1 // SPDX-License-Identifier: GPL-2.0-only
3 * IOMMU API for Rockchip
5 * Module Authors: Simon Xue <xxm@rock-chips.com>
6 * Daniel Kurtz <djkurtz@chromium.org>
10 #include <linux/compiler.h>
11 #include <linux/delay.h>
12 #include <linux/device.h>
13 #include <linux/dma-mapping.h>
14 #include <linux/errno.h>
15 #include <linux/interrupt.h>
17 #include <linux/iommu.h>
18 #include <linux/iopoll.h>
19 #include <linux/list.h>
21 #include <linux/init.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/spinlock.h>
29 /** MMU register offsets */
30 #define RK_MMU_DTE_ADDR 0x00 /* Directory table address */
31 #define RK_MMU_STATUS 0x04
32 #define RK_MMU_COMMAND 0x08
33 #define RK_MMU_PAGE_FAULT_ADDR 0x0C /* IOVA of last page fault */
34 #define RK_MMU_ZAP_ONE_LINE 0x10 /* Shootdown one IOTLB entry */
35 #define RK_MMU_INT_RAWSTAT 0x14 /* IRQ status ignoring mask */
36 #define RK_MMU_INT_CLEAR 0x18 /* Acknowledge and re-arm irq */
37 #define RK_MMU_INT_MASK 0x1C /* IRQ enable */
38 #define RK_MMU_INT_STATUS 0x20 /* IRQ status after masking */
39 #define RK_MMU_AUTO_GATING 0x24
41 #define DTE_ADDR_DUMMY 0xCAFEBABE
43 #define RK_MMU_POLL_PERIOD_US 100
44 #define RK_MMU_FORCE_RESET_TIMEOUT_US 100000
45 #define RK_MMU_POLL_TIMEOUT_US 1000
47 /* RK_MMU_STATUS fields */
48 #define RK_MMU_STATUS_PAGING_ENABLED BIT(0)
49 #define RK_MMU_STATUS_PAGE_FAULT_ACTIVE BIT(1)
50 #define RK_MMU_STATUS_STALL_ACTIVE BIT(2)
51 #define RK_MMU_STATUS_IDLE BIT(3)
52 #define RK_MMU_STATUS_REPLAY_BUFFER_EMPTY BIT(4)
53 #define RK_MMU_STATUS_PAGE_FAULT_IS_WRITE BIT(5)
54 #define RK_MMU_STATUS_STALL_NOT_ACTIVE BIT(31)
56 /* RK_MMU_COMMAND command values */
57 #define RK_MMU_CMD_ENABLE_PAGING 0 /* Enable memory translation */
58 #define RK_MMU_CMD_DISABLE_PAGING 1 /* Disable memory translation */
59 #define RK_MMU_CMD_ENABLE_STALL 2 /* Stall paging to allow other cmds */
60 #define RK_MMU_CMD_DISABLE_STALL 3 /* Stop stall re-enables paging */
61 #define RK_MMU_CMD_ZAP_CACHE 4 /* Shoot down entire IOTLB */
62 #define RK_MMU_CMD_PAGE_FAULT_DONE 5 /* Clear page fault */
63 #define RK_MMU_CMD_FORCE_RESET 6 /* Reset all registers */
65 /* RK_MMU_INT_* register fields */
66 #define RK_MMU_IRQ_PAGE_FAULT 0x01 /* page fault */
67 #define RK_MMU_IRQ_BUS_ERROR 0x02 /* bus read error */
68 #define RK_MMU_IRQ_MASK (RK_MMU_IRQ_PAGE_FAULT | RK_MMU_IRQ_BUS_ERROR)
70 #define NUM_DT_ENTRIES 1024
71 #define NUM_PT_ENTRIES 1024
73 #define SPAGE_ORDER 12
74 #define SPAGE_SIZE (1 << SPAGE_ORDER)
77 * Support mapping any size that fits in one page table:
80 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
82 struct rk_iommu_domain {
83 struct list_head iommus;
84 u32 *dt; /* page directory table */
86 spinlock_t iommus_lock; /* lock for iommus list */
87 spinlock_t dt_lock; /* lock for modifying page directory table */
89 struct iommu_domain domain;
92 /* list of clocks required by IOMMU */
93 static const char * const rk_iommu_clocks[] = {
98 phys_addr_t (*pt_address)(u32 dte);
99 u32 (*mk_dtentries)(dma_addr_t pt_dma);
100 u32 (*mk_ptentries)(phys_addr_t page, int prot);
107 void __iomem **bases;
110 struct clk_bulk_data *clocks;
113 struct iommu_device iommu;
114 struct list_head node; /* entry in rk_iommu_domain.iommus */
115 struct iommu_domain *domain; /* domain to which iommu is attached */
116 struct iommu_group *group;
119 struct rk_iommudata {
120 struct device_link *link; /* runtime PM link from IOMMU to master */
121 struct rk_iommu *iommu;
124 static struct device *dma_dev;
125 static const struct rk_iommu_ops *rk_ops;
126 static struct iommu_domain rk_identity_domain;
128 static inline void rk_table_flush(struct rk_iommu_domain *dom, dma_addr_t dma,
131 size_t size = count * sizeof(u32); /* count of u32 entry */
133 dma_sync_single_for_device(dma_dev, dma, size, DMA_TO_DEVICE);
136 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
138 return container_of(dom, struct rk_iommu_domain, domain);
142 * The Rockchip rk3288 iommu uses a 2-level page table.
143 * The first level is the "Directory Table" (DT).
144 * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
146 * The second level is the 1024 Page Tables (PT).
147 * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
148 * a 4 KB page of physical memory.
150 * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
151 * Each iommu device has a MMU_DTE_ADDR register that contains the physical
152 * address of the start of the DT page.
154 * The structure of the page table is as follows:
157 * MMU_DTE_ADDR -> +-----+
163 * | | | PTE | -> +-----+
164 * +-----+ +-----+ | |
174 * Each DTE has a PT address and a valid bit:
175 * +---------------------+-----------+-+
176 * | PT address | Reserved |V|
177 * +---------------------+-----------+-+
178 * 31:12 - PT address (PTs always starts on a 4 KB boundary)
180 * 0 - 1 if PT @ PT address is valid
182 #define RK_DTE_PT_ADDRESS_MASK 0xfffff000
183 #define RK_DTE_PT_VALID BIT(0)
185 static inline phys_addr_t rk_dte_pt_address(u32 dte)
187 return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
192 * 31:12 - PT address bit 31:0
193 * 11: 8 - PT address bit 35:32
194 * 7: 4 - PT address bit 39:36
196 * 0 - 1 if PT @ PT address is valid
198 #define RK_DTE_PT_ADDRESS_MASK_V2 GENMASK_ULL(31, 4)
199 #define DTE_HI_MASK1 GENMASK(11, 8)
200 #define DTE_HI_MASK2 GENMASK(7, 4)
201 #define DTE_HI_SHIFT1 24 /* shift bit 8 to bit 32 */
202 #define DTE_HI_SHIFT2 32 /* shift bit 4 to bit 36 */
203 #define PAGE_DESC_HI_MASK1 GENMASK_ULL(35, 32)
204 #define PAGE_DESC_HI_MASK2 GENMASK_ULL(39, 36)
206 static inline phys_addr_t rk_dte_pt_address_v2(u32 dte)
210 dte_v2 = ((dte_v2 & DTE_HI_MASK2) << DTE_HI_SHIFT2) |
211 ((dte_v2 & DTE_HI_MASK1) << DTE_HI_SHIFT1) |
212 (dte_v2 & RK_DTE_PT_ADDRESS_MASK);
214 return (phys_addr_t)dte_v2;
217 static inline bool rk_dte_is_pt_valid(u32 dte)
219 return dte & RK_DTE_PT_VALID;
222 static inline u32 rk_mk_dte(dma_addr_t pt_dma)
224 return (pt_dma & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
227 static inline u32 rk_mk_dte_v2(dma_addr_t pt_dma)
229 pt_dma = (pt_dma & RK_DTE_PT_ADDRESS_MASK) |
230 ((pt_dma & PAGE_DESC_HI_MASK1) >> DTE_HI_SHIFT1) |
231 (pt_dma & PAGE_DESC_HI_MASK2) >> DTE_HI_SHIFT2;
233 return (pt_dma & RK_DTE_PT_ADDRESS_MASK_V2) | RK_DTE_PT_VALID;
237 * Each PTE has a Page address, some flags and a valid bit:
238 * +---------------------+---+-------+-+
239 * | Page address |Rsv| Flags |V|
240 * +---------------------+---+-------+-+
241 * 31:12 - Page address (Pages always start on a 4 KB boundary)
244 * 8 - Read allocate - allocate cache space on read misses
245 * 7 - Read cache - enable cache & prefetch of data
246 * 6 - Write buffer - enable delaying writes on their way to memory
247 * 5 - Write allocate - allocate cache space on write misses
248 * 4 - Write cache - different writes can be merged together
249 * 3 - Override cache attributes
250 * if 1, bits 4-8 control cache attributes
251 * if 0, the system bus defaults are used
254 * 0 - 1 if Page @ Page address is valid
256 #define RK_PTE_PAGE_ADDRESS_MASK 0xfffff000
257 #define RK_PTE_PAGE_FLAGS_MASK 0x000001fe
258 #define RK_PTE_PAGE_WRITABLE BIT(2)
259 #define RK_PTE_PAGE_READABLE BIT(1)
260 #define RK_PTE_PAGE_VALID BIT(0)
262 static inline bool rk_pte_is_page_valid(u32 pte)
264 return pte & RK_PTE_PAGE_VALID;
267 /* TODO: set cache flags per prot IOMMU_CACHE */
268 static u32 rk_mk_pte(phys_addr_t page, int prot)
271 flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
272 flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
273 page &= RK_PTE_PAGE_ADDRESS_MASK;
274 return page | flags | RK_PTE_PAGE_VALID;
279 * 31:12 - Page address bit 31:0
280 * 11: 8 - Page address bit 35:32
281 * 7: 4 - Page address bit 39:36
285 * 0 - 1 if Page @ Page address is valid
288 static u32 rk_mk_pte_v2(phys_addr_t page, int prot)
292 flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
293 flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
295 return rk_mk_dte_v2(page) | flags;
298 static u32 rk_mk_pte_invalid(u32 pte)
300 return pte & ~RK_PTE_PAGE_VALID;
304 * rk3288 iova (IOMMU Virtual Address) format
306 * +-----------+-----------+-------------+
307 * | DTE index | PTE index | Page offset |
308 * +-----------+-----------+-------------+
309 * 31:22 - DTE index - index of DTE in DT
310 * 21:12 - PTE index - index of PTE in PT @ DTE.pt_address
311 * 11: 0 - Page offset - offset into page @ PTE.page_address
313 #define RK_IOVA_DTE_MASK 0xffc00000
314 #define RK_IOVA_DTE_SHIFT 22
315 #define RK_IOVA_PTE_MASK 0x003ff000
316 #define RK_IOVA_PTE_SHIFT 12
317 #define RK_IOVA_PAGE_MASK 0x00000fff
318 #define RK_IOVA_PAGE_SHIFT 0
320 static u32 rk_iova_dte_index(dma_addr_t iova)
322 return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
325 static u32 rk_iova_pte_index(dma_addr_t iova)
327 return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
330 static u32 rk_iova_page_offset(dma_addr_t iova)
332 return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
335 static u32 rk_iommu_read(void __iomem *base, u32 offset)
337 return readl(base + offset);
340 static void rk_iommu_write(void __iomem *base, u32 offset, u32 value)
342 writel(value, base + offset);
345 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
349 for (i = 0; i < iommu->num_mmu; i++)
350 writel(command, iommu->bases[i] + RK_MMU_COMMAND);
353 static void rk_iommu_base_command(void __iomem *base, u32 command)
355 writel(command, base + RK_MMU_COMMAND);
357 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova_start,
361 dma_addr_t iova_end = iova_start + size;
363 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
364 * entire iotlb rather than iterate over individual iovas.
366 for (i = 0; i < iommu->num_mmu; i++) {
369 for (iova = iova_start; iova < iova_end; iova += SPAGE_SIZE)
370 rk_iommu_write(iommu->bases[i], RK_MMU_ZAP_ONE_LINE, iova);
374 static bool rk_iommu_is_stall_active(struct rk_iommu *iommu)
379 for (i = 0; i < iommu->num_mmu; i++)
380 active &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
381 RK_MMU_STATUS_STALL_ACTIVE);
386 static bool rk_iommu_is_paging_enabled(struct rk_iommu *iommu)
391 for (i = 0; i < iommu->num_mmu; i++)
392 enable &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
393 RK_MMU_STATUS_PAGING_ENABLED);
398 static bool rk_iommu_is_reset_done(struct rk_iommu *iommu)
403 for (i = 0; i < iommu->num_mmu; i++)
404 done &= rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR) == 0;
409 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
414 if (rk_iommu_is_stall_active(iommu))
417 /* Stall can only be enabled if paging is enabled */
418 if (!rk_iommu_is_paging_enabled(iommu))
421 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
423 ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
424 val, RK_MMU_POLL_PERIOD_US,
425 RK_MMU_POLL_TIMEOUT_US);
427 for (i = 0; i < iommu->num_mmu; i++)
428 dev_err(iommu->dev, "Enable stall request timed out, status: %#08x\n",
429 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
434 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
439 if (!rk_iommu_is_stall_active(iommu))
442 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
444 ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
445 !val, RK_MMU_POLL_PERIOD_US,
446 RK_MMU_POLL_TIMEOUT_US);
448 for (i = 0; i < iommu->num_mmu; i++)
449 dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
450 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
455 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
460 if (rk_iommu_is_paging_enabled(iommu))
463 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
465 ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
466 val, RK_MMU_POLL_PERIOD_US,
467 RK_MMU_POLL_TIMEOUT_US);
469 for (i = 0; i < iommu->num_mmu; i++)
470 dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
471 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
476 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
481 if (!rk_iommu_is_paging_enabled(iommu))
484 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
486 ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
487 !val, RK_MMU_POLL_PERIOD_US,
488 RK_MMU_POLL_TIMEOUT_US);
490 for (i = 0; i < iommu->num_mmu; i++)
491 dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
492 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
497 static int rk_iommu_force_reset(struct rk_iommu *iommu)
503 if (iommu->reset_disabled)
507 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
508 * and verifying that upper 5 (v1) or 7 (v2) nybbles are read back.
510 for (i = 0; i < iommu->num_mmu; i++) {
511 dte_addr = rk_ops->pt_address(DTE_ADDR_DUMMY);
512 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, dte_addr);
514 if (dte_addr != rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR)) {
515 dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
520 rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
522 ret = readx_poll_timeout(rk_iommu_is_reset_done, iommu, val,
523 val, RK_MMU_FORCE_RESET_TIMEOUT_US,
524 RK_MMU_POLL_TIMEOUT_US);
526 dev_err(iommu->dev, "FORCE_RESET command timed out\n");
533 static void log_iova(struct rk_iommu *iommu, int index, dma_addr_t iova)
535 void __iomem *base = iommu->bases[index];
536 u32 dte_index, pte_index, page_offset;
538 phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
541 phys_addr_t pte_addr_phys = 0;
542 u32 *pte_addr = NULL;
544 phys_addr_t page_addr_phys = 0;
547 dte_index = rk_iova_dte_index(iova);
548 pte_index = rk_iova_pte_index(iova);
549 page_offset = rk_iova_page_offset(iova);
551 mmu_dte_addr = rk_iommu_read(base, RK_MMU_DTE_ADDR);
552 mmu_dte_addr_phys = rk_ops->pt_address(mmu_dte_addr);
554 dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
555 dte_addr = phys_to_virt(dte_addr_phys);
558 if (!rk_dte_is_pt_valid(dte))
561 pte_addr_phys = rk_ops->pt_address(dte) + (pte_index * 4);
562 pte_addr = phys_to_virt(pte_addr_phys);
565 if (!rk_pte_is_page_valid(pte))
568 page_addr_phys = rk_ops->pt_address(pte) + page_offset;
569 page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
572 dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
573 &iova, dte_index, pte_index, page_offset);
574 dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
575 &mmu_dte_addr_phys, &dte_addr_phys, dte,
576 rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
577 rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
580 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
582 struct rk_iommu *iommu = dev_id;
586 irqreturn_t ret = IRQ_NONE;
589 err = pm_runtime_get_if_in_use(iommu->dev);
590 if (!err || WARN_ON_ONCE(err < 0))
593 if (WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks)))
596 for (i = 0; i < iommu->num_mmu; i++) {
597 int_status = rk_iommu_read(iommu->bases[i], RK_MMU_INT_STATUS);
602 iova = rk_iommu_read(iommu->bases[i], RK_MMU_PAGE_FAULT_ADDR);
604 if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
607 status = rk_iommu_read(iommu->bases[i], RK_MMU_STATUS);
608 flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
609 IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
611 dev_err(iommu->dev, "Page fault at %pad of type %s\n",
613 (flags == IOMMU_FAULT_WRITE) ? "write" : "read");
615 log_iova(iommu, i, iova);
618 * Report page fault to any installed handlers.
619 * Ignore the return code, though, since we always zap cache
620 * and clear the page fault anyway.
622 if (iommu->domain != &rk_identity_domain)
623 report_iommu_fault(iommu->domain, iommu->dev, iova,
626 dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
628 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
629 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_PAGE_FAULT_DONE);
632 if (int_status & RK_MMU_IRQ_BUS_ERROR)
633 dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
635 if (int_status & ~RK_MMU_IRQ_MASK)
636 dev_err(iommu->dev, "unexpected int_status: %#08x\n",
639 rk_iommu_write(iommu->bases[i], RK_MMU_INT_CLEAR, int_status);
642 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
645 pm_runtime_put(iommu->dev);
649 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
652 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
654 phys_addr_t pt_phys, phys = 0;
658 spin_lock_irqsave(&rk_domain->dt_lock, flags);
660 dte = rk_domain->dt[rk_iova_dte_index(iova)];
661 if (!rk_dte_is_pt_valid(dte))
664 pt_phys = rk_ops->pt_address(dte);
665 page_table = (u32 *)phys_to_virt(pt_phys);
666 pte = page_table[rk_iova_pte_index(iova)];
667 if (!rk_pte_is_page_valid(pte))
670 phys = rk_ops->pt_address(pte) + rk_iova_page_offset(iova);
672 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
677 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
678 dma_addr_t iova, size_t size)
680 struct list_head *pos;
683 /* shootdown these iova from all iommus using this domain */
684 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
685 list_for_each(pos, &rk_domain->iommus) {
686 struct rk_iommu *iommu;
689 iommu = list_entry(pos, struct rk_iommu, node);
691 /* Only zap TLBs of IOMMUs that are powered on. */
692 ret = pm_runtime_get_if_in_use(iommu->dev);
693 if (WARN_ON_ONCE(ret < 0))
696 WARN_ON(clk_bulk_enable(iommu->num_clocks,
698 rk_iommu_zap_lines(iommu, iova, size);
699 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
700 pm_runtime_put(iommu->dev);
703 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
706 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
707 dma_addr_t iova, size_t size)
709 rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
710 if (size > SPAGE_SIZE)
711 rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
715 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
718 u32 *page_table, *dte_addr;
723 assert_spin_locked(&rk_domain->dt_lock);
725 dte_index = rk_iova_dte_index(iova);
726 dte_addr = &rk_domain->dt[dte_index];
728 if (rk_dte_is_pt_valid(dte))
731 page_table = (u32 *)get_zeroed_page(GFP_ATOMIC | rk_ops->gfp_flags);
733 return ERR_PTR(-ENOMEM);
735 pt_dma = dma_map_single(dma_dev, page_table, SPAGE_SIZE, DMA_TO_DEVICE);
736 if (dma_mapping_error(dma_dev, pt_dma)) {
737 dev_err(dma_dev, "DMA mapping error while allocating page table\n");
738 free_page((unsigned long)page_table);
739 return ERR_PTR(-ENOMEM);
742 dte = rk_ops->mk_dtentries(pt_dma);
745 rk_table_flush(rk_domain,
746 rk_domain->dt_dma + dte_index * sizeof(u32), 1);
748 pt_phys = rk_ops->pt_address(dte);
749 return (u32 *)phys_to_virt(pt_phys);
752 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
753 u32 *pte_addr, dma_addr_t pte_dma,
756 unsigned int pte_count;
757 unsigned int pte_total = size / SPAGE_SIZE;
759 assert_spin_locked(&rk_domain->dt_lock);
761 for (pte_count = 0; pte_count < pte_total; pte_count++) {
762 u32 pte = pte_addr[pte_count];
763 if (!rk_pte_is_page_valid(pte))
766 pte_addr[pte_count] = rk_mk_pte_invalid(pte);
769 rk_table_flush(rk_domain, pte_dma, pte_count);
771 return pte_count * SPAGE_SIZE;
774 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
775 dma_addr_t pte_dma, dma_addr_t iova,
776 phys_addr_t paddr, size_t size, int prot)
778 unsigned int pte_count;
779 unsigned int pte_total = size / SPAGE_SIZE;
780 phys_addr_t page_phys;
782 assert_spin_locked(&rk_domain->dt_lock);
784 for (pte_count = 0; pte_count < pte_total; pte_count++) {
785 u32 pte = pte_addr[pte_count];
787 if (rk_pte_is_page_valid(pte))
790 pte_addr[pte_count] = rk_ops->mk_ptentries(paddr, prot);
795 rk_table_flush(rk_domain, pte_dma, pte_total);
798 * Zap the first and last iova to evict from iotlb any previously
799 * mapped cachelines holding stale values for its dte and pte.
800 * We only zap the first and last iova, since only they could have
801 * dte or pte shared with an existing mapping.
803 rk_iommu_zap_iova_first_last(rk_domain, iova, size);
807 /* Unmap the range of iovas that we just mapped */
808 rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma,
809 pte_count * SPAGE_SIZE);
811 iova += pte_count * SPAGE_SIZE;
812 page_phys = rk_ops->pt_address(pte_addr[pte_count]);
813 pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
814 &iova, &page_phys, &paddr, prot);
819 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
820 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
822 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
824 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
825 u32 *page_table, *pte_addr;
826 u32 dte_index, pte_index;
829 spin_lock_irqsave(&rk_domain->dt_lock, flags);
832 * pgsize_bitmap specifies iova sizes that fit in one page table
833 * (1024 4-KiB pages = 4 MiB).
834 * So, size will always be 4096 <= size <= 4194304.
835 * Since iommu_map() guarantees that both iova and size will be
836 * aligned, we will always only be mapping from a single dte here.
838 page_table = rk_dte_get_page_table(rk_domain, iova);
839 if (IS_ERR(page_table)) {
840 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
841 return PTR_ERR(page_table);
844 dte_index = rk_domain->dt[rk_iova_dte_index(iova)];
845 pte_index = rk_iova_pte_index(iova);
846 pte_addr = &page_table[pte_index];
848 pte_dma = rk_ops->pt_address(dte_index) + pte_index * sizeof(u32);
849 ret = rk_iommu_map_iova(rk_domain, pte_addr, pte_dma, iova,
852 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
857 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
858 size_t size, struct iommu_iotlb_gather *gather)
860 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
862 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
868 spin_lock_irqsave(&rk_domain->dt_lock, flags);
871 * pgsize_bitmap specifies iova sizes that fit in one page table
872 * (1024 4-KiB pages = 4 MiB).
873 * So, size will always be 4096 <= size <= 4194304.
874 * Since iommu_unmap() guarantees that both iova and size will be
875 * aligned, we will always only be unmapping from a single dte here.
877 dte = rk_domain->dt[rk_iova_dte_index(iova)];
878 /* Just return 0 if iova is unmapped */
879 if (!rk_dte_is_pt_valid(dte)) {
880 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
884 pt_phys = rk_ops->pt_address(dte);
885 pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
886 pte_dma = pt_phys + rk_iova_pte_index(iova) * sizeof(u32);
887 unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma, size);
889 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
891 /* Shootdown iotlb entries for iova range that was just unmapped */
892 rk_iommu_zap_iova(rk_domain, iova, unmap_size);
897 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
899 struct rk_iommudata *data = dev_iommu_priv_get(dev);
901 return data ? data->iommu : NULL;
904 /* Must be called with iommu powered on and attached */
905 static void rk_iommu_disable(struct rk_iommu *iommu)
909 /* Ignore error while disabling, just keep going */
910 WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks));
911 rk_iommu_enable_stall(iommu);
912 rk_iommu_disable_paging(iommu);
913 for (i = 0; i < iommu->num_mmu; i++) {
914 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, 0);
915 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, 0);
917 rk_iommu_disable_stall(iommu);
918 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
921 /* Must be called with iommu powered on and attached */
922 static int rk_iommu_enable(struct rk_iommu *iommu)
924 struct iommu_domain *domain = iommu->domain;
925 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
928 ret = clk_bulk_enable(iommu->num_clocks, iommu->clocks);
932 ret = rk_iommu_enable_stall(iommu);
934 goto out_disable_clocks;
936 ret = rk_iommu_force_reset(iommu);
938 goto out_disable_stall;
940 for (i = 0; i < iommu->num_mmu; i++) {
941 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
942 rk_ops->mk_dtentries(rk_domain->dt_dma));
943 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
944 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
947 ret = rk_iommu_enable_paging(iommu);
950 rk_iommu_disable_stall(iommu);
952 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
956 static int rk_iommu_identity_attach(struct iommu_domain *identity_domain,
959 struct rk_iommu *iommu;
960 struct rk_iommu_domain *rk_domain;
964 /* Allow 'virtual devices' (eg drm) to detach from domain */
965 iommu = rk_iommu_from_dev(dev);
969 rk_domain = to_rk_domain(iommu->domain);
971 dev_dbg(dev, "Detaching from iommu domain\n");
973 if (iommu->domain == identity_domain)
976 iommu->domain = identity_domain;
978 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
979 list_del_init(&iommu->node);
980 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
982 ret = pm_runtime_get_if_in_use(iommu->dev);
983 WARN_ON_ONCE(ret < 0);
985 rk_iommu_disable(iommu);
986 pm_runtime_put(iommu->dev);
992 static void rk_iommu_identity_free(struct iommu_domain *domain)
996 static struct iommu_domain_ops rk_identity_ops = {
997 .attach_dev = rk_iommu_identity_attach,
998 .free = rk_iommu_identity_free,
1001 static struct iommu_domain rk_identity_domain = {
1002 .type = IOMMU_DOMAIN_IDENTITY,
1003 .ops = &rk_identity_ops,
1007 static void rk_iommu_set_platform_dma(struct device *dev)
1009 WARN_ON(rk_iommu_identity_attach(&rk_identity_domain, dev));
1013 static int rk_iommu_attach_device(struct iommu_domain *domain,
1016 struct rk_iommu *iommu;
1017 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1018 unsigned long flags;
1022 * Allow 'virtual devices' (e.g., drm) to attach to domain.
1023 * Such a device does not belong to an iommu group.
1025 iommu = rk_iommu_from_dev(dev);
1029 dev_dbg(dev, "Attaching to iommu domain\n");
1031 /* iommu already attached */
1032 if (iommu->domain == domain)
1035 ret = rk_iommu_identity_attach(&rk_identity_domain, dev);
1039 iommu->domain = domain;
1041 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
1042 list_add_tail(&iommu->node, &rk_domain->iommus);
1043 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
1045 ret = pm_runtime_get_if_in_use(iommu->dev);
1046 if (!ret || WARN_ON_ONCE(ret < 0))
1049 ret = rk_iommu_enable(iommu);
1051 WARN_ON(rk_iommu_identity_attach(&rk_identity_domain, dev));
1053 pm_runtime_put(iommu->dev);
1058 static struct iommu_domain *rk_iommu_domain_alloc(unsigned type)
1060 struct rk_iommu_domain *rk_domain;
1062 if (type == IOMMU_DOMAIN_IDENTITY)
1063 return &rk_identity_domain;
1065 if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
1071 rk_domain = kzalloc(sizeof(*rk_domain), GFP_KERNEL);
1076 * rk32xx iommus use a 2 level pagetable.
1077 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
1078 * Allocate one 4 KiB page for each table.
1080 rk_domain->dt = (u32 *)get_zeroed_page(GFP_KERNEL | rk_ops->gfp_flags);
1082 goto err_free_domain;
1084 rk_domain->dt_dma = dma_map_single(dma_dev, rk_domain->dt,
1085 SPAGE_SIZE, DMA_TO_DEVICE);
1086 if (dma_mapping_error(dma_dev, rk_domain->dt_dma)) {
1087 dev_err(dma_dev, "DMA map error for DT\n");
1091 spin_lock_init(&rk_domain->iommus_lock);
1092 spin_lock_init(&rk_domain->dt_lock);
1093 INIT_LIST_HEAD(&rk_domain->iommus);
1095 rk_domain->domain.geometry.aperture_start = 0;
1096 rk_domain->domain.geometry.aperture_end = DMA_BIT_MASK(32);
1097 rk_domain->domain.geometry.force_aperture = true;
1099 return &rk_domain->domain;
1102 free_page((unsigned long)rk_domain->dt);
1109 static void rk_iommu_domain_free(struct iommu_domain *domain)
1111 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1114 WARN_ON(!list_empty(&rk_domain->iommus));
1116 for (i = 0; i < NUM_DT_ENTRIES; i++) {
1117 u32 dte = rk_domain->dt[i];
1118 if (rk_dte_is_pt_valid(dte)) {
1119 phys_addr_t pt_phys = rk_ops->pt_address(dte);
1120 u32 *page_table = phys_to_virt(pt_phys);
1121 dma_unmap_single(dma_dev, pt_phys,
1122 SPAGE_SIZE, DMA_TO_DEVICE);
1123 free_page((unsigned long)page_table);
1127 dma_unmap_single(dma_dev, rk_domain->dt_dma,
1128 SPAGE_SIZE, DMA_TO_DEVICE);
1129 free_page((unsigned long)rk_domain->dt);
1134 static struct iommu_device *rk_iommu_probe_device(struct device *dev)
1136 struct rk_iommudata *data;
1137 struct rk_iommu *iommu;
1139 data = dev_iommu_priv_get(dev);
1141 return ERR_PTR(-ENODEV);
1143 iommu = rk_iommu_from_dev(dev);
1145 data->link = device_link_add(dev, iommu->dev,
1146 DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME);
1148 return &iommu->iommu;
1151 static void rk_iommu_release_device(struct device *dev)
1153 struct rk_iommudata *data = dev_iommu_priv_get(dev);
1155 device_link_del(data->link);
1158 static struct iommu_group *rk_iommu_device_group(struct device *dev)
1160 struct rk_iommu *iommu;
1162 iommu = rk_iommu_from_dev(dev);
1164 return iommu_group_ref_get(iommu->group);
1167 static int rk_iommu_of_xlate(struct device *dev,
1168 struct of_phandle_args *args)
1170 struct platform_device *iommu_dev;
1171 struct rk_iommudata *data;
1173 data = devm_kzalloc(dma_dev, sizeof(*data), GFP_KERNEL);
1177 iommu_dev = of_find_device_by_node(args->np);
1179 data->iommu = platform_get_drvdata(iommu_dev);
1180 data->iommu->domain = &rk_identity_domain;
1181 dev_iommu_priv_set(dev, data);
1183 platform_device_put(iommu_dev);
1188 static const struct iommu_ops rk_iommu_ops = {
1189 .domain_alloc = rk_iommu_domain_alloc,
1190 .probe_device = rk_iommu_probe_device,
1191 .release_device = rk_iommu_release_device,
1192 .device_group = rk_iommu_device_group,
1194 .set_platform_dma_ops = rk_iommu_set_platform_dma,
1196 .pgsize_bitmap = RK_IOMMU_PGSIZE_BITMAP,
1197 .of_xlate = rk_iommu_of_xlate,
1198 .default_domain_ops = &(const struct iommu_domain_ops) {
1199 .attach_dev = rk_iommu_attach_device,
1200 .map = rk_iommu_map,
1201 .unmap = rk_iommu_unmap,
1202 .iova_to_phys = rk_iommu_iova_to_phys,
1203 .free = rk_iommu_domain_free,
1207 static int rk_iommu_probe(struct platform_device *pdev)
1209 struct device *dev = &pdev->dev;
1210 struct rk_iommu *iommu;
1211 struct resource *res;
1212 const struct rk_iommu_ops *ops;
1213 int num_res = pdev->num_resources;
1216 iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
1220 platform_set_drvdata(pdev, iommu);
1224 ops = of_device_get_match_data(dev);
1229 * That should not happen unless different versions of the
1230 * hardware block are embedded the same SoC
1232 if (WARN_ON(rk_ops != ops))
1235 iommu->bases = devm_kcalloc(dev, num_res, sizeof(*iommu->bases),
1240 for (i = 0; i < num_res; i++) {
1241 res = platform_get_resource(pdev, IORESOURCE_MEM, i);
1244 iommu->bases[i] = devm_ioremap_resource(&pdev->dev, res);
1245 if (IS_ERR(iommu->bases[i]))
1249 if (iommu->num_mmu == 0)
1250 return PTR_ERR(iommu->bases[0]);
1252 iommu->num_irq = platform_irq_count(pdev);
1253 if (iommu->num_irq < 0)
1254 return iommu->num_irq;
1256 iommu->reset_disabled = device_property_read_bool(dev,
1257 "rockchip,disable-mmu-reset");
1259 iommu->num_clocks = ARRAY_SIZE(rk_iommu_clocks);
1260 iommu->clocks = devm_kcalloc(iommu->dev, iommu->num_clocks,
1261 sizeof(*iommu->clocks), GFP_KERNEL);
1265 for (i = 0; i < iommu->num_clocks; ++i)
1266 iommu->clocks[i].id = rk_iommu_clocks[i];
1269 * iommu clocks should be present for all new devices and devicetrees
1270 * but there are older devicetrees without clocks out in the wild.
1271 * So clocks as optional for the time being.
1273 err = devm_clk_bulk_get(iommu->dev, iommu->num_clocks, iommu->clocks);
1275 iommu->num_clocks = 0;
1279 err = clk_bulk_prepare(iommu->num_clocks, iommu->clocks);
1283 iommu->group = iommu_group_alloc();
1284 if (IS_ERR(iommu->group)) {
1285 err = PTR_ERR(iommu->group);
1286 goto err_unprepare_clocks;
1289 err = iommu_device_sysfs_add(&iommu->iommu, dev, NULL, dev_name(dev));
1293 err = iommu_device_register(&iommu->iommu, &rk_iommu_ops, dev);
1295 goto err_remove_sysfs;
1298 * Use the first registered IOMMU device for domain to use with DMA
1299 * API, since a domain might not physically correspond to a single
1303 dma_dev = &pdev->dev;
1305 pm_runtime_enable(dev);
1307 for (i = 0; i < iommu->num_irq; i++) {
1308 int irq = platform_get_irq(pdev, i);
1312 goto err_pm_disable;
1315 err = devm_request_irq(iommu->dev, irq, rk_iommu_irq,
1316 IRQF_SHARED, dev_name(dev), iommu);
1318 goto err_pm_disable;
1321 dma_set_mask_and_coherent(dev, rk_ops->dma_bit_mask);
1325 pm_runtime_disable(dev);
1327 iommu_device_sysfs_remove(&iommu->iommu);
1329 iommu_group_put(iommu->group);
1330 err_unprepare_clocks:
1331 clk_bulk_unprepare(iommu->num_clocks, iommu->clocks);
1335 static void rk_iommu_shutdown(struct platform_device *pdev)
1337 struct rk_iommu *iommu = platform_get_drvdata(pdev);
1340 for (i = 0; i < iommu->num_irq; i++) {
1341 int irq = platform_get_irq(pdev, i);
1343 devm_free_irq(iommu->dev, irq, iommu);
1346 pm_runtime_force_suspend(&pdev->dev);
1349 static int __maybe_unused rk_iommu_suspend(struct device *dev)
1351 struct rk_iommu *iommu = dev_get_drvdata(dev);
1353 if (iommu->domain == &rk_identity_domain)
1356 rk_iommu_disable(iommu);
1360 static int __maybe_unused rk_iommu_resume(struct device *dev)
1362 struct rk_iommu *iommu = dev_get_drvdata(dev);
1364 if (iommu->domain == &rk_identity_domain)
1367 return rk_iommu_enable(iommu);
1370 static const struct dev_pm_ops rk_iommu_pm_ops = {
1371 SET_RUNTIME_PM_OPS(rk_iommu_suspend, rk_iommu_resume, NULL)
1372 SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
1373 pm_runtime_force_resume)
1376 static struct rk_iommu_ops iommu_data_ops_v1 = {
1377 .pt_address = &rk_dte_pt_address,
1378 .mk_dtentries = &rk_mk_dte,
1379 .mk_ptentries = &rk_mk_pte,
1380 .dma_bit_mask = DMA_BIT_MASK(32),
1381 .gfp_flags = GFP_DMA32,
1384 static struct rk_iommu_ops iommu_data_ops_v2 = {
1385 .pt_address = &rk_dte_pt_address_v2,
1386 .mk_dtentries = &rk_mk_dte_v2,
1387 .mk_ptentries = &rk_mk_pte_v2,
1388 .dma_bit_mask = DMA_BIT_MASK(40),
1392 static const struct of_device_id rk_iommu_dt_ids[] = {
1393 { .compatible = "rockchip,iommu",
1394 .data = &iommu_data_ops_v1,
1396 { .compatible = "rockchip,rk3568-iommu",
1397 .data = &iommu_data_ops_v2,
1402 static struct platform_driver rk_iommu_driver = {
1403 .probe = rk_iommu_probe,
1404 .shutdown = rk_iommu_shutdown,
1407 .of_match_table = rk_iommu_dt_ids,
1408 .pm = &rk_iommu_pm_ops,
1409 .suppress_bind_attrs = true,
1412 builtin_platform_driver(rk_iommu_driver);