/drivers/xen/swiotlb-xen.c

https://bitbucket.org/cyanogenmod/android_kernel_asus_tf300t · C · 522 lines · 315 code · 70 blank · 137 comment · 38 complexity · a3dff569cf25a2ec5daf5b4e4a3e0221 MD5 · raw file

  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. #include <linux/bootmem.h>
  36. #include <linux/dma-mapping.h>
  37. #include <xen/swiotlb-xen.h>
  38. #include <xen/page.h>
  39. #include <xen/xen-ops.h>
  40. /*
  41. * Used to do a quick range check in swiotlb_tbl_unmap_single and
  42. * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
  43. * API.
  44. */
  45. static char *xen_io_tlb_start, *xen_io_tlb_end;
  46. static unsigned long xen_io_tlb_nslabs;
  47. /*
  48. * Quick lookup value of the bus address of the IOTLB.
  49. */
  50. u64 start_dma_addr;
  51. static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
  52. {
  53. return phys_to_machine(XPADDR(paddr)).maddr;
  54. }
  55. static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
  56. {
  57. return machine_to_phys(XMADDR(baddr)).paddr;
  58. }
  59. static dma_addr_t xen_virt_to_bus(void *address)
  60. {
  61. return xen_phys_to_bus(virt_to_phys(address));
  62. }
  63. static int check_pages_physically_contiguous(unsigned long pfn,
  64. unsigned int offset,
  65. size_t length)
  66. {
  67. unsigned long next_mfn;
  68. int i;
  69. int nr_pages;
  70. next_mfn = pfn_to_mfn(pfn);
  71. nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
  72. for (i = 1; i < nr_pages; i++) {
  73. if (pfn_to_mfn(++pfn) != ++next_mfn)
  74. return 0;
  75. }
  76. return 1;
  77. }
  78. static int range_straddles_page_boundary(phys_addr_t p, size_t size)
  79. {
  80. unsigned long pfn = PFN_DOWN(p);
  81. unsigned int offset = p & ~PAGE_MASK;
  82. if (offset + size <= PAGE_SIZE)
  83. return 0;
  84. if (check_pages_physically_contiguous(pfn, offset, size))
  85. return 0;
  86. return 1;
  87. }
  88. static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
  89. {
  90. unsigned long mfn = PFN_DOWN(dma_addr);
  91. unsigned long pfn = mfn_to_local_pfn(mfn);
  92. phys_addr_t paddr;
  93. /* If the address is outside our domain, it CAN
  94. * have the same virtual address as another address
  95. * in our domain. Therefore _only_ check address within our domain.
  96. */
  97. if (pfn_valid(pfn)) {
  98. paddr = PFN_PHYS(pfn);
  99. return paddr >= virt_to_phys(xen_io_tlb_start) &&
  100. paddr < virt_to_phys(xen_io_tlb_end);
  101. }
  102. return 0;
  103. }
  104. static int max_dma_bits = 32;
  105. static int
  106. xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
  107. {
  108. int i, rc;
  109. int dma_bits;
  110. dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
  111. i = 0;
  112. do {
  113. int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
  114. do {
  115. rc = xen_create_contiguous_region(
  116. (unsigned long)buf + (i << IO_TLB_SHIFT),
  117. get_order(slabs << IO_TLB_SHIFT),
  118. dma_bits);
  119. } while (rc && dma_bits++ < max_dma_bits);
  120. if (rc)
  121. return rc;
  122. i += slabs;
  123. } while (i < nslabs);
  124. return 0;
  125. }
  126. void __init xen_swiotlb_init(int verbose)
  127. {
  128. unsigned long bytes;
  129. int rc;
  130. unsigned long nr_tbl;
  131. nr_tbl = swioltb_nr_tbl();
  132. if (nr_tbl)
  133. xen_io_tlb_nslabs = nr_tbl;
  134. else {
  135. xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
  136. xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
  137. }
  138. bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
  139. /*
  140. * Get IO TLB memory from any location.
  141. */
  142. xen_io_tlb_start = alloc_bootmem(bytes);
  143. if (!xen_io_tlb_start)
  144. panic("Cannot allocate SWIOTLB buffer");
  145. xen_io_tlb_end = xen_io_tlb_start + bytes;
  146. /*
  147. * And replace that memory with pages under 4GB.
  148. */
  149. rc = xen_swiotlb_fixup(xen_io_tlb_start,
  150. bytes,
  151. xen_io_tlb_nslabs);
  152. if (rc)
  153. goto error;
  154. start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
  155. swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
  156. return;
  157. error:
  158. panic("DMA(%d): Failed to exchange pages allocated for DMA with Xen! "\
  159. "We either don't have the permission or you do not have enough"\
  160. "free memory under 4GB!\n", rc);
  161. }
  162. void *
  163. xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
  164. dma_addr_t *dma_handle, gfp_t flags)
  165. {
  166. void *ret;
  167. int order = get_order(size);
  168. u64 dma_mask = DMA_BIT_MASK(32);
  169. unsigned long vstart;
  170. /*
  171. * Ignore region specifiers - the kernel's ideas of
  172. * pseudo-phys memory layout has nothing to do with the
  173. * machine physical layout. We can't allocate highmem
  174. * because we can't return a pointer to it.
  175. */
  176. flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
  177. if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
  178. return ret;
  179. vstart = __get_free_pages(flags, order);
  180. ret = (void *)vstart;
  181. if (hwdev && hwdev->coherent_dma_mask)
  182. dma_mask = dma_alloc_coherent_mask(hwdev, flags);
  183. if (ret) {
  184. if (xen_create_contiguous_region(vstart, order,
  185. fls64(dma_mask)) != 0) {
  186. free_pages(vstart, order);
  187. return NULL;
  188. }
  189. memset(ret, 0, size);
  190. *dma_handle = virt_to_machine(ret).maddr;
  191. }
  192. return ret;
  193. }
  194. EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
  195. void
  196. xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
  197. dma_addr_t dev_addr)
  198. {
  199. int order = get_order(size);
  200. if (dma_release_from_coherent(hwdev, order, vaddr))
  201. return;
  202. xen_destroy_contiguous_region((unsigned long)vaddr, order);
  203. free_pages((unsigned long)vaddr, order);
  204. }
  205. EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
  206. /*
  207. * Map a single buffer of the indicated size for DMA in streaming mode. The
  208. * physical address to use is returned.
  209. *
  210. * Once the device is given the dma address, the device owns this memory until
  211. * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
  212. */
  213. dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
  214. unsigned long offset, size_t size,
  215. enum dma_data_direction dir,
  216. struct dma_attrs *attrs)
  217. {
  218. phys_addr_t phys = page_to_phys(page) + offset;
  219. dma_addr_t dev_addr = xen_phys_to_bus(phys);
  220. void *map;
  221. BUG_ON(dir == DMA_NONE);
  222. /*
  223. * If the address happens to be in the device's DMA window,
  224. * we can safely return the device addr and not worry about bounce
  225. * buffering it.
  226. */
  227. if (dma_capable(dev, dev_addr, size) &&
  228. !range_straddles_page_boundary(phys, size) && !swiotlb_force)
  229. return dev_addr;
  230. /*
  231. * Oh well, have to allocate and map a bounce buffer.
  232. */
  233. map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
  234. if (!map)
  235. return DMA_ERROR_CODE;
  236. dev_addr = xen_virt_to_bus(map);
  237. /*
  238. * Ensure that the address returned is DMA'ble
  239. */
  240. if (!dma_capable(dev, dev_addr, size)) {
  241. swiotlb_tbl_unmap_single(dev, map, size, dir);
  242. dev_addr = 0;
  243. }
  244. return dev_addr;
  245. }
  246. EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
  247. /*
  248. * Unmap a single streaming mode DMA translation. The dma_addr and size must
  249. * match what was provided for in a previous xen_swiotlb_map_page call. All
  250. * other usages are undefined.
  251. *
  252. * After this call, reads by the cpu to the buffer are guaranteed to see
  253. * whatever the device wrote there.
  254. */
  255. static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
  256. size_t size, enum dma_data_direction dir)
  257. {
  258. phys_addr_t paddr = xen_bus_to_phys(dev_addr);
  259. BUG_ON(dir == DMA_NONE);
  260. /* NOTE: We use dev_addr here, not paddr! */
  261. if (is_xen_swiotlb_buffer(dev_addr)) {
  262. swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
  263. return;
  264. }
  265. if (dir != DMA_FROM_DEVICE)
  266. return;
  267. /*
  268. * phys_to_virt doesn't work with hihgmem page but we could
  269. * call dma_mark_clean() with hihgmem page here. However, we
  270. * are fine since dma_mark_clean() is null on POWERPC. We can
  271. * make dma_mark_clean() take a physical address if necessary.
  272. */
  273. dma_mark_clean(phys_to_virt(paddr), size);
  274. }
  275. void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
  276. size_t size, enum dma_data_direction dir,
  277. struct dma_attrs *attrs)
  278. {
  279. xen_unmap_single(hwdev, dev_addr, size, dir);
  280. }
  281. EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
  282. /*
  283. * Make physical memory consistent for a single streaming mode DMA translation
  284. * after a transfer.
  285. *
  286. * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
  287. * using the cpu, yet do not wish to teardown the dma mapping, you must
  288. * call this function before doing so. At the next point you give the dma
  289. * address back to the card, you must first perform a
  290. * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
  291. */
  292. static void
  293. xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
  294. size_t size, enum dma_data_direction dir,
  295. enum dma_sync_target target)
  296. {
  297. phys_addr_t paddr = xen_bus_to_phys(dev_addr);
  298. BUG_ON(dir == DMA_NONE);
  299. /* NOTE: We use dev_addr here, not paddr! */
  300. if (is_xen_swiotlb_buffer(dev_addr)) {
  301. swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
  302. target);
  303. return;
  304. }
  305. if (dir != DMA_FROM_DEVICE)
  306. return;
  307. dma_mark_clean(phys_to_virt(paddr), size);
  308. }
  309. void
  310. xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
  311. size_t size, enum dma_data_direction dir)
  312. {
  313. xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
  314. }
  315. EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
  316. void
  317. xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
  318. size_t size, enum dma_data_direction dir)
  319. {
  320. xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
  321. }
  322. EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
  323. /*
  324. * Map a set of buffers described by scatterlist in streaming mode for DMA.
  325. * This is the scatter-gather version of the above xen_swiotlb_map_page
  326. * interface. Here the scatter gather list elements are each tagged with the
  327. * appropriate dma address and length. They are obtained via
  328. * sg_dma_{address,length}(SG).
  329. *
  330. * NOTE: An implementation may be able to use a smaller number of
  331. * DMA address/length pairs than there are SG table elements.
  332. * (for example via virtual mapping capabilities)
  333. * The routine returns the number of addr/length pairs actually
  334. * used, at most nents.
  335. *
  336. * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
  337. * same here.
  338. */
  339. int
  340. xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
  341. int nelems, enum dma_data_direction dir,
  342. struct dma_attrs *attrs)
  343. {
  344. struct scatterlist *sg;
  345. int i;
  346. BUG_ON(dir == DMA_NONE);
  347. for_each_sg(sgl, sg, nelems, i) {
  348. phys_addr_t paddr = sg_phys(sg);
  349. dma_addr_t dev_addr = xen_phys_to_bus(paddr);
  350. if (swiotlb_force ||
  351. !dma_capable(hwdev, dev_addr, sg->length) ||
  352. range_straddles_page_boundary(paddr, sg->length)) {
  353. void *map = swiotlb_tbl_map_single(hwdev,
  354. start_dma_addr,
  355. sg_phys(sg),
  356. sg->length, dir);
  357. if (!map) {
  358. /* Don't panic here, we expect map_sg users
  359. to do proper error handling. */
  360. xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
  361. attrs);
  362. sgl[0].dma_length = 0;
  363. return DMA_ERROR_CODE;
  364. }
  365. sg->dma_address = xen_virt_to_bus(map);
  366. } else
  367. sg->dma_address = dev_addr;
  368. sg->dma_length = sg->length;
  369. }
  370. return nelems;
  371. }
  372. EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
  373. int
  374. xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
  375. enum dma_data_direction dir)
  376. {
  377. return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
  378. }
  379. EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
  380. /*
  381. * Unmap a set of streaming mode DMA translations. Again, cpu read rules
  382. * concerning calls here are the same as for swiotlb_unmap_page() above.
  383. */
  384. void
  385. xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
  386. int nelems, enum dma_data_direction dir,
  387. struct dma_attrs *attrs)
  388. {
  389. struct scatterlist *sg;
  390. int i;
  391. BUG_ON(dir == DMA_NONE);
  392. for_each_sg(sgl, sg, nelems, i)
  393. xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
  394. }
  395. EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
  396. void
  397. xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
  398. enum dma_data_direction dir)
  399. {
  400. return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
  401. }
  402. EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
  403. /*
  404. * Make physical memory consistent for a set of streaming mode DMA translations
  405. * after a transfer.
  406. *
  407. * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
  408. * and usage.
  409. */
  410. static void
  411. xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
  412. int nelems, enum dma_data_direction dir,
  413. enum dma_sync_target target)
  414. {
  415. struct scatterlist *sg;
  416. int i;
  417. for_each_sg(sgl, sg, nelems, i)
  418. xen_swiotlb_sync_single(hwdev, sg->dma_address,
  419. sg->dma_length, dir, target);
  420. }
  421. void
  422. xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
  423. int nelems, enum dma_data_direction dir)
  424. {
  425. xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
  426. }
  427. EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
  428. void
  429. xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
  430. int nelems, enum dma_data_direction dir)
  431. {
  432. xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
  433. }
  434. EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
  435. int
  436. xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
  437. {
  438. return !dma_addr;
  439. }
  440. EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
  441. /*
  442. * Return whether the given device DMA address mask can be supported
  443. * properly. For example, if your device can only drive the low 24-bits
  444. * during bus mastering, then you would pass 0x00ffffff as the mask to
  445. * this function.
  446. */
  447. int
  448. xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
  449. {
  450. return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
  451. }
  452. EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);