Linux 4.19.133
[linux/fpc-iii.git] / drivers / iommu / rockchip-iommu.c
blob140b287e886c843289e90ba211ae115148d7b49f
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/clk.h>
8 #include <linux/compiler.h>
9 #include <linux/delay.h>
10 #include <linux/device.h>
11 #include <linux/dma-iommu.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/errno.h>
14 #include <linux/interrupt.h>
15 #include <linux/io.h>
16 #include <linux/iommu.h>
17 #include <linux/iopoll.h>
18 #include <linux/list.h>
19 #include <linux/mm.h>
20 #include <linux/module.h>
21 #include <linux/of.h>
22 #include <linux/of_iommu.h>
23 #include <linux/of_platform.h>
24 #include <linux/platform_device.h>
25 #include <linux/pm_runtime.h>
26 #include <linux/slab.h>
27 #include <linux/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:
78 * 4 KiB to 4 MiB
80 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
82 struct rk_iommu_domain {
83 struct list_head iommus;
84 u32 *dt; /* page directory table */
85 dma_addr_t dt_dma;
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[] = {
94 "aclk", "iface",
97 struct rk_iommu {
98 struct device *dev;
99 void __iomem **bases;
100 int num_mmu;
101 struct clk_bulk_data *clocks;
102 int num_clocks;
103 bool reset_disabled;
104 struct iommu_device iommu;
105 struct list_head node; /* entry in rk_iommu_domain.iommus */
106 struct iommu_domain *domain; /* domain to which iommu is attached */
107 struct iommu_group *group;
110 struct rk_iommudata {
111 struct device_link *link; /* runtime PM link from IOMMU to master */
112 struct rk_iommu *iommu;
115 static struct device *dma_dev;
117 static inline void rk_table_flush(struct rk_iommu_domain *dom, dma_addr_t dma,
118 unsigned int count)
120 size_t size = count * sizeof(u32); /* count of u32 entry */
122 dma_sync_single_for_device(dma_dev, dma, size, DMA_TO_DEVICE);
125 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
127 return container_of(dom, struct rk_iommu_domain, domain);
131 * The Rockchip rk3288 iommu uses a 2-level page table.
132 * The first level is the "Directory Table" (DT).
133 * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
134 * to a "Page Table".
135 * The second level is the 1024 Page Tables (PT).
136 * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
137 * a 4 KB page of physical memory.
139 * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
140 * Each iommu device has a MMU_DTE_ADDR register that contains the physical
141 * address of the start of the DT page.
143 * The structure of the page table is as follows:
145 * DT
146 * MMU_DTE_ADDR -> +-----+
147 * | |
148 * +-----+ PT
149 * | DTE | -> +-----+
150 * +-----+ | | Memory
151 * | | +-----+ Page
152 * | | | PTE | -> +-----+
153 * +-----+ +-----+ | |
154 * | | | |
155 * | | | |
156 * +-----+ | |
157 * | |
158 * | |
159 * +-----+
163 * Each DTE has a PT address and a valid bit:
164 * +---------------------+-----------+-+
165 * | PT address | Reserved |V|
166 * +---------------------+-----------+-+
167 * 31:12 - PT address (PTs always starts on a 4 KB boundary)
168 * 11: 1 - Reserved
169 * 0 - 1 if PT @ PT address is valid
171 #define RK_DTE_PT_ADDRESS_MASK 0xfffff000
172 #define RK_DTE_PT_VALID BIT(0)
174 static inline phys_addr_t rk_dte_pt_address(u32 dte)
176 return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
179 static inline bool rk_dte_is_pt_valid(u32 dte)
181 return dte & RK_DTE_PT_VALID;
184 static inline u32 rk_mk_dte(dma_addr_t pt_dma)
186 return (pt_dma & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
190 * Each PTE has a Page address, some flags and a valid bit:
191 * +---------------------+---+-------+-+
192 * | Page address |Rsv| Flags |V|
193 * +---------------------+---+-------+-+
194 * 31:12 - Page address (Pages always start on a 4 KB boundary)
195 * 11: 9 - Reserved
196 * 8: 1 - Flags
197 * 8 - Read allocate - allocate cache space on read misses
198 * 7 - Read cache - enable cache & prefetch of data
199 * 6 - Write buffer - enable delaying writes on their way to memory
200 * 5 - Write allocate - allocate cache space on write misses
201 * 4 - Write cache - different writes can be merged together
202 * 3 - Override cache attributes
203 * if 1, bits 4-8 control cache attributes
204 * if 0, the system bus defaults are used
205 * 2 - Writable
206 * 1 - Readable
207 * 0 - 1 if Page @ Page address is valid
209 #define RK_PTE_PAGE_ADDRESS_MASK 0xfffff000
210 #define RK_PTE_PAGE_FLAGS_MASK 0x000001fe
211 #define RK_PTE_PAGE_WRITABLE BIT(2)
212 #define RK_PTE_PAGE_READABLE BIT(1)
213 #define RK_PTE_PAGE_VALID BIT(0)
215 static inline phys_addr_t rk_pte_page_address(u32 pte)
217 return (phys_addr_t)pte & RK_PTE_PAGE_ADDRESS_MASK;
220 static inline bool rk_pte_is_page_valid(u32 pte)
222 return pte & RK_PTE_PAGE_VALID;
225 /* TODO: set cache flags per prot IOMMU_CACHE */
226 static u32 rk_mk_pte(phys_addr_t page, int prot)
228 u32 flags = 0;
229 flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
230 flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
231 page &= RK_PTE_PAGE_ADDRESS_MASK;
232 return page | flags | RK_PTE_PAGE_VALID;
235 static u32 rk_mk_pte_invalid(u32 pte)
237 return pte & ~RK_PTE_PAGE_VALID;
241 * rk3288 iova (IOMMU Virtual Address) format
242 * 31 22.21 12.11 0
243 * +-----------+-----------+-------------+
244 * | DTE index | PTE index | Page offset |
245 * +-----------+-----------+-------------+
246 * 31:22 - DTE index - index of DTE in DT
247 * 21:12 - PTE index - index of PTE in PT @ DTE.pt_address
248 * 11: 0 - Page offset - offset into page @ PTE.page_address
250 #define RK_IOVA_DTE_MASK 0xffc00000
251 #define RK_IOVA_DTE_SHIFT 22
252 #define RK_IOVA_PTE_MASK 0x003ff000
253 #define RK_IOVA_PTE_SHIFT 12
254 #define RK_IOVA_PAGE_MASK 0x00000fff
255 #define RK_IOVA_PAGE_SHIFT 0
257 static u32 rk_iova_dte_index(dma_addr_t iova)
259 return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
262 static u32 rk_iova_pte_index(dma_addr_t iova)
264 return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
267 static u32 rk_iova_page_offset(dma_addr_t iova)
269 return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
272 static u32 rk_iommu_read(void __iomem *base, u32 offset)
274 return readl(base + offset);
277 static void rk_iommu_write(void __iomem *base, u32 offset, u32 value)
279 writel(value, base + offset);
282 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
284 int i;
286 for (i = 0; i < iommu->num_mmu; i++)
287 writel(command, iommu->bases[i] + RK_MMU_COMMAND);
290 static void rk_iommu_base_command(void __iomem *base, u32 command)
292 writel(command, base + RK_MMU_COMMAND);
294 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova_start,
295 size_t size)
297 int i;
298 dma_addr_t iova_end = iova_start + size;
300 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
301 * entire iotlb rather than iterate over individual iovas.
303 for (i = 0; i < iommu->num_mmu; i++) {
304 dma_addr_t iova;
306 for (iova = iova_start; iova < iova_end; iova += SPAGE_SIZE)
307 rk_iommu_write(iommu->bases[i], RK_MMU_ZAP_ONE_LINE, iova);
311 static bool rk_iommu_is_stall_active(struct rk_iommu *iommu)
313 bool active = true;
314 int i;
316 for (i = 0; i < iommu->num_mmu; i++)
317 active &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
318 RK_MMU_STATUS_STALL_ACTIVE);
320 return active;
323 static bool rk_iommu_is_paging_enabled(struct rk_iommu *iommu)
325 bool enable = true;
326 int i;
328 for (i = 0; i < iommu->num_mmu; i++)
329 enable &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
330 RK_MMU_STATUS_PAGING_ENABLED);
332 return enable;
335 static bool rk_iommu_is_reset_done(struct rk_iommu *iommu)
337 bool done = true;
338 int i;
340 for (i = 0; i < iommu->num_mmu; i++)
341 done &= rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR) == 0;
343 return done;
346 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
348 int ret, i;
349 bool val;
351 if (rk_iommu_is_stall_active(iommu))
352 return 0;
354 /* Stall can only be enabled if paging is enabled */
355 if (!rk_iommu_is_paging_enabled(iommu))
356 return 0;
358 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
360 ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
361 val, RK_MMU_POLL_PERIOD_US,
362 RK_MMU_POLL_TIMEOUT_US);
363 if (ret)
364 for (i = 0; i < iommu->num_mmu; i++)
365 dev_err(iommu->dev, "Enable stall request timed out, status: %#08x\n",
366 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
368 return ret;
371 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
373 int ret, i;
374 bool val;
376 if (!rk_iommu_is_stall_active(iommu))
377 return 0;
379 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
381 ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
382 !val, RK_MMU_POLL_PERIOD_US,
383 RK_MMU_POLL_TIMEOUT_US);
384 if (ret)
385 for (i = 0; i < iommu->num_mmu; i++)
386 dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
387 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
389 return ret;
392 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
394 int ret, i;
395 bool val;
397 if (rk_iommu_is_paging_enabled(iommu))
398 return 0;
400 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
402 ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
403 val, RK_MMU_POLL_PERIOD_US,
404 RK_MMU_POLL_TIMEOUT_US);
405 if (ret)
406 for (i = 0; i < iommu->num_mmu; i++)
407 dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
408 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
410 return ret;
413 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
415 int ret, i;
416 bool val;
418 if (!rk_iommu_is_paging_enabled(iommu))
419 return 0;
421 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
423 ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
424 !val, RK_MMU_POLL_PERIOD_US,
425 RK_MMU_POLL_TIMEOUT_US);
426 if (ret)
427 for (i = 0; i < iommu->num_mmu; i++)
428 dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
429 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
431 return ret;
434 static int rk_iommu_force_reset(struct rk_iommu *iommu)
436 int ret, i;
437 u32 dte_addr;
438 bool val;
440 if (iommu->reset_disabled)
441 return 0;
444 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
445 * and verifying that upper 5 nybbles are read back.
447 for (i = 0; i < iommu->num_mmu; i++) {
448 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, DTE_ADDR_DUMMY);
450 dte_addr = rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR);
451 if (dte_addr != (DTE_ADDR_DUMMY & RK_DTE_PT_ADDRESS_MASK)) {
452 dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
453 return -EFAULT;
457 rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
459 ret = readx_poll_timeout(rk_iommu_is_reset_done, iommu, val,
460 val, RK_MMU_FORCE_RESET_TIMEOUT_US,
461 RK_MMU_POLL_TIMEOUT_US);
462 if (ret) {
463 dev_err(iommu->dev, "FORCE_RESET command timed out\n");
464 return ret;
467 return 0;
470 static void log_iova(struct rk_iommu *iommu, int index, dma_addr_t iova)
472 void __iomem *base = iommu->bases[index];
473 u32 dte_index, pte_index, page_offset;
474 u32 mmu_dte_addr;
475 phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
476 u32 *dte_addr;
477 u32 dte;
478 phys_addr_t pte_addr_phys = 0;
479 u32 *pte_addr = NULL;
480 u32 pte = 0;
481 phys_addr_t page_addr_phys = 0;
482 u32 page_flags = 0;
484 dte_index = rk_iova_dte_index(iova);
485 pte_index = rk_iova_pte_index(iova);
486 page_offset = rk_iova_page_offset(iova);
488 mmu_dte_addr = rk_iommu_read(base, RK_MMU_DTE_ADDR);
489 mmu_dte_addr_phys = (phys_addr_t)mmu_dte_addr;
491 dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
492 dte_addr = phys_to_virt(dte_addr_phys);
493 dte = *dte_addr;
495 if (!rk_dte_is_pt_valid(dte))
496 goto print_it;
498 pte_addr_phys = rk_dte_pt_address(dte) + (pte_index * 4);
499 pte_addr = phys_to_virt(pte_addr_phys);
500 pte = *pte_addr;
502 if (!rk_pte_is_page_valid(pte))
503 goto print_it;
505 page_addr_phys = rk_pte_page_address(pte) + page_offset;
506 page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
508 print_it:
509 dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
510 &iova, dte_index, pte_index, page_offset);
511 dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
512 &mmu_dte_addr_phys, &dte_addr_phys, dte,
513 rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
514 rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
517 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
519 struct rk_iommu *iommu = dev_id;
520 u32 status;
521 u32 int_status;
522 dma_addr_t iova;
523 irqreturn_t ret = IRQ_NONE;
524 int i, err;
526 err = pm_runtime_get_if_in_use(iommu->dev);
527 if (WARN_ON_ONCE(err <= 0))
528 return ret;
530 if (WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks)))
531 goto out;
533 for (i = 0; i < iommu->num_mmu; i++) {
534 int_status = rk_iommu_read(iommu->bases[i], RK_MMU_INT_STATUS);
535 if (int_status == 0)
536 continue;
538 ret = IRQ_HANDLED;
539 iova = rk_iommu_read(iommu->bases[i], RK_MMU_PAGE_FAULT_ADDR);
541 if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
542 int flags;
544 status = rk_iommu_read(iommu->bases[i], RK_MMU_STATUS);
545 flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
546 IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
548 dev_err(iommu->dev, "Page fault at %pad of type %s\n",
549 &iova,
550 (flags == IOMMU_FAULT_WRITE) ? "write" : "read");
552 log_iova(iommu, i, iova);
555 * Report page fault to any installed handlers.
556 * Ignore the return code, though, since we always zap cache
557 * and clear the page fault anyway.
559 if (iommu->domain)
560 report_iommu_fault(iommu->domain, iommu->dev, iova,
561 flags);
562 else
563 dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
565 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
566 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_PAGE_FAULT_DONE);
569 if (int_status & RK_MMU_IRQ_BUS_ERROR)
570 dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
572 if (int_status & ~RK_MMU_IRQ_MASK)
573 dev_err(iommu->dev, "unexpected int_status: %#08x\n",
574 int_status);
576 rk_iommu_write(iommu->bases[i], RK_MMU_INT_CLEAR, int_status);
579 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
581 out:
582 pm_runtime_put(iommu->dev);
583 return ret;
586 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
587 dma_addr_t iova)
589 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
590 unsigned long flags;
591 phys_addr_t pt_phys, phys = 0;
592 u32 dte, pte;
593 u32 *page_table;
595 spin_lock_irqsave(&rk_domain->dt_lock, flags);
597 dte = rk_domain->dt[rk_iova_dte_index(iova)];
598 if (!rk_dte_is_pt_valid(dte))
599 goto out;
601 pt_phys = rk_dte_pt_address(dte);
602 page_table = (u32 *)phys_to_virt(pt_phys);
603 pte = page_table[rk_iova_pte_index(iova)];
604 if (!rk_pte_is_page_valid(pte))
605 goto out;
607 phys = rk_pte_page_address(pte) + rk_iova_page_offset(iova);
608 out:
609 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
611 return phys;
614 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
615 dma_addr_t iova, size_t size)
617 struct list_head *pos;
618 unsigned long flags;
620 /* shootdown these iova from all iommus using this domain */
621 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
622 list_for_each(pos, &rk_domain->iommus) {
623 struct rk_iommu *iommu;
624 int ret;
626 iommu = list_entry(pos, struct rk_iommu, node);
628 /* Only zap TLBs of IOMMUs that are powered on. */
629 ret = pm_runtime_get_if_in_use(iommu->dev);
630 if (WARN_ON_ONCE(ret < 0))
631 continue;
632 if (ret) {
633 WARN_ON(clk_bulk_enable(iommu->num_clocks,
634 iommu->clocks));
635 rk_iommu_zap_lines(iommu, iova, size);
636 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
637 pm_runtime_put(iommu->dev);
640 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
643 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
644 dma_addr_t iova, size_t size)
646 rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
647 if (size > SPAGE_SIZE)
648 rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
649 SPAGE_SIZE);
652 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
653 dma_addr_t iova)
655 u32 *page_table, *dte_addr;
656 u32 dte_index, dte;
657 phys_addr_t pt_phys;
658 dma_addr_t pt_dma;
660 assert_spin_locked(&rk_domain->dt_lock);
662 dte_index = rk_iova_dte_index(iova);
663 dte_addr = &rk_domain->dt[dte_index];
664 dte = *dte_addr;
665 if (rk_dte_is_pt_valid(dte))
666 goto done;
668 page_table = (u32 *)get_zeroed_page(GFP_ATOMIC | GFP_DMA32);
669 if (!page_table)
670 return ERR_PTR(-ENOMEM);
672 pt_dma = dma_map_single(dma_dev, page_table, SPAGE_SIZE, DMA_TO_DEVICE);
673 if (dma_mapping_error(dma_dev, pt_dma)) {
674 dev_err(dma_dev, "DMA mapping error while allocating page table\n");
675 free_page((unsigned long)page_table);
676 return ERR_PTR(-ENOMEM);
679 dte = rk_mk_dte(pt_dma);
680 *dte_addr = dte;
682 rk_table_flush(rk_domain, pt_dma, NUM_PT_ENTRIES);
683 rk_table_flush(rk_domain,
684 rk_domain->dt_dma + dte_index * sizeof(u32), 1);
685 done:
686 pt_phys = rk_dte_pt_address(dte);
687 return (u32 *)phys_to_virt(pt_phys);
690 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
691 u32 *pte_addr, dma_addr_t pte_dma,
692 size_t size)
694 unsigned int pte_count;
695 unsigned int pte_total = size / SPAGE_SIZE;
697 assert_spin_locked(&rk_domain->dt_lock);
699 for (pte_count = 0; pte_count < pte_total; pte_count++) {
700 u32 pte = pte_addr[pte_count];
701 if (!rk_pte_is_page_valid(pte))
702 break;
704 pte_addr[pte_count] = rk_mk_pte_invalid(pte);
707 rk_table_flush(rk_domain, pte_dma, pte_count);
709 return pte_count * SPAGE_SIZE;
712 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
713 dma_addr_t pte_dma, dma_addr_t iova,
714 phys_addr_t paddr, size_t size, int prot)
716 unsigned int pte_count;
717 unsigned int pte_total = size / SPAGE_SIZE;
718 phys_addr_t page_phys;
720 assert_spin_locked(&rk_domain->dt_lock);
722 for (pte_count = 0; pte_count < pte_total; pte_count++) {
723 u32 pte = pte_addr[pte_count];
725 if (rk_pte_is_page_valid(pte))
726 goto unwind;
728 pte_addr[pte_count] = rk_mk_pte(paddr, prot);
730 paddr += SPAGE_SIZE;
733 rk_table_flush(rk_domain, pte_dma, pte_total);
736 * Zap the first and last iova to evict from iotlb any previously
737 * mapped cachelines holding stale values for its dte and pte.
738 * We only zap the first and last iova, since only they could have
739 * dte or pte shared with an existing mapping.
741 rk_iommu_zap_iova_first_last(rk_domain, iova, size);
743 return 0;
744 unwind:
745 /* Unmap the range of iovas that we just mapped */
746 rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma,
747 pte_count * SPAGE_SIZE);
749 iova += pte_count * SPAGE_SIZE;
750 page_phys = rk_pte_page_address(pte_addr[pte_count]);
751 pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
752 &iova, &page_phys, &paddr, prot);
754 return -EADDRINUSE;
757 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
758 phys_addr_t paddr, size_t size, int prot)
760 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
761 unsigned long flags;
762 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
763 u32 *page_table, *pte_addr;
764 u32 dte_index, pte_index;
765 int ret;
767 spin_lock_irqsave(&rk_domain->dt_lock, flags);
770 * pgsize_bitmap specifies iova sizes that fit in one page table
771 * (1024 4-KiB pages = 4 MiB).
772 * So, size will always be 4096 <= size <= 4194304.
773 * Since iommu_map() guarantees that both iova and size will be
774 * aligned, we will always only be mapping from a single dte here.
776 page_table = rk_dte_get_page_table(rk_domain, iova);
777 if (IS_ERR(page_table)) {
778 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
779 return PTR_ERR(page_table);
782 dte_index = rk_domain->dt[rk_iova_dte_index(iova)];
783 pte_index = rk_iova_pte_index(iova);
784 pte_addr = &page_table[pte_index];
785 pte_dma = rk_dte_pt_address(dte_index) + pte_index * sizeof(u32);
786 ret = rk_iommu_map_iova(rk_domain, pte_addr, pte_dma, iova,
787 paddr, size, prot);
789 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
791 return ret;
794 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
795 size_t size)
797 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
798 unsigned long flags;
799 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
800 phys_addr_t pt_phys;
801 u32 dte;
802 u32 *pte_addr;
803 size_t unmap_size;
805 spin_lock_irqsave(&rk_domain->dt_lock, flags);
808 * pgsize_bitmap specifies iova sizes that fit in one page table
809 * (1024 4-KiB pages = 4 MiB).
810 * So, size will always be 4096 <= size <= 4194304.
811 * Since iommu_unmap() guarantees that both iova and size will be
812 * aligned, we will always only be unmapping from a single dte here.
814 dte = rk_domain->dt[rk_iova_dte_index(iova)];
815 /* Just return 0 if iova is unmapped */
816 if (!rk_dte_is_pt_valid(dte)) {
817 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
818 return 0;
821 pt_phys = rk_dte_pt_address(dte);
822 pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
823 pte_dma = pt_phys + rk_iova_pte_index(iova) * sizeof(u32);
824 unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma, size);
826 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
828 /* Shootdown iotlb entries for iova range that was just unmapped */
829 rk_iommu_zap_iova(rk_domain, iova, unmap_size);
831 return unmap_size;
834 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
836 struct rk_iommudata *data = dev->archdata.iommu;
838 return data ? data->iommu : NULL;
841 /* Must be called with iommu powered on and attached */
842 static void rk_iommu_disable(struct rk_iommu *iommu)
844 int i;
846 /* Ignore error while disabling, just keep going */
847 WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks));
848 rk_iommu_enable_stall(iommu);
849 rk_iommu_disable_paging(iommu);
850 for (i = 0; i < iommu->num_mmu; i++) {
851 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, 0);
852 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, 0);
854 rk_iommu_disable_stall(iommu);
855 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
858 /* Must be called with iommu powered on and attached */
859 static int rk_iommu_enable(struct rk_iommu *iommu)
861 struct iommu_domain *domain = iommu->domain;
862 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
863 int ret, i;
865 ret = clk_bulk_enable(iommu->num_clocks, iommu->clocks);
866 if (ret)
867 return ret;
869 ret = rk_iommu_enable_stall(iommu);
870 if (ret)
871 goto out_disable_clocks;
873 ret = rk_iommu_force_reset(iommu);
874 if (ret)
875 goto out_disable_stall;
877 for (i = 0; i < iommu->num_mmu; i++) {
878 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
879 rk_domain->dt_dma);
880 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
881 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
884 ret = rk_iommu_enable_paging(iommu);
886 out_disable_stall:
887 rk_iommu_disable_stall(iommu);
888 out_disable_clocks:
889 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
890 return ret;
893 static void rk_iommu_detach_device(struct iommu_domain *domain,
894 struct device *dev)
896 struct rk_iommu *iommu;
897 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
898 unsigned long flags;
899 int ret;
901 /* Allow 'virtual devices' (eg drm) to detach from domain */
902 iommu = rk_iommu_from_dev(dev);
903 if (!iommu)
904 return;
906 dev_dbg(dev, "Detaching from iommu domain\n");
908 /* iommu already detached */
909 if (iommu->domain != domain)
910 return;
912 iommu->domain = NULL;
914 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
915 list_del_init(&iommu->node);
916 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
918 ret = pm_runtime_get_if_in_use(iommu->dev);
919 WARN_ON_ONCE(ret < 0);
920 if (ret > 0) {
921 rk_iommu_disable(iommu);
922 pm_runtime_put(iommu->dev);
926 static int rk_iommu_attach_device(struct iommu_domain *domain,
927 struct device *dev)
929 struct rk_iommu *iommu;
930 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
931 unsigned long flags;
932 int ret;
935 * Allow 'virtual devices' (e.g., drm) to attach to domain.
936 * Such a device does not belong to an iommu group.
938 iommu = rk_iommu_from_dev(dev);
939 if (!iommu)
940 return 0;
942 dev_dbg(dev, "Attaching to iommu domain\n");
944 /* iommu already attached */
945 if (iommu->domain == domain)
946 return 0;
948 if (iommu->domain)
949 rk_iommu_detach_device(iommu->domain, dev);
951 iommu->domain = domain;
953 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
954 list_add_tail(&iommu->node, &rk_domain->iommus);
955 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
957 ret = pm_runtime_get_if_in_use(iommu->dev);
958 if (!ret || WARN_ON_ONCE(ret < 0))
959 return 0;
961 ret = rk_iommu_enable(iommu);
962 if (ret)
963 rk_iommu_detach_device(iommu->domain, dev);
965 pm_runtime_put(iommu->dev);
967 return ret;
970 static struct iommu_domain *rk_iommu_domain_alloc(unsigned type)
972 struct rk_iommu_domain *rk_domain;
974 if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
975 return NULL;
977 if (!dma_dev)
978 return NULL;
980 rk_domain = kzalloc(sizeof(*rk_domain), GFP_KERNEL);
981 if (!rk_domain)
982 return NULL;
984 if (type == IOMMU_DOMAIN_DMA &&
985 iommu_get_dma_cookie(&rk_domain->domain))
986 goto err_free_domain;
989 * rk32xx iommus use a 2 level pagetable.
990 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
991 * Allocate one 4 KiB page for each table.
993 rk_domain->dt = (u32 *)get_zeroed_page(GFP_KERNEL | GFP_DMA32);
994 if (!rk_domain->dt)
995 goto err_put_cookie;
997 rk_domain->dt_dma = dma_map_single(dma_dev, rk_domain->dt,
998 SPAGE_SIZE, DMA_TO_DEVICE);
999 if (dma_mapping_error(dma_dev, rk_domain->dt_dma)) {
1000 dev_err(dma_dev, "DMA map error for DT\n");
1001 goto err_free_dt;
1004 rk_table_flush(rk_domain, rk_domain->dt_dma, NUM_DT_ENTRIES);
1006 spin_lock_init(&rk_domain->iommus_lock);
1007 spin_lock_init(&rk_domain->dt_lock);
1008 INIT_LIST_HEAD(&rk_domain->iommus);
1010 rk_domain->domain.geometry.aperture_start = 0;
1011 rk_domain->domain.geometry.aperture_end = DMA_BIT_MASK(32);
1012 rk_domain->domain.geometry.force_aperture = true;
1014 return &rk_domain->domain;
1016 err_free_dt:
1017 free_page((unsigned long)rk_domain->dt);
1018 err_put_cookie:
1019 if (type == IOMMU_DOMAIN_DMA)
1020 iommu_put_dma_cookie(&rk_domain->domain);
1021 err_free_domain:
1022 kfree(rk_domain);
1024 return NULL;
1027 static void rk_iommu_domain_free(struct iommu_domain *domain)
1029 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1030 int i;
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);
1051 kfree(rk_domain);
1054 static int rk_iommu_add_device(struct device *dev)
1056 struct iommu_group *group;
1057 struct rk_iommu *iommu;
1058 struct rk_iommudata *data;
1060 data = dev->archdata.iommu;
1061 if (!data)
1062 return -ENODEV;
1064 iommu = rk_iommu_from_dev(dev);
1066 group = iommu_group_get_for_dev(dev);
1067 if (IS_ERR(group))
1068 return PTR_ERR(group);
1069 iommu_group_put(group);
1071 iommu_device_link(&iommu->iommu, dev);
1072 data->link = device_link_add(dev, iommu->dev, DL_FLAG_PM_RUNTIME);
1074 return 0;
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);
1105 if (!data)
1106 return -ENOMEM;
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);
1115 return 0;
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;
1139 int err, i, irq;
1141 iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
1142 if (!iommu)
1143 return -ENOMEM;
1145 platform_set_drvdata(pdev, iommu);
1146 iommu->dev = dev;
1147 iommu->num_mmu = 0;
1149 iommu->bases = devm_kcalloc(dev, num_res, sizeof(*iommu->bases),
1150 GFP_KERNEL);
1151 if (!iommu->bases)
1152 return -ENOMEM;
1154 for (i = 0; i < num_res; i++) {
1155 res = platform_get_resource(pdev, IORESOURCE_MEM, i);
1156 if (!res)
1157 continue;
1158 iommu->bases[i] = devm_ioremap_resource(&pdev->dev, res);
1159 if (IS_ERR(iommu->bases[i]))
1160 continue;
1161 iommu->num_mmu++;
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);
1172 if (!iommu->clocks)
1173 return -ENOMEM;
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);
1184 if (err == -ENOENT)
1185 iommu->num_clocks = 0;
1186 else if (err)
1187 return err;
1189 err = clk_bulk_prepare(iommu->num_clocks, iommu->clocks);
1190 if (err)
1191 return err;
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));
1200 if (err)
1201 goto err_put_group;
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);
1207 if (err)
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
1213 * IOMMU device..
1215 if (!dma_dev)
1216 dma_dev = &pdev->dev;
1218 bus_set_iommu(&platform_bus_type, &rk_iommu_ops);
1220 pm_runtime_enable(dev);
1222 i = 0;
1223 while ((irq = platform_get_irq(pdev, i++)) != -ENXIO) {
1224 if (irq < 0)
1225 return irq;
1227 err = devm_request_irq(iommu->dev, irq, rk_iommu_irq,
1228 IRQF_SHARED, dev_name(dev), iommu);
1229 if (err) {
1230 pm_runtime_disable(dev);
1231 goto err_remove_sysfs;
1235 return 0;
1236 err_remove_sysfs:
1237 iommu_device_sysfs_remove(&iommu->iommu);
1238 err_put_group:
1239 iommu_group_put(iommu->group);
1240 err_unprepare_clocks:
1241 clk_bulk_unprepare(iommu->num_clocks, iommu->clocks);
1242 return err;
1245 static void rk_iommu_shutdown(struct platform_device *pdev)
1247 struct rk_iommu *iommu = platform_get_drvdata(pdev);
1248 int i = 0, irq;
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);
1260 if (!iommu->domain)
1261 return 0;
1263 rk_iommu_disable(iommu);
1264 return 0;
1267 static int __maybe_unused rk_iommu_resume(struct device *dev)
1269 struct rk_iommu *iommu = dev_get_drvdata(dev);
1271 if (!iommu->domain)
1272 return 0;
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" },
1285 { /* sentinel */ }
1287 MODULE_DEVICE_TABLE(of, rk_iommu_dt_ids);
1289 static struct platform_driver rk_iommu_driver = {
1290 .probe = rk_iommu_probe,
1291 .shutdown = rk_iommu_shutdown,
1292 .driver = {
1293 .name = "rk_iommu",
1294 .of_match_table = rk_iommu_dt_ids,
1295 .pm = &rk_iommu_pm_ops,
1296 .suppress_bind_attrs = true,
1300 static int __init rk_iommu_init(void)
1302 return platform_driver_register(&rk_iommu_driver);
1304 subsys_initcall(rk_iommu_init);
1306 MODULE_DESCRIPTION("IOMMU API for Rockchip");
1307 MODULE_AUTHOR("Simon Xue <xxm@rock-chips.com> and Daniel Kurtz <djkurtz@chromium.org>");
1308 MODULE_ALIAS("platform:rockchip-iommu");
1309 MODULE_LICENSE("GPL v2");