1 // SPDX-License-Identifier: GPL-2.0-only
4 * by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
6 * This code provides a IOMMU for Xen PV guests with PCI passthrough.
8 * PV guests under Xen are running in an non-contiguous memory architecture.
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
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.
27 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
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>
35 #include <xen/xen-ops.h>
36 #include <xen/hvc-console.h>
38 #include <asm/dma-mapping.h>
40 #include <trace/events/swiotlb.h>
41 #define MAX_DMA_BITS 32
44 * Quick lookup value of the bus address of the IOTLB.
47 static inline phys_addr_t xen_phys_to_bus(struct device *dev, phys_addr_t paddr)
49 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
50 phys_addr_t baddr = (phys_addr_t)bfn << XEN_PAGE_SHIFT;
52 baddr |= paddr & ~XEN_PAGE_MASK;
56 static inline dma_addr_t xen_phys_to_dma(struct device *dev, phys_addr_t paddr)
58 return phys_to_dma(dev, xen_phys_to_bus(dev, paddr));
61 static inline phys_addr_t xen_bus_to_phys(struct device *dev,
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);
71 static inline phys_addr_t xen_dma_to_phys(struct device *dev,
74 return xen_bus_to_phys(dev, dma_to_phys(dev, dma_addr));
77 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
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);
82 next_bfn = pfn_to_bfn(xen_pfn);
84 for (i = 1; i < nr_pages; i++)
85 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
91 static int is_xen_swiotlb_buffer(struct device *dev, dma_addr_t dma_addr)
93 unsigned long bfn = XEN_PFN_DOWN(dma_to_phys(dev, dma_addr));
94 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
95 phys_addr_t paddr = (phys_addr_t)xen_pfn << XEN_PAGE_SHIFT;
97 /* If the address is outside our domain, it CAN
98 * have the same virtual address as another address
99 * in our domain. Therefore _only_ check address within our domain.
101 if (pfn_valid(PFN_DOWN(paddr)))
102 return is_swiotlb_buffer(dev, paddr);
107 int xen_swiotlb_fixup(void *buf, unsigned long nslabs)
110 unsigned int order = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT);
111 unsigned int i, dma_bits = order + PAGE_SHIFT;
112 dma_addr_t dma_handle;
113 phys_addr_t p = virt_to_phys(buf);
115 BUILD_BUG_ON(IO_TLB_SEGSIZE & (IO_TLB_SEGSIZE - 1));
116 BUG_ON(nslabs % IO_TLB_SEGSIZE);
121 rc = xen_create_contiguous_region(
122 p + (i << IO_TLB_SHIFT), order,
123 dma_bits, &dma_handle);
124 } while (rc && dma_bits++ < MAX_DMA_BITS);
129 } while (i < nslabs);
134 xen_swiotlb_alloc_coherent(struct device *dev, size_t size,
135 dma_addr_t *dma_handle, gfp_t flags, unsigned long attrs)
137 u64 dma_mask = dev->coherent_dma_mask;
138 int order = get_order(size);
142 /* Align the allocation to the Xen page size */
143 size = 1UL << (order + XEN_PAGE_SHIFT);
145 ret = (void *)__get_free_pages(flags, get_order(size));
148 phys = virt_to_phys(ret);
150 *dma_handle = xen_phys_to_dma(dev, phys);
151 if (*dma_handle + size - 1 > dma_mask ||
152 range_straddles_page_boundary(phys, size)) {
153 if (xen_create_contiguous_region(phys, order, fls64(dma_mask),
156 SetPageXenRemapped(virt_to_page(ret));
159 memset(ret, 0, size);
163 free_pages((unsigned long)ret, get_order(size));
168 xen_swiotlb_free_coherent(struct device *dev, size_t size, void *vaddr,
169 dma_addr_t dma_handle, unsigned long attrs)
171 phys_addr_t phys = virt_to_phys(vaddr);
172 int order = get_order(size);
174 /* Convert the size to actually allocated. */
175 size = 1UL << (order + XEN_PAGE_SHIFT);
177 if (WARN_ON_ONCE(dma_handle + size - 1 > dev->coherent_dma_mask) ||
178 WARN_ON_ONCE(range_straddles_page_boundary(phys, size)))
181 if (TestClearPageXenRemapped(virt_to_page(vaddr)))
182 xen_destroy_contiguous_region(phys, order);
183 free_pages((unsigned long)vaddr, get_order(size));
185 #endif /* CONFIG_X86 */
188 * Map a single buffer of the indicated size for DMA in streaming mode. The
189 * physical address to use is returned.
191 * Once the device is given the dma address, the device owns this memory until
192 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
194 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
195 unsigned long offset, size_t size,
196 enum dma_data_direction dir,
199 phys_addr_t map, phys = page_to_phys(page) + offset;
200 dma_addr_t dev_addr = xen_phys_to_dma(dev, phys);
202 BUG_ON(dir == DMA_NONE);
204 * If the address happens to be in the device's DMA window,
205 * we can safely return the device addr and not worry about bounce
208 if (dma_capable(dev, dev_addr, size, true) &&
209 !range_straddles_page_boundary(phys, size) &&
210 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
211 !is_swiotlb_force_bounce(dev))
215 * Oh well, have to allocate and map a bounce buffer.
217 trace_swiotlb_bounced(dev, dev_addr, size);
219 map = swiotlb_tbl_map_single(dev, phys, size, size, 0, dir, attrs);
220 if (map == (phys_addr_t)DMA_MAPPING_ERROR)
221 return DMA_MAPPING_ERROR;
224 dev_addr = xen_phys_to_dma(dev, map);
227 * Ensure that the address returned is DMA'ble
229 if (unlikely(!dma_capable(dev, dev_addr, size, true))) {
230 swiotlb_tbl_unmap_single(dev, map, size, dir,
231 attrs | DMA_ATTR_SKIP_CPU_SYNC);
232 return DMA_MAPPING_ERROR;
236 if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
237 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr))))
238 arch_sync_dma_for_device(phys, size, dir);
240 xen_dma_sync_for_device(dev, dev_addr, size, dir);
246 * Unmap a single streaming mode DMA translation. The dma_addr and size must
247 * match what was provided for in a previous xen_swiotlb_map_page call. All
248 * other usages are undefined.
250 * After this call, reads by the cpu to the buffer are guaranteed to see
251 * whatever the device wrote there.
253 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
254 size_t size, enum dma_data_direction dir, unsigned long attrs)
256 phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr);
258 BUG_ON(dir == DMA_NONE);
260 if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
261 if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr))))
262 arch_sync_dma_for_cpu(paddr, size, dir);
264 xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir);
267 /* NOTE: We use dev_addr here, not paddr! */
268 if (is_xen_swiotlb_buffer(hwdev, dev_addr))
269 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
273 xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
274 size_t size, enum dma_data_direction dir)
276 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
278 if (!dev_is_dma_coherent(dev)) {
279 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
280 arch_sync_dma_for_cpu(paddr, size, dir);
282 xen_dma_sync_for_cpu(dev, dma_addr, size, dir);
285 if (is_xen_swiotlb_buffer(dev, dma_addr))
286 swiotlb_sync_single_for_cpu(dev, paddr, size, dir);
290 xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
291 size_t size, enum dma_data_direction dir)
293 phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
295 if (is_xen_swiotlb_buffer(dev, dma_addr))
296 swiotlb_sync_single_for_device(dev, paddr, size, dir);
298 if (!dev_is_dma_coherent(dev)) {
299 if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
300 arch_sync_dma_for_device(paddr, size, dir);
302 xen_dma_sync_for_device(dev, dma_addr, size, dir);
307 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
308 * concerning calls here are the same as for swiotlb_unmap_page() above.
311 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
312 enum dma_data_direction dir, unsigned long attrs)
314 struct scatterlist *sg;
317 BUG_ON(dir == DMA_NONE);
319 for_each_sg(sgl, sg, nelems, i)
320 xen_swiotlb_unmap_page(hwdev, sg->dma_address, sg_dma_len(sg),
326 xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
327 enum dma_data_direction dir, unsigned long attrs)
329 struct scatterlist *sg;
332 BUG_ON(dir == DMA_NONE);
334 for_each_sg(sgl, sg, nelems, i) {
335 sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg),
336 sg->offset, sg->length, dir, attrs);
337 if (sg->dma_address == DMA_MAPPING_ERROR)
339 sg_dma_len(sg) = sg->length;
344 xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
350 xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
351 int nelems, enum dma_data_direction dir)
353 struct scatterlist *sg;
356 for_each_sg(sgl, sg, nelems, i) {
357 xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address,
363 xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
364 int nelems, enum dma_data_direction dir)
366 struct scatterlist *sg;
369 for_each_sg(sgl, sg, nelems, i) {
370 xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
376 * Return whether the given device DMA address mask can be supported
377 * properly. For example, if your device can only drive the low 24-bits
378 * during bus mastering, then you would pass 0x00ffffff as the mask to
382 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
384 return xen_phys_to_dma(hwdev, io_tlb_default_mem.end - 1) <= mask;
387 const struct dma_map_ops xen_swiotlb_dma_ops = {
389 .alloc = xen_swiotlb_alloc_coherent,
390 .free = xen_swiotlb_free_coherent,
392 .alloc = dma_direct_alloc,
393 .free = dma_direct_free,
395 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
396 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
397 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
398 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
399 .map_sg = xen_swiotlb_map_sg,
400 .unmap_sg = xen_swiotlb_unmap_sg,
401 .map_page = xen_swiotlb_map_page,
402 .unmap_page = xen_swiotlb_unmap_page,
403 .dma_supported = xen_swiotlb_dma_supported,
404 .mmap = dma_common_mmap,
405 .get_sgtable = dma_common_get_sgtable,
406 .alloc_pages = dma_common_alloc_pages,
407 .free_pages = dma_common_free_pages,