1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) 2015 Broadcom
4 */
5
6/**
7 * DOC: VC4 KMS
8 *
9 * This is the general code for implementing KMS mode setting that
10 * doesn't clearly associate with any of the other objects (plane,
11 * crtc, HDMI encoder).
12 */
13
14#include <linux/clk.h>
15#include <linux/sort.h>
16
17#include <drm/drm_atomic.h>
18#include <drm/drm_atomic_helper.h>
19#include <drm/drm_crtc.h>
20#include <drm/drm_fourcc.h>
21#include <drm/drm_gem_framebuffer_helper.h>
22#include <drm/drm_probe_helper.h>
23#include <drm/drm_vblank.h>
24
25#include "vc4_drv.h"
26#include "vc4_regs.h"
27
28struct vc4_ctm_state {
29	struct drm_private_state base;
30	struct drm_color_ctm *ctm;
31	int fifo;
32};
33
34#define to_vc4_ctm_state(_state)				\
35	container_of_const(_state, struct vc4_ctm_state, base)
36
37struct vc4_load_tracker_state {
38	struct drm_private_state base;
39	u64 hvs_load;
40	u64 membus_load;
41};
42
43#define to_vc4_load_tracker_state(_state)				\
44	container_of_const(_state, struct vc4_load_tracker_state, base)
45
46static struct vc4_ctm_state *vc4_get_ctm_state(struct drm_atomic_state *state,
47					       struct drm_private_obj *manager)
48{
49	struct drm_device *dev = state->dev;
50	struct vc4_dev *vc4 = to_vc4_dev(dev);
51	struct drm_private_state *priv_state;
52	int ret;
53
54	ret = drm_modeset_lock(&vc4->ctm_state_lock, state->acquire_ctx);
55	if (ret)
56		return ERR_PTR(ret);
57
58	priv_state = drm_atomic_get_private_obj_state(state, manager);
59	if (IS_ERR(priv_state))
60		return ERR_CAST(priv_state);
61
62	return to_vc4_ctm_state(priv_state);
63}
64
65static struct drm_private_state *
66vc4_ctm_duplicate_state(struct drm_private_obj *obj)
67{
68	struct vc4_ctm_state *state;
69
70	state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
71	if (!state)
72		return NULL;
73
74	__drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
75
76	return &state->base;
77}
78
79static void vc4_ctm_destroy_state(struct drm_private_obj *obj,
80				  struct drm_private_state *state)
81{
82	struct vc4_ctm_state *ctm_state = to_vc4_ctm_state(state);
83
84	kfree(ctm_state);
85}
86
87static const struct drm_private_state_funcs vc4_ctm_state_funcs = {
88	.atomic_duplicate_state = vc4_ctm_duplicate_state,
89	.atomic_destroy_state = vc4_ctm_destroy_state,
90};
91
92static void vc4_ctm_obj_fini(struct drm_device *dev, void *unused)
93{
94	struct vc4_dev *vc4 = to_vc4_dev(dev);
95
96	drm_atomic_private_obj_fini(&vc4->ctm_manager);
97}
98
99static int vc4_ctm_obj_init(struct vc4_dev *vc4)
100{
101	struct vc4_ctm_state *ctm_state;
102
103	drm_modeset_lock_init(&vc4->ctm_state_lock);
104
105	ctm_state = kzalloc(sizeof(*ctm_state), GFP_KERNEL);
106	if (!ctm_state)
107		return -ENOMEM;
108
109	drm_atomic_private_obj_init(&vc4->base, &vc4->ctm_manager, &ctm_state->base,
110				    &vc4_ctm_state_funcs);
111
112	return drmm_add_action_or_reset(&vc4->base, vc4_ctm_obj_fini, NULL);
113}
114
115/* Converts a DRM S31.32 value to the HW S0.9 format. */
116static u16 vc4_ctm_s31_32_to_s0_9(u64 in)
117{
118	u16 r;
119
120	/* Sign bit. */
121	r = in & BIT_ULL(63) ? BIT(9) : 0;
122
123	if ((in & GENMASK_ULL(62, 32)) > 0) {
124		/* We have zero integer bits so we can only saturate here. */
125		r |= GENMASK(8, 0);
126	} else {
127		/* Otherwise take the 9 most important fractional bits. */
128		r |= (in >> 23) & GENMASK(8, 0);
129	}
130
131	return r;
132}
133
134static void
135vc4_ctm_commit(struct vc4_dev *vc4, struct drm_atomic_state *state)
136{
137	struct vc4_hvs *hvs = vc4->hvs;
138	struct vc4_ctm_state *ctm_state = to_vc4_ctm_state(vc4->ctm_manager.state);
139	struct drm_color_ctm *ctm = ctm_state->ctm;
140
141	if (ctm_state->fifo) {
142		HVS_WRITE(SCALER_OLEDCOEF2,
143			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[0]),
144					SCALER_OLEDCOEF2_R_TO_R) |
145			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[3]),
146					SCALER_OLEDCOEF2_R_TO_G) |
147			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[6]),
148					SCALER_OLEDCOEF2_R_TO_B));
149		HVS_WRITE(SCALER_OLEDCOEF1,
150			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[1]),
151					SCALER_OLEDCOEF1_G_TO_R) |
152			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[4]),
153					SCALER_OLEDCOEF1_G_TO_G) |
154			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[7]),
155					SCALER_OLEDCOEF1_G_TO_B));
156		HVS_WRITE(SCALER_OLEDCOEF0,
157			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[2]),
158					SCALER_OLEDCOEF0_B_TO_R) |
159			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[5]),
160					SCALER_OLEDCOEF0_B_TO_G) |
161			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[8]),
162					SCALER_OLEDCOEF0_B_TO_B));
163	}
164
165	HVS_WRITE(SCALER_OLEDOFFS,
166		  VC4_SET_FIELD(ctm_state->fifo, SCALER_OLEDOFFS_DISPFIFO));
167}
168
169struct vc4_hvs_state *
170vc4_hvs_get_new_global_state(const struct drm_atomic_state *state)
171{
172	struct vc4_dev *vc4 = to_vc4_dev(state->dev);
173	struct drm_private_state *priv_state;
174
175	priv_state = drm_atomic_get_new_private_obj_state(state, &vc4->hvs_channels);
176	if (!priv_state)
177		return ERR_PTR(-EINVAL);
178
179	return to_vc4_hvs_state(priv_state);
180}
181
182struct vc4_hvs_state *
183vc4_hvs_get_old_global_state(const struct drm_atomic_state *state)
184{
185	struct vc4_dev *vc4 = to_vc4_dev(state->dev);
186	struct drm_private_state *priv_state;
187
188	priv_state = drm_atomic_get_old_private_obj_state(state, &vc4->hvs_channels);
189	if (!priv_state)
190		return ERR_PTR(-EINVAL);
191
192	return to_vc4_hvs_state(priv_state);
193}
194
195struct vc4_hvs_state *
196vc4_hvs_get_global_state(struct drm_atomic_state *state)
197{
198	struct vc4_dev *vc4 = to_vc4_dev(state->dev);
199	struct drm_private_state *priv_state;
200
201	priv_state = drm_atomic_get_private_obj_state(state, &vc4->hvs_channels);
202	if (IS_ERR(priv_state))
203		return ERR_CAST(priv_state);
204
205	return to_vc4_hvs_state(priv_state);
206}
207
208static void vc4_hvs_pv_muxing_commit(struct vc4_dev *vc4,
209				     struct drm_atomic_state *state)
210{
211	struct vc4_hvs *hvs = vc4->hvs;
212	struct drm_crtc_state *crtc_state;
213	struct drm_crtc *crtc;
214	unsigned int i;
215
216	for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
217		struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
218		struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc_state);
219		u32 dispctrl;
220		u32 dsp3_mux;
221
222		if (!crtc_state->active)
223			continue;
224
225		if (vc4_state->assigned_channel != 2)
226			continue;
227
228		/*
229		 * SCALER_DISPCTRL_DSP3 = X, where X < 2 means 'connect DSP3 to
230		 * FIFO X'.
231		 * SCALER_DISPCTRL_DSP3 = 3 means 'disable DSP 3'.
232		 *
233		 * DSP3 is connected to FIFO2 unless the transposer is
234		 * enabled. In this case, FIFO 2 is directly accessed by the
235		 * TXP IP, and we need to disable the FIFO2 -> pixelvalve1
236		 * route.
237		 */
238		if (vc4_crtc->feeds_txp)
239			dsp3_mux = VC4_SET_FIELD(3, SCALER_DISPCTRL_DSP3_MUX);
240		else
241			dsp3_mux = VC4_SET_FIELD(2, SCALER_DISPCTRL_DSP3_MUX);
242
243		dispctrl = HVS_READ(SCALER_DISPCTRL) &
244			   ~SCALER_DISPCTRL_DSP3_MUX_MASK;
245		HVS_WRITE(SCALER_DISPCTRL, dispctrl | dsp3_mux);
246	}
247}
248
249static void vc5_hvs_pv_muxing_commit(struct vc4_dev *vc4,
250				     struct drm_atomic_state *state)
251{
252	struct vc4_hvs *hvs = vc4->hvs;
253	struct drm_crtc_state *crtc_state;
254	struct drm_crtc *crtc;
255	unsigned char mux;
256	unsigned int i;
257	u32 reg;
258
259	for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
260		struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc_state);
261		struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
262		unsigned int channel = vc4_state->assigned_channel;
263
264		if (!vc4_state->update_muxing)
265			continue;
266
267		switch (vc4_crtc->data->hvs_output) {
268		case 2:
269			drm_WARN_ON(&vc4->base,
270				    VC4_GET_FIELD(HVS_READ(SCALER_DISPCTRL),
271						  SCALER_DISPCTRL_DSP3_MUX) == channel);
272
273			mux = (channel == 2) ? 0 : 1;
274			reg = HVS_READ(SCALER_DISPECTRL);
275			HVS_WRITE(SCALER_DISPECTRL,
276				  (reg & ~SCALER_DISPECTRL_DSP2_MUX_MASK) |
277				  VC4_SET_FIELD(mux, SCALER_DISPECTRL_DSP2_MUX));
278			break;
279
280		case 3:
281			if (channel == VC4_HVS_CHANNEL_DISABLED)
282				mux = 3;
283			else
284				mux = channel;
285
286			reg = HVS_READ(SCALER_DISPCTRL);
287			HVS_WRITE(SCALER_DISPCTRL,
288				  (reg & ~SCALER_DISPCTRL_DSP3_MUX_MASK) |
289				  VC4_SET_FIELD(mux, SCALER_DISPCTRL_DSP3_MUX));
290			break;
291
292		case 4:
293			if (channel == VC4_HVS_CHANNEL_DISABLED)
294				mux = 3;
295			else
296				mux = channel;
297
298			reg = HVS_READ(SCALER_DISPEOLN);
299			HVS_WRITE(SCALER_DISPEOLN,
300				  (reg & ~SCALER_DISPEOLN_DSP4_MUX_MASK) |
301				  VC4_SET_FIELD(mux, SCALER_DISPEOLN_DSP4_MUX));
302
303			break;
304
305		case 5:
306			if (channel == VC4_HVS_CHANNEL_DISABLED)
307				mux = 3;
308			else
309				mux = channel;
310
311			reg = HVS_READ(SCALER_DISPDITHER);
312			HVS_WRITE(SCALER_DISPDITHER,
313				  (reg & ~SCALER_DISPDITHER_DSP5_MUX_MASK) |
314				  VC4_SET_FIELD(mux, SCALER_DISPDITHER_DSP5_MUX));
315			break;
316
317		default:
318			break;
319		}
320	}
321}
322
323static void vc4_atomic_commit_tail(struct drm_atomic_state *state)
324{
325	struct drm_device *dev = state->dev;
326	struct vc4_dev *vc4 = to_vc4_dev(dev);
327	struct vc4_hvs *hvs = vc4->hvs;
328	struct drm_crtc_state *new_crtc_state;
329	struct vc4_hvs_state *new_hvs_state;
330	struct drm_crtc *crtc;
331	struct vc4_hvs_state *old_hvs_state;
332	unsigned int channel;
333	int i;
334
335	old_hvs_state = vc4_hvs_get_old_global_state(state);
336	if (WARN_ON(IS_ERR(old_hvs_state)))
337		return;
338
339	new_hvs_state = vc4_hvs_get_new_global_state(state);
340	if (WARN_ON(IS_ERR(new_hvs_state)))
341		return;
342
343	for_each_new_crtc_in_state(state, crtc, new_crtc_state, i) {
344		struct vc4_crtc_state *vc4_crtc_state;
345
346		if (!new_crtc_state->commit)
347			continue;
348
349		vc4_crtc_state = to_vc4_crtc_state(new_crtc_state);
350		vc4_hvs_mask_underrun(hvs, vc4_crtc_state->assigned_channel);
351	}
352
353	for (channel = 0; channel < HVS_NUM_CHANNELS; channel++) {
354		struct drm_crtc_commit *commit;
355		int ret;
356
357		if (!old_hvs_state->fifo_state[channel].in_use)
358			continue;
359
360		commit = old_hvs_state->fifo_state[channel].pending_commit;
361		if (!commit)
362			continue;
363
364		ret = drm_crtc_commit_wait(commit);
365		if (ret)
366			drm_err(dev, "Timed out waiting for commit\n");
367
368		drm_crtc_commit_put(commit);
369		old_hvs_state->fifo_state[channel].pending_commit = NULL;
370	}
371
372	if (vc4->is_vc5) {
373		unsigned long state_rate = max(old_hvs_state->core_clock_rate,
374					       new_hvs_state->core_clock_rate);
375		unsigned long core_rate = clamp_t(unsigned long, state_rate,
376						  500000000, hvs->max_core_rate);
377
378		drm_dbg(dev, "Raising the core clock at %lu Hz\n", core_rate);
379
380		/*
381		 * Do a temporary request on the core clock during the
382		 * modeset.
383		 */
384		WARN_ON(clk_set_min_rate(hvs->core_clk, core_rate));
385	}
386
387	drm_atomic_helper_commit_modeset_disables(dev, state);
388
389	vc4_ctm_commit(vc4, state);
390
391	if (vc4->is_vc5)
392		vc5_hvs_pv_muxing_commit(vc4, state);
393	else
394		vc4_hvs_pv_muxing_commit(vc4, state);
395
396	drm_atomic_helper_commit_planes(dev, state,
397					DRM_PLANE_COMMIT_ACTIVE_ONLY);
398
399	drm_atomic_helper_commit_modeset_enables(dev, state);
400
401	drm_atomic_helper_fake_vblank(state);
402
403	drm_atomic_helper_commit_hw_done(state);
404
405	drm_atomic_helper_wait_for_flip_done(dev, state);
406
407	drm_atomic_helper_cleanup_planes(dev, state);
408
409	if (vc4->is_vc5) {
410		unsigned long core_rate = min_t(unsigned long,
411						hvs->max_core_rate,
412						new_hvs_state->core_clock_rate);
413
414		drm_dbg(dev, "Running the core clock at %lu Hz\n", core_rate);
415
416		/*
417		 * Request a clock rate based on the current HVS
418		 * requirements.
419		 */
420		WARN_ON(clk_set_min_rate(hvs->core_clk, core_rate));
421
422		drm_dbg(dev, "Core clock actual rate: %lu Hz\n",
423			clk_get_rate(hvs->core_clk));
424	}
425}
426
427static int vc4_atomic_commit_setup(struct drm_atomic_state *state)
428{
429	struct drm_crtc_state *crtc_state;
430	struct vc4_hvs_state *hvs_state;
431	struct drm_crtc *crtc;
432	unsigned int i;
433
434	hvs_state = vc4_hvs_get_new_global_state(state);
435	if (WARN_ON(IS_ERR(hvs_state)))
436		return PTR_ERR(hvs_state);
437
438	for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
439		struct vc4_crtc_state *vc4_crtc_state =
440			to_vc4_crtc_state(crtc_state);
441		unsigned int channel =
442			vc4_crtc_state->assigned_channel;
443
444		if (channel == VC4_HVS_CHANNEL_DISABLED)
445			continue;
446
447		if (!hvs_state->fifo_state[channel].in_use)
448			continue;
449
450		hvs_state->fifo_state[channel].pending_commit =
451			drm_crtc_commit_get(crtc_state->commit);
452	}
453
454	return 0;
455}
456
457static struct drm_framebuffer *vc4_fb_create(struct drm_device *dev,
458					     struct drm_file *file_priv,
459					     const struct drm_mode_fb_cmd2 *mode_cmd)
460{
461	struct vc4_dev *vc4 = to_vc4_dev(dev);
462	struct drm_mode_fb_cmd2 mode_cmd_local;
463
464	if (WARN_ON_ONCE(vc4->is_vc5))
465		return ERR_PTR(-ENODEV);
466
467	/* If the user didn't specify a modifier, use the
468	 * vc4_set_tiling_ioctl() state for the BO.
469	 */
470	if (!(mode_cmd->flags & DRM_MODE_FB_MODIFIERS)) {
471		struct drm_gem_object *gem_obj;
472		struct vc4_bo *bo;
473
474		gem_obj = drm_gem_object_lookup(file_priv,
475						mode_cmd->handles[0]);
476		if (!gem_obj) {
477			DRM_DEBUG("Failed to look up GEM BO %d\n",
478				  mode_cmd->handles[0]);
479			return ERR_PTR(-ENOENT);
480		}
481		bo = to_vc4_bo(gem_obj);
482
483		mode_cmd_local = *mode_cmd;
484
485		if (bo->t_format) {
486			mode_cmd_local.modifier[0] =
487				DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
488		} else {
489			mode_cmd_local.modifier[0] = DRM_FORMAT_MOD_NONE;
490		}
491
492		drm_gem_object_put(gem_obj);
493
494		mode_cmd = &mode_cmd_local;
495	}
496
497	return drm_gem_fb_create(dev, file_priv, mode_cmd);
498}
499
500/* Our CTM has some peculiar limitations: we can only enable it for one CRTC
501 * at a time and the HW only supports S0.9 scalars. To account for the latter,
502 * we don't allow userland to set a CTM that we have no hope of approximating.
503 */
504static int
505vc4_ctm_atomic_check(struct drm_device *dev, struct drm_atomic_state *state)
506{
507	struct vc4_dev *vc4 = to_vc4_dev(dev);
508	struct vc4_ctm_state *ctm_state = NULL;
509	struct drm_crtc *crtc;
510	struct drm_crtc_state *old_crtc_state, *new_crtc_state;
511	struct drm_color_ctm *ctm;
512	int i;
513
514	for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
515		/* CTM is being disabled. */
516		if (!new_crtc_state->ctm && old_crtc_state->ctm) {
517			ctm_state = vc4_get_ctm_state(state, &vc4->ctm_manager);
518			if (IS_ERR(ctm_state))
519				return PTR_ERR(ctm_state);
520			ctm_state->fifo = 0;
521		}
522	}
523
524	for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
525		if (new_crtc_state->ctm == old_crtc_state->ctm)
526			continue;
527
528		if (!ctm_state) {
529			ctm_state = vc4_get_ctm_state(state, &vc4->ctm_manager);
530			if (IS_ERR(ctm_state))
531				return PTR_ERR(ctm_state);
532		}
533
534		/* CTM is being enabled or the matrix changed. */
535		if (new_crtc_state->ctm) {
536			struct vc4_crtc_state *vc4_crtc_state =
537				to_vc4_crtc_state(new_crtc_state);
538
539			/* fifo is 1-based since 0 disables CTM. */
540			int fifo = vc4_crtc_state->assigned_channel + 1;
541
542			/* Check userland isn't trying to turn on CTM for more
543			 * than one CRTC at a time.
544			 */
545			if (ctm_state->fifo && ctm_state->fifo != fifo) {
546				DRM_DEBUG_DRIVER("Too many CTM configured\n");
547				return -EINVAL;
548			}
549
550			/* Check we can approximate the specified CTM.
551			 * We disallow scalars |c| > 1.0 since the HW has
552			 * no integer bits.
553			 */
554			ctm = new_crtc_state->ctm->data;
555			for (i = 0; i < ARRAY_SIZE(ctm->matrix); i++) {
556				u64 val = ctm->matrix[i];
557
558				val &= ~BIT_ULL(63);
559				if (val > BIT_ULL(32))
560					return -EINVAL;
561			}
562
563			ctm_state->fifo = fifo;
564			ctm_state->ctm = ctm;
565		}
566	}
567
568	return 0;
569}
570
571static int vc4_load_tracker_atomic_check(struct drm_atomic_state *state)
572{
573	struct drm_plane_state *old_plane_state, *new_plane_state;
574	struct vc4_dev *vc4 = to_vc4_dev(state->dev);
575	struct vc4_load_tracker_state *load_state;
576	struct drm_private_state *priv_state;
577	struct drm_plane *plane;
578	int i;
579
580	priv_state = drm_atomic_get_private_obj_state(state,
581						      &vc4->load_tracker);
582	if (IS_ERR(priv_state))
583		return PTR_ERR(priv_state);
584
585	load_state = to_vc4_load_tracker_state(priv_state);
586	for_each_oldnew_plane_in_state(state, plane, old_plane_state,
587				       new_plane_state, i) {
588		struct vc4_plane_state *vc4_plane_state;
589
590		if (old_plane_state->fb && old_plane_state->crtc) {
591			vc4_plane_state = to_vc4_plane_state(old_plane_state);
592			load_state->membus_load -= vc4_plane_state->membus_load;
593			load_state->hvs_load -= vc4_plane_state->hvs_load;
594		}
595
596		if (new_plane_state->fb && new_plane_state->crtc) {
597			vc4_plane_state = to_vc4_plane_state(new_plane_state);
598			load_state->membus_load += vc4_plane_state->membus_load;
599			load_state->hvs_load += vc4_plane_state->hvs_load;
600		}
601	}
602
603	/* Don't check the load when the tracker is disabled. */
604	if (!vc4->load_tracker_enabled)
605		return 0;
606
607	/* The absolute limit is 2Gbyte/sec, but let's take a margin to let
608	 * the system work when other blocks are accessing the memory.
609	 */
610	if (load_state->membus_load > SZ_1G + SZ_512M)
611		return -ENOSPC;
612
613	/* HVS clock is supposed to run @ 250Mhz, let's take a margin and
614	 * consider the maximum number of cycles is 240M.
615	 */
616	if (load_state->hvs_load > 240000000ULL)
617		return -ENOSPC;
618
619	return 0;
620}
621
622static struct drm_private_state *
623vc4_load_tracker_duplicate_state(struct drm_private_obj *obj)
624{
625	struct vc4_load_tracker_state *state;
626
627	state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
628	if (!state)
629		return NULL;
630
631	__drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
632
633	return &state->base;
634}
635
636static void vc4_load_tracker_destroy_state(struct drm_private_obj *obj,
637					   struct drm_private_state *state)
638{
639	struct vc4_load_tracker_state *load_state;
640
641	load_state = to_vc4_load_tracker_state(state);
642	kfree(load_state);
643}
644
645static const struct drm_private_state_funcs vc4_load_tracker_state_funcs = {
646	.atomic_duplicate_state = vc4_load_tracker_duplicate_state,
647	.atomic_destroy_state = vc4_load_tracker_destroy_state,
648};
649
650static void vc4_load_tracker_obj_fini(struct drm_device *dev, void *unused)
651{
652	struct vc4_dev *vc4 = to_vc4_dev(dev);
653
654	drm_atomic_private_obj_fini(&vc4->load_tracker);
655}
656
657static int vc4_load_tracker_obj_init(struct vc4_dev *vc4)
658{
659	struct vc4_load_tracker_state *load_state;
660
661	load_state = kzalloc(sizeof(*load_state), GFP_KERNEL);
662	if (!load_state)
663		return -ENOMEM;
664
665	drm_atomic_private_obj_init(&vc4->base, &vc4->load_tracker,
666				    &load_state->base,
667				    &vc4_load_tracker_state_funcs);
668
669	return drmm_add_action_or_reset(&vc4->base, vc4_load_tracker_obj_fini, NULL);
670}
671
672static struct drm_private_state *
673vc4_hvs_channels_duplicate_state(struct drm_private_obj *obj)
674{
675	struct vc4_hvs_state *old_state = to_vc4_hvs_state(obj->state);
676	struct vc4_hvs_state *state;
677	unsigned int i;
678
679	state = kzalloc(sizeof(*state), GFP_KERNEL);
680	if (!state)
681		return NULL;
682
683	__drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
684
685	for (i = 0; i < HVS_NUM_CHANNELS; i++) {
686		state->fifo_state[i].in_use = old_state->fifo_state[i].in_use;
687		state->fifo_state[i].fifo_load = old_state->fifo_state[i].fifo_load;
688	}
689
690	state->core_clock_rate = old_state->core_clock_rate;
691
692	return &state->base;
693}
694
695static void vc4_hvs_channels_destroy_state(struct drm_private_obj *obj,
696					   struct drm_private_state *state)
697{
698	struct vc4_hvs_state *hvs_state = to_vc4_hvs_state(state);
699	unsigned int i;
700
701	for (i = 0; i < HVS_NUM_CHANNELS; i++) {
702		if (!hvs_state->fifo_state[i].pending_commit)
703			continue;
704
705		drm_crtc_commit_put(hvs_state->fifo_state[i].pending_commit);
706	}
707
708	kfree(hvs_state);
709}
710
711static void vc4_hvs_channels_print_state(struct drm_printer *p,
712					 const struct drm_private_state *state)
713{
714	const struct vc4_hvs_state *hvs_state = to_vc4_hvs_state(state);
715	unsigned int i;
716
717	drm_printf(p, "HVS State\n");
718	drm_printf(p, "\tCore Clock Rate: %lu\n", hvs_state->core_clock_rate);
719
720	for (i = 0; i < HVS_NUM_CHANNELS; i++) {
721		drm_printf(p, "\tChannel %d\n", i);
722		drm_printf(p, "\t\tin use=%d\n", hvs_state->fifo_state[i].in_use);
723		drm_printf(p, "\t\tload=%lu\n", hvs_state->fifo_state[i].fifo_load);
724	}
725}
726
727static const struct drm_private_state_funcs vc4_hvs_state_funcs = {
728	.atomic_duplicate_state = vc4_hvs_channels_duplicate_state,
729	.atomic_destroy_state = vc4_hvs_channels_destroy_state,
730	.atomic_print_state = vc4_hvs_channels_print_state,
731};
732
733static void vc4_hvs_channels_obj_fini(struct drm_device *dev, void *unused)
734{
735	struct vc4_dev *vc4 = to_vc4_dev(dev);
736
737	drm_atomic_private_obj_fini(&vc4->hvs_channels);
738}
739
740static int vc4_hvs_channels_obj_init(struct vc4_dev *vc4)
741{
742	struct vc4_hvs_state *state;
743
744	state = kzalloc(sizeof(*state), GFP_KERNEL);
745	if (!state)
746		return -ENOMEM;
747
748	drm_atomic_private_obj_init(&vc4->base, &vc4->hvs_channels,
749				    &state->base,
750				    &vc4_hvs_state_funcs);
751
752	return drmm_add_action_or_reset(&vc4->base, vc4_hvs_channels_obj_fini, NULL);
753}
754
755static int cmp_vc4_crtc_hvs_output(const void *a, const void *b)
756{
757	const struct vc4_crtc *crtc_a =
758		to_vc4_crtc(*(const struct drm_crtc **)a);
759	const struct vc4_crtc_data *data_a =
760		vc4_crtc_to_vc4_crtc_data(crtc_a);
761	const struct vc4_crtc *crtc_b =
762		to_vc4_crtc(*(const struct drm_crtc **)b);
763	const struct vc4_crtc_data *data_b =
764		vc4_crtc_to_vc4_crtc_data(crtc_b);
765
766	return data_a->hvs_output - data_b->hvs_output;
767}
768
769/*
770 * The BCM2711 HVS has up to 7 outputs connected to the pixelvalves and
771 * the TXP (and therefore all the CRTCs found on that platform).
772 *
773 * The naive (and our initial) implementation would just iterate over
774 * all the active CRTCs, try to find a suitable FIFO, and then remove it
775 * from the pool of available FIFOs. However, there are a few corner
776 * cases that need to be considered:
777 *
778 * - When running in a dual-display setup (so with two CRTCs involved),
779 *   we can update the state of a single CRTC (for example by changing
780 *   its mode using xrandr under X11) without affecting the other. In
781 *   this case, the other CRTC wouldn't be in the state at all, so we
782 *   need to consider all the running CRTCs in the DRM device to assign
783 *   a FIFO, not just the one in the state.
784 *
785 * - To fix the above, we can't use drm_atomic_get_crtc_state on all
786 *   enabled CRTCs to pull their CRTC state into the global state, since
787 *   a page flip would start considering their vblank to complete. Since
788 *   we don't have a guarantee that they are actually active, that
789 *   vblank might never happen, and shouldn't even be considered if we
790 *   want to do a page flip on a single CRTC. That can be tested by
791 *   doing a modetest -v first on HDMI1 and then on HDMI0.
792 *
793 * - Since we need the pixelvalve to be disabled and enabled back when
794 *   the FIFO is changed, we should keep the FIFO assigned for as long
795 *   as the CRTC is enabled, only considering it free again once that
796 *   CRTC has been disabled. This can be tested by booting X11 on a
797 *   single display, and changing the resolution down and then back up.
798 */
799static int vc4_pv_muxing_atomic_check(struct drm_device *dev,
800				      struct drm_atomic_state *state)
801{
802	struct vc4_hvs_state *hvs_new_state;
803	struct drm_crtc **sorted_crtcs;
804	struct drm_crtc *crtc;
805	unsigned int unassigned_channels = 0;
806	unsigned int i;
807	int ret;
808
809	hvs_new_state = vc4_hvs_get_global_state(state);
810	if (IS_ERR(hvs_new_state))
811		return PTR_ERR(hvs_new_state);
812
813	for (i = 0; i < ARRAY_SIZE(hvs_new_state->fifo_state); i++)
814		if (!hvs_new_state->fifo_state[i].in_use)
815			unassigned_channels |= BIT(i);
816
817	/*
818	 * The problem we have to solve here is that we have up to 7
819	 * encoders, connected to up to 6 CRTCs.
820	 *
821	 * Those CRTCs, depending on the instance, can be routed to 1, 2
822	 * or 3 HVS FIFOs, and we need to set the muxing between FIFOs and
823	 * outputs in the HVS accordingly.
824	 *
825	 * It would be pretty hard to come up with an algorithm that
826	 * would generically solve this. However, the current routing
827	 * trees we support allow us to simplify a bit the problem.
828	 *
829	 * Indeed, with the current supported layouts, if we try to
830	 * assign in the ascending crtc index order the FIFOs, we can't
831	 * fall into the situation where an earlier CRTC that had
832	 * multiple routes is assigned one that was the only option for
833	 * a later CRTC.
834	 *
835	 * If the layout changes and doesn't give us that in the future,
836	 * we will need to have something smarter, but it works so far.
837	 */
838	sorted_crtcs = kmalloc_array(dev->num_crtcs, sizeof(*sorted_crtcs), GFP_KERNEL);
839	if (!sorted_crtcs)
840		return -ENOMEM;
841
842	i = 0;
843	drm_for_each_crtc(crtc, dev)
844		sorted_crtcs[i++] = crtc;
845
846	sort(sorted_crtcs, i, sizeof(*sorted_crtcs), cmp_vc4_crtc_hvs_output, NULL);
847
848	for (i = 0; i < dev->num_crtcs; i++) {
849		struct vc4_crtc_state *old_vc4_crtc_state, *new_vc4_crtc_state;
850		struct drm_crtc_state *old_crtc_state, *new_crtc_state;
851		struct vc4_crtc *vc4_crtc;
852		unsigned int matching_channels;
853		unsigned int channel;
854
855		crtc = sorted_crtcs[i];
856		if (!crtc)
857			continue;
858		vc4_crtc = to_vc4_crtc(crtc);
859
860		old_crtc_state = drm_atomic_get_old_crtc_state(state, crtc);
861		if (!old_crtc_state)
862			continue;
863		old_vc4_crtc_state = to_vc4_crtc_state(old_crtc_state);
864
865		new_crtc_state = drm_atomic_get_new_crtc_state(state, crtc);
866		if (!new_crtc_state)
867			continue;
868		new_vc4_crtc_state = to_vc4_crtc_state(new_crtc_state);
869
870		drm_dbg(dev, "%s: Trying to find a channel.\n", crtc->name);
871
872		/* Nothing to do here, let's skip it */
873		if (old_crtc_state->enable == new_crtc_state->enable) {
874			if (new_crtc_state->enable)
875				drm_dbg(dev, "%s: Already enabled, reusing channel %d.\n",
876					crtc->name, new_vc4_crtc_state->assigned_channel);
877			else
878				drm_dbg(dev, "%s: Disabled, ignoring.\n", crtc->name);
879
880			continue;
881		}
882
883		/* Muxing will need to be modified, mark it as such */
884		new_vc4_crtc_state->update_muxing = true;
885
886		/* If we're disabling our CRTC, we put back our channel */
887		if (!new_crtc_state->enable) {
888			channel = old_vc4_crtc_state->assigned_channel;
889
890			drm_dbg(dev, "%s: Disabling, Freeing channel %d\n",
891				crtc->name, channel);
892
893			hvs_new_state->fifo_state[channel].in_use = false;
894			new_vc4_crtc_state->assigned_channel = VC4_HVS_CHANNEL_DISABLED;
895			continue;
896		}
897
898		matching_channels = unassigned_channels & vc4_crtc->data->hvs_available_channels;
899		if (!matching_channels) {
900			ret = -EINVAL;
901			goto err_free_crtc_array;
902		}
903
904		channel = ffs(matching_channels) - 1;
905
906		drm_dbg(dev, "Assigned HVS channel %d to CRTC %s\n", channel, crtc->name);
907		new_vc4_crtc_state->assigned_channel = channel;
908		unassigned_channels &= ~BIT(channel);
909		hvs_new_state->fifo_state[channel].in_use = true;
910	}
911
912	kfree(sorted_crtcs);
913	return 0;
914
915err_free_crtc_array:
916	kfree(sorted_crtcs);
917	return ret;
918}
919
920static int
921vc4_core_clock_atomic_check(struct drm_atomic_state *state)
922{
923	struct vc4_dev *vc4 = to_vc4_dev(state->dev);
924	struct drm_private_state *priv_state;
925	struct vc4_hvs_state *hvs_new_state;
926	struct vc4_load_tracker_state *load_state;
927	struct drm_crtc_state *old_crtc_state, *new_crtc_state;
928	struct drm_crtc *crtc;
929	unsigned int num_outputs;
930	unsigned long pixel_rate;
931	unsigned long cob_rate;
932	unsigned int i;
933
934	priv_state = drm_atomic_get_private_obj_state(state,
935						      &vc4->load_tracker);
936	if (IS_ERR(priv_state))
937		return PTR_ERR(priv_state);
938
939	load_state = to_vc4_load_tracker_state(priv_state);
940
941	hvs_new_state = vc4_hvs_get_global_state(state);
942	if (IS_ERR(hvs_new_state))
943		return PTR_ERR(hvs_new_state);
944
945	for_each_oldnew_crtc_in_state(state, crtc,
946				      old_crtc_state,
947				      new_crtc_state,
948				      i) {
949		if (old_crtc_state->active) {
950			struct vc4_crtc_state *old_vc4_state =
951				to_vc4_crtc_state(old_crtc_state);
952			unsigned int channel = old_vc4_state->assigned_channel;
953
954			hvs_new_state->fifo_state[channel].fifo_load = 0;
955		}
956
957		if (new_crtc_state->active) {
958			struct vc4_crtc_state *new_vc4_state =
959				to_vc4_crtc_state(new_crtc_state);
960			unsigned int channel = new_vc4_state->assigned_channel;
961
962			hvs_new_state->fifo_state[channel].fifo_load =
963				new_vc4_state->hvs_load;
964		}
965	}
966
967	cob_rate = 0;
968	num_outputs = 0;
969	for (i = 0; i < HVS_NUM_CHANNELS; i++) {
970		if (!hvs_new_state->fifo_state[i].in_use)
971			continue;
972
973		num_outputs++;
974		cob_rate = max_t(unsigned long,
975				 hvs_new_state->fifo_state[i].fifo_load,
976				 cob_rate);
977	}
978
979	pixel_rate = load_state->hvs_load;
980	if (num_outputs > 1) {
981		pixel_rate = (pixel_rate * 40) / 100;
982	} else {
983		pixel_rate = (pixel_rate * 60) / 100;
984	}
985
986	hvs_new_state->core_clock_rate = max(cob_rate, pixel_rate);
987
988	return 0;
989}
990
991
992static int
993vc4_atomic_check(struct drm_device *dev, struct drm_atomic_state *state)
994{
995	int ret;
996
997	ret = vc4_pv_muxing_atomic_check(dev, state);
998	if (ret)
999		return ret;
1000
1001	ret = vc4_ctm_atomic_check(dev, state);
1002	if (ret < 0)
1003		return ret;
1004
1005	ret = drm_atomic_helper_check(dev, state);
1006	if (ret)
1007		return ret;
1008
1009	ret = vc4_load_tracker_atomic_check(state);
1010	if (ret)
1011		return ret;
1012
1013	return vc4_core_clock_atomic_check(state);
1014}
1015
1016static struct drm_mode_config_helper_funcs vc4_mode_config_helpers = {
1017	.atomic_commit_setup	= vc4_atomic_commit_setup,
1018	.atomic_commit_tail	= vc4_atomic_commit_tail,
1019};
1020
1021static const struct drm_mode_config_funcs vc4_mode_funcs = {
1022	.atomic_check = vc4_atomic_check,
1023	.atomic_commit = drm_atomic_helper_commit,
1024	.fb_create = vc4_fb_create,
1025};
1026
1027static const struct drm_mode_config_funcs vc5_mode_funcs = {
1028	.atomic_check = vc4_atomic_check,
1029	.atomic_commit = drm_atomic_helper_commit,
1030	.fb_create = drm_gem_fb_create,
1031};
1032
1033int vc4_kms_load(struct drm_device *dev)
1034{
1035	struct vc4_dev *vc4 = to_vc4_dev(dev);
1036	int ret;
1037
1038	/*
1039	 * The limits enforced by the load tracker aren't relevant for
1040	 * the BCM2711, but the load tracker computations are used for
1041	 * the core clock rate calculation.
1042	 */
1043	if (!vc4->is_vc5) {
1044		/* Start with the load tracker enabled. Can be
1045		 * disabled through the debugfs load_tracker file.
1046		 */
1047		vc4->load_tracker_enabled = true;
1048	}
1049
1050	/* Set support for vblank irq fast disable, before drm_vblank_init() */
1051	dev->vblank_disable_immediate = true;
1052
1053	ret = drm_vblank_init(dev, dev->mode_config.num_crtc);
1054	if (ret < 0) {
1055		dev_err(dev->dev, "failed to initialize vblank\n");
1056		return ret;
1057	}
1058
1059	if (vc4->is_vc5) {
1060		dev->mode_config.max_width = 7680;
1061		dev->mode_config.max_height = 7680;
1062	} else {
1063		dev->mode_config.max_width = 2048;
1064		dev->mode_config.max_height = 2048;
1065	}
1066
1067	dev->mode_config.funcs = vc4->is_vc5 ? &vc5_mode_funcs : &vc4_mode_funcs;
1068	dev->mode_config.helper_private = &vc4_mode_config_helpers;
1069	dev->mode_config.preferred_depth = 24;
1070	dev->mode_config.async_page_flip = true;
1071	dev->mode_config.normalize_zpos = true;
1072
1073	ret = vc4_ctm_obj_init(vc4);
1074	if (ret)
1075		return ret;
1076
1077	ret = vc4_load_tracker_obj_init(vc4);
1078	if (ret)
1079		return ret;
1080
1081	ret = vc4_hvs_channels_obj_init(vc4);
1082	if (ret)
1083		return ret;
1084
1085	drm_mode_config_reset(dev);
1086
1087	drm_kms_helper_poll_init(dev);
1088
1089	return 0;
1090}
1091