xref: /linux/drivers/iommu/rockchip-iommu.c (revision 8d4346ecd4950ae08cc76a6de327c264e846758c)
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 #include <linux/string_choices.h>
29 
30 #include "iommu-pages.h"
31 
32 /** MMU register offsets */
33 #define RK_MMU_DTE_ADDR		0x00	/* Directory table address */
34 #define RK_MMU_STATUS		0x04
35 #define RK_MMU_COMMAND		0x08
36 #define RK_MMU_PAGE_FAULT_ADDR	0x0C	/* IOVA of last page fault */
37 #define RK_MMU_ZAP_ONE_LINE	0x10	/* Shootdown one IOTLB entry */
38 #define RK_MMU_INT_RAWSTAT	0x14	/* IRQ status ignoring mask */
39 #define RK_MMU_INT_CLEAR	0x18	/* Acknowledge and re-arm irq */
40 #define RK_MMU_INT_MASK		0x1C	/* IRQ enable */
41 #define RK_MMU_INT_STATUS	0x20	/* IRQ status after masking */
42 #define RK_MMU_AUTO_GATING	0x24
43 
44 #define DTE_ADDR_DUMMY		0xCAFEBABE
45 
46 #define RK_MMU_POLL_PERIOD_US		100
47 #define RK_MMU_FORCE_RESET_TIMEOUT_US	100000
48 #define RK_MMU_POLL_TIMEOUT_US		1000
49 
50 /* RK_MMU_STATUS fields */
51 #define RK_MMU_STATUS_PAGING_ENABLED       BIT(0)
52 #define RK_MMU_STATUS_PAGE_FAULT_ACTIVE    BIT(1)
53 #define RK_MMU_STATUS_STALL_ACTIVE         BIT(2)
54 #define RK_MMU_STATUS_IDLE                 BIT(3)
55 #define RK_MMU_STATUS_REPLAY_BUFFER_EMPTY  BIT(4)
56 #define RK_MMU_STATUS_PAGE_FAULT_IS_WRITE  BIT(5)
57 #define RK_MMU_STATUS_STALL_NOT_ACTIVE     BIT(31)
58 
59 /* RK_MMU_COMMAND command values */
60 #define RK_MMU_CMD_ENABLE_PAGING    0  /* Enable memory translation */
61 #define RK_MMU_CMD_DISABLE_PAGING   1  /* Disable memory translation */
62 #define RK_MMU_CMD_ENABLE_STALL     2  /* Stall paging to allow other cmds */
63 #define RK_MMU_CMD_DISABLE_STALL    3  /* Stop stall re-enables paging */
64 #define RK_MMU_CMD_ZAP_CACHE        4  /* Shoot down entire IOTLB */
65 #define RK_MMU_CMD_PAGE_FAULT_DONE  5  /* Clear page fault */
66 #define RK_MMU_CMD_FORCE_RESET      6  /* Reset all registers */
67 
68 /* RK_MMU_INT_* register fields */
69 #define RK_MMU_IRQ_PAGE_FAULT    0x01  /* page fault */
70 #define RK_MMU_IRQ_BUS_ERROR     0x02  /* bus read error */
71 #define RK_MMU_IRQ_MASK          (RK_MMU_IRQ_PAGE_FAULT | RK_MMU_IRQ_BUS_ERROR)
72 
73 #define NUM_DT_ENTRIES 1024
74 #define NUM_PT_ENTRIES 1024
75 
76 #define SPAGE_ORDER 12
77 #define SPAGE_SIZE (1 << SPAGE_ORDER)
78 
79 #define DISABLE_FETCH_DTE_TIME_LIMIT BIT(31)
80 
81  /*
82   * Support mapping any size that fits in one page table:
83   *   4 KiB to 4 MiB
84   */
85 #define RK_IOMMU_PGSIZE_BITMAP 0x007ff000
86 
87 struct rk_iommu_domain {
88 	struct list_head iommus;
89 	u32 *dt; /* page directory table */
90 	dma_addr_t dt_dma;
91 	spinlock_t iommus_lock; /* lock for iommus list */
92 	spinlock_t dt_lock; /* lock for modifying page directory table */
93 	struct device *dma_dev;
94 
95 	struct iommu_domain domain;
96 };
97 
98 /* list of clocks required by IOMMU */
99 static const char * const rk_iommu_clocks[] = {
100 	"aclk", "iface",
101 };
102 
103 struct rk_iommu_ops {
104 	phys_addr_t (*pt_address)(u32 dte);
105 	u32 (*mk_dtentries)(dma_addr_t pt_dma);
106 	u32 (*mk_ptentries)(phys_addr_t page, int prot);
107 	u64 dma_bit_mask;
108 	gfp_t gfp_flags;
109 };
110 
111 struct rk_iommu {
112 	struct device *dev;
113 	void __iomem **bases;
114 	int num_mmu;
115 	int num_irq;
116 	struct clk_bulk_data *clocks;
117 	int num_clocks;
118 	bool reset_disabled;
119 	struct iommu_device iommu;
120 	struct list_head node; /* entry in rk_iommu_domain.iommus */
121 	struct iommu_domain *domain; /* domain to which iommu is attached */
122 };
123 
124 struct rk_iommudata {
125 	struct device_link *link; /* runtime PM link from IOMMU to master */
126 	struct rk_iommu *iommu;
127 };
128 
129 static const struct rk_iommu_ops *rk_ops;
130 static struct iommu_domain rk_identity_domain;
131 
132 static inline void rk_table_flush(struct rk_iommu_domain *dom, dma_addr_t dma,
133 				  unsigned int count)
134 {
135 	size_t size = count * sizeof(u32); /* count of u32 entry */
136 
137 	dma_sync_single_for_device(dom->dma_dev, dma, size, DMA_TO_DEVICE);
138 }
139 
140 static struct rk_iommu_domain *to_rk_domain(struct iommu_domain *dom)
141 {
142 	return container_of(dom, struct rk_iommu_domain, domain);
143 }
144 
145 /*
146  * The Rockchip rk3288 iommu uses a 2-level page table.
147  * The first level is the "Directory Table" (DT).
148  * The DT consists of 1024 4-byte Directory Table Entries (DTEs), each pointing
149  * to a "Page Table".
150  * The second level is the 1024 Page Tables (PT).
151  * Each PT consists of 1024 4-byte Page Table Entries (PTEs), each pointing to
152  * a 4 KB page of physical memory.
153  *
154  * The DT and each PT fits in a single 4 KB page (4-bytes * 1024 entries).
155  * Each iommu device has a MMU_DTE_ADDR register that contains the physical
156  * address of the start of the DT page.
157  *
158  * The structure of the page table is as follows:
159  *
160  *                   DT
161  * MMU_DTE_ADDR -> +-----+
162  *                 |     |
163  *                 +-----+     PT
164  *                 | DTE | -> +-----+
165  *                 +-----+    |     |     Memory
166  *                 |     |    +-----+     Page
167  *                 |     |    | PTE | -> +-----+
168  *                 +-----+    +-----+    |     |
169  *                            |     |    |     |
170  *                            |     |    |     |
171  *                            +-----+    |     |
172  *                                       |     |
173  *                                       |     |
174  *                                       +-----+
175  */
176 
177 /*
178  * Each DTE has a PT address and a valid bit:
179  * +---------------------+-----------+-+
180  * | PT address          | Reserved  |V|
181  * +---------------------+-----------+-+
182  *  31:12 - PT address (PTs always starts on a 4 KB boundary)
183  *  11: 1 - Reserved
184  *      0 - 1 if PT @ PT address is valid
185  */
186 #define RK_DTE_PT_ADDRESS_MASK    0xfffff000
187 #define RK_DTE_PT_VALID           BIT(0)
188 
189 static inline phys_addr_t rk_dte_pt_address(u32 dte)
190 {
191 	return (phys_addr_t)dte & RK_DTE_PT_ADDRESS_MASK;
192 }
193 
194 /*
195  * In v2:
196  * 31:12 - PT address bit 31:0
197  * 11: 8 - PT address bit 35:32
198  *  7: 4 - PT address bit 39:36
199  *  3: 1 - Reserved
200  *     0 - 1 if PT @ PT address is valid
201  */
202 #define RK_DTE_PT_ADDRESS_MASK_V2 GENMASK_ULL(31, 4)
203 #define DTE_HI_MASK1	GENMASK(11, 8)
204 #define DTE_HI_MASK2	GENMASK(7, 4)
205 #define DTE_HI_SHIFT1	24 /* shift bit 8 to bit 32 */
206 #define DTE_HI_SHIFT2	32 /* shift bit 4 to bit 36 */
207 #define PAGE_DESC_HI_MASK1	GENMASK_ULL(35, 32)
208 #define PAGE_DESC_HI_MASK2	GENMASK_ULL(39, 36)
209 
210 static inline phys_addr_t rk_dte_pt_address_v2(u32 dte)
211 {
212 	u64 dte_v2 = dte;
213 
214 	dte_v2 = ((dte_v2 & DTE_HI_MASK2) << DTE_HI_SHIFT2) |
215 		 ((dte_v2 & DTE_HI_MASK1) << DTE_HI_SHIFT1) |
216 		 (dte_v2 & RK_DTE_PT_ADDRESS_MASK);
217 
218 	return (phys_addr_t)dte_v2;
219 }
220 
221 static inline bool rk_dte_is_pt_valid(u32 dte)
222 {
223 	return dte & RK_DTE_PT_VALID;
224 }
225 
226 static inline u32 rk_mk_dte(dma_addr_t pt_dma)
227 {
228 	return (pt_dma & RK_DTE_PT_ADDRESS_MASK) | RK_DTE_PT_VALID;
229 }
230 
231 static inline u32 rk_mk_dte_v2(dma_addr_t pt_dma)
232 {
233 	pt_dma = (pt_dma & RK_DTE_PT_ADDRESS_MASK) |
234 		 ((pt_dma & PAGE_DESC_HI_MASK1) >> DTE_HI_SHIFT1) |
235 		 (pt_dma & PAGE_DESC_HI_MASK2) >> DTE_HI_SHIFT2;
236 
237 	return (pt_dma & RK_DTE_PT_ADDRESS_MASK_V2) | RK_DTE_PT_VALID;
238 }
239 
240 /*
241  * Each PTE has a Page address, some flags and a valid bit:
242  * +---------------------+---+-------+-+
243  * | Page address        |Rsv| Flags |V|
244  * +---------------------+---+-------+-+
245  *  31:12 - Page address (Pages always start on a 4 KB boundary)
246  *  11: 9 - Reserved
247  *   8: 1 - Flags
248  *      8 - Read allocate - allocate cache space on read misses
249  *      7 - Read cache - enable cache & prefetch of data
250  *      6 - Write buffer - enable delaying writes on their way to memory
251  *      5 - Write allocate - allocate cache space on write misses
252  *      4 - Write cache - different writes can be merged together
253  *      3 - Override cache attributes
254  *          if 1, bits 4-8 control cache attributes
255  *          if 0, the system bus defaults are used
256  *      2 - Writable
257  *      1 - Readable
258  *      0 - 1 if Page @ Page address is valid
259  */
260 #define RK_PTE_PAGE_ADDRESS_MASK  0xfffff000
261 #define RK_PTE_PAGE_FLAGS_MASK    0x000001fe
262 #define RK_PTE_PAGE_WRITABLE      BIT(2)
263 #define RK_PTE_PAGE_READABLE      BIT(1)
264 #define RK_PTE_PAGE_VALID         BIT(0)
265 
266 static inline bool rk_pte_is_page_valid(u32 pte)
267 {
268 	return pte & RK_PTE_PAGE_VALID;
269 }
270 
271 /* TODO: set cache flags per prot IOMMU_CACHE */
272 static u32 rk_mk_pte(phys_addr_t page, int prot)
273 {
274 	u32 flags = 0;
275 	flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
276 	flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
277 	page &= RK_PTE_PAGE_ADDRESS_MASK;
278 	return page | flags | RK_PTE_PAGE_VALID;
279 }
280 
281 /*
282  * In v2:
283  * 31:12 - Page address bit 31:0
284  * 11: 8 - Page address bit 35:32
285  *  7: 4 - Page address bit 39:36
286  *     3 - Security
287  *     2 - Writable
288  *     1 - Readable
289  *     0 - 1 if Page @ Page address is valid
290  */
291 
292 static u32 rk_mk_pte_v2(phys_addr_t page, int prot)
293 {
294 	u32 flags = 0;
295 
296 	flags |= (prot & IOMMU_READ) ? RK_PTE_PAGE_READABLE : 0;
297 	flags |= (prot & IOMMU_WRITE) ? RK_PTE_PAGE_WRITABLE : 0;
298 
299 	return rk_mk_dte_v2(page) | flags;
300 }
301 
302 static u32 rk_mk_pte_invalid(u32 pte)
303 {
304 	return pte & ~RK_PTE_PAGE_VALID;
305 }
306 
307 /*
308  * rk3288 iova (IOMMU Virtual Address) format
309  *  31       22.21       12.11          0
310  * +-----------+-----------+-------------+
311  * | DTE index | PTE index | Page offset |
312  * +-----------+-----------+-------------+
313  *  31:22 - DTE index   - index of DTE in DT
314  *  21:12 - PTE index   - index of PTE in PT @ DTE.pt_address
315  *  11: 0 - Page offset - offset into page @ PTE.page_address
316  */
317 #define RK_IOVA_DTE_MASK    0xffc00000
318 #define RK_IOVA_DTE_SHIFT   22
319 #define RK_IOVA_PTE_MASK    0x003ff000
320 #define RK_IOVA_PTE_SHIFT   12
321 #define RK_IOVA_PAGE_MASK   0x00000fff
322 #define RK_IOVA_PAGE_SHIFT  0
323 
324 static u32 rk_iova_dte_index(dma_addr_t iova)
325 {
326 	return (u32)(iova & RK_IOVA_DTE_MASK) >> RK_IOVA_DTE_SHIFT;
327 }
328 
329 static u32 rk_iova_pte_index(dma_addr_t iova)
330 {
331 	return (u32)(iova & RK_IOVA_PTE_MASK) >> RK_IOVA_PTE_SHIFT;
332 }
333 
334 static u32 rk_iova_page_offset(dma_addr_t iova)
335 {
336 	return (u32)(iova & RK_IOVA_PAGE_MASK) >> RK_IOVA_PAGE_SHIFT;
337 }
338 
339 static u32 rk_iommu_read(void __iomem *base, u32 offset)
340 {
341 	return readl(base + offset);
342 }
343 
344 static void rk_iommu_write(void __iomem *base, u32 offset, u32 value)
345 {
346 	writel(value, base + offset);
347 }
348 
349 static void rk_iommu_command(struct rk_iommu *iommu, u32 command)
350 {
351 	int i;
352 
353 	for (i = 0; i < iommu->num_mmu; i++)
354 		writel(command, iommu->bases[i] + RK_MMU_COMMAND);
355 }
356 
357 static void rk_iommu_base_command(void __iomem *base, u32 command)
358 {
359 	writel(command, base + RK_MMU_COMMAND);
360 }
361 static void rk_iommu_zap_lines(struct rk_iommu *iommu, dma_addr_t iova_start,
362 			       size_t size)
363 {
364 	int i;
365 	dma_addr_t iova_end = iova_start + size;
366 	/*
367 	 * TODO(djkurtz): Figure out when it is more efficient to shootdown the
368 	 * entire iotlb rather than iterate over individual iovas.
369 	 */
370 	for (i = 0; i < iommu->num_mmu; i++) {
371 		dma_addr_t iova;
372 
373 		for (iova = iova_start; iova < iova_end; iova += SPAGE_SIZE)
374 			rk_iommu_write(iommu->bases[i], RK_MMU_ZAP_ONE_LINE, iova);
375 	}
376 }
377 
378 static bool rk_iommu_is_stall_active(struct rk_iommu *iommu)
379 {
380 	bool active = true;
381 	int i;
382 
383 	for (i = 0; i < iommu->num_mmu; i++)
384 		active &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
385 					   RK_MMU_STATUS_STALL_ACTIVE);
386 
387 	return active;
388 }
389 
390 static bool rk_iommu_is_paging_enabled(struct rk_iommu *iommu)
391 {
392 	bool enable = true;
393 	int i;
394 
395 	for (i = 0; i < iommu->num_mmu; i++)
396 		enable &= !!(rk_iommu_read(iommu->bases[i], RK_MMU_STATUS) &
397 					   RK_MMU_STATUS_PAGING_ENABLED);
398 
399 	return enable;
400 }
401 
402 static bool rk_iommu_is_reset_done(struct rk_iommu *iommu)
403 {
404 	bool done = true;
405 	int i;
406 
407 	for (i = 0; i < iommu->num_mmu; i++)
408 		done &= rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR) == 0;
409 
410 	return done;
411 }
412 
413 static int rk_iommu_enable_stall(struct rk_iommu *iommu)
414 {
415 	int ret, i;
416 	bool val;
417 
418 	if (rk_iommu_is_stall_active(iommu))
419 		return 0;
420 
421 	/* Stall can only be enabled if paging is enabled */
422 	if (!rk_iommu_is_paging_enabled(iommu))
423 		return 0;
424 
425 	rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_STALL);
426 
427 	ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
428 				 val, RK_MMU_POLL_PERIOD_US,
429 				 RK_MMU_POLL_TIMEOUT_US);
430 	if (ret)
431 		for (i = 0; i < iommu->num_mmu; i++)
432 			dev_err(iommu->dev, "Enable stall request timed out, status: %#08x\n",
433 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
434 
435 	return ret;
436 }
437 
438 static int rk_iommu_disable_stall(struct rk_iommu *iommu)
439 {
440 	int ret, i;
441 	bool val;
442 
443 	if (!rk_iommu_is_stall_active(iommu))
444 		return 0;
445 
446 	rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_STALL);
447 
448 	ret = readx_poll_timeout(rk_iommu_is_stall_active, iommu, val,
449 				 !val, RK_MMU_POLL_PERIOD_US,
450 				 RK_MMU_POLL_TIMEOUT_US);
451 	if (ret)
452 		for (i = 0; i < iommu->num_mmu; i++)
453 			dev_err(iommu->dev, "Disable stall request timed out, status: %#08x\n",
454 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
455 
456 	return ret;
457 }
458 
459 static int rk_iommu_enable_paging(struct rk_iommu *iommu)
460 {
461 	int ret, i;
462 	bool val;
463 
464 	if (rk_iommu_is_paging_enabled(iommu))
465 		return 0;
466 
467 	rk_iommu_command(iommu, RK_MMU_CMD_ENABLE_PAGING);
468 
469 	ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
470 				 val, RK_MMU_POLL_PERIOD_US,
471 				 RK_MMU_POLL_TIMEOUT_US);
472 	if (ret)
473 		for (i = 0; i < iommu->num_mmu; i++)
474 			dev_err(iommu->dev, "Enable paging request timed out, status: %#08x\n",
475 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
476 
477 	return ret;
478 }
479 
480 static int rk_iommu_disable_paging(struct rk_iommu *iommu)
481 {
482 	int ret, i;
483 	bool val;
484 
485 	if (!rk_iommu_is_paging_enabled(iommu))
486 		return 0;
487 
488 	rk_iommu_command(iommu, RK_MMU_CMD_DISABLE_PAGING);
489 
490 	ret = readx_poll_timeout(rk_iommu_is_paging_enabled, iommu, val,
491 				 !val, RK_MMU_POLL_PERIOD_US,
492 				 RK_MMU_POLL_TIMEOUT_US);
493 	if (ret)
494 		for (i = 0; i < iommu->num_mmu; i++)
495 			dev_err(iommu->dev, "Disable paging request timed out, status: %#08x\n",
496 				rk_iommu_read(iommu->bases[i], RK_MMU_STATUS));
497 
498 	return ret;
499 }
500 
501 static int rk_iommu_force_reset(struct rk_iommu *iommu)
502 {
503 	int ret, i;
504 	u32 dte_addr;
505 	bool val;
506 
507 	if (iommu->reset_disabled)
508 		return 0;
509 
510 	/*
511 	 * Check if register DTE_ADDR is working by writing DTE_ADDR_DUMMY
512 	 * and verifying that upper 5 (v1) or 7 (v2) nybbles are read back.
513 	 */
514 	for (i = 0; i < iommu->num_mmu; i++) {
515 		dte_addr = rk_ops->pt_address(DTE_ADDR_DUMMY);
516 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, dte_addr);
517 
518 		if (dte_addr != rk_iommu_read(iommu->bases[i], RK_MMU_DTE_ADDR)) {
519 			dev_err(iommu->dev, "Error during raw reset. MMU_DTE_ADDR is not functioning\n");
520 			return -EFAULT;
521 		}
522 	}
523 
524 	rk_iommu_command(iommu, RK_MMU_CMD_FORCE_RESET);
525 
526 	ret = readx_poll_timeout(rk_iommu_is_reset_done, iommu, val,
527 				 val, RK_MMU_FORCE_RESET_TIMEOUT_US,
528 				 RK_MMU_POLL_TIMEOUT_US);
529 	if (ret) {
530 		dev_err(iommu->dev, "FORCE_RESET command timed out\n");
531 		return ret;
532 	}
533 
534 	return 0;
535 }
536 
537 static void log_iova(struct rk_iommu *iommu, int index, dma_addr_t iova)
538 {
539 	void __iomem *base = iommu->bases[index];
540 	u32 dte_index, pte_index, page_offset;
541 	u32 mmu_dte_addr;
542 	phys_addr_t mmu_dte_addr_phys, dte_addr_phys;
543 	u32 *dte_addr;
544 	u32 dte;
545 	phys_addr_t pte_addr_phys = 0;
546 	u32 *pte_addr = NULL;
547 	u32 pte = 0;
548 	phys_addr_t page_addr_phys = 0;
549 	u32 page_flags = 0;
550 
551 	dte_index = rk_iova_dte_index(iova);
552 	pte_index = rk_iova_pte_index(iova);
553 	page_offset = rk_iova_page_offset(iova);
554 
555 	mmu_dte_addr = rk_iommu_read(base, RK_MMU_DTE_ADDR);
556 	mmu_dte_addr_phys = rk_ops->pt_address(mmu_dte_addr);
557 
558 	dte_addr_phys = mmu_dte_addr_phys + (4 * dte_index);
559 	dte_addr = phys_to_virt(dte_addr_phys);
560 	dte = *dte_addr;
561 
562 	if (!rk_dte_is_pt_valid(dte))
563 		goto print_it;
564 
565 	pte_addr_phys = rk_ops->pt_address(dte) + (pte_index * 4);
566 	pte_addr = phys_to_virt(pte_addr_phys);
567 	pte = *pte_addr;
568 
569 	if (!rk_pte_is_page_valid(pte))
570 		goto print_it;
571 
572 	page_addr_phys = rk_ops->pt_address(pte) + page_offset;
573 	page_flags = pte & RK_PTE_PAGE_FLAGS_MASK;
574 
575 print_it:
576 	dev_err(iommu->dev, "iova = %pad: dte_index: %#03x pte_index: %#03x page_offset: %#03x\n",
577 		&iova, dte_index, pte_index, page_offset);
578 	dev_err(iommu->dev, "mmu_dte_addr: %pa dte@%pa: %#08x valid: %u pte@%pa: %#08x valid: %u page@%pa flags: %#03x\n",
579 		&mmu_dte_addr_phys, &dte_addr_phys, dte,
580 		rk_dte_is_pt_valid(dte), &pte_addr_phys, pte,
581 		rk_pte_is_page_valid(pte), &page_addr_phys, page_flags);
582 }
583 
584 static irqreturn_t rk_iommu_irq(int irq, void *dev_id)
585 {
586 	struct rk_iommu *iommu = dev_id;
587 	u32 status;
588 	u32 int_status;
589 	dma_addr_t iova;
590 	irqreturn_t ret = IRQ_NONE;
591 	int i, err;
592 
593 	err = pm_runtime_get_if_in_use(iommu->dev);
594 	if (!err || WARN_ON_ONCE(err < 0))
595 		return ret;
596 
597 	if (WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks)))
598 		goto out;
599 
600 	for (i = 0; i < iommu->num_mmu; i++) {
601 		int_status = rk_iommu_read(iommu->bases[i], RK_MMU_INT_STATUS);
602 		if (int_status == 0)
603 			continue;
604 
605 		ret = IRQ_HANDLED;
606 		iova = rk_iommu_read(iommu->bases[i], RK_MMU_PAGE_FAULT_ADDR);
607 
608 		if (int_status & RK_MMU_IRQ_PAGE_FAULT) {
609 			int flags;
610 
611 			status = rk_iommu_read(iommu->bases[i], RK_MMU_STATUS);
612 			flags = (status & RK_MMU_STATUS_PAGE_FAULT_IS_WRITE) ?
613 					IOMMU_FAULT_WRITE : IOMMU_FAULT_READ;
614 
615 			dev_err(iommu->dev, "Page fault at %pad of type %s\n",
616 				&iova,
617 				str_write_read(flags == IOMMU_FAULT_WRITE));
618 
619 			log_iova(iommu, i, iova);
620 
621 			/*
622 			 * Report page fault to any installed handlers.
623 			 * Ignore the return code, though, since we always zap cache
624 			 * and clear the page fault anyway.
625 			 */
626 			if (iommu->domain != &rk_identity_domain)
627 				report_iommu_fault(iommu->domain, iommu->dev, iova,
628 						   flags);
629 			else
630 				dev_err(iommu->dev, "Page fault while iommu not attached to domain?\n");
631 
632 			rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
633 			rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_PAGE_FAULT_DONE);
634 		}
635 
636 		if (int_status & RK_MMU_IRQ_BUS_ERROR)
637 			dev_err(iommu->dev, "BUS_ERROR occurred at %pad\n", &iova);
638 
639 		if (int_status & ~RK_MMU_IRQ_MASK)
640 			dev_err(iommu->dev, "unexpected int_status: %#08x\n",
641 				int_status);
642 
643 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_CLEAR, int_status);
644 	}
645 
646 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
647 
648 out:
649 	pm_runtime_put(iommu->dev);
650 	return ret;
651 }
652 
653 static phys_addr_t rk_iommu_iova_to_phys(struct iommu_domain *domain,
654 					 dma_addr_t iova)
655 {
656 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
657 	unsigned long flags;
658 	phys_addr_t pt_phys, phys = 0;
659 	u32 dte, pte;
660 	u32 *page_table;
661 
662 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
663 
664 	dte = rk_domain->dt[rk_iova_dte_index(iova)];
665 	if (!rk_dte_is_pt_valid(dte))
666 		goto out;
667 
668 	pt_phys = rk_ops->pt_address(dte);
669 	page_table = (u32 *)phys_to_virt(pt_phys);
670 	pte = page_table[rk_iova_pte_index(iova)];
671 	if (!rk_pte_is_page_valid(pte))
672 		goto out;
673 
674 	phys = rk_ops->pt_address(pte) + rk_iova_page_offset(iova);
675 out:
676 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
677 
678 	return phys;
679 }
680 
681 static void rk_iommu_zap_iova(struct rk_iommu_domain *rk_domain,
682 			      dma_addr_t iova, size_t size)
683 {
684 	struct list_head *pos;
685 	unsigned long flags;
686 
687 	/* shootdown these iova from all iommus using this domain */
688 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
689 	list_for_each(pos, &rk_domain->iommus) {
690 		struct rk_iommu *iommu;
691 		int ret;
692 
693 		iommu = list_entry(pos, struct rk_iommu, node);
694 
695 		/* Only zap TLBs of IOMMUs that are powered on. */
696 		ret = pm_runtime_get_if_in_use(iommu->dev);
697 		if (WARN_ON_ONCE(ret < 0))
698 			continue;
699 		if (ret) {
700 			WARN_ON(clk_bulk_enable(iommu->num_clocks,
701 						iommu->clocks));
702 			rk_iommu_zap_lines(iommu, iova, size);
703 			clk_bulk_disable(iommu->num_clocks, iommu->clocks);
704 			pm_runtime_put(iommu->dev);
705 		}
706 	}
707 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
708 }
709 
710 static void rk_iommu_zap_iova_first_last(struct rk_iommu_domain *rk_domain,
711 					 dma_addr_t iova, size_t size)
712 {
713 	rk_iommu_zap_iova(rk_domain, iova, SPAGE_SIZE);
714 	if (size > SPAGE_SIZE)
715 		rk_iommu_zap_iova(rk_domain, iova + size - SPAGE_SIZE,
716 					SPAGE_SIZE);
717 }
718 
719 static u32 *rk_dte_get_page_table(struct rk_iommu_domain *rk_domain,
720 				  dma_addr_t iova)
721 {
722 	u32 *page_table, *dte_addr;
723 	u32 dte_index, dte;
724 	phys_addr_t pt_phys;
725 	dma_addr_t pt_dma;
726 
727 	assert_spin_locked(&rk_domain->dt_lock);
728 
729 	dte_index = rk_iova_dte_index(iova);
730 	dte_addr = &rk_domain->dt[dte_index];
731 	dte = *dte_addr;
732 	if (rk_dte_is_pt_valid(dte))
733 		goto done;
734 
735 	page_table = iommu_alloc_pages_sz(GFP_ATOMIC | rk_ops->gfp_flags,
736 					  SPAGE_SIZE);
737 	if (!page_table)
738 		return ERR_PTR(-ENOMEM);
739 
740 	pt_dma = dma_map_single(rk_domain->dma_dev, page_table, SPAGE_SIZE, DMA_TO_DEVICE);
741 	if (dma_mapping_error(rk_domain->dma_dev, pt_dma)) {
742 		dev_err(rk_domain->dma_dev, "DMA mapping error while allocating page table\n");
743 		iommu_free_pages(page_table);
744 		return ERR_PTR(-ENOMEM);
745 	}
746 
747 	dte = rk_ops->mk_dtentries(pt_dma);
748 	*dte_addr = dte;
749 
750 	rk_table_flush(rk_domain,
751 		       rk_domain->dt_dma + dte_index * sizeof(u32), 1);
752 done:
753 	pt_phys = rk_ops->pt_address(dte);
754 	return (u32 *)phys_to_virt(pt_phys);
755 }
756 
757 static size_t rk_iommu_unmap_iova(struct rk_iommu_domain *rk_domain,
758 				  u32 *pte_addr, dma_addr_t pte_dma,
759 				  size_t size)
760 {
761 	unsigned int pte_count;
762 	unsigned int pte_total = size / SPAGE_SIZE;
763 
764 	assert_spin_locked(&rk_domain->dt_lock);
765 
766 	for (pte_count = 0; pte_count < pte_total; pte_count++) {
767 		u32 pte = pte_addr[pte_count];
768 		if (!rk_pte_is_page_valid(pte))
769 			break;
770 
771 		pte_addr[pte_count] = rk_mk_pte_invalid(pte);
772 	}
773 
774 	rk_table_flush(rk_domain, pte_dma, pte_count);
775 
776 	return pte_count * SPAGE_SIZE;
777 }
778 
779 static int rk_iommu_map_iova(struct rk_iommu_domain *rk_domain, u32 *pte_addr,
780 			     dma_addr_t pte_dma, dma_addr_t iova,
781 			     phys_addr_t paddr, size_t size, int prot)
782 {
783 	unsigned int pte_count;
784 	unsigned int pte_total = size / SPAGE_SIZE;
785 	phys_addr_t page_phys;
786 
787 	assert_spin_locked(&rk_domain->dt_lock);
788 
789 	for (pte_count = 0; pte_count < pte_total; pte_count++) {
790 		u32 pte = pte_addr[pte_count];
791 
792 		if (rk_pte_is_page_valid(pte))
793 			goto unwind;
794 
795 		pte_addr[pte_count] = rk_ops->mk_ptentries(paddr, prot);
796 
797 		paddr += SPAGE_SIZE;
798 	}
799 
800 	rk_table_flush(rk_domain, pte_dma, pte_total);
801 
802 	/*
803 	 * Zap the first and last iova to evict from iotlb any previously
804 	 * mapped cachelines holding stale values for its dte and pte.
805 	 * We only zap the first and last iova, since only they could have
806 	 * dte or pte shared with an existing mapping.
807 	 */
808 	rk_iommu_zap_iova_first_last(rk_domain, iova, size);
809 
810 	return 0;
811 unwind:
812 	/* Unmap the range of iovas that we just mapped */
813 	rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma,
814 			    pte_count * SPAGE_SIZE);
815 
816 	iova += pte_count * SPAGE_SIZE;
817 	page_phys = rk_ops->pt_address(pte_addr[pte_count]);
818 	pr_err("iova: %pad already mapped to %pa cannot remap to phys: %pa prot: %#x\n",
819 	       &iova, &page_phys, &paddr, prot);
820 
821 	return -EADDRINUSE;
822 }
823 
824 static int rk_iommu_map(struct iommu_domain *domain, unsigned long _iova,
825 			phys_addr_t paddr, size_t size, size_t count,
826 			int prot, gfp_t gfp, size_t *mapped)
827 {
828 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
829 	unsigned long flags;
830 	dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
831 	u32 *page_table, *pte_addr;
832 	u32 dte_index, pte_index;
833 	int ret;
834 
835 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
836 
837 	/*
838 	 * pgsize_bitmap specifies iova sizes that fit in one page table
839 	 * (1024 4-KiB pages = 4 MiB).
840 	 * So, size will always be 4096 <= size <= 4194304.
841 	 * Since iommu_map() guarantees that both iova and size will be
842 	 * aligned, we will always only be mapping from a single dte here.
843 	 */
844 	page_table = rk_dte_get_page_table(rk_domain, iova);
845 	if (IS_ERR(page_table)) {
846 		spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
847 		return PTR_ERR(page_table);
848 	}
849 
850 	dte_index = rk_domain->dt[rk_iova_dte_index(iova)];
851 	pte_index = rk_iova_pte_index(iova);
852 	pte_addr = &page_table[pte_index];
853 
854 	pte_dma = rk_ops->pt_address(dte_index) + pte_index * sizeof(u32);
855 	ret = rk_iommu_map_iova(rk_domain, pte_addr, pte_dma, iova,
856 				paddr, size, prot);
857 
858 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
859 	if (!ret)
860 		*mapped = size;
861 
862 	return ret;
863 }
864 
865 static size_t rk_iommu_unmap(struct iommu_domain *domain, unsigned long _iova,
866 			     size_t size, size_t count, struct iommu_iotlb_gather *gather)
867 {
868 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
869 	unsigned long flags;
870 	dma_addr_t pte_dma, iova = (dma_addr_t)_iova;
871 	phys_addr_t pt_phys;
872 	u32 dte;
873 	u32 *pte_addr;
874 	size_t unmap_size;
875 
876 	spin_lock_irqsave(&rk_domain->dt_lock, flags);
877 
878 	/*
879 	 * pgsize_bitmap specifies iova sizes that fit in one page table
880 	 * (1024 4-KiB pages = 4 MiB).
881 	 * So, size will always be 4096 <= size <= 4194304.
882 	 * Since iommu_unmap() guarantees that both iova and size will be
883 	 * aligned, we will always only be unmapping from a single dte here.
884 	 */
885 	dte = rk_domain->dt[rk_iova_dte_index(iova)];
886 	/* Just return 0 if iova is unmapped */
887 	if (!rk_dte_is_pt_valid(dte)) {
888 		spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
889 		return 0;
890 	}
891 
892 	pt_phys = rk_ops->pt_address(dte);
893 	pte_addr = (u32 *)phys_to_virt(pt_phys) + rk_iova_pte_index(iova);
894 	pte_dma = pt_phys + rk_iova_pte_index(iova) * sizeof(u32);
895 	unmap_size = rk_iommu_unmap_iova(rk_domain, pte_addr, pte_dma, size);
896 
897 	spin_unlock_irqrestore(&rk_domain->dt_lock, flags);
898 
899 	/* Shootdown iotlb entries for iova range that was just unmapped */
900 	rk_iommu_zap_iova(rk_domain, iova, unmap_size);
901 
902 	return unmap_size;
903 }
904 
905 static struct rk_iommu *rk_iommu_from_dev(struct device *dev)
906 {
907 	struct rk_iommudata *data = dev_iommu_priv_get(dev);
908 
909 	return data ? data->iommu : NULL;
910 }
911 
912 /* Must be called with iommu powered on and attached */
913 static void rk_iommu_disable(struct rk_iommu *iommu)
914 {
915 	int i;
916 
917 	/* Ignore error while disabling, just keep going */
918 	WARN_ON(clk_bulk_enable(iommu->num_clocks, iommu->clocks));
919 	rk_iommu_enable_stall(iommu);
920 	rk_iommu_disable_paging(iommu);
921 	for (i = 0; i < iommu->num_mmu; i++) {
922 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, 0);
923 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR, 0);
924 	}
925 	rk_iommu_disable_stall(iommu);
926 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
927 }
928 
929 /* Must be called with iommu powered on and attached */
930 static int rk_iommu_enable(struct rk_iommu *iommu)
931 {
932 	struct iommu_domain *domain = iommu->domain;
933 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
934 	int ret, i;
935 	u32 auto_gate;
936 
937 	ret = clk_bulk_enable(iommu->num_clocks, iommu->clocks);
938 	if (ret)
939 		return ret;
940 
941 	ret = rk_iommu_enable_stall(iommu);
942 	if (ret)
943 		goto out_disable_clocks;
944 
945 	ret = rk_iommu_force_reset(iommu);
946 	if (ret)
947 		goto out_disable_stall;
948 
949 	for (i = 0; i < iommu->num_mmu; i++) {
950 		rk_iommu_write(iommu->bases[i], RK_MMU_DTE_ADDR,
951 			       rk_ops->mk_dtentries(rk_domain->dt_dma));
952 		rk_iommu_base_command(iommu->bases[i], RK_MMU_CMD_ZAP_CACHE);
953 		rk_iommu_write(iommu->bases[i], RK_MMU_INT_MASK, RK_MMU_IRQ_MASK);
954 
955 		/* Workaround for iommu blocked, BIT(31) default to 1 */
956 		auto_gate = rk_iommu_read(iommu->bases[i], RK_MMU_AUTO_GATING);
957 		auto_gate |= DISABLE_FETCH_DTE_TIME_LIMIT;
958 		rk_iommu_write(iommu->bases[i], RK_MMU_AUTO_GATING, auto_gate);
959 	}
960 
961 	ret = rk_iommu_enable_paging(iommu);
962 
963 out_disable_stall:
964 	rk_iommu_disable_stall(iommu);
965 out_disable_clocks:
966 	clk_bulk_disable(iommu->num_clocks, iommu->clocks);
967 	return ret;
968 }
969 
970 static int rk_iommu_identity_attach(struct iommu_domain *identity_domain,
971 				    struct device *dev,
972 				    struct iommu_domain *old)
973 {
974 	struct rk_iommu *iommu;
975 	struct rk_iommu_domain *rk_domain;
976 	unsigned long flags;
977 	int ret;
978 
979 	/* Allow 'virtual devices' (eg drm) to detach from domain */
980 	iommu = rk_iommu_from_dev(dev);
981 	if (!iommu)
982 		return -ENODEV;
983 
984 	rk_domain = to_rk_domain(iommu->domain);
985 
986 	dev_dbg(dev, "Detaching from iommu domain\n");
987 
988 	if (iommu->domain == identity_domain)
989 		return 0;
990 
991 	iommu->domain = identity_domain;
992 
993 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
994 	list_del_init(&iommu->node);
995 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
996 
997 	ret = pm_runtime_get_if_in_use(iommu->dev);
998 	WARN_ON_ONCE(ret < 0);
999 	if (ret > 0) {
1000 		rk_iommu_disable(iommu);
1001 		pm_runtime_put(iommu->dev);
1002 	}
1003 
1004 	return 0;
1005 }
1006 
1007 static struct iommu_domain_ops rk_identity_ops = {
1008 	.attach_dev = rk_iommu_identity_attach,
1009 };
1010 
1011 static struct iommu_domain rk_identity_domain = {
1012 	.type = IOMMU_DOMAIN_IDENTITY,
1013 	.ops = &rk_identity_ops,
1014 };
1015 
1016 static int rk_iommu_attach_device(struct iommu_domain *domain,
1017 				  struct device *dev, struct iommu_domain *old)
1018 {
1019 	struct rk_iommu *iommu;
1020 	struct rk_iommu_domain *rk_domain = to_rk_domain(domain);
1021 	unsigned long flags;
1022 	int ret;
1023 
1024 	/*
1025 	 * Allow 'virtual devices' (e.g., drm) to attach to domain.
1026 	 * Such a device does not belong to an iommu group.
1027 	 */
1028 	iommu = rk_iommu_from_dev(dev);
1029 	if (!iommu)
1030 		return 0;
1031 
1032 	dev_dbg(dev, "Attaching to iommu domain\n");
1033 
1034 	/* iommu already attached */
1035 	if (iommu->domain == domain)
1036 		return 0;
1037 
1038 	ret = rk_iommu_identity_attach(&rk_identity_domain, dev, old);
1039 	if (ret)
1040 		return ret;
1041 
1042 	iommu->domain = domain;
1043 
1044 	spin_lock_irqsave(&rk_domain->iommus_lock, flags);
1045 	list_add_tail(&iommu->node, &rk_domain->iommus);
1046 	spin_unlock_irqrestore(&rk_domain->iommus_lock, flags);
1047 
1048 	ret = pm_runtime_get_if_in_use(iommu->dev);
1049 	if (!ret || WARN_ON_ONCE(ret < 0))
1050 		return 0;
1051 
1052 	ret = rk_iommu_enable(iommu);
1053 	if (ret) {
1054 		/*
1055 		 * Note rk_iommu_identity_attach() might fail before physically
1056 		 * attaching the dev to iommu->domain, in which case the actual
1057 		 * old domain for this revert should be rk_identity_domain v.s.
1058 		 * iommu->domain. Since rk_iommu_identity_attach() does not care
1059 		 * about the old domain argument for now, this is not a problem.
1060 		 */
1061 		WARN_ON(rk_iommu_identity_attach(&rk_identity_domain, dev,
1062 						 iommu->domain));
1063 	}
1064 
1065 	pm_runtime_put(iommu->dev);
1066 
1067 	return ret;
1068 }
1069 
1070 static struct iommu_domain *rk_iommu_domain_alloc_paging(struct device *dev)
1071 {
1072 	struct rk_iommu_domain *rk_domain;
1073 	struct rk_iommu *iommu;
1074 
1075 	rk_domain = kzalloc_obj(*rk_domain);
1076 	if (!rk_domain)
1077 		return NULL;
1078 
1079 	/*
1080 	 * rk32xx iommus use a 2 level pagetable.
1081 	 * Each level1 (dt) and level2 (pt) table has 1024 4-byte entries.
1082 	 * Allocate one 4 KiB page for each table.
1083 	 */
1084 	rk_domain->dt = iommu_alloc_pages_sz(GFP_KERNEL | rk_ops->gfp_flags,
1085 					     SPAGE_SIZE);
1086 	if (!rk_domain->dt)
1087 		goto err_free_domain;
1088 
1089 	iommu = rk_iommu_from_dev(dev);
1090 	rk_domain->dma_dev = iommu->dev;
1091 	rk_domain->dt_dma = dma_map_single(rk_domain->dma_dev, rk_domain->dt,
1092 					   SPAGE_SIZE, DMA_TO_DEVICE);
1093 	if (dma_mapping_error(rk_domain->dma_dev, rk_domain->dt_dma)) {
1094 		dev_err(rk_domain->dma_dev, "DMA map error for DT\n");
1095 		goto err_free_dt;
1096 	}
1097 
1098 	spin_lock_init(&rk_domain->iommus_lock);
1099 	spin_lock_init(&rk_domain->dt_lock);
1100 	INIT_LIST_HEAD(&rk_domain->iommus);
1101 
1102 	rk_domain->domain.pgsize_bitmap = RK_IOMMU_PGSIZE_BITMAP;
1103 
1104 	rk_domain->domain.geometry.aperture_start = 0;
1105 	rk_domain->domain.geometry.aperture_end   = DMA_BIT_MASK(32);
1106 	rk_domain->domain.geometry.force_aperture = true;
1107 
1108 	return &rk_domain->domain;
1109 
1110 err_free_dt:
1111 	iommu_free_pages(rk_domain->dt);
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(rk_domain->dma_dev, pt_phys,
1131 					 SPAGE_SIZE, DMA_TO_DEVICE);
1132 			iommu_free_pages(page_table);
1133 		}
1134 	}
1135 
1136 	dma_unmap_single(rk_domain->dma_dev, rk_domain->dt_dma,
1137 			 SPAGE_SIZE, DMA_TO_DEVICE);
1138 	iommu_free_pages(rk_domain->dt);
1139 
1140 	kfree(rk_domain);
1141 }
1142 
1143 static struct iommu_device *rk_iommu_probe_device(struct device *dev)
1144 {
1145 	struct rk_iommudata *data;
1146 	struct rk_iommu *iommu;
1147 
1148 	data = dev_iommu_priv_get(dev);
1149 	if (!data)
1150 		return ERR_PTR(-ENODEV);
1151 
1152 	iommu = rk_iommu_from_dev(dev);
1153 
1154 	data->link = device_link_add(dev, iommu->dev,
1155 				     DL_FLAG_STATELESS | DL_FLAG_PM_RUNTIME);
1156 
1157 	return &iommu->iommu;
1158 }
1159 
1160 static void rk_iommu_release_device(struct device *dev)
1161 {
1162 	struct rk_iommudata *data = dev_iommu_priv_get(dev);
1163 
1164 	device_link_del(data->link);
1165 }
1166 
1167 static int rk_iommu_of_xlate(struct device *dev,
1168 			     const struct of_phandle_args *args)
1169 {
1170 	struct platform_device *iommu_dev;
1171 	struct rk_iommudata *data;
1172 
1173 	iommu_dev = of_find_device_by_node(args->np);
1174 
1175 	data = devm_kzalloc(&iommu_dev->dev, sizeof(*data), GFP_KERNEL);
1176 	if (!data)
1177 		return -ENOMEM;
1178 
1179 	data->iommu = platform_get_drvdata(iommu_dev);
1180 	dev_iommu_priv_set(dev, data);
1181 
1182 	platform_device_put(iommu_dev);
1183 
1184 	return 0;
1185 }
1186 
1187 static const struct iommu_ops rk_iommu_ops = {
1188 	.identity_domain = &rk_identity_domain,
1189 	.domain_alloc_paging = rk_iommu_domain_alloc_paging,
1190 	.probe_device = rk_iommu_probe_device,
1191 	.release_device = rk_iommu_release_device,
1192 	.device_group = generic_single_device_group,
1193 	.of_xlate = rk_iommu_of_xlate,
1194 	.default_domain_ops = &(const struct iommu_domain_ops) {
1195 		.attach_dev	= rk_iommu_attach_device,
1196 		.map_pages	= rk_iommu_map,
1197 		.unmap_pages	= rk_iommu_unmap,
1198 		.iova_to_phys	= rk_iommu_iova_to_phys,
1199 		.free		= rk_iommu_domain_free,
1200 	}
1201 };
1202 
1203 static int rk_iommu_probe(struct platform_device *pdev)
1204 {
1205 	struct device *dev = &pdev->dev;
1206 	struct rk_iommu *iommu;
1207 	struct resource *res;
1208 	const struct rk_iommu_ops *ops;
1209 	int num_res = pdev->num_resources;
1210 	int err, i;
1211 
1212 	iommu = devm_kzalloc(dev, sizeof(*iommu), GFP_KERNEL);
1213 	if (!iommu)
1214 		return -ENOMEM;
1215 
1216 	iommu->domain = &rk_identity_domain;
1217 
1218 	platform_set_drvdata(pdev, iommu);
1219 	iommu->dev = dev;
1220 	iommu->num_mmu = 0;
1221 
1222 	ops = of_device_get_match_data(dev);
1223 	if (!rk_ops)
1224 		rk_ops = ops;
1225 
1226 	/*
1227 	 * That should not happen unless different versions of the
1228 	 * hardware block are embedded the same SoC
1229 	 */
1230 	if (WARN_ON(rk_ops != ops))
1231 		return -EINVAL;
1232 
1233 	iommu->bases = devm_kcalloc(dev, num_res, sizeof(*iommu->bases),
1234 				    GFP_KERNEL);
1235 	if (!iommu->bases)
1236 		return -ENOMEM;
1237 
1238 	for (i = 0; i < num_res; i++) {
1239 		res = platform_get_resource(pdev, IORESOURCE_MEM, i);
1240 		if (!res)
1241 			continue;
1242 		iommu->bases[i] = devm_ioremap_resource(&pdev->dev, res);
1243 		if (IS_ERR(iommu->bases[i]))
1244 			continue;
1245 		iommu->num_mmu++;
1246 	}
1247 	if (iommu->num_mmu == 0)
1248 		return PTR_ERR(iommu->bases[0]);
1249 
1250 	iommu->num_irq = platform_irq_count(pdev);
1251 	if (iommu->num_irq < 0)
1252 		return iommu->num_irq;
1253 
1254 	iommu->reset_disabled = device_property_read_bool(dev,
1255 					"rockchip,disable-mmu-reset");
1256 
1257 	iommu->num_clocks = ARRAY_SIZE(rk_iommu_clocks);
1258 	iommu->clocks = devm_kcalloc(iommu->dev, iommu->num_clocks,
1259 				     sizeof(*iommu->clocks), GFP_KERNEL);
1260 	if (!iommu->clocks)
1261 		return -ENOMEM;
1262 
1263 	for (i = 0; i < iommu->num_clocks; ++i)
1264 		iommu->clocks[i].id = rk_iommu_clocks[i];
1265 
1266 	/*
1267 	 * iommu clocks should be present for all new devices and devicetrees
1268 	 * but there are older devicetrees without clocks out in the wild.
1269 	 * So clocks as optional for the time being.
1270 	 */
1271 	err = devm_clk_bulk_get(iommu->dev, iommu->num_clocks, iommu->clocks);
1272 	if (err == -ENOENT)
1273 		iommu->num_clocks = 0;
1274 	else if (err)
1275 		return err;
1276 
1277 	err = clk_bulk_prepare(iommu->num_clocks, iommu->clocks);
1278 	if (err)
1279 		return err;
1280 
1281 	pm_runtime_enable(dev);
1282 
1283 	for (i = 0; i < iommu->num_irq; i++) {
1284 		int irq = platform_get_irq(pdev, i);
1285 
1286 		if (irq < 0) {
1287 			err = irq;
1288 			goto err_pm_disable;
1289 		}
1290 
1291 		err = devm_request_irq(iommu->dev, irq, rk_iommu_irq,
1292 				       IRQF_SHARED, dev_name(dev), iommu);
1293 		if (err)
1294 			goto err_pm_disable;
1295 	}
1296 
1297 	dma_set_mask_and_coherent(dev, rk_ops->dma_bit_mask);
1298 
1299 	err = iommu_device_sysfs_add(&iommu->iommu, dev, NULL, dev_name(dev));
1300 	if (err)
1301 		goto err_pm_disable;
1302 
1303 	err = iommu_device_register(&iommu->iommu, &rk_iommu_ops, dev);
1304 	if (err)
1305 		goto err_remove_sysfs;
1306 
1307 	return 0;
1308 err_remove_sysfs:
1309 	iommu_device_sysfs_remove(&iommu->iommu);
1310 err_pm_disable:
1311 	pm_runtime_disable(dev);
1312 	clk_bulk_unprepare(iommu->num_clocks, iommu->clocks);
1313 	return err;
1314 }
1315 
1316 static void rk_iommu_shutdown(struct platform_device *pdev)
1317 {
1318 	struct rk_iommu *iommu = platform_get_drvdata(pdev);
1319 	int i;
1320 
1321 	for (i = 0; i < iommu->num_irq; i++) {
1322 		int irq = platform_get_irq(pdev, i);
1323 
1324 		devm_free_irq(iommu->dev, irq, iommu);
1325 	}
1326 
1327 	pm_runtime_force_suspend(&pdev->dev);
1328 }
1329 
1330 static int __maybe_unused rk_iommu_suspend(struct device *dev)
1331 {
1332 	struct rk_iommu *iommu = dev_get_drvdata(dev);
1333 
1334 	if (iommu->domain == &rk_identity_domain)
1335 		return 0;
1336 
1337 	rk_iommu_disable(iommu);
1338 	return 0;
1339 }
1340 
1341 static int __maybe_unused rk_iommu_resume(struct device *dev)
1342 {
1343 	struct rk_iommu *iommu = dev_get_drvdata(dev);
1344 
1345 	if (iommu->domain == &rk_identity_domain)
1346 		return 0;
1347 
1348 	return rk_iommu_enable(iommu);
1349 }
1350 
1351 static const struct dev_pm_ops rk_iommu_pm_ops = {
1352 	SET_RUNTIME_PM_OPS(rk_iommu_suspend, rk_iommu_resume, NULL)
1353 	SET_SYSTEM_SLEEP_PM_OPS(pm_runtime_force_suspend,
1354 				pm_runtime_force_resume)
1355 };
1356 
1357 static struct rk_iommu_ops iommu_data_ops_v1 = {
1358 	.pt_address = &rk_dte_pt_address,
1359 	.mk_dtentries = &rk_mk_dte,
1360 	.mk_ptentries = &rk_mk_pte,
1361 	.dma_bit_mask = DMA_BIT_MASK(32),
1362 	.gfp_flags = GFP_DMA32,
1363 };
1364 
1365 static struct rk_iommu_ops iommu_data_ops_v2 = {
1366 	.pt_address = &rk_dte_pt_address_v2,
1367 	.mk_dtentries = &rk_mk_dte_v2,
1368 	.mk_ptentries = &rk_mk_pte_v2,
1369 	.dma_bit_mask = DMA_BIT_MASK(40),
1370 	.gfp_flags = 0,
1371 };
1372 
1373 static const struct of_device_id rk_iommu_dt_ids[] = {
1374 	{	.compatible = "rockchip,iommu",
1375 		.data = &iommu_data_ops_v1,
1376 	},
1377 	{	.compatible = "rockchip,rk3568-iommu",
1378 		.data = &iommu_data_ops_v2,
1379 	},
1380 	{ /* sentinel */ }
1381 };
1382 
1383 static struct platform_driver rk_iommu_driver = {
1384 	.probe = rk_iommu_probe,
1385 	.shutdown = rk_iommu_shutdown,
1386 	.driver = {
1387 		   .name = "rk_iommu",
1388 		   .of_match_table = rk_iommu_dt_ids,
1389 		   .pm = &rk_iommu_pm_ops,
1390 		   .suppress_bind_attrs = true,
1391 	},
1392 };
1393 builtin_platform_driver(rk_iommu_driver);
1394