xref: /openbmc/linux/drivers/gpu/drm/vc4/vc4_v3d.c (revision 423997ff)
1 /*
2  * Copyright (c) 2014 The Linux Foundation. All rights reserved.
3  * Copyright (C) 2013 Red Hat
4  * Author: Rob Clark <robdclark@gmail.com>
5  *
6  * This program is free software; you can redistribute it and/or modify it
7  * under the terms of the GNU General Public License version 2 as published by
8  * the Free Software Foundation.
9  *
10  * This program is distributed in the hope that it will be useful, but WITHOUT
11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
13  * more details.
14  *
15  * You should have received a copy of the GNU General Public License along with
16  * this program.  If not, see <http://www.gnu.org/licenses/>.
17  */
18 
19 #include <linux/clk.h>
20 #include <linux/component.h>
21 #include <linux/pm_runtime.h>
22 #include "vc4_drv.h"
23 #include "vc4_regs.h"
24 
25 static const struct debugfs_reg32 v3d_regs[] = {
26 	VC4_REG32(V3D_IDENT0),
27 	VC4_REG32(V3D_IDENT1),
28 	VC4_REG32(V3D_IDENT2),
29 	VC4_REG32(V3D_SCRATCH),
30 	VC4_REG32(V3D_L2CACTL),
31 	VC4_REG32(V3D_SLCACTL),
32 	VC4_REG32(V3D_INTCTL),
33 	VC4_REG32(V3D_INTENA),
34 	VC4_REG32(V3D_INTDIS),
35 	VC4_REG32(V3D_CT0CS),
36 	VC4_REG32(V3D_CT1CS),
37 	VC4_REG32(V3D_CT0EA),
38 	VC4_REG32(V3D_CT1EA),
39 	VC4_REG32(V3D_CT0CA),
40 	VC4_REG32(V3D_CT1CA),
41 	VC4_REG32(V3D_CT00RA0),
42 	VC4_REG32(V3D_CT01RA0),
43 	VC4_REG32(V3D_CT0LC),
44 	VC4_REG32(V3D_CT1LC),
45 	VC4_REG32(V3D_CT0PC),
46 	VC4_REG32(V3D_CT1PC),
47 	VC4_REG32(V3D_PCS),
48 	VC4_REG32(V3D_BFC),
49 	VC4_REG32(V3D_RFC),
50 	VC4_REG32(V3D_BPCA),
51 	VC4_REG32(V3D_BPCS),
52 	VC4_REG32(V3D_BPOA),
53 	VC4_REG32(V3D_BPOS),
54 	VC4_REG32(V3D_BXCF),
55 	VC4_REG32(V3D_SQRSV0),
56 	VC4_REG32(V3D_SQRSV1),
57 	VC4_REG32(V3D_SQCNTL),
58 	VC4_REG32(V3D_SRQPC),
59 	VC4_REG32(V3D_SRQUA),
60 	VC4_REG32(V3D_SRQUL),
61 	VC4_REG32(V3D_SRQCS),
62 	VC4_REG32(V3D_VPACNTL),
63 	VC4_REG32(V3D_VPMBASE),
64 	VC4_REG32(V3D_PCTRC),
65 	VC4_REG32(V3D_PCTRE),
66 	VC4_REG32(V3D_PCTR(0)),
67 	VC4_REG32(V3D_PCTRS(0)),
68 	VC4_REG32(V3D_PCTR(1)),
69 	VC4_REG32(V3D_PCTRS(1)),
70 	VC4_REG32(V3D_PCTR(2)),
71 	VC4_REG32(V3D_PCTRS(2)),
72 	VC4_REG32(V3D_PCTR(3)),
73 	VC4_REG32(V3D_PCTRS(3)),
74 	VC4_REG32(V3D_PCTR(4)),
75 	VC4_REG32(V3D_PCTRS(4)),
76 	VC4_REG32(V3D_PCTR(5)),
77 	VC4_REG32(V3D_PCTRS(5)),
78 	VC4_REG32(V3D_PCTR(6)),
79 	VC4_REG32(V3D_PCTRS(6)),
80 	VC4_REG32(V3D_PCTR(7)),
81 	VC4_REG32(V3D_PCTRS(7)),
82 	VC4_REG32(V3D_PCTR(8)),
83 	VC4_REG32(V3D_PCTRS(8)),
84 	VC4_REG32(V3D_PCTR(9)),
85 	VC4_REG32(V3D_PCTRS(9)),
86 	VC4_REG32(V3D_PCTR(10)),
87 	VC4_REG32(V3D_PCTRS(10)),
88 	VC4_REG32(V3D_PCTR(11)),
89 	VC4_REG32(V3D_PCTRS(11)),
90 	VC4_REG32(V3D_PCTR(12)),
91 	VC4_REG32(V3D_PCTRS(12)),
92 	VC4_REG32(V3D_PCTR(13)),
93 	VC4_REG32(V3D_PCTRS(13)),
94 	VC4_REG32(V3D_PCTR(14)),
95 	VC4_REG32(V3D_PCTRS(14)),
96 	VC4_REG32(V3D_PCTR(15)),
97 	VC4_REG32(V3D_PCTRS(15)),
98 	VC4_REG32(V3D_DBGE),
99 	VC4_REG32(V3D_FDBGO),
100 	VC4_REG32(V3D_FDBGB),
101 	VC4_REG32(V3D_FDBGR),
102 	VC4_REG32(V3D_FDBGS),
103 	VC4_REG32(V3D_ERRSTAT),
104 };
105 
106 static int vc4_v3d_debugfs_ident(struct seq_file *m, void *unused)
107 {
108 	struct drm_info_node *node = (struct drm_info_node *)m->private;
109 	struct drm_device *dev = node->minor->dev;
110 	struct vc4_dev *vc4 = to_vc4_dev(dev);
111 	int ret = vc4_v3d_pm_get(vc4);
112 
113 	if (ret == 0) {
114 		uint32_t ident1 = V3D_READ(V3D_IDENT1);
115 		uint32_t nslc = VC4_GET_FIELD(ident1, V3D_IDENT1_NSLC);
116 		uint32_t tups = VC4_GET_FIELD(ident1, V3D_IDENT1_TUPS);
117 		uint32_t qups = VC4_GET_FIELD(ident1, V3D_IDENT1_QUPS);
118 
119 		seq_printf(m, "Revision:   %d\n",
120 			   VC4_GET_FIELD(ident1, V3D_IDENT1_REV));
121 		seq_printf(m, "Slices:     %d\n", nslc);
122 		seq_printf(m, "TMUs:       %d\n", nslc * tups);
123 		seq_printf(m, "QPUs:       %d\n", nslc * qups);
124 		seq_printf(m, "Semaphores: %d\n",
125 			   VC4_GET_FIELD(ident1, V3D_IDENT1_NSEM));
126 		vc4_v3d_pm_put(vc4);
127 	}
128 
129 	return 0;
130 }
131 
132 /**
133  * Wraps pm_runtime_get_sync() in a refcount, so that we can reliably
134  * get the pm_runtime refcount to 0 in vc4_reset().
135  */
136 int
137 vc4_v3d_pm_get(struct vc4_dev *vc4)
138 {
139 	mutex_lock(&vc4->power_lock);
140 	if (vc4->power_refcount++ == 0) {
141 		int ret = pm_runtime_get_sync(&vc4->v3d->pdev->dev);
142 
143 		if (ret < 0) {
144 			vc4->power_refcount--;
145 			mutex_unlock(&vc4->power_lock);
146 			return ret;
147 		}
148 	}
149 	mutex_unlock(&vc4->power_lock);
150 
151 	return 0;
152 }
153 
154 void
155 vc4_v3d_pm_put(struct vc4_dev *vc4)
156 {
157 	mutex_lock(&vc4->power_lock);
158 	if (--vc4->power_refcount == 0) {
159 		pm_runtime_mark_last_busy(&vc4->v3d->pdev->dev);
160 		pm_runtime_put_autosuspend(&vc4->v3d->pdev->dev);
161 	}
162 	mutex_unlock(&vc4->power_lock);
163 }
164 
165 static void vc4_v3d_init_hw(struct drm_device *dev)
166 {
167 	struct vc4_dev *vc4 = to_vc4_dev(dev);
168 
169 	/* Take all the memory that would have been reserved for user
170 	 * QPU programs, since we don't have an interface for running
171 	 * them, anyway.
172 	 */
173 	V3D_WRITE(V3D_VPMBASE, 0);
174 }
175 
176 int vc4_v3d_get_bin_slot(struct vc4_dev *vc4)
177 {
178 	struct drm_device *dev = vc4->dev;
179 	unsigned long irqflags;
180 	int slot;
181 	uint64_t seqno = 0;
182 	struct vc4_exec_info *exec;
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 CMA 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->dev, 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 CMA 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.paddr & 0xf0000000) ==
272 		    ((bo->base.paddr + 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_unlocked(&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 	mutex_lock(&vc4->bin_bo_lock);
329 
330 	if (used && *used)
331 		goto complete;
332 
333 	if (vc4->bin_bo)
334 		kref_get(&vc4->bin_bo_kref);
335 	else
336 		ret = bin_bo_alloc(vc4);
337 
338 	if (ret == 0 && used)
339 		*used = true;
340 
341 complete:
342 	mutex_unlock(&vc4->bin_bo_lock);
343 
344 	return ret;
345 }
346 
347 static void bin_bo_release(struct kref *ref)
348 {
349 	struct vc4_dev *vc4 = container_of(ref, struct vc4_dev, bin_bo_kref);
350 
351 	if (WARN_ON_ONCE(!vc4->bin_bo))
352 		return;
353 
354 	drm_gem_object_put_unlocked(&vc4->bin_bo->base.base);
355 	vc4->bin_bo = NULL;
356 }
357 
358 void vc4_v3d_bin_bo_put(struct vc4_dev *vc4)
359 {
360 	mutex_lock(&vc4->bin_bo_lock);
361 	kref_put(&vc4->bin_bo_kref, bin_bo_release);
362 	mutex_unlock(&vc4->bin_bo_lock);
363 }
364 
365 #ifdef CONFIG_PM
366 static int vc4_v3d_runtime_suspend(struct device *dev)
367 {
368 	struct vc4_v3d *v3d = dev_get_drvdata(dev);
369 	struct vc4_dev *vc4 = v3d->vc4;
370 
371 	vc4_irq_uninstall(vc4->dev);
372 
373 	clk_disable_unprepare(v3d->clk);
374 
375 	return 0;
376 }
377 
378 static int vc4_v3d_runtime_resume(struct device *dev)
379 {
380 	struct vc4_v3d *v3d = dev_get_drvdata(dev);
381 	struct vc4_dev *vc4 = v3d->vc4;
382 	int ret;
383 
384 	ret = clk_prepare_enable(v3d->clk);
385 	if (ret != 0)
386 		return ret;
387 
388 	vc4_v3d_init_hw(vc4->dev);
389 
390 	/* We disabled the IRQ as part of vc4_irq_uninstall in suspend. */
391 	enable_irq(vc4->dev->irq);
392 	vc4_irq_postinstall(vc4->dev);
393 
394 	return 0;
395 }
396 #endif
397 
398 static int vc4_v3d_bind(struct device *dev, struct device *master, void *data)
399 {
400 	struct platform_device *pdev = to_platform_device(dev);
401 	struct drm_device *drm = dev_get_drvdata(master);
402 	struct vc4_dev *vc4 = to_vc4_dev(drm);
403 	struct vc4_v3d *v3d = NULL;
404 	int ret;
405 
406 	v3d = devm_kzalloc(&pdev->dev, sizeof(*v3d), GFP_KERNEL);
407 	if (!v3d)
408 		return -ENOMEM;
409 
410 	dev_set_drvdata(dev, v3d);
411 
412 	v3d->pdev = pdev;
413 
414 	v3d->regs = vc4_ioremap_regs(pdev, 0);
415 	if (IS_ERR(v3d->regs))
416 		return PTR_ERR(v3d->regs);
417 	v3d->regset.base = v3d->regs;
418 	v3d->regset.regs = v3d_regs;
419 	v3d->regset.nregs = ARRAY_SIZE(v3d_regs);
420 
421 	vc4->v3d = v3d;
422 	v3d->vc4 = vc4;
423 
424 	v3d->clk = devm_clk_get(dev, NULL);
425 	if (IS_ERR(v3d->clk)) {
426 		int ret = PTR_ERR(v3d->clk);
427 
428 		if (ret == -ENOENT) {
429 			/* bcm2835 didn't have a clock reference in the DT. */
430 			ret = 0;
431 			v3d->clk = NULL;
432 		} else {
433 			if (ret != -EPROBE_DEFER)
434 				dev_err(dev, "Failed to get V3D clock: %d\n",
435 					ret);
436 			return ret;
437 		}
438 	}
439 
440 	if (V3D_READ(V3D_IDENT0) != V3D_EXPECTED_IDENT0) {
441 		DRM_ERROR("V3D_IDENT0 read 0x%08x instead of 0x%08x\n",
442 			  V3D_READ(V3D_IDENT0), V3D_EXPECTED_IDENT0);
443 		return -EINVAL;
444 	}
445 
446 	ret = clk_prepare_enable(v3d->clk);
447 	if (ret != 0)
448 		return ret;
449 
450 	/* Reset the binner overflow address/size at setup, to be sure
451 	 * we don't reuse an old one.
452 	 */
453 	V3D_WRITE(V3D_BPOA, 0);
454 	V3D_WRITE(V3D_BPOS, 0);
455 
456 	vc4_v3d_init_hw(drm);
457 
458 	ret = drm_irq_install(drm, platform_get_irq(pdev, 0));
459 	if (ret) {
460 		DRM_ERROR("Failed to install IRQ handler\n");
461 		return ret;
462 	}
463 
464 	pm_runtime_set_active(dev);
465 	pm_runtime_use_autosuspend(dev);
466 	pm_runtime_set_autosuspend_delay(dev, 40); /* a little over 2 frames. */
467 	pm_runtime_enable(dev);
468 
469 	vc4_debugfs_add_file(drm, "v3d_ident", vc4_v3d_debugfs_ident, NULL);
470 	vc4_debugfs_add_regset32(drm, "v3d_regs", &v3d->regset);
471 
472 	return 0;
473 }
474 
475 static void vc4_v3d_unbind(struct device *dev, struct device *master,
476 			   void *data)
477 {
478 	struct drm_device *drm = dev_get_drvdata(master);
479 	struct vc4_dev *vc4 = to_vc4_dev(drm);
480 
481 	pm_runtime_disable(dev);
482 
483 	drm_irq_uninstall(drm);
484 
485 	/* Disable the binner's overflow memory address, so the next
486 	 * driver probe (if any) doesn't try to reuse our old
487 	 * allocation.
488 	 */
489 	V3D_WRITE(V3D_BPOA, 0);
490 	V3D_WRITE(V3D_BPOS, 0);
491 
492 	vc4->v3d = NULL;
493 }
494 
495 static const struct dev_pm_ops vc4_v3d_pm_ops = {
496 	SET_RUNTIME_PM_OPS(vc4_v3d_runtime_suspend, vc4_v3d_runtime_resume, NULL)
497 };
498 
499 static const struct component_ops vc4_v3d_ops = {
500 	.bind   = vc4_v3d_bind,
501 	.unbind = vc4_v3d_unbind,
502 };
503 
504 static int vc4_v3d_dev_probe(struct platform_device *pdev)
505 {
506 	return component_add(&pdev->dev, &vc4_v3d_ops);
507 }
508 
509 static int vc4_v3d_dev_remove(struct platform_device *pdev)
510 {
511 	component_del(&pdev->dev, &vc4_v3d_ops);
512 	return 0;
513 }
514 
515 const struct of_device_id vc4_v3d_dt_match[] = {
516 	{ .compatible = "brcm,bcm2835-v3d" },
517 	{ .compatible = "brcm,cygnus-v3d" },
518 	{ .compatible = "brcm,vc4-v3d" },
519 	{}
520 };
521 
522 struct platform_driver vc4_v3d_driver = {
523 	.probe = vc4_v3d_dev_probe,
524 	.remove = vc4_v3d_dev_remove,
525 	.driver = {
526 		.name = "vc4_v3d",
527 		.of_match_table = vc4_v3d_dt_match,
528 		.pm = &vc4_v3d_pm_ops,
529 	},
530 };
531