xref: /openbmc/linux/drivers/iommu/exynos-iommu.c (revision 0c874100)
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/kmemleak.h>
21 #include <linux/list.h>
22 #include <linux/of.h>
23 #include <linux/of_iommu.h>
24 #include <linux/of_platform.h>
25 #include <linux/platform_device.h>
26 #include <linux/pm_runtime.h>
27 #include <linux/slab.h>
28 #include <linux/dma-iommu.h>
29 
30 typedef u32 sysmmu_iova_t;
31 typedef u32 sysmmu_pte_t;
32 
33 /* We do not consider super section mapping (16MB) */
34 #define SECT_ORDER 20
35 #define LPAGE_ORDER 16
36 #define SPAGE_ORDER 12
37 
38 #define SECT_SIZE (1 << SECT_ORDER)
39 #define LPAGE_SIZE (1 << LPAGE_ORDER)
40 #define SPAGE_SIZE (1 << SPAGE_ORDER)
41 
42 #define SECT_MASK (~(SECT_SIZE - 1))
43 #define LPAGE_MASK (~(LPAGE_SIZE - 1))
44 #define SPAGE_MASK (~(SPAGE_SIZE - 1))
45 
46 #define lv1ent_fault(sent) ((*(sent) == ZERO_LV2LINK) || \
47 			   ((*(sent) & 3) == 0) || ((*(sent) & 3) == 3))
48 #define lv1ent_zero(sent) (*(sent) == ZERO_LV2LINK)
49 #define lv1ent_page_zero(sent) ((*(sent) & 3) == 1)
50 #define lv1ent_page(sent) ((*(sent) != ZERO_LV2LINK) && \
51 			  ((*(sent) & 3) == 1))
52 #define lv1ent_section(sent) ((*(sent) & 3) == 2)
53 
54 #define lv2ent_fault(pent) ((*(pent) & 3) == 0)
55 #define lv2ent_small(pent) ((*(pent) & 2) == 2)
56 #define lv2ent_large(pent) ((*(pent) & 3) == 1)
57 
58 /*
59  * v1.x - v3.x SYSMMU supports 32bit physical and 32bit virtual address spaces
60  * v5.0 introduced support for 36bit physical address space by shifting
61  * all page entry values by 4 bits.
62  * All SYSMMU controllers in the system support the address spaces of the same
63  * size, so PG_ENT_SHIFT can be initialized on first SYSMMU probe to proper
64  * value (0 or 4).
65  */
66 static short PG_ENT_SHIFT = -1;
67 #define SYSMMU_PG_ENT_SHIFT 0
68 #define SYSMMU_V5_PG_ENT_SHIFT 4
69 
70 static const sysmmu_pte_t *LV1_PROT;
71 static const sysmmu_pte_t SYSMMU_LV1_PROT[] = {
72 	((0 << 15) | (0 << 10)), /* no access */
73 	((1 << 15) | (1 << 10)), /* IOMMU_READ only */
74 	((0 << 15) | (1 << 10)), /* IOMMU_WRITE not supported, use read/write */
75 	((0 << 15) | (1 << 10)), /* IOMMU_READ | IOMMU_WRITE */
76 };
77 static const sysmmu_pte_t SYSMMU_V5_LV1_PROT[] = {
78 	(0 << 4), /* no access */
79 	(1 << 4), /* IOMMU_READ only */
80 	(2 << 4), /* IOMMU_WRITE only */
81 	(3 << 4), /* IOMMU_READ | IOMMU_WRITE */
82 };
83 
84 static const sysmmu_pte_t *LV2_PROT;
85 static const sysmmu_pte_t SYSMMU_LV2_PROT[] = {
86 	((0 << 9) | (0 << 4)), /* no access */
87 	((1 << 9) | (1 << 4)), /* IOMMU_READ only */
88 	((0 << 9) | (1 << 4)), /* IOMMU_WRITE not supported, use read/write */
89 	((0 << 9) | (1 << 4)), /* IOMMU_READ | IOMMU_WRITE */
90 };
91 static const sysmmu_pte_t SYSMMU_V5_LV2_PROT[] = {
92 	(0 << 2), /* no access */
93 	(1 << 2), /* IOMMU_READ only */
94 	(2 << 2), /* IOMMU_WRITE only */
95 	(3 << 2), /* IOMMU_READ | IOMMU_WRITE */
96 };
97 
98 #define SYSMMU_SUPPORTED_PROT_BITS (IOMMU_READ | IOMMU_WRITE)
99 
100 #define sect_to_phys(ent) (((phys_addr_t) ent) << PG_ENT_SHIFT)
101 #define section_phys(sent) (sect_to_phys(*(sent)) & SECT_MASK)
102 #define section_offs(iova) (iova & (SECT_SIZE - 1))
103 #define lpage_phys(pent) (sect_to_phys(*(pent)) & LPAGE_MASK)
104 #define lpage_offs(iova) (iova & (LPAGE_SIZE - 1))
105 #define spage_phys(pent) (sect_to_phys(*(pent)) & SPAGE_MASK)
106 #define spage_offs(iova) (iova & (SPAGE_SIZE - 1))
107 
108 #define NUM_LV1ENTRIES 4096
109 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
110 
111 static u32 lv1ent_offset(sysmmu_iova_t iova)
112 {
113 	return iova >> SECT_ORDER;
114 }
115 
116 static u32 lv2ent_offset(sysmmu_iova_t iova)
117 {
118 	return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
119 }
120 
121 #define LV1TABLE_SIZE (NUM_LV1ENTRIES * sizeof(sysmmu_pte_t))
122 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
123 
124 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
125 #define lv2table_base(sent) (sect_to_phys(*(sent) & 0xFFFFFFC0))
126 
127 #define mk_lv1ent_sect(pa, prot) ((pa >> PG_ENT_SHIFT) | LV1_PROT[prot] | 2)
128 #define mk_lv1ent_page(pa) ((pa >> PG_ENT_SHIFT) | 1)
129 #define mk_lv2ent_lpage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 1)
130 #define mk_lv2ent_spage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 2)
131 
132 #define CTRL_ENABLE	0x5
133 #define CTRL_BLOCK	0x7
134 #define CTRL_DISABLE	0x0
135 
136 #define CFG_LRU		0x1
137 #define CFG_EAP		(1 << 2)
138 #define CFG_QOS(n)	((n & 0xF) << 7)
139 #define CFG_ACGEN	(1 << 24) /* System MMU 3.3 only */
140 #define CFG_SYSSEL	(1 << 22) /* System MMU 3.2 only */
141 #define CFG_FLPDCACHE	(1 << 20) /* System MMU 3.2+ only */
142 
143 /* common registers */
144 #define REG_MMU_CTRL		0x000
145 #define REG_MMU_CFG		0x004
146 #define REG_MMU_STATUS		0x008
147 #define REG_MMU_VERSION		0x034
148 
149 #define MMU_MAJ_VER(val)	((val) >> 7)
150 #define MMU_MIN_VER(val)	((val) & 0x7F)
151 #define MMU_RAW_VER(reg)	(((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
152 
153 #define MAKE_MMU_VER(maj, min)	((((maj) & 0xF) << 7) | ((min) & 0x7F))
154 
155 /* v1.x - v3.x registers */
156 #define REG_MMU_FLUSH		0x00C
157 #define REG_MMU_FLUSH_ENTRY	0x010
158 #define REG_PT_BASE_ADDR	0x014
159 #define REG_INT_STATUS		0x018
160 #define REG_INT_CLEAR		0x01C
161 
162 #define REG_PAGE_FAULT_ADDR	0x024
163 #define REG_AW_FAULT_ADDR	0x028
164 #define REG_AR_FAULT_ADDR	0x02C
165 #define REG_DEFAULT_SLAVE_ADDR	0x030
166 
167 /* v5.x registers */
168 #define REG_V5_PT_BASE_PFN	0x00C
169 #define REG_V5_MMU_FLUSH_ALL	0x010
170 #define REG_V5_MMU_FLUSH_ENTRY	0x014
171 #define REG_V5_MMU_FLUSH_RANGE	0x018
172 #define REG_V5_MMU_FLUSH_START	0x020
173 #define REG_V5_MMU_FLUSH_END	0x024
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 	struct device_link *link;	/* runtime PM link to master */
268 	void __iomem *sfrbase;		/* our registers */
269 	struct clk *clk;		/* SYSMMU's clock */
270 	struct clk *aclk;		/* SYSMMU's aclk clock */
271 	struct clk *pclk;		/* SYSMMU's pclk clock */
272 	struct clk *clk_master;		/* master's device clock */
273 	spinlock_t lock;		/* lock for modyfying state */
274 	bool active;			/* current status */
275 	struct exynos_iommu_domain *domain; /* domain we belong to */
276 	struct list_head domain_node;	/* node for domain clients list */
277 	struct list_head owner_node;	/* node for owner controllers list */
278 	phys_addr_t pgtable;		/* assigned page table structure */
279 	unsigned int version;		/* our version */
280 
281 	struct iommu_device iommu;	/* IOMMU core handle */
282 };
283 
284 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
285 {
286 	return container_of(dom, struct exynos_iommu_domain, domain);
287 }
288 
289 static void sysmmu_unblock(struct sysmmu_drvdata *data)
290 {
291 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
292 }
293 
294 static bool sysmmu_block(struct sysmmu_drvdata *data)
295 {
296 	int i = 120;
297 
298 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
299 	while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
300 		--i;
301 
302 	if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
303 		sysmmu_unblock(data);
304 		return false;
305 	}
306 
307 	return true;
308 }
309 
310 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
311 {
312 	if (MMU_MAJ_VER(data->version) < 5)
313 		writel(0x1, data->sfrbase + REG_MMU_FLUSH);
314 	else
315 		writel(0x1, data->sfrbase + REG_V5_MMU_FLUSH_ALL);
316 }
317 
318 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
319 				sysmmu_iova_t iova, unsigned int num_inv)
320 {
321 	unsigned int i;
322 
323 	if (MMU_MAJ_VER(data->version) < 5) {
324 		for (i = 0; i < num_inv; i++) {
325 			writel((iova & SPAGE_MASK) | 1,
326 				     data->sfrbase + REG_MMU_FLUSH_ENTRY);
327 			iova += SPAGE_SIZE;
328 		}
329 	} else {
330 		if (num_inv == 1) {
331 			writel((iova & SPAGE_MASK) | 1,
332 				     data->sfrbase + REG_V5_MMU_FLUSH_ENTRY);
333 		} else {
334 			writel((iova & SPAGE_MASK),
335 				     data->sfrbase + REG_V5_MMU_FLUSH_START);
336 			writel((iova & SPAGE_MASK) + (num_inv - 1) * SPAGE_SIZE,
337 				     data->sfrbase + REG_V5_MMU_FLUSH_END);
338 			writel(1, data->sfrbase + REG_V5_MMU_FLUSH_RANGE);
339 		}
340 	}
341 }
342 
343 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
344 {
345 	if (MMU_MAJ_VER(data->version) < 5)
346 		writel(pgd, data->sfrbase + REG_PT_BASE_ADDR);
347 	else
348 		writel(pgd >> PAGE_SHIFT,
349 			     data->sfrbase + REG_V5_PT_BASE_PFN);
350 
351 	__sysmmu_tlb_invalidate(data);
352 }
353 
354 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
355 {
356 	BUG_ON(clk_prepare_enable(data->clk_master));
357 	BUG_ON(clk_prepare_enable(data->clk));
358 	BUG_ON(clk_prepare_enable(data->pclk));
359 	BUG_ON(clk_prepare_enable(data->aclk));
360 }
361 
362 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
363 {
364 	clk_disable_unprepare(data->aclk);
365 	clk_disable_unprepare(data->pclk);
366 	clk_disable_unprepare(data->clk);
367 	clk_disable_unprepare(data->clk_master);
368 }
369 
370 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
371 {
372 	u32 ver;
373 
374 	__sysmmu_enable_clocks(data);
375 
376 	ver = readl(data->sfrbase + REG_MMU_VERSION);
377 
378 	/* controllers on some SoCs don't report proper version */
379 	if (ver == 0x80000001u)
380 		data->version = MAKE_MMU_VER(1, 0);
381 	else
382 		data->version = MMU_RAW_VER(ver);
383 
384 	dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
385 		MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
386 
387 	__sysmmu_disable_clocks(data);
388 }
389 
390 static void show_fault_information(struct sysmmu_drvdata *data,
391 				   const struct sysmmu_fault_info *finfo,
392 				   sysmmu_iova_t fault_addr)
393 {
394 	sysmmu_pte_t *ent;
395 
396 	dev_err(data->sysmmu, "%s: %s FAULT occurred at %#x\n",
397 		dev_name(data->master), finfo->name, fault_addr);
398 	dev_dbg(data->sysmmu, "Page table base: %pa\n", &data->pgtable);
399 	ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
400 	dev_dbg(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
401 	if (lv1ent_page(ent)) {
402 		ent = page_entry(ent, fault_addr);
403 		dev_dbg(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
404 	}
405 }
406 
407 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
408 {
409 	/* SYSMMU is in blocked state when interrupt occurred. */
410 	struct sysmmu_drvdata *data = dev_id;
411 	const struct sysmmu_fault_info *finfo;
412 	unsigned int i, n, itype;
413 	sysmmu_iova_t fault_addr = -1;
414 	unsigned short reg_status, reg_clear;
415 	int ret = -ENOSYS;
416 
417 	WARN_ON(!data->active);
418 
419 	if (MMU_MAJ_VER(data->version) < 5) {
420 		reg_status = REG_INT_STATUS;
421 		reg_clear = REG_INT_CLEAR;
422 		finfo = sysmmu_faults;
423 		n = ARRAY_SIZE(sysmmu_faults);
424 	} else {
425 		reg_status = REG_V5_INT_STATUS;
426 		reg_clear = REG_V5_INT_CLEAR;
427 		finfo = sysmmu_v5_faults;
428 		n = ARRAY_SIZE(sysmmu_v5_faults);
429 	}
430 
431 	spin_lock(&data->lock);
432 
433 	clk_enable(data->clk_master);
434 
435 	itype = __ffs(readl(data->sfrbase + reg_status));
436 	for (i = 0; i < n; i++, finfo++)
437 		if (finfo->bit == itype)
438 			break;
439 	/* unknown/unsupported fault */
440 	BUG_ON(i == n);
441 
442 	/* print debug message */
443 	fault_addr = readl(data->sfrbase + finfo->addr_reg);
444 	show_fault_information(data, finfo, fault_addr);
445 
446 	if (data->domain)
447 		ret = report_iommu_fault(&data->domain->domain,
448 					data->master, fault_addr, finfo->type);
449 	/* fault is not recovered by fault handler */
450 	BUG_ON(ret != 0);
451 
452 	writel(1 << itype, data->sfrbase + reg_clear);
453 
454 	sysmmu_unblock(data);
455 
456 	clk_disable(data->clk_master);
457 
458 	spin_unlock(&data->lock);
459 
460 	return IRQ_HANDLED;
461 }
462 
463 static void __sysmmu_disable(struct sysmmu_drvdata *data)
464 {
465 	unsigned long flags;
466 
467 	clk_enable(data->clk_master);
468 
469 	spin_lock_irqsave(&data->lock, flags);
470 	writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
471 	writel(0, data->sfrbase + REG_MMU_CFG);
472 	data->active = false;
473 	spin_unlock_irqrestore(&data->lock, flags);
474 
475 	__sysmmu_disable_clocks(data);
476 }
477 
478 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
479 {
480 	unsigned int cfg;
481 
482 	if (data->version <= MAKE_MMU_VER(3, 1))
483 		cfg = CFG_LRU | CFG_QOS(15);
484 	else if (data->version <= MAKE_MMU_VER(3, 2))
485 		cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
486 	else
487 		cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
488 
489 	cfg |= CFG_EAP; /* enable access protection bits check */
490 
491 	writel(cfg, data->sfrbase + REG_MMU_CFG);
492 }
493 
494 static void __sysmmu_enable(struct sysmmu_drvdata *data)
495 {
496 	unsigned long flags;
497 
498 	__sysmmu_enable_clocks(data);
499 
500 	spin_lock_irqsave(&data->lock, flags);
501 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
502 	__sysmmu_init_config(data);
503 	__sysmmu_set_ptbase(data, data->pgtable);
504 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
505 	data->active = true;
506 	spin_unlock_irqrestore(&data->lock, flags);
507 
508 	/*
509 	 * SYSMMU driver keeps master's clock enabled only for the short
510 	 * time, while accessing the registers. For performing address
511 	 * translation during DMA transaction it relies on the client
512 	 * driver to enable it.
513 	 */
514 	clk_disable(data->clk_master);
515 }
516 
517 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
518 					    sysmmu_iova_t iova)
519 {
520 	unsigned long flags;
521 
522 	spin_lock_irqsave(&data->lock, flags);
523 	if (data->active && data->version >= MAKE_MMU_VER(3, 3)) {
524 		clk_enable(data->clk_master);
525 		if (sysmmu_block(data)) {
526 			if (data->version >= MAKE_MMU_VER(5, 0))
527 				__sysmmu_tlb_invalidate(data);
528 			else
529 				__sysmmu_tlb_invalidate_entry(data, iova, 1);
530 			sysmmu_unblock(data);
531 		}
532 		clk_disable(data->clk_master);
533 	}
534 	spin_unlock_irqrestore(&data->lock, flags);
535 }
536 
537 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
538 					sysmmu_iova_t iova, size_t size)
539 {
540 	unsigned long flags;
541 
542 	spin_lock_irqsave(&data->lock, flags);
543 	if (data->active) {
544 		unsigned int num_inv = 1;
545 
546 		clk_enable(data->clk_master);
547 
548 		/*
549 		 * L2TLB invalidation required
550 		 * 4KB page: 1 invalidation
551 		 * 64KB page: 16 invalidations
552 		 * 1MB page: 64 invalidations
553 		 * because it is set-associative TLB
554 		 * with 8-way and 64 sets.
555 		 * 1MB page can be cached in one of all sets.
556 		 * 64KB page can be one of 16 consecutive sets.
557 		 */
558 		if (MMU_MAJ_VER(data->version) == 2)
559 			num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
560 
561 		if (sysmmu_block(data)) {
562 			__sysmmu_tlb_invalidate_entry(data, iova, num_inv);
563 			sysmmu_unblock(data);
564 		}
565 		clk_disable(data->clk_master);
566 	}
567 	spin_unlock_irqrestore(&data->lock, flags);
568 }
569 
570 static const struct iommu_ops exynos_iommu_ops;
571 
572 static int __init exynos_sysmmu_probe(struct platform_device *pdev)
573 {
574 	int irq, ret;
575 	struct device *dev = &pdev->dev;
576 	struct sysmmu_drvdata *data;
577 	struct resource *res;
578 
579 	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
580 	if (!data)
581 		return -ENOMEM;
582 
583 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
584 	data->sfrbase = devm_ioremap_resource(dev, res);
585 	if (IS_ERR(data->sfrbase))
586 		return PTR_ERR(data->sfrbase);
587 
588 	irq = platform_get_irq(pdev, 0);
589 	if (irq <= 0) {
590 		dev_err(dev, "Unable to find IRQ resource\n");
591 		return irq;
592 	}
593 
594 	ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
595 				dev_name(dev), data);
596 	if (ret) {
597 		dev_err(dev, "Unabled to register handler of irq %d\n", irq);
598 		return ret;
599 	}
600 
601 	data->clk = devm_clk_get(dev, "sysmmu");
602 	if (PTR_ERR(data->clk) == -ENOENT)
603 		data->clk = NULL;
604 	else if (IS_ERR(data->clk))
605 		return PTR_ERR(data->clk);
606 
607 	data->aclk = devm_clk_get(dev, "aclk");
608 	if (PTR_ERR(data->aclk) == -ENOENT)
609 		data->aclk = NULL;
610 	else if (IS_ERR(data->aclk))
611 		return PTR_ERR(data->aclk);
612 
613 	data->pclk = devm_clk_get(dev, "pclk");
614 	if (PTR_ERR(data->pclk) == -ENOENT)
615 		data->pclk = NULL;
616 	else if (IS_ERR(data->pclk))
617 		return PTR_ERR(data->pclk);
618 
619 	if (!data->clk && (!data->aclk || !data->pclk)) {
620 		dev_err(dev, "Failed to get device clock(s)!\n");
621 		return -ENOSYS;
622 	}
623 
624 	data->clk_master = devm_clk_get(dev, "master");
625 	if (PTR_ERR(data->clk_master) == -ENOENT)
626 		data->clk_master = NULL;
627 	else if (IS_ERR(data->clk_master))
628 		return PTR_ERR(data->clk_master);
629 
630 	data->sysmmu = dev;
631 	spin_lock_init(&data->lock);
632 
633 	ret = iommu_device_sysfs_add(&data->iommu, &pdev->dev, NULL,
634 				     dev_name(data->sysmmu));
635 	if (ret)
636 		return ret;
637 
638 	iommu_device_set_ops(&data->iommu, &exynos_iommu_ops);
639 	iommu_device_set_fwnode(&data->iommu, &dev->of_node->fwnode);
640 
641 	ret = iommu_device_register(&data->iommu);
642 	if (ret)
643 		return ret;
644 
645 	platform_set_drvdata(pdev, data);
646 
647 	__sysmmu_get_version(data);
648 	if (PG_ENT_SHIFT < 0) {
649 		if (MMU_MAJ_VER(data->version) < 5) {
650 			PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
651 			LV1_PROT = SYSMMU_LV1_PROT;
652 			LV2_PROT = SYSMMU_LV2_PROT;
653 		} else {
654 			PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
655 			LV1_PROT = SYSMMU_V5_LV1_PROT;
656 			LV2_PROT = SYSMMU_V5_LV2_PROT;
657 		}
658 	}
659 
660 	/*
661 	 * use the first registered sysmmu device for performing
662 	 * dma mapping operations on iommu page tables (cpu cache flush)
663 	 */
664 	if (!dma_dev)
665 		dma_dev = &pdev->dev;
666 
667 	pm_runtime_enable(dev);
668 
669 	return 0;
670 }
671 
672 static int __maybe_unused exynos_sysmmu_suspend(struct device *dev)
673 {
674 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
675 	struct device *master = data->master;
676 
677 	if (master) {
678 		struct exynos_iommu_owner *owner = master->archdata.iommu;
679 
680 		mutex_lock(&owner->rpm_lock);
681 		if (data->domain) {
682 			dev_dbg(data->sysmmu, "saving state\n");
683 			__sysmmu_disable(data);
684 		}
685 		mutex_unlock(&owner->rpm_lock);
686 	}
687 	return 0;
688 }
689 
690 static int __maybe_unused exynos_sysmmu_resume(struct device *dev)
691 {
692 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
693 	struct device *master = data->master;
694 
695 	if (master) {
696 		struct exynos_iommu_owner *owner = master->archdata.iommu;
697 
698 		mutex_lock(&owner->rpm_lock);
699 		if (data->domain) {
700 			dev_dbg(data->sysmmu, "restoring state\n");
701 			__sysmmu_enable(data);
702 		}
703 		mutex_unlock(&owner->rpm_lock);
704 	}
705 	return 0;
706 }
707 
708 static const struct dev_pm_ops sysmmu_pm_ops = {
709 	SET_RUNTIME_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume, NULL)
710 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
711 				pm_runtime_force_resume)
712 };
713 
714 static const struct of_device_id sysmmu_of_match[] = {
715 	{ .compatible	= "samsung,exynos-sysmmu", },
716 	{ },
717 };
718 
719 static struct platform_driver exynos_sysmmu_driver __refdata = {
720 	.probe	= exynos_sysmmu_probe,
721 	.driver	= {
722 		.name		= "exynos-sysmmu",
723 		.of_match_table	= sysmmu_of_match,
724 		.pm		= &sysmmu_pm_ops,
725 		.suppress_bind_attrs = true,
726 	}
727 };
728 
729 static inline void update_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
730 {
731 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
732 				DMA_TO_DEVICE);
733 	*ent = cpu_to_le32(val);
734 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
735 				   DMA_TO_DEVICE);
736 }
737 
738 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
739 {
740 	struct exynos_iommu_domain *domain;
741 	dma_addr_t handle;
742 	int i;
743 
744 	/* Check if correct PTE offsets are initialized */
745 	BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
746 
747 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
748 	if (!domain)
749 		return NULL;
750 
751 	if (type == IOMMU_DOMAIN_DMA) {
752 		if (iommu_get_dma_cookie(&domain->domain) != 0)
753 			goto err_pgtable;
754 	} else if (type != IOMMU_DOMAIN_UNMANAGED) {
755 		goto err_pgtable;
756 	}
757 
758 	domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
759 	if (!domain->pgtable)
760 		goto err_dma_cookie;
761 
762 	domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
763 	if (!domain->lv2entcnt)
764 		goto err_counter;
765 
766 	/* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
767 	for (i = 0; i < NUM_LV1ENTRIES; i++)
768 		domain->pgtable[i] = ZERO_LV2LINK;
769 
770 	handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
771 				DMA_TO_DEVICE);
772 	/* For mapping page table entries we rely on dma == phys */
773 	BUG_ON(handle != virt_to_phys(domain->pgtable));
774 	if (dma_mapping_error(dma_dev, handle))
775 		goto err_lv2ent;
776 
777 	spin_lock_init(&domain->lock);
778 	spin_lock_init(&domain->pgtablelock);
779 	INIT_LIST_HEAD(&domain->clients);
780 
781 	domain->domain.geometry.aperture_start = 0;
782 	domain->domain.geometry.aperture_end   = ~0UL;
783 	domain->domain.geometry.force_aperture = true;
784 
785 	return &domain->domain;
786 
787 err_lv2ent:
788 	free_pages((unsigned long)domain->lv2entcnt, 1);
789 err_counter:
790 	free_pages((unsigned long)domain->pgtable, 2);
791 err_dma_cookie:
792 	if (type == IOMMU_DOMAIN_DMA)
793 		iommu_put_dma_cookie(&domain->domain);
794 err_pgtable:
795 	kfree(domain);
796 	return NULL;
797 }
798 
799 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
800 {
801 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
802 	struct sysmmu_drvdata *data, *next;
803 	unsigned long flags;
804 	int i;
805 
806 	WARN_ON(!list_empty(&domain->clients));
807 
808 	spin_lock_irqsave(&domain->lock, flags);
809 
810 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
811 		spin_lock(&data->lock);
812 		__sysmmu_disable(data);
813 		data->pgtable = 0;
814 		data->domain = NULL;
815 		list_del_init(&data->domain_node);
816 		spin_unlock(&data->lock);
817 	}
818 
819 	spin_unlock_irqrestore(&domain->lock, flags);
820 
821 	if (iommu_domain->type == IOMMU_DOMAIN_DMA)
822 		iommu_put_dma_cookie(iommu_domain);
823 
824 	dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
825 			 DMA_TO_DEVICE);
826 
827 	for (i = 0; i < NUM_LV1ENTRIES; i++)
828 		if (lv1ent_page(domain->pgtable + i)) {
829 			phys_addr_t base = lv2table_base(domain->pgtable + i);
830 
831 			dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
832 					 DMA_TO_DEVICE);
833 			kmem_cache_free(lv2table_kmem_cache,
834 					phys_to_virt(base));
835 		}
836 
837 	free_pages((unsigned long)domain->pgtable, 2);
838 	free_pages((unsigned long)domain->lv2entcnt, 1);
839 	kfree(domain);
840 }
841 
842 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
843 				    struct device *dev)
844 {
845 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
846 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
847 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
848 	struct sysmmu_drvdata *data, *next;
849 	unsigned long flags;
850 
851 	if (!has_sysmmu(dev) || owner->domain != iommu_domain)
852 		return;
853 
854 	mutex_lock(&owner->rpm_lock);
855 
856 	list_for_each_entry(data, &owner->controllers, owner_node) {
857 		pm_runtime_get_noresume(data->sysmmu);
858 		if (pm_runtime_active(data->sysmmu))
859 			__sysmmu_disable(data);
860 		pm_runtime_put(data->sysmmu);
861 	}
862 
863 	spin_lock_irqsave(&domain->lock, flags);
864 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
865 		spin_lock(&data->lock);
866 		data->pgtable = 0;
867 		data->domain = NULL;
868 		list_del_init(&data->domain_node);
869 		spin_unlock(&data->lock);
870 	}
871 	owner->domain = NULL;
872 	spin_unlock_irqrestore(&domain->lock, flags);
873 
874 	mutex_unlock(&owner->rpm_lock);
875 
876 	dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n", __func__,
877 		&pagetable);
878 }
879 
880 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
881 				   struct device *dev)
882 {
883 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
884 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
885 	struct sysmmu_drvdata *data;
886 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
887 	unsigned long flags;
888 
889 	if (!has_sysmmu(dev))
890 		return -ENODEV;
891 
892 	if (owner->domain)
893 		exynos_iommu_detach_device(owner->domain, dev);
894 
895 	mutex_lock(&owner->rpm_lock);
896 
897 	spin_lock_irqsave(&domain->lock, flags);
898 	list_for_each_entry(data, &owner->controllers, owner_node) {
899 		spin_lock(&data->lock);
900 		data->pgtable = pagetable;
901 		data->domain = domain;
902 		list_add_tail(&data->domain_node, &domain->clients);
903 		spin_unlock(&data->lock);
904 	}
905 	owner->domain = iommu_domain;
906 	spin_unlock_irqrestore(&domain->lock, flags);
907 
908 	list_for_each_entry(data, &owner->controllers, owner_node) {
909 		pm_runtime_get_noresume(data->sysmmu);
910 		if (pm_runtime_active(data->sysmmu))
911 			__sysmmu_enable(data);
912 		pm_runtime_put(data->sysmmu);
913 	}
914 
915 	mutex_unlock(&owner->rpm_lock);
916 
917 	dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa\n", __func__,
918 		&pagetable);
919 
920 	return 0;
921 }
922 
923 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
924 		sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
925 {
926 	if (lv1ent_section(sent)) {
927 		WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
928 		return ERR_PTR(-EADDRINUSE);
929 	}
930 
931 	if (lv1ent_fault(sent)) {
932 		dma_addr_t handle;
933 		sysmmu_pte_t *pent;
934 		bool need_flush_flpd_cache = lv1ent_zero(sent);
935 
936 		pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
937 		BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
938 		if (!pent)
939 			return ERR_PTR(-ENOMEM);
940 
941 		update_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
942 		kmemleak_ignore(pent);
943 		*pgcounter = NUM_LV2ENTRIES;
944 		handle = dma_map_single(dma_dev, pent, LV2TABLE_SIZE,
945 					DMA_TO_DEVICE);
946 		if (dma_mapping_error(dma_dev, handle)) {
947 			kmem_cache_free(lv2table_kmem_cache, pent);
948 			return ERR_PTR(-EADDRINUSE);
949 		}
950 
951 		/*
952 		 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
953 		 * FLPD cache may cache the address of zero_l2_table. This
954 		 * function replaces the zero_l2_table with new L2 page table
955 		 * to write valid mappings.
956 		 * Accessing the valid area may cause page fault since FLPD
957 		 * cache may still cache zero_l2_table for the valid area
958 		 * instead of new L2 page table that has the mapping
959 		 * information of the valid area.
960 		 * Thus any replacement of zero_l2_table with other valid L2
961 		 * page table must involve FLPD cache invalidation for System
962 		 * MMU v3.3.
963 		 * FLPD cache invalidation is performed with TLB invalidation
964 		 * by VPN without blocking. It is safe to invalidate TLB without
965 		 * blocking because the target address of TLB invalidation is
966 		 * not currently mapped.
967 		 */
968 		if (need_flush_flpd_cache) {
969 			struct sysmmu_drvdata *data;
970 
971 			spin_lock(&domain->lock);
972 			list_for_each_entry(data, &domain->clients, domain_node)
973 				sysmmu_tlb_invalidate_flpdcache(data, iova);
974 			spin_unlock(&domain->lock);
975 		}
976 	}
977 
978 	return page_entry(sent, iova);
979 }
980 
981 static int lv1set_section(struct exynos_iommu_domain *domain,
982 			  sysmmu_pte_t *sent, sysmmu_iova_t iova,
983 			  phys_addr_t paddr, int prot, short *pgcnt)
984 {
985 	if (lv1ent_section(sent)) {
986 		WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
987 			iova);
988 		return -EADDRINUSE;
989 	}
990 
991 	if (lv1ent_page(sent)) {
992 		if (*pgcnt != NUM_LV2ENTRIES) {
993 			WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
994 				iova);
995 			return -EADDRINUSE;
996 		}
997 
998 		kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
999 		*pgcnt = 0;
1000 	}
1001 
1002 	update_pte(sent, mk_lv1ent_sect(paddr, prot));
1003 
1004 	spin_lock(&domain->lock);
1005 	if (lv1ent_page_zero(sent)) {
1006 		struct sysmmu_drvdata *data;
1007 		/*
1008 		 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
1009 		 * entry by speculative prefetch of SLPD which has no mapping.
1010 		 */
1011 		list_for_each_entry(data, &domain->clients, domain_node)
1012 			sysmmu_tlb_invalidate_flpdcache(data, iova);
1013 	}
1014 	spin_unlock(&domain->lock);
1015 
1016 	return 0;
1017 }
1018 
1019 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
1020 		       int prot, short *pgcnt)
1021 {
1022 	if (size == SPAGE_SIZE) {
1023 		if (WARN_ON(!lv2ent_fault(pent)))
1024 			return -EADDRINUSE;
1025 
1026 		update_pte(pent, mk_lv2ent_spage(paddr, prot));
1027 		*pgcnt -= 1;
1028 	} else { /* size == LPAGE_SIZE */
1029 		int i;
1030 		dma_addr_t pent_base = virt_to_phys(pent);
1031 
1032 		dma_sync_single_for_cpu(dma_dev, pent_base,
1033 					sizeof(*pent) * SPAGES_PER_LPAGE,
1034 					DMA_TO_DEVICE);
1035 		for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1036 			if (WARN_ON(!lv2ent_fault(pent))) {
1037 				if (i > 0)
1038 					memset(pent - i, 0, sizeof(*pent) * i);
1039 				return -EADDRINUSE;
1040 			}
1041 
1042 			*pent = mk_lv2ent_lpage(paddr, prot);
1043 		}
1044 		dma_sync_single_for_device(dma_dev, pent_base,
1045 					   sizeof(*pent) * SPAGES_PER_LPAGE,
1046 					   DMA_TO_DEVICE);
1047 		*pgcnt -= SPAGES_PER_LPAGE;
1048 	}
1049 
1050 	return 0;
1051 }
1052 
1053 /*
1054  * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1055  *
1056  * System MMU v3.x has advanced logic to improve address translation
1057  * performance with caching more page table entries by a page table walk.
1058  * However, the logic has a bug that while caching faulty page table entries,
1059  * System MMU reports page fault if the cached fault entry is hit even though
1060  * the fault entry is updated to a valid entry after the entry is cached.
1061  * To prevent caching faulty page table entries which may be updated to valid
1062  * entries later, the virtual memory manager should care about the workaround
1063  * for the problem. The following describes the workaround.
1064  *
1065  * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1066  * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1067  *
1068  * Precisely, any start address of I/O virtual region must be aligned with
1069  * the following sizes for System MMU v3.1 and v3.2.
1070  * System MMU v3.1: 128KiB
1071  * System MMU v3.2: 256KiB
1072  *
1073  * Because System MMU v3.3 caches page table entries more aggressively, it needs
1074  * more workarounds.
1075  * - Any two consecutive I/O virtual regions must have a hole of size larger
1076  *   than or equal to 128KiB.
1077  * - Start address of an I/O virtual region must be aligned by 128KiB.
1078  */
1079 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1080 			    unsigned long l_iova, phys_addr_t paddr, size_t size,
1081 			    int prot)
1082 {
1083 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1084 	sysmmu_pte_t *entry;
1085 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1086 	unsigned long flags;
1087 	int ret = -ENOMEM;
1088 
1089 	BUG_ON(domain->pgtable == NULL);
1090 	prot &= SYSMMU_SUPPORTED_PROT_BITS;
1091 
1092 	spin_lock_irqsave(&domain->pgtablelock, flags);
1093 
1094 	entry = section_entry(domain->pgtable, iova);
1095 
1096 	if (size == SECT_SIZE) {
1097 		ret = lv1set_section(domain, entry, iova, paddr, prot,
1098 				     &domain->lv2entcnt[lv1ent_offset(iova)]);
1099 	} else {
1100 		sysmmu_pte_t *pent;
1101 
1102 		pent = alloc_lv2entry(domain, entry, iova,
1103 				      &domain->lv2entcnt[lv1ent_offset(iova)]);
1104 
1105 		if (IS_ERR(pent))
1106 			ret = PTR_ERR(pent);
1107 		else
1108 			ret = lv2set_page(pent, paddr, size, prot,
1109 				       &domain->lv2entcnt[lv1ent_offset(iova)]);
1110 	}
1111 
1112 	if (ret)
1113 		pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1114 			__func__, ret, size, iova);
1115 
1116 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1117 
1118 	return ret;
1119 }
1120 
1121 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1122 					      sysmmu_iova_t iova, size_t size)
1123 {
1124 	struct sysmmu_drvdata *data;
1125 	unsigned long flags;
1126 
1127 	spin_lock_irqsave(&domain->lock, flags);
1128 
1129 	list_for_each_entry(data, &domain->clients, domain_node)
1130 		sysmmu_tlb_invalidate_entry(data, iova, size);
1131 
1132 	spin_unlock_irqrestore(&domain->lock, flags);
1133 }
1134 
1135 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1136 				 unsigned long l_iova, size_t size)
1137 {
1138 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1139 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1140 	sysmmu_pte_t *ent;
1141 	size_t err_pgsize;
1142 	unsigned long flags;
1143 
1144 	BUG_ON(domain->pgtable == NULL);
1145 
1146 	spin_lock_irqsave(&domain->pgtablelock, flags);
1147 
1148 	ent = section_entry(domain->pgtable, iova);
1149 
1150 	if (lv1ent_section(ent)) {
1151 		if (WARN_ON(size < SECT_SIZE)) {
1152 			err_pgsize = SECT_SIZE;
1153 			goto err;
1154 		}
1155 
1156 		/* workaround for h/w bug in System MMU v3.3 */
1157 		update_pte(ent, ZERO_LV2LINK);
1158 		size = SECT_SIZE;
1159 		goto done;
1160 	}
1161 
1162 	if (unlikely(lv1ent_fault(ent))) {
1163 		if (size > SECT_SIZE)
1164 			size = SECT_SIZE;
1165 		goto done;
1166 	}
1167 
1168 	/* lv1ent_page(sent) == true here */
1169 
1170 	ent = page_entry(ent, iova);
1171 
1172 	if (unlikely(lv2ent_fault(ent))) {
1173 		size = SPAGE_SIZE;
1174 		goto done;
1175 	}
1176 
1177 	if (lv2ent_small(ent)) {
1178 		update_pte(ent, 0);
1179 		size = SPAGE_SIZE;
1180 		domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1181 		goto done;
1182 	}
1183 
1184 	/* lv1ent_large(ent) == true here */
1185 	if (WARN_ON(size < LPAGE_SIZE)) {
1186 		err_pgsize = LPAGE_SIZE;
1187 		goto err;
1188 	}
1189 
1190 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1191 				sizeof(*ent) * SPAGES_PER_LPAGE,
1192 				DMA_TO_DEVICE);
1193 	memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1194 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1195 				   sizeof(*ent) * SPAGES_PER_LPAGE,
1196 				   DMA_TO_DEVICE);
1197 	size = LPAGE_SIZE;
1198 	domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1199 done:
1200 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1201 
1202 	exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1203 
1204 	return size;
1205 err:
1206 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1207 
1208 	pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1209 		__func__, size, iova, err_pgsize);
1210 
1211 	return 0;
1212 }
1213 
1214 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1215 					  dma_addr_t iova)
1216 {
1217 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1218 	sysmmu_pte_t *entry;
1219 	unsigned long flags;
1220 	phys_addr_t phys = 0;
1221 
1222 	spin_lock_irqsave(&domain->pgtablelock, flags);
1223 
1224 	entry = section_entry(domain->pgtable, iova);
1225 
1226 	if (lv1ent_section(entry)) {
1227 		phys = section_phys(entry) + section_offs(iova);
1228 	} else if (lv1ent_page(entry)) {
1229 		entry = page_entry(entry, iova);
1230 
1231 		if (lv2ent_large(entry))
1232 			phys = lpage_phys(entry) + lpage_offs(iova);
1233 		else if (lv2ent_small(entry))
1234 			phys = spage_phys(entry) + spage_offs(iova);
1235 	}
1236 
1237 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1238 
1239 	return phys;
1240 }
1241 
1242 static int exynos_iommu_add_device(struct device *dev)
1243 {
1244 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1245 	struct sysmmu_drvdata *data;
1246 	struct iommu_group *group;
1247 
1248 	if (!has_sysmmu(dev))
1249 		return -ENODEV;
1250 
1251 	group = iommu_group_get_for_dev(dev);
1252 
1253 	if (IS_ERR(group))
1254 		return PTR_ERR(group);
1255 
1256 	list_for_each_entry(data, &owner->controllers, owner_node) {
1257 		/*
1258 		 * SYSMMU will be runtime activated via device link
1259 		 * (dependency) to its master device, so there are no
1260 		 * direct calls to pm_runtime_get/put in this driver.
1261 		 */
1262 		data->link = device_link_add(dev, data->sysmmu,
1263 					     DL_FLAG_PM_RUNTIME);
1264 	}
1265 	iommu_group_put(group);
1266 
1267 	return 0;
1268 }
1269 
1270 static void exynos_iommu_remove_device(struct device *dev)
1271 {
1272 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1273 	struct sysmmu_drvdata *data;
1274 
1275 	if (!has_sysmmu(dev))
1276 		return;
1277 
1278 	if (owner->domain) {
1279 		struct iommu_group *group = iommu_group_get(dev);
1280 
1281 		if (group) {
1282 			WARN_ON(owner->domain !=
1283 				iommu_group_default_domain(group));
1284 			exynos_iommu_detach_device(owner->domain, dev);
1285 			iommu_group_put(group);
1286 		}
1287 	}
1288 	iommu_group_remove_device(dev);
1289 
1290 	list_for_each_entry(data, &owner->controllers, owner_node)
1291 		device_link_del(data->link);
1292 }
1293 
1294 static int exynos_iommu_of_xlate(struct device *dev,
1295 				 struct of_phandle_args *spec)
1296 {
1297 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1298 	struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1299 	struct sysmmu_drvdata *data, *entry;
1300 
1301 	if (!sysmmu)
1302 		return -ENODEV;
1303 
1304 	data = platform_get_drvdata(sysmmu);
1305 	if (!data)
1306 		return -ENODEV;
1307 
1308 	if (!owner) {
1309 		owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1310 		if (!owner)
1311 			return -ENOMEM;
1312 
1313 		INIT_LIST_HEAD(&owner->controllers);
1314 		mutex_init(&owner->rpm_lock);
1315 		dev->archdata.iommu = owner;
1316 	}
1317 
1318 	list_for_each_entry(entry, &owner->controllers, owner_node)
1319 		if (entry == data)
1320 			return 0;
1321 
1322 	list_add_tail(&data->owner_node, &owner->controllers);
1323 	data->master = dev;
1324 
1325 	return 0;
1326 }
1327 
1328 static const struct iommu_ops exynos_iommu_ops = {
1329 	.domain_alloc = exynos_iommu_domain_alloc,
1330 	.domain_free = exynos_iommu_domain_free,
1331 	.attach_dev = exynos_iommu_attach_device,
1332 	.detach_dev = exynos_iommu_detach_device,
1333 	.map = exynos_iommu_map,
1334 	.unmap = exynos_iommu_unmap,
1335 	.iova_to_phys = exynos_iommu_iova_to_phys,
1336 	.device_group = generic_device_group,
1337 	.add_device = exynos_iommu_add_device,
1338 	.remove_device = exynos_iommu_remove_device,
1339 	.pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1340 	.of_xlate = exynos_iommu_of_xlate,
1341 };
1342 
1343 static int __init exynos_iommu_init(void)
1344 {
1345 	struct device_node *np;
1346 	int ret;
1347 
1348 	np = of_find_matching_node(NULL, sysmmu_of_match);
1349 	if (!np)
1350 		return 0;
1351 
1352 	of_node_put(np);
1353 
1354 	lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1355 				LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1356 	if (!lv2table_kmem_cache) {
1357 		pr_err("%s: Failed to create kmem cache\n", __func__);
1358 		return -ENOMEM;
1359 	}
1360 
1361 	ret = platform_driver_register(&exynos_sysmmu_driver);
1362 	if (ret) {
1363 		pr_err("%s: Failed to register driver\n", __func__);
1364 		goto err_reg_driver;
1365 	}
1366 
1367 	zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1368 	if (zero_lv2_table == NULL) {
1369 		pr_err("%s: Failed to allocate zero level2 page table\n",
1370 			__func__);
1371 		ret = -ENOMEM;
1372 		goto err_zero_lv2;
1373 	}
1374 
1375 	ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1376 	if (ret) {
1377 		pr_err("%s: Failed to register exynos-iommu driver.\n",
1378 								__func__);
1379 		goto err_set_iommu;
1380 	}
1381 
1382 	return 0;
1383 err_set_iommu:
1384 	kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1385 err_zero_lv2:
1386 	platform_driver_unregister(&exynos_sysmmu_driver);
1387 err_reg_driver:
1388 	kmem_cache_destroy(lv2table_kmem_cache);
1389 	return ret;
1390 }
1391 core_initcall(exynos_iommu_init);
1392