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 int range_straddles_page_boundary(phys_addr_t p, size_t size) 78 { 79 unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p); 80 unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size); 81 phys_addr_t algn = 1ULL << (get_order(size) + PAGE_SHIFT); 82 83 next_bfn = pfn_to_bfn(xen_pfn); 84 85 /* If buffer is physically aligned, ensure DMA alignment. */ 86 if (IS_ALIGNED(p, algn) && 87 !IS_ALIGNED((phys_addr_t)next_bfn << XEN_PAGE_SHIFT, algn)) 88 return 1; 89 90 for (i = 1; i < nr_pages; i++) 91 if (pfn_to_bfn(++xen_pfn) != ++next_bfn) 92 return 1; 93 94 return 0; 95 } 96 97 static int is_xen_swiotlb_buffer(struct device *dev, dma_addr_t dma_addr) 98 { 99 unsigned long bfn = XEN_PFN_DOWN(dma_to_phys(dev, dma_addr)); 100 unsigned long xen_pfn = bfn_to_local_pfn(bfn); 101 phys_addr_t paddr = (phys_addr_t)xen_pfn << XEN_PAGE_SHIFT; 102 103 /* If the address is outside our domain, it CAN 104 * have the same virtual address as another address 105 * in our domain. Therefore _only_ check address within our domain. 106 */ 107 if (pfn_valid(PFN_DOWN(paddr))) 108 return is_swiotlb_buffer(dev, paddr); 109 return 0; 110 } 111 112 #ifdef CONFIG_X86 113 int xen_swiotlb_fixup(void *buf, unsigned long nslabs) 114 { 115 int rc; 116 unsigned int order = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT); 117 unsigned int i, dma_bits = order + PAGE_SHIFT; 118 dma_addr_t dma_handle; 119 phys_addr_t p = virt_to_phys(buf); 120 121 BUILD_BUG_ON(IO_TLB_SEGSIZE & (IO_TLB_SEGSIZE - 1)); 122 BUG_ON(nslabs % IO_TLB_SEGSIZE); 123 124 i = 0; 125 do { 126 do { 127 rc = xen_create_contiguous_region( 128 p + (i << IO_TLB_SHIFT), order, 129 dma_bits, &dma_handle); 130 } while (rc && dma_bits++ < MAX_DMA_BITS); 131 if (rc) 132 return rc; 133 134 i += IO_TLB_SEGSIZE; 135 } while (i < nslabs); 136 return 0; 137 } 138 139 static void * 140 xen_swiotlb_alloc_coherent(struct device *dev, size_t size, 141 dma_addr_t *dma_handle, gfp_t flags, unsigned long attrs) 142 { 143 u64 dma_mask = dev->coherent_dma_mask; 144 int order = get_order(size); 145 phys_addr_t phys; 146 void *ret; 147 148 /* Align the allocation to the Xen page size */ 149 size = ALIGN(size, XEN_PAGE_SIZE); 150 151 ret = (void *)__get_free_pages(flags, get_order(size)); 152 if (!ret) 153 return ret; 154 phys = virt_to_phys(ret); 155 156 *dma_handle = xen_phys_to_dma(dev, phys); 157 if (*dma_handle + size - 1 > dma_mask || 158 range_straddles_page_boundary(phys, size)) { 159 if (xen_create_contiguous_region(phys, order, fls64(dma_mask), 160 dma_handle) != 0) 161 goto out_free_pages; 162 SetPageXenRemapped(virt_to_page(ret)); 163 } 164 165 memset(ret, 0, size); 166 return ret; 167 168 out_free_pages: 169 free_pages((unsigned long)ret, get_order(size)); 170 return NULL; 171 } 172 173 static void 174 xen_swiotlb_free_coherent(struct device *dev, size_t size, void *vaddr, 175 dma_addr_t dma_handle, unsigned long attrs) 176 { 177 phys_addr_t phys = virt_to_phys(vaddr); 178 int order = get_order(size); 179 180 /* Convert the size to actually allocated. */ 181 size = ALIGN(size, XEN_PAGE_SIZE); 182 183 if (WARN_ON_ONCE(dma_handle + size - 1 > dev->coherent_dma_mask) || 184 WARN_ON_ONCE(range_straddles_page_boundary(phys, size))) 185 return; 186 187 if (TestClearPageXenRemapped(virt_to_page(vaddr))) 188 xen_destroy_contiguous_region(phys, order); 189 free_pages((unsigned long)vaddr, get_order(size)); 190 } 191 #endif /* CONFIG_X86 */ 192 193 /* 194 * Map a single buffer of the indicated size for DMA in streaming mode. The 195 * physical address to use is returned. 196 * 197 * Once the device is given the dma address, the device owns this memory until 198 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed. 199 */ 200 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page, 201 unsigned long offset, size_t size, 202 enum dma_data_direction dir, 203 unsigned long attrs) 204 { 205 phys_addr_t map, phys = page_to_phys(page) + offset; 206 dma_addr_t dev_addr = xen_phys_to_dma(dev, phys); 207 208 BUG_ON(dir == DMA_NONE); 209 /* 210 * If the address happens to be in the device's DMA window, 211 * we can safely return the device addr and not worry about bounce 212 * buffering it. 213 */ 214 if (dma_capable(dev, dev_addr, size, true) && 215 !range_straddles_page_boundary(phys, size) && 216 !xen_arch_need_swiotlb(dev, phys, dev_addr) && 217 !is_swiotlb_force_bounce(dev)) 218 goto done; 219 220 /* 221 * Oh well, have to allocate and map a bounce buffer. 222 */ 223 trace_swiotlb_bounced(dev, dev_addr, size); 224 225 map = swiotlb_tbl_map_single(dev, phys, size, size, 0, dir, attrs); 226 if (map == (phys_addr_t)DMA_MAPPING_ERROR) 227 return DMA_MAPPING_ERROR; 228 229 phys = map; 230 dev_addr = xen_phys_to_dma(dev, map); 231 232 /* 233 * Ensure that the address returned is DMA'ble 234 */ 235 if (unlikely(!dma_capable(dev, dev_addr, size, true))) { 236 swiotlb_tbl_unmap_single(dev, map, size, dir, 237 attrs | DMA_ATTR_SKIP_CPU_SYNC); 238 return DMA_MAPPING_ERROR; 239 } 240 241 done: 242 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) { 243 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr)))) 244 arch_sync_dma_for_device(phys, size, dir); 245 else 246 xen_dma_sync_for_device(dev, dev_addr, size, dir); 247 } 248 return dev_addr; 249 } 250 251 /* 252 * Unmap a single streaming mode DMA translation. The dma_addr and size must 253 * match what was provided for in a previous xen_swiotlb_map_page call. All 254 * other usages are undefined. 255 * 256 * After this call, reads by the cpu to the buffer are guaranteed to see 257 * whatever the device wrote there. 258 */ 259 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr, 260 size_t size, enum dma_data_direction dir, unsigned long attrs) 261 { 262 phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr); 263 264 BUG_ON(dir == DMA_NONE); 265 266 if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) { 267 if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr)))) 268 arch_sync_dma_for_cpu(paddr, size, dir); 269 else 270 xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir); 271 } 272 273 /* NOTE: We use dev_addr here, not paddr! */ 274 if (is_xen_swiotlb_buffer(hwdev, dev_addr)) 275 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs); 276 } 277 278 static void 279 xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr, 280 size_t size, enum dma_data_direction dir) 281 { 282 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr); 283 284 if (!dev_is_dma_coherent(dev)) { 285 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr)))) 286 arch_sync_dma_for_cpu(paddr, size, dir); 287 else 288 xen_dma_sync_for_cpu(dev, dma_addr, size, dir); 289 } 290 291 if (is_xen_swiotlb_buffer(dev, dma_addr)) 292 swiotlb_sync_single_for_cpu(dev, paddr, size, dir); 293 } 294 295 static void 296 xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr, 297 size_t size, enum dma_data_direction dir) 298 { 299 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr); 300 301 if (is_xen_swiotlb_buffer(dev, dma_addr)) 302 swiotlb_sync_single_for_device(dev, paddr, size, dir); 303 304 if (!dev_is_dma_coherent(dev)) { 305 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr)))) 306 arch_sync_dma_for_device(paddr, size, dir); 307 else 308 xen_dma_sync_for_device(dev, dma_addr, size, dir); 309 } 310 } 311 312 /* 313 * Unmap a set of streaming mode DMA translations. Again, cpu read rules 314 * concerning calls here are the same as for swiotlb_unmap_page() above. 315 */ 316 static void 317 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems, 318 enum dma_data_direction dir, unsigned long attrs) 319 { 320 struct scatterlist *sg; 321 int i; 322 323 BUG_ON(dir == DMA_NONE); 324 325 for_each_sg(sgl, sg, nelems, i) 326 xen_swiotlb_unmap_page(hwdev, sg->dma_address, sg_dma_len(sg), 327 dir, attrs); 328 329 } 330 331 static int 332 xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems, 333 enum dma_data_direction dir, unsigned long attrs) 334 { 335 struct scatterlist *sg; 336 int i; 337 338 BUG_ON(dir == DMA_NONE); 339 340 for_each_sg(sgl, sg, nelems, i) { 341 sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg), 342 sg->offset, sg->length, dir, attrs); 343 if (sg->dma_address == DMA_MAPPING_ERROR) 344 goto out_unmap; 345 sg_dma_len(sg) = sg->length; 346 } 347 348 return nelems; 349 out_unmap: 350 xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC); 351 sg_dma_len(sgl) = 0; 352 return -EIO; 353 } 354 355 static void 356 xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl, 357 int nelems, enum dma_data_direction dir) 358 { 359 struct scatterlist *sg; 360 int i; 361 362 for_each_sg(sgl, sg, nelems, i) { 363 xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address, 364 sg->length, dir); 365 } 366 } 367 368 static void 369 xen_swiotlb_sync_sg_for_device(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_device(dev, sg->dma_address, 377 sg->length, dir); 378 } 379 } 380 381 /* 382 * Return whether the given device DMA address mask can be supported 383 * properly. For example, if your device can only drive the low 24-bits 384 * during bus mastering, then you would pass 0x00ffffff as the mask to 385 * this function. 386 */ 387 static int 388 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask) 389 { 390 return xen_phys_to_dma(hwdev, default_swiotlb_limit()) <= mask; 391 } 392 393 const struct dma_map_ops xen_swiotlb_dma_ops = { 394 #ifdef CONFIG_X86 395 .alloc = xen_swiotlb_alloc_coherent, 396 .free = xen_swiotlb_free_coherent, 397 #else 398 .alloc = dma_direct_alloc, 399 .free = dma_direct_free, 400 #endif 401 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu, 402 .sync_single_for_device = xen_swiotlb_sync_single_for_device, 403 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu, 404 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device, 405 .map_sg = xen_swiotlb_map_sg, 406 .unmap_sg = xen_swiotlb_unmap_sg, 407 .map_page = xen_swiotlb_map_page, 408 .unmap_page = xen_swiotlb_unmap_page, 409 .dma_supported = xen_swiotlb_dma_supported, 410 .mmap = dma_common_mmap, 411 .get_sgtable = dma_common_get_sgtable, 412 .alloc_pages = dma_common_alloc_pages, 413 .free_pages = dma_common_free_pages, 414 .max_mapping_size = swiotlb_max_mapping_size, 415 }; 416