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 #define XEN_SWIOTLB_ERROR_CODE (~(dma_addr_t)0x0)
58 static char *xen_io_tlb_start, *xen_io_tlb_end;
59 static unsigned long xen_io_tlb_nslabs;
61 * Quick lookup value of the bus address of the IOTLB.
64 static u64 start_dma_addr;
67 * Both of these functions should avoid XEN_PFN_PHYS because phys_addr_t
68 * can be 32bit when dma_addr_t is 64bit leading to a loss in
69 * information if the shift is done before casting to 64bit.
71 static inline dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
73 unsigned long bfn = pfn_to_bfn(XEN_PFN_DOWN(paddr));
74 dma_addr_t dma = (dma_addr_t)bfn << XEN_PAGE_SHIFT;
76 dma |= paddr & ~XEN_PAGE_MASK;
81 static inline phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
83 unsigned long xen_pfn = bfn_to_pfn(XEN_PFN_DOWN(baddr));
84 dma_addr_t dma = (dma_addr_t)xen_pfn << XEN_PAGE_SHIFT;
85 phys_addr_t paddr = dma;
87 paddr |= baddr & ~XEN_PAGE_MASK;
92 static inline dma_addr_t xen_virt_to_bus(void *address)
94 return xen_phys_to_bus(virt_to_phys(address));
97 static int check_pages_physically_contiguous(unsigned long xen_pfn,
101 unsigned long next_bfn;
105 next_bfn = pfn_to_bfn(xen_pfn);
106 nr_pages = (offset + length + XEN_PAGE_SIZE-1) >> XEN_PAGE_SHIFT;
108 for (i = 1; i < nr_pages; i++) {
109 if (pfn_to_bfn(++xen_pfn) != ++next_bfn)
115 static inline int range_straddles_page_boundary(phys_addr_t p, size_t size)
117 unsigned long xen_pfn = XEN_PFN_DOWN(p);
118 unsigned int offset = p & ~XEN_PAGE_MASK;
120 if (offset + size <= XEN_PAGE_SIZE)
122 if (check_pages_physically_contiguous(xen_pfn, offset, size))
127 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
129 unsigned long bfn = XEN_PFN_DOWN(dma_addr);
130 unsigned long xen_pfn = bfn_to_local_pfn(bfn);
131 phys_addr_t paddr = XEN_PFN_PHYS(xen_pfn);
133 /* If the address is outside our domain, it CAN
134 * have the same virtual address as another address
135 * in our domain. Therefore _only_ check address within our domain.
137 if (pfn_valid(PFN_DOWN(paddr))) {
138 return paddr >= virt_to_phys(xen_io_tlb_start) &&
139 paddr < virt_to_phys(xen_io_tlb_end);
144 static int max_dma_bits = 32;
147 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
151 dma_addr_t dma_handle;
152 phys_addr_t p = virt_to_phys(buf);
154 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
158 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
161 rc = xen_create_contiguous_region(
162 p + (i << IO_TLB_SHIFT),
163 get_order(slabs << IO_TLB_SHIFT),
164 dma_bits, &dma_handle);
165 } while (rc && dma_bits++ < max_dma_bits);
170 } while (i < nslabs);
173 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
176 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
177 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
179 xen_io_tlb_nslabs = nr_tbl;
181 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
184 enum xen_swiotlb_err {
185 XEN_SWIOTLB_UNKNOWN = 0,
190 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
193 case XEN_SWIOTLB_ENOMEM:
194 return "Cannot allocate Xen-SWIOTLB buffer\n";
195 case XEN_SWIOTLB_EFIXUP:
196 return "Failed to get contiguous memory for DMA from Xen!\n"\
197 "You either: don't have the permissions, do not have"\
198 " enough free memory under 4GB, or the hypervisor memory"\
199 " is too fragmented!";
205 int __ref xen_swiotlb_init(int verbose, bool early)
207 unsigned long bytes, order;
209 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
210 unsigned int repeat = 3;
212 xen_io_tlb_nslabs = swiotlb_nr_tbl();
214 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
215 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
217 * Get IO TLB memory from any location.
220 xen_io_tlb_start = memblock_alloc(PAGE_ALIGN(bytes),
223 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
224 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
225 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
226 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
227 if (xen_io_tlb_start)
231 if (order != get_order(bytes)) {
232 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
233 (PAGE_SIZE << order) >> 20);
234 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
235 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
238 if (!xen_io_tlb_start) {
239 m_ret = XEN_SWIOTLB_ENOMEM;
242 xen_io_tlb_end = xen_io_tlb_start + bytes;
244 * And replace that memory with pages under 4GB.
246 rc = xen_swiotlb_fixup(xen_io_tlb_start,
251 memblock_free(__pa(xen_io_tlb_start),
254 free_pages((unsigned long)xen_io_tlb_start, order);
255 xen_io_tlb_start = NULL;
257 m_ret = XEN_SWIOTLB_EFIXUP;
260 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
262 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
264 panic("Cannot allocate SWIOTLB buffer");
267 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
270 swiotlb_set_max_segment(PAGE_SIZE);
275 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
276 (xen_io_tlb_nslabs >> 1));
277 pr_info("Lowering to %luMB\n",
278 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
281 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
283 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
285 free_pages((unsigned long)xen_io_tlb_start, order);
290 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
291 dma_addr_t *dma_handle, gfp_t flags,
295 int order = get_order(size);
296 u64 dma_mask = DMA_BIT_MASK(32);
301 * Ignore region specifiers - the kernel's ideas of
302 * pseudo-phys memory layout has nothing to do with the
303 * machine physical layout. We can't allocate highmem
304 * because we can't return a pointer to it.
306 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
308 /* Convert the size to actually allocated. */
309 size = 1UL << (order + XEN_PAGE_SHIFT);
311 /* On ARM this function returns an ioremap'ped virtual address for
312 * which virt_to_phys doesn't return the corresponding physical
313 * address. In fact on ARM virt_to_phys only works for kernel direct
314 * mapped RAM memory. Also see comment below.
316 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
321 if (hwdev && hwdev->coherent_dma_mask)
322 dma_mask = hwdev->coherent_dma_mask;
324 /* At this point dma_handle is the physical address, next we are
325 * going to set it to the machine address.
326 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
329 dev_addr = xen_phys_to_bus(phys);
330 if (((dev_addr + size - 1 <= dma_mask)) &&
331 !range_straddles_page_boundary(phys, size))
332 *dma_handle = dev_addr;
334 if (xen_create_contiguous_region(phys, order,
335 fls64(dma_mask), dma_handle) != 0) {
336 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
340 memset(ret, 0, size);
345 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
346 dma_addr_t dev_addr, unsigned long attrs)
348 int order = get_order(size);
350 u64 dma_mask = DMA_BIT_MASK(32);
352 if (hwdev && hwdev->coherent_dma_mask)
353 dma_mask = hwdev->coherent_dma_mask;
355 /* do not use virt_to_phys because on ARM it doesn't return you the
356 * physical address */
357 phys = xen_bus_to_phys(dev_addr);
359 /* Convert the size to actually allocated. */
360 size = 1UL << (order + XEN_PAGE_SHIFT);
362 if (((dev_addr + size - 1 <= dma_mask)) ||
363 range_straddles_page_boundary(phys, size))
364 xen_destroy_contiguous_region(phys, order);
366 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
370 * Map a single buffer of the indicated size for DMA in streaming mode. The
371 * physical address to use is returned.
373 * Once the device is given the dma address, the device owns this memory until
374 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
376 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
377 unsigned long offset, size_t size,
378 enum dma_data_direction dir,
381 phys_addr_t map, phys = page_to_phys(page) + offset;
382 dma_addr_t dev_addr = xen_phys_to_bus(phys);
384 BUG_ON(dir == DMA_NONE);
386 * If the address happens to be in the device's DMA window,
387 * we can safely return the device addr and not worry about bounce
390 if (dma_capable(dev, dev_addr, size) &&
391 !range_straddles_page_boundary(phys, size) &&
392 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
393 (swiotlb_force != SWIOTLB_FORCE)) {
394 /* we are not interested in the dma_addr returned by
395 * xen_dma_map_page, only in the potential cache flushes executed
396 * by the function. */
397 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
402 * Oh well, have to allocate and map a bounce buffer.
404 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
406 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
408 if (map == SWIOTLB_MAP_ERROR)
409 return XEN_SWIOTLB_ERROR_CODE;
411 dev_addr = xen_phys_to_bus(map);
412 xen_dma_map_page(dev, pfn_to_page(map >> PAGE_SHIFT),
413 dev_addr, map & ~PAGE_MASK, size, dir, attrs);
416 * Ensure that the address returned is DMA'ble
418 if (dma_capable(dev, dev_addr, size))
421 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
422 swiotlb_tbl_unmap_single(dev, map, size, dir, attrs);
424 return XEN_SWIOTLB_ERROR_CODE;
428 * Unmap a single streaming mode DMA translation. The dma_addr and size must
429 * match what was provided for in a previous xen_swiotlb_map_page call. All
430 * other usages are undefined.
432 * After this call, reads by the cpu to the buffer are guaranteed to see
433 * whatever the device wrote there.
435 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
436 size_t size, enum dma_data_direction dir,
439 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
441 BUG_ON(dir == DMA_NONE);
443 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
445 /* NOTE: We use dev_addr here, not paddr! */
446 if (is_xen_swiotlb_buffer(dev_addr)) {
447 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
451 if (dir != DMA_FROM_DEVICE)
455 * phys_to_virt doesn't work with hihgmem page but we could
456 * call dma_mark_clean() with hihgmem page here. However, we
457 * are fine since dma_mark_clean() is null on POWERPC. We can
458 * make dma_mark_clean() take a physical address if necessary.
460 dma_mark_clean(phys_to_virt(paddr), size);
463 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
464 size_t size, enum dma_data_direction dir,
467 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
471 * Make physical memory consistent for a single streaming mode DMA translation
474 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
475 * using the cpu, yet do not wish to teardown the dma mapping, you must
476 * call this function before doing so. At the next point you give the dma
477 * address back to the card, you must first perform a
478 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
481 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
482 size_t size, enum dma_data_direction dir,
483 enum dma_sync_target target)
485 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
487 BUG_ON(dir == DMA_NONE);
489 if (target == SYNC_FOR_CPU)
490 xen_dma_sync_single_for_cpu(hwdev, dev_addr, size, dir);
492 /* NOTE: We use dev_addr here, not paddr! */
493 if (is_xen_swiotlb_buffer(dev_addr))
494 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
496 if (target == SYNC_FOR_DEVICE)
497 xen_dma_sync_single_for_device(hwdev, dev_addr, size, dir);
499 if (dir != DMA_FROM_DEVICE)
502 dma_mark_clean(phys_to_virt(paddr), size);
506 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
507 size_t size, enum dma_data_direction dir)
509 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
513 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
514 size_t size, enum dma_data_direction dir)
516 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
520 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
521 * concerning calls here are the same as for swiotlb_unmap_page() above.
524 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
525 int nelems, enum dma_data_direction dir,
528 struct scatterlist *sg;
531 BUG_ON(dir == DMA_NONE);
533 for_each_sg(sgl, sg, nelems, i)
534 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
539 * Map a set of buffers described by scatterlist in streaming mode for DMA.
540 * This is the scatter-gather version of the above xen_swiotlb_map_page
541 * interface. Here the scatter gather list elements are each tagged with the
542 * appropriate dma address and length. They are obtained via
543 * sg_dma_{address,length}(SG).
545 * NOTE: An implementation may be able to use a smaller number of
546 * DMA address/length pairs than there are SG table elements.
547 * (for example via virtual mapping capabilities)
548 * The routine returns the number of addr/length pairs actually
549 * used, at most nents.
551 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
555 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
556 int nelems, enum dma_data_direction dir,
559 struct scatterlist *sg;
562 BUG_ON(dir == DMA_NONE);
564 for_each_sg(sgl, sg, nelems, i) {
565 phys_addr_t paddr = sg_phys(sg);
566 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
568 if (swiotlb_force == SWIOTLB_FORCE ||
569 xen_arch_need_swiotlb(hwdev, paddr, dev_addr) ||
570 !dma_capable(hwdev, dev_addr, sg->length) ||
571 range_straddles_page_boundary(paddr, sg->length)) {
572 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
577 if (map == SWIOTLB_MAP_ERROR) {
578 dev_warn(hwdev, "swiotlb buffer is full\n");
579 /* Don't panic here, we expect map_sg users
580 to do proper error handling. */
581 attrs |= DMA_ATTR_SKIP_CPU_SYNC;
582 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
587 dev_addr = xen_phys_to_bus(map);
588 xen_dma_map_page(hwdev, pfn_to_page(map >> PAGE_SHIFT),
594 sg->dma_address = dev_addr;
596 /* we are not interested in the dma_addr returned by
597 * xen_dma_map_page, only in the potential cache flushes executed
598 * by the function. */
599 xen_dma_map_page(hwdev, pfn_to_page(paddr >> PAGE_SHIFT),
605 sg->dma_address = dev_addr;
607 sg_dma_len(sg) = sg->length;
613 * Make physical memory consistent for a set of streaming mode DMA translations
616 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
620 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
621 int nelems, enum dma_data_direction dir,
622 enum dma_sync_target target)
624 struct scatterlist *sg;
627 for_each_sg(sgl, sg, nelems, i)
628 xen_swiotlb_sync_single(hwdev, sg->dma_address,
629 sg_dma_len(sg), dir, target);
633 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
634 int nelems, enum dma_data_direction dir)
636 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
640 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
641 int nelems, enum dma_data_direction dir)
643 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
647 * Return whether the given device DMA address mask can be supported
648 * properly. For example, if your device can only drive the low 24-bits
649 * during bus mastering, then you would pass 0x00ffffff as the mask to
653 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
655 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
659 * Create userspace mapping for the DMA-coherent memory.
660 * This function should be called with the pages from the current domain only,
661 * passing pages mapped from other domains would lead to memory corruption.
664 xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
665 void *cpu_addr, dma_addr_t dma_addr, size_t size,
668 #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
669 if (xen_get_dma_ops(dev)->mmap)
670 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
671 dma_addr, size, attrs);
673 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
677 * This function should be called with the pages from the current domain only,
678 * passing pages mapped from other domains would lead to memory corruption.
681 xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
682 void *cpu_addr, dma_addr_t handle, size_t size,
685 #if defined(CONFIG_ARM) || defined(CONFIG_ARM64)
686 if (xen_get_dma_ops(dev)->get_sgtable) {
689 * This check verifies that the page belongs to the current domain and
690 * is not one mapped from another domain.
691 * This check is for debug only, and should not go to production build
693 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
694 BUG_ON (!page_is_ram(bfn));
696 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
697 handle, size, attrs);
700 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs);
703 static int xen_swiotlb_mapping_error(struct device *dev, dma_addr_t dma_addr)
705 return dma_addr == XEN_SWIOTLB_ERROR_CODE;
708 const struct dma_map_ops xen_swiotlb_dma_ops = {
709 .alloc = xen_swiotlb_alloc_coherent,
710 .free = xen_swiotlb_free_coherent,
711 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
712 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
713 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
714 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
715 .map_sg = xen_swiotlb_map_sg_attrs,
716 .unmap_sg = xen_swiotlb_unmap_sg_attrs,
717 .map_page = xen_swiotlb_map_page,
718 .unmap_page = xen_swiotlb_unmap_page,
719 .dma_supported = xen_swiotlb_dma_supported,
720 .mmap = xen_swiotlb_dma_mmap,
721 .get_sgtable = xen_swiotlb_get_sgtable,
722 .mapping_error = xen_swiotlb_mapping_error,