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),
220 if (!xen_io_tlb_start)
221 panic("%s: Failed to allocate %lu bytes align=0x%lx\n",
222 __func__, PAGE_ALIGN(bytes), PAGE_SIZE);
224 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
225 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
226 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
227 xen_io_tlb_start = (void *)xen_get_swiotlb_free_pages(order);
228 if (xen_io_tlb_start)
232 if (order != get_order(bytes)) {
233 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
234 (PAGE_SIZE << order) >> 20);
235 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
236 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
239 if (!xen_io_tlb_start) {
240 m_ret = XEN_SWIOTLB_ENOMEM;
243 xen_io_tlb_end = xen_io_tlb_start + bytes;
245 * And replace that memory with pages under 4GB.
247 rc = xen_swiotlb_fixup(xen_io_tlb_start,
252 memblock_free(__pa(xen_io_tlb_start),
255 free_pages((unsigned long)xen_io_tlb_start, order);
256 xen_io_tlb_start = NULL;
258 m_ret = XEN_SWIOTLB_EFIXUP;
261 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
263 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
265 panic("Cannot allocate SWIOTLB buffer");
268 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
271 swiotlb_set_max_segment(PAGE_SIZE);
276 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
277 (xen_io_tlb_nslabs >> 1));
278 pr_info("Lowering to %luMB\n",
279 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
282 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
284 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
286 free_pages((unsigned long)xen_io_tlb_start, order);
291 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
292 dma_addr_t *dma_handle, gfp_t flags,
296 int order = get_order(size);
297 u64 dma_mask = DMA_BIT_MASK(32);
302 * Ignore region specifiers - the kernel's ideas of
303 * pseudo-phys memory layout has nothing to do with the
304 * machine physical layout. We can't allocate highmem
305 * because we can't return a pointer to it.
307 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
309 /* Convert the size to actually allocated. */
310 size = 1UL << (order + XEN_PAGE_SHIFT);
312 /* On ARM this function returns an ioremap'ped virtual address for
313 * which virt_to_phys doesn't return the corresponding physical
314 * address. In fact on ARM virt_to_phys only works for kernel direct
315 * mapped RAM memory. Also see comment below.
317 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
322 if (hwdev && hwdev->coherent_dma_mask)
323 dma_mask = hwdev->coherent_dma_mask;
325 /* At this point dma_handle is the physical address, next we are
326 * going to set it to the machine address.
327 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
330 dev_addr = xen_phys_to_bus(phys);
331 if (((dev_addr + size - 1 <= dma_mask)) &&
332 !range_straddles_page_boundary(phys, size))
333 *dma_handle = dev_addr;
335 if (xen_create_contiguous_region(phys, order,
336 fls64(dma_mask), dma_handle) != 0) {
337 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
341 memset(ret, 0, size);
346 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
347 dma_addr_t dev_addr, unsigned long attrs)
349 int order = get_order(size);
351 u64 dma_mask = DMA_BIT_MASK(32);
353 if (hwdev && hwdev->coherent_dma_mask)
354 dma_mask = hwdev->coherent_dma_mask;
356 /* do not use virt_to_phys because on ARM it doesn't return you the
357 * physical address */
358 phys = xen_bus_to_phys(dev_addr);
360 /* Convert the size to actually allocated. */
361 size = 1UL << (order + XEN_PAGE_SHIFT);
363 if (((dev_addr + size - 1 <= dma_mask)) ||
364 range_straddles_page_boundary(phys, size))
365 xen_destroy_contiguous_region(phys, order);
367 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
371 * Map a single buffer of the indicated size for DMA in streaming mode. The
372 * physical address to use is returned.
374 * Once the device is given the dma address, the device owns this memory until
375 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
377 static dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
378 unsigned long offset, size_t size,
379 enum dma_data_direction dir,
382 phys_addr_t map, phys = page_to_phys(page) + offset;
383 dma_addr_t dev_addr = xen_phys_to_bus(phys);
385 BUG_ON(dir == DMA_NONE);
387 * If the address happens to be in the device's DMA window,
388 * we can safely return the device addr and not worry about bounce
391 if (dma_capable(dev, dev_addr, size) &&
392 !range_straddles_page_boundary(phys, size) &&
393 !xen_arch_need_swiotlb(dev, phys, dev_addr) &&
394 swiotlb_force != SWIOTLB_FORCE)
398 * Oh well, have to allocate and map a bounce buffer.
400 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
402 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir,
404 if (map == DMA_MAPPING_ERROR)
405 return DMA_MAPPING_ERROR;
407 dev_addr = xen_phys_to_bus(map);
410 * Ensure that the address returned is DMA'ble
412 if (unlikely(!dma_capable(dev, dev_addr, size))) {
413 swiotlb_tbl_unmap_single(dev, map, size, dir,
414 attrs | DMA_ATTR_SKIP_CPU_SYNC);
415 return DMA_MAPPING_ERROR;
418 page = pfn_to_page(map >> PAGE_SHIFT);
419 offset = map & ~PAGE_MASK;
422 * we are not interested in the dma_addr returned by xen_dma_map_page,
423 * only in the potential cache flushes executed by the function.
425 xen_dma_map_page(dev, page, dev_addr, offset, size, dir, attrs);
430 * Unmap a single streaming mode DMA translation. The dma_addr and size must
431 * match what was provided for in a previous xen_swiotlb_map_page call. All
432 * other usages are undefined.
434 * After this call, reads by the cpu to the buffer are guaranteed to see
435 * whatever the device wrote there.
437 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
438 size_t size, enum dma_data_direction dir,
441 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
443 BUG_ON(dir == DMA_NONE);
445 xen_dma_unmap_page(hwdev, dev_addr, size, dir, attrs);
447 /* NOTE: We use dev_addr here, not paddr! */
448 if (is_xen_swiotlb_buffer(dev_addr))
449 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir, attrs);
452 static void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
453 size_t size, enum dma_data_direction dir,
456 xen_unmap_single(hwdev, dev_addr, size, dir, attrs);
460 xen_swiotlb_sync_single_for_cpu(struct device *dev, dma_addr_t dma_addr,
461 size_t size, enum dma_data_direction dir)
463 phys_addr_t paddr = xen_bus_to_phys(dma_addr);
465 xen_dma_sync_single_for_cpu(dev, dma_addr, size, dir);
467 if (is_xen_swiotlb_buffer(dma_addr))
468 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_CPU);
472 xen_swiotlb_sync_single_for_device(struct device *dev, dma_addr_t dma_addr,
473 size_t size, enum dma_data_direction dir)
475 phys_addr_t paddr = xen_bus_to_phys(dma_addr);
477 if (is_xen_swiotlb_buffer(dma_addr))
478 swiotlb_tbl_sync_single(dev, paddr, size, dir, SYNC_FOR_DEVICE);
480 xen_dma_sync_single_for_device(dev, dma_addr, size, dir);
484 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
485 * concerning calls here are the same as for swiotlb_unmap_page() above.
488 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
489 enum dma_data_direction dir, unsigned long attrs)
491 struct scatterlist *sg;
494 BUG_ON(dir == DMA_NONE);
496 for_each_sg(sgl, sg, nelems, i)
497 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir, attrs);
502 xen_swiotlb_map_sg(struct device *dev, struct scatterlist *sgl, int nelems,
503 enum dma_data_direction dir, unsigned long attrs)
505 struct scatterlist *sg;
508 BUG_ON(dir == DMA_NONE);
510 for_each_sg(sgl, sg, nelems, i) {
511 sg->dma_address = xen_swiotlb_map_page(dev, sg_page(sg),
512 sg->offset, sg->length, dir, attrs);
513 if (sg->dma_address == DMA_MAPPING_ERROR)
515 sg_dma_len(sg) = sg->length;
520 xen_swiotlb_unmap_sg(dev, sgl, i, dir, attrs | DMA_ATTR_SKIP_CPU_SYNC);
526 xen_swiotlb_sync_sg_for_cpu(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_cpu(dev, sg->dma_address,
539 xen_swiotlb_sync_sg_for_device(struct device *dev, struct scatterlist *sgl,
540 int nelems, enum dma_data_direction dir)
542 struct scatterlist *sg;
545 for_each_sg(sgl, sg, nelems, i) {
546 xen_swiotlb_sync_single_for_device(dev, sg->dma_address,
552 * Return whether the given device DMA address mask can be supported
553 * properly. For example, if your device can only drive the low 24-bits
554 * during bus mastering, then you would pass 0x00ffffff as the mask to
558 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
560 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
564 * Create userspace mapping for the DMA-coherent memory.
565 * This function should be called with the pages from the current domain only,
566 * passing pages mapped from other domains would lead to memory corruption.
569 xen_swiotlb_dma_mmap(struct device *dev, struct vm_area_struct *vma,
570 void *cpu_addr, dma_addr_t dma_addr, size_t size,
574 if (xen_get_dma_ops(dev)->mmap)
575 return xen_get_dma_ops(dev)->mmap(dev, vma, cpu_addr,
576 dma_addr, size, attrs);
578 return dma_common_mmap(dev, vma, cpu_addr, dma_addr, size, attrs);
582 * This function should be called with the pages from the current domain only,
583 * passing pages mapped from other domains would lead to memory corruption.
586 xen_swiotlb_get_sgtable(struct device *dev, struct sg_table *sgt,
587 void *cpu_addr, dma_addr_t handle, size_t size,
591 if (xen_get_dma_ops(dev)->get_sgtable) {
594 * This check verifies that the page belongs to the current domain and
595 * is not one mapped from another domain.
596 * This check is for debug only, and should not go to production build
598 unsigned long bfn = PHYS_PFN(dma_to_phys(dev, handle));
599 BUG_ON (!page_is_ram(bfn));
601 return xen_get_dma_ops(dev)->get_sgtable(dev, sgt, cpu_addr,
602 handle, size, attrs);
605 return dma_common_get_sgtable(dev, sgt, cpu_addr, handle, size, attrs);
608 const struct dma_map_ops xen_swiotlb_dma_ops = {
609 .alloc = xen_swiotlb_alloc_coherent,
610 .free = xen_swiotlb_free_coherent,
611 .sync_single_for_cpu = xen_swiotlb_sync_single_for_cpu,
612 .sync_single_for_device = xen_swiotlb_sync_single_for_device,
613 .sync_sg_for_cpu = xen_swiotlb_sync_sg_for_cpu,
614 .sync_sg_for_device = xen_swiotlb_sync_sg_for_device,
615 .map_sg = xen_swiotlb_map_sg,
616 .unmap_sg = xen_swiotlb_unmap_sg,
617 .map_page = xen_swiotlb_map_page,
618 .unmap_page = xen_swiotlb_unmap_page,
619 .dma_supported = xen_swiotlb_dma_supported,
620 .mmap = xen_swiotlb_dma_mmap,
621 .get_sgtable = xen_swiotlb_get_sgtable,