xref: /linux/arch/sparc/kernel/iommu.c (revision 3d0fe49454652117522f60bfbefb978ba0e5300b)
1 // SPDX-License-Identifier: GPL-2.0
2 /* iommu.c: Generic sparc64 IOMMU support.
3  *
4  * Copyright (C) 1999, 2007, 2008 David S. Miller (davem@davemloft.net)
5  * Copyright (C) 1999, 2000 Jakub Jelinek (jakub@redhat.com)
6  */
7 
8 #include <linux/kernel.h>
9 #include <linux/export.h>
10 #include <linux/slab.h>
11 #include <linux/delay.h>
12 #include <linux/device.h>
13 #include <linux/dma-map-ops.h>
14 #include <linux/errno.h>
15 #include <linux/iommu-helper.h>
16 #include <linux/bitmap.h>
17 #include <asm/iommu-common.h>
18 
19 #ifdef CONFIG_PCI
20 #include <linux/pci.h>
21 #endif
22 
23 #include <asm/iommu.h>
24 
25 #include "iommu_common.h"
26 #include "kernel.h"
27 
28 #define STC_CTXMATCH_ADDR(STC, CTX)	\
29 	((STC)->strbuf_ctxmatch_base + ((CTX) << 3))
30 #define STC_FLUSHFLAG_INIT(STC) \
31 	(*((STC)->strbuf_flushflag) = 0UL)
32 #define STC_FLUSHFLAG_SET(STC) \
33 	(*((STC)->strbuf_flushflag) != 0UL)
34 
35 #define iommu_read(__reg) \
36 ({	u64 __ret; \
37 	__asm__ __volatile__("ldxa [%1] %2, %0" \
38 			     : "=r" (__ret) \
39 			     : "r" (__reg), "i" (ASI_PHYS_BYPASS_EC_E) \
40 			     : "memory"); \
41 	__ret; \
42 })
43 #define iommu_write(__reg, __val) \
44 	__asm__ __volatile__("stxa %0, [%1] %2" \
45 			     : /* no outputs */ \
46 			     : "r" (__val), "r" (__reg), \
47 			       "i" (ASI_PHYS_BYPASS_EC_E))
48 
49 /* Must be invoked under the IOMMU lock. */
50 static void iommu_flushall(struct iommu_map_table *iommu_map_table)
51 {
52 	struct iommu *iommu = container_of(iommu_map_table, struct iommu, tbl);
53 	if (iommu->iommu_flushinv) {
54 		iommu_write(iommu->iommu_flushinv, ~(u64)0);
55 	} else {
56 		unsigned long tag;
57 		int entry;
58 
59 		tag = iommu->iommu_tags;
60 		for (entry = 0; entry < 16; entry++) {
61 			iommu_write(tag, 0);
62 			tag += 8;
63 		}
64 
65 		/* Ensure completion of previous PIO writes. */
66 		(void) iommu_read(iommu->write_complete_reg);
67 	}
68 }
69 
70 #define IOPTE_CONSISTENT(CTX) \
71 	(IOPTE_VALID | IOPTE_CACHE | \
72 	 (((CTX) << 47) & IOPTE_CONTEXT))
73 
74 #define IOPTE_STREAMING(CTX) \
75 	(IOPTE_CONSISTENT(CTX) | IOPTE_STBUF)
76 
77 /* Existing mappings are never marked invalid, instead they
78  * are pointed to a dummy page.
79  */
80 #define IOPTE_IS_DUMMY(iommu, iopte)	\
81 	((iopte_val(*iopte) & IOPTE_PAGE) == (iommu)->dummy_page_pa)
82 
83 static inline void iopte_make_dummy(struct iommu *iommu, iopte_t *iopte)
84 {
85 	unsigned long val = iopte_val(*iopte);
86 
87 	val &= ~IOPTE_PAGE;
88 	val |= iommu->dummy_page_pa;
89 
90 	iopte_val(*iopte) = val;
91 }
92 
93 int iommu_table_init(struct iommu *iommu, int tsbsize,
94 		     u32 dma_offset, u32 dma_addr_mask,
95 		     int numa_node)
96 {
97 	unsigned long i, order, sz, num_tsb_entries;
98 	struct page *page;
99 
100 	num_tsb_entries = tsbsize / sizeof(iopte_t);
101 
102 	/* Setup initial software IOMMU state. */
103 	spin_lock_init(&iommu->lock);
104 	iommu->ctx_lowest_free = 1;
105 	iommu->tbl.table_map_base = dma_offset;
106 	iommu->dma_addr_mask = dma_addr_mask;
107 
108 	/* Allocate and initialize the free area map.  */
109 	sz = num_tsb_entries / 8;
110 	sz = (sz + 7UL) & ~7UL;
111 	iommu->tbl.map = kzalloc_node(sz, GFP_KERNEL, numa_node);
112 	if (!iommu->tbl.map)
113 		return -ENOMEM;
114 
115 	iommu_tbl_pool_init(&iommu->tbl, num_tsb_entries, IO_PAGE_SHIFT,
116 			    (tlb_type != hypervisor ? iommu_flushall : NULL),
117 			    false, 1, false);
118 
119 	/* Allocate and initialize the dummy page which we
120 	 * set inactive IO PTEs to point to.
121 	 */
122 	page = alloc_pages_node(numa_node, GFP_KERNEL, 0);
123 	if (!page) {
124 		printk(KERN_ERR "IOMMU: Error, gfp(dummy_page) failed.\n");
125 		goto out_free_map;
126 	}
127 	iommu->dummy_page = (unsigned long) page_address(page);
128 	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
129 	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
130 
131 	/* Now allocate and setup the IOMMU page table itself.  */
132 	order = get_order(tsbsize);
133 	page = alloc_pages_node(numa_node, GFP_KERNEL, order);
134 	if (!page) {
135 		printk(KERN_ERR "IOMMU: Error, gfp(tsb) failed.\n");
136 		goto out_free_dummy_page;
137 	}
138 	iommu->page_table = (iopte_t *)page_address(page);
139 
140 	for (i = 0; i < num_tsb_entries; i++)
141 		iopte_make_dummy(iommu, &iommu->page_table[i]);
142 
143 	return 0;
144 
145 out_free_dummy_page:
146 	free_page(iommu->dummy_page);
147 	iommu->dummy_page = 0UL;
148 
149 out_free_map:
150 	kfree(iommu->tbl.map);
151 	iommu->tbl.map = NULL;
152 
153 	return -ENOMEM;
154 }
155 
156 static inline iopte_t *alloc_npages(struct device *dev,
157 				    struct iommu *iommu,
158 				    unsigned long npages)
159 {
160 	unsigned long entry;
161 
162 	entry = iommu_tbl_range_alloc(dev, &iommu->tbl, npages, NULL,
163 				      (unsigned long)(-1), 0);
164 	if (unlikely(entry == IOMMU_ERROR_CODE))
165 		return NULL;
166 
167 	return iommu->page_table + entry;
168 }
169 
170 static int iommu_alloc_ctx(struct iommu *iommu)
171 {
172 	int lowest = iommu->ctx_lowest_free;
173 	int n = find_next_zero_bit(iommu->ctx_bitmap, IOMMU_NUM_CTXS, lowest);
174 
175 	if (unlikely(n == IOMMU_NUM_CTXS)) {
176 		n = find_next_zero_bit(iommu->ctx_bitmap, lowest, 1);
177 		if (unlikely(n == lowest)) {
178 			printk(KERN_WARNING "IOMMU: Ran out of contexts.\n");
179 			n = 0;
180 		}
181 	}
182 	if (n)
183 		__set_bit(n, iommu->ctx_bitmap);
184 
185 	return n;
186 }
187 
188 static inline void iommu_free_ctx(struct iommu *iommu, int ctx)
189 {
190 	if (likely(ctx)) {
191 		__clear_bit(ctx, iommu->ctx_bitmap);
192 		if (ctx < iommu->ctx_lowest_free)
193 			iommu->ctx_lowest_free = ctx;
194 	}
195 }
196 
197 static void *dma_4u_alloc_coherent(struct device *dev, size_t size,
198 				   dma_addr_t *dma_addrp, gfp_t gfp,
199 				   unsigned long attrs)
200 {
201 	unsigned long order, first_page;
202 	struct iommu *iommu;
203 	struct page *page;
204 	int npages, nid;
205 	iopte_t *iopte;
206 	void *ret;
207 
208 	size = IO_PAGE_ALIGN(size);
209 	order = get_order(size);
210 	if (order >= 10)
211 		return NULL;
212 
213 	nid = dev->archdata.numa_node;
214 	page = alloc_pages_node(nid, gfp, order);
215 	if (unlikely(!page))
216 		return NULL;
217 
218 	first_page = (unsigned long) page_address(page);
219 	memset((char *)first_page, 0, PAGE_SIZE << order);
220 
221 	iommu = dev->archdata.iommu;
222 
223 	iopte = alloc_npages(dev, iommu, size >> IO_PAGE_SHIFT);
224 
225 	if (unlikely(iopte == NULL)) {
226 		free_pages(first_page, order);
227 		return NULL;
228 	}
229 
230 	*dma_addrp = (iommu->tbl.table_map_base +
231 		      ((iopte - iommu->page_table) << IO_PAGE_SHIFT));
232 	ret = (void *) first_page;
233 	npages = size >> IO_PAGE_SHIFT;
234 	first_page = __pa(first_page);
235 	while (npages--) {
236 		iopte_val(*iopte) = (IOPTE_CONSISTENT(0UL) |
237 				     IOPTE_WRITE |
238 				     (first_page & IOPTE_PAGE));
239 		iopte++;
240 		first_page += IO_PAGE_SIZE;
241 	}
242 
243 	return ret;
244 }
245 
246 static void dma_4u_free_coherent(struct device *dev, size_t size,
247 				 void *cpu, dma_addr_t dvma,
248 				 unsigned long attrs)
249 {
250 	struct iommu *iommu;
251 	unsigned long order, npages;
252 
253 	npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
254 	iommu = dev->archdata.iommu;
255 
256 	iommu_tbl_range_free(&iommu->tbl, dvma, npages, IOMMU_ERROR_CODE);
257 
258 	order = get_order(size);
259 	if (order < 10)
260 		free_pages((unsigned long)cpu, order);
261 }
262 
263 static dma_addr_t dma_4u_map_page(struct device *dev, struct page *page,
264 				  unsigned long offset, size_t sz,
265 				  enum dma_data_direction direction,
266 				  unsigned long attrs)
267 {
268 	struct iommu *iommu;
269 	struct strbuf *strbuf;
270 	iopte_t *base;
271 	unsigned long flags, npages, oaddr;
272 	unsigned long i, base_paddr, ctx;
273 	u32 bus_addr, ret;
274 	unsigned long iopte_protection;
275 
276 	iommu = dev->archdata.iommu;
277 	strbuf = dev->archdata.stc;
278 
279 	if (unlikely(direction == DMA_NONE))
280 		goto bad_no_ctx;
281 
282 	oaddr = (unsigned long)(page_address(page) + offset);
283 	npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
284 	npages >>= IO_PAGE_SHIFT;
285 
286 	base = alloc_npages(dev, iommu, npages);
287 	spin_lock_irqsave(&iommu->lock, flags);
288 	ctx = 0;
289 	if (iommu->iommu_ctxflush)
290 		ctx = iommu_alloc_ctx(iommu);
291 	spin_unlock_irqrestore(&iommu->lock, flags);
292 
293 	if (unlikely(!base))
294 		goto bad;
295 
296 	bus_addr = (iommu->tbl.table_map_base +
297 		    ((base - iommu->page_table) << IO_PAGE_SHIFT));
298 	ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
299 	base_paddr = __pa(oaddr & IO_PAGE_MASK);
300 	if (strbuf->strbuf_enabled)
301 		iopte_protection = IOPTE_STREAMING(ctx);
302 	else
303 		iopte_protection = IOPTE_CONSISTENT(ctx);
304 	if (direction != DMA_TO_DEVICE)
305 		iopte_protection |= IOPTE_WRITE;
306 
307 	for (i = 0; i < npages; i++, base++, base_paddr += IO_PAGE_SIZE)
308 		iopte_val(*base) = iopte_protection | base_paddr;
309 
310 	return ret;
311 
312 bad:
313 	iommu_free_ctx(iommu, ctx);
314 bad_no_ctx:
315 	if (printk_ratelimit())
316 		WARN_ON(1);
317 	return DMA_MAPPING_ERROR;
318 }
319 
320 static void strbuf_flush(struct strbuf *strbuf, struct iommu *iommu,
321 			 u32 vaddr, unsigned long ctx, unsigned long npages,
322 			 enum dma_data_direction direction)
323 {
324 	int limit;
325 
326 	if (strbuf->strbuf_ctxflush &&
327 	    iommu->iommu_ctxflush) {
328 		unsigned long matchreg, flushreg;
329 		u64 val;
330 
331 		flushreg = strbuf->strbuf_ctxflush;
332 		matchreg = STC_CTXMATCH_ADDR(strbuf, ctx);
333 
334 		iommu_write(flushreg, ctx);
335 		val = iommu_read(matchreg);
336 		val &= 0xffff;
337 		if (!val)
338 			goto do_flush_sync;
339 
340 		while (val) {
341 			if (val & 0x1)
342 				iommu_write(flushreg, ctx);
343 			val >>= 1;
344 		}
345 		val = iommu_read(matchreg);
346 		if (unlikely(val)) {
347 			printk(KERN_WARNING "strbuf_flush: ctx flush "
348 			       "timeout matchreg[%llx] ctx[%lx]\n",
349 			       val, ctx);
350 			goto do_page_flush;
351 		}
352 	} else {
353 		unsigned long i;
354 
355 	do_page_flush:
356 		for (i = 0; i < npages; i++, vaddr += IO_PAGE_SIZE)
357 			iommu_write(strbuf->strbuf_pflush, vaddr);
358 	}
359 
360 do_flush_sync:
361 	/* If the device could not have possibly put dirty data into
362 	 * the streaming cache, no flush-flag synchronization needs
363 	 * to be performed.
364 	 */
365 	if (direction == DMA_TO_DEVICE)
366 		return;
367 
368 	STC_FLUSHFLAG_INIT(strbuf);
369 	iommu_write(strbuf->strbuf_fsync, strbuf->strbuf_flushflag_pa);
370 	(void) iommu_read(iommu->write_complete_reg);
371 
372 	limit = 100000;
373 	while (!STC_FLUSHFLAG_SET(strbuf)) {
374 		limit--;
375 		if (!limit)
376 			break;
377 		udelay(1);
378 		rmb();
379 	}
380 	if (!limit)
381 		printk(KERN_WARNING "strbuf_flush: flushflag timeout "
382 		       "vaddr[%08x] ctx[%lx] npages[%ld]\n",
383 		       vaddr, ctx, npages);
384 }
385 
386 static void dma_4u_unmap_page(struct device *dev, dma_addr_t bus_addr,
387 			      size_t sz, enum dma_data_direction direction,
388 			      unsigned long attrs)
389 {
390 	struct iommu *iommu;
391 	struct strbuf *strbuf;
392 	iopte_t *base;
393 	unsigned long flags, npages, ctx, i;
394 
395 	if (unlikely(direction == DMA_NONE)) {
396 		if (printk_ratelimit())
397 			WARN_ON(1);
398 		return;
399 	}
400 
401 	iommu = dev->archdata.iommu;
402 	strbuf = dev->archdata.stc;
403 
404 	npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
405 	npages >>= IO_PAGE_SHIFT;
406 	base = iommu->page_table +
407 		((bus_addr - iommu->tbl.table_map_base) >> IO_PAGE_SHIFT);
408 	bus_addr &= IO_PAGE_MASK;
409 
410 	spin_lock_irqsave(&iommu->lock, flags);
411 
412 	/* Record the context, if any. */
413 	ctx = 0;
414 	if (iommu->iommu_ctxflush)
415 		ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL;
416 
417 	/* Step 1: Kick data out of streaming buffers if necessary. */
418 	if (strbuf->strbuf_enabled && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
419 		strbuf_flush(strbuf, iommu, bus_addr, ctx,
420 			     npages, direction);
421 
422 	/* Step 2: Clear out TSB entries. */
423 	for (i = 0; i < npages; i++)
424 		iopte_make_dummy(iommu, base + i);
425 
426 	iommu_free_ctx(iommu, ctx);
427 	spin_unlock_irqrestore(&iommu->lock, flags);
428 
429 	iommu_tbl_range_free(&iommu->tbl, bus_addr, npages, IOMMU_ERROR_CODE);
430 }
431 
432 static int dma_4u_map_sg(struct device *dev, struct scatterlist *sglist,
433 			 int nelems, enum dma_data_direction direction,
434 			 unsigned long attrs)
435 {
436 	struct scatterlist *s, *outs, *segstart;
437 	unsigned long flags, handle, prot, ctx;
438 	dma_addr_t dma_next = 0, dma_addr;
439 	unsigned int max_seg_size;
440 	unsigned long seg_boundary_size;
441 	int outcount, incount, i;
442 	struct strbuf *strbuf;
443 	struct iommu *iommu;
444 	unsigned long base_shift;
445 
446 	BUG_ON(direction == DMA_NONE);
447 
448 	iommu = dev->archdata.iommu;
449 	strbuf = dev->archdata.stc;
450 	if (nelems == 0 || !iommu)
451 		return -EINVAL;
452 
453 	spin_lock_irqsave(&iommu->lock, flags);
454 
455 	ctx = 0;
456 	if (iommu->iommu_ctxflush)
457 		ctx = iommu_alloc_ctx(iommu);
458 
459 	if (strbuf->strbuf_enabled)
460 		prot = IOPTE_STREAMING(ctx);
461 	else
462 		prot = IOPTE_CONSISTENT(ctx);
463 	if (direction != DMA_TO_DEVICE)
464 		prot |= IOPTE_WRITE;
465 
466 	outs = s = segstart = &sglist[0];
467 	outcount = 1;
468 	incount = nelems;
469 	handle = 0;
470 
471 	/* Init first segment length for backout at failure */
472 	outs->dma_length = 0;
473 
474 	max_seg_size = dma_get_max_seg_size(dev);
475 	seg_boundary_size = dma_get_seg_boundary_nr_pages(dev, IO_PAGE_SHIFT);
476 	base_shift = iommu->tbl.table_map_base >> IO_PAGE_SHIFT;
477 	for_each_sg(sglist, s, nelems, i) {
478 		unsigned long paddr, npages, entry, out_entry = 0, slen;
479 		iopte_t *base;
480 
481 		slen = s->length;
482 		/* Sanity check */
483 		if (slen == 0) {
484 			dma_next = 0;
485 			continue;
486 		}
487 		/* Allocate iommu entries for that segment */
488 		paddr = (unsigned long) SG_ENT_PHYS_ADDRESS(s);
489 		npages = iommu_num_pages(paddr, slen, IO_PAGE_SIZE);
490 		entry = iommu_tbl_range_alloc(dev, &iommu->tbl, npages,
491 					      &handle, (unsigned long)(-1), 0);
492 
493 		/* Handle failure */
494 		if (unlikely(entry == IOMMU_ERROR_CODE)) {
495 			if (printk_ratelimit())
496 				printk(KERN_INFO "iommu_alloc failed, iommu %p paddr %lx"
497 				       " npages %lx\n", iommu, paddr, npages);
498 			goto iommu_map_failed;
499 		}
500 
501 		base = iommu->page_table + entry;
502 
503 		/* Convert entry to a dma_addr_t */
504 		dma_addr = iommu->tbl.table_map_base +
505 			(entry << IO_PAGE_SHIFT);
506 		dma_addr |= (s->offset & ~IO_PAGE_MASK);
507 
508 		/* Insert into HW table */
509 		paddr &= IO_PAGE_MASK;
510 		while (npages--) {
511 			iopte_val(*base) = prot | paddr;
512 			base++;
513 			paddr += IO_PAGE_SIZE;
514 		}
515 
516 		/* If we are in an open segment, try merging */
517 		if (segstart != s) {
518 			/* We cannot merge if:
519 			 * - allocated dma_addr isn't contiguous to previous allocation
520 			 */
521 			if ((dma_addr != dma_next) ||
522 			    (outs->dma_length + s->length > max_seg_size) ||
523 			    (is_span_boundary(out_entry, base_shift,
524 					      seg_boundary_size, outs, s))) {
525 				/* Can't merge: create a new segment */
526 				segstart = s;
527 				outcount++;
528 				outs = sg_next(outs);
529 			} else {
530 				outs->dma_length += s->length;
531 			}
532 		}
533 
534 		if (segstart == s) {
535 			/* This is a new segment, fill entries */
536 			outs->dma_address = dma_addr;
537 			outs->dma_length = slen;
538 			out_entry = entry;
539 		}
540 
541 		/* Calculate next page pointer for contiguous check */
542 		dma_next = dma_addr + slen;
543 	}
544 
545 	spin_unlock_irqrestore(&iommu->lock, flags);
546 
547 	if (outcount < incount) {
548 		outs = sg_next(outs);
549 		outs->dma_length = 0;
550 	}
551 
552 	return outcount;
553 
554 iommu_map_failed:
555 	for_each_sg(sglist, s, nelems, i) {
556 		if (s->dma_length != 0) {
557 			unsigned long vaddr, npages, entry, j;
558 			iopte_t *base;
559 
560 			vaddr = s->dma_address & IO_PAGE_MASK;
561 			npages = iommu_num_pages(s->dma_address, s->dma_length,
562 						 IO_PAGE_SIZE);
563 
564 			entry = (vaddr - iommu->tbl.table_map_base)
565 				>> IO_PAGE_SHIFT;
566 			base = iommu->page_table + entry;
567 
568 			for (j = 0; j < npages; j++)
569 				iopte_make_dummy(iommu, base + j);
570 
571 			iommu_tbl_range_free(&iommu->tbl, vaddr, npages,
572 					     IOMMU_ERROR_CODE);
573 
574 			s->dma_length = 0;
575 		}
576 		if (s == outs)
577 			break;
578 	}
579 	spin_unlock_irqrestore(&iommu->lock, flags);
580 
581 	return -EINVAL;
582 }
583 
584 /* If contexts are being used, they are the same in all of the mappings
585  * we make for a particular SG.
586  */
587 static unsigned long fetch_sg_ctx(struct iommu *iommu, struct scatterlist *sg)
588 {
589 	unsigned long ctx = 0;
590 
591 	if (iommu->iommu_ctxflush) {
592 		iopte_t *base;
593 		u32 bus_addr;
594 		struct iommu_map_table *tbl = &iommu->tbl;
595 
596 		bus_addr = sg->dma_address & IO_PAGE_MASK;
597 		base = iommu->page_table +
598 			((bus_addr - tbl->table_map_base) >> IO_PAGE_SHIFT);
599 
600 		ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL;
601 	}
602 	return ctx;
603 }
604 
605 static void dma_4u_unmap_sg(struct device *dev, struct scatterlist *sglist,
606 			    int nelems, enum dma_data_direction direction,
607 			    unsigned long attrs)
608 {
609 	unsigned long flags, ctx;
610 	struct scatterlist *sg;
611 	struct strbuf *strbuf;
612 	struct iommu *iommu;
613 
614 	BUG_ON(direction == DMA_NONE);
615 
616 	iommu = dev->archdata.iommu;
617 	strbuf = dev->archdata.stc;
618 
619 	ctx = fetch_sg_ctx(iommu, sglist);
620 
621 	spin_lock_irqsave(&iommu->lock, flags);
622 
623 	sg = sglist;
624 	while (nelems--) {
625 		dma_addr_t dma_handle = sg->dma_address;
626 		unsigned int len = sg->dma_length;
627 		unsigned long npages, entry;
628 		iopte_t *base;
629 		int i;
630 
631 		if (!len)
632 			break;
633 		npages = iommu_num_pages(dma_handle, len, IO_PAGE_SIZE);
634 
635 		entry = ((dma_handle - iommu->tbl.table_map_base)
636 			 >> IO_PAGE_SHIFT);
637 		base = iommu->page_table + entry;
638 
639 		dma_handle &= IO_PAGE_MASK;
640 		if (strbuf->strbuf_enabled && !(attrs & DMA_ATTR_SKIP_CPU_SYNC))
641 			strbuf_flush(strbuf, iommu, dma_handle, ctx,
642 				     npages, direction);
643 
644 		for (i = 0; i < npages; i++)
645 			iopte_make_dummy(iommu, base + i);
646 
647 		iommu_tbl_range_free(&iommu->tbl, dma_handle, npages,
648 				     IOMMU_ERROR_CODE);
649 		sg = sg_next(sg);
650 	}
651 
652 	iommu_free_ctx(iommu, ctx);
653 
654 	spin_unlock_irqrestore(&iommu->lock, flags);
655 }
656 
657 static void dma_4u_sync_single_for_cpu(struct device *dev,
658 				       dma_addr_t bus_addr, size_t sz,
659 				       enum dma_data_direction direction)
660 {
661 	struct iommu *iommu;
662 	struct strbuf *strbuf;
663 	unsigned long flags, ctx, npages;
664 
665 	iommu = dev->archdata.iommu;
666 	strbuf = dev->archdata.stc;
667 
668 	if (!strbuf->strbuf_enabled)
669 		return;
670 
671 	spin_lock_irqsave(&iommu->lock, flags);
672 
673 	npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
674 	npages >>= IO_PAGE_SHIFT;
675 	bus_addr &= IO_PAGE_MASK;
676 
677 	/* Step 1: Record the context, if any. */
678 	ctx = 0;
679 	if (iommu->iommu_ctxflush &&
680 	    strbuf->strbuf_ctxflush) {
681 		iopte_t *iopte;
682 		struct iommu_map_table *tbl = &iommu->tbl;
683 
684 		iopte = iommu->page_table +
685 			((bus_addr - tbl->table_map_base)>>IO_PAGE_SHIFT);
686 		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
687 	}
688 
689 	/* Step 2: Kick data out of streaming buffers. */
690 	strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
691 
692 	spin_unlock_irqrestore(&iommu->lock, flags);
693 }
694 
695 static void dma_4u_sync_sg_for_cpu(struct device *dev,
696 				   struct scatterlist *sglist, int nelems,
697 				   enum dma_data_direction direction)
698 {
699 	struct iommu *iommu;
700 	struct strbuf *strbuf;
701 	unsigned long flags, ctx, npages, i;
702 	struct scatterlist *sg, *sgprv;
703 	u32 bus_addr;
704 
705 	iommu = dev->archdata.iommu;
706 	strbuf = dev->archdata.stc;
707 
708 	if (!strbuf->strbuf_enabled)
709 		return;
710 
711 	spin_lock_irqsave(&iommu->lock, flags);
712 
713 	/* Step 1: Record the context, if any. */
714 	ctx = 0;
715 	if (iommu->iommu_ctxflush &&
716 	    strbuf->strbuf_ctxflush) {
717 		iopte_t *iopte;
718 		struct iommu_map_table *tbl = &iommu->tbl;
719 
720 		iopte = iommu->page_table + ((sglist[0].dma_address -
721 			tbl->table_map_base) >> IO_PAGE_SHIFT);
722 		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
723 	}
724 
725 	/* Step 2: Kick data out of streaming buffers. */
726 	bus_addr = sglist[0].dma_address & IO_PAGE_MASK;
727 	sgprv = NULL;
728 	for_each_sg(sglist, sg, nelems, i) {
729 		if (sg->dma_length == 0)
730 			break;
731 		sgprv = sg;
732 	}
733 
734 	npages = (IO_PAGE_ALIGN(sgprv->dma_address + sgprv->dma_length)
735 		  - bus_addr) >> IO_PAGE_SHIFT;
736 	strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
737 
738 	spin_unlock_irqrestore(&iommu->lock, flags);
739 }
740 
741 static int dma_4u_supported(struct device *dev, u64 device_mask)
742 {
743 	struct iommu *iommu = dev->archdata.iommu;
744 
745 	if (ali_sound_dma_hack(dev, device_mask))
746 		return 1;
747 
748 	if (device_mask < iommu->dma_addr_mask)
749 		return 0;
750 	return 1;
751 }
752 
753 static const struct dma_map_ops sun4u_dma_ops = {
754 	.alloc			= dma_4u_alloc_coherent,
755 	.free			= dma_4u_free_coherent,
756 	.map_page		= dma_4u_map_page,
757 	.unmap_page		= dma_4u_unmap_page,
758 	.map_sg			= dma_4u_map_sg,
759 	.unmap_sg		= dma_4u_unmap_sg,
760 	.sync_single_for_cpu	= dma_4u_sync_single_for_cpu,
761 	.sync_sg_for_cpu	= dma_4u_sync_sg_for_cpu,
762 	.dma_supported		= dma_4u_supported,
763 };
764 
765 const struct dma_map_ops *dma_ops = &sun4u_dma_ops;
766 EXPORT_SYMBOL(dma_ops);
767