xref: /openbmc/linux/drivers/iommu/exynos-iommu.c (revision 0edbfea5)
1 /*
2  * Copyright (c) 2011,2016 Samsung Electronics Co., Ltd.
3  *		http://www.samsung.com
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License version 2 as
7  * published by the Free Software Foundation.
8  */
9 
10 #ifdef CONFIG_EXYNOS_IOMMU_DEBUG
11 #define DEBUG
12 #endif
13 
14 #include <linux/clk.h>
15 #include <linux/dma-mapping.h>
16 #include <linux/err.h>
17 #include <linux/io.h>
18 #include <linux/iommu.h>
19 #include <linux/interrupt.h>
20 #include <linux/list.h>
21 #include <linux/of.h>
22 #include <linux/of_iommu.h>
23 #include <linux/of_platform.h>
24 #include <linux/platform_device.h>
25 #include <linux/pm_runtime.h>
26 #include <linux/slab.h>
27 #include <linux/dma-iommu.h>
28 
29 typedef u32 sysmmu_iova_t;
30 typedef u32 sysmmu_pte_t;
31 
32 /* We do not consider super section mapping (16MB) */
33 #define SECT_ORDER 20
34 #define LPAGE_ORDER 16
35 #define SPAGE_ORDER 12
36 
37 #define SECT_SIZE (1 << SECT_ORDER)
38 #define LPAGE_SIZE (1 << LPAGE_ORDER)
39 #define SPAGE_SIZE (1 << SPAGE_ORDER)
40 
41 #define SECT_MASK (~(SECT_SIZE - 1))
42 #define LPAGE_MASK (~(LPAGE_SIZE - 1))
43 #define SPAGE_MASK (~(SPAGE_SIZE - 1))
44 
45 #define lv1ent_fault(sent) ((*(sent) == ZERO_LV2LINK) || \
46 			   ((*(sent) & 3) == 0) || ((*(sent) & 3) == 3))
47 #define lv1ent_zero(sent) (*(sent) == ZERO_LV2LINK)
48 #define lv1ent_page_zero(sent) ((*(sent) & 3) == 1)
49 #define lv1ent_page(sent) ((*(sent) != ZERO_LV2LINK) && \
50 			  ((*(sent) & 3) == 1))
51 #define lv1ent_section(sent) ((*(sent) & 3) == 2)
52 
53 #define lv2ent_fault(pent) ((*(pent) & 3) == 0)
54 #define lv2ent_small(pent) ((*(pent) & 2) == 2)
55 #define lv2ent_large(pent) ((*(pent) & 3) == 1)
56 
57 /*
58  * v1.x - v3.x SYSMMU supports 32bit physical and 32bit virtual address spaces
59  * v5.0 introduced support for 36bit physical address space by shifting
60  * all page entry values by 4 bits.
61  * All SYSMMU controllers in the system support the address spaces of the same
62  * size, so PG_ENT_SHIFT can be initialized on first SYSMMU probe to proper
63  * value (0 or 4).
64  */
65 static short PG_ENT_SHIFT = -1;
66 #define SYSMMU_PG_ENT_SHIFT 0
67 #define SYSMMU_V5_PG_ENT_SHIFT 4
68 
69 #define sect_to_phys(ent) (((phys_addr_t) ent) << PG_ENT_SHIFT)
70 #define section_phys(sent) (sect_to_phys(*(sent)) & SECT_MASK)
71 #define section_offs(iova) (iova & (SECT_SIZE - 1))
72 #define lpage_phys(pent) (sect_to_phys(*(pent)) & LPAGE_MASK)
73 #define lpage_offs(iova) (iova & (LPAGE_SIZE - 1))
74 #define spage_phys(pent) (sect_to_phys(*(pent)) & SPAGE_MASK)
75 #define spage_offs(iova) (iova & (SPAGE_SIZE - 1))
76 
77 #define NUM_LV1ENTRIES 4096
78 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
79 
80 static u32 lv1ent_offset(sysmmu_iova_t iova)
81 {
82 	return iova >> SECT_ORDER;
83 }
84 
85 static u32 lv2ent_offset(sysmmu_iova_t iova)
86 {
87 	return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
88 }
89 
90 #define LV1TABLE_SIZE (NUM_LV1ENTRIES * sizeof(sysmmu_pte_t))
91 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
92 
93 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
94 #define lv2table_base(sent) (sect_to_phys(*(sent) & 0xFFFFFFC0))
95 
96 #define mk_lv1ent_sect(pa) ((pa >> PG_ENT_SHIFT) | 2)
97 #define mk_lv1ent_page(pa) ((pa >> PG_ENT_SHIFT) | 1)
98 #define mk_lv2ent_lpage(pa) ((pa >> PG_ENT_SHIFT) | 1)
99 #define mk_lv2ent_spage(pa) ((pa >> PG_ENT_SHIFT) | 2)
100 
101 #define CTRL_ENABLE	0x5
102 #define CTRL_BLOCK	0x7
103 #define CTRL_DISABLE	0x0
104 
105 #define CFG_LRU		0x1
106 #define CFG_QOS(n)	((n & 0xF) << 7)
107 #define CFG_ACGEN	(1 << 24) /* System MMU 3.3 only */
108 #define CFG_SYSSEL	(1 << 22) /* System MMU 3.2 only */
109 #define CFG_FLPDCACHE	(1 << 20) /* System MMU 3.2+ only */
110 
111 /* common registers */
112 #define REG_MMU_CTRL		0x000
113 #define REG_MMU_CFG		0x004
114 #define REG_MMU_STATUS		0x008
115 #define REG_MMU_VERSION		0x034
116 
117 #define MMU_MAJ_VER(val)	((val) >> 7)
118 #define MMU_MIN_VER(val)	((val) & 0x7F)
119 #define MMU_RAW_VER(reg)	(((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
120 
121 #define MAKE_MMU_VER(maj, min)	((((maj) & 0xF) << 7) | ((min) & 0x7F))
122 
123 /* v1.x - v3.x registers */
124 #define REG_MMU_FLUSH		0x00C
125 #define REG_MMU_FLUSH_ENTRY	0x010
126 #define REG_PT_BASE_ADDR	0x014
127 #define REG_INT_STATUS		0x018
128 #define REG_INT_CLEAR		0x01C
129 
130 #define REG_PAGE_FAULT_ADDR	0x024
131 #define REG_AW_FAULT_ADDR	0x028
132 #define REG_AR_FAULT_ADDR	0x02C
133 #define REG_DEFAULT_SLAVE_ADDR	0x030
134 
135 /* v5.x registers */
136 #define REG_V5_PT_BASE_PFN	0x00C
137 #define REG_V5_MMU_FLUSH_ALL	0x010
138 #define REG_V5_MMU_FLUSH_ENTRY	0x014
139 #define REG_V5_INT_STATUS	0x060
140 #define REG_V5_INT_CLEAR	0x064
141 #define REG_V5_FAULT_AR_VA	0x070
142 #define REG_V5_FAULT_AW_VA	0x080
143 
144 #define has_sysmmu(dev)		(dev->archdata.iommu != NULL)
145 
146 static struct device *dma_dev;
147 static struct kmem_cache *lv2table_kmem_cache;
148 static sysmmu_pte_t *zero_lv2_table;
149 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
150 
151 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
152 {
153 	return pgtable + lv1ent_offset(iova);
154 }
155 
156 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
157 {
158 	return (sysmmu_pte_t *)phys_to_virt(
159 				lv2table_base(sent)) + lv2ent_offset(iova);
160 }
161 
162 /*
163  * IOMMU fault information register
164  */
165 struct sysmmu_fault_info {
166 	unsigned int bit;	/* bit number in STATUS register */
167 	unsigned short addr_reg; /* register to read VA fault address */
168 	const char *name;	/* human readable fault name */
169 	unsigned int type;	/* fault type for report_iommu_fault */
170 };
171 
172 static const struct sysmmu_fault_info sysmmu_faults[] = {
173 	{ 0, REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
174 	{ 1, REG_AR_FAULT_ADDR, "AR MULTI-HIT", IOMMU_FAULT_READ },
175 	{ 2, REG_AW_FAULT_ADDR, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
176 	{ 3, REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
177 	{ 4, REG_AR_FAULT_ADDR, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
178 	{ 5, REG_AR_FAULT_ADDR, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
179 	{ 6, REG_AW_FAULT_ADDR, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
180 	{ 7, REG_AW_FAULT_ADDR, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
181 };
182 
183 static const struct sysmmu_fault_info sysmmu_v5_faults[] = {
184 	{ 0, REG_V5_FAULT_AR_VA, "AR PTW", IOMMU_FAULT_READ },
185 	{ 1, REG_V5_FAULT_AR_VA, "AR PAGE", IOMMU_FAULT_READ },
186 	{ 2, REG_V5_FAULT_AR_VA, "AR MULTI-HIT", IOMMU_FAULT_READ },
187 	{ 3, REG_V5_FAULT_AR_VA, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
188 	{ 4, REG_V5_FAULT_AR_VA, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
189 	{ 16, REG_V5_FAULT_AW_VA, "AW PTW", IOMMU_FAULT_WRITE },
190 	{ 17, REG_V5_FAULT_AW_VA, "AW PAGE", IOMMU_FAULT_WRITE },
191 	{ 18, REG_V5_FAULT_AW_VA, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
192 	{ 19, REG_V5_FAULT_AW_VA, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
193 	{ 20, REG_V5_FAULT_AW_VA, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
194 };
195 
196 /*
197  * This structure is attached to dev.archdata.iommu of the master device
198  * on device add, contains a list of SYSMMU controllers defined by device tree,
199  * which are bound to given master device. It is usually referenced by 'owner'
200  * pointer.
201 */
202 struct exynos_iommu_owner {
203 	struct list_head controllers;	/* list of sysmmu_drvdata.owner_node */
204 	struct iommu_domain *domain;	/* domain this device is attached */
205 };
206 
207 /*
208  * This structure exynos specific generalization of struct iommu_domain.
209  * It contains list of SYSMMU controllers from all master devices, which has
210  * been attached to this domain and page tables of IO address space defined by
211  * it. It is usually referenced by 'domain' pointer.
212  */
213 struct exynos_iommu_domain {
214 	struct list_head clients; /* list of sysmmu_drvdata.domain_node */
215 	sysmmu_pte_t *pgtable;	/* lv1 page table, 16KB */
216 	short *lv2entcnt;	/* free lv2 entry counter for each section */
217 	spinlock_t lock;	/* lock for modyfying list of clients */
218 	spinlock_t pgtablelock;	/* lock for modifying page table @ pgtable */
219 	struct iommu_domain domain; /* generic domain data structure */
220 };
221 
222 /*
223  * This structure hold all data of a single SYSMMU controller, this includes
224  * hw resources like registers and clocks, pointers and list nodes to connect
225  * it to all other structures, internal state and parameters read from device
226  * tree. It is usually referenced by 'data' pointer.
227  */
228 struct sysmmu_drvdata {
229 	struct device *sysmmu;		/* SYSMMU controller device */
230 	struct device *master;		/* master device (owner) */
231 	void __iomem *sfrbase;		/* our registers */
232 	struct clk *clk;		/* SYSMMU's clock */
233 	struct clk *aclk;		/* SYSMMU's aclk clock */
234 	struct clk *pclk;		/* SYSMMU's pclk clock */
235 	struct clk *clk_master;		/* master's device clock */
236 	int activations;		/* number of calls to sysmmu_enable */
237 	spinlock_t lock;		/* lock for modyfying state */
238 	struct exynos_iommu_domain *domain; /* domain we belong to */
239 	struct list_head domain_node;	/* node for domain clients list */
240 	struct list_head owner_node;	/* node for owner controllers list */
241 	phys_addr_t pgtable;		/* assigned page table structure */
242 	unsigned int version;		/* our version */
243 };
244 
245 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
246 {
247 	return container_of(dom, struct exynos_iommu_domain, domain);
248 }
249 
250 static bool set_sysmmu_active(struct sysmmu_drvdata *data)
251 {
252 	/* return true if the System MMU was not active previously
253 	   and it needs to be initialized */
254 	return ++data->activations == 1;
255 }
256 
257 static bool set_sysmmu_inactive(struct sysmmu_drvdata *data)
258 {
259 	/* return true if the System MMU is needed to be disabled */
260 	BUG_ON(data->activations < 1);
261 	return --data->activations == 0;
262 }
263 
264 static bool is_sysmmu_active(struct sysmmu_drvdata *data)
265 {
266 	return data->activations > 0;
267 }
268 
269 static void sysmmu_unblock(struct sysmmu_drvdata *data)
270 {
271 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
272 }
273 
274 static bool sysmmu_block(struct sysmmu_drvdata *data)
275 {
276 	int i = 120;
277 
278 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
279 	while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
280 		--i;
281 
282 	if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
283 		sysmmu_unblock(data);
284 		return false;
285 	}
286 
287 	return true;
288 }
289 
290 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
291 {
292 	if (MMU_MAJ_VER(data->version) < 5)
293 		writel(0x1, data->sfrbase + REG_MMU_FLUSH);
294 	else
295 		writel(0x1, data->sfrbase + REG_V5_MMU_FLUSH_ALL);
296 }
297 
298 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
299 				sysmmu_iova_t iova, unsigned int num_inv)
300 {
301 	unsigned int i;
302 
303 	for (i = 0; i < num_inv; i++) {
304 		if (MMU_MAJ_VER(data->version) < 5)
305 			writel((iova & SPAGE_MASK) | 1,
306 				     data->sfrbase + REG_MMU_FLUSH_ENTRY);
307 		else
308 			writel((iova & SPAGE_MASK) | 1,
309 				     data->sfrbase + REG_V5_MMU_FLUSH_ENTRY);
310 		iova += SPAGE_SIZE;
311 	}
312 }
313 
314 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
315 {
316 	if (MMU_MAJ_VER(data->version) < 5)
317 		writel(pgd, data->sfrbase + REG_PT_BASE_ADDR);
318 	else
319 		writel(pgd >> PAGE_SHIFT,
320 			     data->sfrbase + REG_V5_PT_BASE_PFN);
321 
322 	__sysmmu_tlb_invalidate(data);
323 }
324 
325 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
326 {
327 	u32 ver;
328 
329 	clk_enable(data->clk_master);
330 	clk_enable(data->clk);
331 	clk_enable(data->pclk);
332 	clk_enable(data->aclk);
333 
334 	ver = readl(data->sfrbase + REG_MMU_VERSION);
335 
336 	/* controllers on some SoCs don't report proper version */
337 	if (ver == 0x80000001u)
338 		data->version = MAKE_MMU_VER(1, 0);
339 	else
340 		data->version = MMU_RAW_VER(ver);
341 
342 	dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
343 		MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
344 
345 	clk_disable(data->aclk);
346 	clk_disable(data->pclk);
347 	clk_disable(data->clk);
348 	clk_disable(data->clk_master);
349 }
350 
351 static void show_fault_information(struct sysmmu_drvdata *data,
352 				   const struct sysmmu_fault_info *finfo,
353 				   sysmmu_iova_t fault_addr)
354 {
355 	sysmmu_pte_t *ent;
356 
357 	dev_err(data->sysmmu, "%s FAULT occurred at %#x (page table base: %pa)\n",
358 		finfo->name, fault_addr, &data->pgtable);
359 	ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
360 	dev_err(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
361 	if (lv1ent_page(ent)) {
362 		ent = page_entry(ent, fault_addr);
363 		dev_err(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
364 	}
365 }
366 
367 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
368 {
369 	/* SYSMMU is in blocked state when interrupt occurred. */
370 	struct sysmmu_drvdata *data = dev_id;
371 	const struct sysmmu_fault_info *finfo;
372 	unsigned int i, n, itype;
373 	sysmmu_iova_t fault_addr = -1;
374 	unsigned short reg_status, reg_clear;
375 	int ret = -ENOSYS;
376 
377 	WARN_ON(!is_sysmmu_active(data));
378 
379 	if (MMU_MAJ_VER(data->version) < 5) {
380 		reg_status = REG_INT_STATUS;
381 		reg_clear = REG_INT_CLEAR;
382 		finfo = sysmmu_faults;
383 		n = ARRAY_SIZE(sysmmu_faults);
384 	} else {
385 		reg_status = REG_V5_INT_STATUS;
386 		reg_clear = REG_V5_INT_CLEAR;
387 		finfo = sysmmu_v5_faults;
388 		n = ARRAY_SIZE(sysmmu_v5_faults);
389 	}
390 
391 	spin_lock(&data->lock);
392 
393 	clk_enable(data->clk_master);
394 
395 	itype = __ffs(readl(data->sfrbase + reg_status));
396 	for (i = 0; i < n; i++, finfo++)
397 		if (finfo->bit == itype)
398 			break;
399 	/* unknown/unsupported fault */
400 	BUG_ON(i == n);
401 
402 	/* print debug message */
403 	fault_addr = readl(data->sfrbase + finfo->addr_reg);
404 	show_fault_information(data, finfo, fault_addr);
405 
406 	if (data->domain)
407 		ret = report_iommu_fault(&data->domain->domain,
408 					data->master, fault_addr, finfo->type);
409 	/* fault is not recovered by fault handler */
410 	BUG_ON(ret != 0);
411 
412 	writel(1 << itype, data->sfrbase + reg_clear);
413 
414 	sysmmu_unblock(data);
415 
416 	clk_disable(data->clk_master);
417 
418 	spin_unlock(&data->lock);
419 
420 	return IRQ_HANDLED;
421 }
422 
423 static void __sysmmu_disable_nocount(struct sysmmu_drvdata *data)
424 {
425 	clk_enable(data->clk_master);
426 
427 	writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
428 	writel(0, data->sfrbase + REG_MMU_CFG);
429 
430 	clk_disable(data->aclk);
431 	clk_disable(data->pclk);
432 	clk_disable(data->clk);
433 	clk_disable(data->clk_master);
434 }
435 
436 static bool __sysmmu_disable(struct sysmmu_drvdata *data)
437 {
438 	bool disabled;
439 	unsigned long flags;
440 
441 	spin_lock_irqsave(&data->lock, flags);
442 
443 	disabled = set_sysmmu_inactive(data);
444 
445 	if (disabled) {
446 		data->pgtable = 0;
447 		data->domain = NULL;
448 
449 		__sysmmu_disable_nocount(data);
450 
451 		dev_dbg(data->sysmmu, "Disabled\n");
452 	} else  {
453 		dev_dbg(data->sysmmu, "%d times left to disable\n",
454 					data->activations);
455 	}
456 
457 	spin_unlock_irqrestore(&data->lock, flags);
458 
459 	return disabled;
460 }
461 
462 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
463 {
464 	unsigned int cfg;
465 
466 	if (data->version <= MAKE_MMU_VER(3, 1))
467 		cfg = CFG_LRU | CFG_QOS(15);
468 	else if (data->version <= MAKE_MMU_VER(3, 2))
469 		cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
470 	else
471 		cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
472 
473 	writel(cfg, data->sfrbase + REG_MMU_CFG);
474 }
475 
476 static void __sysmmu_enable_nocount(struct sysmmu_drvdata *data)
477 {
478 	clk_enable(data->clk_master);
479 	clk_enable(data->clk);
480 	clk_enable(data->pclk);
481 	clk_enable(data->aclk);
482 
483 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
484 
485 	__sysmmu_init_config(data);
486 
487 	__sysmmu_set_ptbase(data, data->pgtable);
488 
489 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
490 
491 	clk_disable(data->clk_master);
492 }
493 
494 static int __sysmmu_enable(struct sysmmu_drvdata *data, phys_addr_t pgtable,
495 			   struct exynos_iommu_domain *domain)
496 {
497 	int ret = 0;
498 	unsigned long flags;
499 
500 	spin_lock_irqsave(&data->lock, flags);
501 	if (set_sysmmu_active(data)) {
502 		data->pgtable = pgtable;
503 		data->domain = domain;
504 
505 		__sysmmu_enable_nocount(data);
506 
507 		dev_dbg(data->sysmmu, "Enabled\n");
508 	} else {
509 		ret = (pgtable == data->pgtable) ? 1 : -EBUSY;
510 
511 		dev_dbg(data->sysmmu, "already enabled\n");
512 	}
513 
514 	if (WARN_ON(ret < 0))
515 		set_sysmmu_inactive(data); /* decrement count */
516 
517 	spin_unlock_irqrestore(&data->lock, flags);
518 
519 	return ret;
520 }
521 
522 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
523 					    sysmmu_iova_t iova)
524 {
525 	unsigned long flags;
526 
527 	clk_enable(data->clk_master);
528 
529 	spin_lock_irqsave(&data->lock, flags);
530 	if (is_sysmmu_active(data)) {
531 		if (data->version >= MAKE_MMU_VER(3, 3))
532 			__sysmmu_tlb_invalidate_entry(data, iova, 1);
533 	}
534 	spin_unlock_irqrestore(&data->lock, flags);
535 
536 	clk_disable(data->clk_master);
537 }
538 
539 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
540 					sysmmu_iova_t iova, size_t size)
541 {
542 	unsigned long flags;
543 
544 	spin_lock_irqsave(&data->lock, flags);
545 	if (is_sysmmu_active(data)) {
546 		unsigned int num_inv = 1;
547 
548 		clk_enable(data->clk_master);
549 
550 		/*
551 		 * L2TLB invalidation required
552 		 * 4KB page: 1 invalidation
553 		 * 64KB page: 16 invalidations
554 		 * 1MB page: 64 invalidations
555 		 * because it is set-associative TLB
556 		 * with 8-way and 64 sets.
557 		 * 1MB page can be cached in one of all sets.
558 		 * 64KB page can be one of 16 consecutive sets.
559 		 */
560 		if (MMU_MAJ_VER(data->version) == 2)
561 			num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
562 
563 		if (sysmmu_block(data)) {
564 			__sysmmu_tlb_invalidate_entry(data, iova, num_inv);
565 			sysmmu_unblock(data);
566 		}
567 		clk_disable(data->clk_master);
568 	} else {
569 		dev_dbg(data->master,
570 			"disabled. Skipping TLB invalidation @ %#x\n", iova);
571 	}
572 	spin_unlock_irqrestore(&data->lock, flags);
573 }
574 
575 static int __init exynos_sysmmu_probe(struct platform_device *pdev)
576 {
577 	int irq, ret;
578 	struct device *dev = &pdev->dev;
579 	struct sysmmu_drvdata *data;
580 	struct resource *res;
581 
582 	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
583 	if (!data)
584 		return -ENOMEM;
585 
586 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
587 	data->sfrbase = devm_ioremap_resource(dev, res);
588 	if (IS_ERR(data->sfrbase))
589 		return PTR_ERR(data->sfrbase);
590 
591 	irq = platform_get_irq(pdev, 0);
592 	if (irq <= 0) {
593 		dev_err(dev, "Unable to find IRQ resource\n");
594 		return irq;
595 	}
596 
597 	ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
598 				dev_name(dev), data);
599 	if (ret) {
600 		dev_err(dev, "Unabled to register handler of irq %d\n", irq);
601 		return ret;
602 	}
603 
604 	data->clk = devm_clk_get(dev, "sysmmu");
605 	if (!IS_ERR(data->clk)) {
606 		ret = clk_prepare(data->clk);
607 		if (ret) {
608 			dev_err(dev, "Failed to prepare clk\n");
609 			return ret;
610 		}
611 	} else {
612 		data->clk = NULL;
613 	}
614 
615 	data->aclk = devm_clk_get(dev, "aclk");
616 	if (!IS_ERR(data->aclk)) {
617 		ret = clk_prepare(data->aclk);
618 		if (ret) {
619 			dev_err(dev, "Failed to prepare aclk\n");
620 			return ret;
621 		}
622 	} else {
623 		data->aclk = NULL;
624 	}
625 
626 	data->pclk = devm_clk_get(dev, "pclk");
627 	if (!IS_ERR(data->pclk)) {
628 		ret = clk_prepare(data->pclk);
629 		if (ret) {
630 			dev_err(dev, "Failed to prepare pclk\n");
631 			return ret;
632 		}
633 	} else {
634 		data->pclk = NULL;
635 	}
636 
637 	if (!data->clk && (!data->aclk || !data->pclk)) {
638 		dev_err(dev, "Failed to get device clock(s)!\n");
639 		return -ENOSYS;
640 	}
641 
642 	data->clk_master = devm_clk_get(dev, "master");
643 	if (!IS_ERR(data->clk_master)) {
644 		ret = clk_prepare(data->clk_master);
645 		if (ret) {
646 			dev_err(dev, "Failed to prepare master's clk\n");
647 			return ret;
648 		}
649 	} else {
650 		data->clk_master = NULL;
651 	}
652 
653 	data->sysmmu = dev;
654 	spin_lock_init(&data->lock);
655 
656 	platform_set_drvdata(pdev, data);
657 
658 	__sysmmu_get_version(data);
659 	if (PG_ENT_SHIFT < 0) {
660 		if (MMU_MAJ_VER(data->version) < 5)
661 			PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
662 		else
663 			PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
664 	}
665 
666 	pm_runtime_enable(dev);
667 
668 	return 0;
669 }
670 
671 #ifdef CONFIG_PM_SLEEP
672 static int exynos_sysmmu_suspend(struct device *dev)
673 {
674 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
675 
676 	dev_dbg(dev, "suspend\n");
677 	if (is_sysmmu_active(data)) {
678 		__sysmmu_disable_nocount(data);
679 		pm_runtime_put(dev);
680 	}
681 	return 0;
682 }
683 
684 static int exynos_sysmmu_resume(struct device *dev)
685 {
686 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
687 
688 	dev_dbg(dev, "resume\n");
689 	if (is_sysmmu_active(data)) {
690 		pm_runtime_get_sync(dev);
691 		__sysmmu_enable_nocount(data);
692 	}
693 	return 0;
694 }
695 #endif
696 
697 static const struct dev_pm_ops sysmmu_pm_ops = {
698 	SET_LATE_SYSTEM_SLEEP_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume)
699 };
700 
701 static const struct of_device_id sysmmu_of_match[] __initconst = {
702 	{ .compatible	= "samsung,exynos-sysmmu", },
703 	{ },
704 };
705 
706 static struct platform_driver exynos_sysmmu_driver __refdata = {
707 	.probe	= exynos_sysmmu_probe,
708 	.driver	= {
709 		.name		= "exynos-sysmmu",
710 		.of_match_table	= sysmmu_of_match,
711 		.pm		= &sysmmu_pm_ops,
712 	}
713 };
714 
715 static inline void update_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
716 {
717 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
718 				DMA_TO_DEVICE);
719 	*ent = val;
720 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
721 				   DMA_TO_DEVICE);
722 }
723 
724 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
725 {
726 	struct exynos_iommu_domain *domain;
727 	dma_addr_t handle;
728 	int i;
729 
730 	/* Check if correct PTE offsets are initialized */
731 	BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
732 
733 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
734 	if (!domain)
735 		return NULL;
736 
737 	if (type == IOMMU_DOMAIN_DMA) {
738 		if (iommu_get_dma_cookie(&domain->domain) != 0)
739 			goto err_pgtable;
740 	} else if (type != IOMMU_DOMAIN_UNMANAGED) {
741 		goto err_pgtable;
742 	}
743 
744 	domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
745 	if (!domain->pgtable)
746 		goto err_dma_cookie;
747 
748 	domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
749 	if (!domain->lv2entcnt)
750 		goto err_counter;
751 
752 	/* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
753 	for (i = 0; i < NUM_LV1ENTRIES; i += 8) {
754 		domain->pgtable[i + 0] = ZERO_LV2LINK;
755 		domain->pgtable[i + 1] = ZERO_LV2LINK;
756 		domain->pgtable[i + 2] = ZERO_LV2LINK;
757 		domain->pgtable[i + 3] = ZERO_LV2LINK;
758 		domain->pgtable[i + 4] = ZERO_LV2LINK;
759 		domain->pgtable[i + 5] = ZERO_LV2LINK;
760 		domain->pgtable[i + 6] = ZERO_LV2LINK;
761 		domain->pgtable[i + 7] = ZERO_LV2LINK;
762 	}
763 
764 	handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
765 				DMA_TO_DEVICE);
766 	/* For mapping page table entries we rely on dma == phys */
767 	BUG_ON(handle != virt_to_phys(domain->pgtable));
768 
769 	spin_lock_init(&domain->lock);
770 	spin_lock_init(&domain->pgtablelock);
771 	INIT_LIST_HEAD(&domain->clients);
772 
773 	domain->domain.geometry.aperture_start = 0;
774 	domain->domain.geometry.aperture_end   = ~0UL;
775 	domain->domain.geometry.force_aperture = true;
776 
777 	return &domain->domain;
778 
779 err_counter:
780 	free_pages((unsigned long)domain->pgtable, 2);
781 err_dma_cookie:
782 	if (type == IOMMU_DOMAIN_DMA)
783 		iommu_put_dma_cookie(&domain->domain);
784 err_pgtable:
785 	kfree(domain);
786 	return NULL;
787 }
788 
789 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
790 {
791 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
792 	struct sysmmu_drvdata *data, *next;
793 	unsigned long flags;
794 	int i;
795 
796 	WARN_ON(!list_empty(&domain->clients));
797 
798 	spin_lock_irqsave(&domain->lock, flags);
799 
800 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
801 		if (__sysmmu_disable(data))
802 			data->master = NULL;
803 		list_del_init(&data->domain_node);
804 	}
805 
806 	spin_unlock_irqrestore(&domain->lock, flags);
807 
808 	if (iommu_domain->type == IOMMU_DOMAIN_DMA)
809 		iommu_put_dma_cookie(iommu_domain);
810 
811 	dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
812 			 DMA_TO_DEVICE);
813 
814 	for (i = 0; i < NUM_LV1ENTRIES; i++)
815 		if (lv1ent_page(domain->pgtable + i)) {
816 			phys_addr_t base = lv2table_base(domain->pgtable + i);
817 
818 			dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
819 					 DMA_TO_DEVICE);
820 			kmem_cache_free(lv2table_kmem_cache,
821 					phys_to_virt(base));
822 		}
823 
824 	free_pages((unsigned long)domain->pgtable, 2);
825 	free_pages((unsigned long)domain->lv2entcnt, 1);
826 	kfree(domain);
827 }
828 
829 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
830 				    struct device *dev)
831 {
832 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
833 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
834 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
835 	struct sysmmu_drvdata *data, *next;
836 	unsigned long flags;
837 	bool found = false;
838 
839 	if (!has_sysmmu(dev) || owner->domain != iommu_domain)
840 		return;
841 
842 	spin_lock_irqsave(&domain->lock, flags);
843 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
844 		if (data->master == dev) {
845 			if (__sysmmu_disable(data)) {
846 				data->master = NULL;
847 				list_del_init(&data->domain_node);
848 			}
849 			pm_runtime_put(data->sysmmu);
850 			found = true;
851 		}
852 	}
853 	spin_unlock_irqrestore(&domain->lock, flags);
854 
855 	owner->domain = NULL;
856 
857 	if (found)
858 		dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n",
859 					__func__, &pagetable);
860 	else
861 		dev_err(dev, "%s: No IOMMU is attached\n", __func__);
862 }
863 
864 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
865 				   struct device *dev)
866 {
867 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
868 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
869 	struct sysmmu_drvdata *data;
870 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
871 	unsigned long flags;
872 	int ret = -ENODEV;
873 
874 	if (!has_sysmmu(dev))
875 		return -ENODEV;
876 
877 	if (owner->domain)
878 		exynos_iommu_detach_device(owner->domain, dev);
879 
880 	list_for_each_entry(data, &owner->controllers, owner_node) {
881 		pm_runtime_get_sync(data->sysmmu);
882 		ret = __sysmmu_enable(data, pagetable, domain);
883 		if (ret >= 0) {
884 			data->master = dev;
885 
886 			spin_lock_irqsave(&domain->lock, flags);
887 			list_add_tail(&data->domain_node, &domain->clients);
888 			spin_unlock_irqrestore(&domain->lock, flags);
889 		}
890 	}
891 
892 	if (ret < 0) {
893 		dev_err(dev, "%s: Failed to attach IOMMU with pgtable %pa\n",
894 					__func__, &pagetable);
895 		return ret;
896 	}
897 
898 	owner->domain = iommu_domain;
899 	dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa %s\n",
900 		__func__, &pagetable, (ret == 0) ? "" : ", again");
901 
902 	return ret;
903 }
904 
905 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
906 		sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
907 {
908 	if (lv1ent_section(sent)) {
909 		WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
910 		return ERR_PTR(-EADDRINUSE);
911 	}
912 
913 	if (lv1ent_fault(sent)) {
914 		sysmmu_pte_t *pent;
915 		bool need_flush_flpd_cache = lv1ent_zero(sent);
916 
917 		pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
918 		BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
919 		if (!pent)
920 			return ERR_PTR(-ENOMEM);
921 
922 		update_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
923 		kmemleak_ignore(pent);
924 		*pgcounter = NUM_LV2ENTRIES;
925 		dma_map_single(dma_dev, pent, LV2TABLE_SIZE, DMA_TO_DEVICE);
926 
927 		/*
928 		 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
929 		 * FLPD cache may cache the address of zero_l2_table. This
930 		 * function replaces the zero_l2_table with new L2 page table
931 		 * to write valid mappings.
932 		 * Accessing the valid area may cause page fault since FLPD
933 		 * cache may still cache zero_l2_table for the valid area
934 		 * instead of new L2 page table that has the mapping
935 		 * information of the valid area.
936 		 * Thus any replacement of zero_l2_table with other valid L2
937 		 * page table must involve FLPD cache invalidation for System
938 		 * MMU v3.3.
939 		 * FLPD cache invalidation is performed with TLB invalidation
940 		 * by VPN without blocking. It is safe to invalidate TLB without
941 		 * blocking because the target address of TLB invalidation is
942 		 * not currently mapped.
943 		 */
944 		if (need_flush_flpd_cache) {
945 			struct sysmmu_drvdata *data;
946 
947 			spin_lock(&domain->lock);
948 			list_for_each_entry(data, &domain->clients, domain_node)
949 				sysmmu_tlb_invalidate_flpdcache(data, iova);
950 			spin_unlock(&domain->lock);
951 		}
952 	}
953 
954 	return page_entry(sent, iova);
955 }
956 
957 static int lv1set_section(struct exynos_iommu_domain *domain,
958 			  sysmmu_pte_t *sent, sysmmu_iova_t iova,
959 			  phys_addr_t paddr, short *pgcnt)
960 {
961 	if (lv1ent_section(sent)) {
962 		WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
963 			iova);
964 		return -EADDRINUSE;
965 	}
966 
967 	if (lv1ent_page(sent)) {
968 		if (*pgcnt != NUM_LV2ENTRIES) {
969 			WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
970 				iova);
971 			return -EADDRINUSE;
972 		}
973 
974 		kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
975 		*pgcnt = 0;
976 	}
977 
978 	update_pte(sent, mk_lv1ent_sect(paddr));
979 
980 	spin_lock(&domain->lock);
981 	if (lv1ent_page_zero(sent)) {
982 		struct sysmmu_drvdata *data;
983 		/*
984 		 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
985 		 * entry by speculative prefetch of SLPD which has no mapping.
986 		 */
987 		list_for_each_entry(data, &domain->clients, domain_node)
988 			sysmmu_tlb_invalidate_flpdcache(data, iova);
989 	}
990 	spin_unlock(&domain->lock);
991 
992 	return 0;
993 }
994 
995 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
996 								short *pgcnt)
997 {
998 	if (size == SPAGE_SIZE) {
999 		if (WARN_ON(!lv2ent_fault(pent)))
1000 			return -EADDRINUSE;
1001 
1002 		update_pte(pent, mk_lv2ent_spage(paddr));
1003 		*pgcnt -= 1;
1004 	} else { /* size == LPAGE_SIZE */
1005 		int i;
1006 		dma_addr_t pent_base = virt_to_phys(pent);
1007 
1008 		dma_sync_single_for_cpu(dma_dev, pent_base,
1009 					sizeof(*pent) * SPAGES_PER_LPAGE,
1010 					DMA_TO_DEVICE);
1011 		for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1012 			if (WARN_ON(!lv2ent_fault(pent))) {
1013 				if (i > 0)
1014 					memset(pent - i, 0, sizeof(*pent) * i);
1015 				return -EADDRINUSE;
1016 			}
1017 
1018 			*pent = mk_lv2ent_lpage(paddr);
1019 		}
1020 		dma_sync_single_for_device(dma_dev, pent_base,
1021 					   sizeof(*pent) * SPAGES_PER_LPAGE,
1022 					   DMA_TO_DEVICE);
1023 		*pgcnt -= SPAGES_PER_LPAGE;
1024 	}
1025 
1026 	return 0;
1027 }
1028 
1029 /*
1030  * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1031  *
1032  * System MMU v3.x has advanced logic to improve address translation
1033  * performance with caching more page table entries by a page table walk.
1034  * However, the logic has a bug that while caching faulty page table entries,
1035  * System MMU reports page fault if the cached fault entry is hit even though
1036  * the fault entry is updated to a valid entry after the entry is cached.
1037  * To prevent caching faulty page table entries which may be updated to valid
1038  * entries later, the virtual memory manager should care about the workaround
1039  * for the problem. The following describes the workaround.
1040  *
1041  * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1042  * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1043  *
1044  * Precisely, any start address of I/O virtual region must be aligned with
1045  * the following sizes for System MMU v3.1 and v3.2.
1046  * System MMU v3.1: 128KiB
1047  * System MMU v3.2: 256KiB
1048  *
1049  * Because System MMU v3.3 caches page table entries more aggressively, it needs
1050  * more workarounds.
1051  * - Any two consecutive I/O virtual regions must have a hole of size larger
1052  *   than or equal to 128KiB.
1053  * - Start address of an I/O virtual region must be aligned by 128KiB.
1054  */
1055 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1056 			    unsigned long l_iova, phys_addr_t paddr, size_t size,
1057 			    int prot)
1058 {
1059 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1060 	sysmmu_pte_t *entry;
1061 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1062 	unsigned long flags;
1063 	int ret = -ENOMEM;
1064 
1065 	BUG_ON(domain->pgtable == NULL);
1066 
1067 	spin_lock_irqsave(&domain->pgtablelock, flags);
1068 
1069 	entry = section_entry(domain->pgtable, iova);
1070 
1071 	if (size == SECT_SIZE) {
1072 		ret = lv1set_section(domain, entry, iova, paddr,
1073 				     &domain->lv2entcnt[lv1ent_offset(iova)]);
1074 	} else {
1075 		sysmmu_pte_t *pent;
1076 
1077 		pent = alloc_lv2entry(domain, entry, iova,
1078 				      &domain->lv2entcnt[lv1ent_offset(iova)]);
1079 
1080 		if (IS_ERR(pent))
1081 			ret = PTR_ERR(pent);
1082 		else
1083 			ret = lv2set_page(pent, paddr, size,
1084 				       &domain->lv2entcnt[lv1ent_offset(iova)]);
1085 	}
1086 
1087 	if (ret)
1088 		pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1089 			__func__, ret, size, iova);
1090 
1091 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1092 
1093 	return ret;
1094 }
1095 
1096 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1097 					      sysmmu_iova_t iova, size_t size)
1098 {
1099 	struct sysmmu_drvdata *data;
1100 	unsigned long flags;
1101 
1102 	spin_lock_irqsave(&domain->lock, flags);
1103 
1104 	list_for_each_entry(data, &domain->clients, domain_node)
1105 		sysmmu_tlb_invalidate_entry(data, iova, size);
1106 
1107 	spin_unlock_irqrestore(&domain->lock, flags);
1108 }
1109 
1110 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1111 				 unsigned long l_iova, size_t size)
1112 {
1113 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1114 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1115 	sysmmu_pte_t *ent;
1116 	size_t err_pgsize;
1117 	unsigned long flags;
1118 
1119 	BUG_ON(domain->pgtable == NULL);
1120 
1121 	spin_lock_irqsave(&domain->pgtablelock, flags);
1122 
1123 	ent = section_entry(domain->pgtable, iova);
1124 
1125 	if (lv1ent_section(ent)) {
1126 		if (WARN_ON(size < SECT_SIZE)) {
1127 			err_pgsize = SECT_SIZE;
1128 			goto err;
1129 		}
1130 
1131 		/* workaround for h/w bug in System MMU v3.3 */
1132 		update_pte(ent, ZERO_LV2LINK);
1133 		size = SECT_SIZE;
1134 		goto done;
1135 	}
1136 
1137 	if (unlikely(lv1ent_fault(ent))) {
1138 		if (size > SECT_SIZE)
1139 			size = SECT_SIZE;
1140 		goto done;
1141 	}
1142 
1143 	/* lv1ent_page(sent) == true here */
1144 
1145 	ent = page_entry(ent, iova);
1146 
1147 	if (unlikely(lv2ent_fault(ent))) {
1148 		size = SPAGE_SIZE;
1149 		goto done;
1150 	}
1151 
1152 	if (lv2ent_small(ent)) {
1153 		update_pte(ent, 0);
1154 		size = SPAGE_SIZE;
1155 		domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1156 		goto done;
1157 	}
1158 
1159 	/* lv1ent_large(ent) == true here */
1160 	if (WARN_ON(size < LPAGE_SIZE)) {
1161 		err_pgsize = LPAGE_SIZE;
1162 		goto err;
1163 	}
1164 
1165 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1166 				sizeof(*ent) * SPAGES_PER_LPAGE,
1167 				DMA_TO_DEVICE);
1168 	memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1169 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1170 				   sizeof(*ent) * SPAGES_PER_LPAGE,
1171 				   DMA_TO_DEVICE);
1172 	size = LPAGE_SIZE;
1173 	domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1174 done:
1175 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1176 
1177 	exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1178 
1179 	return size;
1180 err:
1181 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1182 
1183 	pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1184 		__func__, size, iova, err_pgsize);
1185 
1186 	return 0;
1187 }
1188 
1189 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1190 					  dma_addr_t iova)
1191 {
1192 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1193 	sysmmu_pte_t *entry;
1194 	unsigned long flags;
1195 	phys_addr_t phys = 0;
1196 
1197 	spin_lock_irqsave(&domain->pgtablelock, flags);
1198 
1199 	entry = section_entry(domain->pgtable, iova);
1200 
1201 	if (lv1ent_section(entry)) {
1202 		phys = section_phys(entry) + section_offs(iova);
1203 	} else if (lv1ent_page(entry)) {
1204 		entry = page_entry(entry, iova);
1205 
1206 		if (lv2ent_large(entry))
1207 			phys = lpage_phys(entry) + lpage_offs(iova);
1208 		else if (lv2ent_small(entry))
1209 			phys = spage_phys(entry) + spage_offs(iova);
1210 	}
1211 
1212 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1213 
1214 	return phys;
1215 }
1216 
1217 static struct iommu_group *get_device_iommu_group(struct device *dev)
1218 {
1219 	struct iommu_group *group;
1220 
1221 	group = iommu_group_get(dev);
1222 	if (!group)
1223 		group = iommu_group_alloc();
1224 
1225 	return group;
1226 }
1227 
1228 static int exynos_iommu_add_device(struct device *dev)
1229 {
1230 	struct iommu_group *group;
1231 
1232 	if (!has_sysmmu(dev))
1233 		return -ENODEV;
1234 
1235 	group = iommu_group_get_for_dev(dev);
1236 
1237 	if (IS_ERR(group))
1238 		return PTR_ERR(group);
1239 
1240 	iommu_group_put(group);
1241 
1242 	return 0;
1243 }
1244 
1245 static void exynos_iommu_remove_device(struct device *dev)
1246 {
1247 	if (!has_sysmmu(dev))
1248 		return;
1249 
1250 	iommu_group_remove_device(dev);
1251 }
1252 
1253 static int exynos_iommu_of_xlate(struct device *dev,
1254 				 struct of_phandle_args *spec)
1255 {
1256 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1257 	struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1258 	struct sysmmu_drvdata *data;
1259 
1260 	if (!sysmmu)
1261 		return -ENODEV;
1262 
1263 	data = platform_get_drvdata(sysmmu);
1264 	if (!data)
1265 		return -ENODEV;
1266 
1267 	if (!owner) {
1268 		owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1269 		if (!owner)
1270 			return -ENOMEM;
1271 
1272 		INIT_LIST_HEAD(&owner->controllers);
1273 		dev->archdata.iommu = owner;
1274 	}
1275 
1276 	list_add_tail(&data->owner_node, &owner->controllers);
1277 	return 0;
1278 }
1279 
1280 static struct iommu_ops exynos_iommu_ops = {
1281 	.domain_alloc = exynos_iommu_domain_alloc,
1282 	.domain_free = exynos_iommu_domain_free,
1283 	.attach_dev = exynos_iommu_attach_device,
1284 	.detach_dev = exynos_iommu_detach_device,
1285 	.map = exynos_iommu_map,
1286 	.unmap = exynos_iommu_unmap,
1287 	.map_sg = default_iommu_map_sg,
1288 	.iova_to_phys = exynos_iommu_iova_to_phys,
1289 	.device_group = get_device_iommu_group,
1290 	.add_device = exynos_iommu_add_device,
1291 	.remove_device = exynos_iommu_remove_device,
1292 	.pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1293 	.of_xlate = exynos_iommu_of_xlate,
1294 };
1295 
1296 static bool init_done;
1297 
1298 static int __init exynos_iommu_init(void)
1299 {
1300 	int ret;
1301 
1302 	lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1303 				LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1304 	if (!lv2table_kmem_cache) {
1305 		pr_err("%s: Failed to create kmem cache\n", __func__);
1306 		return -ENOMEM;
1307 	}
1308 
1309 	ret = platform_driver_register(&exynos_sysmmu_driver);
1310 	if (ret) {
1311 		pr_err("%s: Failed to register driver\n", __func__);
1312 		goto err_reg_driver;
1313 	}
1314 
1315 	zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1316 	if (zero_lv2_table == NULL) {
1317 		pr_err("%s: Failed to allocate zero level2 page table\n",
1318 			__func__);
1319 		ret = -ENOMEM;
1320 		goto err_zero_lv2;
1321 	}
1322 
1323 	ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1324 	if (ret) {
1325 		pr_err("%s: Failed to register exynos-iommu driver.\n",
1326 								__func__);
1327 		goto err_set_iommu;
1328 	}
1329 
1330 	init_done = true;
1331 
1332 	return 0;
1333 err_set_iommu:
1334 	kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1335 err_zero_lv2:
1336 	platform_driver_unregister(&exynos_sysmmu_driver);
1337 err_reg_driver:
1338 	kmem_cache_destroy(lv2table_kmem_cache);
1339 	return ret;
1340 }
1341 
1342 static int __init exynos_iommu_of_setup(struct device_node *np)
1343 {
1344 	struct platform_device *pdev;
1345 
1346 	if (!init_done)
1347 		exynos_iommu_init();
1348 
1349 	pdev = of_platform_device_create(np, NULL, platform_bus_type.dev_root);
1350 	if (IS_ERR(pdev))
1351 		return PTR_ERR(pdev);
1352 
1353 	/*
1354 	 * use the first registered sysmmu device for performing
1355 	 * dma mapping operations on iommu page tables (cpu cache flush)
1356 	 */
1357 	if (!dma_dev)
1358 		dma_dev = &pdev->dev;
1359 
1360 	of_iommu_set_ops(np, &exynos_iommu_ops);
1361 	return 0;
1362 }
1363 
1364 IOMMU_OF_DECLARE(exynos_iommu_of, "samsung,exynos-sysmmu",
1365 		 exynos_iommu_of_setup);
1366