1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (C) 2015 Broadcom
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).
14 #include <linux/clk.h>
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>
27 struct vc4_ctm_state {
28 struct drm_private_state base;
29 struct drm_color_ctm *ctm;
33 static struct vc4_ctm_state *to_vc4_ctm_state(struct drm_private_state *priv)
35 return container_of(priv, struct vc4_ctm_state, base);
38 struct vc4_load_tracker_state {
39 struct drm_private_state base;
44 static struct vc4_load_tracker_state *
45 to_vc4_load_tracker_state(struct drm_private_state *priv)
47 return container_of(priv, struct vc4_load_tracker_state, base);
50 static struct vc4_ctm_state *vc4_get_ctm_state(struct drm_atomic_state *state,
51 struct drm_private_obj *manager)
53 struct drm_device *dev = state->dev;
54 struct vc4_dev *vc4 = dev->dev_private;
55 struct drm_private_state *priv_state;
58 ret = drm_modeset_lock(&vc4->ctm_state_lock, state->acquire_ctx);
62 priv_state = drm_atomic_get_private_obj_state(state, manager);
63 if (IS_ERR(priv_state))
64 return ERR_CAST(priv_state);
66 return to_vc4_ctm_state(priv_state);
69 static struct drm_private_state *
70 vc4_ctm_duplicate_state(struct drm_private_obj *obj)
72 struct vc4_ctm_state *state;
74 state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
78 __drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
83 static void vc4_ctm_destroy_state(struct drm_private_obj *obj,
84 struct drm_private_state *state)
86 struct vc4_ctm_state *ctm_state = to_vc4_ctm_state(state);
91 static const struct drm_private_state_funcs vc4_ctm_state_funcs = {
92 .atomic_duplicate_state = vc4_ctm_duplicate_state,
93 .atomic_destroy_state = vc4_ctm_destroy_state,
96 /* Converts a DRM S31.32 value to the HW S0.9 format. */
97 static u16 vc4_ctm_s31_32_to_s0_9(u64 in)
102 r = in & BIT_ULL(63) ? BIT(9) : 0;
104 if ((in & GENMASK_ULL(62, 32)) > 0) {
105 /* We have zero integer bits so we can only saturate here. */
108 /* Otherwise take the 9 most important fractional bits. */
109 r |= (in >> 23) & GENMASK(8, 0);
116 vc4_ctm_commit(struct vc4_dev *vc4, struct drm_atomic_state *state)
118 struct vc4_ctm_state *ctm_state = to_vc4_ctm_state(vc4->ctm_manager.state);
119 struct drm_color_ctm *ctm = ctm_state->ctm;
121 if (ctm_state->fifo) {
122 HVS_WRITE(SCALER_OLEDCOEF2,
123 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[0]),
124 SCALER_OLEDCOEF2_R_TO_R) |
125 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[3]),
126 SCALER_OLEDCOEF2_R_TO_G) |
127 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[6]),
128 SCALER_OLEDCOEF2_R_TO_B));
129 HVS_WRITE(SCALER_OLEDCOEF1,
130 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[1]),
131 SCALER_OLEDCOEF1_G_TO_R) |
132 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[4]),
133 SCALER_OLEDCOEF1_G_TO_G) |
134 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[7]),
135 SCALER_OLEDCOEF1_G_TO_B));
136 HVS_WRITE(SCALER_OLEDCOEF0,
137 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[2]),
138 SCALER_OLEDCOEF0_B_TO_R) |
139 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[5]),
140 SCALER_OLEDCOEF0_B_TO_G) |
141 VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[8]),
142 SCALER_OLEDCOEF0_B_TO_B));
145 HVS_WRITE(SCALER_OLEDOFFS,
146 VC4_SET_FIELD(ctm_state->fifo, SCALER_OLEDOFFS_DISPFIFO));
149 static void vc4_hvs_pv_muxing_commit(struct vc4_dev *vc4,
150 struct drm_atomic_state *state)
152 struct drm_crtc_state *crtc_state;
153 struct drm_crtc *crtc;
156 for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
157 struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc_state);
161 if (!crtc_state->active)
164 if (vc4_state->assigned_channel != 2)
168 * SCALER_DISPCTRL_DSP3 = X, where X < 2 means 'connect DSP3 to
170 * SCALER_DISPCTRL_DSP3 = 3 means 'disable DSP 3'.
172 * DSP3 is connected to FIFO2 unless the transposer is
173 * enabled. In this case, FIFO 2 is directly accessed by the
174 * TXP IP, and we need to disable the FIFO2 -> pixelvalve1
177 if (vc4_state->feed_txp)
178 dsp3_mux = VC4_SET_FIELD(3, SCALER_DISPCTRL_DSP3_MUX);
180 dsp3_mux = VC4_SET_FIELD(2, SCALER_DISPCTRL_DSP3_MUX);
182 dispctrl = HVS_READ(SCALER_DISPCTRL) &
183 ~SCALER_DISPCTRL_DSP3_MUX_MASK;
184 HVS_WRITE(SCALER_DISPCTRL, dispctrl | dsp3_mux);
188 static void vc5_hvs_pv_muxing_commit(struct vc4_dev *vc4,
189 struct drm_atomic_state *state)
191 struct drm_crtc_state *crtc_state;
192 struct drm_crtc *crtc;
193 unsigned char dsp2_mux = 0;
194 unsigned char dsp3_mux = 3;
195 unsigned char dsp4_mux = 3;
196 unsigned char dsp5_mux = 3;
200 for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
201 struct vc4_crtc_state *vc4_state = to_vc4_crtc_state(crtc_state);
202 struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
204 if (!crtc_state->active)
207 switch (vc4_crtc->data->hvs_output) {
209 dsp2_mux = (vc4_state->assigned_channel == 2) ? 0 : 1;
213 dsp3_mux = vc4_state->assigned_channel;
217 dsp4_mux = vc4_state->assigned_channel;
221 dsp5_mux = vc4_state->assigned_channel;
229 reg = HVS_READ(SCALER_DISPECTRL);
230 HVS_WRITE(SCALER_DISPECTRL,
231 (reg & ~SCALER_DISPECTRL_DSP2_MUX_MASK) |
232 VC4_SET_FIELD(dsp2_mux, SCALER_DISPECTRL_DSP2_MUX));
234 reg = HVS_READ(SCALER_DISPCTRL);
235 HVS_WRITE(SCALER_DISPCTRL,
236 (reg & ~SCALER_DISPCTRL_DSP3_MUX_MASK) |
237 VC4_SET_FIELD(dsp3_mux, SCALER_DISPCTRL_DSP3_MUX));
239 reg = HVS_READ(SCALER_DISPEOLN);
240 HVS_WRITE(SCALER_DISPEOLN,
241 (reg & ~SCALER_DISPEOLN_DSP4_MUX_MASK) |
242 VC4_SET_FIELD(dsp4_mux, SCALER_DISPEOLN_DSP4_MUX));
244 reg = HVS_READ(SCALER_DISPDITHER);
245 HVS_WRITE(SCALER_DISPDITHER,
246 (reg & ~SCALER_DISPDITHER_DSP5_MUX_MASK) |
247 VC4_SET_FIELD(dsp5_mux, SCALER_DISPDITHER_DSP5_MUX));
251 vc4_atomic_complete_commit(struct drm_atomic_state *state)
253 struct drm_device *dev = state->dev;
254 struct vc4_dev *vc4 = to_vc4_dev(dev);
255 struct vc4_hvs *hvs = vc4->hvs;
256 struct drm_crtc_state *new_crtc_state;
257 struct drm_crtc *crtc;
260 for_each_new_crtc_in_state(state, crtc, new_crtc_state, i) {
261 struct vc4_crtc_state *vc4_crtc_state;
263 if (!new_crtc_state->commit)
266 vc4_crtc_state = to_vc4_crtc_state(new_crtc_state);
267 vc4_hvs_mask_underrun(dev, vc4_crtc_state->assigned_channel);
271 clk_set_min_rate(hvs->core_clk, 500000000);
273 drm_atomic_helper_wait_for_fences(dev, state, false);
275 drm_atomic_helper_wait_for_dependencies(state);
277 drm_atomic_helper_commit_modeset_disables(dev, state);
279 vc4_ctm_commit(vc4, state);
282 vc5_hvs_pv_muxing_commit(vc4, state);
284 vc4_hvs_pv_muxing_commit(vc4, state);
286 drm_atomic_helper_commit_planes(dev, state, 0);
288 drm_atomic_helper_commit_modeset_enables(dev, state);
290 drm_atomic_helper_fake_vblank(state);
292 drm_atomic_helper_commit_hw_done(state);
294 drm_atomic_helper_wait_for_flip_done(dev, state);
296 drm_atomic_helper_cleanup_planes(dev, state);
298 drm_atomic_helper_commit_cleanup_done(state);
301 clk_set_min_rate(hvs->core_clk, 0);
303 drm_atomic_state_put(state);
305 up(&vc4->async_modeset);
308 static void commit_work(struct work_struct *work)
310 struct drm_atomic_state *state = container_of(work,
311 struct drm_atomic_state,
313 vc4_atomic_complete_commit(state);
317 * vc4_atomic_commit - commit validated state object
319 * @state: the driver state object
320 * @nonblock: nonblocking commit
322 * This function commits a with drm_atomic_helper_check() pre-validated state
323 * object. This can still fail when e.g. the framebuffer reservation fails. For
324 * now this doesn't implement asynchronous commits.
327 * Zero for success or -errno.
329 static int vc4_atomic_commit(struct drm_device *dev,
330 struct drm_atomic_state *state,
333 struct vc4_dev *vc4 = to_vc4_dev(dev);
336 if (state->async_update) {
337 ret = down_interruptible(&vc4->async_modeset);
341 ret = drm_atomic_helper_prepare_planes(dev, state);
343 up(&vc4->async_modeset);
347 drm_atomic_helper_async_commit(dev, state);
349 drm_atomic_helper_cleanup_planes(dev, state);
351 up(&vc4->async_modeset);
356 /* We know for sure we don't want an async update here. Set
357 * state->legacy_cursor_update to false to prevent
358 * drm_atomic_helper_setup_commit() from auto-completing
361 state->legacy_cursor_update = false;
362 ret = drm_atomic_helper_setup_commit(state, nonblock);
366 INIT_WORK(&state->commit_work, commit_work);
368 ret = down_interruptible(&vc4->async_modeset);
372 ret = drm_atomic_helper_prepare_planes(dev, state);
374 up(&vc4->async_modeset);
379 ret = drm_atomic_helper_wait_for_fences(dev, state, true);
381 drm_atomic_helper_cleanup_planes(dev, state);
382 up(&vc4->async_modeset);
388 * This is the point of no return - everything below never fails except
389 * when the hw goes bonghits. Which means we can commit the new state on
390 * the software side now.
393 BUG_ON(drm_atomic_helper_swap_state(state, false) < 0);
396 * Everything below can be run asynchronously without the need to grab
397 * any modeset locks at all under one condition: It must be guaranteed
398 * that the asynchronous work has either been cancelled (if the driver
399 * supports it, which at least requires that the framebuffers get
400 * cleaned up with drm_atomic_helper_cleanup_planes()) or completed
401 * before the new state gets committed on the software side with
402 * drm_atomic_helper_swap_state().
404 * This scheme allows new atomic state updates to be prepared and
405 * checked in parallel to the asynchronous completion of the previous
406 * update. Which is important since compositors need to figure out the
407 * composition of the next frame right after having submitted the
411 drm_atomic_state_get(state);
413 queue_work(system_unbound_wq, &state->commit_work);
415 vc4_atomic_complete_commit(state);
420 static struct drm_framebuffer *vc4_fb_create(struct drm_device *dev,
421 struct drm_file *file_priv,
422 const struct drm_mode_fb_cmd2 *mode_cmd)
424 struct drm_mode_fb_cmd2 mode_cmd_local;
426 /* If the user didn't specify a modifier, use the
427 * vc4_set_tiling_ioctl() state for the BO.
429 if (!(mode_cmd->flags & DRM_MODE_FB_MODIFIERS)) {
430 struct drm_gem_object *gem_obj;
433 gem_obj = drm_gem_object_lookup(file_priv,
434 mode_cmd->handles[0]);
436 DRM_DEBUG("Failed to look up GEM BO %d\n",
437 mode_cmd->handles[0]);
438 return ERR_PTR(-ENOENT);
440 bo = to_vc4_bo(gem_obj);
442 mode_cmd_local = *mode_cmd;
445 mode_cmd_local.modifier[0] =
446 DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
448 mode_cmd_local.modifier[0] = DRM_FORMAT_MOD_NONE;
451 drm_gem_object_put(gem_obj);
453 mode_cmd = &mode_cmd_local;
456 return drm_gem_fb_create(dev, file_priv, mode_cmd);
459 /* Our CTM has some peculiar limitations: we can only enable it for one CRTC
460 * at a time and the HW only supports S0.9 scalars. To account for the latter,
461 * we don't allow userland to set a CTM that we have no hope of approximating.
464 vc4_ctm_atomic_check(struct drm_device *dev, struct drm_atomic_state *state)
466 struct vc4_dev *vc4 = to_vc4_dev(dev);
467 struct vc4_ctm_state *ctm_state = NULL;
468 struct drm_crtc *crtc;
469 struct drm_crtc_state *old_crtc_state, *new_crtc_state;
470 struct drm_color_ctm *ctm;
473 for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
474 /* CTM is being disabled. */
475 if (!new_crtc_state->ctm && old_crtc_state->ctm) {
476 ctm_state = vc4_get_ctm_state(state, &vc4->ctm_manager);
477 if (IS_ERR(ctm_state))
478 return PTR_ERR(ctm_state);
483 for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
484 if (new_crtc_state->ctm == old_crtc_state->ctm)
488 ctm_state = vc4_get_ctm_state(state, &vc4->ctm_manager);
489 if (IS_ERR(ctm_state))
490 return PTR_ERR(ctm_state);
493 /* CTM is being enabled or the matrix changed. */
494 if (new_crtc_state->ctm) {
495 struct vc4_crtc_state *vc4_crtc_state =
496 to_vc4_crtc_state(new_crtc_state);
498 /* fifo is 1-based since 0 disables CTM. */
499 int fifo = vc4_crtc_state->assigned_channel + 1;
501 /* Check userland isn't trying to turn on CTM for more
502 * than one CRTC at a time.
504 if (ctm_state->fifo && ctm_state->fifo != fifo) {
505 DRM_DEBUG_DRIVER("Too many CTM configured\n");
509 /* Check we can approximate the specified CTM.
510 * We disallow scalars |c| > 1.0 since the HW has
513 ctm = new_crtc_state->ctm->data;
514 for (i = 0; i < ARRAY_SIZE(ctm->matrix); i++) {
515 u64 val = ctm->matrix[i];
518 if (val > BIT_ULL(32))
522 ctm_state->fifo = fifo;
523 ctm_state->ctm = ctm;
530 static int vc4_load_tracker_atomic_check(struct drm_atomic_state *state)
532 struct drm_plane_state *old_plane_state, *new_plane_state;
533 struct vc4_dev *vc4 = to_vc4_dev(state->dev);
534 struct vc4_load_tracker_state *load_state;
535 struct drm_private_state *priv_state;
536 struct drm_plane *plane;
539 if (!vc4->load_tracker_available)
542 priv_state = drm_atomic_get_private_obj_state(state,
544 if (IS_ERR(priv_state))
545 return PTR_ERR(priv_state);
547 load_state = to_vc4_load_tracker_state(priv_state);
548 for_each_oldnew_plane_in_state(state, plane, old_plane_state,
549 new_plane_state, i) {
550 struct vc4_plane_state *vc4_plane_state;
552 if (old_plane_state->fb && old_plane_state->crtc) {
553 vc4_plane_state = to_vc4_plane_state(old_plane_state);
554 load_state->membus_load -= vc4_plane_state->membus_load;
555 load_state->hvs_load -= vc4_plane_state->hvs_load;
558 if (new_plane_state->fb && new_plane_state->crtc) {
559 vc4_plane_state = to_vc4_plane_state(new_plane_state);
560 load_state->membus_load += vc4_plane_state->membus_load;
561 load_state->hvs_load += vc4_plane_state->hvs_load;
565 /* Don't check the load when the tracker is disabled. */
566 if (!vc4->load_tracker_enabled)
569 /* The absolute limit is 2Gbyte/sec, but let's take a margin to let
570 * the system work when other blocks are accessing the memory.
572 if (load_state->membus_load > SZ_1G + SZ_512M)
575 /* HVS clock is supposed to run @ 250Mhz, let's take a margin and
576 * consider the maximum number of cycles is 240M.
578 if (load_state->hvs_load > 240000000ULL)
584 static struct drm_private_state *
585 vc4_load_tracker_duplicate_state(struct drm_private_obj *obj)
587 struct vc4_load_tracker_state *state;
589 state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
593 __drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
598 static void vc4_load_tracker_destroy_state(struct drm_private_obj *obj,
599 struct drm_private_state *state)
601 struct vc4_load_tracker_state *load_state;
603 load_state = to_vc4_load_tracker_state(state);
607 static const struct drm_private_state_funcs vc4_load_tracker_state_funcs = {
608 .atomic_duplicate_state = vc4_load_tracker_duplicate_state,
609 .atomic_destroy_state = vc4_load_tracker_destroy_state,
612 #define NUM_OUTPUTS 6
613 #define NUM_CHANNELS 3
616 vc4_atomic_check(struct drm_device *dev, struct drm_atomic_state *state)
618 unsigned long unassigned_channels = GENMASK(NUM_CHANNELS - 1, 0);
619 struct drm_crtc_state *crtc_state;
620 struct drm_crtc *crtc;
624 * Since the HVS FIFOs are shared across all the pixelvalves and
625 * the TXP (and thus all the CRTCs), we need to pull the current
626 * state of all the enabled CRTCs so that an update to a single
627 * CRTC still keeps the previous FIFOs enabled and assigned to
628 * the same CRTCs, instead of evaluating only the CRTC being
631 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
632 if (!crtc->state->enable)
635 crtc_state = drm_atomic_get_crtc_state(state, crtc);
636 if (IS_ERR(crtc_state))
637 return PTR_ERR(crtc_state);
640 for_each_new_crtc_in_state(state, crtc, crtc_state, i) {
641 struct vc4_crtc_state *vc4_crtc_state =
642 to_vc4_crtc_state(crtc_state);
643 struct vc4_crtc *vc4_crtc = to_vc4_crtc(crtc);
644 unsigned int matching_channels;
646 if (!crtc_state->active)
650 * The problem we have to solve here is that we have
651 * up to 7 encoders, connected to up to 6 CRTCs.
653 * Those CRTCs, depending on the instance, can be
654 * routed to 1, 2 or 3 HVS FIFOs, and we need to set
655 * the change the muxing between FIFOs and outputs in
656 * the HVS accordingly.
658 * It would be pretty hard to come up with an
659 * algorithm that would generically solve
660 * this. However, the current routing trees we support
661 * allow us to simplify a bit the problem.
663 * Indeed, with the current supported layouts, if we
664 * try to assign in the ascending crtc index order the
665 * FIFOs, we can't fall into the situation where an
666 * earlier CRTC that had multiple routes is assigned
667 * one that was the only option for a later CRTC.
669 * If the layout changes and doesn't give us that in
670 * the future, we will need to have something smarter,
671 * but it works so far.
673 matching_channels = unassigned_channels & vc4_crtc->data->hvs_available_channels;
674 if (matching_channels) {
675 unsigned int channel = ffs(matching_channels) - 1;
677 vc4_crtc_state->assigned_channel = channel;
678 unassigned_channels &= ~BIT(channel);
684 ret = vc4_ctm_atomic_check(dev, state);
688 ret = drm_atomic_helper_check(dev, state);
692 return vc4_load_tracker_atomic_check(state);
695 static const struct drm_mode_config_funcs vc4_mode_funcs = {
696 .atomic_check = vc4_atomic_check,
697 .atomic_commit = vc4_atomic_commit,
698 .fb_create = vc4_fb_create,
701 int vc4_kms_load(struct drm_device *dev)
703 struct vc4_dev *vc4 = to_vc4_dev(dev);
704 struct vc4_ctm_state *ctm_state;
705 struct vc4_load_tracker_state *load_state;
706 bool is_vc5 = of_device_is_compatible(dev->dev->of_node,
711 vc4->load_tracker_available = true;
713 /* Start with the load tracker enabled. Can be
714 * disabled through the debugfs load_tracker file.
716 vc4->load_tracker_enabled = true;
719 sema_init(&vc4->async_modeset, 1);
721 /* Set support for vblank irq fast disable, before drm_vblank_init() */
722 dev->vblank_disable_immediate = true;
724 dev->irq_enabled = true;
725 ret = drm_vblank_init(dev, dev->mode_config.num_crtc);
727 dev_err(dev->dev, "failed to initialize vblank\n");
732 dev->mode_config.max_width = 7680;
733 dev->mode_config.max_height = 7680;
735 dev->mode_config.max_width = 2048;
736 dev->mode_config.max_height = 2048;
739 dev->mode_config.funcs = &vc4_mode_funcs;
740 dev->mode_config.preferred_depth = 24;
741 dev->mode_config.async_page_flip = true;
742 dev->mode_config.allow_fb_modifiers = true;
744 drm_modeset_lock_init(&vc4->ctm_state_lock);
746 ctm_state = kzalloc(sizeof(*ctm_state), GFP_KERNEL);
750 drm_atomic_private_obj_init(dev, &vc4->ctm_manager, &ctm_state->base,
751 &vc4_ctm_state_funcs);
753 if (vc4->load_tracker_available) {
754 load_state = kzalloc(sizeof(*load_state), GFP_KERNEL);
756 drm_atomic_private_obj_fini(&vc4->ctm_manager);
760 drm_atomic_private_obj_init(dev, &vc4->load_tracker,
762 &vc4_load_tracker_state_funcs);
765 drm_mode_config_reset(dev);
767 drm_kms_helper_poll_init(dev);