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/bootmem.h>
39 #include <linux/dma-mapping.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
57 static unsigned long dma_alloc_coherent_mask(struct device *dev,
60 unsigned long dma_mask = 0;
62 dma_mask = dev->coherent_dma_mask;
64 dma_mask = (gfp & GFP_DMA) ? DMA_BIT_MASK(24) : DMA_BIT_MASK(32);
70 static char *xen_io_tlb_start, *xen_io_tlb_end;
71 static unsigned long xen_io_tlb_nslabs;
73 * Quick lookup value of the bus address of the IOTLB.
76 static u64 start_dma_addr;
79 * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
80 * can be 32bit when dma_addr_t is 64bit leading to a loss in
81 * information if the shift is done before casting to 64bit.
83 static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
85 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
86 dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
88 dma |= paddr & ~XEN_PAGE_MASK;
93 static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
95 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
96 dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
97 phys_addr_t paddr = dma;
99 paddr |= baddr & ~XEN_PAGE_MASK;
104 static inline dma_addr_t xen_virt_to_bus(void *address)
106 return xen_phys_to_bus(virt_to_phys(address));
109 static int check_pages_physically_contiguous(unsigned long xen_pfn,
113 unsigned long next_bfn;
117 next_bfn = pfn_to_bfn(xen_pfn);
118 nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT;
120 for (i = 1; i < nr_pages; i++) {
121 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
127 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
129 unsigned long xen_pfn = XEN_PFN_DOWN(p);
130 unsigned int offset = p & ~XEN_PAGE_MASK;
132 if (offset + size <= XEN_PAGE_SIZE)
134 if (check_pages_physically_contiguous(xen_pfn, offset, size))
139 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
141 unsigned long bfn = XEN_PFN_DOWN(dma_addr);
142 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
143 phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
145 /* If the address is outside our domain, it CAN
146 * have the same virtual address as another address
147 * in our domain. Therefore _only_ check address within our domain.
149 if (pfn_valid(PFN_DOWN(paddr))) {
150 return paddr >= virt_to_phys(xen_io_tlb_start) &&
151 paddr < virt_to_phys(xen_io_tlb_end);
156 static int max_dma_bits = 32;
159 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
163 dma_addr_t dma_handle;
164 phys_addr_t p = virt_to_phys(buf);
166 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
170 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
173 rc = xen_create_contiguous_region(
174 p + (i << IO_TLB_SHIFT),
175 get_order(slabs << IO_TLB_SHIFT),
176 dma_bits, &dma_handle);
177 } while (rc && dma_bits++ < max_dma_bits);
182 } while (i < nslabs);
185 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
188 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
189 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
191 xen_io_tlb_nslabs = nr_tbl;
193 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
196 enum xen_swiotlb_err {
197 XEN_SWIOTLB_UNKNOWN = 0,
202 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
205 case XEN_SWIOTLB_ENOMEM:
206 return "Cannot allocate Xen-SWIOTLB buffer\n";
207 case XEN_SWIOTLB_EFIXUP:
208 return "Failed to get contiguous memory for DMA from Xen!\n"\
209 "You either: don't have the permissions, do not have"\
210 " enough free memory under 4GB, or the hypervisor memory"\
211 " is too fragmented!";
217 int __ref xen_swiotlb_init(int verbose, bool early)
219 unsigned long bytes, order;
221 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
222 unsigned int repeat = 3;
224 xen_io_tlb_nslabs = swiotlb_nr_tbl();
226 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
227 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
229 * Get IO TLB memory from any location.
232 xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
234 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
235 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
236 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
237 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
238 if (xen_io_tlb_start)
242 if (order != get_order(bytes)) {
243 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
244 (PAGE_SIZE << order) >> 20);
245 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
246 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
249 if (!xen_io_tlb_start) {
250 m_ret = XEN_SWIOTLB_ENOMEM;
253 xen_io_tlb_end = xen_io_tlb_start + bytes;
255 * And replace that memory with pages under 4GB.
257 rc = xen_swiotlb_fixup(xen_io_tlb_start,
262 free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
264 free_pages((unsigned long)xen_io_tlb_start, order);
265 xen_io_tlb_start = NULL;
267 m_ret = XEN_SWIOTLB_EFIXUP;
270 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
272 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
274 panic("Cannot allocate SWIOTLB buffer");
277 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
280 swiotlb_set_max_segment(PAGE_SIZE);
285 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
286 (xen_io_tlb_nslabs >> 1));
287 pr_info("Lowering to %luMB\n",
288 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
291 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
293 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
295 free_pages((unsigned long)xen_io_tlb_start, order);
300 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
301 dma_addr_t *dma_handle, gfp_t flags,
305 int order = get_order(size);
306 u64 dma_mask = DMA_BIT_MASK(32);
311 * Ignore region specifiers - the kernel's ideas of
312 * pseudo-phys memory layout has nothing to do with the
313 * machine physical layout. We can't allocate highmem
314 * because we can't return a pointer to it.
316 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
318 /* On ARM this function returns an ioremap'ped virtual address for
319 * which virt_to_phys doesn't return the corresponding physical
320 * address. In fact on ARM virt_to_phys only works for kernel direct
321 * mapped RAM memory. Also see comment below.
323 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
328 if (hwdev && hwdev->coherent_dma_mask)
329 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
331 /* At this point dma_handle is the physical address, next we are
332 * going to set it to the machine address.
333 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
336 dev_addr = xen_phys_to_bus(phys);
337 if (((dev_addr + size - 1 <= dma_mask)) &&
338 !range_straddles_page_boundary(phys, size))
339 *dma_handle = dev_addr;
341 if (xen_create_contiguous_region(phys, order,
342 fls64(dma_mask), dma_handle) != 0) {
343 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
347 memset(ret, 0, size);
352 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
353 dma_addr_t dev_addr, unsigned long attrs)
355 int order = get_order(size);
357 u64 dma_mask = DMA_BIT_MASK(32);
359 if (hwdev && hwdev->coherent_dma_mask)
360 dma_mask = hwdev->coherent_dma_mask;
362 /* do not use virt_to_phys because on ARM it doesn't return you the
363 * physical address */
364 phys = xen_bus_to_phys(dev_addr);
366 if (((dev_addr + size - 1 > dma_mask)) ||
367 range_straddles_page_boundary(phys, size))
368 xen_destroy_contiguous_region(phys, order);
370 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
374 * Map a single buffer of the indicated size for DMA in streaming mode. The
375 * physical address to use is returned.
377 * Once the device is given the dma address, the device owns this memory until
378 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
380 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
381 unsigned long offset, size_t size,
382 enum dma_data_direction dir,
385 phys_addr_t map, phys = page_to_phys(page) + offset;
386 dma_addr_t dev_addr = xen_phys_to_bus(phys);
388 BUG_ON(dir == DMA_NONE);
390 * If the address happens to be in the device's DMA window,
391 * we can safely return the device addr and not worry about bounce
394 if (dma_capable(dev, dev_addr, size) &&
395 !range_straddles_page_boundary(phys, size) &&
396 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
397 (swiotlb_force != SWIOTLB_FORCE)) {
398 /* we are not interested in the dma_addr returned by
399 * xen_dma_map_page, only in the potential cache flushes executed
400 * by the function. */
401 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
406 * Oh well, have to allocate and map a bounce buffer.
408 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
410 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
412 if (map == SWIOTLB_MAP_ERROR)
413 return DMA_ERROR_CODE;
415 dev_addr = xen_phys_to_bus(map);
416 xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
417 dev_addr, map & ~PAGE_MASK, size, dir, attrs);
420 * Ensure that the address returned is DMA'ble
422 if (dma_capable(dev, dev_addr, size))
425 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
426 swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
428 return DMA_ERROR_CODE;
432 * Unmap a single streaming mode DMA translation. The dma_addr and size must
433 * match what was provided for in a previous xen_swiotlb_map_page call. All
434 * other usages are undefined.
436 * After this call, reads by the cpu to the buffer are guaranteed to see
437 * whatever the device wrote there.
439 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
440 size_t size, enum dma_data_direction dir,
443 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
445 BUG_ON(dir == DMA_NONE);
447 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
449 /* NOTE: We use dev_addr here, not paddr! */
450 if (is_xen_swiotlb_buffer(dev_addr)) {
451 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
455 if (dir != DMA_FROM_DEVICE)
459 * phys_to_virt doesn't work with hihgmem page but we could
460 * call dma_mark_clean() with hihgmem page here. However, we
461 * are fine since dma_mark_clean() is null on POWERPC. We can
462 * make dma_mark_clean() take a physical address if necessary.
464 dma_mark_clean(phys_to_virt(paddr), size);
467 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
468 size_t size, enum dma_data_direction dir,
471 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
475 * Make physical memory consistent for a single streaming mode DMA translation
478 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
479 * using the cpu, yet do not wish to teardown the dma mapping, you must
480 * call this function before doing so. At the next point you give the dma
481 * address back to the card, you must first perform a
482 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
485 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
486 size_t size, enum dma_data_direction dir,
487 enum dma_sync_target target)
489 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
491 BUG_ON(dir == DMA_NONE);
493 if (target == SYNC_FOR_CPU)
494 xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
496 /* NOTE: We use dev_addr here, not paddr! */
497 if (is_xen_swiotlb_buffer(dev_addr))
498 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
500 if (target == SYNC_FOR_DEVICE)
501 xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
503 if (dir != DMA_FROM_DEVICE)
506 dma_mark_clean(phys_to_virt(paddr), size);
510 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
511 size_t size, enum dma_data_direction dir)
513 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
517 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
518 size_t size, enum dma_data_direction dir)
520 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
524 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
525 * concerning calls here are the same as for swiotlb_unmap_page() above.
528 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
529 int nelems, enum dma_data_direction dir,
532 struct scatterlist *sg;
535 BUG_ON(dir == DMA_NONE);
537 for_each_sg(sgl, sg, nelems, i)
538 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
543 * Map a set of buffers described by scatterlist in streaming mode for DMA.
544 * This is the scatter-gather version of the above xen_swiotlb_map_page
545 * interface. Here the scatter gather list elements are each tagged with the
546 * appropriate dma address and length. They are obtained via
547 * sg_dma_{address,length}(SG).
549 * NOTE: An implementation may be able to use a smaller number of
550 * DMA address/length pairs than there are SG table elements.
551 * (for example via virtual mapping capabilities)
552 * The routine returns the number of addr/length pairs actually
553 * used, at most nents.
555 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
559 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
560 int nelems, enum dma_data_direction dir,
563 struct scatterlist *sg;
566 BUG_ON(dir == DMA_NONE);
568 for_each_sg(sgl, sg, nelems, i) {
569 phys_addr_t paddr = sg_phys(sg);
570 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
572 if (swiotlb_force == SWIOTLB_FORCE ||
573 xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
574 !dma_capable(hwdev, dev_addr, sg->length) ||
575 range_straddles_page_boundary(paddr, sg->length)) {
576 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
581 if (map == SWIOTLB_MAP_ERROR) {
582 dev_warn(hwdev, "swiotlb buffer is full\n");
583 /* Don't panic here, we expect map_sg users
584 to do proper error handling. */
585 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
586 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
591 dev_addr = xen_phys_to_bus(map);
592 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
598 sg->dma_address = dev_addr;
600 /* we are not interested in the dma_addr returned by
601 * xen_dma_map_page, only in the potential cache flushes executed
602 * by the function. */
603 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
609 sg->dma_address = dev_addr;
611 sg_dma_len(sg) = sg->length;
617 * Make physical memory consistent for a set of streaming mode DMA translations
620 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
624 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
625 int nelems, enum dma_data_direction dir,
626 enum dma_sync_target target)
628 struct scatterlist *sg;
631 for_each_sg(sgl, sg, nelems, i)
632 xen_swiotlb_sync_single(hwdev, sg->dma_address,
633 sg_dma_len(sg), dir, target);
637 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
638 int nelems, enum dma_data_direction dir)
640 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
644 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
645 int nelems, enum dma_data_direction dir)
647 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
651 * Return whether the given device DMA address mask can be supported
652 * properly. For example, if your device can only drive the low 24-bits
653 * during bus mastering, then you would pass 0x00ffffff as the mask to
657 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
659 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
663 xen_swiotlb_set_dma_mask(struct device *dev, u64 dma_mask)
665 if (!dev->dma_mask || !xen_swiotlb_dma_supported(dev, dma_mask))
668 *dev->dma_mask = dma_mask;
674 * Create userspace mapping for the DMA-coherent memory.
675 * This function should be called with the pages from the current domain only,
676 * passing pages mapped from other domains would lead to memory corruption.
679 xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
680 void *cpu_addr, dma_addr_t dma_addr, size_t size,
683 #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
684 if (xen_get_dma_ops(dev)->mmap)
685 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
686 dma_addr, size, attrs);
688 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size);
692 * This function should be called with the pages from the current domain only,
693 * passing pages mapped from other domains would lead to memory corruption.
696 xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
697 void *cpu_addr, dma_addr_t handle, size_t size,
700 #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
701 if (xen_get_dma_ops(dev)->get_sgtable) {
704 * This check verifies that the page belongs to the current domain and
705 * is not one mapped from another domain.
706 * This check is for debug only, and should not go to production build
708 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
709 BUG_ON (!page_is_ram(bfn));
711 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
712 handle, size, attrs);
715 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size);
718 const struct dma_map_ops xen_swiotlb_dma_ops = {
719 .alloc = xen_swiotlb_alloc_coherent,
720 .free = xen_swiotlb_free_coherent,
721 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
722 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
723 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
724 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
725 .map_sg = xen_swiotlb_map_sg_attrs,
726 .unmap_sg = xen_swiotlb_unmap_sg_attrs,
727 .map_page = xen_swiotlb_map_page,
728 .unmap_page = xen_swiotlb_unmap_page,
729 .dma_supported = xen_swiotlb_dma_supported,
730 .set_dma_mask = xen_swiotlb_set_dma_mask,
731 .mmap = xen_swiotlb_dma_mmap,
732 .get_sgtable = xen_swiotlb_get_sgtable,