xref: /openbmc/linux/drivers/xen/swiotlb-xen.c (revision 552b8b36)
1 /*
2  *  Copyright 2010
3  *  by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
4  *
5  * This code provides a IOMMU for Xen PV guests with PCI passthrough.
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License v2.0 as published by
9  * the Free Software Foundation
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * PV guests under Xen are running in an non-contiguous memory architecture.
17  *
18  * When PCI pass-through is utilized, this necessitates an IOMMU for
19  * translating bus (DMA) to virtual and vice-versa and also providing a
20  * mechanism to have contiguous pages for device drivers operations (say DMA
21  * operations).
22  *
23  * Specifically, under Xen the Linux idea of pages is an illusion. It
24  * assumes that pages start at zero and go up to the available memory. To
25  * help with that, the Linux Xen MMU provides a lookup mechanism to
26  * translate the page frame numbers (PFN) to machine frame numbers (MFN)
27  * and vice-versa. The MFN are the "real" frame numbers. Furthermore
28  * memory is not contiguous. Xen hypervisor stitches memory for guests
29  * from different pools, which means there is no guarantee that PFN==MFN
30  * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
31  * allocated in descending order (high to low), meaning the guest might
32  * never get any MFN's under the 4GB mark.
33  *
34  */
35 
36 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
37 
38 #include <linux/bootmem.h>
39 #include <linux/dma-mapping.h>
40 #include <linux/export.h>
41 #include <xen/swiotlb-xen.h>
42 #include <xen/page.h>
43 #include <xen/xen-ops.h>
44 #include <xen/hvc-console.h>
45 
46 #include <asm/dma-mapping.h>
47 #include <asm/xen/page-coherent.h>
48 
49 #include <trace/events/swiotlb.h>
50 /*
51  * Used to do a quick range check in swiotlb_tbl_unmap_single and
52  * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
53  * API.
54  */
55 
56 #ifndef CONFIG_X86
57 static unsigned long dma_alloc_coherent_mask(struct device *dev,
58 					    gfp_t gfp)
59 {
60 	unsigned long dma_mask = 0;
61 
62 	dma_mask = dev->coherent_dma_mask;
63 	if (!dma_mask)
64 		dma_mask = (gfp & GFP_DMA) ? DMA_BIT_MASK(24) : DMA_BIT_MASK(32);
65 
66 	return dma_mask;
67 }
68 #endif
69 
70 static char *xen_io_tlb_start, *xen_io_tlb_end;
71 static unsigned long xen_io_tlb_nslabs;
72 /*
73  * Quick lookup value of the bus address of the IOTLB.
74  */
75 
76 static u64 start_dma_addr;
77 
78 /*
79  * Both of these functions should avoid PFN_PHYS because phys_addr_t
80  * can be 32bit when dma_addr_t is 64bit leading to a loss in
81  * information if the shift is done before casting to 64bit.
82  */
83 static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
84 {
85 	unsigned long bfn = pfn_to_bfn(PFN_DOWN(paddr));
86 	dma_addr_t dma = (dma_addr_t)bfn << PAGE_SHIFT;
87 
88 	dma |= paddr & ~PAGE_MASK;
89 
90 	return dma;
91 }
92 
93 static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
94 {
95 	unsigned long pfn = bfn_to_pfn(PFN_DOWN(baddr));
96 	dma_addr_t dma = (dma_addr_t)pfn << PAGE_SHIFT;
97 	phys_addr_t paddr = dma;
98 
99 	paddr |= baddr & ~PAGE_MASK;
100 
101 	return paddr;
102 }
103 
104 static inline dma_addr_t xen_virt_to_bus(void *address)
105 {
106 	return xen_phys_to_bus(virt_to_phys(address));
107 }
108 
109 static int check_pages_physically_contiguous(unsigned long pfn,
110 					     unsigned int offset,
111 					     size_t length)
112 {
113 	unsigned long next_bfn;
114 	int i;
115 	int nr_pages;
116 
117 	next_bfn = pfn_to_bfn(pfn);
118 	nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
119 
120 	for (i = 1; i < nr_pages; i++) {
121 		if (pfn_to_bfn(++pfn) != ++next_bfn)
122 			return 0;
123 	}
124 	return 1;
125 }
126 
127 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
128 {
129 	unsigned long pfn = PFN_DOWN(p);
130 	unsigned int offset = p & ~PAGE_MASK;
131 
132 	if (offset + size <= PAGE_SIZE)
133 		return 0;
134 	if (check_pages_physically_contiguous(pfn, offset, size))
135 		return 0;
136 	return 1;
137 }
138 
139 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
140 {
141 	unsigned long bfn = PFN_DOWN(dma_addr);
142 	unsigned long pfn = bfn_to_local_pfn(bfn);
143 	phys_addr_t paddr;
144 
145 	/* If the address is outside our domain, it CAN
146 	 * have the same virtual address as another address
147 	 * in our domain. Therefore _only_ check address within our domain.
148 	 */
149 	if (pfn_valid(pfn)) {
150 		paddr = PFN_PHYS(pfn);
151 		return paddr >= virt_to_phys(xen_io_tlb_start) &&
152 		       paddr < virt_to_phys(xen_io_tlb_end);
153 	}
154 	return 0;
155 }
156 
157 static int max_dma_bits = 32;
158 
159 static int
160 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
161 {
162 	int i, rc;
163 	int dma_bits;
164 	dma_addr_t dma_handle;
165 	phys_addr_t p = virt_to_phys(buf);
166 
167 	dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
168 
169 	i = 0;
170 	do {
171 		int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
172 
173 		do {
174 			rc = xen_create_contiguous_region(
175 				p + (i << IO_TLB_SHIFT),
176 				get_order(slabs << IO_TLB_SHIFT),
177 				dma_bits, &dma_handle);
178 		} while (rc && dma_bits++ < max_dma_bits);
179 		if (rc)
180 			return rc;
181 
182 		i += slabs;
183 	} while (i < nslabs);
184 	return 0;
185 }
186 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
187 {
188 	if (!nr_tbl) {
189 		xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
190 		xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
191 	} else
192 		xen_io_tlb_nslabs = nr_tbl;
193 
194 	return xen_io_tlb_nslabs << IO_TLB_SHIFT;
195 }
196 
197 enum xen_swiotlb_err {
198 	XEN_SWIOTLB_UNKNOWN = 0,
199 	XEN_SWIOTLB_ENOMEM,
200 	XEN_SWIOTLB_EFIXUP
201 };
202 
203 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
204 {
205 	switch (err) {
206 	case XEN_SWIOTLB_ENOMEM:
207 		return "Cannot allocate Xen-SWIOTLB buffer\n";
208 	case XEN_SWIOTLB_EFIXUP:
209 		return "Failed to get contiguous memory for DMA from Xen!\n"\
210 		    "You either: don't have the permissions, do not have"\
211 		    " enough free memory under 4GB, or the hypervisor memory"\
212 		    " is too fragmented!";
213 	default:
214 		break;
215 	}
216 	return "";
217 }
218 int __ref xen_swiotlb_init(int verbose, bool early)
219 {
220 	unsigned long bytes, order;
221 	int rc = -ENOMEM;
222 	enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
223 	unsigned int repeat = 3;
224 
225 	xen_io_tlb_nslabs = swiotlb_nr_tbl();
226 retry:
227 	bytes = xen_set_nslabs(xen_io_tlb_nslabs);
228 	order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
229 	/*
230 	 * Get IO TLB memory from any location.
231 	 */
232 	if (early)
233 		xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
234 	else {
235 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
236 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
237 		while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
238 			xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
239 			if (xen_io_tlb_start)
240 				break;
241 			order--;
242 		}
243 		if (order != get_order(bytes)) {
244 			pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
245 				(PAGE_SIZE << order) >> 20);
246 			xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
247 			bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
248 		}
249 	}
250 	if (!xen_io_tlb_start) {
251 		m_ret = XEN_SWIOTLB_ENOMEM;
252 		goto error;
253 	}
254 	xen_io_tlb_end = xen_io_tlb_start + bytes;
255 	/*
256 	 * And replace that memory with pages under 4GB.
257 	 */
258 	rc = xen_swiotlb_fixup(xen_io_tlb_start,
259 			       bytes,
260 			       xen_io_tlb_nslabs);
261 	if (rc) {
262 		if (early)
263 			free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
264 		else {
265 			free_pages((unsigned long)xen_io_tlb_start, order);
266 			xen_io_tlb_start = NULL;
267 		}
268 		m_ret = XEN_SWIOTLB_EFIXUP;
269 		goto error;
270 	}
271 	start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
272 	if (early) {
273 		if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
274 			 verbose))
275 			panic("Cannot allocate SWIOTLB buffer");
276 		rc = 0;
277 	} else
278 		rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
279 	return rc;
280 error:
281 	if (repeat--) {
282 		xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
283 					(xen_io_tlb_nslabs >> 1));
284 		pr_info("Lowering to %luMB\n",
285 			(xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
286 		goto retry;
287 	}
288 	pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
289 	if (early)
290 		panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
291 	else
292 		free_pages((unsigned long)xen_io_tlb_start, order);
293 	return rc;
294 }
295 void *
296 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
297 			   dma_addr_t *dma_handle, gfp_t flags,
298 			   struct dma_attrs *attrs)
299 {
300 	void *ret;
301 	int order = get_order(size);
302 	u64 dma_mask = DMA_BIT_MASK(32);
303 	phys_addr_t phys;
304 	dma_addr_t dev_addr;
305 
306 	/*
307 	* Ignore region specifiers - the kernel's ideas of
308 	* pseudo-phys memory layout has nothing to do with the
309 	* machine physical layout.  We can't allocate highmem
310 	* because we can't return a pointer to it.
311 	*/
312 	flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
313 
314 	/* On ARM this function returns an ioremap'ped virtual address for
315 	 * which virt_to_phys doesn't return the corresponding physical
316 	 * address. In fact on ARM virt_to_phys only works for kernel direct
317 	 * mapped RAM memory. Also see comment below.
318 	 */
319 	ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
320 
321 	if (!ret)
322 		return ret;
323 
324 	if (hwdev && hwdev->coherent_dma_mask)
325 		dma_mask = dma_alloc_coherent_mask(hwdev, flags);
326 
327 	/* At this point dma_handle is the physical address, next we are
328 	 * going to set it to the machine address.
329 	 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
330 	 * to *dma_handle. */
331 	phys = *dma_handle;
332 	dev_addr = xen_phys_to_bus(phys);
333 	if (((dev_addr + size - 1 <= dma_mask)) &&
334 	    !range_straddles_page_boundary(phys, size))
335 		*dma_handle = dev_addr;
336 	else {
337 		if (xen_create_contiguous_region(phys, order,
338 						 fls64(dma_mask), dma_handle) != 0) {
339 			xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
340 			return NULL;
341 		}
342 	}
343 	memset(ret, 0, size);
344 	return ret;
345 }
346 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
347 
348 void
349 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
350 			  dma_addr_t dev_addr, struct dma_attrs *attrs)
351 {
352 	int order = get_order(size);
353 	phys_addr_t phys;
354 	u64 dma_mask = DMA_BIT_MASK(32);
355 
356 	if (hwdev && hwdev->coherent_dma_mask)
357 		dma_mask = hwdev->coherent_dma_mask;
358 
359 	/* do not use virt_to_phys because on ARM it doesn't return you the
360 	 * physical address */
361 	phys = xen_bus_to_phys(dev_addr);
362 
363 	if (((dev_addr + size - 1 > dma_mask)) ||
364 	    range_straddles_page_boundary(phys, size))
365 		xen_destroy_contiguous_region(phys, order);
366 
367 	xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
368 }
369 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
370 
371 
372 /*
373  * Map a single buffer of the indicated size for DMA in streaming mode.  The
374  * physical address to use is returned.
375  *
376  * Once the device is given the dma address, the device owns this memory until
377  * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
378  */
379 dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
380 				unsigned long offset, size_t size,
381 				enum dma_data_direction dir,
382 				struct dma_attrs *attrs)
383 {
384 	phys_addr_t map, phys = page_to_phys(page) + offset;
385 	dma_addr_t dev_addr = xen_phys_to_bus(phys);
386 
387 	BUG_ON(dir == DMA_NONE);
388 	/*
389 	 * If the address happens to be in the device's DMA window,
390 	 * we can safely return the device addr and not worry about bounce
391 	 * buffering it.
392 	 */
393 	if (dma_capable(dev, dev_addr, size) &&
394 	    !range_straddles_page_boundary(phys, size) &&
395 		!xen_arch_need_swiotlb(dev, PFN_DOWN(phys), PFN_DOWN(dev_addr)) &&
396 		!swiotlb_force) {
397 		/* we are not interested in the dma_addr returned by
398 		 * xen_dma_map_page, only in the potential cache flushes executed
399 		 * by the function. */
400 		xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
401 		return dev_addr;
402 	}
403 
404 	/*
405 	 * Oh well, have to allocate and map a bounce buffer.
406 	 */
407 	trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
408 
409 	map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
410 	if (map == SWIOTLB_MAP_ERROR)
411 		return DMA_ERROR_CODE;
412 
413 	xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
414 					dev_addr, map & ~PAGE_MASK, size, dir, attrs);
415 	dev_addr = xen_phys_to_bus(map);
416 
417 	/*
418 	 * Ensure that the address returned is DMA'ble
419 	 */
420 	if (!dma_capable(dev, dev_addr, size)) {
421 		swiotlb_tbl_unmap_single(dev, map, size, dir);
422 		dev_addr = 0;
423 	}
424 	return dev_addr;
425 }
426 EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
427 
428 /*
429  * Unmap a single streaming mode DMA translation.  The dma_addr and size must
430  * match what was provided for in a previous xen_swiotlb_map_page call.  All
431  * other usages are undefined.
432  *
433  * After this call, reads by the cpu to the buffer are guaranteed to see
434  * whatever the device wrote there.
435  */
436 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
437 			     size_t size, enum dma_data_direction dir,
438 				 struct dma_attrs *attrs)
439 {
440 	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
441 
442 	BUG_ON(dir == DMA_NONE);
443 
444 	xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
445 
446 	/* NOTE: We use dev_addr here, not paddr! */
447 	if (is_xen_swiotlb_buffer(dev_addr)) {
448 		swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
449 		return;
450 	}
451 
452 	if (dir != DMA_FROM_DEVICE)
453 		return;
454 
455 	/*
456 	 * phys_to_virt doesn't work with hihgmem page but we could
457 	 * call dma_mark_clean() with hihgmem page here. However, we
458 	 * are fine since dma_mark_clean() is null on POWERPC. We can
459 	 * make dma_mark_clean() take a physical address if necessary.
460 	 */
461 	dma_mark_clean(phys_to_virt(paddr), size);
462 }
463 
464 void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
465 			    size_t size, enum dma_data_direction dir,
466 			    struct dma_attrs *attrs)
467 {
468 	xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
469 }
470 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
471 
472 /*
473  * Make physical memory consistent for a single streaming mode DMA translation
474  * after a transfer.
475  *
476  * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
477  * using the cpu, yet do not wish to teardown the dma mapping, you must
478  * call this function before doing so.  At the next point you give the dma
479  * address back to the card, you must first perform a
480  * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
481  */
482 static void
483 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
484 			size_t size, enum dma_data_direction dir,
485 			enum dma_sync_target target)
486 {
487 	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
488 
489 	BUG_ON(dir == DMA_NONE);
490 
491 	if (target == SYNC_FOR_CPU)
492 		xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
493 
494 	/* NOTE: We use dev_addr here, not paddr! */
495 	if (is_xen_swiotlb_buffer(dev_addr))
496 		swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
497 
498 	if (target == SYNC_FOR_DEVICE)
499 		xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
500 
501 	if (dir != DMA_FROM_DEVICE)
502 		return;
503 
504 	dma_mark_clean(phys_to_virt(paddr), size);
505 }
506 
507 void
508 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
509 				size_t size, enum dma_data_direction dir)
510 {
511 	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
512 }
513 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
514 
515 void
516 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
517 				   size_t size, enum dma_data_direction dir)
518 {
519 	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
520 }
521 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
522 
523 /*
524  * Map a set of buffers described by scatterlist in streaming mode for DMA.
525  * This is the scatter-gather version of the above xen_swiotlb_map_page
526  * interface.  Here the scatter gather list elements are each tagged with the
527  * appropriate dma address and length.  They are obtained via
528  * sg_dma_{address,length}(SG).
529  *
530  * NOTE: An implementation may be able to use a smaller number of
531  *       DMA address/length pairs than there are SG table elements.
532  *       (for example via virtual mapping capabilities)
533  *       The routine returns the number of addr/length pairs actually
534  *       used, at most nents.
535  *
536  * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
537  * same here.
538  */
539 int
540 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
541 			 int nelems, enum dma_data_direction dir,
542 			 struct dma_attrs *attrs)
543 {
544 	struct scatterlist *sg;
545 	int i;
546 
547 	BUG_ON(dir == DMA_NONE);
548 
549 	for_each_sg(sgl, sg, nelems, i) {
550 		phys_addr_t paddr = sg_phys(sg);
551 		dma_addr_t dev_addr = xen_phys_to_bus(paddr);
552 
553 		if (swiotlb_force ||
554 		    xen_arch_need_swiotlb(hwdev, PFN_DOWN(paddr), PFN_DOWN(dev_addr)) ||
555 		    !dma_capable(hwdev, dev_addr, sg->length) ||
556 		    range_straddles_page_boundary(paddr, sg->length)) {
557 			phys_addr_t map = swiotlb_tbl_map_single(hwdev,
558 								 start_dma_addr,
559 								 sg_phys(sg),
560 								 sg->length,
561 								 dir);
562 			if (map == SWIOTLB_MAP_ERROR) {
563 				dev_warn(hwdev, "swiotlb buffer is full\n");
564 				/* Don't panic here, we expect map_sg users
565 				   to do proper error handling. */
566 				xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
567 							   attrs);
568 				sg_dma_len(sgl) = 0;
569 				return 0;
570 			}
571 			xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
572 						dev_addr,
573 						map & ~PAGE_MASK,
574 						sg->length,
575 						dir,
576 						attrs);
577 			sg->dma_address = xen_phys_to_bus(map);
578 		} else {
579 			/* we are not interested in the dma_addr returned by
580 			 * xen_dma_map_page, only in the potential cache flushes executed
581 			 * by the function. */
582 			xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
583 						dev_addr,
584 						paddr & ~PAGE_MASK,
585 						sg->length,
586 						dir,
587 						attrs);
588 			sg->dma_address = dev_addr;
589 		}
590 		sg_dma_len(sg) = sg->length;
591 	}
592 	return nelems;
593 }
594 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
595 
596 /*
597  * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
598  * concerning calls here are the same as for swiotlb_unmap_page() above.
599  */
600 void
601 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
602 			   int nelems, enum dma_data_direction dir,
603 			   struct dma_attrs *attrs)
604 {
605 	struct scatterlist *sg;
606 	int i;
607 
608 	BUG_ON(dir == DMA_NONE);
609 
610 	for_each_sg(sgl, sg, nelems, i)
611 		xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
612 
613 }
614 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
615 
616 /*
617  * Make physical memory consistent for a set of streaming mode DMA translations
618  * after a transfer.
619  *
620  * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
621  * and usage.
622  */
623 static void
624 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
625 		    int nelems, enum dma_data_direction dir,
626 		    enum dma_sync_target target)
627 {
628 	struct scatterlist *sg;
629 	int i;
630 
631 	for_each_sg(sgl, sg, nelems, i)
632 		xen_swiotlb_sync_single(hwdev, sg->dma_address,
633 					sg_dma_len(sg), dir, target);
634 }
635 
636 void
637 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
638 			    int nelems, enum dma_data_direction dir)
639 {
640 	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
641 }
642 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
643 
644 void
645 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
646 			       int nelems, enum dma_data_direction dir)
647 {
648 	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
649 }
650 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
651 
652 int
653 xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
654 {
655 	return !dma_addr;
656 }
657 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
658 
659 /*
660  * Return whether the given device DMA address mask can be supported
661  * properly.  For example, if your device can only drive the low 24-bits
662  * during bus mastering, then you would pass 0x00ffffff as the mask to
663  * this function.
664  */
665 int
666 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
667 {
668 	return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
669 }
670 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
671 
672 int
673 xen_swiotlb_set_dma_mask(struct device *dev, u64 dma_mask)
674 {
675 	if (!dev->dma_mask || !xen_swiotlb_dma_supported(dev, dma_mask))
676 		return -EIO;
677 
678 	*dev->dma_mask = dma_mask;
679 
680 	return 0;
681 }
682 EXPORT_SYMBOL_GPL(xen_swiotlb_set_dma_mask);
683