xref: /openbmc/linux/drivers/iommu/ipmmu-vmsa.c (revision a06c488d)
1 /*
2  * IPMMU VMSA
3  *
4  * Copyright (C) 2014 Renesas Electronics Corporation
5  *
6  * This program is free software; you can redistribute it and/or modify
7  * it under the terms of the GNU General Public License as published by
8  * the Free Software Foundation; version 2 of the License.
9  */
10 
11 #include <linux/delay.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/err.h>
14 #include <linux/export.h>
15 #include <linux/interrupt.h>
16 #include <linux/io.h>
17 #include <linux/iommu.h>
18 #include <linux/module.h>
19 #include <linux/of.h>
20 #include <linux/platform_device.h>
21 #include <linux/sizes.h>
22 #include <linux/slab.h>
23 
24 #include <asm/dma-iommu.h>
25 #include <asm/pgalloc.h>
26 
27 #include "io-pgtable.h"
28 
29 struct ipmmu_vmsa_device {
30 	struct device *dev;
31 	void __iomem *base;
32 	struct list_head list;
33 
34 	unsigned int num_utlbs;
35 
36 	struct dma_iommu_mapping *mapping;
37 };
38 
39 struct ipmmu_vmsa_domain {
40 	struct ipmmu_vmsa_device *mmu;
41 	struct iommu_domain io_domain;
42 
43 	struct io_pgtable_cfg cfg;
44 	struct io_pgtable_ops *iop;
45 
46 	unsigned int context_id;
47 	spinlock_t lock;			/* Protects mappings */
48 };
49 
50 struct ipmmu_vmsa_archdata {
51 	struct ipmmu_vmsa_device *mmu;
52 	unsigned int *utlbs;
53 	unsigned int num_utlbs;
54 };
55 
56 static DEFINE_SPINLOCK(ipmmu_devices_lock);
57 static LIST_HEAD(ipmmu_devices);
58 
59 static struct ipmmu_vmsa_domain *to_vmsa_domain(struct iommu_domain *dom)
60 {
61 	return container_of(dom, struct ipmmu_vmsa_domain, io_domain);
62 }
63 
64 #define TLB_LOOP_TIMEOUT		100	/* 100us */
65 
66 /* -----------------------------------------------------------------------------
67  * Registers Definition
68  */
69 
70 #define IM_NS_ALIAS_OFFSET		0x800
71 
72 #define IM_CTX_SIZE			0x40
73 
74 #define IMCTR				0x0000
75 #define IMCTR_TRE			(1 << 17)
76 #define IMCTR_AFE			(1 << 16)
77 #define IMCTR_RTSEL_MASK		(3 << 4)
78 #define IMCTR_RTSEL_SHIFT		4
79 #define IMCTR_TREN			(1 << 3)
80 #define IMCTR_INTEN			(1 << 2)
81 #define IMCTR_FLUSH			(1 << 1)
82 #define IMCTR_MMUEN			(1 << 0)
83 
84 #define IMCAAR				0x0004
85 
86 #define IMTTBCR				0x0008
87 #define IMTTBCR_EAE			(1 << 31)
88 #define IMTTBCR_PMB			(1 << 30)
89 #define IMTTBCR_SH1_NON_SHAREABLE	(0 << 28)
90 #define IMTTBCR_SH1_OUTER_SHAREABLE	(2 << 28)
91 #define IMTTBCR_SH1_INNER_SHAREABLE	(3 << 28)
92 #define IMTTBCR_SH1_MASK		(3 << 28)
93 #define IMTTBCR_ORGN1_NC		(0 << 26)
94 #define IMTTBCR_ORGN1_WB_WA		(1 << 26)
95 #define IMTTBCR_ORGN1_WT		(2 << 26)
96 #define IMTTBCR_ORGN1_WB		(3 << 26)
97 #define IMTTBCR_ORGN1_MASK		(3 << 26)
98 #define IMTTBCR_IRGN1_NC		(0 << 24)
99 #define IMTTBCR_IRGN1_WB_WA		(1 << 24)
100 #define IMTTBCR_IRGN1_WT		(2 << 24)
101 #define IMTTBCR_IRGN1_WB		(3 << 24)
102 #define IMTTBCR_IRGN1_MASK		(3 << 24)
103 #define IMTTBCR_TSZ1_MASK		(7 << 16)
104 #define IMTTBCR_TSZ1_SHIFT		16
105 #define IMTTBCR_SH0_NON_SHAREABLE	(0 << 12)
106 #define IMTTBCR_SH0_OUTER_SHAREABLE	(2 << 12)
107 #define IMTTBCR_SH0_INNER_SHAREABLE	(3 << 12)
108 #define IMTTBCR_SH0_MASK		(3 << 12)
109 #define IMTTBCR_ORGN0_NC		(0 << 10)
110 #define IMTTBCR_ORGN0_WB_WA		(1 << 10)
111 #define IMTTBCR_ORGN0_WT		(2 << 10)
112 #define IMTTBCR_ORGN0_WB		(3 << 10)
113 #define IMTTBCR_ORGN0_MASK		(3 << 10)
114 #define IMTTBCR_IRGN0_NC		(0 << 8)
115 #define IMTTBCR_IRGN0_WB_WA		(1 << 8)
116 #define IMTTBCR_IRGN0_WT		(2 << 8)
117 #define IMTTBCR_IRGN0_WB		(3 << 8)
118 #define IMTTBCR_IRGN0_MASK		(3 << 8)
119 #define IMTTBCR_SL0_LVL_2		(0 << 4)
120 #define IMTTBCR_SL0_LVL_1		(1 << 4)
121 #define IMTTBCR_TSZ0_MASK		(7 << 0)
122 #define IMTTBCR_TSZ0_SHIFT		O
123 
124 #define IMBUSCR				0x000c
125 #define IMBUSCR_DVM			(1 << 2)
126 #define IMBUSCR_BUSSEL_SYS		(0 << 0)
127 #define IMBUSCR_BUSSEL_CCI		(1 << 0)
128 #define IMBUSCR_BUSSEL_IMCAAR		(2 << 0)
129 #define IMBUSCR_BUSSEL_CCI_IMCAAR	(3 << 0)
130 #define IMBUSCR_BUSSEL_MASK		(3 << 0)
131 
132 #define IMTTLBR0			0x0010
133 #define IMTTUBR0			0x0014
134 #define IMTTLBR1			0x0018
135 #define IMTTUBR1			0x001c
136 
137 #define IMSTR				0x0020
138 #define IMSTR_ERRLVL_MASK		(3 << 12)
139 #define IMSTR_ERRLVL_SHIFT		12
140 #define IMSTR_ERRCODE_TLB_FORMAT	(1 << 8)
141 #define IMSTR_ERRCODE_ACCESS_PERM	(4 << 8)
142 #define IMSTR_ERRCODE_SECURE_ACCESS	(5 << 8)
143 #define IMSTR_ERRCODE_MASK		(7 << 8)
144 #define IMSTR_MHIT			(1 << 4)
145 #define IMSTR_ABORT			(1 << 2)
146 #define IMSTR_PF			(1 << 1)
147 #define IMSTR_TF			(1 << 0)
148 
149 #define IMMAIR0				0x0028
150 #define IMMAIR1				0x002c
151 #define IMMAIR_ATTR_MASK		0xff
152 #define IMMAIR_ATTR_DEVICE		0x04
153 #define IMMAIR_ATTR_NC			0x44
154 #define IMMAIR_ATTR_WBRWA		0xff
155 #define IMMAIR_ATTR_SHIFT(n)		((n) << 3)
156 #define IMMAIR_ATTR_IDX_NC		0
157 #define IMMAIR_ATTR_IDX_WBRWA		1
158 #define IMMAIR_ATTR_IDX_DEV		2
159 
160 #define IMEAR				0x0030
161 
162 #define IMPCTR				0x0200
163 #define IMPSTR				0x0208
164 #define IMPEAR				0x020c
165 #define IMPMBA(n)			(0x0280 + ((n) * 4))
166 #define IMPMBD(n)			(0x02c0 + ((n) * 4))
167 
168 #define IMUCTR(n)			(0x0300 + ((n) * 16))
169 #define IMUCTR_FIXADDEN			(1 << 31)
170 #define IMUCTR_FIXADD_MASK		(0xff << 16)
171 #define IMUCTR_FIXADD_SHIFT		16
172 #define IMUCTR_TTSEL_MMU(n)		((n) << 4)
173 #define IMUCTR_TTSEL_PMB		(8 << 4)
174 #define IMUCTR_TTSEL_MASK		(15 << 4)
175 #define IMUCTR_FLUSH			(1 << 1)
176 #define IMUCTR_MMUEN			(1 << 0)
177 
178 #define IMUASID(n)			(0x0308 + ((n) * 16))
179 #define IMUASID_ASID8_MASK		(0xff << 8)
180 #define IMUASID_ASID8_SHIFT		8
181 #define IMUASID_ASID0_MASK		(0xff << 0)
182 #define IMUASID_ASID0_SHIFT		0
183 
184 /* -----------------------------------------------------------------------------
185  * Read/Write Access
186  */
187 
188 static u32 ipmmu_read(struct ipmmu_vmsa_device *mmu, unsigned int offset)
189 {
190 	return ioread32(mmu->base + offset);
191 }
192 
193 static void ipmmu_write(struct ipmmu_vmsa_device *mmu, unsigned int offset,
194 			u32 data)
195 {
196 	iowrite32(data, mmu->base + offset);
197 }
198 
199 static u32 ipmmu_ctx_read(struct ipmmu_vmsa_domain *domain, unsigned int reg)
200 {
201 	return ipmmu_read(domain->mmu, domain->context_id * IM_CTX_SIZE + reg);
202 }
203 
204 static void ipmmu_ctx_write(struct ipmmu_vmsa_domain *domain, unsigned int reg,
205 			    u32 data)
206 {
207 	ipmmu_write(domain->mmu, domain->context_id * IM_CTX_SIZE + reg, data);
208 }
209 
210 /* -----------------------------------------------------------------------------
211  * TLB and microTLB Management
212  */
213 
214 /* Wait for any pending TLB invalidations to complete */
215 static void ipmmu_tlb_sync(struct ipmmu_vmsa_domain *domain)
216 {
217 	unsigned int count = 0;
218 
219 	while (ipmmu_ctx_read(domain, IMCTR) & IMCTR_FLUSH) {
220 		cpu_relax();
221 		if (++count == TLB_LOOP_TIMEOUT) {
222 			dev_err_ratelimited(domain->mmu->dev,
223 			"TLB sync timed out -- MMU may be deadlocked\n");
224 			return;
225 		}
226 		udelay(1);
227 	}
228 }
229 
230 static void ipmmu_tlb_invalidate(struct ipmmu_vmsa_domain *domain)
231 {
232 	u32 reg;
233 
234 	reg = ipmmu_ctx_read(domain, IMCTR);
235 	reg |= IMCTR_FLUSH;
236 	ipmmu_ctx_write(domain, IMCTR, reg);
237 
238 	ipmmu_tlb_sync(domain);
239 }
240 
241 /*
242  * Enable MMU translation for the microTLB.
243  */
244 static void ipmmu_utlb_enable(struct ipmmu_vmsa_domain *domain,
245 			      unsigned int utlb)
246 {
247 	struct ipmmu_vmsa_device *mmu = domain->mmu;
248 
249 	/*
250 	 * TODO: Reference-count the microTLB as several bus masters can be
251 	 * connected to the same microTLB.
252 	 */
253 
254 	/* TODO: What should we set the ASID to ? */
255 	ipmmu_write(mmu, IMUASID(utlb), 0);
256 	/* TODO: Do we need to flush the microTLB ? */
257 	ipmmu_write(mmu, IMUCTR(utlb),
258 		    IMUCTR_TTSEL_MMU(domain->context_id) | IMUCTR_FLUSH |
259 		    IMUCTR_MMUEN);
260 }
261 
262 /*
263  * Disable MMU translation for the microTLB.
264  */
265 static void ipmmu_utlb_disable(struct ipmmu_vmsa_domain *domain,
266 			       unsigned int utlb)
267 {
268 	struct ipmmu_vmsa_device *mmu = domain->mmu;
269 
270 	ipmmu_write(mmu, IMUCTR(utlb), 0);
271 }
272 
273 static void ipmmu_tlb_flush_all(void *cookie)
274 {
275 	struct ipmmu_vmsa_domain *domain = cookie;
276 
277 	ipmmu_tlb_invalidate(domain);
278 }
279 
280 static void ipmmu_tlb_add_flush(unsigned long iova, size_t size,
281 				size_t granule, bool leaf, void *cookie)
282 {
283 	/* The hardware doesn't support selective TLB flush. */
284 }
285 
286 static struct iommu_gather_ops ipmmu_gather_ops = {
287 	.tlb_flush_all = ipmmu_tlb_flush_all,
288 	.tlb_add_flush = ipmmu_tlb_add_flush,
289 	.tlb_sync = ipmmu_tlb_flush_all,
290 };
291 
292 /* -----------------------------------------------------------------------------
293  * Domain/Context Management
294  */
295 
296 static int ipmmu_domain_init_context(struct ipmmu_vmsa_domain *domain)
297 {
298 	u64 ttbr;
299 
300 	/*
301 	 * Allocate the page table operations.
302 	 *
303 	 * VMSA states in section B3.6.3 "Control of Secure or Non-secure memory
304 	 * access, Long-descriptor format" that the NStable bit being set in a
305 	 * table descriptor will result in the NStable and NS bits of all child
306 	 * entries being ignored and considered as being set. The IPMMU seems
307 	 * not to comply with this, as it generates a secure access page fault
308 	 * if any of the NStable and NS bits isn't set when running in
309 	 * non-secure mode.
310 	 */
311 	domain->cfg.quirks = IO_PGTABLE_QUIRK_ARM_NS;
312 	domain->cfg.pgsize_bitmap = SZ_1G | SZ_2M | SZ_4K,
313 	domain->cfg.ias = 32;
314 	domain->cfg.oas = 40;
315 	domain->cfg.tlb = &ipmmu_gather_ops;
316 	/*
317 	 * TODO: Add support for coherent walk through CCI with DVM and remove
318 	 * cache handling. For now, delegate it to the io-pgtable code.
319 	 */
320 	domain->cfg.iommu_dev = domain->mmu->dev;
321 
322 	domain->iop = alloc_io_pgtable_ops(ARM_32_LPAE_S1, &domain->cfg,
323 					   domain);
324 	if (!domain->iop)
325 		return -EINVAL;
326 
327 	/*
328 	 * TODO: When adding support for multiple contexts, find an unused
329 	 * context.
330 	 */
331 	domain->context_id = 0;
332 
333 	/* TTBR0 */
334 	ttbr = domain->cfg.arm_lpae_s1_cfg.ttbr[0];
335 	ipmmu_ctx_write(domain, IMTTLBR0, ttbr);
336 	ipmmu_ctx_write(domain, IMTTUBR0, ttbr >> 32);
337 
338 	/*
339 	 * TTBCR
340 	 * We use long descriptors with inner-shareable WBWA tables and allocate
341 	 * the whole 32-bit VA space to TTBR0.
342 	 */
343 	ipmmu_ctx_write(domain, IMTTBCR, IMTTBCR_EAE |
344 			IMTTBCR_SH0_INNER_SHAREABLE | IMTTBCR_ORGN0_WB_WA |
345 			IMTTBCR_IRGN0_WB_WA | IMTTBCR_SL0_LVL_1);
346 
347 	/* MAIR0 */
348 	ipmmu_ctx_write(domain, IMMAIR0, domain->cfg.arm_lpae_s1_cfg.mair[0]);
349 
350 	/* IMBUSCR */
351 	ipmmu_ctx_write(domain, IMBUSCR,
352 			ipmmu_ctx_read(domain, IMBUSCR) &
353 			~(IMBUSCR_DVM | IMBUSCR_BUSSEL_MASK));
354 
355 	/*
356 	 * IMSTR
357 	 * Clear all interrupt flags.
358 	 */
359 	ipmmu_ctx_write(domain, IMSTR, ipmmu_ctx_read(domain, IMSTR));
360 
361 	/*
362 	 * IMCTR
363 	 * Enable the MMU and interrupt generation. The long-descriptor
364 	 * translation table format doesn't use TEX remapping. Don't enable AF
365 	 * software management as we have no use for it. Flush the TLB as
366 	 * required when modifying the context registers.
367 	 */
368 	ipmmu_ctx_write(domain, IMCTR, IMCTR_INTEN | IMCTR_FLUSH | IMCTR_MMUEN);
369 
370 	return 0;
371 }
372 
373 static void ipmmu_domain_destroy_context(struct ipmmu_vmsa_domain *domain)
374 {
375 	/*
376 	 * Disable the context. Flush the TLB as required when modifying the
377 	 * context registers.
378 	 *
379 	 * TODO: Is TLB flush really needed ?
380 	 */
381 	ipmmu_ctx_write(domain, IMCTR, IMCTR_FLUSH);
382 	ipmmu_tlb_sync(domain);
383 }
384 
385 /* -----------------------------------------------------------------------------
386  * Fault Handling
387  */
388 
389 static irqreturn_t ipmmu_domain_irq(struct ipmmu_vmsa_domain *domain)
390 {
391 	const u32 err_mask = IMSTR_MHIT | IMSTR_ABORT | IMSTR_PF | IMSTR_TF;
392 	struct ipmmu_vmsa_device *mmu = domain->mmu;
393 	u32 status;
394 	u32 iova;
395 
396 	status = ipmmu_ctx_read(domain, IMSTR);
397 	if (!(status & err_mask))
398 		return IRQ_NONE;
399 
400 	iova = ipmmu_ctx_read(domain, IMEAR);
401 
402 	/*
403 	 * Clear the error status flags. Unlike traditional interrupt flag
404 	 * registers that must be cleared by writing 1, this status register
405 	 * seems to require 0. The error address register must be read before,
406 	 * otherwise its value will be 0.
407 	 */
408 	ipmmu_ctx_write(domain, IMSTR, 0);
409 
410 	/* Log fatal errors. */
411 	if (status & IMSTR_MHIT)
412 		dev_err_ratelimited(mmu->dev, "Multiple TLB hits @0x%08x\n",
413 				    iova);
414 	if (status & IMSTR_ABORT)
415 		dev_err_ratelimited(mmu->dev, "Page Table Walk Abort @0x%08x\n",
416 				    iova);
417 
418 	if (!(status & (IMSTR_PF | IMSTR_TF)))
419 		return IRQ_NONE;
420 
421 	/*
422 	 * Try to handle page faults and translation faults.
423 	 *
424 	 * TODO: We need to look up the faulty device based on the I/O VA. Use
425 	 * the IOMMU device for now.
426 	 */
427 	if (!report_iommu_fault(&domain->io_domain, mmu->dev, iova, 0))
428 		return IRQ_HANDLED;
429 
430 	dev_err_ratelimited(mmu->dev,
431 			    "Unhandled fault: status 0x%08x iova 0x%08x\n",
432 			    status, iova);
433 
434 	return IRQ_HANDLED;
435 }
436 
437 static irqreturn_t ipmmu_irq(int irq, void *dev)
438 {
439 	struct ipmmu_vmsa_device *mmu = dev;
440 	struct iommu_domain *io_domain;
441 	struct ipmmu_vmsa_domain *domain;
442 
443 	if (!mmu->mapping)
444 		return IRQ_NONE;
445 
446 	io_domain = mmu->mapping->domain;
447 	domain = to_vmsa_domain(io_domain);
448 
449 	return ipmmu_domain_irq(domain);
450 }
451 
452 /* -----------------------------------------------------------------------------
453  * IOMMU Operations
454  */
455 
456 static struct iommu_domain *ipmmu_domain_alloc(unsigned type)
457 {
458 	struct ipmmu_vmsa_domain *domain;
459 
460 	if (type != IOMMU_DOMAIN_UNMANAGED)
461 		return NULL;
462 
463 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
464 	if (!domain)
465 		return NULL;
466 
467 	spin_lock_init(&domain->lock);
468 
469 	return &domain->io_domain;
470 }
471 
472 static void ipmmu_domain_free(struct iommu_domain *io_domain)
473 {
474 	struct ipmmu_vmsa_domain *domain = to_vmsa_domain(io_domain);
475 
476 	/*
477 	 * Free the domain resources. We assume that all devices have already
478 	 * been detached.
479 	 */
480 	ipmmu_domain_destroy_context(domain);
481 	free_io_pgtable_ops(domain->iop);
482 	kfree(domain);
483 }
484 
485 static int ipmmu_attach_device(struct iommu_domain *io_domain,
486 			       struct device *dev)
487 {
488 	struct ipmmu_vmsa_archdata *archdata = dev->archdata.iommu;
489 	struct ipmmu_vmsa_device *mmu = archdata->mmu;
490 	struct ipmmu_vmsa_domain *domain = to_vmsa_domain(io_domain);
491 	unsigned long flags;
492 	unsigned int i;
493 	int ret = 0;
494 
495 	if (!mmu) {
496 		dev_err(dev, "Cannot attach to IPMMU\n");
497 		return -ENXIO;
498 	}
499 
500 	spin_lock_irqsave(&domain->lock, flags);
501 
502 	if (!domain->mmu) {
503 		/* The domain hasn't been used yet, initialize it. */
504 		domain->mmu = mmu;
505 		ret = ipmmu_domain_init_context(domain);
506 	} else if (domain->mmu != mmu) {
507 		/*
508 		 * Something is wrong, we can't attach two devices using
509 		 * different IOMMUs to the same domain.
510 		 */
511 		dev_err(dev, "Can't attach IPMMU %s to domain on IPMMU %s\n",
512 			dev_name(mmu->dev), dev_name(domain->mmu->dev));
513 		ret = -EINVAL;
514 	}
515 
516 	spin_unlock_irqrestore(&domain->lock, flags);
517 
518 	if (ret < 0)
519 		return ret;
520 
521 	for (i = 0; i < archdata->num_utlbs; ++i)
522 		ipmmu_utlb_enable(domain, archdata->utlbs[i]);
523 
524 	return 0;
525 }
526 
527 static void ipmmu_detach_device(struct iommu_domain *io_domain,
528 				struct device *dev)
529 {
530 	struct ipmmu_vmsa_archdata *archdata = dev->archdata.iommu;
531 	struct ipmmu_vmsa_domain *domain = to_vmsa_domain(io_domain);
532 	unsigned int i;
533 
534 	for (i = 0; i < archdata->num_utlbs; ++i)
535 		ipmmu_utlb_disable(domain, archdata->utlbs[i]);
536 
537 	/*
538 	 * TODO: Optimize by disabling the context when no device is attached.
539 	 */
540 }
541 
542 static int ipmmu_map(struct iommu_domain *io_domain, unsigned long iova,
543 		     phys_addr_t paddr, size_t size, int prot)
544 {
545 	struct ipmmu_vmsa_domain *domain = to_vmsa_domain(io_domain);
546 
547 	if (!domain)
548 		return -ENODEV;
549 
550 	return domain->iop->map(domain->iop, iova, paddr, size, prot);
551 }
552 
553 static size_t ipmmu_unmap(struct iommu_domain *io_domain, unsigned long iova,
554 			  size_t size)
555 {
556 	struct ipmmu_vmsa_domain *domain = to_vmsa_domain(io_domain);
557 
558 	return domain->iop->unmap(domain->iop, iova, size);
559 }
560 
561 static phys_addr_t ipmmu_iova_to_phys(struct iommu_domain *io_domain,
562 				      dma_addr_t iova)
563 {
564 	struct ipmmu_vmsa_domain *domain = to_vmsa_domain(io_domain);
565 
566 	/* TODO: Is locking needed ? */
567 
568 	return domain->iop->iova_to_phys(domain->iop, iova);
569 }
570 
571 static int ipmmu_find_utlbs(struct ipmmu_vmsa_device *mmu, struct device *dev,
572 			    unsigned int *utlbs, unsigned int num_utlbs)
573 {
574 	unsigned int i;
575 
576 	for (i = 0; i < num_utlbs; ++i) {
577 		struct of_phandle_args args;
578 		int ret;
579 
580 		ret = of_parse_phandle_with_args(dev->of_node, "iommus",
581 						 "#iommu-cells", i, &args);
582 		if (ret < 0)
583 			return ret;
584 
585 		of_node_put(args.np);
586 
587 		if (args.np != mmu->dev->of_node || args.args_count != 1)
588 			return -EINVAL;
589 
590 		utlbs[i] = args.args[0];
591 	}
592 
593 	return 0;
594 }
595 
596 static int ipmmu_add_device(struct device *dev)
597 {
598 	struct ipmmu_vmsa_archdata *archdata;
599 	struct ipmmu_vmsa_device *mmu;
600 	struct iommu_group *group = NULL;
601 	unsigned int *utlbs;
602 	unsigned int i;
603 	int num_utlbs;
604 	int ret = -ENODEV;
605 
606 	if (dev->archdata.iommu) {
607 		dev_warn(dev, "IOMMU driver already assigned to device %s\n",
608 			 dev_name(dev));
609 		return -EINVAL;
610 	}
611 
612 	/* Find the master corresponding to the device. */
613 
614 	num_utlbs = of_count_phandle_with_args(dev->of_node, "iommus",
615 					       "#iommu-cells");
616 	if (num_utlbs < 0)
617 		return -ENODEV;
618 
619 	utlbs = kcalloc(num_utlbs, sizeof(*utlbs), GFP_KERNEL);
620 	if (!utlbs)
621 		return -ENOMEM;
622 
623 	spin_lock(&ipmmu_devices_lock);
624 
625 	list_for_each_entry(mmu, &ipmmu_devices, list) {
626 		ret = ipmmu_find_utlbs(mmu, dev, utlbs, num_utlbs);
627 		if (!ret) {
628 			/*
629 			 * TODO Take a reference to the MMU to protect
630 			 * against device removal.
631 			 */
632 			break;
633 		}
634 	}
635 
636 	spin_unlock(&ipmmu_devices_lock);
637 
638 	if (ret < 0)
639 		return -ENODEV;
640 
641 	for (i = 0; i < num_utlbs; ++i) {
642 		if (utlbs[i] >= mmu->num_utlbs) {
643 			ret = -EINVAL;
644 			goto error;
645 		}
646 	}
647 
648 	/* Create a device group and add the device to it. */
649 	group = iommu_group_alloc();
650 	if (IS_ERR(group)) {
651 		dev_err(dev, "Failed to allocate IOMMU group\n");
652 		ret = PTR_ERR(group);
653 		goto error;
654 	}
655 
656 	ret = iommu_group_add_device(group, dev);
657 	iommu_group_put(group);
658 
659 	if (ret < 0) {
660 		dev_err(dev, "Failed to add device to IPMMU group\n");
661 		group = NULL;
662 		goto error;
663 	}
664 
665 	archdata = kzalloc(sizeof(*archdata), GFP_KERNEL);
666 	if (!archdata) {
667 		ret = -ENOMEM;
668 		goto error;
669 	}
670 
671 	archdata->mmu = mmu;
672 	archdata->utlbs = utlbs;
673 	archdata->num_utlbs = num_utlbs;
674 	dev->archdata.iommu = archdata;
675 
676 	/*
677 	 * Create the ARM mapping, used by the ARM DMA mapping core to allocate
678 	 * VAs. This will allocate a corresponding IOMMU domain.
679 	 *
680 	 * TODO:
681 	 * - Create one mapping per context (TLB).
682 	 * - Make the mapping size configurable ? We currently use a 2GB mapping
683 	 *   at a 1GB offset to ensure that NULL VAs will fault.
684 	 */
685 	if (!mmu->mapping) {
686 		struct dma_iommu_mapping *mapping;
687 
688 		mapping = arm_iommu_create_mapping(&platform_bus_type,
689 						   SZ_1G, SZ_2G);
690 		if (IS_ERR(mapping)) {
691 			dev_err(mmu->dev, "failed to create ARM IOMMU mapping\n");
692 			ret = PTR_ERR(mapping);
693 			goto error;
694 		}
695 
696 		mmu->mapping = mapping;
697 	}
698 
699 	/* Attach the ARM VA mapping to the device. */
700 	ret = arm_iommu_attach_device(dev, mmu->mapping);
701 	if (ret < 0) {
702 		dev_err(dev, "Failed to attach device to VA mapping\n");
703 		goto error;
704 	}
705 
706 	return 0;
707 
708 error:
709 	arm_iommu_release_mapping(mmu->mapping);
710 
711 	kfree(dev->archdata.iommu);
712 	kfree(utlbs);
713 
714 	dev->archdata.iommu = NULL;
715 
716 	if (!IS_ERR_OR_NULL(group))
717 		iommu_group_remove_device(dev);
718 
719 	return ret;
720 }
721 
722 static void ipmmu_remove_device(struct device *dev)
723 {
724 	struct ipmmu_vmsa_archdata *archdata = dev->archdata.iommu;
725 
726 	arm_iommu_detach_device(dev);
727 	iommu_group_remove_device(dev);
728 
729 	kfree(archdata->utlbs);
730 	kfree(archdata);
731 
732 	dev->archdata.iommu = NULL;
733 }
734 
735 static const struct iommu_ops ipmmu_ops = {
736 	.domain_alloc = ipmmu_domain_alloc,
737 	.domain_free = ipmmu_domain_free,
738 	.attach_dev = ipmmu_attach_device,
739 	.detach_dev = ipmmu_detach_device,
740 	.map = ipmmu_map,
741 	.unmap = ipmmu_unmap,
742 	.map_sg = default_iommu_map_sg,
743 	.iova_to_phys = ipmmu_iova_to_phys,
744 	.add_device = ipmmu_add_device,
745 	.remove_device = ipmmu_remove_device,
746 	.pgsize_bitmap = SZ_1G | SZ_2M | SZ_4K,
747 };
748 
749 /* -----------------------------------------------------------------------------
750  * Probe/remove and init
751  */
752 
753 static void ipmmu_device_reset(struct ipmmu_vmsa_device *mmu)
754 {
755 	unsigned int i;
756 
757 	/* Disable all contexts. */
758 	for (i = 0; i < 4; ++i)
759 		ipmmu_write(mmu, i * IM_CTX_SIZE + IMCTR, 0);
760 }
761 
762 static int ipmmu_probe(struct platform_device *pdev)
763 {
764 	struct ipmmu_vmsa_device *mmu;
765 	struct resource *res;
766 	int irq;
767 	int ret;
768 
769 	if (!IS_ENABLED(CONFIG_OF) && !pdev->dev.platform_data) {
770 		dev_err(&pdev->dev, "missing platform data\n");
771 		return -EINVAL;
772 	}
773 
774 	mmu = devm_kzalloc(&pdev->dev, sizeof(*mmu), GFP_KERNEL);
775 	if (!mmu) {
776 		dev_err(&pdev->dev, "cannot allocate device data\n");
777 		return -ENOMEM;
778 	}
779 
780 	mmu->dev = &pdev->dev;
781 	mmu->num_utlbs = 32;
782 
783 	/* Map I/O memory and request IRQ. */
784 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
785 	mmu->base = devm_ioremap_resource(&pdev->dev, res);
786 	if (IS_ERR(mmu->base))
787 		return PTR_ERR(mmu->base);
788 
789 	/*
790 	 * The IPMMU has two register banks, for secure and non-secure modes.
791 	 * The bank mapped at the beginning of the IPMMU address space
792 	 * corresponds to the running mode of the CPU. When running in secure
793 	 * mode the non-secure register bank is also available at an offset.
794 	 *
795 	 * Secure mode operation isn't clearly documented and is thus currently
796 	 * not implemented in the driver. Furthermore, preliminary tests of
797 	 * non-secure operation with the main register bank were not successful.
798 	 * Offset the registers base unconditionally to point to the non-secure
799 	 * alias space for now.
800 	 */
801 	mmu->base += IM_NS_ALIAS_OFFSET;
802 
803 	irq = platform_get_irq(pdev, 0);
804 	if (irq < 0) {
805 		dev_err(&pdev->dev, "no IRQ found\n");
806 		return irq;
807 	}
808 
809 	ret = devm_request_irq(&pdev->dev, irq, ipmmu_irq, 0,
810 			       dev_name(&pdev->dev), mmu);
811 	if (ret < 0) {
812 		dev_err(&pdev->dev, "failed to request IRQ %d\n", irq);
813 		return ret;
814 	}
815 
816 	ipmmu_device_reset(mmu);
817 
818 	/*
819 	 * We can't create the ARM mapping here as it requires the bus to have
820 	 * an IOMMU, which only happens when bus_set_iommu() is called in
821 	 * ipmmu_init() after the probe function returns.
822 	 */
823 
824 	spin_lock(&ipmmu_devices_lock);
825 	list_add(&mmu->list, &ipmmu_devices);
826 	spin_unlock(&ipmmu_devices_lock);
827 
828 	platform_set_drvdata(pdev, mmu);
829 
830 	return 0;
831 }
832 
833 static int ipmmu_remove(struct platform_device *pdev)
834 {
835 	struct ipmmu_vmsa_device *mmu = platform_get_drvdata(pdev);
836 
837 	spin_lock(&ipmmu_devices_lock);
838 	list_del(&mmu->list);
839 	spin_unlock(&ipmmu_devices_lock);
840 
841 	arm_iommu_release_mapping(mmu->mapping);
842 
843 	ipmmu_device_reset(mmu);
844 
845 	return 0;
846 }
847 
848 static const struct of_device_id ipmmu_of_ids[] = {
849 	{ .compatible = "renesas,ipmmu-vmsa", },
850 	{ }
851 };
852 
853 static struct platform_driver ipmmu_driver = {
854 	.driver = {
855 		.name = "ipmmu-vmsa",
856 		.of_match_table = of_match_ptr(ipmmu_of_ids),
857 	},
858 	.probe = ipmmu_probe,
859 	.remove	= ipmmu_remove,
860 };
861 
862 static int __init ipmmu_init(void)
863 {
864 	int ret;
865 
866 	ret = platform_driver_register(&ipmmu_driver);
867 	if (ret < 0)
868 		return ret;
869 
870 	if (!iommu_present(&platform_bus_type))
871 		bus_set_iommu(&platform_bus_type, &ipmmu_ops);
872 
873 	return 0;
874 }
875 
876 static void __exit ipmmu_exit(void)
877 {
878 	return platform_driver_unregister(&ipmmu_driver);
879 }
880 
881 subsys_initcall(ipmmu_init);
882 module_exit(ipmmu_exit);
883 
884 MODULE_DESCRIPTION("IOMMU API for Renesas VMSA-compatible IPMMU");
885 MODULE_AUTHOR("Laurent Pinchart <laurent.pinchart@ideasonboard.com>");
886 MODULE_LICENSE("GPL v2");
887