xref: /openbmc/linux/drivers/gpu/host1x/dev.c (revision 82df5b73)
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/dma-mapping.h>
10 #include <linux/io.h>
11 #include <linux/list.h>
12 #include <linux/module.h>
13 #include <linux/of_device.h>
14 #include <linux/of.h>
15 #include <linux/slab.h>
16 
17 #define CREATE_TRACE_POINTS
18 #include <trace/events/host1x.h>
19 #undef CREATE_TRACE_POINTS
20 
21 #include "bus.h"
22 #include "channel.h"
23 #include "debug.h"
24 #include "dev.h"
25 #include "intr.h"
26 
27 #include "hw/host1x01.h"
28 #include "hw/host1x02.h"
29 #include "hw/host1x04.h"
30 #include "hw/host1x05.h"
31 #include "hw/host1x06.h"
32 #include "hw/host1x07.h"
33 
34 void host1x_hypervisor_writel(struct host1x *host1x, u32 v, u32 r)
35 {
36 	writel(v, host1x->hv_regs + r);
37 }
38 
39 u32 host1x_hypervisor_readl(struct host1x *host1x, u32 r)
40 {
41 	return readl(host1x->hv_regs + r);
42 }
43 
44 void host1x_sync_writel(struct host1x *host1x, u32 v, u32 r)
45 {
46 	void __iomem *sync_regs = host1x->regs + host1x->info->sync_offset;
47 
48 	writel(v, sync_regs + r);
49 }
50 
51 u32 host1x_sync_readl(struct host1x *host1x, u32 r)
52 {
53 	void __iomem *sync_regs = host1x->regs + host1x->info->sync_offset;
54 
55 	return readl(sync_regs + r);
56 }
57 
58 void host1x_ch_writel(struct host1x_channel *ch, u32 v, u32 r)
59 {
60 	writel(v, ch->regs + r);
61 }
62 
63 u32 host1x_ch_readl(struct host1x_channel *ch, u32 r)
64 {
65 	return readl(ch->regs + r);
66 }
67 
68 static const struct host1x_info host1x01_info = {
69 	.nb_channels = 8,
70 	.nb_pts = 32,
71 	.nb_mlocks = 16,
72 	.nb_bases = 8,
73 	.init = host1x01_init,
74 	.sync_offset = 0x3000,
75 	.dma_mask = DMA_BIT_MASK(32),
76 	.has_wide_gather = false,
77 	.has_hypervisor = false,
78 	.num_sid_entries = 0,
79 	.sid_table = NULL,
80 };
81 
82 static const struct host1x_info host1x02_info = {
83 	.nb_channels = 9,
84 	.nb_pts = 32,
85 	.nb_mlocks = 16,
86 	.nb_bases = 12,
87 	.init = host1x02_init,
88 	.sync_offset = 0x3000,
89 	.dma_mask = DMA_BIT_MASK(32),
90 	.has_wide_gather = false,
91 	.has_hypervisor = false,
92 	.num_sid_entries = 0,
93 	.sid_table = NULL,
94 };
95 
96 static const struct host1x_info host1x04_info = {
97 	.nb_channels = 12,
98 	.nb_pts = 192,
99 	.nb_mlocks = 16,
100 	.nb_bases = 64,
101 	.init = host1x04_init,
102 	.sync_offset = 0x2100,
103 	.dma_mask = DMA_BIT_MASK(34),
104 	.has_wide_gather = false,
105 	.has_hypervisor = false,
106 	.num_sid_entries = 0,
107 	.sid_table = NULL,
108 };
109 
110 static const struct host1x_info host1x05_info = {
111 	.nb_channels = 14,
112 	.nb_pts = 192,
113 	.nb_mlocks = 16,
114 	.nb_bases = 64,
115 	.init = host1x05_init,
116 	.sync_offset = 0x2100,
117 	.dma_mask = DMA_BIT_MASK(34),
118 	.has_wide_gather = false,
119 	.has_hypervisor = false,
120 	.num_sid_entries = 0,
121 	.sid_table = NULL,
122 };
123 
124 static const struct host1x_sid_entry tegra186_sid_table[] = {
125 	{
126 		/* VIC */
127 		.base = 0x1af0,
128 		.offset = 0x30,
129 		.limit = 0x34
130 	},
131 };
132 
133 static const struct host1x_info host1x06_info = {
134 	.nb_channels = 63,
135 	.nb_pts = 576,
136 	.nb_mlocks = 24,
137 	.nb_bases = 16,
138 	.init = host1x06_init,
139 	.sync_offset = 0x0,
140 	.dma_mask = DMA_BIT_MASK(40),
141 	.has_wide_gather = true,
142 	.has_hypervisor = true,
143 	.num_sid_entries = ARRAY_SIZE(tegra186_sid_table),
144 	.sid_table = tegra186_sid_table,
145 };
146 
147 static const struct host1x_sid_entry tegra194_sid_table[] = {
148 	{
149 		/* VIC */
150 		.base = 0x1af0,
151 		.offset = 0x30,
152 		.limit = 0x34
153 	},
154 };
155 
156 static const struct host1x_info host1x07_info = {
157 	.nb_channels = 63,
158 	.nb_pts = 704,
159 	.nb_mlocks = 32,
160 	.nb_bases = 0,
161 	.init = host1x07_init,
162 	.sync_offset = 0x0,
163 	.dma_mask = DMA_BIT_MASK(40),
164 	.has_wide_gather = true,
165 	.has_hypervisor = true,
166 	.num_sid_entries = ARRAY_SIZE(tegra194_sid_table),
167 	.sid_table = tegra194_sid_table,
168 };
169 
170 static const struct of_device_id host1x_of_match[] = {
171 	{ .compatible = "nvidia,tegra194-host1x", .data = &host1x07_info, },
172 	{ .compatible = "nvidia,tegra186-host1x", .data = &host1x06_info, },
173 	{ .compatible = "nvidia,tegra210-host1x", .data = &host1x05_info, },
174 	{ .compatible = "nvidia,tegra124-host1x", .data = &host1x04_info, },
175 	{ .compatible = "nvidia,tegra114-host1x", .data = &host1x02_info, },
176 	{ .compatible = "nvidia,tegra30-host1x", .data = &host1x01_info, },
177 	{ .compatible = "nvidia,tegra20-host1x", .data = &host1x01_info, },
178 	{ },
179 };
180 MODULE_DEVICE_TABLE(of, host1x_of_match);
181 
182 static void host1x_setup_sid_table(struct host1x *host)
183 {
184 	const struct host1x_info *info = host->info;
185 	unsigned int i;
186 
187 	for (i = 0; i < info->num_sid_entries; i++) {
188 		const struct host1x_sid_entry *entry = &info->sid_table[i];
189 
190 		host1x_hypervisor_writel(host, entry->offset, entry->base);
191 		host1x_hypervisor_writel(host, entry->limit, entry->base + 4);
192 	}
193 }
194 
195 static bool host1x_wants_iommu(struct host1x *host1x)
196 {
197 	/*
198 	 * If we support addressing a maximum of 32 bits of physical memory
199 	 * and if the host1x firewall is enabled, there's no need to enable
200 	 * IOMMU support. This can happen for example on Tegra20, Tegra30
201 	 * and Tegra114.
202 	 *
203 	 * Tegra124 and later can address up to 34 bits of physical memory and
204 	 * many platforms come equipped with more than 2 GiB of system memory,
205 	 * which requires crossing the 4 GiB boundary. But there's a catch: on
206 	 * SoCs before Tegra186 (i.e. Tegra124 and Tegra210), the host1x can
207 	 * only address up to 32 bits of memory in GATHER opcodes, which means
208 	 * that command buffers need to either be in the first 2 GiB of system
209 	 * memory (which could quickly lead to memory exhaustion), or command
210 	 * buffers need to be treated differently from other buffers (which is
211 	 * not possible with the current ABI).
212 	 *
213 	 * A third option is to use the IOMMU in these cases to make sure all
214 	 * buffers will be mapped into a 32-bit IOVA space that host1x can
215 	 * address. This allows all of the system memory to be used and works
216 	 * within the limitations of the host1x on these SoCs.
217 	 *
218 	 * In summary, default to enable IOMMU on Tegra124 and later. For any
219 	 * of the earlier SoCs, only use the IOMMU for additional safety when
220 	 * the host1x firewall is disabled.
221 	 */
222 	if (host1x->info->dma_mask <= DMA_BIT_MASK(32)) {
223 		if (IS_ENABLED(CONFIG_TEGRA_HOST1X_FIREWALL))
224 			return false;
225 	}
226 
227 	return true;
228 }
229 
230 static struct iommu_domain *host1x_iommu_attach(struct host1x *host)
231 {
232 	struct iommu_domain *domain = iommu_get_domain_for_dev(host->dev);
233 	int err;
234 
235 	/*
236 	 * We may not always want to enable IOMMU support (for example if the
237 	 * host1x firewall is already enabled and we don't support addressing
238 	 * more than 32 bits of physical memory), so check for that first.
239 	 *
240 	 * Similarly, if host1x is already attached to an IOMMU (via the DMA
241 	 * API), don't try to attach again.
242 	 */
243 	if (!host1x_wants_iommu(host) || domain)
244 		return domain;
245 
246 	host->group = iommu_group_get(host->dev);
247 	if (host->group) {
248 		struct iommu_domain_geometry *geometry;
249 		dma_addr_t start, end;
250 		unsigned long order;
251 
252 		err = iova_cache_get();
253 		if (err < 0)
254 			goto put_group;
255 
256 		host->domain = iommu_domain_alloc(&platform_bus_type);
257 		if (!host->domain) {
258 			err = -ENOMEM;
259 			goto put_cache;
260 		}
261 
262 		err = iommu_attach_group(host->domain, host->group);
263 		if (err) {
264 			if (err == -ENODEV)
265 				err = 0;
266 
267 			goto free_domain;
268 		}
269 
270 		geometry = &host->domain->geometry;
271 		start = geometry->aperture_start & host->info->dma_mask;
272 		end = geometry->aperture_end & host->info->dma_mask;
273 
274 		order = __ffs(host->domain->pgsize_bitmap);
275 		init_iova_domain(&host->iova, 1UL << order, start >> order);
276 		host->iova_end = end;
277 
278 		domain = host->domain;
279 	}
280 
281 	return domain;
282 
283 free_domain:
284 	iommu_domain_free(host->domain);
285 	host->domain = NULL;
286 put_cache:
287 	iova_cache_put();
288 put_group:
289 	iommu_group_put(host->group);
290 	host->group = NULL;
291 
292 	return ERR_PTR(err);
293 }
294 
295 static int host1x_iommu_init(struct host1x *host)
296 {
297 	u64 mask = host->info->dma_mask;
298 	struct iommu_domain *domain;
299 	int err;
300 
301 	domain = host1x_iommu_attach(host);
302 	if (IS_ERR(domain)) {
303 		err = PTR_ERR(domain);
304 		dev_err(host->dev, "failed to attach to IOMMU: %d\n", err);
305 		return err;
306 	}
307 
308 	/*
309 	 * If we're not behind an IOMMU make sure we don't get push buffers
310 	 * that are allocated outside of the range addressable by the GATHER
311 	 * opcode.
312 	 *
313 	 * Newer generations of Tegra (Tegra186 and later) support a wide
314 	 * variant of the GATHER opcode that allows addressing more bits.
315 	 */
316 	if (!domain && !host->info->has_wide_gather)
317 		mask = DMA_BIT_MASK(32);
318 
319 	err = dma_coerce_mask_and_coherent(host->dev, mask);
320 	if (err < 0) {
321 		dev_err(host->dev, "failed to set DMA mask: %d\n", err);
322 		return err;
323 	}
324 
325 	return 0;
326 }
327 
328 static void host1x_iommu_exit(struct host1x *host)
329 {
330 	if (host->domain) {
331 		put_iova_domain(&host->iova);
332 		iommu_detach_group(host->domain, host->group);
333 
334 		iommu_domain_free(host->domain);
335 		host->domain = NULL;
336 
337 		iova_cache_put();
338 
339 		iommu_group_put(host->group);
340 		host->group = NULL;
341 	}
342 }
343 
344 static int host1x_probe(struct platform_device *pdev)
345 {
346 	struct host1x *host;
347 	struct resource *regs, *hv_regs = NULL;
348 	int syncpt_irq;
349 	int err;
350 
351 	host = devm_kzalloc(&pdev->dev, sizeof(*host), GFP_KERNEL);
352 	if (!host)
353 		return -ENOMEM;
354 
355 	host->info = of_device_get_match_data(&pdev->dev);
356 
357 	if (host->info->has_hypervisor) {
358 		regs = platform_get_resource_byname(pdev, IORESOURCE_MEM, "vm");
359 		if (!regs) {
360 			dev_err(&pdev->dev, "failed to get vm registers\n");
361 			return -ENXIO;
362 		}
363 
364 		hv_regs = platform_get_resource_byname(pdev, IORESOURCE_MEM,
365 						       "hypervisor");
366 		if (!hv_regs) {
367 			dev_err(&pdev->dev,
368 				"failed to get hypervisor registers\n");
369 			return -ENXIO;
370 		}
371 	} else {
372 		regs = platform_get_resource(pdev, IORESOURCE_MEM, 0);
373 		if (!regs) {
374 			dev_err(&pdev->dev, "failed to get registers\n");
375 			return -ENXIO;
376 		}
377 	}
378 
379 	syncpt_irq = platform_get_irq(pdev, 0);
380 	if (syncpt_irq < 0)
381 		return syncpt_irq;
382 
383 	mutex_init(&host->devices_lock);
384 	INIT_LIST_HEAD(&host->devices);
385 	INIT_LIST_HEAD(&host->list);
386 	host->dev = &pdev->dev;
387 
388 	/* set common host1x device data */
389 	platform_set_drvdata(pdev, host);
390 
391 	host->regs = devm_ioremap_resource(&pdev->dev, regs);
392 	if (IS_ERR(host->regs))
393 		return PTR_ERR(host->regs);
394 
395 	if (host->info->has_hypervisor) {
396 		host->hv_regs = devm_ioremap_resource(&pdev->dev, hv_regs);
397 		if (IS_ERR(host->hv_regs))
398 			return PTR_ERR(host->hv_regs);
399 	}
400 
401 	host->dev->dma_parms = &host->dma_parms;
402 	dma_set_max_seg_size(host->dev, UINT_MAX);
403 
404 	if (host->info->init) {
405 		err = host->info->init(host);
406 		if (err)
407 			return err;
408 	}
409 
410 	host->clk = devm_clk_get(&pdev->dev, NULL);
411 	if (IS_ERR(host->clk)) {
412 		err = PTR_ERR(host->clk);
413 
414 		if (err != -EPROBE_DEFER)
415 			dev_err(&pdev->dev, "failed to get clock: %d\n", err);
416 
417 		return err;
418 	}
419 
420 	host->rst = devm_reset_control_get(&pdev->dev, "host1x");
421 	if (IS_ERR(host->rst)) {
422 		err = PTR_ERR(host->rst);
423 		dev_err(&pdev->dev, "failed to get reset: %d\n", err);
424 		return err;
425 	}
426 
427 	err = host1x_iommu_init(host);
428 	if (err < 0) {
429 		dev_err(&pdev->dev, "failed to setup IOMMU: %d\n", err);
430 		return err;
431 	}
432 
433 	err = host1x_channel_list_init(&host->channel_list,
434 				       host->info->nb_channels);
435 	if (err) {
436 		dev_err(&pdev->dev, "failed to initialize channel list\n");
437 		goto iommu_exit;
438 	}
439 
440 	err = clk_prepare_enable(host->clk);
441 	if (err < 0) {
442 		dev_err(&pdev->dev, "failed to enable clock\n");
443 		goto free_channels;
444 	}
445 
446 	err = reset_control_deassert(host->rst);
447 	if (err < 0) {
448 		dev_err(&pdev->dev, "failed to deassert reset: %d\n", err);
449 		goto unprepare_disable;
450 	}
451 
452 	err = host1x_syncpt_init(host);
453 	if (err) {
454 		dev_err(&pdev->dev, "failed to initialize syncpts\n");
455 		goto reset_assert;
456 	}
457 
458 	err = host1x_intr_init(host, syncpt_irq);
459 	if (err) {
460 		dev_err(&pdev->dev, "failed to initialize interrupts\n");
461 		goto deinit_syncpt;
462 	}
463 
464 	host1x_debug_init(host);
465 
466 	if (host->info->has_hypervisor)
467 		host1x_setup_sid_table(host);
468 
469 	err = host1x_register(host);
470 	if (err < 0)
471 		goto deinit_intr;
472 
473 	return 0;
474 
475 deinit_intr:
476 	host1x_intr_deinit(host);
477 deinit_syncpt:
478 	host1x_syncpt_deinit(host);
479 reset_assert:
480 	reset_control_assert(host->rst);
481 unprepare_disable:
482 	clk_disable_unprepare(host->clk);
483 free_channels:
484 	host1x_channel_list_free(&host->channel_list);
485 iommu_exit:
486 	host1x_iommu_exit(host);
487 
488 	return err;
489 }
490 
491 static int host1x_remove(struct platform_device *pdev)
492 {
493 	struct host1x *host = platform_get_drvdata(pdev);
494 
495 	host1x_unregister(host);
496 	host1x_debug_deinit(host);
497 	host1x_intr_deinit(host);
498 	host1x_syncpt_deinit(host);
499 	reset_control_assert(host->rst);
500 	clk_disable_unprepare(host->clk);
501 	host1x_iommu_exit(host);
502 
503 	return 0;
504 }
505 
506 static struct platform_driver tegra_host1x_driver = {
507 	.driver = {
508 		.name = "tegra-host1x",
509 		.of_match_table = host1x_of_match,
510 	},
511 	.probe = host1x_probe,
512 	.remove = host1x_remove,
513 };
514 
515 static struct platform_driver * const drivers[] = {
516 	&tegra_host1x_driver,
517 	&tegra_mipi_driver,
518 };
519 
520 static int __init tegra_host1x_init(void)
521 {
522 	int err;
523 
524 	err = bus_register(&host1x_bus_type);
525 	if (err < 0)
526 		return err;
527 
528 	err = platform_register_drivers(drivers, ARRAY_SIZE(drivers));
529 	if (err < 0)
530 		bus_unregister(&host1x_bus_type);
531 
532 	return err;
533 }
534 module_init(tegra_host1x_init);
535 
536 static void __exit tegra_host1x_exit(void)
537 {
538 	platform_unregister_drivers(drivers, ARRAY_SIZE(drivers));
539 	bus_unregister(&host1x_bus_type);
540 }
541 module_exit(tegra_host1x_exit);
542 
543 /**
544  * host1x_get_dma_mask() - query the supported DMA mask for host1x
545  * @host1x: host1x instance
546  *
547  * Note that this returns the supported DMA mask for host1x, which can be
548  * different from the applicable DMA mask under certain circumstances.
549  */
550 u64 host1x_get_dma_mask(struct host1x *host1x)
551 {
552 	return host1x->info->dma_mask;
553 }
554 EXPORT_SYMBOL(host1x_get_dma_mask);
555 
556 MODULE_AUTHOR("Thierry Reding <thierry.reding@avionic-design.de>");
557 MODULE_AUTHOR("Terje Bergstrom <tbergstrom@nvidia.com>");
558 MODULE_DESCRIPTION("Host1x driver for Tegra products");
559 MODULE_LICENSE("GPL");
560