xref: /openbmc/linux/drivers/gpu/host1x/dev.c (revision 99a15348)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Tegra host1x driver
4  *
5  * Copyright (c) 2010-2013, NVIDIA Corporation.
6  */
7 
8 #include <linux/clk.h>
9 #include <linux/delay.h>
10 #include <linux/dma-mapping.h>
11 #include <linux/io.h>
12 #include <linux/list.h>
13 #include <linux/module.h>
14 #include <linux/of_device.h>
15 #include <linux/of.h>
16 #include <linux/pm_runtime.h>
17 #include <linux/slab.h>
18 
19 #include <soc/tegra/common.h>
20 
21 #define CREATE_TRACE_POINTS
22 #include <trace/events/host1x.h>
23 #undef CREATE_TRACE_POINTS
24 
25 #if IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU)
26 #include <asm/dma-iommu.h>
27 #endif
28 
29 #include "bus.h"
30 #include "channel.h"
31 #include "debug.h"
32 #include "dev.h"
33 #include "intr.h"
34 
35 #include "hw/host1x01.h"
36 #include "hw/host1x02.h"
37 #include "hw/host1x04.h"
38 #include "hw/host1x05.h"
39 #include "hw/host1x06.h"
40 #include "hw/host1x07.h"
41 
42 void host1x_hypervisor_writel(struct host1x *host1x, u32 v, u32 r)
43 {
44 	writel(v, host1x->hv_regs + r);
45 }
46 
47 u32 host1x_hypervisor_readl(struct host1x *host1x, u32 r)
48 {
49 	return readl(host1x->hv_regs + r);
50 }
51 
52 void host1x_sync_writel(struct host1x *host1x, u32 v, u32 r)
53 {
54 	void __iomem *sync_regs = host1x->regs + host1x->info->sync_offset;
55 
56 	writel(v, sync_regs + r);
57 }
58 
59 u32 host1x_sync_readl(struct host1x *host1x, u32 r)
60 {
61 	void __iomem *sync_regs = host1x->regs + host1x->info->sync_offset;
62 
63 	return readl(sync_regs + r);
64 }
65 
66 void host1x_ch_writel(struct host1x_channel *ch, u32 v, u32 r)
67 {
68 	writel(v, ch->regs + r);
69 }
70 
71 u32 host1x_ch_readl(struct host1x_channel *ch, u32 r)
72 {
73 	return readl(ch->regs + r);
74 }
75 
76 static const struct host1x_info host1x01_info = {
77 	.nb_channels = 8,
78 	.nb_pts = 32,
79 	.nb_mlocks = 16,
80 	.nb_bases = 8,
81 	.init = host1x01_init,
82 	.sync_offset = 0x3000,
83 	.dma_mask = DMA_BIT_MASK(32),
84 	.has_wide_gather = false,
85 	.has_hypervisor = false,
86 	.num_sid_entries = 0,
87 	.sid_table = NULL,
88 	.reserve_vblank_syncpts = true,
89 };
90 
91 static const struct host1x_info host1x02_info = {
92 	.nb_channels = 9,
93 	.nb_pts = 32,
94 	.nb_mlocks = 16,
95 	.nb_bases = 12,
96 	.init = host1x02_init,
97 	.sync_offset = 0x3000,
98 	.dma_mask = DMA_BIT_MASK(32),
99 	.has_wide_gather = false,
100 	.has_hypervisor = false,
101 	.num_sid_entries = 0,
102 	.sid_table = NULL,
103 	.reserve_vblank_syncpts = true,
104 };
105 
106 static const struct host1x_info host1x04_info = {
107 	.nb_channels = 12,
108 	.nb_pts = 192,
109 	.nb_mlocks = 16,
110 	.nb_bases = 64,
111 	.init = host1x04_init,
112 	.sync_offset = 0x2100,
113 	.dma_mask = DMA_BIT_MASK(34),
114 	.has_wide_gather = false,
115 	.has_hypervisor = false,
116 	.num_sid_entries = 0,
117 	.sid_table = NULL,
118 	.reserve_vblank_syncpts = false,
119 };
120 
121 static const struct host1x_info host1x05_info = {
122 	.nb_channels = 14,
123 	.nb_pts = 192,
124 	.nb_mlocks = 16,
125 	.nb_bases = 64,
126 	.init = host1x05_init,
127 	.sync_offset = 0x2100,
128 	.dma_mask = DMA_BIT_MASK(34),
129 	.has_wide_gather = false,
130 	.has_hypervisor = false,
131 	.num_sid_entries = 0,
132 	.sid_table = NULL,
133 	.reserve_vblank_syncpts = false,
134 };
135 
136 static const struct host1x_sid_entry tegra186_sid_table[] = {
137 	{
138 		/* VIC */
139 		.base = 0x1af0,
140 		.offset = 0x30,
141 		.limit = 0x34
142 	},
143 	{
144 		/* NVDEC */
145 		.base = 0x1b00,
146 		.offset = 0x30,
147 		.limit = 0x34
148 	},
149 };
150 
151 static const struct host1x_info host1x06_info = {
152 	.nb_channels = 63,
153 	.nb_pts = 576,
154 	.nb_mlocks = 24,
155 	.nb_bases = 16,
156 	.init = host1x06_init,
157 	.sync_offset = 0x0,
158 	.dma_mask = DMA_BIT_MASK(40),
159 	.has_wide_gather = true,
160 	.has_hypervisor = true,
161 	.num_sid_entries = ARRAY_SIZE(tegra186_sid_table),
162 	.sid_table = tegra186_sid_table,
163 	.reserve_vblank_syncpts = false,
164 };
165 
166 static const struct host1x_sid_entry tegra194_sid_table[] = {
167 	{
168 		/* VIC */
169 		.base = 0x1af0,
170 		.offset = 0x30,
171 		.limit = 0x34
172 	},
173 	{
174 		/* NVDEC */
175 		.base = 0x1b00,
176 		.offset = 0x30,
177 		.limit = 0x34
178 	},
179 	{
180 		/* NVDEC1 */
181 		.base = 0x1bc0,
182 		.offset = 0x30,
183 		.limit = 0x34
184 	},
185 };
186 
187 static const struct host1x_info host1x07_info = {
188 	.nb_channels = 63,
189 	.nb_pts = 704,
190 	.nb_mlocks = 32,
191 	.nb_bases = 0,
192 	.init = host1x07_init,
193 	.sync_offset = 0x0,
194 	.dma_mask = DMA_BIT_MASK(40),
195 	.has_wide_gather = true,
196 	.has_hypervisor = true,
197 	.num_sid_entries = ARRAY_SIZE(tegra194_sid_table),
198 	.sid_table = tegra194_sid_table,
199 	.reserve_vblank_syncpts = false,
200 };
201 
202 static const struct of_device_id host1x_of_match[] = {
203 	{ .compatible = "nvidia,tegra194-host1x", .data = &host1x07_info, },
204 	{ .compatible = "nvidia,tegra186-host1x", .data = &host1x06_info, },
205 	{ .compatible = "nvidia,tegra210-host1x", .data = &host1x05_info, },
206 	{ .compatible = "nvidia,tegra124-host1x", .data = &host1x04_info, },
207 	{ .compatible = "nvidia,tegra114-host1x", .data = &host1x02_info, },
208 	{ .compatible = "nvidia,tegra30-host1x", .data = &host1x01_info, },
209 	{ .compatible = "nvidia,tegra20-host1x", .data = &host1x01_info, },
210 	{ },
211 };
212 MODULE_DEVICE_TABLE(of, host1x_of_match);
213 
214 static void host1x_setup_sid_table(struct host1x *host)
215 {
216 	const struct host1x_info *info = host->info;
217 	unsigned int i;
218 
219 	if (!info->has_hypervisor)
220 		return;
221 
222 	for (i = 0; i < info->num_sid_entries; i++) {
223 		const struct host1x_sid_entry *entry = &info->sid_table[i];
224 
225 		host1x_hypervisor_writel(host, entry->offset, entry->base);
226 		host1x_hypervisor_writel(host, entry->limit, entry->base + 4);
227 	}
228 }
229 
230 static bool host1x_wants_iommu(struct host1x *host1x)
231 {
232 	/*
233 	 * If we support addressing a maximum of 32 bits of physical memory
234 	 * and if the host1x firewall is enabled, there's no need to enable
235 	 * IOMMU support. This can happen for example on Tegra20, Tegra30
236 	 * and Tegra114.
237 	 *
238 	 * Tegra124 and later can address up to 34 bits of physical memory and
239 	 * many platforms come equipped with more than 2 GiB of system memory,
240 	 * which requires crossing the 4 GiB boundary. But there's a catch: on
241 	 * SoCs before Tegra186 (i.e. Tegra124 and Tegra210), the host1x can
242 	 * only address up to 32 bits of memory in GATHER opcodes, which means
243 	 * that command buffers need to either be in the first 2 GiB of system
244 	 * memory (which could quickly lead to memory exhaustion), or command
245 	 * buffers need to be treated differently from other buffers (which is
246 	 * not possible with the current ABI).
247 	 *
248 	 * A third option is to use the IOMMU in these cases to make sure all
249 	 * buffers will be mapped into a 32-bit IOVA space that host1x can
250 	 * address. This allows all of the system memory to be used and works
251 	 * within the limitations of the host1x on these SoCs.
252 	 *
253 	 * In summary, default to enable IOMMU on Tegra124 and later. For any
254 	 * of the earlier SoCs, only use the IOMMU for additional safety when
255 	 * the host1x firewall is disabled.
256 	 */
257 	if (host1x->info->dma_mask <= DMA_BIT_MASK(32)) {
258 		if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
259 			return false;
260 	}
261 
262 	return true;
263 }
264 
265 static struct iommu_domain *host1x_iommu_attach(struct host1x *host)
266 {
267 	struct iommu_domain *domain = iommu_get_domain_for_dev(host->dev);
268 	int err;
269 
270 #if IS_ENABLED(CONFIG_ARM_DMA_USE_IOMMU)
271 	if (host->dev->archdata.mapping) {
272 		struct dma_iommu_mapping *mapping =
273 				to_dma_iommu_mapping(host->dev);
274 		arm_iommu_detach_device(host->dev);
275 		arm_iommu_release_mapping(mapping);
276 
277 		domain = iommu_get_domain_for_dev(host->dev);
278 	}
279 #endif
280 
281 	/*
282 	 * We may not always want to enable IOMMU support (for example if the
283 	 * host1x firewall is already enabled and we don't support addressing
284 	 * more than 32 bits of physical memory), so check for that first.
285 	 *
286 	 * Similarly, if host1x is already attached to an IOMMU (via the DMA
287 	 * API), don't try to attach again.
288 	 */
289 	if (!host1x_wants_iommu(host) || domain)
290 		return domain;
291 
292 	host->group = iommu_group_get(host->dev);
293 	if (host->group) {
294 		struct iommu_domain_geometry *geometry;
295 		dma_addr_t start, end;
296 		unsigned long order;
297 
298 		err = iova_cache_get();
299 		if (err < 0)
300 			goto put_group;
301 
302 		host->domain = iommu_domain_alloc(&platform_bus_type);
303 		if (!host->domain) {
304 			err = -ENOMEM;
305 			goto put_cache;
306 		}
307 
308 		err = iommu_attach_group(host->domain, host->group);
309 		if (err) {
310 			if (err == -ENODEV)
311 				err = 0;
312 
313 			goto free_domain;
314 		}
315 
316 		geometry = &host->domain->geometry;
317 		start = geometry->aperture_start & host->info->dma_mask;
318 		end = geometry->aperture_end & host->info->dma_mask;
319 
320 		order = __ffs(host->domain->pgsize_bitmap);
321 		init_iova_domain(&host->iova, 1UL << order, start >> order);
322 		host->iova_end = end;
323 
324 		domain = host->domain;
325 	}
326 
327 	return domain;
328 
329 free_domain:
330 	iommu_domain_free(host->domain);
331 	host->domain = NULL;
332 put_cache:
333 	iova_cache_put();
334 put_group:
335 	iommu_group_put(host->group);
336 	host->group = NULL;
337 
338 	return ERR_PTR(err);
339 }
340 
341 static int host1x_iommu_init(struct host1x *host)
342 {
343 	u64 mask = host->info->dma_mask;
344 	struct iommu_domain *domain;
345 	int err;
346 
347 	domain = host1x_iommu_attach(host);
348 	if (IS_ERR(domain)) {
349 		err = PTR_ERR(domain);
350 		dev_err(host->dev, "failed to attach to IOMMU: %d\n", err);
351 		return err;
352 	}
353 
354 	/*
355 	 * If we're not behind an IOMMU make sure we don't get push buffers
356 	 * that are allocated outside of the range addressable by the GATHER
357 	 * opcode.
358 	 *
359 	 * Newer generations of Tegra (Tegra186 and later) support a wide
360 	 * variant of the GATHER opcode that allows addressing more bits.
361 	 */
362 	if (!domain && !host->info->has_wide_gather)
363 		mask = DMA_BIT_MASK(32);
364 
365 	err = dma_coerce_mask_and_coherent(host->dev, mask);
366 	if (err < 0) {
367 		dev_err(host->dev, "failed to set DMA mask: %d\n", err);
368 		return err;
369 	}
370 
371 	return 0;
372 }
373 
374 static void host1x_iommu_exit(struct host1x *host)
375 {
376 	if (host->domain) {
377 		put_iova_domain(&host->iova);
378 		iommu_detach_group(host->domain, host->group);
379 
380 		iommu_domain_free(host->domain);
381 		host->domain = NULL;
382 
383 		iova_cache_put();
384 
385 		iommu_group_put(host->group);
386 		host->group = NULL;
387 	}
388 }
389 
390 static int host1x_get_resets(struct host1x *host)
391 {
392 	int err;
393 
394 	host->resets[0].id = "mc";
395 	host->resets[1].id = "host1x";
396 	host->nresets = ARRAY_SIZE(host->resets);
397 
398 	err = devm_reset_control_bulk_get_optional_exclusive_released(
399 				host->dev, host->nresets, host->resets);
400 	if (err) {
401 		dev_err(host->dev, "failed to get reset: %d\n", err);
402 		return err;
403 	}
404 
405 	if (WARN_ON(!host->resets[1].rstc))
406 		return -ENOENT;
407 
408 	return 0;
409 }
410 
411 static int host1x_probe(struct platform_device *pdev)
412 {
413 	struct host1x *host;
414 	struct resource *regs, *hv_regs = NULL;
415 	int syncpt_irq;
416 	int err;
417 
418 	host = devm_kzalloc(&pdev->dev, sizeof(*host), GFP_KERNEL);
419 	if (!host)
420 		return -ENOMEM;
421 
422 	host->info = of_device_get_match_data(&pdev->dev);
423 
424 	if (host->info->has_hypervisor) {
425 		regs = platform_get_resource_byname(pdev, IORESOURCE_MEM, "vm");
426 		if (!regs) {
427 			dev_err(&pdev->dev, "failed to get vm registers\n");
428 			return -ENXIO;
429 		}
430 
431 		hv_regs = platform_get_resource_byname(pdev, IORESOURCE_MEM,
432 						       "hypervisor");
433 		if (!hv_regs) {
434 			dev_err(&pdev->dev,
435 				"failed to get hypervisor registers\n");
436 			return -ENXIO;
437 		}
438 	} else {
439 		regs = platform_get_resource(pdev, IORESOURCE_MEM, 0);
440 		if (!regs) {
441 			dev_err(&pdev->dev, "failed to get registers\n");
442 			return -ENXIO;
443 		}
444 	}
445 
446 	syncpt_irq = platform_get_irq(pdev, 0);
447 	if (syncpt_irq < 0)
448 		return syncpt_irq;
449 
450 	mutex_init(&host->devices_lock);
451 	INIT_LIST_HEAD(&host->devices);
452 	INIT_LIST_HEAD(&host->list);
453 	host->dev = &pdev->dev;
454 
455 	/* set common host1x device data */
456 	platform_set_drvdata(pdev, host);
457 
458 	host->regs = devm_ioremap_resource(&pdev->dev, regs);
459 	if (IS_ERR(host->regs))
460 		return PTR_ERR(host->regs);
461 
462 	if (host->info->has_hypervisor) {
463 		host->hv_regs = devm_ioremap_resource(&pdev->dev, hv_regs);
464 		if (IS_ERR(host->hv_regs))
465 			return PTR_ERR(host->hv_regs);
466 	}
467 
468 	host->dev->dma_parms = &host->dma_parms;
469 	dma_set_max_seg_size(host->dev, UINT_MAX);
470 
471 	if (host->info->init) {
472 		err = host->info->init(host);
473 		if (err)
474 			return err;
475 	}
476 
477 	host->clk = devm_clk_get(&pdev->dev, NULL);
478 	if (IS_ERR(host->clk)) {
479 		err = PTR_ERR(host->clk);
480 
481 		if (err != -EPROBE_DEFER)
482 			dev_err(&pdev->dev, "failed to get clock: %d\n", err);
483 
484 		return err;
485 	}
486 
487 	err = host1x_get_resets(host);
488 	if (err)
489 		return err;
490 
491 	host1x_bo_cache_init(&host->cache);
492 
493 	err = host1x_iommu_init(host);
494 	if (err < 0) {
495 		dev_err(&pdev->dev, "failed to setup IOMMU: %d\n", err);
496 		goto destroy_cache;
497 	}
498 
499 	err = host1x_channel_list_init(&host->channel_list,
500 				       host->info->nb_channels);
501 	if (err) {
502 		dev_err(&pdev->dev, "failed to initialize channel list\n");
503 		goto iommu_exit;
504 	}
505 
506 	err = host1x_syncpt_init(host);
507 	if (err) {
508 		dev_err(&pdev->dev, "failed to initialize syncpts\n");
509 		goto free_channels;
510 	}
511 
512 	err = host1x_intr_init(host, syncpt_irq);
513 	if (err) {
514 		dev_err(&pdev->dev, "failed to initialize interrupts\n");
515 		goto deinit_syncpt;
516 	}
517 
518 	pm_runtime_enable(&pdev->dev);
519 
520 	err = devm_tegra_core_dev_init_opp_table_common(&pdev->dev);
521 	if (err)
522 		goto pm_disable;
523 
524 	/* the driver's code isn't ready yet for the dynamic RPM */
525 	err = pm_runtime_resume_and_get(&pdev->dev);
526 	if (err)
527 		goto pm_disable;
528 
529 	host1x_debug_init(host);
530 
531 	err = host1x_register(host);
532 	if (err < 0)
533 		goto deinit_debugfs;
534 
535 	err = devm_of_platform_populate(&pdev->dev);
536 	if (err < 0)
537 		goto unregister;
538 
539 	return 0;
540 
541 unregister:
542 	host1x_unregister(host);
543 deinit_debugfs:
544 	host1x_debug_deinit(host);
545 
546 	pm_runtime_put_sync_suspend(&pdev->dev);
547 pm_disable:
548 	pm_runtime_disable(&pdev->dev);
549 
550 	host1x_intr_deinit(host);
551 deinit_syncpt:
552 	host1x_syncpt_deinit(host);
553 free_channels:
554 	host1x_channel_list_free(&host->channel_list);
555 iommu_exit:
556 	host1x_iommu_exit(host);
557 destroy_cache:
558 	host1x_bo_cache_destroy(&host->cache);
559 
560 	return err;
561 }
562 
563 static int host1x_remove(struct platform_device *pdev)
564 {
565 	struct host1x *host = platform_get_drvdata(pdev);
566 
567 	host1x_unregister(host);
568 	host1x_debug_deinit(host);
569 
570 	pm_runtime_force_suspend(&pdev->dev);
571 
572 	host1x_intr_deinit(host);
573 	host1x_syncpt_deinit(host);
574 	host1x_channel_list_free(&host->channel_list);
575 	host1x_iommu_exit(host);
576 	host1x_bo_cache_destroy(&host->cache);
577 
578 	return 0;
579 }
580 
581 static int __maybe_unused host1x_runtime_suspend(struct device *dev)
582 {
583 	struct host1x *host = dev_get_drvdata(dev);
584 	int err;
585 
586 	host1x_intr_stop(host);
587 	host1x_syncpt_save(host);
588 
589 	err = reset_control_bulk_assert(host->nresets, host->resets);
590 	if (err) {
591 		dev_err(dev, "failed to assert reset: %d\n", err);
592 		goto resume_host1x;
593 	}
594 
595 	usleep_range(1000, 2000);
596 
597 	clk_disable_unprepare(host->clk);
598 	reset_control_bulk_release(host->nresets, host->resets);
599 
600 	return 0;
601 
602 resume_host1x:
603 	host1x_setup_sid_table(host);
604 	host1x_syncpt_restore(host);
605 	host1x_intr_start(host);
606 
607 	return err;
608 }
609 
610 static int __maybe_unused host1x_runtime_resume(struct device *dev)
611 {
612 	struct host1x *host = dev_get_drvdata(dev);
613 	int err;
614 
615 	err = reset_control_bulk_acquire(host->nresets, host->resets);
616 	if (err) {
617 		dev_err(dev, "failed to acquire reset: %d\n", err);
618 		return err;
619 	}
620 
621 	err = clk_prepare_enable(host->clk);
622 	if (err) {
623 		dev_err(dev, "failed to enable clock: %d\n", err);
624 		goto release_reset;
625 	}
626 
627 	err = reset_control_bulk_deassert(host->nresets, host->resets);
628 	if (err < 0) {
629 		dev_err(dev, "failed to deassert reset: %d\n", err);
630 		goto disable_clk;
631 	}
632 
633 	host1x_setup_sid_table(host);
634 	host1x_syncpt_restore(host);
635 	host1x_intr_start(host);
636 
637 	return 0;
638 
639 disable_clk:
640 	clk_disable_unprepare(host->clk);
641 release_reset:
642 	reset_control_bulk_release(host->nresets, host->resets);
643 
644 	return err;
645 }
646 
647 static const struct dev_pm_ops host1x_pm_ops = {
648 	SET_RUNTIME_PM_OPS(host1x_runtime_suspend, host1x_runtime_resume,
649 			   NULL)
650 	/* TODO: add system suspend-resume once driver will be ready for that */
651 };
652 
653 static struct platform_driver tegra_host1x_driver = {
654 	.driver = {
655 		.name = "tegra-host1x",
656 		.of_match_table = host1x_of_match,
657 		.pm = &host1x_pm_ops,
658 	},
659 	.probe = host1x_probe,
660 	.remove = host1x_remove,
661 };
662 
663 static struct platform_driver * const drivers[] = {
664 	&tegra_host1x_driver,
665 	&tegra_mipi_driver,
666 };
667 
668 static int __init tegra_host1x_init(void)
669 {
670 	int err;
671 
672 	err = bus_register(&host1x_bus_type);
673 	if (err < 0)
674 		return err;
675 
676 	err = platform_register_drivers(drivers, ARRAY_SIZE(drivers));
677 	if (err < 0)
678 		bus_unregister(&host1x_bus_type);
679 
680 	return err;
681 }
682 module_init(tegra_host1x_init);
683 
684 static void __exit tegra_host1x_exit(void)
685 {
686 	platform_unregister_drivers(drivers, ARRAY_SIZE(drivers));
687 	bus_unregister(&host1x_bus_type);
688 }
689 module_exit(tegra_host1x_exit);
690 
691 /**
692  * host1x_get_dma_mask() - query the supported DMA mask for host1x
693  * @host1x: host1x instance
694  *
695  * Note that this returns the supported DMA mask for host1x, which can be
696  * different from the applicable DMA mask under certain circumstances.
697  */
698 u64 host1x_get_dma_mask(struct host1x *host1x)
699 {
700 	return host1x->info->dma_mask;
701 }
702 EXPORT_SYMBOL(host1x_get_dma_mask);
703 
704 MODULE_AUTHOR("Thierry Reding <thierry.reding@avionic-design.de>");
705 MODULE_AUTHOR("Terje Bergstrom <tbergstrom@nvidia.com>");
706 MODULE_DESCRIPTION("Host1x driver for Tegra products");
707 MODULE_LICENSE("GPL");
708