1 // SPDX-License-Identifier: GPL-2.0-only
3 * IOMMU API for Rockchip
10 #include <linux/compiler.h>
11 #include <linux/delay.h>
12 #include <linux/device.h>
13 #include <linux/dma-iommu.h>
14 #include <linux/dma-mapping.h>
15 #include <linux/errno.h>
16 #include <linux/interrupt.h>
18 #include <linux/iommu.h>
19 #include <linux/iopoll.h>
20 #include <linux/list.h>
22 #include <linux/init.h>
24 #include <linux/of_iommu.h>
25 #include <linux/of_platform.h>
26 #include <linux/platform_device.h>
27 #include <linux/pm_runtime.h>
28 #include <linux/slab.h>
29 #include <linux/spinlock.h>
31 /** MMU register offsets */
32 #define RK_MMU_DTE_ADDR 0x00 /* Directory table address */
33 #define RK_MMU_STATUS 0x04
34 #define RK_MMU_COMMAND 0x08
35 #define RK_MMU_PAGE_FAULT_ADDR 0x0C /* IOVA of last page fault */
36 #define RK_MMU_ZAP_ONE_LINE 0x10 /* Shootdown one IOTLB entry */
37 #define RK_MMU_INT_RAWSTAT 0x14 /* IRQ status ignoring mask */
38 #define RK_MMU_INT_CLEAR 0x18 /* Acknowledge and re-arm irq */
39 #define RK_MMU_INT_MASK 0x1C /* IRQ enable */
40 #define RK_MMU_INT_STATUS 0x20 /* IRQ status after masking */
41 #define RK_MMU_AUTO_GATING 0x24
43 #define DTE_ADDR_DUMMY 0xCAFEBABE
45 #define RK_MMU_POLL_PERIOD_US 100
46 #define RK_MMU_FORCE_RESET_TIMEOUT_US 100000
47 #define RK_MMU_POLL_TIMEOUT_US 1000
49 /* RK_MMU_STATUS fields */
50 #define RK_MMU_STATUS_PAGING_ENABLED BIT(0)
51 #define RK_MMU_STATUS_PAGE_FAULT_ACTIVE BIT(1)
52 #define RK_MMU_STATUS_STALL_ACTIVE BIT(2)
53 #define RK_MMU_STATUS_IDLE BIT(3)
54 #define RK_MMU_STATUS_REPLAY_BUFFER_EMPTY BIT(4)
55 #define RK_MMU_STATUS_PAGE_FAULT_IS_WRITE BIT(5)
56 #define RK_MMU_STATUS_STALL_NOT_ACTIVE BIT(31)
58 /* RK_MMU_COMMAND command values */
59 #define RK_MMU_CMD_ENABLE_PAGING 0 /* Enable memory translation */
60 #define RK_MMU_CMD_DISABLE_PAGING 1 /* Disable memory translation */
61 #define RK_MMU_CMD_ENABLE_STALL 2 /* Stall paging to allow other cmds */
62 #define RK_MMU_CMD_DISABLE_STALL 3 /* Stop stall re-enables paging */
63 #define RK_MMU_CMD_ZAP_CACHE 4 /* Shoot down entire IOTLB */
64 #define RK_MMU_CMD_PAGE_FAULT_DONE 5 /* Clear page fault */
65 #define RK_MMU_CMD_FORCE_RESET 6 /* Reset all registers */
67 /* RK_MMU_INT_* register fields */
68 #define RK_MMU_IRQ_PAGE_FAULT 0x01 /* page fault */
69 #define RK_MMU_IRQ_BUS_ERROR 0x02 /* bus read error */
70 #define RK_MMU_IRQ_MASK (RK_MMU_IRQ_PAGE_FAULT | RK_MMU_IRQ_BUS_ERROR)
72 #define NUM_DT_ENTRIES 1024
73 #define NUM_PT_ENTRIES 1024
75 #define SPAGE_ORDER 12
76 #define SPAGE_SIZE (1 << SPAGE_ORDER)
79 * Support mapping any size that fits in one page table:
82 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
84 struct rk_iommu_domain {
85 struct list_head iommus;
86 u32 *dt; /* page directory table */
88 spinlock_t iommus_lock; /* lock for iommus list */
89 spinlock_t dt_lock; /* lock for modifying page directory table */
91 struct iommu_domain domain;
94 /* list of clocks required by IOMMU */
95 static const char * const rk_iommu_clocks[] = {
101 void __iomem **bases;
103 struct clk_bulk_data *clocks;
106 struct iommu_device iommu;
107 struct list_head node; /* entry in rk_iommu_domain.iommus */
108 struct iommu_domain *domain; /* domain to which iommu is attached */
109 struct iommu_group *group;
112 struct rk_iommudata {
113 struct device_link *link; /* runtime PM link from IOMMU to master */
114 struct rk_iommu *iommu;
117 static struct device *dma_dev;
119 static inline void rk_table_flush(struct rk_iommu_domain *dom, dma_addr_t dma,
122 size_t size = count * sizeof(u32); /* count of u32 entry */
124 dma_sync_single_for_device(dma_dev, dma, size, DMA_TO_DEVICE);
127 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
129 return container_of(dom, struct rk_iommu_domain, domain);
133 * The Rockchip rk3288 iommu uses a 2-level page table.
134 * The first level is the "Directory Table" (DT).
135 * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
137 * The second level is the 1024 Page Tables (PT).
138 * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
139 * a 4 KB page of physical memory.
141 * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
142 * Each iommu device has a MMU_DTE_ADDR register that contains the physical
143 * address of the start of the DT page.
145 * The structure of the page table is as follows:
148 * MMU_DTE_ADDR -> +-----+
154 * | | | PTE | -> +-----+
155 * +-----+ +-----+ | |
165 * Each DTE has a PT address and a valid bit:
166 * +---------------------+-----------+-+
167 * | PT address | Reserved |V|
168 * +---------------------+-----------+-+
169 * 31:12 - PT address (PTs always starts on a 4 KB boundary)
171 * 0 - 1 if PT @ PT address is valid
173 #define RK_DTE_PT_ADDRESS_MASK 0xfffff000
174 #define RK_DTE_PT_VALID BIT(0)
176 static inline phys_addr_t rk_dte_pt_address(u32 dte)
178 return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
181 static inline bool rk_dte_is_pt_valid(u32 dte)
183 return dte & RK_DTE_PT_VALID;
186 static inline u32 rk_mk_dte(dma_addr_t pt_dma)
188 return (pt_dma & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
192 * Each PTE has a Page address, some flags and a valid bit:
193 * +---------------------+---+-------+-+
194 * | Page address |Rsv| Flags |V|
195 * +---------------------+---+-------+-+
196 * 31:12 - Page address (Pages always start on a 4 KB boundary)
199 * 8 - Read allocate - allocate cache space on read misses
200 * 7 - Read cache - enable cache & prefetch of data
201 * 6 - Write buffer - enable delaying writes on their way to memory
202 * 5 - Write allocate - allocate cache space on write misses
203 * 4 - Write cache - different writes can be merged together
204 * 3 - Override cache attributes
205 * if 1, bits 4-8 control cache attributes
206 * if 0, the system bus defaults are used
209 * 0 - 1 if Page @ Page address is valid
211 #define RK_PTE_PAGE_ADDRESS_MASK 0xfffff000
212 #define RK_PTE_PAGE_FLAGS_MASK 0x000001fe
213 #define RK_PTE_PAGE_WRITABLE BIT(2)
214 #define RK_PTE_PAGE_READABLE BIT(1)
215 #define RK_PTE_PAGE_VALID BIT(0)
217 static inline phys_addr_t rk_pte_page_address(u32 pte)
219 return (phys_addr_t)pte & RK_PTE_PAGE_ADDRESS_MASK;
222 static inline bool rk_pte_is_page_valid(u32 pte)
224 return pte & RK_PTE_PAGE_VALID;
227 /* TODO: set cache flags per prot IOMMU_CACHE */
228 static u32 rk_mk_pte(phys_addr_t page, int prot)
231 flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
232 flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
233 page &= RK_PTE_PAGE_ADDRESS_MASK;
234 return page | flags | RK_PTE_PAGE_VALID;
237 static u32 rk_mk_pte_invalid(u32 pte)
239 return pte & ~RK_PTE_PAGE_VALID;
243 * rk3288 iova (IOMMU Virtual Address) format
245 * +-----------+-----------+-------------+
246 * | DTE index | PTE index | Page offset |
247 * +-----------+-----------+-------------+
248 * 31:22 - DTE index - index of DTE in DT
249 * 21:12 - PTE index - index of PTE in PT @ DTE.pt_address
250 * 11: 0 - Page offset - offset into page @ PTE.page_address
252 #define RK_IOVA_DTE_MASK 0xffc00000
253 #define RK_IOVA_DTE_SHIFT 22
254 #define RK_IOVA_PTE_MASK 0x003ff000
255 #define RK_IOVA_PTE_SHIFT 12
256 #define RK_IOVA_PAGE_MASK 0x00000fff
257 #define RK_IOVA_PAGE_SHIFT 0
259 static u32 rk_iova_dte_index(dma_addr_t iova)
261 return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
264 static u32 rk_iova_pte_index(dma_addr_t iova)
266 return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
269 static u32 rk_iova_page_offset(dma_addr_t iova)
271 return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
274 static u32 rk_iommu_read(void __iomem *base, u32 offset)
276 return readl(base + offset);
279 static void rk_iommu_write(void __iomem *base, u32 offset, u32 value)
281 writel(value, base + offset);
284 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
288 for (i = 0; i < iommu->num_mmu; i++)
289 writel(command, iommu->bases[i] + RK_MMU_COMMAND);
292 static void rk_iommu_base_command(void __iomem *base, u32 command)
294 writel(command, base + RK_MMU_COMMAND);
296 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova_start,
300 dma_addr_t iova_end = iova_start + size;
302 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
303 * entire iotlb rather than iterate over individual iovas.
305 for (i = 0; i < iommu->num_mmu; i++) {
308 for (iova = iova_start; iova < iova_end; iova += SPAGE_SIZE)
309 rk_iommu_write(iommu->bases[i], RK_MMU_ZAP_ONE_LINE, iova);
313 static bool rk_iommu_is_stall_active(struct rk_iommu *iommu)
318 for (i = 0; i < iommu->num_mmu; i++)
319 active &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
320 RK_MMU_STATUS_STALL_ACTIVE);
325 static bool rk_iommu_is_paging_enabled(struct rk_iommu *iommu)
330 for (i = 0; i < iommu->num_mmu; i++)
331 enable &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
332 RK_MMU_STATUS_PAGING_ENABLED);
337 static bool rk_iommu_is_reset_done(struct rk_iommu *iommu)
342 for (i = 0; i < iommu->num_mmu; i++)
343 done &= rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR) == 0;
348 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
353 if (rk_iommu_is_stall_active(iommu))
356 /* Stall can only be enabled if paging is enabled */
357 if (!rk_iommu_is_paging_enabled(iommu))
360 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
362 ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
363 val, RK_MMU_POLL_PERIOD_US,
364 RK_MMU_POLL_TIMEOUT_US);
366 for (i = 0; i < iommu->num_mmu; i++)
367 dev_err(iommu->dev, "Enable stall request timed out, status: %#08x\n",
368 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
373 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
378 if (!rk_iommu_is_stall_active(iommu))
381 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
383 ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
384 !val, RK_MMU_POLL_PERIOD_US,
385 RK_MMU_POLL_TIMEOUT_US);
387 for (i = 0; i < iommu->num_mmu; i++)
388 dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
389 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
394 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
399 if (rk_iommu_is_paging_enabled(iommu))
402 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
404 ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
405 val, RK_MMU_POLL_PERIOD_US,
406 RK_MMU_POLL_TIMEOUT_US);
408 for (i = 0; i < iommu->num_mmu; i++)
409 dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
410 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
415 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
420 if (!rk_iommu_is_paging_enabled(iommu))
423 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
425 ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
426 !val, RK_MMU_POLL_PERIOD_US,
427 RK_MMU_POLL_TIMEOUT_US);
429 for (i = 0; i < iommu->num_mmu; i++)
430 dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
431 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
436 static int rk_iommu_force_reset(struct rk_iommu *iommu)
442 if (iommu->reset_disabled)
446 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
447 * and verifying that upper 5 nybbles are read back.
449 for (i = 0; i < iommu->num_mmu; i++) {
450 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, DTE_ADDR_DUMMY);
452 dte_addr = rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR);
453 if (dte_addr != (DTE_ADDR_DUMMY & RK_DTE_PT_ADDRESS_MASK)) {
454 dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
459 rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
461 ret = readx_poll_timeout(rk_iommu_is_reset_done, iommu, val,
462 val, RK_MMU_FORCE_RESET_TIMEOUT_US,
463 RK_MMU_POLL_TIMEOUT_US);
465 dev_err(iommu->dev, "FORCE_RESET command timed out\n");
472 static void log_iova(struct rk_iommu *iommu, int index, dma_addr_t iova)
474 void __iomem *base = iommu->bases[index];
475 u32 dte_index, pte_index, page_offset;
477 phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
480 phys_addr_t pte_addr_phys = 0;
481 u32 *pte_addr = NULL;
483 phys_addr_t page_addr_phys = 0;
486 dte_index = rk_iova_dte_index(iova);
487 pte_index = rk_iova_pte_index(iova);
488 page_offset = rk_iova_page_offset(iova);
490 mmu_dte_addr = rk_iommu_read(base, RK_MMU_DTE_ADDR);
491 mmu_dte_addr_phys = (phys_addr_t)mmu_dte_addr;
493 dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
494 dte_addr = phys_to_virt(dte_addr_phys);
497 if (!rk_dte_is_pt_valid(dte))
500 pte_addr_phys = rk_dte_pt_address(dte) + (pte_index * 4);
501 pte_addr = phys_to_virt(pte_addr_phys);
504 if (!rk_pte_is_page_valid(pte))
507 page_addr_phys = rk_pte_page_address(pte) + page_offset;
508 page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
511 dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
512 &iova, dte_index, pte_index, page_offset);
513 dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
514 &mmu_dte_addr_phys, &dte_addr_phys, dte,
515 rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
516 rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
519 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
521 struct rk_iommu *iommu = dev_id;
525 irqreturn_t ret = IRQ_NONE;
528 err = pm_runtime_get_if_in_use(iommu->dev);
529 if (WARN_ON_ONCE(err <= 0))
532 if (WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks)))
535 for (i = 0; i < iommu->num_mmu; i++) {
536 int_status = rk_iommu_read(iommu->bases[i], RK_MMU_INT_STATUS);
541 iova = rk_iommu_read(iommu->bases[i], RK_MMU_PAGE_FAULT_ADDR);
543 if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
546 status = rk_iommu_read(iommu->bases[i], RK_MMU_STATUS);
547 flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
548 IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
550 dev_err(iommu->dev, "Page fault at %pad of type %s\n",
552 (flags == IOMMU_FAULT_WRITE) ? "write" : "read");
554 log_iova(iommu, i, iova);
557 * Report page fault to any installed handlers.
558 * Ignore the return code, though, since we always zap cache
559 * and clear the page fault anyway.
562 report_iommu_fault(iommu->domain, iommu->dev, iova,
565 dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
567 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
568 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_PAGE_FAULT_DONE);
571 if (int_status & RK_MMU_IRQ_BUS_ERROR)
572 dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
574 if (int_status & ~RK_MMU_IRQ_MASK)
575 dev_err(iommu->dev, "unexpected int_status: %#08x\n",
578 rk_iommu_write(iommu->bases[i], RK_MMU_INT_CLEAR, int_status);
581 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
584 pm_runtime_put(iommu->dev);
588 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
591 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
593 phys_addr_t pt_phys, phys = 0;
597 spin_lock_irqsave(&rk_domain->dt_lock, flags);
599 dte = rk_domain->dt[rk_iova_dte_index(iova)];
600 if (!rk_dte_is_pt_valid(dte))
603 pt_phys = rk_dte_pt_address(dte);
604 page_table = (u32 *)phys_to_virt(pt_phys);
605 pte = page_table[rk_iova_pte_index(iova)];
606 if (!rk_pte_is_page_valid(pte))
609 phys = rk_pte_page_address(pte) + rk_iova_page_offset(iova);
611 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
616 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
617 dma_addr_t iova, size_t size)
619 struct list_head *pos;
622 /* shootdown these iova from all iommus using this domain */
623 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
624 list_for_each(pos, &rk_domain->iommus) {
625 struct rk_iommu *iommu;
628 iommu = list_entry(pos, struct rk_iommu, node);
630 /* Only zap TLBs of IOMMUs that are powered on. */
631 ret = pm_runtime_get_if_in_use(iommu->dev);
632 if (WARN_ON_ONCE(ret < 0))
635 WARN_ON(clk_bulk_enable(iommu->num_clocks,
637 rk_iommu_zap_lines(iommu, iova, size);
638 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
639 pm_runtime_put(iommu->dev);
642 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
645 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
646 dma_addr_t iova, size_t size)
648 rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
649 if (size > SPAGE_SIZE)
650 rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
654 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
657 u32 *page_table, *dte_addr;
662 assert_spin_locked(&rk_domain->dt_lock);
664 dte_index = rk_iova_dte_index(iova);
665 dte_addr = &rk_domain->dt[dte_index];
667 if (rk_dte_is_pt_valid(dte))
670 page_table = (u32 *)get_zeroed_page(GFP_ATOMIC | GFP_DMA32);
672 return ERR_PTR(-ENOMEM);
674 pt_dma = dma_map_single(dma_dev, page_table, SPAGE_SIZE, DMA_TO_DEVICE);
675 if (dma_mapping_error(dma_dev, pt_dma)) {
676 dev_err(dma_dev, "DMA mapping error while allocating page table\n");
677 free_page((unsigned long)page_table);
678 return ERR_PTR(-ENOMEM);
681 dte = rk_mk_dte(pt_dma);
684 rk_table_flush(rk_domain, pt_dma, NUM_PT_ENTRIES);
685 rk_table_flush(rk_domain,
686 rk_domain->dt_dma + dte_index * sizeof(u32), 1);
688 pt_phys = rk_dte_pt_address(dte);
689 return (u32 *)phys_to_virt(pt_phys);
692 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
693 u32 *pte_addr, dma_addr_t pte_dma,
696 unsigned int pte_count;
697 unsigned int pte_total = size / SPAGE_SIZE;
699 assert_spin_locked(&rk_domain->dt_lock);
701 for (pte_count = 0; pte_count < pte_total; pte_count++) {
702 u32 pte = pte_addr[pte_count];
703 if (!rk_pte_is_page_valid(pte))
706 pte_addr[pte_count] = rk_mk_pte_invalid(pte);
709 rk_table_flush(rk_domain, pte_dma, pte_count);
711 return pte_count * SPAGE_SIZE;
714 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
715 dma_addr_t pte_dma, dma_addr_t iova,
716 phys_addr_t paddr, size_t size, int prot)
718 unsigned int pte_count;
719 unsigned int pte_total = size / SPAGE_SIZE;
720 phys_addr_t page_phys;
722 assert_spin_locked(&rk_domain->dt_lock);
724 for (pte_count = 0; pte_count < pte_total; pte_count++) {
725 u32 pte = pte_addr[pte_count];
727 if (rk_pte_is_page_valid(pte))
730 pte_addr[pte_count] = rk_mk_pte(paddr, prot);
735 rk_table_flush(rk_domain, pte_dma, pte_total);
738 * Zap the first and last iova to evict from iotlb any previously
739 * mapped cachelines holding stale values for its dte and pte.
740 * We only zap the first and last iova, since only they could have
741 * dte or pte shared with an existing mapping.
743 rk_iommu_zap_iova_first_last(rk_domain, iova, size);
747 /* Unmap the range of iovas that we just mapped */
748 rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma,
749 pte_count * SPAGE_SIZE);
751 iova += pte_count * SPAGE_SIZE;
752 page_phys = rk_pte_page_address(pte_addr[pte_count]);
753 pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
754 &iova, &page_phys, &paddr, prot);
759 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
760 phys_addr_t paddr, size_t size, int prot)
762 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
764 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
765 u32 *page_table, *pte_addr;
766 u32 dte_index, pte_index;
769 spin_lock_irqsave(&rk_domain->dt_lock, flags);
772 * pgsize_bitmap specifies iova sizes that fit in one page table
773 * (1024 4-KiB pages = 4 MiB).
774 * So, size will always be 4096 <= size <= 4194304.
775 * Since iommu_map() guarantees that both iova and size will be
776 * aligned, we will always only be mapping from a single dte here.
778 page_table = rk_dte_get_page_table(rk_domain, iova);
779 if (IS_ERR(page_table)) {
780 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
781 return PTR_ERR(page_table);
784 dte_index = rk_domain->dt[rk_iova_dte_index(iova)];
785 pte_index = rk_iova_pte_index(iova);
786 pte_addr = &page_table[pte_index];
787 pte_dma = rk_dte_pt_address(dte_index) + pte_index * sizeof(u32);
788 ret = rk_iommu_map_iova(rk_domain, pte_addr, pte_dma, iova,
791 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
796 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
797 size_t size, struct iommu_iotlb_gather *gather)
799 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
801 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
807 spin_lock_irqsave(&rk_domain->dt_lock, flags);
810 * pgsize_bitmap specifies iova sizes that fit in one page table
811 * (1024 4-KiB pages = 4 MiB).
812 * So, size will always be 4096 <= size <= 4194304.
813 * Since iommu_unmap() guarantees that both iova and size will be
814 * aligned, we will always only be unmapping from a single dte here.
816 dte = rk_domain->dt[rk_iova_dte_index(iova)];
817 /* Just return 0 if iova is unmapped */
818 if (!rk_dte_is_pt_valid(dte)) {
819 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
823 pt_phys = rk_dte_pt_address(dte);
824 pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
825 pte_dma = pt_phys + rk_iova_pte_index(iova) * sizeof(u32);
826 unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma, size);
828 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
830 /* Shootdown iotlb entries for iova range that was just unmapped */
831 rk_iommu_zap_iova(rk_domain, iova, unmap_size);
836 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
838 struct rk_iommudata *data = dev->archdata.iommu;
840 return data ? data->iommu : NULL;
843 /* Must be called with iommu powered on and attached */
844 static void rk_iommu_disable(struct rk_iommu *iommu)
848 /* Ignore error while disabling, just keep going */
849 WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks));
850 rk_iommu_enable_stall(iommu);
851 rk_iommu_disable_paging(iommu);
852 for (i = 0; i < iommu->num_mmu; i++) {
853 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, 0);
854 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, 0);
856 rk_iommu_disable_stall(iommu);
857 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
860 /* Must be called with iommu powered on and attached */
861 static int rk_iommu_enable(struct rk_iommu *iommu)
863 struct iommu_domain *domain = iommu->domain;
864 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
867 ret = clk_bulk_enable(iommu->num_clocks, iommu->clocks);
871 ret = rk_iommu_enable_stall(iommu);
873 goto out_disable_clocks;
875 ret = rk_iommu_force_reset(iommu);
877 goto out_disable_stall;
879 for (i = 0; i < iommu->num_mmu; i++) {
880 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
882 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
883 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
886 ret = rk_iommu_enable_paging(iommu);
889 rk_iommu_disable_stall(iommu);
891 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
895 static void rk_iommu_detach_device(struct iommu_domain *domain,
898 struct rk_iommu *iommu;
899 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
903 /* Allow 'virtual devices' (eg drm) to detach from domain */
904 iommu = rk_iommu_from_dev(dev);
908 dev_dbg(dev, "Detaching from iommu domain\n");
910 /* iommu already detached */
911 if (iommu->domain != domain)
914 iommu->domain = NULL;
916 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
917 list_del_init(&iommu->node);
918 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
920 ret = pm_runtime_get_if_in_use(iommu->dev);
921 WARN_ON_ONCE(ret < 0);
923 rk_iommu_disable(iommu);
924 pm_runtime_put(iommu->dev);
928 static int rk_iommu_attach_device(struct iommu_domain *domain,
931 struct rk_iommu *iommu;
932 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
937 * Allow 'virtual devices' (e.g., drm) to attach to domain.
938 * Such a device does not belong to an iommu group.
940 iommu = rk_iommu_from_dev(dev);
944 dev_dbg(dev, "Attaching to iommu domain\n");
946 /* iommu already attached */
947 if (iommu->domain == domain)
951 rk_iommu_detach_device(iommu->domain, dev);
953 iommu->domain = domain;
955 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
956 list_add_tail(&iommu->node, &rk_domain->iommus);
957 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
959 ret = pm_runtime_get_if_in_use(iommu->dev);
960 if (!ret || WARN_ON_ONCE(ret < 0))
963 ret = rk_iommu_enable(iommu);
965 rk_iommu_detach_device(iommu->domain, dev);
967 pm_runtime_put(iommu->dev);
972 static struct iommu_domain *rk_iommu_domain_alloc(unsigned type)
974 struct rk_iommu_domain *rk_domain;
976 if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
982 rk_domain = devm_kzalloc(dma_dev, sizeof(*rk_domain), GFP_KERNEL);
986 if (type == IOMMU_DOMAIN_DMA &&
987 iommu_get_dma_cookie(&rk_domain->domain))
991 * rk32xx iommus use a 2 level pagetable.
992 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
993 * Allocate one 4 KiB page for each table.
995 rk_domain->dt = (u32 *)get_zeroed_page(GFP_KERNEL | GFP_DMA32);
999 rk_domain->dt_dma = dma_map_single(dma_dev, rk_domain->dt,
1000 SPAGE_SIZE, DMA_TO_DEVICE);
1001 if (dma_mapping_error(dma_dev, rk_domain->dt_dma)) {
1002 dev_err(dma_dev, "DMA map error for DT\n");
1006 rk_table_flush(rk_domain, rk_domain->dt_dma, NUM_DT_ENTRIES);
1008 spin_lock_init(&rk_domain->iommus_lock);
1009 spin_lock_init(&rk_domain->dt_lock);
1010 INIT_LIST_HEAD(&rk_domain->iommus);
1012 rk_domain->domain.geometry.aperture_start = 0;
1013 rk_domain->domain.geometry.aperture_end = DMA_BIT_MASK(32);
1014 rk_domain->domain.geometry.force_aperture = true;
1016 return &rk_domain->domain;
1019 free_page((unsigned long)rk_domain->dt);
1021 if (type == IOMMU_DOMAIN_DMA)
1022 iommu_put_dma_cookie(&rk_domain->domain);
1027 static void rk_iommu_domain_free(struct iommu_domain *domain)
1029 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1032 WARN_ON(!list_empty(&rk_domain->iommus));
1034 for (i = 0; i < NUM_DT_ENTRIES; i++) {
1035 u32 dte = rk_domain->dt[i];
1036 if (rk_dte_is_pt_valid(dte)) {
1037 phys_addr_t pt_phys = rk_dte_pt_address(dte);
1038 u32 *page_table = phys_to_virt(pt_phys);
1039 dma_unmap_single(dma_dev, pt_phys,
1040 SPAGE_SIZE, DMA_TO_DEVICE);
1041 free_page((unsigned long)page_table);
1045 dma_unmap_single(dma_dev, rk_domain->dt_dma,
1046 SPAGE_SIZE, DMA_TO_DEVICE);
1047 free_page((unsigned long)rk_domain->dt);
1049 if (domain->type == IOMMU_DOMAIN_DMA)
1050 iommu_put_dma_cookie(&rk_domain->domain);
1053 static int rk_iommu_add_device(struct device *dev)
1055 struct iommu_group *group;
1056 struct rk_iommu *iommu;
1057 struct rk_iommudata *data;
1059 data = dev->archdata.iommu;
1063 iommu = rk_iommu_from_dev(dev);
1065 group = iommu_group_get_for_dev(dev);
1067 return PTR_ERR(group);
1068 iommu_group_put(group);
1070 iommu_device_link(&iommu->iommu, dev);
1071 data->link = device_link_add(dev, iommu->dev,
1072 DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME);
1077 static void rk_iommu_remove_device(struct device *dev)
1079 struct rk_iommu *iommu;
1080 struct rk_iommudata *data = dev->archdata.iommu;
1082 iommu = rk_iommu_from_dev(dev);
1084 device_link_del(data->link);
1085 iommu_device_unlink(&iommu->iommu, dev);
1086 iommu_group_remove_device(dev);
1089 static struct iommu_group *rk_iommu_device_group(struct device *dev)
1091 struct rk_iommu *iommu;
1093 iommu = rk_iommu_from_dev(dev);
1095 return iommu_group_ref_get(iommu->group);
1098 static int rk_iommu_of_xlate(struct device *dev,
1099 struct of_phandle_args *args)
1101 struct platform_device *iommu_dev;
1102 struct rk_iommudata *data;
1104 data = devm_kzalloc(dma_dev, sizeof(*data), GFP_KERNEL);
1108 iommu_dev = of_find_device_by_node(args->np);
1110 data->iommu = platform_get_drvdata(iommu_dev);
1111 dev->archdata.iommu = data;
1113 platform_device_put(iommu_dev);
1118 static const struct iommu_ops rk_iommu_ops = {
1119 .domain_alloc = rk_iommu_domain_alloc,
1120 .domain_free = rk_iommu_domain_free,
1121 .attach_dev = rk_iommu_attach_device,
1122 .detach_dev = rk_iommu_detach_device,
1123 .map = rk_iommu_map,
1124 .unmap = rk_iommu_unmap,
1125 .add_device = rk_iommu_add_device,
1126 .remove_device = rk_iommu_remove_device,
1127 .iova_to_phys = rk_iommu_iova_to_phys,
1128 .device_group = rk_iommu_device_group,
1129 .pgsize_bitmap = RK_IOMMU_PGSIZE_BITMAP,
1130 .of_xlate = rk_iommu_of_xlate,
1133 static int rk_iommu_probe(struct platform_device *pdev)
1135 struct device *dev = &pdev->dev;
1136 struct rk_iommu *iommu;
1137 struct resource *res;
1138 int num_res = pdev->num_resources;
1141 iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
1145 platform_set_drvdata(pdev, iommu);
1149 iommu->bases = devm_kcalloc(dev, num_res, sizeof(*iommu->bases),
1154 for (i = 0; i < num_res; i++) {
1155 res = platform_get_resource(pdev, IORESOURCE_MEM, i);
1158 iommu->bases[i] = devm_ioremap_resource(&pdev->dev, res);
1159 if (IS_ERR(iommu->bases[i]))
1163 if (iommu->num_mmu == 0)
1164 return PTR_ERR(iommu->bases[0]);
1166 iommu->reset_disabled = device_property_read_bool(dev,
1167 "rockchip,disable-mmu-reset");
1169 iommu->num_clocks = ARRAY_SIZE(rk_iommu_clocks);
1170 iommu->clocks = devm_kcalloc(iommu->dev, iommu->num_clocks,
1171 sizeof(*iommu->clocks), GFP_KERNEL);
1175 for (i = 0; i < iommu->num_clocks; ++i)
1176 iommu->clocks[i].id = rk_iommu_clocks[i];
1179 * iommu clocks should be present for all new devices and devicetrees
1180 * but there are older devicetrees without clocks out in the wild.
1181 * So clocks as optional for the time being.
1183 err = devm_clk_bulk_get(iommu->dev, iommu->num_clocks, iommu->clocks);
1185 iommu->num_clocks = 0;
1189 err = clk_bulk_prepare(iommu->num_clocks, iommu->clocks);
1193 iommu->group = iommu_group_alloc();
1194 if (IS_ERR(iommu->group)) {
1195 err = PTR_ERR(iommu->group);
1196 goto err_unprepare_clocks;
1199 err = iommu_device_sysfs_add(&iommu->iommu, dev, NULL, dev_name(dev));
1203 iommu_device_set_ops(&iommu->iommu, &rk_iommu_ops);
1204 iommu_device_set_fwnode(&iommu->iommu, &dev->of_node->fwnode);
1206 err = iommu_device_register(&iommu->iommu);
1208 goto err_remove_sysfs;
1211 * Use the first registered IOMMU device for domain to use with DMA
1212 * API, since a domain might not physically correspond to a single
1216 dma_dev = &pdev->dev;
1218 bus_set_iommu(&platform_bus_type, &rk_iommu_ops);
1220 pm_runtime_enable(dev);
1223 while ((irq = platform_get_irq(pdev, i++)) != -ENXIO) {
1227 err = devm_request_irq(iommu->dev, irq, rk_iommu_irq,
1228 IRQF_SHARED, dev_name(dev), iommu);
1230 pm_runtime_disable(dev);
1231 goto err_remove_sysfs;
1237 iommu_device_sysfs_remove(&iommu->iommu);
1239 iommu_group_put(iommu->group);
1240 err_unprepare_clocks:
1241 clk_bulk_unprepare(iommu->num_clocks, iommu->clocks);
1245 static void rk_iommu_shutdown(struct platform_device *pdev)
1247 struct rk_iommu *iommu = platform_get_drvdata(pdev);
1250 while ((irq = platform_get_irq(pdev, i++)) != -ENXIO)
1251 devm_free_irq(iommu->dev, irq, iommu);
1253 pm_runtime_force_suspend(&pdev->dev);
1256 static int __maybe_unused rk_iommu_suspend(struct device *dev)
1258 struct rk_iommu *iommu = dev_get_drvdata(dev);
1263 rk_iommu_disable(iommu);
1267 static int __maybe_unused rk_iommu_resume(struct device *dev)
1269 struct rk_iommu *iommu = dev_get_drvdata(dev);
1274 return rk_iommu_enable(iommu);
1277 static const struct dev_pm_ops rk_iommu_pm_ops = {
1278 SET_RUNTIME_PM_OPS(rk_iommu_suspend, rk_iommu_resume, NULL)
1279 SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
1280 pm_runtime_force_resume)
1283 static const struct of_device_id rk_iommu_dt_ids[] = {
1284 { .compatible = "rockchip,iommu" },
1288 static struct platform_driver rk_iommu_driver = {
1289 .probe = rk_iommu_probe,
1290 .shutdown = rk_iommu_shutdown,
1293 .of_match_table = rk_iommu_dt_ids,
1294 .pm = &rk_iommu_pm_ops,
1295 .suppress_bind_attrs = true,
1299 static int __init rk_iommu_init(void)
1301 return platform_driver_register(&rk_iommu_driver);
1303 subsys_initcall(rk_iommu_init);