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