45fd4850bacbdbc53d58faa0f87da6ed50563236
[platform/kernel/linux-starfive.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_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 /*
189  * IOMMU fault information register
190  */
191 struct sysmmu_fault_info {
192         unsigned int bit;       /* bit number in STATUS register */
193         unsigned short addr_reg; /* register to read VA fault address */
194         const char *name;       /* human readable fault name */
195         unsigned int type;      /* fault type for report_iommu_fault */
196 };
197
198 static const struct sysmmu_fault_info sysmmu_faults[] = {
199         { 0, REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
200         { 1, REG_AR_FAULT_ADDR, "AR MULTI-HIT", IOMMU_FAULT_READ },
201         { 2, REG_AW_FAULT_ADDR, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
202         { 3, REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
203         { 4, REG_AR_FAULT_ADDR, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
204         { 5, REG_AR_FAULT_ADDR, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
205         { 6, REG_AW_FAULT_ADDR, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
206         { 7, REG_AW_FAULT_ADDR, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
207 };
208
209 static const struct sysmmu_fault_info sysmmu_v5_faults[] = {
210         { 0, REG_V5_FAULT_AR_VA, "AR PTW", IOMMU_FAULT_READ },
211         { 1, REG_V5_FAULT_AR_VA, "AR PAGE", IOMMU_FAULT_READ },
212         { 2, REG_V5_FAULT_AR_VA, "AR MULTI-HIT", IOMMU_FAULT_READ },
213         { 3, REG_V5_FAULT_AR_VA, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
214         { 4, REG_V5_FAULT_AR_VA, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
215         { 16, REG_V5_FAULT_AW_VA, "AW PTW", IOMMU_FAULT_WRITE },
216         { 17, REG_V5_FAULT_AW_VA, "AW PAGE", IOMMU_FAULT_WRITE },
217         { 18, REG_V5_FAULT_AW_VA, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
218         { 19, REG_V5_FAULT_AW_VA, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
219         { 20, REG_V5_FAULT_AW_VA, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
220 };
221
222 /*
223  * This structure is attached to dev->iommu->priv of the master device
224  * on device add, contains a list of SYSMMU controllers defined by device tree,
225  * which are bound to given master device. It is usually referenced by 'owner'
226  * pointer.
227 */
228 struct exynos_iommu_owner {
229         struct list_head controllers;   /* list of sysmmu_drvdata.owner_node */
230         struct iommu_domain *domain;    /* domain this device is attached */
231         struct mutex rpm_lock;          /* for runtime pm of all sysmmus */
232 };
233
234 /*
235  * This structure exynos specific generalization of struct iommu_domain.
236  * It contains list of SYSMMU controllers from all master devices, which has
237  * been attached to this domain and page tables of IO address space defined by
238  * it. It is usually referenced by 'domain' pointer.
239  */
240 struct exynos_iommu_domain {
241         struct list_head clients; /* list of sysmmu_drvdata.domain_node */
242         sysmmu_pte_t *pgtable;  /* lv1 page table, 16KB */
243         short *lv2entcnt;       /* free lv2 entry counter for each section */
244         spinlock_t lock;        /* lock for modyfying list of clients */
245         spinlock_t pgtablelock; /* lock for modifying page table @ pgtable */
246         struct iommu_domain domain; /* generic domain data structure */
247 };
248
249 /*
250  * SysMMU version specific data. Contains offsets for the registers which can
251  * be found in different SysMMU variants, but have different offset values.
252  */
253 struct sysmmu_variant {
254         u32 pt_base;            /* page table base address (physical) */
255         u32 flush_all;          /* invalidate all TLB entries */
256         u32 flush_entry;        /* invalidate specific TLB entry */
257         u32 flush_range;        /* invalidate TLB entries in specified range */
258         u32 flush_start;        /* start address of range invalidation */
259         u32 flush_end;          /* end address of range invalidation */
260         u32 int_status;         /* interrupt status information */
261         u32 int_clear;          /* clear the interrupt */
262 };
263
264 /*
265  * This structure hold all data of a single SYSMMU controller, this includes
266  * hw resources like registers and clocks, pointers and list nodes to connect
267  * it to all other structures, internal state and parameters read from device
268  * tree. It is usually referenced by 'data' pointer.
269  */
270 struct sysmmu_drvdata {
271         struct device *sysmmu;          /* SYSMMU controller device */
272         struct device *master;          /* master device (owner) */
273         struct device_link *link;       /* runtime PM link to master */
274         void __iomem *sfrbase;          /* our registers */
275         struct clk *clk;                /* SYSMMU's clock */
276         struct clk *aclk;               /* SYSMMU's aclk clock */
277         struct clk *pclk;               /* SYSMMU's pclk clock */
278         struct clk *clk_master;         /* master's device clock */
279         spinlock_t lock;                /* lock for modyfying state */
280         bool active;                    /* current status */
281         struct exynos_iommu_domain *domain; /* domain we belong to */
282         struct list_head domain_node;   /* node for domain clients list */
283         struct list_head owner_node;    /* node for owner controllers list */
284         phys_addr_t pgtable;            /* assigned page table structure */
285         unsigned int version;           /* our version */
286
287         struct iommu_device iommu;      /* IOMMU core handle */
288         const struct sysmmu_variant *variant; /* version specific data */
289
290         /* v7 fields */
291         bool has_vcr;                   /* virtual machine control register */
292 };
293
294 #define SYSMMU_REG(data, reg) ((data)->sfrbase + (data)->variant->reg)
295
296 /* SysMMU v1..v3 */
297 static const struct sysmmu_variant sysmmu_v1_variant = {
298         .flush_all      = 0x0c,
299         .flush_entry    = 0x10,
300         .pt_base        = 0x14,
301         .int_status     = 0x18,
302         .int_clear      = 0x1c,
303 };
304
305 /* SysMMU v5 and v7 (non-VM capable) */
306 static const struct sysmmu_variant sysmmu_v5_variant = {
307         .pt_base        = 0x0c,
308         .flush_all      = 0x10,
309         .flush_entry    = 0x14,
310         .flush_range    = 0x18,
311         .flush_start    = 0x20,
312         .flush_end      = 0x24,
313         .int_status     = 0x60,
314         .int_clear      = 0x64,
315 };
316
317 /* SysMMU v7: VM capable register set */
318 static const struct sysmmu_variant sysmmu_v7_vm_variant = {
319         .pt_base        = 0x800c,
320         .flush_all      = 0x8010,
321         .flush_entry    = 0x8014,
322         .flush_range    = 0x8018,
323         .flush_start    = 0x8020,
324         .flush_end      = 0x8024,
325         .int_status     = 0x60,
326         .int_clear      = 0x64,
327 };
328
329 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
330 {
331         return container_of(dom, struct exynos_iommu_domain, domain);
332 }
333
334 static void sysmmu_unblock(struct sysmmu_drvdata *data)
335 {
336         writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
337 }
338
339 static bool sysmmu_block(struct sysmmu_drvdata *data)
340 {
341         int i = 120;
342
343         writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
344         while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
345                 --i;
346
347         if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
348                 sysmmu_unblock(data);
349                 return false;
350         }
351
352         return true;
353 }
354
355 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
356 {
357         writel(0x1, SYSMMU_REG(data, flush_all));
358 }
359
360 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
361                                 sysmmu_iova_t iova, unsigned int num_inv)
362 {
363         unsigned int i;
364
365         if (MMU_MAJ_VER(data->version) < 5 || num_inv == 1) {
366                 for (i = 0; i < num_inv; i++) {
367                         writel((iova & SPAGE_MASK) | 1,
368                                SYSMMU_REG(data, flush_entry));
369                         iova += SPAGE_SIZE;
370                 }
371         } else {
372                 writel(iova & SPAGE_MASK, SYSMMU_REG(data, flush_start));
373                 writel((iova & SPAGE_MASK) + (num_inv - 1) * SPAGE_SIZE,
374                        SYSMMU_REG(data, flush_end));
375                 writel(0x1, SYSMMU_REG(data, flush_range));
376         }
377 }
378
379 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
380 {
381         u32 pt_base;
382
383         if (MMU_MAJ_VER(data->version) < 5)
384                 pt_base = pgd;
385         else
386                 pt_base = pgd >> SPAGE_ORDER;
387
388         writel(pt_base, SYSMMU_REG(data, pt_base));
389         __sysmmu_tlb_invalidate(data);
390 }
391
392 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
393 {
394         BUG_ON(clk_prepare_enable(data->clk_master));
395         BUG_ON(clk_prepare_enable(data->clk));
396         BUG_ON(clk_prepare_enable(data->pclk));
397         BUG_ON(clk_prepare_enable(data->aclk));
398 }
399
400 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
401 {
402         clk_disable_unprepare(data->aclk);
403         clk_disable_unprepare(data->pclk);
404         clk_disable_unprepare(data->clk);
405         clk_disable_unprepare(data->clk_master);
406 }
407
408 static bool __sysmmu_has_capa1(struct sysmmu_drvdata *data)
409 {
410         u32 capa0 = readl(data->sfrbase + REG_V7_CAPA0);
411
412         return capa0 & CAPA0_CAPA1_EXIST;
413 }
414
415 static void __sysmmu_get_vcr(struct sysmmu_drvdata *data)
416 {
417         u32 capa1 = readl(data->sfrbase + REG_V7_CAPA1);
418
419         data->has_vcr = capa1 & CAPA1_VCR_ENABLED;
420 }
421
422 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
423 {
424         u32 ver;
425
426         __sysmmu_enable_clocks(data);
427
428         ver = readl(data->sfrbase + REG_MMU_VERSION);
429
430         /* controllers on some SoCs don't report proper version */
431         if (ver == 0x80000001u)
432                 data->version = MAKE_MMU_VER(1, 0);
433         else
434                 data->version = MMU_RAW_VER(ver);
435
436         dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
437                 MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
438
439         if (MMU_MAJ_VER(data->version) < 5) {
440                 data->variant = &sysmmu_v1_variant;
441         } else if (MMU_MAJ_VER(data->version) < 7) {
442                 data->variant = &sysmmu_v5_variant;
443         } else {
444                 if (__sysmmu_has_capa1(data))
445                         __sysmmu_get_vcr(data);
446                 if (data->has_vcr)
447                         data->variant = &sysmmu_v7_vm_variant;
448                 else
449                         data->variant = &sysmmu_v5_variant;
450         }
451
452         __sysmmu_disable_clocks(data);
453 }
454
455 static void show_fault_information(struct sysmmu_drvdata *data,
456                                    const struct sysmmu_fault_info *finfo,
457                                    sysmmu_iova_t fault_addr)
458 {
459         sysmmu_pte_t *ent;
460
461         dev_err(data->sysmmu, "%s: %s FAULT occurred at %#x\n",
462                 dev_name(data->master), finfo->name, fault_addr);
463         dev_dbg(data->sysmmu, "Page table base: %pa\n", &data->pgtable);
464         ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
465         dev_dbg(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
466         if (lv1ent_page(ent)) {
467                 ent = page_entry(ent, fault_addr);
468                 dev_dbg(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
469         }
470 }
471
472 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
473 {
474         /* SYSMMU is in blocked state when interrupt occurred. */
475         struct sysmmu_drvdata *data = dev_id;
476         const struct sysmmu_fault_info *finfo;
477         unsigned int i, n, itype;
478         sysmmu_iova_t fault_addr;
479         int ret = -ENOSYS;
480
481         WARN_ON(!data->active);
482
483         if (MMU_MAJ_VER(data->version) < 5) {
484                 finfo = sysmmu_faults;
485                 n = ARRAY_SIZE(sysmmu_faults);
486         } else {
487                 finfo = sysmmu_v5_faults;
488                 n = ARRAY_SIZE(sysmmu_v5_faults);
489         }
490
491         spin_lock(&data->lock);
492
493         clk_enable(data->clk_master);
494
495         itype = __ffs(readl(SYSMMU_REG(data, int_status)));
496         for (i = 0; i < n; i++, finfo++)
497                 if (finfo->bit == itype)
498                         break;
499         /* unknown/unsupported fault */
500         BUG_ON(i == n);
501
502         /* print debug message */
503         fault_addr = readl(data->sfrbase + finfo->addr_reg);
504         show_fault_information(data, finfo, fault_addr);
505
506         if (data->domain)
507                 ret = report_iommu_fault(&data->domain->domain,
508                                         data->master, fault_addr, finfo->type);
509         /* fault is not recovered by fault handler */
510         BUG_ON(ret != 0);
511
512         writel(1 << itype, SYSMMU_REG(data, int_clear));
513
514         sysmmu_unblock(data);
515
516         clk_disable(data->clk_master);
517
518         spin_unlock(&data->lock);
519
520         return IRQ_HANDLED;
521 }
522
523 static void __sysmmu_disable(struct sysmmu_drvdata *data)
524 {
525         unsigned long flags;
526
527         clk_enable(data->clk_master);
528
529         spin_lock_irqsave(&data->lock, flags);
530         writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
531         writel(0, data->sfrbase + REG_MMU_CFG);
532         data->active = false;
533         spin_unlock_irqrestore(&data->lock, flags);
534
535         __sysmmu_disable_clocks(data);
536 }
537
538 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
539 {
540         unsigned int cfg;
541
542         if (data->version <= MAKE_MMU_VER(3, 1))
543                 cfg = CFG_LRU | CFG_QOS(15);
544         else if (data->version <= MAKE_MMU_VER(3, 2))
545                 cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
546         else
547                 cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
548
549         cfg |= CFG_EAP; /* enable access protection bits check */
550
551         writel(cfg, data->sfrbase + REG_MMU_CFG);
552 }
553
554 static void __sysmmu_enable_vid(struct sysmmu_drvdata *data)
555 {
556         u32 ctrl;
557
558         if (MMU_MAJ_VER(data->version) < 7 || !data->has_vcr)
559                 return;
560
561         ctrl = readl(data->sfrbase + REG_V7_CTRL_VM);
562         ctrl |= CTRL_VM_ENABLE | CTRL_VM_FAULT_MODE_STALL;
563         writel(ctrl, data->sfrbase + REG_V7_CTRL_VM);
564 }
565
566 static void __sysmmu_enable(struct sysmmu_drvdata *data)
567 {
568         unsigned long flags;
569
570         __sysmmu_enable_clocks(data);
571
572         spin_lock_irqsave(&data->lock, flags);
573         writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
574         __sysmmu_init_config(data);
575         __sysmmu_set_ptbase(data, data->pgtable);
576         __sysmmu_enable_vid(data);
577         writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
578         data->active = true;
579         spin_unlock_irqrestore(&data->lock, flags);
580
581         /*
582          * SYSMMU driver keeps master's clock enabled only for the short
583          * time, while accessing the registers. For performing address
584          * translation during DMA transaction it relies on the client
585          * driver to enable it.
586          */
587         clk_disable(data->clk_master);
588 }
589
590 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
591                                             sysmmu_iova_t iova)
592 {
593         unsigned long flags;
594
595         spin_lock_irqsave(&data->lock, flags);
596         if (data->active && data->version >= MAKE_MMU_VER(3, 3)) {
597                 clk_enable(data->clk_master);
598                 if (sysmmu_block(data)) {
599                         if (data->version >= MAKE_MMU_VER(5, 0))
600                                 __sysmmu_tlb_invalidate(data);
601                         else
602                                 __sysmmu_tlb_invalidate_entry(data, iova, 1);
603                         sysmmu_unblock(data);
604                 }
605                 clk_disable(data->clk_master);
606         }
607         spin_unlock_irqrestore(&data->lock, flags);
608 }
609
610 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
611                                         sysmmu_iova_t iova, size_t size)
612 {
613         unsigned long flags;
614
615         spin_lock_irqsave(&data->lock, flags);
616         if (data->active) {
617                 unsigned int num_inv = 1;
618
619                 clk_enable(data->clk_master);
620
621                 /*
622                  * L2TLB invalidation required
623                  * 4KB page: 1 invalidation
624                  * 64KB page: 16 invalidations
625                  * 1MB page: 64 invalidations
626                  * because it is set-associative TLB
627                  * with 8-way and 64 sets.
628                  * 1MB page can be cached in one of all sets.
629                  * 64KB page can be one of 16 consecutive sets.
630                  */
631                 if (MMU_MAJ_VER(data->version) == 2)
632                         num_inv = min_t(unsigned int, size / SPAGE_SIZE, 64);
633
634                 if (sysmmu_block(data)) {
635                         __sysmmu_tlb_invalidate_entry(data, iova, num_inv);
636                         sysmmu_unblock(data);
637                 }
638                 clk_disable(data->clk_master);
639         }
640         spin_unlock_irqrestore(&data->lock, flags);
641 }
642
643 static const struct iommu_ops exynos_iommu_ops;
644
645 static int exynos_sysmmu_probe(struct platform_device *pdev)
646 {
647         int irq, ret;
648         struct device *dev = &pdev->dev;
649         struct sysmmu_drvdata *data;
650         struct resource *res;
651
652         data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
653         if (!data)
654                 return -ENOMEM;
655
656         res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
657         data->sfrbase = devm_ioremap_resource(dev, res);
658         if (IS_ERR(data->sfrbase))
659                 return PTR_ERR(data->sfrbase);
660
661         irq = platform_get_irq(pdev, 0);
662         if (irq <= 0)
663                 return irq;
664
665         ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
666                                 dev_name(dev), data);
667         if (ret) {
668                 dev_err(dev, "Unabled to register handler of irq %d\n", irq);
669                 return ret;
670         }
671
672         data->clk = devm_clk_get(dev, "sysmmu");
673         if (PTR_ERR(data->clk) == -ENOENT)
674                 data->clk = NULL;
675         else if (IS_ERR(data->clk))
676                 return PTR_ERR(data->clk);
677
678         data->aclk = devm_clk_get(dev, "aclk");
679         if (PTR_ERR(data->aclk) == -ENOENT)
680                 data->aclk = NULL;
681         else if (IS_ERR(data->aclk))
682                 return PTR_ERR(data->aclk);
683
684         data->pclk = devm_clk_get(dev, "pclk");
685         if (PTR_ERR(data->pclk) == -ENOENT)
686                 data->pclk = NULL;
687         else if (IS_ERR(data->pclk))
688                 return PTR_ERR(data->pclk);
689
690         if (!data->clk && (!data->aclk || !data->pclk)) {
691                 dev_err(dev, "Failed to get device clock(s)!\n");
692                 return -ENOSYS;
693         }
694
695         data->clk_master = devm_clk_get(dev, "master");
696         if (PTR_ERR(data->clk_master) == -ENOENT)
697                 data->clk_master = NULL;
698         else if (IS_ERR(data->clk_master))
699                 return PTR_ERR(data->clk_master);
700
701         data->sysmmu = dev;
702         spin_lock_init(&data->lock);
703
704         __sysmmu_get_version(data);
705
706         ret = iommu_device_sysfs_add(&data->iommu, &pdev->dev, NULL,
707                                      dev_name(data->sysmmu));
708         if (ret)
709                 return ret;
710
711         ret = iommu_device_register(&data->iommu, &exynos_iommu_ops, dev);
712         if (ret)
713                 goto err_iommu_register;
714
715         platform_set_drvdata(pdev, data);
716
717         if (PG_ENT_SHIFT < 0) {
718                 if (MMU_MAJ_VER(data->version) < 5) {
719                         PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
720                         LV1_PROT = SYSMMU_LV1_PROT;
721                         LV2_PROT = SYSMMU_LV2_PROT;
722                 } else {
723                         PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
724                         LV1_PROT = SYSMMU_V5_LV1_PROT;
725                         LV2_PROT = SYSMMU_V5_LV2_PROT;
726                 }
727         }
728
729         if (MMU_MAJ_VER(data->version) >= 5) {
730                 ret = dma_set_mask(dev, DMA_BIT_MASK(36));
731                 if (ret) {
732                         dev_err(dev, "Unable to set DMA mask: %d\n", ret);
733                         goto err_dma_set_mask;
734                 }
735         }
736
737         /*
738          * use the first registered sysmmu device for performing
739          * dma mapping operations on iommu page tables (cpu cache flush)
740          */
741         if (!dma_dev)
742                 dma_dev = &pdev->dev;
743
744         pm_runtime_enable(dev);
745
746         return 0;
747
748 err_dma_set_mask:
749         iommu_device_unregister(&data->iommu);
750 err_iommu_register:
751         iommu_device_sysfs_remove(&data->iommu);
752         return ret;
753 }
754
755 static int __maybe_unused exynos_sysmmu_suspend(struct device *dev)
756 {
757         struct sysmmu_drvdata *data = dev_get_drvdata(dev);
758         struct device *master = data->master;
759
760         if (master) {
761                 struct exynos_iommu_owner *owner = dev_iommu_priv_get(master);
762
763                 mutex_lock(&owner->rpm_lock);
764                 if (data->domain) {
765                         dev_dbg(data->sysmmu, "saving state\n");
766                         __sysmmu_disable(data);
767                 }
768                 mutex_unlock(&owner->rpm_lock);
769         }
770         return 0;
771 }
772
773 static int __maybe_unused exynos_sysmmu_resume(struct device *dev)
774 {
775         struct sysmmu_drvdata *data = dev_get_drvdata(dev);
776         struct device *master = data->master;
777
778         if (master) {
779                 struct exynos_iommu_owner *owner = dev_iommu_priv_get(master);
780
781                 mutex_lock(&owner->rpm_lock);
782                 if (data->domain) {
783                         dev_dbg(data->sysmmu, "restoring state\n");
784                         __sysmmu_enable(data);
785                 }
786                 mutex_unlock(&owner->rpm_lock);
787         }
788         return 0;
789 }
790
791 static const struct dev_pm_ops sysmmu_pm_ops = {
792         SET_RUNTIME_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume, NULL)
793         SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
794                                 pm_runtime_force_resume)
795 };
796
797 static const struct of_device_id sysmmu_of_match[] = {
798         { .compatible   = "samsung,exynos-sysmmu", },
799         { },
800 };
801
802 static struct platform_driver exynos_sysmmu_driver __refdata = {
803         .probe  = exynos_sysmmu_probe,
804         .driver = {
805                 .name           = "exynos-sysmmu",
806                 .of_match_table = sysmmu_of_match,
807                 .pm             = &sysmmu_pm_ops,
808                 .suppress_bind_attrs = true,
809         }
810 };
811
812 static inline void exynos_iommu_set_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
813 {
814         dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
815                                 DMA_TO_DEVICE);
816         *ent = cpu_to_le32(val);
817         dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
818                                    DMA_TO_DEVICE);
819 }
820
821 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
822 {
823         struct exynos_iommu_domain *domain;
824         dma_addr_t handle;
825         int i;
826
827         /* Check if correct PTE offsets are initialized */
828         BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
829
830         if (type != IOMMU_DOMAIN_DMA && type != IOMMU_DOMAIN_UNMANAGED)
831                 return NULL;
832
833         domain = kzalloc(sizeof(*domain), GFP_KERNEL);
834         if (!domain)
835                 return NULL;
836
837         domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
838         if (!domain->pgtable)
839                 goto err_pgtable;
840
841         domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
842         if (!domain->lv2entcnt)
843                 goto err_counter;
844
845         /* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
846         for (i = 0; i < NUM_LV1ENTRIES; i++)
847                 domain->pgtable[i] = ZERO_LV2LINK;
848
849         handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
850                                 DMA_TO_DEVICE);
851         /* For mapping page table entries we rely on dma == phys */
852         BUG_ON(handle != virt_to_phys(domain->pgtable));
853         if (dma_mapping_error(dma_dev, handle))
854                 goto err_lv2ent;
855
856         spin_lock_init(&domain->lock);
857         spin_lock_init(&domain->pgtablelock);
858         INIT_LIST_HEAD(&domain->clients);
859
860         domain->domain.geometry.aperture_start = 0;
861         domain->domain.geometry.aperture_end   = ~0UL;
862         domain->domain.geometry.force_aperture = true;
863
864         return &domain->domain;
865
866 err_lv2ent:
867         free_pages((unsigned long)domain->lv2entcnt, 1);
868 err_counter:
869         free_pages((unsigned long)domain->pgtable, 2);
870 err_pgtable:
871         kfree(domain);
872         return NULL;
873 }
874
875 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
876 {
877         struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
878         struct sysmmu_drvdata *data, *next;
879         unsigned long flags;
880         int i;
881
882         WARN_ON(!list_empty(&domain->clients));
883
884         spin_lock_irqsave(&domain->lock, flags);
885
886         list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
887                 spin_lock(&data->lock);
888                 __sysmmu_disable(data);
889                 data->pgtable = 0;
890                 data->domain = NULL;
891                 list_del_init(&data->domain_node);
892                 spin_unlock(&data->lock);
893         }
894
895         spin_unlock_irqrestore(&domain->lock, flags);
896
897         dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
898                          DMA_TO_DEVICE);
899
900         for (i = 0; i < NUM_LV1ENTRIES; i++)
901                 if (lv1ent_page(domain->pgtable + i)) {
902                         phys_addr_t base = lv2table_base(domain->pgtable + i);
903
904                         dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
905                                          DMA_TO_DEVICE);
906                         kmem_cache_free(lv2table_kmem_cache,
907                                         phys_to_virt(base));
908                 }
909
910         free_pages((unsigned long)domain->pgtable, 2);
911         free_pages((unsigned long)domain->lv2entcnt, 1);
912         kfree(domain);
913 }
914
915 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
916                                     struct device *dev)
917 {
918         struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
919         struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
920         phys_addr_t pagetable = virt_to_phys(domain->pgtable);
921         struct sysmmu_drvdata *data, *next;
922         unsigned long flags;
923
924         if (!has_sysmmu(dev) || owner->domain != iommu_domain)
925                 return;
926
927         mutex_lock(&owner->rpm_lock);
928
929         list_for_each_entry(data, &owner->controllers, owner_node) {
930                 pm_runtime_get_noresume(data->sysmmu);
931                 if (pm_runtime_active(data->sysmmu))
932                         __sysmmu_disable(data);
933                 pm_runtime_put(data->sysmmu);
934         }
935
936         spin_lock_irqsave(&domain->lock, flags);
937         list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
938                 spin_lock(&data->lock);
939                 data->pgtable = 0;
940                 data->domain = NULL;
941                 list_del_init(&data->domain_node);
942                 spin_unlock(&data->lock);
943         }
944         owner->domain = NULL;
945         spin_unlock_irqrestore(&domain->lock, flags);
946
947         mutex_unlock(&owner->rpm_lock);
948
949         dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n", __func__,
950                 &pagetable);
951 }
952
953 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
954                                    struct device *dev)
955 {
956         struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
957         struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
958         struct sysmmu_drvdata *data;
959         phys_addr_t pagetable = virt_to_phys(domain->pgtable);
960         unsigned long flags;
961
962         if (!has_sysmmu(dev))
963                 return -ENODEV;
964
965         if (owner->domain)
966                 exynos_iommu_detach_device(owner->domain, dev);
967
968         mutex_lock(&owner->rpm_lock);
969
970         spin_lock_irqsave(&domain->lock, flags);
971         list_for_each_entry(data, &owner->controllers, owner_node) {
972                 spin_lock(&data->lock);
973                 data->pgtable = pagetable;
974                 data->domain = domain;
975                 list_add_tail(&data->domain_node, &domain->clients);
976                 spin_unlock(&data->lock);
977         }
978         owner->domain = iommu_domain;
979         spin_unlock_irqrestore(&domain->lock, flags);
980
981         list_for_each_entry(data, &owner->controllers, owner_node) {
982                 pm_runtime_get_noresume(data->sysmmu);
983                 if (pm_runtime_active(data->sysmmu))
984                         __sysmmu_enable(data);
985                 pm_runtime_put(data->sysmmu);
986         }
987
988         mutex_unlock(&owner->rpm_lock);
989
990         dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa\n", __func__,
991                 &pagetable);
992
993         return 0;
994 }
995
996 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
997                 sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
998 {
999         if (lv1ent_section(sent)) {
1000                 WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
1001                 return ERR_PTR(-EADDRINUSE);
1002         }
1003
1004         if (lv1ent_fault(sent)) {
1005                 dma_addr_t handle;
1006                 sysmmu_pte_t *pent;
1007                 bool need_flush_flpd_cache = lv1ent_zero(sent);
1008
1009                 pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
1010                 BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
1011                 if (!pent)
1012                         return ERR_PTR(-ENOMEM);
1013
1014                 exynos_iommu_set_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
1015                 kmemleak_ignore(pent);
1016                 *pgcounter = NUM_LV2ENTRIES;
1017                 handle = dma_map_single(dma_dev, pent, LV2TABLE_SIZE,
1018                                         DMA_TO_DEVICE);
1019                 if (dma_mapping_error(dma_dev, handle)) {
1020                         kmem_cache_free(lv2table_kmem_cache, pent);
1021                         return ERR_PTR(-EADDRINUSE);
1022                 }
1023
1024                 /*
1025                  * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
1026                  * FLPD cache may cache the address of zero_l2_table. This
1027                  * function replaces the zero_l2_table with new L2 page table
1028                  * to write valid mappings.
1029                  * Accessing the valid area may cause page fault since FLPD
1030                  * cache may still cache zero_l2_table for the valid area
1031                  * instead of new L2 page table that has the mapping
1032                  * information of the valid area.
1033                  * Thus any replacement of zero_l2_table with other valid L2
1034                  * page table must involve FLPD cache invalidation for System
1035                  * MMU v3.3.
1036                  * FLPD cache invalidation is performed with TLB invalidation
1037                  * by VPN without blocking. It is safe to invalidate TLB without
1038                  * blocking because the target address of TLB invalidation is
1039                  * not currently mapped.
1040                  */
1041                 if (need_flush_flpd_cache) {
1042                         struct sysmmu_drvdata *data;
1043
1044                         spin_lock(&domain->lock);
1045                         list_for_each_entry(data, &domain->clients, domain_node)
1046                                 sysmmu_tlb_invalidate_flpdcache(data, iova);
1047                         spin_unlock(&domain->lock);
1048                 }
1049         }
1050
1051         return page_entry(sent, iova);
1052 }
1053
1054 static int lv1set_section(struct exynos_iommu_domain *domain,
1055                           sysmmu_pte_t *sent, sysmmu_iova_t iova,
1056                           phys_addr_t paddr, int prot, short *pgcnt)
1057 {
1058         if (lv1ent_section(sent)) {
1059                 WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
1060                         iova);
1061                 return -EADDRINUSE;
1062         }
1063
1064         if (lv1ent_page(sent)) {
1065                 if (*pgcnt != NUM_LV2ENTRIES) {
1066                         WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
1067                                 iova);
1068                         return -EADDRINUSE;
1069                 }
1070
1071                 kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
1072                 *pgcnt = 0;
1073         }
1074
1075         exynos_iommu_set_pte(sent, mk_lv1ent_sect(paddr, prot));
1076
1077         spin_lock(&domain->lock);
1078         if (lv1ent_page_zero(sent)) {
1079                 struct sysmmu_drvdata *data;
1080                 /*
1081                  * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
1082                  * entry by speculative prefetch of SLPD which has no mapping.
1083                  */
1084                 list_for_each_entry(data, &domain->clients, domain_node)
1085                         sysmmu_tlb_invalidate_flpdcache(data, iova);
1086         }
1087         spin_unlock(&domain->lock);
1088
1089         return 0;
1090 }
1091
1092 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
1093                        int prot, short *pgcnt)
1094 {
1095         if (size == SPAGE_SIZE) {
1096                 if (WARN_ON(!lv2ent_fault(pent)))
1097                         return -EADDRINUSE;
1098
1099                 exynos_iommu_set_pte(pent, mk_lv2ent_spage(paddr, prot));
1100                 *pgcnt -= 1;
1101         } else { /* size == LPAGE_SIZE */
1102                 int i;
1103                 dma_addr_t pent_base = virt_to_phys(pent);
1104
1105                 dma_sync_single_for_cpu(dma_dev, pent_base,
1106                                         sizeof(*pent) * SPAGES_PER_LPAGE,
1107                                         DMA_TO_DEVICE);
1108                 for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1109                         if (WARN_ON(!lv2ent_fault(pent))) {
1110                                 if (i > 0)
1111                                         memset(pent - i, 0, sizeof(*pent) * i);
1112                                 return -EADDRINUSE;
1113                         }
1114
1115                         *pent = mk_lv2ent_lpage(paddr, prot);
1116                 }
1117                 dma_sync_single_for_device(dma_dev, pent_base,
1118                                            sizeof(*pent) * SPAGES_PER_LPAGE,
1119                                            DMA_TO_DEVICE);
1120                 *pgcnt -= SPAGES_PER_LPAGE;
1121         }
1122
1123         return 0;
1124 }
1125
1126 /*
1127  * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1128  *
1129  * System MMU v3.x has advanced logic to improve address translation
1130  * performance with caching more page table entries by a page table walk.
1131  * However, the logic has a bug that while caching faulty page table entries,
1132  * System MMU reports page fault if the cached fault entry is hit even though
1133  * the fault entry is updated to a valid entry after the entry is cached.
1134  * To prevent caching faulty page table entries which may be updated to valid
1135  * entries later, the virtual memory manager should care about the workaround
1136  * for the problem. The following describes the workaround.
1137  *
1138  * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1139  * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1140  *
1141  * Precisely, any start address of I/O virtual region must be aligned with
1142  * the following sizes for System MMU v3.1 and v3.2.
1143  * System MMU v3.1: 128KiB
1144  * System MMU v3.2: 256KiB
1145  *
1146  * Because System MMU v3.3 caches page table entries more aggressively, it needs
1147  * more workarounds.
1148  * - Any two consecutive I/O virtual regions must have a hole of size larger
1149  *   than or equal to 128KiB.
1150  * - Start address of an I/O virtual region must be aligned by 128KiB.
1151  */
1152 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1153                             unsigned long l_iova, phys_addr_t paddr, size_t size,
1154                             int prot, gfp_t gfp)
1155 {
1156         struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1157         sysmmu_pte_t *entry;
1158         sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1159         unsigned long flags;
1160         int ret = -ENOMEM;
1161
1162         BUG_ON(domain->pgtable == NULL);
1163         prot &= SYSMMU_SUPPORTED_PROT_BITS;
1164
1165         spin_lock_irqsave(&domain->pgtablelock, flags);
1166
1167         entry = section_entry(domain->pgtable, iova);
1168
1169         if (size == SECT_SIZE) {
1170                 ret = lv1set_section(domain, entry, iova, paddr, prot,
1171                                      &domain->lv2entcnt[lv1ent_offset(iova)]);
1172         } else {
1173                 sysmmu_pte_t *pent;
1174
1175                 pent = alloc_lv2entry(domain, entry, iova,
1176                                       &domain->lv2entcnt[lv1ent_offset(iova)]);
1177
1178                 if (IS_ERR(pent))
1179                         ret = PTR_ERR(pent);
1180                 else
1181                         ret = lv2set_page(pent, paddr, size, prot,
1182                                        &domain->lv2entcnt[lv1ent_offset(iova)]);
1183         }
1184
1185         if (ret)
1186                 pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1187                         __func__, ret, size, iova);
1188
1189         spin_unlock_irqrestore(&domain->pgtablelock, flags);
1190
1191         return ret;
1192 }
1193
1194 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1195                                               sysmmu_iova_t iova, size_t size)
1196 {
1197         struct sysmmu_drvdata *data;
1198         unsigned long flags;
1199
1200         spin_lock_irqsave(&domain->lock, flags);
1201
1202         list_for_each_entry(data, &domain->clients, domain_node)
1203                 sysmmu_tlb_invalidate_entry(data, iova, size);
1204
1205         spin_unlock_irqrestore(&domain->lock, flags);
1206 }
1207
1208 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1209                                  unsigned long l_iova, size_t size,
1210                                  struct iommu_iotlb_gather *gather)
1211 {
1212         struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1213         sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1214         sysmmu_pte_t *ent;
1215         size_t err_pgsize;
1216         unsigned long flags;
1217
1218         BUG_ON(domain->pgtable == NULL);
1219
1220         spin_lock_irqsave(&domain->pgtablelock, flags);
1221
1222         ent = section_entry(domain->pgtable, iova);
1223
1224         if (lv1ent_section(ent)) {
1225                 if (WARN_ON(size < SECT_SIZE)) {
1226                         err_pgsize = SECT_SIZE;
1227                         goto err;
1228                 }
1229
1230                 /* workaround for h/w bug in System MMU v3.3 */
1231                 exynos_iommu_set_pte(ent, ZERO_LV2LINK);
1232                 size = SECT_SIZE;
1233                 goto done;
1234         }
1235
1236         if (unlikely(lv1ent_fault(ent))) {
1237                 if (size > SECT_SIZE)
1238                         size = SECT_SIZE;
1239                 goto done;
1240         }
1241
1242         /* lv1ent_page(sent) == true here */
1243
1244         ent = page_entry(ent, iova);
1245
1246         if (unlikely(lv2ent_fault(ent))) {
1247                 size = SPAGE_SIZE;
1248                 goto done;
1249         }
1250
1251         if (lv2ent_small(ent)) {
1252                 exynos_iommu_set_pte(ent, 0);
1253                 size = SPAGE_SIZE;
1254                 domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1255                 goto done;
1256         }
1257
1258         /* lv1ent_large(ent) == true here */
1259         if (WARN_ON(size < LPAGE_SIZE)) {
1260                 err_pgsize = LPAGE_SIZE;
1261                 goto err;
1262         }
1263
1264         dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1265                                 sizeof(*ent) * SPAGES_PER_LPAGE,
1266                                 DMA_TO_DEVICE);
1267         memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1268         dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1269                                    sizeof(*ent) * SPAGES_PER_LPAGE,
1270                                    DMA_TO_DEVICE);
1271         size = LPAGE_SIZE;
1272         domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1273 done:
1274         spin_unlock_irqrestore(&domain->pgtablelock, flags);
1275
1276         exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1277
1278         return size;
1279 err:
1280         spin_unlock_irqrestore(&domain->pgtablelock, flags);
1281
1282         pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1283                 __func__, size, iova, err_pgsize);
1284
1285         return 0;
1286 }
1287
1288 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1289                                           dma_addr_t iova)
1290 {
1291         struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1292         sysmmu_pte_t *entry;
1293         unsigned long flags;
1294         phys_addr_t phys = 0;
1295
1296         spin_lock_irqsave(&domain->pgtablelock, flags);
1297
1298         entry = section_entry(domain->pgtable, iova);
1299
1300         if (lv1ent_section(entry)) {
1301                 phys = section_phys(entry) + section_offs(iova);
1302         } else if (lv1ent_page(entry)) {
1303                 entry = page_entry(entry, iova);
1304
1305                 if (lv2ent_large(entry))
1306                         phys = lpage_phys(entry) + lpage_offs(iova);
1307                 else if (lv2ent_small(entry))
1308                         phys = spage_phys(entry) + spage_offs(iova);
1309         }
1310
1311         spin_unlock_irqrestore(&domain->pgtablelock, flags);
1312
1313         return phys;
1314 }
1315
1316 static struct iommu_device *exynos_iommu_probe_device(struct device *dev)
1317 {
1318         struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1319         struct sysmmu_drvdata *data;
1320
1321         if (!has_sysmmu(dev))
1322                 return ERR_PTR(-ENODEV);
1323
1324         list_for_each_entry(data, &owner->controllers, owner_node) {
1325                 /*
1326                  * SYSMMU will be runtime activated via device link
1327                  * (dependency) to its master device, so there are no
1328                  * direct calls to pm_runtime_get/put in this driver.
1329                  */
1330                 data->link = device_link_add(dev, data->sysmmu,
1331                                              DL_FLAG_STATELESS |
1332                                              DL_FLAG_PM_RUNTIME);
1333         }
1334
1335         /* There is always at least one entry, see exynos_iommu_of_xlate() */
1336         data = list_first_entry(&owner->controllers,
1337                                 struct sysmmu_drvdata, owner_node);
1338
1339         return &data->iommu;
1340 }
1341
1342 static void exynos_iommu_release_device(struct device *dev)
1343 {
1344         struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1345         struct sysmmu_drvdata *data;
1346
1347         if (owner->domain) {
1348                 struct iommu_group *group = iommu_group_get(dev);
1349
1350                 if (group) {
1351                         WARN_ON(owner->domain !=
1352                                 iommu_group_default_domain(group));
1353                         exynos_iommu_detach_device(owner->domain, dev);
1354                         iommu_group_put(group);
1355                 }
1356         }
1357
1358         list_for_each_entry(data, &owner->controllers, owner_node)
1359                 device_link_del(data->link);
1360 }
1361
1362 static int exynos_iommu_of_xlate(struct device *dev,
1363                                  struct of_phandle_args *spec)
1364 {
1365         struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1366         struct exynos_iommu_owner *owner = dev_iommu_priv_get(dev);
1367         struct sysmmu_drvdata *data, *entry;
1368
1369         if (!sysmmu)
1370                 return -ENODEV;
1371
1372         data = platform_get_drvdata(sysmmu);
1373         if (!data) {
1374                 put_device(&sysmmu->dev);
1375                 return -ENODEV;
1376         }
1377
1378         if (!owner) {
1379                 owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1380                 if (!owner) {
1381                         put_device(&sysmmu->dev);
1382                         return -ENOMEM;
1383                 }
1384
1385                 INIT_LIST_HEAD(&owner->controllers);
1386                 mutex_init(&owner->rpm_lock);
1387                 dev_iommu_priv_set(dev, owner);
1388         }
1389
1390         list_for_each_entry(entry, &owner->controllers, owner_node)
1391                 if (entry == data)
1392                         return 0;
1393
1394         list_add_tail(&data->owner_node, &owner->controllers);
1395         data->master = dev;
1396
1397         return 0;
1398 }
1399
1400 static const struct iommu_ops exynos_iommu_ops = {
1401         .domain_alloc = exynos_iommu_domain_alloc,
1402         .device_group = generic_device_group,
1403         .probe_device = exynos_iommu_probe_device,
1404         .release_device = exynos_iommu_release_device,
1405         .pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1406         .of_xlate = exynos_iommu_of_xlate,
1407         .default_domain_ops = &(const struct iommu_domain_ops) {
1408                 .attach_dev     = exynos_iommu_attach_device,
1409                 .detach_dev     = exynos_iommu_detach_device,
1410                 .map            = exynos_iommu_map,
1411                 .unmap          = exynos_iommu_unmap,
1412                 .iova_to_phys   = exynos_iommu_iova_to_phys,
1413                 .free           = exynos_iommu_domain_free,
1414         }
1415 };
1416
1417 static int __init exynos_iommu_init(void)
1418 {
1419         struct device_node *np;
1420         int ret;
1421
1422         np = of_find_matching_node(NULL, sysmmu_of_match);
1423         if (!np)
1424                 return 0;
1425
1426         of_node_put(np);
1427
1428         lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1429                                 LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1430         if (!lv2table_kmem_cache) {
1431                 pr_err("%s: Failed to create kmem cache\n", __func__);
1432                 return -ENOMEM;
1433         }
1434
1435         ret = platform_driver_register(&exynos_sysmmu_driver);
1436         if (ret) {
1437                 pr_err("%s: Failed to register driver\n", __func__);
1438                 goto err_reg_driver;
1439         }
1440
1441         zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1442         if (zero_lv2_table == NULL) {
1443                 pr_err("%s: Failed to allocate zero level2 page table\n",
1444                         __func__);
1445                 ret = -ENOMEM;
1446                 goto err_zero_lv2;
1447         }
1448
1449         return 0;
1450 err_zero_lv2:
1451         platform_driver_unregister(&exynos_sysmmu_driver);
1452 err_reg_driver:
1453         kmem_cache_destroy(lv2table_kmem_cache);
1454         return ret;
1455 }
1456 core_initcall(exynos_iommu_init);