xref: /openbmc/linux/drivers/iommu/exynos-iommu.c (revision bf070bb0)
1 /*
2  * Copyright (c) 2011,2016 Samsung Electronics Co., Ltd.
3  *		http://www.samsung.com
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License version 2 as
7  * published by the Free Software Foundation.
8  */
9 
10 #ifdef CONFIG_EXYNOS_IOMMU_DEBUG
11 #define DEBUG
12 #endif
13 
14 #include <linux/clk.h>
15 #include <linux/dma-mapping.h>
16 #include <linux/err.h>
17 #include <linux/io.h>
18 #include <linux/iommu.h>
19 #include <linux/interrupt.h>
20 #include <linux/list.h>
21 #include <linux/of.h>
22 #include <linux/of_iommu.h>
23 #include <linux/of_platform.h>
24 #include <linux/platform_device.h>
25 #include <linux/pm_runtime.h>
26 #include <linux/slab.h>
27 #include <linux/dma-iommu.h>
28 
29 typedef u32 sysmmu_iova_t;
30 typedef u32 sysmmu_pte_t;
31 
32 /* We do not consider super section mapping (16MB) */
33 #define SECT_ORDER 20
34 #define LPAGE_ORDER 16
35 #define SPAGE_ORDER 12
36 
37 #define SECT_SIZE (1 << SECT_ORDER)
38 #define LPAGE_SIZE (1 << LPAGE_ORDER)
39 #define SPAGE_SIZE (1 << SPAGE_ORDER)
40 
41 #define SECT_MASK (~(SECT_SIZE - 1))
42 #define LPAGE_MASK (~(LPAGE_SIZE - 1))
43 #define SPAGE_MASK (~(SPAGE_SIZE - 1))
44 
45 #define lv1ent_fault(sent) ((*(sent) == ZERO_LV2LINK) || \
46 			   ((*(sent) & 3) == 0) || ((*(sent) & 3) == 3))
47 #define lv1ent_zero(sent) (*(sent) == ZERO_LV2LINK)
48 #define lv1ent_page_zero(sent) ((*(sent) & 3) == 1)
49 #define lv1ent_page(sent) ((*(sent) != ZERO_LV2LINK) && \
50 			  ((*(sent) & 3) == 1))
51 #define lv1ent_section(sent) ((*(sent) & 3) == 2)
52 
53 #define lv2ent_fault(pent) ((*(pent) & 3) == 0)
54 #define lv2ent_small(pent) ((*(pent) & 2) == 2)
55 #define lv2ent_large(pent) ((*(pent) & 3) == 1)
56 
57 /*
58  * v1.x - v3.x SYSMMU supports 32bit physical and 32bit virtual address spaces
59  * v5.0 introduced support for 36bit physical address space by shifting
60  * all page entry values by 4 bits.
61  * All SYSMMU controllers in the system support the address spaces of the same
62  * size, so PG_ENT_SHIFT can be initialized on first SYSMMU probe to proper
63  * value (0 or 4).
64  */
65 static short PG_ENT_SHIFT = -1;
66 #define SYSMMU_PG_ENT_SHIFT 0
67 #define SYSMMU_V5_PG_ENT_SHIFT 4
68 
69 static const sysmmu_pte_t *LV1_PROT;
70 static const sysmmu_pte_t SYSMMU_LV1_PROT[] = {
71 	((0 << 15) | (0 << 10)), /* no access */
72 	((1 << 15) | (1 << 10)), /* IOMMU_READ only */
73 	((0 << 15) | (1 << 10)), /* IOMMU_WRITE not supported, use read/write */
74 	((0 << 15) | (1 << 10)), /* IOMMU_READ | IOMMU_WRITE */
75 };
76 static const sysmmu_pte_t SYSMMU_V5_LV1_PROT[] = {
77 	(0 << 4), /* no access */
78 	(1 << 4), /* IOMMU_READ only */
79 	(2 << 4), /* IOMMU_WRITE only */
80 	(3 << 4), /* IOMMU_READ | IOMMU_WRITE */
81 };
82 
83 static const sysmmu_pte_t *LV2_PROT;
84 static const sysmmu_pte_t SYSMMU_LV2_PROT[] = {
85 	((0 << 9) | (0 << 4)), /* no access */
86 	((1 << 9) | (1 << 4)), /* IOMMU_READ only */
87 	((0 << 9) | (1 << 4)), /* IOMMU_WRITE not supported, use read/write */
88 	((0 << 9) | (1 << 4)), /* IOMMU_READ | IOMMU_WRITE */
89 };
90 static const sysmmu_pte_t SYSMMU_V5_LV2_PROT[] = {
91 	(0 << 2), /* no access */
92 	(1 << 2), /* IOMMU_READ only */
93 	(2 << 2), /* IOMMU_WRITE only */
94 	(3 << 2), /* IOMMU_READ | IOMMU_WRITE */
95 };
96 
97 #define SYSMMU_SUPPORTED_PROT_BITS (IOMMU_READ | IOMMU_WRITE)
98 
99 #define sect_to_phys(ent) (((phys_addr_t) ent) << PG_ENT_SHIFT)
100 #define section_phys(sent) (sect_to_phys(*(sent)) & SECT_MASK)
101 #define section_offs(iova) (iova & (SECT_SIZE - 1))
102 #define lpage_phys(pent) (sect_to_phys(*(pent)) & LPAGE_MASK)
103 #define lpage_offs(iova) (iova & (LPAGE_SIZE - 1))
104 #define spage_phys(pent) (sect_to_phys(*(pent)) & SPAGE_MASK)
105 #define spage_offs(iova) (iova & (SPAGE_SIZE - 1))
106 
107 #define NUM_LV1ENTRIES 4096
108 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
109 
110 static u32 lv1ent_offset(sysmmu_iova_t iova)
111 {
112 	return iova >> SECT_ORDER;
113 }
114 
115 static u32 lv2ent_offset(sysmmu_iova_t iova)
116 {
117 	return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
118 }
119 
120 #define LV1TABLE_SIZE (NUM_LV1ENTRIES * sizeof(sysmmu_pte_t))
121 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
122 
123 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
124 #define lv2table_base(sent) (sect_to_phys(*(sent) & 0xFFFFFFC0))
125 
126 #define mk_lv1ent_sect(pa, prot) ((pa >> PG_ENT_SHIFT) | LV1_PROT[prot] | 2)
127 #define mk_lv1ent_page(pa) ((pa >> PG_ENT_SHIFT) | 1)
128 #define mk_lv2ent_lpage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 1)
129 #define mk_lv2ent_spage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 2)
130 
131 #define CTRL_ENABLE	0x5
132 #define CTRL_BLOCK	0x7
133 #define CTRL_DISABLE	0x0
134 
135 #define CFG_LRU		0x1
136 #define CFG_EAP		(1 << 2)
137 #define CFG_QOS(n)	((n & 0xF) << 7)
138 #define CFG_ACGEN	(1 << 24) /* System MMU 3.3 only */
139 #define CFG_SYSSEL	(1 << 22) /* System MMU 3.2 only */
140 #define CFG_FLPDCACHE	(1 << 20) /* System MMU 3.2+ only */
141 
142 /* common registers */
143 #define REG_MMU_CTRL		0x000
144 #define REG_MMU_CFG		0x004
145 #define REG_MMU_STATUS		0x008
146 #define REG_MMU_VERSION		0x034
147 
148 #define MMU_MAJ_VER(val)	((val) >> 7)
149 #define MMU_MIN_VER(val)	((val) & 0x7F)
150 #define MMU_RAW_VER(reg)	(((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
151 
152 #define MAKE_MMU_VER(maj, min)	((((maj) & 0xF) << 7) | ((min) & 0x7F))
153 
154 /* v1.x - v3.x registers */
155 #define REG_MMU_FLUSH		0x00C
156 #define REG_MMU_FLUSH_ENTRY	0x010
157 #define REG_PT_BASE_ADDR	0x014
158 #define REG_INT_STATUS		0x018
159 #define REG_INT_CLEAR		0x01C
160 
161 #define REG_PAGE_FAULT_ADDR	0x024
162 #define REG_AW_FAULT_ADDR	0x028
163 #define REG_AR_FAULT_ADDR	0x02C
164 #define REG_DEFAULT_SLAVE_ADDR	0x030
165 
166 /* v5.x registers */
167 #define REG_V5_PT_BASE_PFN	0x00C
168 #define REG_V5_MMU_FLUSH_ALL	0x010
169 #define REG_V5_MMU_FLUSH_ENTRY	0x014
170 #define REG_V5_MMU_FLUSH_RANGE	0x018
171 #define REG_V5_MMU_FLUSH_START	0x020
172 #define REG_V5_MMU_FLUSH_END	0x024
173 #define REG_V5_INT_STATUS	0x060
174 #define REG_V5_INT_CLEAR	0x064
175 #define REG_V5_FAULT_AR_VA	0x070
176 #define REG_V5_FAULT_AW_VA	0x080
177 
178 #define has_sysmmu(dev)		(dev->archdata.iommu != NULL)
179 
180 static struct device *dma_dev;
181 static struct kmem_cache *lv2table_kmem_cache;
182 static sysmmu_pte_t *zero_lv2_table;
183 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
184 
185 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
186 {
187 	return pgtable + lv1ent_offset(iova);
188 }
189 
190 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
191 {
192 	return (sysmmu_pte_t *)phys_to_virt(
193 				lv2table_base(sent)) + lv2ent_offset(iova);
194 }
195 
196 /*
197  * IOMMU fault information register
198  */
199 struct sysmmu_fault_info {
200 	unsigned int bit;	/* bit number in STATUS register */
201 	unsigned short addr_reg; /* register to read VA fault address */
202 	const char *name;	/* human readable fault name */
203 	unsigned int type;	/* fault type for report_iommu_fault */
204 };
205 
206 static const struct sysmmu_fault_info sysmmu_faults[] = {
207 	{ 0, REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
208 	{ 1, REG_AR_FAULT_ADDR, "AR MULTI-HIT", IOMMU_FAULT_READ },
209 	{ 2, REG_AW_FAULT_ADDR, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
210 	{ 3, REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
211 	{ 4, REG_AR_FAULT_ADDR, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
212 	{ 5, REG_AR_FAULT_ADDR, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
213 	{ 6, REG_AW_FAULT_ADDR, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
214 	{ 7, REG_AW_FAULT_ADDR, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
215 };
216 
217 static const struct sysmmu_fault_info sysmmu_v5_faults[] = {
218 	{ 0, REG_V5_FAULT_AR_VA, "AR PTW", IOMMU_FAULT_READ },
219 	{ 1, REG_V5_FAULT_AR_VA, "AR PAGE", IOMMU_FAULT_READ },
220 	{ 2, REG_V5_FAULT_AR_VA, "AR MULTI-HIT", IOMMU_FAULT_READ },
221 	{ 3, REG_V5_FAULT_AR_VA, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
222 	{ 4, REG_V5_FAULT_AR_VA, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
223 	{ 16, REG_V5_FAULT_AW_VA, "AW PTW", IOMMU_FAULT_WRITE },
224 	{ 17, REG_V5_FAULT_AW_VA, "AW PAGE", IOMMU_FAULT_WRITE },
225 	{ 18, REG_V5_FAULT_AW_VA, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
226 	{ 19, REG_V5_FAULT_AW_VA, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
227 	{ 20, REG_V5_FAULT_AW_VA, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
228 };
229 
230 /*
231  * This structure is attached to dev.archdata.iommu of the master device
232  * on device add, contains a list of SYSMMU controllers defined by device tree,
233  * which are bound to given master device. It is usually referenced by 'owner'
234  * pointer.
235 */
236 struct exynos_iommu_owner {
237 	struct list_head controllers;	/* list of sysmmu_drvdata.owner_node */
238 	struct iommu_domain *domain;	/* domain this device is attached */
239 	struct mutex rpm_lock;		/* for runtime pm of all sysmmus */
240 };
241 
242 /*
243  * This structure exynos specific generalization of struct iommu_domain.
244  * It contains list of SYSMMU controllers from all master devices, which has
245  * been attached to this domain and page tables of IO address space defined by
246  * it. It is usually referenced by 'domain' pointer.
247  */
248 struct exynos_iommu_domain {
249 	struct list_head clients; /* list of sysmmu_drvdata.domain_node */
250 	sysmmu_pte_t *pgtable;	/* lv1 page table, 16KB */
251 	short *lv2entcnt;	/* free lv2 entry counter for each section */
252 	spinlock_t lock;	/* lock for modyfying list of clients */
253 	spinlock_t pgtablelock;	/* lock for modifying page table @ pgtable */
254 	struct iommu_domain domain; /* generic domain data structure */
255 };
256 
257 /*
258  * This structure hold all data of a single SYSMMU controller, this includes
259  * hw resources like registers and clocks, pointers and list nodes to connect
260  * it to all other structures, internal state and parameters read from device
261  * tree. It is usually referenced by 'data' pointer.
262  */
263 struct sysmmu_drvdata {
264 	struct device *sysmmu;		/* SYSMMU controller device */
265 	struct device *master;		/* master device (owner) */
266 	struct device_link *link;	/* runtime PM link to master */
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 	struct iommu_device iommu;	/* IOMMU core handle */
281 };
282 
283 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
284 {
285 	return container_of(dom, struct exynos_iommu_domain, domain);
286 }
287 
288 static void sysmmu_unblock(struct sysmmu_drvdata *data)
289 {
290 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
291 }
292 
293 static bool sysmmu_block(struct sysmmu_drvdata *data)
294 {
295 	int i = 120;
296 
297 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
298 	while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
299 		--i;
300 
301 	if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
302 		sysmmu_unblock(data);
303 		return false;
304 	}
305 
306 	return true;
307 }
308 
309 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
310 {
311 	if (MMU_MAJ_VER(data->version) < 5)
312 		writel(0x1, data->sfrbase + REG_MMU_FLUSH);
313 	else
314 		writel(0x1, data->sfrbase + REG_V5_MMU_FLUSH_ALL);
315 }
316 
317 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
318 				sysmmu_iova_t iova, unsigned int num_inv)
319 {
320 	unsigned int i;
321 
322 	if (MMU_MAJ_VER(data->version) < 5) {
323 		for (i = 0; i < num_inv; i++) {
324 			writel((iova & SPAGE_MASK) | 1,
325 				     data->sfrbase + REG_MMU_FLUSH_ENTRY);
326 			iova += SPAGE_SIZE;
327 		}
328 	} else {
329 		if (num_inv == 1) {
330 			writel((iova & SPAGE_MASK) | 1,
331 				     data->sfrbase + REG_V5_MMU_FLUSH_ENTRY);
332 		} else {
333 			writel((iova & SPAGE_MASK),
334 				     data->sfrbase + REG_V5_MMU_FLUSH_START);
335 			writel((iova & SPAGE_MASK) + (num_inv - 1) * SPAGE_SIZE,
336 				     data->sfrbase + REG_V5_MMU_FLUSH_END);
337 			writel(1, data->sfrbase + REG_V5_MMU_FLUSH_RANGE);
338 		}
339 	}
340 }
341 
342 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
343 {
344 	if (MMU_MAJ_VER(data->version) < 5)
345 		writel(pgd, data->sfrbase + REG_PT_BASE_ADDR);
346 	else
347 		writel(pgd >> PAGE_SHIFT,
348 			     data->sfrbase + REG_V5_PT_BASE_PFN);
349 
350 	__sysmmu_tlb_invalidate(data);
351 }
352 
353 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
354 {
355 	BUG_ON(clk_prepare_enable(data->clk_master));
356 	BUG_ON(clk_prepare_enable(data->clk));
357 	BUG_ON(clk_prepare_enable(data->pclk));
358 	BUG_ON(clk_prepare_enable(data->aclk));
359 }
360 
361 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
362 {
363 	clk_disable_unprepare(data->aclk);
364 	clk_disable_unprepare(data->pclk);
365 	clk_disable_unprepare(data->clk);
366 	clk_disable_unprepare(data->clk_master);
367 }
368 
369 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
370 {
371 	u32 ver;
372 
373 	__sysmmu_enable_clocks(data);
374 
375 	ver = readl(data->sfrbase + REG_MMU_VERSION);
376 
377 	/* controllers on some SoCs don't report proper version */
378 	if (ver == 0x80000001u)
379 		data->version = MAKE_MMU_VER(1, 0);
380 	else
381 		data->version = MMU_RAW_VER(ver);
382 
383 	dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
384 		MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
385 
386 	__sysmmu_disable_clocks(data);
387 }
388 
389 static void show_fault_information(struct sysmmu_drvdata *data,
390 				   const struct sysmmu_fault_info *finfo,
391 				   sysmmu_iova_t fault_addr)
392 {
393 	sysmmu_pte_t *ent;
394 
395 	dev_err(data->sysmmu, "%s: %s FAULT occurred at %#x\n",
396 		dev_name(data->master), finfo->name, fault_addr);
397 	dev_dbg(data->sysmmu, "Page table base: %pa\n", &data->pgtable);
398 	ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
399 	dev_dbg(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
400 	if (lv1ent_page(ent)) {
401 		ent = page_entry(ent, fault_addr);
402 		dev_dbg(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
403 	}
404 }
405 
406 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
407 {
408 	/* SYSMMU is in blocked state when interrupt occurred. */
409 	struct sysmmu_drvdata *data = dev_id;
410 	const struct sysmmu_fault_info *finfo;
411 	unsigned int i, n, itype;
412 	sysmmu_iova_t fault_addr = -1;
413 	unsigned short reg_status, reg_clear;
414 	int ret = -ENOSYS;
415 
416 	WARN_ON(!data->active);
417 
418 	if (MMU_MAJ_VER(data->version) < 5) {
419 		reg_status = REG_INT_STATUS;
420 		reg_clear = REG_INT_CLEAR;
421 		finfo = sysmmu_faults;
422 		n = ARRAY_SIZE(sysmmu_faults);
423 	} else {
424 		reg_status = REG_V5_INT_STATUS;
425 		reg_clear = REG_V5_INT_CLEAR;
426 		finfo = sysmmu_v5_faults;
427 		n = ARRAY_SIZE(sysmmu_v5_faults);
428 	}
429 
430 	spin_lock(&data->lock);
431 
432 	clk_enable(data->clk_master);
433 
434 	itype = __ffs(readl(data->sfrbase + reg_status));
435 	for (i = 0; i < n; i++, finfo++)
436 		if (finfo->bit == itype)
437 			break;
438 	/* unknown/unsupported fault */
439 	BUG_ON(i == n);
440 
441 	/* print debug message */
442 	fault_addr = readl(data->sfrbase + finfo->addr_reg);
443 	show_fault_information(data, finfo, fault_addr);
444 
445 	if (data->domain)
446 		ret = report_iommu_fault(&data->domain->domain,
447 					data->master, fault_addr, finfo->type);
448 	/* fault is not recovered by fault handler */
449 	BUG_ON(ret != 0);
450 
451 	writel(1 << itype, data->sfrbase + reg_clear);
452 
453 	sysmmu_unblock(data);
454 
455 	clk_disable(data->clk_master);
456 
457 	spin_unlock(&data->lock);
458 
459 	return IRQ_HANDLED;
460 }
461 
462 static void __sysmmu_disable(struct sysmmu_drvdata *data)
463 {
464 	unsigned long flags;
465 
466 	clk_enable(data->clk_master);
467 
468 	spin_lock_irqsave(&data->lock, flags);
469 	writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
470 	writel(0, data->sfrbase + REG_MMU_CFG);
471 	data->active = false;
472 	spin_unlock_irqrestore(&data->lock, flags);
473 
474 	__sysmmu_disable_clocks(data);
475 }
476 
477 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
478 {
479 	unsigned int cfg;
480 
481 	if (data->version <= MAKE_MMU_VER(3, 1))
482 		cfg = CFG_LRU | CFG_QOS(15);
483 	else if (data->version <= MAKE_MMU_VER(3, 2))
484 		cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
485 	else
486 		cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
487 
488 	cfg |= CFG_EAP; /* enable access protection bits check */
489 
490 	writel(cfg, data->sfrbase + REG_MMU_CFG);
491 }
492 
493 static void __sysmmu_enable(struct sysmmu_drvdata *data)
494 {
495 	unsigned long flags;
496 
497 	__sysmmu_enable_clocks(data);
498 
499 	spin_lock_irqsave(&data->lock, flags);
500 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
501 	__sysmmu_init_config(data);
502 	__sysmmu_set_ptbase(data, data->pgtable);
503 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
504 	data->active = true;
505 	spin_unlock_irqrestore(&data->lock, flags);
506 
507 	/*
508 	 * SYSMMU driver keeps master's clock enabled only for the short
509 	 * time, while accessing the registers. For performing address
510 	 * translation during DMA transaction it relies on the client
511 	 * driver to enable it.
512 	 */
513 	clk_disable(data->clk_master);
514 }
515 
516 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
517 					    sysmmu_iova_t iova)
518 {
519 	unsigned long flags;
520 
521 	spin_lock_irqsave(&data->lock, flags);
522 	if (data->active && data->version >= MAKE_MMU_VER(3, 3)) {
523 		clk_enable(data->clk_master);
524 		if (sysmmu_block(data)) {
525 			if (data->version >= MAKE_MMU_VER(5, 0))
526 				__sysmmu_tlb_invalidate(data);
527 			else
528 				__sysmmu_tlb_invalidate_entry(data, iova, 1);
529 			sysmmu_unblock(data);
530 		}
531 		clk_disable(data->clk_master);
532 	}
533 	spin_unlock_irqrestore(&data->lock, flags);
534 }
535 
536 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
537 					sysmmu_iova_t iova, size_t size)
538 {
539 	unsigned long flags;
540 
541 	spin_lock_irqsave(&data->lock, flags);
542 	if (data->active) {
543 		unsigned int num_inv = 1;
544 
545 		clk_enable(data->clk_master);
546 
547 		/*
548 		 * L2TLB invalidation required
549 		 * 4KB page: 1 invalidation
550 		 * 64KB page: 16 invalidations
551 		 * 1MB page: 64 invalidations
552 		 * because it is set-associative TLB
553 		 * with 8-way and 64 sets.
554 		 * 1MB page can be cached in one of all sets.
555 		 * 64KB page can be one of 16 consecutive sets.
556 		 */
557 		if (MMU_MAJ_VER(data->version) == 2)
558 			num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
559 
560 		if (sysmmu_block(data)) {
561 			__sysmmu_tlb_invalidate_entry(data, iova, num_inv);
562 			sysmmu_unblock(data);
563 		}
564 		clk_disable(data->clk_master);
565 	}
566 	spin_unlock_irqrestore(&data->lock, flags);
567 }
568 
569 static const struct iommu_ops exynos_iommu_ops;
570 
571 static int __init exynos_sysmmu_probe(struct platform_device *pdev)
572 {
573 	int irq, ret;
574 	struct device *dev = &pdev->dev;
575 	struct sysmmu_drvdata *data;
576 	struct resource *res;
577 
578 	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
579 	if (!data)
580 		return -ENOMEM;
581 
582 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
583 	data->sfrbase = devm_ioremap_resource(dev, res);
584 	if (IS_ERR(data->sfrbase))
585 		return PTR_ERR(data->sfrbase);
586 
587 	irq = platform_get_irq(pdev, 0);
588 	if (irq <= 0) {
589 		dev_err(dev, "Unable to find IRQ resource\n");
590 		return irq;
591 	}
592 
593 	ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
594 				dev_name(dev), data);
595 	if (ret) {
596 		dev_err(dev, "Unabled to register handler of irq %d\n", irq);
597 		return ret;
598 	}
599 
600 	data->clk = devm_clk_get(dev, "sysmmu");
601 	if (PTR_ERR(data->clk) == -ENOENT)
602 		data->clk = NULL;
603 	else if (IS_ERR(data->clk))
604 		return PTR_ERR(data->clk);
605 
606 	data->aclk = devm_clk_get(dev, "aclk");
607 	if (PTR_ERR(data->aclk) == -ENOENT)
608 		data->aclk = NULL;
609 	else if (IS_ERR(data->aclk))
610 		return PTR_ERR(data->aclk);
611 
612 	data->pclk = devm_clk_get(dev, "pclk");
613 	if (PTR_ERR(data->pclk) == -ENOENT)
614 		data->pclk = NULL;
615 	else if (IS_ERR(data->pclk))
616 		return PTR_ERR(data->pclk);
617 
618 	if (!data->clk && (!data->aclk || !data->pclk)) {
619 		dev_err(dev, "Failed to get device clock(s)!\n");
620 		return -ENOSYS;
621 	}
622 
623 	data->clk_master = devm_clk_get(dev, "master");
624 	if (PTR_ERR(data->clk_master) == -ENOENT)
625 		data->clk_master = NULL;
626 	else if (IS_ERR(data->clk_master))
627 		return PTR_ERR(data->clk_master);
628 
629 	data->sysmmu = dev;
630 	spin_lock_init(&data->lock);
631 
632 	ret = iommu_device_sysfs_add(&data->iommu, &pdev->dev, NULL,
633 				     dev_name(data->sysmmu));
634 	if (ret)
635 		return ret;
636 
637 	iommu_device_set_ops(&data->iommu, &exynos_iommu_ops);
638 	iommu_device_set_fwnode(&data->iommu, &dev->of_node->fwnode);
639 
640 	ret = iommu_device_register(&data->iommu);
641 	if (ret)
642 		return ret;
643 
644 	platform_set_drvdata(pdev, data);
645 
646 	__sysmmu_get_version(data);
647 	if (PG_ENT_SHIFT < 0) {
648 		if (MMU_MAJ_VER(data->version) < 5) {
649 			PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
650 			LV1_PROT = SYSMMU_LV1_PROT;
651 			LV2_PROT = SYSMMU_LV2_PROT;
652 		} else {
653 			PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
654 			LV1_PROT = SYSMMU_V5_LV1_PROT;
655 			LV2_PROT = SYSMMU_V5_LV2_PROT;
656 		}
657 	}
658 
659 	/*
660 	 * use the first registered sysmmu device for performing
661 	 * dma mapping operations on iommu page tables (cpu cache flush)
662 	 */
663 	if (!dma_dev)
664 		dma_dev = &pdev->dev;
665 
666 	pm_runtime_enable(dev);
667 
668 	return 0;
669 }
670 
671 static int __maybe_unused exynos_sysmmu_suspend(struct device *dev)
672 {
673 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
674 	struct device *master = data->master;
675 
676 	if (master) {
677 		struct exynos_iommu_owner *owner = master->archdata.iommu;
678 
679 		mutex_lock(&owner->rpm_lock);
680 		if (data->domain) {
681 			dev_dbg(data->sysmmu, "saving state\n");
682 			__sysmmu_disable(data);
683 		}
684 		mutex_unlock(&owner->rpm_lock);
685 	}
686 	return 0;
687 }
688 
689 static int __maybe_unused exynos_sysmmu_resume(struct device *dev)
690 {
691 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
692 	struct device *master = data->master;
693 
694 	if (master) {
695 		struct exynos_iommu_owner *owner = master->archdata.iommu;
696 
697 		mutex_lock(&owner->rpm_lock);
698 		if (data->domain) {
699 			dev_dbg(data->sysmmu, "restoring state\n");
700 			__sysmmu_enable(data);
701 		}
702 		mutex_unlock(&owner->rpm_lock);
703 	}
704 	return 0;
705 }
706 
707 static const struct dev_pm_ops sysmmu_pm_ops = {
708 	SET_RUNTIME_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume, NULL)
709 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
710 				pm_runtime_force_resume)
711 };
712 
713 static const struct of_device_id sysmmu_of_match[] = {
714 	{ .compatible	= "samsung,exynos-sysmmu", },
715 	{ },
716 };
717 
718 static struct platform_driver exynos_sysmmu_driver __refdata = {
719 	.probe	= exynos_sysmmu_probe,
720 	.driver	= {
721 		.name		= "exynos-sysmmu",
722 		.of_match_table	= sysmmu_of_match,
723 		.pm		= &sysmmu_pm_ops,
724 		.suppress_bind_attrs = true,
725 	}
726 };
727 
728 static inline void update_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
729 {
730 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
731 				DMA_TO_DEVICE);
732 	*ent = cpu_to_le32(val);
733 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
734 				   DMA_TO_DEVICE);
735 }
736 
737 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
738 {
739 	struct exynos_iommu_domain *domain;
740 	dma_addr_t handle;
741 	int i;
742 
743 	/* Check if correct PTE offsets are initialized */
744 	BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
745 
746 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
747 	if (!domain)
748 		return NULL;
749 
750 	if (type == IOMMU_DOMAIN_DMA) {
751 		if (iommu_get_dma_cookie(&domain->domain) != 0)
752 			goto err_pgtable;
753 	} else if (type != IOMMU_DOMAIN_UNMANAGED) {
754 		goto err_pgtable;
755 	}
756 
757 	domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
758 	if (!domain->pgtable)
759 		goto err_dma_cookie;
760 
761 	domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
762 	if (!domain->lv2entcnt)
763 		goto err_counter;
764 
765 	/* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
766 	for (i = 0; i < NUM_LV1ENTRIES; i++)
767 		domain->pgtable[i] = ZERO_LV2LINK;
768 
769 	handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
770 				DMA_TO_DEVICE);
771 	/* For mapping page table entries we rely on dma == phys */
772 	BUG_ON(handle != virt_to_phys(domain->pgtable));
773 	if (dma_mapping_error(dma_dev, handle))
774 		goto err_lv2ent;
775 
776 	spin_lock_init(&domain->lock);
777 	spin_lock_init(&domain->pgtablelock);
778 	INIT_LIST_HEAD(&domain->clients);
779 
780 	domain->domain.geometry.aperture_start = 0;
781 	domain->domain.geometry.aperture_end   = ~0UL;
782 	domain->domain.geometry.force_aperture = true;
783 
784 	return &domain->domain;
785 
786 err_lv2ent:
787 	free_pages((unsigned long)domain->lv2entcnt, 1);
788 err_counter:
789 	free_pages((unsigned long)domain->pgtable, 2);
790 err_dma_cookie:
791 	if (type == IOMMU_DOMAIN_DMA)
792 		iommu_put_dma_cookie(&domain->domain);
793 err_pgtable:
794 	kfree(domain);
795 	return NULL;
796 }
797 
798 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
799 {
800 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
801 	struct sysmmu_drvdata *data, *next;
802 	unsigned long flags;
803 	int i;
804 
805 	WARN_ON(!list_empty(&domain->clients));
806 
807 	spin_lock_irqsave(&domain->lock, flags);
808 
809 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
810 		spin_lock(&data->lock);
811 		__sysmmu_disable(data);
812 		data->pgtable = 0;
813 		data->domain = NULL;
814 		list_del_init(&data->domain_node);
815 		spin_unlock(&data->lock);
816 	}
817 
818 	spin_unlock_irqrestore(&domain->lock, flags);
819 
820 	if (iommu_domain->type == IOMMU_DOMAIN_DMA)
821 		iommu_put_dma_cookie(iommu_domain);
822 
823 	dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
824 			 DMA_TO_DEVICE);
825 
826 	for (i = 0; i < NUM_LV1ENTRIES; i++)
827 		if (lv1ent_page(domain->pgtable + i)) {
828 			phys_addr_t base = lv2table_base(domain->pgtable + i);
829 
830 			dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
831 					 DMA_TO_DEVICE);
832 			kmem_cache_free(lv2table_kmem_cache,
833 					phys_to_virt(base));
834 		}
835 
836 	free_pages((unsigned long)domain->pgtable, 2);
837 	free_pages((unsigned long)domain->lv2entcnt, 1);
838 	kfree(domain);
839 }
840 
841 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
842 				    struct device *dev)
843 {
844 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
845 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
846 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
847 	struct sysmmu_drvdata *data, *next;
848 	unsigned long flags;
849 
850 	if (!has_sysmmu(dev) || owner->domain != iommu_domain)
851 		return;
852 
853 	mutex_lock(&owner->rpm_lock);
854 
855 	list_for_each_entry(data, &owner->controllers, owner_node) {
856 		pm_runtime_get_noresume(data->sysmmu);
857 		if (pm_runtime_active(data->sysmmu))
858 			__sysmmu_disable(data);
859 		pm_runtime_put(data->sysmmu);
860 	}
861 
862 	spin_lock_irqsave(&domain->lock, flags);
863 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
864 		spin_lock(&data->lock);
865 		data->pgtable = 0;
866 		data->domain = NULL;
867 		list_del_init(&data->domain_node);
868 		spin_unlock(&data->lock);
869 	}
870 	owner->domain = NULL;
871 	spin_unlock_irqrestore(&domain->lock, flags);
872 
873 	mutex_unlock(&owner->rpm_lock);
874 
875 	dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n", __func__,
876 		&pagetable);
877 }
878 
879 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
880 				   struct device *dev)
881 {
882 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
883 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
884 	struct sysmmu_drvdata *data;
885 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
886 	unsigned long flags;
887 
888 	if (!has_sysmmu(dev))
889 		return -ENODEV;
890 
891 	if (owner->domain)
892 		exynos_iommu_detach_device(owner->domain, dev);
893 
894 	mutex_lock(&owner->rpm_lock);
895 
896 	spin_lock_irqsave(&domain->lock, flags);
897 	list_for_each_entry(data, &owner->controllers, owner_node) {
898 		spin_lock(&data->lock);
899 		data->pgtable = pagetable;
900 		data->domain = domain;
901 		list_add_tail(&data->domain_node, &domain->clients);
902 		spin_unlock(&data->lock);
903 	}
904 	owner->domain = iommu_domain;
905 	spin_unlock_irqrestore(&domain->lock, flags);
906 
907 	list_for_each_entry(data, &owner->controllers, owner_node) {
908 		pm_runtime_get_noresume(data->sysmmu);
909 		if (pm_runtime_active(data->sysmmu))
910 			__sysmmu_enable(data);
911 		pm_runtime_put(data->sysmmu);
912 	}
913 
914 	mutex_unlock(&owner->rpm_lock);
915 
916 	dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa\n", __func__,
917 		&pagetable);
918 
919 	return 0;
920 }
921 
922 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
923 		sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
924 {
925 	if (lv1ent_section(sent)) {
926 		WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
927 		return ERR_PTR(-EADDRINUSE);
928 	}
929 
930 	if (lv1ent_fault(sent)) {
931 		dma_addr_t handle;
932 		sysmmu_pte_t *pent;
933 		bool need_flush_flpd_cache = lv1ent_zero(sent);
934 
935 		pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
936 		BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
937 		if (!pent)
938 			return ERR_PTR(-ENOMEM);
939 
940 		update_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
941 		kmemleak_ignore(pent);
942 		*pgcounter = NUM_LV2ENTRIES;
943 		handle = dma_map_single(dma_dev, pent, LV2TABLE_SIZE,
944 					DMA_TO_DEVICE);
945 		if (dma_mapping_error(dma_dev, handle)) {
946 			kmem_cache_free(lv2table_kmem_cache, pent);
947 			return ERR_PTR(-EADDRINUSE);
948 		}
949 
950 		/*
951 		 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
952 		 * FLPD cache may cache the address of zero_l2_table. This
953 		 * function replaces the zero_l2_table with new L2 page table
954 		 * to write valid mappings.
955 		 * Accessing the valid area may cause page fault since FLPD
956 		 * cache may still cache zero_l2_table for the valid area
957 		 * instead of new L2 page table that has the mapping
958 		 * information of the valid area.
959 		 * Thus any replacement of zero_l2_table with other valid L2
960 		 * page table must involve FLPD cache invalidation for System
961 		 * MMU v3.3.
962 		 * FLPD cache invalidation is performed with TLB invalidation
963 		 * by VPN without blocking. It is safe to invalidate TLB without
964 		 * blocking because the target address of TLB invalidation is
965 		 * not currently mapped.
966 		 */
967 		if (need_flush_flpd_cache) {
968 			struct sysmmu_drvdata *data;
969 
970 			spin_lock(&domain->lock);
971 			list_for_each_entry(data, &domain->clients, domain_node)
972 				sysmmu_tlb_invalidate_flpdcache(data, iova);
973 			spin_unlock(&domain->lock);
974 		}
975 	}
976 
977 	return page_entry(sent, iova);
978 }
979 
980 static int lv1set_section(struct exynos_iommu_domain *domain,
981 			  sysmmu_pte_t *sent, sysmmu_iova_t iova,
982 			  phys_addr_t paddr, int prot, short *pgcnt)
983 {
984 	if (lv1ent_section(sent)) {
985 		WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
986 			iova);
987 		return -EADDRINUSE;
988 	}
989 
990 	if (lv1ent_page(sent)) {
991 		if (*pgcnt != NUM_LV2ENTRIES) {
992 			WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
993 				iova);
994 			return -EADDRINUSE;
995 		}
996 
997 		kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
998 		*pgcnt = 0;
999 	}
1000 
1001 	update_pte(sent, mk_lv1ent_sect(paddr, prot));
1002 
1003 	spin_lock(&domain->lock);
1004 	if (lv1ent_page_zero(sent)) {
1005 		struct sysmmu_drvdata *data;
1006 		/*
1007 		 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
1008 		 * entry by speculative prefetch of SLPD which has no mapping.
1009 		 */
1010 		list_for_each_entry(data, &domain->clients, domain_node)
1011 			sysmmu_tlb_invalidate_flpdcache(data, iova);
1012 	}
1013 	spin_unlock(&domain->lock);
1014 
1015 	return 0;
1016 }
1017 
1018 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
1019 		       int prot, short *pgcnt)
1020 {
1021 	if (size == SPAGE_SIZE) {
1022 		if (WARN_ON(!lv2ent_fault(pent)))
1023 			return -EADDRINUSE;
1024 
1025 		update_pte(pent, mk_lv2ent_spage(paddr, prot));
1026 		*pgcnt -= 1;
1027 	} else { /* size == LPAGE_SIZE */
1028 		int i;
1029 		dma_addr_t pent_base = virt_to_phys(pent);
1030 
1031 		dma_sync_single_for_cpu(dma_dev, pent_base,
1032 					sizeof(*pent) * SPAGES_PER_LPAGE,
1033 					DMA_TO_DEVICE);
1034 		for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1035 			if (WARN_ON(!lv2ent_fault(pent))) {
1036 				if (i > 0)
1037 					memset(pent - i, 0, sizeof(*pent) * i);
1038 				return -EADDRINUSE;
1039 			}
1040 
1041 			*pent = mk_lv2ent_lpage(paddr, prot);
1042 		}
1043 		dma_sync_single_for_device(dma_dev, pent_base,
1044 					   sizeof(*pent) * SPAGES_PER_LPAGE,
1045 					   DMA_TO_DEVICE);
1046 		*pgcnt -= SPAGES_PER_LPAGE;
1047 	}
1048 
1049 	return 0;
1050 }
1051 
1052 /*
1053  * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1054  *
1055  * System MMU v3.x has advanced logic to improve address translation
1056  * performance with caching more page table entries by a page table walk.
1057  * However, the logic has a bug that while caching faulty page table entries,
1058  * System MMU reports page fault if the cached fault entry is hit even though
1059  * the fault entry is updated to a valid entry after the entry is cached.
1060  * To prevent caching faulty page table entries which may be updated to valid
1061  * entries later, the virtual memory manager should care about the workaround
1062  * for the problem. The following describes the workaround.
1063  *
1064  * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1065  * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1066  *
1067  * Precisely, any start address of I/O virtual region must be aligned with
1068  * the following sizes for System MMU v3.1 and v3.2.
1069  * System MMU v3.1: 128KiB
1070  * System MMU v3.2: 256KiB
1071  *
1072  * Because System MMU v3.3 caches page table entries more aggressively, it needs
1073  * more workarounds.
1074  * - Any two consecutive I/O virtual regions must have a hole of size larger
1075  *   than or equal to 128KiB.
1076  * - Start address of an I/O virtual region must be aligned by 128KiB.
1077  */
1078 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1079 			    unsigned long l_iova, phys_addr_t paddr, size_t size,
1080 			    int prot)
1081 {
1082 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1083 	sysmmu_pte_t *entry;
1084 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1085 	unsigned long flags;
1086 	int ret = -ENOMEM;
1087 
1088 	BUG_ON(domain->pgtable == NULL);
1089 	prot &= SYSMMU_SUPPORTED_PROT_BITS;
1090 
1091 	spin_lock_irqsave(&domain->pgtablelock, flags);
1092 
1093 	entry = section_entry(domain->pgtable, iova);
1094 
1095 	if (size == SECT_SIZE) {
1096 		ret = lv1set_section(domain, entry, iova, paddr, prot,
1097 				     &domain->lv2entcnt[lv1ent_offset(iova)]);
1098 	} else {
1099 		sysmmu_pte_t *pent;
1100 
1101 		pent = alloc_lv2entry(domain, entry, iova,
1102 				      &domain->lv2entcnt[lv1ent_offset(iova)]);
1103 
1104 		if (IS_ERR(pent))
1105 			ret = PTR_ERR(pent);
1106 		else
1107 			ret = lv2set_page(pent, paddr, size, prot,
1108 				       &domain->lv2entcnt[lv1ent_offset(iova)]);
1109 	}
1110 
1111 	if (ret)
1112 		pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1113 			__func__, ret, size, iova);
1114 
1115 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1116 
1117 	return ret;
1118 }
1119 
1120 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1121 					      sysmmu_iova_t iova, size_t size)
1122 {
1123 	struct sysmmu_drvdata *data;
1124 	unsigned long flags;
1125 
1126 	spin_lock_irqsave(&domain->lock, flags);
1127 
1128 	list_for_each_entry(data, &domain->clients, domain_node)
1129 		sysmmu_tlb_invalidate_entry(data, iova, size);
1130 
1131 	spin_unlock_irqrestore(&domain->lock, flags);
1132 }
1133 
1134 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1135 				 unsigned long l_iova, size_t size)
1136 {
1137 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1138 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1139 	sysmmu_pte_t *ent;
1140 	size_t err_pgsize;
1141 	unsigned long flags;
1142 
1143 	BUG_ON(domain->pgtable == NULL);
1144 
1145 	spin_lock_irqsave(&domain->pgtablelock, flags);
1146 
1147 	ent = section_entry(domain->pgtable, iova);
1148 
1149 	if (lv1ent_section(ent)) {
1150 		if (WARN_ON(size < SECT_SIZE)) {
1151 			err_pgsize = SECT_SIZE;
1152 			goto err;
1153 		}
1154 
1155 		/* workaround for h/w bug in System MMU v3.3 */
1156 		update_pte(ent, ZERO_LV2LINK);
1157 		size = SECT_SIZE;
1158 		goto done;
1159 	}
1160 
1161 	if (unlikely(lv1ent_fault(ent))) {
1162 		if (size > SECT_SIZE)
1163 			size = SECT_SIZE;
1164 		goto done;
1165 	}
1166 
1167 	/* lv1ent_page(sent) == true here */
1168 
1169 	ent = page_entry(ent, iova);
1170 
1171 	if (unlikely(lv2ent_fault(ent))) {
1172 		size = SPAGE_SIZE;
1173 		goto done;
1174 	}
1175 
1176 	if (lv2ent_small(ent)) {
1177 		update_pte(ent, 0);
1178 		size = SPAGE_SIZE;
1179 		domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1180 		goto done;
1181 	}
1182 
1183 	/* lv1ent_large(ent) == true here */
1184 	if (WARN_ON(size < LPAGE_SIZE)) {
1185 		err_pgsize = LPAGE_SIZE;
1186 		goto err;
1187 	}
1188 
1189 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1190 				sizeof(*ent) * SPAGES_PER_LPAGE,
1191 				DMA_TO_DEVICE);
1192 	memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1193 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1194 				   sizeof(*ent) * SPAGES_PER_LPAGE,
1195 				   DMA_TO_DEVICE);
1196 	size = LPAGE_SIZE;
1197 	domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1198 done:
1199 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1200 
1201 	exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1202 
1203 	return size;
1204 err:
1205 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1206 
1207 	pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1208 		__func__, size, iova, err_pgsize);
1209 
1210 	return 0;
1211 }
1212 
1213 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1214 					  dma_addr_t iova)
1215 {
1216 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1217 	sysmmu_pte_t *entry;
1218 	unsigned long flags;
1219 	phys_addr_t phys = 0;
1220 
1221 	spin_lock_irqsave(&domain->pgtablelock, flags);
1222 
1223 	entry = section_entry(domain->pgtable, iova);
1224 
1225 	if (lv1ent_section(entry)) {
1226 		phys = section_phys(entry) + section_offs(iova);
1227 	} else if (lv1ent_page(entry)) {
1228 		entry = page_entry(entry, iova);
1229 
1230 		if (lv2ent_large(entry))
1231 			phys = lpage_phys(entry) + lpage_offs(iova);
1232 		else if (lv2ent_small(entry))
1233 			phys = spage_phys(entry) + spage_offs(iova);
1234 	}
1235 
1236 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1237 
1238 	return phys;
1239 }
1240 
1241 static struct iommu_group *get_device_iommu_group(struct device *dev)
1242 {
1243 	struct iommu_group *group;
1244 
1245 	group = iommu_group_get(dev);
1246 	if (!group)
1247 		group = iommu_group_alloc();
1248 
1249 	return group;
1250 }
1251 
1252 static int exynos_iommu_add_device(struct device *dev)
1253 {
1254 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1255 	struct sysmmu_drvdata *data;
1256 	struct iommu_group *group;
1257 
1258 	if (!has_sysmmu(dev))
1259 		return -ENODEV;
1260 
1261 	group = iommu_group_get_for_dev(dev);
1262 
1263 	if (IS_ERR(group))
1264 		return PTR_ERR(group);
1265 
1266 	list_for_each_entry(data, &owner->controllers, owner_node) {
1267 		/*
1268 		 * SYSMMU will be runtime activated via device link
1269 		 * (dependency) to its master device, so there are no
1270 		 * direct calls to pm_runtime_get/put in this driver.
1271 		 */
1272 		data->link = device_link_add(dev, data->sysmmu,
1273 					     DL_FLAG_PM_RUNTIME);
1274 	}
1275 	iommu_group_put(group);
1276 
1277 	return 0;
1278 }
1279 
1280 static void exynos_iommu_remove_device(struct device *dev)
1281 {
1282 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1283 	struct sysmmu_drvdata *data;
1284 
1285 	if (!has_sysmmu(dev))
1286 		return;
1287 
1288 	if (owner->domain) {
1289 		struct iommu_group *group = iommu_group_get(dev);
1290 
1291 		if (group) {
1292 			WARN_ON(owner->domain !=
1293 				iommu_group_default_domain(group));
1294 			exynos_iommu_detach_device(owner->domain, dev);
1295 			iommu_group_put(group);
1296 		}
1297 	}
1298 	iommu_group_remove_device(dev);
1299 
1300 	list_for_each_entry(data, &owner->controllers, owner_node)
1301 		device_link_del(data->link);
1302 }
1303 
1304 static int exynos_iommu_of_xlate(struct device *dev,
1305 				 struct of_phandle_args *spec)
1306 {
1307 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1308 	struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1309 	struct sysmmu_drvdata *data, *entry;
1310 
1311 	if (!sysmmu)
1312 		return -ENODEV;
1313 
1314 	data = platform_get_drvdata(sysmmu);
1315 	if (!data)
1316 		return -ENODEV;
1317 
1318 	if (!owner) {
1319 		owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1320 		if (!owner)
1321 			return -ENOMEM;
1322 
1323 		INIT_LIST_HEAD(&owner->controllers);
1324 		mutex_init(&owner->rpm_lock);
1325 		dev->archdata.iommu = owner;
1326 	}
1327 
1328 	list_for_each_entry(entry, &owner->controllers, owner_node)
1329 		if (entry == data)
1330 			return 0;
1331 
1332 	list_add_tail(&data->owner_node, &owner->controllers);
1333 	data->master = dev;
1334 
1335 	return 0;
1336 }
1337 
1338 static const struct iommu_ops exynos_iommu_ops = {
1339 	.domain_alloc = exynos_iommu_domain_alloc,
1340 	.domain_free = exynos_iommu_domain_free,
1341 	.attach_dev = exynos_iommu_attach_device,
1342 	.detach_dev = exynos_iommu_detach_device,
1343 	.map = exynos_iommu_map,
1344 	.unmap = exynos_iommu_unmap,
1345 	.map_sg = default_iommu_map_sg,
1346 	.iova_to_phys = exynos_iommu_iova_to_phys,
1347 	.device_group = get_device_iommu_group,
1348 	.add_device = exynos_iommu_add_device,
1349 	.remove_device = exynos_iommu_remove_device,
1350 	.pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1351 	.of_xlate = exynos_iommu_of_xlate,
1352 };
1353 
1354 static int __init exynos_iommu_init(void)
1355 {
1356 	int ret;
1357 
1358 	lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1359 				LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1360 	if (!lv2table_kmem_cache) {
1361 		pr_err("%s: Failed to create kmem cache\n", __func__);
1362 		return -ENOMEM;
1363 	}
1364 
1365 	ret = platform_driver_register(&exynos_sysmmu_driver);
1366 	if (ret) {
1367 		pr_err("%s: Failed to register driver\n", __func__);
1368 		goto err_reg_driver;
1369 	}
1370 
1371 	zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1372 	if (zero_lv2_table == NULL) {
1373 		pr_err("%s: Failed to allocate zero level2 page table\n",
1374 			__func__);
1375 		ret = -ENOMEM;
1376 		goto err_zero_lv2;
1377 	}
1378 
1379 	ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1380 	if (ret) {
1381 		pr_err("%s: Failed to register exynos-iommu driver.\n",
1382 								__func__);
1383 		goto err_set_iommu;
1384 	}
1385 
1386 	return 0;
1387 err_set_iommu:
1388 	kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1389 err_zero_lv2:
1390 	platform_driver_unregister(&exynos_sysmmu_driver);
1391 err_reg_driver:
1392 	kmem_cache_destroy(lv2table_kmem_cache);
1393 	return ret;
1394 }
1395 core_initcall(exynos_iommu_init);
1396 
1397 IOMMU_OF_DECLARE(exynos_iommu_of, "samsung,exynos-sysmmu", NULL);
1398