xref: /openbmc/linux/drivers/iommu/exynos-iommu.c (revision 2d99a7ec)
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 #define CTRL_VM_ENABLE			BIT(0)
139 #define CTRL_VM_FAULT_MODE_STALL	BIT(3)
140 #define CAPA0_CAPA1_EXIST		BIT(11)
141 #define CAPA1_VCR_ENABLED		BIT(14)
142 
143 /* common registers */
144 #define REG_MMU_CTRL		0x000
145 #define REG_MMU_CFG		0x004
146 #define REG_MMU_STATUS		0x008
147 #define REG_MMU_VERSION		0x034
148 
149 #define MMU_MAJ_VER(val)	((val) >> 7)
150 #define MMU_MIN_VER(val)	((val) & 0x7F)
151 #define MMU_RAW_VER(reg)	(((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
152 
153 #define MAKE_MMU_VER(maj, min)	((((maj) & 0xF) << 7) | ((min) & 0x7F))
154 
155 /* v1.x - v3.x registers */
156 #define REG_PAGE_FAULT_ADDR	0x024
157 #define REG_AW_FAULT_ADDR	0x028
158 #define REG_AR_FAULT_ADDR	0x02C
159 #define REG_DEFAULT_SLAVE_ADDR	0x030
160 
161 /* v5.x registers */
162 #define REG_V5_FAULT_AR_VA	0x070
163 #define REG_V5_FAULT_AW_VA	0x080
164 
165 /* v7.x registers */
166 #define REG_V7_CAPA0		0x870
167 #define REG_V7_CAPA1		0x874
168 #define REG_V7_CTRL_VM		0x8000
169 
170 #define has_sysmmu(dev)		(dev_iommu_priv_get(dev) != NULL)
171 
172 static struct device *dma_dev;
173 static struct kmem_cache *lv2table_kmem_cache;
174 static sysmmu_pte_t *zero_lv2_table;
175 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
176 
177 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
178 {
179 	return pgtable + lv1ent_offset(iova);
180 }
181 
182 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
183 {
184 	return (sysmmu_pte_t *)phys_to_virt(
185 				lv2table_base(sent)) + lv2ent_offset(iova);
186 }
187 
188 struct sysmmu_fault {
189 	sysmmu_iova_t addr;	/* IOVA address that caused fault */
190 	const char *name;	/* human readable fault name */
191 	unsigned int type;	/* fault type for report_iommu_fault() */
192 };
193 
194 struct sysmmu_v1_fault_info {
195 	unsigned short addr_reg; /* register to read IOVA fault address */
196 	const char *name;	/* human readable fault name */
197 	unsigned int type;	/* fault type for report_iommu_fault */
198 };
199 
200 static const struct sysmmu_v1_fault_info sysmmu_v1_faults[] = {
201 	{ REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
202 	{ REG_AR_FAULT_ADDR, "MULTI-HIT", IOMMU_FAULT_READ },
203 	{ REG_AW_FAULT_ADDR, "MULTI-HIT", IOMMU_FAULT_WRITE },
204 	{ REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
205 	{ REG_AR_FAULT_ADDR, "SECURITY PROTECTION", IOMMU_FAULT_READ },
206 	{ REG_AR_FAULT_ADDR, "ACCESS PROTECTION", IOMMU_FAULT_READ },
207 	{ REG_AW_FAULT_ADDR, "SECURITY PROTECTION", IOMMU_FAULT_WRITE },
208 	{ REG_AW_FAULT_ADDR, "ACCESS PROTECTION", IOMMU_FAULT_WRITE },
209 };
210 
211 /* SysMMU v5 has the same faults for AR (0..4 bits) and AW (16..20 bits) */
212 static const char * const sysmmu_v5_fault_names[] = {
213 	"PTW",
214 	"PAGE",
215 	"MULTI-HIT",
216 	"ACCESS PROTECTION",
217 	"SECURITY PROTECTION"
218 };
219 
220 static const char * const sysmmu_v7_fault_names[] = {
221 	"PTW",
222 	"PAGE",
223 	"ACCESS PROTECTION",
224 	"RESERVED"
225 };
226 
227 /*
228  * This structure is attached to dev->iommu->priv of the master device
229  * on device add, contains a list of SYSMMU controllers defined by device tree,
230  * which are bound to given master device. It is usually referenced by 'owner'
231  * pointer.
232 */
233 struct exynos_iommu_owner {
234 	struct list_head controllers;	/* list of sysmmu_drvdata.owner_node */
235 	struct iommu_domain *domain;	/* domain this device is attached */
236 	struct mutex rpm_lock;		/* for runtime pm of all sysmmus */
237 };
238 
239 /*
240  * This structure exynos specific generalization of struct iommu_domain.
241  * It contains list of SYSMMU controllers from all master devices, which has
242  * been attached to this domain and page tables of IO address space defined by
243  * it. It is usually referenced by 'domain' pointer.
244  */
245 struct exynos_iommu_domain {
246 	struct list_head clients; /* list of sysmmu_drvdata.domain_node */
247 	sysmmu_pte_t *pgtable;	/* lv1 page table, 16KB */
248 	short *lv2entcnt;	/* free lv2 entry counter for each section */
249 	spinlock_t lock;	/* lock for modyfying list of clients */
250 	spinlock_t pgtablelock;	/* lock for modifying page table @ pgtable */
251 	struct iommu_domain domain; /* generic domain data structure */
252 };
253 
254 struct sysmmu_drvdata;
255 
256 /*
257  * SysMMU version specific data. Contains offsets for the registers which can
258  * be found in different SysMMU variants, but have different offset values.
259  * Also contains version specific callbacks to abstract the hardware.
260  */
261 struct sysmmu_variant {
262 	u32 pt_base;		/* page table base address (physical) */
263 	u32 flush_all;		/* invalidate all TLB entries */
264 	u32 flush_entry;	/* invalidate specific TLB entry */
265 	u32 flush_range;	/* invalidate TLB entries in specified range */
266 	u32 flush_start;	/* start address of range invalidation */
267 	u32 flush_end;		/* end address of range invalidation */
268 	u32 int_status;		/* interrupt status information */
269 	u32 int_clear;		/* clear the interrupt */
270 	u32 fault_va;		/* IOVA address that caused fault */
271 	u32 fault_info;		/* fault transaction info */
272 
273 	int (*get_fault_info)(struct sysmmu_drvdata *data, unsigned int itype,
274 			      struct sysmmu_fault *fault);
275 };
276 
277 /*
278  * This structure hold all data of a single SYSMMU controller, this includes
279  * hw resources like registers and clocks, pointers and list nodes to connect
280  * it to all other structures, internal state and parameters read from device
281  * tree. It is usually referenced by 'data' pointer.
282  */
283 struct sysmmu_drvdata {
284 	struct device *sysmmu;		/* SYSMMU controller device */
285 	struct device *master;		/* master device (owner) */
286 	struct device_link *link;	/* runtime PM link to master */
287 	void __iomem *sfrbase;		/* our registers */
288 	struct clk *clk;		/* SYSMMU's clock */
289 	struct clk *aclk;		/* SYSMMU's aclk clock */
290 	struct clk *pclk;		/* SYSMMU's pclk clock */
291 	struct clk *clk_master;		/* master's device clock */
292 	spinlock_t lock;		/* lock for modyfying state */
293 	bool active;			/* current status */
294 	struct exynos_iommu_domain *domain; /* domain we belong to */
295 	struct list_head domain_node;	/* node for domain clients list */
296 	struct list_head owner_node;	/* node for owner controllers list */
297 	phys_addr_t pgtable;		/* assigned page table structure */
298 	unsigned int version;		/* our version */
299 
300 	struct iommu_device iommu;	/* IOMMU core handle */
301 	const struct sysmmu_variant *variant; /* version specific data */
302 
303 	/* v7 fields */
304 	bool has_vcr;			/* virtual machine control register */
305 };
306 
307 #define SYSMMU_REG(data, reg) ((data)->sfrbase + (data)->variant->reg)
308 
309 static int exynos_sysmmu_v1_get_fault_info(struct sysmmu_drvdata *data,
310 					   unsigned int itype,
311 					   struct sysmmu_fault *fault)
312 {
313 	const struct sysmmu_v1_fault_info *finfo;
314 
315 	if (itype >= ARRAY_SIZE(sysmmu_v1_faults))
316 		return -ENXIO;
317 
318 	finfo = &sysmmu_v1_faults[itype];
319 	fault->addr = readl(data->sfrbase + finfo->addr_reg);
320 	fault->name = finfo->name;
321 	fault->type = finfo->type;
322 
323 	return 0;
324 }
325 
326 static int exynos_sysmmu_v5_get_fault_info(struct sysmmu_drvdata *data,
327 					   unsigned int itype,
328 					   struct sysmmu_fault *fault)
329 {
330 	unsigned int addr_reg;
331 
332 	if (itype < ARRAY_SIZE(sysmmu_v5_fault_names)) {
333 		fault->type = IOMMU_FAULT_READ;
334 		addr_reg = REG_V5_FAULT_AR_VA;
335 	} else if (itype >= 16 && itype <= 20) {
336 		fault->type = IOMMU_FAULT_WRITE;
337 		addr_reg = REG_V5_FAULT_AW_VA;
338 		itype -= 16;
339 	} else {
340 		return -ENXIO;
341 	}
342 
343 	fault->name = sysmmu_v5_fault_names[itype];
344 	fault->addr = readl(data->sfrbase + addr_reg);
345 
346 	return 0;
347 }
348 
349 static int exynos_sysmmu_v7_get_fault_info(struct sysmmu_drvdata *data,
350 					   unsigned int itype,
351 					   struct sysmmu_fault *fault)
352 {
353 	u32 info = readl(SYSMMU_REG(data, fault_info));
354 
355 	fault->addr = readl(SYSMMU_REG(data, fault_va));
356 	fault->name = sysmmu_v7_fault_names[itype % 4];
357 	fault->type = (info & BIT(20)) ? IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
358 
359 	return 0;
360 }
361 
362 /* SysMMU v1..v3 */
363 static const struct sysmmu_variant sysmmu_v1_variant = {
364 	.flush_all	= 0x0c,
365 	.flush_entry	= 0x10,
366 	.pt_base	= 0x14,
367 	.int_status	= 0x18,
368 	.int_clear	= 0x1c,
369 
370 	.get_fault_info	= exynos_sysmmu_v1_get_fault_info,
371 };
372 
373 /* SysMMU v5 */
374 static const struct sysmmu_variant sysmmu_v5_variant = {
375 	.pt_base	= 0x0c,
376 	.flush_all	= 0x10,
377 	.flush_entry	= 0x14,
378 	.flush_range	= 0x18,
379 	.flush_start	= 0x20,
380 	.flush_end	= 0x24,
381 	.int_status	= 0x60,
382 	.int_clear	= 0x64,
383 
384 	.get_fault_info	= exynos_sysmmu_v5_get_fault_info,
385 };
386 
387 /* SysMMU v7: non-VM capable register layout */
388 static const struct sysmmu_variant sysmmu_v7_variant = {
389 	.pt_base	= 0x0c,
390 	.flush_all	= 0x10,
391 	.flush_entry	= 0x14,
392 	.flush_range	= 0x18,
393 	.flush_start	= 0x20,
394 	.flush_end	= 0x24,
395 	.int_status	= 0x60,
396 	.int_clear	= 0x64,
397 	.fault_va	= 0x70,
398 	.fault_info	= 0x78,
399 
400 	.get_fault_info	= exynos_sysmmu_v7_get_fault_info,
401 };
402 
403 /* SysMMU v7: VM capable register layout */
404 static const struct sysmmu_variant sysmmu_v7_vm_variant = {
405 	.pt_base	= 0x800c,
406 	.flush_all	= 0x8010,
407 	.flush_entry	= 0x8014,
408 	.flush_range	= 0x8018,
409 	.flush_start	= 0x8020,
410 	.flush_end	= 0x8024,
411 	.int_status	= 0x60,
412 	.int_clear	= 0x64,
413 	.fault_va	= 0x1000,
414 	.fault_info	= 0x1004,
415 
416 	.get_fault_info	= exynos_sysmmu_v7_get_fault_info,
417 };
418 
419 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
420 {
421 	return container_of(dom, struct exynos_iommu_domain, domain);
422 }
423 
424 static void sysmmu_unblock(struct sysmmu_drvdata *data)
425 {
426 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
427 }
428 
429 static bool sysmmu_block(struct sysmmu_drvdata *data)
430 {
431 	int i = 120;
432 
433 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
434 	while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
435 		--i;
436 
437 	if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
438 		sysmmu_unblock(data);
439 		return false;
440 	}
441 
442 	return true;
443 }
444 
445 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
446 {
447 	writel(0x1, SYSMMU_REG(data, flush_all));
448 }
449 
450 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
451 				sysmmu_iova_t iova, unsigned int num_inv)
452 {
453 	unsigned int i;
454 
455 	if (MMU_MAJ_VER(data->version) < 5 || num_inv == 1) {
456 		for (i = 0; i < num_inv; i++) {
457 			writel((iova & SPAGE_MASK) | 1,
458 			       SYSMMU_REG(data, flush_entry));
459 			iova += SPAGE_SIZE;
460 		}
461 	} else {
462 		writel(iova & SPAGE_MASK, SYSMMU_REG(data, flush_start));
463 		writel((iova & SPAGE_MASK) + (num_inv - 1) * SPAGE_SIZE,
464 		       SYSMMU_REG(data, flush_end));
465 		writel(0x1, SYSMMU_REG(data, flush_range));
466 	}
467 }
468 
469 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
470 {
471 	u32 pt_base;
472 
473 	if (MMU_MAJ_VER(data->version) < 5)
474 		pt_base = pgd;
475 	else
476 		pt_base = pgd >> SPAGE_ORDER;
477 
478 	writel(pt_base, SYSMMU_REG(data, pt_base));
479 	__sysmmu_tlb_invalidate(data);
480 }
481 
482 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
483 {
484 	BUG_ON(clk_prepare_enable(data->clk_master));
485 	BUG_ON(clk_prepare_enable(data->clk));
486 	BUG_ON(clk_prepare_enable(data->pclk));
487 	BUG_ON(clk_prepare_enable(data->aclk));
488 }
489 
490 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
491 {
492 	clk_disable_unprepare(data->aclk);
493 	clk_disable_unprepare(data->pclk);
494 	clk_disable_unprepare(data->clk);
495 	clk_disable_unprepare(data->clk_master);
496 }
497 
498 static bool __sysmmu_has_capa1(struct sysmmu_drvdata *data)
499 {
500 	u32 capa0 = readl(data->sfrbase + REG_V7_CAPA0);
501 
502 	return capa0 & CAPA0_CAPA1_EXIST;
503 }
504 
505 static void __sysmmu_get_vcr(struct sysmmu_drvdata *data)
506 {
507 	u32 capa1 = readl(data->sfrbase + REG_V7_CAPA1);
508 
509 	data->has_vcr = capa1 & CAPA1_VCR_ENABLED;
510 }
511 
512 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
513 {
514 	u32 ver;
515 
516 	__sysmmu_enable_clocks(data);
517 
518 	ver = readl(data->sfrbase + REG_MMU_VERSION);
519 
520 	/* controllers on some SoCs don't report proper version */
521 	if (ver == 0x80000001u)
522 		data->version = MAKE_MMU_VER(1, 0);
523 	else
524 		data->version = MMU_RAW_VER(ver);
525 
526 	dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
527 		MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
528 
529 	if (MMU_MAJ_VER(data->version) < 5) {
530 		data->variant = &sysmmu_v1_variant;
531 	} else if (MMU_MAJ_VER(data->version) < 7) {
532 		data->variant = &sysmmu_v5_variant;
533 	} else {
534 		if (__sysmmu_has_capa1(data))
535 			__sysmmu_get_vcr(data);
536 		if (data->has_vcr)
537 			data->variant = &sysmmu_v7_vm_variant;
538 		else
539 			data->variant = &sysmmu_v7_variant;
540 	}
541 
542 	__sysmmu_disable_clocks(data);
543 }
544 
545 static void show_fault_information(struct sysmmu_drvdata *data,
546 				   const struct sysmmu_fault *fault)
547 {
548 	sysmmu_pte_t *ent;
549 
550 	dev_err(data->sysmmu, "%s: [%s] %s FAULT occurred at %#x\n",
551 		dev_name(data->master),
552 		fault->type == IOMMU_FAULT_READ ? "READ" : "WRITE",
553 		fault->name, fault->addr);
554 	dev_dbg(data->sysmmu, "Page table base: %pa\n", &data->pgtable);
555 	ent = section_entry(phys_to_virt(data->pgtable), fault->addr);
556 	dev_dbg(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
557 	if (lv1ent_page(ent)) {
558 		ent = page_entry(ent, fault->addr);
559 		dev_dbg(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
560 	}
561 }
562 
563 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
564 {
565 	struct sysmmu_drvdata *data = dev_id;
566 	unsigned int itype;
567 	struct sysmmu_fault fault;
568 	int ret = -ENOSYS;
569 
570 	WARN_ON(!data->active);
571 
572 	spin_lock(&data->lock);
573 	clk_enable(data->clk_master);
574 
575 	itype = __ffs(readl(SYSMMU_REG(data, int_status)));
576 	ret = data->variant->get_fault_info(data, itype, &fault);
577 	if (ret) {
578 		dev_err(data->sysmmu, "Unhandled interrupt bit %u\n", itype);
579 		goto out;
580 	}
581 	show_fault_information(data, &fault);
582 
583 	if (data->domain) {
584 		ret = report_iommu_fault(&data->domain->domain, data->master,
585 					 fault.addr, fault.type);
586 	}
587 	if (ret)
588 		panic("Unrecoverable System MMU Fault!");
589 
590 out:
591 	writel(1 << itype, SYSMMU_REG(data, int_clear));
592 
593 	/* SysMMU is in blocked state when interrupt occurred */
594 	sysmmu_unblock(data);
595 	clk_disable(data->clk_master);
596 	spin_unlock(&data->lock);
597 
598 	return IRQ_HANDLED;
599 }
600 
601 static void __sysmmu_disable(struct sysmmu_drvdata *data)
602 {
603 	unsigned long flags;
604 
605 	clk_enable(data->clk_master);
606 
607 	spin_lock_irqsave(&data->lock, flags);
608 	writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
609 	writel(0, data->sfrbase + REG_MMU_CFG);
610 	data->active = false;
611 	spin_unlock_irqrestore(&data->lock, flags);
612 
613 	__sysmmu_disable_clocks(data);
614 }
615 
616 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
617 {
618 	unsigned int cfg;
619 
620 	if (data->version <= MAKE_MMU_VER(3, 1))
621 		cfg = CFG_LRU | CFG_QOS(15);
622 	else if (data->version <= MAKE_MMU_VER(3, 2))
623 		cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
624 	else
625 		cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
626 
627 	cfg |= CFG_EAP; /* enable access protection bits check */
628 
629 	writel(cfg, data->sfrbase + REG_MMU_CFG);
630 }
631 
632 static void __sysmmu_enable_vid(struct sysmmu_drvdata *data)
633 {
634 	u32 ctrl;
635 
636 	if (MMU_MAJ_VER(data->version) < 7 || !data->has_vcr)
637 		return;
638 
639 	ctrl = readl(data->sfrbase + REG_V7_CTRL_VM);
640 	ctrl |= CTRL_VM_ENABLE | CTRL_VM_FAULT_MODE_STALL;
641 	writel(ctrl, data->sfrbase + REG_V7_CTRL_VM);
642 }
643 
644 static void __sysmmu_enable(struct sysmmu_drvdata *data)
645 {
646 	unsigned long flags;
647 
648 	__sysmmu_enable_clocks(data);
649 
650 	spin_lock_irqsave(&data->lock, flags);
651 	writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
652 	__sysmmu_init_config(data);
653 	__sysmmu_set_ptbase(data, data->pgtable);
654 	__sysmmu_enable_vid(data);
655 	writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
656 	data->active = true;
657 	spin_unlock_irqrestore(&data->lock, flags);
658 
659 	/*
660 	 * SYSMMU driver keeps master's clock enabled only for the short
661 	 * time, while accessing the registers. For performing address
662 	 * translation during DMA transaction it relies on the client
663 	 * driver to enable it.
664 	 */
665 	clk_disable(data->clk_master);
666 }
667 
668 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
669 					    sysmmu_iova_t iova)
670 {
671 	unsigned long flags;
672 
673 	spin_lock_irqsave(&data->lock, flags);
674 	if (data->active && data->version >= MAKE_MMU_VER(3, 3)) {
675 		clk_enable(data->clk_master);
676 		if (sysmmu_block(data)) {
677 			if (data->version >= MAKE_MMU_VER(5, 0))
678 				__sysmmu_tlb_invalidate(data);
679 			else
680 				__sysmmu_tlb_invalidate_entry(data, iova, 1);
681 			sysmmu_unblock(data);
682 		}
683 		clk_disable(data->clk_master);
684 	}
685 	spin_unlock_irqrestore(&data->lock, flags);
686 }
687 
688 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
689 					sysmmu_iova_t iova, size_t size)
690 {
691 	unsigned long flags;
692 
693 	spin_lock_irqsave(&data->lock, flags);
694 	if (data->active) {
695 		unsigned int num_inv = 1;
696 
697 		clk_enable(data->clk_master);
698 
699 		/*
700 		 * L2TLB invalidation required
701 		 * 4KB page: 1 invalidation
702 		 * 64KB page: 16 invalidations
703 		 * 1MB page: 64 invalidations
704 		 * because it is set-associative TLB
705 		 * with 8-way and 64 sets.
706 		 * 1MB page can be cached in one of all sets.
707 		 * 64KB page can be one of 16 consecutive sets.
708 		 */
709 		if (MMU_MAJ_VER(data->version) == 2)
710 			num_inv = min_t(unsigned int, size / SPAGE_SIZE, 64);
711 
712 		if (sysmmu_block(data)) {
713 			__sysmmu_tlb_invalidate_entry(data, iova, num_inv);
714 			sysmmu_unblock(data);
715 		}
716 		clk_disable(data->clk_master);
717 	}
718 	spin_unlock_irqrestore(&data->lock, flags);
719 }
720 
721 static const struct iommu_ops exynos_iommu_ops;
722 
723 static int exynos_sysmmu_probe(struct platform_device *pdev)
724 {
725 	int irq, ret;
726 	struct device *dev = &pdev->dev;
727 	struct sysmmu_drvdata *data;
728 	struct resource *res;
729 
730 	data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
731 	if (!data)
732 		return -ENOMEM;
733 
734 	res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
735 	data->sfrbase = devm_ioremap_resource(dev, res);
736 	if (IS_ERR(data->sfrbase))
737 		return PTR_ERR(data->sfrbase);
738 
739 	irq = platform_get_irq(pdev, 0);
740 	if (irq <= 0)
741 		return irq;
742 
743 	ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
744 				dev_name(dev), data);
745 	if (ret) {
746 		dev_err(dev, "Unabled to register handler of irq %d\n", irq);
747 		return ret;
748 	}
749 
750 	data->clk = devm_clk_get(dev, "sysmmu");
751 	if (PTR_ERR(data->clk) == -ENOENT)
752 		data->clk = NULL;
753 	else if (IS_ERR(data->clk))
754 		return PTR_ERR(data->clk);
755 
756 	data->aclk = devm_clk_get(dev, "aclk");
757 	if (PTR_ERR(data->aclk) == -ENOENT)
758 		data->aclk = NULL;
759 	else if (IS_ERR(data->aclk))
760 		return PTR_ERR(data->aclk);
761 
762 	data->pclk = devm_clk_get(dev, "pclk");
763 	if (PTR_ERR(data->pclk) == -ENOENT)
764 		data->pclk = NULL;
765 	else if (IS_ERR(data->pclk))
766 		return PTR_ERR(data->pclk);
767 
768 	if (!data->clk && (!data->aclk || !data->pclk)) {
769 		dev_err(dev, "Failed to get device clock(s)!\n");
770 		return -ENOSYS;
771 	}
772 
773 	data->clk_master = devm_clk_get(dev, "master");
774 	if (PTR_ERR(data->clk_master) == -ENOENT)
775 		data->clk_master = NULL;
776 	else if (IS_ERR(data->clk_master))
777 		return PTR_ERR(data->clk_master);
778 
779 	data->sysmmu = dev;
780 	spin_lock_init(&data->lock);
781 
782 	__sysmmu_get_version(data);
783 
784 	ret = iommu_device_sysfs_add(&data->iommu, &pdev->dev, NULL,
785 				     dev_name(data->sysmmu));
786 	if (ret)
787 		return ret;
788 
789 	platform_set_drvdata(pdev, data);
790 
791 	if (PG_ENT_SHIFT < 0) {
792 		if (MMU_MAJ_VER(data->version) < 5) {
793 			PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
794 			LV1_PROT = SYSMMU_LV1_PROT;
795 			LV2_PROT = SYSMMU_LV2_PROT;
796 		} else {
797 			PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
798 			LV1_PROT = SYSMMU_V5_LV1_PROT;
799 			LV2_PROT = SYSMMU_V5_LV2_PROT;
800 		}
801 	}
802 
803 	if (MMU_MAJ_VER(data->version) >= 5) {
804 		ret = dma_set_mask(dev, DMA_BIT_MASK(36));
805 		if (ret) {
806 			dev_err(dev, "Unable to set DMA mask: %d\n", ret);
807 			goto err_dma_set_mask;
808 		}
809 	}
810 
811 	/*
812 	 * use the first registered sysmmu device for performing
813 	 * dma mapping operations on iommu page tables (cpu cache flush)
814 	 */
815 	if (!dma_dev)
816 		dma_dev = &pdev->dev;
817 
818 	pm_runtime_enable(dev);
819 
820 	ret = iommu_device_register(&data->iommu, &exynos_iommu_ops, dev);
821 	if (ret)
822 		goto err_dma_set_mask;
823 
824 	return 0;
825 
826 err_dma_set_mask:
827 	iommu_device_sysfs_remove(&data->iommu);
828 	return ret;
829 }
830 
831 static int __maybe_unused exynos_sysmmu_suspend(struct device *dev)
832 {
833 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
834 	struct device *master = data->master;
835 
836 	if (master) {
837 		struct exynos_iommu_owner *owner = dev_iommu_priv_get(master);
838 
839 		mutex_lock(&owner->rpm_lock);
840 		if (data->domain) {
841 			dev_dbg(data->sysmmu, "saving state\n");
842 			__sysmmu_disable(data);
843 		}
844 		mutex_unlock(&owner->rpm_lock);
845 	}
846 	return 0;
847 }
848 
849 static int __maybe_unused exynos_sysmmu_resume(struct device *dev)
850 {
851 	struct sysmmu_drvdata *data = dev_get_drvdata(dev);
852 	struct device *master = data->master;
853 
854 	if (master) {
855 		struct exynos_iommu_owner *owner = dev_iommu_priv_get(master);
856 
857 		mutex_lock(&owner->rpm_lock);
858 		if (data->domain) {
859 			dev_dbg(data->sysmmu, "restoring state\n");
860 			__sysmmu_enable(data);
861 		}
862 		mutex_unlock(&owner->rpm_lock);
863 	}
864 	return 0;
865 }
866 
867 static const struct dev_pm_ops sysmmu_pm_ops = {
868 	SET_RUNTIME_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume, NULL)
869 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
870 				pm_runtime_force_resume)
871 };
872 
873 static const struct of_device_id sysmmu_of_match[] = {
874 	{ .compatible	= "samsung,exynos-sysmmu", },
875 	{ },
876 };
877 
878 static struct platform_driver exynos_sysmmu_driver __refdata = {
879 	.probe	= exynos_sysmmu_probe,
880 	.driver	= {
881 		.name		= "exynos-sysmmu",
882 		.of_match_table	= sysmmu_of_match,
883 		.pm		= &sysmmu_pm_ops,
884 		.suppress_bind_attrs = true,
885 	}
886 };
887 
888 static inline void exynos_iommu_set_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
889 {
890 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
891 				DMA_TO_DEVICE);
892 	*ent = cpu_to_le32(val);
893 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
894 				   DMA_TO_DEVICE);
895 }
896 
897 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
898 {
899 	struct exynos_iommu_domain *domain;
900 	dma_addr_t handle;
901 	int i;
902 
903 	/* Check if correct PTE offsets are initialized */
904 	BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
905 
906 	if (type != IOMMU_DOMAIN_DMA && type != IOMMU_DOMAIN_UNMANAGED)
907 		return NULL;
908 
909 	domain = kzalloc(sizeof(*domain), GFP_KERNEL);
910 	if (!domain)
911 		return NULL;
912 
913 	domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
914 	if (!domain->pgtable)
915 		goto err_pgtable;
916 
917 	domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
918 	if (!domain->lv2entcnt)
919 		goto err_counter;
920 
921 	/* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
922 	for (i = 0; i < NUM_LV1ENTRIES; i++)
923 		domain->pgtable[i] = ZERO_LV2LINK;
924 
925 	handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
926 				DMA_TO_DEVICE);
927 	/* For mapping page table entries we rely on dma == phys */
928 	BUG_ON(handle != virt_to_phys(domain->pgtable));
929 	if (dma_mapping_error(dma_dev, handle))
930 		goto err_lv2ent;
931 
932 	spin_lock_init(&domain->lock);
933 	spin_lock_init(&domain->pgtablelock);
934 	INIT_LIST_HEAD(&domain->clients);
935 
936 	domain->domain.geometry.aperture_start = 0;
937 	domain->domain.geometry.aperture_end   = ~0UL;
938 	domain->domain.geometry.force_aperture = true;
939 
940 	return &domain->domain;
941 
942 err_lv2ent:
943 	free_pages((unsigned long)domain->lv2entcnt, 1);
944 err_counter:
945 	free_pages((unsigned long)domain->pgtable, 2);
946 err_pgtable:
947 	kfree(domain);
948 	return NULL;
949 }
950 
951 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
952 {
953 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
954 	struct sysmmu_drvdata *data, *next;
955 	unsigned long flags;
956 	int i;
957 
958 	WARN_ON(!list_empty(&domain->clients));
959 
960 	spin_lock_irqsave(&domain->lock, flags);
961 
962 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
963 		spin_lock(&data->lock);
964 		__sysmmu_disable(data);
965 		data->pgtable = 0;
966 		data->domain = NULL;
967 		list_del_init(&data->domain_node);
968 		spin_unlock(&data->lock);
969 	}
970 
971 	spin_unlock_irqrestore(&domain->lock, flags);
972 
973 	dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
974 			 DMA_TO_DEVICE);
975 
976 	for (i = 0; i < NUM_LV1ENTRIES; i++)
977 		if (lv1ent_page(domain->pgtable + i)) {
978 			phys_addr_t base = lv2table_base(domain->pgtable + i);
979 
980 			dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
981 					 DMA_TO_DEVICE);
982 			kmem_cache_free(lv2table_kmem_cache,
983 					phys_to_virt(base));
984 		}
985 
986 	free_pages((unsigned long)domain->pgtable, 2);
987 	free_pages((unsigned long)domain->lv2entcnt, 1);
988 	kfree(domain);
989 }
990 
991 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
992 				    struct device *dev)
993 {
994 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
995 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
996 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
997 	struct sysmmu_drvdata *data, *next;
998 	unsigned long flags;
999 
1000 	if (!has_sysmmu(dev) || owner->domain != iommu_domain)
1001 		return;
1002 
1003 	mutex_lock(&owner->rpm_lock);
1004 
1005 	list_for_each_entry(data, &owner->controllers, owner_node) {
1006 		pm_runtime_get_noresume(data->sysmmu);
1007 		if (pm_runtime_active(data->sysmmu))
1008 			__sysmmu_disable(data);
1009 		pm_runtime_put(data->sysmmu);
1010 	}
1011 
1012 	spin_lock_irqsave(&domain->lock, flags);
1013 	list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
1014 		spin_lock(&data->lock);
1015 		data->pgtable = 0;
1016 		data->domain = NULL;
1017 		list_del_init(&data->domain_node);
1018 		spin_unlock(&data->lock);
1019 	}
1020 	owner->domain = NULL;
1021 	spin_unlock_irqrestore(&domain->lock, flags);
1022 
1023 	mutex_unlock(&owner->rpm_lock);
1024 
1025 	dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n", __func__,
1026 		&pagetable);
1027 }
1028 
1029 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
1030 				   struct device *dev)
1031 {
1032 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1033 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1034 	struct sysmmu_drvdata *data;
1035 	phys_addr_t pagetable = virt_to_phys(domain->pgtable);
1036 	unsigned long flags;
1037 
1038 	if (!has_sysmmu(dev))
1039 		return -ENODEV;
1040 
1041 	if (owner->domain)
1042 		exynos_iommu_detach_device(owner->domain, dev);
1043 
1044 	mutex_lock(&owner->rpm_lock);
1045 
1046 	spin_lock_irqsave(&domain->lock, flags);
1047 	list_for_each_entry(data, &owner->controllers, owner_node) {
1048 		spin_lock(&data->lock);
1049 		data->pgtable = pagetable;
1050 		data->domain = domain;
1051 		list_add_tail(&data->domain_node, &domain->clients);
1052 		spin_unlock(&data->lock);
1053 	}
1054 	owner->domain = iommu_domain;
1055 	spin_unlock_irqrestore(&domain->lock, flags);
1056 
1057 	list_for_each_entry(data, &owner->controllers, owner_node) {
1058 		pm_runtime_get_noresume(data->sysmmu);
1059 		if (pm_runtime_active(data->sysmmu))
1060 			__sysmmu_enable(data);
1061 		pm_runtime_put(data->sysmmu);
1062 	}
1063 
1064 	mutex_unlock(&owner->rpm_lock);
1065 
1066 	dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa\n", __func__,
1067 		&pagetable);
1068 
1069 	return 0;
1070 }
1071 
1072 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
1073 		sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
1074 {
1075 	if (lv1ent_section(sent)) {
1076 		WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
1077 		return ERR_PTR(-EADDRINUSE);
1078 	}
1079 
1080 	if (lv1ent_fault(sent)) {
1081 		dma_addr_t handle;
1082 		sysmmu_pte_t *pent;
1083 		bool need_flush_flpd_cache = lv1ent_zero(sent);
1084 
1085 		pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
1086 		BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
1087 		if (!pent)
1088 			return ERR_PTR(-ENOMEM);
1089 
1090 		exynos_iommu_set_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
1091 		kmemleak_ignore(pent);
1092 		*pgcounter = NUM_LV2ENTRIES;
1093 		handle = dma_map_single(dma_dev, pent, LV2TABLE_SIZE,
1094 					DMA_TO_DEVICE);
1095 		if (dma_mapping_error(dma_dev, handle)) {
1096 			kmem_cache_free(lv2table_kmem_cache, pent);
1097 			return ERR_PTR(-EADDRINUSE);
1098 		}
1099 
1100 		/*
1101 		 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
1102 		 * FLPD cache may cache the address of zero_l2_table. This
1103 		 * function replaces the zero_l2_table with new L2 page table
1104 		 * to write valid mappings.
1105 		 * Accessing the valid area may cause page fault since FLPD
1106 		 * cache may still cache zero_l2_table for the valid area
1107 		 * instead of new L2 page table that has the mapping
1108 		 * information of the valid area.
1109 		 * Thus any replacement of zero_l2_table with other valid L2
1110 		 * page table must involve FLPD cache invalidation for System
1111 		 * MMU v3.3.
1112 		 * FLPD cache invalidation is performed with TLB invalidation
1113 		 * by VPN without blocking. It is safe to invalidate TLB without
1114 		 * blocking because the target address of TLB invalidation is
1115 		 * not currently mapped.
1116 		 */
1117 		if (need_flush_flpd_cache) {
1118 			struct sysmmu_drvdata *data;
1119 
1120 			spin_lock(&domain->lock);
1121 			list_for_each_entry(data, &domain->clients, domain_node)
1122 				sysmmu_tlb_invalidate_flpdcache(data, iova);
1123 			spin_unlock(&domain->lock);
1124 		}
1125 	}
1126 
1127 	return page_entry(sent, iova);
1128 }
1129 
1130 static int lv1set_section(struct exynos_iommu_domain *domain,
1131 			  sysmmu_pte_t *sent, sysmmu_iova_t iova,
1132 			  phys_addr_t paddr, int prot, short *pgcnt)
1133 {
1134 	if (lv1ent_section(sent)) {
1135 		WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
1136 			iova);
1137 		return -EADDRINUSE;
1138 	}
1139 
1140 	if (lv1ent_page(sent)) {
1141 		if (*pgcnt != NUM_LV2ENTRIES) {
1142 			WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
1143 				iova);
1144 			return -EADDRINUSE;
1145 		}
1146 
1147 		kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
1148 		*pgcnt = 0;
1149 	}
1150 
1151 	exynos_iommu_set_pte(sent, mk_lv1ent_sect(paddr, prot));
1152 
1153 	spin_lock(&domain->lock);
1154 	if (lv1ent_page_zero(sent)) {
1155 		struct sysmmu_drvdata *data;
1156 		/*
1157 		 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
1158 		 * entry by speculative prefetch of SLPD which has no mapping.
1159 		 */
1160 		list_for_each_entry(data, &domain->clients, domain_node)
1161 			sysmmu_tlb_invalidate_flpdcache(data, iova);
1162 	}
1163 	spin_unlock(&domain->lock);
1164 
1165 	return 0;
1166 }
1167 
1168 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
1169 		       int prot, short *pgcnt)
1170 {
1171 	if (size == SPAGE_SIZE) {
1172 		if (WARN_ON(!lv2ent_fault(pent)))
1173 			return -EADDRINUSE;
1174 
1175 		exynos_iommu_set_pte(pent, mk_lv2ent_spage(paddr, prot));
1176 		*pgcnt -= 1;
1177 	} else { /* size == LPAGE_SIZE */
1178 		int i;
1179 		dma_addr_t pent_base = virt_to_phys(pent);
1180 
1181 		dma_sync_single_for_cpu(dma_dev, pent_base,
1182 					sizeof(*pent) * SPAGES_PER_LPAGE,
1183 					DMA_TO_DEVICE);
1184 		for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1185 			if (WARN_ON(!lv2ent_fault(pent))) {
1186 				if (i > 0)
1187 					memset(pent - i, 0, sizeof(*pent) * i);
1188 				return -EADDRINUSE;
1189 			}
1190 
1191 			*pent = mk_lv2ent_lpage(paddr, prot);
1192 		}
1193 		dma_sync_single_for_device(dma_dev, pent_base,
1194 					   sizeof(*pent) * SPAGES_PER_LPAGE,
1195 					   DMA_TO_DEVICE);
1196 		*pgcnt -= SPAGES_PER_LPAGE;
1197 	}
1198 
1199 	return 0;
1200 }
1201 
1202 /*
1203  * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1204  *
1205  * System MMU v3.x has advanced logic to improve address translation
1206  * performance with caching more page table entries by a page table walk.
1207  * However, the logic has a bug that while caching faulty page table entries,
1208  * System MMU reports page fault if the cached fault entry is hit even though
1209  * the fault entry is updated to a valid entry after the entry is cached.
1210  * To prevent caching faulty page table entries which may be updated to valid
1211  * entries later, the virtual memory manager should care about the workaround
1212  * for the problem. The following describes the workaround.
1213  *
1214  * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1215  * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1216  *
1217  * Precisely, any start address of I/O virtual region must be aligned with
1218  * the following sizes for System MMU v3.1 and v3.2.
1219  * System MMU v3.1: 128KiB
1220  * System MMU v3.2: 256KiB
1221  *
1222  * Because System MMU v3.3 caches page table entries more aggressively, it needs
1223  * more workarounds.
1224  * - Any two consecutive I/O virtual regions must have a hole of size larger
1225  *   than or equal to 128KiB.
1226  * - Start address of an I/O virtual region must be aligned by 128KiB.
1227  */
1228 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1229 			    unsigned long l_iova, phys_addr_t paddr, size_t size,
1230 			    int prot, gfp_t gfp)
1231 {
1232 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1233 	sysmmu_pte_t *entry;
1234 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1235 	unsigned long flags;
1236 	int ret = -ENOMEM;
1237 
1238 	BUG_ON(domain->pgtable == NULL);
1239 	prot &= SYSMMU_SUPPORTED_PROT_BITS;
1240 
1241 	spin_lock_irqsave(&domain->pgtablelock, flags);
1242 
1243 	entry = section_entry(domain->pgtable, iova);
1244 
1245 	if (size == SECT_SIZE) {
1246 		ret = lv1set_section(domain, entry, iova, paddr, prot,
1247 				     &domain->lv2entcnt[lv1ent_offset(iova)]);
1248 	} else {
1249 		sysmmu_pte_t *pent;
1250 
1251 		pent = alloc_lv2entry(domain, entry, iova,
1252 				      &domain->lv2entcnt[lv1ent_offset(iova)]);
1253 
1254 		if (IS_ERR(pent))
1255 			ret = PTR_ERR(pent);
1256 		else
1257 			ret = lv2set_page(pent, paddr, size, prot,
1258 				       &domain->lv2entcnt[lv1ent_offset(iova)]);
1259 	}
1260 
1261 	if (ret)
1262 		pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1263 			__func__, ret, size, iova);
1264 
1265 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1266 
1267 	return ret;
1268 }
1269 
1270 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1271 					      sysmmu_iova_t iova, size_t size)
1272 {
1273 	struct sysmmu_drvdata *data;
1274 	unsigned long flags;
1275 
1276 	spin_lock_irqsave(&domain->lock, flags);
1277 
1278 	list_for_each_entry(data, &domain->clients, domain_node)
1279 		sysmmu_tlb_invalidate_entry(data, iova, size);
1280 
1281 	spin_unlock_irqrestore(&domain->lock, flags);
1282 }
1283 
1284 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1285 				 unsigned long l_iova, size_t size,
1286 				 struct iommu_iotlb_gather *gather)
1287 {
1288 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1289 	sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1290 	sysmmu_pte_t *ent;
1291 	size_t err_pgsize;
1292 	unsigned long flags;
1293 
1294 	BUG_ON(domain->pgtable == NULL);
1295 
1296 	spin_lock_irqsave(&domain->pgtablelock, flags);
1297 
1298 	ent = section_entry(domain->pgtable, iova);
1299 
1300 	if (lv1ent_section(ent)) {
1301 		if (WARN_ON(size < SECT_SIZE)) {
1302 			err_pgsize = SECT_SIZE;
1303 			goto err;
1304 		}
1305 
1306 		/* workaround for h/w bug in System MMU v3.3 */
1307 		exynos_iommu_set_pte(ent, ZERO_LV2LINK);
1308 		size = SECT_SIZE;
1309 		goto done;
1310 	}
1311 
1312 	if (unlikely(lv1ent_fault(ent))) {
1313 		if (size > SECT_SIZE)
1314 			size = SECT_SIZE;
1315 		goto done;
1316 	}
1317 
1318 	/* lv1ent_page(sent) == true here */
1319 
1320 	ent = page_entry(ent, iova);
1321 
1322 	if (unlikely(lv2ent_fault(ent))) {
1323 		size = SPAGE_SIZE;
1324 		goto done;
1325 	}
1326 
1327 	if (lv2ent_small(ent)) {
1328 		exynos_iommu_set_pte(ent, 0);
1329 		size = SPAGE_SIZE;
1330 		domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1331 		goto done;
1332 	}
1333 
1334 	/* lv1ent_large(ent) == true here */
1335 	if (WARN_ON(size < LPAGE_SIZE)) {
1336 		err_pgsize = LPAGE_SIZE;
1337 		goto err;
1338 	}
1339 
1340 	dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1341 				sizeof(*ent) * SPAGES_PER_LPAGE,
1342 				DMA_TO_DEVICE);
1343 	memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1344 	dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1345 				   sizeof(*ent) * SPAGES_PER_LPAGE,
1346 				   DMA_TO_DEVICE);
1347 	size = LPAGE_SIZE;
1348 	domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1349 done:
1350 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1351 
1352 	exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1353 
1354 	return size;
1355 err:
1356 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1357 
1358 	pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1359 		__func__, size, iova, err_pgsize);
1360 
1361 	return 0;
1362 }
1363 
1364 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1365 					  dma_addr_t iova)
1366 {
1367 	struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1368 	sysmmu_pte_t *entry;
1369 	unsigned long flags;
1370 	phys_addr_t phys = 0;
1371 
1372 	spin_lock_irqsave(&domain->pgtablelock, flags);
1373 
1374 	entry = section_entry(domain->pgtable, iova);
1375 
1376 	if (lv1ent_section(entry)) {
1377 		phys = section_phys(entry) + section_offs(iova);
1378 	} else if (lv1ent_page(entry)) {
1379 		entry = page_entry(entry, iova);
1380 
1381 		if (lv2ent_large(entry))
1382 			phys = lpage_phys(entry) + lpage_offs(iova);
1383 		else if (lv2ent_small(entry))
1384 			phys = spage_phys(entry) + spage_offs(iova);
1385 	}
1386 
1387 	spin_unlock_irqrestore(&domain->pgtablelock, flags);
1388 
1389 	return phys;
1390 }
1391 
1392 static struct iommu_device *exynos_iommu_probe_device(struct device *dev)
1393 {
1394 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1395 	struct sysmmu_drvdata *data;
1396 
1397 	if (!has_sysmmu(dev))
1398 		return ERR_PTR(-ENODEV);
1399 
1400 	list_for_each_entry(data, &owner->controllers, owner_node) {
1401 		/*
1402 		 * SYSMMU will be runtime activated via device link
1403 		 * (dependency) to its master device, so there are no
1404 		 * direct calls to pm_runtime_get/put in this driver.
1405 		 */
1406 		data->link = device_link_add(dev, data->sysmmu,
1407 					     DL_FLAG_STATELESS |
1408 					     DL_FLAG_PM_RUNTIME);
1409 	}
1410 
1411 	/* There is always at least one entry, see exynos_iommu_of_xlate() */
1412 	data = list_first_entry(&owner->controllers,
1413 				struct sysmmu_drvdata, owner_node);
1414 
1415 	return &data->iommu;
1416 }
1417 
1418 static void exynos_iommu_release_device(struct device *dev)
1419 {
1420 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1421 	struct sysmmu_drvdata *data;
1422 
1423 	if (owner->domain) {
1424 		struct iommu_group *group = iommu_group_get(dev);
1425 
1426 		if (group) {
1427 #ifndef CONFIG_ARM
1428 			WARN_ON(owner->domain !=
1429 				iommu_group_default_domain(group));
1430 #endif
1431 			exynos_iommu_detach_device(owner->domain, dev);
1432 			iommu_group_put(group);
1433 		}
1434 	}
1435 
1436 	list_for_each_entry(data, &owner->controllers, owner_node)
1437 		device_link_del(data->link);
1438 }
1439 
1440 static int exynos_iommu_of_xlate(struct device *dev,
1441 				 struct of_phandle_args *spec)
1442 {
1443 	struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1444 	struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1445 	struct sysmmu_drvdata *data, *entry;
1446 
1447 	if (!sysmmu)
1448 		return -ENODEV;
1449 
1450 	data = platform_get_drvdata(sysmmu);
1451 	if (!data) {
1452 		put_device(&sysmmu->dev);
1453 		return -ENODEV;
1454 	}
1455 
1456 	if (!owner) {
1457 		owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1458 		if (!owner) {
1459 			put_device(&sysmmu->dev);
1460 			return -ENOMEM;
1461 		}
1462 
1463 		INIT_LIST_HEAD(&owner->controllers);
1464 		mutex_init(&owner->rpm_lock);
1465 		dev_iommu_priv_set(dev, owner);
1466 	}
1467 
1468 	list_for_each_entry(entry, &owner->controllers, owner_node)
1469 		if (entry == data)
1470 			return 0;
1471 
1472 	list_add_tail(&data->owner_node, &owner->controllers);
1473 	data->master = dev;
1474 
1475 	return 0;
1476 }
1477 
1478 static const struct iommu_ops exynos_iommu_ops = {
1479 	.domain_alloc = exynos_iommu_domain_alloc,
1480 	.device_group = generic_device_group,
1481 #ifdef CONFIG_ARM
1482 	.set_platform_dma_ops = exynos_iommu_release_device,
1483 #endif
1484 	.probe_device = exynos_iommu_probe_device,
1485 	.release_device = exynos_iommu_release_device,
1486 	.pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1487 	.of_xlate = exynos_iommu_of_xlate,
1488 	.default_domain_ops = &(const struct iommu_domain_ops) {
1489 		.attach_dev	= exynos_iommu_attach_device,
1490 		.map		= exynos_iommu_map,
1491 		.unmap		= exynos_iommu_unmap,
1492 		.iova_to_phys	= exynos_iommu_iova_to_phys,
1493 		.free		= exynos_iommu_domain_free,
1494 	}
1495 };
1496 
1497 static int __init exynos_iommu_init(void)
1498 {
1499 	struct device_node *np;
1500 	int ret;
1501 
1502 	np = of_find_matching_node(NULL, sysmmu_of_match);
1503 	if (!np)
1504 		return 0;
1505 
1506 	of_node_put(np);
1507 
1508 	lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1509 				LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1510 	if (!lv2table_kmem_cache) {
1511 		pr_err("%s: Failed to create kmem cache\n", __func__);
1512 		return -ENOMEM;
1513 	}
1514 
1515 	zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1516 	if (zero_lv2_table == NULL) {
1517 		pr_err("%s: Failed to allocate zero level2 page table\n",
1518 			__func__);
1519 		ret = -ENOMEM;
1520 		goto err_zero_lv2;
1521 	}
1522 
1523 	ret = platform_driver_register(&exynos_sysmmu_driver);
1524 	if (ret) {
1525 		pr_err("%s: Failed to register driver\n", __func__);
1526 		goto err_reg_driver;
1527 	}
1528 
1529 	return 0;
1530 err_reg_driver:
1531 	kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1532 err_zero_lv2:
1533 	kmem_cache_destroy(lv2table_kmem_cache);
1534 	return ret;
1535 }
1536 core_initcall(exynos_iommu_init);
1537