2 * Copyright (C) 2015 Broadcom
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License version 2 as
6 * published by the Free Software Foundation.
10 * DOC: VC4 plane module
12 * Each DRM plane is a layer of pixels being scanned out by the HVS.
14 * At atomic modeset check time, we compute the HVS display element
15 * state that would be necessary for displaying the plane (giving us a
16 * chance to figure out if a plane configuration is invalid), then at
17 * atomic flush time the CRTC will ask us to write our element state
18 * into the region of the HVS that it has allocated for us.
21 #include <drm/drm_atomic.h>
22 #include <drm/drm_atomic_helper.h>
23 #include <drm/drm_fb_cma_helper.h>
24 #include <drm/drm_plane_helper.h>
25 #include <drm/drm_atomic_uapi.h>
27 #include "uapi/drm/vc4_drm.h"
31 static const struct hvs_format {
32 u32 drm; /* DRM_FORMAT_* */
33 u32 hvs; /* HVS_FORMAT_* */
37 .drm = DRM_FORMAT_XRGB8888, .hvs = HVS_PIXEL_FORMAT_RGBA8888,
38 .pixel_order = HVS_PIXEL_ORDER_ABGR,
41 .drm = DRM_FORMAT_ARGB8888, .hvs = HVS_PIXEL_FORMAT_RGBA8888,
42 .pixel_order = HVS_PIXEL_ORDER_ABGR,
45 .drm = DRM_FORMAT_ABGR8888, .hvs = HVS_PIXEL_FORMAT_RGBA8888,
46 .pixel_order = HVS_PIXEL_ORDER_ARGB,
49 .drm = DRM_FORMAT_XBGR8888, .hvs = HVS_PIXEL_FORMAT_RGBA8888,
50 .pixel_order = HVS_PIXEL_ORDER_ARGB,
53 .drm = DRM_FORMAT_RGB565, .hvs = HVS_PIXEL_FORMAT_RGB565,
54 .pixel_order = HVS_PIXEL_ORDER_XRGB,
57 .drm = DRM_FORMAT_BGR565, .hvs = HVS_PIXEL_FORMAT_RGB565,
58 .pixel_order = HVS_PIXEL_ORDER_XBGR,
61 .drm = DRM_FORMAT_ARGB1555, .hvs = HVS_PIXEL_FORMAT_RGBA5551,
62 .pixel_order = HVS_PIXEL_ORDER_ABGR,
65 .drm = DRM_FORMAT_XRGB1555, .hvs = HVS_PIXEL_FORMAT_RGBA5551,
66 .pixel_order = HVS_PIXEL_ORDER_ABGR,
69 .drm = DRM_FORMAT_RGB888, .hvs = HVS_PIXEL_FORMAT_RGB888,
70 .pixel_order = HVS_PIXEL_ORDER_XRGB,
73 .drm = DRM_FORMAT_BGR888, .hvs = HVS_PIXEL_FORMAT_RGB888,
74 .pixel_order = HVS_PIXEL_ORDER_XBGR,
77 .drm = DRM_FORMAT_YUV422,
78 .hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_3PLANE,
79 .pixel_order = HVS_PIXEL_ORDER_XYCBCR,
82 .drm = DRM_FORMAT_YVU422,
83 .hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_3PLANE,
84 .pixel_order = HVS_PIXEL_ORDER_XYCRCB,
87 .drm = DRM_FORMAT_YUV420,
88 .hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_3PLANE,
89 .pixel_order = HVS_PIXEL_ORDER_XYCBCR,
92 .drm = DRM_FORMAT_YVU420,
93 .hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_3PLANE,
94 .pixel_order = HVS_PIXEL_ORDER_XYCRCB,
97 .drm = DRM_FORMAT_NV12,
98 .hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_2PLANE,
99 .pixel_order = HVS_PIXEL_ORDER_XYCBCR,
102 .drm = DRM_FORMAT_NV21,
103 .hvs = HVS_PIXEL_FORMAT_YCBCR_YUV420_2PLANE,
104 .pixel_order = HVS_PIXEL_ORDER_XYCRCB,
107 .drm = DRM_FORMAT_NV16,
108 .hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_2PLANE,
109 .pixel_order = HVS_PIXEL_ORDER_XYCBCR,
112 .drm = DRM_FORMAT_NV61,
113 .hvs = HVS_PIXEL_FORMAT_YCBCR_YUV422_2PLANE,
114 .pixel_order = HVS_PIXEL_ORDER_XYCRCB,
118 static const struct hvs_format *vc4_get_hvs_format(u32 drm_format)
122 for (i = 0; i < ARRAY_SIZE(hvs_formats); i++) {
123 if (hvs_formats[i].drm == drm_format)
124 return &hvs_formats[i];
130 static enum vc4_scaling_mode vc4_get_scaling_mode(u32 src, u32 dst)
133 return VC4_SCALING_NONE;
134 if (3 * dst >= 2 * src)
135 return VC4_SCALING_PPF;
137 return VC4_SCALING_TPZ;
140 static bool plane_enabled(struct drm_plane_state *state)
142 return state->fb && state->crtc;
145 static struct drm_plane_state *vc4_plane_duplicate_state(struct drm_plane *plane)
147 struct vc4_plane_state *vc4_state;
149 if (WARN_ON(!plane->state))
152 vc4_state = kmemdup(plane->state, sizeof(*vc4_state), GFP_KERNEL);
156 memset(&vc4_state->lbm, 0, sizeof(vc4_state->lbm));
157 vc4_state->dlist_initialized = 0;
159 __drm_atomic_helper_plane_duplicate_state(plane, &vc4_state->base);
161 if (vc4_state->dlist) {
162 vc4_state->dlist = kmemdup(vc4_state->dlist,
163 vc4_state->dlist_count * 4,
165 if (!vc4_state->dlist) {
169 vc4_state->dlist_size = vc4_state->dlist_count;
172 return &vc4_state->base;
175 static void vc4_plane_destroy_state(struct drm_plane *plane,
176 struct drm_plane_state *state)
178 struct vc4_dev *vc4 = to_vc4_dev(plane->dev);
179 struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
181 if (vc4_state->lbm.allocated) {
182 unsigned long irqflags;
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);
189 kfree(vc4_state->dlist);
190 __drm_atomic_helper_plane_destroy_state(&vc4_state->base);
194 /* Called during init to allocate the plane's atomic state. */
195 static void vc4_plane_reset(struct drm_plane *plane)
197 struct vc4_plane_state *vc4_state;
199 WARN_ON(plane->state);
201 vc4_state = kzalloc(sizeof(*vc4_state), GFP_KERNEL);
205 __drm_atomic_helper_plane_reset(plane, &vc4_state->base);
208 static void vc4_dlist_write(struct vc4_plane_state *vc4_state, u32 val)
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);
216 memcpy(new_dlist, vc4_state->dlist, vc4_state->dlist_count * 4);
218 kfree(vc4_state->dlist);
219 vc4_state->dlist = new_dlist;
220 vc4_state->dlist_size = new_size;
223 vc4_state->dlist[vc4_state->dlist_count++] = val;
226 /* Returns the scl0/scl1 field based on whether the dimensions need to
227 * be up/down/non-scaled.
229 * This is a replication of a table from the spec.
231 static u32 vc4_get_scl_field(struct drm_plane_state *state, int plane)
233 struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
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;
253 case VC4_SCALING_NONE << 2 | VC4_SCALING_NONE:
254 /* The unity case is independently handled by
261 static int vc4_plane_margins_adj(struct drm_plane_state *pstate)
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;
267 crtc_state = drm_atomic_get_new_crtc_state(pstate->state,
270 vc4_crtc_get_margins(crtc_state, &left, &right, &top, &bottom);
271 if (!left && !right && !top && !bottom)
274 if (left + right >= crtc_state->mode.hdisplay ||
275 top + bottom >= crtc_state->mode.vdisplay)
278 adjhdisplay = crtc_state->mode.hdisplay - (left + right);
279 vc4_pstate->crtc_x = DIV_ROUND_CLOSEST(vc4_pstate->crtc_x *
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;
286 adjvdisplay = crtc_state->mode.vdisplay - (top + bottom);
287 vc4_pstate->crtc_y = DIV_ROUND_CLOSEST(vc4_pstate->crtc_y *
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;
294 vc4_pstate->crtc_w = DIV_ROUND_CLOSEST(vc4_pstate->crtc_w *
296 crtc_state->mode.hdisplay);
297 vc4_pstate->crtc_h = DIV_ROUND_CLOSEST(vc4_pstate->crtc_h *
299 crtc_state->mode.vdisplay);
301 if (!vc4_pstate->crtc_w || !vc4_pstate->crtc_h)
307 static int vc4_plane_setup_clipping_and_scaling(struct drm_plane_state *state)
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 u32 format = fb->format->format;
314 int num_planes = fb->format->num_planes;
315 struct drm_crtc_state *crtc_state;
316 u32 h_subsample, v_subsample;
319 crtc_state = drm_atomic_get_existing_crtc_state(state->state,
322 DRM_DEBUG_KMS("Invalid crtc state\n");
326 ret = drm_atomic_helper_check_plane_state(state, crtc_state, 1,
327 INT_MAX, true, true);
331 h_subsample = drm_format_horz_chroma_subsampling(format);
332 v_subsample = drm_format_vert_chroma_subsampling(format);
334 for (i = 0; i < num_planes; i++)
335 vc4_state->offsets[i] = bo->paddr + fb->offsets[i];
337 /* We don't support subpixel source positioning for scaling. */
338 if ((state->src.x1 & subpixel_src_mask) ||
339 (state->src.x2 & subpixel_src_mask) ||
340 (state->src.y1 & subpixel_src_mask) ||
341 (state->src.y2 & subpixel_src_mask)) {
345 vc4_state->src_x = state->src.x1 >> 16;
346 vc4_state->src_y = state->src.y1 >> 16;
347 vc4_state->src_w[0] = (state->src.x2 - state->src.x1) >> 16;
348 vc4_state->src_h[0] = (state->src.y2 - state->src.y1) >> 16;
350 vc4_state->crtc_x = state->dst.x1;
351 vc4_state->crtc_y = state->dst.y1;
352 vc4_state->crtc_w = state->dst.x2 - state->dst.x1;
353 vc4_state->crtc_h = state->dst.y2 - state->dst.y1;
355 ret = vc4_plane_margins_adj(state);
359 vc4_state->x_scaling[0] = vc4_get_scaling_mode(vc4_state->src_w[0],
361 vc4_state->y_scaling[0] = vc4_get_scaling_mode(vc4_state->src_h[0],
364 vc4_state->is_unity = (vc4_state->x_scaling[0] == VC4_SCALING_NONE &&
365 vc4_state->y_scaling[0] == VC4_SCALING_NONE);
367 if (num_planes > 1) {
368 vc4_state->is_yuv = true;
370 vc4_state->src_w[1] = vc4_state->src_w[0] / h_subsample;
371 vc4_state->src_h[1] = vc4_state->src_h[0] / v_subsample;
373 vc4_state->x_scaling[1] =
374 vc4_get_scaling_mode(vc4_state->src_w[1],
376 vc4_state->y_scaling[1] =
377 vc4_get_scaling_mode(vc4_state->src_h[1],
380 /* YUV conversion requires that horizontal scaling be enabled
381 * on the UV plane even if vc4_get_scaling_mode() returned
382 * VC4_SCALING_NONE (which can happen when the down-scaling
383 * ratio is 0.5). Let's force it to VC4_SCALING_PPF in this
386 if (vc4_state->x_scaling[1] == VC4_SCALING_NONE)
387 vc4_state->x_scaling[1] = VC4_SCALING_PPF;
389 vc4_state->is_yuv = false;
390 vc4_state->x_scaling[1] = VC4_SCALING_NONE;
391 vc4_state->y_scaling[1] = VC4_SCALING_NONE;
397 static void vc4_write_tpz(struct vc4_plane_state *vc4_state, u32 src, u32 dst)
401 scale = (1 << 16) * src / dst;
403 /* The specs note that while the reciprocal would be defined
404 * as (1<<32)/scale, ~0 is close enough.
408 vc4_dlist_write(vc4_state,
409 VC4_SET_FIELD(scale, SCALER_TPZ0_SCALE) |
410 VC4_SET_FIELD(0, SCALER_TPZ0_IPHASE));
411 vc4_dlist_write(vc4_state,
412 VC4_SET_FIELD(recip, SCALER_TPZ1_RECIP));
415 static void vc4_write_ppf(struct vc4_plane_state *vc4_state, u32 src, u32 dst)
417 u32 scale = (1 << 16) * src / dst;
419 vc4_dlist_write(vc4_state,
421 VC4_SET_FIELD(scale, SCALER_PPF_SCALE) |
422 VC4_SET_FIELD(0, SCALER_PPF_IPHASE));
425 static u32 vc4_lbm_size(struct drm_plane_state *state)
427 struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
428 /* This is the worst case number. One of the two sizes will
429 * be used depending on the scaling configuration.
431 u32 pix_per_line = max(vc4_state->src_w[0], (u32)vc4_state->crtc_w);
434 /* LBM is not needed when there's no vertical scaling. */
435 if (vc4_state->y_scaling[0] == VC4_SCALING_NONE &&
436 vc4_state->y_scaling[1] == VC4_SCALING_NONE)
439 if (!vc4_state->is_yuv) {
440 if (vc4_state->y_scaling[0] == VC4_SCALING_TPZ)
441 lbm = pix_per_line * 8;
443 /* In special cases, this multiplier might be 12. */
444 lbm = pix_per_line * 16;
447 /* There are cases for this going down to a multiplier
448 * of 2, but according to the firmware source, the
449 * table in the docs is somewhat wrong.
451 lbm = pix_per_line * 16;
454 lbm = roundup(lbm, 32);
459 static void vc4_write_scaling_parameters(struct drm_plane_state *state,
462 struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
464 /* Ch0 H-PPF Word 0: Scaling Parameters */
465 if (vc4_state->x_scaling[channel] == VC4_SCALING_PPF) {
466 vc4_write_ppf(vc4_state,
467 vc4_state->src_w[channel], vc4_state->crtc_w);
470 /* Ch0 V-PPF Words 0-1: Scaling Parameters, Context */
471 if (vc4_state->y_scaling[channel] == VC4_SCALING_PPF) {
472 vc4_write_ppf(vc4_state,
473 vc4_state->src_h[channel], vc4_state->crtc_h);
474 vc4_dlist_write(vc4_state, 0xc0c0c0c0);
477 /* Ch0 H-TPZ Words 0-1: Scaling Parameters, Recip */
478 if (vc4_state->x_scaling[channel] == VC4_SCALING_TPZ) {
479 vc4_write_tpz(vc4_state,
480 vc4_state->src_w[channel], vc4_state->crtc_w);
483 /* Ch0 V-TPZ Words 0-2: Scaling Parameters, Recip, Context */
484 if (vc4_state->y_scaling[channel] == VC4_SCALING_TPZ) {
485 vc4_write_tpz(vc4_state,
486 vc4_state->src_h[channel], vc4_state->crtc_h);
487 vc4_dlist_write(vc4_state, 0xc0c0c0c0);
491 static void vc4_plane_calc_load(struct drm_plane_state *state)
493 unsigned int hvs_load_shift, vrefresh, i;
494 struct drm_framebuffer *fb = state->fb;
495 struct vc4_plane_state *vc4_state;
496 struct drm_crtc_state *crtc_state;
497 unsigned int vscale_factor;
499 vc4_state = to_vc4_plane_state(state);
500 crtc_state = drm_atomic_get_existing_crtc_state(state->state,
502 vrefresh = drm_mode_vrefresh(&crtc_state->adjusted_mode);
504 /* The HVS is able to process 2 pixels/cycle when scaling the source,
505 * 4 pixels/cycle otherwise.
506 * Alpha blending step seems to be pipelined and it's always operating
507 * at 4 pixels/cycle, so the limiting aspect here seems to be the
509 * HVS load is expressed in clk-cycles/sec (AKA Hz).
511 if (vc4_state->x_scaling[0] != VC4_SCALING_NONE ||
512 vc4_state->x_scaling[1] != VC4_SCALING_NONE ||
513 vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
514 vc4_state->y_scaling[1] != VC4_SCALING_NONE)
519 vc4_state->membus_load = 0;
520 vc4_state->hvs_load = 0;
521 for (i = 0; i < fb->format->num_planes; i++) {
522 /* Even if the bandwidth/plane required for a single frame is
524 * vc4_state->src_w[i] * vc4_state->src_h[i] * cpp * vrefresh
526 * when downscaling, we have to read more pixels per line in
527 * the time frame reserved for a single line, so the bandwidth
528 * demand can be punctually higher. To account for that, we
529 * calculate the down-scaling factor and multiply the plane
530 * load by this number. We're likely over-estimating the read
531 * demand, but that's better than under-estimating it.
533 vscale_factor = DIV_ROUND_UP(vc4_state->src_h[i],
535 vc4_state->membus_load += vc4_state->src_w[i] *
536 vc4_state->src_h[i] * vscale_factor *
538 vc4_state->hvs_load += vc4_state->crtc_h * vc4_state->crtc_w;
541 vc4_state->hvs_load *= vrefresh;
542 vc4_state->hvs_load >>= hvs_load_shift;
543 vc4_state->membus_load *= vrefresh;
546 static int vc4_plane_allocate_lbm(struct drm_plane_state *state)
548 struct vc4_dev *vc4 = to_vc4_dev(state->plane->dev);
549 struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
550 unsigned long irqflags;
553 lbm_size = vc4_lbm_size(state);
557 if (WARN_ON(!vc4_state->lbm_offset))
560 /* Allocate the LBM memory that the HVS will use for temporary
561 * storage due to our scaling/format conversion.
563 if (!vc4_state->lbm.allocated) {
566 spin_lock_irqsave(&vc4->hvs->mm_lock, irqflags);
567 ret = drm_mm_insert_node_generic(&vc4->hvs->lbm_mm,
570 spin_unlock_irqrestore(&vc4->hvs->mm_lock, irqflags);
575 WARN_ON_ONCE(lbm_size != vc4_state->lbm.size);
578 vc4_state->dlist[vc4_state->lbm_offset] = vc4_state->lbm.start;
583 /* Writes out a full display list for an active plane to the plane's
584 * private dlist state.
586 static int vc4_plane_mode_set(struct drm_plane *plane,
587 struct drm_plane_state *state)
589 struct vc4_dev *vc4 = to_vc4_dev(plane->dev);
590 struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
591 struct drm_framebuffer *fb = state->fb;
592 u32 ctl0_offset = vc4_state->dlist_count;
593 const struct hvs_format *format = vc4_get_hvs_format(fb->format->format);
594 u64 base_format_mod = fourcc_mod_broadcom_mod(fb->modifier);
595 int num_planes = fb->format->num_planes;
596 u32 h_subsample, v_subsample;
597 bool mix_plane_alpha;
599 u32 scl0, scl1, pitch0;
601 u32 hvs_format = format->hvs;
602 unsigned int rotation;
605 if (vc4_state->dlist_initialized)
608 ret = vc4_plane_setup_clipping_and_scaling(state);
612 /* SCL1 is used for Cb/Cr scaling of planar formats. For RGB
613 * and 4:4:4, scl1 should be set to scl0 so both channels of
614 * the scaler do the same thing. For YUV, the Y plane needs
615 * to be put in channel 1 and Cb/Cr in channel 0, so we swap
616 * the scl fields here.
618 if (num_planes == 1) {
619 scl0 = vc4_get_scl_field(state, 0);
622 scl0 = vc4_get_scl_field(state, 1);
623 scl1 = vc4_get_scl_field(state, 0);
626 h_subsample = drm_format_horz_chroma_subsampling(format->drm);
627 v_subsample = drm_format_vert_chroma_subsampling(format->drm);
629 rotation = drm_rotation_simplify(state->rotation,
634 /* We must point to the last line when Y reflection is enabled. */
635 src_y = vc4_state->src_y;
636 if (rotation & DRM_MODE_REFLECT_Y)
637 src_y += vc4_state->src_h[0] - 1;
639 switch (base_format_mod) {
640 case DRM_FORMAT_MOD_LINEAR:
641 tiling = SCALER_CTL0_TILING_LINEAR;
642 pitch0 = VC4_SET_FIELD(fb->pitches[0], SCALER_SRC_PITCH);
644 /* Adjust the base pointer to the first pixel to be scanned
647 for (i = 0; i < num_planes; i++) {
648 vc4_state->offsets[i] += src_y /
649 (i ? v_subsample : 1) *
652 vc4_state->offsets[i] += vc4_state->src_x /
653 (i ? h_subsample : 1) *
659 case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED: {
660 u32 tile_size_shift = 12; /* T tiles are 4kb */
661 /* Whole-tile offsets, mostly for setting the pitch. */
662 u32 tile_w_shift = fb->format->cpp[0] == 2 ? 6 : 5;
663 u32 tile_h_shift = 5; /* 16 and 32bpp are 32 pixels high */
664 u32 tile_w_mask = (1 << tile_w_shift) - 1;
665 /* The height mask on 32-bit-per-pixel tiles is 63, i.e. twice
666 * the height (in pixels) of a 4k tile.
668 u32 tile_h_mask = (2 << tile_h_shift) - 1;
669 /* For T-tiled, the FB pitch is "how many bytes from one row to
670 * the next, such that
672 * pitch * tile_h == tile_size * tiles_per_row
674 u32 tiles_w = fb->pitches[0] >> (tile_size_shift - tile_h_shift);
675 u32 tiles_l = vc4_state->src_x >> tile_w_shift;
676 u32 tiles_r = tiles_w - tiles_l;
677 u32 tiles_t = src_y >> tile_h_shift;
678 /* Intra-tile offsets, which modify the base address (the
679 * SCALER_PITCH0_TILE_Y_OFFSET tells HVS how to walk from that
682 u32 tile_y = (src_y >> 4) & 1;
683 u32 subtile_y = (src_y >> 2) & 3;
684 u32 utile_y = src_y & 3;
685 u32 x_off = vc4_state->src_x & tile_w_mask;
686 u32 y_off = src_y & tile_h_mask;
688 /* When Y reflection is requested we must set the
689 * SCALER_PITCH0_TILE_LINE_DIR flag to tell HVS that all lines
690 * after the initial one should be fetched in descending order,
691 * which makes sense since we start from the last line and go
693 * Don't know why we need y_off = max_y_off - y_off, but it's
694 * definitely required (I guess it's also related to the "going
695 * backward" situation).
697 if (rotation & DRM_MODE_REFLECT_Y) {
698 y_off = tile_h_mask - y_off;
699 pitch0 = SCALER_PITCH0_TILE_LINE_DIR;
704 tiling = SCALER_CTL0_TILING_256B_OR_T;
705 pitch0 |= (VC4_SET_FIELD(x_off, SCALER_PITCH0_SINK_PIX) |
706 VC4_SET_FIELD(y_off, SCALER_PITCH0_TILE_Y_OFFSET) |
707 VC4_SET_FIELD(tiles_l, SCALER_PITCH0_TILE_WIDTH_L) |
708 VC4_SET_FIELD(tiles_r, SCALER_PITCH0_TILE_WIDTH_R));
709 vc4_state->offsets[0] += tiles_t * (tiles_w << tile_size_shift);
710 vc4_state->offsets[0] += subtile_y << 8;
711 vc4_state->offsets[0] += utile_y << 4;
713 /* Rows of tiles alternate left-to-right and right-to-left. */
715 pitch0 |= SCALER_PITCH0_TILE_INITIAL_LINE_DIR;
716 vc4_state->offsets[0] += (tiles_w - tiles_l) <<
718 vc4_state->offsets[0] -= (1 + !tile_y) << 10;
720 vc4_state->offsets[0] += tiles_l << tile_size_shift;
721 vc4_state->offsets[0] += tile_y << 10;
727 case DRM_FORMAT_MOD_BROADCOM_SAND64:
728 case DRM_FORMAT_MOD_BROADCOM_SAND128:
729 case DRM_FORMAT_MOD_BROADCOM_SAND256: {
730 uint32_t param = fourcc_mod_broadcom_param(fb->modifier);
731 u32 tile_w, tile, x_off, pix_per_tile;
733 hvs_format = HVS_PIXEL_FORMAT_H264;
735 switch (base_format_mod) {
736 case DRM_FORMAT_MOD_BROADCOM_SAND64:
737 tiling = SCALER_CTL0_TILING_64B;
740 case DRM_FORMAT_MOD_BROADCOM_SAND128:
741 tiling = SCALER_CTL0_TILING_128B;
744 case DRM_FORMAT_MOD_BROADCOM_SAND256:
745 tiling = SCALER_CTL0_TILING_256B_OR_T;
752 if (param > SCALER_TILE_HEIGHT_MASK) {
753 DRM_DEBUG_KMS("SAND height too large (%d)\n", param);
757 pix_per_tile = tile_w / fb->format->cpp[0];
758 tile = vc4_state->src_x / pix_per_tile;
759 x_off = vc4_state->src_x % pix_per_tile;
761 /* Adjust the base pointer to the first pixel to be scanned
764 for (i = 0; i < num_planes; i++) {
765 vc4_state->offsets[i] += param * tile_w * tile;
766 vc4_state->offsets[i] += src_y /
767 (i ? v_subsample : 1) *
769 vc4_state->offsets[i] += x_off /
770 (i ? h_subsample : 1) *
774 pitch0 = VC4_SET_FIELD(param, SCALER_TILE_HEIGHT);
779 DRM_DEBUG_KMS("Unsupported FB tiling flag 0x%16llx",
780 (long long)fb->modifier);
785 vc4_dlist_write(vc4_state,
787 (rotation & DRM_MODE_REFLECT_X ? SCALER_CTL0_HFLIP : 0) |
788 (rotation & DRM_MODE_REFLECT_Y ? SCALER_CTL0_VFLIP : 0) |
789 VC4_SET_FIELD(SCALER_CTL0_RGBA_EXPAND_ROUND, SCALER_CTL0_RGBA_EXPAND) |
790 (format->pixel_order << SCALER_CTL0_ORDER_SHIFT) |
791 (hvs_format << SCALER_CTL0_PIXEL_FORMAT_SHIFT) |
792 VC4_SET_FIELD(tiling, SCALER_CTL0_TILING) |
793 (vc4_state->is_unity ? SCALER_CTL0_UNITY : 0) |
794 VC4_SET_FIELD(scl0, SCALER_CTL0_SCL0) |
795 VC4_SET_FIELD(scl1, SCALER_CTL0_SCL1));
797 /* Position Word 0: Image Positions and Alpha Value */
798 vc4_state->pos0_offset = vc4_state->dlist_count;
799 vc4_dlist_write(vc4_state,
800 VC4_SET_FIELD(state->alpha >> 8, SCALER_POS0_FIXED_ALPHA) |
801 VC4_SET_FIELD(vc4_state->crtc_x, SCALER_POS0_START_X) |
802 VC4_SET_FIELD(vc4_state->crtc_y, SCALER_POS0_START_Y));
804 /* Position Word 1: Scaled Image Dimensions. */
805 if (!vc4_state->is_unity) {
806 vc4_dlist_write(vc4_state,
807 VC4_SET_FIELD(vc4_state->crtc_w,
808 SCALER_POS1_SCL_WIDTH) |
809 VC4_SET_FIELD(vc4_state->crtc_h,
810 SCALER_POS1_SCL_HEIGHT));
813 /* Don't waste cycles mixing with plane alpha if the set alpha
814 * is opaque or there is no per-pixel alpha information.
815 * In any case we use the alpha property value as the fixed alpha.
817 mix_plane_alpha = state->alpha != DRM_BLEND_ALPHA_OPAQUE &&
818 fb->format->has_alpha;
820 /* Position Word 2: Source Image Size, Alpha */
821 vc4_state->pos2_offset = vc4_state->dlist_count;
822 vc4_dlist_write(vc4_state,
823 VC4_SET_FIELD(fb->format->has_alpha ?
824 SCALER_POS2_ALPHA_MODE_PIPELINE :
825 SCALER_POS2_ALPHA_MODE_FIXED,
826 SCALER_POS2_ALPHA_MODE) |
827 (mix_plane_alpha ? SCALER_POS2_ALPHA_MIX : 0) |
828 (fb->format->has_alpha ? SCALER_POS2_ALPHA_PREMULT : 0) |
829 VC4_SET_FIELD(vc4_state->src_w[0], SCALER_POS2_WIDTH) |
830 VC4_SET_FIELD(vc4_state->src_h[0], SCALER_POS2_HEIGHT));
832 /* Position Word 3: Context. Written by the HVS. */
833 vc4_dlist_write(vc4_state, 0xc0c0c0c0);
836 /* Pointer Word 0/1/2: RGB / Y / Cb / Cr Pointers
838 * The pointers may be any byte address.
840 vc4_state->ptr0_offset = vc4_state->dlist_count;
841 for (i = 0; i < num_planes; i++)
842 vc4_dlist_write(vc4_state, vc4_state->offsets[i]);
844 /* Pointer Context Word 0/1/2: Written by the HVS */
845 for (i = 0; i < num_planes; i++)
846 vc4_dlist_write(vc4_state, 0xc0c0c0c0);
849 vc4_dlist_write(vc4_state, pitch0);
852 for (i = 1; i < num_planes; i++) {
853 if (hvs_format != HVS_PIXEL_FORMAT_H264) {
854 vc4_dlist_write(vc4_state,
855 VC4_SET_FIELD(fb->pitches[i],
858 vc4_dlist_write(vc4_state, pitch0);
862 /* Colorspace conversion words */
863 if (vc4_state->is_yuv) {
864 vc4_dlist_write(vc4_state, SCALER_CSC0_ITR_R_601_5);
865 vc4_dlist_write(vc4_state, SCALER_CSC1_ITR_R_601_5);
866 vc4_dlist_write(vc4_state, SCALER_CSC2_ITR_R_601_5);
869 vc4_state->lbm_offset = 0;
871 if (vc4_state->x_scaling[0] != VC4_SCALING_NONE ||
872 vc4_state->x_scaling[1] != VC4_SCALING_NONE ||
873 vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
874 vc4_state->y_scaling[1] != VC4_SCALING_NONE) {
875 /* Reserve a slot for the LBM Base Address. The real value will
876 * be set when calling vc4_plane_allocate_lbm().
878 if (vc4_state->y_scaling[0] != VC4_SCALING_NONE ||
879 vc4_state->y_scaling[1] != VC4_SCALING_NONE)
880 vc4_state->lbm_offset = vc4_state->dlist_count++;
882 if (num_planes > 1) {
883 /* Emit Cb/Cr as channel 0 and Y as channel
884 * 1. This matches how we set up scl0/scl1
887 vc4_write_scaling_parameters(state, 1);
889 vc4_write_scaling_parameters(state, 0);
891 /* If any PPF setup was done, then all the kernel
892 * pointers get uploaded.
894 if (vc4_state->x_scaling[0] == VC4_SCALING_PPF ||
895 vc4_state->y_scaling[0] == VC4_SCALING_PPF ||
896 vc4_state->x_scaling[1] == VC4_SCALING_PPF ||
897 vc4_state->y_scaling[1] == VC4_SCALING_PPF) {
898 u32 kernel = VC4_SET_FIELD(vc4->hvs->mitchell_netravali_filter.start,
899 SCALER_PPF_KERNEL_OFFSET);
902 vc4_dlist_write(vc4_state, kernel);
904 vc4_dlist_write(vc4_state, kernel);
906 vc4_dlist_write(vc4_state, kernel);
908 vc4_dlist_write(vc4_state, kernel);
912 vc4_state->dlist[ctl0_offset] |=
913 VC4_SET_FIELD(vc4_state->dlist_count, SCALER_CTL0_SIZE);
915 /* crtc_* are already clipped coordinates. */
916 covers_screen = vc4_state->crtc_x == 0 && vc4_state->crtc_y == 0 &&
917 vc4_state->crtc_w == state->crtc->mode.hdisplay &&
918 vc4_state->crtc_h == state->crtc->mode.vdisplay;
919 /* Background fill might be necessary when the plane has per-pixel
920 * alpha content or a non-opaque plane alpha and could blend from the
921 * background or does not cover the entire screen.
923 vc4_state->needs_bg_fill = fb->format->has_alpha || !covers_screen ||
924 state->alpha != DRM_BLEND_ALPHA_OPAQUE;
926 /* Flag the dlist as initialized to avoid checking it twice in case
927 * the async update check already called vc4_plane_mode_set() and
928 * decided to fallback to sync update because async update was not
931 vc4_state->dlist_initialized = 1;
933 vc4_plane_calc_load(state);
938 /* If a modeset involves changing the setup of a plane, the atomic
939 * infrastructure will call this to validate a proposed plane setup.
940 * However, if a plane isn't getting updated, this (and the
941 * corresponding vc4_plane_atomic_update) won't get called. Thus, we
942 * compute the dlist here and have all active plane dlists get updated
943 * in the CRTC's flush.
945 static int vc4_plane_atomic_check(struct drm_plane *plane,
946 struct drm_plane_state *state)
948 struct vc4_plane_state *vc4_state = to_vc4_plane_state(state);
951 vc4_state->dlist_count = 0;
953 if (!plane_enabled(state))
956 ret = vc4_plane_mode_set(plane, state);
960 return vc4_plane_allocate_lbm(state);
963 static void vc4_plane_atomic_update(struct drm_plane *plane,
964 struct drm_plane_state *old_state)
966 /* No contents here. Since we don't know where in the CRTC's
967 * dlist we should be stored, our dlist is uploaded to the
968 * hardware with vc4_plane_write_dlist() at CRTC atomic_flush
973 u32 vc4_plane_write_dlist(struct drm_plane *plane, u32 __iomem *dlist)
975 struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
978 vc4_state->hw_dlist = dlist;
980 /* Can't memcpy_toio() because it needs to be 32-bit writes. */
981 for (i = 0; i < vc4_state->dlist_count; i++)
982 writel(vc4_state->dlist[i], &dlist[i]);
984 return vc4_state->dlist_count;
987 u32 vc4_plane_dlist_size(const struct drm_plane_state *state)
989 const struct vc4_plane_state *vc4_state =
990 container_of(state, typeof(*vc4_state), base);
992 return vc4_state->dlist_count;
995 /* Updates the plane to immediately (well, once the FIFO needs
996 * refilling) scan out from at a new framebuffer.
998 void vc4_plane_async_set_fb(struct drm_plane *plane, struct drm_framebuffer *fb)
1000 struct vc4_plane_state *vc4_state = to_vc4_plane_state(plane->state);
1001 struct drm_gem_cma_object *bo = drm_fb_cma_get_gem_obj(fb, 0);
1004 /* We're skipping the address adjustment for negative origin,
1005 * because this is only called on the primary plane.
1007 WARN_ON_ONCE(plane->state->crtc_x < 0 || plane->state->crtc_y < 0);
1008 addr = bo->paddr + fb->offsets[0];
1010 /* Write the new address into the hardware immediately. The
1011 * scanout will start from this address as soon as the FIFO
1012 * needs to refill with pixels.
1014 writel(addr, &vc4_state->hw_dlist[vc4_state->ptr0_offset]);
1016 /* Also update the CPU-side dlist copy, so that any later
1017 * atomic updates that don't do a new modeset on our plane
1018 * also use our updated address.
1020 vc4_state->dlist[vc4_state->ptr0_offset] = addr;
1023 static void vc4_plane_atomic_async_update(struct drm_plane *plane,
1024 struct drm_plane_state *state)
1026 struct vc4_plane_state *vc4_state, *new_vc4_state;
1028 drm_atomic_set_fb_for_plane(plane->state, state->fb);
1029 plane->state->crtc_x = state->crtc_x;
1030 plane->state->crtc_y = state->crtc_y;
1031 plane->state->crtc_w = state->crtc_w;
1032 plane->state->crtc_h = state->crtc_h;
1033 plane->state->src_x = state->src_x;
1034 plane->state->src_y = state->src_y;
1035 plane->state->src_w = state->src_w;
1036 plane->state->src_h = state->src_h;
1037 plane->state->src_h = state->src_h;
1038 plane->state->alpha = state->alpha;
1039 plane->state->pixel_blend_mode = state->pixel_blend_mode;
1040 plane->state->rotation = state->rotation;
1041 plane->state->zpos = state->zpos;
1042 plane->state->normalized_zpos = state->normalized_zpos;
1043 plane->state->color_encoding = state->color_encoding;
1044 plane->state->color_range = state->color_range;
1045 plane->state->src = state->src;
1046 plane->state->dst = state->dst;
1047 plane->state->visible = state->visible;
1049 new_vc4_state = to_vc4_plane_state(state);
1050 vc4_state = to_vc4_plane_state(plane->state);
1052 vc4_state->crtc_x = new_vc4_state->crtc_x;
1053 vc4_state->crtc_y = new_vc4_state->crtc_y;
1054 vc4_state->crtc_h = new_vc4_state->crtc_h;
1055 vc4_state->crtc_w = new_vc4_state->crtc_w;
1056 vc4_state->src_x = new_vc4_state->src_x;
1057 vc4_state->src_y = new_vc4_state->src_y;
1058 memcpy(vc4_state->src_w, new_vc4_state->src_w,
1059 sizeof(vc4_state->src_w));
1060 memcpy(vc4_state->src_h, new_vc4_state->src_h,
1061 sizeof(vc4_state->src_h));
1062 memcpy(vc4_state->x_scaling, new_vc4_state->x_scaling,
1063 sizeof(vc4_state->x_scaling));
1064 memcpy(vc4_state->y_scaling, new_vc4_state->y_scaling,
1065 sizeof(vc4_state->y_scaling));
1066 vc4_state->is_unity = new_vc4_state->is_unity;
1067 vc4_state->is_yuv = new_vc4_state->is_yuv;
1068 memcpy(vc4_state->offsets, new_vc4_state->offsets,
1069 sizeof(vc4_state->offsets));
1070 vc4_state->needs_bg_fill = new_vc4_state->needs_bg_fill;
1072 /* Update the current vc4_state pos0, pos2 and ptr0 dlist entries. */
1073 vc4_state->dlist[vc4_state->pos0_offset] =
1074 new_vc4_state->dlist[vc4_state->pos0_offset];
1075 vc4_state->dlist[vc4_state->pos2_offset] =
1076 new_vc4_state->dlist[vc4_state->pos2_offset];
1077 vc4_state->dlist[vc4_state->ptr0_offset] =
1078 new_vc4_state->dlist[vc4_state->ptr0_offset];
1080 /* Note that we can't just call vc4_plane_write_dlist()
1081 * because that would smash the context data that the HVS is
1084 writel(vc4_state->dlist[vc4_state->pos0_offset],
1085 &vc4_state->hw_dlist[vc4_state->pos0_offset]);
1086 writel(vc4_state->dlist[vc4_state->pos2_offset],
1087 &vc4_state->hw_dlist[vc4_state->pos2_offset]);
1088 writel(vc4_state->dlist[vc4_state->ptr0_offset],
1089 &vc4_state->hw_dlist[vc4_state->ptr0_offset]);
1092 static int vc4_plane_atomic_async_check(struct drm_plane *plane,
1093 struct drm_plane_state *state)
1095 struct vc4_plane_state *old_vc4_state, *new_vc4_state;
1099 ret = vc4_plane_mode_set(plane, state);
1103 old_vc4_state = to_vc4_plane_state(plane->state);
1104 new_vc4_state = to_vc4_plane_state(state);
1105 if (old_vc4_state->dlist_count != new_vc4_state->dlist_count ||
1106 old_vc4_state->pos0_offset != new_vc4_state->pos0_offset ||
1107 old_vc4_state->pos2_offset != new_vc4_state->pos2_offset ||
1108 old_vc4_state->ptr0_offset != new_vc4_state->ptr0_offset ||
1109 vc4_lbm_size(plane->state) != vc4_lbm_size(state))
1112 /* Only pos0, pos2 and ptr0 DWORDS can be updated in an async update
1113 * if anything else has changed, fallback to a sync update.
1115 for (i = 0; i < new_vc4_state->dlist_count; i++) {
1116 if (i == new_vc4_state->pos0_offset ||
1117 i == new_vc4_state->pos2_offset ||
1118 i == new_vc4_state->ptr0_offset ||
1119 (new_vc4_state->lbm_offset &&
1120 i == new_vc4_state->lbm_offset))
1123 if (new_vc4_state->dlist[i] != old_vc4_state->dlist[i])
1130 static int vc4_prepare_fb(struct drm_plane *plane,
1131 struct drm_plane_state *state)
1134 struct dma_fence *fence;
1140 bo = to_vc4_bo(&drm_fb_cma_get_gem_obj(state->fb, 0)->base);
1142 fence = reservation_object_get_excl_rcu(bo->base.base.resv);
1143 drm_atomic_set_fence_for_plane(state, fence);
1145 if (plane->state->fb == state->fb)
1148 ret = vc4_bo_inc_usecnt(bo);
1155 static void vc4_cleanup_fb(struct drm_plane *plane,
1156 struct drm_plane_state *state)
1160 if (plane->state->fb == state->fb || !state->fb)
1163 bo = to_vc4_bo(&drm_fb_cma_get_gem_obj(state->fb, 0)->base);
1164 vc4_bo_dec_usecnt(bo);
1167 static const struct drm_plane_helper_funcs vc4_plane_helper_funcs = {
1168 .atomic_check = vc4_plane_atomic_check,
1169 .atomic_update = vc4_plane_atomic_update,
1170 .prepare_fb = vc4_prepare_fb,
1171 .cleanup_fb = vc4_cleanup_fb,
1172 .atomic_async_check = vc4_plane_atomic_async_check,
1173 .atomic_async_update = vc4_plane_atomic_async_update,
1176 static void vc4_plane_destroy(struct drm_plane *plane)
1178 drm_plane_cleanup(plane);
1181 static bool vc4_format_mod_supported(struct drm_plane *plane,
1185 /* Support T_TILING for RGB formats only. */
1187 case DRM_FORMAT_XRGB8888:
1188 case DRM_FORMAT_ARGB8888:
1189 case DRM_FORMAT_ABGR8888:
1190 case DRM_FORMAT_XBGR8888:
1191 case DRM_FORMAT_RGB565:
1192 case DRM_FORMAT_BGR565:
1193 case DRM_FORMAT_ARGB1555:
1194 case DRM_FORMAT_XRGB1555:
1195 switch (fourcc_mod_broadcom_mod(modifier)) {
1196 case DRM_FORMAT_MOD_LINEAR:
1197 case DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED:
1202 case DRM_FORMAT_NV12:
1203 case DRM_FORMAT_NV21:
1204 switch (fourcc_mod_broadcom_mod(modifier)) {
1205 case DRM_FORMAT_MOD_LINEAR:
1206 case DRM_FORMAT_MOD_BROADCOM_SAND64:
1207 case DRM_FORMAT_MOD_BROADCOM_SAND128:
1208 case DRM_FORMAT_MOD_BROADCOM_SAND256:
1213 case DRM_FORMAT_YUV422:
1214 case DRM_FORMAT_YVU422:
1215 case DRM_FORMAT_YUV420:
1216 case DRM_FORMAT_YVU420:
1217 case DRM_FORMAT_NV16:
1218 case DRM_FORMAT_NV61:
1220 return (modifier == DRM_FORMAT_MOD_LINEAR);
1224 static const struct drm_plane_funcs vc4_plane_funcs = {
1225 .update_plane = drm_atomic_helper_update_plane,
1226 .disable_plane = drm_atomic_helper_disable_plane,
1227 .destroy = vc4_plane_destroy,
1228 .set_property = NULL,
1229 .reset = vc4_plane_reset,
1230 .atomic_duplicate_state = vc4_plane_duplicate_state,
1231 .atomic_destroy_state = vc4_plane_destroy_state,
1232 .format_mod_supported = vc4_format_mod_supported,
1235 struct drm_plane *vc4_plane_init(struct drm_device *dev,
1236 enum drm_plane_type type)
1238 struct drm_plane *plane = NULL;
1239 struct vc4_plane *vc4_plane;
1240 u32 formats[ARRAY_SIZE(hvs_formats)];
1243 static const uint64_t modifiers[] = {
1244 DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED,
1245 DRM_FORMAT_MOD_BROADCOM_SAND128,
1246 DRM_FORMAT_MOD_BROADCOM_SAND64,
1247 DRM_FORMAT_MOD_BROADCOM_SAND256,
1248 DRM_FORMAT_MOD_LINEAR,
1249 DRM_FORMAT_MOD_INVALID
1252 vc4_plane = devm_kzalloc(dev->dev, sizeof(*vc4_plane),
1255 return ERR_PTR(-ENOMEM);
1257 for (i = 0; i < ARRAY_SIZE(hvs_formats); i++)
1258 formats[i] = hvs_formats[i].drm;
1260 plane = &vc4_plane->base;
1261 ret = drm_universal_plane_init(dev, plane, 0,
1263 formats, ARRAY_SIZE(formats),
1264 modifiers, type, NULL);
1266 drm_plane_helper_add(plane, &vc4_plane_helper_funcs);
1268 drm_plane_create_alpha_property(plane);
1269 drm_plane_create_rotation_property(plane, DRM_MODE_ROTATE_0,
1271 DRM_MODE_ROTATE_180 |
1272 DRM_MODE_REFLECT_X |
1273 DRM_MODE_REFLECT_Y);