
154 
155 static bool plane_enabled(struct drm_plane_state *state)
156 {
157 	return state->fb && !WARN_ON(!state->crtc);
158 }
159 
160 static struct drm_plane_state *vc4_plane_duplicate_state(struct drm_plane *plane)
161 {
162 	struct vc4_plane_state *vc4_state;
163 
164 	if (WARN_ON(!plane->state))
165 		return NULL;
166 
167 	vc4_state = kmemdup(plane->state, sizeof(*vc4_state), GFP_KERNEL);
168 	if (!vc4_state)
169 		return NULL;
170 
171 	memset(&vc4_state->lbm, 0, sizeof(vc4_state->lbm));
172 	vc4_state->dlist_initialized = 0;
173 
174 	__drm_atomic_helper_plane_duplicate_state(plane, &vc4_state->base);
175 
176 	if (vc4_state->dlist) {
177 		vc4_state->dlist = kmemdup(vc4_state->dlist,
178 					   vc4_state->dlist_count * 4,
179 					   GFP_KERNEL);
180 		if (!vc4_state->dlist) {
181 			kfree(vc4_state);
182 			return NULL;
183 		}
184 		vc4_state->dlist_size = vc4_state->dlist_count;
185 	}
186 
187 	return &vc4_state->base;
188 }
189 
190 static void vc4_plane_destroy_state(struct drm_plane *plane,
191 				    struct drm_plane_state *state)
192 {
193 	struct vc4_dev *vc4 = to_vc4_dev(plane->dev);
194 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
195 
196 	if (drm_mm_node_allocated(&vc4_state->lbm)) {
197 		unsigned long irqflags;
198 
199 		spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
200 		drm_mm_remove_node(&vc4_state->lbm);
201 		spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
202 	}
203 
204 	kfree(vc4_state->dlist);
205 	__drm_atomic_helper_plane_destroy_state(&vc4_state->base);
206 	kfree(state);
207 }
208 
209 /* Called during init to allocate the plane's atomic state. */
210 static void vc4_plane_reset(struct drm_plane *plane)
211 {
212 	struct vc4_plane_state *vc4_state;
213 
214 	WARN_ON(plane->state);
215 
216 	vc4_state = kzalloc(sizeof(*vc4_state), GFP_KERNEL);
217 	if (!vc4_state)
218 		return;
219 
220 	__drm_atomic_helper_plane_reset(plane, &vc4_state->base);
221 }
222 
223 static void vc4_dlist_write(struct vc4_plane_state *vc4_state, u32 val)
224 {
225 	if (vc4_state->dlist_count == vc4_state->dlist_size) {
226 		u32 new_size = max(4u, vc4_state->dlist_count * 2);
227 		u32 *new_dlist = kmalloc_array(new_size, 4, GFP_KERNEL);
228 
229 		if (!new_dlist)
230 			return;
231 		memcpy(new_dlist, vc4_state->dlist, vc4_state->dlist_count * 4);
232 
233 		kfree(vc4_state->dlist);
234 		vc4_state->dlist = new_dlist;
235 		vc4_state->dlist_size = new_size;
236 	}
237 
238 	vc4_state->dlist[vc4_state->dlist_count++] = val;
239 }
240 
241 /* Returns the scl0/scl1 field based on whether the dimensions need to
242  * be up/down/non-scaled.
243  *
244  * This is a replication of a table from the spec.
245  */
246 static u32 vc4_get_scl_field(struct drm_plane_state *state, int plane)
247 {
248 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
249 
250 	switch (vc4_state->x_scaling[plane] << 2 | vc4_state->y_scaling[plane]) {
251 	case VC4_SCALING_PPF << 2 | VC4_SCALING_PPF:
252 		return SCALER_CTL0_SCL_H_PPF_V_PPF;
253 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_PPF:
254 		return SCALER_CTL0_SCL_H_TPZ_V_PPF;
255 	case VC4_SCALING_PPF << 2 | VC4_SCALING_TPZ:
256 		return SCALER_CTL0_SCL_H_PPF_V_TPZ;
257 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_TPZ:
258 		return SCALER_CTL0_SCL_H_TPZ_V_TPZ;
259 	case VC4_SCALING_PPF << 2 | VC4_SCALING_NONE:
260 		return SCALER_CTL0_SCL_H_PPF_V_NONE;
261 	case VC4_SCALING_NONE << 2 | VC4_SCALING_PPF:
262 		return SCALER_CTL0_SCL_H_NONE_V_PPF;
263 	case VC4_SCALING_NONE << 2 | VC4_SCALING_TPZ:
264 		return SCALER_CTL0_SCL_H_NONE_V_TPZ;
265 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_NONE:
266 		return SCALER_CTL0_SCL_H_TPZ_V_NONE;
267 	default:
268 	case VC4_SCALING_NONE << 2 | VC4_SCALING_NONE:
269 		/* The unity case is independently handled by
270 		 * SCALER_CTL0_UNITY.
271 		 */
272 		return 0;
273 	}
274 }
275 
276 static int vc4_plane_margins_adj(struct drm_plane_state *pstate)
277 {
278 	struct vc4_plane_state *vc4_pstate = to_vc4_plane_state(pstate);
279 	unsigned int left, right, top, bottom, adjhdisplay, adjvdisplay;
280 	struct drm_crtc_state *crtc_state;
281 
282 	crtc_state = drm_atomic_get_new_crtc_state(pstate->state,
283 						   pstate->crtc);
284 
285 	vc4_crtc_get_margins(crtc_state, &left, &right, &top, &bottom);
286 	if (!left && !right && !top && !bottom)
287 		return 0;
288 
289 	if (left + right >= crtc_state->mode.hdisplay ||
290 	    top + bottom >= crtc_state->mode.vdisplay)
291 		return -EINVAL;
292 
293 	adjhdisplay = crtc_state->mode.hdisplay - (left + right);
294 	vc4_pstate->crtc_x = DIV_ROUND_CLOSEST(vc4_pstate->crtc_x *
295 					       adjhdisplay,
296 					       crtc_state->mode.hdisplay);
297 	vc4_pstate->crtc_x += left;
298 	if (vc4_pstate->crtc_x > crtc_state->mode.hdisplay - left)
299 		vc4_pstate->crtc_x = crtc_state->mode.hdisplay - left;
300 
301 	adjvdisplay = crtc_state->mode.vdisplay - (top + bottom);
302 	vc4_pstate->crtc_y = DIV_ROUND_CLOSEST(vc4_pstate->crtc_y *
303 					       adjvdisplay,
304 					       crtc_state->mode.vdisplay);
305 	vc4_pstate->crtc_y += top;
306 	if (vc4_pstate->crtc_y > crtc_state->mode.vdisplay - top)
307 		vc4_pstate->crtc_y = crtc_state->mode.vdisplay - top;
308 
309 	vc4_pstate->crtc_w = DIV_ROUND_CLOSEST(vc4_pstate->crtc_w *
310 					       adjhdisplay,
311 					       crtc_state->mode.hdisplay);
312 	vc4_pstate->crtc_h = DIV_ROUND_CLOSEST(vc4_pstate->crtc_h *
313 					       adjvdisplay,
314 					       crtc_state->mode.vdisplay);
315 
316 	if (!vc4_pstate->crtc_w || !vc4_pstate->crtc_h)
317 		return -EINVAL;
318 
319 	return 0;
320 }
321 
322 static int vc4_plane_setup_clipping_and_scaling(struct drm_plane_state *state)
323 {
324 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
325 	struct drm_framebuffer *fb = state->fb;
326 	struct drm_gem_cma_object *bo = drm_fb_cma_get_gem_obj(fb, 0);
327 	u32 subpixel_src_mask = (1 << 16) - 1;
328 	int num_planes = fb->format->num_planes;
329 	struct drm_crtc_state *crtc_state;
330 	u32 h_subsample = fb->format->hsub;
331 	u32 v_subsample = fb->format->vsub;
332 	int i, ret;
333 
334 	crtc_state = drm_atomic_get_existing_crtc_state(state->state,
335 							state->crtc);
336 	if (!crtc_state) {
337 		DRM_DEBUG_KMS("Invalid crtc state\n");
338 		return -EINVAL;
339 	}
340 
341 	ret = drm_atomic_helper_check_plane_state(state, crtc_state, 1,
342 						  INT_MAX, true, true);
343 	if (ret)
344 		return ret;
345 
346 	for (i = 0; i < num_planes; i++)
347 		vc4_state->offsets[i] = bo->paddr + fb->offsets[i];
348 
349 	/* We don't support subpixel source positioning for scaling. */
350 	if ((state->src.x1 & subpixel_src_mask) ||
351 	    (state->src.x2 & subpixel_src_mask) ||
352 	    (state->src.y1 & subpixel_src_mask) ||
353 	    (state->src.y2 & subpixel_src_mask)) {
354 		return -EINVAL;
355 	}
356 
357 	vc4_state->src_x = state->src.x1 >> 16;
358 	vc4_state->src_y = state->src.y1 >> 16;
359 	vc4_state->src_w[0] = (state->src.x2 - state->src.x1) >> 16;
360 	vc4_state->src_h[0] = (state->src.y2 - state->src.y1) >> 16;
361 
362 	vc4_state->crtc_x = state->dst.x1;
363 	vc4_state->crtc_y = state->dst.y1;
364 	vc4_state->crtc_w = state->dst.x2 - state->dst.x1;
365 	vc4_state->crtc_h = state->dst.y2 - state->dst.y1;
366 
367 	ret = vc4_plane_margins_adj(state);
368 	if (ret)
369 		return ret;
370 
371 	vc4_state->x_scaling[0] = vc4_get_scaling_mode(vc4_state->src_w[0],
372 						       vc4_state->crtc_w);
373 	vc4_state->y_scaling[0] = vc4_get_scaling_mode(vc4_state->src_h[0],
374 						       vc4_state->crtc_h);
375 
376 	vc4_state->is_unity = (vc4_state->x_scaling[0] == VC4_SCALING_NONE &&
377 			       vc4_state->y_scaling[0] == VC4_SCALING_NONE);
378 
379 	if (num_planes > 1) {
380 		vc4_state->is_yuv = true;
381 
382 		vc4_state->src_w[1] = vc4_state->src_w[0] / h_subsample;
383 		vc4_state->src_h[1] = vc4_state->src_h[0] / v_subsample;
384 
385 		vc4_state->x_scaling[1] =
386 			vc4_get_scaling_mode(vc4_state->src_w[1],
387 					     vc4_state->crtc_w);
388 		vc4_state->y_scaling[1] =
389 			vc4_get_scaling_mode(vc4_state->src_h[1],
390 					     vc4_state->crtc_h);
391 
392 		/* YUV conversion requires that horizontal scaling be enabled
393 		 * on the UV plane even if vc4_get_scaling_mode() returned
394 		 * VC4_SCALING_NONE (which can happen when the down-scaling
395 		 * ratio is 0.5). Let's force it to VC4_SCALING_PPF in this
396 		 * case.
397 		 */
398 		if (vc4_state->x_scaling[1] == VC4_SCALING_NONE)
399 			vc4_state->x_scaling[1] = VC4_SCALING_PPF;
400 	} else {
401 		vc4_state->is_yuv = false;
402 		vc4_state->x_scaling[1] = VC4_SCALING_NONE;
403 		vc4_state->y_scaling[1] = VC4_SCALING_NONE;
404 	}
405 
406 	return 0;
407 }
408 
409 static void vc4_write_tpz(struct vc4_plane_state *vc4_state, u32 src, u32 dst)
410 {
411 	u32 scale, recip;
412 
413 	scale = (1 << 16) * src / dst;
414 
415 	/* The specs note that while the reciprocal would be defined
416 	 * as (1<<32)/scale, ~0 is close enough.
417 	 */
418 	recip = ~0 / scale;
419 
420 	vc4_dlist_write(vc4_state,
421 			VC4_SET_FIELD(scale, SCALER_TPZ0_SCALE) |
422 			VC4_SET_FIELD(0, SCALER_TPZ0_IPHASE));
423 	vc4_dlist_write(vc4_state,
424 			VC4_SET_FIELD(recip, SCALER_TPZ1_RECIP));
425 }
426 
427 static void vc4_write_ppf(struct vc4_plane_state *vc4_state, u32 src, u32 dst)
428 {
429 	u32 scale = (1 << 16) * src / dst;
430 
431 	vc4_dlist_write(vc4_state,
432 			SCALER_PPF_AGC |
433 			VC4_SET_FIELD(scale, SCALER_PPF_SCALE) |
434 			VC4_SET_FIELD(0, SCALER_PPF_IPHASE));
435 }
436 
437 static u32 vc4_lbm_size(struct drm_plane_state *state)
438 {
439 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
440 	u32 pix_per_line;
441 	u32 lbm;
442 
443 	/* LBM is not needed when there's no vertical scaling. */
444 	if (vc4_state->y_scaling[0] == VC4_SCALING_NONE &&
445 	    vc4_state->y_scaling[1] == VC4_SCALING_NONE)
446 		return 0;
447 
448 	/*
449 	 * This can be further optimized in the RGB/YUV444 case if the PPF
450 	 * decimation factor is between 0.5 and 1.0 by using crtc_w.
451 	 *
452 	 * It's not an issue though, since in that case since src_w[0] is going
453 	 * to be greater than or equal to crtc_w.
454 	 */
455 	if (vc4_state->x_scaling[0] == VC4_SCALING_TPZ)
456 		pix_per_line = vc4_state->crtc_w;
457 	else
458 		pix_per_line = vc4_state->src_w[0];
459 
460 	if (!vc4_state->is_yuv) {
461 		if (vc4_state->y_scaling[0] == VC4_SCALING_TPZ)
462 			lbm = pix_per_line * 8;
463 		else {
464 			/* In special cases, this multiplier might be 12. */
465 			lbm = pix_per_line * 16;
466 		}
467 	} else {
468 		/* There are cases for this going down to a multiplier
469 		 * of 2, but according to the firmware source, the
470 		 * table in the docs is somewhat wrong.
471 		 */
472 		lbm = pix_per_line * 16;
473 	}
474 
475 	lbm = roundup(lbm, 32);
476 
477 	return lbm;
478 }
479 
480 static void vc4_write_scaling_parameters(struct drm_plane_state *state,
481 					 int channel)
482 {
483 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
484 
485 	/* Ch0 H-PPF Word 0: Scaling Parameters */
486 	if (vc4_state->x_scaling[channel] == VC4_SCALING_PPF) {
487 		vc4_write_ppf(vc4_state,
488 			      vc4_state->src_w[channel], vc4_state->crtc_w);
489 	}
490 
491 	/* Ch0 V-PPF Words 0-1: Scaling Parameters, Context */
492 	if (vc4_state->y_scaling[channel] == VC4_SCALING_PPF) {
493 		vc4_write_ppf(vc4_state,
494 			      vc4_state->src_h[channel], vc4_state->crtc_h);
495 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
496 	}
497 
498 	/* Ch0 H-TPZ Words 0-1: Scaling Parameters, Recip */
499 	if (vc4_state->x_scaling[channel] == VC4_SCALING_TPZ) {
500 		vc4_write_tpz(vc4_state,
501 			      vc4_state->src_w[channel], vc4_state->crtc_w);
502 	}
503 
504 	/* Ch0 V-TPZ Words 0-2: Scaling Parameters, Recip, Context */
505 	if (vc4_state->y_scaling[channel] == VC4_SCALING_TPZ) {
506 		vc4_write_tpz(vc4_state,
507 			      vc4_state->src_h[channel], vc4_state->crtc_h);
508 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
509 	}
510 }
511 
512 static void vc4_plane_calc_load(struct drm_plane_state *state)
513 {
514 	unsigned int hvs_load_shift, vrefresh, i;
515 	struct drm_framebuffer *fb = state->fb;
516 	struct vc4_plane_state *vc4_state;
517 	struct drm_crtc_state *crtc_state;
518 	unsigned int vscale_factor;
519 	struct vc4_dev *vc4;
520 
521 	vc4 = to_vc4_dev(state->plane->dev);
522 	if (!vc4->load_tracker_available)
523 		return;
524 
525 	vc4_state = to_vc4_plane_state(state);
526 	crtc_state = drm_atomic_get_existing_crtc_state(state->state,
527 							state->crtc);
528 	vrefresh = drm_mode_vrefresh(&crtc_state->adjusted_mode);
529 
530 	/* The HVS is able to process 2 pixels/cycle when scaling the source,
531 	 * 4 pixels/cycle otherwise.
532 	 * Alpha blending step seems to be pipelined and it's always operating
533 	 * at 4 pixels/cycle, so the limiting aspect here seems to be the
534 	 * scaler block.
535 	 * HVS load is expressed in clk-cycles/sec (AKA Hz).
536 	 */
537 	if (vc4_state->x_scaling[0] != VC4_SCALING_NONE ||
538 	    vc4_state->x_scaling[1] != VC4_SCALING_NONE ||
539 	    vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
540 	    vc4_state->y_scaling[1] != VC4_SCALING_NONE)
541 		hvs_load_shift = 1;
542 	else
543 		hvs_load_shift = 2;
544 
545 	vc4_state->membus_load = 0;
546 	vc4_state->hvs_load = 0;
547 	for (i = 0; i < fb->format->num_planes; i++) {
548 		/* Even if the bandwidth/plane required for a single frame is
549 		 *
550 		 * vc4_state->src_w[i] * vc4_state->src_h[i] * cpp * vrefresh
551 		 *
552 		 * when downscaling, we have to read more pixels per line in
553 		 * the time frame reserved for a single line, so the bandwidth
554 		 * demand can be punctually higher. To account for that, we
555 		 * calculate the down-scaling factor and multiply the plane
556 		 * load by this number. We're likely over-estimating the read
557 		 * demand, but that's better than under-estimating it.
558 		 */
559 		vscale_factor = DIV_ROUND_UP(vc4_state->src_h[i],
560 					     vc4_state->crtc_h);
561 		vc4_state->membus_load += vc4_state->src_w[i] *
562 					  vc4_state->src_h[i] * vscale_factor *
563 					  fb->format->cpp[i];
564 		vc4_state->hvs_load += vc4_state->crtc_h * vc4_state->crtc_w;
565 	}
566 
567 	vc4_state->hvs_load *= vrefresh;
568 	vc4_state->hvs_load >>= hvs_load_shift;
569 	vc4_state->membus_load *= vrefresh;
570 }
571 
572 static int vc4_plane_allocate_lbm(struct drm_plane_state *state)
573 {
574 	struct vc4_dev *vc4 = to_vc4_dev(state->plane->dev);
575 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
576 	unsigned long irqflags;
577 	u32 lbm_size;
578 
579 	lbm_size = vc4_lbm_size(state);
580 	if (!lbm_size)
581 		return 0;
582 
583 	if (WARN_ON(!vc4_state->lbm_offset))
584 		return -EINVAL;
585 
586 	/* Allocate the LBM memory that the HVS will use for temporary
587 	 * storage due to our scaling/format conversion.
588 	 */
589 	if (!drm_mm_node_allocated(&vc4_state->lbm)) {
590 		int ret;
591 
592 		spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
593 		ret = drm_mm_insert_node_generic(&vc4->hvs->lbm_mm,
594 						 &vc4_state->lbm,
595 						 lbm_size,
596 						 vc4->hvs->hvs5 ? 64 : 32,
597 						 0, 0);
598 		spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
599 
600 		if (ret)
601 			return ret;
602 	} else {
603 		WARN_ON_ONCE(lbm_size != vc4_state->lbm.size);
604 	}
605 
606 	vc4_state->dlist[vc4_state->lbm_offset] = vc4_state->lbm.start;
607 
608 	return 0;
609 }
610 
611 /* Writes out a full display list for an active plane to the plane's
612  * private dlist state.
613  */
614 static int vc4_plane_mode_set(struct drm_plane *plane,
615 			      struct drm_plane_state *state)
616 {
617 	struct vc4_dev *vc4 = to_vc4_dev(plane->dev);
618 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
619 	struct drm_framebuffer *fb = state->fb;
620 	u32 ctl0_offset = vc4_state->dlist_count;
621 	const struct hvs_format *format = vc4_get_hvs_format(fb->format->format);
622 	u64 base_format_mod = fourcc_mod_broadcom_mod(fb->modifier);
623 	int num_planes = fb->format->num_planes;
624 	u32 h_subsample = fb->format->hsub;
625 	u32 v_subsample = fb->format->vsub;
626 	bool mix_plane_alpha;
627 	bool covers_screen;
628 	u32 scl0, scl1, pitch0;
629 	u32 tiling, src_y;
630 	u32 hvs_format = format->hvs;
631 	unsigned int rotation;
632 	int ret, i;
633 
634 	if (vc4_state->dlist_initialized)
635 		return 0;
636 
637 	ret = vc4_plane_setup_clipping_and_scaling(state);
638 	if (ret)
639 		return ret;
640 
641 	/* SCL1 is used for Cb/Cr scaling of planar formats.  For RGB
642 	 * and 4:4:4, scl1 should be set to scl0 so both channels of
643 	 * the scaler do the same thing.  For YUV, the Y plane needs
644 	 * to be put in channel 1 and Cb/Cr in channel 0, so we swap
645 	 * the scl fields here.
646 	 */
647 	if (num_planes == 1) {
648 		scl0 = vc4_get_scl_field(state, 0);
649 		scl1 = scl0;
650 	} else {
651 		scl0 = vc4_get_scl_field(state, 1);
652 		scl1 = vc4_get_scl_field(state, 0);
653 	}
654 
655 	rotation = drm_rotation_simplify(state->rotation,
656 					 DRM_MODE_ROTATE_0 |
657 					 DRM_MODE_REFLECT_X |
658 					 DRM_MODE_REFLECT_Y);
659 
660 	/* We must point to the last line when Y reflection is enabled. */
661 	src_y = vc4_state->src_y;
662 	if (rotation & DRM_MODE_REFLECT_Y)
663 		src_y += vc4_state->src_h[0] - 1;
664 
665 	switch (base_format_mod) {
666 	case DRM_FORMAT_MOD_LINEAR:
667 		tiling = SCALER_CTL0_TILING_LINEAR;
668 		pitch0 = VC4_SET_FIELD(fb->pitches[0], SCALER_SRC_PITCH);
669 
670 		/* Adjust the base pointer to the first pixel to be scanned
671 		 * out.
672 		 */
673 		for (i = 0; i < num_planes; i++) {
674 			vc4_state->offsets[i] += src_y /
675 						 (i ? v_subsample : 1) *
676 						 fb->pitches[i];
677 
678 			vc4_state->offsets[i] += vc4_state->src_x /
679 						 (i ? h_subsample : 1) *
680 						 fb->format->cpp[i];
681 		}
682 
683 		break;
684 
685 	case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED: {
686 		u32 tile_size_shift = 12; /* T tiles are 4kb */
687 		/* Whole-tile offsets, mostly for setting the pitch. */
688 		u32 tile_w_shift = fb->format->cpp[0] == 2 ? 6 : 5;
689 		u32 tile_h_shift = 5; /* 16 and 32bpp are 32 pixels high */
690 		u32 tile_w_mask = (1 << tile_w_shift) - 1;
691 		/* The height mask on 32-bit-per-pixel tiles is 63, i.e. twice
692 		 * the height (in pixels) of a 4k tile.
693 		 */
694 		u32 tile_h_mask = (2 << tile_h_shift) - 1;
695 		/* For T-tiled, the FB pitch is "how many bytes from one row to
696 		 * the next, such that
697 		 *
698 		 *	pitch * tile_h == tile_size * tiles_per_row
699 		 */
700 		u32 tiles_w = fb->pitches[0] >> (tile_size_shift - tile_h_shift);
701 		u32 tiles_l = vc4_state->src_x >> tile_w_shift;
702 		u32 tiles_r = tiles_w - tiles_l;
703 		u32 tiles_t = src_y >> tile_h_shift;
704 		/* Intra-tile offsets, which modify the base address (the
705 		 * SCALER_PITCH0_TILE_Y_OFFSET tells HVS how to walk from that
706 		 * base address).
707 		 */
708 		u32 tile_y = (src_y >> 4) & 1;
709 		u32 subtile_y = (src_y >> 2) & 3;
710 		u32 utile_y = src_y & 3;
711 		u32 x_off = vc4_state->src_x & tile_w_mask;
712 		u32 y_off = src_y & tile_h_mask;
713 
714 		/* When Y reflection is requested we must set the
715 		 * SCALER_PITCH0_TILE_LINE_DIR flag to tell HVS that all lines
716 		 * after the initial one should be fetched in descending order,
717 		 * which makes sense since we start from the last line and go
718 		 * backward.
719 		 * Don't know why we need y_off = max_y_off - y_off, but it's
720 		 * definitely required (I guess it's also related to the "going
721 		 * backward" situation).
722 		 */
723 		if (rotation & DRM_MODE_REFLECT_Y) {
724 			y_off = tile_h_mask - y_off;
725 			pitch0 = SCALER_PITCH0_TILE_LINE_DIR;
726 		} else {
727 			pitch0 = 0;
728 		}
729 
730 		tiling = SCALER_CTL0_TILING_256B_OR_T;
731 		pitch0 |= (VC4_SET_FIELD(x_off, SCALER_PITCH0_SINK_PIX) |
732 			   VC4_SET_FIELD(y_off, SCALER_PITCH0_TILE_Y_OFFSET) |
733 			   VC4_SET_FIELD(tiles_l, SCALER_PITCH0_TILE_WIDTH_L) |
734 			   VC4_SET_FIELD(tiles_r, SCALER_PITCH0_TILE_WIDTH_R));
735 		vc4_state->offsets[0] += tiles_t * (tiles_w << tile_size_shift);
736 		vc4_state->offsets[0] += subtile_y << 8;
737 		vc4_state->offsets[0] += utile_y << 4;
738 
739 		/* Rows of tiles alternate left-to-right and right-to-left. */
740 		if (tiles_t & 1) {
741 			pitch0 |= SCALER_PITCH0_TILE_INITIAL_LINE_DIR;
742 			vc4_state->offsets[0] += (tiles_w - tiles_l) <<
743 						 tile_size_shift;
744 			vc4_state->offsets[0] -= (1 + !tile_y) << 10;
745 		} else {
746 			vc4_state->offsets[0] += tiles_l << tile_size_shift;
747 			vc4_state->offsets[0] += tile_y << 10;
748 		}
749 
750 		break;
751 	}
752 
753 	case DRM_FORMAT_MOD_BROADCOM_SAND64:
754 	case DRM_FORMAT_MOD_BROADCOM_SAND128:
755 	case DRM_FORMAT_MOD_BROADCOM_SAND256: {
756 		uint32_t param = fourcc_mod_broadcom_param(fb->modifier);
757 		u32 tile_w, tile, x_off, pix_per_tile;
758 
759 		hvs_format = HVS_PIXEL_FORMAT_H264;
760 
761 		switch (base_format_mod) {
762 		case DRM_FORMAT_MOD_BROADCOM_SAND64:
763 			tiling = SCALER_CTL0_TILING_64B;
764 			tile_w = 64;
765 			break;
766 		case DRM_FORMAT_MOD_BROADCOM_SAND128:
767 			tiling = SCALER_CTL0_TILING_128B;
768 			tile_w = 128;
769 			break;
770 		case DRM_FORMAT_MOD_BROADCOM_SAND256:
771 			tiling = SCALER_CTL0_TILING_256B_OR_T;
772 			tile_w = 256;
773 			break;
774 		default:
775 			break;
776 		}
777 
778 		if (param > SCALER_TILE_HEIGHT_MASK) {
779 			DRM_DEBUG_KMS("SAND height too large (%d)\n", param);
780 			return -EINVAL;
781 		}
782 
783 		pix_per_tile = tile_w / fb->format->cpp[0];
784 		tile = vc4_state->src_x / pix_per_tile;
785 		x_off = vc4_state->src_x % pix_per_tile;
786 
787 		/* Adjust the base pointer to the first pixel to be scanned
788 		 * out.
789 		 */
790 		for (i = 0; i < num_planes; i++) {
791 			vc4_state->offsets[i] += param * tile_w * tile;
792 			vc4_state->offsets[i] += src_y /
793 						 (i ? v_subsample : 1) *
794 						 tile_w;
795 			vc4_state->offsets[i] += x_off /
796 						 (i ? h_subsample : 1) *
797 						 fb->format->cpp[i];
798 		}
799 
800 		pitch0 = VC4_SET_FIELD(param, SCALER_TILE_HEIGHT);
801 		break;
802 	}
803 
804 	default:
805 		DRM_DEBUG_KMS("Unsupported FB tiling flag 0x%16llx",
806 			      (long long)fb->modifier);
807 		return -EINVAL;
808 	}
809 
810 	/* Don't waste cycles mixing with plane alpha if the set alpha
811 	 * is opaque or there is no per-pixel alpha information.
812 	 * In any case we use the alpha property value as the fixed alpha.
813 	 */
814 	mix_plane_alpha = state->alpha != DRM_BLEND_ALPHA_OPAQUE &&
815 			  fb->format->has_alpha;
816 
817 	if (!vc4->hvs->hvs5) {
818 	/* Control word */
819 		vc4_dlist_write(vc4_state,
820 				SCALER_CTL0_VALID |
821 				(rotation & DRM_MODE_REFLECT_X ? SCALER_CTL0_HFLIP : 0) |
822 				(rotation & DRM_MODE_REFLECT_Y ? SCALER_CTL0_VFLIP : 0) |
823 				VC4_SET_FIELD(SCALER_CTL0_RGBA_EXPAND_ROUND, SCALER_CTL0_RGBA_EXPAND) |
824 				(format->pixel_order << SCALER_CTL0_ORDER_SHIFT) |
825 				(hvs_format << SCALER_CTL0_PIXEL_FORMAT_SHIFT) |
826 				VC4_SET_FIELD(tiling, SCALER_CTL0_TILING) |
827 				(vc4_state->is_unity ? SCALER_CTL0_UNITY : 0) |
828 				VC4_SET_FIELD(scl0, SCALER_CTL0_SCL0) |
829 				VC4_SET_FIELD(scl1, SCALER_CTL0_SCL1));
830 
831 		/* Position Word 0: Image Positions and Alpha Value */
832 		vc4_state->pos0_offset = vc4_state->dlist_count;
833 		vc4_dlist_write(vc4_state,
834 				VC4_SET_FIELD(state->alpha >> 8, SCALER_POS0_FIXED_ALPHA) |
835 				VC4_SET_FIELD(vc4_state->crtc_x, SCALER_POS0_START_X) |
836 				VC4_SET_FIELD(vc4_state->crtc_y, SCALER_POS0_START_Y));
837 
838 		/* Position Word 1: Scaled Image Dimensions. */
839 		if (!vc4_state->is_unity) {
840 			vc4_dlist_write(vc4_state,
841 					VC4_SET_FIELD(vc4_state->crtc_w,
842 						      SCALER_POS1_SCL_WIDTH) |
843 					VC4_SET_FIELD(vc4_state->crtc_h,
844 						      SCALER_POS1_SCL_HEIGHT));
845 		}
846 
847 		/* Position Word 2: Source Image Size, Alpha */
848 		vc4_state->pos2_offset = vc4_state->dlist_count;
849 		vc4_dlist_write(vc4_state,
850 				VC4_SET_FIELD(fb->format->has_alpha ?
851 					      SCALER_POS2_ALPHA_MODE_PIPELINE :
852 					      SCALER_POS2_ALPHA_MODE_FIXED,
853 					      SCALER_POS2_ALPHA_MODE) |
854 				(mix_plane_alpha ? SCALER_POS2_ALPHA_MIX : 0) |
855 				(fb->format->has_alpha ?
856 						SCALER_POS2_ALPHA_PREMULT : 0) |
857 				VC4_SET_FIELD(vc4_state->src_w[0],
858 					      SCALER_POS2_WIDTH) |
859 				VC4_SET_FIELD(vc4_state->src_h[0],
860 					      SCALER_POS2_HEIGHT));
861 
862 		/* Position Word 3: Context.  Written by the HVS. */
863 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
864 
865 	} else {
866 		u32 hvs_pixel_order = format->pixel_order;
867 
868 		if (format->pixel_order_hvs5)
869 			hvs_pixel_order = format->pixel_order_hvs5;
870 
871 		/* Control word */
872 		vc4_dlist_write(vc4_state,
873 				SCALER_CTL0_VALID |
874 				(hvs_pixel_order << SCALER_CTL0_ORDER_SHIFT) |
875 				(hvs_format << SCALER_CTL0_PIXEL_FORMAT_SHIFT) |
876 				VC4_SET_FIELD(tiling, SCALER_CTL0_TILING) |
877 				(vc4_state->is_unity ?
878 						SCALER5_CTL0_UNITY : 0) |
879 				VC4_SET_FIELD(scl0, SCALER_CTL0_SCL0) |
880 				VC4_SET_FIELD(scl1, SCALER_CTL0_SCL1) |
881 				SCALER5_CTL0_ALPHA_EXPAND |
882 				SCALER5_CTL0_RGB_EXPAND);
883 
884 		/* Position Word 0: Image Positions and Alpha Value */
885 		vc4_state->pos0_offset = vc4_state->dlist_count;
886 		vc4_dlist_write(vc4_state,
887 				(rotation & DRM_MODE_REFLECT_Y ?
888 						SCALER5_POS0_VFLIP : 0) |
889 				VC4_SET_FIELD(vc4_state->crtc_x,
890 					      SCALER_POS0_START_X) |
891 				(rotation & DRM_MODE_REFLECT_X ?
892 					      SCALER5_POS0_HFLIP : 0) |
893 				VC4_SET_FIELD(vc4_state->crtc_y,
894 					      SCALER5_POS0_START_Y)
895 			       );
896 
897 		/* Control Word 2 */
898 		vc4_dlist_write(vc4_state,
899 				VC4_SET_FIELD(state->alpha >> 4,
900 					      SCALER5_CTL2_ALPHA) |
901 				(fb->format->has_alpha ?
902 					SCALER5_CTL2_ALPHA_PREMULT : 0) |
903 				(mix_plane_alpha ?
904 					SCALER5_CTL2_ALPHA_MIX : 0) |
905 				VC4_SET_FIELD(fb->format->has_alpha ?
906 				      SCALER5_CTL2_ALPHA_MODE_PIPELINE :
907 				      SCALER5_CTL2_ALPHA_MODE_FIXED,
908 				      SCALER5_CTL2_ALPHA_MODE)
909 			       );
910 
911 		/* Position Word 1: Scaled Image Dimensions. */
912 		if (!vc4_state->is_unity) {
913 			vc4_dlist_write(vc4_state,
914 					VC4_SET_FIELD(vc4_state->crtc_w,
915 						      SCALER_POS1_SCL_WIDTH) |
916 					VC4_SET_FIELD(vc4_state->crtc_h,
917 						      SCALER_POS1_SCL_HEIGHT));
918 		}
919 
920 		/* Position Word 2: Source Image Size */
921 		vc4_state->pos2_offset = vc4_state->dlist_count;
922 		vc4_dlist_write(vc4_state,
923 				VC4_SET_FIELD(vc4_state->src_w[0],
924 					      SCALER5_POS2_WIDTH) |
925 				VC4_SET_FIELD(vc4_state->src_h[0],
926 					      SCALER5_POS2_HEIGHT));
927 
928 		/* Position Word 3: Context.  Written by the HVS. */
929 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
930 	}
931 
932 
933 	/* Pointer Word 0/1/2: RGB / Y / Cb / Cr Pointers
934 	 *
935 	 * The pointers may be any byte address.
936 	 */
937 	vc4_state->ptr0_offset = vc4_state->dlist_count;
938 	for (i = 0; i < num_planes; i++)
939 		vc4_dlist_write(vc4_state, vc4_state->offsets[i]);
940 
941 	/* Pointer Context Word 0/1/2: Written by the HVS */
942 	for (i = 0; i < num_planes; i++)
943 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
944 
945 	/* Pitch word 0 */
946 	vc4_dlist_write(vc4_state, pitch0);
947 
948 	/* Pitch word 1/2 */
949 	for (i = 1; i < num_planes; i++) {
950 		if (hvs_format != HVS_PIXEL_FORMAT_H264) {
951 			vc4_dlist_write(vc4_state,
952 					VC4_SET_FIELD(fb->pitches[i],
953 						      SCALER_SRC_PITCH));
954 		} else {
955 			vc4_dlist_write(vc4_state, pitch0);
956 		}
957 	}
958 
959 	/* Colorspace conversion words */
960 	if (vc4_state->is_yuv) {
961 		vc4_dlist_write(vc4_state, SCALER_CSC0_ITR_R_601_5);
962 		vc4_dlist_write(vc4_state, SCALER_CSC1_ITR_R_601_5);
963 		vc4_dlist_write(vc4_state, SCALER_CSC2_ITR_R_601_5);
964 	}
965 
966 	vc4_state->lbm_offset = 0;
967 
968 	if (vc4_state->x_scaling[0] != VC4_SCALING_NONE ||
969 	    vc4_state->x_scaling[1] != VC4_SCALING_NONE ||
970 	    vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
971 	    vc4_state->y_scaling[1] != VC4_SCALING_NONE) {
972 		/* Reserve a slot for the LBM Base Address. The real value will
973 		 * be set when calling vc4_plane_allocate_lbm().
974 		 */
975 		if (vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
976 		    vc4_state->y_scaling[1] != VC4_SCALING_NONE)
977 			vc4_state->lbm_offset = vc4_state->dlist_count++;
978 
979 		if (num_planes > 1) {
980 			/* Emit Cb/Cr as channel 0 and Y as channel
981 			 * 1. This matches how we set up scl0/scl1
982 			 * above.
983 			 */
984 			vc4_write_scaling_parameters(state, 1);
985 		}
986 		vc4_write_scaling_parameters(state, 0);
987 
988 		/* If any PPF setup was done, then all the kernel
989 		 * pointers get uploaded.
990 		 */
991 		if (vc4_state->x_scaling[0] == VC4_SCALING_PPF ||
992 		    vc4_state->y_scaling[0] == VC4_SCALING_PPF ||
993 		    vc4_state->x_scaling[1] == VC4_SCALING_PPF ||
994 		    vc4_state->y_scaling[1] == VC4_SCALING_PPF) {
995 			u32 kernel = VC4_SET_FIELD(vc4->hvs->mitchell_netravali_filter.start,
996 						   SCALER_PPF_KERNEL_OFFSET);
997 
998 			/* HPPF plane 0 */
999 			vc4_dlist_write(vc4_state, kernel);
1000 			/* VPPF plane 0 */
1001 			vc4_dlist_write(vc4_state, kernel);
1002 			/* HPPF plane 1 */
1003 			vc4_dlist_write(vc4_state, kernel);
1004 			/* VPPF plane 1 */
1005 			vc4_dlist_write(vc4_state, kernel);
1006 		}
1007 	}
1008 
1009 	vc4_state->dlist[ctl0_offset] |=
1010 		VC4_SET_FIELD(vc4_state->dlist_count, SCALER_CTL0_SIZE);
1011 
1012 	/* crtc_* are already clipped coordinates. */
1013 	covers_screen = vc4_state->crtc_x == 0 && vc4_state->crtc_y == 0 &&
1014 			vc4_state->crtc_w == state->crtc->mode.hdisplay &&
1015 			vc4_state->crtc_h == state->crtc->mode.vdisplay;
1016 	/* Background fill might be necessary when the plane has per-pixel
1017 	 * alpha content or a non-opaque plane alpha and could blend from the
1018 	 * background or does not cover the entire screen.
1019 	 */
1020 	vc4_state->needs_bg_fill = fb->format->has_alpha || !covers_screen ||
1021 				   state->alpha != DRM_BLEND_ALPHA_OPAQUE;
1022 
1023 	/* Flag the dlist as initialized to avoid checking it twice in case
1024 	 * the async update check already called vc4_plane_mode_set() and
1025 	 * decided to fallback to sync update because async update was not
1026 	 * possible.
1027 	 */
1028 	vc4_state->dlist_initialized = 1;
1029 
1030 	vc4_plane_calc_load(state);
1031 
1032 	return 0;
1033 }
1034 
1035 /* If a modeset involves changing the setup of a plane, the atomic
1036  * infrastructure will call this to validate a proposed plane setup.
1037  * However, if a plane isn't getting updated, this (and the
1038  * corresponding vc4_plane_atomic_update) won't get called.  Thus, we
1039  * compute the dlist here and have all active plane dlists get updated
1040  * in the CRTC's flush.
1041  */
1042 static int vc4_plane_atomic_check(struct drm_plane *plane,
1043 				  struct drm_plane_state *state)
1044 {
1045 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
1046 	int ret;
1047 
1048 	vc4_state->dlist_count = 0;
1049 
1050 	if (!plane_enabled(state))
1051 		return 0;
1052 
1053 	ret = vc4_plane_mode_set(plane, state);
1054 	if (ret)
1055 		return ret;
1056 
1057 	return vc4_plane_allocate_lbm(state);
1058 }
1059 
1060 static void vc4_plane_atomic_update(struct drm_plane *plane,
1061 				    struct drm_plane_state *old_state)
1062 {
1063 	/* No contents here.  Since we don't know where in the CRTC's
1064 	 * dlist we should be stored, our dlist is uploaded to the
1065 	 * hardware with vc4_plane_write_dlist() at CRTC atomic_flush
1066 	 * time.
1067 	 */
1068 }
1069 
1070 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist)
1071 {
1072 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
1073 	int i;
1074 
1075 	vc4_state->hw_dlist = dlist;
1076 
1077 	/* Can't memcpy_toio() because it needs to be 32-bit writes. */
1078 	for (i = 0; i < vc4_state->dlist_count; i++)
1079 		writel(vc4_state->dlist[i], &dlist[i]);
1080 
1081 	return vc4_state->dlist_count;
1082 }
1083 
1084 u32 vc4_plane_dlist_size(const struct drm_plane_state *state)
1085 {
1086 	const struct vc4_plane_state *vc4_state =
1087 		container_of(state, typeof(*vc4_state), base);
1088 
1089 	return vc4_state->dlist_count;
1090 }
1091 
1092 /* Updates the plane to immediately (well, once the FIFO needs
1093  * refilling) scan out from at a new framebuffer.
1094  */
1095 void vc4_plane_async_set_fb(struct drm_plane *plane, struct drm_framebuffer *fb)
1096 {
1097 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
1098 	struct drm_gem_cma_object *bo = drm_fb_cma_get_gem_obj(fb, 0);
1099 	uint32_t addr;
1100 
1101 	/* We're skipping the address adjustment for negative origin,
1102 	 * because this is only called on the primary plane.
1103 	 */
1104 	WARN_ON_ONCE(plane->state->crtc_x < 0 || plane->state->crtc_y < 0);
1105 	addr = bo->paddr + fb->offsets[0];
1106 
1107 	/* Write the new address into the hardware immediately.  The
1108 	 * scanout will start from this address as soon as the FIFO
1109 	 * needs to refill with pixels.
1110 	 */
1111 	writel(addr, &vc4_state->hw_dlist[vc4_state->ptr0_offset]);
1112 
1113 	/* Also update the CPU-side dlist copy, so that any later
1114 	 * atomic updates that don't do a new modeset on our plane
1115 	 * also use our updated address.
1116 	 */
1117 	vc4_state->dlist[vc4_state->ptr0_offset] = addr;
1118 }
1119 
1120 static void vc4_plane_atomic_async_update(struct drm_plane *plane,
1121 					  struct drm_plane_state *state)
1122 {
1123 	struct vc4_plane_state *vc4_state, *new_vc4_state;
1124 
1125 	swap(plane->state->fb, state->fb);
1126 	plane->state->crtc_x = state->crtc_x;
1127 	plane->state->crtc_y = state->crtc_y;
1128 	plane->state->crtc_w = state->crtc_w;
1129 	plane->state->crtc_h = state->crtc_h;
1130 	plane->state->src_x = state->src_x;
1131 	plane->state->src_y = state->src_y;
1132 	plane->state->src_w = state->src_w;
1133 	plane->state->src_h = state->src_h;
1134 	plane->state->src_h = state->src_h;
1135 	plane->state->alpha = state->alpha;
1136 	plane->state->pixel_blend_mode = state->pixel_blend_mode;
1137 	plane->state->rotation = state->rotation;
1138 	plane->state->zpos = state->zpos;
1139 	plane->state->normalized_zpos = state->normalized_zpos;
1140 	plane->state->color_encoding = state->color_encoding;
1141 	plane->state->color_range = state->color_range;
1142 	plane->state->src = state->src;
1143 	plane->state->dst = state->dst;
1144 	plane->state->visible = state->visible;
1145 
1146 	new_vc4_state = to_vc4_plane_state(state);
1147 	vc4_state = to_vc4_plane_state(plane->state);
1148 
1149 	vc4_state->crtc_x = new_vc4_state->crtc_x;
1150 	vc4_state->crtc_y = new_vc4_state->crtc_y;
1151 	vc4_state->crtc_h = new_vc4_state->crtc_h;
1152 	vc4_state->crtc_w = new_vc4_state->crtc_w;
1153 	vc4_state->src_x = new_vc4_state->src_x;
1154 	vc4_state->src_y = new_vc4_state->src_y;
1155 	memcpy(vc4_state->src_w, new_vc4_state->src_w,
1156 	       sizeof(vc4_state->src_w));
1157 	memcpy(vc4_state->src_h, new_vc4_state->src_h,
1158 	       sizeof(vc4_state->src_h));
1159 	memcpy(vc4_state->x_scaling, new_vc4_state->x_scaling,
1160 	       sizeof(vc4_state->x_scaling));
1161 	memcpy(vc4_state->y_scaling, new_vc4_state->y_scaling,
1162 	       sizeof(vc4_state->y_scaling));
1163 	vc4_state->is_unity = new_vc4_state->is_unity;
1164 	vc4_state->is_yuv = new_vc4_state->is_yuv;
1165 	memcpy(vc4_state->offsets, new_vc4_state->offsets,
1166 	       sizeof(vc4_state->offsets));
1167 	vc4_state->needs_bg_fill = new_vc4_state->needs_bg_fill;
1168 
1169 	/* Update the current vc4_state pos0, pos2 and ptr0 dlist entries. */
1170 	vc4_state->dlist[vc4_state->pos0_offset] =
1171 		new_vc4_state->dlist[vc4_state->pos0_offset];
1172 	vc4_state->dlist[vc4_state->pos2_offset] =
1173 		new_vc4_state->dlist[vc4_state->pos2_offset];
1174 	vc4_state->dlist[vc4_state->ptr0_offset] =
1175 		new_vc4_state->dlist[vc4_state->ptr0_offset];
1176 
1177 	/* Note that we can't just call vc4_plane_write_dlist()
1178 	 * because that would smash the context data that the HVS is
1179 	 * currently using.
1180 	 */
1181 	writel(vc4_state->dlist[vc4_state->pos0_offset],
1182 	       &vc4_state->hw_dlist[vc4_state->pos0_offset]);
1183 	writel(vc4_state->dlist[vc4_state->pos2_offset],
1184 	       &vc4_state->hw_dlist[vc4_state->pos2_offset]);
1185 	writel(vc4_state->dlist[vc4_state->ptr0_offset],
1186 	       &vc4_state->hw_dlist[vc4_state->ptr0_offset]);
1187 }
1188 
1189 static int vc4_plane_atomic_async_check(struct drm_plane *plane,
1190 					struct drm_plane_state *state)
1191 {
1192 	struct vc4_plane_state *old_vc4_state, *new_vc4_state;
1193 	int ret;
1194 	u32 i;
1195 
1196 	ret = vc4_plane_mode_set(plane, state);
1197 	if (ret)
1198 		return ret;
1199 
1200 	old_vc4_state = to_vc4_plane_state(plane->state);
1201 	new_vc4_state = to_vc4_plane_state(state);
1202 	if (old_vc4_state->dlist_count != new_vc4_state->dlist_count ||
1203 	    old_vc4_state->pos0_offset != new_vc4_state->pos0_offset ||
1204 	    old_vc4_state->pos2_offset != new_vc4_state->pos2_offset ||
1205 	    old_vc4_state->ptr0_offset != new_vc4_state->ptr0_offset ||
1206 	    vc4_lbm_size(plane->state) != vc4_lbm_size(state))
1207 		return -EINVAL;
1208 
1209 	/* Only pos0, pos2 and ptr0 DWORDS can be updated in an async update
1210 	 * if anything else has changed, fallback to a sync update.
1211 	 */
1212 	for (i = 0; i < new_vc4_state->dlist_count; i++) {
1213 		if (i == new_vc4_state->pos0_offset ||
1214 		    i == new_vc4_state->pos2_offset ||
1215 		    i == new_vc4_state->ptr0_offset ||
1216 		    (new_vc4_state->lbm_offset &&
1217 		     i == new_vc4_state->lbm_offset))
1218 			continue;
1219 
1220 		if (new_vc4_state->dlist[i] != old_vc4_state->dlist[i])
1221 			return -EINVAL;
1222 	}
1223 
1224 	return 0;
1225 }
1226 
1227 static int vc4_prepare_fb(struct drm_plane *plane,
1228 			  struct drm_plane_state *state)
1229 {
1230 	struct vc4_bo *bo;
1231 	int ret;
1232 
1233 	if (!state->fb)
1234 		return 0;
1235 
1236 	bo = to_vc4_bo(&drm_fb_cma_get_gem_obj(state->fb, 0)->base);
1237 
1238 	drm_gem_fb_prepare_fb(plane, state);
1239 
1240 	if (plane->state->fb == state->fb)
1241 		return 0;
1242 
1243 	ret = vc4_bo_inc_usecnt(bo);
1244 	if (ret)
1245 		return ret;
1246 
1247 	return 0;
1248 }
1249 
1250 static void vc4_cleanup_fb(struct drm_plane *plane,
1251 			   struct drm_plane_state *state)
1252 {
1253 	struct vc4_bo *bo;
1254 
1255 	if (plane->state->fb == state->fb || !state->fb)
1256 		return;
1257 
1258 	bo = to_vc4_bo(&drm_fb_cma_get_gem_obj(state->fb, 0)->base);
1259 	vc4_bo_dec_usecnt(bo);
1260 }
1261 
1262 static const struct drm_plane_helper_funcs vc4_plane_helper_funcs = {
1263 	.atomic_check = vc4_plane_atomic_check,
1264 	.atomic_update = vc4_plane_atomic_update,
1265 	.prepare_fb = vc4_prepare_fb,
1266 	.cleanup_fb = vc4_cleanup_fb,
1267 	.atomic_async_check = vc4_plane_atomic_async_check,
1268 	.atomic_async_update = vc4_plane_atomic_async_update,
1269 };
1270 
1271 static bool vc4_format_mod_supported(struct drm_plane *plane,
1272 				     uint32_t format,
1273 				     uint64_t modifier)
1274 {
1275 	/* Support T_TILING for RGB formats only. */
1276 	switch (format) {
1277 	case DRM_FORMAT_XRGB8888:
1278 	case DRM_FORMAT_ARGB8888:
1279 	case DRM_FORMAT_ABGR8888:
1280 	case DRM_FORMAT_XBGR8888:
1281 	case DRM_FORMAT_RGB565:
1282 	case DRM_FORMAT_BGR565:
1283 	case DRM_FORMAT_ARGB1555:
1284 	case DRM_FORMAT_XRGB1555:
1285 		switch (fourcc_mod_broadcom_mod(modifier)) {
1286 		case DRM_FORMAT_MOD_LINEAR:
1287 		case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
1288 			return true;
1289 		default:
1290 			return false;
1291 		}
1292 	case DRM_FORMAT_NV12:
1293 	case DRM_FORMAT_NV21:
1294 		switch (fourcc_mod_broadcom_mod(modifier)) {
1295 		case DRM_FORMAT_MOD_LINEAR:
1296 		case DRM_FORMAT_MOD_BROADCOM_SAND64:
1297 		case DRM_FORMAT_MOD_BROADCOM_SAND128:
1298 		case DRM_FORMAT_MOD_BROADCOM_SAND256:
1299 			return true;
1300 		default:
1301 			return false;
1302 		}
1303 	case DRM_FORMAT_RGBX1010102:
1304 	case DRM_FORMAT_BGRX1010102:
1305 	case DRM_FORMAT_RGBA1010102:
1306 	case DRM_FORMAT_BGRA1010102:
1307 	case DRM_FORMAT_YUV422:
1308 	case DRM_FORMAT_YVU422:
1309 	case DRM_FORMAT_YUV420:
1310 	case DRM_FORMAT_YVU420:
1311 	case DRM_FORMAT_NV16:
1312 	case DRM_FORMAT_NV61:
1313 	default:
1314 		return (modifier == DRM_FORMAT_MOD_LINEAR);
1315 	}
1316 }
1317 
1318 static const struct drm_plane_funcs vc4_plane_funcs = {
1319 	.update_plane = drm_atomic_helper_update_plane,
1320 	.disable_plane = drm_atomic_helper_disable_plane,
1321 	.destroy = drm_plane_cleanup,
1322 	.set_property = NULL,
1323 	.reset = vc4_plane_reset,
1324 	.atomic_duplicate_state = vc4_plane_duplicate_state,
1325 	.atomic_destroy_state = vc4_plane_destroy_state,
1326 	.format_mod_supported = vc4_format_mod_supported,
1327 };
1328 
1329 struct drm_plane *vc4_plane_init(struct drm_device *dev,
1330 				 enum drm_plane_type type)
1331 {
1332 	struct drm_plane *plane = NULL;
1333 	struct vc4_plane *vc4_plane;
1334 	u32 formats[ARRAY_SIZE(hvs_formats)];
1335 	int ret = 0;
1336 	unsigned i;
1337 	static const uint64_t modifiers[] = {
1338 		DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED,
1339 		DRM_FORMAT_MOD_BROADCOM_SAND128,
1340 		DRM_FORMAT_MOD_BROADCOM_SAND64,
1341 		DRM_FORMAT_MOD_BROADCOM_SAND256,
1342 		DRM_FORMAT_MOD_LINEAR,
1343 		DRM_FORMAT_MOD_INVALID
1344 	};
1345 
1346 	vc4_plane = devm_kzalloc(dev->dev, sizeof(*vc4_plane),
1347 				 GFP_KERNEL);
1348 	if (!vc4_plane)
1349 		return ERR_PTR(-ENOMEM);
1350 
1351 	for (i = 0; i < ARRAY_SIZE(hvs_formats); i++)
1352 		formats[i] = hvs_formats[i].drm;
1353 
1354 	plane = &vc4_plane->base;
1355 	ret = drm_universal_plane_init(dev, plane, 0,
1356 				       &vc4_plane_funcs,
1357 				       formats, ARRAY_SIZE(formats),
1358 				       modifiers, type, NULL);
1359 	if (ret)
1360 		return ERR_PTR(ret);
1361 
1362 	drm_plane_helper_add(plane, &vc4_plane_helper_funcs);
1363 
1364 	drm_plane_create_alpha_property(plane);
1365 	drm_plane_create_rotation_property(plane, DRM_MODE_ROTATE_0,
1366 					   DRM_MODE_ROTATE_0 |
1367 					   DRM_MODE_ROTATE_180 |
1368 					   DRM_MODE_REFLECT_X |
1369 					   DRM_MODE_REFLECT_Y);
1370 
1371 	return plane;
1372 }
1373 
1374 int vc4_plane_create_additional_planes(struct drm_device *drm)
1375 {
1376 	struct drm_plane *cursor_plane;
1377 	struct drm_crtc *crtc;
1378 	unsigned int i;
1379 
1380 	/* Set up some arbitrary number of planes.  We're not limited
1381 	 * by a set number of physical registers, just the space in
1382 	 * the HVS (16k) and how small an plane can be (28 bytes).
1383 	 * However, each plane we set up takes up some memory, and
1384 	 * increases the cost of looping over planes, which atomic
1385 	 * modesetting does quite a bit.  As a result, we pick a
1386 	 * modest number of planes to expose, that should hopefully
1387 	 * still cover any sane usecase.
1388 	 */
1389 	for (i = 0; i < 16; i++) {
1390 		struct drm_plane *plane =
1391 			vc4_plane_init(drm, DRM_PLANE_TYPE_OVERLAY);
1392 
1393 		if (IS_ERR(plane))
1394 			continue;
1395 
1396 		plane->possible_crtcs =
1397 			GENMASK(drm->mode_config.num_crtc - 1, 0);
1398 	}
1399 
1400 	drm_for_each_crtc(crtc, drm) {
1401 		/* Set up the legacy cursor after overlay initialization,
1402 		 * since we overlay planes on the CRTC in the order they were
1403 		 * initialized.
1404 		 */
1405 		cursor_plane = vc4_plane_init(drm, DRM_PLANE_TYPE_CURSOR);
1406 		if (!IS_ERR(cursor_plane)) {
1407 			cursor_plane->possible_crtcs = drm_crtc_mask(crtc);
1408 			crtc->cursor = cursor_plane;
1409 		}
1410 	}
1411 
1412 	return 0;
1413 }
1414