1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * IOMMU API for Rockchip
4  *
5  * Module Authors:	Simon Xue <xxm@rock-chips.com>
6  *			Daniel Kurtz <djkurtz@chromium.org>
7  */
8 
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_platform.h>
25 #include <linux/platform_device.h>
26 #include <linux/pm_runtime.h>
27 #include <linux/slab.h>
28 #include <linux/spinlock.h>
29 
30 /** MMU register offsets */
31 #define RK_MMU_DTE_ADDR		0x00	/* Directory table address */
32 #define RK_MMU_STATUS		0x04
33 #define RK_MMU_COMMAND		0x08
34 #define RK_MMU_PAGE_FAULT_ADDR	0x0C	/* IOVA of last page fault */
35 #define RK_MMU_ZAP_ONE_LINE	0x10	/* Shootdown one IOTLB entry */
36 #define RK_MMU_INT_RAWSTAT	0x14	/* IRQ status ignoring mask */
37 #define RK_MMU_INT_CLEAR	0x18	/* Acknowledge and re-arm irq */
38 #define RK_MMU_INT_MASK		0x1C	/* IRQ enable */
39 #define RK_MMU_INT_STATUS	0x20	/* IRQ status after masking */
40 #define RK_MMU_AUTO_GATING	0x24
41 
42 #define DTE_ADDR_DUMMY		0xCAFEBABE
43 
44 #define RK_MMU_POLL_PERIOD_US		100
45 #define RK_MMU_FORCE_RESET_TIMEOUT_US	100000
46 #define RK_MMU_POLL_TIMEOUT_US		1000
47 
48 /* RK_MMU_STATUS fields */
49 #define RK_MMU_STATUS_PAGING_ENABLED       BIT(0)
50 #define RK_MMU_STATUS_PAGE_FAULT_ACTIVE    BIT(1)
51 #define RK_MMU_STATUS_STALL_ACTIVE         BIT(2)
52 #define RK_MMU_STATUS_IDLE                 BIT(3)
53 #define RK_MMU_STATUS_REPLAY_BUFFER_EMPTY  BIT(4)
54 #define RK_MMU_STATUS_PAGE_FAULT_IS_WRITE  BIT(5)
55 #define RK_MMU_STATUS_STALL_NOT_ACTIVE     BIT(31)
56 
57 /* RK_MMU_COMMAND command values */
58 #define RK_MMU_CMD_ENABLE_PAGING    0  /* Enable memory translation */
59 #define RK_MMU_CMD_DISABLE_PAGING   1  /* Disable memory translation */
60 #define RK_MMU_CMD_ENABLE_STALL     2  /* Stall paging to allow other cmds */
61 #define RK_MMU_CMD_DISABLE_STALL    3  /* Stop stall re-enables paging */
62 #define RK_MMU_CMD_ZAP_CACHE        4  /* Shoot down entire IOTLB */
63 #define RK_MMU_CMD_PAGE_FAULT_DONE  5  /* Clear page fault */
64 #define RK_MMU_CMD_FORCE_RESET      6  /* Reset all registers */
65 
66 /* RK_MMU_INT_* register fields */
67 #define RK_MMU_IRQ_PAGE_FAULT    0x01  /* page fault */
68 #define RK_MMU_IRQ_BUS_ERROR     0x02  /* bus read error */
69 #define RK_MMU_IRQ_MASK          (RK_MMU_IRQ_PAGE_FAULT | RK_MMU_IRQ_BUS_ERROR)
70 
71 #define NUM_DT_ENTRIES 1024
72 #define NUM_PT_ENTRIES 1024
73 
74 #define SPAGE_ORDER 12
75 #define SPAGE_SIZE (1 << SPAGE_ORDER)
76 
77  /*
78   * Support mapping any size that fits in one page table:
79   *   4 KiB to 4 MiB
80   */
81 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
82 
83 struct rk_iommu_domain {
84 	struct list_head iommus;
85 	u32 *dt; /* page directory table */
86 	dma_addr_t dt_dma;
87 	spinlock_t iommus_lock; /* lock for iommus list */
88 	spinlock_t dt_lock; /* lock for modifying page directory table */
89 
90 	struct iommu_domain domain;
91 };
92 
93 /* list of clocks required by IOMMU */
94 static const char * const rk_iommu_clocks[] = {
95 	"aclk", "iface",
96 };
97 
98 struct rk_iommu_ops {
99 	phys_addr_t (*pt_address)(u32 dte);
100 	u32 (*mk_dtentries)(dma_addr_t pt_dma);
101 	u32 (*mk_ptentries)(phys_addr_t page, int prot);
102 	phys_addr_t (*dte_addr_phys)(u32 addr);
103 	u32 (*dma_addr_dte)(dma_addr_t dt_dma);
104 	u64 dma_bit_mask;
105 };
106 
107 struct rk_iommu {
108 	struct device *dev;
109 	void __iomem **bases;
110 	int num_mmu;
111 	int num_irq;
112 	struct clk_bulk_data *clocks;
113 	int num_clocks;
114 	bool reset_disabled;
115 	struct iommu_device iommu;
116 	struct list_head node; /* entry in rk_iommu_domain.iommus */
117 	struct iommu_domain *domain; /* domain to which iommu is attached */
118 	struct iommu_group *group;
119 };
120 
121 struct rk_iommudata {
122 	struct device_link *link; /* runtime PM link from IOMMU to master */
123 	struct rk_iommu *iommu;
124 };
125 
126 static struct device *dma_dev;
127 static const struct rk_iommu_ops *rk_ops;
128 
129 static inline void rk_table_flush(struct rk_iommu_domain *dom, dma_addr_t dma,
130 				  unsigned int count)
131 {
132 	size_t size = count * sizeof(u32); /* count of u32 entry */
133 
134 	dma_sync_single_for_device(dma_dev, dma, size, DMA_TO_DEVICE);
135 }
136 
137 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
138 {
139 	return container_of(dom, struct rk_iommu_domain, domain);
140 }
141 
142 /*
143  * The Rockchip rk3288 iommu uses a 2-level page table.
144  * The first level is the "Directory Table" (DT).
145  * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
146  * to a "Page Table".
147  * The second level is the 1024 Page Tables (PT).
148  * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
149  * a 4 KB page of physical memory.
150  *
151  * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
152  * Each iommu device has a MMU_DTE_ADDR register that contains the physical
153  * address of the start of the DT page.
154  *
155  * The structure of the page table is as follows:
156  *
157  *                   DT
158  * MMU_DTE_ADDR -> +-----+
159  *                 |     |
160  *                 +-----+     PT
161  *                 | DTE | -> +-----+
162  *                 +-----+    |     |     Memory
163  *                 |     |    +-----+     Page
164  *                 |     |    | PTE | -> +-----+
165  *                 +-----+    +-----+    |     |
166  *                            |     |    |     |
167  *                            |     |    |     |
168  *                            +-----+    |     |
169  *                                       |     |
170  *                                       |     |
171  *                                       +-----+
172  */
173 
174 /*
175  * Each DTE has a PT address and a valid bit:
176  * +---------------------+-----------+-+
177  * | PT address          | Reserved  |V|
178  * +---------------------+-----------+-+
179  *  31:12 - PT address (PTs always starts on a 4 KB boundary)
180  *  11: 1 - Reserved
181  *      0 - 1 if PT @ PT address is valid
182  */
183 #define RK_DTE_PT_ADDRESS_MASK    0xfffff000
184 #define RK_DTE_PT_VALID           BIT(0)
185 
186 static inline phys_addr_t rk_dte_pt_address(u32 dte)
187 {
188 	return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
189 }
190 
191 /*
192  * In v2:
193  * 31:12 - PT address bit 31:0
194  * 11: 8 - PT address bit 35:32
195  *  7: 4 - PT address bit 39:36
196  *  3: 1 - Reserved
197  *     0 - 1 if PT @ PT address is valid
198  */
199 #define RK_DTE_PT_ADDRESS_MASK_V2 GENMASK_ULL(31, 4)
200 #define DTE_HI_MASK1	GENMASK(11, 8)
201 #define DTE_HI_MASK2	GENMASK(7, 4)
202 #define DTE_HI_SHIFT1	24 /* shift bit 8 to bit 32 */
203 #define DTE_HI_SHIFT2	32 /* shift bit 4 to bit 36 */
204 #define PAGE_DESC_HI_MASK1	GENMASK_ULL(39, 36)
205 #define PAGE_DESC_HI_MASK2	GENMASK_ULL(35, 32)
206 
207 static inline phys_addr_t rk_dte_pt_address_v2(u32 dte)
208 {
209 	u64 dte_v2 = dte;
210 
211 	dte_v2 = ((dte_v2 & DTE_HI_MASK2) << DTE_HI_SHIFT2) |
212 		 ((dte_v2 & DTE_HI_MASK1) << DTE_HI_SHIFT1) |
213 		 (dte_v2 & RK_DTE_PT_ADDRESS_MASK);
214 
215 	return (phys_addr_t)dte_v2;
216 }
217 
218 static inline bool rk_dte_is_pt_valid(u32 dte)
219 {
220 	return dte & RK_DTE_PT_VALID;
221 }
222 
223 static inline u32 rk_mk_dte(dma_addr_t pt_dma)
224 {
225 	return (pt_dma & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
226 }
227 
228 static inline u32 rk_mk_dte_v2(dma_addr_t pt_dma)
229 {
230 	pt_dma = (pt_dma & RK_DTE_PT_ADDRESS_MASK) |
231 		 ((pt_dma & PAGE_DESC_HI_MASK1) >> DTE_HI_SHIFT1) |
232 		 (pt_dma & PAGE_DESC_HI_MASK2) >> DTE_HI_SHIFT2;
233 
234 	return (pt_dma & RK_DTE_PT_ADDRESS_MASK_V2) | RK_DTE_PT_VALID;
235 }
236 
237 /*
238  * Each PTE has a Page address, some flags and a valid bit:
239  * +---------------------+---+-------+-+
240  * | Page address        |Rsv| Flags |V|
241  * +---------------------+---+-------+-+
242  *  31:12 - Page address (Pages always start on a 4 KB boundary)
243  *  11: 9 - Reserved
244  *   8: 1 - Flags
245  *      8 - Read allocate - allocate cache space on read misses
246  *      7 - Read cache - enable cache & prefetch of data
247  *      6 - Write buffer - enable delaying writes on their way to memory
248  *      5 - Write allocate - allocate cache space on write misses
249  *      4 - Write cache - different writes can be merged together
250  *      3 - Override cache attributes
251  *          if 1, bits 4-8 control cache attributes
252  *          if 0, the system bus defaults are used
253  *      2 - Writable
254  *      1 - Readable
255  *      0 - 1 if Page @ Page address is valid
256  */
257 #define RK_PTE_PAGE_ADDRESS_MASK  0xfffff000
258 #define RK_PTE_PAGE_FLAGS_MASK    0x000001fe
259 #define RK_PTE_PAGE_WRITABLE      BIT(2)
260 #define RK_PTE_PAGE_READABLE      BIT(1)
261 #define RK_PTE_PAGE_VALID         BIT(0)
262 
263 static inline bool rk_pte_is_page_valid(u32 pte)
264 {
265 	return pte & RK_PTE_PAGE_VALID;
266 }
267 
268 /* TODO: set cache flags per prot IOMMU_CACHE */
269 static u32 rk_mk_pte(phys_addr_t page, int prot)
270 {
271 	u32 flags = 0;
272 	flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
273 	flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
274 	page &= RK_PTE_PAGE_ADDRESS_MASK;
275 	return page | flags | RK_PTE_PAGE_VALID;
276 }
277 
278 /*
279  * In v2:
280  * 31:12 - Page address bit 31:0
281  *  11:9 - Page address bit 34:32
282  *   8:4 - Page address bit 39:35
283  *     3 - Security
284  *     2 - Readable
285  *     1 - Writable
286  *     0 - 1 if Page @ Page address is valid
287  */
288 #define RK_PTE_PAGE_READABLE_V2      BIT(2)
289 #define RK_PTE_PAGE_WRITABLE_V2      BIT(1)
290 
291 static u32 rk_mk_pte_v2(phys_addr_t page, int prot)
292 {
293 	u32 flags = 0;
294 
295 	flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE_V2 : 0;
296 	flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE_V2 : 0;
297 
298 	return rk_mk_dte_v2(page) | flags;
299 }
300 
301 static u32 rk_mk_pte_invalid(u32 pte)
302 {
303 	return pte & ~RK_PTE_PAGE_VALID;
304 }
305 
306 /*
307  * rk3288 iova (IOMMU Virtual Address) format
308  *  31       22.21       12.11          0
309  * +-----------+-----------+-------------+
310  * | DTE index | PTE index | Page offset |
311  * +-----------+-----------+-------------+
312  *  31:22 - DTE index   - index of DTE in DT
313  *  21:12 - PTE index   - index of PTE in PT @ DTE.pt_address
314  *  11: 0 - Page offset - offset into page @ PTE.page_address
315  */
316 #define RK_IOVA_DTE_MASK    0xffc00000
317 #define RK_IOVA_DTE_SHIFT   22
318 #define RK_IOVA_PTE_MASK    0x003ff000
319 #define RK_IOVA_PTE_SHIFT   12
320 #define RK_IOVA_PAGE_MASK   0x00000fff
321 #define RK_IOVA_PAGE_SHIFT  0
322 
323 static u32 rk_iova_dte_index(dma_addr_t iova)
324 {
325 	return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
326 }
327 
328 static u32 rk_iova_pte_index(dma_addr_t iova)
329 {
330 	return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
331 }
332 
333 static u32 rk_iova_page_offset(dma_addr_t iova)
334 {
335 	return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
336 }
337 
338 static u32 rk_iommu_read(void __iomem *base, u32 offset)
339 {
340 	return readl(base + offset);
341 }
342 
343 static void rk_iommu_write(void __iomem *base, u32 offset, u32 value)
344 {
345 	writel(value, base + offset);
346 }
347 
348 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
349 {
350 	int i;
351 
352 	for (i = 0; i < iommu->num_mmu; i++)
353 		writel(command, iommu->bases[i] + RK_MMU_COMMAND);
354 }
355 
356 static void rk_iommu_base_command(void __iomem *base, u32 command)
357 {
358 	writel(command, base + RK_MMU_COMMAND);
359 }
360 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova_start,
361 			       size_t size)
362 {
363 	int i;
364 	dma_addr_t iova_end = iova_start + size;
365 	/*
366 	 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
367 	 * entire iotlb rather than iterate over individual iovas.
368 	 */
369 	for (i = 0; i < iommu->num_mmu; i++) {
370 		dma_addr_t iova;
371 
372 		for (iova = iova_start; iova < iova_end; iova += SPAGE_SIZE)
373 			rk_iommu_write(iommu->bases[i], RK_MMU_ZAP_ONE_LINE, iova);
374 	}
375 }
376 
377 static bool rk_iommu_is_stall_active(struct rk_iommu *iommu)
378 {
379 	bool active = true;
380 	int i;
381 
382 	for (i = 0; i < iommu->num_mmu; i++)
383 		active &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
384 					   RK_MMU_STATUS_STALL_ACTIVE);
385 
386 	return active;
387 }
388 
389 static bool rk_iommu_is_paging_enabled(struct rk_iommu *iommu)
390 {
391 	bool enable = true;
392 	int i;
393 
394 	for (i = 0; i < iommu->num_mmu; i++)
395 		enable &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
396 					   RK_MMU_STATUS_PAGING_ENABLED);
397 
398 	return enable;
399 }
400 
401 static bool rk_iommu_is_reset_done(struct rk_iommu *iommu)
402 {
403 	bool done = true;
404 	int i;
405 
406 	for (i = 0; i < iommu->num_mmu; i++)
407 		done &= rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR) == 0;
408 
409 	return done;
410 }
411 
412 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
413 {
414 	int ret, i;
415 	bool val;
416 
417 	if (rk_iommu_is_stall_active(iommu))
418 		return 0;
419 
420 	/* Stall can only be enabled if paging is enabled */
421 	if (!rk_iommu_is_paging_enabled(iommu))
422 		return 0;
423 
424 	rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
425 
426 	ret = readx_poll_timeout(rk_iommu_is_stall_active, 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, "Enable stall request timed out, status: %#08x\n",
432 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
433 
434 	return ret;
435 }
436 
437 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
438 {
439 	int ret, i;
440 	bool val;
441 
442 	if (!rk_iommu_is_stall_active(iommu))
443 		return 0;
444 
445 	rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
446 
447 	ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
448 				 !val, RK_MMU_POLL_PERIOD_US,
449 				 RK_MMU_POLL_TIMEOUT_US);
450 	if (ret)
451 		for (i = 0; i < iommu->num_mmu; i++)
452 			dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
453 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
454 
455 	return ret;
456 }
457 
458 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
459 {
460 	int ret, i;
461 	bool val;
462 
463 	if (rk_iommu_is_paging_enabled(iommu))
464 		return 0;
465 
466 	rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
467 
468 	ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
469 				 val, RK_MMU_POLL_PERIOD_US,
470 				 RK_MMU_POLL_TIMEOUT_US);
471 	if (ret)
472 		for (i = 0; i < iommu->num_mmu; i++)
473 			dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
474 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
475 
476 	return ret;
477 }
478 
479 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
480 {
481 	int ret, i;
482 	bool val;
483 
484 	if (!rk_iommu_is_paging_enabled(iommu))
485 		return 0;
486 
487 	rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
488 
489 	ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
490 				 !val, RK_MMU_POLL_PERIOD_US,
491 				 RK_MMU_POLL_TIMEOUT_US);
492 	if (ret)
493 		for (i = 0; i < iommu->num_mmu; i++)
494 			dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
495 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
496 
497 	return ret;
498 }
499 
500 static int rk_iommu_force_reset(struct rk_iommu *iommu)
501 {
502 	int ret, i;
503 	u32 dte_addr;
504 	bool val;
505 
506 	if (iommu->reset_disabled)
507 		return 0;
508 
509 	/*
510 	 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
511 	 * and verifying that upper 5 nybbles are read back.
512 	 */
513 	for (i = 0; i < iommu->num_mmu; i++) {
514 		dte_addr = rk_ops->pt_address(DTE_ADDR_DUMMY);
515 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, dte_addr);
516 
517 		if (dte_addr != rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR)) {
518 			dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
519 			return -EFAULT;
520 		}
521 	}
522 
523 	rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
524 
525 	ret = readx_poll_timeout(rk_iommu_is_reset_done, iommu, val,
526 				 val, RK_MMU_FORCE_RESET_TIMEOUT_US,
527 				 RK_MMU_POLL_TIMEOUT_US);
528 	if (ret) {
529 		dev_err(iommu->dev, "FORCE_RESET command timed out\n");
530 		return ret;
531 	}
532 
533 	return 0;
534 }
535 
536 static inline phys_addr_t rk_dte_addr_phys(u32 addr)
537 {
538 	return (phys_addr_t)addr;
539 }
540 
541 static inline u32 rk_dma_addr_dte(dma_addr_t dt_dma)
542 {
543 	return dt_dma;
544 }
545 
546 #define DT_HI_MASK GENMASK_ULL(39, 32)
547 #define DTE_BASE_HI_MASK GENMASK(11, 4)
548 #define DT_SHIFT   28
549 
550 static inline phys_addr_t rk_dte_addr_phys_v2(u32 addr)
551 {
552 	u64 addr64 = addr;
553 	return (phys_addr_t)(addr64 & RK_DTE_PT_ADDRESS_MASK) |
554 	       ((addr64 & DTE_BASE_HI_MASK) << DT_SHIFT);
555 }
556 
557 static inline u32 rk_dma_addr_dte_v2(dma_addr_t dt_dma)
558 {
559 	return (dt_dma & RK_DTE_PT_ADDRESS_MASK) |
560 	       ((dt_dma & DT_HI_MASK) >> DT_SHIFT);
561 }
562 
563 static void log_iova(struct rk_iommu *iommu, int index, dma_addr_t iova)
564 {
565 	void __iomem *base = iommu->bases[index];
566 	u32 dte_index, pte_index, page_offset;
567 	u32 mmu_dte_addr;
568 	phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
569 	u32 *dte_addr;
570 	u32 dte;
571 	phys_addr_t pte_addr_phys = 0;
572 	u32 *pte_addr = NULL;
573 	u32 pte = 0;
574 	phys_addr_t page_addr_phys = 0;
575 	u32 page_flags = 0;
576 
577 	dte_index = rk_iova_dte_index(iova);
578 	pte_index = rk_iova_pte_index(iova);
579 	page_offset = rk_iova_page_offset(iova);
580 
581 	mmu_dte_addr = rk_iommu_read(base, RK_MMU_DTE_ADDR);
582 	mmu_dte_addr_phys = rk_ops->dte_addr_phys(mmu_dte_addr);
583 
584 	dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
585 	dte_addr = phys_to_virt(dte_addr_phys);
586 	dte = *dte_addr;
587 
588 	if (!rk_dte_is_pt_valid(dte))
589 		goto print_it;
590 
591 	pte_addr_phys = rk_ops->pt_address(dte) + (pte_index * 4);
592 	pte_addr = phys_to_virt(pte_addr_phys);
593 	pte = *pte_addr;
594 
595 	if (!rk_pte_is_page_valid(pte))
596 		goto print_it;
597 
598 	page_addr_phys = rk_ops->pt_address(pte) + page_offset;
599 	page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
600 
601 print_it:
602 	dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
603 		&iova, dte_index, pte_index, page_offset);
604 	dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
605 		&mmu_dte_addr_phys, &dte_addr_phys, dte,
606 		rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
607 		rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
608 }
609 
610 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
611 {
612 	struct rk_iommu *iommu = dev_id;
613 	u32 status;
614 	u32 int_status;
615 	dma_addr_t iova;
616 	irqreturn_t ret = IRQ_NONE;
617 	int i, err;
618 
619 	err = pm_runtime_get_if_in_use(iommu->dev);
620 	if (!err || WARN_ON_ONCE(err < 0))
621 		return ret;
622 
623 	if (WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks)))
624 		goto out;
625 
626 	for (i = 0; i < iommu->num_mmu; i++) {
627 		int_status = rk_iommu_read(iommu->bases[i], RK_MMU_INT_STATUS);
628 		if (int_status == 0)
629 			continue;
630 
631 		ret = IRQ_HANDLED;
632 		iova = rk_iommu_read(iommu->bases[i], RK_MMU_PAGE_FAULT_ADDR);
633 
634 		if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
635 			int flags;
636 
637 			status = rk_iommu_read(iommu->bases[i], RK_MMU_STATUS);
638 			flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
639 					IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
640 
641 			dev_err(iommu->dev, "Page fault at %pad of type %s\n",
642 				&iova,
643 				(flags == IOMMU_FAULT_WRITE) ? "write" : "read");
644 
645 			log_iova(iommu, i, iova);
646 
647 			/*
648 			 * Report page fault to any installed handlers.
649 			 * Ignore the return code, though, since we always zap cache
650 			 * and clear the page fault anyway.
651 			 */
652 			if (iommu->domain)
653 				report_iommu_fault(iommu->domain, iommu->dev, iova,
654 						   flags);
655 			else
656 				dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
657 
658 			rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
659 			rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_PAGE_FAULT_DONE);
660 		}
661 
662 		if (int_status & RK_MMU_IRQ_BUS_ERROR)
663 			dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
664 
665 		if (int_status & ~RK_MMU_IRQ_MASK)
666 			dev_err(iommu->dev, "unexpected int_status: %#08x\n",
667 				int_status);
668 
669 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_CLEAR, int_status);
670 	}
671 
672 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
673 
674 out:
675 	pm_runtime_put(iommu->dev);
676 	return ret;
677 }
678 
679 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
680 					 dma_addr_t iova)
681 {
682 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
683 	unsigned long flags;
684 	phys_addr_t pt_phys, phys = 0;
685 	u32 dte, pte;
686 	u32 *page_table;
687 
688 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
689 
690 	dte = rk_domain->dt[rk_iova_dte_index(iova)];
691 	if (!rk_dte_is_pt_valid(dte))
692 		goto out;
693 
694 	pt_phys = rk_ops->pt_address(dte);
695 	page_table = (u32 *)phys_to_virt(pt_phys);
696 	pte = page_table[rk_iova_pte_index(iova)];
697 	if (!rk_pte_is_page_valid(pte))
698 		goto out;
699 
700 	phys = rk_ops->pt_address(pte) + rk_iova_page_offset(iova);
701 out:
702 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
703 
704 	return phys;
705 }
706 
707 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
708 			      dma_addr_t iova, size_t size)
709 {
710 	struct list_head *pos;
711 	unsigned long flags;
712 
713 	/* shootdown these iova from all iommus using this domain */
714 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
715 	list_for_each(pos, &rk_domain->iommus) {
716 		struct rk_iommu *iommu;
717 		int ret;
718 
719 		iommu = list_entry(pos, struct rk_iommu, node);
720 
721 		/* Only zap TLBs of IOMMUs that are powered on. */
722 		ret = pm_runtime_get_if_in_use(iommu->dev);
723 		if (WARN_ON_ONCE(ret < 0))
724 			continue;
725 		if (ret) {
726 			WARN_ON(clk_bulk_enable(iommu->num_clocks,
727 						iommu->clocks));
728 			rk_iommu_zap_lines(iommu, iova, size);
729 			clk_bulk_disable(iommu->num_clocks, iommu->clocks);
730 			pm_runtime_put(iommu->dev);
731 		}
732 	}
733 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
734 }
735 
736 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
737 					 dma_addr_t iova, size_t size)
738 {
739 	rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
740 	if (size > SPAGE_SIZE)
741 		rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
742 					SPAGE_SIZE);
743 }
744 
745 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
746 				  dma_addr_t iova)
747 {
748 	u32 *page_table, *dte_addr;
749 	u32 dte_index, dte;
750 	phys_addr_t pt_phys;
751 	dma_addr_t pt_dma;
752 
753 	assert_spin_locked(&rk_domain->dt_lock);
754 
755 	dte_index = rk_iova_dte_index(iova);
756 	dte_addr = &rk_domain->dt[dte_index];
757 	dte = *dte_addr;
758 	if (rk_dte_is_pt_valid(dte))
759 		goto done;
760 
761 	page_table = (u32 *)get_zeroed_page(GFP_ATOMIC | GFP_DMA32);
762 	if (!page_table)
763 		return ERR_PTR(-ENOMEM);
764 
765 	pt_dma = dma_map_single(dma_dev, page_table, SPAGE_SIZE, DMA_TO_DEVICE);
766 	if (dma_mapping_error(dma_dev, pt_dma)) {
767 		dev_err(dma_dev, "DMA mapping error while allocating page table\n");
768 		free_page((unsigned long)page_table);
769 		return ERR_PTR(-ENOMEM);
770 	}
771 
772 	dte = rk_ops->mk_dtentries(pt_dma);
773 	*dte_addr = dte;
774 
775 	rk_table_flush(rk_domain,
776 		       rk_domain->dt_dma + dte_index * sizeof(u32), 1);
777 done:
778 	pt_phys = rk_ops->pt_address(dte);
779 	return (u32 *)phys_to_virt(pt_phys);
780 }
781 
782 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
783 				  u32 *pte_addr, dma_addr_t pte_dma,
784 				  size_t size)
785 {
786 	unsigned int pte_count;
787 	unsigned int pte_total = size / SPAGE_SIZE;
788 
789 	assert_spin_locked(&rk_domain->dt_lock);
790 
791 	for (pte_count = 0; pte_count < pte_total; pte_count++) {
792 		u32 pte = pte_addr[pte_count];
793 		if (!rk_pte_is_page_valid(pte))
794 			break;
795 
796 		pte_addr[pte_count] = rk_mk_pte_invalid(pte);
797 	}
798 
799 	rk_table_flush(rk_domain, pte_dma, pte_count);
800 
801 	return pte_count * SPAGE_SIZE;
802 }
803 
804 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
805 			     dma_addr_t pte_dma, dma_addr_t iova,
806 			     phys_addr_t paddr, size_t size, int prot)
807 {
808 	unsigned int pte_count;
809 	unsigned int pte_total = size / SPAGE_SIZE;
810 	phys_addr_t page_phys;
811 
812 	assert_spin_locked(&rk_domain->dt_lock);
813 
814 	for (pte_count = 0; pte_count < pte_total; pte_count++) {
815 		u32 pte = pte_addr[pte_count];
816 
817 		if (rk_pte_is_page_valid(pte))
818 			goto unwind;
819 
820 		pte_addr[pte_count] = rk_ops->mk_ptentries(paddr, prot);
821 
822 		paddr += SPAGE_SIZE;
823 	}
824 
825 	rk_table_flush(rk_domain, pte_dma, pte_total);
826 
827 	/*
828 	 * Zap the first and last iova to evict from iotlb any previously
829 	 * mapped cachelines holding stale values for its dte and pte.
830 	 * We only zap the first and last iova, since only they could have
831 	 * dte or pte shared with an existing mapping.
832 	 */
833 	rk_iommu_zap_iova_first_last(rk_domain, iova, size);
834 
835 	return 0;
836 unwind:
837 	/* Unmap the range of iovas that we just mapped */
838 	rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma,
839 			    pte_count * SPAGE_SIZE);
840 
841 	iova += pte_count * SPAGE_SIZE;
842 	page_phys = rk_ops->pt_address(pte_addr[pte_count]);
843 	pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
844 	       &iova, &page_phys, &paddr, prot);
845 
846 	return -EADDRINUSE;
847 }
848 
849 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
850 			phys_addr_t paddr, size_t size, int prot, gfp_t gfp)
851 {
852 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
853 	unsigned long flags;
854 	dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
855 	u32 *page_table, *pte_addr;
856 	u32 dte_index, pte_index;
857 	int ret;
858 
859 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
860 
861 	/*
862 	 * pgsize_bitmap specifies iova sizes that fit in one page table
863 	 * (1024 4-KiB pages = 4 MiB).
864 	 * So, size will always be 4096 <= size <= 4194304.
865 	 * Since iommu_map() guarantees that both iova and size will be
866 	 * aligned, we will always only be mapping from a single dte here.
867 	 */
868 	page_table = rk_dte_get_page_table(rk_domain, iova);
869 	if (IS_ERR(page_table)) {
870 		spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
871 		return PTR_ERR(page_table);
872 	}
873 
874 	dte_index = rk_domain->dt[rk_iova_dte_index(iova)];
875 	pte_index = rk_iova_pte_index(iova);
876 	pte_addr = &page_table[pte_index];
877 
878 	pte_dma = rk_ops->pt_address(dte_index) + pte_index * sizeof(u32);
879 	ret = rk_iommu_map_iova(rk_domain, pte_addr, pte_dma, iova,
880 				paddr, size, prot);
881 
882 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
883 
884 	return ret;
885 }
886 
887 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
888 			     size_t size, struct iommu_iotlb_gather *gather)
889 {
890 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
891 	unsigned long flags;
892 	dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
893 	phys_addr_t pt_phys;
894 	u32 dte;
895 	u32 *pte_addr;
896 	size_t unmap_size;
897 
898 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
899 
900 	/*
901 	 * pgsize_bitmap specifies iova sizes that fit in one page table
902 	 * (1024 4-KiB pages = 4 MiB).
903 	 * So, size will always be 4096 <= size <= 4194304.
904 	 * Since iommu_unmap() guarantees that both iova and size will be
905 	 * aligned, we will always only be unmapping from a single dte here.
906 	 */
907 	dte = rk_domain->dt[rk_iova_dte_index(iova)];
908 	/* Just return 0 if iova is unmapped */
909 	if (!rk_dte_is_pt_valid(dte)) {
910 		spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
911 		return 0;
912 	}
913 
914 	pt_phys = rk_ops->pt_address(dte);
915 	pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
916 	pte_dma = pt_phys + rk_iova_pte_index(iova) * sizeof(u32);
917 	unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma, size);
918 
919 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
920 
921 	/* Shootdown iotlb entries for iova range that was just unmapped */
922 	rk_iommu_zap_iova(rk_domain, iova, unmap_size);
923 
924 	return unmap_size;
925 }
926 
927 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
928 {
929 	struct rk_iommudata *data = dev_iommu_priv_get(dev);
930 
931 	return data ? data->iommu : NULL;
932 }
933 
934 /* Must be called with iommu powered on and attached */
935 static void rk_iommu_disable(struct rk_iommu *iommu)
936 {
937 	int i;
938 
939 	/* Ignore error while disabling, just keep going */
940 	WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks));
941 	rk_iommu_enable_stall(iommu);
942 	rk_iommu_disable_paging(iommu);
943 	for (i = 0; i < iommu->num_mmu; i++) {
944 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, 0);
945 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, 0);
946 	}
947 	rk_iommu_disable_stall(iommu);
948 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
949 }
950 
951 /* Must be called with iommu powered on and attached */
952 static int rk_iommu_enable(struct rk_iommu *iommu)
953 {
954 	struct iommu_domain *domain = iommu->domain;
955 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
956 	int ret, i;
957 
958 	ret = clk_bulk_enable(iommu->num_clocks, iommu->clocks);
959 	if (ret)
960 		return ret;
961 
962 	ret = rk_iommu_enable_stall(iommu);
963 	if (ret)
964 		goto out_disable_clocks;
965 
966 	ret = rk_iommu_force_reset(iommu);
967 	if (ret)
968 		goto out_disable_stall;
969 
970 	for (i = 0; i < iommu->num_mmu; i++) {
971 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
972 			       rk_ops->dma_addr_dte(rk_domain->dt_dma));
973 		rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
974 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
975 	}
976 
977 	ret = rk_iommu_enable_paging(iommu);
978 
979 out_disable_stall:
980 	rk_iommu_disable_stall(iommu);
981 out_disable_clocks:
982 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
983 	return ret;
984 }
985 
986 static void rk_iommu_detach_device(struct iommu_domain *domain,
987 				   struct device *dev)
988 {
989 	struct rk_iommu *iommu;
990 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
991 	unsigned long flags;
992 	int ret;
993 
994 	/* Allow 'virtual devices' (eg drm) to detach from domain */
995 	iommu = rk_iommu_from_dev(dev);
996 	if (!iommu)
997 		return;
998 
999 	dev_dbg(dev, "Detaching from iommu domain\n");
1000 
1001 	/* iommu already detached */
1002 	if (iommu->domain != domain)
1003 		return;
1004 
1005 	iommu->domain = NULL;
1006 
1007 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
1008 	list_del_init(&iommu->node);
1009 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
1010 
1011 	ret = pm_runtime_get_if_in_use(iommu->dev);
1012 	WARN_ON_ONCE(ret < 0);
1013 	if (ret > 0) {
1014 		rk_iommu_disable(iommu);
1015 		pm_runtime_put(iommu->dev);
1016 	}
1017 }
1018 
1019 static int rk_iommu_attach_device(struct iommu_domain *domain,
1020 		struct device *dev)
1021 {
1022 	struct rk_iommu *iommu;
1023 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1024 	unsigned long flags;
1025 	int ret;
1026 
1027 	/*
1028 	 * Allow 'virtual devices' (e.g., drm) to attach to domain.
1029 	 * Such a device does not belong to an iommu group.
1030 	 */
1031 	iommu = rk_iommu_from_dev(dev);
1032 	if (!iommu)
1033 		return 0;
1034 
1035 	dev_dbg(dev, "Attaching to iommu domain\n");
1036 
1037 	/* iommu already attached */
1038 	if (iommu->domain == domain)
1039 		return 0;
1040 
1041 	if (iommu->domain)
1042 		rk_iommu_detach_device(iommu->domain, dev);
1043 
1044 	iommu->domain = domain;
1045 
1046 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
1047 	list_add_tail(&iommu->node, &rk_domain->iommus);
1048 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
1049 
1050 	ret = pm_runtime_get_if_in_use(iommu->dev);
1051 	if (!ret || WARN_ON_ONCE(ret < 0))
1052 		return 0;
1053 
1054 	ret = rk_iommu_enable(iommu);
1055 	if (ret)
1056 		rk_iommu_detach_device(iommu->domain, dev);
1057 
1058 	pm_runtime_put(iommu->dev);
1059 
1060 	return ret;
1061 }
1062 
1063 static struct iommu_domain *rk_iommu_domain_alloc(unsigned type)
1064 {
1065 	struct rk_iommu_domain *rk_domain;
1066 
1067 	if (type != IOMMU_DOMAIN_UNMANAGED && type != IOMMU_DOMAIN_DMA)
1068 		return NULL;
1069 
1070 	if (!dma_dev)
1071 		return NULL;
1072 
1073 	rk_domain = kzalloc(sizeof(*rk_domain), GFP_KERNEL);
1074 	if (!rk_domain)
1075 		return NULL;
1076 
1077 	if (type == IOMMU_DOMAIN_DMA &&
1078 	    iommu_get_dma_cookie(&rk_domain->domain))
1079 		goto err_free_domain;
1080 
1081 	/*
1082 	 * rk32xx iommus use a 2 level pagetable.
1083 	 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
1084 	 * Allocate one 4 KiB page for each table.
1085 	 */
1086 	rk_domain->dt = (u32 *)get_zeroed_page(GFP_KERNEL | GFP_DMA32);
1087 	if (!rk_domain->dt)
1088 		goto err_put_cookie;
1089 
1090 	rk_domain->dt_dma = dma_map_single(dma_dev, rk_domain->dt,
1091 					   SPAGE_SIZE, DMA_TO_DEVICE);
1092 	if (dma_mapping_error(dma_dev, rk_domain->dt_dma)) {
1093 		dev_err(dma_dev, "DMA map error for DT\n");
1094 		goto err_free_dt;
1095 	}
1096 
1097 	spin_lock_init(&rk_domain->iommus_lock);
1098 	spin_lock_init(&rk_domain->dt_lock);
1099 	INIT_LIST_HEAD(&rk_domain->iommus);
1100 
1101 	rk_domain->domain.geometry.aperture_start = 0;
1102 	rk_domain->domain.geometry.aperture_end   = DMA_BIT_MASK(32);
1103 	rk_domain->domain.geometry.force_aperture = true;
1104 
1105 	return &rk_domain->domain;
1106 
1107 err_free_dt:
1108 	free_page((unsigned long)rk_domain->dt);
1109 err_put_cookie:
1110 	if (type == IOMMU_DOMAIN_DMA)
1111 		iommu_put_dma_cookie(&rk_domain->domain);
1112 err_free_domain:
1113 	kfree(rk_domain);
1114 
1115 	return NULL;
1116 }
1117 
1118 static void rk_iommu_domain_free(struct iommu_domain *domain)
1119 {
1120 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1121 	int i;
1122 
1123 	WARN_ON(!list_empty(&rk_domain->iommus));
1124 
1125 	for (i = 0; i < NUM_DT_ENTRIES; i++) {
1126 		u32 dte = rk_domain->dt[i];
1127 		if (rk_dte_is_pt_valid(dte)) {
1128 			phys_addr_t pt_phys = rk_ops->pt_address(dte);
1129 			u32 *page_table = phys_to_virt(pt_phys);
1130 			dma_unmap_single(dma_dev, pt_phys,
1131 					 SPAGE_SIZE, DMA_TO_DEVICE);
1132 			free_page((unsigned long)page_table);
1133 		}
1134 	}
1135 
1136 	dma_unmap_single(dma_dev, rk_domain->dt_dma,
1137 			 SPAGE_SIZE, DMA_TO_DEVICE);
1138 	free_page((unsigned long)rk_domain->dt);
1139 
1140 	if (domain->type == IOMMU_DOMAIN_DMA)
1141 		iommu_put_dma_cookie(&rk_domain->domain);
1142 	kfree(rk_domain);
1143 }
1144 
1145 static struct iommu_device *rk_iommu_probe_device(struct device *dev)
1146 {
1147 	struct rk_iommudata *data;
1148 	struct rk_iommu *iommu;
1149 
1150 	data = dev_iommu_priv_get(dev);
1151 	if (!data)
1152 		return ERR_PTR(-ENODEV);
1153 
1154 	iommu = rk_iommu_from_dev(dev);
1155 
1156 	data->link = device_link_add(dev, iommu->dev,
1157 				     DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME);
1158 
1159 	return &iommu->iommu;
1160 }
1161 
1162 static void rk_iommu_release_device(struct device *dev)
1163 {
1164 	struct rk_iommudata *data = dev_iommu_priv_get(dev);
1165 
1166 	device_link_del(data->link);
1167 }
1168 
1169 static struct iommu_group *rk_iommu_device_group(struct device *dev)
1170 {
1171 	struct rk_iommu *iommu;
1172 
1173 	iommu = rk_iommu_from_dev(dev);
1174 
1175 	return iommu_group_ref_get(iommu->group);
1176 }
1177 
1178 static int rk_iommu_of_xlate(struct device *dev,
1179 			     struct of_phandle_args *args)
1180 {
1181 	struct platform_device *iommu_dev;
1182 	struct rk_iommudata *data;
1183 
1184 	data = devm_kzalloc(dma_dev, sizeof(*data), GFP_KERNEL);
1185 	if (!data)
1186 		return -ENOMEM;
1187 
1188 	iommu_dev = of_find_device_by_node(args->np);
1189 
1190 	data->iommu = platform_get_drvdata(iommu_dev);
1191 	dev_iommu_priv_set(dev, data);
1192 
1193 	platform_device_put(iommu_dev);
1194 
1195 	return 0;
1196 }
1197 
1198 static const struct iommu_ops rk_iommu_ops = {
1199 	.domain_alloc = rk_iommu_domain_alloc,
1200 	.domain_free = rk_iommu_domain_free,
1201 	.attach_dev = rk_iommu_attach_device,
1202 	.detach_dev = rk_iommu_detach_device,
1203 	.map = rk_iommu_map,
1204 	.unmap = rk_iommu_unmap,
1205 	.probe_device = rk_iommu_probe_device,
1206 	.release_device = rk_iommu_release_device,
1207 	.iova_to_phys = rk_iommu_iova_to_phys,
1208 	.device_group = rk_iommu_device_group,
1209 	.pgsize_bitmap = RK_IOMMU_PGSIZE_BITMAP,
1210 	.of_xlate = rk_iommu_of_xlate,
1211 };
1212 
1213 static int rk_iommu_probe(struct platform_device *pdev)
1214 {
1215 	struct device *dev = &pdev->dev;
1216 	struct rk_iommu *iommu;
1217 	struct resource *res;
1218 	const struct rk_iommu_ops *ops;
1219 	int num_res = pdev->num_resources;
1220 	int err, i;
1221 
1222 	iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
1223 	if (!iommu)
1224 		return -ENOMEM;
1225 
1226 	platform_set_drvdata(pdev, iommu);
1227 	iommu->dev = dev;
1228 	iommu->num_mmu = 0;
1229 
1230 	ops = of_device_get_match_data(dev);
1231 	if (!rk_ops)
1232 		rk_ops = ops;
1233 
1234 	/*
1235 	 * That should not happen unless different versions of the
1236 	 * hardware block are embedded the same SoC
1237 	 */
1238 	if (WARN_ON(rk_ops != ops))
1239 		return -EINVAL;
1240 
1241 	iommu->bases = devm_kcalloc(dev, num_res, sizeof(*iommu->bases),
1242 				    GFP_KERNEL);
1243 	if (!iommu->bases)
1244 		return -ENOMEM;
1245 
1246 	for (i = 0; i < num_res; i++) {
1247 		res = platform_get_resource(pdev, IORESOURCE_MEM, i);
1248 		if (!res)
1249 			continue;
1250 		iommu->bases[i] = devm_ioremap_resource(&pdev->dev, res);
1251 		if (IS_ERR(iommu->bases[i]))
1252 			continue;
1253 		iommu->num_mmu++;
1254 	}
1255 	if (iommu->num_mmu == 0)
1256 		return PTR_ERR(iommu->bases[0]);
1257 
1258 	iommu->num_irq = platform_irq_count(pdev);
1259 	if (iommu->num_irq < 0)
1260 		return iommu->num_irq;
1261 
1262 	iommu->reset_disabled = device_property_read_bool(dev,
1263 					"rockchip,disable-mmu-reset");
1264 
1265 	iommu->num_clocks = ARRAY_SIZE(rk_iommu_clocks);
1266 	iommu->clocks = devm_kcalloc(iommu->dev, iommu->num_clocks,
1267 				     sizeof(*iommu->clocks), GFP_KERNEL);
1268 	if (!iommu->clocks)
1269 		return -ENOMEM;
1270 
1271 	for (i = 0; i < iommu->num_clocks; ++i)
1272 		iommu->clocks[i].id = rk_iommu_clocks[i];
1273 
1274 	/*
1275 	 * iommu clocks should be present for all new devices and devicetrees
1276 	 * but there are older devicetrees without clocks out in the wild.
1277 	 * So clocks as optional for the time being.
1278 	 */
1279 	err = devm_clk_bulk_get(iommu->dev, iommu->num_clocks, iommu->clocks);
1280 	if (err == -ENOENT)
1281 		iommu->num_clocks = 0;
1282 	else if (err)
1283 		return err;
1284 
1285 	err = clk_bulk_prepare(iommu->num_clocks, iommu->clocks);
1286 	if (err)
1287 		return err;
1288 
1289 	iommu->group = iommu_group_alloc();
1290 	if (IS_ERR(iommu->group)) {
1291 		err = PTR_ERR(iommu->group);
1292 		goto err_unprepare_clocks;
1293 	}
1294 
1295 	err = iommu_device_sysfs_add(&iommu->iommu, dev, NULL, dev_name(dev));
1296 	if (err)
1297 		goto err_put_group;
1298 
1299 	err = iommu_device_register(&iommu->iommu, &rk_iommu_ops, dev);
1300 	if (err)
1301 		goto err_remove_sysfs;
1302 
1303 	/*
1304 	 * Use the first registered IOMMU device for domain to use with DMA
1305 	 * API, since a domain might not physically correspond to a single
1306 	 * IOMMU device..
1307 	 */
1308 	if (!dma_dev)
1309 		dma_dev = &pdev->dev;
1310 
1311 	bus_set_iommu(&platform_bus_type, &rk_iommu_ops);
1312 
1313 	pm_runtime_enable(dev);
1314 
1315 	for (i = 0; i < iommu->num_irq; i++) {
1316 		int irq = platform_get_irq(pdev, i);
1317 
1318 		if (irq < 0)
1319 			return irq;
1320 
1321 		err = devm_request_irq(iommu->dev, irq, rk_iommu_irq,
1322 				       IRQF_SHARED, dev_name(dev), iommu);
1323 		if (err) {
1324 			pm_runtime_disable(dev);
1325 			goto err_remove_sysfs;
1326 		}
1327 	}
1328 
1329 	dma_set_mask_and_coherent(dev, rk_ops->dma_bit_mask);
1330 
1331 	return 0;
1332 err_remove_sysfs:
1333 	iommu_device_sysfs_remove(&iommu->iommu);
1334 err_put_group:
1335 	iommu_group_put(iommu->group);
1336 err_unprepare_clocks:
1337 	clk_bulk_unprepare(iommu->num_clocks, iommu->clocks);
1338 	return err;
1339 }
1340 
1341 static void rk_iommu_shutdown(struct platform_device *pdev)
1342 {
1343 	struct rk_iommu *iommu = platform_get_drvdata(pdev);
1344 	int i;
1345 
1346 	for (i = 0; i < iommu->num_irq; i++) {
1347 		int irq = platform_get_irq(pdev, i);
1348 
1349 		devm_free_irq(iommu->dev, irq, iommu);
1350 	}
1351 
1352 	pm_runtime_force_suspend(&pdev->dev);
1353 }
1354 
1355 static int __maybe_unused rk_iommu_suspend(struct device *dev)
1356 {
1357 	struct rk_iommu *iommu = dev_get_drvdata(dev);
1358 
1359 	if (!iommu->domain)
1360 		return 0;
1361 
1362 	rk_iommu_disable(iommu);
1363 	return 0;
1364 }
1365 
1366 static int __maybe_unused rk_iommu_resume(struct device *dev)
1367 {
1368 	struct rk_iommu *iommu = dev_get_drvdata(dev);
1369 
1370 	if (!iommu->domain)
1371 		return 0;
1372 
1373 	return rk_iommu_enable(iommu);
1374 }
1375 
1376 static const struct dev_pm_ops rk_iommu_pm_ops = {
1377 	SET_RUNTIME_PM_OPS(rk_iommu_suspend, rk_iommu_resume, NULL)
1378 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
1379 				pm_runtime_force_resume)
1380 };
1381 
1382 static struct rk_iommu_ops iommu_data_ops_v1 = {
1383 	.pt_address = &rk_dte_pt_address,
1384 	.mk_dtentries = &rk_mk_dte,
1385 	.mk_ptentries = &rk_mk_pte,
1386 	.dte_addr_phys = &rk_dte_addr_phys,
1387 	.dma_addr_dte = &rk_dma_addr_dte,
1388 	.dma_bit_mask = DMA_BIT_MASK(32),
1389 };
1390 
1391 static struct rk_iommu_ops iommu_data_ops_v2 = {
1392 	.pt_address = &rk_dte_pt_address_v2,
1393 	.mk_dtentries = &rk_mk_dte_v2,
1394 	.mk_ptentries = &rk_mk_pte_v2,
1395 	.dte_addr_phys = &rk_dte_addr_phys_v2,
1396 	.dma_addr_dte = &rk_dma_addr_dte_v2,
1397 	.dma_bit_mask = DMA_BIT_MASK(40),
1398 };
1399 
1400 static const struct of_device_id rk_iommu_dt_ids[] = {
1401 	{	.compatible = "rockchip,iommu",
1402 		.data = &iommu_data_ops_v1,
1403 	},
1404 	{	.compatible = "rockchip,rk3568-iommu",
1405 		.data = &iommu_data_ops_v2,
1406 	},
1407 	{ /* sentinel */ }
1408 };
1409 
1410 static struct platform_driver rk_iommu_driver = {
1411 	.probe = rk_iommu_probe,
1412 	.shutdown = rk_iommu_shutdown,
1413 	.driver = {
1414 		   .name = "rk_iommu",
1415 		   .of_match_table = rk_iommu_dt_ids,
1416 		   .pm = &rk_iommu_pm_ops,
1417 		   .suppress_bind_attrs = true,
1418 	},
1419 };
1420 
1421 static int __init rk_iommu_init(void)
1422 {
1423 	return platform_driver_register(&rk_iommu_driver);
1424 }
1425 subsys_initcall(rk_iommu_init);
1426