1 // SPDX-License-Identifier: GPL-2.0-only 2 /* 3 * Copyright 2010 4 * by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com> 5 * 6 * This code provides a IOMMU for Xen PV guests with PCI passthrough. 7 * 8 * PV guests under Xen are running in an non-contiguous memory architecture. 9 * 10 * When PCI pass-through is utilized, this necessitates an IOMMU for 11 * translating bus (DMA) to virtual and vice-versa and also providing a 12 * mechanism to have contiguous pages for device drivers operations (say DMA 13 * operations). 14 * 15 * Specifically, under Xen the Linux idea of pages is an illusion. It 16 * assumes that pages start at zero and go up to the available memory. To 17 * help with that, the Linux Xen MMU provides a lookup mechanism to 18 * translate the page frame numbers (PFN) to machine frame numbers (MFN) 19 * and vice-versa. The MFN are the "real" frame numbers. Furthermore 20 * memory is not contiguous. Xen hypervisor stitches memory for guests 21 * from different pools, which means there is no guarantee that PFN==MFN 22 * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are 23 * allocated in descending order (high to low), meaning the guest might 24 * never get any MFN's under the 4GB mark. 25 */ 26 27 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt 28 29 #include <linux/memblock.h> 30 #include <linux/dma-direct.h> 31 #include <linux/dma-map-ops.h> 32 #include <linux/export.h> 33 #include <xen/swiotlb-xen.h> 34 #include <xen/page.h> 35 #include <xen/xen-ops.h> 36 #include <xen/hvc-console.h> 37 38 #include <asm/dma-mapping.h> 39 40 #include <trace/events/swiotlb.h> 41 #define MAX_DMA_BITS 32 42 43 /* 44 * Quick lookup value of the bus address of the IOTLB. 45 */ 46 47 static inline phys_addr_t xen_phys_to_bus(struct device *dev, phys_addr_t paddr) 48 { 49 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr)); 50 phys_addr_t baddr = (phys_addr_t)bfn << XEN_PAGE_SHIFT; 51 52 baddr |= paddr & ~XEN_PAGE_MASK; 53 return baddr; 54 } 55 56 static inline dma_addr_t xen_phys_to_dma(struct device *dev, phys_addr_t paddr) 57 { 58 return phys_to_dma(dev, xen_phys_to_bus(dev, paddr)); 59 } 60 61 static inline phys_addr_t xen_bus_to_phys(struct device *dev, 62 phys_addr_t baddr) 63 { 64 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr)); 65 phys_addr_t paddr = (xen_pfn << XEN_PAGE_SHIFT) | 66 (baddr & ~XEN_PAGE_MASK); 67 68 return paddr; 69 } 70 71 static inline phys_addr_t xen_dma_to_phys(struct device *dev, 72 dma_addr_t dma_addr) 73 { 74 return xen_bus_to_phys(dev, dma_to_phys(dev, dma_addr)); 75 } 76 77 static inline bool range_requires_alignment(phys_addr_t p, size_t size) 78 { 79 phys_addr_t algn = 1ULL << (get_order(size) + PAGE_SHIFT); 80 phys_addr_t bus_addr = pfn_to_bfn(XEN_PFN_DOWN(p)) << XEN_PAGE_SHIFT; 81 82 return IS_ALIGNED(p, algn) && !IS_ALIGNED(bus_addr, algn); 83 } 84 85 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size) 86 { 87 unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p); 88 unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size); 89 90 next_bfn = pfn_to_bfn(xen_pfn); 91 92 for (i = 1; i < nr_pages; i++) 93 if (pfn_to_bfn(++xen_pfn) != ++next_bfn) 94 return 1; 95 96 return 0; 97 } 98 99 static struct io_tlb_pool *xen_swiotlb_find_pool(struct device *dev, 100 dma_addr_t dma_addr) 101 { 102 unsigned long bfn = XEN_PFN_DOWN(dma_to_phys(dev, dma_addr)); 103 unsigned long xen_pfn = bfn_to_local_pfn(bfn); 104 phys_addr_t paddr = (phys_addr_t)xen_pfn << XEN_PAGE_SHIFT; 105 106 /* If the address is outside our domain, it CAN 107 * have the same virtual address as another address 108 * in our domain. Therefore _only_ check address within our domain. 109 */ 110 if (pfn_valid(PFN_DOWN(paddr))) 111 return swiotlb_find_pool(dev, paddr); 112 return NULL; 113 } 114 115 #ifdef CONFIG_X86 116 int __init xen_swiotlb_fixup(void *buf, unsigned long nslabs) 117 { 118 int rc; 119 unsigned int order = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT); 120 unsigned int i, dma_bits = order + PAGE_SHIFT; 121 dma_addr_t dma_handle; 122 phys_addr_t p = virt_to_phys(buf); 123 124 BUILD_BUG_ON(IO_TLB_SEGSIZE & (IO_TLB_SEGSIZE - 1)); 125 BUG_ON(nslabs % IO_TLB_SEGSIZE); 126 127 i = 0; 128 do { 129 do { 130 rc = xen_create_contiguous_region( 131 p + (i << IO_TLB_SHIFT), order, 132 dma_bits, &dma_handle); 133 } while (rc && dma_bits++ < MAX_DMA_BITS); 134 if (rc) 135 return rc; 136 137 i += IO_TLB_SEGSIZE; 138 } while (i < nslabs); 139 return 0; 140 } 141 142 static void * 143 xen_swiotlb_alloc_coherent(struct device *dev, size_t size, 144 dma_addr_t *dma_handle, gfp_t flags, unsigned long attrs) 145 { 146 u64 dma_mask = dev->coherent_dma_mask; 147 int order = get_order(size); 148 phys_addr_t phys; 149 void *ret; 150 151 /* Align the allocation to the Xen page size */ 152 size = ALIGN(size, XEN_PAGE_SIZE); 153 154 ret = (void *)__get_free_pages(flags, get_order(size)); 155 if (!ret) 156 return ret; 157 phys = virt_to_phys(ret); 158 159 *dma_handle = xen_phys_to_dma(dev, phys); 160 if (*dma_handle + size - 1 > dma_mask || 161 range_straddles_page_boundary(phys, size) || 162 range_requires_alignment(phys, size)) { 163 if (xen_create_contiguous_region(phys, order, fls64(dma_mask), 164 dma_handle) != 0) 165 goto out_free_pages; 166 SetPageXenRemapped(virt_to_page(ret)); 167 } 168 169 memset(ret, 0, size); 170 return ret; 171 172 out_free_pages: 173 free_pages((unsigned long)ret, get_order(size)); 174 return NULL; 175 } 176 177 static void 178 xen_swiotlb_free_coherent(struct device *dev, size_t size, void *vaddr, 179 dma_addr_t dma_handle, unsigned long attrs) 180 { 181 phys_addr_t phys = virt_to_phys(vaddr); 182 int order = get_order(size); 183 184 /* Convert the size to actually allocated. */ 185 size = ALIGN(size, XEN_PAGE_SIZE); 186 187 if (WARN_ON_ONCE(dma_handle + size - 1 > dev->coherent_dma_mask) || 188 WARN_ON_ONCE(range_straddles_page_boundary(phys, size) || 189 range_requires_alignment(phys, size))) 190 return; 191 192 if (TestClearPageXenRemapped(virt_to_page(vaddr))) 193 xen_destroy_contiguous_region(phys, order); 194 free_pages((unsigned long)vaddr, get_order(size)); 195 } 196 #endif /* CONFIG_X86 */ 197 198 /* 199 * Map a single buffer of the indicated size for DMA in streaming mode. The 200 * physical address to use is returned. 201 * 202 * Once the device is given the dma address, the device owns this memory until 203 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed. 204 */ 205 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page, 206 unsigned long offset, size_t size, 207 enum dma_data_direction dir, 208 unsigned long attrs) 209 { 210 phys_addr_t map, phys = page_to_phys(page) + offset; 211 dma_addr_t dev_addr = xen_phys_to_dma(dev, phys); 212 213 BUG_ON(dir == DMA_NONE); 214 /* 215 * If the address happens to be in the device's DMA window, 216 * we can safely return the device addr and not worry about bounce 217 * buffering it. 218 */ 219 if (dma_capable(dev, dev_addr, size, true) && 220 !range_straddles_page_boundary(phys, size) && 221 !xen_arch_need_swiotlb(dev, phys, dev_addr) && 222 !is_swiotlb_force_bounce(dev)) 223 goto done; 224 225 /* 226 * Oh well, have to allocate and map a bounce buffer. 227 */ 228 trace_swiotlb_bounced(dev, dev_addr, size); 229 230 map = swiotlb_tbl_map_single(dev, phys, size, 0, dir, attrs); 231 if (map == (phys_addr_t)DMA_MAPPING_ERROR) 232 return DMA_MAPPING_ERROR; 233 234 phys = map; 235 dev_addr = xen_phys_to_dma(dev, map); 236 237 /* 238 * Ensure that the address returned is DMA'ble 239 */ 240 if (unlikely(!dma_capable(dev, dev_addr, size, true))) { 241 __swiotlb_tbl_unmap_single(dev, map, size, dir, 242 attrs | DMA_ATTR_SKIP_CPU_SYNC, 243 swiotlb_find_pool(dev, map)); 244 return DMA_MAPPING_ERROR; 245 } 246 247 done: 248 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) { 249 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr)))) 250 arch_sync_dma_for_device(phys, size, dir); 251 else 252 xen_dma_sync_for_device(dev, dev_addr, size, dir); 253 } 254 return dev_addr; 255 } 256 257 /* 258 * Unmap a single streaming mode DMA translation. The dma_addr and size must 259 * match what was provided for in a previous xen_swiotlb_map_page call. All 260 * other usages are undefined. 261 * 262 * After this call, reads by the cpu to the buffer are guaranteed to see 263 * whatever the device wrote there. 264 */ 265 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr, 266 size_t size, enum dma_data_direction dir, unsigned long attrs) 267 { 268 phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr); 269 struct io_tlb_pool *pool; 270 271 BUG_ON(dir == DMA_NONE); 272 273 if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) { 274 if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr)))) 275 arch_sync_dma_for_cpu(paddr, size, dir); 276 else 277 xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir); 278 } 279 280 /* NOTE: We use dev_addr here, not paddr! */ 281 pool = xen_swiotlb_find_pool(hwdev, dev_addr); 282 if (pool) 283 __swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, 284 attrs, pool); 285 } 286 287 static void 288 xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr, 289 size_t size, enum dma_data_direction dir) 290 { 291 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr); 292 struct io_tlb_pool *pool; 293 294 if (!dev_is_dma_coherent(dev)) { 295 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr)))) 296 arch_sync_dma_for_cpu(paddr, size, dir); 297 else 298 xen_dma_sync_for_cpu(dev, dma_addr, size, dir); 299 } 300 301 pool = xen_swiotlb_find_pool(dev, dma_addr); 302 if (pool) 303 __swiotlb_sync_single_for_cpu(dev, paddr, size, dir, pool); 304 } 305 306 static void 307 xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr, 308 size_t size, enum dma_data_direction dir) 309 { 310 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr); 311 struct io_tlb_pool *pool; 312 313 pool = xen_swiotlb_find_pool(dev, dma_addr); 314 if (pool) 315 __swiotlb_sync_single_for_device(dev, paddr, size, dir, pool); 316 317 if (!dev_is_dma_coherent(dev)) { 318 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr)))) 319 arch_sync_dma_for_device(paddr, size, dir); 320 else 321 xen_dma_sync_for_device(dev, dma_addr, size, dir); 322 } 323 } 324 325 /* 326 * Unmap a set of streaming mode DMA translations. Again, cpu read rules 327 * concerning calls here are the same as for swiotlb_unmap_page() above. 328 */ 329 static void 330 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems, 331 enum dma_data_direction dir, unsigned long attrs) 332 { 333 struct scatterlist *sg; 334 int i; 335 336 BUG_ON(dir == DMA_NONE); 337 338 for_each_sg(sgl, sg, nelems, i) 339 xen_swiotlb_unmap_page(hwdev, sg->dma_address, sg_dma_len(sg), 340 dir, attrs); 341 342 } 343 344 static int 345 xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems, 346 enum dma_data_direction dir, unsigned long attrs) 347 { 348 struct scatterlist *sg; 349 int i; 350 351 BUG_ON(dir == DMA_NONE); 352 353 for_each_sg(sgl, sg, nelems, i) { 354 sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg), 355 sg->offset, sg->length, dir, attrs); 356 if (sg->dma_address == DMA_MAPPING_ERROR) 357 goto out_unmap; 358 sg_dma_len(sg) = sg->length; 359 } 360 361 return nelems; 362 out_unmap: 363 xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC); 364 sg_dma_len(sgl) = 0; 365 return -EIO; 366 } 367 368 static void 369 xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl, 370 int nelems, enum dma_data_direction dir) 371 { 372 struct scatterlist *sg; 373 int i; 374 375 for_each_sg(sgl, sg, nelems, i) { 376 xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address, 377 sg->length, dir); 378 } 379 } 380 381 static void 382 xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl, 383 int nelems, enum dma_data_direction dir) 384 { 385 struct scatterlist *sg; 386 int i; 387 388 for_each_sg(sgl, sg, nelems, i) { 389 xen_swiotlb_sync_single_for_device(dev, sg->dma_address, 390 sg->length, dir); 391 } 392 } 393 394 /* 395 * Return whether the given device DMA address mask can be supported 396 * properly. For example, if your device can only drive the low 24-bits 397 * during bus mastering, then you would pass 0x00ffffff as the mask to 398 * this function. 399 */ 400 static int 401 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask) 402 { 403 return xen_phys_to_dma(hwdev, default_swiotlb_limit()) <= mask; 404 } 405 406 const struct dma_map_ops xen_swiotlb_dma_ops = { 407 #ifdef CONFIG_X86 408 .alloc = xen_swiotlb_alloc_coherent, 409 .free = xen_swiotlb_free_coherent, 410 #else 411 .alloc = dma_direct_alloc, 412 .free = dma_direct_free, 413 #endif 414 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu, 415 .sync_single_for_device = xen_swiotlb_sync_single_for_device, 416 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu, 417 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device, 418 .map_sg = xen_swiotlb_map_sg, 419 .unmap_sg = xen_swiotlb_unmap_sg, 420 .map_page = xen_swiotlb_map_page, 421 .unmap_page = xen_swiotlb_unmap_page, 422 .dma_supported = xen_swiotlb_dma_supported, 423 .mmap = dma_common_mmap, 424 .get_sgtable = dma_common_get_sgtable, 425 .alloc_pages_op = dma_common_alloc_pages, 426 .free_pages = dma_common_free_pages, 427 .max_mapping_size = swiotlb_max_mapping_size, 428 }; 429