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-direct.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 #define XEN_SWIOTLB_ERROR_CODE	(~(dma_addr_t)0x0)
57 
58 static char *xen_io_tlb_start, *xen_io_tlb_end;
59 static unsigned long xen_io_tlb_nslabs;
60 /*
61  * Quick lookup value of the bus address of the IOTLB.
62  */
63 
64 static u64 start_dma_addr;
65 
66 /*
67  * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
68  * can be 32bit when dma_addr_t is 64bit leading to a loss in
69  * information if the shift is done before casting to 64bit.
70  */
xen_phys_to_bus(phys_addr_t paddr)71 static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
72 {
73 	unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
74 	dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
75 
76 	dma |= paddr & ~XEN_PAGE_MASK;
77 
78 	return dma;
79 }
80 
xen_bus_to_phys(dma_addr_t baddr)81 static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
82 {
83 	unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
84 	dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
85 	phys_addr_t paddr = dma;
86 
87 	paddr |= baddr & ~XEN_PAGE_MASK;
88 
89 	return paddr;
90 }
91 
xen_virt_to_bus(void * address)92 static inline dma_addr_t xen_virt_to_bus(void *address)
93 {
94 	return xen_phys_to_bus(virt_to_phys(address));
95 }
96 
check_pages_physically_contiguous(unsigned long xen_pfn,unsigned int offset,size_t length)97 static int check_pages_physically_contiguous(unsigned long xen_pfn,
98 					     unsigned int offset,
99 					     size_t length)
100 {
101 	unsigned long next_bfn;
102 	int i;
103 	int nr_pages;
104 
105 	next_bfn = pfn_to_bfn(xen_pfn);
106 	nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT;
107 
108 	for (i = 1; i < nr_pages; i++) {
109 		if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
110 			return 0;
111 	}
112 	return 1;
113 }
114 
range_straddles_page_boundary(phys_addr_t p,size_t size)115 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
116 {
117 	unsigned long xen_pfn = XEN_PFN_DOWN(p);
118 	unsigned int offset = p & ~XEN_PAGE_MASK;
119 
120 	if (offset + size <= XEN_PAGE_SIZE)
121 		return 0;
122 	if (check_pages_physically_contiguous(xen_pfn, offset, size))
123 		return 0;
124 	return 1;
125 }
126 
is_xen_swiotlb_buffer(dma_addr_t dma_addr)127 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
128 {
129 	unsigned long bfn = XEN_PFN_DOWN(dma_addr);
130 	unsigned long xen_pfn = bfn_to_local_pfn(bfn);
131 	phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
132 
133 	/* If the address is outside our domain, it CAN
134 	 * have the same virtual address as another address
135 	 * in our domain. Therefore _only_ check address within our domain.
136 	 */
137 	if (pfn_valid(PFN_DOWN(paddr))) {
138 		return paddr >= virt_to_phys(xen_io_tlb_start) &&
139 		       paddr < virt_to_phys(xen_io_tlb_end);
140 	}
141 	return 0;
142 }
143 
144 static int max_dma_bits = 32;
145 
146 static int
xen_swiotlb_fixup(void * buf,size_t size,unsigned long nslabs)147 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
148 {
149 	int i, rc;
150 	int dma_bits;
151 	dma_addr_t dma_handle;
152 	phys_addr_t p = virt_to_phys(buf);
153 
154 	dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
155 
156 	i = 0;
157 	do {
158 		int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
159 
160 		do {
161 			rc = xen_create_contiguous_region(
162 				p + (i << IO_TLB_SHIFT),
163 				get_order(slabs << IO_TLB_SHIFT),
164 				dma_bits, &dma_handle);
165 		} while (rc && dma_bits++ < max_dma_bits);
166 		if (rc)
167 			return rc;
168 
169 		i += slabs;
170 	} while (i < nslabs);
171 	return 0;
172 }
xen_set_nslabs(unsigned long nr_tbl)173 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
174 {
175 	if (!nr_tbl) {
176 		xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
177 		xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
178 	} else
179 		xen_io_tlb_nslabs = nr_tbl;
180 
181 	return xen_io_tlb_nslabs << IO_TLB_SHIFT;
182 }
183 
184 enum xen_swiotlb_err {
185 	XEN_SWIOTLB_UNKNOWN = 0,
186 	XEN_SWIOTLB_ENOMEM,
187 	XEN_SWIOTLB_EFIXUP
188 };
189 
xen_swiotlb_error(enum xen_swiotlb_err err)190 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
191 {
192 	switch (err) {
193 	case XEN_SWIOTLB_ENOMEM:
194 		return "Cannot allocate Xen-SWIOTLB buffer\n";
195 	case XEN_SWIOTLB_EFIXUP:
196 		return "Failed to get contiguous memory for DMA from Xen!\n"\
197 		    "You either: don't have the permissions, do not have"\
198 		    " enough free memory under 4GB, or the hypervisor memory"\
199 		    " is too fragmented!";
200 	default:
201 		break;
202 	}
203 	return "";
204 }
xen_swiotlb_init(int verbose,bool early)205 int __ref xen_swiotlb_init(int verbose, bool early)
206 {
207 	unsigned long bytes, order;
208 	int rc = -ENOMEM;
209 	enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
210 	unsigned int repeat = 3;
211 
212 	xen_io_tlb_nslabs = swiotlb_nr_tbl();
213 retry:
214 	bytes = xen_set_nslabs(xen_io_tlb_nslabs);
215 	order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
216 	/*
217 	 * Get IO TLB memory from any location.
218 	 */
219 	if (early)
220 		xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
221 	else {
222 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
223 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
224 		while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
225 			xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
226 			if (xen_io_tlb_start)
227 				break;
228 			order--;
229 		}
230 		if (order != get_order(bytes)) {
231 			pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
232 				(PAGE_SIZE << order) >> 20);
233 			xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
234 			bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
235 		}
236 	}
237 	if (!xen_io_tlb_start) {
238 		m_ret = XEN_SWIOTLB_ENOMEM;
239 		goto error;
240 	}
241 	xen_io_tlb_end = xen_io_tlb_start + bytes;
242 	/*
243 	 * And replace that memory with pages under 4GB.
244 	 */
245 	rc = xen_swiotlb_fixup(xen_io_tlb_start,
246 			       bytes,
247 			       xen_io_tlb_nslabs);
248 	if (rc) {
249 		if (early)
250 			free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
251 		else {
252 			free_pages((unsigned long)xen_io_tlb_start, order);
253 			xen_io_tlb_start = NULL;
254 		}
255 		m_ret = XEN_SWIOTLB_EFIXUP;
256 		goto error;
257 	}
258 	start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
259 	if (early) {
260 		if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
261 			 verbose))
262 			panic("Cannot allocate SWIOTLB buffer");
263 		rc = 0;
264 	} else
265 		rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
266 
267 	if (!rc)
268 		swiotlb_set_max_segment(PAGE_SIZE);
269 
270 	return rc;
271 error:
272 	if (repeat--) {
273 		xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
274 					(xen_io_tlb_nslabs >> 1));
275 		pr_info("Lowering to %luMB\n",
276 			(xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
277 		goto retry;
278 	}
279 	pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
280 	if (early)
281 		panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
282 	else
283 		free_pages((unsigned long)xen_io_tlb_start, order);
284 	return rc;
285 }
286 
287 static void *
xen_swiotlb_alloc_coherent(struct device * hwdev,size_t size,dma_addr_t * dma_handle,gfp_t flags,unsigned long attrs)288 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
289 			   dma_addr_t *dma_handle, gfp_t flags,
290 			   unsigned long attrs)
291 {
292 	void *ret;
293 	int order = get_order(size);
294 	u64 dma_mask = DMA_BIT_MASK(32);
295 	phys_addr_t phys;
296 	dma_addr_t dev_addr;
297 
298 	/*
299 	* Ignore region specifiers - the kernel's ideas of
300 	* pseudo-phys memory layout has nothing to do with the
301 	* machine physical layout.  We can't allocate highmem
302 	* because we can't return a pointer to it.
303 	*/
304 	flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
305 
306 	/* Convert the size to actually allocated. */
307 	size = 1UL << (order + XEN_PAGE_SHIFT);
308 
309 	/* On ARM this function returns an ioremap'ped virtual address for
310 	 * which virt_to_phys doesn't return the corresponding physical
311 	 * address. In fact on ARM virt_to_phys only works for kernel direct
312 	 * mapped RAM memory. Also see comment below.
313 	 */
314 	ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
315 
316 	if (!ret)
317 		return ret;
318 
319 	if (hwdev && hwdev->coherent_dma_mask)
320 		dma_mask = hwdev->coherent_dma_mask;
321 
322 	/* At this point dma_handle is the physical address, next we are
323 	 * going to set it to the machine address.
324 	 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
325 	 * to *dma_handle. */
326 	phys = *dma_handle;
327 	dev_addr = xen_phys_to_bus(phys);
328 	if (((dev_addr + size - 1 <= dma_mask)) &&
329 	    !range_straddles_page_boundary(phys, size))
330 		*dma_handle = dev_addr;
331 	else {
332 		if (xen_create_contiguous_region(phys, order,
333 						 fls64(dma_mask), dma_handle) != 0) {
334 			xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
335 			return NULL;
336 		}
337 	}
338 	memset(ret, 0, size);
339 	return ret;
340 }
341 
342 static void
xen_swiotlb_free_coherent(struct device * hwdev,size_t size,void * vaddr,dma_addr_t dev_addr,unsigned long attrs)343 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
344 			  dma_addr_t dev_addr, unsigned long attrs)
345 {
346 	int order = get_order(size);
347 	phys_addr_t phys;
348 	u64 dma_mask = DMA_BIT_MASK(32);
349 
350 	if (hwdev && hwdev->coherent_dma_mask)
351 		dma_mask = hwdev->coherent_dma_mask;
352 
353 	/* do not use virt_to_phys because on ARM it doesn't return you the
354 	 * physical address */
355 	phys = xen_bus_to_phys(dev_addr);
356 
357 	/* Convert the size to actually allocated. */
358 	size = 1UL << (order + XEN_PAGE_SHIFT);
359 
360 	if (!WARN_ON((dev_addr + size - 1 > dma_mask) ||
361 		     range_straddles_page_boundary(phys, size)))
362 		xen_destroy_contiguous_region(phys, order);
363 
364 	xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
365 }
366 
367 /*
368  * Map a single buffer of the indicated size for DMA in streaming mode.  The
369  * physical address to use is returned.
370  *
371  * Once the device is given the dma address, the device owns this memory until
372  * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
373  */
xen_swiotlb_map_page(struct device * dev,struct page * page,unsigned long offset,size_t size,enum dma_data_direction dir,unsigned long attrs)374 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
375 				unsigned long offset, size_t size,
376 				enum dma_data_direction dir,
377 				unsigned long attrs)
378 {
379 	phys_addr_t map, phys = page_to_phys(page) + offset;
380 	dma_addr_t dev_addr = xen_phys_to_bus(phys);
381 
382 	BUG_ON(dir == DMA_NONE);
383 	/*
384 	 * If the address happens to be in the device's DMA window,
385 	 * we can safely return the device addr and not worry about bounce
386 	 * buffering it.
387 	 */
388 	if (dma_capable(dev, dev_addr, size) &&
389 	    !range_straddles_page_boundary(phys, size) &&
390 		!xen_arch_need_swiotlb(dev, phys, dev_addr) &&
391 		(swiotlb_force != SWIOTLB_FORCE)) {
392 		/* we are not interested in the dma_addr returned by
393 		 * xen_dma_map_page, only in the potential cache flushes executed
394 		 * by the function. */
395 		xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
396 		return dev_addr;
397 	}
398 
399 	/*
400 	 * Oh well, have to allocate and map a bounce buffer.
401 	 */
402 	trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
403 
404 	map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
405 				     attrs);
406 	if (map == SWIOTLB_MAP_ERROR)
407 		return XEN_SWIOTLB_ERROR_CODE;
408 
409 	dev_addr = xen_phys_to_bus(map);
410 	xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
411 					dev_addr, map & ~PAGE_MASK, size, dir, attrs);
412 
413 	/*
414 	 * Ensure that the address returned is DMA'ble
415 	 */
416 	if (dma_capable(dev, dev_addr, size))
417 		return dev_addr;
418 
419 	attrs |= DMA_ATTR_SKIP_CPU_SYNC;
420 	swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
421 
422 	return XEN_SWIOTLB_ERROR_CODE;
423 }
424 
425 /*
426  * Unmap a single streaming mode DMA translation.  The dma_addr and size must
427  * match what was provided for in a previous xen_swiotlb_map_page call.  All
428  * other usages are undefined.
429  *
430  * After this call, reads by the cpu to the buffer are guaranteed to see
431  * whatever the device wrote there.
432  */
xen_unmap_single(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir,unsigned long attrs)433 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
434 			     size_t size, enum dma_data_direction dir,
435 			     unsigned long attrs)
436 {
437 	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
438 
439 	BUG_ON(dir == DMA_NONE);
440 
441 	xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
442 
443 	/* NOTE: We use dev_addr here, not paddr! */
444 	if (is_xen_swiotlb_buffer(dev_addr)) {
445 		swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
446 		return;
447 	}
448 
449 	if (dir != DMA_FROM_DEVICE)
450 		return;
451 
452 	/*
453 	 * phys_to_virt doesn't work with hihgmem page but we could
454 	 * call dma_mark_clean() with hihgmem page here. However, we
455 	 * are fine since dma_mark_clean() is null on POWERPC. We can
456 	 * make dma_mark_clean() take a physical address if necessary.
457 	 */
458 	dma_mark_clean(phys_to_virt(paddr), size);
459 }
460 
xen_swiotlb_unmap_page(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir,unsigned long attrs)461 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
462 			    size_t size, enum dma_data_direction dir,
463 			    unsigned long attrs)
464 {
465 	xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
466 }
467 
468 /*
469  * Make physical memory consistent for a single streaming mode DMA translation
470  * after a transfer.
471  *
472  * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
473  * using the cpu, yet do not wish to teardown the dma mapping, you must
474  * call this function before doing so.  At the next point you give the dma
475  * address back to the card, you must first perform a
476  * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
477  */
478 static void
xen_swiotlb_sync_single(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir,enum dma_sync_target target)479 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
480 			size_t size, enum dma_data_direction dir,
481 			enum dma_sync_target target)
482 {
483 	phys_addr_t paddr = xen_bus_to_phys(dev_addr);
484 
485 	BUG_ON(dir == DMA_NONE);
486 
487 	if (target == SYNC_FOR_CPU)
488 		xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
489 
490 	/* NOTE: We use dev_addr here, not paddr! */
491 	if (is_xen_swiotlb_buffer(dev_addr))
492 		swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
493 
494 	if (target == SYNC_FOR_DEVICE)
495 		xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
496 
497 	if (dir != DMA_FROM_DEVICE)
498 		return;
499 
500 	dma_mark_clean(phys_to_virt(paddr), size);
501 }
502 
503 void
xen_swiotlb_sync_single_for_cpu(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir)504 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
505 				size_t size, enum dma_data_direction dir)
506 {
507 	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
508 }
509 
510 void
xen_swiotlb_sync_single_for_device(struct device * hwdev,dma_addr_t dev_addr,size_t size,enum dma_data_direction dir)511 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
512 				   size_t size, enum dma_data_direction dir)
513 {
514 	xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
515 }
516 
517 /*
518  * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
519  * concerning calls here are the same as for swiotlb_unmap_page() above.
520  */
521 static void
xen_swiotlb_unmap_sg_attrs(struct device * hwdev,struct scatterlist * sgl,int nelems,enum dma_data_direction dir,unsigned long attrs)522 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
523 			   int nelems, enum dma_data_direction dir,
524 			   unsigned long attrs)
525 {
526 	struct scatterlist *sg;
527 	int i;
528 
529 	BUG_ON(dir == DMA_NONE);
530 
531 	for_each_sg(sgl, sg, nelems, i)
532 		xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
533 
534 }
535 
536 /*
537  * Map a set of buffers described by scatterlist in streaming mode for DMA.
538  * This is the scatter-gather version of the above xen_swiotlb_map_page
539  * interface.  Here the scatter gather list elements are each tagged with the
540  * appropriate dma address and length.  They are obtained via
541  * sg_dma_{address,length}(SG).
542  *
543  * NOTE: An implementation may be able to use a smaller number of
544  *       DMA address/length pairs than there are SG table elements.
545  *       (for example via virtual mapping capabilities)
546  *       The routine returns the number of addr/length pairs actually
547  *       used, at most nents.
548  *
549  * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
550  * same here.
551  */
552 static int
xen_swiotlb_map_sg_attrs(struct device * hwdev,struct scatterlist * sgl,int nelems,enum dma_data_direction dir,unsigned long attrs)553 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
554 			 int nelems, enum dma_data_direction dir,
555 			 unsigned long attrs)
556 {
557 	struct scatterlist *sg;
558 	int i;
559 
560 	BUG_ON(dir == DMA_NONE);
561 
562 	for_each_sg(sgl, sg, nelems, i) {
563 		phys_addr_t paddr = sg_phys(sg);
564 		dma_addr_t dev_addr = xen_phys_to_bus(paddr);
565 
566 		if (swiotlb_force == SWIOTLB_FORCE ||
567 		    xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
568 		    !dma_capable(hwdev, dev_addr, sg->length) ||
569 		    range_straddles_page_boundary(paddr, sg->length)) {
570 			phys_addr_t map = swiotlb_tbl_map_single(hwdev,
571 								 start_dma_addr,
572 								 sg_phys(sg),
573 								 sg->length,
574 								 dir, attrs);
575 			if (map == SWIOTLB_MAP_ERROR) {
576 				dev_warn(hwdev, "swiotlb buffer is full\n");
577 				/* Don't panic here, we expect map_sg users
578 				   to do proper error handling. */
579 				attrs |= DMA_ATTR_SKIP_CPU_SYNC;
580 				xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
581 							   attrs);
582 				sg_dma_len(sgl) = 0;
583 				return 0;
584 			}
585 			dev_addr = xen_phys_to_bus(map);
586 			xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
587 						dev_addr,
588 						map & ~PAGE_MASK,
589 						sg->length,
590 						dir,
591 						attrs);
592 			sg->dma_address = dev_addr;
593 		} else {
594 			/* we are not interested in the dma_addr returned by
595 			 * xen_dma_map_page, only in the potential cache flushes executed
596 			 * by the function. */
597 			xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
598 						dev_addr,
599 						paddr & ~PAGE_MASK,
600 						sg->length,
601 						dir,
602 						attrs);
603 			sg->dma_address = dev_addr;
604 		}
605 		sg_dma_len(sg) = sg->length;
606 	}
607 	return nelems;
608 }
609 
610 /*
611  * Make physical memory consistent for a set of streaming mode DMA translations
612  * after a transfer.
613  *
614  * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
615  * and usage.
616  */
617 static void
xen_swiotlb_sync_sg(struct device * hwdev,struct scatterlist * sgl,int nelems,enum dma_data_direction dir,enum dma_sync_target target)618 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
619 		    int nelems, enum dma_data_direction dir,
620 		    enum dma_sync_target target)
621 {
622 	struct scatterlist *sg;
623 	int i;
624 
625 	for_each_sg(sgl, sg, nelems, i)
626 		xen_swiotlb_sync_single(hwdev, sg->dma_address,
627 					sg_dma_len(sg), dir, target);
628 }
629 
630 static void
xen_swiotlb_sync_sg_for_cpu(struct device * hwdev,struct scatterlist * sg,int nelems,enum dma_data_direction dir)631 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
632 			    int nelems, enum dma_data_direction dir)
633 {
634 	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
635 }
636 
637 static void
xen_swiotlb_sync_sg_for_device(struct device * hwdev,struct scatterlist * sg,int nelems,enum dma_data_direction dir)638 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
639 			       int nelems, enum dma_data_direction dir)
640 {
641 	xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
642 }
643 
644 /*
645  * Return whether the given device DMA address mask can be supported
646  * properly.  For example, if your device can only drive the low 24-bits
647  * during bus mastering, then you would pass 0x00ffffff as the mask to
648  * this function.
649  */
650 static int
xen_swiotlb_dma_supported(struct device * hwdev,u64 mask)651 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
652 {
653 	return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
654 }
655 
656 /*
657  * Create userspace mapping for the DMA-coherent memory.
658  * This function should be called with the pages from the current domain only,
659  * passing pages mapped from other domains would lead to memory corruption.
660  */
661 static int
xen_swiotlb_dma_mmap(struct device * dev,struct vm_area_struct * vma,void * cpu_addr,dma_addr_t dma_addr,size_t size,unsigned long attrs)662 xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
663 		     void *cpu_addr, dma_addr_t dma_addr, size_t size,
664 		     unsigned long attrs)
665 {
666 #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
667 	if (xen_get_dma_ops(dev)->mmap)
668 		return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
669 						    dma_addr, size, attrs);
670 #endif
671 	return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size);
672 }
673 
674 /*
675  * This function should be called with the pages from the current domain only,
676  * passing pages mapped from other domains would lead to memory corruption.
677  */
678 static int
xen_swiotlb_get_sgtable(struct device * dev,struct sg_table * sgt,void * cpu_addr,dma_addr_t handle,size_t size,unsigned long attrs)679 xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
680 			void *cpu_addr, dma_addr_t handle, size_t size,
681 			unsigned long attrs)
682 {
683 #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
684 	if (xen_get_dma_ops(dev)->get_sgtable) {
685 #if 0
686 	/*
687 	 * This check verifies that the page belongs to the current domain and
688 	 * is not one mapped from another domain.
689 	 * This check is for debug only, and should not go to production build
690 	 */
691 		unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
692 		BUG_ON (!page_is_ram(bfn));
693 #endif
694 		return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
695 							   handle, size, attrs);
696 	}
697 #endif
698 	return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size);
699 }
700 
xen_swiotlb_mapping_error(struct device * dev,dma_addr_t dma_addr)701 static int xen_swiotlb_mapping_error(struct device *dev, dma_addr_t dma_addr)
702 {
703 	return dma_addr == XEN_SWIOTLB_ERROR_CODE;
704 }
705 
706 const struct dma_map_ops xen_swiotlb_dma_ops = {
707 	.alloc = xen_swiotlb_alloc_coherent,
708 	.free = xen_swiotlb_free_coherent,
709 	.sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
710 	.sync_single_for_device = xen_swiotlb_sync_single_for_device,
711 	.sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
712 	.sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
713 	.map_sg = xen_swiotlb_map_sg_attrs,
714 	.unmap_sg = xen_swiotlb_unmap_sg_attrs,
715 	.map_page = xen_swiotlb_map_page,
716 	.unmap_page = xen_swiotlb_unmap_page,
717 	.dma_supported = xen_swiotlb_dma_supported,
718 	.mmap = xen_swiotlb_dma_mmap,
719 	.get_sgtable = xen_swiotlb_get_sgtable,
720 	.mapping_error	= xen_swiotlb_mapping_error,
721 };
722