Merge tag 'io_uring-5.11-2021-01-16' of git://git.kernel.dk/linux-block
[linux/fpc-iii.git] / drivers / iommu / rockchip-iommu.c
blobe5d86b7177dec8aa42833fb9cdb587623886b76f
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * IOMMU API for Rockchip
5 * Module Authors: Simon Xue <xxm@rock-chips.com>
6 * Daniel Kurtz <djkurtz@chromium.org>
7 */
9 #include <linux/clk.h>
10 #include <linux/compiler.h>
11 #include <linux/delay.h>
12 #include <linux/device.h>
13 #include <linux/dma-iommu.h>
14 #include <linux/dma-mapping.h>
15 #include <linux/errno.h>
16 #include <linux/interrupt.h>
17 #include <linux/io.h>
18 #include <linux/iommu.h>
19 #include <linux/iopoll.h>
20 #include <linux/list.h>
21 #include <linux/mm.h>
22 #include <linux/init.h>
23 #include <linux/of.h>
24 #include <linux/of_iommu.h>
25 #include <linux/of_platform.h>
26 #include <linux/platform_device.h>
27 #include <linux/pm_runtime.h>
28 #include <linux/slab.h>
29 #include <linux/spinlock.h>
31 /** MMU register offsets */
32 #define RK_MMU_DTE_ADDR 0x00 /* Directory table address */
33 #define RK_MMU_STATUS 0x04
34 #define RK_MMU_COMMAND 0x08
35 #define RK_MMU_PAGE_FAULT_ADDR 0x0C /* IOVA of last page fault */
36 #define RK_MMU_ZAP_ONE_LINE 0x10 /* Shootdown one IOTLB entry */
37 #define RK_MMU_INT_RAWSTAT 0x14 /* IRQ status ignoring mask */
38 #define RK_MMU_INT_CLEAR 0x18 /* Acknowledge and re-arm irq */
39 #define RK_MMU_INT_MASK 0x1C /* IRQ enable */
40 #define RK_MMU_INT_STATUS 0x20 /* IRQ status after masking */
41 #define RK_MMU_AUTO_GATING 0x24
43 #define DTE_ADDR_DUMMY 0xCAFEBABE
45 #define RK_MMU_POLL_PERIOD_US 100
46 #define RK_MMU_FORCE_RESET_TIMEOUT_US 100000
47 #define RK_MMU_POLL_TIMEOUT_US 1000
49 /* RK_MMU_STATUS fields */
50 #define RK_MMU_STATUS_PAGING_ENABLED BIT(0)
51 #define RK_MMU_STATUS_PAGE_FAULT_ACTIVE BIT(1)
52 #define RK_MMU_STATUS_STALL_ACTIVE BIT(2)
53 #define RK_MMU_STATUS_IDLE BIT(3)
54 #define RK_MMU_STATUS_REPLAY_BUFFER_EMPTY BIT(4)
55 #define RK_MMU_STATUS_PAGE_FAULT_IS_WRITE BIT(5)
56 #define RK_MMU_STATUS_STALL_NOT_ACTIVE BIT(31)
58 /* RK_MMU_COMMAND command values */
59 #define RK_MMU_CMD_ENABLE_PAGING 0 /* Enable memory translation */
60 #define RK_MMU_CMD_DISABLE_PAGING 1 /* Disable memory translation */
61 #define RK_MMU_CMD_ENABLE_STALL 2 /* Stall paging to allow other cmds */
62 #define RK_MMU_CMD_DISABLE_STALL 3 /* Stop stall re-enables paging */
63 #define RK_MMU_CMD_ZAP_CACHE 4 /* Shoot down entire IOTLB */
64 #define RK_MMU_CMD_PAGE_FAULT_DONE 5 /* Clear page fault */
65 #define RK_MMU_CMD_FORCE_RESET 6 /* Reset all registers */
67 /* RK_MMU_INT_* register fields */
68 #define RK_MMU_IRQ_PAGE_FAULT 0x01 /* page fault */
69 #define RK_MMU_IRQ_BUS_ERROR 0x02 /* bus read error */
70 #define RK_MMU_IRQ_MASK (RK_MMU_IRQ_PAGE_FAULT | RK_MMU_IRQ_BUS_ERROR)
72 #define NUM_DT_ENTRIES 1024
73 #define NUM_PT_ENTRIES 1024
75 #define SPAGE_ORDER 12
76 #define SPAGE_SIZE (1 << SPAGE_ORDER)
79 * Support mapping any size that fits in one page table:
80 * 4 KiB to 4 MiB
82 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
84 struct rk_iommu_domain {
85 struct list_head iommus;
86 u32 *dt; /* page directory table */
87 dma_addr_t dt_dma;
88 spinlock_t iommus_lock; /* lock for iommus list */
89 spinlock_t dt_lock; /* lock for modifying page directory table */
91 struct iommu_domain domain;
94 /* list of clocks required by IOMMU */
95 static const char * const rk_iommu_clocks[] = {
96 "aclk", "iface",
99 struct rk_iommu {
100 struct device *dev;
101 void __iomem **bases;
102 int num_mmu;
103 int num_irq;
104 struct clk_bulk_data *clocks;
105 int num_clocks;
106 bool reset_disabled;
107 struct iommu_device iommu;
108 struct list_head node; /* entry in rk_iommu_domain.iommus */
109 struct iommu_domain *domain; /* domain to which iommu is attached */
110 struct iommu_group *group;
113 struct rk_iommudata {
114 struct device_link *link; /* runtime PM link from IOMMU to master */
115 struct rk_iommu *iommu;
118 static struct device *dma_dev;
120 static inline void rk_table_flush(struct rk_iommu_domain *dom, dma_addr_t dma,
121 unsigned int count)
123 size_t size = count * sizeof(u32); /* count of u32 entry */
125 dma_sync_single_for_device(dma_dev, dma, size, DMA_TO_DEVICE);
128 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
130 return container_of(dom, struct rk_iommu_domain, domain);
134 * The Rockchip rk3288 iommu uses a 2-level page table.
135 * The first level is the "Directory Table" (DT).
136 * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
137 * to a "Page Table".
138 * The second level is the 1024 Page Tables (PT).
139 * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
140 * a 4 KB page of physical memory.
142 * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
143 * Each iommu device has a MMU_DTE_ADDR register that contains the physical
144 * address of the start of the DT page.
146 * The structure of the page table is as follows:
148 * DT
149 * MMU_DTE_ADDR -> +-----+
150 * | |
151 * +-----+ PT
152 * | DTE | -> +-----+
153 * +-----+ | | Memory
154 * | | +-----+ Page
155 * | | | PTE | -> +-----+
156 * +-----+ +-----+ | |
157 * | | | |
158 * | | | |
159 * +-----+ | |
160 * | |
161 * | |
162 * +-----+
166 * Each DTE has a PT address and a valid bit:
167 * +---------------------+-----------+-+
168 * | PT address | Reserved |V|
169 * +---------------------+-----------+-+
170 * 31:12 - PT address (PTs always starts on a 4 KB boundary)
171 * 11: 1 - Reserved
172 * 0 - 1 if PT @ PT address is valid
174 #define RK_DTE_PT_ADDRESS_MASK 0xfffff000
175 #define RK_DTE_PT_VALID BIT(0)
177 static inline phys_addr_t rk_dte_pt_address(u32 dte)
179 return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
182 static inline bool rk_dte_is_pt_valid(u32 dte)
184 return dte & RK_DTE_PT_VALID;
187 static inline u32 rk_mk_dte(dma_addr_t pt_dma)
189 return (pt_dma & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
193 * Each PTE has a Page address, some flags and a valid bit:
194 * +---------------------+---+-------+-+
195 * | Page address |Rsv| Flags |V|
196 * +---------------------+---+-------+-+
197 * 31:12 - Page address (Pages always start on a 4 KB boundary)
198 * 11: 9 - Reserved
199 * 8: 1 - Flags
200 * 8 - Read allocate - allocate cache space on read misses
201 * 7 - Read cache - enable cache & prefetch of data
202 * 6 - Write buffer - enable delaying writes on their way to memory
203 * 5 - Write allocate - allocate cache space on write misses
204 * 4 - Write cache - different writes can be merged together
205 * 3 - Override cache attributes
206 * if 1, bits 4-8 control cache attributes
207 * if 0, the system bus defaults are used
208 * 2 - Writable
209 * 1 - Readable
210 * 0 - 1 if Page @ Page address is valid
212 #define RK_PTE_PAGE_ADDRESS_MASK 0xfffff000
213 #define RK_PTE_PAGE_FLAGS_MASK 0x000001fe
214 #define RK_PTE_PAGE_WRITABLE BIT(2)
215 #define RK_PTE_PAGE_READABLE BIT(1)
216 #define RK_PTE_PAGE_VALID BIT(0)
218 static inline phys_addr_t rk_pte_page_address(u32 pte)
220 return (phys_addr_t)pte & RK_PTE_PAGE_ADDRESS_MASK;
223 static inline bool rk_pte_is_page_valid(u32 pte)
225 return pte & RK_PTE_PAGE_VALID;
228 /* TODO: set cache flags per prot IOMMU_CACHE */
229 static u32 rk_mk_pte(phys_addr_t page, int prot)
231 u32 flags = 0;
232 flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
233 flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
234 page &= RK_PTE_PAGE_ADDRESS_MASK;
235 return page | flags | RK_PTE_PAGE_VALID;
238 static u32 rk_mk_pte_invalid(u32 pte)
240 return pte & ~RK_PTE_PAGE_VALID;
244 * rk3288 iova (IOMMU Virtual Address) format
245 * 31 22.21 12.11 0
246 * +-----------+-----------+-------------+
247 * | DTE index | PTE index | Page offset |
248 * +-----------+-----------+-------------+
249 * 31:22 - DTE index - index of DTE in DT
250 * 21:12 - PTE index - index of PTE in PT @ DTE.pt_address
251 * 11: 0 - Page offset - offset into page @ PTE.page_address
253 #define RK_IOVA_DTE_MASK 0xffc00000
254 #define RK_IOVA_DTE_SHIFT 22
255 #define RK_IOVA_PTE_MASK 0x003ff000
256 #define RK_IOVA_PTE_SHIFT 12
257 #define RK_IOVA_PAGE_MASK 0x00000fff
258 #define RK_IOVA_PAGE_SHIFT 0
260 static u32 rk_iova_dte_index(dma_addr_t iova)
262 return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
265 static u32 rk_iova_pte_index(dma_addr_t iova)
267 return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
270 static u32 rk_iova_page_offset(dma_addr_t iova)
272 return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
275 static u32 rk_iommu_read(void __iomem *base, u32 offset)
277 return readl(base + offset);
280 static void rk_iommu_write(void __iomem *base, u32 offset, u32 value)
282 writel(value, base + offset);
285 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
287 int i;
289 for (i = 0; i < iommu->num_mmu; i++)
290 writel(command, iommu->bases[i] + RK_MMU_COMMAND);
293 static void rk_iommu_base_command(void __iomem *base, u32 command)
295 writel(command, base + RK_MMU_COMMAND);
297 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova_start,
298 size_t size)
300 int i;
301 dma_addr_t iova_end = iova_start + size;
303 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
304 * entire iotlb rather than iterate over individual iovas.
306 for (i = 0; i < iommu->num_mmu; i++) {
307 dma_addr_t iova;
309 for (iova = iova_start; iova < iova_end; iova += SPAGE_SIZE)
310 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 bool rk_iommu_is_reset_done(struct rk_iommu *iommu)
340 bool done = true;
341 int i;
343 for (i = 0; i < iommu->num_mmu; i++)
344 done &= rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR) == 0;
346 return done;
349 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
351 int ret, i;
352 bool val;
354 if (rk_iommu_is_stall_active(iommu))
355 return 0;
357 /* Stall can only be enabled if paging is enabled */
358 if (!rk_iommu_is_paging_enabled(iommu))
359 return 0;
361 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
363 ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
364 val, RK_MMU_POLL_PERIOD_US,
365 RK_MMU_POLL_TIMEOUT_US);
366 if (ret)
367 for (i = 0; i < iommu->num_mmu; i++)
368 dev_err(iommu->dev, "Enable stall request timed out, status: %#08x\n",
369 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
371 return ret;
374 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
376 int ret, i;
377 bool val;
379 if (!rk_iommu_is_stall_active(iommu))
380 return 0;
382 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
384 ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
385 !val, RK_MMU_POLL_PERIOD_US,
386 RK_MMU_POLL_TIMEOUT_US);
387 if (ret)
388 for (i = 0; i < iommu->num_mmu; i++)
389 dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
390 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
392 return ret;
395 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
397 int ret, i;
398 bool val;
400 if (rk_iommu_is_paging_enabled(iommu))
401 return 0;
403 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
405 ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
406 val, RK_MMU_POLL_PERIOD_US,
407 RK_MMU_POLL_TIMEOUT_US);
408 if (ret)
409 for (i = 0; i < iommu->num_mmu; i++)
410 dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
411 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
413 return ret;
416 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
418 int ret, i;
419 bool val;
421 if (!rk_iommu_is_paging_enabled(iommu))
422 return 0;
424 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
426 ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
427 !val, RK_MMU_POLL_PERIOD_US,
428 RK_MMU_POLL_TIMEOUT_US);
429 if (ret)
430 for (i = 0; i < iommu->num_mmu; i++)
431 dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
432 rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
434 return ret;
437 static int rk_iommu_force_reset(struct rk_iommu *iommu)
439 int ret, i;
440 u32 dte_addr;
441 bool val;
443 if (iommu->reset_disabled)
444 return 0;
447 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
448 * and verifying that upper 5 nybbles are read back.
450 for (i = 0; i < iommu->num_mmu; i++) {
451 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, DTE_ADDR_DUMMY);
453 dte_addr = rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR);
454 if (dte_addr != (DTE_ADDR_DUMMY & RK_DTE_PT_ADDRESS_MASK)) {
455 dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
456 return -EFAULT;
460 rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
462 ret = readx_poll_timeout(rk_iommu_is_reset_done, iommu, val,
463 val, RK_MMU_FORCE_RESET_TIMEOUT_US,
464 RK_MMU_POLL_TIMEOUT_US);
465 if (ret) {
466 dev_err(iommu->dev, "FORCE_RESET command timed out\n");
467 return ret;
470 return 0;
473 static void log_iova(struct rk_iommu *iommu, int index, dma_addr_t iova)
475 void __iomem *base = iommu->bases[index];
476 u32 dte_index, pte_index, page_offset;
477 u32 mmu_dte_addr;
478 phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
479 u32 *dte_addr;
480 u32 dte;
481 phys_addr_t pte_addr_phys = 0;
482 u32 *pte_addr = NULL;
483 u32 pte = 0;
484 phys_addr_t page_addr_phys = 0;
485 u32 page_flags = 0;
487 dte_index = rk_iova_dte_index(iova);
488 pte_index = rk_iova_pte_index(iova);
489 page_offset = rk_iova_page_offset(iova);
491 mmu_dte_addr = rk_iommu_read(base, RK_MMU_DTE_ADDR);
492 mmu_dte_addr_phys = (phys_addr_t)mmu_dte_addr;
494 dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
495 dte_addr = phys_to_virt(dte_addr_phys);
496 dte = *dte_addr;
498 if (!rk_dte_is_pt_valid(dte))
499 goto print_it;
501 pte_addr_phys = rk_dte_pt_address(dte) + (pte_index * 4);
502 pte_addr = phys_to_virt(pte_addr_phys);
503 pte = *pte_addr;
505 if (!rk_pte_is_page_valid(pte))
506 goto print_it;
508 page_addr_phys = rk_pte_page_address(pte) + page_offset;
509 page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
511 print_it:
512 dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
513 &iova, dte_index, pte_index, page_offset);
514 dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
515 &mmu_dte_addr_phys, &dte_addr_phys, dte,
516 rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
517 rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
520 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
522 struct rk_iommu *iommu = dev_id;
523 u32 status;
524 u32 int_status;
525 dma_addr_t iova;
526 irqreturn_t ret = IRQ_NONE;
527 int i, err;
529 err = pm_runtime_get_if_in_use(iommu->dev);
530 if (!err || WARN_ON_ONCE(err < 0))
531 return ret;
533 if (WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks)))
534 goto out;
536 for (i = 0; i < iommu->num_mmu; i++) {
537 int_status = rk_iommu_read(iommu->bases[i], RK_MMU_INT_STATUS);
538 if (int_status == 0)
539 continue;
541 ret = IRQ_HANDLED;
542 iova = rk_iommu_read(iommu->bases[i], RK_MMU_PAGE_FAULT_ADDR);
544 if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
545 int flags;
547 status = rk_iommu_read(iommu->bases[i], RK_MMU_STATUS);
548 flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
549 IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
551 dev_err(iommu->dev, "Page fault at %pad of type %s\n",
552 &iova,
553 (flags == IOMMU_FAULT_WRITE) ? "write" : "read");
555 log_iova(iommu, i, iova);
558 * Report page fault to any installed handlers.
559 * Ignore the return code, though, since we always zap cache
560 * and clear the page fault anyway.
562 if (iommu->domain)
563 report_iommu_fault(iommu->domain, iommu->dev, iova,
564 flags);
565 else
566 dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
568 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
569 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_PAGE_FAULT_DONE);
572 if (int_status & RK_MMU_IRQ_BUS_ERROR)
573 dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
575 if (int_status & ~RK_MMU_IRQ_MASK)
576 dev_err(iommu->dev, "unexpected int_status: %#08x\n",
577 int_status);
579 rk_iommu_write(iommu->bases[i], RK_MMU_INT_CLEAR, int_status);
582 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
584 out:
585 pm_runtime_put(iommu->dev);
586 return ret;
589 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
590 dma_addr_t iova)
592 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
593 unsigned long flags;
594 phys_addr_t pt_phys, phys = 0;
595 u32 dte, pte;
596 u32 *page_table;
598 spin_lock_irqsave(&rk_domain->dt_lock, flags);
600 dte = rk_domain->dt[rk_iova_dte_index(iova)];
601 if (!rk_dte_is_pt_valid(dte))
602 goto out;
604 pt_phys = rk_dte_pt_address(dte);
605 page_table = (u32 *)phys_to_virt(pt_phys);
606 pte = page_table[rk_iova_pte_index(iova)];
607 if (!rk_pte_is_page_valid(pte))
608 goto out;
610 phys = rk_pte_page_address(pte) + rk_iova_page_offset(iova);
611 out:
612 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
614 return phys;
617 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
618 dma_addr_t iova, size_t size)
620 struct list_head *pos;
621 unsigned long flags;
623 /* shootdown these iova from all iommus using this domain */
624 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
625 list_for_each(pos, &rk_domain->iommus) {
626 struct rk_iommu *iommu;
627 int ret;
629 iommu = list_entry(pos, struct rk_iommu, node);
631 /* Only zap TLBs of IOMMUs that are powered on. */
632 ret = pm_runtime_get_if_in_use(iommu->dev);
633 if (WARN_ON_ONCE(ret < 0))
634 continue;
635 if (ret) {
636 WARN_ON(clk_bulk_enable(iommu->num_clocks,
637 iommu->clocks));
638 rk_iommu_zap_lines(iommu, iova, size);
639 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
640 pm_runtime_put(iommu->dev);
643 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
646 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
647 dma_addr_t iova, size_t size)
649 rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
650 if (size > SPAGE_SIZE)
651 rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
652 SPAGE_SIZE);
655 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
656 dma_addr_t iova)
658 u32 *page_table, *dte_addr;
659 u32 dte_index, dte;
660 phys_addr_t pt_phys;
661 dma_addr_t pt_dma;
663 assert_spin_locked(&rk_domain->dt_lock);
665 dte_index = rk_iova_dte_index(iova);
666 dte_addr = &rk_domain->dt[dte_index];
667 dte = *dte_addr;
668 if (rk_dte_is_pt_valid(dte))
669 goto done;
671 page_table = (u32 *)get_zeroed_page(GFP_ATOMIC | GFP_DMA32);
672 if (!page_table)
673 return ERR_PTR(-ENOMEM);
675 pt_dma = dma_map_single(dma_dev, page_table, SPAGE_SIZE, DMA_TO_DEVICE);
676 if (dma_mapping_error(dma_dev, pt_dma)) {
677 dev_err(dma_dev, "DMA mapping error while allocating page table\n");
678 free_page((unsigned long)page_table);
679 return ERR_PTR(-ENOMEM);
682 dte = rk_mk_dte(pt_dma);
683 *dte_addr = dte;
685 rk_table_flush(rk_domain, pt_dma, NUM_PT_ENTRIES);
686 rk_table_flush(rk_domain,
687 rk_domain->dt_dma + dte_index * sizeof(u32), 1);
688 done:
689 pt_phys = rk_dte_pt_address(dte);
690 return (u32 *)phys_to_virt(pt_phys);
693 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
694 u32 *pte_addr, dma_addr_t pte_dma,
695 size_t size)
697 unsigned int pte_count;
698 unsigned int pte_total = size / SPAGE_SIZE;
700 assert_spin_locked(&rk_domain->dt_lock);
702 for (pte_count = 0; pte_count < pte_total; pte_count++) {
703 u32 pte = pte_addr[pte_count];
704 if (!rk_pte_is_page_valid(pte))
705 break;
707 pte_addr[pte_count] = rk_mk_pte_invalid(pte);
710 rk_table_flush(rk_domain, pte_dma, pte_count);
712 return pte_count * SPAGE_SIZE;
715 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
716 dma_addr_t pte_dma, dma_addr_t iova,
717 phys_addr_t paddr, size_t size, int prot)
719 unsigned int pte_count;
720 unsigned int pte_total = size / SPAGE_SIZE;
721 phys_addr_t page_phys;
723 assert_spin_locked(&rk_domain->dt_lock);
725 for (pte_count = 0; pte_count < pte_total; pte_count++) {
726 u32 pte = pte_addr[pte_count];
728 if (rk_pte_is_page_valid(pte))
729 goto unwind;
731 pte_addr[pte_count] = rk_mk_pte(paddr, prot);
733 paddr += SPAGE_SIZE;
736 rk_table_flush(rk_domain, pte_dma, pte_total);
739 * Zap the first and last iova to evict from iotlb any previously
740 * mapped cachelines holding stale values for its dte and pte.
741 * We only zap the first and last iova, since only they could have
742 * dte or pte shared with an existing mapping.
744 rk_iommu_zap_iova_first_last(rk_domain, iova, size);
746 return 0;
747 unwind:
748 /* Unmap the range of iovas that we just mapped */
749 rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma,
750 pte_count * SPAGE_SIZE);
752 iova += pte_count * SPAGE_SIZE;
753 page_phys = rk_pte_page_address(pte_addr[pte_count]);
754 pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
755 &iova, &page_phys, &paddr, prot);
757 return -EADDRINUSE;
760 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
761 phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
763 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
764 unsigned long flags;
765 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
766 u32 *page_table, *pte_addr;
767 u32 dte_index, pte_index;
768 int ret;
770 spin_lock_irqsave(&rk_domain->dt_lock, flags);
773 * pgsize_bitmap specifies iova sizes that fit in one page table
774 * (1024 4-KiB pages = 4 MiB).
775 * So, size will always be 4096 <= size <= 4194304.
776 * Since iommu_map() guarantees that both iova and size will be
777 * aligned, we will always only be mapping from a single dte here.
779 page_table = rk_dte_get_page_table(rk_domain, iova);
780 if (IS_ERR(page_table)) {
781 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
782 return PTR_ERR(page_table);
785 dte_index = rk_domain->dt[rk_iova_dte_index(iova)];
786 pte_index = rk_iova_pte_index(iova);
787 pte_addr = &page_table[pte_index];
788 pte_dma = rk_dte_pt_address(dte_index) + pte_index * sizeof(u32);
789 ret = rk_iommu_map_iova(rk_domain, pte_addr, pte_dma, iova,
790 paddr, size, prot);
792 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
794 return ret;
797 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
798 size_t size, struct iommu_iotlb_gather *gather)
800 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
801 unsigned long flags;
802 dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
803 phys_addr_t pt_phys;
804 u32 dte;
805 u32 *pte_addr;
806 size_t unmap_size;
808 spin_lock_irqsave(&rk_domain->dt_lock, flags);
811 * pgsize_bitmap specifies iova sizes that fit in one page table
812 * (1024 4-KiB pages = 4 MiB).
813 * So, size will always be 4096 <= size <= 4194304.
814 * Since iommu_unmap() guarantees that both iova and size will be
815 * aligned, we will always only be unmapping from a single dte here.
817 dte = rk_domain->dt[rk_iova_dte_index(iova)];
818 /* Just return 0 if iova is unmapped */
819 if (!rk_dte_is_pt_valid(dte)) {
820 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
821 return 0;
824 pt_phys = rk_dte_pt_address(dte);
825 pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
826 pte_dma = pt_phys + rk_iova_pte_index(iova) * sizeof(u32);
827 unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma, size);
829 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
831 /* Shootdown iotlb entries for iova range that was just unmapped */
832 rk_iommu_zap_iova(rk_domain, iova, unmap_size);
834 return unmap_size;
837 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
839 struct rk_iommudata *data = dev_iommu_priv_get(dev);
841 return data ? data->iommu : NULL;
844 /* Must be called with iommu powered on and attached */
845 static void rk_iommu_disable(struct rk_iommu *iommu)
847 int i;
849 /* Ignore error while disabling, just keep going */
850 WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks));
851 rk_iommu_enable_stall(iommu);
852 rk_iommu_disable_paging(iommu);
853 for (i = 0; i < iommu->num_mmu; i++) {
854 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, 0);
855 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, 0);
857 rk_iommu_disable_stall(iommu);
858 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
861 /* Must be called with iommu powered on and attached */
862 static int rk_iommu_enable(struct rk_iommu *iommu)
864 struct iommu_domain *domain = iommu->domain;
865 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
866 int ret, i;
868 ret = clk_bulk_enable(iommu->num_clocks, iommu->clocks);
869 if (ret)
870 return ret;
872 ret = rk_iommu_enable_stall(iommu);
873 if (ret)
874 goto out_disable_clocks;
876 ret = rk_iommu_force_reset(iommu);
877 if (ret)
878 goto out_disable_stall;
880 for (i = 0; i < iommu->num_mmu; i++) {
881 rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
882 rk_domain->dt_dma);
883 rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
884 rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
887 ret = rk_iommu_enable_paging(iommu);
889 out_disable_stall:
890 rk_iommu_disable_stall(iommu);
891 out_disable_clocks:
892 clk_bulk_disable(iommu->num_clocks, iommu->clocks);
893 return ret;
896 static void rk_iommu_detach_device(struct iommu_domain *domain,
897 struct device *dev)
899 struct rk_iommu *iommu;
900 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
901 unsigned long flags;
902 int ret;
904 /* Allow 'virtual devices' (eg drm) to detach from domain */
905 iommu = rk_iommu_from_dev(dev);
906 if (!iommu)
907 return;
909 dev_dbg(dev, "Detaching from iommu domain\n");
911 /* iommu already detached */
912 if (iommu->domain != domain)
913 return;
915 iommu->domain = NULL;
917 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
918 list_del_init(&iommu->node);
919 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
921 ret = pm_runtime_get_if_in_use(iommu->dev);
922 WARN_ON_ONCE(ret < 0);
923 if (ret > 0) {
924 rk_iommu_disable(iommu);
925 pm_runtime_put(iommu->dev);
929 static int rk_iommu_attach_device(struct iommu_domain *domain,
930 struct device *dev)
932 struct rk_iommu *iommu;
933 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
934 unsigned long flags;
935 int ret;
938 * Allow 'virtual devices' (e.g., drm) to attach to domain.
939 * Such a device does not belong to an iommu group.
941 iommu = rk_iommu_from_dev(dev);
942 if (!iommu)
943 return 0;
945 dev_dbg(dev, "Attaching to iommu domain\n");
947 /* iommu already attached */
948 if (iommu->domain == domain)
949 return 0;
951 if (iommu->domain)
952 rk_iommu_detach_device(iommu->domain, dev);
954 iommu->domain = domain;
956 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
957 list_add_tail(&iommu->node, &rk_domain->iommus);
958 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
960 ret = pm_runtime_get_if_in_use(iommu->dev);
961 if (!ret || WARN_ON_ONCE(ret < 0))
962 return 0;
964 ret = rk_iommu_enable(iommu);
965 if (ret)
966 rk_iommu_detach_device(iommu->domain, dev);
968 pm_runtime_put(iommu->dev);
970 return ret;
973 static struct iommu_domain *rk_iommu_domain_alloc(unsigned type)
975 struct rk_iommu_domain *rk_domain;
977 if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
978 return NULL;
980 if (!dma_dev)
981 return NULL;
983 rk_domain = kzalloc(sizeof(*rk_domain), GFP_KERNEL);
984 if (!rk_domain)
985 return NULL;
987 if (type == IOMMU_DOMAIN_DMA &&
988 iommu_get_dma_cookie(&rk_domain->domain))
989 goto err_free_domain;
992 * rk32xx iommus use a 2 level pagetable.
993 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
994 * Allocate one 4 KiB page for each table.
996 rk_domain->dt = (u32 *)get_zeroed_page(GFP_KERNEL | GFP_DMA32);
997 if (!rk_domain->dt)
998 goto err_put_cookie;
1000 rk_domain->dt_dma = dma_map_single(dma_dev, rk_domain->dt,
1001 SPAGE_SIZE, DMA_TO_DEVICE);
1002 if (dma_mapping_error(dma_dev, rk_domain->dt_dma)) {
1003 dev_err(dma_dev, "DMA map error for DT\n");
1004 goto err_free_dt;
1007 rk_table_flush(rk_domain, rk_domain->dt_dma, NUM_DT_ENTRIES);
1009 spin_lock_init(&rk_domain->iommus_lock);
1010 spin_lock_init(&rk_domain->dt_lock);
1011 INIT_LIST_HEAD(&rk_domain->iommus);
1013 rk_domain->domain.geometry.aperture_start = 0;
1014 rk_domain->domain.geometry.aperture_end = DMA_BIT_MASK(32);
1015 rk_domain->domain.geometry.force_aperture = true;
1017 return &rk_domain->domain;
1019 err_free_dt:
1020 free_page((unsigned long)rk_domain->dt);
1021 err_put_cookie:
1022 if (type == IOMMU_DOMAIN_DMA)
1023 iommu_put_dma_cookie(&rk_domain->domain);
1024 err_free_domain:
1025 kfree(rk_domain);
1027 return NULL;
1030 static void rk_iommu_domain_free(struct iommu_domain *domain)
1032 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1033 int i;
1035 WARN_ON(!list_empty(&rk_domain->iommus));
1037 for (i = 0; i < NUM_DT_ENTRIES; i++) {
1038 u32 dte = rk_domain->dt[i];
1039 if (rk_dte_is_pt_valid(dte)) {
1040 phys_addr_t pt_phys = rk_dte_pt_address(dte);
1041 u32 *page_table = phys_to_virt(pt_phys);
1042 dma_unmap_single(dma_dev, pt_phys,
1043 SPAGE_SIZE, DMA_TO_DEVICE);
1044 free_page((unsigned long)page_table);
1048 dma_unmap_single(dma_dev, rk_domain->dt_dma,
1049 SPAGE_SIZE, DMA_TO_DEVICE);
1050 free_page((unsigned long)rk_domain->dt);
1052 if (domain->type == IOMMU_DOMAIN_DMA)
1053 iommu_put_dma_cookie(&rk_domain->domain);
1054 kfree(rk_domain);
1057 static struct iommu_device *rk_iommu_probe_device(struct device *dev)
1059 struct rk_iommudata *data;
1060 struct rk_iommu *iommu;
1062 data = dev_iommu_priv_get(dev);
1063 if (!data)
1064 return ERR_PTR(-ENODEV);
1066 iommu = rk_iommu_from_dev(dev);
1068 data->link = device_link_add(dev, iommu->dev,
1069 DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME);
1071 return &iommu->iommu;
1074 static void rk_iommu_release_device(struct device *dev)
1076 struct rk_iommudata *data = dev_iommu_priv_get(dev);
1078 device_link_del(data->link);
1081 static struct iommu_group *rk_iommu_device_group(struct device *dev)
1083 struct rk_iommu *iommu;
1085 iommu = rk_iommu_from_dev(dev);
1087 return iommu_group_ref_get(iommu->group);
1090 static int rk_iommu_of_xlate(struct device *dev,
1091 struct of_phandle_args *args)
1093 struct platform_device *iommu_dev;
1094 struct rk_iommudata *data;
1096 data = devm_kzalloc(dma_dev, sizeof(*data), GFP_KERNEL);
1097 if (!data)
1098 return -ENOMEM;
1100 iommu_dev = of_find_device_by_node(args->np);
1102 data->iommu = platform_get_drvdata(iommu_dev);
1103 dev_iommu_priv_set(dev, data);
1105 platform_device_put(iommu_dev);
1107 return 0;
1110 static const struct iommu_ops rk_iommu_ops = {
1111 .domain_alloc = rk_iommu_domain_alloc,
1112 .domain_free = rk_iommu_domain_free,
1113 .attach_dev = rk_iommu_attach_device,
1114 .detach_dev = rk_iommu_detach_device,
1115 .map = rk_iommu_map,
1116 .unmap = rk_iommu_unmap,
1117 .probe_device = rk_iommu_probe_device,
1118 .release_device = rk_iommu_release_device,
1119 .iova_to_phys = rk_iommu_iova_to_phys,
1120 .device_group = rk_iommu_device_group,
1121 .pgsize_bitmap = RK_IOMMU_PGSIZE_BITMAP,
1122 .of_xlate = rk_iommu_of_xlate,
1125 static int rk_iommu_probe(struct platform_device *pdev)
1127 struct device *dev = &pdev->dev;
1128 struct rk_iommu *iommu;
1129 struct resource *res;
1130 int num_res = pdev->num_resources;
1131 int err, i;
1133 iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
1134 if (!iommu)
1135 return -ENOMEM;
1137 platform_set_drvdata(pdev, iommu);
1138 iommu->dev = dev;
1139 iommu->num_mmu = 0;
1141 iommu->bases = devm_kcalloc(dev, num_res, sizeof(*iommu->bases),
1142 GFP_KERNEL);
1143 if (!iommu->bases)
1144 return -ENOMEM;
1146 for (i = 0; i < num_res; i++) {
1147 res = platform_get_resource(pdev, IORESOURCE_MEM, i);
1148 if (!res)
1149 continue;
1150 iommu->bases[i] = devm_ioremap_resource(&pdev->dev, res);
1151 if (IS_ERR(iommu->bases[i]))
1152 continue;
1153 iommu->num_mmu++;
1155 if (iommu->num_mmu == 0)
1156 return PTR_ERR(iommu->bases[0]);
1158 iommu->num_irq = platform_irq_count(pdev);
1159 if (iommu->num_irq < 0)
1160 return iommu->num_irq;
1162 iommu->reset_disabled = device_property_read_bool(dev,
1163 "rockchip,disable-mmu-reset");
1165 iommu->num_clocks = ARRAY_SIZE(rk_iommu_clocks);
1166 iommu->clocks = devm_kcalloc(iommu->dev, iommu->num_clocks,
1167 sizeof(*iommu->clocks), GFP_KERNEL);
1168 if (!iommu->clocks)
1169 return -ENOMEM;
1171 for (i = 0; i < iommu->num_clocks; ++i)
1172 iommu->clocks[i].id = rk_iommu_clocks[i];
1175 * iommu clocks should be present for all new devices and devicetrees
1176 * but there are older devicetrees without clocks out in the wild.
1177 * So clocks as optional for the time being.
1179 err = devm_clk_bulk_get(iommu->dev, iommu->num_clocks, iommu->clocks);
1180 if (err == -ENOENT)
1181 iommu->num_clocks = 0;
1182 else if (err)
1183 return err;
1185 err = clk_bulk_prepare(iommu->num_clocks, iommu->clocks);
1186 if (err)
1187 return err;
1189 iommu->group = iommu_group_alloc();
1190 if (IS_ERR(iommu->group)) {
1191 err = PTR_ERR(iommu->group);
1192 goto err_unprepare_clocks;
1195 err = iommu_device_sysfs_add(&iommu->iommu, dev, NULL, dev_name(dev));
1196 if (err)
1197 goto err_put_group;
1199 iommu_device_set_ops(&iommu->iommu, &rk_iommu_ops);
1200 iommu_device_set_fwnode(&iommu->iommu, &dev->of_node->fwnode);
1202 err = iommu_device_register(&iommu->iommu);
1203 if (err)
1204 goto err_remove_sysfs;
1207 * Use the first registered IOMMU device for domain to use with DMA
1208 * API, since a domain might not physically correspond to a single
1209 * IOMMU device..
1211 if (!dma_dev)
1212 dma_dev = &pdev->dev;
1214 bus_set_iommu(&platform_bus_type, &rk_iommu_ops);
1216 pm_runtime_enable(dev);
1218 for (i = 0; i < iommu->num_irq; i++) {
1219 int irq = platform_get_irq(pdev, i);
1221 if (irq < 0)
1222 return irq;
1224 err = devm_request_irq(iommu->dev, irq, rk_iommu_irq,
1225 IRQF_SHARED, dev_name(dev), iommu);
1226 if (err) {
1227 pm_runtime_disable(dev);
1228 goto err_remove_sysfs;
1232 return 0;
1233 err_remove_sysfs:
1234 iommu_device_sysfs_remove(&iommu->iommu);
1235 err_put_group:
1236 iommu_group_put(iommu->group);
1237 err_unprepare_clocks:
1238 clk_bulk_unprepare(iommu->num_clocks, iommu->clocks);
1239 return err;
1242 static void rk_iommu_shutdown(struct platform_device *pdev)
1244 struct rk_iommu *iommu = platform_get_drvdata(pdev);
1245 int i;
1247 for (i = 0; i < iommu->num_irq; i++) {
1248 int irq = platform_get_irq(pdev, i);
1250 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 */ }
1288 static struct platform_driver rk_iommu_driver = {
1289 .probe = rk_iommu_probe,
1290 .shutdown = rk_iommu_shutdown,
1291 .driver = {
1292 .name = "rk_iommu",
1293 .of_match_table = rk_iommu_dt_ids,
1294 .pm = &rk_iommu_pm_ops,
1295 .suppress_bind_attrs = true,
1299 static int __init rk_iommu_init(void)
1301 return platform_driver_register(&rk_iommu_driver);
1303 subsys_initcall(rk_iommu_init);