Commit | Line | Data |
---|---|---|
b097186f KRW |
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 | ||
283c0972 JP |
36 | #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt |
37 | ||
b097186f KRW |
38 | #include <linux/bootmem.h> |
39 | #include <linux/dma-mapping.h> | |
63c9744b | 40 | #include <linux/export.h> |
b097186f KRW |
41 | #include <xen/swiotlb-xen.h> |
42 | #include <xen/page.h> | |
43 | #include <xen/xen-ops.h> | |
f4b2f07b | 44 | #include <xen/hvc-console.h> |
83862ccf | 45 | #include <asm/dma-mapping.h> |
1b65c4e5 | 46 | #include <asm/xen/page-coherent.h> |
b097186f KRW |
47 | /* |
48 | * Used to do a quick range check in swiotlb_tbl_unmap_single and | |
49 | * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this | |
50 | * API. | |
51 | */ | |
52 | ||
83862ccf SS |
53 | #ifndef CONFIG_X86 |
54 | static unsigned long dma_alloc_coherent_mask(struct device *dev, | |
55 | gfp_t gfp) | |
56 | { | |
57 | unsigned long dma_mask = 0; | |
58 | ||
59 | dma_mask = dev->coherent_dma_mask; | |
60 | if (!dma_mask) | |
61 | dma_mask = (gfp & GFP_DMA) ? DMA_BIT_MASK(24) : DMA_BIT_MASK(32); | |
62 | ||
63 | return dma_mask; | |
64 | } | |
65 | #endif | |
66 | ||
b097186f KRW |
67 | static char *xen_io_tlb_start, *xen_io_tlb_end; |
68 | static unsigned long xen_io_tlb_nslabs; | |
69 | /* | |
70 | * Quick lookup value of the bus address of the IOTLB. | |
71 | */ | |
72 | ||
b8b0f559 | 73 | static u64 start_dma_addr; |
b097186f KRW |
74 | |
75 | static dma_addr_t xen_phys_to_bus(phys_addr_t paddr) | |
76 | { | |
6eab04a8 | 77 | return phys_to_machine(XPADDR(paddr)).maddr; |
b097186f KRW |
78 | } |
79 | ||
80 | static phys_addr_t xen_bus_to_phys(dma_addr_t baddr) | |
81 | { | |
82 | return machine_to_phys(XMADDR(baddr)).paddr; | |
83 | } | |
84 | ||
85 | static dma_addr_t xen_virt_to_bus(void *address) | |
86 | { | |
87 | return xen_phys_to_bus(virt_to_phys(address)); | |
88 | } | |
89 | ||
90 | static int check_pages_physically_contiguous(unsigned long pfn, | |
91 | unsigned int offset, | |
92 | size_t length) | |
93 | { | |
94 | unsigned long next_mfn; | |
95 | int i; | |
96 | int nr_pages; | |
97 | ||
98 | next_mfn = pfn_to_mfn(pfn); | |
99 | nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT; | |
100 | ||
101 | for (i = 1; i < nr_pages; i++) { | |
102 | if (pfn_to_mfn(++pfn) != ++next_mfn) | |
103 | return 0; | |
104 | } | |
105 | return 1; | |
106 | } | |
107 | ||
108 | static int range_straddles_page_boundary(phys_addr_t p, size_t size) | |
109 | { | |
110 | unsigned long pfn = PFN_DOWN(p); | |
111 | unsigned int offset = p & ~PAGE_MASK; | |
112 | ||
113 | if (offset + size <= PAGE_SIZE) | |
114 | return 0; | |
115 | if (check_pages_physically_contiguous(pfn, offset, size)) | |
116 | return 0; | |
117 | return 1; | |
118 | } | |
119 | ||
120 | static int is_xen_swiotlb_buffer(dma_addr_t dma_addr) | |
121 | { | |
122 | unsigned long mfn = PFN_DOWN(dma_addr); | |
123 | unsigned long pfn = mfn_to_local_pfn(mfn); | |
124 | phys_addr_t paddr; | |
125 | ||
126 | /* If the address is outside our domain, it CAN | |
127 | * have the same virtual address as another address | |
128 | * in our domain. Therefore _only_ check address within our domain. | |
129 | */ | |
130 | if (pfn_valid(pfn)) { | |
131 | paddr = PFN_PHYS(pfn); | |
132 | return paddr >= virt_to_phys(xen_io_tlb_start) && | |
133 | paddr < virt_to_phys(xen_io_tlb_end); | |
134 | } | |
135 | return 0; | |
136 | } | |
137 | ||
138 | static int max_dma_bits = 32; | |
139 | ||
140 | static int | |
141 | xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs) | |
142 | { | |
143 | int i, rc; | |
144 | int dma_bits; | |
69908907 | 145 | dma_addr_t dma_handle; |
1b65c4e5 | 146 | phys_addr_t p = virt_to_phys(buf); |
b097186f KRW |
147 | |
148 | dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT; | |
149 | ||
150 | i = 0; | |
151 | do { | |
152 | int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE); | |
153 | ||
154 | do { | |
155 | rc = xen_create_contiguous_region( | |
1b65c4e5 | 156 | p + (i << IO_TLB_SHIFT), |
b097186f | 157 | get_order(slabs << IO_TLB_SHIFT), |
69908907 | 158 | dma_bits, &dma_handle); |
b097186f KRW |
159 | } while (rc && dma_bits++ < max_dma_bits); |
160 | if (rc) | |
161 | return rc; | |
162 | ||
163 | i += slabs; | |
164 | } while (i < nslabs); | |
165 | return 0; | |
166 | } | |
1cef36a5 KRW |
167 | static unsigned long xen_set_nslabs(unsigned long nr_tbl) |
168 | { | |
169 | if (!nr_tbl) { | |
170 | xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT); | |
171 | xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE); | |
172 | } else | |
173 | xen_io_tlb_nslabs = nr_tbl; | |
b097186f | 174 | |
1cef36a5 KRW |
175 | return xen_io_tlb_nslabs << IO_TLB_SHIFT; |
176 | } | |
b097186f | 177 | |
5bab7864 KRW |
178 | enum xen_swiotlb_err { |
179 | XEN_SWIOTLB_UNKNOWN = 0, | |
180 | XEN_SWIOTLB_ENOMEM, | |
181 | XEN_SWIOTLB_EFIXUP | |
182 | }; | |
183 | ||
184 | static const char *xen_swiotlb_error(enum xen_swiotlb_err err) | |
185 | { | |
186 | switch (err) { | |
187 | case XEN_SWIOTLB_ENOMEM: | |
188 | return "Cannot allocate Xen-SWIOTLB buffer\n"; | |
189 | case XEN_SWIOTLB_EFIXUP: | |
190 | return "Failed to get contiguous memory for DMA from Xen!\n"\ | |
191 | "You either: don't have the permissions, do not have"\ | |
192 | " enough free memory under 4GB, or the hypervisor memory"\ | |
193 | " is too fragmented!"; | |
194 | default: | |
195 | break; | |
196 | } | |
197 | return ""; | |
198 | } | |
b8277600 | 199 | int __ref xen_swiotlb_init(int verbose, bool early) |
b097186f | 200 | { |
b8277600 | 201 | unsigned long bytes, order; |
f4b2f07b | 202 | int rc = -ENOMEM; |
5bab7864 | 203 | enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN; |
f4b2f07b | 204 | unsigned int repeat = 3; |
5f98ecdb | 205 | |
1cef36a5 | 206 | xen_io_tlb_nslabs = swiotlb_nr_tbl(); |
f4b2f07b | 207 | retry: |
1cef36a5 | 208 | bytes = xen_set_nslabs(xen_io_tlb_nslabs); |
b8277600 | 209 | order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT); |
b097186f KRW |
210 | /* |
211 | * Get IO TLB memory from any location. | |
212 | */ | |
b8277600 KRW |
213 | if (early) |
214 | xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes)); | |
215 | else { | |
216 | #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT)) | |
217 | #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT) | |
218 | while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) { | |
219 | xen_io_tlb_start = (void *)__get_free_pages(__GFP_NOWARN, order); | |
220 | if (xen_io_tlb_start) | |
221 | break; | |
222 | order--; | |
223 | } | |
224 | if (order != get_order(bytes)) { | |
283c0972 JP |
225 | pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n", |
226 | (PAGE_SIZE << order) >> 20); | |
b8277600 KRW |
227 | xen_io_tlb_nslabs = SLABS_PER_PAGE << order; |
228 | bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT; | |
229 | } | |
230 | } | |
f4b2f07b | 231 | if (!xen_io_tlb_start) { |
5bab7864 | 232 | m_ret = XEN_SWIOTLB_ENOMEM; |
f4b2f07b KRW |
233 | goto error; |
234 | } | |
b097186f KRW |
235 | xen_io_tlb_end = xen_io_tlb_start + bytes; |
236 | /* | |
237 | * And replace that memory with pages under 4GB. | |
238 | */ | |
239 | rc = xen_swiotlb_fixup(xen_io_tlb_start, | |
240 | bytes, | |
241 | xen_io_tlb_nslabs); | |
f4b2f07b | 242 | if (rc) { |
b8277600 KRW |
243 | if (early) |
244 | free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes)); | |
245 | else { | |
246 | free_pages((unsigned long)xen_io_tlb_start, order); | |
247 | xen_io_tlb_start = NULL; | |
248 | } | |
5bab7864 | 249 | m_ret = XEN_SWIOTLB_EFIXUP; |
b097186f | 250 | goto error; |
f4b2f07b | 251 | } |
b097186f | 252 | start_dma_addr = xen_virt_to_bus(xen_io_tlb_start); |
c468bdee | 253 | if (early) { |
ac2cbab2 YL |
254 | if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, |
255 | verbose)) | |
256 | panic("Cannot allocate SWIOTLB buffer"); | |
c468bdee KRW |
257 | rc = 0; |
258 | } else | |
b8277600 KRW |
259 | rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs); |
260 | return rc; | |
b097186f | 261 | error: |
f4b2f07b KRW |
262 | if (repeat--) { |
263 | xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */ | |
264 | (xen_io_tlb_nslabs >> 1)); | |
283c0972 JP |
265 | pr_info("Lowering to %luMB\n", |
266 | (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20); | |
f4b2f07b KRW |
267 | goto retry; |
268 | } | |
283c0972 | 269 | pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc); |
b8277600 KRW |
270 | if (early) |
271 | panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc); | |
272 | else | |
273 | free_pages((unsigned long)xen_io_tlb_start, order); | |
274 | return rc; | |
b097186f | 275 | } |
b097186f KRW |
276 | void * |
277 | xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size, | |
baa676fc AP |
278 | dma_addr_t *dma_handle, gfp_t flags, |
279 | struct dma_attrs *attrs) | |
b097186f KRW |
280 | { |
281 | void *ret; | |
282 | int order = get_order(size); | |
283 | u64 dma_mask = DMA_BIT_MASK(32); | |
6810df88 KRW |
284 | phys_addr_t phys; |
285 | dma_addr_t dev_addr; | |
b097186f KRW |
286 | |
287 | /* | |
288 | * Ignore region specifiers - the kernel's ideas of | |
289 | * pseudo-phys memory layout has nothing to do with the | |
290 | * machine physical layout. We can't allocate highmem | |
291 | * because we can't return a pointer to it. | |
292 | */ | |
293 | flags &= ~(__GFP_DMA | __GFP_HIGHMEM); | |
294 | ||
295 | if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret)) | |
296 | return ret; | |
297 | ||
1b65c4e5 SS |
298 | /* On ARM this function returns an ioremap'ped virtual address for |
299 | * which virt_to_phys doesn't return the corresponding physical | |
300 | * address. In fact on ARM virt_to_phys only works for kernel direct | |
301 | * mapped RAM memory. Also see comment below. | |
302 | */ | |
303 | ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs); | |
b097186f | 304 | |
6810df88 KRW |
305 | if (!ret) |
306 | return ret; | |
307 | ||
b097186f | 308 | if (hwdev && hwdev->coherent_dma_mask) |
b5031ed1 | 309 | dma_mask = dma_alloc_coherent_mask(hwdev, flags); |
b097186f | 310 | |
1b65c4e5 SS |
311 | /* At this point dma_handle is the physical address, next we are |
312 | * going to set it to the machine address. | |
313 | * Do not use virt_to_phys(ret) because on ARM it doesn't correspond | |
314 | * to *dma_handle. */ | |
315 | phys = *dma_handle; | |
6810df88 KRW |
316 | dev_addr = xen_phys_to_bus(phys); |
317 | if (((dev_addr + size - 1 <= dma_mask)) && | |
318 | !range_straddles_page_boundary(phys, size)) | |
319 | *dma_handle = dev_addr; | |
320 | else { | |
1b65c4e5 | 321 | if (xen_create_contiguous_region(phys, order, |
69908907 | 322 | fls64(dma_mask), dma_handle) != 0) { |
1b65c4e5 | 323 | xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs); |
b097186f KRW |
324 | return NULL; |
325 | } | |
b097186f | 326 | } |
6810df88 | 327 | memset(ret, 0, size); |
b097186f KRW |
328 | return ret; |
329 | } | |
330 | EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent); | |
331 | ||
332 | void | |
333 | xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr, | |
baa676fc | 334 | dma_addr_t dev_addr, struct dma_attrs *attrs) |
b097186f KRW |
335 | { |
336 | int order = get_order(size); | |
6810df88 KRW |
337 | phys_addr_t phys; |
338 | u64 dma_mask = DMA_BIT_MASK(32); | |
b097186f KRW |
339 | |
340 | if (dma_release_from_coherent(hwdev, order, vaddr)) | |
341 | return; | |
342 | ||
6810df88 KRW |
343 | if (hwdev && hwdev->coherent_dma_mask) |
344 | dma_mask = hwdev->coherent_dma_mask; | |
345 | ||
1b65c4e5 SS |
346 | /* do not use virt_to_phys because on ARM it doesn't return you the |
347 | * physical address */ | |
348 | phys = xen_bus_to_phys(dev_addr); | |
6810df88 KRW |
349 | |
350 | if (((dev_addr + size - 1 > dma_mask)) || | |
351 | range_straddles_page_boundary(phys, size)) | |
1b65c4e5 | 352 | xen_destroy_contiguous_region(phys, order); |
6810df88 | 353 | |
1b65c4e5 | 354 | xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs); |
b097186f KRW |
355 | } |
356 | EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent); | |
357 | ||
358 | ||
359 | /* | |
360 | * Map a single buffer of the indicated size for DMA in streaming mode. The | |
361 | * physical address to use is returned. | |
362 | * | |
363 | * Once the device is given the dma address, the device owns this memory until | |
364 | * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed. | |
365 | */ | |
366 | dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page, | |
367 | unsigned long offset, size_t size, | |
368 | enum dma_data_direction dir, | |
369 | struct dma_attrs *attrs) | |
370 | { | |
e05ed4d1 | 371 | phys_addr_t map, phys = page_to_phys(page) + offset; |
b097186f | 372 | dma_addr_t dev_addr = xen_phys_to_bus(phys); |
b097186f KRW |
373 | |
374 | BUG_ON(dir == DMA_NONE); | |
375 | /* | |
376 | * If the address happens to be in the device's DMA window, | |
377 | * we can safely return the device addr and not worry about bounce | |
378 | * buffering it. | |
379 | */ | |
380 | if (dma_capable(dev, dev_addr, size) && | |
6cf05463 SS |
381 | !range_straddles_page_boundary(phys, size) && !swiotlb_force) { |
382 | /* we are not interested in the dma_addr returned by | |
383 | * xen_dma_map_page, only in the potential cache flushes executed | |
384 | * by the function. */ | |
385 | xen_dma_map_page(dev, page, offset, size, dir, attrs); | |
b097186f | 386 | return dev_addr; |
6cf05463 | 387 | } |
b097186f KRW |
388 | |
389 | /* | |
390 | * Oh well, have to allocate and map a bounce buffer. | |
391 | */ | |
392 | map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir); | |
e05ed4d1 | 393 | if (map == SWIOTLB_MAP_ERROR) |
b097186f KRW |
394 | return DMA_ERROR_CODE; |
395 | ||
6cf05463 SS |
396 | xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT), |
397 | map & ~PAGE_MASK, size, dir, attrs); | |
e05ed4d1 | 398 | dev_addr = xen_phys_to_bus(map); |
b097186f KRW |
399 | |
400 | /* | |
401 | * Ensure that the address returned is DMA'ble | |
402 | */ | |
ab2a47bd | 403 | if (!dma_capable(dev, dev_addr, size)) { |
61ca08c3 | 404 | swiotlb_tbl_unmap_single(dev, map, size, dir); |
ab2a47bd KRW |
405 | dev_addr = 0; |
406 | } | |
b097186f KRW |
407 | return dev_addr; |
408 | } | |
409 | EXPORT_SYMBOL_GPL(xen_swiotlb_map_page); | |
410 | ||
411 | /* | |
412 | * Unmap a single streaming mode DMA translation. The dma_addr and size must | |
413 | * match what was provided for in a previous xen_swiotlb_map_page call. All | |
414 | * other usages are undefined. | |
415 | * | |
416 | * After this call, reads by the cpu to the buffer are guaranteed to see | |
417 | * whatever the device wrote there. | |
418 | */ | |
419 | static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr, | |
6cf05463 SS |
420 | size_t size, enum dma_data_direction dir, |
421 | struct dma_attrs *attrs) | |
b097186f KRW |
422 | { |
423 | phys_addr_t paddr = xen_bus_to_phys(dev_addr); | |
424 | ||
425 | BUG_ON(dir == DMA_NONE); | |
426 | ||
6cf05463 SS |
427 | xen_dma_unmap_page(hwdev, paddr, size, dir, attrs); |
428 | ||
b097186f KRW |
429 | /* NOTE: We use dev_addr here, not paddr! */ |
430 | if (is_xen_swiotlb_buffer(dev_addr)) { | |
61ca08c3 | 431 | swiotlb_tbl_unmap_single(hwdev, paddr, size, dir); |
b097186f KRW |
432 | return; |
433 | } | |
434 | ||
435 | if (dir != DMA_FROM_DEVICE) | |
436 | return; | |
437 | ||
438 | /* | |
439 | * phys_to_virt doesn't work with hihgmem page but we could | |
440 | * call dma_mark_clean() with hihgmem page here. However, we | |
441 | * are fine since dma_mark_clean() is null on POWERPC. We can | |
442 | * make dma_mark_clean() take a physical address if necessary. | |
443 | */ | |
444 | dma_mark_clean(phys_to_virt(paddr), size); | |
445 | } | |
446 | ||
447 | void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr, | |
448 | size_t size, enum dma_data_direction dir, | |
449 | struct dma_attrs *attrs) | |
450 | { | |
6cf05463 | 451 | xen_unmap_single(hwdev, dev_addr, size, dir, attrs); |
b097186f KRW |
452 | } |
453 | EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page); | |
454 | ||
455 | /* | |
456 | * Make physical memory consistent for a single streaming mode DMA translation | |
457 | * after a transfer. | |
458 | * | |
459 | * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer | |
460 | * using the cpu, yet do not wish to teardown the dma mapping, you must | |
461 | * call this function before doing so. At the next point you give the dma | |
462 | * address back to the card, you must first perform a | |
463 | * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer | |
464 | */ | |
465 | static void | |
466 | xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr, | |
467 | size_t size, enum dma_data_direction dir, | |
468 | enum dma_sync_target target) | |
469 | { | |
470 | phys_addr_t paddr = xen_bus_to_phys(dev_addr); | |
471 | ||
472 | BUG_ON(dir == DMA_NONE); | |
473 | ||
6cf05463 SS |
474 | if (target == SYNC_FOR_CPU) |
475 | xen_dma_sync_single_for_cpu(hwdev, paddr, size, dir); | |
476 | ||
b097186f | 477 | /* NOTE: We use dev_addr here, not paddr! */ |
6cf05463 | 478 | if (is_xen_swiotlb_buffer(dev_addr)) |
fbfda893 | 479 | swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target); |
6cf05463 SS |
480 | |
481 | if (target == SYNC_FOR_DEVICE) | |
482 | xen_dma_sync_single_for_cpu(hwdev, paddr, size, dir); | |
b097186f KRW |
483 | |
484 | if (dir != DMA_FROM_DEVICE) | |
485 | return; | |
486 | ||
487 | dma_mark_clean(phys_to_virt(paddr), size); | |
488 | } | |
489 | ||
490 | void | |
491 | xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr, | |
492 | size_t size, enum dma_data_direction dir) | |
493 | { | |
494 | xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU); | |
495 | } | |
496 | EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu); | |
497 | ||
498 | void | |
499 | xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr, | |
500 | size_t size, enum dma_data_direction dir) | |
501 | { | |
502 | xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE); | |
503 | } | |
504 | EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device); | |
505 | ||
506 | /* | |
507 | * Map a set of buffers described by scatterlist in streaming mode for DMA. | |
508 | * This is the scatter-gather version of the above xen_swiotlb_map_page | |
509 | * interface. Here the scatter gather list elements are each tagged with the | |
510 | * appropriate dma address and length. They are obtained via | |
511 | * sg_dma_{address,length}(SG). | |
512 | * | |
513 | * NOTE: An implementation may be able to use a smaller number of | |
514 | * DMA address/length pairs than there are SG table elements. | |
515 | * (for example via virtual mapping capabilities) | |
516 | * The routine returns the number of addr/length pairs actually | |
517 | * used, at most nents. | |
518 | * | |
519 | * Device ownership issues as mentioned above for xen_swiotlb_map_page are the | |
520 | * same here. | |
521 | */ | |
522 | int | |
523 | xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, | |
524 | int nelems, enum dma_data_direction dir, | |
525 | struct dma_attrs *attrs) | |
526 | { | |
527 | struct scatterlist *sg; | |
528 | int i; | |
529 | ||
530 | BUG_ON(dir == DMA_NONE); | |
531 | ||
532 | for_each_sg(sgl, sg, nelems, i) { | |
533 | phys_addr_t paddr = sg_phys(sg); | |
534 | dma_addr_t dev_addr = xen_phys_to_bus(paddr); | |
535 | ||
536 | if (swiotlb_force || | |
537 | !dma_capable(hwdev, dev_addr, sg->length) || | |
538 | range_straddles_page_boundary(paddr, sg->length)) { | |
e05ed4d1 AD |
539 | phys_addr_t map = swiotlb_tbl_map_single(hwdev, |
540 | start_dma_addr, | |
541 | sg_phys(sg), | |
542 | sg->length, | |
543 | dir); | |
544 | if (map == SWIOTLB_MAP_ERROR) { | |
783d0281 | 545 | dev_warn(hwdev, "swiotlb buffer is full\n"); |
b097186f KRW |
546 | /* Don't panic here, we expect map_sg users |
547 | to do proper error handling. */ | |
548 | xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir, | |
549 | attrs); | |
781575cd | 550 | sg_dma_len(sgl) = 0; |
b097186f KRW |
551 | return DMA_ERROR_CODE; |
552 | } | |
e05ed4d1 | 553 | sg->dma_address = xen_phys_to_bus(map); |
6cf05463 SS |
554 | } else { |
555 | /* we are not interested in the dma_addr returned by | |
556 | * xen_dma_map_page, only in the potential cache flushes executed | |
557 | * by the function. */ | |
558 | xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT), | |
559 | paddr & ~PAGE_MASK, | |
560 | sg->length, | |
561 | dir, | |
562 | attrs); | |
b097186f | 563 | sg->dma_address = dev_addr; |
6cf05463 | 564 | } |
781575cd | 565 | sg_dma_len(sg) = sg->length; |
b097186f KRW |
566 | } |
567 | return nelems; | |
568 | } | |
569 | EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs); | |
570 | ||
b097186f KRW |
571 | /* |
572 | * Unmap a set of streaming mode DMA translations. Again, cpu read rules | |
573 | * concerning calls here are the same as for swiotlb_unmap_page() above. | |
574 | */ | |
575 | void | |
576 | xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl, | |
577 | int nelems, enum dma_data_direction dir, | |
578 | struct dma_attrs *attrs) | |
579 | { | |
580 | struct scatterlist *sg; | |
581 | int i; | |
582 | ||
583 | BUG_ON(dir == DMA_NONE); | |
584 | ||
585 | for_each_sg(sgl, sg, nelems, i) | |
6cf05463 | 586 | xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs); |
b097186f KRW |
587 | |
588 | } | |
589 | EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs); | |
590 | ||
b097186f KRW |
591 | /* |
592 | * Make physical memory consistent for a set of streaming mode DMA translations | |
593 | * after a transfer. | |
594 | * | |
595 | * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules | |
596 | * and usage. | |
597 | */ | |
598 | static void | |
599 | xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl, | |
600 | int nelems, enum dma_data_direction dir, | |
601 | enum dma_sync_target target) | |
602 | { | |
603 | struct scatterlist *sg; | |
604 | int i; | |
605 | ||
606 | for_each_sg(sgl, sg, nelems, i) | |
607 | xen_swiotlb_sync_single(hwdev, sg->dma_address, | |
781575cd | 608 | sg_dma_len(sg), dir, target); |
b097186f KRW |
609 | } |
610 | ||
611 | void | |
612 | xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg, | |
613 | int nelems, enum dma_data_direction dir) | |
614 | { | |
615 | xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU); | |
616 | } | |
617 | EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu); | |
618 | ||
619 | void | |
620 | xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg, | |
621 | int nelems, enum dma_data_direction dir) | |
622 | { | |
623 | xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE); | |
624 | } | |
625 | EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device); | |
626 | ||
627 | int | |
628 | xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr) | |
629 | { | |
630 | return !dma_addr; | |
631 | } | |
632 | EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error); | |
633 | ||
634 | /* | |
635 | * Return whether the given device DMA address mask can be supported | |
636 | * properly. For example, if your device can only drive the low 24-bits | |
637 | * during bus mastering, then you would pass 0x00ffffff as the mask to | |
638 | * this function. | |
639 | */ | |
640 | int | |
641 | xen_swiotlb_dma_supported(struct device *hwdev, u64 mask) | |
642 | { | |
643 | return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask; | |
644 | } | |
645 | EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported); | |
eb1ddc00 SS |
646 | |
647 | int | |
648 | xen_swiotlb_set_dma_mask(struct device *dev, u64 dma_mask) | |
649 | { | |
650 | if (!dev->dma_mask || !xen_swiotlb_dma_supported(dev, dma_mask)) | |
651 | return -EIO; | |
652 | ||
653 | *dev->dma_mask = dma_mask; | |
654 | ||
655 | return 0; | |
656 | } | |
657 | EXPORT_SYMBOL_GPL(xen_swiotlb_set_dma_mask); |