xref: /openbmc/linux/drivers/iommu/exynos-iommu.c (revision 4da722ca)
1 /*
2  * Copyright (c) 2011,2016 Samsung Electronics Co., Ltd.
3  *		http://www.samsung.com
4  *
5  * This program is free software; you can redistribute it and/or modify
6  * it under the terms of the GNU General Public License version 2 as
7  * published by the Free Software Foundation.
8  */
9 
10 #ifdef CONFIG_EXYNOS_IOMMU_DEBUG
11 #define DEBUG
12 #endif
13 
14 #include <linux/clk.h>
15 #include <linux/dma-mapping.h>
16 #include <linux/err.h>
17 #include <linux/io.h>
18 #include <linux/iommu.h>
19 #include <linux/interrupt.h>
20 #include <linux/list.h>
21 #include <linux/of.h>
22 #include <linux/of_iommu.h>
23 #include <linux/of_platform.h>
24 #include <linux/platform_device.h>
25 #include <linux/pm_runtime.h>
26 #include <linux/slab.h>
27 #include <linux/dma-iommu.h>
28 
29 typedef u32 sysmmu_iova_t;
30 typedef u32 sysmmu_pte_t;
31 
32 /* We do not consider super section mapping (16MB) */
33 #define SECT_ORDER 20
34 #define LPAGE_ORDER 16
35 #define SPAGE_ORDER 12
36 
37 #define SECT_SIZE (1 << SECT_ORDER)
38 #define LPAGE_SIZE (1 << LPAGE_ORDER)
39 #define SPAGE_SIZE (1 << SPAGE_ORDER)
40 
41 #define SECT_MASK (~(SECT_SIZE - 1))
42 #define LPAGE_MASK (~(LPAGE_SIZE - 1))
43 #define SPAGE_MASK (~(SPAGE_SIZE - 1))
44 
45 #define lv1ent_fault(sent) ((*(sent) == ZERO_LV2LINK) || \
46 			   ((*(sent) & 3) == 0) || ((*(sent) & 3) == 3))
47 #define lv1ent_zero(sent) (*(sent) == ZERO_LV2LINK)
48 #define lv1ent_page_zero(sent) ((*(sent) & 3) == 1)
49 #define lv1ent_page(sent) ((*(sent) != ZERO_LV2LINK) && \
50 			  ((*(sent) & 3) == 1))
51 #define lv1ent_section(sent) ((*(sent) & 3) == 2)
52 
53 #define lv2ent_fault(pent) ((*(pent) & 3) == 0)
54 #define lv2ent_small(pent) ((*(pent) & 2) == 2)
55 #define lv2ent_large(pent) ((*(pent) & 3) == 1)
56 
57 #ifdef CONFIG_BIG_ENDIAN
58 #warning "revisit driver if we can enable big-endian ptes"
59 #endif
60 
61 /*
62  * v1.x - v3.x SYSMMU supports 32bit physical and 32bit virtual address spaces
63  * v5.0 introduced support for 36bit physical address space by shifting
64  * all page entry values by 4 bits.
65  * All SYSMMU controllers in the system support the address spaces of the same
66  * size, so PG_ENT_SHIFT can be initialized on first SYSMMU probe to proper
67  * value (0 or 4).
68  */
69 static short PG_ENT_SHIFT = -1;
70 #define SYSMMU_PG_ENT_SHIFT 0
71 #define SYSMMU_V5_PG_ENT_SHIFT 4
72 
73 static const sysmmu_pte_t *LV1_PROT;
74 static const sysmmu_pte_t SYSMMU_LV1_PROT[] = {
75 	((0 << 15) | (0 << 10)), /* no access */
76 	((1 << 15) | (1 << 10)), /* IOMMU_READ only */
77 	((0 << 15) | (1 << 10)), /* IOMMU_WRITE not supported, use read/write */
78 	((0 << 15) | (1 << 10)), /* IOMMU_READ | IOMMU_WRITE */
79 };
80 static const sysmmu_pte_t SYSMMU_V5_LV1_PROT[] = {
81 	(0 << 4), /* no access */
82 	(1 << 4), /* IOMMU_READ only */
83 	(2 << 4), /* IOMMU_WRITE only */
84 	(3 << 4), /* IOMMU_READ | IOMMU_WRITE */
85 };
86 
87 static const sysmmu_pte_t *LV2_PROT;
88 static const sysmmu_pte_t SYSMMU_LV2_PROT[] = {
89 	((0 << 9) | (0 << 4)), /* no access */
90 	((1 << 9) | (1 << 4)), /* IOMMU_READ only */
91 	((0 << 9) | (1 << 4)), /* IOMMU_WRITE not supported, use read/write */
92 	((0 << 9) | (1 << 4)), /* IOMMU_READ | IOMMU_WRITE */
93 };
94 static const sysmmu_pte_t SYSMMU_V5_LV2_PROT[] = {
95 	(0 << 2), /* no access */
96 	(1 << 2), /* IOMMU_READ only */
97 	(2 << 2), /* IOMMU_WRITE only */
98 	(3 << 2), /* IOMMU_READ | IOMMU_WRITE */
99 };
100 
101 #define SYSMMU_SUPPORTED_PROT_BITS (IOMMU_READ | IOMMU_WRITE)
102 
103 #define sect_to_phys(ent) (((phys_addr_t) ent) << PG_ENT_SHIFT)
104 #define section_phys(sent) (sect_to_phys(*(sent)) & SECT_MASK)
105 #define section_offs(iova) (iova & (SECT_SIZE - 1))
106 #define lpage_phys(pent) (sect_to_phys(*(pent)) & LPAGE_MASK)
107 #define lpage_offs(iova) (iova & (LPAGE_SIZE - 1))
108 #define spage_phys(pent) (sect_to_phys(*(pent)) & SPAGE_MASK)
109 #define spage_offs(iova) (iova & (SPAGE_SIZE - 1))
110 
111 #define NUM_LV1ENTRIES 4096
112 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
113 
114 static u32 lv1ent_offset(sysmmu_iova_t iova)
115 {
116 	return iova >> SECT_ORDER;
117 }
118 
119 static u32 lv2ent_offset(sysmmu_iova_t iova)
120 {
121 	return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
122 }
123 
124 #define LV1TABLE_SIZE (NUM_LV1ENTRIES * sizeof(sysmmu_pte_t))
125 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
126 
127 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
128 #define lv2table_base(sent) (sect_to_phys(*(sent) & 0xFFFFFFC0))
129 
130 #define mk_lv1ent_sect(pa, prot) ((pa >> PG_ENT_SHIFT) | LV1_PROT[prot] | 2)
131 #define mk_lv1ent_page(pa) ((pa >> PG_ENT_SHIFT) | 1)
132 #define mk_lv2ent_lpage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 1)
133 #define mk_lv2ent_spage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 2)
134 
135 #define CTRL_ENABLE	0x5
136 #define CTRL_BLOCK	0x7
137 #define CTRL_DISABLE	0x0
138 
139 #define CFG_LRU		0x1
140 #define CFG_EAP		(1 << 2)
141 #define CFG_QOS(n)	((n & 0xF) << 7)
142 #define CFG_ACGEN	(1 << 24) /* System MMU 3.3 only */
143 #define CFG_SYSSEL	(1 << 22) /* System MMU 3.2 only */
144 #define CFG_FLPDCACHE	(1 << 20) /* System MMU 3.2+ only */
145 
146 /* common registers */
147 #define REG_MMU_CTRL		0x000
148 #define REG_MMU_CFG		0x004
149 #define REG_MMU_STATUS		0x008
150 #define REG_MMU_VERSION		0x034
151 
152 #define MMU_MAJ_VER(val)	((val) >> 7)
153 #define MMU_MIN_VER(val)	((val) & 0x7F)
154 #define MMU_RAW_VER(reg)	(((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
155 
156 #define MAKE_MMU_VER(maj, min)	((((maj) & 0xF) << 7) | ((min) & 0x7F))
157 
158 /* v1.x - v3.x registers */
159 #define REG_MMU_FLUSH		0x00C
160 #define REG_MMU_FLUSH_ENTRY	0x010
161 #define REG_PT_BASE_ADDR	0x014
162 #define REG_INT_STATUS		0x018
163 #define REG_INT_CLEAR		0x01C
164 
165 #define REG_PAGE_FAULT_ADDR	0x024
166 #define REG_AW_FAULT_ADDR	0x028
167 #define REG_AR_FAULT_ADDR	0x02C
168 #define REG_DEFAULT_SLAVE_ADDR	0x030
169 
170 /* v5.x registers */
171 #define REG_V5_PT_BASE_PFN	0x00C
172 #define REG_V5_MMU_FLUSH_ALL	0x010
173 #define REG_V5_MMU_FLUSH_ENTRY	0x014
174 #define REG_V5_MMU_FLUSH_RANGE	0x018
175 #define REG_V5_MMU_FLUSH_START	0x020
176 #define REG_V5_MMU_FLUSH_END	0x024
177 #define REG_V5_INT_STATUS	0x060
178 #define REG_V5_INT_CLEAR	0x064
179 #define REG_V5_FAULT_AR_VA	0x070
180 #define REG_V5_FAULT_AW_VA	0x080
181 
182 #define has_sysmmu(dev)		(dev->archdata.iommu != NULL)
183 
184 static struct device *dma_dev;
185 static struct kmem_cache *lv2table_kmem_cache;
186 static sysmmu_pte_t *zero_lv2_table;
187 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
188 
189 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
190 {
191 	return pgtable + lv1ent_offset(iova);
192 }
193 
194 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
195 {
196 	return (sysmmu_pte_t *)phys_to_virt(
197 				lv2table_base(sent)) + lv2ent_offset(iova);
198 }
199 
200 /*
201  * IOMMU fault information register
202  */
203 struct sysmmu_fault_info {
204 	unsigned int bit;	/* bit number in STATUS register */
205 	unsigned short addr_reg; /* register to read VA fault address */
206 	const char *name;	/* human readable fault name */
207 	unsigned int type;	/* fault type for report_iommu_fault */
208 };
209 
210 static const struct sysmmu_fault_info sysmmu_faults[] = {
211 	{ 0, REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
212 	{ 1, REG_AR_FAULT_ADDR, "AR MULTI-HIT", IOMMU_FAULT_READ },
213 	{ 2, REG_AW_FAULT_ADDR, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
214 	{ 3, REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
215 	{ 4, REG_AR_FAULT_ADDR, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
216 	{ 5, REG_AR_FAULT_ADDR, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
217 	{ 6, REG_AW_FAULT_ADDR, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
218 	{ 7, REG_AW_FAULT_ADDR, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
219 };
220 
221 static const struct sysmmu_fault_info sysmmu_v5_faults[] = {
222 	{ 0, REG_V5_FAULT_AR_VA, "AR PTW", IOMMU_FAULT_READ },
223 	{ 1, REG_V5_FAULT_AR_VA, "AR PAGE", IOMMU_FAULT_READ },
224 	{ 2, REG_V5_FAULT_AR_VA, "AR MULTI-HIT", IOMMU_FAULT_READ },
225 	{ 3, REG_V5_FAULT_AR_VA, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
226 	{ 4, REG_V5_FAULT_AR_VA, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
227 	{ 16, REG_V5_FAULT_AW_VA, "AW PTW", IOMMU_FAULT_WRITE },
228 	{ 17, REG_V5_FAULT_AW_VA, "AW PAGE", IOMMU_FAULT_WRITE },
229 	{ 18, REG_V5_FAULT_AW_VA, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
230 	{ 19, REG_V5_FAULT_AW_VA, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
231 	{ 20, REG_V5_FAULT_AW_VA, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
232 };
233 
234 /*
235  * This structure is attached to dev.archdata.iommu of the master device
236  * on device add, contains a list of SYSMMU controllers defined by device tree,
237  * which are bound to given master device. It is usually referenced by 'owner'
238  * pointer.
239 */
240 struct exynos_iommu_owner {
241 	struct list_head controllers;	/* list of sysmmu_drvdata.owner_node */
242 	struct iommu_domain *domain;	/* domain this device is attached */
243 	struct mutex rpm_lock;		/* for runtime pm of all sysmmus */
244 };
245 
246 /*
247  * This structure exynos specific generalization of struct iommu_domain.
248  * It contains list of SYSMMU controllers from all master devices, which has
249  * been attached to this domain and page tables of IO address space defined by
250  * it. It is usually referenced by 'domain' pointer.
251  */
252 struct exynos_iommu_domain {
253 	struct list_head clients; /* list of sysmmu_drvdata.domain_node */
254 	sysmmu_pte_t *pgtable;	/* lv1 page table, 16KB */
255 	short *lv2entcnt;	/* free lv2 entry counter for each section */
256 	spinlock_t lock;	/* lock for modyfying list of clients */
257 	spinlock_t pgtablelock;	/* lock for modifying page table @ pgtable */
258 	struct iommu_domain domain; /* generic domain data structure */
259 };
260 
261 /*
262  * This structure hold all data of a single SYSMMU controller, this includes
263  * hw resources like registers and clocks, pointers and list nodes to connect
264  * it to all other structures, internal state and parameters read from device
265  * tree. It is usually referenced by 'data' pointer.
266  */
267 struct sysmmu_drvdata {
268 	struct device *sysmmu;		/* SYSMMU controller device */
269 	struct device *master;		/* master device (owner) */
270 	void __iomem *sfrbase;		/* our registers */
271 	struct clk *clk;		/* SYSMMU's clock */
272 	struct clk *aclk;		/* SYSMMU's aclk clock */
273 	struct clk *pclk;		/* SYSMMU's pclk clock */
274 	struct clk *clk_master;		/* master's device clock */
275 	spinlock_t lock;		/* lock for modyfying state */
276 	bool active;			/* current status */
277 	struct exynos_iommu_domain *domain; /* domain we belong to */
278 	struct list_head domain_node;	/* node for domain clients list */
279 	struct list_head owner_node;	/* node for owner controllers list */
280 	phys_addr_t pgtable;		/* assigned page table structure */
281 	unsigned int version;		/* our version */
282 
283 	struct iommu_device iommu;	/* IOMMU core handle */
284 };
285 
286 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
287 {
288 	return container_of(dom, struct exynos_iommu_domain, domain);
289 }
290 
291 static void sysmmu_unblock(struct sysmmu_drvdata *data)
292 {
293 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
294 }
295 
296 static bool sysmmu_block(struct sysmmu_drvdata *data)
297 {
298 	int i = 120;
299 
300 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
301 	while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
302 		--i;
303 
304 	if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
305 		sysmmu_unblock(data);
306 		return false;
307 	}
308 
309 	return true;
310 }
311 
312 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
313 {
314 	if (MMU_MAJ_VER(data->version) < 5)
315 		writel(0x1, data->sfrbase + REG_MMU_FLUSH);
316 	else
317 		writel(0x1, data->sfrbase + REG_V5_MMU_FLUSH_ALL);
318 }
319 
320 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
321 				sysmmu_iova_t iova, unsigned int num_inv)
322 {
323 	unsigned int i;
324 
325 	if (MMU_MAJ_VER(data->version) < 5) {
326 		for (i = 0; i < num_inv; i++) {
327 			writel((iova & SPAGE_MASK) | 1,
328 				     data->sfrbase + REG_MMU_FLUSH_ENTRY);
329 			iova += SPAGE_SIZE;
330 		}
331 	} else {
332 		if (num_inv == 1) {
333 			writel((iova & SPAGE_MASK) | 1,
334 				     data->sfrbase + REG_V5_MMU_FLUSH_ENTRY);
335 		} else {
336 			writel((iova & SPAGE_MASK),
337 				     data->sfrbase + REG_V5_MMU_FLUSH_START);
338 			writel((iova & SPAGE_MASK) + (num_inv - 1) * SPAGE_SIZE,
339 				     data->sfrbase + REG_V5_MMU_FLUSH_END);
340 			writel(1, data->sfrbase + REG_V5_MMU_FLUSH_RANGE);
341 		}
342 	}
343 }
344 
345 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
346 {
347 	if (MMU_MAJ_VER(data->version) < 5)
348 		writel(pgd, data->sfrbase + REG_PT_BASE_ADDR);
349 	else
350 		writel(pgd >> PAGE_SHIFT,
351 			     data->sfrbase + REG_V5_PT_BASE_PFN);
352 
353 	__sysmmu_tlb_invalidate(data);
354 }
355 
356 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
357 {
358 	BUG_ON(clk_prepare_enable(data->clk_master));
359 	BUG_ON(clk_prepare_enable(data->clk));
360 	BUG_ON(clk_prepare_enable(data->pclk));
361 	BUG_ON(clk_prepare_enable(data->aclk));
362 }
363 
364 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
365 {
366 	clk_disable_unprepare(data->aclk);
367 	clk_disable_unprepare(data->pclk);
368 	clk_disable_unprepare(data->clk);
369 	clk_disable_unprepare(data->clk_master);
370 }
371 
372 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
373 {
374 	u32 ver;
375 
376 	__sysmmu_enable_clocks(data);
377 
378 	ver = readl(data->sfrbase + REG_MMU_VERSION);
379 
380 	/* controllers on some SoCs don't report proper version */
381 	if (ver == 0x80000001u)
382 		data->version = MAKE_MMU_VER(1, 0);
383 	else
384 		data->version = MMU_RAW_VER(ver);
385 
386 	dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
387 		MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
388 
389 	__sysmmu_disable_clocks(data);
390 }
391 
392 static void show_fault_information(struct sysmmu_drvdata *data,
393 				   const struct sysmmu_fault_info *finfo,
394 				   sysmmu_iova_t fault_addr)
395 {
396 	sysmmu_pte_t *ent;
397 
398 	dev_err(data->sysmmu, "%s: %s FAULT occurred at %#x\n",
399 		dev_name(data->master), finfo->name, fault_addr);
400 	dev_dbg(data->sysmmu, "Page table base: %pa\n", &data->pgtable);
401 	ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
402 	dev_dbg(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
403 	if (lv1ent_page(ent)) {
404 		ent = page_entry(ent, fault_addr);
405 		dev_dbg(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
406 	}
407 }
408 
409 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
410 {
411 	/* SYSMMU is in blocked state when interrupt occurred. */
412 	struct sysmmu_drvdata *data = dev_id;
413 	const struct sysmmu_fault_info *finfo;
414 	unsigned int i, n, itype;
415 	sysmmu_iova_t fault_addr = -1;
416 	unsigned short reg_status, reg_clear;
417 	int ret = -ENOSYS;
418 
419 	WARN_ON(!data->active);
420 
421 	if (MMU_MAJ_VER(data->version) < 5) {
422 		reg_status = REG_INT_STATUS;
423 		reg_clear = REG_INT_CLEAR;
424 		finfo = sysmmu_faults;
425 		n = ARRAY_SIZE(sysmmu_faults);
426 	} else {
427 		reg_status = REG_V5_INT_STATUS;
428 		reg_clear = REG_V5_INT_CLEAR;
429 		finfo = sysmmu_v5_faults;
430 		n = ARRAY_SIZE(sysmmu_v5_faults);
431 	}
432 
433 	spin_lock(&data->lock);
434 
435 	clk_enable(data->clk_master);
436 
437 	itype = __ffs(readl(data->sfrbase + reg_status));
438 	for (i = 0; i < n; i++, finfo++)
439 		if (finfo->bit == itype)
440 			break;
441 	/* unknown/unsupported fault */
442 	BUG_ON(i == n);
443 
444 	/* print debug message */
445 	fault_addr = readl(data->sfrbase + finfo->addr_reg);
446 	show_fault_information(data, finfo, fault_addr);
447 
448 	if (data->domain)
449 		ret = report_iommu_fault(&data->domain->domain,
450 					data->master, fault_addr, finfo->type);
451 	/* fault is not recovered by fault handler */
452 	BUG_ON(ret != 0);
453 
454 	writel(1 << itype, data->sfrbase + reg_clear);
455 
456 	sysmmu_unblock(data);
457 
458 	clk_disable(data->clk_master);
459 
460 	spin_unlock(&data->lock);
461 
462 	return IRQ_HANDLED;
463 }
464 
465 static void __sysmmu_disable(struct sysmmu_drvdata *data)
466 {
467 	unsigned long flags;
468 
469 	clk_enable(data->clk_master);
470 
471 	spin_lock_irqsave(&data->lock, flags);
472 	writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
473 	writel(0, data->sfrbase + REG_MMU_CFG);
474 	data->active = false;
475 	spin_unlock_irqrestore(&data->lock, flags);
476 
477 	__sysmmu_disable_clocks(data);
478 }
479 
480 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
481 {
482 	unsigned int cfg;
483 
484 	if (data->version <= MAKE_MMU_VER(3, 1))
485 		cfg = CFG_LRU | CFG_QOS(15);
486 	else if (data->version <= MAKE_MMU_VER(3, 2))
487 		cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
488 	else
489 		cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
490 
491 	cfg |= CFG_EAP; /* enable access protection bits check */
492 
493 	writel(cfg, data->sfrbase + REG_MMU_CFG);
494 }
495 
496 static void __sysmmu_enable(struct sysmmu_drvdata *data)
497 {
498 	unsigned long flags;
499 
500 	__sysmmu_enable_clocks(data);
501 
502 	spin_lock_irqsave(&data->lock, flags);
503 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
504 	__sysmmu_init_config(data);
505 	__sysmmu_set_ptbase(data, data->pgtable);
506 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
507 	data->active = true;
508 	spin_unlock_irqrestore(&data->lock, flags);
509 
510 	/*
511 	 * SYSMMU driver keeps master's clock enabled only for the short
512 	 * time, while accessing the registers. For performing address
513 	 * translation during DMA transaction it relies on the client
514 	 * driver to enable it.
515 	 */
516 	clk_disable(data->clk_master);
517 }
518 
519 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
520 					    sysmmu_iova_t iova)
521 {
522 	unsigned long flags;
523 
524 	spin_lock_irqsave(&data->lock, flags);
525 	if (data->active && data->version >= MAKE_MMU_VER(3, 3)) {
526 		clk_enable(data->clk_master);
527 		if (sysmmu_block(data)) {
528 			if (data->version >= MAKE_MMU_VER(5, 0))
529 				__sysmmu_tlb_invalidate(data);
530 			else
531 				__sysmmu_tlb_invalidate_entry(data, iova, 1);
532 			sysmmu_unblock(data);
533 		}
534 		clk_disable(data->clk_master);
535 	}
536 	spin_unlock_irqrestore(&data->lock, flags);
537 }
538 
539 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
540 					sysmmu_iova_t iova, size_t size)
541 {
542 	unsigned long flags;
543 
544 	spin_lock_irqsave(&data->lock, flags);
545 	if (data->active) {
546 		unsigned int num_inv = 1;
547 
548 		clk_enable(data->clk_master);
549 
550 		/*
551 		 * L2TLB invalidation required
552 		 * 4KB page: 1 invalidation
553 		 * 64KB page: 16 invalidations
554 		 * 1MB page: 64 invalidations
555 		 * because it is set-associative TLB
556 		 * with 8-way and 64 sets.
557 		 * 1MB page can be cached in one of all sets.
558 		 * 64KB page can be one of 16 consecutive sets.
559 		 */
560 		if (MMU_MAJ_VER(data->version) == 2)
561 			num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
562 
563 		if (sysmmu_block(data)) {
564 			__sysmmu_tlb_invalidate_entry(data, iova, num_inv);
565 			sysmmu_unblock(data);
566 		}
567 		clk_disable(data->clk_master);
568 	}
569 	spin_unlock_irqrestore(&data->lock, flags);
570 }
571 
572 static struct iommu_ops exynos_iommu_ops;
573 
574 static int __init exynos_sysmmu_probe(struct platform_device *pdev)
575 {
576 	int irq, ret;
577 	struct device *dev = &pdev->dev;
578 	struct sysmmu_drvdata *data;
579 	struct resource *res;
580 
581 	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
582 	if (!data)
583 		return -ENOMEM;
584 
585 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
586 	data->sfrbase = devm_ioremap_resource(dev, res);
587 	if (IS_ERR(data->sfrbase))
588 		return PTR_ERR(data->sfrbase);
589 
590 	irq = platform_get_irq(pdev, 0);
591 	if (irq <= 0) {
592 		dev_err(dev, "Unable to find IRQ resource\n");
593 		return irq;
594 	}
595 
596 	ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
597 				dev_name(dev), data);
598 	if (ret) {
599 		dev_err(dev, "Unabled to register handler of irq %d\n", irq);
600 		return ret;
601 	}
602 
603 	data->clk = devm_clk_get(dev, "sysmmu");
604 	if (PTR_ERR(data->clk) == -ENOENT)
605 		data->clk = NULL;
606 	else if (IS_ERR(data->clk))
607 		return PTR_ERR(data->clk);
608 
609 	data->aclk = devm_clk_get(dev, "aclk");
610 	if (PTR_ERR(data->aclk) == -ENOENT)
611 		data->aclk = NULL;
612 	else if (IS_ERR(data->aclk))
613 		return PTR_ERR(data->aclk);
614 
615 	data->pclk = devm_clk_get(dev, "pclk");
616 	if (PTR_ERR(data->pclk) == -ENOENT)
617 		data->pclk = NULL;
618 	else if (IS_ERR(data->pclk))
619 		return PTR_ERR(data->pclk);
620 
621 	if (!data->clk && (!data->aclk || !data->pclk)) {
622 		dev_err(dev, "Failed to get device clock(s)!\n");
623 		return -ENOSYS;
624 	}
625 
626 	data->clk_master = devm_clk_get(dev, "master");
627 	if (PTR_ERR(data->clk_master) == -ENOENT)
628 		data->clk_master = NULL;
629 	else if (IS_ERR(data->clk_master))
630 		return PTR_ERR(data->clk_master);
631 
632 	data->sysmmu = dev;
633 	spin_lock_init(&data->lock);
634 
635 	ret = iommu_device_sysfs_add(&data->iommu, &pdev->dev, NULL,
636 				     dev_name(data->sysmmu));
637 	if (ret)
638 		return ret;
639 
640 	iommu_device_set_ops(&data->iommu, &exynos_iommu_ops);
641 	iommu_device_set_fwnode(&data->iommu, &dev->of_node->fwnode);
642 
643 	ret = iommu_device_register(&data->iommu);
644 	if (ret)
645 		return ret;
646 
647 	platform_set_drvdata(pdev, data);
648 
649 	__sysmmu_get_version(data);
650 	if (PG_ENT_SHIFT < 0) {
651 		if (MMU_MAJ_VER(data->version) < 5) {
652 			PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
653 			LV1_PROT = SYSMMU_LV1_PROT;
654 			LV2_PROT = SYSMMU_LV2_PROT;
655 		} else {
656 			PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
657 			LV1_PROT = SYSMMU_V5_LV1_PROT;
658 			LV2_PROT = SYSMMU_V5_LV2_PROT;
659 		}
660 	}
661 
662 	pm_runtime_enable(dev);
663 
664 	return 0;
665 }
666 
667 static int __maybe_unused exynos_sysmmu_suspend(struct device *dev)
668 {
669 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
670 	struct device *master = data->master;
671 
672 	if (master) {
673 		struct exynos_iommu_owner *owner = master->archdata.iommu;
674 
675 		mutex_lock(&owner->rpm_lock);
676 		if (data->domain) {
677 			dev_dbg(data->sysmmu, "saving state\n");
678 			__sysmmu_disable(data);
679 		}
680 		mutex_unlock(&owner->rpm_lock);
681 	}
682 	return 0;
683 }
684 
685 static int __maybe_unused exynos_sysmmu_resume(struct device *dev)
686 {
687 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
688 	struct device *master = data->master;
689 
690 	if (master) {
691 		struct exynos_iommu_owner *owner = master->archdata.iommu;
692 
693 		mutex_lock(&owner->rpm_lock);
694 		if (data->domain) {
695 			dev_dbg(data->sysmmu, "restoring state\n");
696 			__sysmmu_enable(data);
697 		}
698 		mutex_unlock(&owner->rpm_lock);
699 	}
700 	return 0;
701 }
702 
703 static const struct dev_pm_ops sysmmu_pm_ops = {
704 	SET_RUNTIME_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume, NULL)
705 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
706 				pm_runtime_force_resume)
707 };
708 
709 static const struct of_device_id sysmmu_of_match[] __initconst = {
710 	{ .compatible	= "samsung,exynos-sysmmu", },
711 	{ },
712 };
713 
714 static struct platform_driver exynos_sysmmu_driver __refdata = {
715 	.probe	= exynos_sysmmu_probe,
716 	.driver	= {
717 		.name		= "exynos-sysmmu",
718 		.of_match_table	= sysmmu_of_match,
719 		.pm		= &sysmmu_pm_ops,
720 		.suppress_bind_attrs = true,
721 	}
722 };
723 
724 static inline void update_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
725 {
726 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
727 				DMA_TO_DEVICE);
728 	*ent = cpu_to_le32(val);
729 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
730 				   DMA_TO_DEVICE);
731 }
732 
733 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
734 {
735 	struct exynos_iommu_domain *domain;
736 	dma_addr_t handle;
737 	int i;
738 
739 	/* Check if correct PTE offsets are initialized */
740 	BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
741 
742 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
743 	if (!domain)
744 		return NULL;
745 
746 	if (type == IOMMU_DOMAIN_DMA) {
747 		if (iommu_get_dma_cookie(&domain->domain) != 0)
748 			goto err_pgtable;
749 	} else if (type != IOMMU_DOMAIN_UNMANAGED) {
750 		goto err_pgtable;
751 	}
752 
753 	domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
754 	if (!domain->pgtable)
755 		goto err_dma_cookie;
756 
757 	domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
758 	if (!domain->lv2entcnt)
759 		goto err_counter;
760 
761 	/* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
762 	for (i = 0; i < NUM_LV1ENTRIES; i++)
763 		domain->pgtable[i] = ZERO_LV2LINK;
764 
765 	handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
766 				DMA_TO_DEVICE);
767 	/* For mapping page table entries we rely on dma == phys */
768 	BUG_ON(handle != virt_to_phys(domain->pgtable));
769 	if (dma_mapping_error(dma_dev, handle))
770 		goto err_lv2ent;
771 
772 	spin_lock_init(&domain->lock);
773 	spin_lock_init(&domain->pgtablelock);
774 	INIT_LIST_HEAD(&domain->clients);
775 
776 	domain->domain.geometry.aperture_start = 0;
777 	domain->domain.geometry.aperture_end   = ~0UL;
778 	domain->domain.geometry.force_aperture = true;
779 
780 	return &domain->domain;
781 
782 err_lv2ent:
783 	free_pages((unsigned long)domain->lv2entcnt, 1);
784 err_counter:
785 	free_pages((unsigned long)domain->pgtable, 2);
786 err_dma_cookie:
787 	if (type == IOMMU_DOMAIN_DMA)
788 		iommu_put_dma_cookie(&domain->domain);
789 err_pgtable:
790 	kfree(domain);
791 	return NULL;
792 }
793 
794 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
795 {
796 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
797 	struct sysmmu_drvdata *data, *next;
798 	unsigned long flags;
799 	int i;
800 
801 	WARN_ON(!list_empty(&domain->clients));
802 
803 	spin_lock_irqsave(&domain->lock, flags);
804 
805 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
806 		spin_lock(&data->lock);
807 		__sysmmu_disable(data);
808 		data->pgtable = 0;
809 		data->domain = NULL;
810 		list_del_init(&data->domain_node);
811 		spin_unlock(&data->lock);
812 	}
813 
814 	spin_unlock_irqrestore(&domain->lock, flags);
815 
816 	if (iommu_domain->type == IOMMU_DOMAIN_DMA)
817 		iommu_put_dma_cookie(iommu_domain);
818 
819 	dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
820 			 DMA_TO_DEVICE);
821 
822 	for (i = 0; i < NUM_LV1ENTRIES; i++)
823 		if (lv1ent_page(domain->pgtable + i)) {
824 			phys_addr_t base = lv2table_base(domain->pgtable + i);
825 
826 			dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
827 					 DMA_TO_DEVICE);
828 			kmem_cache_free(lv2table_kmem_cache,
829 					phys_to_virt(base));
830 		}
831 
832 	free_pages((unsigned long)domain->pgtable, 2);
833 	free_pages((unsigned long)domain->lv2entcnt, 1);
834 	kfree(domain);
835 }
836 
837 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
838 				    struct device *dev)
839 {
840 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
841 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
842 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
843 	struct sysmmu_drvdata *data, *next;
844 	unsigned long flags;
845 
846 	if (!has_sysmmu(dev) || owner->domain != iommu_domain)
847 		return;
848 
849 	mutex_lock(&owner->rpm_lock);
850 
851 	list_for_each_entry(data, &owner->controllers, owner_node) {
852 		pm_runtime_get_noresume(data->sysmmu);
853 		if (pm_runtime_active(data->sysmmu))
854 			__sysmmu_disable(data);
855 		pm_runtime_put(data->sysmmu);
856 	}
857 
858 	spin_lock_irqsave(&domain->lock, flags);
859 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
860 		spin_lock(&data->lock);
861 		data->pgtable = 0;
862 		data->domain = NULL;
863 		list_del_init(&data->domain_node);
864 		spin_unlock(&data->lock);
865 	}
866 	owner->domain = NULL;
867 	spin_unlock_irqrestore(&domain->lock, flags);
868 
869 	mutex_unlock(&owner->rpm_lock);
870 
871 	dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n", __func__,
872 		&pagetable);
873 }
874 
875 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
876 				   struct device *dev)
877 {
878 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
879 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
880 	struct sysmmu_drvdata *data;
881 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
882 	unsigned long flags;
883 
884 	if (!has_sysmmu(dev))
885 		return -ENODEV;
886 
887 	if (owner->domain)
888 		exynos_iommu_detach_device(owner->domain, dev);
889 
890 	mutex_lock(&owner->rpm_lock);
891 
892 	spin_lock_irqsave(&domain->lock, flags);
893 	list_for_each_entry(data, &owner->controllers, owner_node) {
894 		spin_lock(&data->lock);
895 		data->pgtable = pagetable;
896 		data->domain = domain;
897 		list_add_tail(&data->domain_node, &domain->clients);
898 		spin_unlock(&data->lock);
899 	}
900 	owner->domain = iommu_domain;
901 	spin_unlock_irqrestore(&domain->lock, flags);
902 
903 	list_for_each_entry(data, &owner->controllers, owner_node) {
904 		pm_runtime_get_noresume(data->sysmmu);
905 		if (pm_runtime_active(data->sysmmu))
906 			__sysmmu_enable(data);
907 		pm_runtime_put(data->sysmmu);
908 	}
909 
910 	mutex_unlock(&owner->rpm_lock);
911 
912 	dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa\n", __func__,
913 		&pagetable);
914 
915 	return 0;
916 }
917 
918 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
919 		sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
920 {
921 	if (lv1ent_section(sent)) {
922 		WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
923 		return ERR_PTR(-EADDRINUSE);
924 	}
925 
926 	if (lv1ent_fault(sent)) {
927 		dma_addr_t handle;
928 		sysmmu_pte_t *pent;
929 		bool need_flush_flpd_cache = lv1ent_zero(sent);
930 
931 		pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
932 		BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
933 		if (!pent)
934 			return ERR_PTR(-ENOMEM);
935 
936 		update_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
937 		kmemleak_ignore(pent);
938 		*pgcounter = NUM_LV2ENTRIES;
939 		handle = dma_map_single(dma_dev, pent, LV2TABLE_SIZE,
940 					DMA_TO_DEVICE);
941 		if (dma_mapping_error(dma_dev, handle)) {
942 			kmem_cache_free(lv2table_kmem_cache, pent);
943 			return ERR_PTR(-EADDRINUSE);
944 		}
945 
946 		/*
947 		 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
948 		 * FLPD cache may cache the address of zero_l2_table. This
949 		 * function replaces the zero_l2_table with new L2 page table
950 		 * to write valid mappings.
951 		 * Accessing the valid area may cause page fault since FLPD
952 		 * cache may still cache zero_l2_table for the valid area
953 		 * instead of new L2 page table that has the mapping
954 		 * information of the valid area.
955 		 * Thus any replacement of zero_l2_table with other valid L2
956 		 * page table must involve FLPD cache invalidation for System
957 		 * MMU v3.3.
958 		 * FLPD cache invalidation is performed with TLB invalidation
959 		 * by VPN without blocking. It is safe to invalidate TLB without
960 		 * blocking because the target address of TLB invalidation is
961 		 * not currently mapped.
962 		 */
963 		if (need_flush_flpd_cache) {
964 			struct sysmmu_drvdata *data;
965 
966 			spin_lock(&domain->lock);
967 			list_for_each_entry(data, &domain->clients, domain_node)
968 				sysmmu_tlb_invalidate_flpdcache(data, iova);
969 			spin_unlock(&domain->lock);
970 		}
971 	}
972 
973 	return page_entry(sent, iova);
974 }
975 
976 static int lv1set_section(struct exynos_iommu_domain *domain,
977 			  sysmmu_pte_t *sent, sysmmu_iova_t iova,
978 			  phys_addr_t paddr, int prot, short *pgcnt)
979 {
980 	if (lv1ent_section(sent)) {
981 		WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
982 			iova);
983 		return -EADDRINUSE;
984 	}
985 
986 	if (lv1ent_page(sent)) {
987 		if (*pgcnt != NUM_LV2ENTRIES) {
988 			WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
989 				iova);
990 			return -EADDRINUSE;
991 		}
992 
993 		kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
994 		*pgcnt = 0;
995 	}
996 
997 	update_pte(sent, mk_lv1ent_sect(paddr, prot));
998 
999 	spin_lock(&domain->lock);
1000 	if (lv1ent_page_zero(sent)) {
1001 		struct sysmmu_drvdata *data;
1002 		/*
1003 		 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
1004 		 * entry by speculative prefetch of SLPD which has no mapping.
1005 		 */
1006 		list_for_each_entry(data, &domain->clients, domain_node)
1007 			sysmmu_tlb_invalidate_flpdcache(data, iova);
1008 	}
1009 	spin_unlock(&domain->lock);
1010 
1011 	return 0;
1012 }
1013 
1014 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
1015 		       int prot, short *pgcnt)
1016 {
1017 	if (size == SPAGE_SIZE) {
1018 		if (WARN_ON(!lv2ent_fault(pent)))
1019 			return -EADDRINUSE;
1020 
1021 		update_pte(pent, mk_lv2ent_spage(paddr, prot));
1022 		*pgcnt -= 1;
1023 	} else { /* size == LPAGE_SIZE */
1024 		int i;
1025 		dma_addr_t pent_base = virt_to_phys(pent);
1026 
1027 		dma_sync_single_for_cpu(dma_dev, pent_base,
1028 					sizeof(*pent) * SPAGES_PER_LPAGE,
1029 					DMA_TO_DEVICE);
1030 		for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1031 			if (WARN_ON(!lv2ent_fault(pent))) {
1032 				if (i > 0)
1033 					memset(pent - i, 0, sizeof(*pent) * i);
1034 				return -EADDRINUSE;
1035 			}
1036 
1037 			*pent = mk_lv2ent_lpage(paddr, prot);
1038 		}
1039 		dma_sync_single_for_device(dma_dev, pent_base,
1040 					   sizeof(*pent) * SPAGES_PER_LPAGE,
1041 					   DMA_TO_DEVICE);
1042 		*pgcnt -= SPAGES_PER_LPAGE;
1043 	}
1044 
1045 	return 0;
1046 }
1047 
1048 /*
1049  * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1050  *
1051  * System MMU v3.x has advanced logic to improve address translation
1052  * performance with caching more page table entries by a page table walk.
1053  * However, the logic has a bug that while caching faulty page table entries,
1054  * System MMU reports page fault if the cached fault entry is hit even though
1055  * the fault entry is updated to a valid entry after the entry is cached.
1056  * To prevent caching faulty page table entries which may be updated to valid
1057  * entries later, the virtual memory manager should care about the workaround
1058  * for the problem. The following describes the workaround.
1059  *
1060  * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1061  * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1062  *
1063  * Precisely, any start address of I/O virtual region must be aligned with
1064  * the following sizes for System MMU v3.1 and v3.2.
1065  * System MMU v3.1: 128KiB
1066  * System MMU v3.2: 256KiB
1067  *
1068  * Because System MMU v3.3 caches page table entries more aggressively, it needs
1069  * more workarounds.
1070  * - Any two consecutive I/O virtual regions must have a hole of size larger
1071  *   than or equal to 128KiB.
1072  * - Start address of an I/O virtual region must be aligned by 128KiB.
1073  */
1074 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1075 			    unsigned long l_iova, phys_addr_t paddr, size_t size,
1076 			    int prot)
1077 {
1078 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1079 	sysmmu_pte_t *entry;
1080 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1081 	unsigned long flags;
1082 	int ret = -ENOMEM;
1083 
1084 	BUG_ON(domain->pgtable == NULL);
1085 	prot &= SYSMMU_SUPPORTED_PROT_BITS;
1086 
1087 	spin_lock_irqsave(&domain->pgtablelock, flags);
1088 
1089 	entry = section_entry(domain->pgtable, iova);
1090 
1091 	if (size == SECT_SIZE) {
1092 		ret = lv1set_section(domain, entry, iova, paddr, prot,
1093 				     &domain->lv2entcnt[lv1ent_offset(iova)]);
1094 	} else {
1095 		sysmmu_pte_t *pent;
1096 
1097 		pent = alloc_lv2entry(domain, entry, iova,
1098 				      &domain->lv2entcnt[lv1ent_offset(iova)]);
1099 
1100 		if (IS_ERR(pent))
1101 			ret = PTR_ERR(pent);
1102 		else
1103 			ret = lv2set_page(pent, paddr, size, prot,
1104 				       &domain->lv2entcnt[lv1ent_offset(iova)]);
1105 	}
1106 
1107 	if (ret)
1108 		pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1109 			__func__, ret, size, iova);
1110 
1111 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1112 
1113 	return ret;
1114 }
1115 
1116 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1117 					      sysmmu_iova_t iova, size_t size)
1118 {
1119 	struct sysmmu_drvdata *data;
1120 	unsigned long flags;
1121 
1122 	spin_lock_irqsave(&domain->lock, flags);
1123 
1124 	list_for_each_entry(data, &domain->clients, domain_node)
1125 		sysmmu_tlb_invalidate_entry(data, iova, size);
1126 
1127 	spin_unlock_irqrestore(&domain->lock, flags);
1128 }
1129 
1130 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1131 				 unsigned long l_iova, size_t size)
1132 {
1133 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1134 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1135 	sysmmu_pte_t *ent;
1136 	size_t err_pgsize;
1137 	unsigned long flags;
1138 
1139 	BUG_ON(domain->pgtable == NULL);
1140 
1141 	spin_lock_irqsave(&domain->pgtablelock, flags);
1142 
1143 	ent = section_entry(domain->pgtable, iova);
1144 
1145 	if (lv1ent_section(ent)) {
1146 		if (WARN_ON(size < SECT_SIZE)) {
1147 			err_pgsize = SECT_SIZE;
1148 			goto err;
1149 		}
1150 
1151 		/* workaround for h/w bug in System MMU v3.3 */
1152 		update_pte(ent, ZERO_LV2LINK);
1153 		size = SECT_SIZE;
1154 		goto done;
1155 	}
1156 
1157 	if (unlikely(lv1ent_fault(ent))) {
1158 		if (size > SECT_SIZE)
1159 			size = SECT_SIZE;
1160 		goto done;
1161 	}
1162 
1163 	/* lv1ent_page(sent) == true here */
1164 
1165 	ent = page_entry(ent, iova);
1166 
1167 	if (unlikely(lv2ent_fault(ent))) {
1168 		size = SPAGE_SIZE;
1169 		goto done;
1170 	}
1171 
1172 	if (lv2ent_small(ent)) {
1173 		update_pte(ent, 0);
1174 		size = SPAGE_SIZE;
1175 		domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1176 		goto done;
1177 	}
1178 
1179 	/* lv1ent_large(ent) == true here */
1180 	if (WARN_ON(size < LPAGE_SIZE)) {
1181 		err_pgsize = LPAGE_SIZE;
1182 		goto err;
1183 	}
1184 
1185 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1186 				sizeof(*ent) * SPAGES_PER_LPAGE,
1187 				DMA_TO_DEVICE);
1188 	memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1189 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1190 				   sizeof(*ent) * SPAGES_PER_LPAGE,
1191 				   DMA_TO_DEVICE);
1192 	size = LPAGE_SIZE;
1193 	domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1194 done:
1195 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1196 
1197 	exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1198 
1199 	return size;
1200 err:
1201 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1202 
1203 	pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1204 		__func__, size, iova, err_pgsize);
1205 
1206 	return 0;
1207 }
1208 
1209 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1210 					  dma_addr_t iova)
1211 {
1212 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1213 	sysmmu_pte_t *entry;
1214 	unsigned long flags;
1215 	phys_addr_t phys = 0;
1216 
1217 	spin_lock_irqsave(&domain->pgtablelock, flags);
1218 
1219 	entry = section_entry(domain->pgtable, iova);
1220 
1221 	if (lv1ent_section(entry)) {
1222 		phys = section_phys(entry) + section_offs(iova);
1223 	} else if (lv1ent_page(entry)) {
1224 		entry = page_entry(entry, iova);
1225 
1226 		if (lv2ent_large(entry))
1227 			phys = lpage_phys(entry) + lpage_offs(iova);
1228 		else if (lv2ent_small(entry))
1229 			phys = spage_phys(entry) + spage_offs(iova);
1230 	}
1231 
1232 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1233 
1234 	return phys;
1235 }
1236 
1237 static struct iommu_group *get_device_iommu_group(struct device *dev)
1238 {
1239 	struct iommu_group *group;
1240 
1241 	group = iommu_group_get(dev);
1242 	if (!group)
1243 		group = iommu_group_alloc();
1244 
1245 	return group;
1246 }
1247 
1248 static int exynos_iommu_add_device(struct device *dev)
1249 {
1250 	struct iommu_group *group;
1251 
1252 	if (!has_sysmmu(dev))
1253 		return -ENODEV;
1254 
1255 	group = iommu_group_get_for_dev(dev);
1256 
1257 	if (IS_ERR(group))
1258 		return PTR_ERR(group);
1259 
1260 	iommu_group_put(group);
1261 
1262 	return 0;
1263 }
1264 
1265 static void exynos_iommu_remove_device(struct device *dev)
1266 {
1267 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1268 
1269 	if (!has_sysmmu(dev))
1270 		return;
1271 
1272 	if (owner->domain) {
1273 		struct iommu_group *group = iommu_group_get(dev);
1274 
1275 		if (group) {
1276 			WARN_ON(owner->domain !=
1277 				iommu_group_default_domain(group));
1278 			exynos_iommu_detach_device(owner->domain, dev);
1279 			iommu_group_put(group);
1280 		}
1281 	}
1282 	iommu_group_remove_device(dev);
1283 }
1284 
1285 static int exynos_iommu_of_xlate(struct device *dev,
1286 				 struct of_phandle_args *spec)
1287 {
1288 	struct exynos_iommu_owner *owner = dev->archdata.iommu;
1289 	struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1290 	struct sysmmu_drvdata *data, *entry;
1291 
1292 	if (!sysmmu)
1293 		return -ENODEV;
1294 
1295 	data = platform_get_drvdata(sysmmu);
1296 	if (!data)
1297 		return -ENODEV;
1298 
1299 	if (!owner) {
1300 		owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1301 		if (!owner)
1302 			return -ENOMEM;
1303 
1304 		INIT_LIST_HEAD(&owner->controllers);
1305 		mutex_init(&owner->rpm_lock);
1306 		dev->archdata.iommu = owner;
1307 	}
1308 
1309 	list_for_each_entry(entry, &owner->controllers, owner_node)
1310 		if (entry == data)
1311 			return 0;
1312 
1313 	list_add_tail(&data->owner_node, &owner->controllers);
1314 	data->master = dev;
1315 
1316 	/*
1317 	 * SYSMMU will be runtime activated via device link (dependency) to its
1318 	 * master device, so there are no direct calls to pm_runtime_get/put
1319 	 * in this driver.
1320 	 */
1321 	device_link_add(dev, data->sysmmu, DL_FLAG_PM_RUNTIME);
1322 
1323 	return 0;
1324 }
1325 
1326 static struct iommu_ops exynos_iommu_ops = {
1327 	.domain_alloc = exynos_iommu_domain_alloc,
1328 	.domain_free = exynos_iommu_domain_free,
1329 	.attach_dev = exynos_iommu_attach_device,
1330 	.detach_dev = exynos_iommu_detach_device,
1331 	.map = exynos_iommu_map,
1332 	.unmap = exynos_iommu_unmap,
1333 	.map_sg = default_iommu_map_sg,
1334 	.iova_to_phys = exynos_iommu_iova_to_phys,
1335 	.device_group = get_device_iommu_group,
1336 	.add_device = exynos_iommu_add_device,
1337 	.remove_device = exynos_iommu_remove_device,
1338 	.pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1339 	.of_xlate = exynos_iommu_of_xlate,
1340 };
1341 
1342 static bool init_done;
1343 
1344 static int __init exynos_iommu_init(void)
1345 {
1346 	int ret;
1347 
1348 	lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1349 				LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1350 	if (!lv2table_kmem_cache) {
1351 		pr_err("%s: Failed to create kmem cache\n", __func__);
1352 		return -ENOMEM;
1353 	}
1354 
1355 	ret = platform_driver_register(&exynos_sysmmu_driver);
1356 	if (ret) {
1357 		pr_err("%s: Failed to register driver\n", __func__);
1358 		goto err_reg_driver;
1359 	}
1360 
1361 	zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1362 	if (zero_lv2_table == NULL) {
1363 		pr_err("%s: Failed to allocate zero level2 page table\n",
1364 			__func__);
1365 		ret = -ENOMEM;
1366 		goto err_zero_lv2;
1367 	}
1368 
1369 	ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1370 	if (ret) {
1371 		pr_err("%s: Failed to register exynos-iommu driver.\n",
1372 								__func__);
1373 		goto err_set_iommu;
1374 	}
1375 
1376 	init_done = true;
1377 
1378 	return 0;
1379 err_set_iommu:
1380 	kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1381 err_zero_lv2:
1382 	platform_driver_unregister(&exynos_sysmmu_driver);
1383 err_reg_driver:
1384 	kmem_cache_destroy(lv2table_kmem_cache);
1385 	return ret;
1386 }
1387 
1388 static int __init exynos_iommu_of_setup(struct device_node *np)
1389 {
1390 	struct platform_device *pdev;
1391 
1392 	if (!init_done)
1393 		exynos_iommu_init();
1394 
1395 	pdev = of_platform_device_create(np, NULL, platform_bus_type.dev_root);
1396 	if (!pdev)
1397 		return -ENODEV;
1398 
1399 	/*
1400 	 * use the first registered sysmmu device for performing
1401 	 * dma mapping operations on iommu page tables (cpu cache flush)
1402 	 */
1403 	if (!dma_dev)
1404 		dma_dev = &pdev->dev;
1405 
1406 	return 0;
1407 }
1408 
1409 IOMMU_OF_DECLARE(exynos_iommu_of, "samsung,exynos-sysmmu",
1410 		 exynos_iommu_of_setup);
1411