3 * by Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
5 * This code provides a IOMMU for Xen PV guests with PCI passthrough.
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License v2.0 as published by
9 * the Free Software Foundation
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * PV guests under Xen are running in an non-contiguous memory architecture.
18 * When PCI pass-through is utilized, this necessitates an IOMMU for
19 * translating bus (DMA) to virtual and vice-versa and also providing a
20 * mechanism to have contiguous pages for device drivers operations (say DMA
23 * Specifically, under Xen the Linux idea of pages is an illusion. It
24 * assumes that pages start at zero and go up to the available memory. To
25 * help with that, the Linux Xen MMU provides a lookup mechanism to
26 * translate the page frame numbers (PFN) to machine frame numbers (MFN)
27 * and vice-versa. The MFN are the "real" frame numbers. Furthermore
28 * memory is not contiguous. Xen hypervisor stitches memory for guests
29 * from different pools, which means there is no guarantee that PFN==MFN
30 * and PFN+1==MFN+1. Lastly with Xen 4.0, pages (in debug mode) are
31 * allocated in descending order (high to low), meaning the guest might
32 * never get any MFN's under the 4GB mark.
36 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
38 #include <linux/memblock.h>
39 #include <linux/dma-direct.h>
40 #include <linux/export.h>
41 #include <xen/swiotlb-xen.h>
43 #include <xen/xen-ops.h>
44 #include <xen/hvc-console.h>
46 #include <asm/dma-mapping.h>
47 #include <asm/xen/page-coherent.h>
49 #include <trace/events/swiotlb.h>
51 * Used to do a quick range check in swiotlb_tbl_unmap_single and
52 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
56 static char *xen_io_tlb_start, *xen_io_tlb_end;
57 static unsigned long xen_io_tlb_nslabs;
59 * Quick lookup value of the bus address of the IOTLB.
62 static u64 start_dma_addr;
65 * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
66 * can be 32bit when dma_addr_t is 64bit leading to a loss in
67 * information if the shift is done before casting to 64bit.
69 static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
71 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
72 dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
74 dma |= paddr & ~XEN_PAGE_MASK;
79 static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
81 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
82 dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
83 phys_addr_t paddr = dma;
85 paddr |= baddr & ~XEN_PAGE_MASK;
90 static inline dma_addr_t xen_virt_to_bus(void *address)
92 return xen_phys_to_bus(virt_to_phys(address));
95 static int check_pages_physically_contiguous(unsigned long xen_pfn,
99 unsigned long next_bfn;
103 next_bfn = pfn_to_bfn(xen_pfn);
104 nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT;
106 for (i = 1; i < nr_pages; i++) {
107 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
113 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
115 unsigned long xen_pfn = XEN_PFN_DOWN(p);
116 unsigned int offset = p & ~XEN_PAGE_MASK;
118 if (offset + size <= XEN_PAGE_SIZE)
120 if (check_pages_physically_contiguous(xen_pfn, offset, size))
125 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
127 unsigned long bfn = XEN_PFN_DOWN(dma_addr);
128 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
129 phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
131 /* If the address is outside our domain, it CAN
132 * have the same virtual address as another address
133 * in our domain. Therefore _only_ check address within our domain.
135 if (pfn_valid(PFN_DOWN(paddr))) {
136 return paddr >= virt_to_phys(xen_io_tlb_start) &&
137 paddr < virt_to_phys(xen_io_tlb_end);
142 static int max_dma_bits = 32;
145 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
149 dma_addr_t dma_handle;
150 phys_addr_t p = virt_to_phys(buf);
152 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
156 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
159 rc = xen_create_contiguous_region(
160 p + (i << IO_TLB_SHIFT),
161 get_order(slabs << IO_TLB_SHIFT),
162 dma_bits, &dma_handle);
163 } while (rc && dma_bits++ < max_dma_bits);
168 } while (i < nslabs);
171 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
174 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
175 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
177 xen_io_tlb_nslabs = nr_tbl;
179 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
182 enum xen_swiotlb_err {
183 XEN_SWIOTLB_UNKNOWN = 0,
188 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
191 case XEN_SWIOTLB_ENOMEM:
192 return "Cannot allocate Xen-SWIOTLB buffer\n";
193 case XEN_SWIOTLB_EFIXUP:
194 return "Failed to get contiguous memory for DMA from Xen!\n"\
195 "You either: don't have the permissions, do not have"\
196 " enough free memory under 4GB, or the hypervisor memory"\
197 " is too fragmented!";
203 int __ref xen_swiotlb_init(int verbose, bool early)
205 unsigned long bytes, order;
207 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
208 unsigned int repeat = 3;
210 xen_io_tlb_nslabs = swiotlb_nr_tbl();
212 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
213 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
215 * Get IO TLB memory from any location.
218 xen_io_tlb_start = memblock_alloc(PAGE_ALIGN(bytes),
221 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
222 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
223 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
224 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
225 if (xen_io_tlb_start)
229 if (order != get_order(bytes)) {
230 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
231 (PAGE_SIZE << order) >> 20);
232 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
233 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
236 if (!xen_io_tlb_start) {
237 m_ret = XEN_SWIOTLB_ENOMEM;
240 xen_io_tlb_end = xen_io_tlb_start + bytes;
242 * And replace that memory with pages under 4GB.
244 rc = xen_swiotlb_fixup(xen_io_tlb_start,
249 memblock_free(__pa(xen_io_tlb_start),
252 free_pages((unsigned long)xen_io_tlb_start, order);
253 xen_io_tlb_start = NULL;
255 m_ret = XEN_SWIOTLB_EFIXUP;
258 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
260 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
262 panic("Cannot allocate SWIOTLB buffer");
265 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
268 swiotlb_set_max_segment(PAGE_SIZE);
273 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
274 (xen_io_tlb_nslabs >> 1));
275 pr_info("Lowering to %luMB\n",
276 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
279 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
281 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
283 free_pages((unsigned long)xen_io_tlb_start, order);
288 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
289 dma_addr_t *dma_handle, gfp_t flags,
293 int order = get_order(size);
294 u64 dma_mask = DMA_BIT_MASK(32);
299 * Ignore region specifiers - the kernel's ideas of
300 * pseudo-phys memory layout has nothing to do with the
301 * machine physical layout. We can't allocate highmem
302 * because we can't return a pointer to it.
304 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
306 /* Convert the size to actually allocated. */
307 size = 1UL << (order + XEN_PAGE_SHIFT);
309 /* On ARM this function returns an ioremap'ped virtual address for
310 * which virt_to_phys doesn't return the corresponding physical
311 * address. In fact on ARM virt_to_phys only works for kernel direct
312 * mapped RAM memory. Also see comment below.
314 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
319 if (hwdev && hwdev->coherent_dma_mask)
320 dma_mask = hwdev->coherent_dma_mask;
322 /* At this point dma_handle is the physical address, next we are
323 * going to set it to the machine address.
324 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
327 dev_addr = xen_phys_to_bus(phys);
328 if (((dev_addr + size - 1 <= dma_mask)) &&
329 !range_straddles_page_boundary(phys, size))
330 *dma_handle = dev_addr;
332 if (xen_create_contiguous_region(phys, order,
333 fls64(dma_mask), dma_handle) != 0) {
334 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
338 memset(ret, 0, size);
343 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
344 dma_addr_t dev_addr, unsigned long attrs)
346 int order = get_order(size);
348 u64 dma_mask = DMA_BIT_MASK(32);
350 if (hwdev && hwdev->coherent_dma_mask)
351 dma_mask = hwdev->coherent_dma_mask;
353 /* do not use virt_to_phys because on ARM it doesn't return you the
354 * physical address */
355 phys = xen_bus_to_phys(dev_addr);
357 /* Convert the size to actually allocated. */
358 size = 1UL << (order + XEN_PAGE_SHIFT);
360 if (((dev_addr + size - 1 <= dma_mask)) ||
361 range_straddles_page_boundary(phys, size))
362 xen_destroy_contiguous_region(phys, order);
364 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
368 * Map a single buffer of the indicated size for DMA in streaming mode. The
369 * physical address to use is returned.
371 * Once the device is given the dma address, the device owns this memory until
372 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
374 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
375 unsigned long offset, size_t size,
376 enum dma_data_direction dir,
379 phys_addr_t map, phys = page_to_phys(page) + offset;
380 dma_addr_t dev_addr = xen_phys_to_bus(phys);
382 BUG_ON(dir == DMA_NONE);
384 * If the address happens to be in the device's DMA window,
385 * we can safely return the device addr and not worry about bounce
388 if (dma_capable(dev, dev_addr, size) &&
389 !range_straddles_page_boundary(phys, size) &&
390 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
391 (swiotlb_force != SWIOTLB_FORCE)) {
392 /* we are not interested in the dma_addr returned by
393 * xen_dma_map_page, only in the potential cache flushes executed
394 * by the function. */
395 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
400 * Oh well, have to allocate and map a bounce buffer.
402 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
404 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
406 if (map == DMA_MAPPING_ERROR)
407 return DMA_MAPPING_ERROR;
409 dev_addr = xen_phys_to_bus(map);
410 xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
411 dev_addr, map & ~PAGE_MASK, size, dir, attrs);
414 * Ensure that the address returned is DMA'ble
416 if (dma_capable(dev, dev_addr, size))
419 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
420 swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
422 return DMA_MAPPING_ERROR;
426 * Unmap a single streaming mode DMA translation. The dma_addr and size must
427 * match what was provided for in a previous xen_swiotlb_map_page call. All
428 * other usages are undefined.
430 * After this call, reads by the cpu to the buffer are guaranteed to see
431 * whatever the device wrote there.
433 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
434 size_t size, enum dma_data_direction dir,
437 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
439 BUG_ON(dir == DMA_NONE);
441 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
443 /* NOTE: We use dev_addr here, not paddr! */
444 if (is_xen_swiotlb_buffer(dev_addr))
445 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
448 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
449 size_t size, enum dma_data_direction dir,
452 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
456 * Make physical memory consistent for a single streaming mode DMA translation
459 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
460 * using the cpu, yet do not wish to teardown the dma mapping, you must
461 * call this function before doing so. At the next point you give the dma
462 * address back to the card, you must first perform a
463 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
466 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
467 size_t size, enum dma_data_direction dir,
468 enum dma_sync_target target)
470 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
472 BUG_ON(dir == DMA_NONE);
474 if (target == SYNC_FOR_CPU)
475 xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
477 /* NOTE: We use dev_addr here, not paddr! */
478 if (is_xen_swiotlb_buffer(dev_addr))
479 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
481 if (target == SYNC_FOR_DEVICE)
482 xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
486 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
487 size_t size, enum dma_data_direction dir)
489 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
493 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
494 size_t size, enum dma_data_direction dir)
496 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
500 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
501 * concerning calls here are the same as for swiotlb_unmap_page() above.
504 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
505 enum dma_data_direction dir, unsigned long attrs)
507 struct scatterlist *sg;
510 BUG_ON(dir == DMA_NONE);
512 for_each_sg(sgl, sg, nelems, i)
513 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
518 * Map a set of buffers described by scatterlist in streaming mode for DMA.
519 * This is the scatter-gather version of the above xen_swiotlb_map_page
520 * interface. Here the scatter gather list elements are each tagged with the
521 * appropriate dma address and length. They are obtained via
522 * sg_dma_{address,length}(SG).
524 * NOTE: An implementation may be able to use a smaller number of
525 * DMA address/length pairs than there are SG table elements.
526 * (for example via virtual mapping capabilities)
527 * The routine returns the number of addr/length pairs actually
528 * used, at most nents.
530 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
534 xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
535 enum dma_data_direction dir, unsigned long attrs)
537 struct scatterlist *sg;
540 BUG_ON(dir == DMA_NONE);
542 for_each_sg(sgl, sg, nelems, i) {
543 phys_addr_t paddr = sg_phys(sg);
544 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
546 if (swiotlb_force == SWIOTLB_FORCE ||
547 xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
548 !dma_capable(hwdev, dev_addr, sg->length) ||
549 range_straddles_page_boundary(paddr, sg->length)) {
550 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
555 if (map == DMA_MAPPING_ERROR) {
556 dev_warn(hwdev, "swiotlb buffer is full\n");
557 /* Don't panic here, we expect map_sg users
558 to do proper error handling. */
559 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
560 xen_swiotlb_unmap_sg(hwdev, sgl, i, dir, attrs);
564 dev_addr = xen_phys_to_bus(map);
565 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
571 sg->dma_address = dev_addr;
573 /* we are not interested in the dma_addr returned by
574 * xen_dma_map_page, only in the potential cache flushes executed
575 * by the function. */
576 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
582 sg->dma_address = dev_addr;
584 sg_dma_len(sg) = sg->length;
590 * Make physical memory consistent for a set of streaming mode DMA translations
593 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
597 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
598 int nelems, enum dma_data_direction dir,
599 enum dma_sync_target target)
601 struct scatterlist *sg;
604 for_each_sg(sgl, sg, nelems, i)
605 xen_swiotlb_sync_single(hwdev, sg->dma_address,
606 sg_dma_len(sg), dir, target);
610 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
611 int nelems, enum dma_data_direction dir)
613 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
617 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
618 int nelems, enum dma_data_direction dir)
620 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
624 * Return whether the given device DMA address mask can be supported
625 * properly. For example, if your device can only drive the low 24-bits
626 * during bus mastering, then you would pass 0x00ffffff as the mask to
630 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
632 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
636 * Create userspace mapping for the DMA-coherent memory.
637 * This function should be called with the pages from the current domain only,
638 * passing pages mapped from other domains would lead to memory corruption.
641 xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
642 void *cpu_addr, dma_addr_t dma_addr, size_t size,
646 if (xen_get_dma_ops(dev)->mmap)
647 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
648 dma_addr, size, attrs);
650 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
654 * This function should be called with the pages from the current domain only,
655 * passing pages mapped from other domains would lead to memory corruption.
658 xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
659 void *cpu_addr, dma_addr_t handle, size_t size,
663 if (xen_get_dma_ops(dev)->get_sgtable) {
666 * This check verifies that the page belongs to the current domain and
667 * is not one mapped from another domain.
668 * This check is for debug only, and should not go to production build
670 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
671 BUG_ON (!page_is_ram(bfn));
673 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
674 handle, size, attrs);
677 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs);
680 const struct dma_map_ops xen_swiotlb_dma_ops = {
681 .alloc = xen_swiotlb_alloc_coherent,
682 .free = xen_swiotlb_free_coherent,
683 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
684 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
685 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
686 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
687 .map_sg = xen_swiotlb_map_sg,
688 .unmap_sg = xen_swiotlb_unmap_sg,
689 .map_page = xen_swiotlb_map_page,
690 .unmap_page = xen_swiotlb_unmap_page,
691 .dma_supported = xen_swiotlb_dma_supported,
692 .mmap = xen_swiotlb_dma_mmap,
693 .get_sgtable = xen_swiotlb_get_sgtable,