xref: /linux/drivers/gpu/drm/vc4/vc4_v3d.c (revision b4db9f840283caca0d904436f187ef56a9126eaa)
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 "vc4_drv.h"
14 #include "vc4_regs.h"
15 
16 static const struct debugfs_reg32 v3d_regs[] = {
17 	VC4_REG32(V3D_IDENT0),
18 	VC4_REG32(V3D_IDENT1),
19 	VC4_REG32(V3D_IDENT2),
20 	VC4_REG32(V3D_SCRATCH),
21 	VC4_REG32(V3D_L2CACTL),
22 	VC4_REG32(V3D_SLCACTL),
23 	VC4_REG32(V3D_INTCTL),
24 	VC4_REG32(V3D_INTENA),
25 	VC4_REG32(V3D_INTDIS),
26 	VC4_REG32(V3D_CT0CS),
27 	VC4_REG32(V3D_CT1CS),
28 	VC4_REG32(V3D_CT0EA),
29 	VC4_REG32(V3D_CT1EA),
30 	VC4_REG32(V3D_CT0CA),
31 	VC4_REG32(V3D_CT1CA),
32 	VC4_REG32(V3D_CT00RA0),
33 	VC4_REG32(V3D_CT01RA0),
34 	VC4_REG32(V3D_CT0LC),
35 	VC4_REG32(V3D_CT1LC),
36 	VC4_REG32(V3D_CT0PC),
37 	VC4_REG32(V3D_CT1PC),
38 	VC4_REG32(V3D_PCS),
39 	VC4_REG32(V3D_BFC),
40 	VC4_REG32(V3D_RFC),
41 	VC4_REG32(V3D_BPCA),
42 	VC4_REG32(V3D_BPCS),
43 	VC4_REG32(V3D_BPOA),
44 	VC4_REG32(V3D_BPOS),
45 	VC4_REG32(V3D_BXCF),
46 	VC4_REG32(V3D_SQRSV0),
47 	VC4_REG32(V3D_SQRSV1),
48 	VC4_REG32(V3D_SQCNTL),
49 	VC4_REG32(V3D_SRQPC),
50 	VC4_REG32(V3D_SRQUA),
51 	VC4_REG32(V3D_SRQUL),
52 	VC4_REG32(V3D_SRQCS),
53 	VC4_REG32(V3D_VPACNTL),
54 	VC4_REG32(V3D_VPMBASE),
55 	VC4_REG32(V3D_PCTRC),
56 	VC4_REG32(V3D_PCTRE),
57 	VC4_REG32(V3D_PCTR(0)),
58 	VC4_REG32(V3D_PCTRS(0)),
59 	VC4_REG32(V3D_PCTR(1)),
60 	VC4_REG32(V3D_PCTRS(1)),
61 	VC4_REG32(V3D_PCTR(2)),
62 	VC4_REG32(V3D_PCTRS(2)),
63 	VC4_REG32(V3D_PCTR(3)),
64 	VC4_REG32(V3D_PCTRS(3)),
65 	VC4_REG32(V3D_PCTR(4)),
66 	VC4_REG32(V3D_PCTRS(4)),
67 	VC4_REG32(V3D_PCTR(5)),
68 	VC4_REG32(V3D_PCTRS(5)),
69 	VC4_REG32(V3D_PCTR(6)),
70 	VC4_REG32(V3D_PCTRS(6)),
71 	VC4_REG32(V3D_PCTR(7)),
72 	VC4_REG32(V3D_PCTRS(7)),
73 	VC4_REG32(V3D_PCTR(8)),
74 	VC4_REG32(V3D_PCTRS(8)),
75 	VC4_REG32(V3D_PCTR(9)),
76 	VC4_REG32(V3D_PCTRS(9)),
77 	VC4_REG32(V3D_PCTR(10)),
78 	VC4_REG32(V3D_PCTRS(10)),
79 	VC4_REG32(V3D_PCTR(11)),
80 	VC4_REG32(V3D_PCTRS(11)),
81 	VC4_REG32(V3D_PCTR(12)),
82 	VC4_REG32(V3D_PCTRS(12)),
83 	VC4_REG32(V3D_PCTR(13)),
84 	VC4_REG32(V3D_PCTRS(13)),
85 	VC4_REG32(V3D_PCTR(14)),
86 	VC4_REG32(V3D_PCTRS(14)),
87 	VC4_REG32(V3D_PCTR(15)),
88 	VC4_REG32(V3D_PCTRS(15)),
89 	VC4_REG32(V3D_DBGE),
90 	VC4_REG32(V3D_FDBGO),
91 	VC4_REG32(V3D_FDBGB),
92 	VC4_REG32(V3D_FDBGR),
93 	VC4_REG32(V3D_FDBGS),
94 	VC4_REG32(V3D_ERRSTAT),
95 };
96 
97 static int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused)
98 {
99 	struct drm_debugfs_entry *entry = m->private;
100 	struct drm_device *dev = entry->dev;
101 	struct vc4_dev *vc4 = to_vc4_dev(dev);
102 	int ret = vc4_v3d_pm_get(vc4);
103 
104 	if (ret == 0) {
105 		uint32_t ident1 = V3D_READ(V3D_IDENT1);
106 		uint32_t nslc = VC4_GET_FIELD(ident1, V3D_IDENT1_NSLC);
107 		uint32_t tups = VC4_GET_FIELD(ident1, V3D_IDENT1_TUPS);
108 		uint32_t qups = VC4_GET_FIELD(ident1, V3D_IDENT1_QUPS);
109 
110 		seq_printf(m, "Revision:   %d\n",
111 			   VC4_GET_FIELD(ident1, V3D_IDENT1_REV));
112 		seq_printf(m, "Slices:     %d\n", nslc);
113 		seq_printf(m, "TMUs:       %d\n", nslc * tups);
114 		seq_printf(m, "QPUs:       %d\n", nslc * qups);
115 		seq_printf(m, "Semaphores: %d\n",
116 			   VC4_GET_FIELD(ident1, V3D_IDENT1_NSEM));
117 		vc4_v3d_pm_put(vc4);
118 	}
119 
120 	return 0;
121 }
122 
123 /*
124  * Wraps pm_runtime_get_sync() in a refcount, so that we can reliably
125  * get the pm_runtime refcount to 0 in vc4_reset().
126  */
127 int
128 vc4_v3d_pm_get(struct vc4_dev *vc4)
129 {
130 	if (WARN_ON_ONCE(vc4->is_vc5))
131 		return -ENODEV;
132 
133 	mutex_lock(&vc4->power_lock);
134 	if (vc4->power_refcount++ == 0) {
135 		int ret = pm_runtime_get_sync(&vc4->v3d->pdev->dev);
136 
137 		if (ret < 0) {
138 			vc4->power_refcount--;
139 			mutex_unlock(&vc4->power_lock);
140 			return ret;
141 		}
142 	}
143 	mutex_unlock(&vc4->power_lock);
144 
145 	return 0;
146 }
147 
148 void
149 vc4_v3d_pm_put(struct vc4_dev *vc4)
150 {
151 	if (WARN_ON_ONCE(vc4->is_vc5))
152 		return;
153 
154 	mutex_lock(&vc4->power_lock);
155 	if (--vc4->power_refcount == 0) {
156 		pm_runtime_mark_last_busy(&vc4->v3d->pdev->dev);
157 		pm_runtime_put_autosuspend(&vc4->v3d->pdev->dev);
158 	}
159 	mutex_unlock(&vc4->power_lock);
160 }
161 
162 static void vc4_v3d_init_hw(struct drm_device *dev)
163 {
164 	struct vc4_dev *vc4 = to_vc4_dev(dev);
165 
166 	/* Take all the memory that would have been reserved for user
167 	 * QPU programs, since we don't have an interface for running
168 	 * them, anyway.
169 	 */
170 	V3D_WRITE(V3D_VPMBASE, 0);
171 }
172 
173 int vc4_v3d_get_bin_slot(struct vc4_dev *vc4)
174 {
175 	struct drm_device *dev = &vc4->base;
176 	unsigned long irqflags;
177 	int slot;
178 	uint64_t seqno = 0;
179 	struct vc4_exec_info *exec;
180 
181 	if (WARN_ON_ONCE(vc4->is_vc5))
182 		return -ENODEV;
183 
184 try_again:
185 	spin_lock_irqsave(&vc4->job_lock, irqflags);
186 	slot = ffs(~vc4->bin_alloc_used);
187 	if (slot != 0) {
188 		/* Switch from ffs() bit index to a 0-based index. */
189 		slot--;
190 		vc4->bin_alloc_used |= BIT(slot);
191 		spin_unlock_irqrestore(&vc4->job_lock, irqflags);
192 		return slot;
193 	}
194 
195 	/* Couldn't find an open slot.  Wait for render to complete
196 	 * and try again.
197 	 */
198 	exec = vc4_last_render_job(vc4);
199 	if (exec)
200 		seqno = exec->seqno;
201 	spin_unlock_irqrestore(&vc4->job_lock, irqflags);
202 
203 	if (seqno) {
204 		int ret = vc4_wait_for_seqno(dev, seqno, ~0ull, true);
205 
206 		if (ret == 0)
207 			goto try_again;
208 
209 		return ret;
210 	}
211 
212 	return -ENOMEM;
213 }
214 
215 /*
216  * bin_bo_alloc() - allocates the memory that will be used for
217  * tile binning.
218  *
219  * The binner has a limitation that the addresses in the tile state
220  * buffer that point into the tile alloc buffer or binner overflow
221  * memory only have 28 bits (256MB), and the top 4 on the bus for
222  * tile alloc references end up coming from the tile state buffer's
223  * address.
224  *
225  * To work around this, we allocate a single large buffer while V3D is
226  * in use, make sure that it has the top 4 bits constant across its
227  * entire extent, and then put the tile state, tile alloc, and binner
228  * overflow memory inside that buffer.
229  *
230  * This creates a limitation where we may not be able to execute a job
231  * if it doesn't fit within the buffer that we allocated up front.
232  * However, it turns out that 16MB is "enough for anybody", and
233  * real-world applications run into allocation failures from the
234  * overall DMA pool before they make scenes complicated enough to run
235  * out of bin space.
236  */
237 static int bin_bo_alloc(struct vc4_dev *vc4)
238 {
239 	struct vc4_v3d *v3d = vc4->v3d;
240 	uint32_t size = 16 * 1024 * 1024;
241 	int ret = 0;
242 	struct list_head list;
243 
244 	if (!v3d)
245 		return -ENODEV;
246 
247 	/* We may need to try allocating more than once to get a BO
248 	 * that doesn't cross 256MB.  Track the ones we've allocated
249 	 * that failed so far, so that we can free them when we've got
250 	 * one that succeeded (if we freed them right away, our next
251 	 * allocation would probably be the same chunk of memory).
252 	 */
253 	INIT_LIST_HEAD(&list);
254 
255 	while (true) {
256 		struct vc4_bo *bo = vc4_bo_create(&vc4->base, size, true,
257 						  VC4_BO_TYPE_BIN);
258 
259 		if (IS_ERR(bo)) {
260 			ret = PTR_ERR(bo);
261 
262 			dev_err(&v3d->pdev->dev,
263 				"Failed to allocate memory for tile binning: "
264 				"%d. You may need to enable DMA or give it "
265 				"more memory.",
266 				ret);
267 			break;
268 		}
269 
270 		/* Check if this BO won't trigger the addressing bug. */
271 		if ((bo->base.dma_addr & 0xf0000000) ==
272 		    ((bo->base.dma_addr + bo->base.base.size - 1) & 0xf0000000)) {
273 			vc4->bin_bo = bo;
274 
275 			/* Set up for allocating 512KB chunks of
276 			 * binner memory.  The biggest allocation we
277 			 * need to do is for the initial tile alloc +
278 			 * tile state buffer.  We can render to a
279 			 * maximum of ((2048*2048) / (32*32) = 4096
280 			 * tiles in a frame (until we do floating
281 			 * point rendering, at which point it would be
282 			 * 8192).  Tile state is 48b/tile (rounded to
283 			 * a page), and tile alloc is 32b/tile
284 			 * (rounded to a page), plus a page of extra,
285 			 * for a total of 320kb for our worst-case.
286 			 * We choose 512kb so that it divides evenly
287 			 * into our 16MB, and the rest of the 512kb
288 			 * will be used as storage for the overflow
289 			 * from the initial 32b CL per bin.
290 			 */
291 			vc4->bin_alloc_size = 512 * 1024;
292 			vc4->bin_alloc_used = 0;
293 			vc4->bin_alloc_overflow = 0;
294 			WARN_ON_ONCE(sizeof(vc4->bin_alloc_used) * 8 !=
295 				     bo->base.base.size / vc4->bin_alloc_size);
296 
297 			kref_init(&vc4->bin_bo_kref);
298 
299 			/* Enable the out-of-memory interrupt to set our
300 			 * newly-allocated binner BO, potentially from an
301 			 * already-pending-but-masked interrupt.
302 			 */
303 			V3D_WRITE(V3D_INTENA, V3D_INT_OUTOMEM);
304 
305 			break;
306 		}
307 
308 		/* Put it on the list to free later, and try again. */
309 		list_add(&bo->unref_head, &list);
310 	}
311 
312 	/* Free all the BOs we allocated but didn't choose. */
313 	while (!list_empty(&list)) {
314 		struct vc4_bo *bo = list_last_entry(&list,
315 						    struct vc4_bo, unref_head);
316 
317 		list_del(&bo->unref_head);
318 		drm_gem_object_put(&bo->base.base);
319 	}
320 
321 	return ret;
322 }
323 
324 int vc4_v3d_bin_bo_get(struct vc4_dev *vc4, bool *used)
325 {
326 	int ret = 0;
327 
328 	if (WARN_ON_ONCE(vc4->is_vc5))
329 		return -ENODEV;
330 
331 	mutex_lock(&vc4->bin_bo_lock);
332 
333 	if (used && *used)
334 		goto complete;
335 
336 	if (vc4->bin_bo)
337 		kref_get(&vc4->bin_bo_kref);
338 	else
339 		ret = bin_bo_alloc(vc4);
340 
341 	if (ret == 0 && used)
342 		*used = true;
343 
344 complete:
345 	mutex_unlock(&vc4->bin_bo_lock);
346 
347 	return ret;
348 }
349 
350 static void bin_bo_release(struct kref *ref)
351 {
352 	struct vc4_dev *vc4 = container_of(ref, struct vc4_dev, bin_bo_kref);
353 
354 	if (WARN_ON_ONCE(!vc4->bin_bo))
355 		return;
356 
357 	drm_gem_object_put(&vc4->bin_bo->base.base);
358 	vc4->bin_bo = NULL;
359 }
360 
361 void vc4_v3d_bin_bo_put(struct vc4_dev *vc4)
362 {
363 	if (WARN_ON_ONCE(vc4->is_vc5))
364 		return;
365 
366 	mutex_lock(&vc4->bin_bo_lock);
367 	kref_put(&vc4->bin_bo_kref, bin_bo_release);
368 	mutex_unlock(&vc4->bin_bo_lock);
369 }
370 
371 #ifdef CONFIG_PM
372 static int vc4_v3d_runtime_suspend(struct device *dev)
373 {
374 	struct vc4_v3d *v3d = dev_get_drvdata(dev);
375 	struct vc4_dev *vc4 = v3d->vc4;
376 
377 	vc4_irq_disable(&vc4->base);
378 
379 	clk_disable_unprepare(v3d->clk);
380 
381 	return 0;
382 }
383 
384 static int vc4_v3d_runtime_resume(struct device *dev)
385 {
386 	struct vc4_v3d *v3d = dev_get_drvdata(dev);
387 	struct vc4_dev *vc4 = v3d->vc4;
388 	int ret;
389 
390 	ret = clk_prepare_enable(v3d->clk);
391 	if (ret != 0)
392 		return ret;
393 
394 	vc4_v3d_init_hw(&vc4->base);
395 
396 	vc4_irq_enable(&vc4->base);
397 
398 	return 0;
399 }
400 #endif
401 
402 int vc4_v3d_debugfs_init(struct drm_minor *minor)
403 {
404 	struct drm_device *drm = minor->dev;
405 	struct vc4_dev *vc4 = to_vc4_dev(drm);
406 	struct vc4_v3d *v3d = vc4->v3d;
407 
408 	if (!vc4->v3d)
409 		return -ENODEV;
410 
411 	drm_debugfs_add_file(drm, "v3d_ident", vc4_v3d_debugfs_ident, NULL);
412 
413 	vc4_debugfs_add_regset32(drm, "v3d_regs", &v3d->regset);
414 
415 	return 0;
416 }
417 
418 static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
419 {
420 	struct platform_device *pdev = to_platform_device(dev);
421 	struct drm_device *drm = dev_get_drvdata(master);
422 	struct vc4_dev *vc4 = to_vc4_dev(drm);
423 	struct vc4_v3d *v3d = NULL;
424 	int ret;
425 
426 	v3d = devm_kzalloc(&pdev->dev, sizeof(*v3d), GFP_KERNEL);
427 	if (!v3d)
428 		return -ENOMEM;
429 
430 	dev_set_drvdata(dev, v3d);
431 
432 	v3d->pdev = pdev;
433 
434 	v3d->regs = vc4_ioremap_regs(pdev, 0);
435 	if (IS_ERR(v3d->regs))
436 		return PTR_ERR(v3d->regs);
437 	v3d->regset.base = v3d->regs;
438 	v3d->regset.regs = v3d_regs;
439 	v3d->regset.nregs = ARRAY_SIZE(v3d_regs);
440 
441 	vc4->v3d = v3d;
442 	v3d->vc4 = vc4;
443 
444 	v3d->clk = devm_clk_get(dev, NULL);
445 	if (IS_ERR(v3d->clk)) {
446 		int ret = PTR_ERR(v3d->clk);
447 
448 		if (ret == -ENOENT) {
449 			/* bcm2835 didn't have a clock reference in the DT. */
450 			ret = 0;
451 			v3d->clk = NULL;
452 		} else {
453 			if (ret != -EPROBE_DEFER)
454 				dev_err(dev, "Failed to get V3D clock: %d\n",
455 					ret);
456 			return ret;
457 		}
458 	}
459 
460 	ret = platform_get_irq(pdev, 0);
461 	if (ret < 0)
462 		return ret;
463 	vc4->irq = ret;
464 
465 	ret = devm_pm_runtime_enable(dev);
466 	if (ret)
467 		return ret;
468 
469 	ret = pm_runtime_resume_and_get(dev);
470 	if (ret)
471 		return ret;
472 
473 	if (V3D_READ(V3D_IDENT0) != V3D_EXPECTED_IDENT0) {
474 		DRM_ERROR("V3D_IDENT0 read 0x%08x instead of 0x%08x\n",
475 			  V3D_READ(V3D_IDENT0), V3D_EXPECTED_IDENT0);
476 		ret = -EINVAL;
477 		goto err_put_runtime_pm;
478 	}
479 
480 	/* Reset the binner overflow address/size at setup, to be sure
481 	 * we don't reuse an old one.
482 	 */
483 	V3D_WRITE(V3D_BPOA, 0);
484 	V3D_WRITE(V3D_BPOS, 0);
485 
486 	ret = vc4_irq_install(drm, vc4->irq);
487 	if (ret) {
488 		DRM_ERROR("Failed to install IRQ handler\n");
489 		goto err_put_runtime_pm;
490 	}
491 
492 	pm_runtime_use_autosuspend(dev);
493 	pm_runtime_set_autosuspend_delay(dev, 40); /* a little over 2 frames. */
494 
495 	return 0;
496 
497 err_put_runtime_pm:
498 	pm_runtime_put(dev);
499 
500 	return ret;
501 }
502 
503 static void vc4_v3d_unbind(struct device *dev, struct device *master,
504 			   void *data)
505 {
506 	struct drm_device *drm = dev_get_drvdata(master);
507 	struct vc4_dev *vc4 = to_vc4_dev(drm);
508 
509 	vc4_irq_uninstall(drm);
510 
511 	/* Disable the binner's overflow memory address, so the next
512 	 * driver probe (if any) doesn't try to reuse our old
513 	 * allocation.
514 	 */
515 	V3D_WRITE(V3D_BPOA, 0);
516 	V3D_WRITE(V3D_BPOS, 0);
517 
518 	vc4->v3d = NULL;
519 }
520 
521 static const struct dev_pm_ops vc4_v3d_pm_ops = {
522 	SET_RUNTIME_PM_OPS(vc4_v3d_runtime_suspend, vc4_v3d_runtime_resume, NULL)
523 };
524 
525 static const struct component_ops vc4_v3d_ops = {
526 	.bind   = vc4_v3d_bind,
527 	.unbind = vc4_v3d_unbind,
528 };
529 
530 static int vc4_v3d_dev_probe(struct platform_device *pdev)
531 {
532 	return component_add(&pdev->dev, &vc4_v3d_ops);
533 }
534 
535 static void vc4_v3d_dev_remove(struct platform_device *pdev)
536 {
537 	component_del(&pdev->dev, &vc4_v3d_ops);
538 }
539 
540 const struct of_device_id vc4_v3d_dt_match[] = {
541 	{ .compatible = "brcm,bcm2835-v3d" },
542 	{ .compatible = "brcm,cygnus-v3d" },
543 	{ .compatible = "brcm,vc4-v3d" },
544 	{}
545 };
546 
547 struct platform_driver vc4_v3d_driver = {
548 	.probe = vc4_v3d_dev_probe,
549 	.remove_new = vc4_v3d_dev_remove,
550 	.driver = {
551 		.name = "vc4_v3d",
552 		.of_match_table = vc4_v3d_dt_match,
553 		.pm = &vc4_v3d_pm_ops,
554 	},
555 };
556