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/slab.h>
10 #include <linux/delay.h>
11 #include <linux/device.h>
12 #include <linux/dma-mapping.h>
13 #include <linux/errno.h>
14 #include <linux/iommu-helper.h>
15 #include <linux/bitmap.h>
16 
17 #ifdef CONFIG_PCI
18 #include <linux/pci.h>
19 #endif
20 
21 #include <asm/iommu.h>
22 
23 #include "iommu_common.h"
24 
25 #define STC_CTXMATCH_ADDR(STC, CTX)	\
26 	((STC)->strbuf_ctxmatch_base + ((CTX) << 3))
27 #define STC_FLUSHFLAG_INIT(STC) \
28 	(*((STC)->strbuf_flushflag) = 0UL)
29 #define STC_FLUSHFLAG_SET(STC) \
30 	(*((STC)->strbuf_flushflag) != 0UL)
31 
32 #define iommu_read(__reg) \
33 ({	u64 __ret; \
34 	__asm__ __volatile__("ldxa [%1] %2, %0" \
35 			     : "=r" (__ret) \
36 			     : "r" (__reg), "i" (ASI_PHYS_BYPASS_EC_E) \
37 			     : "memory"); \
38 	__ret; \
39 })
40 #define iommu_write(__reg, __val) \
41 	__asm__ __volatile__("stxa %0, [%1] %2" \
42 			     : /* no outputs */ \
43 			     : "r" (__val), "r" (__reg), \
44 			       "i" (ASI_PHYS_BYPASS_EC_E))
45 
46 /* Must be invoked under the IOMMU lock. */
iommu_flushall(struct iommu * iommu)47 static void iommu_flushall(struct iommu *iommu)
48 {
49 	if (iommu->iommu_flushinv) {
50 		iommu_write(iommu->iommu_flushinv, ~(u64)0);
51 	} else {
52 		unsigned long tag;
53 		int entry;
54 
55 		tag = iommu->iommu_tags;
56 		for (entry = 0; entry < 16; entry++) {
57 			iommu_write(tag, 0);
58 			tag += 8;
59 		}
60 
61 		/* Ensure completion of previous PIO writes. */
62 		(void) iommu_read(iommu->write_complete_reg);
63 	}
64 }
65 
66 #define IOPTE_CONSISTENT(CTX) \
67 	(IOPTE_VALID | IOPTE_CACHE | \
68 	 (((CTX) << 47) & IOPTE_CONTEXT))
69 
70 #define IOPTE_STREAMING(CTX) \
71 	(IOPTE_CONSISTENT(CTX) | IOPTE_STBUF)
72 
73 /* Existing mappings are never marked invalid, instead they
74  * are pointed to a dummy page.
75  */
76 #define IOPTE_IS_DUMMY(iommu, iopte)	\
77 	((iopte_val(*iopte) & IOPTE_PAGE) == (iommu)->dummy_page_pa)
78 
iopte_make_dummy(struct iommu * iommu,iopte_t * iopte)79 static inline void iopte_make_dummy(struct iommu *iommu, iopte_t *iopte)
80 {
81 	unsigned long val = iopte_val(*iopte);
82 
83 	val &= ~IOPTE_PAGE;
84 	val |= iommu->dummy_page_pa;
85 
86 	iopte_val(*iopte) = val;
87 }
88 
89 /* Based almost entirely upon the ppc64 iommu allocator.  If you use the 'handle'
90  * facility it must all be done in one pass while under the iommu lock.
91  *
92  * On sun4u platforms, we only flush the IOMMU once every time we've passed
93  * over the entire page table doing allocations.  Therefore we only ever advance
94  * the hint and cannot backtrack it.
95  */
iommu_range_alloc(struct device * dev,struct iommu * iommu,unsigned long npages,unsigned long * handle)96 unsigned long iommu_range_alloc(struct device *dev,
97 				struct iommu *iommu,
98 				unsigned long npages,
99 				unsigned long *handle)
100 {
101 	unsigned long n, end, start, limit, boundary_size;
102 	struct iommu_arena *arena = &iommu->arena;
103 	int pass = 0;
104 
105 	/* This allocator was derived from x86_64's bit string search */
106 
107 	/* Sanity check */
108 	if (unlikely(npages == 0)) {
109 		if (printk_ratelimit())
110 			WARN_ON(1);
111 		return DMA_ERROR_CODE;
112 	}
113 
114 	if (handle && *handle)
115 		start = *handle;
116 	else
117 		start = arena->hint;
118 
119 	limit = arena->limit;
120 
121 	/* The case below can happen if we have a small segment appended
122 	 * to a large, or when the previous alloc was at the very end of
123 	 * the available space. If so, go back to the beginning and flush.
124 	 */
125 	if (start >= limit) {
126 		start = 0;
127 		if (iommu->flush_all)
128 			iommu->flush_all(iommu);
129 	}
130 
131  again:
132 
133 	if (dev)
134 		boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
135 				      1 << IO_PAGE_SHIFT);
136 	else
137 		boundary_size = ALIGN(1UL << 32, 1 << IO_PAGE_SHIFT);
138 
139 	n = iommu_area_alloc(arena->map, limit, start, npages,
140 			     iommu->page_table_map_base >> IO_PAGE_SHIFT,
141 			     boundary_size >> IO_PAGE_SHIFT, 0);
142 	if (n == -1) {
143 		if (likely(pass < 1)) {
144 			/* First failure, rescan from the beginning.  */
145 			start = 0;
146 			if (iommu->flush_all)
147 				iommu->flush_all(iommu);
148 			pass++;
149 			goto again;
150 		} else {
151 			/* Second failure, give up */
152 			return DMA_ERROR_CODE;
153 		}
154 	}
155 
156 	end = n + npages;
157 
158 	arena->hint = end;
159 
160 	/* Update handle for SG allocations */
161 	if (handle)
162 		*handle = end;
163 
164 	return n;
165 }
166 
iommu_range_free(struct iommu * iommu,dma_addr_t dma_addr,unsigned long npages)167 void iommu_range_free(struct iommu *iommu, dma_addr_t dma_addr, unsigned long npages)
168 {
169 	struct iommu_arena *arena = &iommu->arena;
170 	unsigned long entry;
171 
172 	entry = (dma_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT;
173 
174 	bitmap_clear(arena->map, entry, npages);
175 }
176 
iommu_table_init(struct iommu * iommu,int tsbsize,u32 dma_offset,u32 dma_addr_mask,int numa_node)177 int iommu_table_init(struct iommu *iommu, int tsbsize,
178 		     u32 dma_offset, u32 dma_addr_mask,
179 		     int numa_node)
180 {
181 	unsigned long i, order, sz, num_tsb_entries;
182 	struct page *page;
183 
184 	num_tsb_entries = tsbsize / sizeof(iopte_t);
185 
186 	/* Setup initial software IOMMU state. */
187 	spin_lock_init(&iommu->lock);
188 	iommu->ctx_lowest_free = 1;
189 	iommu->page_table_map_base = dma_offset;
190 	iommu->dma_addr_mask = dma_addr_mask;
191 
192 	/* Allocate and initialize the free area map.  */
193 	sz = num_tsb_entries / 8;
194 	sz = (sz + 7UL) & ~7UL;
195 	iommu->arena.map = kmalloc_node(sz, GFP_KERNEL, numa_node);
196 	if (!iommu->arena.map) {
197 		printk(KERN_ERR "IOMMU: Error, kmalloc(arena.map) failed.\n");
198 		return -ENOMEM;
199 	}
200 	memset(iommu->arena.map, 0, sz);
201 	iommu->arena.limit = num_tsb_entries;
202 
203 	if (tlb_type != hypervisor)
204 		iommu->flush_all = iommu_flushall;
205 
206 	/* Allocate and initialize the dummy page which we
207 	 * set inactive IO PTEs to point to.
208 	 */
209 	page = alloc_pages_node(numa_node, GFP_KERNEL, 0);
210 	if (!page) {
211 		printk(KERN_ERR "IOMMU: Error, gfp(dummy_page) failed.\n");
212 		goto out_free_map;
213 	}
214 	iommu->dummy_page = (unsigned long) page_address(page);
215 	memset((void *)iommu->dummy_page, 0, PAGE_SIZE);
216 	iommu->dummy_page_pa = (unsigned long) __pa(iommu->dummy_page);
217 
218 	/* Now allocate and setup the IOMMU page table itself.  */
219 	order = get_order(tsbsize);
220 	page = alloc_pages_node(numa_node, GFP_KERNEL, order);
221 	if (!page) {
222 		printk(KERN_ERR "IOMMU: Error, gfp(tsb) failed.\n");
223 		goto out_free_dummy_page;
224 	}
225 	iommu->page_table = (iopte_t *)page_address(page);
226 
227 	for (i = 0; i < num_tsb_entries; i++)
228 		iopte_make_dummy(iommu, &iommu->page_table[i]);
229 
230 	return 0;
231 
232 out_free_dummy_page:
233 	free_page(iommu->dummy_page);
234 	iommu->dummy_page = 0UL;
235 
236 out_free_map:
237 	kfree(iommu->arena.map);
238 	iommu->arena.map = NULL;
239 
240 	return -ENOMEM;
241 }
242 
alloc_npages(struct device * dev,struct iommu * iommu,unsigned long npages)243 static inline iopte_t *alloc_npages(struct device *dev, struct iommu *iommu,
244 				    unsigned long npages)
245 {
246 	unsigned long entry;
247 
248 	entry = iommu_range_alloc(dev, iommu, npages, NULL);
249 	if (unlikely(entry == DMA_ERROR_CODE))
250 		return NULL;
251 
252 	return iommu->page_table + entry;
253 }
254 
iommu_alloc_ctx(struct iommu * iommu)255 static int iommu_alloc_ctx(struct iommu *iommu)
256 {
257 	int lowest = iommu->ctx_lowest_free;
258 	int n = find_next_zero_bit(iommu->ctx_bitmap, IOMMU_NUM_CTXS, lowest);
259 
260 	if (unlikely(n == IOMMU_NUM_CTXS)) {
261 		n = find_next_zero_bit(iommu->ctx_bitmap, lowest, 1);
262 		if (unlikely(n == lowest)) {
263 			printk(KERN_WARNING "IOMMU: Ran out of contexts.\n");
264 			n = 0;
265 		}
266 	}
267 	if (n)
268 		__set_bit(n, iommu->ctx_bitmap);
269 
270 	return n;
271 }
272 
iommu_free_ctx(struct iommu * iommu,int ctx)273 static inline void iommu_free_ctx(struct iommu *iommu, int ctx)
274 {
275 	if (likely(ctx)) {
276 		__clear_bit(ctx, iommu->ctx_bitmap);
277 		if (ctx < iommu->ctx_lowest_free)
278 			iommu->ctx_lowest_free = ctx;
279 	}
280 }
281 
dma_4u_alloc_coherent(struct device * dev,size_t size,dma_addr_t * dma_addrp,gfp_t gfp)282 static void *dma_4u_alloc_coherent(struct device *dev, size_t size,
283 				   dma_addr_t *dma_addrp, gfp_t gfp)
284 {
285 	unsigned long flags, order, first_page;
286 	struct iommu *iommu;
287 	struct page *page;
288 	int npages, nid;
289 	iopte_t *iopte;
290 	void *ret;
291 
292 	size = IO_PAGE_ALIGN(size);
293 	order = get_order(size);
294 	if (order >= 10)
295 		return NULL;
296 
297 	nid = dev->archdata.numa_node;
298 	page = alloc_pages_node(nid, gfp, order);
299 	if (unlikely(!page))
300 		return NULL;
301 
302 	first_page = (unsigned long) page_address(page);
303 	memset((char *)first_page, 0, PAGE_SIZE << order);
304 
305 	iommu = dev->archdata.iommu;
306 
307 	spin_lock_irqsave(&iommu->lock, flags);
308 	iopte = alloc_npages(dev, iommu, size >> IO_PAGE_SHIFT);
309 	spin_unlock_irqrestore(&iommu->lock, flags);
310 
311 	if (unlikely(iopte == NULL)) {
312 		free_pages(first_page, order);
313 		return NULL;
314 	}
315 
316 	*dma_addrp = (iommu->page_table_map_base +
317 		      ((iopte - iommu->page_table) << IO_PAGE_SHIFT));
318 	ret = (void *) first_page;
319 	npages = size >> IO_PAGE_SHIFT;
320 	first_page = __pa(first_page);
321 	while (npages--) {
322 		iopte_val(*iopte) = (IOPTE_CONSISTENT(0UL) |
323 				     IOPTE_WRITE |
324 				     (first_page & IOPTE_PAGE));
325 		iopte++;
326 		first_page += IO_PAGE_SIZE;
327 	}
328 
329 	return ret;
330 }
331 
dma_4u_free_coherent(struct device * dev,size_t size,void * cpu,dma_addr_t dvma)332 static void dma_4u_free_coherent(struct device *dev, size_t size,
333 				 void *cpu, dma_addr_t dvma)
334 {
335 	struct iommu *iommu;
336 	unsigned long flags, order, npages;
337 
338 	npages = IO_PAGE_ALIGN(size) >> IO_PAGE_SHIFT;
339 	iommu = dev->archdata.iommu;
340 
341 	spin_lock_irqsave(&iommu->lock, flags);
342 
343 	iommu_range_free(iommu, dvma, npages);
344 
345 	spin_unlock_irqrestore(&iommu->lock, flags);
346 
347 	order = get_order(size);
348 	if (order < 10)
349 		free_pages((unsigned long)cpu, order);
350 }
351 
dma_4u_map_page(struct device * dev,struct page * page,unsigned long offset,size_t sz,enum dma_data_direction direction,struct dma_attrs * attrs)352 static dma_addr_t dma_4u_map_page(struct device *dev, struct page *page,
353 				  unsigned long offset, size_t sz,
354 				  enum dma_data_direction direction,
355 				  struct dma_attrs *attrs)
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)(page_address(page) + offset);
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 
strbuf_flush(struct strbuf * strbuf,struct iommu * iommu,u32 vaddr,unsigned long ctx,unsigned long npages,enum dma_data_direction direction)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 
dma_4u_unmap_page(struct device * dev,dma_addr_t bus_addr,size_t sz,enum dma_data_direction direction,struct dma_attrs * attrs)475 static void dma_4u_unmap_page(struct device *dev, dma_addr_t bus_addr,
476 			      size_t sz, enum dma_data_direction direction,
477 			      struct dma_attrs *attrs)
478 {
479 	struct iommu *iommu;
480 	struct strbuf *strbuf;
481 	iopte_t *base;
482 	unsigned long flags, npages, ctx, i;
483 
484 	if (unlikely(direction == DMA_NONE)) {
485 		if (printk_ratelimit())
486 			WARN_ON(1);
487 		return;
488 	}
489 
490 	iommu = dev->archdata.iommu;
491 	strbuf = dev->archdata.stc;
492 
493 	npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
494 	npages >>= IO_PAGE_SHIFT;
495 	base = iommu->page_table +
496 		((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
497 	bus_addr &= IO_PAGE_MASK;
498 
499 	spin_lock_irqsave(&iommu->lock, flags);
500 
501 	/* Record the context, if any. */
502 	ctx = 0;
503 	if (iommu->iommu_ctxflush)
504 		ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL;
505 
506 	/* Step 1: Kick data out of streaming buffers if necessary. */
507 	if (strbuf->strbuf_enabled)
508 		strbuf_flush(strbuf, iommu, bus_addr, ctx,
509 			     npages, direction);
510 
511 	/* Step 2: Clear out TSB entries. */
512 	for (i = 0; i < npages; i++)
513 		iopte_make_dummy(iommu, base + i);
514 
515 	iommu_range_free(iommu, bus_addr, npages);
516 
517 	iommu_free_ctx(iommu, ctx);
518 
519 	spin_unlock_irqrestore(&iommu->lock, flags);
520 }
521 
dma_4u_map_sg(struct device * dev,struct scatterlist * sglist,int nelems,enum dma_data_direction direction,struct dma_attrs * attrs)522 static int dma_4u_map_sg(struct device *dev, struct scatterlist *sglist,
523 			 int nelems, enum dma_data_direction direction,
524 			 struct dma_attrs *attrs)
525 {
526 	struct scatterlist *s, *outs, *segstart;
527 	unsigned long flags, handle, prot, ctx;
528 	dma_addr_t dma_next = 0, dma_addr;
529 	unsigned int max_seg_size;
530 	unsigned long seg_boundary_size;
531 	int outcount, incount, i;
532 	struct strbuf *strbuf;
533 	struct iommu *iommu;
534 	unsigned long base_shift;
535 
536 	BUG_ON(direction == DMA_NONE);
537 
538 	iommu = dev->archdata.iommu;
539 	strbuf = dev->archdata.stc;
540 	if (nelems == 0 || !iommu)
541 		return 0;
542 
543 	spin_lock_irqsave(&iommu->lock, flags);
544 
545 	ctx = 0;
546 	if (iommu->iommu_ctxflush)
547 		ctx = iommu_alloc_ctx(iommu);
548 
549 	if (strbuf->strbuf_enabled)
550 		prot = IOPTE_STREAMING(ctx);
551 	else
552 		prot = IOPTE_CONSISTENT(ctx);
553 	if (direction != DMA_TO_DEVICE)
554 		prot |= IOPTE_WRITE;
555 
556 	outs = s = segstart = &sglist[0];
557 	outcount = 1;
558 	incount = nelems;
559 	handle = 0;
560 
561 	/* Init first segment length for backout at failure */
562 	outs->dma_length = 0;
563 
564 	max_seg_size = dma_get_max_seg_size(dev);
565 	seg_boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
566 				  IO_PAGE_SIZE) >> IO_PAGE_SHIFT;
567 	base_shift = iommu->page_table_map_base >> IO_PAGE_SHIFT;
568 	for_each_sg(sglist, s, nelems, i) {
569 		unsigned long paddr, npages, entry, out_entry = 0, slen;
570 		iopte_t *base;
571 
572 		slen = s->length;
573 		/* Sanity check */
574 		if (slen == 0) {
575 			dma_next = 0;
576 			continue;
577 		}
578 		/* Allocate iommu entries for that segment */
579 		paddr = (unsigned long) SG_ENT_PHYS_ADDRESS(s);
580 		npages = iommu_num_pages(paddr, slen, IO_PAGE_SIZE);
581 		entry = iommu_range_alloc(dev, iommu, npages, &handle);
582 
583 		/* Handle failure */
584 		if (unlikely(entry == DMA_ERROR_CODE)) {
585 			if (printk_ratelimit())
586 				printk(KERN_INFO "iommu_alloc failed, iommu %p paddr %lx"
587 				       " npages %lx\n", iommu, paddr, npages);
588 			goto iommu_map_failed;
589 		}
590 
591 		base = iommu->page_table + entry;
592 
593 		/* Convert entry to a dma_addr_t */
594 		dma_addr = iommu->page_table_map_base +
595 			(entry << IO_PAGE_SHIFT);
596 		dma_addr |= (s->offset & ~IO_PAGE_MASK);
597 
598 		/* Insert into HW table */
599 		paddr &= IO_PAGE_MASK;
600 		while (npages--) {
601 			iopte_val(*base) = prot | paddr;
602 			base++;
603 			paddr += IO_PAGE_SIZE;
604 		}
605 
606 		/* If we are in an open segment, try merging */
607 		if (segstart != s) {
608 			/* We cannot merge if:
609 			 * - allocated dma_addr isn't contiguous to previous allocation
610 			 */
611 			if ((dma_addr != dma_next) ||
612 			    (outs->dma_length + s->length > max_seg_size) ||
613 			    (is_span_boundary(out_entry, base_shift,
614 					      seg_boundary_size, outs, s))) {
615 				/* Can't merge: create a new segment */
616 				segstart = s;
617 				outcount++;
618 				outs = sg_next(outs);
619 			} else {
620 				outs->dma_length += s->length;
621 			}
622 		}
623 
624 		if (segstart == s) {
625 			/* This is a new segment, fill entries */
626 			outs->dma_address = dma_addr;
627 			outs->dma_length = slen;
628 			out_entry = entry;
629 		}
630 
631 		/* Calculate next page pointer for contiguous check */
632 		dma_next = dma_addr + slen;
633 	}
634 
635 	spin_unlock_irqrestore(&iommu->lock, flags);
636 
637 	if (outcount < incount) {
638 		outs = sg_next(outs);
639 		outs->dma_address = DMA_ERROR_CODE;
640 		outs->dma_length = 0;
641 	}
642 
643 	return outcount;
644 
645 iommu_map_failed:
646 	for_each_sg(sglist, s, nelems, i) {
647 		if (s->dma_length != 0) {
648 			unsigned long vaddr, npages, entry, j;
649 			iopte_t *base;
650 
651 			vaddr = s->dma_address & IO_PAGE_MASK;
652 			npages = iommu_num_pages(s->dma_address, s->dma_length,
653 						 IO_PAGE_SIZE);
654 			iommu_range_free(iommu, vaddr, npages);
655 
656 			entry = (vaddr - iommu->page_table_map_base)
657 				>> IO_PAGE_SHIFT;
658 			base = iommu->page_table + entry;
659 
660 			for (j = 0; j < npages; j++)
661 				iopte_make_dummy(iommu, base + j);
662 
663 			s->dma_address = DMA_ERROR_CODE;
664 			s->dma_length = 0;
665 		}
666 		if (s == outs)
667 			break;
668 	}
669 	spin_unlock_irqrestore(&iommu->lock, flags);
670 
671 	return 0;
672 }
673 
674 /* If contexts are being used, they are the same in all of the mappings
675  * we make for a particular SG.
676  */
fetch_sg_ctx(struct iommu * iommu,struct scatterlist * sg)677 static unsigned long fetch_sg_ctx(struct iommu *iommu, struct scatterlist *sg)
678 {
679 	unsigned long ctx = 0;
680 
681 	if (iommu->iommu_ctxflush) {
682 		iopte_t *base;
683 		u32 bus_addr;
684 
685 		bus_addr = sg->dma_address & IO_PAGE_MASK;
686 		base = iommu->page_table +
687 			((bus_addr - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
688 
689 		ctx = (iopte_val(*base) & IOPTE_CONTEXT) >> 47UL;
690 	}
691 	return ctx;
692 }
693 
dma_4u_unmap_sg(struct device * dev,struct scatterlist * sglist,int nelems,enum dma_data_direction direction,struct dma_attrs * attrs)694 static void dma_4u_unmap_sg(struct device *dev, struct scatterlist *sglist,
695 			    int nelems, enum dma_data_direction direction,
696 			    struct dma_attrs *attrs)
697 {
698 	unsigned long flags, ctx;
699 	struct scatterlist *sg;
700 	struct strbuf *strbuf;
701 	struct iommu *iommu;
702 
703 	BUG_ON(direction == DMA_NONE);
704 
705 	iommu = dev->archdata.iommu;
706 	strbuf = dev->archdata.stc;
707 
708 	ctx = fetch_sg_ctx(iommu, sglist);
709 
710 	spin_lock_irqsave(&iommu->lock, flags);
711 
712 	sg = sglist;
713 	while (nelems--) {
714 		dma_addr_t dma_handle = sg->dma_address;
715 		unsigned int len = sg->dma_length;
716 		unsigned long npages, entry;
717 		iopte_t *base;
718 		int i;
719 
720 		if (!len)
721 			break;
722 		npages = iommu_num_pages(dma_handle, len, IO_PAGE_SIZE);
723 		iommu_range_free(iommu, dma_handle, npages);
724 
725 		entry = ((dma_handle - iommu->page_table_map_base)
726 			 >> IO_PAGE_SHIFT);
727 		base = iommu->page_table + entry;
728 
729 		dma_handle &= IO_PAGE_MASK;
730 		if (strbuf->strbuf_enabled)
731 			strbuf_flush(strbuf, iommu, dma_handle, ctx,
732 				     npages, direction);
733 
734 		for (i = 0; i < npages; i++)
735 			iopte_make_dummy(iommu, base + i);
736 
737 		sg = sg_next(sg);
738 	}
739 
740 	iommu_free_ctx(iommu, ctx);
741 
742 	spin_unlock_irqrestore(&iommu->lock, flags);
743 }
744 
dma_4u_sync_single_for_cpu(struct device * dev,dma_addr_t bus_addr,size_t sz,enum dma_data_direction direction)745 static void dma_4u_sync_single_for_cpu(struct device *dev,
746 				       dma_addr_t bus_addr, size_t sz,
747 				       enum dma_data_direction direction)
748 {
749 	struct iommu *iommu;
750 	struct strbuf *strbuf;
751 	unsigned long flags, ctx, npages;
752 
753 	iommu = dev->archdata.iommu;
754 	strbuf = dev->archdata.stc;
755 
756 	if (!strbuf->strbuf_enabled)
757 		return;
758 
759 	spin_lock_irqsave(&iommu->lock, flags);
760 
761 	npages = IO_PAGE_ALIGN(bus_addr + sz) - (bus_addr & IO_PAGE_MASK);
762 	npages >>= IO_PAGE_SHIFT;
763 	bus_addr &= IO_PAGE_MASK;
764 
765 	/* Step 1: Record the context, if any. */
766 	ctx = 0;
767 	if (iommu->iommu_ctxflush &&
768 	    strbuf->strbuf_ctxflush) {
769 		iopte_t *iopte;
770 
771 		iopte = iommu->page_table +
772 			((bus_addr - iommu->page_table_map_base)>>IO_PAGE_SHIFT);
773 		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
774 	}
775 
776 	/* Step 2: Kick data out of streaming buffers. */
777 	strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
778 
779 	spin_unlock_irqrestore(&iommu->lock, flags);
780 }
781 
dma_4u_sync_sg_for_cpu(struct device * dev,struct scatterlist * sglist,int nelems,enum dma_data_direction direction)782 static void dma_4u_sync_sg_for_cpu(struct device *dev,
783 				   struct scatterlist *sglist, int nelems,
784 				   enum dma_data_direction direction)
785 {
786 	struct iommu *iommu;
787 	struct strbuf *strbuf;
788 	unsigned long flags, ctx, npages, i;
789 	struct scatterlist *sg, *sgprv;
790 	u32 bus_addr;
791 
792 	iommu = dev->archdata.iommu;
793 	strbuf = dev->archdata.stc;
794 
795 	if (!strbuf->strbuf_enabled)
796 		return;
797 
798 	spin_lock_irqsave(&iommu->lock, flags);
799 
800 	/* Step 1: Record the context, if any. */
801 	ctx = 0;
802 	if (iommu->iommu_ctxflush &&
803 	    strbuf->strbuf_ctxflush) {
804 		iopte_t *iopte;
805 
806 		iopte = iommu->page_table +
807 			((sglist[0].dma_address - iommu->page_table_map_base) >> IO_PAGE_SHIFT);
808 		ctx = (iopte_val(*iopte) & IOPTE_CONTEXT) >> 47UL;
809 	}
810 
811 	/* Step 2: Kick data out of streaming buffers. */
812 	bus_addr = sglist[0].dma_address & IO_PAGE_MASK;
813 	sgprv = NULL;
814 	for_each_sg(sglist, sg, nelems, i) {
815 		if (sg->dma_length == 0)
816 			break;
817 		sgprv = sg;
818 	}
819 
820 	npages = (IO_PAGE_ALIGN(sgprv->dma_address + sgprv->dma_length)
821 		  - bus_addr) >> IO_PAGE_SHIFT;
822 	strbuf_flush(strbuf, iommu, bus_addr, ctx, npages, direction);
823 
824 	spin_unlock_irqrestore(&iommu->lock, flags);
825 }
826 
827 static struct dma_map_ops sun4u_dma_ops = {
828 	.alloc_coherent		= dma_4u_alloc_coherent,
829 	.free_coherent		= dma_4u_free_coherent,
830 	.map_page		= dma_4u_map_page,
831 	.unmap_page		= dma_4u_unmap_page,
832 	.map_sg			= dma_4u_map_sg,
833 	.unmap_sg		= dma_4u_unmap_sg,
834 	.sync_single_for_cpu	= dma_4u_sync_single_for_cpu,
835 	.sync_sg_for_cpu	= dma_4u_sync_sg_for_cpu,
836 };
837 
838 struct dma_map_ops *dma_ops = &sun4u_dma_ops;
839 EXPORT_SYMBOL(dma_ops);
840 
841 extern int pci64_dma_supported(struct pci_dev *pdev, u64 device_mask);
842 
dma_supported(struct device * dev,u64 device_mask)843 int dma_supported(struct device *dev, u64 device_mask)
844 {
845 	struct iommu *iommu = dev->archdata.iommu;
846 	u64 dma_addr_mask = iommu->dma_addr_mask;
847 
848 	if (device_mask >= (1UL << 32UL))
849 		return 0;
850 
851 	if ((device_mask & dma_addr_mask) == dma_addr_mask)
852 		return 1;
853 
854 #ifdef CONFIG_PCI
855 	if (dev->bus == &pci_bus_type)
856 		return pci64_dma_supported(to_pci_dev(dev), device_mask);
857 #endif
858 
859 	return 0;
860 }
861 EXPORT_SYMBOL(dma_supported);
862