2 * Copyright (c) 2011,2016 Samsung Electronics Co., Ltd.
3 * http://www.samsung.com
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License version 2 as
7 * published by the Free Software Foundation.
10 #ifdef CONFIG_EXYNOS_IOMMU_DEBUG
14 #include <linux/clk.h>
15 #include <linux/dma-mapping.h>
16 #include <linux/err.h>
18 #include <linux/iommu.h>
19 #include <linux/interrupt.h>
20 #include <linux/list.h>
22 #include <linux/of_iommu.h>
23 #include <linux/of_platform.h>
24 #include <linux/platform_device.h>
25 #include <linux/pm_runtime.h>
26 #include <linux/slab.h>
27 #include <linux/dma-iommu.h>
29 typedef u32 sysmmu_iova_t;
30 typedef u32 sysmmu_pte_t;
32 /* We do not consider super section mapping (16MB) */
34 #define LPAGE_ORDER 16
35 #define SPAGE_ORDER 12
37 #define SECT_SIZE (1 << SECT_ORDER)
38 #define LPAGE_SIZE (1 << LPAGE_ORDER)
39 #define SPAGE_SIZE (1 << SPAGE_ORDER)
41 #define SECT_MASK (~(SECT_SIZE - 1))
42 #define LPAGE_MASK (~(LPAGE_SIZE - 1))
43 #define SPAGE_MASK (~(SPAGE_SIZE - 1))
45 #define lv1ent_fault(sent) ((*(sent) == ZERO_LV2LINK) || \
46 ((*(sent) & 3) == 0) || ((*(sent) & 3) == 3))
47 #define lv1ent_zero(sent) (*(sent) == ZERO_LV2LINK)
48 #define lv1ent_page_zero(sent) ((*(sent) & 3) == 1)
49 #define lv1ent_page(sent) ((*(sent) != ZERO_LV2LINK) && \
51 #define lv1ent_section(sent) ((*(sent) & 3) == 2)
53 #define lv2ent_fault(pent) ((*(pent) & 3) == 0)
54 #define lv2ent_small(pent) ((*(pent) & 2) == 2)
55 #define lv2ent_large(pent) ((*(pent) & 3) == 1)
57 #ifdef CONFIG_BIG_ENDIAN
58 #warning "revisit driver if we can enable big-endian ptes"
62 * v1.x - v3.x SYSMMU supports 32bit physical and 32bit virtual address spaces
63 * v5.0 introduced support for 36bit physical address space by shifting
64 * all page entry values by 4 bits.
65 * All SYSMMU controllers in the system support the address spaces of the same
66 * size, so PG_ENT_SHIFT can be initialized on first SYSMMU probe to proper
69 static short PG_ENT_SHIFT = -1;
70 #define SYSMMU_PG_ENT_SHIFT 0
71 #define SYSMMU_V5_PG_ENT_SHIFT 4
73 static const sysmmu_pte_t *LV1_PROT;
74 static const sysmmu_pte_t SYSMMU_LV1_PROT[] = {
75 ((0 << 15) | (0 << 10)), /* no access */
76 ((1 << 15) | (1 << 10)), /* IOMMU_READ only */
77 ((0 << 15) | (1 << 10)), /* IOMMU_WRITE not supported, use read/write */
78 ((0 << 15) | (1 << 10)), /* IOMMU_READ | IOMMU_WRITE */
80 static const sysmmu_pte_t SYSMMU_V5_LV1_PROT[] = {
81 (0 << 4), /* no access */
82 (1 << 4), /* IOMMU_READ only */
83 (2 << 4), /* IOMMU_WRITE only */
84 (3 << 4), /* IOMMU_READ | IOMMU_WRITE */
87 static const sysmmu_pte_t *LV2_PROT;
88 static const sysmmu_pte_t SYSMMU_LV2_PROT[] = {
89 ((0 << 9) | (0 << 4)), /* no access */
90 ((1 << 9) | (1 << 4)), /* IOMMU_READ only */
91 ((0 << 9) | (1 << 4)), /* IOMMU_WRITE not supported, use read/write */
92 ((0 << 9) | (1 << 4)), /* IOMMU_READ | IOMMU_WRITE */
94 static const sysmmu_pte_t SYSMMU_V5_LV2_PROT[] = {
95 (0 << 2), /* no access */
96 (1 << 2), /* IOMMU_READ only */
97 (2 << 2), /* IOMMU_WRITE only */
98 (3 << 2), /* IOMMU_READ | IOMMU_WRITE */
101 #define SYSMMU_SUPPORTED_PROT_BITS (IOMMU_READ | IOMMU_WRITE)
103 #define sect_to_phys(ent) (((phys_addr_t) ent) << PG_ENT_SHIFT)
104 #define section_phys(sent) (sect_to_phys(*(sent)) & SECT_MASK)
105 #define section_offs(iova) (iova & (SECT_SIZE - 1))
106 #define lpage_phys(pent) (sect_to_phys(*(pent)) & LPAGE_MASK)
107 #define lpage_offs(iova) (iova & (LPAGE_SIZE - 1))
108 #define spage_phys(pent) (sect_to_phys(*(pent)) & SPAGE_MASK)
109 #define spage_offs(iova) (iova & (SPAGE_SIZE - 1))
111 #define NUM_LV1ENTRIES 4096
112 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
114 static u32 lv1ent_offset(sysmmu_iova_t iova)
116 return iova >> SECT_ORDER;
119 static u32 lv2ent_offset(sysmmu_iova_t iova)
121 return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
124 #define LV1TABLE_SIZE (NUM_LV1ENTRIES * sizeof(sysmmu_pte_t))
125 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
127 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
128 #define lv2table_base(sent) (sect_to_phys(*(sent) & 0xFFFFFFC0))
130 #define mk_lv1ent_sect(pa, prot) ((pa >> PG_ENT_SHIFT) | LV1_PROT[prot] | 2)
131 #define mk_lv1ent_page(pa) ((pa >> PG_ENT_SHIFT) | 1)
132 #define mk_lv2ent_lpage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 1)
133 #define mk_lv2ent_spage(pa, prot) ((pa >> PG_ENT_SHIFT) | LV2_PROT[prot] | 2)
135 #define CTRL_ENABLE 0x5
136 #define CTRL_BLOCK 0x7
137 #define CTRL_DISABLE 0x0
140 #define CFG_EAP (1 << 2)
141 #define CFG_QOS(n) ((n & 0xF) << 7)
142 #define CFG_ACGEN (1 << 24) /* System MMU 3.3 only */
143 #define CFG_SYSSEL (1 << 22) /* System MMU 3.2 only */
144 #define CFG_FLPDCACHE (1 << 20) /* System MMU 3.2+ only */
146 /* common registers */
147 #define REG_MMU_CTRL 0x000
148 #define REG_MMU_CFG 0x004
149 #define REG_MMU_STATUS 0x008
150 #define REG_MMU_VERSION 0x034
152 #define MMU_MAJ_VER(val) ((val) >> 7)
153 #define MMU_MIN_VER(val) ((val) & 0x7F)
154 #define MMU_RAW_VER(reg) (((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
156 #define MAKE_MMU_VER(maj, min) ((((maj) & 0xF) << 7) | ((min) & 0x7F))
158 /* v1.x - v3.x registers */
159 #define REG_MMU_FLUSH 0x00C
160 #define REG_MMU_FLUSH_ENTRY 0x010
161 #define REG_PT_BASE_ADDR 0x014
162 #define REG_INT_STATUS 0x018
163 #define REG_INT_CLEAR 0x01C
165 #define REG_PAGE_FAULT_ADDR 0x024
166 #define REG_AW_FAULT_ADDR 0x028
167 #define REG_AR_FAULT_ADDR 0x02C
168 #define REG_DEFAULT_SLAVE_ADDR 0x030
171 #define REG_V5_PT_BASE_PFN 0x00C
172 #define REG_V5_MMU_FLUSH_ALL 0x010
173 #define REG_V5_MMU_FLUSH_ENTRY 0x014
174 #define REG_V5_INT_STATUS 0x060
175 #define REG_V5_INT_CLEAR 0x064
176 #define REG_V5_FAULT_AR_VA 0x070
177 #define REG_V5_FAULT_AW_VA 0x080
179 #define has_sysmmu(dev) (dev->archdata.iommu != NULL)
181 static struct device *dma_dev;
182 static struct kmem_cache *lv2table_kmem_cache;
183 static sysmmu_pte_t *zero_lv2_table;
184 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
186 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
188 return pgtable + lv1ent_offset(iova);
191 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
193 return (sysmmu_pte_t *)phys_to_virt(
194 lv2table_base(sent)) + lv2ent_offset(iova);
198 * IOMMU fault information register
200 struct sysmmu_fault_info {
201 unsigned int bit; /* bit number in STATUS register */
202 unsigned short addr_reg; /* register to read VA fault address */
203 const char *name; /* human readable fault name */
204 unsigned int type; /* fault type for report_iommu_fault */
207 static const struct sysmmu_fault_info sysmmu_faults[] = {
208 { 0, REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
209 { 1, REG_AR_FAULT_ADDR, "AR MULTI-HIT", IOMMU_FAULT_READ },
210 { 2, REG_AW_FAULT_ADDR, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
211 { 3, REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
212 { 4, REG_AR_FAULT_ADDR, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
213 { 5, REG_AR_FAULT_ADDR, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
214 { 6, REG_AW_FAULT_ADDR, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
215 { 7, REG_AW_FAULT_ADDR, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
218 static const struct sysmmu_fault_info sysmmu_v5_faults[] = {
219 { 0, REG_V5_FAULT_AR_VA, "AR PTW", IOMMU_FAULT_READ },
220 { 1, REG_V5_FAULT_AR_VA, "AR PAGE", IOMMU_FAULT_READ },
221 { 2, REG_V5_FAULT_AR_VA, "AR MULTI-HIT", IOMMU_FAULT_READ },
222 { 3, REG_V5_FAULT_AR_VA, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
223 { 4, REG_V5_FAULT_AR_VA, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
224 { 16, REG_V5_FAULT_AW_VA, "AW PTW", IOMMU_FAULT_WRITE },
225 { 17, REG_V5_FAULT_AW_VA, "AW PAGE", IOMMU_FAULT_WRITE },
226 { 18, REG_V5_FAULT_AW_VA, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
227 { 19, REG_V5_FAULT_AW_VA, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
228 { 20, REG_V5_FAULT_AW_VA, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
232 * This structure is attached to dev.archdata.iommu of the master device
233 * on device add, contains a list of SYSMMU controllers defined by device tree,
234 * which are bound to given master device. It is usually referenced by 'owner'
237 struct exynos_iommu_owner {
238 struct list_head controllers; /* list of sysmmu_drvdata.owner_node */
239 struct iommu_domain *domain; /* domain this device is attached */
240 struct mutex rpm_lock; /* for runtime pm of all sysmmus */
244 * This structure exynos specific generalization of struct iommu_domain.
245 * It contains list of SYSMMU controllers from all master devices, which has
246 * been attached to this domain and page tables of IO address space defined by
247 * it. It is usually referenced by 'domain' pointer.
249 struct exynos_iommu_domain {
250 struct list_head clients; /* list of sysmmu_drvdata.domain_node */
251 sysmmu_pte_t *pgtable; /* lv1 page table, 16KB */
252 short *lv2entcnt; /* free lv2 entry counter for each section */
253 spinlock_t lock; /* lock for modyfying list of clients */
254 spinlock_t pgtablelock; /* lock for modifying page table @ pgtable */
255 struct iommu_domain domain; /* generic domain data structure */
259 * This structure hold all data of a single SYSMMU controller, this includes
260 * hw resources like registers and clocks, pointers and list nodes to connect
261 * it to all other structures, internal state and parameters read from device
262 * tree. It is usually referenced by 'data' pointer.
264 struct sysmmu_drvdata {
265 struct device *sysmmu; /* SYSMMU controller device */
266 struct device *master; /* master device (owner) */
267 void __iomem *sfrbase; /* our registers */
268 struct clk *clk; /* SYSMMU's clock */
269 struct clk *aclk; /* SYSMMU's aclk clock */
270 struct clk *pclk; /* SYSMMU's pclk clock */
271 struct clk *clk_master; /* master's device clock */
272 spinlock_t lock; /* lock for modyfying state */
273 bool active; /* current status */
274 struct exynos_iommu_domain *domain; /* domain we belong to */
275 struct list_head domain_node; /* node for domain clients list */
276 struct list_head owner_node; /* node for owner controllers list */
277 phys_addr_t pgtable; /* assigned page table structure */
278 unsigned int version; /* our version */
281 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
283 return container_of(dom, struct exynos_iommu_domain, domain);
286 static void sysmmu_unblock(struct sysmmu_drvdata *data)
288 writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
291 static bool sysmmu_block(struct sysmmu_drvdata *data)
295 writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
296 while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
299 if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
300 sysmmu_unblock(data);
307 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
309 if (MMU_MAJ_VER(data->version) < 5)
310 writel(0x1, data->sfrbase + REG_MMU_FLUSH);
312 writel(0x1, data->sfrbase + REG_V5_MMU_FLUSH_ALL);
315 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
316 sysmmu_iova_t iova, unsigned int num_inv)
320 for (i = 0; i < num_inv; i++) {
321 if (MMU_MAJ_VER(data->version) < 5)
322 writel((iova & SPAGE_MASK) | 1,
323 data->sfrbase + REG_MMU_FLUSH_ENTRY);
325 writel((iova & SPAGE_MASK) | 1,
326 data->sfrbase + REG_V5_MMU_FLUSH_ENTRY);
331 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
333 if (MMU_MAJ_VER(data->version) < 5)
334 writel(pgd, data->sfrbase + REG_PT_BASE_ADDR);
336 writel(pgd >> PAGE_SHIFT,
337 data->sfrbase + REG_V5_PT_BASE_PFN);
339 __sysmmu_tlb_invalidate(data);
342 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
344 BUG_ON(clk_prepare_enable(data->clk_master));
345 BUG_ON(clk_prepare_enable(data->clk));
346 BUG_ON(clk_prepare_enable(data->pclk));
347 BUG_ON(clk_prepare_enable(data->aclk));
350 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
352 clk_disable_unprepare(data->aclk);
353 clk_disable_unprepare(data->pclk);
354 clk_disable_unprepare(data->clk);
355 clk_disable_unprepare(data->clk_master);
358 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
362 __sysmmu_enable_clocks(data);
364 ver = readl(data->sfrbase + REG_MMU_VERSION);
366 /* controllers on some SoCs don't report proper version */
367 if (ver == 0x80000001u)
368 data->version = MAKE_MMU_VER(1, 0);
370 data->version = MMU_RAW_VER(ver);
372 dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
373 MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
375 __sysmmu_disable_clocks(data);
378 static void show_fault_information(struct sysmmu_drvdata *data,
379 const struct sysmmu_fault_info *finfo,
380 sysmmu_iova_t fault_addr)
384 dev_err(data->sysmmu, "%s: %s FAULT occurred at %#x\n",
385 dev_name(data->master), finfo->name, fault_addr);
386 dev_dbg(data->sysmmu, "Page table base: %pa\n", &data->pgtable);
387 ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
388 dev_dbg(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
389 if (lv1ent_page(ent)) {
390 ent = page_entry(ent, fault_addr);
391 dev_dbg(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
395 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
397 /* SYSMMU is in blocked state when interrupt occurred. */
398 struct sysmmu_drvdata *data = dev_id;
399 const struct sysmmu_fault_info *finfo;
400 unsigned int i, n, itype;
401 sysmmu_iova_t fault_addr = -1;
402 unsigned short reg_status, reg_clear;
405 WARN_ON(!data->active);
407 if (MMU_MAJ_VER(data->version) < 5) {
408 reg_status = REG_INT_STATUS;
409 reg_clear = REG_INT_CLEAR;
410 finfo = sysmmu_faults;
411 n = ARRAY_SIZE(sysmmu_faults);
413 reg_status = REG_V5_INT_STATUS;
414 reg_clear = REG_V5_INT_CLEAR;
415 finfo = sysmmu_v5_faults;
416 n = ARRAY_SIZE(sysmmu_v5_faults);
419 spin_lock(&data->lock);
421 clk_enable(data->clk_master);
423 itype = __ffs(readl(data->sfrbase + reg_status));
424 for (i = 0; i < n; i++, finfo++)
425 if (finfo->bit == itype)
427 /* unknown/unsupported fault */
430 /* print debug message */
431 fault_addr = readl(data->sfrbase + finfo->addr_reg);
432 show_fault_information(data, finfo, fault_addr);
435 ret = report_iommu_fault(&data->domain->domain,
436 data->master, fault_addr, finfo->type);
437 /* fault is not recovered by fault handler */
440 writel(1 << itype, data->sfrbase + reg_clear);
442 sysmmu_unblock(data);
444 clk_disable(data->clk_master);
446 spin_unlock(&data->lock);
451 static void __sysmmu_disable(struct sysmmu_drvdata *data)
455 clk_enable(data->clk_master);
457 spin_lock_irqsave(&data->lock, flags);
458 writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
459 writel(0, data->sfrbase + REG_MMU_CFG);
460 data->active = false;
461 spin_unlock_irqrestore(&data->lock, flags);
463 __sysmmu_disable_clocks(data);
466 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
470 if (data->version <= MAKE_MMU_VER(3, 1))
471 cfg = CFG_LRU | CFG_QOS(15);
472 else if (data->version <= MAKE_MMU_VER(3, 2))
473 cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
475 cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
477 cfg |= CFG_EAP; /* enable access protection bits check */
479 writel(cfg, data->sfrbase + REG_MMU_CFG);
482 static void __sysmmu_enable(struct sysmmu_drvdata *data)
486 __sysmmu_enable_clocks(data);
488 spin_lock_irqsave(&data->lock, flags);
489 writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
490 __sysmmu_init_config(data);
491 __sysmmu_set_ptbase(data, data->pgtable);
492 writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
494 spin_unlock_irqrestore(&data->lock, flags);
497 * SYSMMU driver keeps master's clock enabled only for the short
498 * time, while accessing the registers. For performing address
499 * translation during DMA transaction it relies on the client
500 * driver to enable it.
502 clk_disable(data->clk_master);
505 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
510 spin_lock_irqsave(&data->lock, flags);
511 if (data->active && data->version >= MAKE_MMU_VER(3, 3)) {
512 clk_enable(data->clk_master);
513 __sysmmu_tlb_invalidate_entry(data, iova, 1);
514 clk_disable(data->clk_master);
516 spin_unlock_irqrestore(&data->lock, flags);
519 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
520 sysmmu_iova_t iova, size_t size)
524 spin_lock_irqsave(&data->lock, flags);
526 unsigned int num_inv = 1;
528 clk_enable(data->clk_master);
531 * L2TLB invalidation required
532 * 4KB page: 1 invalidation
533 * 64KB page: 16 invalidations
534 * 1MB page: 64 invalidations
535 * because it is set-associative TLB
536 * with 8-way and 64 sets.
537 * 1MB page can be cached in one of all sets.
538 * 64KB page can be one of 16 consecutive sets.
540 if (MMU_MAJ_VER(data->version) == 2)
541 num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
543 if (sysmmu_block(data)) {
544 __sysmmu_tlb_invalidate_entry(data, iova, num_inv);
545 sysmmu_unblock(data);
547 clk_disable(data->clk_master);
549 spin_unlock_irqrestore(&data->lock, flags);
552 static struct iommu_ops exynos_iommu_ops;
554 static int __init exynos_sysmmu_probe(struct platform_device *pdev)
557 struct device *dev = &pdev->dev;
558 struct sysmmu_drvdata *data;
559 struct resource *res;
561 data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
565 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
566 data->sfrbase = devm_ioremap_resource(dev, res);
567 if (IS_ERR(data->sfrbase))
568 return PTR_ERR(data->sfrbase);
570 irq = platform_get_irq(pdev, 0);
572 dev_err(dev, "Unable to find IRQ resource\n");
576 ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
577 dev_name(dev), data);
579 dev_err(dev, "Unabled to register handler of irq %d\n", irq);
583 data->clk = devm_clk_get(dev, "sysmmu");
584 if (PTR_ERR(data->clk) == -ENOENT)
586 else if (IS_ERR(data->clk))
587 return PTR_ERR(data->clk);
589 data->aclk = devm_clk_get(dev, "aclk");
590 if (PTR_ERR(data->aclk) == -ENOENT)
592 else if (IS_ERR(data->aclk))
593 return PTR_ERR(data->aclk);
595 data->pclk = devm_clk_get(dev, "pclk");
596 if (PTR_ERR(data->pclk) == -ENOENT)
598 else if (IS_ERR(data->pclk))
599 return PTR_ERR(data->pclk);
601 if (!data->clk && (!data->aclk || !data->pclk)) {
602 dev_err(dev, "Failed to get device clock(s)!\n");
606 data->clk_master = devm_clk_get(dev, "master");
607 if (PTR_ERR(data->clk_master) == -ENOENT)
608 data->clk_master = NULL;
609 else if (IS_ERR(data->clk_master))
610 return PTR_ERR(data->clk_master);
613 spin_lock_init(&data->lock);
615 platform_set_drvdata(pdev, data);
617 __sysmmu_get_version(data);
618 if (PG_ENT_SHIFT < 0) {
619 if (MMU_MAJ_VER(data->version) < 5) {
620 PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
621 LV1_PROT = SYSMMU_LV1_PROT;
622 LV2_PROT = SYSMMU_LV2_PROT;
624 PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
625 LV1_PROT = SYSMMU_V5_LV1_PROT;
626 LV2_PROT = SYSMMU_V5_LV2_PROT;
630 pm_runtime_enable(dev);
632 of_iommu_set_ops(dev->of_node, &exynos_iommu_ops);
637 static int __maybe_unused exynos_sysmmu_suspend(struct device *dev)
639 struct sysmmu_drvdata *data = dev_get_drvdata(dev);
640 struct device *master = data->master;
643 struct exynos_iommu_owner *owner = master->archdata.iommu;
645 mutex_lock(&owner->rpm_lock);
647 dev_dbg(data->sysmmu, "saving state\n");
648 __sysmmu_disable(data);
650 mutex_unlock(&owner->rpm_lock);
655 static int __maybe_unused exynos_sysmmu_resume(struct device *dev)
657 struct sysmmu_drvdata *data = dev_get_drvdata(dev);
658 struct device *master = data->master;
661 struct exynos_iommu_owner *owner = master->archdata.iommu;
663 mutex_lock(&owner->rpm_lock);
665 dev_dbg(data->sysmmu, "restoring state\n");
666 __sysmmu_enable(data);
668 mutex_unlock(&owner->rpm_lock);
673 static const struct dev_pm_ops sysmmu_pm_ops = {
674 SET_RUNTIME_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume, NULL)
675 SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
676 pm_runtime_force_resume)
679 static const struct of_device_id sysmmu_of_match[] __initconst = {
680 { .compatible = "samsung,exynos-sysmmu", },
684 static struct platform_driver exynos_sysmmu_driver __refdata = {
685 .probe = exynos_sysmmu_probe,
687 .name = "exynos-sysmmu",
688 .of_match_table = sysmmu_of_match,
689 .pm = &sysmmu_pm_ops,
690 .suppress_bind_attrs = true,
694 static inline void update_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
696 dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
698 *ent = cpu_to_le32(val);
699 dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
703 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
705 struct exynos_iommu_domain *domain;
709 /* Check if correct PTE offsets are initialized */
710 BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
712 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
716 if (type == IOMMU_DOMAIN_DMA) {
717 if (iommu_get_dma_cookie(&domain->domain) != 0)
719 } else if (type != IOMMU_DOMAIN_UNMANAGED) {
723 domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
724 if (!domain->pgtable)
727 domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
728 if (!domain->lv2entcnt)
731 /* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
732 for (i = 0; i < NUM_LV1ENTRIES; i += 8) {
733 domain->pgtable[i + 0] = ZERO_LV2LINK;
734 domain->pgtable[i + 1] = ZERO_LV2LINK;
735 domain->pgtable[i + 2] = ZERO_LV2LINK;
736 domain->pgtable[i + 3] = ZERO_LV2LINK;
737 domain->pgtable[i + 4] = ZERO_LV2LINK;
738 domain->pgtable[i + 5] = ZERO_LV2LINK;
739 domain->pgtable[i + 6] = ZERO_LV2LINK;
740 domain->pgtable[i + 7] = ZERO_LV2LINK;
743 handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
745 /* For mapping page table entries we rely on dma == phys */
746 BUG_ON(handle != virt_to_phys(domain->pgtable));
747 if (dma_mapping_error(dma_dev, handle))
750 spin_lock_init(&domain->lock);
751 spin_lock_init(&domain->pgtablelock);
752 INIT_LIST_HEAD(&domain->clients);
754 domain->domain.geometry.aperture_start = 0;
755 domain->domain.geometry.aperture_end = ~0UL;
756 domain->domain.geometry.force_aperture = true;
758 return &domain->domain;
761 free_pages((unsigned long)domain->lv2entcnt, 1);
763 free_pages((unsigned long)domain->pgtable, 2);
765 if (type == IOMMU_DOMAIN_DMA)
766 iommu_put_dma_cookie(&domain->domain);
772 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
774 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
775 struct sysmmu_drvdata *data, *next;
779 WARN_ON(!list_empty(&domain->clients));
781 spin_lock_irqsave(&domain->lock, flags);
783 list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
784 spin_lock(&data->lock);
785 __sysmmu_disable(data);
788 list_del_init(&data->domain_node);
789 spin_unlock(&data->lock);
792 spin_unlock_irqrestore(&domain->lock, flags);
794 if (iommu_domain->type == IOMMU_DOMAIN_DMA)
795 iommu_put_dma_cookie(iommu_domain);
797 dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
800 for (i = 0; i < NUM_LV1ENTRIES; i++)
801 if (lv1ent_page(domain->pgtable + i)) {
802 phys_addr_t base = lv2table_base(domain->pgtable + i);
804 dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
806 kmem_cache_free(lv2table_kmem_cache,
810 free_pages((unsigned long)domain->pgtable, 2);
811 free_pages((unsigned long)domain->lv2entcnt, 1);
815 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
818 struct exynos_iommu_owner *owner = dev->archdata.iommu;
819 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
820 phys_addr_t pagetable = virt_to_phys(domain->pgtable);
821 struct sysmmu_drvdata *data, *next;
824 if (!has_sysmmu(dev) || owner->domain != iommu_domain)
827 mutex_lock(&owner->rpm_lock);
829 list_for_each_entry(data, &owner->controllers, owner_node) {
830 pm_runtime_get_noresume(data->sysmmu);
831 if (pm_runtime_active(data->sysmmu))
832 __sysmmu_disable(data);
833 pm_runtime_put(data->sysmmu);
836 spin_lock_irqsave(&domain->lock, flags);
837 list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
838 spin_lock(&data->lock);
841 list_del_init(&data->domain_node);
842 spin_unlock(&data->lock);
844 owner->domain = NULL;
845 spin_unlock_irqrestore(&domain->lock, flags);
847 mutex_unlock(&owner->rpm_lock);
849 dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n", __func__,
853 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
856 struct exynos_iommu_owner *owner = dev->archdata.iommu;
857 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
858 struct sysmmu_drvdata *data;
859 phys_addr_t pagetable = virt_to_phys(domain->pgtable);
862 if (!has_sysmmu(dev))
866 exynos_iommu_detach_device(owner->domain, dev);
868 mutex_lock(&owner->rpm_lock);
870 spin_lock_irqsave(&domain->lock, flags);
871 list_for_each_entry(data, &owner->controllers, owner_node) {
872 spin_lock(&data->lock);
873 data->pgtable = pagetable;
874 data->domain = domain;
875 list_add_tail(&data->domain_node, &domain->clients);
876 spin_unlock(&data->lock);
878 owner->domain = iommu_domain;
879 spin_unlock_irqrestore(&domain->lock, flags);
881 list_for_each_entry(data, &owner->controllers, owner_node) {
882 pm_runtime_get_noresume(data->sysmmu);
883 if (pm_runtime_active(data->sysmmu))
884 __sysmmu_enable(data);
885 pm_runtime_put(data->sysmmu);
888 mutex_unlock(&owner->rpm_lock);
890 dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa\n", __func__,
896 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
897 sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
899 if (lv1ent_section(sent)) {
900 WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
901 return ERR_PTR(-EADDRINUSE);
904 if (lv1ent_fault(sent)) {
907 bool need_flush_flpd_cache = lv1ent_zero(sent);
909 pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
910 BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
912 return ERR_PTR(-ENOMEM);
914 update_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
915 kmemleak_ignore(pent);
916 *pgcounter = NUM_LV2ENTRIES;
917 handle = dma_map_single(dma_dev, pent, LV2TABLE_SIZE,
919 if (dma_mapping_error(dma_dev, handle)) {
920 kmem_cache_free(lv2table_kmem_cache, pent);
921 return ERR_PTR(-EADDRINUSE);
925 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
926 * FLPD cache may cache the address of zero_l2_table. This
927 * function replaces the zero_l2_table with new L2 page table
928 * to write valid mappings.
929 * Accessing the valid area may cause page fault since FLPD
930 * cache may still cache zero_l2_table for the valid area
931 * instead of new L2 page table that has the mapping
932 * information of the valid area.
933 * Thus any replacement of zero_l2_table with other valid L2
934 * page table must involve FLPD cache invalidation for System
936 * FLPD cache invalidation is performed with TLB invalidation
937 * by VPN without blocking. It is safe to invalidate TLB without
938 * blocking because the target address of TLB invalidation is
939 * not currently mapped.
941 if (need_flush_flpd_cache) {
942 struct sysmmu_drvdata *data;
944 spin_lock(&domain->lock);
945 list_for_each_entry(data, &domain->clients, domain_node)
946 sysmmu_tlb_invalidate_flpdcache(data, iova);
947 spin_unlock(&domain->lock);
951 return page_entry(sent, iova);
954 static int lv1set_section(struct exynos_iommu_domain *domain,
955 sysmmu_pte_t *sent, sysmmu_iova_t iova,
956 phys_addr_t paddr, int prot, short *pgcnt)
958 if (lv1ent_section(sent)) {
959 WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
964 if (lv1ent_page(sent)) {
965 if (*pgcnt != NUM_LV2ENTRIES) {
966 WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
971 kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
975 update_pte(sent, mk_lv1ent_sect(paddr, prot));
977 spin_lock(&domain->lock);
978 if (lv1ent_page_zero(sent)) {
979 struct sysmmu_drvdata *data;
981 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
982 * entry by speculative prefetch of SLPD which has no mapping.
984 list_for_each_entry(data, &domain->clients, domain_node)
985 sysmmu_tlb_invalidate_flpdcache(data, iova);
987 spin_unlock(&domain->lock);
992 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
993 int prot, short *pgcnt)
995 if (size == SPAGE_SIZE) {
996 if (WARN_ON(!lv2ent_fault(pent)))
999 update_pte(pent, mk_lv2ent_spage(paddr, prot));
1001 } else { /* size == LPAGE_SIZE */
1003 dma_addr_t pent_base = virt_to_phys(pent);
1005 dma_sync_single_for_cpu(dma_dev, pent_base,
1006 sizeof(*pent) * SPAGES_PER_LPAGE,
1008 for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1009 if (WARN_ON(!lv2ent_fault(pent))) {
1011 memset(pent - i, 0, sizeof(*pent) * i);
1015 *pent = mk_lv2ent_lpage(paddr, prot);
1017 dma_sync_single_for_device(dma_dev, pent_base,
1018 sizeof(*pent) * SPAGES_PER_LPAGE,
1020 *pgcnt -= SPAGES_PER_LPAGE;
1027 * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1029 * System MMU v3.x has advanced logic to improve address translation
1030 * performance with caching more page table entries by a page table walk.
1031 * However, the logic has a bug that while caching faulty page table entries,
1032 * System MMU reports page fault if the cached fault entry is hit even though
1033 * the fault entry is updated to a valid entry after the entry is cached.
1034 * To prevent caching faulty page table entries which may be updated to valid
1035 * entries later, the virtual memory manager should care about the workaround
1036 * for the problem. The following describes the workaround.
1038 * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1039 * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1041 * Precisely, any start address of I/O virtual region must be aligned with
1042 * the following sizes for System MMU v3.1 and v3.2.
1043 * System MMU v3.1: 128KiB
1044 * System MMU v3.2: 256KiB
1046 * Because System MMU v3.3 caches page table entries more aggressively, it needs
1048 * - Any two consecutive I/O virtual regions must have a hole of size larger
1049 * than or equal to 128KiB.
1050 * - Start address of an I/O virtual region must be aligned by 128KiB.
1052 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1053 unsigned long l_iova, phys_addr_t paddr, size_t size,
1056 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1057 sysmmu_pte_t *entry;
1058 sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1059 unsigned long flags;
1062 BUG_ON(domain->pgtable == NULL);
1063 prot &= SYSMMU_SUPPORTED_PROT_BITS;
1065 spin_lock_irqsave(&domain->pgtablelock, flags);
1067 entry = section_entry(domain->pgtable, iova);
1069 if (size == SECT_SIZE) {
1070 ret = lv1set_section(domain, entry, iova, paddr, prot,
1071 &domain->lv2entcnt[lv1ent_offset(iova)]);
1075 pent = alloc_lv2entry(domain, entry, iova,
1076 &domain->lv2entcnt[lv1ent_offset(iova)]);
1079 ret = PTR_ERR(pent);
1081 ret = lv2set_page(pent, paddr, size, prot,
1082 &domain->lv2entcnt[lv1ent_offset(iova)]);
1086 pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1087 __func__, ret, size, iova);
1089 spin_unlock_irqrestore(&domain->pgtablelock, flags);
1094 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *domain,
1095 sysmmu_iova_t iova, size_t size)
1097 struct sysmmu_drvdata *data;
1098 unsigned long flags;
1100 spin_lock_irqsave(&domain->lock, flags);
1102 list_for_each_entry(data, &domain->clients, domain_node)
1103 sysmmu_tlb_invalidate_entry(data, iova, size);
1105 spin_unlock_irqrestore(&domain->lock, flags);
1108 static size_t exynos_iommu_unmap(struct iommu_domain *iommu_domain,
1109 unsigned long l_iova, size_t size)
1111 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1112 sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1115 unsigned long flags;
1117 BUG_ON(domain->pgtable == NULL);
1119 spin_lock_irqsave(&domain->pgtablelock, flags);
1121 ent = section_entry(domain->pgtable, iova);
1123 if (lv1ent_section(ent)) {
1124 if (WARN_ON(size < SECT_SIZE)) {
1125 err_pgsize = SECT_SIZE;
1129 /* workaround for h/w bug in System MMU v3.3 */
1130 update_pte(ent, ZERO_LV2LINK);
1135 if (unlikely(lv1ent_fault(ent))) {
1136 if (size > SECT_SIZE)
1141 /* lv1ent_page(sent) == true here */
1143 ent = page_entry(ent, iova);
1145 if (unlikely(lv2ent_fault(ent))) {
1150 if (lv2ent_small(ent)) {
1153 domain->lv2entcnt[lv1ent_offset(iova)] += 1;
1157 /* lv1ent_large(ent) == true here */
1158 if (WARN_ON(size < LPAGE_SIZE)) {
1159 err_pgsize = LPAGE_SIZE;
1163 dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent),
1164 sizeof(*ent) * SPAGES_PER_LPAGE,
1166 memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1167 dma_sync_single_for_device(dma_dev, virt_to_phys(ent),
1168 sizeof(*ent) * SPAGES_PER_LPAGE,
1171 domain->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1173 spin_unlock_irqrestore(&domain->pgtablelock, flags);
1175 exynos_iommu_tlb_invalidate_entry(domain, iova, size);
1179 spin_unlock_irqrestore(&domain->pgtablelock, flags);
1181 pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1182 __func__, size, iova, err_pgsize);
1187 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *iommu_domain,
1190 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1191 sysmmu_pte_t *entry;
1192 unsigned long flags;
1193 phys_addr_t phys = 0;
1195 spin_lock_irqsave(&domain->pgtablelock, flags);
1197 entry = section_entry(domain->pgtable, iova);
1199 if (lv1ent_section(entry)) {
1200 phys = section_phys(entry) + section_offs(iova);
1201 } else if (lv1ent_page(entry)) {
1202 entry = page_entry(entry, iova);
1204 if (lv2ent_large(entry))
1205 phys = lpage_phys(entry) + lpage_offs(iova);
1206 else if (lv2ent_small(entry))
1207 phys = spage_phys(entry) + spage_offs(iova);
1210 spin_unlock_irqrestore(&domain->pgtablelock, flags);
1215 static struct iommu_group *get_device_iommu_group(struct device *dev)
1217 struct iommu_group *group;
1219 group = iommu_group_get(dev);
1221 group = iommu_group_alloc();
1226 static int exynos_iommu_add_device(struct device *dev)
1228 struct iommu_group *group;
1230 if (!has_sysmmu(dev))
1233 group = iommu_group_get_for_dev(dev);
1236 return PTR_ERR(group);
1238 iommu_group_put(group);
1243 static void exynos_iommu_remove_device(struct device *dev)
1245 if (!has_sysmmu(dev))
1248 iommu_group_remove_device(dev);
1251 static int exynos_iommu_of_xlate(struct device *dev,
1252 struct of_phandle_args *spec)
1254 struct exynos_iommu_owner *owner = dev->archdata.iommu;
1255 struct platform_device *sysmmu = of_find_device_by_node(spec->np);
1256 struct sysmmu_drvdata *data;
1261 data = platform_get_drvdata(sysmmu);
1266 owner = kzalloc(sizeof(*owner), GFP_KERNEL);
1270 INIT_LIST_HEAD(&owner->controllers);
1271 mutex_init(&owner->rpm_lock);
1272 dev->archdata.iommu = owner;
1275 list_add_tail(&data->owner_node, &owner->controllers);
1279 * SYSMMU will be runtime activated via device link (dependency) to its
1280 * master device, so there are no direct calls to pm_runtime_get/put
1283 device_link_add(dev, data->sysmmu, DL_FLAG_PM_RUNTIME);
1288 static struct iommu_ops exynos_iommu_ops = {
1289 .domain_alloc = exynos_iommu_domain_alloc,
1290 .domain_free = exynos_iommu_domain_free,
1291 .attach_dev = exynos_iommu_attach_device,
1292 .detach_dev = exynos_iommu_detach_device,
1293 .map = exynos_iommu_map,
1294 .unmap = exynos_iommu_unmap,
1295 .map_sg = default_iommu_map_sg,
1296 .iova_to_phys = exynos_iommu_iova_to_phys,
1297 .device_group = get_device_iommu_group,
1298 .add_device = exynos_iommu_add_device,
1299 .remove_device = exynos_iommu_remove_device,
1300 .pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1301 .of_xlate = exynos_iommu_of_xlate,
1304 static bool init_done;
1306 static int __init exynos_iommu_init(void)
1310 lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1311 LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1312 if (!lv2table_kmem_cache) {
1313 pr_err("%s: Failed to create kmem cache\n", __func__);
1317 ret = platform_driver_register(&exynos_sysmmu_driver);
1319 pr_err("%s: Failed to register driver\n", __func__);
1320 goto err_reg_driver;
1323 zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1324 if (zero_lv2_table == NULL) {
1325 pr_err("%s: Failed to allocate zero level2 page table\n",
1331 ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1333 pr_err("%s: Failed to register exynos-iommu driver.\n",
1342 kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1344 platform_driver_unregister(&exynos_sysmmu_driver);
1346 kmem_cache_destroy(lv2table_kmem_cache);
1350 static int __init exynos_iommu_of_setup(struct device_node *np)
1352 struct platform_device *pdev;
1355 exynos_iommu_init();
1357 pdev = of_platform_device_create(np, NULL, platform_bus_type.dev_root);
1362 * use the first registered sysmmu device for performing
1363 * dma mapping operations on iommu page tables (cpu cache flush)
1366 dma_dev = &pdev->dev;
1371 IOMMU_OF_DECLARE(exynos_iommu_of, "samsung,exynos-sysmmu",
1372 exynos_iommu_of_setup);