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_print.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
vc4_v3d_debugfs_ident(struct seq_file * m,void * unused)99 static int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused)
100 {
101 struct drm_debugfs_entry *entry = m->private;
102 struct drm_device *dev = entry->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
vc4_v3d_pm_get(struct vc4_dev * vc4)130 vc4_v3d_pm_get(struct vc4_dev *vc4)
131 {
132 if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
133 return -ENODEV;
134
135 mutex_lock(&vc4->power_lock);
136 if (vc4->power_refcount++ == 0) {
137 int ret = pm_runtime_get_sync(&vc4->v3d->pdev->dev);
138
139 if (ret < 0) {
140 vc4->power_refcount--;
141 mutex_unlock(&vc4->power_lock);
142 return ret;
143 }
144 }
145 mutex_unlock(&vc4->power_lock);
146
147 return 0;
148 }
149
150 void
vc4_v3d_pm_put(struct vc4_dev * vc4)151 vc4_v3d_pm_put(struct vc4_dev *vc4)
152 {
153 if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
154 return;
155
156 mutex_lock(&vc4->power_lock);
157 if (--vc4->power_refcount == 0) {
158 pm_runtime_mark_last_busy(&vc4->v3d->pdev->dev);
159 pm_runtime_put_autosuspend(&vc4->v3d->pdev->dev);
160 }
161 mutex_unlock(&vc4->power_lock);
162 }
163
vc4_v3d_init_hw(struct drm_device * dev)164 static void vc4_v3d_init_hw(struct drm_device *dev)
165 {
166 struct vc4_dev *vc4 = to_vc4_dev(dev);
167
168 /* Take all the memory that would have been reserved for user
169 * QPU programs, since we don't have an interface for running
170 * them, anyway.
171 */
172 V3D_WRITE(V3D_VPMBASE, 0);
173 }
174
vc4_v3d_get_bin_slot(struct vc4_dev * vc4)175 int vc4_v3d_get_bin_slot(struct vc4_dev *vc4)
176 {
177 struct drm_device *dev = &vc4->base;
178 unsigned long irqflags;
179 int slot;
180 uint64_t seqno = 0;
181 struct vc4_exec_info *exec;
182
183 if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
184 return -ENODEV;
185
186 try_again:
187 spin_lock_irqsave(&vc4->job_lock, irqflags);
188 slot = ffs(~vc4->bin_alloc_used);
189 if (slot != 0) {
190 /* Switch from ffs() bit index to a 0-based index. */
191 slot--;
192 vc4->bin_alloc_used |= BIT(slot);
193 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
194 return slot;
195 }
196
197 /* Couldn't find an open slot. Wait for render to complete
198 * and try again.
199 */
200 exec = vc4_last_render_job(vc4);
201 if (exec)
202 seqno = exec->seqno;
203 spin_unlock_irqrestore(&vc4->job_lock, irqflags);
204
205 if (seqno) {
206 int ret = vc4_wait_for_seqno(dev, seqno, ~0ull, true);
207
208 if (ret == 0)
209 goto try_again;
210
211 return ret;
212 }
213
214 return -ENOMEM;
215 }
216
217 /*
218 * bin_bo_alloc() - allocates the memory that will be used for
219 * tile binning.
220 *
221 * The binner has a limitation that the addresses in the tile state
222 * buffer that point into the tile alloc buffer or binner overflow
223 * memory only have 28 bits (256MB), and the top 4 on the bus for
224 * tile alloc references end up coming from the tile state buffer's
225 * address.
226 *
227 * To work around this, we allocate a single large buffer while V3D is
228 * in use, make sure that it has the top 4 bits constant across its
229 * entire extent, and then put the tile state, tile alloc, and binner
230 * overflow memory inside that buffer.
231 *
232 * This creates a limitation where we may not be able to execute a job
233 * if it doesn't fit within the buffer that we allocated up front.
234 * However, it turns out that 16MB is "enough for anybody", and
235 * real-world applications run into allocation failures from the
236 * overall DMA pool before they make scenes complicated enough to run
237 * out of bin space.
238 */
bin_bo_alloc(struct vc4_dev * vc4)239 static int bin_bo_alloc(struct vc4_dev *vc4)
240 {
241 struct vc4_v3d *v3d = vc4->v3d;
242 uint32_t size = 16 * 1024 * 1024;
243 int ret = 0;
244 struct list_head list;
245
246 if (!v3d)
247 return -ENODEV;
248
249 /* We may need to try allocating more than once to get a BO
250 * that doesn't cross 256MB. Track the ones we've allocated
251 * that failed so far, so that we can free them when we've got
252 * one that succeeded (if we freed them right away, our next
253 * allocation would probably be the same chunk of memory).
254 */
255 INIT_LIST_HEAD(&list);
256
257 while (true) {
258 struct vc4_bo *bo = vc4_bo_create(&vc4->base, size, true,
259 VC4_BO_TYPE_BIN);
260
261 if (IS_ERR(bo)) {
262 ret = PTR_ERR(bo);
263
264 dev_err(&v3d->pdev->dev,
265 "Failed to allocate memory for tile binning: "
266 "%d. You may need to enable DMA or give it "
267 "more memory.",
268 ret);
269 break;
270 }
271
272 /* Check if this BO won't trigger the addressing bug. */
273 if ((bo->base.dma_addr & 0xf0000000) ==
274 ((bo->base.dma_addr + bo->base.base.size - 1) & 0xf0000000)) {
275 vc4->bin_bo = bo;
276
277 /* Set up for allocating 512KB chunks of
278 * binner memory. The biggest allocation we
279 * need to do is for the initial tile alloc +
280 * tile state buffer. We can render to a
281 * maximum of ((2048*2048) / (32*32) = 4096
282 * tiles in a frame (until we do floating
283 * point rendering, at which point it would be
284 * 8192). Tile state is 48b/tile (rounded to
285 * a page), and tile alloc is 32b/tile
286 * (rounded to a page), plus a page of extra,
287 * for a total of 320kb for our worst-case.
288 * We choose 512kb so that it divides evenly
289 * into our 16MB, and the rest of the 512kb
290 * will be used as storage for the overflow
291 * from the initial 32b CL per bin.
292 */
293 vc4->bin_alloc_size = 512 * 1024;
294 vc4->bin_alloc_used = 0;
295 vc4->bin_alloc_overflow = 0;
296 WARN_ON_ONCE(sizeof(vc4->bin_alloc_used) * 8 !=
297 bo->base.base.size / vc4->bin_alloc_size);
298
299 kref_init(&vc4->bin_bo_kref);
300
301 /* Enable the out-of-memory interrupt to set our
302 * newly-allocated binner BO, potentially from an
303 * already-pending-but-masked interrupt.
304 */
305 V3D_WRITE(V3D_INTENA, V3D_INT_OUTOMEM);
306
307 break;
308 }
309
310 /* Put it on the list to free later, and try again. */
311 list_add(&bo->unref_head, &list);
312 }
313
314 /* Free all the BOs we allocated but didn't choose. */
315 while (!list_empty(&list)) {
316 struct vc4_bo *bo = list_last_entry(&list,
317 struct vc4_bo, unref_head);
318
319 list_del(&bo->unref_head);
320 drm_gem_object_put(&bo->base.base);
321 }
322
323 return ret;
324 }
325
vc4_v3d_bin_bo_get(struct vc4_dev * vc4,bool * used)326 int vc4_v3d_bin_bo_get(struct vc4_dev *vc4, bool *used)
327 {
328 int ret = 0;
329
330 if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
331 return -ENODEV;
332
333 mutex_lock(&vc4->bin_bo_lock);
334
335 if (used && *used)
336 goto complete;
337
338 if (vc4->bin_bo)
339 kref_get(&vc4->bin_bo_kref);
340 else
341 ret = bin_bo_alloc(vc4);
342
343 if (ret == 0 && used)
344 *used = true;
345
346 complete:
347 mutex_unlock(&vc4->bin_bo_lock);
348
349 return ret;
350 }
351
bin_bo_release(struct kref * ref)352 static void bin_bo_release(struct kref *ref)
353 {
354 struct vc4_dev *vc4 = container_of(ref, struct vc4_dev, bin_bo_kref);
355
356 if (WARN_ON_ONCE(!vc4->bin_bo))
357 return;
358
359 drm_gem_object_put(&vc4->bin_bo->base.base);
360 vc4->bin_bo = NULL;
361 }
362
vc4_v3d_bin_bo_put(struct vc4_dev * vc4)363 void vc4_v3d_bin_bo_put(struct vc4_dev *vc4)
364 {
365 if (WARN_ON_ONCE(vc4->gen > VC4_GEN_4))
366 return;
367
368 mutex_lock(&vc4->bin_bo_lock);
369 kref_put(&vc4->bin_bo_kref, bin_bo_release);
370 mutex_unlock(&vc4->bin_bo_lock);
371 }
372
373 #ifdef CONFIG_PM
vc4_v3d_runtime_suspend(struct device * dev)374 static int vc4_v3d_runtime_suspend(struct device *dev)
375 {
376 struct vc4_v3d *v3d = dev_get_drvdata(dev);
377 struct vc4_dev *vc4 = v3d->vc4;
378
379 vc4_irq_disable(&vc4->base);
380
381 clk_disable_unprepare(v3d->clk);
382
383 return 0;
384 }
385
vc4_v3d_runtime_resume(struct device * dev)386 static int vc4_v3d_runtime_resume(struct device *dev)
387 {
388 struct vc4_v3d *v3d = dev_get_drvdata(dev);
389 struct vc4_dev *vc4 = v3d->vc4;
390 int ret;
391
392 ret = clk_prepare_enable(v3d->clk);
393 if (ret != 0)
394 return ret;
395
396 vc4_v3d_init_hw(&vc4->base);
397
398 vc4_irq_enable(&vc4->base);
399
400 return 0;
401 }
402 #endif
403
vc4_v3d_debugfs_init(struct drm_minor * minor)404 int vc4_v3d_debugfs_init(struct drm_minor *minor)
405 {
406 struct drm_device *drm = minor->dev;
407 struct vc4_dev *vc4 = to_vc4_dev(drm);
408 struct vc4_v3d *v3d = vc4->v3d;
409
410 if (!vc4->v3d)
411 return -ENODEV;
412
413 drm_debugfs_add_file(drm, "v3d_ident", vc4_v3d_debugfs_ident, NULL);
414
415 vc4_debugfs_add_regset32(drm, "v3d_regs", &v3d->regset);
416
417 return 0;
418 }
419
vc4_v3d_bind(struct device * dev,struct device * master,void * data)420 static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
421 {
422 struct platform_device *pdev = to_platform_device(dev);
423 struct drm_device *drm = dev_get_drvdata(master);
424 struct vc4_dev *vc4 = to_vc4_dev(drm);
425 struct vc4_v3d *v3d = NULL;
426 int ret;
427
428 v3d = devm_kzalloc(&pdev->dev, sizeof(*v3d), GFP_KERNEL);
429 if (!v3d)
430 return -ENOMEM;
431
432 dev_set_drvdata(dev, v3d);
433
434 v3d->pdev = pdev;
435
436 v3d->regs = vc4_ioremap_regs(pdev, 0);
437 if (IS_ERR(v3d->regs))
438 return PTR_ERR(v3d->regs);
439 v3d->regset.base = v3d->regs;
440 v3d->regset.regs = v3d_regs;
441 v3d->regset.nregs = ARRAY_SIZE(v3d_regs);
442
443 vc4->v3d = v3d;
444 v3d->vc4 = vc4;
445
446 v3d->clk = devm_clk_get_optional(dev, NULL);
447 if (IS_ERR(v3d->clk))
448 return dev_err_probe(dev, PTR_ERR(v3d->clk), "Failed to get V3D clock\n");
449
450 ret = platform_get_irq(pdev, 0);
451 if (ret < 0)
452 return ret;
453 vc4->irq = ret;
454
455 ret = devm_pm_runtime_enable(dev);
456 if (ret)
457 return ret;
458
459 ret = pm_runtime_resume_and_get(dev);
460 if (ret)
461 return ret;
462
463 if (V3D_READ(V3D_IDENT0) != V3D_EXPECTED_IDENT0) {
464 drm_err(drm, "V3D_IDENT0 read 0x%08x instead of 0x%08x\n",
465 V3D_READ(V3D_IDENT0), V3D_EXPECTED_IDENT0);
466 ret = -EINVAL;
467 goto err_put_runtime_pm;
468 }
469
470 /* Reset the binner overflow address/size at setup, to be sure
471 * we don't reuse an old one.
472 */
473 V3D_WRITE(V3D_BPOA, 0);
474 V3D_WRITE(V3D_BPOS, 0);
475
476 ret = vc4_irq_install(drm, vc4->irq);
477 if (ret) {
478 drm_err(drm, "Failed to install IRQ handler\n");
479 goto err_put_runtime_pm;
480 }
481
482 pm_runtime_use_autosuspend(dev);
483 pm_runtime_set_autosuspend_delay(dev, 40); /* a little over 2 frames. */
484
485 return 0;
486
487 err_put_runtime_pm:
488 pm_runtime_put(dev);
489
490 return ret;
491 }
492
vc4_v3d_unbind(struct device * dev,struct device * master,void * data)493 static void vc4_v3d_unbind(struct device *dev, struct device *master,
494 void *data)
495 {
496 struct drm_device *drm = dev_get_drvdata(master);
497 struct vc4_dev *vc4 = to_vc4_dev(drm);
498
499 vc4_irq_uninstall(drm);
500
501 /* Disable the binner's overflow memory address, so the next
502 * driver probe (if any) doesn't try to reuse our old
503 * allocation.
504 */
505 V3D_WRITE(V3D_BPOA, 0);
506 V3D_WRITE(V3D_BPOS, 0);
507
508 vc4->v3d = NULL;
509 }
510
511 static const struct dev_pm_ops vc4_v3d_pm_ops = {
512 SET_RUNTIME_PM_OPS(vc4_v3d_runtime_suspend, vc4_v3d_runtime_resume, NULL)
513 };
514
515 static const struct component_ops vc4_v3d_ops = {
516 .bind = vc4_v3d_bind,
517 .unbind = vc4_v3d_unbind,
518 };
519
vc4_v3d_dev_probe(struct platform_device * pdev)520 static int vc4_v3d_dev_probe(struct platform_device *pdev)
521 {
522 return component_add(&pdev->dev, &vc4_v3d_ops);
523 }
524
vc4_v3d_dev_remove(struct platform_device * pdev)525 static void vc4_v3d_dev_remove(struct platform_device *pdev)
526 {
527 component_del(&pdev->dev, &vc4_v3d_ops);
528 }
529
530 const struct of_device_id vc4_v3d_dt_match[] = {
531 { .compatible = "brcm,bcm2835-v3d" },
532 { .compatible = "brcm,cygnus-v3d" },
533 { .compatible = "brcm,vc4-v3d" },
534 {}
535 };
536
537 struct platform_driver vc4_v3d_driver = {
538 .probe = vc4_v3d_dev_probe,
539 .remove = vc4_v3d_dev_remove,
540 .driver = {
541 .name = "vc4_v3d",
542 .of_match_table = vc4_v3d_dt_match,
543 .pm = &vc4_v3d_pm_ops,
544 },
545 };
546