xref: /openbmc/linux/drivers/iommu/exynos-iommu.c (revision 3c8c1539)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (c) 2011,2016 Samsung Electronics Co., Ltd.
4  *		http://www.samsung.com
5  */
6 
7 #ifdef CONFIG_EXYNOS_IOMMU_DEBUG
8 #define DEBUG
9 #endif
10 
11 #include <linux/clk.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/err.h>
14 #include <linux/io.h>
15 #include <linux/iommu.h>
16 #include <linux/interrupt.h>
17 #include <linux/kmemleak.h>
18 #include <linux/list.h>
19 #include <linux/of.h>
20 #include <linux/of_platform.h>
21 #include <linux/platform_device.h>
22 #include <linux/pm_runtime.h>
23 #include <linux/slab.h>
24 
25 typedef u32 sysmmu_iova_t;
26 typedef u32 sysmmu_pte_t;
27 
28 /* We do not consider super section mapping (16MB) */
29 #define SECT_ORDER 20
30 #define LPAGE_ORDER 16
31 #define SPAGE_ORDER 12
32 
33 #define SECT_SIZE (1 << SECT_ORDER)
34 #define LPAGE_SIZE (1 << LPAGE_ORDER)
35 #define SPAGE_SIZE (1 << SPAGE_ORDER)
36 
37 #define SECT_MASK (~(SECT_SIZE - 1))
38 #define LPAGE_MASK (~(LPAGE_SIZE - 1))
39 #define SPAGE_MASK (~(SPAGE_SIZE - 1))
40 
41 #define lv1ent_fault(sent) ((*(sent) == ZERO_LV2LINK) || \
42 			   ((*(sent) & 3) == 0) || ((*(sent) & 3) == 3))
43 #define lv1ent_zero(sent) (*(sent) == ZERO_LV2LINK)
44 #define lv1ent_page_zero(sent) ((*(sent) & 3) == 1)
45 #define lv1ent_page(sent) ((*(sent) != ZERO_LV2LINK) && \
46 			  ((*(sent) & 3) == 1))
47 #define lv1ent_section(sent) ((*(sent) & 3) == 2)
48 
49 #define lv2ent_fault(pent) ((*(pent) & 3) == 0)
50 #define lv2ent_small(pent) ((*(pent) & 2) == 2)
51 #define lv2ent_large(pent) ((*(pent) & 3) == 1)
52 
53 /*
54  * v1.x - v3.x SYSMMU supports 32bit physical and 32bit virtual address spaces
55  * v5.0 introduced support for 36bit physical address space by shifting
56  * all page entry values by 4 bits.
57  * All SYSMMU controllers in the system support the address spaces of the same
58  * size, so PG_ENT_SHIFT can be initialized on first SYSMMU probe to proper
59  * value (0 or 4).
60  */
61 static short PG_ENT_SHIFT = -1;
62 #define SYSMMU_PG_ENT_SHIFT 0
63 #define SYSMMU_V5_PG_ENT_SHIFT 4
64 
65 static const sysmmu_pte_t *LV1_PROT;
66 static const sysmmu_pte_t SYSMMU_LV1_PROT[] = {
67 	((0 << 15) | (0 << 10)), /* no access */
68 	((1 << 15) | (1 << 10)), /* IOMMU_READ only */
69 	((0 << 15) | (1 << 10)), /* IOMMU_WRITE not supported, use read/write */
70 	((0 << 15) | (1 << 10)), /* IOMMU_READ | IOMMU_WRITE */
71 };
72 static const sysmmu_pte_t SYSMMU_V5_LV1_PROT[] = {
73 	(0 << 4), /* no access */
74 	(1 << 4), /* IOMMU_READ only */
75 	(2 << 4), /* IOMMU_WRITE only */
76 	(3 << 4), /* IOMMU_READ | IOMMU_WRITE */
77 };
78 
79 static const sysmmu_pte_t *LV2_PROT;
80 static const sysmmu_pte_t SYSMMU_LV2_PROT[] = {
81 	((0 << 9) | (0 << 4)), /* no access */
82 	((1 << 9) | (1 << 4)), /* IOMMU_READ only */
83 	((0 << 9) | (1 << 4)), /* IOMMU_WRITE not supported, use read/write */
84 	((0 << 9) | (1 << 4)), /* IOMMU_READ | IOMMU_WRITE */
85 };
86 static const sysmmu_pte_t SYSMMU_V5_LV2_PROT[] = {
87 	(0 << 2), /* no access */
88 	(1 << 2), /* IOMMU_READ only */
89 	(2 << 2), /* IOMMU_WRITE only */
90 	(3 << 2), /* IOMMU_READ | IOMMU_WRITE */
91 };
92 
93 #define SYSMMU_SUPPORTED_PROT_BITS (IOMMU_READ | IOMMU_WRITE)
94 
95 #define sect_to_phys(ent) (((phys_addr_t) ent) << PG_ENT_SHIFT)
96 #define section_phys(sent) (sect_to_phys(*(sent)) & SECT_MASK)
97 #define section_offs(iova) (iova & (SECT_SIZE - 1))
98 #define lpage_phys(pent) (sect_to_phys(*(pent)) & LPAGE_MASK)
99 #define lpage_offs(iova) (iova & (LPAGE_SIZE - 1))
100 #define spage_phys(pent) (sect_to_phys(*(pent)) & SPAGE_MASK)
101 #define spage_offs(iova) (iova & (SPAGE_SIZE - 1))
102 
103 #define NUM_LV1ENTRIES 4096
104 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
105 
106 static u32 lv1ent_offset(sysmmu_iova_t iova)
107 {
108 	return iova >> SECT_ORDER;
109 }
110 
111 static u32 lv2ent_offset(sysmmu_iova_t iova)
112 {
113 	return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
114 }
115 
116 #define LV1TABLE_SIZE (NUM_LV1ENTRIES * sizeof(sysmmu_pte_t))
117 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
118 
119 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
120 #define lv2table_base(sent) (sect_to_phys(*(sent) & 0xFFFFFFC0))
121 
122 #define mk_lv1ent_sect(pa, prot) ((pa >> PG_ENT_SHIFT) | LV1_PROT[prot] | 2)
123 #define mk_lv1ent_page(pa) ((pa >> PG_ENT_SHIFT) | 1)
124 #define mk_lv2ent_lpage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 1)
125 #define mk_lv2ent_spage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 2)
126 
127 #define CTRL_ENABLE	0x5
128 #define CTRL_BLOCK	0x7
129 #define CTRL_DISABLE	0x0
130 
131 #define CFG_LRU		0x1
132 #define CFG_EAP		(1 << 2)
133 #define CFG_QOS(n)	((n & 0xF) << 7)
134 #define CFG_ACGEN	(1 << 24) /* System MMU 3.3 only */
135 #define CFG_SYSSEL	(1 << 22) /* System MMU 3.2 only */
136 #define CFG_FLPDCACHE	(1 << 20) /* System MMU 3.2+ only */
137 
138 /* common registers */
139 #define REG_MMU_CTRL		0x000
140 #define REG_MMU_CFG		0x004
141 #define REG_MMU_STATUS		0x008
142 #define REG_MMU_VERSION		0x034
143 
144 #define MMU_MAJ_VER(val)	((val) >> 7)
145 #define MMU_MIN_VER(val)	((val) & 0x7F)
146 #define MMU_RAW_VER(reg)	(((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
147 
148 #define MAKE_MMU_VER(maj, min)	((((maj) & 0xF) << 7) | ((min) & 0x7F))
149 
150 /* v1.x - v3.x registers */
151 #define REG_MMU_FLUSH		0x00C
152 #define REG_MMU_FLUSH_ENTRY	0x010
153 #define REG_PT_BASE_ADDR	0x014
154 #define REG_INT_STATUS		0x018
155 #define REG_INT_CLEAR		0x01C
156 
157 #define REG_PAGE_FAULT_ADDR	0x024
158 #define REG_AW_FAULT_ADDR	0x028
159 #define REG_AR_FAULT_ADDR	0x02C
160 #define REG_DEFAULT_SLAVE_ADDR	0x030
161 
162 /* v5.x registers */
163 #define REG_V5_PT_BASE_PFN	0x00C
164 #define REG_V5_MMU_FLUSH_ALL	0x010
165 #define REG_V5_MMU_FLUSH_ENTRY	0x014
166 #define REG_V5_MMU_FLUSH_RANGE	0x018
167 #define REG_V5_MMU_FLUSH_START	0x020
168 #define REG_V5_MMU_FLUSH_END	0x024
169 #define REG_V5_INT_STATUS	0x060
170 #define REG_V5_INT_CLEAR	0x064
171 #define REG_V5_FAULT_AR_VA	0x070
172 #define REG_V5_FAULT_AW_VA	0x080
173 
174 #define has_sysmmu(dev)		(dev_iommu_priv_get(dev) != NULL)
175 
176 static struct device *dma_dev;
177 static struct kmem_cache *lv2table_kmem_cache;
178 static sysmmu_pte_t *zero_lv2_table;
179 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
180 
181 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
182 {
183 	return pgtable + lv1ent_offset(iova);
184 }
185 
186 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
187 {
188 	return (sysmmu_pte_t *)phys_to_virt(
189 				lv2table_base(sent)) + lv2ent_offset(iova);
190 }
191 
192 /*
193  * IOMMU fault information register
194  */
195 struct sysmmu_fault_info {
196 	unsigned int bit;	/* bit number in STATUS register */
197 	unsigned short addr_reg; /* register to read VA fault address */
198 	const char *name;	/* human readable fault name */
199 	unsigned int type;	/* fault type for report_iommu_fault */
200 };
201 
202 static const struct sysmmu_fault_info sysmmu_faults[] = {
203 	{ 0, REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
204 	{ 1, REG_AR_FAULT_ADDR, "AR MULTI-HIT", IOMMU_FAULT_READ },
205 	{ 2, REG_AW_FAULT_ADDR, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
206 	{ 3, REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
207 	{ 4, REG_AR_FAULT_ADDR, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
208 	{ 5, REG_AR_FAULT_ADDR, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
209 	{ 6, REG_AW_FAULT_ADDR, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
210 	{ 7, REG_AW_FAULT_ADDR, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
211 };
212 
213 static const struct sysmmu_fault_info sysmmu_v5_faults[] = {
214 	{ 0, REG_V5_FAULT_AR_VA, "AR PTW", IOMMU_FAULT_READ },
215 	{ 1, REG_V5_FAULT_AR_VA, "AR PAGE", IOMMU_FAULT_READ },
216 	{ 2, REG_V5_FAULT_AR_VA, "AR MULTI-HIT", IOMMU_FAULT_READ },
217 	{ 3, REG_V5_FAULT_AR_VA, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
218 	{ 4, REG_V5_FAULT_AR_VA, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
219 	{ 16, REG_V5_FAULT_AW_VA, "AW PTW", IOMMU_FAULT_WRITE },
220 	{ 17, REG_V5_FAULT_AW_VA, "AW PAGE", IOMMU_FAULT_WRITE },
221 	{ 18, REG_V5_FAULT_AW_VA, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
222 	{ 19, REG_V5_FAULT_AW_VA, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
223 	{ 20, REG_V5_FAULT_AW_VA, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
224 };
225 
226 /*
227  * This structure is attached to dev->iommu->priv of the master device
228  * on device add, contains a list of SYSMMU controllers defined by device tree,
229  * which are bound to given master device. It is usually referenced by 'owner'
230  * pointer.
231 */
232 struct exynos_iommu_owner {
233 	struct list_head controllers;	/* list of sysmmu_drvdata.owner_node */
234 	struct iommu_domain *domain;	/* domain this device is attached */
235 	struct mutex rpm_lock;		/* for runtime pm of all sysmmus */
236 };
237 
238 /*
239  * This structure exynos specific generalization of struct iommu_domain.
240  * It contains list of SYSMMU controllers from all master devices, which has
241  * been attached to this domain and page tables of IO address space defined by
242  * it. It is usually referenced by 'domain' pointer.
243  */
244 struct exynos_iommu_domain {
245 	struct list_head clients; /* list of sysmmu_drvdata.domain_node */
246 	sysmmu_pte_t *pgtable;	/* lv1 page table, 16KB */
247 	short *lv2entcnt;	/* free lv2 entry counter for each section */
248 	spinlock_t lock;	/* lock for modyfying list of clients */
249 	spinlock_t pgtablelock;	/* lock for modifying page table @ pgtable */
250 	struct iommu_domain domain; /* generic domain data structure */
251 };
252 
253 /*
254  * This structure hold all data of a single SYSMMU controller, this includes
255  * hw resources like registers and clocks, pointers and list nodes to connect
256  * it to all other structures, internal state and parameters read from device
257  * tree. It is usually referenced by 'data' pointer.
258  */
259 struct sysmmu_drvdata {
260 	struct device *sysmmu;		/* SYSMMU controller device */
261 	struct device *master;		/* master device (owner) */
262 	struct device_link *link;	/* runtime PM link to master */
263 	void __iomem *sfrbase;		/* our registers */
264 	struct clk *clk;		/* SYSMMU's clock */
265 	struct clk *aclk;		/* SYSMMU's aclk clock */
266 	struct clk *pclk;		/* SYSMMU's pclk clock */
267 	struct clk *clk_master;		/* master's device clock */
268 	spinlock_t lock;		/* lock for modyfying state */
269 	bool active;			/* current status */
270 	struct exynos_iommu_domain *domain; /* domain we belong to */
271 	struct list_head domain_node;	/* node for domain clients list */
272 	struct list_head owner_node;	/* node for owner controllers list */
273 	phys_addr_t pgtable;		/* assigned page table structure */
274 	unsigned int version;		/* our version */
275 
276 	struct iommu_device iommu;	/* IOMMU core handle */
277 };
278 
279 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
280 {
281 	return container_of(dom, struct exynos_iommu_domain, domain);
282 }
283 
284 static void sysmmu_unblock(struct sysmmu_drvdata *data)
285 {
286 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
287 }
288 
289 static bool sysmmu_block(struct sysmmu_drvdata *data)
290 {
291 	int i = 120;
292 
293 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
294 	while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
295 		--i;
296 
297 	if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
298 		sysmmu_unblock(data);
299 		return false;
300 	}
301 
302 	return true;
303 }
304 
305 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
306 {
307 	if (MMU_MAJ_VER(data->version) < 5)
308 		writel(0x1, data->sfrbase + REG_MMU_FLUSH);
309 	else
310 		writel(0x1, data->sfrbase + REG_V5_MMU_FLUSH_ALL);
311 }
312 
313 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
314 				sysmmu_iova_t iova, unsigned int num_inv)
315 {
316 	unsigned int i;
317 
318 	if (MMU_MAJ_VER(data->version) < 5) {
319 		for (i = 0; i < num_inv; i++) {
320 			writel((iova & SPAGE_MASK) | 1,
321 				     data->sfrbase + REG_MMU_FLUSH_ENTRY);
322 			iova += SPAGE_SIZE;
323 		}
324 	} else {
325 		if (num_inv == 1) {
326 			writel((iova & SPAGE_MASK) | 1,
327 				     data->sfrbase + REG_V5_MMU_FLUSH_ENTRY);
328 		} else {
329 			writel((iova & SPAGE_MASK),
330 				     data->sfrbase + REG_V5_MMU_FLUSH_START);
331 			writel((iova & SPAGE_MASK) + (num_inv - 1) * SPAGE_SIZE,
332 				     data->sfrbase + REG_V5_MMU_FLUSH_END);
333 			writel(1, data->sfrbase + REG_V5_MMU_FLUSH_RANGE);
334 		}
335 	}
336 }
337 
338 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
339 {
340 	if (MMU_MAJ_VER(data->version) < 5)
341 		writel(pgd, data->sfrbase + REG_PT_BASE_ADDR);
342 	else
343 		writel(pgd >> PAGE_SHIFT,
344 			     data->sfrbase + REG_V5_PT_BASE_PFN);
345 
346 	__sysmmu_tlb_invalidate(data);
347 }
348 
349 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
350 {
351 	BUG_ON(clk_prepare_enable(data->clk_master));
352 	BUG_ON(clk_prepare_enable(data->clk));
353 	BUG_ON(clk_prepare_enable(data->pclk));
354 	BUG_ON(clk_prepare_enable(data->aclk));
355 }
356 
357 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
358 {
359 	clk_disable_unprepare(data->aclk);
360 	clk_disable_unprepare(data->pclk);
361 	clk_disable_unprepare(data->clk);
362 	clk_disable_unprepare(data->clk_master);
363 }
364 
365 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
366 {
367 	u32 ver;
368 
369 	__sysmmu_enable_clocks(data);
370 
371 	ver = readl(data->sfrbase + REG_MMU_VERSION);
372 
373 	/* controllers on some SoCs don't report proper version */
374 	if (ver == 0x80000001u)
375 		data->version = MAKE_MMU_VER(1, 0);
376 	else
377 		data->version = MMU_RAW_VER(ver);
378 
379 	dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
380 		MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
381 
382 	__sysmmu_disable_clocks(data);
383 }
384 
385 static void show_fault_information(struct sysmmu_drvdata *data,
386 				   const struct sysmmu_fault_info *finfo,
387 				   sysmmu_iova_t fault_addr)
388 {
389 	sysmmu_pte_t *ent;
390 
391 	dev_err(data->sysmmu, "%s: %s FAULT occurred at %#x\n",
392 		dev_name(data->master), finfo->name, fault_addr);
393 	dev_dbg(data->sysmmu, "Page table base: %pa\n", &data->pgtable);
394 	ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
395 	dev_dbg(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
396 	if (lv1ent_page(ent)) {
397 		ent = page_entry(ent, fault_addr);
398 		dev_dbg(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
399 	}
400 }
401 
402 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
403 {
404 	/* SYSMMU is in blocked state when interrupt occurred. */
405 	struct sysmmu_drvdata *data = dev_id;
406 	const struct sysmmu_fault_info *finfo;
407 	unsigned int i, n, itype;
408 	sysmmu_iova_t fault_addr;
409 	unsigned short reg_status, reg_clear;
410 	int ret = -ENOSYS;
411 
412 	WARN_ON(!data->active);
413 
414 	if (MMU_MAJ_VER(data->version) < 5) {
415 		reg_status = REG_INT_STATUS;
416 		reg_clear = REG_INT_CLEAR;
417 		finfo = sysmmu_faults;
418 		n = ARRAY_SIZE(sysmmu_faults);
419 	} else {
420 		reg_status = REG_V5_INT_STATUS;
421 		reg_clear = REG_V5_INT_CLEAR;
422 		finfo = sysmmu_v5_faults;
423 		n = ARRAY_SIZE(sysmmu_v5_faults);
424 	}
425 
426 	spin_lock(&data->lock);
427 
428 	clk_enable(data->clk_master);
429 
430 	itype = __ffs(readl(data->sfrbase + reg_status));
431 	for (i = 0; i < n; i++, finfo++)
432 		if (finfo->bit == itype)
433 			break;
434 	/* unknown/unsupported fault */
435 	BUG_ON(i == n);
436 
437 	/* print debug message */
438 	fault_addr = readl(data->sfrbase + finfo->addr_reg);
439 	show_fault_information(data, finfo, fault_addr);
440 
441 	if (data->domain)
442 		ret = report_iommu_fault(&data->domain->domain,
443 					data->master, fault_addr, finfo->type);
444 	/* fault is not recovered by fault handler */
445 	BUG_ON(ret != 0);
446 
447 	writel(1 << itype, data->sfrbase + reg_clear);
448 
449 	sysmmu_unblock(data);
450 
451 	clk_disable(data->clk_master);
452 
453 	spin_unlock(&data->lock);
454 
455 	return IRQ_HANDLED;
456 }
457 
458 static void __sysmmu_disable(struct sysmmu_drvdata *data)
459 {
460 	unsigned long flags;
461 
462 	clk_enable(data->clk_master);
463 
464 	spin_lock_irqsave(&data->lock, flags);
465 	writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
466 	writel(0, data->sfrbase + REG_MMU_CFG);
467 	data->active = false;
468 	spin_unlock_irqrestore(&data->lock, flags);
469 
470 	__sysmmu_disable_clocks(data);
471 }
472 
473 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
474 {
475 	unsigned int cfg;
476 
477 	if (data->version <= MAKE_MMU_VER(3, 1))
478 		cfg = CFG_LRU | CFG_QOS(15);
479 	else if (data->version <= MAKE_MMU_VER(3, 2))
480 		cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
481 	else
482 		cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
483 
484 	cfg |= CFG_EAP; /* enable access protection bits check */
485 
486 	writel(cfg, data->sfrbase + REG_MMU_CFG);
487 }
488 
489 static void __sysmmu_enable(struct sysmmu_drvdata *data)
490 {
491 	unsigned long flags;
492 
493 	__sysmmu_enable_clocks(data);
494 
495 	spin_lock_irqsave(&data->lock, flags);
496 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
497 	__sysmmu_init_config(data);
498 	__sysmmu_set_ptbase(data, data->pgtable);
499 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
500 	data->active = true;
501 	spin_unlock_irqrestore(&data->lock, flags);
502 
503 	/*
504 	 * SYSMMU driver keeps master's clock enabled only for the short
505 	 * time, while accessing the registers. For performing address
506 	 * translation during DMA transaction it relies on the client
507 	 * driver to enable it.
508 	 */
509 	clk_disable(data->clk_master);
510 }
511 
512 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
513 					    sysmmu_iova_t iova)
514 {
515 	unsigned long flags;
516 
517 	spin_lock_irqsave(&data->lock, flags);
518 	if (data->active && data->version >= MAKE_MMU_VER(3, 3)) {
519 		clk_enable(data->clk_master);
520 		if (sysmmu_block(data)) {
521 			if (data->version >= MAKE_MMU_VER(5, 0))
522 				__sysmmu_tlb_invalidate(data);
523 			else
524 				__sysmmu_tlb_invalidate_entry(data, iova, 1);
525 			sysmmu_unblock(data);
526 		}
527 		clk_disable(data->clk_master);
528 	}
529 	spin_unlock_irqrestore(&data->lock, flags);
530 }
531 
532 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
533 					sysmmu_iova_t iova, size_t size)
534 {
535 	unsigned long flags;
536 
537 	spin_lock_irqsave(&data->lock, flags);
538 	if (data->active) {
539 		unsigned int num_inv = 1;
540 
541 		clk_enable(data->clk_master);
542 
543 		/*
544 		 * L2TLB invalidation required
545 		 * 4KB page: 1 invalidation
546 		 * 64KB page: 16 invalidations
547 		 * 1MB page: 64 invalidations
548 		 * because it is set-associative TLB
549 		 * with 8-way and 64 sets.
550 		 * 1MB page can be cached in one of all sets.
551 		 * 64KB page can be one of 16 consecutive sets.
552 		 */
553 		if (MMU_MAJ_VER(data->version) == 2)
554 			num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
555 
556 		if (sysmmu_block(data)) {
557 			__sysmmu_tlb_invalidate_entry(data, iova, num_inv);
558 			sysmmu_unblock(data);
559 		}
560 		clk_disable(data->clk_master);
561 	}
562 	spin_unlock_irqrestore(&data->lock, flags);
563 }
564 
565 static const struct iommu_ops exynos_iommu_ops;
566 
567 static int exynos_sysmmu_probe(struct platform_device *pdev)
568 {
569 	int irq, ret;
570 	struct device *dev = &pdev->dev;
571 	struct sysmmu_drvdata *data;
572 	struct resource *res;
573 
574 	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
575 	if (!data)
576 		return -ENOMEM;
577 
578 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
579 	data->sfrbase = devm_ioremap_resource(dev, res);
580 	if (IS_ERR(data->sfrbase))
581 		return PTR_ERR(data->sfrbase);
582 
583 	irq = platform_get_irq(pdev, 0);
584 	if (irq <= 0)
585 		return irq;
586 
587 	ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
588 				dev_name(dev), data);
589 	if (ret) {
590 		dev_err(dev, "Unabled to register handler of irq %d\n", irq);
591 		return ret;
592 	}
593 
594 	data->clk = devm_clk_get(dev, "sysmmu");
595 	if (PTR_ERR(data->clk) == -ENOENT)
596 		data->clk = NULL;
597 	else if (IS_ERR(data->clk))
598 		return PTR_ERR(data->clk);
599 
600 	data->aclk = devm_clk_get(dev, "aclk");
601 	if (PTR_ERR(data->aclk) == -ENOENT)
602 		data->aclk = NULL;
603 	else if (IS_ERR(data->aclk))
604 		return PTR_ERR(data->aclk);
605 
606 	data->pclk = devm_clk_get(dev, "pclk");
607 	if (PTR_ERR(data->pclk) == -ENOENT)
608 		data->pclk = NULL;
609 	else if (IS_ERR(data->pclk))
610 		return PTR_ERR(data->pclk);
611 
612 	if (!data->clk && (!data->aclk || !data->pclk)) {
613 		dev_err(dev, "Failed to get device clock(s)!\n");
614 		return -ENOSYS;
615 	}
616 
617 	data->clk_master = devm_clk_get(dev, "master");
618 	if (PTR_ERR(data->clk_master) == -ENOENT)
619 		data->clk_master = NULL;
620 	else if (IS_ERR(data->clk_master))
621 		return PTR_ERR(data->clk_master);
622 
623 	data->sysmmu = dev;
624 	spin_lock_init(&data->lock);
625 
626 	ret = iommu_device_sysfs_add(&data->iommu, &pdev->dev, NULL,
627 				     dev_name(data->sysmmu));
628 	if (ret)
629 		return ret;
630 
631 	ret = iommu_device_register(&data->iommu, &exynos_iommu_ops, dev);
632 	if (ret)
633 		return ret;
634 
635 	platform_set_drvdata(pdev, data);
636 
637 	__sysmmu_get_version(data);
638 	if (PG_ENT_SHIFT < 0) {
639 		if (MMU_MAJ_VER(data->version) < 5) {
640 			PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
641 			LV1_PROT = SYSMMU_LV1_PROT;
642 			LV2_PROT = SYSMMU_LV2_PROT;
643 		} else {
644 			PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
645 			LV1_PROT = SYSMMU_V5_LV1_PROT;
646 			LV2_PROT = SYSMMU_V5_LV2_PROT;
647 		}
648 	}
649 
650 	/*
651 	 * use the first registered sysmmu device for performing
652 	 * dma mapping operations on iommu page tables (cpu cache flush)
653 	 */
654 	if (!dma_dev)
655 		dma_dev = &pdev->dev;
656 
657 	pm_runtime_enable(dev);
658 
659 	return 0;
660 }
661 
662 static int __maybe_unused exynos_sysmmu_suspend(struct device *dev)
663 {
664 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
665 	struct device *master = data->master;
666 
667 	if (master) {
668 		struct exynos_iommu_owner *owner = dev_iommu_priv_get(master);
669 
670 		mutex_lock(&owner->rpm_lock);
671 		if (data->domain) {
672 			dev_dbg(data->sysmmu, "saving state\n");
673 			__sysmmu_disable(data);
674 		}
675 		mutex_unlock(&owner->rpm_lock);
676 	}
677 	return 0;
678 }
679 
680 static int __maybe_unused exynos_sysmmu_resume(struct device *dev)
681 {
682 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
683 	struct device *master = data->master;
684 
685 	if (master) {
686 		struct exynos_iommu_owner *owner = dev_iommu_priv_get(master);
687 
688 		mutex_lock(&owner->rpm_lock);
689 		if (data->domain) {
690 			dev_dbg(data->sysmmu, "restoring state\n");
691 			__sysmmu_enable(data);
692 		}
693 		mutex_unlock(&owner->rpm_lock);
694 	}
695 	return 0;
696 }
697 
698 static const struct dev_pm_ops sysmmu_pm_ops = {
699 	SET_RUNTIME_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume, NULL)
700 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
701 				pm_runtime_force_resume)
702 };
703 
704 static const struct of_device_id sysmmu_of_match[] = {
705 	{ .compatible	= "samsung,exynos-sysmmu", },
706 	{ },
707 };
708 
709 static struct platform_driver exynos_sysmmu_driver __refdata = {
710 	.probe	= exynos_sysmmu_probe,
711 	.driver	= {
712 		.name		= "exynos-sysmmu",
713 		.of_match_table	= sysmmu_of_match,
714 		.pm		= &sysmmu_pm_ops,
715 		.suppress_bind_attrs = true,
716 	}
717 };
718 
719 static inline void exynos_iommu_set_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
720 {
721 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
722 				DMA_TO_DEVICE);
723 	*ent = cpu_to_le32(val);
724 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
725 				   DMA_TO_DEVICE);
726 }
727 
728 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
729 {
730 	struct exynos_iommu_domain *domain;
731 	dma_addr_t handle;
732 	int i;
733 
734 	/* Check if correct PTE offsets are initialized */
735 	BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
736 
737 	if (type != IOMMU_DOMAIN_DMA && type != IOMMU_DOMAIN_UNMANAGED)
738 		return NULL;
739 
740 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
741 	if (!domain)
742 		return NULL;
743 
744 	domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
745 	if (!domain->pgtable)
746 		goto err_pgtable;
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++)
754 		domain->pgtable[i] = ZERO_LV2LINK;
755 
756 	handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
757 				DMA_TO_DEVICE);
758 	/* For mapping page table entries we rely on dma == phys */
759 	BUG_ON(handle != virt_to_phys(domain->pgtable));
760 	if (dma_mapping_error(dma_dev, handle))
761 		goto err_lv2ent;
762 
763 	spin_lock_init(&domain->lock);
764 	spin_lock_init(&domain->pgtablelock);
765 	INIT_LIST_HEAD(&domain->clients);
766 
767 	domain->domain.geometry.aperture_start = 0;
768 	domain->domain.geometry.aperture_end   = ~0UL;
769 	domain->domain.geometry.force_aperture = true;
770 
771 	return &domain->domain;
772 
773 err_lv2ent:
774 	free_pages((unsigned long)domain->lv2entcnt, 1);
775 err_counter:
776 	free_pages((unsigned long)domain->pgtable, 2);
777 err_pgtable:
778 	kfree(domain);
779 	return NULL;
780 }
781 
782 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
783 {
784 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
785 	struct sysmmu_drvdata *data, *next;
786 	unsigned long flags;
787 	int i;
788 
789 	WARN_ON(!list_empty(&domain->clients));
790 
791 	spin_lock_irqsave(&domain->lock, flags);
792 
793 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
794 		spin_lock(&data->lock);
795 		__sysmmu_disable(data);
796 		data->pgtable = 0;
797 		data->domain = NULL;
798 		list_del_init(&data->domain_node);
799 		spin_unlock(&data->lock);
800 	}
801 
802 	spin_unlock_irqrestore(&domain->lock, flags);
803 
804 	dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
805 			 DMA_TO_DEVICE);
806 
807 	for (i = 0; i < NUM_LV1ENTRIES; i++)
808 		if (lv1ent_page(domain->pgtable + i)) {
809 			phys_addr_t base = lv2table_base(domain->pgtable + i);
810 
811 			dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
812 					 DMA_TO_DEVICE);
813 			kmem_cache_free(lv2table_kmem_cache,
814 					phys_to_virt(base));
815 		}
816 
817 	free_pages((unsigned long)domain->pgtable, 2);
818 	free_pages((unsigned long)domain->lv2entcnt, 1);
819 	kfree(domain);
820 }
821 
822 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
823 				    struct device *dev)
824 {
825 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
826 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
827 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
828 	struct sysmmu_drvdata *data, *next;
829 	unsigned long flags;
830 
831 	if (!has_sysmmu(dev) || owner->domain != iommu_domain)
832 		return;
833 
834 	mutex_lock(&owner->rpm_lock);
835 
836 	list_for_each_entry(data, &owner->controllers, owner_node) {
837 		pm_runtime_get_noresume(data->sysmmu);
838 		if (pm_runtime_active(data->sysmmu))
839 			__sysmmu_disable(data);
840 		pm_runtime_put(data->sysmmu);
841 	}
842 
843 	spin_lock_irqsave(&domain->lock, flags);
844 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
845 		spin_lock(&data->lock);
846 		data->pgtable = 0;
847 		data->domain = NULL;
848 		list_del_init(&data->domain_node);
849 		spin_unlock(&data->lock);
850 	}
851 	owner->domain = NULL;
852 	spin_unlock_irqrestore(&domain->lock, flags);
853 
854 	mutex_unlock(&owner->rpm_lock);
855 
856 	dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n", __func__,
857 		&pagetable);
858 }
859 
860 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
861 				   struct device *dev)
862 {
863 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
864 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
865 	struct sysmmu_drvdata *data;
866 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
867 	unsigned long flags;
868 
869 	if (!has_sysmmu(dev))
870 		return -ENODEV;
871 
872 	if (owner->domain)
873 		exynos_iommu_detach_device(owner->domain, dev);
874 
875 	mutex_lock(&owner->rpm_lock);
876 
877 	spin_lock_irqsave(&domain->lock, flags);
878 	list_for_each_entry(data, &owner->controllers, owner_node) {
879 		spin_lock(&data->lock);
880 		data->pgtable = pagetable;
881 		data->domain = domain;
882 		list_add_tail(&data->domain_node, &domain->clients);
883 		spin_unlock(&data->lock);
884 	}
885 	owner->domain = iommu_domain;
886 	spin_unlock_irqrestore(&domain->lock, flags);
887 
888 	list_for_each_entry(data, &owner->controllers, owner_node) {
889 		pm_runtime_get_noresume(data->sysmmu);
890 		if (pm_runtime_active(data->sysmmu))
891 			__sysmmu_enable(data);
892 		pm_runtime_put(data->sysmmu);
893 	}
894 
895 	mutex_unlock(&owner->rpm_lock);
896 
897 	dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa\n", __func__,
898 		&pagetable);
899 
900 	return 0;
901 }
902 
903 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
904 		sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
905 {
906 	if (lv1ent_section(sent)) {
907 		WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
908 		return ERR_PTR(-EADDRINUSE);
909 	}
910 
911 	if (lv1ent_fault(sent)) {
912 		dma_addr_t handle;
913 		sysmmu_pte_t *pent;
914 		bool need_flush_flpd_cache = lv1ent_zero(sent);
915 
916 		pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
917 		BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
918 		if (!pent)
919 			return ERR_PTR(-ENOMEM);
920 
921 		exynos_iommu_set_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
922 		kmemleak_ignore(pent);
923 		*pgcounter = NUM_LV2ENTRIES;
924 		handle = dma_map_single(dma_dev, pent, LV2TABLE_SIZE,
925 					DMA_TO_DEVICE);
926 		if (dma_mapping_error(dma_dev, handle)) {
927 			kmem_cache_free(lv2table_kmem_cache, pent);
928 			return ERR_PTR(-EADDRINUSE);
929 		}
930 
931 		/*
932 		 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
933 		 * FLPD cache may cache the address of zero_l2_table. This
934 		 * function replaces the zero_l2_table with new L2 page table
935 		 * to write valid mappings.
936 		 * Accessing the valid area may cause page fault since FLPD
937 		 * cache may still cache zero_l2_table for the valid area
938 		 * instead of new L2 page table that has the mapping
939 		 * information of the valid area.
940 		 * Thus any replacement of zero_l2_table with other valid L2
941 		 * page table must involve FLPD cache invalidation for System
942 		 * MMU v3.3.
943 		 * FLPD cache invalidation is performed with TLB invalidation
944 		 * by VPN without blocking. It is safe to invalidate TLB without
945 		 * blocking because the target address of TLB invalidation is
946 		 * not currently mapped.
947 		 */
948 		if (need_flush_flpd_cache) {
949 			struct sysmmu_drvdata *data;
950 
951 			spin_lock(&domain->lock);
952 			list_for_each_entry(data, &domain->clients, domain_node)
953 				sysmmu_tlb_invalidate_flpdcache(data, iova);
954 			spin_unlock(&domain->lock);
955 		}
956 	}
957 
958 	return page_entry(sent, iova);
959 }
960 
961 static int lv1set_section(struct exynos_iommu_domain *domain,
962 			  sysmmu_pte_t *sent, sysmmu_iova_t iova,
963 			  phys_addr_t paddr, int prot, short *pgcnt)
964 {
965 	if (lv1ent_section(sent)) {
966 		WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
967 			iova);
968 		return -EADDRINUSE;
969 	}
970 
971 	if (lv1ent_page(sent)) {
972 		if (*pgcnt != NUM_LV2ENTRIES) {
973 			WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
974 				iova);
975 			return -EADDRINUSE;
976 		}
977 
978 		kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
979 		*pgcnt = 0;
980 	}
981 
982 	exynos_iommu_set_pte(sent, mk_lv1ent_sect(paddr, prot));
983 
984 	spin_lock(&domain->lock);
985 	if (lv1ent_page_zero(sent)) {
986 		struct sysmmu_drvdata *data;
987 		/*
988 		 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
989 		 * entry by speculative prefetch of SLPD which has no mapping.
990 		 */
991 		list_for_each_entry(data, &domain->clients, domain_node)
992 			sysmmu_tlb_invalidate_flpdcache(data, iova);
993 	}
994 	spin_unlock(&domain->lock);
995 
996 	return 0;
997 }
998 
999 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
1000 		       int prot, short *pgcnt)
1001 {
1002 	if (size == SPAGE_SIZE) {
1003 		if (WARN_ON(!lv2ent_fault(pent)))
1004 			return -EADDRINUSE;
1005 
1006 		exynos_iommu_set_pte(pent, mk_lv2ent_spage(paddr, prot));
1007 		*pgcnt -= 1;
1008 	} else { /* size == LPAGE_SIZE */
1009 		int i;
1010 		dma_addr_t pent_base = virt_to_phys(pent);
1011 
1012 		dma_sync_single_for_cpu(dma_dev, pent_base,
1013 					sizeof(*pent) * SPAGES_PER_LPAGE,
1014 					DMA_TO_DEVICE);
1015 		for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1016 			if (WARN_ON(!lv2ent_fault(pent))) {
1017 				if (i > 0)
1018 					memset(pent - i, 0, sizeof(*pent) * i);
1019 				return -EADDRINUSE;
1020 			}
1021 
1022 			*pent = mk_lv2ent_lpage(paddr, prot);
1023 		}
1024 		dma_sync_single_for_device(dma_dev, pent_base,
1025 					   sizeof(*pent) * SPAGES_PER_LPAGE,
1026 					   DMA_TO_DEVICE);
1027 		*pgcnt -= SPAGES_PER_LPAGE;
1028 	}
1029 
1030 	return 0;
1031 }
1032 
1033 /*
1034  * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1035  *
1036  * System MMU v3.x has advanced logic to improve address translation
1037  * performance with caching more page table entries by a page table walk.
1038  * However, the logic has a bug that while caching faulty page table entries,
1039  * System MMU reports page fault if the cached fault entry is hit even though
1040  * the fault entry is updated to a valid entry after the entry is cached.
1041  * To prevent caching faulty page table entries which may be updated to valid
1042  * entries later, the virtual memory manager should care about the workaround
1043  * for the problem. The following describes the workaround.
1044  *
1045  * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1046  * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1047  *
1048  * Precisely, any start address of I/O virtual region must be aligned with
1049  * the following sizes for System MMU v3.1 and v3.2.
1050  * System MMU v3.1: 128KiB
1051  * System MMU v3.2: 256KiB
1052  *
1053  * Because System MMU v3.3 caches page table entries more aggressively, it needs
1054  * more workarounds.
1055  * - Any two consecutive I/O virtual regions must have a hole of size larger
1056  *   than or equal to 128KiB.
1057  * - Start address of an I/O virtual region must be aligned by 128KiB.
1058  */
1059 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1060 			    unsigned long l_iova, phys_addr_t paddr, size_t size,
1061 			    int prot, gfp_t gfp)
1062 {
1063 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1064 	sysmmu_pte_t *entry;
1065 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1066 	unsigned long flags;
1067 	int ret = -ENOMEM;
1068 
1069 	BUG_ON(domain->pgtable == NULL);
1070 	prot &= SYSMMU_SUPPORTED_PROT_BITS;
1071 
1072 	spin_lock_irqsave(&domain->pgtablelock, flags);
1073 
1074 	entry = section_entry(domain->pgtable, iova);
1075 
1076 	if (size == SECT_SIZE) {
1077 		ret = lv1set_section(domain, entry, iova, paddr, prot,
1078 				     &domain->lv2entcnt[lv1ent_offset(iova)]);
1079 	} else {
1080 		sysmmu_pte_t *pent;
1081 
1082 		pent = alloc_lv2entry(domain, entry, iova,
1083 				      &domain->lv2entcnt[lv1ent_offset(iova)]);
1084 
1085 		if (IS_ERR(pent))
1086 			ret = PTR_ERR(pent);
1087 		else
1088 			ret = lv2set_page(pent, paddr, size, prot,
1089 				       &domain->lv2entcnt[lv1ent_offset(iova)]);
1090 	}
1091 
1092 	if (ret)
1093 		pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1094 			__func__, ret, size, iova);
1095 
1096 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1097 
1098 	return ret;
1099 }
1100 
1101 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1102 					      sysmmu_iova_t iova, size_t size)
1103 {
1104 	struct sysmmu_drvdata *data;
1105 	unsigned long flags;
1106 
1107 	spin_lock_irqsave(&domain->lock, flags);
1108 
1109 	list_for_each_entry(data, &domain->clients, domain_node)
1110 		sysmmu_tlb_invalidate_entry(data, iova, size);
1111 
1112 	spin_unlock_irqrestore(&domain->lock, flags);
1113 }
1114 
1115 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1116 				 unsigned long l_iova, size_t size,
1117 				 struct iommu_iotlb_gather *gather)
1118 {
1119 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1120 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1121 	sysmmu_pte_t *ent;
1122 	size_t err_pgsize;
1123 	unsigned long flags;
1124 
1125 	BUG_ON(domain->pgtable == NULL);
1126 
1127 	spin_lock_irqsave(&domain->pgtablelock, flags);
1128 
1129 	ent = section_entry(domain->pgtable, iova);
1130 
1131 	if (lv1ent_section(ent)) {
1132 		if (WARN_ON(size < SECT_SIZE)) {
1133 			err_pgsize = SECT_SIZE;
1134 			goto err;
1135 		}
1136 
1137 		/* workaround for h/w bug in System MMU v3.3 */
1138 		exynos_iommu_set_pte(ent, ZERO_LV2LINK);
1139 		size = SECT_SIZE;
1140 		goto done;
1141 	}
1142 
1143 	if (unlikely(lv1ent_fault(ent))) {
1144 		if (size > SECT_SIZE)
1145 			size = SECT_SIZE;
1146 		goto done;
1147 	}
1148 
1149 	/* lv1ent_page(sent) == true here */
1150 
1151 	ent = page_entry(ent, iova);
1152 
1153 	if (unlikely(lv2ent_fault(ent))) {
1154 		size = SPAGE_SIZE;
1155 		goto done;
1156 	}
1157 
1158 	if (lv2ent_small(ent)) {
1159 		exynos_iommu_set_pte(ent, 0);
1160 		size = SPAGE_SIZE;
1161 		domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1162 		goto done;
1163 	}
1164 
1165 	/* lv1ent_large(ent) == true here */
1166 	if (WARN_ON(size < LPAGE_SIZE)) {
1167 		err_pgsize = LPAGE_SIZE;
1168 		goto err;
1169 	}
1170 
1171 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1172 				sizeof(*ent) * SPAGES_PER_LPAGE,
1173 				DMA_TO_DEVICE);
1174 	memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1175 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1176 				   sizeof(*ent) * SPAGES_PER_LPAGE,
1177 				   DMA_TO_DEVICE);
1178 	size = LPAGE_SIZE;
1179 	domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1180 done:
1181 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1182 
1183 	exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1184 
1185 	return size;
1186 err:
1187 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1188 
1189 	pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1190 		__func__, size, iova, err_pgsize);
1191 
1192 	return 0;
1193 }
1194 
1195 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1196 					  dma_addr_t iova)
1197 {
1198 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1199 	sysmmu_pte_t *entry;
1200 	unsigned long flags;
1201 	phys_addr_t phys = 0;
1202 
1203 	spin_lock_irqsave(&domain->pgtablelock, flags);
1204 
1205 	entry = section_entry(domain->pgtable, iova);
1206 
1207 	if (lv1ent_section(entry)) {
1208 		phys = section_phys(entry) + section_offs(iova);
1209 	} else if (lv1ent_page(entry)) {
1210 		entry = page_entry(entry, iova);
1211 
1212 		if (lv2ent_large(entry))
1213 			phys = lpage_phys(entry) + lpage_offs(iova);
1214 		else if (lv2ent_small(entry))
1215 			phys = spage_phys(entry) + spage_offs(iova);
1216 	}
1217 
1218 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1219 
1220 	return phys;
1221 }
1222 
1223 static struct iommu_device *exynos_iommu_probe_device(struct device *dev)
1224 {
1225 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1226 	struct sysmmu_drvdata *data;
1227 
1228 	if (!has_sysmmu(dev))
1229 		return ERR_PTR(-ENODEV);
1230 
1231 	list_for_each_entry(data, &owner->controllers, owner_node) {
1232 		/*
1233 		 * SYSMMU will be runtime activated via device link
1234 		 * (dependency) to its master device, so there are no
1235 		 * direct calls to pm_runtime_get/put in this driver.
1236 		 */
1237 		data->link = device_link_add(dev, data->sysmmu,
1238 					     DL_FLAG_STATELESS |
1239 					     DL_FLAG_PM_RUNTIME);
1240 	}
1241 
1242 	/* There is always at least one entry, see exynos_iommu_of_xlate() */
1243 	data = list_first_entry(&owner->controllers,
1244 				struct sysmmu_drvdata, owner_node);
1245 
1246 	return &data->iommu;
1247 }
1248 
1249 static void exynos_iommu_release_device(struct device *dev)
1250 {
1251 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1252 	struct sysmmu_drvdata *data;
1253 
1254 	if (!has_sysmmu(dev))
1255 		return;
1256 
1257 	if (owner->domain) {
1258 		struct iommu_group *group = iommu_group_get(dev);
1259 
1260 		if (group) {
1261 			WARN_ON(owner->domain !=
1262 				iommu_group_default_domain(group));
1263 			exynos_iommu_detach_device(owner->domain, dev);
1264 			iommu_group_put(group);
1265 		}
1266 	}
1267 
1268 	list_for_each_entry(data, &owner->controllers, owner_node)
1269 		device_link_del(data->link);
1270 }
1271 
1272 static int exynos_iommu_of_xlate(struct device *dev,
1273 				 struct of_phandle_args *spec)
1274 {
1275 	struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1276 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1277 	struct sysmmu_drvdata *data, *entry;
1278 
1279 	if (!sysmmu)
1280 		return -ENODEV;
1281 
1282 	data = platform_get_drvdata(sysmmu);
1283 	if (!data) {
1284 		put_device(&sysmmu->dev);
1285 		return -ENODEV;
1286 	}
1287 
1288 	if (!owner) {
1289 		owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1290 		if (!owner) {
1291 			put_device(&sysmmu->dev);
1292 			return -ENOMEM;
1293 		}
1294 
1295 		INIT_LIST_HEAD(&owner->controllers);
1296 		mutex_init(&owner->rpm_lock);
1297 		dev_iommu_priv_set(dev, owner);
1298 	}
1299 
1300 	list_for_each_entry(entry, &owner->controllers, owner_node)
1301 		if (entry == data)
1302 			return 0;
1303 
1304 	list_add_tail(&data->owner_node, &owner->controllers);
1305 	data->master = dev;
1306 
1307 	return 0;
1308 }
1309 
1310 static const struct iommu_ops exynos_iommu_ops = {
1311 	.domain_alloc = exynos_iommu_domain_alloc,
1312 	.domain_free = exynos_iommu_domain_free,
1313 	.attach_dev = exynos_iommu_attach_device,
1314 	.detach_dev = exynos_iommu_detach_device,
1315 	.map = exynos_iommu_map,
1316 	.unmap = exynos_iommu_unmap,
1317 	.iova_to_phys = exynos_iommu_iova_to_phys,
1318 	.device_group = generic_device_group,
1319 	.probe_device = exynos_iommu_probe_device,
1320 	.release_device = exynos_iommu_release_device,
1321 	.pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1322 	.of_xlate = exynos_iommu_of_xlate,
1323 };
1324 
1325 static int __init exynos_iommu_init(void)
1326 {
1327 	struct device_node *np;
1328 	int ret;
1329 
1330 	np = of_find_matching_node(NULL, sysmmu_of_match);
1331 	if (!np)
1332 		return 0;
1333 
1334 	of_node_put(np);
1335 
1336 	lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1337 				LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1338 	if (!lv2table_kmem_cache) {
1339 		pr_err("%s: Failed to create kmem cache\n", __func__);
1340 		return -ENOMEM;
1341 	}
1342 
1343 	ret = platform_driver_register(&exynos_sysmmu_driver);
1344 	if (ret) {
1345 		pr_err("%s: Failed to register driver\n", __func__);
1346 		goto err_reg_driver;
1347 	}
1348 
1349 	zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1350 	if (zero_lv2_table == NULL) {
1351 		pr_err("%s: Failed to allocate zero level2 page table\n",
1352 			__func__);
1353 		ret = -ENOMEM;
1354 		goto err_zero_lv2;
1355 	}
1356 
1357 	ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1358 	if (ret) {
1359 		pr_err("%s: Failed to register exynos-iommu driver.\n",
1360 								__func__);
1361 		goto err_set_iommu;
1362 	}
1363 
1364 	return 0;
1365 err_set_iommu:
1366 	kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1367 err_zero_lv2:
1368 	platform_driver_unregister(&exynos_sysmmu_driver);
1369 err_reg_driver:
1370 	kmem_cache_destroy(lv2table_kmem_cache);
1371 	return ret;
1372 }
1373 core_initcall(exynos_iommu_init);
1374