Linux 4.16.11
[linux/fpc-iii.git] / drivers / iommu / rockchip-iommu.c
blob9d991c2d87674c1dc19d3dd1d64f4c287041f4db
1 /*
2 * This program is free software; you can redistribute it and/or modify
3 * it under the terms of the GNU General Public License version 2 as
4 * published by the Free Software Foundation.
5 */
7 #include <linux/compiler.h>
8 #include <linux/delay.h>
9 #include <linux/device.h>
10 #include <linux/dma-iommu.h>
11 #include <linux/dma-mapping.h>
12 #include <linux/errno.h>
13 #include <linux/interrupt.h>
14 #include <linux/io.h>
15 #include <linux/iommu.h>
16 #include <linux/jiffies.h>
17 #include <linux/list.h>
18 #include <linux/mm.h>
19 #include <linux/module.h>
20 #include <linux/of.h>
21 #include <linux/of_platform.h>
22 #include <linux/platform_device.h>
23 #include <linux/slab.h>
24 #include <linux/spinlock.h>
26 /** MMU register offsets */
27 #define RK_MMU_DTE_ADDR 0x00 /* Directory table address */
28 #define RK_MMU_STATUS 0x04
29 #define RK_MMU_COMMAND 0x08
30 #define RK_MMU_PAGE_FAULT_ADDR 0x0C /* IOVA of last page fault */
31 #define RK_MMU_ZAP_ONE_LINE 0x10 /* Shootdown one IOTLB entry */
32 #define RK_MMU_INT_RAWSTAT 0x14 /* IRQ status ignoring mask */
33 #define RK_MMU_INT_CLEAR 0x18 /* Acknowledge and re-arm irq */
34 #define RK_MMU_INT_MASK 0x1C /* IRQ enable */
35 #define RK_MMU_INT_STATUS 0x20 /* IRQ status after masking */
36 #define RK_MMU_AUTO_GATING 0x24
38 #define DTE_ADDR_DUMMY 0xCAFEBABE
39 #define FORCE_RESET_TIMEOUT 100 /* ms */
41 /* RK_MMU_STATUS fields */
42 #define RK_MMU_STATUS_PAGING_ENABLED BIT(0)
43 #define RK_MMU_STATUS_PAGE_FAULT_ACTIVE BIT(1)
44 #define RK_MMU_STATUS_STALL_ACTIVE BIT(2)
45 #define RK_MMU_STATUS_IDLE BIT(3)
46 #define RK_MMU_STATUS_REPLAY_BUFFER_EMPTY BIT(4)
47 #define RK_MMU_STATUS_PAGE_FAULT_IS_WRITE BIT(5)
48 #define RK_MMU_STATUS_STALL_NOT_ACTIVE BIT(31)
50 /* RK_MMU_COMMAND command values */
51 #define RK_MMU_CMD_ENABLE_PAGING 0 /* Enable memory translation */
52 #define RK_MMU_CMD_DISABLE_PAGING 1 /* Disable memory translation */
53 #define RK_MMU_CMD_ENABLE_STALL 2 /* Stall paging to allow other cmds */
54 #define RK_MMU_CMD_DISABLE_STALL 3 /* Stop stall re-enables paging */
55 #define RK_MMU_CMD_ZAP_CACHE 4 /* Shoot down entire IOTLB */
56 #define RK_MMU_CMD_PAGE_FAULT_DONE 5 /* Clear page fault */
57 #define RK_MMU_CMD_FORCE_RESET 6 /* Reset all registers */
59 /* RK_MMU_INT_* register fields */
60 #define RK_MMU_IRQ_PAGE_FAULT 0x01 /* page fault */
61 #define RK_MMU_IRQ_BUS_ERROR 0x02 /* bus read error */
62 #define RK_MMU_IRQ_MASK (RK_MMU_IRQ_PAGE_FAULT | RK_MMU_IRQ_BUS_ERROR)
64 #define NUM_DT_ENTRIES 1024
65 #define NUM_PT_ENTRIES 1024
67 #define SPAGE_ORDER 12
68 #define SPAGE_SIZE (1 << SPAGE_ORDER)
71 * Support mapping any size that fits in one page table:
72 * 4 KiB to 4 MiB
74 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
76 #define IOMMU_REG_POLL_COUNT_FAST 1000
78 struct rk_iommu_domain {
79 struct list_head iommus;
80 struct platform_device *pdev;
81 u32 *dt; /* page directory table */
82 dma_addr_t dt_dma;
83 spinlock_t iommus_lock; /* lock for iommus list */
84 spinlock_t dt_lock; /* lock for modifying page directory table */
86 struct iommu_domain domain;
89 struct rk_iommu {
90 struct device *dev;
91 void __iomem **bases;
92 int num_mmu;
93 int *irq;
94 int num_irq;
95 bool reset_disabled;
96 struct iommu_device iommu;
97 struct list_head node; /* entry in rk_iommu_domain.iommus */
98 struct iommu_domain *domain; /* domain to which iommu is attached */
101 static inline void rk_table_flush(struct rk_iommu_domain *dom, dma_addr_t dma,
102 unsigned int count)
104 size_t size = count * sizeof(u32); /* count of u32 entry */
106 dma_sync_single_for_device(&dom->pdev->dev, dma, size, DMA_TO_DEVICE);
109 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
111 return container_of(dom, struct rk_iommu_domain, domain);
115 * Inspired by _wait_for in intel_drv.h
116 * This is NOT safe for use in interrupt context.
118 * Note that it's important that we check the condition again after having
119 * timed out, since the timeout could be due to preemption or similar and
120 * we've never had a chance to check the condition before the timeout.
122 #define rk_wait_for(COND, MS) ({ \
123 unsigned long timeout__ = jiffies + msecs_to_jiffies(MS) + 1; \
124 int ret__ = 0; \
125 while (!(COND)) { \
126 if (time_after(jiffies, timeout__)) { \
127 ret__ = (COND) ? 0 : -ETIMEDOUT; \
128 break; \
130 usleep_range(50, 100); \
132 ret__; \
136 * The Rockchip rk3288 iommu uses a 2-level page table.
137 * The first level is the "Directory Table" (DT).
138 * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
139 * to a "Page Table".
140 * The second level is the 1024 Page Tables (PT).
141 * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
142 * a 4 KB page of physical memory.
144 * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
145 * Each iommu device has a MMU_DTE_ADDR register that contains the physical
146 * address of the start of the DT page.
148 * The structure of the page table is as follows:
150 * DT
151 * MMU_DTE_ADDR -> +-----+
152 * | |
153 * +-----+ PT
154 * | DTE | -> +-----+
155 * +-----+ | | Memory
156 * | | +-----+ Page
157 * | | | PTE | -> +-----+
158 * +-----+ +-----+ | |
159 * | | | |
160 * | | | |
161 * +-----+ | |
162 * | |
163 * | |
164 * +-----+
168 * Each DTE has a PT address and a valid bit:
169 * +---------------------+-----------+-+
170 * | PT address | Reserved |V|
171 * +---------------------+-----------+-+
172 * 31:12 - PT address (PTs always starts on a 4 KB boundary)
173 * 11: 1 - Reserved
174 * 0 - 1 if PT @ PT address is valid
176 #define RK_DTE_PT_ADDRESS_MASK 0xfffff000
177 #define RK_DTE_PT_VALID BIT(0)
179 static inline phys_addr_t rk_dte_pt_address(u32 dte)
181 return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
184 static inline bool rk_dte_is_pt_valid(u32 dte)
186 return dte & RK_DTE_PT_VALID;
189 static inline u32 rk_mk_dte(dma_addr_t pt_dma)
191 return (pt_dma & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
195 * Each PTE has a Page address, some flags and a valid bit:
196 * +---------------------+---+-------+-+
197 * | Page address |Rsv| Flags |V|
198 * +---------------------+---+-------+-+
199 * 31:12 - Page address (Pages always start on a 4 KB boundary)
200 * 11: 9 - Reserved
201 * 8: 1 - Flags
202 * 8 - Read allocate - allocate cache space on read misses
203 * 7 - Read cache - enable cache & prefetch of data
204 * 6 - Write buffer - enable delaying writes on their way to memory
205 * 5 - Write allocate - allocate cache space on write misses
206 * 4 - Write cache - different writes can be merged together
207 * 3 - Override cache attributes
208 * if 1, bits 4-8 control cache attributes
209 * if 0, the system bus defaults are used
210 * 2 - Writable
211 * 1 - Readable
212 * 0 - 1 if Page @ Page address is valid
214 #define RK_PTE_PAGE_ADDRESS_MASK 0xfffff000
215 #define RK_PTE_PAGE_FLAGS_MASK 0x000001fe
216 #define RK_PTE_PAGE_WRITABLE BIT(2)
217 #define RK_PTE_PAGE_READABLE BIT(1)
218 #define RK_PTE_PAGE_VALID BIT(0)
220 static inline phys_addr_t rk_pte_page_address(u32 pte)
222 return (phys_addr_t)pte & RK_PTE_PAGE_ADDRESS_MASK;
225 static inline bool rk_pte_is_page_valid(u32 pte)
227 return pte & RK_PTE_PAGE_VALID;
230 /* TODO: set cache flags per prot IOMMU_CACHE */
231 static u32 rk_mk_pte(phys_addr_t page, int prot)
233 u32 flags = 0;
234 flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
235 flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
236 page &= RK_PTE_PAGE_ADDRESS_MASK;
237 return page | flags | RK_PTE_PAGE_VALID;
240 static u32 rk_mk_pte_invalid(u32 pte)
242 return pte & ~RK_PTE_PAGE_VALID;
246 * rk3288 iova (IOMMU Virtual Address) format
247 * 31 22.21 12.11 0
248 * +-----------+-----------+-------------+
249 * | DTE index | PTE index | Page offset |
250 * +-----------+-----------+-------------+
251 * 31:22 - DTE index - index of DTE in DT
252 * 21:12 - PTE index - index of PTE in PT @ DTE.pt_address
253 * 11: 0 - Page offset - offset into page @ PTE.page_address
255 #define RK_IOVA_DTE_MASK 0xffc00000
256 #define RK_IOVA_DTE_SHIFT 22
257 #define RK_IOVA_PTE_MASK 0x003ff000
258 #define RK_IOVA_PTE_SHIFT 12
259 #define RK_IOVA_PAGE_MASK 0x00000fff
260 #define RK_IOVA_PAGE_SHIFT 0
262 static u32 rk_iova_dte_index(dma_addr_t iova)
264 return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
267 static u32 rk_iova_pte_index(dma_addr_t iova)
269 return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
272 static u32 rk_iova_page_offset(dma_addr_t iova)
274 return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
277 static u32 rk_iommu_read(void __iomem *base, u32 offset)
279 return readl(base + offset);
282 static void rk_iommu_write(void __iomem *base, u32 offset, u32 value)
284 writel(value, base + offset);
287 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
289 int i;
291 for (i = 0; i < iommu->num_mmu; i++)
292 writel(command, iommu->bases[i] + RK_MMU_COMMAND);
295 static void rk_iommu_base_command(void __iomem *base, u32 command)
297 writel(command, base + RK_MMU_COMMAND);
299 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova,
300 size_t size)
302 int i;
304 dma_addr_t iova_end = iova + size;
306 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
307 * entire iotlb rather than iterate over individual iovas.
309 for (i = 0; i < iommu->num_mmu; i++)
310 for (; iova < iova_end; iova += SPAGE_SIZE)
311 rk_iommu_write(iommu->bases[i], RK_MMU_ZAP_ONE_LINE, iova);
314 static bool rk_iommu_is_stall_active(struct rk_iommu *iommu)
316 bool active = true;
317 int i;
319 for (i = 0; i < iommu->num_mmu; i++)
320 active &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
321 RK_MMU_STATUS_STALL_ACTIVE);
323 return active;
326 static bool rk_iommu_is_paging_enabled(struct rk_iommu *iommu)
328 bool enable = true;
329 int i;
331 for (i = 0; i < iommu->num_mmu; i++)
332 enable &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
333 RK_MMU_STATUS_PAGING_ENABLED);
335 return enable;
338 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
340 int ret, i;
342 if (rk_iommu_is_stall_active(iommu))
343 return 0;
345 /* Stall can only be enabled if paging is enabled */
346 if (!rk_iommu_is_paging_enabled(iommu))
347 return 0;
349 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
351 ret = rk_wait_for(rk_iommu_is_stall_active(iommu), 1);
352 if (ret)
353 for (i = 0; i < iommu->num_mmu; i++)
354 dev_err(iommu->dev, "Enable stall request timed out, status: %#08x\n",
355 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
357 return ret;
360 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
362 int ret, i;
364 if (!rk_iommu_is_stall_active(iommu))
365 return 0;
367 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
369 ret = rk_wait_for(!rk_iommu_is_stall_active(iommu), 1);
370 if (ret)
371 for (i = 0; i < iommu->num_mmu; i++)
372 dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
373 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
375 return ret;
378 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
380 int ret, i;
382 if (rk_iommu_is_paging_enabled(iommu))
383 return 0;
385 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
387 ret = rk_wait_for(rk_iommu_is_paging_enabled(iommu), 1);
388 if (ret)
389 for (i = 0; i < iommu->num_mmu; i++)
390 dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
391 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
393 return ret;
396 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
398 int ret, i;
400 if (!rk_iommu_is_paging_enabled(iommu))
401 return 0;
403 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
405 ret = rk_wait_for(!rk_iommu_is_paging_enabled(iommu), 1);
406 if (ret)
407 for (i = 0; i < iommu->num_mmu; i++)
408 dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
409 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
411 return ret;
414 static int rk_iommu_force_reset(struct rk_iommu *iommu)
416 int ret, i;
417 u32 dte_addr;
419 if (iommu->reset_disabled)
420 return 0;
423 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
424 * and verifying that upper 5 nybbles are read back.
426 for (i = 0; i < iommu->num_mmu; i++) {
427 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, DTE_ADDR_DUMMY);
429 dte_addr = rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR);
430 if (dte_addr != (DTE_ADDR_DUMMY & RK_DTE_PT_ADDRESS_MASK)) {
431 dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
432 return -EFAULT;
436 rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
438 for (i = 0; i < iommu->num_mmu; i++) {
439 ret = rk_wait_for(rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR) == 0x00000000,
440 FORCE_RESET_TIMEOUT);
441 if (ret) {
442 dev_err(iommu->dev, "FORCE_RESET command timed out\n");
443 return ret;
447 return 0;
450 static void log_iova(struct rk_iommu *iommu, int index, dma_addr_t iova)
452 void __iomem *base = iommu->bases[index];
453 u32 dte_index, pte_index, page_offset;
454 u32 mmu_dte_addr;
455 phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
456 u32 *dte_addr;
457 u32 dte;
458 phys_addr_t pte_addr_phys = 0;
459 u32 *pte_addr = NULL;
460 u32 pte = 0;
461 phys_addr_t page_addr_phys = 0;
462 u32 page_flags = 0;
464 dte_index = rk_iova_dte_index(iova);
465 pte_index = rk_iova_pte_index(iova);
466 page_offset = rk_iova_page_offset(iova);
468 mmu_dte_addr = rk_iommu_read(base, RK_MMU_DTE_ADDR);
469 mmu_dte_addr_phys = (phys_addr_t)mmu_dte_addr;
471 dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
472 dte_addr = phys_to_virt(dte_addr_phys);
473 dte = *dte_addr;
475 if (!rk_dte_is_pt_valid(dte))
476 goto print_it;
478 pte_addr_phys = rk_dte_pt_address(dte) + (pte_index * 4);
479 pte_addr = phys_to_virt(pte_addr_phys);
480 pte = *pte_addr;
482 if (!rk_pte_is_page_valid(pte))
483 goto print_it;
485 page_addr_phys = rk_pte_page_address(pte) + page_offset;
486 page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
488 print_it:
489 dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
490 &iova, dte_index, pte_index, page_offset);
491 dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
492 &mmu_dte_addr_phys, &dte_addr_phys, dte,
493 rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
494 rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
497 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
499 struct rk_iommu *iommu = dev_id;
500 u32 status;
501 u32 int_status;
502 dma_addr_t iova;
503 irqreturn_t ret = IRQ_NONE;
504 int i;
506 for (i = 0; i < iommu->num_mmu; i++) {
507 int_status = rk_iommu_read(iommu->bases[i], RK_MMU_INT_STATUS);
508 if (int_status == 0)
509 continue;
511 ret = IRQ_HANDLED;
512 iova = rk_iommu_read(iommu->bases[i], RK_MMU_PAGE_FAULT_ADDR);
514 if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
515 int flags;
517 status = rk_iommu_read(iommu->bases[i], RK_MMU_STATUS);
518 flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
519 IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
521 dev_err(iommu->dev, "Page fault at %pad of type %s\n",
522 &iova,
523 (flags == IOMMU_FAULT_WRITE) ? "write" : "read");
525 log_iova(iommu, i, iova);
528 * Report page fault to any installed handlers.
529 * Ignore the return code, though, since we always zap cache
530 * and clear the page fault anyway.
532 if (iommu->domain)
533 report_iommu_fault(iommu->domain, iommu->dev, iova,
534 flags);
535 else
536 dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
538 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
539 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_PAGE_FAULT_DONE);
542 if (int_status & RK_MMU_IRQ_BUS_ERROR)
543 dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
545 if (int_status & ~RK_MMU_IRQ_MASK)
546 dev_err(iommu->dev, "unexpected int_status: %#08x\n",
547 int_status);
549 rk_iommu_write(iommu->bases[i], RK_MMU_INT_CLEAR, int_status);
552 return ret;
555 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
556 dma_addr_t iova)
558 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
559 unsigned long flags;
560 phys_addr_t pt_phys, phys = 0;
561 u32 dte, pte;
562 u32 *page_table;
564 spin_lock_irqsave(&rk_domain->dt_lock, flags);
566 dte = rk_domain->dt[rk_iova_dte_index(iova)];
567 if (!rk_dte_is_pt_valid(dte))
568 goto out;
570 pt_phys = rk_dte_pt_address(dte);
571 page_table = (u32 *)phys_to_virt(pt_phys);
572 pte = page_table[rk_iova_pte_index(iova)];
573 if (!rk_pte_is_page_valid(pte))
574 goto out;
576 phys = rk_pte_page_address(pte) + rk_iova_page_offset(iova);
577 out:
578 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
580 return phys;
583 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
584 dma_addr_t iova, size_t size)
586 struct list_head *pos;
587 unsigned long flags;
589 /* shootdown these iova from all iommus using this domain */
590 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
591 list_for_each(pos, &rk_domain->iommus) {
592 struct rk_iommu *iommu;
593 iommu = list_entry(pos, struct rk_iommu, node);
594 rk_iommu_zap_lines(iommu, iova, size);
596 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
599 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
600 dma_addr_t iova, size_t size)
602 rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
603 if (size > SPAGE_SIZE)
604 rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
605 SPAGE_SIZE);
608 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
609 dma_addr_t iova)
611 struct device *dev = &rk_domain->pdev->dev;
612 u32 *page_table, *dte_addr;
613 u32 dte_index, dte;
614 phys_addr_t pt_phys;
615 dma_addr_t pt_dma;
617 assert_spin_locked(&rk_domain->dt_lock);
619 dte_index = rk_iova_dte_index(iova);
620 dte_addr = &rk_domain->dt[dte_index];
621 dte = *dte_addr;
622 if (rk_dte_is_pt_valid(dte))
623 goto done;
625 page_table = (u32 *)get_zeroed_page(GFP_ATOMIC | GFP_DMA32);
626 if (!page_table)
627 return ERR_PTR(-ENOMEM);
629 pt_dma = dma_map_single(dev, page_table, SPAGE_SIZE, DMA_TO_DEVICE);
630 if (dma_mapping_error(dev, pt_dma)) {
631 dev_err(dev, "DMA mapping error while allocating page table\n");
632 free_page((unsigned long)page_table);
633 return ERR_PTR(-ENOMEM);
636 dte = rk_mk_dte(pt_dma);
637 *dte_addr = dte;
639 rk_table_flush(rk_domain, pt_dma, NUM_PT_ENTRIES);
640 rk_table_flush(rk_domain,
641 rk_domain->dt_dma + dte_index * sizeof(u32), 1);
642 done:
643 pt_phys = rk_dte_pt_address(dte);
644 return (u32 *)phys_to_virt(pt_phys);
647 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
648 u32 *pte_addr, dma_addr_t pte_dma,
649 size_t size)
651 unsigned int pte_count;
652 unsigned int pte_total = size / SPAGE_SIZE;
654 assert_spin_locked(&rk_domain->dt_lock);
656 for (pte_count = 0; pte_count < pte_total; pte_count++) {
657 u32 pte = pte_addr[pte_count];
658 if (!rk_pte_is_page_valid(pte))
659 break;
661 pte_addr[pte_count] = rk_mk_pte_invalid(pte);
664 rk_table_flush(rk_domain, pte_dma, pte_count);
666 return pte_count * SPAGE_SIZE;
669 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
670 dma_addr_t pte_dma, dma_addr_t iova,
671 phys_addr_t paddr, size_t size, int prot)
673 unsigned int pte_count;
674 unsigned int pte_total = size / SPAGE_SIZE;
675 phys_addr_t page_phys;
677 assert_spin_locked(&rk_domain->dt_lock);
679 for (pte_count = 0; pte_count < pte_total; pte_count++) {
680 u32 pte = pte_addr[pte_count];
682 if (rk_pte_is_page_valid(pte))
683 goto unwind;
685 pte_addr[pte_count] = rk_mk_pte(paddr, prot);
687 paddr += SPAGE_SIZE;
690 rk_table_flush(rk_domain, pte_dma, pte_total);
693 * Zap the first and last iova to evict from iotlb any previously
694 * mapped cachelines holding stale values for its dte and pte.
695 * We only zap the first and last iova, since only they could have
696 * dte or pte shared with an existing mapping.
698 rk_iommu_zap_iova_first_last(rk_domain, iova, size);
700 return 0;
701 unwind:
702 /* Unmap the range of iovas that we just mapped */
703 rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma,
704 pte_count * SPAGE_SIZE);
706 iova += pte_count * SPAGE_SIZE;
707 page_phys = rk_pte_page_address(pte_addr[pte_count]);
708 pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
709 &iova, &page_phys, &paddr, prot);
711 return -EADDRINUSE;
714 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
715 phys_addr_t paddr, size_t size, int prot)
717 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
718 unsigned long flags;
719 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
720 u32 *page_table, *pte_addr;
721 u32 dte_index, pte_index;
722 int ret;
724 spin_lock_irqsave(&rk_domain->dt_lock, flags);
727 * pgsize_bitmap specifies iova sizes that fit in one page table
728 * (1024 4-KiB pages = 4 MiB).
729 * So, size will always be 4096 <= size <= 4194304.
730 * Since iommu_map() guarantees that both iova and size will be
731 * aligned, we will always only be mapping from a single dte here.
733 page_table = rk_dte_get_page_table(rk_domain, iova);
734 if (IS_ERR(page_table)) {
735 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
736 return PTR_ERR(page_table);
739 dte_index = rk_domain->dt[rk_iova_dte_index(iova)];
740 pte_index = rk_iova_pte_index(iova);
741 pte_addr = &page_table[pte_index];
742 pte_dma = rk_dte_pt_address(dte_index) + pte_index * sizeof(u32);
743 ret = rk_iommu_map_iova(rk_domain, pte_addr, pte_dma, iova,
744 paddr, size, prot);
746 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
748 return ret;
751 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
752 size_t size)
754 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
755 unsigned long flags;
756 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
757 phys_addr_t pt_phys;
758 u32 dte;
759 u32 *pte_addr;
760 size_t unmap_size;
762 spin_lock_irqsave(&rk_domain->dt_lock, flags);
765 * pgsize_bitmap specifies iova sizes that fit in one page table
766 * (1024 4-KiB pages = 4 MiB).
767 * So, size will always be 4096 <= size <= 4194304.
768 * Since iommu_unmap() guarantees that both iova and size will be
769 * aligned, we will always only be unmapping from a single dte here.
771 dte = rk_domain->dt[rk_iova_dte_index(iova)];
772 /* Just return 0 if iova is unmapped */
773 if (!rk_dte_is_pt_valid(dte)) {
774 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
775 return 0;
778 pt_phys = rk_dte_pt_address(dte);
779 pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
780 pte_dma = pt_phys + rk_iova_pte_index(iova) * sizeof(u32);
781 unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma, size);
783 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
785 /* Shootdown iotlb entries for iova range that was just unmapped */
786 rk_iommu_zap_iova(rk_domain, iova, unmap_size);
788 return unmap_size;
791 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
793 struct iommu_group *group;
794 struct device *iommu_dev;
795 struct rk_iommu *rk_iommu;
797 group = iommu_group_get(dev);
798 if (!group)
799 return NULL;
800 iommu_dev = iommu_group_get_iommudata(group);
801 rk_iommu = dev_get_drvdata(iommu_dev);
802 iommu_group_put(group);
804 return rk_iommu;
807 static int rk_iommu_attach_device(struct iommu_domain *domain,
808 struct device *dev)
810 struct rk_iommu *iommu;
811 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
812 unsigned long flags;
813 int ret, i;
816 * Allow 'virtual devices' (e.g., drm) to attach to domain.
817 * Such a device does not belong to an iommu group.
819 iommu = rk_iommu_from_dev(dev);
820 if (!iommu)
821 return 0;
823 ret = rk_iommu_enable_stall(iommu);
824 if (ret)
825 return ret;
827 ret = rk_iommu_force_reset(iommu);
828 if (ret)
829 return ret;
831 iommu->domain = domain;
833 for (i = 0; i < iommu->num_irq; i++) {
834 ret = devm_request_irq(iommu->dev, iommu->irq[i], rk_iommu_irq,
835 IRQF_SHARED, dev_name(dev), iommu);
836 if (ret)
837 return ret;
840 for (i = 0; i < iommu->num_mmu; i++) {
841 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
842 rk_domain->dt_dma);
843 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
844 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
847 ret = rk_iommu_enable_paging(iommu);
848 if (ret)
849 return ret;
851 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
852 list_add_tail(&iommu->node, &rk_domain->iommus);
853 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
855 dev_dbg(dev, "Attached to iommu domain\n");
857 rk_iommu_disable_stall(iommu);
859 return 0;
862 static void rk_iommu_detach_device(struct iommu_domain *domain,
863 struct device *dev)
865 struct rk_iommu *iommu;
866 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
867 unsigned long flags;
868 int i;
870 /* Allow 'virtual devices' (eg drm) to detach from domain */
871 iommu = rk_iommu_from_dev(dev);
872 if (!iommu)
873 return;
875 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
876 list_del_init(&iommu->node);
877 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
879 /* Ignore error while disabling, just keep going */
880 rk_iommu_enable_stall(iommu);
881 rk_iommu_disable_paging(iommu);
882 for (i = 0; i < iommu->num_mmu; i++) {
883 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, 0);
884 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, 0);
886 rk_iommu_disable_stall(iommu);
888 for (i = 0; i < iommu->num_irq; i++)
889 devm_free_irq(iommu->dev, iommu->irq[i], iommu);
891 iommu->domain = NULL;
893 dev_dbg(dev, "Detached from iommu domain\n");
896 static struct iommu_domain *rk_iommu_domain_alloc(unsigned type)
898 struct rk_iommu_domain *rk_domain;
899 struct platform_device *pdev;
900 struct device *iommu_dev;
902 if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
903 return NULL;
905 /* Register a pdev per domain, so DMA API can base on this *dev
906 * even some virtual master doesn't have an iommu slave
908 pdev = platform_device_register_simple("rk_iommu_domain",
909 PLATFORM_DEVID_AUTO, NULL, 0);
910 if (IS_ERR(pdev))
911 return NULL;
913 rk_domain = devm_kzalloc(&pdev->dev, sizeof(*rk_domain), GFP_KERNEL);
914 if (!rk_domain)
915 goto err_unreg_pdev;
917 rk_domain->pdev = pdev;
919 if (type == IOMMU_DOMAIN_DMA &&
920 iommu_get_dma_cookie(&rk_domain->domain))
921 goto err_unreg_pdev;
924 * rk32xx iommus use a 2 level pagetable.
925 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
926 * Allocate one 4 KiB page for each table.
928 rk_domain->dt = (u32 *)get_zeroed_page(GFP_KERNEL | GFP_DMA32);
929 if (!rk_domain->dt)
930 goto err_put_cookie;
932 iommu_dev = &pdev->dev;
933 rk_domain->dt_dma = dma_map_single(iommu_dev, rk_domain->dt,
934 SPAGE_SIZE, DMA_TO_DEVICE);
935 if (dma_mapping_error(iommu_dev, rk_domain->dt_dma)) {
936 dev_err(iommu_dev, "DMA map error for DT\n");
937 goto err_free_dt;
940 rk_table_flush(rk_domain, rk_domain->dt_dma, NUM_DT_ENTRIES);
942 spin_lock_init(&rk_domain->iommus_lock);
943 spin_lock_init(&rk_domain->dt_lock);
944 INIT_LIST_HEAD(&rk_domain->iommus);
946 rk_domain->domain.geometry.aperture_start = 0;
947 rk_domain->domain.geometry.aperture_end = DMA_BIT_MASK(32);
948 rk_domain->domain.geometry.force_aperture = true;
950 return &rk_domain->domain;
952 err_free_dt:
953 free_page((unsigned long)rk_domain->dt);
954 err_put_cookie:
955 if (type == IOMMU_DOMAIN_DMA)
956 iommu_put_dma_cookie(&rk_domain->domain);
957 err_unreg_pdev:
958 platform_device_unregister(pdev);
960 return NULL;
963 static void rk_iommu_domain_free(struct iommu_domain *domain)
965 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
966 int i;
968 WARN_ON(!list_empty(&rk_domain->iommus));
970 for (i = 0; i < NUM_DT_ENTRIES; i++) {
971 u32 dte = rk_domain->dt[i];
972 if (rk_dte_is_pt_valid(dte)) {
973 phys_addr_t pt_phys = rk_dte_pt_address(dte);
974 u32 *page_table = phys_to_virt(pt_phys);
975 dma_unmap_single(&rk_domain->pdev->dev, pt_phys,
976 SPAGE_SIZE, DMA_TO_DEVICE);
977 free_page((unsigned long)page_table);
981 dma_unmap_single(&rk_domain->pdev->dev, rk_domain->dt_dma,
982 SPAGE_SIZE, DMA_TO_DEVICE);
983 free_page((unsigned long)rk_domain->dt);
985 if (domain->type == IOMMU_DOMAIN_DMA)
986 iommu_put_dma_cookie(&rk_domain->domain);
988 platform_device_unregister(rk_domain->pdev);
991 static bool rk_iommu_is_dev_iommu_master(struct device *dev)
993 struct device_node *np = dev->of_node;
994 int ret;
997 * An iommu master has an iommus property containing a list of phandles
998 * to iommu nodes, each with an #iommu-cells property with value 0.
1000 ret = of_count_phandle_with_args(np, "iommus", "#iommu-cells");
1001 return (ret > 0);
1004 static int rk_iommu_group_set_iommudata(struct iommu_group *group,
1005 struct device *dev)
1007 struct device_node *np = dev->of_node;
1008 struct platform_device *pd;
1009 int ret;
1010 struct of_phandle_args args;
1013 * An iommu master has an iommus property containing a list of phandles
1014 * to iommu nodes, each with an #iommu-cells property with value 0.
1016 ret = of_parse_phandle_with_args(np, "iommus", "#iommu-cells", 0,
1017 &args);
1018 if (ret) {
1019 dev_err(dev, "of_parse_phandle_with_args(%pOF) => %d\n",
1020 np, ret);
1021 return ret;
1023 if (args.args_count != 0) {
1024 dev_err(dev, "incorrect number of iommu params found for %pOF (found %d, expected 0)\n",
1025 args.np, args.args_count);
1026 return -EINVAL;
1029 pd = of_find_device_by_node(args.np);
1030 of_node_put(args.np);
1031 if (!pd) {
1032 dev_err(dev, "iommu %pOF not found\n", args.np);
1033 return -EPROBE_DEFER;
1036 /* TODO(djkurtz): handle multiple slave iommus for a single master */
1037 iommu_group_set_iommudata(group, &pd->dev, NULL);
1039 return 0;
1042 static int rk_iommu_add_device(struct device *dev)
1044 struct iommu_group *group;
1045 struct rk_iommu *iommu;
1046 int ret;
1048 if (!rk_iommu_is_dev_iommu_master(dev))
1049 return -ENODEV;
1051 group = iommu_group_get(dev);
1052 if (!group) {
1053 group = iommu_group_alloc();
1054 if (IS_ERR(group)) {
1055 dev_err(dev, "Failed to allocate IOMMU group\n");
1056 return PTR_ERR(group);
1060 ret = iommu_group_add_device(group, dev);
1061 if (ret)
1062 goto err_put_group;
1064 ret = rk_iommu_group_set_iommudata(group, dev);
1065 if (ret)
1066 goto err_remove_device;
1068 iommu = rk_iommu_from_dev(dev);
1069 if (iommu)
1070 iommu_device_link(&iommu->iommu, dev);
1072 iommu_group_put(group);
1074 return 0;
1076 err_remove_device:
1077 iommu_group_remove_device(dev);
1078 err_put_group:
1079 iommu_group_put(group);
1080 return ret;
1083 static void rk_iommu_remove_device(struct device *dev)
1085 struct rk_iommu *iommu;
1087 if (!rk_iommu_is_dev_iommu_master(dev))
1088 return;
1090 iommu = rk_iommu_from_dev(dev);
1091 if (iommu)
1092 iommu_device_unlink(&iommu->iommu, dev);
1094 iommu_group_remove_device(dev);
1097 static const struct iommu_ops rk_iommu_ops = {
1098 .domain_alloc = rk_iommu_domain_alloc,
1099 .domain_free = rk_iommu_domain_free,
1100 .attach_dev = rk_iommu_attach_device,
1101 .detach_dev = rk_iommu_detach_device,
1102 .map = rk_iommu_map,
1103 .unmap = rk_iommu_unmap,
1104 .map_sg = default_iommu_map_sg,
1105 .add_device = rk_iommu_add_device,
1106 .remove_device = rk_iommu_remove_device,
1107 .iova_to_phys = rk_iommu_iova_to_phys,
1108 .pgsize_bitmap = RK_IOMMU_PGSIZE_BITMAP,
1111 static int rk_iommu_domain_probe(struct platform_device *pdev)
1113 struct device *dev = &pdev->dev;
1115 dev->dma_parms = devm_kzalloc(dev, sizeof(*dev->dma_parms), GFP_KERNEL);
1116 if (!dev->dma_parms)
1117 return -ENOMEM;
1119 /* Set dma_ops for dev, otherwise it would be dummy_dma_ops */
1120 arch_setup_dma_ops(dev, 0, DMA_BIT_MASK(32), NULL, false);
1122 dma_set_max_seg_size(dev, DMA_BIT_MASK(32));
1123 dma_coerce_mask_and_coherent(dev, DMA_BIT_MASK(32));
1125 return 0;
1128 static struct platform_driver rk_iommu_domain_driver = {
1129 .probe = rk_iommu_domain_probe,
1130 .driver = {
1131 .name = "rk_iommu_domain",
1135 static int rk_iommu_probe(struct platform_device *pdev)
1137 struct device *dev = &pdev->dev;
1138 struct rk_iommu *iommu;
1139 struct resource *res;
1140 int num_res = pdev->num_resources;
1141 int err, i;
1143 iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
1144 if (!iommu)
1145 return -ENOMEM;
1147 platform_set_drvdata(pdev, iommu);
1148 iommu->dev = dev;
1149 iommu->num_mmu = 0;
1151 iommu->bases = devm_kzalloc(dev, sizeof(*iommu->bases) * num_res,
1152 GFP_KERNEL);
1153 if (!iommu->bases)
1154 return -ENOMEM;
1156 for (i = 0; i < num_res; i++) {
1157 res = platform_get_resource(pdev, IORESOURCE_MEM, i);
1158 if (!res)
1159 continue;
1160 iommu->bases[i] = devm_ioremap_resource(&pdev->dev, res);
1161 if (IS_ERR(iommu->bases[i]))
1162 continue;
1163 iommu->num_mmu++;
1165 if (iommu->num_mmu == 0)
1166 return PTR_ERR(iommu->bases[0]);
1168 iommu->num_irq = platform_irq_count(pdev);
1169 if (iommu->num_irq < 0)
1170 return iommu->num_irq;
1171 if (iommu->num_irq == 0)
1172 return -ENXIO;
1174 iommu->irq = devm_kcalloc(dev, iommu->num_irq, sizeof(*iommu->irq),
1175 GFP_KERNEL);
1176 if (!iommu->irq)
1177 return -ENOMEM;
1179 for (i = 0; i < iommu->num_irq; i++) {
1180 iommu->irq[i] = platform_get_irq(pdev, i);
1181 if (iommu->irq[i] < 0) {
1182 dev_err(dev, "Failed to get IRQ, %d\n", iommu->irq[i]);
1183 return -ENXIO;
1187 iommu->reset_disabled = device_property_read_bool(dev,
1188 "rockchip,disable-mmu-reset");
1190 err = iommu_device_sysfs_add(&iommu->iommu, dev, NULL, dev_name(dev));
1191 if (err)
1192 return err;
1194 iommu_device_set_ops(&iommu->iommu, &rk_iommu_ops);
1195 err = iommu_device_register(&iommu->iommu);
1197 return err;
1200 static int rk_iommu_remove(struct platform_device *pdev)
1202 struct rk_iommu *iommu = platform_get_drvdata(pdev);
1204 if (iommu) {
1205 iommu_device_sysfs_remove(&iommu->iommu);
1206 iommu_device_unregister(&iommu->iommu);
1209 return 0;
1212 static const struct of_device_id rk_iommu_dt_ids[] = {
1213 { .compatible = "rockchip,iommu" },
1214 { /* sentinel */ }
1216 MODULE_DEVICE_TABLE(of, rk_iommu_dt_ids);
1218 static struct platform_driver rk_iommu_driver = {
1219 .probe = rk_iommu_probe,
1220 .remove = rk_iommu_remove,
1221 .driver = {
1222 .name = "rk_iommu",
1223 .of_match_table = rk_iommu_dt_ids,
1227 static int __init rk_iommu_init(void)
1229 struct device_node *np;
1230 int ret;
1232 np = of_find_matching_node(NULL, rk_iommu_dt_ids);
1233 if (!np)
1234 return 0;
1236 of_node_put(np);
1238 ret = bus_set_iommu(&platform_bus_type, &rk_iommu_ops);
1239 if (ret)
1240 return ret;
1242 ret = platform_driver_register(&rk_iommu_domain_driver);
1243 if (ret)
1244 return ret;
1246 ret = platform_driver_register(&rk_iommu_driver);
1247 if (ret)
1248 platform_driver_unregister(&rk_iommu_domain_driver);
1249 return ret;
1251 static void __exit rk_iommu_exit(void)
1253 platform_driver_unregister(&rk_iommu_driver);
1254 platform_driver_unregister(&rk_iommu_domain_driver);
1257 subsys_initcall(rk_iommu_init);
1258 module_exit(rk_iommu_exit);
1260 MODULE_DESCRIPTION("IOMMU API for Rockchip");
1261 MODULE_AUTHOR("Simon Xue <xxm@rock-chips.com> and Daniel Kurtz <djkurtz@chromium.org>");
1262 MODULE_ALIAS("platform:rockchip-iommu");
1263 MODULE_LICENSE("GPL v2");