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 #include <linux/bootmem.h>
37 #include <linux/dma-mapping.h>
38 #include <xen/swiotlb-xen.h>
40 #include <xen/xen-ops.h>
41 #include <xen/hvc-console.h>
43 * Used to do a quick range check in swiotlb_tbl_unmap_single and
44 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
48 static char *xen_io_tlb_start, *xen_io_tlb_end;
49 static unsigned long xen_io_tlb_nslabs;
51 * Quick lookup value of the bus address of the IOTLB.
56 static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
58 return phys_to_machine(XPADDR(paddr)).maddr;
61 static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
63 return machine_to_phys(XMADDR(baddr)).paddr;
66 static dma_addr_t xen_virt_to_bus(void *address)
68 return xen_phys_to_bus(virt_to_phys(address));
71 static int check_pages_physically_contiguous(unsigned long pfn,
75 unsigned long next_mfn;
79 next_mfn = pfn_to_mfn(pfn);
80 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
82 for (i = 1; i < nr_pages; i++) {
83 if (pfn_to_mfn(++pfn) != ++next_mfn)
89 static int range_straddles_page_boundary(phys_addr_t p, size_t size)
91 unsigned long pfn = PFN_DOWN(p);
92 unsigned int offset = p & ~PAGE_MASK;
94 if (offset + size <= PAGE_SIZE)
96 if (check_pages_physically_contiguous(pfn, offset, size))
101 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
103 unsigned long mfn = PFN_DOWN(dma_addr);
104 unsigned long pfn = mfn_to_local_pfn(mfn);
107 /* If the address is outside our domain, it CAN
108 * have the same virtual address as another address
109 * in our domain. Therefore _only_ check address within our domain.
111 if (pfn_valid(pfn)) {
112 paddr = PFN_PHYS(pfn);
113 return paddr >= virt_to_phys(xen_io_tlb_start) &&
114 paddr < virt_to_phys(xen_io_tlb_end);
119 static int max_dma_bits = 32;
122 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
127 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
131 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
134 rc = xen_create_contiguous_region(
135 (unsigned long)buf + (i << IO_TLB_SHIFT),
136 get_order(slabs << IO_TLB_SHIFT),
138 } while (rc && dma_bits++ < max_dma_bits);
143 } while (i < nslabs);
147 void __init xen_swiotlb_init(int verbose)
151 unsigned long nr_tbl;
153 unsigned int repeat = 3;
155 nr_tbl = swioltb_nr_tbl();
157 xen_io_tlb_nslabs = nr_tbl;
159 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
160 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
163 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
166 * Get IO TLB memory from any location.
168 xen_io_tlb_start = alloc_bootmem(bytes);
169 if (!xen_io_tlb_start) {
170 m = "Cannot allocate Xen-SWIOTLB buffer!\n";
173 xen_io_tlb_end = xen_io_tlb_start + bytes;
175 * And replace that memory with pages under 4GB.
177 rc = xen_swiotlb_fixup(xen_io_tlb_start,
181 free_bootmem(__pa(xen_io_tlb_start), bytes);
182 m = "Failed to get contiguous memory for DMA from Xen!\n"\
183 "You either: don't have the permissions, do not have"\
184 " enough free memory under 4GB, or the hypervisor memory"\
185 "is too fragmented!";
188 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
189 swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs, verbose);
194 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
195 (xen_io_tlb_nslabs >> 1));
196 printk(KERN_INFO "Xen-SWIOTLB: Lowering to %luMB\n",
197 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
200 xen_raw_printk("%s (rc:%d)", m, rc);
201 panic("%s (rc:%d)", m, rc);
205 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
206 dma_addr_t *dma_handle, gfp_t flags)
209 int order = get_order(size);
210 u64 dma_mask = DMA_BIT_MASK(32);
211 unsigned long vstart;
216 * Ignore region specifiers - the kernel's ideas of
217 * pseudo-phys memory layout has nothing to do with the
218 * machine physical layout. We can't allocate highmem
219 * because we can't return a pointer to it.
221 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
223 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
226 vstart = __get_free_pages(flags, order);
227 ret = (void *)vstart;
232 if (hwdev && hwdev->coherent_dma_mask)
233 dma_mask = hwdev->coherent_dma_mask;
235 phys = virt_to_phys(ret);
236 dev_addr = xen_phys_to_bus(phys);
237 if (((dev_addr + size - 1 <= dma_mask)) &&
238 !range_straddles_page_boundary(phys, size))
239 *dma_handle = dev_addr;
241 if (xen_create_contiguous_region(vstart, order,
242 fls64(dma_mask)) != 0) {
243 free_pages(vstart, order);
246 *dma_handle = virt_to_machine(ret).maddr;
248 memset(ret, 0, size);
251 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
254 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
257 int order = get_order(size);
259 u64 dma_mask = DMA_BIT_MASK(32);
261 if (dma_release_from_coherent(hwdev, order, vaddr))
264 if (hwdev && hwdev->coherent_dma_mask)
265 dma_mask = hwdev->coherent_dma_mask;
267 phys = virt_to_phys(vaddr);
269 if (((dev_addr + size - 1 > dma_mask)) ||
270 range_straddles_page_boundary(phys, size))
271 xen_destroy_contiguous_region((unsigned long)vaddr, order);
273 free_pages((unsigned long)vaddr, order);
275 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
279 * Map a single buffer of the indicated size for DMA in streaming mode. The
280 * physical address to use is returned.
282 * Once the device is given the dma address, the device owns this memory until
283 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
285 dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
286 unsigned long offset, size_t size,
287 enum dma_data_direction dir,
288 struct dma_attrs *attrs)
290 phys_addr_t phys = page_to_phys(page) + offset;
291 dma_addr_t dev_addr = xen_phys_to_bus(phys);
294 BUG_ON(dir == DMA_NONE);
296 * If the address happens to be in the device's DMA window,
297 * we can safely return the device addr and not worry about bounce
300 if (dma_capable(dev, dev_addr, size) &&
301 !range_straddles_page_boundary(phys, size) && !swiotlb_force)
305 * Oh well, have to allocate and map a bounce buffer.
307 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
309 return DMA_ERROR_CODE;
311 dev_addr = xen_virt_to_bus(map);
314 * Ensure that the address returned is DMA'ble
316 if (!dma_capable(dev, dev_addr, size)) {
317 swiotlb_tbl_unmap_single(dev, map, size, dir);
322 EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
325 * Unmap a single streaming mode DMA translation. The dma_addr and size must
326 * match what was provided for in a previous xen_swiotlb_map_page call. All
327 * other usages are undefined.
329 * After this call, reads by the cpu to the buffer are guaranteed to see
330 * whatever the device wrote there.
332 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
333 size_t size, enum dma_data_direction dir)
335 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
337 BUG_ON(dir == DMA_NONE);
339 /* NOTE: We use dev_addr here, not paddr! */
340 if (is_xen_swiotlb_buffer(dev_addr)) {
341 swiotlb_tbl_unmap_single(hwdev, phys_to_virt(paddr), size, dir);
345 if (dir != DMA_FROM_DEVICE)
349 * phys_to_virt doesn't work with hihgmem page but we could
350 * call dma_mark_clean() with hihgmem page here. However, we
351 * are fine since dma_mark_clean() is null on POWERPC. We can
352 * make dma_mark_clean() take a physical address if necessary.
354 dma_mark_clean(phys_to_virt(paddr), size);
357 void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
358 size_t size, enum dma_data_direction dir,
359 struct dma_attrs *attrs)
361 xen_unmap_single(hwdev, dev_addr, size, dir);
363 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
366 * Make physical memory consistent for a single streaming mode DMA translation
369 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
370 * using the cpu, yet do not wish to teardown the dma mapping, you must
371 * call this function before doing so. At the next point you give the dma
372 * address back to the card, you must first perform a
373 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
376 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
377 size_t size, enum dma_data_direction dir,
378 enum dma_sync_target target)
380 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
382 BUG_ON(dir == DMA_NONE);
384 /* NOTE: We use dev_addr here, not paddr! */
385 if (is_xen_swiotlb_buffer(dev_addr)) {
386 swiotlb_tbl_sync_single(hwdev, phys_to_virt(paddr), size, dir,
391 if (dir != DMA_FROM_DEVICE)
394 dma_mark_clean(phys_to_virt(paddr), size);
398 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
399 size_t size, enum dma_data_direction dir)
401 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
403 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
406 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
407 size_t size, enum dma_data_direction dir)
409 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
411 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
414 * Map a set of buffers described by scatterlist in streaming mode for DMA.
415 * This is the scatter-gather version of the above xen_swiotlb_map_page
416 * interface. Here the scatter gather list elements are each tagged with the
417 * appropriate dma address and length. They are obtained via
418 * sg_dma_{address,length}(SG).
420 * NOTE: An implementation may be able to use a smaller number of
421 * DMA address/length pairs than there are SG table elements.
422 * (for example via virtual mapping capabilities)
423 * The routine returns the number of addr/length pairs actually
424 * used, at most nents.
426 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
430 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
431 int nelems, enum dma_data_direction dir,
432 struct dma_attrs *attrs)
434 struct scatterlist *sg;
437 BUG_ON(dir == DMA_NONE);
439 for_each_sg(sgl, sg, nelems, i) {
440 phys_addr_t paddr = sg_phys(sg);
441 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
444 !dma_capable(hwdev, dev_addr, sg->length) ||
445 range_straddles_page_boundary(paddr, sg->length)) {
446 void *map = swiotlb_tbl_map_single(hwdev,
451 /* Don't panic here, we expect map_sg users
452 to do proper error handling. */
453 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
455 sgl[0].dma_length = 0;
456 return DMA_ERROR_CODE;
458 sg->dma_address = xen_virt_to_bus(map);
460 sg->dma_address = dev_addr;
461 sg->dma_length = sg->length;
465 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
468 xen_swiotlb_map_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
469 enum dma_data_direction dir)
471 return xen_swiotlb_map_sg_attrs(hwdev, sgl, nelems, dir, NULL);
473 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg);
476 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
477 * concerning calls here are the same as for swiotlb_unmap_page() above.
480 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
481 int nelems, enum dma_data_direction dir,
482 struct dma_attrs *attrs)
484 struct scatterlist *sg;
487 BUG_ON(dir == DMA_NONE);
489 for_each_sg(sgl, sg, nelems, i)
490 xen_unmap_single(hwdev, sg->dma_address, sg->dma_length, dir);
493 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
496 xen_swiotlb_unmap_sg(struct device *hwdev, struct scatterlist *sgl, int nelems,
497 enum dma_data_direction dir)
499 return xen_swiotlb_unmap_sg_attrs(hwdev, sgl, nelems, dir, NULL);
501 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg);
504 * Make physical memory consistent for a set of streaming mode DMA translations
507 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
511 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
512 int nelems, enum dma_data_direction dir,
513 enum dma_sync_target target)
515 struct scatterlist *sg;
518 for_each_sg(sgl, sg, nelems, i)
519 xen_swiotlb_sync_single(hwdev, sg->dma_address,
520 sg->dma_length, dir, target);
524 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
525 int nelems, enum dma_data_direction dir)
527 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
529 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
532 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
533 int nelems, enum dma_data_direction dir)
535 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
537 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
540 xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
544 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
547 * Return whether the given device DMA address mask can be supported
548 * properly. For example, if your device can only drive the low 24-bits
549 * during bus mastering, then you would pass 0x00ffffff as the mask to
553 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
555 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
557 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);