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>
45 #include <asm/dma-mapping.h>
46 #include <asm/xen/page-coherent.h>
48 * Used to do a quick range check in swiotlb_tbl_unmap_single and
49 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
54 static unsigned long dma_alloc_coherent_mask(struct device *dev,
57 unsigned long dma_mask = 0;
59 dma_mask = dev->coherent_dma_mask;
61 dma_mask = (gfp & GFP_DMA) ? DMA_BIT_MASK(24) : DMA_BIT_MASK(32);
67 static char *xen_io_tlb_start, *xen_io_tlb_end;
68 static unsigned long xen_io_tlb_nslabs;
70 * Quick lookup value of the bus address of the IOTLB.
73 static u64 start_dma_addr;
75 static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
77 return phys_to_machine(XPADDR(paddr)).maddr;
80 static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
82 return machine_to_phys(XMADDR(baddr)).paddr;
85 static dma_addr_t xen_virt_to_bus(void *address)
87 return xen_phys_to_bus(virt_to_phys(address));
90 static int check_pages_physically_contiguous(unsigned long pfn,
94 unsigned long next_mfn;
98 next_mfn = pfn_to_mfn(pfn);
99 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
101 for (i = 1; i < nr_pages; i++) {
102 if (pfn_to_mfn(++pfn) != ++next_mfn)
108 static int range_straddles_page_boundary(phys_addr_t p, size_t size)
110 unsigned long pfn = PFN_DOWN(p);
111 unsigned int offset = p & ~PAGE_MASK;
113 if (offset + size <= PAGE_SIZE)
115 if (check_pages_physically_contiguous(pfn, offset, size))
120 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
122 unsigned long mfn = PFN_DOWN(dma_addr);
123 unsigned long pfn = mfn_to_local_pfn(mfn);
126 /* If the address is outside our domain, it CAN
127 * have the same virtual address as another address
128 * in our domain. Therefore _only_ check address within our domain.
130 if (pfn_valid(pfn)) {
131 paddr = PFN_PHYS(pfn);
132 return paddr >= virt_to_phys(xen_io_tlb_start) &&
133 paddr < virt_to_phys(xen_io_tlb_end);
138 static int max_dma_bits = 32;
141 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
145 dma_addr_t dma_handle;
146 phys_addr_t p = virt_to_phys(buf);
148 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
152 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
155 rc = xen_create_contiguous_region(
156 p + (i << IO_TLB_SHIFT),
157 get_order(slabs << IO_TLB_SHIFT),
158 dma_bits, &dma_handle);
159 } while (rc && dma_bits++ < max_dma_bits);
164 } while (i < nslabs);
167 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
170 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
171 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
173 xen_io_tlb_nslabs = nr_tbl;
175 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
178 enum xen_swiotlb_err {
179 XEN_SWIOTLB_UNKNOWN = 0,
184 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
187 case XEN_SWIOTLB_ENOMEM:
188 return "Cannot allocate Xen-SWIOTLB buffer\n";
189 case XEN_SWIOTLB_EFIXUP:
190 return "Failed to get contiguous memory for DMA from Xen!\n"\
191 "You either: don't have the permissions, do not have"\
192 " enough free memory under 4GB, or the hypervisor memory"\
193 " is too fragmented!";
199 int __ref xen_swiotlb_init(int verbose, bool early)
201 unsigned long bytes, order;
203 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
204 unsigned int repeat = 3;
206 xen_io_tlb_nslabs = swiotlb_nr_tbl();
208 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
209 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
211 * Get IO TLB memory from any location.
214 xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
216 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
217 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
218 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
219 xen_io_tlb_start = (void *)__get_free_pages(__GFP_NOWARN, order);
220 if (xen_io_tlb_start)
224 if (order != get_order(bytes)) {
225 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
226 (PAGE_SIZE << order) >> 20);
227 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
228 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
231 if (!xen_io_tlb_start) {
232 m_ret = XEN_SWIOTLB_ENOMEM;
235 xen_io_tlb_end = xen_io_tlb_start + bytes;
237 * And replace that memory with pages under 4GB.
239 rc = xen_swiotlb_fixup(xen_io_tlb_start,
244 free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
246 free_pages((unsigned long)xen_io_tlb_start, order);
247 xen_io_tlb_start = NULL;
249 m_ret = XEN_SWIOTLB_EFIXUP;
252 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
254 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
256 panic("Cannot allocate SWIOTLB buffer");
259 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
263 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
264 (xen_io_tlb_nslabs >> 1));
265 pr_info("Lowering to %luMB\n",
266 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
269 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
271 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
273 free_pages((unsigned long)xen_io_tlb_start, order);
277 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
278 dma_addr_t *dma_handle, gfp_t flags,
279 struct dma_attrs *attrs)
282 int order = get_order(size);
283 u64 dma_mask = DMA_BIT_MASK(32);
288 * Ignore region specifiers - the kernel's ideas of
289 * pseudo-phys memory layout has nothing to do with the
290 * machine physical layout. We can't allocate highmem
291 * because we can't return a pointer to it.
293 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
295 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
298 /* On ARM this function returns an ioremap'ped virtual address for
299 * which virt_to_phys doesn't return the corresponding physical
300 * address. In fact on ARM virt_to_phys only works for kernel direct
301 * mapped RAM memory. Also see comment below.
303 ret = xen_alloc_coherent_pages(hwdev, size, dma_handle, flags, attrs);
308 if (hwdev && hwdev->coherent_dma_mask)
309 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
311 /* At this point dma_handle is the physical address, next we are
312 * going to set it to the machine address.
313 * Do not use virt_to_phys(ret) because on ARM it doesn't correspond
316 dev_addr = xen_phys_to_bus(phys);
317 if (((dev_addr + size - 1 <= dma_mask)) &&
318 !range_straddles_page_boundary(phys, size))
319 *dma_handle = dev_addr;
321 if (xen_create_contiguous_region(phys, order,
322 fls64(dma_mask), dma_handle) != 0) {
323 xen_free_coherent_pages(hwdev, size, ret, (dma_addr_t)phys, attrs);
327 memset(ret, 0, size);
330 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
333 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
334 dma_addr_t dev_addr, struct dma_attrs *attrs)
336 int order = get_order(size);
338 u64 dma_mask = DMA_BIT_MASK(32);
340 if (dma_release_from_coherent(hwdev, order, vaddr))
343 if (hwdev && hwdev->coherent_dma_mask)
344 dma_mask = hwdev->coherent_dma_mask;
346 /* do not use virt_to_phys because on ARM it doesn't return you the
347 * physical address */
348 phys = xen_bus_to_phys(dev_addr);
350 if (((dev_addr + size - 1 > dma_mask)) ||
351 range_straddles_page_boundary(phys, size))
352 xen_destroy_contiguous_region(phys, order);
354 xen_free_coherent_pages(hwdev, size, vaddr, (dma_addr_t)phys, attrs);
356 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
360 * Map a single buffer of the indicated size for DMA in streaming mode. The
361 * physical address to use is returned.
363 * Once the device is given the dma address, the device owns this memory until
364 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
366 dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
367 unsigned long offset, size_t size,
368 enum dma_data_direction dir,
369 struct dma_attrs *attrs)
371 phys_addr_t map, phys = page_to_phys(page) + offset;
372 dma_addr_t dev_addr = xen_phys_to_bus(phys);
374 BUG_ON(dir == DMA_NONE);
376 * If the address happens to be in the device's DMA window,
377 * we can safely return the device addr and not worry about bounce
380 if (dma_capable(dev, dev_addr, size) &&
381 !range_straddles_page_boundary(phys, size) && !swiotlb_force)
385 * Oh well, have to allocate and map a bounce buffer.
387 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
388 if (map == SWIOTLB_MAP_ERROR)
389 return DMA_ERROR_CODE;
391 dev_addr = xen_phys_to_bus(map);
394 * Ensure that the address returned is DMA'ble
396 if (!dma_capable(dev, dev_addr, size)) {
397 swiotlb_tbl_unmap_single(dev, map, size, dir);
402 EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
405 * Unmap a single streaming mode DMA translation. The dma_addr and size must
406 * match what was provided for in a previous xen_swiotlb_map_page call. All
407 * other usages are undefined.
409 * After this call, reads by the cpu to the buffer are guaranteed to see
410 * whatever the device wrote there.
412 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
413 size_t size, enum dma_data_direction dir)
415 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
417 BUG_ON(dir == DMA_NONE);
419 /* NOTE: We use dev_addr here, not paddr! */
420 if (is_xen_swiotlb_buffer(dev_addr)) {
421 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
425 if (dir != DMA_FROM_DEVICE)
429 * phys_to_virt doesn't work with hihgmem page but we could
430 * call dma_mark_clean() with hihgmem page here. However, we
431 * are fine since dma_mark_clean() is null on POWERPC. We can
432 * make dma_mark_clean() take a physical address if necessary.
434 dma_mark_clean(phys_to_virt(paddr), size);
437 void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
438 size_t size, enum dma_data_direction dir,
439 struct dma_attrs *attrs)
441 xen_unmap_single(hwdev, dev_addr, size, dir);
443 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
446 * Make physical memory consistent for a single streaming mode DMA translation
449 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
450 * using the cpu, yet do not wish to teardown the dma mapping, you must
451 * call this function before doing so. At the next point you give the dma
452 * address back to the card, you must first perform a
453 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
456 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
457 size_t size, enum dma_data_direction dir,
458 enum dma_sync_target target)
460 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
462 BUG_ON(dir == DMA_NONE);
464 /* NOTE: We use dev_addr here, not paddr! */
465 if (is_xen_swiotlb_buffer(dev_addr)) {
466 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
470 if (dir != DMA_FROM_DEVICE)
473 dma_mark_clean(phys_to_virt(paddr), size);
477 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
478 size_t size, enum dma_data_direction dir)
480 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
482 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
485 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
486 size_t size, enum dma_data_direction dir)
488 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
490 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
493 * Map a set of buffers described by scatterlist in streaming mode for DMA.
494 * This is the scatter-gather version of the above xen_swiotlb_map_page
495 * interface. Here the scatter gather list elements are each tagged with the
496 * appropriate dma address and length. They are obtained via
497 * sg_dma_{address,length}(SG).
499 * NOTE: An implementation may be able to use a smaller number of
500 * DMA address/length pairs than there are SG table elements.
501 * (for example via virtual mapping capabilities)
502 * The routine returns the number of addr/length pairs actually
503 * used, at most nents.
505 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
509 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
510 int nelems, enum dma_data_direction dir,
511 struct dma_attrs *attrs)
513 struct scatterlist *sg;
516 BUG_ON(dir == DMA_NONE);
518 for_each_sg(sgl, sg, nelems, i) {
519 phys_addr_t paddr = sg_phys(sg);
520 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
523 !dma_capable(hwdev, dev_addr, sg->length) ||
524 range_straddles_page_boundary(paddr, sg->length)) {
525 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
530 if (map == SWIOTLB_MAP_ERROR) {
531 /* Don't panic here, we expect map_sg users
532 to do proper error handling. */
533 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
536 return DMA_ERROR_CODE;
538 sg->dma_address = xen_phys_to_bus(map);
540 sg->dma_address = dev_addr;
541 sg_dma_len(sg) = sg->length;
545 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
548 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
549 * concerning calls here are the same as for swiotlb_unmap_page() above.
552 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
553 int nelems, enum dma_data_direction dir,
554 struct dma_attrs *attrs)
556 struct scatterlist *sg;
559 BUG_ON(dir == DMA_NONE);
561 for_each_sg(sgl, sg, nelems, i)
562 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
565 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
568 * Make physical memory consistent for a set of streaming mode DMA translations
571 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
575 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
576 int nelems, enum dma_data_direction dir,
577 enum dma_sync_target target)
579 struct scatterlist *sg;
582 for_each_sg(sgl, sg, nelems, i)
583 xen_swiotlb_sync_single(hwdev, sg->dma_address,
584 sg_dma_len(sg), dir, target);
588 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
589 int nelems, enum dma_data_direction dir)
591 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
593 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
596 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
597 int nelems, enum dma_data_direction dir)
599 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
601 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
604 xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
608 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
611 * Return whether the given device DMA address mask can be supported
612 * properly. For example, if your device can only drive the low 24-bits
613 * during bus mastering, then you would pass 0x00ffffff as the mask to
617 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
619 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
621 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);
624 xen_swiotlb_set_dma_mask(struct device *dev, u64 dma_mask)
626 if (!dev->dma_mask || !xen_swiotlb_dma_supported(dev, dma_mask))
629 *dev->dma_mask = dma_mask;
633 EXPORT_SYMBOL_GPL(xen_swiotlb_set_dma_mask);