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 #define sect_to_phys(ent) (((phys_addr_t) ent) << PG_ENT_SHIFT)
74 #define section_phys(sent) (sect_to_phys(*(sent)) & SECT_MASK)
75 #define section_offs(iova) (iova & (SECT_SIZE - 1))
76 #define lpage_phys(pent) (sect_to_phys(*(pent)) & LPAGE_MASK)
77 #define lpage_offs(iova) (iova & (LPAGE_SIZE - 1))
78 #define spage_phys(pent) (sect_to_phys(*(pent)) & SPAGE_MASK)
79 #define spage_offs(iova) (iova & (SPAGE_SIZE - 1))
81 #define NUM_LV1ENTRIES 4096
82 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
84 static u32 lv1ent_offset(sysmmu_iova_t iova)
86 return iova >> SECT_ORDER;
89 static u32 lv2ent_offset(sysmmu_iova_t iova)
91 return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
94 #define LV1TABLE_SIZE (NUM_LV1ENTRIES * sizeof(sysmmu_pte_t))
95 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
97 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
98 #define lv2table_base(sent) (sect_to_phys(*(sent) & 0xFFFFFFC0))
100 #define mk_lv1ent_sect(pa) ((pa >> PG_ENT_SHIFT) | 2)
101 #define mk_lv1ent_page(pa) ((pa >> PG_ENT_SHIFT) | 1)
102 #define mk_lv2ent_lpage(pa) ((pa >> PG_ENT_SHIFT) | 1)
103 #define mk_lv2ent_spage(pa) ((pa >> PG_ENT_SHIFT) | 2)
105 #define CTRL_ENABLE 0x5
106 #define CTRL_BLOCK 0x7
107 #define CTRL_DISABLE 0x0
110 #define CFG_QOS(n) ((n & 0xF) << 7)
111 #define CFG_ACGEN (1 << 24) /* System MMU 3.3 only */
112 #define CFG_SYSSEL (1 << 22) /* System MMU 3.2 only */
113 #define CFG_FLPDCACHE (1 << 20) /* System MMU 3.2+ only */
115 /* common registers */
116 #define REG_MMU_CTRL 0x000
117 #define REG_MMU_CFG 0x004
118 #define REG_MMU_STATUS 0x008
119 #define REG_MMU_VERSION 0x034
121 #define MMU_MAJ_VER(val) ((val) >> 7)
122 #define MMU_MIN_VER(val) ((val) & 0x7F)
123 #define MMU_RAW_VER(reg) (((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
125 #define MAKE_MMU_VER(maj, min) ((((maj) & 0xF) << 7) | ((min) & 0x7F))
127 /* v1.x - v3.x registers */
128 #define REG_MMU_FLUSH 0x00C
129 #define REG_MMU_FLUSH_ENTRY 0x010
130 #define REG_PT_BASE_ADDR 0x014
131 #define REG_INT_STATUS 0x018
132 #define REG_INT_CLEAR 0x01C
134 #define REG_PAGE_FAULT_ADDR 0x024
135 #define REG_AW_FAULT_ADDR 0x028
136 #define REG_AR_FAULT_ADDR 0x02C
137 #define REG_DEFAULT_SLAVE_ADDR 0x030
140 #define REG_V5_PT_BASE_PFN 0x00C
141 #define REG_V5_MMU_FLUSH_ALL 0x010
142 #define REG_V5_MMU_FLUSH_ENTRY 0x014
143 #define REG_V5_INT_STATUS 0x060
144 #define REG_V5_INT_CLEAR 0x064
145 #define REG_V5_FAULT_AR_VA 0x070
146 #define REG_V5_FAULT_AW_VA 0x080
148 #define has_sysmmu(dev) (dev->archdata.iommu != NULL)
150 static struct device *dma_dev;
151 static struct kmem_cache *lv2table_kmem_cache;
152 static sysmmu_pte_t *zero_lv2_table;
153 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
155 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
157 return pgtable + lv1ent_offset(iova);
160 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
162 return (sysmmu_pte_t *)phys_to_virt(
163 lv2table_base(sent)) + lv2ent_offset(iova);
167 * IOMMU fault information register
169 struct sysmmu_fault_info {
170 unsigned int bit; /* bit number in STATUS register */
171 unsigned short addr_reg; /* register to read VA fault address */
172 const char *name; /* human readable fault name */
173 unsigned int type; /* fault type for report_iommu_fault */
176 static const struct sysmmu_fault_info sysmmu_faults[] = {
177 { 0, REG_PAGE_FAULT_ADDR, "PAGE", IOMMU_FAULT_READ },
178 { 1, REG_AR_FAULT_ADDR, "AR MULTI-HIT", IOMMU_FAULT_READ },
179 { 2, REG_AW_FAULT_ADDR, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
180 { 3, REG_DEFAULT_SLAVE_ADDR, "BUS ERROR", IOMMU_FAULT_READ },
181 { 4, REG_AR_FAULT_ADDR, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
182 { 5, REG_AR_FAULT_ADDR, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
183 { 6, REG_AW_FAULT_ADDR, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
184 { 7, REG_AW_FAULT_ADDR, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
187 static const struct sysmmu_fault_info sysmmu_v5_faults[] = {
188 { 0, REG_V5_FAULT_AR_VA, "AR PTW", IOMMU_FAULT_READ },
189 { 1, REG_V5_FAULT_AR_VA, "AR PAGE", IOMMU_FAULT_READ },
190 { 2, REG_V5_FAULT_AR_VA, "AR MULTI-HIT", IOMMU_FAULT_READ },
191 { 3, REG_V5_FAULT_AR_VA, "AR ACCESS PROTECTION", IOMMU_FAULT_READ },
192 { 4, REG_V5_FAULT_AR_VA, "AR SECURITY PROTECTION", IOMMU_FAULT_READ },
193 { 16, REG_V5_FAULT_AW_VA, "AW PTW", IOMMU_FAULT_WRITE },
194 { 17, REG_V5_FAULT_AW_VA, "AW PAGE", IOMMU_FAULT_WRITE },
195 { 18, REG_V5_FAULT_AW_VA, "AW MULTI-HIT", IOMMU_FAULT_WRITE },
196 { 19, REG_V5_FAULT_AW_VA, "AW ACCESS PROTECTION", IOMMU_FAULT_WRITE },
197 { 20, REG_V5_FAULT_AW_VA, "AW SECURITY PROTECTION", IOMMU_FAULT_WRITE },
201 * This structure is attached to dev.archdata.iommu of the master device
202 * on device add, contains a list of SYSMMU controllers defined by device tree,
203 * which are bound to given master device. It is usually referenced by 'owner'
206 struct exynos_iommu_owner {
207 struct list_head controllers; /* list of sysmmu_drvdata.owner_node */
208 struct iommu_domain *domain; /* domain this device is attached */
212 * This structure exynos specific generalization of struct iommu_domain.
213 * It contains list of SYSMMU controllers from all master devices, which has
214 * been attached to this domain and page tables of IO address space defined by
215 * it. It is usually referenced by 'domain' pointer.
217 struct exynos_iommu_domain {
218 struct list_head clients; /* list of sysmmu_drvdata.domain_node */
219 sysmmu_pte_t *pgtable; /* lv1 page table, 16KB */
220 short *lv2entcnt; /* free lv2 entry counter for each section */
221 spinlock_t lock; /* lock for modyfying list of clients */
222 spinlock_t pgtablelock; /* lock for modifying page table @ pgtable */
223 struct iommu_domain domain; /* generic domain data structure */
227 * This structure hold all data of a single SYSMMU controller, this includes
228 * hw resources like registers and clocks, pointers and list nodes to connect
229 * it to all other structures, internal state and parameters read from device
230 * tree. It is usually referenced by 'data' pointer.
232 struct sysmmu_drvdata {
233 struct device *sysmmu; /* SYSMMU controller device */
234 struct device *master; /* master device (owner) */
235 void __iomem *sfrbase; /* our registers */
236 struct clk *clk; /* SYSMMU's clock */
237 struct clk *aclk; /* SYSMMU's aclk clock */
238 struct clk *pclk; /* SYSMMU's pclk clock */
239 struct clk *clk_master; /* master's device clock */
240 int activations; /* number of calls to sysmmu_enable */
241 spinlock_t lock; /* lock for modyfying state */
242 struct exynos_iommu_domain *domain; /* domain we belong to */
243 struct list_head domain_node; /* node for domain clients list */
244 struct list_head owner_node; /* node for owner controllers list */
245 phys_addr_t pgtable; /* assigned page table structure */
246 unsigned int version; /* our version */
249 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
251 return container_of(dom, struct exynos_iommu_domain, domain);
254 static bool set_sysmmu_active(struct sysmmu_drvdata *data)
256 /* return true if the System MMU was not active previously
257 and it needs to be initialized */
258 return ++data->activations == 1;
261 static bool set_sysmmu_inactive(struct sysmmu_drvdata *data)
263 /* return true if the System MMU is needed to be disabled */
264 BUG_ON(data->activations < 1);
265 return --data->activations == 0;
268 static bool is_sysmmu_active(struct sysmmu_drvdata *data)
270 return data->activations > 0;
273 static void sysmmu_unblock(struct sysmmu_drvdata *data)
275 writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
278 static bool sysmmu_block(struct sysmmu_drvdata *data)
282 writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
283 while ((i > 0) && !(readl(data->sfrbase + REG_MMU_STATUS) & 1))
286 if (!(readl(data->sfrbase + REG_MMU_STATUS) & 1)) {
287 sysmmu_unblock(data);
294 static void __sysmmu_tlb_invalidate(struct sysmmu_drvdata *data)
296 if (MMU_MAJ_VER(data->version) < 5)
297 writel(0x1, data->sfrbase + REG_MMU_FLUSH);
299 writel(0x1, data->sfrbase + REG_V5_MMU_FLUSH_ALL);
302 static void __sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
303 sysmmu_iova_t iova, unsigned int num_inv)
307 for (i = 0; i < num_inv; i++) {
308 if (MMU_MAJ_VER(data->version) < 5)
309 writel((iova & SPAGE_MASK) | 1,
310 data->sfrbase + REG_MMU_FLUSH_ENTRY);
312 writel((iova & SPAGE_MASK) | 1,
313 data->sfrbase + REG_V5_MMU_FLUSH_ENTRY);
318 static void __sysmmu_set_ptbase(struct sysmmu_drvdata *data, phys_addr_t pgd)
320 if (MMU_MAJ_VER(data->version) < 5)
321 writel(pgd, data->sfrbase + REG_PT_BASE_ADDR);
323 writel(pgd >> PAGE_SHIFT,
324 data->sfrbase + REG_V5_PT_BASE_PFN);
326 __sysmmu_tlb_invalidate(data);
329 static void __sysmmu_enable_clocks(struct sysmmu_drvdata *data)
331 BUG_ON(clk_prepare_enable(data->clk_master));
332 BUG_ON(clk_prepare_enable(data->clk));
333 BUG_ON(clk_prepare_enable(data->pclk));
334 BUG_ON(clk_prepare_enable(data->aclk));
337 static void __sysmmu_disable_clocks(struct sysmmu_drvdata *data)
339 clk_disable_unprepare(data->aclk);
340 clk_disable_unprepare(data->pclk);
341 clk_disable_unprepare(data->clk);
342 clk_disable_unprepare(data->clk_master);
345 static void __sysmmu_get_version(struct sysmmu_drvdata *data)
349 __sysmmu_enable_clocks(data);
351 ver = readl(data->sfrbase + REG_MMU_VERSION);
353 /* controllers on some SoCs don't report proper version */
354 if (ver == 0x80000001u)
355 data->version = MAKE_MMU_VER(1, 0);
357 data->version = MMU_RAW_VER(ver);
359 dev_dbg(data->sysmmu, "hardware version: %d.%d\n",
360 MMU_MAJ_VER(data->version), MMU_MIN_VER(data->version));
362 __sysmmu_disable_clocks(data);
365 static void show_fault_information(struct sysmmu_drvdata *data,
366 const struct sysmmu_fault_info *finfo,
367 sysmmu_iova_t fault_addr)
371 dev_err(data->sysmmu, "%s FAULT occurred at %#x (page table base: %pa)\n",
372 finfo->name, fault_addr, &data->pgtable);
373 ent = section_entry(phys_to_virt(data->pgtable), fault_addr);
374 dev_err(data->sysmmu, "\tLv1 entry: %#x\n", *ent);
375 if (lv1ent_page(ent)) {
376 ent = page_entry(ent, fault_addr);
377 dev_err(data->sysmmu, "\t Lv2 entry: %#x\n", *ent);
381 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
383 /* SYSMMU is in blocked state when interrupt occurred. */
384 struct sysmmu_drvdata *data = dev_id;
385 const struct sysmmu_fault_info *finfo;
386 unsigned int i, n, itype;
387 sysmmu_iova_t fault_addr = -1;
388 unsigned short reg_status, reg_clear;
391 WARN_ON(!is_sysmmu_active(data));
393 if (MMU_MAJ_VER(data->version) < 5) {
394 reg_status = REG_INT_STATUS;
395 reg_clear = REG_INT_CLEAR;
396 finfo = sysmmu_faults;
397 n = ARRAY_SIZE(sysmmu_faults);
399 reg_status = REG_V5_INT_STATUS;
400 reg_clear = REG_V5_INT_CLEAR;
401 finfo = sysmmu_v5_faults;
402 n = ARRAY_SIZE(sysmmu_v5_faults);
405 spin_lock(&data->lock);
407 clk_enable(data->clk_master);
409 itype = __ffs(readl(data->sfrbase + reg_status));
410 for (i = 0; i < n; i++, finfo++)
411 if (finfo->bit == itype)
413 /* unknown/unsupported fault */
416 /* print debug message */
417 fault_addr = readl(data->sfrbase + finfo->addr_reg);
418 show_fault_information(data, finfo, fault_addr);
421 ret = report_iommu_fault(&data->domain->domain,
422 data->master, fault_addr, finfo->type);
423 /* fault is not recovered by fault handler */
426 writel(1 << itype, data->sfrbase + reg_clear);
428 sysmmu_unblock(data);
430 clk_disable(data->clk_master);
432 spin_unlock(&data->lock);
437 static void __sysmmu_disable_nocount(struct sysmmu_drvdata *data)
439 clk_enable(data->clk_master);
441 writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
442 writel(0, data->sfrbase + REG_MMU_CFG);
444 __sysmmu_disable_clocks(data);
447 static bool __sysmmu_disable(struct sysmmu_drvdata *data)
452 spin_lock_irqsave(&data->lock, flags);
454 disabled = set_sysmmu_inactive(data);
460 __sysmmu_disable_nocount(data);
462 dev_dbg(data->sysmmu, "Disabled\n");
464 dev_dbg(data->sysmmu, "%d times left to disable\n",
468 spin_unlock_irqrestore(&data->lock, flags);
473 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
477 if (data->version <= MAKE_MMU_VER(3, 1))
478 cfg = CFG_LRU | CFG_QOS(15);
479 else if (data->version <= MAKE_MMU_VER(3, 2))
480 cfg = CFG_LRU | CFG_QOS(15) | CFG_FLPDCACHE | CFG_SYSSEL;
482 cfg = CFG_QOS(15) | CFG_FLPDCACHE | CFG_ACGEN;
484 writel(cfg, data->sfrbase + REG_MMU_CFG);
487 static void __sysmmu_enable_nocount(struct sysmmu_drvdata *data)
489 __sysmmu_enable_clocks(data);
491 writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
493 __sysmmu_init_config(data);
495 __sysmmu_set_ptbase(data, data->pgtable);
497 writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
500 * SYSMMU driver keeps master's clock enabled only for the short
501 * time, while accessing the registers. For performing address
502 * translation during DMA transaction it relies on the client
503 * driver to enable it.
505 clk_disable(data->clk_master);
508 static int __sysmmu_enable(struct sysmmu_drvdata *data, phys_addr_t pgtable,
509 struct exynos_iommu_domain *domain)
514 spin_lock_irqsave(&data->lock, flags);
515 if (set_sysmmu_active(data)) {
516 data->pgtable = pgtable;
517 data->domain = domain;
519 __sysmmu_enable_nocount(data);
521 dev_dbg(data->sysmmu, "Enabled\n");
523 ret = (pgtable == data->pgtable) ? 1 : -EBUSY;
525 dev_dbg(data->sysmmu, "already enabled\n");
528 if (WARN_ON(ret < 0))
529 set_sysmmu_inactive(data); /* decrement count */
531 spin_unlock_irqrestore(&data->lock, flags);
536 static void sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
542 spin_lock_irqsave(&data->lock, flags);
543 if (is_sysmmu_active(data) && data->version >= MAKE_MMU_VER(3, 3)) {
544 clk_enable(data->clk_master);
545 __sysmmu_tlb_invalidate_entry(data, iova, 1);
546 clk_disable(data->clk_master);
548 spin_unlock_irqrestore(&data->lock, flags);
552 static void sysmmu_tlb_invalidate_entry(struct sysmmu_drvdata *data,
553 sysmmu_iova_t iova, size_t size)
557 spin_lock_irqsave(&data->lock, flags);
558 if (is_sysmmu_active(data)) {
559 unsigned int num_inv = 1;
561 clk_enable(data->clk_master);
564 * L2TLB invalidation required
565 * 4KB page: 1 invalidation
566 * 64KB page: 16 invalidations
567 * 1MB page: 64 invalidations
568 * because it is set-associative TLB
569 * with 8-way and 64 sets.
570 * 1MB page can be cached in one of all sets.
571 * 64KB page can be one of 16 consecutive sets.
573 if (MMU_MAJ_VER(data->version) == 2)
574 num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
576 if (sysmmu_block(data)) {
577 __sysmmu_tlb_invalidate_entry(data, iova, num_inv);
578 sysmmu_unblock(data);
580 clk_disable(data->clk_master);
582 dev_dbg(data->master,
583 "disabled. Skipping TLB invalidation @ %#x\n", iova);
585 spin_unlock_irqrestore(&data->lock, flags);
588 static struct iommu_ops exynos_iommu_ops;
590 static int __init exynos_sysmmu_probe(struct platform_device *pdev)
593 struct device *dev = &pdev->dev;
594 struct sysmmu_drvdata *data;
595 struct resource *res;
597 data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
601 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
602 data->sfrbase = devm_ioremap_resource(dev, res);
603 if (IS_ERR(data->sfrbase))
604 return PTR_ERR(data->sfrbase);
606 irq = platform_get_irq(pdev, 0);
608 dev_err(dev, "Unable to find IRQ resource\n");
612 ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
613 dev_name(dev), data);
615 dev_err(dev, "Unabled to register handler of irq %d\n", irq);
619 data->clk = devm_clk_get(dev, "sysmmu");
620 if (PTR_ERR(data->clk) == -ENOENT)
622 else if (IS_ERR(data->clk))
623 return PTR_ERR(data->clk);
625 data->aclk = devm_clk_get(dev, "aclk");
626 if (PTR_ERR(data->aclk) == -ENOENT)
628 else if (IS_ERR(data->aclk))
629 return PTR_ERR(data->aclk);
631 data->pclk = devm_clk_get(dev, "pclk");
632 if (PTR_ERR(data->pclk) == -ENOENT)
634 else if (IS_ERR(data->pclk))
635 return PTR_ERR(data->pclk);
637 if (!data->clk && (!data->aclk || !data->pclk)) {
638 dev_err(dev, "Failed to get device clock(s)!\n");
642 data->clk_master = devm_clk_get(dev, "master");
643 if (PTR_ERR(data->clk_master) == -ENOENT)
644 data->clk_master = NULL;
645 else if (IS_ERR(data->clk_master))
646 return PTR_ERR(data->clk_master);
649 spin_lock_init(&data->lock);
651 platform_set_drvdata(pdev, data);
653 __sysmmu_get_version(data);
654 if (PG_ENT_SHIFT < 0) {
655 if (MMU_MAJ_VER(data->version) < 5)
656 PG_ENT_SHIFT = SYSMMU_PG_ENT_SHIFT;
658 PG_ENT_SHIFT = SYSMMU_V5_PG_ENT_SHIFT;
661 pm_runtime_enable(dev);
663 of_iommu_set_ops(dev->of_node, &exynos_iommu_ops);
668 #ifdef CONFIG_PM_SLEEP
669 static int exynos_sysmmu_suspend(struct device *dev)
671 struct sysmmu_drvdata *data = dev_get_drvdata(dev);
673 dev_dbg(dev, "suspend\n");
674 if (is_sysmmu_active(data)) {
675 __sysmmu_disable_nocount(data);
681 static int exynos_sysmmu_resume(struct device *dev)
683 struct sysmmu_drvdata *data = dev_get_drvdata(dev);
685 dev_dbg(dev, "resume\n");
686 if (is_sysmmu_active(data)) {
687 pm_runtime_get_sync(dev);
688 __sysmmu_enable_nocount(data);
694 static const struct dev_pm_ops sysmmu_pm_ops = {
695 SET_LATE_SYSTEM_SLEEP_PM_OPS(exynos_sysmmu_suspend, exynos_sysmmu_resume)
698 static const struct of_device_id sysmmu_of_match[] __initconst = {
699 { .compatible = "samsung,exynos-sysmmu", },
703 static struct platform_driver exynos_sysmmu_driver __refdata = {
704 .probe = exynos_sysmmu_probe,
706 .name = "exynos-sysmmu",
707 .of_match_table = sysmmu_of_match,
708 .pm = &sysmmu_pm_ops,
709 .suppress_bind_attrs = true,
713 static inline void update_pte(sysmmu_pte_t *ent, sysmmu_pte_t val)
715 dma_sync_single_for_cpu(dma_dev, virt_to_phys(ent), sizeof(*ent),
717 *ent = cpu_to_le32(val);
718 dma_sync_single_for_device(dma_dev, virt_to_phys(ent), sizeof(*ent),
722 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
724 struct exynos_iommu_domain *domain;
728 /* Check if correct PTE offsets are initialized */
729 BUG_ON(PG_ENT_SHIFT < 0 || !dma_dev);
731 domain = kzalloc(sizeof(*domain), GFP_KERNEL);
735 if (type == IOMMU_DOMAIN_DMA) {
736 if (iommu_get_dma_cookie(&domain->domain) != 0)
738 } else if (type != IOMMU_DOMAIN_UNMANAGED) {
742 domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
743 if (!domain->pgtable)
746 domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
747 if (!domain->lv2entcnt)
750 /* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
751 for (i = 0; i < NUM_LV1ENTRIES; i += 8) {
752 domain->pgtable[i + 0] = ZERO_LV2LINK;
753 domain->pgtable[i + 1] = ZERO_LV2LINK;
754 domain->pgtable[i + 2] = ZERO_LV2LINK;
755 domain->pgtable[i + 3] = ZERO_LV2LINK;
756 domain->pgtable[i + 4] = ZERO_LV2LINK;
757 domain->pgtable[i + 5] = ZERO_LV2LINK;
758 domain->pgtable[i + 6] = ZERO_LV2LINK;
759 domain->pgtable[i + 7] = ZERO_LV2LINK;
762 handle = dma_map_single(dma_dev, domain->pgtable, LV1TABLE_SIZE,
764 /* For mapping page table entries we rely on dma == phys */
765 BUG_ON(handle != virt_to_phys(domain->pgtable));
767 spin_lock_init(&domain->lock);
768 spin_lock_init(&domain->pgtablelock);
769 INIT_LIST_HEAD(&domain->clients);
771 domain->domain.geometry.aperture_start = 0;
772 domain->domain.geometry.aperture_end = ~0UL;
773 domain->domain.geometry.force_aperture = true;
775 return &domain->domain;
778 free_pages((unsigned long)domain->pgtable, 2);
780 if (type == IOMMU_DOMAIN_DMA)
781 iommu_put_dma_cookie(&domain->domain);
787 static void exynos_iommu_domain_free(struct iommu_domain *iommu_domain)
789 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
790 struct sysmmu_drvdata *data, *next;
794 WARN_ON(!list_empty(&domain->clients));
796 spin_lock_irqsave(&domain->lock, flags);
798 list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
799 if (__sysmmu_disable(data))
801 list_del_init(&data->domain_node);
804 spin_unlock_irqrestore(&domain->lock, flags);
806 if (iommu_domain->type == IOMMU_DOMAIN_DMA)
807 iommu_put_dma_cookie(iommu_domain);
809 dma_unmap_single(dma_dev, virt_to_phys(domain->pgtable), LV1TABLE_SIZE,
812 for (i = 0; i < NUM_LV1ENTRIES; i++)
813 if (lv1ent_page(domain->pgtable + i)) {
814 phys_addr_t base = lv2table_base(domain->pgtable + i);
816 dma_unmap_single(dma_dev, base, LV2TABLE_SIZE,
818 kmem_cache_free(lv2table_kmem_cache,
822 free_pages((unsigned long)domain->pgtable, 2);
823 free_pages((unsigned long)domain->lv2entcnt, 1);
827 static void exynos_iommu_detach_device(struct iommu_domain *iommu_domain,
830 struct exynos_iommu_owner *owner = dev->archdata.iommu;
831 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
832 phys_addr_t pagetable = virt_to_phys(domain->pgtable);
833 struct sysmmu_drvdata *data, *next;
837 if (!has_sysmmu(dev) || owner->domain != iommu_domain)
840 spin_lock_irqsave(&domain->lock, flags);
841 list_for_each_entry_safe(data, next, &domain->clients, domain_node) {
842 if (data->master == dev) {
843 if (__sysmmu_disable(data)) {
845 list_del_init(&data->domain_node);
847 pm_runtime_put(data->sysmmu);
851 spin_unlock_irqrestore(&domain->lock, flags);
853 owner->domain = NULL;
856 dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n",
857 __func__, &pagetable);
859 dev_err(dev, "%s: No IOMMU is attached\n", __func__);
862 static int exynos_iommu_attach_device(struct iommu_domain *iommu_domain,
865 struct exynos_iommu_owner *owner = dev->archdata.iommu;
866 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
867 struct sysmmu_drvdata *data;
868 phys_addr_t pagetable = virt_to_phys(domain->pgtable);
872 if (!has_sysmmu(dev))
876 exynos_iommu_detach_device(owner->domain, dev);
878 list_for_each_entry(data, &owner->controllers, owner_node) {
879 pm_runtime_get_sync(data->sysmmu);
880 ret = __sysmmu_enable(data, pagetable, domain);
884 spin_lock_irqsave(&domain->lock, flags);
885 list_add_tail(&data->domain_node, &domain->clients);
886 spin_unlock_irqrestore(&domain->lock, flags);
891 dev_err(dev, "%s: Failed to attach IOMMU with pgtable %pa\n",
892 __func__, &pagetable);
896 owner->domain = iommu_domain;
897 dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa %s\n",
898 __func__, &pagetable, (ret == 0) ? "" : ", again");
903 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *domain,
904 sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
906 if (lv1ent_section(sent)) {
907 WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
908 return ERR_PTR(-EADDRINUSE);
911 if (lv1ent_fault(sent)) {
913 bool need_flush_flpd_cache = lv1ent_zero(sent);
915 pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
916 BUG_ON((uintptr_t)pent & (LV2TABLE_SIZE - 1));
918 return ERR_PTR(-ENOMEM);
920 update_pte(sent, mk_lv1ent_page(virt_to_phys(pent)));
921 kmemleak_ignore(pent);
922 *pgcounter = NUM_LV2ENTRIES;
923 dma_map_single(dma_dev, pent, LV2TABLE_SIZE, DMA_TO_DEVICE);
926 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
927 * FLPD cache may cache the address of zero_l2_table. This
928 * function replaces the zero_l2_table with new L2 page table
929 * to write valid mappings.
930 * Accessing the valid area may cause page fault since FLPD
931 * cache may still cache zero_l2_table for the valid area
932 * instead of new L2 page table that has the mapping
933 * information of the valid area.
934 * Thus any replacement of zero_l2_table with other valid L2
935 * page table must involve FLPD cache invalidation for System
937 * FLPD cache invalidation is performed with TLB invalidation
938 * by VPN without blocking. It is safe to invalidate TLB without
939 * blocking because the target address of TLB invalidation is
940 * not currently mapped.
942 if (need_flush_flpd_cache) {
943 struct sysmmu_drvdata *data;
945 spin_lock(&domain->lock);
946 list_for_each_entry(data, &domain->clients, domain_node)
947 sysmmu_tlb_invalidate_flpdcache(data, iova);
948 spin_unlock(&domain->lock);
952 return page_entry(sent, iova);
955 static int lv1set_section(struct exynos_iommu_domain *domain,
956 sysmmu_pte_t *sent, sysmmu_iova_t iova,
957 phys_addr_t paddr, short *pgcnt)
959 if (lv1ent_section(sent)) {
960 WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
965 if (lv1ent_page(sent)) {
966 if (*pgcnt != NUM_LV2ENTRIES) {
967 WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
972 kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
976 update_pte(sent, mk_lv1ent_sect(paddr));
978 spin_lock(&domain->lock);
979 if (lv1ent_page_zero(sent)) {
980 struct sysmmu_drvdata *data;
982 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
983 * entry by speculative prefetch of SLPD which has no mapping.
985 list_for_each_entry(data, &domain->clients, domain_node)
986 sysmmu_tlb_invalidate_flpdcache(data, iova);
988 spin_unlock(&domain->lock);
993 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
996 if (size == SPAGE_SIZE) {
997 if (WARN_ON(!lv2ent_fault(pent)))
1000 update_pte(pent, mk_lv2ent_spage(paddr));
1002 } else { /* size == LPAGE_SIZE */
1004 dma_addr_t pent_base = virt_to_phys(pent);
1006 dma_sync_single_for_cpu(dma_dev, pent_base,
1007 sizeof(*pent) * SPAGES_PER_LPAGE,
1009 for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
1010 if (WARN_ON(!lv2ent_fault(pent))) {
1012 memset(pent - i, 0, sizeof(*pent) * i);
1016 *pent = mk_lv2ent_lpage(paddr);
1018 dma_sync_single_for_device(dma_dev, pent_base,
1019 sizeof(*pent) * SPAGES_PER_LPAGE,
1021 *pgcnt -= SPAGES_PER_LPAGE;
1028 * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
1030 * System MMU v3.x has advanced logic to improve address translation
1031 * performance with caching more page table entries by a page table walk.
1032 * However, the logic has a bug that while caching faulty page table entries,
1033 * System MMU reports page fault if the cached fault entry is hit even though
1034 * the fault entry is updated to a valid entry after the entry is cached.
1035 * To prevent caching faulty page table entries which may be updated to valid
1036 * entries later, the virtual memory manager should care about the workaround
1037 * for the problem. The following describes the workaround.
1039 * Any two consecutive I/O virtual address regions must have a hole of 128KiB
1040 * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
1042 * Precisely, any start address of I/O virtual region must be aligned with
1043 * the following sizes for System MMU v3.1 and v3.2.
1044 * System MMU v3.1: 128KiB
1045 * System MMU v3.2: 256KiB
1047 * Because System MMU v3.3 caches page table entries more aggressively, it needs
1049 * - Any two consecutive I/O virtual regions must have a hole of size larger
1050 * than or equal to 128KiB.
1051 * - Start address of an I/O virtual region must be aligned by 128KiB.
1053 static int exynos_iommu_map(struct iommu_domain *iommu_domain,
1054 unsigned long l_iova, phys_addr_t paddr, size_t size,
1057 struct exynos_iommu_domain *domain = to_exynos_domain(iommu_domain);
1058 sysmmu_pte_t *entry;
1059 sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1060 unsigned long flags;
1063 BUG_ON(domain->pgtable == NULL);
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,
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,
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 dev->archdata.iommu = owner;
1274 list_add_tail(&data->owner_node, &owner->controllers);
1278 static struct iommu_ops exynos_iommu_ops = {
1279 .domain_alloc = exynos_iommu_domain_alloc,
1280 .domain_free = exynos_iommu_domain_free,
1281 .attach_dev = exynos_iommu_attach_device,
1282 .detach_dev = exynos_iommu_detach_device,
1283 .map = exynos_iommu_map,
1284 .unmap = exynos_iommu_unmap,
1285 .map_sg = default_iommu_map_sg,
1286 .iova_to_phys = exynos_iommu_iova_to_phys,
1287 .device_group = get_device_iommu_group,
1288 .add_device = exynos_iommu_add_device,
1289 .remove_device = exynos_iommu_remove_device,
1290 .pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1291 .of_xlate = exynos_iommu_of_xlate,
1294 static bool init_done;
1296 static int __init exynos_iommu_init(void)
1300 lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1301 LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1302 if (!lv2table_kmem_cache) {
1303 pr_err("%s: Failed to create kmem cache\n", __func__);
1307 ret = platform_driver_register(&exynos_sysmmu_driver);
1309 pr_err("%s: Failed to register driver\n", __func__);
1310 goto err_reg_driver;
1313 zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1314 if (zero_lv2_table == NULL) {
1315 pr_err("%s: Failed to allocate zero level2 page table\n",
1321 ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1323 pr_err("%s: Failed to register exynos-iommu driver.\n",
1332 kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1334 platform_driver_unregister(&exynos_sysmmu_driver);
1336 kmem_cache_destroy(lv2table_kmem_cache);
1340 static int __init exynos_iommu_of_setup(struct device_node *np)
1342 struct platform_device *pdev;
1345 exynos_iommu_init();
1347 pdev = of_platform_device_create(np, NULL, platform_bus_type.dev_root);
1349 return PTR_ERR(pdev);
1352 * use the first registered sysmmu device for performing
1353 * dma mapping operations on iommu page tables (cpu cache flush)
1356 dma_dev = &pdev->dev;
1361 IOMMU_OF_DECLARE(exynos_iommu_of, "samsung,exynos-sysmmu",
1362 exynos_iommu_of_setup);