xref: /openbmc/linux/drivers/gpu/drm/vc4/vc4_plane.c (revision dbf563ee)
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 } hvs_formats[] = {
36 	{
37 		.drm = DRM_FORMAT_XRGB8888, .hvs = HVS_PIXEL_FORMAT_RGBA8888,
38 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
39 	},
40 	{
41 		.drm = DRM_FORMAT_ARGB8888, .hvs = HVS_PIXEL_FORMAT_RGBA8888,
42 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
43 	},
44 	{
45 		.drm = DRM_FORMAT_ABGR8888, .hvs = HVS_PIXEL_FORMAT_RGBA8888,
46 		.pixel_order = HVS_PIXEL_ORDER_ARGB,
47 	},
48 	{
49 		.drm = DRM_FORMAT_XBGR8888, .hvs = HVS_PIXEL_FORMAT_RGBA8888,
50 		.pixel_order = HVS_PIXEL_ORDER_ARGB,
51 	},
52 	{
53 		.drm = DRM_FORMAT_RGB565, .hvs = HVS_PIXEL_FORMAT_RGB565,
54 		.pixel_order = HVS_PIXEL_ORDER_XRGB,
55 	},
56 	{
57 		.drm = DRM_FORMAT_BGR565, .hvs = HVS_PIXEL_FORMAT_RGB565,
58 		.pixel_order = HVS_PIXEL_ORDER_XBGR,
59 	},
60 	{
61 		.drm = DRM_FORMAT_ARGB1555, .hvs = HVS_PIXEL_FORMAT_RGBA5551,
62 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
63 	},
64 	{
65 		.drm = DRM_FORMAT_XRGB1555, .hvs = HVS_PIXEL_FORMAT_RGBA5551,
66 		.pixel_order = HVS_PIXEL_ORDER_ABGR,
67 	},
68 	{
69 		.drm = DRM_FORMAT_RGB888, .hvs = HVS_PIXEL_FORMAT_RGB888,
70 		.pixel_order = HVS_PIXEL_ORDER_XRGB,
71 	},
72 	{
73 		.drm = DRM_FORMAT_BGR888, .hvs = HVS_PIXEL_FORMAT_RGB888,
74 		.pixel_order = HVS_PIXEL_ORDER_XBGR,
75 	},
76 	{
77 		.drm = DRM_FORMAT_YUV422,
78 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_3PLANE,
79 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
80 	},
81 	{
82 		.drm = DRM_FORMAT_YVU422,
83 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_3PLANE,
84 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
85 	},
86 	{
87 		.drm = DRM_FORMAT_YUV420,
88 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_3PLANE,
89 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
90 	},
91 	{
92 		.drm = DRM_FORMAT_YVU420,
93 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_3PLANE,
94 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
95 	},
96 	{
97 		.drm = DRM_FORMAT_NV12,
98 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_2PLANE,
99 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
100 	},
101 	{
102 		.drm = DRM_FORMAT_NV21,
103 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_2PLANE,
104 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
105 	},
106 	{
107 		.drm = DRM_FORMAT_NV16,
108 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_2PLANE,
109 		.pixel_order = HVS_PIXEL_ORDER_XYCBCR,
110 	},
111 	{
112 		.drm = DRM_FORMAT_NV61,
113 		.hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_2PLANE,
114 		.pixel_order = HVS_PIXEL_ORDER_XYCRCB,
115 	},
116 };
117 
118 static const struct hvs_format *vc4_get_hvs_format(u32 drm_format)
119 {
120 	unsigned i;
121 
122 	for (i = 0; i < ARRAY_SIZE(hvs_formats); i++) {
123 		if (hvs_formats[i].drm == drm_format)
124 			return &hvs_formats[i];
125 	}
126 
127 	return NULL;
128 }
129 
130 static enum vc4_scaling_mode vc4_get_scaling_mode(u32 src, u32 dst)
131 {
132 	if (dst == src)
133 		return VC4_SCALING_NONE;
134 	if (3 * dst >= 2 * src)
135 		return VC4_SCALING_PPF;
136 	else
137 		return VC4_SCALING_TPZ;
138 }
139 
140 static bool plane_enabled(struct drm_plane_state *state)
141 {
142 	return state->fb && !WARN_ON(!state->crtc);
143 }
144 
145 static struct drm_plane_state *vc4_plane_duplicate_state(struct drm_plane *plane)
146 {
147 	struct vc4_plane_state *vc4_state;
148 
149 	if (WARN_ON(!plane->state))
150 		return NULL;
151 
152 	vc4_state = kmemdup(plane->state, sizeof(*vc4_state), GFP_KERNEL);
153 	if (!vc4_state)
154 		return NULL;
155 
156 	memset(&vc4_state->lbm, 0, sizeof(vc4_state->lbm));
157 	vc4_state->dlist_initialized = 0;
158 
159 	__drm_atomic_helper_plane_duplicate_state(plane, &vc4_state->base);
160 
161 	if (vc4_state->dlist) {
162 		vc4_state->dlist = kmemdup(vc4_state->dlist,
163 					   vc4_state->dlist_count * 4,
164 					   GFP_KERNEL);
165 		if (!vc4_state->dlist) {
166 			kfree(vc4_state);
167 			return NULL;
168 		}
169 		vc4_state->dlist_size = vc4_state->dlist_count;
170 	}
171 
172 	return &vc4_state->base;
173 }
174 
175 static void vc4_plane_destroy_state(struct drm_plane *plane,
176 				    struct drm_plane_state *state)
177 {
178 	struct vc4_dev *vc4 = to_vc4_dev(plane->dev);
179 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
180 
181 	if (drm_mm_node_allocated(&vc4_state->lbm)) {
182 		unsigned long irqflags;
183 
184 		spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
185 		drm_mm_remove_node(&vc4_state->lbm);
186 		spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
187 	}
188 
189 	kfree(vc4_state->dlist);
190 	__drm_atomic_helper_plane_destroy_state(&vc4_state->base);
191 	kfree(state);
192 }
193 
194 /* Called during init to allocate the plane's atomic state. */
195 static void vc4_plane_reset(struct drm_plane *plane)
196 {
197 	struct vc4_plane_state *vc4_state;
198 
199 	WARN_ON(plane->state);
200 
201 	vc4_state = kzalloc(sizeof(*vc4_state), GFP_KERNEL);
202 	if (!vc4_state)
203 		return;
204 
205 	__drm_atomic_helper_plane_reset(plane, &vc4_state->base);
206 }
207 
208 static void vc4_dlist_write(struct vc4_plane_state *vc4_state, u32 val)
209 {
210 	if (vc4_state->dlist_count == vc4_state->dlist_size) {
211 		u32 new_size = max(4u, vc4_state->dlist_count * 2);
212 		u32 *new_dlist = kmalloc_array(new_size, 4, GFP_KERNEL);
213 
214 		if (!new_dlist)
215 			return;
216 		memcpy(new_dlist, vc4_state->dlist, vc4_state->dlist_count * 4);
217 
218 		kfree(vc4_state->dlist);
219 		vc4_state->dlist = new_dlist;
220 		vc4_state->dlist_size = new_size;
221 	}
222 
223 	vc4_state->dlist[vc4_state->dlist_count++] = val;
224 }
225 
226 /* Returns the scl0/scl1 field based on whether the dimensions need to
227  * be up/down/non-scaled.
228  *
229  * This is a replication of a table from the spec.
230  */
231 static u32 vc4_get_scl_field(struct drm_plane_state *state, int plane)
232 {
233 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
234 
235 	switch (vc4_state->x_scaling[plane] << 2 | vc4_state->y_scaling[plane]) {
236 	case VC4_SCALING_PPF << 2 | VC4_SCALING_PPF:
237 		return SCALER_CTL0_SCL_H_PPF_V_PPF;
238 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_PPF:
239 		return SCALER_CTL0_SCL_H_TPZ_V_PPF;
240 	case VC4_SCALING_PPF << 2 | VC4_SCALING_TPZ:
241 		return SCALER_CTL0_SCL_H_PPF_V_TPZ;
242 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_TPZ:
243 		return SCALER_CTL0_SCL_H_TPZ_V_TPZ;
244 	case VC4_SCALING_PPF << 2 | VC4_SCALING_NONE:
245 		return SCALER_CTL0_SCL_H_PPF_V_NONE;
246 	case VC4_SCALING_NONE << 2 | VC4_SCALING_PPF:
247 		return SCALER_CTL0_SCL_H_NONE_V_PPF;
248 	case VC4_SCALING_NONE << 2 | VC4_SCALING_TPZ:
249 		return SCALER_CTL0_SCL_H_NONE_V_TPZ;
250 	case VC4_SCALING_TPZ << 2 | VC4_SCALING_NONE:
251 		return SCALER_CTL0_SCL_H_TPZ_V_NONE;
252 	default:
253 	case VC4_SCALING_NONE << 2 | VC4_SCALING_NONE:
254 		/* The unity case is independently handled by
255 		 * SCALER_CTL0_UNITY.
256 		 */
257 		return 0;
258 	}
259 }
260 
261 static int vc4_plane_margins_adj(struct drm_plane_state *pstate)
262 {
263 	struct vc4_plane_state *vc4_pstate = to_vc4_plane_state(pstate);
264 	unsigned int left, right, top, bottom, adjhdisplay, adjvdisplay;
265 	struct drm_crtc_state *crtc_state;
266 
267 	crtc_state = drm_atomic_get_new_crtc_state(pstate->state,
268 						   pstate->crtc);
269 
270 	vc4_crtc_get_margins(crtc_state, &left, &right, &top, &bottom);
271 	if (!left && !right && !top && !bottom)
272 		return 0;
273 
274 	if (left + right >= crtc_state->mode.hdisplay ||
275 	    top + bottom >= crtc_state->mode.vdisplay)
276 		return -EINVAL;
277 
278 	adjhdisplay = crtc_state->mode.hdisplay - (left + right);
279 	vc4_pstate->crtc_x = DIV_ROUND_CLOSEST(vc4_pstate->crtc_x *
280 					       adjhdisplay,
281 					       crtc_state->mode.hdisplay);
282 	vc4_pstate->crtc_x += left;
283 	if (vc4_pstate->crtc_x > crtc_state->mode.hdisplay - left)
284 		vc4_pstate->crtc_x = crtc_state->mode.hdisplay - left;
285 
286 	adjvdisplay = crtc_state->mode.vdisplay - (top + bottom);
287 	vc4_pstate->crtc_y = DIV_ROUND_CLOSEST(vc4_pstate->crtc_y *
288 					       adjvdisplay,
289 					       crtc_state->mode.vdisplay);
290 	vc4_pstate->crtc_y += top;
291 	if (vc4_pstate->crtc_y > crtc_state->mode.vdisplay - top)
292 		vc4_pstate->crtc_y = crtc_state->mode.vdisplay - top;
293 
294 	vc4_pstate->crtc_w = DIV_ROUND_CLOSEST(vc4_pstate->crtc_w *
295 					       adjhdisplay,
296 					       crtc_state->mode.hdisplay);
297 	vc4_pstate->crtc_h = DIV_ROUND_CLOSEST(vc4_pstate->crtc_h *
298 					       adjvdisplay,
299 					       crtc_state->mode.vdisplay);
300 
301 	if (!vc4_pstate->crtc_w || !vc4_pstate->crtc_h)
302 		return -EINVAL;
303 
304 	return 0;
305 }
306 
307 static int vc4_plane_setup_clipping_and_scaling(struct drm_plane_state *state)
308 {
309 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
310 	struct drm_framebuffer *fb = state->fb;
311 	struct drm_gem_cma_object *bo = drm_fb_cma_get_gem_obj(fb, 0);
312 	u32 subpixel_src_mask = (1 << 16) - 1;
313 	int num_planes = fb->format->num_planes;
314 	struct drm_crtc_state *crtc_state;
315 	u32 h_subsample = fb->format->hsub;
316 	u32 v_subsample = fb->format->vsub;
317 	int i, ret;
318 
319 	crtc_state = drm_atomic_get_existing_crtc_state(state->state,
320 							state->crtc);
321 	if (!crtc_state) {
322 		DRM_DEBUG_KMS("Invalid crtc state\n");
323 		return -EINVAL;
324 	}
325 
326 	ret = drm_atomic_helper_check_plane_state(state, crtc_state, 1,
327 						  INT_MAX, true, true);
328 	if (ret)
329 		return ret;
330 
331 	for (i = 0; i < num_planes; i++)
332 		vc4_state->offsets[i] = bo->paddr + fb->offsets[i];
333 
334 	/* We don't support subpixel source positioning for scaling. */
335 	if ((state->src.x1 & subpixel_src_mask) ||
336 	    (state->src.x2 & subpixel_src_mask) ||
337 	    (state->src.y1 & subpixel_src_mask) ||
338 	    (state->src.y2 & subpixel_src_mask)) {
339 		return -EINVAL;
340 	}
341 
342 	vc4_state->src_x = state->src.x1 >> 16;
343 	vc4_state->src_y = state->src.y1 >> 16;
344 	vc4_state->src_w[0] = (state->src.x2 - state->src.x1) >> 16;
345 	vc4_state->src_h[0] = (state->src.y2 - state->src.y1) >> 16;
346 
347 	vc4_state->crtc_x = state->dst.x1;
348 	vc4_state->crtc_y = state->dst.y1;
349 	vc4_state->crtc_w = state->dst.x2 - state->dst.x1;
350 	vc4_state->crtc_h = state->dst.y2 - state->dst.y1;
351 
352 	ret = vc4_plane_margins_adj(state);
353 	if (ret)
354 		return ret;
355 
356 	vc4_state->x_scaling[0] = vc4_get_scaling_mode(vc4_state->src_w[0],
357 						       vc4_state->crtc_w);
358 	vc4_state->y_scaling[0] = vc4_get_scaling_mode(vc4_state->src_h[0],
359 						       vc4_state->crtc_h);
360 
361 	vc4_state->is_unity = (vc4_state->x_scaling[0] == VC4_SCALING_NONE &&
362 			       vc4_state->y_scaling[0] == VC4_SCALING_NONE);
363 
364 	if (num_planes > 1) {
365 		vc4_state->is_yuv = true;
366 
367 		vc4_state->src_w[1] = vc4_state->src_w[0] / h_subsample;
368 		vc4_state->src_h[1] = vc4_state->src_h[0] / v_subsample;
369 
370 		vc4_state->x_scaling[1] =
371 			vc4_get_scaling_mode(vc4_state->src_w[1],
372 					     vc4_state->crtc_w);
373 		vc4_state->y_scaling[1] =
374 			vc4_get_scaling_mode(vc4_state->src_h[1],
375 					     vc4_state->crtc_h);
376 
377 		/* YUV conversion requires that horizontal scaling be enabled
378 		 * on the UV plane even if vc4_get_scaling_mode() returned
379 		 * VC4_SCALING_NONE (which can happen when the down-scaling
380 		 * ratio is 0.5). Let's force it to VC4_SCALING_PPF in this
381 		 * case.
382 		 */
383 		if (vc4_state->x_scaling[1] == VC4_SCALING_NONE)
384 			vc4_state->x_scaling[1] = VC4_SCALING_PPF;
385 	} else {
386 		vc4_state->is_yuv = false;
387 		vc4_state->x_scaling[1] = VC4_SCALING_NONE;
388 		vc4_state->y_scaling[1] = VC4_SCALING_NONE;
389 	}
390 
391 	return 0;
392 }
393 
394 static void vc4_write_tpz(struct vc4_plane_state *vc4_state, u32 src, u32 dst)
395 {
396 	u32 scale, recip;
397 
398 	scale = (1 << 16) * src / dst;
399 
400 	/* The specs note that while the reciprocal would be defined
401 	 * as (1<<32)/scale, ~0 is close enough.
402 	 */
403 	recip = ~0 / scale;
404 
405 	vc4_dlist_write(vc4_state,
406 			VC4_SET_FIELD(scale, SCALER_TPZ0_SCALE) |
407 			VC4_SET_FIELD(0, SCALER_TPZ0_IPHASE));
408 	vc4_dlist_write(vc4_state,
409 			VC4_SET_FIELD(recip, SCALER_TPZ1_RECIP));
410 }
411 
412 static void vc4_write_ppf(struct vc4_plane_state *vc4_state, u32 src, u32 dst)
413 {
414 	u32 scale = (1 << 16) * src / dst;
415 
416 	vc4_dlist_write(vc4_state,
417 			SCALER_PPF_AGC |
418 			VC4_SET_FIELD(scale, SCALER_PPF_SCALE) |
419 			VC4_SET_FIELD(0, SCALER_PPF_IPHASE));
420 }
421 
422 static u32 vc4_lbm_size(struct drm_plane_state *state)
423 {
424 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
425 	/* This is the worst case number.  One of the two sizes will
426 	 * be used depending on the scaling configuration.
427 	 */
428 	u32 pix_per_line = max(vc4_state->src_w[0], (u32)vc4_state->crtc_w);
429 	u32 lbm;
430 
431 	/* LBM is not needed when there's no vertical scaling. */
432 	if (vc4_state->y_scaling[0] == VC4_SCALING_NONE &&
433 	    vc4_state->y_scaling[1] == VC4_SCALING_NONE)
434 		return 0;
435 
436 	if (!vc4_state->is_yuv) {
437 		if (vc4_state->y_scaling[0] == VC4_SCALING_TPZ)
438 			lbm = pix_per_line * 8;
439 		else {
440 			/* In special cases, this multiplier might be 12. */
441 			lbm = pix_per_line * 16;
442 		}
443 	} else {
444 		/* There are cases for this going down to a multiplier
445 		 * of 2, but according to the firmware source, the
446 		 * table in the docs is somewhat wrong.
447 		 */
448 		lbm = pix_per_line * 16;
449 	}
450 
451 	lbm = roundup(lbm, 32);
452 
453 	return lbm;
454 }
455 
456 static void vc4_write_scaling_parameters(struct drm_plane_state *state,
457 					 int channel)
458 {
459 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
460 
461 	/* Ch0 H-PPF Word 0: Scaling Parameters */
462 	if (vc4_state->x_scaling[channel] == VC4_SCALING_PPF) {
463 		vc4_write_ppf(vc4_state,
464 			      vc4_state->src_w[channel], vc4_state->crtc_w);
465 	}
466 
467 	/* Ch0 V-PPF Words 0-1: Scaling Parameters, Context */
468 	if (vc4_state->y_scaling[channel] == VC4_SCALING_PPF) {
469 		vc4_write_ppf(vc4_state,
470 			      vc4_state->src_h[channel], vc4_state->crtc_h);
471 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
472 	}
473 
474 	/* Ch0 H-TPZ Words 0-1: Scaling Parameters, Recip */
475 	if (vc4_state->x_scaling[channel] == VC4_SCALING_TPZ) {
476 		vc4_write_tpz(vc4_state,
477 			      vc4_state->src_w[channel], vc4_state->crtc_w);
478 	}
479 
480 	/* Ch0 V-TPZ Words 0-2: Scaling Parameters, Recip, Context */
481 	if (vc4_state->y_scaling[channel] == VC4_SCALING_TPZ) {
482 		vc4_write_tpz(vc4_state,
483 			      vc4_state->src_h[channel], vc4_state->crtc_h);
484 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
485 	}
486 }
487 
488 static void vc4_plane_calc_load(struct drm_plane_state *state)
489 {
490 	unsigned int hvs_load_shift, vrefresh, i;
491 	struct drm_framebuffer *fb = state->fb;
492 	struct vc4_plane_state *vc4_state;
493 	struct drm_crtc_state *crtc_state;
494 	unsigned int vscale_factor;
495 
496 	vc4_state = to_vc4_plane_state(state);
497 	crtc_state = drm_atomic_get_existing_crtc_state(state->state,
498 							state->crtc);
499 	vrefresh = drm_mode_vrefresh(&crtc_state->adjusted_mode);
500 
501 	/* The HVS is able to process 2 pixels/cycle when scaling the source,
502 	 * 4 pixels/cycle otherwise.
503 	 * Alpha blending step seems to be pipelined and it's always operating
504 	 * at 4 pixels/cycle, so the limiting aspect here seems to be the
505 	 * scaler block.
506 	 * HVS load is expressed in clk-cycles/sec (AKA Hz).
507 	 */
508 	if (vc4_state->x_scaling[0] != VC4_SCALING_NONE ||
509 	    vc4_state->x_scaling[1] != VC4_SCALING_NONE ||
510 	    vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
511 	    vc4_state->y_scaling[1] != VC4_SCALING_NONE)
512 		hvs_load_shift = 1;
513 	else
514 		hvs_load_shift = 2;
515 
516 	vc4_state->membus_load = 0;
517 	vc4_state->hvs_load = 0;
518 	for (i = 0; i < fb->format->num_planes; i++) {
519 		/* Even if the bandwidth/plane required for a single frame is
520 		 *
521 		 * vc4_state->src_w[i] * vc4_state->src_h[i] * cpp * vrefresh
522 		 *
523 		 * when downscaling, we have to read more pixels per line in
524 		 * the time frame reserved for a single line, so the bandwidth
525 		 * demand can be punctually higher. To account for that, we
526 		 * calculate the down-scaling factor and multiply the plane
527 		 * load by this number. We're likely over-estimating the read
528 		 * demand, but that's better than under-estimating it.
529 		 */
530 		vscale_factor = DIV_ROUND_UP(vc4_state->src_h[i],
531 					     vc4_state->crtc_h);
532 		vc4_state->membus_load += vc4_state->src_w[i] *
533 					  vc4_state->src_h[i] * vscale_factor *
534 					  fb->format->cpp[i];
535 		vc4_state->hvs_load += vc4_state->crtc_h * vc4_state->crtc_w;
536 	}
537 
538 	vc4_state->hvs_load *= vrefresh;
539 	vc4_state->hvs_load >>= hvs_load_shift;
540 	vc4_state->membus_load *= vrefresh;
541 }
542 
543 static int vc4_plane_allocate_lbm(struct drm_plane_state *state)
544 {
545 	struct vc4_dev *vc4 = to_vc4_dev(state->plane->dev);
546 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
547 	unsigned long irqflags;
548 	u32 lbm_size;
549 
550 	lbm_size = vc4_lbm_size(state);
551 	if (!lbm_size)
552 		return 0;
553 
554 	if (WARN_ON(!vc4_state->lbm_offset))
555 		return -EINVAL;
556 
557 	/* Allocate the LBM memory that the HVS will use for temporary
558 	 * storage due to our scaling/format conversion.
559 	 */
560 	if (!drm_mm_node_allocated(&vc4_state->lbm)) {
561 		int ret;
562 
563 		spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
564 		ret = drm_mm_insert_node_generic(&vc4->hvs->lbm_mm,
565 						 &vc4_state->lbm,
566 						 lbm_size, 32, 0, 0);
567 		spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
568 
569 		if (ret)
570 			return ret;
571 	} else {
572 		WARN_ON_ONCE(lbm_size != vc4_state->lbm.size);
573 	}
574 
575 	vc4_state->dlist[vc4_state->lbm_offset] = vc4_state->lbm.start;
576 
577 	return 0;
578 }
579 
580 /* Writes out a full display list for an active plane to the plane's
581  * private dlist state.
582  */
583 static int vc4_plane_mode_set(struct drm_plane *plane,
584 			      struct drm_plane_state *state)
585 {
586 	struct vc4_dev *vc4 = to_vc4_dev(plane->dev);
587 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
588 	struct drm_framebuffer *fb = state->fb;
589 	u32 ctl0_offset = vc4_state->dlist_count;
590 	const struct hvs_format *format = vc4_get_hvs_format(fb->format->format);
591 	u64 base_format_mod = fourcc_mod_broadcom_mod(fb->modifier);
592 	int num_planes = fb->format->num_planes;
593 	u32 h_subsample = fb->format->hsub;
594 	u32 v_subsample = fb->format->vsub;
595 	bool mix_plane_alpha;
596 	bool covers_screen;
597 	u32 scl0, scl1, pitch0;
598 	u32 tiling, src_y;
599 	u32 hvs_format = format->hvs;
600 	unsigned int rotation;
601 	int ret, i;
602 
603 	if (vc4_state->dlist_initialized)
604 		return 0;
605 
606 	ret = vc4_plane_setup_clipping_and_scaling(state);
607 	if (ret)
608 		return ret;
609 
610 	/* SCL1 is used for Cb/Cr scaling of planar formats.  For RGB
611 	 * and 4:4:4, scl1 should be set to scl0 so both channels of
612 	 * the scaler do the same thing.  For YUV, the Y plane needs
613 	 * to be put in channel 1 and Cb/Cr in channel 0, so we swap
614 	 * the scl fields here.
615 	 */
616 	if (num_planes == 1) {
617 		scl0 = vc4_get_scl_field(state, 0);
618 		scl1 = scl0;
619 	} else {
620 		scl0 = vc4_get_scl_field(state, 1);
621 		scl1 = vc4_get_scl_field(state, 0);
622 	}
623 
624 	rotation = drm_rotation_simplify(state->rotation,
625 					 DRM_MODE_ROTATE_0 |
626 					 DRM_MODE_REFLECT_X |
627 					 DRM_MODE_REFLECT_Y);
628 
629 	/* We must point to the last line when Y reflection is enabled. */
630 	src_y = vc4_state->src_y;
631 	if (rotation & DRM_MODE_REFLECT_Y)
632 		src_y += vc4_state->src_h[0] - 1;
633 
634 	switch (base_format_mod) {
635 	case DRM_FORMAT_MOD_LINEAR:
636 		tiling = SCALER_CTL0_TILING_LINEAR;
637 		pitch0 = VC4_SET_FIELD(fb->pitches[0], SCALER_SRC_PITCH);
638 
639 		/* Adjust the base pointer to the first pixel to be scanned
640 		 * out.
641 		 */
642 		for (i = 0; i < num_planes; i++) {
643 			vc4_state->offsets[i] += src_y /
644 						 (i ? v_subsample : 1) *
645 						 fb->pitches[i];
646 
647 			vc4_state->offsets[i] += vc4_state->src_x /
648 						 (i ? h_subsample : 1) *
649 						 fb->format->cpp[i];
650 		}
651 
652 		break;
653 
654 	case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED: {
655 		u32 tile_size_shift = 12; /* T tiles are 4kb */
656 		/* Whole-tile offsets, mostly for setting the pitch. */
657 		u32 tile_w_shift = fb->format->cpp[0] == 2 ? 6 : 5;
658 		u32 tile_h_shift = 5; /* 16 and 32bpp are 32 pixels high */
659 		u32 tile_w_mask = (1 << tile_w_shift) - 1;
660 		/* The height mask on 32-bit-per-pixel tiles is 63, i.e. twice
661 		 * the height (in pixels) of a 4k tile.
662 		 */
663 		u32 tile_h_mask = (2 << tile_h_shift) - 1;
664 		/* For T-tiled, the FB pitch is "how many bytes from one row to
665 		 * the next, such that
666 		 *
667 		 *	pitch * tile_h == tile_size * tiles_per_row
668 		 */
669 		u32 tiles_w = fb->pitches[0] >> (tile_size_shift - tile_h_shift);
670 		u32 tiles_l = vc4_state->src_x >> tile_w_shift;
671 		u32 tiles_r = tiles_w - tiles_l;
672 		u32 tiles_t = src_y >> tile_h_shift;
673 		/* Intra-tile offsets, which modify the base address (the
674 		 * SCALER_PITCH0_TILE_Y_OFFSET tells HVS how to walk from that
675 		 * base address).
676 		 */
677 		u32 tile_y = (src_y >> 4) & 1;
678 		u32 subtile_y = (src_y >> 2) & 3;
679 		u32 utile_y = src_y & 3;
680 		u32 x_off = vc4_state->src_x & tile_w_mask;
681 		u32 y_off = src_y & tile_h_mask;
682 
683 		/* When Y reflection is requested we must set the
684 		 * SCALER_PITCH0_TILE_LINE_DIR flag to tell HVS that all lines
685 		 * after the initial one should be fetched in descending order,
686 		 * which makes sense since we start from the last line and go
687 		 * backward.
688 		 * Don't know why we need y_off = max_y_off - y_off, but it's
689 		 * definitely required (I guess it's also related to the "going
690 		 * backward" situation).
691 		 */
692 		if (rotation & DRM_MODE_REFLECT_Y) {
693 			y_off = tile_h_mask - y_off;
694 			pitch0 = SCALER_PITCH0_TILE_LINE_DIR;
695 		} else {
696 			pitch0 = 0;
697 		}
698 
699 		tiling = SCALER_CTL0_TILING_256B_OR_T;
700 		pitch0 |= (VC4_SET_FIELD(x_off, SCALER_PITCH0_SINK_PIX) |
701 			   VC4_SET_FIELD(y_off, SCALER_PITCH0_TILE_Y_OFFSET) |
702 			   VC4_SET_FIELD(tiles_l, SCALER_PITCH0_TILE_WIDTH_L) |
703 			   VC4_SET_FIELD(tiles_r, SCALER_PITCH0_TILE_WIDTH_R));
704 		vc4_state->offsets[0] += tiles_t * (tiles_w << tile_size_shift);
705 		vc4_state->offsets[0] += subtile_y << 8;
706 		vc4_state->offsets[0] += utile_y << 4;
707 
708 		/* Rows of tiles alternate left-to-right and right-to-left. */
709 		if (tiles_t & 1) {
710 			pitch0 |= SCALER_PITCH0_TILE_INITIAL_LINE_DIR;
711 			vc4_state->offsets[0] += (tiles_w - tiles_l) <<
712 						 tile_size_shift;
713 			vc4_state->offsets[0] -= (1 + !tile_y) << 10;
714 		} else {
715 			vc4_state->offsets[0] += tiles_l << tile_size_shift;
716 			vc4_state->offsets[0] += tile_y << 10;
717 		}
718 
719 		break;
720 	}
721 
722 	case DRM_FORMAT_MOD_BROADCOM_SAND64:
723 	case DRM_FORMAT_MOD_BROADCOM_SAND128:
724 	case DRM_FORMAT_MOD_BROADCOM_SAND256: {
725 		uint32_t param = fourcc_mod_broadcom_param(fb->modifier);
726 		u32 tile_w, tile, x_off, pix_per_tile;
727 
728 		hvs_format = HVS_PIXEL_FORMAT_H264;
729 
730 		switch (base_format_mod) {
731 		case DRM_FORMAT_MOD_BROADCOM_SAND64:
732 			tiling = SCALER_CTL0_TILING_64B;
733 			tile_w = 64;
734 			break;
735 		case DRM_FORMAT_MOD_BROADCOM_SAND128:
736 			tiling = SCALER_CTL0_TILING_128B;
737 			tile_w = 128;
738 			break;
739 		case DRM_FORMAT_MOD_BROADCOM_SAND256:
740 			tiling = SCALER_CTL0_TILING_256B_OR_T;
741 			tile_w = 256;
742 			break;
743 		default:
744 			break;
745 		}
746 
747 		if (param > SCALER_TILE_HEIGHT_MASK) {
748 			DRM_DEBUG_KMS("SAND height too large (%d)\n", param);
749 			return -EINVAL;
750 		}
751 
752 		pix_per_tile = tile_w / fb->format->cpp[0];
753 		tile = vc4_state->src_x / pix_per_tile;
754 		x_off = vc4_state->src_x % pix_per_tile;
755 
756 		/* Adjust the base pointer to the first pixel to be scanned
757 		 * out.
758 		 */
759 		for (i = 0; i < num_planes; i++) {
760 			vc4_state->offsets[i] += param * tile_w * tile;
761 			vc4_state->offsets[i] += src_y /
762 						 (i ? v_subsample : 1) *
763 						 tile_w;
764 			vc4_state->offsets[i] += x_off /
765 						 (i ? h_subsample : 1) *
766 						 fb->format->cpp[i];
767 		}
768 
769 		pitch0 = VC4_SET_FIELD(param, SCALER_TILE_HEIGHT);
770 		break;
771 	}
772 
773 	default:
774 		DRM_DEBUG_KMS("Unsupported FB tiling flag 0x%16llx",
775 			      (long long)fb->modifier);
776 		return -EINVAL;
777 	}
778 
779 	/* Control word */
780 	vc4_dlist_write(vc4_state,
781 			SCALER_CTL0_VALID |
782 			(rotation & DRM_MODE_REFLECT_X ? SCALER_CTL0_HFLIP : 0) |
783 			(rotation & DRM_MODE_REFLECT_Y ? SCALER_CTL0_VFLIP : 0) |
784 			VC4_SET_FIELD(SCALER_CTL0_RGBA_EXPAND_ROUND, SCALER_CTL0_RGBA_EXPAND) |
785 			(format->pixel_order << SCALER_CTL0_ORDER_SHIFT) |
786 			(hvs_format << SCALER_CTL0_PIXEL_FORMAT_SHIFT) |
787 			VC4_SET_FIELD(tiling, SCALER_CTL0_TILING) |
788 			(vc4_state->is_unity ? SCALER_CTL0_UNITY : 0) |
789 			VC4_SET_FIELD(scl0, SCALER_CTL0_SCL0) |
790 			VC4_SET_FIELD(scl1, SCALER_CTL0_SCL1));
791 
792 	/* Position Word 0: Image Positions and Alpha Value */
793 	vc4_state->pos0_offset = vc4_state->dlist_count;
794 	vc4_dlist_write(vc4_state,
795 			VC4_SET_FIELD(state->alpha >> 8, SCALER_POS0_FIXED_ALPHA) |
796 			VC4_SET_FIELD(vc4_state->crtc_x, SCALER_POS0_START_X) |
797 			VC4_SET_FIELD(vc4_state->crtc_y, SCALER_POS0_START_Y));
798 
799 	/* Position Word 1: Scaled Image Dimensions. */
800 	if (!vc4_state->is_unity) {
801 		vc4_dlist_write(vc4_state,
802 				VC4_SET_FIELD(vc4_state->crtc_w,
803 					      SCALER_POS1_SCL_WIDTH) |
804 				VC4_SET_FIELD(vc4_state->crtc_h,
805 					      SCALER_POS1_SCL_HEIGHT));
806 	}
807 
808 	/* Don't waste cycles mixing with plane alpha if the set alpha
809 	 * is opaque or there is no per-pixel alpha information.
810 	 * In any case we use the alpha property value as the fixed alpha.
811 	 */
812 	mix_plane_alpha = state->alpha != DRM_BLEND_ALPHA_OPAQUE &&
813 			  fb->format->has_alpha;
814 
815 	/* Position Word 2: Source Image Size, Alpha */
816 	vc4_state->pos2_offset = vc4_state->dlist_count;
817 	vc4_dlist_write(vc4_state,
818 			VC4_SET_FIELD(fb->format->has_alpha ?
819 				      SCALER_POS2_ALPHA_MODE_PIPELINE :
820 				      SCALER_POS2_ALPHA_MODE_FIXED,
821 				      SCALER_POS2_ALPHA_MODE) |
822 			(mix_plane_alpha ? SCALER_POS2_ALPHA_MIX : 0) |
823 			(fb->format->has_alpha ? SCALER_POS2_ALPHA_PREMULT : 0) |
824 			VC4_SET_FIELD(vc4_state->src_w[0], SCALER_POS2_WIDTH) |
825 			VC4_SET_FIELD(vc4_state->src_h[0], SCALER_POS2_HEIGHT));
826 
827 	/* Position Word 3: Context.  Written by the HVS. */
828 	vc4_dlist_write(vc4_state, 0xc0c0c0c0);
829 
830 
831 	/* Pointer Word 0/1/2: RGB / Y / Cb / Cr Pointers
832 	 *
833 	 * The pointers may be any byte address.
834 	 */
835 	vc4_state->ptr0_offset = vc4_state->dlist_count;
836 	for (i = 0; i < num_planes; i++)
837 		vc4_dlist_write(vc4_state, vc4_state->offsets[i]);
838 
839 	/* Pointer Context Word 0/1/2: Written by the HVS */
840 	for (i = 0; i < num_planes; i++)
841 		vc4_dlist_write(vc4_state, 0xc0c0c0c0);
842 
843 	/* Pitch word 0 */
844 	vc4_dlist_write(vc4_state, pitch0);
845 
846 	/* Pitch word 1/2 */
847 	for (i = 1; i < num_planes; i++) {
848 		if (hvs_format != HVS_PIXEL_FORMAT_H264) {
849 			vc4_dlist_write(vc4_state,
850 					VC4_SET_FIELD(fb->pitches[i],
851 						      SCALER_SRC_PITCH));
852 		} else {
853 			vc4_dlist_write(vc4_state, pitch0);
854 		}
855 	}
856 
857 	/* Colorspace conversion words */
858 	if (vc4_state->is_yuv) {
859 		vc4_dlist_write(vc4_state, SCALER_CSC0_ITR_R_601_5);
860 		vc4_dlist_write(vc4_state, SCALER_CSC1_ITR_R_601_5);
861 		vc4_dlist_write(vc4_state, SCALER_CSC2_ITR_R_601_5);
862 	}
863 
864 	vc4_state->lbm_offset = 0;
865 
866 	if (vc4_state->x_scaling[0] != VC4_SCALING_NONE ||
867 	    vc4_state->x_scaling[1] != VC4_SCALING_NONE ||
868 	    vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
869 	    vc4_state->y_scaling[1] != VC4_SCALING_NONE) {
870 		/* Reserve a slot for the LBM Base Address. The real value will
871 		 * be set when calling vc4_plane_allocate_lbm().
872 		 */
873 		if (vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
874 		    vc4_state->y_scaling[1] != VC4_SCALING_NONE)
875 			vc4_state->lbm_offset = vc4_state->dlist_count++;
876 
877 		if (num_planes > 1) {
878 			/* Emit Cb/Cr as channel 0 and Y as channel
879 			 * 1. This matches how we set up scl0/scl1
880 			 * above.
881 			 */
882 			vc4_write_scaling_parameters(state, 1);
883 		}
884 		vc4_write_scaling_parameters(state, 0);
885 
886 		/* If any PPF setup was done, then all the kernel
887 		 * pointers get uploaded.
888 		 */
889 		if (vc4_state->x_scaling[0] == VC4_SCALING_PPF ||
890 		    vc4_state->y_scaling[0] == VC4_SCALING_PPF ||
891 		    vc4_state->x_scaling[1] == VC4_SCALING_PPF ||
892 		    vc4_state->y_scaling[1] == VC4_SCALING_PPF) {
893 			u32 kernel = VC4_SET_FIELD(vc4->hvs->mitchell_netravali_filter.start,
894 						   SCALER_PPF_KERNEL_OFFSET);
895 
896 			/* HPPF plane 0 */
897 			vc4_dlist_write(vc4_state, kernel);
898 			/* VPPF plane 0 */
899 			vc4_dlist_write(vc4_state, kernel);
900 			/* HPPF plane 1 */
901 			vc4_dlist_write(vc4_state, kernel);
902 			/* VPPF plane 1 */
903 			vc4_dlist_write(vc4_state, kernel);
904 		}
905 	}
906 
907 	vc4_state->dlist[ctl0_offset] |=
908 		VC4_SET_FIELD(vc4_state->dlist_count, SCALER_CTL0_SIZE);
909 
910 	/* crtc_* are already clipped coordinates. */
911 	covers_screen = vc4_state->crtc_x == 0 && vc4_state->crtc_y == 0 &&
912 			vc4_state->crtc_w == state->crtc->mode.hdisplay &&
913 			vc4_state->crtc_h == state->crtc->mode.vdisplay;
914 	/* Background fill might be necessary when the plane has per-pixel
915 	 * alpha content or a non-opaque plane alpha and could blend from the
916 	 * background or does not cover the entire screen.
917 	 */
918 	vc4_state->needs_bg_fill = fb->format->has_alpha || !covers_screen ||
919 				   state->alpha != DRM_BLEND_ALPHA_OPAQUE;
920 
921 	/* Flag the dlist as initialized to avoid checking it twice in case
922 	 * the async update check already called vc4_plane_mode_set() and
923 	 * decided to fallback to sync update because async update was not
924 	 * possible.
925 	 */
926 	vc4_state->dlist_initialized = 1;
927 
928 	vc4_plane_calc_load(state);
929 
930 	return 0;
931 }
932 
933 /* If a modeset involves changing the setup of a plane, the atomic
934  * infrastructure will call this to validate a proposed plane setup.
935  * However, if a plane isn't getting updated, this (and the
936  * corresponding vc4_plane_atomic_update) won't get called.  Thus, we
937  * compute the dlist here and have all active plane dlists get updated
938  * in the CRTC's flush.
939  */
940 static int vc4_plane_atomic_check(struct drm_plane *plane,
941 				  struct drm_plane_state *state)
942 {
943 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
944 	int ret;
945 
946 	vc4_state->dlist_count = 0;
947 
948 	if (!plane_enabled(state))
949 		return 0;
950 
951 	ret = vc4_plane_mode_set(plane, state);
952 	if (ret)
953 		return ret;
954 
955 	return vc4_plane_allocate_lbm(state);
956 }
957 
958 static void vc4_plane_atomic_update(struct drm_plane *plane,
959 				    struct drm_plane_state *old_state)
960 {
961 	/* No contents here.  Since we don't know where in the CRTC's
962 	 * dlist we should be stored, our dlist is uploaded to the
963 	 * hardware with vc4_plane_write_dlist() at CRTC atomic_flush
964 	 * time.
965 	 */
966 }
967 
968 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist)
969 {
970 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
971 	int i;
972 
973 	vc4_state->hw_dlist = dlist;
974 
975 	/* Can't memcpy_toio() because it needs to be 32-bit writes. */
976 	for (i = 0; i < vc4_state->dlist_count; i++)
977 		writel(vc4_state->dlist[i], &dlist[i]);
978 
979 	return vc4_state->dlist_count;
980 }
981 
982 u32 vc4_plane_dlist_size(const struct drm_plane_state *state)
983 {
984 	const struct vc4_plane_state *vc4_state =
985 		container_of(state, typeof(*vc4_state), base);
986 
987 	return vc4_state->dlist_count;
988 }
989 
990 /* Updates the plane to immediately (well, once the FIFO needs
991  * refilling) scan out from at a new framebuffer.
992  */
993 void vc4_plane_async_set_fb(struct drm_plane *plane, struct drm_framebuffer *fb)
994 {
995 	struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
996 	struct drm_gem_cma_object *bo = drm_fb_cma_get_gem_obj(fb, 0);
997 	uint32_t addr;
998 
999 	/* We're skipping the address adjustment for negative origin,
1000 	 * because this is only called on the primary plane.
1001 	 */
1002 	WARN_ON_ONCE(plane->state->crtc_x < 0 || plane->state->crtc_y < 0);
1003 	addr = bo->paddr + fb->offsets[0];
1004 
1005 	/* Write the new address into the hardware immediately.  The
1006 	 * scanout will start from this address as soon as the FIFO
1007 	 * needs to refill with pixels.
1008 	 */
1009 	writel(addr, &vc4_state->hw_dlist[vc4_state->ptr0_offset]);
1010 
1011 	/* Also update the CPU-side dlist copy, so that any later
1012 	 * atomic updates that don't do a new modeset on our plane
1013 	 * also use our updated address.
1014 	 */
1015 	vc4_state->dlist[vc4_state->ptr0_offset] = addr;
1016 }
1017 
1018 static void vc4_plane_atomic_async_update(struct drm_plane *plane,
1019 					  struct drm_plane_state *state)
1020 {
1021 	struct vc4_plane_state *vc4_state, *new_vc4_state;
1022 
1023 	swap(plane->state->fb, state->fb);
1024 	plane->state->crtc_x = state->crtc_x;
1025 	plane->state->crtc_y = state->crtc_y;
1026 	plane->state->crtc_w = state->crtc_w;
1027 	plane->state->crtc_h = state->crtc_h;
1028 	plane->state->src_x = state->src_x;
1029 	plane->state->src_y = state->src_y;
1030 	plane->state->src_w = state->src_w;
1031 	plane->state->src_h = state->src_h;
1032 	plane->state->src_h = state->src_h;
1033 	plane->state->alpha = state->alpha;
1034 	plane->state->pixel_blend_mode = state->pixel_blend_mode;
1035 	plane->state->rotation = state->rotation;
1036 	plane->state->zpos = state->zpos;
1037 	plane->state->normalized_zpos = state->normalized_zpos;
1038 	plane->state->color_encoding = state->color_encoding;
1039 	plane->state->color_range = state->color_range;
1040 	plane->state->src = state->src;
1041 	plane->state->dst = state->dst;
1042 	plane->state->visible = state->visible;
1043 
1044 	new_vc4_state = to_vc4_plane_state(state);
1045 	vc4_state = to_vc4_plane_state(plane->state);
1046 
1047 	vc4_state->crtc_x = new_vc4_state->crtc_x;
1048 	vc4_state->crtc_y = new_vc4_state->crtc_y;
1049 	vc4_state->crtc_h = new_vc4_state->crtc_h;
1050 	vc4_state->crtc_w = new_vc4_state->crtc_w;
1051 	vc4_state->src_x = new_vc4_state->src_x;
1052 	vc4_state->src_y = new_vc4_state->src_y;
1053 	memcpy(vc4_state->src_w, new_vc4_state->src_w,
1054 	       sizeof(vc4_state->src_w));
1055 	memcpy(vc4_state->src_h, new_vc4_state->src_h,
1056 	       sizeof(vc4_state->src_h));
1057 	memcpy(vc4_state->x_scaling, new_vc4_state->x_scaling,
1058 	       sizeof(vc4_state->x_scaling));
1059 	memcpy(vc4_state->y_scaling, new_vc4_state->y_scaling,
1060 	       sizeof(vc4_state->y_scaling));
1061 	vc4_state->is_unity = new_vc4_state->is_unity;
1062 	vc4_state->is_yuv = new_vc4_state->is_yuv;
1063 	memcpy(vc4_state->offsets, new_vc4_state->offsets,
1064 	       sizeof(vc4_state->offsets));
1065 	vc4_state->needs_bg_fill = new_vc4_state->needs_bg_fill;
1066 
1067 	/* Update the current vc4_state pos0, pos2 and ptr0 dlist entries. */
1068 	vc4_state->dlist[vc4_state->pos0_offset] =
1069 		new_vc4_state->dlist[vc4_state->pos0_offset];
1070 	vc4_state->dlist[vc4_state->pos2_offset] =
1071 		new_vc4_state->dlist[vc4_state->pos2_offset];
1072 	vc4_state->dlist[vc4_state->ptr0_offset] =
1073 		new_vc4_state->dlist[vc4_state->ptr0_offset];
1074 
1075 	/* Note that we can't just call vc4_plane_write_dlist()
1076 	 * because that would smash the context data that the HVS is
1077 	 * currently using.
1078 	 */
1079 	writel(vc4_state->dlist[vc4_state->pos0_offset],
1080 	       &vc4_state->hw_dlist[vc4_state->pos0_offset]);
1081 	writel(vc4_state->dlist[vc4_state->pos2_offset],
1082 	       &vc4_state->hw_dlist[vc4_state->pos2_offset]);
1083 	writel(vc4_state->dlist[vc4_state->ptr0_offset],
1084 	       &vc4_state->hw_dlist[vc4_state->ptr0_offset]);
1085 }
1086 
1087 static int vc4_plane_atomic_async_check(struct drm_plane *plane,
1088 					struct drm_plane_state *state)
1089 {
1090 	struct vc4_plane_state *old_vc4_state, *new_vc4_state;
1091 	int ret;
1092 	u32 i;
1093 
1094 	ret = vc4_plane_mode_set(plane, state);
1095 	if (ret)
1096 		return ret;
1097 
1098 	old_vc4_state = to_vc4_plane_state(plane->state);
1099 	new_vc4_state = to_vc4_plane_state(state);
1100 	if (old_vc4_state->dlist_count != new_vc4_state->dlist_count ||
1101 	    old_vc4_state->pos0_offset != new_vc4_state->pos0_offset ||
1102 	    old_vc4_state->pos2_offset != new_vc4_state->pos2_offset ||
1103 	    old_vc4_state->ptr0_offset != new_vc4_state->ptr0_offset ||
1104 	    vc4_lbm_size(plane->state) != vc4_lbm_size(state))
1105 		return -EINVAL;
1106 
1107 	/* Only pos0, pos2 and ptr0 DWORDS can be updated in an async update
1108 	 * if anything else has changed, fallback to a sync update.
1109 	 */
1110 	for (i = 0; i < new_vc4_state->dlist_count; i++) {
1111 		if (i == new_vc4_state->pos0_offset ||
1112 		    i == new_vc4_state->pos2_offset ||
1113 		    i == new_vc4_state->ptr0_offset ||
1114 		    (new_vc4_state->lbm_offset &&
1115 		     i == new_vc4_state->lbm_offset))
1116 			continue;
1117 
1118 		if (new_vc4_state->dlist[i] != old_vc4_state->dlist[i])
1119 			return -EINVAL;
1120 	}
1121 
1122 	return 0;
1123 }
1124 
1125 static int vc4_prepare_fb(struct drm_plane *plane,
1126 			  struct drm_plane_state *state)
1127 {
1128 	struct vc4_bo *bo;
1129 	int ret;
1130 
1131 	if (!state->fb)
1132 		return 0;
1133 
1134 	bo = to_vc4_bo(&drm_fb_cma_get_gem_obj(state->fb, 0)->base);
1135 
1136 	drm_gem_fb_prepare_fb(plane, state);
1137 
1138 	if (plane->state->fb == state->fb)
1139 		return 0;
1140 
1141 	ret = vc4_bo_inc_usecnt(bo);
1142 	if (ret)
1143 		return ret;
1144 
1145 	return 0;
1146 }
1147 
1148 static void vc4_cleanup_fb(struct drm_plane *plane,
1149 			   struct drm_plane_state *state)
1150 {
1151 	struct vc4_bo *bo;
1152 
1153 	if (plane->state->fb == state->fb || !state->fb)
1154 		return;
1155 
1156 	bo = to_vc4_bo(&drm_fb_cma_get_gem_obj(state->fb, 0)->base);
1157 	vc4_bo_dec_usecnt(bo);
1158 }
1159 
1160 static const struct drm_plane_helper_funcs vc4_plane_helper_funcs = {
1161 	.atomic_check = vc4_plane_atomic_check,
1162 	.atomic_update = vc4_plane_atomic_update,
1163 	.prepare_fb = vc4_prepare_fb,
1164 	.cleanup_fb = vc4_cleanup_fb,
1165 	.atomic_async_check = vc4_plane_atomic_async_check,
1166 	.atomic_async_update = vc4_plane_atomic_async_update,
1167 };
1168 
1169 static void vc4_plane_destroy(struct drm_plane *plane)
1170 {
1171 	drm_plane_cleanup(plane);
1172 }
1173 
1174 static bool vc4_format_mod_supported(struct drm_plane *plane,
1175 				     uint32_t format,
1176 				     uint64_t modifier)
1177 {
1178 	/* Support T_TILING for RGB formats only. */
1179 	switch (format) {
1180 	case DRM_FORMAT_XRGB8888:
1181 	case DRM_FORMAT_ARGB8888:
1182 	case DRM_FORMAT_ABGR8888:
1183 	case DRM_FORMAT_XBGR8888:
1184 	case DRM_FORMAT_RGB565:
1185 	case DRM_FORMAT_BGR565:
1186 	case DRM_FORMAT_ARGB1555:
1187 	case DRM_FORMAT_XRGB1555:
1188 		switch (fourcc_mod_broadcom_mod(modifier)) {
1189 		case DRM_FORMAT_MOD_LINEAR:
1190 		case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
1191 			return true;
1192 		default:
1193 			return false;
1194 		}
1195 	case DRM_FORMAT_NV12:
1196 	case DRM_FORMAT_NV21:
1197 		switch (fourcc_mod_broadcom_mod(modifier)) {
1198 		case DRM_FORMAT_MOD_LINEAR:
1199 		case DRM_FORMAT_MOD_BROADCOM_SAND64:
1200 		case DRM_FORMAT_MOD_BROADCOM_SAND128:
1201 		case DRM_FORMAT_MOD_BROADCOM_SAND256:
1202 			return true;
1203 		default:
1204 			return false;
1205 		}
1206 	case DRM_FORMAT_YUV422:
1207 	case DRM_FORMAT_YVU422:
1208 	case DRM_FORMAT_YUV420:
1209 	case DRM_FORMAT_YVU420:
1210 	case DRM_FORMAT_NV16:
1211 	case DRM_FORMAT_NV61:
1212 	default:
1213 		return (modifier == DRM_FORMAT_MOD_LINEAR);
1214 	}
1215 }
1216 
1217 static const struct drm_plane_funcs vc4_plane_funcs = {
1218 	.update_plane = drm_atomic_helper_update_plane,
1219 	.disable_plane = drm_atomic_helper_disable_plane,
1220 	.destroy = vc4_plane_destroy,
1221 	.set_property = NULL,
1222 	.reset = vc4_plane_reset,
1223 	.atomic_duplicate_state = vc4_plane_duplicate_state,
1224 	.atomic_destroy_state = vc4_plane_destroy_state,
1225 	.format_mod_supported = vc4_format_mod_supported,
1226 };
1227 
1228 struct drm_plane *vc4_plane_init(struct drm_device *dev,
1229 				 enum drm_plane_type type)
1230 {
1231 	struct drm_plane *plane = NULL;
1232 	struct vc4_plane *vc4_plane;
1233 	u32 formats[ARRAY_SIZE(hvs_formats)];
1234 	int ret = 0;
1235 	unsigned i;
1236 	static const uint64_t modifiers[] = {
1237 		DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED,
1238 		DRM_FORMAT_MOD_BROADCOM_SAND128,
1239 		DRM_FORMAT_MOD_BROADCOM_SAND64,
1240 		DRM_FORMAT_MOD_BROADCOM_SAND256,
1241 		DRM_FORMAT_MOD_LINEAR,
1242 		DRM_FORMAT_MOD_INVALID
1243 	};
1244 
1245 	vc4_plane = devm_kzalloc(dev->dev, sizeof(*vc4_plane),
1246 				 GFP_KERNEL);
1247 	if (!vc4_plane)
1248 		return ERR_PTR(-ENOMEM);
1249 
1250 	for (i = 0; i < ARRAY_SIZE(hvs_formats); i++)
1251 		formats[i] = hvs_formats[i].drm;
1252 
1253 	plane = &vc4_plane->base;
1254 	ret = drm_universal_plane_init(dev, plane, 0,
1255 				       &vc4_plane_funcs,
1256 				       formats, ARRAY_SIZE(formats),
1257 				       modifiers, type, NULL);
1258 
1259 	drm_plane_helper_add(plane, &vc4_plane_helper_funcs);
1260 
1261 	drm_plane_create_alpha_property(plane);
1262 	drm_plane_create_rotation_property(plane, DRM_MODE_ROTATE_0,
1263 					   DRM_MODE_ROTATE_0 |
1264 					   DRM_MODE_ROTATE_180 |
1265 					   DRM_MODE_REFLECT_X |
1266 					   DRM_MODE_REFLECT_Y);
1267 
1268 	return plane;
1269 }
1270 
1271 int vc4_plane_create_additional_planes(struct drm_device *drm)
1272 {
1273 	struct drm_plane *cursor_plane;
1274 	struct drm_crtc *crtc;
1275 	unsigned int i;
1276 
1277 	/* Set up some arbitrary number of planes.  We're not limited
1278 	 * by a set number of physical registers, just the space in
1279 	 * the HVS (16k) and how small an plane can be (28 bytes).
1280 	 * However, each plane we set up takes up some memory, and
1281 	 * increases the cost of looping over planes, which atomic
1282 	 * modesetting does quite a bit.  As a result, we pick a
1283 	 * modest number of planes to expose, that should hopefully
1284 	 * still cover any sane usecase.
1285 	 */
1286 	for (i = 0; i < 8; i++) {
1287 		struct drm_plane *plane =
1288 			vc4_plane_init(drm, DRM_PLANE_TYPE_OVERLAY);
1289 
1290 		if (IS_ERR(plane))
1291 			continue;
1292 
1293 		plane->possible_crtcs =
1294 			GENMASK(drm->mode_config.num_crtc - 1, 0);
1295 	}
1296 
1297 	drm_for_each_crtc(crtc, drm) {
1298 		/* Set up the legacy cursor after overlay initialization,
1299 		 * since we overlay planes on the CRTC in the order they were
1300 		 * initialized.
1301 		 */
1302 		cursor_plane = vc4_plane_init(drm, DRM_PLANE_TYPE_CURSOR);
1303 		if (!IS_ERR(cursor_plane)) {
1304 			cursor_plane->possible_crtcs = drm_crtc_mask(crtc);
1305 			crtc->cursor = cursor_plane;
1306 		}
1307 	}
1308 
1309 	return 0;
1310 }
1311