drivers/xen/swiotlb-xen.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  *  Copyright 2010
0004  *  by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
0005  *
0006  * This code provides a IOMMU for Xen PV guests with PCI passthrough.
0007  *
0008  * PV guests under Xen are running in an non-contiguous memory architecture.
0009  *
0010  * When PCI pass-through is utilized, this necessitates an IOMMU for
0011  * translating bus (DMA) to virtual and vice-versa and also providing a
0012  * mechanism to have contiguous pages for device drivers operations (say DMA
0013  * operations).
0014  *
0015  * Specifically, under Xen the Linux idea of pages is an illusion. It
0016  * assumes that pages start at zero and go up to the available memory. To
0017  * help with that, the Linux Xen MMU provides a lookup mechanism to
0018  * translate the page frame numbers (PFN) to machine frame numbers (MFN)
0019  * and vice-versa. The MFN are the "real" frame numbers. Furthermore
0020  * memory is not contiguous. Xen hypervisor stitches memory for guests
0021  * from different pools, which means there is no guarantee that PFN==MFN
0022  * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
0023  * allocated in descending order (high to low), meaning the guest might
0024  * never get any MFN's under the 4GB mark.
0025  */
0026
0027 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
0028
0029 #include <linux/memblock.h>
0030 #include <linux/dma-direct.h>
0031 #include <linux/dma-map-ops.h>
0032 #include <linux/export.h>
0033 #include <xen/swiotlb-xen.h>
0034 #include <xen/page.h>
0035 #include <xen/xen-ops.h>
0036 #include <xen/hvc-console.h>
0037
0038 #include <asm/dma-mapping.h>
0039
0040 #include <trace/events/swiotlb.h>
0041 #define MAX_DMA_BITS 32
0042
0043 /*
0044  * Quick lookup value of the bus address of the IOTLB.
0045  */
0046
0047 static inline phys_addr_t xen_phys_to_bus(struct device *dev, phys_addr_t paddr)
0048 {
0049     unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
0050     phys_addr_t baddr = (phys_addr_t)bfn << XEN_PAGE_SHIFT;
0051
0052     baddr |= paddr & ~XEN_PAGE_MASK;
0053     return baddr;
0054 }
0055
0056 static inline dma_addr_t xen_phys_to_dma(struct device *dev, phys_addr_t paddr)
0057 {
0058     return phys_to_dma(dev, xen_phys_to_bus(dev, paddr));
0059 }
0060
0061 static inline phys_addr_t xen_bus_to_phys(struct device *dev,
0062                       phys_addr_t baddr)
0063 {
0064     unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
0065     phys_addr_t paddr = (xen_pfn << XEN_PAGE_SHIFT) |
0066                 (baddr & ~XEN_PAGE_MASK);
0067
0068     return paddr;
0069 }
0070
0071 static inline phys_addr_t xen_dma_to_phys(struct device *dev,
0072                       dma_addr_t dma_addr)
0073 {
0074     return xen_bus_to_phys(dev, dma_to_phys(dev, dma_addr));
0075 }
0076
0077 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
0078 {
0079     unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p);
0080     unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size);
0081
0082     next_bfn = pfn_to_bfn(xen_pfn);
0083
0084     for (i = 1; i < nr_pages; i++)
0085         if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
0086             return 1;
0087
0088     return 0;
0089 }
0090
0091 static int is_xen_swiotlb_buffer(struct device *dev, dma_addr_t dma_addr)
0092 {
0093     unsigned long bfn = XEN_PFN_DOWN(dma_to_phys(dev, dma_addr));
0094     unsigned long xen_pfn = bfn_to_local_pfn(bfn);
0095     phys_addr_t paddr = (phys_addr_t)xen_pfn << XEN_PAGE_SHIFT;
0096
0097     /* If the address is outside our domain, it CAN
0098      * have the same virtual address as another address
0099      * in our domain. Therefore _only_ check address within our domain.
0100      */
0101     if (pfn_valid(PFN_DOWN(paddr)))
0102         return is_swiotlb_buffer(dev, paddr);
0103     return 0;
0104 }
0105
0106 #ifdef CONFIG_X86
0107 int xen_swiotlb_fixup(void *buf, unsigned long nslabs)
0108 {
0109     int rc;
0110     unsigned int order = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT);
0111     unsigned int i, dma_bits = order + PAGE_SHIFT;
0112     dma_addr_t dma_handle;
0113     phys_addr_t p = virt_to_phys(buf);
0114
0115     BUILD_BUG_ON(IO_TLB_SEGSIZE & (IO_TLB_SEGSIZE - 1));
0116     BUG_ON(nslabs % IO_TLB_SEGSIZE);
0117
0118     i = 0;
0119     do {
0120         do {
0121             rc = xen_create_contiguous_region(
0122                 p + (i << IO_TLB_SHIFT), order,
0123                 dma_bits, &dma_handle);
0124         } while (rc && dma_bits++ < MAX_DMA_BITS);
0125         if (rc)
0126             return rc;
0127
0128         i += IO_TLB_SEGSIZE;
0129     } while (i < nslabs);
0130     return 0;
0131 }
0132
0133 static void *
0134 xen_swiotlb_alloc_coherent(struct device *dev, size_t size,
0135         dma_addr_t *dma_handle, gfp_t flags, unsigned long attrs)
0136 {
0137     u64 dma_mask = dev->coherent_dma_mask;
0138     int order = get_order(size);
0139     phys_addr_t phys;
0140     void *ret;
0141
0142     /* Align the allocation to the Xen page size */
0143     size = 1UL << (order + XEN_PAGE_SHIFT);
0144
0145     ret = (void *)__get_free_pages(flags, get_order(size));
0146     if (!ret)
0147         return ret;
0148     phys = virt_to_phys(ret);
0149
0150     *dma_handle = xen_phys_to_dma(dev, phys);
0151     if (*dma_handle + size - 1 > dma_mask ||
0152         range_straddles_page_boundary(phys, size)) {
0153         if (xen_create_contiguous_region(phys, order, fls64(dma_mask),
0154                 dma_handle) != 0)
0155             goto out_free_pages;
0156         SetPageXenRemapped(virt_to_page(ret));
0157     }
0158
0159     memset(ret, 0, size);
0160     return ret;
0161
0162 out_free_pages:
0163     free_pages((unsigned long)ret, get_order(size));
0164     return NULL;
0165 }
0166
0167 static void
0168 xen_swiotlb_free_coherent(struct device *dev, size_t size, void *vaddr,
0169         dma_addr_t dma_handle, unsigned long attrs)
0170 {
0171     phys_addr_t phys = virt_to_phys(vaddr);
0172     int order = get_order(size);
0173
0174     /* Convert the size to actually allocated. */
0175     size = 1UL << (order + XEN_PAGE_SHIFT);
0176
0177     if (WARN_ON_ONCE(dma_handle + size - 1 > dev->coherent_dma_mask) ||
0178         WARN_ON_ONCE(range_straddles_page_boundary(phys, size)))
0179             return;
0180
0181     if (TestClearPageXenRemapped(virt_to_page(vaddr)))
0182         xen_destroy_contiguous_region(phys, order);
0183     free_pages((unsigned long)vaddr, get_order(size));
0184 }
0185 #endif /* CONFIG_X86 */
0186
0187 /*
0188  * Map a single buffer of the indicated size for DMA in streaming mode.  The
0189  * physical address to use is returned.
0190  *
0191  * Once the device is given the dma address, the device owns this memory until
0192  * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
0193  */
0194 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
0195                 unsigned long offset, size_t size,
0196                 enum dma_data_direction dir,
0197                 unsigned long attrs)
0198 {
0199     phys_addr_t map, phys = page_to_phys(page) + offset;
0200     dma_addr_t dev_addr = xen_phys_to_dma(dev, phys);
0201
0202     BUG_ON(dir == DMA_NONE);
0203     /*
0204      * If the address happens to be in the device's DMA window,
0205      * we can safely return the device addr and not worry about bounce
0206      * buffering it.
0207      */
0208     if (dma_capable(dev, dev_addr, size, true) &&
0209         !range_straddles_page_boundary(phys, size) &&
0210         !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
0211         !is_swiotlb_force_bounce(dev))
0212         goto done;
0213
0214     /*
0215      * Oh well, have to allocate and map a bounce buffer.
0216      */
0217     trace_swiotlb_bounced(dev, dev_addr, size);
0218
0219     map = swiotlb_tbl_map_single(dev, phys, size, size, 0, dir, attrs);
0220     if (map == (phys_addr_t)DMA_MAPPING_ERROR)
0221         return DMA_MAPPING_ERROR;
0222
0223     phys = map;
0224     dev_addr = xen_phys_to_dma(dev, map);
0225
0226     /*
0227      * Ensure that the address returned is DMA'ble
0228      */
0229     if (unlikely(!dma_capable(dev, dev_addr, size, true))) {
0230         swiotlb_tbl_unmap_single(dev, map, size, dir,
0231                 attrs | DMA_ATTR_SKIP_CPU_SYNC);
0232         return DMA_MAPPING_ERROR;
0233     }
0234
0235 done:
0236     if (!dev_is_dma_coherent(dev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
0237         if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dev_addr))))
0238             arch_sync_dma_for_device(phys, size, dir);
0239         else
0240             xen_dma_sync_for_device(dev, dev_addr, size, dir);
0241     }
0242     return dev_addr;
0243 }
0244
0245 /*
0246  * Unmap a single streaming mode DMA translation.  The dma_addr and size must
0247  * match what was provided for in a previous xen_swiotlb_map_page call.  All
0248  * other usages are undefined.
0249  *
0250  * After this call, reads by the cpu to the buffer are guaranteed to see
0251  * whatever the device wrote there.
0252  */
0253 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
0254         size_t size, enum dma_data_direction dir, unsigned long attrs)
0255 {
0256     phys_addr_t paddr = xen_dma_to_phys(hwdev, dev_addr);
0257
0258     BUG_ON(dir == DMA_NONE);
0259
0260     if (!dev_is_dma_coherent(hwdev) && !(attrs & DMA_ATTR_SKIP_CPU_SYNC)) {
0261         if (pfn_valid(PFN_DOWN(dma_to_phys(hwdev, dev_addr))))
0262             arch_sync_dma_for_cpu(paddr, size, dir);
0263         else
0264             xen_dma_sync_for_cpu(hwdev, dev_addr, size, dir);
0265     }
0266
0267     /* NOTE: We use dev_addr here, not paddr! */
0268     if (is_xen_swiotlb_buffer(hwdev, dev_addr))
0269         swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
0270 }
0271
0272 static void
0273 xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
0274         size_t size, enum dma_data_direction dir)
0275 {
0276     phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
0277
0278     if (!dev_is_dma_coherent(dev)) {
0279         if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
0280             arch_sync_dma_for_cpu(paddr, size, dir);
0281         else
0282             xen_dma_sync_for_cpu(dev, dma_addr, size, dir);
0283     }
0284
0285     if (is_xen_swiotlb_buffer(dev, dma_addr))
0286         swiotlb_sync_single_for_cpu(dev, paddr, size, dir);
0287 }
0288
0289 static void
0290 xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
0291         size_t size, enum dma_data_direction dir)
0292 {
0293     phys_addr_t paddr = xen_dma_to_phys(dev, dma_addr);
0294
0295     if (is_xen_swiotlb_buffer(dev, dma_addr))
0296         swiotlb_sync_single_for_device(dev, paddr, size, dir);
0297
0298     if (!dev_is_dma_coherent(dev)) {
0299         if (pfn_valid(PFN_DOWN(dma_to_phys(dev, dma_addr))))
0300             arch_sync_dma_for_device(paddr, size, dir);
0301         else
0302             xen_dma_sync_for_device(dev, dma_addr, size, dir);
0303     }
0304 }
0305
0306 /*
0307  * Unmap a set of streaming mode DMA translations.  Again, cpu read rules
0308  * concerning calls here are the same as for swiotlb_unmap_page() above.
0309  */
0310 static void
0311 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
0312         enum dma_data_direction dir, unsigned long attrs)
0313 {
0314     struct scatterlist *sg;
0315     int i;
0316
0317     BUG_ON(dir == DMA_NONE);
0318
0319     for_each_sg(sgl, sg, nelems, i)
0320         xen_swiotlb_unmap_page(hwdev, sg->dma_address, sg_dma_len(sg),
0321                 dir, attrs);
0322
0323 }
0324
0325 static int
0326 xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
0327         enum dma_data_direction dir, unsigned long attrs)
0328 {
0329     struct scatterlist *sg;
0330     int i;
0331
0332     BUG_ON(dir == DMA_NONE);
0333
0334     for_each_sg(sgl, sg, nelems, i) {
0335         sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg),
0336                 sg->offset, sg->length, dir, attrs);
0337         if (sg->dma_address == DMA_MAPPING_ERROR)
0338             goto out_unmap;
0339         sg_dma_len(sg) = sg->length;
0340     }
0341
0342     return nelems;
0343 out_unmap:
0344     xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
0345     sg_dma_len(sgl) = 0;
0346     return -EIO;
0347 }
0348
0349 static void
0350 xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
0351                 int nelems, enum dma_data_direction dir)
0352 {
0353     struct scatterlist *sg;
0354     int i;
0355
0356     for_each_sg(sgl, sg, nelems, i) {
0357         xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address,
0358                 sg->length, dir);
0359     }
0360 }
0361
0362 static void
0363 xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
0364                    int nelems, enum dma_data_direction dir)
0365 {
0366     struct scatterlist *sg;
0367     int i;
0368
0369     for_each_sg(sgl, sg, nelems, i) {
0370         xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
0371                 sg->length, dir);
0372     }
0373 }
0374
0375 /*
0376  * Return whether the given device DMA address mask can be supported
0377  * properly.  For example, if your device can only drive the low 24-bits
0378  * during bus mastering, then you would pass 0x00ffffff as the mask to
0379  * this function.
0380  */
0381 static int
0382 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
0383 {
0384     return xen_phys_to_dma(hwdev, io_tlb_default_mem.end - 1) <= mask;
0385 }
0386
0387 const struct dma_map_ops xen_swiotlb_dma_ops = {
0388 #ifdef CONFIG_X86
0389     .alloc = xen_swiotlb_alloc_coherent,
0390     .free = xen_swiotlb_free_coherent,
0391 #else
0392     .alloc = dma_direct_alloc,
0393     .free = dma_direct_free,
0394 #endif
0395     .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
0396     .sync_single_for_device = xen_swiotlb_sync_single_for_device,
0397     .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
0398     .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
0399     .map_sg = xen_swiotlb_map_sg,
0400     .unmap_sg = xen_swiotlb_unmap_sg,
0401     .map_page = xen_swiotlb_map_page,
0402     .unmap_page = xen_swiotlb_unmap_page,
0403     .dma_supported = xen_swiotlb_dma_supported,
0404     .mmap = dma_common_mmap,
0405     .get_sgtable = dma_common_get_sgtable,
0406     .alloc_pages = dma_common_alloc_pages,
0407     .free_pages = dma_common_free_pages,
0408 };