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 #define CREATE_TRACE_POINTS
47 #include <trace/events/swiotlb.h>
49 * Used to do a quick range check in swiotlb_tbl_unmap_single and
50 * swiotlb_tbl_sync_single_*, to see if the memory was in fact allocated by this
54 static char *xen_io_tlb_start, *xen_io_tlb_end;
55 static unsigned long xen_io_tlb_nslabs;
57 * Quick lookup value of the bus address of the IOTLB.
60 static u64 start_dma_addr;
62 static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
64 return phys_to_machine(XPADDR(paddr)).maddr;
67 static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
69 return machine_to_phys(XMADDR(baddr)).paddr;
72 static dma_addr_t xen_virt_to_bus(void *address)
74 return xen_phys_to_bus(virt_to_phys(address));
77 static int check_pages_physically_contiguous(unsigned long pfn,
81 unsigned long next_mfn;
85 next_mfn = pfn_to_mfn(pfn);
86 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
88 for (i = 1; i < nr_pages; i++) {
89 if (pfn_to_mfn(++pfn) != ++next_mfn)
95 static int range_straddles_page_boundary(phys_addr_t p, size_t size)
97 unsigned long pfn = PFN_DOWN(p);
98 unsigned int offset = p & ~PAGE_MASK;
100 if (offset + size <= PAGE_SIZE)
102 if (check_pages_physically_contiguous(pfn, offset, size))
107 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
109 unsigned long mfn = PFN_DOWN(dma_addr);
110 unsigned long pfn = mfn_to_local_pfn(mfn);
113 /* If the address is outside our domain, it CAN
114 * have the same virtual address as another address
115 * in our domain. Therefore _only_ check address within our domain.
117 if (pfn_valid(pfn)) {
118 paddr = PFN_PHYS(pfn);
119 return paddr >= virt_to_phys(xen_io_tlb_start) &&
120 paddr < virt_to_phys(xen_io_tlb_end);
125 static int max_dma_bits = 32;
128 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
133 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
137 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
140 rc = xen_create_contiguous_region(
141 (unsigned long)buf + (i << IO_TLB_SHIFT),
142 get_order(slabs << IO_TLB_SHIFT),
144 } while (rc && dma_bits++ < max_dma_bits);
149 } while (i < nslabs);
152 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
155 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
156 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
158 xen_io_tlb_nslabs = nr_tbl;
160 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
163 enum xen_swiotlb_err {
164 XEN_SWIOTLB_UNKNOWN = 0,
169 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
172 case XEN_SWIOTLB_ENOMEM:
173 return "Cannot allocate Xen-SWIOTLB buffer\n";
174 case XEN_SWIOTLB_EFIXUP:
175 return "Failed to get contiguous memory for DMA from Xen!\n"\
176 "You either: don't have the permissions, do not have"\
177 " enough free memory under 4GB, or the hypervisor memory"\
178 " is too fragmented!";
184 int __ref xen_swiotlb_init(int verbose, bool early)
186 unsigned long bytes, order;
188 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
189 unsigned int repeat = 3;
191 xen_io_tlb_nslabs = swiotlb_nr_tbl();
193 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
194 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
196 * Get IO TLB memory from any location.
199 xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
201 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
202 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
203 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
204 xen_io_tlb_start = (void *)__get_free_pages(__GFP_NOWARN, order);
205 if (xen_io_tlb_start)
209 if (order != get_order(bytes)) {
210 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
211 (PAGE_SIZE << order) >> 20);
212 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
213 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
216 if (!xen_io_tlb_start) {
217 m_ret = XEN_SWIOTLB_ENOMEM;
220 xen_io_tlb_end = xen_io_tlb_start + bytes;
222 * And replace that memory with pages under 4GB.
224 rc = xen_swiotlb_fixup(xen_io_tlb_start,
229 free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
231 free_pages((unsigned long)xen_io_tlb_start, order);
232 xen_io_tlb_start = NULL;
234 m_ret = XEN_SWIOTLB_EFIXUP;
237 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
239 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
241 panic("Cannot allocate SWIOTLB buffer");
244 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
248 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
249 (xen_io_tlb_nslabs >> 1));
250 pr_info("Lowering to %luMB\n",
251 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
254 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
256 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
258 free_pages((unsigned long)xen_io_tlb_start, order);
262 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
263 dma_addr_t *dma_handle, gfp_t flags,
264 struct dma_attrs *attrs)
267 int order = get_order(size);
268 u64 dma_mask = DMA_BIT_MASK(32);
269 unsigned long vstart;
274 * Ignore region specifiers - the kernel's ideas of
275 * pseudo-phys memory layout has nothing to do with the
276 * machine physical layout. We can't allocate highmem
277 * because we can't return a pointer to it.
279 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
281 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
284 vstart = __get_free_pages(flags, order);
285 ret = (void *)vstart;
290 if (hwdev && hwdev->coherent_dma_mask)
291 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
293 phys = virt_to_phys(ret);
294 dev_addr = xen_phys_to_bus(phys);
295 if (((dev_addr + size - 1 <= dma_mask)) &&
296 !range_straddles_page_boundary(phys, size))
297 *dma_handle = dev_addr;
299 if (xen_create_contiguous_region(vstart, order,
300 fls64(dma_mask)) != 0) {
301 free_pages(vstart, order);
304 *dma_handle = virt_to_machine(ret).maddr;
306 memset(ret, 0, size);
309 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
312 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
313 dma_addr_t dev_addr, struct dma_attrs *attrs)
315 int order = get_order(size);
317 u64 dma_mask = DMA_BIT_MASK(32);
319 if (dma_release_from_coherent(hwdev, order, vaddr))
322 if (hwdev && hwdev->coherent_dma_mask)
323 dma_mask = hwdev->coherent_dma_mask;
325 phys = virt_to_phys(vaddr);
327 if (((dev_addr + size - 1 > dma_mask)) ||
328 range_straddles_page_boundary(phys, size))
329 xen_destroy_contiguous_region((unsigned long)vaddr, order);
331 free_pages((unsigned long)vaddr, order);
333 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
337 * Map a single buffer of the indicated size for DMA in streaming mode. The
338 * physical address to use is returned.
340 * Once the device is given the dma address, the device owns this memory until
341 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
343 dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
344 unsigned long offset, size_t size,
345 enum dma_data_direction dir,
346 struct dma_attrs *attrs)
348 phys_addr_t map, phys = page_to_phys(page) + offset;
349 dma_addr_t dev_addr = xen_phys_to_bus(phys);
351 BUG_ON(dir == DMA_NONE);
353 * If the address happens to be in the device's DMA window,
354 * we can safely return the device addr and not worry about bounce
357 if (dma_capable(dev, dev_addr, size) &&
358 !range_straddles_page_boundary(phys, size) && !swiotlb_force)
362 * Oh well, have to allocate and map a bounce buffer.
364 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
366 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
367 if (map == SWIOTLB_MAP_ERROR)
368 return DMA_ERROR_CODE;
370 dev_addr = xen_phys_to_bus(map);
373 * Ensure that the address returned is DMA'ble
375 if (!dma_capable(dev, dev_addr, size)) {
376 swiotlb_tbl_unmap_single(dev, map, size, dir);
381 EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
384 * Unmap a single streaming mode DMA translation. The dma_addr and size must
385 * match what was provided for in a previous xen_swiotlb_map_page call. All
386 * other usages are undefined.
388 * After this call, reads by the cpu to the buffer are guaranteed to see
389 * whatever the device wrote there.
391 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
392 size_t size, enum dma_data_direction dir)
394 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
396 BUG_ON(dir == DMA_NONE);
398 /* NOTE: We use dev_addr here, not paddr! */
399 if (is_xen_swiotlb_buffer(dev_addr)) {
400 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
404 if (dir != DMA_FROM_DEVICE)
408 * phys_to_virt doesn't work with hihgmem page but we could
409 * call dma_mark_clean() with hihgmem page here. However, we
410 * are fine since dma_mark_clean() is null on POWERPC. We can
411 * make dma_mark_clean() take a physical address if necessary.
413 dma_mark_clean(phys_to_virt(paddr), size);
416 void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
417 size_t size, enum dma_data_direction dir,
418 struct dma_attrs *attrs)
420 xen_unmap_single(hwdev, dev_addr, size, dir);
422 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
425 * Make physical memory consistent for a single streaming mode DMA translation
428 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
429 * using the cpu, yet do not wish to teardown the dma mapping, you must
430 * call this function before doing so. At the next point you give the dma
431 * address back to the card, you must first perform a
432 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
435 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
436 size_t size, enum dma_data_direction dir,
437 enum dma_sync_target target)
439 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
441 BUG_ON(dir == DMA_NONE);
443 /* NOTE: We use dev_addr here, not paddr! */
444 if (is_xen_swiotlb_buffer(dev_addr)) {
445 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
449 if (dir != DMA_FROM_DEVICE)
452 dma_mark_clean(phys_to_virt(paddr), size);
456 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
457 size_t size, enum dma_data_direction dir)
459 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
461 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
464 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
465 size_t size, enum dma_data_direction dir)
467 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
469 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
472 * Map a set of buffers described by scatterlist in streaming mode for DMA.
473 * This is the scatter-gather version of the above xen_swiotlb_map_page
474 * interface. Here the scatter gather list elements are each tagged with the
475 * appropriate dma address and length. They are obtained via
476 * sg_dma_{address,length}(SG).
478 * NOTE: An implementation may be able to use a smaller number of
479 * DMA address/length pairs than there are SG table elements.
480 * (for example via virtual mapping capabilities)
481 * The routine returns the number of addr/length pairs actually
482 * used, at most nents.
484 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
488 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
489 int nelems, enum dma_data_direction dir,
490 struct dma_attrs *attrs)
492 struct scatterlist *sg;
495 BUG_ON(dir == DMA_NONE);
497 for_each_sg(sgl, sg, nelems, i) {
498 phys_addr_t paddr = sg_phys(sg);
499 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
502 !dma_capable(hwdev, dev_addr, sg->length) ||
503 range_straddles_page_boundary(paddr, sg->length)) {
504 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
509 if (map == SWIOTLB_MAP_ERROR) {
510 /* Don't panic here, we expect map_sg users
511 to do proper error handling. */
512 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
515 return DMA_ERROR_CODE;
517 sg->dma_address = xen_phys_to_bus(map);
519 sg->dma_address = dev_addr;
520 sg_dma_len(sg) = sg->length;
524 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
527 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
528 * concerning calls here are the same as for swiotlb_unmap_page() above.
531 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
532 int nelems, enum dma_data_direction dir,
533 struct dma_attrs *attrs)
535 struct scatterlist *sg;
538 BUG_ON(dir == DMA_NONE);
540 for_each_sg(sgl, sg, nelems, i)
541 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
544 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
547 * Make physical memory consistent for a set of streaming mode DMA translations
550 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
554 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
555 int nelems, enum dma_data_direction dir,
556 enum dma_sync_target target)
558 struct scatterlist *sg;
561 for_each_sg(sgl, sg, nelems, i)
562 xen_swiotlb_sync_single(hwdev, sg->dma_address,
563 sg_dma_len(sg), dir, target);
567 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
568 int nelems, enum dma_data_direction dir)
570 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
572 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
575 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
576 int nelems, enum dma_data_direction dir)
578 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
580 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
583 xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
587 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
590 * Return whether the given device DMA address mask can be supported
591 * properly. For example, if your device can only drive the low 24-bits
592 * during bus mastering, then you would pass 0x00ffffff as the mask to
596 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
598 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
600 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);