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 #include <trace/events/swiotlb.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
53 static char *xen_io_tlb_start, *xen_io_tlb_end;
54 static unsigned long xen_io_tlb_nslabs;
56 * Quick lookup value of the bus address of the IOTLB.
59 static u64 start_dma_addr;
61 static dma_addr_t xen_phys_to_bus(phys_addr_t paddr)
63 return phys_to_machine(XPADDR(paddr)).maddr;
66 static phys_addr_t xen_bus_to_phys(dma_addr_t baddr)
68 return machine_to_phys(XMADDR(baddr)).paddr;
71 static dma_addr_t xen_virt_to_bus(void *address)
73 return xen_phys_to_bus(virt_to_phys(address));
76 static int check_pages_physically_contiguous(unsigned long pfn,
80 unsigned long next_mfn;
84 next_mfn = pfn_to_mfn(pfn);
85 nr_pages = (offset + length + PAGE_SIZE-1) >> PAGE_SHIFT;
87 for (i = 1; i < nr_pages; i++) {
88 if (pfn_to_mfn(++pfn) != ++next_mfn)
94 static int range_straddles_page_boundary(phys_addr_t p, size_t size)
96 unsigned long pfn = PFN_DOWN(p);
97 unsigned int offset = p & ~PAGE_MASK;
99 if (offset + size <= PAGE_SIZE)
101 if (check_pages_physically_contiguous(pfn, offset, size))
106 static int is_xen_swiotlb_buffer(dma_addr_t dma_addr)
108 unsigned long mfn = PFN_DOWN(dma_addr);
109 unsigned long pfn = mfn_to_local_pfn(mfn);
112 /* If the address is outside our domain, it CAN
113 * have the same virtual address as another address
114 * in our domain. Therefore _only_ check address within our domain.
116 if (pfn_valid(pfn)) {
117 paddr = PFN_PHYS(pfn);
118 return paddr >= virt_to_phys(xen_io_tlb_start) &&
119 paddr < virt_to_phys(xen_io_tlb_end);
124 static int max_dma_bits = 32;
127 xen_swiotlb_fixup(void *buf, size_t size, unsigned long nslabs)
132 dma_bits = get_order(IO_TLB_SEGSIZE << IO_TLB_SHIFT) + PAGE_SHIFT;
136 int slabs = min(nslabs - i, (unsigned long)IO_TLB_SEGSIZE);
139 rc = xen_create_contiguous_region(
140 (unsigned long)buf + (i << IO_TLB_SHIFT),
141 get_order(slabs << IO_TLB_SHIFT),
143 } while (rc && dma_bits++ < max_dma_bits);
148 } while (i < nslabs);
151 static unsigned long xen_set_nslabs(unsigned long nr_tbl)
154 xen_io_tlb_nslabs = (64 * 1024 * 1024 >> IO_TLB_SHIFT);
155 xen_io_tlb_nslabs = ALIGN(xen_io_tlb_nslabs, IO_TLB_SEGSIZE);
157 xen_io_tlb_nslabs = nr_tbl;
159 return xen_io_tlb_nslabs << IO_TLB_SHIFT;
162 enum xen_swiotlb_err {
163 XEN_SWIOTLB_UNKNOWN = 0,
168 static const char *xen_swiotlb_error(enum xen_swiotlb_err err)
171 case XEN_SWIOTLB_ENOMEM:
172 return "Cannot allocate Xen-SWIOTLB buffer\n";
173 case XEN_SWIOTLB_EFIXUP:
174 return "Failed to get contiguous memory for DMA from Xen!\n"\
175 "You either: don't have the permissions, do not have"\
176 " enough free memory under 4GB, or the hypervisor memory"\
177 " is too fragmented!";
183 int __ref xen_swiotlb_init(int verbose, bool early)
185 unsigned long bytes, order;
187 enum xen_swiotlb_err m_ret = XEN_SWIOTLB_UNKNOWN;
188 unsigned int repeat = 3;
190 xen_io_tlb_nslabs = swiotlb_nr_tbl();
192 bytes = xen_set_nslabs(xen_io_tlb_nslabs);
193 order = get_order(xen_io_tlb_nslabs << IO_TLB_SHIFT);
195 * Get IO TLB memory from any location.
198 xen_io_tlb_start = alloc_bootmem_pages(PAGE_ALIGN(bytes));
200 #define SLABS_PER_PAGE (1 << (PAGE_SHIFT - IO_TLB_SHIFT))
201 #define IO_TLB_MIN_SLABS ((1<<20) >> IO_TLB_SHIFT)
202 while ((SLABS_PER_PAGE << order) > IO_TLB_MIN_SLABS) {
203 xen_io_tlb_start = (void *)__get_free_pages(__GFP_NOWARN, order);
204 if (xen_io_tlb_start)
208 if (order != get_order(bytes)) {
209 pr_warn("Warning: only able to allocate %ld MB for software IO TLB\n",
210 (PAGE_SIZE << order) >> 20);
211 xen_io_tlb_nslabs = SLABS_PER_PAGE << order;
212 bytes = xen_io_tlb_nslabs << IO_TLB_SHIFT;
215 if (!xen_io_tlb_start) {
216 m_ret = XEN_SWIOTLB_ENOMEM;
219 xen_io_tlb_end = xen_io_tlb_start + bytes;
221 * And replace that memory with pages under 4GB.
223 rc = xen_swiotlb_fixup(xen_io_tlb_start,
228 free_bootmem(__pa(xen_io_tlb_start), PAGE_ALIGN(bytes));
230 free_pages((unsigned long)xen_io_tlb_start, order);
231 xen_io_tlb_start = NULL;
233 m_ret = XEN_SWIOTLB_EFIXUP;
236 start_dma_addr = xen_virt_to_bus(xen_io_tlb_start);
238 if (swiotlb_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs,
240 panic("Cannot allocate SWIOTLB buffer");
243 rc = swiotlb_late_init_with_tbl(xen_io_tlb_start, xen_io_tlb_nslabs);
247 xen_io_tlb_nslabs = max(1024UL, /* Min is 2MB */
248 (xen_io_tlb_nslabs >> 1));
249 pr_info("Lowering to %luMB\n",
250 (xen_io_tlb_nslabs << IO_TLB_SHIFT) >> 20);
253 pr_err("%s (rc:%d)\n", xen_swiotlb_error(m_ret), rc);
255 panic("%s (rc:%d)", xen_swiotlb_error(m_ret), rc);
257 free_pages((unsigned long)xen_io_tlb_start, order);
261 xen_swiotlb_alloc_coherent(struct device *hwdev, size_t size,
262 dma_addr_t *dma_handle, gfp_t flags,
263 struct dma_attrs *attrs)
266 int order = get_order(size);
267 u64 dma_mask = DMA_BIT_MASK(32);
268 unsigned long vstart;
273 * Ignore region specifiers - the kernel's ideas of
274 * pseudo-phys memory layout has nothing to do with the
275 * machine physical layout. We can't allocate highmem
276 * because we can't return a pointer to it.
278 flags &= ~(__GFP_DMA | __GFP_HIGHMEM);
280 if (dma_alloc_from_coherent(hwdev, size, dma_handle, &ret))
283 vstart = __get_free_pages(flags, order);
284 ret = (void *)vstart;
289 if (hwdev && hwdev->coherent_dma_mask)
290 dma_mask = dma_alloc_coherent_mask(hwdev, flags);
292 phys = virt_to_phys(ret);
293 dev_addr = xen_phys_to_bus(phys);
294 if (((dev_addr + size - 1 <= dma_mask)) &&
295 !range_straddles_page_boundary(phys, size))
296 *dma_handle = dev_addr;
298 if (xen_create_contiguous_region(vstart, order,
299 fls64(dma_mask)) != 0) {
300 free_pages(vstart, order);
303 *dma_handle = virt_to_machine(ret).maddr;
305 memset(ret, 0, size);
308 EXPORT_SYMBOL_GPL(xen_swiotlb_alloc_coherent);
311 xen_swiotlb_free_coherent(struct device *hwdev, size_t size, void *vaddr,
312 dma_addr_t dev_addr, struct dma_attrs *attrs)
314 int order = get_order(size);
316 u64 dma_mask = DMA_BIT_MASK(32);
318 if (dma_release_from_coherent(hwdev, order, vaddr))
321 if (hwdev && hwdev->coherent_dma_mask)
322 dma_mask = hwdev->coherent_dma_mask;
324 phys = virt_to_phys(vaddr);
326 if (((dev_addr + size - 1 > dma_mask)) ||
327 range_straddles_page_boundary(phys, size))
328 xen_destroy_contiguous_region((unsigned long)vaddr, order);
330 free_pages((unsigned long)vaddr, order);
332 EXPORT_SYMBOL_GPL(xen_swiotlb_free_coherent);
336 * Map a single buffer of the indicated size for DMA in streaming mode. The
337 * physical address to use is returned.
339 * Once the device is given the dma address, the device owns this memory until
340 * either xen_swiotlb_unmap_page or xen_swiotlb_dma_sync_single is performed.
342 dma_addr_t xen_swiotlb_map_page(struct device *dev, struct page *page,
343 unsigned long offset, size_t size,
344 enum dma_data_direction dir,
345 struct dma_attrs *attrs)
347 phys_addr_t map, phys = page_to_phys(page) + offset;
348 dma_addr_t dev_addr = xen_phys_to_bus(phys);
350 BUG_ON(dir == DMA_NONE);
352 * If the address happens to be in the device's DMA window,
353 * we can safely return the device addr and not worry about bounce
356 if (dma_capable(dev, dev_addr, size) &&
357 !range_straddles_page_boundary(phys, size) && !swiotlb_force)
361 * Oh well, have to allocate and map a bounce buffer.
363 trace_swiotlb_bounced(dev, dev_addr, size, swiotlb_force);
365 map = swiotlb_tbl_map_single(dev, start_dma_addr, phys, size, dir);
366 if (map == SWIOTLB_MAP_ERROR)
367 return DMA_ERROR_CODE;
369 dev_addr = xen_phys_to_bus(map);
372 * Ensure that the address returned is DMA'ble
374 if (!dma_capable(dev, dev_addr, size)) {
375 swiotlb_tbl_unmap_single(dev, map, size, dir);
380 EXPORT_SYMBOL_GPL(xen_swiotlb_map_page);
383 * Unmap a single streaming mode DMA translation. The dma_addr and size must
384 * match what was provided for in a previous xen_swiotlb_map_page call. All
385 * other usages are undefined.
387 * After this call, reads by the cpu to the buffer are guaranteed to see
388 * whatever the device wrote there.
390 static void xen_unmap_single(struct device *hwdev, dma_addr_t dev_addr,
391 size_t size, enum dma_data_direction dir)
393 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
395 BUG_ON(dir == DMA_NONE);
397 /* NOTE: We use dev_addr here, not paddr! */
398 if (is_xen_swiotlb_buffer(dev_addr)) {
399 swiotlb_tbl_unmap_single(hwdev, paddr, size, dir);
403 if (dir != DMA_FROM_DEVICE)
407 * phys_to_virt doesn't work with hihgmem page but we could
408 * call dma_mark_clean() with hihgmem page here. However, we
409 * are fine since dma_mark_clean() is null on POWERPC. We can
410 * make dma_mark_clean() take a physical address if necessary.
412 dma_mark_clean(phys_to_virt(paddr), size);
415 void xen_swiotlb_unmap_page(struct device *hwdev, dma_addr_t dev_addr,
416 size_t size, enum dma_data_direction dir,
417 struct dma_attrs *attrs)
419 xen_unmap_single(hwdev, dev_addr, size, dir);
421 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_page);
424 * Make physical memory consistent for a single streaming mode DMA translation
427 * If you perform a xen_swiotlb_map_page() but wish to interrogate the buffer
428 * using the cpu, yet do not wish to teardown the dma mapping, you must
429 * call this function before doing so. At the next point you give the dma
430 * address back to the card, you must first perform a
431 * xen_swiotlb_dma_sync_for_device, and then the device again owns the buffer
434 xen_swiotlb_sync_single(struct device *hwdev, dma_addr_t dev_addr,
435 size_t size, enum dma_data_direction dir,
436 enum dma_sync_target target)
438 phys_addr_t paddr = xen_bus_to_phys(dev_addr);
440 BUG_ON(dir == DMA_NONE);
442 /* NOTE: We use dev_addr here, not paddr! */
443 if (is_xen_swiotlb_buffer(dev_addr)) {
444 swiotlb_tbl_sync_single(hwdev, paddr, size, dir, target);
448 if (dir != DMA_FROM_DEVICE)
451 dma_mark_clean(phys_to_virt(paddr), size);
455 xen_swiotlb_sync_single_for_cpu(struct device *hwdev, dma_addr_t dev_addr,
456 size_t size, enum dma_data_direction dir)
458 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_CPU);
460 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_cpu);
463 xen_swiotlb_sync_single_for_device(struct device *hwdev, dma_addr_t dev_addr,
464 size_t size, enum dma_data_direction dir)
466 xen_swiotlb_sync_single(hwdev, dev_addr, size, dir, SYNC_FOR_DEVICE);
468 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_single_for_device);
471 * Map a set of buffers described by scatterlist in streaming mode for DMA.
472 * This is the scatter-gather version of the above xen_swiotlb_map_page
473 * interface. Here the scatter gather list elements are each tagged with the
474 * appropriate dma address and length. They are obtained via
475 * sg_dma_{address,length}(SG).
477 * NOTE: An implementation may be able to use a smaller number of
478 * DMA address/length pairs than there are SG table elements.
479 * (for example via virtual mapping capabilities)
480 * The routine returns the number of addr/length pairs actually
481 * used, at most nents.
483 * Device ownership issues as mentioned above for xen_swiotlb_map_page are the
487 xen_swiotlb_map_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
488 int nelems, enum dma_data_direction dir,
489 struct dma_attrs *attrs)
491 struct scatterlist *sg;
494 BUG_ON(dir == DMA_NONE);
496 for_each_sg(sgl, sg, nelems, i) {
497 phys_addr_t paddr = sg_phys(sg);
498 dma_addr_t dev_addr = xen_phys_to_bus(paddr);
501 !dma_capable(hwdev, dev_addr, sg->length) ||
502 range_straddles_page_boundary(paddr, sg->length)) {
503 phys_addr_t map = swiotlb_tbl_map_single(hwdev,
508 if (map == SWIOTLB_MAP_ERROR) {
509 /* Don't panic here, we expect map_sg users
510 to do proper error handling. */
511 xen_swiotlb_unmap_sg_attrs(hwdev, sgl, i, dir,
514 return DMA_ERROR_CODE;
516 sg->dma_address = xen_phys_to_bus(map);
518 sg->dma_address = dev_addr;
519 sg_dma_len(sg) = sg->length;
523 EXPORT_SYMBOL_GPL(xen_swiotlb_map_sg_attrs);
526 * Unmap a set of streaming mode DMA translations. Again, cpu read rules
527 * concerning calls here are the same as for swiotlb_unmap_page() above.
530 xen_swiotlb_unmap_sg_attrs(struct device *hwdev, struct scatterlist *sgl,
531 int nelems, enum dma_data_direction dir,
532 struct dma_attrs *attrs)
534 struct scatterlist *sg;
537 BUG_ON(dir == DMA_NONE);
539 for_each_sg(sgl, sg, nelems, i)
540 xen_unmap_single(hwdev, sg->dma_address, sg_dma_len(sg), dir);
543 EXPORT_SYMBOL_GPL(xen_swiotlb_unmap_sg_attrs);
546 * Make physical memory consistent for a set of streaming mode DMA translations
549 * The same as swiotlb_sync_single_* but for a scatter-gather list, same rules
553 xen_swiotlb_sync_sg(struct device *hwdev, struct scatterlist *sgl,
554 int nelems, enum dma_data_direction dir,
555 enum dma_sync_target target)
557 struct scatterlist *sg;
560 for_each_sg(sgl, sg, nelems, i)
561 xen_swiotlb_sync_single(hwdev, sg->dma_address,
562 sg_dma_len(sg), dir, target);
566 xen_swiotlb_sync_sg_for_cpu(struct device *hwdev, struct scatterlist *sg,
567 int nelems, enum dma_data_direction dir)
569 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_CPU);
571 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_cpu);
574 xen_swiotlb_sync_sg_for_device(struct device *hwdev, struct scatterlist *sg,
575 int nelems, enum dma_data_direction dir)
577 xen_swiotlb_sync_sg(hwdev, sg, nelems, dir, SYNC_FOR_DEVICE);
579 EXPORT_SYMBOL_GPL(xen_swiotlb_sync_sg_for_device);
582 xen_swiotlb_dma_mapping_error(struct device *hwdev, dma_addr_t dma_addr)
586 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_mapping_error);
589 * Return whether the given device DMA address mask can be supported
590 * properly. For example, if your device can only drive the low 24-bits
591 * during bus mastering, then you would pass 0x00ffffff as the mask to
595 xen_swiotlb_dma_supported(struct device *hwdev, u64 mask)
597 return xen_virt_to_bus(xen_io_tlb_end - 1) <= mask;
599 EXPORT_SYMBOL_GPL(xen_swiotlb_dma_supported);