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
16#include <drm/drm_atomic.h>
17#include <drm/drm_atomic_helper.h>
18#include <drm/drm_crtc.h>
19#include <drm/drm_gem_framebuffer_helper.h>
20#include <drm/drm_plane_helper.h>
21#include <drm/drm_probe_helper.h>
22#include <drm/drm_vblank.h>
23
24#include "vc4_drv.h"
25#include "vc4_regs.h"
26
27#define HVS_NUM_CHANNELS 3
28
29struct vc4_ctm_state {
30	struct drm_private_state base;
31	struct drm_color_ctm *ctm;
32	int fifo;
33};
34
35static struct vc4_ctm_state *to_vc4_ctm_state(struct drm_private_state *priv)
36{
37	return container_of(priv, struct vc4_ctm_state, base);
38}
39
40struct vc4_hvs_state {
41	struct drm_private_state base;
42	unsigned int unassigned_channels;
43};
44
45static struct vc4_hvs_state *
46to_vc4_hvs_state(struct drm_private_state *priv)
47{
48	return container_of(priv, struct vc4_hvs_state, base);
49}
50
51struct vc4_load_tracker_state {
52	struct drm_private_state base;
53	u64 hvs_load;
54	u64 membus_load;
55};
56
57static struct vc4_load_tracker_state *
58to_vc4_load_tracker_state(struct drm_private_state *priv)
59{
60	return container_of(priv, struct vc4_load_tracker_state, base);
61}
62
63static struct vc4_ctm_state *vc4_get_ctm_state(struct drm_atomic_state *state,
64					       struct drm_private_obj *manager)
65{
66	struct drm_device *dev = state->dev;
67	struct vc4_dev *vc4 = to_vc4_dev(dev);
68	struct drm_private_state *priv_state;
69	int ret;
70
71	ret = drm_modeset_lock(&vc4->ctm_state_lock, state->acquire_ctx);
72	if (ret)
73		return ERR_PTR(ret);
74
75	priv_state = drm_atomic_get_private_obj_state(state, manager);
76	if (IS_ERR(priv_state))
77		return ERR_CAST(priv_state);
78
79	return to_vc4_ctm_state(priv_state);
80}
81
82static struct drm_private_state *
83vc4_ctm_duplicate_state(struct drm_private_obj *obj)
84{
85	struct vc4_ctm_state *state;
86
87	state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
88	if (!state)
89		return NULL;
90
91	__drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
92
93	return &state->base;
94}
95
96static void vc4_ctm_destroy_state(struct drm_private_obj *obj,
97				  struct drm_private_state *state)
98{
99	struct vc4_ctm_state *ctm_state = to_vc4_ctm_state(state);
100
101	kfree(ctm_state);
102}
103
104static const struct drm_private_state_funcs vc4_ctm_state_funcs = {
105	.atomic_duplicate_state = vc4_ctm_duplicate_state,
106	.atomic_destroy_state = vc4_ctm_destroy_state,
107};
108
109static void vc4_ctm_obj_fini(struct drm_device *dev, void *unused)
110{
111	struct vc4_dev *vc4 = to_vc4_dev(dev);
112
113	drm_atomic_private_obj_fini(&vc4->ctm_manager);
114}
115
116static int vc4_ctm_obj_init(struct vc4_dev *vc4)
117{
118	struct vc4_ctm_state *ctm_state;
119
120	drm_modeset_lock_init(&vc4->ctm_state_lock);
121
122	ctm_state = kzalloc(sizeof(*ctm_state), GFP_KERNEL);
123	if (!ctm_state)
124		return -ENOMEM;
125
126	drm_atomic_private_obj_init(&vc4->base, &vc4->ctm_manager, &ctm_state->base,
127				    &vc4_ctm_state_funcs);
128
129	return drmm_add_action_or_reset(&vc4->base, vc4_ctm_obj_fini, NULL);
130}
131
132/* Converts a DRM S31.32 value to the HW S0.9 format. */
133static u16 vc4_ctm_s31_32_to_s0_9(u64 in)
134{
135	u16 r;
136
137	/* Sign bit. */
138	r = in & BIT_ULL(63) ? BIT(9) : 0;
139
140	if ((in & GENMASK_ULL(62, 32)) > 0) {
141		/* We have zero integer bits so we can only saturate here. */
142		r |= GENMASK(8, 0);
143	} else {
144		/* Otherwise take the 9 most important fractional bits. */
145		r |= (in >> 23) & GENMASK(8, 0);
146	}
147
148	return r;
149}
150
151static void
152vc4_ctm_commit(struct vc4_dev *vc4, struct drm_atomic_state *state)
153{
154	struct vc4_ctm_state *ctm_state = to_vc4_ctm_state(vc4->ctm_manager.state);
155	struct drm_color_ctm *ctm = ctm_state->ctm;
156
157	if (ctm_state->fifo) {
158		HVS_WRITE(SCALER_OLEDCOEF2,
159			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[0]),
160					SCALER_OLEDCOEF2_R_TO_R) |
161			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[3]),
162					SCALER_OLEDCOEF2_R_TO_G) |
163			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[6]),
164					SCALER_OLEDCOEF2_R_TO_B));
165		HVS_WRITE(SCALER_OLEDCOEF1,
166			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[1]),
167					SCALER_OLEDCOEF1_G_TO_R) |
168			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[4]),
169					SCALER_OLEDCOEF1_G_TO_G) |
170			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[7]),
171					SCALER_OLEDCOEF1_G_TO_B));
172		HVS_WRITE(SCALER_OLEDCOEF0,
173			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[2]),
174					SCALER_OLEDCOEF0_B_TO_R) |
175			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[5]),
176					SCALER_OLEDCOEF0_B_TO_G) |
177			  VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[8]),
178					SCALER_OLEDCOEF0_B_TO_B));
179	}
180
181	HVS_WRITE(SCALER_OLEDOFFS,
182		  VC4_SET_FIELD(ctm_state->fifo, SCALER_OLEDOFFS_DISPFIFO));
183}
184
185static struct vc4_hvs_state *
186vc4_hvs_get_global_state(struct drm_atomic_state *state)
187{
188	struct vc4_dev *vc4 = to_vc4_dev(state->dev);
189	struct drm_private_state *priv_state;
190
191	priv_state = drm_atomic_get_private_obj_state(state, &vc4->hvs_channels);
192	if (IS_ERR(priv_state))
193		return ERR_CAST(priv_state);
194
195	return to_vc4_hvs_state(priv_state);
196}
197
198static void vc4_hvs_pv_muxing_commit(struct vc4_dev *vc4,
199				     struct drm_atomic_state *state)
200{
201	struct drm_crtc_state *crtc_state;
202	struct drm_crtc *crtc;
203	unsigned int i;
204
205	for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
206		struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc_state);
207		u32 dispctrl;
208		u32 dsp3_mux;
209
210		if (!crtc_state->active)
211			continue;
212
213		if (vc4_state->assigned_channel != 2)
214			continue;
215
216		/*
217		 * SCALER_DISPCTRL_DSP3 = X, where X < 2 means 'connect DSP3 to
218		 * FIFO X'.
219		 * SCALER_DISPCTRL_DSP3 = 3 means 'disable DSP 3'.
220		 *
221		 * DSP3 is connected to FIFO2 unless the transposer is
222		 * enabled. In this case, FIFO 2 is directly accessed by the
223		 * TXP IP, and we need to disable the FIFO2 -> pixelvalve1
224		 * route.
225		 */
226		if (vc4_state->feed_txp)
227			dsp3_mux = VC4_SET_FIELD(3, SCALER_DISPCTRL_DSP3_MUX);
228		else
229			dsp3_mux = VC4_SET_FIELD(2, SCALER_DISPCTRL_DSP3_MUX);
230
231		dispctrl = HVS_READ(SCALER_DISPCTRL) &
232			   ~SCALER_DISPCTRL_DSP3_MUX_MASK;
233		HVS_WRITE(SCALER_DISPCTRL, dispctrl | dsp3_mux);
234	}
235}
236
237static void vc5_hvs_pv_muxing_commit(struct vc4_dev *vc4,
238				     struct drm_atomic_state *state)
239{
240	struct drm_crtc_state *crtc_state;
241	struct drm_crtc *crtc;
242	unsigned char mux;
243	unsigned int i;
244	u32 reg;
245
246	for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
247		struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc_state);
248		struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
249
250		if (!vc4_state->update_muxing)
251			continue;
252
253		switch (vc4_crtc->data->hvs_output) {
254		case 2:
255			mux = (vc4_state->assigned_channel == 2) ? 0 : 1;
256			reg = HVS_READ(SCALER_DISPECTRL);
257			HVS_WRITE(SCALER_DISPECTRL,
258				  (reg & ~SCALER_DISPECTRL_DSP2_MUX_MASK) |
259				  VC4_SET_FIELD(mux, SCALER_DISPECTRL_DSP2_MUX));
260			break;
261
262		case 3:
263			if (vc4_state->assigned_channel == VC4_HVS_CHANNEL_DISABLED)
264				mux = 3;
265			else
266				mux = vc4_state->assigned_channel;
267
268			reg = HVS_READ(SCALER_DISPCTRL);
269			HVS_WRITE(SCALER_DISPCTRL,
270				  (reg & ~SCALER_DISPCTRL_DSP3_MUX_MASK) |
271				  VC4_SET_FIELD(mux, SCALER_DISPCTRL_DSP3_MUX));
272			break;
273
274		case 4:
275			if (vc4_state->assigned_channel == VC4_HVS_CHANNEL_DISABLED)
276				mux = 3;
277			else
278				mux = vc4_state->assigned_channel;
279
280			reg = HVS_READ(SCALER_DISPEOLN);
281			HVS_WRITE(SCALER_DISPEOLN,
282				  (reg & ~SCALER_DISPEOLN_DSP4_MUX_MASK) |
283				  VC4_SET_FIELD(mux, SCALER_DISPEOLN_DSP4_MUX));
284
285			break;
286
287		case 5:
288			if (vc4_state->assigned_channel == VC4_HVS_CHANNEL_DISABLED)
289				mux = 3;
290			else
291				mux = vc4_state->assigned_channel;
292
293			reg = HVS_READ(SCALER_DISPDITHER);
294			HVS_WRITE(SCALER_DISPDITHER,
295				  (reg & ~SCALER_DISPDITHER_DSP5_MUX_MASK) |
296				  VC4_SET_FIELD(mux, SCALER_DISPDITHER_DSP5_MUX));
297			break;
298
299		default:
300			break;
301		}
302	}
303}
304
305static void
306vc4_atomic_complete_commit(struct drm_atomic_state *state)
307{
308	struct drm_device *dev = state->dev;
309	struct vc4_dev *vc4 = to_vc4_dev(dev);
310	struct vc4_hvs *hvs = vc4->hvs;
311	struct drm_crtc_state *new_crtc_state;
312	struct drm_crtc *crtc;
313	int i;
314
315	for_each_new_crtc_in_state(state, crtc, new_crtc_state, i) {
316		struct vc4_crtc_state *vc4_crtc_state;
317
318		if (!new_crtc_state->commit)
319			continue;
320
321		vc4_crtc_state = to_vc4_crtc_state(new_crtc_state);
322		vc4_hvs_mask_underrun(dev, vc4_crtc_state->assigned_channel);
323	}
324
325	if (vc4->hvs->hvs5)
326		clk_set_min_rate(hvs->core_clk, 500000000);
327
328	drm_atomic_helper_wait_for_fences(dev, state, false);
329
330	drm_atomic_helper_wait_for_dependencies(state);
331
332	drm_atomic_helper_commit_modeset_disables(dev, state);
333
334	vc4_ctm_commit(vc4, state);
335
336	if (vc4->hvs->hvs5)
337		vc5_hvs_pv_muxing_commit(vc4, state);
338	else
339		vc4_hvs_pv_muxing_commit(vc4, state);
340
341	drm_atomic_helper_commit_planes(dev, state, 0);
342
343	drm_atomic_helper_commit_modeset_enables(dev, state);
344
345	drm_atomic_helper_fake_vblank(state);
346
347	drm_atomic_helper_commit_hw_done(state);
348
349	drm_atomic_helper_wait_for_flip_done(dev, state);
350
351	drm_atomic_helper_cleanup_planes(dev, state);
352
353	drm_atomic_helper_commit_cleanup_done(state);
354
355	if (vc4->hvs->hvs5)
356		clk_set_min_rate(hvs->core_clk, 0);
357
358	drm_atomic_state_put(state);
359
360	up(&vc4->async_modeset);
361}
362
363static void commit_work(struct work_struct *work)
364{
365	struct drm_atomic_state *state = container_of(work,
366						      struct drm_atomic_state,
367						      commit_work);
368	vc4_atomic_complete_commit(state);
369}
370
371/**
372 * vc4_atomic_commit - commit validated state object
373 * @dev: DRM device
374 * @state: the driver state object
375 * @nonblock: nonblocking commit
376 *
377 * This function commits a with drm_atomic_helper_check() pre-validated state
378 * object. This can still fail when e.g. the framebuffer reservation fails. For
379 * now this doesn't implement asynchronous commits.
380 *
381 * RETURNS
382 * Zero for success or -errno.
383 */
384static int vc4_atomic_commit(struct drm_device *dev,
385			     struct drm_atomic_state *state,
386			     bool nonblock)
387{
388	struct vc4_dev *vc4 = to_vc4_dev(dev);
389	int ret;
390
391	if (state->async_update) {
392		ret = down_interruptible(&vc4->async_modeset);
393		if (ret)
394			return ret;
395
396		ret = drm_atomic_helper_prepare_planes(dev, state);
397		if (ret) {
398			up(&vc4->async_modeset);
399			return ret;
400		}
401
402		drm_atomic_helper_async_commit(dev, state);
403
404		drm_atomic_helper_cleanup_planes(dev, state);
405
406		up(&vc4->async_modeset);
407
408		return 0;
409	}
410
411	/* We know for sure we don't want an async update here. Set
412	 * state->legacy_cursor_update to false to prevent
413	 * drm_atomic_helper_setup_commit() from auto-completing
414	 * commit->flip_done.
415	 */
416	state->legacy_cursor_update = false;
417	ret = drm_atomic_helper_setup_commit(state, nonblock);
418	if (ret)
419		return ret;
420
421	INIT_WORK(&state->commit_work, commit_work);
422
423	ret = down_interruptible(&vc4->async_modeset);
424	if (ret)
425		return ret;
426
427	ret = drm_atomic_helper_prepare_planes(dev, state);
428	if (ret) {
429		up(&vc4->async_modeset);
430		return ret;
431	}
432
433	if (!nonblock) {
434		ret = drm_atomic_helper_wait_for_fences(dev, state, true);
435		if (ret) {
436			drm_atomic_helper_cleanup_planes(dev, state);
437			up(&vc4->async_modeset);
438			return ret;
439		}
440	}
441
442	/*
443	 * This is the point of no return - everything below never fails except
444	 * when the hw goes bonghits. Which means we can commit the new state on
445	 * the software side now.
446	 */
447
448	BUG_ON(drm_atomic_helper_swap_state(state, false) < 0);
449
450	/*
451	 * Everything below can be run asynchronously without the need to grab
452	 * any modeset locks at all under one condition: It must be guaranteed
453	 * that the asynchronous work has either been cancelled (if the driver
454	 * supports it, which at least requires that the framebuffers get
455	 * cleaned up with drm_atomic_helper_cleanup_planes()) or completed
456	 * before the new state gets committed on the software side with
457	 * drm_atomic_helper_swap_state().
458	 *
459	 * This scheme allows new atomic state updates to be prepared and
460	 * checked in parallel to the asynchronous completion of the previous
461	 * update. Which is important since compositors need to figure out the
462	 * composition of the next frame right after having submitted the
463	 * current layout.
464	 */
465
466	drm_atomic_state_get(state);
467	if (nonblock)
468		queue_work(system_unbound_wq, &state->commit_work);
469	else
470		vc4_atomic_complete_commit(state);
471
472	return 0;
473}
474
475static struct drm_framebuffer *vc4_fb_create(struct drm_device *dev,
476					     struct drm_file *file_priv,
477					     const struct drm_mode_fb_cmd2 *mode_cmd)
478{
479	struct drm_mode_fb_cmd2 mode_cmd_local;
480
481	/* If the user didn't specify a modifier, use the
482	 * vc4_set_tiling_ioctl() state for the BO.
483	 */
484	if (!(mode_cmd->flags & DRM_MODE_FB_MODIFIERS)) {
485		struct drm_gem_object *gem_obj;
486		struct vc4_bo *bo;
487
488		gem_obj = drm_gem_object_lookup(file_priv,
489						mode_cmd->handles[0]);
490		if (!gem_obj) {
491			DRM_DEBUG("Failed to look up GEM BO %d\n",
492				  mode_cmd->handles[0]);
493			return ERR_PTR(-ENOENT);
494		}
495		bo = to_vc4_bo(gem_obj);
496
497		mode_cmd_local = *mode_cmd;
498
499		if (bo->t_format) {
500			mode_cmd_local.modifier[0] =
501				DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
502		} else {
503			mode_cmd_local.modifier[0] = DRM_FORMAT_MOD_NONE;
504		}
505
506		drm_gem_object_put(gem_obj);
507
508		mode_cmd = &mode_cmd_local;
509	}
510
511	return drm_gem_fb_create(dev, file_priv, mode_cmd);
512}
513
514/* Our CTM has some peculiar limitations: we can only enable it for one CRTC
515 * at a time and the HW only supports S0.9 scalars. To account for the latter,
516 * we don't allow userland to set a CTM that we have no hope of approximating.
517 */
518static int
519vc4_ctm_atomic_check(struct drm_device *dev, struct drm_atomic_state *state)
520{
521	struct vc4_dev *vc4 = to_vc4_dev(dev);
522	struct vc4_ctm_state *ctm_state = NULL;
523	struct drm_crtc *crtc;
524	struct drm_crtc_state *old_crtc_state, *new_crtc_state;
525	struct drm_color_ctm *ctm;
526	int i;
527
528	for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
529		/* CTM is being disabled. */
530		if (!new_crtc_state->ctm && old_crtc_state->ctm) {
531			ctm_state = vc4_get_ctm_state(state, &vc4->ctm_manager);
532			if (IS_ERR(ctm_state))
533				return PTR_ERR(ctm_state);
534			ctm_state->fifo = 0;
535		}
536	}
537
538	for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
539		if (new_crtc_state->ctm == old_crtc_state->ctm)
540			continue;
541
542		if (!ctm_state) {
543			ctm_state = vc4_get_ctm_state(state, &vc4->ctm_manager);
544			if (IS_ERR(ctm_state))
545				return PTR_ERR(ctm_state);
546		}
547
548		/* CTM is being enabled or the matrix changed. */
549		if (new_crtc_state->ctm) {
550			struct vc4_crtc_state *vc4_crtc_state =
551				to_vc4_crtc_state(new_crtc_state);
552
553			/* fifo is 1-based since 0 disables CTM. */
554			int fifo = vc4_crtc_state->assigned_channel + 1;
555
556			/* Check userland isn't trying to turn on CTM for more
557			 * than one CRTC at a time.
558			 */
559			if (ctm_state->fifo && ctm_state->fifo != fifo) {
560				DRM_DEBUG_DRIVER("Too many CTM configured\n");
561				return -EINVAL;
562			}
563
564			/* Check we can approximate the specified CTM.
565			 * We disallow scalars |c| > 1.0 since the HW has
566			 * no integer bits.
567			 */
568			ctm = new_crtc_state->ctm->data;
569			for (i = 0; i < ARRAY_SIZE(ctm->matrix); i++) {
570				u64 val = ctm->matrix[i];
571
572				val &= ~BIT_ULL(63);
573				if (val > BIT_ULL(32))
574					return -EINVAL;
575			}
576
577			ctm_state->fifo = fifo;
578			ctm_state->ctm = ctm;
579		}
580	}
581
582	return 0;
583}
584
585static int vc4_load_tracker_atomic_check(struct drm_atomic_state *state)
586{
587	struct drm_plane_state *old_plane_state, *new_plane_state;
588	struct vc4_dev *vc4 = to_vc4_dev(state->dev);
589	struct vc4_load_tracker_state *load_state;
590	struct drm_private_state *priv_state;
591	struct drm_plane *plane;
592	int i;
593
594	if (!vc4->load_tracker_available)
595		return 0;
596
597	priv_state = drm_atomic_get_private_obj_state(state,
598						      &vc4->load_tracker);
599	if (IS_ERR(priv_state))
600		return PTR_ERR(priv_state);
601
602	load_state = to_vc4_load_tracker_state(priv_state);
603	for_each_oldnew_plane_in_state(state, plane, old_plane_state,
604				       new_plane_state, i) {
605		struct vc4_plane_state *vc4_plane_state;
606
607		if (old_plane_state->fb && old_plane_state->crtc) {
608			vc4_plane_state = to_vc4_plane_state(old_plane_state);
609			load_state->membus_load -= vc4_plane_state->membus_load;
610			load_state->hvs_load -= vc4_plane_state->hvs_load;
611		}
612
613		if (new_plane_state->fb && new_plane_state->crtc) {
614			vc4_plane_state = to_vc4_plane_state(new_plane_state);
615			load_state->membus_load += vc4_plane_state->membus_load;
616			load_state->hvs_load += vc4_plane_state->hvs_load;
617		}
618	}
619
620	/* Don't check the load when the tracker is disabled. */
621	if (!vc4->load_tracker_enabled)
622		return 0;
623
624	/* The absolute limit is 2Gbyte/sec, but let's take a margin to let
625	 * the system work when other blocks are accessing the memory.
626	 */
627	if (load_state->membus_load > SZ_1G + SZ_512M)
628		return -ENOSPC;
629
630	/* HVS clock is supposed to run @ 250Mhz, let's take a margin and
631	 * consider the maximum number of cycles is 240M.
632	 */
633	if (load_state->hvs_load > 240000000ULL)
634		return -ENOSPC;
635
636	return 0;
637}
638
639static struct drm_private_state *
640vc4_load_tracker_duplicate_state(struct drm_private_obj *obj)
641{
642	struct vc4_load_tracker_state *state;
643
644	state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
645	if (!state)
646		return NULL;
647
648	__drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
649
650	return &state->base;
651}
652
653static void vc4_load_tracker_destroy_state(struct drm_private_obj *obj,
654					   struct drm_private_state *state)
655{
656	struct vc4_load_tracker_state *load_state;
657
658	load_state = to_vc4_load_tracker_state(state);
659	kfree(load_state);
660}
661
662static const struct drm_private_state_funcs vc4_load_tracker_state_funcs = {
663	.atomic_duplicate_state = vc4_load_tracker_duplicate_state,
664	.atomic_destroy_state = vc4_load_tracker_destroy_state,
665};
666
667static void vc4_load_tracker_obj_fini(struct drm_device *dev, void *unused)
668{
669	struct vc4_dev *vc4 = to_vc4_dev(dev);
670
671	if (!vc4->load_tracker_available)
672		return;
673
674	drm_atomic_private_obj_fini(&vc4->load_tracker);
675}
676
677static int vc4_load_tracker_obj_init(struct vc4_dev *vc4)
678{
679	struct vc4_load_tracker_state *load_state;
680
681	if (!vc4->load_tracker_available)
682		return 0;
683
684	load_state = kzalloc(sizeof(*load_state), GFP_KERNEL);
685	if (!load_state)
686		return -ENOMEM;
687
688	drm_atomic_private_obj_init(&vc4->base, &vc4->load_tracker,
689				    &load_state->base,
690				    &vc4_load_tracker_state_funcs);
691
692	return drmm_add_action_or_reset(&vc4->base, vc4_load_tracker_obj_fini, NULL);
693}
694
695static struct drm_private_state *
696vc4_hvs_channels_duplicate_state(struct drm_private_obj *obj)
697{
698	struct vc4_hvs_state *old_state = to_vc4_hvs_state(obj->state);
699	struct vc4_hvs_state *state;
700
701	state = kzalloc(sizeof(*state), GFP_KERNEL);
702	if (!state)
703		return NULL;
704
705	__drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
706
707	state->unassigned_channels = old_state->unassigned_channels;
708
709	return &state->base;
710}
711
712static void vc4_hvs_channels_destroy_state(struct drm_private_obj *obj,
713					   struct drm_private_state *state)
714{
715	struct vc4_hvs_state *hvs_state = to_vc4_hvs_state(state);
716
717	kfree(hvs_state);
718}
719
720static const struct drm_private_state_funcs vc4_hvs_state_funcs = {
721	.atomic_duplicate_state = vc4_hvs_channels_duplicate_state,
722	.atomic_destroy_state = vc4_hvs_channels_destroy_state,
723};
724
725static void vc4_hvs_channels_obj_fini(struct drm_device *dev, void *unused)
726{
727	struct vc4_dev *vc4 = to_vc4_dev(dev);
728
729	drm_atomic_private_obj_fini(&vc4->hvs_channels);
730}
731
732static int vc4_hvs_channels_obj_init(struct vc4_dev *vc4)
733{
734	struct vc4_hvs_state *state;
735
736	state = kzalloc(sizeof(*state), GFP_KERNEL);
737	if (!state)
738		return -ENOMEM;
739
740	state->unassigned_channels = GENMASK(HVS_NUM_CHANNELS - 1, 0);
741	drm_atomic_private_obj_init(&vc4->base, &vc4->hvs_channels,
742				    &state->base,
743				    &vc4_hvs_state_funcs);
744
745	return drmm_add_action_or_reset(&vc4->base, vc4_hvs_channels_obj_fini, NULL);
746}
747
748/*
749 * The BCM2711 HVS has up to 7 outputs connected to the pixelvalves and
750 * the TXP (and therefore all the CRTCs found on that platform).
751 *
752 * The naive (and our initial) implementation would just iterate over
753 * all the active CRTCs, try to find a suitable FIFO, and then remove it
754 * from the pool of available FIFOs. However, there are a few corner
755 * cases that need to be considered:
756 *
757 * - When running in a dual-display setup (so with two CRTCs involved),
758 *   we can update the state of a single CRTC (for example by changing
759 *   its mode using xrandr under X11) without affecting the other. In
760 *   this case, the other CRTC wouldn't be in the state at all, so we
761 *   need to consider all the running CRTCs in the DRM device to assign
762 *   a FIFO, not just the one in the state.
763 *
764 * - To fix the above, we can't use drm_atomic_get_crtc_state on all
765 *   enabled CRTCs to pull their CRTC state into the global state, since
766 *   a page flip would start considering their vblank to complete. Since
767 *   we don't have a guarantee that they are actually active, that
768 *   vblank might never happen, and shouldn't even be considered if we
769 *   want to do a page flip on a single CRTC. That can be tested by
770 *   doing a modetest -v first on HDMI1 and then on HDMI0.
771 *
772 * - Since we need the pixelvalve to be disabled and enabled back when
773 *   the FIFO is changed, we should keep the FIFO assigned for as long
774 *   as the CRTC is enabled, only considering it free again once that
775 *   CRTC has been disabled. This can be tested by booting X11 on a
776 *   single display, and changing the resolution down and then back up.
777 */
778static int vc4_pv_muxing_atomic_check(struct drm_device *dev,
779				      struct drm_atomic_state *state)
780{
781	struct vc4_hvs_state *hvs_new_state;
782	struct drm_crtc_state *old_crtc_state, *new_crtc_state;
783	struct drm_crtc *crtc;
784	unsigned int i;
785
786	hvs_new_state = vc4_hvs_get_global_state(state);
787	if (!hvs_new_state)
788		return -EINVAL;
789
790	for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
791		struct vc4_crtc_state *old_vc4_crtc_state =
792			to_vc4_crtc_state(old_crtc_state);
793		struct vc4_crtc_state *new_vc4_crtc_state =
794			to_vc4_crtc_state(new_crtc_state);
795		struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
796		unsigned int matching_channels;
797
798		/* Nothing to do here, let's skip it */
799		if (old_crtc_state->enable == new_crtc_state->enable)
800			continue;
801
802		/* Muxing will need to be modified, mark it as such */
803		new_vc4_crtc_state->update_muxing = true;
804
805		/* If we're disabling our CRTC, we put back our channel */
806		if (!new_crtc_state->enable) {
807			hvs_new_state->unassigned_channels |= BIT(old_vc4_crtc_state->assigned_channel);
808			new_vc4_crtc_state->assigned_channel = VC4_HVS_CHANNEL_DISABLED;
809			continue;
810		}
811
812		/*
813		 * The problem we have to solve here is that we have
814		 * up to 7 encoders, connected to up to 6 CRTCs.
815		 *
816		 * Those CRTCs, depending on the instance, can be
817		 * routed to 1, 2 or 3 HVS FIFOs, and we need to set
818		 * the change the muxing between FIFOs and outputs in
819		 * the HVS accordingly.
820		 *
821		 * It would be pretty hard to come up with an
822		 * algorithm that would generically solve
823		 * this. However, the current routing trees we support
824		 * allow us to simplify a bit the problem.
825		 *
826		 * Indeed, with the current supported layouts, if we
827		 * try to assign in the ascending crtc index order the
828		 * FIFOs, we can't fall into the situation where an
829		 * earlier CRTC that had multiple routes is assigned
830		 * one that was the only option for a later CRTC.
831		 *
832		 * If the layout changes and doesn't give us that in
833		 * the future, we will need to have something smarter,
834		 * but it works so far.
835		 */
836		matching_channels = hvs_new_state->unassigned_channels & vc4_crtc->data->hvs_available_channels;
837		if (matching_channels) {
838			unsigned int channel = ffs(matching_channels) - 1;
839
840			new_vc4_crtc_state->assigned_channel = channel;
841			hvs_new_state->unassigned_channels &= ~BIT(channel);
842		} else {
843			return -EINVAL;
844		}
845	}
846
847	return 0;
848}
849
850static int
851vc4_atomic_check(struct drm_device *dev, struct drm_atomic_state *state)
852{
853	int ret;
854
855	ret = vc4_pv_muxing_atomic_check(dev, state);
856	if (ret)
857		return ret;
858
859	ret = vc4_ctm_atomic_check(dev, state);
860	if (ret < 0)
861		return ret;
862
863	ret = drm_atomic_helper_check(dev, state);
864	if (ret)
865		return ret;
866
867	return vc4_load_tracker_atomic_check(state);
868}
869
870static const struct drm_mode_config_funcs vc4_mode_funcs = {
871	.atomic_check = vc4_atomic_check,
872	.atomic_commit = vc4_atomic_commit,
873	.fb_create = vc4_fb_create,
874};
875
876int vc4_kms_load(struct drm_device *dev)
877{
878	struct vc4_dev *vc4 = to_vc4_dev(dev);
879	bool is_vc5 = of_device_is_compatible(dev->dev->of_node,
880					      "brcm,bcm2711-vc5");
881	int ret;
882
883	if (!is_vc5) {
884		vc4->load_tracker_available = true;
885
886		/* Start with the load tracker enabled. Can be
887		 * disabled through the debugfs load_tracker file.
888		 */
889		vc4->load_tracker_enabled = true;
890	}
891
892	sema_init(&vc4->async_modeset, 1);
893
894	/* Set support for vblank irq fast disable, before drm_vblank_init() */
895	dev->vblank_disable_immediate = true;
896
897	dev->irq_enabled = true;
898	ret = drm_vblank_init(dev, dev->mode_config.num_crtc);
899	if (ret < 0) {
900		dev_err(dev->dev, "failed to initialize vblank\n");
901		return ret;
902	}
903
904	if (is_vc5) {
905		dev->mode_config.max_width = 7680;
906		dev->mode_config.max_height = 7680;
907	} else {
908		dev->mode_config.max_width = 2048;
909		dev->mode_config.max_height = 2048;
910	}
911
912	dev->mode_config.funcs = &vc4_mode_funcs;
913	dev->mode_config.preferred_depth = 24;
914	dev->mode_config.async_page_flip = true;
915	dev->mode_config.allow_fb_modifiers = true;
916
917	ret = vc4_ctm_obj_init(vc4);
918	if (ret)
919		return ret;
920
921	ret = vc4_load_tracker_obj_init(vc4);
922	if (ret)
923		return ret;
924
925	ret = vc4_hvs_channels_obj_init(vc4);
926	if (ret)
927		return ret;
928
929	drm_mode_config_reset(dev);
930
931	drm_kms_helper_poll_init(dev);
932
933	return 0;
934}
935