1 /* linux/drivers/iommu/exynos_iommu.c
3 * Copyright (c) 2011 Samsung Electronics Co., Ltd.
4 * http://www.samsung.com
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 as
8 * published by the Free Software Foundation.
11 #ifdef CONFIG_EXYNOS_IOMMU_DEBUG
16 #include <linux/interrupt.h>
17 #include <linux/platform_device.h>
18 #include <linux/slab.h>
19 #include <linux/pm_runtime.h>
20 #include <linux/clk.h>
21 #include <linux/err.h>
23 #include <linux/iommu.h>
24 #include <linux/errno.h>
25 #include <linux/list.h>
26 #include <linux/memblock.h>
27 #include <linux/export.h>
29 #include <asm/cacheflush.h>
30 #include <asm/pgtable.h>
32 typedef u32 sysmmu_iova_t;
33 typedef u32 sysmmu_pte_t;
35 /* We do not consider super section mapping (16MB) */
37 #define LPAGE_ORDER 16
38 #define SPAGE_ORDER 12
40 #define SECT_SIZE (1 << SECT_ORDER)
41 #define LPAGE_SIZE (1 << LPAGE_ORDER)
42 #define SPAGE_SIZE (1 << SPAGE_ORDER)
44 #define SECT_MASK (~(SECT_SIZE - 1))
45 #define LPAGE_MASK (~(LPAGE_SIZE - 1))
46 #define SPAGE_MASK (~(SPAGE_SIZE - 1))
48 #define lv1ent_fault(sent) ((*(sent) == ZERO_LV2LINK) || \
49 ((*(sent) & 3) == 0) || ((*(sent) & 3) == 3))
50 #define lv1ent_zero(sent) (*(sent) == ZERO_LV2LINK)
51 #define lv1ent_page_zero(sent) ((*(sent) & 3) == 1)
52 #define lv1ent_page(sent) ((*(sent) != ZERO_LV2LINK) && \
54 #define lv1ent_section(sent) ((*(sent) & 3) == 2)
56 #define lv2ent_fault(pent) ((*(pent) & 3) == 0)
57 #define lv2ent_small(pent) ((*(pent) & 2) == 2)
58 #define lv2ent_large(pent) ((*(pent) & 3) == 1)
60 static u32 sysmmu_page_offset(sysmmu_iova_t iova, u32 size)
62 return iova & (size - 1);
65 #define section_phys(sent) (*(sent) & SECT_MASK)
66 #define section_offs(iova) sysmmu_page_offset((iova), SECT_SIZE)
67 #define lpage_phys(pent) (*(pent) & LPAGE_MASK)
68 #define lpage_offs(iova) sysmmu_page_offset((iova), LPAGE_SIZE)
69 #define spage_phys(pent) (*(pent) & SPAGE_MASK)
70 #define spage_offs(iova) sysmmu_page_offset((iova), SPAGE_SIZE)
72 #define NUM_LV1ENTRIES 4096
73 #define NUM_LV2ENTRIES (SECT_SIZE / SPAGE_SIZE)
75 static u32 lv1ent_offset(sysmmu_iova_t iova)
77 return iova >> SECT_ORDER;
80 static u32 lv2ent_offset(sysmmu_iova_t iova)
82 return (iova >> SPAGE_ORDER) & (NUM_LV2ENTRIES - 1);
85 #define LV2TABLE_SIZE (NUM_LV2ENTRIES * sizeof(sysmmu_pte_t))
87 #define SPAGES_PER_LPAGE (LPAGE_SIZE / SPAGE_SIZE)
89 #define lv2table_base(sent) (*(sent) & 0xFFFFFC00)
91 #define mk_lv1ent_sect(pa) ((pa) | 2)
92 #define mk_lv1ent_page(pa) ((pa) | 1)
93 #define mk_lv2ent_lpage(pa) ((pa) | 1)
94 #define mk_lv2ent_spage(pa) ((pa) | 2)
96 #define CTRL_ENABLE 0x5
97 #define CTRL_BLOCK 0x7
98 #define CTRL_DISABLE 0x0
101 #define CFG_QOS(n) ((n & 0xF) << 7)
102 #define CFG_MASK 0x0150FFFF /* Selecting bit 0-15, 20, 22 and 24 */
103 #define CFG_ACGEN (1 << 24) /* System MMU 3.3 only */
104 #define CFG_SYSSEL (1 << 22) /* System MMU 3.2 only */
105 #define CFG_FLPDCACHE (1 << 20) /* System MMU 3.2+ only */
107 #define REG_MMU_CTRL 0x000
108 #define REG_MMU_CFG 0x004
109 #define REG_MMU_STATUS 0x008
110 #define REG_MMU_FLUSH 0x00C
111 #define REG_MMU_FLUSH_ENTRY 0x010
112 #define REG_PT_BASE_ADDR 0x014
113 #define REG_INT_STATUS 0x018
114 #define REG_INT_CLEAR 0x01C
116 #define REG_PAGE_FAULT_ADDR 0x024
117 #define REG_AW_FAULT_ADDR 0x028
118 #define REG_AR_FAULT_ADDR 0x02C
119 #define REG_DEFAULT_SLAVE_ADDR 0x030
121 #define REG_MMU_VERSION 0x034
123 #define MMU_MAJ_VER(val) ((val) >> 7)
124 #define MMU_MIN_VER(val) ((val) & 0x7F)
125 #define MMU_RAW_VER(reg) (((reg) >> 21) & ((1 << 11) - 1)) /* 11 bits */
127 #define MAKE_MMU_VER(maj, min) ((((maj) & 0xF) << 7) | ((min) & 0x7F))
129 #define REG_PB0_SADDR 0x04C
130 #define REG_PB0_EADDR 0x050
131 #define REG_PB1_SADDR 0x054
132 #define REG_PB1_EADDR 0x058
134 #define has_sysmmu(dev) (dev->archdata.iommu != NULL)
136 static struct kmem_cache *lv2table_kmem_cache;
137 static sysmmu_pte_t *zero_lv2_table;
138 #define ZERO_LV2LINK mk_lv1ent_page(virt_to_phys(zero_lv2_table))
140 static sysmmu_pte_t *section_entry(sysmmu_pte_t *pgtable, sysmmu_iova_t iova)
142 return pgtable + lv1ent_offset(iova);
145 static sysmmu_pte_t *page_entry(sysmmu_pte_t *sent, sysmmu_iova_t iova)
147 return (sysmmu_pte_t *)phys_to_virt(
148 lv2table_base(sent)) + lv2ent_offset(iova);
151 enum exynos_sysmmu_inttype {
159 SYSMMU_AW_PROTECTION, /* 7 */
160 SYSMMU_FAULT_UNKNOWN,
164 static unsigned short fault_reg_offset[SYSMMU_FAULTS_NUM] = {
168 REG_DEFAULT_SLAVE_ADDR,
175 static char *sysmmu_fault_name[SYSMMU_FAULTS_NUM] = {
177 "AR MULTI-HIT FAULT",
178 "AW MULTI-HIT FAULT",
180 "AR SECURITY PROTECTION FAULT",
181 "AR ACCESS PROTECTION FAULT",
182 "AW SECURITY PROTECTION FAULT",
183 "AW ACCESS PROTECTION FAULT",
187 /* attached to dev.archdata.iommu of the master device */
188 struct exynos_iommu_owner {
189 struct list_head client; /* entry of exynos_iommu_domain.clients */
191 struct device *sysmmu;
192 struct iommu_domain *domain;
193 void *vmm_data; /* IO virtual memory manager's data */
194 spinlock_t lock; /* Lock to preserve consistency of System MMU */
197 struct exynos_iommu_domain {
198 struct list_head clients; /* list of sysmmu_drvdata.node */
199 sysmmu_pte_t *pgtable; /* lv1 page table, 16KB */
200 short *lv2entcnt; /* free lv2 entry counter for each section */
201 spinlock_t lock; /* lock for this structure */
202 spinlock_t pgtablelock; /* lock for modifying page table @ pgtable */
203 struct iommu_domain domain; /* generic domain data structure */
206 struct sysmmu_drvdata {
207 struct device *sysmmu; /* System MMU's device descriptor */
208 struct device *master; /* Owner of system MMU */
209 void __iomem *sfrbase;
211 struct clk *clk_master;
214 struct iommu_domain *domain;
218 static struct exynos_iommu_domain *to_exynos_domain(struct iommu_domain *dom)
220 return container_of(dom, struct exynos_iommu_domain, domain);
223 static bool set_sysmmu_active(struct sysmmu_drvdata *data)
225 /* return true if the System MMU was not active previously
226 and it needs to be initialized */
227 return ++data->activations == 1;
230 static bool set_sysmmu_inactive(struct sysmmu_drvdata *data)
232 /* return true if the System MMU is needed to be disabled */
233 BUG_ON(data->activations < 1);
234 return --data->activations == 0;
237 static bool is_sysmmu_active(struct sysmmu_drvdata *data)
239 return data->activations > 0;
242 static void sysmmu_unblock(void __iomem *sfrbase)
244 __raw_writel(CTRL_ENABLE, sfrbase + REG_MMU_CTRL);
247 static unsigned int __raw_sysmmu_version(struct sysmmu_drvdata *data)
249 return MMU_RAW_VER(__raw_readl(data->sfrbase + REG_MMU_VERSION));
252 static bool sysmmu_block(void __iomem *sfrbase)
256 __raw_writel(CTRL_BLOCK, sfrbase + REG_MMU_CTRL);
257 while ((i > 0) && !(__raw_readl(sfrbase + REG_MMU_STATUS) & 1))
260 if (!(__raw_readl(sfrbase + REG_MMU_STATUS) & 1)) {
261 sysmmu_unblock(sfrbase);
268 static void __sysmmu_tlb_invalidate(void __iomem *sfrbase)
270 __raw_writel(0x1, sfrbase + REG_MMU_FLUSH);
273 static void __sysmmu_tlb_invalidate_entry(void __iomem *sfrbase,
274 sysmmu_iova_t iova, unsigned int num_inv)
278 for (i = 0; i < num_inv; i++) {
279 __raw_writel((iova & SPAGE_MASK) | 1,
280 sfrbase + REG_MMU_FLUSH_ENTRY);
285 static void __sysmmu_set_ptbase(void __iomem *sfrbase,
288 __raw_writel(pgd, sfrbase + REG_PT_BASE_ADDR);
290 __sysmmu_tlb_invalidate(sfrbase);
293 static void show_fault_information(const char *name,
294 enum exynos_sysmmu_inttype itype,
295 phys_addr_t pgtable_base, sysmmu_iova_t fault_addr)
299 if ((itype >= SYSMMU_FAULTS_NUM) || (itype < SYSMMU_PAGEFAULT))
300 itype = SYSMMU_FAULT_UNKNOWN;
302 pr_err("%s occurred at %#x by %s(Page table base: %pa)\n",
303 sysmmu_fault_name[itype], fault_addr, name, &pgtable_base);
305 ent = section_entry(phys_to_virt(pgtable_base), fault_addr);
306 pr_err("\tLv1 entry: %#x\n", *ent);
308 if (lv1ent_page(ent)) {
309 ent = page_entry(ent, fault_addr);
310 pr_err("\t Lv2 entry: %#x\n", *ent);
314 static irqreturn_t exynos_sysmmu_irq(int irq, void *dev_id)
316 /* SYSMMU is in blocked state when interrupt occurred. */
317 struct sysmmu_drvdata *data = dev_id;
318 enum exynos_sysmmu_inttype itype;
319 sysmmu_iova_t addr = -1;
322 WARN_ON(!is_sysmmu_active(data));
324 spin_lock(&data->lock);
326 if (!IS_ERR(data->clk_master))
327 clk_enable(data->clk_master);
329 itype = (enum exynos_sysmmu_inttype)
330 __ffs(__raw_readl(data->sfrbase + REG_INT_STATUS));
331 if (WARN_ON(!((itype >= 0) && (itype < SYSMMU_FAULT_UNKNOWN))))
332 itype = SYSMMU_FAULT_UNKNOWN;
334 addr = __raw_readl(data->sfrbase + fault_reg_offset[itype]);
336 if (itype == SYSMMU_FAULT_UNKNOWN) {
337 pr_err("%s: Fault is not occurred by System MMU '%s'!\n",
338 __func__, dev_name(data->sysmmu));
339 pr_err("%s: Please check if IRQ is correctly configured.\n",
344 __raw_readl(data->sfrbase + REG_PT_BASE_ADDR);
345 show_fault_information(dev_name(data->sysmmu),
348 ret = report_iommu_fault(data->domain,
349 data->master, addr, itype);
352 /* fault is not recovered by fault handler */
355 __raw_writel(1 << itype, data->sfrbase + REG_INT_CLEAR);
357 sysmmu_unblock(data->sfrbase);
359 if (!IS_ERR(data->clk_master))
360 clk_disable(data->clk_master);
362 spin_unlock(&data->lock);
367 static void __sysmmu_disable_nocount(struct sysmmu_drvdata *data)
369 if (!IS_ERR(data->clk_master))
370 clk_enable(data->clk_master);
372 __raw_writel(CTRL_DISABLE, data->sfrbase + REG_MMU_CTRL);
373 __raw_writel(0, data->sfrbase + REG_MMU_CFG);
375 clk_disable(data->clk);
376 if (!IS_ERR(data->clk_master))
377 clk_disable(data->clk_master);
380 static bool __sysmmu_disable(struct sysmmu_drvdata *data)
385 spin_lock_irqsave(&data->lock, flags);
387 disabled = set_sysmmu_inactive(data);
393 __sysmmu_disable_nocount(data);
395 dev_dbg(data->sysmmu, "Disabled\n");
397 dev_dbg(data->sysmmu, "%d times left to disable\n",
401 spin_unlock_irqrestore(&data->lock, flags);
406 static void __sysmmu_init_config(struct sysmmu_drvdata *data)
408 unsigned int cfg = CFG_LRU | CFG_QOS(15);
411 ver = __raw_sysmmu_version(data);
412 if (MMU_MAJ_VER(ver) == 3) {
413 if (MMU_MIN_VER(ver) >= 2) {
414 cfg |= CFG_FLPDCACHE;
415 if (MMU_MIN_VER(ver) == 3) {
424 __raw_writel(cfg, data->sfrbase + REG_MMU_CFG);
427 static void __sysmmu_enable_nocount(struct sysmmu_drvdata *data)
429 if (!IS_ERR(data->clk_master))
430 clk_enable(data->clk_master);
431 clk_enable(data->clk);
433 __raw_writel(CTRL_BLOCK, data->sfrbase + REG_MMU_CTRL);
435 __sysmmu_init_config(data);
437 __sysmmu_set_ptbase(data->sfrbase, data->pgtable);
439 __raw_writel(CTRL_ENABLE, data->sfrbase + REG_MMU_CTRL);
441 if (!IS_ERR(data->clk_master))
442 clk_disable(data->clk_master);
445 static int __sysmmu_enable(struct sysmmu_drvdata *data,
446 phys_addr_t pgtable, struct iommu_domain *domain)
451 spin_lock_irqsave(&data->lock, flags);
452 if (set_sysmmu_active(data)) {
453 data->pgtable = pgtable;
454 data->domain = domain;
456 __sysmmu_enable_nocount(data);
458 dev_dbg(data->sysmmu, "Enabled\n");
460 ret = (pgtable == data->pgtable) ? 1 : -EBUSY;
462 dev_dbg(data->sysmmu, "already enabled\n");
465 if (WARN_ON(ret < 0))
466 set_sysmmu_inactive(data); /* decrement count */
468 spin_unlock_irqrestore(&data->lock, flags);
473 /* __exynos_sysmmu_enable: Enables System MMU
475 * returns -error if an error occurred and System MMU is not enabled,
476 * 0 if the System MMU has been just enabled and 1 if System MMU was already
479 static int __exynos_sysmmu_enable(struct device *dev, phys_addr_t pgtable,
480 struct iommu_domain *domain)
484 struct exynos_iommu_owner *owner = dev->archdata.iommu;
485 struct sysmmu_drvdata *data;
487 BUG_ON(!has_sysmmu(dev));
489 spin_lock_irqsave(&owner->lock, flags);
491 data = dev_get_drvdata(owner->sysmmu);
493 ret = __sysmmu_enable(data, pgtable, domain);
497 spin_unlock_irqrestore(&owner->lock, flags);
502 int exynos_sysmmu_enable(struct device *dev, phys_addr_t pgtable)
504 BUG_ON(!memblock_is_memory(pgtable));
506 return __exynos_sysmmu_enable(dev, pgtable, NULL);
509 static bool exynos_sysmmu_disable(struct device *dev)
512 bool disabled = true;
513 struct exynos_iommu_owner *owner = dev->archdata.iommu;
514 struct sysmmu_drvdata *data;
516 BUG_ON(!has_sysmmu(dev));
518 spin_lock_irqsave(&owner->lock, flags);
520 data = dev_get_drvdata(owner->sysmmu);
522 disabled = __sysmmu_disable(data);
526 spin_unlock_irqrestore(&owner->lock, flags);
531 static void __sysmmu_tlb_invalidate_flpdcache(struct sysmmu_drvdata *data,
534 if (__raw_sysmmu_version(data) == MAKE_MMU_VER(3, 3))
535 __raw_writel(iova | 0x1, data->sfrbase + REG_MMU_FLUSH_ENTRY);
538 static void sysmmu_tlb_invalidate_flpdcache(struct device *dev,
542 struct exynos_iommu_owner *owner = dev->archdata.iommu;
543 struct sysmmu_drvdata *data = dev_get_drvdata(owner->sysmmu);
545 if (!IS_ERR(data->clk_master))
546 clk_enable(data->clk_master);
548 spin_lock_irqsave(&data->lock, flags);
549 if (is_sysmmu_active(data))
550 __sysmmu_tlb_invalidate_flpdcache(data, iova);
551 spin_unlock_irqrestore(&data->lock, flags);
553 if (!IS_ERR(data->clk_master))
554 clk_disable(data->clk_master);
557 static void sysmmu_tlb_invalidate_entry(struct device *dev, sysmmu_iova_t iova,
560 struct exynos_iommu_owner *owner = dev->archdata.iommu;
562 struct sysmmu_drvdata *data;
564 data = dev_get_drvdata(owner->sysmmu);
566 spin_lock_irqsave(&data->lock, flags);
567 if (is_sysmmu_active(data)) {
568 unsigned int num_inv = 1;
570 if (!IS_ERR(data->clk_master))
571 clk_enable(data->clk_master);
574 * L2TLB invalidation required
575 * 4KB page: 1 invalidation
576 * 64KB page: 16 invalidations
577 * 1MB page: 64 invalidations
578 * because it is set-associative TLB
579 * with 8-way and 64 sets.
580 * 1MB page can be cached in one of all sets.
581 * 64KB page can be one of 16 consecutive sets.
583 if (MMU_MAJ_VER(__raw_sysmmu_version(data)) == 2)
584 num_inv = min_t(unsigned int, size / PAGE_SIZE, 64);
586 if (sysmmu_block(data->sfrbase)) {
587 __sysmmu_tlb_invalidate_entry(
588 data->sfrbase, iova, num_inv);
589 sysmmu_unblock(data->sfrbase);
591 if (!IS_ERR(data->clk_master))
592 clk_disable(data->clk_master);
594 dev_dbg(dev, "disabled. Skipping TLB invalidation @ %#x\n",
597 spin_unlock_irqrestore(&data->lock, flags);
600 void exynos_sysmmu_tlb_invalidate(struct device *dev)
602 struct exynos_iommu_owner *owner = dev->archdata.iommu;
604 struct sysmmu_drvdata *data;
606 data = dev_get_drvdata(owner->sysmmu);
608 spin_lock_irqsave(&data->lock, flags);
609 if (is_sysmmu_active(data)) {
610 if (!IS_ERR(data->clk_master))
611 clk_enable(data->clk_master);
612 if (sysmmu_block(data->sfrbase)) {
613 __sysmmu_tlb_invalidate(data->sfrbase);
614 sysmmu_unblock(data->sfrbase);
616 if (!IS_ERR(data->clk_master))
617 clk_disable(data->clk_master);
619 dev_dbg(dev, "disabled. Skipping TLB invalidation\n");
621 spin_unlock_irqrestore(&data->lock, flags);
624 static int __init exynos_sysmmu_probe(struct platform_device *pdev)
627 struct device *dev = &pdev->dev;
628 struct sysmmu_drvdata *data;
629 struct resource *res;
631 data = devm_kzalloc(dev, sizeof(*data), GFP_KERNEL);
635 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
636 data->sfrbase = devm_ioremap_resource(dev, res);
637 if (IS_ERR(data->sfrbase))
638 return PTR_ERR(data->sfrbase);
640 irq = platform_get_irq(pdev, 0);
642 dev_err(dev, "Unable to find IRQ resource\n");
646 ret = devm_request_irq(dev, irq, exynos_sysmmu_irq, 0,
647 dev_name(dev), data);
649 dev_err(dev, "Unabled to register handler of irq %d\n", irq);
653 data->clk = devm_clk_get(dev, "sysmmu");
654 if (IS_ERR(data->clk)) {
655 dev_err(dev, "Failed to get clock!\n");
656 return PTR_ERR(data->clk);
658 ret = clk_prepare(data->clk);
660 dev_err(dev, "Failed to prepare clk\n");
665 data->clk_master = devm_clk_get(dev, "master");
666 if (!IS_ERR(data->clk_master)) {
667 ret = clk_prepare(data->clk_master);
669 clk_unprepare(data->clk);
670 dev_err(dev, "Failed to prepare master's clk\n");
676 spin_lock_init(&data->lock);
678 platform_set_drvdata(pdev, data);
680 pm_runtime_enable(dev);
685 static const struct of_device_id sysmmu_of_match[] __initconst = {
686 { .compatible = "samsung,exynos-sysmmu", },
690 static struct platform_driver exynos_sysmmu_driver __refdata = {
691 .probe = exynos_sysmmu_probe,
693 .name = "exynos-sysmmu",
694 .of_match_table = sysmmu_of_match,
698 static inline void pgtable_flush(void *vastart, void *vaend)
700 dmac_flush_range(vastart, vaend);
701 outer_flush_range(virt_to_phys(vastart),
702 virt_to_phys(vaend));
705 static struct iommu_domain *exynos_iommu_domain_alloc(unsigned type)
707 struct exynos_iommu_domain *exynos_domain;
710 if (type != IOMMU_DOMAIN_UNMANAGED)
713 exynos_domain = kzalloc(sizeof(*exynos_domain), GFP_KERNEL);
717 exynos_domain->pgtable = (sysmmu_pte_t *)__get_free_pages(GFP_KERNEL, 2);
718 if (!exynos_domain->pgtable)
721 exynos_domain->lv2entcnt = (short *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, 1);
722 if (!exynos_domain->lv2entcnt)
725 /* Workaround for System MMU v3.3 to prevent caching 1MiB mapping */
726 for (i = 0; i < NUM_LV1ENTRIES; i += 8) {
727 exynos_domain->pgtable[i + 0] = ZERO_LV2LINK;
728 exynos_domain->pgtable[i + 1] = ZERO_LV2LINK;
729 exynos_domain->pgtable[i + 2] = ZERO_LV2LINK;
730 exynos_domain->pgtable[i + 3] = ZERO_LV2LINK;
731 exynos_domain->pgtable[i + 4] = ZERO_LV2LINK;
732 exynos_domain->pgtable[i + 5] = ZERO_LV2LINK;
733 exynos_domain->pgtable[i + 6] = ZERO_LV2LINK;
734 exynos_domain->pgtable[i + 7] = ZERO_LV2LINK;
737 pgtable_flush(exynos_domain->pgtable, exynos_domain->pgtable + NUM_LV1ENTRIES);
739 spin_lock_init(&exynos_domain->lock);
740 spin_lock_init(&exynos_domain->pgtablelock);
741 INIT_LIST_HEAD(&exynos_domain->clients);
743 exynos_domain->domain.geometry.aperture_start = 0;
744 exynos_domain->domain.geometry.aperture_end = ~0UL;
745 exynos_domain->domain.geometry.force_aperture = true;
747 return &exynos_domain->domain;
750 free_pages((unsigned long)exynos_domain->pgtable, 2);
752 kfree(exynos_domain);
756 static void exynos_iommu_domain_free(struct iommu_domain *domain)
758 struct exynos_iommu_domain *priv = to_exynos_domain(domain);
759 struct exynos_iommu_owner *owner;
763 WARN_ON(!list_empty(&priv->clients));
765 spin_lock_irqsave(&priv->lock, flags);
767 list_for_each_entry(owner, &priv->clients, client) {
768 while (!exynos_sysmmu_disable(owner->dev))
769 ; /* until System MMU is actually disabled */
772 while (!list_empty(&priv->clients))
773 list_del_init(priv->clients.next);
775 spin_unlock_irqrestore(&priv->lock, flags);
777 for (i = 0; i < NUM_LV1ENTRIES; i++)
778 if (lv1ent_page(priv->pgtable + i))
779 kmem_cache_free(lv2table_kmem_cache,
780 phys_to_virt(lv2table_base(priv->pgtable + i)));
782 free_pages((unsigned long)priv->pgtable, 2);
783 free_pages((unsigned long)priv->lv2entcnt, 1);
787 static int exynos_iommu_attach_device(struct iommu_domain *domain,
790 struct exynos_iommu_owner *owner = dev->archdata.iommu;
791 struct exynos_iommu_domain *priv = to_exynos_domain(domain);
792 phys_addr_t pagetable = virt_to_phys(priv->pgtable);
796 spin_lock_irqsave(&priv->lock, flags);
798 ret = __exynos_sysmmu_enable(dev, pagetable, domain);
800 list_add_tail(&owner->client, &priv->clients);
801 owner->domain = domain;
804 spin_unlock_irqrestore(&priv->lock, flags);
807 dev_err(dev, "%s: Failed to attach IOMMU with pgtable %pa\n",
808 __func__, &pagetable);
812 dev_dbg(dev, "%s: Attached IOMMU with pgtable %pa %s\n",
813 __func__, &pagetable, (ret == 0) ? "" : ", again");
818 static void exynos_iommu_detach_device(struct iommu_domain *domain,
821 struct exynos_iommu_owner *owner;
822 struct exynos_iommu_domain *priv = to_exynos_domain(domain);
823 phys_addr_t pagetable = virt_to_phys(priv->pgtable);
826 spin_lock_irqsave(&priv->lock, flags);
828 list_for_each_entry(owner, &priv->clients, client) {
829 if (owner == dev->archdata.iommu) {
830 if (exynos_sysmmu_disable(dev)) {
831 list_del_init(&owner->client);
832 owner->domain = NULL;
838 spin_unlock_irqrestore(&priv->lock, flags);
840 if (owner == dev->archdata.iommu)
841 dev_dbg(dev, "%s: Detached IOMMU with pgtable %pa\n",
842 __func__, &pagetable);
844 dev_err(dev, "%s: No IOMMU is attached\n", __func__);
847 static sysmmu_pte_t *alloc_lv2entry(struct exynos_iommu_domain *priv,
848 sysmmu_pte_t *sent, sysmmu_iova_t iova, short *pgcounter)
850 if (lv1ent_section(sent)) {
851 WARN(1, "Trying mapping on %#08x mapped with 1MiB page", iova);
852 return ERR_PTR(-EADDRINUSE);
855 if (lv1ent_fault(sent)) {
857 bool need_flush_flpd_cache = lv1ent_zero(sent);
859 pent = kmem_cache_zalloc(lv2table_kmem_cache, GFP_ATOMIC);
860 BUG_ON((unsigned int)pent & (LV2TABLE_SIZE - 1));
862 return ERR_PTR(-ENOMEM);
864 *sent = mk_lv1ent_page(virt_to_phys(pent));
865 *pgcounter = NUM_LV2ENTRIES;
866 pgtable_flush(pent, pent + NUM_LV2ENTRIES);
867 pgtable_flush(sent, sent + 1);
870 * If pre-fetched SLPD is a faulty SLPD in zero_l2_table,
871 * FLPD cache may cache the address of zero_l2_table. This
872 * function replaces the zero_l2_table with new L2 page table
873 * to write valid mappings.
874 * Accessing the valid area may cause page fault since FLPD
875 * cache may still cache zero_l2_table for the valid area
876 * instead of new L2 page table that has the mapping
877 * information of the valid area.
878 * Thus any replacement of zero_l2_table with other valid L2
879 * page table must involve FLPD cache invalidation for System
881 * FLPD cache invalidation is performed with TLB invalidation
882 * by VPN without blocking. It is safe to invalidate TLB without
883 * blocking because the target address of TLB invalidation is
884 * not currently mapped.
886 if (need_flush_flpd_cache) {
887 struct exynos_iommu_owner *owner;
889 spin_lock(&priv->lock);
890 list_for_each_entry(owner, &priv->clients, client)
891 sysmmu_tlb_invalidate_flpdcache(
893 spin_unlock(&priv->lock);
897 return page_entry(sent, iova);
900 static int lv1set_section(struct exynos_iommu_domain *priv,
901 sysmmu_pte_t *sent, sysmmu_iova_t iova,
902 phys_addr_t paddr, short *pgcnt)
904 if (lv1ent_section(sent)) {
905 WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
910 if (lv1ent_page(sent)) {
911 if (*pgcnt != NUM_LV2ENTRIES) {
912 WARN(1, "Trying mapping on 1MiB@%#08x that is mapped",
917 kmem_cache_free(lv2table_kmem_cache, page_entry(sent, 0));
921 *sent = mk_lv1ent_sect(paddr);
923 pgtable_flush(sent, sent + 1);
925 spin_lock(&priv->lock);
926 if (lv1ent_page_zero(sent)) {
927 struct exynos_iommu_owner *owner;
929 * Flushing FLPD cache in System MMU v3.3 that may cache a FLPD
930 * entry by speculative prefetch of SLPD which has no mapping.
932 list_for_each_entry(owner, &priv->clients, client)
933 sysmmu_tlb_invalidate_flpdcache(owner->dev, iova);
935 spin_unlock(&priv->lock);
940 static int lv2set_page(sysmmu_pte_t *pent, phys_addr_t paddr, size_t size,
943 if (size == SPAGE_SIZE) {
944 if (WARN_ON(!lv2ent_fault(pent)))
947 *pent = mk_lv2ent_spage(paddr);
948 pgtable_flush(pent, pent + 1);
950 } else { /* size == LPAGE_SIZE */
953 for (i = 0; i < SPAGES_PER_LPAGE; i++, pent++) {
954 if (WARN_ON(!lv2ent_fault(pent))) {
956 memset(pent - i, 0, sizeof(*pent) * i);
960 *pent = mk_lv2ent_lpage(paddr);
962 pgtable_flush(pent - SPAGES_PER_LPAGE, pent);
963 *pgcnt -= SPAGES_PER_LPAGE;
970 * *CAUTION* to the I/O virtual memory managers that support exynos-iommu:
972 * System MMU v3.x has advanced logic to improve address translation
973 * performance with caching more page table entries by a page table walk.
974 * However, the logic has a bug that while caching faulty page table entries,
975 * System MMU reports page fault if the cached fault entry is hit even though
976 * the fault entry is updated to a valid entry after the entry is cached.
977 * To prevent caching faulty page table entries which may be updated to valid
978 * entries later, the virtual memory manager should care about the workaround
979 * for the problem. The following describes the workaround.
981 * Any two consecutive I/O virtual address regions must have a hole of 128KiB
982 * at maximum to prevent misbehavior of System MMU 3.x (workaround for h/w bug).
984 * Precisely, any start address of I/O virtual region must be aligned with
985 * the following sizes for System MMU v3.1 and v3.2.
986 * System MMU v3.1: 128KiB
987 * System MMU v3.2: 256KiB
989 * Because System MMU v3.3 caches page table entries more aggressively, it needs
991 * - Any two consecutive I/O virtual regions must have a hole of size larger
992 * than or equal to 128KiB.
993 * - Start address of an I/O virtual region must be aligned by 128KiB.
995 static int exynos_iommu_map(struct iommu_domain *domain, unsigned long l_iova,
996 phys_addr_t paddr, size_t size, int prot)
998 struct exynos_iommu_domain *priv = to_exynos_domain(domain);
1000 sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1001 unsigned long flags;
1004 BUG_ON(priv->pgtable == NULL);
1006 spin_lock_irqsave(&priv->pgtablelock, flags);
1008 entry = section_entry(priv->pgtable, iova);
1010 if (size == SECT_SIZE) {
1011 ret = lv1set_section(priv, entry, iova, paddr,
1012 &priv->lv2entcnt[lv1ent_offset(iova)]);
1016 pent = alloc_lv2entry(priv, entry, iova,
1017 &priv->lv2entcnt[lv1ent_offset(iova)]);
1020 ret = PTR_ERR(pent);
1022 ret = lv2set_page(pent, paddr, size,
1023 &priv->lv2entcnt[lv1ent_offset(iova)]);
1027 pr_err("%s: Failed(%d) to map %#zx bytes @ %#x\n",
1028 __func__, ret, size, iova);
1030 spin_unlock_irqrestore(&priv->pgtablelock, flags);
1035 static void exynos_iommu_tlb_invalidate_entry(struct exynos_iommu_domain *priv,
1036 sysmmu_iova_t iova, size_t size)
1038 struct exynos_iommu_owner *owner;
1039 unsigned long flags;
1041 spin_lock_irqsave(&priv->lock, flags);
1043 list_for_each_entry(owner, &priv->clients, client)
1044 sysmmu_tlb_invalidate_entry(owner->dev, iova, size);
1046 spin_unlock_irqrestore(&priv->lock, flags);
1049 static size_t exynos_iommu_unmap(struct iommu_domain *domain,
1050 unsigned long l_iova, size_t size)
1052 struct exynos_iommu_domain *priv = to_exynos_domain(domain);
1053 sysmmu_iova_t iova = (sysmmu_iova_t)l_iova;
1056 unsigned long flags;
1058 BUG_ON(priv->pgtable == NULL);
1060 spin_lock_irqsave(&priv->pgtablelock, flags);
1062 ent = section_entry(priv->pgtable, iova);
1064 if (lv1ent_section(ent)) {
1065 if (WARN_ON(size < SECT_SIZE)) {
1066 err_pgsize = SECT_SIZE;
1070 /* workaround for h/w bug in System MMU v3.3 */
1071 *ent = ZERO_LV2LINK;
1072 pgtable_flush(ent, ent + 1);
1077 if (unlikely(lv1ent_fault(ent))) {
1078 if (size > SECT_SIZE)
1083 /* lv1ent_page(sent) == true here */
1085 ent = page_entry(ent, iova);
1087 if (unlikely(lv2ent_fault(ent))) {
1092 if (lv2ent_small(ent)) {
1095 pgtable_flush(ent, ent + 1);
1096 priv->lv2entcnt[lv1ent_offset(iova)] += 1;
1100 /* lv1ent_large(ent) == true here */
1101 if (WARN_ON(size < LPAGE_SIZE)) {
1102 err_pgsize = LPAGE_SIZE;
1106 memset(ent, 0, sizeof(*ent) * SPAGES_PER_LPAGE);
1107 pgtable_flush(ent, ent + SPAGES_PER_LPAGE);
1110 priv->lv2entcnt[lv1ent_offset(iova)] += SPAGES_PER_LPAGE;
1112 spin_unlock_irqrestore(&priv->pgtablelock, flags);
1114 exynos_iommu_tlb_invalidate_entry(priv, iova, size);
1118 spin_unlock_irqrestore(&priv->pgtablelock, flags);
1120 pr_err("%s: Failed: size(%#zx) @ %#x is smaller than page size %#zx\n",
1121 __func__, size, iova, err_pgsize);
1126 static phys_addr_t exynos_iommu_iova_to_phys(struct iommu_domain *domain,
1129 struct exynos_iommu_domain *priv = to_exynos_domain(domain);
1130 sysmmu_pte_t *entry;
1131 unsigned long flags;
1132 phys_addr_t phys = 0;
1134 spin_lock_irqsave(&priv->pgtablelock, flags);
1136 entry = section_entry(priv->pgtable, iova);
1138 if (lv1ent_section(entry)) {
1139 phys = section_phys(entry) + section_offs(iova);
1140 } else if (lv1ent_page(entry)) {
1141 entry = page_entry(entry, iova);
1143 if (lv2ent_large(entry))
1144 phys = lpage_phys(entry) + lpage_offs(iova);
1145 else if (lv2ent_small(entry))
1146 phys = spage_phys(entry) + spage_offs(iova);
1149 spin_unlock_irqrestore(&priv->pgtablelock, flags);
1154 static int exynos_iommu_add_device(struct device *dev)
1156 struct iommu_group *group;
1159 group = iommu_group_get(dev);
1162 group = iommu_group_alloc();
1163 if (IS_ERR(group)) {
1164 dev_err(dev, "Failed to allocate IOMMU group\n");
1165 return PTR_ERR(group);
1169 ret = iommu_group_add_device(group, dev);
1170 iommu_group_put(group);
1175 static void exynos_iommu_remove_device(struct device *dev)
1177 iommu_group_remove_device(dev);
1180 static const struct iommu_ops exynos_iommu_ops = {
1181 .domain_alloc = exynos_iommu_domain_alloc,
1182 .domain_free = exynos_iommu_domain_free,
1183 .attach_dev = exynos_iommu_attach_device,
1184 .detach_dev = exynos_iommu_detach_device,
1185 .map = exynos_iommu_map,
1186 .unmap = exynos_iommu_unmap,
1187 .map_sg = default_iommu_map_sg,
1188 .iova_to_phys = exynos_iommu_iova_to_phys,
1189 .add_device = exynos_iommu_add_device,
1190 .remove_device = exynos_iommu_remove_device,
1191 .pgsize_bitmap = SECT_SIZE | LPAGE_SIZE | SPAGE_SIZE,
1194 static int __init exynos_iommu_init(void)
1196 struct device_node *np;
1199 np = of_find_matching_node(NULL, sysmmu_of_match);
1205 lv2table_kmem_cache = kmem_cache_create("exynos-iommu-lv2table",
1206 LV2TABLE_SIZE, LV2TABLE_SIZE, 0, NULL);
1207 if (!lv2table_kmem_cache) {
1208 pr_err("%s: Failed to create kmem cache\n", __func__);
1212 ret = platform_driver_register(&exynos_sysmmu_driver);
1214 pr_err("%s: Failed to register driver\n", __func__);
1215 goto err_reg_driver;
1218 zero_lv2_table = kmem_cache_zalloc(lv2table_kmem_cache, GFP_KERNEL);
1219 if (zero_lv2_table == NULL) {
1220 pr_err("%s: Failed to allocate zero level2 page table\n",
1226 ret = bus_set_iommu(&platform_bus_type, &exynos_iommu_ops);
1228 pr_err("%s: Failed to register exynos-iommu driver.\n",
1235 kmem_cache_free(lv2table_kmem_cache, zero_lv2_table);
1237 platform_driver_unregister(&exynos_sysmmu_driver);
1239 kmem_cache_destroy(lv2table_kmem_cache);
1242 subsys_initcall(exynos_iommu_init);