xref: /linux/drivers/iommu/apple-dart.c (revision f2745dc0ba3dadd8fa2b2c33f48253d78e133a12)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Apple DART (Device Address Resolution Table) IOMMU driver
4  *
5  * Copyright (C) 2021 The Asahi Linux Contributors
6  *
7  * Based on arm/arm-smmu/arm-ssmu.c and arm/arm-smmu-v3/arm-smmu-v3.c
8  *  Copyright (C) 2013 ARM Limited
9  *  Copyright (C) 2015 ARM Limited
10  * and on exynos-iommu.c
11  *  Copyright (c) 2011,2016 Samsung Electronics Co., Ltd.
12  */
13 
14 #include <linux/atomic.h>
15 #include <linux/bitfield.h>
16 #include <linux/clk.h>
17 #include <linux/dev_printk.h>
18 #include <linux/dma-iommu.h>
19 #include <linux/dma-mapping.h>
20 #include <linux/err.h>
21 #include <linux/interrupt.h>
22 #include <linux/io-pgtable.h>
23 #include <linux/iommu.h>
24 #include <linux/iopoll.h>
25 #include <linux/module.h>
26 #include <linux/of.h>
27 #include <linux/of_address.h>
28 #include <linux/of_iommu.h>
29 #include <linux/of_platform.h>
30 #include <linux/pci.h>
31 #include <linux/platform_device.h>
32 #include <linux/slab.h>
33 #include <linux/swab.h>
34 #include <linux/types.h>
35 
36 #define DART_MAX_STREAMS 16
37 #define DART_MAX_TTBR 4
38 #define MAX_DARTS_PER_DEVICE 2
39 
40 #define DART_STREAM_ALL 0xffff
41 
42 #define DART_PARAMS1 0x00
43 #define DART_PARAMS_PAGE_SHIFT GENMASK(27, 24)
44 
45 #define DART_PARAMS2 0x04
46 #define DART_PARAMS_BYPASS_SUPPORT BIT(0)
47 
48 #define DART_STREAM_COMMAND 0x20
49 #define DART_STREAM_COMMAND_BUSY BIT(2)
50 #define DART_STREAM_COMMAND_INVALIDATE BIT(20)
51 
52 #define DART_STREAM_SELECT 0x34
53 
54 #define DART_ERROR 0x40
55 #define DART_ERROR_STREAM GENMASK(27, 24)
56 #define DART_ERROR_CODE GENMASK(11, 0)
57 #define DART_ERROR_FLAG BIT(31)
58 
59 #define DART_ERROR_READ_FAULT BIT(4)
60 #define DART_ERROR_WRITE_FAULT BIT(3)
61 #define DART_ERROR_NO_PTE BIT(2)
62 #define DART_ERROR_NO_PMD BIT(1)
63 #define DART_ERROR_NO_TTBR BIT(0)
64 
65 #define DART_CONFIG 0x60
66 #define DART_CONFIG_LOCK BIT(15)
67 
68 #define DART_STREAM_COMMAND_BUSY_TIMEOUT 100
69 
70 #define DART_ERROR_ADDR_HI 0x54
71 #define DART_ERROR_ADDR_LO 0x50
72 
73 #define DART_STREAMS_ENABLE 0xfc
74 
75 #define DART_TCR(sid) (0x100 + 4 * (sid))
76 #define DART_TCR_TRANSLATE_ENABLE BIT(7)
77 #define DART_TCR_BYPASS0_ENABLE BIT(8)
78 #define DART_TCR_BYPASS1_ENABLE BIT(12)
79 
80 #define DART_TTBR(sid, idx) (0x200 + 16 * (sid) + 4 * (idx))
81 #define DART_TTBR_VALID BIT(31)
82 #define DART_TTBR_SHIFT 12
83 
84 /*
85  * Private structure associated with each DART device.
86  *
87  * @dev: device struct
88  * @regs: mapped MMIO region
89  * @irq: interrupt number, can be shared with other DARTs
90  * @clks: clocks associated with this DART
91  * @num_clks: number of @clks
92  * @lock: lock for hardware operations involving this dart
93  * @pgsize: pagesize supported by this DART
94  * @supports_bypass: indicates if this DART supports bypass mode
95  * @force_bypass: force bypass mode due to pagesize mismatch?
96  * @sid2group: maps stream ids to iommu_groups
97  * @iommu: iommu core device
98  */
99 struct apple_dart {
100 	struct device *dev;
101 
102 	void __iomem *regs;
103 
104 	int irq;
105 	struct clk_bulk_data *clks;
106 	int num_clks;
107 
108 	spinlock_t lock;
109 
110 	u32 pgsize;
111 	u32 supports_bypass : 1;
112 	u32 force_bypass : 1;
113 
114 	struct iommu_group *sid2group[DART_MAX_STREAMS];
115 	struct iommu_device iommu;
116 };
117 
118 /*
119  * Convenience struct to identify streams.
120  *
121  * The normal variant is used inside apple_dart_master_cfg which isn't written
122  * to concurrently.
123  * The atomic variant is used inside apple_dart_domain where we have to guard
124  * against races from potential parallel calls to attach/detach_device.
125  * Note that even inside the atomic variant the apple_dart pointer is not
126  * protected: This pointer is initialized once under the domain init mutex
127  * and never changed again afterwards. Devices with different dart pointers
128  * cannot be attached to the same domain.
129  *
130  * @dart dart pointer
131  * @sid stream id bitmap
132  */
133 struct apple_dart_stream_map {
134 	struct apple_dart *dart;
135 	unsigned long sidmap;
136 };
137 struct apple_dart_atomic_stream_map {
138 	struct apple_dart *dart;
139 	atomic64_t sidmap;
140 };
141 
142 /*
143  * This structure is attached to each iommu domain handled by a DART.
144  *
145  * @pgtbl_ops: pagetable ops allocated by io-pgtable
146  * @finalized: true if the domain has been completely initialized
147  * @init_lock: protects domain initialization
148  * @stream_maps: streams attached to this domain (valid for DMA/UNMANAGED only)
149  * @domain: core iommu domain pointer
150  */
151 struct apple_dart_domain {
152 	struct io_pgtable_ops *pgtbl_ops;
153 
154 	bool finalized;
155 	struct mutex init_lock;
156 	struct apple_dart_atomic_stream_map stream_maps[MAX_DARTS_PER_DEVICE];
157 
158 	struct iommu_domain domain;
159 };
160 
161 /*
162  * This structure is attached to devices with dev_iommu_priv_set() on of_xlate
163  * and contains a list of streams bound to this device.
164  * So far the worst case seen is a single device with two streams
165  * from different darts, such that this simple static array is enough.
166  *
167  * @streams: streams for this device
168  */
169 struct apple_dart_master_cfg {
170 	struct apple_dart_stream_map stream_maps[MAX_DARTS_PER_DEVICE];
171 };
172 
173 /*
174  * Helper macro to iterate over apple_dart_master_cfg.stream_maps and
175  * apple_dart_domain.stream_maps
176  *
177  * @i int used as loop variable
178  * @base pointer to base struct (apple_dart_master_cfg or apple_dart_domain)
179  * @stream pointer to the apple_dart_streams struct for each loop iteration
180  */
181 #define for_each_stream_map(i, base, stream_map)                               \
182 	for (i = 0, stream_map = &(base)->stream_maps[0];                      \
183 	     i < MAX_DARTS_PER_DEVICE && stream_map->dart;                     \
184 	     stream_map = &(base)->stream_maps[++i])
185 
186 static struct platform_driver apple_dart_driver;
187 static const struct iommu_ops apple_dart_iommu_ops;
188 
189 static struct apple_dart_domain *to_dart_domain(struct iommu_domain *dom)
190 {
191 	return container_of(dom, struct apple_dart_domain, domain);
192 }
193 
194 static void
195 apple_dart_hw_enable_translation(struct apple_dart_stream_map *stream_map)
196 {
197 	int sid;
198 
199 	for_each_set_bit(sid, &stream_map->sidmap, DART_MAX_STREAMS)
200 		writel(DART_TCR_TRANSLATE_ENABLE,
201 		       stream_map->dart->regs + DART_TCR(sid));
202 }
203 
204 static void apple_dart_hw_disable_dma(struct apple_dart_stream_map *stream_map)
205 {
206 	int sid;
207 
208 	for_each_set_bit(sid, &stream_map->sidmap, DART_MAX_STREAMS)
209 		writel(0, stream_map->dart->regs + DART_TCR(sid));
210 }
211 
212 static void
213 apple_dart_hw_enable_bypass(struct apple_dart_stream_map *stream_map)
214 {
215 	int sid;
216 
217 	WARN_ON(!stream_map->dart->supports_bypass);
218 	for_each_set_bit(sid, &stream_map->sidmap, DART_MAX_STREAMS)
219 		writel(DART_TCR_BYPASS0_ENABLE | DART_TCR_BYPASS1_ENABLE,
220 		       stream_map->dart->regs + DART_TCR(sid));
221 }
222 
223 static void apple_dart_hw_set_ttbr(struct apple_dart_stream_map *stream_map,
224 				   u8 idx, phys_addr_t paddr)
225 {
226 	int sid;
227 
228 	WARN_ON(paddr & ((1 << DART_TTBR_SHIFT) - 1));
229 	for_each_set_bit(sid, &stream_map->sidmap, DART_MAX_STREAMS)
230 		writel(DART_TTBR_VALID | (paddr >> DART_TTBR_SHIFT),
231 		       stream_map->dart->regs + DART_TTBR(sid, idx));
232 }
233 
234 static void apple_dart_hw_clear_ttbr(struct apple_dart_stream_map *stream_map,
235 				     u8 idx)
236 {
237 	int sid;
238 
239 	for_each_set_bit(sid, &stream_map->sidmap, DART_MAX_STREAMS)
240 		writel(0, stream_map->dart->regs + DART_TTBR(sid, idx));
241 }
242 
243 static void
244 apple_dart_hw_clear_all_ttbrs(struct apple_dart_stream_map *stream_map)
245 {
246 	int i;
247 
248 	for (i = 0; i < DART_MAX_TTBR; ++i)
249 		apple_dart_hw_clear_ttbr(stream_map, i);
250 }
251 
252 static int
253 apple_dart_hw_stream_command(struct apple_dart_stream_map *stream_map,
254 			     u32 command)
255 {
256 	unsigned long flags;
257 	int ret;
258 	u32 command_reg;
259 
260 	spin_lock_irqsave(&stream_map->dart->lock, flags);
261 
262 	writel(stream_map->sidmap, stream_map->dart->regs + DART_STREAM_SELECT);
263 	writel(command, stream_map->dart->regs + DART_STREAM_COMMAND);
264 
265 	ret = readl_poll_timeout_atomic(
266 		stream_map->dart->regs + DART_STREAM_COMMAND, command_reg,
267 		!(command_reg & DART_STREAM_COMMAND_BUSY), 1,
268 		DART_STREAM_COMMAND_BUSY_TIMEOUT);
269 
270 	spin_unlock_irqrestore(&stream_map->dart->lock, flags);
271 
272 	if (ret) {
273 		dev_err(stream_map->dart->dev,
274 			"busy bit did not clear after command %x for streams %lx\n",
275 			command, stream_map->sidmap);
276 		return ret;
277 	}
278 
279 	return 0;
280 }
281 
282 static int
283 apple_dart_hw_invalidate_tlb(struct apple_dart_stream_map *stream_map)
284 {
285 	return apple_dart_hw_stream_command(stream_map,
286 					    DART_STREAM_COMMAND_INVALIDATE);
287 }
288 
289 static int apple_dart_hw_reset(struct apple_dart *dart)
290 {
291 	u32 config;
292 	struct apple_dart_stream_map stream_map;
293 
294 	config = readl(dart->regs + DART_CONFIG);
295 	if (config & DART_CONFIG_LOCK) {
296 		dev_err(dart->dev, "DART is locked down until reboot: %08x\n",
297 			config);
298 		return -EINVAL;
299 	}
300 
301 	stream_map.dart = dart;
302 	stream_map.sidmap = DART_STREAM_ALL;
303 	apple_dart_hw_disable_dma(&stream_map);
304 	apple_dart_hw_clear_all_ttbrs(&stream_map);
305 
306 	/* enable all streams globally since TCR is used to control isolation */
307 	writel(DART_STREAM_ALL, dart->regs + DART_STREAMS_ENABLE);
308 
309 	/* clear any pending errors before the interrupt is unmasked */
310 	writel(readl(dart->regs + DART_ERROR), dart->regs + DART_ERROR);
311 
312 	return apple_dart_hw_invalidate_tlb(&stream_map);
313 }
314 
315 static void apple_dart_domain_flush_tlb(struct apple_dart_domain *domain)
316 {
317 	int i;
318 	struct apple_dart_atomic_stream_map *domain_stream_map;
319 	struct apple_dart_stream_map stream_map;
320 
321 	for_each_stream_map(i, domain, domain_stream_map) {
322 		stream_map.dart = domain_stream_map->dart;
323 		stream_map.sidmap = atomic64_read(&domain_stream_map->sidmap);
324 		apple_dart_hw_invalidate_tlb(&stream_map);
325 	}
326 }
327 
328 static void apple_dart_flush_iotlb_all(struct iommu_domain *domain)
329 {
330 	apple_dart_domain_flush_tlb(to_dart_domain(domain));
331 }
332 
333 static void apple_dart_iotlb_sync(struct iommu_domain *domain,
334 				  struct iommu_iotlb_gather *gather)
335 {
336 	apple_dart_domain_flush_tlb(to_dart_domain(domain));
337 }
338 
339 static void apple_dart_iotlb_sync_map(struct iommu_domain *domain,
340 				      unsigned long iova, size_t size)
341 {
342 	apple_dart_domain_flush_tlb(to_dart_domain(domain));
343 }
344 
345 static phys_addr_t apple_dart_iova_to_phys(struct iommu_domain *domain,
346 					   dma_addr_t iova)
347 {
348 	struct apple_dart_domain *dart_domain = to_dart_domain(domain);
349 	struct io_pgtable_ops *ops = dart_domain->pgtbl_ops;
350 
351 	if (!ops)
352 		return 0;
353 
354 	return ops->iova_to_phys(ops, iova);
355 }
356 
357 static int apple_dart_map_pages(struct iommu_domain *domain, unsigned long iova,
358 				phys_addr_t paddr, size_t pgsize,
359 				size_t pgcount, int prot, gfp_t gfp,
360 				size_t *mapped)
361 {
362 	struct apple_dart_domain *dart_domain = to_dart_domain(domain);
363 	struct io_pgtable_ops *ops = dart_domain->pgtbl_ops;
364 
365 	if (!ops)
366 		return -ENODEV;
367 
368 	return ops->map_pages(ops, iova, paddr, pgsize, pgcount, prot, gfp,
369 			      mapped);
370 }
371 
372 static size_t apple_dart_unmap_pages(struct iommu_domain *domain,
373 				     unsigned long iova, size_t pgsize,
374 				     size_t pgcount,
375 				     struct iommu_iotlb_gather *gather)
376 {
377 	struct apple_dart_domain *dart_domain = to_dart_domain(domain);
378 	struct io_pgtable_ops *ops = dart_domain->pgtbl_ops;
379 
380 	return ops->unmap_pages(ops, iova, pgsize, pgcount, gather);
381 }
382 
383 static void
384 apple_dart_setup_translation(struct apple_dart_domain *domain,
385 			     struct apple_dart_stream_map *stream_map)
386 {
387 	int i;
388 	struct io_pgtable_cfg *pgtbl_cfg =
389 		&io_pgtable_ops_to_pgtable(domain->pgtbl_ops)->cfg;
390 
391 	for (i = 0; i < pgtbl_cfg->apple_dart_cfg.n_ttbrs; ++i)
392 		apple_dart_hw_set_ttbr(stream_map, i,
393 				       pgtbl_cfg->apple_dart_cfg.ttbr[i]);
394 	for (; i < DART_MAX_TTBR; ++i)
395 		apple_dart_hw_clear_ttbr(stream_map, i);
396 
397 	apple_dart_hw_enable_translation(stream_map);
398 	apple_dart_hw_invalidate_tlb(stream_map);
399 }
400 
401 static int apple_dart_finalize_domain(struct iommu_domain *domain,
402 				      struct apple_dart_master_cfg *cfg)
403 {
404 	struct apple_dart_domain *dart_domain = to_dart_domain(domain);
405 	struct apple_dart *dart = cfg->stream_maps[0].dart;
406 	struct io_pgtable_cfg pgtbl_cfg;
407 	int ret = 0;
408 	int i;
409 
410 	mutex_lock(&dart_domain->init_lock);
411 
412 	if (dart_domain->finalized)
413 		goto done;
414 
415 	for (i = 0; i < MAX_DARTS_PER_DEVICE; ++i) {
416 		dart_domain->stream_maps[i].dart = cfg->stream_maps[i].dart;
417 		atomic64_set(&dart_domain->stream_maps[i].sidmap,
418 			     cfg->stream_maps[i].sidmap);
419 	}
420 
421 	pgtbl_cfg = (struct io_pgtable_cfg){
422 		.pgsize_bitmap = dart->pgsize,
423 		.ias = 32,
424 		.oas = 36,
425 		.coherent_walk = 1,
426 		.iommu_dev = dart->dev,
427 	};
428 
429 	dart_domain->pgtbl_ops =
430 		alloc_io_pgtable_ops(APPLE_DART, &pgtbl_cfg, domain);
431 	if (!dart_domain->pgtbl_ops) {
432 		ret = -ENOMEM;
433 		goto done;
434 	}
435 
436 	domain->pgsize_bitmap = pgtbl_cfg.pgsize_bitmap;
437 	domain->geometry.aperture_start = 0;
438 	domain->geometry.aperture_end = DMA_BIT_MASK(32);
439 	domain->geometry.force_aperture = true;
440 
441 	dart_domain->finalized = true;
442 
443 done:
444 	mutex_unlock(&dart_domain->init_lock);
445 	return ret;
446 }
447 
448 static int
449 apple_dart_mod_streams(struct apple_dart_atomic_stream_map *domain_maps,
450 		       struct apple_dart_stream_map *master_maps,
451 		       bool add_streams)
452 {
453 	int i;
454 
455 	for (i = 0; i < MAX_DARTS_PER_DEVICE; ++i) {
456 		if (domain_maps[i].dart != master_maps[i].dart)
457 			return -EINVAL;
458 	}
459 
460 	for (i = 0; i < MAX_DARTS_PER_DEVICE; ++i) {
461 		if (!domain_maps[i].dart)
462 			break;
463 		if (add_streams)
464 			atomic64_or(master_maps[i].sidmap,
465 				    &domain_maps[i].sidmap);
466 		else
467 			atomic64_and(~master_maps[i].sidmap,
468 				     &domain_maps[i].sidmap);
469 	}
470 
471 	return 0;
472 }
473 
474 static int apple_dart_domain_add_streams(struct apple_dart_domain *domain,
475 					 struct apple_dart_master_cfg *cfg)
476 {
477 	return apple_dart_mod_streams(domain->stream_maps, cfg->stream_maps,
478 				      true);
479 }
480 
481 static int apple_dart_domain_remove_streams(struct apple_dart_domain *domain,
482 					    struct apple_dart_master_cfg *cfg)
483 {
484 	return apple_dart_mod_streams(domain->stream_maps, cfg->stream_maps,
485 				      false);
486 }
487 
488 static int apple_dart_attach_dev(struct iommu_domain *domain,
489 				 struct device *dev)
490 {
491 	int ret, i;
492 	struct apple_dart_stream_map *stream_map;
493 	struct apple_dart_master_cfg *cfg = dev_iommu_priv_get(dev);
494 	struct apple_dart_domain *dart_domain = to_dart_domain(domain);
495 
496 	if (cfg->stream_maps[0].dart->force_bypass &&
497 	    domain->type != IOMMU_DOMAIN_IDENTITY)
498 		return -EINVAL;
499 	if (!cfg->stream_maps[0].dart->supports_bypass &&
500 	    domain->type == IOMMU_DOMAIN_IDENTITY)
501 		return -EINVAL;
502 
503 	ret = apple_dart_finalize_domain(domain, cfg);
504 	if (ret)
505 		return ret;
506 
507 	switch (domain->type) {
508 	case IOMMU_DOMAIN_DMA:
509 	case IOMMU_DOMAIN_UNMANAGED:
510 		ret = apple_dart_domain_add_streams(dart_domain, cfg);
511 		if (ret)
512 			return ret;
513 
514 		for_each_stream_map(i, cfg, stream_map)
515 			apple_dart_setup_translation(dart_domain, stream_map);
516 		break;
517 	case IOMMU_DOMAIN_BLOCKED:
518 		for_each_stream_map(i, cfg, stream_map)
519 			apple_dart_hw_disable_dma(stream_map);
520 		break;
521 	case IOMMU_DOMAIN_IDENTITY:
522 		for_each_stream_map(i, cfg, stream_map)
523 			apple_dart_hw_enable_bypass(stream_map);
524 		break;
525 	}
526 
527 	return ret;
528 }
529 
530 static void apple_dart_detach_dev(struct iommu_domain *domain,
531 				  struct device *dev)
532 {
533 	int i;
534 	struct apple_dart_stream_map *stream_map;
535 	struct apple_dart_master_cfg *cfg = dev_iommu_priv_get(dev);
536 	struct apple_dart_domain *dart_domain = to_dart_domain(domain);
537 
538 	for_each_stream_map(i, cfg, stream_map)
539 		apple_dart_hw_disable_dma(stream_map);
540 
541 	if (domain->type == IOMMU_DOMAIN_DMA ||
542 	    domain->type == IOMMU_DOMAIN_UNMANAGED)
543 		apple_dart_domain_remove_streams(dart_domain, cfg);
544 }
545 
546 static struct iommu_device *apple_dart_probe_device(struct device *dev)
547 {
548 	struct apple_dart_master_cfg *cfg = dev_iommu_priv_get(dev);
549 	struct apple_dart_stream_map *stream_map;
550 	int i;
551 
552 	if (!cfg)
553 		return ERR_PTR(-ENODEV);
554 
555 	for_each_stream_map(i, cfg, stream_map)
556 		device_link_add(
557 			dev, stream_map->dart->dev,
558 			DL_FLAG_PM_RUNTIME | DL_FLAG_AUTOREMOVE_SUPPLIER);
559 
560 	return &cfg->stream_maps[0].dart->iommu;
561 }
562 
563 static void apple_dart_release_device(struct device *dev)
564 {
565 	struct apple_dart_master_cfg *cfg = dev_iommu_priv_get(dev);
566 
567 	if (!cfg)
568 		return;
569 
570 	dev_iommu_priv_set(dev, NULL);
571 	kfree(cfg);
572 }
573 
574 static struct iommu_domain *apple_dart_domain_alloc(unsigned int type)
575 {
576 	struct apple_dart_domain *dart_domain;
577 
578 	if (type != IOMMU_DOMAIN_DMA && type != IOMMU_DOMAIN_UNMANAGED &&
579 	    type != IOMMU_DOMAIN_IDENTITY && type != IOMMU_DOMAIN_BLOCKED)
580 		return NULL;
581 
582 	dart_domain = kzalloc(sizeof(*dart_domain), GFP_KERNEL);
583 	if (!dart_domain)
584 		return NULL;
585 
586 	mutex_init(&dart_domain->init_lock);
587 
588 	/* no need to allocate pgtbl_ops or do any other finalization steps */
589 	if (type == IOMMU_DOMAIN_IDENTITY || type == IOMMU_DOMAIN_BLOCKED)
590 		dart_domain->finalized = true;
591 
592 	return &dart_domain->domain;
593 }
594 
595 static void apple_dart_domain_free(struct iommu_domain *domain)
596 {
597 	struct apple_dart_domain *dart_domain = to_dart_domain(domain);
598 
599 	if (dart_domain->pgtbl_ops)
600 		free_io_pgtable_ops(dart_domain->pgtbl_ops);
601 
602 	kfree(dart_domain);
603 }
604 
605 static int apple_dart_of_xlate(struct device *dev, struct of_phandle_args *args)
606 {
607 	struct apple_dart_master_cfg *cfg = dev_iommu_priv_get(dev);
608 	struct platform_device *iommu_pdev = of_find_device_by_node(args->np);
609 	struct apple_dart *dart = platform_get_drvdata(iommu_pdev);
610 	struct apple_dart *cfg_dart;
611 	int i, sid;
612 
613 	if (args->args_count != 1)
614 		return -EINVAL;
615 	sid = args->args[0];
616 
617 	if (!cfg)
618 		cfg = kzalloc(sizeof(*cfg), GFP_KERNEL);
619 	if (!cfg)
620 		return -ENOMEM;
621 	dev_iommu_priv_set(dev, cfg);
622 
623 	cfg_dart = cfg->stream_maps[0].dart;
624 	if (cfg_dart) {
625 		if (cfg_dart->supports_bypass != dart->supports_bypass)
626 			return -EINVAL;
627 		if (cfg_dart->force_bypass != dart->force_bypass)
628 			return -EINVAL;
629 		if (cfg_dart->pgsize != dart->pgsize)
630 			return -EINVAL;
631 	}
632 
633 	for (i = 0; i < MAX_DARTS_PER_DEVICE; ++i) {
634 		if (cfg->stream_maps[i].dart == dart) {
635 			cfg->stream_maps[i].sidmap |= 1 << sid;
636 			return 0;
637 		}
638 	}
639 	for (i = 0; i < MAX_DARTS_PER_DEVICE; ++i) {
640 		if (!cfg->stream_maps[i].dart) {
641 			cfg->stream_maps[i].dart = dart;
642 			cfg->stream_maps[i].sidmap = 1 << sid;
643 			return 0;
644 		}
645 	}
646 
647 	return -EINVAL;
648 }
649 
650 static DEFINE_MUTEX(apple_dart_groups_lock);
651 
652 static void apple_dart_release_group(void *iommu_data)
653 {
654 	int i, sid;
655 	struct apple_dart_stream_map *stream_map;
656 	struct apple_dart_master_cfg *group_master_cfg = iommu_data;
657 
658 	mutex_lock(&apple_dart_groups_lock);
659 
660 	for_each_stream_map(i, group_master_cfg, stream_map)
661 		for_each_set_bit(sid, &stream_map->sidmap, DART_MAX_STREAMS)
662 			stream_map->dart->sid2group[sid] = NULL;
663 
664 	kfree(iommu_data);
665 	mutex_unlock(&apple_dart_groups_lock);
666 }
667 
668 static struct iommu_group *apple_dart_device_group(struct device *dev)
669 {
670 	int i, sid;
671 	struct apple_dart_master_cfg *cfg = dev_iommu_priv_get(dev);
672 	struct apple_dart_stream_map *stream_map;
673 	struct apple_dart_master_cfg *group_master_cfg;
674 	struct iommu_group *group = NULL;
675 	struct iommu_group *res = ERR_PTR(-EINVAL);
676 
677 	mutex_lock(&apple_dart_groups_lock);
678 
679 	for_each_stream_map(i, cfg, stream_map) {
680 		for_each_set_bit(sid, &stream_map->sidmap, DART_MAX_STREAMS) {
681 			struct iommu_group *stream_group =
682 				stream_map->dart->sid2group[sid];
683 
684 			if (group && group != stream_group) {
685 				res = ERR_PTR(-EINVAL);
686 				goto out;
687 			}
688 
689 			group = stream_group;
690 		}
691 	}
692 
693 	if (group) {
694 		res = iommu_group_ref_get(group);
695 		goto out;
696 	}
697 
698 #ifdef CONFIG_PCI
699 	if (dev_is_pci(dev))
700 		group = pci_device_group(dev);
701 	else
702 #endif
703 		group = generic_device_group(dev);
704 
705 	res = ERR_PTR(-ENOMEM);
706 	if (!group)
707 		goto out;
708 
709 	group_master_cfg = kmemdup(cfg, sizeof(*group_master_cfg), GFP_KERNEL);
710 	if (!group_master_cfg) {
711 		iommu_group_put(group);
712 		goto out;
713 	}
714 
715 	iommu_group_set_iommudata(group, group_master_cfg,
716 		apple_dart_release_group);
717 
718 	for_each_stream_map(i, cfg, stream_map)
719 		for_each_set_bit(sid, &stream_map->sidmap, DART_MAX_STREAMS)
720 			stream_map->dart->sid2group[sid] = group;
721 
722 	res = group;
723 
724 out:
725 	mutex_unlock(&apple_dart_groups_lock);
726 	return res;
727 }
728 
729 static int apple_dart_def_domain_type(struct device *dev)
730 {
731 	struct apple_dart_master_cfg *cfg = dev_iommu_priv_get(dev);
732 
733 	if (cfg->stream_maps[0].dart->force_bypass)
734 		return IOMMU_DOMAIN_IDENTITY;
735 	if (!cfg->stream_maps[0].dart->supports_bypass)
736 		return IOMMU_DOMAIN_DMA;
737 
738 	return 0;
739 }
740 
741 #ifndef CONFIG_PCIE_APPLE_MSI_DOORBELL_ADDR
742 /* Keep things compiling when CONFIG_PCI_APPLE isn't selected */
743 #define CONFIG_PCIE_APPLE_MSI_DOORBELL_ADDR	0
744 #endif
745 #define DOORBELL_ADDR	(CONFIG_PCIE_APPLE_MSI_DOORBELL_ADDR & PAGE_MASK)
746 
747 static void apple_dart_get_resv_regions(struct device *dev,
748 					struct list_head *head)
749 {
750 	if (IS_ENABLED(CONFIG_PCIE_APPLE) && dev_is_pci(dev)) {
751 		struct iommu_resv_region *region;
752 		int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
753 
754 		region = iommu_alloc_resv_region(DOORBELL_ADDR,
755 						 PAGE_SIZE, prot,
756 						 IOMMU_RESV_MSI);
757 		if (!region)
758 			return;
759 
760 		list_add_tail(&region->list, head);
761 	}
762 
763 	iommu_dma_get_resv_regions(dev, head);
764 }
765 
766 static const struct iommu_ops apple_dart_iommu_ops = {
767 	.domain_alloc = apple_dart_domain_alloc,
768 	.probe_device = apple_dart_probe_device,
769 	.release_device = apple_dart_release_device,
770 	.device_group = apple_dart_device_group,
771 	.of_xlate = apple_dart_of_xlate,
772 	.def_domain_type = apple_dart_def_domain_type,
773 	.get_resv_regions = apple_dart_get_resv_regions,
774 	.put_resv_regions = generic_iommu_put_resv_regions,
775 	.pgsize_bitmap = -1UL, /* Restricted during dart probe */
776 	.owner = THIS_MODULE,
777 	.default_domain_ops = &(const struct iommu_domain_ops) {
778 		.attach_dev	= apple_dart_attach_dev,
779 		.detach_dev	= apple_dart_detach_dev,
780 		.map_pages	= apple_dart_map_pages,
781 		.unmap_pages	= apple_dart_unmap_pages,
782 		.flush_iotlb_all = apple_dart_flush_iotlb_all,
783 		.iotlb_sync	= apple_dart_iotlb_sync,
784 		.iotlb_sync_map	= apple_dart_iotlb_sync_map,
785 		.iova_to_phys	= apple_dart_iova_to_phys,
786 		.free		= apple_dart_domain_free,
787 	}
788 };
789 
790 static irqreturn_t apple_dart_irq(int irq, void *dev)
791 {
792 	struct apple_dart *dart = dev;
793 	const char *fault_name = NULL;
794 	u32 error = readl(dart->regs + DART_ERROR);
795 	u32 error_code = FIELD_GET(DART_ERROR_CODE, error);
796 	u32 addr_lo = readl(dart->regs + DART_ERROR_ADDR_LO);
797 	u32 addr_hi = readl(dart->regs + DART_ERROR_ADDR_HI);
798 	u64 addr = addr_lo | (((u64)addr_hi) << 32);
799 	u8 stream_idx = FIELD_GET(DART_ERROR_STREAM, error);
800 
801 	if (!(error & DART_ERROR_FLAG))
802 		return IRQ_NONE;
803 
804 	/* there should only be a single bit set but let's use == to be sure */
805 	if (error_code == DART_ERROR_READ_FAULT)
806 		fault_name = "READ FAULT";
807 	else if (error_code == DART_ERROR_WRITE_FAULT)
808 		fault_name = "WRITE FAULT";
809 	else if (error_code == DART_ERROR_NO_PTE)
810 		fault_name = "NO PTE FOR IOVA";
811 	else if (error_code == DART_ERROR_NO_PMD)
812 		fault_name = "NO PMD FOR IOVA";
813 	else if (error_code == DART_ERROR_NO_TTBR)
814 		fault_name = "NO TTBR FOR IOVA";
815 	else
816 		fault_name = "unknown";
817 
818 	dev_err_ratelimited(
819 		dart->dev,
820 		"translation fault: status:0x%x stream:%d code:0x%x (%s) at 0x%llx",
821 		error, stream_idx, error_code, fault_name, addr);
822 
823 	writel(error, dart->regs + DART_ERROR);
824 	return IRQ_HANDLED;
825 }
826 
827 static int apple_dart_set_bus_ops(const struct iommu_ops *ops)
828 {
829 	int ret;
830 
831 	if (!iommu_present(&platform_bus_type)) {
832 		ret = bus_set_iommu(&platform_bus_type, ops);
833 		if (ret)
834 			return ret;
835 	}
836 #ifdef CONFIG_PCI
837 	if (!iommu_present(&pci_bus_type)) {
838 		ret = bus_set_iommu(&pci_bus_type, ops);
839 		if (ret) {
840 			bus_set_iommu(&platform_bus_type, NULL);
841 			return ret;
842 		}
843 	}
844 #endif
845 	return 0;
846 }
847 
848 static int apple_dart_probe(struct platform_device *pdev)
849 {
850 	int ret;
851 	u32 dart_params[2];
852 	struct resource *res;
853 	struct apple_dart *dart;
854 	struct device *dev = &pdev->dev;
855 
856 	dart = devm_kzalloc(dev, sizeof(*dart), GFP_KERNEL);
857 	if (!dart)
858 		return -ENOMEM;
859 
860 	dart->dev = dev;
861 	spin_lock_init(&dart->lock);
862 
863 	dart->regs = devm_platform_get_and_ioremap_resource(pdev, 0, &res);
864 	if (IS_ERR(dart->regs))
865 		return PTR_ERR(dart->regs);
866 
867 	if (resource_size(res) < 0x4000) {
868 		dev_err(dev, "MMIO region too small (%pr)\n", res);
869 		return -EINVAL;
870 	}
871 
872 	dart->irq = platform_get_irq(pdev, 0);
873 	if (dart->irq < 0)
874 		return -ENODEV;
875 
876 	ret = devm_clk_bulk_get_all(dev, &dart->clks);
877 	if (ret < 0)
878 		return ret;
879 	dart->num_clks = ret;
880 
881 	ret = clk_bulk_prepare_enable(dart->num_clks, dart->clks);
882 	if (ret)
883 		return ret;
884 
885 	ret = apple_dart_hw_reset(dart);
886 	if (ret)
887 		goto err_clk_disable;
888 
889 	dart_params[0] = readl(dart->regs + DART_PARAMS1);
890 	dart_params[1] = readl(dart->regs + DART_PARAMS2);
891 	dart->pgsize = 1 << FIELD_GET(DART_PARAMS_PAGE_SHIFT, dart_params[0]);
892 	dart->supports_bypass = dart_params[1] & DART_PARAMS_BYPASS_SUPPORT;
893 	dart->force_bypass = dart->pgsize > PAGE_SIZE;
894 
895 	ret = request_irq(dart->irq, apple_dart_irq, IRQF_SHARED,
896 			  "apple-dart fault handler", dart);
897 	if (ret)
898 		goto err_clk_disable;
899 
900 	platform_set_drvdata(pdev, dart);
901 
902 	ret = apple_dart_set_bus_ops(&apple_dart_iommu_ops);
903 	if (ret)
904 		goto err_free_irq;
905 
906 	ret = iommu_device_sysfs_add(&dart->iommu, dev, NULL, "apple-dart.%s",
907 				     dev_name(&pdev->dev));
908 	if (ret)
909 		goto err_remove_bus_ops;
910 
911 	ret = iommu_device_register(&dart->iommu, &apple_dart_iommu_ops, dev);
912 	if (ret)
913 		goto err_sysfs_remove;
914 
915 	dev_info(
916 		&pdev->dev,
917 		"DART [pagesize %x, bypass support: %d, bypass forced: %d] initialized\n",
918 		dart->pgsize, dart->supports_bypass, dart->force_bypass);
919 	return 0;
920 
921 err_sysfs_remove:
922 	iommu_device_sysfs_remove(&dart->iommu);
923 err_remove_bus_ops:
924 	apple_dart_set_bus_ops(NULL);
925 err_free_irq:
926 	free_irq(dart->irq, dart);
927 err_clk_disable:
928 	clk_bulk_disable_unprepare(dart->num_clks, dart->clks);
929 
930 	return ret;
931 }
932 
933 static int apple_dart_remove(struct platform_device *pdev)
934 {
935 	struct apple_dart *dart = platform_get_drvdata(pdev);
936 
937 	apple_dart_hw_reset(dart);
938 	free_irq(dart->irq, dart);
939 	apple_dart_set_bus_ops(NULL);
940 
941 	iommu_device_unregister(&dart->iommu);
942 	iommu_device_sysfs_remove(&dart->iommu);
943 
944 	clk_bulk_disable_unprepare(dart->num_clks, dart->clks);
945 
946 	return 0;
947 }
948 
949 static const struct of_device_id apple_dart_of_match[] = {
950 	{ .compatible = "apple,t8103-dart", .data = NULL },
951 	{},
952 };
953 MODULE_DEVICE_TABLE(of, apple_dart_of_match);
954 
955 static struct platform_driver apple_dart_driver = {
956 	.driver	= {
957 		.name			= "apple-dart",
958 		.of_match_table		= apple_dart_of_match,
959 		.suppress_bind_attrs    = true,
960 	},
961 	.probe	= apple_dart_probe,
962 	.remove	= apple_dart_remove,
963 };
964 
965 module_platform_driver(apple_dart_driver);
966 
967 MODULE_DESCRIPTION("IOMMU API for Apple's DART");
968 MODULE_AUTHOR("Sven Peter <sven@svenpeter.dev>");
969 MODULE_LICENSE("GPL v2");
970