xref: /openbmc/linux/drivers/iommu/exynos-iommu.c (revision 8730046c)
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 #ifdef CONFIG_BIG_ENDIAN
58 #warning "revisit driver if we can enable big-endian ptes"
59 #endif
60 
61 /*
62  * v1.x - v3.x SYSMMU supports 32bit physical and 32bit virtual address spaces
63  * v5.0 introduced support for 36bit physical address space by shifting
64  * all page entry values by 4 bits.
65  * All SYSMMU controllers in the system support the address spaces of the same
66  * size, so PG_ENT_SHIFT can be initialized on first SYSMMU probe to proper
67  * value (0 or 4).
68  */
69 static short PG_ENT_SHIFT = -1;
70 #define SYSMMU_PG_ENT_SHIFT 0
71 #define SYSMMU_V5_PG_ENT_SHIFT 4
72 
73 static const sysmmu_pte_t *LV1_PROT;
74 static const sysmmu_pte_t SYSMMU_LV1_PROT[] = {
75 	((0 << 15) | (0 << 10)), /* no access */
76 	((1 << 15) | (1 << 10)), /* IOMMU_READ only */
77 	((0 << 15) | (1 << 10)), /* IOMMU_WRITE not supported, use read/write */
78 	((0 << 15) | (1 << 10)), /* IOMMU_READ | IOMMU_WRITE */
79 };
80 static const sysmmu_pte_t SYSMMU_V5_LV1_PROT[] = {
81 	(0 << 4), /* no access */
82 	(1 << 4), /* IOMMU_READ only */
83 	(2 << 4), /* IOMMU_WRITE only */
84 	(3 << 4), /* IOMMU_READ | IOMMU_WRITE */
85 };
86 
87 static const sysmmu_pte_t *LV2_PROT;
88 static const sysmmu_pte_t SYSMMU_LV2_PROT[] = {
89 	((0 << 9) | (0 << 4)), /* no access */
90 	((1 << 9) | (1 << 4)), /* IOMMU_READ only */
91 	((0 << 9) | (1 << 4)), /* IOMMU_WRITE not supported, use read/write */
92 	((0 << 9) | (1 << 4)), /* IOMMU_READ | IOMMU_WRITE */
93 };
94 static const sysmmu_pte_t SYSMMU_V5_LV2_PROT[] = {
95 	(0 << 2), /* no access */
96 	(1 << 2), /* IOMMU_READ only */
97 	(2 << 2), /* IOMMU_WRITE only */
98 	(3 << 2), /* IOMMU_READ | IOMMU_WRITE */
99 };
100 
101 #define SYSMMU_SUPPORTED_PROT_BITS (IOMMU_READ | IOMMU_WRITE)
102 
103 #define sect_to_phys(ent) (((phys_addr_t) ent) << PG_ENT_SHIFT)
104 #define section_phys(sent) (sect_to_phys(*(sent)) & SECT_MASK)
105 #define section_offs(iova) (iova & (SECT_SIZE - 1))
106 #define lpage_phys(pent) (sect_to_phys(*(pent)) & LPAGE_MASK)
107 #define lpage_offs(iova) (iova & (LPAGE_SIZE - 1))
108 #define spage_phys(pent) (sect_to_phys(*(pent)) & SPAGE_MASK)
109 #define spage_offs(iova) (iova & (SPAGE_SIZE - 1))
110 
111 #define NUM_LV1ENTRIES 4096
112 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
113 
114 static u32 lv1ent_offset(sysmmu_iova_t iova)
115 {
116 	return iova >> SECT_ORDER;
117 }
118 
119 static u32 lv2ent_offset(sysmmu_iova_t iova)
120 {
121 	return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
122 }
123 
124 #define LV1TABLE_SIZE (NUM_LV1ENTRIES * sizeof(sysmmu_pte_t))
125 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
126 
127 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
128 #define lv2table_base(sent) (sect_to_phys(*(sent) & 0xFFFFFFC0))
129 
130 #define mk_lv1ent_sect(pa, prot) ((pa >> PG_ENT_SHIFT) | LV1_PROT[prot] | 2)
131 #define mk_lv1ent_page(pa) ((pa >> PG_ENT_SHIFT) | 1)
132 #define mk_lv2ent_lpage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 1)
133 #define mk_lv2ent_spage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 2)
134 
135 #define CTRL_ENABLE	0x5
136 #define CTRL_BLOCK	0x7
137 #define CTRL_DISABLE	0x0
138 
139 #define CFG_LRU		0x1
140 #define CFG_EAP		(1 << 2)
141 #define CFG_QOS(n)	((n & 0xF) << 7)
142 #define CFG_ACGEN	(1 << 24) /* System MMU 3.3 only */
143 #define CFG_SYSSEL	(1 << 22) /* System MMU 3.2 only */
144 #define CFG_FLPDCACHE	(1 << 20) /* System MMU 3.2+ only */
145 
146 /* common registers */
147 #define REG_MMU_CTRL		0x000
148 #define REG_MMU_CFG		0x004
149 #define REG_MMU_STATUS		0x008
150 #define REG_MMU_VERSION		0x034
151 
152 #define MMU_MAJ_VER(val)	((val) >> 7)
153 #define MMU_MIN_VER(val)	((val) & 0x7F)
154 #define MMU_RAW_VER(reg)	(((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
155 
156 #define MAKE_MMU_VER(maj, min)	((((maj) & 0xF) << 7) | ((min) & 0x7F))
157 
158 /* v1.x - v3.x registers */
159 #define REG_MMU_FLUSH		0x00C
160 #define REG_MMU_FLUSH_ENTRY	0x010
161 #define REG_PT_BASE_ADDR	0x014
162 #define REG_INT_STATUS		0x018
163 #define REG_INT_CLEAR		0x01C
164 
165 #define REG_PAGE_FAULT_ADDR	0x024
166 #define REG_AW_FAULT_ADDR	0x028
167 #define REG_AR_FAULT_ADDR	0x02C
168 #define REG_DEFAULT_SLAVE_ADDR	0x030
169 
170 /* v5.x registers */
171 #define REG_V5_PT_BASE_PFN	0x00C
172 #define REG_V5_MMU_FLUSH_ALL	0x010
173 #define REG_V5_MMU_FLUSH_ENTRY	0x014
174 #define REG_V5_INT_STATUS	0x060
175 #define REG_V5_INT_CLEAR	0x064
176 #define REG_V5_FAULT_AR_VA	0x070
177 #define REG_V5_FAULT_AW_VA	0x080
178 
179 #define has_sysmmu(dev)		(dev->archdata.iommu != NULL)
180 
181 static struct device *dma_dev;
182 static struct kmem_cache *lv2table_kmem_cache;
183 static sysmmu_pte_t *zero_lv2_table;
184 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
185 
186 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
187 {
188 	return pgtable + lv1ent_offset(iova);
189 }
190 
191 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
192 {
193 	return (sysmmu_pte_t *)phys_to_virt(
194 				lv2table_base(sent)) + lv2ent_offset(iova);
195 }
196 
197 /*
198  * IOMMU fault information register
199  */
200 struct sysmmu_fault_info {
201 	unsigned int bit;	/* bit number in STATUS register */
202 	unsigned short addr_reg; /* register to read VA fault address */
203 	const char *name;	/* human readable fault name */
204 	unsigned int type;	/* fault type for report_iommu_fault */
205 };
206 
207 static const struct sysmmu_fault_info sysmmu_faults[] = {
208 	{ 0, REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
209 	{ 1, REG_AR_FAULT_ADDR, "AR MULTI-HIT", IOMMU_FAULT_READ },
210 	{ 2, REG_AW_FAULT_ADDR, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
211 	{ 3, REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
212 	{ 4, REG_AR_FAULT_ADDR, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
213 	{ 5, REG_AR_FAULT_ADDR, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
214 	{ 6, REG_AW_FAULT_ADDR, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
215 	{ 7, REG_AW_FAULT_ADDR, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
216 };
217 
218 static const struct sysmmu_fault_info sysmmu_v5_faults[] = {
219 	{ 0, REG_V5_FAULT_AR_VA, "AR PTW", IOMMU_FAULT_READ },
220 	{ 1, REG_V5_FAULT_AR_VA, "AR PAGE", IOMMU_FAULT_READ },
221 	{ 2, REG_V5_FAULT_AR_VA, "AR MULTI-HIT", IOMMU_FAULT_READ },
222 	{ 3, REG_V5_FAULT_AR_VA, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
223 	{ 4, REG_V5_FAULT_AR_VA, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
224 	{ 16, REG_V5_FAULT_AW_VA, "AW PTW", IOMMU_FAULT_WRITE },
225 	{ 17, REG_V5_FAULT_AW_VA, "AW PAGE", IOMMU_FAULT_WRITE },
226 	{ 18, REG_V5_FAULT_AW_VA, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
227 	{ 19, REG_V5_FAULT_AW_VA, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
228 	{ 20, REG_V5_FAULT_AW_VA, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
229 };
230 
231 /*
232  * This structure is attached to dev.archdata.iommu of the master device
233  * on device add, contains a list of SYSMMU controllers defined by device tree,
234  * which are bound to given master device. It is usually referenced by 'owner'
235  * pointer.
236 */
237 struct exynos_iommu_owner {
238 	struct list_head controllers;	/* list of sysmmu_drvdata.owner_node */
239 	struct iommu_domain *domain;	/* domain this device is attached */
240 	struct mutex rpm_lock;		/* for runtime pm of all sysmmus */
241 };
242 
243 /*
244  * This structure exynos specific generalization of struct iommu_domain.
245  * It contains list of SYSMMU controllers from all master devices, which has
246  * been attached to this domain and page tables of IO address space defined by
247  * it. It is usually referenced by 'domain' pointer.
248  */
249 struct exynos_iommu_domain {
250 	struct list_head clients; /* list of sysmmu_drvdata.domain_node */
251 	sysmmu_pte_t *pgtable;	/* lv1 page table, 16KB */
252 	short *lv2entcnt;	/* free lv2 entry counter for each section */
253 	spinlock_t lock;	/* lock for modyfying list of clients */
254 	spinlock_t pgtablelock;	/* lock for modifying page table @ pgtable */
255 	struct iommu_domain domain; /* generic domain data structure */
256 };
257 
258 /*
259  * This structure hold all data of a single SYSMMU controller, this includes
260  * hw resources like registers and clocks, pointers and list nodes to connect
261  * it to all other structures, internal state and parameters read from device
262  * tree. It is usually referenced by 'data' pointer.
263  */
264 struct sysmmu_drvdata {
265 	struct device *sysmmu;		/* SYSMMU controller device */
266 	struct device *master;		/* master device (owner) */
267 	void __iomem *sfrbase;		/* our registers */
268 	struct clk *clk;		/* SYSMMU's clock */
269 	struct clk *aclk;		/* SYSMMU's aclk clock */
270 	struct clk *pclk;		/* SYSMMU's pclk clock */
271 	struct clk *clk_master;		/* master's device clock */
272 	spinlock_t lock;		/* lock for modyfying state */
273 	bool active;			/* current status */
274 	struct exynos_iommu_domain *domain; /* domain we belong to */
275 	struct list_head domain_node;	/* node for domain clients list */
276 	struct list_head owner_node;	/* node for owner controllers list */
277 	phys_addr_t pgtable;		/* assigned page table structure */
278 	unsigned int version;		/* our version */
279 };
280 
281 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
282 {
283 	return container_of(dom, struct exynos_iommu_domain, domain);
284 }
285 
286 static void sysmmu_unblock(struct sysmmu_drvdata *data)
287 {
288 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
289 }
290 
291 static bool sysmmu_block(struct sysmmu_drvdata *data)
292 {
293 	int i = 120;
294 
295 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
296 	while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
297 		--i;
298 
299 	if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
300 		sysmmu_unblock(data);
301 		return false;
302 	}
303 
304 	return true;
305 }
306 
307 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
308 {
309 	if (MMU_MAJ_VER(data->version) < 5)
310 		writel(0x1, data->sfrbase + REG_MMU_FLUSH);
311 	else
312 		writel(0x1, data->sfrbase + REG_V5_MMU_FLUSH_ALL);
313 }
314 
315 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
316 				sysmmu_iova_t iova, unsigned int num_inv)
317 {
318 	unsigned int i;
319 
320 	for (i = 0; i < num_inv; i++) {
321 		if (MMU_MAJ_VER(data->version) < 5)
322 			writel((iova & SPAGE_MASK) | 1,
323 				     data->sfrbase + REG_MMU_FLUSH_ENTRY);
324 		else
325 			writel((iova & SPAGE_MASK) | 1,
326 				     data->sfrbase + REG_V5_MMU_FLUSH_ENTRY);
327 		iova += SPAGE_SIZE;
328 	}
329 }
330 
331 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
332 {
333 	if (MMU_MAJ_VER(data->version) < 5)
334 		writel(pgd, data->sfrbase + REG_PT_BASE_ADDR);
335 	else
336 		writel(pgd >> PAGE_SHIFT,
337 			     data->sfrbase + REG_V5_PT_BASE_PFN);
338 
339 	__sysmmu_tlb_invalidate(data);
340 }
341 
342 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
343 {
344 	BUG_ON(clk_prepare_enable(data->clk_master));
345 	BUG_ON(clk_prepare_enable(data->clk));
346 	BUG_ON(clk_prepare_enable(data->pclk));
347 	BUG_ON(clk_prepare_enable(data->aclk));
348 }
349 
350 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
351 {
352 	clk_disable_unprepare(data->aclk);
353 	clk_disable_unprepare(data->pclk);
354 	clk_disable_unprepare(data->clk);
355 	clk_disable_unprepare(data->clk_master);
356 }
357 
358 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
359 {
360 	u32 ver;
361 
362 	__sysmmu_enable_clocks(data);
363 
364 	ver = readl(data->sfrbase + REG_MMU_VERSION);
365 
366 	/* controllers on some SoCs don't report proper version */
367 	if (ver == 0x80000001u)
368 		data->version = MAKE_MMU_VER(1, 0);
369 	else
370 		data->version = MMU_RAW_VER(ver);
371 
372 	dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
373 		MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
374 
375 	__sysmmu_disable_clocks(data);
376 }
377 
378 static void show_fault_information(struct sysmmu_drvdata *data,
379 				   const struct sysmmu_fault_info *finfo,
380 				   sysmmu_iova_t fault_addr)
381 {
382 	sysmmu_pte_t *ent;
383 
384 	dev_err(data->sysmmu, "%s FAULT occurred at %#x (page table base: %pa)\n",
385 		finfo->name, fault_addr, &data->pgtable);
386 	ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
387 	dev_err(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
388 	if (lv1ent_page(ent)) {
389 		ent = page_entry(ent, fault_addr);
390 		dev_err(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
391 	}
392 }
393 
394 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
395 {
396 	/* SYSMMU is in blocked state when interrupt occurred. */
397 	struct sysmmu_drvdata *data = dev_id;
398 	const struct sysmmu_fault_info *finfo;
399 	unsigned int i, n, itype;
400 	sysmmu_iova_t fault_addr = -1;
401 	unsigned short reg_status, reg_clear;
402 	int ret = -ENOSYS;
403 
404 	WARN_ON(!data->active);
405 
406 	if (MMU_MAJ_VER(data->version) < 5) {
407 		reg_status = REG_INT_STATUS;
408 		reg_clear = REG_INT_CLEAR;
409 		finfo = sysmmu_faults;
410 		n = ARRAY_SIZE(sysmmu_faults);
411 	} else {
412 		reg_status = REG_V5_INT_STATUS;
413 		reg_clear = REG_V5_INT_CLEAR;
414 		finfo = sysmmu_v5_faults;
415 		n = ARRAY_SIZE(sysmmu_v5_faults);
416 	}
417 
418 	spin_lock(&data->lock);
419 
420 	clk_enable(data->clk_master);
421 
422 	itype = __ffs(readl(data->sfrbase + reg_status));
423 	for (i = 0; i < n; i++, finfo++)
424 		if (finfo->bit == itype)
425 			break;
426 	/* unknown/unsupported fault */
427 	BUG_ON(i == n);
428 
429 	/* print debug message */
430 	fault_addr = readl(data->sfrbase + finfo->addr_reg);
431 	show_fault_information(data, finfo, fault_addr);
432 
433 	if (data->domain)
434 		ret = report_iommu_fault(&data->domain->domain,
435 					data->master, fault_addr, finfo->type);
436 	/* fault is not recovered by fault handler */
437 	BUG_ON(ret != 0);
438 
439 	writel(1 << itype, data->sfrbase + reg_clear);
440 
441 	sysmmu_unblock(data);
442 
443 	clk_disable(data->clk_master);
444 
445 	spin_unlock(&data->lock);
446 
447 	return IRQ_HANDLED;
448 }
449 
450 static void __sysmmu_disable(struct sysmmu_drvdata *data)
451 {
452 	unsigned long flags;
453 
454 	clk_enable(data->clk_master);
455 
456 	spin_lock_irqsave(&data->lock, flags);
457 	writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
458 	writel(0, data->sfrbase + REG_MMU_CFG);
459 	data->active = false;
460 	spin_unlock_irqrestore(&data->lock, flags);
461 
462 	__sysmmu_disable_clocks(data);
463 }
464 
465 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
466 {
467 	unsigned int cfg;
468 
469 	if (data->version <= MAKE_MMU_VER(3, 1))
470 		cfg = CFG_LRU | CFG_QOS(15);
471 	else if (data->version <= MAKE_MMU_VER(3, 2))
472 		cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
473 	else
474 		cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
475 
476 	cfg |= CFG_EAP; /* enable access protection bits check */
477 
478 	writel(cfg, data->sfrbase + REG_MMU_CFG);
479 }
480 
481 static void __sysmmu_enable(struct sysmmu_drvdata *data)
482 {
483 	unsigned long flags;
484 
485 	__sysmmu_enable_clocks(data);
486 
487 	spin_lock_irqsave(&data->lock, flags);
488 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
489 	__sysmmu_init_config(data);
490 	__sysmmu_set_ptbase(data, data->pgtable);
491 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
492 	data->active = true;
493 	spin_unlock_irqrestore(&data->lock, flags);
494 
495 	/*
496 	 * SYSMMU driver keeps master's clock enabled only for the short
497 	 * time, while accessing the registers. For performing address
498 	 * translation during DMA transaction it relies on the client
499 	 * driver to enable it.
500 	 */
501 	clk_disable(data->clk_master);
502 }
503 
504 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
505 					    sysmmu_iova_t iova)
506 {
507 	unsigned long flags;
508 
509 	spin_lock_irqsave(&data->lock, flags);
510 	if (data->active && data->version >= MAKE_MMU_VER(3, 3)) {
511 		clk_enable(data->clk_master);
512 		__sysmmu_tlb_invalidate_entry(data, iova, 1);
513 		clk_disable(data->clk_master);
514 	}
515 	spin_unlock_irqrestore(&data->lock, flags);
516 }
517 
518 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
519 					sysmmu_iova_t iova, size_t size)
520 {
521 	unsigned long flags;
522 
523 	spin_lock_irqsave(&data->lock, flags);
524 	if (data->active) {
525 		unsigned int num_inv = 1;
526 
527 		clk_enable(data->clk_master);
528 
529 		/*
530 		 * L2TLB invalidation required
531 		 * 4KB page: 1 invalidation
532 		 * 64KB page: 16 invalidations
533 		 * 1MB page: 64 invalidations
534 		 * because it is set-associative TLB
535 		 * with 8-way and 64 sets.
536 		 * 1MB page can be cached in one of all sets.
537 		 * 64KB page can be one of 16 consecutive sets.
538 		 */
539 		if (MMU_MAJ_VER(data->version) == 2)
540 			num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
541 
542 		if (sysmmu_block(data)) {
543 			__sysmmu_tlb_invalidate_entry(data, iova, num_inv);
544 			sysmmu_unblock(data);
545 		}
546 		clk_disable(data->clk_master);
547 	}
548 	spin_unlock_irqrestore(&data->lock, flags);
549 }
550 
551 static struct iommu_ops exynos_iommu_ops;
552 
553 static int __init exynos_sysmmu_probe(struct platform_device *pdev)
554 {
555 	int irq, ret;
556 	struct device *dev = &pdev->dev;
557 	struct sysmmu_drvdata *data;
558 	struct resource *res;
559 
560 	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
561 	if (!data)
562 		return -ENOMEM;
563 
564 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
565 	data->sfrbase = devm_ioremap_resource(dev, res);
566 	if (IS_ERR(data->sfrbase))
567 		return PTR_ERR(data->sfrbase);
568 
569 	irq = platform_get_irq(pdev, 0);
570 	if (irq <= 0) {
571 		dev_err(dev, "Unable to find IRQ resource\n");
572 		return irq;
573 	}
574 
575 	ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
576 				dev_name(dev), data);
577 	if (ret) {
578 		dev_err(dev, "Unabled to register handler of irq %d\n", irq);
579 		return ret;
580 	}
581 
582 	data->clk = devm_clk_get(dev, "sysmmu");
583 	if (PTR_ERR(data->clk) == -ENOENT)
584 		data->clk = NULL;
585 	else if (IS_ERR(data->clk))
586 		return PTR_ERR(data->clk);
587 
588 	data->aclk = devm_clk_get(dev, "aclk");
589 	if (PTR_ERR(data->aclk) == -ENOENT)
590 		data->aclk = NULL;
591 	else if (IS_ERR(data->aclk))
592 		return PTR_ERR(data->aclk);
593 
594 	data->pclk = devm_clk_get(dev, "pclk");
595 	if (PTR_ERR(data->pclk) == -ENOENT)
596 		data->pclk = NULL;
597 	else if (IS_ERR(data->pclk))
598 		return PTR_ERR(data->pclk);
599 
600 	if (!data->clk && (!data->aclk || !data->pclk)) {
601 		dev_err(dev, "Failed to get device clock(s)!\n");
602 		return -ENOSYS;
603 	}
604 
605 	data->clk_master = devm_clk_get(dev, "master");
606 	if (PTR_ERR(data->clk_master) == -ENOENT)
607 		data->clk_master = NULL;
608 	else if (IS_ERR(data->clk_master))
609 		return PTR_ERR(data->clk_master);
610 
611 	data->sysmmu = dev;
612 	spin_lock_init(&data->lock);
613 
614 	platform_set_drvdata(pdev, data);
615 
616 	__sysmmu_get_version(data);
617 	if (PG_ENT_SHIFT < 0) {
618 		if (MMU_MAJ_VER(data->version) < 5) {
619 			PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
620 			LV1_PROT = SYSMMU_LV1_PROT;
621 			LV2_PROT = SYSMMU_LV2_PROT;
622 		} else {
623 			PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
624 			LV1_PROT = SYSMMU_V5_LV1_PROT;
625 			LV2_PROT = SYSMMU_V5_LV2_PROT;
626 		}
627 	}
628 
629 	pm_runtime_enable(dev);
630 
631 	of_iommu_set_ops(dev->of_node, &exynos_iommu_ops);
632 
633 	return 0;
634 }
635 
636 static int __maybe_unused exynos_sysmmu_suspend(struct device *dev)
637 {
638 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
639 	struct device *master = data->master;
640 
641 	if (master) {
642 		struct exynos_iommu_owner *owner = master->archdata.iommu;
643 
644 		mutex_lock(&owner->rpm_lock);
645 		if (data->domain) {
646 			dev_dbg(data->sysmmu, "saving state\n");
647 			__sysmmu_disable(data);
648 		}
649 		mutex_unlock(&owner->rpm_lock);
650 	}
651 	return 0;
652 }
653 
654 static int __maybe_unused exynos_sysmmu_resume(struct device *dev)
655 {
656 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
657 	struct device *master = data->master;
658 
659 	if (master) {
660 		struct exynos_iommu_owner *owner = master->archdata.iommu;
661 
662 		mutex_lock(&owner->rpm_lock);
663 		if (data->domain) {
664 			dev_dbg(data->sysmmu, "restoring state\n");
665 			__sysmmu_enable(data);
666 		}
667 		mutex_unlock(&owner->rpm_lock);
668 	}
669 	return 0;
670 }
671 
672 static const struct dev_pm_ops sysmmu_pm_ops = {
673 	SET_RUNTIME_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume, NULL)
674 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
675 				pm_runtime_force_resume)
676 };
677 
678 static const struct of_device_id sysmmu_of_match[] __initconst = {
679 	{ .compatible	= "samsung,exynos-sysmmu", },
680 	{ },
681 };
682 
683 static struct platform_driver exynos_sysmmu_driver __refdata = {
684 	.probe	= exynos_sysmmu_probe,
685 	.driver	= {
686 		.name		= "exynos-sysmmu",
687 		.of_match_table	= sysmmu_of_match,
688 		.pm		= &sysmmu_pm_ops,
689 		.suppress_bind_attrs = true,
690 	}
691 };
692 
693 static inline void update_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
694 {
695 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
696 				DMA_TO_DEVICE);
697 	*ent = cpu_to_le32(val);
698 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
699 				   DMA_TO_DEVICE);
700 }
701 
702 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
703 {
704 	struct exynos_iommu_domain *domain;
705 	dma_addr_t handle;
706 	int i;
707 
708 	/* Check if correct PTE offsets are initialized */
709 	BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
710 
711 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
712 	if (!domain)
713 		return NULL;
714 
715 	if (type == IOMMU_DOMAIN_DMA) {
716 		if (iommu_get_dma_cookie(&domain->domain) != 0)
717 			goto err_pgtable;
718 	} else if (type != IOMMU_DOMAIN_UNMANAGED) {
719 		goto err_pgtable;
720 	}
721 
722 	domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
723 	if (!domain->pgtable)
724 		goto err_dma_cookie;
725 
726 	domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
727 	if (!domain->lv2entcnt)
728 		goto err_counter;
729 
730 	/* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
731 	for (i = 0; i < NUM_LV1ENTRIES; i += 8) {
732 		domain->pgtable[i + 0] = ZERO_LV2LINK;
733 		domain->pgtable[i + 1] = ZERO_LV2LINK;
734 		domain->pgtable[i + 2] = ZERO_LV2LINK;
735 		domain->pgtable[i + 3] = ZERO_LV2LINK;
736 		domain->pgtable[i + 4] = ZERO_LV2LINK;
737 		domain->pgtable[i + 5] = ZERO_LV2LINK;
738 		domain->pgtable[i + 6] = ZERO_LV2LINK;
739 		domain->pgtable[i + 7] = ZERO_LV2LINK;
740 	}
741 
742 	handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
743 				DMA_TO_DEVICE);
744 	/* For mapping page table entries we rely on dma == phys */
745 	BUG_ON(handle != virt_to_phys(domain->pgtable));
746 
747 	spin_lock_init(&domain->lock);
748 	spin_lock_init(&domain->pgtablelock);
749 	INIT_LIST_HEAD(&domain->clients);
750 
751 	domain->domain.geometry.aperture_start = 0;
752 	domain->domain.geometry.aperture_end   = ~0UL;
753 	domain->domain.geometry.force_aperture = true;
754 
755 	return &domain->domain;
756 
757 err_counter:
758 	free_pages((unsigned long)domain->pgtable, 2);
759 err_dma_cookie:
760 	if (type == IOMMU_DOMAIN_DMA)
761 		iommu_put_dma_cookie(&domain->domain);
762 err_pgtable:
763 	kfree(domain);
764 	return NULL;
765 }
766 
767 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
768 {
769 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
770 	struct sysmmu_drvdata *data, *next;
771 	unsigned long flags;
772 	int i;
773 
774 	WARN_ON(!list_empty(&domain->clients));
775 
776 	spin_lock_irqsave(&domain->lock, flags);
777 
778 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
779 		spin_lock(&data->lock);
780 		__sysmmu_disable(data);
781 		data->pgtable = 0;
782 		data->domain = NULL;
783 		list_del_init(&data->domain_node);
784 		spin_unlock(&data->lock);
785 	}
786 
787 	spin_unlock_irqrestore(&domain->lock, flags);
788 
789 	if (iommu_domain->type == IOMMU_DOMAIN_DMA)
790 		iommu_put_dma_cookie(iommu_domain);
791 
792 	dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
793 			 DMA_TO_DEVICE);
794 
795 	for (i = 0; i < NUM_LV1ENTRIES; i++)
796 		if (lv1ent_page(domain->pgtable + i)) {
797 			phys_addr_t base = lv2table_base(domain->pgtable + i);
798 
799 			dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
800 					 DMA_TO_DEVICE);
801 			kmem_cache_free(lv2table_kmem_cache,
802 					phys_to_virt(base));
803 		}
804 
805 	free_pages((unsigned long)domain->pgtable, 2);
806 	free_pages((unsigned long)domain->lv2entcnt, 1);
807 	kfree(domain);
808 }
809 
810 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
811 				    struct device *dev)
812 {
813 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
814 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
815 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
816 	struct sysmmu_drvdata *data, *next;
817 	unsigned long flags;
818 
819 	if (!has_sysmmu(dev) || owner->domain != iommu_domain)
820 		return;
821 
822 	mutex_lock(&owner->rpm_lock);
823 
824 	list_for_each_entry(data, &owner->controllers, owner_node) {
825 		pm_runtime_get_noresume(data->sysmmu);
826 		if (pm_runtime_active(data->sysmmu))
827 			__sysmmu_disable(data);
828 		pm_runtime_put(data->sysmmu);
829 	}
830 
831 	spin_lock_irqsave(&domain->lock, flags);
832 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
833 		spin_lock(&data->lock);
834 		data->pgtable = 0;
835 		data->domain = NULL;
836 		list_del_init(&data->domain_node);
837 		spin_unlock(&data->lock);
838 	}
839 	owner->domain = NULL;
840 	spin_unlock_irqrestore(&domain->lock, flags);
841 
842 	mutex_unlock(&owner->rpm_lock);
843 
844 	dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n", __func__,
845 		&pagetable);
846 }
847 
848 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
849 				   struct device *dev)
850 {
851 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
852 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
853 	struct sysmmu_drvdata *data;
854 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
855 	unsigned long flags;
856 
857 	if (!has_sysmmu(dev))
858 		return -ENODEV;
859 
860 	if (owner->domain)
861 		exynos_iommu_detach_device(owner->domain, dev);
862 
863 	mutex_lock(&owner->rpm_lock);
864 
865 	spin_lock_irqsave(&domain->lock, flags);
866 	list_for_each_entry(data, &owner->controllers, owner_node) {
867 		spin_lock(&data->lock);
868 		data->pgtable = pagetable;
869 		data->domain = domain;
870 		list_add_tail(&data->domain_node, &domain->clients);
871 		spin_unlock(&data->lock);
872 	}
873 	owner->domain = iommu_domain;
874 	spin_unlock_irqrestore(&domain->lock, flags);
875 
876 	list_for_each_entry(data, &owner->controllers, owner_node) {
877 		pm_runtime_get_noresume(data->sysmmu);
878 		if (pm_runtime_active(data->sysmmu))
879 			__sysmmu_enable(data);
880 		pm_runtime_put(data->sysmmu);
881 	}
882 
883 	mutex_unlock(&owner->rpm_lock);
884 
885 	dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa\n", __func__,
886 		&pagetable);
887 
888 	return 0;
889 }
890 
891 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
892 		sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
893 {
894 	if (lv1ent_section(sent)) {
895 		WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
896 		return ERR_PTR(-EADDRINUSE);
897 	}
898 
899 	if (lv1ent_fault(sent)) {
900 		sysmmu_pte_t *pent;
901 		bool need_flush_flpd_cache = lv1ent_zero(sent);
902 
903 		pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
904 		BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
905 		if (!pent)
906 			return ERR_PTR(-ENOMEM);
907 
908 		update_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
909 		kmemleak_ignore(pent);
910 		*pgcounter = NUM_LV2ENTRIES;
911 		dma_map_single(dma_dev, pent, LV2TABLE_SIZE, DMA_TO_DEVICE);
912 
913 		/*
914 		 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
915 		 * FLPD cache may cache the address of zero_l2_table. This
916 		 * function replaces the zero_l2_table with new L2 page table
917 		 * to write valid mappings.
918 		 * Accessing the valid area may cause page fault since FLPD
919 		 * cache may still cache zero_l2_table for the valid area
920 		 * instead of new L2 page table that has the mapping
921 		 * information of the valid area.
922 		 * Thus any replacement of zero_l2_table with other valid L2
923 		 * page table must involve FLPD cache invalidation for System
924 		 * MMU v3.3.
925 		 * FLPD cache invalidation is performed with TLB invalidation
926 		 * by VPN without blocking. It is safe to invalidate TLB without
927 		 * blocking because the target address of TLB invalidation is
928 		 * not currently mapped.
929 		 */
930 		if (need_flush_flpd_cache) {
931 			struct sysmmu_drvdata *data;
932 
933 			spin_lock(&domain->lock);
934 			list_for_each_entry(data, &domain->clients, domain_node)
935 				sysmmu_tlb_invalidate_flpdcache(data, iova);
936 			spin_unlock(&domain->lock);
937 		}
938 	}
939 
940 	return page_entry(sent, iova);
941 }
942 
943 static int lv1set_section(struct exynos_iommu_domain *domain,
944 			  sysmmu_pte_t *sent, sysmmu_iova_t iova,
945 			  phys_addr_t paddr, int prot, short *pgcnt)
946 {
947 	if (lv1ent_section(sent)) {
948 		WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
949 			iova);
950 		return -EADDRINUSE;
951 	}
952 
953 	if (lv1ent_page(sent)) {
954 		if (*pgcnt != NUM_LV2ENTRIES) {
955 			WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
956 				iova);
957 			return -EADDRINUSE;
958 		}
959 
960 		kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
961 		*pgcnt = 0;
962 	}
963 
964 	update_pte(sent, mk_lv1ent_sect(paddr, prot));
965 
966 	spin_lock(&domain->lock);
967 	if (lv1ent_page_zero(sent)) {
968 		struct sysmmu_drvdata *data;
969 		/*
970 		 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
971 		 * entry by speculative prefetch of SLPD which has no mapping.
972 		 */
973 		list_for_each_entry(data, &domain->clients, domain_node)
974 			sysmmu_tlb_invalidate_flpdcache(data, iova);
975 	}
976 	spin_unlock(&domain->lock);
977 
978 	return 0;
979 }
980 
981 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
982 		       int prot, short *pgcnt)
983 {
984 	if (size == SPAGE_SIZE) {
985 		if (WARN_ON(!lv2ent_fault(pent)))
986 			return -EADDRINUSE;
987 
988 		update_pte(pent, mk_lv2ent_spage(paddr, prot));
989 		*pgcnt -= 1;
990 	} else { /* size == LPAGE_SIZE */
991 		int i;
992 		dma_addr_t pent_base = virt_to_phys(pent);
993 
994 		dma_sync_single_for_cpu(dma_dev, pent_base,
995 					sizeof(*pent) * SPAGES_PER_LPAGE,
996 					DMA_TO_DEVICE);
997 		for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
998 			if (WARN_ON(!lv2ent_fault(pent))) {
999 				if (i > 0)
1000 					memset(pent - i, 0, sizeof(*pent) * i);
1001 				return -EADDRINUSE;
1002 			}
1003 
1004 			*pent = mk_lv2ent_lpage(paddr, prot);
1005 		}
1006 		dma_sync_single_for_device(dma_dev, pent_base,
1007 					   sizeof(*pent) * SPAGES_PER_LPAGE,
1008 					   DMA_TO_DEVICE);
1009 		*pgcnt -= SPAGES_PER_LPAGE;
1010 	}
1011 
1012 	return 0;
1013 }
1014 
1015 /*
1016  * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1017  *
1018  * System MMU v3.x has advanced logic to improve address translation
1019  * performance with caching more page table entries by a page table walk.
1020  * However, the logic has a bug that while caching faulty page table entries,
1021  * System MMU reports page fault if the cached fault entry is hit even though
1022  * the fault entry is updated to a valid entry after the entry is cached.
1023  * To prevent caching faulty page table entries which may be updated to valid
1024  * entries later, the virtual memory manager should care about the workaround
1025  * for the problem. The following describes the workaround.
1026  *
1027  * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1028  * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1029  *
1030  * Precisely, any start address of I/O virtual region must be aligned with
1031  * the following sizes for System MMU v3.1 and v3.2.
1032  * System MMU v3.1: 128KiB
1033  * System MMU v3.2: 256KiB
1034  *
1035  * Because System MMU v3.3 caches page table entries more aggressively, it needs
1036  * more workarounds.
1037  * - Any two consecutive I/O virtual regions must have a hole of size larger
1038  *   than or equal to 128KiB.
1039  * - Start address of an I/O virtual region must be aligned by 128KiB.
1040  */
1041 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1042 			    unsigned long l_iova, phys_addr_t paddr, size_t size,
1043 			    int prot)
1044 {
1045 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1046 	sysmmu_pte_t *entry;
1047 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1048 	unsigned long flags;
1049 	int ret = -ENOMEM;
1050 
1051 	BUG_ON(domain->pgtable == NULL);
1052 	prot &= SYSMMU_SUPPORTED_PROT_BITS;
1053 
1054 	spin_lock_irqsave(&domain->pgtablelock, flags);
1055 
1056 	entry = section_entry(domain->pgtable, iova);
1057 
1058 	if (size == SECT_SIZE) {
1059 		ret = lv1set_section(domain, entry, iova, paddr, prot,
1060 				     &domain->lv2entcnt[lv1ent_offset(iova)]);
1061 	} else {
1062 		sysmmu_pte_t *pent;
1063 
1064 		pent = alloc_lv2entry(domain, entry, iova,
1065 				      &domain->lv2entcnt[lv1ent_offset(iova)]);
1066 
1067 		if (IS_ERR(pent))
1068 			ret = PTR_ERR(pent);
1069 		else
1070 			ret = lv2set_page(pent, paddr, size, prot,
1071 				       &domain->lv2entcnt[lv1ent_offset(iova)]);
1072 	}
1073 
1074 	if (ret)
1075 		pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1076 			__func__, ret, size, iova);
1077 
1078 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1079 
1080 	return ret;
1081 }
1082 
1083 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1084 					      sysmmu_iova_t iova, size_t size)
1085 {
1086 	struct sysmmu_drvdata *data;
1087 	unsigned long flags;
1088 
1089 	spin_lock_irqsave(&domain->lock, flags);
1090 
1091 	list_for_each_entry(data, &domain->clients, domain_node)
1092 		sysmmu_tlb_invalidate_entry(data, iova, size);
1093 
1094 	spin_unlock_irqrestore(&domain->lock, flags);
1095 }
1096 
1097 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1098 				 unsigned long l_iova, size_t size)
1099 {
1100 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1101 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1102 	sysmmu_pte_t *ent;
1103 	size_t err_pgsize;
1104 	unsigned long flags;
1105 
1106 	BUG_ON(domain->pgtable == NULL);
1107 
1108 	spin_lock_irqsave(&domain->pgtablelock, flags);
1109 
1110 	ent = section_entry(domain->pgtable, iova);
1111 
1112 	if (lv1ent_section(ent)) {
1113 		if (WARN_ON(size < SECT_SIZE)) {
1114 			err_pgsize = SECT_SIZE;
1115 			goto err;
1116 		}
1117 
1118 		/* workaround for h/w bug in System MMU v3.3 */
1119 		update_pte(ent, ZERO_LV2LINK);
1120 		size = SECT_SIZE;
1121 		goto done;
1122 	}
1123 
1124 	if (unlikely(lv1ent_fault(ent))) {
1125 		if (size > SECT_SIZE)
1126 			size = SECT_SIZE;
1127 		goto done;
1128 	}
1129 
1130 	/* lv1ent_page(sent) == true here */
1131 
1132 	ent = page_entry(ent, iova);
1133 
1134 	if (unlikely(lv2ent_fault(ent))) {
1135 		size = SPAGE_SIZE;
1136 		goto done;
1137 	}
1138 
1139 	if (lv2ent_small(ent)) {
1140 		update_pte(ent, 0);
1141 		size = SPAGE_SIZE;
1142 		domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1143 		goto done;
1144 	}
1145 
1146 	/* lv1ent_large(ent) == true here */
1147 	if (WARN_ON(size < LPAGE_SIZE)) {
1148 		err_pgsize = LPAGE_SIZE;
1149 		goto err;
1150 	}
1151 
1152 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1153 				sizeof(*ent) * SPAGES_PER_LPAGE,
1154 				DMA_TO_DEVICE);
1155 	memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1156 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1157 				   sizeof(*ent) * SPAGES_PER_LPAGE,
1158 				   DMA_TO_DEVICE);
1159 	size = LPAGE_SIZE;
1160 	domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1161 done:
1162 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1163 
1164 	exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1165 
1166 	return size;
1167 err:
1168 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1169 
1170 	pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1171 		__func__, size, iova, err_pgsize);
1172 
1173 	return 0;
1174 }
1175 
1176 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1177 					  dma_addr_t iova)
1178 {
1179 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1180 	sysmmu_pte_t *entry;
1181 	unsigned long flags;
1182 	phys_addr_t phys = 0;
1183 
1184 	spin_lock_irqsave(&domain->pgtablelock, flags);
1185 
1186 	entry = section_entry(domain->pgtable, iova);
1187 
1188 	if (lv1ent_section(entry)) {
1189 		phys = section_phys(entry) + section_offs(iova);
1190 	} else if (lv1ent_page(entry)) {
1191 		entry = page_entry(entry, iova);
1192 
1193 		if (lv2ent_large(entry))
1194 			phys = lpage_phys(entry) + lpage_offs(iova);
1195 		else if (lv2ent_small(entry))
1196 			phys = spage_phys(entry) + spage_offs(iova);
1197 	}
1198 
1199 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1200 
1201 	return phys;
1202 }
1203 
1204 static struct iommu_group *get_device_iommu_group(struct device *dev)
1205 {
1206 	struct iommu_group *group;
1207 
1208 	group = iommu_group_get(dev);
1209 	if (!group)
1210 		group = iommu_group_alloc();
1211 
1212 	return group;
1213 }
1214 
1215 static int exynos_iommu_add_device(struct device *dev)
1216 {
1217 	struct iommu_group *group;
1218 
1219 	if (!has_sysmmu(dev))
1220 		return -ENODEV;
1221 
1222 	group = iommu_group_get_for_dev(dev);
1223 
1224 	if (IS_ERR(group))
1225 		return PTR_ERR(group);
1226 
1227 	iommu_group_put(group);
1228 
1229 	return 0;
1230 }
1231 
1232 static void exynos_iommu_remove_device(struct device *dev)
1233 {
1234 	if (!has_sysmmu(dev))
1235 		return;
1236 
1237 	iommu_group_remove_device(dev);
1238 }
1239 
1240 static int exynos_iommu_of_xlate(struct device *dev,
1241 				 struct of_phandle_args *spec)
1242 {
1243 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1244 	struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1245 	struct sysmmu_drvdata *data;
1246 
1247 	if (!sysmmu)
1248 		return -ENODEV;
1249 
1250 	data = platform_get_drvdata(sysmmu);
1251 	if (!data)
1252 		return -ENODEV;
1253 
1254 	if (!owner) {
1255 		owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1256 		if (!owner)
1257 			return -ENOMEM;
1258 
1259 		INIT_LIST_HEAD(&owner->controllers);
1260 		mutex_init(&owner->rpm_lock);
1261 		dev->archdata.iommu = owner;
1262 	}
1263 
1264 	list_add_tail(&data->owner_node, &owner->controllers);
1265 	data->master = dev;
1266 
1267 	/*
1268 	 * SYSMMU will be runtime activated via device link (dependency) to its
1269 	 * master device, so there are no direct calls to pm_runtime_get/put
1270 	 * in this driver.
1271 	 */
1272 	device_link_add(dev, data->sysmmu, DL_FLAG_PM_RUNTIME);
1273 
1274 	return 0;
1275 }
1276 
1277 static struct iommu_ops exynos_iommu_ops = {
1278 	.domain_alloc = exynos_iommu_domain_alloc,
1279 	.domain_free = exynos_iommu_domain_free,
1280 	.attach_dev = exynos_iommu_attach_device,
1281 	.detach_dev = exynos_iommu_detach_device,
1282 	.map = exynos_iommu_map,
1283 	.unmap = exynos_iommu_unmap,
1284 	.map_sg = default_iommu_map_sg,
1285 	.iova_to_phys = exynos_iommu_iova_to_phys,
1286 	.device_group = get_device_iommu_group,
1287 	.add_device = exynos_iommu_add_device,
1288 	.remove_device = exynos_iommu_remove_device,
1289 	.pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1290 	.of_xlate = exynos_iommu_of_xlate,
1291 };
1292 
1293 static bool init_done;
1294 
1295 static int __init exynos_iommu_init(void)
1296 {
1297 	int ret;
1298 
1299 	lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1300 				LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1301 	if (!lv2table_kmem_cache) {
1302 		pr_err("%s: Failed to create kmem cache\n", __func__);
1303 		return -ENOMEM;
1304 	}
1305 
1306 	ret = platform_driver_register(&exynos_sysmmu_driver);
1307 	if (ret) {
1308 		pr_err("%s: Failed to register driver\n", __func__);
1309 		goto err_reg_driver;
1310 	}
1311 
1312 	zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1313 	if (zero_lv2_table == NULL) {
1314 		pr_err("%s: Failed to allocate zero level2 page table\n",
1315 			__func__);
1316 		ret = -ENOMEM;
1317 		goto err_zero_lv2;
1318 	}
1319 
1320 	ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1321 	if (ret) {
1322 		pr_err("%s: Failed to register exynos-iommu driver.\n",
1323 								__func__);
1324 		goto err_set_iommu;
1325 	}
1326 
1327 	init_done = true;
1328 
1329 	return 0;
1330 err_set_iommu:
1331 	kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1332 err_zero_lv2:
1333 	platform_driver_unregister(&exynos_sysmmu_driver);
1334 err_reg_driver:
1335 	kmem_cache_destroy(lv2table_kmem_cache);
1336 	return ret;
1337 }
1338 
1339 static int __init exynos_iommu_of_setup(struct device_node *np)
1340 {
1341 	struct platform_device *pdev;
1342 
1343 	if (!init_done)
1344 		exynos_iommu_init();
1345 
1346 	pdev = of_platform_device_create(np, NULL, platform_bus_type.dev_root);
1347 	if (!pdev)
1348 		return -ENODEV;
1349 
1350 	/*
1351 	 * use the first registered sysmmu device for performing
1352 	 * dma mapping operations on iommu page tables (cpu cache flush)
1353 	 */
1354 	if (!dma_dev)
1355 		dma_dev = &pdev->dev;
1356 
1357 	return 0;
1358 }
1359 
1360 IOMMU_OF_DECLARE(exynos_iommu_of, "samsung,exynos-sysmmu",
1361 		 exynos_iommu_of_setup);
1362