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/export.h>
32 #include <xen/swiotlb-xen.h>
34 #include <xen/xen-ops.h>
35 #include <xen/hvc-console.h>
37 #include <asm/dma-mapping.h>
38 #include <asm/xen/page-coherent.h>
40 #include <trace/events/swiotlb.h>
42 * Used to do a quick range check in swiotlb_tbl_unmap_single and
43 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
47 static char *xen_io_tlb_start, *xen_io_tlb_end;
48 static unsigned long xen_io_tlb_nslabs;
50 * Quick lookup value of the bus address of the IOTLB.
53 static u64 start_dma_addr;
56 * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
57 * can be 32bit when dma_addr_t is 64bit leading to a loss in
58 * information if the shift is done before casting to 64bit.
60 static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
62 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
63 dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
65 dma |= paddr & ~XEN_PAGE_MASK;
70 static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
72 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
73 dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
74 phys_addr_t paddr = dma;
76 paddr |= baddr & ~XEN_PAGE_MASK;
81 static inline dma_addr_t xen_virt_to_bus(void *address)
83 return xen_phys_to_bus(virt_to_phys(address));
86 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
88 unsigned long next_bfn, xen_pfn = XEN_PFN_DOWN(p);
89 unsigned int i, nr_pages = XEN_PFN_UP(xen_offset_in_page(p) + size);
91 next_bfn = pfn_to_bfn(xen_pfn);
93 for (i = 1; i < nr_pages; i++)
94 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
100 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
102 unsigned long bfn = XEN_PFN_DOWN(dma_addr);
103 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
104 phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
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.
110 if (pfn_valid(PFN_DOWN(paddr))) {
111 return paddr >= virt_to_phys(xen_io_tlb_start) &&
112 paddr < virt_to_phys(xen_io_tlb_end);
117 static int max_dma_bits = 32;
120 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
124 dma_addr_t dma_handle;
125 phys_addr_t p = virt_to_phys(buf);
127 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
131 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
134 rc = xen_create_contiguous_region(
135 p + (i << IO_TLB_SHIFT),
136 get_order(slabs << IO_TLB_SHIFT),
137 dma_bits, &dma_handle);
138 } while (rc && dma_bits++ < max_dma_bits);
143 } while (i < nslabs);
146 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
149 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
150 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
152 xen_io_tlb_nslabs = nr_tbl;
154 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
157 enum xen_swiotlb_err {
158 XEN_SWIOTLB_UNKNOWN = 0,
163 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
166 case XEN_SWIOTLB_ENOMEM:
167 return "Cannot allocate Xen-SWIOTLB buffer\n";
168 case XEN_SWIOTLB_EFIXUP:
169 return "Failed to get contiguous memory for DMA from Xen!\n"\
170 "You either: don't have the permissions, do not have"\
171 " enough free memory under 4GB, or the hypervisor memory"\
172 " is too fragmented!";
178 int __ref xen_swiotlb_init(int verbose, bool early)
180 unsigned long bytes, order;
182 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
183 unsigned int repeat = 3;
185 xen_io_tlb_nslabs = swiotlb_nr_tbl();
187 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
188 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
191 * IO TLB memory already allocated. Just use it.
193 if (io_tlb_start != 0) {
194 xen_io_tlb_start = phys_to_virt(io_tlb_start);
199 * Get IO TLB memory from any location.
202 xen_io_tlb_start = memblock_alloc(PAGE_ALIGN(bytes),
204 if (!xen_io_tlb_start)
205 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
206 __func__, PAGE_ALIGN(bytes), PAGE_SIZE);
208 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
209 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
210 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
211 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
212 if (xen_io_tlb_start)
216 if (order != get_order(bytes)) {
217 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
218 (PAGE_SIZE << order) >> 20);
219 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
220 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
223 if (!xen_io_tlb_start) {
224 m_ret = XEN_SWIOTLB_ENOMEM;
228 * And replace that memory with pages under 4GB.
230 rc = xen_swiotlb_fixup(xen_io_tlb_start,
235 memblock_free(__pa(xen_io_tlb_start),
238 free_pages((unsigned long)xen_io_tlb_start, order);
239 xen_io_tlb_start = NULL;
241 m_ret = XEN_SWIOTLB_EFIXUP;
244 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
246 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
248 panic("Cannot allocate SWIOTLB buffer");
251 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
254 xen_io_tlb_end = xen_io_tlb_start + bytes;
256 swiotlb_set_max_segment(PAGE_SIZE);
261 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
262 (xen_io_tlb_nslabs >> 1));
263 pr_info("Lowering to %luMB\n",
264 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
267 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
269 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
271 free_pages((unsigned long)xen_io_tlb_start, order);
276 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
277 dma_addr_t *dma_handle, gfp_t flags,
281 int order = get_order(size);
282 u64 dma_mask = DMA_BIT_MASK(32);
287 * Ignore region specifiers - the kernel's ideas of
288 * pseudo-phys memory layout has nothing to do with the
289 * machine physical layout. We can't allocate highmem
290 * because we can't return a pointer to it.
292 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
294 /* Convert the size to actually allocated. */
295 size = 1UL << (order + XEN_PAGE_SHIFT);
297 /* On ARM this function returns an ioremap'ped virtual address for
298 * which virt_to_phys doesn't return the corresponding physical
299 * address. In fact on ARM virt_to_phys only works for kernel direct
300 * mapped RAM memory. Also see comment below.
302 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
307 if (hwdev && hwdev->coherent_dma_mask)
308 dma_mask = hwdev->coherent_dma_mask;
310 /* At this point dma_handle is the physical address, next we are
311 * going to set it to the machine address.
312 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
315 dev_addr = xen_phys_to_bus(phys);
316 if (((dev_addr + size - 1 <= dma_mask)) &&
317 !range_straddles_page_boundary(phys, size))
318 *dma_handle = dev_addr;
320 if (xen_create_contiguous_region(phys, order,
321 fls64(dma_mask), dma_handle) != 0) {
322 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
325 SetPageXenRemapped(virt_to_page(ret));
327 memset(ret, 0, size);
332 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
333 dma_addr_t dev_addr, unsigned long attrs)
335 int order = get_order(size);
337 u64 dma_mask = DMA_BIT_MASK(32);
339 if (hwdev && hwdev->coherent_dma_mask)
340 dma_mask = hwdev->coherent_dma_mask;
342 /* do not use virt_to_phys because on ARM it doesn't return you the
343 * physical address */
344 phys = xen_bus_to_phys(dev_addr);
346 /* Convert the size to actually allocated. */
347 size = 1UL << (order + XEN_PAGE_SHIFT);
349 if (!WARN_ON((dev_addr + size - 1 > dma_mask) ||
350 range_straddles_page_boundary(phys, size)) &&
351 TestClearPageXenRemapped(virt_to_page(vaddr)))
352 xen_destroy_contiguous_region(phys, order);
354 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
358 * Map a single buffer of the indicated size for DMA in streaming mode. The
359 * physical address to use is returned.
361 * Once the device is given the dma address, the device owns this memory until
362 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
364 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
365 unsigned long offset, size_t size,
366 enum dma_data_direction dir,
369 phys_addr_t map, phys = page_to_phys(page) + offset;
370 dma_addr_t dev_addr = xen_phys_to_bus(phys);
372 BUG_ON(dir == DMA_NONE);
374 * If the address happens to be in the device's DMA window,
375 * we can safely return the device addr and not worry about bounce
378 if (dma_capable(dev, dev_addr, size) &&
379 !range_straddles_page_boundary(phys, size) &&
380 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
381 swiotlb_force != SWIOTLB_FORCE)
385 * Oh well, have to allocate and map a bounce buffer.
387 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
389 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
391 if (map == (phys_addr_t)DMA_MAPPING_ERROR)
392 return DMA_MAPPING_ERROR;
394 dev_addr = xen_phys_to_bus(map);
397 * Ensure that the address returned is DMA'ble
399 if (unlikely(!dma_capable(dev, dev_addr, size))) {
400 swiotlb_tbl_unmap_single(dev, map, size, dir,
401 attrs | DMA_ATTR_SKIP_CPU_SYNC);
402 return DMA_MAPPING_ERROR;
405 page = pfn_to_page(map >> PAGE_SHIFT);
406 offset = map & ~PAGE_MASK;
409 * we are not interested in the dma_addr returned by xen_dma_map_page,
410 * only in the potential cache flushes executed by the function.
412 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
417 * Unmap a single streaming mode DMA translation. The dma_addr and size must
418 * match what was provided for in a previous xen_swiotlb_map_page call. All
419 * other usages are undefined.
421 * After this call, reads by the cpu to the buffer are guaranteed to see
422 * whatever the device wrote there.
424 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
425 size_t size, enum dma_data_direction dir,
428 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
430 BUG_ON(dir == DMA_NONE);
432 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
434 /* NOTE: We use dev_addr here, not paddr! */
435 if (is_xen_swiotlb_buffer(dev_addr))
436 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
439 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
440 size_t size, enum dma_data_direction dir,
443 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
447 xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
448 size_t size, enum dma_data_direction dir)
450 phys_addr_t paddr = xen_bus_to_phys(dma_addr);
452 xen_dma_sync_single_for_cpu(dev, dma_addr, size, dir);
454 if (is_xen_swiotlb_buffer(dma_addr))
455 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_CPU);
459 xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
460 size_t size, enum dma_data_direction dir)
462 phys_addr_t paddr = xen_bus_to_phys(dma_addr);
464 if (is_xen_swiotlb_buffer(dma_addr))
465 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_DEVICE);
467 xen_dma_sync_single_for_device(dev, dma_addr, size, dir);
471 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
472 * concerning calls here are the same as for swiotlb_unmap_page() above.
475 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
476 enum dma_data_direction dir, unsigned long attrs)
478 struct scatterlist *sg;
481 BUG_ON(dir == DMA_NONE);
483 for_each_sg(sgl, sg, nelems, i)
484 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
489 xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
490 enum dma_data_direction dir, unsigned long attrs)
492 struct scatterlist *sg;
495 BUG_ON(dir == DMA_NONE);
497 for_each_sg(sgl, sg, nelems, i) {
498 sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg),
499 sg->offset, sg->length, dir, attrs);
500 if (sg->dma_address == DMA_MAPPING_ERROR)
502 sg_dma_len(sg) = sg->length;
507 xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
513 xen_swiotlb_sync_sg_for_cpu(struct device *dev, struct scatterlist *sgl,
514 int nelems, enum dma_data_direction dir)
516 struct scatterlist *sg;
519 for_each_sg(sgl, sg, nelems, i) {
520 xen_swiotlb_sync_single_for_cpu(dev, sg->dma_address,
526 xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
527 int nelems, enum dma_data_direction dir)
529 struct scatterlist *sg;
532 for_each_sg(sgl, sg, nelems, i) {
533 xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
539 * Return whether the given device DMA address mask can be supported
540 * properly. For example, if your device can only drive the low 24-bits
541 * during bus mastering, then you would pass 0x00ffffff as the mask to
545 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
547 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
551 * Create userspace mapping for the DMA-coherent memory.
552 * This function should be called with the pages from the current domain only,
553 * passing pages mapped from other domains would lead to memory corruption.
556 xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
557 void *cpu_addr, dma_addr_t dma_addr, size_t size,
561 if (xen_get_dma_ops(dev)->mmap)
562 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
563 dma_addr, size, attrs);
565 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
569 * This function should be called with the pages from the current domain only,
570 * passing pages mapped from other domains would lead to memory corruption.
573 xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
574 void *cpu_addr, dma_addr_t handle, size_t size,
578 if (xen_get_dma_ops(dev)->get_sgtable) {
581 * This check verifies that the page belongs to the current domain and
582 * is not one mapped from another domain.
583 * This check is for debug only, and should not go to production build
585 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
586 BUG_ON (!page_is_ram(bfn));
588 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
589 handle, size, attrs);
592 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs);
595 const struct dma_map_ops xen_swiotlb_dma_ops = {
596 .alloc = xen_swiotlb_alloc_coherent,
597 .free = xen_swiotlb_free_coherent,
598 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
599 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
600 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
601 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
602 .map_sg = xen_swiotlb_map_sg,
603 .unmap_sg = xen_swiotlb_unmap_sg,
604 .map_page = xen_swiotlb_map_page,
605 .unmap_page = xen_swiotlb_unmap_page,
606 .dma_supported = xen_swiotlb_dma_supported,
607 .mmap = xen_swiotlb_dma_mmap,
608 .get_sgtable = xen_swiotlb_get_sgtable,