drivers/gpu/drm/vc4/vc4_kms.c

   1 /*
   2  * Copyright (C) 2015 Broadcom
   3  *
   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.
   7  */
   8
   9 /**
  10  * DOC: VC4 KMS
  11  *
  12  * This is the general code for implementing KMS mode setting that
  13  * doesn't clearly associate with any of the other objects (plane,
  14  * crtc, HDMI encoder).
  15  */
  16
  17 #include <drm/drm_crtc.h>
  18 #include <drm/drm_atomic.h>
  19 #include <drm/drm_atomic_helper.h>
  20 #include <drm/drm_gem_framebuffer_helper.h>
  21 #include <drm/drm_plane_helper.h>
  22 #include <drm/drm_probe_helper.h>
  23 #include "vc4_drv.h"
  24 #include "vc4_regs.h"
  25
  26 struct vc4_ctm_state {
  27         struct drm_private_state base;
  28         struct drm_color_ctm *ctm;
  29         int fifo;
  30 };
  31
  32 static struct vc4_ctm_state *to_vc4_ctm_state(struct drm_private_state *priv)
  33 {
  34         return container_of(priv, struct vc4_ctm_state, base);
  35 }
  36
  37 struct vc4_load_tracker_state {
  38         struct drm_private_state base;
  39         u64 hvs_load;
  40         u64 membus_load;
  41 };
  42
  43 static struct vc4_load_tracker_state *
  44 to_vc4_load_tracker_state(struct drm_private_state *priv)
  45 {
  46         return container_of(priv, struct vc4_load_tracker_state, base);
  47 }
  48
  49 static struct vc4_ctm_state *vc4_get_ctm_state(struct drm_atomic_state *state,
  50                                                struct drm_private_obj *manager)
  51 {
  52         struct drm_device *dev = state->dev;
  53         struct vc4_dev *vc4 = dev->dev_private;
  54         struct drm_private_state *priv_state;
  55         int ret;
  56
  57         ret = drm_modeset_lock(&vc4->ctm_state_lock, state->acquire_ctx);
  58         if (ret)
  59                 return ERR_PTR(ret);
  60
  61         priv_state = drm_atomic_get_private_obj_state(state, manager);
  62         if (IS_ERR(priv_state))
  63                 return ERR_CAST(priv_state);
  64
  65         return to_vc4_ctm_state(priv_state);
  66 }
  67
  68 static struct drm_private_state *
  69 vc4_ctm_duplicate_state(struct drm_private_obj *obj)
  70 {
  71         struct vc4_ctm_state *state;
  72
  73         state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
  74         if (!state)
  75                 return NULL;
  76
  77         __drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
  78
  79         return &state->base;
  80 }
  81
  82 static void vc4_ctm_destroy_state(struct drm_private_obj *obj,
  83                                   struct drm_private_state *state)
  84 {
  85         struct vc4_ctm_state *ctm_state = to_vc4_ctm_state(state);
  86
  87         kfree(ctm_state);
  88 }
  89
  90 static const struct drm_private_state_funcs vc4_ctm_state_funcs = {
  91         .atomic_duplicate_state = vc4_ctm_duplicate_state,
  92         .atomic_destroy_state = vc4_ctm_destroy_state,
  93 };
  94
  95 /* Converts a DRM S31.32 value to the HW S0.9 format. */
  96 static u16 vc4_ctm_s31_32_to_s0_9(u64 in)
  97 {
  98         u16 r;
  99
 100         /* Sign bit. */
 101         r = in & BIT_ULL(63) ? BIT(9) : 0;
 102
 103         if ((in & GENMASK_ULL(62, 32)) > 0) {
 104                 /* We have zero integer bits so we can only saturate here. */
 105                 r |= GENMASK(8, 0);
 106         } else {
 107                 /* Otherwise take the 9 most important fractional bits. */
 108                 r |= (in >> 23) & GENMASK(8, 0);
 109         }
 110
 111         return r;
 112 }
 113
 114 static void
 115 vc4_ctm_commit(struct vc4_dev *vc4, struct drm_atomic_state *state)
 116 {
 117         struct vc4_ctm_state *ctm_state = to_vc4_ctm_state(vc4->ctm_manager.state);
 118         struct drm_color_ctm *ctm = ctm_state->ctm;
 119
 120         if (ctm_state->fifo) {
 121                 HVS_WRITE(SCALER_OLEDCOEF2,
 122                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[0]),
 123                                         SCALER_OLEDCOEF2_R_TO_R) |
 124                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[3]),
 125                                         SCALER_OLEDCOEF2_R_TO_G) |
 126                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[6]),
 127                                         SCALER_OLEDCOEF2_R_TO_B));
 128                 HVS_WRITE(SCALER_OLEDCOEF1,
 129                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[1]),
 130                                         SCALER_OLEDCOEF1_G_TO_R) |
 131                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[4]),
 132                                         SCALER_OLEDCOEF1_G_TO_G) |
 133                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[7]),
 134                                         SCALER_OLEDCOEF1_G_TO_B));
 135                 HVS_WRITE(SCALER_OLEDCOEF0,
 136                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[2]),
 137                                         SCALER_OLEDCOEF0_B_TO_R) |
 138                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[5]),
 139                                         SCALER_OLEDCOEF0_B_TO_G) |
 140                           VC4_SET_FIELD(vc4_ctm_s31_32_to_s0_9(ctm->matrix[8]),
 141                                         SCALER_OLEDCOEF0_B_TO_B));
 142         }
 143
 144         HVS_WRITE(SCALER_OLEDOFFS,
 145                   VC4_SET_FIELD(ctm_state->fifo, SCALER_OLEDOFFS_DISPFIFO));
 146 }
 147
 148 static void
 149 vc4_atomic_complete_commit(struct drm_atomic_state *state)
 150 {
 151         struct drm_device *dev = state->dev;
 152         struct vc4_dev *vc4 = to_vc4_dev(dev);
 153         struct vc4_crtc *vc4_crtc;
 154         int i;
 155
 156         for (i = 0; i < dev->mode_config.num_crtc; i++) {
 157                 if (!state->crtcs[i].ptr || !state->crtcs[i].commit)
 158                         continue;
 159
 160                 vc4_crtc = to_vc4_crtc(state->crtcs[i].ptr);
 161                 vc4_hvs_mask_underrun(dev, vc4_crtc->channel);
 162         }
 163
 164         drm_atomic_helper_wait_for_fences(dev, state, false);
 165
 166         drm_atomic_helper_wait_for_dependencies(state);
 167
 168         drm_atomic_helper_commit_modeset_disables(dev, state);
 169
 170         vc4_ctm_commit(vc4, state);
 171
 172         drm_atomic_helper_commit_planes(dev, state, 0);
 173
 174         drm_atomic_helper_commit_modeset_enables(dev, state);
 175
 176         drm_atomic_helper_fake_vblank(state);
 177
 178         drm_atomic_helper_commit_hw_done(state);
 179
 180         drm_atomic_helper_wait_for_flip_done(dev, state);
 181
 182         drm_atomic_helper_cleanup_planes(dev, state);
 183
 184         drm_atomic_helper_commit_cleanup_done(state);
 185
 186         drm_atomic_state_put(state);
 187
 188         up(&vc4->async_modeset);
 189 }
 190
 191 static void commit_work(struct work_struct *work)
 192 {
 193         struct drm_atomic_state *state = container_of(work,
 194                                                       struct drm_atomic_state,
 195                                                       commit_work);
 196         vc4_atomic_complete_commit(state);
 197 }
 198
 199 /**
 200  * vc4_atomic_commit - commit validated state object
 201  * @dev: DRM device
 202  * @state: the driver state object
 203  * @nonblock: nonblocking commit
 204  *
 205  * This function commits a with drm_atomic_helper_check() pre-validated state
 206  * object. This can still fail when e.g. the framebuffer reservation fails. For
 207  * now this doesn't implement asynchronous commits.
 208  *
 209  * RETURNS
 210  * Zero for success or -errno.
 211  */
 212 static int vc4_atomic_commit(struct drm_device *dev,
 213                              struct drm_atomic_state *state,
 214                              bool nonblock)
 215 {
 216         struct vc4_dev *vc4 = to_vc4_dev(dev);
 217         int ret;
 218
 219         if (state->async_update) {
 220                 ret = down_interruptible(&vc4->async_modeset);
 221                 if (ret)
 222                         return ret;
 223
 224                 ret = drm_atomic_helper_prepare_planes(dev, state);
 225                 if (ret) {
 226                         up(&vc4->async_modeset);
 227                         return ret;
 228                 }
 229
 230                 drm_atomic_helper_async_commit(dev, state);
 231
 232                 drm_atomic_helper_cleanup_planes(dev, state);
 233
 234                 up(&vc4->async_modeset);
 235
 236                 return 0;
 237         }
 238
 239         /* We know for sure we don't want an async update here. Set
 240          * state->legacy_cursor_update to false to prevent
 241          * drm_atomic_helper_setup_commit() from auto-completing
 242          * commit->flip_done.
 243          */
 244         state->legacy_cursor_update = false;
 245         ret = drm_atomic_helper_setup_commit(state, nonblock);
 246         if (ret)
 247                 return ret;
 248
 249         INIT_WORK(&state->commit_work, commit_work);
 250
 251         ret = down_interruptible(&vc4->async_modeset);
 252         if (ret)
 253                 return ret;
 254
 255         ret = drm_atomic_helper_prepare_planes(dev, state);
 256         if (ret) {
 257                 up(&vc4->async_modeset);
 258                 return ret;
 259         }
 260
 261         if (!nonblock) {
 262                 ret = drm_atomic_helper_wait_for_fences(dev, state, true);
 263                 if (ret) {
 264                         drm_atomic_helper_cleanup_planes(dev, state);
 265                         up(&vc4->async_modeset);
 266                         return ret;
 267                 }
 268         }
 269
 270         /*
 271          * This is the point of no return - everything below never fails except
 272          * when the hw goes bonghits. Which means we can commit the new state on
 273          * the software side now.
 274          */
 275
 276         BUG_ON(drm_atomic_helper_swap_state(state, false) < 0);
 277
 278         /*
 279          * Everything below can be run asynchronously without the need to grab
 280          * any modeset locks at all under one condition: It must be guaranteed
 281          * that the asynchronous work has either been cancelled (if the driver
 282          * supports it, which at least requires that the framebuffers get
 283          * cleaned up with drm_atomic_helper_cleanup_planes()) or completed
 284          * before the new state gets committed on the software side with
 285          * drm_atomic_helper_swap_state().
 286          *
 287          * This scheme allows new atomic state updates to be prepared and
 288          * checked in parallel to the asynchronous completion of the previous
 289          * update. Which is important since compositors need to figure out the
 290          * composition of the next frame right after having submitted the
 291          * current layout.
 292          */
 293
 294         drm_atomic_state_get(state);
 295         if (nonblock)
 296                 queue_work(system_unbound_wq, &state->commit_work);
 297         else
 298                 vc4_atomic_complete_commit(state);
 299
 300         return 0;
 301 }
 302
 303 static struct drm_framebuffer *vc4_fb_create(struct drm_device *dev,
 304                                              struct drm_file *file_priv,
 305                                              const struct drm_mode_fb_cmd2 *mode_cmd)
 306 {
 307         struct drm_mode_fb_cmd2 mode_cmd_local;
 308
 309         /* If the user didn't specify a modifier, use the
 310          * vc4_set_tiling_ioctl() state for the BO.
 311          */
 312         if (!(mode_cmd->flags & DRM_MODE_FB_MODIFIERS)) {
 313                 struct drm_gem_object *gem_obj;
 314                 struct vc4_bo *bo;
 315
 316                 gem_obj = drm_gem_object_lookup(file_priv,
 317                                                 mode_cmd->handles[0]);
 318                 if (!gem_obj) {
 319                         DRM_DEBUG("Failed to look up GEM BO %d\n",
 320                                   mode_cmd->handles[0]);
 321                         return ERR_PTR(-ENOENT);
 322                 }
 323                 bo = to_vc4_bo(gem_obj);
 324
 325                 mode_cmd_local = *mode_cmd;
 326
 327                 if (bo->t_format) {
 328                         mode_cmd_local.modifier[0] =
 329                                 DRM_FORMAT_MOD_BROADCOM_VC4_T_TILED;
 330                 } else {
 331                         mode_cmd_local.modifier[0] = DRM_FORMAT_MOD_NONE;
 332                 }
 333
 334                 drm_gem_object_put_unlocked(gem_obj);
 335
 336                 mode_cmd = &mode_cmd_local;
 337         }
 338
 339         return drm_gem_fb_create(dev, file_priv, mode_cmd);
 340 }
 341
 342 /* Our CTM has some peculiar limitations: we can only enable it for one CRTC
 343  * at a time and the HW only supports S0.9 scalars. To account for the latter,
 344  * we don't allow userland to set a CTM that we have no hope of approximating.
 345  */
 346 static int
 347 vc4_ctm_atomic_check(struct drm_device *dev, struct drm_atomic_state *state)
 348 {
 349         struct vc4_dev *vc4 = to_vc4_dev(dev);
 350         struct vc4_ctm_state *ctm_state = NULL;
 351         struct drm_crtc *crtc;
 352         struct drm_crtc_state *old_crtc_state, *new_crtc_state;
 353         struct drm_color_ctm *ctm;
 354         int i;
 355
 356         for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
 357                 /* CTM is being disabled. */
 358                 if (!new_crtc_state->ctm && old_crtc_state->ctm) {
 359                         ctm_state = vc4_get_ctm_state(state, &vc4->ctm_manager);
 360                         if (IS_ERR(ctm_state))
 361                                 return PTR_ERR(ctm_state);
 362                         ctm_state->fifo = 0;
 363                 }
 364         }
 365
 366         for_each_oldnew_crtc_in_state(state, crtc, old_crtc_state, new_crtc_state, i) {
 367                 if (new_crtc_state->ctm == old_crtc_state->ctm)
 368                         continue;
 369
 370                 if (!ctm_state) {
 371                         ctm_state = vc4_get_ctm_state(state, &vc4->ctm_manager);
 372                         if (IS_ERR(ctm_state))
 373                                 return PTR_ERR(ctm_state);
 374                 }
 375
 376                 /* CTM is being enabled or the matrix changed. */
 377                 if (new_crtc_state->ctm) {
 378                         /* fifo is 1-based since 0 disables CTM. */
 379                         int fifo = to_vc4_crtc(crtc)->channel + 1;
 380
 381                         /* Check userland isn't trying to turn on CTM for more
 382                          * than one CRTC at a time.
 383                          */
 384                         if (ctm_state->fifo && ctm_state->fifo != fifo) {
 385                                 DRM_DEBUG_DRIVER("Too many CTM configured\n");
 386                                 return -EINVAL;
 387                         }
 388
 389                         /* Check we can approximate the specified CTM.
 390                          * We disallow scalars |c| > 1.0 since the HW has
 391                          * no integer bits.
 392                          */
 393                         ctm = new_crtc_state->ctm->data;
 394                         for (i = 0; i < ARRAY_SIZE(ctm->matrix); i++) {
 395                                 u64 val = ctm->matrix[i];
 396
 397                                 val &= ~BIT_ULL(63);
 398                                 if (val > BIT_ULL(32))
 399                                         return -EINVAL;
 400                         }
 401
 402                         ctm_state->fifo = fifo;
 403                         ctm_state->ctm = ctm;
 404                 }
 405         }
 406
 407         return 0;
 408 }
 409
 410 static int vc4_load_tracker_atomic_check(struct drm_atomic_state *state)
 411 {
 412         struct drm_plane_state *old_plane_state, *new_plane_state;
 413         struct vc4_dev *vc4 = to_vc4_dev(state->dev);
 414         struct vc4_load_tracker_state *load_state;
 415         struct drm_private_state *priv_state;
 416         struct drm_plane *plane;
 417         int i;
 418
 419         priv_state = drm_atomic_get_private_obj_state(state,
 420                                                       &vc4->load_tracker);
 421         if (IS_ERR(priv_state))
 422                 return PTR_ERR(priv_state);
 423
 424         load_state = to_vc4_load_tracker_state(priv_state);
 425         for_each_oldnew_plane_in_state(state, plane, old_plane_state,
 426                                        new_plane_state, i) {
 427                 struct vc4_plane_state *vc4_plane_state;
 428
 429                 if (old_plane_state->fb && old_plane_state->crtc) {
 430                         vc4_plane_state = to_vc4_plane_state(old_plane_state);
 431                         load_state->membus_load -= vc4_plane_state->membus_load;
 432                         load_state->hvs_load -= vc4_plane_state->hvs_load;
 433                 }
 434
 435                 if (new_plane_state->fb && new_plane_state->crtc) {
 436                         vc4_plane_state = to_vc4_plane_state(new_plane_state);
 437                         load_state->membus_load += vc4_plane_state->membus_load;
 438                         load_state->hvs_load += vc4_plane_state->hvs_load;
 439                 }
 440         }
 441
 442         /* Don't check the load when the tracker is disabled. */
 443         if (!vc4->load_tracker_enabled)
 444                 return 0;
 445
 446         /* The absolute limit is 2Gbyte/sec, but let's take a margin to let
 447          * the system work when other blocks are accessing the memory.
 448          */
 449         if (load_state->membus_load > SZ_1G + SZ_512M)
 450                 return -ENOSPC;
 451
 452         /* HVS clock is supposed to run @ 250Mhz, let's take a margin and
 453          * consider the maximum number of cycles is 240M.
 454          */
 455         if (load_state->hvs_load > 240000000ULL)
 456                 return -ENOSPC;
 457
 458         return 0;
 459 }
 460
 461 static struct drm_private_state *
 462 vc4_load_tracker_duplicate_state(struct drm_private_obj *obj)
 463 {
 464         struct vc4_load_tracker_state *state;
 465
 466         state = kmemdup(obj->state, sizeof(*state), GFP_KERNEL);
 467         if (!state)
 468                 return NULL;
 469
 470         __drm_atomic_helper_private_obj_duplicate_state(obj, &state->base);
 471
 472         return &state->base;
 473 }
 474
 475 static void vc4_load_tracker_destroy_state(struct drm_private_obj *obj,
 476                                            struct drm_private_state *state)
 477 {
 478         struct vc4_load_tracker_state *load_state;
 479
 480         load_state = to_vc4_load_tracker_state(state);
 481         kfree(load_state);
 482 }
 483
 484 static const struct drm_private_state_funcs vc4_load_tracker_state_funcs = {
 485         .atomic_duplicate_state = vc4_load_tracker_duplicate_state,
 486         .atomic_destroy_state = vc4_load_tracker_destroy_state,
 487 };
 488
 489 static int
 490 vc4_atomic_check(struct drm_device *dev, struct drm_atomic_state *state)
 491 {
 492         int ret;
 493
 494         ret = vc4_ctm_atomic_check(dev, state);
 495         if (ret < 0)
 496                 return ret;
 497
 498         ret = drm_atomic_helper_check(dev, state);
 499         if (ret)
 500                 return ret;
 501
 502         return vc4_load_tracker_atomic_check(state);
 503 }
 504
 505 static const struct drm_mode_config_funcs vc4_mode_funcs = {
 506         .atomic_check = vc4_atomic_check,
 507         .atomic_commit = vc4_atomic_commit,
 508         .fb_create = vc4_fb_create,
 509 };
 510
 511 int vc4_kms_load(struct drm_device *dev)
 512 {
 513         struct vc4_dev *vc4 = to_vc4_dev(dev);
 514         struct vc4_ctm_state *ctm_state;
 515         struct vc4_load_tracker_state *load_state;
 516         int ret;
 517
 518         /* Start with the load tracker enabled. Can be disabled through the
 519          * debugfs load_tracker file.
 520          */
 521         vc4->load_tracker_enabled = true;
 522
 523         sema_init(&vc4->async_modeset, 1);
 524
 525         /* Set support for vblank irq fast disable, before drm_vblank_init() */
 526         dev->vblank_disable_immediate = true;
 527
 528         dev->irq_enabled = true;
 529         ret = drm_vblank_init(dev, dev->mode_config.num_crtc);
 530         if (ret < 0) {
 531                 dev_err(dev->dev, "failed to initialize vblank\n");
 532                 return ret;
 533         }
 534
 535         dev->mode_config.max_width = 2048;
 536         dev->mode_config.max_height = 2048;
 537         dev->mode_config.funcs = &vc4_mode_funcs;
 538         dev->mode_config.preferred_depth = 24;
 539         dev->mode_config.async_page_flip = true;
 540         dev->mode_config.allow_fb_modifiers = true;
 541
 542         drm_modeset_lock_init(&vc4->ctm_state_lock);
 543
 544         ctm_state = kzalloc(sizeof(*ctm_state), GFP_KERNEL);
 545         if (!ctm_state)
 546                 return -ENOMEM;
 547
 548         drm_atomic_private_obj_init(dev, &vc4->ctm_manager, &ctm_state->base,
 549                                     &vc4_ctm_state_funcs);
 550
 551         load_state = kzalloc(sizeof(*load_state), GFP_KERNEL);
 552         if (!load_state) {
 553                 drm_atomic_private_obj_fini(&vc4->ctm_manager);
 554                 return -ENOMEM;
 555         }
 556
 557         drm_atomic_private_obj_init(dev, &vc4->load_tracker, &load_state->base,
 558                                     &vc4_load_tracker_state_funcs);
 559
 560         drm_mode_config_reset(dev);
 561
 562         drm_kms_helper_poll_init(dev);
 563
 564         return 0;
 565 }