xref: /openbmc/linux/arch/sparc/kernel/iommu.c (revision b627b4ed)
1 /* iommu.c: Generic sparc64 IOMMU support.
2  *
3  * Copyright (C) 1999, 2007, 2008 David S. Miller (davem@davemloft.net)
4  * Copyright (C) 1999, 2000 Jakub Jelinek (jakub@redhat.com)
5  */
6 
7 #include <linux/kernel.h>
8 #include <linux/module.h>
9 #include <linux/delay.h>
10 #include <linux/device.h>
11 #include <linux/dma-mapping.h>
12 #include <linux/errno.h>
13 #include <linux/iommu-helper.h>
14 
15 #ifdef CONFIG_PCI
16 #include <linux/pci.h>
17 #endif
18 
19 #include <asm/iommu.h>
20 
21 #include "iommu_common.h"
22 
23 #define STC_CTXMATCH_ADDR(STC, CTX)	\
24 	((STC)->strbuf_ctxmatch_base + ((CTX) << 3))
25 #define STC_FLUSHFLAG_INIT(STC) \
26 	(*((STC)->strbuf_flushflag) = 0UL)
27 #define STC_FLUSHFLAG_SET(STC) \
28 	(*((STC)->strbuf_flushflag) != 0UL)
29 
30 #define iommu_read(__reg) \
31 ({	u64 __ret; \
32 	__asm__ __volatile__("ldxa [%1] %2, %0" \
33 			     : "=r" (__ret) \
34 			     : "r" (__reg), "i" (ASI_PHYS_BYPASS_EC_E) \
35 			     : "memory"); \
36 	__ret; \
37 })
38 #define iommu_write(__reg, __val) \
39 	__asm__ __volatile__("stxa %0, [%1] %2" \
40 			     : /* no outputs */ \
41 			     : "r" (__val), "r" (__reg), \
42 			       "i" (ASI_PHYS_BYPASS_EC_E))
43 
44 /* Must be invoked under the IOMMU lock. */
45 static void iommu_flushall(struct iommu *iommu)
46 {
47 	if (iommu->iommu_flushinv) {
48 		iommu_write(iommu->iommu_flushinv, ~(u64)0);
49 	} else {
50 		unsigned long tag;
51 		int entry;
52 
53 		tag = iommu->iommu_tags;
54 		for (entry = 0; entry < 16; entry++) {
55 			iommu_write(tag, 0);
56 			tag += 8;
57 		}
58 
59 		/* Ensure completion of previous PIO writes. */
60 		(void) iommu_read(iommu->write_complete_reg);
61 	}
62 }
63 
64 #define IOPTE_CONSISTENT(CTX) \
65 	(IOPTE_VALID | IOPTE_CACHE | \
66 	 (((CTX) << 47) & IOPTE_CONTEXT))
67 
68 #define IOPTE_STREAMING(CTX) \
69 	(IOPTE_CONSISTENT(CTX) | IOPTE_STBUF)
70 
71 /* Existing mappings are never marked invalid, instead they
72  * are pointed to a dummy page.
73  */
74 #define IOPTE_IS_DUMMY(iommu, iopte)	\
75 	((iopte_val(*iopte) & IOPTE_PAGE) == (iommu)->dummy_page_pa)
76 
77 static inline void iopte_make_dummy(struct iommu *iommu, iopte_t *iopte)
78 {
79 	unsigned long val = iopte_val(*iopte);
80 
81 	val &= ~IOPTE_PAGE;
82 	val |= iommu->dummy_page_pa;
83 
84 	iopte_val(*iopte) = val;
85 }
86 
87 /* Based almost entirely upon the ppc64 iommu allocator.  If you use the 'handle'
88  * facility it must all be done in one pass while under the iommu lock.
89  *
90  * On sun4u platforms, we only flush the IOMMU once every time we've passed
91  * over the entire page table doing allocations.  Therefore we only ever advance
92  * the hint and cannot backtrack it.
93  */
94 unsigned long iommu_range_alloc(struct device *dev,
95 				struct iommu *iommu,
96 				unsigned long npages,
97 				unsigned long *handle)
98 {
99 	unsigned long n, end, start, limit, boundary_size;
100 	struct iommu_arena *arena = &iommu->arena;
101 	int pass = 0;
102 
103 	/* This allocator was derived from x86_64's bit string search */
104 
105 	/* Sanity check */
106 	if (unlikely(npages == 0)) {
107 		if (printk_ratelimit())
108 			WARN_ON(1);
109 		return DMA_ERROR_CODE;
110 	}
111 
112 	if (handle && *handle)
113 		start = *handle;
114 	else
115 		start = arena->hint;
116 
117 	limit = arena->limit;
118 
119 	/* The case below can happen if we have a small segment appended
120 	 * to a large, or when the previous alloc was at the very end of
121 	 * the available space. If so, go back to the beginning and flush.
122 	 */
123 	if (start >= limit) {
124 		start = 0;
125 		if (iommu->flush_all)
126 			iommu->flush_all(iommu);
127 	}
128 
129  again:
130 
131 	if (dev)
132 		boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
133 				      1 << IO_PAGE_SHIFT);
134 	else
135 		boundary_size = ALIGN(1UL << 32, 1 << IO_PAGE_SHIFT);
136 
137 	n = iommu_area_alloc(arena->map, limit, start, npages,
138 			     iommu->page_table_map_base >> IO_PAGE_SHIFT,
139 			     boundary_size >> IO_PAGE_SHIFT, 0);
140 	if (n == -1) {
141 		if (likely(pass < 1)) {
142 			/* First failure, rescan from the beginning.  */
143 			start = 0;
144 			if (iommu->flush_all)
145 				iommu->flush_all(iommu);
146 			pass++;
147 			goto again;
148 		} else {
149 			/* Second failure, give up */
150 			return DMA_ERROR_CODE;
151 		}
152 	}
153 
154 	end = n + npages;
155 
156 	arena->hint = end;
157 
158 	/* Update handle for SG allocations */
159 	if (handle)
160 		*handle = end;
161 
162 	return n;
163 }
164 
165 void iommu_range_free(struct iommu *iommu, dma_addr_t dma_addr, unsigned long npages)
166 {
167 	struct iommu_arena *arena = &iommu->arena;
168 	unsigned long entry;
169 
170 	entry = (dma_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT;
171 
172 	iommu_area_free(arena->map, entry, npages);
173 }
174 
175 int iommu_table_init(struct iommu *iommu, int tsbsize,
176 		     u32 dma_offset, u32 dma_addr_mask,
177 		     int numa_node)
178 {
179 	unsigned long i, order, sz, num_tsb_entries;
180 	struct page *page;
181 
182 	num_tsb_entries = tsbsize / sizeof(iopte_t);
183 
184 	/* Setup initial software IOMMU state. */
185 	spin_lock_init(&iommu->lock);
186 	iommu->ctx_lowest_free = 1;
187 	iommu->page_table_map_base = dma_offset;
188 	iommu->dma_addr_mask = dma_addr_mask;
189 
190 	/* Allocate and initialize the free area map.  */
191 	sz = num_tsb_entries / 8;
192 	sz = (sz + 7UL) & ~7UL;
193 	iommu->arena.map = kmalloc_node(sz, GFP_KERNEL, numa_node);
194 	if (!iommu->arena.map) {
195 		printk(KERN_ERR "IOMMU: Error, kmalloc(arena.map) failed.\n");
196 		return -ENOMEM;
197 	}
198 	memset(iommu->arena.map, 0, sz);
199 	iommu->arena.limit = num_tsb_entries;
200 
201 	if (tlb_type != hypervisor)
202 		iommu->flush_all = iommu_flushall;
203 
204 	/* Allocate and initialize the dummy page which we
205 	 * set inactive IO PTEs to point to.
206 	 */
207 	page = alloc_pages_node(numa_node, GFP_KERNEL, 0);
208 	if (!page) {
209 		printk(KERN_ERR "IOMMU: Error, gfp(dummy_page) failed.\n");
210 		goto out_free_map;
211 	}
212 	iommu->dummy_page = (unsigned long) page_address(page);
213 	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
214 	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
215 
216 	/* Now allocate and setup the IOMMU page table itself.  */
217 	order = get_order(tsbsize);
218 	page = alloc_pages_node(numa_node, GFP_KERNEL, order);
219 	if (!page) {
220 		printk(KERN_ERR "IOMMU: Error, gfp(tsb) failed.\n");
221 		goto out_free_dummy_page;
222 	}
223 	iommu->page_table = (iopte_t *)page_address(page);
224 
225 	for (i = 0; i < num_tsb_entries; i++)
226 		iopte_make_dummy(iommu, &iommu->page_table[i]);
227 
228 	return 0;
229 
230 out_free_dummy_page:
231 	free_page(iommu->dummy_page);
232 	iommu->dummy_page = 0UL;
233 
234 out_free_map:
235 	kfree(iommu->arena.map);
236 	iommu->arena.map = NULL;
237 
238 	return -ENOMEM;
239 }
240 
241 static inline iopte_t *alloc_npages(struct device *dev, struct iommu *iommu,
242 				    unsigned long npages)
243 {
244 	unsigned long entry;
245 
246 	entry = iommu_range_alloc(dev, iommu, npages, NULL);
247 	if (unlikely(entry == DMA_ERROR_CODE))
248 		return NULL;
249 
250 	return iommu->page_table + entry;
251 }
252 
253 static int iommu_alloc_ctx(struct iommu *iommu)
254 {
255 	int lowest = iommu->ctx_lowest_free;
256 	int sz = IOMMU_NUM_CTXS - lowest;
257 	int n = find_next_zero_bit(iommu->ctx_bitmap, sz, lowest);
258 
259 	if (unlikely(n == sz)) {
260 		n = find_next_zero_bit(iommu->ctx_bitmap, lowest, 1);
261 		if (unlikely(n == lowest)) {
262 			printk(KERN_WARNING "IOMMU: Ran out of contexts.\n");
263 			n = 0;
264 		}
265 	}
266 	if (n)
267 		__set_bit(n, iommu->ctx_bitmap);
268 
269 	return n;
270 }
271 
272 static inline void iommu_free_ctx(struct iommu *iommu, int ctx)
273 {
274 	if (likely(ctx)) {
275 		__clear_bit(ctx, iommu->ctx_bitmap);
276 		if (ctx < iommu->ctx_lowest_free)
277 			iommu->ctx_lowest_free = ctx;
278 	}
279 }
280 
281 static void *dma_4u_alloc_coherent(struct device *dev, size_t size,
282 				   dma_addr_t *dma_addrp, gfp_t gfp)
283 {
284 	unsigned long flags, order, first_page;
285 	struct iommu *iommu;
286 	struct page *page;
287 	int npages, nid;
288 	iopte_t *iopte;
289 	void *ret;
290 
291 	size = IO_PAGE_ALIGN(size);
292 	order = get_order(size);
293 	if (order >= 10)
294 		return NULL;
295 
296 	nid = dev->archdata.numa_node;
297 	page = alloc_pages_node(nid, gfp, order);
298 	if (unlikely(!page))
299 		return NULL;
300 
301 	first_page = (unsigned long) page_address(page);
302 	memset((char *)first_page, 0, PAGE_SIZE << order);
303 
304 	iommu = dev->archdata.iommu;
305 
306 	spin_lock_irqsave(&iommu->lock, flags);
307 	iopte = alloc_npages(dev, iommu, size >> IO_PAGE_SHIFT);
308 	spin_unlock_irqrestore(&iommu->lock, flags);
309 
310 	if (unlikely(iopte == NULL)) {
311 		free_pages(first_page, order);
312 		return NULL;
313 	}
314 
315 	*dma_addrp = (iommu->page_table_map_base +
316 		      ((iopte - iommu->page_table) << IO_PAGE_SHIFT));
317 	ret = (void *) first_page;
318 	npages = size >> IO_PAGE_SHIFT;
319 	first_page = __pa(first_page);
320 	while (npages--) {
321 		iopte_val(*iopte) = (IOPTE_CONSISTENT(0UL) |
322 				     IOPTE_WRITE |
323 				     (first_page & IOPTE_PAGE));
324 		iopte++;
325 		first_page += IO_PAGE_SIZE;
326 	}
327 
328 	return ret;
329 }
330 
331 static void dma_4u_free_coherent(struct device *dev, size_t size,
332 				 void *cpu, dma_addr_t dvma)
333 {
334 	struct iommu *iommu;
335 	iopte_t *iopte;
336 	unsigned long flags, order, npages;
337 
338 	npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
339 	iommu = dev->archdata.iommu;
340 	iopte = iommu->page_table +
341 		((dvma - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
342 
343 	spin_lock_irqsave(&iommu->lock, flags);
344 
345 	iommu_range_free(iommu, dvma, npages);
346 
347 	spin_unlock_irqrestore(&iommu->lock, flags);
348 
349 	order = get_order(size);
350 	if (order < 10)
351 		free_pages((unsigned long)cpu, order);
352 }
353 
354 static dma_addr_t dma_4u_map_single(struct device *dev, void *ptr, size_t sz,
355 				    enum dma_data_direction direction)
356 {
357 	struct iommu *iommu;
358 	struct strbuf *strbuf;
359 	iopte_t *base;
360 	unsigned long flags, npages, oaddr;
361 	unsigned long i, base_paddr, ctx;
362 	u32 bus_addr, ret;
363 	unsigned long iopte_protection;
364 
365 	iommu = dev->archdata.iommu;
366 	strbuf = dev->archdata.stc;
367 
368 	if (unlikely(direction == DMA_NONE))
369 		goto bad_no_ctx;
370 
371 	oaddr = (unsigned long)ptr;
372 	npages = IO_PAGE_ALIGN(oaddr + sz) - (oaddr & IO_PAGE_MASK);
373 	npages >>= IO_PAGE_SHIFT;
374 
375 	spin_lock_irqsave(&iommu->lock, flags);
376 	base = alloc_npages(dev, iommu, npages);
377 	ctx = 0;
378 	if (iommu->iommu_ctxflush)
379 		ctx = iommu_alloc_ctx(iommu);
380 	spin_unlock_irqrestore(&iommu->lock, flags);
381 
382 	if (unlikely(!base))
383 		goto bad;
384 
385 	bus_addr = (iommu->page_table_map_base +
386 		    ((base - iommu->page_table) << IO_PAGE_SHIFT));
387 	ret = bus_addr | (oaddr & ~IO_PAGE_MASK);
388 	base_paddr = __pa(oaddr & IO_PAGE_MASK);
389 	if (strbuf->strbuf_enabled)
390 		iopte_protection = IOPTE_STREAMING(ctx);
391 	else
392 		iopte_protection = IOPTE_CONSISTENT(ctx);
393 	if (direction != DMA_TO_DEVICE)
394 		iopte_protection |= IOPTE_WRITE;
395 
396 	for (i = 0; i < npages; i++, base++, base_paddr += IO_PAGE_SIZE)
397 		iopte_val(*base) = iopte_protection | base_paddr;
398 
399 	return ret;
400 
401 bad:
402 	iommu_free_ctx(iommu, ctx);
403 bad_no_ctx:
404 	if (printk_ratelimit())
405 		WARN_ON(1);
406 	return DMA_ERROR_CODE;
407 }
408 
409 static void strbuf_flush(struct strbuf *strbuf, struct iommu *iommu,
410 			 u32 vaddr, unsigned long ctx, unsigned long npages,
411 			 enum dma_data_direction direction)
412 {
413 	int limit;
414 
415 	if (strbuf->strbuf_ctxflush &&
416 	    iommu->iommu_ctxflush) {
417 		unsigned long matchreg, flushreg;
418 		u64 val;
419 
420 		flushreg = strbuf->strbuf_ctxflush;
421 		matchreg = STC_CTXMATCH_ADDR(strbuf, ctx);
422 
423 		iommu_write(flushreg, ctx);
424 		val = iommu_read(matchreg);
425 		val &= 0xffff;
426 		if (!val)
427 			goto do_flush_sync;
428 
429 		while (val) {
430 			if (val & 0x1)
431 				iommu_write(flushreg, ctx);
432 			val >>= 1;
433 		}
434 		val = iommu_read(matchreg);
435 		if (unlikely(val)) {
436 			printk(KERN_WARNING "strbuf_flush: ctx flush "
437 			       "timeout matchreg[%llx] ctx[%lx]\n",
438 			       val, ctx);
439 			goto do_page_flush;
440 		}
441 	} else {
442 		unsigned long i;
443 
444 	do_page_flush:
445 		for (i = 0; i < npages; i++, vaddr += IO_PAGE_SIZE)
446 			iommu_write(strbuf->strbuf_pflush, vaddr);
447 	}
448 
449 do_flush_sync:
450 	/* If the device could not have possibly put dirty data into
451 	 * the streaming cache, no flush-flag synchronization needs
452 	 * to be performed.
453 	 */
454 	if (direction == DMA_TO_DEVICE)
455 		return;
456 
457 	STC_FLUSHFLAG_INIT(strbuf);
458 	iommu_write(strbuf->strbuf_fsync, strbuf->strbuf_flushflag_pa);
459 	(void) iommu_read(iommu->write_complete_reg);
460 
461 	limit = 100000;
462 	while (!STC_FLUSHFLAG_SET(strbuf)) {
463 		limit--;
464 		if (!limit)
465 			break;
466 		udelay(1);
467 		rmb();
468 	}
469 	if (!limit)
470 		printk(KERN_WARNING "strbuf_flush: flushflag timeout "
471 		       "vaddr[%08x] ctx[%lx] npages[%ld]\n",
472 		       vaddr, ctx, npages);
473 }
474 
475 static void dma_4u_unmap_single(struct device *dev, dma_addr_t bus_addr,
476 				size_t sz, enum dma_data_direction direction)
477 {
478 	struct iommu *iommu;
479 	struct strbuf *strbuf;
480 	iopte_t *base;
481 	unsigned long flags, npages, ctx, i;
482 
483 	if (unlikely(direction == DMA_NONE)) {
484 		if (printk_ratelimit())
485 			WARN_ON(1);
486 		return;
487 	}
488 
489 	iommu = dev->archdata.iommu;
490 	strbuf = dev->archdata.stc;
491 
492 	npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
493 	npages >>= IO_PAGE_SHIFT;
494 	base = iommu->page_table +
495 		((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
496 	bus_addr &= IO_PAGE_MASK;
497 
498 	spin_lock_irqsave(&iommu->lock, flags);
499 
500 	/* Record the context, if any. */
501 	ctx = 0;
502 	if (iommu->iommu_ctxflush)
503 		ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL;
504 
505 	/* Step 1: Kick data out of streaming buffers if necessary. */
506 	if (strbuf->strbuf_enabled)
507 		strbuf_flush(strbuf, iommu, bus_addr, ctx,
508 			     npages, direction);
509 
510 	/* Step 2: Clear out TSB entries. */
511 	for (i = 0; i < npages; i++)
512 		iopte_make_dummy(iommu, base + i);
513 
514 	iommu_range_free(iommu, bus_addr, npages);
515 
516 	iommu_free_ctx(iommu, ctx);
517 
518 	spin_unlock_irqrestore(&iommu->lock, flags);
519 }
520 
521 static int dma_4u_map_sg(struct device *dev, struct scatterlist *sglist,
522 			 int nelems, enum dma_data_direction direction)
523 {
524 	struct scatterlist *s, *outs, *segstart;
525 	unsigned long flags, handle, prot, ctx;
526 	dma_addr_t dma_next = 0, dma_addr;
527 	unsigned int max_seg_size;
528 	unsigned long seg_boundary_size;
529 	int outcount, incount, i;
530 	struct strbuf *strbuf;
531 	struct iommu *iommu;
532 	unsigned long base_shift;
533 
534 	BUG_ON(direction == DMA_NONE);
535 
536 	iommu = dev->archdata.iommu;
537 	strbuf = dev->archdata.stc;
538 	if (nelems == 0 || !iommu)
539 		return 0;
540 
541 	spin_lock_irqsave(&iommu->lock, flags);
542 
543 	ctx = 0;
544 	if (iommu->iommu_ctxflush)
545 		ctx = iommu_alloc_ctx(iommu);
546 
547 	if (strbuf->strbuf_enabled)
548 		prot = IOPTE_STREAMING(ctx);
549 	else
550 		prot = IOPTE_CONSISTENT(ctx);
551 	if (direction != DMA_TO_DEVICE)
552 		prot |= IOPTE_WRITE;
553 
554 	outs = s = segstart = &sglist[0];
555 	outcount = 1;
556 	incount = nelems;
557 	handle = 0;
558 
559 	/* Init first segment length for backout at failure */
560 	outs->dma_length = 0;
561 
562 	max_seg_size = dma_get_max_seg_size(dev);
563 	seg_boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
564 				  IO_PAGE_SIZE) >> IO_PAGE_SHIFT;
565 	base_shift = iommu->page_table_map_base >> IO_PAGE_SHIFT;
566 	for_each_sg(sglist, s, nelems, i) {
567 		unsigned long paddr, npages, entry, out_entry = 0, slen;
568 		iopte_t *base;
569 
570 		slen = s->length;
571 		/* Sanity check */
572 		if (slen == 0) {
573 			dma_next = 0;
574 			continue;
575 		}
576 		/* Allocate iommu entries for that segment */
577 		paddr = (unsigned long) SG_ENT_PHYS_ADDRESS(s);
578 		npages = iommu_num_pages(paddr, slen, IO_PAGE_SIZE);
579 		entry = iommu_range_alloc(dev, iommu, npages, &handle);
580 
581 		/* Handle failure */
582 		if (unlikely(entry == DMA_ERROR_CODE)) {
583 			if (printk_ratelimit())
584 				printk(KERN_INFO "iommu_alloc failed, iommu %p paddr %lx"
585 				       " npages %lx\n", iommu, paddr, npages);
586 			goto iommu_map_failed;
587 		}
588 
589 		base = iommu->page_table + entry;
590 
591 		/* Convert entry to a dma_addr_t */
592 		dma_addr = iommu->page_table_map_base +
593 			(entry << IO_PAGE_SHIFT);
594 		dma_addr |= (s->offset & ~IO_PAGE_MASK);
595 
596 		/* Insert into HW table */
597 		paddr &= IO_PAGE_MASK;
598 		while (npages--) {
599 			iopte_val(*base) = prot | paddr;
600 			base++;
601 			paddr += IO_PAGE_SIZE;
602 		}
603 
604 		/* If we are in an open segment, try merging */
605 		if (segstart != s) {
606 			/* We cannot merge if:
607 			 * - allocated dma_addr isn't contiguous to previous allocation
608 			 */
609 			if ((dma_addr != dma_next) ||
610 			    (outs->dma_length + s->length > max_seg_size) ||
611 			    (is_span_boundary(out_entry, base_shift,
612 					      seg_boundary_size, outs, s))) {
613 				/* Can't merge: create a new segment */
614 				segstart = s;
615 				outcount++;
616 				outs = sg_next(outs);
617 			} else {
618 				outs->dma_length += s->length;
619 			}
620 		}
621 
622 		if (segstart == s) {
623 			/* This is a new segment, fill entries */
624 			outs->dma_address = dma_addr;
625 			outs->dma_length = slen;
626 			out_entry = entry;
627 		}
628 
629 		/* Calculate next page pointer for contiguous check */
630 		dma_next = dma_addr + slen;
631 	}
632 
633 	spin_unlock_irqrestore(&iommu->lock, flags);
634 
635 	if (outcount < incount) {
636 		outs = sg_next(outs);
637 		outs->dma_address = DMA_ERROR_CODE;
638 		outs->dma_length = 0;
639 	}
640 
641 	return outcount;
642 
643 iommu_map_failed:
644 	for_each_sg(sglist, s, nelems, i) {
645 		if (s->dma_length != 0) {
646 			unsigned long vaddr, npages, entry, j;
647 			iopte_t *base;
648 
649 			vaddr = s->dma_address & IO_PAGE_MASK;
650 			npages = iommu_num_pages(s->dma_address, s->dma_length,
651 						 IO_PAGE_SIZE);
652 			iommu_range_free(iommu, vaddr, npages);
653 
654 			entry = (vaddr - iommu->page_table_map_base)
655 				>> IO_PAGE_SHIFT;
656 			base = iommu->page_table + entry;
657 
658 			for (j = 0; j < npages; j++)
659 				iopte_make_dummy(iommu, base + j);
660 
661 			s->dma_address = DMA_ERROR_CODE;
662 			s->dma_length = 0;
663 		}
664 		if (s == outs)
665 			break;
666 	}
667 	spin_unlock_irqrestore(&iommu->lock, flags);
668 
669 	return 0;
670 }
671 
672 /* If contexts are being used, they are the same in all of the mappings
673  * we make for a particular SG.
674  */
675 static unsigned long fetch_sg_ctx(struct iommu *iommu, struct scatterlist *sg)
676 {
677 	unsigned long ctx = 0;
678 
679 	if (iommu->iommu_ctxflush) {
680 		iopte_t *base;
681 		u32 bus_addr;
682 
683 		bus_addr = sg->dma_address & IO_PAGE_MASK;
684 		base = iommu->page_table +
685 			((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
686 
687 		ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL;
688 	}
689 	return ctx;
690 }
691 
692 static void dma_4u_unmap_sg(struct device *dev, struct scatterlist *sglist,
693 			    int nelems, enum dma_data_direction direction)
694 {
695 	unsigned long flags, ctx;
696 	struct scatterlist *sg;
697 	struct strbuf *strbuf;
698 	struct iommu *iommu;
699 
700 	BUG_ON(direction == DMA_NONE);
701 
702 	iommu = dev->archdata.iommu;
703 	strbuf = dev->archdata.stc;
704 
705 	ctx = fetch_sg_ctx(iommu, sglist);
706 
707 	spin_lock_irqsave(&iommu->lock, flags);
708 
709 	sg = sglist;
710 	while (nelems--) {
711 		dma_addr_t dma_handle = sg->dma_address;
712 		unsigned int len = sg->dma_length;
713 		unsigned long npages, entry;
714 		iopte_t *base;
715 		int i;
716 
717 		if (!len)
718 			break;
719 		npages = iommu_num_pages(dma_handle, len, IO_PAGE_SIZE);
720 		iommu_range_free(iommu, dma_handle, npages);
721 
722 		entry = ((dma_handle - iommu->page_table_map_base)
723 			 >> IO_PAGE_SHIFT);
724 		base = iommu->page_table + entry;
725 
726 		dma_handle &= IO_PAGE_MASK;
727 		if (strbuf->strbuf_enabled)
728 			strbuf_flush(strbuf, iommu, dma_handle, ctx,
729 				     npages, direction);
730 
731 		for (i = 0; i < npages; i++)
732 			iopte_make_dummy(iommu, base + i);
733 
734 		sg = sg_next(sg);
735 	}
736 
737 	iommu_free_ctx(iommu, ctx);
738 
739 	spin_unlock_irqrestore(&iommu->lock, flags);
740 }
741 
742 static void dma_4u_sync_single_for_cpu(struct device *dev,
743 				       dma_addr_t bus_addr, size_t sz,
744 				       enum dma_data_direction direction)
745 {
746 	struct iommu *iommu;
747 	struct strbuf *strbuf;
748 	unsigned long flags, ctx, npages;
749 
750 	iommu = dev->archdata.iommu;
751 	strbuf = dev->archdata.stc;
752 
753 	if (!strbuf->strbuf_enabled)
754 		return;
755 
756 	spin_lock_irqsave(&iommu->lock, flags);
757 
758 	npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
759 	npages >>= IO_PAGE_SHIFT;
760 	bus_addr &= IO_PAGE_MASK;
761 
762 	/* Step 1: Record the context, if any. */
763 	ctx = 0;
764 	if (iommu->iommu_ctxflush &&
765 	    strbuf->strbuf_ctxflush) {
766 		iopte_t *iopte;
767 
768 		iopte = iommu->page_table +
769 			((bus_addr - iommu->page_table_map_base)>>IO_PAGE_SHIFT);
770 		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
771 	}
772 
773 	/* Step 2: Kick data out of streaming buffers. */
774 	strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
775 
776 	spin_unlock_irqrestore(&iommu->lock, flags);
777 }
778 
779 static void dma_4u_sync_sg_for_cpu(struct device *dev,
780 				   struct scatterlist *sglist, int nelems,
781 				   enum dma_data_direction direction)
782 {
783 	struct iommu *iommu;
784 	struct strbuf *strbuf;
785 	unsigned long flags, ctx, npages, i;
786 	struct scatterlist *sg, *sgprv;
787 	u32 bus_addr;
788 
789 	iommu = dev->archdata.iommu;
790 	strbuf = dev->archdata.stc;
791 
792 	if (!strbuf->strbuf_enabled)
793 		return;
794 
795 	spin_lock_irqsave(&iommu->lock, flags);
796 
797 	/* Step 1: Record the context, if any. */
798 	ctx = 0;
799 	if (iommu->iommu_ctxflush &&
800 	    strbuf->strbuf_ctxflush) {
801 		iopte_t *iopte;
802 
803 		iopte = iommu->page_table +
804 			((sglist[0].dma_address - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
805 		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
806 	}
807 
808 	/* Step 2: Kick data out of streaming buffers. */
809 	bus_addr = sglist[0].dma_address & IO_PAGE_MASK;
810 	sgprv = NULL;
811 	for_each_sg(sglist, sg, nelems, i) {
812 		if (sg->dma_length == 0)
813 			break;
814 		sgprv = sg;
815 	}
816 
817 	npages = (IO_PAGE_ALIGN(sgprv->dma_address + sgprv->dma_length)
818 		  - bus_addr) >> IO_PAGE_SHIFT;
819 	strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
820 
821 	spin_unlock_irqrestore(&iommu->lock, flags);
822 }
823 
824 static const struct dma_ops sun4u_dma_ops = {
825 	.alloc_coherent		= dma_4u_alloc_coherent,
826 	.free_coherent		= dma_4u_free_coherent,
827 	.map_single		= dma_4u_map_single,
828 	.unmap_single		= dma_4u_unmap_single,
829 	.map_sg			= dma_4u_map_sg,
830 	.unmap_sg		= dma_4u_unmap_sg,
831 	.sync_single_for_cpu	= dma_4u_sync_single_for_cpu,
832 	.sync_sg_for_cpu	= dma_4u_sync_sg_for_cpu,
833 };
834 
835 const struct dma_ops *dma_ops = &sun4u_dma_ops;
836 EXPORT_SYMBOL(dma_ops);
837 
838 int dma_supported(struct device *dev, u64 device_mask)
839 {
840 	struct iommu *iommu = dev->archdata.iommu;
841 	u64 dma_addr_mask = iommu->dma_addr_mask;
842 
843 	if (device_mask >= (1UL << 32UL))
844 		return 0;
845 
846 	if ((device_mask & dma_addr_mask) == dma_addr_mask)
847 		return 1;
848 
849 #ifdef CONFIG_PCI
850 	if (dev->bus == &pci_bus_type)
851 		return pci_dma_supported(to_pci_dev(dev), device_mask);
852 #endif
853 
854 	return 0;
855 }
856 EXPORT_SYMBOL(dma_supported);
857 
858 int dma_set_mask(struct device *dev, u64 dma_mask)
859 {
860 #ifdef CONFIG_PCI
861 	if (dev->bus == &pci_bus_type)
862 		return pci_set_dma_mask(to_pci_dev(dev), dma_mask);
863 #endif
864 	return -EINVAL;
865 }
866 EXPORT_SYMBOL(dma_set_mask);
867