swiotlb-xen.c 21 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723
  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. #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
  36. #include <linux/memblock.h>
  37. #include <linux/dma-direct.h>
  38. #include <linux/export.h>
  39. #include <xen/swiotlb-xen.h>
  40. #include <xen/page.h>
  41. #include <xen/xen-ops.h>
  42. #include <xen/hvc-console.h>
  43. #include <asm/dma-mapping.h>
  44. #include <asm/xen/page-coherent.h>
  45. #include <trace/events/swiotlb.h>
  46. /*
  47. * Used to do a quick range check in swiotlb_tbl_unmap_single and
  48. * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
  49. * API.
  50. */
  51. #define XEN_SWIOTLB_ERROR_CODE (~(dma_addr_t)0x0)
  52. static char *xen_io_tlb_start, *xen_io_tlb_end;
  53. static unsigned long xen_io_tlb_nslabs;
  54. /*
  55. * Quick lookup value of the bus address of the IOTLB.
  56. */
  57. static u64 start_dma_addr;
  58. /*
  59. * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
  60. * can be 32bit when dma_addr_t is 64bit leading to a loss in
  61. * information if the shift is done before casting to 64bit.
  62. */
  63. static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
  64. {
  65. unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
  66. dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
  67. dma |= paddr & ~XEN_PAGE_MASK;
  68. return dma;
  69. }
  70. static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
  71. {
  72. unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
  73. dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
  74. phys_addr_t paddr = dma;
  75. paddr |= baddr & ~XEN_PAGE_MASK;
  76. return paddr;
  77. }
  78. static inline dma_addr_t xen_virt_to_bus(void *address)
  79. {
  80. return xen_phys_to_bus(virt_to_phys(address));
  81. }
  82. static int check_pages_physically_contiguous(unsigned long xen_pfn,
  83. unsigned int offset,
  84. size_t length)
  85. {
  86. unsigned long next_bfn;
  87. int i;
  88. int nr_pages;
  89. next_bfn = pfn_to_bfn(xen_pfn);
  90. nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT;
  91. for (i = 1; i < nr_pages; i++) {
  92. if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
  93. return 0;
  94. }
  95. return 1;
  96. }
  97. static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
  98. {
  99. unsigned long xen_pfn = XEN_PFN_DOWN(p);
  100. unsigned int offset = p & ~XEN_PAGE_MASK;
  101. if (offset + size <= XEN_PAGE_SIZE)
  102. return 0;
  103. if (check_pages_physically_contiguous(xen_pfn, offset, size))
  104. return 0;
  105. return 1;
  106. }
  107. static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
  108. {
  109. unsigned long bfn = XEN_PFN_DOWN(dma_addr);
  110. unsigned long xen_pfn = bfn_to_local_pfn(bfn);
  111. phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
  112. /* If the address is outside our domain, it CAN
  113. * have the same virtual address as another address
  114. * in our domain. Therefore _only_ check address within our domain.
  115. */
  116. if (pfn_valid(PFN_DOWN(paddr))) {
  117. return paddr >= virt_to_phys(xen_io_tlb_start) &&
  118. paddr < virt_to_phys(xen_io_tlb_end);
  119. }
  120. return 0;
  121. }
  122. static int max_dma_bits = 32;
  123. static int
  124. xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
  125. {
  126. int i, rc;
  127. int dma_bits;
  128. dma_addr_t dma_handle;
  129. phys_addr_t p = virt_to_phys(buf);
  130. dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
  131. i = 0;
  132. do {
  133. int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
  134. do {
  135. rc = xen_create_contiguous_region(
  136. p + (i << IO_TLB_SHIFT),
  137. get_order(slabs << IO_TLB_SHIFT),
  138. dma_bits, &dma_handle);
  139. } while (rc && dma_bits++ < max_dma_bits);
  140. if (rc)
  141. return rc;
  142. i += slabs;
  143. } while (i < nslabs);
  144. return 0;
  145. }
  146. static unsigned long xen_set_nslabs(unsigned long nr_tbl)
  147. {
  148. if (!nr_tbl) {
  149. xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
  150. xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
  151. } else
  152. xen_io_tlb_nslabs = nr_tbl;
  153. return xen_io_tlb_nslabs << IO_TLB_SHIFT;
  154. }
  155. enum xen_swiotlb_err {
  156. XEN_SWIOTLB_UNKNOWN = 0,
  157. XEN_SWIOTLB_ENOMEM,
  158. XEN_SWIOTLB_EFIXUP
  159. };
  160. static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
  161. {
  162. switch (err) {
  163. case XEN_SWIOTLB_ENOMEM:
  164. return "Cannot allocate Xen-SWIOTLB buffer\n";
  165. case XEN_SWIOTLB_EFIXUP:
  166. return "Failed to get contiguous memory for DMA from Xen!\n"\
  167. "You either: don't have the permissions, do not have"\
  168. " enough free memory under 4GB, or the hypervisor memory"\
  169. " is too fragmented!";
  170. default:
  171. break;
  172. }
  173. return "";
  174. }
  175. int __ref xen_swiotlb_init(int verbose, bool early)
  176. {
  177. unsigned long bytes, order;
  178. int rc = -ENOMEM;
  179. enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
  180. unsigned int repeat = 3;
  181. xen_io_tlb_nslabs = swiotlb_nr_tbl();
  182. retry:
  183. bytes = xen_set_nslabs(xen_io_tlb_nslabs);
  184. order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
  185. /*
  186. * Get IO TLB memory from any location.
  187. */
  188. if (early)
  189. xen_io_tlb_start = memblock_alloc(PAGE_ALIGN(bytes),
  190. PAGE_SIZE);
  191. else {
  192. #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
  193. #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
  194. while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
  195. xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
  196. if (xen_io_tlb_start)
  197. break;
  198. order--;
  199. }
  200. if (order != get_order(bytes)) {
  201. pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
  202. (PAGE_SIZE << order) >> 20);
  203. xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
  204. bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
  205. }
  206. }
  207. if (!xen_io_tlb_start) {
  208. m_ret = XEN_SWIOTLB_ENOMEM;
  209. goto error;
  210. }
  211. xen_io_tlb_end = xen_io_tlb_start + bytes;
  212. /*
  213. * And replace that memory with pages under 4GB.
  214. */
  215. rc = xen_swiotlb_fixup(xen_io_tlb_start,
  216. bytes,
  217. xen_io_tlb_nslabs);
  218. if (rc) {
  219. if (early)
  220. memblock_free(__pa(xen_io_tlb_start),
  221. PAGE_ALIGN(bytes));
  222. else {
  223. free_pages((unsigned long)xen_io_tlb_start, order);
  224. xen_io_tlb_start = NULL;
  225. }
  226. m_ret = XEN_SWIOTLB_EFIXUP;
  227. goto error;
  228. }
  229. start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
  230. if (early) {
  231. if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
  232. verbose))
  233. panic("Cannot allocate SWIOTLB buffer");
  234. rc = 0;
  235. } else
  236. rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
  237. if (!rc)
  238. swiotlb_set_max_segment(PAGE_SIZE);
  239. return rc;
  240. error:
  241. if (repeat--) {
  242. xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
  243. (xen_io_tlb_nslabs >> 1));
  244. pr_info("Lowering to %luMB\n",
  245. (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
  246. goto retry;
  247. }
  248. pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
  249. if (early)
  250. panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
  251. else
  252. free_pages((unsigned long)xen_io_tlb_start, order);
  253. return rc;
  254. }
  255. static void *
  256. xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
  257. dma_addr_t *dma_handle, gfp_t flags,
  258. unsigned long attrs)
  259. {
  260. void *ret;
  261. int order = get_order(size);
  262. u64 dma_mask = DMA_BIT_MASK(32);
  263. phys_addr_t phys;
  264. dma_addr_t dev_addr;
  265. /*
  266. * Ignore region specifiers - the kernel's ideas of
  267. * pseudo-phys memory layout has nothing to do with the
  268. * machine physical layout. We can't allocate highmem
  269. * because we can't return a pointer to it.
  270. */
  271. flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
  272. /* Convert the size to actually allocated. */
  273. size = 1UL << (order + XEN_PAGE_SHIFT);
  274. /* On ARM this function returns an ioremap'ped virtual address for
  275. * which virt_to_phys doesn't return the corresponding physical
  276. * address. In fact on ARM virt_to_phys only works for kernel direct
  277. * mapped RAM memory. Also see comment below.
  278. */
  279. ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
  280. if (!ret)
  281. return ret;
  282. if (hwdev && hwdev->coherent_dma_mask)
  283. dma_mask = hwdev->coherent_dma_mask;
  284. /* At this point dma_handle is the physical address, next we are
  285. * going to set it to the machine address.
  286. * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
  287. * to *dma_handle. */
  288. phys = *dma_handle;
  289. dev_addr = xen_phys_to_bus(phys);
  290. if (((dev_addr + size - 1 <= dma_mask)) &&
  291. !range_straddles_page_boundary(phys, size))
  292. *dma_handle = dev_addr;
  293. else {
  294. if (xen_create_contiguous_region(phys, order,
  295. fls64(dma_mask), dma_handle) != 0) {
  296. xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
  297. return NULL;
  298. }
  299. }
  300. memset(ret, 0, size);
  301. return ret;
  302. }
  303. static void
  304. xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
  305. dma_addr_t dev_addr, unsigned long attrs)
  306. {
  307. int order = get_order(size);
  308. phys_addr_t phys;
  309. u64 dma_mask = DMA_BIT_MASK(32);
  310. if (hwdev && hwdev->coherent_dma_mask)
  311. dma_mask = hwdev->coherent_dma_mask;
  312. /* do not use virt_to_phys because on ARM it doesn't return you the
  313. * physical address */
  314. phys = xen_bus_to_phys(dev_addr);
  315. /* Convert the size to actually allocated. */
  316. size = 1UL << (order + XEN_PAGE_SHIFT);
  317. if (((dev_addr + size - 1 <= dma_mask)) ||
  318. range_straddles_page_boundary(phys, size))
  319. xen_destroy_contiguous_region(phys, order);
  320. xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
  321. }
  322. /*
  323. * Map a single buffer of the indicated size for DMA in streaming mode. The
  324. * physical address to use is returned.
  325. *
  326. * Once the device is given the dma address, the device owns this memory until
  327. * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
  328. */
  329. static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
  330. unsigned long offset, size_t size,
  331. enum dma_data_direction dir,
  332. unsigned long attrs)
  333. {
  334. phys_addr_t map, phys = page_to_phys(page) + offset;
  335. dma_addr_t dev_addr = xen_phys_to_bus(phys);
  336. BUG_ON(dir == DMA_NONE);
  337. /*
  338. * If the address happens to be in the device's DMA window,
  339. * we can safely return the device addr and not worry about bounce
  340. * buffering it.
  341. */
  342. if (dma_capable(dev, dev_addr, size) &&
  343. !range_straddles_page_boundary(phys, size) &&
  344. !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
  345. (swiotlb_force != SWIOTLB_FORCE)) {
  346. /* we are not interested in the dma_addr returned by
  347. * xen_dma_map_page, only in the potential cache flushes executed
  348. * by the function. */
  349. xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
  350. return dev_addr;
  351. }
  352. /*
  353. * Oh well, have to allocate and map a bounce buffer.
  354. */
  355. trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
  356. map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
  357. attrs);
  358. if (map == SWIOTLB_MAP_ERROR)
  359. return XEN_SWIOTLB_ERROR_CODE;
  360. dev_addr = xen_phys_to_bus(map);
  361. xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
  362. dev_addr, map & ~PAGE_MASK, size, dir, attrs);
  363. /*
  364. * Ensure that the address returned is DMA'ble
  365. */
  366. if (dma_capable(dev, dev_addr, size))
  367. return dev_addr;
  368. attrs |= DMA_ATTR_SKIP_CPU_SYNC;
  369. swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
  370. return XEN_SWIOTLB_ERROR_CODE;
  371. }
  372. /*
  373. * Unmap a single streaming mode DMA translation. The dma_addr and size must
  374. * match what was provided for in a previous xen_swiotlb_map_page call. All
  375. * other usages are undefined.
  376. *
  377. * After this call, reads by the cpu to the buffer are guaranteed to see
  378. * whatever the device wrote there.
  379. */
  380. static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
  381. size_t size, enum dma_data_direction dir,
  382. unsigned long attrs)
  383. {
  384. phys_addr_t paddr = xen_bus_to_phys(dev_addr);
  385. BUG_ON(dir == DMA_NONE);
  386. xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
  387. /* NOTE: We use dev_addr here, not paddr! */
  388. if (is_xen_swiotlb_buffer(dev_addr)) {
  389. swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
  390. return;
  391. }
  392. if (dir != DMA_FROM_DEVICE)
  393. return;
  394. /*
  395. * phys_to_virt doesn't work with hihgmem page but we could
  396. * call dma_mark_clean() with hihgmem page here. However, we
  397. * are fine since dma_mark_clean() is null on POWERPC. We can
  398. * make dma_mark_clean() take a physical address if necessary.
  399. */
  400. dma_mark_clean(phys_to_virt(paddr), size);
  401. }
  402. static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
  403. size_t size, enum dma_data_direction dir,
  404. unsigned long attrs)
  405. {
  406. xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
  407. }
  408. /*
  409. * Make physical memory consistent for a single streaming mode DMA translation
  410. * after a transfer.
  411. *
  412. * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
  413. * using the cpu, yet do not wish to teardown the dma mapping, you must
  414. * call this function before doing so. At the next point you give the dma
  415. * address back to the card, you must first perform a
  416. * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
  417. */
  418. static void
  419. xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
  420. size_t size, enum dma_data_direction dir,
  421. enum dma_sync_target target)
  422. {
  423. phys_addr_t paddr = xen_bus_to_phys(dev_addr);
  424. BUG_ON(dir == DMA_NONE);
  425. if (target == SYNC_FOR_CPU)
  426. xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
  427. /* NOTE: We use dev_addr here, not paddr! */
  428. if (is_xen_swiotlb_buffer(dev_addr))
  429. swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
  430. if (target == SYNC_FOR_DEVICE)
  431. xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
  432. if (dir != DMA_FROM_DEVICE)
  433. return;
  434. dma_mark_clean(phys_to_virt(paddr), size);
  435. }
  436. void
  437. xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
  438. size_t size, enum dma_data_direction dir)
  439. {
  440. xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
  441. }
  442. void
  443. xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
  444. size_t size, enum dma_data_direction dir)
  445. {
  446. xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
  447. }
  448. /*
  449. * Unmap a set of streaming mode DMA translations. Again, cpu read rules
  450. * concerning calls here are the same as for swiotlb_unmap_page() above.
  451. */
  452. static void
  453. xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
  454. int nelems, enum dma_data_direction dir,
  455. unsigned long attrs)
  456. {
  457. struct scatterlist *sg;
  458. int i;
  459. BUG_ON(dir == DMA_NONE);
  460. for_each_sg(sgl, sg, nelems, i)
  461. xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
  462. }
  463. /*
  464. * Map a set of buffers described by scatterlist in streaming mode for DMA.
  465. * This is the scatter-gather version of the above xen_swiotlb_map_page
  466. * interface. Here the scatter gather list elements are each tagged with the
  467. * appropriate dma address and length. They are obtained via
  468. * sg_dma_{address,length}(SG).
  469. *
  470. * NOTE: An implementation may be able to use a smaller number of
  471. * DMA address/length pairs than there are SG table elements.
  472. * (for example via virtual mapping capabilities)
  473. * The routine returns the number of addr/length pairs actually
  474. * used, at most nents.
  475. *
  476. * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
  477. * same here.
  478. */
  479. static int
  480. xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
  481. int nelems, enum dma_data_direction dir,
  482. unsigned long attrs)
  483. {
  484. struct scatterlist *sg;
  485. int i;
  486. BUG_ON(dir == DMA_NONE);
  487. for_each_sg(sgl, sg, nelems, i) {
  488. phys_addr_t paddr = sg_phys(sg);
  489. dma_addr_t dev_addr = xen_phys_to_bus(paddr);
  490. if (swiotlb_force == SWIOTLB_FORCE ||
  491. xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
  492. !dma_capable(hwdev, dev_addr, sg->length) ||
  493. range_straddles_page_boundary(paddr, sg->length)) {
  494. phys_addr_t map = swiotlb_tbl_map_single(hwdev,
  495. start_dma_addr,
  496. sg_phys(sg),
  497. sg->length,
  498. dir, attrs);
  499. if (map == SWIOTLB_MAP_ERROR) {
  500. dev_warn(hwdev, "swiotlb buffer is full\n");
  501. /* Don't panic here, we expect map_sg users
  502. to do proper error handling. */
  503. attrs |= DMA_ATTR_SKIP_CPU_SYNC;
  504. xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
  505. attrs);
  506. sg_dma_len(sgl) = 0;
  507. return 0;
  508. }
  509. dev_addr = xen_phys_to_bus(map);
  510. xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
  511. dev_addr,
  512. map & ~PAGE_MASK,
  513. sg->length,
  514. dir,
  515. attrs);
  516. sg->dma_address = dev_addr;
  517. } else {
  518. /* we are not interested in the dma_addr returned by
  519. * xen_dma_map_page, only in the potential cache flushes executed
  520. * by the function. */
  521. xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
  522. dev_addr,
  523. paddr & ~PAGE_MASK,
  524. sg->length,
  525. dir,
  526. attrs);
  527. sg->dma_address = dev_addr;
  528. }
  529. sg_dma_len(sg) = sg->length;
  530. }
  531. return nelems;
  532. }
  533. /*
  534. * Make physical memory consistent for a set of streaming mode DMA translations
  535. * after a transfer.
  536. *
  537. * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
  538. * and usage.
  539. */
  540. static void
  541. xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
  542. int nelems, enum dma_data_direction dir,
  543. enum dma_sync_target target)
  544. {
  545. struct scatterlist *sg;
  546. int i;
  547. for_each_sg(sgl, sg, nelems, i)
  548. xen_swiotlb_sync_single(hwdev, sg->dma_address,
  549. sg_dma_len(sg), dir, target);
  550. }
  551. static void
  552. xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
  553. int nelems, enum dma_data_direction dir)
  554. {
  555. xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
  556. }
  557. static void
  558. xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
  559. int nelems, enum dma_data_direction dir)
  560. {
  561. xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
  562. }
  563. /*
  564. * Return whether the given device DMA address mask can be supported
  565. * properly. For example, if your device can only drive the low 24-bits
  566. * during bus mastering, then you would pass 0x00ffffff as the mask to
  567. * this function.
  568. */
  569. static int
  570. xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
  571. {
  572. return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
  573. }
  574. /*
  575. * Create userspace mapping for the DMA-coherent memory.
  576. * This function should be called with the pages from the current domain only,
  577. * passing pages mapped from other domains would lead to memory corruption.
  578. */
  579. static int
  580. xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
  581. void *cpu_addr, dma_addr_t dma_addr, size_t size,
  582. unsigned long attrs)
  583. {
  584. #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
  585. if (xen_get_dma_ops(dev)->mmap)
  586. return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
  587. dma_addr, size, attrs);
  588. #endif
  589. return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
  590. }
  591. /*
  592. * This function should be called with the pages from the current domain only,
  593. * passing pages mapped from other domains would lead to memory corruption.
  594. */
  595. static int
  596. xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
  597. void *cpu_addr, dma_addr_t handle, size_t size,
  598. unsigned long attrs)
  599. {
  600. #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
  601. if (xen_get_dma_ops(dev)->get_sgtable) {
  602. #if 0
  603. /*
  604. * This check verifies that the page belongs to the current domain and
  605. * is not one mapped from another domain.
  606. * This check is for debug only, and should not go to production build
  607. */
  608. unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
  609. BUG_ON (!page_is_ram(bfn));
  610. #endif
  611. return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
  612. handle, size, attrs);
  613. }
  614. #endif
  615. return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs);
  616. }
  617. static int xen_swiotlb_mapping_error(struct device *dev, dma_addr_t dma_addr)
  618. {
  619. return dma_addr == XEN_SWIOTLB_ERROR_CODE;
  620. }
  621. const struct dma_map_ops xen_swiotlb_dma_ops = {
  622. .alloc = xen_swiotlb_alloc_coherent,
  623. .free = xen_swiotlb_free_coherent,
  624. .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
  625. .sync_single_for_device = xen_swiotlb_sync_single_for_device,
  626. .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
  627. .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
  628. .map_sg = xen_swiotlb_map_sg_attrs,
  629. .unmap_sg = xen_swiotlb_unmap_sg_attrs,
  630. .map_page = xen_swiotlb_map_page,
  631. .unmap_page = xen_swiotlb_unmap_page,
  632. .dma_supported = xen_swiotlb_dma_supported,
  633. .mmap = xen_swiotlb_dma_mmap,
  634. .get_sgtable = xen_swiotlb_get_sgtable,
  635. .mapping_error = xen_swiotlb_mapping_error,
  636. };