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/memblock.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 static char *xen_io_tlb_start, *xen_io_tlb_end; 57 static unsigned long xen_io_tlb_nslabs; 58 /* 59 * Quick lookup value of the bus address of the IOTLB. 60 */ 61 62 static u64 start_dma_addr; 63 64 /* 65 * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t 66 * can be 32bit when dma_addr_t is 64bit leading to a loss in 67 * information if the shift is done before casting to 64bit. 68 */ 69 static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr) 70 { 71 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr)); 72 dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT; 73 74 dma |= paddr & ~XEN_PAGE_MASK; 75 76 return dma; 77 } 78 79 static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr) 80 { 81 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr)); 82 dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT; 83 phys_addr_t paddr = dma; 84 85 paddr |= baddr & ~XEN_PAGE_MASK; 86 87 return paddr; 88 } 89 90 static inline dma_addr_t xen_virt_to_bus(void *address) 91 { 92 return xen_phys_to_bus(virt_to_phys(address)); 93 } 94 95 static int check_pages_physically_contiguous(unsigned long xen_pfn, 96 unsigned int offset, 97 size_t length) 98 { 99 unsigned long next_bfn; 100 int i; 101 int nr_pages; 102 103 next_bfn = pfn_to_bfn(xen_pfn); 104 nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT; 105 106 for (i = 1; i < nr_pages; i++) { 107 if (pfn_to_bfn(++xen_pfn) != ++next_bfn) 108 return 0; 109 } 110 return 1; 111 } 112 113 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size) 114 { 115 unsigned long xen_pfn = XEN_PFN_DOWN(p); 116 unsigned int offset = p & ~XEN_PAGE_MASK; 117 118 if (offset + size <= XEN_PAGE_SIZE) 119 return 0; 120 if (check_pages_physically_contiguous(xen_pfn, offset, size)) 121 return 0; 122 return 1; 123 } 124 125 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr) 126 { 127 unsigned long bfn = XEN_PFN_DOWN(dma_addr); 128 unsigned long xen_pfn = bfn_to_local_pfn(bfn); 129 phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn); 130 131 /* If the address is outside our domain, it CAN 132 * have the same virtual address as another address 133 * in our domain. Therefore _only_ check address within our domain. 134 */ 135 if (pfn_valid(PFN_DOWN(paddr))) { 136 return paddr >= virt_to_phys(xen_io_tlb_start) && 137 paddr < virt_to_phys(xen_io_tlb_end); 138 } 139 return 0; 140 } 141 142 static int max_dma_bits = 32; 143 144 static int 145 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs) 146 { 147 int i, rc; 148 int dma_bits; 149 dma_addr_t dma_handle; 150 phys_addr_t p = virt_to_phys(buf); 151 152 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT; 153 154 i = 0; 155 do { 156 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE); 157 158 do { 159 rc = xen_create_contiguous_region( 160 p + (i << IO_TLB_SHIFT), 161 get_order(slabs << IO_TLB_SHIFT), 162 dma_bits, &dma_handle); 163 } while (rc && dma_bits++ < max_dma_bits); 164 if (rc) 165 return rc; 166 167 i += slabs; 168 } while (i < nslabs); 169 return 0; 170 } 171 static unsigned long xen_set_nslabs(unsigned long nr_tbl) 172 { 173 if (!nr_tbl) { 174 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT); 175 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE); 176 } else 177 xen_io_tlb_nslabs = nr_tbl; 178 179 return xen_io_tlb_nslabs << IO_TLB_SHIFT; 180 } 181 182 enum xen_swiotlb_err { 183 XEN_SWIOTLB_UNKNOWN = 0, 184 XEN_SWIOTLB_ENOMEM, 185 XEN_SWIOTLB_EFIXUP 186 }; 187 188 static const char *xen_swiotlb_error(enum xen_swiotlb_err err) 189 { 190 switch (err) { 191 case XEN_SWIOTLB_ENOMEM: 192 return "Cannot allocate Xen-SWIOTLB buffer\n"; 193 case XEN_SWIOTLB_EFIXUP: 194 return "Failed to get contiguous memory for DMA from Xen!\n"\ 195 "You either: don't have the permissions, do not have"\ 196 " enough free memory under 4GB, or the hypervisor memory"\ 197 " is too fragmented!"; 198 default: 199 break; 200 } 201 return ""; 202 } 203 int __ref xen_swiotlb_init(int verbose, bool early) 204 { 205 unsigned long bytes, order; 206 int rc = -ENOMEM; 207 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN; 208 unsigned int repeat = 3; 209 210 xen_io_tlb_nslabs = swiotlb_nr_tbl(); 211 retry: 212 bytes = xen_set_nslabs(xen_io_tlb_nslabs); 213 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT); 214 /* 215 * Get IO TLB memory from any location. 216 */ 217 if (early) 218 xen_io_tlb_start = memblock_alloc(PAGE_ALIGN(bytes), 219 PAGE_SIZE); 220 else { 221 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT)) 222 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT) 223 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) { 224 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order); 225 if (xen_io_tlb_start) 226 break; 227 order--; 228 } 229 if (order != get_order(bytes)) { 230 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n", 231 (PAGE_SIZE << order) >> 20); 232 xen_io_tlb_nslabs = SLABS_PER_PAGE << order; 233 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT; 234 } 235 } 236 if (!xen_io_tlb_start) { 237 m_ret = XEN_SWIOTLB_ENOMEM; 238 goto error; 239 } 240 xen_io_tlb_end = xen_io_tlb_start + bytes; 241 /* 242 * And replace that memory with pages under 4GB. 243 */ 244 rc = xen_swiotlb_fixup(xen_io_tlb_start, 245 bytes, 246 xen_io_tlb_nslabs); 247 if (rc) { 248 if (early) 249 memblock_free(__pa(xen_io_tlb_start), 250 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 * 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 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 (((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 */ 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 == DMA_MAPPING_ERROR) 407 return DMA_MAPPING_ERROR; 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 DMA_MAPPING_ERROR; 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 */ 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 } 447 448 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr, 449 size_t size, enum dma_data_direction dir, 450 unsigned long attrs) 451 { 452 xen_unmap_single(hwdev, dev_addr, size, dir, attrs); 453 } 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); 473 474 if (target == SYNC_FOR_CPU) 475 xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir); 476 477 /* NOTE: We use dev_addr here, not paddr! */ 478 if (is_xen_swiotlb_buffer(dev_addr)) 479 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target); 480 481 if (target == SYNC_FOR_DEVICE) 482 xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir); 483 } 484 485 void 486 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr, 487 size_t size, enum dma_data_direction dir) 488 { 489 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU); 490 } 491 492 void 493 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr, 494 size_t size, enum dma_data_direction dir) 495 { 496 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE); 497 } 498 499 /* 500 * Unmap a set of streaming mode DMA translations. Again, cpu read rules 501 * concerning calls here are the same as for swiotlb_unmap_page() above. 502 */ 503 static void 504 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl, 505 int nelems, enum dma_data_direction dir, 506 unsigned long attrs) 507 { 508 struct scatterlist *sg; 509 int i; 510 511 BUG_ON(dir == DMA_NONE); 512 513 for_each_sg(sgl, sg, nelems, i) 514 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs); 515 516 } 517 518 /* 519 * Map a set of buffers described by scatterlist in streaming mode for DMA. 520 * This is the scatter-gather version of the above xen_swiotlb_map_page 521 * interface. Here the scatter gather list elements are each tagged with the 522 * appropriate dma address and length. They are obtained via 523 * sg_dma_{address,length}(SG). 524 * 525 * NOTE: An implementation may be able to use a smaller number of 526 * DMA address/length pairs than there are SG table elements. 527 * (for example via virtual mapping capabilities) 528 * The routine returns the number of addr/length pairs actually 529 * used, at most nents. 530 * 531 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the 532 * same here. 533 */ 534 static int 535 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl, 536 int nelems, enum dma_data_direction dir, 537 unsigned long attrs) 538 { 539 struct scatterlist *sg; 540 int i; 541 542 BUG_ON(dir == DMA_NONE); 543 544 for_each_sg(sgl, sg, nelems, i) { 545 phys_addr_t paddr = sg_phys(sg); 546 dma_addr_t dev_addr = xen_phys_to_bus(paddr); 547 548 if (swiotlb_force == SWIOTLB_FORCE || 549 xen_arch_need_swiotlb(hwdev, paddr, dev_addr) || 550 !dma_capable(hwdev, dev_addr, sg->length) || 551 range_straddles_page_boundary(paddr, sg->length)) { 552 phys_addr_t map = swiotlb_tbl_map_single(hwdev, 553 start_dma_addr, 554 sg_phys(sg), 555 sg->length, 556 dir, attrs); 557 if (map == DMA_MAPPING_ERROR) { 558 dev_warn(hwdev, "swiotlb buffer is full\n"); 559 /* Don't panic here, we expect map_sg users 560 to do proper error handling. */ 561 attrs |= DMA_ATTR_SKIP_CPU_SYNC; 562 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir, 563 attrs); 564 sg_dma_len(sgl) = 0; 565 return 0; 566 } 567 dev_addr = xen_phys_to_bus(map); 568 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT), 569 dev_addr, 570 map & ~PAGE_MASK, 571 sg->length, 572 dir, 573 attrs); 574 sg->dma_address = dev_addr; 575 } else { 576 /* we are not interested in the dma_addr returned by 577 * xen_dma_map_page, only in the potential cache flushes executed 578 * by the function. */ 579 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT), 580 dev_addr, 581 paddr & ~PAGE_MASK, 582 sg->length, 583 dir, 584 attrs); 585 sg->dma_address = dev_addr; 586 } 587 sg_dma_len(sg) = sg->length; 588 } 589 return nelems; 590 } 591 592 /* 593 * Make physical memory consistent for a set of streaming mode DMA translations 594 * after a transfer. 595 * 596 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules 597 * and usage. 598 */ 599 static void 600 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl, 601 int nelems, enum dma_data_direction dir, 602 enum dma_sync_target target) 603 { 604 struct scatterlist *sg; 605 int i; 606 607 for_each_sg(sgl, sg, nelems, i) 608 xen_swiotlb_sync_single(hwdev, sg->dma_address, 609 sg_dma_len(sg), dir, target); 610 } 611 612 static void 613 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg, 614 int nelems, enum dma_data_direction dir) 615 { 616 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU); 617 } 618 619 static 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 } 625 626 /* 627 * Return whether the given device DMA address mask can be supported 628 * properly. For example, if your device can only drive the low 24-bits 629 * during bus mastering, then you would pass 0x00ffffff as the mask to 630 * this function. 631 */ 632 static int 633 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask) 634 { 635 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask; 636 } 637 638 /* 639 * Create userspace mapping for the DMA-coherent memory. 640 * This function should be called with the pages from the current domain only, 641 * passing pages mapped from other domains would lead to memory corruption. 642 */ 643 static int 644 xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma, 645 void *cpu_addr, dma_addr_t dma_addr, size_t size, 646 unsigned long attrs) 647 { 648 #if defined(CONFIG_ARM) || defined(CONFIG_ARM64) 649 if (xen_get_dma_ops(dev)->mmap) 650 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr, 651 dma_addr, size, attrs); 652 #endif 653 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs); 654 } 655 656 /* 657 * This function should be called with the pages from the current domain only, 658 * passing pages mapped from other domains would lead to memory corruption. 659 */ 660 static int 661 xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt, 662 void *cpu_addr, dma_addr_t handle, size_t size, 663 unsigned long attrs) 664 { 665 #if defined(CONFIG_ARM) || defined(CONFIG_ARM64) 666 if (xen_get_dma_ops(dev)->get_sgtable) { 667 #if 0 668 /* 669 * This check verifies that the page belongs to the current domain and 670 * is not one mapped from another domain. 671 * This check is for debug only, and should not go to production build 672 */ 673 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle)); 674 BUG_ON (!page_is_ram(bfn)); 675 #endif 676 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr, 677 handle, size, attrs); 678 } 679 #endif 680 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs); 681 } 682 683 const struct dma_map_ops xen_swiotlb_dma_ops = { 684 .alloc = xen_swiotlb_alloc_coherent, 685 .free = xen_swiotlb_free_coherent, 686 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu, 687 .sync_single_for_device = xen_swiotlb_sync_single_for_device, 688 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu, 689 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device, 690 .map_sg = xen_swiotlb_map_sg_attrs, 691 .unmap_sg = xen_swiotlb_unmap_sg_attrs, 692 .map_page = xen_swiotlb_map_page, 693 .unmap_page = xen_swiotlb_unmap_page, 694 .dma_supported = xen_swiotlb_dma_supported, 695 .mmap = xen_swiotlb_dma_mmap, 696 .get_sgtable = xen_swiotlb_get_sgtable, 697 }; 698