xref: /linux/drivers/iommu/rockchip-iommu.c (revision c48a7c44a1d02516309015b6134c9bb982e17008)
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-mapping.h>
14 #include <linux/errno.h>
15 #include <linux/interrupt.h>
16 #include <linux/io.h>
17 #include <linux/iommu.h>
18 #include <linux/iopoll.h>
19 #include <linux/list.h>
20 #include <linux/mm.h>
21 #include <linux/init.h>
22 #include <linux/of.h>
23 #include <linux/of_platform.h>
24 #include <linux/platform_device.h>
25 #include <linux/pm_runtime.h>
26 #include <linux/slab.h>
27 #include <linux/spinlock.h>
28 
29 /** MMU register offsets */
30 #define RK_MMU_DTE_ADDR		0x00	/* Directory table address */
31 #define RK_MMU_STATUS		0x04
32 #define RK_MMU_COMMAND		0x08
33 #define RK_MMU_PAGE_FAULT_ADDR	0x0C	/* IOVA of last page fault */
34 #define RK_MMU_ZAP_ONE_LINE	0x10	/* Shootdown one IOTLB entry */
35 #define RK_MMU_INT_RAWSTAT	0x14	/* IRQ status ignoring mask */
36 #define RK_MMU_INT_CLEAR	0x18	/* Acknowledge and re-arm irq */
37 #define RK_MMU_INT_MASK		0x1C	/* IRQ enable */
38 #define RK_MMU_INT_STATUS	0x20	/* IRQ status after masking */
39 #define RK_MMU_AUTO_GATING	0x24
40 
41 #define DTE_ADDR_DUMMY		0xCAFEBABE
42 
43 #define RK_MMU_POLL_PERIOD_US		100
44 #define RK_MMU_FORCE_RESET_TIMEOUT_US	100000
45 #define RK_MMU_POLL_TIMEOUT_US		1000
46 
47 /* RK_MMU_STATUS fields */
48 #define RK_MMU_STATUS_PAGING_ENABLED       BIT(0)
49 #define RK_MMU_STATUS_PAGE_FAULT_ACTIVE    BIT(1)
50 #define RK_MMU_STATUS_STALL_ACTIVE         BIT(2)
51 #define RK_MMU_STATUS_IDLE                 BIT(3)
52 #define RK_MMU_STATUS_REPLAY_BUFFER_EMPTY  BIT(4)
53 #define RK_MMU_STATUS_PAGE_FAULT_IS_WRITE  BIT(5)
54 #define RK_MMU_STATUS_STALL_NOT_ACTIVE     BIT(31)
55 
56 /* RK_MMU_COMMAND command values */
57 #define RK_MMU_CMD_ENABLE_PAGING    0  /* Enable memory translation */
58 #define RK_MMU_CMD_DISABLE_PAGING   1  /* Disable memory translation */
59 #define RK_MMU_CMD_ENABLE_STALL     2  /* Stall paging to allow other cmds */
60 #define RK_MMU_CMD_DISABLE_STALL    3  /* Stop stall re-enables paging */
61 #define RK_MMU_CMD_ZAP_CACHE        4  /* Shoot down entire IOTLB */
62 #define RK_MMU_CMD_PAGE_FAULT_DONE  5  /* Clear page fault */
63 #define RK_MMU_CMD_FORCE_RESET      6  /* Reset all registers */
64 
65 /* RK_MMU_INT_* register fields */
66 #define RK_MMU_IRQ_PAGE_FAULT    0x01  /* page fault */
67 #define RK_MMU_IRQ_BUS_ERROR     0x02  /* bus read error */
68 #define RK_MMU_IRQ_MASK          (RK_MMU_IRQ_PAGE_FAULT | RK_MMU_IRQ_BUS_ERROR)
69 
70 #define NUM_DT_ENTRIES 1024
71 #define NUM_PT_ENTRIES 1024
72 
73 #define SPAGE_ORDER 12
74 #define SPAGE_SIZE (1 << SPAGE_ORDER)
75 
76  /*
77   * Support mapping any size that fits in one page table:
78   *   4 KiB to 4 MiB
79   */
80 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
81 
82 struct rk_iommu_domain {
83 	struct list_head iommus;
84 	u32 *dt; /* page directory table */
85 	dma_addr_t dt_dma;
86 	spinlock_t iommus_lock; /* lock for iommus list */
87 	spinlock_t dt_lock; /* lock for modifying page directory table */
88 
89 	struct iommu_domain domain;
90 };
91 
92 /* list of clocks required by IOMMU */
93 static const char * const rk_iommu_clocks[] = {
94 	"aclk", "iface",
95 };
96 
97 struct rk_iommu_ops {
98 	phys_addr_t (*pt_address)(u32 dte);
99 	u32 (*mk_dtentries)(dma_addr_t pt_dma);
100 	u32 (*mk_ptentries)(phys_addr_t page, int prot);
101 	u64 dma_bit_mask;
102 	gfp_t gfp_flags;
103 };
104 
105 struct rk_iommu {
106 	struct device *dev;
107 	void __iomem **bases;
108 	int num_mmu;
109 	int num_irq;
110 	struct clk_bulk_data *clocks;
111 	int num_clocks;
112 	bool reset_disabled;
113 	struct iommu_device iommu;
114 	struct list_head node; /* entry in rk_iommu_domain.iommus */
115 	struct iommu_domain *domain; /* domain to which iommu is attached */
116 };
117 
118 struct rk_iommudata {
119 	struct device_link *link; /* runtime PM link from IOMMU to master */
120 	struct rk_iommu *iommu;
121 };
122 
123 static struct device *dma_dev;
124 static const struct rk_iommu_ops *rk_ops;
125 static struct iommu_domain rk_identity_domain;
126 
127 static inline void rk_table_flush(struct rk_iommu_domain *dom, dma_addr_t dma,
128 				  unsigned int count)
129 {
130 	size_t size = count * sizeof(u32); /* count of u32 entry */
131 
132 	dma_sync_single_for_device(dma_dev, dma, size, DMA_TO_DEVICE);
133 }
134 
135 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
136 {
137 	return container_of(dom, struct rk_iommu_domain, domain);
138 }
139 
140 /*
141  * The Rockchip rk3288 iommu uses a 2-level page table.
142  * The first level is the "Directory Table" (DT).
143  * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
144  * to a "Page Table".
145  * The second level is the 1024 Page Tables (PT).
146  * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
147  * a 4 KB page of physical memory.
148  *
149  * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
150  * Each iommu device has a MMU_DTE_ADDR register that contains the physical
151  * address of the start of the DT page.
152  *
153  * The structure of the page table is as follows:
154  *
155  *                   DT
156  * MMU_DTE_ADDR -> +-----+
157  *                 |     |
158  *                 +-----+     PT
159  *                 | DTE | -> +-----+
160  *                 +-----+    |     |     Memory
161  *                 |     |    +-----+     Page
162  *                 |     |    | PTE | -> +-----+
163  *                 +-----+    +-----+    |     |
164  *                            |     |    |     |
165  *                            |     |    |     |
166  *                            +-----+    |     |
167  *                                       |     |
168  *                                       |     |
169  *                                       +-----+
170  */
171 
172 /*
173  * Each DTE has a PT address and a valid bit:
174  * +---------------------+-----------+-+
175  * | PT address          | Reserved  |V|
176  * +---------------------+-----------+-+
177  *  31:12 - PT address (PTs always starts on a 4 KB boundary)
178  *  11: 1 - Reserved
179  *      0 - 1 if PT @ PT address is valid
180  */
181 #define RK_DTE_PT_ADDRESS_MASK    0xfffff000
182 #define RK_DTE_PT_VALID           BIT(0)
183 
184 static inline phys_addr_t rk_dte_pt_address(u32 dte)
185 {
186 	return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
187 }
188 
189 /*
190  * In v2:
191  * 31:12 - PT address bit 31:0
192  * 11: 8 - PT address bit 35:32
193  *  7: 4 - PT address bit 39:36
194  *  3: 1 - Reserved
195  *     0 - 1 if PT @ PT address is valid
196  */
197 #define RK_DTE_PT_ADDRESS_MASK_V2 GENMASK_ULL(31, 4)
198 #define DTE_HI_MASK1	GENMASK(11, 8)
199 #define DTE_HI_MASK2	GENMASK(7, 4)
200 #define DTE_HI_SHIFT1	24 /* shift bit 8 to bit 32 */
201 #define DTE_HI_SHIFT2	32 /* shift bit 4 to bit 36 */
202 #define PAGE_DESC_HI_MASK1	GENMASK_ULL(35, 32)
203 #define PAGE_DESC_HI_MASK2	GENMASK_ULL(39, 36)
204 
205 static inline phys_addr_t rk_dte_pt_address_v2(u32 dte)
206 {
207 	u64 dte_v2 = dte;
208 
209 	dte_v2 = ((dte_v2 & DTE_HI_MASK2) << DTE_HI_SHIFT2) |
210 		 ((dte_v2 & DTE_HI_MASK1) << DTE_HI_SHIFT1) |
211 		 (dte_v2 & RK_DTE_PT_ADDRESS_MASK);
212 
213 	return (phys_addr_t)dte_v2;
214 }
215 
216 static inline bool rk_dte_is_pt_valid(u32 dte)
217 {
218 	return dte & RK_DTE_PT_VALID;
219 }
220 
221 static inline u32 rk_mk_dte(dma_addr_t pt_dma)
222 {
223 	return (pt_dma & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
224 }
225 
226 static inline u32 rk_mk_dte_v2(dma_addr_t pt_dma)
227 {
228 	pt_dma = (pt_dma & RK_DTE_PT_ADDRESS_MASK) |
229 		 ((pt_dma & PAGE_DESC_HI_MASK1) >> DTE_HI_SHIFT1) |
230 		 (pt_dma & PAGE_DESC_HI_MASK2) >> DTE_HI_SHIFT2;
231 
232 	return (pt_dma & RK_DTE_PT_ADDRESS_MASK_V2) | RK_DTE_PT_VALID;
233 }
234 
235 /*
236  * Each PTE has a Page address, some flags and a valid bit:
237  * +---------------------+---+-------+-+
238  * | Page address        |Rsv| Flags |V|
239  * +---------------------+---+-------+-+
240  *  31:12 - Page address (Pages always start on a 4 KB boundary)
241  *  11: 9 - Reserved
242  *   8: 1 - Flags
243  *      8 - Read allocate - allocate cache space on read misses
244  *      7 - Read cache - enable cache & prefetch of data
245  *      6 - Write buffer - enable delaying writes on their way to memory
246  *      5 - Write allocate - allocate cache space on write misses
247  *      4 - Write cache - different writes can be merged together
248  *      3 - Override cache attributes
249  *          if 1, bits 4-8 control cache attributes
250  *          if 0, the system bus defaults are used
251  *      2 - Writable
252  *      1 - Readable
253  *      0 - 1 if Page @ Page address is valid
254  */
255 #define RK_PTE_PAGE_ADDRESS_MASK  0xfffff000
256 #define RK_PTE_PAGE_FLAGS_MASK    0x000001fe
257 #define RK_PTE_PAGE_WRITABLE      BIT(2)
258 #define RK_PTE_PAGE_READABLE      BIT(1)
259 #define RK_PTE_PAGE_VALID         BIT(0)
260 
261 static inline bool rk_pte_is_page_valid(u32 pte)
262 {
263 	return pte & RK_PTE_PAGE_VALID;
264 }
265 
266 /* TODO: set cache flags per prot IOMMU_CACHE */
267 static u32 rk_mk_pte(phys_addr_t page, int prot)
268 {
269 	u32 flags = 0;
270 	flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
271 	flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
272 	page &= RK_PTE_PAGE_ADDRESS_MASK;
273 	return page | flags | RK_PTE_PAGE_VALID;
274 }
275 
276 /*
277  * In v2:
278  * 31:12 - Page address bit 31:0
279  * 11: 8 - Page address bit 35:32
280  *  7: 4 - Page address bit 39:36
281  *     3 - Security
282  *     2 - Writable
283  *     1 - Readable
284  *     0 - 1 if Page @ Page address is valid
285  */
286 
287 static u32 rk_mk_pte_v2(phys_addr_t page, int prot)
288 {
289 	u32 flags = 0;
290 
291 	flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
292 	flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
293 
294 	return rk_mk_dte_v2(page) | flags;
295 }
296 
297 static u32 rk_mk_pte_invalid(u32 pte)
298 {
299 	return pte & ~RK_PTE_PAGE_VALID;
300 }
301 
302 /*
303  * rk3288 iova (IOMMU Virtual Address) format
304  *  31       22.21       12.11          0
305  * +-----------+-----------+-------------+
306  * | DTE index | PTE index | Page offset |
307  * +-----------+-----------+-------------+
308  *  31:22 - DTE index   - index of DTE in DT
309  *  21:12 - PTE index   - index of PTE in PT @ DTE.pt_address
310  *  11: 0 - Page offset - offset into page @ PTE.page_address
311  */
312 #define RK_IOVA_DTE_MASK    0xffc00000
313 #define RK_IOVA_DTE_SHIFT   22
314 #define RK_IOVA_PTE_MASK    0x003ff000
315 #define RK_IOVA_PTE_SHIFT   12
316 #define RK_IOVA_PAGE_MASK   0x00000fff
317 #define RK_IOVA_PAGE_SHIFT  0
318 
319 static u32 rk_iova_dte_index(dma_addr_t iova)
320 {
321 	return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
322 }
323 
324 static u32 rk_iova_pte_index(dma_addr_t iova)
325 {
326 	return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
327 }
328 
329 static u32 rk_iova_page_offset(dma_addr_t iova)
330 {
331 	return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
332 }
333 
334 static u32 rk_iommu_read(void __iomem *base, u32 offset)
335 {
336 	return readl(base + offset);
337 }
338 
339 static void rk_iommu_write(void __iomem *base, u32 offset, u32 value)
340 {
341 	writel(value, base + offset);
342 }
343 
344 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
345 {
346 	int i;
347 
348 	for (i = 0; i < iommu->num_mmu; i++)
349 		writel(command, iommu->bases[i] + RK_MMU_COMMAND);
350 }
351 
352 static void rk_iommu_base_command(void __iomem *base, u32 command)
353 {
354 	writel(command, base + RK_MMU_COMMAND);
355 }
356 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova_start,
357 			       size_t size)
358 {
359 	int i;
360 	dma_addr_t iova_end = iova_start + size;
361 	/*
362 	 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
363 	 * entire iotlb rather than iterate over individual iovas.
364 	 */
365 	for (i = 0; i < iommu->num_mmu; i++) {
366 		dma_addr_t iova;
367 
368 		for (iova = iova_start; iova < iova_end; iova += SPAGE_SIZE)
369 			rk_iommu_write(iommu->bases[i], RK_MMU_ZAP_ONE_LINE, iova);
370 	}
371 }
372 
373 static bool rk_iommu_is_stall_active(struct rk_iommu *iommu)
374 {
375 	bool active = true;
376 	int i;
377 
378 	for (i = 0; i < iommu->num_mmu; i++)
379 		active &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
380 					   RK_MMU_STATUS_STALL_ACTIVE);
381 
382 	return active;
383 }
384 
385 static bool rk_iommu_is_paging_enabled(struct rk_iommu *iommu)
386 {
387 	bool enable = true;
388 	int i;
389 
390 	for (i = 0; i < iommu->num_mmu; i++)
391 		enable &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
392 					   RK_MMU_STATUS_PAGING_ENABLED);
393 
394 	return enable;
395 }
396 
397 static bool rk_iommu_is_reset_done(struct rk_iommu *iommu)
398 {
399 	bool done = true;
400 	int i;
401 
402 	for (i = 0; i < iommu->num_mmu; i++)
403 		done &= rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR) == 0;
404 
405 	return done;
406 }
407 
408 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
409 {
410 	int ret, i;
411 	bool val;
412 
413 	if (rk_iommu_is_stall_active(iommu))
414 		return 0;
415 
416 	/* Stall can only be enabled if paging is enabled */
417 	if (!rk_iommu_is_paging_enabled(iommu))
418 		return 0;
419 
420 	rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
421 
422 	ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
423 				 val, RK_MMU_POLL_PERIOD_US,
424 				 RK_MMU_POLL_TIMEOUT_US);
425 	if (ret)
426 		for (i = 0; i < iommu->num_mmu; i++)
427 			dev_err(iommu->dev, "Enable stall request timed out, status: %#08x\n",
428 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
429 
430 	return ret;
431 }
432 
433 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
434 {
435 	int ret, i;
436 	bool val;
437 
438 	if (!rk_iommu_is_stall_active(iommu))
439 		return 0;
440 
441 	rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
442 
443 	ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
444 				 !val, RK_MMU_POLL_PERIOD_US,
445 				 RK_MMU_POLL_TIMEOUT_US);
446 	if (ret)
447 		for (i = 0; i < iommu->num_mmu; i++)
448 			dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
449 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
450 
451 	return ret;
452 }
453 
454 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
455 {
456 	int ret, i;
457 	bool val;
458 
459 	if (rk_iommu_is_paging_enabled(iommu))
460 		return 0;
461 
462 	rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
463 
464 	ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
465 				 val, RK_MMU_POLL_PERIOD_US,
466 				 RK_MMU_POLL_TIMEOUT_US);
467 	if (ret)
468 		for (i = 0; i < iommu->num_mmu; i++)
469 			dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
470 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
471 
472 	return ret;
473 }
474 
475 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
476 {
477 	int ret, i;
478 	bool val;
479 
480 	if (!rk_iommu_is_paging_enabled(iommu))
481 		return 0;
482 
483 	rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
484 
485 	ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
486 				 !val, RK_MMU_POLL_PERIOD_US,
487 				 RK_MMU_POLL_TIMEOUT_US);
488 	if (ret)
489 		for (i = 0; i < iommu->num_mmu; i++)
490 			dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
491 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
492 
493 	return ret;
494 }
495 
496 static int rk_iommu_force_reset(struct rk_iommu *iommu)
497 {
498 	int ret, i;
499 	u32 dte_addr;
500 	bool val;
501 
502 	if (iommu->reset_disabled)
503 		return 0;
504 
505 	/*
506 	 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
507 	 * and verifying that upper 5 (v1) or 7 (v2) nybbles are read back.
508 	 */
509 	for (i = 0; i < iommu->num_mmu; i++) {
510 		dte_addr = rk_ops->pt_address(DTE_ADDR_DUMMY);
511 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, dte_addr);
512 
513 		if (dte_addr != rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR)) {
514 			dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
515 			return -EFAULT;
516 		}
517 	}
518 
519 	rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
520 
521 	ret = readx_poll_timeout(rk_iommu_is_reset_done, iommu, val,
522 				 val, RK_MMU_FORCE_RESET_TIMEOUT_US,
523 				 RK_MMU_POLL_TIMEOUT_US);
524 	if (ret) {
525 		dev_err(iommu->dev, "FORCE_RESET command timed out\n");
526 		return ret;
527 	}
528 
529 	return 0;
530 }
531 
532 static void log_iova(struct rk_iommu *iommu, int index, dma_addr_t iova)
533 {
534 	void __iomem *base = iommu->bases[index];
535 	u32 dte_index, pte_index, page_offset;
536 	u32 mmu_dte_addr;
537 	phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
538 	u32 *dte_addr;
539 	u32 dte;
540 	phys_addr_t pte_addr_phys = 0;
541 	u32 *pte_addr = NULL;
542 	u32 pte = 0;
543 	phys_addr_t page_addr_phys = 0;
544 	u32 page_flags = 0;
545 
546 	dte_index = rk_iova_dte_index(iova);
547 	pte_index = rk_iova_pte_index(iova);
548 	page_offset = rk_iova_page_offset(iova);
549 
550 	mmu_dte_addr = rk_iommu_read(base, RK_MMU_DTE_ADDR);
551 	mmu_dte_addr_phys = rk_ops->pt_address(mmu_dte_addr);
552 
553 	dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
554 	dte_addr = phys_to_virt(dte_addr_phys);
555 	dte = *dte_addr;
556 
557 	if (!rk_dte_is_pt_valid(dte))
558 		goto print_it;
559 
560 	pte_addr_phys = rk_ops->pt_address(dte) + (pte_index * 4);
561 	pte_addr = phys_to_virt(pte_addr_phys);
562 	pte = *pte_addr;
563 
564 	if (!rk_pte_is_page_valid(pte))
565 		goto print_it;
566 
567 	page_addr_phys = rk_ops->pt_address(pte) + page_offset;
568 	page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
569 
570 print_it:
571 	dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
572 		&iova, dte_index, pte_index, page_offset);
573 	dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
574 		&mmu_dte_addr_phys, &dte_addr_phys, dte,
575 		rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
576 		rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
577 }
578 
579 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
580 {
581 	struct rk_iommu *iommu = dev_id;
582 	u32 status;
583 	u32 int_status;
584 	dma_addr_t iova;
585 	irqreturn_t ret = IRQ_NONE;
586 	int i, err;
587 
588 	err = pm_runtime_get_if_in_use(iommu->dev);
589 	if (!err || WARN_ON_ONCE(err < 0))
590 		return ret;
591 
592 	if (WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks)))
593 		goto out;
594 
595 	for (i = 0; i < iommu->num_mmu; i++) {
596 		int_status = rk_iommu_read(iommu->bases[i], RK_MMU_INT_STATUS);
597 		if (int_status == 0)
598 			continue;
599 
600 		ret = IRQ_HANDLED;
601 		iova = rk_iommu_read(iommu->bases[i], RK_MMU_PAGE_FAULT_ADDR);
602 
603 		if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
604 			int flags;
605 
606 			status = rk_iommu_read(iommu->bases[i], RK_MMU_STATUS);
607 			flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
608 					IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
609 
610 			dev_err(iommu->dev, "Page fault at %pad of type %s\n",
611 				&iova,
612 				(flags == IOMMU_FAULT_WRITE) ? "write" : "read");
613 
614 			log_iova(iommu, i, iova);
615 
616 			/*
617 			 * Report page fault to any installed handlers.
618 			 * Ignore the return code, though, since we always zap cache
619 			 * and clear the page fault anyway.
620 			 */
621 			if (iommu->domain != &rk_identity_domain)
622 				report_iommu_fault(iommu->domain, iommu->dev, iova,
623 						   flags);
624 			else
625 				dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
626 
627 			rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
628 			rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_PAGE_FAULT_DONE);
629 		}
630 
631 		if (int_status & RK_MMU_IRQ_BUS_ERROR)
632 			dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
633 
634 		if (int_status & ~RK_MMU_IRQ_MASK)
635 			dev_err(iommu->dev, "unexpected int_status: %#08x\n",
636 				int_status);
637 
638 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_CLEAR, int_status);
639 	}
640 
641 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
642 
643 out:
644 	pm_runtime_put(iommu->dev);
645 	return ret;
646 }
647 
648 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
649 					 dma_addr_t iova)
650 {
651 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
652 	unsigned long flags;
653 	phys_addr_t pt_phys, phys = 0;
654 	u32 dte, pte;
655 	u32 *page_table;
656 
657 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
658 
659 	dte = rk_domain->dt[rk_iova_dte_index(iova)];
660 	if (!rk_dte_is_pt_valid(dte))
661 		goto out;
662 
663 	pt_phys = rk_ops->pt_address(dte);
664 	page_table = (u32 *)phys_to_virt(pt_phys);
665 	pte = page_table[rk_iova_pte_index(iova)];
666 	if (!rk_pte_is_page_valid(pte))
667 		goto out;
668 
669 	phys = rk_ops->pt_address(pte) + rk_iova_page_offset(iova);
670 out:
671 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
672 
673 	return phys;
674 }
675 
676 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
677 			      dma_addr_t iova, size_t size)
678 {
679 	struct list_head *pos;
680 	unsigned long flags;
681 
682 	/* shootdown these iova from all iommus using this domain */
683 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
684 	list_for_each(pos, &rk_domain->iommus) {
685 		struct rk_iommu *iommu;
686 		int ret;
687 
688 		iommu = list_entry(pos, struct rk_iommu, node);
689 
690 		/* Only zap TLBs of IOMMUs that are powered on. */
691 		ret = pm_runtime_get_if_in_use(iommu->dev);
692 		if (WARN_ON_ONCE(ret < 0))
693 			continue;
694 		if (ret) {
695 			WARN_ON(clk_bulk_enable(iommu->num_clocks,
696 						iommu->clocks));
697 			rk_iommu_zap_lines(iommu, iova, size);
698 			clk_bulk_disable(iommu->num_clocks, iommu->clocks);
699 			pm_runtime_put(iommu->dev);
700 		}
701 	}
702 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
703 }
704 
705 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
706 					 dma_addr_t iova, size_t size)
707 {
708 	rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
709 	if (size > SPAGE_SIZE)
710 		rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
711 					SPAGE_SIZE);
712 }
713 
714 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
715 				  dma_addr_t iova)
716 {
717 	u32 *page_table, *dte_addr;
718 	u32 dte_index, dte;
719 	phys_addr_t pt_phys;
720 	dma_addr_t pt_dma;
721 
722 	assert_spin_locked(&rk_domain->dt_lock);
723 
724 	dte_index = rk_iova_dte_index(iova);
725 	dte_addr = &rk_domain->dt[dte_index];
726 	dte = *dte_addr;
727 	if (rk_dte_is_pt_valid(dte))
728 		goto done;
729 
730 	page_table = (u32 *)get_zeroed_page(GFP_ATOMIC | rk_ops->gfp_flags);
731 	if (!page_table)
732 		return ERR_PTR(-ENOMEM);
733 
734 	pt_dma = dma_map_single(dma_dev, page_table, SPAGE_SIZE, DMA_TO_DEVICE);
735 	if (dma_mapping_error(dma_dev, pt_dma)) {
736 		dev_err(dma_dev, "DMA mapping error while allocating page table\n");
737 		free_page((unsigned long)page_table);
738 		return ERR_PTR(-ENOMEM);
739 	}
740 
741 	dte = rk_ops->mk_dtentries(pt_dma);
742 	*dte_addr = dte;
743 
744 	rk_table_flush(rk_domain,
745 		       rk_domain->dt_dma + dte_index * sizeof(u32), 1);
746 done:
747 	pt_phys = rk_ops->pt_address(dte);
748 	return (u32 *)phys_to_virt(pt_phys);
749 }
750 
751 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
752 				  u32 *pte_addr, dma_addr_t pte_dma,
753 				  size_t size)
754 {
755 	unsigned int pte_count;
756 	unsigned int pte_total = size / SPAGE_SIZE;
757 
758 	assert_spin_locked(&rk_domain->dt_lock);
759 
760 	for (pte_count = 0; pte_count < pte_total; pte_count++) {
761 		u32 pte = pte_addr[pte_count];
762 		if (!rk_pte_is_page_valid(pte))
763 			break;
764 
765 		pte_addr[pte_count] = rk_mk_pte_invalid(pte);
766 	}
767 
768 	rk_table_flush(rk_domain, pte_dma, pte_count);
769 
770 	return pte_count * SPAGE_SIZE;
771 }
772 
773 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
774 			     dma_addr_t pte_dma, dma_addr_t iova,
775 			     phys_addr_t paddr, size_t size, int prot)
776 {
777 	unsigned int pte_count;
778 	unsigned int pte_total = size / SPAGE_SIZE;
779 	phys_addr_t page_phys;
780 
781 	assert_spin_locked(&rk_domain->dt_lock);
782 
783 	for (pte_count = 0; pte_count < pte_total; pte_count++) {
784 		u32 pte = pte_addr[pte_count];
785 
786 		if (rk_pte_is_page_valid(pte))
787 			goto unwind;
788 
789 		pte_addr[pte_count] = rk_ops->mk_ptentries(paddr, prot);
790 
791 		paddr += SPAGE_SIZE;
792 	}
793 
794 	rk_table_flush(rk_domain, pte_dma, pte_total);
795 
796 	/*
797 	 * Zap the first and last iova to evict from iotlb any previously
798 	 * mapped cachelines holding stale values for its dte and pte.
799 	 * We only zap the first and last iova, since only they could have
800 	 * dte or pte shared with an existing mapping.
801 	 */
802 	rk_iommu_zap_iova_first_last(rk_domain, iova, size);
803 
804 	return 0;
805 unwind:
806 	/* Unmap the range of iovas that we just mapped */
807 	rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma,
808 			    pte_count * SPAGE_SIZE);
809 
810 	iova += pte_count * SPAGE_SIZE;
811 	page_phys = rk_ops->pt_address(pte_addr[pte_count]);
812 	pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
813 	       &iova, &page_phys, &paddr, prot);
814 
815 	return -EADDRINUSE;
816 }
817 
818 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
819 			phys_addr_t paddr, size_t size, size_t count,
820 			int prot, gfp_t gfp, size_t *mapped)
821 {
822 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
823 	unsigned long flags;
824 	dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
825 	u32 *page_table, *pte_addr;
826 	u32 dte_index, pte_index;
827 	int ret;
828 
829 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
830 
831 	/*
832 	 * pgsize_bitmap specifies iova sizes that fit in one page table
833 	 * (1024 4-KiB pages = 4 MiB).
834 	 * So, size will always be 4096 <= size <= 4194304.
835 	 * Since iommu_map() guarantees that both iova and size will be
836 	 * aligned, we will always only be mapping from a single dte here.
837 	 */
838 	page_table = rk_dte_get_page_table(rk_domain, iova);
839 	if (IS_ERR(page_table)) {
840 		spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
841 		return PTR_ERR(page_table);
842 	}
843 
844 	dte_index = rk_domain->dt[rk_iova_dte_index(iova)];
845 	pte_index = rk_iova_pte_index(iova);
846 	pte_addr = &page_table[pte_index];
847 
848 	pte_dma = rk_ops->pt_address(dte_index) + pte_index * sizeof(u32);
849 	ret = rk_iommu_map_iova(rk_domain, pte_addr, pte_dma, iova,
850 				paddr, size, prot);
851 
852 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
853 	if (!ret)
854 		*mapped = size;
855 
856 	return ret;
857 }
858 
859 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
860 			     size_t size, size_t count, struct iommu_iotlb_gather *gather)
861 {
862 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
863 	unsigned long flags;
864 	dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
865 	phys_addr_t pt_phys;
866 	u32 dte;
867 	u32 *pte_addr;
868 	size_t unmap_size;
869 
870 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
871 
872 	/*
873 	 * pgsize_bitmap specifies iova sizes that fit in one page table
874 	 * (1024 4-KiB pages = 4 MiB).
875 	 * So, size will always be 4096 <= size <= 4194304.
876 	 * Since iommu_unmap() guarantees that both iova and size will be
877 	 * aligned, we will always only be unmapping from a single dte here.
878 	 */
879 	dte = rk_domain->dt[rk_iova_dte_index(iova)];
880 	/* Just return 0 if iova is unmapped */
881 	if (!rk_dte_is_pt_valid(dte)) {
882 		spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
883 		return 0;
884 	}
885 
886 	pt_phys = rk_ops->pt_address(dte);
887 	pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
888 	pte_dma = pt_phys + rk_iova_pte_index(iova) * sizeof(u32);
889 	unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma, size);
890 
891 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
892 
893 	/* Shootdown iotlb entries for iova range that was just unmapped */
894 	rk_iommu_zap_iova(rk_domain, iova, unmap_size);
895 
896 	return unmap_size;
897 }
898 
899 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
900 {
901 	struct rk_iommudata *data = dev_iommu_priv_get(dev);
902 
903 	return data ? data->iommu : NULL;
904 }
905 
906 /* Must be called with iommu powered on and attached */
907 static void rk_iommu_disable(struct rk_iommu *iommu)
908 {
909 	int i;
910 
911 	/* Ignore error while disabling, just keep going */
912 	WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks));
913 	rk_iommu_enable_stall(iommu);
914 	rk_iommu_disable_paging(iommu);
915 	for (i = 0; i < iommu->num_mmu; i++) {
916 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, 0);
917 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, 0);
918 	}
919 	rk_iommu_disable_stall(iommu);
920 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
921 }
922 
923 /* Must be called with iommu powered on and attached */
924 static int rk_iommu_enable(struct rk_iommu *iommu)
925 {
926 	struct iommu_domain *domain = iommu->domain;
927 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
928 	int ret, i;
929 
930 	ret = clk_bulk_enable(iommu->num_clocks, iommu->clocks);
931 	if (ret)
932 		return ret;
933 
934 	ret = rk_iommu_enable_stall(iommu);
935 	if (ret)
936 		goto out_disable_clocks;
937 
938 	ret = rk_iommu_force_reset(iommu);
939 	if (ret)
940 		goto out_disable_stall;
941 
942 	for (i = 0; i < iommu->num_mmu; i++) {
943 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
944 			       rk_ops->mk_dtentries(rk_domain->dt_dma));
945 		rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
946 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
947 	}
948 
949 	ret = rk_iommu_enable_paging(iommu);
950 
951 out_disable_stall:
952 	rk_iommu_disable_stall(iommu);
953 out_disable_clocks:
954 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
955 	return ret;
956 }
957 
958 static int rk_iommu_identity_attach(struct iommu_domain *identity_domain,
959 				    struct device *dev)
960 {
961 	struct rk_iommu *iommu;
962 	struct rk_iommu_domain *rk_domain;
963 	unsigned long flags;
964 	int ret;
965 
966 	/* Allow 'virtual devices' (eg drm) to detach from domain */
967 	iommu = rk_iommu_from_dev(dev);
968 	if (!iommu)
969 		return -ENODEV;
970 
971 	rk_domain = to_rk_domain(iommu->domain);
972 
973 	dev_dbg(dev, "Detaching from iommu domain\n");
974 
975 	if (iommu->domain == identity_domain)
976 		return 0;
977 
978 	iommu->domain = identity_domain;
979 
980 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
981 	list_del_init(&iommu->node);
982 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
983 
984 	ret = pm_runtime_get_if_in_use(iommu->dev);
985 	WARN_ON_ONCE(ret < 0);
986 	if (ret > 0) {
987 		rk_iommu_disable(iommu);
988 		pm_runtime_put(iommu->dev);
989 	}
990 
991 	return 0;
992 }
993 
994 static struct iommu_domain_ops rk_identity_ops = {
995 	.attach_dev = rk_iommu_identity_attach,
996 };
997 
998 static struct iommu_domain rk_identity_domain = {
999 	.type = IOMMU_DOMAIN_IDENTITY,
1000 	.ops = &rk_identity_ops,
1001 };
1002 
1003 static int rk_iommu_attach_device(struct iommu_domain *domain,
1004 		struct device *dev)
1005 {
1006 	struct rk_iommu *iommu;
1007 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1008 	unsigned long flags;
1009 	int ret;
1010 
1011 	/*
1012 	 * Allow 'virtual devices' (e.g., drm) to attach to domain.
1013 	 * Such a device does not belong to an iommu group.
1014 	 */
1015 	iommu = rk_iommu_from_dev(dev);
1016 	if (!iommu)
1017 		return 0;
1018 
1019 	dev_dbg(dev, "Attaching to iommu domain\n");
1020 
1021 	/* iommu already attached */
1022 	if (iommu->domain == domain)
1023 		return 0;
1024 
1025 	ret = rk_iommu_identity_attach(&rk_identity_domain, dev);
1026 	if (ret)
1027 		return ret;
1028 
1029 	iommu->domain = domain;
1030 
1031 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
1032 	list_add_tail(&iommu->node, &rk_domain->iommus);
1033 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
1034 
1035 	ret = pm_runtime_get_if_in_use(iommu->dev);
1036 	if (!ret || WARN_ON_ONCE(ret < 0))
1037 		return 0;
1038 
1039 	ret = rk_iommu_enable(iommu);
1040 	if (ret)
1041 		WARN_ON(rk_iommu_identity_attach(&rk_identity_domain, dev));
1042 
1043 	pm_runtime_put(iommu->dev);
1044 
1045 	return ret;
1046 }
1047 
1048 static struct iommu_domain *rk_iommu_domain_alloc_paging(struct device *dev)
1049 {
1050 	struct rk_iommu_domain *rk_domain;
1051 
1052 	if (!dma_dev)
1053 		return NULL;
1054 
1055 	rk_domain = kzalloc(sizeof(*rk_domain), GFP_KERNEL);
1056 	if (!rk_domain)
1057 		return NULL;
1058 
1059 	/*
1060 	 * rk32xx iommus use a 2 level pagetable.
1061 	 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
1062 	 * Allocate one 4 KiB page for each table.
1063 	 */
1064 	rk_domain->dt = (u32 *)get_zeroed_page(GFP_KERNEL | rk_ops->gfp_flags);
1065 	if (!rk_domain->dt)
1066 		goto err_free_domain;
1067 
1068 	rk_domain->dt_dma = dma_map_single(dma_dev, rk_domain->dt,
1069 					   SPAGE_SIZE, DMA_TO_DEVICE);
1070 	if (dma_mapping_error(dma_dev, rk_domain->dt_dma)) {
1071 		dev_err(dma_dev, "DMA map error for DT\n");
1072 		goto err_free_dt;
1073 	}
1074 
1075 	spin_lock_init(&rk_domain->iommus_lock);
1076 	spin_lock_init(&rk_domain->dt_lock);
1077 	INIT_LIST_HEAD(&rk_domain->iommus);
1078 
1079 	rk_domain->domain.geometry.aperture_start = 0;
1080 	rk_domain->domain.geometry.aperture_end   = DMA_BIT_MASK(32);
1081 	rk_domain->domain.geometry.force_aperture = true;
1082 
1083 	return &rk_domain->domain;
1084 
1085 err_free_dt:
1086 	free_page((unsigned long)rk_domain->dt);
1087 err_free_domain:
1088 	kfree(rk_domain);
1089 
1090 	return NULL;
1091 }
1092 
1093 static void rk_iommu_domain_free(struct iommu_domain *domain)
1094 {
1095 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1096 	int i;
1097 
1098 	WARN_ON(!list_empty(&rk_domain->iommus));
1099 
1100 	for (i = 0; i < NUM_DT_ENTRIES; i++) {
1101 		u32 dte = rk_domain->dt[i];
1102 		if (rk_dte_is_pt_valid(dte)) {
1103 			phys_addr_t pt_phys = rk_ops->pt_address(dte);
1104 			u32 *page_table = phys_to_virt(pt_phys);
1105 			dma_unmap_single(dma_dev, pt_phys,
1106 					 SPAGE_SIZE, DMA_TO_DEVICE);
1107 			free_page((unsigned long)page_table);
1108 		}
1109 	}
1110 
1111 	dma_unmap_single(dma_dev, rk_domain->dt_dma,
1112 			 SPAGE_SIZE, DMA_TO_DEVICE);
1113 	free_page((unsigned long)rk_domain->dt);
1114 
1115 	kfree(rk_domain);
1116 }
1117 
1118 static struct iommu_device *rk_iommu_probe_device(struct device *dev)
1119 {
1120 	struct rk_iommudata *data;
1121 	struct rk_iommu *iommu;
1122 
1123 	data = dev_iommu_priv_get(dev);
1124 	if (!data)
1125 		return ERR_PTR(-ENODEV);
1126 
1127 	iommu = rk_iommu_from_dev(dev);
1128 
1129 	data->link = device_link_add(dev, iommu->dev,
1130 				     DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME);
1131 
1132 	return &iommu->iommu;
1133 }
1134 
1135 static void rk_iommu_release_device(struct device *dev)
1136 {
1137 	struct rk_iommudata *data = dev_iommu_priv_get(dev);
1138 
1139 	device_link_del(data->link);
1140 }
1141 
1142 static int rk_iommu_of_xlate(struct device *dev,
1143 			     struct of_phandle_args *args)
1144 {
1145 	struct platform_device *iommu_dev;
1146 	struct rk_iommudata *data;
1147 
1148 	data = devm_kzalloc(dma_dev, sizeof(*data), GFP_KERNEL);
1149 	if (!data)
1150 		return -ENOMEM;
1151 
1152 	iommu_dev = of_find_device_by_node(args->np);
1153 
1154 	data->iommu = platform_get_drvdata(iommu_dev);
1155 	data->iommu->domain = &rk_identity_domain;
1156 	dev_iommu_priv_set(dev, data);
1157 
1158 	platform_device_put(iommu_dev);
1159 
1160 	return 0;
1161 }
1162 
1163 static const struct iommu_ops rk_iommu_ops = {
1164 	.identity_domain = &rk_identity_domain,
1165 	.domain_alloc_paging = rk_iommu_domain_alloc_paging,
1166 	.probe_device = rk_iommu_probe_device,
1167 	.release_device = rk_iommu_release_device,
1168 	.device_group = generic_single_device_group,
1169 	.pgsize_bitmap = RK_IOMMU_PGSIZE_BITMAP,
1170 	.of_xlate = rk_iommu_of_xlate,
1171 	.default_domain_ops = &(const struct iommu_domain_ops) {
1172 		.attach_dev	= rk_iommu_attach_device,
1173 		.map_pages	= rk_iommu_map,
1174 		.unmap_pages	= rk_iommu_unmap,
1175 		.iova_to_phys	= rk_iommu_iova_to_phys,
1176 		.free		= rk_iommu_domain_free,
1177 	}
1178 };
1179 
1180 static int rk_iommu_probe(struct platform_device *pdev)
1181 {
1182 	struct device *dev = &pdev->dev;
1183 	struct rk_iommu *iommu;
1184 	struct resource *res;
1185 	const struct rk_iommu_ops *ops;
1186 	int num_res = pdev->num_resources;
1187 	int err, i;
1188 
1189 	iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
1190 	if (!iommu)
1191 		return -ENOMEM;
1192 
1193 	platform_set_drvdata(pdev, iommu);
1194 	iommu->dev = dev;
1195 	iommu->num_mmu = 0;
1196 
1197 	ops = of_device_get_match_data(dev);
1198 	if (!rk_ops)
1199 		rk_ops = ops;
1200 
1201 	/*
1202 	 * That should not happen unless different versions of the
1203 	 * hardware block are embedded the same SoC
1204 	 */
1205 	if (WARN_ON(rk_ops != ops))
1206 		return -EINVAL;
1207 
1208 	iommu->bases = devm_kcalloc(dev, num_res, sizeof(*iommu->bases),
1209 				    GFP_KERNEL);
1210 	if (!iommu->bases)
1211 		return -ENOMEM;
1212 
1213 	for (i = 0; i < num_res; i++) {
1214 		res = platform_get_resource(pdev, IORESOURCE_MEM, i);
1215 		if (!res)
1216 			continue;
1217 		iommu->bases[i] = devm_ioremap_resource(&pdev->dev, res);
1218 		if (IS_ERR(iommu->bases[i]))
1219 			continue;
1220 		iommu->num_mmu++;
1221 	}
1222 	if (iommu->num_mmu == 0)
1223 		return PTR_ERR(iommu->bases[0]);
1224 
1225 	iommu->num_irq = platform_irq_count(pdev);
1226 	if (iommu->num_irq < 0)
1227 		return iommu->num_irq;
1228 
1229 	iommu->reset_disabled = device_property_read_bool(dev,
1230 					"rockchip,disable-mmu-reset");
1231 
1232 	iommu->num_clocks = ARRAY_SIZE(rk_iommu_clocks);
1233 	iommu->clocks = devm_kcalloc(iommu->dev, iommu->num_clocks,
1234 				     sizeof(*iommu->clocks), GFP_KERNEL);
1235 	if (!iommu->clocks)
1236 		return -ENOMEM;
1237 
1238 	for (i = 0; i < iommu->num_clocks; ++i)
1239 		iommu->clocks[i].id = rk_iommu_clocks[i];
1240 
1241 	/*
1242 	 * iommu clocks should be present for all new devices and devicetrees
1243 	 * but there are older devicetrees without clocks out in the wild.
1244 	 * So clocks as optional for the time being.
1245 	 */
1246 	err = devm_clk_bulk_get(iommu->dev, iommu->num_clocks, iommu->clocks);
1247 	if (err == -ENOENT)
1248 		iommu->num_clocks = 0;
1249 	else if (err)
1250 		return err;
1251 
1252 	err = clk_bulk_prepare(iommu->num_clocks, iommu->clocks);
1253 	if (err)
1254 		return err;
1255 
1256 	err = iommu_device_sysfs_add(&iommu->iommu, dev, NULL, dev_name(dev));
1257 	if (err)
1258 		goto err_unprepare_clocks;
1259 
1260 	err = iommu_device_register(&iommu->iommu, &rk_iommu_ops, dev);
1261 	if (err)
1262 		goto err_remove_sysfs;
1263 
1264 	/*
1265 	 * Use the first registered IOMMU device for domain to use with DMA
1266 	 * API, since a domain might not physically correspond to a single
1267 	 * IOMMU device..
1268 	 */
1269 	if (!dma_dev)
1270 		dma_dev = &pdev->dev;
1271 
1272 	pm_runtime_enable(dev);
1273 
1274 	for (i = 0; i < iommu->num_irq; i++) {
1275 		int irq = platform_get_irq(pdev, i);
1276 
1277 		if (irq < 0) {
1278 			err = irq;
1279 			goto err_pm_disable;
1280 		}
1281 
1282 		err = devm_request_irq(iommu->dev, irq, rk_iommu_irq,
1283 				       IRQF_SHARED, dev_name(dev), iommu);
1284 		if (err)
1285 			goto err_pm_disable;
1286 	}
1287 
1288 	dma_set_mask_and_coherent(dev, rk_ops->dma_bit_mask);
1289 
1290 	return 0;
1291 err_pm_disable:
1292 	pm_runtime_disable(dev);
1293 err_remove_sysfs:
1294 	iommu_device_sysfs_remove(&iommu->iommu);
1295 err_unprepare_clocks:
1296 	clk_bulk_unprepare(iommu->num_clocks, iommu->clocks);
1297 	return err;
1298 }
1299 
1300 static void rk_iommu_shutdown(struct platform_device *pdev)
1301 {
1302 	struct rk_iommu *iommu = platform_get_drvdata(pdev);
1303 	int i;
1304 
1305 	for (i = 0; i < iommu->num_irq; i++) {
1306 		int irq = platform_get_irq(pdev, i);
1307 
1308 		devm_free_irq(iommu->dev, irq, iommu);
1309 	}
1310 
1311 	pm_runtime_force_suspend(&pdev->dev);
1312 }
1313 
1314 static int __maybe_unused rk_iommu_suspend(struct device *dev)
1315 {
1316 	struct rk_iommu *iommu = dev_get_drvdata(dev);
1317 
1318 	if (iommu->domain == &rk_identity_domain)
1319 		return 0;
1320 
1321 	rk_iommu_disable(iommu);
1322 	return 0;
1323 }
1324 
1325 static int __maybe_unused rk_iommu_resume(struct device *dev)
1326 {
1327 	struct rk_iommu *iommu = dev_get_drvdata(dev);
1328 
1329 	if (iommu->domain == &rk_identity_domain)
1330 		return 0;
1331 
1332 	return rk_iommu_enable(iommu);
1333 }
1334 
1335 static const struct dev_pm_ops rk_iommu_pm_ops = {
1336 	SET_RUNTIME_PM_OPS(rk_iommu_suspend, rk_iommu_resume, NULL)
1337 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
1338 				pm_runtime_force_resume)
1339 };
1340 
1341 static struct rk_iommu_ops iommu_data_ops_v1 = {
1342 	.pt_address = &rk_dte_pt_address,
1343 	.mk_dtentries = &rk_mk_dte,
1344 	.mk_ptentries = &rk_mk_pte,
1345 	.dma_bit_mask = DMA_BIT_MASK(32),
1346 	.gfp_flags = GFP_DMA32,
1347 };
1348 
1349 static struct rk_iommu_ops iommu_data_ops_v2 = {
1350 	.pt_address = &rk_dte_pt_address_v2,
1351 	.mk_dtentries = &rk_mk_dte_v2,
1352 	.mk_ptentries = &rk_mk_pte_v2,
1353 	.dma_bit_mask = DMA_BIT_MASK(40),
1354 	.gfp_flags = 0,
1355 };
1356 
1357 static const struct of_device_id rk_iommu_dt_ids[] = {
1358 	{	.compatible = "rockchip,iommu",
1359 		.data = &iommu_data_ops_v1,
1360 	},
1361 	{	.compatible = "rockchip,rk3568-iommu",
1362 		.data = &iommu_data_ops_v2,
1363 	},
1364 	{ /* sentinel */ }
1365 };
1366 
1367 static struct platform_driver rk_iommu_driver = {
1368 	.probe = rk_iommu_probe,
1369 	.shutdown = rk_iommu_shutdown,
1370 	.driver = {
1371 		   .name = "rk_iommu",
1372 		   .of_match_table = rk_iommu_dt_ids,
1373 		   .pm = &rk_iommu_pm_ops,
1374 		   .suppress_bind_attrs = true,
1375 	},
1376 };
1377 builtin_platform_driver(rk_iommu_driver);
1378