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 */
6
7 #include <asm/cacheflush.h>
8 #include <asm/pgtable.h>
9 #include <linux/compiler.h>
10 #include <linux/delay.h>
11 #include <linux/device.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>
25
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
37
38 #define DTE_ADDR_DUMMY 0xCAFEBABE
39 #define FORCE_RESET_TIMEOUT 100 /* ms */
40
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)
49
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 */
58
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)
63
64 #define NUM_DT_ENTRIES 1024
65 #define NUM_PT_ENTRIES 1024
66
67 #define SPAGE_ORDER 12
68 #define SPAGE_SIZE (1 << SPAGE_ORDER)
69
70 /*
71 * Support mapping any size that fits in one page table:
72 * 4 KiB to 4 MiB
73 */
74 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
75
76 #define IOMMU_REG_POLL_COUNT_FAST 1000
77
78 struct rk_iommu_domain {
79 struct list_head iommus;
80 u32 *dt; /* page directory table */
81 spinlock_t iommus_lock; /* lock for iommus list */
82 spinlock_t dt_lock; /* lock for modifying page directory table */
83
84 struct iommu_domain domain;
85 };
86
87 struct rk_iommu {
88 struct device *dev;
89 void __iomem *base;
90 int irq;
91 struct list_head node; /* entry in rk_iommu_domain.iommus */
92 struct iommu_domain *domain; /* domain to which iommu is attached */
93 };
94
rk_table_flush(u32 * va,unsigned int count)95 static inline void rk_table_flush(u32 *va, unsigned int count)
96 {
97 phys_addr_t pa_start = virt_to_phys(va);
98 phys_addr_t pa_end = virt_to_phys(va + count);
99 size_t size = pa_end - pa_start;
100
101 __cpuc_flush_dcache_area(va, size);
102 outer_flush_range(pa_start, pa_end);
103 }
104
to_rk_domain(struct iommu_domain * dom)105 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
106 {
107 return container_of(dom, struct rk_iommu_domain, domain);
108 }
109
110 /**
111 * Inspired by _wait_for in intel_drv.h
112 * This is NOT safe for use in interrupt context.
113 *
114 * Note that it's important that we check the condition again after having
115 * timed out, since the timeout could be due to preemption or similar and
116 * we've never had a chance to check the condition before the timeout.
117 */
118 #define rk_wait_for(COND, MS) ({ \
119 unsigned long timeout__ = jiffies + msecs_to_jiffies(MS) + 1; \
120 int ret__ = 0; \
121 while (!(COND)) { \
122 if (time_after(jiffies, timeout__)) { \
123 ret__ = (COND) ? 0 : -ETIMEDOUT; \
124 break; \
125 } \
126 usleep_range(50, 100); \
127 } \
128 ret__; \
129 })
130
131 /*
132 * The Rockchip rk3288 iommu uses a 2-level page table.
133 * The first level is the "Directory Table" (DT).
134 * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
135 * to a "Page Table".
136 * The second level is the 1024 Page Tables (PT).
137 * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
138 * a 4 KB page of physical memory.
139 *
140 * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
141 * Each iommu device has a MMU_DTE_ADDR register that contains the physical
142 * address of the start of the DT page.
143 *
144 * The structure of the page table is as follows:
145 *
146 * DT
147 * MMU_DTE_ADDR -> +-----+
148 * | |
149 * +-----+ PT
150 * | DTE | -> +-----+
151 * +-----+ | | Memory
152 * | | +-----+ Page
153 * | | | PTE | -> +-----+
154 * +-----+ +-----+ | |
155 * | | | |
156 * | | | |
157 * +-----+ | |
158 * | |
159 * | |
160 * +-----+
161 */
162
163 /*
164 * Each DTE has a PT address and a valid bit:
165 * +---------------------+-----------+-+
166 * | PT address | Reserved |V|
167 * +---------------------+-----------+-+
168 * 31:12 - PT address (PTs always starts on a 4 KB boundary)
169 * 11: 1 - Reserved
170 * 0 - 1 if PT @ PT address is valid
171 */
172 #define RK_DTE_PT_ADDRESS_MASK 0xfffff000
173 #define RK_DTE_PT_VALID BIT(0)
174
rk_dte_pt_address(u32 dte)175 static inline phys_addr_t rk_dte_pt_address(u32 dte)
176 {
177 return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
178 }
179
rk_dte_is_pt_valid(u32 dte)180 static inline bool rk_dte_is_pt_valid(u32 dte)
181 {
182 return dte & RK_DTE_PT_VALID;
183 }
184
rk_mk_dte(u32 * pt)185 static u32 rk_mk_dte(u32 *pt)
186 {
187 phys_addr_t pt_phys = virt_to_phys(pt);
188 return (pt_phys & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
189 }
190
191 /*
192 * Each PTE has a Page address, some flags and a valid bit:
193 * +---------------------+---+-------+-+
194 * | Page address |Rsv| Flags |V|
195 * +---------------------+---+-------+-+
196 * 31:12 - Page address (Pages always start on a 4 KB boundary)
197 * 11: 9 - Reserved
198 * 8: 1 - Flags
199 * 8 - Read allocate - allocate cache space on read misses
200 * 7 - Read cache - enable cache & prefetch of data
201 * 6 - Write buffer - enable delaying writes on their way to memory
202 * 5 - Write allocate - allocate cache space on write misses
203 * 4 - Write cache - different writes can be merged together
204 * 3 - Override cache attributes
205 * if 1, bits 4-8 control cache attributes
206 * if 0, the system bus defaults are used
207 * 2 - Writable
208 * 1 - Readable
209 * 0 - 1 if Page @ Page address is valid
210 */
211 #define RK_PTE_PAGE_ADDRESS_MASK 0xfffff000
212 #define RK_PTE_PAGE_FLAGS_MASK 0x000001fe
213 #define RK_PTE_PAGE_WRITABLE BIT(2)
214 #define RK_PTE_PAGE_READABLE BIT(1)
215 #define RK_PTE_PAGE_VALID BIT(0)
216
rk_pte_page_address(u32 pte)217 static inline phys_addr_t rk_pte_page_address(u32 pte)
218 {
219 return (phys_addr_t)pte & RK_PTE_PAGE_ADDRESS_MASK;
220 }
221
rk_pte_is_page_valid(u32 pte)222 static inline bool rk_pte_is_page_valid(u32 pte)
223 {
224 return pte & RK_PTE_PAGE_VALID;
225 }
226
227 /* TODO: set cache flags per prot IOMMU_CACHE */
rk_mk_pte(phys_addr_t page,int prot)228 static u32 rk_mk_pte(phys_addr_t page, int prot)
229 {
230 u32 flags = 0;
231 flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
232 flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
233 page &= RK_PTE_PAGE_ADDRESS_MASK;
234 return page | flags | RK_PTE_PAGE_VALID;
235 }
236
rk_mk_pte_invalid(u32 pte)237 static u32 rk_mk_pte_invalid(u32 pte)
238 {
239 return pte & ~RK_PTE_PAGE_VALID;
240 }
241
242 /*
243 * rk3288 iova (IOMMU Virtual Address) format
244 * 31 22.21 12.11 0
245 * +-----------+-----------+-------------+
246 * | DTE index | PTE index | Page offset |
247 * +-----------+-----------+-------------+
248 * 31:22 - DTE index - index of DTE in DT
249 * 21:12 - PTE index - index of PTE in PT @ DTE.pt_address
250 * 11: 0 - Page offset - offset into page @ PTE.page_address
251 */
252 #define RK_IOVA_DTE_MASK 0xffc00000
253 #define RK_IOVA_DTE_SHIFT 22
254 #define RK_IOVA_PTE_MASK 0x003ff000
255 #define RK_IOVA_PTE_SHIFT 12
256 #define RK_IOVA_PAGE_MASK 0x00000fff
257 #define RK_IOVA_PAGE_SHIFT 0
258
rk_iova_dte_index(dma_addr_t iova)259 static u32 rk_iova_dte_index(dma_addr_t iova)
260 {
261 return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
262 }
263
rk_iova_pte_index(dma_addr_t iova)264 static u32 rk_iova_pte_index(dma_addr_t iova)
265 {
266 return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
267 }
268
rk_iova_page_offset(dma_addr_t iova)269 static u32 rk_iova_page_offset(dma_addr_t iova)
270 {
271 return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
272 }
273
rk_iommu_read(struct rk_iommu * iommu,u32 offset)274 static u32 rk_iommu_read(struct rk_iommu *iommu, u32 offset)
275 {
276 return readl(iommu->base + offset);
277 }
278
rk_iommu_write(struct rk_iommu * iommu,u32 offset,u32 value)279 static void rk_iommu_write(struct rk_iommu *iommu, u32 offset, u32 value)
280 {
281 writel(value, iommu->base + offset);
282 }
283
rk_iommu_command(struct rk_iommu * iommu,u32 command)284 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
285 {
286 writel(command, iommu->base + RK_MMU_COMMAND);
287 }
288
rk_iommu_zap_lines(struct rk_iommu * iommu,dma_addr_t iova,size_t size)289 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova,
290 size_t size)
291 {
292 dma_addr_t iova_end = iova + size;
293 /*
294 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
295 * entire iotlb rather than iterate over individual iovas.
296 */
297 for (; iova < iova_end; iova += SPAGE_SIZE)
298 rk_iommu_write(iommu, RK_MMU_ZAP_ONE_LINE, iova);
299 }
300
rk_iommu_is_stall_active(struct rk_iommu * iommu)301 static bool rk_iommu_is_stall_active(struct rk_iommu *iommu)
302 {
303 return rk_iommu_read(iommu, RK_MMU_STATUS) & RK_MMU_STATUS_STALL_ACTIVE;
304 }
305
rk_iommu_is_paging_enabled(struct rk_iommu * iommu)306 static bool rk_iommu_is_paging_enabled(struct rk_iommu *iommu)
307 {
308 return rk_iommu_read(iommu, RK_MMU_STATUS) &
309 RK_MMU_STATUS_PAGING_ENABLED;
310 }
311
rk_iommu_enable_stall(struct rk_iommu * iommu)312 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
313 {
314 int ret;
315
316 if (rk_iommu_is_stall_active(iommu))
317 return 0;
318
319 /* Stall can only be enabled if paging is enabled */
320 if (!rk_iommu_is_paging_enabled(iommu))
321 return 0;
322
323 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
324
325 ret = rk_wait_for(rk_iommu_is_stall_active(iommu), 1);
326 if (ret)
327 dev_err(iommu->dev, "Enable stall request timed out, status: %#08x\n",
328 rk_iommu_read(iommu, RK_MMU_STATUS));
329
330 return ret;
331 }
332
rk_iommu_disable_stall(struct rk_iommu * iommu)333 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
334 {
335 int ret;
336
337 if (!rk_iommu_is_stall_active(iommu))
338 return 0;
339
340 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
341
342 ret = rk_wait_for(!rk_iommu_is_stall_active(iommu), 1);
343 if (ret)
344 dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
345 rk_iommu_read(iommu, RK_MMU_STATUS));
346
347 return ret;
348 }
349
rk_iommu_enable_paging(struct rk_iommu * iommu)350 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
351 {
352 int ret;
353
354 if (rk_iommu_is_paging_enabled(iommu))
355 return 0;
356
357 rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
358
359 ret = rk_wait_for(rk_iommu_is_paging_enabled(iommu), 1);
360 if (ret)
361 dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
362 rk_iommu_read(iommu, RK_MMU_STATUS));
363
364 return ret;
365 }
366
rk_iommu_disable_paging(struct rk_iommu * iommu)367 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
368 {
369 int ret;
370
371 if (!rk_iommu_is_paging_enabled(iommu))
372 return 0;
373
374 rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
375
376 ret = rk_wait_for(!rk_iommu_is_paging_enabled(iommu), 1);
377 if (ret)
378 dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
379 rk_iommu_read(iommu, RK_MMU_STATUS));
380
381 return ret;
382 }
383
rk_iommu_force_reset(struct rk_iommu * iommu)384 static int rk_iommu_force_reset(struct rk_iommu *iommu)
385 {
386 int ret;
387 u32 dte_addr;
388
389 /*
390 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
391 * and verifying that upper 5 nybbles are read back.
392 */
393 rk_iommu_write(iommu, RK_MMU_DTE_ADDR, DTE_ADDR_DUMMY);
394
395 dte_addr = rk_iommu_read(iommu, RK_MMU_DTE_ADDR);
396 if (dte_addr != (DTE_ADDR_DUMMY & RK_DTE_PT_ADDRESS_MASK)) {
397 dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
398 return -EFAULT;
399 }
400
401 rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
402
403 ret = rk_wait_for(rk_iommu_read(iommu, RK_MMU_DTE_ADDR) == 0x00000000,
404 FORCE_RESET_TIMEOUT);
405 if (ret)
406 dev_err(iommu->dev, "FORCE_RESET command timed out\n");
407
408 return ret;
409 }
410
log_iova(struct rk_iommu * iommu,dma_addr_t iova)411 static void log_iova(struct rk_iommu *iommu, dma_addr_t iova)
412 {
413 u32 dte_index, pte_index, page_offset;
414 u32 mmu_dte_addr;
415 phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
416 u32 *dte_addr;
417 u32 dte;
418 phys_addr_t pte_addr_phys = 0;
419 u32 *pte_addr = NULL;
420 u32 pte = 0;
421 phys_addr_t page_addr_phys = 0;
422 u32 page_flags = 0;
423
424 dte_index = rk_iova_dte_index(iova);
425 pte_index = rk_iova_pte_index(iova);
426 page_offset = rk_iova_page_offset(iova);
427
428 mmu_dte_addr = rk_iommu_read(iommu, RK_MMU_DTE_ADDR);
429 mmu_dte_addr_phys = (phys_addr_t)mmu_dte_addr;
430
431 dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
432 dte_addr = phys_to_virt(dte_addr_phys);
433 dte = *dte_addr;
434
435 if (!rk_dte_is_pt_valid(dte))
436 goto print_it;
437
438 pte_addr_phys = rk_dte_pt_address(dte) + (pte_index * 4);
439 pte_addr = phys_to_virt(pte_addr_phys);
440 pte = *pte_addr;
441
442 if (!rk_pte_is_page_valid(pte))
443 goto print_it;
444
445 page_addr_phys = rk_pte_page_address(pte) + page_offset;
446 page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
447
448 print_it:
449 dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
450 &iova, dte_index, pte_index, page_offset);
451 dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
452 &mmu_dte_addr_phys, &dte_addr_phys, dte,
453 rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
454 rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
455 }
456
rk_iommu_irq(int irq,void * dev_id)457 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
458 {
459 struct rk_iommu *iommu = dev_id;
460 u32 status;
461 u32 int_status;
462 dma_addr_t iova;
463
464 int_status = rk_iommu_read(iommu, RK_MMU_INT_STATUS);
465 if (int_status == 0)
466 return IRQ_NONE;
467
468 iova = rk_iommu_read(iommu, RK_MMU_PAGE_FAULT_ADDR);
469
470 if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
471 int flags;
472
473 status = rk_iommu_read(iommu, RK_MMU_STATUS);
474 flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
475 IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
476
477 dev_err(iommu->dev, "Page fault at %pad of type %s\n",
478 &iova,
479 (flags == IOMMU_FAULT_WRITE) ? "write" : "read");
480
481 log_iova(iommu, iova);
482
483 /*
484 * Report page fault to any installed handlers.
485 * Ignore the return code, though, since we always zap cache
486 * and clear the page fault anyway.
487 */
488 if (iommu->domain)
489 report_iommu_fault(iommu->domain, iommu->dev, iova,
490 flags);
491 else
492 dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
493
494 rk_iommu_command(iommu, RK_MMU_CMD_ZAP_CACHE);
495 rk_iommu_command(iommu, RK_MMU_CMD_PAGE_FAULT_DONE);
496 }
497
498 if (int_status & RK_MMU_IRQ_BUS_ERROR)
499 dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
500
501 if (int_status & ~RK_MMU_IRQ_MASK)
502 dev_err(iommu->dev, "unexpected int_status: %#08x\n",
503 int_status);
504
505 rk_iommu_write(iommu, RK_MMU_INT_CLEAR, int_status);
506
507 return IRQ_HANDLED;
508 }
509
rk_iommu_iova_to_phys(struct iommu_domain * domain,dma_addr_t iova)510 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
511 dma_addr_t iova)
512 {
513 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
514 unsigned long flags;
515 phys_addr_t pt_phys, phys = 0;
516 u32 dte, pte;
517 u32 *page_table;
518
519 spin_lock_irqsave(&rk_domain->dt_lock, flags);
520
521 dte = rk_domain->dt[rk_iova_dte_index(iova)];
522 if (!rk_dte_is_pt_valid(dte))
523 goto out;
524
525 pt_phys = rk_dte_pt_address(dte);
526 page_table = (u32 *)phys_to_virt(pt_phys);
527 pte = page_table[rk_iova_pte_index(iova)];
528 if (!rk_pte_is_page_valid(pte))
529 goto out;
530
531 phys = rk_pte_page_address(pte) + rk_iova_page_offset(iova);
532 out:
533 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
534
535 return phys;
536 }
537
rk_iommu_zap_iova(struct rk_iommu_domain * rk_domain,dma_addr_t iova,size_t size)538 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
539 dma_addr_t iova, size_t size)
540 {
541 struct list_head *pos;
542 unsigned long flags;
543
544 /* shootdown these iova from all iommus using this domain */
545 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
546 list_for_each(pos, &rk_domain->iommus) {
547 struct rk_iommu *iommu;
548 iommu = list_entry(pos, struct rk_iommu, node);
549 rk_iommu_zap_lines(iommu, iova, size);
550 }
551 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
552 }
553
rk_iommu_zap_iova_first_last(struct rk_iommu_domain * rk_domain,dma_addr_t iova,size_t size)554 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
555 dma_addr_t iova, size_t size)
556 {
557 rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
558 if (size > SPAGE_SIZE)
559 rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
560 SPAGE_SIZE);
561 }
562
rk_dte_get_page_table(struct rk_iommu_domain * rk_domain,dma_addr_t iova)563 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
564 dma_addr_t iova)
565 {
566 u32 *page_table, *dte_addr;
567 u32 dte;
568 phys_addr_t pt_phys;
569
570 assert_spin_locked(&rk_domain->dt_lock);
571
572 dte_addr = &rk_domain->dt[rk_iova_dte_index(iova)];
573 dte = *dte_addr;
574 if (rk_dte_is_pt_valid(dte))
575 goto done;
576
577 page_table = (u32 *)get_zeroed_page(GFP_ATOMIC | GFP_DMA32);
578 if (!page_table)
579 return ERR_PTR(-ENOMEM);
580
581 dte = rk_mk_dte(page_table);
582 *dte_addr = dte;
583
584 rk_table_flush(page_table, NUM_PT_ENTRIES);
585 rk_table_flush(dte_addr, 1);
586
587 done:
588 pt_phys = rk_dte_pt_address(dte);
589 return (u32 *)phys_to_virt(pt_phys);
590 }
591
rk_iommu_unmap_iova(struct rk_iommu_domain * rk_domain,u32 * pte_addr,dma_addr_t iova,size_t size)592 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
593 u32 *pte_addr, dma_addr_t iova, size_t size)
594 {
595 unsigned int pte_count;
596 unsigned int pte_total = size / SPAGE_SIZE;
597
598 assert_spin_locked(&rk_domain->dt_lock);
599
600 for (pte_count = 0; pte_count < pte_total; pte_count++) {
601 u32 pte = pte_addr[pte_count];
602 if (!rk_pte_is_page_valid(pte))
603 break;
604
605 pte_addr[pte_count] = rk_mk_pte_invalid(pte);
606 }
607
608 rk_table_flush(pte_addr, pte_count);
609
610 return pte_count * SPAGE_SIZE;
611 }
612
rk_iommu_map_iova(struct rk_iommu_domain * rk_domain,u32 * pte_addr,dma_addr_t iova,phys_addr_t paddr,size_t size,int prot)613 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
614 dma_addr_t iova, phys_addr_t paddr, size_t size,
615 int prot)
616 {
617 unsigned int pte_count;
618 unsigned int pte_total = size / SPAGE_SIZE;
619 phys_addr_t page_phys;
620
621 assert_spin_locked(&rk_domain->dt_lock);
622
623 for (pte_count = 0; pte_count < pte_total; pte_count++) {
624 u32 pte = pte_addr[pte_count];
625
626 if (rk_pte_is_page_valid(pte))
627 goto unwind;
628
629 pte_addr[pte_count] = rk_mk_pte(paddr, prot);
630
631 paddr += SPAGE_SIZE;
632 }
633
634 rk_table_flush(pte_addr, pte_count);
635
636 /*
637 * Zap the first and last iova to evict from iotlb any previously
638 * mapped cachelines holding stale values for its dte and pte.
639 * We only zap the first and last iova, since only they could have
640 * dte or pte shared with an existing mapping.
641 */
642 rk_iommu_zap_iova_first_last(rk_domain, iova, size);
643
644 return 0;
645 unwind:
646 /* Unmap the range of iovas that we just mapped */
647 rk_iommu_unmap_iova(rk_domain, pte_addr, iova, pte_count * SPAGE_SIZE);
648
649 iova += pte_count * SPAGE_SIZE;
650 page_phys = rk_pte_page_address(pte_addr[pte_count]);
651 pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
652 &iova, &page_phys, &paddr, prot);
653
654 return -EADDRINUSE;
655 }
656
rk_iommu_map(struct iommu_domain * domain,unsigned long _iova,phys_addr_t paddr,size_t size,int prot)657 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
658 phys_addr_t paddr, size_t size, int prot)
659 {
660 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
661 unsigned long flags;
662 dma_addr_t iova = (dma_addr_t)_iova;
663 u32 *page_table, *pte_addr;
664 int ret;
665
666 spin_lock_irqsave(&rk_domain->dt_lock, flags);
667
668 /*
669 * pgsize_bitmap specifies iova sizes that fit in one page table
670 * (1024 4-KiB pages = 4 MiB).
671 * So, size will always be 4096 <= size <= 4194304.
672 * Since iommu_map() guarantees that both iova and size will be
673 * aligned, we will always only be mapping from a single dte here.
674 */
675 page_table = rk_dte_get_page_table(rk_domain, iova);
676 if (IS_ERR(page_table)) {
677 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
678 return PTR_ERR(page_table);
679 }
680
681 pte_addr = &page_table[rk_iova_pte_index(iova)];
682 ret = rk_iommu_map_iova(rk_domain, pte_addr, iova, paddr, size, prot);
683 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
684
685 return ret;
686 }
687
rk_iommu_unmap(struct iommu_domain * domain,unsigned long _iova,size_t size)688 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
689 size_t size)
690 {
691 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
692 unsigned long flags;
693 dma_addr_t iova = (dma_addr_t)_iova;
694 phys_addr_t pt_phys;
695 u32 dte;
696 u32 *pte_addr;
697 size_t unmap_size;
698
699 spin_lock_irqsave(&rk_domain->dt_lock, flags);
700
701 /*
702 * pgsize_bitmap specifies iova sizes that fit in one page table
703 * (1024 4-KiB pages = 4 MiB).
704 * So, size will always be 4096 <= size <= 4194304.
705 * Since iommu_unmap() guarantees that both iova and size will be
706 * aligned, we will always only be unmapping from a single dte here.
707 */
708 dte = rk_domain->dt[rk_iova_dte_index(iova)];
709 /* Just return 0 if iova is unmapped */
710 if (!rk_dte_is_pt_valid(dte)) {
711 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
712 return 0;
713 }
714
715 pt_phys = rk_dte_pt_address(dte);
716 pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
717 unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, iova, size);
718
719 spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
720
721 /* Shootdown iotlb entries for iova range that was just unmapped */
722 rk_iommu_zap_iova(rk_domain, iova, unmap_size);
723
724 return unmap_size;
725 }
726
rk_iommu_from_dev(struct device * dev)727 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
728 {
729 struct iommu_group *group;
730 struct device *iommu_dev;
731 struct rk_iommu *rk_iommu;
732
733 group = iommu_group_get(dev);
734 if (!group)
735 return NULL;
736 iommu_dev = iommu_group_get_iommudata(group);
737 rk_iommu = dev_get_drvdata(iommu_dev);
738 iommu_group_put(group);
739
740 return rk_iommu;
741 }
742
rk_iommu_attach_device(struct iommu_domain * domain,struct device * dev)743 static int rk_iommu_attach_device(struct iommu_domain *domain,
744 struct device *dev)
745 {
746 struct rk_iommu *iommu;
747 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
748 unsigned long flags;
749 int ret;
750 phys_addr_t dte_addr;
751
752 /*
753 * Allow 'virtual devices' (e.g., drm) to attach to domain.
754 * Such a device does not belong to an iommu group.
755 */
756 iommu = rk_iommu_from_dev(dev);
757 if (!iommu)
758 return 0;
759
760 ret = rk_iommu_enable_stall(iommu);
761 if (ret)
762 return ret;
763
764 ret = rk_iommu_force_reset(iommu);
765 if (ret)
766 return ret;
767
768 iommu->domain = domain;
769
770 ret = devm_request_irq(dev, iommu->irq, rk_iommu_irq,
771 IRQF_SHARED, dev_name(dev), iommu);
772 if (ret)
773 return ret;
774
775 dte_addr = virt_to_phys(rk_domain->dt);
776 rk_iommu_write(iommu, RK_MMU_DTE_ADDR, dte_addr);
777 rk_iommu_command(iommu, RK_MMU_CMD_ZAP_CACHE);
778 rk_iommu_write(iommu, RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
779
780 ret = rk_iommu_enable_paging(iommu);
781 if (ret)
782 return ret;
783
784 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
785 list_add_tail(&iommu->node, &rk_domain->iommus);
786 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
787
788 dev_dbg(dev, "Attached to iommu domain\n");
789
790 rk_iommu_disable_stall(iommu);
791
792 return 0;
793 }
794
rk_iommu_detach_device(struct iommu_domain * domain,struct device * dev)795 static void rk_iommu_detach_device(struct iommu_domain *domain,
796 struct device *dev)
797 {
798 struct rk_iommu *iommu;
799 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
800 unsigned long flags;
801
802 /* Allow 'virtual devices' (eg drm) to detach from domain */
803 iommu = rk_iommu_from_dev(dev);
804 if (!iommu)
805 return;
806
807 spin_lock_irqsave(&rk_domain->iommus_lock, flags);
808 list_del_init(&iommu->node);
809 spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
810
811 /* Ignore error while disabling, just keep going */
812 rk_iommu_enable_stall(iommu);
813 rk_iommu_disable_paging(iommu);
814 rk_iommu_write(iommu, RK_MMU_INT_MASK, 0);
815 rk_iommu_write(iommu, RK_MMU_DTE_ADDR, 0);
816 rk_iommu_disable_stall(iommu);
817
818 devm_free_irq(dev, iommu->irq, iommu);
819
820 iommu->domain = NULL;
821
822 dev_dbg(dev, "Detached from iommu domain\n");
823 }
824
rk_iommu_domain_alloc(unsigned type)825 static struct iommu_domain *rk_iommu_domain_alloc(unsigned type)
826 {
827 struct rk_iommu_domain *rk_domain;
828
829 if (type != IOMMU_DOMAIN_UNMANAGED)
830 return NULL;
831
832 rk_domain = kzalloc(sizeof(*rk_domain), GFP_KERNEL);
833 if (!rk_domain)
834 return NULL;
835
836 /*
837 * rk32xx iommus use a 2 level pagetable.
838 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
839 * Allocate one 4 KiB page for each table.
840 */
841 rk_domain->dt = (u32 *)get_zeroed_page(GFP_KERNEL | GFP_DMA32);
842 if (!rk_domain->dt)
843 goto err_dt;
844
845 rk_table_flush(rk_domain->dt, NUM_DT_ENTRIES);
846
847 spin_lock_init(&rk_domain->iommus_lock);
848 spin_lock_init(&rk_domain->dt_lock);
849 INIT_LIST_HEAD(&rk_domain->iommus);
850
851 return &rk_domain->domain;
852
853 err_dt:
854 kfree(rk_domain);
855 return NULL;
856 }
857
rk_iommu_domain_free(struct iommu_domain * domain)858 static void rk_iommu_domain_free(struct iommu_domain *domain)
859 {
860 struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
861 int i;
862
863 WARN_ON(!list_empty(&rk_domain->iommus));
864
865 for (i = 0; i < NUM_DT_ENTRIES; i++) {
866 u32 dte = rk_domain->dt[i];
867 if (rk_dte_is_pt_valid(dte)) {
868 phys_addr_t pt_phys = rk_dte_pt_address(dte);
869 u32 *page_table = phys_to_virt(pt_phys);
870 free_page((unsigned long)page_table);
871 }
872 }
873
874 free_page((unsigned long)rk_domain->dt);
875 kfree(rk_domain);
876 }
877
rk_iommu_is_dev_iommu_master(struct device * dev)878 static bool rk_iommu_is_dev_iommu_master(struct device *dev)
879 {
880 struct device_node *np = dev->of_node;
881 int ret;
882
883 /*
884 * An iommu master has an iommus property containing a list of phandles
885 * to iommu nodes, each with an #iommu-cells property with value 0.
886 */
887 ret = of_count_phandle_with_args(np, "iommus", "#iommu-cells");
888 return (ret > 0);
889 }
890
rk_iommu_group_set_iommudata(struct iommu_group * group,struct device * dev)891 static int rk_iommu_group_set_iommudata(struct iommu_group *group,
892 struct device *dev)
893 {
894 struct device_node *np = dev->of_node;
895 struct platform_device *pd;
896 int ret;
897 struct of_phandle_args args;
898
899 /*
900 * An iommu master has an iommus property containing a list of phandles
901 * to iommu nodes, each with an #iommu-cells property with value 0.
902 */
903 ret = of_parse_phandle_with_args(np, "iommus", "#iommu-cells", 0,
904 &args);
905 if (ret) {
906 dev_err(dev, "of_parse_phandle_with_args(%s) => %d\n",
907 np->full_name, ret);
908 return ret;
909 }
910 if (args.args_count != 0) {
911 dev_err(dev, "incorrect number of iommu params found for %s (found %d, expected 0)\n",
912 args.np->full_name, args.args_count);
913 return -EINVAL;
914 }
915
916 pd = of_find_device_by_node(args.np);
917 of_node_put(args.np);
918 if (!pd) {
919 dev_err(dev, "iommu %s not found\n", args.np->full_name);
920 return -EPROBE_DEFER;
921 }
922
923 /* TODO(djkurtz): handle multiple slave iommus for a single master */
924 iommu_group_set_iommudata(group, &pd->dev, NULL);
925
926 return 0;
927 }
928
rk_iommu_add_device(struct device * dev)929 static int rk_iommu_add_device(struct device *dev)
930 {
931 struct iommu_group *group;
932 int ret;
933
934 if (!rk_iommu_is_dev_iommu_master(dev))
935 return -ENODEV;
936
937 group = iommu_group_get(dev);
938 if (!group) {
939 group = iommu_group_alloc();
940 if (IS_ERR(group)) {
941 dev_err(dev, "Failed to allocate IOMMU group\n");
942 return PTR_ERR(group);
943 }
944 }
945
946 ret = iommu_group_add_device(group, dev);
947 if (ret)
948 goto err_put_group;
949
950 ret = rk_iommu_group_set_iommudata(group, dev);
951 if (ret)
952 goto err_remove_device;
953
954 iommu_group_put(group);
955
956 return 0;
957
958 err_remove_device:
959 iommu_group_remove_device(dev);
960 err_put_group:
961 iommu_group_put(group);
962 return ret;
963 }
964
rk_iommu_remove_device(struct device * dev)965 static void rk_iommu_remove_device(struct device *dev)
966 {
967 if (!rk_iommu_is_dev_iommu_master(dev))
968 return;
969
970 iommu_group_remove_device(dev);
971 }
972
973 static const struct iommu_ops rk_iommu_ops = {
974 .domain_alloc = rk_iommu_domain_alloc,
975 .domain_free = rk_iommu_domain_free,
976 .attach_dev = rk_iommu_attach_device,
977 .detach_dev = rk_iommu_detach_device,
978 .map = rk_iommu_map,
979 .unmap = rk_iommu_unmap,
980 .add_device = rk_iommu_add_device,
981 .remove_device = rk_iommu_remove_device,
982 .iova_to_phys = rk_iommu_iova_to_phys,
983 .pgsize_bitmap = RK_IOMMU_PGSIZE_BITMAP,
984 };
985
rk_iommu_probe(struct platform_device * pdev)986 static int rk_iommu_probe(struct platform_device *pdev)
987 {
988 struct device *dev = &pdev->dev;
989 struct rk_iommu *iommu;
990 struct resource *res;
991
992 iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
993 if (!iommu)
994 return -ENOMEM;
995
996 platform_set_drvdata(pdev, iommu);
997 iommu->dev = dev;
998
999 res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1000 iommu->base = devm_ioremap_resource(&pdev->dev, res);
1001 if (IS_ERR(iommu->base))
1002 return PTR_ERR(iommu->base);
1003
1004 iommu->irq = platform_get_irq(pdev, 0);
1005 if (iommu->irq < 0) {
1006 dev_err(dev, "Failed to get IRQ, %d\n", iommu->irq);
1007 return -ENXIO;
1008 }
1009
1010 return 0;
1011 }
1012
rk_iommu_remove(struct platform_device * pdev)1013 static int rk_iommu_remove(struct platform_device *pdev)
1014 {
1015 return 0;
1016 }
1017
1018 static const struct of_device_id rk_iommu_dt_ids[] = {
1019 { .compatible = "rockchip,iommu" },
1020 { /* sentinel */ }
1021 };
1022 MODULE_DEVICE_TABLE(of, rk_iommu_dt_ids);
1023
1024 static struct platform_driver rk_iommu_driver = {
1025 .probe = rk_iommu_probe,
1026 .remove = rk_iommu_remove,
1027 .driver = {
1028 .name = "rk_iommu",
1029 .of_match_table = rk_iommu_dt_ids,
1030 },
1031 };
1032
rk_iommu_init(void)1033 static int __init rk_iommu_init(void)
1034 {
1035 struct device_node *np;
1036 int ret;
1037
1038 np = of_find_matching_node(NULL, rk_iommu_dt_ids);
1039 if (!np)
1040 return 0;
1041
1042 of_node_put(np);
1043
1044 ret = bus_set_iommu(&platform_bus_type, &rk_iommu_ops);
1045 if (ret)
1046 return ret;
1047
1048 return platform_driver_register(&rk_iommu_driver);
1049 }
rk_iommu_exit(void)1050 static void __exit rk_iommu_exit(void)
1051 {
1052 platform_driver_unregister(&rk_iommu_driver);
1053 }
1054
1055 subsys_initcall(rk_iommu_init);
1056 module_exit(rk_iommu_exit);
1057
1058 MODULE_DESCRIPTION("IOMMU API for Rockchip");
1059 MODULE_AUTHOR("Simon Xue <xxm@rock-chips.com> and Daniel Kurtz <djkurtz@chromium.org>");
1060 MODULE_ALIAS("platform:rockchip-iommu");
1061 MODULE_LICENSE("GPL v2");
1062