drivers/gpu/drm/vc4/vc4_v3d.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
   4  * Copyright (C) 2013 Red Hat
   5  * Author: Rob Clark <robdclark@gmail.com>
   6  */
   7
   8 #include <linux/clk.h>
   9 #include <linux/component.h>
  10 #include <linux/platform_device.h>
  11 #include <linux/pm_runtime.h>
  12
  13 #include <drm/drm_irq.h>
  14
  15 #include "vc4_drv.h"
  16 #include "vc4_regs.h"
  17
  18 static const struct debugfs_reg32 v3d_regs[] = {
  19         VC4_REG32(V3D_IDENT0),
  20         VC4_REG32(V3D_IDENT1),
  21         VC4_REG32(V3D_IDENT2),
  22         VC4_REG32(V3D_SCRATCH),
  23         VC4_REG32(V3D_L2CACTL),
  24         VC4_REG32(V3D_SLCACTL),
  25         VC4_REG32(V3D_INTCTL),
  26         VC4_REG32(V3D_INTENA),
  27         VC4_REG32(V3D_INTDIS),
  28         VC4_REG32(V3D_CT0CS),
  29         VC4_REG32(V3D_CT1CS),
  30         VC4_REG32(V3D_CT0EA),
  31         VC4_REG32(V3D_CT1EA),
  32         VC4_REG32(V3D_CT0CA),
  33         VC4_REG32(V3D_CT1CA),
  34         VC4_REG32(V3D_CT00RA0),
  35         VC4_REG32(V3D_CT01RA0),
  36         VC4_REG32(V3D_CT0LC),
  37         VC4_REG32(V3D_CT1LC),
  38         VC4_REG32(V3D_CT0PC),
  39         VC4_REG32(V3D_CT1PC),
  40         VC4_REG32(V3D_PCS),
  41         VC4_REG32(V3D_BFC),
  42         VC4_REG32(V3D_RFC),
  43         VC4_REG32(V3D_BPCA),
  44         VC4_REG32(V3D_BPCS),
  45         VC4_REG32(V3D_BPOA),
  46         VC4_REG32(V3D_BPOS),
  47         VC4_REG32(V3D_BXCF),
  48         VC4_REG32(V3D_SQRSV0),
  49         VC4_REG32(V3D_SQRSV1),
  50         VC4_REG32(V3D_SQCNTL),
  51         VC4_REG32(V3D_SRQPC),
  52         VC4_REG32(V3D_SRQUA),
  53         VC4_REG32(V3D_SRQUL),
  54         VC4_REG32(V3D_SRQCS),
  55         VC4_REG32(V3D_VPACNTL),
  56         VC4_REG32(V3D_VPMBASE),
  57         VC4_REG32(V3D_PCTRC),
  58         VC4_REG32(V3D_PCTRE),
  59         VC4_REG32(V3D_PCTR(0)),
  60         VC4_REG32(V3D_PCTRS(0)),
  61         VC4_REG32(V3D_PCTR(1)),
  62         VC4_REG32(V3D_PCTRS(1)),
  63         VC4_REG32(V3D_PCTR(2)),
  64         VC4_REG32(V3D_PCTRS(2)),
  65         VC4_REG32(V3D_PCTR(3)),
  66         VC4_REG32(V3D_PCTRS(3)),
  67         VC4_REG32(V3D_PCTR(4)),
  68         VC4_REG32(V3D_PCTRS(4)),
  69         VC4_REG32(V3D_PCTR(5)),
  70         VC4_REG32(V3D_PCTRS(5)),
  71         VC4_REG32(V3D_PCTR(6)),
  72         VC4_REG32(V3D_PCTRS(6)),
  73         VC4_REG32(V3D_PCTR(7)),
  74         VC4_REG32(V3D_PCTRS(7)),
  75         VC4_REG32(V3D_PCTR(8)),
  76         VC4_REG32(V3D_PCTRS(8)),
  77         VC4_REG32(V3D_PCTR(9)),
  78         VC4_REG32(V3D_PCTRS(9)),
  79         VC4_REG32(V3D_PCTR(10)),
  80         VC4_REG32(V3D_PCTRS(10)),
  81         VC4_REG32(V3D_PCTR(11)),
  82         VC4_REG32(V3D_PCTRS(11)),
  83         VC4_REG32(V3D_PCTR(12)),
  84         VC4_REG32(V3D_PCTRS(12)),
  85         VC4_REG32(V3D_PCTR(13)),
  86         VC4_REG32(V3D_PCTRS(13)),
  87         VC4_REG32(V3D_PCTR(14)),
  88         VC4_REG32(V3D_PCTRS(14)),
  89         VC4_REG32(V3D_PCTR(15)),
  90         VC4_REG32(V3D_PCTRS(15)),
  91         VC4_REG32(V3D_DBGE),
  92         VC4_REG32(V3D_FDBGO),
  93         VC4_REG32(V3D_FDBGB),
  94         VC4_REG32(V3D_FDBGR),
  95         VC4_REG32(V3D_FDBGS),
  96         VC4_REG32(V3D_ERRSTAT),
  97 };
  98
  99 static int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused)
 100 {
 101         struct drm_info_node *node = (struct drm_info_node *)m->private;
 102         struct drm_device *dev = node->minor->dev;
 103         struct vc4_dev *vc4 = to_vc4_dev(dev);
 104         int ret = vc4_v3d_pm_get(vc4);
 105
 106         if (ret == 0) {
 107                 uint32_t ident1 = V3D_READ(V3D_IDENT1);
 108                 uint32_t nslc = VC4_GET_FIELD(ident1, V3D_IDENT1_NSLC);
 109                 uint32_t tups = VC4_GET_FIELD(ident1, V3D_IDENT1_TUPS);
 110                 uint32_t qups = VC4_GET_FIELD(ident1, V3D_IDENT1_QUPS);
 111
 112                 seq_printf(m, "Revision:   %d\n",
 113                            VC4_GET_FIELD(ident1, V3D_IDENT1_REV));
 114                 seq_printf(m, "Slices:     %d\n", nslc);
 115                 seq_printf(m, "TMUs:       %d\n", nslc * tups);
 116                 seq_printf(m, "QPUs:       %d\n", nslc * qups);
 117                 seq_printf(m, "Semaphores: %d\n",
 118                            VC4_GET_FIELD(ident1, V3D_IDENT1_NSEM));
 119                 vc4_v3d_pm_put(vc4);
 120         }
 121
 122         return 0;
 123 }
 124
 125 /*
 126  * Wraps pm_runtime_get_sync() in a refcount, so that we can reliably
 127  * get the pm_runtime refcount to 0 in vc4_reset().
 128  */
 129 int
 130 vc4_v3d_pm_get(struct vc4_dev *vc4)
 131 {
 132         mutex_lock(&vc4->power_lock);
 133         if (vc4->power_refcount++ == 0) {
 134                 int ret = pm_runtime_get_sync(&vc4->v3d->pdev->dev);
 135
 136                 if (ret < 0) {
 137                         vc4->power_refcount--;
 138                         mutex_unlock(&vc4->power_lock);
 139                         return ret;
 140                 }
 141         }
 142         mutex_unlock(&vc4->power_lock);
 143
 144         return 0;
 145 }
 146
 147 void
 148 vc4_v3d_pm_put(struct vc4_dev *vc4)
 149 {
 150         mutex_lock(&vc4->power_lock);
 151         if (--vc4->power_refcount == 0) {
 152                 pm_runtime_mark_last_busy(&vc4->v3d->pdev->dev);
 153                 pm_runtime_put_autosuspend(&vc4->v3d->pdev->dev);
 154         }
 155         mutex_unlock(&vc4->power_lock);
 156 }
 157
 158 static void vc4_v3d_init_hw(struct drm_device *dev)
 159 {
 160         struct vc4_dev *vc4 = to_vc4_dev(dev);
 161
 162         /* Take all the memory that would have been reserved for user
 163          * QPU programs, since we don't have an interface for running
 164          * them, anyway.
 165          */
 166         V3D_WRITE(V3D_VPMBASE, 0);
 167 }
 168
 169 int vc4_v3d_get_bin_slot(struct vc4_dev *vc4)
 170 {
 171         struct drm_device *dev = &vc4->base;
 172         unsigned long irqflags;
 173         int slot;
 174         uint64_t seqno = 0;
 175         struct vc4_exec_info *exec;
 176
 177 try_again:
 178         spin_lock_irqsave(&vc4->job_lock, irqflags);
 179         slot = ffs(~vc4->bin_alloc_used);
 180         if (slot != 0) {
 181                 /* Switch from ffs() bit index to a 0-based index. */
 182                 slot--;
 183                 vc4->bin_alloc_used |= BIT(slot);
 184                 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 185                 return slot;
 186         }
 187
 188         /* Couldn't find an open slot.  Wait for render to complete
 189          * and try again.
 190          */
 191         exec = vc4_last_render_job(vc4);
 192         if (exec)
 193                 seqno = exec->seqno;
 194         spin_unlock_irqrestore(&vc4->job_lock, irqflags);
 195
 196         if (seqno) {
 197                 int ret = vc4_wait_for_seqno(dev, seqno, ~0ull, true);
 198
 199                 if (ret == 0)
 200                         goto try_again;
 201
 202                 return ret;
 203         }
 204
 205         return -ENOMEM;
 206 }
 207
 208 /*
 209  * bin_bo_alloc() - allocates the memory that will be used for
 210  * tile binning.
 211  *
 212  * The binner has a limitation that the addresses in the tile state
 213  * buffer that point into the tile alloc buffer or binner overflow
 214  * memory only have 28 bits (256MB), and the top 4 on the bus for
 215  * tile alloc references end up coming from the tile state buffer's
 216  * address.
 217  *
 218  * To work around this, we allocate a single large buffer while V3D is
 219  * in use, make sure that it has the top 4 bits constant across its
 220  * entire extent, and then put the tile state, tile alloc, and binner
 221  * overflow memory inside that buffer.
 222  *
 223  * This creates a limitation where we may not be able to execute a job
 224  * if it doesn't fit within the buffer that we allocated up front.
 225  * However, it turns out that 16MB is "enough for anybody", and
 226  * real-world applications run into allocation failures from the
 227  * overall CMA pool before they make scenes complicated enough to run
 228  * out of bin space.
 229  */
 230 static int bin_bo_alloc(struct vc4_dev *vc4)
 231 {
 232         struct vc4_v3d *v3d = vc4->v3d;
 233         uint32_t size = 16 * 1024 * 1024;
 234         int ret = 0;
 235         struct list_head list;
 236
 237         if (!v3d)
 238                 return -ENODEV;
 239
 240         /* We may need to try allocating more than once to get a BO
 241          * that doesn't cross 256MB.  Track the ones we've allocated
 242          * that failed so far, so that we can free them when we've got
 243          * one that succeeded (if we freed them right away, our next
 244          * allocation would probably be the same chunk of memory).
 245          */
 246         INIT_LIST_HEAD(&list);
 247
 248         while (true) {
 249                 struct vc4_bo *bo = vc4_bo_create(&vc4->base, size, true,
 250                                                   VC4_BO_TYPE_BIN);
 251
 252                 if (IS_ERR(bo)) {
 253                         ret = PTR_ERR(bo);
 254
 255                         dev_err(&v3d->pdev->dev,
 256                                 "Failed to allocate memory for tile binning: "
 257                                 "%d. You may need to enable CMA or give it "
 258                                 "more memory.",
 259                                 ret);
 260                         break;
 261                 }
 262
 263                 /* Check if this BO won't trigger the addressing bug. */
 264                 if ((bo->base.paddr & 0xf0000000) ==
 265                     ((bo->base.paddr + bo->base.base.size - 1) & 0xf0000000)) {
 266                         vc4->bin_bo = bo;
 267
 268                         /* Set up for allocating 512KB chunks of
 269                          * binner memory.  The biggest allocation we
 270                          * need to do is for the initial tile alloc +
 271                          * tile state buffer.  We can render to a
 272                          * maximum of ((2048*2048) / (32*32) = 4096
 273                          * tiles in a frame (until we do floating
 274                          * point rendering, at which point it would be
 275                          * 8192).  Tile state is 48b/tile (rounded to
 276                          * a page), and tile alloc is 32b/tile
 277                          * (rounded to a page), plus a page of extra,
 278                          * for a total of 320kb for our worst-case.
 279                          * We choose 512kb so that it divides evenly
 280                          * into our 16MB, and the rest of the 512kb
 281                          * will be used as storage for the overflow
 282                          * from the initial 32b CL per bin.
 283                          */
 284                         vc4->bin_alloc_size = 512 * 1024;
 285                         vc4->bin_alloc_used = 0;
 286                         vc4->bin_alloc_overflow = 0;
 287                         WARN_ON_ONCE(sizeof(vc4->bin_alloc_used) * 8 !=
 288                                      bo->base.base.size / vc4->bin_alloc_size);
 289
 290                         kref_init(&vc4->bin_bo_kref);
 291
 292                         /* Enable the out-of-memory interrupt to set our
 293                          * newly-allocated binner BO, potentially from an
 294                          * already-pending-but-masked interrupt.
 295                          */
 296                         V3D_WRITE(V3D_INTENA, V3D_INT_OUTOMEM);
 297
 298                         break;
 299                 }
 300
 301                 /* Put it on the list to free later, and try again. */
 302                 list_add(&bo->unref_head, &list);
 303         }
 304
 305         /* Free all the BOs we allocated but didn't choose. */
 306         while (!list_empty(&list)) {
 307                 struct vc4_bo *bo = list_last_entry(&list,
 308                                                     struct vc4_bo, unref_head);
 309
 310                 list_del(&bo->unref_head);
 311                 drm_gem_object_put(&bo->base.base);
 312         }
 313
 314         return ret;
 315 }
 316
 317 int vc4_v3d_bin_bo_get(struct vc4_dev *vc4, bool *used)
 318 {
 319         int ret = 0;
 320
 321         mutex_lock(&vc4->bin_bo_lock);
 322
 323         if (used && *used)
 324                 goto complete;
 325
 326         if (vc4->bin_bo)
 327                 kref_get(&vc4->bin_bo_kref);
 328         else
 329                 ret = bin_bo_alloc(vc4);
 330
 331         if (ret == 0 && used)
 332                 *used = true;
 333
 334 complete:
 335         mutex_unlock(&vc4->bin_bo_lock);
 336
 337         return ret;
 338 }
 339
 340 static void bin_bo_release(struct kref *ref)
 341 {
 342         struct vc4_dev *vc4 = container_of(ref, struct vc4_dev, bin_bo_kref);
 343
 344         if (WARN_ON_ONCE(!vc4->bin_bo))
 345                 return;
 346
 347         drm_gem_object_put(&vc4->bin_bo->base.base);
 348         vc4->bin_bo = NULL;
 349 }
 350
 351 void vc4_v3d_bin_bo_put(struct vc4_dev *vc4)
 352 {
 353         mutex_lock(&vc4->bin_bo_lock);
 354         kref_put(&vc4->bin_bo_kref, bin_bo_release);
 355         mutex_unlock(&vc4->bin_bo_lock);
 356 }
 357
 358 #ifdef CONFIG_PM
 359 static int vc4_v3d_runtime_suspend(struct device *dev)
 360 {
 361         struct vc4_v3d *v3d = dev_get_drvdata(dev);
 362         struct vc4_dev *vc4 = v3d->vc4;
 363
 364         vc4_irq_uninstall(&vc4->base);
 365
 366         clk_disable_unprepare(v3d->clk);
 367
 368         return 0;
 369 }
 370
 371 static int vc4_v3d_runtime_resume(struct device *dev)
 372 {
 373         struct vc4_v3d *v3d = dev_get_drvdata(dev);
 374         struct vc4_dev *vc4 = v3d->vc4;
 375         int ret;
 376
 377         ret = clk_prepare_enable(v3d->clk);
 378         if (ret != 0)
 379                 return ret;
 380
 381         vc4_v3d_init_hw(&vc4->base);
 382
 383         /* We disabled the IRQ as part of vc4_irq_uninstall in suspend. */
 384         enable_irq(vc4->base.irq);
 385         vc4_irq_postinstall(&vc4->base);
 386
 387         return 0;
 388 }
 389 #endif
 390
 391 static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
 392 {
 393         struct platform_device *pdev = to_platform_device(dev);
 394         struct drm_device *drm = dev_get_drvdata(master);
 395         struct vc4_dev *vc4 = to_vc4_dev(drm);
 396         struct vc4_v3d *v3d = NULL;
 397         int ret;
 398
 399         v3d = devm_kzalloc(&pdev->dev, sizeof(*v3d), GFP_KERNEL);
 400         if (!v3d)
 401                 return -ENOMEM;
 402
 403         dev_set_drvdata(dev, v3d);
 404
 405         v3d->pdev = pdev;
 406
 407         v3d->regs = vc4_ioremap_regs(pdev, 0);
 408         if (IS_ERR(v3d->regs))
 409                 return PTR_ERR(v3d->regs);
 410         v3d->regset.base = v3d->regs;
 411         v3d->regset.regs = v3d_regs;
 412         v3d->regset.nregs = ARRAY_SIZE(v3d_regs);
 413
 414         vc4->v3d = v3d;
 415         v3d->vc4 = vc4;
 416
 417         v3d->clk = devm_clk_get(dev, NULL);
 418         if (IS_ERR(v3d->clk)) {
 419                 int ret = PTR_ERR(v3d->clk);
 420
 421                 if (ret == -ENOENT) {
 422                         /* bcm2835 didn't have a clock reference in the DT. */
 423                         ret = 0;
 424                         v3d->clk = NULL;
 425                 } else {
 426                         if (ret != -EPROBE_DEFER)
 427                                 dev_err(dev, "Failed to get V3D clock: %d\n",
 428                                         ret);
 429                         return ret;
 430                 }
 431         }
 432
 433         if (V3D_READ(V3D_IDENT0) != V3D_EXPECTED_IDENT0) {
 434                 DRM_ERROR("V3D_IDENT0 read 0x%08x instead of 0x%08x\n",
 435                           V3D_READ(V3D_IDENT0), V3D_EXPECTED_IDENT0);
 436                 return -EINVAL;
 437         }
 438
 439         ret = clk_prepare_enable(v3d->clk);
 440         if (ret != 0)
 441                 return ret;
 442
 443         /* Reset the binner overflow address/size at setup, to be sure
 444          * we don't reuse an old one.
 445          */
 446         V3D_WRITE(V3D_BPOA, 0);
 447         V3D_WRITE(V3D_BPOS, 0);
 448
 449         vc4_v3d_init_hw(drm);
 450
 451         ret = drm_irq_install(drm, platform_get_irq(pdev, 0));
 452         if (ret) {
 453                 DRM_ERROR("Failed to install IRQ handler\n");
 454                 return ret;
 455         }
 456
 457         pm_runtime_set_active(dev);
 458         pm_runtime_use_autosuspend(dev);
 459         pm_runtime_set_autosuspend_delay(dev, 40); /* a little over 2 frames. */
 460         pm_runtime_enable(dev);
 461
 462         vc4_debugfs_add_file(drm, "v3d_ident", vc4_v3d_debugfs_ident, NULL);
 463         vc4_debugfs_add_regset32(drm, "v3d_regs", &v3d->regset);
 464
 465         return 0;
 466 }
 467
 468 static void vc4_v3d_unbind(struct device *dev, struct device *master,
 469                            void *data)
 470 {
 471         struct drm_device *drm = dev_get_drvdata(master);
 472         struct vc4_dev *vc4 = to_vc4_dev(drm);
 473
 474         pm_runtime_disable(dev);
 475
 476         drm_irq_uninstall(drm);
 477
 478         /* Disable the binner's overflow memory address, so the next
 479          * driver probe (if any) doesn't try to reuse our old
 480          * allocation.
 481          */
 482         V3D_WRITE(V3D_BPOA, 0);
 483         V3D_WRITE(V3D_BPOS, 0);
 484
 485         vc4->v3d = NULL;
 486 }
 487
 488 static const struct dev_pm_ops vc4_v3d_pm_ops = {
 489         SET_RUNTIME_PM_OPS(vc4_v3d_runtime_suspend, vc4_v3d_runtime_resume, NULL)
 490 };
 491
 492 static const struct component_ops vc4_v3d_ops = {
 493         .bind   = vc4_v3d_bind,
 494         .unbind = vc4_v3d_unbind,
 495 };
 496
 497 static int vc4_v3d_dev_probe(struct platform_device *pdev)
 498 {
 499         return component_add(&pdev->dev, &vc4_v3d_ops);
 500 }
 501
 502 static int vc4_v3d_dev_remove(struct platform_device *pdev)
 503 {
 504         component_del(&pdev->dev, &vc4_v3d_ops);
 505         return 0;
 506 }
 507
 508 const struct of_device_id vc4_v3d_dt_match[] = {
 509         { .compatible = "brcm,bcm2835-v3d" },
 510         { .compatible = "brcm,cygnus-v3d" },
 511         { .compatible = "brcm,vc4-v3d" },
 512         {}
 513 };
 514
 515 struct platform_driver vc4_v3d_driver = {
 516         .probe = vc4_v3d_dev_probe,
 517         .remove = vc4_v3d_dev_remove,
 518         .driver = {
 519                 .name = "vc4_v3d",
 520                 .of_match_table = vc4_v3d_dt_match,
 521                 .pm = &vc4_v3d_pm_ops,
 522         },
 523 };